ALGEBRAIC GRAPH THEORY

ALGEBRAIC GRAPH THEORY

Second Edition

NORMAN BIGGS
London School of Economics

CAMBRIDGE
UNIVERSITY PRESS
CAMBRIDGE u n i v e r s i t y p r e s s
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,
São Paulo, Delhi, Dubai, Tokyo, Mexico City

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK

Published in the United States of America by

Cambridge University Press, New York

www.cambridge.org
Information on this title: www.cambridge.org/9780521458979

© Cambridge University Press 1974, 1993

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 1974

Second edition 1993
Reprinted 1996

A catalogue recordfor this publication is available from the British Library

Library of Congress cataloguing in publication data available

ISBN 978-0-521-20335-7 Hardback

ISBN 978-0-521-45897-9 Paperback

Cambridge University Press has no responsibility for the persistence or

accuracy of URLs for external or third-party internet websites referred to in
this publication, and does not guarantee that any content on such websites is,
or will remain, accurate or appropriate. Information regarding prices, travel
timetables, and other factual information given in this work is correct at
the time of first printing but Cambridge University Press does not guarantee
the accuracy of such information thereafter.
 Contents

Preface vii
1 Introduction 1
PART ONE - LINEAR ALGEBRA IN GRAPH THEORY
2 The spectrum of a graph 7
3 Regular graphs and line graphs 14
4 Cycles and cuts 23
5 Spanning trees and associated structures 31
6 The tree-number 38
7 Deteminant expansions 44
8 Vertex-partitions and the spectrum 52
PART TWO - COLOURING PROBLEMS
9 The chromatic polynomial 63
10 Subgraph expansions 73
11 The multiplicative expansion 81
12 The induced subgraph expansion 89
13 The Tutte polynomial 97
14 Chromatic polynomials and spanning trees 106
PART THREE - SYMMETRY AND REGULARITY
15 Automorphisms of graphs 115
16 Vertex-transitive graphs 122
17 Symmetric graphs 130
vi Contents

18 Symmetric graphs of degree three 138

19 The covering-graph construction 149
20 Distance-transitive graphs 155
21 Feasibility of intersection arrays 164
22 Imprimitivity 173
23 Minimal regular graphs with given girth 180
References 191
Index 202
 Preface

This book is a substantially enlarged version of the Cambridge Tract

with the same title published in 1974. There are two major changes.
• The main text has been thoroughly revised in order to clarify the
exposition, and to bring the notation into line with current practice.
In the course of revision it was a pleasant surprise to find that the
original text remained a fairly good introduction to the subject, both
in outline and in detail. For this reason I have resisted the temptation
to reorganise the material in order to make the book rather more like a
standard textbook.
• Many Additional Results are now included at the end of each
chapter. These replace the rather patchy selection in the old version,
and they are intended to cover most of the major advances in the last
twenty years. It is hoped that the combination of the revised text and
the additional results will render the book of service to a wide range of
readers.
I am grateful to all those people who have helped by commenting upon
the old version and the draft of the new one. Particular thanks are due
to Peter Rowlinson, Tony Gardiner, Ian Anderson, Robin Wilson, and
Graham Brightwell. On the practical side, I thank Alison Adcock, who
prepared a TgX version of the old book, and David Tranah of Cambridge
University Press, who has been constant in his support.

Norman Biggs March 1993

 Introduction to algebraic graph theory

About the book

This book is concerned with the use of algebraic techniques in the study
of graphs. The aim is to translate properties of graphs into algebraic
properties and then, using the results and methods of algebra, to deduce
theorems about graphs.
It is fortunate that the basic terminology of graph theory has now be-
come part of the vocabulary of most people who have a serious interest
in studying mathematics at this level. A few basic definitions are gath-
ered together at the end of this chapter for the sake of convenience and
standardization. Brief explanations of other graph-theoretical terms are
included as they are needed. A small number of concepts from matrix
theory, permutation-group theory, and other areas of mathematics, are
used, and these are also accompanied by a brief explanation.
The literature of algebraic graph theory itself has grown enormously
since 1974, when the original version of this book was published. Liter-
ally thousands of research papers have appeared, and the most relevant
ones are cited here, both in the main text and in the Additional Re-
sults at the end of each chapter. But no attempt has been made to
provide a complete bibliography, partly because there are now several
books dealing with aspects of this subject. In particular there are two
books which contain massive quantities of information, and on which it
is convenient to rely for 'amplification and exemplification' of the main
results discussed here.
2 Introduction to algebraic graph theory

These are:
Spectra of Graphs: D.M. Cvetkovic, M. Doob, and H. Sachs, Academic
Press (New York) 1980.
Distance-Regular Graphs: A.E. Brouwer, A.M. Cohen, and A. Neumaier,
Springer-Verlag (Berlin) 1989.
References to these two books are given in the form [CvDS, p. 777], and
[BCN, p. 888].
CD. Godsil's recent book Algebraic Combinatorics (Chapman and
Hall, 1993) arrived too late to be quoted as reference. It is in many
ways complementary to this book, since it covers several of the same
topics from a different point of view. Finally, the long-awaited Handbook
of Combinatorics will contain authoritative accounts of many subjects
discussed in these pages.

Outline of the book

The book is in three parts, each divided into a number of short chap-
ters. The first part deals with the applications of linear algebra and
matrix theory to the study of graphs. We begin by introducing the ad-
jacency matrix of a graph; this matrix completely determines the graph,
and its spectral properties are shown to be related to properties of the
graph. For example, if a graph is regular, then the eigenvalues of its
adjacency matrix are bounded in absolute value by the degree of the
graph. In the case of a line graph, there is a strong lower bound for the
eigenvalues. Another matrix which completely describes a graph is the
incidence matrix of the graph. This matrix represents a linear mapping
which determines the homology of the graph. The problem of choosing
a basis for the homology of a graph is just that of finding a fundamental
system of cycles, and this problem is solved by using a spanning tree.
At the same time we study cuts in the graph. These ideas are then
applied to the systematic solution of network equations, a topic which
supplied the stimulus for the original theoretical development. We then
investigate formulae for the number of spanning trees in a graph, and
results which are derived from the expansion of determinants. These
expansions illuminate the relationship between a graph and the charac-
teristic polynomial of its adjacency matrix. The first part ends with a
discussion of how spectral techniques can be used in problems involving
partitions of the vertex-set, such as the vertex-colouring problem.
The second part of the book deals with the colouring problem from a
different point of view. The algebraic technique for counting the colour-
ings of a graph is founded on a polynomial known as the chromatic
 Introduction to algebraic graph theory 3

polynomial. We first discuss some simple ways of calculating this poly-

nomial, and show how these can be applied in several important cases.
Many important properties of the chromatic polynomial of a graph stem
from its connection with the family of subgraphs of the graph, and we
show how the chromatic polynomial can be expanded in terms of sub-
graphs. From the first (additive) expansion another (multiplicative)
expansion can be derived, and the latter depends upon a very restricted
class of subgraphs. This leads to efficient methods for approximating
the chromatic polynomials of large graphs. A completely different kind
of expansion relates the chromatic polynomial to the spanning trees of a
graph; this expansion has several remarkable features and leads to new
ways of looking at the colouring problems, and some new properties of
chromatic polynomials.
The third part of the book is concerned with symmetry and regularity
properties. A symmetry property of a graph is related to the existence
of automorphisms - that is, permutations of the vertices which pre-
serve adjacency. A regularity property is defined in purely numerical
terms. Consequently, symmetry properties induce regularity properties,
but the converse is not necessarily true. We first study the elementary
properties of automorphisms, and explain the connection between the
automorphisms of a graph and the eigenvalues of its adjacency matrix.
We then introduce a hierarchy of symmetry conditions which can be
imposed on a graph, and proceed to investigate their consequences. The
condition that all vertices be alike (under the action of the group of auto-
morphisms) turns out to be rather a weak one, but a slight strengthening
of it leads to highly non-trivial conclusions. In fact, under certain condi-
tions, there is an absolute bound to the level of symmetry which a graph
can possess. A strong symmetry property, called distance-transitivity,
and the consequent regularity property, called distance-regularity, are
then introduced. We return to the methods of linear algebra to derive
numerical constraints upon the existence of graphs with these properties.
Finally, these constraints are applied to the problem of finding minimal
regular graphs whose degree and girth are given.

Basic definitions and notation

Formally, a general graph F consists of three things: a set VT, a set £T,
and an incidence relation, that is, a subset of VT x ET. An element
of VT is called a vertex, an element of £T is called an edge, and the
incidence relation is required to be such that an edge is incident with
either one vertex (in which case it is a loop) or two vertices. If every
4 Introduction to algebraic graph theory

edge is incident with two vertices, and no two edges are incident with
the same pair of vertices, then we say that F is a strict graph or briefly,
a graph. In this case, ET can be regarded as a subset of the set of
unordered pairs of vertices. We shall deal mainly with graphs (that is,
strict graphs), except in Part Two, where it is sometimes essential to
consider general graphs.
If v and w are vertices of a graph F, and e = {v, w] is an edge of F,
then we say that e joins v and w, and that v and w are the ends of e.
The number of edges of which v is an end is called the degree of v. A
subgraph of F is constructed by taking a subset S of ET together with
all vertices incident in F with some edge belonging to S. An induced
subgraph of F is obtained by taking a subset U of VT together with
all edges which are incident in F only with vertices belonging to U. In
both cases the incidence relation in the subgraph is inherited from the
incidence relation in F. We shall use the notation (S)r, (U)r for these
subgraphs, and usually, when the context is clear, the subscript F will
be omitted.
 PART ONE

Linear algebra in graph theory

 The spectrum of a graph

We begin by defining a matrix which will play an important role in many

parts of this book. Suppose that F is a graph whose vertex-set VT is
the set {vi,V2,-.,vn}, and consider ET as a set of unordered pairs of
elements of VT. If {vi,Vj} is in ET, then we say that Vi and Vj are
adjacent.
Definition 2.1 The adjacency matrix of F is the n x n matrix A —
A(F) whose entries a^ are given by
1 if Vi and Vj are adjacent;
I' otherwise.
to.
For the sake of definiteness we consider A as a matrix over the complex
field. Of course, it follows directly from the definition that A is a real
symmetric matrix, and that the trace of A is zero. Since the rows and
columns of A correspond to an arbitrary labelling of the vertices of
F, it is clear that we shall be interested primarily in those properties
of the adjacency matrix which are invariant under permutations of the
rows and columns. Foremost among such properties are the spectral
properties of A.
Suppose that A is an eigenvalue of A. Then, since A is real and sym-
metric, it follows that A is real, and the multiplicity of A as a root of
the equation det(AI — A) = 0 is equal to the dimension of the space of
eigenvectors corresponding to A.
8 Linear algebra in graph theory

Definition 2.2 The spectrum of a graph F is the set of numbers which

are eigenvalues of A(F), together with their multiplicities. If the distinct
eigenvalues of A(F) are Ao > Aj > . . . > A s _i, and their multiplicities
are m(Ao),m(Ai),... ,m(A s _!), then we shall write

SpecF= ( A ° Al
"• As
~J
\m(X0) m(Ai)
For example, the complete graph Kn is the graph with n vertices in
which each distinct pair are adjacent. Thus the graph K4 has adjacency
matrix
-0 1 1 1-
1 0 1 1
A =
1 1 0 1
.1 1 1 0 .
and an easy calculation shows that the spectrum of K4 is
'3 - 1 "
Spec K4 = , „

We shall usually refer to the eigenvalues of A = A(F) as the eigenval-

ues ofT. Also, the characteristic polynomial det(AI —A) will be referred
to as the characteristic polynomial o/F, and denoted by x(F; A). Let us
suppose that the characteristic polynomial of F is
X(F; A) = An + dA"" 1 + c2A"~2 + C3A"-3 + ... + cn.
In this form we know that — c\ is the sum of the zeros, that is, the sum
of the eigenvalues. This is also the trace of A which, as we have already
noted, is zero. Thus c\ — 0. More generally, it is proved in the theory
of matrices that all the coefficients can be expressed in terms of the
principal minors of A, where a principal minor is the determinant of a
submatrix obtained by taking a subset of the rows and the same subset
of the columns. This leads to the following simple result.

Proposition 2.3 The coefficients of the characteristic polynomial of

a graph F satisfy:
(1) ci = 0;
(2) — c-i is the number of edges of F;
(3) —C3 is twice the number of triangles in F.
Proof For each i s {1,2, ...,n}, the number (—l)*Ci is the sum of
those principal minors of A which have i rows and columns. So we can
argue as follows.
(1) Since the diagonal elements of A are all zero, c\ = 0.
(2) A principal minor with two rows and columns, and which has a
 The spectrum of a graph

non-zero entry, must be of the form

0 1
1 0
There is one such minor for each pair of adjacent vertices of F, and each
has value —1. Hence (-l) 2 c 2 = — \ET\, giving the result.
(3) There are essentially three possibilities for non-trivial principal
minors with three rows and columns:
0 1 0 0 1 1 0 1 1
1 0 0 5 1 0 0 1 0 1
0 0 0 1 0 0 1 1 0
and, of these, the only non-zero one is the last (whose value is 2). This
principal minor corresponds to three mutually adjacent vertices in F,
and so we have the required description of C3. •

These simple results indicate that the characteristic polynomial of a

graph is an object of the kind we study in algebraic graph theory: it is
an algebraic construction which contains graphical information. Propo-
sition 2.3 is just a pointer, and we shall obtain a more comprehensive
result on the coefficients of the characteristic polynomial in Chapter 7.
Suppose A is the adjacency matrix of a graph F. Then the set of
polynomials in A, with complex coefficients, forms an algebra under
the usual matrix operations. This algebra has finite dimension as a
complex vector space. Indeed, the Cayley-Hamilton theorem asserts
that A satisfies its own characteristic equation, so the dimension is at
most n, the number of vertices in F.
Definition 2.4 The adjacency algebra of a graph F is the algebra of
polynomials in the adjacency matrix A = A(F). We shall denote the
adjacency algebra of F by A(T).
Since every element of the adjacency algebra is a linear combination
of powers of A, we can obtain results about -4(F) from a study of these
powers. We define a walk of length I in F, from Vi to Vj, to be a finite
sequence of vertices of F,
Vi = U0,Ui,...,Ui = Vj,

such that Ut_i and are adjacent for 1 < t <

Lemma 2.5 The number of walks of length I in F, from Vi to Vj, is
the entry in position (i, j) of the matrix A1.
Proof The result is true for I = 0 (since A 0 = I) and for / = 1 (since
A 1 = A is the adjacency matrix). Suppose that the result is true for
/ = L. The set of walks of length L + 1 from Vi to Vj is in bijective
10 Linear algebra in graph theory

correspondence with the set of walks of length L from Vi to vertices Vh

adjacent to Vj. Thus the number of such walks is

J2 (A L k = JT(AL)ihahj = {AL+l)i3.
{vh,vj}€ ET h=l
It follows that the number of walks of length L + 1 joining Vi to Vj is
(AL+1)ij. The general result follows by induction. •

A graph is said to be connected if each pair of vertices is joined by

a walk. The number of edges traversed in the shortest walk joining v^
and Vj is called the distance in F between vt and Vj and is denoted by
d(vi,Vj). The maximum value of the distance function in a connected
graph F is called the diameter of F.
Proposition 2.6 Let F be a connected graph with adjacency algebra
A(T) and diameter d. Then the dimension of A{T) is at least d+1.
Proof Let x and y be vertices of F such that d(x, y) = d, and suppose
that
x = wo,wi,...,Wd = y
is a walk of length d. Then, for each i 6 {1,2,..., d}, there is at least one
walk of length i, but no shorter walk, joining WQ to Wi. Consequently,
A* has a non-zero entry in a position where the corresponding entries of
I, A, A 2 , . . . , A1'1 are zero. It follows that A* is not linearly dependent
on {I, A , . . . , A*^ 1 }, and that {I, A , . . . , A d } is a linearly independent
set in A(T). Since this set has d+1 members, the proposition is proved.
•
There is a close connection between the adjacency algebra and the
spectrum of F. If the adjacency matrix has s distinct eigenvalues then,
since it is a real symmetric matrix, its minimum polynomial (the monic
polynomial of least degree which annihilates it) has degree s. Conse-
quently the dimension of the adjacency algebra is equal to s. Thus we
have the following bound for the number of distinct eigenvalues.
Corollary 2.7 A connected graph with diameter d has at least d+1
distinct eigenvalues. •
One of the major topics of the last part of this book is the study of
a class of 'highly regular' connected graphs which have the minimum
number d + 1 of distinct eigenvalues. In the following chapters we shall
encounter several other examples of the link between structural regular-
ity and the spectrum.
 The spectrum of a graph 11

Notation The eigenvalues of a graph may be be listed in two ways: in

strictly decreasing order of the distinct values, as in Definition 2.2, or in
weakly decreasing order (with repeated values) Ao > Aj > ... > An_i,
where n = |VT|. We shall use either method, as appropriate.

Additional Results
2a A reduction formula for x Suppose F is a graph with a vertex
vi of degree 1, and let V2 be the vertex adjacent to v\. Let T\ be
the induced subgraph obtained by removing v\, and Fi2 the induced
subgraph obtained by removing {ui,^}- Then

This formula can be used to calculate the characteristic polynomial of

any tree, because a tree always has a vertex of degree 1. A more general
reduction formula was found by Rowlinson (1987).
2b The characteristic polynomial of a path Let Pn be the path graph
with vertex-set {vi, V2, • • •, vn} and edges {vi, Vi+i} (1 < i < n — 1). For
n > 3 we have
X(Pn; A) = A X (P n -i; A) - X (P n _ 2 ; A).
Hence x(Pni A) = Un(X/2), where Un denotes the Chebyshev polynomial
of the second kind.
2c The spectrum of a bipartite graph A graph is bipartite if its vertex-
set can be partitioned into two parts \\ and V% such that each edge has
one vertex in Vi and one vertex in V-i- If we order the vertices so that
those in V\ come first, then the adjacency matrix of a bipartite graph
takes the form
I B

If x is an eigenvector corresponding to the eigenvalue A, and x is obtained

from x by changing the signs of the entries corresponding to vertices in
V2, then x is an eigenvector corresponding to the eigenvalue —A. It
follows that the spectrum of a bipartite graph is symmetric with respect
to 0, a result originally obtained by Coulson and Rushbrooke (1940) in
the context of theoretical chemistry.
2d The derivative of \ F° r i = 1, 2,..., n let Fj denote the induced
subgraph (VT\«i). Then
12 Linear algebra in graph theory

2e The eigenvalue 0 Suppose that a graph has two vertices Vi and Vj

such that the set of vertices adjacent to v^ is the same as the set of ver-
tices adjacent to Vj. Then the vector x whose only non-zero components
are Xi = 1 and Xj = —1 is an eigenvector of the adjacency matrix, with
eigenvalue 0. If F has a set of r vertices, all of which have the same set
of neighbours, then the multiplicity of 0 is at least r — 1. (An alternative
argument uses the observation that there are r equal columns of A, and
so its rank is at most n — r + 1.)

2f Cospectral graphs Two non-isomorphic graphs are said to be cospec-

tral if they have the same eigenvalues with the same multiplicities. The
first example of this phenomenon was given by Collatz and Sinogowitz
(1957), and many examples are given in [CvDS, pp. 156-161]. Two con-
nected graphs with 6 vertices, both having the characteristic polynomial
A6 - 7A4 - 4A3 + 7A2 + 4A - 1, are shown in Figure 1.

Figure 1: two cospectral graphs.

2g The walk-generating matrix Let gij (r) denote the number of walks
of length r in F from Vi to Vj. If we write G(z) for the matrix

r=l
l
then G(z) = (I — zA) , where A is the adjacency matrix of F. This
may be regarded as a matrix over the ring of formal power series in z,
or as a real matrix defined whenever z fi SpecF. From the formula for
the inverse matrix and 2e, we obtain

trG(z) =

2h Closed walks and sums of powers of eigenvalues A closed walk is

one whose initial and final vertices coincide. By Lemma 2.5 the total
number of closed walks of length I is equal to tr A'. Since the trace of a
matrix is the sum of its eigenvalues, an alternative expression is Yl^H-
In particular, the sum of the eigenvalues is zero, the sum of the squares
 The spectrum of a graph 13

is twice the number of edges, and the sum of the cubes is six times the
number of triangles.

2i An upper bound for the largest eigenvalue Suppose that the eigen-
values of F are Ao > Ai > ... > An_i, where F has n vertices and m
edges. Prom 2h we obtain 52 A^ = 0 and J2 ^t = ^rn- I* follows that

Another bound of the same type is Ao < \Z2ni — n + 1 (Yuan 1988).

2j The spectral decomposition (Godsil and Mohar 1988) The adjacency
matrix has a spectral decomposition A = 52^aE a , where the matrices
E a are idempotent and mutually orthogonal. It is easy to check that,
given a set of mutually orthonormal eigenvectors x a , we can take
E a = x a x*, that is (Ea)ij = (x a )j(x a )j.
It follows that if / is any function for which /(A) is defined, then /(A) =
]P /(A a )E a . For example, the walk-generating matrix G(z) — (I-^A)" 1
is defined whenever z $. SpecF, and it can be expressed in the form

a=0
This yields the following expression for the individual walk-generating
functions:
n-l

a=0

2k The distance matrices For a graph with diameter d the distance

matrices A/j (0 < h < d) are defined as follows:
(A )••3 = •[ 1 ' if ®(Vu Vj^ ~ h]
\ 0, otherwise.
It follows that
Ao = I, Ai = A, A o + Ai + A2 + ... + Ad = J,
where J is the matrix in which each entry is 1. The distance matrix
Ah can be expressed as a polynomial of degree h in A, for each h in
{0,1,..., d}, if and only if the graph is distance-regular (see Chapter
20). For such a graph the adjacency algebra has the minimum possible
dimension d+1.
 Regular graphs and line graphs

In this chapter we discuss graphs which possess some kinds of combi-

natorial regularity, and whose spectra, in consequence, have distinctive
features. A graph is said to be regular of degree k (or k-regular) if each of
its vertices has degree k. This is the most obvious kind of combinatorial
regularity, and it has interesting consequences for the eigenvalues.
Proposition 3.1 Let T be a regular graph of degree k. Then:
(1) k is an eigenvalue o/F;
(2) ifT is connected, then the multiplicity of k is 1;
(3) for any eigenvalue A ofT, we have \X\ < k.
Proof (1) Let u = [ 1 , 1 , . . . , 1]*; then if A is the adjacency matrix of
F we have A u = fcu, since there are k l's in each row. Thus fc is an
eigenvalue of F.
(2) Let x = [aJi,X2,...,a; n ]* denote any non-zero vector for which
A x = fcx, and suppose that Xj is an entry of x with the largest absolute
value. Since (Ax)j = kxj, we have
ZJ X$ ~-— KXj

where £ ' denotes summation over those k vertices Vi which are adjacent
to Vj. By the maximal property of Xj, it follows that xt = Xj for all
these vertices. If F is connected we may proceed successively in this way,
eventually showing that all entries of x are equal. Thus x is a multiple
of u, and the space of eigenvectors associated with the eigenvalue k has
dimension 1.
(3) Suppose that Ay = Ay, y ^ 0, and let yj denote an entry of y
 Regular graphs and line graphs 15

which is largest in absolute value. By the same argument as in (2), we

have 'S'yi = Xyj, and so
\\\\yj\ = \X'yi\<V\yi\<k\yj\.
Thus |A| < k, as required. •

The adjacency algebra of a regular connected graph also has a distinc-

tive property, related to the results of Proposition 3.1. Let J denote the
matrix each of whose entries is +1. Then, if A is the adjacency matrix
of a regular graph of degree k, we have AJ = JA = kJ. This is the point
of departure for the following result.
Proposition 3.2 (Hoffman 1963) The matrix J belongs to the adja-
cency algebra A(T) if and only if T is a regular connected graph.
Proof Suppose J is in A(T). By the definition of A(T), J is a polyno-
mial in A; consequently AJ = JA. Now, if k^ denotes the degree of the
vertex vit then (AJ)^ = k^ and (AJ),., = k^\ so that all the degrees
are equal and F is regular. Further, if F were disconnected we could
find two vertices with no walks joining them, so that the corresponding
entry of A1 would be zero for all I > 0. Then every polynomial in A
would have a zero entry, contradicting the fact that J € A(F). Thus F
is connected.
Conversely, suppose that F is connected and regular of degree k. Then,
by part (1) of Proposition 3.1, k is an eigenvalue of F, and so the mini-
mum polynomial of A is of the form p{\) = (A — k)q{\). Since p{A) = 0,
we obtain Aq(A) = kq(A); that is, each column of q(A) is an eigenvec-
tor of A corresponding to the eigenvalue k. By part (2) of Proposition
3.1, it follows that each column of q(A) is a multiple of u, and since q(A)
is a symmetric matrix, it is a multiple of J. Thus J is a polynomial in
A. •

Corollary 3.3 Let T be a k-regular connected graph with n vertices,

and let the distinct eigenvalues ofT be k > Xi > ... > As_i. Then if
q(X) = ]J(X - Xi), where the product is over the range 1 < i < s - 1, we
have

Proof It follows from the proof of Proposition 3.2 that q(A) = aJ,
for some constant a. Now the eigenvalues of q(A) are q(k) and q(Xi) for
1 < i < s — 1, and all of these except q(k) are zero. The only non-zero
eigenvalue of aJ is an, hence a = q(k)/n. •
16 Linear algebra in graph theory

For some classes of regular graphs, such as the strongly regular graphs
(3c), it is possible to determine the polynomial function / for which
/(A) = J by direct means, based on Lemma 2.5. This provides a pow-
erful method for determining the spectra of these graphs. At a more
basic level, there is a special class of regular graphs whose spectra can
be found by means of a well-known technique in matrix theory. As
this class contains several important families of graphs, we shall briefly
review the relevant theory.
An nxn matrix S is said to be a circulant matrix if its entries satisfy
Sij — Sij-i+i, where the subscripts are reduced modulo n and lie in the
set {1,2,..., n}. In other words, row i of S is obtained from the first
row of S by a cyclic shift of i — 1 steps, and so any circulant matrix is
determined by its first row. Let W denote the circulant matrix whose
first row is [0,1,0,..., 0], and let S denote a general circulant matrix
whose first row is [si,S2,...,s n ]. Then a straightforward calculation
shows that

Since the eigenvalues of W are 1, w, w2,..., un~1, where ui = exp(27ri/n),

it follows that the eigenvalues of S are

Definition 3.4 A circulant graph is a graph T whose vertices can be

ordered so that the adjacency matrix A(F) is a circulant matrix.

The adjacency matrix is a symmetric matrix with zero entries on the

main diagonal. It follows that, if the first row of the adjacency matrix
of a circulant graph is [a\, 02,...,a n ], then a: — 0 and o^ = an-i+2 for
i-2,...,n.
Proposition 3.5 Suppose that [0,d2, • • • ,an] is the first row of the
adjacency matrix of a circulant graph F. Then the eigenvalues ofT are

3=2
Proof This result follows directly from the expression for the eigen-
values of a circulant matrix. •

We remark that the n eigenvalues given by the formula of Proposition

3.5 are not necessarily all distinct.
 Regular graphs and line graphs 17

We shall give three examples of this technique. First, the complete

graph Kn is a circulant graph: the first row of its adjacency matrix is
[ 0 , l , l , . . . , l ] . Since
1 + ujr + ... + o; (n ~ 1)r = 0 for r € {1,2,..., n - 1},
it follows from Proposition 3.5 that the spectrum of Kn is:
_(n-\ -1 \
Spec Kn
~\ 1 n-l)-
Our second example is the cycle graph C n , whose adjacency matrix
is a circulant matrix with first row [0,1,0,... ,0,1]. In the notation
of Proposition 3.5, the eigenvalues are Ar = 2cos(2?rr/n), but these
numbers are not all distinct; taking account of coincidences the complete
description of the spectrum is:
c n _ (^ ^cos 2Kjri ... 2 cos(n —
bpecCn-^j 2 2

n (2 2cos27r/n ... 2cos(n-2)

bpec L n — I 1
VI
o
Z ...
o
L 1J
A third family of circulant graphs are the graphs Hs obtained by re-
moving s disjoint edges from K2S- The graph Hs is sometimes known as
a hyperoctahedral graph, because it is the skeleton of a hyperoctahedron
in s dimensions. It is also known as the cocktail-party graph CP(s)y so
called because it is alleged that if there are s married couples at a cock-
tail party each person talks to everyone except their spouse. It is also
a special kind of complete multipartite graph, to be defined in Chapter
6. Clearly, the graph Hs is a circulant graph; we may take the first row
of its adjacency matrix to be [oi,..., a2S], where each entry is 1, except
that a\ = a s +i = 0. It follows that the eigenvalues of Hs are
Ao = 2 s - 2 , Ar = - l - a ; r s (1 < r < 2s - 1),
where UJ2S = 1 and u) ^ 1. Consequently,

s-2 0 -2

We now turn to another structural property which has implications for

the spectrum of a graph. The line graph L(T) of a graph F is constructed
by taking the edges of F as vertices of L(F), and joining two vertices in
L(V) whenever the corresponding edges in F have a common vertex.
The spectra of line graphs were investigated extensively by Hoffman
(1969) and others. Here we outline the basic results; more recent work
is described in the Additional Results at the end of the chapter.
18 Linear algebra in graph theory

We shall continue to suppose that F has n vertices, V\,V2,-. • ,vn. We

shall need to label the edges of F also; that is, ET = {e\, e^,..., em}. For
the purposes of this chapter only we define a n n x m matrix X = X(F)
as follows:
f l , Xvi and ej are incident;
\ 0, otherwise.
othe

Lemma 3.6 Suppose that F and X are as above. Let A denote the
adjacency matrix ofT and AL the adjacency matrix of L(T). Then:
(1) X*X = AL + 2I m ;
(2) if F is regular of degree k, then XX* = A + kln.
The subscripts denote the sizes of the identity matrices.
Proof (1) We have

from which it follows that (X*X)jj is the number of vertices vi of F

which are incident with both the edges e; and ej. The required result is
now a consequence of the definitions of L(T) and A^,.
(2) This part is proved by a similar counting argument. D

Proposition 3.7 / / A is an eigenvalue of a line graph L(T), then

A> - 2 .
Proof The matrix X*X is non-negative definite, since we have z*X*Xz
= ||Xz|| 2 > 0, for any vector z. Thus the eigenvalues of X*X are non-
negative. But Ai = X*X — 2I m , so the eigenvalues of AL are not less
than - 2 . •

The condition that all eigenvalues of a graph be not less than —2

is a restrictive one, but it is not sufficient to characterize line graphs.
For example, the hyperoctahedral graphs Hs satisfy this condition, but
these graphs are not line graphs. Seidel (1968, see 3g) gave examples
of regular graphs which have least eigenvalue —2 and are neither line
graphs nor hyperoctahedral graphs. Subsequently, a characterization of
all graphs with least eigenvalue —2 was obtained by Cameron, Goethals,
Seidel, and Shult (1976, see 3i).
When F is a regular graph of degree k, its line graph L(T) is regular
of degree 2/c — 2. We can think of this as a connection between the
maximum eigenvalues of F and L(T), and in fact the connection extends
to all eigenvalues, by virtue of the following result.
 Regular graphs and line graphs 19

Theorem 3.8 (Sachs 1967) IfT is a regular graph of degree k with n

vertices and m = |nfc edges, then
X(L(T); A) = (A + 2)m-nX(T; X + 2-k).
Proof We shall use the notation and results of Lemma 3.6. Define
two partitioned matrices with n + m rows and columns as follows:
TT _ AIn —X _. _ Ira X
LU lm J [A Aim
Then we have
A I™« - X-"--<»•
X* 0 ] TAIn 0
TTV —
[ -<v
" A/TT —
Al m J
J L
[ AJV Aljn - A A

Since det(UV) = det(VU) we deduce that

Amdet(AIn - XX*) = A"det(AIm - X*X).
Thus we may argue as follows:
x(£(r);A)=det(AIm-AL)
= det((A + 2)I m - X*X)
= (A + 2) m -"det((A + 2)I n - XX 4 )
= (A + 2) m -"det((A + 2- jfe)In - A)

It follows from Theorem 3.8 that, if the spectrum of T is

k X\ ... As_
Spec r = .

then the spectrum of L(T) is

Spec LCT) = ,
1
1 mi ... ms-i m—n

For example, the line graph L(Kt) is sometimes called the triangle
graph and denoted by At. Its vertices correspond to the ^t(t — 1) pairs of
numbers from the set {1,2,..., t}, two vertices being adjacent whenever
the corresponding pairs have just one common member. From the known
spectrum of Kt, and Theorem 3.8, we have

2t-4 i-4 -2
20 Linear algebra in graph theory

Additional results

3a The complement of a regular graph Let F be a graph with n vertices

and let F c denote its complement, that is, the graph with the same
vertex-set whose edge-set is complementary to that of F. Let A c denote
the adjacency matrix of F c . Then A + A c = J — I. It was proved by
Sachs (1962) that if F is connected and regular of degree fc, then
(A + fc + 1)X(FC; A) - (-l) n (A - n + k + 1)X(F; -A - 1).
3b The Petersen graph The complement of the line graph of K5 is
known as the Petersen graph. It occurs in many contexts throughout
graph theory. We shall denote it by the symbol O3, as it is the case
k = 3 of the family {Ok} of odd graphs, to be defined later (8f). We
have
'3 1 - 2 s
Spec O3 = , 1 g
In particular, the least eigenvalue is —2, although O3 is neither a line
graph nor a hyperoctahedral graph.
3c Strongly regular graphs A fc-regular graph is said to be strongly
regular, with parameters (k, a, c), if the following conditions hold. Each
pair of adjacent vertices has the same number a > 0 of common neigh-
bours, and each pair of non-adjacent vertices has the same number c > 1
of common neighbours. It follows from Lemma 2.5 that the adjacency
matrix of such a graph satisfies
A 2 + (c - o)A + (c - fc)I = cJ.
In other words, the polynomial function / whose existence is guaranteed
by Proposition 3.2 is f{x) = (l/c)(a;2 + (c - a)x + (c - A;)).
3d The spectrum of a strongly regular graph Since the eigenvalues of
the n x n matrix J are n (with multiplicity 1) and 0 (with multiplicity
n— 1), it follows from 3c that the eigenvalues of a strongly regular graph
are k (with multiplicity 1) and the two roots Ai, A2 of the quadratic
equation /(A) = 0 (with total multiplicity n — 1). The multiplicities
mi = m(Ai) and m^ — m{^2) can be determined from the equations
m\ + mi = n — 1, k + miAi +TO2A2= 0,
the second of which follows from 2h. For example, the Petersen graph
(3b) is strongly regular with parameters (3,0,1), and this gives an al-
ternative method of determining its spectrum.
3e The Mobius ladders The Mobius ladder M^ is a regular graph of
degree 3 with 2h vertices (h > 3). It is constructed from the cycle graph
 Regular graphs and line graphs 21

Cih by adding new edges joining each pair of 'opposite' vertices, and so
it is a circulant graph. The eigenvalues are the numbers
Xj = 2 COS(TTj/h) + (-l)J (0<j<2h- 1).
3f Graphs characterized by their spectra Although there are many ex-
amples of cospectral graphs, there are also cases where there is a unique
graph with a given spectrum. We give two instances.
(o) The spectrum of the triangle graph A t = L(Kt) is given above. If
F is a graph for which SpecF = SpecA^, and t ^ 8, then F = A t . In
the case t = 8 there are three exceptional graphs, not isomorphic with
As, but having the same spectrum as As (Chang 1959, Hoffman 1960).
(b) The complete bipartite graph Ka<a is constructed by taking two
sets of a vertices and joining every vertex in the first set to every vertex
in the second. If F is a graph for which SpecF = Speci(X a i O ), and
a 7^ 4, then F = L{Ka^a). In the case a = 4 there is one exceptional
graph; this graph is depicted in Figure 2 (Shrikhande 1959).

13

Figure 2: Shrikhande's exceptional graph.

3g Regular graphs with least eigenvalue - 2 The following graphs hav-

ing least eigenvalue —2 were noted by Seidel (1968). They are neither
line graphs nor hyperoctahedral graphs:
(a) the Petersen graph;
(b) a 5-regular graph with 16 vertices;
(c) a 16-regular graph with 27 vertices (see p. 57);
(d) the exceptional graphs mentioned in 3f.
3h Generalized line graphs The cocktail party graph CP(s) is defined
on page 17. For any graph F with vertices vi,V2,--.,vn, and any non-
negative integers a\, a-i,..., an, we construct the generalized line graph
L(T; ai,a2,. • • ,an) as follows. The vertex-set is the union of the vertex-
sets of L(F), C P ( a i ) , CP(a2), • • •, CP(an), and the edge-set is the union
22 Linear algebra in graph theory

of the edge-sets, together with edges joining all vertices of CP(ai) to

every vertex of L(T) corresponding to an edge of F containing Vi, for
1 < i < n. A generalized line graph constructed in this way has least
eigenvalue —2.
3i All graphs with least eigenvalue —2 If F is a graph with least eigen-
value not less than —2, then A + 21 is non-negative definite, and so
A + 21 = MM* for some matrix M. By establishing a correspon-
dence between the rows of M and sets of vectors known as root systems,
Cameron, Goethals, Seidel, and Shult (1976) showed that all graphs
with least eigenvalue not less than —2 fall into three classes: (a) the line
graphs of bipartite graphs; (b) the generalized line graphs described in
3h; (c) a finite class of graphs arising from the root systems E^, Er, E$.
3j Perfect codes in regular graphs For any vertex v of a graph F define
the e-neighbourhood of v to be Ne(v) = {u € VT | d(u, v) < e}. A
perfect e-code in F is a set C C VT such that the e-neighbourhoods
Ne(c) with c € C form a partition of VT. Suppose that C is a perfect
1-code in a ^-regular graph F. Then the vector c which takes the value
1 on vertices in C and 0 on other vertices satisfies Ac = u — c. It follows
that u — (k + l)c is an eigenvector of F, with eigenvalue —1. Thus a
necessary condition for a regular graph to have a perfect 1-code is that
—1 is an eigenvalue. (See also 21j.)
3k Spectral bounds for the diameter Suppose that T is connected and
A;-regular, so that XQ = k and Ai < k. Alon and Milman (1985) proved
that the diameter d is bounded above by a function of n, k, and the
'gap' k — \\; specifically

d < 2 2k V log 2 n

Mohar (1991) improved this to

Using the spectral decomposition of Ar (see 2j), Chung (1989) obtained

a bound involving the second largest eigenvalue in absolute value A* =
max(Ai, -A n _i). She showed that if (k/A*)r > n — 1 then (A r ) f • > 0
for all i,j. It follows that
"ln(n-l)'
d <
 Cycles and cuts

Let C denote the field of complex numbers, and let X be anyfiniteset.

Then the set of all functions from X to C has the structure of a finite-
dimensional vector space; if / : X —> C and g : X —>• <C, then the vector
space operations are defined by the rules
(/ + 9)(x) = f{x) + g(x), (a/)(x) = af(x) (xGX,a€ C).
The dimension of this vector space is equal to the number of members
of X.
Definition 4.1 The vertex-space Cb(F) of a graph is the vector space
of all functions from VT to C The edge-space C\ (F) of F is the vector
space of all functions from ET to C
Taking VT = {vi,v2,... ,vn} and ET = {ei,e 2 ,... ,e m }, it follows
that Co(F) is a vector space of dimension n and Ci(F) is a vector space
of dimension m. Any function r\ : VT —• C can be represented by a
column vector
y = [yi,y2,---,yn\t,
where j/j = r/(vi) (1 < i < n). This representation corresponds to choos-
ing as a basis for Co(F) the set of functions {wi,o>2,-. -,w n }, defined
by
i \ _ / 1) if * = j ;
Ui{Vj)
~\0, otherwise.
In a similar way, we may choose the basis {ej,e2,... ,e m } f° r Ci(F)
24 Linear algebra in graph theory

defined by

0, otherwise;
and hence represent a function £ : ET -» C by a column vector x =
[xi,X2,---,xm]t such that x» = £(e») (1 < i < m). We shall refer to
the bases {u>i,u;2,... ,wn} and {ei, £2,..., e m } as the standard bases for
Co(r)andC,(r).
We now introduce a useful device. For each edge ea = {vrr,vT} of F,
we shall choose one of va,vT to be the positive end of e a , and the other
one to be the negative end. We refer to this procedure by saying that F
has been given an orientation. Although this device is employed in the
proofs of several results, the results themselves are independent of it.
Definition 4.2 The incidence matrix D of F, with respect to a given
orientation of F, is the n xm matrix (d^) whose entries are

{
+1, if Vi is the positive end of e^;
— 1, if Vi is the negative end of ej;
0, otherwise.
The rows of the incidence matrix correspond to the vertices of F, and
its columns correspond to the edges of F; each column contains just two
non-zero entries, +1 and — 1, representing the positive and negative ends
of the corresponding edge.
We remark that D is the representation, with respect to the standard
bases, of a linear mapping from Ci(F) to CQ(T). This mapping will be
called the incidence mapping, and be denoted by D. For each £ : ET —> C
the function Df; : VT —* C is defined by

For the rest of this chapter we shall let c denote the number of con-
nected components of F.
Proposition 4.3 The incidence matrix D o / F has rank n — c.
Proof The incidence matrix can be written in the partitioned form
D(1) 0 ... 0
2
0 D< > . . . 0

0 0 . . . D^c
by a suitable labelling of the vertices and edges of F, where the matrix
£>(*) (1 < i < c) is the incidence matrix of a component fM of F. We
shall show that the rank of D^-1 is n, — 1, where rij = |VT^|, from which
the required result follows by addition.
 Cycles and cuts 25

Let dj denote the row of D ^ corresponding to the vertex Vj of T^\

Since there is just one +1 and just one —1 in each column of D' 1 ', it
follows that the sum of the rows of D' 1 ' is the zero row vector, and that
the rank of D^' is at most rii — \. Suppose we have a linear relation
^Zctjdj = 0, where the summation is over all rows of D ^ , and not all
the coefficients otj are zero. Choose a row dk for which a^ ^ 0; this
row has non-zero entries in those columns corresponding to the edges
incident with Vk- For each such column, there is just one other row d ;
with a non-zero entry in that column, and in order that the given linear
relation should hold, we must have a; = a^. Thus, if a/t ^ 0, then
a; = ajt for all vertices vi adjacent to v^. Since F ^ is connected, it
follows that all coefficients OLJ are equal, and so the given linear relation
is just a multiple of J^ dj = 0. Consequently, the rank of D ^ is rii — 1.
•
The following definition applies to a general graph F with n vertices, m
edges, and c components, although for the time being we shall continue
to deal with strict graphs, rather than general graphs.
Definition 4.4 The rank of F and the co-rank of F are, respectively,
r(F) = n — c; s(F) = m — n + c.
We now investigate the kernel of the incidence mapping D, and its
relationship with graph-theoretical properties of F. Let Q be a set of
edges such that the subgraph {Q) is a cycle graph. We say that Q is a
cycle in F; the two possible cyclic orderings of the vertices of (Q) induce
two possible cycle-orientations of the edges Q. Let us choose one of these
cycle-orientations, and define a function £Q in C\ (F) as follows. We put
£q(e) = +1 if e belongs to Q and its cycle-orientation coincides with its
orientation in F, ^cj(e) = — 1 if e belongs to Q and its cycle-orientation is
the reverse of its orientation in F, while if e is not in Q we put ^Q(e) = 0.
Theorem 4.5 The kernel of the incidence mapping D ofT is a vector
space whose dimension is equal to the co-rank o/F. If Q is a cycle in F,
then £Q belongs to the kernel of D.
Proof Since the rank of D is n — c, and the dimension of C\ (F) is m,
it follows that the kernel of D has dimension m — n + c = s(F). With
respect to the standard bases for Ci(F) and Co(F), we may take D to be
the incidence matrix, and £Q to be represented by a column vector XQ.
Now (DXQ)J is the inner product of the row d; of D and the vector XQ.
If vt is not incident with some edges of Q, then this inner product is 0;
if Vi is incident with some edges of Q, then it is incident with precisely
26 Linear algebra in graph theory

two edges, and the choice of signs in the definition of £Q implies that the
inner product is again 0. Thus Dxg = 0, and £Q belongs to the kernel
of D. U

If p and a are two elements of the edge-space of F (that is, functions

from ET to <C), then we may define their inner product

where the over line indicates the complex conjugate. When p and a are
represented by coordinate vectors, with respect to the standard basis
of Ci(F), this inner product corresponds to the usual inner product
of vectors in the complex vector space C m . (In practice we use only
functions with real values, so the conjugation is irrelevant.)
Definition 4.6 The cycle-subspace of F is the kernel of the incidence
mapping of F. The cut-subspace of F is the orthogonal complement of
the cycle-subspace in Ci(F), with respect to the inner product defined
above.
The first part of this definition is justified by the result of Theorem 4.5,
which says that vectors representing cycles belong to the cycle-subspace:
indeed, in the next chapter we shall show how to construct a basis for the
cycle-subspace consisting entirely of cycles. We now proceed to justify
the second part of the definition.
Let VT = V\ U V2 be a partition of VT into non-empty disjoint subsets.
If the set H of edges of F which have one vertex in Vj and one vertex in
V2 is non-empty, then we say that if is a cut in F. We may choose one of
the two possible cut-orientations for H, by specifying that one of Vi, V2
contains the positive ends of all edges in H, while the other contains
the negative ends. We now define a function £H in Ci(F) by putting
£tf (e) = +1 if e belongs to H and its cut-orientation coincides with its
orientation in F, £# (e) = — 1 if e belongs to H and its cut-orientation is
the reverse of its orientation in F, and £#(e) = 0 if e is not in H.
Proposition 4.7 The cut-subspace of F is a vector space whose di-
mension is equal to the rank ofT. If H is a cut in F, then £# belongs
to the cut-subspace.
Proof Since the dimension of the cycle-subspace is m — n 4- c, its
orthogonal complement, the cut-subspace, has dimension n — c = r(F).
If if is a cut in F, we have VT — Vi U V2, where V\ and V2 are disjoint
and non-empty, and H consists precisely of those edges which have one
vertex in V\ and one vertex in V2- Thus, if x # is the column vector
 Cycles and cuts 27

representing £#, we have

where dj is the row of the incidence matrix corresponding to Vi. The

sign on the right-hand side of this equation depends only on which of
the two possible cut-orientations has been chosen for H. Now if Dz = 0,
then djZ = 0 for each v^ £ V, and we deduce that x ^z = 0. In other
words, £H belongs to the orthogonal complement of the cycle-subspace,
and, by definition, this is the cut-subspace. D

The proof of Proposition 4.7 indicates one way of choosing a basis

{£ii£2j • • • ,£n-c} for the cut-subspace of F. The set of edges incident
with a vertex Vj of F forms a cut whose representative vector is d*-.
If, for each component F ^ (1 < i < c) of F, we delete one row of D
corresponding to a vertex in T^\ then the remaining n - c rows are
linearly independent. Furthermore, the transpose of any vector xH,
representing a cut H, can be expressed as a linear combination of these
n — c rows, by using the equation displayed in the proof of Proposition
4.7 and the fact that the sum of rows corresponding to each component
isO.
This basis has the desirable property that each member represents an
actual cut, rather than a 'linear combination' of cuts. It is, however,
rather clumsy to work with, and in the next chapter we shall investigate
a more elegant procedure which has the added advantage that it provides
a basis for the cycle-subspace as well.
We end this chapter by proving a simple relationship between the
Laplacian matrix Q = DD* and the adjacency matrix of F.
Proposition 4.8 Let D be the incidence matrix (with respect to some
orientation) of a graph F, and let A be the adjacency matrix ofT. Then
the Laplacian matrix Q satisifies
Q = DD* = A - A,
where A is the diagonal matrix whose ith diagonal entry is the degree
of the vertex V{ (1 < i < n). Consequently, Q is independent of the
orientation given to F.
Proof (DD*)jj is the inner product of the rows d, and dj of D. If
i 7^ j , then these rows have a non-zero entry in the same column if and
only if there is an edge joining u, and Vj. In this case, the two non-zero
entries are +1 and - 1 , so that (DD%- = - 1 . Similarly, (DD*^ is the
inner product of dj with itself, and, since the number of entries ±1 in
di is equal to the degree of vt, the result follows. D
28 Linear algebra in graph theory

Additional Results

4a The coboundary mapping The linear mapping from Co(F) to Ci(F)

defined (with respect to the standard bases) by x >-» D*x is sometimes
called the coboundary mapping for F. The kernel of the coboundary map-
ping is a vector space of dimension c, and the image of the coboundary
mapping is the cut-subspace of F.

4b The isoperimetric number For any set X C VT the cut defined by

the partition of VT into X and its complement is denoted by 6X. The
isoperimetric number of F is defined to be
i(T) = min T^T-
|X|<|vr|/2 \X\
For example, it is easy to check that i(Kn) = [n-/2], i(03) = 1.

4c Small cycles The girth of a graph is the number g of edges in a

smallest cycle. For example, g(Kn) = 3 (n > 3), g(Ka,a) — 4 (a > 2),
and 5(03) = 5. If F has girth g > 2r + 1 then for each pair of vertices
v and w such that d(v, w) = q < r there is a unique walk of length q
from v to w. In the A;-regular case this leads to the following relations
between the adjacency matrix and the distance matrices A, (2 < q < r)
defined in 2k:
A 2 = A 2 - fcl, Aq = AA,_, - (fc - 1)A,_2 (3 < q < r).
It follows that a distance matrix Aq with q < r is expressible as a
polynomial in A. Explicitly, Aq = fq(A), where
fo(x) = 1, fx(x)=x, f2(x) = x2-k,
fq(x) - xfq-x{x) ~(k- l)/,-2(i) (q > 3).

4d Girift and excess It is an elementary exercise (see Chapter 23)

to show that the number of vertices in a fc-regular graph with girth
g — 2r + 1 is at least
no(k,g) = l + k + k(k-l)+k(k-l)2 + ... + k(k- l)r.
The Petersen graph O3 achieves the lower bound no for the case k = 3
and g = 5, but in the general case graphs which achieve the lower bound
are rare (Chapter 23). For anyfc-regulargraph F with girth g we define
the excess to be the amount e by which the lower bound is exceeded,
that is, e = n — no(k, g), where n is the number of vertices in F. Using
the equations given in 4c, Biggs (1980) established a lower bound for e
in terms of the eigenvalues of F. Define the polynomials gj by
9i{x) = /0(a:) + fi{x) + ... + fi(x),
 Cycles and cuts 29

where the polynomials / , are defined above. Then for any fc-regular
graph with girth g = 2r + 1 the excess e satisfies
e>|<7r(A)|, (A e Spec F, X ^ k).

4e The Laplacian spectrum Let /io < /^i < • • • < M«-i D e the eigen-
values of the Laplacian matrix Q. Then:
(a) (J,Q = 0, with eigenvector [ 1 , 1 , . . . , 1];
(b) if F is connected, /J,\ > 0;
(c) if F is regular of degree k, then /u, = k - A*, where the Aj are
the (ordinary) eigenvalues of F, in weakly decreasing order.

4f Planar graphs and duality A planar graph is one which can be

drawn in the plane in the usual way, without extraneous crossings of the
edges. The dual of a graph so drawn is the graph whose vertices are
the resulting regions of the plane, two being adjacent when they have a
common edge. Let F be a connected planar graph, and F* a dual of F.
If F is given an orientation and D is the incidence matrix of F, then F*
can be given an orientation so that its incidence matrix D* satisfies:
(a) rank (D) + rank (D*) = |£T|;
(b) D*D* = 0.

4g The image of the incidence mapping Let w be an element of C 0 (F),

where F is a connected graph. Then u> is in the image of D if and only
if

vevr
A more sophisticated way of expressing this result is as follows. Let
S : Co(F) —> C denote the linear map defined by S^w) = ^Zw(u); this is
known as the augmentation map. Then the sequence of linear maps
Ci(r) - ^ co(F) -£> c —> o
is exact. In particular, this means that the image of D is equal to the
kernel of 5.

4 h Flows An element <j) of the cycle-subspace of F is called a flow on F.

The support of <f), written S(4>), is the set of edges e for which ^(e) ^ 0;
a subset S of ET is a minimal support if S = S(<f>) for some flow 0, and
the only flow whose support is properly contained in S is the zero flow.
We have the following basic facts.
(a) The set of flows with a given minimal support (together with
the zero flow) forms a one-dimensional space.
30 Linear algebra in graph theory

(b) A minimal support is a cycle.

(c) If <fi is a flow whose support is minimal, then |(/>(e)| is constant
on S((f>).

4i Integral flows The flow 4> is integral if each <fi(e) is an integer; it is

primitive if S(<p) is minimal and each (j>(e) is 0,1 or —1. We say that the
flow 0 conforms to the flow \ if S(6) C S(x) and 8(e)x(e) > 0 for e in
S(6). Tutte (1956) showed that:
(a) for a given integral flow <fr there is a primitive flow which con-
forms to <f>;
(b) any integral flow <j> is the sum of integer multiples of primitive
flows, each of which conforms to <f>.

4j Modular flows Suppose the entries 0,1, —1 of D are taken to be

elements of the ring TLU = Z / u Z of residue classes of integers modulo u.
A flow mod u on F is a vector x with components in Z u for which D x
= 0, where 0 is the zero vector over Z u . The results in 4i imply that, if
x is a given flow mod u, then there is an integral flow y, each of whose
components j/j satisfies y, £ Xi and — u < yi < u. Consequently, if F has
a flow mod u, then it has a flow mod (u + 1) (Tutte 1956).

4k The 5-flow conjecture A nowhere-zero f-flow <j> on T is a flow mod

/ for which S(<f>) = ET. Tutte (1954) conjectured that every graph with
no isthmus has a nowhere-zero 5-flow. (An isthmus is a cut consisting
of a single edge.) The following results are known.
(a) Every planar graph with no isthmus has a nowhere-zero 4-flow.
(b) The Petersen graph does not have a nowhere-zero 4-flow.
(c) Every graph with no isthmus has a nowhere-zero 6-flow (Sey-
mour 1981).
Spanning trees and associated structures

The problem of finding bases for the cycle-subspace and the cut-subspace
is of great practical and theoretical importance. It was originally solved
by Kirchhoff (1847) in his studies of electrical networks, and we shall
give a brief exposition of that topic at the end of the chapter.
We shall restrict our attention to connected graphs, because the cycle-
subspace and the cut-subspace of a disconnected graph are the direct
sums of the corresponding spaces for the components. Throughout this
chapter, F will denote a connected graph with n vertices and m edges,
so that r(F) = n — 1 and s(F) = m — n + 1. We shall also assume that
F has been given an orientation.
A spanning tree in F is a subgraph which has n — 1 edges and contains
no cycles. It follows that a spanning tree is connected. We shall use the
symbol T to denote both the spanning tree itself and its edge-set. The
following simple lemma is a direct consequence of the definition.
Lemma 5.1 Let T be a spanning tree in a connected graph F. Then:
(1) for each edge gofT which is not in T there is a unique cycle in F
containing g and edges in T only.
(2) for each edge h of T which is in T, there is a unique cut in F
containing h and edges not in T only. D

We write cyc(T, g) and cut(T, h) for the unique cycle and cut whose
existence is guaranteed by Lemma 5.1. We give cyc(T,g) and cut(T, h)
the cycle-orientation and cut-orientation which coincide, on g and h
respectively, with the orientation in F. Then we have elements
32 Linear algebra in graph theory

and £{T,h) of the edge-space Ci(F); these elements axe defined (in terms
of the given cycle and cut) as in Chapter 4.
Theorem 5.2 With the same hypothesis as in Lemma 5.1, we have:
(1) as g runs through the set £T — T, the m — n + 1 elements £(T, 9 )
form a basis for the cycle-sub space of F.
(2) as h runs through the set T, the n — 1 elements €(T,h) form a basis
for the cut-subspace of F.
Proof (1) Since the elements £(T,g) correspond to cycles, it follows
from Theorem 4.5 that they belong to the cycle-subspace. They form
a linearly independent set, because a given edge g in EF — T belongs
to cyc(T, g) but to no other cyc(T, g') for g' ^ g. Finally, since there
are m - n + 1 of these elements, and this is the dimension of the cycle-
subspace, it follows that we have a basis.
(2) This is proved by arguments analogous to those used in the proof
of the first part. •

We shall now put the foregoing ideas into a form which will show
explicitly how cycles and cuts can be derived from the incidence matrix,
by means of simple matrix operations. To do this, we shall require some
properties of submatrices of the incidence matrix.
Proposition 5.3 (Poincare 1901) Any square submatrix of the inci-
dence matrix D of a graph F has determinant equal to 0 or +1 or — 1.
Proof Let S denote a square submatrix of D. If every column of S
has two non-zero entries, then these entries must be +1 and —1 and so,
since each column has sum zero, S is singular and det S = 0. Also, if
every column of S has no non-zero entries, then det S = 0.
The remaining case occurs when a column of S has precisely one non-
zero entry. In this case we can expand det S in terms of this column,
obtaining det S = ± det S', where S' has one row and column fewer
than S. Continuing this process, we eventually arrive at either a zero
determinant or a single entry of D, and so the result is proved. •

Proposition 5.4 Let U be a subset of ET with \U\ = n - 1. Let T>u

denote an (n— 1) x (n — 1) submatrix ofT>, consisting of the intersection
of those n — 1 columns of D corresponding to the edges in U and any set
of n — 1 rows of D. Then Du is invertible if and only if the subgraph
{U) is a spanning tree ofT.
Proof Suppose that (U) is a spanning tree of F. Then the submatrix
Dy consists of n - 1 rows of the incidence matrix D' of U. Since (U) is
connected, the rank of D' is n — 1, and so T>u is invertible.
 Spanning trees and associated structures 33

Conversely, suppose that T>u is invertible. Then the incidence matrix

D' of (U) has an invertible (n — 1) x (n-1) submatrix, and consequently
the rank of D' is (n — 1). Since \U\ = n — 1, this means that the cycle-
subspace of (U) has dimension zero, and so (U) is a spanning tree of
r. •
Suppose that VT = {v\, v<i, • • •, vn} and ET = {e\, 62, • • •, em}, where
the labelling has been chosen so that e\, e<i,..., e n _i are the edges of a
given spanning tree T of T. The incidence matrix of F is then partitioned
as follows:

where D T is an (n — 1) x (n — 1) square matrix, invertible by Proposition

5.4, and the last row d n is linearly dependent on the other rows.
Let C denote the matrix whose columns are the vectors representing
the elements £(T,ej) (^ < j < m) with respect to the standard basis of
Ci(F). Then C can be written in the partitioned form
CT 1
c=,
•lm-n+1
Since every column of C represents a cycle, and consequently belongs to
the kernel of D, we have DC = 0. Thus
T"\ — 1 T^
T1 — —JLJrp • * ?y -

In a similar fashion, the matrix K whose columns represent the elements

£(T,ej) (1 < 3' < n ~ 1) c a n De written in the form
X-i
r

Since each column of K belongs to the orthogonal complement of the

cycle-subspace, we have CK* = 0; that is, Cx + K T = 0. Thus
=
J\.X (Uy UN) •

Our equations for Cx and Kx show how the basic cycles and cuts asso-
ciated with T can be deduced from the incidence matrix. We also have
an algebraic proof of the following proposition.
Proposition 5.5 Let T be a spanning tree of T and let a and b be
edges of T such that a€T,b&T. Then
b 6 cut(r, a)<*ae cyc(T, b).
Proof This result follows immediately from the definitions of Cx and
K T , and the fact that C T + K^ = 0. •
34 Linear algebra in graph theory

We end this chapter with a brief exposition of the solution of network

equations; this application provided the stimulus for Kirchhoff's devel-
opment of the foregoing theory in the middle of the nineteenth century.
An electrical network is a connected graph F (with an arbitrary ori-
entation) which has certain physical characteristics, specified by two
vectors in the edge-space of F. These vectors are the current vector w
and the voltage vector z. These vectors are related by a linear equa-
tion z = Mw + n, where M is a diagonal matrix whose entries are the
conductances of the edges, and n represents externally applied voltages.
Further, w and z satisfy the equations
Dw = 0, C'z = 0,
which are known as Kirchhoff's laws. If we choose a spanning tree T in
F, and partition D and C as before, then the same partition on w and
z gives
[ wj-1
w= , z=
Now, from Dw = 0 we have D T W ^ + DJVWJV = 0, and since Cx =
—D^D/y it follows that
wj- = CXWJV and w = CWJV-
In other words, all the entries of the current vector are determined by the
entries corresponding to edges not in T. Substituting in z = Mw + n,
and premultiplying by C*, we obtain
(C t MC)w JV = - C ' n .
Since C*MC is a square matrix with size and rank both equal to m—n+1
it is invertible.
So this equation determines WN, and consequently both w (from
w = CWAT) and z (from z = Mw + n) in turn. Thus we have a system-
atic method of solving network equations, which distinguishes clearly
between the essential unknowns and the redundant ones.

Additional Results

5a Total unimodularity A matrix is said to be totally unimodular if ev-

ery square submatrix of it has determinant 0, 1, or —1; thus Proposition
5.3 states that D is totally unimodular. A generalisation of this result
was proved by Heller and Tompkins (1956). They showed that if M is
a matrix with elements 0, 1, or —1 such that every column contains at
most two non-zero elements, then M is totally unimodular if and only
if its rows can be partitioned into two disjoint parts satisfying:
 Spanning trees and associated structures 35

(i) if a column has two non-zero elements with the same sign, then
their rows are in different parts;
(ii) if a column has two non-zero elements with opposite signs, then
their rows are in the same part.

5b Integral solutions of LP problems Hoffman and Kruskal (1956)

proved the following result. If M is a totally unimodular matrix and
b is an integral vector then, for each objective function c, the linear
programming problem (LP)
maximise c'x subject to Mx < b
has an optimal solution which is integral, provided that there is a finite
solution.
Several optimization problems on graphs have LP formulations in
which M is the incidence matrix, or a modified form of it. Among
them are the maximum flow problem and the shortest path problem,
the details of which are given in the standard text of Grotschel, Lovasz
and Schrijver (1988). Hoffman and Kruskal's theorem leads to 'integral-
ity' results, such as the fact that if the capacities are integral then there
is a maximum flow which is also integral.

5c The unoriented incidence matrix As in Chapter 3, let X denote

the matrix obtained from the incidence matrix D of F by replacing each
entry ±1 by +1. It follows from the result of Heller and Tompkins
quoted in 5a that F is bipartite if and only if X is totally unimodular.
This was first observed by Egervary (1931).

5d The image of D again With the notation of 4g, if a; is integer-valued

and S(w) = 0 then there is an integer-valued £ such that £>(£) = w.

5e The inverse of Dx Let T be a spanning tree for F and let T>T

denote the corresponding (n — 1) x (n — 1) matrix. Then (D^ 1 )^ = ±1
if the edge ej occurs in the unique path in T joining Vj to vn. Otherwise
( D r % = 0.

5f The Laplacian formulation of network equations For simplicity, con-

sider the case of a network in. which each edge has conductance 1. Then
the network equations are
z = w + n, Dw = 0, C*z = 0.
The last equation says that z is orthogonal to the cycle-subspace and
so, by Definition 4.6, it belongs to the cut-subspace. It follows from 4a
36 Linear algebra in graph theory

that z = D*</> for some potential <f> in the vertex-space. Using the other
two equations we obtain
D D V = Dn; that is, Qcj> = n,
where Q is the Laplacian matrix and 77 is a vector in which t)v is the
current flowing into the network at the vertex v. In particular, defining

{
+1, iiv = x;
-1, Hv = y;
0, otherwise;
we see that the solution of the network equations when a current /
enters at x and leaves at y is given by finding the potential satisfying
Q0 = Irfv.

5g Existence and uniqueness of the solution; Thomson's principle Sim-

ple proofs of the results in the following paragraphs may be be found in
a paper by Thomassen (1990). If x and y are vertices of a finite graph
then there is a unique solution <j> to the network equations for the case
when a positive real-valued current I enters at x and leaves at y. The
current vector z = D'</> is the vector satisfying Dz = Irfy for which the
power ||z|| 2 is a minimum. (This is known as Thomson's principle.)

5h An explicit solution for the network equations Suppose that x and

y are adjacent vertices of a connected graph F, and let K denote the
total number of spanning trees of F. (See Chapter 6 for more about K.)
For each spanning tree T of F send a current I/K along the unique path
in T from x to y. Then the current vector z which solves the network
equations for a current I entering at x and leaving at y is the sum of these
currents taken over all T. This result goes back to Kirchhoff (1847). For
historical details and an algebraic proof, see Nerode and Shank (1961).

5i The effective resistance For any two vertices x and y let <f> be the
potential satisfying Q</> = Irfv. Following Ohm's law, the effective
resistance from x to y is defined to be (4>x — 4>y)/I- If x and y are
adjacent vertices this is equal to KXV/K, where Kxy is the number of
spanning trees which contain the edge {x,y}.
For example, it can be shown (see p. 39) that the number of spanning
trees of the complete graph Kn is n n ~ 2 ; since each one contains n — 1
of the n(n — l)/2 edges, there are 2n"~ 3 spanning trees containing a
given edge. It follows that the effective resistance across an edge of Kn
 Spanning trees and associated structures 37

is 2/n. In general, if a graph has n vertices and m edges, and it is edge-

transitive (see Chapter 15), then the effective resistance across an edge
is (n - l)/m.

5j Monotonicity results Let R(x, y, T) denote the effective resistance

of F from x to y. If T' is obtained from F by removing an edge (the
cutting operation), then
R(x,y,V) > R(x,y,T).
The inequality is reversed if I" is obtained from T by identifying two
vertices (the shorting operation). These results are known as Rayleigh's
monotonicity law.
 6

The tree-number

Several famous results in algebraic graph theory, including one of the

oldest, are formulae for the numbers of spanning trees of certain graphs.
Many formulae of this kind were given in the monograph written by
Moon (1970). We shall show how such results can be derived from the
Laplacian matrix Q introduced in Chapter 4.
Definition 6.1 The number of spanning trees of a graph F is its
tree-number, denoted by n(T).
Of course, if F is disconnected, then K(T) = 0. For the connected
case, Theorem 6.3 below is a version of a formula for K(F) which has
been discovered many times. We need a preparatory lemma concerning
the matrix of cofactors (adjugate) of Q.
Lemma 6.2 Let D be the incidence matrix of a graph T, and let
Q = DD* be the Laplacian matrix. Then the adjugate of Q is a multiple
ofJ.
Proof Let n be the number of vertices of F. If F is disconnected, then
rank (Q) = rank (D) < n - 1,
and so every cofactor of Q is zero. That is, adj Q = 0 = 0J.
If F is connected, then the ranks of D and Q are n — 1. Since
Q adj Q = (det Q)I = 0,
it follows that each column of adj Q belongs to the kernel of Q. But this
kernel is a one-dimensional space, spanned by u = [1,1,..., 1]*. Thus,
 The tree-number 39

each column of adj Q is a multiple of u. Since Q is symmetric, so is adj

Q, and all the multipliers must be equal. Hence, adj Q is a multiple of
J. •
Theorem 6.3 Every cofactor of Q is equal to the tree-number of F,
that is,
adj Q = K(F)J.
Proof By Lemma 6.2, it is sufficient to show that one cofactor of Q is
equal to «(F). Let DQ denote the matrix obtained from D by removing
the last row; then det D 0 D Q is a cofactor of Q. This determinant can
be expanded by the Binet-Cauchy theorem (see Theory of Matrices by
P. Lancaster (Academic Press) 1969, p. 38). The expansion is
det(D[/)det(D*£/),

where Df/ denotes the square submatrix of Do whose n — 1 columns

correspond to the edges in a subset U of £T. Now, by Proposition 5.4,
det D[/ is non-zero if and only if the subgraph (U) is a spanning tree for
F, and then detDy takes the values ±1. Since detDfy = detDy, we
have det(DoDo) = K(F), and the result follows. •

For the complete graph Kn we have Q = n l - J . A simple determinant

manipulation on nl — J with one row and column removed shows that
K(Kn) = nn~2. This result was first obtained, for small values of n, by
Cayley (1889).
We can dispense with the rather arbitrary procedure of removing one
row and column from Q, by means of the following result.
Proposition 6.4 (Temperley 1964) The tree-number of a graph F
with n vertices is given by the formula
/6(D = n- 2 det (J + Q).
Proof Since nJ = J 2 and JQ = Owe have the following equation:
(nl - J)(J + Q) = nJ + nQ - J 2 - J Q = nQ.
Thus, taking adjugates and using Theorem 6.3, we can argue as follows,
where K = K(F):
adj (J + Q)adj (nl - J) = adj nQ,
adj (J + Q)n n ~ 2 J - n n - J adj Q,
adj (J + Q)J = n«J,
(J + Q) adj (J + Q)J = (J + Q)n«J,
40 Linear algebra in graph theory

det (J + Q)J = n2Ki.

It follows that det(J + Q) = n2K, as required. D

The next result uses the Laplacian spectrum introduced in 4e.

Corollary 6.5 Let 0 < /xi < . . . < /x n -i be the Laplacian spectrum of
a graph F with n vertices. Then

IfT is connected and k-regular, and its spectrum is

\ mi ... ms-i
then

«(T) = n-'fiik - \r)mr = n-V(r; k),

r=l
where x' denotes the derivative of the characteristic polynomial \-
Proof Since Q and J commute, the eigenvalues of J + Q are the
sums of corresponding eigenvalues of J and Q. The eigenvalues of J are
n, 0 , 0 , . . . , 0, so the eigenvalues of J+Q are n, pi, fj,%,..., nn-i- Since the
determinant is the product of the eigenvalues, the first formula follows.
In the case of a regular graph of degree k, an (ordinary) eigenvalue
A is k — fi, where /x is a Laplacian eigenvalue. The result follows by
collecting the eigenvalues according to their multiplicities, and recalling
that k — A is a simple factor of \ m the connected case. •

Later in this book, when we have developed techniques for calculating

the spectra of highly regular graphs, we shall be able to use this Corollary
to write down the tree-numbers of many well-known families of graphs.
For the moment, we shall consider applications of Corollary 6.5 in some
simple, but important, cases. If F is a regular graph of degree k, then
the characteristic polynomial of its line graph L(T) is known in terms
of that of F (Theorem 3.8). If F has n vertices and m edges, so that
2m = nk, then we have

Differentiating the result of Theorem 3.8 and putting A = 2k — 2, we get

x'(L(r);2k-2) = (2k)m-nx'(T;k).
Hence we obtain the tree-number of F in terms of that of L(T):
 The tree-number 41

For example, the tree-number of the triangle graph At = L(Kt) is

The complete multipartite graph KaXta2,...,aa has a vertex-set which is

partitioned into s parts A\, A2,..., As, where \At\ = Oj (1 < i < s); two
vertices axe joined by an edge if and only if they belong to different parts.
In general this graph is not regular, but its complement (as defined in
3a) consists of regular connected components. The tree-number of such
graphs can be found by a modification of Proposition 6.4, due to Moon
(1967). This is based on the properties of the characteristic function of
the Laplacian matrix:
<r(r;/i)=det(/iI-Q).
Proposition 6.6 (1) IfT is disconnected, then the a function for F
is the product of the a functions for the components ofT.
(2) IfT is a k-regular graph, then a(T; //) = (—l) n x(F; k - fi), where
X is the characteristic polynomial of the adjacency matrix.
(3) IfTc is the complement ofT, and F has n vertices, then
K(T) =n-2a{Tc\n).
Proof (1) This follows directly from the definition of a.
(2) In the fc-regular case, we have
det(/il - Q) = det(/<I - (fcl - A)) = (-1)" det((fc - n)I - A),
whence the result.
(3) Let Q c denote the Laplacian matrix for F c , so that Q+Q c = nl—J.
Then using Proposition 6.4, we have
K (D = n" 2 det(J + Q) = n~ 2 det(nl - Q c ) = TC2<T(YC; n).
•
Consider the complete multipartite graph /fai,a2,.••><*»> where a\ +
a% + • • • + as = n, the complement of which consists of s compo-
=
nents isomorphic with Kai, Ka2,..., Kas. We know that x(Kn\ty
(A -I- l ) n - 1 ( A — n + 1), and, using part (2) of Proposition 6.6, we obtain
a{Ka- n) = (-l)ax(Ka; a - 1 - fj) = ^ - a) 0 " 1 .
Consequently, applying parts (1) and (3) of Proposition 6.6,
K(Kaua2,...,a.) = n-2.(n)(n - a^1'1... (n)(n - CL,) 0 "- 1
2 ai 1
= n ~ (n - ai) - ... (n - a.
This result was originally found (by different means) by Austin (1960).
We note the special cases:
K{Ka,b) = ab-lba-\ K(HS) = 2 2 s - V - 1 ( s - l) s .
42 Linear algebra in graph theory

Additional Results

6a A bound for the tree-number of a regular graph If F is a connected

fc-regular graph with n vertices, then applying the arithmetic-geometric
mean inequality to the product formula in Corollary 6.5 we obtain
-t / I N n—1

< 1- I nk
n \n — 1
with equality if and only if F = Kn.

6b More bounds for the tree-number Grimmett (1976) showed that

the bound in 6a can be extended to non-regular graphs. The result for
any graph with m edges is
n-l

n \n — 1 /
This is clearly a generalisation of result 6a, since 2m = nk in the fc-
regular case. Grone and Merris (1988) showed that if TT(F) is the product
of the vertex-degrees then

with equality if and only if F = Kn.

6c A recursion for the tree-number For any (general) graph F, and any
edge e which is not a loop, we define the graph F ^ to be the subgraph
obtained by removing e, and F(e) to be the graph obtained from F^e) by
identifying the vertices of e. Note that even if F itself is a graph (rather
than a general graph) this process may produce a general graph. We
have

6d Tree-number of a Mb'bius ladder The tree-number of the Mobius

ladder Mh denned in 3e may be computed in two ways. Using the
spectral formula 6.5 we obtain
2h 1
1 ~

An alternative is to use 6c to obtain a recursion formula. Sedlacek

(1970) used this method to obtain
n{Mh) = ~[(2 + v/3)" + (2 - v/3)h] + h.
The recursive method was discussed in greater generality by Biggs,
Damerell and Sands (1972); see 9i.
 The tree-number 43

6e Almost-complete graphs Let F be a graph constructed by removing

q disjoint edges from Kn where n > 2q. Then

In particular, taking n = 2q, we have the formula for the tree-number

of if,.

6f Tree-numbers of planar duals Let F and F* be dual planar graphs

(as defined in 4f) and let D and D* be the corresponding incidence
matrices. Suppose that F has n vertices, F* has n* vertices, and \ET\ =
\ET*\ = m; then (n - 1) + (n* - 1) = m. If Dy is a square submatrix of
D, whose n — 1 columns correspond to the edges of a subset U of ET,
and U* denotes the complementary subset of ET* = £T, then D[/ is
non-singular if and only if D^ is non-singular. Consequently,

6g The octahedron and the cube The octahedron graph is H3 = -^2,2,2;

it is planar and the cube graph Qs is its dual. We have
0 -2\ o _. /3 1 -1
3 2 ] ' SP««»=(i 3 3
Hence K(H3) = K(QS) = 384, in agreement with 6e.

6h The a function of the complement Prom the equation Q + Q c =

nl — J, we obtain
fil - Q c = [(n - M J - ' J - l][(n - M)I - Q].
Taking determinants we have
(n - M)<r(Fc;M) = (-1)"

6i Spectral characterization of complete multipartite graphs The com-

plete multipartite graphs defined on page 41 are the only connected
graphs for which the second largest eigenvalue Ai is not positive (Smith
1970).
 Determinant expansions

In this chapter we shall investigate the characteristic polynomial x> and

the polynomial a introduced in Chapter 6, by means of determinant ex-
pansions. We begin by considering the determinant of the adjacency ma-
trix A of a graph F. We suppose, as before, that VT = {v\, v%,..., vn}
and that the rows and columns of A are labelled to conform with this
notation. The expansion which is useful here is the usual definition of a
determinant: if A = (a^), then
det A = ] P sgn(7r)ai!7rla2,7r2 • • • an,™,
where the summation is over all permutations TT of the set { 1 , 2 , . . . , n}.
In order to express the quantities which appear in the above expansion
in graph-theoretical terms, it is helpful to introduce a new definition.
Definition 7.1 An elementary graph is a simple graph, each compo-
nent of which is regular and has degree 1 or 2. In other words, each
component is a single edge (K2) or a cycle (C r ). A spanning elementary
subgraph of F is an elementary subgraph which contains all vertices of
r.
We observe that the co-rank of an elementary graph is just the number
of its components which are cycles.
Proposition 7.2 (Harary 1962) Let A be the adjacency matrix of a
graph F. Then
 Determinant expansions 45

where the summation is over all spanning elementary subgraphs A o/F.

Proof Consider a term sgn(7r)aii7ria2,T2 • • • o«,7rn in the expansion of
det A. This term vanishes if, for some i € {1,2,..., n}, a*,™ = 0; that
is, if {vi,vni} is not an edge of F. In particular, the term vanishes if n
fixes any symbol. Thus, if the term corresponding to a permutation n is
non-zero, then 7r can be expressed uniquely as the composition of disjoint
cycles of length at least two. Each cycle (ij) of length two corresponds
to the factors a^aji, and signifies a single edge {vi,Vj} in F. Each
cycle (pqr... t) of length greater than two corresponds to the factors
apqaqr ... atp, and signifies a cycle {vp, vq,...,vt} in F. Consequently,
each non-vanishing term in the determinant expansion gives rise to an
elementary subgraph A of F, with VA = VF.
The sign of a permutation n is (—l)N', where Ne is the number of
even cycles in IT. If there are cj cycles of length I, then the equation
E/c; = n shows that the number No of odd cycles is congruent to n
modulo 2. Hence,
r(A) =n~(No + Ne) == Ne (mod 2),
so the sign of n is equal to (—l)r(A).
Each elementary subgraph A with n vertices gives rise to several per-
mutations n for which the corresponding term in the determinant ex-
pansion does not vanish. The number of such TT arising from a given A is
2S^A', since for each cycle-component in A there are two ways of choosing
the corresponding cycle in 7r. Thus each A contributes (—l) r ( A )2 3 ^ to
the determinant, and we have the result. •

For example, in the complete graph K± there are just two kinds of ele-
mentary subgraph with four vertices: pairs of disjoint edges (for which
r — 2 and s = 0) and 4-cycles (for which r = 3 and s = 1. There are
three subgraphs of each kind so we have
det A(K4) = 3(-l) 2 2° + 3(-l) 3 2 1 = - 3 .
At the beginning of this book we obtained a description of the first
few coefficients of the characteristic polynomial of F, in terms of some
small subgraphs of F (Proposition 2.3). We shall now extend that result
to all the coefficients. We shall suppose, as before, that
X (F; A) - Xn + cxA""1 + c2A"-2 + ... + c n .

Proposition 7.3 The coefficients of the characteristic polynomial are

given by
46 Linear algebra in graph theory

where the summation is over all elementary subgraphs A of F with i

vertices.
Proof The number {-\)lCi is the sum of all principal minors of A
with i rows and columns. Each such minor is the determinant of the
adjacency matrix of an induced subgraph of F with i vertices. Any
elementary subgraph with i vertices is contained in precisely one of these
induced subgraphs, and so, by applying Proposition 7.2 to each minor,
we obtain the required result. •

The only elementary graphs with fewer than four vertices are: K2 (an
edge), and C3 (a triangle). Thus, we can immediately regain the results
of Proposition 2.3 from the general formula of Proposition 7.3. We
can also use Proposition 7.3 to derive explicit expressions for the other
coefficients, for example, c\. Since the only elementary graphs with four
vertices are the cycle graph C4 and the graph having two disjoint edges,
it follows that
C4 = na - 2rib,
where na is the number of pairs of disjoint edges in F, and rib is the
number of 4-cycles in F. (See 7i.)
As well as giving explicit expressions for the coefficients of the charac-
teristic polynomial, Proposition 7.3 throws some light on the problem of
cospectral graphs (2f). The fact that elementary subgraphs are rather
loosely related to the structure of a graph helps to explain why there are
many pairs of non-isomorphic graphs having the same spectrum. This
is particularly so in the case of trees (see 7b and 7c).
We now turn to an expansion of the characteristic function of the
Laplacian matrix
<x(F; H) = detOil - Q).
Although the Laplacian matrix Q differs from —A only in its diagonal
entries, the ideas involved in this expansion are quite different from those
which we have used to investigate the characteristic polynomial of A.
One reason for this is that a principal submatrix of Q is (in general) not
the Laplacian matrix of an induced subgraph of F (the diagonal entries
give the degrees in F, rather than in the subgraph).
We shall write
CT(F; H) = det(/tl - Q) = nn + q-i/i""1 + ... + ?«-iM + qn.
The coefficient (—1)*^ is the sum of the principal minors of Q which
have i rows and columns. Using results from Chapter 6 and some simple
 Determinant expansions 47

observations we obtain
qi = -2\ET\, gn_x = (-ly^n/cOT), gn = 0.
We shall find a general expression for qi which subsumes these results.
The method is based on the expansion of a principal minor of Q = DD*
by means of the Binet-Cauchy theorem, as in the proof of Theorem 6.3.
Let X be a non-empty subset of the vertex-set of F, and Y a non-empty
subset of the edge-set of F. We denote by D(X, Y) the submatrix of the
incidence matrix D of F defined by the rows corresponding to vertices in
X and the columns corresponding to edges in Y. The following lemma
amplifies the results of Propositions 5.3 and 5.4.
Lemma 7.4 Let X and Y be as above, with \X\ = \Y\, and let VQ
denote the vertex-set of the subgraph (Y). Then D(X,Y) is invertible if
and only if the following conditions are satisfied:
(1) X is a subset ofVo;
(2) (Y) contains no cycles;
(3) VQ\X contains precisely one vertex from each component of (Y).
Proof Suppose that D(X,Y) is invertible. If X were not a subset
of VQ, then T)(X,Y) would contain a row of zeros and would not be
invertible; hence condition (1) holds. The matrix ~D(Vo,Y) is the inci-
dence matrix of (Y), and if (Y) contains a cycle then D(Vo, Y)z = 0 for
the vector z representing this cycle. Consequently D(X, Y)z = 0 and
T>(X,Y) is not invertible. Thus condition (2) holds. It follows that the
co-rank of (Y) is zero; that is,

where c is the number of components of (Y). Since \X\ = \Y\ we have

]Ti/o\X| = c. If X contained all the vertices from some component of (V),
then the corresponding rows of D(X, Y) would sum to 0, and D(X, Y)
would not be invertible. Thus VQ\X contains some vertices from each
component of (Y), and since \VQ \ X\ = c, it must contain precisely one
vertex from each component, and condition (3) is verified.
The converse is proved by reversing the argument. •
A graph $ whose co-rank is zero is a forest; it is the union of compo-
nents each of which is a tree. We shall use the symbol p(<£) to denote the
product of the numbers of vertices in the components of $. In particular,
if $ is connected it is a tree, and we have
Theorem 7.5 The coefficients qi of the polynomial <r(F; n) are given
by the formula
48 Linear algebra in graph theory

where the summation is over all sub-forests $ o/F which have i edges.
Proof Let Qx denote the principal submatrix of Q whose rows and
columns correspond to the vertices in a subset X of VF. Then <jj =
J^detQx, where the summation is over all X with \X\ = i. Using the
notation of Lemma 7.4 and the fact that Q — DD', it follows from the
Binet-Cauchy theorem that
)2 2
det Qx = ^2 det T)(X, Y) det D{X, Yf = ^ ( d e t D(X, Y))
This summation is over all subsets Y of ET with \Y\ = \X\ = i. Thus,

X,Y
By Proposition 5.3, (detD(X, Y))2 is either 0 or 1, and it takes the
value 1 if and only if the three conditions of Lemma 7.4 hold. For each
forest $ = (Y) there are p($) ways of omitting one vertex from each
component of $, and consequently there are y>($) summands equal to 1
in the expression for qi. This is the result.

Corollary 7.6 The tree-number of a graph T is given by the formula

where the summation is over all forests $ which are subgraphs of the
complement ofT.
Proof The result of Proposition 6.6, part (3), expresses K(T) in terms
of the a function of F c . The stated result follows from the formula of
Theorem 7.5 for the coefficients of a. •

This formula can be useful when the complement of F is relatively

small; examples of this situation are given in 6e and 7d. In the case of a
regular graph F, the relationship between a and x leads to an interesting
consequence of Theorem 7.5.
Proposition 7.7 LetT be a regular graph of degree k, and let\^ (0 <
i < n) denote the ith derivative of the characteristic polynomial of F.
Then

where the summation is over all forests $ which are subgraphs of F with
\E&\ = n-i.
Proof Prom part (2) of Proposition 6.6, we have
 Determinant expansions 49

The Taylor expansion of \ at the value k can be written in the form

i=0
Comparing this with a(T; (i) = J2 Qn-i/J-1, we have the result. •

We notice that the case i = 1 of Proposition 7.7 gives

which is just the formula given in Corollary 6.5.

Additional Results

7a Odd cycles (Sachs 1964) Let x(f; A) = Yl Cn-i\l and suppose

C3 = C5 = • • • = C2r-1 = 0, C 2r +1 # 0.
Then the shortest odd cycle in F has length 2r + 1, and there are
—C2r+i/2 such cycles.

7b The characteristic polynomial of a tree Suppose that J^ CjA™~1 is

the characteristic polynomial of a tree with n vertices. Then the odd
coefficients c-zr+i are zero, and the even coefficients c<iT are given by the
rule that (—l)rC2r is the number of ways of choosing r disjoint edges in
the tree.

7c Cospectral trees The result 7b facilitates the construction of pairs

of cospectral trees. For example, there are two different trees with eight
vertices and characteristic polynomial A8 — 7A6 + 10A4. Schwenk (1973)
proved that if we select a tree T with n vertices, all such trees being
equally likely, then the probability that T belongs to a cospectral pair
tends to 1 as n tends to infinity.

7d The a function of a star graph A star graph is a complete bipartite

graph Kifi. For such a graph we can calculate a explicitly from the
formula of Theorem 7.5: the result is

Consequently if F is the graph obtained by removing a star K\$ from

Kn, where n > b + 1, we have
50 Linear algebra in graph theory

7e Complete matchings, Hamiltonian cycles, and the determinant for-

mula We may write the formula for det A as

where f(r, s) is the number of spanning elementary subgraphs with rank

r and co-rank s. Two terms in this formula have special significance.
The number finji, 0) is the number of disjoint edges which cover all
the vertices - the complete matchings. The number f(n — 1,1) is the
number of spanning elementary subgraphs which are connected, that is,
the number of single cycles which cover all the vertices - the Hamiltonian
cycles.

7f Reconstruction - Kelly's lemma For each vertex v 6 VF let Fv

denote the induced subgraph (VT \ v). The deck of F is the set of
(unlabelled) induced subgraphs {Tv | v S VT}. The graph is said to be
reconstructible if every graph with the same deck as F is isomorphic to
F. The reconstruction conjecture is that every graph with at least three
vertices is reconstructible.
A function defined on graphs is said to be reconstructible if it takes
the same value on all graphs with the same deck. For any graphs F and
A, let n(F, A) be the number of subgraphs of F which are isomorphic to
A. Standard double counting arguments lead to the formula
(rA) ( r A )
£
From this formula it follows that n(F, A) is reconstructible whenever
|VA| < |VT| (Kelly 1957).

7g Reconstruction - Kocay's Lemma A sequence of graphs

is said to be a cover of the graph F if there are subgraphs A* of F such

that A, is isomorphic to <f>i (1 < i < I) and the union of the subgraphs
is F. The number of covers of F by T is denoted by c(F,^ r ). Kocay
(1981) proved that, provided all the members of the sequence T have
fewer vertices than F, the function

x
is reconstructible, where the sum is taken over all isomorphism classes
of graphs X such that \VX\ = |VT|.
 Determinant expansions 51

7h The reconstructibility of the characteristic polynomial Using the

lemmas of Kelly and Kocay, and the formula in Proposition 7.3, it can
be shown that the coefficients c, of the characteristic polynomial are
reconstructible. In particular (—l)nCn — det A is reconstructible. These
results were first established by Tutte (1979) using a different method.
His proof, and that using Kocay's lemma (as given by Bondy (1991))
both depend on showing that the number of Hamiltonian cycles, denoted
by f(n - 1,1) in 7e, is reconstructible.

7i Angles and the number of 4-cycles The number of 3-cycles in a graph

is determined by the spectrum (see 2h), but the number of 4-cycles is
not, except in special cases, such as when the graph is regular. However,
the number of 4-cycles is determined by the spectrum and the angles,
which are defined as follows. Let ]T] AaEo be the spectral decomposition
of the adjacency matrix A, as denned in 2j, and let ei, e2, •.., e n be the
standard orthonormal basis for Euclidean n-space. Then the angles are
the numbers
aij=Eiej (l<i,j<n).
More about this construction, and an explicit formula for the number of
4-cycles, can be found in a paper by Cvetkovic and Rowlinson (1988).

7j The Shannon capacity of a graph Let FA denote the product of

graphs F and A obtained by taking the vertex-set to be Cartesian prod-
uct of their vertex-sets, and defining two distinct vertices to be adjacent
if both coordinates are equal or adjacent. Let F r denote the product
of r copies of F, and let a(F r ) denote the maximum number of mutu-
ally non-adjacent vertices of F r . A construction in coding theory due to
Shannon involves the quantity
8(F) = lim(a(F r )) 1 / r ,
r—>oo
and this is known as the Shannon capacity of F. Since a(F) r < a(F r ), it
follows that a(T) < 9(F), but in general equality does not hold. Lovasz
(1979) showed that O(F) is bounded above by the largest eigenvalue of
any real symmetric matrix C for which c^ = 1 whenever Vi and Vj are
not adjacent. In particular, this yields the result 0(Cs) = Vo.
 8

Vertex-partitions and the spectrum

One of the oldest problems in graph theory is the vertex-colouring prob-

lem, which involves the assignment of colours to the vertices in such
a way that adjacent vertices have different colours. This can be inter-
preted as a problem about a special kind of partition of the vertex-set,
as described in the first definition below. In this chapter we shall apply
spectral techniques to the vertex-colouring problem, using inequalities
involving the eigenvalues of a graph. Similar techniques can also be ap-
plied to other problems about vertex-partitions, and some of these are
mentioned in the Additional Results at the end of the chapter.
Definition 8.1 A colour-partition of a general graph F is a partition
of VT into subsets, called colour-classes,
vr = v1uv2u...uvl,
such that each Vj (1 < i < I) contains no pair of adjacent vertices. In
other words, the induced subgraphs (Vi) have no edges. The chromatic
number of F, written f(F), is the least natural number I for which such
a partition is possible.
We define a vertex-colouring of F to be an assignment of colours to the
vertices, with the property that adjacent vertices have different colours,
so clearly, a vertex-colouring in which I colours are used gives rise to a
colour-partition with I colour-classes.
We note that if F has a loop, then it has a self-adjacent vertex, and
consequently no colour-partitions. Also, if F has several edges joining
the same pair of vertices then only one of these edges is relevant to
 Vertex-partitions and the spectrum 53

the definition of a colour-partition, since the definition depends only on

whether vertices are adjacent or not. Thus we can continue, for the
moment, to deal with strict graphs. However, this is allowable only for
the purposes of the present chapter; some of the constructions used in
Part Two require the introduction of general graphs.
If i'(r) = 1, then F has no edges. If v{T) — 2 then F is a bipartite
graph, as denned in 2c. Since a cycle of odd length cannot be coloured
with two colours, it follows that a bipartite graph contains no odd cycles.
This observation leads to another proof of the result established in 2c.
Proposition 8.2 Suppose the bipartite graph T has an eigenvalue A
of multiplicity m(A). Then — A is also an eigenvalue ofT, and m(—X) =
m(A).
Proof The formula of Proposition 7.3 expresses the characteristic
polynomial of a graph F in terms of the elementary subgraphs of F.
If F is bipartite then F has no odd cycles, and consequently no ele-
mentary subgraphs with an odd number of vertices. It follows that the
characteristic polynomial of F has the form
n n 2 1 6 2
X (F; z) = z + c2z ~ + CiZ* -* + ... = z p{z ),
where 6 = 0 or 1, and p is a polynomial function. Thus the eigenvalues,
which are the zeros of \, have the required property. •
The spectrum of the complete bipartite graph Ka<b can be found in
the following manner. We may suppose that the vertices of Ka^ are
labelled in such a way that its adjacency matrix is
J
A~- f[J*
A ° 0
where J is the a x b matrix having all entries +1. The matrix A has
just two linearly independent rows, and so its rank is 2. Consequently,
0 is an eigenvalue of A with multiplicity a + b — 2. The characteristic
polynomial is thus of the form za+b~2(z2 +C2). By Proposition 2.3, — c^
is equal the number of edges of Ka^, that is, 06. Hence

This example illustrates the fact (Proposition 8.2) that the spectrum
of a bipartite graph is symmetrical with respect to the origin. Indeed,
the converse of this result is also true [CvDS, p. 87]. But if i/(r) > 2
the spectrum of F does not have a distinctive property, as it does in
the bipartite case. However, as we shall see, it is possible to make
useful deductions about the chromatic number from a knowledge of the
maximum and minimum eigenvalues of F.
54 Linear algebra in graph theory

For any real symmetric matrix M , we shall denote the maximum and
minimum eigenvalues of M by A max (M) and A m i n (M). If M is the
adjacency matrix of a graph F we shall also use the notation Amax(F)
and A m ; n (F). It follows from Proposition 8.2 that, for a bipartite graph
F, we have A m i n (r) = - A m a x ( F ) .
We need a useful technique from matrix theory. Let (x, y) denote the
inner product of the column vectors x, y. For any real n x n symmetric
matrix X, and any real non-zero n x l column vector z, the number
(z, Xz)/(z, z) is known as the Rayleigh quotient, and written i?(X;z).
In matrix theory it is proved that
A m . x (X)> J R(X;*)>A I I l l l l (X) for all z jk 0,
a result which has important applications in spectral graph theory.

Proposition 8.3 (1) If A is an induced subgraph ofT, then

A m ax(A) < A r a a x (F); A m i n (A) > A r a i n (F).
(2) If the greatest and least degrees among the vertices ofT are fcmax(F)
and fcmjn(F), and the average degree is fcaVe(F), then
) > A max (F) > Aw e (r) > fcmin(F).
Proof (1) We may suppose that the vertices of F are labelled so
that the adjacency matrix A of F has a leading principal submatrix
Ao, which is the adjacency matrix of A. Let Zo be chosen such that
A o zo = A max (Ao)zo and (zo.Zo) = 1- Further, let z be the column
vector with |VT| rows formed by adjoining zero entries to Zo- Then
A m a x (A 0 ) = -R(A 0 ;z 0 ) = R(A;z) < A m a x (A).
That is, A max (A) < A max (F). The other inequality is proved similarly.
(2) Let u be the column vector each of whose entries is + 1 . Then, if
n = |VT| and fcW is the degree of the vertex Vi, we have

R(A; u) = ± £ ay = i £ *« = fcave(F).
i,j i
The Rayleigh quotient i?(A; u) is at most A m a x (A), that is A max (F), and
it is clear that the average degree is not less than the minimum degree.
Hence
A max (F) > fcave(F) > kmin(T).
Finally, let x be an eigenvector corresponding to the eigenvalue Ao =
Amax(F), and let Xj be a largest positive entry of x. By an argument
similar to that used in Proposition 3.1, we have
= (Aox)j = E'xi < k^Xj < kmax(r)Xj,
 Vertex-partitions and the spectrum 55

where the sum E' is taken over the vertices Vi adjacent to Vj. Thus
(). n
We shall now bound the chromatic number of F in terms of A max (r)
and A min (r). A graph F is l-critical if i/(F) = /, and for all induced
subgraphs A ^ F w e have u(A) < I.
Lemma 8.4 Suppose F is a graph with chromatic number I > 2. Then
F has an l-critical induced subgraph A, and every vertex of A has degree
at least I - 1 in A.
Proof The set of all induced subgraphs of F is non-empty and contains
some graphs (for example, F itself) whose chromatic number is I, and
also some graphs (for example, those with one vertex) whose chromatic
number is not /. Let A be an induced subgraph whose chromatic number
is I, and which is minimal with respect to the number of vertices; then
clearly A is Z-critical. If v is any vertex of A, then {VA \ v) is an induced
subgraph of A and has a vertex-colouring with l — l colours. If the degree
of v in A were less than l — l, then we could extend this vertex-colouring
to A, contradicting the fact that v(A) = I. Thus the degree of v is at
least I - 1. •

Proposition 8.5 (Wilf 1967) For any graph T we have

u{T) < 1 + A m a x (r).
Proof It follows from Lemma 8.4 that there is an induced subgraph
A of T such that v(A) = u(T), and kmin(A) > i/(T) - 1. Thus, using the
inequalities of Proposition 8.3, we have
K r ) < i +fcmin(A)< i + Amax(A) < i + A max (r).
•
Wilf's bound may be compared with the simple bound u < 1 + fcmax,
which is proved by an obvious argument. There is also a nontrivial
refinement of the simple bound, known as Brooks's theorem: v < kma,x
unless F is a complete graph or an odd cycle. For example, for the
complete bipartite graph Ka$ we have
kma.x{Ka,b) = max(a, b), Xmax(Katb) = y/(ab).
When a is large in comparison with 6 the second number is much smaller
than the first, but it is still a poor bound for the chromatic number, 2.
Our next major result is complementary to the previous one, in that
it provides a lower bound for the chromatic number. We require a pre-
liminary lemma and a corollary.
56 Linear algebra in graph theory

L e m m a 8.6 Let X be a real symmetric matrix, partitioned in the form

P Ql
Q' R j '
where P and R are square symmetric matrices. Then

Proof Let A = Amjn(X) and take an arbitrary e > 0. Then X* =

X — (A — e)I is a positive-definite symmetric matrix, partitioned in the
same way as X, with
P* = P - (A - e)I, Q* = Q, R* = R - (A - e)I.
By applying the method of Rayleigh quotients to the matrix X*, it can
be shown that
Amax(X ) < A m a x (" ) + Amax(rL ).
(See for instance Linear Transformations by H. L. Hamburger and M. E.
Grimshaw (Cambridge, 1956), p. 77.) Thus, in terms of X, P and R,
we have
Amax(X) - (A - e) < A max (P) - (A - e) + A max (R) - (A - c),
and since c is arbitrary and A = Amin(X) we have the result. •

Corollary 8.7 Let A be a real symmetric matrix, partitioned into t2

submatrices A,j in such a way that the row and column partitions are
the same; in other words, each diagonal sub-matrix An (1 < i < t) is
square. Then
t
: - l)A min (A) < V A m a x ( A i i ) .

Proof We prove this result by induction on t. It is true when t = 2, by

the lemma. Suppose that it is true when t = T — 1; then we shall show
that it holds when t = T. Let A be partitioned into T2 submatrices,
in the manner stated, and let B be the matrix A with the last row and
column of submatrices deleted. By the lemma,
Amax(A) + Amin(A) < A max (B) + A m a x (A T T ),
and by the induction hypothesis,
T-l
Amax(B) + (T - 2)Amin(B) < ] T A max (A ii ).

Now Amin(B) > Amjn(A), as in the proof of Proposition 8.3. Thus,

adding the two inequalities, we have the result for t = T, and the general
result follows by induction. •

We can now establish a lower bound for the chromatic number.

 Vertex-partitions and the spectrum 57

Theorem 8.8 (Hoffman 1970) For any graph F, whose edge-set is

non-empty,

Proof The vertex-set VF can be partitioned into v = i/(F) colour-

classes; consequently the adjacency matrix A of F can be partitioned
into v2 submatrices, as in the preceding corollary. In this case, the
diagonal submatrices An (1 < i < u) consist entirely of zeros, and so
Amax(Aj,) = 0 (1 < i < v). Applying Corollary 8.7 we have
Amax(A) + (i/ - l)Amin(A) < 0.
But, if F has at least one edge, then Amin(A) = Amjn(r) < 0. The result
now follows. •

In cases where the spectrum of a graph is known, Hoffman's bound

can be very useful. Consider, for example, the graph E which arises
from the classical configuration of 27 lines on a general cubic surface, in
which each line meets 10 other lines. The vertices of £ represent lines,
and adjacent vertices represent skew lines, so that £ is a regular graph
with 27 vertices and degree 16. This is the graph with least eigenvalue
—2 mentioned in 3g. Since Amax(£) = 16 and Amjn(£) = - 2 , Hoffman's
bound is v(E) > 1 + 16/2 = 9, a result which would be difficult to
establish by direct means. On the other hand, it is fairly easy to find
a vertex-colouring using 9 colours (Haemers 1979), so Hoffman's bound
leads to the exact answer v(£) = 9 in this case.

Additional Results

8a The eigenvalues of a planar graph Let F be a planar connected

graph. Then it follows from Theorem 8.8 and the four-colour theorem
that
Amin(A ) S 7jA m a x (l )•

8b Another bound for the chromatic number Let F be a regular graph

of degree k with n vertices. In any colour-partition of F each colour-class
has at most n — k vertices; consequently i/(F) > n/(n — k). Cvetkovic
(1972, see also 8h) proved a corresponding result for any, not necessarily
regular, graph:

<r> s
58 Linear algebra in graph theory

8c The second eigenvalue of the Laplacian The eigenvalues of a real

symmetric matrix may be characterised in terms of the Rayleigh quo-
tient. In particular, for the first non-zero eigenvalue MI of the Laplacian
matrix Q, the characterisation asserts that
Mi = min i?(Q;x),
u*x=O
where u is the all-1 vector, the eigenvector corresponding to Mo- This
provides a powerful method for finding upper bounds for Mi- If we think
of x as a real-valued function £ defined on the vertex-set, the condi-
tion u*x = 0 becomes J2 €(v) — 0? an<3 for any function satisfying this
condition we have (by a simple manipulation of the Rayleigh quotient)

{v,w}eE ' v

8d A spectral bound for the isoperimeiric number Let 8X be the cut

defined by X C VT, and let x = \X\, n = |V"17|. Define f (v) to be I/a: if
v £ X, and — l/(n — x) otherwise. Then 8c implies that

Mi < |**| (-
\x n — x/
It follows that, for the isoperimetric number defined in 4b, we have
*(F) > Mi/2, and in the fc-regular case, i(T) > (k — X\)/2 (Alon and
Milman 1985).

8e Equipartitions Suppose that T is a graph with n vertices, and X is

a partition of its vertex-set. Let 6X denote the set of edges whose ends
are in different parts. We say that X is an equipartition if each part has
the same size p; then n = pq where q is the number of parts. Biggs,
Brightwell, and Tsoubelis (1992) showed that in this case
( 2
Mi < \SX\

8f The odd graphs (Biggs 1979) Let k be a natural number greater than
1, and let 5 be a set of cardinality 2k — 1. The odd graph Ok is defined as
follows: its vertices correspond to the subsets of S of cardinality k — 1,
and two vertices are adjacent if and only if the corresponding subsets
are disjoint. (For example, O2 = -K3, and O3 is the Petersen graph.) Ok
is a regular graph of degree k; when k — 2 its girth is 3, when k = 3 it
is 5, and when k > 4 it is 6,
The spectrum of Ok can be obtained by the methods described in
Chapters 20 and 21 (see 21b). In particular, the largest eigenvalue Amax
 Vertex-partitions and the spectrum 59

is k, the next largest is A: — 2, and the least is 1 — k. Using Theorem 8.8

and 8d respectively, we get the following lower bounds for the chromatic
and isoperirnetric numbers:

=1.
To see how good these bounds are, let V[a,/?] denote the set of vertices
containing a given pair a,0 € S; let V[a, /?] denote the set of vertices
containing a but not /3; and so on. Then the three sets V[a, 0\, V\a, 0\,
and V[a, /3]uV[a, /?] form a colour partition, so v(Ok) = 3. Furthermore,
the cut denned by X — V[a, /?] U V[a, 0\ and its complement satisfies

2( 2 *r 2 3 ) k

Thus i(Ofc) < 1 + (k — I ) " 1 . Further results about the odd graphs may
be found in 17d, 20b, and 21b.

8g The Motzkin-Straus formula Consider the quadratic programming

problem (QP)
maximize x'Ax subject to u*x = 1, x > 0,
where A is the adjacency matrix of a graph F. Define the support of a
feasible vector x to be the set of vertices vt for which x» ^= 0. It can
be shown that, for an optimal x with minimal support, the support is a
clique (a complete subgraph) in T. It follows that the maximum value
for the QP is 1 — l/w(r), where co(F) is the size of the largest clique in
F. This is the formula of Motzkin and Straus (1965). Putting Xi = 1/n
for i = 1,2,...,n = |VT|, and letting m - \ET\, we get

1 >
w(r) - n2 '
In particular, we have Turdn's Theorem: if F has no triangles, then
m < n 2 /4.

8h Another spectral bound Let s be the sum of the entries of the nor-
malized eigenvector corresponding to Amax. Wilf (1985) observed that
the Motzkin-Straus formula leads to the result w > s2/(s2 — A max ). Since
s2 < n it follows that u > n/(n — A max ). Furthermore, the chromatic
number v cannot be less than w, so this strengthens the result 8b of
Cvetkovic.
 PART TWO

Colouring problems
 9

The chromatic polynomial

Part Two is concerned with polynomial functions which represent certain

numbers associated with graphs. The best-known example, the chro-
matic polynomial, is introduced in this chapter. It should be stressed
that here we have to deal with general graphs, because some of the
constructions fail when restricted to strict graphs.

Definition 9.1 Let F be a general graph with n vertices, and let u

be a complex number. For each natural number r, let m r (F) denote
the number of distinct colour-partitions of VT into r colour-classes, and
define «(r) to be the complex number u(u — l)(u — 2 ) . . . (u — r +1). The
chromatic polynomial of F is the function defined by

Proposition 9.2 If s is a natural number, then C(T;s) is the number

of vertex-colourings of F using at most s colours.
Proof Every vertex-colouring of F in which exactly r colours are used
gives rise to a colour-partition into r colour-classes. Conversely, for each
colour-partition into r colours we can assign s colours to the colour-
classes in s(s — 1)... (s — r + 1) ways. Hence the number of vertex-
colourings in which s colours are available is J2m r (F)s( r ) = C(F; s).
•
64 Colouring problems

The simplest example is the chromatic polynomial of the complete

graph Kn. Since every vertex of Kn is adjacent to every other one, the
numbers of colour-partitions are
mi{Kn) = m2(Kn) = ... = mn-i{Kn) = 0; mn(Kn) = 1.
Hence
C(Kn; u) = u{u - l)(w - 2 ) . . . (u - n + 1).
Possibly the most important fact about the chromatic polynomial is
that it is indeed a polynomial; in other words, the number of vertex-
colourings of a graph, with a given number of colours available, is the
value of a polynomial function. This is because the expressions U(r)
which occur in the definition are themselves polynomials.
Some simple properties of the chromatic polynomial follow directly
from its definition. For example, if F has n vertices, then mn(T) = 1;
hence C(F; u) is a monic polynomial of degree n. Other results follow
directly from Proposition 9.2 and the principle that a polynomial is
uniquely determined by its values at an infinite set of natural numbers.
For instance, if F is disconnected, with two components Fi and F2, then
we can colour the vertices of Fi and F2 independently, and it follows that
C(T;s) = C(Fi; s)C(F2; s) for any natural number s. Consequently
C(r;u)=C(ri;u)C(r2;u),
as elements of the ring of polynomials with integer coefficients.
Since u is a factor of U(r) for all r > 1, it follows that C(F;0) = 0
for any general graph F. If F has c components, then the coefficients of
1 = u°, u1,..., uc~l are all zero, by virtue of the result on disconnected
graphs in the previous paragraph. Also, if ET / 0 then F has no vertex-
colouring with just one colour, and so C(F; 1) = 0 and u — 1 is a factor
of C(r;u).
The problem of finding the chromatic number of a graph is part of the
general problem of locating the zeros of its chromatic polynomial, be-
cause the chromatic number ^(F) is the smallest natural number u which
is not a zero of C(F; u). This fact has stimulated some interesting work
(see 9i, 9j and 9k for example), but, as yet, elementary methods have
proved more useful in answering questions about chromatic numbers.
The simplest method of calculating chromatic polynomials is a recur-
sive technique. Suppose that F is a general graph and that e is an edge
of F which is not a loop. The graph F^e' whose edge-set is ET \ {e} and
whose vertex-set is VF is said to be obtained by deleting e; while the
graph F(ej constructed from F^e^ by identifying the two vertices incident
with e in F, is said to be obtained by contracting e. We note that F ^
 The chromatic polynomial 65

has one edge fewer than F, and F(ej has one edge and one vertex fewer
than F, and so the following Proposition provides a method for calculat-
ing the chromatic polynomial by repeated reduction to 'smaller' graphs.
This is known as the deletion-contraction method.

Proposition 9.3 The chromatic polynomial satisfies the relation

C(T;u) = C(r<e>;«) - C(T(e);u).
Proof Consider the vertex-colourings of F ^ with s colours available.
These colourings fall into two disjoint sets: those in which the ends of
e are coloured differently, and those in which the ends of e are coloured
alike. The first set is in bijective correspondence with the colourings of
F, and the second set is in bijective correspondence with the colourings
of F(e). Hence C(T^; s) = C(F; s) + C(F(e); s) for each natural number
s, and the result follows. •

Corollary 9.4 If T is a tree with n vertices then

C(T;u)=u{u-l)n-\
Proof We prove this by induction, using the elementary fact that any
tree with n > 2 vertices has a vertex (in fact, at least two vertices) of
degree 1. The result is clearly true when n = 1. Suppose it is true when
n = N — 1, and let T be a tree with TV vertices, e an edge of T incident
with a vertex of degree 1. Then T^ has two components: an isolated
vertex, and a tree with N — 1 vertices, the latter being T(e). Hence

and using Proposition 9.3 and the induction hypothesis

C{T;«) = (u - l)C(T (e) ; u) = (« - l)u(« - I ) " " 2 = u(u - \)N~X.
Hence the result is true when n = N, and for all n by the principle of
induction. •
The deletion-contraction method also yields the chromatic polynomial
of a cycle graph Cn. If n > 3 the deletion of any edge from Cn results
in a path graph Pn, which is a tree with n vertices, and the contraction
of any edge results in a cycle graph Cn-\- Hence
C(Cn;u) = u(u - I ) " - 1 - C(C n _ i ; u).
Since C3 = K3, we have
C(C3; u) = u{u - \){u - 2) = (u - I) 3 - (w - 1).
We can solve the recursion given above, with this 'initial condition', to
obtain the formula
66 Colouring problems

We now describe two other useful techniques for calculating chromatic

polynomials. The first is concerned with the 'join' operation for graphs.
Suppose Fi and F2 are two graphs; then we define their join Fj + F2 to
be the graph with vertex-set and edge-set given by

+ F2) = ETi U ET2 U {{x, y}\xeVTuye VT2}.

In other words, Fi + F2 consists of copies of Fi and F2 with additional
edges joining every vertex of Fi to every vertex of F2.

Proposition 9.5 The numbers of colour-partitions of F = Ti + T2

are given by
mi(T) =

Proof Since every vertex of Fj is adjacent (in F) to every vertex of

F2, any colour-class of vertices in F is either a colour-class in Fj or a
colour-class in F 2 . Hence the result. •

Corollary 9.6 The chromatic polynomial of the join Ti + F 2 is

C(Ti + F 2 ; u) = C(Ti;u) o C(F 2 ;u),
where the o operation on polynomials signifies that we write each poly-
nomial in the form 2ZTOiM(i) and multiply as ifu^ were the power ul.
D

For example, the complete bipartite graph -^3,3 is the join ./V3 + JV3,
where Nn is the graph with n vertices and no edges. From Corollary
9.6, we have

+ 3U(2) + U(i)) O (u (3 ) + 3U(2) + U

6U(5) + llU(4) + 6W(3) + W(2)
M5 + 36w4 - 75u3 + 78u2 - 31u.
The chromatic polynomials of all complete multipartite graphs can be
found in this way.

Another application of the method yields the chromatic polynomials

of the graphs N\ + F and N2 + F, sometimes known as the cone and
suspension of F, and denoted by cT and sT respectively.
 The chromatic polynomial 67

Proposition 9.7 The chromatic polynomials of a cone and a suspen-

sion are given by
C(cT; u) = uC(F; u - 1),
C(sT; u) = u{u - 1)C(F; u - 2) + uC(T; u - 1).
Proof Let C(T; u) = ^2 miU^). Using Corollary 9.6 and the fact that
U(i+i) = u{u - l)(j), we have
C(cT; u) = C(Nx +T;u)=uo C(T; u) = u (1) o

j(w - l)(j) = uC(T;u - 1).

The second part is proved in a similar way, using the identity u2 —

D

Another useful technique for the calculation of chromatic polynomials

applies to graphs of the kind described in the next definition.
Definition 9.8 The general graph V is quasi-separable if there is a
subset K of VT such that the induced subgraph (K) is a complete graph
and the induced subgraph (VT \ K) is disconnected. T is separable if
| AT | < 1; in this case either K = 0 so that T is in fact disconnected, or
\K\ = 1, in which case we say that the single vertex of K is a cut-vertex.
It follows that in a quasi-separable graph V we have VT = V\ U V2,
where (Vi Pi V2) is complete and there are no edges in F joining V\ \
(Vi n V2) to V2 \ (V! n V2). We shall refer to the pair (Vi, V2) as a
quasi-separation of F, or simply a separation if |Vi PI V2I < 1.
A graph which is quasi-separable but not separable is shown in Figure
3; the relevant quasi-separation is given by Vi = {1, 2,4}, V2 = {2,3,4}.

Figure 3: a quasi-separable graph.

Proposition 9.9 If the graph F has quasi-separation (Vi, V2) then

C(F;u) =
68 Colouring problems

Proof If V\ ("1 Vi is empty we make the convention that the denomina-

tor is 1, and the result is a consequence of the remark about disconnected
graphs following Proposition 9.2. Suppose that (Vi n V2) is a complete
graph Kt, t > 1. Since F contains this complete graph, F has no vertex-
colouring with fewer than t colours, and so u^ is a factor of C(T;u).
For each natural number s > t, C(F; s)/s( t ) is the number of ways of
extending a given vertex-colouring of {Vi D V2) to the whole of F, us-
ing at most s colours. Also, both {Vi} and (V2) contain the complete
graph Kt = (V\ n V2), so C({Vi}; s)/s( t ), i € {1,2}, has a corresponding
interpretation. Since there are no edges in F joining V\ \ {V\ D Vj) to
V2 \ (Vi n V2), the extensions of a vertex-colouring of {Vi n V"2) to {Vi)
and to {V2) are independent. Hence
C(T;a)

for all s >t. The corresponding identity for the polynomials follows.
•
The formula of Proposition 9.9 is often useful in working out chromatic
polynomials of small graphs. For instance, the graph shown in Figure 3
is 'two K3's with a common K2'; hence its chromatic polynomial is
" ( * - ! ) ( « - 2 ) « ( u - ! ) ( « - 2 ) = u(u _ 1 ) ( n _ 2)2_

An important theoretical application of Proposition 9.9 will be described

in Chapter 12.

Additional Results

9a Wheels and pyramids The cone of the cycle graph Cn-\ is the wheel
or pyramid Wn\ the suspension of Cn-i is the double pyramid Un. The
chromatic polynomials of these graphs are
C(Wn; u) = u(u - 2) 71 - 1 -(- ( - l ) " - 1 ^ - 2);
I n ; u) = u(u - l)(u - 3) n - 2 + u(u - 2)n~2 + (-l)nu(u2 - 3u + 1).

9b The cocktail-party graphs Let ps(u) — C(Hs;u), where Hs is the

cocktail-party graph .#2,2,...,2 with 2s vertices. The polynomials ps(u)
can be found from the recursion
pi(u)=u2, ps(u) = u(u - I)p3-i(u - 2) + ups-i(u - 1) (s > 2).
 The chromatic polynomial 69

9c Ladders and Mobius ladders The ladder Lh (h > 3) is a regular

graph of degree 3 with 2h vertices u\, v.2, • • •, «h, v\, v%,..., Vh\ the ver-
tices u\,... ,Uh form a cycle of length h, as do the vertices v\,..., Vh,
and the remaining edges are of the form {u.j,i>,}, 1 < i < h. The Mobius
ladders Mh were defined in 3e. By systematic use of the deletion-
contraction method Biggs, Damerell, and Sands (1972, see also 9i)
showed that
C(Lh; u) = (u2 -3u + 3)h + (u- 1){(3 - u)h + (1 - u)h) + u2 -
C(Mh; u) = (u2 -3u + 3)h + (u - 1){(3 - u)h - (1 - u)h} - 1.

9d The chromatic polynomial characterizes trees Corollary 9.4 implies

that different graphs may have the same chromatic polynomial, since any
two trees with the same number of vertices have this property. However,
if F is a simple graph with n vertices, and C(T; u) = u{u - I)™"1, then
T is a tree (Read 1968).

9e Chromatically unique graphs A graph is said to be chromatically

unique if it is the only graph with its chromatic polynomial. From
Corollary 9.4 we know that any tree with more than three vertices is
not chromatically unique. Several families of graphs are known to be
chromatically unique, among them the following.
(a) The complete graphs, Kn.
(b) The cycle graphs, Cn (n > 3).
(c) The wheel graphs Wn for odd n.
It is known that W§ and W$ are not chromatically unique, but W\o is.
See Li and Whitehead (1992) for this result and additional references.

9f The chromatic polynomials of the regular polyhedra The chromatic

polynomials of the graphs formed by the vertices and edges of the five
regular polyhedra in three dimensions are known. The graph of the
tetrahedron is K4, the graph of the octahedron is H3 = K2,2,2 (9b),
and the graph of the cube is L4 (9c). The chromatic polynomial of
the icosahedron was computed by Whitney (1932b); after removing the
factors u(u — l)(u — 2)(u — 3), it is
u8 -24u 7 +260u 6 - 1670u5+6999w4 - 19698u3+36408u2 -40240u+20170.
The computation of the chromatic polynomial of the dodecahedron was
first attempted by D.A. Sands (in an unpublished thesis, 1972) and
Haggard (1976). In order to reduce the size of the coefficients it is
70 Colouring problems

convenient to express the result in the form

17
-u(u - l)(w - 2) ] T Ci(l - u)\

In this form the coefficients Ci are all positive, and they are:
1, 10, 56, 230, 759, 2112, 5104, 10912, 20880, 35972,

55768, 77152, 93538, 96396, 80572, 50808, 21302, 4412.

9g Interpolation formulae Suppose that two finite sequences of real

numbers mo, m i , . . . , mn and Po,Pi, • • • ,pn are related by the rule
k
Pk = ^ m r f c ( r ) .
r=0
Then there is an inverse formula giving the m's in terms of the p's, and
this in turn leads to a formula for the polynomial p(u) of degree n whose
value at k € {0,1,..., n} is pk'-

fc=0 V
' r=0
2.
fc=0 V ;

In particular we have formulae for the numbers of colour-partitions, and

the chromatic polynomial, in terms of the numbers of fc-colourings.
9h Acyclic orientations An orientation of a graph, as defined in Chapter
4, is said to be acyclic if it has no directed cycles. For example, on a
tree with n vertices any orientation is acyclic, so there are 2n~l acyclic
orientations. Stanley (1973) showed that in general the number of acyclic
orientations of F is the absolute value of C(F; —1).
9i Recursive families and chromatic roots As was remarked at the
beginning of this chapter, the location of the zeros of a chromatic poly-
nomial is a fundamental problem, because it subsumes the problem of
finding the chromatic number. One of the few positive results in this
direction is that the zeros for some families of graphs lie near certain
curves in the complex plane. Biggs, Damerell and Sands (1972) defined
a recursive family of graphs {Tn} to be a sequence of graphs in which the
polynomials C(Tn; u) are related by a linear homogeneous recurrence,
in which the coefficients are polynomials in u. In this case C(Tn; u) can
be expressed in the form
fe

where the functions a* and Aj are not necessarily polynomials. For

 The chromatic polynomial 71

example, the ladders form a recursive family, and as in 9c we have

a\(u) = 1, 0:2(11) = u - 1, as(u) =u—l, a4(u) = u2 - 3u + 3;
Ai(u) = u2 — Zu + 3, A2(«) = 3 - u, Xaiu) = 1 - u, A4(u) = 1.
Define a chromatic root of the family {Fn} to be a complex number (
for which there is an infinite sequence (un) such that un is a zero of
C(Tn;u) and limu n = £. Beraha, Kahane and Weiss (1980) obtained
necessary and sufficient conditions for ( to be a chromatic root, and
Read (1990) explained how their results confirm empirical observations
of Biggs, Damerell and Sands concerning the chromatic roots of the
ladder graphs. It turns out that the chromatic roots of the ladders are
0, 1, together with the points lying on parts of two quartic curves and
the line Hu-2.
9j Planar graphs It is clear that the integers 0,1,2,3 are zeros of
C(T; u) for suitable planar graphs F, but the four-colour theorem tells
us that the integer 4 is never a zero. The first result about non-integral
zeros was obtained by Tutte (1970). He observed that there is often a
zero close to (3 + v/5)/2 = 2.6180 ..., and he proved that for any graph
F with n vertices which triangulates the plane
n
/ /— \ ~5
C(F; (3 •

Figure 4: the iterated octahedron.

However this does not imply that (3 + \/5)/2 is a chromatic root of

every family of plane triangulations. For example, the 'iterated octahe-
dron' A n (Figure 4) has chromatic polynomial
C(A n ; u) = u(u - l)(u - 2)(u3 - 9«2 + 29M - 32)n.
Tutte's result applies to this family, and consequently all the graphs
have a zero 'near' 2.6180 But this zero is a constant, 2.5466...,
independent of n; there is no zero which tends to 2.6180... as n —* 00.
72 Colouring problems

9k Chromatic roots of planar graphs Tutte's result (9j) led to specula-

tion concerning the numbers bn = 2 + 2cos(27r/n) as chromatic roots of
families of planar graphs, based on the observations that b? = 0, 63 = 1,
64 == 2, 65 = 2.6180..., b6 = 3, and bn -> 4. Beraha and Kahane (1979)
proved that 4 is indeed a chromatic root of a family of planar graphs, and
Beraha, Kahane and Weiss (1980) proved the the same thing for 65,67,
and 610. Concerning the numbers bn in general, Tutte (1984) observes
that 'their significance is not yet properly understood'.

91 Zero-free intervals If F is a graph which triangulates the plane,

there are no non-integral zeros £ of C(T; u) for which C, < 2.5466 ..., the
zero of the octahedron mentioned in 9j (Woodall 1992).

9m Confluence of the deletion-contraction method In the method of

deletion and contraction we are free to choose any edge at each step.
The fact that the order of choosing edges does not affect the final result
is obvious, given the concrete interpretation of the chromatic polynomial
in terms of colourings. However, the deletion-contraction method may
be applied formally, as a set of so-called rewriting rules, and then it is
necessary to prove that there is a 'normal form', independent of the order
in which the rules are applied. This follows from two general properties
of the rewriting rules, known as 'well-foundedness' and 'local confluence'
(Yetter 1990).

9n The umbral chromatic polynomial If P is a partition of an n-set in

which there are a, parts of size i, then we define the formal expression

Given a graph F, let

p
where the sum is over all colour-partitions P of VT with r parts. Clearly,
putting <f>i — <f>2 — • • • = (j>n-i = 1 we obtain the ordinary m r (F), as
defined on p. 63.
Ray and Wright (1992) show that the corresponding generalization of
the chromatic polynomial is obtained by replacing the expressions U(r)
by what are known as the conjugate Bell polynomials bf{u). Thus they
define the umbral chromatic polynomial

r=l
They obtain interpolation formulae like those in 9g and analogues of
other properties of the ordinary chromatic polynomial.
 10

Subgraph expansions

It is clear that calculating the chromatic polynomial of a graph is at

least as hard as finding its chromatic number. The latter problem is
known to be difficult, in a technical sense which appears to correspond
with practical experience. (More details may be found in the Additional
Results at the end of Chapter 13.) There are nevertheless good reasons,
both theoretical and practical, for studying methods of calculating the
chromatic polynomial which are more sophisticated than those discussed
in the previous chapter. These methods are based on the idea of an
'expansion' in terms of certain subgraphs.
Definition 10.1 The rank polynomial of a general graph F is the
function defined by
r
R(T;x,y) = E*
scEr
where r(S) and s(S) are the rank and co-rank of the subgraph (5) off.
If we write R(T; x, y) — £ prsxrys then prs is the number of subgraphs
of F with rank r and co-rank s, and we say that the matrix (prs) is the
rank matrix of T.
For example, the rank matrix of the graph K3>3 is
1
9
36
84 9
117 45 6
L 81 78 36 9 1J
74 Colouring problems

Here the rows are labelled by the values of the rank r from 0 to 5, and the
columns are labelled by the values of the co-rank s from 0 to 4. We notice
that since r(S)+s{S) — \S\ for all S C ET, an antidiagonal (sloping from
bottom left to top right) corresponds to subgraphs with a fixed number
t of edges, and consequently sums to the binomial coefficient (^). We
observe also that the number in the bottom left-hand corner (generally
Pn-i,o, where n is the number of vertices) is just the tree-number of the
graph. These facts mean that in this case very few entries need to be
calculated explicitly.
As we shall see, several interesting functions can be obtained by as-
signing particular values to the indeterminates x and y in the rank poly-
nomial. Trivially, putting y = x gives R(T;x,x) = (x + l)'^ 1 "', because
r{S) + s(S) = \S\ for all S C ET. The main result to be proved in
this chapter is that by assigning certain values to x and y we obtain the
chromatic polynomial.
For any natural number u, let [u] denote the set {1,2,... ,u}, which
we shall think of as a set of u colours, and let [u]x denote the set of all
functions u> : X —» [u]. For a general graph F, the set [u]vr contains
some functions which are vertex-colourings of F with u colours available,
and some functions which are not vertex-colourings since they violate the
condition that adjacent vertices must receive different colours. In order
to pick out the vertex-colourings we make the following definition.

Definition 10.2 For each w € [v]vr we define the indicator function

Q:ET-* {0,1} as follows.
~, . _ f 1, if e has vertices v\,V2 such that w(vi)
\0 otherwise.
In particular, cD(e) = 0 if e is a loop.
Lemma 10.3 IfT is a general graph and u is a natural number, then

n
Proof The product n ^ ( e ) *s z e r 0 unless Q(e) = 1 for all e € EF,
and this is so only if u> is a vertex-colouring of F. Thus the sum of these
products is the number of vertex-colourings of F using at most u colours.
The result follows from Proposition 9.2. •

Theorem 10.4 The chromatic polynomial of a graph F with n vertices

has an expansion in terms of subgraphs as follows:
C(T;u)=
SCET
 Subgraph expansions 75

Proof For any natural number u we have

e =
^) E
Expanding the product of terms 1 + /(e) we obtain a sum of
expressions n/( e )> o n e f° r e a c n subset S C ET. That is,

c(r;«)= X) E
We now switch the order in the double sum. For each 5 C ET let
VS — V(S); then any function from VS to [u] is the restriction to VS
of u l v r \ v s l functions from VT to [u]. Thus

|yrxvsi 1
E E n(-(«)-!)= E-
vr
w6[u] scsrees sccr
E nw- )'
ue[u]v's ces
Consider the product n ( ^ ( e ) ~ 1) o v e r a ^ edges e € 5. If the product
is non-zero, uj(e) must be 0 for each e G S, which means that w is
constant on each component of (S). In this case the value of the product
is (—1)ISL If {S) has c components there are uc such functions ui; hence
the sum of the product over all v)vs\ functions u> : VS —> [u] is (—l)'slufi.
The result follows from the equation
|VT \ VS\ + c = n - \VS\ + c = n- r{S).
D

Corollary 10.5 The chromatic polynomial and the rank polynomial

of a general graph T with n vertices are related by the identity

If the chromatic polynomial is

C(T; u) = boun + hun-1 +...+ bn^u + bn,
then the coefficients bi can be expressed in terms of the entries in the
rank matrix, as follows:

j
Proof The identity between the polynomials follows directly from
Theorem 10.4 and the definition of the rank polynomial. In terms of the
76 Colouring problems

coefficients, we have
M " " ' = C(T;u) = unR(T; -u~\ -1)

Equating coefficients of powers of u, and rearranging the signs, we have

the result stated above. •

The formula for the coefficients expresses bi as an alternating sum of

the entries in the zth row of the rank matrix. This formula was first
studied by Birkhoff (1912), in the original paper on chromatic polyno-
mials, and Whitney (1932a). For example, from the rank matrix for
^3,3 given above we have
bx = - 9 , 62 = 36, h = -84 + 9 = - 7 5, b4 .= 117 - 45 + 6 = 78,
65 = - 8 1 + 78 - 36 + 9 - 1 = - 3 1 .
This checks with the result obtained in Chapter 9 by a different method:
3t3; u)=u6 - 9M5 + 36M4 - 75u3 + 78u2

Proposition 10.6 Let F be a strict graph of girth g, having m edges

and r\ cycles of length g. Then, with the above notation for the coeffi-
cients of the chromatic polynomial of T, we have
(1) (-l) < 6,-=(7) for i= 0,l,...,g-2;
(2) ( - l ) » - V i = (fl-i)-?'-
Proof A subgraph of F with rank i < g — 2 must have co-rank zero,
since F has no cycles with fewer than g edges. Thus, for alH < g — 2,
we have p^ = (™) and ptj = 0 if j > 0. Further, the only subgraphs
of F with rank g — 1 are the ( ™J forests with g — 1 edges (which have
co-rank zero), and the r\ cycles with g edges (which have co-rank 1).
Thus,

Pg-i,o = ( . ) , pg-i,i=ri, pg-u=--0, if j > 1.

The result follows from the expression for the coefficients of the chro-
matic polynomial. •

We observe that for a strict graph the girth g is at least 3, so the coef-
ficient of un~1 in the chromatic polynomial is — m, where n and m are
the numbers of vertices and edges respectively.
 Subgraph expansions 77

The formula for the coefficients of the chromatic polynomial is an

alternating sum, and its use involves counting many subgraphs which
'cancel out' in the final result. Whitney (1932a) discovered a reduction
which involves counting fewer subgraphs. His result also shows that the
non-zero coefficients of the chromatic polynomial alternate in sign; that
is, (—l)lbi is always positive. Let F be a simple graph whose edge-set
ET = {e\, e2,..., em} is ordered by the natural order of subscripts. This
ordering is to remain fixed throughout our discussion, A broken cycle
in F is the result of removing the first edge from some cycle; in other
words, it is a subset B of ET such that for some edge e; we have
(1) B U {e{\ is a cycle in F; (2) i > I for each edge e, G B.
The next proposition expresses the coefficients of the chromatic polyno-
mial in terms of the subgraphs which contain no broken cycles; clearly
such subgraphs contain no cycles, and so they are forests.
Proposition 10.7 (Whitney 1932a) Let T be a strict graph whose
edge-set is ordered as above, and let C(T;u) = ~^2,biUn~l. Then (— l)lbi
is the number of subgraphs ofT which have i edges and contain no broken
cycles.
Proof Suppose Bi, 52> • • • ,Bt is a list of the broken cycles of F, in
dictionary order based on the ordering of ET. Let /» (1 <i <t) denote
the edge which, when added to Bi, completes a cycle. The edges /, are
not necessarily all different, but, because of the way in which the broken
cycles are ordered, it follows that /j is not in Bj when j > i.
Define So to be the set of subgraphs of F containing no broken cycle,
and for 1 < h < t define S^ to be the set of subgraphs containing Bh
but not Bh+i,Bh+2, • • • ,Bt- Then So, S i , . . . , S^ is a partition of the
set of all subgraphs of F. We claim that, in the expression

the total contribution to the sum from S i , . . . , S t is zero.

Suppose 5 is a subset of ET not containing /^; then S contains Bh if
and only if S U {fh} contains Bh. Further, S contains Bi (i > h) if and
only if 5 U {fh} contains Bi, since fh is not in Bi. Thus if one of the
subgraphs {S), {S U {fh}) is in S/,, then both are in S^. They have the
same rank, but their co-ranks differ by one, and so their contributions
to the alternating sum cancel. Consequently, we need only consider the
contribution of S o to J2(-iyPij• Since a subgraph (S) in S o is a forest,
it has co-rank j = 0 and rank i = \S\, whence the result. •

Corollary 10.8 Let T be a strict graph with rank r. Then the co-
78 Colouring problems

efficients of C(T;u) alternate strictly in sign; that is, (—l)lbi > 0 for
i = O,l...,r.
Proof The characterization of Proposition 10.7 shows that (—l)*6j >
0 for 0 < i < n. In order to obtain the strict inequality we must show
that there is a subgraph with i edges and containing no broken cycle, for
1 = 0 , 1 , . . . , r. Suppose we successively remove edges from F in such a
way that at least one cycle is destroyed at each stage; this process stops
when we reach a subgraph (F) of F with \F\ = r and s(F) = 0. Let us
order the edges of F so that the edges in F come first. Then (F) contains
no broken cycle, and any subset of F generates a subgraph containing
no broken cycle. Thus we have produced the required subgraphs, and
the result follows. •

Recall that, at the beginning of Chapter 9, we observed that 6j = 0 if

i — n,n—1,..., n—(c—1), where n = \VT\ and F has c components. That
is, bi = 0 if i = r + 1 , . . . , n. Thus we have shown that the coefficients
of the chromatic polynomial alternate strictly, and then become zero.

Additional Results

10a Inequalities for the coefficients of the chromatic polynomial If F

is a connected strict graph with n vertices and m edges, and C(F; u) =
S6jUn""J, then

10b Codichromatic graphs An example of two non-isomorphic general

graphs having the same rank matrix was found in the 1930's by Marion
C. Gray (see Figure 5).

Figure 5: two general graphs with the same rank matrix.

 Subgraph expansions 79

Tutte (1974) drew attention to this work and constructed pairs of strict
graphs which have the same rank matrix.

10c V-functions A function / defined on isomorphism classes of graphs

and taking values in a ring A is a V-function if it satisfies the following
conditions.
(a) If F is empty then f(T) = l.
(b) If T is the union of disjoint graphs Fj, F 2 , then /(F) = / ( r i ) / ( r 2 ) .
(c) If e is any edge of F which is not a loop, then / satisfies the deletion-
contraction formula

It is easy to see that the chromatic polynomial and the rank polyno-
mial, suitably normalized, are V-functions. The most general V-function
is constructed as follows. For any sequence i = i o , i i , i 2 , . . . of non-
negative integers, with finite sum, let v(T, i) be the number of spanning
subgraphs of F which have i& components of co-rank k, for k > 0. Let
s = (SQ, SI, «2, • • •) be any infinite sequence of elements of A and let

Then / s is a V-function, and every V-function can be written in this

way (Tutte 1947b).

lOd The rank polynomial as a V-function By taking the ring A to be

the ring of polynomials with integer coefficients in two indeterminates
x,y and s to be the sequence defined by s, = xyl, we obtain the rank
polynomial (with suitable normalization) as a V-function.

lOe Homeomorphic graphs The operation of replacing an edge with

ends u, v by two edges with ends u, w and w, v, where w is a new vertex,
is known as subdividing the edge. Regarding a graph as a topological
space in the obvious way, it is clear that subdividing an edge results in
a graph homeomorphic to the original one. In general, two graphs are
homeomorphic if they can both be obtained from the same graph by
sequences of subdivisions. A graph function is said to be a topological
invariant if its values on homeomorphic graphs are equal. It can be
shown that a non-trivial V-function / is a topological invariant if and
only if / ( # ! ) = - 1 .

lOf Interaction models The formula obtained in Lemma 10.3 arises

naturally in theoretical physics. The vertices of the graph F = (V, E)
80 Colouring problems

are 'particles', each of which which can have one of u attributes, so that
a 'state' of the system is a function UJ : V —> [u]. Each pair of adjacent
vertices, corresponding to an edge e & E, has an interaction ^ ( e ) which
depends on the state w, and the 'weight' I(w) is the product of the
interactions iuj(e). The partition function is the sum of all the weights:

The chromatic polynomial is the special case arising when iu is the

indicator function a; as in Definition 10.2: that is, iu{e) is 1 if the ends
of e have different attributes, and is 0 otherwise. In general, the value
of u and the function iu determine an interaction model. An account of
the properties of such models was given by Biggs (1977b).

lOg The Ising and Potts models Using the 'interaction model' termi-
nology, suppose iu (e) is a if the ends of e have the same attribute in the
state w, and 1 otherwise. For general u this is known as the Potts model,
and the special case u = 2 is known as the Ising model. The partition
function for the Potts model can be expanded as a rank function:

£ ,a - 1).
U
SCEV

lOh A general form of the subgraph expansion An interaction model

for which iu(e) takes only two values, one when the ends of e have
the same attribute and another when they have different attributes,
is said to be a resonant model. The expansions in terms of the rank
function described above can be generalized to any resonant model in
the following way. Let F and G be resonant models for which the two
(distinct) values of iu(e) are / i , / o and gi,go respectively, and let fa =
19i +6, i = 0,1. Then Zp(T) can be expanded in terms of the values of
Z<3 on the subgraphs of F as follows:

lOi Another expansion of the chromatic polynomial Nagle (1971) ob-

tained the following expansion:
C(T;u)= ] T ulvr\vsl(l-«-1)lBr\sliV«5>;«),
seer
where the function N has the following properties: (a) if F has an isth-
mus, then iV(F; u) = 0; (b) TV is a topological invariant.
 11

The multiplicative expansion

In this chapter and the next one we shall investigate expansions of the
chromatic polynomial which involve relatively few subgraphs in compar-
ison with the expansion of Chapter 10. The idea first appeared in the
work of Whitney (1932b), and it was developed independently by Tutte
(1967) and researchers in theoretical physics who described the method
as a 'linked-cluster expansion' (Baker 1971). The simple version given
here is based on a paper by the present author (Biggs 1973a). There are
other approaches which use more algebraic machinery; see Biggs (1978)
and lie.
We begin with some definitions. Recall that if a connected graph F is
separable then it has a certain number of cut-vertices, and the removal
of any cut-vertex disconnects the graph. A non-separable subgraph of T
which is non-empty and maximal (considered as a subset of the edges)
is known as a block. Every edge is in just one block, and we may think
of F as a set of blocks 'stuck together' at the cut-vertices. In the case
of a disconnected graph we define the blocks to be the blocks of the
components. It is worth remarking that this means that isolated vertices
are disregarded, since every block must have at least one edge.
Let Y be a real-valued function defined for all graphs, and having the
following two properties.
PI: Y(T) = 1 if T has no edges;
P2: Y(T) is the product of the numbers Y(B) taken over all blocks
B ofF.
82 Colouring problems

Given such a Y, let X be the real-valued function defined by

X(T) - ]T Y{S).
SCET
An example of a function satisfying P I and P2 is obtained by taking
Y(T) = xr(r^ys(r\ where x and y are a given pair of real numbers, in
which case the corresponding X is (an evaluation of) the rank polyno-
mial of F. The fact that this Y satisfies P2 is a consequence of the
equations
r(r) = £>(£), *(r) = £>(£),
where the sums are taken over the set of blocks B of F.
Lemma 11.1 If the function Y satisfies P I and P2, then the corre-
sponding function X satisfies the same properties.
Proof (PI) If r has no edges then the sum occurring in the definition
of X contains only one term, F(0), which is 1.
(P2) Suppose F has just two blocks Fi and F2, with edge-sets E\ and
E2. Then for any S C ET, the sets Si = S fl Ex and S2 = S D E2, are
such that S = St U S2 and St n S2 = 0. Thus the blocks of (S) in F
are the blocks of (Si), regarded as a subgraph of Fi, together with the
blocks of (^2) regarded as a subset of F2- By P2 we have
Y(S)r = Y(S1)r1Y(S2)r2-
(This equation remains true if either or both of Si, S2 are empty, by
PI.) Consequently,
X(T) = ]T Y(S)r = Y, E
SCET SlCE1S2CE2

S1CE1 S2CE2
If F has b > 2 blocks we have a similar argument, taking Ti to be the
first 6—1 blocks. Hence the general result follows by induction. •

We shall now transform the sum X(T) into a product, using exponen-
tial and logarithmic functions. We require also the fundamental identity
underlying the 'principle of inclusion and exclusion': that is

ICJ
provided that J is not the empty set.
Definition 11.2 Let (X, Y) be a pair of functions as above, and sup-
pose that the values of X are positive. Then the logarithmic transform
 The multiplicative expansion 83

of the pair (X, Y) is the pair of functions (X, Y) denned by

X(T) = (-l)l £ r l 5 2 (-1)1*1 logX{5), Y(T) = expX(F).
SCET

Proposition 11.3 Let F be a general graph. IfT has no edges, or if

F is separable and has no isolated vertices, then X(T) = 0.
Proof If ET is empty then X(T) = 1, and consequently X(T) = 0.
Suppose that F has no isolated vertices and is separable. Then either
it is disconnected or it is connected and has at least one cut-vertex. In
either case it can be expressed as the union of two subgraphs (E\) and
(E2) with E\ and E2 non-empty and disjoint. For S C £T, we have
X(S) = X(S1)X(S2),
where Si = S n E\ and S2 = S n E2, and so logX(S) = log
log X(S2}- This justifies the following calculation.

SCET

5Z E (-l) |Sll+|S2| (log A-<5,) + log

S^Ej S2C£2

(-l)ls'llogX{51) 52 (-
x S2C£2

52 (-i)|S2| io g x(s2
Both E\ and £^2 are non-empty, so the fundamental inclusion-exclusion
identity stated above implies that the entire expression is zero, and we
have the result. •

Theorem 11.4 Let F be a non-separable graph, and let (X, Y) be the

logarithmic transform of the pair (X, Y). Then X(T) has a multiplicative
expansion
X{T)= n Y{S),
SCEF
in which Y is equal to 1 (and so may be ignored) for separable subgraphs
ofT.
Proof The fact that Y{S) = expX(S) = 1 for separable subgraphs
(S) follows from the previous lemma, since, by definition, a subgraph
has no isolated vertices.
84 Colouring problems

We shall prove that

iogx(r)=
SCET

from which the theorem follows by taking exponentials. Now, from the
definition of X,
E x(S)=
SCEV scEr RCS
and (R) as a subgraph of (5) is identical with (R) as a subgraph of T.
Writing Y = S\R the right-hand side becomes
E E (-D|iJ|+|y|(-i)|fl|iogx(i?>
RCET YCEV\R

= E ^gx(R) E (-D|v|-
RCEV YCEV\R

The inner sum is non-zero only when EF\R = 0; that is, when R = ET.
Thus the expression reduces to log X(ET) = \ogX(T), as required.
•
We now apply the general theory of the logarithmic transform to the
particular case of the chromatic polynomial. We take the function Y to
be

This satisfies P I and P2 and, by Theorem 10.4, the corresponding X

function is

Lemma 11.5 Let (XU,YU) denote the particular pair of functions

given above. Then, for a given graph T, Xu(Y)and YU(T) can be defined
for all sufficiently large integers u.
Proof In order to define XU{T) satisfactorily we must ensure that
\ogXu(S) is denned for all subsets S C £T. Now if u is an integer
greater than the chromatic number of F, it is clear that C((S)\u) is
positive and so the logarithm of XU(S) = u~^v^s^C((S);u) is defined.
•
We can eliminate the logarithmic and exponential functions from the
general definition of Y, obtaining
Y(T)= I ] {*<S»e(S), where e(5) = (-l)l £ r \ s | .
SCET
 The multiplicative expansion 85

For the particular case Yu we get

YU(T)= JJ {u-^C((S);u)y(s\
scEr
which is valid for all sufficiently large positive integers u. The product
formula shows that Yu is a rational function in its domain of definition.
We shall find it convenient to deal separately with the Yu function
for a single edge, that is YU(K.2)- It is easy work this out explicitly:
Yu(K2) = \-vTl.
Proposition 11.6 For every non-separable graph A having more than
one edge there is a rational function q(A; u) such that the chromatic
polynomial of a graph T has a multiplicative expansion

where the product is taken over all those non-separable subgraphs A o / F

which have more than one edge.
Proof We have seen that, if (XU,YU) is the pair defined by
Yu(r) = (-1)1*%-^), XU(T) = u^vr\C(T; u),
then Yu is defined Jbr all subgraphs of F provided u is a^sufficiently
large integer, and YU{K2) = (1 — M" 1 ). Setting q(A;u) = YU{A) when
\EA\ > 1 and applying Theorem 11.4 we see that the identity holds for
an infinite set of values of u. Since both sides are holomorphic functions,
they are identical. •

The functions q(T; u) can be found explicitly for certain standard

graphs. For example, for the cycle graph Cn the only subgraph occuring
in the product is Cn itself; hence
C(Cn;u)=un(l-u-x)nq(Cn;u).
By a result of Chapter 9, the left-hand side is (u - l ) n + (-l) n (u - 1),
so that

This simple calculation highlights an apparent circularity which arises

if we propose to use the multiplicative expansion to calculate chromatic
polynomials. The difficulty is that the right-hand side of the multiplica-
tive expansion of C(T; u) contains a term q(T;u), and we have, as yet,
no way of finding q(T; u) without prior knowledge of C(T;u). In the
next chapter it will be shown that this seemingly fundamental objection
can be surmounted by means of a few simple observations. We shall also
obtain a version of Proposition 11.6 in which the number of subgraphs
involved is reduced still further.
86 Colouring problems

Additional Results

l l a The q function of a crossed cycle Let C+ denote a graph con-

structed from the cycle graph Cn by the addition of one edge joining
two distinct vertices which are not adjacent in Cn. Then

l i b Theta graphs (Baker 1971) Let Qr,s,t denote the graph consisting
of two vertices joined by three disjoint paths of length r, s, and t. ©r,s,t
has n = r + s + t— 1 vertices and r + s + t edges, and q(Qr,s,t',u) is

1 - (1 - u)r-n - (1 - u)s~n - (1 - it)*-" + (2 - u)(l - u)~n

(1 - (1 - u)r-n)(l - (1 - u)s~n)(l - (1 - uf-n) '

l i e The multiplicative expansion of the rank polynomial If Y(T) =

x r ( r )y"( r ) then X(T) = R(T; x, y), and the logarithmic transform applied
to the pair (X, Y) leads to a multiplicative expansion

where the product is over all non-separable subgraphs A of F which have

more than one edge (Tutte 1967).

l i d Whitney's theorem on counting subgraphs In Chapter 10 we ob-

tained a formula for the coefficients of the chromatic polynomial which
involved counting all the subgraphs. In this chapter we have shown
that, in theory, only the non-separable subgraphs are needed. Whitney
(1932b) obtained this result in a different way, by showing that there is a
general expression for the number of subgraphs of any particular type in
terms of the numbers of non-separable subgraphs. Specifically, let fit(F)
be the number of subgraphs of F which have a given type t, where a
'type' is determined by the number of blocks of each isomorphism class.
Then there is a polynomial function <J?t, independent of F, with rational
coefficients and no constant term, such that
n t (r)=* t (n <y (r),n T (r),... ),
where a, T, . . . are the nonseparable types with not more edges than t.
For example, if A|| denotes the type with one block isomorphic to Kz
and two blocks isomorphic to K2 we have
7 1 2
2 +
 The multiplicative expansion 87

where 0 is the type of the theta graph ©2,2,1 and the other notation is
self-explanatory.

l i e An algebraic framework In order to unify the theory of the mul-

tiplicative expansion and Whitney's theorem described above, Biggs
(1977b, 1978) introduced the following algebraic framework. Define St,
the set of star types, to be the set of isomorphism classes of non-separable
graphs, and Gr, the set of graph types, to be the set of functions from St
to the non-negative integers, with finite support. Let X and Y respec-
tively be the vector spaces of real-valued functions defined on St and
Gr. When St is regarded as a subset of Gr in the obvious way, we have
a projection J : Y —> X.
For a given graph F, of type g, define cg € Y by the rule that cg(t) is
the number of subgraphs of F which are of type t. Then Jcg represents
the numbers of non-separable subgraphs of F. Whitney's theorem asserts
that there is an operator W : X —• Y, such that
W(Jcg) = cg for all g € Gr.
In the papers quoted, it is proved that W = B~1U, where B is a
linear operator defined by a certain infinite matrix, and U : X —* Y is
the 'monomial' mapping defined by

l l f Expansions as linear functional Denote the subspaces of X and

Y consisting of vectors with finite support by X o and Yo respectively.
The real vector spaces Xo and Y o admit scalar products defined in the
usual way:
((xi,x 2 » = 53XI(<T)X2(<T); (yi.yz) = 53yi(*)y2(*)-
a t
For any given m € Y o there is a linear functional M defined by M(y) =
(y, m). On vectors cg representing 'real graphs', M{cs) is, by definition
of the scalar product, a sum over subgraphs in which each subgraph of
type t contributes m(i). In the author's papers quoted above it is shown
that under certain conditions there is a corresponding linear functional
C on Xo such that
expC(Jcg) = M(cg) for all g e Gr.
Explicitly, we have
£(x) = ((x, 1)), where 1 =
88 Colouring problems

l l g The Hopf algebra framework There is clearly a substantial amount

of algebraic structure underlying Whitney's theorem and the multiplica-
tive expansion. Schmitt (1993) carries this idea to its logical conclusion
by introducing coalgebras and Hopf algebras. He shows that the algebra
of formal power series with rational coefficients over St can be given the
structure of a Hopf algebra, and that it is isomorphic to the dual of the
free module with rational coefficients over Gr. Whitney's theorem is a
direct consequence of the isomorphism.
Another approach using Hopf algebras is discussed by Ray (1992).
 12

The induced subgraph expansion

In this chapter we shall modify the multiplicative expansion of the chro-

matic polynomial in such a way that the induced subgraphs are the
only ones occurring in the formula. This procedure has two advantages.
First, there are fewer induced subgraphs than subgraphs in general; and
secondly, the function which takes the place of the q function (in the
notation of Proposition 11.6) turns out to be trivial for a wider class of
graphs.
The formal details of the transition to induced subgraphs are quite
straightforward. For any non-separable graph A define
Q(A;«)=JIg(A;ti),
where the product is over the set of spanning subgraphs A of A, that
is, those for which VA — VA. It follows immediately that Q is a ratio-
nal function of u. For example, the cycle graph Cn has just one non-
separable spanning subgraph, which is Cn itself. Thus the definition of
Q gives
Q(Cn;u) = q(Cn;u) =

Proposition 12.1 The chromatic polynomial has a multiplicative ex-

pansion

where the product is over all non-separable induced subgraphs ofT having
more than one edge.
90 Colouring problems

Proof The factors which appear in Proposition 11.6 can be grouped

in such a way that each group contains those subgraphs of F which have
a given vertex-set. This grouping of factors corresponds precisely to that
given in the definition of Q and the resulting expression for C, since each
subgraph A of F is a subgraph of exactly one induced subgraph A of F
(the one for which VA = VA), and conversely, each subgraph of A is a
subgraph of F. •

The crucial fact which makes the multiplicative expansion useful in

practice is that the q and Q functions are rational functions of a special
kind. Specifically, it can be shown that

where v and 6 are polynomials whose degrees satisfy

deg S - deg v > \VT\ - 1.
The first satisfactory proof of this important fact was given by Tutte
(1967), using the notion of 'tree mappings'. An algebraic proof was
given by Biggs (1978, see also l i e and llf) .
Given this result, we can prove the same thing for Q.

Proposition 12.2 LetT be a non-separable graph. ThenQ(T;u) may

be written in the form

where v and 6 are polynomials such that deg 6 — deg v > \VT\ — 1.
Proof The function Q is defined to be the product of functions q, over
a set of graphs with the same number of vertices. Thus the result for q
implies the result for Q. •

We are now in a position to overcome the circularity mentioned at

the end of the previous chapter. It is possible, using Proposition 12.2,
to calculate both C(F; u) and Q(T\ u) provided only that we know the
Q functions for all proper induced subgraphs of F, that is, the induced
subgraphs not including F itself. To see this we write the formula of
Proposition 12.1 as

where P(u) is a product of rational functions corresponding to the proper

induced subgraphs, including the vertices (for each of which we have
factor u) and the edges (for each of which we have a factor 1 — u~1).
 The induced subgraph expansion 91

It follows that P(u) can be written as a polynomial of degree n — \VT\

plus a power series in u~l:
P(u) =un + a i u n - 1 + . . . + a n _ i u + an + ctn+ivT1 + ... .
But, following Proposition 12.2, the function Q(F; u) can be written
Q(F; u) = 1 + (30u~n+l + /? l U - n + ... .
It follows that multiplying P(u) by this expression does not alter the
coefficients of un,un~1,... ,u2 in P(u). Thus the polynomial part of
P(u) is a correct expression for C(T; u), except for the coefficients of u
and 1. But these coefficients in C(F; u) are easily found, by noting that
u(u - 1) is a factor of C(T; u). It follows that both C(T; u) and Q{T; u)
are determined by the known function P(u).
An example will elucidate this argument. Take F = K\\ then the only
proper induced subgraphs of T having more than one edge are the four
copies of K3 = C3. Thus

Q(K4;u)

Dividing (u - I) 2 into u2(u - 2)4 gives P(u) = u4 - 6u3 + llu2

and so
C(K4; u) = u4- 6u3 + llu 2 - au + b.
Since u(u — 1) is a factor of C(K4; u) it follows that a = 6, 6 = 0 and
C(K4; u) = u4 - 6w3 + llu 2 - 6u = u(u - l)(u - 2){u - 3).
We can also find Q{K4\ u) by substituting back, obtaining
2u-3

The technique which we have just described has the important conse-
quence that we can calculate chromatic polynomials merely by counting
induced subgraphs, without knowing any C and Q functions in advance.
In particular, it implies that the chromatic polynomial is reconstructible,
in the sense of 7f.
To make this explicit, suppose that Ai, A2,..., A; is a list of the iso-
morphism types of non-separable induced subgraphs of F, where K\ =
Ai and K2 = A2 axe included, for the sake of uniformity, and F = A;.
Then we define a matrix N = (n^), by putting mj equal to the number
of induced subgraphs of A» which are isomorphic with Aj. We may sup-
pose that the list has been ordered in such a way that N is a triangular
matrix each of whose diagonal entries is +1.
92 Colouring problems

Proposition 12.3 The matrix N completely determines the chromatic

polynomial ofT.
Proof We know the C and Q functions for all the graphs with at
most three vertices. Now, suppose we know the C and Q functions for
the induced subgraphs of F with at most t vertices; then we can find
the C and Q functions for each induced subgraph with t + 1 vertices,
by using the technique previously explained. Thus, using this procedure
recursively leads to the chromatic polynomial of F. •

For example, the following is a complete list of the non-separable

isomorphism types of induced subgraphs of the ladder graph L3. (The
graph itself occurs as A6 in Figure 6.)

A, Aa A., A5

Figure 6: the induced subgraphs of L3.

The N matrix for F is:

ri
2
3
4
5
.6 lj
To see how the method works, suppose that we have completed the
calculations for subgraphs with at most four vertices. The C and Q
functions for these graphs are as follows.
A2 A3 A4

C: u(u-l) u(u - l)(u 2 - 3w + 3)

Q : (u - l)/u u(u - 2)/(u - I) 2 u(u2 -3u + 3)/(u - I) 3

The remainder of the calculation now proceeds in the following way. We
have C(A5;u) = P5(u)Q(A5;u), where

= u(u - l)(u - 2)(w2 - 3w + 3).

 The induced subgraph expansion 93

Here (atypically) -Ps(w) is a polynomial divisible by u(u — 1), and so

C(A5;u) = u(u- l)(u - 2)(u2 -3u + 3) and Q(A5;u) = l.
At the next stage, we have C(As;u) = PQ(u)Q(Ae;u), where

(I) 6
= u6 - 9u5 + 34u4 - 67u3 + 67w2
Here Pe(w) is not a polynomial. Extending the terms in u2 and above
to a polynonial divisible by u(u — 1) we get C(T; u) = u6 — 9u5 + 34u4 —
67u3 + 67u2 - 26u.

One noteworthy feature of the preceding calculation is that Q(A$; u) =

1, although A5 is a non-separable graph. This means that we could
have ignored A5 completely, both in setting up the matrix N and in
the subsequent calculations. The next proposition shows that there is a
large class of non-separable graphs T for which Q(F; u) = 1.

Proposition 12.4 (Baker 1971) / / the graph T is quasi-separable, in

the sense of Definition 9.8, then Q(T;u) = 1.
Proof We prove this result by induction on the number of vertices
of F. The result is true for all quasi-separable graphs with at most
four vertices. For this set contains only one graph (the graph shown in
Fig.3, p. 67) which is not in fact separable, and the claim can be readily
checked for that graph.
Suppose that the result is true for all quasi-separable graphs with at
most L vertices, and let T be a quasi-separable graph with L+1 vertices.
We have a quasi-separation (Vi,^ ) of T, where (Vi (~l V2) is complete
and (VT - (V\ n V2)) is disconnected. The expansion of Proposition 12.1
can be written in the form
) = P(u)Q(T;u),
where P(u) is a product of factors corresponding to the proper non-
separable induced subgraphs of F. If U is any proper subset of VT
for which U % V\ and U % V2, then (U) is a quasi-separable graph,
with quasi-separation (Vi n U), (V2 CiU). By the induction hypothesis,

Thus the non-trivial terms in the product P(u) correspond to the

subsets of V\ and the subsets of V2. However, a subset of V\ D V2 occurs
just once, rather than twice. It follows that

P(u) =
C((V1DV2};u)
94 Colouring problems

Since Proposition 9.9 tells us that C(T; u) is also equal to this expression,
it follows that Q(F; u) = 1, and the induction step is verified. •

We observe that the graph A5, in the example preceding the propo-
sition, is in fact quasi-separable, and so the fact that Q(A5;u) = 1 is
explained.
The following theorem is the culmination of the theory developed in
Chapters 10-12.

Theorem 12.5 The chromatic polynomial of a graph is determined

by its proper induced subgraphs which are not quasi-separable.
Proof This theorem follows from Propositions 12.1 and 12.4. •

We close this chapter with a brief explanation of how the theory can be
used to study the 'chromatic polynomial' of an infinite graph. Suppose
\P is an infinite graph which can be regarded in some way as the limit of
a sequence of finite graphs 9n, with (FvE'nl = vn say. The appropriate
definition of the 'chromatic polynomial' of <fr is

provided the limit exists for a suitable range of values of u. In theoreti-

cal physics this is known as taking the thermodynamic limit, and some
existence results have been proved for 'interaction models', as defined in
lOf. Grimmett (1978) obtained strong results for the rank polynomial,
but for our present purposes, blind faith and ignorance will suffice.
If *„ has reasonable regularity properties, then the number of induced
subgraphs of a given type in \t n is avn, where a is a constant represent-
ing the density, that is the number of induced subgraphs of that type
per vertex. For example, if \&n is regular of degree k, the number of
edges is (k/2)vn, and so the density of edges is k/2. If we now take the
(l/w n )th root of the multiplicative Q-formula for C(\l/ra; u) we get a term
u (corresponding to the vertices), a term (1 — u~x)k/2 (corresponding to
the edges), and in general a term Q{A; u)a for each induced subgraph
A of density a. This leads to a definition of the 'chromatic polynomial'
which does not depend on the approximating sequence \&n. Unfortu-
nately, nothing is known about the convergence of the infinite product,
although it is clear that the smallest induced subgraphs, which are the
easiest to count, contribute the largest terms.
A good illustration is provided by the infinite plane square lattice
graph. Here the only induced subgraphs which are not quasi-separable
and have not more than eight vertices are: the vertices, edges, CVs
 The induced subgraph expansion 95

and Cs's, with densities 1,2,1 and 1 respectively. It follows that an

approximation to C ^ in this case is

The correct value when u = 3 is known to be (4/3)3//2 = 1.540... (Lieb

1967), whereas the approximation gives 1.512 For larger values of u
it seems likely that the approximation is better, but no general results
are known. (See also 12f.)

Additional Results

12a The Q function for complete graphs We have

Q(Kn;u)= 1Q (u-i)" ( i \
0<i<n-l

where ^ ) = (-ir- 1 -("T 1 ).

12b The Q functions for all graphs with less than six vertices The only
graphs with less than five vertices which are not quasi-separable are K<i,
K3, K4, and d, and we have already found Q for all these. Writing
Q(T; u) = 1 4- r(T; u), the r functions are as follows:
r(K2;u) = 1/u
r(K3;u)= -l/{u-lf
r(K4; u) = -(2u - 3)/u(u - 2)2
r(C 4 ;u)= l/(w-l)3-
The relevant graphs with five vertices are: K$, W5, W*~ (the wheel with
one 'spoke' removed), X2>3 and C5. The r functions are:
r(K5; u) = -(6u4 - 48w3 + 140u2 - 176u + 81)/(u - l) 4 (u - 3)4
r(W5; u) = (3u2 - 9u + 7)/u(u - 2f(u2 - 3u + 3)
r(W*-;u) = (2u2 - 6u + 5)/u(u - 2)(M 2 - 3u + 3)2
r(K2,3; u) = (u3 - 6u2 + llu - 7)/u(u2 - Zu + 3)3
r(C5;u)= -l/(u-l)\

12c Petersen's graph The only non-quasi-separable induced subgraphs

of Petersen's graph 0 3 have 2,5,6,7,8,9,10 vertices, and there is one
96 Colouring problems

isomorphism class in each case. The N matrix is

5 1
6 0 1
8 2 1 1
10 4 2 4 1
12 6 4 9 3 1
.15 12 10 30 15 10 1
Using the method described on pp. 92-93 this gives the chromatic poly-
nomial of O3:
u(u - 1)(u - 2)(u7 - 12w6 + 67u5 - 230u4 + 529u3 - 814u2 + 775u - 352).
12d The first non-trivial coefficient in q and Q If F is non-separable
and has n vertices and m edges, then the coefficient of u~^n~^ in the
expression for q(T; u) in descending powers of u is equal to ( - l ) m . The
corresponding coefficient in Q{T;u) is therefore XX~1)'BA'> where the
summation is over all non-separable spanning subgraphs A of F (Tutte
1967).
12e Chromatic powers Let <rm(F) denote the sum of the mth powers
of the zeros of C(F; u). Suppose that

where the expansion is valid for |u| sufficiently large. If n(F; A) denotes
the number of induced subgraphs of F which are isomorphic with A, we
have

where the sum is taken over isomorphism classes of non-quasi-separable

graphs (Tutte 1967).
12f Approximations for the infinite square lattice There have been
many attempts to determine the 'chromatic polynomial' Coo(w) of the
infinite square lattice. Biggs and Meredith (1976) obtained the estimate
2
hu
z - 3 + y/u - 2u + 5).
Using the 'transfer matrix' method, Biggs (1977a) obtained the bounds
< \{u 2 + v ^ - 4 u + 8).
Kim and Enting (1979) obtained a series approximation in terms of
x = \/{u — 1): apart from a simple factor it is
+ x7 + 3xs + 4x9 + 3x10 Ux 12 + 24a;
13
8z
14

- 91a;15 - 261a;16 - 290x 17 254x 18

 13

The Tutte polynomial

There is a remarkable relationship between the rank polynomial and the

spanning trees of a graph. In this chapter we shall develop this theory
by giving an explicit definition of what is known as the Tutte polynomial
T(F; x, y) of a graph F in terms of its spanning trees, and then proving
an identity between the Tutte polynomial and the rank polynomial.
An alternative approach to the Tutte polynomial is to define it recur-
sively by the deletion-contraction property,

where e is neither a loop nor an isthmus. This rule, together with a

'boundary condition' (see 13c), does in fact define T completely. How-
ever, it is not immediately obvious that the method leads to a result
which is independent of the order in which edges are deleted and con-
tracted, and it provides no insight into the remarkable properties of T.
For these reasons we shall follow the constructive route given below.
The definition of the rank polynomial depends upon the assignment
of the ordered pair (rank, co-rank) of non-negative integers to each sub-
graph; we shall call such an assignment a bigrading of the set of sub-
graphs. If F is connected, the set of subgraphs whose bigrading is
(r(r),0) is just the set of spanning trees of F. We shall introduce a
new bigrading of subgraphs which has the property that, if it is given
only for the spanning trees of F, then the entire rank polynomial of F
is determined. Our procedure is based initially upon an ordering of the
edge-set ET, although a consequence of our main result is the fact that
98 Colouring problems

this arbitrary ordering is essentially irrelevant. Another consequence of

the main result is an expansion of the chromatic polynomial in terms of
spanning trees; this will be the subject of Chapter 14.
We now fix some hypotheses and conventions which will remain in
force throughout this chapter. The graph T is a connected general graph,
and ET has a fixed total ordering denoted by <. If X C ET we shall
use the symbol X (rather than (X)) to denote the corresponding edge-
subgraph of F, and the singleton sets {x} C ET will be denoted by x
instead of {x}. The rank of T will be denoted by ro; thus ro = r(T) =
\vr\ - 1 .
If X C ET and x £ ET \ X, then the rank of X U x is either r(X)
or r(X) + 1, and in the latter case we say that x is independent of X.
Now if r(X) •£ ro, there will certainly be some edges of T which are
independent of X, and we shall denote the first of these (in the ordering
<) by X(X). We note that, since
r(Y) + s(Y) = \Y\ for all Y C ET,
we have the equations
r(X U X(X)) = r(X) + 1, s(X U X(X)) = s(X).
Similarly, if s(X) ^ 0 then there are some edges x for which s(X \ x) =
s(X) - 1, and we denote the first of these by n(X). We have
r(X \ n(X)) = r(X), s(X \ M(X)) = s(X) - 1.
Definition 13.1 The A operator on subsets of ET assigns to each
set X C ET the set Xx derived from X by successively adjoining the
edges \(X), X(X U X(X)),..., until no further increase in the rank is
possible. The /x operator takes X to the set XM, which is derived from
X by successively removing the edges fi(X), fi(X \ n(X)),..., until no
further decrease in the co-rank is possible.
We notice the following properties of the A and /x operators:
XCX\ r(Xx) = r0, s(Xx) = s(X).
X» C X, r(X") = r(X), a(X") = 0.
We shall exploit the obvious similarity between the two operators by
giving proofs only for one of them. The first lemma says that the edges
which must be added to a subgraph A to form Ax can be added in any
order. (In what follows the notation x < y will mean x < y and x ^ y.)
Lemma 13.2 If AQB CAX, then Bx = Ax.
Proof If A = Ax, the result is trivial. Suppose
AX\A =X = {xi,x2)...,xt},
 The Tutte polynomial 99

where xi < x2 < ... < xt, and let B = A U Y, where Y C X. If Y = X,

then B = Ax and Bx = Axx = Ax. If Y / X, let xa be the first edge
in X \ Y. Then, if an edge x is independent of B, it follows that x
is independent of A U {xi,... ,xa-i} (which is contained in B), and so
xa < x, since xa is the first edge independent of
>lU{xi,...,xa_i}.
But xa itself is certainly independent of B, since when we add the edges
in X to A, the rank must increase by exactly one at each step. Thus
xa = X(B), and by successively repeating the argument with B' = B U
X(B), B" = B'U X(B'),..., we have the result. •

Lemma 13.3 IfACB and r(B) ^ r 0 , then X(B) e ^4A.

Proof Since r(B) ^ ro, there is a first edge X(B) independent of B,
and consequently independent of A. Suppose X(B) is not in Ax. Then
each edge x in Ax \ A must satisfy x < X(B), and so x is not independent
of B; also, since A C B, no edge in A is independent of S. Thus all
edges in Ax are not independent of B, and r(B) = r(Ax) — ro. This is
a contradiction, so our hypothesis was false and X(B) is in Ax. D

We note the analogous properties of the \i operator:

A* C B C A => B" = A"; B <Z A and s(B) ^ 0 =» /i(B) ^ ^ .
The next definition introduces a new bigrading of the subsets of ET.
Definition 13.4 Let X be a subset of ET. An edge e in ET\X is said
to be externally active with respect to X if n(X U e) = e. An edge / in
X is said to be internally active with respect to X if X(X \f) = f. The
number of edges which are externally (internally) active with respect to
X is called the external (internal) activity of X.
We shall denote the sets of edges which are externally and internally
active with respect to X by Xe and X1 respectively, and use the notation
X+=XUX€, X~=X\XL.
These concepts are motivated by their interpretation in the case of a
spanning tree, because in that case they are related to the systems of
basic cycles and cuts which were discussed in Chapter 5.
Proposition 13.5 For any spanning tree TofTwe have:
(1) the edge e is externally active with respect to T if and only if e is the
first edge (in the ordering <) of cyc(T, e);
(2) the edge f is internally active with respect to T if and only iff is the
first edge (in the ordering <) of cut(T, / ) .
100 Colouring problems

Proof By definition, e is externally active if and only if e is the first

edge whose removal decreases the co-rank of TU e. But T Ue contains
just one cycle, which is cyc(T, e), and any edge whose removal decreases
the co-rank must belong to this cycle.
The second part is proved by a parallel argument. •

Definition 13.6 The Tutte polynomial of a connected graph F, with

respect to an ordering < of ET, is denned as follows. Suppose ty is the
number of spanning trees of F whose internal activity is i and whose
external activity is j . Then the Tutte polynomial is

Remarkably, it will turn out that T is independent of the chosen ordering.

In order to obtain the main result we shall investigate the relation-
ship between the concepts just defined and the following diagram of
operators.

A -^ B

V
C -^ V
Here A denotes all subsets of ET, B denotes subsets Z with r(Z) = r0,
C denotes subsets W with s(W) = 0, and V denotes subsets T with
r(T) = ro and s(T) = 0 (that is, spanning trees). It is worth remarking
that the diagram is commutative, although we shall not need this result
(see 13g).
Proposition 13.7 Let X be any subset in the image of the X operator,
so that r(X) = r0 and Xx = X. Then

Proof Suppose X = YX. Then Y C Yx = X, so Y C X. If / is an

edge of X~, then certainly / is in X — Yx. If / were in Yx \ Y, then by
Lemma 13.2, X(YX \ f) = f; but this means that / is internally active
with respect to X = Yx, contradicting / € X~. Thus / is in Y, and
X~ QY.
Suppose X- C Y C X. If X = Y, then we have X = Xx = Yx.
Now if / € X \ Y, then / is internally active with respect to X and so
X(X\f) = f. Prom F C I \ / w e have (by Lemma 13.3) X(X\f) e Yx,
that is, / G Yx. Since this is true for all / in X \ Y, it follows that
X\Y CYX, and consequently X C Yx. Finally, from Definition 13.1
and Y C X C Yx we deduce that Xx = Yx, that is X = Yx. D
 The Tutte polynomial 101

We note the analogous result: if X is in the image of the /i operator,

then

Proposition 13.8 Let T be a spanning tree (that is, T e V), and

suppose W £ C is such that Wx = T. Then We = Te.
Proof Suppose that the edge e is externally active with respect to
T. We shall show that the whole of cyc(T, e) belongs to W, whence it
follows that e is externally active with respect to W. If there is an edge
f ^ em cyc(T, e) which is not in W, then, since (by Proposition 13.7)
we have T~ C W C T, / must be internally active with respect to T.
Now / 6 cyc(T, e) implies that e € cut(T,/), and the internally active
property of / means that / < e. This contradicts the externally active
property of e. Hence cyc(T, e) C W, and e is externally active with
respect to W.
Conversely, if e is externally active with respect to W it follows im-
mediately that e is externally active with respect to T. •

We now set up the main theorem, using the portion A —* C —> T>
of the operator diagram. Define
Pii = \{X € A | r(X) = r 0 - t, s(X) = j}\,
ni = \{W 6 C | r(W) = r 0 - t, |W~| = j}|,

Of course, the last line merely repeats Definition 13.6. We have three
corresponding two-variable polynomials.
R{T; x,y)=Yl Pa^V3> P ( r > £;*.») = £ wy*V>
T(r, <;!,») = ^ tyiV,
where the modified rank polynomial R is related to the usual one (Defi-
nition 10.1) by R(T;x,y) = a
Theorem 13.9 Let T be a connected graph with n vertices and let <
be any ordering of ET. Then the Tutte polynomial is related to the rank
polynomial as follows:
T(r,<;x + l,y + l) = R(r;x,y) = xn-lR(T;x-\y).
Proof We shall use the intermediate polynomial P defined above, and
prove the equalities
T{T, <;x+l,y + l) = P(T, <;x,y+l) = R(T; x,y),
which are equivalent to the following relationships among the coeffi-
102 Colouring problems

cients:

x 7
k i
For the first identity, consider A : C -* V. By Proposition 13.7, if T is
in V, then
T = WX if and only if T~ CWCT.
Also, by Proposition 13.8, the external activities of T and W are the
same. Consequently, for each one of the tkj spanning trees T with |X'| =
k and \X^\ = j there are (*) subgraphs W in C with r(W) = r0 -i and
\We\ — j . These subgraphs are obtained by removing from T any set of
i edges contained in the k internally active edges of T. This proves the
first identity.
For the second identity, we consider /i : A -+ C. By the analogue of
Proposition 13.7 for /x, if X is in C, then
X^Y* if and only if XCYCX+.
Consequently, for each one of the iru subgraphs X in C with r(X) =
ro — i and \Xe\ = I, there are ('.) subgraphs Y with r(Y) = r 0 - i and
s(Y) = j . These subgraphs are obtained by adding to X any set of j
edges contained in the I externally active edges of X. This proves the
second identity. •

Corollary 13.10 The Tutte polynomial of a connected graph T is

independent of the ordering used in its definition.
Proof This statement follows from Theorem 13.9 and the fact that
the rank polynomial is independent of the ordering. •

The original proof of Theorem 13.9 by Tutte (1954) was inductive; the
proof given above is a simplification of the first constructive proof by
Crapo (1969). In the light of the Corollary we can write T(T;x,y) for
the Tutte polynomial of T. It should be noted that, although each
coefficient Uj is independent of the ordering, the corresponding set of
spanning trees (having internal activity i and external activity j) does
depend on the ordering.

Additional Results

13a Tutte polynomials of cycles By listing the spanning trees of Cn

and calculating their internal and external activities we obtain
T 1 //"* \ i i 2 i t n —1
 The Tutte polynomial 103

13b The Tutte matrix of Petersen's graph (Biggs 1973b) The matrix
(Uj) of coefficients of the Tutte polynomial for Petersen's graph is
r 0 36 84 75 35 9 1
36 168 171 65 10
120 240 105 15
180 170 30
170 70
114 12
56
21
6
L l

13c The deletion-contraction property The following two properties

completely define the Tutte polynomial for connected graphs.
(1) If e is an edge of the connected graph F which is neither a loop
nor an isthmus, then T(T;x,y) = T(r(e);;c,y) + T(r ( e ) ;x,y).
(2) If Ajj is formed from a tree with i edges by adding j loops,
T(AiJ',x,y)=xiy*.
13d Recursive families (Biggs, Damerell and Sands 1972) Using the
deletion-contraction property we can obtain a second-order recurrence
for the Tutte polynomials of the cycle graphs:
T(Cn+2;x,y) - (x + l)T(Cn+1;x,y) + xT(Cn;x,y) = 0.
Generally, a family {Ti} of graphs is said to be a recursive family if there
is a linear recurrence of the form
T(Ti+p; x, y) + aiT{Yl+p^ x,y) + ...+ o p T(r,; x, y) = 0,
where the coefficients a j , . . . , ap are polynomial functions of (x, y), and
are independent of /. Thus the cycle graphs form a recursive family
with p = 2. The families {Lh}, {Mh} of ladders and Mobius ladders
are recursive families with p — 6; they have the same recurrence, whose
auxiliary equation is
(t - l)(t - x)(t2 -(x + y + 2)t + xy){t2 ~ (x2 + x + y+ l)t + x2y) = 0.
From this we can deduce the tree-numbers and the chromatic polyno-
mials for these graphs. (See also 9c.)

13e Tutte polynomials of complete graphs Let T(X, y, a) and p(y, a)

be the exponential generating functions for the polynomials T{Kn\ x, y)
and y(2'(y — l)~n respectively. Then
r(x,y,a) =
x-1
104 Colouring problems

13f Inversions of trees A labelled tree on n vertices is a spanning tree

A of Kn with the vertex-set {1,2,..., n}. Let inv(A) denote the number
of edges {i,j} of A for which i < j and j is on the path in A from 1 to
i. Then we have

A
where the sum is over all labelled trees on n vertices.

13g The commutative diagram If X C ET, define

T = X» U (Xx \ X) = Xx \ (X \ X").
Then Xx» = T = X^x (Crapo 1969).

13h Counting forests If we write T(T; 1,1 + t) = £&<*, then fa is

the number of forests in V which have | ^ r | — i — 1 edges. It follows
that T(T; 1,2) is the total number of forests in T, and T(T; 1,1) is the
tree-number of T.

13i Planar graphs If T and T* are dual planar graphs, then there is
a bijective correspondence between their spanning trees which switches
internal and external activity. It follows that t%j = t*^ and
T{Y;x,y)=T{T*;y,x).

13j The medial graph Let F be a connected graph which is embed-

ded in the plane. For each e € E(T) choose an interior point m(e)
on e. The medial graph M(T) associated with the given embedding of
F has vertex-set {m(e) | e € E(T)} and edge-set defined as follows.
For each face of the embedded graph T there is a cycle with edges
ei,e2,. •. ,ek bounding that face; we create a corresponding sequence
f^i, fi2, • •.,Hk of edges of M(T) which (i) forms a cycle in M(T) with
vertices m(ei),m(e 2 ),... ,m(efc) and (ii) is topologically identical with
the original cycle. M(T) is a 4-regular graph and as such it has at least
one Eulerian partition; that is, a partition of its edge-set into cycles with-
out repeated edges. Let fk denote the number of Eulerian partitions of
M(G) into k cycles such that, at any any vertex of M(G), the two cy-
cles passing through that vertex do not cross, in the obvious topological
sense. Las Vergnas (1978) proved that

fc>0
See also Jaeger (1988) and Las Vergnas (1988).
 The Tutte polynomial 105

13k Tutte polynomials for knots and links (Thistlethwaite 1987) A

knot or link L is usually represented by a diagram in the plane; the
diagram is said to be alternating if the 'crossings' are alternately over and
under as we traverse each component. Associated with an alternating
diagram is a graph DL such that the Jones polynomial of L is given by
VL{t) = {-t)-KT{DL;-t,-rl\
where K is a number depending on L.
This relationship leads to a simple proof of a conjecture made by Tait
in the 19th century: the number of crossings in any alternating diagram
of a given link is invariant, provided there are no 'nugatory' crossings.

131 Intractability of calculating the Tutte polynomial A counting prob-

lem is said to be #P-hard if it has a certain technical property which,
it is believed, is equivalent to computational intractability. Jaeger, Ver-
tigan and Welsh (1990) showed that computing T(T;x,y) is #P-hard,
except for a few points and curves in the complex {x, y)-plane. In par-
ticular, computing the Jones polynomial (13k) of an alternating link is
#P-hard.
 14

Chromatic polynomials and spanning trees

In this chapter we shall study the relationship between the Tutte poly-
nomial and the chromatic polynomial of a connected graph. The main
result is as follows.
Theorem 14.1 Let F be a connected graph with n vertices. Then
n-l
1
C(T; u) = (-l)"- ^ ]T *«(1 - ")*'
t=i
where £JO is the number of spanning trees ofT which have internal activity
i and external activity zero (with respect to any fixed ordering of EY).
Proof We have only to invoke some identities derived in earlier chap-
ters. The chromatic polynomial is related to the rank polynomial as in
Corollary 10.5, and the rank polynomial is related to the Tutte polyno-
mial as in Theorem 13.9. Thus we have
C(T;u) =unR(F;-u-1,-l)

The result follows from the definition of the Tutte polynomial. •

This theorem indicates a purely algebraic way of calculating chro-

matic polynomials. If we are given the incidence matrix of a graph F,
then the basic cycles and cuts associated with each spanning tree T of
F can be found by matrix operations, as explained in Chapter 5. Prom
this information we can compute the internal and external activities of
 Chromatic polynomials and spanning trees 107

T, using the results of Proposition 13.5. The method is impracticable

for hand calculation, but it is well-adapted to automatic computation in
view of the availability of sophisticated programs for carrying out ma-
trix algebra. Furthermore, it is demonstrably better than the deletion-
contraction method (see 14h).

Theorem 14.1 also has theoretical implications for the study of chro-
matic polynomials, and the remainder of this chapter is devoted to some
of these consequences. First, we observe that if the chromatic polyno-
mial is expressed in the 'reduced form'
n-2
C(F; u) = ±w(w — 1) V J diW1 where w = 1 — u,
i=0
then the coefficients a, are all non-negative. In fact a, is the number
U+ifi- It is convenient to use the reduced form to record chromatic
polynomials, because the coefficients have fixed sign and are relatively
small.

Proposition 14.2 Let T be a connected graph, and let (£„•) denote

the matrix of coefficients of its Tutte polynomial. Then

Proof Suppose that the ordering of ET = {ei, e 2 , . . . , e m } is the nat-

ural order of the subscripts. If T is a spanning tree with internal activity
1 and external activity 0, then ei must be an edge of T, otherwise it
would be externally active. Further, e% is not an edge of T, otherwise
both ei and t2 would be internally active. Also e\ is in cyc(T, e2), oth-
erwise e2 would be externally active. Consequently T« = (T \ e{) U e% is
a spanning tree, with internal activity 0 and external activity 1.
Reversing the argument shows that T >-+ T* is a bijection, and hence
tio (the number of spanning trees T with |T"i = 1 and \T€\ = 0) is equal
to toi (the number of spanning trees T* with \T^\ = 0 and |T*| = 1).
•
The number tw has appeared in the work of several authors, for exam-
ple Crapo (1967) and Essam (1971). We note that it is the coefficient
ao in the reduced form of the chromatic polynomial. It is sufficiently
important to warrant a name.
Definition 14.3 The chromatic invariant 9(T) of a connected graph
F is the number of spanning trees of T which have internal activity 1
and external activity 0.
108 Colouring problems

Theorem 14.1 provides another interpretation of 8(T) in terms of the

chromatic polynomial of F. Let C denote the derivative of C; then a
simple calculation shows that

When F is non-separable it has at least one spanning tree with internal

activity 1 and external activity 0 (14b). Thus, for a non-separable graph
with an even number of vertices C is increasing at its zero u = 1, whereas
if the graph has an odd number of vertices it is decreasing.
The link with the chromatic polynomial can also be used to justify
the use of the name 'invariant' for 0(F). Recall that two graphs are said
to be homeomorphic if they can both be obtained from the same graph
by inserting extra vertices of degree two in its edges.
Proposition 14.4 If I \ and F2 are homeomorphic connected graphs
with at least two edges, then
0(1^) = 0(T2).
Proof Let F be a graph which has at least three edges, and a vertex
of degree two. Let e and / be the edges incident with this vertex. The
deletion of either e or / , say e, results in a graph r( e ) in which the
edge / is attached at a cut-vertex to a graph To with at least one edge.
Hence C(r^;u) is of the form (u - 1)C(TO; u), where C(F 0 ; 1) = 0. The
contraction of e in F results in a graph homeomorphic with F. We have

= («-i)C(r o ;u)-c(r (e); u),

and, on differentiating and putting u — 1, we find
C"(r;l) = - C " ( r ( e ) ; l ) .
Since F has one more vertex than F(e), it follows that

Now, if two graphs are homeomorphic, then they are related to some
graph by a sequence of operations like that by which F(e) was obtained
from F; hence we have the result. •

It is worth remarking that both the proof and the result fail in the case
where one of the graphs is K2; we have 6{K2) = 1, whereas any path
graph Pn (n > 3) is homeomorphic with K2 but 0(Pn) = 0.
We end this chapter with an application of Theorem 14.1 to the uni-
modal conjecture of Read (1968). This is the conjecture that if
;u) = un - Clu
n1
' + ... + {-l)n-lcn_lU,
 Chromatic polynomials and spanning trees 109

then, for some number M in the range 1 < M < n — 1, we have

Cl < C2 < . . . < CM > CM+l > • • • > Cn-l-

There is strong numerical evidence to support this conjecture, but a

proof seems surprisingly elusive. The following partial result was ob-
tained by Heron (1972).
Proposition 14.5 Using the above notation for the chromatic poly-
nomial of a connected graph F with n vertices, we have
Ci-i < c i for all i < -{n- 1).
Proof The result of Theorem 14.1 leads to the following expression
for the coefficients of the chromatic polynomial:
^ /n-l-A ^ (n~l-l\
U \n-l-tj ^ \ *-l J
Now if i<\{n- 1), then i - I < \{n - 1 - /) for all I > 0. Hence, by
the unimodal property of the binomial coefficients, we have

Thus, since each number £n_i_j)o is a non-negative integer, it follows

that Ci > c,_i for i < | ( n — 1), as required.

Additional Results

14a A product formula for 8 If F has a quasi-separation (Vi, V2) with

jVx 0 Vaj = t t h e n

This formula is particularly useful when t — 2.

14b Graphs with a given value of 9 A connected graph F is separable

if and only if 9(T) = 0. It is a series-parallel graph if and only if
0(F) < 1 (Brylawski 1971). One graph with 0 = 2 is #4, and it follows
from Brylawski's result on series-parallel graphs that if F contains a
subgraph homeomorphic to K4 then 0(F) > 2. In order to show that
all values of 6 can occur we need only remark that for the wheel Wn we
have 0(Wn) = n — 2. Using the product formula 14a we can construct
infinitely many graphs with any given value of 9, by gluing any edge of
any series-parallel graph to any edge of the appropriate wheel.
110 Colouring problems

14c The chromatic invariants of dual graphs Let F and F* be dual

planar connected graphs. Then
0(r) = 0(r*).
For instance,
0(Q3) = 0(#2,2,2) = 11; O(Icosahedron) = 6(Dodecahedron) = 4412.

14d Some explicit formulae For the complete graphs ifn, the ladders
Lh, and the Mobius ladders Mh, we have
0(Kn) = (n - 2)! (n > 2),
0(Lh) = 2fe - ft - 1 (ft > 3),
0(Mh) = 2h - ft (ft > 2).

14e Tfte /Zoiu polynomial Let C*(F; w) denote the number of nowhere-
zero u-flows (see 4k) on a connected graph F with n vertices and m
edges. Then
C*(F; u) = {-l)mR(T; - 1 , -u) = ( - l ) m - n + 1 T ( F ; 0,1 - «).
If F is planar and F* is its dual, then (Tutte 1954):
C(T*;u)=uC*(r;u).
Thus the problem of finding the flow polynomial of a planar graph is
equivalent tofindingthe chromatic polynomial of its dual. For example,
theflowpolynomial of a ladder graph can be derived from the chromatic
polynomial of its dual, a double pyramid (9a).
The general relationship between theflowpolynomial of a graph and
an 'interaction model' is discussed in Biggs (1977b, Chapter 3).

14f Theflowpolynomials of Kz,z and O3. From the rank matrix of

^3,3 (Chapter 10) and the Tutte matrix of O3 (13b) we can obtain the
flow polynomials for these (non-planar) graphs:
C*(K3,s; «) = ( « - l)(u - 2)(M2 - %u + 10);
C*(O3;u) = {u- l)(u - 2)(it - 3)(« - 4)(u2 -5u + 10).
In both cases there is no graph whose chromatic polynomial is uC*.

14g Expansions of the flow polynomial Jaeger (1991) obtained an ex-

pansion of the flow polynomial of a graph F of degree 3 imbedded in
the plane. Define an even subgraph to be a subgraph (C) in which ev-
ery vertex has even degree. Since T has degree 3 this means that every
component of (C) is a cycle, and so each component can be oriented in
 Chromatic polynomials and spanning trees 111

one of two ways. Associated with every oriented even subgraph (C) is a
weight w(C), such that
C*(F; (u + y - 1 ) 2 ) = ] £ ( V ,
c
where p(C) is a 'rotation number' depending on the relative orientation
of the cycles of (C) with respect to the plane in which F is embedded.
14h The superiority of the matrix method It follows from the result of
Jaeger, Vertigan and Welsh (131) that computing the chromatic polyno-
mial is, in general, #P-hard. However, there is some interest in compar-
ing methods of computation, even though they are all 'bad' in theoretical
terms.
The matrix method (call it Method A), described in our comments on
Theorem 14.1, has been used only rarely (Biggs 1973b). However, An-
thony (1990) showed that it is more efficient than the method of deletion
and contraction (Method B), even when that method incorporates rules
for curtailing the computation. Specifically, the worst-case running time
of Method A for a graph with n vertices and m edges is of the order of
( n ™i) n2m - ^ T_A(TI) and Te{n) denote the worst-case running times of
the respective methods for any sequence of graphs such that F n has n
vertices and the average degree A(n) —» oo as n —> oo, we have
log TB(n)
• oo as n —> oo.
log TA{n)
 PART THREE

Symmetry and regularity

 15
Automorphisms of graphs

An automorphism of a (simple) graph F is a permutation n of VT which

has the property that {u, v} is an edge of F if and only if {ir(u),Tr(v)}
is an edge of F. The set of all automorphisms of F, with the operation
of composition, is the automorphism group of F, denoted by Aut(F).
Some basic properties of automorphisms are direct consequences of
the definitions. For example, if two vertices x and y belong to the same
orbit, that is, if there is an automorphism a such that a(x) = y, then
x and y have the same degree. This, and other similar results, will be
taken for granted in our exposition.
We say that F is vertex-transitive if Aut(F) acts transitively on VT,
that is, if there is just one orbit. This means that given any two vertices
u and v there is an automorphism TT € Aut(F) such that TT(U) = v.
The action of Aut(F) on VT induces an action on ET, by the rule
n{x, y} = {n(x),7r(y)}, and we say that F is edge-transitive if this action
is transitive, in other words, if given any pair of edges there is an auto-
morphism which transforms one into the other. It is easy to construct
graphs which are vertex-transitive but not edge-transitive: the ladder
graph L3 is a simple example. In the opposite direction we have the
following result.
Proposition 15.1 If a connected graph is edge-transitive but not
vertex-transitive, then it is bipartite.
Proof Let {x, y} be an edge of F, and let X and Y denote the orbits
containing x and y respectively under the action of Aut(F) on the ver-
tices. It follows from the definition of an orbit that X and Y are either
116 Symmetry and regularity

disjoint or identical. Since F is connected, every vertex z is in some

edge {z,w}, and since F is edge-transitive, z belongs to either X or Y.
Thus XUY = VT. If X = Y = VT then F would be vertex-transitive,
contrary to hypothesis; consequently X n Y is empty. Every edge of F
has one end in X and one end in Y, so F is bipartite. D

The complete bipartite graph Ka^ with a ^ b is an obvious example of

a graph which is edge-transitive but not vertex-transitive. In this case
the graph is not regular, and it is not vertex-transitive for that reason,
because it is clear that in a vertex-transitive graph each vertex must have
the same degree. Examples of regular graphs which are edge-transitive
but not vertex-transitive are not quite so obvious, but examples are
known (see 15c).
The next proposition establishes a link between the spectrum of a
graph and its automorphism group. We shall suppose that VT is the
set {vi,i>2, • • • ,vn}, and that the rows and columns of the adjacency
matrix of F are labelled in the usual way. A permutation -K of VT can
be represented by a "permutation matrix P = (Pij), where Pij = 1 if
Vi = IT(VJ), and Pij = 0 otherwise.
Proposition 15.2 Let A be the adjacency matrix of a graph F, and
ft a permutation of VT. Then TT is an automorphism ofT if and only if
PA = AP, where P is the permutation matrix representing n.
Proof Let Vh — TT(VJ) and Vk = K(VJ). Then we have

(AP) hj = ZciMpij = ahk.

Consequently, AP = PA if and only if Uj and VJ are adjacent whenever
Vh and Vk are adjacent; that is, if and only if TT is an automorphism of
r. D

A consequence of this result is that, loosely speaking, automorphisms

produce multiple eigenvectors corresponding to a given eigenvalue. To be
precise, suppose x is an eigenvector of A corresponding to the eigenvalue
A. Then we have
APx = PAx = PAx = APx.
This means that Px is also an eigenvector of A corresponding to the
eigenvalue A. If x and Px are linearly independent we conclude that
A is not a simple eigenvalue. The following results provide a complete
description of what happens when A is simple.
 Automorphisms of graphs 117

Lemma 15.3 Let A be an simple eigenvalue ofT, and let x be a cor-

responding eigenvector, with real components. If the permutation matrix
P represents an automorphism of F then Px = ±x.
Proof If A has multiplicity one, x and Px are linearly dependent;
that is Px = [j,x for some complex number ft. Since x and P are real,
/i is real; and, since P" = I for some natural number s > 1, it follows
that p is an sth root of unity. Consequently /z = ±1 and the lemma is
proved. •

Theorem 15.4 (Mowshowitz 1969, Petersdorf and Sachs 1969) //

all the eigenvalues of the graph F are simple, every automorphism of F
(apart from the identity) has order 2.
Proof Suppose that every eigenvalue of F has multiplicity one. Then,
for any permutation matrix P representing an automorphism of F, and
any eigenvector x, we have P 2 x = x. The space spanned by the eigen-
vectors is the whole space of column vectors, and so P 2 = I. D

Theorem 15.4 characterizes the group of a graph which has the maximum
number n = |VT| of distinct eigenvalues: every element of the group
is an involution, and so the group is an elementary abelian 2-group.
For example, the theta graph ©2,2,1 (K4, with one edge deleted) has
automorphism group Z2 x Z2. The characteristic polynomial is

and so every eigenvalue is simple. On the other hand if we know that

a graph has an automorphism of order at least three, then it must have
a multiple eigenvalue. In particular, this means that the 2/i numbers
obtained in 3e as the eigenvalues of the Mobius ladder M2/1 cannot all
be distinct.

The question of which groups can be the automorphism group of some

graph was answered by Frucht (1938). He showed that, for every ab-
stract finite group G, there is a graph F whose automorphism group is
isomorphic to G. He also proved that the same result holds with F re-
stricted to be a regular graph of degree 3 (Frucht 1949). Although there
are some gaps in the original proof, satisfactory proofs of the result are
now available. For an overview of this subject the reader is referred to
Babai (1981). He describes how Frucht's work stimulated a great deal
of research, and how it has been extended by several authors to show
that the conclusion remains true, even if we specify in advance that F
must satisfy a number of graph-theoretical conditions.
118 Symmetry and regularity

If we strengthen the question by asking whether every group of per-

mutations of a set X is the automorphism group of some graph with
vertex-set X, then the answer is negative. For example, the cyclic
permutation-group of order 3 is not the automorphism group of any
graph with three vertices. (It is, of course, a subgroup of the group of
K3.) This tends to confirm our intuitive impression that there must be
some constraints upon the possible symmetry of graphs. One such con-
straint is the following. If F is a connected graph and d(u, v) denotes the
distance in F between the vertices u and v then, for any automorphism
a we have
d{u,v) = d{a{u),a{v)).
Thus there can be no automorphism which transforms a pair of vertices
at distance r into a pair at distance s ^ r. The following definition
frames conditions which are, in a sense, partially converse to this result.
Definition 15.5 Let F be a graph with automorphism group Aut(F).
We say that F is symmetric if, for all vertices u,v,x,y of F such that u
and v are adjacent, and x and y are adjacent, there is an automorphism
a in Aut(F) for which a(u) = x and a(v) = y. We say that F is distance-
transitive if, for all vertices u,v,x,y of F such that d(u, v) — d(x, y), there
is an automorphism a in Aut(F) satisfying a(u) = x and a(v) — y.

It is clear that we have a hierarchy of conditions:

distance-transitive =>• symmetric => vertex-transitive.

In the following chapters we shall investigate these conditions in turn,

beginning with the weakest one.

Additional Results

15a How large can an automorphism group be? For any value of n,
the automorphism group of the complete graph Kn contains all the n!
permutations of its n vertices: it is the symmetric group Sn- Any other
graph on n vertices has an automorphism group which is a subgroup of
Sn. Since the complete graph is the only connected graph in which each
pair of distinct vertices is at the same distance, it is the only connected
graph for which the automorphism group can act doubly-transitively on
the vertex-set.
 Automorphisms of graphs 119

15b How small can an automorphism group be? Except for very small
values of n, it is easy to construct a graph with n vertices which has the
trivial automorphism group, containing only the identity permutation.
For n > 7 the tree with n vertices shown in Figure 7 is an example.

Figure 7: a tree with no non-trivial automorphisms.

In fact, almost all graphs have the trivial automorphism group. The full
story is described by Bollobas (1985, Chapter 9).

15c A regular graph which is edge-transitive but not vertex-transitive

Consider a cube divided into 27 equal cubes, in the manner of Rubik's
cube, and let us say that a row is a set of three cubes in a row parallel
to a side of the big cube. Define a graph whose vertices are the 27 cubes
and the 27 rows, a cube-vertex being adjacent to the three row-vertices
to which it belongs. This example of a regular edge-transitive graph
which is not vertex-transitive is the first of a family of examples due to
Bouwer (1972).

15d The automorphism groups of trees (Jordan 1869) Let T be a finite

tree. Then either (i) T has a vertex v, known as the centroid, which is
fixed by every automorphism of T, or (ii) T has an edge {x, t/}, known
as the bicentroid, which is fixed (setwise) by every automorphism of T.

15e The graphs P{h,t) The 'generalized Petersen graph' P(ft,£) is

a 3-regular graph with 2h vertices xo, Xi,..., XH-U Vo, 2/i, • • •, Vh-i and
edges {xi, y<}, {xi, xi+i}, {yi, y%+th f o r a11 * e {0,1,..., /i » 1}, where
the subscripts are reduced modulo h. For example, P(h, 1) is the ladder
graph Lh, and P(5,2) is Petersen's graph. Frucht, Graver and Watkins
(1971) showed that:
(a) P(ft, t) is vertex-transitive if and only if t2 = ±1 (mod ft), or (ft, t) =
(10,2).
(b) P(h,t) is symmetric if and only if (ft, t) is one of (4,1), (5,2), (8,3),
(10,2), (10,3), (12,5), (24,5).
120 Symmetry and regularity

Case-by-case checking of the latter result shows that P(h, t) is distance-

transitive if and only if (h,t) is one of (4,1), (5,2), (10,3).

15f The connection between Aut(r) and Aut(£(F)) (Whitney 1932c)

The automorphism groups of F and its line graph L(F) are not necessar-
ily isomorphic: for example Ki = L(K2), so in this case the first group
is trivial but the second is not. However, this is a rare phenomenon.
There is a group homomorphism 0 : Aut(r) —> Aut(Z,(F)) defined by
{9g){u, v} = {0u,Ov) where g € Aut(r), {u, v} € E{T),
and we have: (i) 6 is a monomorphism provided F ^ K2; (ii) 0 is an
epimorphism provided F is not K4, K4 with one edge deleted, or K4
with two adjacent edges deleted.

15g Homogeneous graphs A graph F is said to be weakly homogeneous

if, whenever two subsets U\, U2 of VF are such that (Ui) and (f/jj) are
isomorphic, then at least one isomorphism between them extends to an
automorphism of F. The complete list of weakly homogeneous graphs is
as follows.
(a) The cycle graph C 5 .
(b) The disjoint union of t > 1 copies of the complete graph Kn.
(c) The complete multipartite graphs iiTSjS>...,s with t > 2 parts of
equal size s.
(d) The line graph L(K3>3).
A graph is homogeneous if, whenever two subsets U\, U2 of VT are such
that (Ui) and (U2) are isomorphic, then every isomorphism between
them extends to an automorphism of F. It is obvious that a homo-
geneous graph is weakly homogeneous, and somewhat surprisingly the
converse is also true. This result has a contorted history. The 1974
version of this book caused some confusion by attributing to Sheehan
the classification of weakly homogeneous graphs given above. In fact,
Sheehan (1974) obtained the classification of homogeneous graphs. Gar-
diner observed the error in the book, and then (1976) obtained the same
list for the weakly homogeneous case by an independent method. Fi-
nally Ronse (1978) showed directly that a weakly homogeneous graph is
homogeneous.

15h Graphs which are transitive on vertices and edges Let F be a graph
for which Aut(F) acts transitively on both vertices and edges. Then F
is a regular graph, and if its degree is odd, it is symmetric (Tutte 1966).
If its degree is even, the conclusion may be false, as was first shown
 Automorphisms of graphs 121

by Bouwer (1970). Holt (1981) gave an example of a 4-regular graph

with 27 vertices which is vertex-transitive and edge-transitive, but not
symmetric, and Alspach, Marusic and Nowitz (1993) showed that Holt's
example is the smallest possible.

15i Graphs with a given group (Izbicki 1960) Let an abstract finite
group G and natural numbers r and s satisfying r > 3, 2 < s < r be
given. Then there are infinitely many graphs F with the properties:
(a) Aut(F) is isomorphic to G;
(b) F is regular of degree r;
(c) the chromatic number of F is s.
 16

Vertex-transitive graphs

In this chapter we study graphs F for which the automorphism group acts
transitively on VT. As we have already noted in the previous chapter,
vertex-transitivity implies that every vertex has the same degree, so F
is a regular graph.
We shall use thefollowingstandard results on transitive permutation
groups. Let G — Aut(F) and let Gv denote the stabilizer subgroup for
the vertex v; that is, the subgroup of G containing those automorphisms
which fix v. In the vertex-transitive case all stabilizer subgroups Gv (v 6
VF) are conjugate in G, and consequently isomorphic. The index of Gv
in G is given by the equation
\G : Gv\ = \G\/\GV\ = |VT|.
If each stabilizer Gv is the identity group, then every element of G
(except the identity) does not fix any vertex, and we say that G acts
regularly on VT. In this case the order of G is equal to the number of
vertices.
There is a standard construction, due originally to Cayley (1878),
which enables us to construct many, but not all, vertex-transitive graphs.
We shall give a streamlined version which has proved to be well-adapted
to the needs of algebraic graph theory. Let G be any abstract finite
group, with identity 1; and suppose ft is a set of generators for G, with
the properties:

(i) x g Q => re"1 G fi; (ii) 1 £ fi.

 Vertex-transitive graphs 123

Definition 16.1 The Cayley graph T = F(G, £2) is the simple graph
whose vertex-set and edge-set are defined as follows:
VT = G; EY = {{g,h}\g-lh€tt}.
Simple verifications show that ET is well-defined, and that T(G, £2) is
a connected graph. For example, if G is the symmetric group S3, and
£2 = {(12), (23), (13)}, then the Cayley graph T{G,Q) is isomorphic to
#3,3 (Figure 8).
l, ,(12)

(123)

(132)' '(23)
Figure 8: K3t3 as a Cayley graph for 53.

Proposition 16.2 (1) The Cayley graph T(G, £2) is vertex-transitive.

(2) Suppose that ir is an automorphism of the group G such that TT(£2) =
£2. Then n, regarded as a permutation of the vertices ofT(G,Cl), is a
graph automorphism fixing the vertex 1.
Proof (1) For each g in G we may define a permutation ~g of VT = G
by the rule g(h) = gh (h € G). This permutation is an automorphism
of T, for
{h,k}GET=>h~lkefl
=* (gh)~lgk € Q

The set of all g (g € G) constitutes a group G (isomorphic with G),

which is a subgroup of the full group of automorphisms of F(G, £2), and
acts transitively on the vertices.
(2) Since IT is a group automorphism it must fix the vertex 1. Fur-
thermore, n is a graph automorphism, since
{h, jfc} € J5r => /i-1fc € £2 => Ttih^k) € £2

a
The second part of this proposition implies that the automorphism
group of a Cayley graph F(G, 0) will often be strictly larger than G.
124 Symmetry and regularity

In the example illustrated in Figure 8, every group automorphism of

S3 fixes fl setwise, and so it follows that the stabilizer of the vertex
1 has order at least 6. In fact, the order of the stabilizer is 12, and
|Aut(K 3 , 3 )| = 72.
Not every vertex-transitive graph is a Cayley graph; for example Pe-
tersen's graph O3 is not a Cayley graph. This statement can be checked
by noting that there are only two groups of order 10 and they have
few generating sets of size three satisfying the conditions in Definition
16.1. An exhaustive check of all the possibilities confirms that Petersen's
graph does not arise as a Cayley graph in this way.
We begin our study of the hierarchy of symmetry conditions with the
case when Aut(F) acts regularly on V(T).
Lemma 16.3 Let F be a connected graph. Then a subgroup H of
Aut(F) acts regularly on the vertices if and only if F is isomorphic to a
Cayley graph T(H, Q), for some set Q which generates H.
Proof Suppose VT = {v\, v 2 , . . . , vn}, and H is a subgroup of Aut(F)
acting regularly on VT. Then, for 1 < i < n there is a unique hi € H
such that hi(vi) = i>j. Let
Cl = {hi € H I Vi is adjacent to vi in F}.
Simple checks show that Q satisfies the two conditions required by Def-
inition 16.1 and that the bijection Vi <-> hi is a graph isomorphism of F
with T(H,Q). Conversely if F = F(#,Q) then the groupjf? defined in
the proof of Proposition 16.2 acts regularly on VT, and H « H. •

Lemma 16.3 shows that if Aut(F) itself acts regularly on VT, then F
is a Cayley graph F(Aut(F), O).
Definition 16.4 A finite abstract group G admits a graphical regular
representation, or GRR, if there is a graph F such that G is isomorphic
with Aut(F), and Aut(F) acts regularly on VT.
The question of which abstract groups admit a GRR was answered
completely in the late 1970's (see 16g). It turns out that the second
part of Proposition 16.2 is essentially the only obstacle to there being a
GRR for G. In other words, a group G has no GRR if and only if every
generating set Cl for G which satisfies conditions (i) and (ii) is such that
there is an automorphism of G fixing Q setwise.
As an example of the ideas involved we show that the group S3 admits
no graphical regular representation. If there were a suitable graph F,
then it would be a Cayley graph F(5 3 , f2). Now, it is easy to check by an
 Vertex-transitive graphs 125

exhaustive search that, for any generating set 0 satisfying conditions (i)
and (ii) on p. 122, there is some automorphism of 53 fixing Q setwise.
Thus, by part (2) of Proposition 16.2, the automorphism group of a
Cayley graph T(S$, 0) is strictly larger than S3.
In the case of transitive abelian groups, precise information is provided
by the next proposition.
Proposition 16.5 Let F be a vertex-transitive graph whose automor-
phism group G = Aut(F) is abelian. Then G acts regularly on VT, and
G is an elementary abelian 2-group.
Proof If g and h are elements of the abelian group G, and g fixes v,
then gh(v) = hg(v) = h(v) so that g fixes h(v) also. If G is transitive,
every vertex is of the form h(v) for some h in G, so g fixes every vertex.
That is, g = 1.
Thus G acts regularly on VT and so, by Lemma 16.3, F is a Cayley
graph F(G, fi). Now since G is Abelian, the function g »-> g~x is an
automorphism of G, and it fixes Q setwise. If this automorphism were
non-trivial, then part (2) of Proposition 16.2 would imply that G is not
regular. Thus g = g"1 for all g € G, and every element of G has order
2. •
We now turn to a discussion of some simple spectral properties of
vertex-transitive graphs. A vertex-transitive graph F is necessarily a
regular graph, and so its spectrum has the properties which are stated
in Proposition 3.1. In particular, if F is connected and regular of degree
k, then k is a simple eigenvalue of F. It turns out that we can use the
vertex-transitivity property to characterize the simple eigenvalues of F.
Proposition 16.6 (Petersdorf and Sachs 1969) Let F be a vertex-
transitive graph which has degree k, and let X be a simple eigenvalue of
F. / / |VT| is odd, then X = k. If \VT\ is even, then X is one of the
integers 2a — k (0 < a < k).
Proof Let x be a real eigenvector corresponding to the simple eigen-
value A, and let P be a permutation matrix representing an automor-
phism "K of F. If ir(vi) = Vj, then, by Lemma 15.3,
Xi = (PX),- = ±Xj.
Since F is vertex-transitive, we deduce that all the entries of x have
the same absolute value. Now, since u = [ 1 , 1 . . . , 1]* is an eigenvector
corresponding to the eigenvalue k, if A ^ k we must have u*x = 0, that
is Yl xi = 0- This is impossible for an odd number of summands of equal
absolute value, and so our first statement is proved.
126 Symmetry and regularity

If F has an even number of vertices, choose a vertex Vi of T and suppose

that, of the vertices Vj adjacent to vu a number a have Xj = x% while
k — a have Xj = —X{. Since (Ax)i = Xxi it follows that Y^ xj = ^xi>
where the sum is taken over vertices adjacent to «*. Thus
axi — (k — a)xi = Xxi.
whence A = 2a — k. •

For example, the only numbers which can be simple eigenvalues of a 3-

regular vertex-transitive graph are 3,1, —1, —3. This statement is false if
we assume merely that the graph is regular of degree 3: many examples
can be found in [CvDS, pp. 292-305].
If we strengthen the assumptions by postulating that T is symmetric,
then the simple eigenvalues are restricted still further.

Proposition 16.7 Let T be a symmetric graph of degree k, and let A

be a simple eigenvalue ofT. Then A = ±k.
Proof We continue to use the notation of the previous proof. Let Vj
and vi be any two vertices adjacent to «»; then there is an automorphism
7T of F such that n(vi) = Vi and K(VJ) = uj. If P is the permutation
matrix representing n, then n(vi) = Vi implies that P x = x and so
Xj = x\. Thus a — 0 or k, and A = ±k.
We remark that the eigenvalue -k occurs, and is necessarily simple,
if and only if T is bipartite.
Additional Results

16a Circulant graphs A circulant graph is vertex-transitive, and a

connected circulant graph is a Cayley graph F(Z n , 0) for a cyclic group
Z n . Adam (1967) conjectured that if two such graphs r(Z n ,fl) and
F ( Z n , 0 ' ) are isomorphic then O' = zQ, for some invertible element z
in Z n . Elspas and Turner (1970) showed that the conjecture is true if
n is a prime, or if the graphs have only simple eigenvalues, but false
in general. Parsons (1980) showed that it is true if both graphs have
vertex-neighbourhoods isomorphic to the cycle C^.

16b The ladder graphs as Cayley graphs The dihedral group Z?2n of
order 2n is defined by the presentation

The Cayley graph of £>2n with respect to the generating set {x,x~~l,y}
is the ladder graph Ln.
 Vertex-transitive graphs 127

16c Cayley graphs for the tetrahedral and icosahedral groups The al-
ternating group An is the subgroup of index two in Sn containing all
the even permutations. The groups A4 and A5 are sometimes known as
the tetrahedral and icosahedral groups because they are isomorphic with
groups of rotations of the respective polyhedra. Both groups can be
represented by planar Cayley graphs. A Cayley graph for At is shown
in Figure 9.

Figure 9: a Cayley graph for A*.

A Cayley graph for A5 is the skeleton of the famous carbon-60 structure,

also known as buckminsterfullerene, or the 'buckie-ball', or the 'soccer
ball'.

16d The stabilizer of a vertex-neighbourhood Suppose that F is a

vertex-transitive graph with G = Aut(F). For any vertex v of F, de-
fine
Lv = {g e Gv I gfixeseach vertex adjacent to v}.
Then Lv is a normal subgroup of Gv. More explicitly, there is a homo-
morphism from Gv into the group of all permutations of the neighbours
of v, with kernel Lv. It follows from this that \GV : Lv\ < k\, where k is
the degree.

16e The order of the vertex-stabilizer Let Hn be the graph formed by

linking together n 'units' of the form shown in Figure 10 so that they
form a complete circuit. Then the graphs Hn are vertex-transitive and
the order of the vertex-stabilizer (2™) is not bounded in terms of the
degree.
On the other hand, in a symmetric graph the order of the vertex-
stabilizer is bounded in terms of the degree. See 17g.
128 Symmetry and regularity

Figure 10: the vertex-stabilizer is not bounded.

16f Coset graphs Let G be an abstract finite group, H a subgroup of G,

and fl a subset of G\H such that 1 $ O, fi"1 = fi, and if UQ generates
G. The simplest way of denning a (general) graph whose vertices are the
right cosets of H in G is to make Hg\ and Hgi adjacent whenever g29\X
is in fi. The graph so constructed is connected and vertex-transitive.
There are other ways of defining a graph whose vertices are cosets,
and some of them result in a symmetric graph. Examples and further
references may be found in a paper by Conder and Lorirfier (1989).

16g Graphical regular representations Hetzel (1976) proved that the

only solvable groups which have no GRR are:
(a) abelian groups of exponent greater than 2;
(b) generalized dicyclic groups;
(c) thirteen 'exceptional' groups, such as the elementary abelian groups
Z | , Z 2 , Z | , the dihedral groups De,Ds,Di0, and the alternating group
A,.
This work subsumed earlier results by several other authors. Godsil
(1981) showed that every non-solvable group has a GRR, so the list
given above is the complete list of groups which have no GRR.

16h The eigenvalues of a Cayley graph (Babai 1979) Let T(G, Q) be

a Cayley graph and suppose that the irreducible characters of G are
Xij X2> • • •. Xc with degrees m, n2,..., nc respectively. Then the eigen-
values of F fall into families (A*)j, 1 < i < c, 1 < j < n, such that
each (Ai)j contains rii eigenvalues, all with a common value Ajj. (Note
that the total number of eigenvalues is thus ^Znf, which is the correct
number \G\.) The sum of the tth powers of the Ay corresponding to a
given character \% satisfies

where the sum on the right-hand side is taken over all products of t
elements of 0.
 Vertex-transitive graphs 129

16i The Paley graphs Denote the additive group of the field GF(q)
by Gq and let fi be the set of non-zero squares in GF(q). If q = 1 (mod
4) then fi generates Gq and satisfies the conditions at the foot of p. 122
(remembering that the identity of Gq is the zero element of the field).
The Paley graph P(q) is the Cayley graph F(G 9 ,O). These graphs are
strongly regular and self-complementary. If q is the rth power of a prime,
the order of Aut(P(g)) is rq(q - l)/2.

16j Graphs with a specified vertex-neighbourhood A graph is said to be

locally K if, for each vertex v, the subgraph induced by the neighbours of
v is isomorphic to K. For example, the graphs which are locally Petersen
were determined by Hall (1980): there are just three of them, having
21, 63, and 65 vertices. Many other papers on this topic are listed by
Blokhuis and Brouwer (1992).

16k Generators for the automorphism group Let V be a connected

vertex-transitive graph and let Gv denote the stabilizer of the vertex
v. If h is any automorphism of T for which d(v,h(v)) = 1, and T is
symmetric, then h and Gv generate Aut(F).
 17

Symmetric graphs

The condition of vertex-transitivity is not a very powerful one, as is

demonstrated by the fact that we can construct at least one vertex-
transitive graph from each finite group, by means of the Cayley graph
construction. A vertex-transitive graph is symmetric if and only if each
vertex-stabilizer Gv acts transitively on the set of vertices adjacent to v.
For example, there are just two distinct 3-regular graphs with 6 vertices;
one is 1(3,3 and the other is the ladder L3. Both these graphs are vertex-
transitive, and -^3,3 is symmetric, but L3 is not because there are two
'kinds' of edges at each vertex.
Although the property of being symmetric is apparently only slightly
stronger than vertex-transitivity, symmetric graphs do have distinctive
properties which are not shared by all vertex-transitive graphs. This was
first demonstrated by Tutte (1947a) in the case of 3-regular graphs. More
recently his results have been extended to graphs of higher degree, and it
has become apparent that the results are closely related to fundamental
classification theorems in group theory. (See 17a, 17f, 17g.)
We begin by defining a t-arc [a] in a graph F to be a sequence
(ao, c*i,..., at) of t+ 1 vertices of F, with the properties that {a,_i, a,}
is in ET for 1 < i < t, and on-\ •=£ on+\ for 1 < i < t — 1. A t-arc is not
quite the same thing as the sequence of vertices underlying a path of
length t, because it is convenient to allow repeated vertices. We regard
a single vertex u a s a 0-arc [v]. If /? = (0o,fii,... ,/3s) is an s-arc in
F, then we write [a.(3} for the sequence (ao, • • ., at, /3o, • • •, Ps), provided
 Symmetric graphs 131

that this is a (t + s + l)-arc; that is, provided at is adjacent to /30 and

Qf-i ¥" 0o, at T^A-
Definition 17.1 A graph F is t-transitive (t > 1) if its automorphism
group is transitive on the set of i-arcs in F, but not transitive on the set
of (t + l)-arcs in F.
There is little risk of confusion with the concept of multiple transitivity
used in the general theory of permutation groups, since (as was noted in
15a) the only graphs which are multiply transitive in that sense are the
complete graphs. We observe that the automorphism group is transitive
on 1-arcs if and only if F is symmetric (since a 1-arc is just a pair of
adjacent vertices). Consequently, any symmetric graph is i-transitive
for some t > 1.
The only connected graph of degree one is K2, and this graph is 1-
transitive. The only connected graphs of degree two are the cycle graphs
Cn {n > 3), and these are anomalous in that they are transitive on t-
arcs for all t > 1. Prom now on, we shall usually assume that the graphs
under consideration are connected and regular of degree not less than
three. For such graphs we have the following elementary inequality.
Proposition 17.2 Let F be a t-transitive graph whose degree is at
least three and whose girth is g. Then

Proof F contains a cycle of length g, which is, in particular, a g-axc.

Because the degree is at least three, we can alter one edge of this g-avc
to obtain a <?-arc whose ends do not coincide. Clearly no automorphism
of F can take a g-axc of the first kind to a g-axc of the second kind; so
it follows that t < g.

Pg-t=O0-

[a]

Figure 11: illustrating the proof of Proposition 17.2.

Consequently, if we select a cycle of length g in F, then there is a t-arc

132 Symmetry and regularity

[a], without repeated vertices, contained in it. Let [/?] be the (g - t)-arc
beginning at at and ending at Qo which completes the cycle of length
g. Also let v be a vertex adjacent to at~i, but which is not at-2 or at\
this situation is depicted in Figure 11. Since F is ^-transitive, there is
an automorphism taking the t-arc [a] to the t-axc (ao,a\,..., Qt-i,v).
This automorphism must take the (g — t + l)-arc [at-i./3] to another
(g — t + l)-arc [at-1-7], where 70 = v and ^g-t — &o- The two arcs
{ott-i-fi] and [at-1.7] may overlap, but they define a cycle of length at
most 2(g-t + 1). Hence g < 2(g - t + 1), that is, g>2t-2. •

Definition 17.3 Let [a] and [/?] be any two s-arcs in a graph F. We
say that [0\ is a successor of [a] if /3< = aj+i ( 0 < i < s — 1).

It is helpful to think of the operation of taking a successor of [a] in

terms of 'shunting' [a] through one step in F. Suppose we ask whether
repeated shunting will transform a given s-arc into any other. If there
are vertices of degree one in F then our shunting might be halted in
a 'siding', while if all vertices have degree two we cannot reverse the
direction of our 'train'. However, if each vertex of F has degree not
less than three, and F is connected, then our intuition is correct and
the shunting procedure always works. The proof of this requires careful
examination of several cases, and may be found in Tutte's book (Tutte
1966, pp. 56-58). Formally the result is as follows.

Lemma 17.4 Let F be a connected graph in which the degree of each

vertex is at least three. If s > 1 and [a], [/?] are any two s-arcs in F,
then there is a finite sequence [ a ^ ] (1 < i < I) of s-arcs in F such
that [a*1*] = [a], [a (/) ] = [/?], and [a(i+1>] is a successor of [aW] for

We can now state and prove a convenient test for ^-transitivity. Let
F be a connected graph in which the degree of each vertex is at least
three, and let [a] be a i-arc in F.

Figure 12: a i-arc and its successors.

 Symmetric graphs 133

Suppose (as in Figure 12) that the vertices adjacent to at are at-i and
v ( 1 ) ,v< 2 >,...,v ( i ) , and let |/?W] denote the t-arc ( a i , a 2 , . . . ,at, v^) for
1 <i < I, so that each [0^] is a successor of [a].

T h e o r e m 17.5 Let T be a connected k-regular graph with I = k — 1 >

3, and let [a] be a t-arc in T. Then Aut(F) is transitive on t-arcs if
and only if it contains automorphisms gi,g%,..-,gi such that gi[a] =
[/3(i)] (1 < i < I).
Proof The condition is clearly satisfied if Aut(r) is transitive on t-
arcs. Conversely, suppose the relevant automorphisms <?i,<?2> • • • ,gi can
be found; then they generate a subgroup H = (gi, g2,..., gi) of Aut(F),
and we shall show that H is transitive on t-arcs.
Let [8] be a i-arc in the orbit of [a] under H; thus [8] —ft[a] for some
h e H. If [4>] is any successor of [8], then h~x[<j>] is a successor of [a],
and so [<j>] = hgt[a] for some i € { 1 , 2 , . . . , / } . That is, [<j>] is also in the
orbit of [a] under H. Now Lemma 17.4 tells us that all t-arcs can be
obtained from [a] by repeatedly taking successors, and so all t-arcs are
in the orbit of [a] under if. •

As an example, consider Petersen's graph O3, whose vertices are

the unordered pairs from the set {1,2,3,4,5} with disjoint pairs be-
ing adjacent. The automorphism group is the group of all permuta-
tions of {1,2,3,4,5}, acting in the obvious way on the vertices. Since
the girth of O3 is 5, Proposition 17.2 tells us that the graph is at
most 3-transitive. The 3-arc [a] — (12,34,15,23) has two successors:
[/3<x)] = (34,15,23,14) and [/3<2>] = (34,15,23,45). The automorphism
(13)(245) takes [a] to [/?(1)] and the automorphism (13524) takes [a] to
^ hence O3 is 3-transitive.
In addition to its usefulness as a test for t-transitivity, Theorem 17.5
also provides a starting point for theoretical investigations into the struc-
ture of t-transitive graphs. Suppose that T is a connected t-transitive
graph (t > 1), which is regular of degree fc > 3, and let [a] be a given
t-arc in F.
Definition 17.6 The stabilizer sequence of [a] is the sequence
Aut(r) = G > Ft > Ft-t > ...>Fi>F0
of subgroups of Aut(r), where F, (0 < i < t) is defined to be the
pointwise stabilizer of the set {QO, a i , . . . , ctt-i}-
In the case of Petersen's graph, with respect to the 3-arc (12,34,15,23),
the group FQ is trivial, Fi is the group of order 2 generated by (34), F2
134 Symmetry and regularity

is the group of order 4 generated by (34) and (12), and F 3 is the group
of order 12 generated by (34), (12) and (345).
In general, since G is transitive on s-arcs (1 < s < t), all stabilizer
sequences of £~arcs are conjugate in G, and consequently we shall often
omit explicit reference to [a].
The order of each group occuring in the stabilizer sequence is de-
termined by the order of FQ, as follows. Since Ft is the stabilizer of
the single vertex ao in the vertex-transitive group G, it follows that
\G : Ft\ = n = |VT|. Since G is transitive on 1-arcs, Ft acts transitively
on the k vertices adjacent to ao and Ft-\ is the stabilizer of the vertex
Qi in this action; consequently \Ft : Ft-i\ = k. Since G is transitive
on s-arcs (2 < s < t), the group Ft-s+i acts transitively on the k — 1
vertices adjacent to a s _i (other than a s _2), and F t _ s is the stabilizer
of the vertex as in this action; consequently |F<_S+1 : F t _ s | = k — 1 for
2 < s < t.
Thus we have

\G\=nk(k-l)t-1\F0\.
This confirms our earlier observations about Petersen's graph, where we
have t = 3 and |F 0 | = 1, so that |Fi| = 2, |F 2 | = 4, |F 3 | = 12 and
\G\ = 120.
We shall now explain how the properties of the stabilizer sequence can
be conveniently discussed in terms of the set {g\, gi,..., gi} of / = k — 1
automorphisms whose existence is guaranteed by Theorem 17.5. Define
an increasing sequence of subsets of G = Aut(F), denoted by {1} = YoC.
Yi C y 2 C . . . , as follows:
Yt = {g-'gl | a,b € {1,2,... ,1} and 1 < j < i}.

Proposition 17.7 (1) If 1 < i < t, then Yi is a subset of Fit but not
a subset o/Fj_i. (2) IfO<i< t, then Fj is the subgroup of G generated
by Yi and FQ.
Proof (1) For 1 < a < I, we have g^{<Uj) = Q-j+r provided that both
j and j + r lie between 0 and t. Also, <?£~-'+1(Q;j) — v^- I* follows
that gZ*§bfixesao, ot\,..., ctt-i for all j < i, and so Yi C F». If it were
true that Yi C Fj_i, then g^%g\ would fix at-i+i> but this means that
gla(at-i+i) — glb(at-i+i), that is v^ = v^bh Since this is false for a^b,
we have 1^ % Fj_i.
(2) Suppose / € Fi, and /[a] = ( a o , a i , • - • , a t - i , 7 i , • • • ,7»)- Pick any
 Symmetric graphs 135

gb; since 7! is adjacent to at-i, glili) is adjacent to gi(at-i) - at, and

so fl^i(7i) = « (o) for some a € {1,2,..., /}. Then
9al9lf\a\ = (ao,ai,...,at-.i+1,62,-..,6i) say.
By applying the same method with i replaced by i - 1, we can find
an automorphism gc{l~X) g1^1, which belongs to both Y^i and y», and
takes 62 to at-i+2 while fixing a o , a i , . . . , a t _ i + i . Continuing in this
way, we construct g in Yt such that gf[a] — [a], that is, gf is in Fo.
Consequently / is in the group generated by Yi and Fo. Conversely,
both Yi and Fo are contained in Fi so we have the result. •

All members of the sets YQ, Y\,..., Yt fix the vertex a 0 and so belong
to Ft, the stabilizer of ao; further, we have shown that Ft is generated
by Yt and Fo. In the case of Yj+i, we note that this set contains some
automorphisms not fixing ao, and we may ask whether Yj+i and Fo
suffice to generate the entire automorphism group G. The following
proposition shows that the answer is ' y e s \ unless the graph is bipartite.
The reason why bipartite graphs are exceptional in this respect is that
if F is a symmetric bipartite graph in which VT is partitioned into
two colour-classes V\ and V2, then the automorphisms which fix V\ and
V2 setwise form a subgroup of index two in Aut(F). We say that this
subgroup preserves the bipartition.
Proposition 17.8 Let T be a t-transitive graph with t > 2 and girth
greater than 3. Let G* denote the subgroup of G = Aut(F) generated by
Yt+i and Fo. Then either (1) G* = G; or (2) F is bipartite, \G:G*\ = 2,
and G* is the subgroup of G preserving the bipartition.
Proof Let u be any vertex of F such that d(u, ao) = 2; we show
first that there is some g* in G* taking ao to u. Since the girth of
F is greater than 3, the vertices w(a) = #a +1 ("o) and u(6) = g£ +1 (a 0 )
satisfy d{v^a\v^) = 2. Consequently, the distance between ao and
t+1 +1
ga^ ^gl (cto) is also 2. Now G* contains Ft (since the latter is gen-
erated by Yt, which is a subset of Yt+i, and Fo), and Ft is transitive
on the 2-arcs which begin at a 0 (since t > 2). Thus G* contains
an automorphism / fixing ao and taking ga + 9b+1(ao) to u, and
g* = fga 9b+1 takes ao to u.
Let U denote the orbit of ao under the action of G*. U contains all
vertices whose distance from a 0 is two, and consequently all vertices
whose distance from a 0 is even. If U = VT, then G* is transitive on
VT, and since it contains Ft, the stabilizer of the vertex ao in (G*) is
Ft. Thus |G*| = |VT||F t | = \G\, and so G* = G. If U =£ V, then U
136 Symmetry and regularity

consists precisely of those vertices whose distance from a^ is even, and

F is bipartite, with colour-classes U and VT \ U. Since G* fixes them
setwise, G* is the subgroup of G preserving the bipartition. •

We remark that the only connected graphs of girth three whose auto-
morphism group is transitive on 2-arcs are the complete graphs. Thus
the girth constraint in Proposition 17.8 is not very restrictive.
In the next chapter we shall specialize the results of Propositions
17.7 and 17.8 to 3-regular graphs; our results will lead to very precise
information about the stabilizer sequence.
Additional Results
17a The significance of the condition t > 2 In 16d we observed that
the vertex-stabilizer Gv has a normal subgroup Lv such that Gv/Lv is
a group of permutations of the vertices adjacent to D. In the case of
a symmetric graph with t > 2 this group of permutations is doubly-
transitive. Since all doubly-transitive permutation groups are 'known',
this observation links the problem of classifying symmetric graphs with
the classification theorems of group theory. See also 17f and 17g.

17b The stabilizer of an edge-neighbourhood Suppose that F is a sym-

metric graph of degree k with G — Aut(F). For any edge {v,w} of F,
define Gvw — GvnGw, Lvw = LVC\LW, where Lv and Lw are the stabi-
lizers of the respective vertex-neighbourhoods, as defined in 16d. Then
we have the following subgroup relationships among these groups.
(a) Lv is a normal subgroup of Gv and Gvw;
(b) Lvw is a normal subgroup of Lv and Gvw.
It follows from standard theorems of group theory that
Jjy LtyL/iU

jLjyyj J-JW

and LVLW/LW is a normal subgroup of Gvw/Lw. The last group is a

group of permutations of the neighbours of w, fixing v. Thus we have
\LV : Lvw\ < (k - 1)!, and
\GV\ < k\(k - 1)\\LVW\.

17c The full automorphism group of Kn^n It is clear that the graph
2
Kn^n has at least 2(n!) automorphisms. Simple arguments suffice to
show that there are no others, but, for the sake of example, we can use
17b. In this case the neighbourhood of an edge is the whole graph, so
Lvw = 1. It follows that
\G\ < 2n\Gv\ < 2nn\ (n - 1)! = 2 (n!)2.
 Symmetric graphs 137

17d The automorphism group of Ok A more substantial application

of 17b shows that the symmetric group S21C-1 is the full automorphism
group of Ok- When k > 3 every 3-arc in Ok determines a unique 6-cycle,
and it follows from this that if g e Lvw then g € Lwx for all vertices x
adjacent to w. Hence Lvw = 1, and the order of the full automorphism
group is at most

An alternative proof using the Erdos-Ko-Rado theorem may be found

in Biggs (1979).

17e The stabilizer sequence for odd graphs The odd graphs Ok are
3-transitive, for all k > 3. The stabilizer sequence is
G = S2k-u F3 = SkxSk-i, F2 = Sk-i x Sk-i,
Fi = S*_i x Sk-2, Fo = Sk-2 x Sk-2-

17f Lvw is a p-group (Gardiner 1973) For any t-transitive graph with
t > 2 the edge-neighbourhood stabilizer Lvw is a p-group, for some prime
p. If t > 4 and the degree is p + 1 it follows that the order of a vertex-
stabilizer Gv is (p + l)pt~1m, where t = 4,5 or 7 and m is a divisor of
{p-lf.

17g There are no 8-transitive graphs Weiss (1983) extended the results
of Gardiner and others, and using the classification theorems of group
theory he showed that there are no finite graphs (apart from the cycles)
for which a group of automorphisms can act transitively on the £-arcs
for t > 8. 7-transitive graphs do exist: the smallest is a 4-regular graph
with 728 vertices [BCN, p. 222].

17h Symmetric cycles A cycle with vertices VQ, V\, ..., vi~\ in a graph
F is symmetric if there is an automorphism g of F such that g(vi) = Vi+i,
where the subscripts are taken modulo /. J.H. Conway observed that
in a symmetric graph of degree k the symmetric cycles fall into fc — 1
equivalence classes under the action of the automorphism group. The
details may be found in Biggs (1981a). For example, the two classes
in Petersen's graph contain 5-cycles and 6-cycles, and in general, the
classes in Ok have lengths 6,10,..., 4fc — 6 and 2k — 1.
 18

Symmetric graphs of degree three

In this chapter we shall use the traditional term cubic graph to denote
a simple, connected graph which is regular of degree three. As we shall
see, the theory of symmetric cubic graphs is full of strange delights.
Suppose that F is a t-transitive graph so that, by definition, Aut(F)
is transitive on the f-arcs of F but not transitive on the (t + l)-arcs of
F. The distinctive feature of the cubic case is that Aut(F) acts regularly
on the t-arcs.
Proposition 18.1 Let [a] be a t-arc in a cubic t-transitive graph T.
Then an automorphism of F which fixes [a] must be the identity.
Proof Suppose / is an automorphism fixing each vertex «0; «i> • • • > ®t-
If / is not the identity, then / does not fix all t-arcs in F. It follows
from Lemma 17.4 that there is some t-arc [/?] such that / fixes [/?],
but / does not fix both successors of [/?]. Clearly, if 0t-i,u^1\u^2') are
the vertices adjacent to /%, then / must interchange u^ and u^2\ Let
w ^ /?i be a vertex adjacent to /3o- Since F is t-transitive there is an
automorphism h € Aut(F) taking the t-arc (w, fio,..., f3t-i) to [/?], and
we may suppose the notation chosen so that h((3t) = u^ • Then h
and fh are automorphisms of F taking the (t + l)-arc [w.0\ to its two
successors, and, by Theorem 17.5, Aut(F) is transitive on (t + l)-arcs.
This contradicts our hypothesis, and so we must have / = 1. •

From now on we shall suppose that we are dealing with a cubic t-

transitive graph F, and that we have chosen an arbitrary t-arc [a] in F.
 Symmetric graphs of degree three 139

If the stabilizer sequence of this t-axc is

Aut(r) = G > Ft > Ft-i > ... > FQ,
then Proposition 18.1 implies that |F 0 | = 1. Consequently we know the
orders of all the groups in the stabilizer sequence:
1*11 = 2* (0<t<t-l),
|F t | = 3x2*- 1 ,
\G\ = n x 3 x 2*-1 (n = |VT|).
The structure of these groups can be elucidated by investigating cer-
tain sets of generators for them. These generators are derived from the
sets Yi defined for the general case in Chapter 17. Let at-i,v^\v^ be
the vertices adjacent to at, and let gr (r = 1,2) denote automorphisms
taking [a] to (ai, ai,..., at, v^). We shall use the following notation:
9 = 9i, ^o = 9T192, xi=g~lxag% (i = 1 , 2 , . . . , * ) .
The effect of these automorphisms on the basic i-arc [a] is indicated in
Figure 13. We note that these automorphisms are unique, as a conse-
quence of Proposition 18.1.

• . • /
[a]

-• • • •. • • •
92 [«1 N^ xo[a]

Figure 13: the effect of 51,32 and xo on [a].

In this chapter, (X) will denote the subgroup of Aut(F) generated by

the set X.
Proposition 18.2 The stabilizer sequence of a cubic t-transitive graph
with t > 2 has the following properties:
(1) Fi = (xo,Xi,...,Xi-i) fori = 1,2,...,*;
(2) ifG* = (xo,xi,...,xt) then\G:G*\ < 2;

Proof We shall use the notation and results of Propositions 17.7 and
17.8. In the cubic case we have Fo = 1, and the set Yi consists of the
elements g^gi and their inverses g^ g\ for 1 < j < i.
(1) It follows from part (2) of Proposition 17.7 that Fi = (Fi). Now
140 Symmetry and regularity

and so Ft = (xo,xi,... ,x»-i).

(2) It follows from Proposition 17.8 that the group G* = (Yi+i), that
is (xo, x\,. •.,xt), is a subgroup of index 1 or 2 in G, provided that the
girth of F is greater than three. If the girth is three, then it is easy to
see that the only possibility is t = 2,F = K4, and we may verify the
conclusion explicitly in that case.
(3) If G = G*, then (xo,g) contains (xo,xi,..., xt) = G* — G. If
|G : G*\ = 2, then T is bipartite, and each element g* of G* moves
vertices of F through an even distance in F. But the element g = g\
moves some vertices to adjacent vertices, and so g £ G*. Thus, adjoining
g to G* must enlarge the group, and since G* is a maximal subgroup of
G (because it has index 2) we have (G*,g) = (xo,g) = G. •

In the previous chapter we considered Petersen's graph, obtaining for

the 3-arc [a] = (12,34,15,23) the automorphisms gx = (13)(245),#2 =
(13524). Hence
x0 = (34), Xl = (12), x2 = (35), x3 = (14).
We know that this graph is not bipartite, since it has cycles of length 5,
and so in this case G* = (xo,xi,X2,X3) = G « 5 5 .
Another simple example is the 2-transitive graph Q3, the (ordinary)
cube graph, depicted in Figure 14. Taking [a] = (1,2,3) we have the
automorphisms as listed.

/
5 (i

8 7

, /
4 ;s
Figure 14: the cube graph Q3.

51 = (1234)(5678), g2 = (123785)(46),
xo - (36)(45), X! = (16)(47), x2 =
In this case the graph is bipartite and G* = (aro,xi,X2) preserves the
bipartition
VQ3 = {1,3,5,7}U{2,4,6,8}.
It follows that \G : G*\ = 2.
 Symmetric graphs of degree three 141

The main result on i-transitive cubic graphs is that there are no finite
examples with t > 5. The proof of this very important result is due to
Tutte (1947a), with later improvements by Sims (1967) and Djokovic
(1972). Following these authors we shall obtain the result as an alge-
braic consequence of the presentation of the stabilizer sequence given in
Proposition 18.2. A rather more streamlined proof, using 'geometrical'
arguments to replace some of the algebraic calculations, has been given
by Weiss (1974).
We shall suppose that t > 4, as this assumption helps to avoid vacuous
statements. We observe that each generator Xi (i > 0) is an involution,
and that each element of Fi (1 < i < t — 1) has a unique expression in
the form
xpxa ... xT, where 0<p<a<...<T<i — 1,
where we allow the empty set of subscripts to represent the identity
element. The uniqueness of the expression is a consequence of the fact
that there are 2l such expressions, and \Fi\ = 2l for 1 < i < t — 1.
The key idea is to determine which stabilizers are abelian and which
are non-abelian. It is immediate that Fi and F2 are abelian, since |Fi | =
2 and (i^t = 4. Let A denote the largest natural number such that F\
is abelian.
Proposition 18.3 Ift > 4, then 2 < A < \{t + 2).
Proof We have already remarked that A > 2. Suppose that F\ =
(XQ,...,XX-I) is abelian, so that its conjugate g~t+x~1F\gt~x+1, that
is (xt-\+i,- • • ,xt), is also abelian. If
A - l > £ - A + l,
then both these groups contain ZA-I, and together they generate G*;
hence ZA-I commutes with every element of G*. Now g2 e G* (since
g e G and \G : G*\ < 2) and so
ZA-I = 9~2xx~i92 = x\+i,
whence x0 = X2- This is false, given t > 4, since IF3I > |F 2 |, and so we
must have
A-l<i-A+l, that is A < - ( < + 2),
as claimed. D

Proposition 18.3 gives an upper bound for A in terms of t. We shall

find a lower bound of the same kind by means of arguments involving
the commutators [a,b] — a~1b~1ab of the canonical generators a;,. Note
that since these generators are involutions, we have
[Xi,Xj] = (XiXj)2.
142 Symmetry and regularity

Lemma 18.4 The generators x% satisfy the following conditions.

(1) [xuXj] = 1 if \j - i| < A, but {xi,Xj} ^ 1 if \j - i\ = A.
(2) The centre of Fj = (xo, • • • ,Xj-i) is the group {XJ-\, • • • , ^ A - I ) (A <
3 < 2A).
(3) The commutator subgroup of F i + 1 is a subgroup of (xi,.. .,Xj_i)
= g-lFi^g{l<i<t-2).
Proof (1) We may suppose without loss that j > i; then [xi,Xj] =
g~ [xo,Xj-i]gt and so [xi,Xj] = 1 if and only if xo and Xj-i commute.
t

The result follows from the fact that F\ = (xo, • • •, XA-I) is the largest
abelian stabilizer.
(2) If the non-identity element x of Fj is written in the form

xpxa ...xT (0 < p < a < ... < T < j - 1),

then x does not commute with xp+\. Further, if p + A < j then xp+\
belongs to Fj. Similarly, x does not commute with xT~\, and if r — A >
— 1, then xT-\ belongs to Fj. Thus, if x is in the centre of Fj then
p > J' — A and r < A - 1, so that x is in (XJ-\, . . . , x^-i). Conversely, it
follows from (1) that every element of this group is in the centre of Fj.
(3) Provided that 1 < i < t - 2, the groups Fj = (x0,... ,Xj_i)
and g~lFig = (xi,...,x») are different, and they are both of index
two in Fi+i, and consequently normal in Fi+i. Thus their intersection
(xi,... ,Xi-i) = g~1Fi^ig is normal in Fi+i, and the quotient group
Fi+i I'{g~lFi-ig) is abelian, since it has order 4. Hence the commutator
subgroup of Fi+i is contained in g~1Fi-.ig. •

Since [xo, x\] belongs to the commutator subgroup of Fx-i, it follows

(from part (3) of Lemma 18.4 with i = A) that [xoiX^] belongs to the
group ( x i , . . . , x\_i). In other words, there is a unique expression

[xo,x\]=xli...xv (1 < fi < v < A - 1).

L e m m a 18.5 With the above notation, we have:

(1) fj, + \>t-l; (2) 2X-i/>t-l.
Proof (1) Suppose that p, + A < t — 2. Then (by part (3) of Lemma
18.4) the element [XO,XM+A] of the commutator subgroup of F M + A + I
is contained in ( x i , . . . , x^+x-i)- The centre of ( x i , . . . , X ^ + A - I ) is the
group (x M ,..., x\), and since this contains both x,\ and [xo, x\] it follows
that [XO,XM+A] commutes with x\ and with [XQJXA]. Also x^ commutes
 Symmetric graphs of degree three 143

with x^+x, since /J, < A — 1. Hence we have the following calculation:

1
[xo,xx][xo, O;M
= [Xo,Xx].
This implies that xM+.\ commutes with [xo,xx] = x^.-.Xv But this
is false, since X^+A does not commute with £M but does commute with
any other term in the expression for [xo>£.\]- Thus our hypothesis was
wrong, and /i. + A > t — 1.
(2) If 2A - u < t — 2, then using arguments parallel to those in (1), we
may prove that [x2\-v,Xo] commutes with xx-u and with [xx-v,X2\-v]\
also Xx-v commutes with xo, since v > 1. A calculation like that in (1)
then implies that XQ commutes with
[x\-v, X2\-u] = Xp+X-v • • • Xx,
which is false. Hence 2A — is >t — 1. •

Theorem 18.6 (Tutte 1947a) There is no finite t-transitive cubic

graph with t > 5.
Proof If t is at least four, then Proposition 18.3 tells us that A <
\{t-\- 2). However, the results of Lemma 18.5 show that t — 1 — A < / i <
is < 2A — t + 1; that is, A > \{t — 1). Now, if t > 4 there is an integer A
such that
| ( « - l ) < A < | ( t + 2)
only when t = 4,5,7. It remains to exclude the possibility t — 7, which
is done by means of the following special argument.
If F is a 7-transitive cubic graph, then the inequalities for A, // and
is imply that A = 4, \i = is = 2; thus [£0,0:4] = £2- Also, by part
(3) of Lemma 18.4, [:ro,x5] belongs to the group ( i j , 12,^3,14). If the
standard expression for [£0,3:5] actually contains x±, then we can write
[xo>X5] = hx±, where h e (xi,x2,x3) so that h commutes with XQ and
X4. Hence
x2 - \XQ,X^ = (x0x4)2 = (xohxt)2 = (xo(xox5)2)2
= (x5x0x5)2 = x5xlx5 = 1.
Since this is absurd, [xo,X5] = (xox5)2 must belong to (zi,£2,0:3).
Now the original definitions show that Xi,X2, and X3 fix the vertex
03 of the 7-arc [a], and so £0X5(0:3) = £5£o(a3) = £5(03). That is,
xo fixes £5(0:3). Further, since x 5 fixes a\ but not a 2 we have a 7-
arc [0] — (x5(a3),x5(a2),ai,a2,a3,a4,a5,a6) in T. The three vertices
144 Symmetry and regularity

adjacent to ai are ao,a2 and £5(0:2), and since XQ fixes ao,ai and a 2
it must fix x^(a2) also. Consequently XQ fixes the whole 7-arc [0], and
this contradicts Proposition 18.1. Hence t = 7 cannot occur. •

Goldschmidt (1980) proved an important extension of this result.

The simplest example of a 5-transitive cubic graph is constructed as
follows. Let the symmetric group SQ act on the 6 symbols {a, b, c, d, e, / } ,
and take the vertices of a graph Q to be the 15 permutations of shape
(ab) and the 15 permutations of shape (ab)(cd)(ef). Join two vertices
by an edge if and only if the corresponding permutations have different
shape and they commute. For instance, (ab) is joined to the vertices
(ab)(cd)(ef), (ab)(ce)(df) and (ab)(cf)(de), while (ab)(cd)(ef) is joined
to (ab), (cd) and (e/). Clearly, any automorphism of the group S$ is an
automorphism of fi, and so
|Aut(fi)| = |AutS 6 | = 1440 = 30 x 3 x 2 4 ,
as we expect for a 5-transitive cubic graph with 30 vertices. We can
verify that O is indeed 5-transitive by working out generators in terms
of the following 5-arc:
(06), (ab)(cd)(ef), (cd), (ae)(bf)(cd), (ae), (ae)(bd)(cf).
If 7T is an element of 56, denote the corresponding inner automorphism
(conjugation) of 56 by |TT|. Then the generators for the stabilizer se-
quence may be chosen as follows:
xQ = \(cd)\, *i = \(ab)(cd)(ef)\, x 2 = |(a&)|,
x3 = |(o6)(c/)(de)|, x4 = |(c/)|.
The groups which occur in the stabilizer sequence are
F5 = 5 4 xZ 2 , F4 = D8x Z2, F3 = (Z 2 ) 3 ,
F 2 = (Z 2 ) 2 , F,=Z2.
Finally, we may choose x5 so that G* = (x0,... ,x5) is isomorphic to
5e, and so \G : G*\ = 2 in accordance with the fact that the graph is
bipartite.
Additional Results

18a A non-bipartite 5-transitive cubic graph A 5-transitive cubic graph

with 234 vertices, which is not bipartite, can be constructed as follows.
The vertices correspond to the 234 triangles in PG(2,3) and two vertices
are adjacent whenever the corresponding triangles have one common
point and their remaining four points are distinct and collinear. The
automorphism group is the group Aut PSL(3,3), of order 11232 = 234 x
3x24.
 Symmetric graphs of degree three 145

18b The sextet graphs (Biggs and Hoare 1983) Let q be an odd prime
power. Define a duet to be an unordered pair of points ab on the pro-
jective line PG(l,q) = GF(q) U {oo}, and a quartet to be an unordered
pair of duets {ab | cd} such that the cross-ratio

^
(a - d)(b - c)
(The usual conventions about oo apply here.) A sextet is an unordered
triple of duets {ab | cd | ef} such that each of {ab | cd}, {cd \ ef} and
{ef | ab} is a quartet. There are q(q2 — l)/24 sextets if q = 1 (mod 4),
and none if q = 3 (mod 4).
When q = 1 (mod 8) it is possible to define 'adjacency' of sextets
in such a way that each sextet is adjacent to three others. Thus we
obtain a regular graph T,(q) of degree 3, whose components £o(<?) are all
isomorphic. The sextet graph S(p) is defined to be £o(p) if p = 1 (mod
8) and T,0(p2) if p = 3,5,7 (mod 8).
The sextet graphs S(p) so defined form an infinite family of cubic
graphs, one for each odd prime p. The graph S(p) is 5-transitive when
p = 3 or 5 (mod 8), and 4-transitive otherwise. The order of S(p)
depends on the congruence class of p modulo 16, as follows:
n = —pip2 — 1) when p = 1,15 (mod 16);

n = ^rp(p 2 - 1) when p = 7,9 (mod 16);

n = ^p2ipA - 1) when p = 3,5,11,13 (mod 16).

The group AutS(p) is PSL(2,p), PGL(2,p), PTL(2,p2) in the respec-
tive cases. The two smallest 5-transitive sextet graphs are 5(3), which is
isomorphic to the graph 0 described above, and 5(5), which is a graph
with 650 vertices.

18c Conway 's presentations and the seven types Given an arbitrary t-
arc [a], let a and b denote the automorphisms taking [a] to its successors
(so a = pi and b = g2 in the notation described at the beginning of this
chapter). Also, let a be the automorphism which reverses [a]; that is,
o-(cti) = at-i (0<i< t).
Since we know that Aut(f) acts regularly on the i-arcs, it follows that
a2 is the identity and aacr is either a" 1 or 6" 1 . We denote the case when
cracr = a" 1 by t+ and the case when aaa = b~1 by t~. It turns out that
the t+ case can occur only when t = 2,3,4,5 and the t~ case only when
t = l,2,4.
In each of the cases it can be shown, by analysis of the action of
146 Symmetry and regularity

suitable combinations of a,b and a on [a], that certain relations must

hold in Aut(F). For example, in the 2+ case these relations are:
a2 = 1, (era)2 = 1, (ab)2 = 1, (a^b)2 = 1, abaa2 = b2.
In the 5 + case they are:
a2 = 1, {aaf = 1, (crb)2 = 1, (a"1?))2 = 1, ( a ~ V ) 2 = 1,
(a- 3 6 3 ) 2 = 1, a4b'4a4 = ba% a4baab = ba3b.
Let us denote the groups generated by a, b and <r, subject to the
appropriate relations, by
G2 , G3 , G 4 , Gh , Gx , G2 , G 4 •
Each of these groups is an infinite group of automorphisms of the infinite
cubic tree T3, acting regularly on the i-arcs, for the relevant value of t,
and they are the only such groups, up to conjugacy in Aut(Ta). More
detailed information about the seven groups, using different presenta-
tions, is given by Djokovic and Miller (1980), and Conder and Lorimer
(1989).

18d Finite cubic graphs and groups Any group acting regularly on the
i-arcs of a finite cubic graph F is a quotient of one of the seven groups in
18c. The quotient is defined by adding relations which represent cycles
in T, a cycle of length I in T corresponding to a word of length I in a and
b which represents the identity. For example, adding the relation a4 — 1
to the relations for G j defines a group G^ia4). This is the group of the
cube Qs, as can be verified by showing that the permutations
o = (1234)(5678), b = (123785)(46), a = (13)(57),
satisfy the defining relations for G^ia4) and represent automorphisms
of Qz acting in the prescribed way on the 2-arc (1,2,3) (see Figure 14).

18e Coset enumeration In the notation of the Conway presentations,

the stabilizer of a f-arc is
Ft = (a-ibi\i= 1,2,...,t).
If G is a quotient of G^ or GJ, then the index \G : Ft\ is the cardinality
of a cubic graph for which G is a t-transitive group of automorphisms.
The index may be finite or infinite, but if it is finite the method of coset
enumeration will (in principle) determine its value. This is a power-
ful method for constructing finite ^-transitive cubic graphs. See Biggs
(1984a) for further details.

18f The structure of a stabilizer sequence The groups occuring in the

 Symmetric graphs of degree three 147

stabilizer sequence are determined up to isomorphism, as in the following

table. (Note that when £ = 2,4 both the t+ and t~ cases can occur, but
the abstract groups are the same.)
t Fl F2 F3 F4 F5
1 Z3
2 Z2 S3
3 Z 2 (Z 2 ) 2 D12
4 Z 2 (Z 2 ) 2 As
5 Z 2 (Z 22) 2

18g Symmetric Y and H graphs Let Y and H denote the trees whose
pictorial representations correspond to the respective letters. Both of
these trees have vertices of degree 1 (leaves) and 3 only. Given any such
tree T we can form an expansion of T by taking a number n of disjoint
copies of T and joining each set of corresponding leaves by a cycle of
length n; each cycle has a constant 'step', and different cycles will in
general have different steps. For example, when T = K2, we get the
graphs P(n, t) described in 15e by joining one set of leaves with step 1
and the other set with step t.
Clearly an expansion of T is a cubic graph. The result quoted in 15e
implies that only seven expansions of K2 are symmetric. Horton and
Bouwer (1991) showed that there are only six other expansions which
are symmetric. Four of them are expansions of Y: n = 7, steps 1,2,4;
n = 14, steps 1,3,5; n = 28, steps 1,3,9; n = 56, steps 1,9,25. The
other two are expansions of H: n = 17, steps 1,2,4,8; and n — 34, steps
1,9,13,15.

18h Foster's census of symmetric cubic graphs (More details and bibli-
ographical references relating to the following sketch are given by Bouwer
(1988).) In 1920 two electrical engineers, G.A. Campbell and R.M. Fos-
ter, wrote a paper in which the graph ^3,3 was used in the context of
'telephone substation and repeater circuits'. Twelve years later Foster
published drawings of nine symmetric cubic graphs. He continued to
work on the subject, and in 1966 he spoke at a conference at the Uni-
versity of Waterloo, where he distributed a mimeographed list of such
graphs with up to 400 vertices. In 1988, when Foster was just 92, Bouwer
and his colleagues published Foster's census for graphs with up to 512
vertices. Remarkably, only five graphs (out of 198) are known to have
been missed by Foster, and workers in this field are convinced that there
can be very few others, if any.
The graphs with n < 30 vertices are as follows.
148 Symmetry and regularity

Ki, #3,3, Q3, Petersen's graph, Heawood's graph 5(7), P(8,3)

(see 15e), the Pappus graph (see 19h), P(10,3) or the Desargues
graph (see 19b), the dodecahedron, P(12,5), Y(7 : 1,2,4) (see
18g), and fi = 5(3).

18i All 5-transitive cubic graphs with less than 5000 vertices Coset
enumerations based on the Conway presentations and other techniques
have established that the following list of 5-transitive cubic graphs with
n < 5000 vertices is 'almost certainly' complete.
n - 30: the sextet graph 5(3), group G£(a8);
n = 90: a threefold cover of 5(3) (see 19c), group G£ (610);
n — 234: the graph described in 18a, group Gg"(a13);
n — 468: a double covering of the previous graph, group G$(b12);
n = 650: the sextet graph 5(5), group G£(a12);
n = 2352: a graph to be described in 19e, group Gg"(a14);
n = 4704: a double covering of the previous graph, group G$((ab)8).

18j The symmetric group 5io is a quotient of G$ (Conder 1987) The

following permutations of { 1 , . . . , 9, X} satisfy the Conway relations for
G j , as given in 18c:
a = (12) (34675) (89X), b = (1246853)(79X), a = (12)(34)(56)(9JsT).
Since these permutations generate the symmetric group 5io, it follows
that there is a 5-transitive cubic graph with 101/48 = 75600 vertices.
The graph can be constructed in a way which shows that it is closely
related to the simplest 5-transitive cubic graph, the graph fi = 5(3)
(Lorimer 1989).
 19

The covering graph construction

In this chapter we shall study a 'covering graph' technique which, in

certain circumstances, enables us to manufacture new symmetric graphs
from a given one. The method was first used in this context by J.H. Con-
way, who used the simple version discussed in Theorem 19.5 to show that
there are infinitely many connected cubic graphs which are 5-transitive.
The general version given here was developed in the original 1974 edi-
tion of this book, and has since found several other applications, some of
which are described in the Additional Results at the end of the chapter.
The related technique of 'voltage-graphs' (see Gross 1974) is much used
in the theory of graph embeddings.
We shall use the symbol ST to denote the set of 1-arcs or sides of a
graph F; each edge {u, v} of F gives rise to two sides, (u, v) and (v,u).
For any group K, we define a K-chain on F to be a function <j>: SF —* K
such that <j)(u,v) = (^(i^u))" 1 for all sides (u,v) of F .
Definition 19.1 The covering graph F = T(K, <p) of F, with respect
to a given Jif-chain 0 on F, is defined as follows. The vertex-set of F is
K x VT, and two vertices (KI, wi), {K2, V2) are joined by an edge if and
only if
(vitV?) € SF and K2 — Ki4>{v\, v2)-
It is easy to check that the definition of adjacency depends only on the
unordered pair of vertices.
As an example let F = K4, and let K be the group Z2 whose elements
150 Symmetry and regularity

r f
Figure 15: Q3 as a double covering of K4.

are 1 and z\ the function <^> which assigns z to each side of K4 is a Z 2 -

chain on K4. The covering graph F(Z 2 , <j>) is isomorphic to the cube Q3,
as depicted in Figure 15.
Suppose that a group G acts as a group of automorphisms of a group
K; that is, for each g in G we have an automorphism g of K such that
the function g »—> p is a group homomorphism from G to Autif. In this
situation we define the semi-direct product of K by G, denoted by KxG,
to be the group whose elements are the ordered pairs (n,g), with the
group operation given by

Let T be a graph, cj> a if-chain on T, and let G = Aut(F). Then G

acts on the sides of T by the rule g(u,v) = (g(u),g(v)), and we may
postulate a special relationship between the action of G on K and its
action on ST.
Definition 19.2 The .ftT-chain <j> is compatible with the given actions
of G on K and ST, if the following diagram is commutative for each g
inG.

ST K

ST -1+ K
Proposition 19.3 Suppose that T is a graph whose automorphism
group G = Aut(F) acts as a group of automorphisms of a group K.
Suppose further that there is a K-chain <f> onT which is compatible with
the actions of G on K and ST. Then the semi-direct product KxG is a
group of automorphisms of the covering graph T = T(K, <f>).
 The covering graph construction 151

Proof Define the effect of an element («, g) of Kx G on a vertex («', v)

of F by the rule
~ (Kg(n'),g(v)).
(K,9)(K',V)
Using the definition of compatibility, a simple calculation shows that
this permutation of VT is an automorphism of F. •

The usefulness of the covering graph construction lies in the fact that
a much stronger version of Proposition 19.3 is true.
Proposition 19.4 With the notation and hypotheses of Proposition
19.3, suppose also that G is transitive on the t-arcs ofT. Then KxG is
transitive on the t-arcs ofF.
Proof ^ Let ((KO,VQ), ..., (Kt,vt)) and ((K'O,V'O), ..., («J>ut)) b e t w o l~
arcs in F. Then (vo,.. •, vt) and (v'o,..., v't) are f-arcs in F, and so there
is some g in G such that g(vi) = v\ (0 < i < t). Suppose we choose
K* in K such that (n*,g) takes (KO,VQ) to (K'O,V'O); that is, we choose
K* = KQI^KQ))"1 . Then we claim that (n*,g) takes {ni,Vi) to (K^,^)
foralH€{0,l,...,0-
The claim is true when i = 0, and we make the inductive hypothesis
that it is true when i — j - 1, so that
(«j-_i,uj_i) = (K*,5)(KJ_I,VJ_I) = {^giKj^givj-i)).
Since (KJ,VJ) is adjacent to (K,J-I,VJ-I) we have Kj — KJ-I4>(VJ-I,VJ),
and the corresponding equation holds for the primed symbols as well.
Thus:

Consequently, («*,</) takes (KJ,VJ) to (K^,^-), and the result follows by

the principle of induction. •

The requirement that a compatible if-chain should exist is rather

restrictive. In fact, for a given graph F and group K, it is very likely that
the only covering graph is the trivial one consisting of \K\ components
each isomorphic with F. However, it is possible to choose K (depending
on F) in such a way that a non-trivial covering graph always exists.
Let us suppose that a t-transitive graph F is given. We define K to be
the free Z2-module on the set £T; thus, K is the direct product of \ET\
copies of %2, and its elements are the formal products Ylea » where
a(a) = 0 or 1 and the product is over all ea in ET. The automorphism
group G = Aut(F) acts on K through its action on ET, and furthermore
152 Symmetry and regularity

there is a fC-chain (f> onT defined by the rule 4>{u,v) = e i ; where e% =

{u, v} regarded as an element of K. This K-chain is compatible with
the actions of G on K and ST, and so the covering graph F = T(K, <j>)
exists and (by Proposition 19.4) its automorphism group is transitive on
i-arcs. •

Theorem 19.5 Let T be a t-transitive graph whose rank and co-rank

are r(T) and s(T). Then, with the special choices ofK and <f> given above,
the covering graph T consists of2r^ connected components, each having
2 a ( r ) |VT| vertices.
Proof Pick a vertex v of F, and let F o denote the component of F
which contains the vertex (l,v). If
V - U0,Ui,... ,U[ = V

are the vertices of a cycle in F, with edges e^ = {UJ_I, Ui} , then we have
the following path in F o :
(l,v), (ei,ux), (eie2,u2), . . . , ( e i e 2 .. -et,v).
Conversely, the vertex (n, v) is in Fo only if « represents the edges of a
cycle in F. Since there are s(F) independent cycles in F, there are 2S^
elements K in K such that (K, V) is in Fo- It follows that Fo has 2s^r-* | VT
vertices; further, F is vertex-transitive and so each component has this
number of vertices. Finally, since
|VT | = \K\\VT\ = 2 l E r | | F F | and r(F) + s(F) = |JET|,
there must be 2 r ^ r ' components. D

Corollary 19.6 There are infinitely many cubic 5-transitive graphs.

Proof We know that there is at least one cubic 5-transitive graph,
the graph ft constructed at the end of the previous chapter. Applying
the construction of Theorem 19.5 to O, we obtain a cubic 5-transitive
graph fi0 with 2s^n^|FJ7| vertices, and since s(Cl) > 0 this graph is not
isomorphic with il. We may repeat this process as often as we please,
obtaining an infinite sequence of graphs with the required properties.
•

Of course the number of vertices used in Corollary 19.6 quickly be-

comes astronomical; for instance, the two graphs which follow fi in the
sequence have about 2 21 and 2 100000 vertices respectively. Biggs and
Hoare (1983) have given an explicit construction for infinitely many cu-
bic 5-transitive graphs which involves much smaller numbers (see 18b).
 The covering graph construction 153

Additional Results

19a Double coverings Let G be the automorphism group of a con-

nected graph T, and let G act on the group Z 2 by the rule that g is the
identity automorphism of Z2 for each g in G. Then the Z2-chain <fi on F
which assigns the non-identity element of Z2 to each side of F is compat-
ible with the actions of G on SV and Z 2 . The covering graph F(Z2, (f>)
is connected if and only if F is not bipartite. For example, applying the
construction to the graph with 234 vertices described in 18a we get a
connected 5-transitive cubic graph with 468 vertices.

19b The Desargues graph The construction of 19a applied to Pe-

tersen's graph results in a cubic 3-transitive graph with 20 vertices. The
vertices of this graph correspond to the points and lines in the Desargues
configuration, with two vertices being adjacent if they correspond to an
incident (point, line) pair. This graph was described by Coxeter (1950),
together with several others derived from geometrical configurations.

19c A threefold covering of 5(3) The second 5-transitive cubic graph

in order of magnitude is a graph with 90 vertices which is a threefold
covering of the sextet graph 5(3) (see 18i). Ito (1982) constructed an
explicit Z3-chain on 5(3), which shows that the graph is a covering graph
of 5(3) in the sense of this chapter.
19d Another covering construction for 5-transitive cubic graphs Sup-
pose that F is a cubic graph and G = Aut(F) is a group of type 4 + .
Then the automorphism a~lb fixes the vertices ao, ai, a.i, and 0:3 of the
basic 4-arc [a], and (because the degree is 3) it must fix the other ver-
tices /3i,/32 adjacent to oc\,a.2 respectively. By considerations of order
we see that this is the only non-identity automorphism with this prop-
erty. So for each e = {v, w} € ET the group Lvw has order 2; in other
words, there is a unique involution je which fixes e and the four vertices
adjacent to e. The involutions je generate the group G* (Proposition
18.2) which is normal, of index 1 or 2 in G. Consequently G acts by
conjugation as a group of automorphisms of G*.
If we take K = G* and define a If-chain on F by
<f>(v,w)=je,
then the compatibility condition is satisfied and, by Proposition 19.4,
we have a graph F on which KxG acts 4-transitively. However, there is
a bonus. As shown by Biggs (1982b) there is an 'extra' automorphism
F, so that F is in fact a 5-transitive graph.
154 Symmetry and regularity

19e A 5-transitive cubic graph with 2352 vertices The simplest case
of 19d is when F = 5(7), a graph with 14 vertices also known as the
Heawood graph. In this case both F and its 5-transitive covering graph F
with 2352 vertices can be constructed directly in terms of the seven-point
plane PG(2,2) (Biggs 1982a).

19f Conway generators for the covering graph Let a be the Conway
generator for the t-transitive group G of F, with respect to the i-arc
[a], and suppose <>/ is a compatible if-chain. Then the corresponding
generator a for the group KxG of T(K, <j>) is (A, a), where A = </>(c*o, oi\).

19g Homological coverings Let F be a graph with co-rank s and let R

be a ring. The 'first homology group with coefficients in K1 of a graph F is
the direct product R" of s copies of R. (This a just a mild generalization
of the cycle space denned in Chapter 4.) The functorial properties of
homology imply that the automorphism group of F acts as a group of
automorphisms of the homology group, and so a covering graph F can be
constructed using K = Rs. Biggs (1984b) gave an explicit form of this
construction and showed that when R = TL the number of components
of F is equal to the tree-number of F.

19h The Pappus graph In the homological covering construction, take

R — Z 3 as the coefficient group and F = #3,3, so that s = 4 and
K = Z 3 . The covering graph in this case has 27 components, each
with 6 x 3 4 /27 = 18 vertices. Each component is a copy of the Pappus
graph, whose vertices correspond to the points and lines of the Pappus
configuration, with adjacent vertices corresponding to an incident (point,
line) pair. See also Coxeter (1950).
 20

Distance-transitive graphs

In Chapter 15 a connected graph F was defined to be distance-transitive

if, for any vertices u,v,x,y of F satisfying d(u, v) = d(x,y), there is an
automorphism g of F which takes a t o i and v to y.

ro(») r,(») rs(t>)

Figure 16: a distance-partition of K%^.

It is helpful to recast the definition. For any vertex v of a connected

graph F we define

where i is a non-negative integer not exceeding d, the diameter of F. It

is clear that F0(i>) = {v}, and VT is partitioned into the disjoint subsets
ro(v),... ,Td(v), for each v in VT. Small graphs may be depicted in a
manner which emphasises this partition by arranging their vertices in
columns, according to distance from an arbitrary vertex v. For example,
^3,3 is displayed in this way in Figure 16.
156 Symmetry and regularity

Lemma 20.1 A connected graph F with diameter d and automor-

phism group G = Aut(F) is distance-transitive if and only if it is vertex-
transitive and the vertex-stabilizer Gv is transitive on the set Fj(i>), for
each i € { 0 , 1 , . . . , d}, and each v € VT.
Proof Suppose that F is distance-transitive. Taking u = v and x = y
in the definition (as given above), we see that F is vertex-transitive.
Taking y = v, we see that Gv is transitive on Ti(v) (0 < i < d).
Conversely, suppose vertices u, v, x, y are given such that d(u, v) =
d(x, y) — i. Let g be an automorphism such that g(v) = y and let
h 6 Gy be such that h(g(u)) = x. Then hg takes u to x and v to y.
D

As we shall see, the adjacency algebra (defined in Chapter 2) plays a

major part in the study of distance-transitive graphs. In preparation for
the algebraic theory we begin by investigating some simple combinatorial
consequences of the definition.
For any connected graph F, any vertices u,v of F, and any non-
negative integers h and i, define Shi(u, v) to be the number of vertices
of F whose distance from u is h and whose distance from v is i. That is,
8hi{u,v) = \{w e VT | d(u,w) = h and d(v,w) = i}\.
In a distance-transitive graph the numbers Shi{u, v) depend, not on the
individual pair (u, v), but only on the distance d(u, v). So if d(u, v) = j
we shall write
Shij = Shi(u,v).

Definition 20.2 The intersection numbers of a distance-transitive

graph with diameter d are the numbers Shij, where h,i and j belong to
the set {0,1,...,d}.

Clearly there are (d + I) 3 intersection numbers, but it turns out that

there are many identities relating them; and in due course we shall show
that just 2d of them are sufficient to determine the rest.
Consider the intersection numbers with h = 1. For a fixed j , SUJ is
the number of vertices w such that w is adjacent to u and d(v, w) = i,
when d(u, v) — j . Now, if w is adjacent to u and d(u, v) = j , then
d(v, w) must be one of the numbers j — l,j,j + 1; in other words
aiij=0 if i^j-l,j,j + l.
More generally, s^j = 0 if the largest of h,i,j is greater than the sum
of the other two.
 Distance-transitive graphs 157

For the intersection numbers SUJ which are not identically zero we
shall use the notation
C
J= 8
l,j-l,j a
3 — sl,j,j> fy = Sij+itj,
where 0 < j < d, and it is convenient to leave c$ and bd undefined.
The numbers Cj,aj, bj have the following simple interpretation in terms
of the diagrammatic representation of F introduced at the beginning of
this chapter. If we pick an arbitrary vertex v and a vertex u in Tj(v),
then u is adjacent to Cj vertices in Tj-i(v),aj vertices in Tj(v), and bj
vertices in Fj + i (y). These numbers are independent of u and v, provided
that d(u, v) = j .
Definition 20.3 The intersection array of a distance-transitive graph
is

{ *
a0 ai ... a, ...
Ci

bo bi . . . bj . . .
... Cj ...

For example, consider the cube Qz, which is a distance-transitive

graph with diameter 3. From the representation in Figure 17 we may
write down its intersection array.

* 1 2 3
t(Q3) = { 0 0 0 0
3 2 1 *
Figure 17: Qz as a distance-transitive graph.

We observe that a distance-transitive graph is vertex-transitive, and

consequently regular, of degree k say. Clearly we have bo = k and
ao = 0, C\ — 1. Further, since each column of the intersection array
sums to k, if we are given the first and third rows we can calculate
the middle row. Thus it is both logically sufficient and typographically
convenient to use the alternative notation
t(r) = {fc, &!,..., & d _i;l,C2,...,Cd}.
However, the original notation of Definition 20.3 is intuitively helpful,
158 Symmetry and regularity

and we shall continue to use it whenever it seems appropriate. In due

course we shall see that the intersection array determines all the inter-
section numbers Shij.
Many well-known families of graphs are distance-transitive, although
this apparent profusion of examples is rather misleading, because the
property is, in some senses, very rare. The complete graphs Kn and the
complete bipartite graphs Kk,k are distance-transitive. Their diameters
are 1 and 2 respectively, and the intersection arrays are:
f * 1 ) (* I k
t{Kn) ~{ 0 n - 2 \ , i(Kkik) ={0 0 0
[n-1 * J [k fc-1 *
The triangle graphs A t = L(Kt) (p. 21) are distance-transitive, with
diameter 2, and for t > 4:
* 1 4

!
0 t-2 2*-8
2f - 4 t - 3 *
Many other distance-transitive graphs will be described in the following
chapters.
Denote by ki (0 < i < d) the number of vertices in Fj(u) for any
vertex v; in particular &o = 1 and k\ = k.
Proposition 20.4 Let T be a distance-transitive graph whose intersec-
tion array is {k,b\,... ,bd-i; 1,C2,... ,Cd}. Then we have the following
equations and inequalities:
(1) ki^lH-i = ha (1 < i < d).
(2) 1 < c2 < c3 < . . . < cd.
(3) k>h >b2 >...>bd-i.
Proof (1) For any v in VT, there are fcj_i vertices in Fi-i(v) and each
is joined to 6j_i vertices in Fi(v). Also there are &j vertices in Ti(v) and
each is joined to Cj vertices in Fj_i(i;). Thus the number of edges with
one end in Fi_i(t;) and one end in Fi(v) is fc;_i&t_i = kiCi.
(2) Suppose u is in Ti+i(v) (1 < i < d - 1). Pick a path v,x,...,u of
length i + l; then d(x,u) —i. liw is in Fi_i(a;)nr 1 (u), then d(v,w) = i,
and so w is in Fi(v) n Fi(u). It follows that
Ci = tri_i(a;) n Fi(u)| < |Fi(t;) n Ti(u)\ = ci+1.
(3) This is proved by an argument analogous to that used in (2). •

Proposition 20.4 provides some simple constraints which must be sat-

isfied if an arbitrary array is to be the intersection array of some distance-
transitive graph. We shall obtain much more restrictive conditions in the
 Distance-transitive graphs 159

next chapter. However, in order to derive these conditions, we need not

postulate that the graph is distance-transitive, but merely that it has
the combinatorial regularity implied by the existence of an intersection
array. This is the justification for the following definition.
Definition 20.5 A distance-regular graph is a regular connected graph
with degree k and diameter d, for which following holds. There are
natural numbers
bo = k, bi,...,bd-i, c\ — 1 , C2,...,Cd,
such that for each pair (u, v) of vertices satisfying d(u, v) = j we have
(1) the number of vertices in r j _ 1 (t)) adjacent to u is Cj (1 < j < d)\
(2) the number of vertices in Fj+i (v) adjacent to u is bj (0 < j < d — 1).
The array {k,bi,... ,&<j_i; 1,C2,... ,Cd} is the intersection array of F.
Note that a distance-regular graph with diameter d = 2 is simply a
strongly regular graph, as denned in 3c. In terms of the general defini-
tion, the parameters a and c of a strongly regular graph are given by
a — k — l — bi a n d c = c<x-
It is clear that a distance-transitive graph is distance-regular, but the
converse is not true. Although many 'familiar' examples of distance-
regular graphs are distance-transitive, it is possible to construct arbi-
trarily large families of distance-regular graphs which are not distance-
transitive. Several examples will be given in the course of the following
chapters.
We shall now construct a basis for the adjacency algebra of a distance-
regular graph. Given a graph F with vertex-set {v\,..., vn} and diame-
ter d, define a set {Ao, A i , . . . , Ad} o f n x n distance matrices as follows:
1
-I X d{vr,va) = h;
— 1 0 otherwise.
In particular Ao = I, and Ai is the usual adjacency matrix A of F. We
notice that Ao + Ai + . . . + Ad = J, where J is the all-1 matrix.
Lemma 20.6 Let F be a distance-regular graph and let
{k,h,..., bd-i; 1, c 2 , . . . , ca}
be its intersection array. For 1 < i < d — 1, define a* = k — bi — Ci; then
AAi = 6£_iAi_i + a,Ai + Ci+iAi+i (1 < i < d - 1).
Proof Prom the definition of A and Ai it follows that (AAj) rs is
the number of vertices w of F such that d(vr,w) = 1 and d(vs,w) =
i. If there are any such vertices w, then d(vr, vs) must be one of the
numbers i — 1, i, i +1, and the number of vertices w in these three cases
160 Symmetry and regularity

is 6i_i,Oi,Ci+i, respectively. Thus (AA,) r s is equal to the (r,s)-entry

of the matrix on the right-hand side. •

Theorem 20.7 (Damerell 1973) Let F be a distance-regular graph

with diameter d. Then {Ao, A x , . . . , Ad} is a basis for the adjacency
algebra A{T), and consequently the dimension of A(T) is d+1.
Proof By recursive applications of the lemma we see that Aj is a
polynomial Pi(A), for i = 2,...,d. The form of the recursion shows
that the degree of pi is at most i, and since Ao, A i , . . . , A^ are linearly
independent (exactly one of them has a non-zero entry in any given
position) the degree of p; is exactly i.
Since Ao + Ai + . . . 4- Ad = 3 and T isfc-regularwe have
(A - AI)(Ao + Ai + . . . + A d ) = 0.
The left-hand side is a polynomial in A of degree d+1, so the dimension
of -4(r) is at most d+1. However, since {Ao, A j , . . . , A^} is a set of d+1
linearly independent members of A(T), it is a basis, and the dimension
is equal to d + 1. •

It follows from Theorem 20.7 that a distance-regular graph has just

d+1 distinct eigenvalues, the minimum number possible for a graph of
diameter d. These eigenvalues, and a remarkable formula for calculating
their multiplicities, form the subject of the next chapter.
The full set of (d + I) 3 intersection numbers can be defined for a
distance-regular graph; this is a trivial remark for a distance-transitive
graph, but it requires proof in the distance-regular case. In the course
of the proof we shall relate these intersection numbers to the basis

Proposition 20.8 Let F be a distance-regular graph with diameter d.

(1) The numbers Sih(u,v), h,i € {0,1,...,d}, depend only on d(u,v).
(2) If Shi(u,v) = Shij when d(u,v) = j , then
d

j=o
Proof We prove both parts in one argument. Since {A o , A 1 ; . . . ,
is a basis for A(T), the product A^Aj is a linear combination
Now
rs — Shi(vr,vs),

and there is just one member of the basis whose (r, s)-entry is 1: it
is that Aj for which d(vr,vs) = j . Thus Shi(vr,vs) = thij, and so
 Distance-transitive graphs 161

Shi{vr,vs) depends only on d(vT,vs). Further, the coefficient thij is just

the intersection number Shij- Q

At this point a few historical remarks are in order. The theory which
underlies our treatment of the adjacency algebra of a distance-regular
graph was developed in two quite different contexts. First, the associ-
ation schemes used by Bose in the statistical design of experiments led
to an association algebra (Bose and Mesner 1959), which corresponds
to our adjacency algebra. Bose and others also studied strongly regular
graphs which, as we have noted, are just distance-regular graphs with
diameter 2. Secondly, the work of Schur (1933) and Wielandt (1964) on
the commuting algebra, or centralizer ring, of a permutation group, cul-
minated in the paper of Higman (1967) which employs graph-theoretic
ideas very closely related to those of this chapter. The discovery of
sporadic simple groups as the automorphism groups of strongly regu-
lar graphs (for example by Higman and Sims (1968)) gave a powerful
impetus to work in this area. The formulation in terms of the proper-
ties of distance-transitivity and distance-regularity was developed by the
present author and some of his colleagues in the years 1969-1973, and
a consolidated account appeared in the first edition of this book (1974).
In the last twenty years an extensive literature has been accumulating.
The reader is referred to the now-standard text of Brouwer, Cohen and
Neumaier [BCN], which admirably covers the state of the art up to 1989,
and contains a bibliography of 800 items.

Additional Results
20a The cube graphs Thefc-cube,Qk, is the graph defined as follows:
the vertices of Qk are the 2fc symbols (ei,C2, • • • ,Cfc), where e* = 0 or
1 (1 < i < A:), and two vertices are adjacent when the symbols differ
in exactly one coordinate. The graph Qk (k > 2) is distance-transitive,
with degree k and diameter k, and the intersection array is
~\ K ^ tv "™ J. j K ™" Z j • . • j i } 1 ] Z ) O ) « • • J / C J .

20b The odd graphs yet again The odd graphs Ok (k > 2) are distance-
transitive, with degree k and diameter k — 1. The intersection array, in
the cases k = 21 — 1 and k = 21 respectively, is
{21 - 1,21 - 2,21 - 2,... ,1 + 1,1 + 1,1; 1,1,2,2,... ,1 - 1,1 - 1},

{21,21 - 1,21 - 1 , . . . , I + 1,1 + 1; 1,1,2,2,..., I - 1,1 - 1, / } .

162 Symmetry and regularity

20c A distance-regular graph which is not distance-transitive Let ^

denote the graph whose vertices are the 26 symbols a*, 6* (where i is an
integer modulo 13), and in which:
di and aj are adjacent •&• \i — j \ = 1,3,4;
bi and bj are adjacent «• \i — j \ — 2,5,6;
a* and bj are adjacent <=> i — j = 0,1,3,9.
Then $ is distance-regular, with diameter 2, and its intersection array
is {10,6; 1,4}. But ty is not distance-transitive; in fact there is no au-
tomorphism taking a vertex a, to a vertex bj (Adel'son-Velskii et al.
1969).
20d Strengthening the distance-transitivity condition A connected sim-
ple graph is r-ply transitive if, for any two ordered r-tuples of vertices
(xi,...,xr) and (yi,...,yr) satisfying d{xuxj) - #(&,%•) for all i,j,
there is an automorphism g for which g(xi) = yi (1 <i < r). Clearly, a
1-ply transitive graph is vertex-transitive, and a 2-ply transitive graph
is distance-transitive. Meredith (1976) showed that the only 3-ply tran-
sitive graphs with girth greater than 4 (equivalently ci = 1) are the
cycles.
20e 6-ply transitive graphs (Cameron 1980) The following is a com-
plete list of all 6-ply transitive graphs.
(i) The complete multipartite graphs with parts of equal size (including
the complete graphs as the case when the parts have size 1).
(ii) The complete bipartite graphs with the edges of a complete matching
deleted.
(iii) The cycles.
(iv) L(K3t3).
(v) The icosahedron.
(vi) The graph whose vertices are the 3-subsets of a 6-set, two vertices
being adjacent whenever they have two common members.

20f Strongly regular graphs and partial geometries A partial geometry

pg(s, t, a) is an incidence structure of points and lines such that every
line has s +1 points, every point is on t +1 lines, two distinct lines meet
in at most one point, and for every non-incident (point, line) pair (p, I)
there are a lines through p that meet I. The graph whose vertices are the
points, two being adjacent if they are collinear, is strongly regular with
parameters k = s(t +1), a — t(a — 1) + s — 1, c = a(t +1). Equivalently,
it is a distance-regular graph with intersection array
{s(t + 1), (s - a + l)t; 1, a(t + 1)}.
 Distance-transitive graphs 163

20g Symmetric designs as distance-regular graphs A symmetric design

with parameters (v,fc,A) is a set P of points and a set B of blocks such
that \P\ — \B\ = v, each block has k points and each point is infcblocks,
and each pair of points is in A blocks. It follows from the definition that
(v — 1)A = k(k— 1). When A = 1 a symmetric design is called a protective
plane.
The graph whose vertices are the points and blocks of a symmetric
design, two being adjacent when they are incident, is distance-regular
with intersection array
^/C,A/ I, K A; 1, A, fe j .

For example, when A = 1 we have the incidence graph of a projective

plane; the case fc = 3 is Heawood's graph 5(7) mentioned in 18h. If
the projective plane is Desarguesian (that is, if it can be coordinatized
using a finite field) then the corresponding graph is distance-transitive.

20h The classification problem for DT and DR graphs For each k > 3
there are only finitely many DT graphs with degree k. This has been
proved in several ways: see Cameron (1982) and Weiss (1985), for ex-
ample. For DR graphs the result has been established only in the case
fc = 3 (Biggs, Boshier and Shawe-Taylor (1986); see 21i).
For the general DR case the problem is to find an upper bound for
the diameter d in terms of fc. Such a result could be regarded as a
strengthening of the monotonicity conditions (2) and (3) of Proposition
20.4, in which we seek to bound the number of repeated values among
the columns (c,, ai,bi) of the intersection array. An important result on
these lines was obtained by Ivanov (1983).
 21

Feasibility of intersection arrays

In this chapter we shall study the following question. Suppose that an

arbitrary array of integers {k,bi,..., bd-i; 1, C2,..., c^} is given: when
is there a distance-regular graph with this as its intersection array?
The results obtained in the previous chapter provide some simple nec-
essary conditions. For example, part (1) of Proposition 20.4 yields an
explicit formula for the numbers hi = \Ti(v)\:
h == (kh ... k-O/tecs ...a) (2 < i < d).
These numbers must be integers, so we have a non-trivial constraint on
the intersection array. Similarly, the monotonicity conditions in parts
(2) and (3) of Proposition 20.4 must be satisfied.
There are also some elementary parity conditions. Let n = 1 + &i +
... + kd be the number of vertices of the putative graph, then if k is odd,
n must be even. That is, nk = 0 (mod 2). Similarly, considering the
induced subgraph defined by the vertices in I\(v), we see that feoj 5 0
(mod 2) for 1 < i < d, where Oj = k — 6j — Cj.
These conditions are quite restrictive, yet they are satisfied by many
arrays which are not realised by any graph. For example, {3,2,1; 1,1,3}
passes all these tests, and would represent a graph with degree 3, diam-
eter 3, and 12 vertices. In this case, simple (but special) arguments can
be used to prove that there is no graph. The main result of this chapter
is a general condition which rules out a multitude of examples of this
kind.
Recall that the adjacency algebra A(T) of a distance-regular graph T
 Feasibility of intersection arrays 165

has as a basis the d+1 distance matrices {A0) A x , . . . , A d }, which satisfy

AftAj = J2 ShijAj. This equation can be interpreted as saying that left-
multiplication by Ah, regarded as a linear mapping of A(T) with respect
to the given basis, is faithfully represented by the (d+ l)x(d+1) matrix
Bfc defined by

(This representation seems natural for our purposes although it is the

transpose of the one most commonly employed. Since the algebra A(F)
is commutative, the difference is immaterial.) The existence of this
representation is sufficiently important to justify a formal statement.
Proposition 21.1 The adjacency algebra A(T) of a distance-regular
graph F with diameter d can be faithfully represented by an algebra of
matrices with d+1 rows and columns. A basis for this representation is
the set {Bo,Bi,...,Bd}, where )ij is the intersection number
forh,i,j € {0, l,...,d}. •
The members of -4(F) can now be regarded as square matrices of size
d+1 (instead of n), a considerable simplification. What is more, the
matrix Bi alone is sufficient. To see this, we notice first that, since
(Bi)jj = suj, the matrix Bi is tridiagonal:
TO 1
k di C2
h a2 •

• • cd
• a dA
We shall often write B for Bi, and refer to B as the intersection matrix
of F. Note that it is just another way of writing the intersection array.
Now, since the matrices Bj are images of the matrices Ai under a faithful
representation, the equation obtained in Lemma 20.6 carries over:
BBi = 6i-iBi_i + aiBi + c i + i B i + 1 (1 < i < d - 1).
Consequently each B, is a polynomial in B with coefficients which de-
pend only on the entries of B. It follows from this (in theory) that A(T)
and the spectrum of F are determined by B, which in turn is determined
by the intersection array t(F). We shall now give an explicit demonstra-
tion of this fact.
Proposition 21.2 Let T be a distance-regular graph with degree k and
diameter d. Then V has d+1 distinct eigenvalues k — \Q, AI, ..., A^,
which are the eigenvalues of the intersection matrix B.
166 Symmetry and regularity

Proof We noted in Chapter 20 that T has exactly d + 1 distinct

eigenvalues. Since B is the image of the adjacency matrix A under a
faithful representation, the minimum polynomials of A and B coincide,
and so the eigenvalues of A are the same as those of B. D

Each eigenvalue A, common to A and B, is a simple eigenvalue of B,

since B is a matrix of size d + 1. However, the multiplicity m(A) of A
as an eigenvalue of A will usually be greater than one, since the sum of
the multiplicities is n, the number of vertices. We shall show how m(A)
can be calculated from B alone.
Let us regard A as an indeterminate, and define a sequence of polyno-
mials in A, with rational coefficients, by the recursion:
vo(\) = 1,

d+lvi+i{\) + (OJ - A)vi(A) + bi-iVi-i{X) = 0 (i = 1,2,...,d ~ 1).

The polynomial Vi(\) has degree i in A, and comparing the definition
with Lemma 20.6 we see that

Another interpretation of the sequence {UJ(A)} is as follows. If we

introduce the column vector v(A) = [«o(A),ui(A),... ^(A)]*, then the
defining equations are those which arise when we put t;o(A) = 1 and solve
the system Bv(A) = Av(A), using one row of B at a time, and stopping
at row d — 1. The last row of B gives rise to an equation representing
the condition that v(A) is an eigenvector of B corresponding to the
eigenvalue of A. The roots of this equation in A are the eigenvalues
Ao, A i , . . . , Ad of B, and so a right eigenvector v, corresponding to Aj
has components (VJ).,- = Vj(Xi).
It is convenient to consider also the left eigenvector Uj corresponding
to AJ; this is a row vector satisfying UjB = AjUj. We shall say that a
vector x is standard when XQ — 1.

Lemma 21.3 Suppose that m and Vj are standard left and right eigen-
vectors corresponding to the eigenvalue Aj o / B . Then (VJ)J = kj(\ii)j,
for alii, je {0,1,...,d}.
Proof Each eigenvalue of B is simple, and so there is a one-dimensional
space of corresponding eigenvectors. It follows that there are unique
standard eigenvectors u* and v*. (If (u,)o or (vi) 0 were zero, then the
tridiagonal form of B would imply that u» = 0, Vj = 0.)
Let K denote the diagonal matrix with diagonal entriesfco,fci,• . . , fe^.
Using the equations bi-ik^i = c,fcj (2 < i < d) we may check that B K
 Feasibility of intersection arrays 167

is a symmetric matrix; that is,

BK = (BK) ( - KB*.
Thus, if UjB = XiUi (0 < i < d), we have
B K u | = KB'ul = K(u i B) t = K(A*in)* = A<KuJ.
In other words, Ku- is a right eigenvector of B corresponding to A*. Also
(Ku')o = 1, and so by the uniqueness of v$, it follows that Ku* = v*.
D

We notice that, when i ^ I, the inner product (u,,v;) is zero, since

Ai(Uj, Vj) = UjBvj = Aj(Uj, Vj).

Our main result is that the inner product with i = I determines the
multiplicity m(Aj).
Theorem 21.4 With the notation above, the multiplicity of the eigen-
value Aj of a distance-regular graph with n vertices is
^) io<i<d)
Proof For i = 0 , 1 , . . . , d define
d

3=0
We can calculate the trace of Lj in two ways. First, the trace of Aj is
zero (j 7^ 0), and A o = I, so that
tr(L4) = (u i ) o tr(I) = n.
On the other hand, since Aj = u,-(A), the eigenvalues of Aj are Vj(X0),
..., Vj(Xd), with multiplicities m(A 0 ),..., m(X<i); consequently the trace
of Aj is £m(Aj)uj(A0- Thus

which gives the required result. D

In the context of our question about the realisability of a given ar-

ray, we shall view Theorem 21.4 in the following way. The numbers
n/(ui, Vj), which are completely determined by the array, represent mul-
tiplicities of the eigenvalues of the adjacency matrix of a supposed graph,
and consequently if there is such a graph, they must be positive integers.
This turns out to be a very powerful condition.
168 Symmetry and regularity

Definition 21.5 The array {k, bit..., bd-i; 1, c 2 , . . . , Cd} is feasible if

the following conditions are satisfied.
(1) The numbers kt = (kbi... &i-i)/(c 2 c 3 . . . Cj) are integers (2 < i < d).
(2) fc > 6i > . . . > 6d_i and 1 < c2 < . . . < c d .
(3) If n - 1 + k + &2 + . • • + kd and <n = k - bt - a (1 < i < d - 1),
ad — k — Cd, then nk = 0 (mod 2) and fc,aj = 0 (mod 2).
(4) The numbers n/(u$, v») are positive integers (0 < i < d).
It should be noted that the definition of feasibility given above is
a matter of convention. The conditions stated are not sufficient for
the existence of a graph with the given array, and indeed there are
many other, independent, 'feasibility' conditions. Some useful ones are
given in 21c, 21d, and 21e; the standard reference [BCN] provides
a comprehensive treatment. The four conditions which comprise our
definition of feasibility are chosen because they are particularly useful,
and any reasonable way of testing a given array will surely include them.

The four conditions are easy to apply in practice. The calculation of

n/(uj, Vi) is facilitated by Lemma 21.3, which implies that

n
3
For example, consider the array {3,2,1; 1,1,3} which, as we have al-
ready noted, satisfies the first three conditions. The eigenvalues of B
are 3 , - 1 and the roots of the quadratic equation A2 + A — 3 = 0. If
0 is one of the quadratic eigenvalues the corresponding eigenvector is
[1,0,-0,-1]*, and the 'multiplicity' is

12
/ ( l + J + J + I) = 24/(3 + °2) = 24/(6 " 0)'
which is clearly not an integer. Thus there is no graph with the given
array.
For a positive example, consider the array {2r,r — 1; 1,4} (r > 2),
for which the corresponding B matrix is

0 1 0
2r r 4
0 r - 1 2r - 4
It is easy to verify that k = 2r, k2 = \r{r - 1), n = \{r + l)(r + 2), so
that conditions (1), (2) and (3) of Definition 21.5 are fulfilled.
 Feasibility of intersection arrays 169

The eigenvalues of B are Ao = 2r, Ai = r - 2, A2 = - 2 , and the

calculation of the multiplicities goes as follows:
1 1 ' 1"
v0 = 2r , Vi = r - 2 v2 = - 2
>(r-l) 1-r 1

m(Ai) n
(ui, 1 + (r - 2) 2 /2r + (1 - r)2/\r{r - 1)

m(A2) = r 1
1 / i r ( r _ 1 N = 2~( ~
r
)(
(u 2 ,v 2 ) 1-
Since these values are integers, condition (4) is satisfied and the array is
feasible. In fact the array is realized by the triangle graph A r+ 2, as we
noted in Chapter 20. (The eigenvalues and multiplicities of this graph
were found in a different way in Chapter 3.)
Another example is the graph S representing the 27 lines on a cubic
surface (Chapter 8, p. 57). This is a distance-regular graph with diam-
eter 2 and intersection array {16,5; 1,8}, from which we may calculate
the spectrum:
_ _ / 1 6 4 -2s
Spec £ = ^ x 6 2 0
These examples have diameter 2, and so they are strongly regular
graphs. In that case the multiplicities can also be obtained by more
elementary methods (see 3d). But for a general distance-regular graph
the multiplicity formula is invaluable.

Additional Results

21a The spectra o/Qk and the Hamming graphs The eigenvalues of the
fc-cube Qk are A» = k - 2i (0 < i < k), with multiplicities m(Xi) - (*).
The fc-cube is the case q = 2 of the Hamming graph H(d,q), whose
vertices are the qd d-vectors with elements in a set of size q, two being
adjacent when they differ in just one coordinate. The graph H(d, q) is
distance-transitive, with intersection array
{d(q-l),(d-l)(q-l),...,(q-l); 1,2,..., d}.
The eigenvalues are d(q - 1) - qi, i = 0,1,...,d with multiplicities
(d\(q — 1)\ The intersection array determines the Hamming graph
H(d, q) uniquely, except when q = 4; in that case there are other graphs
with the same intersection array [BCN, p. 262].
170 Symmetry and regularity

21b The spectrum of Ok The eigenvalues of the odd graph Ok are

\i = (-lY(k-i) (0 < i <fe—1), and
m{ i) =

21c Elementary conditions on the intersection array The following

conditions must be satisfied by the intersection array of any distance-
regular graph. Proofs may be found in Biggs (1976).
(1) If a\ = 0 and a^^Q then 02 > C2.
(2) If ax = 1 theno 2 > c2.
(3) If a-i — 2 and fc is not a multiple of 3, then c2 > 2.
21d Integrality of all intersection numbers Since the matrices Bi are
the images of the A* under a faithful representation it follows that they
satisfy the relation B, = u*(B) (0 < i < d). Since (Bh)%j is the number
Shij, it follows that each of the matrices computed by means of this
formula must have integral entries.

21e The Krein conditions Define

(0 < i < d),
n
where the L4 are as in the proof of Theorem 21.4. The E, are mutually
orthogonal, idempotent, and form a basis for the adjacency algebra.
This algebra is closed under the pointwise product o of matrices, because
Ai o Aj — SijAj. It follows that there are real numbers qhij such that

Scott (1973) observed that these Krein parameters must be non-negative.

Thus we have a new set of 'feasibility' conditions, which can be stated
explicitly as follows:

=E-( r=0

21f An array which is not realisable The array {9,8; 1,4} is feasible
in the sense of Definition 21.5. We have

and the 'eigenvalues' are 9,1, —5 with 'multiplicities' 1,21,6 respectively.

The conditions given in 21c are satisfied, and also 21d since
0 1 0 0 0 1
B— 9 0 4 , B 2 = «a(B) = 0 8 5
0 8 5 18 10 12
 Feasibility of intersection arrays 171

However the Krein condition g222 > 0, in the notation of 21e, does not
hold. An elementary proof that this array is not realisable was given by
Biggs (1970).

21g Feasibility conditions for strongly regular graphs A strongly regu-

lar graph, as defined in 3c, is a distance-regular graph with intersection
array {k, k — a — 1; l,c}. The eigenvalues and their multiplicities can
be computed by the elementary methods described in 3d, or by the
general methods described in this chapter. A good survey is given by
Seidel (1979). In addition to the 'feasibility' conditions which hold for
distance-regular graphs in general, there is a useful 'absolute bound'
n < -m(m + 3),
where n is the number of vertices and m is the multiplicity of either one
of the eigenvalues A ^ k. For example, this test shows that the array
considered in 21f is not realisable.

21hThe friendship theorem If, in a finite set of people, each pair of

people has precisely one common friend, then someone is everyone's
friend. (Friendship is interpreted as a symmetric, irreflexive relation.)
The result may be proved as follows. Let F denote the graph whose
vertices represent people and whose edges join friends. Then F is either
a graph consisting of a number of triangles all with a common vertex or a
strongly regular graph with intersection array {k,k — 2;1,1}. The array
is not feasible, so the first possibility must hold. This is an unpublished
proof of G. Higman; for other proofs see Hammersley (1981).

21i Distance-regular and distance-transitive graphs with degree 3 Biggs

and Smith (1971) proved that there are exactly 12 distance-transitive
graphs with degree 3. They are: (i) the symmetric cubic graphs with
n < 30 vertices listed in 18h, with the exception of P(8,3) and P(12,5);
(ii) the threefold covering of 5(3) with n = 90 vertices described in 19c;
(iii) the expansion of H with n = 102 vertices described in 18g.
Biggs, Boshier and Shawe-Taylor (1986) showed that in the distance-
regular case there is just one other graph, which has 126 vertices (see
23b).

21j Perfect codes in distance-regular graphs The definition of a per-

fect e-code in a graph was given in 3k. Let Vi(X) be the polynomials
associated with a distance-regular graph F, and let
Xi(X) = «o(A) + «i(A) + . . . + Vi(X) (0<i< d).
172 Symmetry and regularity

If there is a perfect e-code in F then xe(A) is a factor of x<i(A) in the ring

of polynomials with rational coefficients. This implies that the zeros of
xe(X) must be eigenvalues of F, This result was first established by S.P.
Lloyd in the 'classical' case of a cube or Hamming graph. Biggs (1973c)
gave a proof for the general distance-transitive case and Delsarte (1973)
proved similar results in a more general context.

21k Sporadic groups and graphs Several of the sporadic simple groups
can be represented as the automorphism group of a distance-transitive
graph. A typical example is the distance-transitive graph with 266 ver-
tices which has degree 11, diameter 4, and intersection array {11,10,6,1;
1,1,5,11}. The automorphism group of this graph is Janko's simple
group of order 175 560. As usual, the reader should consult [BCN] for a
full account.

211 The permutation character If F is a distance-transitive graph with

diameter d, then the permutation character \ corresponding to the rep-
resentation of Aut(F) on VT is the sum of d + 1 irreducible characters:
X = 1 + Xi + • • • + Xd
and the labelling can be chosen so that the degree of \i is m(Ai) (0 <
i < d). This can be deduced from the results of Wielandt (1964); see
also [BCN, p. 137].
 22

Imprimitivity

In this chapter we investigate the relationship between primitivity and

distance-transitivity. We shall prove that the automorphism group of a
distance-transitive graph can act imprimitively in only two ways, both
of which have simple characterizations in terms of the structure of the
graph.
We begin by summarizing some terminology. If G is a group of per-
mutations of a set X, a block B is a subset of X such that B and g(B)
are either disjoint or identical, for each g in G. If G is transitive on X,
then we say that the permutation group (X, G) is primitive if the only
blocks are the trivial blocks, that is, those with cardinality 0, 1 or \X\.
If B is a non-trivial block and G is transitive on X, then each g(B) is a
block, and the distinct blocks g(B) form a partition of X which we refer
to as a block system. Further, G acts transitively on these blocks.
A graph F is said to be primitive or imprimitive according as the
group G = Aut(F) acting on VT has the corresponding property. For
example, the ladder graph L3 is imprimitive: there is a block system
with two blocks, the vertices of the triangles in L3.
Proposition 22.1 Let T be a connected graph for which the group
of automorphisms acts imprimitively and symmetrically (in the sense of
Definition 15.5). Then a block system for the action of Aut(F) on VT
must be a colour-partition ofT.
Proof Suppose that VT is partitioned by the block system
174 Symmetry and regularity

Then we may select one block, call it C, and elements g^ in Aut(r),

such that
£«> = gMc (1 < i < I).
Suppose C contains two adjacent vertices u and v. Since F is symmetric,
for each vertex w adjacent to u there is an automorphism g such that
g(u) = u and g(v) — w. Then u belongs to C n <7(C), and C is a block,
so C = g(C) and w belongs to C. Since w was any vertex adjacent to v,
the set Fi (u) is contained in C, and by repeating the argument we can
show that F 2 (u),T 3 (u),... are contained in C. Since F is connected, we
have C = VT. This contradicts the hypothesis of imprimitivity, and so
our assumption that C contains a pair of adjacent vertices is false. Thus
C is a colour-class and since each block B^ is the image of C under an
automorphism, the block system is a colour-partition. •

This result is false if we assume only that the graph is vertex-transitive,

rather than symmetric. The ladder graph L3 mentioned above provides
a counter-example.

The rest of this chapter is devoted to an investigation of the relation-

ship between primitivity and distance-transitivity. We shall show that,
in an imprimitive distance-transitive graph, the vertex-colouring induced
by a block system is either a 2-colouring or a colouring of another quite
specific kind.
Lemma 22.2 Let F be a distance-transitive graph with diameter d,
and suppose B is a block for the action o/Aut(F) on VT. If B contains
two vertices u and v such that d(u,v) = j (1 < j < d), then B contains
all the sets rrj(u), where r is an integer satisfying 0 < rj < d.
Proof Let w be any vertex in Tj(u). Since F is distance-transitive
there is an automorphism g such that g(u) — u and g(v) = w. Thus u
is in B fl g(B), and since B is a block, B = g(B) and w is in B. So
r » c B.
If z is in F2j(u), there is a vertex y e Tj(u) for which d(y,z) = j .
Since d(z,y) = d(u,y), and both u and y are in B, it follows by a
repetition of the argument in the previous paragraph that z is in B,
and so F2j(u) C B. Further repetitions of the argument show that
TTj (u) C B for each r such that rj <d. •

For the rest of this chapter we use the symbol d' to denote the largest
even integer not exceeding d.
 Imprimitivity 175

Proposition 22.3 Let F be a distance-transitive graph with diameter

d and degree k > 3. Then a non-trivial block for the action of Aut(F)
on VT which contains the vertex u must be one of the following sets:
Ba(u) = {u}urd(u), Bb(u) = {u}ur2(«)ur4(«)u...urv(u).
Proof Suppose B is a non-trivial block containing u, and is not the
set Ba(u). Then B contains a vertex v ^ u such that d(u, v) — j < d,
and consequently Tj(u) C B.
Consider the numbers Cj,aj,bj in the intersection array of F. We
must have aj = 0, because if a, were non-zero then B would contain
two adjacent vertices, which is impossible by Proposition 22.1. Since
c
j + a-j +bj = k >Z,
one of Cj,bj is at least 2. Prom parts (2) and (3) of Proposition 20.4
it follows that one of Cj+\,bj-i is at least 2, and consequently Tj(u)
contains a pair of vertices at distance 2. Thus B contains the set Bb(u).
If it contained any other vertices, it would contain two adjacent vertices
and would be the trivial block VT. We deduce that B — Bb(u), as
required. •

The cube Qz is an example of an imprimitive distance-transitive graph

with diameter d — 3, so d! — 2 here. One block system consists of four
sets of the form {u} U ^ ( u ) of size two, while another block system
consists of two sets of the form {u} U F2(u) of size four. This example
illustrates the fact that both types of imprimitivity allowed by Proposi-
tion 22.3 can occur in the same graph.
Another instructive example is the cocktail-party graph CP(s), as
defined on p. 17. Here there are s blocks u U ^ ( u ) , each of size two,
and since d' = d = 2 these blocks are simultaneously of type Ba (u) and
Bb(u). The next lemma clears up this case.

Lemma 22.4 Let T be a distance-transitive graph with girth 3 and

diameter d > 2, in which the set
Bb(u) ={«}ur 2 (ti)U...U Td, (u)
is a block. Then d — 2 and consequently Bb(u) = {u} U ^ ( u ) = Ba(u).
Proof Since F contains triangles and is distance-transitive, every or-
dered pair of adjacent vertices belongs to a triangle. Choose adjacent
vertices v\ € T\{u), V2 € F2(u); then there is some vertex z such that
v\v2z is a triangle. If z were in F2(u), then 2?&(u) would contain adjacent
vertices, contrary to Proposition 22.1. Thus z must be in Fi(w).
If d > 3, we can find a vertex v3 6 F3(M) which is adjacent to v2
176 Symmetry and regularity

Figure 18: illustrating the proof of Lemma 22.4.

(Figure 18). But then ^(1*3) contains the adjacent vertices v\ and z,
and if h is an automorphism of F taking u to v3, h(Bb(u)) is a block
containing adjacent vertices, again contradicting Proposition 22.1. Thus
we must have d = 2. •

Proposition 22.5 Let T be a distance-transitive graph with diameter

d>3 and degree k > 3. Then
x - Bb(u) = {u} u r2(u) u . . . u vd- (u)
is a block if and only ifTis bipartite.
Proof Suppose F is bipartite. If X is not a block, then there is
an automorphism g of F such that X and g(X) intersect but are not
identical. This would imply that there are vertices x and y in X, for
which g(x) € X but g(y) £ X, so that d(x,y) is even and d(g(x),g(y))
is odd. From this contradiction we conclude that X is a block.
Conversely, suppose X is a block. A minimal odd cycle in F has length
2j + 1 greater than 3, by Lemma 22.4. We may suppose this cycle to be
uu\... W1V1V2W2 • •. U2U, where
ui,u-2 e Ti(u), wi,w2 € Fj_i(w), vi,v2 6 Tj(u),
and if j = 2, then «i = w\ and u2 — w2. If j is even, then X contains
the adjacent vertices vi and v2, and so X = VT, a contradiction. If
j is odd we have, for i — 1,2, d(u,Wi) — d(ui,Vi), and so there is an
automorphism hi taking u to u* and Wi to w*. Thus Yt — hi{X)\s& block
containing Ui and Uj. But, since F contains no triangles, d(ui,u2) = 2
and so u2 £ Vi. Consequently Fi = Y2 and we have adjacent vertices
^i,^2 in Y\, so that Yx = FF, X = VF. From this contradiction it
follows that F has no odd cycles and is bipartite. •

Lemma 22.4 and Proposition 22.5 lead to the conclusion that, if a

block of the type Bb(u) exists in a distance-transitive graph F, then
either d = 2, in which case the block is also of type Ba(u), or d > 3 and
F is bipartite. The complete tripartite graphs KT,T,r are examples of the
first case and are clearly not bipartite.
 Imprimitivity 177

We shall now show that graphs which have blocks of type Ba (u) can
also be given a simple graph-theoretical characterization.
Definition 22.6 A graph of diameter d is said to be antipodal if,
for any vertices u, v, w such that d(u, v) = d(u, w) = d, it follows that
d(v, w) = d or v = w.
The cubes Qk are trivially antipodal, since every vertex has a unique
vertex at maximum distance from it; these graphs are at the same time
bipartite. The dodecahedron is also trivially antipodal, but it is not
bipartite. Examples of graphs which are non-trivially antipodal and not
bipartite are the complete tripartite graphs Kr>rir, which have diameter
2, and the line graph of Petersen's graph, which has diameter 3.
Proposition 22.7 A distance-transitive graph F of diameter d has a
block Ba(u) = {u} U r<i(u) if and only ifT is antipodal.
Proof Suppose F is antipodal. Then if x is in Ba(u), it follows that
Ba(u) = {x} U Td(x) = Ba(x). Consequently, if g is any automorphism
of F, and z is in Ba(u) r\g(Ba(u)) then
Ba(u) = {z}uTd(z)=g(Ba(u)),
so that Ba{u) is a block.
Conversely, suppose Ba(u) is a block, and v,w belong to F^u) (v ^
w). Let d{v,w) = j (1 < j < d), and let h be any automorphism
of F such that h(v) — u. Then h(w) is in Fj(w). Also h(w) belongs
to h(Ba(u)) — Ba(u), since h{Ba(u)) intersects Ba(u) (u is in both
sets) and Ba(u) is a block. This is impossible for 1 < j < d, so that
d(v, w) — d, and F is antipodal.
Theorem 22.8 (Smith 1971) An imprimitive distance-transitive graph
with degree k > 3 is either bipartite or antipodal. (Both possibilities can
occur in the same graph.)
Proof A non-trivial block is either of the type Ba(u) or Bb{u). In the
case of a block of type Bb(u), Proposition 22.5 tells us that either the
graph is bipartite, or its diameter is less than 3. If the diameter is 1,
then the graph is complete, and consequently primitive. If the diameter
is 2, a block of type Bb(u) is also of type Ba(u). Consequently, if the
graph is not bipartite, it must be antipodal. •

The notion of primitivity can be defined without reference to a group

action, in the following way. Given a graph F with diameter d, let
Fj (1 < i < d) be the graph whose vertices are the same as those of F,
two vertices being adjacent in F* if and only if they are at distance i in F.
178 Symmetry and regularity

Then F is said to be imprimitive if any of the graphs Fi is disconnected.

It is easy to see that for a bipartite graph F2 has two components,
and for an antipodal graph F^ is the disjoint union of complete graphs.
Using this definition Smith's theorem and its proof can be extended to
distance-regular graphs (see [BCN, p. 140]).
The complete graphs are primitive and distance-transitive. Other
families with the same properties are line graphs of a certain kind. Apart
from these families, primitive distance-transitive graphs are scarce, and
we give them a special name.
Definition 22.9 An automorphic graph is a distance-transitive graph
which is primitive and not a complete graph or a line graph.
For instance, of the 12 distance-transitive graphs with degree 3 (21i),
only three are automorphic. They are Petersen's graph, Coxeter's graph
(the expansion of Y with 28 vertices), and the expansion of H with 102
vertices. The odd graph O4 is the only automorphic graph with degree
4. Many more details may be found in [BCN].

Additional Results
22a The derived graph of an antipodal graph Let F be a distance-
transitive antipodal graph, with degree k and diameter d > 2. Define
the derived graph F' by taking the vertices of F' to be the blocks {u} U
Td(u) in F, two blocks being joined in F' whenever they contain adjacent
vertices of F. Then F' is a distance-transitive graph with degree k and
diameter equal to [d/2j (Smith 1971).
22b The icosahedron and the dodecahedron The icosahedron / and the
dodecahedron D are distance-transitive with
i(I) = {5,2,1; 1,2,5}; i(D) = {3,2,1,1,1; 1,1,1,2,3}.
Both graphs are antipodal, and the derived graphs are K6 and O3.
22c The intersection array of an antipodal covering We can look at
the construction in 22a from the opposite point of view, as follows. A
distance-regular graph F is an antipodal r-fold covering of the distance-
regular graph F if F is antipodal, its derived graph is F, and |VF| =
r|VT|. It turns out that the intersection array of F is related to the
intersection array {k, b\,..., 6^-1; 1,C2,...,Cd} of F in one of two ways.
Either (i) F has even diameter 2d > 2 and
 Imprimitivity 179

or (ii) F has odd diameter Id + 1 and, for some positive integer t such
that (r — l)t < min(6<j_i,a(j) and cd < t, we have
t(f) = {k, &!,..., &d_i, (r - lji.cj.cd-i, •. •, c 2 ,1;
l , c 2 , . . . ,Cd,t,bd-i, ••• ,bi,k}.
Clearly, the total number of possibilities is finite, and r < k in any case.
22d Antipodal coverings of Kk,k Let F be a distance-regular graph
which is an antipodal r-fold covering of Kk,k- Then it follows from 22c
that r must divide k, and if rt — k the intersection array for F is
{k,k - l,k - t,l;l,t,k - l,k}.
This array is feasible (provided that r divides k) and the spectrum of F

SpecF=(* ^ 2k°_1 fc;^ "'

In the case r = k, the existence of F implies the existence of a projective
plane of order k (Gardiner 1974).
22e Distance-regular graphs with diameter three A distance-regular
graph with diameter three is antipodal, bipartite, or primitive (in the
extended sense defined on p. 177). In the antipodal case the intersection
array is of the form {k, (r —1)7,1; 1,7, k}, and the graph is an antipodal
r-fold covering of Kk+i. This case has been the subject of several papers:
see Biggs (1982c), Cameron (1991), Godsil and Hensel (1992). In the
bipartite case the intersection array is of the form {k, k— l,k-X; 1, X,k},
and the graph is the incidence graph of a symmetric 2-design with pa-
rameters (v, k, A), where v — k(k — 1)/A+1. Several families of primitive
graphs are known, and some sporadic ones [BCN, pp. 425-431].
22f An automorphic graph with k = 5 and d — 3 Let L = {a, b, c, d, e, / } ,
and TV = {1,2,3,4,5,6}. The following table establishes a bijection be-
tween the 15 single-transpositions on L, and the 15 triple-transpositions
on N.
(O6)H + (15)(23)(46) (ac)i--> (14)(26)(35) (ad)v-•+ (13)(24)(56)
(ae) t-- (12)(36)(45) -» (16)(25)(34) (6c) H-> (12)(34)(56)
(bd)^ - (14)(25)(36) («*)- ^ (16)(24)(35) (bf)- -, (13)(26)(45)
(cd)^ -> (16)(23)(45) (ce) H- (13)(25)(46) (c/)" -. (15)(24)(36)
(de)K - (15)(26)(34) - (12)(35)(46) ( e / ) - - (14)(23)(56)
Define a graph F whose vertex-set is L x N, and in which (h,n\) is
adjacent to (h,n2) if and only if the transposition (ri\,n2) is one of
those corresponding to (h,fo)- Then F is an automorphic graph with
degree 5 and diameter 3. Its intersection array is {5,4,2; 1,1,4} and its
automorphism group is Aut S$.
 23

Minimal regular graphs with given girth

Results on the feasibility of intersection arrays can be applied to a wide

range of combinatorial problems. The last chapter of this book deals
with a graph-theoretical problem which has been the subject of much
research. We shall study regular graphs whose degree (k > 3) and girth
(g > 3) are given. For all such values of k and g there is at least one
graph with these properties (Sachs 1963), and so it makes sense to ask for
the smallest one. We note that when k = 2 the cycle graphs provide the
complete answer to the problem, and so we shall be concerned primarily
with the case k > 3.

Proposition 23.1 (1) The number of vertices in a graph with degree

k and odd girth g — 2d + 1 is at least
no(k,g) = l + k + k(k-l) + ... + k(k- l)^g-3l
If there is such a graph, having exactly no(k,g) vertices, then it is
distance-regular with diameter d, and its intersection array is
{fc,fc-l,fc-l,...,fc-l;l,l,l,...,l}.
(2) The number of vertices in a graph with degree k and even girth g = 2d
is at least
no(k,g) = l + k
If there is such a graph, having exactly no(k,g) vertices, then it is bipar-
tite and distance-regular with diameter d; its intersection array is
\k, k ~ 1, k 1,..., K 1 ; 1 , 1 , 1 , . . . , 1 , kj.
 Minimal regular graphs with given girth 181

Proof (1) Suppose that F is a graph with degree k and girth g = 2d+l,
and let (u, v) be any pair of vertices such that d(u, v) = j (1 < j < d).
The number of vertices in Tj^i(v) adjacent to u is 1, otherwise we
should have a cycle of length at most 2j < 2d + 1 in F. Using the
standard notation (Definition 20.5) we have shown the existence of the
numbers c\ = l,...,Cd = 1. Similarly, if 1 < j < d, then there are
no vertices in Tj(v) adjacent to u, otherwise we should have a cycle of
length at most 2j + l < 2cf+l. This means that a,j = 0 and consequently
bj = k — aj — Cj = k — 1, for 1 < j < d. It follows that the diameter of
F is at least d, and that F has at least no(k, g) vertices. If F has just
n
o{k,g) vertices, its diameter must be precisely d, which implies that
a<j = 0, and F has the stated intersection array.
(2) In this case the argument proceeds as in (1), except that c<j may
be greater than one. Now the recurrence for the numbers ki = |Fj(v)|
shows that kd is smallest when a = k; if this is so, then F has at least
no(k,g) vertices. If F has exactly no(fc, g) vertices, then its diameter is
d, and it has the stated intersection array. The form of this array shows
that F has no odd cycles, and so it is bipartite. •

Definition 23.2 A graph with degree k, girth g, and such that there
are no smaller graphs with the same degree and girth, is called a (k, g)-
cage. A (k,g)-cage with no(k,g) vertices is said to be a Moore graph if
g is odd, and a generalized polygon graph if g is even. (The reasons for
the apparently bizarre terminology are historical, and may be found in
the references given below.)
We have already remarked that a (k, g)-cage exists for all k > 3 and
g > 3. For example, Petersen's graph O3 is the unique (3,5)-cage: it
has 10 vertices and no(3,5) = 10, so it is a Moore graph. On the other
hand, the unique (3,7)-cage has 24 vertices (see 23c) and no(3,7) = 22,
so there is no Moore graph in this case. The main result of this chapter
is that Moore graphs and generalized polygon graphs are very rare.
In the cases g = 3 and g — 4 the intersection arrays in question are
{fc;l} and {fc,fc-l;l,jfc}
and these are feasible for all k > 3. It is very easy to see that each
array has a unique realisation - the complete graph K^+i and the com-
plete bipartite graph Kkik, respectively. Thus, when g = 3 we have a
unique Moore graph Kk+i, and when g = 4 we have a unique generalized
polygon graph Kk,k-
When g > 5 the problem is much more subtle, both in the technical
details and in the nature of the solution. The results are due to a number
182 Symmetry and regularity

of mathematicians. The generalized polygon case was essentially solved

by Feit and Higman (1964); the Moore graph case was investigated by
Hoffman and Singleton (1960), Vijayan (1972), Damerell (1973) and
Bannai and Ito (1973).
We shall apply the algebraic techniques developed in Chapter 21 to
both cases, in a uniform manner. Specifically we investigate the feasi-
bility of the intersection matrix
ro 1
k 0 1
Jfc-1 0
-1
1
0 c
fc-1 k-cj
which subsumes, by putting c = 1 and c = k, the intersection matrices
of Moore graphs and generalized polygon graphs.
Suppose that A is an eigenvalue of B and that the corresponding
standard left eigenvector is u(A) = [UQ(A), «I(A), ..., u<j(A)]. Then, from
the equations u(A)B = Au(A) and uo(A) = 1, we deduce that u\(X) =
X/k and

(**) cud-i(X) + (k - c - \)ud{\) = 0.

The equations (*) give a recursion which enables us to express Ui(\)
as a polynomial of degree i in A for 0 < i < d. The equation (**)
then becomes a polynomial equation of degree d + 1 in A. In fact (**)
represents the condition that A is an eigenvalue; it is the characteristic
equation of B.
Put q = y/k — 1 and suppose that |A| < 2q, so that we may write
A = 2q cos a for some a, 0 < a < ir (this assumption will be justified in
the course of the ensuing argument). The solution to the recursion (*)
can be found explicity:
q2 sin(i + l ) a — sin(i —
(1 < i < d).
kq1 sin a
Lemma 23.3 With the above notation, the number 2qcosa is an
eigenvalue of B if and only if
'c-V
qsm(d+ sin(d - l)a = 0.
Proof The stated equation results from substituting the explicit forms
of Ud-i and Ud in the equation (**), which is the characteristic equation
of B. •
 Minimal regular graphs with given girth 183

Proposition 23.4 (1) Let g = 2d and suppose F is a generalized

polygon graph for the values (k,g). Then F has d+l distinct eigenvalues:
k,-k, 2qcosirj/d (j = 1,2,... ,d - 1).
(2) Let g = 2d + 1 and suppose F is a Moore graph for the values (k, g).
Then F has d+l distinct eigenvalues:
k, 2gcosa, (j = 1,2,... ,d),
where the numbers QJ, ... ,otd o-re the distinct solutions in the interval
0 < a < 7T of the equation qsin(d + l)a + sin da = 0.
Proof (1) The existence of the eigenvalues k and — k follows from the
fact that F isfc-regularand bipartite. Now the eigenvalues of F are (by
Proposition 21.2) the d + l eigenvalues of its intersection matrix, which
is the matrix given above with c = k. In that case, A = 2qcosa is an
eigenvalue of B if and only if
qsin(d + l)a + A; sin da + qsin(d — l)a = 0.
This reduces to (2gcosa + k) sin da — 0, and since \k/2q\ > 1 when k >
3, the only possibility is that sin da = 0. Thus in the range 0 < a < TT
there are d — 1 solutions a = irj/d, corresponding to j = 1,..., d — 1,
and we have the required total of d + 1 eigenvalues in all.
(2) Since F is fc-regular, k is an eigenvalue. As in (1), we now seek
eigenvalues A = 2gcosa of B, this time with c = 1. The equation of
Lemma 23.3 reduces to
A = q sin(d + l)a + sin da = 0.
For 1 < j < d, A is strictly positive at 6j = (j — ^)7r/(d+1) and strictly
negative at cpj = (j + ^)n/(d +1). Hence there is a zero aj of A in each
one of the d intervals (#,, 4>j). Thus we have the required total of d + 1
eigenvalues in all. •
We now have enough information to calculate the multiplicities of the
eigenvalues and to test the feasibility of the corresponding intersection
array. Suppose that A is an eigenvalue of B. The multiplicity of A as
an eigenvalue of the putative graph is given by Theorem 21.4: m(A) =
n/(u(A), v(A)). We shall use this in the form m(A) = n/^fcjU,(A) 2 .
For our matrix B we have ko = 1, h = k(k - I)1"1 (1 < i < d — 1), and
kd = c~lk{k — I)"*"1. Also, for an eigenvalue A = 2qcosa we have
2

kq1 sin a
= (2hksin2a)~1(E + Fcos2ia + Gsin2ia) (1 < i < d),
where we have written
h = q =fc— 1, E = (h +1) — 2/icos2a,
184 Symmetry and regularity

F = 2h-(h2 + l)cos2a, G = (h2 - 1)sin2a.

Allowing for the anomalous form of kd by means of a compensating term,
we can sum the trigonometric series involved in ^kiUi{X)2 and obtain:

1 + (2hksin2 a ) - 1 \dE + {Fcos(d + l)a + Gsin(d + 1

sin a

) .

Fortunately this expression can be simplified considerably in the two

cases, c = 1 and c = k, which are of particular interest.

Proposition 23.5 IfA ^ ±fc is an eigenvalue of a generalized polygon

graph with girth g = 2d, then its multiplicity is given by
... nk (4h-\2

If X ^ k is an eigenvalue of a Moore graph with girth g = 2d+ 1, then

its multiplicity is given by

)
* 7
Proof In the case of even girth, c = k and we know that A = 2q cos a
is an eigenvalue if and only if sin da = 0. In this case the expression for
J2kiUi(\)2 becomes
1 + (2/ifc sin2 a)-l[dE + hk'1 (E + F)} = (2hk sin2 a^dE.
On putting 2d = g, A = 2qcosa this leads to the formula given.
In the case of odd girth, c = 1, and we know that A = 2gcosa is an
eigenvalue if and only if
qsin(d+ l)a + sin da = 0.
From this equation we have
—osina —osina
tan da = 1 + q cos a; sin da =

sin(d + l)a = —? ; cos(d+l)a =

Vk + A Vfc + A
Substituting for the relevant quantities in the general expression, and
putting g = 2d + 1, we obtain, after some algebraic manipulation, the
stated formula. •

We are now ready for the main theorem, which is the result of the
combined efforts of the mathematicians mentioned earlier in this chapter.
 Minimal regular graphs with given girth 185

Theorem 23.6 The intersection array for a generalized polygon graph

with k > 3, g > 4 is feasible if and only if g € {4,6,8,12}. The inter-
section array for a Moore graph with k > 3, g > 5 is feasible if and only
if g = 5 and k£ {3,7,57}.
Proof Suppose g is even, g = Id. Then a generalized polygon graph
has d *- 1 eigenvalues Xj = 2q COS(TTj/d) with multiplicities

m(Ajj
~ g \k*-Xj)-
If m(Aj) is a positive integer, Af is rational, which means that cos27r/d
is rational. But it is well known (see for example, Irrational Numbers by
I. Niven (Wiley, 1956), p. 37) that this is so if and only if d G {2,3,4,6}.
The case when g is odd presents more problems. We shall deal with
g = 5 and g = 7 separately, and then dispose of g > 9. Suppose g = 5.
Then the characteristic equation
g sin 3a + sin 2a = 0
reduces, in terms of A = 2qcosa, to A2 + A — (k — 1) = 0. Thus there
are two eigenvalues Ai = |(—1 + >/D) and A2 = \{—1 — VAD)! where
Z? = 4fc — 3. We have n = 1 + k2 and putting this in the formula for
m(A) we get
m _ (fc + fc3)(4fc-4-A2)
W
(Jfe-A)(6*-2 + 5A)'
If v^D is irrational, we multiply out the expression above, substitute
A = \ ( - 1 ± V^D) and equate the coefficients of v^D. This gives 5m + /c —
2 = k + k3, where m = m(Ai) = m(A2). But there are three eigenvalues
in all: k, Ai, A2 with multiplicities 1,m,m; hence 1 + 2m = n = 1 + k2.
Thus 5fc2 — 4 = 2k3, which has no solution for A; > 3. Consequently VD
must be rational, s = y/~D, say. Then k = \(s2 + 3 ) and substituting for
Ai and k in terms of s in the expression for mi = m(Ai) we obtain the
following polynomial equation in s:
s5 + s4 + 6s 3 - 2s 2 + (9 - mi)s - 15 = 0.
It follows that s must be a divisor of 15, and the possibilities are s =
1,3,5,15, giving k = 1,3,5,57. The first possibility is clearly absurd,
but the three others do lead to feasible intersection arrays.
Suppose g = 7. Then the characteristic equation
q sin 4a + sin 3a = 0
reduces, in terms of A = 2qcosa, to A3 + A2 — 2(k-l)X-(k-l) = 0. This
equation has no rational roots (and consequently no integral roots), since
we may write it in the form k — 1 = A2(A + 1)/(2A + 1 ) , and if any prime
186 Symmetry and regularity

divisor of 2 A + 1 divides x = X or A + 1 it must divide 2 A + 1 — a: = A + 1 or

A, which is impossible. So the roots Ai, A2, A3 are all irrational, and their
multiplicities are all equal, to m say. Then 1 + 3m = n = l + k — k2 + k3,
whereas k + m(Xi + A2 + A3) = trA = 0. But Ai + A2 + A3 = — 1, hence

m = k = -(A; 3 - fc2 + fc),

which is impossible for k > 3. Thus there are no Moore graphs when
5 = 7.
Suppose g > 9. We obtain a contradiction here by proving first that
— 1 < Ai + A<f < 0, and then showing that all eigenvalues must in fact
be integers. (The argument just fails in the case fc = 3, g — 9, but this
can be discarded by an explicit calculation of the 'multiplicities'.)
Let ai (1 < i < d) be the roots of
A = qsin(d + \)a + sin da = 0,
and set u> = n/(d+1). The proof of Proposition 23.4 showed that a\ lies
between ui/2 and 3w/2, and these bounds can be improved by noting
that A is positive at w and negative at w(l + l/2q). Thus w < ai <
<JJ(1 + l/2q) and
0 < 2gcoso; — 2q'cosa:i < 2gcosw - 2gcosw(l + l/2g)
= 2gcosw(l - cosuj/2q) + 2qsinojsm(uj/2q)
2
<2qx - (

In a similar way it can be shown that dw < ad < u>(d + l/2g), and
0 < 2q cos duj — 2q cos ad < us2.
Adding the two inequalities, and noting that
Ai = 2qcosa\, \d = 2qcosad, cosdu =—cosui,
we have
-9w 2 /4 < Ai + Xd < 0.
2 2 2 2 2
Now w = ?r /(d + I) < TT /5 < 4/9, so - 1 < Ai + Xd < 0, as
promised.
To show that the eigenvalues must be integers we note first that since
the characteristic equation is monic with integer coefficients, the eigen-
values are algebraic integers. The formula for m(X) is the quotient of
two quadratic expressions in A, and so m(X) is integral only if A is, at
worst, a quadratic irrational. Suppose A is a quadratic irrational. Then
R(X) = gm(X)/nk = (4ft - X2)/(k - X)(f + A)
 Minimal regular graphs with given girth 187

is rational number, and this equation can be written in the form

(R(X) - 1)A2 + R(X)(f - k)X - (R(X)fk - Ah) = 0.
But this must be a multiple of the minimal equation for A, which is
monic with integer coefficients. In particular,
(/ - k)R{X) Ah - A* fk-Ah
Where
** ~ /-fc '
must be an integer. However, / = k + (k — 2)/g > fc, so t > {k2 —
Ah)/(f — k) = <?(fc — 2), and consequently |A — t\ > g(k — 2) — k, since
|A| < k. Thus

for all k > 3, g > 9 (except when A; = 3, # = 9, as we have already

noted). Since S(X) is to be an integer, we must have S(X) = 0, which
leads to the absurdity R(X) = m(X) = 0. Thus all eigenvalues A must be
integers, which is incompatible with the inequality — 1 < Ai + A,j < 0,
and consequently disposes of all cases with g > 9. •

The question of the existence of graphs allowed by Theorem 23.6 is a

difficult one, and it contains some celebrated unsolved problems. In the
case of even girth g = 2d, we can relate the problem to existence of a
structure known as a generalized d-gon, defined as follows.
Let (P, L, I) be an incidence system consisting of two disjoint finite
sets, P (points) and L (lines), and an incidence relation I between points
and lines. A sequence whose terms are alternately points and lines, each
term being incident with its successor, is called a chain; it is a proper
chain if there are no repeated terms, except possibly when the first
and last terms are identical (when we speak of a closed chain). A (non-
degenerate) generalized d-gon is an incidence system with the properties:
(a) each pair of elements of P U L is joined by a chain of length at most
d; (b) there is a pair of elements of P U L for which there is no proper
chain of length less than d joining them; (c) there are no closed chains
of length less than 2d.
Denote by Gd{s, t) a generalized d-gon with s points on each line and
t lines through each point. Given a Gd(k, fc), the graph whose vertex-set
is P U L and whose edge-set consists of incident pairs is a (fc, 2d)-cage
with no(fc, 2d) vertices. The converse is also true. Thus our 'generalized
polygon graphs' are just the incidence graphs of generalized d-gons with
s = t.
188 Symmetry and regularity

It is easy to construct a G2{k,k) for all k > 2; the corresponding

graph is the complete bipartite graph Kk^. A G$(k,k) is simply a
projective plane with k points on each line. So the existence problem
for generalized polygon graphs of girth 6 is covered by the known results
on projective planes, a fact noted by Singleton (1966). There is at least
one such plane whenever k — 1 is a prime power, and none are known for
which k — 1 is not a prime power. Generalized quadrangles G±{k, k) are
also known to exist for all prime power values of k — 1, and generalized
hexagons G§{k, k) exist whenever k - 1 is an odd power of 3. Benson
(1966) was the first to construct the graphs corresponding to the the
last two cases.
In the case of odd girth g > 3, the only Moore graphs allowed by
Theorem 23.6 are those with g = 5 and k € {3,7,57}. The graph with
k = 3 is Petersen's graph. The graph with k = 7 was constructed and
proved unique by Hoffman and Singleton (1960); a construction is given
in 23d. The existence of a graph with k = 57 remains an enigma: the
results of Aschbacher (1971) show that such a graph cannot be distance-
transitive, and so the construction, if there is one, is certain to be very
complicated.

Additional Results

23a Moore graphs and generalized polygon graphs with degree 3 In the
case k = 3, the Moore graphs of girth 3 and girth 5 (K4 and O3) exist and
are unique. There are no other Moore graphs of degree 3, by Theorem
23.6. The generalized polygon graphs of girth 4, 6, 8 and 12 exist and
are unique. They are ^3,3, Heawood's graph 5(7), Tutte's graph Cl, and
the incidence graph of the unique generalized hexagon with 63 points
and 63 lines (see 23b).

23b The (3,12)-cage A direct construction of the generalized hexagon

graph of degree 3 is as follows. Given a unitary polarity of the projective
plane PG(2,32), there are 63 points of the plane which do not lie on
their polar lines, and they form 63 self-polar triangles (Edge 1963). The
(3,12)-cage is the graph whose 126 vertices are these 63 points and 63
triangles, with adjacent vertices corresponding to an incident (point,
triangle) pair.
This graph is not vertex-transitive, since there is no automorphism
taking a 'point' vertex to a 'triangle' vertex. However, it follows from
Proposition 23.1 that it is distance-regular.
 Minimal regular graphs with given girth 189

23c Cages with degree 3 and g < 12 All cases except g = 7,9,10 and 11
have been covered above. In these cases we know from the general theory
that a (3, <7)-cage must have more than no(3, g) vertices. The (3,7)-cage
is a graph with 24 vertices, and it is unique; details are given by Tutte
(1966). There are numerous (3,9)-cages; they have 58 vertices and the
first one was found by Biggs and Hoare (1980). The fact that no smaller
graph has degree 3 and girth 9 is the result of a computer search by
B. McKay. There are three (3,10)-cages; they have 70 vertices (O'Keefe
and Wong 1980). The size of the (3, ll)-cage is as yet unknown. Since
it is not a Moore graph it must have at least 96 vertices; the smallest
known graph with degree 3 and girth 11 has 112 vertices.

23d The Hoffman-Singleton graph The unique (7,5)-cage may be con-

structed by extending the graph described in 22f as follows. Add 14 new
vertices, called L, N, a, b, c, d, e, f, 1,2,3,4,5,6; join L to a, b, c, d, e, /
and N; join N to 1,2,3,4,5,6 and L. Also, join the vertex denoted by
(l,n) in 22f to I and n. The automorphism group of this graph is the
group of order 252 000 obtained from PSU(3,52) by adjoining the field
automorphism of GF(b2) (Hoffman and Singleton 1960).

23e Cages of girth 5 with 4 < k < 6 In these cases we know that a cage
is not a Moore graph. There is a unique (4,5)-cage with 19 vertices, due
to Robertson (1964). There are several (5,5)-cages having 30 vertices;
see [BCN, p. 210]. There is a unique (6,5)-cage (O'Keefe and Wong
1979); it has 40 vertices and it is the induced subgraph obtained by
deleting the vertices of a Petersen graph from the Hoffman-Singleton
graph.

23f Cages of girth 6 Recall (4d) that the excess of a fc-regular graph
with n vertices and girth g is e = n — no(k,g). Biggs and Ito (1980)
showed that, for small values of e, a fc-regular graph with girth 6 and
excess e = 2(77 — 1) is an 77-fold covering of the incidence graph of a
symmetric (u,fc,77)-design.
When 77 = 1 such a design is a projective plane, and we have the
generalized polygon graph as discussed above. When 77 = 2 such a
design is called a biplane. In this case it can be shown that a necessary
condition for the existence of a graph is that either k or k — 2 must be
a perfect square (see Biggs 1981b). Such graphs with k = 3 and k = A
do exist, but they they are not (k, 6)-cages, because for these values of k
there is a generalized polygon graph. The first significant case is k = 11,
because here it is now known that there is no projective plane, so the
190 Symmetry and regularity

graph (if it exists) would be an (11,6)-cage. There are several biplanes,

but the existence of a 2-fold covering has not been settled.
When 77 = 3, coverings have been constructed for k = 4,7 and 12. The
case k = 7 is particularly important, because there is no projective plane
or biplane in this case, and so the graph is a (7,6)-cage (see O'Keefe and
Wong (1981) and Ito (1981)). This is the last of the known cages.

23g Families of graphs with large girth Graphs with small excess are
very special, and we therefore adopt a wider definition of what is 'inter-
esting' in this context. Let {F r } be a family of fc-regular graphs such
that F r has nr vertices and girth gr. We say that the family has large
girth if nr and gr both tend to infinity as r —• 00, in such a way that

lim —- is a finite constant c.

r-*oo gr
It follows from the explicit form of no(k, g) that c cannot be less than 0.5.
For many years the existence of families with large girth was established
only by non-constructive means; these arguments showed that there are
families with c = 1. Weiss (1984) showed that, in the case k = 3, the
family of sextet graphs S(p) defined in 18b has c = 0.75, and Lubotzky,
Phillips and Sarnak (see 23h) constructed families which attain the same
value for infinitely many values of k. A simple construction for cubic
graphs with large girth (but with c > 1) was given by Biggs (1987).

23h The graphs of Lubotzky, Phillips and Sarnak Let p be a prime

congruent to 1 modulo 4 and let H denote the set of integral quaternions
a = (00,01,02,03). Define A(2) to be the set of it-equivalence classes
of elements a of H with a = 1 mod 2 and ||a|| a power of p, where
aR(3 if ±.pra = psf3. Denote by 5" the set of elements of H satisfying
||a|| = p, a = 1 mod 2, and ao > 0. There are (p + l)/2 conjugate
pairs {a, a} in S, and the Cayley graph of A(2) with respect to S is the
infinite (p + l)-regular tree.
Now let q be another prime congruent to 1 modulo 4, such that q >
y/p and (p I q) = —1. Denote by A(2g) the normal subgroup of A(2)
consisting of those classes represented by a with at, 02,03 divisible by 2q.
The Cayley graph of S/A(2q) with respect to A(2)/A(2o) is a bipartite
(p + l)-regular graph with q{q2 — 1) vertices and girth approximately
41ogpg. For further details see Lubotzky, Phillips and Sarnak (1988),
Biggs and Boshier (1990).
 REFERENCES

Adam, A. (1967). Research problem 2-10, JCT 2, 393. [126]

Adel'son-Velskii, G.M., et al. (1969). Example of a graph without a
transitive automorphism group [in Russian], Soviet Math. Dokl. 10,
440-441. [162]
Alon, N. and V.D. Milman (1985). Ai, isoperimetric inequalities for
graphs and superconcentrators, JCT(B) 38, 73-88. [22,58]
Alspach, B., D. Marusic, and L. Nowitz (1993). Constructing graphs
which are 1/2-transitive, J. Austral. Math. Soc, to appear. [120]
Anthony, M.H.G. (1990). Computing chromatic polynomials, Ars Com-
binatoria 29C, 216-220. [Ill]
Aschbacher, M. (1971). The non-existence of rank three permutation
groups of degree 3250 and subdegree 57, J. Algebra 19, 538-540. [188]
Austin, T.L. (1960). The enumeration of point labelled chromatic graphs
and trees, Canad. J. Math. 12, 535-545. [41]
Babai, L. (1979). Spectra of Cayley graphs, JCT(B) 27, 180-189. [128]
Babai, L. (1981). On the abstract group of automorphisms, Combina-
torics (ed. H.N. Temperley, Cambridge U.P.), 1-40. [117]
Baker, G.A. (1971). Linked-cluster expansion for the graph-vertex col-
oration problem, JCT(B) 10, 217-231. [81,86,93]
Bannai, E. and T. Ito (1973). On Moore graphs, J. Fac. Sci. Univ.
Tokyo Sec. 1A 20, 191-208. [182]
Benson, C.T. (1966). Minimal regular graphs of girth eight and twelve,
Canad. J. Math. 18, 1091-1094. [188]
192 References

Beraha, S. and J. Kahane (1979). Is the four-color conjecture almost

false? JCT(B) 27, 1-12. [71]
Beraha, S., J. Kahane and N.J. Weiss (1980). Limits of chromatic zeros
of some families of maps, JCT(B), 28, 52-65. [71,72]
Biggs, N.L. (1970). Finite Groups of Automorphisms (Cambridge Uni-
versity Press). [171]
Biggs, N.L. (1973a). Expansions of the chromatic polynomial, Discrete
Math. 6, 105-113. [81]
Biggs, N.L. (1973b). Three remarkable graphs, Canadian J. Math. 25,
397-411. [103,111]
Biggs, N.L. (1973c). Perfect codes in graphs, JCT(B) 15, 289-296. [172]
Biggs, N.L. (1976). Automorphic graphs and the Krein condition, Geom.
Dedicata 5, 117-127. [170]
Biggs, N.L. (1977a). Colouring square lattice graphs, Bull. London
Math. Soc. 9, 54-56. [96]
Biggs, N.L. (1977b). Interaction Models (Cambridge U.P.). [ 80,87,110]
Biggs, N.L. (1978). On cluster expansions in graph theory and physics,
Quart. J. Math. (Oxford) 29, 159-173. [81,87,90]
Biggs, N.L. (1979). Some odd graph theory, Ann. New York Acad. Sci.
319, 71-81. [58,137]
Biggs, N.L. (1980). Girth, valency and excess, Lin. Alg. Appl. 31,
55-59. [28]
Biggs, N.L. (1981a). Aspects of symmetry in graphs, Algebraic Methods
in Graph Theory ed. L. Lovasz and V. Sos (North-Holland) 27-35. [137]
Biggs, N.L. (1981b). Covering biplanes, Theory and Applications of
Graphs ed. G. Chartrand (Wiley), 73-82. [189]
Biggs, N.L. (1982a). A new 5-arc-transitive cubic graph, J. Graph The-
ory 6, 447-451. [154]
Biggs, N.L. (1982b). Constructing 5-arc-transitive cubic graphs, J. Lon-
don Math. Soc. (2) 26, 193-200. [153]
Biggs, N.L. (1982c). Distance-regular graphs with diameter three, Ann.
Discrete Math. 15, 68-90. [179]
Biggs, N.L. (1984a). Presentations for cubic graphs, Computational
Group Theory ed. M.D. Atkinson (Academic Press), 57-63. [146]
Biggs, N.L. (1984b). Homological coverings of graphs, J. London Math.
Soc. (2) 30, 1-14. [154]
Biggs, N.L. (1987). Graphs with large girth, Ars Combinatoria 25C,
73-80. [190]
Biggs, N.L. and A.G. Boshier (1990). Note on the girth of Ramanujan
graphs, JCT(B) 49, 190-194. [190]
 References 193

Biggs, N.L., A.G. Boshier and J. Shawe-Taylor (1986). Cubic distance-

regular graphs, J. London Math. Soc. (2) 33, 385-394. [163,171]
Biggs, N.L., G.R. Brightwell and D. Tsoubelis (1992). Theoretical and
practical studies of a competitive learning process, Network 3 285-301.
[58]
Biggs, N.L., R.M. Damerell and D.A. Sands (1972). Recursive families
of graphs, JCT(B) 12, 123-131. [42, 69, 70, 103]
Biggs, N.L. and M.J. Hoare (1980). A trivalent graph with 58 vertices
and girth 9, Discrete Math. 30, 299-301. [189]
Biggs, N.L. and M.J. Hoare (1983). The sextet construction for cubic
graphs, Combinatorica 3, 153-165. [144,152]
Biggs, N.L. and T. Ito (1980). Graphs with even girth and small excess,
Math. Proc. Cambridge Philos. Soc. 88, 1-10. [189]
Biggs, N.L. and G.H.J. Meredith (1976). Approximations for chromatic
polynomials, JCT(B) 20, 5-19. [96]
Biggs, N.L. and D.H. Smith (1971). On trivalent graphs, Bull. London
Math. Soc. 3, 155-158. [171]
Birkhoff, G.D. (1912). A determinant formula for the number of ways
of coloring a map, Ann. Math. 14, 42-46. [76]
Blokhuis, A. and A.E. Brouwer (1993). Locally K3t3 or Petersen graphs,
Discrete Math. 106/7, 53-60. [129]
Bollobas, B. (1985). Random Graphs (Academic Press). [119]
Bondy, J.A. (1991). A graph reconstructor's manual, Surveys in Com-
binatorics ed. A.D. Keedwell (Cambridge U.P.), 221-252. [51]
Bose, R.C. and D.M. Mesner (1959). Linear associative algebras, Ann.
Math. Statist. 30, 21-38. [161]
Bouwer, I.Z. (1970). Vertex- and edge-transitive but not 1-transitive
graphs, Canad. Math. Bull. 13, 231-237. [120]
Bouwer, I.Z. (1972). On edge- but not vertex-transitive regular graphs,
JCT(B) 12, 32-40. [119]
Bouwer, I.Z. (1988). The Foster Census (Charles Babbage Research
Centre, Winnipeg). [147]
Brylawski, T. (1971). A combinatorial model for series-parallel net-
works, Trans. Amer. Math. Soc. 154, 1-22. [109]
Cameron, P.J. (1980). 6-transitive graphs, JCT(B) 28, 168-179. [162]
Cameron, P.J. (1982). There are only finitely many distance-transitive
graphs of given valency greater than two, Combinatorica 2, 9-13. [163]
Cameron, P.J. (1991). Covers of graphs and EGQs, Discrete Math. 97,
83-92. [179]
194 References

Cameron, P.J., J.M. Goethals, J.J. Seidel, E.E. Shult (1976). Line
graphs, root systems, and elliptic geometry, J. Algebra 43, 305-327.
[18,22]
Cayley, A (1878). On theory of groups, Proc. London Math. Soc. 9,
126-133. [122]
Cayley, A. (1889). A theorem on trees, Quart. J. Math. 23, 376-378.
{Collected papers, 13, 2628). [39]
Chang, S. (1959). The uniqueness and nonuniqueness of the triangular
association scheme, Sci. Record 3, 604-613. [21]
Chung, F-K. (1989). Diameters and eigenvalues, J. Amer. Math. Soc.
2, 187-196. [22]
Collatz, L. and U. Sinogowitz (1957) Spektren endlicher Grafen, Abh.
Math. Sem. Univ. Hamburg 2 1, 63-77 [12]
Conder, M. (1987). A new 5-arc-transitive cubic graph, J. Graph Theory
11, 303-307. [148]
Conder, M. and P. Lorimer (1989). Automorphism groups of symmetric
graphs of valency 3, JCT(B) 47, 60-72. [128,146]
Coulson, C.A. and G.S. Rushbrooke (1940). Note on the method of
molecular orbitals, Proc. Camb. Philos. Soc. 36, 193-200. [11]
Coxeter, H.S.M. (1950). Self-dual configurations and regular graphs,
Bull. Amer. Math. Soc. 59, 413-455. [153,154]
Crapo, H.H. (1967). A higher invariant for matroids, JCT 2, 406-417.
[107]
Crapo, H.H. (1969). The Tutte polynomial, Aeq. Math. 3, 211-229.
[102,104]
Cvetkovic, D.M. (1972). Chromatic number and the spectrum of a
graph, Publ. Inst. Math. (Beograd) 14, 25-38. [57]
Cvetkovic, D.M. and P. Rowlinson (1988). Further properties of graph
angles, Scientia A: Math. Sci. 1, 41-51. [51]
Damerell, R.M. (1973). On Moore graphs, Proc. Cambridge Philos.
Soc. 74, 227-236. [160,182]
Delsarte, P. (1973). An algebraic approach to the association schemes
of coding theory, Philips Research Reports Suppl. 10. [172]
Djokovic, D.Z. (1972). On regular graphs II, JCT(B) 12, 252-259. [140]
Djokovic, D.Z. and G.L. Miller (1980). Regular groups of automor-
phisms of cubic graphs, JCT(B) 29, 195-230. [146]
Edge, W.L. (1963). A second note on the simple group of order 6048,
Proc. Cambridge Philos. Soc. 59, 1-9. [188]
Egervary, E. (1931). Matrizok kombinatorius tulajdonsagairol, Mat.
Fiz. Lapok 38, 16-28 [35]
 References 195

Elspas, B. and J.Turner (1970). Graphs with circulant adjacency ma-

trices, JCT 9, 297-307. [126]
Essam, J.W. (1971). Graph theory and statistical physics, Discrete
Math. 1, 83-112. [107]
Feit, W. and G. Higman (1964). The non-existence of certain generalized
polygons, J. Algebra 1, 114-131. [182]
Frucht, R. (1938). Herstellung von Graphen mit vorgegebener abstrakter
Gruppe, Compositio Math. 6, 239-250. [117]
Frucht, R. (1949). Graphs of degree three with a given abstract group,
Canad. J. Math. 1, 365-378. [117]
Frucht, R., J.E. Graver and M.E. Watkins (1971). The groups of the
generalized Petersen graphs, Proc. Cambridge Philos. Soc. 70,211-218.
[119]
Gardiner, A.D. (1973). Arc transitivity in graphs, Quart. J. Math.
Oxford (2) 24, 399-407. [137]
Gardiner, A.D. (1974). Antipodal covering graphs, JCT(B) 16, 255-273.
[179]
Gardiner, A.D. (1976). Homogeneous graphs, JCT(B) 20, 94-102. [120]
Godsil, C.D. (1981). GRR's for solvable groups, Algebraic Methods in
Graph Theory ed. L. Lovasz and V. Sos (North-Holland), 221-239. [128]
Godsil, C.D. and A.D. Hensel (1992). Distance-regular covers of the
complete graph, JCT(B) 56, 205-238. [179]
Godsil, C.D. and B. Mohar (1988). Walk-generating functions and spec-
tral measures of infinite graphs, Lin. Alg. Appl. 107, 191-206. [13]
Goldschmidt, D. (1980). Automorphisms of trivalent graphs, Ann. of
Math. I l l , 377-406. [143]
Grimmett, G.R. (1976). An upper bound for the number of spanning
trees of a graph, Discrete Math. 16, 323-324. [42]
Grimmett, G.R. (1978). The rank polynomials of large random lattices,
J. London Math. Soc. 18, 567-575. [94]
Grone, R. and R. Merris (1988). A bound for the complexity of a simple
graph, Discrete Math. 69, 97-99. [42]
Gross, J.L. (1974). Voltage graphs, Discrete Math. 9, 239-246. [149]
Grotschel, M., L. Lovasz, and A. Schrijver (1988). Geometric Algorithms
and Combinatorial Optimization (Springer, New York). [35]
Haemers, W. (1979). Eigenvalue Techniques in Design and Graph The-
ory (Math. Centrum, Amsterdam). [57]
Haggard, G. (1976). The chromatic polynomial of the dodecahedron,
Technical Report No. 11 (University of Maine). [69]
196 References

Hall, J.I. (1980). Locally Petersen graphs, J. Graph Theory, 4, 173-187.

[129]
Hammersley, J. (1981). The friendship theorem and the love problem,
Surveys in Combinatorics ed. E.K. Lloyd (Cambridge University Press),
31-54. [171]
Harary, F. (1962). The determinant of the adjacency matrix of a graph,
SIAM Review 4, 202-210. [44]
Heller, I. and C.B. Tompkins. An extension of a theorem of Dantzig's,
Linear Inequalities and Related Systems ed. H.W. Kuhn and
A.W. Tucker (Princeton U.P.). [34,35]
Heron, A. (1972). Matroid polynomials, Combinatorics ed.
D.J.A. Welsh and D.R. Woodall (Institute of Mathematics and its Ap-
plications) 164-202. [109]
Hetzel, D. (1976). Uber regulare graphische Darstellung von auflosbaren
gruppen, Diplomarbeit, Technische Universitdt Berlin. [128]
Higman, D.G. (1967). Intersection matrices for finite permutation
groups, J. Algebra 6, 22-42. [161]
Higman, D.G. and C.C. Sims (1968). A simple group of order 44,352,000,
Math. Zeitschr. 105, 110-113. [161]
Hoffman, A.J. (1960). On the exceptional case in the characterization
of the arcs of a complete graph, IBM J. Res. Dev. 4, 487-496. [21]
Hoffman, A.J. (1963). On the polynomial of a graph, Amer. Math.
Monthly 70, 30-36. [15]
Hoffman, A.J. (1969). The eigenvalues of the adjacency matrix of a
graph, Combinatorial Mathematics and its Applications (Univ. of N.
Carolina Press, Chapel Hill), 578-584. [17]
Hoffman, A.J. (1970). On eigenvalues and colorings of graphs, Graph
Theory and its Applications (Academic Press, New York), 79-92. [57]
Hoffman, A.J. and J.B. Kruskal (1956). Integer boundary points of con-
vex polyhedra, Linear Inequalities and Related Systems ed. H.W. Kuhn
and A.W. Tucker (Princeton U.P.), 223-246. [35]
Hoffman, A.J. and R.R. Singleton (1960). On Moore graphs with diam-
eters 2 and 3, IBM J. Res. Dev. 4, 497-504. [182,188,189]
Holt, D. (1981). A graph which is edge-transitive but not arc-transitive,
J. Graph Theory 5, 201-204. [120]
Horton, J.D. and I.Z. Bouwer (1991). Symmetric Y-graphs and H-
graphs, JCT(B) 53, 114-129. [147]
Ito, T. (1981). On a graph of O'Keefe and Wong, J. Graph Theory 5,
87-94. * [190]
Ito, T. (1982). Bipartite distance-regular graphs of valency three, Lin.
Alg. Appl. 46, 195-213. [153]
 References 197

Ivanov, A. A. (1983). Bounding the diameter of a distance-regular graph,

Soviet Math. Dokl. 28, 149-152. [161]
Izbicki, H. (1960). Regulare Graphen beliebigen Grades mit vorgegebe-
nen Eigenschaften, Monatshefte fur Math. 64, 15-21. [121]
Jaeger, F. (1988). On Tutte polynomials and cycles of planar graphs,
JCT(B) 44, 129-146. [104]
Jaeger, F. (1991). Even subgraph expansions for the flow polynomial of
cubic plane maps, JCT(B) 52, 259-273. [110]
Jaeger, F., D.L. Vertigan and D.J.A. Welsh (1990). The computational
complexity of the Jones and Tutte polynomials, Math. Proc. Camb.
Philos. Soc. 108, 35-53. [105,111]
Jordan, C. (1869). Sur les assemblages des lignes, J. f. d. Reine u.
Angew. Math. 70, 185-190. [119]
Kelly, P.J. (1957). A congruence theorem for trees. Pacific J. Math. 7,
961-968. [50]
Kim, D. and I.G. Enting (1979). The limit of chromatic polynomials,
JCT(B) 26, 327-336. [96]
Kirchhoff, G. (1847). Uber die Auflosung der Gleichungen, auf welche
man bei der Untersuchung der linearen Verteilung galvanischer Strome
gefiirht wird, Ann. Phys. Chem. 72, 497-508. [31,36]
Kocay, W.L. (1981). On reconstructing spanning subgraphs, Ars Com-
binatoria 11, 301-313. [50]
Las Vergnas, M. (1978). Eulerian circuits of 4-valent graphs imbedded in
surfaces, Algebraic Methods in Graph Theory ed. L. Lovasz and V. Sos
(North-Holland, Amsterdam), 451-477. [104]
Las Vergnas, M. (1988). On the evaluation at (3,3) of the Tutte poly-
nomial of a graph, JCT(B) 45, 367-372. [104]
Li, N-Z., and E.G. Whitehead (1992). The chromatic uniqueness of Ww,
Discrete Math. 104, 197-199. [69]
Lieb, E.H. (1967). Exact solution of the problem of the entropy of two-
dimensional ice, Phys. Rev. Letters 18, 692-694. [95]
Lorimer, P. (1989). The construction of Tutte's 8-cage and the Conder
graph, J. Graph Theory 13, 553-557. [148]
Lovasz, L. (1979). On the Shannon capacity of a graph, IEEE Trans.
Inform. Theory 25, 1-7. [51]
Lubotzky, A., R. Phillips and P. Sarnak (1988). Ramanujan graphs,
Combinatorica 8, 261-277. [190]
Meredith, G.H.J. (1976). Triple-transitive graphs, J. London Math. Soc.
(2) 13, 249-257. [162]
Mohar, B. (1991). Eigenvalues, diameter, and mean distance in graphs,
Graphs Comb. 7 53-64. [22]
198 References

Moon, J. (1967). Enumerating labelled trees, Graph Theory and Theo-

retical Physics (Academic Press, New York), 261-272. [41]
Moon, J. (1970). Counting labelled trees (Canadian Math. Congress,
Montreal). [38]
Motzkin, T., and E.G. Straus (1965). Maxima for graphs and a new
proof of a theorem of Turan, Canad. J. Math. 17, 533-540. [59]
Mowshowitz, A. (1969). The group of a graph whose adjacency matrix
has all distinct eigenvalues, Proof Techniques in Graph Theory (Aca-
demic Press, New York), 109-110. [117]
Nagle, J.F. (1971). A new subgraph expansion for obtaining coloring
polynomials for graphs, JCT(B) 10, 42-59. [80]
Nerode, A. and H. Shank (1961). An algebraic proof of Kirchhoff's
network theorem, Amer. Math. Monthly 68, 244-247. [36]
O'Keefe, M. and P.K. Wong (1979). A smallest graph of girth 5 and
valency 6, JCT(B) 26, 145-149. [189]
O'Keefe, M. and P.K. Wong (1980). A smallest graph of girth 10 and
valency 3, JCT(B) 29, 91-105. [189]
O'Keefe, M. and P.K. Wong (1981). The smallest graph of girth 6 and
valency 7, J. Graph Theory 5, 79-85. [190]
Parsons, T.D. (1980). Circulant graph imbeddings, JCT(B) 29, 310-
320. [126]
Petersdorf, M. and H. Sachs (1969). Spektrum und Automorphismen-
gruppe eines Graphen, Combinatorial Theory and its Applications, III
(North-Holland, Amsterdam), 891-907. [117,125]
Poincare, H. (1901). Second complement a l'analysis situs, Proc. London
Math. Soc. 32, 277-308. [32]
Ray, N. (1992). Tutte algebras of graphs and formal group theory, Proc.
London Math. Soc. 65, 23-45. [88]
Ray, N. and C. Wright (1992). Umbral interpolation and the addi-
tion/contraction tree for graphs, Discrete Math. 103, 67-74. [72]
Read, R.C. (1968). An introduction to chromatic polynomials, /. Comb.
Theory 4, 52-71. [69,108]
Read, R.C. (1990). Recent advances in chromatic polynomial theory,
Proc. 5th Carib. Conf. on Graph Theory and Computing. [71]
Robertson, N. (1964). The smallest graph of girth 5 and valency 4, Bull.
Amer. Math. Soc. 70, 824-825. [189]
Ronse, C. (1978). On homogeneous graphs, J. London Math. Soc. (2)
17, 375-379. [120]
Rowlinson, P. (1987). A deletion-contraction algorithm for the char-
acteristic polynomial of a multigraph, Proc. Roy. Soc. Edin. 105A,
153-166. [11]
 References 199

Sachs, H. (1962). Uber selbstkomplementare Graphen, Publ. Math.

Debrecen 9, 270-288. [20]
Sachs, H. (1963). Regular graphs with given girth and restricted circuits,
J. London Math. Soc. 38, 423-429. [180]
Sachs, H. (1964). Beziehungen zwischen den in einem Graphen enthal-
tenen Kreisen und seinem charakterischen Polynom, Publ. Math. De-
brecen. 11, 119-134. [49]
Sachs, H. (1967). Uber Teiler, Faktoren und characterische Polynome
von Graphen II, Wiss. Z. Techn. Hosch. Ilmenau 13, 405-412. [19]
Schmitt, W. (1993). Hopf algebra methods in graph theory, Pure Appl.
Algebra, to appear. [88]
Schur, I. (1933). Zur Theorie der einfach transitiven Permutationsgrup-
pen, S.B. Preuss. Akad. Wiss., 598-623. [161]
Schwenk, A.J. (1973). Almost all trees are cospectral, New Directions
in the Theory of Graphs (Academic Press, New York). [49]
Scott, L.L. (1973). A condition on Higman's parameters, Notices Amer.
Math. Soc. 20, A97 (701-20-45). [170]
Sedlacek, J. (1970). On the skeletons of a graph or digraph, Combi-
natorial Structures and Applications (Gordon and Breach, New York),
387-391. [42]
Seidel, J.J. (1968). Strongly regular graphs with ( — 1,1,0) adjacency
matrix having eigenvalue 3, Lin. Alg. Appl. 1, 281-298. [18,21]
Seidel, J.J. (1979). Strongly regular graphs, Surveys in Combinatorics
ed. B. Bollobas (Cambridge U.P.), 157-180. [171]
Seymour, P.D. (1981). Nowhere-zero 6-flows. JCT(B) 30, 130-135. [30]
Sheehan, J. (1974). Smoothly embeddable subgraphs, J. London Math.
Soc. (2) 9, 212-218. [120]
Shrikhande, S.S. (1959). The uniqueness of the L2 association scheme,
Ann. Math. Stat. 30, 781-798. [21]
Sims, C.C.(1967).Graphs and finite permutation groups, Math. Zeitschr.
95, 76-86. [141]
Singleton, R.R. (1966). On minimal graphs of maximum even girth, J.
Comb. Theory 1, 306-332. [188]
Smith, D.H. (1971). Primitive and imprirnitive graphs, Quart. J. Math.
Oxford (2) 22, 551-557. [177,178]
Smith, J.H. (1970). Some properties of the spectrum of a graph, Com-
binatorial Structures and their Applications ed. R. Guy et al. (Gordon
and Breach, New York), 403-406. [43]
Stanley, R.P. (1973). Acyclic orientations of graphs, Discrete Math. 5,
171-178. [70]
200 References

Temperley, H.N.V. (1964). On the mutual cancellation of cluster inte-

grals in Mayer's fugacity series, Proc. Phys. Soc. 83, 3-16. [39]
Thistlethwaite, M.B. (1987). A spanning tree expansion of the Jones
polynomial, Topology 26, 297-309. [105]
Thomassen, C. (1990). Resistances and currents in infinite electrical
networks, JCT(B) 49, 87-102. [36]
Tutte, W.T. (1947a). A family of cubical graphs, Proc. Cambridge
Philos. Soc. 45, 459-474. [130,141,143]
Tutte, W.T. (1947b). A ring in graph theory, Proc. Cambridge Philos.
Soc. 43, 26-40. [79]
Tutte, W.T. (1954). A contribution to the theory of chromatic polyno-
mials, Canad. J. Math. 6, 80-91. [30,102,110]
Tutte, W.T. (1956). A class of Abelian groups, Canad. J. Math. 8,
13-28. [30]
Tutte, W.T. (1966). Connectivity in Graphs (Toronto University Press).
[120,132, 189]
Tutte,W.T.(1967). On dichromatic polynomials, JCT2, 301-320.
[81,86,90,96]
Tutte, W.T. (1970). On chromatic polynomials and the golden ratio,
JCT, 9 289-296. [71]
Tutte, W.T. (1974). Codichromatic graphs, JCT(B) 16, 168-174. [79]
Tutte, W.T. (1979). All the king's horses - a guide to reconstruction,
Graph Theory and Related Topics ed. J.A. Bondy and U.S.R. Murty
(Academic Press), 15-33. [57]
Tutte, W.T. (1984). Graph Theory (Addison-Wesley). [72]
Vijayan, K.S. (1972). Association schemes and Moore graphs, Notices
Amer. Math. Soc. 19, A-685. [182]
Weiss, A. (1984). Girths of bipartite sextet graphs, Combinatorica 4,
241-245. [190]
Weiss, R. (1974). Uber s-regulare Graphen, JCT(B) 16, 229-233. [141]
Weiss, R. (1983). The non-existence of 8-transitive graphs, Combina-
torica 1, 309-311. [137]
Weiss, R. (1985) On distance-transitive graphs, Bull. London Math.
Soc. 17, 253-256. [163]
Whitney, H. (1932a). A logical expansion in mathematics, Bull. Amer.
Math. Soc. 38, 572-579. [76,77]
Whitney, H. (1932b). The coloring of graphs, Ann. Math. 33, 688-
718. [69,81,86]
Whitney, H. (1932c). Congruent graphs and the connectivity of graphs,
Amer. J. Math. 54, 150-168. [120]
 References 201

Wielandt, H. (1964). Finite Permutation Groups (Academic Press).

[161,172]
Wilf, H.S. (1967). The eigenvalues of a graph and its chromatic number,
J. London Math. Soc. 42, 330-332. [55]
Wilf, H.S. (1986). Spectral bounds for the clique and independence
numbers of graphs, JCT(B) 40, 113-117. [59]
Woodall, D.R. (1992). A zero-free interval for chromatic polynomials,
Discrete Math. 101, 333-341. [72]
Yetter, D.N. (1990). On graph invariants given by linear recurrence
relations, JCT(B) 48, 6-18. [72]
Yuan, Hong (1988). A bound on the spectral radius of graphs, Lin. Alg.
Appl. 108, 135-139. [13]
 Index

acyclic orientation 70 complete graph 8

adjacent 7 complete matching 50
adjacency algebra 9 complete multipartite graph 41
adjacency matrix 7 conductance 34
almost-complete 43 cone 66
alternating knot 105 confluence 72
angles 51 conforms 30
antipodal 177 conjugate Bell polynomials 72
antipodal r-fold covering 178 connected 10
augmentation 29 contracting 64
automorphic 178 Conway's presentations 145
automorphism 115 co-rank 25, 97
automorphism group 115 coset graph 128
cospectral graphs 12, 49
bicentroid 119 cover 50
bigrading 97 covering graph 149
bipartite 11
cube 43, 69, 140, 157, 161, 169
biplane 189
cubic graph 138
block 81
current 34
block system 173
cut 26
broken cycle 77
Brooks's theorem 55 cut-orientation 26
buckminsterfullerene 127 cut-subspace 26
cut-vertex 67
cage 181, 188, 189 cycle 25
Cayley graph 123 cycle graph 17, 65
centroid 119 cycle-orientation 25
characteristic polynomial 8 cycle-subspace 26
chromatically unique 69
chromatic invariant 107 degree 4
chromatic number 52 deletion-contraction 65, 72
chromatic polynomial 63 density 94
chromatic root 71 derived graph 178
circulant graph 16, 126 Desargues graph 148, 153
circulant matrix 16 diameter 10
closed walk 12 dihedral group 126
coboundary mapping 28 distance 10
cocktail-party graph 17, 68 distance matrices 13, 159
colour-class 52 distance-regular 13, 159
colour-partition 52 distance-transitive 118, 155
compatible 150 dodecahedron 69, 178
complete bipartite graph 21 double pyramid 68
 Index 203

double-transitivity 118 incidence mapping 24, 29

dual 29,43 incidence matrix 24
independent 98
edge 3 indicator function 74
edge space 23 induced subgraph 4
edge-transitive 115, 118, 120 interaction model 80
effective resistance 36 internal activity 99
eigenvalue 8 internally active 99
electrical network 34 intersection array 157, 159
elementary 44 intersection matrix 165
ends 4 intersection numbers 156
equipartition 58 Ising model 80
even subgraph 110 isoperimetric number 28, 58
excess 28, 189 isthmus 30
expansion 147
external activity 99 join 66
externally active 99 Jones polynomial 105

feasible array 168 K-chain 149

flow 29 Kelly's lemma 50
flow polynomial 110 Kirchhoff's laws 34
forest 47 Kocay's lemma 50
Foster's census 147 Krein parameters 170
friendship theorem 171
labelled tree 104
generalized d-gon 187 ladder 69, 126
generalized line graph 21 Laplacian matrix 27
generalized polygon graph 181 Laplacian spectrum 29, 40
general graph 3 line graph 17, 120
girth 28, 76, 131, 180 logarithmic transform 82
graph 4 loop 3
graphical regular representation medial graph 104
124, 128 minimal support 29
graph types 87 Mobius ladder 20, 42, 69, 110
modified rank polynomial 101
Hamiltonian cycle 50
modular flow 30
Hamming graph 169
Moore graph 181
Heawood graph 148, 154, 163
Motzkin-Straus formula 59
Hoffman-Singleton graph 189
homeomorphic 79, 108 negative end 24
homogeneous 120 nowhere-zero 30
homological covering 154
Hopf algebra 88 octahedron 43
hyperoctahedral graph 17 odd graphs 20, 58, 137, 161, 170
orbit 115
icosahedral group 127 orientation 24
icosahedron 69, 178
imprimitive 177 Paley graph 129
204 Index

Pappus graph 148, 154 spectrum 8

partial geometry 162 sporadic groups 172
partition function 80 square lattice 96
path graph 11 stabilizer 122
perfect code 22, 171 stabilizer sequence 133, 137, 147
permutation character 172 standard bases 24
permutation matrix 116 star graph 49
Petersen graph 20, 95, 103, 133 star types 87
planar 29 strict graph 4
positive end 24 strongly regular graph 16, 20, 159,
potential 36 171
Potts model 80 subdividing 79
power 36 subgraph 4
primitive 30, 173 successor 132
principal minors 8 support 29
projective plane 163 suspension 66
proper 90 symmetric 118, 126
pyramid 68 symmetric cycle 137
symmetric design 163
quasi-separable 67 symmetric group 118, 148
quasi-separation 67
t-arc 130
rank 25 tetrahedral group 127
rank matrix 73 thermodynamic limit 94
rank polynomial 73 theta graph 86
Rayleigh quotient 54 Thomson's principle 36
Rayleigh's monotonicity law 37 topological invariant 79
reconstructible 50, 91 totally unimodular 34
reconstruction conjecture 50 tree 47, 49, 65, 119
recursive family 70, 103 tree-number 38
regular graph 14 triangle graph 19, 169
regular action 122 tridiagonal 165
resonant model 80 t-transitive 131
rewriting rules 72 Turan's Theorem 59
root systems 22 Tutte polynomial 97, 100
r-ply transitive 162
umbral chromatic polynomial 72
semi-direct product 150 unimodal conjecture 108
separable 67
separation 67 vertex 3
series-parallel 109 vertex-colouring 52
sextet graph 145 vertex space 23
Shannon capacity 51 vertex-stabilizer 122, 127
sides 149 vertex-transitive 115, 120, 125
simple eigenvalues 116, 125 V-function 79
spanning elementary subgraph 44 voltage 34
spanning tree 31
spectral decomposition 13 walk 9
 Index 205

walk-generating function 13 weakly homogeneous 120

walk-generating matrix 12 wheel 68