A spectral approach to bandwidth

and separator problems in graphs

Christoph Helmberg

Bojan Mohar

Svatopluk Poljak

Franz Rendl

Abstract
Lower bounds on the bandwidth, the size of a vertex separator of general undirected graphs,
and the largest common subgraph of two undirected (weighted) graphs are obtained. The
bounds are based on a projection technique developed for the quadratic assignment problem,
and once more demonstrate the importance of the extreme eigenvalues of the Laplacian.
They will be shown to be strict for certain classes of graphs and compare favourably to
bounds already known in literature. Further improvement is gained by applying nonlinear
optimization methods.
Keywords: bandwidth, vertex separator, Laplacian eigenvalues
Math. Subj. Class. (1991): 05 C 50, 05 C 78, 90 C 35.
1 Introduction
The bandwidth of a graph is a graph invariant that is computationally extremely dicult to
determine. A survey on the bandwidth of graphs and related problems is contained in e.g. [2].
Heuristics to nd good labelings of vertices of graphs were developed by engineers as early as in
the 60s to facilitate Finite Element calculations of structural design problems, see e.g. [17]. To
estimate how close these labelings come to the minimal bandwidth, it is important to have good
lower bounds on the minimal bandwidth. Recently, spectral properties of graphs were used to
bound the bandwidth from below, see e.g [10, 11].

University of Technology Graz, Department of Mathematics, Kopernikusgasse 24, A - 8010 Graz, Austria.

University of Ljubljana, Department of Mathematics, Jadranska 19, 61111 Ljubljana, Slovenia.

Charles University Prague, Department of Applied Mathematics, Malostranske n. 25, 118 00 Prague, Czech
Republic.
1
In the present paper we strengthen the approach described in [10, 11] and derive new lower
bounds for graph invariants related to the bandwidth. In particular, we get bounds on the size
of balanced vertex separators in graphs.
The paper is organized as follows. In Section 2 we derive the mathematical tools necessary
to obtain the new bounds. These tools are applied in Section 3 to the bandwidth problem,
and in Section 4 to the vertex separator problem. In particular, we present two bounds on the
bandwidth problem. The rst bound, given in Theorem 8, uses only two extreme eigenvalues of
the Laplacian matrix. The second bound, given in Theorem 9, is in the form of an eciently
solvable convex optimization problem, and makes use of the whole Laplacian spectrum. In Section
5 we shortly discuss the problem of nding the largest common subgraph of two given graphs
whose special cases can also be applied to get some information on the bandwidth problem.
Section 6 contains applications of our approach to the closely related 1-sum problem. We show
that our new bound improves an earlier bound by Juvan and Mohar [10]. We close with some
computational experiments in Section 7.
2 The tools
Let G be an undirected weighted graph having (weighted) adjacency matrix A. For S
1
, S
2
V (G)
denote by
cut(S
1
, S
2
) :=

iS
1
,jS
2
a
ij
the total weight of edges between the sets S
1
and S
2
. Lower bounds on cut(S
1
, S
2
) for all (disjoint)
sets S
1
, S
2
of specied cardinalities [S
i
[ = m
i
(i = 1, 2) play a fundamental role in the analysis of
several graph problems, like bounding the bandwidth, or nding vertex separators. The following
theorem provides a lower bound on cut(S
1
, S
2
) in terms of m
1
, m
2
and the extreme Laplacian
eigenvalues of G. The approach is based on partitioning results from [16]. Given numbers n,
2
,
and
n
dene the following symmetric function on positive integers t
1
, t
2
smaller than n:
f(t
1
, t
2
) :=

t
1
t
2
2n
_
_

t
1
t
2
+
_
(n t
1
)(n t
2
)
_

2
+
_

t
1
t
2

_
(n t
1
)(n t
2
)
_

n
_
(1)
Theorem 1 Let G be an undirected graph of order n. Let
2
and
n
denote the second small-
est and the largest Laplacian eigenvalue of G, respectively. If m
1
, m
2
, m
3
are positive integers
summing up to n, then
cut(S
1
, S
2
) f(m
1
, m
2
)
for all partitions (S
1
, S
2
, S
3
) of V (G) into subsets of cardinalities [S
i
[ = m
i
, i = 1, 2, 3.
Before proving this theorem, we need some auxiliary results. Following [16], we represent a
partition (S
1
, S
2
, S
3
) of V (G) by an associated n 3 matrix X, where
x
ij
=
_
1 if i S
j
0 if i , S
j
.
2
The columns of X will be denoted by x
1
, x
2
, x
3
. Let
B =
_
_
_
_
0 1 0
1 0 0
0 0 0
_
_
_
_
.
It is easy to see that tr X
t
AXB = x
t
1
Ax
2
+x
t
2
Ax
1
which in turn implies that
cut(S
1
, S
2
) =
1
2
tr X
t
AXB. (2)
In [16] it is shown that X describes a partition of V (G) into subsets of specied sizes m =
(m
1
, m
2
, m
3
)
t
if and only if
X
t
u
n
= m, Xu
3
= u
n
, (3)
X
t
X = M, (4)
X 0 elementwise. (5)
Throughout the paper we will denote by u, (or u
n
to indicate the size) a vector of all ones, I
n
the identity matrix of order n, M := diag(m), and V will be an n (n 1) matrix satisfying:
V
t
V = I
n1
, V
t
u
n
= 0. (6)
The eigenvalues of a symmetric matrix Q of order n will be denoted by
1
(Q) . . .
n
(Q).
Finally, let m := (

m
1
,

m
2
,

m
3
)
t
, and let W be a 3 2 matrix satisfying:
W
t
W = I
2
, W
t
m = 0.
Then
WW
t
= I
3

1
n
mm
t
. (7)
The following result was proved in [16].
Proposition 2 A matrix X satises (3), (4) if and only if there exists an (n 1) 2 matrix Z
with Z
t
Z = I
2
such that
X =
1
n
u
n
u
t
3
M +V ZW
t
M
1/2
.
It is also easy to see that
tr X
t
(A+D)XB = tr X
t
AXB (8)
for all diagonal matrices D, and all matrices X representing partitions. Let s(A) := u
t
Au, and
r(A) := Au. The following choice of D will turn out to be particularly useful:
D :=
s(A)
n
I
n
diag(r(A)). (9)
Note that tr D = 0. To prove Theorem 1 we need to know the eigenvalues of the 2 2 matrix

M :=

m
1
m
2
W
t
BW.
3
Lemma 3 The eigenvalues
1

2
of the matrix

M are given by

1,2
=
1
n
_
m
1
m
2

_
m
1
m
2
(n m
1
)(n m
2
)
_
.
Proof. We observe that
1
and
2
are characterized by

1
+
2
= tr

M =
2m
1
m
2
n
,

2
1
+
2
2
= tr

M
2
=
2m
1
m
2
n
_
m
3
+
2m
1
m
2
n
_
.
After substituting the given values for
1
and
2
, the lemma follows.
We point out that
f(m
1
, m
2
) =
1
2

2

1
2

n
. (10)
Next we relate the eigenvalues of

A := V
t
(A+D)V
to the Laplacian eigenvalues of G. Let
1
. . .
n
denote the eigenvalues of the Laplacian
L := diag(r(A)) A of our weighted graph G.
Lemma 4 The eigenvalues of the matrix V
t
(A+D)V are

i
=
s(A)
n

i
, i = 2, 3, . . . , n.
Proof. First note that V
t
(A+D)V = V
t
(
s(A)
n
I L)V . Let x be an eigenvector for
i
, where
i > 1. Then x u, since u is eigenvector of
1
. By (6), there exists a nonzero vector y such that
x = V y. We conclude that
V
t
(A+D)V y =
s(A)
n
y V
t
L(V y)
=
s(A)
n
y V
t
(
i
V y)
=
_
s(A)
n

i
_
y.
We will also make use of the following two variations of the HomanWielandt inequality. We
formulate these results as optimization problems, where the optimal objective function value is
given by the eigenvalues of the underlying matrices. These results were rst used to get tractable
relaxations of graph partition problems and quadratic assignment problems.
Theorem 5 Let A and B be symmetric matrices of order n and k, respectively, and suppose that
k n. Then
mintr AXBX
t
: X
t
X = I
k
= min
_
k

i=1

i
(B)
(i)
(A) : injection
_
.
4
Proof. A similar result has been shown in [12] and is also implicitly contained in e.g. [6],
where the theorem is proved for the case k = n, and both matrices positive denite. The proof
easily generalizes to the present situation. For the sake of completeness we include the following
argument and note that a similar result is also proved in [14].
We use the spectral decompositions of A and B,
A = PEP
t
, B = QFQ
t
,
with orthogonal matrices P, Q and diagonal matrices E, F of appropriate sizes. Note that the
matrix Y := P
t
XQ has orthonormal columns, so we get:
tr AXBX
t
= tr EY FY
t
(11)
=
n

i=1
k

j=1

i
(A)
j
(B)y
2
ij
(12)
min
_

i,j

i
(A)
j
(B)z
ij
:

i
z
ij
= 1,

j
z
ij
1, z
ij
0
_
(13)
= min
_
k

i=1

i
(B)
(i)
(A) : injective
_
. (14)
The last equality follows from the fact that the transportation constraints on the z
ij
are
totally unimodular. Therefore there exists an optimal 0-1 solution to the minimization problem,
and these are characterized by injections : 1, . . . , k 1, . . . , n.
To see that equality holds, it is sucient to take an ordering of the eigenvalues and eigenvectors
in E, F, P and Q given by the minimal injection and using X = PQ
t
to get attainment for the
lower bound. In particular, P contains only the eigenvectors corresponding to the k eigenvalues
of the image of in the correct ordering.
In the subsequent theorem we use the following sets of matrices: The set O := X : X
t
X = I
n

contains orthogonal matrices, and c := X : Xu = X

t
u = u contains all matrices with row and
column sums equal to one. Finally, the set of permutation matrices is denoted by .
Theorem 6 [8]. Let A and B be symmetric matrices of order n, and suppose that Au = u.
Then
maxtr AXBX
t
: X maxtr AXBX
t
: X O c
=
n1

i=1

i
(V
t
AV )
i
(V
t
BV ) +
s(B)
n
. (15)
We are now ready to prove Theorem 1.
Proof of Theorem 1. Using equalities (2), (9), and (8), and the characterization (3), (4), (5) for
partitions X, we conclude that
cut(S
1
, S
2
) min
_
1
2
tr X
t
(A+D)XB : X satises (3), (4)
_
=: g(m
1
, m
2
) (16)
5
for all partitions (S
1
, S
2
, S
3
) with subsets of specied sizes m
1
, m
2
, m
3
. We will show that
g(m
1
, m
2
) = f(m
1
, m
2
). Substituting the parametrization of X from Proposition 2 we obtain
after some rearrangement of terms
g(m
1
, m
2
) = min
_
m
1
m
2
s(A)
n
2
+
1
2
tr

AZ

MZ
t
: Z
t
Z = I
2
_
.
We note in particular that due to the special choice of D we have
V
t
(A+D)u = 0 .
Therefore the terms linear in Z vanish. The minimum is given by Theorem 5. By Lemma 3,

2
0
1
. Thus, using Theorem 6 we get
mintr

AZ

MZ
t
: Z
t
Z = I
2
=
2

max
(

A) +
1

min
(

A).
A simple calculation using the spectral information from Lemma 4 shows that the terms
containing s(A) cancel, leaving the sum (10) of products of eigenvalues.
In the sequel, Theorem 6 will be applied several times in situations when one of the matrices
A or B is the Laplacian matrix L of a (weighted) graph. In such cases, expression (15) becomes
particularly simple, since the spectrum of V
t
LV consists of the eigenvalues of L but the eigenvalue
zero. Let us recall that the Laplacian matrix L = L(G) of a (weighted) graph G is dened by
L(G) = diag(r(A)) A where A is the (weighted) adjacency matrix of G, and r(A) is the vector
of the row sums of A. The following lemma follows easily from the proof of Lemma 4.
Lemma 7 The Laplacian matrix L(G) has the following properties.
(i) If the edge weights of G are nonnegative, then L(G) is positive semidenite.
(ii) u = (1, . . . , 1)
t
is an eigenvector of L(G) corresponding to the eigenvalue 0.
(iii) The spectrum of V
t
L(G)V (where V is the projection matrix satisfying (6)) consists of
the eigenvalues of L where the multiplicity of the eigenvalue zero is decreased by one.
3 Bounding the bandwidth
The bandwidth

(G) of an nvertex graph G is dened as

(G) := minmax
ijE
[(i) (j)[ : bijection V (G) 1, . . . , n . (17)
Determining

(G) is computationally an extremely dicult task. Several lower bounds on

(G) expressed in terms of the Laplacian eigenvalues of G are obtained in [10] and further
6
developed and improved in [11]. A basic proof technique to derive lower bounds on

(G)
consists in showing that
cut(S
1
, S
2
) > 0
for all S
1
, S
2
with [S
1
[ = m
1
> 0, [S
2
[ = m
2
> 0. If this is the case, it follows that

(G) (n m
1
m
2
) + 1, (18)
because using a bijection where the minimum in (17) is attained, there are edges between
S
1
:=
1
(1), . . . ,
1
(m
1
) and S
2
:=
1
(n m
2
+ 1), . . . ,
1
(n). Using Theorem 1, we
obtain the following lower bound on

(G).
Theorem 8 Let G be an undirected graph on n vertices with at least one edge. Let
2
and
n
denote the second smallest and the largest Laplacian eigenvalue of G, respectively. Let be the
largest integer smaller than n
2
/
n
.
(a) If n 2, then G = K
n
and

(G) = n 1.
(b) If n 2 and n = 0 mod(2), then

(G) + 1.
(c) Otherwise

(G) .
Proof. To prove (a) we note that we can set m
1
= m
2
= 1 to get
cut(S
1
, S
2
) f(1, 1) =

n
2n
_
n
2

n
(n 2)
_
> 0
and therefore the bandwidth is n 1, so G is the complete graph.
In case (b) we rst note that in the special case = 0 the theorem is trivially true because
G contains edges. Therefore consider n 2 > 0 and n = 2s > 0. Applying Theorem 1
with m
1
= m
2
= s > 0 and m
3
= > 0 we obtain
cut(S
1
, S
2
) f(s, s) =
s
n
2n
_
n
2

_
> 0.
Therefore

(G) + 1.
Finally suppose n = 2s +1 and s > 0. We set m
1
= m
2
= s +1 and m
3
= 1 to show
similarly that in this case

(G) .
We note that (c) could be improved slightly by testing whether f(s + 1, s) > 0 in which case

(G) + 1.
An alternate way to derive bounds on the bandwidth consists in using the Quadratic Assign-
ment Problem (QAP) as a relaxation. The QAP consists in optimizing
tr AXBX
t
over the set of permutation matrices X. We assume that the matrices A and B, dening QAP,
are symmetric.
7
As before, we denote by L the Laplacian matrix of a given graph G. Although G is unweighted,
we may assign arbitrary weights to its edges, and then let L be the weighted Laplacian. To test
whether

(G) > k, we dene the following set of n n matrices:

B
k
:= B : B = B
t
, b
ij
= 0 if [i j[ k.
Theorem 9 Let L denote an arbitrary weighted Laplacian matrix of a graph G. If there exists
a matrix B B
k
such that
n1

i=1

i+1
(L)
i
(V
t
BV ) < 0 , (19)
then

(G) > k.
Proof. Suppose that

(G) k and let B B

k
be arbitrary. Let the vertices of G be labeled
such that l
ij
= 0 whenever [i j[ > k. Then we can apply Theorem 6 to the matrices L and B,
since = 0 is eigenvalue of L for the eigenvector u. Using Lemma 7 (iii) we get
0 = tr LB maxtr LXBX
t
: X O c =
n1

i=1

i+1
(L)
i
(V
t
BV ) .
This shows that (19) is not possible.
Consider the following matrix

B =
_
_
_
_
0
m
1
m
1
0
m
1
m
3
1
m
1
m
2
0
m
3
m
1
0
m
3
m
3
0
m
3
m
2
1
m
2
m
1
0
m
2
m
3
0
m
2
m
2
_
_
_
_
where 0
ij
and 1
ij
denotes the i j matrix of all zeros and all ones, respectively. Obviously,

B B
m
3
. Using an argument similar to the proof of Lemma 3, one easily nds that the nonzero
eigenvalues of the matrix V
t

BV are equal to
1
and
2
as given by Lemma 3. Therefore, we
conclude that
min
BB
m
3
QAP(L, B) QAP(L,

B) =
2

2
+
1

n
= 2f(m
1
, m
2
),
where the last equation is just (10). This shows that the bound given by Theorem 8 is contained
in Theorem 9.
We conclude with another application of the Quadratic Assignment model to derive an upper
bound on the bandwidth.
Let G be a nonempty graph of order n with m edges. Let A = (a
ij
) denote the adjacency
matrix of G, and let B = (b
ij
) be an n n symmetric matrix dened by b
ij
= m
|ij|1
for i ,= j
and b
ii
= 0. Let
QAP(A, B) := max

i,j
a
ij
b
(i)(j)
denote the value of the optimum solution of the quadratic assignment problem with matrices A
and B. We have the following result.
8
Proposition 10 The bandwidth of G is at most k if and only if QAP(A, B) 2m
k
. Hence,

(G) =
_
log
m
1
2
QAP(A, B)
_

_
log
m
1
2
n1

i=1

i+1
(L(G))
i
(V
t
BV )
_
.
Proof. The statement is obvious for m = 1. Hence we will assume that G has at least two
edges. Assume that the bandwidth is at most k. Then there exists a permutation such that
[(i)(j)[ k for all edges ij. Thus, a
ij
b
(i)(j)
m
k1
for every pair i, j, and

i,j
a
ij
b
(i)(j)

2m m
k1
= 2m
k
. Conversely, assume that

(G) > k. Then, for every permutation , there

exists an edge i
0
j
0
such that [(i
0
) (j
0
)[ > k, i.e. a
i
0
j
0
b
(i
0
)(j
0
)
m
k
. Since m > 1,
we conclude that

i,j
a
ij
b
(i)(j)
> 2m
k
. This proves the rst statement. As an immediate
consequence we get

(G) =
_
log
m
1
2
QAP(A, B)
_
. We will estimate QAP(A, B) as follows. Let
L = L(G) denote the Laplacian matrix of G. Since the diagonal entries of B are zero, we have
QAP(A, B) = QAP(L, B). By Lemma 7 (iii), the eigenvalues of V
t
LV are
2
, . . . ,
n
. Since
Lu = 0 u, applying Theorem 6 (with = 0) yields the upper bound
QAP(L, B) = maxtr AXBX
t
: X
n1

i=1

ni+1
(L(G))
i
(V
t
BV ).
4 Lower Bounds on Vertex Separators
A vertex separator S V partitions the vertices of a graph G into three sets S
1
, S
2
and S
3
= S
such that no vertex of S
1
is adjacent to a vertex of S
2
. In applications S
3
should be a small
set separating two large sets of roughly the same size. Here, S
3
is called a balanced separator if
2[S
1
[ [S
2
[ and 2[S
2
[ [S
1
[. By means of Theorem 1 we shall derive a new lower bound on the
cardinality of such separators.
Suppose S
3
separates S
1
and S
2
and let m
1
= [S
1
[ and m
2
= [S
2
[. Then cut(S
1
, S
2
) = 0, and
Theorem 1 yields a lower bound on [S
3
[.
Theorem 11 Let G be a graph on n vertices. Let
2
and
n
denote the second smallest and the
largest Laplacian eigenvalue of G, respectively. If S
3
V separates vertex sets S
1
, S
2
then
[S
3
[
4
n

2
[S
1
[[S
2
[
n(
n

2
)
2
. (20)
Proof. Let m
i
= [S
i
[ for i 1, 2, 3 and suppose S
3
separates S
1
and S
2
. By Theorem 1,
0 = cut(S
1
, S
2
) f(m
1
, m
2
). Therefore we conclude:
(m
1
+m
3
)(m
2
+m
3
)(
n

2
)
2
m
1
m
2
(
n
+
2
)
2
or equivalently
nm
3

m
1
m
2
((
n
+
2
)
2
(
n

2
)
2
)
(
n

2
)
2
,
9
and (20) follows immediately.
For a discussion of the tightness of this bound we shall need the following binary graph
operations. Let G
1
and G
2
be unweighted vertex disjoint graphs. Denote by G
1
G
2
their union,
and let G
1
G
2
be their join (obtained from G
1
G
2
by joining every vertex of G
1
with every
vertex of G
2
). The characteristic polynomial (G, ) =

n
i=1
(
i
(A(G))) of the resulting graph
G can be expressed in terms of the characteristic polynomials of G
1
and G
2
. With n = n
1
+n
2
,
(G
1
G
2
, ) = (G
1
, ) (G
2
, ) ,
(G
1
G
2
, ) =
( n)
( n
1
)( n
2
)
(G
1
, n
2
) (G
2
, n
1
) .
Consider the class of graphs of the form G = (G
1
G
2
) G
3
where the G
i
are graphs of orders
m
1
, m
2
, m
3
, respectively. The smallest possible separator is of size m
3
provided m
3
m
1
+ m
2
.
By the above formulas, if m
3
m
1
+ m
2
, then
n
(G) = n = m
1
+m
2
+m
3
and
2
(G) = m
3
.
Thus, (20) is tight whenever m
1
= m
2
.
To deduce a lower bound for balanced separators from (20), we parameterize m
1
and m
2
by
m
1
= t(n m
3
) and m
2
= (1 t)(n m
3
) for t [
1
3
,
2
3
]. The minimum of the right hand side of
(20) with respect to t is obtained for t
1
3
,
2
3
. For balanced separators the bound now reads
(m
3
= [S
3
[):
m
3

8
9

2
(
n

2
)
2
(n m
3
)
2
n
.
By solving the quadratic equation for m
3
, we get the following:
Corollary 12 Let G be a graph on n vertices. Let
2
and
n
denote the second smallest and the
largest Laplacian eigenvalue of G, respectively. If S
3
V is a balanced vertex separator of G,
then
m
3
n
_
1 +
9
16
(G)
_
_
1 +
9
16
(G)
_
2
1
_
where
(G) =
(
n

2
)
2

2
.
If the graph G is connected, the lower bounds derived above are closely related to the condition
number cond
2
(V
t
LV ) of the matrix V
t
LV , because in this case
min
(V
t
LV ) =
2
(L) > 0 (cf.
Lemma 7 (iii)) and
(G) := cond
2
(V
t
LV ) =

n

2
.
Theorem 8 shows that

(G)
_
n
(G)
_
and Corollary 12 yields the following bound for balanced separators:
[S
3
[ n h((G))
10
where
h(x) := 1 +
9
16
(x 1)
2
x

_
1 +
9
16
(x 1)
2
x
_
2
1 .
Other lower bounds for the size of vertex separators can be found in [15] and in [13, Theorem
2.8]. With = (G) being the maximal (weighted) degree of G, the latter bound reads
[S
3
[
4
2
[S
1
[[S
2
[
n
2
[S
1
S
2
[
. (21)
It is easy to see that (20) is superior to (21) if and only if

n
+
2
_
1 +
m
3
n


2

n
_
. (22)
For a probabilistic comparison of the obtained bounds, we state another result from [11]. Let
(
n,p
be the probability space of random graphs of order n with edge probability p.
Theorem 13 For xed edge probability p (0 < p < 1) and any > 0, almost all graphs in (
n,p
have their Laplacian eigenvalues
2
(G) and
n
(G) bounded by
pn f

(n) <
2
< pn f

(n)
and
pn +f

(n) >
n
> pn +f

(n)
where f

(n) =
_
(2 +)p(1 p)nlog n.
Furthermore, we observe that for xed edge probability p, almost all graphs have np.
First we want to show that
(1 +
m
3
n


2

n
)
p
2
for almost all graphs. Note that 1
2
/
n
. By a result of Fiedler [4, Theorem 4.1],
2
m
3
and m
3
/n can be bounded from below by
lim
n

2
n
p lim
n
f

n
= p
which proves the claim. We are now ready to show that for xed edge probability p, (22) is true
for almost all graphs. It follows from
pn +f

(n) pn + (pn f

(n))
p
2
which is equivalent to
f

(n)(1 +
p
2
) n
p
2
2
.
The latter inequality is true for almost all graphs as f

(n) grows slower than n. This proves the

following:
Theorem 14 For xed edge probability p (0 < p < 1), almost all graphs in (
n,p
satisfy
4
n

2
[S
1
[[S
2
[
n(
n

2
)
2

4
2
[S
1
[[S
2
[
n
2
[S
1
S
2
[
.
11
5 Testing for Subgraph Isomorphism
The problem of subgraph isomorphism can be stated as follows. Given two undirected, unweighted
graphs G = (V, E) and H = (V
H
, E
H
), does G contain a subgraph G

= (V

, E

) which is
isomorphic to H. This problem contains a number of other well known problems such as maximum
clique-size, maximum independent set, hamiltonian circuit, etc. In the following, we assume that
both graphs G and H are of order n. (Otherwise we add as many isolated vertices as necessary
to the smaller graph.) The following lemma is easily proved.
Lemma 15 Let G and H be unweighted graphs of order n. Let A
G
and A
H
be the corresponding
adjacency matrices. Then G contains a subgraph isomorphic to H if and only if
max
X
tr A
G
XA
H
X
t
= 2[E(H)[.
The lemma suggests the following denitions. A graph is a common subgraph of two graphs G
and H if it is isomorphic to subgraphs of G and H. A graph G

is a largest common subgraph of

two graphs G and H if it is a common subgraph of G and H and there is no common subgraph
with more edges than G

. It is convenient to dene the cardinality of the edge set of a common

subgraph as the size of the common subgraph. Lemma 15 now generalizes to:
Lemma 16 Let G and H be unweighted graphs of order n. Let A
G
and A
H
be the corresponding
adjacency matrices. Then the size S of a largest common subgraph satises
S =
1
2
max
X
tr A
G
XA
H
X
t
. (23)
Note that (23) is symmetric in G and H. If we use the Laplacian L
G
instead of A
G
, (23)
transforms to
S =
1
2
min
X
tr L
G
XA
H
X
t
. (24)
We can now state the main result of this section.
Theorem 17 Given unweighted graphs G and H of order n, let L(G) and L(H) denote their
Laplacian matrices, and let A(G), A(H) be their adjacency matrices, respectively. Then the size
S of a largest common subgraph of G and H is bounded by
S min
_
1
2
n1

i=1

i+1
(L(G))
i
(V
t
A(H)V ),
1
2
n1

i=1

i+1
(L(H))
i
(V
t
A(G)V )
_
(25)
Proof. The argument is similar to the proof of Theorem 9 and is therefore omitted.
The problem of nding the largest, or at least a large common subgraph, has practical ap-
plications in distributed computing. Bokhari [1] studies the mapping of nite element grids to
an array of processors. If adjacent modules of the grid should also be mapped to directly linked
12
processors, this leads precisely to the problem of nding the largest common subgraph of the
connection structure of the processors and the nite element grid.
Another typical question in graph theory is, given a set E of forbidden edges between vertices
labeled from 1 to n, is it possible to label the vertices of a given graph G in such a way that the
resulting graph does not contain any edge from E. Let H be the graph of forbidden edges, and
let A
G
and A
H
be the adjacency matrices of G and H, respectively. Then such a labeling for G
exists if and only if
min
X
tr A
G
XA
H
X
t
= 0 . (26)
6 Lower bound on the 1-sum problem
The 1-sum
1
(G) of a graph G of order n is dened as

1
(G) := min
_

(i,j)E
[(i) (j)[ : bijection V (G) 1, . . . , n
_
. (27)
Juvan and Mohar [10] formulated the following eigenvalue lower bound depending on the second
smallest Laplacian eigenvalue:

1
(G)
2
(L)
n
2
1
6
. (28)
We will derive a tighter lower bound based on Theorem 6. Let B = (b
ij
) be an n n matrix
dened by b
ij
= [i j[ for every i, j. We recall that L = L(G) denotes the (weighted) Laplacian
matrix of a graph G, and V is an n (n 1) projection matrix satisfying (6). We rst present
some properties of the matrix B.
Lemma 18 For every n, the above matrix B has the following properties:
(i) tr(V
t
BV ) =
1
3
(n
2
1).
(ii) The matrix V
t
BV is negative denite.
Proof. (i) Using the properties of V , we have
tr V
t
BV = tr BV V
t
= tr B(I
uu
t
n
) =
1
n
u
t
Bu.
It is easily seen that u
t
Bu =

i,j
b
ij
=
1
3
(n
3
n). Thus (i) follows.
(ii) Let x be a nonzero vector, and x u. We have to show that x
t
Bx < 0. We rst note
that the vectors
i
:= e
i
e
i+1
= (0, . . . , 0, 1, 1, 0, . . . , 0)
t
for i = 1, . . . , n 1 form a basis of
u

. Thus x can be written as x =

i
a
i

i
and

i
a
2
i
> 0. Now note that
t
i
B
j
= 2, if i = j,
and 0 otherwise. Therefore,
x
t
Bx =

i,j
a
i
a
j

t
i
B
j
= 2

i
a
2
i
< 0.
Now we show that Theorem 6 leads to a lower bound on the 1-sum that dominates the
JuvanMohar bound.
13
Theorem 19 Let G be a graph of order n. Then

1
(G)
1
2
n1

i=1

i+1
(L)
i
(V
t
BV )
n
2
1
6

2
(L). (29)
Moreover, if G is dierent from the empty and the complete graph, then the second inequality is
strict.
Proof. Let
i
=
i+1
(L), i = 1, . . . , n 1. By Theorem 6,

1
(G) =
1
2
maxtr LXBX
t
: X
1
2
n1

i=1

i
(V
t
BV ) .
To demonstrate the second inequality we use the fact that the eigenvalues of V
t
BV are negative
(Lemma 18(ii)) to conclude

n1

i=1

i
(V
t
BV )
1
n1

i=1

i
(V
t
BV ) =
1
n
2
1
3
. (30)
The last equality follows from Lemma 18(i). Note that
1
,=
n1
if G is neither complete nor
empty, showing that strict inequality holds in this case.
6.1 The 1-sum of Q
n
Let Q
n
denote the n-dimensional Cartesian cube. We compare the lower bound of Theorem 19
with the actual value of
1
(Q
n
), and with the lower bound (28). The actual value

1
(Q
n
) = 2
n1
(2
n
1) (31)
was determined by Harper [9]. The Laplacian eigenvalues of Q
n
are recalled in Subsection 7.3.
The Juvan-Mohar bound states

1
(Q
n
)
1
3
(2
2n
1).
The computation of the lower bound of Theorem 19 was done in Matlab, and the results are
compared in Table 1. The columns are: the dimension of the cube, Juvan-Mohar bound, bound
by Theorem 19, and the exact value by (31).
6.2 The 1-sum in randomly generated graphs
We compare the Juvan-Mohar bound (28) and the bound of Theorem 19 on randomly generated
graphs with edge probability 0.5. We have generated 100 graphs for each order n = 20, 30, . . . , 100.
For each generated graph, we compared the Juvan-Mohar bound (28) and the bound of Theorem
19, respectively. In all cases, the new bound was better than the old one. For each size n =
20, 30, . . . , 100, we have computed the average, minimum, and maximum gain of Theorem 19 over
14
n Juvan-Mohar Theorem 19 1-sum
1 1 1 1
2 5 6 6
3 21 24 28
4 85 99 120
5 341 392 496
6 1365 1542 2016
7 5461 6074 8128
Table 1: 1-sum of the cube Q
n
(28). The results are summarized in Table 2. They show the percentage of the improvement,
e.g., the number 23.37 in Table 2 means that on 20-vertex examples, the new bound gave 23.37%
higher values in average. Let us comment that the high value in the column max for small n is
caused by the fact that
2
is very small when G is not suciently connected. (For that reason,
only connected graphs were generated.)
n average max min
20 23.37 90.55 9.48
30 13.55 34.99 7.07
40 10.22 35.38 4.61
50 8.07 18.41 3.50
60 6.52 19.82 3.62
70 5.64 16.26 3.07
80 4.81 12.89 2.34
90 4.11 11.08 2.30
100 3.84 8.64 2.05
Table 2: Comparison of bounds on 1-sum for randomly generated graphs
7 Computation and performance of the bounds
In this section we discuss the computability and the performance of the bounds on the bandwidth.
We compare the new bounds given in this paper with several previously known results.
15
7.1 Convexity and persymmetric matrices
Theorem 9 provides a lower bound on the bandwidth. Given a graph G, it is natural to ask
what is the best possible lower bound ensured by this theorem. We show that the bound can be
computed eciently for any weighted Laplacian of G. This follows from the fact that, for every
k, we can decide in polynomial time whether (9) holds for some B B
k
. The latter question can
be decided by solving the optimization problem
min
BB
k
f(B) (32)
where the function f(B) is dened as
f(B) =
n1

i=1

i
(V
t
BV ) (33)
where
i
=
i+1
(L) are the non-trivial eigenvalues of the Laplacian under an arbitrary (but xed)
weighting of the edges. Using the Fans theorem and the Rayleighs principle, it is not dicult
to show that f(B) is a convex function. Since the set B
k
is convex as well, problem (32) can be
solved in polynomial time by the ellipsoid method (see [7]).
We recall that a matrix B = (b
ij
) is called persymmetric if b
ij
= b
nj+1,ni+1
for every i and
j. Using the permutation matrix T = (t
ij
) with t
i,n+1i
= 1, i = 1, . . . , n, it is clear that B is
persymmetric if and only if B = T
t
BT. A consequence of the convexity is that it is enough to
consider (32) only on the subclass of persymmetric matrices instead of the whole B
k
. This can
be seen as follows: Given B, let B

be dened as b

i,j
= b
nj+1,ni+1
for every i, j = 1, . . . , n.
Clearly, B B
k
if and only if B

B
k
. Moreover, f(B) = f(B

) because V
t
BV and V
t
B

V have
the same spectrum. Hence f(
1
2
(B + B

))
1
2
(f(B) + f(B

)) = f(B) by the convexity. Since

1
2
(B +B

) is persymmetric, the claim follows.

Theorem 9 provides us with a possibility to formulate explicit lower bounds on the bandwidth
by specifying a subclass of B
k
. Possible candidates for these subclasses could be either Toeplitz
matrices (b
ij
= c
|ij|
), or circulant matrices. The main computational advantage in using circu-
lant matrices consists in the fact that the eigenvalues of a circulant matrix are given by a simple
formula (and hence can be computed in an easier way). The disadvantage of course lies in the
fact that we have only n 2k degrees of freedom to choose the circulants.
7.2 The bandwidth of the Petersen graph
We use our results to show that the bandwidth of the Petersen graph is 5. This can be seen as
follows. The Laplacian eigenvalues of the Petersen graph are 0, 2, and 5 with multiplicities 1, 5,
and 4, respectively. The symmetric Toeplitz matrix having the rst row (0 0 0 0 0 4 8 11 14 20)
lies in B
4
and makes (19) negative. This shows a lower bound of 5. On the other hand, it is easy
to nd a labeling realizing this bound. For comparison, Theorem 8 gives a weaker lower bound
3 on the bandwidth of the Petersen graph.
16
7.3 The bandwidth of Q
n
Let Q
n
denote the n-dimensional Cartesian cube. We compare the lower bound of Theorem 8
with the actual value of

(Q
n
). We show that the lower bound is

(Q
n
)
2
n
n
1 while the
actual value is

(Q
n
) 2
n
/

n, which is a relatively narrow gap.

Here are more details. The actual value

(Q
n
) =
n1

k=0
_
k
k/2|
_
(34)
was determined by Harper [9]. Since
_
k
k/2
_
2
k
/

k, (34) can be estimated by the double of the

largest term. The Laplacian spectrum of Q
n
consists of n+1 distinct eigenvalues
i
= 2i, each of
multiplicity
_
n
i
_
, i = 0, . . . , n. (This follows easily from the description of the adjacency spectra
of Q
n
, see, e.g., [18].) In particular, we have
2
= 2 and
max
= 2n. Hence

(Q
n
)
2
n
n
1 by
Theorem 8.
7.4 The bandwidth of random regular graphs
Though Theorem 8 provides a weaker lower bound on the bandwidth than Theorem 9, it is
sucient to derive the following result of F. de la Vega on the bandwidth of random regular
graphs of xed degree.
Corollary 20 [3] Almost all d-regular graphs on n vertices have bandwidth at least b
d
n where b
d
tends to 1 as d .
Proof. It is known [5] that random dregular graphs have
2
bounded away from zero.
By Theorem 8 this implies the existence of the constant b
d
since
n
2d. Friedman, Kahn, and
Szemeredi [5] also proved that for a random d-regular graph G we have
2
(G) = d O(

d) and

n
(G) = d +O(

d) with probability tending to 1 as n grows. This shows that b

d
1 as d .
Acknowledgements: The second author was supported in part by the Ministry of Research
and Technology of Slovenia, Research Project P1021010192. The work of the third author
was partially done during his stay in the Mathematical Institute of Academia Sinica in Taipei,
under a support of National Science Council of R.O.C. We thank the Christian Doppler Labor
- Diskrete Optimierung for partial nancial support.
References
[1] S. H. BOKHARI. On the mapping problem. IEEE Transactions on Computers, C30:207214,
1981.
17
[2] P.Z. CHINN, J. CHVATALOVA, A.K. DEWDNEY, and N.E. GIBBS. The bandwidth
problem for graphs and matrices A survey. J. Graph Theory, 6:223254, 1982.
[3] W. FERNANDEZ DE LA VEGA. On the bandwidth of random graphs, in: Combinatorial
Mathematics, (eds. C. Berge at al. ), Annals of Discrete Mathematics 17: 633638, 1983.
[4] M. FIEDLER. Laplacian of Graphs and Algebraic Connectivity, in Combinatorics and
Graph Theory, Banach Center Publications, 25:5770, PWN-Polish Scientic Publishers,
Warsaw, 1989.
[5] J. FRIEDMAN, J. KAHN, and E. SZEMER

EDI. On the second eigenvalue in random

regular graphs. Proc. 21st Annual ACM Symp. Theory Comput., Seattle, 1989 (ACM, New
York, 1989) pp. 587598.
[6] N. GAFFKE and O. KRAFFT. Matrix inequalities in the Loewner ordering, in Mod-
ern Applied Mathematics - Optimization and Operations Research, B. Korte (ed.), North
Holland, 1982, pp. 576622.
[7] M. GR

OTSCHEL, L. LOV

ASZ, and A. SCHRIJVER, Geometric algorithms and combina-

torial optimization, Springer-Verlag, Berlin, 1988.
[8] S. HADLEY, F. RENDL and H. WOLKOWICZ. A new lower bound via projection for the
quadratic assignment problem. Mathematics of Operations Research, 17:727739, 1992.
[9] L. H. HARPER. Optimal assignments of numbers to vertices J. SIAM, 12:131135, 1964.
[10] M. JUVAN and B. MOHAR. Optimal linear labelings and eigenvalues of graphs. Discrete
Applied Mathematics, 36:153168, 1992.
[11] M. JUVAN and B. MOHAR. Laplace eigenvalues and bandwidth-type invariants of graphs.
J. Graph Theory, 17:393407, 1993.
[12] J. von NEUMANN. Some matrix inequalities and metrization of matrix space. Tomsk Univ.
Rev., 1:286300, 1937 (reprinted in: John von Neumann: Collected Works, Vol. 4, A. H.
Taub, ed., Macmillan, New York, 205219, 1962).
[13] B. MOHAR and S. POLJAK. Eigenvalues in combinatorial optimization, in Combinatorial
and graph-theoretic problems in linear algebra, R. Brualdi, S. Friedland, V. Klee, Eds.,
IMA Volumes in Mathematics and Its Applications, Vol. 50, Springer-Verlag, 1993.
[14] M. J. OVERTON and R. S. WOMMERSLEY. On the sum of the largest eigenvalues of a
symmetric matrix. SIAM J. Matrix Anal. Appl., 13:4145, 1992.
[15] A. POTHEN, H. D. SIMON and K. LIOU. Partitioning sparse matrices with eigenvectors
of graphs. SIAM J. Matrix Anal. Appl., 11:430452, 1990.
18
[16] F. RENDL and H. WOLKOWICZ. A projection technique for partitioning the nodes of a
graph. Technical Report CORR 90-20, University of Waterloo, Waterloo, Canada, 1990.
[17] R. ROSEN. Matrix bandwidth minimization. Proc. ACM Nat. Conf., Princeton, 1968,
pp. 585595.
[18] A. J. SCHWENK and R. J. WILSON. On the eigenvalues of a graph, in: Selected Topics in
Graph Theory 1 (eds. L.W. Beineke and R.J. Wilson), Academic Press 1978, pp. 307336.
19