Daniel J Gross, John T Saccoman, Charles L Suffel-Spanning Tree Results For Graphs and Multigraphs - A Matrix-Theoretic Approach-World Scientific Publishing Company (2014)

8963_9789814566032_TP.
indd 1 19/8/14 12:21 pm

May 2, 2013 14:6 BC: 8831 - Probability and Statistical Theory PST˙ws
This page intentionally left blank

World Scientific
8963_9789814566032_TP.indd 2 19/8/14 12:21 pm

Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.
SPANNING â•›TREE â•›RESULTS â•›FOR â•›GRAPHS â•›AND â•›MULTIGRAPHS

A Matrix-Theoretic Approach
Copyright © 2015 by World Scientific Publishing Co. Pte. Ltd.
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance
Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy
is not required from the publisher.
ISBN 978-981-4566-03-2
Printed in Singapore
LaiFun - Spanning Tree Results for Graphs.indd 1 20/8/2014 9:51:45 AM

This work could not have been undertaken without the influence of our
departed friend and colleague, Frank T. Boesch.
To Mary, the smartest woman I know

--DJG
To John J. Saccoman, my father

--JTS
To my mother Mildred and my father Charles

--CLS

Preface
This book is concerned with the calculation of the number of spanning

trees of a multigraph using algebraic and analytic techniques. We also
include several results on optimizing the number of spanning trees
among all multigraphs in a class, i.e., those having a specified number of
nodes, n, and edges, e, denoted Ω ( n, e ) . The problem has some practical
use in network reliability theory. Some of the material in this book has
appeared elsewhere in individual publications and has been collected
here for the purpose of exposition. A formal probabilistic reliability
model can be described as follows: the edges of a multigraph are
assumed to have equal and independent probabilities of operation p, and
the reliability R of a multigraph is defined to be the probability that a
spanning connected subgraph operates. If ς i denotes the number of
spanning connected subgraphs having i edges, then it is easily verified
that
e
e −i
R= ∑ ς p (1 − p )
i = n −1
i
i
.
For small values of p, the reliability polynomial is dominated by the

ς n−1 term, and since ς n−1 is the number of spanning trees, graphs with a
vii
viii Preface
larger number of spanning trees will have greater reliability for such p. In
the study of graph theory, most of the results regarding the number of
spanning trees have only been proven for simple graphs, so herein, we
investigate the problem for the extended class of multigraphs. It should be
noted that, while extensions to multigraphs make the optimal solution
readily apparent in many problems, it is not the case for the spanning tree
problem.
In Chapter 0, we present some graph theory and matrix theory

background material so that the reader will be familiar with the
terminology used in the sequel. In Chapter 1, we introduce many
algebraic results for both simple graphs and multigraphs regarding the
calculation of their number of spanning trees. In Chapter 2, we present
and extend a classical optimization formulation of Cheng that was useful
in optimizing the number of spanning trees for certain graphs. In Chapter
3, we present a heretofore unpublished result outlined by the late Frank
Boesch in the area of spanning tree enumeration of threshold graphs. In
Chapter 4, we show that a complete graph minus a matching, previously
shown to have the greatest number of spanning trees among all simple
graphs having the same number of nodes and edges, is also optimal when
the class is extended to include most multigraphs having a single multiple
edge of multiplicity two. We also present an argument using degree
sequences that demonstrates the optimality of this simple graph for almost
all simple graphs in the class. In Chapter 5, we discuss graphs and
multigraphs having all of their Laplacian eigenvalues as integers.
Contents
Preface ................................................................................................................ vii
0 An Introduction to Relevant Graph Theory and Matrix Theory .................. 1

0.1 Graph Theory ............................................................................................. 1
0.2 Matrix Theory .............................................................................................. 19
1 Calculating the Number of Spanning Trees: The Algebraic Approach ......... 37

1.1 The Node-Arc Incidence Matrix .................................................................. 38
1.2 Laplacian Matrix .......................................................................................... 41
1.3 Special Graphs ............................................................................................. 50
1.4 Temperley’s B-Matrix .................................................................................. 72
1.5 Multigraphs ................................................................................................. 77
1.6 Eigenvalue Bounds for Multigraphs............................................................. 79
1.7 Multigraph Complements ............................................................................. 84
1.8 Two Maximum Tree Results ........................................................................ 89
2 Multigraphs with the Maximum Number of Spanning Trees: An Analytic

Approach ............................................................................................................. 93
2.1 The Maximum Spanning Tree Problem ....................................................... 93
2.2 Two Maximum Spanning Tree Results ........................................................ 105
3 Threshold Graphs .............................................................................................. 111

3.1 Characteristic Polynomials of Threshold Graphs ......................................... 111
3.2 Minimum Number of Spanning Trees .......................................................... 123
4 Approaches to the Multigraph Problem ........................................................... 147
ix
x Contents
5 Laplacian Integral Graphs and Multigraphs ................................................... 159

5.1 Complete Graphs and Related Structures ..................................................... 159
5.2 Split Graphs and Related Structures ............................................................. 161
5.3 Laplacian Integral Multigraphs .................................................................... 165
Bibliography ....................................................................................................... 169
Index .................................................................................................................... 173

Chapter 0
An Introduction to Relevant
Graph Theory and Matrix Theory
This book is concerned with the calculation of the number of spanning

trees of a multigraph using algebraic and analytic techniques. We also
include several results on optimizing the number of spanning trees
among all multigraphs in a class, i.e., those having a specified number of
nodes, n, and edges, e, denoted Ω ( n, e) . The problem has some practical
use in network reliability theory. Some of the material in this book has
appeared elsewhere in individual publications and has been collected
here for the purpose of exposition. In preparation, we first collect some
relevant graph theoretical and matrix theoretical results.
0.1 Graph Theory
Though we assume that the reader of this work is well versed in Graph
Theory, in this section we provide the primary graph theoretic definitions
and operations that are used in the body of the succeeding chapters. For
any other terminology and notation not provided here we refer the reader
to Chartrand, Lesniak and Zhang [Chartrand, 2011].
A multigraph is a pair M = (V , m) , where m is a nonnegative integer-
valued function defined on the collection of all two-element subsets
of V, denoted V(2) . In the case there are several multigraphs under
1
2 Spanning Tree Results for Graphs and Multigraphs
consideration we use the notation M = (V ( M ) , mM ) . The elements of V

are called the nodes of the multigraph. We assume that V is a finite set
and n = V is the order of the multigraph, i.e. the order is the number of
nodes. A multiedge is an element {u , v} ∈ V(2) such that m ({u, v}) ≠ 0 ;
m({u, v} ) is called the multiplicity of the multiedge. If for each {u, v} ,
m ({u, v} ) ∈{0,1} , then M is a graph, and m−1 ({1}) = E is called the edge
set. In this case, we use G instead of M and employ the alternate notation
G = (V , E ) .
Remark 0.1 The set of multigraphs includes the set of graphs. Thus any
result that is stated for multigraphs also holds for graphs. On the other
hand, results that are stated for graphs are only applicable to graphs and
do not hold for multigraphs.
A multigraph has a geometric representation in which each element

(node) of V is depicted by a point, and two points u and v are joined by
m ({u, v} ) curves. Figure 0.1 shows the geometric representation of a
multigraph M and a graph G. It is traditional to refer to each point in the
representation as a node and the collection of all such points as V. Also
we refer to each curve in the representation as an edge, and the collection
of all such curves as E. In the case that m ({u, v} ) ≥ 1 we represent any
single edge between the nodes u and v by uv. The number of edges,
denoted by e, is the size of the multigraph, i.e. e = E = ∑ m ({u , v}) .
{u , v}∈V( 2 )
An Introduction to Relevant Graph Theory and Matrix Theory 3
We denote the class of all multigraphs of order n and size e by Ω( n, e) .
In Figure 0.1 M ∈ Ω(5,10) and G ∈Ω ( 5,7 )
M G
Figure 0.1: A multigraph M and a graph G.
Note that a multigraph may have several geometric representations

that look different. We say that the multigraphs M ′ = (V ′, m′ ) and
M ′′ = (V ′′, m′′) are isomorphic if there is a bijection f : V ′ → V ′′ such
that for every {v1 , v2 } ∈V(2)′ , m ′ ({v1 , v2 }) = m ′′ ({ f (v ) , f (v )}) .

1 2 The
two multigraphs depicted in Figure 0.2 are isomorphic, under the

bijection f ( vi ) = ui .
Figure 0.2: Two isomorphic multigraphs.

Edges are incident at its end-nodes, and two nodes which have an
edge between them are called adjacent. Node u has degree deg(u) equal
to the number of edges having that node as an end-node, i.e.
deg(u ) = ∑ m ({u, v}) .
v ≠u
The following theorem, the First Theorem of Multigraph Theory, was

proved by Euler in 1736.
Theorem 0.1 (The First Theorem of Multigraph Theory) The sum

of the degrees of all the nodes of a multigraph M is equal to twice the
number of edges.
Proof Let x = vi v j be an edge of M, then x contributes 1 to both
deg ( v1 ) and deg ( v2 ) . Thus when summing the degrees of the nodes of
M, each edge is counted twice. □
Another way to represent a multigraph is by a matrix. Let M = (V , m)
be a multigraph of order n, with node set V = {v1 , v2 ,…, vn } . The
adjacency matrix of M, denoted A ( M ) or simply A is the n × n matrix
( )
 aij  where aij = m {vi , v j } . Note if A is the adjacency matrix of M,
n
then ∑a
j =1
ij = deg ( vi ) .
Example 0.1 The adjacency matrix of the multigraph in Figure 0.1 is

0 1 0 0 2
1 0 1 1 1

A = 0 1 0 3 1 .
 
0 1 3 0 0
 2 1 1 0 0
A multigraph M is r-regular if each node has degree equal to r. When

the specific value of r is not needed we say the multigraph is regular.
The multigraph M ′ = (V ′, m′) is called a submultigraph of
M = (V , m) if V ′ ⊆ V and m′ ({u, v} ) ≤ m ({u, v} ) for all {u , v} ∈V(2)

′ . If
V ′ = V , then M ′ is called a spanning submultigraph. If M is a graph

then we use the terms subgraph and spanning subgraph. If
m′ ({u, v} ) = m ({u , v} ) for all {u, v} ∈V(2)′ , then M ′ = (V ′, m′ ) is an
induced submultigraph and is denoted by V ′ . If V ′ contains no
edges, i.e. m ({u, v} ) = 0 for all {u, v} ∈V(2)′ then V ′ is called an
independent set of nodes. If V ′ = V and m′ ({u, v} ) = min m ({u, v} ) ,1{ }

′ , then M ′ = (V ′, m′ ) is a spanning subgraph called the
for all {u , v} ∈V(2)
underlying graph of the multigraph (V , m) . In Figure 0.1 G is the

underlying graph of multigraph M.
A path in a multigraph M is an alternating sequence of nodes and
edges v1 , x1 , v2 , x2 , v3 ,…, xk −1 , vk , where {v1 , v2 ,…, vk } are distinct nodes
and xi is an edge between vi and vi +1 . If the multigraph is in fact a graph
then it is only necessary to list the nodes. The length of a path is the
number of edges in the path. It is evident from the definition that a path
on k nodes has length k − 1 . A cycle in a multigraph M is an alternating
sequence of nodes and edges v1 , x1 , v2 , x2 , v3 ,… , xk −1 , vk , xk , v1 , where
{v0 , v1 ,…, vk } are distinct nodes and x i is an edge between vi −1 and v i ,
1 ≤ i ≤ k − 1 , and xk is an edge between vk and v0 . The length of a cycle

is the number of edges in the cycle, so a cycle on k nodes has length k.
In a multigraph it is possible to have cycles of length 2, i.e. if there are
multiple edges between a pair of nodes, but in a graph the minimum
cycle length is 3. A graph G is acyclic if it has no subgraphs which are
cycles.
A multigraph is connected when every partition of the node set
V = V1 ∪ V2 , V1 ,V2 ≠ ∅ , and V1 ∩ V2 = ∅ has at least one multiedge with
one endpoint in V1 and the other in V2 . Alternately, a multigraph is
connected if there is at least one path between every pair of nodes. A
multigraph which is not connected is disconnected. A component of a
multigraph M is a maximal connected submultigraph M ′ , i.e. if M ′′ is a
submultigraph of M that properly contains M ′ , then M ′′ is disconnected.
A disconnected multigraph contains at least two components, thus a
multigraph is connected if and only if it has one component.
A tree is a connected graph with n nodes and n − 1 edges.
Considering a longest path in a tree it is easy to see that a tree has at least
two nodes of degree 1, called leaves or pendant edges. A spanning
subgraph that is also a tree is called a spanning tree. Figure 0.3 depicts a
multigraph and one of its spanning trees. It is easy to see that a
multigraph has spanning trees if and only if it is connected.
M T
Figure 0.3: A multigraph M and a spanning tree T.
One of the purposes of this book is to determine which multigraph

with n nodes and e edges, i.e. in   n, e  , has the greatest number of
spanning trees among all multigraphs in   n, e  . Since spanning trees

exist only if the multigraph is connected, given a connected multigraph
M  (V , m) , the following procedure will produce a spanning tree.
Choose a node u at random, set V1  u , V2  V  V1 , and E   . If
V2   then since M is connected there is at least one edge with one

endpoint in V1 and the other in V2 . Choose one such edge, add the edge
to E and remove the endpoint from set V2 and add it to V1 . Continue this
process until V2   and therefore V1  V . The resulting graph
T  V1 , E  is a spanning tree.

The process found above can be used to find spanning trees but it is
inefficient for finding all spanning trees of a multigraph; the multigraph
in Figure 0.3 has 34 other spanning trees.
  
A directed multigraph is a pair M  V , m  , where m is a
 
nonnegative integer-valued function defined on V  V   u , u  | u V  .
As for multigraphs, we refer to the elements of V as nodes. A multi-arc is

an element ( u, v ) ∈ (V × V − {( u, u ) | u ∈V }) such that m ( ( u, v ) ) ≠ 0 ;
m ( ( u, v ) ) is called the multiplicity of the multi-arc. If for each pair
( u, v ) , m ( ( u, v ) ) ∈{0,1} , then (V , m) is a directed graph or a digraph,
and m −1 ({1} ) = A is called the arc-set. We sometimes employ the
alternate notation D = (V , A) .
As in the case of multigraphs, directed multigraphs have geometric
representations as well. Each element of V is depicted by a point, and
two points u and v are joined by m ( ( u , v ) ) directed curves, directed from
u to v. Figure 0.4 shows a directed multigraph and a directed graph. It is

traditional to refer to each directed curve in the representation as an arc,
and the collection of all such arcs as A. An arc from u to v is incident
from u and incident to v; if such an arc exists we say u is adjacent to v
and v is adjacent from u. Node u has indegree, indeg(u), equal to the
number of arcs incident to u and outdegree, outdeg(u), equal to the
number of arcs incident from u, i.e. indeg(u ) = ∑ m ( ( v, u ) ) and
v≠u
outdeg(u ) = ∑ m ( ( u, v ) ) .
v ≠u
M D
Figure 0.4: A directed multigraph M and a directed graph D .

A theorem for directed multigraphs analogous to the First Theorem of

Multigraph Theory follows from the observation that each arc
contributes 1 to the indegree of a node and 1 to the outdegree of another
node.
Theorem 0.2 (The First Theorem of Directed Multigraph Theory)

The sum of the indegrees of all the nodes of a directed multigraph is
equal to the sum of the outdegrees, and these both equal the number of
arcs.
Let M = (V , m ) be a directed multigraph and define
m ({u , v}) = m ( ( u , v ) ) + m ( ( v, u ) ) for each two element subset
{u , v} ∈V(2) . The multigraph M = (V , m ) is called the underlying
multigraph of M . Alternatively, if M is a given multigraph, then any

directed multigraph having M as its underlying multigraph is referred to
as an orientation of M.
In Figure 0.5, M is the underlying multigraph of the multi directed
graph M , while M can be considered an orientation of M.
Figure 0.5
We now introduce some special graphs that we will encounter in this

work.
A complete graph is a graph of the form (V ,V(2) ) , i.e. there is an
edge between every pair of nodes. If V = n we denote the complete

graph of order n by K n . We depict two complete graphs in Figure 0.6.
K4 K5
Figure 0.6: The complete graphs K 4 and K 5 .
A subgraph G ′ = (V ′, E ′ ) of a graph G = (V , E ) is a clique if it is a

maximal complete subgraph of G, i.e. E ′ = V(2)
′ and for any subset of
nodes V ′′ that properly contain V ′ , V ′′ is not complete. In the graph
depicted in Figure 0.7, {v1 , v2 , v3 , v4 } is a clique, while {v1 , v2 , v3} is
complete but not a clique.
v1 v2
v4 v3
Figure 0.7: A clique inside a graph. The clique edges are thicker.
A graph G = (V , E ) is bipartite if V can be partitioned into two sets

V1 and V2 such that any edge in E has one endpoint in V1 and the other
in V2 . A bipartite graph is a complete bipartite graph if for every u ∈ V1
and every v ∈ V2 , uv ∈ E . If V1 = p and V2 = q we denote the complete
bipartite graph by K p,q . When p =1, we refer to the graph K1, q as a star.
More generally, a graph G = (V , E ) is k-partite if V can be partitioned

into k sets V1 ,V2 ,… ,Vk , such that any edge in E has one endpoint in Vi
and the other in V j , for some 1 ≤ i < j ≤ k . Note a 2-partite graph is
bipartite. A k-partite graph is a complete k-partite graph if for every

u ∈ Vi and every v ∈V j , uv ∈ E whenever i ≠ j . If a complete k-partite
graph is regular, then each part has the same order. If Vi = ni for
1 ≤ i ≤ k , then we denote the complete k-partite graph by K n1 , n2 ,…, nk .
Figure 0.8 depicts the complete bipartite graph K 2,3 , the star K1,4 and
the complete 3-partite graph K 2,2,3 .
Figure 0.8: k-partite graphs.

Figure 0.8: k-partite graphs (continued).
Suppose n1 , n2 ,..., n k and n are positive integers such that

n
n1 < n2 < … < nk ≤ . The circulant graph Cn ( n1 , n2 ,..., n k ) is the graph
2
having node set V = {0,1, 2, …, n − 1} with node i adjacent to each node
( i ± n ) mod n
j for each j, 1 ≤ j ≤ k . The n1 , n2 ,..., n k denote the “jump”
sizes. If n j = j for all j, 1 ≤ j ≤ k , we denote the circulant by Cnk .
Figure 0.9 shows two circulants.
0 0
7 1 7 1
6 2 6 2
5 3 5 3
4 4
Figure 0.9: The circulants C82 and C8 (1,3) .

Given a graph G = (V , E ) , the complement of G, denoted G , is the
graph (V ,V(2) − E ) , i.e. G has the same node set as G and uv is an edge
of G if and only if uv is not an edge of G. A graph and its complement

are shown in Figure 0.10.
Figure 0.10: A graph and its complement.
On occasion it is necessary to delete a set of nodes or a set

of edges from a multigraph. Let M = (V , m) be a multigraph with
edge set E. If U ⊆ V is a set of nodes, then M − U is the multigraph
obtained by deleting the node set U and all edges incident on a
node in U, i.e. M − U = (V − U , m ') where m ' ({u , v} ) = m ({u, v} ) for all
{u, v} ∈ (V − U )(2) . If U = {u} we write M − u instead of M − {u} . If
F ⊆ E is a set of edges, then, M − F is the multigraph obtained by

deleting the edges in F, but not the end-nodes. If x is an edge of M and
F = { x} we write M − x instead of M − { x} .
Remark 0.2 There is an ambiguous case, i.e. M − {u, v} , where {u, v}

can be either the set consisting of the two nodes u and v or a multiedge
incident on the two nodes u and v. If we are deleting the two node set we
will write M − u − v and if we are deleting an edge incident on u and v

we will write M − uv .
Remark 0.3 One can also think of the complement of a graph

G = (V , E ) as the graph obtained by deleting the edge set E from the
complete graph on node set V.
Example 0.2 Consider the multigraph M, depicted in Figure 0.11a, with

specified nodes labeled u and v and specified edges x and y. The
multigraphs M − u − v and M − { x, y} are shown in Figure 0.11b and
0.11c, respectively.
u
x
v
a) A multigraph M
b) c)
Figure 0.11
A set of edges in a multigraph is a matching or an independent set of

edges if no two edges in the set are incident at the same node. A
matching consists of k edges, which we denote by kK 2 , since each edge
is isomorphic to K 2 . We note that by the definition of a matching,
n
k ≤   . A special case of edge deletion that we will encounter is K n
2
minus a matching, denoted K n − kK 2 . In Figure 0.12 we depict
K5 − 2K2 . Note, the two edges deleted are arbitrary as long as they are
not incident on a node, as all such graphs are isomorphic.
Figure 0.12: K 5 − 2 K 2 . The deleted edges are indicated by the dotted edges.
At times it may be necessary to add an edge to a multigraph. Let

M = (V , m) be a multigraph and u , v ∈ V , then M + k ⋅uv is the multigraph
obtained by adding k edges between the nodes u and v, i.e.
m ({u , v}) + k if {u ′, v ′} = {u , v}

M + k ⋅uv = (V , m ′ ) , where m ′ ({u ′, v ′}) =  .
 m ({u ′, v ′}) otherwise
When k = 1, we write M + uv . If x is an arbitrary edge, then M + kx denotes

the multigraph with k edges added between a pair of nodes.
Example 0.3 Consider the multigraph M, depicted in Figure 0.13a,

with specified nodes labeled u, v and w. The multigraphs M +2⋅uv and
M + uw are shown in Figure 0.13b and 0.13c, respectively. In Figure 0.14

we show K 5+ x , the placement of the edge x is arbitrary.
v
a) A multigraph M
u u
w w
v v
b) M +2{u ,v} c) M +{u , w}
Figure 0.13
Figure 0.14: K 5+ x .
If x = uv is an edge of a multigraph M, then the contraction M | x is

the multigraph obtained from coalescing u and v, the end-nodes of x, and
deleting any loops which result. The contraction results in a multigraph
with one fewer node and m u, v  fewer edges.
Example 0.4 Consider the multigraph M, depicted in Figure 0.15a,

with specified edges labeled x and y. The multigraphs M | x and M | y
are shown in Figure 0.15b and 0.15c, respectively.
a) A multigraph M
b) M | x c) M | y
Figure 0.15: Contractions.
We end this section by introducing some binary operations on graphs.

Let G1  V1 , E1  and G2  V2 , E2  with V1  V2   . The union of the
graphs G1 and G2 , denoted G1  G2 , is the graph V1  V2 , E1  E2  .

The join of the graphs G1 and G2 , denoted G1 + G2 , consists of G1 ∪ G2

and all possible edges between the nodes in G1 and those in G2 (see
Figure 0.16b). The product of the graphs G1 and G2 , denoted G1 × G2 ,
is the graph (V1 × V2 , E ) where {( u1 , v1 ) , ( u2 , v2 )} ∈ E if and only if
u1 = u2 and v1v2 ∈ E2 or v1 = v2 and u1u2 ∈ E1 (see Figure 0.16c).
v1
v3
a)
u1 v1
u2
v2 v3
b)
c) K 2 × K3
Figure 0.16: The join and the product of two graphs.
0.2 Matrix Theory
Some fundamental but perhaps less well-known results from matrix

theory to be used in the sequel are included with proofs for the sake of
completeness. We assume that the reader is familiar with basic matrix
definitions and operations, such as size of a matrix, matrix addition,
matrix multiplication, scalar multiplication transpose, and determinant.
For definitions of these and any other concepts not explicitly defined
herein we refer the reader to Lancaster and Tismenetsky [Lancaster,
1985].
We will first explain some matrix notation that will be used. If P is a
 p11 p12 ⋯ p1s 
p p22 ⋯ p2 s 
matrix of size r × s , then P =  21 =  pij  r ×s . If the
 ⋮ ⋮ ⋱ ⋮ 
 
 pr1 pr 2 ⋯ prs 
size is understood we will suppress the subscript in the latter. We refer to
a matrix of size 1× s or r × 1 (i.e., a row matrix or a column matrix,
respectively) as a vector. Unless explicitly stated, all vectors are column
 x1 
x 
matrices and we write x =   . Finally, I r denotes the r × r identity
2
⋮
 
 xr 
matrix, Or×s denotes the r × s matrix zero matrix with all entries 0, 0r
the zero vector with all entries 0, and 1r the vector with all entries 1. If
the size is understood we suppress the subscripts. We note that all
matrices and vectors are bold-faced.
The first result we present concerns the determinant of a product of

two matrices.
Theorem 0.3 – The Binèt-Cauchy Theorem [Lancaster, 1985]. If

P =  pij  and Q =  qij  are matrices over an arbitrary field, with
k ×l l ×k
k ≤ l , then det(PQ) =
 1 2 3 ... k     j1 j2 j3 ... jk  
∑ det  P 
j3 ...
 det  Q
jk     1 2 3 ...
,
k  
1≤ j1 < j2 <...< jk ≤l   j1 j2
 i1 i2 i3 ⋯ i p 
where the minor C 
j3 ⋯ j p 
denotes the submatrix of C
 j1 j2
obtained by deleting all rows except i1 , i2 ,…, i p and all columns except
j1 , j2 ,…, j p . In other words, the determinant of the product PQ is equal

to the sum of the products of all possible minors of (the maximal) order k
of P with corresponding minors of Q of the same order.
Proof Observe that

 l l l

 ∑
p1τ1 qτ11 ∑ p1τ 2 qτ 2 2 ⋯ ∑ p τ qτ 1 k kk
 τ1 =1 τ2 =1 k =1 τ 
 l l l 

∑
p2τ qτ 1
PQ = τ1 =1 1 1
∑ p τ qτ
τ
2 2 2 2 ⋯
τ k =1
∑
p2τ k qτ k k 
.
2 =1
 
 ⋮ ⋮ ⋱ ⋮ 
 l l l 
 pkτ k qτ k k 
∑
p q
τ =1 kτ1 τ1 1 ∑ p τ qτ k 2 2 2 ⋯ ∑ 
1 τ 2 =1 τ k =1 
Now, by successively applying the linearity of the determinant for a

specific column, we obtain
 p1τ1 qτ11 p1τ 2 qτ 2 2 ⋯ p1τ k qτ k k 

 
l l l
 p2τ qτ 1 p2τ 2 qτ 2 2 ⋯ p2τ k qτ k k 
det(PQ) = ∑∑ ∑
⋅ ⋅ ⋅ det  1 1
⋮ 
τ1 =1τ 2 =1 τ k =1  ⋮ ⋮ ⋱
 pkτ1 qτ11 pkτ 2 qτ 2 2 ⋯ pkτ k qτ k k 

which is a sum of l k determinants. Of course, if τ a = τ b , where a ≠ b ,

then the corresponding term in the sum vanishes and the sum reduces to
only those terms for which τ1 ,τ 2 ,…,τ k are distinct. Now consider a
specific list 1 ≤ j1 < j2 < … < jk ≤ l , and using σ to denote permutation,
it follows that det ( PQ ) =

 
∑ ∑ det  1 1
⋮ ⋮ ⋱ ⋮ 
.
1≤ j1 < j2 <...< jk ≤ l σ ( j1 , j2 ,..., jk ) = (τ1 ,τ 2 ,...,τ k )
 
 pkτ1 qτ11 pkτ 2 qτ 2 2 ⋯ pkτ k qτ k k 
But

 
det  1 1 =
 ⋮ ⋮ ⋱ ⋮ 
 pkτ1 qτ11 pkτ 2 qτ 2 2 ⋯ pkτ k qτ k k 

 p1 j1 q j11 p1 j2 q j2 2 ⋯ p1 jk q jk k 
 
Iσ  p2 j q j 1 p2 j2 q j2 2 ⋯ p2 jk q jk k 
( −1) det  1 1 ,
⋮ ⋮ ⋱ ⋮
 
 pkj1 q j11 pkj2 q j2 2 ⋯ pkjk q jk k 

where I σ denotes the number of inversions in the permutation
σ ( j1 , j2 ,…, jk ) = (τ 1 ,τ 2 ,…,τ k ) . Again, successively applying the

linearity of determinants for a fixed column,
det(PQ) = ∑ ∑τ τ
1≤ j1 < j2 <...< jk ≤l σ ( j1 , j2 ,..., jk ) = ( 1 , 2 ,...,τ k )
(−1) Iσ q j11q j2 2 ⋅⋅⋅ q jk k i
 1 2 ⋅⋅⋅ k 
det  P  .
  j1 j2 ⋅⋅⋅ jk  
Finally, the result follows from the observation that
 1 2 ⋅⋅⋅ k 
det  Q  = ∑ ( −1) Iσ q j11q j2 2 ⋅ ⋅ ⋅ q jk k . □
  j1 j2 ⋅⋅⋅ jk   σ
4 1 
 2 −1 0   
Example 0.5 Let P =   and Q =  2 −1 , then applying
 3 2 2   3 2 
Binèt-Cauchy:
 2 1 4 1   1 0   2 1
det( PQ )  det   det    det   det  
3 2   2 1  2 2 3 2 
2 0 4 1 
+ det   det     7  6    2  7    4  5   36 .
3 2 3 2
 6 3  6 3
Multiplying the matrices, PQ =   and det    36.
 22 5  22 5
Before we proceed with the remainder of this section, we will review

some pertinent material from Linear Algebra. A set of vectors W   n is
a subspace of  n if the following three conditions hold: (i) 0 W , (ii)

x  y W for every x, y W , and (iii) cx W for every c   , x   n .
Given a set of vectors x , x ,…, x 

1 2 j the span of the vectors,
span x1 , x 2 ,…, x j  is the set consisting of all linear combinations of the
vectors, i.e.
span x1 , x 2 ,…, x j   c1x1  c2 x 2  …  c j x j | c1 , c2 , , c j   .
It is an easy exercise to show that span x1 , x 2 ,…, x j  is a subspace of
 n . A set of vectors x , x ,…, x 

1 2 j is linearly independent if
c1x1  c2 x2  …  c j x j  0 only when ci  0, for all 1 < i < j. If the set of

vectors is not linearly independent, then it is linearly dependent, i.e. there
exist c1 , c2 ,, c j   at least one of which is not 0 such that
c1x1  c2 x2  …  c j x j  0 . A basis of a subspace W is a linearly

independent set of vectors that span W. The dimension of a subspace W,
dim W  , is the number of vectors in any basis of W. Given two vectors
 x1   y1 
x  y  n
 2
and y =   , the dot product x ⋅ y = ∑x y
2
x= . We will also use
⋮ ⋮ i =1
i i
   
 xn   yn 
the notation xT y . Two vectors x and y are orthogonal if x ⋅ y = 0 .
The orthogonal complement of a subspace W ⊆ ℝn , W⊥ =
{y ∈ℝ n
}
| y ⋅ x = 0, for all x ∈ W . The Euclidean norm or just norm of
a vector, x e
or simply x , is given by x e = x ⋅ x . A set of vectors
{x , x ,…, x } is orthonormal if the vectors each have norm 1 and are

1 2 j
1, i = k
pairwise orthogonal, i.e. xi ⋅ x k =  . Any set of orthonormal
0, i ≠ k
vectors is necessarily linearly independent. An orthonormal basis is a
basis whose vectors form an orthonormal set.
Let A be an n × n matrix. An eigenvalue of A is a scalar λ such that
Ax = λ x for some non-zero vector x. Any non-zero vector x such that
Ax = λ x is an eigenvector associated with the eigenvalue λ . The
characteristic polynomial of A, PA ( λ ) = det ( λI − A ) . Eigenvalues are
the roots of the equation PA ( λ ) = 0 .
Proposition 0.4 If A is an n × n real symmetric matrix then it has n real

eigenvalues and an orthonormal basis of associated eigenvectors.
Proof Suppose A is an n × n real symmetric matrix, W is a subspace of

ℝ n and W is invariant under A and AT , i.e. A (W ) ⊂ W and
AT (W ) ⊂ W . If {x1 , x2 ,…, xk } is an orthonormal basis for W then there

k
exists unique real k × k matrices B and C such that if x = ∑ α i x i ,
i =1
 β1   α1 
 ⋮  = B ⋮ 
k k
Ax = ∑ β i x i , and A x = ∑ χi xi ,
T
then     and
i =1 i =1
 β k  α k 
 χ1   α1 
 ⋮  = C  ⋮  . It is easy to see that BT = C so that if A is symmetric
   
 χ k  α k 
and A (W ) ⊂ W , then AT (W ) ⊂ W and B is symmetric.

Our first claim is that if A is symmetric then the linear operator
x → Ax has k real eigenvalues in W (including multiplicities). Indeed,
the eigenvalues of B are those of x → Ax . Now det ( λI − B ) splits into
linear factors over the complex numbers and therefore has k roots
(including multiplicities) which are readily seen to be real.
Next we claim that each subspace W of dimension k which is
invariant under A has an orthonormal basis of eigenvectors
{y1 , y 2 ,…, y k } corresponding to the eigenvalues λ1 , λ2 ,…, λk . We prove
this by induction on k. If k = 1 then W is spanned by the unit vector y1

and since W is invariant under A we have that Ay1 = λ1y1 , i.e. λ1 is a real
eigenvalue and {y1} is the orthonormal basis. Now suppose 2 ≤ k ≤ n

and the claim is true for all invariant W of dimension < k. Let dim(W) = k
and A (W ) ⊆ W . Let y1 ∈W be an eigenvalue of norm one associated
with the real eigenvalue λ1 . Then W1 = span {y1} is invariant under A.
But then W1⊥ ∩ W , the collection of all vectors in W orthogonal to W1 , is

also invariant under the symmetric matrix A, i.e. if x ∈ W1⊥ then
0 = ( y T A ) x = y T ( Ax ) . Since dim (W1⊥ ∩ W ) = k − 1 , there exists k − 1
real eigenvalues and associated orthonormal eigenvectors y 2 , y3 ,…, y k

for A on W1⊥ ∩ W , thereby establishing the claim.
Finally we may apply this claim to k = n since A ( ℝ n ) ⊆ ℝ n . □
Theorem 0.5 (Courant-Fischer Max-Min Theorem) [Lancaster, 1985]

If the eigenvalues of the symmetric matrix A are named in order of
x T Ax
increasing value, i.e. λ1 ≤ λ2 ≤ … ≤ λn , then λ j = max min T where
Lj 0 ≠ x∈ L j x x
Lj varies over all ( n − j + 1) -dimensional subspaces of ℝ n .
To prove the Courant-Fischer Theorem, we first require a Lemma.
Lemma 0.6 For each 1 ≤ j ≤ n let L j denote an arbitrary ( n − j + 1) -
dimensional subspace of ℝ n , and consider a symmetric matrix A with

eigenvalues λ1 ≤ λ 2 ≤ … ≤ λn . Then
x T Ax x T Ax
λ j ≥ min T
and λn − j +1 ≤ max T .
0 ≠ x∈ L j x x 0 ≠ x∈L j x x
Proof (of Lemma 0.6) Let x1 , x 2 ,…, xn be an orthonormal system of

eigenvectors for the corresponding eigenvalues λ1 ≤ λ 2 ≤ … ≤ λn of A,
i.e., Axi = λi xi , and let Lˆ j = span {x1 , x 2 ,…, x j } for j = 1, 2,…, n . Fix j,
( )
and observe that, since dim ( L j ) + dim Lˆ j = ( n − j + 1) + j > n , there
n
exists a nonzero vector x0 ∈ ℝ belonging to both of these subspaces. In
j
particular, since x0 ∈ Lˆ j , it follows that x 0 = ∑ α k x k and thus,
k =1
T
 j   j  j
x 0T Ax0 =  ∑ α k x k  A  ∑α k x k  = ∑ λk α k .
2
 k =1   k =1  k =1
Hence,
j j
xT0 Ax 0
λ1 ∑ α k ≤ xT0 Ax 0 ≤ λ j ∑ α k , or λ1 ≤
2 2
≤ λj .
k =1 k =1 xT0 x0
The first result must follow because x0 ∈ L j , and so
xT Ax xT Ax
min ≤ T ≤ λj .
0 ≠ x∈L j xT x x0 x 0
The second result is a consequence of replacing Lj by
{
span x1 , x2 ,..., x j . } □
Proof (of Courant-Fischer) Because of Lemma 0.5, it suffices to

show the existence of an (n − j+1)-dimensional subspace such that
{
min ( Ax , x ) = λ j . Indeed, taking L j = span x j , x j +1 ,..., x n , we obtain
0 ≠ x∈ L j
}
such a subspace. □
A matrix A is positive semi-definite if xT Ax ≥ 0 for all vectors x.

Note: a matrix with all its principal minors non-negative is positive semi-
definite. A matrix A is positive definite if xT Ax > 0 for all vectors x ≠ 0 .
Proposition 0.7 If A is a positive semi-definite real symmetric matrix,

then its eigenvalues satisfy 0 ≤ λ1 ≤ λ2 ≤ … ≤ λn . If A is positive definite,
then its eigenvalues satisfy 0 < λ1 ≤ λ2 ≤ … ≤ λn .
Proof Let A be a positive semi-definite real symmetric matrix. We

know from Proposition 0.4 that all λi are real. Next consider Ax = λ x ,
where x ≠ 0 , so that 0 ≤ xT Ax and hence λ ≥ 0 . The result for positive

definite matrices follows by replacing the inequality 0 ≤ xT Ax with
0 < xT Ax . □
We refer to the list of eigenvalues of a matrix as the spectrum of the

matrix.
Corollary 0.8 Suppose that A is a positive semi-definite real symmetric

n × n matrix, and B is a real symmetric n × n matrix. Let α1 ,α 2 ,…,α n
denote the eigenvalues of A + B and β1 , β 2 ,…, β n denote the eigenvalues
of B. Then α k ≥ β k for all k = 1, 2,…, n .
Proof Note that xT ( A + B ) x = xT Ax + xT Bx ≥ xT Bx so that
xT ( A + B )x x T Bx
α n = max ≥ max = β n , and
Lk xT x Lk xT x
xT ( A + B ) x
α k = max min ≥
Lk 0 ≠ x∈Lk xT x
xT Bx
max min = min max xT Bx = β k
Lk 0 ≠ x∈Lk xT x y j =1
e
x j =1, xT y j = 0
e
j = k +1,... n j = k +1,... n
for each k = 1, 2,…, n. □

The following Lemma, the contrapositive of the Levy-Desplanques-

Haddamard Theorem [Lancaster, 1985], presents a useful property of
singular matrices:
Theorem 0.9 (Dominant Matrix Theorem) If A is a singular matrix,

then there exists a row i such that aii ≤ ∑ aij .
j ≠i
Proof If a matrix A is singular, then there is a non-zero vector x so that

Ax = 0. Hence, ∑a x
j
ij j = 0 for all i, or aii xi = −∑ aij x j for all i, and
j ≠i
ai ,i x i ≤ ∑ aij x j . Clearly, ai ,i x i ≤ max x j

j ≠i
( j ≠i
)∑ a
j ≠i
ij for all i. Let i
now be chosen so that xi is the largest value of xk for k = 1, 2,…, n ;
hence, max x j ≤ xi = max xk for all k. Further, since x≠0,

j ≠i k
xi = max xk > 0 . Therefore, aii ≤ ∑ aij .

k
□
j ≠i
Consider the matrix A = M − λ I which is singular for any λ , an

eigenvalue of M. Applying Theorem 0.7 we see that there exists an i
such that mii − λ = λ − mii ≤ ∑ mij . Now, using the second triangle
j ≠i
inequality x − y ≤ x − y , we get mii − λ ≤ λ − mii ≤ ∑ mij and

j ≠i
λ − mii ≤ λ − mii ≤ ∑ mij . Since these inequalities are true for each
j ≠i
eigenvalue λ, and some i, it follows that

   
min  mii −
i 
∑m ij  ≤ λ min ≤ λ max ≤ max  mii +
 i 
∑m ij .

 j ≠i   j ≠i 
Of course, if we are dealing with positive semi-definite matrices, then
λ = λ . This leads to the following corollary:
Corollary 0.10 (Gersgorin’s Theorem) Given an n-square complex

matrix M =  mij  we have eigenvalue bounds
   
λmin ≥ min  mii − ∑ mij  and λmax ≥ max  mii − ∑ mij  .
i   i  
 j ≠i   j ≠i 
Our next result gives a relationship between eigenvalues of a square

matrix to the trace and to the determinant of the matrix.
Theorem 0.11 Let A be an n × n matrix, with eigenvalues λ1 , λ2 ,…, λn ,

n n
then ∑λ
i =1
i = trace ( A ) and ∏λ
i =1
i = det ( A ) .
Proof Let A =  aij  be an n × n matrix and PA ( λ ) = det ( λI − A ) be
the characteristic polynomial of A. Since PA ( λ ) is a polynomial

of degree n we can write it in the form in the
n
PA ( λ ) = λ n − pn−1λ n −1 + … + ( −1) p0 . From this form we see that
n
PA (0) = ( −1) p0 . But plugging 0 into PA ( λ ) = det ( λI − A ) we get
n
PA ( 0) = det ( − A ) = ( −1) det( A) , thus p0 = det ( A ) . Now evaluating
det ( λI − A ) and collecting like terms we see that the coefficient of
λ n−1 , − pn −1 , is − ( a11 + a22 + … + ann ) = −trace ( A ) . If λ1 , λ2 ,…, λn are

the eigenvalues of A, i.e. the roots of PA ( λ ) = 0 , then we can also
express PA ( λ ) in the form PA ( λ ) = ( λ − λ1 )( λ − λ2 )⋯ ( λ − λn ) .

Multiplying out we get
n
PA ( λ ) = λ n − ( λ1 + λ2 + … + λn ) λ n −1 + … + ( −1) λ1λ2 ⋯λn .
Since the coefficients of the polynomial are unique, the results follow
from equating the coefficient of λ n −1 and the constant term in both
expansions. □
The following development of some algebraic graph theoretical ideas

will be useful in the sequel. If B and C are n × k and m × l matrices over
an arbitrary field, then the Kroenecker product of B and C, B ⊗ C, is the
nm × kl matrix with block form
 b11C b12 C ⋯ b1k C 
b C b C ⋯ b C 
 21 22 2k 
.
 ⋮ ⋮ ⋱ ⋮ 
 
 bn1C bn 2C ⋯ bnk C 
The Kroenecker sum of an n × n matrix B and m × m matrix C, denoted
B ⊕ C, is simply B ⊗ I m + I n ⊗ C .
Example 0.6
 1 0 3 1 0 3 
  
 2 2 −1 1 −1  2 −1
 1 
 1 0 3  
 2 −1    2 3 2   2 3 2  
 1 3 ⊗  2 −1 1 =  =
   2 3 2   1 0 3 1 0 3 
   
 1  2 −1 1 3  2 −1
 1 
  2 3 2   2 3 2 
 2 0 6 −1 0 −3
 4 −2 2 −2 1 −1

 4 6 4 −2 −3 −2 
 
 1 0 3 3 0 9
 2 −1 1 6 −3 3
 
 2 3 2 6 9 6 
and
1 0 3 1 0 3
 2 −1    2 −1  
1 3  ⊕  2 −1 1  =  1 3  ⊗ I 3 + I 2 ⊗  2 −1 1  =
    
 2 3 2   2 3 2 
 1 0 3 1 0 3 
    
1  2 −1 1  0  2 −1 1  
 2I 3 −1I 3    2 3 2   2 3 2  
 1I  + =
 3 3I3   1 0 3  1 0 3 
0  2 −1 1  1  2 −1 1  
   
  2 3 2   2 3 2  
2 0 0 −1 0 0  1 0 3 0 0 0
0 2 0 0 −1 0   2 −1 1 0 0 0 

0 0 2 0 0 −1  2 3 2 0 0 0
 + =
1 0 0 3 0 0  0 0 0 1 0 3
0 1 0 0 3 0  0 0 0 2 −1 1
   
0 0 1 0 0 3   0 0 0 2 3 2 
3 0 3 −1 0 0 
2 1 1 0 −1 0 

2 3 4 0 0 −1
 .
1 0 0 4 0 3
0 1 0 2 2 1
 
0 0 1 2 3 5 
Theorem 0.12 (An Associative Law of Products) If the products EF

and UV are defined, then (EF ) ⊗ ( UV ) = (E ⊗ U )(F ⊗ V ).
Proof Without loss of generality, suppose that EF is k × m and UV is

l × n . Then (( EF ) ⊗ ( UV ) )ij = ( EF )αβ ( UV ) pq where i = (α − 1) l + p
with 1 ≤ p ≤ l ,1 ≤ α ≤ k ; and 1 ≤ q ≤ n ,1 ≤ β ≤ m . On the other hand,

if E has s columns and U has t columns, then
((E ⊗ U)(F ⊗ V )ij = ∑ (E ⊗ U )

z
iz (F ⊗ V ) zj = ∑ ( Eαγ U )( Fγβ V ) ,
z
pr rq
where i and j are as above, and z = ( γ − 1) m + r with 1 ≤ γ ≤ s, 1 ≤ r ≤ t.

s t
Hence, (E ⊗ U)(F ⊗ V)ij = ∑ ( Eαγ Fγβ )∑ ( U pr Vrq ) = ( EF )
αβ ( UV ) pq . □
γ =1 r =1
An immediate consequence concerns the inverse of a Kroenecker

product:
Corollary 0.13 (Inverse of a Kroenecker Product) If P and Q are

−1
invertible, then so is P ⊗ Q and ( P ⊗ Q ) = P −1 ⊗ Q −1 .
Proof Consider (P ⊗ Q)(P−1 ⊗ Q−1 ) = PP−1 ⊗ QQ−1 = Im ⊗ In = Imn ,

where P is m × m and Q is n × n . □
Theorem 0.14 (Eigenvalues of a Kroenecker Polynomial) Let B and

C be square matrices, and set φ (B; C) = ∑ α ij B i ⊗ C j ( ) a finite sum.
i, j
Then the eigenvalues of φ (B; C) are given by φ ( β ; γ ) = ∑α

i, j
ij β iγ j ,
where β and γ run through the eigenvalues of B and C, respectively.
Proof Since the entries of B and C are complex, there exist invertible
matrices P and Q such that B1 = PBP−1 and C1 = QCQ−1 are upper
triangular. Of course, B1i = PBi P−1 and C1j = QC j Q−1 are also upper
triangular, and therefore so is each matrix αij B1i ⊗ C1j ( ) as well as the
sum ∑α ( B
i, j
ij
i
1 )
⊗ C1j . Now it follows from Theorem 0.12 and Corollary
0.13 that
(B ⊗ C ) = P
i
1 1
j −1
( )
Bi P ⊗ Q −1C j Q = ( P ⊗ Q) −1 Bi ⊗ C j = ( P ⊗ Q )
and thus φ (B1 ; C1 ) = ∑ α ij B1i ⊗ C1j = ( P ⊗ Q ) − which is equal to

1
( )
i, j
 
∑α ( B
−1

 ij
i
 )
⊗ C j  ( P ⊗ Q ) = ( P ⊗ Q ) φ ( B; C )( P ⊗ Q ) .
 i, j 
Therefore, φ ( B1; C1 ) and φ ( B; C) have the same eigenvalues, but
φ ( B1; C1 ) is upper triangular with diagonal entries given by

φ ( β ; γ ) = ∑ α ij β iγ j
as β and γ vary through the eigenvalues of B and
i, j
C, respectively. Hence, the result follows. □
Corollary 0.15 If B and C are square matrices, then the eigenvalues of

B ⊗ C and B ⊕ C are given by βγ and β + γ , respectively, as β and
γ vary through the eigenvalues of B and C.
Proof This result follows from Theorem 0.14 after observing

that B ⊗ C and B ⊕ C are obtained from φ p ( b; c ) = bc and
φs ( b; c ) = b0c + bc0 , respectively, upon substitution of B for b and C for

c, and the replacement of ordinary multiplication by Kroenecker product.
□

Chapter 1
Calculating the Number of

Spanning Trees: The Algebraic
Approach
In this chapter, we explore the algebraic method of computing t ( M ) ,
the number of spanning trees of a multigraph M. We then apply these

techniques to some special families of graphs.
A formal probabilistic reliability model, All Terminal Reliability
(ATR), can be described as follows: the edges of a multigraph M are
assumed to have equal and independent probabilities of operation p, and
the reliability R ( M , p ) of a multigraph is defined to be the probability
that a spanning connected subgraph operates. If ς i denotes the number

of spanning connected subgraphs having i edges, then it is easily verified
that
e
∑ ς p (1 − p )
e −i
R ( M , p) = i
i
.
i = n −1
For small values of p, the polynomial will be dominated by the ς n −1
term, and since ς n −1 is the number of spanning trees, graphs with a larger
number of spanning trees will have greater reliability for such p.
37
In graph theory, most of the results regarding the number of spanning

trees have only been proven for graphs, so in addition to an exposition of
results for graphs, we will investigate the problem for the extended class
of multigraphs. It should be noted that, while extensions to multigraphs
make the optimal solution readily apparent in many problems, this is not
the case for the spanning tree problem.
1.1 The Node-Arc Incidence Matrix
Consider a directed multigraph M with n nodes, {v1 , v2 ,..., vn } , and e
arcs, { x1 , x2 ,..., xe } . We define an n × e matrix, called the node-arc
incidence matrix S M =  sij  , which uniquely characterizes M by

( )
 1 if arc x j is incident to node vi


sij =  −1 if arc x j is incident from node vi .
 0 otherwise

Example 1.1 The directed multigraph depicted in Figure 1.1 has node-
arc incidence matrix
e1 e2 e3 e4 e5 e6 e7
v1  −1 1 0 0 0 0 −1
v2  1 −1 −1 0 0 0 0 
( )
S M =
v3  0 0 1 −1 1 1 0 
 
v4  0 0 0 1 −1 −1 1 
Calculating the Number of Spanning Trees: The Algebraic Approach 39
e1
v1 v2
e2
e7 e4 e3
e5
v4 e6 v3
Figure 1.1: A directed multigraph.
Since each arc has exactly two end-nodes, the sum of the row vectors of
S is 0Te . Clearly, any single row can be deleted from S without losing
any information about M . The matrix resulting from the deletion of the
ith row is called a reduced node-arc incidence matrix, and is denoted
( )
S r (i ) M , i.e., in terms of the notation of Chapter 0, this is the submatrix
 1 2 3 ⋯ i −1 i +1 ⋯ n
S 
1 2 ⋯ i − 1 i i + 1 ⋯ e 
or alternatively, if Ei denotes the matrix that results from deleting row i
( )
of the unit matrix, S r ( i ) M = Ei S. Note that it is often convenient to
renumber the nodes of the graph so that node n is deleted and thus node
k of the renumbered nodes corresponds to row k of S r ( n ) .
Theorem 1.1 The node-arc incidence matrix S is totally unimodular, i.e.

every square submatrix has determinant +1, − 1, or 0. Furthermore,
every n × n submatrix has determinant ±1 if and only if the n − 1
selected arcs form a tree in the underlying multigraph. [Seshu and Reed,
1961]
Proof Let x1 , x2 ,… , xe denote the edges of the underlying multigraph

M, and s1 , s2 ,..., se denote the corresponding columns of S . It suffices to
show that {s , s ,..., s } is linearly dependent if and only if

j1 j2 jk
{x j1 , x j2 ,..., x jk } (i.e., the corresponding undirected edges) contains a

cycle.
(Necessity) Assume there is a directed cycle with a counterclockwise
orientation. (See Figure 1.2.)
Figure 1.2: A directed cycle with counter-clockwise orientation.
k
Then ∑α s
i =1
i ji = 0 where α i = 1 if the jth arc is in the cycle, and α i = 0
otherwise. If the cycle is not counterclockwise, then we obtain a

nontrivial linear combination of the columns of S summing to zero by
multiplying those s values that correspond to arcs oriented opposite the
counterclockwise direction by −1 and thereby obtaining
k
∑α s
i =1
i ji =0.
(Sufficiency) If M is acyclic, we can add edges if necessary and direct

them arbitrarily so that a tree T is formed. It suffices to show that the
reduced node-arc incidence matrix for a tree, S r (i ) T ( ) has determinant
±1 . The proof follows by induction on the number of nodes involved.

( )
First assume the directed tree T has two nodes. Then S r (i ) T is an 1× 1
matrix with entry ±1 ; hence the determinant is ±1 . Now assume that any
directed tree on n − 1 nodes has corresponding reduced node-arc
incidence matrix determinant equal to ±1 . Now consider any reduced
node-arc incidence matrix of a directed tree T on n nodes. As T has at
least two pendant nodes, S r (i ) T( ) has at least one row having a single
nonzero entry in it which, of course, corresponds to a pendant node, say
v. Then expansion of ( ( ))
det S r (i ) T along that row yields
( ( ) ) so, by the induction hypothesis, the result follows.

±1 ⋅ det S r (i ) T □
1.2 Laplacian Matrix

T
Given a multigraph M of order n, the product S M ( )( S(M ) ) is the
same for any orientation M . We denote this product by H ( M ) , or H ,

and call it the Laplacian (also known as the nodal admittance matrix) of
M. The terminology “nodal admittance matrix” has its origins in the
theory of electrical circuit analysis.
Example 1.2 Let M be the multigraph depicted in Figure 1.3, and
( )
consider the orientation M and S M given in Example 1.1.
v1 v2
v4 v3
Figure 1.3
T
Then H ( M ) = S M ( )( S(M ) ) =
 −1 1 0 0 
 1 −1 0 0 
 −1 1 0 0 0 0 −1  
   0 − 1 1 0 
 1 −1 −1 0 0 0 0   0 0 −1 1 =
 0 0 1 −1 1 1 0   
   0 0 1 −1 
 0 0 0 1 −1 −1 1   0 0 1 −1
 
 −1 0 0 1
 
 3 −2 0 −1
 −2 3 −1 0 
  .
 0 −1 4 −3
 
 −1 0 −3 4 
It is easily shown that H is the n × n matrix  hij  where
 deg(vi ) if i = j
hij =  .
 − m(vi , v j ) if i ≠ j
Note also that if A ( M ) is the adjacency matrix of M and D( M ) is the

degree matrix, i.e. a diagonal matrix whose diagonal entries are the
degrees of the nodes of M , then H ( M )  D( M )  A ( M ) .
Example 1.2 continued
 3 2 0 1  3 0 0 0 0 2 0 1
 2 3 1 0  0 3 0 0   2 0 1 0 
H   DA.
 0 1 4 3 0 0 4 0 0 1 0 3
     
 1 0 3 4  0 0 0 4 1 0 3 0
The following proposition outlines some of the properties of the

Laplacian matrix H  M  for a multigraph M    n, e  .
Proposition 1.2 The n  n Laplacian matrix H  M  of the multigraph
M    n, e  has the following properties:
(a) H  M  is a real, symmetric matrix, and has real eigenvalues.
(b) H  M  is positive semi-definite, and has all nonnegative

eigenvalues.
n
(c) For each i, 1  i  n ,   H  M 
j 1
ij
 0.
(d) H  M  is singular, and its minimum eigenvalue is 0.
(e) H  M  is hyperdominant, i.e., ( H ( M ))ii   | H ( M ) | , for all i.

ji
ij
(f) ( H ( M )) ij  0 , whenever i  j .
n
(g) trace  H  M      H  M  ii  2e .
i 1
Proof (a) Follows trivially from the definitions and by Proposition 0.7.
(b) We have x T Hx = xT SS T x = x e ≥ 0. By Proposition 0.7, the
eigenvalues are all nonnegative.
(c) Follows trivially from the fact H = D − A .
(d) Since H1 = 0 by (c) we have that 0 is an eigenvalue, and H
is singular.
(e) and (f) follow trivially from the fact H = D − A .
(g) Follows from the fact that H = D − A and the First Theorem of
Multigraph Theory. □
Remark 1.1 From Proposition 1.2 (a), (b) and (e) we get the
eigenvalues of H ( M ) can be ordered
0 = λ1 ( H ( M ) ) ≤ λ2 ( H ( M ) ) ≤ … ≤ λn ( H ( M ) ) .
Unless otherwise stated, we assume that eigenvalues are the

eigenvalues of the Laplacian H ( M ) . In order to simplify notation we
will use λi ( M ) or simply λi when the multigraph is understood, in
place of λi ( H ( M ) ) .
We denote by H ij the matrix resulting from deleting row i and
column j from matrix H , and the (i, j)-minor of H is det(Hij ) . The (i, j)-
i+ j
cofactor is given by Cij ( H ) = ( −1) ( )
det H ij . The adjugate matrix of
H , adj( H ) , is the transpose of the matrix formed by the cofactors of

H, where entry ( i, j ) is Cij ( H ) . Determinant computation by cofactor
expansion yields Hadj(H ) = det ( H ) I .
Next, we prove the well-known Kirchhoff's Matrix Tree Theorem

[Kirchhoff, 1847]:
Theorem 1.3 (Kirchhoff) Let H be the Laplacian matrix of a multigraph

of order n. Then all cofactors of H are equal and their common value is
the number of spanning trees.
Proof First observe that Theorem 1.1 together with the Binet-Cauchy
Theorem yields t ( M ) = Cii ( H ) [Brooks, 1940]. Next, observe that by
Theorem 1.2, the rows of S sum to zero and H is symmetric,
thus H1 = 0 and 1T H = 0 , and so rank( H ) is strictly less than n. We
want to show that adj( H ) = α J , where J is the n × n matrix of all ones
and α is a scalar. If H is n × n , with rank less than n − 1 , then adj( H ) =
0, and the result is true, i.e. α = 0 . Thus, the case that remains is when the
rank is n − 1 or equivalently, the null space of H has dimension 1. Now
H adj( H ) = adj( H ) H = 0I , so each column of adj( H ) is an eigenvector
corresponding to the eigenvalue 0 and so adj ( H ) = α1 1, α 2 1,..., α n 1 for
scalars α1 , α 2 ,..., α n . However, since H is symmetric, adj( H ) is
α1 1T 
 
symmetric, and thus, α11,α 2 1,..., α n 1 =  ⋮  . Hence, all α i are
α 1T 
 n 
equal, and adj( H ) = α J. □
Remark 1.2 A careful examination of the proof that all cofactors are
equal clearly indicates that all cofactors of a matrix are equal whenever
its rows and columns sum to zero, regardless of whether the matrix is
symmetric. Conversely, it is easily shown from the fact L adj ( L ) =
adj ( L ) L = det ( L ) I that every equicofactor matrix having nonzero

cofactor has both row and column sum zero.
T
(
Remark 1.3 Consider H = SS T . Let H r ( n ) = S r ( n ) S r ( n ) ) denote the
reduced Laplacian matrix. Then Theorem 1.3 implies det(H r ( n ) ) = ( )=
det( ) = t ( M ) = Cij ( H ) . We also note that, for a connected multigraph, H r ( n ) is

non-singular.
The next proposition is only applicable to graphs, as the complement of

the graph G, G , is not defined for multigraphs.
Proposition 1.4
(a) If 0 = λ1 (G ) ≤ λ2 (G ) ≤ ... ≤ λn (G ) and 0 = λ1 (G ) ≤ λ2 (G ) ≤ ... ≤ λn (G )
are the eigenvalues of the Laplacian matrices HH

(G(G andHH
) and
)and (G(G
), ),
respectively, then λk + 2 (G) = n − λn − k (G) for k = 0,…, n − 2.
(b) The largest eigenvalue λn ( G ) = n if and only if G is disconnected.
(c) [Fiedler, 1973] 2δ (G ) − (n − 2) ≤ λ2 (G ) ≤ κ (G ) , where δ ( G ) is
the minimum degree of the nodes in G, and κ ( G ) is the node

connectivity of G.
Proof (a) We begin by showing that if λ is a non-zero eigenvalue of

H (G ) with associated eigenvector x, then n − λ is an eigenvalue of
H (G ) is also associated with the eigenvector x viz. the equation
( )
H G x = nx − H(G )x = (n − λ )x .
To this end, first realize that
( )
H (G )x + H G x = H( K n )x = (nI n − J n )x = nx − J n x .
But 1Tn H ( G ) = 0Tn , and H ( G ) x = λ x implies
(1Tn x) = 1Tn ( λx) =1Tn ( H (G ) x) = 0Tn x = 0.
As λ ≠ 0, it follows that 1Tn x = 0, and therefore, J n x = 0Tn . Hence,
( )
H(G ) x + H G x = nx , and H (G ) x = ( n − λ ) x follows immediately.
Now, suppose first that G is connected. Denoting the orthonormal

basis associated with these eigenvalues by x1 , x2 , x3 ,..., x n , it follows that
1
x1 must be . Since 1, x2 , x3 ,..., x n also correspond to the
1e
eigenvalues, 0, n − λ 2 ( G ) ,..., n − λ n ( G ) of H (G ) , they constitute the
complete set of eigenvalues of H (G ) . If, on the other hand, G is
disconnected, so that G is necessarily connected, the result follows by

applying the same argument to G .
(b) Observe that G is disconnected if and only if rank (H(G)) ≤ n − 2 .
But this is the case if and only if 0 is a multiple eigenvalue of H (G ) ,
( )
i.e., 0 = λ2 G = n − λn ( G ) .
(c) We defer the proof of this result until Section 1.6, where Proposition
1.19 provides a more general result. □
Remark 1.4 Kelmans stated in an earlier publication [Kelmans, 1967]

that λ2 ( G ) ≤ δ ( G ) , which is clearly subsumed by Proposition 1.4(c)
since κ ( G ) ≤ δ ( G ) .
Another formula for the number of spanning trees of a multigraph

involves the eigenvalues of H ( M ) [Kelmans, 1974]. Given a matrix C
we denote by PC ( λ ) the characteristic polynomial of C.
Theorem 1.5: The number of spanning trees of a multigraph is related to

the eigenvalues of its Laplacian matrix as follows:
n
1
t (M ) = ∏ λ (M ),
i where 0 = λ1 ( M ) ≤ λ2 ( M ) ≤ ... ≤ λn ( M ) .
n i=2
Proof Let PH ( M ) ( λ ) = det ( λ I n − H ( M ) ) , the characteristic polynomial
of H(M). The theorem is established by showing that both sides of the

n −1
( −1) PH′ ( M ) ( 0 )
equation in the statement equal .
n
n
Of course, PH ( M ) ( λ ) = λ ∏ ( λ − λ ( M ) ) so that
i =2
i
n n
n −1
PH′ ( M ) (0) = ∏ ( −λi (M ) ) = ( −1) ∏ ( λi (M ) ).
i=2 i=2
On the other hand, we may differentiate det ( λ I n − H ( M ) ) by adding the
determinants of the matrices obtained by differentiating a single row. A

typical determinant in the sum has ith row consisting of a one in the ith
position and other entries equal to 0; the remaining rows are those of
λ I n − H ( M ) . Thus, when λ is set equal to 0, we obtain the determinant
of − Hi ( M ) , i.e., − H ( M ) with ith row and column deleted. Now, an

application of the Matrix Tree Theorem (Theorem 1.3) yields the
□
n −1
conclusion that PH′ (M) (0) = n ( −1) t ( M ) and the result follows.
Example 1.3 Consider the multigraph M given in Figure 1.4. We will

compute t ( M ) two ways, by Theorem 1.3 and by Theorem 1.5.
v1 v2
v4 v3
M
Figure 1.4
 4 −2 −1 −1
 −2 3 −1 0 
Its Laplacian is H ( M ) =   . Using the (1,1)-cofactor
 −1 −1 5 −3
 
 −1 0 −3 4 
 3 −1 0 
det  −1 5 −3 = 29 . The eigenvalues of H ( M ) are
1+1
t ( M ) = ( −1)
 0 −3 4 
0, 2.676595722, 5.642973632, and 7.681330643. Thus
2.676595722i5.642973632i7.681330643
t(M ) = = 29 .
4
We leave it to the reader to identify the 29 spanning trees.
Remark 1.5 The presence of more than one eigenvalue equal to zero
implies a disconnected multigraph. In fact it is easily shown by
examining the block form of the Laplacian that the number of zero
eigenvalues is the same as the number of components. Thus, the formula
is valid for disconnected multigraphs as well.
1.3 Special Graphs
We are now in position to find the number of spanning trees for several
different families of graphs.
Proposition 1.6
(a) (Cayley’s Theorem) [Cayley, 1889] The number of spanning
trees of K n is n n− 2 .
(b) The number of spanning trees of a graph that is regular of
degree r and having adjacency matrix eigenvalues
n
1
α1 ≥ α 2 ≥ ... ≥ α n is
n ∏ ( r − α (G )) .
i=2
i
Proof (a) Since H ( K n ) = nI n − J n , it can be shown that
n −1 1 n −1
PH ( K n ) ( λ ) = λ ( λ − n ) . By Theorem 1.5, t ( K n ) = ( n ) = nn−2 .
n
(b) As previously noted, H ( G ) = D ( G ) − A ( G ) . Since A ( G ) is a real,

symmetric matrix, then by Proposition 0.4, there exists an orthonormal
basis of eigenvectors associated with the eigenvalues of A ( G ) . If G is a
regular graph of degree r, then it is easy to show that the eigenvectors of
A ( G ) are also the eigenvectors associated with H ( G ) , and the
eigenvalues are related by λi ( G ) = r − αi ( G ) for i = 2,3,..., n, where
r ≥ α1 ( G ) ≥ ... ≥ α n ( G ) are the eigenvalues of A ( G ) . Application of

Theorem 1.5 yields the result. □
Our next task is to determine the number of spanning trees for the
regular graph K n × K m using Proposition 1.6(b). To do this, we first
require a result for the eigenvalues of the adjacency matrix of the product
of two graphs based on the eigenvalues of the individual graphs.
Lemma 1.7 If the nodes of G1 and G2 are labeled u1 , u2 ,..., un and
v1 , v2 ,..., vm , respectively, and the nodes of G1 × G2 are ordered
(
lexicographically (i.e., label of ui , v j ) is smaller than that of ( u , v ) if
k ℓ
and only if i < k or i = k , j < ℓ,) , then
A(G1 × G 2 ) = A ( G1 ) ⊕ A ( G2 ) .
Proof Recall from Chapter 0 the Kroenecker sum:

A ( G1 ) ⊕ A ( G2 ) = I n ⊗ A ( G2 ) + A ( G1 ) ⊗ I m and consider A ( G1 ) ⊕
A ( G2 ) as an n × n matrix of blocks B χβ ,1 ≤ χ , β ≤ n, i.e.

 B11 B12 ⋯ B1n 

B B22 ⋯ B2 n 
A ( G1 ) ⊕ A ( G2 ) =  .
21
 ⋮ ⋮ ⋱ ⋮ 
 
 Bn1 Bn 2 ⋯ Bnn 
Since the χχ block of I n ⊗ A ( G2 ) is A ( G2 ) and the χχ block of
A ( G1 ) ⊗ I m is ( A ( G1 ) ) ⋅ I m = 0 , it follows that B χχ = A ( G2 ) . Thus,

χχ
the entry of A ( G1 ) ⊕ A ( G2 ) in the row corresponding to node uχ , vi ( )

and the column corresponding to u χ , v j ( ) is ( A ( G ) )
2 ij
. In other words,
the entry is 1 if vi and v j are adjacent in G2 and 0 otherwise. Similarly,
if χ ≠ β , it is easily seen that B χβ = A ( G1 ) ( ) χβ ⋅ I m; thus the entry of
A ( G1 ) ⊗ A ( G2 ) in the row corresponding to ( uχ , v ) i and column
(
corresponding to u β , v j ) is
 { }
 0 if i ≠ j or if i = j but u χ , u β ∉ E1
 .
{
1 if i = j and u χ , uβ ∈ E1 }
In conclusion, A ( G1 ) ⊕ A ( G2 ) is the adjacency matrix of G1 × G2 . □
Corollary 1.8 The number of spanning trees of a product of complete

graphs K n × K m is (m + n)( n −1)( m −1) mm − 2 n n − 2 .
Proof By Lemma 1.7, A ( K n × K m ) = A ( K n ) ⊕ A ( K m ) . Since the

eigenvalues of a Kroenecker sum are obtained by summing those of the
( )
summands, we require the eigenvalues of A K p . But the eigenvalues
of H K p( ) are p with multiplicity p − 1 , and 0 with multiplicity 1, by
Proposition 1.6(a), so A K p ( ) has eigenvalues −1 with multiplicity
p − 1 and p − 1 with multiplicity 1. It follows that the eigenvalues of
A ( K n × K m ) are:
 −2 with multplicity (n − 1)(m − 1)

n − 2
 with multiplicity m − 1
α ( Kn × Km ) =  .
m − 2 with multiplicity n − 1
 m + n − 2 with multiplicity 1
Finally, application of Proposition 1.6(b) yields
t ( Kn × K m )
1 ( n −1)( m −1) m −1 n −1
=
mn
( m + n − 2 − ( −2) ) ( m + n − 2 − (n − 2) ) ( m + n − 2 − (m − 2) )
and so t ( K n × K m ) = (m + n)( n−1)( m−1) mm − 2 n n− 2 . □
Recall the definition of a circulant graph given in Chapter 0. For

2
example, C8 (see Figure 0.9) has adjacency matrix
0 1 1 0 0 0 1 1
1 0 1 1 0 0 0 1 

1 1 0 1 1 0 0 0
 
0 1 1 0 1 1 0 0
.
0 0 1 1 0 1 1 0
 
0 0 0 1 1 0 1 1
1 0 0 0 1 1 0 1
 
1 1 0 0 0 1 1 0 
Its rows are successive shifts to the right of the row vector
[0 1 1 0 0 0 1 1] . This is true in general for any circulant
graph, i.e. the adjacency matrix is obtained by successively shifting the
first row vector to the right by one place. More formally, row k of the
circulant matrix A generated by the vector a = a1 ,..., an  is
( )
ε k −1  a0,..., an −1  , where ε represents the shift operator, i.e.,
ε ( b1 ,..., bn ) = ( bn , b1 ,..., bn −1 ) . Furthermore, by setting En equal to the
n × n circulant matrix generated by [ 0 1 0 . . . 0] , i.e.
0 1 0 0 ⋯ 0
0 0 1 0 ⋯ 0 

0 0 0 1 ⋯ 0
En =  
⋮ ⋮ ⋮ ⋮ ⋱ ⋮
0 0 0 0 ⋯ 1
 
1 0 0 0 ⋯ 0 
∑a E
i
we obtain A = i
i
n where Ein = ( E n ) . Those nonzero circulant
i =1
matrices which are the adjacency matrices of circulant graphs are

distinguished by the fact that a0 = 0, ai = 0, 1 for i ≤ 1 and ai = 1
if and only if an − i = 1, i.e., A is symmetric. Observe that
( )
A C82 = E8 + E82 + E86 + E87 . To compute the eigenvalues of a circulant
matrix,
det ( λ I n − En ) =
λ −1 0 0 ⋯ 0 
0 λ −1 0 ⋯ 0 
 
0 0 λ −1 ⋯ 0  n −1 1+ n n −1 n
det   = λ ⋅ λ − 1( −1) ( −1) = λ − 1
⋮ ⋮ ⋮ ⋮ ⋱ ⋮ 
0 0 0 0 ⋯ −1
 
 −1 0 0 0 ⋯ λ 
so that the eigenvalues of En are the nth roots of unity. Hence, if

k
n
nk < , the eigenvalues of A Cn ( n1 , n2 ,..., nk ) =
2
( ) ∑(E ni
n + E nn − ni ) are
i =1
given by
n − ni
k  i 2π j  i  i 2π j  k  2π ni j ni 
∑  e n  +  e n  =2 ∑ cos 
 n  for j = 0,1,… , n − 1 .
i =1     i =1  
n
On the other hand, if nk = , the eigenvalues of
2
n
  n  k
A  Cn  n1 , n2 ,..., nk −1 ,   =
  2  ∑(
i =1
E nni + E nn − ni )+ E n2
are given by
k
 2π ni j 
∑ cos 
j
2 + ( −1) , j = 0,1,..., n − 1 .
i =1 n 
The preceding discussion proves the following proposition:
Proposition 1.9 The number of spanning trees of a circulant is
(
t Cn ( n1 ,..., nk ) = )
 2n −1 n −1  k
 2π ni j  
∏ ∑ cos 
n
 k −  if nk <
 n j =1  i =1
n   2

 n
−1
n .
 2n −1 2  k
 4π ni m   2  k
 2 ( 2m − 1) π ni  
∏ ∑ ∏ ∑
n
 k − cos   k − 1 − cos   if nk =
 n 
m =1 
 n     n   2
i =1  m =1  i =1
The next result, which follows from Theorem 1.5, gives the number of
spanning trees for G in terms of the Laplacian matrix of the complement
G of G.
Corollary 1.10 If G is a graph on n nodes with complement

G and PH( G ) ( λ ) is the characteristic polynomial of H G , the ( )
1
Laplacian matrix of G, then t ( G ) = P ( n).
n 2 H (G )
n
1
Proof Applying Theorem 1.5, t ( G ) =
n ∏ λ ( G ) . By Proposition 1.4,
i =2
i
n
1
λi ( G ) = n − λn −i ( G ) for i = 1,..., n − 1. Thus, t (G ) =
n ∏ ( n − λ (G )).
i=2
i
n n
1
Now, PH ( G ) ( λ ) = λ ∏ ( λ − λi ( G ) ) so that
i =2
P
n H(G )
(n) = ∏ ( n − λ (G )) ,
i =2
i
and the result follows. □
This theorem can be refined after observing that, whenever G

can be decomposed into a node-disjoint union of subgraphs
k
C1 , C2 ,..., Ck , then PH ( G ) ( λ ) = ∏ P ( ) (λ ).
i =1
H Ci
We record this refinement
in the next result:
Corollary 1.11 If the complement G of G is the node disjoint union of

k
1
subgraphs C1 , C2 ,..., Ck , then t(G) =
n2
∏ P ( ) ( n ).
i =1
H Ci
In particular, if
G has n − χ isolated nodes, and the deletion of these nodes leaves
Gχ , then t ( G ) = n n − χ − 2 PH □
( Gχ ) (
n).
The use of the graph complement makes it easier to calculate the

number of spanning trees for many types of graphs. One such example is
“ K n minus a matching of size k”, i.e., K n − kK 2 . The formula for the
number of spanning trees for this and some other interesting graphs are
established in the following proposition:
Proposition 1.12 The formulas for the number of spanning trees for
some special types of graphs are as follows:
k
 2
(a) K n minus a matching (Figure 1.5): t ( K n − kK 2 ) = n n − 2  1 −  .
 n
(b) k-partite graphs (Figure 1.6)
(i) (
complete bipartite graph: t K p , q = q p −1 p q −1 ; )
(ii) regular complete k-partite graph:
 
(r −1)k
t  K r ,r ,...,r  = r rk − 2 k k − 2 ( k − 1) ;
 
k terms
(iii) complete bipartite with equal part sizes minus a full

p −1
( )
matching: t K p , p − pK 2 = ( p − 1) p p − 2 ( p − 2 ) .
(c)
(i) complete graph of order n minus a path of order p, n > p .
(Figure 1.7):
( )
t K n − Pp =
  2 
p +1
 2  
p +1
 n n − p −2  n − 2 +  n − 2  − 1  −  n − 2 −  n − 2  − 1   if n > 4
  2     2    
2
 2   2 
 2  n − 2  − 1      .

  2  
 3− p
 4 ⋅ p if n = 4
(ii) complete graph of order n minus a cycle of order p

(Figure 1.8):
( )
t Kn − C p =
  2 
p
n−2 2 
p

 n − p −2  n − 2 n−2   n−2  p +1 
 n
 2 +  2  − 1 + −  2  − 1 + 2 ( − 1) n
 if > 4
     2   
 
     

 p
 2 − 2 ( − 1 ) if n = 4 and p = 3, 4
 4
Proof (a) Suppose that E is a matching in K n consisting of k node
disjoint edges and G = K n − E. Then G = kK 2 ∪ ( n − 2k ) K1 , a node
disjoint union. Since the characteristic polynomial of H ( K 2 ) is
λ ( λ − 2 ) , the formula follows by Corollary 1.11.

Figure 1.5: K 6 − 2 K 2 . The dashed edges indicate the 2K 2 .
(b) Consider the most general complete k-partite graph G with

k
complement G = ∪ K χi ∪ ( n − χ ) K1 where the union is node-
i =1
disjoint. Two special cases are the complete bipartite case and the
complete k-partite case where all parts have the same order. In the
first case, k = 2 and n = χ . Setting χ1 = p and χ 2 = q , we see that
( )
t K p , q = q p −1 p q −1 .
In the second case, n = χ and χ1 = χ 2 = ⋅ ⋅ ⋅ = χ k = r . Thus,
 
(r −1)k
t  K r ,r ,..., r  = r rk − 2 k k − 2 ( k − 1) .
 
k terms 
Finally, if G = K p, p − pK 2 , then G consists of two copies of K p

joined by a matching of size p. Hence, with the node set
( )
appropriately labeled, H G has the block form given below:
 ( p + 1) I p − J p −I p 
( )
H G =
−I p ( p + 1) I p − J p 
.

Of course,
( λ − ( p + 1) ) I p + J p Ip 
λI2 p − H ( G ) =  .
 Ip ( λ − ( p + 1) ) I p + J p 

A B
Since any matrix   with square blocks A, B, C, and D has
C D 
determinant equal to det (AD-BC) if A and C commute, we may
write:
( ) ) = det ( ( ( λ − ( p + 1) ) I )
2
(
det λ I 2 p − H G p + Jp ) − Ip =
det ((( λ − ( p + 1)) − 1) I 2

p − ( p + 2 − 2λ ) J p . )
Now consider the general determinant det aI p − bJ p . ( )
(
If b = 0, then det aI p − bJ p = a p . )
a 
If b ≠ 0, then det(aI p − bJ p ) = b p ⋅ det  I p − J p  .
 b 
( )
But det λ I p − J p = λ p −1 ( λ − p ) , so that
p −1
a a 
(
det aI p − bJ p ) =b  
b
p
 
p −1
 b − p  = a ( a − bp ) .
Thus, whether b = 0 or not,
( )
det aI p − bJ p = a p −1 ( a − bp ) .
Applying this result to the problem at hand, we obtain

p −1
(
det λ I 2 p − H G ( )) = ((λ − p) 2
−1 ) ((λ − p) 2
− 1 − ( p − 2λ ) p )
p −1 p −1
= ( λ − p − 2) (λ − p) (λ 2
− 2λ . )
Thus, by Corollary 1.10,
p −1
( )
t K p , p − pK 2 = ( p − 1) p p − 2 ( p − 2 ) . □
a) K 2,3
b) K 2,2,2
c) K3,3 − 3K 2
Figure 1.6: k-partite graphs.
(c) (i) Observe that the Laplacian matrix of a path is
1 −1 0 0 ⋯ 0 
 −1 2 −1 0 ⋯ 0 
 
0 −1 2 −1 ⋱ 0 
( )
H Pp = .
⋮ ⋱ ⋱ ⋱ ⋱ 0
0 ⋯ 0 −1 2 −1
 
 0 ⋯ 0 0 −1 1 
We let f p ( x ) denote the determinant of the general p × p matrix with
the following form:

x −1 0 0 ⋯ 0
 −1 x + 1 −1 0 ⋯ 0 

0 −1 x + 1 −1 ⋱ 0
 .
⋮ ⋱ ⋱ ⋱ ⋱ 0
0 ⋯ 0 −1 x + 1 −1
 
 0 ⋯ 0 0 −1 x 
After subsequent expansion about the first and last rows (respectively),
denote by g p ( y ) the determinant of the p × p matrix
y −1 0 0 ⋯ 0
 −1 y −1 0 ⋯ 0 

0 − 1 y −1 ⋱ 0
 .
⋮ ⋱ ⋱ ⋱ ⋱ 0
0 ⋯ 0 −1 y −1
 
 0 ⋯ 0 0 −1 y 
Expanding g p ( y ) about the first row yields g p ( y ) = yg p −1 ( y ) − g p − 2 ( y ) .
Of course, this is equivalent to g p −1 ( y ) − yg p − 2 ( y ) + g p −3 ( y ) = 0 .
Setting y = x + 1 in this equation yields
g p −1 (1 + x) − (1 + x) g p − 2 (1 + x) + g p −3 (1 + x) = 0 .
Setting y = x + 1 in the first equation and multiplying the result by
(1 − x ) yields
(1 − x) g p (1 + x) + ( x 2 − 1) g p −1 (1 + x) + (1 − x) g p − 2 (1 + x) = 0 .
Adding the previous two equations and simplifying further gives
x 2 g p −1 ( x + 1) − 2 xg p − 2 ( x + 1) + g p − 3 ( x + 1) − ( x − 1) g p ( x + 1) = 0,
from which we obtain

f p ( x ) = x 2 g p − 2 ( x + 1) − 2 xg p −3 ( x + 1) + g p − 4 ( x + 1) =
( x − 1) g p −1 ( x + 1) . (*)
Now, if U p ( z ) denotes the Chebyshev polynomial of the second kind,
which is defined by the recursion:
U p + 2 ( z ) − 2 zU p +1 ( z ) + U p ( z ) = 0, where U 0 ( z ) = 1 and U1 ( z ) = 2 z ,
 y
then g p ( y ) = U p   . It is known [Seshu, 1961] that
2
p +1 p +1
U ( z) =
p
( z+ z2 − 1 ) −(z − z2 − 1 )
2 z2 − 1
for z 2 ≠ 1, U p (1) = p + 1, and U p (−1) = (−1) p ( p + 1) . Thus, from (*)
 y p
and g p ( y ) = U p   , we obtain, f p ( x ) = 0 if x = 1 , 4 p ( −1) if x = −3 ,
2
and
 2
p
  x +1 2  
p
x −1  x + 1  x +1  x + 1  
 +   −1 −  −   −1 
 2  2    2  2  
2
 x +1
2  − 1     

 2 
ifif x ≠ 1, − 3 . The substitution x = 1 − λ yields the characteristic
(
polynomial det λ I p − H Pp ( ) ) = ( −1) p
f p (1 − λ ) = 0 if λ = 0, 4p if λ = 4 ,
and if λ ≠ 0,4 the determinant is

 p +1 2 
p
2−λ 2  
p
( −1)  2 − λ 2−λ 
λ    2 − λ   
 2 +  2  −1 −  2 −  2  −1  .
2
    
 2−λ 
2  −1     

 2 
Finally, the result of Corollary 1.10 yields t ( G ) = 43− p ⋅ p if n = 4 , and
 2 
p
 2  
p
n n − p −1
 n − 2  n−2   n − 2  n − 2   
(n n − p −1
) 2  2 +  2  − 1  −  2 −  2  − 1  
   
 n−2   
2   − 1   
 2 
if n > 4 .
(ii) We first consider the Laplacian matrix of the cycle,
 2 −1 0 0 ⋯ −1
 − 1 2 −1 0 ⋯ 0 
 
 0 −1 2 −1 ⋱ 0 
H Cp =  ( ) .
⋮ ⋱ ⋱ ⋱ ⋱ 0
 0 ⋯ 0 −1 2 −1
 
 −1 0 ⋯ 0 −1 2 
We observe, in a similar fashion to the proof in (c)(i), that this matrix has
the form
 y −1 0 0 ⋯ −1
 − 1 y −1 0 ⋯ 0 
 
 0 − 1 y −1 ⋱ 0 
 
⋮ ⋱ ⋱ ⋱ ⋱ 0
 0 ⋯ 0 −1 y −1
 
 −1 0 ⋯ 0 −1 y 
and its determinant, hp ( y ) , can be written hp ( y) = y g p −1 ( y)
− 2 g p − 2 ( y) − 2 , where g p ( y) is the determinant of the matrix described
in the proof of (c)(i). Similar calculations lead to
( ( ) ) = ( −1)
det λ I p − H C p
p
hp (2 − λ ) .
Corollary 1.10, and direct substitution yield the formula t(G) =

p
2 − 2 ( −1)
if n = 4, and p = 3, 4 , and
4
 2 
p
 2 
p 
n− p−2  n − 2  n − 2  n − 2  n − 2  p +1 
n   2 +  2  − 1  +  2 −  2  − 1  + 2 ( −1) 
     
if n > 4 .
Figure 1.7: K5 − P3 . The dashed edges indicate the deleted P3 .
Figure 1.8: K5 − C3 . The dashed edges indicate the deleted C3 .
The following example checks the validity of the formulas for the
number of spanning trees found in this section. We do not include K n
here, as Cayley’s formula is well known.
Example 1.4 We consider the graphs depicted in Figures 0.16c, 0.9, and
1.5-1.8. For each we find t ( G ) by first applying the formulas found in
Corollary 1.8, Proposition 1.9 and Proposition 1.12, and second by
computing the (1,1)-cofactor of the Laplacian of G.
Example 1.4 a – Product of complete graphs K n × K m .
By Corollary 1.8 t ( K n × K m ) = (m + n)( n −1)( m −1) mm − 2 n n − 2 .

For K 2 × K 3 (Figure 0.16c) n = 2 and m = 3, so
t ( K 2 × K3 ) = (3 + 2)(2 −1)(3−1) 33− 2 22 − 2 = 523120 = 75.
 3 −1 −1 −1 0 0 
 −1 3 −1 0 −1 0 
 
 −1 −1 3 0 0 −1
One Laplacian matrix for K 2 × K 3 is   which
 −1 0 0 3 −1 −1
 0 −1 0 −1 3 −1
 
 0 0 −1 −1 −1 3
has (1,1)-cofactor 75 as well, confirming the result of the formula.
Example 1.4 b – Circulant graph Cn ( n1 ,..., nk ) .
(
By Proposition 1.9: t Cn ( n1 ,..., nk ) = )
 2n −1 n −1  k
 2π ni j  
∏  k − ∑ cos 
n
  if nk <
 n j =1 i =1
n   2

 n
−1
n .
 2n −1 2  k
 4π ni m   2  k
 2 ( 2m − 1) π ni 
∏ ∑ ∏ ∑
n
 k − cos   k − 1 − cos   if nk =
 n m =1 
 n   m =1   n   2
 i =1   i =1
n
For C8 (1,2 ) (Figure 0.7) k = 2, n = 8, n1 = 1, and n2 = 2. Thus, nk <
2
so we use the top formula, and
28−1 8 −1
 2
 2π ni j  
t ( C8 (1, 2 ) ) =
8 ∏ 2 −
j =1 
∑ cos 
i =1 8  
=
2 7 7
   2π n1 j   2π n2 j   
8 ∏  2 −  cos  8   + cos 
 8  
  =
j =1 
27    2π (1)(1)   2π ( 2 )(1)   
 2 −  cos   + cos     ×
8    8   8   

   2π (1)( 2 )   2π ( 2 )( 2 )   
 2 −  cos   + cos     ×
   8   8   
   2π (1)( 3)   2π ( 2 )( 3 )   
 2 −  cos   + cos     ×
   8   8   
   2π (1)( 4 )   2π ( 2 )( 4 )   
 2 −  cos   + cos     ×
   8   8   
   2π (1)( 5 )   2π ( 2 )( 5 )   
 2 −  cos   + cos     ×
   8   8   
   2π (1)( 6 )   2π ( 2 )( 6 )   
 2 −  cos   + cos     ×

    
 8   8
   2π (1)( 7 )   2π ( 2 )( 7 )   
 2 −  cos   + cos     = 3528.

    
 8   8
One Laplacian matrix for C8 (1,2 ) is

 4 −1 − 1 0 0 0 −1 −1
 −1 4 −1 −1 0 0 0 −1

 −1 −1 4 −1 −1 0 0 0
 
 0 −1 − 1 4 −1 −1 0 0 
 0 0 −1 −1 4 −1 − 1 0 
 
 0 0 0 −1 −1 4 −1 −1
 −1 0 0 0 −1 −1 4 −1
 
 −1 −1 0 0 0 −1 −1 4 
which has (1,1)-cofactor 3528 as well, confirming the result of the

formula.
Example 1.4 c – K n minus a matching.

k
 2
By Proposition 1.12 a): t ( K n − kK 2 ) = n n − 2  1 −  .
 n 
For K 6 − 2 K 2 (Figure 1.5), n = 6 and k = 2 , so
2
 2 4
t ( K 6 − 2 K 2 ) = 6 6 − 2  1 −  = 6 4   = 576.
 6 9
4 −1 −1 −1
−1 0
 −1 4 −1 −1
0 −1

 −1 −1 5 −1
−1 −1
One Laplacian matrix for K 6 − 2 K 2 is  
 −1 0 −1 4 −1
−1
0 −1 −1 −1
−1 4
 
 −1 −1 −1 −1 −1 5 
formula.
Example 1.4 d – Complete Bipartite K p ,q .
( )
By Proposition 1.12 b i): t K p , q = q p −1 p q −1 .
( )
For K 2,3 (Figure 1.6 a) p = 2 and q = 3, so t K 2,3 = 32−123−1 = 12.
3 0 −1 −1 −1
0 3 −1 −1 −1

One Laplacian matrix for K 2,3 is  −1 −1 2 0 0  which has
 
 −1 −1 0 2 0 
 −1 −1 0 0 2 
(1,1)-cofactor 12 as well, confirming the result of the formula.
Example 1.4 e – Regular complete k-partite K r ,r ,...r .
 
(r −1)k
By Proposition 1.12 b ii): t  K  = r rk − 2 k k −2 ( k − 1) .
 r ,r ,...r 
k terms
For K 2,2,2 (Figure 1.6 b) k =3 and r =2, so
(2 −1)3
( )
t K 2,2,2 = 22(3) − 233− 2 ( 3 − 1) = 24 ⋅ 3 ⋅ 23 = 384.
 4 0 −1 −1 −1 −1
 0 4 −1 −1 −1 −1

 −1 − 1 4 0 −1 −1
One Laplacian matrix for K 2,2,2 is   which
 −1 − 1 0 4 −1 −1
 −1 −1 −1 −1 4 0
 
 −1 −1 −1 −1 0 4 
has (1,1) cofactor 384 as well, confirming the result of the

formula.
Example 1.4 f – Complete bipartite with equal part sizes minus a full
matching K p , p − pK 2 .
p −1
( )
By Proposition 1.12 b iii): t K p , p − pK 2 = ( p − 1) p p − 2 ( p − 2 ) .
For K 3,3 − 3K 2 (Figure 1.6 c) p = 2 so
3 −1
( )
t K3,3 − 3K 2 = ( 3 − 1) ⋅ 33− 2 ⋅ ( 3 − 2 ) = 6.
 2 0 0 0 −1 −1
 0 2 0 −1 0 −1
 
 0 0 2 −1 −1 0 
One Laplacian matrix for K3,3 − 3K 2 is  
 0 −1 −1 2 0 0 
 −1 0 −1 0 2 0 
 
 −1 −1 0 0 0 2 
which has (1,1)-cofactor 6 as well, confirming the result of the formula.
Example 1.4 g – Complete graph of order n minus a path of order p

(n > p) K n − Pp .
By Proposition 1.12 c i):
  2 
p +1
 2  
p +1
 nn− p −2   n − 2 +  n − 2  − 1  −  n − 2 −  n − 2  − 1   if n > 4
  2  2    2  2   
2
=  2  n − 2  − 1      
(
t K n − Pp )  
    .
  2 
 3− p
 4 ⋅ p if n = 4
For K 5 − P3 (Figure 1.7) n = 5 and p = 3 thus the top formula is used, so

t ( K 5 − P3 ) =
 2 
3 +1
5 2 2  
3 +1
55 −3 − 2 5 − 2  5−2   −  5 − 2   
 2 +  2  − 1  −  2 −  2  − 1   = 21.
 5−2 
2
    
2  − 1     

 2 
 3 −1 −1 0 −1
 −1 3 0 −1 −1

One Laplacian matrix for K 5 − P3 is  −1 0 2 0 −1 which has
 
 0 −1 0 2 −1
 −1 −1 −1 −1 4 
Example 1.4 h – Complete graph of order n minus a cycle of order p

(n > p) K n − C p .
By Proposition 1.12 c ii): t ( K n − C p ) =
  2 
p
n−2 2 
p

 n − p −2  n − 2 n−2 n−2
   + 2 ( −1)  if n > 4
p +1
 n  +   − 1 + −   − 1
  2  2    2  2  
     
 p
 2 − 2 ( −1)
 if n = 4 and p = 3, 4
4
For K5 − C3 (Figure 1.8) n = 5 and p = 3, thus the top formula is used,
so t ( K5 − C3 ) =
 2
3
  2  
3
5 5 − 3− 2   5 − 2 +  5 − 2  − 1  −  5 − 2 −  5 − 2  − 1   = 47.
 2  2    2  2   
   
     
2 −1 0 −1 0
 −1 4 −1 −1−1

One Laplacian matrix for K 5 − C3 is  0 −1 2 −1 which has
0
 
0 −1 0 −1 2
 −1 −1 −1 −1 4 
1.4 Temperley’s B-Matrix
By Proposition 1.1 and Theorem 1.3, the Laplacian matrix H is a

singular matrix whose cofactors are all equal. Sometimes it is useful to
work with a non-singular matrix derived from H . Specifically, the
addition of J , where J is a square matrix with all entries equal to 1, to
H will yield such a non-singular matrix, called Temperley’s B-Matrix,
or just B [Temperley, 1964].
In classes of graphs with many edges, B’s eigenvalues are easier to
compute, as there are zero elements wherever a “−1” appears in H . The
eigenvalues of B and H are the same, save a zero eigenvalue of H and
an extra eigenvalue of n in B . This fact is an immediate corollary of the
next matrix-theoretic result:
Theorem 1.13 If L is a matrix whose rows or columns sum to zero,

then the characteristic polynomials of L and L + gJ (g an integer),
respectively, PL ( λ ) and PL + gJ ( λ ) are related as follows:
PL ( λ )( λ − gn ) = PL + gJ ( λ ) λ.
Proof Without loss of generality, assume all rows of L sum to zero.

λ − l11 −l12 −l13 ⋯ −l1n 
 −l λ − l22 −l23 ⋯ −l2 n 
PL ( λ ) = det ( λ I − L ) = det 
21
.
 ⋮ ⋮ ⋮ ⋱ ⋮ 
 
 −ln1 −ln 2 −ln 3 ⋯ λ − lnn 
Addition of all the columns to the first yields:
λ −l12 −l13 ⋯ −l1n 
λ λ − l −l23 ⋯ −l2n 
PL ( λ ) = det 
22
=
⋮ ⋮ ⋮ ⋱ ⋮ 
 
 λ −ln 2 −ln3 ⋯ λ − lnn 
1 −l12 −l13 ⋯ 
−l1n
1 λ − l −l23 ⋯ 
−l2 n n
λ det 
⋮ ⋮
22
⋮ ⋱ ⋮  i =1
∑
 = λ C (λI − L ).
i1
 
1 −ln 2 −ln 3 ⋯ λ − lnn 
Similarly, consider PL + gJ ( λ ) =
λ − l11 − g − l12 − g −l13 − g ⋯ −l1n − g 

 −l − g λ − l22 − g −l23 − g ⋯ −l2 n − g 
det 
21
.
 ⋮ ⋮ ⋮ ⋱ ⋮ 
 
 −ln1 − g −ln 2 − g −ln3 − g ⋯ λ − lnn − g 
Addition of all columns to the first yields:
 λ − gn −l12 − g −l13 − g ⋯ −l1n − g 
 λ − gn λ − l − g −l23 − g ⋯ −l2 n − g 
det 
22
.
 ⋮ ⋮ ⋮ ⋱ ⋮ 
 
 λ − gn −ln 2 − g −ln3 − g ⋯ λ − lnn − g 
Factoring out the first column leaves all ones there, and the subsequent
addition of g times this column to all the others yields:
1 −l12 −l13 ⋯ −l1n 

1 λ − l −l23 ⋯ −l2 n 
(λ − gn) det 
22
.
⋮ ⋮ ⋮ ⋱ ⋮ 
 
1 −ln 2 −ln 3 ⋯ λ − lnn 
n
1 1
Therefore, ∑ C ( λI − L ) =
i =1
i1
λ
PL ( λ ) = P
λ − gn L + gJ
(λ) . □
Corollary 1.14 follows immediately:
Corollary 1.14 The characteristic polynomials for a Laplacian matrix

H and its corresponding B -matrix differ by a single factor in each, i.e.
1 1
PH ( λ ) = P ( B, λ ) .
λ λ −n B
Proof In Theorem 1.13 setting L = H and g = 1, then
PH ( λ )( λ − n ) = PL ( λ )( λ − n ) = PL + J ( λ ) λ = PB ( λ ) λ □
Theorem 1.15 Suppose the complement of the graph G consists of k

node disjoint subgraphs C1 ,C 2 ,..., C k so that G is the join of
C 1 , C 2 ,..., C k , the complements on n1 , n2 ,..., nk nodes, respectively. If

k
1
Gi = C i + K n − ni for each i, then t ( G ) = ( k −1) n − 2( k −1)
n
∏ t ( G ).
i =1
i In
1
particular, if k = 2, t ( G ) = n−2
t ( G1 ) t ( G2 ) .
n
 H (C 1 ) −1 ⋯ −1 
 
 −1 H (C 2 ) −1 ⋮ 
Proof Realize that H ( G ) =   so that
⋮ −1 ⋱ −1 
 
 −1 ⋯ −1 H (C k ) 
B (C1 ) 0 ⋯ 0 
 
0 B (C 2 ) 0 ⋮ 
B (G ) =   and
⋮ 0 ⋱ 0 
 
 0 ⋯ 0 B (C k ) 
k
1 1
t (G ) =
n 2
det ( B ( G ) ) = 2
n
∏ det ( B (C ) ).
i =1
i
B (C i ) 0 
Now, B ( Gi ) = 
 0
( )
 , so det B ( Gi ) = n n − ni det B (C i ) and
nI n − ni 
and ( )
k
1
t ( Gi ) =
n 2 ( ) (
det B ( Gi ) = n n − ni − 2 det B (C i ) . Thus, ) ∏ t (G ) =
i =1
i
k k
∏ ( )
n n − ni − 2 det B (C i ) = n kn − n − 2 k ∏ det ( B (C ) ) =n( i
k −1) n − 2 k
n 2t ( G ) .
i =1 i =1
k
1
Therefore, t ( G ) = ( k −1) n − 2( k −1) ∏ t ( G ). i □
n i =1
The following general theorem suggests a relationship between the

number of spanning trees and the cofactors of B .
Theorem 1.16 Given any equicofactor matrix L , where C denotes the

common value of all cofactors of H, and C ≠ 0, then
det ( L + gJ ) = gn2C ( L ) .
Proof By Remark 1.2 (following the proof of Theorem 1.3), L has row
and column sum zero. Thus, the addition of all columns of L + gJ to the
first yields:
 gn l12 + g l13 + g ⋯ l1n + g 
 gn l + g l + g ⋯ l2 n + g 
det 
22 23
=
⋮ ⋮ ⋮ ⋱ ⋮ 
 
 gn ln 2 + g ln 3 + g ⋯ lnn + g 
1 l12 + g l13 + g ⋯ l1n + g 

1 l + g l + g ⋯ l2 n + g 
gn det 
22 23
.
⋮ ⋮ ⋮ ⋱ ⋮ 
 
1 ln 2 + g ln 3 + g ⋯ lnn + g 
Subtraction of g times column 1 from all the other columns yields
1 l12 l13 ⋯ l1n 

1 l l23 ⋯ l2 n 
gn det 
22
= gn2 C ( L ) upon expansion about column
⋮ ⋮ ⋮ ⋱ ⋮ 
 
1 ln 2 ln 3 ⋯ lnn 
one. □
Theorem 1.16 leads directly to the following result:
Corollary 1.17 The number of spanning trees for a graph can be

determined from its B-matrix by the formula
det ( B ) = n 2 C ( H ) = n 2 ( t ( G ) ) .
1.5 Multigraphs
Most of the classical results regarding the eigenvalues of the Laplacian

matrix are only valid for graphs because they depend on the existence of
the complement of a graph, which is formed by removing the edges of
the graph in question from the complete graph. It is easy to see that the
Laplacian matrix of the complement of a graph is given by
H = nI − J − H . Henceforth, we shall refer to H given as above, as the
algebraic complement of H (or G) even in the event that G is a
multigraph. In the case of a multigraph, H has no graph theoretical
interpretation.
Some of the useful results one obtains concerning H and B in the
graph case are lost when multigraphs are considered. For example, since
the Laplacian matrix of a multigraph is non-negative definite, symmetric
and real, the eigenvalues are non-negative. Hence, in the event that G is a
graph, so that H is the Laplacian matrix of the complement G , we claim
that the maximum eigenvalue of H and B is n. Indeed, the eigenvalues
of B = nI − H are given by n − λ ( H ) where λ ( H ) runs through the
eigenvalues of H .
As the minimum eigenvalue of H is 0, the maximum eigenvalue of
B (and thus of H as well) is n. This fact does not generalize to
multigraphs, as demonstrated by the multigraph given in Figure 1.9.
Figure 1.9: A multigraph in the class Ω ( 5,8 ) .

 3 −1 −1 −1 0 
 −1 3 −1 −1 0 
 
Its Laplacian matrix is H =  −1 −1 3 0 −1 which has
 
 −1 −1 0 4 − 2 
 0 0 −1 −2 3 
maximum eigenvalue approximately equal to 6.214 > 5. Another well-

known result for graphs states that λn = n if and only if the complement
is disconnected (see Proposition 1.6), and the multiplicity of n as an
eigenvalue of H is the number of components of G [Kelmans 1967].
This result has no interpretation in the multigraph setting.
The bounds of Fiedler on the second eigenvalue, namely, λ2 ≤ κ
(node connectivity) and λ2 ≤ δ (minimum node degree), respectively,
(see Proposition 1.4 (c) and Remark 1.4) are invalid in the multigraph
case. A counterexample is given by the multigraph in Figure 1.10. For
this graph, κ = 2 and δ = 4 , and its Laplacian matrix has λ2 = 6 .
Figure 1.10: A multigraph that serves as a counterexample to the bounds

of Fiedler and Kelmans.
Another eigenvalue bound which does not apply to multigraphs states

that λk ≥ χ where Kχ is the largest complete subgraph and
n − χ − 2 ≤ k ≤ n. Consider the multigraph given in Figure 1.11.

Figure 1.11: A multigraph that serves as a counterexample to a

graph eigenvalue bound.
For this graph, χ = 3 and n − χ − 2 = 5 − 3 − 2 ≤ 2 ≤ 5 but we have

λ2 ≈ 1.6571 < χ .
1.6 Eigenvalue Bounds for Multigraphs
We proceed to derive some bounds on the eigenvalues of the Laplacian

matrix of a multigraph, which generalize known results for graphs.
In multigraphs where the largest multiplicity of any edge is g, it is
sometimes useful to form a “new” B-matrix, B̂ by adding g J instead of
J to H . An immediate consequence of Theorem 1.13 is the following
corollary:
Corollary 1.18 Let M be a multigraph such that the largest multiplicity

ˆ = H + gJ , then
of any edge is g. Given H and B
1 1
PH ( λ ) = Pˆ ( λ ) ,
λ λ − gn B
and the number of spanning trees is given by

n
1
gn 2
∏ λ ( Bˆ ) ,
i =1
i ( ) ( )
where λ1 Bˆ ,..., λn Bˆ are the eigenvalues of B̂ .
The second eigenvalue for a graph is known to be bounded by

2δ − ( n − 2 ) ≤ λ2 ≤ κ (Proposition 1.6 (c)) . For multigraphs, we shall
prove that 2δ − g ( n − 2 ) ≤ λ2 ≤ gκ . The proof of λ2 ≤ gκ is deferred

until later in the discussion.
Proposition 1.19 If H ( M ) is the Laplacian matrix of multigraph M
with eigenvalues 0 = λ1 ( M ) ≤ λ2 ( M ) ≤ ⋯ ≤ λn ( M ) , then
λ2 ( M ) ≥ 2δ − g ( n − 2 ) ,
where the largest multiplicity of any edge is g. Furthermore, this bound

is best possible.
Proof Consider bîi − ∑ bîj = hii + g − ∑ hij + g . Note that

j ≠i j ≠i
∑h
j ≠i
ij + g = g ( n − 1) − di , since hij + g = g − α ij , where α ij is the
number of parallel edges between nodes i and j. Since ∑α

j ≠i
ij = di
where hii = d i , we have, by the bound of Gersgorin (Corollary 0.8),
( )
λ0 Bˆ ≥ bîi − ∑ bîj . Thus,
j ≠i
{
λ1 ( M ) ≥ min i di + g − ( g ( n − 1) − di ) = }
min i {2 di − g ( n − 2 )} ≥ 2δ − g ( n − 2 ) .
Now, to see that the bound is best possible consider

g  K k + ( K a ∪ K a )  depicted in Figure 1.12, where gG is the graph
obtained from G by replicating each edge g times. Here,
2δ − g ( n − 2 ) = 2 g ( a − 1 + k ) − g ( k + 2a − 2 ) = gk = gκ . □
gK a
+g
gK k
+g
gK a
Figure 1.12: The graph that realizes the bounds. Note that there are g edges
between each node in gK k and each node in both gK a .
In preparation for our next result, we determine the characteristic

polynomial of the Laplacian of a multistar. A multistar is a multigraph
having a star as its underlying graph. In a graph, ∆ ( G ) is the maximum
degree, which is also the number of leaf nodes in the star of largest order
that is a subgraph of the graph.
Proposition 1.20 Let M be a multistar and let mi be the number of

edges in M between the center node and ith leaf node of the underlying
star. The characteristic polynomial for the Laplacian matrix H of M is

given by
 ∆ ∆ ∆ 
PH ( M ) ( λ ) = λ 
 ∏ ( λ − m ) − ∑ m ∏ ( λ − m )  ,
i i j
 i =1 i =1 i≠ j 
where ∆ represents the maximum degree of the underlying star.
Proof A multistar as described above has Laplacian matrix
 ∆  
  i =1 
∑
  mi  − m1 ⋅ ⋅ − m∆ 

 −m m1 0 ⋅ 0 
H= 1 .
 − m2 0 m2 0 ⋅ 
 
 ⋅ ⋅ ⋅ ⋅ ⋅ 
 − m∆ 0 ⋅ 0 m∆ 
This matrix has characteristic polynomial
∆
 ∆ ∆
(m ) 2 
PH ( M ) ( λ ) = ∏ ( λ − mi ) λ − ∑ mi − ∑ (λ −i m ) 
i =1  i =1 i =1 i 
because
a x1 x2 ⋅ xn 
x y1 0 ⋅ 0 
 1 n n n
∑( x ) ∏ y .
2
det  x2 0 y2 0 ⋅ =a ∏ yi − i j
  i =1 i =1 i≠ j
⋅ ⋅ ⋅ ⋅ ⋅ 
 xn 0 ⋅ 0 yn 

Now, PH ( M ) ( λ ) simplifies to:

∆  ∆
mi   ∆ ∆ ∆ 
∏ ( λ − mi )  λ − λ ∑
  = λ 
( λ − mi )   ∏ ( λ − mi ) − ∑ mi ∏ λ − m j  .
( )
i =1  i =1 i =1 i =1 i≠ j 
□
A multistar with underlying graph K1,q whose q leaf nodes all have
common degree, say f, is denoted f K1,q . Such a graph has eigenvalues 0,
f with multiplicity q − 1 , and (1 + q ) f . Given a multigraph M with

underlying graph G, let u1 , u2 ,… , ur be the nodes of G of maximum
degree ∆ . For each ui consider the multistar centered at ui with all leaf
nodes having degree fi equal to the minimum multiplicity of all the
edges incident at ui . We denote this multistar by fi K1, ∆ .
Theorem 1.21 Let M be a multigraph and ∆ be the maximum degree of

the underlying graph. Furthermore let f be the largest value such that
f K1, ∆ is a submultigraph of M. The lower bound on the largest
eigenvalue of the Laplacian matrix of M is λn ≥ ( ∆ + 1) f .
Proof Express the Laplacian matrix of a multigraph as the sum of the

Laplacian matrix of the spanning subgraph formed by the edges of the
f K1, ∆ contained in that graph and the Laplacian matrix of the spanning
subgraph formed by the remaining edges. The f K1, ∆ Laplacian matrix
has eigenvalues as described above, i.e. 0, f with multiplicity ∆ − 1, and
(1 + ∆ ) f . Referring to Corollary 0.6, let B represent the Laplacian

matrix for the multstar and A the Laplacian matrix for the rest of the
graph. Then λn ( M ) ≥ (1 + ∆ ) f . □
Example 1.5 Consider the multigraph given in Figure 1.13; its

underlying graph has ∆ = 3 and contains 3K1,3 as a submultigraph. The
largest eigenvalue is 13.424 > (1 + ∆ ) f = (1 + 3)3 = 12 .
Figure 1.13: A multigraph containing 3K1,3 .
Remark 1.6 This subsumes Kelmans’ result [Kelmans, 1974], which

states that the largest eigenvalue of the Laplacian of a graph is ∆ + 1 ,
because in a graph, f = 1 .
1.7 Multigraph Complements
Consider a multigraph M = (V , m ) of order n having Laplacian matrix
H , and g the maximum edge multiplicity of M. Consider gK n on V,
and define M ( g ) to be the multigraph obtained by removing all the

edges of M from gK n . The following set of results extend eigenvalue
results to multigraphs, where g is as stated and f the minimum edge
multiplicity.
Theorem 1.22 If 0 = λ1 ≤ λ2 ≤ ... ≤ λn are the eigenvalues of the

Laplacian matrix of some multigraph M with maximum edge multiplicity
( )
g, then λi M ( ) = ng − λn −i + 2 for 2 ≤ n are the eigenvalues of
g
( )
H M ( g ) = g ( nI − J ) − H ( M ) . Furthermore, λn = ng if and only if
M (g) is disconnected.
(
Proof For ease of notation, let H = H ( M ) and H = H M ( g ) . Let λ )
be a nonzero eigenvalue of H , and x an associated eigenvector, i.e.,
Hx = λ x. Then 1T H = 0 implies 1T x = 0. Thus, it follows from
Hx + Hx = g ( nI − J ) x that Hx = ( gn − λ ) x. Likewise, H1+ Η1 =
0 + 0 = 0. Hence, if 1, x 2 , x3 ,..., xn is a basis for the eigenvectors

for H , then they also constitute a basis for the eigenvectors of
H . Thus, the eigenvalues of H are nonnegative and equal
ng − λ n ≤ ng − λ n −1 ≤ ⋅ ⋅ ⋅ ≤ ng − λ1. It is known that a multigraph, in
particular M ( g ) is disconnected if and only if rank ( H ) ≤ n − 2.
Therefore, ng − λn ( M ) = 0 if and only if M is disconnected. □
Theorem 1.23 If M k is obtained by removing any k nodes from

a multigraph M with maximum edge multiplicity g, then
λ2 ( M k ) ≥ λ2 ( M ) − gk .
Proof Without loss of generality, assume M is connected, or else λ2 = 0,

and the inequality follows trivially. The proof proceeds by induction on

k, the number of removed nodes.
For the base case, k = 1 , let M 1 denote the multigraph obtained from M
by removing a single node, and M ′ be obtained from M 1 by adding a
node joined to each of the remaining nodes v2 , v3 ,..., vn by g edges. The
Laplacian matrix of M ′ has (1,1)-entry g (n − 1), and all the other entries
in the first row and column are − g. Furthermore,
H11 ( M ′ ) = H11 ( M ) + gI = H ( M1 ) + gI.
Let x be the eigenvector of H ( M 1 ) corresponding to λ2 , so that
Hx = λ2 x. Now,
0  − g1T x  0

H ( M ′)   =  ( )
 = λ2 ( M 1 ) + g  
x ( )
 λ2 ( M 1 ) + g x  x
as λ2 ≥ 0 implies 1T x = 0. Hence, λ2 ( M1 ) + g is an eigenvalue of
H ( M ′ ) and λ2 ( M 1 ) + g ≥ λ2 ( M ′ ) , by Corollary 0.6.
Next, assume that λ2 ( M k −1 ) + g ( k − 1) ≥ λ2 ( M ) , where M k −1 is the

multigraph obtained by the removal of k − 1 nodes from M. Consider a
multigraph M k obtained from M by the removal of any k nodes from M,
and let M k −1 denote the multigraph obtained by removing any k − 1 of
those nodes from M. It remains to show that λ2 ( M k ) + g ≥ λ2 ( M k −1 ) .

But this follows exactly as in the base case in the event that M k −1 is
connected. If M k −1 is not connected, the inequality follows trivially. □
Note: the preceding proof is a minor variation on Fiedler’s proof, which

applied only to g = 1. [Fieldler, 1973]
Theorem 1.23 has two immediate corollaries:
Corollary 1.24 If M is a connected multigraph with maximum edge

multiplicity g and node connectivity κ , then λ2 ( M ) ≤ gκ .
Proof Let S be a set of nodes of order κ such that M κ = M − S .
Since Mκ is disconnected λ2 ( M κ ) = 0 . By Theorem 1.23
λ2 ( M κ ) ≥ λ2 ( M ) − gκ it follows that that λ2 ( M ) ≤ gκ . □
Remark 1.7 Combining the results of Proposition 1.19 and Theorem

1.23, we have 2δ − g ( n − 2 ) ≤ λ2 ( M ) ≤ g ∆, where ∆ is the maximum
degree of the underlying graph. Examination of the proof of the fact that
2δ − g ( n − 2 ) ≤ λ2 ( M ) is best possible shows that there is a
multigraph for g ∆ = 2δ − g ( n − 2 ) . Hence, λ2 ( M ) ≤ g ∆ is also best

possible.
Corollary 1.25 The largest eigenvalue of a multigraph M is bounded
(
from below: λn ( M ) ≥ g n − ∆ M ( g )( )), where ∆ is the maximum
degree of the underlying graph.
( ) ( ) (
Proof λn ( M ) = ng − λ2 M ( g ) ≥ ng − g ∆ M ( g ) = g n − ∆ M ( g ) ( )). □
There are some results for the Laplacian matrix that rely on some
techniques of matrix theory. The Laplacian matrix, besides being real
and obviously symmetric, is positive semi-definite, which means that
xT Hx ≥ 0 for all n-dimensional column vectors x, and there exists a

non-zero vector x such that xT Hx = 0 . [Boesch, 1984]. This fact leads
to another result:
Theorem 1.26 Let P be a real, symmetric, positive semi-definite n × n

matrix such P1 = 0. Then, if λ1 ( P ) ≤ λ2 ( P ) ≤ ⋯ ≤ λn ( P ) are the
n
eigenvalues of P , we have λ2 ( P ) ≤ min i  pii .
n −1
Proof Observe that λ 2 = min xT Px with x ⊥ 1, x = 1, by Theorem 0.4.
 1 
The reasoning is as follows: let  , x 2 , x 3 ,…, x n  be an orthonormal
 n 
n
basis of eigenvectors, and consider x = ∑α x .
i =2
i i Now xT Px =
n
≥ λ 2 x . Also, xT Px = λ 2 . Consider the matrix Pɶ = P −
2
∑λ α
i =2
i
2
i
λ 2 ( I − n −1J ) . If y ∈ ℝn , write y = c11 + c2 x, where x is orthogonal to
1. Then consider y T Py = c2 xT Pc2 x which equals
( ) (
c22 xT Px − λ2 xT ( I n − n −1J n ) x ≥ c22 λ2 x − λ2 x
2 2
) = 0.
Therefore, P is non-negative definite, and thus,
  1 
min i  mii − λ2  1 −   ≥ 0. □
  n 
There are two immediate corollaries to this result.

n
Corollary 1.27 For any multigraph M, λ2 ( M ) ≤ δ.
n −1
n n
Proof Clearly, λ2 ( M ) ≤ min  hii  = δ. □
n −1 n −1
n
Corollary 1.28 For any multigraph M, λn ( M ) ≥ ∆.
n −1
Proof Since ∆ ( M ) = ( n − 1) g − δ ( M ) ,
(
δ M (g) = ) ( n − 1) g − ∆ ( M ) .
But
n n n
( )
λ2 M ( g ) ≤
n −1
(
δ M (g) = )
n −1
( ( n − 1) g − ∆ ( M ) ) = ng −
n −1
∆ ( M ).
Hence,
n n
( )
ng − λn ( M ) = λ2 M ( ) ≤ ng −
g
n −1
∆ ( M ) , or λn ( M ) ≥
n −1
∆ ( M ).
1.8 Two Maximum Tree Results
It can be said that K n has the greatest number of spanning trees of any
graph in the same class. As a matter of fact, K n is the only graph in its
class, so the result is obvious. However, it is also true that K n has the
greatest number of spanning trees among all multigraphs in its class, and
that it is unique in this regard. This was originally established by
[Kelmans, 1974]. We include a proof for completeness. First, because
the eigenvalues of H ( K n ) are given by 0, n,..., n, Cayley’s Theorem
1 n 1
(Proposition 1.6(a)) follows: t ( K n ) = ∏ λi = n n −1 = n n − 2 . Since
n i =2 n
the diagonal entries of the Laplacian matrix are the degrees of the nodes,
n
trace ( H ) = 2e = ∑ λi , which is the same for all multigraphs in a given
i =1
class. In this case, the sum is n ( n − 1) . By a basic result of optimization

theory, such products are maximized when all factors are equal. Thus, no
 n
other multigraph with e =   edges can have more spanning trees.
 2
Now, prior to demonstrating uniqueness, it is first observed that two
non-isomorphic graphs in the same class can have the same eigenvalues,
so uniqueness of the eigenvalues is not a triviality. Two non-isomorphic
graphs with the same eigenvalues are called cospectral. Consider the
two graphs in Figure 1.14. While both are in the class Ω ( 6,9 ) , and
regular of degree 3, they are not isomorphic. However, calculation of the
eigenvalues of their respective Laplacian matrices yields the identical
sequence (0, 1.438, 3, 4, 4, 5.562).
Figure 1.14: Two non-isomorphic cospectral graphs.

We prove that K n is unique in this regard by contradiction. Suppose

there exists some multigraph which has eigenvalue n with multiplicity
n − 1 . Thus, the null space of (H − nI ) has dimension n − 1 , or
rank ( H − nI ) =1. This fact, together with the symmetric nature of H ,
implies that H − nI = χ J for some constant χ . In the graph K n , all

nodes are adjacent. Any multigraph in the class will have at least one
pair of nonadjacent nodes, i.e., some hij = 0 in H . It follows that
χ J = 0, or χ = 0, which contradicts rank = 1. Hence, we have

established the next theorem:
Theorem 1.29 The complete graph K n is the unique multigraph in its

class having the greatest number of spanning trees.
The following theorem extends this result. After obtaining this result
we learned that apparently it was also proven by Kelmans [Kelmans,
1967]. However, since the original paper has never been translated into
English, we shall provide a proof here, which first appeared in [Heinig,
2013]:
Theorem 1.30 The multigraph gK n has greatest number of spanning

trees over all multigraphs in its class, and it is unique in this regard.
Proof Indeed,
Bˆ ( gK n ) = H ( gK n ) + gJ =
 g ( n − 1) −g ⋯ −g  g g ⋯ g
 
 −g g ( n − 1) ⋯ − g   g g ⋯ g 
+ =
 ⋮ ⋮ ⋱ ⋮  ⋮ ⋮ ⋱ ⋮
   
 − g ⋯ −g g ( n − 1)   g g ⋯ g
 gn 0 ⋯ 0 
 0 gn 0 ⋮ 
 .
⋮ ⋮ ⋱ ⋮ 
 
 0 ⋯ 0 gn 
n
The eigenvalues have product ( gn ) . To prove uniqueness, suppose that
ˆ ( M ) has n eigenvalues equal to

M is a different multigraph, and that B
gn. Then there is an orthonormal matrix such that
S −1Bˆ ( M ) S = gnI = gnSS −1. Then S −1Bˆ ( M ) S = gnI ⇒
SS −1Bˆ ( M ) SS −1 = SgnIS −1 ⇒ Bˆ ( M ) = SgnIS −1 = gnISS −1 = gnI.
ˆ ( M ) = ngSS −1 = ngI, and M = gK .

Therefore, B □
n
Remark 1.8 Of course, this subsumes the case 1K n = K n and provides a

much simpler proof than that given previously.
Chapter 2
Multigraphs with the Maximum

Number of Spanning Trees: An
Analytic Approach
In this chapter, we consider the optimization problem of maximizing the

number of spanning trees among all multigraphs in Ω ( n, e ) , i.e., among
all multigraphs having n nodes and e edges, for certain values of e. Since
a spanning tree is a minimally connected subgraph, multigraphs having
more of these are, in some sense, more immune to disconnection by edge
failure.
A result of Cheng [Cheng, 1981] in optimization theory states that a
product of positive elements having fixed sum A and fixed sum of
squares C is maximized when the elements are all the same or “almost”
the same. He made use of this result to find the multigraph that
maximizes the number of spanning trees for two different values of e,
and the following is an exposition and expansion of his treatment.
2.1 The Maximum Spanning Tree Problem
With t ( M ) denoting the number of spanning trees of the multigraph M,
we define the maximum spanning tree problem, Pt , by
93
Problem Pt :
Maximize t ( M ) subject to
M ∈ Ω ( n, e ) .
Since by Theorem 1.5, the number of spanning trees of a multigraph

M is related to the eigenvalues 0 = λ1 ≤ λ2 ≤ ... ≤ λn of its Laplacian
n
1
matrix H ( M ) as follows: t ( M ) = ∏ λ (M ),
i we observe that the
n i =2
following problem is a relaxation of Pt ,
Problem Ph :
1 n
Maximize ∏ λi subject to
n i=2
1. {0, λ1 , λ2 ,… , λn } is the spectrum of an n × n symmetric matrix
H =  hij  of integers, with each hij ≤ 0 for i ≠ j .

n
2. ∑h
j =1
ij = 0 for each i.
n
3. ∑h
i =1
ii = 2e .
Note that this relaxation contains instances which are not feasible for
Pt because, if H =  hij  is the Laplacian of a multigraph M then the
main diagonal terms must also meet the condition: if hkk = max hii then
1≤ i ≤ n
Multigraphs with the Maximum Number of Spanning Trees: An Analytic Approach 95
hkk ≤ ∑ hii [Hakimi, 1962]. The reader will also observe that both Pt and
i≠k
Ph have finitely many feasible solutions. Since H is a real, symmetric,

positive semi-definite matrix all λi 's ≥ 0 , by Proposition 0.7
n
and trace ( H ) = ∑ λi , by Theorem 0.11. Thus the problem
i=2
Problem Ps :
1 n
Maximize ∏ λi subject to
n i =2
1. λi ≥ 0 for all i = 2,3,…, n
n
2. ∑λ
i=2
i = 2e ,
is a relaxation of Ph with infinitely many instances.

A sometimes fortuitous approach to solving a constrained
optimization problem is to obtain a relaxation that has a readily
obtainable solution which is feasible for the original problem. Even
though Ps is a relaxation of Pt we are not able to use this simple
approach in this case. Rather, we add an additional constraint to both Ph
n
and Ps of the form ∑λ
i=2
i
2
= C . Realize that this condition can be written
as
n
∑h + ∑h
i =1
2
ii
i≠ j
2
ij =C
in the context of Ph . We refer to the newly obtained problems as Ph ,C
and Ps ,C and impose conditions on C and e that guarantee the existence

of a solution obtainable by the Lagrange multiplier technique. Broadly
speaking, we determine that the value of the optimal solution of Ps ,C
increases as C decreases. Thus by determining C m , the minimum C over

the finite collection of C’s which yield feasible solutions for Ph ,C , we
obtain a largest optimal for such a problem. Since each instance of Pt is

feasible for one of the Ph ,C ’s, if there is an instance of Pt that is feasible
for Ph ,Cm then a solution for Pt is obtained. As it happens there are two
such cases that we shall present.
To begin with we give necessary and sufficient conditions on C and e
to insure that Ps ,C is well-posed, i.e. has feasible solutions and an
optimal objective value.
Lemma 2.1 Problem Ps ,C is well-posed and has a solution if and only if
4e 2
≤ C ≤ 4e 2 . Furthermore, the set of feasible solutions contains an n-
n −1
4e 2
tuple with all positive entries if and only if ≤ C < 4e 2 .
n −1
Proof First suppose that λ2 , λ3 ,…, λn satisfies the constraints. Then, with
λ = ( λ 2 , λ3 ,… , λ n ) and 1 = (1,1,… ,1) , the Cauchy-Schwarz Inequality

4e 2
forces 2e = λ ⋅ 1 ≤ λ 1 = C n − 1 . Thus ≤ C . Of course, since
n −1
2
 n 
each λi ≥ 0 and 4e =  ∑ λi  it follows that 4e 2 = C + 2∑ λi λ j ≥ C and
2
 i =2  i< j
4e 2
so ≤ C ≤ 4e 2 . Moreover, if one of the λi ' s is positive, 4e 2 > C .
n −1
Conversely, suppose the inequalities hold. In the event that C = 4e2 , the
n − 1 -tuple ( 2e,0,0,… ,0 ) satisfies the constraints. Next suppose that
4e 2 2e
≤ C < 4e 2 . Realize that λ = 1 terminates on the plane
n −1 n −1
 2e   2e  2e
λ ⋅ 1 = 2e . Now (2e, 0, 0,… ,0 ) − 1 ⋅  1 = 0 so x = 1
 n −1  n −1  n −1
4e 2
and y = C − u , where u is the unit vector in the direction of
n −1
2e
( 2e,0,0,…,0) − 1 , are perpendicular. We claim that λ = x + y is a
n −1
vector with positive components that satisfy the constraints of Ps ,C .
Indeed,
λ ⋅ 1 = x ⋅ 1 + y ⋅ 1 = x ⋅ 1 = 2e .
Also
2
2  2e  4e 2
λ =  + C − =C.
 n −1  n −1
Next
( 2e,0,… ,0 ) − 
2e 2e   1 1 1  1
,… ,  = 2e  1 − ,− ,… , −  = 2e 1 −
 n −1 n −1  n −1 n −1 n −1 n −1
1  1 1 1 
so that u = 1 − ,− ,… , −  . Therefore
1 n −1 n −1 n − 1
1−
n −1
2e 4e 2 1
λ2 = + C− 1− >0
n −1 n −1 n −1
and, for i > 3
 
2e 4e   12
1 
λi = − C− > 0.
n −1 n −1  1 n −1 
 1− 
 n −1 
1 n
Of course the continuous function ∏ λi achieves a maximum value
n i=2
on the compact set defined by the constraints. □
4e 2
Remark 2.1 If ≤ C ≤ 4e2 holds, then Problem Ps ,C has an optimal
n −1
solution, because the objective function is continuous on the nonempty
compact set defined by the constraints.
Since the optimal solution of Problem Ps ,C has positive entries
4e 2
whenever ≤ C < 4e2 , replacement of constraint (1) in Ps ,C by the
n −1
requirement that all λi be strictly positive, yields a problem having the
solution. We designate this problem as Problem Ps,C , p :
Problem Ps ,C , p
n
1
Maximize
n ∏λ i =2
i subject to
1. λi > 0, i = 2,3,..., n
n
2. ∑λ
i =2
i = 2e
n
3. ∑ (λ )
i=2
i
2
=C .
Now, (1) of Ps,C , p defines an open set of ℝ n −1 . Recall that the

LaGrange multiplier method applies to problems of the following type:
those with a continuously differentiable objective function and
continuously differentiable equality constraints, all defined on an open
subset of ℝ n −1 which determine a Jacobian of full rank. Thus, we may
determine the optimal solutions of Ps,C , p with the aid of the LaGrange
multiplier technique if we also assume C < 4e2 to ensure a Jacobian of

full rank.
4e 2
Proposition 2.2 In the event that < C < 4e 2 , and problems Ps ,C and
n −1
Ps,C , p have solutions at a location λ * = ( λ 2* , λ3* ,..., λ n* ), then λ * has
exactly two distinct entries.
Proof To apply the LaGrange multiplier technique, we must consider the

 1 1 ... 1 
Jacobian of the constraints, i.e., ℑ =  . The entries
 2λ2 2λ3 ... 2λn 
are continuous on ℝ n −1 and therefore also in a neighborhood of λ * .

Furthermore, rank ( ℑ ) < 2 implies that all λi are equal, which is
4e 2
impossible, because C > . Therefore, rank ( ℑ ) = 2 .
n −1
n
In addition, f ( λ ) = ∏λ
i =2
i has continuous partials in a neighborhood of
λ * so that, if we consider the function

n
 n
  n

LL((λλ,,α11, α 2 ) = ∏ λi ++αα11 ∑ λλi −−22ee ++αα2 2 ∑ λi2λ− −C C
i =2  i=2   i =2 
then there exists α1* , α 2* such that L has a critical point at ( λ * , α1* , α 2* ) .
∂L
In particular,
∂λi
= ∏λ
j ≠i
j + α1 + 2α 2 λi = 0 for i = 2,..., n must be
satisfied by ( λ * , α1* , α 2* ) . Suppose there exists λi1 * < λi2 * < λi3 * (i.e., three
distinct entries). Then
∂L *
1. = ∏λ j + α1 + 2α 2 λi*1 = 0 ,
∂λi*1 j ≠ i1
∂L *
2. = ∏λ j + α1 + 2α 2 λi*2 = 0 ,
∂λi*2 j ≠ i2
∂L *
3. = ∏λ j + α1 + 2α 2 λi*3 = 0 .
∂λi*3 j ≠ i3
1 *
Subtracting the first two equations leads to α 2 =
2 ∏λ
j ≠ i1 ,i2
j , while
1 *
subtracting the last two equations yields α 2 =
2 ∏λ
j ≠i2 ,i3
j , and putting
these together implies that λi*1 = λi*3 , which contradicts λi*1 < λi*3 . Therefore,
there are only two distinct entries. □
Proposition 2.3
(i) The collection of m-tuples λ in which there are only two distinct
entries and which satisfy
1. λ ⋅ 1ɶ = 2e
2  4e 2 
2. λ = C , λi > 0 (where  where < C < 4e 2  )
 n −1 
is finite in number. The possible multiplicities for the larger entry L are
1, 2,… , n − 2 . For each such multiplicity k,
 ( n − 1)( n − 1 − k )   4e 2 
2e +    C − 
 k  n −1
L(k ,2e, C ) = and
n −1
 (n − i)k  4e 2 
2e −   C − 
 n −1− k  n −1
S (k , 2e, C ) =
n −1
for the larger and smaller entries L ( k , 2e, C ) and S ( k , 2e, C ) ,
respectively, or simply, just L and S.
n
1
(ii) The objective function
n −1 ∏λ
i=2
i achieves its maximum only at
k = 1 , i.e., the solution of Problem Ps ,C is uniquely attained at
λ * = ( L(1, 2e, C ), S (1, 2e, C ),..., S (1, 2e, C )). This location is unique up to
permutation of the entries.
2e = kL + ( n − 1 − k ) S
Proof (i) Consider the system:
C = kL2 + (n − 1 − k ) S 2 .
2e − kL
Substituting S = from the first equation into the second leads
n −1 − k
to ( n − 1) kL2 − 4ekL + [4e2 + C (k − n + 1)] = 0 and application of the
quadratic formula and simplification leads to the formula for L and S.
2e = kL + ( n − 1 − k ) S
(ii) Consider the system as before:
C = kL2 + (n − 1 − k ) S 2 .
∂S S−L
=
∂k 2(n − 1 − k )
Taking partials and solving yields
∂L S − L
= .
∂k 2k
Because there are two distinct entries, k cannot equal zero, and so 1 is
the smallest possible value for k. Given P = S n −1− k Lk , logarithmic
dP  L L S 
differentiation leads to the formula = P ln   − + .
dk   S  2S 2 L 
L x 1
The substitution x = into f ( x ) = ln( x ) − + yields the quantity in
S 2 2x
∂P
the bracketed part of . Since f (1) = 0 and f ′ < 1 for x > 1, the
∂k
value within the bracket is always negative, and therefore, P is a strictly
decreasing function of k. Hence, the maximum value for P is attained
only when k = 1. □
n
1
Proposition 2.4 The optimal value of the objective function
n ∏λ
i =2
i is
strictly decreasing as a function of C (with 2e necessarily held fixed).
Proof It suffices to show that the first derivative of P = S n − 2 L is always

∂P S n −3  L S ( n − 2) 
negative. Indeed, =  + =
∂C 2n − 2  S − 2e L − 2e 
S n −3  L( L − 2e) + S ( m − 1)( S − 2e) 

 < 0. □
2n − 2  ( S − 2e)( L − 2e) 
Before applying this to the specific problem, we require a lemma:
Lemma 2.5 (i) The function L ( k , 2e, C ) is increasing as a function of C

with k and 2e fixed, and decreasing as a function of k with 2e and C
fixed.
(ii) The function P = Lk S m − k is decreasing as a function of k with 2e and
C held fixed, and decreasing as a function of C with k and 2e fixed.
Proof (i) We know
 (n − 1)(n − 1 − k )   4e 2 
2e +    C − 
 k  n −1 
L= =
n −1
 C (n − 1)( n − 1 − k )   4e (n − 1 − k ) 
2
2e +    
 k  k 
.
n −1
2
 2e  C ( n − 1 − k ) 4e 2 ( n − 1 − k )
So,  L − n − 1  = (n − 1)k − , and the partial
  k (n − 1) 2
∂L ( n − 1)( n − 1 − k )
derivative with respect to C is = > 0. Similarly,
∂C 2k ( L(n − 1) − 2e)
∂L 4e 2 − C
= <0.
∂k 2   2e  
( )
k ( n − 1)  2  L −
  n − 1  
(ii) The proof of the first part is the same as that for Proposition
∂L ( n − 1) ( n − 1 − k )
(3.3(ii)). For the second part, consider =
∂C 2k ( L(n − 1) − 2e )
and
∂S
=
( (n − 1)k ) . The partial derivative
∂C 2 ( n − 1 − k )( S ( n − 1) − 2e )
∂P  dL  dS  
= Lk −1S n − k − 2  kS + L  (n − 1 − k ) can be shown to be
∂C  dC  dC  
negative, as the quantity in the brackets reduces to
S (n − 1) ( n − 1 − k )( S (n − 1) − 2e ) + Lk (n − 1) ( L(n − 1) − 2e )
, which has
2 ( L( n − 1) − 2e )( S ( n − 1) − 2e )
positive numerator and negative denominator. □
The following lemma is also needed.
Lemma 2.6 Consider the following problem:

k
Minimize ∑x
i =1
2
i subject to
k
1. ∑x
i =1
i =I
2. xi is a nonnegative integer for each i = 1, 2,… , k .

Then there exists integers s≥0 and 1≤ l ≤ k such that
 s + 1, i = 1,..., l
xi =  is a location of the solution to this problem.
 s, i = l + 1,..., k
Furthermore, this location is unique up to permutation of the xi 's.
Proof Suppose x1 , x2 ,..., xk is feasible, and there exists a and b such

k 2
2
that xa − xb ≥ 2 . Then ∑x
i =1
2
i > ∑ x + (x
i ≠ a ,b
2
i a − 1) + ( xb + 1) . Thus, since
there are only finitely many feasible solutions so that the minimum
k
∑x
i =1
2
i exists, any optimal solution can have at most two distinct entries.
Of course, the values of the larger (l ) and smaller (s) subject to I = sk + l

are uniquely determined by the division algorithm, where 0 ≤ l ≤ k − 1 .
If s divides I, then l = 0, which implies that all entries are equal. If s
does not divide I, then sk + l = I , so s ( k − 1) + l ( s + 1) = I . □
2.2 Two Maximum Spanning Tree Results
Recall that trace of a matrix is the sum of its eigenvalues, and the
eigenvalues of the square of a matrix are the squares of the individual
eigenvalues. We use these observations and Lemma 2.6 to prove the next
two spanning tree results:
Theorem 2.7 The graph K n+ x , i.e. Kn with one edge doubled, has the
greatest number of spanning trees among all multigraphs in its class.
Proof First, we note that there are no graphs in this particular class. In
order to prove the theorem, we require the eigenvalues for the Laplacian
matrix of K n+ x .
Claim: The Laplacian matrix of K n+ x has eigenvalues
λ = ( 0, n, n,..., n, n + 2 ) .
Proof of Claim: Without loss of generality assume the edge x is added
between v1 and v2 . The characteristic polynomial of
 n + 1 −1 0 0 ...
 −1 n + 1 0 0 ...

 0 0 n 0 ...
 
(
B K n+ x ) = . . . . .
 . . . . .
 
 . . . . .
 0 0 0 ... n 

is easily seen to be (λ − n) n −1 (λ − ( n + 2)) . Therefore, by Corollary 1.14,

the Laplacian matrix has eigenvalues λ = (0, n,..., n, n + 2) , which proves
the claim.
Thus, 2e = n( n − 2) + n + 2 = n 2 − 2n + n + 2 = n 2 − n + 2 , and
C = trace[( H ( K n+ x )) 2 ] = (n − 2) n 2 + ( n + 2)2 =
n 3 − 2 n 2 + n 2 + 4 n + 4 = n3 − n 2 + 4 n + 4 ,
4e 2 (n 2 − n + 2)2 4
it follows that = = n3 − n 2 + 4 n + <C. Hence,
n −1 n −1 n −1
by Proposition 2.3(ii), it follows that λ = (n + 2, n,..., n) is the location of
the solution to Problem Ps,C , p .
Next, we show that amongst all the problems Ps , C , p where
 nn    nn  
AA
==   + 1,
2e2=e =
 22 
(
+ 1,C C= =trace (
traceH 2 ( M( M ))
,and
) ) , and
and ΩΩn,n,  + 1+ 1, , (i.e.,
MM∈ ∈
  22  
forcing a multigraph situation) the one with the one with the largest
optimal solution has C = trace( H 2 ( K n+ x )) . We do this by applying
Lemma 2.5 after proving that min trace H 2 ( M ) ( ( ) ) occurs only at

n
(
M = K n+ x . Indeed, trace H 2 ( M ) ( ) ) = ∑ h + ∑ (h )
ii ij
2
. It is easy to see
i =1 i≠ j
  n 
that only the multigraph M ∈ Ω  n,   + 1  which has two distinct
  2 
values of hii that differ by one is M = K n+ x .
n
  n 
Now ∑h
i =1
ii = n2 − n + 2 for all M ∈ Ω  n,   + 1 
  2 
and
∑ h = ∑ ( −h ) .
i =1
ii
i≠ j
ij Since K n+ x has two distinct values of hij , i ≠ j ,
which differ by one, K n+ x minimizes ∑ (h ) = ∑ ( − h )

i≠ j
ij
2
i≠ j
ij
2
over all
  n 
M ∈ Ω  n,   + 1  . □
  2 
Next, we consider the class of multigraphs Ω 2q, q 2 . Cheng ( )

showed that K q , q , the complete bipartite graph, has the most spanning
trees over all multigraphs in this class. We include a proof for
completeness:
Theorem 2.8 The graph K q , q is the unique multigraph having the

greatest number of spanning trees among all multigraphs in its class.
[Cheng 1981]
Proof As in Theorem 2.7, it suffices to show that trace(H ( K p , p )) 2 < trace(H ( M ))2
trace(H ( K ))2 < trace(H ( M ))2 for every multigraph M in its class. However, since
trace(B( M )) 2 = trace(H ( M )) 2 + n 2 , it also suffices to prove that
K q ,q uniquely minimizes trace(B 2 ) over all M ∈ Ω 2q, q 2 . ( )

n n
∑ ∑ ∑b
2
Consider trace( B (G ))2 = bii 2 + bij . Since ii = 2q 2 + 2q, and
i =1 i≠ j i =1
K q ,q has all bii = q + 1 , it follows, as in the proof of Theorem 2.7, that

n
K q ,q minimizes ∑b
i =1
ii
2
(
over Ω 2 q, q 2 . )
 2q 
Next realize that ∑b
i≠ j
ij = 2   − 2q 2
 2
for all M ∈ Ω (2q, q 2 ) .
Furthermore, the off-diagonal entries of B ( K q ,q ) are 0's and 1's, so that it
follows, as in the proof of Theorem 2.7, that ∑b

i≠ j
ij is minimized by
K q ,q . Hence, K q , q has a maximum number of spanning trees in
Ω(2q, q 2 ) .
To see that K q , q is the unique such multigraph in Ω (2q, q 2 ) , first we

claim that no multigraph M which is not a graph can have the same
number of spanning trees as K q ,q . Indeed, as there are at least three
distinct off-diagonal entries of B ( M ) , it follows by Lemma 2.6 that

trace(B( M ))2 > trace(B( K q , q ))2 . Hence, it follows, as in the proof of
Theorem 2.7, that t ( M ) < t ( K q ,q ) .
Next, we claim that no other graph G in Ω (2q, q 2 ) can have the same
number of spanning trees as K q , q . If (H 2 (G )) > trace(H 2 ( K q ,q )) , then as
above, t (G ) < t ( K q , q ) . Hence, it remains to consider the case where
( H 2 (G )) = trace( H 2 ( K q ,q )) , i.e., G provides an instance of the problem
Ps,C , p , which has a unique (up to permutation) solution at

( 0, q, q,…, q, 2q ) by Proposition 2.3.
We claim that the eigenvalues of H ( G ) do not equal 0, q, q,… , q, 2q .

On the contrary, suppose that the eigenvalues of G are 0, q, q,… , q, 2q .
Then the eigenvalues of G are 0, 0, q,… , q by Theorem 1.20. Hence, G

has two components (see the note following Theorem 1.5), say G1 , G2 .
Since the maximum eigenvalue of a graph cannot exceed the number of
nodes in the graph, we have q ≤ n(G1 ) and q ≤ n(G2 ) . However,
n(G1 ) + n(G2 ) = 2q , so n(G1 ) = n(G2 ) = q , and the eigenvalues of
G1 and G2 are 0, p, p,… , p . We know from the proof of Theorem 1.27
that only K q has these eigenvalues. Thus, G = K q ∪ K q , and the
contradiction G = K q ,q follows. □
Remark 2.2 Cheng mentioned but did not prove uniqueness of the
eigenvalues for these results.

Chapter 3
Threshold Graphs
3.1 Characteristic Polynomials of Threshold Graphs
Recall that, if the nodes of G are perfectly reliable, edges operate

independently all with the same probability p and ς i equals the
number of spanning connected subgraphs having i edges, then the
All-Terminal Reliability (ATR) of a graph G can be defined by
e
∑ ς p (1 − p )i e −i
R (G ) = i . For small values of p, the reliability
i = n −1
polynomial will be dominated by the ς n−1 term, and since ς n−1 is the
number of spanning trees, graphs with a larger number of spanning
trees will have greater reliability for such p. Computing the ATR of a
probabilistic graph has been demonstrated to be NP-hard [Ball,
1982]. Thus, it is advantageous to bound from below the ATR of a
network represented by a graph in a particular class Ω ( n, e ) (i.e., all
graphs on n nodes and e edges). It has been conjectured that the best
such bound is provided by a threshold graph [Petingi, 1996]. If a
graph minimizes the number of spanning trees, it will minimize
reliability for small values of p. Let N ( u ) denote the deleted
neighborhood of node u, i.e. the set of all nodes adjacent to u. A
graph G is a threshold graph if, for all pairs of nodes u and v in G,
111
N (u ) − {v} ⊆ N (v ) − {u} whenever deg(u ) ≤ deg( v ) . An alternate

equivalent description of a threshold graph is given in our next result.
We present it without proof.
Proposition 3.1 [Hammer, 1978] A graph G on n nodes is a

threshold graph if and only if the node set V of G is the disjoint union
of sets U and W such that U = K c and W = (n − c) K1 , and if
u , v ∈ W such that deg(u ) ≤ deg(v), then N (u ) ⊆ N ( v ) . In other
words, G consists of a clique and an independent set of nodes such
that the neighborhoods of the nodes of the independent set are
“nested in the clique.”
For convenience, the node set of threshold graphs can be thought

of as having two parts: one that induces a clique, and the other
independent nodes which are adjacent to subsets of the clique in such
a way that their neighborhoods are nested within each other (see
Figure 3.1).
Figure 3.1: A threshold graph - K5 is the clique, and the two independent
nodes have degrees 2 and 3, respectively.
It is somewhat easier to visualize a threshold graph if we depict it

differently. We put the nodes corresponding to the clique in a box
and suppress depicting the edges. Thus the only edges shown are
those from the independent set of vertices to nodes in the clique.
Threshold Graphs 113
Using this convention, Figure 3.2 depicts the same threshold graph as
that given in Figure 3.1. We will use this depiction in the sequel.
Figure 3.2: A threshold graph - K5 is the clique (nodes in the rectangular box),
and the two independent nodes have degrees 2 and 3, respectively.
Because of the way the independent nodes and the edges incident on
them are depicted we refer to the independent set of nodes as cone
points.
In the following discussion, we prove the validity of a formula for
the characteristic polynomial of the Laplacian matrix of a threshold
graph in terms of its degree sequence. In preparation for the proof,
we require two lemmas. Recall PH ( M ) ( λ ) is the characteristic
polynomial of H ( M ) .
Lemma 3.2 If G has an isolated node u, then PH ( G ) (λ ) = λ PH ( G − u ) (λ ).
Proof Label the nodes of G v1 , v2 ,… , vn so that u = v1 . Since u is an
isolated node, the first row of the Laplacian H ( G ) is equal to 0T
and H11 ( G ) = H ( G − u ) . Thus the first row of λ I n − H ( G ) is
[λ ,0,…,0] . Evaluating PH (G ) (λ ) = det ( λ I n − H ( G ) ) by expansion
along the first row, we obtain PH (G ) (λ ) = λ det ( λ I n −1 − H11 ( G ) ) = λ PH (G ) ( λ ) = PH (G −u ) (λ )
(λ ) = λ det λ n −1 − 11 G = λ PH11 (G ) (λ ) = PH (G − u ) (λ ) . □
Lemma 3.3 If G has a full degree node u, i.e., deg ( u ) = n − 1 , then

(λ − 1) PH ( G ) (λ ) = λ (λ − n) PH ( G − u ) (λ − 1) .
Proof Realize u is a full degree node in G if and only if it is isolated

in G . Thus, H ̅ = H ̅ . However, G − u = G − u , so
H ̅ = H . Let 0, λ2 ,..., λn −1 be the eigenvalues of

H( − so that 0, n − 1 − λ 2 ,..., n − 1 − λn −1 are the eigenvalues of
n −1
H (G – u ) . Thus PH ( G − u ) (λ ) = λ ∏ i = 2 (λ − (n − 1 − λi )) and so,
n −1
PH ( G − u ) (λ − 1) = (λ − 1) ∏ i =2
(λ − (n − λi )) . Now, the eigenvalues of
H(G) areare 0,0, λ2 ,..., λn −1 so that those of H (G) are

n −1
0, n, n − λ2 ,..., n − λn −1 and PH ( G ) ( λ ) = λ ( λ − n ) ∏ ( λ − ( n − λ ) ).
i =2 i
Thus, −1 H = − PH(G −u ) (λ − 1) , and the proof is

complete. □
We are now ready to state the formula for the characteristic

polynomial PH (G ) (λ ) where G is a threshold graph and prove the
validity of the formula. The formula has appeared in several

publications [Bapat, 2010; Bleiler, 2007; Bogdanowicz, 1985], but
we present what Paul Erdos would have termed “the proof from the
book.” It is based on an outline due to our friend and colleague, the
late Frank T. Boesch.
Theorem 3.4 If ≤ ≤⋯≤ denotes the degree sequence of a

threshold graph on n nodes, and k is the number of cone nodes,
i.e., = − − 1, ≥ 1, ≥ 1, then the characteristic polynomial

of is given by
k k
d1 d j +1 − d j
PH(G ) (λ ) = λ (λ − n) ∏
i =1
(λ − di ) ∏ (λ − n + j )
j =1
.
Proof The proof proceeds by induction on n ≥ 3 . There is only one

threshold graph when n = 3 that satisfies the hypotheses, namely, P3 ,
where = 1, = 1. In this case, the formula yields
PH(G ) (λ ) = λ (λ − 3)1 (λ − 1)1 (λ − 2)1−1 = λ (λ − 3)(λ − 1) ,
which is equal to det(λ I 3 − H( P3 )) .

Next, we make the induction hypothesis that the formula is valid for
threshold graphs having at most n − 1 but at least 3 nodes that satisfy
the hypotheses. Consider a threshold graph on n nodes with degree
sequence satisfying the hypotheses. Since d1 ≥ 1 , there exists a full
degree node u adjacent to the cone v1 . Removal of u produces one of
the following three cases:
(i) if d1 > 1 , we obtain a threshold graph on n − 1 nodes
with k cone nodes having degrees 1 ≤ d1 − 1 ≤ ... ≤ d k − 1 ;
(ii) if di = 1 , for 1 ≤ i ≤ ℓ < k , and d ℓ +1 ≥ 2, we obtain a
threshold graph on n − ℓ − 1 nodes with k − ℓ cone nodes
degrees 1 ≤ dℓ +1 − 1 ≤ ... ≤ d k − 1 and ℓ isolated nodes;
(iii) if d1 = d2 = ... = d k = 1 , we obtain a complete graph
Kn −k −1 and k isolated nodes.
For case (i), d k +1 − 1 = (n − 1) − k − 1 . Applying the induction hypothesis,

we obtain
PH (G −u ) (λ ) =
k k
d1 −1 d j +1 − d j
λ (λ − (n − 1)) ∏ (λ − ( d
i =1
i − 1)) ∏ (λ − (n − 1) + j)
j =1
so that
k k
d1 −1 d j +1 − d j
PH(G −u ) (λ − 1) = (λ − 1)(λ − n) ∏ (λ − d )∏ (λ − n + j)
i =1
i
j =1
.
Thus, by Lemma 3.3,

(λ − 1) PH(G ) (λ ) = λ (λ − n)PH(G −u ) (λ − 1) =
k k
d j +1 − d j
λ (λ − 1)(λ − n)d1 ∏ (λ − di )∏ (λ − n + j ) ,
i =1 j =1
and the conclusion follows after cancellation of λ − 1 .
For case (ii), observe that

(n − ℓ − 1) − (k − ℓ) − 1 = n − k − 2 = d k +1 − 1,
so, by Lemma 3.2 and the induction hypothesis,
PH(G −u ) (λ ) = λ ℓ+1PH (G −u − v1 − v2 −...− vℓ ) (λ ) =
n k
d j +1 − d j
λ ℓ +1
(λ − (n − ℓ − 1)) dℓ +1 −1
∏ ( λ − (d − 1) ) ∏ ( λ − ( n − ℓ − 1) ) + j − ℓ)
i = ℓ +1
i
j = ℓ +1
so that PH(G −u ) (λ − 1) =
n k
d j +1 − d j
(λ − 1)ℓ +1 (λ − (n − ℓ)) dℓ+1 −1 ∏ ( λ − di ) ∏ (λ − (n − ℓ) + j − ℓ) .
i = ℓ +1 j = ℓ +1
Now, applying Lemma 3.3, and cancelling λ − 1 , we obtain

PH ( G ) ( λ ) =
k k
λ (λ − n)1 (λ − 1)ℓ (λ − (n − ℓ))dℓ+1 −1 ∏ ( λ − di ) ∏ (λ − (n − ℓ) + j − ℓ)d
i = ℓ +1 j = ℓ +1
j +1 − d j
.
k k
Observe that (λ − 1)ℓ ∏ ( λ − d i ) = ∏ (λ − d i ). Also, d j +1 − d j = 0 for
i = ℓ +1 i =1
j = 1,..., ℓ − 1 and dℓ +1 − dℓ = dℓ +1 − 1, so
k k
d j +1 − d j d j +1 − d j
(λ − (n − ℓ)) dℓ+1 −1 ∏ ( λ − ( n − ℓ) + j − ℓ) = ∏ (λ − n + j ) .
j = ℓ +1 j =1
k k
d d j +1 − d j
Hence, PH (G ) (λ ) = λ (λ − n) 1 ∏i =1
(λ − di ) ∏ (λ − n + j)
j =1
, the required
conclusion.
Finally, in case (iii),

n −k − 2
PH(G −u ) (λ ) = λ k PH (K n −( k +1) ) (λ ) = λ k λ ( λ − (n − k − 1) ) , so
k n −k −2
PH ( G − u ) (λ ) = λ k PH (K n−( k +1) ) (λ ) = λ λ ( λ − (n − k − 1) ) .
Again, applying Lemma 3.3, and cancelling λ − 1 , we obtain
PH (G ) ( λ ) = λ (λ − n)1 (λ − 1)k (λ − (n − k ))n − k − 2 .

k
Now, (λ − 1) k = ∏ ( λ − d i ) and
i =1
k
d j +1 − d j
∏ (λ − n + j ) = (λ − n + k )dk +1 −1 = (λ − n + k )n − k −1−1 ,
j =1
thus the result follows. □
Remark 3.1 The material concerning the above formula has been
submitted for publication.
Example 3.1 Consider the threshold graph depicted in Figure 3.2.

For this graph, n =n7,= 7,
k =k2,= 2, and the degree sequence is
d = 2,3, 4, 4,5,6,6 . Applying Theorem 3.4, we obtain
PH ( G ) (λ ) = λ (λ − 7)2 (λ − 2)(λ − 3)(λ − 7 + 1)3− 2 (λ − 7 + 2) 4 −3 =
λ (λ − 7)2 (λ − 2)(λ − 3)(λ − 6)(λ − 5) ,

so the eigenvalues are 0, 2, 3, 5, 6, 7, 7, and the number of spanning trees
2 ⋅3⋅ 5⋅ 6 ⋅ 7 ⋅ 7
is = 1260 by Theorem 1.7.
7
Another way to generate the eigenvalues for the Laplacian matrix of a

threshold graph can be found in [Bapat, 2010] and is based on the work
of Russell Merris [Merris, 1994]. Given a degree sequence of a threshold
graph, d = d1 , d 2 ,⋯ , d n −1 , d n , with d1 ≥ d 2 ≥ ⋯ ≥ d n −1 ≥ d n , we form a
new sequence d ∗ = d1∗ , d 2∗ ,⋯, d n −1∗ , d n ∗ , where di ∗ is the number of

terms in the degree sequence d that are greater than or equal to i. It is
the case that d∗ is in fact the sequence of the eigenvalues
λ1 ≥ λ2 ≥ ⋯ ≥ λn −1 ≥ λn = 0.
Example 3.2 For the threshold graph in Figure 3.2, we reorder the
degree sequence d = 6,6,5,4,4,3,2 , i.e., d1 = 6, d 2 = 6, d 3 = 5, d 4 = 4,
d5 = 4, d6 = 3, and d7 = 2. Thus λ1 = d1* = 7, λ2 = d 2* = 7, λ3 = d 3* = 6,
λ4 = d 4* = 5, λ5 = d 5* = 3, λ6 = d6* = 2, and λ7 = d 7* = 0 .
In their 1996 paper, Laplacian spectra and spanning trees of threshold

graphs [Hammer, 1996], Hammer and Kelmans present the following
formula for the eigenvalues of the Laplacian matrix of a threshold graph:
(n ) (n ) (n )
(*) Let G be a connected threshold graph, and let d = (v1 1 , v2 2 ,..., vs s )
be the degree sequence of G, where vi ≤ vi +1 for i = 1, …, s − 1, and ni is

the multiplicity of vi in the degree sequence. Then
kˆ s
ni nkˆ +1 −1
PH ( G ) (λ ) = ∏ (λ − v ) ∏ (λ − v
i i − 1) ni (λ − vkˆ +1 − 1) and
i =1 i = kˆ + 2
kˆ s −1
nkˆ +1 −1
t (G ) = ∏ (vi ) ni ∏ (v i + 1) ni ( vkˆ +1 + 1) (vs + 1) ns −1 ,
i =1 i = kˆ + 2
 s − 1
where kˆ =   [Hammer, 1996].
 2 
This formula does not work for every threshold graph, as

demonstrated in [Bleiler, 2007]. Consider a threshold graph having a
clique of order 5 and three cones of degree 4, 3, and 2, respectively,
as illustrated in Figure 3.3.
Figure 3.3: A threshold graph that serves as a counterexample to the

Hammer/Kelmans spanning tree formula for threshold graphs.
The degree sequence is

(1) (1) (2) (1) (1) (2)
d = 2, 3, 4, 4, 5, 6, 7, 7 = (2 ,3 , 4 ,5 ,6 ,7 ) ,
so the number of distinct terms in the degree sequence is s = 6.
 s − 1  6 − 1   5 
According to the formula in (*), kˆ =  = = = 2,
 2   2   2 
and
kˆ s
nkˆ+1 −1
PH (G ) (λ ) = ∏ (λ − vi ) ni ∏ (λ − v i − 1) ni (λ − vkˆ +1 − 1) =
i =1 i = kˆ + 2
2 6
ni ni
∏ (λ − v ) ∏ (λ − v − 1)
i i (λ − v2+1 − 1) n2+1 −1 =
i =1 i =4
{(λ − v1 )n1 (λ − v2 )n2 }{(λ − v4 −1)n4 (λ − v5 −1)n5 (λ − v6 −1)n6 (λ − v3 −1)n3 −1} =

{(λ − 2)1 (λ − 3)1}{(λ − 5 − 1)1 (λ − 6 − 1)1 (λ − 7 − 1)2 (λ − 4 − 1)1} =
(λ − 2)1 (λ − 3)1 (λ − 6)1 (λ − 7)1 (λ − 8)2 (λ − 5)1 .
The eigenvalues for the graph’s Laplacian matrix are, by this formula,
2, 3, 5, 6, 7, 8, 8 (plus the zero eigenvalue, which does not factor into
the spanning tree computations for connected graphs). Application of
Theorem 1.5 indicates that this graph would have 10,080 spanning trees.
In fact, by Theorem 0.11, the sum of the eigenvalues and the sum of
the degree sequence terms should both be twice the number of edges.
This is a 19 edge graph, so the sum of the eigenvalues given by (*)
should be 38, but is actually 39, which is a contradiction. Indeed, direct
calculation using the graph’s specific Laplacian matrix reveals that the
eigenvalues are actually 0, 2, 3, 4, 6, 7, 8, 8 and the number of spanning
trees is 8064. The formula of Hammer and Kelmans can be restated as
follows:
Theorem 3.5 [Bleiler, 2007] Let G be a connected threshold graph,

(n )
and let d = (v1( n1 ) , v2( n2 ) ,..., vs s ) be the degree sequence of G, where
vi ≤ vi +1 for i = 1, …, s − 1, and ni is the multiplicity of vi in the degree
 s − 1
sequence, with kˆ =  .
 2 
(a) For s odd, PH (G ) (λ ) =

kˆ s
nkˆ +1 −1
λ ∏ (λ − vi ) n i
∏ (λ − v i − 1)ni (λ − vkˆ +1 − 1) and
i =1 i = kˆ + 2
kˆ s −1
nkˆ +1 −1
t (G ) = ∏ (vi ) ni ∏ (v i + 1) ni (vkˆ +1 + 1) (vs + 1) ns −1 .
i =1 i = kˆ + 2
(b) For s even, PH (G ) (λ ) =

kˆ s
nkˆ+1 −1
λ ∏ (λ − vi ) n i
∏ (λ − v − 1) i
ni
(λ − vkˆ +1 )
i =1 i = kˆ + 2
k s −1
nkˆ+1 −1
t (G ) = ∏ (vi ) ni ∏ (v i + 1) ni (vkˆ +1 ) (vs + 1) ns −1.
i =1 i = kˆ + 2
Remark 3.2 The differences between the two formulas in Theorem

3.5 are based on the fact that, in threshold graphs having an odd
number of distinct degree sequence terms, there is a clear distinction
between the nodes in the clique and the nodes in the independent set.
On the other hand, in threshold graphs having an even number of
distinct degree sequence terms, the maximum degree of the cone
nodes and the minimum degree of the clique nodes is the same.
Furthermore, there can be more than one cone node having this
common degree, but only one clique node.
In his 1985 doctoral thesis, “Spanning Trees in Undirected

Graphs,” Zbignew Rysard Bogdanowicz derived many interesting
spanning tree formulas. He denotes a threshold graph by Q (n, k) where
K n − k is the clique and the k cones have degrees q1 ≥ q2 ≥ ... ≥ qk , and
states the following theorem:
Theorem 3.6 Let q1 , q2 ,..., qk ( qi ≥ qi +1 ) denote the degrees of the

cone nodes of a threshold graph on n nodes, Q(n,k). Then the number
of spanning trees of a Q (n, k) graph is
k −1
t (Q ( n , k )) = qk n qk −1 ( n − k ) n − q1 − k −1 ∏ q (n − k + i)
i
qi − qi +1
i =1
[Bogdanowicz, 1985].
We will apply the various formulas to compute the number of

spanning trees for the threshold graph in Figure 3.3:
Example 3.3 We will apply the theorems of this section to determine

the Laplacian eigenvalues and/or the number of spanning trees of
the threshold graph in Figure 3.3. This graph has parameter values
5
n = 8, k = 3, s = 6 , kˆ =   = 2, and degree sequence
2
d = 2,3, 4, 4,5,6,7,7 = (2(1) ,3(1) , 4(2) ,5(1) ,6(1) , 7 (2) ).
(a) By Theorem 3.4,
k k
d j +1 − d j
PH (G ) (λ ) = λ (λ − n)d1 ∏ (λ − d i ) ∏ (λ − n + j ) =
i =1 j =1
3 k
2 d j +1 − d j
λ (λ − 8) ∏ (λ − d )∏ (λ − 8 + j )
i =1
i
j =1
=
λ (λ − 8) 2 (λ − 2)(λ − 3)(λ − 4)(λ − 7)(λ − 6)(λ − 5)0 =

λ (λ − 8) 2 (λ − 2)(λ − 3)(λ − 4)(λ − 7)(λ − 6).
The number of spanning trees would be
82 ⋅ 7 ⋅ 6 ⋅ 4 ⋅ 3 ⋅ 2
= 8064 .
8
(b) By Theorem 3.5(b),

kˆ s
nkˆ +1 −1
λ ∏ (λ − vi )n i
∏ (λ − v i − 1)ni (λ − vkˆ +1 ) =
i =1 i = kˆ + 2
2 6
λ ∏ (λ − vi )n ∏ (λ − vi − 1)n (λ − v3 )n −1 =
i i 3
i =1 i =4
2
λ ( λ − 2 )( λ − 3)( λ − 6 )( λ − 7 )( λ − 8 ) ( λ − 4 ) .
Once again, the number of spanning trees would be 8064.

(c) Using the notation of Theorem 3.6, the cone point degrees
are q1 = 4, q2 = 3, q3 = 2 . Now, t (Q( n, k )) =
k −1
qk n qk −1 (n − k ) n − q1 − k −1 ∏ q (n − k + i)
i
qi − qi +1
i =1
k −1
= q3 8q3 −1 (8 − 3) 8 − q1 − 3 −1
∏ q (8 − 3 + i)
i
qi − qi +1
i =1
= 2 ⋅ 8 ⋅ 4 ⋅ 6 ⋅ 3 ⋅ 7 = 8064.
3.2 Minimum Number of Spanning Trees
The “lollipop” graph, illustrated in Figure 3.4 has the same ATR and
number of spanning trees as the threshold graph depicted in Figure 3.5.
For our purposes here the placement of the bridges (a bridge is an edge
whose removal disconnects the graph) as either a path joined to the
remainder of the graph or single edges joined to the graph does not
matter as all those edges must appear in every connected spanning
subgraph, in particular every spanning tree.
This graph has the least number of spanning trees of any

connected graph in its class [Bogdanowicz, 2009], and has been
conjectured to form a lower bound on the ATR for all connected graphs
in its class.
⋯ ⋯
Figure 3.4: Lollipop - Kn−k with 1 cone of degree c and a Pk −1 appended to the
cone.
⋯ ⋯
Figure 3.5: Threshold graph having Kn−k with 1 cone of degree c and
k − 1 cones of degree 1.
Direct application of Theorem 3.4 and Theorem 1.5 yields the

number of spanning trees for a lollipop.
Proposition 3.7 Let G be a threshold graph with clique K n − k , one cone

of degree c and k − 1 cones of degree 1. Then
c −1
t ( G ) = c ( n − k + 1) ( n − k )n−k −1−c .
Proof By Theorem 3.4 PH (G ) (λ ) =

k −1 c 1 n − k −1− c
λ ( λ − n )( λ − 1) ( λ − c )( λ − n + (k − 1) ) − ( λ − n + k ) .
The formula follows by applying Theorem 1.5. □
We observe that the formula in Proposition 3.7 is applicable to a

threshold graph in which k = 1 . Such a graph is a special subcase of
the lollipop, called the balloon, depicted in Figure 3.6. When such a
graph has cone degree strictly greater than 1, the balloon graph
resides in a class Ω ( n, e ) in which connected graphs do not have any
edges which are bridges. In these graphs,
( n − 1)( n − 2 )  n
+ 1 < e ≤   (**).
2  2
⋯ ⋯
Figure 3.6: Balloon graph having Kn −1 with 1 cone of degree c.
As stated earlier, it has been proven [Bogdanowicz, 2009] that the

lollipop minimizes the number of spanning trees among all graphs in its
class. In addition, it has been conjectured to have uniformly minimum
ATR among all graphs in its Ω ( n, e ) class. In classes of graphs whose
number of edges satisfy (**), the balloon graph can be shown to
minimize the number of spanning trees.
This result can be proven using a graph transformation called the
swing surgery. This graph transformation is also used in the
approaches to multigraphs in Chapter 4. Results regarding this
surgery were established for its effect on spanning trees in the graph
case by Satyanarayana, Schoppmann and Suffel [Satyanarayana,
1992] and Bogdanowitz [Bogdanowitz 1984]; its effect on a graph’s

reliability was investigated by Brown, Colbourn and Devitt [Brown,
1993] and Kelmans [Kelmans 1981]. Recall that if w and y are
nodes, then m ({ x, y} ) is the multiplicity of edge { x, y} .
Theorem 3.8 (Swing Surgery) Let M be a multigraph with node set

V and w, v ∈ V have the property that, for each y ∈ V − { w, v} ,
m ({w, y} ) ≥ m ({ y, v} ) .
(1) If there exists y such that m ({w, y} ) > m ({ y, v} ) , then for
any such y, the multigraph M ′, obtained from M by deleting

one of the edges incident on w and y and inserting an
additional edge incident on y and v, has at least as many
spanning trees as G, i.e. t ( M ′ ) ≥ t ( M ) . (Figure 3.7a)
(2) Furthermore, if in addition to m ({w, y} ) > m ({ y, v} ) , there
exists an edge incident on y and a node u1 ∈ N ( w ) ∩ N ( v ) ,
and also an edge incident on y and a node u2 ∈ N ( w ) − N ( v ) ,
then t ( M ′ ) > t ( M ) . [Satyanarayana, 1992] (Figure 3.7b)
w y w y
v v
Graph G: t(G) = 34 Graph G′: t( G ′ ) = 45
Figure 3.7a: Example of the swing surgery. One of the multiedges wy

is “swung” down to become edge yv.
w y w y
u1 v u1 v
u2 u2
Graph G: t(G) = 83 Graph G: t( G  ) = 96
Figure 3.7b: Example of the swing surgery. One of the multiedges wy

is “swung” down to become edge wv.
We delay the proof, essentially the same as given in

Schoppmann’s thesis [Schoppmann, 1990], until after the
presentation of three preliminary results. First we require a version
of the well-known Factoring Theorem [Moskowitz, 1958] as it
applies to spanning trees, as well as two lemmas:
Theorem 3.9 (Factoring Theorem) If M x denotes the

multigraph obtained by the deletion of edge x from multigraph M,
and M | x denotes the multigraph obtained by the contraction of edge
x, then t(M) = t(M  x) + t( M | x ).
The proof of this theorem is straight-forward, as the spanning

trees of M  x are in one-to-one correspondence with those of M
that do not include x, while those of M | x are in one-to-one
correspondence with those of M that do include x.
Our next result is a lemma required to prove the theorem.

Lemma 3.10 Let M and M ′ be multigraphs each having node set

V = {w, v, y1 , y2 ,…, ys } with { y1 , y2 ,…, ys } an independent set of vertices.
Assume further that M and M ′ have the same number of edges. We
denote by li and li ′ the number of edges incident on v and yi in M and
M ′, respectively. Likewise, ui and ui ′ denote the number of edges
incident on w and yi in M and M ′, respectively. Finally, we assume

that there are the same number of edges between w and v in both M and
M ′ and let the common value of mM ({w, v}) = mM ′ ({w, v}) be denoted
by h. If u j − l j ≥ u j′ − l j′ and u j + l j = u j′ + l j′ for each j = 1,2,…,s,
then t ( M ′ ) ≥ t ( M ) . Furthermore, if in addition, M ′ is connected and
there exists an integer k such that uk − lk > uk ′ − lk ′ , then t ( M ′ ) > t ( M ) .
Proof (of Lemma 3.10) The proof proceeds by nested induction, first on
the number of nodes in M − w − v, i.e., the yi ' s , and then on the number
of edges incident on y1 . (See Figure 3.8) If s = 1, i.e., there is exactly

one yi, then t(M) = u1l1 + h(u1 + l1 ) and t ( M ′ ) = u ′ l ′ + h(u ′ + l1′ ) . It is a
simple exercise to prove that u1′l1′ ≥ u1l1 when u1′ + l1′ = u1 + l1 and
u1′ − l1′ ≥ u1 − l1 and that the inequality is strict provided
u1′ − l1′ > u1 − l1 .
Assume the lemma is true when there are yi ' s . To prove the lemma for
s + 1 yi ' s , we perform induction on the total number of edges incident

on y1 .
w
u1 u2 u3 u s h
y1 y2 y3 ….. y s
l1 l2 l3 ls
Figure 3.8: The multigraph M at the start of Lemma 3.10.

Note that nodes w and v are labeled, and the edge labels
indicate the multiplicity of each multiple edge.
Base case: Since the case u1 + l1 = 0 is trivial, i.e., t ( M ) = 0 = t ( M ′ ) ,

assume the number of edges incident on y1 is 1, u1 = 1 , and l1 = 0 .
Without loss of generality, we can assume u1′ = 1 and l1′ = 0 . Then
u1 + l1 = u1′ + l1′ and u1′ − l1′ ≥ u1 − l1 . Let x1 ( x1′ ) be the unique
edge incident on w and y1 in M ( M ′ ) . Application of the factoring
theorem to x and x′ yields zero spanning trees when the edge is

deleted and graphs of the type described in the statement of this
lemma with s y i ' s when the edge is contracted. By the induction
hypothesis, t ( M ′ x′ ) ≥ t ( M x ) . Furthermore, if M ′ is connected
and there exists an index k such that | uk − lk |>| uk ′ − lk ′ | , then
necessarily, u1 + l1 ≥ 0 , k ≥ 2, and after the contraction of x1′ , G ′ | x1′

remains connected, and the inequality still holds, so by the induction
hypothesis t ( M ′ x′ ) > t ( M x ) .
Now assume the lemma holds for s  1 yi ' s with u1  l1  p  1 . We

establish the results for u1  l1  p  1  2 by considering two cases.
Case 1. Assume that | u1  l1 |  | u1  l1 | . Without loss of generality,
assume u1  u1 , l1  l1 and u1  l1 . Let x1 be one of the edges incident on
w and y1 in M and x1 be one of the edges incident on w and y1 in M  .
We now apply the factoring theorem to x1 and x1 . Their deletion makes
the sum of the edges incident on y1 equal to p in both graphs. Hence, by
the induction hypothesis, t ( M  x1 )  t ( M   x1 ) . Contraction of x1 and
x1 yields the multigraphs of the lemma, one with l1  h edges joining w
and v and s yi ' s , namely, y2 , y3 ,..., ys 1 in M | x1 and the second with the
analogous description for M  | x1 . Hence, t ( M | x1 )  t ( M  | x1 ) , with
strict inequality if u k  lk  u k   lk  for some k  2 by the primary
induction hypothesis. An application of the factoring theorem completes

the proof in this case.
Case 2. Assume that u1  l1  u1  l1 , and, without loss of generality,
u1  l1 , u1  l1 , u1  u1 , l1  l1 . Consider edges x1 (and x1 ) in M (and
M  ) incident on w and y1 ( w and y1 ). By the secondary induction
hypothesis, t ( M  x1 )  t ( M   x1 ) . On the other hand, contraction
of these edges yields multigraphs M | x1 and M  | x1 where l1  l1

(see Figures 4.3 and 4.4, respectively). Repeated factoring on the
edges incident on w and v in the two multigraphs yields identical
contractions at each application, so it remains to compare the deleted

multigraphs obtained after l1 factorings, say M̂ and M̂ c , where M̂ has
h edges incident on w and v while M̂ c has h l1c l1 edges incident on

w and v. Now the primary induction hypothesis may be invoked to
conclude that the number of spanning trees of M̂ c with l1c l1 edges from

wc and vc removed is at least as great as t Mˆ . Thus, t (Mˆ c) t t (Mˆ ). If
M̂ is connected, then it is readily seen that M̂ c is connected as well, and
since l1c l1 > 0, t M c ! t M . This completes the proof. Ƒ
w
u2 u3 us
y2 y3 ….. ys
h+ l1 l2 l3 ls
v
Figure 3.9: The multigraph M | x1 . Edge labels indicate the multiplicity
of each multiple edge.
wc
u2c u3c u sc
y2c y3c ….. ysc

h+ l1c l2c l3c … ls c
vc
Figure 3.10: The multigraph M c | x1c . Edge labels indicate the multiplicity
of each multiple edge.
Let x1 , x2 ,..., xk be a fixed but arbitrary collection of k edges in a

multigraph G. Consider edge x1 , which can be either deleted or
contracted. If x2 has become a loop, then it is deleted. (This will only
happen if x1 and x2 are edges incident on the same pair of nodes.) If it
has not become a loop, then contract or delete it. Repeat the process
for x3 , and so on. We can represent any such sequence by a k-tuple a
of 0s, 1s and dashes where
1 if xi is contracted

ai   0 if xi is deleted
  if x has become a loop .
 i
A k-tuple of 0's, 1's and dashes is called realizable if each of its

respective operations can be performed on the specified edges of a
graph. We present the following lemma:
Lemma 3.11 Let x1 , x2 ,..., xk be a fixed but arbitrary collection of k

edges in a multigraph M. If A consists of all such realizable k-
tuples corresponding to the edges as described above, then for each
a  A, let M a represent the graph obtained from M by sequentially
performing the operations. Then t ( M )   t ( M a ).

a A
Proof (by induction on k). Let k =1; then
t ( M )  t ( M | x1 )  t ( M  x1 ).
Assume the expansion is valid for k  m. Now, let k = m + 1. By the

induction hypothesis, for the first m edges, t ( M )   t (M
a A
a ), where
A′ consists of all m-tuples corresponding to the sequential

contractions and deletions of the first m edges x1 ,..., xm . If for
a ′ ∈ A′, xm +1 is a loop in M a′ , then extend a′ by appending a dash
onto a ′ and obtain (a ′, − ) ∈ A. Clearly, t ( M a′ ) = t (M (a′,− ) ). If xm +1 is
not a loop, then t ( M a′ ) = t (M (a′,0) ) + t (M (a′,1) ) by the factoring
theorem. Therefore, t ( M ) = ∑ t ( M a ) follows. □

a∈ A
We will now prove Theorem 3.8:
Proof of Theorem 3.8 Let H = M − w − v = M ′ − w′ − v′ and have edges
x1 ,..., xk . We know t ( M ) = ∑ t (M
a∈ A
a ), t (M
′) = ∑ t (M ′ ) ,
a∈ A
a where A
consists of the realizable k-tuples corresponding to the edges of H. Note
that the hypotheses of Lemma 3.10 are satisfied so that t ( M a ) ≤ t ( M a′ ).

Now suppose the conditions of Theorem 3.8 (2) hold. Arrange the edges
of M − v − w in an arbitrary order x1 , x2 ,..., xk . Then consider the k-tuple
a obtained by first contracting x1 and then, for each i in turn, either
contracting xi if it is not a loop or deleting it if it is a loop. The resulting
M a and M a′ are each connected, and they satisfy the condition in
Lemma 3.10 that forces t (M a ) < t (M a′ ) . This completes the proof of

Theorem 3.8 (2). □
Remark 3.3 When w and v satisfy the neighborhood conditions of

Theorem 3.8(1), we say N(w) dominates N (v).
An important consequence of the Swing Surgery is that, given any

connected graph G in Ω ( n, e ) , there exists a threshold graph T also in
Ω ( n, e ) such that t (T ) ≤ t (G ) . [Schoppmann, 1990] This fact, coupled

with judicious application of the Swing Surgery to a threshold graph,
gives the following result:
Theorem 3.12 Consider a class of graphs Ω ( n, e ) , such that
( n − 1)( n − 2 )
e≥ + 1, that admits a unique balloon graph B and another
2
threshold graph T having at least two cone nodes. Denote by R(G) the all
terminal reliability of any graph. Then R ( B ) ≤ R (T ) for all values of p,
i.e., the balloon is a uniform lower bound on ATR for all threshold
graphs in its class. [Petingi, 1996]
The proof of this result is performed by induction on the number

of edges in the complement of the graph. A modified version of the
threshold graph is transformed into a similarly modified version of
the balloon graph in its class by repeated, specific application of
Theorem 3.8. Consequently, since the balloon provides a uniform
lower bound on ATR, it also provides a lower bound on the number
of spanning trees.
3.3 Spanning Trees of Split Graphs
In section 1 of this chapter, we characterized threshold graphs as

having nodes that are one of two types: nodes in a clique, and an
independent set of nodes (i.e., cones) adjacent to subsets of the clique

in such a way that their neighborhoods are nested within each other.
If the neighborhood nesting condition is not met for at least one pair
of nodes, but the graph still maintains the partition between clique
and independent set, the graph is known as a split graph. When
viewed in this scenario, a threshold graph is a special instance of the
split graph, i.e., a split graph with the nesting property for every pair
of nodes.
A non-increasing sequence of nonnegative integers is said to be
graphical if there exists a graph that has the sequence as its degree
sequence. Erdos and Gallai [Erdos, 1960] proved the following theorem:
Theorem 3.13 A sequence of nonnegative integers d n , d n −1 ,..., d1

arranged in non-increasing order is graphical if and only if for each
integer r, 1 ≤ r ≤ n − 1, ∑ ! ≤" "−1 +∑! min '", (.
Example 3.4 Consider the sequence of integers 6, 5, 5, 5, 4, 4, 3. Now,

1
∑d
i =1
i = 6 ≤ 6 = 1(1 − 1) + (1 + 1 + 1 + 1 + 1 + 1) ;
2
∑d
i =1
i = 11 ≤ 12 = 2 ( 2 − 1) + ( 2 + 2 + 2 + 2 + 2 ) ;
3
∑d
i =1
i = 16 ≤ 18 = 3 ( 3 − 1) + ( 3 + 3 + 3 + 3 ) ;
∑d
i =1
i = 21 ≤ 23 = 4 ( 4 − 1) + ( 4 + 4 + 3) ;
5
∑d
i =1
i = 25 ≤ 27 = 5 ( 5 − 1) + ( 4 + 3) ;
6
∑d
i =1
i = 29 ≤ 33 = 6 ( 6 − 1) + ( 3) ;
It satisfies all the inequalities and thus is graphical. However, if we

consider 7, 6, 5, 5, 4, 4, 3,
1
∑d
i =1
i = 7 ≤ 6 = 1(1 − 1) + (1 + 1 + 1 + 1 + 1 + 1) ;
this sequence fails the first inequality and thus is not graphical.
Split graphs are characterized as satisfying the mth Erdos and

Gallai inequality by equality [Hammer, 1981], where m represents
the largest index i in a degree sequence such that d i ≥ i − 1. If this
is the case, then the m nodes of largest degree form a clique in G,
and the remaining nodes constitute the independent set.
We will refer to a split graph as ideal if every node in the clique that
is adjacent to cones is adjacent to the same number of cones, and proper
if all cones have the same degree. The notation for a split graph which is
both ideal and proper will be x-IPS(c,d,b), where x is the number of cone
nodes to which each clique node is adjacent, c is the number of cone
nodes, d is the common degree of each cone, and b is the number of
clique nodes not adjacent to any cone. These parameters uniquely
determine each of these split graphs; -IPS(c,d,b) graphs exist for
cd
positive integers c, d, x and nonnegative integer b such that ∈ ℕ.
x
Figure 3.11: A split graph. We denote this graph 1-IPS(2,3,0), whose clique is of
order 6, and observe that both sides of the 6th Erdos and Gallai inequality are
equal to 36.
Figure 3.12: The split graph 1-IPS(2,3,1).
Figure 3.13: The split graph 2-IPS(4,3,0).
Theorem 1.5 can be applied to a variety split graphs in order to

calculate their number of spanning trees. The following theorem
summarizes those results for the number of spanning trees of some
ideal proper non-threshold split graphs.
Theorem 3.14 The number of spanning trees for some x-IPS(c,d,b)

graphs are as follows:
(1) [Michewicz, 2006] If x = 1 and b = 0, then t(1-IPS(c,d,0)) =
1 n − 2c
( d + 1)( n − c + 1) X c-1Y c −1
n
 [d (c + 1) + 1] + (c − 1)2 d 2 + 2(c + 1) d + 1] 
where X = 
where ,
 2 
 
 [d (c + 1) + 1] − (c − 1) d + 2(c + 1)d + 1] 
2 2
and Y = 
and .
 2 
 
(Figure 3.11)
(2) [Moore, 2013] If x = 1 and b ≠ 0, then t(1-IPS(c,d,b)) =
1
( cd + b ) ( cd + b + 1) ( ) W1 ⋅ W2 ⋅ Z1 ⋅ Z 2 , where W1 =
b −1 c d −1
n
1
2 ( 2 2
( cd + b ) + d + 1 + ( cd + b ) + 2 ( cd + b )( d + 1) + ( d + 1) − 4dn , )
1
( 2 2
W2 = ( cd + b ) + d + 1 − ( cd + b ) + 2 ( cd + b )( d + 1) + ( d + 1) − 4dn ,
2 )
1
( 2
Z1 = ( cd + b ) + d + 1 + ( cd + b ) − 2 ( cd + b )( d − 1) + ( d + 1) ,
2
2
)
and Z 2 =
1
2((cd + b) + d + 1 − (cd + b) − 2 (cd + b)( d − 1) + (d + 1) ).
2 2
(Figure 3.12)
(3) [Fuller, 2014] If x >1 and b = 0, then t(x-IPS(c,d,0) =
c
( d −1)  1
1  cd  x c  1− 
 x
c
−1
c
−1
( x + d ) x + d ( P + Q) x ( P − Q) x ,
cd  x 
c+
x
2
1 cd  1  cd  cd 2
where P =  x + d +  , and Q =  x+d +  −4 .
2 x  2  x  x
(Figure 3.13)
Proof Parts (1) and (2) of Theorem 3.14 were proven in the referenced
publications, while part (3) has been submitted for publication. All of the
proofs follow from elementary methods involving matrix row operations.
To give a flavor for these proofs, we provide a sketch of the proof of (2).
Consider the Temperley’s B-Matrix, B, as defined in Chapter 1. Based
on the structure of the x-IPS(c, d, b) graphs when b > 0, we choose
to arrange the matrix such that the nodes representing each row
(column) are in decreasing order of degree. Hence, B − λI is of the
Q Q4 
form  1 . Submatrix Q 1 is a diagonal ( cd + b ) × ( cd + b )
Q 2 Q 3 
matrix having cd + b + 1 − λ as its first cd entries and cd + b − λ as

its remaining b entries, i.e.,
cd + b + 1 − λ 0 0 ⋯ 0 0 
0 ⋱ 0 ⋯ 0 0 
 
0 0 cd + b + 1 − λ 0 ⋮ ⋮ 
Q1 =  .
⋮ ⋮ 0 cd + b − λ 0 ⋮ 
0 0 ⋯ 0 ⋱ 0 
 
0 ⋯ ⋯ ⋯ 0 cd + b − λ 
Submatrix Q2 is c × ( cd + b ) . Since each cone node is adjacent to d
clique nodes, we have the lefthand part of Q 2 is a cascading vector of d

zeros with the remainder of the entries as ones, and the righthand part
of Q 2 is a c × b matrix of ones, i.e.,
0 0 ⋯0 1 1 ⋯1 ⋯ 1 1 ⋯1 1 ⋯ 1
1 0 0 ⋯ 0 1 1 ⋯1 1 1 ⋯1 ⋮ 1 ⋮ 
Q2 =  .
⋮ 1 ⋱ 1 1 ⋯1 1 ⋮ 1
 
1 ⋯ 1 0 0 ⋯ 0 1 1 1
The bottom right quadrant, which we have labeled submatrix Q 3 , is a

c × c matrix having d + 1 − λ as each entry on its main diagonal and all
ones off its main diagonal, i.e.,
d + 1 − λ 1 ⋯ 1
 1 d +1− λ 1 ⋮ 
Q3 =  .
 ⋮ 1 ⋱ 1 
 
 1 1 1 d +1− λ
Finally, Q4 = Q2T .
We create an upper triangular matrix using row operations in an

algorithmic fashion. First, we use a two-stage process to transform
Q 2 into a submatrix consisting of all zeros. We zero out the left-hand
c × (cd) entries, corresponding to the clique nodes which are adjacent to
cones, by multiplying the rows of Q1 that have diagonal entry
1
cd + b + 1 − λ by − and adding to the appropriate rows of
cd + b + 1 − λ
Q 2 . Then we proceed to the rightmost c × b part of Q 2 , multiplying the
1
rows of Q 1 that have diagonal entry cd + b − λ by − and
cd + b − λ
adding to the appropriate rows of Q 2 . After this is completed, the matrix
has Q1 and Q 4 unchanged, the matrix Q 2 is all zeros, and Q 3 has all
main diagonal entries the same, say, A, and all off- diagonal the terms the
same, say, B, where
d ( c − 1) b
A = (d +1− λ ) − − , and
( cd + b + 1) − λ ( cd + b ) − λ
d (c − 2) b
B =1− − .
( + + ) − ( + b) − λ
cd b 1 λ cd
A B B ⋯ B
B A B ⋯ B 

Since Q 3 =  B B ⋱ B B  , the algorithmic application of row
 
⋮ ⋮ B A B
 B B ⋯ B A 
operations transforms Q 3 into an upper triangular submatrix, which

means that the whole matrix is now upper triangular. Multiplying

down the main diagonal, we obtain the product
cd b  ( A + B )( A − B )   ( A + 2 B )( A − B ) 
( cd + b + 1 − λ ) ( cd + b − λ ) ( A )    
 A  ( A + B) 
 ( A + 3B )( A − B )   A + ( c − 2 ) B ( A − B )   A + ( c − 1) B ( A − B ) 
( ) ( )
 ⋯   .
 ( A + 2B )   (
A + ( c − 3) B 
 )
A + (c − 2) B( 
 )
After cancellation, we have
c −1
( cd + b + 1 − λ ) ( ) ( cd + b − λ ) ( λ 2 − ( cd + b + 1) λ + d ( cd + b ) )
c d −1 b −1
( λ 2 − ( cd + b + d + 1) λ + d ( cd + b + c ) ) ( cd + b + c − λ ) .
The roots of this polynomial are multiplied together, after dividing out
the n 2 (as in Theorem 1.13), t(1-IPS(c,d,b))=
1 b −1 c ( d −1)
× ( cd + b ) × ( cd + b + 1) × W1 × W2 × Z1 × Z 2 ,
n
1
( 2 2
where W1 = ( cd + b ) + d + 1 + ( cd + b ) + 2 ( cd + b )( d + 1) + ( d + 1) − 4dn ,
2 )
(
1 2 2
W2 = ( cd + b ) + d + 1 − ( cd + b ) + 2 ( cd + b )( d + 1) + ( d + 1) − 4dn ,
2 )
1
( 2 2
Z1 = ( cd + b ) + d + 1 + ( cd + b ) − 2 ( cd + b )( d − 1) + ( d + 1) ,
2 )
1
( 2 2
and Z 2 = ( cd + b ) + d + 1 − ( cd + b ) − 2 ( cd + b )( d − 1) + ( d + 1) .
2
□ ) □
We demonstrate the formulas presented in Theorem 3.14 in our next

example.
Example 3.5 a Spanning trees for 1-IPS(2,3,0) (Figure 3.11). By

Theorem 3.14(1), t(1-IPS(c,d,0))=
1 n−2c
( d + 1)( n − c + 1) i
n
c-1
 [d (c + 1) + 1] + (c − 1) 2 d 2 + 2(c + 1)d + 1] 
  i
 2 
 
c −1
 [d (c + 1) + 1] − (c − 1) 2 d 2 + 2(c + 1)d + 1] 
  .
 2 
 
For 1-IPS(2,3,0) , c = 2 , d = 3 and n = 8, so
1
( 3 + 1)( 8 − 2 + 1) − ( ) i
8 2 2
t (1 − IPS (2,3,0) ) =
8
2-1
 [3(2 + 1) + 1] + (2 − 1)2 32 + 2(2 + 1)3 + 1] 
  i
 2 
 
2-1
 [3(2 + 1) + 1] − (2 − 1)2 32 + 2(2 + 1)3 + 1] 
  = 21609.
 2 
 
One Laplacian matrix for 1-IPS(2,3,0) is
 6 −1 −1 −1 −1 −1 −1 0
 −1 6 −1 −1 −1 −1 −1 0 

 −1 −1 6 −1 −1 −1 −1 0
 
 −1 −1 −1 6 −1 −1 0 −1
 −1 −1 −1 −1 6 −1 0 −1
 
 −1 −1 −1 −1 −1 6 0 −1
 −1 −1 −1 0 0 0 3 0
 
 0 0 0 −1 −1 −1 0 3
formula.
Exampl e 3. 5 b Spanning t rees for 1-IPS(2, 3, 0) (Fi gure 3. 12).

By Theorem 3.14(2),
1
t (1 − IPS ( c, d , b )) = (cd + b)b −1 (cd + b + 1)c(d −1) i
n
1
2 (
(cd + b) + d + 1 + (cd + b) + 2 (cd + b)( d + 1) + ( d + 1) − 4dn )i
2 2
2(
(cd + b) + d + 1 − (cd + b) + 2 (cd + b)( d + 1) + ( d + 1) − 4dn )i
1 2 2
2(
(cd + b) + d + 1 + (cd + b) − 2 (cd + b)( d − 1) + ( d + 1) )i
1 2 2
2(
(cd + b ) + d + 1 − (cd + b ) − 2 (cd + b)( d − 1) + ( d + 1) ).
1 2 2
For 1-IPS(2,3,1), c = 2 , d = 3, b = 1 and n = 9, so
1 1−1 2( 3−1)
t (1 − IPS (2,3,1) ) =
9
( ( 2 )( 3) + 1) ( ( 2 )( 3) + 1 + 1) i
1 2 2 
 11 +
2
( ( 2 )( 3) + 1) + 2 ( ( 2 )( 3) + 1) ( 3 + 1) + ( 3 + 1) − 4 ( 3)( 9 ) i

1 2 2 
 11 −
2
( ( 2 )( 3) + 1) + 2 ( ( 2 )( 3) + 1) ( 3 + 1) + ( 3 + 1) − 4 ( 3)( 9 ) i

1 2 2 
 11 +
2
( ( 2 )( 3) + 1) − 2 ( ( 2 )( 3) + 1) ( 3 − 1) + ( 3 + 1) i

1 2 2 
2 
11 − ( ( 2 )( 3) + 1) − 2 ( ( 2 )( 3 ) + 1) ( 3 − 1) + ( 3 + 1)  =

1 4 11 + 13 11 − 13 11 + 37 11 − 37
⋅8 ⋅ ⋅ ⋅ ⋅ = 258048.
9 2 2 2 2

 7 −1 −1 −1 −1 −1 −1 −1 0
 −1 7 −1 −1 −1 −1 −1 −1 0 

 −1 −1 7 −1 −1 −1 −1 −1 0
 
 −1 −1 −1 7 −1 −1 −1 0 −1
 −1 −1 −1 −1 7 −1 −1 0 −1
 
 −1 −1 −1 −1 −1 7 −1 0 −1
 −1 −1 −1 −1 −1 −1 6 0 0
 
 −1 −1 −1 0 0 0 0 3 0
 0 3
 0 0 −1 −1 −1 0 0
formula.
Example 3.5 c Spanning trees for 2-IPS(4,3,0) (Figure 3.13). By

Theorem 3.14(3), t(x-IPS(c,d,0)) =
c
( d −1)  1
1  cd  x c  1− 
 x
( x + d ) x + d
x 
i
cd 
c+
x
c
−1
 2 x
 1  x + d + cd  + 1  x + d + cd  − 4 cd
2
 i
 2 x  2  x  x 
 
c
−1
 2 x
 1  x + d + cd  − 1  x + d + cd  − 4 cd
2
 .
 2 x  2  x  x 
 
For 2-IPS(4,3,0), x = 2, c = 4, d = 3, b = 0, and n = 10, so

t(2-IPS(4,3,0)) =
4
( 3−1)  1
1  ( 4 )( 3)  2 4  1− 
⋅ ( 2 + 3)  2 +  ⋅3  2
i
4+
( 4 )( 3 )  2 
2
4
−1
 2 2 2
 1  2 + 3 + ( 4 )( 3 )  + 1  2 + 3 + ( 4 )( 3)  − 4 ( 4 )( 3)  i
 2  2  2 
 
2 

2 
 
4
−1
 2 2 2
 1  2 + 3 + ( 4 )( 3 )  − 1  2 + 3 + ( 4 )( 3 )  − 4 ( 4 )( 3 )  =
 2  2  2 
 
2 

2 
 
331776.
 7 −1 −1 −1 −1 −1 −1 −1 0 0
 −1 7 −1 −1 −1 −1 −1 −1 0 0 

 −1 −1 7 −1 −1 −1 −1 −1 0 0
 
 −1 −1 −1 7 −1 −1 0 0 −1 −1
 −1 −1 −1 −1 7 −1 0 0 −1 −1
 
 −1 −1 −1 −1 −1 7 0 0 −1 −1
 −1 −1 −1 0 0 0 3 0 0 0
 
 −1 −1 −1 0 0 0 0 3 0 0
 0 0 0 −1 −1 −1 0 0 3 0 

 0 0 0 −1 −1 −1 0 0 0 3

formula.

Chapter 4
Approaches to the Multigraph

Problem
When optimizing the number of spanning trees over a class of graphs,

the problem is fairly straightforward. If we consider the condition that
each pair of nodes has at most one edge between them, and then “relax”
that condition to allow for multigraphs (i.e., allow for multiple edges
between a pair of nodes), one would surmise that the solution to the
problem would be readily attainable. That is simply not the case for the
spanning tree problem. In this chapter, we explore some possible
approaches to one problem in particular.
Kelmans and Chelnokov [Kelmans, 1974] and Shier [Shier, 1974]

n
independently proved that the graph K n − kK 2 , k ≤ has the greatest
2
  n 
number of spanning trees among all graphs in its class Ω  n,   − k  .
  2 
We conjecture that this graph also maximizes the number of spanning
trees over all multigraphs in the class having at most a single multiple
edge of multiplicity equal to two. Based on additional evidence, we
conjecture that K n − kK 2 maximizes the number of spanning trees over
all multigraphs in the class.
147
We define a full neighborhood point (fnp) to be a node w such that

N ( w) = V − {w} . If we consider the case of a connected multigraph with
exactly one double edge, we consider the placement of that multiple edge
in relation to any fnp.
The end nodes of the lone double edge can exist in any of the following
situations:
1. exactly one end node is a fnp;
2. both end nodes are fnp's;
3. neither end node is a fnp.
We denote by Gs the underlying graph of the multigraph M.
Case 1: The multiple edge joins a fnp w and a point v which is not a fnp.
Since N (v ) ≠ V − {v} , there exists a node y ∉ N (v) . Thus, the

neighborhood of w dominates that of y, and, applying Theorem 3.8
(i.e., the Swing Surgery), one of the two edges {w, v} can be replaced
by {v, y} . The number of spanning trees has not decreased; the
multigraph has been transformed into a graph G′ ; and

t ( n, M ) ≤ t (n, G ′) ≤ t ( K n − kK 2 ) . If G ′ ≠ K n − kK 2 , the second inequality
is strict. If G ′ = K n − kK 2 then Gs = K n − ( k + 1) K 2 and exactly one end
node of the multiple edge is the end node of the deleted matching edge.
Let x be the edge {w, v} in M that was removed and replaced by
x ′ = {v, y} from G ′ . Observe that M − x = G ′ − x′ = K n − ( k + 1) K 2 and
Approaches to the Multigraph Problem 149
G  | x is obtained from the connected graph G | x (with loop deleted)

by the addition of an edge. Thus, t ( M | x)  t (G  | x) and t ( M )  t (G ) .
Hence, t ( M )  t ( K n  kK 2 ) .
Case 2: The multiple edge joins two fnp's.

Suppose the multiple edge joins w and v, both of which are of full
 n
neighborhood. Because e     k , there must exist at least two nodes
 2
u and u that are not adjacent. Clearly, N  v  dominates N (u ) , so again
employing the Swing Surgery, edge v, u can be replaced by edge
u , u without reducing the number of spanning trees. Now w and v

satisfy the condition of case (1) in the new graph, and so again we have
t ( M )  t ( K n  kK 2 ) .
Case 3: The multiple edge is incident on two nodes each not of full
neighborhood.
Case 3 is much more involved, but is solved for many instances of
multigraphs in the class. To get a sense of the problem, we require some
further algebraic development. Given two vectors of nonnegative
integers x and y , both of which are of length n, and whose elements are
k k
nonincreasing, x is majorized by y , denoted xx  y , if 
i 1
xi  y
i 1
i
n n
for each 1  k  n  1 and i 1
xi  y
i 1
i [Marshall, 1979]. In other words,
to establish majorization, the direction of the inequalities for the partial

sums cannot change until the final entries in the respective vectors.
Majorization gives an indication of how “spread out” the elements of a

vector are. The more spread out, the smaller the product of the elements,
n n
thus, xx ≺ y implies ∏ i =1
xi ≥ ∏y
i =1
i [Marshall 1979].
Example 4.1 Consider the following three vectors x = (15,10,10,5,3,3) , y = 15,12,11, 4,3,1 and z = 1
x = 15,10,10,5,3,3 , y = (15,12,11, 4,3,1) and z = (16,12,7, 4,4,3) . The table below lists the
partial sums.
k
x = (15,10,10,5,3,3) y = (15,12,11,4,3,1) z = (16,12,7, 4, 4,3)
∑i=1
k=1 15 15 16
k=2 25 27 28
k=3 35 38 35
k=4 40 42 39
k=5 43 45 43
k=6 46 46 46
k k 6 6
Clearly, x ≺ y , as ∑
i =1
xi ≤ ∑ i =1
yi for all k ≤ 5 , and ∑
i =1
xi = ∑y.
i =1
i As
6 6
indicated ∏x
i =1
i = 67500 ≥ 23760 = ∏y .
i =1
i We note that there is no
further majorization relationship among the three vectors. For example,

x ≺/ z and z ≺/ x , as
2 2 4 4
∑
i =1
xi = 25 < 28 = ∑
i =1
zi , but ∑
i =1
xi = 40 > 39 = ∑z
i =1
i .
Similarly, neither x nor z majorize each other.
To prove our next result, namely, for a positive definite symmetric

matrix L , λ (L) ≻ d(L) i.e., the vector of the diagonals of L is
majorized by the vector of its eigenvalues [Marshall, 1979], we
require some further development. Given two non-increasing sequences
of real numbers d = ( d n , d n −1 ,..., d1 ) and d′ = ( d m′ , d m′ −1 ,..., d1′ ) , n > m, d′
is said to interlace d if d i ≤ di′ ≤ d n − m + i , i = 1,..., m.
Example 4.2 Let d = ( 42,32, 22,12, 2 ) and d′ = ( 25, 15, 5 ) , thus n = 5 ,
m = 3 and n − m + i = 2 + i . We note that d1 ≤ d1′ ≤ d3 ( 2 ≤ 5 ≤ 22 ) ,

d 2 ≤ d 2′ ≤ d 4 (12 ≤ 15 ≤ 32 ) , and d3 ≤ d3′ ≤ d5 ( 22 ≤ 25 ≤ 42 ) , so d′
interlaces d, ( 42,32, 25, 22, 15,12, 5, 2 ) .
We use this notion in our next lemma, which is sometimes attributed to

Cauchy.
Lemma 4.1 Let A be a symmetric positive definite ( n − 1) × ( n − 1)

ˆ the
matrix, y an ( n − 1) -dimensional vector, and a ∈ ℝ. Denote by A
ˆ = 
A y
n × n matrix A T  . In addition, let λ1 ≤ λ2 ≤ ... ≤ λn −1 represent
y a
the eigenvalues of A and λˆ1 ≤ λˆ2 ≤ ... ≤ λˆn represent the eigenvalues of
ˆ . Then λˆ ≤ λ ≤ λˆ ≤ λ ... ≤ λˆ ≤ λ ≤ λˆ , i.e., the eigenvalues of A

A 1 1 2 2 n −1 n −1 n
ˆ .
interlace those of A
Proof It suffices to show that λˆ j ≤ λ j ≤ λˆ j +1. Let x ∈ ℝ n −1 , z ∈ ℝ, L j an
x
arbitrary ( n − j + 1) -dimensional subspace of ℝ n and
and xˆ =   . By
z
Theorem 0.5 (Courant-Fischer),
ˆˆ
xˆ T Ax ˆˆ
xˆ T Ax xT Ax
λˆ j = max min ≤ max min = max min = λ j −1.
Lj 0 ≠ x∈L j xˆ T xˆ Ln − j +1 0 ≠ x∈Ln− j +1 xˆ T xˆ Ln − j +1 0 ≠ x∈Ln − j +1 x T x
The upper bound λ j ≤ λˆ j +1 can be demonstrated similarly. □
We use this lemma to prove the theorem:
Theorem 4.2 Let A be a symmetric positive definite n × n matrix, with

diagonal entries arranged in non-decreasing order. Then the vector of
diagonals of A , d = ( d n , d n −1 ,..., d1 ) , majorizes the vector of eigenvalues
of A , λ = ( λn , λn −1 ,..., λ1 ) , i.e., d ≻ λ .
Proof By induction on n. Assume that the result is true for all j × j

matrices satisfying the hypotheses, 2 ≤ j ≤ n − 1. Let A be an n × n
matrix and A n be a submatrix of A obtained by deleting the nth row
and nth column. Let λ1 ≤ λ2 ≤ ... ≤ λn be eigenvalues of A and
λˆ1 ≤ λˆ2 ≤ ... ≤ λˆn −1 the eigenvalues of A n . By the induction hypothesis,

j j
we have ∑
i =1
dn−i ≥ ∑ λˆ
i =1
n −i , j = 1,..., n − 1. By Lemma 4.6, we have
j j
λ1 ≤ λˆ1 ≤ λ2 ≤ λˆ2 ... ≤ λn −1 ≤ λˆn −1 ≤ λn , so ∑ λˆn −i ≥ ∑ λn −i , j = 1,..., n − 1.
i =1 i =1
j j
Therefore, ∑i =1
d n −i +1 ≥ ∑λ
i =1
n −i +1 , j = 1,..., n − 1, and, by Theorem 0.11,
n n
have ∑ d = ∑ λ , i.e., the trace is the sum of the eigenvalues.
i i □
i =1 i =1
Now consider the B-matrix of K n − kK 2 and the B-matrix of a
multigraph M in the same class. Since the B-matrix is positive definite

and symmetric,
n n
∏i =1
di (B( M )) ≥ ∏ λ (B(M )) and
i =1
i
n n
∏ ( i =1
di B ( K n − kK 2 ) ≥ ) ∏λ (B(K
i =1
i n − kK 2 ) . )
Let M be such that
d ( B ( M ) ) ≻ λ B ( K n − kK 2 ) .
( )
Then
n n
∏
i =1
di (B(M )) ≤ ∏ λ (B ( K
i =1
i n − kK 2 ))
which in turn implies

n n
∏ λi (B( M )) ≤ ∏ λi (B( K n − kK )),

i =1 i =1
and thus such a multigraph has fewer spanning trees. Using majorization,
many multigraphs can be shown to have fewer spanning trees than
K n − kK 2 in their respective classes.
Recall that the B-matrix eigenvalue spectrum of K n − kK 2 is n – k

instances of n and k instances of n − 2 . In the overwhelming number
of cases validated by computer, the diagonals of the B-matrix of
these multigraphs majorize the eigenvalues of the B-matrix of the
corresponding K n − kK 2 in its class, and therefore, the multigraphs have
fewer spanning trees. Hence, aside from a relatively small number of
exceptional cases, the majorization argument is valid. The exceptions
can be classified by the following theorem:
Theorem 4.3 The vector of diagonals of the B-matrix of a multigraph M

will not majorize the vector of eigenvalues of a K n − kK 2 in the same
class if and only if
n n
∑ d i ( B ( M ) ) > k (n − 2) = ∑ λi ( B ( K n − kK 2 ) ) .
i = n − k +1 i = n − k +1
i.e., the sum of the last k elements of the vector of diagonals of the
multigraph’s B-matrix is greater than the sum of the corresponding
elements of the vector of K n − kK 2 B-matrix eigenvalues.
Proof To simplify notation, we let d i refer to the diagonals of the B-
matrix of the multigraph and let λi refer to the eigenvalues for the unique
n n
K n − kK 2 in its class. If ∑
i = n − k +1
d i > k ( n − 2) = ∑
i = n − k +1
λi , then
n n n−k n−k n n
∑
i =1
di − ∑
i = n − k +1
di = ∑
i =1
di < ∑ λi = ∑ λi −
i =1 i =1
∑
i = n − k +1
λi .
Hence, the vector of multigraph B-matrix diagonals cannot majorize

the vector of eigenvalues. Now, assume the vector of diagonals of a
multigraph’s B-matrix does not majorize the eigenvalues of K n − kK 2 .
j j
Then there exists a 1 ≤ j < n , so that ∑d < ∑λ
i =1
i
i =1
i . Assume j is the
j −1 j −1
smallest such index. Then d j < λ j , or else ∑ di < ∑ λi , which would
i =1 i =1
contradict the selection of j. Also, j ≤ n − k . If not, then

n n
d j < λ j = n − 2 , and since the d j are nonincreasing, then ∑
i = j +1
di < ∑λ,
i = j +1
i
and so,
n j n j n n
∑ d = ∑ d + ∑ d <∑ λ + ∑ λ = ∑ λ ,
i =1
i
i =1
i
i = j +1
i
i =1
i
i = j +1
i
i =1
i
which contradicts the equality of the sums. Therefore, j ≤ n − k ,

which implies d j < n . In fact, d j must be n − 1, or once again the
sums would not be equal. Now, having established that

n n n −k n−k
∑
i = j +1
di > ∑
i = j +1
λi ,with ∑
i = j +1
di < ∑λ ,
i = j +1
i
it follows that
n n
∑ di < ∑ λi = k (n − 2) . □
i = n − k +1 i = n − k +1
We prove the following corollary to the theorem.

Corollary 4.4 If the vector of diagonals does not majorize the vector of
eigenvalues, then d1 ≤ 2n − k − 2, and d n ≥ n − k .
Proof Given the non-majorization assumption, let j be the smallest index

j j
such that ∑
i =1
di < ∑λ i =1
i . We have shown that j ≤ n − k , and d j = n − 1 ,
so that
j j
∑ i =1
di ≥ d1 + ( j − 1)(n − 1) = d1 + jn − j − n + 1 and ∑λ
i =1
i = jn .
j j
Therefore, since ∑i =1
di < ∑λ
i =1
i , we have
d1 + jn − j − n + 1 < jn, or d1 < n + j − 1 .

Now, since
j ≤ n − k , d1 < n + (n − k ) − 1 = 2n − k − 1 .
Because there is non-majorization,
n n
∑
i = n − k +1
d i > k ( n − 2) = ∑
i = n − k +1
λi , and d n − k +1 ≤ n − 1 .
Thus,
n
∑
i = n − k +1
d i < (k − 1)(n − 1) + d n so (k − 1)(n − 1) + d n > k (n − 2)
implying d n > n − k − 1 . □
Even though majorization handles the vast majority of instances, the

exceptional cases must be handled individually, and this cannot be done
efficiently given the multitude of multigraph realizations for a given
degree sequence. A specific example for which majorization will not

hold is the B-matrix diagonal sequence with two distinct entries, n and
n − 1 , where n occurs n − 2k times, and n − 1 occurs 2k times. The
eigenvalues for B ( Kn − kK 2 ) are ( n,…, n, n − 2,…, n − 2 ) where n
occurs with multiplicity n − k and n − 2 occurs with multiplicity k. The
sum of the last k diagonals is k ( n − 1) which is clearly greater than
k ( n − 2 ) , so by Theorem 4.12, the vector of eigenvalues for the B-matrix

is not majorized. Note that the multigraph specified above has the same
degree sequence as K n − kK 2 in its class. Hence, since the majorization is
in the other direction, it cannot be applied here.

Chapter 5
Laplacian Integral Graphs and

Multigraphs
Of particular interest is the classification of graphs by the nature of

their eigenvalues. In this chapter, we discuss those graphs and
multigraphs whose eigenvalues are all integers. There are many such
graphs and multigraphs, and this chapter is not intended to be an
exhaustive exposition. Much of this material has appeared in earlier
chapters of this book.
5.1 Complete Graphs and Related Structures
The Laplacian eigenvalues for a complete graph K n , as given in the

proof of Cayley’s Theorem (Proposition 1.6(a)) are 0, and n − 1
instances of n. Of course, this leads to the fact that the eigenvalues for
H ( K 2 ) are 0 and 2. This fact can be used to prove the following
proposition, which is used in Chapter 4.
Proposition 5.1 The eigenvalues for H ( Kn − kK 2 ) are n,...n, n –

2,...,n – 2,0, where n occurs with multiplicity n – k – 1 and n − 2 occurs
with multiplicity k.
159
n
Proof Obviously, k ≤ . Now, K n − kK 2 has as its complement k
2
n
copies of K 2 , and if k < , n − 2k copies of K1 . Thus, the
2
eigenvalues of the complement of K n − kK 2 are n − k zeros, and k
instances of 2. By Proposition 1.4(a), if G is the complement of

graph G, and the eigenvalues of G are arranged in non-decreasing
order, then λk + 2 (G) = n − λn −k (G) for k = 0,…, n − 2 . Direct application
yields the result. □
Remark 5.1 The proofs of the spanning tree formulas for complete
bipartite and regular complete k-partite graphs rely heavily on the fact
that the complements of these graphs are disjoint complete graphs. Thus,
these graphs are also Laplacian integral.
For the K n − Pp (n > p) graphs in part (c)(i) of Proposition 1.12,
the complements are Pp , which do not have integer eigenvalues for
all values of p, so therefore, the graphs themselves do not have

integer eigenvalues, either. The same is true for K n − C p graphs in
part (c)(ii); while C3 ,C4 , and C6 all have integer eigenvalues, this is
not the case for all cycles. For example, it can be shown by direct
application of Theorem 1.5 that C8 has eigenvalue spectrum
0, 2 − 2, 2 − 2, 2, 2, 2 + 2, 2 + 2, 4, so its complement will not have

an integral spectrum, either.
Laplacian Integral Graphs and Multigraphs 161
5.2 Split Graphs and Related Structures
In Chapter 3, the Laplacian eigenvalues for various types of split

graphs were introduced. Recall that a split graph is one in which the
nodes can be partitioned into a clique and an independent set, while a
threshold graph is a split graph with the neighborhood domination
property, i.e, for all pairs of nodes u and v in G, N (u ) − {v} ⊆ N (v ) − {u}
whenever deg(u ) ≤ deg(v ) .
As Theorem 3.5 gives the Laplacian characteristic polynomial for
threshold graphs in factored form, we observe that the eigenvalues
are all integers. Until 2010, it was unknown if there existed Laplacian
integral, non-threshold split graphs. Kirkland, Alverez de Freitas, Del
Vecchio and de Abreu [Kirkland, 2010] proved the existence of an
infinite such class of split graphs, the “biregular split graphs,” using
techniques such as Kronecker products, balanced incomplete block
designs and solutions to certain Diophantine equations. The biregular
split graph is the same graph as a certain subclass of the x-IPS(c,d,0)
graphs presented in section 3 of Chapter 3, whose spanning tree formula
can be proven using elementary matrix operations as outlined in the
sketch of the proof of Theorem 3.14(2). The Laplacian integral nature of
these graphs was discovered independently.
Recall that the spanning tree formula is given as
c
( d −1)  1
1  cd  x c  1− 
 x
t ( x − IPS (c, d ,0) = ( x + d ) x + d
x 
i
cd 
c+
x
c
−1
 2 x
 1  x + d + cd  + 1  x + d + cd  − 4 cd
2
 i
 2 x  2  x  x 
 
c
−1
1 2 2 x
  x + d + cd  − 1  x + d + cd  − 4 cd  .
 2 x  2  x  x 
 
The characteristic polynomial giving the result is
c
( d −1)
( λ − ( x + d ) )  cdx + x − λ 
c x
c−
λ (d − λ ) x i
c
−1
 2  cd  cd 2  x
 λ −  x + d + λ −  ,
  x  x 
which has roots
cd c
λ = 0, ( x + d ) , x + with multiplicity ( d − 1) , and d ,
x x
1 2
cd 2 
 cd  1  cd 
 x + d +  +  x + d +  − 4 
 2  x  2  x  x 
and
1 2
cd 2 
 cd  1  cd 
 x + d +  −  x + d +  − 4 
 2  x  2  x  x 
 1
each with multiplicity c 1 −  .
 x
We note that, if the discriminant is an odd perfect square, the graph
will be Laplacian Integral. For example, the graph 3-IPS(6,2,0) has
Laplacian spectrum 0,2,2,2,2,5,7,7, obtained by substituting x = 3,
x c==3,6c = and
6 and d = 2 . Some other examples of Laplacian integral
x-IPS(c,d,0) graphs are 3-IPS(12,2,0), and 2-IPS(4,3,0) . [Fuller, 2010]
There is another class of graphs, based on split graphs, whose

eigenvalues are integral. A graph is defined as almost split if it is not
a split graph, but can be transformed to one by the deletion of one edge.
We will denote by AS(n,2) an almost split graph on n nodes, with a
clique of order n − 2 , two adjacent nodes that are also adjacent to
roughly half the clique nodes, with one clique node in their common
neighborhood. The degrees of the non-clique nodes will be the same in
n +1
the case where n is odd (i.e., their degrees will each be ), and if n
2
is even, the degrees of these nodes will differ by one (i.e., one will be
n n
of degree and the other will be of degree + 1 ). When n is odd,
2 2
n +1 n +1
AS(n,2) has degree sequence , , n − 2,..., n − 2, n − 1. When n is
2 2 n 3 terms
−
n n
even, AS(n,2) has degree sequence , + 1, n − 2,..., n − 2, n − 1. It is
2 2 n 3 terms
−
n2 − 3n + 6
easy to show that such graphs will always have edges,
2
regardless of the parity of n. Figures 5.1 and 5.2 depict two examples of
AS(n,2).
Figure 5.1: The graph AS(11,2).
Figure 5.2: The graph AS(12,2).

Recall that a split graph satisfies the mth Erdos and Gallai inequality
with equality [Hammer, 1981], where m represents the largest index i in
a degree sequence such that d i ≥ i − 1. For the graph in Figure 5.1 (a),
the degree sequence is 10, 9, 9, 9, 9, 9, 9, 9, 9, 6, 6; m = 9, and
9
∑d
i =1
i = 82 ≤ 84 = 9 ( 9 − 1) + ( 6 + 6 ) , so the m = 9th Erdos and Gallai
inequality is not an equality. Therefore, AS(11,2) is not a split graph.

This is the case for any AS(n, 2) graph.
The next proposition gives a formula for their eigenvalues.
Proposition 5.2 (a) The eigenvalues for AS ( n, 2 ) , n odd, are
n +1 n +1
0, , , n − 1,..., n − 1, n, n, all of which are integers.
2 2 n 5 terms
−
n n
(b) The eigenvalues for AS ( n, 2 ) , n even, are 0, , + 1, n − 1,..., n − 1, n, n,
2 2 n 5 terms −
all of which are integers.

Proof The proof proceeds from the observation that the complement
of AS(n,2) is comprised of three components: one K1 , and two
disjoint stars.
(a) If n is odd, the complement of AS ( n, 2 ) is one K1 , and two
disjoint identical stars K n −3 . It can be deduced from

1,
2
Proposition 1.12(b)(i) that a star K n −3 has eigenvalues

1,
2
n−3
0, 1,...1, + 1. Thus, the eigenvalues for these three
n −3
2
-1 terms
2
components, arranged in non-decreasing order, are

n −1 n −1
0,0,0, 1,...,1 , , , and so, by Proposition 1.4 (a),
n − 5 terms
2 2
AS ( n, 2 ) has Laplacian spectrum
n +1 n +1
0, , , n − 1,..., n − 1, n, n.
2 2 n 5 terms −
(b) The proof for n even follows similarly. The only difference is
that, instead of two identical stars, the two stars in the
complement are K n −2 and K n−4 .
1, 1,
2 2
2
 n +1 n −5
Remark 5.2 The graph AS ( n, 2 ) , n even, has   ( n − 1) n
 2 
2
n n−5
spanning trees, while AS ( n, 2 ) , n odd, has   ( n − 1) n spanning
 2
trees.
5.3 Laplacian Integral Multigraphs
There are several multigraphs that are shown to be Laplacian integral.

For example, by Theorem 1.30, gK n (i.e., the complete graph with g
multiple edges between every pair of nodes) has eigenvalue zero, and
n − 1 instances of gn.
Another Laplacian integral multigraph is one with an underlying

threshold graph, first presented by Heinig and Saccoman [Heinig, 2012].
We describe the construction of such a graph as follows: we begin with a
proper threshold graph on n nodes with n1 nodes of degree v1 . Each
edge of the induced complete subgraph induced by the common
neighborhood of the cones is replicated µ times, and we denote this
multigraph by MT µ n ;v n1 . Figure 5.3 depicts MT7;2

3
2.
1
3
Figure 5.3: The multigraph MT7;2 2 which has degree sequence 2, 2, 4, 4, 4, 8, 8.
The inner box in the clique houses a 3K2 .
As in the case of threshold graphs, the eigenvalue spectrum for the

MT graphs is predicated on the parity of the number of unique degree
sequence terms.
Proposition 5.3 [Heinig, 2012] The multigraph MT µ n;v n1 has Laplacian

1
eigenvalues as follows:
(a) (even case: v1 = n − n1 − 1 )
λ = (0, v1n −1 , n, (v2 + µ ) n −1 ) , degree sequence

1 2
v1n1 , v2 n2 .
(b) (odd case: v1 < n − n1 − 1 )
λ = (0, v1n , (v2 + 1)n −1 , n, (v3 + µ ) n −1 , degree sequence

1 2 3
v1n1 , v2 n2 , v3n3 .
The proof of Proposition 5.3 follows from a sequence of elementary
matrix row operations, similar to the proof of Theorem 3.14(2).
Remark 5.3 The graph MT µ n ;v n1 , in the even case, has

1
n2 −1 n3 −1
v1n1 −1 ( v2 + 1) ( v3 + µ ) spanning trees, while MT µ n;v n1 , in the odd
1
n2 −1
case, has v1n1 −1 ( v2 + µ ) spanning trees.

Bibliography
[Ball, 1982] M.O. Ball and S.J. Provan, Bounds on the reliability polynomial for
shellable independence systems, SIAM J. Alg. Disc. Meth. 3, 166-181, 1982.
[Ball, 1983] M. Ball and J.S. Provan, The complexity of counting cuts and of computing
the probability that a graph is connected, SIAM J. on Computing 12, 777-788,1983.
[Bapat, 2010] R. B. Bapat. Graphs and Matrices. Springer: London, Hindustan Book
Agency, New Dehli, India, 2010.
[Bleiler, 2007] S. Bleiler, and J.T. Saccoman. A Correction in the Formula for the
Number of Spanning Trees in Threshold Graphs. Australasian Journal of
Combinatorics 37, 205-213, 2007.
[Bogdanowitz, 1985] Z. Bogdanowicz, Spanning trees in undirected simple graphs,
Doctoral Dissertation, Stevens Tech, Computer Science Department, 1985.
[Bogdanowitz, 2009] Z. Bogdanowicz, Undirected simple connected graphs with
minimum number of spanning trees, Discrete Math 309 (no. 10), 3074-3082, 2009.
[Boesch, 1984] F.T. Boesch and C.L. Suffel, A survey of the algebraic approach to the
study of spanning trees, Technical Report, Stevens Tech, Computer Science
Department, 8401, 1984.
[Brooks, 1940] R.L. Brooks, C.A.B. Smith, A.H. Stone and W.T. Tutte, Dissection of a
Rectangle into Squares, Duke Mathematics Journal 7, 312-340, 1940.
[Brown, 1993]. J.J. Brown, C. Colbourn, and J.S. Devitt, Network transformations and
bounding network reliability, Networks 23, 1- 17, 1993.
[Cayley, 1889] A. Cayley, A theorem on trees, Quart. J. Math 23,376-378: 69, 1889.
[Cheng, 1981] C. Cheng, Maximizing the total number of spanning trees in a graph: two
related problems in graph theory and optimum design theory, Journal of
combinatorial Theory B 31,1981.
[Chartrand, 2011] G. Chartrand, L. Lesniak, and P. Zhang, Graphs and Digraphs (5th ed),
CRC Press, Boca Raton, FL, 2011.
169
170 Bibliography
[Erdos, 1960] P. Erdos and T. Gallai, Graphs with prescribed degrees of vertices
(Hungarian), Mat. Lapok 11, 264-274, 1960.
[Fiedler, 1973] M. Fieldler, Algebraic Connectivity of Graphs, Czech. Math Journal 23,
298-305, 1973.
[Fuller, 2010] J. Fuller. Spanning trees of extended ideal proper split graphs. Honors
Thesis, Seton Hall University, Department of Mathematics and Computer Science,
2010.
[Hakimi, 1962] Hakimi, S. L. On realizability of a set of integers as degrees of the
vertices of a linear graph. I. J. Soc. Indust. Appl. Math. 10, 496–506, 1962.
[Hammer, 1978] P.L. Hammer, T. Ibaraki, and B. Simeone. Degree sequences of
threshold graphs. Congressus Numeratium, Vol. 21, 329-355, 1978.
[Hammer, 1981] P.L. Hammer and B. Simeone. The splittance of a graph. Combinatoria
Vol. 1, 275-284, 1981.
[Heinig, 2012] M. Heinig and J.T. Saccoman, Laplacian integral multigraphs,
Congressus Numerantium 212, 131-143, 2012.
[Kelmans, 1967] A.K. Kelmans, Properties of characteristic polynomial of a graph,
Kibernetiku-na sluzbu kommunizmu 4, Energija, Moskva-Leningrad, 27-41, 1962.
[Kelmans, 1974] A.K. Kelmans and V. M. Chelnokov, A certain polynomial of graph
and graphs with an extremal number of trees, Journal of Combinatorial Theory B 16,
197-214, 1974.
[Kelmans 1981]. On graphs with randomly deleted edges, Acta Math Acad. Sci Hungary
37, 77-88, 1981.
[Kirchhoff, 1847] Über die ausflösung der gleichungen, auf welche man bei der
untersuchung der linearen vertheilung galvanischer ströme geführt wird. (“On the
solution of the equations obtained from the investigation of the linear distribution of
galvanic currents”). Annalen der Physik und Chemie 72,497-508, 1847.
[Lancaster, 1985] P. Lancaster and M. Tismenetsky, Theory of Matrices, Academic
Press, Orlando, FL, 1985.
[Marshall, 1979]. A.W. Marshall and L. Olkin, Inequalities: The Theory of Majorization
and its Applications, Academic Press, New York, 1979.
[Merris, 1994] R. Merris, Degree maximal graphs are Laplacian integral,
Linear Algebra Appl. 199, 381–389, 1994.
[Michewicz, 2007] J. Michewicz and J.T. Saccoman. A formula for the number of
spanning trees of certain non-threshold split graphs. Congressus Numerantium, Vol.
184, 85-96, 2007.
Bibliography 171
[Moore, 2013] K. Moore and J.T. Saccoman. Spanning trees of almost ideal proper split
graphs. Congressus Numerantium, Vol. 216, 129-141, 2013.
[Moskowitz, 1958] F. Moskowitz, The analysis of redundancy networks, AIEE
Transactions on Commumication Electronics 39, 627-632, 1958.
[Petingi, 1996] L. Petingi, J.T. Saccoman. and L. Schoppmann, Uniformly least reliable
graphs, Networks 27(2), 125-131, 1996.
[Seshu, 1961] S. Seshu and M. Reed, Linear Graphs and Electrical Networks, Addison-
Wesley, Reading, MA, 1961.
[Shier, 1974] D. Shier, Maximizing the spanning trees in a graph with n nodes with n
nodes and m edges, Journal of Research of the National Bureau of Standards-B 78B,
193-196, 1974.
[Schoppmann, 1990] L. Schoppmann, Network bounds for edge reliability, Doctoral
dissertation, Stevens Tech, Mathematics Department, 1990.
[Satyanarayana, 1992] A. Satyanarayana, L. Schoppmann, and C.L. Suffel, A reliabilty-
improving graph transformation with applications to network reliability, Networks 22,
209-216, 1992.
[Temperley, 1964] H.N.V. Temperley, On the mutual cancellation of cluster integrals in
Mayer’s fugacity series, Proc. Phys. Soc. 83, 3–16, 1964.

Index
Acylic, 6 graph, 10, 11, 52, 57, 58, 59, 66, 77,
Addition of edges, 15-16 91
Adjacency matrix, 4, 41 k-partite, 11, 59, 69, 160
Adjacent Component, 6
from (a node in a directed graph), 8 Cone points, 113, 114, 115, 121, 122,
to (a node in a directed graph), 8 136, 140, 166
Adjacent Connected multigraph, 6, 7
edges, 4 Contraction, 17, 127, 129, 130
nodes, 4 Cospectral, 90
Adjugate matrix, 44 Courant-Fischer Max-Min Theorem,
Algebraic complement, 77 26-27
All Terminal Reliability (ATR), 37, Cycle, 6, 40, 58, 64, 71, 160
111, 126, 134
Arc, 8 δ (G ) (minimum degree of a graph
Arc-set, 8
G), 46, 48
Balloon graph, 125, 134 ∆ (G ) (maximum degree of a graph
Basis, 24 G), 81, 82, 83, 87
Binet-Cauchy Theorem, 21-23 Degree of a node, 4, 5, 43, 83, 87, 90
Bipartite graph, 11 Degree matrix, 43
Degree sequence, 113, 115, 118, 119,
Cauchy-Schwarz Inequality, 97 136
Cayley’s theorem, 50 Deleted neighborhood, 111
Characteristic polynomial, 48, 56, 58, Deletion
63, 72, 74, 81, 82, 106, 111-115 of nodes, 13-14, 57
Chebyshev polynomial, 63 of edges, 14, 15, 65
Circulant
graph, 12, 53-57, 66-68 Determinant, 20, 21, 22, 30, 40, 41, 49,
matrix, 54 62, 63, 64
Class, 3 Dimension of a subspace (dim), 23, 25,
Clique, 10 27
Cofactor, 44-46, 72, 75 Directed cycle, 38
Complement, 13, 46, 56-57, 59, 74, 77, Directed Graph (digraph), 7
78, 134, 54, 55, 72, 75 Directed multigraph, 7
Complete Disconnected multigraph, 6, 46, 47, 50,
bipartite graph, 11, 57, 58, 64, 67, 78, 85
68, 108 Dominant Matrix Theorem, 30-31
173
174 Index
Dominates, 111, 133 Jacobian, 99

Dot product, 24 Join of two graphs, 18, 74
Double-star
κ (G ) (Node connectivity of a graph
Edge, 2
G), 46, 48
Edge-set, 2
Kirchhoff’s Matrix Tree Theorem, 41,
End-node, 4
45
Eigenvalues, 25-27, 29-32, 35, 36, 105,
k-partite, 11, 57, 59, 69, 160
15-153
K n minus a cycle, 58, 71
Of adjacency matrix, 50-55
Of Laplacian matrix, 43-44, 46-50, K n minus a matching, 15, 57, 58, 68,
94, 106, 109, 110 147, 149, 153, 154, 157
Of Temperley’s B-matrix, 72-75, K n minus a path, 58, 70-71
153-156, 157
Eigenvectors, 24-26, 45, 47, 85, 151 Kroenecker
Equicofactor matrix, 46, 75 polynomial, 35
Euclidean norm, 24, 25 product, 32, 35, 36
sum, 32, 51, 52
Factoring theorem, 127-131, 133
Fiedler bounds, 46, 78, 86 Lagrange multiplier technique, 96, 99
First Theorem of Laplacian matrix, 41, 43-53, 56, 87, 90,
Directed Multigraph Theory, 9 159-166
Multigraph Theory, 4, 42 Leaf (leaves), 6
Full neighborhood point (fnp), 148, 149 Length
Of a cycle, 6
Geometric representation Of a path, 5
Of a directed graph, 8 Linearly
Of a multigraph, 2 Dependent, 23, 40
Gersgorin’s Theorem, 31 Independent, 23, 24
Graph, 2 Lollipop graph, 123-125
Hyperdominant matrix, 43 Majorization, 149, 150, 154, 156, 157

Matching, 15, 56
Identity matrix, 19 Matrix, 19
(i,j)-cofactor, 19, 44, 49, 65, 66, 68, 69, Maximum spanning tree problem, 89-
70-72 92, 94, 147
(i,j)-minor, 44 Minimum spanning tree problem,
Incident edge, 4 111, 123
Incident Minor of a matrix, 27
from (a node in a directed graph), 8 Multi-arc, 7
to (a node in a directed graph), 8 Multiedge, 2
Indegree, 8 Multigraph, 1, 2, 3, 4, 93-95, 106-109,
Independent set 147-157, 165-167
of edges, 15 Multigraph complement, 84, 85
of nodes, 5 Multiplicity of a multiedge, 2
Induced submultigraph, 5 Multistar, 81, 82, 83
Interlace, 151-152
Isomorphic, 3 Neighborhood domination, 133
Neighborhood nesting, 112, 128, 135
Index 175
Neighborhood of a node, 111, 112, Of K p, p − pK2 , 58-60, 61, 70

133,135
Nodal admittance matrix, (see Of K p,q , 57, 59, 61, 69
Laplacian matrix) Of K r , r ,...r , 57, 59-61, 69
Node, 2 k terms
Node-arc incidence matrix, 36
Non-negative definite, 77, 88 Of MT µ n; v n1 , 167
1
Of regular graph, 50-51

Order of a multigraph, 2 Of threshold graph, 122
Orientation of a multigraph, 9, 40 Of x − IPS ( c, d , b ) , 137-145, 162
Orthonormal basis, 25, 26
Outdegree, 8 Spectrum, 28
Split graph, 135
Path, 5, 6, 56, 59 Ideal, 136-137
Pendant node, 6, 39 Proper, 136-137
Positive definite matrix, 27, 28 Star, 11
Positive semi-definite matrix, 27, 28, Subgraph, 5
30, 43, 87, 88, 95 Subspace, 23-27, 152
Product of two graphs, 18, 49, 64 Submultigraph, 5
Swing surgery, 125, 126, 127, 134,
Realizable k-tuple, 132, 133 148, 149
Reduced Symmetric Matrix, 25, 26, 27, 29
Laplacian matrix, 46
node-arc incidence matrix, 39,
40-41 Temperley’s B-matrix, 72, 74, 75, 153,
Regular graph, 5, 49, 50, 51, 55, 90 154, 155, 157
And spanning trees, 76
Size Threshold graph, 111-125, 134-135,
of a multigraph, 2 137
of a matrix, 19 Tree, 6
Span of a set of vectors, 23 Trace, 31, 105
Spanning sub directed graph
Spanning Underlying
subgraph, 5, 6 graph of a multigraph, 5
submultigraph, 5, 83 multigraph of a directed multigraph,
Spanning tree. 6, 7 9
Uniform lower bound, 134
Of AS ( n,2 ) , 165
Union of two graphs, 17, 56-59, 112
Of circulant, 55-56, 66 Unimodular matrix, 39
Of G , 56
Of K n , 50 V(2), 1
Vector, 19, 23, 24, 25, 27, 29, 39, 55,
Of gK n , 91-92 88, 139, 149, 154, 156, 157
Of Kn × K m , 52-53, 66
Zero
Of Kn − kK 2 , 57, 59, 68, 157-160
matrix, 19
Of Kn − C p , 58, 64-65, 71 vector, 19
Of Kn − Pp , 55-56, 66

Daniel J Gross, John T Saccoman, Charles L Suffel-Spanning Tree Results For Graphs and Multigraphs - A Matrix-Theoretic Approach-World Scientific Publishing Company (2014)

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Daniel J Gross, John T Saccoman, Charles L Suffel-Spanning Tree Results For Graphs and Multigraphs - A Matrix-Theoretic Approach-World Scientific Publishing Company (2014)

Загружено:

Авторское право:

Доступные форматы

8963_9789814566032_TP.

indd 1 19/8/14 12:21 pm

This page intentionally left blank

8963_9789814566032_TP.indd 2 19/8/14 12:21 pm

British Library Cataloguing-in-Publication Data

SPANNING â•›TREE â•›RESULTS â•›FOR â•›GRAPHS â•›AND â•›MULTIGRAPHS

LaiFun - Spanning Tree Results for Graphs.indd 1 20/8/2014 9:51:45 AM

To Mary, the smartest woman I know

To John J. Saccoman, my father

To my mother Mildred and my father Charles

This page intentionally left blank

This book is concerned with the calculation of the number of spanning

For small values of p, the reliability polynomial is dominated by the

In Chapter 0, we present some graph theory and matrix theory

Preface ................................................................................................................ vii

0 An Introduction to Relevant Graph Theory and Matrix Theory .................. 1

1 Calculating the Number of Spanning Trees: The Algebraic Approach ......... 37

2 Multigraphs with the Maximum Number of Spanning Trees: An Analytic

3 Threshold Graphs .............................................................................................. 111

4 Approaches to the Multigraph Problem ........................................................... 147

5 Laplacian Integral Graphs and Multigraphs ................................................... 159

Bibliography ....................................................................................................... 169

Index .................................................................................................................... 173

This book is concerned with the calculation of the number of spanning

0.1 Graph Theory

consideration we use the notation M = (V ( M ) , mM ) . The elements of V

nodes. A multiedge is an element {u , v} ∈ V(2) such that m ({u, v}) ≠ 0 ;

m({u, v} ) is called the multiplicity of the multiedge. If for each {u, v} ,

A multigraph has a geometric representation in which each element

We denote the class of all multigraphs of order n and size e by Ω( n, e) .

In Figure 0.1 M ∈ Ω(5,10) and G ∈Ω ( 5,7 )

Figure 0.1: A multigraph M and a graph G.

Note that a multigraph may have several geometric representations

M ′′ = (V ′′, m′′) are isomorphic if there is a bijection f : V ′ → V ′′ such

that for every {v1 , v2 } ∈V(2)′ , m ′ ({v1 , v2 }) = m ′′ ({ f (v ) , f (v )}) .

two multigraphs depicted in Figure 0.2 are isomorphic, under the

Figure 0.2: Two isomorphic multigraphs.

The following theorem, the First Theorem of Multigraph Theory, was

Theorem 0.1 (The First Theorem of Multigraph Theory) The sum

Proof Let x = vi v j be an edge of M, then x contributes 1 to both

Another way to represent a multigraph is by a matrix. Let M = (V , m)

be a multigraph of order n, with node set V = {v1 , v2 ,…, vn } . The

adjacency matrix of M, denoted A ( M ) or simply A is the n × n matrix

Example 0.1 The adjacency matrix of the multigraph in Figure 0.1 is

A multigraph M is r-regular if each node has degree equal to r. When

M = (V , m) if V ′ ⊆ V and m′ ({u, v} ) ≤ m ({u, v} ) for all {u , v} ∈V(2)

V ′ = V , then M ′ is called a spanning submultigraph. If M is a graph

induced submultigraph and is denoted by V ′ . If V ′ contains no

edges, i.e. m ({u, v} ) = 0 for all {u, v} ∈V(2)′ then V ′ is called an

independent set of nodes. If V ′ = V and m′ ({u, v} ) = min m ({u, v} ) ,1{ }

underlying graph of the multigraph (V , m) . In Figure 0.1 G is the

sequence of nodes and edges v1 , x1 , v2 , x2 , v3 ,… , xk −1 , vk , xk , v1 , where

{v0 , v1 ,…, vk } are distinct nodes and x i is an edge between vi −1 and v i ,

1 ≤ i ≤ k − 1 , and xk is an edge between vk and v0 . The length of a cycle

Figure 0.3: A multigraph M and a spanning tree T.

One of the purposes of this book is to determine which multigraph

spanning trees among all multigraphs in   n, e  . Since spanning trees

Choose a node u at random, set V1  u , V2  V  V1 , and E   . If

V2   then since M is connected there is at least one edge with one

T  V1 , E  is a spanning tree.

As for multigraphs, we refer to the elements of V as nodes. A multi-arc is

an element ( u, v ) ∈ (V × V − {( u, u ) | u ∈V }) such that m ( ( u, v ) ) ≠ 0 ;

m ( ( u, v ) ) is called the multiplicity of the multi-arc. If for each pair