Lowell W. Beineke & Robin J. Wilson & Peter J. Cameron - Topics in Algebraic Graph Theory

Topics in Algebraic Graph Theory
The rapidly expanding area of algebraic graph theory uses two different branches of
algebra to explore various aspects of graph theory: linear algebra (for spectral theory) and
group theory (for studying graph symmetry). These areas have links with other areas of
mathematics, such as logic and harmonic analysis, and are increasingly being used in such
areas as computer networks where symmetry is an important feature. Other books cover
portions of this material, but this book is unusual in covering both of these aspects and
there are no other books with such a wide scope.
This book contains ten expository chapters written by acknowledged international
experts in the eld. Their well-written contributions have been carefully edited to enhance
readability and to standardize the chapter structure, terminology and notation throughout
the book. To help the reader, there is an extensive introductory chapter that covers the
basic background material in graph theory, linear algebra and group theory. Each chapter
concludes with an extensive list of references.
LOWELL W. BEI NEKE is Schrey Professor of Mathematics at Indiana University-
Purdue University Fort Wayne. His graph theory interests include topological graph
theory, line graphs, tournaments, decompositions and vulnerability. With Robin J. Wilson
he has edited Selected Topics in Graph Theory (3 volumes), Applications of Graph Theory
and Graph Connections. He is currently the Editor of the College Mathematics Journal.
ROBI N J . WI LSON is Head of the Pure Mathematics Department at the Open University
and Gresham Professor of Geometry, London. He has written and edited many books on
graph theory and combinatorics and on the history of mathematics, including Introduction
to Graph Theory and Four Colours Sufce. His interests include graph coloring, spectral
graph theory and the history of graph theory and combinatorics.
PETER J . CAMERON, internationally recognized for his substantial contributions to the
area, served as academic consultant for this volume. He is a professor of mathematics at
Queen Mary, University of London, and his mathematical interests are in permutation
groups and their operands (logical, algebraic and combinatorial).
i
ENCYCLOPEDI A OF MATHEMATI CS AND I TS APPLI CATI ONS
FOUNDING EDITOR GIAN-CARLO ROTA
Editorial Board
R. Doran, P. Flajolet, M. Ismail, T. Y. Lam, E. Lutwak
Volume 102
The titles below, and earlier volumes in the series, are available from booksellers or
from Cambridge University Press at www.cambridge.org
50 F. Borceux Handbook of Categorical Algebra I
51 F. Borceux Handbook of Categorical Algebra II
52 F. Borceux Handbook of Categorical Algebra III
53 V. F. Kolchin Random Graphs
54 A. Katok and B. Hasselblatt Introduction to the Modern Theory of Dynamical Systems
55 V. N. Sachkov Combinatorial Methods in Discrete Mathematics
56 V. N. Sachkov Probabilistic Methods in Discrete Mathematics
57 P. M. Cohn Skew Fields
58 R. Gardner Geometric Tomography
59 G. A. Baker Jr. and P. Graves-Morris Pade Approximants, 2ed
60 J. Krajicek Bounded Arithmetic, Propositional Logic and Complexity Theory
61 H. Groemer Geometric Applications of Fourier Series and Spherical Harmonics
62 H. O. Fattorini Innite Dimensional Optimization and Control Theory
63 A. C. Thompson Minkowski Geometry
64 R. B. Bapat and T. E. S. Raghavan Nonnegative Matrices with Applications
65 K. Engel Sperner Theory
66 D. Cvetkovic, P. Rowlinson, and S. Simic Eigenspaces of Graphs
67 F. Bergeron, G. Labelle, and P. Leroux Combinatorial Species and Tree-Like Structures
68 R. Goodman and N. Wallach Representations and Invariants of the Classical Groups
69 T. Beth, D. Jungnickel, and H. Lenz Design Theory I, 2ed
70 A. Pietsch and J. Wenzel Orthonormal Systems for Banach Space Geometry
71 G. E. Andrews, R. Askey, and R. Roy Special Functions
72 R. Ticciati Quantum Field Theory for Mathematicians
73 M. Stern Semimodular Lattices
74 I. Lasiecka and R. Triggiani Control Theory for Partial Differential Equations I
75 I. Lasiecka and R. Triggiani Control Theory for Partial Differential Equations II
76 A. A. Ivanov Geometry of Sporadic Groups I
77 A. Schinzel Polynomials with Special Regard to Reducibility
78 H. Lenz, T. Beth, and D. Jungnickel Design Theory II, 2ed
79 T. Palmer Banach Algebras and the General Theory of

-Algebras II
80 O. Stormark Lies Structural Approach to PDE Systems
81 C. F. Dunkl and Y. Xu Orthogonal Polynomials of Several Variables
82 J. P. Mayberry The Foundations of Mathematics in the Theory of Sets
83 C. Foias et al. NavierStokes Equations and Turbulence
84 B. Polster and G. Steinke Geometries on Surfaces
85 R. B. Paris and D. Karninski Asymptotics and MellinBarnes Integrals
86 R. McEliece The Theory of Information and Coding, 2ed
87 B. Magurn Algebraic Introduction to K-Theory
88 T. Mora Systems of Polynomial Equations I
89 K. Bichteler Stochastic Integration with Jumps
90 M. Lothaire Algebraic Combinatorics on Words
91 A. A. Ivanov and S. V. Shpectorov Geometry of Sporadic Groups II
92 P. McMullen and E. Schulte Abstract Regular Polytopes
93 G. Gierz et al. Continuous Lattices and Domains
94 S. Finch Mathematical Constants
95 Y. Jabri The Mountain Pass Theorem
96 G. Gasper, and M. Rahman Basic Hypergeometric Series, 2ed
97 M. C. Pedicchio and W. Tholen Categorical Foundations
99 T. Mora Solving Polynomial Equations Systems II
iii
From the Illustrated London News, 15 September 1883
Topics in Algebraic Graph Theory
Edited by
LOWELL W. BEINEKE
Indiana University-Purdue University
Fort Wayne
ROBIN J. WILSON
The Open University
Academic Consultant
PETER J. CAMERON
Queen Mary,
University of London
v
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, So Paulo, Delhi
Cambridge University Press
32 Avenue of the Americas, New York, NY 10013-2473, USA
www.cambridge.org
Information on this title: www.cambridge.org/9780521801973
Cambridge University Press 2005
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.
First published 2005
Reprinted 2007
Printed in the United States of America
A catalog record for this publication is available from the British Library.
Library of Congress Cataloging in Publication Data
Topics in algebraic graph theory / edited by Lowell W. Beineke and Robin J. Wilson,
academic consultant Peter J. Cameron.
p. cm. (Encyclopedia of mathematics and its applications)
Includes bibliographical references and index.
ISBN 0-521-80197-4
1. Graph theory. I. Beineke, Lowell W. II. Wilson, Robin J. III. Series.
QA166.T64 2004
515.5 dc22 2004045915
ISBN 978-0-521-80197-3 hardback
Cambridge University Press has no responsibility for
the persistence or accuracy of URLs for external or
third-party Internet Web sites referred to in this publication
and does not guarantee that any content on such
Web sites is, or will remain, accurate or appropriate.

Contents
Preface page xi
Foreword by Peter J. Cameron xiii
Introduction 1
LOWELL BEINEKE, ROBIN WILSON AND PETER CAMERON
1. Graph theory 1
2. Linear algebra 10
3. Group theory 19
Eigenvalues of graphs 30
MICHAEL DOOB
1. Introduction 30
2. Some examples 31
3. A little matrix theory 33
4. Eigenvalues and walks 34
5. Eigenvalues and labellings of graphs 39
6. Lower bounds for the eigenvalues 43
7. Upper bounds for the eigenvalues 47
8. Other matrices related to graphs 50
9. Cospectral graphs 51
Graphs and matrices 56
RICHARD A. BRUALDI and BRYAN L. SHADER
1. Introduction 56
2. Some classical theorems 58
3. Digraphs 61
4. Biclique partitions of graphs 67
vii
1
2
viii Contents
5. Bipartite graphs 69
6. Permanents 72
7. Converting the permanent into the determinant 75
8. Chordal graphs and perfect Gaussian
elimination 79
9. Ranking players in tournaments 82
3 Spectral graph theory 88
DRAGO
S CVETKOVI
C and PETER ROWLINSON

1. Introduction 88
2. Angles 89
3. Star sets and star partitions 94
4. Star complements 96
5. Exceptional graphs 99
6. Reconstructing the characteristic polynomial 101
7. Non-complete extended p-sums of graphs 104
8. Integral graphs 107
4 Graph Laplacians 113
BOJAN MOHAR
1. Introduction 113
2. The Laplacian of a graph 115
3. Laplace eigenvalues 117
4. Eigenvalues and vertex partitions of graphs 122
5. The max-cut problem and semi-denite
programming 125
6. Isoperimetric inequalities 127
7. The travelling salesman problem 129
8. Random walks on graphs 130
5 Automorphisms of graphs 137
PETER J. CAMERON
1. Graph automorphisms 137
2. Algorithmic aspects 139
3. Automorphisms of typical graphs 140
4. Permutation groups 141
5. Abstract groups 142
6. Cayley graphs 144
7. Vertex-transitive graphs 145
Contents ix
8. Higher symmetry 148
9. Innite graphs 149
10. Graph homomorphisms 152
Cayley graphs 156
BRIAN ALSPACH
1. Introduction 156
2. Recognition 157
3. Special examples 159
4. Prevalence 160
5. Isomorphism 164
6. Enumeration 167
7. Automorphisms 168
8. Subgraphs 169
9. Hamiltonicity 171
10. Factorization 173
11. Embeddings 174
12. Applications 175
Finite symmetric graphs 179
CHERYL E. PRAEGER
1. Introduction 179
2. s-arc transitive graphs 181
3. Group-theoretic constructions 182
4. Quotient graphs and primitivity 186
5. Distance-transitive graphs 187
6. Local characterizations 189
7. Normal quotients 192
8. Finding automorphism groups 196
9. A geometric approach 198
10. Related families of graphs 199
Strongly regular graphs 203
PETER J. CAMERON
1. An example 203
2. Regularity conditions 205
3. Parameter conditions 206
4. Geometric graphs 208
5. Eigenvalues and their geometry 212
9
7
8
x Contents
6. Rank 3 graphs 214
7. Related classes of graphs 217
9 Distance-transitive graphs 222
ARJEH M. COHEN
1. Introduction 222
2. Distance-transitivity 223
3. Graphs from groups 226
4. Combinatorial properties 230
5. Imprimitivity 233
6. Bounds 235
7. Finite simple groups 236
8. The rst step 238
9. The afne case 240
10. The simple socle case 245
10 Computing with graphs and groups 250
LEONARD H. SOICHER
1. Introduction 250
2. Permutation group algorithms 251
3. Storing and accessing a G-graph 253
4. Constructing G-graphs 254
5. G-breadth-rst search in a G-graph 255
6. Automorphism groups and graph isomorphism 257
7. Computing with vertex-transitive graphs 259
8. Coset enumeration 261
9. Coset enumeration for symmetric graphs 262
Notes on contributors 267
Index of denitions 271
Preface
The eldof graphtheoryhas undergone tremendous growthduringthe past century.
As recently as fty years ago, the graph theory community had few members
and most were in Europe and North America; today there are hundreds of graph
theorists and they span the globe. By the mid-1970s, the eld had reached the
point where we perceived the need for a collection of surveys of the areas of graph
theory: the result was our three-volume series Selected Topics in Graph Theory,
comprising articles written by distinguished experts in a common style. During
the past quarter-century, the transformation of the subject has continued, with
individual areas (such as algebraic graph theory) expanding to the point of having
important sub-branches themselves. This inspired us to conceive of a new series
of books, each a collection of articles within a particular area written by experts
within that area. This is the rst of these books.
One innovative feature of this volume is the engagement of an academic consul-
tant (Peter Cameron) to advise us on topics to be included and authors to be invited.
We believe that this has been successful, the result being chapters covering the full
range of areas within algebraic graph theory written by authors from around the
world. Another important feature is that we have imposed uniform terminology
and notation throughout, as far as possible, in the belief that this will aid readers
in going from one chapter to another. For a similar reason we have not attempted
to remove a small amount of overlap between the various chapters.
We hope that these features will make the book easier to use in an advanced
course or seminar. We heartily thank the authors for cooperating on this, even
though it sometimes required their abandoning some of their favourite conven-
tions for example, most graph theorists use G to denote a graph, whereas for
algebraic graph theorists G usually denotes a group: the graph theorists won on
this one. We also asked our contributors to undergo the ordeal of having their early
versions subjected to detailed critical reading. We believe that the nal product is
thereby signicantly better than it might otherwise have been, simply a collection
xi
xii Preface
of individually authored chapters. We want to express our sincere appreciation to
all of our contributors for their cooperation.
We extend special thanks to Peter Cameron for his willingness to share his
expertise as academic consultant his advice has been invaluable. We are also
grateful to Cambridge University Press for publishing this work; in particular, we
thank Roger Astley for his advice, support and cooperation. Finally, we extend
our appreciation to several universities for the different ways in which they have
assisted with this endeavour: Purdue University awarded the rst editor (LWB)
sabbatical leave during which he was a guest of the Mathematical Institute at
Oxford University, while the second editor (RJW) has had the cooperation of the
Open University and Keble College, Oxford.
LWB, RJW
February 2004
Foreword
The topic of this volume is the connection between graph theory and algebra
more specically, how algebra helps in the study of graphs, and how graph theory
repays the debt.
There are two main connections between graph theory and algebra. These arise
from two algebraic objects associated with a graph: its adjacency matrix and its
automorphism group. Of course, they are related: the automorphism group can
be regarded as the collection of all permutation matrices that commute with the
adjacency matrix. Nevertheless, the two connections involve different algebraic
techniques linear algebra and group theory, respectively and have a somewhat
different avour.
The introductory chapter presents the basic background material on graphs,
linear algebra and groups that is used in the rest of the book.
Turning rst to linear algebra, the rst four chapters survey different aspects of
the interaction between graphs and matrices. The adjacency matrix of a graph is a
real symmetric matrix, and hence is diagonalizable: the eigenvalues of this matrix
are closely connected with a variety of graph parameters. For directed graphs,
the adjacency matrix is not necessarily diagonalizable, and things are much less
straightforward. Moreover, in both cases the adjacency matrix is non-negative,
and so the theory of such matrices developed by Perron and Frobenius can be
applied.
The rst chapter describes the basic results, includingconnections withcounting
closed walks and with clique and colouring problems, while the second gives a
number of rather surprising applications. The third chapter looks in much more
detail at the geometry of the eigenvectors of the adjacency matrix and surveys the
theory of graph angles and star partitions which has been developed largely by the
authors of this chapter.
xiii
xiv Foreword
For some questions about graphs, especiallythose connectedwithrandomwalks
and of approximating manifolds by graphs, a different matrix is more appropriate:
this is the Laplacian matrix, the matrix D A, where Ais the adjacency matrix and
D is the diagonal matrix whose diagonal entries are the vertex-degrees. This matrix
is positive semi-denite with a zero eigenvalue; the smallest non-zero eigenvalue
gives crucial information about the behaviour of a random walk on the graph. This
is discussed in the fourth chapter.
In Chapter 5 the focus switches to automorphisms of graphs. Almost all graphs
admit no non-trivial automorphisms; but often it is the case that graphs with sym-
metry have other important properties. Such graphs as the Petersen graph and the
skeletons of regular polyhedra bear witness to this. Chapter 5 presents the basics
about automorphisms.
Typically we nd that the more symmetry we require of a graph, the more
group theory contributes to the analysis of the graph. A rst level of symmetry
we might require is vertex-transitivity, the property that the automorphism group
acts transitively on vertices. Among such graphs, the most important and widely
studied are the Cayley graphs. A Cayley graph of a group has the group elements
as vertices: the graph is obtained by specifying the neighbours of the identity
element and obtaining the remaining edges by translation. Cayley graphs and
their properties are described in Chapter 6. A stronger symmetry requirement is
that the automorphismgroupacts transitivelyonthe set of adjacent pairs of vertices.
Such graphs are called symmetric, and they form the subject of Chapter 7.
The two algebraic strands, linear algebra and group theory, meet in the study
of graphs satisfying higher regularity conditions: distance-regular graphs (and
their special case, strongly regular graphs) and distance-transitive graphs. A graph
is strongly regular if it is regular and if the number of common neighbours of
two vertices v and w depends only on whether v and w are adjacent. It turns out
that a regular connected graph that is not complete is strongly regular if and only
if its adjacency matrix has just three distinct eigenvalues (the smallest possible
number). Furthermore, if the automorphism group of a graph acts transitively on
the sets of ordered pairs of vertices and of non-adjacent vertices, then the graph is
strongly regular. Such graphs are the subject of Chapter 8.
This last conditionleads naturallytothe denitionof a distance-transitive graph,
one where the automorphism group acts transitively on the set of ordered pairs of
vertices at each xed distance up to the diameter of the graph. This is a very strong
symmetry condition, and the goal of researchers in the area is the determination
of all such graphs. The current state of this research programme is described in
Chapter 9. The arguments here are very group-theoretic, using (for example) the
classication of the nite simple groups.
Foreword xv
If a graph is symmetric, then it is possible to exploit the symmetry to simplify
searching for structures within the graph. Moreover, computational tools in group
theory have reached a high level of sophistication. The nal chapter discusses these
matters with a detailed look at how such computations can be done.
PETER J. CAMERON
Introduction
LOWELL BEINEKE, ROBIN WILSON and PETER CAMERON
1. Graph theory
2. Linear algebra
3. Group theory
References
This introductory chapter is divided into three parts. The rst presents the
basic ideas of graph theory. The second concerns linear algebra (for Chapters
14), while the third concerns group theory (for Chapters 510).
1. Graph theory
This section presents the basic denitions, terminology and notations of graph
theory, along with some fundamental results. Further information can be found in
the many standard books on the subject for example, West [4] or (for a simpler
treatment) Wilson [5].
Graphs
A graph G is a pair of sets (V, E), where V is a nite non-empty set of elements
called vertices, and E is a set of unordered pairs of distinct vertices called edges.
The sets V and E are the vertex-set and the edge-set of G, and are often denoted
by V(G) and E(G), respectively. An example of a graph is shown in Fig. 1.
The number of vertices in a graph is the order of the graph; usually it is denoted
by n and the number of edges by m. Standard notation for the vertex-set is V =
{v
1
, v
2
, . . . , v
n
] and for the edge-set is E = {e
1
, e
2
, . . . , e
m
]. Arbitrary vertices are
frequently represented by u, v, w, . . . and edges by e, f, . . . .
1
2 Lowell Beineke, Robin Wilson and Peter Cameron
5
G:
V = {
1
,
2
,
3
,
4
,
5
}
E = {
1
2
,
1
4
,
2
3
,
2
4
,
3
4
,
4
5
}
Fig. 1.
Variations of graphs
By denition, our graphs are simple, meaning that two vertices are connected by
at most one edge. If several edges, called multiple edges, are allowed between
two vertices, we have a multigraph. Sometimes, loops edges joining vertices
to themselves are also permitted. In a weighted graph, the edges are assigned
numerical values called weights. Finally, if the vertex-set is allowed to be innite,
then G is an innite graph.
Perhaps the most important variation is that of directed graphs; these are dis-
cussed at the end of this section.
Adjacency and degrees
For convenience, the edge {v, w] is commonly written as vw. We say that this edge
joins v and w and that it is incident with v and w. In this case, v and w are adjacent
vertices, or neighbours. The set of neighbours of a vertex v is its neighbourhood
N(v). Two edges are adjacent edges if they have a vertex in common.
The number of neighbours of a vertex v is called its degree, denoted by deg v.
Observe that the sum of the degrees in a graph is twice the number of edges. If
all the degrees of G are equal, then G is regular, or is k-regular if that common
degree is k. The maximum degree in a graph is often denoted by .
Walks
A walk in a graph is a sequence of vertices and edges v
0
, e
1
, v
1
, . . . , e
k
, v
k
, in
which each edge e
i
= v
i 1
v
i
. This walk goes from v
0
to v
k
or connects v
0
and v
k
,
and is called a v
0
-v
k
walk. It is frequently shortened to v
0
v
1
. . . v
k
, since the edges
may be inferred from this. Its length is k, the number of occurrences of edges. If
v
k
= v
0
, the walk is closed.
Some important types of walk are the following:
r
a path is a walk in which no vertex is repeated;
r
a trail is a walk in which no edge is repeated;
r
a cycle is a non-trivial closed trail in which no vertex is repeated.
Introduction 3
Distance
In a connected graph, the distance between two vertices v and w is the minimum
length of a path from v to w, and is denoted by d(v, w). It is easy to see that
distance satises the properties of a metric: for all vertices u, v and w,
r
d(v, w) 0, with equality if and only if v = w;
r
d(v, w) = d(w, v);
r
d(u, w) d(u, v) d(v, w)
The diameter of a graph G is the maximum distance between two vertices
of G. If G has cycles, the girth of G is the length of a shortest cycle, and the
circumference is the length of a longest cycle.
Subgraphs
If G and H are graphs with V(H) V(G) and E(H) E(G), then H is a sub-
graph of G. If, moreover, V(H) = V(G), then H is a spanning subgraph. The
subgraph induced by a non-empty set S of vertices in G is that subgraph H with
vertex-set S whose edge-set consists of those edges of G that join two vertices in
S; it is denoted by S) or G[S]. A subgraph H of G is induced if H = V(H)). In
Fig. 2, H
1
is a spanning subgraph of G, and H
2
is an induced subgraph.
Given a graph G, the deletion of a vertex v results in the subgraph obtained by
excluding v and all edges incident with it. It is denoted by G v and is the subgraph
induced by V {v]. More generally, if S V, we write G S for the graph
obtained from G by deleting all of the vertices of S; that is, G S = V S).
The deletion of an edge e results in the subgraph G e obtained by excluding e
from E; for F E, G F denotes the spanning subgraph with edge-set E F.
Connectedness and connectivity
A graph G is connected if there is a path connecting each pair of vertices. A
(connected) component of G is a maximal connected subgraph of G.
A vertex v of a graph G is a cut-vertex if G v has more components than G.
A connected graph with no cut-vertices is 2-connected or non-separable. The
following statements are equivalent for a graph G with at least three vertices:
G: H
1
: H
2
:
graph spanning subgraph induced subgraph
Fig. 2.
r
G is non-separable;
r
every pair of vertices lie on a cycle;
r
every vertex and edge lie on a cycle;
r
every pair of edges lie on a cycle;
r
for any three vertices u, v, and w, there is a v-w path containing u;
r
for any three vertices u, v, and w, there is a v-w path not containing u;
r
for any two vertices v and w and any edge e, there is a v-w path containing e.
More generally, a graph G is k-connected if there is no set S with fewer than k
vertices for which G S is a connected non-trivial graph. Menger characterized
such graphs.
Mengers theoremA graph G is k-connected if and only if, for each pair of
vertices v and w, there is a set of k v-w paths that pairwise have only v and w in
common.
The connectivity (G) of a graph G is the maximum value of k for which G is
k-connected.
There are similar concepts and results for edges. A cut-edge (or bridge) is any
edge whose deletion produces one more component than before. A non-trivial
graph G is k-edge-connected if the result of removing fewer than k edges is always
connected, and the edge-connectivity (G) is the maximum value of k for which
G is k-edge-connected. We note that Mengers theorem also has an edge version.
Bipartite graphs
If the vertices of a graph G can be partitioned into two non-empty sets so that no
edge joins two vertices in the same set, then G is bipartite. The two sets are called
partite sets, and if they have orders r and s, G may be called an r s bipartite
graph. The most important property of bipartite graphs is that they are the graphs
that contain no cycles of odd length.
Trees
A tree is a connected graph that has no cycles. They have been characterized in
many ways, a few of which we give here. For a graph G of order n:
r
G is connected and has no cycles;
r
G is connected and has n 1 edges;
r
G has no cycles and has n 1 edges.
Any graph without cycles is a forest; note that each component of a forest is a tree.
Introduction 5
Special graphs
We now introduce some individual types of graphs:
r
the complete graph K
n
has n vertices, each of which is adjacent to all of the
others;
r
the null graph N
n
has n vertices and no edges;
r
the path graph P
n
consists of the vertices and edges of a path of length n 1;
r
the cycle graph C
n
consists of the vertices and edges of a cycle of length n;
r
the complete bipartite graph K
r,s
is the r s bipartite graph in which each
vertex is adjacent to all those in the other partite set;
r
in the complete k-partite graph, K
r
1
,r
2
,...,r
n
the vertices are in k sets (having
orders r
1
, r
2
, . . . , r
k
) and each vertex is adjacent to all the others, except those
in the same set. If the k sets all have order r, the graph is denoted by K
k(r)
. The
graph K
k(2)
is sometimes called the k-dimensional octahedral graph or cocktail
party graph, also denoted by CP(k); K
3(2)
is the graph of an octahedron.
r
the d-dimensional cube (or d-cube) Q
d
is the graph whose vertices can be
labelled with the 2
d
binary d-tuples, in such a way that two vertices are
adjacent when their labels differ in exactly one position. It is regular of degree
d, and is isomorphic to the lattice of subgraphs of a set of d elements.
Examples of these graphs are given in Fig. 3.
Operations on graphs
There are several ways to get new graphs from old. We list some of the most
important here.
K
5
:
K
3,3
:
N
5
:
K
3(2)
:
P
5
:
C
5
:
Q
3
:
Fig. 3.
r
The complement G of a graph G has the same vertices as G, but two vertices
are adjacent in G if and only if they are not adjacent in G.
For the other operations, we assume that G and H are graphs with disjoint vertex-
sets, V(G) = {v
1
, v
2
, . . . , v
n
] and V(H) = {w
1
, w
2
, . . . , w
t
]:
r
the union G H has vertex-set V(G) V(H) and edge-set E(G) E(H).
The union of k graphs isomorphic to G is denoted by kG.
r
the join G H is obtained from G H by adding all of the edges from
vertices in G to those in H.
r
the (Cartesian) product G H or G H has vertex-set V(G) V(H), and
(v
i
, w
j
) is adjacent to (v
h
, w
k
) if either (a) v
i
is adjacent to v
h
in G and
w
j
= w
k
, or (b) v
i
= v
h
and w
j
is adjacent to w
k
in H. In less formal terms,
G H can be obtained by taking n copies of H and joining corresponding
vertices in different copies whenever there is an edge in G. Note that, for
d-cubes, Q
d1
= K
2
Q
d
(with Q
1
= K
2
).
Examples of these binary operations are given in Fig. 4.
There are two basic operations involving an edge of a graph. The insertion of
a vertex into an edge e means that the edge e = vw is replaced by a new vertex
u and the two edges vu and uw. Two graphs are homeomorphic if each can be
obtained from a third graph by a sequence of vertex insertions. The contraction of
the edge vw means that v and w are replaced by a new vertex u that is adjacent
to the other neighbours of v and w. If a graph H can be obtained from G by
a sequence of edge contractions and the deletion of isolated vertices, then G is
said to be contractible to H. Finally, H is a minor of G if it can be obtained
from G by a sequence of edge-deletions and edge-contractions and the removal
G:
H:
G + H:
G H:
G H:
Fig. 4.
Introduction 7
insertion contraction
w
u
u
e
w
Fig. 5.
of isolated vertices. The operations of insertion and contraction are illustrated in
Fig. 5.
Traversability
A connected graph G is Eulerian if it has a closed trail containing all of the edges
of G; such a trail is called an Eulerian trail. The following are equivalent for a
connected graph G:
r
G is Eulerian;
r
the degree of each vertex of G is even;
r
the edge-set of G can be partitioned into cycles.
A graph G is Hamiltonian if it has a spanning cycle, and traceable if it has a
spanning path. No good characterizations of these graphs are known.
Planarity
A planar graph is one that can be embedded in the plane in such a way that no
two edges meet except at a vertex incident with both. If a graph G is embedded in
this way, then the points of the plane not on G are partitioned into open sets called
faces or regions. Euler discovered the basic relationship between the numbers of
vertices, edges and faces.
Eulers polyhedron formula Let G be a connected graph embedded in the plane
with n vertices, m edges and f faces. Then n m f = 2.
It follows from this result that a planar graph with n vertices (n 3) has at most
3(n 2) edges, and at most 2(n 2) edges if it is bipartite. From this it follows
that the two graphs K
5
and K
3,3
are non-planar. Kuratowski proved that these two
graphs are the only barriers to planarity.
Kuratowskis theoremThe following statements are equivalent for a graph G:
r
G is planar;
r
G has no subgraph that is homeomorphic to K
5
or K
3,3
;
r
G has no subgraph that is contractible to K
5
or K
3,3
.
Graph colourings
Agraph G is k-colourable if, froma set of k colours, it is possible to assign a colour
to each vertex in such a way that adjacent vertices always have different colours.
The chromatic number (G) is the least value of k for which G is k-chromatic.
It is easy to see that a graph is 2-colourable if and only if it is bipartite, but there
is no good way to determine which graphs are k-colourable for k 3. Brookss
theorem provides one of the best-known bounds on the chromatic number of a
graph.
Brookss theoremIf G is a graph with maximum degree that is neither an odd
cycle nor a complete graph, then (G) .
There are similar concepts for colouring edges. Agraph G is k-edge-colourable
if, from a set of k colours, it is possible to assign a colour to each edge in such a
way that adjacent edges always have different colours. The edge-chromatic number
/
(G) is the least k for which G is k-edge-colourable. Vizing proved that the range
of values of
/
(G) is very limited.
Vizings theoremIf G is a graph with maximum degree , then

/
(G) 1.
Line graphs
The line graph L(G) of a graph G has the edges of G as its vertices, with two of
these vertices adjacent if and only if the corresponding edges are adjacent in G.
An example is given in Fig. 6.
A graph is a line graph if and only if its edges can be partitioned into complete
subgraphs in such a way that no vertex is in more than two of these subgraphs.
Introduction 9
G: L(G):
Fig. 6.
4
D:
V = {
1
,
2
,
3
,
4
}
E = {
1
2
,
1
4
,
2
1
,
3
2
,
3
4
}.
Fig. 7.
Line graphs are also characterized by the property of having none of nine particular
graphs as a forbidden subgraph.
Directed graphs
Digraphs are directed analogues of graphs, and thus have many similarities, as
well as some important differences.
A digraph (or directed graph) D is a pair of sets (V, E) where V is a nite
non-empty set of elements called vertices, and E is a set of ordered pairs of distinct
elements of V called arcs or directed edges. Note that the elements of E are now
ordered, which gives each of them a direction. An example of a digraph is given
in Fig. 7.
Because of the similarities between graphs and digraphs, we mention only the
main differences here and do not redene those concepts that carry over easily.
An arc (v, w) of a digraph may be written as

vw, and is said to go from v to w,
or to go out of v and go into w.
Walks, paths, trails and cycles are understood to be directed, unless otherwise
indicated.
The out-degree d
(v) of a vertex v in a digraph is the number of arcs that go

out of it, and the in-degree d
(v) is the number of arcs that go into it.

Adigraph Dis strongly connected, or strong, if there is a pathfromeachvertexto
each of the others. Astrong component is a maximal strongly connected subgraph.
Connectivity and edge-connectivity are dened in terms of strong connectedness.
A tournament is a digraph in which every pair of vertices are joined by exactly
one arc. One interesting aspect of tournaments is their Hamiltonian properties:
r
every tournament has a spanning path;
r
a tournament has a Hamiltonian cycle if and only if it is strong.
2. Linear algebra
In this section we present the main results on vector spaces and matrices that are
used in Chapters 14. For further details, see [3].
The space R
n
The real n-dimensional space R
n
consists of all n-tuples of real numbers x =
(x
1
, x
2
, . . . , x
n
); in particular, the plane R
2
consists of all pairs (x
1
, x
2
), and three-
dimensional space R
3
consists of all triples (x
1
, x
2
, x
3
). The elements x are vectors,
and the numbers x
i
are the coordinates or components of x.
When x = (x
1
, x
2
, . . . , x
n
) and y = (y
1
, y
2
, . . . , y
n
) are vectors in R
n
, we can
form their sum x y = (x
1
y
1
, x
2
y
2
, . . . , x
n
y
n
), and if is a scalar (real
number), we can form the scalar multiple x = (x
1
, x
2
, . . . , x
n
).
The zero vector is the vector 0 = (0, 0, . . . , 0), and the additive inverse of
x = (x
1
, x
2
, . . . , x
n
) is the vector x = (x
1
, x
2
, . . . , x
n
).
We can similarly dene the complex n-dimensional space C
n
, in which the
vectors are all n-tuples of complex numbers z = (z
1
, z
2
, . . . , z
n
); in this case, we
take the multiplying scalars to be complex numbers.
Metric properties
When x = (x
1
, x
2
, . . . , x
n
) and y = (y
1
, y
2
, . . . , y
n
) are vectors in R
n
, their dot
product is the scalar x y = x
1
y
1
x
2
y
2
x
n
y
n
. The dot product is some-
times called the inner product and denoted by x, y).
The length or norm |x| of a vector x = (x
1
, x
2
, . . . , x
n
) is
(x x)
1/2
=
_
x
2
1
x
2
2
x
n
n
_
1/2
.
A unit vector is a vector u for which |u| = 1, and for any non-zero vector x, the
vector x/|x| is a unit vector.
When x = (x
1
, x
2
, . . . , x
n
) and y = (y
1
, y
2
, . . . , y
n
), the distance between x
and y is d(x, y) = |x y|. The distance function d satises the usual properties
of a metric: for any x, y, z R
n
,
r
d(x, y) 0, and d(x, y) = 0 if and only if x = y;
r
d(x, y) = d(y, x);
r
d(x, z) d(x, y) d(y, z) (triangle inequality).
The following result is usually called the Cauchy-Schwarz inequality:
Cauchy-Schwarz inequality For any x, y R
n
, [x y[ |x| |y|.
Introduction 11
We dene the angle between the non-zero vectors x and y by
cos = x y/|x| |y|.
Two vectors x and y are orthogonal if the angle between them is 0 that is, if
x y = 0. In this case, we have the following celebrated result.
Pythagorass theorem If x and y are orthogonal, then |x y|
2
= |x|
2
|y|
2
.
An orthogonal set of vectors is a set of vectors each pair of which is orthogonal.
An orthonormal set is an orthogonal set in which each vector has length 1.
In a complex space C
n
most of the above concepts are dened as above. One
exception is that the dot product of two complex vectors z = (z
1
, z
2
, . . . , z
n
)
and w = (w
1
, w
2
, . . . , w
n
) is now dened by z w = z
1
w
1
z
2
w
2
z
n
w
n
,
where w is the complex conjugate of w.
Vector spaces
A real vector space V is a set of elements, called vectors, with rules of addition
and scalar multiplication that satisfy the following conditions:
Addition
A1: For all x, y V, x y V;
A2: For all x, y, z V, (x y) z = x (y z);
A3: There is an element 0 V satisfying x 0 = x, for all x V;
A4: For each x V, there is an element x satisfying x (x) = 0;
A5: For all x, y V, x y = y x.
Scalar multiplication
M1: For all x V and R, x V;
M2: For all x V, 1x = x;
M3: For all , R, (x) = ()x;
Distributive laws
D1: For all , R and x V, ( )x = x x;
D2: For all R and x, y V, (x y) = x y.
Examples of real vector spaces are R
n
, C
n
, the set of all real polynomials, the
set of all real innite sequences, and the set of all functions f : R R, each with
the appropriate denitions of addition and scalar multiplication.
Complex vector spaces are dened similarly, except that the scalars are elements
of C, rather than R. More generally, the scalars can come fromany eld, such as the
set Q of rational numbers, the integers Z
p
modulo p, where p is a prime number,
or the nite eld F
q
, where q is a power of a prime.
Subspaces
A non-empty subset W of a vector space V is a subspace of V if W is itself a
vector space with respect to the operations of addition and scalar multiplication in
V. For example, the subspaces of R
3
are {0], the lines and planes through 0, and
R
3
itself.
When X and Y are subspaces of a vector space V, their intersection X Y is
also a subspace of V, as is their sum X Y = {x y : x X, y Y].
When V = X Y and X Y = {0], we call V the direct sum of X and Y, and
write V = X Y.
Bases
Let S = {x
1
, x
2
, . . . , x
r
] be a set of vectors in a vector space V. Then any vector
of the form
1
x
1

2
x
2

r
x
r
,
where
1
,
2
, . . . ,
r
are scalars, is a linear combination of x
1
, x
2
, . . . , x
r
. The set
of all linear combinations of x
1
, x
2
, . . . , x
r
is a subspace of V called the span of
S, denoted by S) or x
1
, x
2
, . . . , x
r
). When S) = V, the set S spans V, or is a
spanning set for V.
The set S = {x
1
, x
2
, . . . , x
r
] is linearly dependent if one of the vectors x
i
is a
linear combination of the others in this case, there are scalars
1
,
2
, . . . ,
r
, not
all zero, for which
1
x
1

2
x
2

r
x
r
= 0.
The set S is linearly independent if it is not linearly dependent that is,
1
x
1

2
x
2

r
x
r
= 0
holds only when
1
=
2
= =
r
= 0.
A basis B is a linearly independent spanning set for V. In this case, each vector
x of V can be written as a linear combination of the vectors in B in exactly one
way; for example, the standard basis for R
3
is {(1, 0, 0), (0, 1, 0), (0, 0, 1)] and a
basis for the set of all real polynomials is {1, x, x
2
, . . .].
Dimension
A vector space V with a nite basis is nite-dimensional. In this situation, any
two bases for V have the same number of elements. This number is the dimension
Introduction 13
of V, denoted by dim V; for example, R
3
has dimension 3. The dimension of a
subspace of V is dened similarly.
When X and Y are subspaces of V, we have the dimension theorem:
dim(X Y) = dim X dimY dim(X Y).
When X Y = {0], this becomes
dim(X Y) = dim X dimY.
Euclidean spaces
Let V be a real vector space, and suppose that with each pair of vectors x and y
in V is associated a scalar x, y). This is an inner product on V if it satises the
following properties: for any x, y, z V,
r
x, x) 0, and x, x) = 0 if and only if x = 0;
r
x, y) = y, x);
r
x y, z) = x, z) y, z).
The vector space V, together with this inner product, is called a real in-
ner product space, or Euclidean space. Examples of Euclidean spaces are R
3
with the dot product as inner product, and the space V of real-valued contin-
uous functions on the interval [1, 1] with the inner product dened for f, g
in V by f, g) =
_
1
1
f(t )g(t ) dt . Analogously to the dot product, we can dene
the metrical notions of length, distance and angle in any Euclidean space, and
we can derive analogues of the Cauchy-Schwarz inequality and Pythagorass
theorem.
An orthogonal basis for a Euclidean space is a basis in which any two distinct
basis vectors are orthogonal. If, further, each basis vector has length 1, then the
basis is anorthonormal basis. If V is a Euclideanspace, the orthogonal complement
W
of a subspace W is the set of all vectors in V that are orthogonal to all vectors
in W that is,
W
= {v V : v, w) = 0 for all w W].

Linear transformations
When V and W are real vector spaces, a function T : V W is a linear transfor-
mation if, for all v
1
, v
2
V and , R,
T(v
1
v
2
) = T(v
1
) T(v
2
).
If V = W, then T is sometimes called a linear operator on V.
The linear transformation T is onto, or surjective, when T(V) = W, and is
one-one, or injective, if T(v
1
) = T(v
2
) only when v
1
= v
2
.
The image of T is the subspace of W dened by
im(T) = {w W : w = T(v), for some v V];
note that T is onto if and only if im(T) = W.
The kernel, or null space, of T is the subspace of V dened by
ker(T) = {v V : T(v) = 0
W
];
note that T is one-one if and only if ker(T) = {0
V
].
Dening the rank and nullity of T by
rank(T) = dimim(T) and nullity(T) = dimker(T),
we obtain the rank-nullity formula:
rank(T) nullity(T) = dimV.
Algebra of linear transformations
When S : U V and T : V W are linear transformations, we can form their
composition T

S : U W, dened by
(T

S)(u) = T(S(u)), for all u U.
The composition of linear transformations is associative.
The linear transformation T : V W is invertible, or non-singular, if there
is a linear transformation T
1
, called the inverse of T, for which T
1
T is the
identity transformation on V and T

T
1
is the identity transformation on W.
Note that a linear transformation is invertible if and only if it is one-one and onto.
The matrix of a linear transformation
Let T : V W be a linear transformation, let {e
1
, e
2
, . . . , e
n
] be a basis for V
and let {f
1
, f
2
, . . . , f
m
] be a basis for W. For each i = 1, 2, . . . , n, we can write
T(e
i
) = a
1i
f
1
a
2i
f
2
a
mi
f
m
,
for some scalars a
1i
, a
2i
, . . . , a
mi
. The rectangular array of scalars
A =
_
_
_
_
a
11
a
12
a
1n
a
21
a
22
a
2n

a
m1
a
m2
a
mn
_
_
_
_
Introduction 15
is the matrix of T associated with the given bases. The scalar a
i j
is the ij-entry of
A, and we abbreviate the above array by writing A = (a
i j
).
Since the matrix A has m rows and n columns, we call A an m n matrix; a
matrix for which m = n is a square matrix of order n. The diagonal of a square
matrix A = (a
i j
) consists of the entries a
i i
down the leading (top-left to bottom-
right) diagonal.
Operations on matrices
Given a matrix A = (a
i j
) and a scalar , we can dene the scalar multiple A =
(a
i j
); note that 1A = A.
Given two m n matrices A = (a
i j
) and B = (b
i j
), we dene their sumA B
to be the matrix C = (c
i j
), where c
i j
= a
i j
b
i j
; matrix addition is commutative
and associative.
Given an m k matrix A = (a
i j
) and a k n matrix B = (b
i j
), we dene
their product AB to be the m n matrix C = (c
i j
), where c
i j
=
k
a
i k
b
kj
; matrix
multiplication is associative, but not commutative in general. The matrix product
AA is written A
2
, with similar notation for higher powers of A.
Given two matrices A = (a
i j
) and B = (b
i j
) of the same size, we dene their
Kronecker product A B to be the matrix C = (c
i j
), where c
i j
= a
i j
b
i j
.
Given an m n matrix A = (a
i j
), we can interchange the rows and columns to
form the n m matrix A
T
= (a
j i
) called its transpose. It follows that
(A
T
)
T
= A and (A)
T
= (A)
T
, for any scalar ,
and that, for matrices A and B of appropriate sizes,
(A B)
T
= A
T
B
T
and (AB)
T
= B
T
A
T
.
The trace tr(A) of a square matrix A is the sum of the diagonal entries of A; for
matrices A and B of appropriate sizes,
tr(A B) = tr(A) tr(B) and tr(AB) = tr(BA).
Types of matrices
Azero matrix 0 is a matrix in which each entry is 0; for matrices Aand 0 of the same
size, A 0 = 0 A = A. An all-1 matrix J is a matrix in which each entry is 1.
Asquare matrix A = (a
i j
) is symmetric if A
T
= A(so a
i j
= a
j i
for all i and j ),
and anti-symmetric or skew-symmetric if A
T
= A(so a
i j
= a
j i
for all i and j ).
A diagonal matrix A is a square matrix in which every non-diagonal entry is
0. The identity matrix I is the diagonal matrix with 1s on its diagonal; for square
matrices A and I of the same order, AI = IA = A. A permutation matrix is a
matrix obtained from I by permuting the rows or columns.
An upper triangular matrix is a square matrix in which every entry below and
to the left of the diagonal is 0; a lower triangular matrix is dened similarly.
Ablock matrix is a matrix arranged in submatrices B
i j
called blocks, as follows:
_
_
_
_
B
11
B
12
B
1s
B
21
B
22
B
2s

B
r1
B
r2
B
rs
_
_
_
_
A circulant matrix A = (a
i j
) is an n n matrix in which each successive row
is obtained by moving the preceding row by one position to the right; thus, for
each i and j , a
i j
= a
i 1, j 1
, where the subscripts are taken modulo n.
A square matrix A is invertible if there is a matrix B for which AB = BA = I;
the matrix B is the inverse of A, denoted by A
1
. Note that (A
1
)
1
= A, and that,
for square matrices A and B of the same order, (AB)
1
= B
1
A
1
.
Determinants
Let A = (a
i j
) be an n n square matrix. The determinant of A, written det A or
[A[, is the sum
(sgn s) a
1s(1)
a
2s(2)
. . . a
ns(n)
,
taken over all permutations s of {1, 2, . . . , n], where sgn s is 1 if s is an even
permutation, and 1 if s is odd; for example, when n = 3, the permutations 123,
231 and 312 are even, and 132, 213 and 321 are odd, and so
_
_
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
_
_
= a
11
a
22
a
33
a
12
a
23
a
31
a
13
a
21
a
32
a
11
a
23
a
32
a
12
a
21
a
33
a
13
a
22
a
31
.
Note that, for any square matrices A and B,
r
det A = det A
T
;
r
det AB = det Adet B;
r
if det A ,= 0, then det A
1
= (det A)
1
.
Given a square matrix A, let the minor M
i j
be the determinant of the matrix
obtained from A by deleting the i th row and the j th column, and let the cofactor
A
i j
= (1)
i j
M
i j
. Then
det A =
i
a
i j
A
i j
=
j
a
i j
A
i j
.
Introduction 17
The matrix adj A = (A
i j
) is the adjoint matrix, and if det A ,= 0, then
A
1
= (det A)
1
(adj A).
Change of basis
Let T : V W be a linear transformation, let {e
1
, e
2
, . . . , e
m
] and {e
/
1
, e
/
2
, . . . ,
e
/
m
] be bases for V, and let {f
1
, f
2
, . . . , f
n
] and {f
1
/
, f
2
/
, . . . , f
n
/
] be bases for W. If
the matrix of T associated with the bases {e
1
, e
2
, . . . , e
m
] and {f
1
, f
2
, . . . , f
n
] is
A, then the matrix of T associated with the bases {e
/
1
, e
/
2
, . . . , e
/
m
] and {f
1
/
, f
2
/
, . . . ,
f
n
/
] is Q
1
AP, where P = ( p
i j
) and Q = (q
i j
) are the transition matrices dened
by e
/
j
=
i
p
i j
e
i
and f
j
/
=
i
q
i j
f
i
.
Now let T : V V be a linear transformation, and let {e
1
, e
2
, . . . , e
m
] and
{e
/
1
, e
/
2
, . . . , e
/
m
] be bases for both V and W. If the matrix of T associated with the
basis {e
1
, e
2
, . . . , e
m
] in V and W is A, then, by the above result, the matrix of T
associated with the basis {e
/
1
, e
/
2
, . . . , e
/
m
] in V and W is P
1
AP, where P = ( p
i j
) is
the transition matrix dened by e
/
j
=
i
p
i j
e
i
. Two matrices A and B are similar
if B = P
1
AP, for some invertible matrix P.
Eigenvalues and eigenvectors
If Ais a square matrix, then the polynomial det(I A) is the characteristic poly-
nomial of A and the equation det(I A) = 0 is its characteristic equation. The
roots of this equation are the eigenvalues of A; a repeated root is a multiple eigen-
value, and a non-repeated root is a simple eigenvalue. The sum of the eigenvalues
of A is equal to the trace of A. When A is a symmetric matrix, the eigenvalues of
A are all real. The set of eigenvalues of A is the spectrum of A.
Every square matrix satises a number of polynomial equations; for example, if
Ais ann n matrix, thenthe matrices I, A, A
2
, . . . , A
n
2
are linearlydependent and
so there is a polynomial equation connecting them. The unique monic polynomial
of lowest degree satised by A is the minimal polynomial of A. The Cayley-
Hamilton theorem states that every matrix satises its characteristic equation.
It follows that the minimum polynomial divides the characteristic polynomial;
moreover, the minimumpolynomial and the characteristic equation share the same
irreducible factors.
Let T : V V be a linear transformation. Regardless of the basis used, all
matrices associated with T have the same eigenvalues: these are the scalars for
which T(v) = v, for some non-zero vector v, and each such vector is an associated
eigenvector. For a given eigenvalue the set of all such eigenvectors, together with
0, is a subspace of V, called the eigenspace associated with .
Diagonalizing matrices
When v
1
, v
2
, . . . , v
r
are eigenvectors associated with distinct eigenvalues
1
,
2
, . . . ,
r
, the set {v
1
, v
2
, . . . , v
r
] is linearly independent. Eigenvectors associ-
ated with the same eigenvalue may be linearly dependent or independent. If we
can nd a basis of eigenvectors, and if P is the matrix with these eigenvectors as
columns, then P
1
AP is a diagonal matrix. Conversely, if A is any square matrix
and if we can nd a matrix P for which P
1
AP is a diagonal matrix, then there
is a basis of eigenvectors, and these eigenvectors form the columns of P; in this
case, we say that A is diagonalizable. Not all square matrices are diagonalizable:
for example, the matrix
_
1 1
0 1
_
is not diagonalizable, since every eigenvector is a multiple of (1 0) and so there
is no basis of eigenvectors.
Every symmetric matrix A has an orthonormal basis of eigenvectors, and so is
diagonalizable. Moreover, the corresponding transition matrix P is then an orthog-
onal matrix (P
1
= P
T
), the matrix P
T
AP is diagonal, and the matrix A is called
orthogonally diagonalizable.
Even when we cannot diagonalize a given matrix, we can always choose a basis
so that the resulting matrix can be decomposed into block diagonal form, with
square matrices B
1
, B
2
, . . . , B
k
arranged down the diagonal, and zeros elsewhere.
There are various existence and uniqueness theorems concerning such decompo-
sitions.
Quadratic forms
A quadratic form is an expression of the form q(x) = x
T
Ax, where x is a column
vector and A is a symmetric matrix; for example,
q(x) = 2x
2
1
3x
2
2
4x
2
3
x
1
x
2
6x
1
x
3
is a quadratic form corresponding to the symmetric matrix
A =
_
_
2
1
/
2
3
1
/
2
3 0
3 0 4
_
_
.
A quadratic form q is positive denite if q(x) > 0, and positive semidenite if
q(x) 0, for every non-zero vector x.
Introduction 19
3. Group theory
This section introduces some basic material about permutation groups and auto-
morphism groups of graphs and explains the notation used for groups. Further
information can be found in the books by Cameron [1] and Dixon and Mortimer
[2] and the many references cited therein.
Groups
A group is a set G with a binary operation satisfying the conditions:
r
for all g, h, k G, (g
h)
k = g
(h
k) (associative law);
r
there exists an element 1 G (the identity) such that 1
g = g
1 = g for all
g G;
r
for each g G, there is an element g
1
G (the inverse of g) such that
g
g
1
= g
1
g = 1.
We usually write g
h more briey as gh.

If in addition we have the condition:
r
for all g, h G, g
h = h
g (commutative law)
then the group G is Abelian (or commutative).
Groups are important here because the set of automorphisms of a graph (with
the operation of composition of mappings) is a group. In many cases, the group
encodes important informationabout the graph; andingeneral, the use of symmetry
can be used to do combinatorial searches in the graph more efciently.
Homomorphisms
Ahomomorphismfroma groupG
1
toa groupG
2
is a function: G
1
G
2
for which
(gh) = (g)(h), for all g, h G
1
. Note two things about this denition. First, in
the dening equation (gh) = (g)(h), the operation gh on the left is the group
operation in G
1
, while the operation (g)(h) on the right is the group operation in
G
2
. Second, it is common among algebraists to write homomorphisms on the right,
so instead of (g), we write g. The advantage of this is that the composition of
with a homomorphism : G
2
G
3
is now , where g() = (g).
An isomorphism fromG
1
to G
2
is a homomorphism that is one-to-one and onto.
If there is an isomorphismfromG
1
to G
2
, then the groups G
1
and G
2
are isomorphic.
Algebraists regard isomorphic groups as being essentially the same, and do not
carefully distinguish between them.
Subgroups
A subgroup of a group is a subset that forms a group in its own right with respect
to the same operation. If H is a subgroup of G (sometimes written H < G), then
there are two natural partitions of G:
r
the partition into sets Hx = {hx : h H], called right cosets of H;
r
the partition into sets xH = {xh : h H], called left cosets of H.
The number of right (or left) cosets is the index of H in G.
Let : G
1
G
2
be a homomorphism, and let 1 denote the identity element of
G
2
. Then the set
N = {g G
1
: g = 1]
is the kernel of . The kernel of a homomorphism is a subgroup of G
1
, and has the
additional property that
g
1
ng N, for all g G
1
, n N.
Such a subgroup is called a normal subgroup. It turns out that every normal sub-
group is the kernel of a homomorphism, and that a subgroup is normal if and only
if the partitions of the group into right and left cosets of the subgroup coincide (so
that we can simply speak of cosets).
Let N be a normal subgroup of G. Then we dene G/N to be the set of cosets
of N in G, and dene an operation on G/N by the rule
(Nx)(Ny) = Nxy.
It can be shown that this operation is indeed well dened (that is, independent of
the chosen representatives x and y of the cosets) and denes a group, which is
called the quotient group or factor group, also denoted by G/N.
Every group G has two trivial normal subgroups: the whole group G and the
identity {1]. A group with no other normal subgroups is simple; any other group
is composite.
Composition series
An important structural result about nite groups is the JordanH older theorem.
Jordan-H older theorem Let G be a nite group. Then the following hold:
(a) there is a chain
{1] = G
r
< G
r1
< < G
1
< G
0
= G
Introduction 21
of subgroups of G such that, for 1 i r, G
i
is a normal subgroup of G
i 1
and the factor group G
i 1
/G
i
is simple.
(b) for any two such chains, the multisets of isomorphism types of simple factor
groups are the same.
The series in (a) is a composition series for G, and the simple factor groups are the
composition factors.
This theorem indicates that the simple groups are the building blocks from
which arbitrary groups can be constructed. Thus nite group theory falls into two
parts, concerned with answering the questions
r
what are the nite simple groups?
r
how can they be put together to build arbitrary groups?
Finite simple groups
The rst question above has recently been completely solved. It is not possible
here to give any hint of the proof, and the result can be stated only in broad terms.
There are four classes of nite simple groups:
r
If p is a prime number, then the cyclic group Z
p
of order p (which can be
realized as the group of rotations of a regular p-gon) is simple; indeed, it has
no subgroups at all except for itself and the identity.
r
For n 2, the set of permutations of {1, 2, . . . , n] that have even parity (that is,
are products of even numbers of transpositions) forms a normal subgroup A
n
of S
n
with index 2, called the alternating group. For n 5, the alternating
group is simple.
r
There is a large collection of groups of Lie type, closely related to groups of
matrices over nite elds. The simplest to describe are the groups PSL
n
(q),
obtained from the group SL
n
(q) of all n n matrices over the nite eld F
q
of
q elements which have determinant 1. The set of scalar matrices of determinant
1 forms a normal subgroup Z of SL
n
(q), and the group PSL
n
(q) is the quotient
group SL
n
(q)/Z. The group PSL
n
(q) is simple for n 2 in all except two
cases, PSL
2
(2) and PSL
2
(3) (which happen to be isomorphic to S
3
and A
4
,
respectively). The other groups of this type are harder to describe.
r
There are just twenty-six further nite simple groups; these are the sporadic
groups.
The Classication of nite simple groups asserts the following.
Classication theorem Any nite simple group is cyclic of prime order, or an
alternating group, or a group of Lie type, or one of the twenty-six sporadic groups.
Permutation groups
A permutation of the set is a bijective mapping g: . We write the image
of the point v under the permutation g as vg, rather than g(v). The composition
g
1
g
2
of two permutations g
1
and g
2
is the permutation obtained by applying g
1
and then g
2
that is,
v(g
1
g
2
) = (vg
1
)g
2
for each v .
A permutation group on is a set G of permutations of satisfying the
following conditions:
r
G is closed under composition: if g
1
, g
2
G then g
1
g
2
G;
r
G contains the identity permutation 1, dened by v1 = v for v .
r
G is closed under inversion, where the inverse of g is the permutation g
1
dened by the rule that vg
1
= w if wg = v.
The degree of the permutation group G is the cardinality of the set .
The simplest example of a permutation group is the set of all permutations of
a set . This is the symmetric group, denoted by Sym(). More generally, an
action of G on is a homomorphism from G to Sym(). The image of the
homomorphism is then a permutation group. The action is faithful if its kernel is
{1] that is, if distinct group elements map to distinct permutations. If the action
is faithful, then G is isomorphic to a permutation group on .
Products
Let G
1
and G
2
be permutation groups on
1
and
2
respectively. We dene two
products:
The direct product G
1
G
2
acts on the disjoint union
1

2
. Each element of
the group is an ordered pair (g
1
, g
2
), with g
1
G
1
, g
2
G
2
, and acts by the rule
v(g
1
, g
2
) =
_
vg
1
if v
1
,
vg
2
if v
2
.
It also has an action as a permutation group on the Cartesian product
1

2
,
where
(v
1
, v
2
)(g
1
, g
2
) = (v
1
g
1
, v
2
g
2
), for v
1

1
, v
2

2
.
The wreath product G
1
: G
2
is a permutation group on
1

2
. Its elements are
compositions of two types of permutations:
Introduction 23
r
for each function f :
2
G
1
, the permutation given by
(v
1
, v
2
) f = (v
1
f (v
2
), v
2
);
r
for each element g G
2
, the permutation given by
(v
1
, v
2
)g = (v
1
, v
2
g).
It also has an action as a permutation group on the set of all functions from
2
to
1
. We do not dene this here, but note that it arises in connection with the
automorphism groups of the Hamming graphs in Chapter 7, Section 5.
The same (abstract) group may act as a permutation group on many different
sets. Two actions of G on sets
1
and
2
are isomorphic if there is a bijection f
from
1
to
2
that commutes with the action of G that is,
(vf )g = (vg) f, for all v
1
and g G.
More generally, the permutation groups G
1
on
1
and G
2
on
2
are equivalent if
there are a bijection f from
1
to
2
and an isomorphism from G
1
to G
2
such
that
(vf ) (g) = (vg) f, for all v
1
and g G
1
.
Note that, although isomorphic actions are equivalent, two actions of the same
group may be equivalent without being isomorphic.
Automorphism groups of graphs
Let G = (V, E) be a simple graph, possibly directed and possibly containing loops.
An automorphism of G is a permutation g of V with the property that {vg, wg]
is an edge if and only if {v, w] is an edge or, if G is a digraph, that (vg, wg) is
an arc if and only if (v, w) is an arc. Now the set of all automorphisms of G is a
permutation group Aut(G), called the automorphism group of G.
The denition of an automorphism of a multigraph is a little more complicated.
The most straightforward approach is to interpret a multigraph as a weighted graph.
If a
v,w
denotes the multiplicity of vw as an edge of G, then an automorphism is
a permutation of V satisfying a
vg,wg
= a
v,w
. Again, the set of automorphisms is a
group.
We note the following results:
Theorem
r
A simple undirected graph and its complement have the same automorphism
group.
r
The automorphism group of the complete graph K
n
or the null graph N
n
is the
symmetric group S
n
.
r
The 5-cycle C
5
has ten automorphisms, realized geometrically as the rotations
and reections of a regular pentagon.
This last group is the dihedral group D
10
. More generally, Aut(C
n
) is the dihedral
group D
2n
, for n 3.
We can describe the automorphism group of a disconnected graph in terms of
the group products introduced in the last section.
Theorem Let G be a graph whose connected components are n
1
copies of G
1
, n
2
copies of G
2
, . . . , n
r
copies of G
r
, where G
1
, G
2
, . . . , G
r
are pairwise non-
isomorphic connected graphs. Then
Aut(G) = (Aut(G
1
) : S
n
1
) (Aut(G
2
) : S
n
2
) (Aut(G
r
) : S
n
r
).
This theorem illustrates the notions of both direct and wreath product. First,
if r = 1, then any automorphism is a composition of n
1
independently chosen
automorphisms of G
1
(acting on the components) and an arbitrary permutation
of the components that is, the group is Aut(G
1
) : S
n
1
. For general values of r,
automorphisms of the subgraphs consisting of copies of each G
i
can be chosen
independently and combined, giving the direct product.
Orbits and stabilizers
Let G be a permutation group on . The relation on , dened by
v w if w = vg for some g G,
is an equivalence relation, and its equivalence classes are the orbits of G. G is
transitive if it has just one orbit; thus, G is transitive if, for any v, w , there
exists g such that vg = w.
The stabilizer G
v
of a point v is the set
H = {g G: vg = v];
it is a subgroup of G. Moreover, if w is a point in the same orbit as v, then the
set
{g G: vg = w]
is a right coset of H in G. This correspondence is a bijection between the orbit
of v and the set G/H of right cosets of H in G. It is, moreover, an isomorphism
between the actions of G on the orbit of v and on the set G/H, where the group
Introduction 25
acts by right multiplication on the latter set. It is thus possible to identify with
the union of the coset spaces of a family of subgroups of G, one subgroup for each
orbit of G.
Group actions
Given a subgroup H of a group G, there is an action of G as a permutation group
on the set of right cosets of H: this action is transitive. However, the induced
permutation group may not be isomorphic to G, since there may be different
elements of G inducing the same permutation (or, equivalently, a non-identity
element of G inducing the identity permutation). Indeed, the stabilizer of the coset
Hg is the conjugate g
1
Hg of H, so the kernel of the action (the subgroup xing
all cosets) is the core of H, the intersection of all the conjugates. We say that His
core-free if
_
gG
g
1
Hg = {1];
thus, the action is faithful if and only if H is core-free.
In particular, if H is the identity subgroup, then it is core-free. In this case, the
cosets are just the singleton subsets of G, and we lose nothing by identifying them
with the elements of G. We obtain the action of G on itself by right multiplication,
where the element g induces the permutation x . xg of G. This action was used
by Cayley to show that every group is isomorphic to a permutation group.
Transitivity
Let G = (V, E) be a graph or digraph, and let G be a group of automorphisms of
G (a subgroup of the automorphism group Aut(G)). G is G-vertex-transitive, or
G acts vertex-transitively on G, if G is transitive on the set V; edge-transitivity is
dened similarly. The graph G is G-symmetric if G acts transitively on the set of
ordered pairs (v, w) of adjacent vertices. (If G is a digraph, this is equivalent to
edge-transitivity, but for anundirectedgraphit is a little stronger, since it implies the
existence of an element of G that interchanges the two ends of an edge.) Finally, G
is vertex-transitive (or edge-transitive, or symmetric) if the appropriate transitivity
holds with G = Aut(G). The term arc-transitive is sometimes used as a synonym
for symmetric. Also, if G is G-symmetric, we call (G, G) a symmetric pair. For
example, the pentagon C
5
is a symmetric graph, and (C
5
, D
10
) is a symmetric
pair. The cyclic subgroup of order 5 acts vertex-transitively, but not symmetrically,
on C
5
.
Orbitals and rank
In the 1960s, Sims introduced graph-theoretic methods into the study of permuta-
tion groups, as follows.
Let G be a permutation group on . Then there is a natural action of G on
, the coordinate-wise action:
(v, w)g = (vg, wg),
for g G, v, w . An orbit of G in this action is an orbital, and the number of
orbitals is the rank. Orbitals are of two types:
r
a diagonal orbital is one of the form O = {(v, v) : v A], where A is an orbit
of G in ;
r
a non-diagonal orbital is one consisting of ordered pairs of distinct elements.
If the degree is greater than 1, there is at least one orbital of each type, and so the
rank is at least 2. It is equal to 2 if and only if G is 2-transitive (or doubly transitive)
that is, for any two pairs (v
1
, v
2
) and (w
1
, w
2
) of distinct elements of , there is
an element g G such that v
1
g = w
1
and v
2
g = w
2
.
There is a natural pairing of the orbitals: the orbital O
paired with O is
{(w, v) : (v, w) O]. If O = O
, the orbital O is self-paired; trivially, a diagonal

orbital is self-paired.
Sims dened the orbital graph G(O) associated with an orbital O as an undi-
rected graph if O is self-paired and a directed graph otherwise, as follows: the
vertex-set is , and the edges are all pairs vw for which (v, w) O. If O is diag-
onal, then G(O) just has a loop at every vertex in the corresponding orbit; and if
O is not self-paired, then G(O
) is the converse digraph of G(O).

Graphs admitting a given group
Sims construction of orbital graphs can in a sense be reversed. Let G be a permu-
tation group on . Which simple graphs or digraphs G with vertex-set admit
G as a group of automorphisms? It is clear that, for any such graph, the set of
ordered pairs of adjacent vertices is preserved by G, and so is a union of orbitals.
Conversely, any union of orbitals is the arc set of a digraph G for which G is a
subgroup of Aut(G).
The construction is very exible. If we want to exclude loops, then we consider
only non-diagonal orbitals. If we insist that the graphs are undirected, then we take
only unions that contain both or neither of each pair O, O
of orbitals. It is also
very practical: instead of having to list all the edges in order to specify a graph, we
need give only a set of orbit representatives. This fact is exploited in Chapter 10.
Introduction 27
This construction can be interpreted within the group. If A is an orbit of G on
, and if H = G
v
is the stabilizer of the point v A, then (as noted earlier) there is
a bijection between A and the set of right cosets of Hin G, as follows: to the point
w A corresponds the set {g G: vg = w]. So we can identify with the set
of all such cosets. Now each orbital O of G is contained in A
1
A
2
, for some
orbits A
1
, A
2
of G on . If H
i
is the stabilizer of a point v
i
A
i
, for i = 1, 2, then
the orbit O consists of all pairs (H
1
x, H
2
y) of cosets such that x H
1
gH
2
y
that is, H
1
x H
1
gH
2
y, for some element g G. Now the double coset H
1
gH
2
corresponds to this orbit. The construction can be reversed: to the double coset
H
1
gH
2
corresponds the above set of pairs of cosets (a subset of A
1
A
2
), and
this set is an orbit of G.
So any graph on the vertex-set which admits G as a group of automorphisms
can be represented by a family of subgroups of G (the stabilizers of vertices in
the G-orbits in ), together with a family of double cosets of these subgroups,
corresponding to the G-orbits on that are edges of the graph.
The description becomes simpler when G is vertex-transitive, and even more
so when G is symmetric. First note that, if G is transitive on , then the stabilizer
H of a point of is core-free in G. So any vertex-transitive graph has the form
G(G, H, S),
where His a core-free subgroup of G, and S is the union of a set of double cosets
of the form HxH; in fact S = {g G: vg v], where H = G
v
. The graph is
undirected if, for all x S, we have x
1
S. The connected component contain-
ing v consists of the vertices whose corresponding cosets belong to the subgroup
generated by S. Such a graph is a coset graph.
Now the graph G(G, H, S) is G-symmetric if and only if S consists of just one
double coset HgH. It is also undirected if and only if g
1
HgH. This leads to
the following result of Sabidussi:
Theorem Let G be an undirected G-symmetric graph. Then G is isomorphic to
(G, H, HgH), for some subgroup Hand some element g of G, where H = G
v
and
vg v for some vertex v, and g
1
HgH. This isomorphism is an equivalence
of G-actions. Moreover, G is connected if and only if H, g) = G.
Note that we may assume, without loss of generality, that the order of g is a power
of 2 and that g interchanges v and vg.
The fact that the given symmetric graph G is isomorphic to a coset graph for the
given group G suggests the need for a clear notion of equivalence of symmetric
pairs. An equivalence between symmetric pairs (G, G) and (G
/
, G
/
) consists of
a graph isomorphism f : G G
/
and a group isomorphism : G G
/
such
that, for all vertices u of G and all g G, ((u) f )g = (u(g)) f . We then write
(G, G)

= (G
/
, G
/
). Thus, every symmetric pair (G, G) is equivalent to the sym-
metric pair (G
, G), where G
is the coset graph G(G, H, HgH).

Primitivity and double transitivity
Let G be a transitive permutation group on . G is primitive on if the only
equivalence relations on that are invariant under the action of G are the trivial
ones the relation of equality, and the universal relation with a single equivalence
class. Now the following assertions hold:
r
a 2-transitive group is primitive;
r
G is primitive on if and only if a stabilizer H = G
v
is a maximal proper
subgroup of G, for v ;
r
If G is primitive on , and N is a non-trivial normal subgroup of , then N is
transitive on .
Motivated by the second assertion, we call a permutation group G on quasi-
primitive if all of its non-trivial normal subgroups are transitive. Thus, a primitive
group is quasiprimitive.
If G is transitive on , then there is a single diagonal orbital {(v, v) : v ];
the corresponding orbital graph consists of a loop at each vertex. Sims proved the
following result:
Theorem The transitive permutation group G on is primitive if and only if every
non-diagonal orbital graph is connected.
For, let G(O) be the orbital graph corresponding to the orbital O containing
(v, w). It can be shown that, if H = G
v
and vg = w, then the connected com-
ponent of G(O) containing v is the orbit of the subgroup H, g). So, if G is
primitive, then H is a maximal subgroup of G, and so H, g) = G for all g / H;
thus all non-diagonal orbital graphs are connected. Conversely, if G is not primi-
tive, then there is a proper subgroup Kof G with H < K. So, if g K, g / H, and
w = vg, then the orbital graph containing the edge (v, w) is neither diagonal nor
connected.
Many classications of particular classes of (quasi)primitive groups relevant to
graph theory have been obtained, using the Classication of nite simple groups.
Among these are:
r
the 2-transitive groups;
r
the primitive permutation groups with rank 3;
r
the primitive permutation groups of odd degree, other than those of afne
type.
Introduction 29
One of the main tools in applying this powerful theorem is the ONan-Scott Theo-
rem, which describes the structure of the socle (the product of the minimal normal
subgroups) of a (quasi)primitive permutation group. See Chapter 7, Section 3, for
an account of this theorem.
References
1. P. J. Cameron, Permutation Groups, London Math. Soc. Student Texts 45, Cambridge
Univ. Press, 1999.
2. J. D. Dixon and B. Mortimer, Permutation Groups, Springer, 1996.
3. S. Lipschutz, Linear Algebra, Schaums Outline Series, McGraw-Hill, 1974.
4. D. B. West, Introduction to Graph Theory (2nd ed.), Prentice-Hall, 2001.
5. R. J. Wilson, Introduction to Graph Theory (4th ed.), Pearson, 1996.
1
Eigenvalues of graphs
MICHAEL DOOB
1. Introduction
2. Some examples
3. A little matrix theory
4. Eigenvalues and walks
5. Eigenvalues and labellings of graphs
6. Lower bounds for the eigenvalues
7. Upper bounds for the eigenvalues
8. Other matrices related to graphs
9. Cospectral graphs
References
This chapter gives a survey of some of the relationships between the proper-
ties of a graph and the spectrum of its adjacency matrix. First we give some
working examples and use them to illustrate the necessary matrix theory
background. Next, the eigenvalues of a graph are related to the path struc-
ture. Then we examine the construction of eigenvalues via graph labellings.
The implications of bounding the eigenvalues (from both above and below)
are also explored. Finally, we examine the question of whether a graph is
determined by its spectrum. In addition, we note some relationships between
a graph and other matrices related to the adjacency matrix.
1. Introduction
The eigenvalues of a graph are the eigenvalues of its usual (0-1)-adjacency matrix.
The relationships between the algebraic properties of these eigenvalues and the
usual (topological and geometric) properties of graphs have been studied quite
intensively. At rst blush it seems implausible that such eigenvalues would have
any relationship with such properties of a graph. In this chapter we shall show
30
1 Eigenvalues of graphs 31
that this is not the case. We shall also lay the groundwork for some deeper results
appearing in subsequent chapters by giving a general background for the theory
and presenting some useful computational tools.
Eigenvalues of other matrices (Laplacian, Seidel matrices, etc.) associated with
a graph are also of interest, and when we talk of such eigenvalues we will specif-
ically note the type of matrix under consideration. Relationships between these
matrices and the adjacency matrix will be discussed.
Agood introduction to the some of the earlier results concerning eigenvalues of
graphs is given in [25]. More advanced work is discussed in [10] and [15]. Alist of
all the graphs with up to seven vertices, with their eigenvalues and characteristic
polynomials, appears in [22].
2. Some examples
It is always good to have some particular graphs with which to work, so rst we
give some examples and their eigenvalues. We start with a specic example, the
graph Q
3
in Fig. 1.
The adjacency matrix A is a square matrix of order 8:
A =
_
_
_
_
_
_
_
_
_
_
_
_
_
0 0 0 0 0 1 1 1
0 0 0 0 1 0 1 1
0 0 0 0 1 1 0 1
0 0 0 0 1 1 1 0
0 1 1 1 0 0 0 0
1 0 1 1 0 0 0 0
1 1 0 1 0 0 0 0
1 1 1 0 0 0 0 0
_
_
_
_
_
_
_
_
_
_
_
_
_
.
The characteristic polynomial of this matrix is easily computed:
det(xI A) = x
8
12x
6
+30x
4
28x
2
+9 = (x 3)(x +3)(x 1)
3
(x +1)
3
,
and so the eigenvalues of Q
3
are 3, 3, 1 and 1, with respective multiplicities
1, 1, 3 and 3.
1 6
8 3
7
2
4
5
Fig. 1.
32 Michael Doob
Table 1.
Graph Eigenvalues Respective multiplicities
K
n
n 1, 1 1, n 1
K
r,s
rs, 0,
rs 1, r +s 2, 1
C
n
2 cos
_
2k
n
_
, k = 1, 2, . . . , n 2, 1, 1, . . . , 1, 2 for n even
1, 1, . . . , 1, 2 for n odd
P
n
2 cos
_
k
n+1
_
, k = 1, 2, . . . , n 1, 1, . . . , 1
L(K
r
) 2r 4, r 4, 2 1, r 1, r(r 3)/2
L(K
r,r
) 2r 2, r 2, 2 1, 2(r 1), (r 1)
2
L(K
r,s
) r +s 2, r 2, s 2, 2 1, s 1, r 1, (r 1)(s 1)
Q
d
d 2k, k = 0, 1, . . . , d
_
d
k
_
, k = 0, 1, . . . , d
CP(r) or K
r(2)
2r 2, 0, 2 1, r, r 1
Table 1 gives some more examples. By the end of this chapter, we shall have
veried all of them. In this table, note that two graphs have the same spectrumonly
if they are isomorphic. This is not true in general, as the following examples show.
The rst pair of non-isomorphic graphs, shown in Fig. 2, is the smallest one with
the same eigenvalues: they are 2, 0 and 2, with respective multiplicities 1, 3 and
1. We see from this example that, in general, connectivity cannot be determined
from the eigenvalues alone.
The next graphs, shown in Fig. 3, are the smallest connected pair with the same
eigenvalues. These two graphs have diameters 2 and 4. Hence, in general, the
diameter of a graph cannot be determined from the eigenvalues.
The trees in Fig. 4, are the smallest pair with the same eigenvalues:
1
2

13
2
,
each of multiplicity 1, and 0 with multiplicity 4. These trees have very different
Fig. 2. Characteristic polynomial: p(x) = x
5
4x
3
6
7x
4
4x
3
+7x
2
+4x 1
8
7x
6
+9x
4
degrees. Hence, in general, we cannot determine much about the degree sequence
from the eigenvalues of a graph.
From these examples we see that there are several properties of graphs that are
not determined by their eigenvalues; more will be said about this in Section 9. On
the other hand, there are other properties (for example, the number of edges and
the number of triangles in a graph) that are determined by the eigenvalues. Some
of these will be given in Section 4. The sorting of graph properties into those
determined by the eigenvalues or not continues to be of research interest.
3. A little matrix theory
Since the adjacency matrix of a graph is real and symmetric, we may use known
results from the theory of matrices to our advantage. We will state some of them
without proofs, which may be found in [14] or [20]. The rst is sometimes called
the principal axis theorem.
Theorem 3.1 (Principal axis theorem) If A is a real symmetric matrix of or-
der n, then A has n real eigenvalues and a corresponding orthonormal set of
eigenvectors.
If U is a matrix with an orthonormal set of eigenvectors as columns, then
U
T
= U
1
and U
T
AU = D, where D is the diagonal matrix whose entries are the
corresponding eigenvalues. This gives a diagonalization theorem.
Theorem 3.2 If A is a real symmetric matrix, then there exists a matrix U such
that U
T
AU = D. In addition, the minimum polynomial is

(x
i
), where the
product is taken over the distinct eigenvalues.
The next theorem, the Perron-Frobenius theorem, gives information about the
largest eigenvalue of a matrix with non-negative entries.
Theorem 3.3 (Perron-Frobenius theorem) If A is a non-negative matrix
with eigenvalues
1

2

n
, then |
1
| |
k
|, for k = 1, 2, . . . , n, and
the eigenvalue
1
has an eigenvector with all entries non-negative. If A is
34 Michael Doob
indecomposable, then the eigenvalue
1
is simple (
1
>
2
), and the eigenvec-
tor has all entries positive.
The adjacency matrix of a graph is indecomposable precisely when the graph
is connected. Hence the largest eigenvalue of a connected graph is simple.
It is a straightforward observation that the spectrum of a graph is the union of
the spectra of its connected components. So, unless otherwise stated, we assume
that all graphs under consideration are connected.
The following result is known as the Interlacing theorem.
Theorem 3.4 (Interlacing theorem) Let A be a real symmetric matrix with
eigenvalues
1

2

n
, and let
1

2

n1
be the eigenval-
ues of a principal submatrix of A. Then
i

i

i +1
, for i = 1, 2, . . . , n 1.
This theoremis of major importance in the study of graph eigenvalues. Aprincipal
submatrix corresponds to an induced subgraph with one fewer vertex. This result
can be visualized as interlacing the eigenvalues on the real axis, as shown in Fig. 5.
The Interlacing theorem and an easy induction yield the following result for
induced subgraphs:
Corollary 3.5 If H is an induced subgraph of G, if
1

2

m
are
the eigenvalues of H, and if
1

2

n
are the eigenvalues of G, then
i +nm

i

i
, for i = 1, 2, . . . , m.
4. Eigenvalues and walks
The oldest, and perhaps the most fundamental, relationship between the eigenval-
ues of a graph and its geometric properties concerns walks particularly, closed
walks. Suppose that A is the adjacency matrix of a graph. Then the powers of A
enumerate the walks in that graph.
Theorem 4.1 If a graph G has adjacency matrix A, then for k = 0, 1, . . . , the
ij-entry of A
k
is the number of v
i
-v
j
walks of length k.
eigenvalues of G
eigenvalues of G {
i
}
Fig. 5.
Proof The case k = 0 is trivial, but will be useful later in this section; the case
k = 1 is the denition of the adjacency matrix. For larger values of k, the proof
follows by an easy induction that uses nothing more than the denition of matrix
multiplication.
This theorem is often applied to enumerate closed walks of length k. These
correspond to diagonal entries in A
k
, and so the number of closed walks of length
k is the sum of the diagonal entries, which in turn is equal to tr (A
k
).
Corollary 4.2 If A is the adjacency matrix of a graph G with degree sequence
d
1
, d
2
, . . . , d
n
, then
r
A
2
i i
= deg v
i
;
r
the number of edges of G is given by
|E(G)| =
1
2
n
i =1
deg v
i
=
1
2
tr (A
2
);
r
the number of triangles in G is
1
6
tr (A
3
).
From this we see that the number of edges and the number of triangles in a graph
are determined by its eigenvalues. This argument does not extend to k-cycles, with
k > 3: the rst example of cospectral graphs in Section 2 shows that the number
of 4-cycles cannot be determined by the eigenvalues alone. However, the degree
sequence adds enough extra information to determine the 4-cycles.
Corollary 4.3 If d
1
, d
2
, . . . , d
n
is the degree sequence of G, then the number of
4-cycles in G is
1
8
(trA
4
+trA
2
n
i =1
d
2
i
).
We can also use Theorem 4.1 to bound the diameter of a graph by the number
of distinct eigenvalues.
Theorem 4.4 If a graph G has diameter D and t distinct eigenvalues, then
D t 1.
Proof Suppose not. Then there exist two vertices v
i
and v
j
such that the distance
between themis exactly t . Thus A
t
i j
> 0, but A
k
i j
= 0 for k = 1, 2, . . . , t 1. So if
q(x) is any polynomial of degree t , the i j -entry of q(A) must be non-zero. But, by
Theorem 3.2, the minimum polynomial is of degree t ; when evaluated at A, this
polynomial has every entry equal to 0, and hence we have a contradiction.
The graphs K
n
, K
r,s
, C
n
, P
n
, L(K
r
), L(K
r,r
), Q
d
and CP(r), in the table of
examples in Section 2, all attain the bound of the last theorem, while L(K
r,s
) does
not whenever r = s. Distance-regular graphs (described later in this chapter) meet
36 Michael Doob
this bound too; however, there is no known general characterization of graphs that
attain this bound.
Matrix equations involving the minimum and other polynomials are often used
to reveal information about the graph, and vice versa. For example, the complete
graph K
n
must have A
2
i j
= n 2 if i = j , and A
2
i i
= n 1. Thus
A
2
(n 2)A = I,
and
x
2
(n 2)x 1 = (x (n 1))(x +1)
is the minimum polynomial. Since the trace of A is 0, the eigenvalues are n 1
(simple) and 1 (n 1 times). Conversely, if a graph with n vertices has the same
eigenvalues as K
n
, then the minimum polynomial forces the adjacency matrix to
satisfy A
2
(n 2)A = I, which makes the graph regular with degree n 1, and
hence K
n
. We say that K
n
is characterized by its spectrum.
Another polynomial is used in the case of regular connected graphs. In this
case, the adjacency matrix A and the all-1 matrix J commute. If the degree of the
graph is r, then the all-1 vector j is an eigenvector of A (with eigenvalue r) and
of J (with eigenvalue n). Any other eigenvector of A is orthogonal to j, and so
is also an eigenvector of J (with eigenvalue 0). If we consider the interpolating
polynomial
h(x) = n
x
i
r
i
,
where the product is taken over all distinct eigenvalues that are not equal to r,
then h(r) = n and h() = 0 for any other eigenvalue . Using our orthonormal
basis of eigenvectors, we see that h(A) J is a symmetric matrix that annihilates
any vector, and hence h(A) J = 0, or h(A) = J. This polynomial is called the
Hoffman polynomial.
Strongly regular graphs
Agraph is strongly regular with parameters (n, r, , ) if it has n vertices, is regular
of degree r, has each pair of adjacent vertices mutually adjacent to other vertices,
and has each pair of non-adjacent vertices mutually adjacent to other vertices.
By convention, we exclude any graph in which each component is complete (that
is, = 0), so that all strongly regular graphs are connected. An example of a
strongly regular graph is the Petersen graph: it clearly has parameters (10, 3, 0, 1).
Strongly regular graphs are considered in detail in Chapter 8.
For any strongly regular graph there are only three types of entries in the
adjacency matrix: those for which two vertices are adjacent, those for which two
vertices are non-adjacent, and the diagonal. We know that the entry in A
2
is ,
and r in each of these cases. From this it is clear that
A
2
+( )A +( r)I = J.
Hence we have essentially the Hoffman polynomial, and we know that the eigen-
values of the graph are r and the two roots of
x
2
+( )x +( r).
This means that strongly regular graphs attain the diameter bound of Theorem 4.4.
Once we have the eigenvalues, it is trivial to compute their multiplicities. If the
distinct eigenvalues of the graph are r,
2
and
3
, with respective multiplicities 1,
m
2
and m
3
, then we have
m
2
+m
3
= n 1 and m
2
2
+m
3
3
= r,
since the adjacency matrix of any graph has trace 0. These multiplicities must
be integers, and, since
2
=
3
, they are determined by the equations displayed
above. The integrality of m
2
and m
3
has been an especially useful tool for showing
the non-existence of potential strongly regular graphs with certain parameters.
We have thus veried the spectra of L(K
n
), L(K
r,r
), K
r,r
and CP(r), as pre-
sented in Section 2, since these graphs are strongly regular.
A regular graph of degree r and diameter d can have at most
1 +r +r(r 1) + +r(r 1)
d
vertices. A graph that attains this bound is called a Moore graph of diameter d. A
Moore graphof diameter d = 2is stronglyregular withparameters (r
2
+1, r, 0, 1).
One of the oldest existence questions in spectral graph theory arises from these
graphs: for which r do these Moore graphs exist?
We follow the logic given above: the eigenvalues are r plus the two solutions
of
x
2
+ x (r 1) = 0
that is,
1
2
(1
4r 3).
38 Michael Doob
Then
2
+
3
= 1 and
2
3
=
4r 3.
In addition,
m
2
+m
3
= r
2
and m
2
2
+m
3
3
= r.
Hence,
(m
2
m
3
)
4r 3 = r(r 2).
If

4r 3 is not an integer, then m
2
= m
3
and r = 2, so that the graph is C
5
.
Otherwise,
4r 3 is an odd integer, 2t +1, so that 2t +1 divides

16r(r 2) = (2t +1)
2
(4t
2
+4t 1) 15.
Hence 2t +1 divides 15, and so t = 0, 1, 2 or 7 which in turn implies that
r = 1, 3, 7 or 57. The case r = 3 gives the Petersen graph, and the case r = 7
also yields a single graph, called the Hoffman-Singleton graph. What happens
in the nal case r = 57 (a question raised in 1960 in [17]) has remained un-
solved, despite much effort over several decades. The Petersen graph contains
two copies of C
5
, the Hoffman-Singleton graph has 50 vertices and contains 10
copies of the Petersen graph. The graph with r = 57, if it exists, would have 3250
vertices.
Much more will be said about strongly regular graphs in Chapter 8.
Distance-regular graphs
Strongly regular graphs can be generalized to distance-regular graphs with di-
ameter greater than 2. We dene n
i
(v) to be the number of vertices at distance
i from v, and for two vertices v and w at distance k from each other, p
k
i j
(v, w)
is the number of vertices at distance i from v and distance j from w. If n
i
(v)
is the same for all v, and if p
k
i j
(v, w) is the same for all vertices v and w at
distance k from each other, then the graph is distance-regular. A strongly regular
graph with parameters (n, r, , ) is such a graph with n
1
(v) = r, n
2
(v) =
n r 1, p
1
11
= and p
2
11
= . The d-dimensional cube Q
d
is distance-regular
with diameter d.
A distance-regular graph of diameter d has an associated intersection array, an
idea introduced in [2]. It has parameters c
i
= p
i
i 1,1
, b
i
= p
i
i +1,1
and a
i
= p
i
i 1
=
r b
i
c
i
, for i = 0, 1, . . . , d. It turns out that the eigenvalues of A are the same
as the eigenvalues of the tridiagonal matrix
_
_
_
_
_
_
_
_
_
a
0
b
0
0 0 0 0
c
1
a
1
b
1
0 0 0
0 c
2
a
2
b
2
0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 a
d1
b
d1
0 0 0 0 c
d
a
d
_
_
_
_
_
_
_
_
_
.
Since the order of this matrix is only d +1, there can be at most d +1 eigenvalues.
ByTheorem4.4, the graphmust have at least d +1eigenvalues. Hence all distance-
regular graphs attain the diameter bound of Theorem 4.4.
We conclude this section with two important distance-regular graphs.
The Hamming graph H(d, l) has the d-tuples (x
1
, x
2
, . . . , x
d
) as its vertices,
where 1 x
i
l. Two vertices are adjacent if, as d-tuples, they agree in all but one
coordinate. This graphhas diameter d, since the distance betweentwovertices is the
number of coordinates in which the corresponding d-tuples differ. The eigenvalues
are kl d, with multiplicity
_
d
k
_
, for k = 0, 1, . . . , d. Notice that if l = 2, we have
the cube graph Q
d
.
The Johnson graph J(d, l) has vertices corresponding to the
_
n
d
_
subsets of
{1, 2, . . . , l}. Two vertices are adjacent if, as subsets, their intersection is of size
d 1. This graph has diameter d, since the distance between two vertices is k if
and only if the corresponding subsets have intersection of size d k. Notice that
if d = 2, we have the graph L(K
l
).
The book [4] is an encyclopaedic reference on distance-regular graphs.
5. Eigenvalues and labellings of graphs
The columns of the adjacency matrix Acorrespond to the vertices of the graph. If A
acts on a vector x, then the entries of that vector can have the same correspondence,
so a vector may be used to label the vertices. When we use an eigenvector, the
eigenvalues will reect the geometric structure of the graph.
Theorem 5.1 If A is the adjacency matrix of a graph G with vertices {v
1
, v
2
, . . . ,
v
n
}, and the vector x is viewed as a labelling of the vertices of G, then, for k =
1, 2, . . . , n, (Ax)
k
is the sum of the labels on the vertices adjacent to v
k
.
This result is convenient for direct calculation. For example, consider the cycle
C
n
, and let
k
= e
2i k/n
(an nth root of unity). If the vertices in order around the
cycle are v
1
, v
2
, . . . , v
n
, let x = (x
1
, x
2
, . . . , x
n
) be dened by letting x
j
=
j
k
.
40 Michael Doob
Then (Ax)
j
= x
j 1
+ x
j +1
(subscripts taken modulo n), and
x
j 1
+ x
j +1
x
j
=
e
2i ( j 1)k/n
+e
2i ( j +1)k/n
e
2i j k/n
= e
2i k/n
+e
2i k/n
=
1
k
+
k
.
Hence, Ax = (
1
k
+
k
)x and, on letting k range from 1 to n, we obtain all the
eigenvalues of C
n
. Notice that the roots of unity
k
and
nk
yield the same eigen-
value, and so, unless k = n or n/2, this eigenvalue has multiplicity 2; otherwise,
the multiplicity is 1.
Similar labellings can be used to nd the eigenvalues of a path. Consider the
path P
n
and the cycle C
2n+2
. Let x be any eigenvector of P
n
and use that vector to
label the rst n vertices of C
2n+2
. Label the next vertex 0, the next n vertices with
x in reverse order, and the last vertex 0. It is now easy to verify that we have an
eigenvector of C
2n+2
. If we shift the labels one step around the cycle (labelling
the rst vertex 0, the next n vertices by x, the next one 0, and the nal n by x in
reverse order), we get a second independent eigenvector. Hence the eigenvalues of
P
n
correspond to the double eigenvalues of C
2n+2
. This technique rst appeared
in [25].
We can also use eigenvalue labellings to deduce an easy result about bipartite
graphs. If G has a bipartition (S, T), and if x is an eigenvector with corresponding
eigenvalue , then we can replace the label x
i
with x
i
for each vertex in S to
get an eigenvector with corresponding eigenvalue . Hence, and have the
same multiplicity, and the spectrum is symmetric about 0. Conversely, if G has a
spectrum that is symmetric about 0, then the trace of A
2k+1
is 0 and the graph has
no odd cycles, and so G is bipartite. This yields one of the oldest results concerning
the spectra of graphs.
Theorem 5.2 A graph is bipartite if and only if its spectrum is symmetric
about 0.
Nowsuppose that v
i
and v
j
are non-adjacent vertices with the same neighbours,
and suppose that x is an eigenvector with eigenvalue . If is the sumof the labels
on the neighbours of v
i
and v
j
, then clearly
x
i
= (Ax)
i
= = (Ax)
j
= x
j
,
and hence (x
i
x
j
) = 0. This yields the following result about the symmetry of
the eigenvectors of a graph.
Corollary 5.3 If v and w are non-adjacent vertices with the same neighbours,
and if x is an eigenvector with eigenvalue , then either v and w have the same
label or = 0.
We can apply this result to K
r,s
. If = 0, then any eigenvector x gives labels that
can have only two possible values, say y and z. We can then say that y = sz and
z = r y, and in particular, y = 0 and z = 0. Hence rsyz = szr y = (y)(z) =
2
yz. So
2
= rs, and the only non-zero eigenvalues are
rs. Since K
r,s
is
connected and has trace 0, these two non-zero values are simple eigenvalues.
A similar result can be obtained for adjacent vertices. Suppose that v
i
and v
j
are adjacent vertices with the same neighbours, and that x is a eigenvector with
eigenvalue . Again, let be the sum of the labels on the neighbours of v
i
and v
j
.
Then
+ x
j
= (Ax)
i
= x
i
and + x
i
= (Ax)
j
= x
j
,
so that ( +1)(x
i
x
j
) = 0.
Corollary 5.4 If v and w are adjacent vertices with the same neighbours, and if
x is an eigenvector with eigenvalue , then either v and w have the same label or
= 1.
We can apply this result to the complete graph K
n
with one edge deleted.
Suppose that G has n vertices, all of which are pairwise joined except for v
1
and
v
2
, and suppose that x is an eigenvector with eigenvalue . For = 0, we see
from Corollary 5.4 that x
3
= x
4
= = x
n
, but Corollary 5.3 tells us nothing.
However, (Ax)
1
= 0 tells us that the common value is 0, and then x
1
= x
2
; thus
= 0 has multiplicity 1. For = 1, we see from Corollary 5.3 that x
1
= x
2
, but
Corollary 5.4 tells us nothing. However,
x
1
= (Ax)
1
=
n
i =3
x
i
= 0 = (Ax)
2
= x
2
,
andsox
1
= x
2
=
n
i =1
x
i
. Onthe other hand, (Ax)
3
= x
3
implies that
n
i =1
x
i
=
0. Hence the eigenspace is
_
(x
1
, x
2
, . . . , x
n
) : x
1
= x
2
= 0 =
n
i =3
x
i
_
,
which has dimension n 3. For any other eigenvalue , x
1
= x
2
= a and x
3
=
x
4
= = x
n
= b. Hence a = (n 2)b and b = (n 3)b +2a. Elimination
gives
=
1
2
_
n 3
_
(n +1)
2
8
_
.
Thus we have completely determined the spectrum.
Starting with a matrix A, the Rayleigh quotient is dened using the usual inner
product for any non-zero vector x as Ax, x/x, x. Since the Rayleigh quotients
42 Michael Doob
of x and cx are identical for c = 0, we can restrict our attention to vectors of unit
length when convenient.
We rst observe that if A is the adjacency matrix of a graph, then the Rayleigh
quotient takes on precisely all values between the largest and smallest eigenvalues.
Theorem 5.5 Let Gbe agraphwithadjacency matrix Aandwitheigenvalues
1

2

n
. Then Ax, x/x, x takes on precisely the values in the interval
[
n
,
1
]. In addition, Ax, x/x, x =
1
or
n
only if x is an eigenvector.
Proof Let y be any non-zero vector, and let {x
1
, x
2
, . . . , x
n
} be an orthonormal
basis of eigenvectors for A. Then y =
r
i
x
i
and
Ay, y
y, y
=
_
A
_
r
i
x
i
_
,
r
i
x
i
_
_
r
j
x
j
,
r
j
x
j
_ =
r
2
i

i
r
2
j
=
r
2
i
r
2
j
i
.
The coefcients of
i
are all non-negative and add to 1 that is, the Rayleigh
quotient is a convex combination of the eigenvalues. Hence, the Rayleigh quotient
of every non-zero vector is in the interval [
n
,
1
], each value in that interval is
attained by the Rayleigh quotient for some vector, and the extreme values of the
interval are attained only by eigenvectors.
Observe also that if
1
is the largest eigenvalue of a connected graph G, x is
a corresponding eigenvector, and Ay, y/y, y =
1
, then we can write y as a
convex combination of eigenvectors of G. Since
1
is simple, this means that
y = rx for some real number r = 0, and so y has all of its coordinates positive or
all of its coordinates negative.
Corollary 5.6 If H is a proper induced subgraph of a connected graph G, if H has
maximum eigenvalue , and if G has maximum eigenvalue , then < .
Proof We know that , from Theorem 3.4. If z is an eigenvector of H with
eigenvalue , we can extend z to y by adding 0 as entries corresponding to vertices
not in H; then Ay, y/y, y = . If = , then by the observation in the previous
paragraph, all the coordinates of y are non-zero. This can happen only if H = G,
and so < .
Corollary 5.7 If e is an edge of a connected graph G, then the largest eigenvalue
of G is strictly greater than the largest eigenvalue of G e.
Proof Let H = G e. If H is disconnected, then its maximum eigenvalue is
attained on one of the components, which is a proper induced subgraph, so Corol-
lary 5.6 applies. If H is connected, let x be an eigenvector of H corresponding
to its largest eigenvalue . If the deleted edge joins v
i
and v
j
, and if A is the
adjacency matrix for G, then the largest eigenvalue of G is bounded below by
Ax, x/x, x = +2x
i
x
j
. Since x has only positive or only negative entries, the
largest eigenvalue of G must be greater than that of H.
The classic proof of the Perron-Frobenius theorem brackets the largest eigen-
value of a non-negative matrix by considering the values of (Ax)
i
/x
i
(i = 1, 2, . . . ,
n), where x is a vector with non-negative coordinates. The largest eigenvalue
1
satises
min
i
_
(Ax)
i
x
i
_

1
max
i
_
(Ax)
i
x
i
_
.
The proof actually shows that if we maximize the left value and minimize the right
value over x 0, then we get the same value fromboth cases. If we let j be the all-1
vector and A be the adjacency matrix of a graph G with degrees d
1
, d
2
, . . . , d
n
,
then we get
min
i
d
i

1
max
i
d
i
.
Onthe other hand, if we use Theorem5.5withthe same vector j, we get
1
n
n
i =1
d
i

. This results in the following theorem.
Theorem 5.8 If Gis a graph with degrees d
1
, d
2
, . . . , d
n
and maximumeigenvalue
1
, then
1
n
n
i =1
d
i

1
max
i
d
i
.
Equality is attained if and only if the graph is regular.
In other words, the maximum eigenvalue lies between the average degree and the
maximum degree, and equals both values if and only if the graph is regular.
6. Lower bounds for the eigenvalues
Finding lower bounds for the eigenvalues of graphs has been a recurring theme
in the study of graph spectra. In this section we use (G) to denote the smallest
eigenvalue.
Proposition 6.1 Let G be a connected graph with least eigenvalue (G). Then
r
(G) 0, with equality for a null graph;
r
if G is not null, then (G) 1, with equality if and only if G is complete;
44 Michael Doob
r
if G is neither complete nor null, then (G)
2, with equality if and only if

G = K
1,2
;
r
if G is neither complete nor K
1,2
, and if (G) 1.5, then G is
(In fact, (G) =
2
3
10 sin(
1
3
arccos(
10
100
) +

6
) +
1
3
1.4811943, for this
graph.)
Proof Since the trace of any adjacency matrix is 0, (G) 0. If a graph has an
edge, then that edge is a two-vertex induced subgraph with least eigenvalue 1.
Theorem 3.5 implies that (G) 1. If G is not a complete graph, then K
1,2
is
an induced subgraph with least eigenvalue
2. Among the graphs G with four

vertices, the one given in the last statement is the only one with least eigenvalue
(G) 1.5.
So we now know that the graphs G with (G) 1.5 are pretty sporadic. The
situation for 2 (G) 1.5 is just the opposite. First, we make an observation
that motivated the original study of graphs for which 2 (G).
The vertex-edge incidence matrix of a graph is a (0, 1)-matrix with rows corre-
sponding to the vertices, columns corresponding to the edges, and an entry equal
to 1 if and only if its row and column correspond to an incident vertex and edge.
Theorem 6.2 Any line graph L(G) satises (L(G)) 2.
Proof Let K be the vertex-edge incidence matrix of G. Evidently,
KK
T
= 2I +A(L(G)).
Since KK
T
is positive semidenite, it has non-negative eigenvalues.
Two non-isomorphic graphs with more than four vertices have non-isomorphic
line graphs. So, in a sense, the number of graphs satisfying 2 (G) < 1.5 is
the same as the number of all line graphs.
Not all graphs satisfying 2 (G) are line graphs. For example, if G is a
cocktail party graph CP(r), then (G) = 2, even though for any r > 2, G is not a
line graph. Sothe determinationof all graphs with 2 (G) is more complicated,
but, as it turns out, leads to one of the most beautiful results in the study of graph
spectra.
Suppose that G is a graph with n vertices and adjacency matrix A, and assume
that 2 (G). Then A +2I is positive semidenite, and hence is of the form
KK
T
for some appropriate matrix K with n rows; the number of columns is the
v
60
v
u v
v
u + v u
u u
Fig. 6.
rank of A +2I. Let k
1
, k
2
, . . . , k
n
be the n rows of K. Then k
i
k
j
= 1 if v
i
and
v
j
are adjacent, k
i
k
j
= 0 if they are not, and k
i
k
i
= 2. So we may think of
these vectors as having length
2 and meeting at angles of 60
or 90
. Now if we
have a set of vectors, and two vectors u and v meet at 60
, then we can add four

other vectors u, v, u v and u +v, to get three equiangular lines in the form
of a star, as shown in Fig. 6. So if we start with vectors k
1
, k
2
, . . . , k
n
, then we
can add vectors when needed to complete the star. Since there can only be a nite
number of vectors with each pair meeting at 60
, 90
or 120
, we can repeat this

process a nite number of times until it eventually stops at some set . This new
set of vectors is called star closed.
Notice that has the following properties:
r
is nite;
r
if x and cx , then c = 1;
r
if x, y , then 2x, y/x, x is an integer;
r
if x and if H is the hyperplane orthogonal to x, then is closed under
reections by H: in other words, x, y implies that y 2
x,y
x,x
x is also in .
Sets of vectors satisfying these properties are called root systems and have been
studied extensively, especially since they are used to classify real semisimple Lie
algebras. Our set has the additional property that all vectors have the same
length. All such root systems are known: complete expositions can be found in
[3, Ch. 6] and [7]. The complete list of such root systems is shown in Table 2, along
with one set of generating vectors; see also Chapter 3, Section 5. Since v if
and only if v , it is customary to count the number of lines in a root system.
The standard basis is denoted by {e
1
, e
2
, . . . , e
n
}, where (e
i
)
j
is 1 if i = j , and 0
otherwise. The beauty of this result is that the arguments are reversible: starting
from the graph theory we can classify all the root systems. The details may be
found in [6].
When the adjacency matrix of a graph satises A = KK
T
2I, where the rows
of Kare taken froma root system, we say that the root systemrepresents the graph.
Suppose that G is a bipartite graph with bipartition (X, Y). For an edge joining
v
i
X and v
j
Y, consider the vector e
i
e
j
. Let K be the matrix with all such
vectors as rows. Then KK
T
= 2I +A(L(G)).
46 Michael Doob
Table 2.
Name Generating vectors Number of lines
A
n
{e
i
e
j
: 1 i < j n +1}
1
2
n(n +1)
D
n
{e
i
e
j
: 1 i < j n} n(n 1)
E
6
A
5
{e
7
e
8
}
_
1
2
8
i =1
i
e
i
:
i
= 1,
6
i =1
i
=
7
+
8
= 0
_ 36
E
7
A
7
_
1
2
8
i =1
i
e
i
:
i
= 1,
8
i =1
i
= 0
_
63
E
8
D
8
_
1
2
8
i =1
i
e
i
:
i
= 1,
8
i =1
i
= 1
_
120
Conversely, if KK
T
has the rows of K taken from the root system A
n
, then any
row is of the form e
i
e
j
; for each such e
i
e
j
, join v
i
and v
j
. This graph G has
vertices corresponding to the columns of K. Multiplying a row vector by 1 when
necessary allows each column to have entries with the same sign. Let the columns
with positive entries form one set of the bipartition, and the columns with negative
entries form the other one. The graph is then bipartite and KK
T
= 2I +A(L(G)).
Theorem 6.3 A graph is represented by A
n
if and only if it is the line graph of a
bipartite graph.
It is easy to see that any line graph can be represented by D
n
. If the edge joins
v
i
and v
j
, let e
i
+e
j
be a row of K; then KK
T
2I = A(L(G)). Notice that the
columns of K correspond to the vertices of G.
If we construct K by taking the vectors e
1
e
j
( j = 2, 3, . . . , n +1) as 2n
rows, then KK
T
2I = A(CP(n)). Since CP(n) is not a line graph for n > 2, we
see that graphs other than line graphs may be represented by D
n
.
The generalized line graph L(G, a
1
, a
2
, . . . , a
n
) is dened for a graph G with
n vertices. It starts with disjoint copies of L(G) and CP(a
1
), CP(a
2
), . . . , CP(a
n
).
Then every vertex in CP(a
i
) is joined to each vertex in L(G) whose corresponding
edge in G has v
i
as an endpoint.
Now consider the following construction of K. The rows for the line graph
L(G) are constructed as before: for each edge v
i
v
j
, let e
i
+e
j
be a row of K. For
each CP(a
i
), add the vectors e
i
e
n+a
1
++a
i 1
+j
, for j = 1, 2, . . . , a
i
. (The last
subscript is just an artifact to increase the column count by 1 each time that a new
pair of vectors is appended to K.) Upon completion of the construction, we have
KK
T
= 2I +A(L(G, a
1
, a
2
, . . . , a
n
)).
Conversely, if we have a set of vectors R taken from D
n
, where the inner
product of any two of them is 0 or 1, we can construct a generalized line graph
in the following way. For 1 i = j n, it may be that two vectors of the form
e
i
e
j
are in R: we take all such vectors and put them aside. For the remaining
vectors, if e
i
e
j
and e
j
e
k
are two vectors in R, then i = k and the e
j
terms have the same sign. Now use these remaining vectors as rows of K. We can
multiply a column by 1 without changing the inner product of any two rows.
Since all entries in a column have the same sign, we may assume with no loss of
generality that all of these vectors are of the form e
i
+e
j
. Now let the vertices of
G correspond to the non-zero columns of K, and join v
i
and v
j
if e
i
+e
j
is a row
of K. Then KK
T
= 2I +A(L(G)) (so far).
Now we go back to the vectors we put aside. They come in pairs and are of the
form e
i
e
j
or e
i
+e
j
; we need consider only the case e
i
e
j
since the other
case is symmetric. We append these vectors as rows of K. Note that the e
j
are the
only non-zero entries in the j th column since an inner product of 1 is forbidden.
The column containing e
i
may contain other entries, but if it does, then the column
corresponds to a vertex in G. The rows corresponding to the newly added vectors
that are non-zero in the i th column must induce a CP(a
i
) subgraph in KK
T
2I.
Finally, we observe that each vertex in the CP(a
i
) is adjacent to v
i
. Thus
KK
T
= 2I +A(L(G, a
1
, a
2
, . . . , a
n
)).
Theorem 6.4 A graph is represented by D
n
if and only if it is a generalized line
graph.
Using the table of generating vectors for root systems given previously, we
immediately deduce the inclusions E
6
E
7
E
8
; hence, we have the following
theorem.
Theorem 6.5 If (G) 2, then Gis a generalized line graph or Gis represented
by E
8
.
Since E
8
is nite, all but a nite number of graphs with (G) 2 are general-
ized line graphs. Of those that are not, it is known that the largest has 36 vertices. If
we consider graphs with (G) > 2, then KK
T
is non-singular, and hence K has
six, seven or eight rows according as G is represented by E
6
, E
7
or E
8
. In fact, there
are 20 such graphs on six vertices, 110 graphs on seven vertices, and 443 graphs
on eight vertices. These were originally found by a mixture of theoretical analysis
and computer search (see [13] and [11]). Further details appear in Chapter 2.
7. Upper bounds for the eigenvalues
In the previous section we saw how to use root systems to nd graphs with eigen-
values bounded from below by 2. A similar analysis can be used to nd graphs
with eigenvalues bounded from above by 2.
48 Michael Doob
i
j
k
Fig. 7. The graph T(i, j, k)
To proceed, we dene the graph T(i, j, k), shown in Fig. 7: take three paths
P
i
, P
j
and P
k
and add a new vertex adjacent to one end-vertex of each path. The
graph is then a tree with i + j +k +1 vertices, three pendant vertices and one
vertex of degree 3.
Now we form a matrix K from a set of vectors, where the inner product of any
twovectors is 0or 1. We canthendene the adjacencymatrixof G bythe equation
KK
T
= 2I A(G)
and we will have a graph whose eigenvalues are bounded from above by 2. Such
sets of vectors have been fully studied: they are called fundamental sets of roots.
Theorem 7.1 If G is a graph with largest eigenvalue
1
= 2, then G is one of the
following graphs: C
n
, K
1,4
, T(2, 2, 2), T(3, 3, 1), T(5, 2, 1) or
Actually, it is an easy exercise in vertex labelling to show that each of these
graphs has maximum eigenvalue 2, and to show that no other graphs can be so
labelled (see [27]). Thus the results concerning fundamental sets of roots can be
derived directly from graph theory.
FromCorollary 5.7 we nowknowall graphs with largest eigenvalue less than 2.
Corollary 7.2 If G is a graph with largest eigenvalue
1
< 2, then G is a path
P
n
, T(1, 1, r), T(1, 2, 4), T(1, 2, 3) or T(1, 2, 2).
We can say more about graphs with largest eigenvalue greater than 2. In [9],
graphs are described whose largest eigenvalue is bounded above by (2 +
5)
1/2
2.058171. It turns out that all such graphs are trees of the form T(i, j, k) or
S( j, k, l), where the latter graph appears in Fig. 8. Notice that the graph given
in the conclusion of Theorem 7.1 is S(1, k, 1).
j
k l
Fig. 8. The graph S( j, k, l)
Theorem 7.3 If G is a graph with largest eigenvalue
1
and if 2 <
1
<
(2 +
5)
1/2
, then G is one of the following:
r
T(1, 2, k), with k > 5;
r
T(1, j, k), with 2 < j < k;
r
T(2, 2, k), with 2 < k;
r
T(2, 3, 3);
r
S( j, k, l), for ( j, l) = (1, 1) and large enough k.
The exact value of large enough in this theorem has been determined (see [5]).
Finding bounds for the second largest eigenvalue
2
has also been of interest.
Obviously
2
(K
n
) = 1, for n > 1. Smith [27] showed what happens for
2
0.
Theorem 7.4
2
(G) 0 if and only if G is a complete multipartite graph.
Proof Consider the three graphs shown in Fig. 9. Each has two positive eigenval-
ues, so by Theorem3.4 none can be a subgraph of a graph G with
2
(G) 0. Since
the graph is connected, the only possibility is for G to be a complete multipartite
graph.
Graphs with
2
(G) 1canbe partiallydescribedinterms of their complements:
these have eigenvalues bounded from below by 2, or have just one eigenvalue
n
with
n
< 2. Details can be found in [8].
Graphs with
2
(G)
2 1 and
2
(G) (
5 1)/2 have also been de-

scribed in terms of minimal forbidden subgraphs that is, graphs not satisfying
the eigenvalue bound and minimal with respect to inclusion. It is known that
the possible values of
2
(G) are dense in (
_
(2 +
5, ), but nowhere dense in

(,
2 1). Further details appear in [12], [18] and [26].

There are several fascinating interrelations between
2
(G) and other areas of
mathematical interest. In particular, this eigenvalue is related to expanders and
superconcentrators [1] and to Ramanujan graphs [19].
Fig. 9.
50 Michael Doob
8. Other matrices related to graphs
There are other matrices obtained from graphs whose eigenvalues are interesting:
as with the adjacency matrix, the rows and columns correspond to the vertices
of the graph. One such is the Seidel matrix: it is 0 on the diagonal, 1 for ad-
jacent vertices, and 1 for non-adjacent vertices. Another is the Laplacian: it has
deg v
i
in the i th diagonal position, 1 for adjacent vertices, and 0 for non-adjacent
vertices. Finally, the complement of a graph has an adjacency matrix with 0 on the
diagonal, 0 for vertices adjacent in the original graph, and 1 for vertices similarly
non-adjacent.
For a regular graph G of degree r, each of these matrices can be written in the
form J +A + I. For the Seidel matrix, = 1, = 2 and = 1; for the
Laplacian, = 0, = 1 and = r; for the complement of G, = 1, = 1
and = 1. It is easy to compute the eigenvalues of these matrices; to avoid
trivialities, we assume that = 0.
Theorem 8.1 If Gis regular of degree r and has eigenvalues
1
= r >
2

n
, then the matrix J +A + I has eigenvalues n +r + and
i
+ , for
i = 2, 3, . . . n.
Proof A regular graph has r as an eigenvalue with j as an eigenvector; this vector
gives the rst eigenvalue. Any other eigenvector x of G is orthogonal to j, so
Jx = 0 and the result follows.
For non-regular graphs this theorem is no longer true. However, the slippage in
the multiplicity is at most 1.
The largest eigenvalue of a graph is called the dominant one; any other eigen-
value is called subdominant.
Theorem 8.2 Let Gbe a graph with as a subdominant eigenvalue of multiplicity
m. Then J +A + I has + as an eigenvalue with multiplicity m, where
m
1 m m
+1.
Proof Let H be the hyperplane orthogonal to vector j, and let E
be the eigenspace
of the eigenvalue . Then
dim(H + E
) +dim(H E
) = dim H +dim E
.
If E
H, then dim(H + E
) = n 1 and dim(H E
) = m
. This implies that

the argument used for the regular case still applies, and so m = m
. Otherwise, we
have dim(H + E
) = n and dim(H E
) = m
1. The argument used in the

regular case now implies that m m
1.
If C = J +A + I, then
A =
1
J +
1
C
1
I.
Using the argument of the previous paragraph on the new matrix, we get m

m 1. Hence m
1 m m
+1.
This result applies immediately to complements and to the Seidel matrix of a
non-regular graph. Although it does not apply to the Laplacian directly, we can
still say something for bipartite graphs. If G is bipartite, we let K be a variant
of the vertex-edge incidence matrix, as follows. Let (X, Y) be the partition of the
vertices. For each edge v
i
v
j
(v
i
X, v
j
Y), let e
i
e
j
be a column of K: this is
the same A
n
representation that we used to prove Theorem 6.3. Then KK
T
= L,
the Laplacian of G. All the entries in a given row have the same sign. Hence,
K
T
K = 2I +A(L(G)). Now L has 0 as a simple eigenvalue (see Section 4), and
KK
T
and K
T
K have the same non-zero eigenvalues.
Theorem 8.3 If Gis a bipartite graph with line graph L(G) and Laplacian L, then
(> 0) is an eigenvalue of L if and only if 2 (> 2) is an eigenvalue of L(G).
The validity of this theorem for paths was rst observed by Haemers [16].
9. Cospectral graphs
From the moment it was realized that different graphs could have the same spec-
trum, the hunt for such graphs was on, and many families of cospectral graphs have
been found. Broadly speaking, two types of constructions are used: one uses op-
erations on graphs (complements, products, etc.) to produce new cospectral ones,
while the other adroitly pastes different graphs together. We look at each of these.
Using graph operations
One easy way to construct cospectral graphs is due to Hoffman (see [21]). Take
two non-isomorphic cospectral regular graphs G
1
and G
2
, and consider the graph
formed by taking k copies of G
1
and s k copies of G
2
. Let H
k
be the complement
of this graph. By Theorem8.1, the graphs H
k
(k = 0, 1, . . . , s) are cospectral. Thus
we can have arbitrarily large sets of non-isomorphic cospectral graphs.
Given graphs G and H with vertex-sets {v
1
, v
2
, . . . , v
n
} and {w
1
, w
2
, . . . , w
m
},
the Cartesian product G H of G and H has as its vertices the pairs {(v
i
, w
j
) :
i = 1, 2, . . . , n, j = 1, 2, . . . , m}. Two vertices (v, w) and (v
, w
) are joined if
either v = v
and w and w
are joined in H, or w = w
and v and v
are joined in G.
We use to denote the Kronecker product of matrices. This product is associative
52 Michael Doob
and so may be extended to the product of more than two factors; for example, the
d-dimensional cube is the Cartesian product of d copies of K
2
.
It is easy to see that
A(G H) = (A(G) I) +(I A(H)).
Let x be an eigenvector of G with eigenvalue , and y be an eigenvector of H with
eigenvalue . Then, using the properties of the Kronecker product,
((A(G) I) +(I A(H)))(x y)
= (A(G) I)(x y) +(I A(H))(x y)
= (A(G)x Iy) +(Ix A(H)y)
= (x y) +(x y)
= (x y) +(x y) = ( +)(x y).
Thus, the eigenvalues of G H are just the pairwise sums of the eigenvalues of
G and H.
An attractive feature of this product is that the factorization is essentially unique
(see [22]). We may take two cospectral graphs G
1
and G
2
and form H
k
, the
Cartesian product of k copies of G
1
and s k copies of G
2
. Then the graphs
H
k
(k = 0, 1, . . . , s) are all cospectral.
The Shrikhande graph is dened in [4]. It is strongly regular with the same
parameters as L(K
4,4
). When G
1
is the Shrikhande graph and G
2
is L(K
4,4
), then
the graphs H
k
(k = 1, 2, . . . , s) are distance-regular with the same parameters,
and therefore cospectral; they are known as Doob graphs. These graphs imply that
arbitrarily large families of cospectral distance-regular graphs with large diameter
exist. Similar constructions for creating cospectral distance-regular graphs can be
found in [4].
This concept can be extended to NEPS (non-extended p-sum) graphs. For this,
we start with n graphs G
1
, G
2
, . . . , G
n
and a set of vectors B = {(
1
,
2
, . . . ,
n
) :
i
= 0 or 1}. The vertex-set of the NEPS graph is the Cartesian product of the
vertex-sets of G
1
, G
2
, . . . , G
n
, just as in the Cartesian product of graphs. Two ver-
tices are adjacent if, as n-tuples, there is a vector (
1
,
2
, . . . ,
n
) Bsothat the co-
ordinates of the n-tuples agree exactly where
i
= 1. When B = {e
1
, e
2
, . . . , e
n
},
the NEPS is the ordinary Cartesian product. If A
1
, A
2
, . . . , A
n
are the respective
adjacency matrices of G
1
, G
2
, . . . , G
n
, and Ais the adjacency matrix of the NEPS
product, then
A =
(
1
,...,
n
)B
A
1
1
A
2
2
A
n
n
.
If x
1
, x
2
, . . . , x
n
are eigenvectors of A
1
, A
2
, . . . , A
n
with eigenvalues
1
,
2
, . . . ,
n
, then
A(x
1
x
2
x
n
) =
(
1
,...,
n
)B
A
1
1
A
n
n
(x
1
x
n
)
=
(
1
,...,
n
)B
_
1
1
x
1
2
2
x
2

n
n
x
n
_
=
(
1
,...,
n
)B
1
1

2
2
. . .
n
n
(x
1
x
2
x
n
).
It follows that we can compute the eigenvalues of the NEPS graphs from the
eigenvalues of the factors. More use of the NEPS graphs will be made in Chapter 3,
Section 7.
Pasting graphs together
A second method of constructing cospectral graphs is by pasting smaller graphs
together. One way is to take two graphs, designate a special vertex (or root) in
each of them, and then identify these two vertices. We denote this new graph by
G H, the roots being understood from the context. Further, we denote by P
G
(x)
the characteristic polynomial of the adjacency matrix of G.
Suppose that {v
1
, v
2
, . . . , v
n
} is the vertex-set of G and that {w
1
, w
2
, . . . , w
m
}
is the vertex-set of H. Without loss of generality we may assume that v
1
and w
1
are
to be identied as a single vertex u. The rows of the adjacency matrix of the new
graph can be ordered by u, v
2
, v
3
, . . . , v
n
, w
2
, w
3
, . . . , w
m
. Then the adjacency
matrix has A(G {v
1
}) and A(H {w
1
}) as blocks on the diagonal.
Let B = xI A(G H). We evaluate
det B =
(1)
sgn
b
1,(1)
b
2,(2)
. . . b
m+n1,(m+n1)
,
by summing over all permutations of the vertices of G H. Let S
1
be the set of
permutations such that (u) = u or (u) is a vertex in G. Similarly, let S
2
be
the set of permutations such that (u) = u or (u) is a vertex in H. Then
S
1
(1)
sgn
b
1,(1)
b
2,(2)
. . . b
m+n1,(m+n1)
= P
G
(x)P
Hw
1
(x).
Similarly,
S
2
(1)
sgn
b
1,(1)
b
2,(2)
. . . b
m+n1,(m+n1)
= P
Gv
1
(x)P
H
(x).
This covers all the permutations, but counts those for which (u) = u twice: this
latter case gives x P
Gv
P
Hw
. In short, we have the following theorem.
54 Michael Doob
w
Fig. 10. The graph H
Theorem 9.1 If G H is formed by identifying the vertex v in G with the vertex
w in H, then
P
GH
(x) = P
G
(x)P
Hw
(x) + P
Gv
(x)P
H
(x) x P
Gv
(x)P
Hw
(x).
Schwenk [23] provides the example in Fig. 10: the graph H is to be used with
roots v and w. One may compute that
P
Hv
(x) = P
Hw
(x) = x
2
(x
2
2)(x
4
4x
2
+2).
If we take any graph G with root u, then we may form G H by identifying u
with v, or u with w, and the two resulting graphs are cospectral. If G happens to
be a tree, then the two resulting graphs are cospectral non-isomorphic trees. By
iterating this process with the same root, we can get arbitrarily large families of
cospectral trees.
Schwenk [24] has also shown that, as the number of vertices gets large, the
probability that a tree contains a copy of H approaches 1; hence almost all trees
have a cospectral mate. One may ask whether the same is true for graphs in general,
but unfortunately almost nothing is known about the answer to this question.
References
1. N. Alon and V. D. Milman,
1
-isoperimetric inequalities for graphs and superconduc-
tors, J. Combin. Theory (B) 38 (1985), 7388.
2. N. Biggs, Intersection matrices for linear graphs, Combinatorial Mathematics and its
Applications (ed. D. J. A. Welsh), Academic Press (1971), 1523.
3. N. Bourbaki, Groupes et Alg` ebres de Lie, Masson, 1981.
4. A. E. Brouwer, A. M. Cohen and A. Neumaier, Distance-Regular Graphs, Springer-
Verlag, 1989.
5. A. E. Brouwer and A. Neumaier, Graphs with spectral radius between 2 and
_
2 +
5,
Linear Alg. Appl. 114/115 (1989), 273276.
6. P. J. Cameron, J. M. Goethals, J. J. Seidel and E. E. Shult, Line graphs, root systems,
and elliptic geometry, J. Algebra 43 (1976), 305327.
7. R. W. Carter, Simple Groups of Lie Type, Wiley, 1989.
8. D. Cvetkovi c, On graphs whose second largest eigenvalue does not exceed 1, Publ.
Inst. Math. (Beograd) 31 (1982), 1520.
9. D. Cvetkovi c, M. Doob and I. Gutman, On graphs whose spectral radius does not exceed
(2 +
5)
1/2
, Ars Combinatorica 14 (1982), 225239.
10. D. Cvetkovi c, M. Doob and H. Sachs, Spectra of Graphs (3rd ed.), Johann Ambrosius
Barth, 1995.
11. D. Cvetkovi c, M. Doob and S. Simi c, Generalized line graphs, J. Graph Theory 5
(1981), 385399.
12. D. Cvetkovi c and S. Simi c, On graphs whose second largest eigenvalue does not exceed
(
5 1)/2, Discrete Math. 139 (1995), 213227.

13. M. Doob and D. Cvetkovi c, On spectral characterizations and embeddings of graphs,
Linear Alg. Appl. 27 (1979), 1726.
14. F. R. Gantmacher, The Theory of Matrices, Chelsea, 1959.
15. C. Godsil and G. Royle, Algebraic Graph Theory, Springer-Verlag, 2001.
16. W. Haemers, Which graphs are determined by their spectrum?, Linear Alg. Appl.
17. A. J. Hoffman and R. R. Singleton, On Moore graphs with diameter 2 and 3, IBM J.
Res. Develop. 4 (1960), 497504.
18. J. Li, Subdominant Eigenvalues of Graphs, PhD thesis, University of Manitoba, 1994.
19. A. Lubotsky, R. Phillips and P. Sarnak, Ramanujan graphs, Combinatorica 8 (1988),
261277.
20. H. Minc and M. Marcus, A Survey of Matrix Theory and Matrix Inequalities, Prindle,
Weber & Schmidt, 1964.
21. A. Mowshowitz, The characteristic polynomial of a graph, J. Combin. Theory (B) 12
(1972), 177193.
22. R. C. Read and R. J. Wilson, An Atlas of Graphs, Oxford Univ. Press, 1998.
23. G. Sabidussi, Graph multiplication, Math. Zeit. 72 (1960), 446457.
24. A. J. Schwenk, Almost all trees are cospectral, New Directions in the Theory of Graphs
(ed. F. Harary), Academic Press (1973), 275307.
25. A. J. Schwenk and R. J. Wilson, On the eigenvalues of a graph, Selected Topics in Graph
Theory (eds. L. W. Beineke and R. J. Wilson), Academic Press (1979), 307336.
26. S. Simi c, Some notes on graphs whose second largest eigenvalue is less than
(
5 1)/2, Linear and Multilinear Algebra 39 (1995), 5971.

27. J. H. Smith, Some properties of the spectrum of a graph, Combinatorial Structures and
their Applications (eds. R. Guy et al.), Gordon and Breach (1970), 403406.
2
Graphs and matrices
RICHARD A. BRUALDI and BRYAN L. SHADER
1. Introduction
2. Some classical theorems
3. Digraphs
4. Biclique partitions of graphs
5. Bipartite graphs
6. Permanents
7. Converting the permanent into the determinant
8. Chordal graphs and perfect Gaussian elimination
9. Ranking players in tournaments
References
Graphs and matrices enjoy a mutually benecial relationship. On one level
this relationship provides different viewpoints and conceptualizations of the
same idea: but it is more than the simple translation of results in one frame-
work into the other. Matrices, and the associated linear algebra, provide
an important tool in graph theory, sometimes the only known tool. On the
other hand, graphs provide a way to understand and take advantage of ma-
trix structure. In this chapter we provide a glimpse of graphs and matrices
working together.
1. Introduction
Matrices provide models for graphs (or models for the same phenomena mod-
elled by graphs) that often illuminate their structure and allow the application of
elementary but powerful linear algebraic techniques to their study. Conversely,
associated with a matrix are graphs that capture in various ways the combinatorial
structure of the matrix. Different graphs may reveal or emphasize different aspects
of a matrixs structure. Taking this structure into account often leads to sharper
theorems, improved inequalities and, sometimes, non-existence conclusions.
56
2 Graphs and matrices 57
Let G be a graph (or a multigraph, general graph, digraph, weighted di-
graph, . . . ). The basic idea behind matrices associated with graphs is the follow-
ing. Let the vertices of G be listed in some order say, v
1
, v
2
, . . . , v
n
; the par-
ticular ordering is of no signicance, but we have to choose one. We dene a
matrix A = (a
i j
) of order n, where the ij-entry is associated with the ordered pair
of vertices (v
i
, v
j
). If G is a graph, then
a
i j
=
_
1, if v
i
and v
j
are joined by an edge
0, otherwise.
As we saw in Chapter 1, the matrix A is the adjacency matrix of G and is a
symmetric (0, 1)-matrix with zeros on the main diagonal.
It is an easy matter to modify the denition of A to take into account these
different variations of the graph G. If G is a multigraph, then we let a
i j
be the
multiplicity of the edge joining v
i
and v
j
, and the adjacency matrix A is a non-
negative integer matrix with zeros on the main diagonal. If loops are permitted,
then Ais a general non-negative integer matrix. If Gis a digraph so that edges have
a direction, then a
i j
does not equal a
j i
in general, and A is a general (0, 1)-matrix,
or a general non-negative integer matrix if edges are allowed to have multiplicity
greater than 1. Finally, if G is a weighted graph (or digraph) where edges have
weights, then we let a
i j
be the weight (positive, negative, zero or complex) of the
edge from v
i
to v
j
.
Now we have closed the connection between graphs and matrices in that A can
be an arbitrary real or complex matrix. Moreover, we can reverse our constructions:
starting with a square matrix A we can dene a graph of one of the types discussed
above whose adjacency matrix is A. So, in a sense, weighted graphs and matrices
are interchangeable; for more on this theme, see [4].
As examples, consider the matrices
A =
_
_
_
_
_
_
_
0 1 1 0 1
1 0 1 0 0
1 1 0 1 1
0 0 1 0 1
1 0 1 1 0
_
_
_
_
_
_
_
and B =
_
_
_
_
_
_
_
0 1 1 0 2
0 1 0 2 0
3 1 0 1 1
0 1 1 0 1
1 0 1 1 1
_
_
_
_
_
_
_
.
The matrix A is a symmetric (0, 1)-matrix of order 5 with zeros on the main
diagonal. Thus A is the adjacency matrix of a graph with 5 vertices. The matrix
B is a non-negative integer matrix with maximum entry equal to 3, and is the
adjacency matrix of a digraph each of whose edges has multiplicity at most 3 (or
a weighted digraph with weights 0, 1, 2 and 3).
58 Richard A. Brualdi and Bryan L. Shader
2. Some classical theorems
In this section we briey mention some of the classic results that illustrate the
powerful and beautiful relationship between graphs and matrices.
As we sawin Chapter 1, one of the earliest theorems to exploit the combinatorial
structure of a matrix was the Perron-Frobenius theorem, concerning the spectral
structure of a square non-negative matrix A (see [16] and [22]). Without any
assumptions on A, the following theorem holds.
Theorem 2.1 (Perron-Frobenius theorem) Let A be a non-negative matrix of
order n. Then A has a non-negative eigenvalue such that every eigenvalue
of A satises || . Moreover, A has an entry-wise non-negative eigenvector x
corresponding to the eigenvalue .
The spectral radius of a complex square matrix is the largest modulus of its
eigenvalues. Thus, the number in Theorem 2.1 is the spectral radius of the
matrix A.
Let Gbe the weighted digraph with adjacency matrix A. One particular assump-
tion on the digraph G leads to more detail concerning the eigenvalue structure of
A. Recall that a digraph is strongly connected if, for each ordered pair (v
i
, v
j
) of
distinct vertices, there is a directed path in G from v
i
to v
j
.
Theorem 2.2 Let Abe a non-negative matrix of order n 2, such that the digraph
G corresponding to A is strongly connected. Then the spectral radius of A is
positive and has multiplicity 1; in addition, A has an entry-wise positive eigenvec-
tor x corresponding to the eigenvalue . The eigenvalues of A with || = are
e
2i j/k
( j = 1, 2, . . . , k), where k is the greatest common divisor of the lengths
of the cycles of G. The only non-negative eigenvectors of A are positive multiples
of x.
Let x = (x
1
, x
2
, . . . , x
n
)
T
. The equation Ax = x implies that
n
j =1
n
i =1
a
i j
x
j
=
n
j =1
x
j
,
showing that is a weighted average of the column sums s
1
, s
2
, . . . , s
n
of A. Hence
min {s
1
, s
2
, . . . , s
n
} max {s
1
, s
2
, . . . , s
n
}.
In particular, this equation implies that if A is a (0, 1)-matrix corresponding to a
digraph G, then the spectral radius of A lies between the minimum and maximum
of the in-degrees of G. Since A and A
T
have the same eigenvalues, a similar
conclusion holds with in-degree replaced by out-degree.
A recent application of the digraph of a complex matrix A = (a
i j
) of order
n to the localization of eigenvalues of A originates in Ger sgorins theorem. This
theoremasserts that the eigenvalues of Alie in that part of the complex plane given
by the union
n
_
i =1
{z C : |z a
i i
| R
i
}
of n closed discs, where R
i
=
j =i
|a
i j
| (i = 1, 2, . . . , n). If Ais irreducible, then
a boundary point of this region can be an eigenvalue of A only if it is a boundary
point of each of the n discs. A better inclusion region takes into account the cycles
of G (see [3] and [7]).
Theorem 2.3 Let A = (a
i j
) be a complex matrix of order n with corresponding
digraph G. The n eigenvalues of A lie in that part of the complex plane given by
the union
_
_
z C :
|z a
i i
|
R
i
_
. (1)
(Here
denotes the union over all cycles of G, and
denotes the product

over those i for which v
i
is a vertex of .) If G is strongly connected, then a
boundary point of (1) can be an eigenvalue of A only if it is a boundary point of
each of the regions in its denition.
Let G be a bipartite graph. Recall that the vertices of G can be partitioned into
two sets Xand Y such that each edge has one vertex in Xand one vertex in Y; the pair
{X, Y} is a bipartition of the vertices of G. We choose a listing x
1
, x
2
, . . . , x
r
for
the vertices in X and a listing y
1
, y
2
, . . . , y
s
for the vertices in Y, and concatenate
these to get a listing x
1
, x
2
, . . . , x
r
, y
1
, y
2
, . . . , y
s
of the vertices of G. Then the
adjacency matrix A of G has the form
A =
_
O
r
B
B
T
O
s
_
where B is an r s matrix, and O
r
and O
s
are zero matrices of orders r and s,
respectively. The matrix B is called the biadjacency matrix of G, and depends on
the listing chosen for the vertices in X and Y. Using our previous observations, we
deduce that bipartite graphs and (rectangular) (0, 1)-matrices are interchangeable,
as are bipartite multigraphs and (rectangular) non-negative integral matrices, and
weighted bipartite graphs and real or complex (rectangular) matrices. Adopting
matrix terminology, we say that the bipartite graph is r s, and is square if r = s.
If the bipartite graph Gis connected, then the bipartition {X, Y} is unique, and thus
squareness is an intrinsic property of G. If G is not connected, then it has more
than one bipartition, some of which may satisfy r = s.
A matching M in a graph G is a collection of independent edges. It is a perfect
matching provided that each vertex of G belongs to some edge in M. Thus, a graph
with an odd number of vertices cannot have a perfect matching. The graph G is
matching covered provided that each edge belongs to at least one perfect matching.
An m n bipartite graph has a perfect matching only if it is square.
Let G be a square bipartite graph. To each edge of G we assign a weight,
where the weights are algebraically independent numbers over the rational eld
Q. We call the biadjacency matrix A of this weighted bipartite graph a generic
biadjacency matrix over Q. A theorem of Frobenius (see [16] and [17]) connects
the structure of a bipartite graph and the determinant of its biadjacency matrix.
Theorem 2.4 Let A be a generic biadjacency matrix corresponding to a square
bipartite graph G. Then G is connected and matching covered if and only if det A
is an irreducible polynomial in the polynomial ring obtained from Q by adjoining
the non-zero elements of A.
Another connection between bipartite graphs and matrices concerns scaling the
rows and columns of a non-negative matrix so that all row and column sums are 1
(see [6] and [43]). Such a matrix is doubly stochastic.
Theorem 2.5 Let Abe a real matrix of order n corresponding to a square bipartite
graph G. Then G is matching covered if and only if there exist positive diagonal
matrices D
1
and D
2
such that D
1
AD
2
is doubly stochastic.
With any multigraph Gwe can also associate a matrix that records the incidences
between the vertices and edges. Let v
1
, v
2
, . . . , v
n
be the vertices of G in some
order, and let e
1
, e
2
, . . . , e
m
be the edges in some order. We arbitrarily assign an
orientation to each edge from one of its vertices (the tail) to the other (the head).
The (oriented) incidence matrix of G, corresponding to the given orientation of G,
is the n m matrix B = (b
i j
) in which
b
i j
=
_
_
_
1, if v
i
is the tail of e
j
1, if v
i
is the head of e
j
0, otherwise.
Each column of B contains exactly one 1 and exactly one 1. Note that BB
T
=
D A, where A is the adjacency matrix of G and D is the diagonal matrix whose
diagonal entries are the degrees of the vertices of G. The matrix BB
T
is the
Laplacian matrix L of G. The following theorem relates the connectedness of a
graph to the algebraic properties of its Laplacian matrix (see Chapter 4 and [7]).
Theorem 2.6 Let Gbe a multigraph with n vertices. Then the rank of the Laplacian
matrix L is at most n 1, with equality if and only if G is connected. If G is
connected, thenthe number of spanningtrees of Gis the modulus of the determinant
of any submatrix of L of order n 1.
We conclude this section by remarking that Seidel [39] has introduced another
matrix into the study of graphs. This is the (0, 1, 1)-matrix S = J
n
I
n
2A,
where A is the adjacency matrix of a graph G with n vertices and J
n
is the all-1
matrix of order n. The matrix S can be obtained from A by replacing each 1 by
1 and each 0 not on the main diagonal by 1. This matrix has been useful in the
study of strongly regular graphs (see Chapter 8).
3. Digraphs
We can viewany graph Gas a digraph, by regarding each edge joining two vertices
as two arcs, one in each direction. If A is the adjacency matrix of G, then A is also
the adjacency matrix of G viewed as a digraph. It is for this reason that we focus
on digraphs in this section.
Let A = (a
i j
) be a matrix of order n, and let G be the weighted digraph corre-
sponding to A. The weight of a directed walk is the product of the weights of its
arcs. A directed v
i
-v
j
walk is a directed walk in G from vertex v
i
to vertex v
j
.
From the denition of matrix multiplication we deduce that, for each positive
integer k, the ij-entry of A
k
is the sum of the weights of all the directed v
i
-v
j
walks of length k in G. If A is a (0, 1)-matrix, then the i j -entry of A is the number
of directed v
i
-v
j
walks of length k. In particular, there is a directed v
i
-v
j
walk
of length k if and only if the i j -entry of A
k
is not 0. This last conclusion holds
whenever Ais a non-negative matrix, an assumption that prevents any cancellation
in summing products of weights.
Let A be a non-negative matrix. Then A is primitive if there is a positive integer
k for which each element of A
k
is positive that is, A
k
is a positive matrix. Note
that if A
k
is positive, then so is A
l
for each integer l > k. It follows from the above
discussion that A is primitive if and only if there exists a positive integer k for
which there are directed v
i
-v
j
walks of length k, for each i and j .
Let P be a permutation matrix of order n, corresponding to the permutation
i
1
, i
2
, . . . , i
n
of {1, 2, . . . , n}. Thus P has 1s in positions ( j, i
j
) (for 1 j n)
and 0s elsewhere. The matrix PAP
T
is obtained from A by permuting rows so
that they are in the order i
1
, i
2
, . . . , i
n
, and then permuting columns so that they
are also in the order i
1
, i
2
, . . . , i
n
: thus, the matrix PAP
T
is obtained from A
by simultaneous permutations of its rows and columns. Note that A and PAP
T
represent the same digraph G, but with a different ordering of its vertices.
The matrix A is reducible if there exists a permutation matrix P such that
PAP
T
=
_
A
1
O
A
21
A
2
_
, (2)
where A
1
and A
2
are matrices of positive orders n
1
and n
2
. Assume that A is
reducible and (2) holds. Then the vertex-set V of G can be partitioned into two
non-empty subsets V
1
and V
2
of sizes n
1
and n
2
, so that all arcs between V
1
and V
2
are directed from V
2
to V
1
, and thus (V
2
, V
1
) is a directed cut of G. Conversely, a
directed cut of G implies that PAP
T
has the formgiven in (2) for some permutation
matrix P. Thus A is reducible if and only if the digraph G has a directed cut,
and irreducible otherwise. Irreducibility of a matrix is equivalent to the strong
connectivity of its digraph. We observe that a digraph is strongly connected if and
only if it has no directed cut.
Dening v w on the vertex-set V to mean that G has a directed v-w walk
and a directed w-v walk, we obtain an equivalence relation that partitions V into
equivalence classes V
1
, V
2
, . . . , V
k
. Each induced digraph G[V
i
] is a maximal
strongly connected subdigraph of G, called a strong component of G. Let G
be
the digraph whose vertices are the sets V
1
, V
2
, . . . , V
k
, with an arc from V
i
to V
j
if and only if there is an arc in G from some vertex in V
i
to some vertex in V
j
(i = j ). Then G
cannot have any closed directed walks. It follows that the sets
V
1
, V
2
, . . . , V
k
can be ordered as V
i
1
, V
i
2
, . . . , V
i
k
in such a way that there is an
arc from V
i
p
to V
i
q
in G
only if 1 q < p k. Ordering the vertices of V so that

those in V
i
1
come rst, then those in V
i
2
, V
i
3
, . . . , V
i
k
, we deduce that there is a
permutation matrix P for which
PAP
T
=
_
_
_
_
_
A
1
0 0
A
21
A
2
0
.
.
.
.
.
.
.
.
.
.
.
.
A
k1
A
k2
A
k
_
_
_
_
_
, (3)
where A
1
, A
2
, . . . , A
k
are irreducible matrices. The matrices A
1
, A
2
, . . . , A
k
,
which correspond to the strong components of the digraph G, are uniquely de-
termined up to simultaneous permutations of their rows and columns. The matrix
in (3) is called the Frobenius normal form of A, and A
1
, A
2
, . . . , A
k
are called
the irreducible components of A. In general, the Frobenius normal form of A is
not unique. The uniqueness depends on the digraph G
that is, on the zero or

non-zero character of the subdiagonal blocks A
i j
. If, for example, all the subdiag-
onal blocks are zero matrices, then the irreducible components can appear in any
order along the diagonal of (3).
If G is a graph, then the matrix A is symmetric, and (3) reduces to
PAP
T
=
_
_
_
_
_
A
1
0 0
0 A
2
0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 A
k
_
_
_
_
_
,
where A
1
, A
2
, . . . , A
k
are symmetric irreducible matrices that correspond to the
connected components of G. Thus the symmetric matrix A is irreducible if and
only if G is connected.
Now assume that A is an irreducible matrix of order n that is, G is a strongly
connected digraph. The index of imprimitivity of A and of G is the greatest
common divisor of the lengths of the cycles of G, or equivalently, of the lengths of
the closed directed walks in G. Since G is strongly connected, it must have a cycle
if n > 1. The properties of strongly connected digraphs collected in the follow-
ing lemma are straightforward to verify and lead to a normal form for irreducible
matrices.
Lemma 3.1 Let G be a strongly connected digraph with index of imprimitivity d.
Then
r
d is the greatest common divisor of the lengths of the closed directed walks
containing any particular vertex;
r
for each pair of vertices v and w, the lengths of all directed vw-walks are
congruent modulo d;
r
the vertex-set of G can be uniquely partitioned into d non-empty sets
U
1
, U
2
, . . . , U
d
in such a way that each arc of G is directed from some U
i
to
U
i +1
(i = 1, 2, . . . , d; U
d+1
= U
1
), and the length of each directed walk from a
vertex in U
i
to a vertex in U
j
is congruent to j i (modulo d), for 1 i, j d.
The sets U
1
, U
2
, . . . , U
d
in Lemma 3.1 are called the imprimitivity sets of the
strongly connected digraph G. If |U
i
| = n
i
(i = 1, 2, . . . , d), then there is a per-
mutation matrix P for which
PAP
T
=
_
_
_
_
_
_
_
0
n
1
A
12
0 0
0 0
n
2
A
23
0
.
.
.
.
.
.
.
.
.
.
.
.
0
0 0 0 A
d1,d
A
d1
0 0 0
n
d
_
_
_
_
_
_
_
. (4)
The following lemma is due to Schur (see [7] for a proof).
Lemma 3.2 A non-empty set of positive integers that is closed under addition
contains all but a nite number of positive multiples of its greatest common divisor.
We next show that primitive matrices are exactly those irreducible matrices
with index of imprimitivity 1.
Theorem 3.3 Let G be a strongly connected digraph with at least two vertices.
Then the index of imprimitivity of G is 1 if and only if there is an integer N such
that, for any two vertices v
i
and v
j
, there are directed v
i
-v
j
walks of each length
l N; equivalently, an irreducible non-negative matrix A is primitive if and only
if its index of imprimitivity is 1.
Proof Let d be the index of imprimitivity of G and assume that there is an integer
N such that, for any two vertices v
i
and v
j
, there is a directed v
i
-v
j
walk of
each length l N. Then there are directed v
1
-v
2
walks and directed v
2
-v
1
walks
of lengths N and N +1. Hence G has closed directed walks of lengths 2N and
2N +1, and so d = 1.
Now assume that d = 1, and let c
1
, c
2
, . . . , c
p
be the distinct lengths of the
cycles of G. The strong connectivity of G implies that, for any two vertices v
i
and v
j
i
-v
j
walk that meets a cycle of each of the lengths
c
1
, c
2
, . . . , c
p
. By Lemma 3.2, there are directed v
j
-v
i
walks of length l, for every
sufciently large integer l.
Now suppose that A is a non-negative irreducible matrix. Since the i j -entry of
A
l
is positive if and only if there is a directed v
i
-v
j
walk in the associated digraph,
the theorem follows.
If G is a connected graph with symmetric adjacency matrix A, then viewing
G as a digraph, we see that G has a directed cycle of length 2; thus the index of
imprimitivity of A is 1 or 2, and A is primitive if and only if G has a cycle of odd
length that is, G is not bipartite.
In our discussion of primitive non-negative matrices, we may restrict ourselves
to (0, 1)-matrices. Let A be a primitive (0, 1)-matrix of order n. The exponent of
A, denoted by exp A, is the smallest positive integer t for which A
t
is positive.
In terms of primitive digraphs, the exponent is the smallest positive integer t for
which there is a directed v
i
v
j
-walk of length t, for all i and j.
There has been considerable work done on determining exponents and related
parameters. The following is a theorem of Dulmage and Mendelsohn [12].
Theorem 3.4 Let A be a primitive (0, 1)-matrix of order n, and let l be the length
of a shortest cycle in its digraph G. Then
exp A n +l(n 2).
Proof Consider the matrix A
l
and its digraph G
l
. There is an arc in G
l
fromvertex
v
i
to vertex v
j
exactly when there is a directed v
i
-v
j
walk of length l in G. Let L
be the set of vertices of G
l
that have loops. Then L has at least l elements, and
the loops at vertices of L imply that there is a directed walk in G
l
of length n 1
from each vertex in L to each vertex v
j
. Hence in G there is a directed walk of
length l(n 1) from each vertex in L to each vertex v
j
. For each vertex v
i
there is
a directed walk in G from v
i
to some vertex in L of length l
i
n l, and hence
a directed v
i
-v
j
walk of length n l +l(n 1) = n +l(n 2). Hence A
n+l(n2)
is a positive matrix.
Since a primitive digraph with n 2 vertices must have a cycle of length
l n 1, we get the following result of Wielandt [47].
Corollary 3.5 The maximum exponent of a primitive matrix of order n ( 2) is
(n 1)
2
+1.
Proof It follows fromTheorem3.4 that the exponent of a primitive matrix of order
n does not exceed (n 1)
2
+1. The digraph obtained from a cycle of length n by
inserting a chord between some pair of vertices at distance 2 on the cycle can be
shown to have exponent (n 1)
2
+1.
It can also be shown that, up to isomorphism, the digraph in the proof is the unique
digraph with n vertices with exponent equal to (n 1)
2
+1 (see, for example,
[7]). The maximum exponent of a primitive symmetric (0, 1)-matrix A of order
n ( 2) equals 2n 2, with equality if and only if the graph of A is a path with a
loop at one end (see [42]).
The notion of primitivity has been generalized (see [15]). Acoloured digraph is
a digraph G in which each arc is labelled by one of a set {c
1
, c
2
, . . . , c
t
} of colours.
Given a directed walk in G, we let denote the column vector whose i th coordi-
nate is the number of arcs of with colour c
i
. Fix one vertex v, and dene a relation
on the vertices of G by v
i
v
j
whenever there exist directed walks
i
and
j
from
v to v
i
and v to v
j
, respectively, for which
i
=
j
. It is easy to verify that, if G is
strongly connected, this relation is an equivalence relation and is independent of v.
The coloured digraph G is called primitive if there is exactly one equivalence class
for . Let
1
,
2
, . . . ,
s
be the cycles of G, let L be the subgroup of Z
t
generated
by {
i
: i = 1, 2, . . . , s}, and let M be the t s matrix whose i th column is
i
.
Theorem 3.6 Let G be a strongly connected coloured digraph with n vertices and
at least one arc of each of t colours. Then the following statements are equivalent:
(1) The coloured digraph G is primitive.
(2) L = Z
t
.
(3) The greatest common divisor of the determinants of the t t submatrices of
M is 1.
(4) There exists a non-negative integer vector h = (h
1
, h
2
, . . . , h
t
)
T
such that,
for any two vertices v
i
and v
j
i
-v
j
walk with h
k
arcs of
colour c
k
, for k = 1, 2, . . . , t .
Proof The theorem follows from the truth of the implications (1) (2), (2)
(3), (2) (4), and (4) (1).
First assume that (1) holds. Let e
i
be the t 1 vector whose i th entry is 1
and whose other entries are 0. The digraph D has an arc vw of colour c
i
. Since
D is primitive, there exist a directed v-w walk and a directed w-v walk
for which =

. Since each closed directed walk of G can be decomposed
into cycles, both and

+e
i
belong to L. Hence, e
i
L for all i, and (2)
holds.
The equivalence of (2) and (3) is a standard result concerning integer matrices,
and can be found in [32].
To show that (2) (4), assume that (2) holds. Since G is strongly con-
nected, there is a closed directed walk that passes through each vertex of D.
Since L = Z
t
, there exist s 1 integer vectors x
( j )
such that Mx
( j )
= e
j
( j =
1, 2, . . . , t ).
Let x be a non-negative integer vector such that x nx
( j )
(entry-wise) for
j = 1, 2, . . . , t, and let h = +t Mx. We claim that, for any two vertices v
and w, there exists a directed v-w walk with = h. To see this, let p be a
path from v to w with p = ( p
1
, p
2
, . . . , p
t
)
T
. Then p = M(
t
j =1
p
j
x
( j )
), and
hence
h = + p +M
_
t
j =1
(x p
j
x
( j )
)
_
.
Note that, for each i , the i th entry n
i
of
t
j =1
(x p
j
x
( j )
) is a non-negative integer.
Thus, the directed walk that starts at v, follows back to v, along the way goes
n
i
times around
i
(i = 1, 2, . . . , t ), and then follows p to w, has = h. Hence
(4) holds.
Clearly (4) implies (1).
The equivalence of (1) and (4) implies that if t = 1, so all arcs are coloured the
same, then G is primitive as a coloured graph if and only if it is primitive as an
uncoloured graph. The notion of exponents of coloured digraphs is introduced and
studied in [41].
4. Biclique partitions of graphs
We next illustrate how elementary but powerful linear algebra can be used in
the study of questions that concern the partitioning of the edges of a graph into
certain types of subgraphs. A biclique partition of a graph G is a partition of
the edge-set of G into complete bipartite subgraphs. More formally, we make the
following denitions. A biclique of a graph G is a subgraph whose edges form a
complete bipartite subgraph. Given two disjoint subsets X and Y, B(X, Y) denotes
the biclique consisting of all edges joining a vertex in X and a vertex in Y. A
biclique partition of G is a collection
B(X
1
, Y
1
), B(X
2
, Y
2
), . . . , B(X
k
, Y
k
) (5)
of bicliques whose edges partition the edges of G. Since a single edge can form a
biclique, every graph has a biclique partition. The biclique partition number bp(G)
of G is the smallest number of bicliques that partition G.
Let A be the adjacency matrix of G. Then it is easy to verify that, for a biclique
partition (5) of G,
A = XY
T
+YX
T
= (X Y)
_
Y
T
X
T
_
,
where X is the n k matrix whose i j -entry is 1 if v
i
X
j
, and 0 otherwise, and
Y is dened analogously. Thus, a partition of G into k bicliques corresponds to a
special type of factorization of A as the product of an n 2k (0, 1)-matrix and a
related 2k n (0, 1)-matrix.
The matrix XY
T
is the adjacency matrix of the digraph obtained from G by
orientating each edge in each B(X
i
, Y
i
) from its vertex in X
i
to its vertex in Y
i
.
Assume that G is the complete graph. A tournament is a digraph obtained by
choosing an orientation for each edge of the complete graph G. Hence, partitions
of the complete graph on n vertices into k bicliques correspond to factorizations
M = XY
T
of the adjacency matrix M of a tournament as the product of an n k
(0, 1)-matrix and a k n (0, 1)-matrix. The following lemma [11] implies that
adjacency matrices of tournaments have special algebraic properties. Additional
properties can be found in [30] and [40].
Lemma 4.1 Let M be the adjacency matrix of a tournament on n vertices. Then
the rank of M is at least n 1.
Proof Let j be the n 1 all-1 vector, and let Nbe the (n +1) n matrix obtained
from M by appending j
T
at the bottom. It sufces to prove that the columns of N
are linearly independent.
Suppose that Nx = 0. Then Mx = 0 and j
T
x = 0. Since M is the adjacency
matrix of a tournament, M+M
T
= J
n
I
n
. Since Mx =0 and x
T
j = 0, we have
0 = x
T
(M+M
T
)x = (x
T
j)
2
x
T
x = x
T
x.
Hence x = 0, and so the columns of N are linearly independent.
Theorem 4.2 The biclique partition number of K
n
is n 1.
Proof A partition of K
n
into k bicliques corresponds to a factorization of the
adjacency matrix M of a tournament on n vertices as the product of an n k
matrix and a k n matrix. The bound bp(K
n
) n 1 follows, since the rank of
M is bounded above by k and below by n 1. Since K
n
can be partitioned into
n 1 stars, bp(K
n
) = n 1.
This theorem, due to H. S. Witsenhausen, is known as the Graham-Pollak
theoremand has several different proofs (see [20], [44], [34] and [9]). Interestingly,
each of these proofs is linear-algebraic in nature. More generally, we have the
following result (see [20]).
Theorem 4.3 Let G be a graph with adjacency matrix A. Then
bp(G) max{n
(A), n
+
(A)},
where n
(A) and n
+
(A) are the number of negative and positive eigenvalues of A.
Proof As argued in [21], the interlacing inequalities for eigenvalues of symmetric
matrices (see Chapter 1 and [22]) imply that n
+
and n
are subadditive functions

that is, for symmetric matrices C and D of the same order,
n
+
(C +D) n
+
(C) +n
+
(D) and n
(C +D) n
(C) +n
(D).
Let A =XY
T
+YX
T
, where X and Y are n k matrices with j th columns

X
j
and

Y
j
. Then
A =
k
j =1
_
X
j
Y
T
j
+

Y
j
X
T
j
_
.
Since, for each j ,
n
+
_
X
j
Y
j
T
+

Y
j
X
j
T
_
= n
X
j
Y
j
T
+

Y
j
X
j
T
_
= 1,
subadditivity implies that n
+
(A) k and n
(A) k. The theorem now follows

fromthe correspondence between partitions of G into k bicliques and factorizations
of A as A = XY
T
+YX
T
.
Note that n
(J
n
I
n
) = n 1 and n
+
(J
n
I
n
) = 1, and hence Theorem 4.3
also shows that bp(K
n
) n 1.
5. Bipartite graphs
As we observed in Section 1, bipartite graphs and rectangular matrices are inter-
changeable. The biadjacency matrix of an r s bipartite graph with bipartition
{X, Y} is an r s matrix. Properties of rectangular matrices can therefore be for-
mulated in terms of bipartite graphs; conversely, the structure of bipartite graphs
can be useful in investigations of rectangular matrices or non-symmetric square
matrices.
Let G be an r s bipartite graph with bipartition {X, Y}, and let A be its
biadjacency matrix. The rows and columns of A are indexed by the elements of
X and Y, respectively, and the edges of G correspond to 1s in A. Given (possibly
empty) subsets U of X and V of Y, the submatrix of A whose row indices are in
U and whose column indices are in V is denoted by A[U, V]; this submatrix is
the biadjacency matrix of the induced subgraph G[U, V] of G on the vertex-set
U V. It is a zero submatrix if and only if each edge of G contains a vertex of
X\U or of Y\V that is, if and only if (X\U) (Y\V) covers the edges of G.
A matching of G corresponds to a collection of 1s of A with no two 1s in the
same row or column. The matching number m(G) is the largest number of edges
in a matching of G. Since the edges of a matching are pairwise vertex-disjoint, the
number of edges in a matching of G does not exceed the cardinality of each subset
of vertices that covers the edges of G. Hence,
m(G) min{r +s (k +l)}, (6)
where the minimum is taken over all pairs (k, l) of non-negative integers for
which A has a k l zero submatrix. Here we consider the empty 0 s and r 0
submatrices as zero submatrices. K onigs theorem asserts that equality holds in
(6).
Theorem 5.1 Let G be an r s bipartite graph with biadjacency matrix A. Then
m(G) equals the minimum of r +s (k +l), taken over all pairs of integers (k, l)
for which A has a k l zero submatrix.
Proof Let d(A) denote the maximum sum of the dimensions of a zero submatrix
of A. It follows from the discussion preceding the theorem that
m(G) (r +s) d(A).
To complete the proof, we show by induction on r +s that G has a matching with
r +s d(A) edges. This is clear if r = 1, s = 1, or A = O, so we assume that
r, s 2 and A = O.
First, suppose that Ahas a zero submatrix A[U, V] for which |U| +|V| = d(A)
and both U and V are non-empty. Each p q zero submatrix of A[X\U, Y] can
be extended to a ( p +|U|) q zero submatrix of A. Hence, by the denition of
d(A),
p +q d(A) |U|.
It follows from the inductive hypothesis that G[X\U, Y] has a matching M
1
with
r |U| edges. Similarly, G[U, Y\V] has a matching M
2
with s |V| edges.
Hence, M
1
M
2
is a matching of G with r +s |U| |V| = r +s d(A)
edges.
Next, suppose that each zero submatrix A[U, V] of A with |U| +|V| = d(A)
has either U = or V = . Without loss of generality we may assume that the
top-left entry of A is 1. Let B = A[X\{x
1
}, Y\{y
1
}]. Since each zero submatrix
of A whose dimensions sum to d(A) is vacuous, the dimensions of each zero
submatrixof B sumtoat most d(A) 1. Hence, byinduction, the inducedsubgraph
G[X\{x
1
}, Y\{y
1
}] has a matching M
of size (r 1) +(s 1) (d(A) 1) =

r +s d(A) 1. Thus, M
along with the edge joining x

1
y
1
is a matching of G
with r +s d(A) edges.
Now assume that G is an r r bipartite graph that is, G is square. A perfect
matching of G corresponds to a permutation matrix P with P A (entry-wise).
K onigs theoremimplies that G has a perfect matching if and only if the dimensions
of each zero submatrix of A sum to at most r.
Suppose further that G is a d-regular graph with d > 0. Then each row and
column of A has precisely d 1s. Consider a zero submatrix A[U, V] of A. Then
the sum (r |V|)d of the elements of A in the columns indexed by Y\V is at
least the sum |U|d of the elements of A in the rows indexed by U. It follows that
|U| +|V| r, and hence that G has a perfect matching M. Removing the edges
of M results in a (d 1)-regular bipartite graph. Hence, by induction, the edges of
the d-regular bipartite graph G can be partitioned into d perfect matchings: such
a partitioning is called a 1-factorization of G. In matrix terms, a square matrix
(0, 1)-matrix A with constant row and column sums is the sum of permutation
matrices.
Theorem 5.1 also implies the following fundamental result of Birkhoff about
doubly stochastic matrices.
Theorem 5.2 The set of all doubly-stochastic matrices of order n is the convex
hull of the set of permutation matrices.
Proof The set of doubly stochastic matrices (regarded as a subset of R
n
2
) is closed
under convex combinations, and contains the permutation matrices of order n.
Let S = (s
i j
) be an n n doubly-stochastic matrix, and let G be the n n
bipartite graph with an edge joining x
i
and y
j
if and only if s
i j
= 0. We prove that
S is a convex combination of permutation matrices, by induction on the number
of non-zero entries of S. Theorem 5.1 implies G has a perfect matching
M = {x
1
y
i
1
, x
2
y
i
2
, . . . , x
m
y
i
m
}.
Let Pbe the permutation matrix corresponding to M, and let = min{s
1i
1
, s
2i
2
, . . . ,
s
mi
m
}. If = 1, then M = P. Otherwise, (S P)/(1 ) is a doubly-stochastic
matrix with fewer non-zero entries than S, and the theorem now follows by
induction.
Theorem2.4 refers to bipartite graphs that are connected and matching-covered,
and we now give a matrix interpretation of these requirements. Let A be an r s
matrix and let G be the associated bipartite graph obtained by ignoring the weights.
Then G is disconnected if and only if there exist permutation matrices P and Q
such that PAQ has the form
_
A
1
0
0 A
2
_
,
for some matrices A
1
and A
2
. An r r matrix is fully indecomposable if the sum
of the dimensions of each of its non-empty zero submatrices is at most r 1.
Assume that G is connected. Applying Theorem 5.1 to the subgraph obtained
by deleting a vertex in each part of the bipartition of G, we deduce that a connected
square bipartite graph is matching-covered if and only if its biadjacency matrix is
fully indecomposable.
Let G be a bipartite graphwitha perfect matching M, andlet Abe its biadjacency
matrix. Using the Frobenius normal form for matrices and digraphs, we can derive
a normal form for G and A. Since G has a perfect matching M, there exists a
permutation matrix S with S A (entry-wise). Each diagonal entry of AS
T
is 1.
Let P be a permutation matrix such that
P(AS
T
)P
T
=
_
_
_
_
_
A
1
0 0
A
21
A
2
0
.
.
.
.
.
.
.
.
.
.
.
.
A
k1
A
k2
A
k
_
_
_
_
_
(7)
is the Frobenius normal form of AS
T
, where A
1
, A
2
, . . . , A
k
are irreducible.
Since the diagonal entries of each A
i
are 1, Theorem 5.1 implies that each A
i
is fully indecomposable. Thus there exist permutation matrices P and Q such that
PAQ has the form given in (7). The matrices A
1
, A
2
, . . . , A
k
are uniquely deter-
mined by A up to (not necessarily simultaneous) permutations of their rows and
columns, and are called the fully indecomposable components of A. The matrix
A is fully indecomposable if and only if it has only one fully indecomposable
component.
6. Permanents
Let G be a square bipartite graph with adjacency matrix A = (a
i j
) of order n. As
already observed, Theorem 5.1 implies that G has a perfect matching if and only
if A does not have a zero matrix, the sum of whose dimensions is n +1. Since
the perfect matchings of G are in one-to-one correspondence with the permuta-
tion matrices P for which P A (entry-wise), the number of perfect matchings
of G equals the number of permutations i
1
, i
2
, . . . , i
n
of {1, 2, . . . , n} for which
a
1i
1
a
2i
2
. . . a
ni
n
= 1.
More generally, if A = (a
i j
) is a real matrix of order n, the permanent of A is
dened by
per A =
i
1
,i
2
,...,i
n
a
1i
1
a
2i
2
. . . a
ni
n
,
where the summation extends over all permutations i
1
, i
2
, . . . , i
n
of {1, 2, . . . , n}.
It can be shown that per A = per PAQ, for all permutation matrices P and Q.
If G is a bipartite graph, then the permanent of A equals the number of perfect
matchings of G. If G is a weighted bipartite graph, we dene the weight of a
perfect matching M of G to be the product of the weights of the edges of M. Then
per A is the sum of the weights of the perfect matchings of G.
The permanent resembles the determinant, but unlike the determinant, which
can be computed in polynomial time using Gaussian elimination, there is no known
way to compute the permanent efciently. In fact, Valiant (see [45] and [46]) has
shown that computing the permanent of a (0, 1)-matrix that is, computing the
number of perfect matchings of a bipartite graph is a #P-complete problem. Thus
computing the permanent is computationally equivalent to many other difcult
counting problems.
The permanent also has meaning for digraphs. A cycle cover of a digraph
is a collection of directed vertex-disjoint cycles containing all the vertices. Let
A = (a
i j
) be a (0, 1)-matrix, and let G be the corresponding digraph. The fact that
each permutation of {1, 2, . . . , n} can be written uniquely as a product of disjoint
permutation cycles implies that the permanent of A counts the number of cycle
covers of G.
We now turn to bounds on the permanent. Let A = (a
i j
) be a non-negative
matrix of order n, and let r
i
=
n
j =1
a
i j
and s
j
=
n
i =1
a
i j
be the row and column
sums of A. Then an elementary bound is
per A min
_
n
i =1
r
i
,
n
i =1
s
i
_
. (8)
Now assume that permutations have been applied to A, so that
r
1
r
2
r
n
and s
1
s
2
s
n
; (9)
then a theorem of Ostrand [33] gives the lower bound
per A max
_
n
i =1
max{1, r
i
i +1},
n
i =1
max{1, s
i
i +1}
_
,
provided that per A = 0. If t = min{r
1
, r
2
, . . . , r
n
, s
1
, s
2
, . . . , s
n
}, we deduce that
per A t ! or per A = 0.
Again, assuming that (9) holds, a theorem of Jurkat and Ryser [23] gives
per A
n
i =1
min{r
i
, s
i
},
thereby improving the bound (8). Following a conjecture of Minc, Br egman [2]
proved, and later Schrijver [38] proved more elegantly, that
per A min
_
n
i =1
(r
i
!)
1/r
i
,
n
i =1
(s
i
!)
1/s
i
_
, (10)
Now assume that G is k-regular, so that A has exactly k 1s in each row and
column. Then (10) implies that
per A (k!)
n/k
.
Suppose that k is a divisor of n, and that J
k
is the all-1 matrix of order k. Then the
matrix
A =
_
_
_
_
_
J
k
0 0
0 J
k
0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 J
k
_
_
_
_
_
has permanent equal to (k!)
n/k
. This implies the truth of a conjecture of Ryser.
Theorem 6.1 Let n be a positive integer, and, let k be a divisor of n. Then the
maximum number of 1-factors of a k-regular n n bipartite graph is (k!)
n/k
.
Let (n, k) be the maximum number of perfect matchings in a k-regular n n
bipartite graph. Then the value (n, k) is given by Theorem 6.1, whenever k
divides n. Also (n, n 1) equals the nth derangement number. In addition,
(n, 2) = 2
n/2
, with equality if and only if G has n/2 connected components
isomorphic to K
2,2
if n is even, and n/2 1 components isomorphic to K
2,2
and one isomorphic to a 6-cycle if n is odd. If n 8, the (n 2)-regular n n
bipartite graphs achieving (n, n 2) are the bipartite complements of the graphs
achieving (n, 2). If k = 3 and n is not divisible by 3, then we have
(3m +1, 3) = 6
m1
9 (m 1), (5, 3) = 13,
and (3m +2, 3) = 6
m2
9
2
(m 2).
We also have (4m +1, 4) = 24
m1
44, for m 1. For references and further
discussion of the above results, see [7]. Finally, McKay and Wanless [31] have
shown that if n = kd, with d 5, then those bipartite graphs achieving (n, n k)
are the bipartite complements of those achieving (n, k).
In a series of papers culminating with [37], Schrijver obtained an exponential
lower bound for the number of perfect matchings of k-regular bipartite multi-
graphs that is, for the permanent of a non-negative integral matrix of order n
with all row and column sums equal to k. Let
(n, k) be the minimum number of

perfect matchings of a k-regular n n bipartite graph, and let
k
= liminf
n
(n, k)
1/n
.
The number
k
gives the best exponential lower bound
n
k
for the number of perfect
matchings of k-regular bipartite graphs. Schrijver proved that
k
= (k 1)
k1
/
k
k2
, from which the following result can be deduced.
Theorem 6.2 Every k regular n n bipartite multigraph has at least
_
(k 1)
k1
k
k2
_
n
perfect matchings.
By Theorem 6.2,
per A
_
(k 1)
k1
k
k2
_
n
,
for each non-negative integral matrix A of order n with all row and column sums
equal to k. For such a matrix A, the matrix (1/k)A is doubly stochastic. Van
der Waerden conjectured, and Egory chev [13] and Falikman [14] independently
proved, a sharp lower bound for the minimum permanent of doubly stochastic
matrices.
Theorem 6.3 If A is a doubly stochastic matrix of order n, then
per A
n!
n
n
,
with equality if and only if A is (1/n)J
n
.
In general, the lower bound of Theorem 6.3 for the number of perfect matchings
in a regular bipartite graph is weaker than that of Theorem 6.2.
Since the denition of per A is similar to that of det A, and since there are
good algorithms for computing the determinant, it is natural to consider whether
there is a simple transformation that converts A to another matrix A
for which
per A = det A
. This possibility is considered in the next section.

7. Converting the permanent into the determinant
A signed digraph is a weighted digraph G
s
with weights 1. Thus, its adjacency
matrix is a (0, 1, 1)-matrix A
s
= (a
i j
). Like the permanent of a (0, 1)-matrix, the
determinant of A
s
can be expressed in terms of cycle covers.
The weight wt( ) of a cycle of G
s
is the product of the weights of its arcs, and
the weight of a cycle cover of G
s
is the product of the weights of its cycles. Each
cycle cover of G
s
determines a permutation of {1, 2, . . . , n} whose permutation
cycles correspond to digraph cycles, but not every permutation of {1, 2, . . . , n}
corresponds to a cycle cover of G. The sign of a permutation is (1)
nk
, where k
is the number of cycles, and we denote this by (1)
. The signed weight s() of a

cycle cover is the product of the weight of andits sign. Usingthe correspondence
between cycle covers of G
s
and non-zero terms in the determinant expansion of
A
s
, we have the formula
det A
s
=
S
n
(1)
i =1
a
i
i
=
C(G
s
)
s(),
where C(G
s
) is the set of cycle covers of G
s
.
Now let G be the digraph obtained from G
s
by ignoring the signs of its arcs,
and let A be the adjacency matrix of G. Since the cycle covers of G
s
are the same
as those of G, the triangle inequality implies that
per A =
C(G)
1 |
C(G
s
)
s()| = |det A
s
|, (11)
with equality if and only if the cycle covers of G
s
all have the same signed weight.
This suggests a possible way of computing the permanent of a (0, 1)-matrix A:
replace certain 1s of A by 1s so as to obtain a matrix A
s
with the property that
all cycle covers of G
s
have the same signed weight, and then calculate |det A
s
|.
An assignment of this type may not be possible. For example, consider J
3
,
the all-1 matrix of order 3. Then per J
3
= 6, and by Hadamards inequality for
determinants, the absolute value of the determinant of each 3 3 matrix of 1s
is at most 3
3/2
, which is less than 6. It is thus impossible to use this technique to
convert the computation of per J
3
into the computation of the determinant of a J
s
3
.
In the case that A is a (0, 1)-matrix, per A = 0 can be efciently checked using
K onigs theorem. If per A = 0 and such an assignment is possible, we say that A
is convertible. More precisely, a square (0, 1)-matrix A is convertible if per A =0
and it is possible to assign minus signs to some non-zero elements of A to obtain
a matrix A
s
such that per A = |det A
s
|; such a matrix A
s
is a conversion of
A. Note that the conversion A
s
applies not only to the permanent of A, but to the
permanent of any matrix obtained fromAby replacing its 1s by arbitrary numbers.
More generally, we call an assignment of minus signs to some of the elements of
A a signing of A.
The matrix J
3
has no conversion, but the matrix
A
s
=
_
1 1
1 1
_
is a conversion of J
2
. We say that a bipartite graph G is convertible if its biadjacency
matrix is convertible.
To test whether a signing A
s
is a conversion of A, we need not examine all cycle
covers of A. The following basic theoremof Bassett, Maybee and Quirk [1] asserts
that if I
n
A(entry-wise) (which implies that there is a loop at each vertex of G
s
),
then one need check only the weights of the cycles of G
s
. We note that if per A = 0,
then there exists a permutation matrix P such that I
n
PA. Clearly, A
s
is a conver-
sion of A if and only if PA
s
is a conversion of PA. Also, if D is a diagonal matrix
each of whose diagonal entries is 1 or 1, then A
s
D is a conversion of AD if and
only if A
s
is a conversion of A. Thus there is no loss of generality in assuming that
I
n
A, and that each element on the main diagonal of A
s
is 1.
Theorem 7.1 Let A be a (0, 1)-matrix with I
n
A, and let A
s
be a signing of A
with every diagonal entry 1. Then A
s
is a conversion of A if and only if each
cycle in the signed digraph of A
s
has weight 1.
Proof Let G be the digraph of A, and let G
s
be the signed digraph of A
s
.
First suppose that A
s
is a conversion of A. Then equality holds in (11), and so
each cycle cover of A
s
has the same signed weight. The cycle cover consisting of
n loops has signed weight (1)
n
. If is an l-cycle of G
s
, then the cycle cover of
G
s
consisting of and n l loops has weight (1)
n1
wt( ), and so wt( ) = 1.
Conversely, suppose that each cycle has weight 1. Let be a cycle cover
consisting of l cycles. Since each cycle has weight 1, the signed weight of is
(1)
n1
(1)
l
= (1)
n
. All cycle covers have the same signed weight (1)
n
, and
by (11), A
s
is a conversion of A.
Given a signed digraph G
s
, we consider the unsigned digraph G
obtained from
G
s
by ignoring the weights and replacing each arc i j with sign +1 by a path
i v
i j
j of length 2, by inserting a new vertex v
i j
. It is easy to verify that G
s
has no cycles of positive weight if and only if G
has no cycles of even length.

Thus, the problem of recognizing whether A
s
is a conversion of A is equivalent
to that of recognizing whether a digraph G
has a cycle of even length. This was a

major unsolved problem until recently. We describe the solution to this even-cycle
problem below.
Theorem 7.1 yields another proof that J
3
is not convertible. If G
s
is a signed
digraph corresponding to a signing A
s
of J
3
with diagonal entries 1, then each
arc of G that is not a loop is contained in exactly two cycles and G
s
has ve
non-trivial cycles, and so not all cycles of G
s
can have weight 1.
As another application of Theorem 7.1, we prove a result of Kasteleyn that
every connected matching-covered planar bipartite graph is convertible. If

G is an
orientation of a plane bipartite graph G, and if is a cycle of G, then a forward
edge of is an edge whose orientation agrees with the clockwise orientation of .
The number of forward edges of is denoted by n().
Corollary 7.2 Let G be a 2-edge-connected matching-covered plane bipartite
graph. If there exists an orientation

G of G for which each cycle G that bounds a
face has an odd number of forward edges, then G is convertible.
Proof Suppose that

G is anorientationof G for whicheachcycle that bounds a face
has an odd number of forward edges. Let M = {x
1
y
1
, x
2
y
2
, . . . , x
n
y
n
} be a perfect
matching of G. By reversing the orientation of each edge with one vertex in a given
subset of vertices, we may assume that each edge of x
i
y
i
of M is oriented in

G from
x
i
to y
i
. Let A
s
= (a
i j
) be the signing of the biadjacency matrix of G with a
i j
= 0
if x
i
is not joined to y
j
in G, a
i j
= 1 if x
i
y
j
in

G, and a
i j
= 1 if y
j
x
i
in

G. Let G
s
be the signed digraph of A
s
, and let x
1
, x
2
, , x
k
, x
1
be a directed
cycle of G
s
. By Theorem 7.1, it sufces to show that the weight of is 1.
Note that
x
1
y
1
x
2
y
2
. . . x
k
y
k
x
1
(12)
is a cycle in the bipartite graph G. Let v, e and f denote the number of vertices,
edges and faces, respectively, in the region R bounded by , and let
1
,
2
, . . . ,
f
be the cycles in G that bound the faces in R. Each edge in R, but not in , is a
forward edge in one face of R and a backward edge in another, and each edge in
belongs to exactly one face bounded by R. This, and the assumption that each
n(
i
) is odd, imply that
n() +e 2k =
f
i =1
n(
i
) f (mod 2).
Hence, n() e + f (mod 2). By Eulers theoremfor planar graphs, v e + f =
1. Since G has a perfect matching containing half of the edges of , v is even. We
conclude that n() is odd.
In (12), the vertices of are either in clockwise or anticlockwise order. In the
former case, each edge of belonging to M is a forward edge and there are an odd
number of backward edges lying in but not in M. Since these edges correspond to
edges of of weight 1, wt( ) = 1. In the latter case, each edge of belonging
to M is a backward edge, and (13) implies that there are an odd number of forward
edges lying in but not in M. Since these edges correspond to edges of of
weight 1, wt( ) = 1.
The existence of such an orientation

G is easily established by induction on
the number of edges. More generally, the biadjacency matrix of a planar bipartite
graph with a perfect matching is convertible (see [24]). Kasteleyn used this result to
determine the number n
pq
of ways of tiling a p q chessboard (with p or q even)
with dominoes. It is easy to see that n
pq
is the number of perfect matchings of the
planar bipartite graph G
pq
whose vertices correspond to the squares of the p q
chessboard and whose edges correspond to adjacent squares. As outlined in [8],
one can use Corollary 7.2 to nd a conversion of the biadjacency matrix of G
pq
,
and then to calculate the eigenvalues of the conversion to attain the closed formula:
n
pq
= 2
pq/2
p
k=1
q
l=1
_
cos
2
_
k
p +1
_
+cos
2
_
l
q +1
__
1/4
e
0.29pq
.
This technique, and more sophisticated algebraic techniques, have been used to
give closed formulas for tilings of other objects (see [28]).
The complete bipartite graph K
3,3
is non-planar, andsince its biadjacencymatrix
is J
3
it is also not convertible. More generally, Little [29] obtained the following
characterization of convertible matrices. An even subdivision of a bipartite graph
G is a graph H obtained by replacing the edges of G by internally disjoint paths,
each with an even number of vertices.
Theorem 7.3 Let G be a bipartite graph with a perfect matching. Then G is
convertible if and only if G does not contain a spanning subgraph H in which one
component is an even subdivision of K
3,3
and other components (if any) are single
edges.
The Heawood graph, which has biadjacency matrix I
7
+C
7
+C
3
7
, where C
7
is the permutation matrix with 1s in positions (1, 2), (2, 3), . . . , (6, 7), (7, 1), is
convertible but not planar. In fact,
per
_
I
7
+C
7
+C
3
7
_
= det
_
I
7
+C
7
+C
3
7
_
= 24.
In 1999, Robertson, Seymour and Thomas [36] gave a different characterization
of convertible graphs, and used it to develop a polynomial-time algorithm for
recognizing convertible matrices, thereby solving the even-cycle problem. Their
characterization asserts that every convertible bipartite graph can be constructed by
piecing together planar bipartite graphs and Heawood graphs in a particular way.
Let G
1
and G
2
be connected matching-covered bipartite graphs, and let x
1
y
1
and x
2
y
2
be edges in G
1
and G
2
, respectively. The 1-join of G
1
and G
2
is the
bipartite graph obtained from G
1
and G
2
by identifying x
1
with x
2
and y
1
with
y
2
, and adjoining all edges of the form xy, where x y
1
is an edge of G
1
and x
2
y
is an edge of G
2
. A brace is a connected bipartite graph G for which each pair of
disjoint edges of G is contained in a perfect matching of G. Now let G
1
and G
2
be
braces, and let w
1
x
1
y
1
z
1
w
1
and w
2
x
2
y
2
z
2
w
2
be cycles in G
1
and G
2
, respectively.
The 2-join of G
1
and G
2
is the bipartite graph obtained by identifying w
1
with
w
2
, x
1
with x
2
, y
1
with y
2
, and z
1
with z
2
. It is not difcult to show that every
1-join or 2-join of convertible bipartite graphs is convertible (see [8]). Also, if G
is convertible and if G
is a bipartite graph which has a perfect matching and is

obtained from G by deleting some edges, then G
is convertible.
Since a matrix is convertible if and only if each of its fully indecomposable
components is convertible, we can restrict ourselves to characterizing convertible
matching-covered bipartite graphs. We now state the characterization of convert-
ible bipartite graphs given in [36].
Theorem 7.4 Let G be a connected matching-covered bipartite graph.
(a) If G is not a brace, then it is convertible if and only if it can be obtained from
the 1-join of convertible graphs by removing a (possibly empty) subset of
edges.
(b) If G is a brace, then it is convertible if and only if it is either a planar
bipartite graph or the Heawood graph, or is a spanning subgraph of a 2-join
of planar bipartite braces.
8. Chordal graphs and perfect Gaussian elimination
A graph G is chordal if each cycle of length greater than 3 has a chord that is,
an edge joining two non-consecutive vertices of . In a chordal graph, no induced
subgraph is a cycle of length greater than 3. Complete graphs, in particular cycles
of length 3, are chordal, as are all trees. Chordal graphs are useful data structures
in solving sparse symmetric systems of linear equations.
A simplicial vertex of a graph G is a vertex v whose neighbours are all joined
to each other. Thus, v is a simplicial vertex of G if and only if the neighbours of v
induce a complete graph. An ordering v
1
, v
2
, . . . , v
n
of the vertices of G is a sim-
plicial (or perfect) elimination ordering if v
i
is a simplicial vertex of the subgraph
induced by {v
i
, v
i +1
, . . . , v
n
}, for i = 1, 2, . . . , n 1. Any ordering of the ver-
tices of a complete graph is a simplicial elimination ordering. Any ordering of the
vertices of a tree obtained by successively deleting end-vertices is a simplicial elim-
ination ordering. Acycle of length 4 or more has no simplicial elimination ordering.
Dirac [11] derived two characterizations of chordal graphs. The rst is the
existence of a simplicial elimination ordering. The other is in terms of vertex
separators of a pair of non-adjacent vertices v and w that is, sets of vertices
S, not containing v or w, for which the subgraph induced on the complement
of S is disconnected, with v and w in different components. If no proper subset
of S separates v and w, then S is a minimal vertex separator of v and w. For
non-adjacent vertices v and w, the complement of {v, w} is a vertex separator, and
hence there is a minimal vertex separator of v and w.
Theorem 8.1 Let G be a graph. The following statements are equivalent.
(1) For each minimal vertex separator S of G, the induced subgraph G[S] is a
complete graph.
(2) G is a chordal graph.
(3) G has a simplicial elimination ordering.
Proof (1) (2) Assume that (1) holds, and consider a cycle = vxwa
1
. . . a
k
v
of length greater than 3, so that k 1. If v and w are adjacent, then has a chord.
Otherwise, let S be a minimal vertex separator of v and w. Then S contains x and
at least one of a
1
, a
2
, . . . , a
k
, and so has a chord. Thus G is chordal.
(2) (1) Assume that G is chordal. Let S be a minimal vertex separator of vertices
v and w, and let G
v
and G
w
be the connected components of G S containing
v and w, respectively. The minimality assumption implies that each vertex in S is
adjacent to some vertex in G
v
and some vertex in G
w
. If S has only one vertex,
then G[S] is a complete graph. Otherwise, let x and y be distinct vertices in S.
Then there exist an x-y path all of whose internal vertices are in G
v
, and an y-x
path all of whose internal vertices are in G
w
. We may choose such paths to have
smallest length. Hence, there is a cycle of length greater than 3, all of whose
vertices (except for x and y) belong to G
v
or G
w
. Since G is chordal, the cycle
has a chord. But our assumptions imply that the only possible chord is an
edge joining x and y. Thus G[S] is a complete graph, and (1) holds.
(3) (2) Let be a cycle of length greater than 3. The vertex v of with the
smallest index in a simplicial elimination scheme has the property that the two
vertices adjacent to v on the cycle are adjacent to each other, and hence determine
a chord of . It follows that G is chordal.
(2) (3) To prove that a chordal graph has a simplicial elimination ordering, it
sufces to show that every chordal graph has a simplicial vertex. Any vertex of
a complete graph is a simplicial vertex. We show by induction that each chordal
graph G which is not complete has at least two non-adjacent simplicial vertices.
Let v and w be non-adjacent vertices of G, and let S be a vertex separator of v and
w. Let G
v
= G[U] and G
w
= G[W]. By what we have proved, G[S] is a complete
graph. If the chordal graph G[S U] is not complete, then it has two non-adjacent
simplicial vertices, by the induction assumption, at least one of which must be in
U, since G[S] is complete. If G[S U] is complete, then any vertex in U is a
simplicial vertex of G. Thus U contains a simplicial vertex of G, and similarly so
does W, and these two vertices are non-adjacent.
Let A = (a
i j
) be a symmetric invertible matrix of order n, all of whose diagonal
entries are non-zero. Gaussian elimination to solve a symmetric system of linear
equations Ax = b proceeds by successively pivoting on a non-zero diagonal ele-
ment (using elementary row operations and the corresponding column operations,
to zero out the non-zero elements in the row and column of the pivot element)
until Ais reduced to a diagonal matrix. If Ais a large sparse matrix, and if the graph
G whose edges correspond to the non-zero off-diagonal elements of A is chordal,
then by following a simplicial elimination ordering of G when choosing pivots
for Gaussian elimination on A, a zero element of A remains zero throughout the
elimination. This is usually described by saying that no ll-in occurs, implying,
in particular, that sparseness is preserved. We are also assuming that the elements
on the main diagonal remain non-zero throughout Gaussian elimination. This im-
plies that a data structure for A can be constructed by using only the positions of
the non-zero elements of A, reducing storage requirements substantially.
Bipartite graphs provide a model for Gaussian elimination for non-symmetric
matrices also. We briey discuss the bipartite analogue of chordal graphs and
simplicial elimination orderings. A bipartite graph with a cycle of length greater
than 3 can never be chordal. A graph G is chordal-bipartite if it is bipartite and
if each cycle of length greater than 4 has a chord. Complete bipartite graphs in
particular, cycles of length 4 are chordal-bipartite. Trees are both chordal and
chordal-bipartite.
Let G be a bipartite graph with bipartition {X, Y}. An edge e joining v in X and
w in Y is bisimplicial if the set Y
v
of neighbours of v, together with the set X
w
of
neighbours of w, induces a complete bipartite graph. Let M = (e
1
, e
2
, . . . , e
k
) be
a sequence of edges of G that form a matching, let e
i
= p
i
q
i
(i = 1, 2, . . . , k), and
let P
i
= { p
1
, p
2
, . . . , p
i 1
} and Q
i
= {q
1
, q
2
, . . . , q
i 1
}, for i = 1, 2, . . . , k +1.
Then M is a bisimplicial elimination orderingfor G if e
i
is a bisimplicial edge of the
subgraph G (P
i
Q
i
) (i = 1, 2, . . . , k) and the induced subgraph G (P
k+1

Q
k+1
) has no edges. In performing Gaussian elimination on a non-symmetric linear
system Ax = b, a bisimplicial elimination ordering provides a sequence of pivots
on the elements of A corresponding to the edges e
1
, e
2
, . . . , e
k
, for which no ll-in
results. Golumbic and Goss [19] proved the following theorem.
Theorem 8.2 Every chordal bipartite graph has a bisimplicial elimination
ordering.
The converse of Theorem8.2 is false, as can be seen by the graph consisting of a 6-
cycle and one end-edge incident to each vertex of the cycle. For more information
on chordal bipartite graphs, see [18] and [19].
9. Ranking players in tournaments
Recall that, as its name suggests, a tournament models the results of a round-robin
tournament on a set of n players, in which each player plays each of the others
exactly once and there are no ties. The vertices correspond to the players, and there
is an arc from v to w if v beats w.
Given the results of a round-robin tournament G, it is natural to try to rank the
players. If the vertices of G can be ordered as (v
i
1
, . . . , v
i
n
), where v
i
j
v
i
k
if
and only if j k, then G is a transitive tournament and there is no ambiguity
in the ranking: player v
i
1
is the strongest, v
i
2
is the second strongest, and so on.
More generally, suppose that G is not strongly connected, so that its adjacency
matrix A is reducible. Then G contains a directed cut (U, W), and since G is a
tournament, v
i
v
j
for all v
i
U and v
j
W. Any ranking of the players must
rank those players in U as stronger than those in W. Hence, in ranking the results
of a round-robin tournament, we can restrict our attention to ranking the results of
the subtournaments corresponding to the irreducible components of A.
Suppose now that G is strongly connected. It is well known that there then
exists a Hamiltonian cycle in G. The existence of such a cycle causes any ranking
to have inconsistencies, since any player v can argue that he is at least as good as
any other player w by considering the path from v to w. Yet it is still desirable to
have some ranking of the players.
Arst attempt is to rank the players according to the number of games they win.
The score of a vertex v
i
is its out-degree, and the score vector s of G is the vector
whose i th entry is the score of v
i
. Thus Aj = s, where j is the n 1 all-1 vec-
tor. This ranking scheme has two drawbacks. First, by the pigeon-hole principle,
each strongly connected tournament has at least two players with the same score,
and so there are always ties in the ranking. Second, this ranking scheme assumes
that all wins are valued equally, but a win against a strong player should ar-
guably be valued more than a win against a weak player. Several power ranking
methods have been proposed that take into account the strength of each players
opponents.
We discuss two such ranking methods. The rst is known as the Kendall-Wei
method, and it can be motivated as follows. The vector s = Aj records the scores
of the players, and so its i th entry is a measure of the strength of player v
i
. The i th
entry
i :i j
S
j
of As is the sum of the strengths (according to s) of the players
that v
i
beats. Thus, As accounts for the strengths of the opponents that player v
i
beats. Now the entries of As can be viewed as a measure of the strength of the
players, and we can consider A
2
s = A(As), the vector whose i th entry is the sum
of the strengths (according to As) of the players that i beats. This can be repeated
to obtain a sequence r
1
= s, r
2
= As, r
3
= Ar
2
, . . . of strength vectors. We nor-
malize this sequence so that the sum of the entries of each vector is 1 to obtain a
sequence
r
1
= s/j
T
s, r
2
= r
2
/j
T
r
2
, . . . .
It is not difcult to show that the limits = lim
n
j
T
r
n
and x = lim
n
r
n
exist and satisfy Ax = x. We conclude from Theorem 2.2 that is the spectral
radius of A and that x is the unique non-negative eigenvector of A whose entries
sum to 1. We are led to the Kendall-Wei method: rank the players of G according
to the entries of x, with stronger players corresponding to larger entries of x.
The variance of the vector x is dened by
var (x) =
i <j
(x
i
x
j
)
2
,
and provides a measure of how evenly matched the players are, according to the
Kendall-Wei method; var(x) is larger when there is a greater disparity among the
strengths of the players. In particular, var(x) = 0 if and only if all the players of G
have equal strength. The following result relates var(x) and (see [30]).
Theorem 9.1 Let G be a strongly connected tournament with adjacency matrix
A. Let be the spectral radius of A, and let x be the positive eigenvector of A
corresponding to with j
T
x = 1. Then:
(1) var(x) = (n 1 2)x
T
x;
(2) (n 1)/2, with equality if and only if n is odd and G is regular.
Proof We have
i <j
(x
i
x
j
)
2
= (n 1)
n
i =1
x
2
i
2
i <j
x
i
x
j
= (n 1)x
T
x x
T
(J
n
I
n
)x
= (n 1)x
T
x x
T
(A +A
T
)x
= (n 1)x
T
x 2x
T
x
= (n 1 2)x
T
x.
Hence (1) holds, and (n 1)/2 with equality if and only if var (x) =0. Since
var (x) =0 if and only if x = (1/n)j, and since j is an eigenvector of A if and only
if A has constant row sums, (2) holds.
In the light of Theorem 9.1, it is natural to ask which strongly connected tour-
naments on n vertices have the smallest and largest spectral radii in other words,
which strongly connected tournaments are the least or most evenly matched, ac-
cording to the Kendall-Wei scheme? Theorem 9.1 implies that, for n odd, the
largest spectral radius is (n 1)/2, and the tournaments achieving this are the
regular tournaments. Brualdi and Li [5] conjectured that, for n = 2k, the largest
spectral radius occurs when the adjacency matrix has the form
_
B B
T
B
T
+I B
_
,
where B is the adjacency matrix of a transitive tournament with k vertices. In
[27] Kirkland shows that, for sufciently large k, a tournament of order 2k with
largest spectral radius must have half of its players with score k, and the other
half with score k 1. Brualdi and Li [5] conjectured that the minimum spec-
tral radius is achieved by the tournament with v
i
v
j
for i and j with i > j +1,
and v
i
v
i +1
for i = 1, 2, . . . , n 1. This conjecture was also proved by
Kirkland [26].
We can also use the positive eigenvector y of A
T
, with j
T
y = 1, to rank the
players of a tournament. This corresponds to taking into account the weaknesses
of those players defeating a given player; smaller entries of y correspond to stronger
players. A ranking scheme of Ramanujacharyula [35] ranks the players according
to the strength-to-weakness ratio x
i
/y
i
, with larger values corresponding to better
players. More about this ranking scheme can be found in [25].
The method Pagerank
TM
, used by the search engine Google to rank web-pages,
is similar to the Kendall-Wei ranking scheme. Unlike many search engines that
work by searching for web-pages having many words in common with a given
query, Google focuses on links between pages and searches for high-quality
pages that are related to the query. The quality of a page is determined by how
many relevant pages refer to it, the quality of those referring pages, how many
relevant pages the page refers to, and the quality of those pages. More precisely,
Google rst uses text-matching techniques to generate a set V of web-pages that
are relevant to the query. Then Google forms a weighted digraph G whose vertices
correspond to the elements of V, and where the weight of the edge vw is the
number of times that web-page v links to web-page w.
Let A be the adjacency matrix of G. Then the i th entry of x
(1)
= A
T
j measures
how often a relevant site refers to site i . Large entries of x
(1)
are potential experts
for the query. Similarly, the i th entry of y
(1)
= Aj measures the number of relevant
sites referred to by i , and large entries of y
(1)
correspond to potential information
sources for the query. But not all links on the web are of equal value. Pagerank
TM
takes into account the quality of the sites linked to and from a site. This is done by
considering the sequences and x
(i )
= A
T
y
(i 1)
and y
(i )
= Ax
(i 1)
for i 2. Thus,
for example, the i th entry of x
(2)
= A
T
y
(1)
is the sum of the weights (according
to y
(1)
) of the sites that refer to page i . It is easy to see that x
(2i )
= (AA
T
)
i
j and
y
(2i )
= (A
T
A)
i
j (i = 1, 2, . . .). Thus, if A
T
A is irreducible, then the normalized
sequences
x
(2i )
/j
T
x
(2i )
and y
(2i )
/j
T
y
(2i )
converge to positive eigenvectors u and v of A
T
A and AA
T
, respectively.
Pagerank
TM
uses the i th entry of u +v as a measure of the relevant expertise
of site i , and then lists the sites in non-increasing order based on the values of the
entries of u. Hence, Google rests mathematically on the Perron-Frobenius theorem,
with which we began our discussion.
References
1. L. Bassett, J. Maybee and J. Quirk, Qualitative economics and the scope of the corre-
spondence principle, Econometrica 36 (1968), 544563.
2. L. M. Br egman, Certain properties of non-negative matrices and their permanents, Dokl.
Akad. Nauk SSSR 211 (1973), 2730 (= Soviet Math. Dokl. 14 (1973), 945949).
3. R. A. Brualdi, Matrices, eigenvalues, and directed graphs, Linear and Multilinear
Algebra 11 (1982), 143165.
4. R. A. Brualdi, The symbiotic relationship of combinatorics and matrix theory, Linear
Alg. Appl. 162164 (1992), 65105.
5. R. A. Brualdi and Q. Li, Problem 31, Discrete Math. 43 (1983), 329330.
6. R. A. Brualdi, S. V. Parter and H. Schneider, The diagonal equivalence of a non-negative
matrix to a stochastic matrix, J. Math. Anal. Appl. 16 (1966), 3150.
7. R. A. Brualdi and H. J. Ryser, Combinatorial Matrix Theory, Encyclopedia of Mathe-
matics and its Applications 39, Cambridge Univ. Press, 1991.
8. R. A. Brualdi and B. L. Shader, Matrices of Sign Solvable Linear Systems, Cambridge
Tracts in Mathematics 116, Cambridge Univ. Press, 1995.
9. F. R. K. Chung, R. L. Grahamand P. M. Winkler, On the addressing problemfor directed
graphs, Graphs Combin. 1 (1985), 4150.
10. D. de Caen, The ranks of tournament matrices, Amer. Math. Monthly 98 (1991), 829
831.
11. G. A. Dirac, On rigid circuit graphs, Abh. Math. Sem. Univ. Hamburg 25 (1961), 7176.
12. A. L. Dulmage and N. S. Mendelsohn. Gaps in the exponent set of primitive matrices,
Illinois J. Math. 8 (1964), 642656.
13. G. P. Egory cev, Asolution of van der Waerdens permanent problem, Dokl. Akad. Nauk
SSSR 258 (1981), 10411044 (= Soviet Math. Dokl. 23 (1981), 619622).
14. D. I. Falikman, A proof of van der Waerdens conjecture on the permanent of a doubly
stochastic matrix, Mat. Zametki 29 (1981), 931938 (= Math. Notes 29 (1981), 475
479).
15. E. Fornasini and M. Valcher, Primitivity of positive matrix pairs: algebraic characteri-
zation, graph theoretic description and 2Dsystems interpretation, SIAMJ. Matrix Anal.
Appl. 19 (1998), 7188.
16. G. Frobenius,

Uber Matrizen aus nicht negativen Elementen, Sitzungsber. Preuss. Akad.
Wiss. Berlin (1912), 456477.
17. G. Frobenius,

Uber zerlegbare Determinanten, Sitzungsber. Preuss. Akad. Wiss. Berlin
(1917), 274277.
18. M. C. Golumbic, Algorithmic GraphTheory andPerfect Graphs, Academic Press, 1980.
19. M. C. Golumbic and C. F. Goss, Perfect elimination and chordal bipartite graphs, J.
Graph Theory 2 (1978), 155163.
20. R. L. Graham and H. O. Pollak, On the addressing problem for loop switching, Bell
System Tech. J. 50 (1971), 24952519.
21. D. Gregory, B. L. Shader and V. Watts, Biclique decompositions and Hermitian rank,
Linear Alg. Appl. 292 (1999), 267280.
22. R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge Univ. Press, 1985.
23. W. B. Jurkat and H. J. Ryser, Term ranks and permanents of non-negative matrices,
J. Algebra 5 (1967), 342357.
24. P. W. Kasteleyn, Dimer statistics and phase transitions, J. Math. Phys. 4 (1963),
287293.
25. S. Kirkland, Spectral radii of tournament matrices whose graphs are related by an arc
reversal, Linear Alg. Appl. 217 (1995), 179202.
26. S. Kirkland, On the minimum Perron value of an irreducible tournament matrix,
Linear Alg. Appl. 244 (1996), 277304.
27. S. Kirkland, A note on Perron vectors for almost regular tournament matrices, Linear
Alg. Appl. 266 (1997), 4347.
28. G. Kuperberg, Symmetries of plane partitions and the permanent-determinant method,
J. Combin. Theory (A) 68 (1994), 115151.
29. C. H. C. Little, A characterization of convertible (0, 1)-matrices, J. Combin. Theory
(B) 18 (1975), 187208.
30. J. S. Maybee and N. J. Pullman, Tournament matrices and their generalizations I,
Linear and Multilinear Algebra 28 (1990), 5770.
31. B. D. McKay and I. M. Wanless. Maximising the permanent of (0, 1)-matrices and the
number of extensions of latin rectangles, Elec. J. Combin. 5 (1998), R11.
32. M. Newman, Integral Matrices, Academic Press, 1972.
33. P. A. Ostrand, Systems of distinct representatives II. J. Math. Anal. Appl. 32 (1970), 14.
34. G. W. Peck, A new proof of a theorem of Graham and Pollak, Discrete Math. 49
(1984), 327328.
35. C. Ramanujacharyula, Analysis of preferential experiments, Psychometrika 29 (1964),
257261.
36. N. Robertson, P. D. Seymour and R. Thomas, Permanents, Pfafan orientations, and
even directed circuits, Ann. of Math. 150 (1999), 929975.
37. A. Schrijver, A short proof of Mincs conjecture, J. Combin. Theory (A) 25 (1978),
8083.
38. A. Schrijver, Counting 1-factors in regular bipartite graphs, J. Combin. Theory (B) 72
(1998), 122135.
39. J. J. Seidel, A survey of two-graphs. Teorie Combinatoire I (ed. B. Segre), Accademia
Nazionale dei Lincei (1976), 481511.
40. B. L. Shader, On tournament matrices, Linear Alg. Appl. 162164 (1992), 335368.
41. B. L. Shader and S. Suwilo, Exponents of non-negative matrix pairs, Linear Alg. Appl.,
363 (2003), 275293.
42. J. Y. Shao, The exponent set of symmetric primitive matrices, Scientia Sinica (A)
XXX, 348358.
43. R. Sinkhorn and P. Knopp, Concerning non-negative matrices and doubly stochastic
matrices, Pacic J. Math. 21 (1967), 343348.
44. H. Tverberg, On the decomposition of K
n
into complete bipartite graphs, J. Graph
Theory 6 (1982), 493494.
45. L. G. Valiant, Completeness classes in algebra, Proc. 11th ACM Symp. Theory of
Computing (1979), 249261.
46. L. G. Valiant, The complexity of computing the permanent, Theoret. Comp. Sci. 8
(1979), 189201.
47. H. Wielandt, Unzerlegbare, nicht negative Matrizen, Math. Zeit. 52 (1950), 642645.
48. T. H. Wei, The Algebraic Foundations of Ranking Theory, PhDthesis, Cambridge, 1952.
3
Spectral graph theory
DRAGO
S CVETKOVI
C and PETER ROWLINSON

1. Introduction
2. Angles
3. Star sets and star partitions
4. Star complements
5. Exceptional graphs
6. Reconstructing the characteristic polynomial
7. Non-complete extended p-sums of graphs
8. Integral graphs
References
In this chapter we are concerned with the spectrum of the adjacency matrix
of a nite graph. We present a selection of recent results related to graph
angles, star sets, star partitions, star complements, and graphs with least
eigenvalue 2. In addition, we consider a spectral reconstruction problem,
graphs that are cospectral or almost cospectral, and graphs for which all
eigenvalues are integers.
1. Introduction
The matrices commonly used to specify a nite graph include the Laplacian matrix
(see Chapter 4), the Seidel matrix (see Section 2 below), and the (0, 1)-adjacency
matrix. In this chapter we are concerned with the spectra of adjacency matrices, a
topic that features prominently in Chapters 1 and 2 and in the monographs [12],
[13] and [21]. The subject has continued to attract the attention of researchers, and
it is impossible to review in one short chapter all of the important results obtained
in the last few years. Accordingly, it was necessary to be selective, and the choice
made here reects the authors own interests.
88
Some basic facts about graph spectra can be found in Chapter 1 and [46, Ch.
11]. Here we extend the discussion of eigenvalues to related algebraic invariants,
such as the angles between eigenspaces and coordinate axes treated in Section 2.
These lead naturally to the notion of a star partition, which provides a one-to-one
correspondence (in general, not unique) between the eigenvalues and the vertices
of a graph (see Section 3). In such a correspondence the vertices associated with a
common eigenvalue are said to forma star set for , while the subgraph induced
by the remaining vertices is called a star complement for . In Section 4 we
discuss star sets and a technique for constructing the graphs with a prescribed star
complement. This technique makes it possible not only to characterize a number
of graphs, but also to complete the description of all graphs with least eigenvalue
2. The problem of describing the exceptional graphs that arise here was one of
some 25 years standing, and it is discussed in Section 5.
Another long-standing problem concerns the reconstructibility of the charac-
teristic polynomial of a graph from the characteristic polynomials of the vertex-
deleted subgraphs. This has recently been solved for trees, and a proof of the recon-
structibility in this case is given in Section 6. On the other hand, a conjecture from
1983 on almost cospectral graphs has recently been disproved. The context is a
constructionknownas a non-complete extended p-sumof graphs, discussedinSec-
tion 7. This construction also provides one means of generating graphs for which
every eigenvalue is an integer: these integral graphs are discussed in Section 8.
The use of computers is an intrinsic feature of research in discrete mathematics
and, in particular, in the theory of graph spectra. The chapter includes computa-
tional results related to minimal cospectral graphs with the same angles, and to
integral graphs with ten or fewer vertices.
2. Angles
Let G be a graph with vertex-set V(G) = {1, 2, . . . , n} and let A be the adjacency
matrix of G. Thus, A = (a
i j
), where a
i j
= 1 if i and j are adjacent and a
i j
= 0
otherwise. The eigenvalues of A are graph invariants, because a relabelling of the
vertices results in a matrix similar to A. Accordingly, we may dene
G
(x), the
characteristic polynomial of G, as det(xI A). We take the distinct eigenvalues of
Ato be
1
,
2
, . . . ,
m
, where
1
>
2
> >
m
. These eigenvalues, together
with their respective multiplicities k
1
, k
2
, . . . , k
m
, constitute the spectrum of G.
Since in general a graph is not determined by its spectrum (see Chapter 1), we
seek further algebraic invariants which might distinguish cospectral graphs. As
an orthogonally diagonalizable matrix, A is determined by n linearly independent
eigenvectors and the corresponding eigenvalues
1

2

n
. Accordingly,
90 Drago s Cvetkovi c and Peter Rowlinson
it is natural to turn our attention to eigenspaces, but we must bear in mind that the
coordinates of vectors are permuted when the vertices are relabelled. This presents
no difculty when all the coordinates coincide, and so, abusing terminology, we
can dene the main angles of G as the cosines of the angles between the all-1
vector j and the eigenspaces of A. For i = 1, 2, . . . , m, let P
i
be the n n matrix
which represents the orthogonal projection of R
n
onto the eigenspace E(
i
) with
respect to the standard basis {e
1
, e
2
, . . . , e
n
} of R
n
. Then the main angles of G are
the numbers
1
,
2
, . . . ,
m
, where
i
= P
i
j/
n. Note that (
1
,
2
, . . . ,
m
) =
(1, 0, . . . , 0) if and only if j E(
1
) that is, if and only if G is regular. In general,
we say that
i
is a main eigenvalue if j / E(
i
)
.
Given the spectrum of G, a knowledge of the invariant (
1
,
2
, . . . ,
m
) is
equivalent (see [39]) to a knowledge of either of two other spectra that appear
in the literature. One is the Seidel spectrum, which is the spectrum of the matrix
J I 2A (where J is the all-1 matrix), and the other is the derived spectrum of
G, as dened by Neumaier [37].
We also observe that, given the spectrum of G, a knowledge of the invariant
(
1
,
2
, . . . ,
m
) is equivalent toa knowledge of the spectrumof the complement G,
or of the cone over G (obtained from G by adding a vertex adjacent to every vertex
of G). For the observation concerning G, note rst that the spectral decomposition
of A is
A =
1
P
1
+
2
P
2
+ +
m
P
m
,
where P
2
i
= P
i
= P
T
i
(i = 1, 2, . . . , m) and P
i
P
j
= 0 (for i = j ). Now the char-
acteristic polynomial of G is given by:
G
(x) = det((x +1)I +A J)
= det((x +1)I +A) j
T
adj((x +1) I +A)j
= (1)
n
G
(x 1)
_
1 j
T
((x +1) I +A)
1
j
_
= (1)
n
G
(x 1)
_
1 n
m
i =1
2
i
x +1 +
i
_
. (1)
We place the second observation in a more general context. For V(G), let
G
denote the graph obtained from G by adding a vertex adjacent to the vertices
in , and let r be the characteristic vector of that is, r =

j
e
j
. Then G
has characteristic polynomial
x r
T
r xI A
= xdet(xI A) r
T
adj(xI A)r
=
G
(x)
_
x
m
i =1
P
i
r
2
x
i
_
.
By [46, Cor. 3.3] we have
G
{ j }
(x) = x
G
(x)
Gj
(x), and so
Gj
(x) =
G
(x)
m
i =1
2
i j
x
i
, (2)
where
i j
= P
i
e
j
. The numbers
i j
are called the angles of G; they are actually
the cosines of the angles between the coordinate axes and the eigenspaces. We may
label the vertices of G so that the columns of the m n matrix (
i j
) are ordered
lexicographically; then (
i j
) is a graph invariant, called the angle matrix of G.
Since the diagonal entries of A
k
are

m
i =1
k
i
2
i j
( j = 1, 2, . . . , n), the columns
of the angle matrix are all the same if and only if, for each k N, the number of
j - j walks of length k is independent of the vertex j ; such a graph is said to be
walk-regular (see [30]).
It follows from (2) that, given the spectrum of G, a knowledge of the angles
of G is equivalent to a knowledge of the spectra of the vertex-deleted subgraphs
of G. (This is the context in which to view Theorem 7.3 of [46]; in particu-
lar, if {x
1
, x
2
, . . . , x
k
i
} is an orthonormal basis for the eigenspace E(
i
), then
k
i
h=l
(e
1
x
h
)
2
=
2
i 1
.)
The angles of G satisfy the following relations (see [21, Ch. 4]):
n
j =1
2
i j
= k
i
and
m
i =1
2
i j
= 1.
Since the number of j - j walks of length k in G is
m
i =1
k
i
2
i j
(see [46, Thm. 2.4]),
the spectrum and angles of G determine the vertex-degrees in G. It is also the case
that the spectrum and angles determine the numbers of 4-cycles and 5-cycles. On
the other hand, the following example shows that a graph may not be determined
by its angles, main angles and spectrum.
Example 1 The two graphs depicted in Fig. 1 are non-isomorphic, but they are
both 4-regular and have the same eigenvalues, the same angles and the same main
angles. The ten vertices are labelled so that the angle sequences (
1 j
,
2 j
, . . . ,
mj
)
1 9 2
7 10 8
3
5 6
4
1 10 2
7 9 8
6
5
3
4
Fig. 1.
coincide for j = 1, 2, . . . , 10; equivalently, for each j , the graphs obtained by
deleting the vertex j are cospectral.
It has been shown by a computer search (see [15]) that graphs with fewer than
10 vertices are characterized by their eigenvalues and angles. However, there are
58 pairs of cospectral graphs on 10 vertices with the property that the graphs within
each pair have the same angles. Moreover, they also have the same main angles (a
fact for which we do not have an explanation), and no multiple eigenvalue is a main
eigenvalue. By (1), the characteristic polynomial of a complementary graph G is
determined by the characteristic polynomial and the main angles of G, and so the
graphs from29 of the 58 pairs are the complements of those fromthe other 29 pairs.
A construction from [21, pp. 113114] shows that there is an innite set of
cospectral trees with the same angles. The trees in the smallest example given
there have order 35, but an exhaustive computer search has revealed that there is
just one example among trees with up to 20 vertices (see [15]). The trees from
this pair have 19 vertices, and it is surprising that there are no examples with 20
vertices. The trees in question are displayed in Fig. 2 as T
1
and T
2
. The subtree T
identied by the heavy lines is well known in constructions of cospectral graphs,
mainly because the graphs T 4 and T 7 are cospectral.
The vertices in T
1
and T
2
are labelled so that T
1
i is cospectral with T
2
i, for
i = 1, 2, . . . , 19. Note that T
1
5 and T
2
5 both have two components with 10
and 8 vertices, which however are not cospectral. In T
1
5 the components have
the following spectra (where non-integer eigenvalues are given to three places of
decimals):
2.074, 1.414, 1.414, 0.835, 0, 0 and 2.222, 1.240, 0.726, 0, 0.
On the other hand, in T
2
5 the two components have the spectra
2.222, 1.414, 1.240, 0.726, 0, 0 and 2.074, 1.414, 0.835, 0, 0.
If we delete vertex 6, the components even have different numbers of vertices, yet
T
1
6 and T
2
6 are still cospectral.
1 2 3 4 5
9
7 8
6
10
11
12 15
14
13 16
18 19
17
T
1
:
1 2 3 4 5
9
7 8
6
16
15
14
13 10
11
12
18 19
17
T
2
:
Fig. 2.
If we try to generalize this example, we encounter difculties. Suppose that we
formthe graph H
1
by attaching any two rooted graphs K and L at vertices 4 and 7 of
T, and then form H
2
by interchanging L and K. The formula for the characteristic
polynomial of a graph obtained by coalescing rooted graphs in this way (see [21, p.
159]) shows that the following pairs are cospectral: H
1
and H
2
, H
1
i and H
2
i
for i = 4, 7 or any vertex i in K or L. However, for other values of i , the pairs are
not cospectral, except for the special graphs K = T
1
and L = T
2
shown in Fig. 2.
An exhaustive search for cospectral graphs on 10 vertices (see [36]) shows that
there exists a set S of 21 cospectral graphs with 10 vertices and 20 edges. The
complements of these graphs are also cospectral and have 25 edges. Computations
show also that, in both cases, the graphs are distinguished by their angles (see
[10]). We reproduce here some data concerning the graphs in S:
eigenvalues:
4.380, 1.686, 1.162, 0.542, 0, 0, 1.295, 1.526, 2.286, 2.663;
coefcients of the characteristic polynomial:
1, 0, 20, 18, 84, 76, 119, 72, 56, 0, 0;
main angles:
0.956, 0.025, 0.066, 0.151, 0.207, 0.044, 0.109, 0.019, 0.032.
These huge sets of cospectral graphs should perhaps be exploited in experiments
to order graphs by their angles, for the following reason. Experience shows that
it is appropriate to order graphs by their eigenvalues or spectral moments; then
cospectral graphs remain to be ordered, and it is natural to use angles for this
purpose because they determine the vertex-degrees.
Although graphs cannot in general be characterized by eigenvalues and angles,
for certain classes of graphs (for example, trees, unicyclic graphs, bicyclic graphs,
tree-like cubic graphs) it is feasible to construct all the graphs in a given class with
prescribed eigenvalues and angles. Details may be found in [21, Ch. 5].
We conclude this section by noting another consequence of (2). Since P
i
= 0
we can always nd a vertex j such that P
i
e
j
= 0, and then the multiplicity of
i
as an eigenvalue of G j is k
i
1. Repeated application of this argument shows
that, for each i {1, 2, . . . , m}, there exists a set Y
i
of vertices such that
|Y
i
| = k
i
and
GY
i
(
i
) = 0. (3)
We shall see in the next section that there always exists a partition of the vertex-set
V(G) = {Y
1
, Y
2
, . . . ,Y
m
} such that (3) holds for each i . Such a partition is called
a polynomial partition for G.
3. Star sets and star partitions
Since the column space of P
i
is E(
i
), we know that, for each i = 1, 2, . . . , m,
there exists a set X
i
of k
i
vertices such that the vectors P
i
e
j
( j X
i
) form a basis
for E(
i
). Such a set is called a star set for
i
in G. (This terminology reects the
fact that the vectors P
i
e
j
( j = 1, . . . , n) form a eutactic star, as dened by Seidel
[47].) A star partition for G is a partition
V(G) = {X
1
, X
2
, . . . , X
m
} (4)
such that X
i
is a star set for
i
. In this situation, if B
i
= {P
i
e
j
: j X
i
}, then
B
1
B
2
. . . B
m
is a basis for R
n
and is called a star basis associated with G.
Note that G is determined by its spectrum and a star basis. The following result
[21, Thm. 7.2.9] shows that the star partitions for G are precisely the polynomial
partitions for G.
Theorem 3.1 Let Gbe a graph and Xbe a subset of V(G). Then for any eigenvalue
i
of G with multiplicity k
i
, the following statements are equivalent:
(1) {P
i
e
j
: j X} is a basis of E(
i
).
(2) R
n
= E(
i
) V, where V = e
j
: j / X.
(3) |X| = k
i
, and
i
is not an eigenvalue of G X.
We now prove the fundamental existence theorem.
Theorem 3.2 Every graph has a star partition.
Proof Let {x
1
, x
2
, . . . , x
n
} be a basis of R
n
, obtained by stringing together arbi-
trary xed bases of E(
1
), E(
2
), . . . , E(
m
); say, E(
i
) has basis {x
h
: h R
i
},
where {R
1
, R
2
,. . ., R
m
} is a xed partition of {1, 2, . . . , n}. Let T be the transition
matrix from the basis {x
1
, x
2
, . . . , x
n
} to the basis {e
1
, e
2
, . . . , e
n
}. Thus T = (t
hj
),
where
e
j
=
n
h=1
t
hj
x
h
( j = 1, 2, . . . , n).
On projecting orthogonally onto E(
i
), we have
P
i
e
j
=
hR
i
t
hj
x
h
. (5)
We say that a partition {C
1
, C
2
, . . . , C
m
} of {1, 2, . . . , n} is feasible if |C
i
| = k
i
,
for i = 1, 2, . . . , m. For such a partition let T
i
be the k
i
k
i
submatrix of Twhose
rows are indexed by R
i
and whose columns are indexed by C
i
. The corresponding
multiple Laplacian development of det T has the form
det T =
i =1
det T
i
_
,
where the sum is taken over all n!/(k
1
! k
2
! k
m
!) feasible partitions. Since T
is invertible, some term
m
i =1
det T
i
is non-zero say, that determined by the
partition {X
1
, X
2
, . . . , X
m
}. This partition is a star partition because, in view of
(5), the invertibility of T
i
guarantees that each x
h
(h R
i
) is a linear combination
of the vectors P
i
e
j
( j X
i
).
This existence theorem may be strengthened as follows: if X is any star set for
an eigenvalue
i
, then there exists a star partition (4) with X
i
= X (see [21, Thm.
7.4.5]). In any case, we can always label the n vertices of G with the n eigenvalues
of G in such a way (in general, not unique) that the vertices labelled
i
constitute
a star set for
i
. This is how the vertices are labelled in the examples shown in
Fig. 3. The second example there is one of 750 star partitions of the Petersen graph;
these fall into ten isomorphism classes determined by the automorphism group of
the graph (see [21, Sec. 7.7]).
An important consequence of Theorem 3.2 is that we can associate with any
graph a star basis which is canonical, in the sense that two graphs are isomorphic
if and only if they have the same spectrum and the same canonical basis. To see
why this is possible, note rst that a graph has only nitely many star partitions
and hence determines only nitely many star bases; then, as a canonical star
basis, we can take one that is extremal in some lexicographical ordering of bases.
This crude approach is grossly inefcient, involving as it does all permutations of
coordinates, and signicant improvements in complexity have been achieved by
the use of a recursive procedure for ordering vertices and star bases simultaneously
(see [21, Ch. 8]). We remark here only that one can always nd a star partition in
3 1
1 1
1
1 1
3
1
1
1 1
1
3
2
2 2
2
Fig. 3.
polynomial time this is a consequence of Edmonds matroid intersection theorem
[28]. If one could nd a canonical star basis in polynomial time, then we would
have a polynomial algorithm for determining whether two graphs are isomorphic.
This graph isomorphism problem was the original motivation for introducing star
partitions, but as we shall see in the next section, it transpires that individual star
sets are important in their own right.
4. Star complements
Let X be a star set for the eigenvalue of the graph G, and let H = G X. We call
H a star complement for in G; it is the subgraph of G induced by the complement
X of X in V(G). A star complement for is called a -basic subgraph in [29].
Here we discuss the inuence of a star complement on the structure of a graph (see
[21, Ch. 7]).
Theorem 4.1 Let X be a star set for in G.
(a) If = 0, then X is a dominating set in G.
(b) If = 1 or 0, then X is a location-dominating set in G that is, the
X-neighbourhoods of vertices in X are distinct and non-empty.
In addition to the notation above, let |X| = |V(G)| t , so that t = |V(H)| =
codimE(). It follows from Theorem 4.1 that if = 1 or 0 then |X| < 2
t
,
and hence |V(G)| < t +2
t
, a bound that can be improved to t +
1
2
(t 1)(t +4)
when t > 1 (see [39]). This bound of the form
1
2
t
2
+ O(t ) is asymptotically
best possible as t , because in the line graph L(K
t
) the eigenspace of 2
has codimension t . When = 0, non-adjacent vertices in X with the same X-
neighbourhood are called duplicate vertices; and when = 1, adjacent vertices
in X with the same X-neighbourhood are called coduplicate vertices (see [29]).
Since duplicate or coduplicate vertices correspond to repeated rows of I A,
it is clear that |V(G)| cannot be bounded in terms of t when = 1 or 0. However,
these exceptional values do not normally obstruct our arguments, because we can
specify a graph to within duplicate or coduplicate vertices. A graph without such
vertices is called a core graph. It follows from a theorem of Kotlov and Lov asz
[35] that if = 1 or 0 and G is a core graph, then |V(G)| = O(2
t /2
).
In view of the foregoing remarks, there are only nitely many graphs (nitely
many core graphs when = 1 or 0) with a star complement of prescribed order,
or equivalently, with an eigenspace of prescribed codimension. (The graphs with
an eigenspace E() ( = 1, 0) of codimension at most 5 are determined in [43].)
There are relatively few graphs with a prescribed star complement, and this is the
basis for characterizing graphs by star complements, as illustrated below.
The next result shows that any graph is determined uniquely by
(i) an eigenvalue ;
(ii) a star complement H for ;
(iii) the embedding of H in G.
Given (ii) and (iii), we know all the edges between vertices in X and all the edges
between X and X; then knowledge of enables us to nd all the edges between
vertices of X. This establishes the role of a single eigenvalue in determining the
structure of a graph.
Theorem 4.2 (Reconstruction theorem) Let X be a star set for in the graph
G. If G X and G X have adjacency matrices A
X
and C, respectively, then G
has an adjacency matrix of the form
_
A
X
B
T
B C
_
,
where
I A
X
= B
T
(I C)
1
B. (6)
Proof Clearly the adjacency matrix of G has the form A =
_
A
X
B
T
B C
_
for some
matrix B. We have
I A =
_
I A
X
B
T
B I C
_
,
where I Cis invertible. In particular, if |X| = k then the matrix (B | I C)
has rank n k. But I Ahas rank n k, and so the rows of (B | I C) form
a basis for the row space of I A. Hence there exists a k (n k) matrix L
such that
(I A
X
| B
T
) = L(B | I C).
Now I A
X
= LB and B
T
= L(I C), and the result then follows by
eliminating L.
The converse of Theorem 4.2 is also true: if (6) holds and A
X
has size k k, then
the null space of I A consists of the vectors
_
x
(I C)
1
Bx
_
, where x R
k
.
We now consider a means of constructing the graphs having H as a star com-
plement for , by adding to H a suitable set X of vertices (the star complement
technique). Such a graph is specied by the H-neighbourhoods of vertices in X,
or equivalently by the columns b
u
(u X) of the matrix B from (6). In view of
Theorem 4.2 and its converse, we may add the set X if and only if we can add each
pair of vertices in X. Explicitly, by equating entries in (5), we see that X is a star
set for if and only if
b
T
u
(I C)
1
b
u
= , for all u X, (7)
and b
T
u
(I C)
1
b
v
{1, 0}, for all distinct u, v X. (8)
Todescribe all the graphs with H as a star complement for , it sufces todetermine
those graphs for which X is maximal, since any graph with H as a star complement
for is an induced subgraph of such a graph.
Example 2 We illustrate the star complement technique in the simple case when
= 1 and H is the 5-cycle 123451. To check (7) and (8), we simply sumthe entries
in an appropriate submatrix of (I C)
1
, which in this case is the circulant matrix
with rst row (1, 0, 1, 1, 0). We nd from (7) that we may add a single vertex
u to H if its H-neighbourhood N(u) consists of either (a) a single vertex (u), or
(b) three consecutive vertices of the 5-cycle. By (8), two distinct vertices u and v
may be added in just four ways as follows (where u v means that the vertices u
and v are adjacent):
r
both are of type (a), (u) (v), u v;
r
both are of type (a), (u) (v), u v;
r
both are of type (b), |N(u) N(v)| = 2, u v;
r
u is of type (a), v is of type (b), (u) v, but (u) is not the middle vertex of
N(v), u v.
It follows that at most two vertices of type (b) may be added, and that the maximal
graphs with H as a star complement for the eigenvalue 1 are those shown in Fig. 4
(with 2, 1, 0 vertices of type (b), respectively): in the gure, the vertices of H are
shown in black.
Fig. 4.
The general problem here is to nd all of the graphs having a given graph as a
star complement, or equivalently, to nd all of the solutions A
X
, B, of (6) for a
given matrix C. The restricted problem is to nd the solutions A
X
, B of (6) for a
given matrix C and a given eigenvalue (see Example 2). We give the results of
six such investigations; in some cases we can characterize certain graphs as those
that are maximal extensions of a prescribed star complement.
(1) [41, Thm. 3.6] If H

= K
1,5
and = 1, then = 1 and G is an induced
subgraph of the Clebsch graph (see [7, p. 35]). If H

= K
1,5
and = 1,
then the core subgraph of G is an induced subgraph of one of two graphs
with 15 and 16 vertices.
(2) [34, Thm. 3.1] If H

= K
2,5
, = 1 and |X| > 1, then = 1 and G is an
induced subgraph of the Schl ai graph (see [7, p. 32]). All of the regular
graphs with K
2,5
as a star complement have been determined.
(3) [42, Thm. 2.2 and Cor. 2.3] If G is r-regular, H

= K
1,r
and 0 < r = , then
r = (
2
+3 +1) and |V(G)| = (
2
+3)
2
. Furthermore, if = 1, then
G is the Clebsch graph; if = 2, then G is the Higman-Sims graph (see [7,
p. 107]).
(4) [33, Thm. 7.3] If H

= 6K
1
K
1,16
, = 2 and |X| is maximal, then G is the
McLaughlin graph (see [3, p. 373]).
(5) [2, Thm. 2.4] If H

= C
t
(t odd, t 5) and = 2, then G is an induced
subgraph of the line graph L(K
t
).
(6) [22, Thm. 2.1] If H

= K
1,t 3
2K
1
(t 4, t = 8) and = 2, then G is an
induced subgraph of a generalized line graph L(K
t u
; u, 0, 0, . . . , 0), for
0 u t 3 (see [7, p. 52]).
In (5) and (6), is equal to the least eigenvalue 2 of G. The star complements
for 2 that can arise in this situation are discussed in [22] and [23]. A connected
graph with least eigenvalue
n
2 which is not a generalized line graph is called
an exceptional graph, and the problem of determining all the exceptional graphs
has only recently been solved (see [17]). The solution, described in the next section,
uses the star complement technique to nd those exceptional graphs that are maxi-
mal, in the sense that any exceptional graph is an induced subgraph of such a graph.
5. Exceptional graphs
If the graph G with adjacency matrix Ahas least eigenvalue
n
2, then A +2I
is the Gram matrix of a set S of n vectors with the property that the angle between
any two of them is 60
or 90
. It is well known that if G is exceptional, then

G is representable in E
8
(see [3, Ch. 3] or [7, Ch. 3]): this means that S can be
taken to be a subset of the root system E
8
dened as follows (see also Chapter 1,
Table 2). If {e
1
, e
2
, . . . , e
8
} is an orthonormal basis for R
8
, then E
8
consists of the
112 vectors of the forme
i
e
j
(i = j ), together with the 128 vectors of the form
1
2
8
i =1
i
e
i
, where
i
1 and
8
i =1
i
= 1. The root system E
7
consists of the
vectors in E
8
orthogonal to a xed vector in E
8
, while E
6
consists of the vectors
in E
8
orthogonal to a pair of vectors at 60
.
The connected graphs with least eigenvalue
n
> 2 were determined by Doob
and Cvetkovi c [27] in 1979; accordingly, we consider only graphs with least eigen-
value 2. It was proved in [23] that such a graph is exceptional if and only if it
has an exceptional star complement for 2. Such a star complement has least
eigenvalue greater than 2 (by interlacing), and hence is one of 573 graphs of the
following types (see [27]):
type I: one of 20 graphs on 6 vertices representable in E
6
;
type II: one of 110 graphs on 7 vertices representable in E
7
, but not E
6
;
type III: one of 443 graphs on 8 vertices representable in E
8
, but not E
7
.
The graphs of type III are one-vertex extensions of graphs of type II, which are in
turn one-vertex extensions of graphs of type I. Vectors in E
8
which represent the
additional vertices are referred to below as extension vectors. The 443 graphs of
type III are describedin[5]. The 110graphs of type II are identiedin[17] bymeans
of the list of 7-vertex graphs in [12]. The 20 graphs of type I are identied in [20]:
they belong to the family F of 31 minimal forbidden subgraphs that characterize
generalized line graphs, the other eleven having 2 as their least eigenvalue (see
[14]). Accordingly, a graphis exceptional if andonlyif its least eigenvalue is greater
than or equal to 2 and it contains a graph of type I as an induced subgraph (see
[23, Prop. 3.1]). Since F was determined in [27] independently of root systems,
the star complement technique may be used to determine the exceptional graphs
without recourse to root systems.
Let G be an exceptional graph with adjacency matrix A, let H be a star comple-
ment for the eigenvalue 2, and let t = 6, 7 or 8. If H is representable in E
t
, then
so is G, since A +2I is the Gram matrix of vectors which span a t -dimensional
space (see [6]). If G is a maximal exceptional graph, then H is of type III, since
otherwise G is representable in E
7
, and then within E
8
we could add an extension
vector to obtain a larger exceptional graph.
We say that a graph is H-maximal if it is maximal with respect to the property
of having H as a star complement for the eigenvalue 2. In order to describe the
exceptional graphs representable in E
6
, it sufces to nd the H-maximal graphs
for star complements H of type I; for those representable in E
7
, but not E
6
, we
take H to be of type II; while if H is of type III, the H-maximal graphs are
precisely the maximal exceptional graphs. Ten H-maximal graphs arise when H
is of type I, and they are described in [17, Ex. 5]. When H is of type II there are
39 H-maximal graphs, and some details are given in [23, Sec. 3]. For H of type
III, Lepovi c used a computer to determine the maximal star sets that can be added
to H (see [29, Algorithm 2.4], and [17]). He found that there are 473 maximal
exceptional graphs in all. Moreover, the results reveal how these graphs can be
constructed independently of a computer search (see [18] and [24]).
To describe some of these graphs, recall from[7, p. 59] that two graphs with the
same vertex-set V are switching-equivalent if, for some partition of V(G) into two
sets P and Q, the non-edges and edges between P and Q are interchanged; for
example, the two graphs of Fig. 1 are switching-equivalent, with P = {1, 2, 7, 8}
and Q = {3, 4, 5, 6, 9, 10}.
First we give the distribution of maximal exceptional graphs over the number
of vertices:
number of vertices: 22 28 29 30 31 32 33 34 36
number of graphs: 1 1 432 25 7 3 1 2 1
The 432 graphs on 29 vertices include 430 cones over graphs switching-equivalent
to the line graph L(K
8
). Of the other 43 maximal exceptional graphs, 37 are exten-
sions of such cones, with maximumdegree 28, while the remaining six graphs have
maximum degree less than 28. None of the maximal exceptional graphs is regular,
but the list includes graphs that are of interest for a variety of reasons. For instance,
several graphs having only integer eigenvalues arise; such graphs are discussed in
Section 8. The examples here include some that have only three distinct eigenval-
ues but are not strongly regular. The two graphs on 34 vertices are cospectral, but
can be distinguished by their angles. All but a few of the graphs have the graph
K
1,2
5K
1
as one of several star complements for 2; in particular, this provides a
means of constructing the maximal exceptional graphs that are not 29-vertex cones.
6. Reconstructing the characteristic polynomial
Here we consider one of four reconstruction problems discussed by Schwenk [45]
(see also [12, Section 3.5] and [46, Section 12]). For any graph G with vertex-set
{1, 2, . . . , n}, let P(G) be the multiset consisting of the characteristic polynomials
Gi
(x) of the vertex-deleted subgraphs G i (i = 1, 2, . . . , n).
Problem 1 Is it true that, for n > 2, the characteristic polynomial
G
(x) of a
graph G is determined uniquely by P(G)?
For n > 2, it is known that
G
(x) is determined by the vertex-deleted subgraphs
themselves: this result is due to Tutte [53]. Problem 1 was posed by the rst author
at the 18th International Scientic Colloquium in Ilmenau in 1973, and the rst
results were obtained by Gutman and Cvetkovi c [31] (see also [13, p. 267] and [12,
pp. 6870]). No examples of non-unique reconstruction of the characteristic poly-
nomial are known. Some relations between Problem1 and the Ulamreconstruction
problem for graphs are described in [21, Sec. 5.4].
Since
G
(x) =

n
i =1
Gi
(x) (see [13, p. 60]), we can readily determine the
characteristic polynomial
G
(x), except for the constant term. If we know just one
eigenvalue of G, then the constant termis determined. In particular, this is the case
if we knowa multiple root of some polynomial
Gi
(x), for then (by interlacing)
is an eigenvalue of G.
In [31], Problem 1 was solved afrmatively for regular graphs and for a broad
class of bipartite graphs including trees without a 1-factor. Here we indicate how
the result was extended to the remaining trees in [11] and [16]. We denote by (G)
the set of distinct eigenvalues of the graph G.
Theorem 6.1 Let H be a graph of order n ( 3) with exactly two connected
components. If these components have different orders, then the characteristic
polynomial of H is determined uniquely by P(H).
Proof Suppose, bywayof contradiction, that there exists at least one graph G = H
suchthat
G
(x) =
H
(x) +a (a = 0) and
Gi
(x) =
Hi
(x), for i = 1, 2, . . . , n.
Let H
1
and H
2
be the two components of H, with H
j
of order n
j
and n
1
> n
2
.
Clearly,
(H i ) = (H
1
) (H
2
i ) (i V(H
2
)). (9)
Since H i has no repeated eigenvalues, the same is true of H
1
, and we let
(H
1
) = {
1
,
2
, . . . ,
n
1
}, where
1
>
2
> >
n
1
.
Let v be a xed vertex of H
2
. Since |(H
2
v)| < n
1
1, there exists at least
one index i = i
0
(1 i
0
n
1
1) such that no eigenvalue of H
2
v lies in the
open interval (
i
0
+1
,
i
0
). We deduce that
_
i
0
+1
,
i
0
_
(H v) = . (10)
Since
Gv
(x) =
Hv
(x), we knowfrom(9) that
i
0
and
i
0
+1
lie in (G v). By
the interlacing theorem, there exists at least one eigenvalue of G in the interval
(
i
0
+1
,
i
0
). Since
G
() = 0 and
G
(
i
0
+1
) =
G
(
i
0
) = a, there exist at least two
eigenvalues , of G in (
i
0
+1
,
i
0
).
Finally, using the interlacing theorem again, we see that G v has at least
one eigenvalue [, ] (
i
0
+1
,
i
0
). Since (G v) = (H v), this is a
contradiction to (10).
In similar vein we have also the following result.
Theorem 6.2 If H is a disconnected graph with at least three components, then
the characteristic polynomial of H is determined uniquely by P(H).
Proof Suppose that H has components H
1
, H
2
, . . . , H
k
(k > 2). We may assume
that all eigenvalues of H are simple, and hence that
1
(H
1
) >
1
(H
2
) > >
1
(H
k
). It follows that
(
1
(H
1
),
1
(H
2
)) (H
k
v) = (v V(H
k
)),
and the proof now follows as before.
In proving the next result, we make use of the fact (see [13, p. 37]) that the
characteristic polynomial of a tree with n vertices has constant term (1)
n/2
or 0,
according as it does or does not have a 1-factor. We write e(G) for the number of
edges of the graph G.
Theorem 6.3 If G is a tree, then its characteristic polynomial is determined
uniquely by P(G).
Proof Suppose, by way of contradiction, that there exists a tree G whose charac-
teristic polynomial is not uniquely determined by P(G). In viewof our introductory
remarks, we know that such a tree G has a 1-factor, and so the number of vertices
is even say n = 2k, where k > 1. Now there exists at least one graph H such
that
H
(x) =
G
(x) +a (a = 0) and
Hi
(x) = P
Gi
(x) (i = 1, 2, . . . , n). Such
a graph H is not connected, since otherwise it is a tree, the number of edges being
known fromthe coefcient of x
n2
in
H
(x). Since a = 0, this tree does not have a
1-factor, and so
H
(0) = 0. By [46, Cor. 3.3], if u is the neighbour of an end-vertex
of H, then 0 is a multiple eigenvalue of H u. Then G has 0 as an eigenvalue,
giving a contradiction.
By Theorems 6.1 and 6.2, H has exactly two connected components H
1
and
H
2
, each with exactly k vertices. Since e(H) = e(G) = 2k 1, we know that one
component, say H
1
, is a unicyclic graph and the other component H
2
is a tree. Let
1
>
2
> >
k
be the eigenvalues of H
2
. If there are a vertex v of H
1
and
an index i
0
such that (
i
0
+1
,
i
0
) (H
1
v) = , then the proof proceeds as in
Theorem 5.1. Otherwise, for any vertex v of H
1
, the eigenvalues
1
,
2
, . . . ,
k1
of H
1
v interlace those of H
2
that is,
i
(
i +1
,
i
) (i = 1, 2, . . . , k 1). (11)
Now, because it has no multiple eigenvalues, the unicyclic graph H
1
is not a cycle,
and so we may choose v to be an end-vertex. Then H
1
v is unicyclic, and we
deduce from (11) the contradiction
2(k 1) = e(H
1
v) =
k1
i =1
2
i
<
k
i =1
(
i
)
2
= 2e(H
2
) = 2(k 1).
This completes the proof.
7. Non-complete extended p-sums of graphs
Next we consider a very general graph operation called the non-complete extended
p-sum of graphs, which we abbreviate as NEPS (see [19]); these graphs were
discussed briey in Chapter 1.
Let B be a set of non-zero binary n-tuples in {0, 1}
n
. The NEPS with basis B of
graphs G
1
, G
2
, . . . , G
n
is the graphwithvertex-set V(G
1
) V(G
2
) V(G
n
),
where the vertices (x
1
, x
2
, . . . , x
n
) and (y
1
, y
2
, . . . , y
n
) are adjacent if and only if
there exists an n-tuple (
1
,
2
, . . . ,
n
) B such that x
i
= y
i
when
i
= 0, and x
i
is adjacent to y
i
in G
i
when
i
= 1.
We identify some special cases where a graph G is the NEPS with basis B of
two graphs G
1
and G
2
:
r
if B = {(0, 1), (1, 0)}, then G is the sum G
1
+ G
2
of G
1
and G
2
;
r
if B = {(1, 1)}, then G is the product G
1
G
2
of G
1
and G
2
;
r
if B = {(0, 1), (1, 0), (1, 1)}, then G is the strong product G
1
G
2
of G
1
and
G
2
.
(A variety of terms and notations for these particular constructions can be found
in the literature.)
The p-sum of n graphs is a NEPS whose basis consists of all n-tuples with
exactly p entries equal to 1. The 2-sum of three graphs G
1
, G
2
and G
3
is denoted
by D(G
1
, G
2
, G
3
).
The foregoing denitions, together with some basic results on NEPS and early
references, can be found in [13]; for a more recent review of results, see [25]. The
proofs of the next two theorems are given in [13, Sec. 2.5].
Theorem 7.1 Let A
1
, A
2
, . . . , A
n
be adjacency matrices of graphs G
1
, G
2
, . . . ,
G
n
, respectively. The NEPS G with basis B of graphs G
1
, G
2
, . . . , G
n
has as
adjacency matrix the matrix A given by
A =
B
A
1
1
A
2
2
. . . A
n
n
. (12)
Here A
0
k
is the identity matrix of the same size as A
k
, A
1
k
= A
k
, and denotes the
Kronecker product of matrices.
One consequence of Theorem 7.1 is the following result.
Theorem 7.2 If p
i
= |V(G
i
)| and if
i 1
,
i 2
, . . . ,
i p
i
is the spectrum of G
i
(i = 1, 2, . . . , n), then the spectrum of the NEPS of G
1
, G
2
, . . . , G
n
with basis
B consists of all possible values
i
1
,i
2
,...,i
n
, where
i
1
,i
2
,...,i
n
=
1
1i
1
2
2i
2
. . .
n
ni
n
(i
k
= 1, 2, . . . , p
k
; k = 1, 2, . . . , n). (13)
Together with expressions (12) and (13) we consider the function
f (x
1
, x
2
, . . . , x
n
) =
B
x
1
1
x
2
2
. . . x
n
n
, (14)
which will be of interest later.
First we quote the following theorem which lists some of the classes of graphs
identied in [25] as being closed under the NEPSoperation. An even graph is one in
which each vertex has even degree, a singular graph is one with 0 as an eigenvalue,
and a transitive graph is one whose automorphism group acts transitively on the
vertices.
Theorem 7.3 The following classes of graphs are closed under the NEPS opera-
tion:
(1) regular graphs;
(2) even graphs;
(3) singular graphs;
(4) integral graphs;
(5) transitive graphs;
(6) walk-regular graphs.
Here, (1) and (2) follow from the fact that the vertex-degrees in a NEPS can be
expressed in terms of the vertex-degrees of the graphs on which the operation
is performed, while (3) and (4) follow from Theorem 6.2. (5) follows from the
observation that the automorphism group of a NEPS contains the direct product
of the automorphism groups of the constituent graphs, while (6) follows from the
characterization of walk-regular graphs in terms of angles (see Section 2 and [25,
Remark 3.5]).
It is a well-known fact that the product of two connected bipartite graphs G
1
and
G
2
is disconnected, with two components. Similar situations arise more generally
in a NEPS, and graph eigenvalues are a good means of investigating the conditions
under which a NEPS is connected or bipartite (see [13, Sec. 7.4]).
Two graphs are almost cospectral if their non-zero eigenvalues (and their
multiplicities) coincide. In [12] it was conjectured that, if a NEPS of bipartite
graphs is disconnected, then its components are almost cospectral. This conjecture
is true for the product of graphs, and we shall show (following [9]) that it is also
true for the 2-sum of three graphs.
Connectedness properties of the NEPS, from Theorems 7.1 and 7.2, are related
to the function (14), as described in [13]; for a recent treatment, see [49]. In the
case of the 2-sumof three graphs, the function (14) is x
1
x
2
+ x
1
x
3
+ x
2
x
3
, which is
even with respect to the variables x
1
, x
2
, x
3
, in the sense that it remains unchanged
when these variables are simultaneously changed in sign. By [13, Thm. 7.12] we
conclude that the 2-sum of three bipartite graphs is a disconnected graph with two
components. For example, D(K
2
, K
2
, K
2
) = 2K
4
; geometrically, this says that if
the vertices of D(K
2
, K
2
, K
2
) are represented by the vertices of a cube, then the
2-sum consists of two tetrahedra.
The proof of the following lemma is straightforward.
Lemma 7.4 Let x
1
, x
2
, . . . , x
r
and y
1
, y
2
, . . . , y
s
be non-increasing sequences of
non-zero real numbers. If
r
i =1
x
k
i
=
s
j =1
y
k
j
for all k N,
then r = s and x
i
= y
i
(i = 1, 2, . . . , r).
We can then prove the following theorem.
Theorem 7.5 Two graphs are almost cospectral if and only if, for each k N,
they have the same number of closed walks of length k.
Proof Assume rst that G and H are two graphs that satisfy the condition con-
cerning walks. If
1
,
2
, . . . ,
m
and
1
,
2
, . . . ,
n
are the eigenvalues of G and
H, respectively, then

m
i =1
k
i
=

n
j =1
k
j
for all k N. By Lemma 7.4, G and
H are almost cospectral. The reverse implication is immediate.
We can use Theorem 7.5 to prove in a new way that if G and H are connected
bipartite graphs, then the two components A and B of the product G H are
almost cospectral. Let (x
1
, y
1
), (x
2
, y
2
), . . . , (x
k
, y
k
) be a closed walk of length k
in A. Then (x
1
, y
2
), (x
2
, y
3
), . . . , (x
k
, y
1
) is a closed walk of length k in B. For
each k 2, this mapping is a bijection between the sets of closed walks of length
k in the graphs A and B. Since no graph has any closed walks of length 1, we
deduce from Theorem 7.5 that the components A and B are almost cospectral. We
apply the same idea to the 2-sum of three graphs.
Let a
1
, a
2
, . . . , a
n
be a nite sequence of real numbers. We count the groups of
mutually equal members successively following one another in this sequence, by
dening the function recursively as follows: (1) = 1, and for i = 2, 3, . . . , n,
(i +1) =
_
(i ), if a
i +1
= a
i
(i ) +1, otherwise.
If a
n
= a
1
and (n) = k, we redene those values of (i ) equal to k to be 1
instead. Let s be the maximum value of and dene b
(i )
= a
i
(i = 1, 2, . . . , n).
The quasi-cyclic shift of the sequence a
1
, a
2
, . . . , a
n
is the sequence obtained from
a
1
, a
2
, . . . , a
n
by replacing each a
i
with b
(i )+1
(where b
s+1
= b
1
).
Theorem 7.6 The two components of the 2-sum of three bipartite graphs are
almost cospectral.
Proof Let H
1
and H
2
be the two components, and let
(x
1
, y
1
, z
1
), (x
2
, y
2
, z
2
), . . . , (x
k
, y
k
, z
k
) (15)
be a closed walk of length k ( 2) in H
1
. We construct from (15) a closed walk of
the same length in H
2
, and this will establish the required one-one correspondence.
The sequence z
1
, z
2
, . . . , z
k
has the properties that z
i
and z
i +1
are either equal
or adjacent (i = 1, 2, . . . , k 1) and z
k
= z
1
.
If not all of the vertices z
1
, z
2
, . . . , z
k
are the same, then the sequence z
1
, z
2
, . . . ,
z
k
determines a closed walk of length at most k in the corresponding graph. Let
w
1
, w
2
, . . . , w
k
be the quasi-cyclic shift of the sequence z
1
, z
2
, . . . , z
k
. It is easy
to see that the sequence
(x
1
, y
1
, w
1
), (x
2
, y
2
, w
2
), . . . , (x
k
, y
k
, w
k
)
is a closed walk of length k in H
2
.
On the other hand, if z
1
= z
2
= = z
k
= z, then
(x
1
, y
2
, z), (x
2
, y
3
, z), . . . , (x
k
, y
1
, z)
is a closed walk in H
2
corresponding to (15).
Thus, for each k, the number of closed walks of length k is the same in H
1
and
H
2
. By Theorem 7.5, the two components are almost cospectral.
It was conjectured that the components of a NEPS of bipartite graphs are always
almost cospectral, but this is not true in general (see [51]). A counter-example is
provided by the graph (P
3
P
3
) + K
2
, and modied conjectures are discussed in
[51].
8. Integral graphs
A graph is integral if all its eigenvalues are integers. The quest for integral graphs
was initiated by Harary and Schwenk [32]. The thirteen connected cubic integral
graphs were obtained by Cvetkovi c and Bussemaker (see [8] and [4]), and inde-
pendently by Schwenk [44]. Cvetkovi c [8] proved that the set of connected regular
integral graphs of any xed degree is nite.
Radosavljevi c and Simi c [48] determined the thirteen connected non-regular
non-bipartite integral graphs with maximumdegree 4. The corresponding problem
for bipartite graphs is not yet solved; see [48] for some details. Recently, Stevanovi c
[52] determined all 24 connected 4-regular integral graphs without 3 or 3 in the
spectrum.
The search for integral graphs becomes easier if we restrict ourselves to the
product of graphs. If G is connected, non-bipartite, r-regular and integral, then the
product G K
2
is connected, bipartite, r-regular and integral, since the eigenval-
ues of K
2
are 1 and 1. Therefore, in determining r-regular integral graphs we
need consider only bipartite graphs, and later extract non-bipartite graphs G from
the decompositions of bipartite graphs of the form G K
2
. On the other hand, if
G is bipartite, then G K
2
= 2G and we cannot obtain new graphs by iterating
the product with K
2
.
A system of Diophantine equations for the spectrum of a connected 4-regular
bipartite integral graph has been obtained in [26]. The equations derive from
an interpretation of the rst six spectral moments in terms of the numbers of
short cycles. The system has a nite number of solutions, and according to [26]
the graphs in question have at most 5040 vertices. The non-existence of graphs
with some of the spectra obtained in [26] was established in [50] using graph
angles. A list of the 65 known 4-regular connected integral graphs is given in
[26].
There are exactly 150 connected integral graphs with up to 10 vertices (see [1]),
and we reproduce here some of the details. The numbers i
n
of connected integral
graphs with n vertices are given for n = 1, 2, . . . , 10 in the following table:
n: 1 2 3 4 5 6 7 8 9 10
i
n
: 1 1 1 2 3 6 7 22 24 83
The connected integral graphs with up to ve vertices are easily identied from
the table of spectra of connected graphs with up to ve vertices, given in [13]: they
are K
1
, K
2
, K
3
, K
4
, C
4
, K
5
, 2K
1
K
3
and K
1,4
.
The six connected integral graphs on six vertices have identication numbers
1, 9, 51, 52, 106 and 109 in the table of connected graphs with six vertices given
in [20]. They are K
6
, the octahedron 3K
2
, C
3
+ K
2
, K
3,3
, C
6
, and the unique tree
with degree sequence (3, 3, 1, 1, 1, 1).
There are seven connected integral graphs on seven vertices, with identica-
tion numbers 4, 191, 235, 300, 710, 841 and 853 in the table of connected graphs
on seven vertices given in [12]. The connected integral graphs on eight, nine
and ten vertices are specied in [1] by their adjacency matrices, and also by
pictures in the case of eight or nine vertices. Their spectra are provided in all
cases.
There are no cospectral (non-isomorphic) connected integral graphs on fewer
than eight vertices. There is just one triplet of connected integral graphs on eight
vertices; there are three pairs on nine vertices; and ten pairs, one triplet, two
quadruplets and one quintuplet on ten vertices. None of the graphs in these sets is
regular, and in all cases the cospectral graphs can be distinguished by angles. We
have already noted that the smallest cospectral graphs with the same angles have
ten vertices (see [15]), and we nd that there are no integral graphs among the 58
pairs of such cospectral graphs.
There are integral cospectral graphs with fewer than eight vertices, but then at
least one of them is disconnected. These cases include the smallest pair of cospec-
tral graphs, K
1,4
and C
4
K
1
, which have the common spectrum (2, 0, 0, 0, 2).
Also, the subdivision graph S(K
1,3
) has as a cospectral mate the disconnected
graph C
6
K
1
, with spectrum (2, 1, 1, 0, 1, 1, 2).
Some cospectral integral graphs have the same main angles as well. This hap-
pens in several pairs: in one pair on nine vertices and in six pairs on ten vertices,
as well as in a triplet on ten vertices. Note that, in each of the two aforementioned
cospectral quadruplets, the main angles are not the same for all four graphs. The
complements of cospectral graphs sharing the same main angles are cospectral.
The complements of graphs from one cospectral quadruplet form two pairs of
cospectral graphs, but not a cospectral quadruplet. Similarly, the complements of
graphs from the other quadruplet do not provide a cospectral quadruplet; in this
case, we obtain only a cospectral triplet. However, among the complements of
graphs in all these sets of cospectral graphs, only the complements of two graphs
on nine vertices are integral. Both complements are disconnected: the rst has
three components, two of them being isolated vertices, while the second has two
components, one an isolated vertex. Cospectral integral graphs with cospectral
integral complements were studied for the rst time in [38], where an example
with eleven vertices is given. It is interesting that we have the same situation with
isolated vertices as above.
Several other graphs in our lists have integral complements; some of these
complements are disconnected, but we can readily identify the pairs of connected
complementary integral graphs. They include one pair on eight vertices, one on
nine vertices, and four on ten vertices. The smallest such pair consists of the 6-
vertex graphs C
6
and C
3
+ K
2
. All of these graphs are regular. Self-complementary
graphs also feature in our lists: examples include a non-regular graph of order 8
and two regular graphs of order 9.
References
1. K. Balinska, D. Cvetkovi c, M. Lepovi c and S. Simi c, There are exactly 150 connected
integral graphs up to 10 vertices, Univ. Beograd, Publ. Elektrotehn. Fak., Ser. Mat. 10
(1999), 95105.
2. F. K. Bell, Characterizing line graphs by star complements, Linear Alg. Appl. 296
(1999), 1525.
Verlag, 1989.
4. F. C. Bussemaker and D. Cvetkovi c, There are exactly 13 connected, cubic, integral
graphs, Univ. Beograd, Publ. Elektrotehn. Fak. Ser. Mat. Fiz. 544576 (1976), 4348.
5. F. C. Bussemaker and A. Neumaier, Exceptional graphs with smallest eigenvalue 2
and related problems, Math. of Computation 59 (1992), 583608.
6. P. J. Cameron, J. M. Goethals, J. J. Seidel and E. E. Shult, Line graphs, root systems,
and elliptic geometry, J. Algebra 43 (1976), 305327.
7. P. J. Cameron and J. H. van Lint, Designs, Graphs, Codes and their Links, Cambridge
Univ. Press, 1991.
8. D. Cvetkovi c, Cubic integral graphs, Univ. Beograd, Publ. Elektrotehn. Fak. Ser. Mat.
Fiz. 498541 (1975), 107113.
9. D. Cvetkovi c, On the 2-sum of three graphs: variations on the graph product discon-
nectedness theme, Bull. Acad. Serbe Sci. Arts, Cl. Sci. Math. Natur., Sci. Math. 118
(24) (1999), 107117.
10. D. Cvetkovi c, Characterizing properties of some graph invariants related to electron
charges in the H uckel molecular orbital theory, Proc. DIMACS Workshop on Discrete
Mathematical Chemistry, March 1998, DIMACS Ser. Discrete Math. Theoret. Comp.
Sci. 51 (2000), 7984.
11. D. Cvetkovi c, Onthe reconstructionof the characteristic polynomial of a graph, Discrete
Math. 212 (2000), 4552.
12. D. Cvetkovi c, M. Doob, I. Gutman and A. Torga sev, Recent Results in the Theory of
Graph Spectra, North-Holland, 1988.
13. D. Cvetkovi c, M. Doob and H. Sachs, Spectra of Graphs (3rd ed.), Johann Ambrosius
Barth Verlag, 1995.
14. D. Cvetkovi c, M. Doob and S. Simi c, Generalized line graphs, J. Graph Theory 5
(1981), 385399.
15. D. Cvetkovi c and M. Lepovi c, Cospectral graphs with the same angles and with a
minimal number of vertices, Univ. Beograd, Publ. Elektrotehn. Fak., Ser. Mat. 8 (1997),
88102.
16. D. Cvetkovi c and M. Lepovi c, Seeking counterexamples to the reconstruction conjec-
ture for characteristic polynomials of graphs, and a positive result, Bull. Acad. Serbe
Sci. Arts, Cl. Sci. Math. Natur., Sci. Math. 116 (23) (1998), 91100.
17. D. Cvetkovi c, M. Lepovi c, P. Rowlinson and S. Simi c, A database of star complements
of graphs, Univ. Beograd, Publ. Elektrotehn. Fak., Ser. Mat. 9 (1998), 103112.
18. D. Cvetkovi c, M. Lepovi c, P. Rowlinson and S. K. Simi c, The maximal exceptional
graphs, J. Combin. Theory (B) 86 (2002), 347363.
19. D. Cvetkovi c andR. Lu ci c, Anewgeneralizationof the p-sumof graphs, Univ. Beograd,
Publ. Elektrotehn. Fak., Ser. Mat. Fiz. 302319 (1970), 6771.
20. D. Cvetkovi c and M. Petri c, Atable of connected graphs on six vertices, Discrete Math.
50 (1984), 3749.
21. D. Cvetkovi c, P. Rowlinson and S. K. Simi c, Eigenspaces of Graphs, Cambridge Univ.
Press, 1997.
22. D. Cvetkovi c, P. Rowlinson and S. K. Simi c, Some characterizations of graphs by star
complements, Linear Alg. Appl. 301 (1999), 8197.
23. D. Cvetkovi c, P. Rowlinson and S. K. Simi c, Graphs with least eigenvalue 2: the star
complement technique, J. Alg. Combin. 14 (2001), 516.
24. D. Cvetkovi c, P. Rowlinson and S. K. Simi c, Constructions of the maximal exceptional
graphs with largest degree less than 28, Bull. Acad. Serbe Sci. Arts, Cl. Sci. Math.
Natur., Sci. Math. 122 (26) (2001), 115131.
25. D. Cvetkovi c and S. Simi c, Non-complete extended p-sum of graphs, graph angles and
star partitions, Publ. Inst. Math. (Beograd) 53(67) (1993), 416.
26. D. Cvetkovi c, S. Simi c and D. Stevanovi c, 4-regular integral graphs, Univ. Beograd,
Publ. Elektrotehn. Fak., Ser. Mat. 9 (1998), 89102.
27. M. Doob and D. Cvetkovi c, On spectral characterizations and embeddings of graphs,
Linear Alg. Appl. 27 (1979), 1726.
28. J. Edmonds, Matroid Intersection, Discrete Optimization 1 (ed. P. L. Hammer et al.)
Annals of Discrete Mathematics 4, North-Holland (1979), 3949.
29. M. N. Ellingham, Basic subgraphs and graph spectra, Austral. J. Combin. 8 (1993),
247265.
30. C. D. Godsil and B. D. McKay, Feasibility conditions for the existence of walk-regular
graphs, Linear Algebra Appl. 30 (1980), 5161.
31. I. Gutman and D. Cvetkovi c, The reconstruction problem for characteristic polynomi-
als of graphs, Univ. Beograd, Publ. Elektrotehn. Fak. Ser. Mat. Fiz. 498541 (1975),
4548.
32. F. Harary and A. J. Schwenk, Which graphs have integral spectra?, Graphs and Com-
binatorics, Proc. Conf. Graph Theory and Combinatorics, George Washington Univ.,
June 1973 (eds. R. Bari and F. Harary), Springer-Verlag (1974), 4551.
33. P. S. Jackson, Star sets and related aspects of algebraic graph theory, PhD thesis,
University of Stirling, 1999.
34. P. S. Jackson and P. Rowlinson, On graphs with complete bipartite star complements,
Linear Alg. Appl. 298 (1999), 920.
35. A. Kotlov and L. Lov asz, The rank and size of graphs, J. Graph Theory 23 (1996),
185189.
36. M. Lepovi c, Some statistical data on graphs with 10 vertices, Univ. Beograd, Publ.
Elektrotehn. Fak., Ser. Mat. 9 (1998), 7988.
37. A. Neumaier, Derived eigenvalues of symmetric matrices, with applications to distance
geometry, Linear Alg. Appl. 134 (1990), 107120.
38. Z. Radosavljevi c and S. Simi c, Computer aided search for all graphs such that graph
and its complement have its spectrum bounded from below by 2, Ars Combin. 24A
(1987), 2127.
39. P. Rowlinson, The characteristic polynomials of modied graphs, Discr. Appl. Math.
67 (1996), 209219.
40. P. Rowlinson, On graphs with multiple eigenvalues, Linear Alg. Appl. 283 (1998),
7585.
41. P. Rowlinson, Star sets and star complements in nite graphs: a spectral construction
technique, Proc. DIMACSWorkshoponDiscrete Mathematical Chemistry, March1998,
DIMACS Ser. Discrete Math. Theoret. Comp. Sci. 51 (2000), 323332.
42. P. Rowlinson, Star sets in regular graphs, J. Combin. Math. Combin. Computing 34
(2000), 322.
43. P. Rowlinson and F. K. Bell, Graph eigenspaces of small codimension, Discrete Math.
220 (2000), 271278.
44. A. J. Schwenk, Exactly thirteen connected cubic graphs have integral spectra, Theory
and Applications of Graphs, Proc. Internat. Conf. Western Michigan Univ., Kalamazoo,
Mich., May 1976 (eds. Y. Alavi and D. Lick), Springer-Verlag (1978), 516533.
45. A. J. Schwenk, Spectral reconstruction problems, Ann. NewYork Acad. Sci. 328 (1979),
183189.
46. A. J. Schwenk and R. J. Wilson, On the eigenvalues of a graph, Selected Topics in Graph
Theory (eds. L. W. Beineke and R. J. Wilson), Academic Press (1978), 307336.
47. J. J. Seidel, Eutactic stars, Combinatorics (eds. A. Hajnal and V. T. S os), North-Holland
(1978), 983989.
48. S. Simi c and Z. Radosavljevi c, The nonregular, nonbipartite, integral graphs with max-
imum degree four, J. Combin. Inf. Syst. Sci. 20 (1995), 926.
49. D. Stevanovi c, When is NEPS of graphs connected?, Linear Alg. Appl. 301 (1999),
137143.
50. D. Stevanovi c: Nonexistence of some 4-regular integral graphs, Univ. Beograd, Publ.
Elektrotehn. Fak., Ser. Mat. 10 (1999), 8186.
51. D. Stevanovi c, When can the components of NEPS of connected bipartite graphs be
almost cospectral?, Linear Alg. Appl. 311 (2000), 3544.
52. D. Stevanovi c, 4-regular integral graphs avoiding 3 in the spectrum, Univ. Beograd,
Publ. Elektrotehn. Fak., Ser. Mat. 14 (2003).
53. W. T. Tutte, All the kings horses: a guide to reconstruction, Graph Theory and Related
Topics (eds. J. A. Bondy and U. S. R. Murty), Academic Press (1979), 1533.
4
Graph Laplacians
BOJAN MOHAR
1. Introduction
2. The Laplacian of a graph
3. Laplace eigenvalues
4. Eigenvalues and vertex partitions of graphs
5. The max-cut problem and semi-denite programming
6. Isoperimetric inequalities
7. The travelling salesman problem
8. Random walks on graphs
References
Eigenvalues of Laplace matrices of graphs are related to various combina-
torial properties of graphs. They can be used to extract information about
some parameters that are hard to compute or estimate, most notably those
related to expansion and vertex partitions. The smallest and largest eigen-
value can be expressed as solutions to a quadratic optimization problem. It
turns out that the right generalized setting for this is semi-denite program-
ming, where duality theory leads to powerful applications, one of which is
given in Section 5. A more general setting of weighted graphs is presented,
which brings us also to the study of simple random walks whose transition
matrix can be expressed via the related Laplacian.
1. Introduction
Eigenvalues of graphs have been extensively studied since the early developments
of graph theory in the 1960s (see, for example, Chapter 1, [4], [11], [12] or [23]).
Most of the early works considered eigenvalues of adjacency matrices of graphs.
New developments in the 1980s made it clear that eigenvalues and eigenvectors
113
114 Bojan Mohar
of the related Laplace matrices of graphs enter the theory in several applications
more naturally than eigenvalues of adjacency matrices (see [8], [37] and [40]), and
therefore deserve their own study.
Many properties of eigenvalues of graphs can be expressed in terms of simple
graph invariants, such as vertex degrees or the number of short cycles. Such results
are most interesting from the point of view of linear algebra. However, from
the viewpoint of applications in combinatorics, it is more interesting to study
relations between eigenvalues and hard graph invariants. Such an approach was
very successful in the study of distance-regular graphs; we refer to the excellent
monographs [4] and [23]. One of the most fascinating aspects of such applications
is that eigenvalues appear as a tool to prove results that appear to have nothing to
do with the eigenvalues themselves.
Applications of eigenvalue methods in combinatorics, graph theory and combi-
natorial optimization have a long history. For example, eigenvalue bounds on the
chromatic number were formulated by Wilf [48] and Hoffman [28] in the 1960s.
Another early application, in the area of graph partition, is due to Fiedler [17] and
Donath and Hoffman [14]. An important use of eigenvalues is Lov aszs notion of
the -function [31].
The next important result was the use of eigenvalues in the construction of
superconcentrators and expanders by Alon and Milman [1]. Isoperimetric prop-
erties of graphs and their eigenvalues play a crucial role in the design of various
randomized algorithms. These applications are based on the so-called rapidly
mixing Markov chains (see Section 8).
There is an increasing interest in the application of eigenvalues to combinato-
rial optimization problems. For example, an eigenvalue approach was used in the
study of the quadratic assignment problem and general graph partition problems
[43], the max-cut problem [13], and labelling problems ([25] and [29]). Spectral
partitioning, which is based on eigenvectors of Laplace eigenvalues of graphs, has
proved to be a successful heuristic approach in the design of partition algorithms
([25] and [27]), in parallel computation [45], clustering [24], ranking ([25] and
[29]), and in graph drawing [19]. We refer to [40] for additional applications.
Laplace matrices of graphs are closely related to the Laplacian, the second
order differential operator ( f ) = div(grad f ). This relation yields an important
bilateral link between the spectral geometry of Riemannian manifolds and graph
theory, and makes it possible to use results about graphs in the study of Laplacians
on manifolds and, conversely, to transfer results about Laplacians on manifolds
to graphs. Each of these directions has given rise to exciting new discoveries (see
Buser [7], Brooks [6] and Chung [8]). An interesting approach, using Laplace
matrices and discrete Schr odinger operators to obtain certain minor monotone
graph invariants, has been developed by Colin de Verdi` ere. His remarkable results,
whichconnect withseveral central areas of mathematics, are nicelypresentedin[9].
In this chapter, only basic information on the Laplace matrices of graphs is
presented. For further and more specialized results, there are several books and
survey papers, such as [4], [8], [9], [11], [33] and [40].
2. The Laplacian of a graph
Let G = (V, E) be an undirected nite graph with n vertices and m edges. We
also consider weighted graphs with a weight function that assigns a non-negative
real weight a
vw
to each pair v, w of vertices. We require that the weights satisfy
the following properties:
r
a
vw
> 0 if vw E, and a
vw
= 0 if vw / E;
r
a
vw
= a
wv
, for all v, w V.
Unweighted graphs can be viewed as weighted graphs in which a
vw
is the number
of edges between v and w.
In the unweighted case, the degree deg v of a vertex v V is the number of
edges of G incident with v. In the weighted case, it is dened by
deg v =
wV
a
vw
.
Recall that the maximum and minimum vertex degrees in G are denoted by (G)
and (G), respectively.
Given a graph G, its (weighted) adjacency matrix A(G) = (a
vw
) is the n n
matrix, with rows and columns indexed by V, whose entries are the edge-weights.
The degree matrix D(G) = diag(deg v : v V) is the diagonal matrix indexed by
V with the vertex-degrees on the diagonal. The difference
L(G) = D(G) A(G) (1)
is the Laplace matrix (or Laplacian) of G.
One of the rst applications of Laplace matrices of graphs is in the statement
of the well-known matrix-tree theorem (see [4, Ch. 6]).
Theorem 2.1 (Matrix-tree theorem) Let G be a graph and let L = L(G) be its
Laplace matrix. If T is a spanning tree of G, let a(T) be the product of the edge-
weights of T. For any vertices v and w, the (weighted) number of spanning trees
of G is
(G) =
{a(T) : T a spanning tree of G] = [det L

v
w
[,
where L
v
w
is the submatrix of L obtained by deleting the column corresponding to
v and the row corresponding to w.
116 Bojan Mohar
The most common proofs of Theorem 2.1 use the related oriented incidence
matrix Q of G, which is dened as follows. Orient arbitrarily each edge of G.
Then the matrix Q is the n m matrix with rows and columns indexed by V and
E, whose (v, e)-entry is
a
vw
if v is the initial vertex of e = vw,
a
wv
if v is
the terminal vertex of e = wv, and 0 if v and e are not incident. If e is a loop, then
the (v, e)-entry is 0, irrespective of its weight.
Theorem 2.2 Let G be a graph, and let Q be an oriented incidence matrix with
respect to some orientation of its edges. Then
L(G) = QQ
T
.
Theorem 2.2 shows that the product QQ
T
is independent of the chosen orientation
of the edges of G.
Let R
V
= {f : V R] be the set of functions fromV to R. If f R
V
and v V,
denote by f
v
the value of f at the vertex v. With the usual operations of sum and
scalar multiplication, R
V
becomes a real vector space of dimension n, endowed
with the inner product
f, g) =
vV
f
v
g
v
.
The corresponding norm in R
V
is
|f| = f,f)
1/2
=
_
vV
f
2
v
_
1/2
.
The matrix L (and other matrices indexed by the vertices of G) acts on R
V
as a
linear operator. Its action is determined by the rule of matrix-vector multiplication,
where g = Lf is the function dened by the formula
g
v
= (Lf)
v
=
wV
L
vw
f
w
=
wV
a
vw
(f
v
f
w
) (v V).
There is a natural quadratic form associated with L.
Theorem 2.3
f, Lf) =
vwE
a
vw
(f
v
f
w
)
2
. (2)
Proof Since L = QQ
T
, the denition of Q implies that
f, Lf) = f, QQ
T
f) = Q
T
f, Q
T
f) =
eE
(Q
T
f)
2
e
=
vwE
a
vw
(f
v
f
w
)
2
.

For graphs without isolated vertices, there is another common denition of
a graph Laplacian. Since it is related to Markov chains and random walks (see
Section 8), it is also known as the transition Laplace matrix,
L
/
(G) = D
1/2
L(G)D
1/2
,
where D is the degree matrix. In particular, we have
L
/
(G)
vw
=
_
_
1
a
vv
deg(v)
, if v = w
a
vw
deg(v) deg(w)
, if vw E(G)
0, otherwise.
(3)
Observe that L
/
(G) is symmetric, and that for a d-regular loopless graph,
L(G) = d L
/
(G).
Notice also that the Laplace matrix L(G) does not change if we add loops to G,
while the transition Laplace matrix L
/
(G) is changed by the addition of loops.
Transition Laplace matrices are extensively treated in [8].
3. Laplace eigenvalues
By denition, L(G) is a real symmetric matrix, and equation (2) (together with the
non-negativity of edge-weights) implies that it is positive semi-denite. It therefore
has n non-negative real eigenvalues
i
=
i
(G), which in this chapter we take in
non-decreasing order
0
1

2

n
.
These eigenvalues are called the Laplace eigenvalues of the graph G. It is easy
to see that 0 is always an eigenvalue of L(G), and that j = (1, 1, . . . , 1)
T
is a
corresponding eigenvector.
Theorem 3.1 The multiplicity of 0 as an eigenvalue of L(G) is equal to the
number of connected components of G.
This result implies that
1
(G) = 0 is a simple eigenvalue of L(G) if and only if
the graph G is connected. It is also obvious, after writing L(G) in block diagonal
form, that the Laplace eigenvalues of G are the union of the eigenvalues of its
connected components.
Since the sum of all the eigenvalues equals the trace of the matrix, we have
n
i =1
i
(G) =
vV
deg(v) = 2[E(G)[, (4)
118 Bojan Mohar
where [E(G)[ should be replaced by the sum of all edge-weights in the case of
weighted graphs.
If M is a matrix with real eigenvalues, we use the notation
i
(M) to denote the
i th smallest eigenvalue of M (respecting the multiplicities). We use the symbol
max
(M) to denote the maximum eigenvalue of M, and sometimes write
min
(M)
instead of
1
(M).
Let G be a (weighted) k-regular graph. Then equation (1) implies that is an
eigenvalue of L(G) if and only if k is an eigenvalue of the weighted adjacency
matrix A(G). More precisely,
i
(L(G)) = k
ni 1
(A(G)) (i = 1, 2, . . . , n).
This result enables us to use known results about the eigenvalues of the adjacency
matrix of a regular graph in the study of its Laplace eigenvalues. For example, the
eigenvalues of the adjacency matrix of the complete graph K
n
are n 1 and 1
(the latter with multiplicity n 1), and therefore,
1
(K
n
) = 0, and
i
(K
n
) = n for
2 i n. Similarly, the Laplace eigenvalues of the n-cycle C
n
are precisely the
numbers
k
= 2 2 cos
_
2k
n
_
(k = 1, 2, . . . , n). (5)
If G is a simple unweighted graph and G is its complement, then
L(G) L(G) = nI J, (6)
where J is the all-1 matrix. The same relation holds in the weighted case if we
dene the edge-weights a
vw
of G to be 1 a
vw
(v, w V, v ,= w). In order to
have non-negative weights in G, we need 0 a
vw
1 for v, w V.
Let f
1
, f
2
, . . . , f
n
be an orthogonal system of eigenvectors of L(G) such that
f
1
= j and L(G)f
i
=
i
f
i
, for i = 1, 2, . . . , n. By (6), we get L(G)f
1
= 0 and
L(G)f
i
= (n
i
)f
i
, for 2 i n. This proves the following result.
Theorem 3.2 Suppose that the edge-weights of a graph G satisfy 0 a
vw
1,
and let G be the graph with edge-weights a
vw
= 1 a
vw
. Then
1
(G) = 0, and
i
(G) = n
ni 2
(G) for 2 i n.
For example, if G = K
r,s
, then G = K
r
K
s
. So, G has eigenvalues 0 (twice),
r (r 1 times), and s (s 1 times). Hence, the Laplace eigenvalues of K
r,s
are 0,
s (r 1 times), r (s 1 times), and r s.
Recall that the Cartesian product G H of graphs G and H has vertex-set
V(G H) = V(G) V(H), where (v
1
, v
2
) is adjacent to (w
1
,w
2
) if and only
if v
1
= w
1
and v
2
w
2
E(H), or v
2
= w
2
and v
1
w
1
E(G). There is a simple
description of the Laplace eigenvalues of G H in terms of the Laplace eigenval-
ues of G and H.
Theorem 3.3 The Laplace eigenvalues of the Cartesian product G H are pre-
cisely the numbers
i
(G)
j
(H),
for i = 1, 2, . . . , [V(G)[ and j = 1, 2, . . . , [V(H)[.
The proof of this result relies on the fact that the Laplace matrix is
L(G H) = L(G) I
V(H)
I
V(G)
L(H),
where denotes the Kronecker product of matrices (see Introduction).
Theorem 3.3 holds also for weighted graphs G and H, if the weight of an edge
of G H joining the vertices (v
1
, v
2
) and (w
1
, w
2
) is taken to be equal to the
weight of the edge v
1
w
1
of G if v
2
= w
2
, and equal to the weight of v
2
w
2
in H if
v
1
= w
1
. As a consequence, we obtain
2
(G H) = min{
2
(G),
2
(H)] and
max
(G H) =
max
(G)
max
(H). (7)
Theorem 3.3 can be used to determine the Laplace spectrum of several well-
known families of graphs. For example, the d-dimensional cube Q
d
is the Cartesian
product of d copies of K
2
. Since the Laplace eigenvalues of K
2
are 0 and 2, the
Laplace spectrum of Q
d
consists of the numbers 0, 2, 4, . . . , 2d. The multiplicity
of 2k in the spectrum of Q
d
is
_
d
k
_
.
Similarly, the Laplace eigenvalues of the path P
k
are (see [2])
(k)
i
= 4 sin
2
_
i
2k
_
(i = 0, 1, . . . , k 1).
Therefore, the grid graph P
m
P
n
has Laplace eigenvalues
i, j
=
(m)
i

(n)
j
= 4 sin
2
_
i
2m
_
4 sin
2
_
j
2n
_
.
It is natural to ask to what extent the Laplace eigenvalues determine the graph.
Several operations are known that change the graph but not its eigenvalues. This
gives rise to examples of families of cospectral graphs (see, for example, [34]).
120 Bojan Mohar
Bounding the Laplace eigenvalues
There are various useful min-max formulas for the eigenvalues of a symmetric
matrix. If M is a real symmetric matrix indexed by V, then
1
(M) = min
_
Mf, f)
f, f)
: 0 ,= f R
V
_
(8)
= min{Mf, f) : f R
V
, |f| = 1],
and similarly
max
(M) = max{Mf, f) : f R
V
, |f| = 1]. (9)
The Rayleigh characterization (8) has a generalization, the min-max characteri-
zation of the kth smallest eigenvalue
k
(M), known also as the Courant-Fisher
formula:
k
(M) = min
U
max
f
{Mf, f) : |f| = 1, f U], (10)
where the minimum is taken over all k-dimensional subspaces U of R
V
. Another
way of expressing (10) is
k
(M) = min{Mf, f) : |f| = 1, f f
i
, 1 i < k], (11)
where f
1
,f
2
, . . . , f
k1
are pairwise orthogonal eigenvectors of
1
,
2
, . . . ,
k1
,
respectively.
Among the Laplace eigenvalues of a connected graph G, the most important are
the extreme non-zero eigenvalues: the second smallest eigenvalue
2
(G) and the
largest eigenvalue
max
(G). Theorem 3.2 shows that
2
(G) = n
max
(G), and it
is thus not surprising that the importance of each of these eigenvalues implies that
of the other. For a (weighted) graph G with Laplace matrix L, (11) implies that
2
(G) = min{Lf, f) : |f| = 1, f j], (12)
since j is an eigenvector corresponding to
1
(G): notice that f is orthogonal to j if
and only if the sum of its coordinates is 0.
Expression (12) can be used to get combinatorial upper bounds on
2
(G). For
example:
Lemma 3.4 For any non-adjacent vertices s and t in G,
2
(G)
1
2
(deg s deg t ).
Proof Let f R
V
be dened by f
v
= 1 if v = s, f
v
= 1 if v = t , and f
v
= 0
otherwise. Since f j, (12) yields
2
(G)
L(G)f, f)
f, f)
=
vw E
a
vw
(f
v
f
w
)
2
vV
f
2
v
=
deg s deg t
2
.
The proof of Lemma 3.4 illustrates a basic technique for exploiting expressions
such as (12). By inserting an appropriate function into the expression, we can give
the right-hand side a meaningful combinatorial interpretation, and this enables
us to relate combinatorial properties to Laplace eigenvalues. To overcome the
orthogonality restriction on f, Fiedler [18] transformed (12) to a more suitable
expression that does not require f to be orthogonal to j.
Theorem 3.5 For any graph G of order n,
2
(G) = 2n min
_
vwE
a
vw
(f
v
f
w
)
2
vV
wV
(f
v
f
w
)
2
: f ,= cj, for c R
_
(13)
and
max
(G) = 2n max
_
vwE
a
vw
(f
v
f
w
)
2
vV
wV
(f
v
f
w
)
2
: f ,= cj, for c R
_
. (14)
The eigenvalues
2
(G) and
max
(G) can be bounded in terms of the maximum
and minimum degrees of G.
Theorem 3.6 For any graph G of order n,
2
(G)
n
n 1
(G) and
n
n 1
(G)
max
(G) 2(G).
If G is an unweighted graph, the last inequality of Theorem3.6 can be strength-
ened to
max
(G) max{deg v deg w : vw E]. If G is connected, then the
equality holds if and only if G is bipartite semi-regular (see [2]). Theorem 3.2
implies that
max
(G) n, with equality if and only if the complement of G is
disconnected.
Let G be a (weighted) graph, and let G
/
= G e be the graph obtained from G
by adding an edge e to G (possibly increasing the weight of an existing edge). Then
L(G
/
) and L(G) differ by a positive semi-denite matrix of rank 1. It follows by
the Courant-Weyl inequalities (see, for example, [11, Thm. 2.1]) that the following
is true.
122 Bojan Mohar
Theorem 3.7 The eigenvalues of G and G
/
= G e interlace:
0 =
1
(G) =
1
(G
/
)
2
(G)
2
(G
/
)
3
(G)
n
(G)
n
(G
/
).
Observe that

n
i =1
(
i
(G
/
)
i
(G)) = 2, by (4), so that at least one inequality
i
(G)
i
(G
/
) is strict. By inserting more than one edge we may lose the inter-
lacing of the eigenvalues.
Theorem 3.8 Suppose that G
1
and G
2
are graphs with the same vertex-set. Let
G = G
1
G
2
be the graph whose edge-weights are the sums of edge-weights of
G
1
and G
2
. Then
r

2
(G)
2
(G
1
)
2
(G
2
);
r
max{
max
(G
1
),
max
(G
2
)]
max
(G)
max
(G
1
)
max
(G
2
).
Eigenvalues of the transition Laplacian
The eigenvalues
/
i
=
i
(L
/
(G)) of the transition Laplace matrix behave similarly
to the eigenvalues of L(G) : they are all real and non-negative,
/
1
= 0, and
/
2
> 0
if and only if G is connected.
Theorem 3.9 Let G be a graph of order n with p isolated vertices. Then
r

n
i =1
/
i
(G) = n p;
r
if G is not the complete graph, then
/
2
(G) 1;
r

/
n
(G) 2, and equality holds if and only if G has a non-trivial bipartite
component;
r
if G is bipartite, then
/
i
(G) = 2
/
ni 1
(G) for i = 1, 2, . . . , n.
We refer to [8] for more details and further results.
4. Eigenvalues and vertex partitions of graphs
We begin this section with some notation. For each subset S V(G), let

S =
V(G) S denote the complement of S in V(G). Given sets of vertices A, B
V(G), let E(A, B) be the set of edges of G with one end in A and the other in B.
We also let
e(A, B) =
vA
wB
a
vw
denote the sumof the weights of the edges in E(A, B), with all edges in G[A B]
counted twice. Note that, for an unweighted graph, e(S,

S) counts the number of
edges in the cut E(S,

S).
The partition problems discussed here are mainly concerned with nding an
appropriate subset S V(G) for whichthe edge-cut E(S,

S) satises some specic
extremal property. In particular, we focus our attention on the weight of a maximum
cut, the bipartition width and the isoperimetric number. All of these problems are
NP-hard, so that non-trivial bounds are desirable, and potentially very important.
The following lemma relates the weight of an edge-cut to the eigenvalues
2
(G)
and
max
(G).
Lemma 4.1 Let G be a (weighted) graph of order n, and let S V(G) with
[S[ = s. Then
2
(G)
s(n s)
n
e(S,

S)
max
(G)
s(n s)
n
.
Proof Let f R
V
be the characteristic function of S : f
v
= 1 if v S, and f
v
= 0
otherwise. Then
vV
wV
(f
v
f
w
)
2
= 2s(n s)
and
vwE
a
vw
(f
v
f
w
)
2
= e(S,

S).
If S ,= and S ,= V(G), then (13) implies that
2
(G) 2n
e(S,

S)
2s(n s)
.
This gives the lower bound, which obviously holds also for S = and S = V(G).
The upper bound is proved analogously, using (14).
It is an immediate, and also an important, consequence of Lemma 4.1 that
in a graph for which all non-trivial Laplace eigenvalues are close together (that
is,
max

2
is small), the weights of all the edge-cuts E(S,

S) corresponding to
vertex-sets S of the same cardinality are approximately the same. In particular, this
property holds in randomgraphs. It is therefore not surprising that many algorithms
dealing with edge-cuts perform well on randomly chosen graphs.
Lemma 4.1 also shows that graphs with large
2
have fast growth rate. This
implies, in particular, that the diameter can be bounded as a function of
2
(G) (see,
for example, [8] and [40] and the references therein).
124 Bojan Mohar
The bipartition width
Loosely speaking, the bipartition width of a graph is the minimum number of
edges in a cut that partitions the vertices into two sets of nearly equal size. More
formally, for a graph G of order n, the bipartition width bw(G) is dened as
bw(G) = min
_
e(S,

S) : S V(G), [S[ =
_
1
2
n
__
.
It is known that, even for unweighted simple graphs, the problem of determining
bw(G) is NP-hard (see, for example, [21, p. 210]).
Since all the sets over which the minimum is taken have size
1
2
n, Lemma 4.1
gives the following lower bound for bw(G).
Corollary 4.2 Let G be a weighted graph of order n. If n is even, then
bw(G)
1
4
n
2
(G).
If n is odd, then
bw(G)
n
2
1
4n

2
(G).
We can further improve the bounds of Corollary 4.2 by introducing a correction
function. A function c R
V
is called a correction function if c j. The following
bound was proved by Boppana [5].
Theorem 4.3 Let G be a weighted graph of even order n. Then
bw(G)
1
4
n max
c
min
f
(L(G) diag(c))f, f)
f, f)
,
where the maximumis taken over all correction functions c R
V
and the minimum
is taken over all non-zero functions f R
V
with f j.
Proof Let S V(G) be a set of cardinality
1
2
n with e(S,

S) = bw(G), and let g
R
V
be its signed characteristic function, dened by g
v
= 1 if v S and g
v
= 1
if v

S. Since [S[ = [
S[, we have g j.
Take an arbitrary correction function c R
V
. Since c j, we have
diag(c)g, g) =
vV
c
v
g
2
v
=
vV
c
v
= 0. (15)
Using (15), and applying (2), we deduce that
(L(G) diag(c))g, g)
g, g)
=
L(G)g, g)
g, g)
=
vwE
a
vw
(g
v
g
w
)
2
vV
g
2
v
=
4e(S,

S)
n
=
4
n
bw(G).
Since c is arbitrary, the bound follows.
For computational purposes, it is convenient toexpress the boundof this theorem
as a maximization of the smallest eigenvalue of an appropriate symmetric matrix.
This can be done as follows. Let Q = (q
1
, q
2
, . . . , q
n1
) be an n (n 1) matrix
whose columns q
i
are pairwise orthogonal unit vectors for which q
i
j (1 i <
n). It is easy to see that, for each x R
n1
, we have Qx, Qx) = x, x) and Qx j.
This implies the following result.
Theorem 4.4 We have
bw(G)
1
4
n max
c

min
(Q
T
(L(G) diag(c))Q),
where the maximum is taken over all correction functions c R
V
.
The bound of Theorem 4.4 can be formulated as a semi-denite program, and can
therefore be computed to an arbitrary precision in polynomial time, using known
polynomial-time methods for solving such programs. For more details, see the
discussion at the end of the next section.
5. The max-cut problem and semi-denite programming
The maximum cut (or max-cut) problem is similar to the maximum version of the
bipartition width problem, except that the restrictions on the sizes of the subsets
over which the maximum is taken are omitted. More precisely, let mc(G) denote
the maximum weight of an edge-cut in G:
mc(G) = max{e(S,

S) : ,= S V(G)].
The problem of determining mc(G) is NP-hard (see [21, p. 210]). Moreover, it
is known that there exists a constant > 0 for which there is no polynomial-time
(1 )-approximation algorithm for the max-cut problem, unless P = NP (see
[3]). On the other hand, it is possible to nd a 0.878-approximation to mc(G) in
polynomial time [22].
Lemma 4.1 implies the following upper bound on mc(G), which was rst ob-
served by Mohar and Poljak (see [40]):
mc(G)
1
4
n
max
(G). (16)
Just as for the bipartition width problem, (16) can be further improved using cor-
rection functions. The corresponding optimized eigenvalue bound was introduced
by Delorme and Poljak [13].
Theorem 5.1 Let G be a weighted graph of order n. Then
mc(G)
1
4
n min
c

max
(L(G) diag(c)), (17)
where the minimum is taken over all correction functions c R
V
.
The proof consists of steps similar to those in the proof of Theorem 4.3.
126 Bojan Mohar
Minimizing the largest eigenvalue of an afne combination of symmetric ma-
trices can be formulated as a semi-denite program (see, for example, [47]). Let
A
0
, A
1
, . . . , A
n
R
mm
be symmetric matrices, and for x R
n
, set
A(x) = A
0

n
i =0
x
i
A
i
.
The problem is to
_
minimize
max
(A(x))
subject to x U,
(LM1)
where U is a linear subspace of R
n
. We can translate this problem into a semi-
denite program by introducing an auxiliary variable t R:
_
minimize t
subject to t I A(x) _ 0, x U,
(LM2)
where the notation X _ 0 means that X is a positive semi-denite matrix. The rst
constraint in (LM2) is equivalent to t
max
(A(x)).
In particular, the bound on the max-cut in Theorem 5.1 is of the form (LM1),
where we take U to be the orthogonal complement of the vector j. Also, the bound
on the bipartition width in Theorem 4.4 can be viewed as a semi-denite program
of the same form, since
max
xU
min
(A(x)) = max
xU
{
max
(A(x))] = min
xU
max
(A(x)).
There exist efcient and practical algorithms for solving semi-denite pro-
grams. Given any > 0, the given semi-denite program can be solved within an
additive error of in polynomial time, where is part of the input, its size being
proportional to log(1/). For this purpose, one can use interior-point methods (see
[41]).
Given a (weighted) graph G, the max-cut problem for G can be formulated as
a quadratic integer program with variables y
v
, v V(G):
_
_
_
maximize
1
2
vwE(G)
a
vw
(1 y
v
y
w
)
subject to y
v
= 1 or 1, for every v V(G).
(MC1)
If (y
v
)
vV
is a feasible solution, then 1 y
v
y
w
is equal to either 0 or 2. Given a
solution of (MC1), the set S with the property that mc(G) = e(S,

S) is determined
by S = {v V(G) : y
v
= 1].
Goemans and Williamson [22] considered the following semi-denite relax-
ation of (MC1), where z
vw
corresponds to the product y
v
y
w
in (MC1):
_
_
maximize
1
2
vwE
a
vw
(1 z
vw
)
subject to z
vv
= 1, for every v V,
Z = [z
vw
]
v,wV
_ 0.
(MC2)
Poljak and Rendl [42] were the rst to realize that the bound of the relaxation
(MC2) is actually equivalent to the eigenvalue upper bound of Theorem 5.1. For
a proof, which uses elegant duality theory of semi-denite programs, we refer
to [42].
Theorem 5.2 Let G be a weighted graph of order n. Then the value of an optimal
solution to (MC2) is equal to the right-hand side of (17).
Goemans andWilliamsonprovedthat (MC2) canbe usedtoobtaina polynomial-
time 0.878-approximation algorithmfor the max-cut problem. Since the announce-
ment of this algorithm, semi-denite programming has been successfully applied
to the design of a number of other approximation algorithms, such as the max-k-cut
(Frieze and Jerrum[20]), a 0.931-approximation algorithmfor MAX2-SAT(Feige
and Goemans [16]), a 0.859-approximation algorithm for MAX DISCUT [16],
and approximate colouring (Karger, Motwani and Sudan [30]). An application to
the travelling salesman problem is described below in Section 7.
6. Isoperimetric inequalities
Isoperimetric problems are related to questions in which one considers the ratio
between the surface area and the volume of some d-dimensional body. In graph
theory, the natural analogue to the volume is the number of vertices or the sum
of the degrees in a set S of vertices of the graph, while the counterpart of the
surface area is the number e(S,

S) of edges with one end in S and the other end
outside S. Problems in which one considers ratios of the form e(S,

S)/[S[ are
called isoperimetric problems for graphs; thus, isoperimetric properties concern
the sizes of the neighborhood of a set of vertices. The related term expansion
usually means that the sizes of the neighborhood can be bounded from below by
some function of the size of the subset. Such isoperimetric properties provide the
foundation for many recent applications of graph theory to theoretical computer
science, as mentioned in the introduction.
128 Bojan Mohar
The isoperimetric number i (G) of a graph G of order n ( 2) is dened as
i (G) = min
_
e(S,

S)
[S[
: S V(G), 0 < [S[
1
2
n
_
.
Computationally, the problem of determining i (G) is NP-hard (see [36]).
A straightforward application of Lemma 4.1 yields the following eigenvalue
lower bound on i (G).
Corollary 6.1 Let G be a weighted graph of order n. Then i (G)
1
2
2
(G).
It is difcult to obtain useful lower bounds on i (G) by combinatorial means.
Corollary 6.1, although easy to prove, gives a non-trivial lower bound on i (G).
For an example where this bound is tight, consider the d-dimensional cube Q
d
as
the product K
2
Q
d1
, and let S be the set of vertices in one copy of Q
d1
. Since
[S[ = e(S,

S) = 2
d1
, we have i (Q
d
) 1. On the other hand, as noted in Section
3,
2
(Q
d
) = 2. Using Corollary 6.1, we conclude that i (Q
d
) = 1.
The following quantity is sometimes easier to deal with than i (G):
i
(G) = min
_
e(S,

S)
[S[[
S[

[V[
2
: ,= S V
_
.
Clearly,
1
2
i (G) i
(G) i (G). It turns out that i
(G) satises a relation similar

to (7):
i
(G H) = min{i
(G), i
(H)].
The above inequality implies that
1
2
min{i (G), i (H)] i (G H) min{i (G), i (H)].
While the inequality on the right can be strict (see [36]), it is not known whether
the factor
1
2
on the left is best possible. We refer to [8, Sec. 2.6] and [26] for further
results.
It is important to note that i (G) also has upper bounds in terms of
2
(G).
Such bounds are known as Cheeger inequalities, since they are discrete analogues
of their continuous counterpart (see [6] and [7]) arising in the study of Laplace
operators on Riemannian manifolds.
The classical form of Cheegers bound adapted to graphs is
i (G)
_
2(G)
2
(G). (18)
An improvement of (18) was obtained by Mohar [36], and for weighted graphs
in [39]:
Theorem 6.2 Let G be a weighted graph of order n ( 3) with maximum degree
and second eigenvalue
2
. Then
i (G)
_
(2
2
)
2
.
Other discrete forms of Cheegers inequality have been proved; see Theorem 8.4,
and [8] and [9] and the references therein.
An important observation, following from the proof of Theorem 6.2, is that
the partition based on the eigenfunction f of
2
is not too far from optimal. Con-
structing partitions based on the eigenfunctions of
2
has proved to be one of the
most successful heuristics in parallel computation and for a divide-and-conquer
approach for processor distribution.
7. The travelling salesman problem
The travelling salesman problem (TSP) is one of the best-known NP-hard com-
binatorial optimization problems, and there is an extensive literature on both its
theoretical and practical aspects. In the symmetric travelling salesman problem
(STSP), it is assumed that the cost of travelling between two points is the same in
both directions.
We shall mention here only one approach, which uses semi-denite program-
ming (see Section 5) to establish a lower bound on the length of an optimal tour.
This bound is obtained by relaxing the STSP and invoking a branch-and-bound
algorithm. Semi-denite relaxations of the STSP were developed by Cvetkovi c,
Cangalovi c and Kova cevi c-Vuj ci c [10], and are based on the Laplace eigenvalues
of graphs.
In what follows we assume that G is the complete graph of order n and that
each edge vw has a cost c
vw
, such that the matrix C = (c
vw
) is symmetric, with
c
vv
= 0. The STSP can now be formulated as follows: nd a Hamiltonian cycle in
G of minimum cost.
The next theorem gives a basis for a discrete semi-denite programming model
of the STSP [10]. Recall from (5) that
2
(C
n
) = 2 2 cos(2/n).
Theorem 7.1 Let H be a 2-regular spanning subgraph of G, and let X = L(H)
J I, where and are real parameters such that >
2
(C
n
)/n and 0 <
2
(C
n
). Then H is a Hamiltonian cycle if and only if the matrix X is positive semi-
denite.
Proof Let 0 =
1

2

n
be the eigenvalues of L(H), and let
f
1
, f
2
, . . . , f
n
be corresponding eigenvectors that are pairwise orthogonal, with
130 Bojan Mohar
f
1
= j. Then f
1
, f
2
, . . . ,f
n
are linearly independent eigenvectors of X, with corre-
sponding eigenvalues n ,
2
, . . . ,
n
. Since H is a union of disjoint
cycles, either H is a Hamiltoniancycle with
2
=
2
(C
n
), or H is disconnectedwith
2
= 0. In either case,
2
is the smallest eigenvalue, since n >
2
(C
n
)
2
.
If H is a Hamiltionian cycle, then
2
=
2
(C
n
) 0 and X is positive
semi-denite. Conversely, if X is positive semi-denite, then
2
(H) =
2
>
0, and so H is connected, and hence a Hamiltonian cycle.
Aresult that is similar to an extension of Theorem7.1 for higher eigenvalues was
used by Mohar [38] to derive a sufcient algebraic condition for non-Hamiltonicity.
Based on that criterion, it can be shown, just by looking at its eigenvalues, that the
Petersen graph is not Hamiltonian.
It follows from Theorem 7.1 that a spanning subgraph H of G is a Hamiltonian
cycle if and only if its Laplacian L(H) = (l
vw
) satises the following conditions:
each l
vv
= 2, X = L(H) J I _ 0, for >
2
(C
n
)/n
and 0 <
2
(C
n
).
This result gives rise to a semi-denite relaxation of the STSP which has the
following equivalent formulation in terms of Laplacians, where L = (l
vw
) is a
symmetric matrix:
_
_
minimize (L) =

vV
wV
_
1
2
c
vw
_
l
vw
,
subject to l
vv
= 2,
wV
l
vw
= 0, 1 l
vw
0 (for v ,= w),
and
2
(L) .
This semi-denite program can be solved in polynomial time and yields a good
approximation to the optimum solution for the STSP.
8. Random walks on graphs
Isoperimetric properties and the eigenvalues treated in previous sections are closely
related to the convergence rates of Markov chains. Several important randomized
algorithms discovered in the last decade have increased the applicability of random
walks and Markov chains to previously intractable problems. Additional reading
on the results related to the presentation in this section can be found in [8] and [44].
Given a weighted graph G (possibly with loops), a simple random walk on
G is a Markov chain with state space V and with transition probabilities p
vw
of
stepping from v to w, dened as follows:
p
vw
=
a
vw
deg v
.
The transition matrix P(G) is
P(G) = D
1
A(G),
where, as before, D is the degree matrix. Although P is not necessarily symmetric,
it has only real eigenvalues. This follows from the following result.
Theorem 8.1 P(G) = I D
1/2
L
/
(G)D
1/2
= I D
1
L(G).
Proof Since I D
1/2
L
/
(G)D
1/2
= I D
1
L(G), the diagonal elements of the
matrix are
1 deg v
1
L
vv
= 1 (deg v a
vv
)/deg v = a
vv
/deg v (v V),
and the off-diagonal entries are
(deg v
1
(L(G))
vw
= a
vw
/deg v = p
vw
(v, w V, v ,= w).
In particular, I P(G) and L
/
(G) have the same eigenvalues, so
i
(P(G)) = 1
/
ni 1
(G).
Theorem 8.1 also implies that P(G) (and thus also P(G)
T
) can be diagonalized,
and that it has n pairwise orthogonal eigenvectors.
Rate of convergence of a random walk
For a random walk on G, let x
(0)
: V R
be the initial probability distribution

on V(G). Then
x
(t )
= P
T
x
(t 1)
= (P
T
)
t
x
(0)
are the probabilities of where the Markov chain will be after t steps. If G is
connected and non-bipartite, then lim
t
x
(t )
exists; it is denoted by x
()
, and is
called the stationary distribution.
There is a close relationship between the eigenvalues of P(G) and the rate of
convergence of a simple random walk on G to the stationary distribution.
Theorem 8.2 Let G be a connected non-bipartite weighted graph, and let x
(0)
be
an initial probability distribution on V(G). If
= max{[
i
(P(G))[ :
i
(P(G)) ,= 1],
then, for a simple random walk on G,
|x
(t )
x
()
| <
t
.
132 Bojan Mohar
Proof Let Q = P(G)
T
, and for i = 1, 2, . . . , n, let y
(i )
be the eigenfunction of
Q corresponding to the eigenvalue
i
=
i
(Q) =
i
(P(G)). We may assume that
|y
(i )
| = 1 and that the vectors y
(i )
are pairwise othogonal. Recall that
n
(Q) = 1,
and note that the assumptions in the theorem imply that < 1. Moreover, the
Perron-Frobenius theorem shows that all the components of y
(n)
are positive.
Let us write
x
(0)
=
n
i =1
i
y
(i )
,
and observe that
n
= x
(0)
, y
(n)
) ,= 0. Then
x
(t )
= Q
t
x
(0)
=
n
i =1
t
i
y
(i )
=
n1
i =1
t
i
y
(i )
n
y
(n)
.
Since < 1, x
(t )
converges to a
n
y
(n)
= x
()
, and so
|x
(t )
x
()
|
2
=
_
_
_
_
n1
i =1
t
i
y
(i )
_
_
_
_
2
=
n1
i =1
|
i
t
i
y
(i )
|
2
=
n1
i =1
2
i

2t
i
|y
(i )
|
2

2t
n1
i =1
2
i
.
Since also
n1
i =1
2
i
<
n
i =1
2
i
= |x
(0)
|
2
1,
we nally get |x
(t )
x
()
| <
t
.
As an example, let us consider a random walk on the d-dimensional hypercube
Q
d
. (Note that Q
d
is a Cayley graph of the group Z
d
2
; Cayley graphs are discussed
in Chapter 6.) Since Q
d
is bipartite, we add a loop to each vertex, and denote the
resulting graph by G. We would like to assign weights to the loops and edges of
G in such a way that becomes as small as possible. Let be the weight assigned
to each loop, and be the weight assigned to each edge of G. We also require that
d = 1. Then
P(G) = I A(Q
d
).
From the Laplace eigenvalues of Q
d
(see Section 3), we see that the eigenvalues
of P(G) are
i
= (d 2i ), i = 0, 1, . . . , d, with multiplicity
_
d
i
_
. Thus,
= max{ (d 1), [ d[]. By balancing both terms, we get equality
when = = (d 1)
1
. Hence,
=
d 1
d 1
= 1
2
d 1
.
By Theorem 8.2, the rate of convergence can be estimated by
t
=
_
d 1
d 1
_
t
=
_
1
2
d 1
_
t
exp
_
2t
d 1
_
.
Similar estimates can be established for other Abelian groups (see, for example,
[44]).
Perhaps the most useful distance between two probability distributions x and y
of a Markov chain is the so-called chi-squared distance, dened as follows:
2
(x, y) =
wV(G)
(x
w
y
w
)
2
y
w
.
Clearly, this denition makes sense only when y
v
> 0, for every v V. A proof
similar to the proof of Theorem 8.2 gives the following result.
Theorem 8.3 For a simple random walk on a non-bipartite weighted graph,
2
(x
(t )
, x
()
)
t

2
(x
(0)
, x
()
).
The quantity (1 )
1
is called the mixing time of the random walk, and is of
interest since the above theorem implies that, in this number of steps, the chi-
squareddistance fromthe stationarydistributionx
()
is cut bya constant factor. The
quantity 1 is related to the second smallest eigenvalue
2
(I P) =
2
(L
/
(G)).
The isoperimetric constant related to the transition Laplacian is also known as
the conductance of the corresponding Markov chain. For a simple random walk
on a graph G, the conductance is dened as
i
P
(G) = min
_
vS
w/ S
p
vw
deg v
vol(S)
: S V(G), 0 < vol(S) m
_
,
where
vol S =
vS
deg v and m =
1
2
vV
deg v =
1
2
vol V.
If G is regular, then the conductance is just the isoperimetric number of G divided
by the degree.
Theorem 8.4 For the simple random walk on a graph G,
1
2
i
2
P
(G)
2
(I P(G)) =
2
(L
/
(G)) 2i
P
(G).
For proofs we refer to [35]; see also [8].
134 Bojan Mohar
Since
2
(I P(G)) 1 , Theorem 8.4 also gives a bound on the mixing
time.
Rapidly mixing Markov chains proved to be an important tool in the design of
polynomial-time randomized algorithms. For example, Dyer, Frieze and Kannan
[15] found a polynomial-time randomized approximation algorithmfor computing
the volume of a convex set K, based on sampling random points in K. Lov asz
and Simonovits [32] improved this algorithm by providing random walk sampling
with better mixing, and by improved isoperimetric estimates.
Another application, somewhat similar to the rst, deals with approximate
counting of perfect matchings in a bipartite graph, and therefore also with the
approximation of the permanent of a (0, 1)-matrix (see Sinclair and Jerrum [46]).
References
1. N. Alon and V. D. Milman,
1
, isoperimetric inequalities for graphs and super-
concentrators, J. Combin. Theory (B) 38 (1985), 7388.
2. W. N. Anderson and T. D. Morley, Eigenvalues of the Laplacian of a graph, Linear and
Multilinear Algebra 18 (1985), 141145.
3. S. Aurora, C. Lund, R. Motwani, M. Sudan and M. Szegedy, Proof verication and
hardness of approximation problems, Proc. 33rd FOCS (1992), 1423.
4. N. L. Biggs, Algebraic Graph Theory (2nd ed.), Cambridge Univ. Press, 1993.
5. R. B. Boppana, Eigenvalues and graph bisection: an average case analysis, 28th Annual
Symp. Found. Comp. Sci., IEEE (1987), 280285.
6. R. Brooks, Spectral geometry and the Cheeger constant, Expanding Graphs (ed. J.
Friedman), Amer. Math. Soc. (1993), 519.
7. P. Buser, Geometry and Spectra of Compact Riemann Surfaces, Birkh auser, 1992.
8. F. R. K. Chung, Spectral Graph Theory, Amer. Math. Soc., 1997.
9. Y. Colin de Verdi` ere, Spectres de Graphes, Cours Sp ecialis es 4, Soc. Math. France,
1998.
10. D. Cvetkovi c, M.

Cangalovi c and V. Kova cevi c-Vuj ci c, Semidenite programming
methods for the symmetric traveling salesman problem, Integer Programming and
Combinatorial Optimization (eds. G. Cornuejols, R. E. Burkard and G. J. Woeginger),
Springer (1999), 126136.
11. D. M. Cvetkovi c, M. DoobandH. Sachs, Spectraof Graphs (3rded.), JohannAmbrosius
Barth Verlag, 1995.
12. D. M. Cvetkovi c, M. Doob, I. Gutman and A. Torga sev, Recent Results in the Theory
of Graph Spectra, Ann. Discr. Math. 36, North-Holland, 1988.
13. C. Delorme and S. Poljak, Laplacian eigenvalues and the maximumcut problem, Math.
Programming 62 (1993), 557574.
14. W. E. Donath and A. J. Hoffman, Lower bounds for the partitioning of graphs, IBM J.
Res. Develop. 17 (1973), 420425.
15. M. E. Dyer, A. Frieze and R. Kannan, A random polynomial-time algorithm for ap-
proximating the volume of convex bodies, J. Assoc. Comput. Mach. 38 (1991), 117.
16. U. Feige and M. X. Goemans, Approximating the value of two prover proof systems,
with applications to MAX2SAT and MAXDICUT, Proc. Third Israel Symp. on Theory
of Computing and Systems (1995), 182189.
17. M. Fiedler, Algebraic connectivity of graphs. Czech. Math. J. 23 (98) (1973), 298305.
18. M. Fiedler, A property of eigenvectors of nonnegative symmetric matrices and its
application to graph theory, Czech. Math. J. 25 (100) (1975), 619633.
19. P. W. Fowler, T. Pisanski and J. Shawe-Taylor, Molecular graph eigenvectors for molec-
ular coordinates, Graph drawing: GD94 (ed. R. Tamassia), Springer-Verlag (1995),
282285.
20. A. Frieze and M. Jerrum, Improved approximation algorithms for MAX k-CUT and MAX
BISECTION, Integer Programming and Combinatorial Optimization; Proc. 4th IPCO
(eds. E. Balas and J. Clausen), Springer-Verlag (1995), 113.
21. M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory
of NP-Completeness, Freeman, 1979.
22. M. X. Goemans and D. P. Williamson, Improved approximation algorithms for maxi-
mum cut and satisability problems using semidenite programming, J. Assoc. Comp.
Mach. 42 (1995), 11151145.
23. C. D. Godsil and G. Royle, Algebraic Graph Theory, Springer-Verlag, 2001.
24. L. Hagen and A. B. Kahng, New spectral methods for ratio cut partitioning and clus-
tering, IEEE Trans. Computer-Aided Design 11 (1992), 10741085.
25. C. Helmberg, B. Mohar, S. Poljak and F. Rendl, A spectral approach to bandwidth and
separator problems in graphs, Linear and Multilinear Algebra 39 (1995), 7390.
26. C. Houdr e and P. Tetali, Concentration of measure for products with Markov kernels
and graphs via functional inequalities, Combin. Probab. Comput. 10 (2001), 128.
27. B. Hendrickson and R. Leland, An improved spectral graph partitioning algorithm for
mapping parallel computations, SIAM J. Sci. Comput. 16 (1995), 452469.
28. A. J. Hoffman, On eigenvalues and colorings of graphs, Graph Theory and its Appli-
cations (ed. B. Harris), Academic Press (1970), 7991.
29. M. Juvan and B. Mohar, Laplace eigenvalues and bandwidth-type invariants of graphs,
J. Graph Theory 17 (1993), 393407.
30. D. Karger, R. Motwani and M. Sudan, Approximate graph coloring by semidenite
programming, Proc. 35th Ann. Symp. FOCS (1994), 213.
31. L. Lov asz, On the Shannon capacity of a graph, IEEE Trans. Inform. Theory IT25
(1979), 17.
32. L. Lov asz and M. Simonovits, Randomwalks in a convex body and an improved volume
algorithm, Random Struct. Algorithms 4 (1993), 359412.
33. R. Merris, Laplacian matrices of graphs: a survey, Linear Alg. Appl. 197/198 (1994),
143176.
34. R. Merris, Large families of Laplacian isospectral graphs, Linear and Multilinear Al-
gebra 43 (1997), 201205.
35. B. Mohar, Isoperimetric inequalities, growth, and the spectrum of graphs, Linear Alg.
Appl. 103 (1988), 119131.
36. B. Mohar, Isoperimetric numbers of graphs, J. Combin. Theory (B) 47 (1989), 274291.
37. B. Mohar, The Laplacian spectrum of graphs, Graph Theory, Combinatorics, and Ap-
plications (eds. Y. Alavi et al.), Wiley (1991), 871898.
38. B. Mohar, A domain monotonicity theorem for graphs and Hamiltonicity, Discr. Appl.
Math. 36 (1992), 169177.
136 Bojan Mohar
39. B. Mohar, Some applications of Laplace eigenvalues of graphs, Graph Symmetry:
Algebraic Methods and Applications (eds. G. Hahn and G. Sabidussi), NATO ASI Ser.
C 497, Kluwer (1997), 225275.
40. B. Mohar and S. Poljak, Eigenvalues in combinatorial optimization, Combinatorial and
Graph-Theoretical Problems in Linear Algebra (eds. R. A. Brualdi, S. Friedland and V.
Klee), IMA Volumes in Mathematics and its Applications 50, Springer-Verlag (1993),
107151.
41. Y. Nesterov and A. Nemirovskii, Interior Point Polynomial Methods in Convex Pro-
gramming, SIAM, 1994.
42. S. Poljak and F. Rendl, Nonpolyhedral relaxations of graph-bisection problems, SIAM
J. Optim. 5 (1995), 467487.
43. F. Rendl and H. Wolkowicz, Applications of parametric programming and eigenvalue
maximization to the quadratic assignment problem, Math. Progr. 53 (1992), 6378.
44. J. S. Rosenthal, Convergence rates for Markov chains, SIAM Review 37 (1995), 387
405.
45. H. D. Simon, Partitioning of unstructured problems for parallel processing, Computing
Systems in Engineering 2 (1991), 135148.
46. A. J. Sinclair and M. R. Jerrum, Approximate counting, uniformgeneration and rapidly
mixing Markov chains, Inform. Comput. 82 (1989), 93133.
47. L. Vandenberghe and S. Boyd, Semidenite programming, SIAM Review 38 (1996),
4995.
48. H. S. Wilf, The eigenvalues of a graph and its chromatic number, J. London Math. Soc.
42 (1967), 330332.
5
Automorphisms of graphs
PETER J. CAMERON
1. Graph automorphisms
2. Algorithmic aspects
3. Automorphisms of typical graphs
4. Permutation groups
5. Abstract groups
6. Cayley graphs
7. Vertex-transitive graphs
8. Higher symmetry
9. Innite graphs
10. Graph homomorphisms
References
This chapter surveys automorphisms of nite graphs, concentrating on the
asymmetry of typical graphs, prescribing automorphism groups (as either
permutation groups or abstract groups), and special properties of vertex-
transitive graphs and related classes. There are short digressions on innite
graphs and graph homomorphisms.
1. Graph automorphisms
An automorphism of a graph G is a permutation g of the vertex-set of G with the
property that, for any vertices u and v, we have vg wg if and only if v w.
(As usual, vg denotes the image of the vertex v under the permutation g: see the
Introduction for the terminology and main results of permutation group theory.)
This simple denition does not sufce for multigraphs, where we need to specify
a permutation of the edges as well as a permutation of the vertices, to ensure that the
multiplicity of edges between two vertices is preserved. Alternatively, a multigraph
137
138 Peter J. Cameron
can be regarded as a weighted graph, where the weight a
v,w
is the number of edges
from v to w; an automorphism must satisfy a
vg,wg
= a
v,w
: this gives a slightly
different description of automorphisms, but the action on the set of vertices is the
same. We consider only simple graphs here.
The set of all automorphisms of a graph G, with the operation of composition
of permutations, is a permutation group on V(G), a subgroup of the symmetric
group on V(G). This is the automorphism group of G, denoted by Aut(G). We
describe any subgroup H of Aut(G) as a group of automorphisms of G, and refer
to Aut(G) as the full automorphism group.
More generally, an isomorphism from a graph G to a graph H is a bijection
f from the vertex-set of G to that of H such that vf wf (in H) if and only if
v w (in G). We say that G and H are isomorphic (written G

= H) if there is an
isomorphism between them.
Among its other jobs, the automorphism group arises in the enumeration of
graphs, specicallyinthe relationbetweencountinglabelledandunlabelledgraphs.
A labelled graph on n vertices is a graph whose vertex-set is {1, 2, . . . , n], while
an unlabelled graph is simply an isomorphism class of n-element graphs.
Now the number of labellings of a given unlabelled graph G on n vertices is
n!/[Aut(G)[. For, a labelling is given by a bijective function F from {1, 2, . . . , n]
to V(G); there are n! such functions, and two of them F
1
and F
2
dene the same
labelled graph if and only if there is an automorphism g such that F
2
(i ) = F
1
(i )g,
for all i {1, 2, . . . , n]. Figure 1 shows the three labellings of the path P
3
, a graph
whose automorphism group has order 2.
The automorphism group is an algebraic invariant of a graph. Before giving
some simple properties, we recall some terminology from the Introduction.
The direct product G
1
G
2
of two permutation groups G
1
and G
2
(acting on sets
1
and
2
) is the permutation group on the disjoint union
1

2
whose elements
are ordered pairs (g
1
, g
2
) for g
i
G
i
; the action is given by
v(g
1
, g
2
) =
vg
1
if v
1
,
vg
2
if v
2
.
This notion extends to the direct product of any number of permutation groups.
If G
2
is a permutation group on {1, 2, . . . , n], then the wreath product G
1
: G
2
is
generated by the direct product of n copies of G
1
, together with the elements of G
2
acting on these n copies of G
1
.
1 2 3 1 1 3 3 2 2
Fig. 1.
Finally, S
n
denotes the symmetric group on {1, 2, . . . , n].
Theorem 1.1
(a) A graph and its complement have the same automorphism group.
(b) Aut(K
n
) = S
n
.
(c) Let the connected components of G consist of n
1
copies of G
1
, n
2
copies
of G
2
, . . . , n
r
copies of G
r
, where G
1
, G
2
, . . . , G
r
are pairwise
non-isomorphic. Then
Aut(G) = (Aut(G
1
) : S
n
1
) (Aut(G
2
) : S
n
2
) (Aut(G
r
) : S
n
r
).
In view of these results, we can reduce questions about automorphism groups to
the case where the graphs are connected.
A recent survey with a somewhat different emphasis is that of Babai and Good-
man [5]. In addition, no serious student should be without Hahn and Sabidussis
book [28], which contains surveys of aspects of graph symmetry.
2. Algorithmic aspects
Two algorithmic questions that arise from the above denitions are graph isomor-
phism and nding the automorphism group. The rst is a decision problem.
Graph isomorphism
Instance: Graphs G and H
Question: Is G

= H?
The second problem requires output. Note that a subgroup of S
n
may be super-
exponentially large in terms of n, but that any subgroup has a generating set of
size O(n), which species it in polynomial space.
Automorphism group
Instance: A graph G
Output: generating permutations for Aut(G)
These two problems are closely related: indeed, the rst has a polynomial
reduction to the second. For, suppose that we are given two graphs G and H. By
taking complements if necessary, we may assume that both G and H are connected.
Now suppose that we can nd generating permutations for Aut(K), where K is
the disjoint union of G and H. Then G and H are isomorphic if and only if some
generator interchanges the two connected components.
Conversely, if we can solve the graph isomorphism problem, we can at least
check whether a graph has a non-trivial automorphism, by attaching distinctive
gadgets at each vertex and checking whether any pair of the resulting graphs
are isomorphic. (Finding generators for the automorphism group may be more
difcult.)
The exact status of these two problems is unresolved. They belong to a select
group of problems that belong to NP but are not known either to belong to P or
to be NP-complete. For some particular classes of graphs notably, graphs of
bounded valency [43] and graphs with bounded eigenvalue multiplicity [7] the
isomorphism problem is known to be polynomial; see Garey and Johnson [23] for
the fundamentals of computational complexity.
In practice, these questions can be resolved for graphs with thousands of ver-
tices. Chapter 10 gives an account of the algorithms used and their implementation.
It turns out that, for almost all graphs, the algorithmic questions can be answered
very quickly. However, almost all does not include some of the most interesting
graphs, including strongly regular graphs (discussed in Chapter 8).
3. Automorphisms of typical graphs
The smallest graph whose automorphism group is trivial (apart from the 1-vertex
graph) is shown in Fig. 2. However, small graphs are (as usual) not a reliable guide
here. Erd os and R enyi [16] proved the following result.
Theorem 3.1 Almost all graphs have no non-trivial automorphisms.
Thus, the proportion of graphs on n vertices that have a non-trivial automorphism
tends to 0 as n : this is true whether we take labelled or unlabelled graphs.
As noted in the introduction, this theorem implies that almost all graphs can be
labelled in n! different ways, so that the number of unlabelled graphs on n vertices
is asymptotically 2
n(n1)/2
/n!. (There are clearly 2
n(n1)/2
labelled graphs on the
vertex-set {1, 2, . . . , n], since we can choose whether or not to join each pair
of vertices by an edge.) There are now good estimates for the error term in the
asymptotic expansion; it arises from graphs with non-trivial symmetry, and so
these estimates quantify the theorem.
In fact, more is true. There are various methods for canonically labelling a
graph for example, choose the lexicographically least labelled graph in the iso-
morphism class. For almost all graphs, the canonical labelling is unique, and can
Fig. 2.
be found in polynomial time; for such graphs, we can verify efciently that their
automorphismgroups are trivial. Typically, graphs with regularity properties, such
as strongly regular graphs (Chapter 8), are hard for canonical labelling algorithms,
even when their automorphism groups are trivial.
The theorem remains true for various special classes of graphs. These include
regular graphs of xed degree k > 2 we can even allow the degree to grow, not
too rapidly, with n(see [59]). They also include the prolic strongly regular graphs
of Latin square or Steiner triple system type discussed in Chapter 8; this uses the
fact that almost all of these structures have no non-trivial automorphisms.
Other methods of quantifying the theoremcan be found. For example, any given
graph can be altered so that some two vertices have the same neighbour sets by
changing at most n/2 adjacencies. The resulting graph has an automorphism that
interchanges the two vertices and xes all the others. Erd os and R enyi [16] showed
that, for almost all graphs, this is the shortest distance to symmetry.
4. Permutation groups
The question, Which permutation groups are the full automorphism groups of
graphs?, has no easy answer. Given a permutation group G on a set , we can
describe all of the graphs on which G acts, as follows. There is a coordinate-wise
action of G on , given by (v, w)g = (vg, wg). Let U be the set of all orbits
of G on that consist of pairs of distinct elements. There is a natural pairing
of orbits in U, where an orbit O is paired with O
= {(v, w) : (w, v) O]. Now

let S be any subset of U that contains the orbit paired with each of its members, and
dene a graph G(S) on the vertex-set by the rule that v w in G(S) if and only
if (v, w) O for some O S. Then G(S) is a simple graph admitting G as a group
of automorphisms; furthermore, every such graph arises by this construction.
The construction may easily be adapted to other classes of graphs. For directed
graphs, we drop the requirement that S is closed under pairing; for graphs with
loops, we include all orbits on , not just those consisting of pairs of distinct
elements; and for multigraphs, we allowmultisets of orbits. It is also very practical:
instead of having to list all the edges in order to specify a graph, we need give only
a set of orbit representatives.
Apermutation group G on is 2-closed if every permutation of that preserves
all the G-orbits in belongs to G. More generally, the 2-closure of G is
the group of permutations that preserve all the G-orbits. These concepts were
introduced by Wielandt [56].
Now, for any graph G, the group Aut(G) is 2-closed, for the edge-set of G is a
union of orbits of Aut(G) and so is preserved by its 2-closure. The converse fails,
but not too badly: in fact, a permutation group G is 2-closed if and only if it is
the full automorphism group of an edge-coloured directed graph. (We associate
a colour with each G-orbit on pairs.) However, it is not easy to decide whether a
given permutation group is 2-closed.
The construction of a graph from a permutation group has been reversed; for
example, many of the sporadic simple groups discovered in the mid-twentieth
century were constructed as groups of automorphisms of particular graphs. The
simplest such construction is that of Higman and Sims [36]: the vertex-set consists
of a special symbol together with the 22 points and 77 blocks of the Witt design
W
22
. We join to all the points; we join a point and a block if they are incident,
and we join two blocks if they are disjoint. It is clear that the automorphism group
of W
22
acts as a group of automorphisms of the graph, xing . It is not hard to
showthat the full automorphismgroup is transitive, and contains the Higman-Sims
simple group as a subgroup of index 2.
In the case of the Fischer groups [18], the group and the graph are even more
closely related. The vertices of the graph are the elements of a particular conjugacy
class of involutions inthe group, twovertices beingjoinedwhenever the involutions
do not commute.
5. Abstract groups
If we consider groups as abstract algebraic objects, rather than as concrete permu-
tation groups, a clear-cut result is possible. Frucht [20] proved the following basic
result.
Theorem 5.1 Every group is the automorphism group of some graph. Moreover,
if the group is nite, then the graph can be taken to be nite.
Subsequently, Frucht [21] showed that every group is the automorphism group
of a trivalent graph, and this has inspired a large number of similar results. We call
a class C of structures universal if every nite group is the automorphism group of
a structure in C. The combined results of Frucht, Sabidussi, Mendelsohn, Babai,
Kantor and others show that the following types of graph or other structures are
universal: regular graphs of degree k for any xed k > 2; bipartite graphs; strongly
regular graphs; Hamiltonian graphs; k-connected graphs for k > 0; k-chromatic
graphs, for k > 1; switching classes of graphs; lattices; projective planes (possibly
innite); Steiner triple systems; symmetric designs (BIBDs).
For example, the universality of Steiner triple systems is shown by using a
graph to construct a triple system with the same symmetry; then the universality
of strongly regular graphs is shown by taking the line graphs of triple systems
(see Chapter 8). The structures are nite in all cases except for projective planes;
Even the question of whether every nite group is a subgroup of the automorphism
group of a nite projective plane is still open.
Another class of results involves structures for which there is some obvious
restriction on the automorphismgroup. For example, a tournament cannot admit an
automorphismof order 2 (for such an automorphismwould necessarily interchange
the ends of some arc), and so the automorphism group has odd order. Moon [46]
showed that every group of odd order is the automorphism group of a tournament,
so this is the only restriction.
P olya observed that not every group is the automorphism group of a tree. More
precisely, the class of automorphism groups of trees is the smallest class that
contains the trivial group and is closed under the direct product and the operation
take the wreath product with the symmetric group S
n
of degree n, for each n > 1.
Automorphism groups of trees are of further importance in group theory. Any
nite tree has either a vertex or an edge which is xed by all automorphisms,
according as it is central or bicentral. Things are very different for innite trees.
We say that a group G acts freely on a tree T if only the identity element of G xes
any vertex or edge of T. Serre [49] has proved the following result.
Theorem 5.2 A group is free if and only if it acts freely on a tree.
This simple observation has led to the Bass-Serre theory in combinatorial group
theory, describing certain group constructions (amalgamated free products and
HNN-extensions) in terms of their actions on a tree; see Dicks and Dunwoody
[15] for an account of this.
Usually such a precise description does not exist. For example, every compo-
sition factor of the automorphism group of a planar graph is either a cyclic group
or an alternating group. More generally, Babai [2] showed the following result.
Theorem 5.3 Given k 0, there is a number f (k) such that, if G is the automor-
phism group of any graph embeddable in the orientable surface of genus k, then
every composition factor of G is cyclic, alternating, or of order at most f (k).
Still more generally, Babai showed that no class of nite graphs which is closed
under subgraphs andcontractions canbe universal, except for the class of all graphs.
We can quantify Fruchts Theorem by asking the question:
given a group G, what is the smallest number of vertices of a graph G with
Aut(G) = G?
This function behaves in a very erratic fashion. For example, the symmetric group
S
n
is the automorphismgroup of the null graph on n vertices, but the smallest graph
whose automorphism group is the alternating group A
n
has about 2
n
vertices; the
exact number was calculated by Liebeck [41]. Clearly this number is not smaller
than the degree of the smallest faithful permutation representation of G; such
questions have been investigated in detail by Babai et al. [6].
Other measures of the size of the graph can be used for example, the number
of edges, or the number of orbits of G on the set of vertices or edges: see [5] for a
survey of results on these questions.
Given that almost all graphs admit only the identity automorphism, we might
wonder whether Fruchts Theorem can be strengthened to state that almost all
graphs that admit a given group G actually have G as their full automorphismgroup.
This holds for some groups, but not all. Cameron [9] proved the following theorem:
Theorem 5.4 Given a group G, consider those n-vertex graphs whose automor-
phism group contains G. The proportion of such graphs whose automorphism
group is precisely G tends to a limit a(G) as n .
The limit a(G) is a rational number, but (unlike the case where G is the trivial
group) is not necessarily 1. For example, if G is the dihedral group of order 10,
then the limit is 1/3; this is because, of those graphs that admit G, almost all have
G acting on a set of ve vertices and xing the rest of the graph. (A random
graph admitting G consists of a random graph on n 5 vertices, and ve special
vertices on which G acts as the symmetry group of a pentagon, all joined to the
same random subset of the other n 5 vertices.) The subgraph induced by the ve
special vertices may be complete, null, or a 5-cycle; only in the third case is G
almost surely the full group.
In fact, a(G) = 1 if and only if G is a direct product of symmetric groups. For
abelian groups, a(G) = 1 if G is an elementary abelian 2-group (possibly trivial),
and a(G) = 0 otherwise. For metabelian groups, the values of a(G) are dense in
the unit interval.
It would be interesting to know whether similar results hold under hypotheses
that tend to work against such very local symmetries; for example, does a similar
result hold for regular graphs?
6. Cayley graphs
A permutation group G is regular if it is transitive and only the identity stabilizes a
point. Any regular action can be identied with the action of the group on itself by
right multiplication, where the group element g induces the permutation x . xg
of G. This is the action used by Cayley to show that every group is isomorphic to
a permutation group.
The orbits on pairs of this group are parametrized by group elements: they
have the form O
g
= {(x, gx) : x G] for g G. The orbit paired with O
g
is O
g
1 .
So our description of a G-invariant graph can be specialized in the following
way.
Let S be a subset of G that is closed under taking inverses and does not contain
the identity. The Cayley graph Cay(G, S) has vertex-set G, and edges
{x, sx], for all s S, x G. Cayley graphs are discussed in Chapter 6.
Since G is a subgroupof Aut(Cay(G, S)), this Cayleygraphis necessarilyvertex-
transitive (see Section 7). It is connected if and only if the set S generates G. Note
that many authors use a different convention, in which the action of the group is
by left multiplication, and the edges have the form {x, xs], for s S, x G. The
difference is immaterial.
A graphical regular representation of a group G is dened to be a graph for
which the regular action of G is the full automorphism group that is, a Cayley
graph Cay(G, S) with Aut(Cay(G, S)) = G. A considerable amount of effort went
into the determination of groups which have such representations, and the problem
was nally solved by Hetzel [35] for soluble groups, and Godsil [25] in general.
First note that an abelian group of exponent greater than 2 never has a graphical
regular representation, since any Cayley graph for an abelian group admits the
automorphism g .g
1
. A generalized dicyclic group is a group that has a cyclic
subgroup H of index 2 and an element g of order 4 for which g
1
hg = h
1
, for
all h H. The quaternion group of order 8 is an example.
The basic theorem is as follows:
Theorem 6.1 A nite group has a graphical regular representation if and only if
it is not an abelian group of exponent greater than 2, a generalized dicyclic group,
or one of thirteen exceptional groups with order at most 32.
What can be said about random Cayley graphs for G, obtained by including
inverse pairs of non-identity elements in S with probability
1
2
? Babai and Godsil [4]
conjecturedthat, except for the twoinnite classes inthe theorem, almost all Cayley
graphs for G are graphical regular representations that is, the probability that a
randomCayley graph is such a representation tends to 1 as [G[ . They proved
that this is true in some cases for example, non-abelian nilpotent groups of odd
order.
RandomCayley graphs have other useful properties; for example, they are often
expanders (see Alon and Roichman [1]).
7. Vertex-transitive graphs
A graph G is vertex-transitive if the automorphism group of G acts transitively on
the vertex-set of G; for example, the Petersen graph is vertex-transitive.
Every vertex-transitive graph has a description as a Schreier coset graph, gen-
eralizing the representation of a Cayley graph discussed above: we replace the
group elements by cosets of a subgroup H of G as vertices of the graph, and for
adjacency, we replace an inverse-closed set of elements by an inverse-closed set
of double cosets of H in G. Sabidussi [47] used this representation to show that
any vertex-transitive graph has a multiple that is a Cayley graph. (Here, a multi-
ple of a graph is obtained by replacing each vertex by an independent set of size
k, and each edge by all possible edges between the corresponding cocliques, for
some k.)
Not every vertex-transitive graph is a Cayley graph. The smallest counter-
example is the Petersen graph: it has no automorphism of order 2 that xes no
vertex. McKay and Praeger [44] have considered the class of vertex-transitive
graphs that are not Cayley graphs.
Maru si c and Jordan independently conjectured that any vertex-transitive graph
has a group of automorphisms that acts semi-regularly on vertices that is, the
stabilizer of any vertex is the identity, but the subgroup is not required to be
transitive. This conjecture was extended by Klin, who conjectured that any 2-
closed permutation group contains such a subgroup. This conjecture is still open,
although Giudici [24] has made substantial progress on it recently.
Obviously, all vertex-transitive graphs are regular. However, they also have
some special properties that are not shared by all regular graphs. From the work
of Mader, Watkins, Little, Grant, Holton, Babai and others, we take the following
list. We need one further denition: a graph is vertex-primitive if no equivalence
relation on the vertex-set is preserved by all automorphisms, apart from the trivial
relations (equality and the universal equivalence).
Theorem 7.1 Let G be a connected k-regular vertex-transitive graph of order n.
Then
r
G is
2
3
(k 1)-connected (and k-connected if it is vertex-primitive);
r
G is k-edge-connected;
r
G has a 1-factor if n is even;
r
G has a cycle of length at least
6n;
r
the product of the clique number and the independence number of G is at
most n.
It has been conjectured that, with nitely many exceptions, every connected
vertex-transitive graph is Hamiltonian. The Petersen graph is one of these excep-
tions: it has a Hamiltonian path but no Hamiltonian cycle. Only four exceptional
graphs are currently known, and all have Hamiltonian paths.
The Hadwiger number of a graph is the smallest number k for which some
component of the graph can be contracted to the complete graph K
k
. A graph
is toroidal if it is embeddable in the torus, and is ring-like if the vertices can
be partitioned into sets S
0
, S
1
, . . . , S
n1
, such that all edges join vertices in the
same set or cyclically in consecutive sets, and the automorphism group induces a
cyclic or dihedral group on this family of sets. Using these concepts, Babai [3] and
Thomassen [50] have recently obtained structure theorems for connected vertex-
transitive graphs with prescribed Hadwiger number.
Theorem 7.2 Every sufciently large connected vertex-transitive graph with
Hadwiger number k is either toroidal or ring-like, with the cardinalities of the
sets S
i
bounded by a function of the Hadwiger number in the ring-like case.
Clearly, arbitrarily large toroidal vertex-transitive graphs can be obtained as
quotients of plane lattices for example, rectangular grids with opposite sides
identied. The proof of this substantial result involves many geometrical ideas,
including isoperimetric inequalities for the hyperbolic plane. A related result of
Thomassen [51] shows that there are only nitely many vertex-transitive graphs
of given genus g 3.
Twoproperties that are weaker thanvertex-transitivity, but stronger thanregular-
ity, are walk-regularity and neighbourhood-regularity. The rst of these is touched
on in Chapter 8; here we consider the second.
Let a graph H be given. A graph G is locally H if, for each vertex v V(G),
the induced subgraph on the set of neighbours of v is isomorphic to H. A graph is
neighbourhood-regular if it is locally H, for some H.
The problem of deciding, for a given graph H, whether there is a graph that is
locally H, is recursively unsolvable (see [8] and [57]). Nevertheless, there are a
number of positive results. For example:
r
For some graphs H, all graphs that are locally H have been determined; see
Hall [32] for locally Petersen graphs, for example.
r
If H is regular and connected with girth at least 6, then there are innite graphs
that are locally H; see Weetman [54].
r
If H is regular with diameter 2, and if it satises some extra conditions, then
every locally H graph is nite; see Weetman [55].
A property of graphs that does not obviously relate to symmetry, but turns out
to imply vertex-transitivity, is compactness. The basic results on this concept are
due to Tinhofer [52]. To dene it, we note that any permutation g of {1, 2, . . . , n]
can be represented by a permutation matrix P(g), and that g Aut(G) if and only
if P(g) commutes with A, where A is the adjacency matrix of G.
A matrix M is doubly stochastic if its entries are non-negative and each
of its row and column sums is 1. By a theorem of Birkhoff (see Chapter 2,
Theorem 5.2), any doubly stochastic matrix is a convex combination of permuta-
tion matrices.
Agraph G with adjacency matrix Ais compact if every doubly stochastic matrix
M that commutes with A is a convex combination of permutation matrices that
commute with A that is, automorphisms of A.
The set of doubly stochastic matrices that commute with A is a polytope. If G
is compact, its automorphisms are precisely the extreme points of this polytope,
and can be found efciently by linear programming. Birkhoffs theorem shows
that the complete graph is compact.
For arbitrary graphs the meaning of compactness is somewhat mysterious, but
for regular graphs, we have the following result. Note that if a graph is compact,
then so is its complement.
Theorem 7.3 If G is a compact connected regular graph, then any two vertices
of G can be interchanged by an automorphism of G. In particular, G is vertex-
transitive.
The converse of this theorem is false. If G is compact and regular of degree k,
then (1/k)A is a doubly stochastic matrix that commutes with A, and is thus
a convex combination of automorphisms. Each such automorphism g has the
property that vg v for all vertices v. However, many vertex-transitive graphs,
such as the Petersen graph, have no non-identity automorphisms with this property.
For more on compact graphs, see Godsil [14]. Amore general concept called weak
compactness has been considered by Evdokimov et al. [17].
8. Higher symmetry
Symmetry conditions related to (and mostly stronger than) vertex-transitivity have
received a lot of attention, often using group-theoretic techniques. One of the
simplest is edge-transitivity, which usually implies vertex-transitivity. Indeed, an
edge-transitive graph that is not vertex-transitive must be bipartite, with the orbits
of the automorphism group as the bipartite blocks; the complete bipartite graph
K
r,s
with r ,= s is a simple example. There are also graphs that are vertex-transitive
and edge-transitive but not arc-transitive (where an arc is a directed edge). For arc-
transitive graphs, the connection between graphs and groups becomes particularly
strong; we refer to Chapter 7 for a survey.
A strengthening of arc-transitivity is distance-transitivity, where we require
the automorphism group to act transitively on pairs of vertices at distance i , for
i = 0, 1, . . . , d, where d is the diameter of the graph. A major research effort
directed at the determination of all such graphs is discussed in Chapter 9.
An even stronger symmetry condition is homogeneity: a graph G is homo-
geneous if each isomorphism between (nite) induced subgraphs of G can be
extended to an automorphism of G. All nite homogeneous graphs have been
determined by Sheehan and Gardiner (see [22]).
Theorem 8.1 A nite graph is homogeneous if and only if it is one of the fol-
lowing:
r
a disjoint union of complete graphs of the same order;
r
a regular complete multipartite graph;
r
the 5-cycle C
5
;
r
the line graph of K
3,3
.
More generally, we say that a graph G is t-homogeneous if any isomorphism
between induced subgraphs of order at most t extends to an automorphism of G.
Now t -homogeneity obviously implies the combinatorial property C(t ) dened in
Chapter 8. The list of graphs that satisfy C(5) is the same as the list of homo-
geneous graphs in the preceding theorem, so the hypothesis can be weakened to
5-homogeneity.
9. Innite graphs
We turn nowto innite graphs. Here, there are two very different areas of research
the rst for locally nite graphs (in which each vertex has nite degree), and the
second for general graphs (but usually requiring homogeneity or some model-
theoretic notions).
For a locally nite graph, the notion of an end (introduced by Halin [30] for
graphs, although used earlier for groups) is crucial. A ray is a one-way innite
path in a graph. K onigs Innity lemma (see [39]) shows that any locally nite
connected innite graph contains a ray. Let R(G) be the set of rays in G. We dene
an equivalence relation on R(G) by saying that R
1
R
2
if there is a ray R
3
that intersects both R
1
and R
2
in innitely many vertices. The equivalence classes
of are the ends of G; we denote the set of ends by E(G).
It can be shown that the number of ends of a locally nite graph G is the
supremum of the number of innite components of G S, taken over all nite
subsets S of V(G). (This result does not distinguish among innite cardinals: we
just say that the supremum of an unbounded set of natural numbers is .) For
example, the integer lattice graph Z
k
has just one end for k > 1, but two ends for
k = 1; an innite tree with maximum degree greater than 2 has innitely many
(indeed, uncountably many) ends.
The main results connecting ends and automorphisms are the following theo-
rems of Halin [31] and Jung [38], respectively.
Theorem 9.1 Every automorphismof a locally nite connected innite graph xes
either an end or a nite subgraph.
Theorem 9.2 Let G be a locally nite connected innite graph, and suppose that
Aut(G) has only nitely many orbits in V(G). Then the number of ends of G is 1,
2 or 2
0
.
For a locally nite graph, we can also consider the rate of growth of the number
a
n
of vertices that are at distance at most n from a xed vertex v. This number
depends on the chosen vertex v, but if the distance from u to v is d, then
a
nd
(v) a
n
(u) a
nd
(v).
Thus, the asymptotics of the rate of growth (for example, polynomial of degree k
or exponential with constant c) do not depend on v, and we can talk of the growth
of G. If G = Cay(G, S), then the choice of a nite generating set S for G does
not affect the asymptotics of growth, and we can again talk of the growth of G.
(Note, however, that a group can act vertex-transitively on each of two graphs with
different growth.) The growth is polynomial of degree k for the integer lattice Z
k
,
and exponential with constant k 1 for the k-regular tree.
This different behaviour is related to the number of ends.
Theorem 9.3 Let G be a locally nite connected innite graph whose automor-
phism group has only nitely many orbits.
(a) If the growth of G is bounded by a polynomial, then
c
1
n
k
a
n
c
2
n
k
,
where k is a positive integer and c
1
, c
2
> 0.
(b) G has linear growth if and only if it has two ends.
(c) If G is a Cayley graph of G, then the growth is polynomial if and only if G is
nilpotent-by-nite.
(d) There exist groups whose growth is faster than polynomial but slower than
exponential.
(e) If G has innitely many ends, then it has exponential growth.
Here part (c) is a celebrated theorem of Wolf [58] and Gromov [27] on groups
of polynomial growth; see Tromov [53] for an extension to vertex-transitive
graphs. Part (b) follows from Gromov and from Seifter and Tromov [48], and an
example for part (d) is the Grigorchuk group [26]. For a survey of vertex-transitive
graphs with polynomial growth, see [37]. There are also connections between the
growth of a graph and harmonic analysis; Lubotzky [42] gives an account of this
material.
Macpherson [45] has determined the locally nite innite graphs that are
distance-transitive. For any integers s, t > 1, there is an innite tree that is semi-
regular, with degrees s and t in the two bipartite blocks. Let M(s, t ) be the graph
with the bipartite block of degree s as its vertex-set; and where two vertices are
adjacent if they lie at distance 2 in the tree.
Theorem 9.4 Every locally nite distance-transitive innite graph is isomorphic
to M(s, t ), for some s, t > 1.
No such result holds without local niteness. Some examples to illustrate this are
given in Cameron [12].
Turning to arbitrary innite graphs, we rst describe the paradoxical result
of Erd os and R enyi [16]: up to isomorphism there is a unique countable random
graph that is, a graph R such that any countable randomgraph is isomorphic to R
with probability 1. Moreover, R is highly symmetric indeed, it is homogeneous,
as dened in the preceding section so the typical asymmetry of nite graphs does
not hold in the countably innite! A survey of this remarkable graph R appears
in [11].
Other aspects of the theory are also very different in the countably innite
case, largely as a result of the random graph R. For example, in any group G, the
square-root set is the set

g = {x G : x
2
= g]; it is non-principal if g ,= 1.
The hypotheses of the following theorem of Cameron and Johnson are very
mild.
Theorem 9.5 Let G be a countable group that is not the union of a nite number
of translates of non-principal square root sets. Then almost all random Cayley
graphs for G are isomorphic to R.
The classication of homogeneous graphs was extended to the innite case by
Lachlan and Woodrow[40]. Agraph is universal K
n
-free if it contains no complete
graph of size n but contains every K
n
-free graph as an induced subgraph. For each
n 2, there is just one countable homogeneous universal K
n
-free graph; these
graphs were rst constructed by Henson [34], but their existence and uniqueness
follow from a general construction method of Frass e [19]. We denote this unique
graph by H
n
.
Theorem 9.6 Every countable homogeneous graph is one of the following:
r
a disjoint union of complete graphs of the same size, or its complement (a
regular complete multipartite graph);
r
Hensons graph H
n
(for n 3), or its complement;
r
the random graph R.
10. Graph homomorphisms
In this nal section we turn to graph homomorphisms. In general, homomorphisms
are more revealingof graphstructure thanautomorphisms, andthe theoryhas devel-
oped in surprising directions. The material here is based on Hahn and Tardif [29],
to which we refer for more details and references. In this section, we consider only
nite simple graphs.
A homomorphism from a graph G to a graph H is a function f from V(G)
to V(H) such that, if v w in G, then vf wf in H. Thus, an isomorphism
is a bijective homomorphism whose inverse is also a homomorphism. We write
G H if there is a homomorphism from G to H.
If G H and H G, then we say that G and H are homomorphically
equivalent, and write G H. We write the homomorphic equivalence class of
G as [G]. The set of such equivalence classes is partially ordered by the rule
that [G] _ [H] if G H. Much is known about this partial order: it is a lattice
order, and it is dense that is, if [G] [H], then there exists a graph for which
[G] [K] [H].
A proper vertex-colouring of G with r colours, being a map from V(G) to
{1, 2, . . . , r] such that adjacent vertices have distinct images, is a homomorphism
from G to the complete graph K
r
; for example, every bipartite graph has a homo-
morphism onto K
2
. We think of G H as saying that G has a H-colouring.
Thus, the existence and enumeration questions for homomorphisms from G to
H generalize the chromatic number and chromatic polynomial of G, and so we
expect these questions to be hard!
The independence ratio i (G) of a graph G is the ratio of the cardinality of the
largest independent set of vertices in G to the total number of vertices of G. The
odd girth g
o
(G) of G is the length of a shortest odd cycle in G, and (G) and (G)
denote the clique number and the chromatic number of G. The following result
gives a necessary condition for the existence of a homomorphism from G to H:
Theorem 10.1 Suppose that G H. Then
r
(G) (H), (G) (H) and g
o
(G) g
o
(H);
r
if H is vertex-transitive, then i (G) i (H).
A retraction of a graph G is a homomorphism f from G onto an induced
subgraph H of G for which the restriction of f to V(H) is the identity map; the
subgraph H is then called a retract of G. It is easy to see that some power of every
homomorphism from G to G is a retraction. Any retraction of a connected graph
can be expressed as the composition of a sequence of foldings, where a folding is
a homomorphism that identies one pair of vertices. The result of Sabidussi [47]
mentioned in Section 7 shows that every vertex-transitive graph is a retract of a
Cayley graph.
Retracts play an important role in topology, based in part on the fact that every
retract H of a graph G is an isometric subgraph of G that is, the distance between
two vertices of H is the same in both graphs.
A graph G is a core if it has no non-trivial retraction that is, if every homo-
morphism is an automorphism. If H is a retract of G and is itself a core, then we
say that H is a core of G. The following result holds.
Theorem 10.2
(a) If G H, then any core of G is isomorphic to any core of H.
(b) In particular, all cores of G are isomorphic, and we can speak of the core
of G.
(c) Up to isomorphism, the core of G is the smallest graph in [G].
(d) The core of a vertex-transitive graph is vertex-transitive.
Many cores are known examples include all circulants of prime order, and the
Kneser graph K(r, s) whose vertices correspond to the s-subsets of {1, 2, . . . , r]
with r > 2s and where two vertices are adjacent if the corresponding sets are
disjoint (the Petersen graph is K(5, 2)). For a survey of cores and their properties,
see Hell and Ne setril [33].
References
1. N. Alon and Y. Roichman, Random Cayley graphs and expanders, Random Structures
Appl. 5 (1994), 271284.
2. L. Babai, Automorphism groups of graphs and edge-contraction, Discrete Math. 8
(1974), 1320.
3. L. Babai, Vertex-transitive graphs and vertex-transitive maps, J. Graph Theory 15
(1991), 587627.
4. L. Babai and C. D. Godsil, On the automorphism groups of almost all Cayley graphs,
Europ. J. Combin. 3 (1982), 915.
5. L. Babai and A. J. Goodman, On the abstract group of automorphisms, Coding Theory,
Design Theory, Group Theory (eds. D. Jungnickel and S. A. Vanstone), Wiley (1993),
121143.
6. L. Babai, A. J. Goodman and L. Pyber, On faithful permutation representations of small
degree, Comm. Algebra 21 (1993), 15871602.
7. L. Babai, D. Yu. Grigoryev and D. M. Mount, Isomorphism of graphs with bounded
eigenvalue multiplicity, Proc. 14th ACM STOC (1982), 310324.
8. V. K. Bulitko, O grafach s zadanymi okruzeniami ver sin (in Russian), Trudy Mat. Inst.
im. Steklova 133 (1973), 7894.
9. P. J. Cameron, On graphs with given automorphism group, Europ. J. Combin. 1 (1980),
9196.
10. P. J. Cameron, 6-transitive graphs, J. Combin. Theory (B) 28 (1980), 168179.
11. P. J. Cameron, The random graph, The Mathematics of Paul Erd os (eds. J. Ne setril and
R. L. Graham), Springer (1996), 331351.
12. P. J. Cameron, A census of innite distance-transitive graphs, Discrete Math. 192
(1998), 1126.
13. P. J. Cameron, Permutation Groups, Cambridge Univ. Press, 1999.
14. A. Chan and C. D. Godsil, Symmetry and eigenvectors, Graph Symmetry: Algebraic
Methods and Applications (eds. G. Hahn and G. Sabidussi), Kluwer (1997), 75106.
15. W. Dicks and M. J. Dunwoody, Groups Acting on Graphs, Cambridge Univ. Press,
1989.
16. P. Erd os and A. R enyi, Asymmetric graphs, Acta Math. Acad. Sci. Hungar. 14 (1963),
295315.
17. S. Evdokimov, M. Karpinski and I. Ponomarenko, Compact cellular algebras and per-
mutation groups, Discrete Math. 197/198 (1999), 247267.
18. B. Fischer, Finite groups generated by 3-transpositions, I, Invent. Math. 13 (1971),
232246.
19. R. Frass e, Sur certains relations qui g en eralisent lordre des nombres rationnels, C. R.
Acad. Sci. Paris 237 (1953), 540542.
20. R. Frucht, Herstellung von Graphen mit vorgegebener abstrakter Gruppe, Compositio
Math. 6 (1938), 239250.
21. R. Frucht, Graphs of degree 3 with given abstract group, Canad. J. Math. 1 (1949),
365378.
22. A. D. Gardiner, Homogeneous graphs, J. Combin. Theory (B) 20 (1976), 94102.
23. M. R. Garey and D. S. Johnson, Computers and Intractibility: A Guide to the Theory
of NP-Completeness, W. H. Freeman, 1979.
24. M. Guidici, Quasiprimitive groups with no xed-point-free elements of prime order,
J. London Math. Soc. (2) 67 (2003), 7384.
25. C. D. Godsil, GRRs for non-solvable groups, Colloq. Math. Soc. J anos Bolyai 25
(1981), 221239.
26. R. I. Grigorchuk, On the Burnside problem for periodic groups, Funkcional Anal. i
Prilozen 14 (1980), 5354; transl. Functional Anal. Appl. 14 (1980), 4143.
27. M. Gromov, Groups of polynomial growth and expanding maps, Publ. Math. IHES 53
(1981), 5378.
28. G. Hahn and G. Sabidussi (eds.), Graph Symmetry: Algebraic Methods and Applica-
tions, Kluwer, 1997.
29. G. Hahn and C. Tardif, Graph homomorphisms: Structure and symmetry, Graph Sym-
metry: Algebraic Methods and Applications (eds. G. Hahn and G. Sabidussi), Kluwer
(1997), 107166.
30. R. Halin,

Uber unendliche Wege in Graphen, Math. Ann. 157 (1964), 125137.
31. R. Halin, Automorphisms and endomorphisms of innite locally nite graphs, Abh.
Math. Sem. Hamburg 39 (1973), 251283.
32. J. I. Hall, Locally Petersen graphs, J. Graph Theory 4 (1980), 173187.
33. P. Hell and J. Ne setril, The core of a graph, Discrete Math. 109 (1992), 117126.
34. C. W. Henson, A family of countable homogeneous graphs, Pacic J. Math. 38 (1971),
6983.
35. D. Hetzel,

Uber regul are graphische Darstellungen von au osbare Grupen, Diplom-
arbeit, Techn. Univ. Berlin, 1976.
36. D. G. Higman and C. C. Sims, A simple group of order 44,352,000, Math. Z. 105
(1968), 110113.
37. W. Imrich and N. Seifter, A survey on graphs with polynomial growth, Discrete Math.
95 (1991), 101117.
38. H. A. Jung, A note on fragments of innite graphs, Combinatorica 1 (1981), 285288.
39. D. K onig, Theorie der endlichen und unendlichen Graphen, Akad. Verlagsgesellschaft,
1936; English translation: Theory of Finite and Innite Graphs (transl. R. McCoart),
Birkh auser, 1990.
40. A. H. Lachlan and R. E. Woodrow, Countable ultrahomogeneous undirected graphs,
Trans. Amer. Math. Soc. 262 (1980), 5194.
41. On graphs whose full automorphism group is an alternating group or a nite classical
group, Proc. London Math. Soc. (3) 47 (1983), 337362.
42. A. Lubotzky, Discrete Groups, ExpandingGraphs, andInvariant Measures, Birkh auser,
1994.
43. E. M. Luks, Isomorphism of graphs of bounded valence can be tested in polynomial
time, J. Comp. Sys. Sci. 25 (1982), 4265.
44. B. D. McKay and C. E. Praeger, Vertex-transitive graphs which are not Cayley graphs,
I, J. Austral. Math. Soc. (A) 56 (1994), 5363.
45. H. D. Macpherson, Innite distance-transitive graphs of nite valency, Combinatorica
2 (1982), 6369.
46. J. W. Moon, Tournaments with a given automorphismgroup, Canad. J. Math. 16 (1964),
485489.
47. G. Sabidussi, Vertex-transitive graphs, Monatsh. Math. 68 (1964), 426438.
48. N. Seifter and V. I. Tromov, Automorphism groups of graphs with quadratic growth,
J. Combin. Theory (B) 71 (1997), 205210.
49. J.-P. Serre, Trees, Springer, 1980.
50. C. Thomassen, The Hadwiger number of innite vertex-transitive graphs, Combinator-
ica 12 (1992), 481491.
51. C. Thomassen, Embeddings of graphs, Discrete Math. 124 (1994), 217228.
52. G. Tinhofer, Graph isomorphism and theorems of Birkhoff type, Computing 36 (1986),
285300.
53. V. I. Tromov, Graphs with polynomial growth, Mat. Sbornik 123 (1984), 407421;
transl. Math. USSR Sbornik 51 (1985), 405417.
54. G. M. Weetman, A construction of locally homogeneous graphs, J. London Math. Soc.
(2) 50 (1994), 6886.
55. G. M. Weetman, Diameter bounds for graph extensions, J. London Math. Soc. (2) 50
(1994), 209221.
56. H. Wielandt, Permutation Groups through Invariant Relations and Invariant Functions,
Lecture Notes, Ohio State University, 1969.
57. P. M. Winkler, Existence of graphs with a given set of r-neighbourhoods, J. Combin.
Theory (B) 34 (1983), 165176.
58. J. Wolf, Growth of nitely generated solvable groups and curvature of Riemannian
manifolds, J. Diff. Geom. 2 (1968), 421446.
59. N. C. Wormald, Models of random regular graphs, Surveys in Combinatorics, 1999
(eds. J. D. Lamb and D. A. Preece), London Math. Soc. Lecture Notes Series 267,
Cambridge Univ. Press (1999), 239298.
6
Cayley graphs
BRIAN ALSPACH
1. Introduction
2. Recognition
3. Special examples
4. Prevalence
5. Isomorphism
6. Enumeration
7. Automorphisms
8. Subgraphs
9. Hamiltonicity
10. Factorization
11. Embeddings
12. Applications
References
Cayley graphs are constructed using groups. This chapter discusses algebraic
and graph-theoretic aspects of nite Cayley graphs. The algebraic aspects in-
clude recognition, isomorphism, prevalence, enumeration and automorphisms.
The graph-theoretic aspects include notable members of the family, subgraphs,
factorization, hamiltonicity and embeddings.
1. Introduction
There are several constructions that produce families of vertex-transitive graphs.
This chapter deals with the family of Cayley graphs, probably the most widely
known and extensively studied family of vertex-transitive graphs. The construction
for Cayley graphs is based on groups. We restrict ourselves to nite groups, but the
156
6 Cayley graphs 157
basic construction is the same for innite groups. While Cayley graphs on nite
groups and those on innite groups share a variety of features, there are aspects of
Cayley graphs on nite groups that do not carry over to those on innite groups,
and vice versa.
Let G be a nite group with identity 1. Let S be a subset of G satisfying 1 / S
and S = S
1
that is, s S if and only if s
1
S. The Cayley graph Cay(G; S)
on G with connection set S is dened as follows:
r
the vertices are the elements of G;
r
there is an edge joining g and h if and only if h = sg for some s S.
The set of all Cayley graphs on G is denoted by Cay(G).
When G is an Abelian group, we use additive notation for the group operation.
Hence, we write S = S for the connection set, and h = s + g (for some s S)
for adjacency.
Cayley digraphs are dened in a similar way, but with the condition S = S
1
removed. Many results about Cayley graphs carry over immediately to Cayley
digraphs, but there are aspects of Cayley digraphs that distinguish them from
Cayley graphs. This chapter does not specically address Cayley digraphs, except
briey in Section 9.
2. Recognition
When G is a nite group and g G, dene g
R
acting on G by hg
R
= hg, for all
h G; clearly, g
R
is a permutation of the elements of G. Dene the group G
R
by
G
R
= {g
R
: g G}. It is easy to see that G
R
acts transitively on the elements of G,
andthat g
R
is anautomorphismof anyCayleygraphonG. These observations prove
the following result, thereby providing a large class of vertex-transitive graphs.
Theorem 2.1 Every Cayley graph is vertex-transitive.
The Cayley graphs on cyclic groups have played a special role in the study of
Cayley graphs. They are widely known as circulant graphs, because their adja-
cency matrices are circulant matrices. We use the notation Circ(n; S) to denote the
circulant graph of order n with connection set S. An example is given in Fig. 1.
How difcult is it to recognize whether an arbitrary graph is a Cayley graph,
or a vertex-transitive graph for that matter? A celebrated theorem of Sabidussi
addresses this recognition question. Before proceeding to the theorem, a denition
is required.
Let G be a transitive permutation group acting on a nite set . It is easy to
show that the following three conditions are equivalent:
158 Brian Alspach
1
2
3
4
5
6
7
Fig. 1. The circulant graph Circ(7; {2, 3, 4, 5})
r
the only element of G that xes an element of is the identity permutation;
r
|G| = ||;
r
for any
1
,
2
, there is a unique element g G satisfying
1
g =
2
.
A transitive permutation group G that satises any one of the these conditions is
said to be regular. We now state the fundamental result of Sabidussi [48].
Theorem 2.2 Agraph G is a Cayley graph if and only if Aut(G) contains a regular
subgroup.
Here is an outline of the proof. To prove the necessity, note that G
R
is regular
and that G
R
Aut(G), for any G Cay(G). To prove the sufciency, assume that
Aut(G) contains a regular subgroup G. Label an arbitrary vertex of G with the
identity element 1 of G. For each of the remaining vertices v V(G), there is a
unique permutation g
v
G that maps the vertex labelled 1 to v. Label v with the
group element g
v
. It is then easy to show that G is a Cayley graph on G.
Essentially all of the work that has been done on establishing whether particular
graphs are Cayley graphs employs Theorem 2.2, and the way in which the theo-
rem is used depends on what we are given. Sometimes we are provided a graph
construction whereby the automorphismgroup of the graph may be computed. We
may then determine whether the group contains a regular subgroup. At other times
we may not be able to produce the automorphismgroup easily, but still may be able
to show that a regular subgroup cannot occur. In other words, we are dealing with
cases in which our description of the graph is special and contains considerable
information regarding recognition.
In the case of arbitrary graphs, we assume that all we have is a description of
the graph as an adjacency matrix, or as a list of the edges in the graph. Recog-
nition now means that we are asking about the existence of an algorithm for
6 Cayley graphs 159
determining whether an arbitrary graph is a Cayley graph. There are good al-
gorithms for determining whether a graph is vertex-transitive, and they work on
graphs with thousands of vertices. They produce generators for the automorphism
group from which verifying vertex-transitivity is easy. However, in trying to rec-
ognize whether or not a vertex-transitive graph is a Cayley graph, we are left with
the problem of determining whether the automorphism group contains a regular
subgroup.
There has been some work on nding efcient algorithms for Cayley graph
recognition, but the classes studied have been limited to circulant tournaments and
special circulant graphs. Efciency is gained by working with association schemes
arising from the graphs (see [43]). We then must show that the association scheme
itself is cyclic, which may be done efciently, so that Theorem 2.2 is still present,
again demonstrating the fundamental nature of Sabidussis Theorem.
We shall ignore the computational problem of recognizing whether an arbitrary
graph is a Cayley graph. Instead, we always assume that Cayley graphs have
been described in terms of the groups on which they are built, together with the
connection sets. For most problems this is not a drawback.
3. Special examples
There are situations involving graphs for which Cayley graphs appear in a natural
way. Often they arise froma small select collection of Cayley graphs. In this section
we present some notable Cayley graphs that appear frequently in the literature. We
dene them in as straightforward a manner as possible, but point out that usually
they may be dened as Cayley graphs in many ways. This aspect will be discussed
in Section 4.
The complete graphs and their complements are Cayley graphs. In particular,
K
n
is a Cayley graph on any group G of order n, where the connection set is the
set of non-identity elements of the group.
The complete multipartite graph K
r(s)
, with r parts each of cardinality s, is
also a Cayley graph. It can be achieved by using the circulant graph of order rs
with the connection set consisting of all the elements that are not congruent to 0
(modulo s).
Of course, complete graphs and complete multipartite graphs appear all the
time, but for the most part their appearances are not related to the fact that they
are Cayley graphs. The rest of the special Cayley graphs we discuss owe much of
their interest to the fact they are Cayley graphs.
The k-dimensional cube Q
k
is the Cayley graph dened on the elementary
Abelian 2-group Z
k
2
, where the connection set is the standard generating set for
160 Brian Alspach
Z
k
2
. The graph Q
k
has generated a long history of interesting questions, and has
become even more important in recent years because of the role it plays in computer
architecture.
The graph formed on the nite eld F
q
, where q 1 (mod 4) and the connection
set is the set of quadratic residues in F
q
, is called a Paley graph; an example appears
in Fig. 2. Here we dene edges additively, since we are using an additive group as
the underlying vertex-set. Paley graphs have many interesting properties.
Let n be even and let S = {1, n/2}. The circulant graph of order n with
connection set S is known as the M obius ladder of order n. These graphs play a
role in topological graph theory.
Problems arising in computer science have brought about a resurgence of in-
terest in grid-like graphs. The corresponding Cayley graphs are the Cartesian
products of cycles of some xed length 3. These graphs are realized as
Cayley graphs by using the group Z
n
, where the connection set is the set of standard

generators of the group.
4. Prevalence
The family of Cayley graphs provides us with a straightforward construction
for vertex-transitive graphs. A natural rst question is to ask whether the fam-
ily of Cayley graphs encompasses all nite vertex-transitive graphs. We turn to the
Petersen graph P for a negative answer.
It is easy to showthat P is vertex-transitive. To showthat it is not a Cayley graph,
we show that Aut(P) does not contain a regular subgroup and apply Theorem 2.2.
1
2
3
4
5
6
8 9
10
11
12
13
7
Fig. 2. The Paley graph of order 13
6 Cayley graphs 161
We can prove this navely using some graph-theoretic facts about P. If G is a
regular subgroup of Aut(P) and
= (v
0
v
1
v
2
v
3
v
4
)(w
0
w
1
w
2
w
3
w
4
)
is an element of order 5 in G, then by using the facts that P is connected, is regular
of degree 3, is not bipartite, has girth 5, and by relabelling (if necessary), we may
assume that v
i
v
i +1
, w
i
w
i +2
and v
i
w
i
are edges, for i 0, 1, 2, 3, 4 (modulo 5).
Let be the element of G for which v
0
= w
0
. Then must have order 2 and
satisfy
1
= . This requires that cannot preserve edges of P, so that no
such can exist. Hence, Aut(P) contains no regular subgroup, and we conclude
that P is not a Cayley graph.
We pursue the preceding example in two directions. The essential reason that
the Petersen graph is not a Cayley graph is that no involution interchanges the two
orbits of . We examine this more carefully.
With the same labelling as before, the mapping dened by f : v
i
w
2i
and
f : w
i
v
2i
is an automorphism of P. Note that f has order 4, and that 2
2
= 1
in F
5
. This is the key to what follows.
For any prime p 1 (mod 4), 1 is a quadratic residue modulo p. Let
2
1 (mod p). Formthe circulant graph H

0
on Z
p
with connection set {1, 1} (recall
that we use additive notation for Abelian groups) and vertex-set {v
0
, v
1
, . . . , v
p1
},
and the circulant graph H
1
on Z
p
with connection set {, } and vertex-set
{w
0
, w
1
, . . . , w
p1
}. Dene the cubic graph GP( p,) as the vertex-disjoint union
of H
0
and H
1
, together with the perfect matching v
i
w
i
, for i = 0,1, . . . , p 1;
GP( p, ) is sometimes called a generalized Petersen graph; GP(13, 5) appears in
Fig. 3.
Fig. 3. The graph GP(13, 5)
162 Brian Alspach
The graph GP( p, ) can be shown to be a non-Cayley vertex-transitive graph,
in a manner similar to that used for P. In contrast, it is known that every vertex-
transitive graph of order 2p, when p is a prime satisfying p 3 (mod 4), is
a Cayley graph. In the same vein, Maru si c [35] proved that all vertex-transitive
graphs of orders p
2
and p
3
, where p is a prime, are Cayley graphs. This prompted
him to pose the problem of determining the set
NC = {n: there exists a non-Cayley vertex-transitive graph of order n}.
An important elementary feature of NC is that kNC = NC, for all positive
integers k. This follows from the observation that if G is a non-Cayley vertex-
transitive graph of order n, then the graph obtained by taking k vertex-disjoint
copies of G is a non-Cayley vertex-transitive graph of order kn. The complement
of the latter graph provides a connected example.
Since kNC = NC, for all positive integers k, Maru si cs problem is solved if
we can nd all the minimal elements of NC: minimal in this context means
that n NC, but no proper divisor of n belongs to NC.
It is convenient to consider the number of distinct prime factors when looking
for minimal elements. When there is only one prime factor that is, n is a prime
power the answer is completely settled. Every vertex-transitive graph of prime
order is a circulant graph, and, as stated above, Maru si c [35] showed that p
2
and
p
3
never belong to NC.
It was known that 16 NC, and for any odd prime p, McKay and Praeger [37]
showed by construction that p
4
NC. Their construction is easy to describe. Let
G( p) be the graph with vertex-set {v
i, j
: i Z
p,
j Z
p
3 } and edge-set
{v
i, j
v
i, j +pk
, v
i, j
v
i +1, j
, v
i, j
v
i +1, j +pa
i , v
i, j
v
i +1, j +a
r p+i },
where i Z
p
, j Z
p
3 , k Z
p
2 , r Z
p
, and a = p +1. Then G( p) does the job.
Thus, for any prime p, p
e
NC if and only if e 4. This implies that n NC
whenever n is divisible by p
4
, for some prime p. However, the latter result was
superseded by a much stronger result obtained by McKay and Praeger [37]. They
proved that any positive integer divisible by a square, other than 12, is in NC.
The preceding results reduce the problem of trying to characterize membership
in NCto the consideration of square-free integers. We already have seen the answer
when n = 2p, where p is an odd prime. The next theorem [38] characterizes
membership in NC for a product of two distinct primes.
Theorem 4.1 Let p and q be distinct primes, with p < q. Then pq NC if and
only if one of the following holds:
r
p
2
divides q 1;
r
q = 2p 1 > 3 or q = ( p
2
+1)/2;
r
q = 2
t
+1, and either p divides 2
t
1 or p = 2
t 1
1;
6 Cayley graphs 163
r
q = 2
t
1 and p = 2
t 1
+1;
r
p = 7 and q = 11.
The situation for a product of three distinct primes becomes more complicated,
but the answer is complete nonetheless (see [26]). We separate even and odd values.
Theorem 4.2 Let p and q be odd primes, with p < q. Then 2pq NCif and only
if one of the following holds:
r
p
2
divides q 1;
r
p 1 (mod 4) or q 1 (mod 4);
r
p = 7 and q = 11 or 19;
r
p q 3 (mod 4), p divides q 1, and p
2
does not divide q 1;
r
p q 3 (mod 4) and p = (q +1)/4.
Theorem 4.3 Let p, q, r be distinct odd primes, with p < q < r. Then pqr NC
if and only if at least one of pq, pr or qr is a member of NC, or none of pq, pr
and qr is a member of NC but one of the following holds:
r
pqr = (2
2
t
+1)(2
2
t +1
+1), for some t;
r
pqr = (2
d1
+1)(2
d
1), for some prime d;
r
pq = 2r +1, 2r 1 or (r +1)/2;
r
pq = (r
2
+1)/2 or pr = (q
2
+1)/2;
r
pq = (r
2
1)/x or pr = (q
2
1)/x, where x = 24, 48 or 120;
r
ab = 2
t
+1 and c divides 2
t
1, where {a, b, c} = { p, q, r};
r
the largest power of p dividing q 1 is p
p
, and the largest power of q dividing
r 1 is q
q
;
r
q = (3p +1)/2 and r = 3p +2, or q = 6p 1 and r = 6p +1;
r
q = (r 1)/2 and p divides r +1, where p > q when p = (r +1)/2;
r
p = (k
d/2
+1)/(k +1), q = (k
d/2
1)/(k 1), r = (k
d1
1)/(k 1),
where k, d 1 and d/2 are primes, and p > q may hold;
r
p = (k
(d1)/2
+1)/(k +1), q = (k
(k1)/2
1)/(k 1), r = (k
d
1)/(k 1),
where k, d and (d 1)/2 are primes, and p > q may hold;
r
p = k
2
k +1, q = (k
5
1)/(k 1), r = (k
7
1)/(k 1), where k is
prime;
r
p = 3, q = (2
d
+1)/3, r = 2
d
1, where d is prime;
r
p = (2
d
+1)/3, q = 2
d
1, r = 2
2d2
+1, where d = 2
t
1 is prime;
r
p = 5, q = 11 and r = 19;
r
p = 7, q = 73 and r = 257.
So far there is no characterization of the members of NC that are products of
four distinct primes. This leaves us with the following intriguing question. Is there
164 Brian Alspach
a number k > 0 for which every product of k distinct primes is in NC? With the
little evidence currently available, a guess in either direction appears hazardous.
The other direction that we wish to pursue, regarding the Petersen graph P as a
non-Cayley vertex-transitive graph, depends on viewing P in a different way. Let
the ten 2-element subsets of a 5-set correspond to the vertices of a graph, and let
two vertices be adjacent whenever their corresponding subsets are disjoint; then
the resulting graph is isomorphic to P.
The preceding is an interpretation of P as a Kneser graph (see Chapter 5). The
general denition of Kneser graphs is as follows. Let r and s be positive integers
satisfying r 2 and s 2r +1. Let the vertices of the graph K(s, r) correspond
to the r-subsets of an s-set, and let two vertices be adjacent if and only if their
corresponding r-subsets are disjoint; then the Petersen graph is K(5, 2). It is easy
to see that K(s, r) is vertex-transitive; in fact, the automorphismgroup of K(s, r) is
the action on the r-subsets induced by the action of S
s
on {1, 2, . . . , s}. Using this
fact, Godsil [20] proved that the only exceptions to K(s, r) being a non-Cayley
vertex-transitive graph occur when r = 2 and s is a prime power congruent to
3 (mod 4), or r = 3 and s = 8 or 32. The fact that

s
r
NC, with the exceptions

just mentioned, is not particularly useful for determining NC.
5. Isomorphism
Some of the most interesting and deepest work on Cayley graphs has revolved
around the question of trying to determine when two Cayley graphs are isomorphic.
For the most part, the question has been conned to Cayley graphs on the same
group, but there has been some work examining the isomorphismof Cayley graphs
on different groups. For a recent excellent survey that goes into far more detail than
this section, see Li [30]. Some preliminary denitions and results are required.
For historical reasons, as well as the levels of difculty involved, much of
the work on isomorphism problems has been concentrated on Cayley graphs on
Abelian groups. In fact, circulant graphs by themselves have garnered considerable
attention.
Let p be a prime and let Z
p
denote the multiplicative group of units of Z
p
. Dene
the permutation T
a,b
on Z
p
for a Z
p
and b Z
p
, by xT
a,b
= ax +b. Using the
notation G Hfor equivalent permutation groups, we have the following classical
theorem of Burnside.
Theorem 5.1 A transitive permutation group G of prime degree p is either doubly
transitive or G {T
a,b
: a H < Z
p
, b Z
p
}.
Burnsides theorem has some interesting applications for circulant graphs of
prime order. If Aut(G) is doubly transitive for a graph G, then G is either complete
6 Cayley graphs 165
or has no edges. Thus, Burnsides theorem reveals considerable information about
other circulant graphs of prime order.
The rst signicant isomorphism result about Cayley graphs was the following
theorem of Turner [52].
Theorem 5.2 Let p be a prime. Two circulant graphs Circ( p; S) and Circ( p; S
)
of order p are isomorphic if and only if S
= aS, for some a Z
p
.
Here is an outline of the proof. If S
= aS for some a Z
p
, then it is easy to
verify that T
a,0
is an isomorphism from Circ( p; S) to Circ( p; S
). On the other
hand, let Circ( p; S) and Circ( p; S
) be isomorphic circulant graphs on Z

p
via an
isomorphism . If Circ( p; S) is complete or has no edges, then the result follows
trivially. In all other cases, Theorem 5.1 implies that their automorphism groups
contain a unique subgroup of order p. Without loss of generality, we may assume
that the subgroup is generated by the p-cycle (0 1 2 p 1). Using the fact that
the p-cycle ((0) (1) (2) ( p 1)) generates the same subgroup, we easily
nd an a Z
p
satisfying S
= aS.
Ad am posed the question of whether two circulant graphs Circ(n; S) and

Circ(n; S
) are isomorphic if and only if there exists an a Z
n
satisfying S
= aS,
where Z
n
denotes the multiplicative group of units in Z
n
.
It is easy to describe a counter-example to

Ad ams conjecture. For n = 25, let
S = {1, 4, 5, 6, 9, 11, 14, 16, 19, 20, 21, 24}
and
S
= {1, 4, 6, 9, 10, 11, 14, 15, 16, 19, 21, 24}.

The two circulant graphs Circ(25, S) and Circ(25, S
) are isomorphic, since each

is a wreath product of a 5-cycle with a 5-cycle. On the other hand, it is easy to see
that there is no a Z
25
for which S
= aS.
Theorem 5.2 set the stage for research on isomorphisms of Cayley graphs. The
observation that multiplication by a Z
n
is an automorphism of the underlying
additive group indicates the direction to follow.
Let G be a nite group. Suppose that two Cayley graphs Cay(G; S) and
Cay(G; S
) on G are isomorphic if and only if there exists a group automorphism

Aut(G) such that S
= S; then we say that G is a CI-group. Rephrasing

Theorem 5.2 in terms of this denition, we see that Z
p
is a CI-group, where p
is a prime. The problem of characterizing nite CI-groups then became, and still
remains, one of the major problems on Cayley graphs. Some notable results have
been obtained, and we examine some of them now.
166 Brian Alspach
Elementary Abelian p-groups have been studied in some detail. Godsil [21]
proved that Z
2
p
is a CI-group, while Dobson [15] proved that Z
3
p
is a CI-group, for
each prime p. Hirasaka and Muzychuk [25] proved that Z
4
p
is a CI-group, for each
prime p. These results lend credence to an earlier conjecture of Babai and Frankl
[7] that Z
n
p
is a CI-group for all primes p and all n 1. However, Nowitz [44]
proved that Z
6
2
is not a CI-group by explicitly constructing two isomorphic Cayley
graphs on Z
6
2
between which no group automorphism acts as an isomorphism.
Nowitzs example is rather intriguing. Its degree is 32, so the degree of the com-
plements, which are themselves counter-examples to the Babai-Frankl conjecture,
is 31. Does there exist an example on Z
6
2
with smaller degree?
It is known that Z
5
2
is a CI-group. In other words, 6 is the smallest rank for
which an elementary 2-group is not a CI-group. Nowitzs example suggests that
there is a smallest rank r
p
for which the elementary p-group of rank r
p
is not a
CI-group, but all of those with rank smaller than r
p
are CI-groups. A recent result
of Muzychuk [42] shows that such a rank does exist.
Theorem 5.3 If p is a prime and n 2p 1 +
2p1
p
, then the elementary

p-group Z
n
p
is not a CI-group.
Note that the bound given in this theorem is tight for p = 2. This makes the
problem of determining the behaviour of the function r
p
tantalizing indeed. Also,
must a counter-example to the Babai-Frankl conjecture of minimal rank have the
propertythat the graphs andtheir complements are dense, as is the case for Nowitzs
example?
Another remarkable result about CI-groups is the following theorem of Li [29].
Theorem 5.4 If G is a nite CI-group, then G is solvable.
It is feasible to determine the CI-groups completely because the list of potential
CI-groups is known. However, given the long struggle to specify which cyclic
groups are CI-groups, this is likely to be difcult. Muzychuk completed the de-
termination of the cyclic CI-groups in two excellent papers [40], [41]. His results
may be stated as follows.
Theorem 5.5 The cyclic group Z
n
is a CI-group if and only n = 8, 9, 18 or 2
e
m,
where m is odd and square-free and e = 0, 1 or 2.
It may be the case that a group G is not a CI-group, but many of the Cayley
graphs on G behave nicely with regard to isomorphism. Accordingly, we say that
a particular Cayley graph Cay(G; S) is a CI-graph if Cay(G; S)

= Cay(G; S
)
implies that there exists an Aut(G) satisfying S
= S. Thus, a group G is a
CI-group if and only if every Cayley graph on G is a CI-graph.
6 Cayley graphs 167
A useful criterion for deciding whether or not a Cayley graph is a CI-graph is
provided by the following theorem of Babai [6].
Theorem 5.6 Let Gbe a Cayley graph on the nite group G. Then Gis a CI-graph
if and only if all regular subgroups of Aut(G) isomorphic to G are conjugate in
Aut(G).
There are many other questions that we can ask about CI-graphs. For example,
we can ask about degree conditions that guarantee that Cayley graphs on some
group are CI-graphs. We can ask about conditions on the automorphismgroups that
guarantee that Cayley graphs are CI-graphs. We can ask about various restrictions
on the connection sets that guarantee that the corresponding Cayley graphs are
CI-graphs. One such restriction is a conjecture of Toida that any circulant graph of
order n whose connection set is contained in the group of units of Z
n
is a CI-graph.
This conjecture has been proved recently (see [16]).
The last isomorphismtopic we discuss is when a given graph can be represented
as a Cayley graph in more than one way. As mentioned earlier, the complete graph
K
n
can be represented as a Cayley graph on any group of order n. Similarly, the
complete multipartite graph K
r(s)
can be represented as a Cayley graph on any
group of order rs that possesses a subgroup of order s.
There are more interesting results than the two simple examples above. We say
that two Cayley graph representations Cay(G
1
, S
1
) and Cay(G
2
, S
2
) are equivalent
when there exists an isomorphism : G
1
G
2
for which S
1
= S
2
. There are
then four non-equivalent representations of the 3-dimensional cube Q
3
as Cayley
graphs: one representation is on Z
3
2
, one is on Z
2
Z
4
, and two are on the dihedral
group of order 8. Moreover, for k = 4, 5 and 6, Dixon [14] has proved that there
are 14, 45 and 238 non-equivalent representations (respectively) of Q
k
.
Morris [39] has determined when circulants of odd prime power order have
representations as Cayley graphs on other Abelian p-groups.
6. Enumeration
Turners motivation for proving Theorem5.2 was his interest in trying to enumerate
the vertex-transitive graphs of order n. His theoremsets the stage for an elementary
enumeration of the vertex-transitive graphs of prime order. Since no primes belong
to NC, every vertex-transitive graph of prime order is a Cayley graph. On the other
hand, for each prime p, the group Z
p
is the unique group of order p. Thus, all
vertex-transitive graphs of prime order are circulant graphs.
Using Theorem 5.2, we see that the isomorphism of circulant graphs of prime
order is completely described by the action of a group on the set of all possible
168 Brian Alspach
connection sets. However, the latter action is induced by the action of a group on
the set {1, 2, . . . , p 1}, and this is the canonical setting for the employment of
P olyas famous enumeration theorem. Since i belongs to a connection set if and
only if p i also belongs to the connection set, and since multiplication by a and
a are the same, the action we have is the cyclic group of order ( p 1)/2 on a
set with ( p 1)/2 elements. Since {i, p i } either belongs, or does not belong,
to a connection set, we obtain the following result of Turner [52].
Theorem 6.1 If p is an odd prime, then the number of isomorphism classes of
vertex-transitive graphs of order p is
2
p 1
d
(d)2
( p1)/2d
,
where the summation runs over all divisors d of ( p 1)/2.
One of the motivations for classifying CI-groups is to extend Theorem 6.1 to
other families of Cayley graphs. Whenever we have a CI-group G, we may employ
P olyas theorem to enumerate the Cayley graphs on G. In order to do this, we have
to determine the so-called cycle index of Aut(G) acting on the set of unordered
pairs {g, g
1
} of non-identity elements of G.
Some work in this direction has been carried out by Alspach and Mishna [5].
Theorem 5.5 classies the cyclic CI-groups and the circulant graphs of the appro-
priate orders are enumerated in [5]. The Cayley graphs on the elementary Abelian
p-groups Z
2
p
have also been enumerated.
Even though the cyclic group of order p
2
is not a CI-group when p is a prime
bigger than 3, the circulant graphs of order p
2
have been enumerated (see [28]).
7. Automorphisms
The problem of determining the full automorphism group of a Cayley graph is
difcult in general. We rst examine the special case of prime order circulants,
where the answer is completely known. The tool that plays an essential role is
Theorem 5.1.
Suppose that p is a prime, and that we are given the circulant graph G =
Circ( p; S). It is nowmore convenient to think of circulant graphs as Cayley graphs
on the additive group Z
p
. The graph is either the complete graph or its complement
if and only if S is all of Z
p
or , respectively. The resulting automorphism group
is S
p
, and this situation is easy to recognize.
When S Z
p
, Theorem 5.1 tells us that Aut(G) has the form {T
a,b
:
a H < Z
p
, b Z
p
}. This implies that the stabilizer of the vertex labelled 0
6 Cayley graphs 169
is T
a,0
, with a H < Z
p
. Thus, if there is an edge joining 0 and k in G, then there
is an edge joining 0 and all of kH, and so the connection set S is a union of cosets
of the multiplicative subgroup Hof Z
p
. On the other hand, if S is a union of cosets
of the subgroup H of Z
p
, but not a union of cosets of any supergroup of H, then
the stabilizer of 0 is {T
a,0
: a H < Z
p
} and we know precisely what Aut(G) is.
If G = Circ( p; S) and S Z
p
, then let e(S) denote the maximum even
order subgroup H of Z
p
for which S is a union of cosets of H. Note that this is
dened because i S if and only if p i S.
The above comments provide an outline of a proof of the following result
in [3].
Theorem 7.1 Let G = Circ( p; S) be a circulant graph of prime order. If S = or
Z
p
, then Aut(G) = S
p
. Otherwise, Aut(G) {T
a,b
: a e(S), b Z
p
}.
We know that the automorphism group of any Cayley graph on a nite group G
contains the left regular representation of G. From the well-known orbit-stabilizer
theorem for permutation groups, we know that the order of the automorphism
group of a Cayley graph on G is the product of |G| and the order of the stabilizer of
any vertex. What makes Theorem 7.1 work is the fact that Theorem 5.1 provides
an exact description of possible stabilizers for circulant graphs of prime order.
Anatural situation to consider is when the stabilizer of any vertex is the identity,
implying that the full automorphismgroup of the Cayley graph on G is the left reg-
ular representation of G. In this case, we say that Cay(G, S) is a graphical regular
representation of G. The rst problem arising in this context is the determination
of the groups that admit a graphical regular representation.
It was known that Abelian groups of exponent greater than 2 and generalized
dicyclic groups do not admit graphical regular representations. Hetzel [24] and
Godsil [19] then independently completed the classication by showing that there
are only nitely many additional groups that do not admit graphical regular rep-
resentations. These additional groups are all of order 32 or less.
A Cayley graph G on a group G is normal when G
R
is a normal subgroup of
Aut(G). If S is the connection set for G and Aut(G)
{s}
= { Aut(G) : S = S},
then Aut(G) = G
R
Aut(G)
{s}
when G is normal. This means that the automorphism
group of G is as small as it can be, because the latter product is always a subgroup
of Aut(G). This is a new area for investigation.
8. Subgraphs
We nowconsider the murky area of the interaction between subgraph structures and
symmetry. There are some problems for which vertex-transitivity has a signicant
170 Brian Alspach
impact, and there are others for which it is unclear what the impact is. We rst
look at an example of the former.
Recall that a 1-factor of a graph G is a spanning subgraph in which each vertex
has degree 1. Since all vertices of a vertex-transitive graph G are the same under
the automorphismgroup, they must all lie in the same class of the Gallai-Edmonds
decomposition of G. The next theorem then follows easily.
Theorem 8.1 Let G be a connected vertex-transitive graph. If G has even order,
then it has a 1-factor. If Ghas odd order, then G v has a 1-factor, for each vertex
v V(G).
If G is a connected vertex-transitive graph of degree d, then G is d-edge-
connected. Thus, if G is a connected vertex-transitive graph of degree d, then for
any two distinct vertices v and w of G, there are d edge-disjoint paths whose
terminal vertices are v and w.
The situation for vertex-connectivity is not as straightforward as for edge-
connectivity. Watkins [53] proved the following theorem.
Theorem 8.2 If G is a connected vertex-transitive regular graph of degree d, and
if (G) denotes the connectivity of G, then (G) > 2d/3. Also, for each > 0,
there exists a connected d-regular vertex-transitive graph H for which
(H) <
+
2
3
d.
Even more interesting than the above lower bound on the connectivity as
a proportion of the degree are the methods introduced in [53] (and indepen-
dently in [34]) to prove the result. For each cutset of cardinality (G), we make
a list of the components that result when we delete the cutset. Any resulting
component of minimum cardinality, taken over all cutsets of cardinality (G),
is called an atom. Mader and Watkins have developed some interesting results
for the atoms in any connected vertex-transitive graph whose degree exceeds its
connectivity.
Probably the most important feature possessed by the atoms is that they form
a system of blocks of imprimitivity for the automorphism group. This lies at the
heart of the work on atoms.
We see from Theorem 8.2 that there are connected vertex-transitive graphs
whose connectivity is close to two-thirds of its degree. Is this still the case for
Cayley graphs? Consider the circulant graph G of order 5d whose connection
set S comprises all integers from {1, 2, . . . , 5d 1} that are congruent to 0, 1 or
6 Cayley graphs 171
2 (modulo 5). The circulant graph G has connectivity 2d and degree 3d 1. Thus,
the ratio is as close to
2
3
as we wish, and there is no improvement in Theorem 8.2
by restricting ourselves to Cayley graphs.
There has been some work on those Cayley graphs whose connectivity is max-
imum. A connection set S is quasi-minimal if the elements of S can be ordered as
s
1
, s
2
, . . . , s
t
, in such a way that
r
if |s
i
| > 2, then s
1
i
is either s
i 1
or s
i +1
;
r
if S
i
is the set {s
1
, s
2
, . . . , s
i
}, then, for each i such that |s
i
| = 2, S
i
is a
proper supergroup of S
i 1
, and for each i such that |s
i
| > 2 and s
1
i
= s
i 1
,
S
i
is a proper supergroup of S
i 2
.
The next theorem is proved using atoms in [4].
Theorem 8.3 If S is a quasi-minimal generating set of the group G, then the
Cayley graph Cay(G; S) has connectivity |S|.
In [10] it is shown that a connected circulant regular graph of degree at least
3 contains cycles of all possible even lengths 4, 6, . . . . In addition, if a circulant
graph has girth 3, then there are cycles of all possible lengths.
There have been studies regarding universality (see [11] and [22]). If H is a
graph of order n, then for any group G of order at least cn
2
, there is a Cayley graph
G on G for which H is an induced subgraph of G. In a different direction, for each
positive integer r, there exists a Paley graph that contains all graphs of order r as
induced subgraphs.
9. Hamiltonicity
Hamiltonicity refers to graph properties that are related to Hamilton cycles and
Hamilton paths, such as the existence and enumeration of Hamilton paths and cy-
cles, whether the graph is Hamilton-connected or Hamilton-laceable, the existence
of Hamilton decompositions, and the structure of the Hamilton space.
The interest in Hamilton cycles in Cayley graphs grewout of a question posed by
Lov asz. He asked whether every connected vertex-transitive graph has a Hamilton
path. In fact, there are only four known non-trivial connected vertex-transitive
graphs that do not possess Hamilton cycles. None of these is a Cayley graph,
which naturally leads one to ask whether every connected Cayley graph has a
Hamilton cycle. This is the rst topic we address.
The following result of Chen and Quimpo [12] is one of the best results on the
topic. We need two denitions before stating it. A graph G is Hamilton-connected
if, for any two distinct vertices v, w in G, there is a Hamilton path whose terminal
172 Brian Alspach
vertices are v and w. Since a vertex-transitive bipartite graph cannot be
Hamilton-connected, we change the denition for bipartite graphs accordingly.
A bipartite graph, with parts A and B satisfying | A| = |B|, is Hamilton-laceable
if, for any vertices v A and w B, there is a Hamilton path whose terminal
vertices are v and w.
Theorem 9.1 Let G be a connected Cayley graph on a nite Abelian group. If
G is bipartite and has degree at least 3, then G is Hamilton-laceable. If G is not
bipartite and has degree at least 3, then G is Hamilton-connected.
One immediate corollary of the Chen-Quimpo theorem is the fact that each
edge of a connected Cayley graph on an Abelian group of order 3 or more lies in
a Hamilton cycle. Theorem 9.1 has been extended to Hamiltonian groups.
This theorem also provides a useful tool for establishing the existence of
Hamilton cycles in Cayley graphs with special structural properties. For exam-
ple, there are situations when the vertex-set can be partitioned in such a way that
the subgraphs induced on the parts are Cayley graphs on Abelian groups, even
though the entire graph is not. Theorem 9.1 gives us a lot of freedom in nding
Hamilton paths in the subgraphs. The problem then becomes one of linking the
paths to form a Hamilton cycle in the original graph.
There also has been some success in looking for Hamilton cycles in Cayley
graphs by considering the order. One general question that arose was whether
every connected Cayley graph of prime power order has a Hamilton cycle. Witte
[55] provided a dramatic answer by proving the corresponding much stronger
result for Cayley digraphs.
Theorem 9.2 Every connected Cayley digraph on a group of order p
e
, where p is
a prime and e 1, has a Hamilton directed cycle.
There have been many particular Cayley graphs that have been shown to have
Hamilton cycles. There is insufcient space to include most of them, and it is not
clear that these special results give much of an indication towards resolving the
general problemanyway. In fact, there is so little current evidence that conjecturing
one way or the other about the general problem seems no better than guessing. We
include one special case, however, because the proof of the existence of a Hamilton
cycle in this case is difcult (see [13]).
Theorem 9.3 The Cayley graph on the symmetric group S
n
(n 3) with connec-
tion set {(1 2), (1 2 n), (n n 1 1)} has a Hamilton cycle.
6 Cayley graphs 173
10. Factorization
We saw earlier that every connected Cayley graph of even order has a 1-factor. In
fact, something much stronger may be the case. A 1-factorization of a graph is a
partition of the edge-set into 1-factors. No example is known of a connected Cayley
graph of even order that is not 1-factorizable. We now present some results about
1-factorizable Cayley graphs, starting with one of Stong [50].
Theorem 10.1 A connected Cayley graph on the group G has a 1-factorization if
one of the following holds:
r
|G| = 2
k
, for some integer k;
r
G is an Abelian group of even order;
r
G is dihedral or dicyclic.
We say that the connection set S is a minimal generating Cayley set for G, if
S generates G, but S {s, s
1
} generates a proper subgroup of G for each s S.
In addition to proving Theorem 10.1, Stong [50] proved that Cayley graphs whose
connection sets are minimal generating Cayley sets have 1-factorizations whenever
the group is a nilpotent group of even order, or the group contains a proper Abelian
normal subgroup of index 2
k
, or the group has order 2
m
p
k
for some prime p
satisfying p > 2
m
.
Cubic graphs are of interest in their own right. If we seek Cayley graphs without
Hamilton cycles, then it seems natural to believe that sparser graphs are less likely
to have a Hamilton cycle. Any cubic graph with a Hamilton cycle clearly has a
1-factorization, but the converse is not true in general. So checking cubic Cayley
graphs for 1-factorizations should be easier than checking them for Hamilton
cycles. Recently, Poto cnik has shown that any cubic Cayley graph G whose auto-
morphism group has a soluble subgroup that acts transitively on the vertex-set of
G has a 1-factorization. In particular, this means that any connected cubic Cayley
graph on a soluble group of even order has a 1-factorization.
A Hamilton decomposition of a graph G is a partition of the edge-set into
Hamilton cycles when the degree is even, or a partition into Hamilton cycles
and a 1-factor when the degree is odd. Many examples of graphs with Hamilton
decompositions are representable as Cayley graphs on Abelian groups. This leads
naturally to ask whether all connected Cayley graphs on Abelian groups have
Hamilton decompositions.
Theorem 9.1 gives an afrmative answer for graphs of degree 2 or 3. The rst
progress came in [8], where it was proved that the answer is yes for degree 4. It
is then easy to show that the answer is yes for degree 5, and most of the cases
for degree 6 are settled. The best result on Hamilton decompositions of Cayley
174 Brian Alspach
graphs on Abelian groups is the next theorem, obtained by Liu [31], [32]. He has
also established Hamilton decompositions for certain special cases.
Theorem 10.2 If G = Cay(G, S) is a connected Cayley graph on an Abelian
group G, and if S is a minimal generating Cayley set, then G has a Hamilton
decomposition.
An isomorphic factorization of a graph G is a partition of the edge-set E of G for
which the subgraphs induced by the edges in each part are isomorphic. In general,
given a graph G and a divisor d of |E|, we are interested in whether there is an
isomorphic factorization of G into d subgraphs. Some early work on isomorphic
factorizations dealt with specic graphs that happen to be Cayley graphs. This leads
naturally to the following question: if d is a divisor of |E| and G is a circulant
graph, a Cayley graph, or a vertex-transitive graph, does there exist an isomorphic
factorization of G into d subgraphs?
One contribution towards these questions is the following result of Fink [17].
Theorem 10.3 If T is any tree with k edges, then the k-cube Q
k
has an isomorphic
factorization by T. Furthermore, there is an isomorphic factorization for which
each copy of T is an induced subgraph.
11. Embeddings
There are a long history and an extensive literature on embedding graphs in sur-
faces. The book of White [54] and a recent excellent survey [47] provide a good
starting point for this topic.
Cayley graphs arise naturally in the context of graph embeddings. Cayley dia-
grams arose more than a hundred years ago out of an interest in trying to nd a
way to picture groups. These are Cayley digraphs in which the arcs are coloured
according to the group element that generates the arc. When we try to draw the
diagrams with no arcs crossing, the colours of the arcs are immaterial, as is whether
there are arcs in both directions between two vertices. Thus, we dene the genus
of a group G as the smallest genus g with the property that there exists a connected
Cayley graph G on G which has an embedding into an orientable surface of genus g.
As early as 1896, planar groups were characterized by Maschke [36]. The
toroidal groups were classied into seventeen innite families, together with some
sporadic cases, by Proulx [46]. Tucker [51] proved that, for any genus g 2, there
are only nitely many groups of genus g. The genus of most Abelian groups was
found by Jungerman and White [27].
A Cayley map is an embedding of a Cayley graph into an orientable sur-
face with the same cyclic rotation scheme at each vertex. This term is relatively
6 Cayley graphs 175
recent apparently rst appearing in [9] but the idea has been around for more
than one hundred years. Cayley maps play a signicant role in the proof of the
Heawood map colouring problem. They also play a role in the determination of
the genus of a group. The long paper [47] unies some earlier scattered work on
Cayley maps, and provides a general framework for studying them.
12. Applications
There are many meaningful applications of Cayley graphs, and we conclude with a
quick outline of a few of them. Any problem for which graphs are used as a model
and the use of edges is being optimized provides a natural setting for Cayley graphs.
For example, suppose that we are to construct a network for which the number
of direct links we may use is restricted, but we want to maximize the probability
that the network remains connected after some links or vertices of the network are
deleted. Then there is a strong tendency for the graph to be either vertex-transitive
or close to it. Consequently, Cayley graphs in particular, the k-dimensional
cube Q
k
have been extensively studied by researchers working with networks.
A recent book on this topic is by Xu [56], and a fundamental paper is [1].
Let C be a circle with circumference r. An r-circular colouring of a graph G is
a mapping c that assigns to each vertex v of G an open arc vc of unit length on C,
with the property that vc wc = whenever v and w are adjacent vertices of G.
The circular chromatic number of G is inf{r : G is r-circular colourable}. Circulant
graphs have played a role in the development of circular chromatic numbers of
graphs. For a general survey of circular chromatic numbers, see [57].
If G is a graph, and if A is a subset of V(G), then let N(A) = {v V(G) A: v
is adjacent to some vertex of A}. The graph G is an (n, d, c)-expander if G has order
n, maximum degree d, and there exists a constant c > 0 such that |N(A)| c| A|,
for all A V(G) satisfying | A| n/2. The construction of expanders and families
of expanders with c xed is of considerable interest, and Cayley graphs play a
signicant role; two excellent references are [2] and [33].
Acknowledgements: I wish to thank Luis Goddyn, Cai Heng Li, Brendan McKay
and Mikhail Muzychuk for valuable conversations.
References
1. S. Akers and B. Krishnamurthy, Agroup-theoretic model for symmetric interconnection
networks, IEEE Trans. Comput. 38 (1989), 555566.
2. N. Alon, Tools from higher algebra, Handbook of Combinatorics Vol II (eds. R. L.
Graham, M. Gr otschel and L. Lov asz), MIT Press and North-Holland (1995), 1749
1783.
176 Brian Alspach
3. B. Alspach, Point-symmetric graphs and digraphs of prime order and transitive permu-
tation groups of prime degree, J. Combin. Theory (B) 15 (1973), 1217.
4. B. Alspach, Cayley graphs with optimal fault tolerance, IEEE Trans. Computers 41
(1992), 13371339.
5. B. Alspach and M. Mishna, Enumeration of Cayley graphs and digraphs, Discrete Math.
256 (2002), 527539.
6. L. Babai, Isomorphism problem for a class of point-symmetric structures, Acta Math.
Acad. Sci. Hungar. 29 (1977), 329336.
7. L. Babai and P. Frankl, Isomorphisms of Cayley graphs I, Combinatorics (Proc. Fifth
Hungarian Colloq., Keszthely, 1976), Colloq. Math. Soc. J anos Bolyai 18, North-
Holland (1978), 3552.
8. J.-C. Bermond, O. Favaron and M. Maheo, Hamiltonian decompositions of Cayley
graphs of degree 4, J. Combin. Theory (B) 46 (1989), 142153.
9. N. Biggs, Cayley maps and symmetrical maps, Proc. Cambridge Philos. Soc. 72 (1972),
381386.
10. Z. Bogdanowicz, Pancyclicity of connected circulant graphs, J. Graph Theory 22
(1996), 167174.
11. B. Bollob as and A. Thomason, Graphs which contain all small graphs, Europ. J. Com-
bin. 2 (1981), 1315.
12. C. C. Chen and N. Quimpo, On strongly hamiltonian abelian group graphs, Combina-
torial Mathematics VIII, Lecture Notes in Mathematics 884, Springer-Verlag (1981),
2334.
13. R. Compton and S. G. Williamson, Doubly adjacent Gray codes for the symmetric
group, Linear and Multilinear Algebra 35 (1993), 237293.
14. J. D. Dixon, Groups with a Cayley graph isomorphic to a hypercube, Bull. Austral.
Math. Soc. 55 (1997), 385393.
15. E. Dobson, Isomorphism problem for Cayley graphs of Z
3
p
, Discrete Math. 147 (1995),
8794.
16. E. Dobson and J. Morris, Toidas Conjecture is true, Electron. J. Combin. 9 (2002),
R35.
17. J. Fink, On the decomposition of n-cubes into isomorphic trees, J. Graph Theory 14
(1990), 405411.
18. G. Gamble and C. E. Praeger, Vertex-primitive groups and graphs of order twice the
product of two distinct odd primes, J. Group Theory 3 (2000), 247269.
19. C. D. Godsil, Graphs with Regular Groups, Ph.D. thesis, University of Melbourne,
1979.
20. C. D. Godsil, More odd graph theory, Discrete Math. 32 (1980), 205207.
21. C. D. Godsil, On Cayley graph isomorphisms, Ars Combin. 15 (1983), 231246.
22. C. D. Godsil and W. Imrich, Embedding graphs in Cayley graphs, Graphs Combin. 3
(1987), 3943.
23. A. Hassani, M. Iranmanesh and C. E. Praeger, On vertex-imprimitive graphs of order a
product of three distinct primes, J. Combin. Math. Combin. Comput. 28 (1998), 187
213.
24. D. Hetzel,

Uber regul are graphische Darstellungen von au osbaren Gruppen, Ph.D.
thesis, Technische Universit at Berlin, 1976.
25. M. Hirasaka and M. Muzychuk, An elementary abelian group of rank 4 is a CI-group,
J. Combin. Theory (A) 94 (2001), 339362.
6 Cayley graphs 177
26. M. Iranmanesh and C. E. Praeger, On non-Cayley vertex-transitive graphs of order a
product of three primes, J. Combin. Theory (B) 81 (2001), 119.
27. M. Jungerman and A. White, On the genus of nite Abelian groups, Europ. J. Combin.
1 (1980), 243251.
28. M. Klin, V. Liskovets and R. P oschel, Analytical enumeration of circulant graphs with
prime-squared number of vertices, S em. Lothar. Combin. 36 (1996), B36d, 36 pp.
(electronic).
29. C. H. Li, Finite CI-groups are soluble, Bull. London Math. Soc. 31 (1999), 419423.
30. C. H. Li, On isomorphisms of nite Cayley graphs a survey, Discrete Math. 256
(2002), 301334.
31. J. Liu, Hamiltonian decompositions of Cayley graphs on abelian groups of odd order,
J. Combin. Theory (B) 66 (1996), 7586.
32. J. Liu, Hamiltonian decompositions of Cayley graphs on abelian groups of even order,
J. Combin. Theory (B) 88 (2003), 305321.
33. A. Lubotzky, Cayley Graphs: Eigenvalues, Expanders and Random Walks, London
Math. Soc. Lecture Notes 218 (1995), 155189.
34. W. Mader, Eine Eigenschaft der Atome endlicher Graphen, Arch. Math. (Basel) 22
(1971), 333336.
35. D. Maru si c, Vertex-transitive graphs and digraphs of order p
k
, Ann. Discrete Math. 27
(1985), 115128.
36. W. Maschke, The representation of nite groups, Amer. J. Math. 18 (1896), 156194.
37. B. D. McKay and C. E. Praeger, Vertex-transitive graphs which are not Cayley graphs,
I, J. Graph. Theory 22 (1996), 321334.
38. A. Miller and C. E. Praeger, Non-Cayley vertex-transitive graphs of order twice the
product of two odd primes, J. Algebraic Combin. 3 (1994), 77111.
39. J. Morris, Isomorphic Cayley graphs on nonisomorphic groups, J. Graph Theory 31
(1999), 345362.
40. M. Muzychuk,

Ad ams conjecture is true in the square-free case, J. Combin. Theory
(A) 72 (1995), 118134.
41. M. Muzychuk, On

Ad ams conjecture for circulant graphs, Discrete Math. 167 (1997),
497510.
42. M. E. Muzychuk, An elementary abelian group of large rank is not a CI-group, Discrete
Math. 264 (13) (2003), 167186.
43. M. Muzychuk and G. Tinhofer, Recognizing circulant graphs in polynomial time: An
application of association schemes, Electron. J. Combin. 8 (2001), R26.
44. L. Nowitz, A non-Cayley-invariant Cayley graph of the elementary Abelian group of
order 64, Discrete Math. 110 (1992), 223228.
45. P. Poto cnik, Edge-colourings of cubic graphs admitting a solvable vertex-transitive
group of automorphisms, preprint.
46. V. Proulx, Classication of the toroidal groups, J. Graph Theory 2 (1978), 269273.
47. R. B. Richter, J.

Sir a n, R. Jajcay, T. Tucker and M. Watkins, Cayley maps, preprint.
48. G. Sabidussi, On a class of xed-point-free graphs, Proc. Amer. Math. Soc. 9 (1958),
800804.
49.

A. Seress, On vertex-transitive, non-Cayley graphs of order pqr, Discrete Math. 182
(1998), 279292.
50. R. Stong, On 1-factorizability of Cayley graphs, J. Combin. Theory (B) 39 (1985),
298307.
178 Brian Alspach
51. T. Tucker, The number of groups of a given genus, Trans. Amer. Math. Soc. 258 (1980),
167179.
52. J. Turner, Point-symmetric graphs with a prime number of points, J. Combin. Theory 3
(1967), 136145.
53. M. Watkins, Connectivity of transitive graphs, J. Combin. Theory 8 (1970), 2329.
54. A. White, Graphs of Groups on Surfaces, Mathematics Studies 188, North-Holland,
2001.
55. D. Witte, Cayley digraphs of prime-power order are hamiltonian, J. Combin. Theory
(B) 40 (1984), 107112.
56. J. Xu, Topological Structure and Analysis of Interconnection Networks, Kluwer, 2001.
57. X. Zhu, Circular chromatic number: a survey, Discrete Math. 229 (2001), 371410.
7
Finite symmetric graphs
CHERYL E. PRAEGER
1. Introduction
2. s-arc transitive graphs
3. Group-theoretic constructions
4. Quotient graphs and primitivity
5. Distance-transitive graphs
6. Local characterizations
7. Normal quotients
8. Finding automorphism groups
9. A geometric approach
10. Related families of graphs
References
The study of nite symmetric graphs goes back to work of Tutte dating from
1947. Because of their transitivity properties, symmetric graphs are regular,
their connected components are isomorphic to one another, and they have
other regularity in their structure. They are thus useful in many applica-
tions. This chapter traces two broad approaches to the study of symmetric
graphs, a local analysis focusing on the structure of a vertex stabilizer, and
a global approach that gives information about the full group of automor-
phisms. These different approaches are explained with reference to distance-
transitive graphs, s-arc transitive graphs, and other families of symmetric
graphs.
1. Introduction
A graph G is symmetric if its automorphism group Aut(G) is transitive on vertices
and on ordered pairs of adjacent vertices that is, on the arcs of G; symmetric
179
180 Cheryl E. Praeger
graphs are also called arc-transitive graphs. They have been studied for over fty
years, beginning with two important papers by Tutte on nite cubic symmetric
graphs in 1947 and 1959, and a paper by Sabidussi in 1964 which suggested a
method for constructing all symmetric graphs from a small amount of information
about their automorphism groups (see [32, Secs. 9 and 11]). As a consequence of
their transitivity properties, symmetric graphs are regular, their connected com-
ponents are isomorphic to one another, and they have other regularity in their
structure. Perhaps for these reasons, they are useful in many applications, such as
modelling interconnection networks (see [20]).
The theory of nite symmetric graphs has two basic threads, one focusing on
a broad global analysis of classes of groups or graphs, and the other concen-
trating on tight local analysis of specic combinatorial or group-theoretic situ-
ations. The theoretical frameworks adopted, as well as the statements of results
and the methodology, all involve a combination of combinatorics and abstract
group theory, often requiring an application of the nite simple group classi-
cation. The purpose of this chapter is to trace these developmental threads for
nite symmetric graphs, in order to understand the power of each, and to appre-
ciate the ways in which they intertwine in elucidating structure, and in charac-
terizing classes of these graphs. In Section 3 we discuss the fundamental links
between graph theory and group theory which underpin the theory of symmetric
graphs.
Alocal analysis of a symmetric graph is a study of the combinatorial structure
or symmetry properties close to a vertex. Sometimes local properties, either
combinatorial or group-theoretic, inuence the structure of the whole graph, and
a major problem is to determine the extent to which this occurs. Tuttes theorem,
discussed in Section 2, provided the rst successful local analysis of a family of
symmetric graphs and was enormously inuential on subsequent research. Further
examples of local analyses of symmetric graphs are given in Section 6.
In contrast to this, global analysis of symmetric graphs focuses on aspects of
graph structure that affect the whole graph. The global approach leads naturally to
a consideration of quotients of symmetric graphs, and thence to a study of symmet-
ric graphs admitting a vertex-primitive or vertex-biprimitive automorphism group
(see Section 4). The most complete manifestation of this approach has been the
classication programme for nite distance-transitive graphs, described briey in
Section5andat greater depthinChapter 9. However, the methodof passingtoprim-
itive quotients is unsatisfactory for studying many families of symmetric graphs
that are given by a local dening property, such as s-arc transitivity. For several
of these families, the appropriate quotients to study are normal quotients, and it is
essential for an understanding of these families to look closely at the quasiprimitive
and bi-quasiprimitive graphs in the family. This approach is illustrated for 2-arc
transitive graphs in Section 7.
7 Finite symmetric graphs 181
For yet other families of symmetric graphs, information about quotients
(whether primitive or normal) is insufcient to describe the overall structure of
typical graphs in the family. A geometric approach that injects additional combi-
natorial information, and offers a framework for studying imprimitive symmetric
graphs, is discussed in Section 9.
Most investigations of symmetric graphs begin with a given graph G and a given
group G acting symmetrically (that is, vertex-transitively and arc-transitively) on
G, and we call (G, G) a symmetric pair. The difcult problem of determining the
full automorphism group of G, or indeed any subgroup of Aut(G) containing G,
often needs to be addressed. Some effective strategies for this are discussed for
primitive and quasiprimitive symmetric graphs in Section 8. Finally, in Section
10, we mention some classes of vertex-transitive graphs that are close to being
symmetric.
2. s-arc transitive graphs
In this section, by way of introduction to the subject, we describe Tuttes seminal
discoveries about the structure of cubic s-arc transitive graphs. For a positive
integer s, an s-arc in a graph G is an (s 1)-tuple (v
0
, v
1
, . . . , v
s
) of vertices
with the properties that v
i 1
v
i
is an edge of G, for 1 i s, and v
i 1
,= v
i 1
for
1 i s 1. The graph G is s-arc transitive if Aut(G) is transitive on s-arcs. If
G is a subgroup of Aut(G), then (G, G) is an s-arc transitive pair if G is transitive
on both vertices and s-arcs of G. Thus, the 1-arc transitive pairs are precisely the
symmetric pairs, and indeed all s-arc transitive pairs are symmetric pairs.
Lemma 2.1 Let Gbe a graph, let G be a subgroup of Aut(G), and let s be a positive
integer. If (G, G) is an s-arc transitive pair, then (G, G) is a symmetric pair.
Proof Let (G, G) be an s-arc transitive pair. Then G is vertex-transitive, by de-
nition, so G is regular of degree k, say. If k = 0, then G has no arcs, and if k = 1,
then each component of G is a complete graph K
2
: in these cases, (G, G) is triv-
ially a symmetric pair. So let k 2. If s > 1, then each 1-arc (v
0
, v
1
) of G can be
extended to an s-arc (v
0
, v
1
, . . . , v
s
). Since G is transitive on s-arcs, it follows that
G is transitive on 1-arcs. Thus, for every s, (G, G) is a 1-arc transitive pair, and as
observed above, the 1-arc transitive pairs are precisely the symmetric pairs.
The property of being s-arc transitive is of special interest for s 2. Many
families of examples are known, and we list a few of them below. The vertices of
the odd graph O
n
are the (n 1)-element subsets of {1, 2, . . . , 2n 1], with two
vertices being adjacent if and only if they are disjoint; O
n
is regular of degree n,
and Aut(O
n
) = S
2n1
(see Chapter 5 and [7, Sec. 9.1]).
Theorem 2.2
(a) For n 3 and s 1, the cycle C
n
is s-arc transitive;
(b) For n 4, the complete graph K
n
is 2-arc transitive, but not 3-arc transitive;
(c) For n 3, the complete bipartite graph K
n,n
is 3-arc transitive, but not 4-arc
transitive;
(d) For an odd integer n 3, O
n
is 3-arc transitive, but not 4-arc transitive.
There are also examples of cubic graphs that are 4-arc transitive but not 5-arc
transitive (for example, the Coxeter graph on 28 vertices), and 5-arc transitive but
not 6-arc transitive (for example, the (3, 8)-cage on 30 vertices) see [7, Secs.
12.3 and 6.9].
Tutte proved that there are no 6-arc transitive cubic graphs (see [32, Thm. 9.2]).
His arguments were an elegant combination of group theory and combinatorial
reasoning.
Theorem 2.3 (Tuttes theorem) If G is an s-arc transitive cubic graph, then
s 5.
Given that a graph G is symmetric, the fact that it is s-arc transitive can be
decided by examining the stabilizer (Aut(G))
v
in Aut(G) of a vertex v; precisely,
we require that (Aut(G))
v
be transitive on the s-arcs with rst vertex v. Thus, s-arc
transitivity may be regarded as a local property. Tuttes result demonstrates that in
a symmetric graph a local property can sometimes control the global structure.
3. Group-theoretic constructions
In 1964, Sabidussi proved that each symmetric graph G is isomorphic to one of the
symmetric coset graphs in the Introduction (Section 3) namely, if G is a subgroup
of Aut(G) and G is G-symmetric, then G

= G(G, H, HgH), where H = G
v
for
some vertex v, and g G with vg v and g
1
HgH (see the Introduction,
Theorem 3.5).
Studying symmetric graphs as coset graphs has proved useful in classication
problems and focuses attention on a given subgroup of automorphisms. It thus
leads naturally to the study of symmetric pairs (G, G), consisting of a symmetric
graph G and a subgroup G of Aut(G) that acts symmetrically on G. Usually infor-
mation about symmetric pairs is required up to equivalence, where equivalence of
symmetric pairs is dened in the Introduction (Section 3), and we note that, for a
G-symmetric graph G, (G, G) is equivalent to (G(G, H, HgH), G).
Shortly after Sabidussis work, a seminal paper of Sims [35] gave a second
group-theoretic characterization of symmetric graphs in terms of orbital graphs.
For a transitive permutation group G and a non-diagonal self-paired orbital O, the
orbital graph G(O), dened in the Introduction (Section 3), is undirected and
is G-symmetric, so (G(O), G) is a symmetric pair. Moreover, up to equivalence,
each symmetric pair arises in this way. Also, a nite transitive permutation group
G has at least one non-diagonal self-paired orbital if and only if [G[ is even (see
[41, Thm. 16.5]).
Sims demonstrated a crucial connection between the primitivity of a transitive
group G and the connectivity of its orbital graphs, showing that G is primitive
if and only if all its non-diagonal orbital graphs are connected see the Intro-
duction, Section 3. However, as the following example demonstrates, if G is not
primitive, it is possible that some of its non-diagonal orbital graphs may still be
connected.
Example Let V = {1, 2, 3, 4, 5, 6] and G = (123456))

= Z
6
. There are ve non-
diagonal orbitals: if O is the orbital containing (1, 2) or (1, 6), then G(O) is
the directed 6-cycle; if O is the orbital containing (1, 3) or (1, 5), then G(O)
has two components, each of which is the directed 3-cycle; if O is the orbital
containing (1, 4), then G(O) has three components, each of which is the undirected
graph K
2
.
Imprimitive permutation groups
Let G be a permutation group on V. A partition P of V is G-invariant if the
elements of G permute the parts of P set-wise that is, pg P for all p P and
g G, where pg = {vg : v p]. The G-invariant partitions of V are in one-one
correspondence with the G-invariant equivalence relations on V namely, a G-
invariant partition P corresponds to the equivalence relation , where v w if
and only if v and w lie in the same part of P. In particular, the relation of equality
corresponds to the partition with all parts of size 1, and the universal relation
corresponds to the partition with a single part. These two partitions are called the
trivial partitions of V, and both are G-invariant for each permutation group G. A
transitive permutation group G is primitive if the only G-invariant partitions are the
trivial ones; otherwise G is imprimitive. This denition of a primitive permutation
group is equivalent to the one given in the Introduction (Section 3).
If P is invariant under a transitive permutation group G, then G induces a
transitive permutation group G
P
on P (by dening g : p pg), and the set-wise
stabilizer G
p
of p P induces a transitive permutation group G
p
p
on p. Moreover,
if p, q P, then the transitive groups G
p
p
and G
q
q
are equivalent. Thus, if the set
V is nite, then each non-trivial G-invariant partition P gives rise to two smaller
transitive permutation groups (up to equivalence) namely, G
P
and G
p
p
. We shall
be especially interested in the cases where one or other of these transitive groups
is as small as possible.
A G-invariant partition P is maximal if {V] is the only G-invariant partition of
V rened by P; in this case, the group G
P
is primitive. Similarly, P is minimal if its
only G-invariant renement is {{v] : v V]; in this case, the group G
p
p
is primitive.
We shall show in the next section that, if (G, G) is a symmetric pair and P is a
non-trivial G-invariant partition of the vertex-set, then the two smaller transitive
groups associated to P correspond to two smaller symmetric pairs. Thus, we shall
be especially interested in symmetric pairs (G, G) for which G is vertex-primitive.
The most useful tool for studying nite primitive permutation groups is the
ONan-Scott Theorem (see [9, Sec. 3]):
Theorem 3.1 (ONan-Scott theorem) Let G be a subgroup of Sym(V). If G is a
nite primitive group, then G belongs to one of several disjoint types of primitive
groups.
The number of types of nite primitive groups varies for different statements
of this theorem. A version that is often used for studying nite symmetric graphs
involves eight types of primitive groups, and each of these types corresponds to
an analogous type of quasiprimitive permutation group, where, as dened in the
Introduction (Section 3), G is quasiprimitive if all its non-trivial normal subgroups
are transitive. A brief description of the ONan-Scott types of primitive groups
(and their generalizations for quasiprimitive groups), suitable for applications to
symmetric graphs, is given in Theorem7.2 (see also [31, Sec. 5]). Many of the types
are described in terms of the structure and action of a minimal normal subgroup
of such a group G, or of its socle (the product of its minimal normal subgroups),
denoted by soc(G).
Two of the ONan-Scott types occur frequently as types of primitive automor-
phism groups of graphs namely, types AS and HA. A primitive or quasiprimitive
group G is said to have type AS, or to be almost simple, if G is a subgroup of
Aut(H), for some non-abelian simple subgroup H of G: here, soc(G) = H is the
unique minimal normal subgroup and is transitive. A primitive group G is said to
have type HA, or to be of afne type, if G has a non-trivial abelian normal subgroup:
here, G is a group of afne transformations of a nite vector space and soc(G) is its
unique minimal normal subgroup and is elementary abelian and regular (transitive
with trivial stabilizers); in fact, soc(G) is the group of translations.
Amalgams and their completions
A group amalgam for a graph aims to capture essential local information about
the symmetry properties of the graph. Several denitions appear in the literature,
and we follow the treatment in [21, Sec. 1.3]. A nite amalgam A of rank n is a
set consisting of n groups H
i
(1 i n) with the property that, for each i ,= j ,
the intersection H
i
H
j
is a subgroup of both H
i
and H
j
.
There is no assumption that the H
i
are subgroups of some common group.
Instead, a group G is called a completion of Aif there exists a map : H
i
G
for which the restriction of to each H
i
is a group homomorphism, and the image
of generates G. If is one-one, then the completion is faithful; for example, if
A is dened as a set of subgroups of a group, then the subgroup generated by the
subgroups in A is a faithful completion of A relative to the inclusion map. For
every amalgam A, there exists a universal completion U(A) (relative to a map )
with the property that, for any completion G (relative to a map ), there is a unique
homomorphism : U(A) G for which is the composition of and . The
universal completion U(A) is dened as the group with generating set H
i
and
with relations all the equalities of the form z = xy, satised by elements x, y and
z of some H
i
. Sometimes (see [21], for example) it is possible to nd not only
U(A), but also all nite completions of A.
For a symmetric pair (G, G) and an edge e = vw, the amalgam A = {G
v
, G
e
]
contains the information that G
v
is transitive on the set of vertices adjacent to v, and
that some element of G interchanges v and w. Thus, A contains the information
that G is symmetric on G, and the set-wise stabilizer in G of the component of G
containing v is a faithful completion of A. The universal completion U(A) acts
symmetrically on an innite regular tree T(A) with vertex-stabilizer and edge-
stabilizer giving the same amalgam A. Further, any symmetric pair (H, H) with
amalgamAis a quotient of (T(A), U(A)), relative to some graph homomorphism
T(A) H and group homomorphism U(A) H. There are innitely many
such pairs (H, H), even with H nite, and one is usually interested in symmetric
pairs with specic additional properties.
Amalgams of rank greater than 2 carry information about additional combi-
natorial structure of the graph, and are key tools for analysing certain classes of
symmetric graphs (see, for example, [21, Ch. 9]). Such amalgams have applica-
tions beyond graph theory, especially to the study of diagram geometries and of
groups that preserve them (see [21] and [29]).
For a symmetric pair (G, G) with G a cubic graph, and an edge e incident with
a vertex v, the structure of possible amalgams {G
v
, G
e
] was deduced from early
work of Tutte and Wong (see Biggs [3, Ch. 18]). This led to the determination
of presentations for universal completions of such pairs. According to Biggs [5],
John Conway discovered such presentations but did not publish them, and also
inaugurated a programme for constructing such [symmetric pairs] from the pre-
sentations; this programme was developed further by Biggs in [4]. Systematic
investigations of these presentations were undertaken in [14] and [40], and these
enabled symmetric pairs for cubic graphs of small girth to be classied (see [28]),
and also led to constructions of arbitrarily large 4-arc transitive and 5-arc transitive
nite cubic graphs (see [12] and [13]).
4. Quotient graphs and primitivity
For every graph G and every partition P of its vertex-set V, the quotient graph G
P
is the graph with vertex-set P for which p, p
/
P are joined by an edge if and
only if there exist v p and v
/
p
/
that are adjacent in G. Also, for each p P,
let G[ p] the induced subgraph of G on p. Here is a simple example.
Example Let G = C
6
on V = {1, 2, 3, 4, 5, 6] with edges 12, 23, . . . , 61, and
let P = { p
1
, p
2
, p
3
] where p
1
= {1, 4], p
2
= {2, 5], p
3
= {3, 6]. Then G
P

= C
3
and, for each i, G[ p
i
] is a null graph with two vertices.
In this example, P is invariant under G = Aut(G) = (123456), (26)(35))

= D
12
,
and we have G
P

= S
3
, and G
p
p

= S
2
, for each p P. All the pairs (G, G), (G
P
,
G
P
), (G[ p], G
p
p
) are symmetric pairs (the last somewhat vacuously, since G[ p]
has no edges). This is a rather general fact, as we see in the next result.
Theorem 4.1 Let (G, G) be a symmetric pair, let P be a G-invariant partition of
V, and let p, p
/
P. Then (G
P
, G
P
), (G[ p], G
p
p
) and (G[ p
/
], G
p
/
p
/
) are all sym-
metric pairs, and (G[ p], G
p
p
) and (G[ p
/
], G
p
/
p
/
) are equivalent.
Proof If ( p, p
/
) and (q, q
/
) are arcs of G
P
, then there exist v p, v
/
p
/
and
w q, w
/
q
/
for which (v, v
/
) and (w, w
/
) are arcs of G. The group G contains
an element that maps (v, v
/
) to (w, w
/
), and hence maps ( p, p
/
) to (q, q
/
). The rest
of the proof that (G
P
, G
P
) and (G[ p], G
p
p
) are symmetric pairs is straight-forward.
For the equivalence, if g G maps p to p
/
, then the restriction of g to p in-
duces an isomorphismfrom G[ p] to G[ p
/
], while the conjugation action of g indu-
ces a group isomorphism from G
p
p
to G
p
/
p
/
, and these together dene the required
equivalence.
If (G, G) is a symmetric pair and G is primitive on vertices, then we say that
(G, G) is a primitive symmetric pair. By choosing P to be a maximal G-invariant
partition, we can ensure that (G
P
, G
P
) is primitive. Similarly by choosing P to
be a minimal G-invariant partition, we can ensure that (G[ p], G
p
p
) is primitive.
However, when G is symmetric on G, either all edges have their ends in the same
block of P, or all edges have their ends in different blocks of P. Thus, one of G
P
and G[ p] is a null graph; in particular, if G is connected, then G[ p] is null.
Suppose nowthat G is connected, so that G[ p] is a null graph. Several strategies
have proved effective for investigating this situation. First, it is sometimes possible
to introduce some combinatorial structure on p, other than G[ p], to aid in the
understanding of (G, G). This was the case for distance-transitive graphs, which
will be discussed in Section 5, and is also an integral part of the geometric approach
to be described in Section 9.
Secondly, for some families of symmetric graphs, we can extract crucial infor-
mation about (G, G) from its primitive quotients (G
P
, G
P
), and in these cases it is
of fundamental importance to study primitive symmetric pairs in the family. In such
a study, the ONan-Scott theorem is essential: see the case of distance-transitive
graphs in Section 5, below.
In one rather trivial case, a primitive quotient gives little useful information
namely, if G is bipartite and P is the bipartition, then (G
P
, G
P
) is equivalent to
(K
2
, S
2
). In this case, it would be helpful to examine quotients (G
P
, G
P
) where the
only G-invariant partitions rened by P are {V] and the bipartition of V. We call
a permutation group with this property biprimitive. Possible structures for biprim-
itive groups can be inferred from the ONan-Scott theorem. However, satisfactory
analyses of families of biprimitive symmetric pairs seem difcult to achieve.
Often we wish to study a family of symmetric graphs with an additional dening
property, sometimes combinatorial and sometimes group-theoretic. Unfortunately,
the extra dening property is rarely inherited by a quotient graph G
P
, so studying
the vertex-primitive graphs in the family does not help us to understand the struc-
ture of typical graphs in the family; for example, if (G, G) is 2-arc transitive and P is
G-invariant, then (G
P
, G
P
) is not in general 2-arc transitive (see Section 7). How-
ever, some families of symmetric pairs (such as the family of 2-arc transitive pairs)
which are not closed under forming primitive quotients are closed under forming
normal quotients. This will be discussed further in Section 7.
5. Distance-transitive graphs
Let G be a connected graph of diameter d 1, and let G be a subgroup of Aut(G).
Then the pair (G, G) is distance-transitive if G is transitive on G
i
= {(v, w):
d(v, w) = i ], for each i = 0, 1, . . . , d. A pair (G, G) is symmetric if and only
if G is transitive on G
0
and G
1
, and so distance-transitive pairs are symmetric. An
equivalent dening condition for distance-transitivity is that (G, G) is distance-
transitive if and only if G is transitive on V and, for a given v V and for
each i = 1, 2, . . . , d, G
v
is transitive on N
i
(v) = {w : d(v, w) = i ]. A distance-
transitive pair (G, G) of degree at least 3 has at most two non-trivial G-invariant
vertex partitions (see [7, Thm. 4.2.1 ff.]). If G is bipartite, then the bipartition is
G-invariant. If G is antipodal that is, if v ,= w and d(u, v) = d(u, w) = d im-
plies that d(v, w) = d then G leaves invariant the antipodal partition consisting
of the antipodal blocks, where the antipodal block containing v is {v] N
d
(v). It
is possible for a distance-transitive pair to be both bipartite and antipodal.
Example 1 For k 2, the complete bipartite graph K
k,k
is regular of degree k,
has diameter 2, and is distance-transitive, bipartite and antipodal with Aut(G) =
S
k
: S
2
. Here the antipodal partition is the same as the bipartition.
Example 2 For d 2, the d-dimensional cube Q
d
is the graph with vertex-set
Z
d
2
, d-tuples v and w being joined by an edge if and only if they differ in exactly
one entry. It has degree and diameter d, and is distance-transitive, bipartite and
antipodal, with Aut(Q
d
) = S
d
2
: S
d
= S
2
: S
d
.
Suppose that (G, G) is distance-transitive of diameter d, and let d
/
= 2d/2.
If G is bipartite and v V, let G
be the graph with vertex-set V
:= {v] N
2
(v)
N
d
/ (v), the part of the bipartition containing v, whose edges are the pairs
from V
at distance 2 in G. Then G
is called a bipartite half of G, (G
, G
) is
distance-transitive, where G
is the index-2 subgroup of G that xes V
set-wise,
and G is called a bipartite double of G
. If G is antipodal of diameter at least 3, with

antipodal partition P, then the antipodal quotient (G
P
, G
P
) is distance-transitive,
and G is a cover of G
P
that is, the map v p(v) that sends each vertex v to the
part p(v) P containing it has the property that, for each v, the restriction to the
neighbourhood N(v) is a bijection whose image is the neighbourhood of p(v) in
G
P
. Thus, if (G, G) is imprimitive, then it is bipartite or antipodal and we obtain
a smaller distance-transitive pair. After at most two such reductions, we reach a
primitive distance-transitive pair.
Thus the classication problem for nite distance-transitive pairs splits into
two separate problems: the major one of classifying the nite primitive distance-
transitive pairs, and the other of nding all distance-transitive pairs obtainable from
the primitive ones by forming antipodal covers and bipartite doubles. There are
many innite families of primitive distance-transitive pairs (see [7]), and among
these the Hamming graphs have a special role.
Example 3 For d 2 and q 2, the Hamming graph H(d, q) is the graph with
vertex-set Z
d
q
, where d-tuples are joined by an edge if and only if they differ in
exactlyone entry. The automorphismgroupAut(H(d, q)) is S
q
: S
d
andis distance-
transitive. For q = 2, H(d, 2) is the d-cube Q
d
and is bipartite and antipodal, while
for q 3, (H(d, q), S
q
: S
d
) is primitive. Furthermore, for q 3, the complement
of H(2, q) is primitive and distance-transitive.
The following theorem is due to Praeger, Saxl and Yokoyama (see [7, Thm.
7.7.1]).
Theorem 5.1 Let (G, G) be a primitive distance-transitive pair. Then either G is
H(d, q) or the complement of H(2, q), or G is primitive of afne type HA or
almost simple type AS.
This reduction theorem led to a programme to classify the primitive distance-
transitive pairs (G, G) of almost simple and afne types see Chapter 9, Sections
5 and 6.
6. Local characterizations
Here we discuss the celebrated theoremof Weiss, that the only nite 8-arc transitive
graphs are cycles. We also mention several other characterizations of families
of symmetric graphs by their local combinatorial or group-theoretic properties,
including locally projective graphs and graphs associated with P-geometries.
s-arc transitive graphs and Weisss theorem
As in Section 2, if G is a subgroup of Aut(G) and s 1, then (G, G) is an s-arc
transitive pair if G is transitive on both vertices and s-arcs of G. We show below
that 2-arc transitivity is characterized by a 2-transitive action of a vertex-stabilizer
G
v
on the neighbourhood N(v). If U V and a subgroup Hof Sym(V) leaves U
invariant, then we denote by H
U
the permutation group on U induced by H.
Lemma 6.1 Let G be a connected graph, let G be a subgroup of Aut(G) that is
vertex-transitive on G, and let v be a vertex. Then (G, G) is a 2-arc transitive pair
if and only if G
N(v)
v
is 2-transitive.
Proof If (G, G) is 2-arc transitive, then G
v
is 2-transitive on N(v), since each
ordered pair (u, w) of distinct vertices of N(v) corresponds to a 2-arc (u, v, w)
with second vertex v. Conversely, let G
v
be 2-transitive on N(v). Then we can map
an arbitrary 2-arc to one with second vertex v, since G is vertex-transitive. Since
G
v
is 2-transitive on N(v), G
v
is transitive on the 2-arcs with second vertex v. Thus
G is transitive on 2-arcs.
A consequence of the classication of the nite simple groups is that all of
the nite 2-transitive permutation groups are known explicitly (see [9, Thm. 5.1]).
Each is primitive and is of afne type HA or almost simple type AS. Building on
work already published in numerous papers, the nal stage of the proof of Weisss
theorem invoked the classication of 2-transitive groups to rule out certain group-
theoretic congurations (see [9, Thm. 5.12]).
Theorem 6.2 If (G, G) is a nite s-arc transitive pair of degree at least 3, then
s < 8.
This result is best possible, since there exists an innite family of 7-arc transitive
graphs associated with the simple groups G
2
(3
n
).
In addition to proving Theorem 6.2, Weiss showed that, if s 4, then G has
degree q 1, for some prime power q, and G
N(v)
v
is an almost simple 2-transitive
group with socle PSL(2, q). Further, if q is even, then s 5, while if q is odd, then
q is a power of 3 and s = 7. Weiss had previously classied all possible amalgams
for 4-arc transitive graphs, and using this classication, Li [22] was able to classify
all primitive and biprimitive 4-arc transitive pairs. His very interesting examples
include, among others, pairs (G, G) constructed in [36] and [39], where G is one
of the sporadic simple groups J
3
, Ru or Th, together with a new example where G
is the Monster simple group.
Theorem 6.3 The nite primitive and biprimitive 4-arc transitive pairs (G, G) are
known explicitly.
2-arc transitive graphs and symmetric 2-designs
Investigations of nite 2-arc transitive graphs were inspired also by Camerons
re-interpretation and generalization of an old result of Manning on primitive per-
mutation groups. Let G be a nite primitive, but not 2-transitive, permutation group
on V. Manning proved that, if a stabilizer G
v
is 2-transitive on an orbit of cardi-
nality k > 2, then G
v
has a second orbit of length m > k, where m is a divisor of
k(k 1) (see [41, 17.7]). If w is a point of the rst G
v
-orbit, and if the G-orbital O
containing (v, w) is self-paired, then by Lemma 6.1, (G(O), G) is a 2-arc transitive
pair, where G(O) is the orbital graph associated with O. Further, all 2-arc transitive
pairs for which the group is primitive on vertices arise in this way. Cameron found
a graph-theoretic meaning for the second G
v
-orbit, as the set N
2
(v) of vertices at
distance 2 from v in G(O) (see [8]).
Theorem 6.4 Let (G, G) be a nite 2-arc transitive pair of degree k > 2, for
which G is primitive on vertices and G is not a complete graph. Then the number
of vertices at distance 2 from a vertex v is k(k 1)/, where is the number of
paths in G of length 2 between a given vertex in N
2
(v) and v.
The largest value of for which an example is known is = 6; it is attained by
a graph of order 100 and degree 22 admitting the Higman-Sims sporadic simple
group. The following conjecture is due to Cameron.
Conjecture If (G, G) is a 2-arc transitive pair of degree k > 2 for which G is
primitive on vertices and G is not a complete graph, then 6.
Although this conjecture has been proved when G
N(v)
v
is almost simple (see [10,
Cor. 3] and [11]), it is still open for afne 2-transitive groups G
N(v)
v
.
If we now drop the primitivity assumption, then examples are known with
arbitrarily large , arising from symmetric 2 ([V[, k, ) designs D (where k
and are as in Theorem 6.4). The incidence graph of D is the graph G(D) with
vertex-set P B, where P is the point-set and B is the block-set, and where the
edges are the incident point-block pairs, called the ags of D. The automorphism
group Aut(D) of Dinduces a group of automorphisms of G(D), and any incidence-
preserving permutation of P B that interchanges P and B (implying that D is
self-dual) induces an automorphismof G(D). Thus, if Dis self-dual and Aut(D) is
ag-transitive, then (G(D), Aut(G(D))) is symmetric with degree k. If, in addition,
the stabilizer of a block is 2-transitive on the points incident with the block, then
(G(D), Aut(G(D))) is 2-arc transitive and, for v P, the set of vertices at distance
2 fromv in G(D) is P{v], and has size v 1 = k(k 1)/. The incidence graph
has girth 6 if D is a projective plane, and otherwise has girth 4. Note that G(D) is
bipartite and so its automorphism group is imprimitive and does not give rise to a
symmetric pair addressed by Camerons conjecture.
Cameron [8, Thms. 4.1 and 4.2] proved that, for > k/2, all 2-arc transitive
pairs arise in this way, and he questioned whether this might be true whenever
> 6. However, it turns out that, when G
N(v)
v
is almost simple, there is a unique
family of examples that do not arise from symmetric designs.
Example For n 3 and q a prime power, the dual orthogonal graph DO
(2n, q)
has as its vertices the n-dimensional totally singular subspaces with respect to a
non-degenerate quadratic form of -type on a 2n-dimensional vector space over
a eld of order q, with two vertices adjacent whenever their intersection has co-
dimension 1 in each.
The following theorem appears in [10] and [11].
Theorem 6.5 Let (G, G) be 2-arc transitive of degree k > 2, with G
N(v)
v
almost
simple, and suppose that, for a vertex v, [N
2
(v)[ = k(k 1)/ with > 6. Then
G is either the incidence graph of a self-dual ag-transitive symmetric 2-design,
or a dual orthogonal graph.
The symmetric designs arising here were described explicitly in [11]; the proof
uses the classication of the almost simple 2-transitive permutation groups. It
is not known whether the conclusion of Theorem 6.5 is also true in the case of
2-transitive afne groups G
N(v)
v
.
Cameron [8] also classied the 2-arc transitive pairs (G, G) of degree k > 2
and girth 4 for which G
N(v)
v
is A
k
or S
k
. The generic examples that arise for any k
are K
k,k
, K
k1,k1
with the edges of a matching removed, or Q
k
and its antipodal
quotient. This classication inspired a similar classication of the odd graphs in
[30], and later a study and eventual classication by Ivanov and Shpectorov of the
class of Petersen and tilde geometries (see [21]).
Locally projective symmetric pairs
Afamily of symmetric pairs that required special attention in the above-mentioned
studies of 2-arc transitive pairs is the family (G, G), where PSL(n, q) G
N(v)
v

PL(n, q), in its natural representation of degree (q
n
1)/(q 1) on the 1-dimen-
sional subspaces of an n-dimensional vector space over a eld of order q. Such
pairs are called locally projective, and the following result of Weiss (which relies
on the classication of nite 2-transitive permutation groups) emphasizes their
special role in the study of 2-arc transitive pairs (see [21, Thm. 9.2.3]).
Theorem 6.6 Let (G, G) be a 2-arc transitive pair such that, for some edge vw,
the pointwise stabilizer in G of N(v) N(w) is non-trivial. Then (G, G) is locally
projective.
As mentioned above, every nite 4-arc transitive pair is locally projective.
Also, when verifying Camerons conjecture for 2-arc transitive pairs (G, G), with
G
N(v)
v
unfaithful and almost simple, it is necessary to characterize an innite fam-
ily of locally projective graphs namely, the dual orthogonal graphs (see [10,
Cor. 1]).
All possibilities for the structure of a vertex-stabilizer G
v
for a locally projective
pair (G, G) have recently been determined in a series of long and difcult papers
by a number of authors; this work is described by Tromov in [37]. A detailed
account of classications and strategies for analysing locally projective pairs can
be found in [21, Ch. 9].
7. Normal quotients
The success of the classication programme for nite distance-transitive pairs
(G, G) was largely due to the fact that antipodal quotient graphs and bipartite
halves of distance-transitive graphs are themselves distance-transitive. Thus, each
distance-transitive pair (G, G) is associated with at least one primitive distance-
transitive pair. However, for families such as the family F
2arc
of 2-arc transitive
pairs, quotients modulo invariant partitions may not lie in the family.
The situation for 2-arc transitive pairs is even more striking. In 1985, Babai [1,
Thm. 1.5] showed that, for each nite regular graph G
0
, there exists a nite 2-arc
transitive pair (G, G) for which G is a cover of G
0
, relative to some vertex-partition
P. Note that P may not be G-invariant, and is certainly not G-invariant if G
0
is not
symmetric. Babais result suggested that F
2arc
might admit no useful organizational
principle. However, it turns out that F
2arc
is closed under forming normal quotients,
and these provide an insightful method for describing the structure of graphs in
this family.
Normal partitions and normal quotients
For a transitive permutation group G on a set V, a partition P of V is G-normal if
P is the set of N-orbits in V, for some normal subgroup N of G. Since elements
of G permute set-wise the orbits of a normal subgroup, each G-normal partition
is G-invariant. Also, the two trivial partitions of V are both G-normal, since they
consist of the orbits either of the identity subgroup or of G itself. If the only G-
normal partitions are the trivial ones, or (equivalently) if every non-trivial normal
subgroup of G is transitive, then G is quasiprimitive on V. It follows that every
primitive permutation group is quasiprimitive, but there are many quasiprimitive
permutation groups that are not primitive; for example, any transitive permutation
representation of a non-abelian simple group, for which a point stabilizer is a
non-maximal subgroup, determines a quasiprimitive permutation group that is not
primitive.
A quotient (G
P
, G
P
) of a symmetric pair (G, G) modulo a G-normal partition P
is called a normal quotient. If P is G-normal, and if [P[ > 1 and the only G-normal
partition rened by P is {V], then G
P
is quasiprimitive. Thus, each symmetric
pair (G, G) has a quasiprimitive normal quotient. However, if G is bipartite, then
the quasiprimitive normal quotient may be simply (K
2
, S
2
), providing very little
information. For bipartite graphs G, the appropriate normal quotients (G
P
, G
P
) to
examine are those where the only G-normal partitions rened by P are {V] and
the bipartition. Such normal quotients, and also the permutation groups G
P
, are
called bi -quasiprimitive. Major questions that arise when investigating a family of
symmetric pairs are whether the family is closed under forming normal quotients,
and if so, whether the quasiprimitive and bi-quasiprimitive pairs in the family can
be analysed successfully. The relationship between a symmetric pair in the family
and its normal quotients is also important.
We saw in Section 6 that a symmetric pair (G, G ) is 2-arc transitive if and only
if the group G
N(v)
v
is 2-transitive. Several other families of interest have dening
properties given in terms of G
N(v)
v
: these include the family F
sym
of all symmetric
pairs, where the property is that G
N(v)
v
is transitive, and the families F
prim
and F
qp
of locally primitive and locally quasiprimitive symmetric pairs which comprise
those (G, G) for which G
N(v)
v
is respectively primitive and quasiprimitive.
We address the above questions for these families. A symmetric pair (G, G) is
a multicover of its quotient (G
P
, G
P
) if there is a constant such that, for each
edge pp
/
of G
P
, each vertex of p is adjacent to vertices of p
/
; we also say that
(G, G) is an -multicover of (G
P
, G
P
) the 1-multicovers are therefore covers.
When P is G-normal, we say that (G, G) is a normal multicover of (G
P
, G
P
), or a
normal cover if = 1. The following theorem can be found in [31, Sec. 4].
Theorem 7.1 Let (G, G) be a nite connected symmetric pair of degree k in F,
where F = F
sym
, F
qp
, F
prim
or F
2arc
, and let G have an intransitive non-trivial
normal subgroup N with orbit-set P in V. Then (G
P
, G
P
) is a connectedsymmetric
pair, and either
G is bipartite and P is the bipartition;
or
(G
P
, G
P
) F, (G, G) is an -multicover of (G
P
, G
P
), and G
P
has degree k/;
further, = 1 if F = F
prim
or F
2arc
.
The general question of constructing normal multicovers of pairs in F
qp
and F
prim
was addressed in [24].
Quasiprimitive groups
In order to analyse the nite quasiprimitive and bi-quasiprimitive pairs in one of the
families of Theorem7.1, we needtounderstandthe structure of nite quasiprimitive
and bi-quasiprimitive permutation groups. It turns out that nite quasiprimitive
permutation groups possess many of the properties of nite primitive groups, in
that similar bounds hold for their orders, minimal degrees and base sizes (see
[34]). In addition, they admit a description similar to that given by Theorem 3.1
for nite primitive groups. We give a version of this result that has proved helpful
in applications involving automorphism groups of vertex-transitive graphs (see
[31, Sec. 5]).
Theorem 7.2 If G is a quasiprimitive permutation group on a nite set V, then G
is of one of the types in Table 1; H denotes a non-abelian nite simple group.
Although some information may be inferred from Theorem 7.2 about nite bi-
quasiprimitive groups, possibilities for their structure are not as well understood
as in the quasiprimitive case (but see [24]).
Table 1. Quasiprimitive types
Type Name Description
HA holomorph of an abelian the unique minimal normal subgroup is elementary
group abelian
HS holomorph of a non-abelian there are two minimal normal subgroups, both
simple group regular, non-abelian and simple
HC holomorph of a composite there are two minimal normal subgroups,
non-abelian group both regular, non-abelian and not simple
AS almost simple H G Aut(H)
TW twisted wreath product the unique minimal normal subgroup H
k
is
regular, non-abelian and not simple
SD simple diagonal the point stabilizer in the unique minimal normal
subgroup H
k
is a diagonal subgroup
CD compound diagonal the point stabilizer in the unique minimal normal
subgroup H
k
is a product of diagonal subgroups
PA product action type the point stabilizer in the unique minimal normal
subgroup H
k
is a sub-direct subgroup of K
k
,
for some proper subgroup K of H
Quasiprimitive and bi-quasiprimitive symmetric pairs
The rst application of Theorem 7.2 was an analysis of the possible types of
quasiprimitive groups G involved in 2-arc transitive quasiprimitive pairs (G, G). It
turns out that, of the eight possible types, only four can occur, and for one of these
types, Ivanov and Praeger classied all the examples (see [31, Sec. 6]).
Theorem 7.3 If (G, G) is a 2-arc transitive quasiprimitive pair, then G has type
HA, AS, TW or PA, and all pairs with G of type HA are known.
Examples are known for each of the types HA, AS, TW, PA, and classications
have been obtained of the 2-arc transitive quasiprimitive pairs (G, G) for some
families of low-rank Lie type almost simple groups G. An extensive analysis of the
situation when G = A
n
or S
n
shows that (G, G) can be described explicitly, unless a
vertex-stabilizer G
v
is itself almost simple, primitive in its action on {1, 2, . . . , n],
and faithful in its action on N(v); this suggests that, for large almost simple
groups G, a complete listing of the 2-arc transitive pairs (G, G) may be infeasible
(see [31, Sec. 6]).
By Theorem 7.1, all non-bipartite 2-arc transitive pairs (G, G) are normal cov-
ers of at least one quasiprimitive 2-arc transitive pair. It would thus be interest-
ing to know the extent to which the 2-arc transitive normal covers of the known
quasiprimitive 2-arc transitive pairs can be described or classied. The situation
for bi-quasiprimitive 2-arc transitive pairs is less well understood, and a better
understanding depends to some extent on a better understanding of bi-quasi-
primitive groups. Currently, the best general analysis is for locally-quasiprimitive
pairs (G, G), in [24, Thm. 1.4]. The bi-quasiprimitive 2-arc transitive pairs (G, G),
in the special case where G is a group of afne transformations, were classied
by Ivanov and Praeger. Studying s-arc transitive pairs (G, G) by considering their
normal quotients has also yielded interesting information in the case where G has
odd order, as described in Li [23].
Theorem 7.4 If (G, G) is an s-arc transitive pair, and if G has odd order, then
s 3.
It may be possible to complete the classication of quasiprimitive 2-arc transitive
and 3-arc transitive pairs of odd order by using the classication of primitive per-
mutation groups of odd degree (see [9, Thm. 5.5]).
In1968, Sims conjectured, ingroup-theoretic language, that for a nite primitive
symmetric pair (G, G), the order of a vertex-stabilizer G
v
must be bounded by some
function of the degree of G. Weiss and the author conjectured that the same should
be true for locally-primitive and locally-quasiprimitive symmetric pairs (see [24,
Sec. 6]).
Bounded stabilizer conjecture There exists an integer function f such that, if
(G, G) is a locally-quasiprimitive symmetric pair of degree k, then [G
v
[ f (k).
Sims conjecture was proved in 1983, using Theorem3.1 together with detailed
information about nite simple groups (see [9, Thm. 5.10]). However, the Bounded
stabilizer conjecture is still open, both for locally-primitive and locally-quasiprimi-
tive pairs. Conder, Li and the author showed that, in the locally-primitive case, this
conjecture is true for non-bipartite graphs if and only if it holds in the special case
when the group G is almost simple (see [24, Sec. 6]). The proof used Theorems
7.1 and 7.2.
8. Finding automorphism groups
It is notoriously difcult to determine the full automorphism group of a graph G,
even if G is symmetric. In some cases, ad hoc combinatorial arguments can be used
to nd a set of congurations (for example, a set of cycles of G) that must be left
invariant by the automorphism group. The fact that Aut(G) is contained in the set-
wise stabilizer in Sym(V) of these congurations may facilitate the determination
of Aut(G).
A more systematic study for some families of symmetric graphs uses group
theory; this has been possible for the class of primitive symmetric pairs. Suppose
that (G, G) is primitive, so that G is a primitive subgroup of Sym(V). The group G
may be much smaller than Aut(G); for example, for the complete graph G = K
8
,
all the groups in the following subgroup chain are primitive on V and symmetric
on G:
PSL(2, 7) < AGL(3, 2) < A
8
< Aut(G) = S
8
.
A general determination of the subgroups of Aut(G) containing a given vertex
primitive subgroup G requires a comprehensive knowledge of the subgroups of
Sym(V) containing G. This latter information is available (see [31, Sec. 7]), and
has been applied to determine the primitive permutation groups that can act sym-
metrically on a graph and have a different socle fromthat of the full automorphism
group, as shown in [26].
Theorem 8.1 Suppose that (G, G) is a primitive symmetric pair. Then either G
and Aut(G) have the same socle, or there exists H such that G < H < Aut(G)
and H = soc(H)G, and G and H can be found explicitly.
However, as we have seen, for many families of symmetric pairs the primitive
members, although interesting, do not give a good guide to the structure of typical
members of the family. Some of these families are closed under forming normal
quotients, and typical graphs in these families are multicovers of their normal
quotients. It follows that the quasiprimitive and bi-quasiprimitive members are of
special interest, and so results like Theorem 8.1 for such graphs would be of great
value. To achieve such results, the rst step would be to determine the subgroups
of Sym(V) that contain a given quasiprimitive or bi-quasiprimitive group G. The
next step would be to apply this to study the subgroups of Aut(G) that contain G.
The situation is more complicated than the primitive case, since such subgroups
need not have the same properties. For example, there are quasiprimitive 2-arc
transitive pairs (G, G) for which Aut(G) is not quasiprimitive (see [31, Sec. 7]).
Given a nite quasiprimitive subgroup G of Sym(V), most primitive subgroups
H containing G are determined in [2]. The cases that are not treated completely
are where G and H are both almost simple, and where H has product action type
PA and G has type TW or PA: dealing with these outstanding cases is a theme of
ongoing research of the author and others. These results are then applied in [33] to
determine the imprimitive quasiprimitive groups G < H < Sym(V) for which G
and H have different quasiprimitive types or socles. Again, the cases that cannot
be treated completely are those where G and H are both almost simple, or where
H is of product action type PA.
The most important unsolved problem in this area is the determination of im-
primitive quasiprimitive groups G < H < Sym(V) for which G and H are both
almost simple. One of the more interesting challenges thrown up by this work is
to make a thorough investigation of Cartesian decompositions of the vertex-sets
that are preserved by quasiprimitive automorphism groups of symmetric graphs.
9. A geometric approach
The structure of distance-transitive graphs presents us with a tantalizing challenge
to nd a similar methodology that might provide a way to analyse other classes
of symmetric graphs. We have seen that the class of connected symmetric pairs
(G, G) is closed under forming quotients, relative to G-invariant vertex-partitions
P; moreover, if P is G-normal and [P[ > 2, then (G, G) is a normal multicover of
(G
P
, G
P
). However, in general, if P is not G-normal, then G is not a multicover
of G
P
and it is unclear how much information about G can be obtained from G
P
.
Also, the induced subgraph on a part p P is a null graph, and this provides no
additional insight into the structure of G.
In [16] it was suggested that, in addition to the quotient G
P
, two other combina-
torial congurations could be used to describe a symmetric pair (G, G). The rst is
the bipartite subgraph G( p, p
/
) induced on the union p p
/
of two adjacent parts
p, p
/
P of G
P
. Since (G
P
, G
P
) is symmetric, this subgraph is symmetric and is
independent (up to equivalence) of the choice of p and p
/
. The second congura-
tion is the incidence structure D
P
( p) induced by the adjacency relation on the pair
( p, N
P
( p)), where N
P
( p) is the set of parts adjacent to p in G
P
: a vertex v p
is incident with p
/
N
P
( p) precisely when v is adjacent to some vertex in p
/
. An
initial study of how these congurations might aid an understanding of symmetric
graphs was made in [16], and a more extensive analysis in the case where G
P
is
a complete graph has been carried out in [17] and [18]. This approach raises a
fundamental question:
Under what conditions is G determined by the triple (G
P
, D
P
( p), G( p, p
/
))?
The incidence structure D
P
( p) may be interpreted as a design with point-set p
and block-set N
P
( p). This suggests certain other questions that might not otherwise
be considered. For example, it is natural to study the case where D
P
( p) has no
repeated blocks that is, where the subsets of p incident with parts p
/
and p
//
in N
P
( p) are distinct when p
/
,= p
//
. The role of this special case is particularly
striking if exactly [ p[ 1 vertices of p are joined to some vertex of p
/
that is,
the blocks of D
P
( p) have size [ p[ 1. An interesting family of examples is as
follows.
Example Let H = (V
H
, E
H
) be a nite connected regular graph of degree k 3,
andlet be a self-paired set of 3-arcs of H: if (u, v, w, x) then(x, w, v, u)
. The 3-arc graph Arc
(H) is the graph whose vertices are the arcs of H, and

where (v, w) and (v
/
, w
/
) are adjacent if and only if (w, v, v
/
, w
/
) .
If G is a subgroup of Aut(H), then (Arc
(H), G) is a symmetric pair if and

only if (H, G) is a symmetric pair and is a G-orbit on the 3-arcs of H; in this
case, P = { p(v) : v V
H
], where p(v) = {(v, w) : w N(v)], is a G-invariant
partition of the arcs of H with the property that D
P
( p(v)) has no repeated blocks,
and has block size [ p(v)[ 1. It turns out that, for [ p[ 3, the 3-arc graphs are
the only examples with this property (see [25]).
Theorem 9.1 Suppose that (G, G) is a nite symmetric pair and that P is a G-
invariant vertex partition such that, for p P, = [ p[ 3 and D
P
( p) has block
size 1. Then D
P
( p) has no repeated blocks if and only if (G
P
, G
P
) is 2-arc
transitive, and in this case, G

= Arc
.
(G
P
) for some self-paired G
P
-orbit on
the 3-arcs of G
P
.
Theorem 9.2 For (G, G) and P as in Theorem 9.1, the following properties are
equivalent: (G
P
, G
P
) is 3-arc transitive, G( p, p
/
)

= K
1,1
, and is the set of
all the 3-arcs of G
P
.
Given our experience to date, one strategy for studying imprimitive symmetric
pairs (G, G) might be rst to investigate proper normal quotients of order at least
3. If there are no such quotients, then (G, G) is quasiprimitive or bi-quasiprimitive,
and in this case we might study a triple (G
P
, D
P
( p), G( p, p
/
)), where P is min-
imal. Here we have the additional information that (G, G) is quasiprimitive or
bi-quasiprimitive, and also that G
N(v)
v
is primitive.
10. Related families of graphs
Some classes of nite graphs possess interesting symmetry properties, but are
not quite symmetric. Although not the subject of this chapter, they have attracted
attention in the literature, and we mention them here to set the chapter in a slightly
wider context.
Aconnected graph G that is edge-transitive need not be vertex-transitive. In this
case, G is bipartite and Aut(G) has two orbits on vertices forming the bipartition.
If such a graph is regular, then it is called semi-symmetric and the rst examples
were constructed by Folkman [15]. Alternatively if, for each vertex v, (Aut(G))
v
is transitive on the set of s-arcs starting with v, then G is locally s-arc transitive.
By studying the possible amalgams of locally s-arc transitive graphs in which each
vertex has degree at least 3, Stellmacher has shown that s 9. Also, a systematic
study of locally s-arc transitive graphs with s 2, using a similar approach to
Section 7, can be found in [19].
Similarly, a connected graph G that is edge-transitive and vertex-transitive need
not be symmetric. If it is not, then it has even degree and Aut(G) has two orbits
on arcs. A pair (G, G) is called half-arc transitive if G is transitive on vertices and
on edges, but not on arcs; examples of half-arc transitive pairs (G, Aut(G)), for
all even degrees, were constructed by Bouwer [6]. For a given transitive subgroup
G of Sym(V), the half-arc transitive pairs (G, G) with vertex-set V are in one-one
correspondence with the orbital digraphs G(O), for non-self-paired G-orbitals O.
However, for a given half-arc transitive pair (G, G), the pair (G, Aut(G)) may be
symmetric, and deciding whether this is true may be difcult. Half-arc transitive
pairs (G, G) of degree 4 for which the stabilizers G
v
have order 2 arise naturally
when studying the embeddings of graphs in surfaces (see [27]).
Innite symmetric graphs of nite degree have already occurred in this chapter
as coset graphs of universal covers of some amalgams in Section 3. The general
study of such graphs is a separate and fruitful area of combinatorics, and a good
introduction may be found in [38].
References
1. L. Babai, Automorphism groups, isomorphism, reconstruction, Handbook of Combi-
natorics (eds. R. Graham, M. Gr otschel and L. Lov asz), Elsevier (1995), 14471540.
2. R. W. Baddeley and C. E. Praeger, On primitive overgroups of quasiprimitive permu-
tation groups, J. Algebra 263 (2003), 294344.
3. N. Biggs, Algebraic Graph Theory, Cambridge Tracts in Mathematics 67, Cambridge
Univ. Press, 1974.
4. N. L. Biggs, Presentations for cubic graphs, Computational Group Theory (Durham,
1982), Academic Press (1984), 5763.
5. N. L. Biggs, Mathematical Reviews, 88c:05063.
6. I. Z. Bouwer, On edge but not vertex transitive regular graphs, J. Combin. Theory (B)
12 (1972), 3240.
Verlag, 1989.
8. P. J. Cameron, Suborbits in transitive permutation groups, Combinatorics (Proc. NATO
Advanced Study Inst., Breukelen, 1974), Part 3: Combinatorial Group Theory, Math.
Centre Tracts 57, Math. Centrum (1974), 98129.
9. P. J. Cameron, Permutation groups, Handbook of Combinatorics, Vol. 1 (eds. R.
Graham, M. Gr otschel and L. Lov asz), Elsevier (1995), 611645.
10. P. J. Cameron and C. E. Praeger, Graphs and permutation groups with projective sub-
constituents, J. London Math. Soc. (2) 25 (1982), 6274.
11. P. J. Cameron and C. E. Praeger, On 2-arc transitive graphs of girth 4, J. Combin. Theory
(B) 35 (1983), 111.
12. M. Conder, An innite family of 5-arc-transitive cubic graphs, Ars Combin. 25(A)
(1988), 95108.
13. M. Conder, An innite family of 4-arc-transitive graphs each with girth 12, Bull. London
Math. Soc. 21 (1989), 375380.
14. M. Conder and P. Lorimer, Automorphism groups of symmetric graphs of valency 3,
J. Combin. Theory (B) 47 (1989), 6072.
15. J. Folkman, Regular line-symmetric graphs, J. Combin. Theory 3 (1967), 215232.
16. A. Gardiner and C. E. Praeger, A geometrical approach to imprimitive graphs, Proc.
London Math. Soc. (3) 71 (1998), 524546.
17. A. Gardiner and C. E. Praeger, Topological covers of complete graphs, Math. Proc.
Cambridge Philos. Soc. 123 (1998), 549559.
18. A. Gardiner, C. E. Praeger and S. Zhou, Cross ratio graphs, J. London Math. Soc. 64
(2001), 257272.
19. M. Giudici, C. H. Li and C. E. Praeger, Analysing nite locally s-arc transitive graphs,
Trans. Amer. Math. Soc. 356 (2004), 291317.
20. M. Heydemann, Cayley graphs and interconnection networks, Graph Symmetry (Mon-
treal, PQ, 1996) (eds. G. Hahn and G. Sabidussi), Kluwer (1997), 167224.
21. A. A. Ivanov, Geometry of Sporadic Groups I, Petersen and Tilde Geometries, Cam-
bridge Univ. Press (1999).
22. C. H. Li, The nite vertex-primitive and vertex-biprimitive s-transitive graphs with
s 4, Trans. Amer. Math. Soc. 353 (2001), 35113529.
23. C. H. Li, On nite s-arc transitive graphs of odd order, J. Combin. Theory (B) 81 (2001),
307317.
24. C. H. Li, C. E. Praeger, A. Venkatesh and S. Zhou, Finite locally-quasiprimitive graphs,
Discrete Math. 246 (2002), 197218.
25. C. H. Li, C. E. Praeger and S. Zhou, A class of nite symmetric graphs with 2-arc
transitive quotients, Math. Proc. Cambridge Philos. Soc. 129 (2000), 1934.
26. M. W. Liebeck, C. E. Praeger and J. Saxl, Primitive permutation groups with a common
suborbit, and edge-transitive graphs, Proc. London Math. Soc. 84 (2002), 405438.
27. D. Marusic and R. Nedela, Maps and half-transitive graphs of valency 4, European J.
Combin. 19 (1998), 345354.
28. M. J. Morton, Classication of 4- and 5-arc-transitive cubic graphs of small girth, J.
Austral. Math. Soc. (A) 50 (1991), 138149.
29. A. Pasini, Diagram Geometries, Oxford Univ. Press, 1994.
30. C. E. Praeger, Primitive permutation groups and a characterization of the odd graphs,
J. Combin. Theory (B) 31 (1981), 117142.
31. C. E. Praeger, Finite quasiprimitive graphs, Surveys in Combinatorics, 1997 (London)
(ed. R. A. Bailey), Cambridge Univ. Press (1997), 6585.
32. C. E. Praeger, Finite transitive permutation groups and nite vertex-transitive graphs,
Graph Symmetry (Montreal, PQ, 1996) (eds. G. Hahn and G. Sabidussi), Kluwer
(1997), 277318.
33. C. E. Praeger, Quotients and inclusions of nite quasiprimitive permutation groups, J.
Algebra 269 (2003), 329346.
34. A. Shalev and C. E. Praeger, Bounds on nite quasiprimitive permutation groups, J.
Austral. Math. Soc. 71 (2001), 243258.
35. C. C. Sims. Graphs and nite permutation groups. Math. Z. 95 (1967), 7686.
36. G. Stroth and R. Weiss, A new construction of the group Ru, Quart. J. Math. (2) 41
(1990), 237243.
37. V. I. Tromov, Vertex stabilisers of locally projective groups of automorphisms of
graphs. A summary, Groups, Combinatorics and Geometry: Durham 2001 (eds. A. A.
Ivanov, M. W. Liebeck and J. Saxl), World Scientic, 2003.
38. M. E. Watkins, Ends and automorphisms of innite graphs, Graph Symmetry (Montreal,
PQ, 1996) (eds., G. Hahn and G. Sabidussi), Kluwer (1997), 379414.
39. R. Weiss, Acharacterization and another construction of Jankos group J
3
, Trans. Amer.
Math. Soc. 298 (1986), 621633.
40. R. Weiss, Presentations for (G, s)-transitive graphs of small valency, Math. Proc. Cam-
bridge Philos. Soc. 101 (1987), 720.
41. H. Wielandt, Finite Permutation Groups, Academic Press, 1964.
8
Strongly regular graphs
PETER J. CAMERON
1. An example
2. Regularity conditions
3. Parameter conditions
4. Geometric graphs
5. Eigenvalues and their geometry
6. Rank 3 graphs
7. Related classes of graphs
References
Strongly regular graphs form an important class of graphs which lie some-
where between the highly structured and the apparently random. This chapter
gives an introduction to these graphs with pointers to more detailed surveys
of particular topics.
1. An example
Consider the Petersen graph, shown in Fig. 1.
Fig. 1.
203
This graph has far too many remarkable properties for even a brief survey here.
(It is the subject of an entire book by Holton and Sheehan [17].) We focus on a
few of its properties: it has ten vertices, is regular of degree 3, and has diameter 2
and girth 5.
These properties are not all independent: simple counting arguments show that
a trivalent graph with diameter 2 has at most ten vertices, with equality if and only
it has girth 5; dually, a trivalent graph with girth 5 has at least ten vertices, with
equality if and only it has diameter 2.
The conditions diameter 2 and girth 5 can be rewritten thus: any two adjacent
vertices have no common neighbours and any two non-adjacent vertices have
exactly one common neighbour. Replacing the particular numbers (10, 3, 0, 1) by
general parameters, we reach the denition of a strongly regular graph.
A strongly regular graph with parameters (n, k, , ) is a graph on n vertices
which is regular of degree k and has the following properties:
r
any two adjacent vertices have exactly common neighbours;
r
any two non-adjacent vertices have exactly common neighbours.
So the Petersen graph is a strongly regular graph with parameters (10, 3, 0, 1).
The complete and null graphs are vacuously strongly regular, the parameters
and (respectively) being undened for them. We often exclude these trivial
cases.
The four parameters are not independent. For any given vertex v, we count in
two ways the ordered pairs (x, y) of adjacent vertices for which x is adjacent to v,
but y is not. We obtain the following result.
Theorem 1.1 The parameters (n, k, , ) of a strongly regular graph satisfy the
equation
k(k 1) = (n k 1).
Later in this chapter we see that the parameters also satisfy various algebraic con-
ditions and inequalities. However, no complete characterization of the parameter
sets of strongly regular graphs is known.
This section concludes with three further basic facts about strongly regular
graphs.
Theorem 1.2
(a) The complement of a strongly regular graph is strongly regular.
(b) A strongly regular graph is disconnected if and only if it is isomorphic to
mK
r
(the disjoint union of m copies of K
r
), for some positive integers m and
r; this occurs if and only if = 0.
(c) Every connected strongly regular graph has diameter 2.
8 Strongly regular graphs 205
2. Regularity conditions
We can put the denition of a strongly regular graph into a more general context as
follows. For each non-negative integer t , let C(t ) be the following graph property.
Let S
1
and S
2
be sets of at most t vertices. If the induced sub-graphs on S
1
and
S
2
are isomorphic, then the number of vertices joined to every vertex in S
1
is
equal to the number joined to every vertex in S
2
.
A graph satisfying C(t ) is sometimes called t-tuple regular. In such a graph, let
(S) denote the number of common neighbours of a set of vertices inducing a
subgraph isomorphic to S, where S runs over all isomorphism types of graphs
with at most t vertices.
The conditions C(t ) become stronger as t increases. C(0) is vacuous, and ()
is just the number of vertices of the graph G. A graph satises C(1) if and only if it
is regular: for each vertex v, (v) is the degree. Agraph satises C(2) if and only if
it is strongly regular: (edge) and (non-edge) are the above parameters and .
In fact, the hierarchy is nite [7], as the following result shows.
Theorem 2.1 A graph that satises C(5) also satises C(t ) for all non-negative
integers t . The only such graphs are nK
r
and its complement for all n and r 1,
the 5-cycle C
5
, and the 3 3 square lattice L(K
3,3
).
There are only two known examples (up to complementation) of graphs that
satisfy C(4) but not C(5): the Schl ai graph on 27 vertices and the McLaughlin
graphon275vertices. Innitelymanyadditional graphs satisfyingC(3) are known;
all of them, except for L(K
n,n
) (for n 4) and nitely many others, are associ-
ated with geometric objects such as quadrics in projective spaces and extremal
generalized quadrangles.
On the other hand, there is no shortage of graphs that satisfy C(0) or C(1). The
number of graphs on n vertices is asymptotic to 2
n(n1)/2
/n!, while the number
of k-regular graphs is asymptotically equal to c
k
n
n(k2)/2
/n!, for 2 < k = o(
n).
(Estimates exist also for k cn; see Wormald [32].) For both graphs and regular
graphs, there are well-developedtheories of randomobjects, includingthe assertion
that almost all of them have no non-trivial automorphisms this explains the n!
in the denominators of the asymptotic formulas).
Strongly regular graphs stand on the cusp between the random and the highly
structured. For example, there is a unique strongly regular graph with parame-
ters (36, 10, 4, 2), but a computation by McKay and Spence [20] showed that
the number of strongly regular graphs with parameters (36, 15, 6, 6) is 32548.
This pattern continues: there is a unique strongly regular graph with parame-
ters (m
2
, 2(m 1), m 2, 2), but more than exponentially many strongly regular
graphs with parameters (m
2
, 3(m 1), m, 6), as we will see. This suggests that
no general asymptotic results are possible, and that, depending on the parameters,
strongly regular graphs can behave in either a highly structured or an apparently
random manner.
Another role of strongly regular graphs is as examples for graph isomorphism
testing algorithms. The global uniformity ensured by the denition makes it harder
to nd a canonical labelling, while the super-exponential number of graphs means
that they cannot be processed as exceptions. Paley graphs and other strongly regular
(and similar) graphs have also been used as models of pseudo-random graphs
(see Thomason [28]).
Recently, Fon-Der-Flaass [14] has observed that an old construction of Wallis
[29] gives rise to more than exponentially many strongly regular graphs with
various parameter sets, which we discuss below. He also used these graphs to
establish the following result about the universality of strongly regular graphs.
Theorem 2.2 Each graph on n vertices is an induced subgraph of a strongly
regular graph on at most 4n
2
vertices. This is within a constant factor of being
best possible.
It is not known whether such a universality result holds for graphs satisfying C(3).
3. Parameter conditions
The parameters of strongly regular graphs satisfy a number of restrictions, the
most important of which are described here.
Theorem 3.1 Let Gbe a strongly regular graph with parameters (v, k, , ). Then
the numbers
f and g =
1
2
v 1
(v 1)( ) 2k
( )
2
4(k )
are non-negative integers.

Proof Let A be the adjacency matrix of G. The fact that G is strongly regular
shows that
A
2
= kI A (J I A),
where J is the all-1 matrix. The all-1 vector j is an eigenvector of Awith eigenvalue
1. Any other eigenvector of A is orthogonal to j, so the corresponding eigenvalue
saties the quadratic equation
x
2
= k x (1 x).
From this, we can calculate the two eigenvalues r and s and (using the fact that the
trace of A is 0) their multiplicities f and g, obtaining the given expressions.
On the basis of this theorem, we can classify strongly regular graphs into two
types:
Type I (or conference) graphs: for these graphs, (v 1)( ) 2k) = 0.
This implies that = 1, k = 2 and n = 4 1. (They are precisely the
strongly regular graphs with the same parameters as their complements.) It is
known that they exist if and only if v is the sum of two squares.
Type II graphs: for these graphs, ( )
2
4(k ) is a perfect square d
2
,
where d divides (v 1)( ) 2k and the quotient is congruent to v 1
(mod 2).
Examples of conference graphs include the Paley graphs P(q): the vertex-set of
P(q) is the nite eld GF(q), where q is a prime power congruent to 1 (mod 4),
and v and w are adjacent if and only if v w is a non-zero square in GF(q) (see
Paley [23]).
The non-principal eigenvalues r and s of a Type II strongly regular graph are
integers with opposite signs. The parameters may be conveniently expressed in
terms of the eigenvalues as follows:
= k r s rs, = k rs.
Of the other conditions satised by the parameters of a strongly regular graph,
the most important are the Kren conditions, rst proved by Scott [24] using a
result of Kren [18] from harmonic analysis. They state that
(r 1)(k r 2rs) (k r)(s 1)
2
and
(s 1)(k s 2rs) (k s)(r 1)
2
.
The rst bound is attained by a graph if and only if the second is attained by its
complement. As we shall see, the two inequalities are associated with the geometry
of the two non-trivial eigenspaces of the adjacency matrix.
Some parameter sets satisfy all known necessary conditions. We mention a few
of these here.
The pseudo-Latin square PL
r
(n), with 1 r n, has parameters v = n
2
,
k = r(n 1), = r
2
3r m, = r(r 1); the signicance of the name
will appear in the next section.
The negative Latin square NL
r
(n), obtained by replacing r and n by their
negatives in the formulas just given. Since this gives = r
2
3r n, we
must have n r(r 3); equality holds if and only if the Kren bound is
attained.
Smith graphs, whose somewhat involved parameters will not be given here
(see [9, p. 111]). These parameters always attain the Kren bound.
These parameters arise in the following theorem of Cameron et al. [8]:
Theorem 3.2 Let G be a graph satisfying C(3). Then either G is the pentagon, or
its parameters are of pseudo-Latin square, negative Latin square, or Smith type.
4. Geometric graphs
The notion of a partial geometry was introduced by Bose [3] as a tool for studying
strongly regular graphs. Subsequently, partial geometries have been studied in their
own right, and the concept has been extended in various ways, not all of which are
related to strongly regular graphs. This section focuses only on the connections.
A partial geometry with parameters (s, t, ) is an incidence structure of points
and lines satisfying the following axioms:
r
each line contains s 1 points, and each point lies on t 1 lines;
r
two lines meet in at most one point, and two points lie on at most one line;
r
if the point p does not lie on the line L, then there are precisely incident pairs
(q, M), where q is a point of L and M is a line through p.
Note that Bose used slightly different parameters: he wrote R, K, T for what we
call s 1, t 1, .
The dual of a partial geometry with parameters (s, t, ) is obtained by inter-
changing the names point and line for the two types of object, and dualizing
the incidence relation. It is a partial geometry with parameters (t, s, ).
The point graph of a partial geometry is the graph whose vertices are the points
of the geometry, with adjacency being dened by collinearity. The line graph is
the point graph of the dual geometry: its vertices are the lines, and adjacency is
given by concurrence.
Theorem 4.1 The point graph of a partial geometry with parameters (s, t, ) is a
strongly regular graph with parameters (n, s(t 1), s 1 t ( 1), (t 1)),
where n = (s 1)(st )/.
The proof of this result is straightforward.
Motivated by this, we say that a strongly regular graph G is geometric if it is the
point graph of a partial geometry, and that G is pseudo-geometric if its parameters
have the form given in Theorem 4.1, for some positive integers s, t and . Some-
times we append the triple (s, t, ) to the term geometric or pseudo-geometric.
Not every pseudo-geometric graph is geometric. Indeed, a pseudo-geometric
(s, t, )-graph is geometric if and only if there is a collection L of (s 1)-cliques
with the property that each edge lies in just one clique of L. If some edge lies in no
(s 1)-clique, then the graph is clearly not geometric, but if there are too many
cliques, it is often not clear whether a suitable collection can be selected.
The major result of Bose [3] can now be stated:
Theorem 4.2 (Boses theorem) Suppose that s, t and are positive integers
satisfying
s >
1
2
(t 2)(t 1 (t
2
1)).
Then any pseudo-geometric (s, t, )-graph is geometric, and each edge of such a
graph lies in a unique (s 1)-clique.
In order to see the power of this theorem, we look at partial geometries a little
more closely. We divide them into six types.
Linear spaces: the case = s 1. A point not on a line L is collinear with every
point of L. It follows that any two points lie on a unique line. Such structures are
also known as 2-designs, pairwise balanced designs or Steiner systems. The point
graph of a linear space is just a complete graph, and is of no interest.
We note in passing the asymptotic existence theoremof Richard Wilson [30]. A
necessary condition for a linear space with n points, and with s 1 points on each
line, is that s divides n 1 and s 1 divides n(n 1). In terms of s and t , this
is the single condition that s 1 divides t (t 1). Wilson showed the existence
of a function f (s) such that a linear space exists if t > f (s) and the necessary
condition is satised.
Dual linear spaces: the case = t 1. Here, the geometries are the duals of those
in the preceding case, but the graphs (the line graphs of linear spaces) are much
more interesting.
We examine two special cases. A linear space with two points on each line is
a complete graph K
n
, and its line graph (the point graph of the dual) is the line
graph L(K
n
). We have s = 1, = 2 and t = n 2, and the inequality in Boses
theorem reduces to n > 8. We have the following result.
Corollary 4.3 The graph L(K
n
) is strongly regular and, for n > 8, it is the unique
strongly regular graph with its parameters.
The conclusion actually holds for all n ,= 8. For n = 8, Chang [10] showed that
there are exactly four strongly regular graphs with parameters (28, 12, 6, 4). In the
literature on strongly regular graphs, L(K
n
) is called the triangular graph T(n).
A linear space with three points on each line is a Steiner triple system. The fact
that a Steiner triple system with n points exists for all n 1 or 3 (mod 6) goes
back to Kirkman in 1847. More recently, Wilson [30] showed that the number v(n)
of Steiner triple systems on an admissible number n of points satises
v(n) exp(n
2
log n/6 cn
2
).
Moreover, a Steiner triple system of order n > 15 can be recovered uniquely
from its line graph. Hence there are super-exponentally many strongly regular
graphs with parameters (n, 3s, s 3, 9), for n = (s 1)(2s 3)/3 and s 0 or
2 (mod 3).
Transversal designs: the case s = . In this case, it can be shown that there is a
partition of the points into subsets traditionally called groups (although there is
no connection with the algebraic notion): each line is a transversal for the family
of groups, and any two points in different groups lie on a line. Thus, the point
graph is complete multipartite, and the partite sets are the groups.
Dual transversal designs: the case t = . For t = 2, the graph is the line graph
of the complete bipartite graph K
n,n
. Boses theorem then gives us the following
result.
Corollary 4.4 The graph L(K
n,n
) is strongly regular, and for n > 4 is the unique
strongly regular graph with its parameters.
This holds for all n ,= 4. Shrikhande showed that there are just two strongly regular
graphs with parameters (16, 6, 2, 2) namely, L(K
4,4
) and one other, now called
the Shrikhande graph and dened below. In the literature on strongly regular
graphs, L(K
n,n
) is called the square lattice graph L
2
(n).
If t > 2, for reasons which will become clear, we use new parameters n and r,
where n = s 1 and r = t 1. The groups dualize to become a partition of the
lines into parallel classes, each parallel class forming a partition of the points.
There are n lines in each parallel class, with n points on each, so the total number
of points is n
2
. The geometry is called a net of order n and degree r.
Select two parallel classes {V
1
, V
2
, . . . , V
n
] and {H
1
, H
2
, . . . , H
n
]. Then the
points can be represented as an n n grid, where the lines V
i
run vertically and
the H
j
run horizontally, and the unique point on V
i
and H
j
can be labelled p
i j
.
Now let {L
1
, L
2
, . . . , L
n
] be another parallel class of lines and construct an
n n array with ij-entry l if p
i j
L
l
: then is a Latin square of order n. By
reversing the construction, we nd that any Latin square of order n gives rise to
a net of order n and degree 3. Since the number of Latin squares of order k is
asymptotically equal to exp(n
2
log n 2n
2
), we obtain super-exponentially many
strongly regular graphs with these parameters.
There are just two non-isomorphic Latin squares of order 4: the Cayley
tables of the Klein group and the cyclic group of order 4. They give rise to two
non-isomorphic strongly regular graphs with parameters (16, 9, 4, 6), whose com-
plements are L
2
(4) and the Shrikhande graph.
If r > 3, we have r 2 additional parallel classes, giving rise to r 2 Latin
squares. It is easily checked that these Latin squares are mutually orthogonal: given
any two squares and
/
and any given entries l, l
/
, there is a unique cell in which
and
/
have entries l and l
/
. Conversely, a set of r 2 mutually orthogonal Latin
squares of order n gives a net of order n and degree r, and hence a strongly regular
graph.
For this reason, the point graph of a net of order n and degree r is called a Latin
square graph, denoted by L
r
(n). A pseudo-geometric graph with the parameters
of L
r
(n) is the same as the pseudo-Latin square graph PL
r
(n) dened earlier
hence the name.
If n is a prime power, then there exists a set of mutually orthogonal Latin
squares of order n and of the maximum possible size n 1; the corresponding
net is an afne plane of order n. Choosing all subsets of r 2 of these squares,
where r cn for 0 < c < 1, we again obtain many strongly regular graphs with
the same parameters, but with only a fractional exponential number in this case.
Generalized quadrangles: the case = 1. In this case the geometry is trivially
recoverable from its point graph, since an edge lies in a unique maximal clique.
There are classical generalized quadrangles, related to the classical groups (the
symplectic, unitary and orthogonal groups), in much the same way as projective
planes are related to the projective groups PGL(3, q), and also non-classical ex-
amples, including some with non-classical parameters. Van Maldegham [21] has
surveyed these geometries.
Fon-Der-Flaass [14] has pointed out that some of Walliss graphs [29] and
variants of them have pseudo-geometric parameters corresponding to generalized
quadrangles with s = q 1 and t = q 1, or with s = t = q, or with s = q 1
and t = q 1, where q is a prime power. So there are super-exponentially many
graphs for these parameter sets.
The rest: with 1 < < min{s, t ]. Here some examples are known, but there is
much less theory.
We conclude this section with a reference to the work of Neumaier [22], which
improves Boses classical results. From Neumaiers work, we quote two of his
most notable results.
Theorem 4.5 A strongly regular graph with parameters (v, k, , ) and eigenval-
ues k, r, s with s < 1 and
r >
1
2
s(s 1)( 1) 1
is the point graph of a dual linear space or a dual transversal design.
The inequality reduces to Boses inequality in the pseudo-geometric case, but
Neumaiers result applies without this assumption.
Theorem 4.6 For each negative integer m, there is a nite set L(m) of strongly
regular graphs with the following property: if G is any connected strongly regular
graph whose adjacency matrix has eigenvalue m, then Gis a complete multipartite
graph with block size m, or the point graph of a dual linear space or a dual
transversal design with t 1 = m, or a member of the list L(m).
For m = 1 the result is trivial. For m = 2 it was proved by Seidel [25] that
the set L(2) contains only the Petersen, Clebsch, Schl ai, Shrikhande, and three
Chang graphs. However, all but twelve of the 32,548 graphs with parameters (36,
15, 6, 6), mentioned earlier, belong to L(3): only twelve come fromLatin squares.
5. Eigenvalues and their geometry
Let G be a strongly regular graph with vertex-set V = {v
1
, v
2
, . . . , v
n
] and ad-
jacency matrix A. As we have seen, A has just three distinct eigenvalues k, r
and s, with multiplicities 1, f and g, respectively (so that 1 f g = n); the
eigenvector associated to the eigenvalue k is the all-1 vector. Thus,
A = kE
0
rE
1
sE
2
,
where E
0
, E
1
, and E
2
are the orthogonal projections of R
n
onto the three eigen-
spaces V
0
, V
1
and V
2
of A.
We nowx attention on one of the non-trivial eigenspaces (V
1
, say) and consider
the projections of the vertices (the basis vectors of R
n
) onto V
1
. Thus, let x
i
= v
i
E
1
,
for i = 1, 2, . . . , n. The basic property of these vectors is the following.
Theorem 5.1 There are real numbers , and , expressible in terms of the
parameters of G, such that the inner products of the vectors x
i
are given by
x
i
, x
j
) =
, if v
i
= v
j
;
, if v
i
v
j
;
, if v
i
,= v
j
and v
i
v
j
.
Moreover, if G is connected and not complete multipartite, then x
i
,= x
j
for i ,= j .
In particular, if G is connected and not complete multipartite (as we assume for
the rest of this section), then the vectors x
1
, x
2
, . . . , x
n
lie on a sphere of radius
in R
f
, and the angular distances between them take one of two possible values:
arccos / (for adjacent vertices) and arccos / (for non-adjacent vertices). It
is convenient to re-scale the vectors by 1/
so that they lie on the unit sphere.

Delsarte et al. [11] have proved the following result:
Theorem 5.2 The cardinality n of a two-distance set on the unit sphere in R
f
satises
n
f 2
2
1.
This result can be translated into an inequality on the parameters of a strongly
regular graph, which is the so-called absolute bound. The same authors also gave
a special bound that depends on the values of , and that is, on the actual
distances realised by the set; it does not apply for all values of the parameters, but
x is sometimes more powerful than the absolute bound.
A set X = {x
1
, x
2
, . . . , x
n
] of vectors lying on the unit sphere = S
f 1
in
Euclidean space R
f
is a spherical t-design if, for any polynomial function F of
degree at most t , we have
1
n
n
i =1
F(x
i
) =
1
vol
F(x)dx.
In other words, the nite set approximates the sphere up to degree t . For small
t , there is a mechanical interpretation. Place unit masses at the points of X. Then
X is a spherical 1-design if and only if the centre of mass is at the origin, and is
a spherical 2-design if, in addition, the inertia ellipsoid is a sphere that is, the
moments of inertia are all equal and the products of inertia are 0.
Theorem 5.3 Let G be a connected strongly regular graph that is not complete
multipartite, and let X be the projection of the vertex-set of G onto a non-trivial
eigenspace, re-scaled to lie on the unit sphere. Then
(a) X is a spherical 2-design.
(b) X is a spherical 3-design if and only if the Kren bound corresponding to this
eigenspace is attained this implies that G satises C(3).
(c) X is a spherical 4-design if and only if the absolute bound is attained this
implies that G satises C(4).
(d) X is never a spherical 5-design.
We saw that the parameters of a graph satisfying C(3) (apart from the pentagon
C
5
) are either of pseudo-Latin square or negative Latin square type, or of Smith
type. Such a graph gives a spherical 3-design if and only if it attains the Kren
bound. All Smith graphs, and also the pentagon, attain this bound, but no graph of
pseudo-Latin square type does. For graphs NL
r
(n) of negative Latin square type,
a 3-design arises if and only if n = r(r 3). Only two such graphs are known, the
Clebsch graph on 16 vertices (r = 1) and the Higman-Sims graph on 100 vertices
(r = 2).
We obtain a spherical 4-design in the smaller eigenspace if and only if the graph
G is C
5
or a so-called extremal Smith graph. Two examples of extremal Smith
graphs are known, the Schl ai graph on 27 vertices and the McLaughlin graph on
275 vertices.
Information on the geometry of eigenspaces for more general classes of graphs
is given in Chapter 3; see also Godsil [15].
6. Rank 3 graphs
Looking again at the picture of the Petersen graph with which we began, we see
that it has ve-fold symmetry, and indeed has the symmetry of a regular pentagon
(the dihedral group of order 10). In fact, there is more symmetry that is not visible
in the diagram. This graph has a well-known representation as the complement
of the line graph of K
5
that is, the vertices can be labelled with the 2-element
subsets of {1, 2, 3, 4, 5] in such a way that two vertices are adjacent if and only
if their labels are disjoint. Now the symmetric group S
5
, in its induced action on
the vertex labels, acts as a group of automorphisms of the graph. It is not hard to
show that S
5
is the full automorphism group. Moreover, S
5
acts transitively on the
set of adjacent pairs of vertices and on the set of non-adjacent pairs of vertices.
A graph G is a rank 3 graph if it admits a group G of automorphisms with the
property that G acts transitively on the set of vertices, on the set of ordered pairs of
adjacent vertices, and on the set of ordered pairs of non-adjacent vertices. The term
comes from permutation group theory, where the rank of a transitive permutation
group G on a set is the number of orbits of G on the set of ordered pairs of
elements of . In the case of a rank 3 graph, with = V(G), the three orbits are
{(v, v) : v V(G)], {(v, w) : v w] and {(v, w) : v ,= w, v w].
Theorem 6.1
(a) Every rank 3 graph is strongly regular.
(b) Let G be a transitive permutation group with rank 3 and even order. Then
there is a rank 3 graph G admitting G as a group of automorphisms.
Proof
(a) This follows from the fact that the number of neighbours of a vertex (or
common neighbours of a pair of vertices) is the same as the number of
(common) neighbours of any image under an automorphism.
(b) The group G has just two orbits O
1
and O
2
on ordered pairs of distinct
elements of . Now, for any orbit O, the set O
= {(w, v) : (v, w) O] is
also an orbit, so either O
1
= O
1
or O
1
= O
2
. However, since G has even
order, it contains an element g of order 2 that interchanges two points v, w of
. If (v, w) O
i
, then O
i
= O
i
, so the rst alternative holds.
Now let the graph G have vertex-set , with v w whenever
(v, w) O
1
. Our argument shows that the graph is undirected: clearly it
admits G as a rank 3 group of automorphisms.
This is a special case of the general construction of G-invariant graphs in Chap-
ter 5. Note that by this construction the orbits O
1
and O
2
give rise to complementary
strongly regular graphs.
A major result in permutation group theory, which relies heavily on the clas-
sication of nite simple groups, is the determination of all rank 3 permutation
groups. We outline the argument here.
Let G be a rank 3 permutation group on . Recall that G is imprimitive if
G preserves a non-trivial equivalence relation, and primitive otherwise. Now, if
G is imprimitive, and is the equivalence relation preserved by G, then the
sets
{(v, w) : v w, v ,= w] and {(v, w) : v , w]
are G-invariant, and so must be the two G-orbits on pairs of distinct points. The cor-
responding graphs are disjoint unions of complete graphs and their complements,
so we may assume that G is primitive.
The basic analysis of such a group is done by considering the socle of G, the
product of its minimal normal subgroups. It follows fromthe ONan-Scott theorem
that one of three possibilities must occur for the socle N of G (see Chapter 7,
Section 3):
(a) N is an elementary abelian group and acts regularly;
(b) N is a non-abelian simple group;
(c) N is the direct product of two isomorphic non-abelian simple groups.
In case (a), because its action is regular, N can be identied with the set of points
permuted, and is the additive group of a vector space V over the eld F
p
, for some
prime p. The subgroup Hthat xes the origin is a group of linear transformations
of V, with two orbits X
1
and X
2
. In our case, the orbits satisfy X
1
= X
1
and
X
2
= X
2
, and the complementary graphs G
1
and G
2
have vertex-set V and
satisfy v w in G
i
if and only if v w X
i
. So the classication in this case
is reduced to nding groups of matrices over F
p
with just two orbits (each closed
under negation) on non-zero vectors. Examples include the following:
r
The multiplicative group of the non-zero squares in F
q
, where q 1 (mod 4):
the orbits are the sets of squares and non-squares in F
q
, and both graphs (which
happen to be isomorphic) are the Paley graph P(q).
r
The orthogonal group preserving a non-degenerate quadratic form Q over F
2
:
the orbits are the sets of non-zero vectors v satisfying Q(v) = 0 or 1. Such
forms can be dened on spaces of even dimension, and there are just two
inequivalent forms; for example, in dimension 4, the quadratic forms
x
1
x
2
x
3
x
4
and x
1
x
2
x
2
3
x
3
x
4
x
2
4
give the graphs L
2
(4) (and its complement) and the Clebsch graph (and its
complement), respectively. In general, these graphs occur among Thomasons
pseudo-random graphs [28]. They are of pseudo- or negative Latin square type,
and satisfy C(3).
The complete list of linear groups with two orbits on non-zero vectors was deter-
mined by Liebeck [19].
In case (b), where the socle N of G is non-abelian and simple, the classica-
tion of nite simple groups shows that it must be an alternating group, a group
of Lie type, or one of the twenty-six sporadic groups. Moreover, the ONan-
Scott theorem gives the extra information that G lies between N and its auto-
morphism group, where N is embedded in Aut(N) as the group of inner auto-
morphisms.
The combined efforts of a number of mathematicians, including Bannai, Kantor,
Liebler, Liebeck and Saxl, have determined all rank 3 actions of almost simple
groups that is, groups lying between a simple group and its automorphismgroup.
Examples include the following.
r
The symmetric group S
n
(for n 5) acts on the set of 2-element subsets of
{1, 2, . . . , n] giving the triangular graph T(n) and its complement.
r
The projective group PGL(n, q) (for n 4) has a rank 3 action on the set of
lines of the projective space: the orbits are the sets of intersecting pairs and
skew pairs of lines.
r
A classical group (one preserving a polarity of a projective space) acts on the
set of self-polar points of the projective space: these form the polar space
associated with the polarity; the action has rank 3, except in a few low-
dimensional cases where it is doubly transitive. For a few cases involving small
elds, the action on the non-self-polar points also has rank 3.
r
There are also various sporadic examples, such as PSU(3, 5
2
) acting on the
vertices of the Hoffman-Singleton graph, or the Higman-Sims group acting on
the vertices of the Higman-Sims graph.
Several of the sporadic simple groups were rst constructed as groups of auto-
morphisms of strongly regular graphs: these were the Hall-Janko, Higman-Sims,
McLaughlin, Suzuki, Fischer and Rudvalis groups.
In case (c) of the ONan-Scott theorem, the socle N of G is the direct product
of two isomorphic simple groups. The analysis leading to this case shows that the
rank 3 graphs that arise are the lattice graphs L
2
(n) and their complements.
7. Related classes of graphs
There are many generalizations and variants of strongly regular graphs. In this
section we introduce a few of these: distance-regular graphs, association schemes,
walk-regular graphs, edge-regular graphs, Deza graphs and strong graphs.
A connected graph G of diameter d is distance-regular if there are constants
c
i
, a
i
and b
i
, for 0 i d, such that, if u and v are vertices at distance i , then the
numbers of vertices w such that w v and w is at distance i 1, i and i 1 from
u are c
i
, a
i
and b
i
, respectively. The numbers c
i
, a
i
and b
i
are the parameters of
the graph; note that c
0
, a
0
and b
d
are 0.
Every distance-regular graph is regular with degree b
0
= k. Also,
c
i
a
i
b
i
= k for all i, and c
1
= 1;
thus there are 2d 3 independent parameters. A distance-regular graph of di-
ameter 2 is the same thing as a connected strongly regular graph and then = a
1
and = c
2
.
A connected graph G is distance-transitive if there is a group G of automor-
phisms of G such that, for any two pairs (v
1
, w
1
) and (v
2
, w
2
) of vertices satisfying
d(v
1
, w
1
) = d(v
2
, w
2
), there is an automorphism g G that maps v
1
to v
2
and w
1
to w
2
. Any distance-transitive graph is distance-regular, and a distance-transitive
graph of diameter 2 is the same thing as a connected rank 3 graph.
The determination of all distance-transitive graphs is not yet complete: this
class of graphs is discussed further in Chapter 9. Further information about
distance-regular and distance-transitive graphs can be found in Brouwer, Cohen
and Neumaier [6].
Many distance-regular graphs are not distance-transitive, but such graphs be-
come less common as the diameter increases. However, as shown by Egawa
[12], there exist distance-regular graphs of arbitrarily large diameter that are not
distance-transitive.
The adjacency matrix of a regular connected graph of diameter d has at least
d 1 distinct eigenvalues, one of which is the degree, and distance-regular graphs
attain this bound. As we have seen, a regular connected graph has precisely three
eigenvalues if and only if it is strongly regular. However, for d 3, there are regular
connected graphs of diameter d with d eigenvalues that are not distance-regular.
The rst examples were constructed by Bridges and Mena [4], but the study of this
interesting class of graphs has not yet progressed much beyond the collection of
examples.
Looking more closely at distance-regular graphs, we can show that there are
constants p
k
i j
, for 0 i, j, k d, with the property that, given vertices u and v
with d(u, v) = k, the number of vertices w such that d(u, w) = i and d(w, v) = j
is precisely p
k
i j
. We can generalize this as follows.
Suppose that the ordered pairs of points of a set are partitioned into s 1
classes C
0
, C
1
, . . . , C
s
with the following properties:
the diagonal {(v, v) : v ] is a single class C
0
;
each class C
i
is symmetric (that is, if (u, v) C
i
then (v, u) C
i
);
given i, j, k {1, 2, . . . , s] and (u, v) C
i
, the number of w such that
(u, w) C
j
and (w, v) C
k
depends only on i, j and k, and not on (u, v).
Such a structure is called an association scheme. Thus, any distance-regular graph
gives rise to an association scheme. More about association schemes can be
found in Bannai and Ito [2], Godsil [15] and Bailey [1]. A still more general
concept is a coherent conguration, where the relations are not required to be
symmetric.
Association schemes were originally used in experimental design by Bose and
his school. Suppose that an experiment is being performed on a number of exper-
imental units that are divided into b blocks of size k for example, k elds on
each of b farms, or k patients in each of b hospitals. We want to apply a number
v of different treatments in such a way that no treatment occurs more than once
in the same block. It is clearly a good idea to arrange that any two treatments
occur together in a block the same number of times, if possible; such a design is
called balanced. However, Fisher showed that, if k < v, this is not possible unless
v b: thus, to test more treatments, we must relax the condition of balance. Bose
observed that the best approach is to have an association scheme on the set of
treatments, and to arrange that the number of times that two treatments u and v
occur together in a block depends only on which associate class C
i
contains (u, v).
Such a design is called partially balanced; see [1] for more information.
Indeed, this is a case where the applications preceded the theory, and the gen-
eralization preceded the special case. Partially balanced designs were dened by
Bose and Nair in 1939. During the 1950s, association schemes became of interest
in their own right, but it was not until Boses 1963 paper [3] that the term strongly
regular graph was introduced.
A graph G is walk-regular if, for each non-negative integer i and vertex v, the
number of closed walks of length i starting at v depends only on i , and not on v;
equivalently, a graph is walk-regular if the characteristic polynomials of all of its
vertex-deleted subgraphs are the same. The class of walk-regular graphs includes
both the vertex-transitive graphs and the distance-regular graphs, and is contained
in the class of regular graphs; see Godsil [15] for more about these graphs.
Recall that a strongly regular graph is dened by three conditions:
(a) each vertex has k neighbours;
(b) any two adjacent vertices have common neighbours;
(c) any two non-adjacent vertices have common neighbours.
We can weaken the denition by requiring only two of these conditions to hold.
A graph satisfying (a) and (b) is edge-regular, and a graph satisfying (b) and (c)
is a Deza graph (see [13]). The class of graphs satisfying (a) and (c) has not been
studied, except in special cases.
More systematically, recall that a graph is t -tuple regular (that is, it satises
the property C(t )) if the number of common neighbours of a set S with at most t
vertices depends only on the isomorphism type of the induced subgraph on S. Let
us say that a graph satises property R(t ) if this condition holds for sets S with
[S[ = t ; thus, a Deza graph satises R(2), but not necessarily R(1). It seems that
no systematic study has been made of the possible sets of integers t for which R(t )
can hold in a graph.
Avariant of Deza graphs was introduced earlier by Seidel, who dened a strong
graph to be one with the property that, for any two vertices u and v, the number
of vertices joined to just one of the two depends only on whether or not u and v
are joined. Using a modied adjacency matrix B with 0 on the diagonal, 1 for
adjacent vertices and 1 for non-adjacent vertices, we nd that
(B
1
I)(B
2
I) = (n 1
1
2
)J,
for some integers
1
and
2
. It follows that, if n 1
1
2
,= 0, then the graph is
regular and thus strongly regular. In the remaining case, when n 1
1
2
= 0,
the graph need not be regular; such special strong graphs are closely connected
with regular two-graphs (see below).
The operation
X
of switching a graph G with respect to a set X of vertices is
dened as follows: edges between X and its complement are switched to non-
edges, and non-edges are switched to edges; adjacencies within X or outside X
remain unaltered. Switching with respect to all subsets generates an equivalence
relation on the class of all graphs on a xed vertex-set V. It is easy to see that, if
T is the set of 3-subsets of V that contain an odd number of edges of G, then T
is unaltered by switching. Moreover, a set T of triples arises from a graph in this
way if and only if each 4-set contains an even number of members of T ; such a
set is called a two-graph. Thus, there is a bijection between the set of two-graphs
on V and the set of switching equivalence classes on V.
Switching a graph G has the effect of pre- and post-multiplying the (0, 1, 1)
adjacency matrix of G (dened above) by a diagonal matrix with entries 1 and 1.
The matrix equation
(B
1
I)(B
2
I) = 0,
satised by special strong graphs, is unaffected by this, so if a graph satises this
equation, then so do all graphs in its switching class. In this case the corresponding
two-graph is called regular. Regular two-graphs are also characterized by the
property that any two vertices in V lie in a constant number of triples in T .
There are many connections between regular two-graphs, strongly regular
graphs, sets of equiangular lines in Euclidean space, doubly transitive permu-
tation groups, antipodal distance-regular graphs of diameter 3, and several other
topics. We refer to Seidels surveys [26] and [27].
References
1. R. A. Bailey, AssociationSchemes: DesignedExperiments, AlgebraandCombinatorics,
Cambridge Univ. Press, 2004.
2. E. Bannai and T. Ito, Algebraic Combinatorics I: Association Schemes, Benjamin, 1984.
3. R. C. Bose, Strongly regular graphs, partial geometries, and partially balanced designs,
Pacic J. Math. 13 (1963), 389419.
4. W. G. Bridges and R. A. Mena, Multiplicative cones a family of three eigenvalue
graphs, Aequat. Math. 22 (1981), 208214.
5. A. E. Brouwer, Strongly regular graphs, The CRCHandbook of Combinatorial Designs
(eds. C. J. Colbourn and J. H. Dinitz), CRC Press (1996), 667685.
6. A. E. Brouwer, A. M. Cohen and A. Neumaier, Distance-regular Graphs, Springer,
1989.
7. P. J. Cameron, 6-transitive graphs, J. Combin. Theory (B) 28 (1980), 168179.
8. P. J. Cameron, J.-M. Goethals and J. J. Seidel, Strongly regular graphs with strongly
regular subconstituents, J. Algebra 55 (1978), 257280.
9. P. J. Cameron and J. H. van Lint, Graphs, Codes, Designs and their Links, Cambridge
Univ. Press, 1991.
10. L.-C. Chang, The uniqueness and non-uniqueness of triangular association schemes,
Science Record 3 (1959), 604613.
11. P. Delsarte, J.-M. Goethals and J. J. Seidel, Spherical codes and designs, Geometriae
Dedicata 6 (1977), 363388.
12. Y. Egawa, Association schemes of quadratic forms, J. Combin. Theory (A) 38 (1985),
114.
13. M. Erickson, S. Fernando, W. H. Haemers, D. Hardy and J. Hemmeter, Deza graphs:
A generalization of strongly regular graphs, J. Combin. Design 7 (1999), 395405.
14. D. G. Fon-Der-Flaass, New prolic constructions of strongly regular graphs, Adv.
Geom. 2 (2002), 301306.
15. C. D. Godsil, Algebraic Combinatorics, Chapman & Hall/CRC Press, 1993.
16. A. J. Hoffman, On the uniqueness of the triangular association scheme, Ann. Math.
Stat. 31 (1960), 492497.
17. D. A. Holton and J. Sheehan, The Petersen graph, Austral. Math. Soc. Lecture Series
7, Cambridge Univ. Press, 1993.
18. M. G. Kren, Hermitian-positive kernels, II, Amer. Math. Soc. Transl. (2) 34 (1963),
109164.
19. M. W. Liebeck, The afne permutation groups of rank 3, Proc. London Math. Soc. (3)
54 (1987), 477516.
20. B. D. McKay and E. Spence, see http:/gauss.maths.gla.ac.uk/~ted/
srgraphs.html.
21. H. Van Maldeghem, Generalized Polygons, Birkh auser, 1998.
22. A. Neumaier, Strongly regular graphs with least eigenvalue m, Arch. Math. 33
(1979), 392400.
23. R. E. A. C. Paley, On orthogonal matrices, J. Math. Phys. 12 (1933), 311320.
24. L. L. Scott, Jr., A condition on Higmans parameters, Notices Amer. Math. Soc. 20
(1973), A-97.
25. J. J. Seidel, Strongly regular graphs with (1, 1, 0) adjacency matrix having eigenvalue
3, Linear Alg. Appl. 1 (1968), 281298.
26. J. J. Seidel, A survey of two-graphs, Proc. Internat. Coll. Teorie Combinatorie (Roma
1973), Accad. Naz. Lincei (1977), 481511.
27. J. J. Seidel and D. E. Taylor, Two-graphs: a second survey, Algebraic methods in graph
theory, II, Conf. Szeged 1978, Colloq. Math. J anos Bolyai 25 (1981), 689711.
28. A. Thomason, Random graphs, strongly regular graphs and pseudorandom graphs,
Surveys in Combinatorics 1987 (ed. C. A. Whitehead), London Math. Soc. Lecture
Note Series 123, Cambridge Univ. Press (1987), 173195.
29. W. D. Wallis, Construction of strongly regular graphs using afne designs, Bull. Austral.
Math. Soc. 4 (1971), 4149.
30. R. M. Wilson, Non-isomorphic Steiner triple systems, Math. Z. 135 (1974), 303313.
31. R. M. Wilson, An existence theory for pairwise balanced designs, III: Proof of the
existence conjectures, J. Combin. Theory (A) 18 (1975), 7279.
32. N. C. Wormald, Models of random regular graphs, Surveys in Combinatorics, 1999
(eds. J. D. Lamb and D. A. Preece), London Math. Soc. Lecture Notes Series 267,
Cambridge Univ. Press (1999), 239298.
9
Distance-transitive graphs
ARJEH M. COHEN
1. Introduction
2. Distance-transitivity
3. Graphs from groups
4. Combinatorial properties
5. Imprimitivity
6. Bounds
7. Finite simple groups
8. The rst step
9. The afne case
10. The simple socle case
11. Conclusion
References
In this chapter we investigate the classication of distance-transitive graphs:
these are graphs whose automorphism groups are transitive on each of the
sets of pairs of vertices at distance i, for i = 0, 1, . . . . We provide an in-
troduction into the eld. By use of the classication of nite simple groups,
it seems possible to nd all distance-transitive graphs. Priority is given to
the classication of the so-called primitive ones. We give an account of the
achievements in this direction.
1. Introduction
An automorphism of a graph reects the symmetry of the graph. A high degree of
symmetry can be related to both aesthetics and efciency. It is often the apparent
symmetry of a graph when pictured in the plane or in space that makes viewing it a
pleasant activity. If the graph has a high degree of symmetry, a set of permutations
222
of the automorphism group of the graph can be used to describe the graph rather
succinctly: we just list a fewedges, and assert that the others can be obtained as im-
ages of those given under successive applications of the generating permutations.
Since groups can be generated by relatively small sets of permutations for any
nite simple group, two generators sufce! this representation of a graph can
be much more efcient than listing all the edges. An extreme example is the
complete graph on the vertex-set {1, 2, . . . , n], that can be described by say-
ing that {1, 2] is an edge and that the permutations (1 2) and (1 2 3 . . . n) are
automorphisms.
In this chapter, we are interested in the case where the automorphism group
of the graph is as large as possible. There are various interpretations of largeness,
most of which use the transitivity of the group on sets of vertices sharing (possibly)
common substructures; for instance, we might require that the group be transitive
on the set of all paths in the graph of a certain length. Here we focus on one
particular interpretation of the group being large.
2. Distance-transitivity
Let G be a connected graph of diameter d, and let G be a group of automorphisms
of G. Recall that d(x, y) is the distance between the vertices x and y of G. Then
the group G is distance-transitive on G if it is transitive on each of the relations
G
i
= {(x, y) G G : d(x, y) = i ], where i = 0, 1, . . . , d. Agraph is distance-
transitive if its automorphism group is distance-transitive on it.
As an example, if G = K
n
is the complete graph on {1, 2, . . . , n] and G is the
symmetric group S
n
, then G is clearly transitive on the vertex-set, and hence on
G
0
= {(x, x) : x G]. Moreover, the diameter of G is 1. Let (x, y) G
1
; then
x ,= y. It is readily seen that the pair (x, y) can be mapped to the pair (1, 2) by means
of an element of S
n
for instance, if x ,= 1, 2 and y ,= 1, 2, then the permutation
(1, x)(2, y) works. So S
n
is transitive on G
1
, and hence is distance-transitive.
A bipartite graph G = K
n,n
with parts of size n is also distance-transitive. Any
permutation that interchanges the two parts is an automorphism, and so is any
permutation that preserves both parts. The diameter is 2, and so we have to check
the transitivity of the automorphism group on the three sets G
i
, for i = 0, 1, 2.
If i = 0, the transitivity assertion means the transitivity on the vertex-set, which
is immediate from the presence of the above-mentioned automorphisms. If i = 1,
then we have a pair (x, y) with x in one part and y in the other. But then we
can move this pair to any other pair of this kind by permutations that preserve
both parts. Finally, if i = 2, then we have two points in the same part. They can
be moved simultaneously to the other part, or to any other pair in the part to
224 Arjeh M. Cohen
which they belong. It follows that the automorphism group of K
n,n
is distance-
transitive.
The diameters of the distance-transitive graphs introduced so far have been
restricted to 1 and 2. The following example gives distance-transitive graphs with
arbitrarily large diameters; recall that the stabilizer of a vertex x in a group G is
the subgroup G
x
:= {g G : xg = x] of G.
Example 2.1 Let d and n be positive integers, with d n/2. The Johnson graph
J(n, d) has as its vertex-set the d-subsets of X = {1, 2, . . . , n], with adjacency
dened by x y if and only if [x y[ = d 1. Its diameter is d. The group
G = S
n
is a distance-transitive group of automorphisms on this graph, with vertex-
stabilizer S
d
S
nd
.
For d = 1, we recover the complete graphs. For d = 2, the complement of
a distance-transitive graph, if connected, is again distance-transitive. The graph
J(4, 2) is the complement of three disjoint copies of K
2
, and J(5, 2) is the com-
plement of the well-known Petersen graph of degree 3 (Fig. 1).
Let us take a closer look at distance-transitive groups of automorphisms of
graphs of diameter 1; here the graph is a complete graph K
n
. A permutation group
on a set V is called doubly transitive (or 2-transitive) if it is transitive on the set
{(x, y) V V: x ,= y] of all pairs from V. The following theorem shows that
S
n
need not be the only distance-transitive automorphism group on K
n
.
Theorem 2.1 A group is a distance-transitive group of automorphisms of K
n
if
and only if it is doubly transitive on its vertex-set.
If n 4, then A
n
is a proper and doubly transitive subgroup of the automor-
phism group S
n
of K
n
. There are many more examples. The doubly transitive
permutation groups have been determined by means of the classication of nite
12
34
45
25
23
13
15
14
35
24
Fig. 1.
simple groups (see Cameron [15]), so the distance-transitive groups of automor-
phisms of graphs with diameter d = 1 are all known.
In this smallest diameter case, we encounter some of the properties that
will recur for all d. One of them is the phenomenon that it is easier to classify
distance-transitive graphs than to classify distance-transitive groups. If G admits
a distance-transitive group, then Aut(G) acts distance-transitively; but when G is
distance-transitive, it is harder to establish which subgroups of Aut(G) are still
distance-transitive on G.
Another property valid for all diameters d is that, when G acts distance-tran-
sitively on G, it has permutation rank d 1 on V(G) that is, the number of orbits
of G on V(G) V(G) is d 1; here, each G
i
(i = 0, 1, . . . , d) is an orbit. As we
have seen, this obvious fact is both necessary and sufcient for the case d = 1. It
does not sufce for higher d, but it is close to being sufcient for d = 2.
We will consider graphs G with vertex-set V for which G is a group of auto-
morphisms of G; such a graph is a G-invariant graph structure on V. The next
example shows that a G-invariant graph structure on a vertex-set V constructed
from a transitive permutation representation of G on V need not be connected.
Example 2.2 Consider the permutation group G on the set V = {1, 2, 3, 4, 5, 6]
generated by the three permutations
a = (1 6 3 2 5 4), b = (1 5 3)(2 4 6), c = (2 6)(3 5).
Clearly, G is transitive on V. The stabilizer of 1 in G is the subgroup Hgenerated
by a
2
b = (2 6 4) and c; this can be veried by using the algorithms explained
in Chapter 10. Besides {1], the subgroup H of G has orbits K = {2, 6, 4] and
L = {3, 5] on V. Since G is transitive on V, its permutation rank equals the number
of H-orbits on V, which is 3.
If G is a connected G-invariant graph structure on V, vertex 1 must have a
neighbour. Since Hxes 1 and consists of automorphisms of G, the presence of a
vertex x in the neighbourhood of 1 implies that all vertices in the H-orbit of x are
adjacent to 1. Thus, apart fromthe complete graph on V, there are two possibilities:
either K or L is the set of all neighbours of 1.
If we let K be the neighbours of 1, then the whole graph G is determined
by the fact that two vertices are adjacent whenever they are in the same G-orbit
as (1 x), for some (and hence all) x K. We nd the other graph by taking L
instead of K. The two graphs are shown in Fig. 2. The graph dened by L is
the disjoint union of two 3-cycles, {1, 3, 5] and {2, 4, 6]. It turns out that the
diameter of a G-invariant graph structure is not necessarily e 1, if e is the
permutation rank. The graph dened by K is the complement of the graph de-
ned by L; it is the complete bipartite graph K
3,3
on V, whose parts are {1, 3, 5]
226 Arjeh M. Cohen
5
3 1
6
4 2
5
2
4
6
3
1
K: L:
Fig. 2.
and {2, 4, 6]. The diameter of K
3,3
is 2, which is one less than the permutation
rank.
The construction of the graph G from the group G using the subgroup orbit K
works well, because the G-orbit of a pair (1, k) with k K is self-paired that is,
it also contains (k, 1). Self-pairedness is necessary for an orbit on pairs to dene
the adjacency of a graph structure, since otherwise the graph denition would lead
to a directed graph.
The group G has size 36, whereas the full automorphism group of K
3,3
has
order 2 6
2
= 72 (see Chapter 5, Theorem 1.1), so this is another example of a
proper distance-transitive subgroup of the full automorphism group of a graph.
In conclusion, when we build distance-transitive graphs on a vertex set V by
use of a permutation group on V, we need criteria for self-pairedness of orbits on
pairs and for connectedness of the graph in terms of the group. This will be the
topic of the next section.
3. Graphs from groups
Applying the denition of distance-transitivity with i = 0 and i = 1, respectively,
we nd that a distance-transitive group G of automorphisms on a graph G is
transitive on both the vertex-set V(G) and on the edge-set E(G) of G. It can
therefore be described in terms of a subgroup H of G and an element r of G. To
see this, take H = G
v
, the stabilizer in G of a vertex v of G, and choose r G in
such a way that vr is adjacent to v. Then G is isomorphic (with equivalence of the
G-actions) to the graph G(G, H, r), whose vertex-set is G/H (the set of cosets of
H in G) and whose adjacency is given by Hx Hy when y HrHx.
Theorem 3.1 Let G be a group of automorphisms of a graph G. If G is transitive
on the vertex-set and the edge-set of G and v is a vertex, then G is isomorphic to
G(G, H, r) where H = G
v
and r G is chosen so that {v, vr] is an edge of G.
In fact, the isomorphismbetween G and G(G, H, r) can even be chosen to preserve
the G-actions on the vertex-sets.
In the identication of V = {1, 2, . . . , 6] of Example 2.2 with cosets of H in
G, the coset Hr with r = a
2
corresponds to 3, a member of L. The subgroup of
G generated by H and r does not coincide with G, since it leaves invariant the
subset {1, 3, 5] of V, whereas G is transitive on V. This implies that the graph on
V built using L cannot be connected: the connected component containing 1 does
not reach an even-labelled vertex.
Suppose that we are given a permutation group G on a vertex-set V, a subgroup
Hof G, and an element r G. Then the above-dened construction for G(G, H, r)
leads to a directed graph that need not be a graph: it is a graph if and only Hx Hy
implies that Hy Hx, for all x, y G. This is readily seen to be equivalent to
self-pairedness of the G-orbit on (G/H) (G/H) containing (H, Hr); in turn, this
is equivalent to r
1
HrH. In Example 2.2, taking r = a, we nd that r
1
=
(a
2
b)r(a
2
b) HrH, so the G-orbit containing (1 1a) = (1 6) is self-paired and
G(G, H, a) is indeed a graph. In conclusion, we have the following criteria for
connectedness of G(G, H, r) (disregarding the directions of the edges) and for
self-pairedness of the adjacency orbit.
Theorem 3.2 Let G be a group with subgroup H, and let r G. Let G be the
directed graph whose vertices are the cosets Hg, for g G, and in which two
vertices Hx, Hy are adjacent if and only if y HrHx. Then:
r
G is connected if and only if H, r) = G;
r
G is undirected if and only if r
1
HrH.
Apart from the anomalies illustrated by the above example, the picture is satis-
factory for d = 2 in that transitive groups with permutation rank 3 give rise to
distance-transitive group actions on graphs with diameter 2.
Theorem 3.3 Let G act transitively on a set V. Then G is distance-transitive on a
G-invariant graph structure on V with diameter 2 if and only if it is transitive of
rank 3 on V, with self-paired orbits on pairs.
In view of Theorem 3.2, the only part of the proof that needs attention is the
existence of a connected G-invariant graph structure in the if part. If G has per-
mutation rank 3 on a subgroup H, then there are vertices r, s G such that
G = H HrH HsH.
Suppose that all orbits on pairs are self-paired. Then, by Theorem 3.2, both
G(G, H, r) and G(G, H, s) are well-dened graphs. Since they are complements
of each other, at least one of them is connected.
As a consequence of the classication of nite simple groups, all rank 3 permu-
tation groups are known (see [13, p. 229]). We next showthat a rank 4 permutation
group does not necessarily lead to a graph on which it acts distance-transitively.
228 Arjeh M. Cohen
Example 3.1 Consider the Mathieu group M
11
on 11 letters, the smallest sporadic
simple group. It can be realized as the permutation group on = {1, 2, . . . , 11]
generated by the permutations
(1 2 3 4 5 6 7 8 9 10 11) and (3 7 11 8)(4 10 5 6).
The M
11
-orbit V of the 6-subset v = {1, 11, 6, 9, 3, 7] of has 66 elements. We
shall analyse the M
11
-invariant graph structures on V. The stabilizer Hin M
11
of v
has size 120 (in accordance with [M
11
[ = 66 120 = 7920), by the famous result
of Lagrange that the cardinality of an orbit is the quotient of the group order and
the stabilizer order. The subgroup H is transitive on v and on v, and is easily
identied with the symmetric group S
5
.
Let V
i
be the set of elements w V for which [v w[ = i . Then
[V
1
[ = 0, [V
2
[ = 15, [V
3
[ = 20, [V
4
[ = 30, [V
5
[ = 0.
Moreover, H is transitive on each of V
2
, V
3
and V
4
, and so M
11
has a transitive
permutation representation on V of permutation rank 4. Since all H-orbit sizes are
distinct, each orbit on pairs must be self-paired. We can therefore dene three non-
trivial M
11
-invariant graph structures on V, according as V
2
, V
3
or V
4
is chosen as
the set of neighbours of v. Denote the resulting graphs by G
2
, G
3
and G
4
. In Fig. 3
we record some basic properties of these graphs. On the left, each of the diagrams
has a small circle, representing the vertex v of G. The larger circles represent the
other H-orbits on V. The number inside each circle is the size of the orbit.
The numbers next to a circle are to be thought of as being attached to a loop
on the circle. The remaining numbers are attached to arrows from one circle to
another. The number f attached to an arrow from circle i to circle j indicates that
there are precisely f vertices of the H-orbit in circle j that are adjacent to a xed
vertex in circle i . We can reformulate this as
f = [G
j
(v) G
1
(w)[ for w G
i
(v).
As suggested by the denition, this number does not depend on the choice of the
vertex w in circle i . For example, the number 10 above the circle containing the
0
5
4
4
4
4
1
1
1
4
8
9
12
12
4
6
9
3
3
3
6
6
6
10
10
10
10
10
30
30
30
30
20
20
20
20
15
15
15
15
15
15
Fig. 3.
number 20 in the middle diagram indicates that a neighbour of v is adjacent to ten
other neighbours of v.
Several important properties of G can be read off directly from these so-called
collapsed adjacency diagrams: the terminology is very close to Soichers termi-
nology (see Chapter 10) in the software GRAPE for the analysis of graphs with
group action in the software system GAP; compare with the double coset diagram
in [13]. One of the key properties of a distance-transitive graph is the fact that,
if G is such a graph, then the collapsed adjacency diagram of G is a path. For
example, the collapsed adjacency diagram of the graph G of Example 2.2 is as
shown in Fig. 4. It is a path of length 2, in accordance with the fact that the graph
is distance-transitive with diameter 2.
The graphs of Example 3.1 are connected and have diameter 2, but they are not
paths. Hence, there is no M
11
-invariant graph structure on V of which M
11
is a
distance-transitive group of automorphisms.
Example 3.2 The Hamming graph H(d, q) has vertex-set V = Z
d
q
, where two
vertices are adjacent if they differ in exactly one coordinate. The wreath product
S
q
: S
d
is a semi-direct product of d copies of the symmetric group S
q
(each acting
on a single component) and S
d
(permuting the coordinates of the vertices), and is
a distance-transitive group of automorphisms of H(d, q). The diameter of H(d,q)
is d, and its collapsed adjacency diagram is shown in Fig. 5.
Example 3.3 Let V be the vector space of all homogeneous polynomials over
F
3
in two variables x, y of degree 2; then V has dimension 3. Dene a graph
G on V by decreeing that two members of V are adjacent if and only if they
differ by an irreducible polynomial. Since a polynomial f in V is irreducible
if and only if f is, this denes a graph on V. The neighbours of 0 in G are
x
2
y
2
, x
2
xy y
2
, x
2
xy y
2
, and their negatives; thus, G(0) has 6 ver-
tices. We shall write G
i
for the graph on V whose edges are those pairs of vertices
1
3
3
3
2
2
Fig. 4.
d
q 1
(q 1)
d
d(q 2)
d(q 1)
d( q 1)
1
(
d
) (q 1)
2
(d 1)(q 1)
2(q 2)
2
q 2
2
Fig. 5.
230 Arjeh M. Cohen
x, y V with d(x, y) = i , where the distance d(x, y) is taken in G. Forming all
differences of irreducible polynomials, we see that G
2
(0) consists of all polynomi-
als in V that are products of two distinct linear factors, so [G
2
(0)[ = 12. Similarly,
all distinct triples of elements of G
1
(0) yield squares of linear terms and their
negatives, so [G
3
(0)[ = 8. Since 1 6 12 8 = 27 = [V[, we have accounted
for all members of V.
Let us now consider the group G of automorphisms of G. The group of trans-
lations of V preserves G and is transitive on V, so each vertex has degree 6.
The stabilizer G
0
of 0 in G(= Aut(G)) contains the linear transformations induced
from a projective linear transformation in x and y. More explicitly, if the latter is
determined by x . ax by and y . cx dy, with ad bc = 1, then the lin-
ear transformation of V determined by x
2
. a
2
x
2
abxy b
2
y
2
, xy . acx
2
(ad bc)xy bdy

2
and y
2
. c
2
x
2
cdxy d
2
y
2
, belongs to G
0
. Moreover,
homotheties in x and y (that is, linear transformations with b = c = 0 and a = d)
act trivially, so PGL(2, 3) acts, and the homotheties on V belong to G
0
. So G
contains the semidirect product of the additive group V and the linear group
PGL(2, 3) 1). It follows directly that G
0
is transitive on G
i
(0), for i = 0, 1, 2,
and hence that G is distance-transitive on G. The collapsed adjacency diagram of
G is shown in Fig. 6. In fact, G is isomorphic with the Hamming graph H(3, 3),
and so its full automorphism group G is the wreath product S
3
: S
3
, of size 6
4
. By
an order argument, this is isomorphic with the group 3
3
(PGL(2, 3) 1)).
4. Combinatorial properties
Distance-transitivity is naturally associated with the presence of a group of
automorphisms. Nevertheless, we can derive some properties of distance-transitive
graphs that are purely combinatorial. Here are the most prominent of these
properties.
Theorem 4.1 If G is a distance-transitive graph with diameter d, then there are
non-negative integers
a
1
, a
2
, . . . , a
d
, b
0
, b
1
, . . . , b
d1
, c
1
, c
2
, . . . , c
d
, k
1
, k
2
, . . . , k
d
1
1
1
6
4
2
2
2
3
3
12 8
6
Fig. 6.
such that the collapsed adjacency diagram of G is
a
1
1
a
2
c
2
c
3
c
d1
c
d
a
d1
a
d
k
1
k
1
b
1
b
2
b
d2
b
d1
k
d1
k
d
k
2
These numbers satisfy the following relations:
r
b
0
= k
1
, c
1
= 1;
r
a
i
b
i
c
i
= k
1
;
r
k
i 1
b
i 1
= k
i
c
i
, for i = 1, 2, . . . , d.
In particular, they are all determined by the array
(b
0
, b
1
, . . . , b
d1
; c
1
, c
2
, . . . , c
d
).
The array mentioned at the end of this theorem is called an intersection array.
A graph is distance-regular if its collapsed adjacency diagram can be drawn in
this manner, with one circle of vertices for each distance from a xed vertex v,
and with the same intersection array for each vertex. Recall from Example 3.3
that G
h
is the graph on the vertex-set of G in which two vertices are adjacent if
and only if they are at distance h in . If G is a graph for which the cardinality
[G
i
(v) G
j
(w)[ does not depend on the choice of the edge {v, w] of G
h
, then G
is distance-regular. In this case, one usually writes p
h
i j
for [G
i
(v) G
j
(w)[.
Such a graph is regular; we usually denote its degree by k, so k = k
1
. If G is
distance-regular, then the numbers p
h
i j
are uniquely determined by the numbers
b
i
= p
i
i 1,1
and c
i 1
= p
i
i 1,1
, for i = 0, 1, . . . , d 1. For example, k = p
0
1,1
=
b
0
, a
i
= k b
i
c
i
, p
h
i, j
= 0 if i j < h, and p
i
i 2,2
= c
i
c
i 1
/c
2
.
The Petersen graph (see Example 2.1) is distance-regular with intersection array
(3, 2; 1, 1); its diagram is shown in Fig. 7.
These combinatorial properties hold in graphs that admit distance-transitive
groups of automorphisms.
Theorem 4.2 Each distance-transitive graph is distance-regular.
1
3
3
2
1
2
6
Fig. 7.
232 Arjeh M. Cohen
Fig. 8.
The result is easily proved, but the converse statement is false. Figure 8 shows a
counter-example due to Adelsson-Velskii, Vesfeiler, Leman and Farad zev (see
[13, p. 136]). It is a distance-regular graph on 26 vertices with intersection array
(10, 6; 1, 4). It has an automorphismof order 13 that can be visualized as a rotation
of the picture, but has no distance-transitive group.
For an algebraic description of this graph, we write V = {a
i
, b
i
: i = 1, 2, . . . ,
13], and dene adjacency on V by
a
i
a
j
[i j [ {1, 3, 4],
b
i
b
j
[i j [ {2, 5, 6],
a
i
b
j
i j {0, 1, 3, 9].
For each diameter larger than 1, distance-regular graphs that are not distance-
transitive are known (see [13, p. 290]).
The intersection array of the Johnson graph J(n, d) is
_
d(n d), (d 1)(n d 1), . . . , (n 2d 1); 1, 4, 9, . . . , d
2
_
.
The group G = S
n
acts distance-transitively on this graph, with vertex-stabilizer
S
d
S
nd
. This illustrates the remarkable fact that a single permutation repre-
sentation of a group G on V may correspond to more than one group-invariant
distance-transitive graph structure on V. To see this, let n = 2d 1 and dene
the odd graph O
d1
to be the graph whose vertex-set consists of the d-subsets
of X = {1, 2, . . . , n], and where adjacency is dened by x y if and only if
x y = that is, the vertices x, y are at distance d in J(2d 1, d). Then O
d1
is a distance-transitive graph with diameter d and with intersection array
(d 1, d, d, d 1, d 1, . . . , {(d 1)/2; 1, 1, 2, 2, 3, . . . , (d 1)/2).
For d = 2, the intersection array is (3, 2; 1, 1), and we recover the Petersen graph.
The permutation action of S
2d1
on the vertex-set of O
d1
is the same as that on
J(2d 1, d).
5. Imprimitivity
Although we are primarily interested in distance-transitive graphs, some notions
can be treated entirely within the wider context of distance-regular graphs. An
example of such a notion is imprimitivity, with which we deal next. For a permu-
tation group G, it means that the stabilizer of a vertex is not a maximal subgroup
of G.
Recall that G
i
denotes the relation being at distance i on the vertex-set of
G; thus, (v, w) G
i
and v G
i
(w) both mean that v is at distance i from w in
G. The following beautiful characterization of distance-regular graphs that are
imprimitive in a combinatorial sense is due to Smith (see [13]).
Theorem 5.1 Let G be a connected distance-regular graph with diameter d and
intersection array (b
0
, . . . , b
d1
; c
1
, . . . , c
d
), and let I be a minimal proper subset
of {0, 1, . . . , d] for which
i I
G
i
is an equivalence relation with I ,= {0]. Then
at least one of the following properties holds:
r
G is antipodal that is, for each vertex v, the distance between any two vertices
from G
d
(v) is d, and b
i
= c
di
for all i ,= d/2; in this case, I = {0, d];
r
G is bipartite that is, a
1
= a
2
= = a
d
= 0, c
i
= k b
i
for i < d, and
c
d
= k; in this case, I = {0, 1, . . . , d] 2Z.
If G acts distance-transitively on G, and if the stabilizer H in G of the vertex
v is contained in a proper subgroup K of G, then the K-orbit of v is a union of
H-orbits and is thus of the form

i I
G
i
(v), for some subset I of {0, 1, . . . , d].
If I ,= {0], then K strictly contains H. Note that I is a proper subset, since K is a
proper subgroup of G containing H. Thus,
i I
G
i
is an equivalence relation on
V(G), and so the above theorem applies.
Conversely, if I is a proper subset of {0, 1, . . . , d] for which

i I
G
i
is an
equivalence relation on V(G), then the stabilizer of the subset

i I
G
i
(v) is a
proper subgroup of G containing H. It coincides with H if and only if I = {0].
Consequently, we have the following result.
Corollary 5.2 Let G be a group acting distance-transitively on a connected graph
G. Then G is imprimitive if and only if G is antipodal or bipartite.
234 Arjeh M. Cohen
The classication problemof distance-transitive graphs can be reduced to primi-
tive (that is, non-imprimitive) distance-transitive graphs, by the following beautiful
result and Theorem 5.4 below.
Theorem 5.3 Suppose that G acts distance-transitively on the connected graph G
with diameter d.
r
If G is antipodal, then G acts distance-transitively on the graph whose vertices
are the equivalence classes of G
0
G
d
, and in which two vertices are adjacent
whenever they contain adjacent vertices in G.
r
If G is bipartite, then G acts distance-transitively on each of the two graphs
obtained from G by taking one of the bipartite classes, and letting two vertices
be adjacent wherever they are at distance 2 in G.
For imprimitive distance-regular graphs, these two constructions lead to distance-
regular graphs. They are called the antipodal quotient in the antipodal case, and
the halved graphs in the bipartite case.
An example of an antipodal quotient comes from the Johnson graph J(2d, d).
The complement in {1, 2, . . . , 2d] of a vertex v is the unique vertex of J(2d, d)
at distance d from v. Thus, G
0
G
d
is an equivalence relation, and J(2d, d) is
antipodal; the antipodal quotients of J(4, 2) and J(6, 3) are the complete graphs
K
3
and K
10
, respectively, The Hamming graph H(d, q) is antipodal if and only if
q = 2, in which case it is also bipartite.
To each imprimitive distance-regular graph G, we can apply a nite sequence
of halvings and antipodal quotients, until we arrive at a primitive distance-regular
graph . We say that G is associated with .
Theorem 5.4 For each primitive distance-regular graph of degree k > 2, the num-
ber of associated imprimitive distance-regular graphs is nite.
The proof of this result is not complicated. As a consequence, the classications of
distance-regular graphs and of distance-transitive graphs can be thought of as two-
stage procedures: rst nd all primitive examples, and then, for each individual
primitive example, determine all the associated imprimitive examples. In what
follows, we focus on the rst stage for distance-transitive graphs. Much has been
done regarding the second stage; see, for example, [13], [25], [37], [17] and [21].
In [21], many interesting examples of antipodal covers of complete graphs can
be found. Complete bipartite graphs K
k,k
with parts of equal size k are distance-
transitive with intersection array (k, k 1; 1, k); antipodal covers of K
k,k
are stud-
ied in [25] and [17].
6. Bounds
The intersection array provides a very useful set of numerical invariants for the
classication of distance-regular graphs; for example, it species all the eigenval-
ues of the adjacency matrix and their multiplicities. We present a selection of the
many inequalities satised by the numbers in the array. As usual, for a graph G,
a vertex v and a distance i , we write k
i
for the cardinality of G
i
(v), the set of
vertices at distance i from v.
Theorem 6.1 Let G be a distance-regular graph with diameter d and degree k.
Then its intersection array (b
0
, b
1
, . . . , b
d1
; c
1
, c
2
, . . . , c
d
) satises the following
conditions:
r
1 = c
1
c
2
c
d
;
r
k = b
0
b
1
b
d1
;
r
if i j d, then b
i
c
j
;
r
c
2
k 2b
1
;
r
if G contains a 4-cycle, then c
i 2
b
i 2
c
i 1
b
i 1
a
1
2, for
i d 3;
r
there exist i and j, with i < j, such that
k = k
1
< k
2
< < k
i 1
< k
i
= k
i 1
= = k
j
> k
j 1
> > k
d
;
r
for these i and j, if = k
i
/k
i 1
and = k
i j
/k
i j 1
, then
[V[
_

1

1
j 1
_
k
i
.
If G is a permutation group on V that leaves invariant a distance-transitive graph
structure on V, then often the only parameters besides [V[ , and sometimes d, that
we can evaluate among those relevant for the intersection array are the numbers
k
i
and frequently just a few of them. In this light, the above inqualities for the k
i
are important when nding the graphs on which G acts distance-transitively.
Recall the denition of a group character: this is the function that assigns to a
group element the trace of a matrix of the element in a given linear representation
of the group. Agroup character is irreducible if the representation is irreducible,
and real if the representation can be dened over the reals.
Theorem 6.2 Suppose that G is distance-transitive on a graph G with diameter
d, with stabilizer H. Then
[H[
_
[G[/(d 1)
_
[G[/(r 1),
where r is the number of irreducible real characters of G. If, moreover, G contains
a 4-cycle, then [H[

[G[/5.
236 Arjeh M. Cohen
The latter part follows directly from the last two parts of Theorem 6.1. The proof
of the rst part rests on the powerful observation below. A group character is
multiplicity-free if no irreducible character occurs more than once in the decom-
position of as a sum of irreducibles. We often say that the representation itself
is multiplicity-free if its character is.
Theorem 6.3 The permutation character of a distance-transitive group action on
a graph is multiplicity-free.
The proof of the theorem follows from the fact that a permutation representation
is multiplicity-free if all of its orbits on pairs are self-paired. The idea of studying
multiplicity-free permutation representations stems from Saxl (see [13, p. 230]).
Baddeley [3] proved an ONan-Scott type theorem for multiplicity-free permuta-
tion representations, which essentially reduces the study of primitive permutation
groups with only self-paired orbits to almost simple groups.
An outstanding question is whether there are only a nite number of distance-
regular graphs of a xed degree k > 2 (see [14]). If, in addition to the degree, the
so-called geometric girth is also xed, then the diameter is known to be bounded
(see [13, p. 184]). This result was used by Weiss to obtain a bound for distance-
transitive graphs of degree k > 2: the diameter does not exceed (k
6
)!2
2k
(see [13,
Cor. 7.3.2]). This implies immediately the following result.
Theorem 6.4 There are only a nite number of distance-transitive graphs of any
given degree k > 2.
The condition k > 2 is necessary in view of polygons that can have arbitrarily
large diameter. For k 13, the explicit list of distance-transitive graphs is known
see [13], for details and other references. Despite the fact that the upper bounds
on the diameter are astronomical, the result indicates that there is some hope for a
classication of distance-transitive graphs. The real progress in the classication
of primitive distance-transitive graphs, however, is due to the classication of nite
simple groups, as we shall see in the next sections.
7. Finite simple groups
In this section, we summarize the classication of nite simple groups. Aclassical
nite simple group is a simple group, obtained from a group H of linear transfor-
mations on a nite vector space V that leave invariant a quadratic or sesquilin-
ear form on V by taking the quotient of the commutator subgroup of H by its
centre.
Table 1.
Name Condition Order
A
n1
(q) n 2 q
n(n1)
2
(q
2
1)(q
3
1) . . . (q
n
1)/ gcd (n, q 1)
2
A
n1
(q) n 3 q
n(n1)
2
(q
2
1)(q
3
1) . . . (q
n
(1)
n
)/ gcd (n, q 1)
B
n
(q) n 2 q
n
2
(q
2
1)(q
4
1) . . . (q
2n
1)/ gcd (2, q 1)
C
n
(q) n 3 q
n
2
(q
2
1)(q
4
1) . . . (q
2n
1)/ gcd (2, q 1)
D
n
(q) n 4 q
n
2
n
(q
2
1)(q
4
1) . . . (q
2n2
1)(q
n
1)/ gcd (4, q
n
1)
2
D
n
(q) n 4 q
n
2
n
(q
2
1)(q
4
1) . . . (q
2n2
1)(q
n
1)/ gcd (4, q
n
1)
Let q be a prime power, and let n ( 2) be an integer.
If the formis identically zero, and if V = F
n
q
, then the classical group is denoted
by A
n1
(q), and if the form is non-degenerate hermitian and V = F
n
q
2
, then the
classical group is denoted by
2
A
n1
(q).
If the form is non-degenerate and quadratic, and if V = F
2n
q
, then the classical
group is denoted by
2
D
n
(q), if the Witt index, the dimension of a maximal linear
subspace of V on which the form vanishes identically, is n 1, and by D
n
(q), if
the Witt index equals n.
If the formis non-degenerate and quadratic, and if V = F
2n1
q
, then the classical
group is denoted by B
n
(q).
If the form is non-degenerate, anti-symmetric, and bilinear, and if V = F
2n
q
,
then the classical group is denoted by C
n
(q).
Each classical simple group is isomorphic to one of those listed in Table 1.
Except for A
1
(2), A
1
(3),
2
A
2
(2), which are solvable, and for B
2
(2), which is iso-
morphic to S
6
, all of these groups are simple.
There are some further series of simple groups obtained from more elaborate
constructions in multilinear algebra or Lie algebras. When q is a prime power,
these are as shown in Table 2. The commutator subgroups of the groups in the
table are simple, with the exception of
2
B
2
(2), which is solvable. All of the other
groups are isomorphic to their commutator subgroups, with the exception of G
2
(2)
(whose commutator subgroupis isomorphic to
2
A
2
(3)),
2
G
2
(3) (whose commutator
subgroup is isomorphic to A
1
(8)) and
2
F
4
(2) (whose commutator subgroup has
index 2). The simple groups described in this paragraph are the exceptional nite
simple groups.
A group is called of Lie type if it is either classical or exceptional; the prime
divisor of q is called the dening characteristic of the group.
Finally, 26 sporadic groups have been found to exist. Among these are the ve
Mathieu groups (Example 3.1 introduced the smallest of these) and the Monster
group.
238 Arjeh M. Cohen
Table 2.
Name Condition Order
2
B
2
(q) q = 2
2m1
q
2
(q 1)(q
2
1)
3
D
4
(q) q
12
(q
2
1)(q
6
1)(q
8
q
4
1)
E
6
(q) gcd (3, q 1)
1
q
36
m{2,5,6,8,9,12]
(q
m
1)
2
E
6
(q) gcd (3, q 1)
1
q
36
m{2,5,6,8,9,12]
(q
m
(1)
m
)
E
7
(q) gcd (2, q 1)
1
q
63
m{2,6,8,10,12,14,18]
(q
m
1)
E
8
(q) q
120
m{2,8,12,14,18,20,24,30]
(q
m
1)
F
4
(q) q
24
(q
2
1)(q
6
1)(q
8
1)(q
12
1)
2
F
4
(q) q = 2
2m1
q
12
(q 1)(q
3
1)(q
4
1)(q
6
1)
G
2
(q) q
6
(q
2
1)(q
6
1)
2
G
2
(q) q = 3
2m1
q
3
(q 1)(q
3
1)
The above-mentioned simple groups are all nite simple groups, according to
the classication of nite simple groups (see [22] for more details).
Theorem 7.1 Let G be a nite non-abelian simple group. Then G is isomorphic
to A
n
(for n 5), to a classical simple group, to an exceptional simple group, or
to a sporadic simple group.
8. The rst step
The rest of this chapter is devoted to the prospective classication of primitive
distance-transitive graphs. Previous overviews can be found in [24] and [10].
The starting point in the classication of primitive distance-transitive graphs is
the following theorem, proved by Praeger, Saxl and Yokoyama (see [4] and [13,
p. 229]). In view of the determination of all permutation groups with permutation
rank 3 (see [13, p. 229]), for the classication of distance-transitive graphs, we
may assume that d 3. Also, since the only connected distance-regular graphs
with degree k = 2 and diameter d are the 2d-gon and the (2d 1)-gon, we may
also assume that k 3.
Theorem 8.1 Let G be a primitive distance-regular graph with a distance-
transitive group G of automorphisms, and assume that k 3 and d 3. Then one
of the following holds:
(i) G is associated with a Hamming graph;
(ii) G has an elementary abelian normal subgroup which is regular on V(G);
(iii) there is a simple non-abelian normal subgroup N of G for which G embeds
canonically in Aut(N) that is, the centralizer C
G
(N) of N in G is trivial.
The basic ingredient of the proof of Theorem 8.1 is the celebrated ONan-Scott
Theorem (see [2]), which does not use the classication of nite simple groups. In
Case (i ), the graph G is well known, but the possibilities for the group G are not
completely determined it is typically a wreath product. Case (ii) has been dealt
with completely, and will be discussed in Section 9; we refer to it as the afne case.
Case (iii), the simple socle case, is the hardest, since no strong general techniques
have been found for ruling out subgroups of small index as stabilizers of distance-
transitive actions on graphs. In Section 10, we go into more detail on this case.
Let us discuss the distinct cases in terms of subgroups. A subnormal subgroup
of a group G is a subgroup N for which there exists a chain
N = N
1
N
2
N
t
= G,
with the property that N
i
is a normal subgroup of N
i 1
, for each i . The generalized
Fitting subgroup F
(G) of G is the product of all minimal non-trivial subnormal

subgroups of G. In Case (ii) of the Praeger-Saxl-Yokoyama theorem, the general-
ized Fitting subgroup satises F
(G)

= F
n
s
for some prime power s, and in Case
(iii), F
(G) is a (single) non-abelian simple group the simple socle. To conclude

this section, we give examples of graphs belonging to each of the Cases (ii) and (iii).
Case (ii): the afne case. The Hamming graphs can be thought of as the eld-size-
1 specializations of the so-called bilinear forms graphs. For n d, the bilinear
forms graph G = BF(n, d, q) has as its vertex-set the (n d)-matrices over F
q
,
and two vertices v and w are adjacent whenever rk(v w) = 1. For each vertex
v, the map x . x v is an automorphism of G. In this way, the additive group
N of the vector space F
nd
q
is a group of automorphisms of G: it is even a regular
permutation group on the vertex-set.
Let H be the subgroup of GL(F
nd
q
) generated by all left-multiplications by
invertible (n n)-matrices and all right-multiplications by invertible (d d)-
matrices. Then the semidirect product G = N H acts distance-transitively on
G. For, N is regular on the vertex-set and so, in order to verify that G is distance-
transitive on G, we need establish only that G
i
(0) is a single H-orbit, for 0 i d.
A matrix lies in G
i
(0) if and only if it is the sum of i matrices of rank 1 and no
fewer; but this is equivalent to having rank i . It is well known that the matrices of
rank i form a single H-orbit for each value of i , and so G
i
(0) is a single H-orbit;
in particular, the diameter of G is d. Moreover, G belongs to the afne case with
socle N; its intersection array is
_
(q
n
1)(q
d
1)
q 1
,
(q
n
q)(q
d
q)
q 1
, . . . ,
(q
n
q
d1
)(q
d
q
d1
)
q 1
;
1,
q
3
q
q 1
, . . . ,
q
2d1
q
d1
q 1
_
.
240 Arjeh M. Cohen
Case (iii): the simple socle case. The Johnson graphs can be thought of as the
eld-size-1 specializations of the so-called Grassmann graphs. For d n/2,
the Grassmann graph G(n, d, q) has as its vertex-set the d-dimensional subspaces
of the vector space F
n
q
of dimension n over F
q
, and two vertices v and w are
adjacent whenever dim(v w) = d 1. The general linear group GL(n, q) acts
distance-transitively on G(n, d, q). The pairs of vertices at mutual distance i form
a single orbit, for each value of i in the range 0 i d, which coincides with
G(n, d, q)
i
; in particular, the diameter of G(n, d, q) is d. The kernel of the action
of GL(n, q) on G(n, d, q) is the centre of the group. For n 3, the quotient group
PGL(n, q) has generalized Fitting subgroup PSL(n, q) = A
n1
(q). So this is a
simple socle case when (n, q) ,= (2, 3) or (2, 4); its intersection array is
_
(q
d
1)(q
n1
q)
(q 1)
2
,
(q
d
q)(q
n
q
2
)
(q 1)
2
, . . . ,
(q
d
q
d1
)(q
nd2
q
d
)
(q 1)
2
; 1,
(q
2
1)
(q 1)
2
2
, . . . ,
(q
d
1)
2
(q 1)
2
_
.
9. The afne case
In this section, G is a graph admitting a primitive group of automorphisms G in
such a way that the vertex-set of G can be identied with a vector space V over the
eld F
s
of order s; here, s = r
b
is some power of a prime r, maximal with respect
to G
0
L(V), where G
0
is the stabilizer in G of 0 V, and L(V) is the group
of all semi-linear transformations of V.
The assumption that G is primitive amounts to requiring that the action of G
0
on
V must be irreducible. In his thesis [4] (see also [6]), van Bon showed that there is a
reduction to the case where G
0
is quasi-simple; more precisely, he proved Theorem
9.1 below. Here, a projective representation of a group H on a vector space is a
linear representation on that vector space of a group

H for which the quotient
H/Z(
H) of

H by its centre Z(
H) coincides with H. For every nite non-abelian

simple group H, there is a minimal nite group

H for which every projective
representation of H comes from a linear representation of

H. This group satises
H/Z(
H)

= H, and is equal to its commutator subgroup. A projective or linear
representation is absolutely irreducible if it is irreducible and remains irreducible
after every nite extension of the eld of scalars of the vector space.
Theorem 9.1 Let G be an afne group that acts primitively and distance-transit-
ively on a connected non-complete graph G with degree and diameter at least 3.
Then, with s and V as above, one of the following must hold:
(i) G is associated with a Hamming graph;
(ii) G is a bilinear forms graph;
(iii) V = F
s
is 1-dimensional, and G
0
is a subgroup of L(F
s
);
(iv) the generalized Fitting subgroup F
(G
0
/Z(G
0
GL(V))) is non-abelian
and simple, and its projective representation on V is absolutely irreducible
and can be realized over no proper subeld of F
s
.
The importance of this theorem lies in the fact that it reduces the afne case
to the study of simple groups and their projective representations. For, in Case
(i v), the group G
0
is up to a nite number of possibilities determined by
projective representations of the generalized Fitting subgroup appearing there, and
this must be a non-abelian nite simple group. Given such a group K, the group
G
0
is contained in the normalizer of the image of K in L(V) for an irreducible
representation space V for K. The number of possibilities for V leading to a
VG
0
-invariant distance-transitive graph structure on V is not yet under control.
Theorem 9.3 takes care of this.
The proof of Theorem 9.1 is based on Aschbachers analysis [1] of subgroups
of classical groups. In the rst two cases, the graphs are fully determined (as
discussed above). The next case was dealt with by Cohen and Ivanov [18], who
proved the following result.
Theorem 9.2 Suppose that Case (iii) of Theorem 9.1 holds. Then s = 64, G
0
is a
semidirect product of the cyclic group of order 9 with a cyclic group of order 3 or
6, and G is the Hamming graph H(4, 3).
The proof uses the Hasse-Weil estimates of the number of vectors (x
1
, x
2
, . . . , x
n
)
in F
n
q
that satisfy x
m
1
x
m
n
= 0.
The remainder of this section is devoted to Case (iv) of Theorem9.1. The general
technique here is to use the bounds of Theorem 6.2 and the following result of van
Bon [4].
Theorem 9.3 Let G, s = r
b
, and V be as in Theorem 9.1. Then:
(i) G
0
is closed under scalar multiplication by F
r
;
(ii) if dim
F
r
V = c, then G
0
has at most c orbits in V;
(iii) [V[ 5[G
0
[.
Part (iii) follows from the observation that G contains a 4-cycle, so Theorem 6.2
can be applied: [G
0
[

[G[/5.
242 Arjeh M. Cohen
The general idea of the approach for a xed simple group K is to determine all
possible projective representations of K on nite vector spaces V for which there
exist G
0
and G, as in Case (iv) of Theorem 9.1, with V as the vertex-set of G and
with K isomorphic to (G
0
GL(V))/Z(G
0
GL(V)).
Up to relatively small factors for the central part Z(G
0
GL(V)) (bounded by
[Z(K)[) and for the non-linear part L(V)/GL(V) (bounded by b), the upper bound
for [V[ in terms of G
0
in Theorem 9.3(iii) can be translated into a slightly weaker
upper bound for [V[ in terms of K. Consequently, when given K, we nd a nite
list of projective representations V for G
0
which is usually short. For each K, there
is a nite list of pairs (G
0
, V) that possibly lead to a G-invariant structure on V for
which the permutation group G = VG
0
is a primitive distance-transitive group
of automorphisms.
Although there are an innite number of possibilities for K, the simple groups
can be dealt with series-by-series corresponding to the lines of the tables in Sec-
tion 6. As we have seen in Section 3, we need to analyze G
0
-orbits on V.
One of the few results of a general nature for deciding whether a subgroup G
0
of L(V) leads to a distance-transitive graph is due to van Bon [7]. It deals with
the case where F
s
is contained in G
0
and Kleaves invariant a unitary or orthogonal
form on V.
The remainder of this section follows the above subdivision of types for a
non-abelian nite simple group K.
Afne alternating groups
For K
= A
n
(n 5), the complete analysis was achieved by Liebeck and Praeger
[34].
Theorem 9.4 Suppose that G, s = r
b
, G and K are as in Theorem 9.1(iv), where
K
= A
n
for some integer n 5. Then G is associated with a Hamming graph.
Afne groups of Lie type in the same characteristic
We next take K to be a group of Lie type with dening characteristic r, the
characteristic of the eld over which the projective representation spaces V to be
considered here are dened. The following result is due to van Bon, Cohen and
Cuypers; see [8] and [11].
b
, V, G, K are as in Theorem 9.1(iv), where
K = F
(G
0
GL(V))/Z(G
0
GL(V)) is a simple group of Lie type with dening
characteristic r. Then one of the following holds:
(i) K
= E
6
(q), V is a 27-dimensional vector space over F
q
, and q = s; the
intersection array of G is
_
(q
12
1)(q
9
1)
q
4
1
, q
8
(q
4
1)(q
5
1), q
16
(q 1); 1, q
8
q
4
,
q
20
q
8
q
4
1
_
;
(ii) K
= A
m1
(q), and G is the bilinear forms graph BF(m, m, q) on F
m
q
, so G
0
contains SL(F
m
q
)/(1)
m
I
m
);
(iii) K
= A
m1
(q), and V is the (m(m 1)/2)-dimensional vector space of
alternating forms on F
m
q
, so G
0
contains SL(F
m
q
)/(1)
m
I
m
);
(iv) K
= A
m1
(q
2
), and V is the (m(m 1)/2)-dimensional vector space over
F
q
of all Hermitian forms on F
m
q
2
, so G
0
contains SL(m, q
2
)/(1)
m
I
m
);
(v) K
= A
1
(8), and G is the primitive distance-regular graph associated with
the Hamming graph H(9, 2);
(vi) K
= A
1
(8), and G is the distance 2 graph
2
of the graph in part (v).
The graphs in the conclusions exist and are distance-transitive. Those in Cases (ii)
and (iii) are known as the alternating and Hermitian forms graphs, denoted by
AF(m, q) and HF(m, q
2
), respectively (see [13]).
We outline the proof. By the use of representation theory of the groups of
Lie type, it is relatively easy to provide a list of small-dimensional projective
representation spaces V for K over F
s
, among which all afne distance-transitive
examples must occur. Here, small means that the somewhat larger group G
0
with F
(G
0
GL(V)/Z(G
0
GL(V)))

= K satises s
m
5[G
0
[, in accordance
with Theorem 9.3(iii).
An intuitive argument (based on closures of orbits in the module over the
algebraic closure of F
s
) tells us that the so-called highest-weight orbit of G
0
in
such a representation space V is among the smallest orbits; the notion highest-
weight orbit comes from the theory of groups of Lie type. In view of Theorem
6.1, the G
0
-orbit that denes adjacency in G must be among the smallest. Thus,
it makes sense to exploit knowledge of the highest-weight orbit to analyse the
possible distance-transitive graph structures on V.
To this end, for a G
0
-orbit O of vectors in V, we consider the following two
properties:
(O1) if v O, then v O for all F
q
;
(O2) for each v, w O with w / v), there exists g G
0,v
with gw w O.
The next theorem shows that the existence of such an orbit O often leads to a
small diameter or is the adjacency orbit in G.
244 Arjeh M. Cohen
b
, V and G are as in Theorem 9.1 (iv), and
let O be a G
0
-orbit satisfying (O1) and (O2). Then, with i = d(0, v), one of the
following holds for any v O:
(i) i = 1;
(ii) i = 2 and, if a
2
,= 0, then there exists w O with v w G
1
(0);
(iii) i = 3 or 4, and there exists w O with v w G
2
(0); moreover,
(a) if a
1
,= 0, then i = d;
(b) if there is no non-trivial G
0
-orbit that consists of sums of two elements
of O of size smaller than [O[, then d 5;
(iv) i = d, a
d
= 0, b
d1
= 1, and G
0
= G
0,v
G
0,w
, for some w G
1
(0).
If O is the smallest of all non-trivial G
0
-orbits, then either d 4 or (i ) or (iv)
holds. But case (iv) leads to a factorization of the group K of Lie type into two
factors, one of which (the one corresponding to G
0,v
) is a parabolic subgroup of
K for our purposes, it sufces to knowthat a parabolic subgroup is a well-studied
subgroup that occurs as the stabilizer in G of the 1-dimensional subspace generated
by a highest-weight vector in a projective representation of G. This situation has
been studied satisfactorily, leaving a limited number of possibilities for G
0,w
; see
[23] and [35].
Afne groups of Lie type of cross characteristic
There is an irreducible projective representation on V = F
6
3
of the sporadic group
M
12
, discovered by Mathieu and realizable as a 5-transitive permutation group on
12 letters. The resulting semi-direct product of F
6
3
and a group G
0
with centre of size
2 and quotient isomorphic to M
12
is a distance-transitive group of automorphisms
on a graph of 729 vertices with intersection array (24, 22, 20; 1, 2, 12) (see [13,
p. 359]). This graph may thus be expected to occur in the classication of afne
distance-transitive graphs with (almost) sporadic vertex-stabilizer. Indeed it does
(see Theorem 9.8, below); however it also appears here. The following result was
obtained by Cohen, Magaard and Shpectorov [19].
b
, V, G and K are as in Theorem 9.1(iv),
where K is a simple group of Lie type whose dening characteristic is a prime p
distinct from r, and suppose that K cannot be dened as a group of Lie type with
dening characteristic r. Then K
= A
1
(11) (so p = 11) and G is the graph on 729
vertices associated with M
12
(discussed above).
As in the other cases, the rst stage in the proof of this theoremis the determina-
tion of pairs K, V = F
n
s
for which the bound of Theorem 9.3(iii) is satised. Here,
the results of Landazuri, Seitz and Zalesskii [39] are used to obtain lower bounds
for non-trivial projective representation spaces for K, in characteristic r ,= p. As
a consequence, a nite (but not really small) list of modules V for groups G
0
with
K = F
(G
0
GL(V)/Z(G
0
GL(V))) is found. The next stage is an analysis of
G
0
-orbits on V. One typically arrives at too many G
0
-orbits on V for G to exist, in
view of Theorem 8.3(ii), or at one or two candidate G
0
-orbits for G
1
(0), in which
case inspection of G
2
(0) often leads to more than one G
0
-orbit.
Afne sporadic groups
The successful classication of those afne distance-transitive graphs for which
K is a simple sporadic group was achieved by van Bon, Ivanov and Sax [12].
Theorem 9.8 Suppose that G, s, V, G, K are as in Theorem 9.1(iv), where K is a
sporadic simple group. Then G and K are as described in Table 3.
All of the graphs appearing in this table have been discussed in [13]. For most of
the 26 sporadic simple groups K, the representation spaces V satisfying the bounds
of Theorem 9.3 could be found by use of the Modular Atlas [27]. For the Conway
groups, special arguments were needed in order to nd a nite list of all modules
V that might occur as regular normal subgroups in Case (iv) of Theorem 9.1.
10. The simple socle case
The classication of nite simple groups can be invoked to make a further sub-
division of the possibilities for F
(G). Knowledge of the maximal subgroups of

F
(G) of moderate index should then help to nish the classication of primitive
distance-transitive graphs. The multiplicity freeness criteria (see Theorem 6.3)
give a method for ruling out several cases.
In cases where the permutation representation of G on the vertex-set of G is
equivalent to conjugation on a conjugacy class of involutions of G, van Bon [5]
has given strong restrictions. This is one of the few necessary conditions for a
Table 3.
[V[ Array Name K
3
6
(26, 24, 19; 1, 3, 8) ternary Golay M
12
2
10
(22, 21, 20; 1, 2, 6) truncated Golay M
22
2
11
(23, 22, 21; 1, 2, 3) perfect Golay M
23
2
12
(24, 23, 22, 21; 1, 2, 3, 24) extended Golay M
24
246 Arjeh M. Cohen
permutation representation to come from a distance-transitive action that rules out
many examples that survive the multiplicity freeness criteria.
Alternating groups
The case F
(G)

= A
n
(n 5) is dealt with by Ivanov as well as by Liebeck, Praeger
and Saxl; the multiplicity-free permutation representations of the alternating and
symmetric groups were found by Saxl (see [13, p. 230]).
Theorem 10.1 Let G be a primitive distance-transitive group on the graph G,
where d 3 and k 3. If F
(G)

= A
n
for some n 5, then G is associated with
a Johnson graph J(n, d) or (if n = 2d 1 is odd) an odd graph O
d1
.
Groups of Lie type
This is the hardest subproblem that remains in the classication of primitive
distance-transitive graphs. The parabolic subgroups for which the permutation
representation is distance-transitive are well understood; roughly speaking, they
correspond to the (possibly imprimitive) distance-transitive permutation represen-
tations of the corresponding Weyl groups that is, eld-size-1 specializations of
the groups of Lie type (see [13]).
The classication is expected to follow the pattern of PSL(n, q), which was
treated by van Bon and Cohen in [9]. The proof uses results of Inglis, Liebeck and
Saxl (see [13, p. 230]) to exclude various permutation representations that are not
multiplicity free.
Theorem 10.2 Let G be a group with PSL(n, q) G Aut(PSL(n, q)), for n 2
and (n, q) ,= (2, 2), (2, 3). If G is a graph with diameter d 3 on which G acts
primitively and distance-transitively, then G is either a Grassmann graph or is as
listed in Table 4.
See [13] for details on these graphs.
For those F
(G) that are of classical Lie type, major progress has been made by
van Bon, Inglis and Saxl. They used Aschbachers list of subgroups of classical
groups (see, for example, [1] or [29]) to distinguish between cases, and they used
known bounds on subgroups to limit the possibilities for vertex-stabilizers when
F
(G) is given. Lawther has analysed several series of pairs of a group of Lie
type and a maximal subgroup with respect to multiplicity freeness (see [30], [31]
and [32]).
For those F
(G) that are of exceptional Lie type, Cohen, Liebeck and Saxl have
obtained partial results. As before, the idea of the proof is to use results of Liebeck
Table 4.
[V[ (n, q) Array Name
28 (2, 7) (3, 2, 2, 1; 1, 1, 1, 2) Coxeter
36 (2, 9) (5, 4, 2; 1, 1, 4) Sylvester
45 (2, 9) (4, 2, 2, 2; 1, 1, 1, 2) generalized
8-gon (2, 1)
68 (2, 16) (12, 10, 3; 1, 3, 8) Doro
102 (2, 17) (3, 2, 2, 2, 1, 1, 1; 1, 1, 1, 1, 1, 3) Biggs-Smith
57 (2, 19) (6, 5, 2; 1, 1, 3) Perkel
65 (2, 25) (10, 6, 4; 1, 2, 5) Hall
q
3
2q
2
2q 1 (3, q) (2q, q, q; 1, 1, 2) generalized
6-gon (q, 1)
280 (3, 4) (9, 8, 6, 3; 1, 1, 3, 8) HF(3, 4)
3
56 (4, 2) (15, 8, 3; 1, 4, 9) J(8, 3)
and Saxl [36] to determine the maximal groups of relatively small index, and to
prune the series that fail to be multiplicity-free (see [30] and [32]). A case-by-case
study for each of the remaining permutation groups then needs to be conducted.
There exists a short list of hard cases that have not yet been dealt with.
Sporadic groups
For F
(G) sporadic, the possible graphs G have been determined by Ivanov, Linton,
Lux, Saxl and Soicher [26]. The proof was based on a computer search for all
possible multiplicity-free permutation characters and subsequent calculation of
the collapsed adjacency diagrams.
Theorem 10.3 If G is a primitive distance-transitive group of automorphisms of a
graph G, with d 3 and a sporadic simple generalized Fitting subgroup K, then
G and K are as described in Table 5.
Details of all six graphs can be found in [13].
Table 5.
[V[ K Array Name
266 J
1
(11, 10, 6, 1; 1, 1, 5, 11) Livingstone graph
315 J
2
(10, 8, 8, 2; 1, 1, 4, 5) near octagon
759 M
24
(30, 28, 24; 1, 3, 15) Witt graph
506 M
23
(15, 14, 12; 1, 1, 9) truncated from Witt
330 M
22
(7, 6, 4, 4; 1, 1, 1, 6) doubly truncated Witt
22880 Suz (280, 243, 144, 10; 1, 8, 90, 280) Patterson graph
248 Arjeh M. Cohen
11. Conclusion
We conclude by providing references to three topics that we have not touched
upon:
r
in view of their high degree of symmetry, it should be no surprise that there
have been some applications of distance-regular graphs: (see [20], [28] and
[38]);
r
innite distance-transitive graphs are treated in [16] and [13];
r
there is a directed version of distance-transitive graphs due to Lam; see [13,
p. 232] and [33] for results on these.
References
1. M. Aschbacher, On the maximal subgroups of the nite classical groups, Invent. Math.
76 (1984), 469514.
2. M. Aschbacher and L. L. Scott, Maximal subgroups of nite groups, J. Algebra 92
(1985), 4480.
3. R. W. Baddeley, Multiplicity-free and self-paired primitive permutation groups, J. Al-
gebra 162 (1993), 482530.
4. J. van Bon, Afne distance-transitive groups, Ph.D. thesis, University of Utrecht, 1990.
5. J. van Bon, On distance-transitive graphs and involutions, Graphs Combin. 7 (1991),
377394.
6. J. van Bon, Afne distance-transitive groups, Proc. London Math. Soc. 67 (1993), 152.
7. J. van Bon, Afne distance-transitive graphs with quadratic forms, Math. Proc. Cam-
bridge Philos. Soc. 112 (1992), 507517.
8. J. van Bon and A. M. Cohen, Afne distance transitive graphs and exceptional Chevalley
groups, Proc. London Math. Soc. 83 (2001), 5170.
9. J. van Bon and A. M. Cohen, Linear groups and distance-transitive graphs, Europ. J.
Combin. 10 (1989), 394411.
10. J. van Bon and A. M. Cohen, Prospective classication of distance-transitive graphs,
Proc. Combinatorics 1988 conference at Ravello (eds. A. Barlotti et al.), Mediterranean
Press (1991), 2538.
11. J. van Bon, A. M. Cohen and H. Cuypers, Afne distance-transitive graphs and classical
groups, preprint TU/e, Eindhoven, 2002.
12. J. van Bon, A. A. Ivanov and J. Saxl, Afne distance-transitive graphs with sporadic
stabilizer. Europ. J. Combin. 20 (1999), 163177.
13. A. E. Brouwer, A. M. Cohen and A. Neumaier, Distance-regular graphs, Ergebnisse
der Math. u.i. Grenzgebiete 3. Folge Band 18, Springer-Verlag, 1989.
14. A. E. Brouwer and J. H. Koolen, The distance-regular graphs of valency four, J. Alg.
Combin. 10 (1999), 524.
15. P. J. Cameron, Finite permutation groups and nite simple groups, Bull. London Math.
Soc. 13 (1981), 122.
16. P. J. Cameron, A census of innite distance-transitive graphs. Discrete Math. 192
(1998), 1126.
17. I. V. Chuvaeva and D. V. Pasechnik, Distance-transitive graphs of type q K
q,q
and
projective planes, Europ. J. Combin. 11 (1990), 341346.
18. A. M. Cohen and A. A. Ivanov, Afne distance-transitive groups of dimension one,
Europ. J. Combin. 21 (2000), 191195.
19. A. M. Cohen, K. Magaard and S. Shpectorov, Afne distance-transitive graphs: the
cross characteristic case, Europ. J. Combin. 20 (1999), 351373.
20. J. R. Driscoll, D. M. Healy, Jr. and D. N. Rockmore, Fast discrete polynomial transforms
with applications to data analysis for distance-transitive graphs, SIAM J. Comput. 26
(1997), 10661099.
21. C. D. Godsil, R. A. Liebler and C. E. Praeger, Antipodal distance-transitive covers of
complete graphs, Europ. J. Combin. 19 (1998), 455478.
22. D. Gorenstein, R. Lyons and R. Solomon, The classication of the nite simple groups,
Mathematical Surveys and Monographs 40, Amer. Math. Soc., 1994.
23. C. Hering, M. W. Liebeck and J. Saxl, The factorizations of the nite exceptional groups
of Lie type, J. Algebra 106 (1987), 517527.
24. A. A. Ivanov, Distance-transitive graphs and their classication, Investigations in Alge-
braic Theory of Combinatorial Objects (eds. I. A. Farad zev, A. A. Ivanov, M. H. Klin
and A. J. Woldar), Math. and its Appl. 84, Kluwer (1994), 283378.
25. A. A. Ivanov, R. A. Liebler, T. Penttila and C. E. Praeger, Antipodal distance-transitive
covers of complete bipartite graphs, Europ. J. Combin. 18 (1997), 1133.
26. A. A. Ivanov, S. A. Linton, K. Lux, J. Saxl and L. H. Soicher, Distance-transitive
representations of the sporadic groups, Comm. Algebra 23 (1995), 33793427.
27. C. Jansen, K. Lux, R. Parker and R. Wilson, An atlas of Brauer characters, London
Mathematical Society Monographs, New Series 11, Oxford Univ. Press, 1995.
28. J. S. Jwo and T. C. Tuan, On transmitting delay in a distance-transitive strongly antipodal
graph, Inform. Process. Lett. 51 (1994), 233235.
29. P. Kleidman and M. Liebeck, The subgroup structure of the nite classical groups,
30. R. Lawther, Some (almost) multiplicity-free coset actions, Groups, Combinatorics and
Geometry (eds. M. W. Liebeck and J. Saxl), London Math. Soc. Notes 165, Cambridge
Univ. Press (1992), 292310.
31. R. Lawther, Folding actions, Bull. London Math. Soc. 25 (1993), 132144.
32. R. Lawther, The action of F
4
(q) on cosets of B
4
(q), J. Algebra 212 (1999), 79118.
33. D. A. Leonard and K. Nomura, The girth of a directed distance-regular graph,
J. Combin. Theory (B) 58 (1993), 3439.
34. M. W. Liebeck and C. E. Praeger, Afne distance-transitive groups with alternating or
symmetric point stabilizer, Europ. J. Combin. 13 (1992), 489501.
35. M. W. Liebeck, C. E. Praeger and J. Saxl, The maximal factorizations of the nite
simple groups and their automorphism groups, Memoirs Ameri. Math. Soc. 432, 1990.
36. M. W. Liebeck and J. Saxl, On the orders of maximal subgroups of the nite exceptional
groups of Lie type, Proc. London Math. Soc. 55 (1987), 299330.
37. R. A. Liebler, The classication of distance-transitive graphs of type q K
q,q
, Europ.
J. Combin. 12 (1991), 125128.
38. J. Rif` a and J. Pujol, Completely transitive codes and distance-transitive graphs, Applied
algebra, algebraic algorithms and error-correcting codes (New Orleans, LA, 1991),
Lecture Notes in Computer Science 539, Springer (1991), 360367.
39. G. Seitz and A. Zalesskii, On the minimal degrees of projective representations of nite
Chevalley groups, II, J. Algebra 158 (1993), 233243.
10
Computing with graphs and groups
LEONARD H. SOICHER
1. Introduction
2. Permutation group algorithms
3. Storing and accessing a G-graph
4. Constructing G-graphs
5. G-breadth-rst search in a G-graph
6. Automorphism groups and graph isomorphism
7. Computing with vertex-transitive graphs
8. Coset enumeration
9. Coset enumeration for symmetric graphs
References
In this chapter we discuss the computational study of graphs with groups
acting on them, and demonstrate various ways in which computational group-
theoretic methods are used in the study of graphs and groups. We place
particular emphasis on the ideas and methods behind the GRAPE and nauty
computer packages.
1. Introduction
The study of graphs with groups acting on themis the study of G-graphs. AG-graph
G = (G, G, ) consists of a graph or digraph G, a group G, and a homomorphism
: G Aut(G). (The study of graphs without groups acting on them is just the
special case of G-graphs for which G is the trivial group.) G-graphs arise naturally
in many areas, most obviously in the study of graphs related to permutation groups,
but also in the study of nite geometries and designs.
An example of a G-graph (where G is the symmetric group S
n
) is the Johnson
graph J(n, k), dened to be the graph whose vertex-set consists of all k-subsets of
{1, 2, . . . , n}, with a vertex v adjacent to a vertex w exactly when |v w| = k 1.
250
Now J(n, k) = (J(n, k), S
n
, ) is an S
n
-graph: if v = {i
1
, i
2
, . . . , i
k
} is a vertex
of J(n, k) and x S
n
, then v(x) = {i
1
x, i
2
x, . . . , i
k
x}.
The GAP package GRAPE (see [13] and [38]) is designed for computing with
G-graphs, and in particular makes use of the group G acting on a G-graph G in
order to construct, store, and compute with G efciently. The nauty package (see
[26]) contains the most powerful programs available for computing automorphism
groups of graphs and testing graph isomorphism; nauty is available as a stand-alone
package, or may be used from within GRAPE or MAGMA (see [27]).
There are many computer systems that are useful for studying graphs and groups
which we do not discuss here. These include the algebra system MAGMA, which
supports computation with graphs and includes the nauty package, the CoCo pack-
age [11] for computing with coherent congurations (see [5] and[12]), andKocays
package Groups & Graphs [21] (for Macintosh and Windows), which includes a
graphical user interface to compute with graphs, geometric congurations, com-
binatorial designs, and their automorphism groups. Also worth mentioning are the
algorithms of Rees and the author for computing fundamental groups and covers
of combinatorial cell complexes, and which have been implemented for simplicial
complexes (see [32]).
This chapter is organized as follows. We rst describe some basic algorithms
for permutation groups that are used in the study of G-graphs. We then discuss
the efcient storage and construction of G-graphs, and how to use a modied
form of breadth-rst search to determine efciently many properties of a G-graph.
Following that, we concentrate on the methods used by nauty. Next, we discuss
the application of computational methods to the study of vertex-transitive graphs.
Then, after a brief discussion of the coset enumeration procedure, we describe
some applications of coset enumeration to the study of symmetric graphs.
Throughout this chapter, a digraph is allowed to have loops, but no multiple
arcs, and by the adjacency set or neighbourhood G(v) of a vertex v in a graph or
digraph G we mean the set of all vertices w such that vw is an edge or arc of G.
2. Permutation group algorithms
In this section, we describe some basic permutation group algorithms that are
important in the study of G-graphs; more detailed elementary expositions can be
found in [4] and [9], and a comprehensive treatment of the state of the art in
permutation group algorithms is the book by Seress [33]. Throughout this section,
V is a nite set of size n, and we are computing with the permutation group G
given by a generating set X of permutations of V. The image of v V under the
permutation x of V is denoted by vx.
252 Leonard H. Soicher
Orbits and Schreier vectors
Let v V. The orbit of v under G is the set vG = {vg : g G}. The calculation
of all of the orbits {vG : v V} of G = X is equivalent to determining the
connected components of the Schreier graph with respect to X, for G acting on
V. This Schreier graph S(V, X) is a digraph with vertex-set V, where (v, w) is an
arc exactly when vx = wfor some x X. Aset of rooted directed spanning trees of
the connected components of S(V, X) describes a set R of orbit representatives (the
roots) and their G-orbits (the vertices in each tree). These trees are called Schreier
trees, and we always direct a Schreier tree so that the direction of each arc is away
from the root. A set of Schreier trees can be compactly encoded using a Schreier
vector, usually implemented as an array. ASchreier vector s = s(V, X), for a set R
of orbit representatives and a set of Schreier trees in S(V, X) that they root, is a map
s : V {0} X,
for which s(v) = 0 if v R, and s(v) = x X means that (vx
1
, v) is an arc of
one of the Schreier trees (so vx
1
is the parent of v in that tree).
The computation of a single orbit rG, and the denition of the Schreier vector
entries s(v) for the vertices v in this orbit, are usually done using a breadth-rst
search of the connected component containing r. We rst dene s(r) = 0 and
perform the breadth-rst search outwards from r, nding the adjacency set of a
vertex u by applying each element of X to u. Whenever we encounter a newvertex
v in our breadth-rst search say by applying x to u we dene s(v) = x.
If we are given a Schreier vector s = s(V, X) and a point v V, then we can
determine a pair (r, w) for which r is the root of the Schreier tree containing v,
and w is a word in X such that rw = v. This calculation proceeds as follows:
1. w := emptyword; r := v; x := s(r);
2. while x = 0 do w := xw; r := r x
1
; x := s(r); end do;
On completion, (r, w) is the required pair.
Some brief remarks are in order. We do not keep evaluating the inverses of
elements of X, since we either make X inverse-closed, and so can access directly
the inverse of a generator in X, or we compute r x
1
by tracing r through its cycle
in x until we nd the point mapped to r by x. Finally, note that our denition of a
Schreier vector (for all the orbits of a group simultaneously) is not the usual one.
Bases and strong generating sets
Asequence B = (b
1
, b
2
, . . . , b
m
) of elements of V is a base for G if the (pointwise)
stabilizer G
B
of B is the trivial group. A base B denes a chain
G = G
(1)
G
(2)
G
(m)
G
(m+1)
= {1}
of subgroups of G, for which G
(i )
is the (pointwise) stabilizer of (b
1
, b
2
, . . . , b
i 1
).
We observe that
|G| =
m
i =1
|G
(i )
: G
(i +1)
| and |G
(i )
: G
(i +1)
| = |b
i
G
(i )
|.
A strong generating set for G, relative to B, is a generating set Y for G with the
property that
Y G
(i )
= G
(i )
, for i = 1, 2, . . . , m +1.
Given a strong generating set relative to B, the orbits b
i
G
(i )
can easily be computed,
and we can thus obtain the order of G. A base and associated strong generating set
are required by most advanced permutation group algorithms, and are very useful
in G-graph computations.
As well as introducing the fundamental concepts of base and strong generating
set, Sims [34] devised an algorithm, now called the Schreier-Sims algorithm, to
construct a base and associated strong generating set for the permutation group
G = X. Modern variants of this algorithm and others, implemented in GAP
and MAGMA, can be used to compute bases and strong generating sets for many
permutation groups of degree 10
5
or more (see [33, Chs. 5, 8]). It is shown in [33,
Ch. 5] that, given a base B for G, a strong generating set for G can be computed in
O(n|B|
2
|X|(log|G|)
3
) time. Of course, we could take B to be a sequence of length
n of all of the elements of V, but in practice we often know (or can compute) a
much shorter base.
Some permutation group algorithms, such as those for determining the G-
stabilizer of a set of points of V or determining the centralizer of a subgroup of G,
currently use backtrack search, and are not polynomial-time algorithms. Modern
implementations of certain permutation group backtrack algorithms can be quite
efcient in practice, however, even when V has size 10
5
or more (see [22]). It
would be extremely interesting if polynomial-time algorithms could be found for
set-stabilizer or subgroup-centralizer, especially since the problem of determining
graph isomorphism is polynomial-time reducible to each of these tasks (see [33,
Ch. 3]).
3. Storing and accessing a G-graph
We now consider how GRAPE stores a G-graph G = (G, G, ) in a compact way
that also enables the efcient recovery of basic information about the graph or
digraph G. In this implementation, we store (G) by a generating set, so we now
assume that G = (G) Aut(G) and let G be given by a generating set X.
Let V
1
, V
2
, . . . , V
k
be the orbits of G on the set V of vertices of G, with re-
spective representatives v
1
, v
2
, . . . , v
k
. We store G using a record data structure,
containing:
r
the permutation generators X for G;
r
the list v
1
, v
2
, . . . , v
k
of orbit representatives;
r
the list G(v
1
), G(v
2
), . . . , G(v
k
) of the adjacency sets of these G-orbit
representatives;
r
a Schreier vector s(V, X).
Note that, even if G is trivial, the above method of storing G is not signicantly
worse than a represention by a list of adjacency sets of the vertices.
We now describe how to use the Schreier vector s = s(V, X) to determine
whether (v, w) is an arc of G, for vertices v and w, and to calculate the adjacency
set G(v) of v. We rst use s to determine a pair (v
i
, w) for which v
i
is the orbit
representative of vG and w is a word in X mapping v
i
to v. Then (v, w) is an arc
of G if and only if ww
1
G(v
i
), and we have G(v) = G(v
i
)w.
Our method of storing a G-graph is space-efcient at the cost of time for the
recovery of the adjacency sets of vertices. It is a good idea for the Schreier trees
encoded by s to be as shallow as is reasonably possible hence the use of breadth-
rst search. Athorough discussion on computing shallowSchreier trees is given in
[33, Ch. 4]. We do remark, however, that our method of storing G-graphs can often
save time when constructing a graph, since we need only calculate the adjacency
sets of orbit representatives for G on V.
We discuss one nal point concerningthe GRAPEdata structure for a G-graph G,
which has turned out to be extremely useful when doing real-life calculations. In-
ternally, the n vertices of G are represented by the integers 1, 2, . . . , n, but each ver-
tex also has a name which can be an object of any GAP type. When constructing
a newgraph (possibly fromanother graph), the vertices of this newgraph are num-
bered 1, 2, . . . , n, but their names are chosen to reect the mathematical nature
of the vertices. Details of this naming for specic functions can be found in the
GRAPE manual. (The use of this naming, which is implemented by a list, was
suggested by P. J. Cameron.)
4. Constructing G-graphs
Although not always necessary, we assume from here on that if we are computing
with a G-graph G = (G, G, ), then we have a base and associated strong gen-
erating set for (G) acting on V = V(G). This allows, for example, the efcient
calculation of the (G)-stabilizer of an arbitrary point in V (see [33, Ch. 5]). We re-
mark that, in many constructions of G-graphs, our knowledge about G often allows
us to compute this base and strong generating set much more efciently than
otherwise. For example, we often know the order of (G), and this allows the use
of a very fast randomized algorithm to compute a base and strong generating set
(see [33, Ch. 8]), often applicable to groups of degree 10
6
or more. We also remark
that the nauty package [26] outputs the automorphism group of a graph as a base
and associated strong generating set for that group (see Section 6).
GRAPE has many functions to construct G-graphs, including Cayley graphs,
orbital graphs and digraphs, induced subgraphs and quotient graphs. However, the
most useful and general way of constructing a G-graph in GRAPEis to use the func-
tion Graph, which behaves as follows. The input is a group G (which may or may
not be a permutation group), a nite set V on which G acts (the action can be one of
the standard actions in GAP or may be one supplied by the user), and a G-invariant
relation rel on V (given as a function of two vertices v and w and which returns
TRUE or FALSE, according to whether or not (v, w) is in the relation). The output
is the G-graph G with vertex-set V, where (v, w) is an arc if and only if rel(v, w).
The rst step is to compute the orbits of G acting on V and the associated
Schreier vector. Then, for each orbit representative r, we need to determine those
vertices v V such that rel(r, v). In fact, we need check rel(r, s) only for those s
in a set of orbit representatives of the orbits on V of the G-stabilizer Hof r, since
rel(r, t ) for each t in sH if and only if rel(r, s).
5. G-breadth-rst search in a G-graph
In this section, for ease of exposition, we assume that our G-graph G = (V, E) is
a simple graph, and that G is a subgroup of Aut(G). We dene G
i
(v) to be the set
of all vertices at distance i from a vertex v of G.
Many properties of G can be determined by (possibly repeated) applications
of breadth-rst search. These include nding the connected components, diameter
and girth, as well as determining various regularity properties such as whether G
is distance-regular. We describe here a version of breadth-rst search, which we
call G-breadth-rst search, which takes into account G as a subgroup of Aut(G).
Let v V, and let H = G
v
be the stabilizer in G of v. The key observation is
that if w is a vertex at distance i from v, then each vertex in the orbit wH is at
distance i from v.
Suppose that V
1
(= {v}),V
2
, . . . , V
k
are the orbits of Hon V, withrespective rep-
resentatives v
1
(= v), v
2
, . . . , v
k
, and let R = {v
1
, v
2
, . . . , v
k
}. In a G-breadth-rst
search from v, we determine the sets R
0
, R
1
, . . . of H-orbit representatives, where
R
i
= G
i
(v) R.
Given R
i
, we then have G
i
(v) as the union of the (already computed) orbits rep-
resented by the elements of R
i
.
Clearly R
0
= {v}. The basic step is to obtain R
i +1
from R
i
. We actually do
more, in order to obtain more information from our G-breadth-rst search. We
start by setting R
i +1
:= {} and then, for each r R
i
, do the following:
1. determine C := G(r) G
i 1
(v), A := G(r) G
i
(v) and
B := G(r) (C A);
2. add to R
i +1
the representatives of the H-orbits that intersect B non-trivially;
3. for later use, store c
i
(v, r) := |C|, a
i
(v, r) := | A|, and b
i
(v, r) := |B|.
The G-breadth-rst search stops with R
0
, R
1
, . . . , R
m
, when R
m+1
is empty but
R
m
is not. If required, it is trivial to recover G
0
(v), G
1
(v), . . . , G
m
(v), the union of
these sets being the vertices in the connected component containing v. Also note
that m is the greatest distance d(v) fromv to any vertex in the connected component
of v. Moreover, if G is connected, then its diameter diam(G) is the maximumvalue
of d(w), as w ranges over a set of representatives of the G-orbits on V.
Let g(v) be the length of a shortest cycle containing v if v is on some cycle
of G, and let g(v) = otherwise. The numbers c
i
(v, r) and a
i
(v, r) computed
above (for each i = 1, 2, . . . , m and each r R
i
) can be used to determine g(v),
as follows.
Let t be the least value of i {1, 2, . . . , m} for which c
i
(v, r) 2 or a
i
(v, r)
1, for some r R
i
, if such an i exists. If no such i exists, then g(v) = ; otherwise,
if c
t
(v, r) 2 for some r R
i
, then g(v) = 2t ; if not, then g(v) = 2t +1.
Note that the girth of G is the minimum value of g(w) as w ranges over a set
of representatives of the G-orbits on V (a girth of means that G has no cycles).
Now dene c
0
(v, v) = 0, a
0
(v, v) = 0 and b
0
(v, v) = |G(v)|. Let 0 i m
and r R
i
. If c
i
(v, r) depends only on i and v (and not on the H-orbit representa-
tive r), then we denote this quantity by c
i
(v) and call it a local parameter of G.
Similarly, if a
i
(v, r) and b
i
(v, r) do not depend on r, then these too are called local
parameters, and are respectively denoted by a
i
(v) and b
i
(v). Such local param-
eters, if and when they exist, are used in the determination of various regularity
properties of G, the strongest of which is distance-regularity. Indeed, the graph G
is distance-regular if and only if G is connected, d(v) is the same for all vertices v
(so each d(v) = diam(G)), and for each vertex v and i = 0, 1, . . . , diam(G), all
local parameters c
i
(v), a
i
(v) and b
i
(v) exist and do not depend on v, thus giv-
ing the parameters c
i
, a
i
and b
i
of a distance-regular graph; see Chapter 9. Of
course, we need to check these conditions only for those vertices v in a set of orbit
representatives of G on V.
Finally, we remark that G acts distance-transitively on G (see Chapter 9) if and
only if G is connected, G has just one orbit on V, and, for some (and hence all)
vertices v, the number of G
v
-orbits on V is diam(G) +1.
G-breadth-rst search is efciently implemented in GRAPE in order to deter-
mine connected components, diameter, girth, and regularity properties of G-graphs.
Although we do not discuss it here, GRAPEincludes many other functions for com-
puting with G-graphs, such as backtrack search functions for classifying complete
subgraphs of given weight-sum in a vertex-weighted G-graph and for classifying
partial linear spaces with given point graph and parameters.
6. Automorphism groups and graph isomorphism
Automorphism groups of graphs are discussed in Chapter 5. The most advanced
algorithms and programs for computing automorphism groups of graphs and for
testing graph isomorphism are those of McKay, freely available as part of his
nauty package [26] and useful for graphs with up to about 10
4
vertices; nauty
also works with digraphs, although we do not consider them in this section. We
shall briey describe McKays method of partition backtrack, which has been
very inuential in computational group theory. Indeed, partition-based backtrack
methods are now used in the most advanced available algorithms for calculating
set-stabilizers, centralizers and normalizers in permutation groups (see [33, Ch. 9]).
Let G = (V, E) be a graph, with V = {1, 2, . . . , n}, and let be an ordered
partition of V. Thus is a sequence (V
1
, V
2
, . . . , V
k
) of distinct subsets of V, for
which {V
1
, V
2
, . . . , V
k
} is a partition of V: the elements of V are called cells. (One
can think of (G, ) as being a (not necessarily properly) vertex-coloured graph,
with vertices v and w having the same colour if and only if they belong to the same
cell of .) Dene the ordered partition c() of V to be
({1, 2, . . . , |V
1
|}, {|V
1
| +1, |V
1
| +2, . . . , |V
1
| +|V
2
|}, . . . ,
{n |V
k
| +1, n |V
k
| +2, . . . , n}).
If x is a permutation of V, dene Gx to be the graph (V, Ex), where Ex =
{{vx, wx} : {v, w} E}, and dene x = (V
1
x, V
2
x, . . . , V
k
x), where V
i
x =
{vx : v V
i
}. The automorphism group Aut(G, ) of (G, ) is the group of all
permutations x of V for which (Gx, x) = (G, ). Thus, when = (V), the
automorphism group of (G, ) is just Aut(G).
The main functions of the nauty package are to determine Aut(G, ), in the
form of a base and associated strong generating set, and in the process, to compute
the image of (G, ) under a canonical labelling map (described below) which is
used for isomorphism testing.
A canonical labelling map is a function C such that, for each graph G with
vertex-set V = {1, 2, . . . , n}, each ordered partition of V, and each permutation
x of V, we have:
r
C(G, ) = Gy, for some permutation y of V such that y = c();
r
C(Gx, x) = C(G, ).
The importance of a canonical labelling map C is this. Suppose that G
1
and G
2
are graphs on the same vertex-set V = {1, 2, . . . , n}, and that
1
and
2
are ordered
partitions of V with c(
1
) = c(
2
). Then there is a permutation y of V such that
(G
1
y,
1
y) = (G
2
,
2
) if and only if C(G
1
,
1
) = C(G
2
,
2
). In particular, G
1
is
isomorphic to G
2
if and only if C(G
1
, (V)) = C(G
2
, (V)).
Partition backtrack
As before, G is a graph with vertex-set V = {1, 2, . . . , n} and = (V
1
, V
2
, . . . ,
V
k
) is an ordered partition of V. The cells of of size 1 are called singletons, and
a discrete partition is one in which all cells are singletons.
For each v V, let u(v, ), denote the index of the cell in containing v; in
other words, u(v, ) = i means that v V
i
. We say that an ordered partition of
V is a renement of if:
r
each cell of is contained in some cell of ;
r
for each v, w V: if u(, v) < u(, w), then u(, v) < u(, w).
A renement process is a function Rsuch that, for each graph G with vertex-set
V, each ordered partition of V, and each permutation x of V,
R(G, ) is a renement of , and R(G, )x = R(Gx, x).
In particular, if x Aut(G), then R(G, )x = R(G, x), and so x maps the
ordered partition to only if x maps the renement R(G, ) to R(G, ). In
particular, if c(R(G, )) = c(R(G, )), then there is no element of Aut(G) that
maps to . Note that, if R(G, ) is discrete, then Aut(G, ) is trivial.
An ordered partition = (V
1
, V
2
, . . . , V
k
) of V is G-equitable if there are
constants d
i j
(1 i, j k) such that |G(v) V
j
| = d
i j
for each vertex v V
i
.
The default renement process Rused by nauty (for simple graphs) maps (G, )
to an ordered G-equitable partition , such that is a renement of and, up to
the order of its cells, is the coarsest equitable partition that is a renement of .
More precise details can be found in McKay [25].
We now outline (roughly) how the nauty procedure nds Aut(G, ). For more
details, together with how a canonical labelling map is found, see [25] (see also
[20], [24] and [28]).
We let Rbe the default renement process used by nauty. The nauty procedure
proceeds by a depth-rst search in a search tree whose nodes are ordered G-
equitable partitions of V; the root is R(G, ) and the leaves are discrete partitions.
We call the rst leaf found . Any other leaf may give rise to a newautomorphism
of (G, ), and we check to see whether the unique permutation of V mapping
to is in Aut(G, ). A non-leaf in the search tree is not discrete, and we obtain
its children as follows. First, we choose a non-singleton cell C of according to
some rule we usually just take the rst non-singleton cell. Then, for each v C
(in ascending order of v), we isolate v that is, we formthe ordered partition v
obtained from by replacing C by {v}, C {v}, and then we add R(G, v) as
a child of . The search tree is pruned in various ways, so as to avoid searching
subtrees providing no new information (see [25]). The search is structured so as to
provide a base and associated strong generating set for Aut(G, ). In determining
the rst leaf , G-vertices w
1
, w
2
, . . . , w
m
, say, are isolated (w
i
at depth i ) for
the formation of the ancestors of (other than the root) and the formation of
itself. It is not difcult to see that (w
1
, w
2
, . . . , w
m
) is a base for Aut(G, ); nauty
computes a strong generating set relative to this base.
7. Computing with vertex-transitive graphs
Recall that a graph or digraph G is vertex-transitive if Aut(G) acts transitively
on V(G). The class of vertex-transitive graphs includes Cayley graphs and sym-
metric graphs (studied in Chapters 6 and 7), which further includes the class of
distance-transitive graphs (studied in Chapter 9). In this section we consider the
computational study of a G-graph G, where G (a subgroup of Aut(G)) acts tran-
sitively on V(G). Note that the GRAPE data structure for storing a G-graph is
especially compact in this case.
Collapsed adjacency matrices
Let G be a transitive permutation group on a nite set V. Then G has a natural ac-
tion on V V, dened by (v, w)x = (vx, wx). The orbits of this action are called
orbitals, and the orbits of the stabilizer G
v
of a point v V are called suborbits. It
is well known that the orbitals for G are in one-to-one correspondence with these
suborbits: this correspondence maps an orbital E to the suborbit {w : (v, w) E}.
The orbital digraph for G associated with an orbital E is simply the digraph (V, E).
If the orbital E is non-diagonal and self-paired (see Chapter 8), then we associate
the orbital graph (V, {{v, w} : (v, w) E}) with E.
Let v
1
V, and suppose that V
1
(= {v
1
}),V
2
, . . . ,V
k
is an ordering of the orbits
of G
v
1
, with respective representatives v
1
,v
2
, . . . ,v
k
; k is the rank of G. Let G =
(V, E) be a graph or digraph on which G acts vertex-transitively, so that E is a
union of orbitals and G(v) is the union of the corresponding suborbits contained
in {V
1
,V
2
, . . . ,V
k
}. For i, j = 1, 2, . . . , k, dene
a
i j
= |G(v
i
) V
j
|.
Note that a
i j
does not depend on the choice v
i
of suborbit representative, and it
can easily be computed using the GRAPE data structure for the G-graph G in
practice, for |V| uptoabout 10
6
. The k k integer matrixA = (a
i j
) is the collapsed
adjacency matrix for G, with respect to G and the ordering of the suborbits. This
matrix (which is extremely compact when k is small) contains at least as much
information as that computed in a G-breadth rst search from v.
Since G acts vertex-transitively, the single collapsed adjacency matrix A for G
can be used to determine whether G is (strongly) connected, and if so, what its
diameter is; whether G is a simple graph, and if so what its girth is; whether G is
distance-regular, and whether G acts distance-transitively on G. See Praeger and
Soicher [31] for more detailed information and applications of collapsed adjacency
matrices, and also Chapter 9 for examples of collapsed adjacency diagrams.
We remark that the collapsed adjacency matrices for the orbital digraphs for
the transitive group G are useful in studying the coherent conguration associated
with G (see [5] and [12]), since, with respect to a xed ordering V
1
,V
2
, . . . ,V
k
of the suborbits as above, the collapsed adjacency matrix for an orbital digraph
(V, E) for G is the transpose of the intersection matrix (as dened in [5, Ch. 3])
corresponding to the orbital paired with E.
Distance-transitive graphs
Chapter 9 provides an overview of the state of the classication of distance-
transitive graphs. Here we discuss the application of computing, which has been
used in the discovery, analysis and classication of certain distance-regular and
distance-transitive graphs (see, for example, [37], [31], [23] and [18]). We also
remark that computing is used in the determination of feasible intersection arrays
for possible distance-regular graphs (see, for example, [3]).
Suppose that G acts distance-transitively on a graph G. Then G must be an
orbital graph for G of the smallest or second-smallest vertex-degree (see [18]).
Furthermore, each orbital for G must be self-paired, which is equivalent to the
property that the permutation character of G on V is the sum of distinct complex
irreducible characters, each with Frobenius-Schur indicator +1 (see [2, p. 64]).
For these, and other reasons, it makes sense to analyse the lower degree orbital
graphs of permutation representations whose character is multiplicity-free that
is, the sum of distinct complex irreducible characters.
Computation has been applied extensively in the classication of the graphs
on which a sporadic simple group or its automorphism group acts primitively and
distance-transitively (see [23] and [18]; the results of this classication are given in
Chapter 9). In the process of this classication, the primitive multiplicity-free per-
mutationcharacters for these sporadic groups were alsodetermined, andfor most of
the corresponding permutation representations of sporadic groups, a collapsed ad-
jacency matrix was computed for the orbital graph of least degree. This built on the
work of Praeger and Soicher [31], where collapsed adjacency matrices were com-
putedfor the orbital digraphs for all permutationrepresentations of rankat most 5of
the sporadic simple groups and their automorphismgroups. The practical computa-
tional determinationof permutationcharacters is describedinsome detail inLinton,
Lux and Soicher [23], which also details randomized techniques for computing col-
lapsed adjacency matrices for certain permutation representations of degree about
10
11
, where it would be impossible to store explicit permutation generators. These
techniques make use of graph algorithms as well as permutation group algorithms.
An ambitious project at Lehrstuhl D f ur Mathematik, RWTH (Aachen), involv-
ing T. Breuer, I. H ohler and J. M uller, has since determined collapsed adjacency
matrices for all orbital digraphs for all multiplicity-free permutation representa-
tions of the sporadic simple groups and their automorphismgroups, and the results
are published on the world-wide web [1] (although what they call collapsed ad-
jacency matrices we would call intersection matrices).
We remark that many permutation and matrix representations of nite simple
groups and related groups can be downloaded from the online ATLAS of group
representations (see [41]). These group representations are very useful for con-
structing related G-graphs and collapsed adjacency matrices.
8. Coset enumeration
Coset enumeration is one of the oldest and most useful methods of computational
group theory (see [30], [35], [14] and their references). For this chapter, we concen-
trate on what coset enumeration does, and on the application of coset enumeration
and related procedures to problems in graph theory.
Let G = X : R be a nitely presented group that is, G is generated by the
nite set X, subject (only) to the nite set R of relators which are words in X X
1
that evaluate to the identity in G (where X
1
= {x
1
: x X}). The input to coset
enumeration is (X, R, Y), where Y is a set of words in X X
1
that generates a
subgroup H of G.
The coset enumeration process attempts to construct a set V, with 1 V, and
a transitive permutation representation : G Sym(V) for which, in this rep-
resentation, H is the stabilizer in G of the point 1. Coset enumeration does this
by using a trial-and-error process for constructing the permutations (X X
1
).
The name coset enumeration comes from the fact that, if a coset enumeration is
successful, then there is a one-to-one correspondence between the elements of V
and the cosets of H in G, with 1 corresponding to H.
If the index of H in G is innite, then the coset enumeration process does not
terminate; if it is nite, then the process terminates, but there can be no computable
general bound (in terms of the size of the input and the putative index) on the time
or store required for termination (see, for example, Neub user [30]).
There is an enormous amount of exibility in the coset enumeration process, and
many different approaches have been suggested and experimented with (see [30],
[35] and [14]). Depending on the presentation and the approach used, there can be
huge variations in the time and store taken. Currently, the most advanced methods
are due to Havas and Ramsay, and these methods are available in their ACEpackage
[15], also available as a GAP package [16] and within the MAGMA system [27].
There are many useful variations on coset enumeration. For example, a mod-
ied Todd-Coxeter enumeration gives a presentation for H, and the low-index
subgroups procedure determines (up to permutation isomorphism) all transitive
representations of G up to some given degree k; see [30] for an excellent introduc-
tion to coset enumeration and its variations.
9. Coset enumeration for symmetric graphs
Recall that a graph G is symmetric if Aut(G) acts transitively on both its ver-
tices and arcs (ordered pairs of adjacent vertices). A subgroup G of Aut(G) acts
symmetrically on G if G acts transitively on both the vertices and arcs of G.
One common way to study connected symmetric graphs with given properties is
by determining the groups that act on themsymmetrically, as quotients of universal
completions of appropriate amalgams (see Chapter 7 and [19] for useful overviews,
and [17] for a more general geometric context). In this approach, we rst use the
given graph-theoretical properties to determine the possible amalgams of the form
A = {G
v
, G
{v,w}
}, where (v, w) is an arc in the putative graph on which G acts
symmetrically. Since the universal completion U(A) of such an amalgam of two
groups (with neither a subgroup of the other) is innite, we need to add further
relations to U(A) to obtain the nite groups of automorphisms we seek. Such extra
relations could come from cycles in our graph or from the local graph structure.
Coset enumeration can then be used in an effort to determine the (hopefully
nite) index of H = G
v
in G, and to construct the representation of G acting on
the set V of right cosets of Hin G. Given such a representation of G on V, we can
reconstruct and study the graph G which may have the properties we seek or
we may have been able to prove theoretically that G has the required properties.
This graph G is simply the orbital graph for G for which the orbital contains
(H, Hg), where g is an element of G
{v,w}
H. Applications of this kind include
those in Morton [29], where the 4-arc- and 5-arc-transitive connected cubic graphs
of girth up to 11 and girth 13 are classied; see Chapter 7 for a discussion of s-
arc transitivity, and Conder and Dobcs anyi (see [6] and [7]), where the connected
symmetric cubic graphs on up to 768 vertices are determined, using a powerful new
low-index subgroups procedure, parallel computation and coset enumeration. Of
course, computational studies of this kind often lead to conjectures and theoretical
results. For more applications of coset enumeration to the study of symmetric
graphs, and for a beautiful and natural characterization and construction of the
sporadic simple group J
3
, see Weiss [39] and also [40], [36] and [10].
Graphs that are locally a given graph
We nowgive an example where additional transitivity assumptions and local struc-
ture specication lead to an amalgam of three groups. We use the ATLAS notation
[8] for group structures.
Let G and H be graphs. Then G is said to be locally-H if, for each vertex v
of G, the induced subgraph on G(v) is isomorphic to H (this situation is discuss-
edbrieyinChapter 5). Givena graph H, coset enumerationcansometimes be used
effectively to study presentations that arise in the classication of the connected
graphs G that are locally-H, for which Aut(G) acts transitively on the ordered
triangles of G. These presentations come from applying the amalgam method to
putative stabilizers of a vertex, incident edge and triangle, contained in a xed
triangle of such a graph G.
A simple, but good, example of this application of coset enumeration is given
in [10], where H is the incidence graph of the unique 2-(11, 5, 2) design, and
where the ordered-triangle-transitive graphs G that are locally-H are classied.
We discuss here the case where a vertex-stabilizer in the automorphismgroup of G
is isomorphic to PGL
2
(11). For such a graph G, it is shown that there is (essentially)
only one possible amalgam A = {X, Y, Z} of the Aut(G)-stabilizers X, Y, Z of
x, {x, y}, and {x, y, z}, contained in a triangle {x,y,z} of G. In this amalgam,
X

= PGL
2
(11), Y

= S
5
, Z

= (A
4
3) : 2, X Y

= A
5
, X Z

= S
4
, and Y
Z

= S
4
. Then Aut(G) is a homomorphic image of U(A) which has the following
presentation, determined in [10]:
a, b, c, d, e : a
3
= b
2
= c
2
= d
2
= (ab)
3
= (ac)
2
= (ad)
2
= a(cd)
4
= (bc)
3
= (bd)
2
= e
2
= (ae)
2
= (be)
2
= (ce)
2
= (de)
3
= 1
.
For this presentation of U(A), X maps onto a, b, c, d, Y maps onto a, b, c, e,
and Z maps onto a, b, d, e; the relations (ac)
2
= (ad)
2
= 1 are consequences
of the others.
Applying coset enumeration, we nd that a, b, c, d has index 432 in U(A).
It is then shown that U(A)

= (3 M
12
) : 2, and that there are just two connected
ordered-triangle-transitive locally-H graphs whose vertex-stabilizer is PGL
2
(11),
having (respectively) 432 and 144 vertices and automorphism groups isomorphic
to U(A) and U(A)/(bcde)
11

= M
12
: 2, the automorphismgroup of the Mathieu
group M
12
.
References
1. T. Breuer and J. M uller, The character tables of endomorphism rings of multiplicity-
free permutation modules of the sporadic simple groups, their automorphism groups,
and their cyclic central extension groups; http://www.math.rwthaachen.
de/
Juergen.Mueller/mferctbl/mferctbl.html.
2. A. E. Brouwer, A. M. Cohen and A. Neumaier, Distance-Regular Graphs, Springer,
1989.
3. A. E. Brouwer and J. H. Koolen, The distance-regular graphs of valency four, J. Alg.
Combin. 10 (1999), 524.
4. G. Butler, Fundamental Algorithms for Permutation Groups, Lecture Notes in Com-
puter Science 559, Springer, 1991.
5. P. J. Cameron, Permutation Groups, London Math. Soc. Student Texts 45, Cambridge
Univ. Press, 1999.
6. M. Conder and P. Dobcs anyi, Trivalent symmetric graphs on up to 768 vertices, J.
Combin. Math. Combin. Comput. 40 (2002), 4163.
7. M. Conder and P. Dobcs anyi, Applications and adaptations of the low index subgroups
procedure, Math. Comput., 2004.
8. J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker and R. A. Wilson, ATLAS of
Finite Groups, Oxford Univ. Press, 1985.
9. H. Cuypers, L. H. Soicher and H. Sterk, Working with nite groups, Some Tapas of
Computer Algebra (eds. A. M. Cohen, H. Cuypers and H. Sterk), Springer (1999),
184207.
10. H. Cuypers, L. H. Soicher and H. Sterk, The small Mathieu groups (Project), Some
Tapas of Computer Algebra (eds. A. M. Cohen, H. Cuypers and H. Sterk), Springer
(1999), 323337.
11. I. A. Farad zev and M. H. Klin, Computer package for computations with coherent
congurations, Proc. ISSAC 91 (ed. S. M. Watt), Assoc. Comp. Mach. (1991), 219
223.
12. I. A. Farad zev, M. H. Klin and M. E. Muzichuk, Cellular rings and groups of automor-
phisms of graphs, Investigations in Algebraic Theory of Combinatorial Objects (eds.
I. A. Farad zev, A. A. Ivanov, M. H. Klin and A. J. Woldar), Kluwer (1994), 1152.
13. The GAP Group, GAP Groups, Algorithms, and Programming, Version 4.3, Aachen,
St Andrews (2002); http://www.gapsystem.org.
14. G. Havas and C. Ramsay, Experiments in coset enumeration, Groups and Computation
III (eds. W. M. Kantor and

A. Seress), de Gruyter (2001), 183192.
15. G. Havas and C. Ramsay, The Advanced Coset Enumerator ACE; http://www.
csee.uq.edu.au/
cram/ce.html.
16. G. Havas, C. Ramsay, G. Gamble and A. Hulpke, The ACE Package for GAP;
http://www.gapsystem.org/Share/ace.html.
17. A. A. Ivanov, Geometry of Sporadic Groups. I: Petersen and Tilde Geometries,
18. A. A. Ivanov, S. A. Linton, K. Lux, J. Saxl and L. H. Soicher, Distance-transitive
representations of the sporadic groups, Comm. Algebra 23 (1995), 33793427.
19. A. A. Ivanov and S. V. Shpectorov, Applications of group amalgams to algebraic
graph theory, Investigations in Algebraic Theory of Combinatorial Objects (eds. I. A.
Farad zev, A. A. Ivanov, M. H. Klin and A. J. Woldar), Kluwer (1994), 417441.
20. W. Kocay, On writing isomorphism programs, Computational and Constructive
Design Theory (ed. W. D. Wallis), Kluwer (1996), 135175.
21. W. Kocay, Groups & Graphs; http://www.paddle.mb.ca/G&G/G&G.html.
22. J. S. Leon, Partitions, renements, and permutation group computation, Groups and
Computation II (eds. L. Finkelstein and W. M. Kantor), DIMACS Series in Discrete
Mathematics and Theoretical Computer Science 28, Amer. Math. Soc. (1997), 123158.
23. S. A. Linton, K. Lux and L. H. Soicher, The primitive distance-transitive representations
of the Fischer groups, Experimental Math. 4 (1995), 235253.
24. B. D. McKay, Computing automorphisms and canonical labellings of graphs, Combin.
Math., Proc. Int. Conf., Canberra 1977, Lecture Notes in Mathematics 686, Springer
(1978), 223232.
25. B. D. McKay, Practical graph isomorphism, Congr. Numerantium 30 (1981), 4587.
26. B. D. McKay, nauty; http://cs.anu.edu.au/people/bdm/nauty.
27. The Magma Computational Algebra System, Computer Algebra Group, School of
Mathematics and Statistics, University of Sydney; http://magma.maths.usyd.
edu.au.
28. T. Miyazaki, The complexity of McKays canonical labelling algorithm, Groups
and Computation II (eds. L. Finkelstein and W. M. Kantor), DIMACS Series in
Discrete Mathematics and Theoretical Computer Science 28, Amer. Math. Soc. (1997),
239256.
29. M. J. Morton, Classication of 4- and 5-arc-transitive cubic graphs of small girth, J.
Austral. Math. Soc. (A) 50 (1991), 138149, and Corrigendum, 52 (1992), 419420.
30. J. Neub user, An elementary introduction to coset table methods in computational
group theory, Groups St. Andrews 1981 (eds. C. M. Campbell and E. F. Robertson),
London Math. Soc. Lecture Note Series 71, Cambridge Univ. Press (1982), 145.
31. C. E. Praeger and L. H. Soicher, Low Rank Representations and Graphs for Sporadic
Groups, Australian Math. Soc. Lecture Series 8, Cambridge Univ. Press, 1997.
32. S. Rees and L. H. Soicher, An algorithmic approach to fundamental groups and covers
of combinatorial cell complexes, J. Symbolic Comp. 29 (2000), 5977; GAP program
available at http://www.maths.qmul.ac.uk/
leonard/fundamental.
33.

A. Seress, Permutation Group Algorithms, Cambridge Univ. Press, 2003.
34. C. C. Sims, Computation with permutation groups, Proc. Second Symposium on
Symbolic and Algebraic Manipulation (ed. S. R. Petrick), Assoc. Comp. Mach. (1971),
2328.
35. C. C. Sims, Computation with Finitely Presented Groups, Cambridge Univ. Press, 1994.
36. L. H. Soicher, On simplicial complexes related to the Suzuki sequence graphs, Groups,
Combinatorics and Geometry (eds. M. W. Liebeck and J. Saxl), London Math. Soc.
Lecture Note Series 165, Cambridge Univ. Press (1992), 240248.
37. L. H. Soicher, Three newdistance-regular graphs, Europ. J. Comb. 14 (1993), 501505.
38. L. H. Soicher, The GRAPE Package for GAP; http://www.gapsystem.
org/Share/grape.html.
39. R. Weiss, A characterization and another construction of Jankos group J
3
, Trans.
Amer. Math. Soc. 298 (1986), 621633.
40. R. Weiss, Presentations for (G, s)-transitive graphs of small valency, Math. Proc.
Camb. Philos. Soc. 101 (1987), 720.
41. R. A. Wilson et al., ATLAS of nite group representations; http://www.mat.
bham.ac.uk/atlas.
Notes on contributors
Brian Alspach [[email protected]] received his Ph.D. under the super-
vision of Paul Kelly in 1966. He spent 196699 at Simon Fraser University in
Vancouver and is now an adjunct professor at the University of Regina. His main
interests are applications of graphs, tournaments, Cayley graphs and graph de-
compositions. Now ofcially retired, he has time for writing, studying piano and
poker.
Lowell Beineke [[email protected]] is Schrey Professor of Mathematics at Indi-
ana University-Purdue University Fort Wayne, where he has been since receiving
his Ph.D. from the University of Michigan under the guidance of Frank Harary.
His graph theory interests are broad, and include topological graph theory, line
graphs, tournaments, decompositions and vulnerability. With Robin Wilson he
edited Selected Topics in Graph Theory (3 volumes), Applications of Graph The-
ory, and Graph Connections. He is currently editor of the College Mathematics
Journal.
Richard Brualdi [[email protected]] is a professor of mathematics at the
University of Wisconsin in Madison. He has written a number of papers on combi-
natorics, graphs and matrices, and is author of Introductory Combinatorics, Com-
binatorial Matrix Theory (with H. J. Ryser), and Matrices of Sign-Solvable Linear
Systems (with B. L. Shader). A past president of the International Linear Alge-
bra Society, he serves as editor-in-chief of the journals Linear Algebra and its
Applications and the Electronic Journal of Combinatorics.
Peter Cameron [[email protected]] is a professor of mathematics at
Queen Mary, University of London. His mathematical interests are in permutation
groups and their operands (which may be logical, algebraic or combinatorial, and
nite or innite). He has written lecture notes on permutation groups, parallelisms,
andprojective andpolar geometry(with J. H. van Lint), a bookongraphs, codes and
267
268 Notes on contributors
designs, and undergraduate textbooks on combinatorics and algebra. He is Presi-
dent of the British Combinatorial Committee.
Arjeh Cohen [[email protected]] studied mathematics and theoretical computer
science at Utrecht University. His research elds are discrete algebra and geometry,
and mathematics on computers. He worked at Rijnmond Authority (Rotterdam),
the University of Twente (Enschede) and CWI (Amsterdam), and at Utrecht Uni-
versity. Since 1992, he has been a professor of discrete mathematics at Eindhoven
University of Technology. He is scientic director of RIACA, chair of the board
of the Research School EIDMA, and president of the OpenMath Society. He has
written about ninety papers, eleven books, and three software packages.
Drago s Cvetkovi c [[email protected]] graduated in electrical engineering at
the University of Belgrade in 1964, where he was awarded a doctoral degree in
mathematics in 1971 and where, since 1986, he has been a professor of mathe-
matics. His major eld of interest is discrete mathematics, and he has published
about 160 scientic papers and about 50 books in graph theory, combinatorics,
linear algebra and articial intelligence, including his best-known monograph,
Spectra of Graphs (with M. Doob and H. Sachs). In 1985 he was admitted to the
Serbian Academy of Sciences and Arts, and in 1991 he was appointed an Honorary
Professor at the University of Stirling, Scotland.
Michael Doob [[email protected]] received his ABdegree fromColum-
bia University and his Ph.D. from the City University of New York. His doctoral
supervisor was AlanHoffman, whointroducedhimtographspectra. After complet-
ing graduate school, he joined the faculty at the University of Manitoba, eventually
rising to the rank of professor. He is the author of dozens of research papers and
three books, two of which are directly related to spectral graph theory.
Bojan Mohar [[email protected]] obtained his Ph.D. from the Univer-
sity of Ljubljana, Slovenia, in 1986. He did his postdoctoral studies at Simon Fraser
University (Vancouver) and Ohio State University, where he was on a Fulbright
Scholarship. He is now a professor in the mathematics department of the Univer-
sity of Ljubljana. His areas of research include topological graph theory, graph
minors, graph colouring, graph algorithms, and algebraic graph theory, and he has
published over a hundred research papers (with over fty co-authors). He has also
co-authored a recent monograph, Graphs on Surfaces. In 1990 he was awarded
the Boris Kidric prize, the Slovenian national award for exceptional achievements
in science.
Cheryl Praeger [[email protected]] is a professor of mathematics at
the University of Western Australia, and a Fellow of the Australian Academy
Notes on contributors 269
of Science. She has published over 200 research articles and two monographs
on permutation groups, algorithms for group computation, and applications to
combinatorics, especially to vertex- and edge-transitive graphs. She initiated the
modern theory of nite quasi-primitive permutation groups and the study of nite
edge-transitive graphs via their normal quotients.
Peter Rowlinson [[email protected]] is a professor of mathematics at the
University of Stirling, Scotland, where he has taught since 1969. His Oxford
D.Phil. thesis was concerned with nite groups, and in 197576 he was a visiting
professor of mathematics at the California Institute of Technology. He has worked
in algebraic graph theory since 1980, and is co-author (with D. Cvetkovi c and S.
Simi c) of Eigenspaces of Graphs and Spectral Generalizations of Line Graphs.
Bryan Shader [[email protected]] is a professor of mathematics at the Univer-
sity of Wyoming in Laramie. His research interests focus on the interplay between
combinatorics, graphs, andmatrices. He is co-author withR. A. Brualdi of Matrices
of Sign-Solvable Linear Systems (1995). He is currently an associate editor of
the journals Linear Algebra and its Applications, and the Electronic Journal of
Linear Algebra, and is an editor of Image, the International Linear Algebra Soci-
etys newsletter.
Leonard Soicher [[email protected]] is a reader in mathematics at Queen
Mary, University of London. His research interests are groups, graphs, designs,
algorithms and computation. He is the author of the GAP package GRAPE for
computing with graphs and groups, and the co-author with Cheryl Praeger of Low
Rank Representations and Graphs for Sporadic Groups.
Robin Wilson [[email protected]] is head of the pure mathematics depart-
ment at the Open University, UK, and Gresham Professor of Geometry, London.
He graduated in mathematics from Oxford University, and received his Ph.D. in
number theory from the University of Pennsylvania. He has written and edited
many books on graph theory and combinatorics and on the history of mathemat-
ics, including Introduction to Graph Theory and Four Colours Sufce, and his
interests include graph colourings, spectral graph theory, and the history of graph
theory and combinatorics.
Index of denitions
Abelian, 19
absolute bound., 213
absolutely irreducible, 240
action, 22
adjacency matrix, 115
adjacency set, 251
adjacent edges, 2
adjacent vertices, 2
afne alternating group, 242
afne case, 239
afne groups of Lie type, 242
afne plane, 211
afne sporadic group, 245
afne type, 184
almost cospectral, 105
almost simple, 184
alternating group, 21, 246
amalgam, 185
angle matrix, 91
angle, 91
antipodal partition, 188
antipodal quotient, 188, 234
arc, 9
arc-transitive, 25, 180
associated, 234
association scheme, 218
associative law, 19
atom, 170
automorphism, 23, 137
automorphism group, 138, 257
balanced, 218
base, 252
Bass-Serre theory, 143
biadjacency matrix, 59
biclique, 67
biclique partition number, 67
bilinear forms graph, 239
bipartite, 4
bipartite double, 188
bipartite half, 188
bipartition, 59
bipartition width, 124
biprimitive, 187
bi-quasiprimitive, 193
brace, 79
bridge, 4
Brookss theorem, 8
canonical labelling map, 257
canonically labelling, 140
Cartesian product, 51, 118
Cayley graph, 145, 156
Cayley map, 174
characteristic polynomial, 89
characterized, 36
Cheeger inequalities, 128
Chen-Quimpo theorem, 172
chord, 79
chordal, 79
chromatic number, 8
CI-graph, 166
circulant graph, 157
circular chromatic number, 175
classical group, 216
Classication of nite simple groups,
21
Clebsch graph, 216
closed, 2
coduplicate vertices, 96
coherent conguration, 218
collapsed adjacency diagram, 229
271
272 Index of denitions
collapsed adjacency matrix, 260
coloured digraph, 65
commutative law, 19
compact, 148
complement, 6
complete bipartite graph, 5
complete graph, 5
complete k-partite graph, 5
completion, 185
component, 3
composite, 20
composition, 22
composition factors, 21
composition series, 21
conductance, 133
cone, 90
conference graph, 207
conjugate, 25
connected, 3
connection set, 157
connectivity, 4
connects, 2
contractible, 6
contraction, 6
convertible, 76
coordinate-wise action, 26
core, 25, 153
core-free, 25
correction function, 124
coset enumeration, 262
coset graph, 27
cospectral graphs, 119
Courant-Fisher formula, 120
cover, 188
covers, 69
cut-edge, 4
cut-vertex, 3
cycle, 2
cycle cover, 72
cycle graph, 5
cycle index, 168
cyclic group, 21
d-cube, 5
d-dimensional cube, 5, 119
degree, 2, 22
degree matrix, 115
deletion of a vertex, 3
deletion of an edge, 3
derived spectrum, 90
Deza graph, 219
diagonal orbital, 26
diameter, 3
digraph, 9
dihedral group, 24
dimensional cube, 188
direct product, 22, 138
directed, 61
directed cut, 62
directed edge, 9
directed graph, 2
discrete, 258
distance, 3
distance-regular, 38, 217, 231
distance-transitive, 187, 223
distance-transitivity, 148
distinct, 141
dominant, 50
Doob graphs, 52
double coset, 27
double coset diagram, 229
doubly stochastic, 60, 147
doubly transitive, 26, 224
dual, 208
dual linear spaces, 209
dual orthogonal graph, 191
duplicate vertices, 96
edge, 1
edge-chromatic number, 8
edge-connectivity, 4
edge-regular, 219
edge-set, 1
edge-transitive, 25
eigenvalue, 17
end, 149
equivalence, 27
equivalent, 23, 167
Eulers polyhedron formula, 7
Eulerian, 7
Eulerian trail, 7
eutactic star, 94
even graph, 105
even subdivision, 78
exceptional nite simple groups, 237
exceptional graph, 99
expander, 175
exponent, 64
extension vectors, 100
face, 7
factor group, 20
faithful, 185
feasible, 94
nding the automorphism group, 139
folding, 152
forest, 4
forward edge, 77
freely, 143
Frobenius normal form, 62
fully indecomposable, 71
fully indecomposable components, 72
fundamental set, 48
G-graph, 250
G-invariant, 183
G-invariant graph structure, 225
G-normal, 193
GAP, 254
generalized dicyclic group, 145
generalized Fitting subgroup, 239
generalized line graph, 46
generalized Petersen graph, 161
generic biadjacency matrix, 60
genus, 174
geometric, 208
Ger sgorins theorem, 59
girth, 3
go into, 9
go out of, 9
Graham-Pollak theorem, 68
Gram matrix, 99
GRAPE, 253
graph, 1, 214, 250
graph isomorphism, 139
graphical regular representation, 145,
169
Grassmann graph, 240
grid graph, 119
Grigorchuk group, 150
group, 19
group character, 235
group of automorphisms, 138
group of Lie type, 21, 246
growth, 150
H-maximal, 100
Hadwiger number, 146
half-arc transitive, 200
halved graph, 234
Hamilton decomposition, 173
Hamilton-connected, 171
Hamilton-laceable, 172
Hamiltonian, 7
Hamming graph, 39, 188, 229
head, 60
Heawood graph, 78
Hoffman polynomial, 36
Hoffman-Singleton graph, 38
homeomorphic, 6
homogeneous, 148
homomorphically equivalent, 152
homomorphism, 19, 152
identity, 19
imprimitive, 183, 215
imprimitivity, 233
imprimitivity sets, 63
in-degree, 9
incidence matrix, 60
incident, 2
independence ratio, 152
index, 20
index of imprimitivity, 63
induced, 3
innite graph, 2
insertion, 6
integral, 107
Interlacing theorem, 34
intersection array, 231
inverse, 19
irreducible, 62, 235
irreducible component, 62
isometric subgraph, 153
isomorphic, 23, 138
isomorphic factorization, 174
isomorphism, 19, 138
isoperimetric number, 128
isoperimetric problem, 127
Johnson graph, 39, 250
joins, 2
JordanH older theorem, 20
k-chromatic, 8
k-colourable, 8
k-connected, 4
k-dimensional octahedral graph, 5
k-edge-colourable, 8
k-edge-connected, 4
k-regular, 2
k-regular graph, 118
kernel, 20
Kneser graph, 153, 164
Krein condition, 207
Kronecker product, 51
Kuratowskis theorem, 8
l-multicover, 194
labelled graph, 138
Laplace eigenvalues, 117
Laplacian, 50, 115
Laplacian matrix, 60, 115
Latin square graph, 211
left coset, 20
length, 2
line graph, 8, 208
linear space, 209
local parameter, 256
locally-H, 147, 263
locally primitive, 194
locally quasiprimitive, 194
locally s-arc transitive, 199
loops, 2
main angle, 90
main eigenvalue, 90
Markov chain, 130
matching, 60
matching covered, 60
matching number, 69
matrix-tree theorem, 115
maximal, 184
maximum cut, 125
McLaughlin graph, 205
Mengers theorem, 4
minimal, 184
minimal generating Cayley set,
173
minimal vertex separator, 80
minor, 6
mixing time, 133
Moore graph, 37
multicover, 194
multigraph, 2
multiple, 146
multiple edges, 2
multiplicity-free, 236, 260
mutually orthogonal, 211
M obius ladder, 160
nauty, 255, 257
neighbour, 2
neighbourhood, 2
neighbourhood-regular, 147
NEPS (non-extended p-sum) of graphs,
52, 104
net, 210
non-complete extended p-sum, 52,
104
non-diagonal orbital, 26
non-separable, 3
normal, 169
normal cover, 194
normal multicover, 194
normal quotient, 193
normal subgroup, 20
null graph, 5
odd girth, 152, 181
of Lie type, 237
orbit, 24, 252
orbital, 26, 259
orbital digraph, 259
orbital graph, 26, 259
order, 1
orthogonal group, 216
out-degree, 9
p-sum, 104
pairing, 26, 141
Paley graph, 160, 207
parameter, 217
partial geometry, 208
partially balanced, 218
partite set, 4
path, 2
path graph, 5
perfect matching, 60
permanent, 72
permutation, 22
permutation group, 22, 138
permutation matrix, 147
permutation rank, 225
Perron-Frobenius theorem, 33,
58
planar graph, 7
point graph, 208
polar space, 216
polynomial partition, 93
positive matrix, 61
primitive, 28, 61, 65, 215
primitive symmetric pair, 186
principal axis theorem, 33
product, 104
projective group, 216
projective representation, 240
pseudo-geometric, 208
quasi-cyclic shift, 107
quasi-minimal, 171
quasiprimitive, 28, 193
quotient graph, 186
quotient group, 20
r-circular colouring, 175
random Cayley graph, 145
rank, 26, 214, 259
Rayleigh characterization, 120
Rayleigh quotient, 41
real, 235
real n-dimensional space, 10
Reconstruction theorem, 97
reducible, 62
renement, 258
renement process, 258
region, 7
regular, 2, 144, 158, 184, 220
retract, 152
right coset, 20
ring-like, 147
root system, 45
Schl ai graph, 205
Schreier coset graph, 146
Schreier graph, 252
Schreier tree, 252
Schreier vector, 252
Schreier-Sims algorithm, 253
Seidel matrix, 50
Seidel spectrum, 90
self-paired, 26, 226, 260
semi-denite program, 126
semi-regularly, 146
semi-symmetric, 199
Shrikhande graph, 52, 210
signed digraph, 75
signed weight, 75
signing, 76
simple graph, 2
simple random walk, 130
simple socle, 239
simplicial, 80
simplicial vertex, 80
simultaneous permutations, 61
singleton, 258
singular graph, 105
socle, 29, 184, 215
spanning, 3
special bound, 213
spectral radius, 58
spectrum, 89
spherical t-design, 213
sporadic groups, 21, 237, 247
square, 59
square lattice graph, 210
square-root, 151
stabilizer, 24
star closed, 45
star complement, 89, 96
star partition, 94
star set, 89, 94
stationary distribution, 131
Steiner system, 209, 210
strong, 9
strong component, 9, 62
strong generating set, 253
strong graph, 219
strong product, 104
strongly connected, 9, 58
strongly regular, 36, 204
subdominant, 50
subgraph, 3
subgroup, 20
subnormal subgroup, 239
suborbit, 259
sum, 104
switching, 219
switching-equivalent, 101
symmetric, 25, 179, 262
symmetric group, 22
symmetric pair, 25
symmetric travelling salesman
problem, 129
symmetrically, 262
t-homogeneous, 149
t-tuple regular, 205
tail, 60
tournament, 9, 67
traceable, 7
trail, 2
transition Laplace matrix, 117
transition probabilities, 130
transitive, 24
transitive graph, 105
transversal design, 210
travelling salesman problem, 129
triangular graph, 216
trivial partition, 183
union, 6
unique countable random graph, 151
universal, 142
universal K
n
-free, 151
universal completion, 185
unlabelled graph, 138
vector, 10
vertex, 1, 9
vertex separator, 80
vertex-edge incidence matrix, 44
vertex-primitive, 146
vertex-set, 1
vertex-transitive, 25, 145, 259
vertex-transitively, 25
Vizings theorem, 8
walk, 2, 61
walk-regular, 91, 218
weight, 2, 61, 75
weight function, 115
weight of a perfect matching, 72
weighted, 115
weighted graph, 2, 115, 138
Weisss theorem, 189
Witt design, 142
wreath product, 22, 138

Lowell W. Beineke & Robin J. Wilson & Peter J. Cameron - Topics in Algebraic Graph Theory

Uploaded by

Copyright:

Available Formats

Lowell W. Beineke & Robin J. Wilson & Peter J. Cameron - Topics in Algebraic Graph Theory

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lowell W. Beineke & Robin J. Wilson & Peter J. Cameron - Topics in Algebraic Graph Theory

Uploaded by

Copyright:

Available Formats

Topics in Algebraic Graph Theory

C and PETER ROWLINSON

(v) of a vertex v in a digraph is the number of arcs that go

(v) is the number of arcs that go into it.

= {v V : v, w) = 0 for all w W].

h more briey as gh.

, the orbital O is self-paired; trivially, a diagonal

) is the converse digraph of G(O).

is the coset graph G(G, H, HgH).

4r 3 is an odd integer, 2t +1, so that 2t +1 divides

2, with equality if and only if

2. Among the graphs G with four

2 and meeting at angles of 60

, then we can add four

, we can repeat this

5 1)/2 have also been de-

5, ), but nowhere dense in

2 1). Further details appear in [12], [18] and [26].

. This implies that

1. The argument used in the

5 1)/2, Discrete Math. 139 (1995), 213227.

5 1)/2, Linear and Multilinear Algebra 39 (1995), 5971.

denotes the union over all cycles of G, and

denotes the product

only if 1 q < p k. Ordering the vertices of V so that

that is, on the zero or

are subadditive functions

(A) k. The theorem now follows

of size (r 1) +(s 1) (d(A) 1) =

along with the edge joining x

(n, k) be the minimum number of

. This possibility is considered in the next section.

. The signed weight s() of a

has no cycles of even length.

has a cycle of even length. This was a

is a bipartite graph which has a perfect matching and is

C and PETER ROWLINSON

has characteristic polynomial

. It is well known that if G is exceptional, then

{a(T) : T a spanning tree of G] = [det L

4 Graph Laplacians 117

(G) i (G). It turns out that i

(G) satises a relation similar

be the initial probability distribution

= {(v, w) : (w, v) O]. Now

, where the connection set is the set of standard

1 (mod p). Formthe circulant graph H

NC, with the exceptions

= aS, for some a Z

) be isomorphic circulant graphs on Z

Ad am posed the question of whether two circulant graphs Circ(n; S) and

) are isomorphic if and only if there exists an a Z

= {1, 4, 6, 9, 10, 11, 14, 15, 16, 19, 21, 24}.

) are isomorphic, since each

) on G are isomorphic if and only if there exists a group automorphism

= S; then we say that G is a CI-group. Rephrasing

, then the elementary

be the graph with vertex-set V

is called a bipartite half of G, (G

is the index-2 subgroup of G that xes V