Lecture Notes in Mathematics 2070

Editors:
J.-M. Morel, Cachan
B. Teissier, Paris

For further volumes:

http://www.springer.com/series/304
•
Alla Detinko  Dane Flannery
Eamonn O’Brien
Editors

Probabilistic Group Theory,

Combinatorics,
and Computing
Lectures from the Fifth de Brún Workshop

123
Editors
Alla Detinko Dane Flannery
Department of Mathematics Department of Mathematics
National University of Ireland, Galway National University of Ireland, Galway
Ireland Ireland

Eamonn O’Brien
Department of Mathematics
University of Auckland
Auckland, New Zealand

ISBN 978-1-4471-4813-5 ISBN 978-1-4471-4814-2 (eBook)

DOI 10.1007/978-1-4471-4814-2
Springer London Heidelberg New York Dordrecht

Lecture Notes in Mathematics ISSN print edition: 0075-8434

ISSN electronic edition: 1617-9692

Library of Congress Control Number: 2012956221

Mathematics Subject Classification (2010): 05-04; 05B05; 20B40; 20D06; 20P05

c Springer-Verlag London 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

This book is inspired by the workshop Groups, Combinatorics, Computing, held

at National University of Ireland, Galway from April 11 to 16, 2011—the Fifth
“de Brún Workshop” run under the auspices of Science Foundation Ireland’s Math-
ematics Initiative Programme. A principal theme of the workshop was interactions
between group theory and combinatorics with algorithmic or computational aspects.
Areas encompassed by this theme are currently the focus of intense research activity.
The core part of the workshop was formed by three lecture courses. These
contained a wide and unique selection of material, for the first time providing an
accessible introduction to frontier research in thematic areas. It became clear that
the courses should be made available to a larger audience.
The book has three chapters, one per lecture course. Each chapter is self-
contained; beginning with background material including historical roots, the reader
is led to the latest results and open problems. Illustrative examples, some proofs and
algorithms, and extensive bibliographies are given.
The first chapter, by Martin Liebeck, is an exposition of recent developments
in probabilistic and asymptotic theory of finite groups, particularly finite simple
groups. The first two sections are on random generation of finite groups and
maximal subgroups. The next topic is representation varieties and character-
theoretic methods. Finally, diameter and growth of Cayley graphs of simple groups
are considered. The chapter traces progress on fundamental conjectures which have
driven the development of this subject.
The second chapter is by Alice Niemeyer, Cheryl Praeger, and Ákos Seress. This
chapter again has a strong probabilistic flavour. It discusses the role of estimation in
the design and analysis of randomised algorithms for computing with finite groups,
and approaches to estimating proportions of important element classes. Among the
latter are geometric methods, the use of generating functions, and theory of Lie
type groups. The chapter also surveys numerous results concerning estimation in
permutation groups and finite classical groups. An application to the construction
of involution centralisers, a key part of the constructive recognition of finite simple
groups, is given. Connections with theoretical computer science are made.

v
vi Preface

In the final chapter, Leonard Soicher presents results from a different area
at the interface of group theory and combinatorics. This chapter emphasises
practical computation. Specifically, it considers how group theory may be used in
the construction, classification, and analysis of combinatorial designs. Statistical
optimality results for semi-Latin squares are reviewed. An account of the new theory
of “uniform” semi-Latin squares and a construction which determines a semi-Latin
square from a transitive permutation group are then given. The chapter describes
use of the GAP package DESIGN. Along with an introduction to the package and
samples of its operation, it is shown how package functions can be used to classify
block designs and semi-Latin squares. In an extended example, new statistically
efficient semi-Latin squares are determined.
We envisage that this book will be a resource for lecture or reading courses or for
self-instruction. Indeed, each chapter is a ready-made graduate lecture course. All
three chapters could serve as the foundation of an advanced graduate programme in
algebra and computing.
The Fifth de Brún Workshop also featured research talks and short presentations.
More details and pdf files of selected talks are available at
http://www.maths.nuigalway.ie/detinko/DeBrun5/
We take this opportunity to record our gratitude to Charles Leedham-Green, who
acted as scientific chair of the workshop. He gave the opening address, and his many
other contributions helped to ensure the event’s success.
In conclusion, we hope that the reader will find these lectures as interesting and
valuable as workshop participants did.

Galway, Ireland Alla Detinko and Dane Flannery

Auckland, New Zealand Eamonn O’Brien
January 2012
Acknowledgements

We thank the authors for agreeing to publication of their courses, and for their efforts
in preparing the courses for this book. The useful and prompt feedback from our
referees is also very much appreciated.
The workshop received funding from Science Foundation Ireland grant 07/MI/
007. NUI Galway provided further support. We are especially indebted to Mary
Kelly and Padraig Ó Catháin, for assistance with many tasks involved in running
the workshop.

vii
•
Contents

1 Probabilistic and Asymptotic Aspects of Finite Simple Groups . . . . . . . . 1

Martin W. Liebeck
1.1 Random Generation of Simple Groups and Maximal Subgroups . . . . 1
1.1.1 Alternating Groups . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1.2 Groups of Lie Type . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
1.1.3 Other Results on Random Generation . . . . .. . . . . . . . . . . . . . . . . . . . 8
1.1.4 Generation of Maximal Subgroups . . . . . . . .. . . . . . . . . . . . . . . . . . . . 10
1.2 Random Generation of Arbitrary Finite Groups .. .. . . . . . . . . . . . . . . . . . . . 11
1.3 Representation Varieties and Character-Theoretic Methods . . . . . . . . . . 15
1.3.1 Fuchsian Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 16
1.3.2 Character Theory . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 17
1.3.3 Symmetric and Alternating Groups .. . . . . . .. . . . . . . . . . . . . . . . . . . . 19
1.3.4 Groups of Lie Type . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 20
1.3.5 Representation Varieties . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
1.3.6 Triangle Groups .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
1.4 Cayley Graphs of Simple Groups: Diameter and Growth .. . . . . . . . . . . . 26
1.4.1 Conjugacy Classes . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 28
1.4.2 Babai’s Conjecture .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 29
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 31
2 Estimation Problems and Randomised Group Algorithms . . . . . . . . . . . . . 35
Alice C. Niemeyer, Cheryl E. Praeger, and Ákos Seress
2.1 Estimation and Randomization .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
2.1.1 Computation with Permutation Groups .. . .. . . . . . . . . . . . . . . . . . . . 35
2.1.2 Recognising the Permutation Group Giants . . . . . . . . . . . . . . . . . . . 36
2.1.3 Monte Carlo Algorithms . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 37
2.1.4 What Kinds of Estimates and in What Groups? .. . . . . . . . . . . . . . 39
2.1.5 What Group is That: Recognising Classical
Groups as Matrix Groups.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
2.1.6 What Group is That: Recognising Lie Type
Groups in Arbitrary Representations . . . . . .. . . . . . . . . . . . . . . . . . . . 42

ix
x Contents

2.2 Proportions of Elements in Symmetric Groups .. . .. . . . . . . . . . . . . . . . . . . . 42

2.2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 42
2.2.2 Historical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 42
2.2.3 Orders of Permutations . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
2.2.4 Number of Cycles. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 44
2.2.5 Generating Functions .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 44
2.2.6 Solutions to x m D 1 in Symmetric and Alternating
Groups .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 48
2.2.7 The Münchausen Method (Bootstrapping) . . . . . . . . . . . . . . . . . . . . 51
2.2.8 Algorithmic Applications of Proportions
in Symmetric Groups .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 54
2.2.9 Restrictions on Cycle Lengths . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 56
2.3 Estimation Techniques in Lie Type Groups.. . . . . . .. . . . . . . . . . . . . . . . . . . . 57
2.3.1 p-Singular Elements in Permutation Groups .. . . . . . . . . . . . . . . . . 57
2.3.2 Quokka Subsets of Finite Groups .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 58
2.3.3 Estimation Theory for Quokka Sets . . . . . . .. . . . . . . . . . . . . . . . . . . . 59
2.3.4 Strong Involutions in Classical Groups .. . .. . . . . . . . . . . . . . . . . . . . 61
2.3.5 More Comments on Strong Involutions . . .. . . . . . . . . . . . . . . . . . . . 63
2.3.6 Regular Semisimple Elements and Generating Functions . . . . 65
2.4 Computing Centralisers of Involutions.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 68
2.4.1 Applications of Centralisers of Involutions Computations . . . 69
2.4.2 Constructive Membership in Lie Type Groups .. . . . . . . . . . . . . . . 70
2.4.3 Constructive Recognition of Lie Type Groups . . . . . . . . . . . . . . . . 72
2.4.4 Computation of an Element Centralising an Involution.. . . . . . 74
2.4.5 Computation of the Full Centraliser . . . . . . .. . . . . . . . . . . . . . . . . . . . 75
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 78
3 Designs, Groups and Computing . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 83
Leonard H. Soicher
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 83
3.2 Background Material.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84
3.2.1 Block Designs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84
3.2.2 Efficiency Measures of 1-Designs . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 85
3.2.3 Permutation Groups . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 86
3.2.4 Latin Squares.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
3.2.5 Semi-Latin Squares .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 88
3.3 Optimality Results for Semi-Latin Squares .. . . . . . .. . . . . . . . . . . . . . . . . . . . 89
3.4 Uniform Semi-Latin Squares . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 90
3.5 Semi-Latin Squares from Transitive Permutation Groups . . . . . . . . . . . . 91
3.5.1 The Canonical Efficiency Factors of SLS.G/ . . . . . . . . . . . . . . . . . 92
3.6 The DESIGN Package .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 94
3.6.1 The BlockDesigns Function . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
3.6.2 The BlockDesignEfficiency Function . . . . . . . . . . . . . . . . 97
Contents xi

3.7 Classifying Semi-Latin Squares .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98

3.8 Efficient Semi-Latin Squares as Subsquares of Uniform
Semi-Latin Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 102
3.9 Some Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 105
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 106
•
Contributors

Martin W. Liebeck Department of Mathematics, Imperial College, London, UK

Alice C. Niemeyer Centre for the Mathematics of Symmetry and Computation,
School of Mathematics and Statistics, The University of Western Australia, Crawley,
WA, Australia
Cheryl E. Praeger Centre for the Mathematics of Symmetry and Computation,
School of Mathematics and Statistics, The University of Western Australia, Crawley,
WA, Australia
King Abdulaziz University, Jeddah, Saudi Arabia
Ákos Seress Centre for the Mathematics of Symmetry and Computation, School
of Mathematics and Statistics, The University of Western Australia, Crawley, WA,
Australia
The Ohio State University, Columbus, OH, USA
Leonard H. Soicher School of Mathematical Sciences, Queen Mary University
of London, London, UK

xiii
•
Chapter 1
Probabilistic and Asymptotic Aspects of Finite
Simple Groups

Martin W. Liebeck

This is a survey of recent developments in the probabilistic and asymptotic theory

of finite groups, with an emphasis on the finite simple groups. The first two sections
are concerned with random generation, while the third section focusses on some
applications of probabilistic methods in representation theory. The final section
deals with asymptotic aspects of the diameter and growth of Cayley graphs.

1.1 Random Generation of Simple Groups and Maximal

Subgroups

1.1.1 Alternating Groups

It is an elementary and well known fact that every alternating group An can be
generated by two elements—for example, by .1 2 3/ and .1 2    n/ if n is odd, and
by .1 2 3/ and .2    n/ if n is even. As long ago as 1892, Netto conjectured that
almost all pairs of elements of An will generate the whole group (see [78, p. 90]).
That is, if for a finite group G we define P .G/ to be the probability that hx; yi D G
for x; y 2 G chosen uniformly at random—so that

jf.x; y/ 2 G  G W hx; yi D Ggj

P .G/ D ;
jGj2

then Netto’s conjecture was that P .An / ! 1 as n ! 1.

It was not until 1969 that Netto’s conjecture was proved by Dixon [16]:

M.W. Liebeck ()

Department of Mathematics, Imperial College, London SW7 2BZ, UK
e-mail: [email protected]

A. Detinko et al. (eds.), Probabilistic Group Theory, Combinatorics, and Computing, 1

Lecture Notes in Mathematics 2070, DOI 10.1007/978-1-4471-4814-2 1,
© Springer-Verlag London 2013
2 M.W. Liebeck

Theorem 1.1. Netto’s conjecture holds—that is, P .An / ! 1 as n ! 1.

8
In fact Dixon proved more than this, showing that P .An / > 1  .log log n/2
for
sufficiently large n.
We shall give a sketch of Dixon’s proof. It relies on two classical results on
permutation groups. The first goes back to Jordan (1873); see [18, p. 84] for a proof.
Lemma 1.2. Suppose that X is a subgroup of Sn which acts primitively on the set
f1; : : : ; ng and contains a p-cycle for some prime p  n  3. Then X D An or Sn .
The second result was deduced by Dixon (see Lemma 3 of [16]) from the paper
[20] of Erdös and Turán—the second of their series of seven pioneering papers on
the statistical theory of the symmetric group.
Lemma 1.3. Let pn be the probability that a permutation in Sn , chosen uniformly
at random, has one of its powers equal to a p-cycle for some prime p  n  3. Then
pn ! 1 as n ! 1.
Sketch proof of Dixon’s Theorem 1.1 Let G D An . Observe that if x; y 2 G do
not generate G, then they both lie in a maximal subgroup M of G. Given M , the
j2
probability that this happens (for random x; y) is jM
jGj2
D jG W M j2 . Hence
X
1  P .G/ D Prob.hx; yi ¤ G for random x; y/  jG W M j2 (1.1)
M max G

where the notation M max G means M is a maximal subgroup of G. Let M be

a set of representatives of the conjugacy classes of maximal subgroups of G. For a
maximal subgroup M , the number of conjugates of M is jG W NG .M /j D jG W M j,
and hence
X
1  P .G/  jG W M j1 : (1.2)
M 2M

The maximal subgroups of G D An fall into three categories, according to their

actions on the set f1; : : : ; ng:
(a) Intransitive subgroups M D .Sk  Snk / \ G for 1  k  Πn2 .
n
(b) Imprimitive subgroups M D .Sr wr Sn=r / \ G, preserving a partition into r
r-subsets, for divisors r of n with 1 < r < n.
(c) Primitive subgroups M .
Denote the contributions to the sum in (1.2) from categories (a), (b), (c) by ˙a , ˙b ,
˙c , respectively, so  1  P .G/  ˙a C ˙b C ˙c . The index in G of a maximal
 that
subgroup in (a) is kn , and hence
!1
X n
Œn=2
˙a D :
k
kD1
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 3

Similarly
X  1
nŠ
˙b D .rŠ/n=r . nr Š/ :
rjn; 1<r<n

Elementary arguments (see Lemmas 1 and 2 in [16]) yield

1 1
˙a D C O. 2 /; ˙b < n  2n=4 :
n n
So clearly
˙a C ˙b ! 0 as n ! 1: (1.3)
Estimating ˙c is less straightforward, however. In fact, instead of estimating this
we deal with the probability Pc that a random pair x; y 2 G generates a proper
primitive subgroup of G. Here Lemmas 1.2 and 1.3 show that Pc ! 0 as n ! 1.
Since 1  P .G/  ˙a C ˙b C Pc , it follows from (1.3) that 1  P .G/ ! 0 as
n ! 1, proving Dixon’s theorem. t
u
The right hand sides of the inequalities (1.1), (1.2) suggest that we define, for any
finite group G, the maximal subgroup zeta function
X X
G .s/ D jG W M js D mn .G/ns
M max G n1

for a real variable s, where mn .G/ denotes the number of maximal subgroups of
index n in G. By the argument for (1.1), we have

P .G/  1  G .2/: (1.4)

For G D An we expressed G .2/ D ˙a C ˙b C ˙c in the above proof. Using

the classification of finite simple groups (CFSG), Babai showed in [4] that the
4
number of conjugacy classes of maximal primitive subgroups of An is at most c log n
for some absolute constant c. Each such subgroup has index at least 12 ΠnC1
2 Š by a
classical result of Bochert (see [18, Theorem 3.3B]), and hence

4 n C 1 1 1
˙c  2c log n .Œ Š/ D O. 2 /:
2 n
This proves [4, 1.2]:
1
Theorem 1.4. For G D An we have G .2/ D n
C O. n12 / and P .G/ D 1  1
n
C
O. n12 /.
A detailed asymptotic expansion of P .An / can be found in [17], and precise
bounds are given in [73].
4 M.W. Liebeck

1.1.2 Groups of Lie Type

At this point we move on to discuss the other non-abelian finite simple groups. By
the classification (CFSG), these are the finite groups of Lie type, together with the
26 sporadic groups. Steinberg proved in [84] that every simple group of Lie type is
2-generated (i.e. can be generated by two elements), and this has also been verified
for the sporadic groups (see [2]). Hence P .G/ > 0 for all finite simple groups
G. In the same paper [16] in which he proved Theorem 1.1, Dixon also made the
following conjecture:
Dixon’s Conjecture. For finite simple groups G we have P .G/ ! 1 as jGj ! 1.
For alternating groups this is of course the content of Theorem 1.1. The
conjecture was proved for classical groups by Kantor and Lubotzky in [37], and
for exceptional groups of Lie type by Liebeck and Shalev in [50]. These proofs
were rather lengthy, but more recent work has led to a much shorter proof, which
we shall now sketch. It is based on the following result, taken from [58].
Theorem 1.5. Fix s > 1. For finite simple groups G, we have G .s/ ! 0 as
jGj ! 1.
Note that the condition s > 1 in the theorem is necessary, since as in (1.2) above,
X
G .s/ D jG W M j1s (1.5)
M 2M

where M is a set of representatives of the conjugacy classes of maximal subgroups

of G. Before discussing the proof of the theorem, here are a couple of immediate
consequences. Firstly, by (1.4) we have
Corollary 1.6. Dixon’s conjecture holds.
P
Next, recalling that G .s/ D n1 mn .G/ns where mn .G/ denotes the number
of maximal subgroups of index n in G, we deduce
Corollary 1.7. Given  > 0, there exists N such that for any n > N and any finite
simple group G, we have mn .G/ < n1C .
Sketch proof of Theorem 1.5
For G D An , the argument given for G .2/ in the proof of Theorem 1.4 works
equally well replacing 2 by any s > 1, to give G .s/ D O.n1s /. So the main task
is to prove Theorem 1.5 for groups of Lie type.

Classical Groups

We begin with an elementary example.

Example. Let G D PSL2 .q/ with q odd. It is well known (see for example [33,
p. 191]) that the maximal subgroups of G are among the following: P (a parabolic
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 5

subgroup of index q C 1), Dq˙1 (dihedral), PSL2 .q0 / or P GL2 .q0 / (where Fq0
is a subfield of Fq ), A4 , S4 or A5 . There are at most 2 conjugacy classes for each
subgroup listed, and the number of subfields of Fq is at most log2 log2 q. It follows
3
that for s > 1 we have G .s/ D .q C 1/1s C O.q 2 .1s/ log log q/ D O.q 1s /. In
particular, G .s/ ! 0 as q ! 1.
Now we sketch the general argument of the proof of Theorem 1.5 for classical
groups, which is essentially that given in [37]. Let G D C ln .q/ denote a simple
classical group over Fq with natural module V of dimension n (so G D PSLn .q/,
PS Un .q/, PSpn .q/ or P ˝n .q/). According to a well known theorem of Asch-
bacher [1], the maximal subgroups of G fall into the following nine families:
C1 : Stabilizers of totally singular or nonsingular subspaces of V (any subspaces
if G D PSLn .q/).
C2 : Stabilizers of direct sum decompositions of V .
C3 : Stabilizers of extension fields of Fq of prime degree.
C4 : Stabilizers of tensor product decompositions V D V1 ˝ V2 .
C5 : Stabilizers of subfields of Fq of prime index.
C6 : Normalizers of r-groups of symplectic type in absolutely irreducible repre-
sentations (r a prime not dividing q).
C7 : Stabilizers of tensor product decompositions V D V1 ˝    ˝ Vm with all Vi
isometric.
C8 : Classical subgroups (of type PSpn .q/, PSOn .q/, PS Un .q 1=2 / in G D
PSLn .q/, or On .q/ in Spn .q/ with q even).
S : Almost simple subgroups with socle acting absolutely irreducibly on V and
defined over no proper subfield of Fq (of Fq 2 if G is unitary).
S
Define C D 8iD1 Ci , and denote by NC (respectively, NS ) the total number of
G-conjugacy classes of maximal subgroups in C (respectively, in S ).
Lemma 1.8. Let G D C ln .q/ as above. There are positive absolute constants c1 ,
c2 , c3 , c4 such that the following hold:
(i) NC  c1 n2 C n log log q;
(ii) NS  f .n/, where f .n/ is a function depending only on n;
2
(iii) also NS  c2 n2 q 6n log q; and jG W M j > q c3 n for all M 2 S ;
n=2
(iv) jG W M j > c4 q for all maximal subgroups M of G.
Precise descriptions of the families Ci can be found in [40, Chap. 4], and the
number of conjugacy classes of subgroups in each family is also given there. Adding
these numbersS up, we easily see that the number of G-conjugacy classes of maximal
subgroups in i ¤5 Ci is less than c1 n2 for some absolute constant c1 , while the
number in C5 is less than n log log q. Part (i) of the lemma follows.
Short arguments for parts (ii) and (iii) can be found in [58, p. 552] and [51, p. 89].
Finally, (iv) follows from [40, 5.2.2] (which gives the subgroups of minimal index
in all simple classical groups).
6 M.W. Liebeck

Corollary 1.9. The total number of conjugacy classes of maximal subgroups of

C ln .q/ is at most f .n/ C c1 n2 C n log log q.
Remark. It is of interest to find good bounds for the function f .n/ in part (ii)
of the lemma. This involves estimating, for each finite quasisimple group S ,
prime p and natural number n, the number of absolutely irreducible representations
of S of dimension n over a field of characteristic p. This is a tough problem,
especially when S is an alternating group or a group of Lie type in characteristic p.
A recent paper [25] of Guralnick, Larsen and Tiep bounds the number of absolutely
irreducible representations of any S of dimension n by the function n3:8 , and uses
this to show that f .n/ < an6 for some absolute constant a. This has been sharpened
by Häsä [29], who has shown that the total number m.G/ of conjugacy classes
of maximal subgroups of any almost simple group G with socle a classical group
C ln .q/ satisfies
m.G/ < 2n5:2 C n log2 log2 q:

Using Lemma 1.8 we can complete the proof of Theorem 1.5 for classical groups
G D C ln .q/. Let s > 1. The argument is divided into the cases where n is bounded
and where n is unbounded. For the case where n is bounded, parts (i), (ii) and (iv)
of the lemma give
X
G .s/ D jG W M j1s < .f .n/ C c1 n2 C n log log q/.c4 q n=2 /1s
M 2M

which tends to 0 as q ! 1. And for the case where n is unbounded, parts (i), (iii)
and (iv) of the lemma give
P P
G .s/  reps: M 2C jG W M j1s C reps: M 2S jG W M j1s
2
 .c1 n2 C n log log q/ .c4 q n=2 /1s C .c2 n2 q 6n log q/ .q c3 n /1s

which tends to 0 as n ! 1.

Exceptional Groups of Lie Type

Now we discuss the proof of Theorem 1.5 when G D G.q/ is a simple exceptional
group of Lie type over Fq —that is, a group in one of the families E8 .q/, E7 .q/,
E6 .q/, 2E6 .q/, F4 .q/, 2F4 .q/, G2 .q/, 3D4 .q/, 2G2 .q/, 2B2 .q/.
Let GN be the simple algebraic group over K D FN q , the algebraic closure of Fq ,
corresponding to G.q/; so if G D E8 .q/ then GN D E8 .K/, if G D 2E6 .q/ then
GN D E6 .K/, and so on. There is a Frobenius endomorphism  of GN such that
G D GN 0 , where GN  denotes the fixed point group fg 2 GN W g  D gg (see [85]).
When G is not a twisted group,  is just a field morphism which acts on root groups
U˛ D fU˛ .t/ W t 2 Kg as U˛ .t/ ! U˛ .t q /; when G is twisted,  also involves a
graph morphism of G. N
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 7

Beginning with the work of Dynkin [19], a great deal of effort has gone into
determining the maximal closed subgroups of positive dimension in the algebraic
N culminating in [49], where this task was completed. The conclusion (see
group G,
[49, Corollary 2]) is that there are only finitely many conjugacy classes of maximal
closed subgroups of positive dimension in G: N

(a) Maximal parabolic subgroups.

(b) Normalizers of reductive subgroups of maximal rank—these are subgroups
N and have root system a subsystem of the root
containing a maximal torus of G,
N
system of G.
(c) A few further classes of (normalizers of) semisimple subgroups.
For example, when GN D E6 .K/ the subgroups under (a) are the parabolics Pi for
1  i  6; those under (b) are the normalizers of the subsystem subgroups A1 A5 ,
A32 , D4 T2 and T6 (a maximal torus); and those under (c) are the normalizers of
subgroups of types F4 , C4 , A2 , G2 , and A2 G2 .
In parallel with this, there are results which relate the subgroup structure of the
finite groups G.q/ with that of G. N The following is taken from [47, Corollary 8]:

Theorem 1.10. There is an absolute constant c such that if M is a maximal

subgroup of the exceptional group G.q/ D GN 0 , then one of the following holds:
(i) jM j < c;
(ii) M is a subfield subgroup G.q0 /, or a twisted subgroup (such as 2E6 .q 1=2 / <
E6 .q/);
(iii) M D MN  for some -stable maximal closed subgroup MN of GN of positive
dimension.
The maximal subgroups M under (ii), (iii) fall into at most d log log q conjugacy
classes of subgroups in G.q/; all satisfy jG.q/ W M j > d 0 q r (here r is the rank of
GN and d; d 0 are positive absolute constants).
This is nice, but it gives no information about the number of conjugacy classes of
bounded maximal subgroups under (i). The possibilities for these subgroups were
determined up to isomorphism in [48], but nothing much was proved about their
conjugacy until the work of Ben Martin [74], which was a major ingredient in the
proof of the following result, taken from [58, Theorem 1.2]:
Theorem 1.11. Let N; R be positive integers, and let G be a finite almost simple
group with socle a group of Lie type of rank at most R. Then the number of
conjugacy classes of maximal subgroups of G of order at most N is bounded by
a function f .N; R/ of N; R only.
Applying this to the finite exceptional groups of Lie type, and combining with
Theorem 1.10, we have:
Corollary 1.12. There is an absolute constant e such that the number of conjugacy
classes of maximal subgroups of any finite exceptional group G.q/ is bounded above
by e log log q.
8 M.W. Liebeck

This leads immediately to the proof of Theorem 1.5 for exceptional groups: for
s > 1, X
G .s/ D jG W M j1s < .e log log q/.d 0 q r /1s
reps: M

which tends to 0 as q ! 1. t
u
We conclude this section with a brief discussion of Theorem 1.11. The proof of
this uses geometric invariant theory, via the theory of strongly reductive subgroups
N a notion due to Richardson. These are closed subgroups H
of the algebraic group G,
which are not contained in any proper parabolic subgroup of CGN .T /, where T is a
maximal torus of CGN .H /; in particular, subgroups lying in no proper parabolic of GN
are strongly reductive. Martin’s main result in [74] is that the number of conjugacy
classes of strongly reductive subgroups of GN of order at most N is bounded by
a function g.N; R/, where R is the rank of G. N To adapt this to the analysis of
N
subgroups of the finite group G.q/ D G , the following was proved in [58, 2.2]:
Lemma 1.13. If F is a finite subgroup of GN which is invariant under , then
either F is strongly reductive, or F is contained in a -invariant proper parabolic
subgroup of G.N

The proof of this involves geometric invariant theory. The lemma implies that
N at which point
maximal non-parabolic subgroups of GN  are strongly reductive in G,
Martin’s result can be used to deduce Theorem 1.11.
We mention finally that Corollaries 1.9 and 1.12 can be used along with
arguments in [52] to prove the following, which is [58, 5.2]:
Theorem 1.14. The symmetric group has no.1/ conjugacy classes of primitive
maximal subgroups, and 12 n C no.1/ classes of maximal subgroups in total.

1.1.3 Other Results on Random Generation

We have discussed at length Dixon’s conjecture that for a finite simple group G,
P .G/ D Prob.hx; yi D G/ ! 1 as jGj ! 1. Since the conjecture was proved,
many variants have been established where one insists on various properties of the
generators x, y. We now briefly discuss some of these.
The first concerns the notion of .2; 3/-generation. A group G is said to be .2; 3/-
generated if it can be generated by two elements x; y such that x 2 D y 3 D 1.
It is well known that such groups are precisely the images of the modular group
PSL2 .Z/ (since PSL2 .Z/ is isomorphic to the free product of the cyclic groups C2
and C3 ). A question which goes back a long way is:
Problem. Which finite simple groups are .2; 3/-generated?
For simple alternating groups this was answered in 1901 by G.A. Miller [77], who
showed that An is .2; 3/-generated if and only if n ¤ 6; 7; 8. There is quite a large
literature on the problem for classical groups—for example, Tamburini et al. [88]
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 9

showed that many classical groups of large dimension are .2; 3/-generated. The
approach in these papers and many others is to produce explicit generators of the
required orders. Is there a probabilistic approach?
In [51], for any finite group G, Liebeck and Shalev defined P2;3 .G/ to be the
probability that two randomly chosen elements x, y of orders 2, 3 generate G. That
is, writing Ir .G/ for the set of elements of order r in G, and ir .G/ D jIr .G/j,

jf.x; y/ 2 I2 .G/  I3 .G/ W hx; yi D Ggj

P2;3 .G/ D :
i2 .G/i3 .G/

For finite simple groups G, we can attempt to estimate P2;3 .G/ using a similar
approach to the proof of Dixon’s conjecture. Given a maximal subgroup M of G,
the probability that a random pair x, y of elements of orders 2, 3 lies in M is
i2 .M /i3 .M /
i2 .G/i3 .G/ , and hence

X i2 .M /i3 .M /
1  P2;3 .G/  :
M max G
i2 .G/i3 .G/

We were able to prove that for alternating groups, and also for classical groups with
some low-dimensional exceptions,

i2 .M / i3 .M /
< cjG W M j2=5 ; < cjG W M j5=8
i2 .G/ i3 .G/

for all maximal subgroups M , where c is a constant. Hence, excluding the low-
dimensional exceptions,
X 41
1  P2;3 .G/  c 2 jG W M j2=55=8 D c 2 G . /;
M max G
40

which tends to 0 as jGj ! 1 by Theorem 1.5. In fact the low-dimensional

exceptions led to some interesting and unexpected counterexamples to .2; 3/-
generation, namely four-dimensional symplectic groups in characteristics 2 and 3.
The final result is [51, 1.4]:
Theorem 1.15. For G an alternating group, or a finite simple classical group not
isomorphic to PSp4 .q/, we have P2;3 .G/ ! 1 as jGj ! 1.
For G D PSp4 .2a / or PSp4 .3a / we have P2;3 .G/ D 0; while for G D PSp4 .p a /
with p > 3 prime, we have P2;3 .G/ ! 12 as p a ! 1.
As a consequence, all but finitely many classical groups (¤ PSp4 .2a /, PSp4 .3a /)
are .2; 3/-generated. The exceptional families PSp4 .2a /, PSp4 .3a / were a surprise
at the time, but are rather easily seen not to be .2; 3/-generated. For example,
consider G D PSp4 .q/ with q D 3a , and regard G as the isomorphic orthogonal
group ˝5 .q/ D ˝.V /. If x; y 2 G are elements of orders 2, 3 respectively, then
10 M.W. Liebeck

the 1-eigenspace of x on V has dimension at least 3, while the 1-eigenspace of y

has dimension 3. Hence these eigenspaces intersect nontrivially, and it follows that
hx; yi ¤ G.
There are many further results of this flavor in the literature. Here is a selection.
For a positive integer k, define Pk; .G/ to be the probability that G is generated by
a random element of order k and a random further element; so

jf.x; y/ 2 Ik .G/  G W hx; yi D Ggj

Pk; .G/ D :
ik .G/jGj

Let PC .G/ be the probability that G is generated by two randomly chosen

conjugates—that is, PC .G/ D Prob.hx; x y i D G for random x; y 2 G/.
Theorem 1.16. For all finite simple groups G,
(i) P2; .G/ ! 1 as jGj ! 1
(ii) P3; .G/ ! 1 as jGj ! 1
(iii) PC .G/ ! 1 as jGj ! 1.
Parts (i) and (ii) are taken from [53], while (iii) is the main result of [27].
We conclude this section by mentioning a couple of results of a slightly different
and rather useful style, in that they provide explicit constants. The first is taken
from [24]:
Theorem 1.17. In any finite simple group G there is a conjugacy class C such that
for any 1 ¤ x 2 G,

1
Prob.hx; ci D G for random c 2 C / > :
10

In other words, for every non-identity x 2 G, we have hx; ci D G for at least a

tenth of the elements c 2 C . As a consequence, for every 1 ¤ x 2 G there exists y
such that hx; yi D G (a property known as the 32 -generation of G).
The final result is taken from [76]:
53
Theorem 1.18. We have P .G/  90 for all finite simple groups G, with equality if
and only if G D A6 .
Further results on random generation of simple groups (such as “random
Fuchsian generation”) can be found in Sect. 1.3.

1.1.4 Generation of Maximal Subgroups

Having discussed the generation of finite simple groups, we move on to the

generation of their maximal subgroups. We know that d.G/ D 2 for every (non-
abelian) simple group, where d.G/ denotes the minimal number of generators of
G. What about d.M / for maximal subgroups M of G?
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 11

We have sketched some results about maximal subgroups of simple groups in

the previous sections. Much more complete information can be found in [40] for
classical groups and in [46] for exceptional groups of Lie type; in particular, the
only unknown maximal subgroups M are almost simple, and for these we have
d.M /  3 by [15]. And for alternating groups, the O’Nan–Scott theorem (see for
example the Appendix of [3]) shows that the primitive maximal subgroups fall into
several classes—affine type, product type, diagonal type and almost simple type—
so again the unknown ones are almost simple.
Despite being “known”, the non-almost simple maximal subgroups can have
quite intricate structures, which makes the proof of the following recent result, taken
from [11, Theorem 1], rather delicate.
Theorem 1.19. If G is a finite simple group, and M is a maximal subgroup of G,
then d.M /  4.
Equality can hold here. For example, let G D P ˝aC2 .q/ with a  2 mod 4
and q  1 mod 4. Then G has a maximal subgroup M of type OaC ˝ OaC , in
the tensor product family C7 (see [40, 4.7.6]), and the precise structure of M is
.P ˝aC .q/  P ˝aC .q//:24 , so clearly d.M /  4.
The theorem has quite a neat consequence concerning primitive permutation
groups (see [11, Theorem 7]): if G is a primitive group with point-stabilizer M ,
then d.M /  d.G/ C 4.
Having found the minimal number of generators, one might ask questions about
the random generation properties of maximal subgroups. Some of these will be
addressed in the next section (see Corollary 1.26).

1.2 Random Generation of Arbitrary Finite Groups

For a finite group G, recall that d.G/ is the minimal number of generators of G. For
k  d.G/ let dk .G/ be the number of generating k-tuples of elements of G. That
is,
dk .G/ D jf.x1 ; : : : ; xk / 2 G k W hx1 ; : : : ; xk i D Ggj:
Set
dk .G/
Pk .G/ D ;
jGjk
so that Pk .G/ is the probability that k randomly chosen elements generate G.
Following Pak [79], define
1
.G/ D minfk 2 N W Pk .G/  g
e
(where e is as usual the base of natural logarithms). The choice of the constant 1=e
here is for convenience, and is not significant; note that for any r  1, we have
Pr.G/ .G/  1  .1  1e /r .
12 M.W. Liebeck

A basic goal is to understand the relationship between .G/ and d.G/ for finite
groups G. We begin with a couple of examples.
Examples. 1. Let G be a p-group for some prime p, and let d D d.G/. For any
k, we have Pk .G/ D Pk .G=˚.G// D Pk .Cpd /. Taking k D d ,

pd  1 pd  p p d  p d 1 Y 1
d
Pd .G/ D Pd .Cpd / D d
: d
::: d
D .1  i /;
p p p 1
p

1 1
and it is easy to see that this product is at least 1  p
 p2
, which is greater than
1
when p > 2, and is when p D 2. A slight refinement gives Pd C1 .G/ > 1e in
e
1
4
all cases, and hence .G/  d.G/ C 1 for p-groups.
2. For simple groups G, Dixon’s conjecture says that P2 .G/ ! 1 as jGj ! 1,
which implies that .G/ D 2 for sufficiently large G. In fact, Theorem 1.18
gives .G/ D 2 for all finite simple groups G.
It is not too hard to generalize Example 1 to all nilpotent groups. This is done in
[79], where the much less easy case of soluble groups is also considered:
Theorem 1.20. (i) For G nilpotent, .G/  d.G/ C 1.
(ii) For G soluble, .G/  3:25 d.G/ C 107 .
Examples due to Mann [69] show that it is not possible to improve the constant
3:25 in part (ii) by much.
One might be tempted to think that for any finite group G, .G/ and d.G/ are
closely related—perhaps .G/ < c  d.G/ for some absolute constant c? This is in
fact far from being true, as is shown by the following result, taken from [37].
Lemma 1.21. For any real number R, there exists a finite group G such that
d.G/ D 2 and .G/ > R.
Proof. To prove the lemma, Kantor and Lubotzky construct such a group G of the
form T N , where T is a non-abelian simple group. A result of Philip Hall [28] shows
that the maximal value of N such that T N is 2-generated is d2 .T /=jAut.T /j. For
example, when T D A5 , Hall calculated that d2 .A5 /=jS5 j D 19, so that A19 5 is
2-generated whereas A20 5 is not.
Now let T D An . By Dixon’s Theorem 1.1, for large n, d2 .T / is at least 23 jT j2 ,
and so d2 .T /=jAut.T /j is at least 23 jAn j2 =jSnj D nŠ=6. Set N D nŠ=8, and define
G D T N . Then d.G/ D 2.
Now consider Pk .G/. The probability that a random k-tuple of elements of
An generates An is at most 1  n1k (since n1k is the probability that all of the k
permutations fix 1). If a random k-tuple in G D AN n generates G, then each of the
N k-tuples in a given coordinate position must generate An , and the probability that
this happens is at most .1  n1k /N D .1  n1k /nŠ=8 . For this to be at least 1e , k must be
of the order of n. Hence .G/ can be arbitrarily large, while d.G/ D 2. t
u
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 13

So it seems that it will be tricky to find a general relationship between .G/ and
d.G/. In [79], Pak proves that .G/  dlog2 jGje C 1 for all finite groups G, and
conjectures that there is a constant C such that .G/ < C  d.G/  log log jGj. This
was proved in a strong form by Lubotzky [66] (see also [15, Theorem 20]):
Theorem 1.22. For all finite groups G,

.G/  d.G/ C 2 log2 log2 jGj C 4:02:

We now discuss Lubotzky’s proof in some detail. Recall that mn .G/ denotes the
number of maximal subgroups of index n in the finite group G. Define

log mn .G/
.G/ D maxn2 :
log n

So mn .G/  n.G/ for all n, and we can think of .G/ as the “polynomial degree”
of the rate of growth of mn .G/.
Lemma 1.23. For any finite group G, .G/  d.G/ C 2:02e.
Proof. For any positive integer k, if x1 ; : : : ; xk denote randomly chosen elements
of G, we have

1  Pk .G/ D Prob.hx1 ; : : : ; xk i ¤ G/
P
 M max G Prob.x1 ; : : : ; xk 2 M /
P jk
D M max G jM
P jGj k

D n2 mn .G/nk
P
 n2 n.G/k :
P
Hence if k  .G/ C 2:02, then 1  Pk .G/  n2 n2:02 , which is less than
1  1e . Therefore Pk .G/  1e for such k, giving the result. t
u
By the lemma, to prove Theorem 1.22, it is sufficient to bound mn .G/ for
arbitrary finite groups G. Each maximal subgroup of index n in G gives a
homomorphism  W G ! Sn T with image .G/ a primitive subgroup of Sn and
kernel Ker./ D coreG .M / D g2G M g .
We call a maximal subgroup M core-free if coreG .M / D 1. According to a result
of Pyber, which appeared later in improved form as [59, Theorem 1.4], the number
of core-free maximal subgroups of index n in G is at most n2 (the improved form is
cn3=2 ). The proof relies on the detailed description of core-free maximal subgroups
given by Aschbacher and Scott in [3].
Now consider the non-core-free maximal subgroups. Each corresponds to a
homomorphism  W G ! Sn with .G/ primitive and Ker./ ¤ 1. To compute
mn .G/, we need to count the number of possibilities for Ker./ (and then multiply
by n2 , by Pyber’s result). To do this, we consider a chief series 1 D N0  N1 
    Nr D G (so Ni G G and each Ni =Ni 1 is minimal normal in G=Ni 1 ).
14 M.W. Liebeck

A finite group can have many chief series, but the number r, and the collection of
chief factors Ni =Ni 1 , are uniquely determined by G; as is the collection of normal
subgroups C1 ; : : : ; Cr , where Ci D CG .Ni =Ni 1 /, the kernel of the action of G on
Ni =Ni 1 .
By the O’Nan–Scott theorem (see [18, Theorems 4.3B, 4.7A]), there are three
possibilities for the structure of the primitive permutation group .G/:
1. .G/ has a unique minimal normal subgroup K, and K Š T k for some non-
abelian simple group T .
2. .G/ has exactly two minimal normal subgroups K1 ; K2 , and K1 Š K2 Š T k
for some non-abelian simple group T .
3. .G/ is an affine group: it has a unique minimal normal subgroup K Š Cpk for
some prime p.
In case (1), Lubotzky showed that Ker./ must be one of the r subgroups Ci ; and
in case (2), Ker./ must be Ci \ Cj for some i; j (see [66, 2.3]). Thus these cases
contribute at most 12 r.r C 1/n2 to mn .G/. Further argument [66, 2.5] shows that
case (3) contributes at most rnd.G/C2 , and hence
Lemma 1.24. For a finite group G which has r chief factors, mn .G/  r 2 nd.G/C2 .
Since r < log jGj, the lemma gives .G/  d.G/ C 2 C 2 log r  d.G/ C 2 C
2 log log jGj, so Theorem 1.22 follows using Lemma 1.23.
Building on Lubotzky’s ideas, and adding a lot more of their own, Jaikin-Zapirain
and Pyber proved the following remarkable result in [34], giving upper and lower
bounds for .G/ which are tight up to a multiplicative constant.
To state the result we need a few definitions. For a non-abelian characteristically
simple group A (i.e. A Š T k with T simple), denote by rkA .G/ the maximal
number r such that G has a normal section which is the product of r chief factors
isomorphic to A. Let l.A/ be the minimal degree of a faithful transitive permutation
representation of A. Finally, define

log rkA .G/

.G/ D maxA :
log l.A/

For example, if G D .An /t (n  5), then all chief factors are isomorphic to An ,
and rkAn .G/ D t, l.An / D n, so .G/ D log t= log n. On the other hand, if G D
An wr At (t  5), then a chief series is 1  .An /t  G, and rkA .G/ D 1 for both
chief factors of G, so .G/ D 0.
Theorem 1.25. There exist absolute constants ˛; ˇ > 0 such that for all finite
groups G,
˛.d.G/ C .G// < .G/ < ˇd.G/ C .G/:
If we return to the example G D .An /nŠ=8 in the proof of Lemma 1.21, the
theorem says that .G/ is of the order of log.nŠ/= log n, hence of the order of n
(while d.G/ D 2). On the other hand, for the wreath product G D An wr At , the
theorem tells us that .G/ is bounded.
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 15

Applications

We conclude this section by discussing a few applications of Theorem 1.25. The

first, taken from [11], is to the random generation of maximal subgroups of finite
simple groups. Recall from Theorem 1.19 that such a maximal subgroup M satisfies
d.M /  4. It is shown also in [11, 8.2] that M has at most three non-abelian chief
factors. Hence .M / is bounded by Theorem 1.25, and so we have
Corollary 1.26. Given  > 0, there exists k D k./ such that Pk .M / > 1   for
any maximal subgroup M of any finite simple group.
One might be tempted to think, in the spirit of Dixon’s conjecture, that there is a
constant k such that Pk .M / ! 1 as jM j ! 1. But this is not the case, as is shown
by the maximal subgroups Sn2 of An , for which Pk .Sn2 /  1  21k for any k.
The second application is to linear groups [34, 9.7]. Let G be a finite subgroup
of GLn .K/ for some field K. A result of Fisher [22] implies that the number of
non-abelian chief factors of G is less than n. Hence Theorem 1.25 gives
Corollary 1.27. There is an absolute constant c such that if G is any finite linear
group in dimension n over some field K, then .G/ < c  d.G/ C log n.
It is striking that the number of random generators does not depend on the
field K.
The third application relates .G/ to the sizes of so-called minimal generating
sets of G—that is, generating sets S such that no proper subset of S generates G.
Define dQ .G/ to be the maximal size of a minimal generating set of G. For example,
if G D Sn then f.1 2/; .2 3/; : : : ; .n1 n/g is a minimal generating set of size n1,
and a result of Whiston [89] shows that dQ .Sn / D n  1.
It is proved in [34, 9.9] that rkA .G/  dQ .G/ for any non-abelian characteristi-
cally simple group A, and hence Theorem 1.25 gives
Corollary 1.28. There is an absolute constant c such that for any finite group G,
.G/ < c  d.G/ C log dQ .G/.
This result has some significance in the analysis of the Product Replacement
Algorithm for choosing random elements in a finite group, since the quantities .G/
and dQ .G/ play a role in this analysis. We refer the reader to [80] for details.

1.3 Representation Varieties and Character-Theoretic

Methods

If is a finitely generated group, K a field and n a natural number, we call

Hom. ; GLn .K//, the set of representations W ! GLn .K/, the representation
variety of in dimension n over K. In this section we shall show how probabilistic
and character-theoretic methods can be brought to bear on the study of such varieties
over algebraically closed fields and also over finite fields, for a particular class
16 M.W. Liebeck

of finitely generated groups . We shall also consider the representation spaces

Hom. ; G.K// and Hom. ; Sn /, where G.K/ is a simple algebraic group over K,
which are of interest in various contexts.

1.3.1 Fuchsian Groups

The class of finitely generated groups we shall consider are the Fuchsian groups,
i.e. finitely generated discrete groups of isometries of the hyperbolic plane. By
classical work of Fricke and Klein, the orientation-preserving Fuchsian groups
have presentations of the following form:

generators: a1 ; b1 ; : : : ; ag ; bg (hyperbolic)
x1 ; : : : ; xd (elliptic)
y1 ; : : : ; ys (parabolic or hyperbolic boundary)

relations: x1m1 D    D xdmd D 1;

x1    xd y1    ys Œa1 ; b1     Œag ; bg  D 1

where g; d; s  0 and mi  2 for all i , and the measure . / > 0, where

X
d
1
. / D 2g  2 C s C .1  /:
1
m i

The number g is called the genus of .

There are also results along the lines discussed below for non-orientation
preserving Fuchsian groups, but for brevity’s sake we shall not mention these.
Examples. 1. Surface groups. These are the groups with s D d D 0 and g  2:
let Y
g D ha1 ; b1 ; : : : ; ag ; bg W Œai ; bi  D 1i
i

so that g D 1 .S /, the fundamental group of a surface S of genus g.

2. Triangle groups. These have g D s D 0, d D 3. For positive integers a, b, c
define
T D Ta;b;c D hx; y; z W x a D y b D zc D xyz D 1i;
where .T / D 1  a1  b1  1c > 0, and call this the .a; b; c/-triangle group.
1
The minimal value of  for a triangle group is 42 , which occurs for the Hurwitz
triangle group T2;3;7 .
3. Free products. When s > 0, is a free product of cyclic groups, such as the free
group Fr of rank r, or a free product Ca Cb D hx; y W x a D y b D 1i (which
has  D 1  a1  b1 ).
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 17

The Fuchsian groups for which s D 0 are said to be co-compact.

Let q be a prime power and K D FN q , the algebraic closure of Fq . In the sections
below, we shall discuss the representation spaces Hom. ; G/, where is a Fuchsian
group and G is GLn .K/, G.K/ (a simple algebraic group over K), G.q/ (a group of
Lie type over Fq ), or Sn . These representation spaces have many connections with
other areas, some of which we shall now point out.
The first connection is with the area of random generation. To say that a finite
group G is generated by two elements is to say that there exists an epimorphism in
the space Hom.F2 ; G/, where F2 is the free group of rank 2. The probability P .G/
(defined in Sect. 1.1.1) that G is generated by two randomly chosen elements is then

jf 2 Hom.F2 ; G/ W epigj
P .G/ D D Prob.random 2 Hom.F2 ; G/ is epi/:
jHom.F2 ; G/j

For any Fuchsian group , we can similarly define

P .G/ D Prob.random 2 Hom. ; G/ is epi/:

Dixon’s conjecture (Corollary 1.6) asserts that P .G/ ! 1 as jGj ! 1 for finite
simple groups G, and one could hope to prove similar results for the probabilities
P .G/. For example, a question posed many years ago asks which finite simple
groups are .2; 3; 7/-generated—that is, generated by three elements of orders 2, 3
and 7 with product equal to 1. There are many results on this question in which the
approach is to construct explicit generators, but one might hope to shed further light
by studying the probabilities P .G/, where G is simple and is the triangle group
T2;3;7 . Results on this and other probabilities P .G/ will be discussed below.
Other connections concern the representation space Hom. ; Sn /. The subgroup
growth of is measured by the growth of the function an . /, the number of
subgroups of index n in , and it is an elementary fact that

an . / D jHomtrans . ; Sn /j=.n  1/Š;

where Homtrans . ; Sn / is the set of homomorphisms ! Sn which have image

which is transitive on f1; : : : ; ng. Another connection is Hurwitz’s theory that
Hom. ; Sn / counts branched coverings of Riemann surfaces, where g is the genus
of the surface and the images of x1 ; : : : ; xd are the monodromy permutations around
the branch points (see [55, Sect. 8] for details).

1.3.2 Character Theory

The connection between the sizes of the spaces Hom. ; G/ and character theory
is given by the following lemma, which goes back to Frobenius and Hurwitz. Let
be a co-compact Fuchsian group with generators ai , bi , xi as above, and let G
18 M.W. Liebeck

be a finite group. For conjugacy classes Ci in G having representatives gi of order

dividing mi (1  i  d ), define C D .C1 ; : : : ; Cd / and

HomC . ; G/ D f 2 Hom. ; G/ W .xi / 2 Ci for 1  i  d g:

Denote by Irr.G/ the set of irreducible (complex) characters of G.

Lemma 1.29. With the above notation,
X .g1 /    .gd /
jHomC . ; G/j D jGj2g1 jC1 j    jCd j :
.1/2g2Cd
2I rr.G/

For a proof, see [55, 3.2]. For example, applying this with g D 0 and d D 3, we
obtain the well known formula of Frobenius that if C1 ; C2 ; C3 are three classes in
G, then the number of solutions to the equation x1 x2 x3 D 1 for xi 2 Ci is

jC1 jjC2 jjC3 j X .g1 / .g2 / .g3 /

: (1.6)
jGj .1/
2I rr.G/

When G is a finite simple group, a major tool in the analysis of Hom. ; G/ is

provided by the representation zeta function of G, defined for a real variable s by
X
 G .s/ D .1/s :
2I rr.G/

The next result is taken from [55, 1.1] for alternating groups, and from [56, 1.1, 1.2]
for groups of Lie type.
Theorem 1.30. (i) If G D An and s > 0, then  G .s/ ! 1 as n ! 1.
(ii) If G D G.q/ is of fixed Lie type with Coxeter number h, and s > h2 , then
 G .s/ ! 1 as q ! 1.
(iii) For any s > 0, there exists r D r.s/ such that for G D G.q/ of rank at least
r, we have  G .s/ ! 1 as jGj ! 1.
In part (ii), the Coxeter number h of G.q/ is defined to be the number of roots
in the root system of G.q/ divided by the rank; for example, if G D PSLn .q/,
PSp2n .q/ or E8 .q/ then h is n, 2n or 30, respectively. The bound h2 is tight, in that
 G.q/ .2= h/ is bounded away from 1. Moreover, if we P write rn .G/ for the number of
irreducible characters of degree n, then  G .s/ D n1 rn .G/ns , so for example
part (ii) implies that for G.q/ of fixed Lie type, rn .G.q// < cn2= h for all n, where
c is an absolute constant.
Theorem 1.30 also holds for the corresponding quasisimple groups—that is,
perfect groups G such that G=Z.G/ is a finite simple group.
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 19

1.3.3 Symmetric and Alternating Groups

Here we discuss some of the ramifications of the representation space

Hom. ; Sn /. Let us begin the story with a well known result of Conder [13],
inspired by the ideas of his supervisor Graham Higman: for n  168, An is
.2; 3; 7/-generated. (There is a reason for the number 168 here: A167 is not .2; 3; 7/-
generated.) This was part of a more general conjecture, attributed to Higman in the
1960s:
Higman’s Conjecture. For any Fuchsian group , there exists N. / such that
surjects onto An for all n > N. /.
Conder’s result covers the case D T2;3;7 . Higman’s conjecture was eventually
proved in 2000 by Everitt [21]. One can quickly reduce to the case of genus 0—
for example, a Fuchsian group of genus g  2 maps onto the free group F2 on
generators x; y (send a1 ! x; a2 ! y and all other generators to 1), and then F2
maps onto An . For the genus 0 case, Everitt’s approach was based on constructing
generators using Higman’s method of coset diagrams.
Inspired by Everitt’s result, Liebeck and Shalev started thinking a few years later
about whether there might be probabilistic approach to Higman’s conjecture: if one
could show that P .An / D Prob. 2 Hom. ; An / is epi/ tends to 1 (or indeed
anything nonzero) as n ! 1, then of course Higman’s conjecture would follow.
Our approach to studying P .An / was similar to the approach to Dixon’s
conjecture. Namely, if 2 Hom. ; An / is not an epimorphism, then . /  M
for some maximal subgroup M of An ; and given M , this happens with probability
jHom. ; M /j=jHom. ; An /j. Therefore
X jHom. ; M /j
1  P .An /  :
M max An
jHom. ; An /j

Hence the task was to estimate jHom. ; An /j and also jHom. ; M /j for maximal
subgroups M . Notice that these estimates must be delicate enough to distinguish
between jHom. ; An /j and jHom. ; An1 /j, since An1 is one possibility for M:
Here is an example of an ingredient of how jHom. ; An /j is estimated, for the
triangle group D T2;3;7 , taken from [55]. For convenience take n to be divisible
by 2  3  7 and such that the conjugacy classes C1 ; C2 ; C3 consisting of fixed-point-
free permutations of shapes .2n=2 /, .3n=3 /, .7n=7 / respectively, all lie in An . One can
prove that for r D 2; 3; 7 the following hold:
1
(i) jCr j .nŠ/1 r .
(ii) For any 2 Irr.An / and cr 2 Cr , we have j .cr /j < cn1=2  .1/1=r .
Part (i) is routine, but (ii) is hard and uses much of the well-developed character
theory of symmetric groups. If we ignore the cn1=2 term in (ii), then, writing C D
.C1 ; C2 ; C2 /, the formula (1.6) gives
20 M.W. Liebeck

1 2
.nŠ/ 2 C 3 C 7
6
X 1 1 1
.1/ 2 C 3 C 7
jHomC . ; An /j  .1  /
nŠ .1/
1¤ 2I rr.An /
43
D .nŠ/ 42 .1  . An . 42
1
/  1//:

1 1
Now . / D .T2;3;7 / D 42 , and  An . 42 / ! 1 as n ! 1 by Theorem 1.30(i),
so for large n this shows that jHom. ; An /j is at least roughly .nŠ/1C. / . Adapting
this calculation to take care of the cn1=2 term in (ii) is a fairly routine technical
matter.
It turns out that .nŠ/1C. / is the correct order of magnitude for jHom. ; An /j
and also jHom. ; Sn /j. The following result is [55, Theorem 1.2].
Theorem 1.31. For any Fuchsian group we have jHom. ; Sn /jD.nŠ/1C. /Co.1/
.
In fact some much more precise estimates were obtained (and were needed, as
remarked above). We were also able to prove that the subgroup growth function
an . / D jHomtrans . ; Sn /j=.n  1/Š satisfies an . / D .nŠ/. /Co.1/ , and establish
the following probabilistic result ([55, Theorem 1.7]).
Theorem 1.32. For any Fuchsian group , the probability that a random homo-
morphism in Homtrans . ; Sn / is an epimorphism tends to 1 as n ! 1.
This implies Higman’s conjecture, which was our original motivation. The
character-theoretic methods and estimates in [55] for symmetric and alternating
groups have been developed and improved in a number of subsequent papers,
notably Larsen–Shalev [42], where very strong estimates on character values are
obtained and a variety of applications given; in particular they solve a number of
longstanding problems concerning mixing times of random walks on symmetric
groups. But we shall not go into this here.

1.3.4 Groups of Lie Type

We now discuss results analogous to Theorems 1.31 and 1.32 for groups G D G.q/
of Lie type. Let be a co-compact Fuchsian group as in Sect. 1.3.1. Again we seek
to apply Lemma 1.29. If gi (1  i  d ) are elements of G of order dividing mi ,
and 2 Irr.G/, then j .g 1 / .gd /j
.1/2g2Cd
 .1/.2g2/ , which quickly yields

X .g1 /    .gd /
2   G .2g  2/    G .2g  2/:
.1/2g2Cd
2I rr.G/

This observation is only useful when g  2, in which case Theorem 1.30 gives
 G .2g  2/ ! 1 as jGj ! 1. In this case, applying Lemma 1.29 and summing
over all classes of elements of orders dividing m1 ; : : : ; md leads to the following
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 21

result, which is part of [57, 1.2, 1.4]. In the statement, jm .G/ denotes the number of
elements of G of order dividing m.
Theorem 1.33. Let be a co-compact Fuchsian group of genus g  2 as in
Sect. 1.3.1.
(i) For all finite quasisimple groups G,

Y
d
jHom. ; G/j D .1 C o.1//  jGj2g1  jmi .G/;
1

where o.1/ refers to a quantity which tends to 0 as jGj ! 1.

1
(ii) For groups G of Lie type of rank r, jHom. ; G/j D jGj1C. /CO. r / .
Part (ii) follows from (i), since one can show that for groups G of large rank,
1
jm .G/ is roughly jGj1 m .
As in the previous section, the analysis of the probabilities P .G/ is much
harder, and for groups of Lie type has only been completed for the case where
g  2. Here is [57, Theorem 1.6].
Theorem 1.34. Let be a co-compact Fuchsian group of genus g  2. Then for
all finite simple groups G, the probability P .G/ ! 1 as jGj ! 1.
Note that the conclusion of the theorem does not remain true for genus 0 or 1,
since there are Fuchsian groups of such genus which do not have all sufficiently
large finite simple groups as quotients. For example, a result of Macbeath [68] says
that PSL2 .q/ can only be an image of the genus 0 group T2;3;7 if q D p or p 3 for
some prime p. Nevertheless, the genus 0 or 1 case does lead to some very interesting
questions which we shall discuss below in Sect. 1.3.6. For the moment, we conclude
this section by stating a conjecture from [57]:
Conjecture. For any Fuchsian group there is an integer f . / such that for finite
simple classical groups G of rank at least f . /, we have P .G/ ! 1 as jGj ! 1.

1.3.5 Representation Varieties

Let be a Fuchsian group, and let K D FN p , the algebraic closure of Fp , where p

is prime. Recall that the representation variety of in dimension n over K is the
variety V D Hom. ; GLn .K//. One of the most basic questions about V is: what
is the dimension of V ?
This question can be attacked using results in the previous section. Here’s
how. Let q be a power of the characteristic p, and define the field morphism
q
 W GLn .K/ ! GLn .K/ to be the map sending the matrix .aij / to .aij /. The fixed
point group of  is GLn .K/ D GLn .q/. Observe that  also acts on the variety
22 M.W. Liebeck

V : for 2 V; 2 , define  . / D . / . Then the fixed point set V D V .q/ is

the finite representation variety Hom. ; GLn .q//.
When the genus of is at least 2, the size of Hom. ; GLn .q// can be estimated
as in the previous section. The connection between this and the dimension of V is
given by a classical result of Lang–Weil [41]:
Lemma 1.35. Let V be an algebraic variety over FN p of dimension f , with e
components of dimension f . For a power q of p, let V .q/ be the set of q-rational
points in V . Then there is a power q0 of p such that jV .q/j D .e C o.1//q f for all
powers q of q0 .
In estimating jV .q/j D jHom. ; GLn .q//j we need to know the limiting be-
haviour of the zeta function  GLn .q/ .s/. This is a little different from Theorem 1.30,
because GLn .q/ has q  1 linear characters (which contribute q  1 to  GLn .q/ .s/).
The outcome is [57, 2.10]: for s  2 and fixed n,

 GLn .q/ .s/ ! q  1 C ı as q ! 1;

where ı D 1 if n D s D 2 and ı D 0 otherwise. Using this, jHom. ; GLn .q//j can

be computed as in the previous section. The conclusion is a little awkward to state
in general, so we just state it for surface groups and refer the reader to [57, 3.8] for
the general case.
Proposition 1.36. If is a surface group of genus g  2, then for fixed n  2,
jHom. ; GLn .q//j D .q  1 C ı C o.1//  jGLn .q/j2g1 .
By Lemma 1.35, this implies
Theorem 1.37. Let be a surface group of genus g  2, and let V be the
representation variety Hom. ; GLn .K//. Then dim V D .2g  1/n2 C 1, and V
has a unique irreducible component of highest dimension.
Arguing similarly for the varieties Hom. ; G/,N where GN D G.K/ is a simple
algebraic group over K, we obtain the following, which is [57, 1.10]. In the
N is the subvariety of elements x 2 GN satisfying x m D 1. The
statement, Jm .G/
dimensions of these subvarieties are studied by Lawther in [45]; in particular,
N tends to .1  1 / dim GN as the rank of GN tends to infinity.
dim Jm .G/ m

Theorem 1.38. Let be a Fuchsian group of genus g  2 as in Sect. 1.3.1, and let
GN D G.K/ be a simple algebraic group. If V is the variety Hom. ; G/,
N then

X
d
dim V D .2g  1/ dim GN C N
dim Jmi .G/:
1

Other results along these lines can be found in [57].

1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 23

1.3.6 Triangle Groups

All the results in the previous two sections concerning the spaces Hom. ; G/ for
G a finite or algebraic group of Lie type assume that the genus of is at least 2.
This is not because the genus 0 or 1 cases are uninteresting, but rather because the
character-theoretic methods in these cases require much more delicate information.
For example, to estimate the sum in the formula (1.6), one needs information on
the character values .g/ for irreducible characters of G, and usable estimates of
character values are hard to come by.
As a result, rather little is known about the spaces Hom. ; G/ and their
ramifications when G is a group of Lie type and has genus 0 or 1. (Note however
that this is not so for G D Sn or An , since the results of Sect. 1.3.3 hold for all
genera.) Nevertheless there are some results and conjectures which make this a very
interesting case.
We shall focus on the case for which most is known—namely that in which is
a triangle group Ta;b;c (of genus 0).

Triangle Generation

Let T D Ta;b;c D hx; y; z W x a D y b D zc D xyz D 1i be a Fuchsian triangle group

(so a1 C b1 C 1c < 1). We shall say that a finite group G is .a; b; c/-generated if it
is an image of T . And for a family of simple groups G.q/ of fixed Lie type, we say
that G.q/ is randomly .a; b; c/-generated if

PT .G.q// D Prob.random 2 Hom.T; G.q// is epi/ ! 1

as q ! 1 through values for which a; b and c divide jG.q/j.

Question. Which finite simple groups of Lie type are (randomly) .a; b; c/-
generated?
This question goes back quite a long way, particularly in the case where
.a; b; c/ D .2; 3; 7/—indeed, .2; 3; 7/-generated groups are also called Hurwitz
groups, and are of interest because they are precisely the groups which realize a
well known upper bound of Hurwitz for the number of automorphisms of a compact
Riemann surface (see [14] for example).
The first substantial result on this question was that of Macbeath [68], who
showed that PSL2 .q/ is .2; 3; 7/-generated if and only if either q D p 
0; ˙1 mod 7, or q D p 3 and p  ˙2; ˙3 mod 7, where p denotes a prime. Many
further results on .2; 3; 7/-generation have followed since—for example, PSL3 .q/
is only .2; 3; 7/-generated if q D 2, while SLn .q/ is .2; 3; 7/-generated for all
n  287 (see [87] for a survey).
Concerning .a; b; c/-generation for more general values, nothing much was done
until the following two results of Marion [70, 72]. In both results, assume a; b; c are
primes and a  b  c.
24 M.W. Liebeck

Theorem 1.39. Given a prime p, there is a unique power p r such that PSL2 .p r /
is .a; b; c/-generated—namely, the minimal power such that a, b and c divide
jPSL2 .p r /j.
A few moments’ thought show that this agrees with Macbeath’s result on
.2; 3; 7/-generation stated above. In particular, the theorem shows that PSL2 .q/
is far from being randomly .a; b; c/-generated.
For three-dimensional classical groups Marion proved
Theorem 1.40. Let G D PSL3 .q/ or PS U3 .q/.
(i) If a > 2, then G is randomly .a; b; c/-generated.
(ii) If a D 2, then given a prime p, there are at most four values of q D p r such
that G is .a; b; c/-generated.
The proofs of these two theorems are character-theoretic. The second theorem
takes a great deal of effort, and appears in a series of three papers [72]. Using the
same methods to tackle higher dimensional groups is not an appetising prospect.
The dichotomy in parts (i) and (ii) of Theorem 1.40 is quite striking. In seeking
to explain it, Marion introduced the idea of rigidity.

Rigidity

N a simple algebraic group over K D FN p , p prime. For

Now switch attention to G,
a  2, define
N
ıa D maxfdim x G W x 2 GN of order ag:
Straightforward matrix calculations give
Lemma 1.41. (i) If GN D PSL2 .K/ then ıa D 2 for all a  2.
(ii) If GN D PSL3 .K/ then ı2 D 4, while ıa D 6 for a > 2.
For example, consider G D PSL3 .K/ with char.K/ ¤ 2. Any involution t 2 G
is conjugate to the image modulo scalars of the diagonal matrix diag.1; 1; 1/;
then dim CG .t/ D 4 and so dim t G D 4. On the other hand, for any odd prime
a ¤ char.K/ there is an element u 2 G of order a having distinct eigenvalues, so
that dim CG .u/ D 2 and dim uG D 6.
The relevance of the above definition to .a; b; c/-generation is given by
N then G.q/ is not .a; b; c/-generated
Proposition 1.42. If ıa C ıb C ıc < 2 dim G,
for any q.
This is [71, Proposition 1]. The proof is an application of a well known result of
Scott [82] which is one of the main tools in this whole area. Here is a sketch for the
N
case where p is a “very good prime” for G—this means that p is not 2 when GN is
symplectic or orthogonal, p is not 2 or 3 when GN is of exceptional type (and also
not 5 when GN D E8 ), and p does not divide n when GN D PSLn .K/.
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 25

We consider the action of GN on its Lie algebra V D L.G/. N Under the assumption
that p is very good, G.q/ acts irreducibly on V for any q, and also dim CV .g/ D
dim CGN .g/ for all g 2 GN (see [12, 1.14]). If G.q/ is generated
P by x1 , x2 , x3 of orders
a, b, c with x1 x2 x3 D 1, then Scott’s result implies that .dim V  dim CV .xi // 
P
2 dim V . As dim V D dim G, N this means that dim x GN  2 dim G, N contradicting
i
the hypothesis that ıa C ıb C ıc < 2 dim G. N
The proposition motivates the following definition.
1 1 1
Definition. We say that a triple .a; b; c/ of primes (with a
C b
C c
< 1) is rigid
for GN if ıa C ıb C ıc D 2 dim G.
N

For example, if GN D PSL2 .K/ then dim GN D 3, so by Lemma 1.41, every triple
N On the other hand, if GN D PSL3 .K/ then dim GN D 8 and triples
is rigid for G.
.2; b; c/ are rigid, while triples .a; b; c/ with a > 2 are not (recall that a  b  c).
In view of these examples, Theorems 1.39 and 1.40 support the following
conjecture, stated in [71].
Conjecture. Let p be a prime and GN a simple algebraic group over K D FN p .
N Then there are only finitely many
Suppose .a; b; c/ is a rigid triple of primes for G.
powers q D p r such that G.q/ is .a; b; c/-generated.
Note that the converse to the conjecture does not hold: for example, it is known
that SL7 .q/ is never a Hurwitz group (see [75]), but .2; 3; 7/ is not a rigid triple for
SL7 .K/. It would be very interesting to find a variant of the conjecture which could
work both ways round.
Theorem 3 of [71] classifies all the rigid triples of primes for simple algebraic
groups. The list is not too daunting: for GN D PSLn or SLn , rigid triples exist only
for n  10; for symplectic groups, only for dimensions up to 26 (and only the
Hurwitz triple .2; 3; 7/ can be rigid beyond dimension 10); for orthogonal groups,
the only rigid triple is .2; 3; 7/ for the groups Spin11;12 ; and the only exceptional
type which has a rigid triple is G2 with triple .2; 5; 5/.
As remarked above, one would not want to adopt the character-theoretic method
of proof of Theorems 1.39 and 1.40 for larger rank cases. Fortunately there is
another tool which can be used to attack the conjecture, namely the classical notion
of rigidity for algebraic groups.

Classical Rigidity

Let GN be a simple algebraic group over an algebraically closed field K, and let
N Define
C1 ; : : : ; Cr be conjugacy classes in G.

C0 D f.x1 ; : : : ; xr / W xi 2 Ci ; x1    xr D 1g:

Following [86], we say that .C1 ; : : : ; Cr / is a rigid tuple of classes if C0 ¤ ; and G

acts transitively on C0 by conjugation.
Rigid tuples play a major role in inverse Galois theory and other areas. Their
relevance to triangle generation and in particular to the above Conjecture is via the
26 M.W. Liebeck

following observations. Assume for convenience of discussion that the characteristic

N
of K is a very good prime for G.
1. If .C1 ; C2 ; C3 / is a rigid triple of classes, and CL.G/ N .x1 ; x2 ; x3 / D 0 for
P
N
.x1 ; x2 ; x3 / 2 C0 , then dim Ci D 2 dim G (see [86, 3.2]).
2. Suppose one can prove aP converse to (1)—namely, that if .C1 ; C2 ; C3 / is a
triple of classes such that dim Ci D 2 dim GN and CL.G/ N .x1 ; x2 ; x3 / D 0 for
.x1 ; x2 ; x3 / 2 C0 , then .C1 ; C2 ; C3 / is a rigid triple.
3. Now let .a; b; c/ be a rigid triple of primes. Simple algebraic groups have only
finitely many classes of elements of any given order, and hence GN has finitely
many triples .C1 ; C2 ; C3 / of classes of elements of orders a; b; c satisfying
dim Ci D 2 dim GN (and the other triples of such classes have dimensions sum-
ming to less than 2 dim G). N

Given (1), (2) and (3), define T to be the set of triples .g1 ; g2 ; g3 / ofPelements of
GN of orders a; b; c with product 1 such that CL.G/N .g ;
1 2g ; g3 / D 0 and dim Ci D
N
N where Ci D g . For .g1 ; g2 ; g3 / 2 T , the triple of classes .C1 ; C2 ; C3 / is
G
2 dim G, i
one of the finitely many in (3), and is rigid by (2). It follows that GN has only finitely
many orbits in its conjugation action on T , and so there can be only finitely many
groups G.q/ generated by such triples g1 ; g2 ; g3 , proving the Conjecture.
Hence, if we can prove the statement in (2), then we can prove the Conjecture.
Unfortunately, the only known case of (2) is the following result of Strambach and
Völklein [86, 2.3] for GN D SLn ; in the case of characteristic zero it goes back to
Katz [39].
N
P Let G D SLn .K/, and let .C1 ; : : : ; Cr / be a tuple of classes in
Theorem 1.43.
N
G such that dim Ci D 2 dim GN and hx1 ; : : : ; xr i is irreducible on the natural
module Vn .K/ for some .x1 ; : : : ; xr / 2 C0 . Then .C1 ; : : : ; Cr / is a rigid tuple.
The Conjecture follows in the case where GN D SLn (see [71] for details). Further
cases are handled in [71], but quite a few remain open. The most elegant way to
finish the proof would be to prove a version of Theorem 1.43 for all types of simple
algebraic groups, but this seems difficult. One further case—that in which GN D
G2 in characteristic 5 and .a; b; c/ D .2; 5; 5/ (the only rigid triple of primes for
exceptional types)—was handled in [62].
Recently, Larsen et al. [43] have introduced the method of deformation theory
into the picture, and used it to prove Marion’s conjecture for all cases where the
underlying characteristic does not divide the product abc of the exponents of the
triangle group.

1.4 Cayley Graphs of Simple Groups: Diameter and Growth

Let G be a finite group with a generating set S which is symmetric—that is, closed
under taking inverses—and does not contain the identity. The Cayley graph .G; S /
is defined to be the graph with vertex set G and edges fg; gsg for all g 2 G, s 2 S .
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 27

It is connected and regular of valency jS j, and G acts regularly on .G; S / by

left multiplication. Because of the transitive action of G, the diameter of .G; S /,
denoted by diam.G; S /, is equal to the maximum distance between the identity
element and any g 2 G, and so

diam.G; S / D maxfl.g/ W g 2 Gg

where l.g/ is the length of the shortest expression

S of g as a product of elements of
S . If d D diam.G; S /, then G D feg [ drD1 S r (where S r D fs1    sr W si 2 S g),
P
and so jGj  drD0 jS jr < jS jd C1 . Hence

log jGj
diam.G; S / >  1: (1.7)
log jS j

Examples. 1. Let G D Cn D hxi, a cyclic group of order n, and let S D fx; x 1 g.

jGj n
Then .G; S / is an n-gon. So diam.G; S / D Πn2 , whereas log log jS j
D log
log 2
.
2. Let G D Sn and S be the set of all transpositions. Here diam.G; S / D n  1,
jGj
while loglog jS j is roughly 2 .
n

3. Let G D Sn and S D f.1 2/; .1 2    n/˙1 g. In this case diam.G; S / is roughly

jGj
n2 , while log log jS j
is of the order of n log n. The same orders of magnitude apply to
a similar generating set for An consisting of a 3-cycle and an n- or .n  1/-cycle
and their inverses.
4. For G D SLn .q/ and S the set of transvections, we have diam.G; S / n and
log jGj n
log jS j 2 .
5. Let G D SLn .p/ (p prime) and S D fx ˙1 ; y ˙1 g where
0 1 0 1
1 1 0 1
B 1 C B 0 0 1 C
B C B C
B C B C
B : C B : C
xDB C; yDB C:
B : C B : C
B C B C
@ : A @ 1A
1 ˙1

log jGj
Then log jS j n2 log p, and also diam.G; S / n2 log p.
All the above examples are elementary except the last, where the fact that
diam.G; S /  C n2 log p for some constant C is a result of Kassabov and Riley
[38].
Define diam.G/ to be the maximum of diam.G; S / over all generating sets S .
The main conjecture in the field is due to Babai, and appears as Conjecture 1.7
in [6]:
Babai’s Conjecture. There is a constant c such that diam.G/ < .log jGj/c for any
non-abelian finite simple group G.
28 M.W. Liebeck

It can be seen from Example 3 above that c must be at least 2 for the conjecture
to hold.
There has been a great deal of recent progress on this conjecture, but before
presenting some of this we shall discuss the special case where S is a union of
conjugacy classes, which has various other connections.

1.4.1 Conjugacy Classes

In the case where the generating set S is a union of classes (which occurs in
Examples 2 and 4 above), a strong form of Babai’s conjecture holds:
Theorem 1.44. There is a constant C such that for any non-abelian finite simple
group G and any non-identity union S of conjugacy classes of G,

log jGj
diam.G; S / < C :
log jS j

jGj
Indeed, G D S k for all k  C log
log jS j
.
This is the main theorem of [54]. In view of (1.7), the diameter bound is best
possible, apart from reduction of the constant C .

Consequences

First we point out an obvious consequence. If S is the set of involutions in a simple

group G, then of course S is a union of classes, and it is not hard to prove that
jS j > cjGj1=2 (see [51, 4.2, 4.3]). Hence Theorem 1.44 implies that every element
of every simple group is a product of k involutions, for some absolute constant k.
The same holds for elements of any fixed order.
A more substantial consequence concerns word maps on simple groups. For a
group G and a nontrivial word w D w.x1 ; : : : ; xd / in the free group of rank d ,
define
w.G/ D fw.g1 ; : : : ; gd / W gi 2 Gg;
the set of w-values in G. For example, if w D Œx1 ; x2  or x1k , then w.G/ is the set of
commutators or kth powers in G. Clearly w.G/ is a union of classes of G.
Given w, it is possible to show that there is a constant c D c.w/ > 0 depending
only on w, such that jw.G/j > jGjc for all simple groups G such that w.G/ ¤ 1
(see [54, 8.2]). Hence Theorem 1.44 gives
Corollary 1.45. For any nontrivial word w, there is a constant c D c.w/ such that
w.G/c D G for every finite simple group G for which w.G/ ¤ 1.
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 29

A result of Jones [36] ensures that w.G/ ¤ 1 provided G is sufficiently large

(i.e. provided jGj > f .w/ where this depends only on w). Recent work of Larsen,
Shalev and Tiep, culminating in [44], shows that the number c.w/ in the corollary
can be replaced by 2. So for example, every element of every sufficiently large
simple group is a product of two commutators or two kth powers, and so on.
Notice that for certain words w, it is definitely not possible to replace c.w/ by
1. For example, the kth power word map x1k is not surjective on any finite group
of order not coprime to k (since it is not injective). To date, there are just a few
instances of word maps which have been shown to be surjective on all simple
groups. The first was the commutator word: it was proved in [61] that every element
of every non-abelian finite simple group is a commutator, a result known as the
p p
Ore conjecture. Some further surjective word maps such as x1 x2 (p prime) are
produced in [26, 63]. One might conjecture that every non-power word map is
surjective on sufficiently large simple groups, but this has recently been shown to be
false in [35]: for example, the word map .x; y/ ! x 2 Œx 2 ; y 1 2 is non-surjective
on PSL2 .p 2rC1 / for all non-negative integers r and all odd primes p such that
p 2 6 1 mod 16 and p 2 6 0; 1 mod 5.
This is one of a number of “width” questions about simple groups. Another is the
conjecture proposed in [64], that if A is any subset of size at least 2 in a finite simple
group G, then G is a product of N conjugates of A for some N  c log jGj= log jAj,
where c is an absolute constant. This has been proved in some cases in [23, 60, 64].

1.4.2 Babai’s Conjecture

There have been spectacular recent developments on Babai’s conjecture, both for
groups of Lie type and for alternating groups. We shall discuss these separately.

Groups of Lie Type

For a long time, even SL2 .p/ (p prime) was a mystery as far as proving Babai’s
conjecture was concerned. Probably the first small (symmetric) generating set one
thinks of for this group is
 ˙1  ˙1
1 1 1 0
S Df ; g:
0 1 1 1

Babai’s conjecture asserts that diam.G; S / < .log p/c for these generators. Surely
this must be easy?
In fact it is not at all easy, and was proved by the following beautiful but
indirect method (see [65]). First observe that the matrices in S , when regarded as
integer matrices, generate SL2 .Z/. Now let .p/ denote the congruence subgroup
30 M.W. Liebeck

which is the kernel of the natural map SL2 .Z/ ! SL2 .p/. If H is the upper half
plane and X.p/ denotes the Riemann surface .p/nH, denote by 1 .X.p// the
smallest eigenvalue for the Laplacian on X.p/. A theorem of Selberg [83] gives
3
1 .X.p//  16 for all p, and this can be used to show that the Cayley graphs
f p D .SL2 .p/; S / W p primeg have their second largest eigenvalues bounded
away from the valency, and hence that they form a family of expander graphs. This
means that there is an expansion constant c > 0, independent of p, such that for
every set A consisting of fewer than half the total number of vertices in p , we have
jıAj > cjAj, where ıA is the boundary of A—that is, the set of vertices not in A
which are joined to some vertex in A. From the expansion property it is easy to
deduce that p has logarithmic diameter, so that diam. .SL2 .p/; S // < c log p, a
strong form of Babai’s conjecture.
One can adopt essentially the same method for the generators
 ˙1  ˙1
1 2 1 0
f ; g
0 1 2 1

of SL2 .p/, since, while these do not generate SL2 .Z/, they do generate a subgroup
of finite index therein. But what if we replace the 2s in these generators with 3s?
Then the matrices generate a subgroup of infinite index in SL2 .Z/, and the above
method breaks down. This question became known as Lubotzky’s 1-2-3 problem,
and was not solved until the breakthrough achieved by Helfgott [30]:
Theorem 1.46. Babai’s conjecture holds for G D SL2 .p/. That is, diam.SL2 .p//
< .log p/c , where c is an absolute constant.
Helfgott deduced this from his key proposition: for any generating set S of
G D SL2 .p/, either jS 3 j > jS j1C , or S k D G, where  > 0 and k do not depend on
p. (Later it was observed that one can take k D 3 here.) The heart of his proof is to
relate the growth of powers of subsets A of G with the growth of the corresponding
set of scalars B D tr.A/ D ftr.x/ W x 2 Ag in Fp under sums and products. By
doing this he could tap into the theory of additive combinatorics, using results such
as the following, taken from [9]: if B is a subset of Fp with p ı < jBj < p 1ı for
some ı > 0, then jB  Bj C jB C Bj > jBj1C , where  > 0 depends only on ı.
Following Helfgott’s result, there was a tremendous surge of progress in this area.
Many new families of expanders were constructed in [8]. Helfgott himself extended
his result to SL3 in [31], and this has now been proved for all groups of Lie type of
bounded rank in [10, 81]. As a consequence, we have
Theorem 1.47. If G D G.q/ is a simple group of Lie type of rank r, then
diam.G/ < .log jGj/c.r/ where c.r/ depends only on r.
Again, the theorem is proved via a growth statement: for any generating set S of
G.q/, either jS 3 j > jS j1C , or S 3 D G, where  > 0 depends only on r.
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 31

These results, and particularly their developments into the theory of expanders,
have many wonderful and surprising applications. For a survey of these develop-
ments and some of the applications, see [67].
Finally, let us remark that Babai’s conjecture remains open for groups of Lie type
of unbounded rank.

Alternating Groups

For the alternating groups An , Babai’s conjecture is that there is a constant C such
that diam.An / < nC . Until very recently, the best bound for diam.An / was that
obtained by Babai and Seress in [5], where it was proved that

diam.An / < exp..1 C o.1//  .n log n/1=2 / D exp..1 C o.1//  .log jAn j/1=2 /:

Babai and Seress also obtained a bound of the same magnitude for the diameter of an
arbitrary subgroup of Sn in [6]; this is best possible, as can be seen by constructing
a cyclic subgroup generated by a permutation with many cycles of different prime
lengths. Various other partial results appeared at regular intervals, such as that in [7],
where it was shown that if the generating set S contains a permutation of degree at
most 0:33n, then diam.An ; S / is polynomially bounded. But no real progress was
made on Babai’s conjecture until a recent breakthrough of Helfgott and Seress [32]:
Theorem 1.48. We have diam.An /  exp.O..log n/4 log log n//, where the im-
plied constant is absolute.
This does not quite prove Babai’s conjecture, but it does prove that diam.An /
is “quasipolynomial” (where a quasipolynomial function f .n/ is one for which
log f .n/ is polynomial in log n), which represents a big step forward. The same
paper also gives a bound of the same magnitude for the diameter of any transitive
subgroup of Sn .

References

1. M. Aschbacher, On the maximal subgroups of the finite classical groups. Invent. Math. 76,
469–514 (1984)
2. M. Aschbacher, R. Guralnick, Some applications of the first cohomology group. J. Algebra 90,
446–460 (1984)
3. M. Aschbacher, L. Scott, Maximal subgroups of finite groups. J. Algebra 92, 44–80 (1985)
4. L. Babai, The probability of generating the symmetric group. J. Comb. Theor. Ser. A 52, 148–
153 (1989)
5. L. Babai, A. Seress, On the diameter of Cayley graphs of the symmetric group. J. Comb. Theor.
Ser. A 49, 175–179 (1988)
6. L. Babai, A. Seress, On the diameter of permutation groups. Eur. J. Comb. 13, 231–243 (1992)
32 M.W. Liebeck

7. L. Babai, R. Beals, A. Seress, in On the Diameter of the Symmetric Group: Polynomial Bounds.
Proceedings of the Fifteenth Annual ACM-SIAM Symposium on Discrete Algorithms (ACM,
New York, 2004), pp. 1108–1112
8. J. Bourgain, A. Gamburd, Uniform expansion bounds for Cayley graphs of SL2 .Fp /. Ann.
Math. 167, 625–642 (2008)
9. J. Bourgain, N. Katz, T. Tao, A sum-product estimate in finite fields, and applications. Geom.
Funct. Anal. 14, 27–57 (2004)
10. E. Breuillard, B. Green, T. Tao, Approximate subgroups of linear groups. Geom. Funct. Anal.
21, 774–819 (2011)
11. T.C. Burness, M.W. Liebeck, A. Shalev, Generation and random generation: from simple
groups to maximal subgroups, preprint, available at http://eprints.soton.ac.uk/151795.
12. R.W. Carter, Finite Groups of Lie Type: Conjugacy Classes and Complex Characters. Pure and
Applied Mathematics (Wiley-Interscience, New York, 1985)
13. M.D.E. Conder, Generators for alternating and symmetric groups. J. Lond. Math. Soc. 22, 75–
86 (1980)
14. M.D.E. Conder, Hurwitz groups: a brief survey. Bull. Am. Math. Soc. 23, 359–370 (1990)
15. F. Dalla Volta, A. Lucchini, Generation of almost simple groups. J. Algebra 178, 194–223
(1995)
16. J.D. Dixon, The probability of generating the symmetric group. Math. Z. 110, 199–205 (1969)
17. J.D. Dixon, Asymptotics of generating the symmetric and alternating groups. Electron. J.
Comb. 12, 1–5 (2005)
18. J.D. Dixon, B. Mortimer, Permutation Groups. Graduate Texts in Mathematics, vol. 163
(Springer, New York, 1996)
19. E.B. Dynkin, Semisimple subalgebras of semisimple Lie algebras. Trans. Am. Math. Soc. 6,
111–244 (1957)
20. P. Erdös, P. Turán, On some problems of a statistical group-theory II. Acta Math. Acad. Sci.
Hung. 18, 151–163 (1967)
21. B. Everitt, Alternating quotients of Fuchsian groups. J. Algebra 223, 457–476 (2000)
22. R.K. Fisher, The number of non-solvable sections in linear groups. J. Lond. Math. Soc. 9,
80–86 (1974)
23. N. Gill, I. Short, L. Pyber, E. Szabó, On the product decomposition conjecture for finite simple
groups, available at arXiv:1111.3497
24. R.M. Guralnick, W.M. Kantor, Probabilistic generation of finite simple groups. J. Algebra 234,
743–792 (2000)
25. R.M. Guralnick, M. Larsen, P.H. Tiep, Representation growth in positive characteristic and
conjugacy classes of maximal subgroups. Duke Math. J. 161, 107–137 (2012)
26. R.M. Guralnick, G. Malle, Products of conjugacy classes and fixed point spaces. J. Am. Math.
Soc. 25, 77–121 (2012)
27. R.M. Guralnick, M.W. Liebeck, J. Saxl, A. Shalev, Random generation of finite simple groups.
J. Algebra 219, 345–355 (1999)
28. P. Hall, The Eulerian functions of a group. Q. J. Math. 7, 134–151 (1936)
29. J. Häsä, Growth of cross-characteristic representations of finite quasisimple groups of Lie type,
preprint, Imperial College London (2012)
30. H.A. Helfgott, Growth and generation in SL2 .Z=pZ/. Ann. Math. 167, 601–623 (2008)
31. H.A. Helfgott, Growth in SL3 .Z=pZ/. J. Eur. Math. Soc. 13, 761–851 (2011)
32. H.A. Helfgott, A. Seress, On the diameter of permutation groups, Ann. Math. (to appear),
available at arXiv:1109.3550
33. B. Huppert, Endliche Gruppen I. Die Grundlehren der Mathematischen Wissenschaften, Band
134 (Springer, Berlin, 1967)
34. A. Jaikin-Zapirain, L. Pyber, Random generation of finite and profinite groups and group
enumeration. Ann. Math. 173, 769–814 (2011)
35. S. Jambor, M.W. Liebeck, E.A. O’Brien, Some word maps that are non-surjective on
infinitely many finite simple groups, Bulletin London Math. Soc. (to appear), available at
arXiv:1205.1952
1 Probabilistic and Asymptotic Aspects of Finite Simple Groups 33

36. G.A. Jones, Varieties and simple groups. J. Austral. Math. Soc. 17, 163–173 (1974)
37. W.M. Kantor, A. Lubotzky, The probability of generating a finite classical group. Geom.
Dedicata 36, 67–87 (1990)
38. M. Kassabov, T.R. Riley, Diameters of Cayley graphs of Chevalley groups. Eur. J. Comb. 28,
791–800 (2007)
39. N. Katz, Rigid Local Systems (Princeton University Press, Princeton, 1996)
40. P. Kleidman, M.W. Liebeck, The Subgroup Structure of the Finite Classical Groups. London
Mathematical Society Lecture Note Series, vol. 129 (Cambridge University Press, Cambridge,
1990)
41. S. Lang, A. Weil, Number of points of varieties over finite fields. Am. J. Math. 76, 819–827
(1954)
42. M. Larsen, A. Shalev, Characters of symmetric groups: sharp bounds and applications. Invent.
Math. 174, 645–687 (2008)
43. M. Larsen, A. Lubotzky, C. Marion, Deformation theory and finite simple quotients of triangle
groups, preprint, Hebrew University, 2012
44. M. Larsen, A. Shalev, P.H. Tiep, Waring problem for finite simple groups. Ann. Math. 174,
1885–1950 (2011)
45. R. Lawther, Elements of specified order in simple algebraic groups. Trans. Am. Math. Soc.
357, 221–245 (2005)
46. M.W. Liebeck, G.M. Seitz, Maximal subgroups of exceptional groups of Lie type, finite and
algebraic. Geom. Dedicata 36, 353–387 (1990)
47. M.W. Liebeck, G.M. Seitz, On the subgroup structure of exceptional groups of Lie type. Trans.
Am. Math. Soc. 350, 3409–3482 (1998)
48. M.W. Liebeck, G.M. Seitz, On finite subgroups of exceptional algebraic groups. J. Reine
Angew. Math. 515, 25–72 (1999)
49. M.W. Liebeck, G.M. Seitz, The maximal subgroups of positive dimension in exceptional
algebraic groups. Mem. Am. Math. Soc. 169(802), 1–227 (2004)
50. M.W. Liebeck, A. Shalev, The probability of generating a finite simple group. Geom. Dedicata
56, 103–113 (1995)
51. M.W. Liebeck, A. Shalev, Classical groups, probabilistic methods, and the .2; 3/-generation
problem. Ann. Math. 144, 77–125 (1996)
52. M.W. Liebeck, A. Shalev, Maximal subgroups of symmetric groups. J. Comb. Theor. Ser. A
75, 341–352 (1996)
53. M.W. Liebeck, A. Shalev, Simple groups, probabilistic methods, and a conjecture of Kantor
and Lubotzky. J. Algebra 184, 31–57 (1996)
54. M.W. Liebeck, A. Shalev, Diameters of simple groups: sharp bounds and applications. Ann.
Math. 154, 383–406 (2001)
55. M.W. Liebeck, A. Shalev, Fuchsian groups, coverings of Riemann surfaces, subgroup growth,
random quotients and random walks. J. Algebra 276, 552–601 (2004)
56. M.W. Liebeck, A. Shalev, Character degrees and random walks in finite groups of Lie type.
Proc. Lond. Math. Soc. 90, 61–86 (2005)
57. M.W. Liebeck, A. Shalev, Fuchsian groups, finite simple groups and representation varieties.
Invent. Math. 159, 317–367 (2005)
58. M.W. Liebeck, B.M.S. Martin, A. Shalev, On conjugacy classes of maximal subgroups of finite
simple groups, and a related zeta function. Duke Math. J. 128, 541–557 (2005)
59. M.W. Liebeck, L. Pyber, A. Shalev, On a conjecture of G.E. Wall. J. Algebra 317, 184–197
(2007)
60. M.W. Liebeck, N. Nikolov, A. Shalev, A conjecture on product decompositions in simple
groups. Groups Geom. Dyn. 4, 799–812 (2010)
61. M.W. Liebeck, E.A. O’Brien, A. Shalev, P.H. Tiep, The Ore conjecture. J. Eur. Math. Soc. 12,
939–1008 (2010)
62. M.W. Liebeck, A.J. Litterick, C. Marion, A rigid triple of conjugacy classes in G2 . J. Group
Theor. 14, 31–35 (2011)
34 M.W. Liebeck

63. M.W. Liebeck, E.A. O’Brien, A. Shalev, P.H. Tiep, Products of squares in finite simple groups.
Proc. Am. Math. Soc. 140, 21–33 (2012)
64. M.W. Liebeck, N. Nikolov, A. Shalev, Product decompositions in finite simple groups. Bull.
Lond. Math. Soc. (to appear), available at arXiv:1107.1528
65. A. Lubotzky, Discrete Groups, Expanding Graphs and Invariant Measures. Progress in
Mathematics, vol. 125 (Birkhäuser, Basel, 1994)
66. A. Lubotzky, The expected number of random elements to generate a finite group. J. Algebra
257, 452–459 (2002)
67. A. Lubotzky, Expander Graphs in Pure and Applied Mathematics, available at arXiv:1105.2389
68. A.M. Macbeath, Generators of the linear fractional groups, in Proceedings of the Symposium
on Pure Math., vol. XII (American Mathematical Society, Providence, Rhode Island, 1969),
pp. 14–32
69. A. Mann, Positively finitely generated groups. Forum Math. 8, 429–459 (1996)
70. C. Marion, Triangle groups and PSL2 .q/. J. Group Theor. 12, 689–708 (2009)
71. C. Marion, On triangle generation of finite groups of Lie type. J. Group Theor. 13, 619–648
(2010)
72. C. Marion, Random and deterministic triangle generation of three-dimensional classical groups
I-III. Comm. Algebra (to appear)
73. A. Maroti, M.C. Tamburini, Bounds for the probability of generating the symmetric and
alternating groups. Arch. Math. 96, 115–121 (2011)
74. B.M.S. Martin, Reductive subgroups of reductive groups in nonzero characteristic. J. Algebra
262, 265–286 (2003)
75. L. Di Martino, M.C. Tamburini, A.E. Zalesskii, On Hurwitz groups of low rank. Comm.
Algebra 28, 5383–5404 (2000)
76. N.E. Menezes, M.R. Quick, C.M. Roney-Dougal, The probability of generating a finite simple
group, Israel J. Math. (to appear)
77. G.A. Miller, On the groups generated by two operators. Bull. Am. Math. Soc. 7, 424–426
(1901)
78. E. Netto, The Theory of Substitutions and Its Applications to Algebra, 2nd edn. (Chelsea
Publishing Co., New York, 1964) (first published in 1892)
79. I. Pak, On probability of generating a finite group, preprint, http://www.math.ucla.edu/pak.
80. I. Pak, What do we know about the product replacement algorithm? in Groups and Computa-
tion, III, vol. 8, Columbus, OH, 1999 (Ohio State University Mathematical Research Institute
Publications/de Gruyter, Berlin, 2001), pp. 301–347
81. L. Pyber, E. Szabó, Growth in finite simple groups of Lie type of bounded rank, available at
arXiv:1005.1858
82. L.L. Scott, Matrices and cohomology. Ann. Math. 105, 473–492 (1977)
83. A. Selberg, On the estimation of Fourier coefficients of modular forms. Proc. Symp. Pure Math.
8, 1–15 (1965)
84. R. Steinberg, Generators for simple groups. Can. J. Math. 14, 277–283 (1962)
85. R. Steinberg, Endomorphisms of linear algebraic groups. Mem. Am. Math. Soc. 80, 1–108
(1968)
86. K. Strambach, H. Völklein, On linearly rigid tuples. J. Reine Angew. Math. 510, 57–62 (1999)
87. M.C. Tamburini, M. Vsemirnov, Hurwitz groups and Hurwitz generation, in Handbook of
Algebra, vol. 4 (Elsevier/North-Holland, Amsterdam, 2006), pp. 385–426
88. M.C. Tamburini, J.S. Wilson, N. Gavioli, On the .2; 3/-generation of some classical groups I.
J. Algebra 168, 353–370 (1994)
89. J. Whiston, Maximal independent generating sets of the symmetric group. J. Algebra 232,
255–268 (2000)
Chapter 2
Estimation Problems and Randomised
Group Algorithms

Alice C. Niemeyer, Cheryl E. Praeger, and Ákos Seress

2.1 Estimation and Randomization

2.1.1 Computation with Permutation Groups

In 1973, Charles Sims [89] proved the existence of the Lyons–Sims sporadic
simple group Ly by constructing its action as a group of permutations of a set of
cardinality 8,835,156 on a computer which could not even store and multiply the two
generators of Ly in this smallest degree permutation representation for the group!
The existence of this finite simple group, together with many of its properties, had
been predicted by Richard Lyons [60], but proof of existence was not established
until Sims’ construction. Leading up to this seminal achievement, Sims [88] had
developed concepts and computational methods that laid the foundation for his
general theory of permutation group computation.

A.C. Niemeyer ()

Centre for the Mathematics of Symmetry and Computation, School of Mathematics and Statistics,
The University of Western Australia, 35 Stirling Highway, Crawley, WA 6009, Australia
e-mail: [email protected]
C.E. Praeger
Centre for the Mathematics of Symmetry and Computation, School of Mathematics and Statistics,
The University of Western Australia, 35 Stirling Highway, Crawley, WA 6009, Australia
King Abdulaziz University, Jeddah, Saudi Arabia
e-mail: [email protected]
Á. Seress
Centre for the Mathematics of Symmetry and Computation, School of Mathematics and Statistics,
The University of Western Australia, 35 Stirling Highway, Crawley, WA 6009, Australia
The Ohio State University, Columbus, OH, USA
e-mail: [email protected]

A. Detinko et al. (eds.), Probabilistic Group Theory, Combinatorics, and Computing, 35

Lecture Notes in Mathematics 2070, DOI 10.1007/978-1-4471-4814-2 2,
© Springer-Verlag London 2013
36 A.C. Niemeyer et al.

Sims introduced the critical concept of a base of a permutation group G on a

finite set ˝, namely a sequence of points ˛1 ; : : : ; ˛b of ˝ such that only the identity
of G fixes all of them. For example, the dihedral group D2n D ha; bi acting on
f1; 2; : : : ; ng, where a D .1; 2; : : : ; n/ and b D .2; n/.3; n  1/ : : : , has a base
B D .1; 2/, since only the identity of D2n fixes both 1 and 2. Moreover the 2n
elements g 2 D2n produce 2n distinct image pairs .1g ; 2g / of the base B—for
example, a maps B to .2; 3/, b maps B to .1; n/.
Sims observed that elements of a permutation group G could always be
represented uniquely by the sequence of images of the points of a given base B.
He exploited this potentially compact representation of group elements, ingeniously
showing how to compute in G with these base images, via a so-called strong
generating set of G relative to B. Sims’ algorithm to construct a base and strong
generating set, called the Schreier–Sims algorithm, is of fundamental importance
for permutation group computation.
For groups possessing a small base, the Schreier–Sims algorithm is extremely
efficient, but for some groups every base has size close to the cardinality n D j˝j
of the point set. For such groups, the methods are not effective. Examples of such
large-base groups include the “giants”: the alternating group Alt.˝/ D An and the
symmetric group Sym.˝/ D Sn , which have minimum-sized bases .1; 2; : : : ; n2/
and .1; 2; : : : ; n  1/ respectively.

2.1.2 Recognising the Permutation Group Giants

For computational purposes, a finite permutation group G on ˝ is given by a

(usually small) set X of generators. The group G consists of all products of arbitrary
length of elements from X . Since the Schreier–Sims algorithm is ineffective for
computation with the giants Alt.˝/ and Sym.˝/, it is important to determine in
advance (that is, before trying to find a base and strong generating set) whether or
not a given permutation group G D hX i is one of these giants. Thus the question
of identifying the giants Alt.˝/ and Sym.˝/, given only a generating set of
permutations, was a central issue in the development of general purpose group
theory computer systems.
Theoretically the problem of detecting these giants had engaged mathematicians
from the earliest studies of group theory. Since the seminal work of Camille Jordan
in the 1870s, it has been known that there are many kinds of permutations such that
the only transitive permutation groups containing them are the giants (we say that
G  Sym.˝/ is transitive if each pair of points of ˝ can be mapped one to the
other by an element of G). The most beautiful of these results that identifies a large
family of such elements is Jordan’s theorem below.
Let us call a permutation g 2 Sn a Jordan element if g contains a p-cycle, for
some prime p with n=2 < p < n  2. For example, g D .1; 2; 3; 4; 5/.6; 7/ 2 S9 is
a Jordan element (with n D 9; p D 5).
2 Estimation and Group Algorithms 37

Theorem 2.1. If a transitive permutation group G  Sn contains a Jordan element

then G is An or Sn .
Given a set of generators for G  Sym.˝/, it is easy to test whether G is
transitive. Hence, recognising the giants boils down to the question: how prevalent
are the Jordan elements in the giants? For a fixed prime p 2 .n=2; n2/, the number
of elements in Sn containing a p-cycle is
!
n nŠ nŠ
.p  1/Š.n  p/Š D .and in An /;
p p 2p

so the proportion of Jordan elements in An or Sn for thisP prime p is 1=p, and

therefore the proportion of Jordan elements in An or Sn is n=2<p<n2 p1  logc n
for some constant c. For n  100, c can be taken to be 1=5, which follows by
applying an inequality by Dusart [25, p. 414] to determine the number of primes p
with n=2 < p < n  2. So roughly c out of every log n independent, uniformly
distributed random elements from Sn or An will be Jordan. That is to say, we should
find a Jordan element with high probability by randomly selecting elements in
a giant.

2.1.3 Monte Carlo Algorithms

How do we turn the comments above into a justifiable algorithm? We want to make
some multiple of log n random selections from a transitive group G on n points
which we suspect may be Sn or An , but as yet we have no proof of this fact. We hope,
and expect, to find a Jordan element, thereby uncovering the secret and proving that
G really is a giant Sn or An .
Formally, we model this process as a Monte Carlo algorithm. The Monte Carlo
method was invented by Stanislaw Ulam in the 1940s; it was named after Monte
Carlo Casino in Monaco which Ulam’s uncle visited often (see the account in [62]).
The characteristics of a Monte Carlo algorithm are that it completes quickly, but
allows a small (user-controlled) probability of “error”, that is, of returning an
incorrect result. In our context, for a Monte Carlo algorithm, we begin with a
prescribed bound on the error probability " 2 .0; 1/. The algorithm typically makes
a number N D N."/ of random selections, depending on ", this number being
determined in advance to guarantee that the probability of an incorrect result is at
most ".
Here is a worked example of a Monte Carlo algorithm to recognise the giants Sn
and An among transitive permutation groups on n points.
38 A.C. Niemeyer et al.

Algorithm 1: JORDAN
Input: A transitive subgroup G D hx1 ; : : : ; xk i  Sn and a real number " 2 .0; 1/ (the error
probability bound);
Output: true or false;
# We hope the algorithm returns true if G is Sn or An – see the comments below;
for up to N D d.log "1 /.log n/=ce random elements g from G do
if g is a Jordan element then
return true;
end
end
Return false;

Comments on the algorithm

1. The procedure completes after at most N repeats of the if statement, so it is an
algorithm! If it returns true then G D An or Sn by Jordan’s Theorem 2.1. On the
other hand, if the algorithm returns false then this may be incorrect, but only if
G does equal An or Sn , and we failed to find a Jordan element.
2. We have
Prob(we do not find a Jordan element, given that G D An or G D Sn ) 
 N
1  logc n < ".
So Algorithm 1 is a Monte Carlo algorithm with error probability less than ".
This is a special kind of Monte Carlo algorithm: the result true is always correct,
and the possibility of an incorrect result is confined to the case where false is
returned. Such algorithms are called one-sided Monte Carlo algorithms.
3. This probability estimate assumes that the random selections made are indepen-
dent and uniformly distributed. There are algorithms available for producing “ap-
proximately random” elements from a group given by a generating set; see [3,18,
24]. We shall not discuss the theoretical details of these algorithms or their prac-
tical performance. Rather we assume in our discussion of randomised algorithms
that we are dealing with independent uniformly distributed random elements.
4. The design and discussion of this simple algorithm used concepts and results
from group theory to prove correctness, and from number theory to establish the
bound on the error probability. It is typical to gather and develop methods from
a variety of mathematical areas to achieve good algorithm design and analysis.
5. Algorithm 1 is essentially the algorithm used in GAP [37] and MAGMA [15]
for testing if a permutation group is a giant. It was first described by Parker and
Nikolai [79], preceding Sims’ work by a decade. The second author (Praeger)
recalls numerous discussions with John Cannon, over a number of years, about
the implementation of this algorithm in connection with his development of the
computer algebra system CAYLEY (a precursor to MAGMA). There was much
to learn about improving the practical performance of the algorithm to avoid its
becoming a bottle-neck for permutation group computation. A wider class of
“good elements” than the Jordan elements was used, based on generalisations of
2 Estimation and Group Algorithms 39

Jordan’s Theorem (see [94, 13.10] and [83]), and better methods were developed
to produce “approximately random” elements.

2.1.4 What Kinds of Estimates and in What Groups?

Notice the role estimates played in Algorithm 1:

a lower bound for the proportion of Jordan elements gives an upper bound on
the error probability.
Does it matter if the estimate is far from the true value? We might, for different
reasons, propose one of two different answers to this question:
1. We might say “no”, because if there are more Jordan elements than our estimates
predict, then we simply find one more quickly and the algorithm confirms that
“G is a giant” more efficiently.
2. We might say “yes”, because if G is not a giant then we force the algorithm to
do needless work in testing too large a number of random elements so that the
algorithm runs more slowly than necessary on non-giants. Note that the algorithm
will never find a Jordan element in a non-giant by Theorem 2.1, so the full quota
of random elements will be tested before completion.
For general purpose algorithms such as Algorithm 1, which are used frequently
on arbitrary permutation groups, the quality of the estimates really does matter. We
should try to make estimates as good as possible, especially when they are used to
analyze a time-critical module of a randomised algorithm.
In the computer algebra systems GAP and MAGMA, new algorithms are under
development for computation with matrix groups and permutation groups. These
employ a tree-like data structure which allows a “divide and conquer” approach,
reducing to computations in normal subgroups and quotient groups. This approach
(see Sect. 2.4.1) reduces many computational problems to the case of finite simple
groups. Accordingly many of the topics chosen in this chapter are of relevance to
computing with finite simple groups.

2.1.5 What Group is That: Recognising Classical Groups

as Matrix Groups

As a more substantial example for group recognition, we describe an algorithm to

recognise a finite classical group in its natural representation. By this, we mean
that the algorithm will return the “name” of the group. We give a broad-brush
description of the classical recognition algorithm developed in [72] generalising the
Neumann–Praeger SL-recognition algorithm in [69].
The algorithm takes as input a subgroup G of a finite n-dimensional classical
group Class.n; q/ over a finite field Fq of order q, such as the general linear group
GL.n; q/ or a symplectic group Sp.n; q/, in its natural representation as a group of
40 A.C. Niemeyer et al.

matrices acting on the underlying vector space V .n; q/. The subgroup G is given
by a generating set of n  n matrices over Fq . The algorithm seeks so-called ppd
elements in G which we describe as follows.
For an integer e > 1, a primitive prime divisor (ppd) of q e  1 is a prime r
dividing q e  1 such that r does not divide q i  1 for any i < e. It has been known
for a long time that primitive prime divisors exist unless q D 2, e D 6, or e D 2
and q C 1 is a power of 2; see [97]. Superficially, primitive prime divisors seem
interesting because the order of the classical group has the form
Y
j Class.n; q/j D q some power .q i ˙ 1/:
various i

We define a ppd-.qI e/ element g 2 Class.n; q/ as an element with order divisible

by a ppd of q e  1. The algorithm in [72] seeks two ppd elements, namely a
ppd-.qI e1/ and a ppd-.qI e2/ element for e1 6D e2 and e1 ; e2 > n=2, which
satisfy various additional minor conditions described in [72, Sects. 2 and 9]. We
call such a pair a good ppd matrix pair. Their importance lies in the following deep
theorem [72, Theorem 4.8], the proof of which relies heavily on the finite simple
group classification.
Theorem 2.2. If G  Class.n; q/ is irreducible on V .n; q/ and G contains a good
ppd matrix pair, then (essentially) G D Class.n; q/ or G is known explicitly.
Thus, provided that (a) we can test efficiently whether G is irreducible on
V .n; q/, (b) good ppd matrix pairs are sufficiently prevalent in Class.n; q/ and are
easily identifiable, and (c) the exceptions in Theorem 2.2 are easy to deal with, the
good ppd matrix pairs could play the role of the Jordan elements used to identify
the permutation group giants in Algorithm 1. We would then have an analogue of
Algorithm 1 for classical groups, underpinned by considerably deeper theory than
Jordan’s Theorem 2.1. It would look like this:

Algorithm 2: RECOGNISECLASSICAL
Input: An irreducible subgroup G D hX1 ; : : : ; Xk i  Class.n; q/ and a real number
" 2 .0; 1/ (the error probability bound).
Output: true or false
# If the output is true, we are certain that G D Class.n; q/;
# the output false may be incorrect;
for Many(depending on n; q; ") random elements g 2 G do
determine if g is a ppd element with the additional properties;
if a good ppd matrix pair is found then
if G is one of the exceptions then
return false
else
return true;
end
end
end
return false;
2 Estimation and Group Algorithms 41

Comments on the algorithm

1. Note that if Algorithm 2 returns true then G really is Class.n; q/ by Theorem 2.2;
while if it returns false then the result may be incorrect (namely if G D
Class.n; q/ and we fail to find the good ppd matrix pair).
2. The missing ingredient is our knowledge of the presence of good ppd matrix
pairs in Class.n; q/, and an estimate of their proportion. We need a positive lower
bound on their proportion in order to decide how Many random elements to test
to ensure an error probability of at most ". This is necessary to prove that we
have a one-sided Monte Carlo algorithm.
Estimating the proportion of ppd-.qI e/ elements in Class.n; q/: For the
details involved in dealing with the additional properties we refer the reader to [72].
For G D Class.n; q/ and e > n=2, let ppd.G; e/ be the proportion of ppd-.qI e/
elements in G. We give a few details for the general linear case.
n 1
Lemma 2.3. Let G D GL.n; q/ and let 2 < e  n. Then eC1  ppd.G; e/  1e .
Proof. Let g 2 G be a ppd-.qI e/ element and let r be a ppd of q e  1 dividing
jgj. By considering a power of g of order r, we can show that g leaves invariant a
unique e-dimensional subspace U of V .n; q/, and acts irreducibly on U . Moreover
the induced element gjU is a ppd-.qI e/ element in GL.U /, and a straightforward
counting argument (see [72, Lemma 5.1]) shows that ppd.G; e/ D ppd.GL.U /; e/.
In other words, we may assume that n D e in the proof. With this assumption,
we have g irreducible on V .n; q/, and each such element lies in a Singer cycle
S D Zq e 1 of G. All Singer cycles are conjugate in G, and distinct Singer cycles
contain disjoint sets of irreducible elements. Moreover the number of Singer cycles
is jG W NG .S /j D jGj=.e.q e  1// (see [69, Lemma 2.1]). Hence ppd.G; e/ is equal
to .1=e/  (the proportion of such elements in the cyclic group S ).
This immediately gives ppd.G; e/  1=e. We need one more observation to
obtain the lower bound. Certainly each element of S of order not divisible by r
lies in the unique subgroup S0 of S of index r. Thus each element of S n S0 has
order divisible by r, and hence ppd.G; e/  .1=e/  .1  1=r/. Now e is the least
positive integer such that q e  1 .mod r/, and so q has order e modulo the prime r.
This implies that e divides r  1, and in particular r  e C 1. Hence ppd.G; e/ 
.1=e/  e=.e C 1/ D 1=.e C 1/. t
u
A similar argument in [72, Theorem 5.7] shows that the bounds of Lemma 2.3
hold for the other classical groups for almost all values of e. Since each ppd element
corresponds to just one e-value (because e > n=2), we can find a lower bound for
the proportion of ppd elements in G P by adding the lower bounds for ppd.G; e/ over
all relevant e. For GL.n; q/, this is n=2<en 1=e log 2 by Lemma 2.3. For the
other classical groups, the values of e occurring all have the same parity (odd for
unitary groups and even for symplectic and orthogonal groups), and for these groups
the proportion of ppd elements is roughly .log 2/=2 [72, Theorem 6.1].
These lower bounds (or rather, the equivalent ones we obtain in [72] after taking
into account the additional conditions on the ppd elements) allow us to decide
42 A.C. Niemeyer et al.

how many random selections to make in order to find a good ppd matrix pair
with probability at least 1  1=", and hence to determine the value for Many in
Algorithm 2.

2.1.6 What Group is That: Recognising Lie Type Groups

in Arbitrary Representations

Of course, we do not only encounter the classical groups in their natural represen-
tation. If G is a simple group of Lie type, given in any permutation or matrix group
representation, and the characteristic p of G is known, then we may proceed by an
extension of Algorithm 2. The procedure that we sketch was developed in [6].
Let e1 and e2 be the two highest ppd exponents, that is, integers e such that
G contains elements of order divisible by a primitive prime divisor of p e  1. It
was shown in [6] that for each pair of integers .e1 ; e2 /, there are at most seven
isomorphism types of Lie type groups of characteristic p with e1 ; e2 as the highest
ppd exponents in the group. Also, ppd elements with ppd exponents e1 and e2 are
frequent enough that we encounter them in a random sample of size polynomial in
the input length.
To distinguish between the possibilities for G with the same values e1 and e2 , we
consider the third highest ppd exponent in G and elements whose order is divisible
by a product of two ppd primes, corresponding to certain chosen ppd exponents. The
result is a polynomial-time Monte Carlo algorithm that names the isomorphism type
of G, with one exception: a polynomial-size random sample may not distinguish the
groups Sp.2m; p f / and O.2m C 1; p f /, for odd primes p. This last ambiguity was
handled by Altseimer and Borovik [1].

2.2 Proportions of Elements in Symmetric Groups

2.2.1 Notation

In this section we fix a set ˝ and consider the symmetric group Sym.˝/ on
˝. When ˝ D f1; : : : ; ng for some positive integer n we write Sn instead of
Sym.f1; : : : ; ng/. Elements of Sn are written in disjoint cycle notation. The number
of cycles of a given element g 2 Sn denotes the number of disjoint cycles g has on
f1; : : : ; ng including fixed points.

2.2.2 Historical Notes

The study of proportions of permutations has been of interest to mathematicians

for a long time. For example, in 1708 Monmort introduced and analyzed a game
2 Estimation and Group Algorithms 43

of 13 cards which he called “jeu de Treize” (the game of thirteen) in his book on
the theory of games [64, pp. 54–64]. He later generalised the game to any number
of cards numbered from 1 to n [65, pp. 130–143]. In the game, a player has n
turns, each time announcing out loud the number of the turn and picking a card at
random from the deck of n cards without replacing it. The game is won if each time
the number of the card and the number announced are different. Leonhard Euler in
Solutio Quaestionis curiosae ex doctrina combinationum [34] describes the game as
follows: Data serie quotcunque litterarum a; b; c; d; e etc., quarum numerus sit n,
invenire quot modis earum ordo immutari possit, ut nullo in eo loco reperiatur, quem
initio occupaverat. This can be translated as Given an arbitrary series (sequence)
of letters a; b; c; d; e; : : :, let the number of which be n, find in how many ways their
order may be changed so that none reappears in the same place which it originally
occupied.1 In [33] Euler showed that the number of solutions is the integer closest
to nŠ=e. Earlier solutions had already been given; for example, Monmort presented
a solution by Nicolas Bernoulli [65, pp. 301–302]. De Moivre also mentions the
game already in the first edition of [23], and gives a solution in [23, Problem 35].
Today this problem is often called the hat-swapping problem: Suppose n men
each put a hat on a hat rack in a restaurant. When they leave they each choose a
random hat. What is the probability that no man chooses his own hat?
Nowadays we call a permutation in Sn which has no fixed points on f1; : : : ; ng a
derangement, and we would rephrase the game of thirteen, Euler’s question or the
hat-swapping problem as: How many derangements are there in Sn ?
In this section we will focus on certain other proportions of elements in Sn .
The proportions that we focus on arise either from algorithmic applications for
permutation groups or from applications to classical groups of Lie type (see
Sect. 2.3.2).

2.2.3 Orders of Permutations

The order of a permutation can easily be read off from its disjoint cycle notation;
namely it is the least common multiple of the cycle lengths. One of the oldest results
on the order of an element in a symmetric group is due to Landau, who determined
how large the order of an element in Sn can be asymptotically.
Theorem 2.4 (Landau [51]).
 
log maxg2Sn .ord.g//
lim p D 1:
n!1 n log.n/

1
Translation by Peter M. Neumann, The Queen’s College, University of Oxford.
44 A.C. Niemeyer et al.

Although the order of an element in Sn can be as large as the previous theorem

suggests, Erdős and Turán were able to prove, in the first of a series of papers [27–
32] on the subject of the statistics of permutations, that most elements have much
smaller order.
Theorem 2.5 (Erdős and Turán [27]). For "; ı > 0 there is a number N0 ."; ı/
such that for all n  N0 ."; ı/,

jfg 2 Sn j .1=2  "/ log2 .n/  log.ord.g//  .1=2 C "/ log2 .n/gj
 1  ı:
nŠ
Erdős and Turán proved many more insightful results on the order of elements
in symmetric groups. For example, in [28] they investigated prime divisors of
the order of elements in symmetric groups. In [29] they described for any x the
limiting behaviour as n tends to infinity of the proportion of elements g in Sn for
which log.ord.g//  12 log2 .n/ C x log3=2 .n/. In [30] they considered, among other
problems, the number of different values that ord.g/ can have as g ranges over the
elements of Sn .
Goh and Schmutz [39] prove that the logarithm of qthe average order of a random
p R1  
permutation in Sn is c n= log.n/, where c D 2 2 0 log log 1ee t dt. This
constant is approximately 2:99.

2.2.4 Number of Cycles

Let a.n/ denote the average number of cycles of the elements in Sn . In a seminal
paper [40], Gončarov examined various properties of random permutations. Among
many other results, he proved that the average number of cycles of a permutation in
Sn is close to log.n/.
Theorem 2.6 (Gončarov [40]).

X
n
1
a.n/ D D log.n/ C C o.1/
i D1
i

for n ! 1.
Plesken and Robertz [82] generalised these results to An and to wreath products
of groups with imprimitive action.

2.2.5 Generating Functions

One very powerful method of obtaining information about certain combinatorial

quantities is to employ generating functions.
2 Estimation and Group Algorithms 45

Given a sequence .an /n2N of real numbers, the Ordinary Generating Function
for an is X
A.z/ WD an zn :
n0

For example, an could be the number of certain elements in Sn .

A very intuitive way to view generating functions is given in the following quote
from Wilf’s aptly named book generatingfunctionology [96]: “A generating function
is a clothesline on which we hang up a sequence of numbers for display.” Here
we just highlight some of the ways in which generating functions can shed light
on some of our problems. To understand the power and beauty of the subject of
generating functions we refer the reader to both Wilf’s book [96] and a recent book
on analytic combinatorics by Flajolet and Sedgewick [35]. Both books also contain
various interesting results on proportions of permutations.
Several types of generating functions can be defined, and the type of generating
function chosen to attack a particular problem depends on the circumstances. In our
situation exponential generating functions are of particular interest. They are of the
form X an
A.z/ WD zn
n0
nŠ

and ensure that the coefficients anŠn of zn are manageable in situations where an
is expected to grow almost as fast as nŠ. For example, if an is the number of
elements with a particular property in Sn ; then this number could grow rapidly and
using an ordinary generating function would quickly produce unwieldy coefficients.
However, dividing by the order of the group Sn ensures that the coefficients an =nŠ
are proportions of elements in Sn and thus all less than 1.
We study generating functions as elements of the ring of formal power series.
Analytic questions, convergence etc. do not concern us just yet. Generating func-
tions can be manipulated in various ways, and this theory is described in the above
mentioned books. Here we just state, as an example, how two generating functions
can be multiplied:
! ! !
X
1 X
1 X
1 X
n
n n
an z  bn z D ak bnk zn :
nD0 nD0 nD0 kD0

The usefulness of taking a generating function approach in our situation can

be highlighted with the following example. A further example, that estimates the
proportion of regular semisimple elements in general linear groups, is given in
Sect. 2.3.6.

2.2.5.1 Example

Let b  1 be a fixed integer and let an denote the number of permutations in Sn all
of whose cycles have length at most b.
46 A.C. Niemeyer et al.

We would like to study the exponential generating function describing the

numbers an . So let X an
A.z/ WD zn :
n0
nŠ

One very effective way of studying a generating function is to start from a

recursive equation for the coefficients an , and we employ this method here. Our
first task is to find a suitable recursion for an . Recall that we write permutations in
disjoint cycle notation. We are interested in finding an expression for the number an
of permutations in Sn all of whose cycles have length at most b in terms of am for
integers m smaller than n. We employ a combinatorial trick which has been used
e.g. in Beals et al. [9, Theorem 3.7].
We enumerate the permutations in Sn all of whose cycles have length at most
b according
  to the length d of the cycle containing the point 1. For a fixed d , we
have dn11
choices for the remaining points of the cycle of length d containing 1.
On these d points we can put any one of .d  1/Š different cycles and we have
and choices for the permutation on the remaining n  d points. Thus we obtain the
recursion
1 X
minfb;ng
an and
D :
nŠ n .n  d /Š
d D1

Note in particular that an D nŠ for n  b, which is in agreement with this recursion.

The recursion implies that
0 1
X1
an n X1
1@ X
minfb;ng
and A n
A.z/ WD z D 1C z
nŠ n .n  d /Š
nD0 nD1 d D1

X
b X
1
1 and n X
b X1
1 an nCd
D 1C z D1C z :
n .n  d /Š nD0
n C d nŠ
d D1 nDd d D1

A very useful trick when working with generating functions is to take the
derivative. This yields in our case

X
b X
1
an X
b X
1
an X
b
A0 .z/ D znCd 1 D zd 1 zn D zd 1 A.z/:
nŠ nŠ
d D1 nD0 d D1 nD0 d D1

Thus
A0 .z/ X b
D zd 1
A.z/
d D1

and so
X
b
zd
log.A.z// D :
d
d D1
2 Estimation and Group Algorithms 47

Therefore we see that our generating function is

X
b
zd
A.z/ D exp. /:
d
d D1

While this has yielded a very succinct way of describing the number of elements
of interest, it does not as yet yield the desired upper and lower bounds for the
proportion of such elements. Thus we would like to know whether generating
functions can tell us about the limiting behaviour of the coefficients.
An elaborate theory of the asymptotic behaviour of the coefficients of the
generating functions exists. We mention here briefly a subject called “Saddlepoint
Analysis”. The theory is described in the above mentioned books (see also the
papers by Moser and Wyman [67, 68] and Bender [11]). We quote here one result
from Flajolet and Sedgewick’s book, which helps in the situation of our example.
The quoted result is based on a more general theorem by W.K. Hayman [43] (see
also Theorem VIII.4 of [35]). In line with the literature, we denote the coefficient
of zn in the generating function A.z/ by Œzn A.z/. The operator z ddz is defined by
z ddz W P .z/ 7! zP 0 .z/.
P
Theorem 2.7 (see Corollary VIII.2 of [35]). Let P .z/ D nj D1 aj zj have non-
negative coefficients and suppose gcd.fj j aj 6D 0g/ D 1. Let F .z/ D exp.P .z//.
Then
1 exp.P .r//
Œzn F .z/ p ;
2 rn
 2
where r is defined by rP 0 .r/ D n and  D z ddz P .r/.

2.2.5.2 Example of Saddlepoint Analysis

P d
Recall that A.z/ D exp. bd D1 zd / is the exponential generating function for the
number of elements all of whose cycles have length at most b.
P d
Let P .z/ D bd D1 zd . Then P .z/ is a polynomial in z with non-negative
coefficients and satisfies gcd.fd j coefficient of zd is nonzerog/ D 1. The first step
in applying Saddlepoint Analysis P is to estimate rPdetermined by the equationpn D
rP 0 .r/. We have n D rP 0 .r/ D r bd D1 r d 1 D bd D1 r d  r b ; and so r  b n.
 r 2 P
The next step is to estimate , where  D r dr P .r/ D r bd D1 dr d 1 D
Pb d
Pb d
d D1 dr  b d D1 r D bn.
P d P
Hence we have r  n1=b ,   bn and P .r/ D bd D1 rd  b1 bd D1 r d D nb ;
implying
1 exp.P .r// 1  e n=b
Œzn A.z/ p  p :
2 rn 2bn n
48 A.C. Niemeyer et al.

2.2.6 Solutions to x m D 1 in Symmetric and Alternating

Groups

The number of solutions to an equation of the form x m D 1 for a fixed integer m

in symmetric and alternating groups of degree n has received quite a lot of attention
in the literature. More recently, interest in such equations has been rekindled due
to algorithmic applications. In particular, it has also been important for algorithmic
applications to find the asymptotic behaviour of the number of solutions of equations
of the form x m D 1 where m is allowed to grow with n.
We begin by outlining some of the results in the literature. For m fixed let

1
c.n; m/ D jfg 2 Sn j g m D 1gj:
nŠ
Let
X
1
Cm .z/ D c.n; m/zn
nD0

be the corresponding generating function.

Theorem 2.8 (Jacobsthal [47]). For a prime p we have

 zp  X
Œn=p
1
Cp .z/ D exp z C and c.n; m/ D :
p .n  p/ŠŠp 
D1

The number nŠc.n; 2/ of solutions to the equation x 2 D 1 in symmetric groups of

degree n deserves particular attention, since it is also the sum of the degrees of the
irreducible representations of Sn . Chowla et al. [20] examined c.n;p2/ and showed
thatpnc.n; 2/ D c.n1; 2/Cc.n2; 2/. Thus they deduced that 1= n  c.n; 2/ 
1= n C n1 and found the dominant term of the asymptotic expansion for c.n; 2/.
Later, Chowla et al. [21] were able to generalise Jacobsthal’s expansion of Cp .x/
to Cm .x/ where m can be an arbitrary fixed integer, and they asked for an asymptotic
formula for c.n; m/.
Theorem 2.9 (Chowla et al. [21]).
0 1
X zd
Cm .z/ D exp @ A:
d
d jm

Moser and Wyman [66, 67] derived an asymptotic formula in terms of a contour
integral for c.n; 2/ and derived the first order term of the asymptotic value of
c.n; p/. Moreover, they were able to obtain corresponding results for alternating
groups.
2 Estimation and Group Algorithms 49

Theorem 2.10 (Moser and Wyman [66, 67]). For a prime p > 2,

1 1  n n.1 p1 / n p1
c.n; p/ p e :
nŠ p e

Herrera [44] gives the following recursive formula for the number nŠb.n; m/ of
elements in Sn of order m:
X .n  1/Š X
nŠb.n; m/ D b.n  s; t/; where gcd.t; s/ D m:
s
.n  s/Š t

Other authors (e.g. Chernoff [19], Mineev and Pavlov [63], and Pavlov [80])
studied the number of elements in Sn or An satisfying an equation of the form
x m D a for some element a 2 Sn .
In 1986 Volynets [92], Wilf [95] and Pavlov [81] independently determined the
limiting behaviour of c.n; m/ for fixed m, and n tending to infinity. The following
theorem is Wilf’s formulation of the result.
Theorem 2.11. Let m be a fixed positive integer. Define ".n; m/ D 0 if m is odd
and ".n; m/ D 1=.2m2n/ if m is even; and let
0 1
1 @ 1 X
 D 1=m 1 C nd=n C ".n; m/A :
n nm
d jm;d <m

Then for n ! 1 we have

n X 1
c.n; m/ p expf g:
2mn dd
d jm

The above result has been generalised in the literature in various directions and
we shall mention some of these.

2.2.6.1 Families of m

Ben-Ezra [10] generalised these formulae as follows. Let ˘ be a set of primes and
let ˘ 0 denote the set of all primes not in ˘ . Further, let C˘ .z/ denote the generating
function for the proportion c.n; ˘ / of all elements whose order only involves
primes in ˘ , and let C˘ 0 .z/ denote the generating function for the proportion
c.n; ˘ 0 / of all elements whose order only involves primes Q in ˘ 0 . For a finite set
B of integers, define jjBjj D 1 if B D ¿ and jjBjj D b2B b otherwise. Then
50 A.C. Niemeyer et al.

Theorem 2.12 (Ben-Ezra [10]).

Q .1/jBjC1
1. C˘ .z/ D B˘ 0 .1  zjjBjj / jjBjj .
jBj<1
Q .1/jBjC1
2. C˘ 0 .z/ D B˘ .1  zjjBjj / jjBjj .
jBj<1

2.2.6.2 Growing m

The first author to consider an equation of the form x m D 1 in symmetric groups of

degree n in which m is not assumed to be fixed was Warlimont [93], who considers
the case m D n. In particular, he shows that
Theorem 2.13 (Warlimont [93]).
 
1 2c 1 2c 1
C 2  c.n; n/  C 2 C O ;
n n n n n3o.1/

where c D 0 if n is odd and c D 1 if n is even.

In 1990 Erdős and Szalay [26] considered the case where m lies in the range
log.n/=.2 log log.n//  m  n.1=4/" , and derived an asymptotic formula for
c.n; m/.
Volynets [92] proved the following result via the Saddlepoint method.
Theorem 2.14 (Volynets [92]). For primes p, and for positive integers n such that
n and p tend to 1 and p=n ! 0,

1  n n.11=p/ 1=2 X .n1=p /mCkp

1
c.n; p/ D p .1 C o.1//;
nŠ e .m C kp/Š
kD0

where m D n  pŒn=p. In particular, if n1=p =p 2 ! 0 then

1  n n.11=p/ 1=2 n1=p

c.n; p/ D p e .1 C o.1//;
nŠ e

while if n1=p =p ! 0 then

1  n n.11=p/ 1=2 nm=p

c.n; p/ D p .1 C o.1//:
nŠ e mŠ
Finally A.V. Kolchin [49] proved the following theorem using the method of
generalised schemes of allocation (see [50, Chap. 5]).
2 Estimation and Group Algorithms 51

Theorem 2.15 (Kolchin [49]). For d , n positive integers such that d log log.n/=
log.n/ ! 0 and for ı D 0 if d is odd and ı D 1=.2d / when d is even, the following
holds:
8 9
1 nn.11=d / 1 <X nj=d =
c.n; d / D p exp  ı .1 C o.1//:
nŠ en d : j ;
j jd

Another generalisation of the above to the case where the cycle lengths are
elements of particular sets can be found in [90]. Finally we would like to refer
the interested reader to V.F. Kolchin’s book on random graphs [50], which contains
many references and notes to the above mentioned, and other, results on random
permutations.

2.2.7 The Münchausen Method (Bootstrapping)

The previous results highlight how difficult it is to obtain the overall limiting
behaviour for c.n; m/ when m  `n for some constant ` and m is allowed to grow
with n. However, for our algorithmic applications (see Sect. 2.2.8 below), we require
good upper bounds for c.n; m/ in the case where m D r.n  k/ for r 2 f1; 2; 3g
and k  6.
To obtain bounds for c.n; m/ in cases where n1  m  `n for some constant `,
we return to more basic methods and highlight some of the ideas in a proof of the
limiting behaviour of c.n; m/ in such cases.
A popular folk tale tells the story of how Baron Münchausen found himself stuck
in a swamp while riding his horse. He then managed to save himself and his horse
by pulling himself out of the swamp by his own ponytail.
We employ a similar strategy to obtain good estimates for our required propor-
tions. We begin by deriving a first crude estimate and then using this to refine our
estimates. This method (also called bootstrapping) was employed in [9] and later in
[73].
The overall estimate for c.n; m/ is obtained in two steps. The first step yields a
very crude estimate. This in turn is employed in a second step to yield a more refined
estimate. 8
<2 for 360 < m
Define .m/ WD 2:5 for 60 < m  360
:
3:345 for m  60.
A first crude estimate for c.n; m/ is given in the following theorem.
Theorem 2.16. Let m; n 2 N with m  n  1. Then

1 .m/m
c.n; m/  C :
n n2
52 A.C. Niemeyer et al.

Proof-Idea for Crude Estimate

The proof of our first crude estimate relies on a simple idea. It divides the problem
of estimating c.n; m/ into several smaller problems by considering the following
proportions in Sn (see [9]) according to how many cycles the numbers 1, 2 and 3 lie
in. Define proportions
1. c .1/ .n; m/ of those g 2 Sn which have 1; 2; 3 in the same g-cycle.
2. c .2/ .n; m/ of those g 2 Sn which have 1; 2; 3 in two g-cycles.
3. c .3/ .n; m/ of those g 2 Sn which have 1; 2; 3 in three g-cycles.
Then it is clear that

c.n; m/ D c .1/ .n; m/ C c .2/ .n; m/ C c .3/ .n; m/:

For each i with i 2 f1; 2; 3g, we can now hope to use the extra knowledge about
the elements that contribute to the proportion c .i / .n; m/ to obtain a first estimate for
this proportion.
For example, we show how we can obtain an estimate for c .1/ .n; m/. Elements
g 2 Sn contributing to this proportion must contain a cycle C of length d with the
following properties:
1. d j m and 3  d .
2. The cycle C of length d contains 1,2,3.
3. The remaining cycles of g all have lengths dividing m.
Now we can obtain an expression for c .1/ .n; m/ by considering all allowable cycle
lengths d and counting the number of cycles C on d points that contain the points
1, 2 and 3 and ensuring that the remaining n  d points all have lengths dividing m.
As C has to contain 1, 2 and 3, we have n  3 points left to choose the remaining
d  3 points of C ; and having chosen a set of d points (which contains the points 1,
2 and 3), we have .d  1/Š ways of arranging them into different cycles. The number
of permutations on the remaining n  d points all of whose cycle lengths divide m
is c.n  d; m/.n  d /Š. Hence
!
.1/ 1 X n3
c .n; m/ D .d  1/Šc.n  d; m/.n  d /Š
nŠ d 3
d jm;d 3

.n  3/Š X
D .d  1/.d  2/c.n  d; m/:
nŠ
d jm;3d n

As we are currently only interested in obtaining a first crude estimate, we apply a

very rough upper bound on c.n  d; m/, by replacing it with the constant 1. We
therefore find
2 Estimation and Group Algorithms 53

.n  3/Š X
c .1/ .n; m/  .d  1/.d  2/
nŠ
d jm;3d n

.n  3/Š X m m
 .  1/.  2/
nŠ t t
m=nt m=3
Z !
m=3
.n  3/Š m2
 .n  1/.n  2/ C dt
nŠ m=n t2
.n  3/Š
 f.n  1/.n  2/ C mn  3mg
nŠ
1 m
< C 2:
n n

We can employ similar estimates to obtain crude upper bounds for c .2/ .n; m/
and c .3/ .n; m/, which we omit here. Having obtained a first crude estimate, we now
insert this estimate when trying to get a better estimate for c.n; m/.

2.2.7.1 The Pull

Enumerating g by the g-cycle of length d on 1 and recalling that n  1  m yields

1 X
c.n; m/ D c.n  d; m/
n d jm
1d n

1 1 X
 C c.n  d; m/:
m n d jm
1d m=2

For example, in the case where m D n or m D n  1, inserting the crude estimate

for c.n  d; m/ in the equations above we find that

1 1 X  1 .m/m

c.n; m/  C C
m n d jm
nd .n  d /2
1d m=2

1 d.m/.2 C 4 .m//
 C ;
m n2

where d.m/ denotes the number of positive integer divisors of m. The above results
allow us to prove the following strong corollaries.
Corollary 2.17. Let n  19. Let f 2 fn  3; n  2g be odd. Then
1. The conditional probability that a random element g has an n-cycle given that it
satisfies g n D 1 is at least 1=2.
54 A.C. Niemeyer et al.

2. The conditional probability that a random element g has an f -cycle given that
it satisfies g 2f D 1 and jg f j D 2 is at least 1=3.
Finally, we highlight some of the results proved in [75] estimating c.n; m/, where
m D rn for a fixed value of r. The proof of this theorem relies on ideas similar to
those outlined above, combined with an idea of Warlimont’s [93] dividing cycles of
permutations into large and small cycles.
Theorem 2.18. For positive integers r; n with r fixed and n sufficiently large,
 
1 a.r/ 1
c.n; rn/ D C 2 CO 5
n n n 2 o.1/
P
where a.r/ D i;j .1 C i Cj 2r
/, 1  i; j  r 2 , ij D r 2 and r C i; r C j divide rn.
Moreover, the conditional probability that an element  g 2 Sn is an n-cycle, given
a.r/ 1
that its order divides rn, is at least 1  n  O n3=2o.1/ .

2.2.8 Algorithmic Applications of Proportions

in Symmetric Groups

Warlimont’s result is very useful for algorithmic purposes. It tells us that most
permutations g satisfying the equation g n D 1 are n-cycles. Moreover, it also
identifies the cycle structure of the second most abundant set of permutations
satisfying the equation g n D 1; namely permutations which consist of two cycles of
length n=2, and these only occur when n is even. This translates into the algorithm
below to find an n-cycle. Note that the algorithm works in any permutation or matrix
group representation of Sn , where we may not easily recognise the cycle structure
of an element in the natural representation. Such algorithms are called black box
group algorithms; for a formal definition, see Sect. 2.4.2.
Suppose we are given a group G and we believe G might be isomorphic to Sn
under a putative, yet unknown, isomorphism  W G ! Sn . We find an element
g 2 G which would map to an n-cycle under  with high probability by Algorithm 3
below.

Algorithm 3: FINDNCYCLE
Input: G a group, n  19 an integer, 0 < " < 1 real;
Output: g or fail;
# If the output is g, then g n D 1;
for up to n log."1 / random elements g 2 G do
if g n D 1 then
return g;
end
end
Return fail;
2 Estimation and Group Algorithms 55

The algorithm takes as input a real " such that 0 < " < 1, and this input is
used to control the probability of failure. We require that the probability that G is
isomorphic to Sn and the algorithm returns fail to be at most ". Note that on each
random selection, the probability of finding an n-cycle is 1=n. Hence the probability
of failing to find an n-cycle in N."/ random selections is .1  1=n/N."/ and we have
.1  1=n/N."/ < " when N."/  log."1 /=. log.1  1=n//. In particular, this is
the case when N."/  n log."1 /.
Thus the above algorithm returns with probability at least 1" an element g 2 G
satisfying g n D 1. Therefore, if G Š Sn then with probability at least 1=2 this
element is an n-cycle, by the above corollary.
Niemeyer and Praeger [74] generalise Warlimont’s result and consider the case
where m  n, namely rn  m < .r C 1/n for fixed positive integers r.
Algorithm 3 is part of a procedure which decides whether a black box group G is
isomorphic to the full symmetric group Sn for a given natural number n. The full al-
gorithm is described in [9]. First, we have to describe a presentation for the group Sn .
Theorem 2.19 (Coxeter and Moser, 1957).

hr; s j r n D s 2 D .rs/.n1/ D Œs; r j 2 D 1 for 2  j  n=2i

is a presentation for Sn . Moreover, if some group G has generators r; s satisfying

this presentation and r 2 ¤ 1 then G is isomorphic to Sn .
Definition 2.20. The transposition y matches the n-cycle x if y moves two adjacent
points in x.
Lemma 2.21. For n  5, an n-cycle and a matching transposition satisfy the
presentation in Theorem 2.19.
Now we are ready to sketch the algorithm BBRECOGNISESN of [9].

Algorithm 4: BBRECOGNISESN
Input: G D hXi a black box group, n  5;
Output: true and a map  W G ! Sn , or fail;
repeat
1. find r 2 G with r n D 1.
# is .r/ an n-cycle?
2. find h 2 G with h2m D 1 where m 2 fn  2; n  3g odd.
# is .hm / a transposition?
3. find a random conjugate s of hm with Œs; s g  6D 1.
# does .s/ interchange two points of .r/?
until repeated too often;
if r or s not found then return fail;
else
define  by
• .r/ D .1; : : : ; n/ and
• .s/ D .1; 2/.
Return true and  W G ! Sn ;
end
56 A.C. Niemeyer et al.

We test whether hr; si Š Sn via the presentation described in Theorem 2.19.

Theorem 2.22. Given a black box group G isomorphic to Sn ; the probability that
the algorithm BBRECOGNISESN.G; n; "/ returns fail is at most ". The cost of the
algorithm is
O..n C n log.n// log."1 //;
where  is the cost of finding a random element in a black box group and  the cost
of a black box group operation.

2.2.9 Restrictions on Cycle Lengths

An extensive amount of literature exists on the topic of random permutations whose

cycle lengths lie in a given set L or lie in a particular arithmetic progression. Early
work includes that of Touchard [91], Gončarov [40] and Gruder [42].
Let L be a set of natural numbers. Let dL .n/ denote the proportion of elements
in Sn all of whose cycle lengths lie in L and let dL .n; k/ denote the proportion of
elements in Sn with exactly k cycles all of whose lengths lie in L . A generating
function for dL .n/ can be found in [91]. This proportion has been studied by many
authors; we just mention briefly some of Gruder’s results.
Theorem 2.23 (Gruder [42]).

1 X 1
dL .n; k/ D :
kŠ x1    xk
.x1 ;:::;xk /2L k
x1 CCxk Dn

P za P1
Put H.z/ D a2L a and let D.z/ D nD0 dL .n/zn .
Theorem 2.24 (Gruder [42]).
1. D.z/ D exp.H.z//. P Pn 
2. D.z/x D exp.xH.z// D 1
nD0 kD0 dL .n; k/x
k n
z .
Bolker and Gleason [13] obtain an explicit asymptotic formula for dL .n/ when
L is an arithmetic progression.
Let pa .n/ denote the proportion of elements in Sn all of whose cycle lengths are
at least a for some a  2.
Theorem 2.25 (Gruder [42]).
1. limn!1 pa1.n/ D exp.1 C 12 C    C a1
1
/.
  Pn 
2. log lima!1 limn!1 pa1.n/ D ; where D limn!1 i D1
1
i
 log.n/ is the
Euler constant.
V.F. Kolchin summarises many of the asymptotic results known about this case
in his book [50]. We refer the interested reader to [50] and references therein.
2 Estimation and Group Algorithms 57

Finally, we mention one particular proportion that has been of considerable

interest in various applications. For positive integers b, let p:b .n/ denote the
proportion of elements in Sn with no cycle of length divisible by b. This proportion
was first studied for primes b in [28], where Erdős and Turán give an explicit
formula for it. This formula immediately generalises to arbitrary positive integers b.
For a prime b, Erdős and Turán also give the limiting distribution of p:b .n/. Many
other authors have also considered this proportion; for example [12], [14, Sect. 2],
[38]. Here we quote a result from [8, Theorem 2.3(b)].
Theorem 2.26. Let n  b. Then
 1=b  1=b
b .1  n1 / b .1 C n2 /
 p :b .n/  :
n .1  b1 / n .1  b1 /

Ben-Ezra [10] obtained a similar result for b D 2. A formula for the proportion
of elements in An with no cycle of length divisible by b is also given in [8]. Maróti
[61] generalises this, and gives a formula for the proportion of elements of order not
divisible by b in arbitrary permutation groups.
The above estimates have proved to be very useful in deriving proportions of
certain elements in finite classical groups of Lie type. Suppose G is a finite classical
group of Lie type given in natural dimension n with n  2. Using the method
outlined in Sect. 2.3.4, [58] shows that the proportion of elements in G for which
some power is an involution with a large 1-eigenspace of dimension d with n=3 
d  2n=3 is at least c= log.n/ for some constant c.

2.3 Estimation Techniques in Lie Type Groups

We start with a seemingly simple result about permutation groups, discuss the deep
Lie-theoretic analysis underpinning it, and indicate how this approach has led to a
powerful estimation technique for Lie type groups.

2.3.1 p-Singular Elements in Permutation Groups

The following beautiful and surprising result of Isaacs et al. [46] was published in
1995.
Theorem 2.27 (Isaacs, Kantor and Spaltenstein [46]). Let G  Sn and let p be a
prime dividing jGj. Then there is at least 1 chance in n that a uniformly distributed
random permutation in G has order a multiple of p.
This result is about any permutation group—not necessarily primitive, nor even
transitive. It is best possible for such a general result, since if n D p then in the
58 A.C. Niemeyer et al.

affine group AGL.1; p/ there are exactly p  1 elements of order divisible by p out
of a total of p.p  1/ elements in the group.
The only known proof of Theorem 2.27 requires the finite simple group
classification. The proof strategy is first to make an elementary reduction to the
case where G is a nonabelian simple group. Then the simple groups are dealt with.
There are no difficulties with the alternating groups An and the sporadic simple
groups. This leaves the finite simple groups of Lie type to be considered, and this
is where the authors of [46] “wave a magic wand” with a sophisticated argument
from the theory of Lie type groups. We (Niemeyer and Praeger) were at first baffled
by this proof, as well as fascinated by what it achieved, so set about trying to
understand it. Along the way there was help from Klaus Lux and Frank Lübeck.
With Frank Lübeck we made our first full-blown application of the theory in [58] to
estimate the proportion of a certain family of even ordered elements in classical
groups. We discovered that this beautiful theory had been introduced by Gus
Lehrer [54,55] to count various element classes and representation theoretic objects
associated with Lie type groups. Recently Arjeh Cohen and Scott Murray [22]
also used this approach to develop algorithms for computing with finite Lie
algebras.
Our objective became: to formalise the ideas into a framework for estimating
proportions of a wide class of subsets of finite Lie type groups. The framework was
first set out in [58] and in general in [76]. We describe it in the next subsection.

2.3.2 Quokka Subsets of Finite Groups

For a finite group G and a prime p dividing jGj, each group element g can be
written uniquely as a commuting product g D us D su, where u is a p-element and
s is a p 0 -element (that is, ord.u/ is a power of p while ord.s/ is coprime to p). This
is called the Jordan p-decomposition of g.
To find this decomposition write ord.g/ D p a b where p − b and a  0. Then
since p a and b are coprime, there are integers r; t such that rp a C tb D 1. It is
a
straightforward to check that the elements u D gt b and s D g rp have the required
properties, and that u; s are independent of the choices for r; t. This decompo-
sition is critical for defining the kinds of subsets amenable to this approach for
estimation.
Definition 2.28. Let G be a finite group and p a prime dividing jGj. A non-empty
subset Q of G is called a quokka set, or a p-quokka set if we wish to emphasise the
prime p, if the following two properties hold:
(a) Q is closed under conjugation by elements of G.
(b) For g 2 G with Jordan p-decomposition g D us D su, g 2 Q if and only if
s 2 Q.
2 Estimation and Group Algorithms 59

A natural place to find p-quokka sets is in finite Lie type groups in characteristic
p; for example, in G D GL.n; q/ with q a power of p. Here, in a Jordan
p-decomposition g D us D su, the element u is unipotent and s is semisimple. The
elements u; s are called the unipotent part and the semisimple part of g, respectively.
Some of the subsets already discussed in this chapter turn out to be quokka sets. We
give an example.
Example 2.29. Let G D GL.n; q/ or SL.n; q/, with q a power of p, let e be an
integer such that e > n=2, and suppose that q e  1 has a primitive prime divisor.
Then the subset Q of ppd-.n; qI e/ elements of G is a p-quokka set. To see this, note
that Q is closed under conjugation since conjugate elements have the same order.
Also, for a Jordan p-decomposition g D us D su, a ppd r of q e  1 divides ord.g/
if and only if r divides ord.s/.

2.3.3 Estimation Theory for Quokka Sets

The standard reference for the concepts discussed below is Roger Carter’s book [17],
and an account of the required theory is given in [76].
The groups: We start with a connected reductive algebraic group G defined over
the algebraic closure Fq of the finite field Fq of order q, where q is a power
of a prime q0 . A Frobenius morphism F W G ! G defines a finite group of
Lie type G F D fg 2 G j F .g/ D gg as its fixed point subgroup. We use the
following example to illustrate the concepts as they arise. For the algebraic group
q
G D SL.n; Fq / and Frobenius morphism F W .aij / 7! .aij /, the finite group of Lie
type is G D SL.n; q/, since the fixed field of the map a 7! aq is Fq .
F

Maximal tori: A torus in an algebraic group is a subgroup T that is isomorphic

to a direct product of a finite number of copies of the multiplicative group of Fq . In
particular, T is abelian. A torus T is F -stable if F .T / D T , and T is a maximal
torus if T is closed and not properly contained in another torus. All F -stable
maximal tori in G are conjugate. In our example G D SL.n; Fq /, the subgroup T0
of diagonal matrices in G is a maximal torus that is isomorphic to a direct product
of n  1 copies of .Fq / .
The Weyl group: Choose an F -stable maximal torus T0 in G. The Weyl group W
is defined as the quotient NG .T0 /=T0 . Since F -stable maximal tori are conjugate,
the group W is independent of the choice of T0 . In our example G D SL.n; Fq /,
with T0 the subgroup of diagonal matrices, NG .T0 / is the subgroup of monomial
matrices in G, and W D NG .T0 /=T0 is isomorphic to the group of nn permutation
matrices, so W Š Sn .
F -conjugacy: Elements v; w 2 W are said to be F -conjugate if there is an
element x 2 W such that v D x 1 wF .x/. Notice that we abuse notation a little
in this definition, since x 2 W is a coset x D x0 T0 and by F .x/ we mean F .x0 /T0
60 A.C. Niemeyer et al.

(which is well defined since T0 is F -stable). In our example G D SL.n; Fq /,

F -conjugation is ordinary conjugation (since each x 2 W has a representative
monomial matrix with entries 0 or ˙1, and hence x is fixed by F ).
A crucial correspondence and the Quokka Theorem: For an F -stable maximal
torus T of G, the intersection T F D T \ G F D fg 2 T j F .g/ D gg is
called a maximal torus of G F ; although all F -stable maximal tori of G are
G-conjugate, there are usually several G F -conjugacy classes of F -stable maximal
tori T F , and the structure of the T F is governed by the Weyl group. There is a
1–1 correspondence between G F -conjugacy classes of F -stable maximal tori and
F -conjugacy classes of the Weyl group. This is a crucial ingredient in proving the
main theorem below. Let C be the set of F -conjugacy classes in W , and for C 2 C ,
let TCF denote a representative F -stable maximal torus of G F corresponding to C .
Theorem 2.30. Let G; F; T0 ; W and C be as above, and let Q G F be a quokka
set. Then
jQj X jC j jT F \ Qj
D  C F :
jG F j jW j jTC j
C 2C

Bounds on proportions: Essentially Theorem 2.30 allows us to separate an

estimation problem within a Lie type group G F into two simpler problems, one
within the Weyl group and the other within various maximal tori. The expression for
jQj
jG F j
in Theorem 2.30 as an exact summation can lead to usable bounds. Suppose
that CO is a union of F -conjugacy classes and that `Q is a positive constant such
jTCF \Qj jQj O
jC j
that jTCF j
 `Q for all C 2 CO. Then Theorem 2.30 implies that jG F j
 `Q jW j
.
jTCF \Qj jQj jC j O
Similarly, if uQ is such that jTCF j
 uQ for all C 2 CO, then jG F j
 uQ jW j.

A worked example: Let G D SL.n; Fq / and let Q be the quokka set of ppd-
.n; qI e/ elements of G, for some e 2 .n=2; n/—see Example 2.29. We use this
“quokka theory” to re-derive Lemma 2.3. The Weyl group is W Š Sn , and each
maximal torus TCF containing an element of Q is of the form

TCF D Zq e 1  other cyclic factors: (2.1)

As we discussed in the last paragraph of the proof of Lemma 2.3, for each such torus,
C jT F \Qj 1
the proportion jT F lies between 1  eC1 and 1. The F -conjugacy class C in W
C j
corresponding to such a torus consists of certain elements of W D Sn containing an
e-cycle, and all classes C with this property correspond to tori TCF as in (2.29). Let
CO be the subset of W of all elements containing an e-cycle. Then jCOj=jW j D 1=e,
1
and as we discussed above, jQj=jG F j lies between .1  eC1 / 1e D eC1
1
and 1e .
r-abundant elements: The original impetus to study the work of Isaacs et al. [46]
so closely came from efforts of Niemeyer and Praeger to understand whether, for a
prime r, the lower bound given in [46] for the proportion of r-singular elements in
2 Estimation and Group Algorithms 61

finite classical groups was close to the true proportion. (An r-singular element is one
with order a multiple of r.) Niemeyer conducted a computer experiment on general
linear groups G D GL.n; p a /, for various dimensions n and primes p and r, where
r divides jGj and r ¤ p, to discover the kinds of r-singular elements in G which
appeared frequently on repeated independent random selections from G. It turned
out that a good proportion of the r-singular elements that we found left invariant,
and acted irreducibly on, a subspace of dimension greater than n=2. Moreover, their
frequency seemed to be roughly proportional to 1=e, where e is the smallest positive
integer such that r divides p ae  1. We decided to call these elements r-abundant. It
seemed at first that the r-abundant elements alone occurred with frequency greater
than the lower bound predicted in [46]. However, it was pointed out to us by
Klaus Lux that, hidden in the proofs in [46] was a lower bound on the proportion
of r-singular elements of the form c=e for some constant c, with e as above. If
e > n=2 then these r-singular elements are the ppd-.n; p a I e/ elements used in
the classical recognition algorithm in [72], and in general r-abundant elements
are as easily recognisable as ppd elements from properties of their characteristic
polynomials: namely, there is an irreducible factor f .x/ of degree greater than n=2
and a multiple of e, such that x has order a multiple of r modulo f .x/ in the
polynomial ring Fpa Œx. A detailed study of r-abundant elements was carried out
by Niemeyer and Praeger with Tomasz Popiel [71] to prove that the experimentally
observed proportion of r-singular elements in general linear groups is correct, and to
find and prove analogues for other finite classical groups. The r-abundant elements
form a quokka set, and their proportion was determined [71, Theorem 1.1] using
Theorem 2.30. For the general linear group GL.n; p a /, the proportion is
 
1 ln.2/
1 t 1

r .r C 1/ e

with an error term of the form c=n for some constant c, where r t is the largest power
of r dividing p ae  1. It would be interesting to know if r-abundant elements could
be useful algorithmically to identify classical groups. To aid our understanding of
such elements, Sabina Pannek is undertaking a Ph.D. project to find which maximal
subgroups of finite classical groups contain elements with an irreducible invariant
subspace of the natural module of more than half the dimension.

2.3.4 Strong Involutions in Classical Groups

In [53], Leedham-Green and O’Brien introduced a new Las Vegas algorithm to

find standard generators for a finite simple n-dimensional classical group H in odd
characteristic in its natural action. (Recall that a randomised algorithm is called
Las Vegas if the output, if it exists, is always correct; the algorithm may report
failure with a small probability.) The algorithm of [53] proceeds by constructing
recursively various centralisers of involutions (elements of order 2), the details of
62 A.C. Niemeyer et al.

Table 2.1 The classical S X n

groups for Theorem 2.31 and
SL.` C 1; q/ GL.` C 1; q/ `C1
Corollary 2.32
SU.` C 1; q/ GU.` C 1; q/ `C1
Sp.2`; q/ GSp.2`; q/ 2`
SO.2` C 1; q/ GO.2` C 1; q/ 2` C 1
SO˙ .2`; q/ GO˙ .2`; q/0 2`

which are discussed further in Sect. 2.4.3. The issue we address here is how to find
an appropriate involution. Leedham-Green and O’Brien wished to work with an
involution whose centraliser would be essentially a product of two smaller classical
groups, each of roughly half the dimension. They called such involutions “strong”:
an involution is strong if its fixed point subspace has dimension in Œn=3; 2n=3/,
or equivalently if its 1-eigenspace has dimension in .n=3; 2n=3. Let I denote
the subset of strong involutions in H . Leedham-Green and O’Brien constructed
elements of I by making independent, uniformly distributed random selections from
H to find an element of even order which powered up to a strong involution. We call
such elements preinvolutions. To estimate how readily a preinvolution can be found
by random selection, we need to estimate the size of the set

P.H; I / D fh 2 H j ord.h/ is even, hord.h/=2 2 I g: (2.2)

Leedham-Green and O’Brien estimated that it would require O.n C n4 log n C

4
n log q/ elementary field operations (that is, additions, multiplications or inver-
sions) to compute a strong involution in H , where  is an upper bound on
the number of elementary field operations required to produce an independent,
uniformly distributed random element of H ; see [53, Theorem 8.27]. Underpinning
this complexity estimate was their estimate that the proportion of preinvolutions in
H was at least c=n, for a constant c.
Niemeyer and Praeger, with Frank Lübeck, used the approach described in
Sect. 2.3.3 to obtain an improved estimate for this proportion [58, Theorem 1.1].
They considered any n-dimensional classical group H satisfying S  H  X ,
where S , X , n are as in one of the lines of Table 2.1 with q odd. Here
GO˙ .2`; q/0 denotes the connected general orthogonal group—the index 2 sub-
group of GO˙ .2`; q/ that does not interchange the two SO˙ .2`; q/-classes of
maximal isotropic subspaces.
Theorem 2.31. Let H satisfy S  H  X , with S , X , n as in one of the lines of
Table 2.1, with q odd and `  2, and let I H be the set of strong involutions.
Then
jP.H; I /j 1
 :
jH j 5000 log2 .`/
The weak constant of 1=5000 arises from the fact that the estimation only
considered one class of elements that power up to a strong involution, and from the
fact that it determined one constant that is valid uniformly for all classical groups.
2 Estimation and Group Algorithms 63

A more detailed analysis taking into account a wider family of preinvolutions would
yield a larger value for the constant.
Lübeck, Niemeyer and Praeger also obtained similar lower bounds for projective
groups: note that, for Z0  Z.X /, since the subset I of involutions in Theorem 2.31
contains no central elements, the set I WD IZ0 =Z0 is a subset of involutions in the
projective group H WD HZ0 =Z0 .
Corollary 2.32. With the above notation, jP.H ; I /j=jH j  1=.5000 log2 `/.
Using this new lower bound reduces the complexity of computing a strong
involution in [53] to O.log.n/ C n4 log n C n4 log q/; that is, replacing the first
summand n by log.n/. It seems to be typical that whenever “quokka theory” is
applicable, it produces superior estimates to more intuitive geometric methods.
In Sect. 2.4.3, the algorithm in [53] will be discussed further. Here we just
mention that the proof of [58, Theorem 1.1] could have been given for a more
general class of involutions called “balanced involutions”. For constants ˛; ˇ such
that 0 < ˛ < 1=2 < ˇ < 1, an .˛; ˇ/-balanced involution in an n-dimensional
classical group H is one with fixed point subspace having dimension in Œ˛n; ˇn/.
The resulting lower bound on the proportion of .˛; ˇ/-balanced involutions in H
would be c= log2 .n/, for a constant c depending only on ˛ and ˇ.

2.3.5 More Comments on Strong Involutions

Before leaving this topic we make some comments about the proof of Theorem 2.31.
First, it is not difficult to see that P.H; I / is a quokka set: it is non-empty since I ¤
;; it is conjugacy closed since I is a union of H -conjugacy classes; and finally, since
q is odd, if g D us D su is the Jordan p-decomposition then gord.g/=2 D s ord.s/=2 ,
and hence g 2 P.H; I / if and only if s 2 P.H; I /.
To obtain the lower bound in Theorem 2.31 we used Theorem 2.30. A special
subset C0 of F -conjugacy classes of W was examined, for which it was possible
both to estimate w0 WD j [C 2C0 C j=jW j and to find a good positive lower bound on
jTCF \ P.H; I /j for each C 2 C0 . To give an understanding of this subset of W ,
while avoiding the technicalities associated with small dimensions and the other
types of classical groups, we confine our attention to H D GL.n; q/ with n  7.
Here C0 is a set of conjugacy classes in W D Sn . We choose a particular positive
integer a as follows, and take W0 WD [C 2C0 C to consist of all permutations with a
single cycle of length 2a k 2 .n=3; 2n=3, for some integer k, and no other cycle of
length divisible by 2a . For a0 D log2 ln 2 C log2 log2 n, we take a to be the integer
in the interval Œa0  1=2; a0 C 1=2/. We note for later use that, since n  7, we have
a  1 and .13=4/  2a  n.
First we show that jTCF \ P.H; I /j=jTCF j  1=2, for C 2 C0 with a cycle of
length 2a k as above. Each torus TCF in the H -conjugacy class of tori corresponding
a
to C is of the form Z  A, where Z is cyclic of order q 2 k  1 leaving invariant
a subspace U of dimension 2a k and acting as a Singer cycle on U , and for each
x 2 A, the 2-part of ord.x/ (that is, the highest power of 2 dividing ord.x/) is
64 A.C. Niemeyer et al.

a
strictly less than the 2-part of q 2 k  1. Now half of the elements z 2 Z are such
a
that the 2-part of ord.z/ is equal to the 2-part of q 2 k  1, and for each such z, and
any x 2 A, the element zx has even order, and .zx/jzxj=2 is the unique involution z0
in Z. The element z0 acts as I on the subspace U and has fixed point subspace
of dimension n  2a k 2 Œn=3; 2n=3/; that is to say, z0 is a strong involution and
zx 2 P.H; I /. Thus jTCF \ P.H; I /j=jTCF j  1=2.
Theorem 2.30 now implies that

jP.H; I /j 1 jW0 j
  ;
jH j 2 jW j

so it remains to estimate the size of W0 . A straightforward counting argument yields

jW0 j X p:2a .n  2a k/
D ; (2.3)
jW j 2a k
k

where the sum is over integers k such that n=3 < 2a k  2n=3, and p:2a .n  2a k/
is the proportion of elements in Sn2a k with no cycle of length divisible by 2a . By
Lemma 4.2(a) of [58], which is based on Theorem 2.26,
1 a 1 a
p:2a .n  2a k/ > .n  2a k/1=2 > n1=2 :
4 4
a
Thus each summand in (2.3) is at least 3=.8n1C1=2 / since 2a k  2n=3. The number
of summands in (2.3) is at least .2n=3  n=3/=2a  1 D n=.3  2a /  1, which is at
least n=.39  2a / (since .13=4/  2a  n). Hence

jP.H; I /j 1 jW0 j 1 n 3 1 1
     D  ;
jH j 2 jW j 2 39  2a 8n1C1=2a 208 2a  n1=2a
1 1 1
which is greater than 208 3 log2 .n/
D 624 log . This proves Theorem 2.31 for H D
2 .n/
GL.n; q/.
The family W0 of elements of the Weyl group W gives a far better lower bound
than bounds obtained by geometric arguments. However we have not considered all
conjugacy classes in W , and indeed it seems that, for this problem, application of
“quokka theory” does not yield an upper bound. It is reasonable to ask how good
the lower bound of Theorem 2.31 is. To attempt to answer this question, we quote a
few sentences from [58, p. 3399].
We did some numerical experiments for small q 2 f3; 5; 9; 13g and groups from the
theorem up to dimension 1000. We computed many pseudo-random elements and checked
if they powered up to an involution with a fixed point space of dimension in the right
range. The proportion of these elements is not a monotonic function in the dimension,
but the trend was that the proportion was about 25% for small dimensions and went
down to about 15% in dimension 1000 (independently of the type of the group and q).
Further, statistical tests on the data from the groups H we sampled strongly indicates that
P .H; I /=jH j D O.1= log.`//. This seems to suggest at least that we cannot expect that
there is a lower bound independent of the rank of the group.
2 Estimation and Group Algorithms 65

2.3.6 Regular Semisimple Elements and Generating Functions

Let H be an n-dimensional classical group in odd characteristic, as in one of the

lines of Table 2.1. The methods described in Sect. 2.3.4 show how to find a strong
involution efficiently, or more generally, how to find an .˛; ˇ/-balanced involution z.
The problem of constructing the centraliser CH .z/ of such an involution will be
discussed in Sect. 2.4. In this section we explore an estimation problem connected
with part of the construction. An essential component in finding CH .z/ is to take
random conjugates zg to find a “nice product” y WD zzg , where “nice” means “close
to regular semisimple”. This procedure is discussed in the seminal paper [78] by
Christopher Parker and Rob Wilson. They estimate that O.n/ random products will
produce a nice product with high probability. The approach taken by Praeger and
Seress [86], and described in this section, shows that only O.log n/ random products
are required.
Written in an appropriate basis, the product y D zzg of an involution z and a
random conjugate zg of z has the following form, where y0 has no ˙1-eigenvectors:
0 1
Ir 0 0
y WD zzg D @ 0 y0 0 A :
0 0 Is

Typically, the dimension r is close to 2m  n, where m is the maximum of the

dimensions of the ˙1-eigenspaces of z, and s is close to 0. The question arises:
what kind of matrix do we expect for y0 “typically”? Let us restrict attention to
the simplest case where H D GL.n; q/ with q odd. By considering the results
of computer experiments on various .˛; ˇ/-balanced involutions and their random
conjugates for various n and odd q, we discovered that often y0 is “regular
semisimple”. For the following discussion, let us assume that y D y0 .
An element y of GL.n; q/ is called semisimple if is diagonalisable over some
extension field of Fq (see [17, p. 11]), and this is equivalent to its minimal
polynomial my .t/ being multiplicity free. Also y is called regular if its centraliser in
the corresponding general linear group over the algebraic closure of Fq has minimal
possible dimension, namely n (see [17, p. 29]). It turns out that an element y of a
general linear group is regular if and only if my .t/ D cy .t/, where cy .t/ denotes the
characteristic polynomial of y. These two conditions for elements of finite classical
groups are discussed and compared in [69, Note 8.1]. The regular semisimple
elements are those which are both regular and semisimple. In fact, for elements y of
H D GL.n; q/, y is regular semisimple if and only if the characteristic polynomial
cy .t/ for its action on V .n; q/ satisfies

cy .t/ D a product of pairwise distinct irreducible polynomials.

Looking into the analysis of this situation in the paper [78], it is clear that Parker and
Wilson recognised that regular semisimple elements y occur frequently. Moreover,
66 A.C. Niemeyer et al.

the proportion of regular semisimple elements in the full n-dimensional matrix

algebra was estimated by Neumann and Praeger [70]. The main result of [86] (cf.
Theorem 2.34) is a strengthening of the estimates in [70, 78].
The characteristic polynomial cy .t/ has two special properties: firstly, when
y D y0 the element y has no ˙1-eigenvectors, so cy .t/ is not divisible by t ˙ 1.
Secondly, since y z D z1 .zzg /z D zg z D y 1 , the characteristic polynomials of
y and y 1 are equal. Now cy 1 .t/ D cy .t/ is the conjugate polynomial of cy .t/
where, for an arbitrary polynomial f .t/ with f .0/ ¤ 0, its conjugate polynomial is
f  .t/ WD f .0/1 t degf f .t 1 /. Thus cy .t/ D cy .t/ is self-conjugate. We have seen
that conjugation by z inverts y, and similarly conjugation by zg inverts y. Inverting
a regular semisimple matrix pins down the conjugacy class of the involution z, as
shown in [86, Lemma 3.1]. For n even and q odd, let C GL.n; q/ denote the the
conjugacy class of involutions with fixed point space of dimension n=2.
Lemma 2.33. Let z; y 2 GL.n; q/ with q odd, such that y is regular semisimple
with characteristic polynomial cy .t/ coprime to t 2  1, and z is an involution
inverting y. Then n is even, z 2 C , and zy is also an involution which inverts y.
By Lemma 2.33, we have a bijection .z0 ; z/ 7! .y; z/ between the sets
 ˇ
ˇ y WD zz0 regular semisimple
X D .z; z0 / 2 C  C ˇ
ˇ with cy .t/ coprime to t 2  1

and 8 ˇ 9
< ˇ y; z 2 GL.n; q/; z2 D 1; y z D y 1 =
ˇ
Y D .y; z/ ˇ y regular semisimple, and :
: ˇ ;
ˇ c .t/ coprime to t 2  1
y

The set X is relevant for algorithmic purposes, while the set Y is more amenable
to estimation techniques. For the algorithm, we are given (that is to say, we have
already found) the involution z 2 C , and we want to know the proportion of z0 2 C
such that .z; z0 / 2 X . This is

jfz0 2 C j .z; z0 / 2 X gj jX j jY j j GL.n; q/j jY j

D 2
D 2
D 2

jC j jC j jC j jC j j GL.n; q/j
4
and the first factor on the right of the equality, namely j GL.n;q/j
jC j2
D j GL.n=2;q/j
j GL.n;q/j , lies
1 7 1 2
between .1  q / and .1  q / . Thus the essential problem is to estimate

jY j
ss.n; q/ WD :
j GL.n; q/j

Parker and Wilson [78] give a heuristic that estimates this quantity as being at least
c=n if we require in addition that y has odd order. Our approach gives a surprisingly
precise answer; see [86, Theorem 1.2]. Since n is even we consider ss.2d; q/.
2 Estimation and Group Algorithms 67

Theorem 2.34. For a fixed odd prime power q, the limit of ss.2d; q/ as d ! 1
exists and
ss.1; q/ WD lim ss.2d; q/ D .1  q 1 /2 :
d !1
p
Moreover jss.2d; q/  ss.1; q/j D o.q0d / for any q0 such that 1 < q0 < q.
Corollary 2.35. There exists c > 0 with the property that for any z 2 C the
proportion of z0 2 C such that .z; z0 / 2 X is bounded below by c.
We use generating functions discussed in Sect. 2.2.5 to study the quantities
ss.2d; q/. We define

X
1
S.u/ D ss.2d; q/ud where ss.0; q/ D 1:
d D0

Since y is regular semisimple, cy .t/ is multiplicity-free, and since y is inverted by

the involution z, we have a factorisation
! 0 s 1
Y r Y
cy .t/ D fi .t/  @ gj .t/gj .t/A (2.4)
i D1 j D1

where each fi D fi  has even degree, and each gj ¤ gj , with the fi ; gj ; gj
pairwise distinct monic irreducibles. We use this decomposition to find in [86,
Lemma 3.2] that the number of pairs .y 0 ; z/ 2 Y such that y 0 has characteristic
polynomial cy .t/ is equal to

j GL.2d; q/j
Q 1
 Q :
r s
2 deg fi deg gj  1/
i D1 .q  1/ j D1 .q

Summing over all possible cy .t/ gives an expression for ss.2d; q/j GL.2d; q/j.
Comparing the expression we obtain for ss.2d; q/ by this process with the
coefficient of ud in the infinite product
!  
Y 1
u 2 deg f Y udeg g
1C 1
 1C ;
q 2 deg f  1 q deg g  1
f Df  ; irred: fg;g  g; g¤g  ; irred:

we see that the two expressions are the same. Hence S.u/ is equal to this infinite
product. The contribution to the infinite product from each irreducible polynomial
f or conjugate pair fg; g g of non-self-conjugate polynomials depends only on the
degrees of the polynomials. Thus

Y um
N  .qI2m/ Y um
M  .qIm/
S.u/ D 1C  1C (2.5)
m1
qm  1 m1
qm  1
68 A.C. Niemeyer et al.

where the exponents are

N  .qI m/ D # monic irreducible self-conjugate polynomials over

Fq of degree m:

M .qI m/ D # (unordered) conjugate pairs of monic irreducible
non-self-conjugate polynomials over Fq of degree m:

It turned out that a somewhat similar infinite product arose when Praeger was
studying separable matrices in finite unitary groups with Jason Fulman and Peter
Neumann in [36]. A similar analysis to that given in [36] for these matrices yielded:
1. S.u/ is analytic for juj < 1 with a simple pole at u D 1.
p
2. S.u/ D .1  u/1 H.u/, with H.u/ analytic for juj < q.
Completing the analysis we found the asymptotic behaviour of the ss.2d; q/, as
in Theorem 2.34.

2.4 Computing Centralisers of Involutions

The results in the previous section play a significant role in the analysis of algo-
rithms to compute centralisers of involutions. In general the problem of computing
centralisers is of great importance in theoretical computer science and in group
theory. In computer science, the main interest stems from the connection with the
graph isomorphism problem.
Problem 2.36. (ISO) Given: graphs 1 .V1 ; E1 / and 2 .V2 ; E2 /.
Find: an edge-preserving bijection between V1 and V2 , or prove that no such
bijection exists.
ISO is polynomial-time reducible to the following computational problems with
permutation groups.
Problem 2.37. (STAB) Given: a permutation group G  Sym.˝/ and a subset
 ˝.
Find: the set stabiliser StabG ./ D fg 2 G j g D g.
Problem 2.38. (INT) Given: permutation groups G; H  Sym.˝/.
Find: the intersection G \ H .
Problem 2.39. (CENT) Given: permutation groups G; H  Sym.˝/.
Find: the centraliser CG .H / D fg 2 G j hg D h for all h 2 H g.
Problems 2.37–2.39 are in the same class of the complexity hierarchy, which
means that they can be reduced to each other in time polynomial in the input
length [59].
2 Estimation and Group Algorithms 69

The reduction of ISO is easiest to STAB or INT. First, we notice that 1 .V1 ; E1 /
and 2 .V2 ; E2 / are isomorphic if and only if 1 [ 2 (disjoint copies of 1 and 2 )
has an automorphism that exchanges V1 and V2 . Therefore, it is enough to compute
automorphism groups of graphs. Given a graph .V; E/, define ˝ as the set of
unordered pairs in V . Then E corresponds to a subset  ˝, and Sym.V / acts as
a group G on ˝. We can compute Aut. / as Aut. / D StabG ./ or Aut. / D
G \ .Sym./  Sym.˝ n //.
Although, using backtrack methods (see e.g. [87, Chap. 9]), ISO and CENT
are usually easy to solve in practice, no polynomial-time solution is known for
Problems 2.36–2.39. Special cases with polynomial-time solutions are of great
theoretical and practical interest.
In group theory, the most important case of centraliser computations is to
construct centralisers of involutions. On the theoretical side, a major tool in the
study and classification of finite simple groups is the investigation of their involution
centralisers [41]. On the computational side, in the last decade involution centraliser
computations became prevalent [1, 7, 45, 53, 56, 78]. In the next subsections, we
describe some applications of centraliser computations; Bray’s algorithm [16] for
computing centralisers of involutions; and efforts to analyze Bray’s algorithm.

2.4.1 Applications of Centralisers of Involutions Computations

A recent active area of computational group theory is the so-called matrix group
recognition project. Let V be a finite dimensional vector space over a finite
field Fq . Given G D hS i  GL.V /, the goal is to compute quantitative and
structural information about G such as the order, a composition series, and important
characteristic subgroups like the largest solvable normal subgroup of G.
There are two main approaches to matrix group recognition. The geometric
approach, initiated by Neumann and Praeger [69] and currently led by Leedham-
Green and O’Brien [52,77], is based on Aschbacher’s classification of matrix groups
[2]. Aschbacher defines nine categories of matrix groups G. In seven of these
categories, there is a natural normal subgroup N C G that can be used to divide
the recognition problem into two smaller subproblems on N and G=N . Based on
that result, the geometric approach tries to find a homomorphism ' W G ! H into
an appropriate permutation or matrix group H , and recursively recognise Im.'/
and Ker.'/. In contrast, the black-box group approach of Babai and Beals [4] aims
for the abstract group theoretic structure of G. Babai and Beals define a series of
characteristic subgroups, present in all finite groups, and initiate a program that
tries to compute a composition series going through these characteristic subgroups.
Both approaches eventually lead to simple (or quasisimple) matrix groups, where
further divide-and-conquer is impossible. For such groups, a major issue is the
solution of the constructive membership problem.
70 A.C. Niemeyer et al.

2.4.2 Constructive Membership in Lie Type Groups

Definition 2.40. A black-box group G is a group whose elements are encoded by

bit strings (strings consisting of 0s and 1s) of uniform length. Moreover, there are
oracles for the following tasks. Given strings representing g; h 2 G, we can compute
a string representing gh; a string for g 1 ; and we can decide whether g D 1.
A black-box algorithm is an algorithm that, given G by a set of generators, uses
only the black-box oracles.
The definition of black-box groups covers the “concrete” representations of
groups as permutation groups or matrix groups defined over finite fields. Note that if
G is a black-box group and N is a recognisable normal subgroup (i.e., given a string
representing some g 2 G, we can decide whether g 2 N ), then G=N is also a black-
box group. This observation plays a crucial role in recursive algorithms, allowing
us to work in factor groups. Also note that we require only that N is recognisable,
but N is not necessarily constructed (i.e., we may not have a generating set for N
in hand). Examples of recognisable normal subgroups that may be hard to construct
are the centre and the largest soluble normal subgroup of G. Black-box groups were
introduced by Babai and Szemerédi [5]. For an introduction to the basic black-box
group algorithms, see [87, Chap. 2].
A black-box group algorithm does not use specific features of the group
representation, nor particulars of how group operations are performed. For example,
we lose all information stored implicitly in the cycle structure of a permutation, or
in the characteristic polynomial of a matrix. In practice, and also in some theoretical
considerations, we often allow oracles for some other operations; an example is an
oracle to compute element orders.
The very reasonable and justified question arises: why do we handicap ourselves
with black-box group algorithms? One answer is that in certain situations, we cannot
do more than the black-box operations. For example, to generate random elements
in a matrix group, so far every algorithm takes repeated products and inverses of
the given generators, and after a while declares the last element constructed as a
random element of the input group [3, 18, 24]. Bray’s algorithm (see Sect. 2.4.4)
for computing centralisers of involutions is another example of a black-box group
algorithm, with a possible enhancement using element order oracles. Another, more
unusual answer is that elements of a permutation group can be described as unique
words in a strong generating set (SGS), constructed in a canonical way. The group
operations are performed using the images of elements of the base associated with
the SGS. For the important class of small-base groups, these group operations
are much faster than permutation multiplication, but the algorithms using this
representation are strictly black-box. For details, we refer to [87, Chap. 5.4].
Next, we define the notion of a straight-line program (SLP). Expressing elements
of a group G in a given set of generators may result in words of length proportional
to jGj; intuitively, SLPs are shortcuts, to reach group elements faster from a set of
generators. By [5], every g 2 G can be reached from any set of generators by an
SLP of length at most .1 C log jGj/2 .
2 Estimation and Group Algorithms 71

Definition 2.41. Given G D hS i and g 2 G, a straight-line program (SLP)

reaching g from S is a sequence of expressions W D .w1 ; : : : ; wm / such that, for
i D 1; 2; : : : ; m,
1. wi is a symbol for some s 2 S ; or
2. wi D .wj ; wk / for some j; k < i ; or
3. wi D .wj ; 1/ for some j < i .
We define the evaluation of W the natural way: eval.wj ; wk / D eval.wj /eval.wk /
and eval.wj ; 1/ D eval.wj /1 ; and require that eval.wm / D g.
Finally, we are ready to define the constructive membership problem.
Definition 2.42. A constructive membership algorithm for a group G is a black-
box group algorithm that, given the black-box group G D hS i and g 2 G,
constructs an SLP reaching g from S .
The main result of this subsection is the following theorem by Holmes et al. [45].
Theorem 2.43 ([45]). Let G be a black-box group equipped with an order oracle.
There is a black-box Monte Carlo algorithm which reduces the constructive
membership problem for G to three instances of the same problem for centralisers
of involutions of G.
Proof. Let G D hS i and g 2 G. An algorithm constructing an SLP reaching g from
S consists of the following steps.
1. Find h 2 G with ord.gh/ D 2`. Define z WD .gh/` .
2. Find an involution x 2 G with ord.xz/ D 2m. Define y WD .xz/m .
3. Construct X D CG .x/.
4. Solve the constructive membership problem for y 2 X .
5. Construct Y D CG .y/.
6. Solve the constructive membership problem for z 2 Y .
7. Construct Z D CG .z/.
8. Solve the constructive membership problem for gh 2 Z.
9. Compute and return an SLP for g.
To prove the correctness of the algorithm, observe that z, constructed in Step 1, is
an involution centralising gh. In Step 2, y is in the centre of the dihedral group
hx; zi, so x is an involution centralising y and y is an involution centralising z.
Hence Steps 3, 5 and 7 compute centralisers of involutions, and the constructive
membership problems in Steps 4, 6 and 8 indeed try to reach elements of G that
are in the appropriate subgroups. Finally, note that the construction of x provides
an SLP reaching x from S and, consequently, we have SLPs reaching y, then z,
then gh from S . Also, in Step 1, we construct an SLP reaching h from S . Hence, in
Step 9, we can construct an SLP reaching g from S . t
u
Remark 2.44. We note that the hypothesis of Theorem 2.43 that G has an order
oracle can be relaxed. The only places in the algorithm where the order oracle is
used are in Steps 1 and 2. For example, at the construction of z in Step 1, we
72 A.C. Niemeyer et al.

can proceed the following way. Instead of computing `, we can raise gh to an

appropriate multiple of the odd part of jGj. To find such a multiple (without knowing
jGj), it is enough to know a superset of primes occurring in jGj or, in the case of
a matrix group G  GL.n; q/, we can work with the set of pseudoprimes: these
are the largest divisors of the numbers q e  1 for e  n, that are relatively prime
to q j  1 for all j < e. The pseudoprimes can be computed in polynomial time
(polynomial in terms of n and log q). For details, see [4]. The use of the order oracle
in Step 2 can be avoided in exactly the same way.
In [45], Holmes et al. show that if G is a simple group of Lie type then the
algorithm described in Theorem 2.43, not counting the time requirement of Steps 4,
6 and 8, runs in polynomial time. However, we cannot apply the theorem recursively
to the groups in these steps, because they are not simple. Therefore, we need a
recursive scheme involving all groups, not only the simple ones. Such a scheme is
designed by Babai et al. in [7]; Theorem 2.43 is a crucial ingredient in the following
result.
Theorem 2.45 ([7]). There is a randomised polynomial-time algorithm, employing
certain number-theoretical oracles, which, given a matrix group G  GL.n; q/ of
odd characteristic, solves the constructive membership problem in G.
The required number-theoretical oracles are the factorisation of integers of the
form q e  1, for 1  e  n, and the solution of the discrete logarithm problem:
given a; b 2 Fq e , decide whether a 2 hbi; and, if the answer is affirmative, then
find an integer x such that a D b x . In polynomial-time algorithms for matrix
groups, it is customary to assume the use of these number theory oracles as they
are already needed in finding a composition series and the order of a 1  1 matrix
group over Fq . We note that Theorem 2.45 extends to matrix groups defined over
fields of characteristic 2, with some restrictions on the composition factors of G. It
is expected that these restrictions will be removed in the near future, as constructive
membership algorithms in all simple groups are in the offing.

2.4.3 Constructive Recognition of Lie Type Groups

Membership testing is an important first step in exploring a permutation or matrix

group G; however, for studying the structure of G and constructing important
subgroups, it is beneficial to identify the composition factors of G with standard
copies of these factor groups. For alternating and classical groups, the standard copy
is the natural permutation and matrix representation, respectively. For exceptional
groups, the definition of a standard copy is not so clear-cut: we may choose the
smallest-dimensional matrix representation, or a Bruhat decomposition, or any other
representation we may be able to control. Here we only give a formal definition for
classical groups, taken from [48].
2 Estimation and Group Algorithms 73

Definition 2.46. Constructive recognition of a black-box group G D hS i isomor-

phic to a simple classical group defined on some vector space over a field of given
characteristic p is an algorithm that verifies that there is, indeed, an isomorphism,
and finds the following:
(i) The field size q D p e ; as well as the type and the dimension d of G.
(ii) A new set S  generating G; a vector space Fdq , and a monomorphism W G !
PSL.d; q/; specified by the image of S  ; such that G acts projectively on Fdq
as a classical group defined on Fdq .
Moreover; the data structures underlying (ii) yield deterministic algorithms for each
of the following:
(iii) Given g 2 G; find g and a straight-line program from S  to g.
(iv) Given h 2 PGL.d; q/; decide whether or not h 2 GI and; if it is; find h1
and a straight-line program from S  to h1 .
(v) Find a form on Fdq involved in the definition of G as a classical group; if G 6Š
PSL.d; q/.
Although Definition 2.46 is formulated in the general context of black-box
groups, of course it can be applied to any given permutation or matrix representation
of G. The simplest but most important case is when G is already given in its
natural representation, and the only task is to find “nice” generators S  such that
each element of G can be reached easily from S  . For classical groups of odd
characteristic, this task has been accomplished by Leedham-Green and O’Brien by
a highly efficient algorithm [53]. A rough outline of their procedure is given in
Algorithm 5.

Algorithm 5: CONSTRUCTIVERECOGNITION
Input: G D hSi  GL.V / Š GL.n; q/, q odd, G is a classical group in its natural
representation;
Output: A data structure for constructive recognition of G;
.1/ repeat
y WD random element of G;
until ord.y/ is even and x WD y ord.y/=2 has ˙1-eigenspaces E1 ; E1 with
dim.Ei / 2 .n=3; 2n=3/;
.2/ Construct H D CG .x/;
.3/ Recursively solve constructive recognition for the restriction of H to its action on E1
and E1 ;
.4/ Use the result of Step .3/ to obtain nice generators and data structure for constructive
recognition of G;

The following simple lemma from [84] implies that Step .3/ is indeed a recursive
call.
74 A.C. Niemeyer et al.

Lemma 2.47. Let G, x, E1 , E1 be as in Algorithm 5, with G classical but not

linear. Then V D E1 ? E1 , and both E1 and E1 are nondegenerate (and of even
dimension if G is symplectic).
Proof. For u 2 E1 and w 2 E1 , we have .u; w/ D .u; w/x D .u; w/ and hence
?
.u; w/ D 0. Thus E1 E1 . Since the bilinear form is nondegenerate, dim.E1 / D
? ? ?
n  dim.E1 / D dim.E1 / and hence E1 D E1 . Therefore, E1 \ E1 D 0 so
E1 , and similarly also E1 , are nondegenerate. In particular, E1 and E1 both have
even dimension if G is symplectic. t
u
Since, for i 2 f1; 1g, x acts as a scalar matrix on Ei , Lemma 2.47 implies that
the restriction of H to Ei is a classical group of the same type as G and Step .3/ is
indeed a recursive call. Note that the requirement dim.Ei / 2 .n=3; 2n=3/ ensures
that CG .x/ can be split into two parts of roughly equal size, thereby ensuring that
the depth of the recursion is logarithmic in n.
To analyze Algorithm 5, for the first two steps we have to estimate (i) the
proportion of elements y as in Step (1); and (ii) give a running time estimate for the
construction of involution centralisers. Task (i) has been accomplished in Sect. 2.3.4.
In the next two subsections, we describe and analyze an algorithm for computing
involution centralisers.

2.4.4 Computation of an Element Centralising an Involution

In this subsection we describe an algorithm by Bray [16] that constructs an element

in the centraliser of a given involution.

Algorithm 6: CENTRALISINGELEMENT
Input: G D hSi and an involution x 2 G;
Output: An element of CG .x/;
.1/ g WD random element of G;
.2/ y WD x  x g and m WD ord.y/;
.3/ if m is even then
return .g/ WD y m=2
else
return .g/ WD y .mC1/=2 g 1
end

We note that the order computation in Step (2) may be avoided, using a superset
of primes occurring in G, or pseudoprimes (see Remark 2.44).
Lemma 2.48. The output of Algorithm 6 is correct: no matter which g 2 G is
chosen in Step .1/, we have .g/ 2 CG .x/.
2 Estimation and Group Algorithms 75

Proof. For any g 2 G, the group D WD hx; x g i is dihedral, of order 2m. If m is

even then .g/ 2 Z.D/; in particular, .g/ centralises x 2 D.
If m is odd then, using that x 2 D 1, we obtain
mC1
m1 mC1
xy D .xg 1 xg/ x.xg 1 xg/ D xg :
2
2 2

mC1
Comparison of the leftmost and rightmost terms gives .g/ D y 2 g 1 2 CG .x/.
t
u
We say that g 2 G is of even type if y D xx g has even order, and g 2 G is of
odd type if y D xx g has odd order. Note that for any c 2 CG .x/, x cg D x g , so
xx g D xx cg and consequently g and cg have the same type. Moreover, .xx g /c D
xx gc so xx g and xx gc are conjugate, have the same order, and g and gc have the
same type. Combining the last two observations, we obtain that in a double coset
CG .x/  g  CG .x/, all elements have the same type.
Lemma 2.49. (i) If g is chosen from the uniform distribution on the set of odd type
elements of G then .g/ is a uniformly distributed random element of CG .x/.
(ii) If g is chosen from the uniform distribution on the set of even type elements
of G and .g/ is in the conjugacy class C of involutions in CG .x/ then .g/ is
a uniformly distributed random element of C .
mC1
Proof. (i) Suppose that g is of odd type. For c 2 CG .x/, we have y 2 .cg/1 D
mC1
y 2 g 1 c 1 and so .cg/ D .g/c 1 . Hence, as cg runs through the coset
mC1
CG .x/ g, y 2 g 1 c 1 runs through CG .x/. This implies that if g runs through
the elements of G of odd type then each element of CG .x/ occurs as .g/
exactly the same number of times.
(ii) Suppose now that g is of even type. Then .g/ D .xx g /m=2 is an involution; let
C denote its conjugacy class in CG .x/. As gc runs through the coset g  CG .x/,
.gc/ D .xx gc /m=2 D ..xx g /m=2 /c covers each element of C the same number
of times. Hence each element of a fixed conjugacy class C of involutions in
CG .x/ has the same chance to occur as .g/ for some g of even type. t
u

2.4.5 Computation of the Full Centraliser

In order to compute a set X of generators of CG .x/ for a given group G and

involution x 2 G, we may construct a sequence .g1 ; : : : ; gm / of random elements
in G and take X WD f.gi / j 1  i  mg. By Lemma 2.48, we always have
hX i  CG .x/, but when can we stop? How large should m be so that, with high
probability, X generates the entire group CG .x/?
By Lemma 2.49, random elements gi of odd type are highly desirable, since then
.gi / is a uniformly distributed random element of CG .x/. Such a random element
76 A.C. Niemeyer et al.

.gi / 2 CG .x/, added to an already constructed proper subgroup H < CG .x/,

increases H with probability 1  1=jCG .x/ W H j  1=2, so if we know an
upper bound ` for the length of subgroup chains in CG .x/ then we may estimate
how many elements gi of odd type we need to encounter. For polynomial-time
computations, the trivial bound `  log2 jGj suffices, but sometimes we have much
better estimates for the number of required random generators. In particular, in the
especially important case when G is a simple group of Lie type defined over a field
of odd characteristic, the structure of involution centralisers is known. Consequently,
for any involution x 2 G, the number of uniformly distributed random elements
needed to generate CG .x/ with probability greater than 1  " can be bounded by a
function of ", independent of G and x [57]. Therefore, the following seminal result
of Parker and Wilson [78] has great importance in the analysis of many matrix group
algorithms.
Theorem 2.50 ([78]). There exists a positive constant c such that:
(i) If G is a simple exceptional group of Lie type defined over a field of odd order,
and x is any involution in G, then the probability that a uniformly distributed
random element g 2 G is of odd type is bounded below by c.
(ii) If G is a simple classical group defined over a field of odd order, with natural
module of dimension n, and x is any involution in G, then the probability that a
uniformly distributed random element g 2 G is of odd type is bounded below by
c=n. Moreover, the order of magnitude 1=n for a lower bound is best possible.
Parker and Wilson [78, p. 886] give an indication of how big the constants can
be: “The constants c that can be obtained from our proofs are of the order of 1=1000,
but we have made no attempt to calculate them explicitly, as we conjecture that the
best possible constants are nearer 1=4.”
The basic idea of the proof of Theorem 2.50 is to identify a set of dihedral
subgroups D of twice odd order in G, each D containing the given involution x.
If the random conjugate x g falls into one of these subgroups D then xx g has odd
order and g is of odd type. In order to avoid double counting, we also require that
generators of the maximal cyclic normal subgroup of D be regular semisimple in a
suitable subgroup H  G. (Here H depends on D but H is also of Lie type. We
require the generators of D to be regular semisimple as elements of this Lie type
group, as defined in Sect. 2.3.6.)
While Theorem 2.50 is sufficient to prove polynomial running time of centraliser
of involution computations in Lie type simple groups, the scarcity of elements of
odd type raises the the following questions. Is there an algorithm that uses the
lower quality random elements .gi / 2 CG .x/, obtained from gi of even type,
to generate CG .x/? Can the asymptotic running time of this algorithm be faster than
the construction of CG .x/ using the uniformly distributed .gi / obtained from gi of
odd type? To formulate this problem precisely, we need some definitions.
We consider finite classical groups H of dimension n over a finite field Fq of
odd order q. We denote by H  the generalized Fitting subgroup of H (for example
H  D SL.n; q/ if H D GL.n; q/). Let ˛; ˇ be real numbers such that 0 < ˛ <
2 Estimation and Group Algorithms 77

1=2 < ˇ < 1, and let x 2 H be of order 2. Recall that x is called an .˛; ˇ/-
balanced involution in H if the subspace E1 .x/ of fixed points of x in the underlying
vector space has dimension r where ˛n  r < ˇn. For a given sequence X D
.C1 ; : : : ; Cm / of conjugacy classes of .˛; ˇ/-balanced involutions in H , a c-tuple
.g1 ; : : : ; gm / is a class-random sequence from X if gi is a uniformly distributed
random element of Ci for each i D 1; : : : ; m, and the gi are mutually independent.
Given a classical group G  GL.n; q/ and an involution x 2 G, the centraliser
CG .x/ modulo x is the direct product of two classical groups H .1/ and H .1/ , acting
on E1 .x/ and E1 .x/, respectively. If g 2 G is of even type then .g/ acts as
an involution g .J / on EJ , for J 2 f1; 1g, and if .g1 ; : : : ; gm / is a sequence of
uniformly distributed random elements of even type in G then Lemma 2.49 implies
.J / .J /
that .g1 ; : : : ; gm / is a class-random sequence from some conjugacy classes of
.J / .J /
involutions X .J / D .C1 ; : : : ; Cm /.
With an application in Algorithm 5 in mind, we propose the following problems.
We use the notation and definitions of the previous paragraphs.
Problem 2.51. Given a classical group G  GL.n; q/ and a .1=3; 2=3/-balanced
involution x in G, estimate the probability p that for a uniformly distributed g 2 G
of even type, g .J / is an .˛; ˇ/-balanced involution in H .J / , for both J 2 f1; 1g.
Here ˛; ˇ are constants, chosen appropriately.
Problem 2.52. Let G  GL.n; q/ be a classical group and let X D .C1 ; : : : ; Cm /
be a sequence of conjugacy classes of .˛; ˇ/-balanced involutions in G. Estimate
the minimum value of m such that, with high probability, a class-random sequence
from X generates a subgroup of G containing G  .
If the product .1=p/m, for the probability p from Problem 2.51 and the minimum
value m from Problem 2.52, satisfies .1=p/m D o.n/ then the elements .g/
obtained from even type g generate CG .x/ asymptotically faster than the elements
.g/ obtained from odd type g.
Problem 2.52 has been solved for all classical groups.
Theorem 2.53 ([84]). Let ˛; ˇ be real numbers such that 0 < ˛ < 1=2 < ˇ < 1.
Then there exist integers m D m.˛; ˇ/ and n.˛; ˇ/ such that, for G, n, q as above,
with q odd, if n > n.˛; ˇ/ and X D .C1 ; : : : ; Cm / is a given sequence of conjugacy
classes of .˛; ˇ/-balanced involutions in G, then a class-random sequence from X
generates a subgroup containing G  with probability at least 1  q n .
The basic idea of the proof of Theorem 2.53 is standard: if a class-random
sequence .g1 ; : : : ; gm / does not generate G  then all gi belong to some maximal
subgroup M < G, with M not containing G  . Since gi is uniformly distributed in
its conjugacy class, we have to estimate the ratios jM \ Ci j=jCi j for all maximal
subgroups M . Maximal subgroups are characterised by Aschbacher’s theorem [2];
it turns out that the most difficult case is when M is reducible (has a proper invariant
subspace).
Much less is known about Problem 2.51. At present, a solution is known only in
the case when G  D SL.n; q/.
78 A.C. Niemeyer et al.

Theorem 2.54 ([85]). There exist c and n0 such that if n > n0 , SL.n; q/ 
G  GL.n; q/, x is a .1=3; 2=3/-balanced involution of G, and g 2 G is a
uniformly distributed random element among the elements of G of even type, then
with probability at least c= log n, g.1/ and g .1/ are .1=6; 2=3/-balanced involutions
on the eigenspaces E1 .x/ and E1 .x/ respectively.
The proof of Theorem 2.54 uses a significant enhancement of the generating
function method described in Sect. 2.3.6, and also some ideas from [58].

Acknowledgements This chapter forms part of our Australian Research Council Discovery
Project DP110101153. Praeger and Seress are supported by an Australian Research Council
Federation Fellowship and Professorial Fellowship, respectively. Niemeyer thanks the Lehrstuhl D
für Mathematik at RWTH Aachen for their hospitality, and acknowledges a DFG grant in SPP1489.
All three of us warmly thank the de Brún Centre for Computational Algebra at National University
of Ireland, Galway, for their hospitality during the Workshop on Groups, Combinatorics and
Computing in April 2011, where we presented the short lecture course that led to the development
of this chapter. We are very grateful to Peter M. Neumann for many thoughtful comments and
advice, and his translation of Euler’s words in Sect. 2.2.2.

References

1. C. Altseimer, A.V. Borovik, Probabilistic Recognition of Orthogonal and Symplectic Groups,

in Groups and Computation, III, vol. 8, Columbus, OH, 1999 (Ohio State University Mathe-
matical Research Institute Publications/de Gruyter, Berlin, 2001), pp. 1–20
2. M. Aschbacher, On the maximal subgroups of the finite classical groups. Invent. Math. 76(3),
469–514 (1984)
3. L. Babai, Local Expansion of Vertex-Transitive Graphs and Random Generation in Finite
Groups, in 23rd ACM Symposium on Theory of Computing (ACM, New York, 1991),
pp. 164–174
4. L. Babai, R. Beals, A Polynomial-Time Theory of Black Box Groups. I, in Groups St. Andrews
1997 in Bath, I. London Mathematical Society Lecture Note Series, vol. 260 (Cambridge
University Press, Cambridge, 1999), pp. 30–64
5. L. Babai, E. Szemerédi, On the Complexity of Matrix Group Problems I, in 25th Annual
Symposium on Foundations of Computer Science (IEEE Computer Society Press, Los
Alamitos, 1984), pp. 229–240
6. L. Babai, W.M. Kantor, P.P. Pálfy, Á. Seress, Black-box recognition of finite simple groups of
Lie type by statistics of element orders. J. Group Theor. 5(4), 383–401 (2002)
7. L. Babai, R. Beals, Á. Seress, Polynomial-Time Theory of Matrix Groups, in 41st ACM
Symposium on Theory of Computing, Bethesda, MD, 2009 (ACM, New York, 2009),
pp. 55–64
8. R. Beals, C.R. Leedham-Green, A.C. Niemeyer, C.E. Praeger, Á. Seress, Permutations with
restricted cycle structure and an algorithmic application. Combin. Probab. Comput. 11(5), 447–
464 (2002)
9. R. Beals, C.R. Leedham-Green, A.C. Niemeyer, C.E. Praeger, Á. Seress, A black-box group
algorithm for recognizing finite symmetric and alternating groups. I. Trans. Am. Math. Soc.
355(5), 2097–2113 (2003)
10. D.E.-C. Ben-Ezra, Counting elements in the symmetric group, Int. J. Algebra Comput. 19(3),
305–313 (2009)
11. E.A. Bender, Asymptotic methods in enumeration. SIAM Rev. 16, 485–515 (1974)
2 Estimation and Group Algorithms 79

12. E.A. Bertram, B. Gordon, Counting special permutations. Eur. J. Comb. 10(3), 221–226 (1989)
13. E.D. Bolker, A.M. Gleason, Counting permutations. J. Comb. Theor. Ser. A 29(2), 236–242
(1980)
14. M. Bóna, A. McLennan, D. White, Permutations with roots. Random Struct. Algorithm 17(2),
157–167 (2000)
15. W. Bosma, J. Cannon, C. Playoust, The Magma algebra system. I. The user language.
J. Symbolic Comput. 24, 235–265 (1997)
16. J.N. Bray, An improved method for generating the centralizer of an involution. Arch. Math.
(Basel) 74(4), 241–245 (2000)
17. R.W. Carter, Finite Groups of Lie Type (Wiley Classics Library, Wiley, Chichester, 1993),
Conjugacy classes and complex characters, Reprint of the 1985 original, A Wiley-Interscience
Publication
18. F. Celler, C.R. Leedham-Green, S.H. Murray, A.C. Niemeyer, E.A. O’Brien, Generating
random elements of a finite group. Comm. Algebra 23(13), 4931–4948 (1995)
19. W.W. Chernoff, Solutions to x r D ˛ in the alternating group. Ars Combin. 29(C), 226–227
(1990) (Twelfth British Combinatorial Conference, Norwich, 1989)
20. S. Chowla, I.N. Herstein, W.K. Moore, On recursions connected with symmetric groups. I. Can.
J. Math. 3, 328–334 (1951)
21. S. Chowla, I.N. Herstein, W.R. Scott, The solutions of x d D 1 in symmetric groups. Norske
Vid. Selsk. Forh. Trondheim 25, 29–31 (1952/1953)
22. A.M. Cohen, S.H. Murray, An algorithm for Lang’s Theorem. J. Algebra 322(3), 675–702
(2009)
23. A. de Moivre, The Doctrine of Chances: Or, A Method of Calculating the Probability of Events
in Play, 2nd edn. (H. Woodfall, London, 1738)
24. J.D. Dixon, Generating random elements in finite groups. Electron. J. Comb. 15(1), Research
Paper 94 (2008)
25. P. Dusart, The kth prime is greater than k.ln k C ln ln k  1/ for k  2. Math. Comp. 68(225),
411–415 (2009)
26. P. Erdős, M. Szalay, On Some Problems of the Statistical Theory of Partitions, in Number
Theory, vol. I, Budapest, 1987. Colloq. Math. Soc. János Bolyai, vol. 51 (North-Holland,
Amsterdam, 1990), pp. 93–110
27. P. Erdős, P. Turán, On some problems of a statistical group-theory. I. Z. Wahrscheinlichkeits-
theorie und Verw. Gebiete 4, 175–186 (1965)
28. P. Erdős, P. Turán, On some problems of a statistical group-theory. II. Acta Math. Acad. Sci.
Hung. 18, 151–163 (1967)
29. P. Erdős, P. Turán, On some problems of a statistical group-theory. III. Acta Math. Acad. Sci.
Hung. 18, 309–320 (1967)
30. P. Erdős, P. Turán, On some problems of a statistical group-theory. IV. Acta Math. Acad. Sci.
Hung. 19, 413–435 (1968)
31. P. Erdős, P. Turán, On some problems of a statistical group theory. VI. J. Indian Math. Soc.
34(3–4), 175–192 (1970/1971)
32. P. Erdős, P. Turán, On some problems of a statistical group theory. V. Period. Math. Hung.
1(1), 5–13 (1971)
33. L. Euler, Calcul de la probabilité dans le jeu de rencontre. Mémoires de l’Academie des
Sciences de Berlin, 7 (1751) 1753, pp. 255–270. Reprinted in Opera Omnia: Series 1, vol.
7, pp. 11–25. Available through The Euler Archive at www.EulerArchive.org.
34. L. Euler, Solutio Quaestionis curiosae ex doctrina combinationum. Mémoires de l’Académie
des Sciences de St.-Petersbourg, 3:57–64, 1811. Reprinted in Opera Omnia: Series 1, vol. 7,
pp. 435–440. Available through The Euler Archive at www.EulerArchive.org.
35. P. Flajolet, R. Sedgewick, Analytic Combinatorics (Cambridge University Press, Cambridge,
2009)
36. J. Fulman, P.M. Neumann, C.E. Praeger, A generating function approach to the enumeration of
matrices in classical groups over finite fields. Mem. Am. Math. Soc. 176(830), vi+90 (2005)
80 A.C. Niemeyer et al.

37. The GAP Group, GAP — Groups, Algorithms, and Programming, Version 4.5.2(beta), 2011,
http://www.gap-system.org/
38. S.P. Glasby, Using recurrence relations to count certain elements in symmetric groups. Eur.
J. Comb. 22(4), 497–501 (2001)
39. W.M.Y. Goh, E. Schmutz, The expected order of a random permutation. Bull. Lond. Math.
Soc. 23(1), 34–42 (1991)
40. V. Gončarov, On the field of combinatory analysis. Am. Math. Soc. Transl. 19(2), 1–46 (1962)
41. D. Gorenstein, R. Lyons, R. Solomon, The Classification of the Finite Simple Groups.
Mathematical Surveys and Monographs, vol. 40 (American Mathematical Society, Providence,
1994)
42. O. Gruder, Zur Theorie der Zerlegung von Permutationen in Zyklen. Ark. Mat. 2(5), 385–414
(1953)
43. W.K. Hayman, A generalisation of Stirling’s formula. J. Reine Angew. Math. 196, 67–95
(1956)
44. R.B. Herrera, The number of elements of given period in finite symmetric groups. Am. Math.
Mon. 64, 488–490 (1957)
45. P.E. Holmes, S.A. Linton, E.A. O’Brien, A.J.E. Ryba, R.A. Wilson, Constructive membership
in black-box groups. J. Group Theor. 11(6), 747–763 (2008)
46. I.M. Isaacs, W.M. Kantor, N. Spaltenstein, On the probability that a group element is
p-singular. J. Algebra 176(1), 139–181 (1995)
47. E. Jacobsthal, Sur le nombre d’éléments du groupe symétrique Sn dont l’ordre est un nombre
premier. Norske Vid. Selsk. Forh. Trondheim 21(12), 49–51 (1949)
48. W.M. Kantor, Á. Seress, Black box classical groups. Mem. Am. Math. Soc. 149(708), viii+168
(2001)
49. A.V. Kolchin, Equations that contain an unknown permutation. Diskret. Mat. 6(1), 100–115
(1994)
50. V.F. Kolchin, Random Graphs. Encyclopedia of Mathematics and Its Applications, vol. 53
(Cambridge University Press, Cambridge, 1999)
51. E. Landau, Handbuch der Lehre von der Verteilung der Primzahlen. 2 Bände, 2nd edn.
(Chelsea Publishing Co., New York, 1953), With an appendix by Paul T. Bateman
52. C.R. Leedham-Green, The Computational Matrix Group Project, in Groups and Computation,
III, vol. 8, Columbus, OH, 1999 (Ohio State University Mathematical Research Institute
Publications/de Gruyter, Berlin, 2001), pp. 229–247
53. C.R. Leedham-Green, E.A. O’Brien, Constructive recognition of classical groups in odd
characteristic. J. Algebra 322(3), 833–881 (2009)
54. G.I. Lehrer, Rational tori, semisimple orbits and the topology of hyperplane complements.
Comment. Math. Helv. 67(2), 226–251 (1992)
55. G.I. Lehrer, The cohomology of the regular semisimple variety. J. Algebra 199(2), 666–689
(1998)
56. M.W. Liebeck, E.A. O’Brien, Finding the characteristic of a group of Lie type. J. Lond. Math.
Soc. (2) 75(3), 741–754 (2007)
57. M.W. Liebeck, A. Shalev, The probability of generating a finite simple group. Geom. Dedicata
56(1), 103–113 (1995)
58. F. Lübeck, A.C. Niemeyer, C.E. Praeger, Finding involutions in finite Lie type groups of odd
characteristic. J. Algebra 321(11), 3397–3417 (2009)
59. E.M. Luks, Permutation Groups and Polynomial-Time Computation, in Groups and compu-
tation, New Brunswick, NJ, 1991. DIMACS Series in Discrete Mathematics and Theoretical
Computer Science, vol. 11 (American Mathematical Society, Providence, 1993), pp. 139–175
60. R. Lyons, Evidence for a new finite simple group. J. Algebra 20, 540–569 (1972)
61. A. Maróti, Symmetric functions, generalized blocks, and permutations with restricted cycle
structure. Eur. J. Comb. 28(3), 942–963 (2007)
62. N. Metropolis, The beginnings of the Monte Carlo method. Los Alamos Sci. 15 (Special Issue),
125–130 (1987)
2 Estimation and Group Algorithms 81

63. M.P. Mineev, A.I. Pavlov, The number of permutations of a special form. Mat. Sb. (N.S.)
99(141)(3), 468–476, 480 (1976)
64. P.R. de Monmort, Essay d’analyse sur les jeux de hazard (J. Quillau, Paris, 1708)
65. P.R. de Monmort, Essay d’analyse sur les jeux de hazard, 2nd edn. (J. Quillau, Paris, 1713)
66. L. Moser, M. Wyman, On solutions of x d D 1 in symmetric groups. Can. J. Math. 7, 159–168
(1955)
67. L. Moser, M. Wyman, Asymptotic expansions. Can. J. Math. 8, 225–233 (1956)
68. L. Moser, M. Wyman, Asymptotic expansions. II. Can. J. Math. 9, 194–209 (1957)
69. P.M. Neumann, C.E. Praeger, A recognition algorithm for special linear groups. Proc. Lond.
Math. Soc. (3) 65 (3), 555–603 (1992)
70. P.M. Neumann, C.E. Praeger, Cyclic matrices over finite fields. J. Lond. Math. Soc. (2) 52,
263–284 (1995)
71. A.C. Niemeyer, T. Popiel, C.E. Praeger, Abundant p-singular elements in finite classical
groups, preprint (2012) http://arxiv.org/abs/1205.1454v2
72. A.C. Niemeyer, C.E. Praeger, A recognition algorithm for classical groups over finite fields.
Proc. Lond. Math. Soc. (3) 77 (1), 117–169 (1998)
73. A.C. Niemeyer, C.E. Praeger, On the frequency of permutations containing a long cycle.
J. Algebra 300(1), 289–304 (2006)
74. A.C. Niemeyer, C.E. Praeger, On permutations of order dividing a given integer. J. Algebr.
Comb. 26(1), 125–142 (2007)
75. A.C. Niemeyer, C.E. Praeger, On the proportion of permutations of order a multiple of the
degree. J. Lond. Math. Soc. (2) 76(3), 622–632 (2007)
76. A.C. Niemeyer, C.E. Praeger, Estimating proportions of elements in finite groups of Lie type.
J. Algebra 324(1), 122–145 (2010)
77. E.A. O’Brien, Algorithms for Matrix Groups, in Groups St. Andrews 2009 in Bath, vol. 2.
London Mathematical Society Lecture Note Series, vol. 388 (Cambridge University Press,
Cambridge, 2011), pp. 297–323
78. C.W. Parker, R.A. Wilson, Recognising simplicity of black-box groups by constructing
involutions and their centralisers. J. Algebra 324(5), 885–915 (2010)
79. E.T. Parker, P.J. Nikolai, A search for analogues of the Mathieu groups. Math. Tables Aids
Comput. 12, 38–43 (1958)
80. A.I. Pavlov, An equation in a symmetric semigroup. Trudy Mat. Inst. Steklov. 177, 114–121,
208 (1986); Proc. Steklov Inst. Math. 1988(4), 121–129, Probabilistic problems of discrete
mathematics
81. A.I. Pavlov, On permutations with cycle lengths from a fixed set. Theor. Probab. Appl. 31,
618–619 (1986)
82. W. Plesken, D. Robertz, The average number of cycles. Arch. Math. (Basel) 93(5), 445–449
(2009)
83. C.E. Praeger, On elements of prime order in primitive permutation groups. J. Algebra 60(1),
126–157 (1979)
84. C.E. Praeger, Á. Seress, Probabilistic generation of finite classical groups in odd characteristic
by involutions. J. Group Theor. 14(4), 521–545 (2011)
85. C.E. Praeger, Á. Seress, Balanced involutions in the centralisers of involutions in finite general
linear groups of odd characteristic (in preparation)
86. C.E. Praeger, Á. Seress, Regular semisimple elements and involutions in finite general linear
groups of odd characteristic. Proc. Am. Math. Soc. 140, 3003–3015 (2012)
87. Á. Seress, Permutation Group Algorithms. Cambridge Tracts in Mathematics, vol. 152
(Cambridge University Press, Cambridge, 2003)
88. C.C. Sims, Computational Methods in the Study of Permutation Groups, in Computational
Problems in Abstract Algebra, Proceedings of the Conference, Oxford, 1967 (Pergamon,
Oxford, 1970), pp. 169–183
89. C.C. Sims, The Existence and Uniqueness of Lyons’ Group, in Finite groups ’72, Proceedings
of the Gainesville Conference, University of Florida, Gainesville, FL, 1972. North–Holland
Mathematical Studies, vol. 7 (North-Holland, Amsterdam, 1973), pp. 138–141
82 A.C. Niemeyer et al.

90. A.N. Timashev, Random permutations with cycle lengths in a given finite set. Diskret. Mat.
20(1), 25–37 (2008)
91. J. Touchard, Sur les cycles des substitutions. Acta Math. 70(1), 243–297 (1939)
92. L.M. Volynets, The number of solutions of the equation x s D e in a symmetric group. Mat.
Zametki 40(2), 155–160, 286 (1986)
93. R. Warlimont, Über die Anzahl der Lösungen von x n D 1 in der symmetrischen Gruppe Sn .
Arch. Math. (Basel) 30 (6), 591–594 (1978)
94. H. Wielandt, Finite Permutation Groups, Translated from the German by R. Bercov (Academic,
New York, 1964)
95. H.S. Wilf, The asymptotics of e P .z/ and the number of elements of each order in Sn . Bull. Am.
Math. Soc. (N.S.) 15 (2), 228–232 (1986)
96. H.S. Wilf, Generatingfunctionology, 2nd edn. (Academic, Boston, 1994)
97. K. Zsigmondy, Zur Theorie der Potenzreste. Monatsh. für Math. U. Phys. 3, 265–284 (1892)
Chapter 3
Designs, Groups and Computing

Leonard H. Soicher

3.1 Introduction

In this chapter we present some applications of groups and computing to the

discovery, construction, classification and analysis of combinatorial designs. The
focus is on certain block designs and their statistical efficiency measures, and in
particular semi-Latin squares, which are certain 1-designs with additional block
structure and which generalise Latin squares.
The chapter starts with background material on block designs, statistical effi-
ciency measures for 1-designs, permutation groups, Latin squares and semi-Latin
squares. We then review statistical optimality results for semi-Latin squares. Next,
we introduce the recent theory of “uniform” semi-Latin squares, which generalise
complete sets of mutually orthogonal Latin squares, and are statistically “Schur-
optimal”. We then describe a recent construction which determines a semi-Latin
square SLS.G/ from a transitive permutation group G, the square being uniform
precisely when G is 2-transitive. Moreover, we show how certain structural proper-
ties of SLS.G/ are determined from the structure of G, and (due to Martin Liebeck)
how statistical efficiency measures of SLS.G/ are determined from the degrees and
multiplicities of the irreducible constituents of the permutation character of G.
We then turn to computation, and discuss the DESIGN package [34] for
GAP [18], focussing on the function BlockDesigns, which can be used for
a wide variety of block design classifications using groups, and the function
BlockDesignEfficiency, for exact information on the efficiency measures of
a given 1-design. We then show how the BlockDesigns function can be applied
to obtain classifications of semi-Latin squares, and of uniform semi-Latin squares
and their subsquares, which can then be analysed using the BlockDesign-

L.H. Soicher ()

School of Mathematical Sciences, Queen Mary University of London, Mile End Road,
London E1 4NS, UK
e-mail: [email protected]

A. Detinko et al. (eds.), Probabilistic Group Theory, Combinatorics, and Computing, 83

Lecture Notes in Mathematics 2070, DOI 10.1007/978-1-4471-4814-2 3,
© Springer-Verlag London 2013
84 L.H. Soicher

Efficiency function. We give an extended example of this, including the

determination of the first published efficient .6  6/=k semi-Latin squares, for
k D 7; 8; 9. It is hoped that our examples of the use of the DESIGN package will
help the reader to use this package in their own investigations of designs. We
conclude the chapter with some open problems.

3.2 Background Material

3.2.1 Block Designs

A block design is an ordered pair .V; B/, such that V is a finite non-empty set of
points, and B is a (disjoint from V ) finite multiset (or collection) of non-empty
subsets of V called blocks, such that every point is in at least one block.
In a multiset (of blocks say), order does not matter, but the number of times
an element occurs (its multiplicity) does indeed matter. We denote a multiset
with elements A1 ; : : : ; Ab (including any repeated elements) by ŒA1 ; : : : ; Ab . For
example, the block design .V; B/, with V D f1; 2; 3g and

B D Œf1; 2g; f1; 2; 3g; f1; 2; 3g; f1; 3g; f2; 3g;

has three points and five blocks.

Let t be a non-negative integer. A t-design, or more specifically a t-.v; k; /
design, is a block design .V; B/ such that v D jV j, each block has the same size k,
and each t-subset of V is contained in the same positive number  of blocks. It is
well known that a t-design is also an s-design, for s D 0; : : : ; t 1 (see, for example,
[23, Theorem 19.3]). For example, the block design .V; B/, with V D f1; : : : ; 7g and

B D Œf1; 2; 3g; f1; 4; 5g; f1; 6; 7g; f2; 4; 7g; f2; 5; 6g; f3; 4; 6g; f3; 5; 7g;

is a 2-.7; 3; 1/ design, and also a 1-.7; 3; 3/ design and a 0-.7; 3; 7/ design.

The dual of a block design is obtained by interchanging the roles of points and
blocks. More precisely, if  D .f˛1 ; : : : ; ˛v g; ŒA1 ; : : : ; Ab / is a block design, then
the dual  of  is the block design .f1; : : : ; bg, ŒB1 ; : : : ; Bv /, with Bi D fj W ˛i 2
Aj g. Note that if  is a 1-.v; k; r/ design then  is a 1-.vr=k; r; k/ design.
Two block designs 1 D .V1 ; B1 / and 2 D .V2 ; B2 / are isomorphic if there is
a bijection from V1 to V2 that maps B1 to B2 . The set of all isomorphisms from a
block design  to itself forms a group, the automorphism group Aut./ of .
Block designs are of interest to both pure mathematicians and statisticians. They
are used by statisticians for the design of comparative experiments: the points
represent “treatments” to be compared, and the blocks represent homogeneous
testing material, so usually only within-blocks information is used in the comparison
of treatment effects. Although statisticians sometimes allow points to be repeated
within a block, here we do not.
3 Designs, Groups and Computing 85

3.2.2 Efficiency Measures of 1-Designs

We now discuss certain statistical efficiency measures of 1-designs. The basic idea
is that once you have decided that your experimental design needs to be in a certain
class C of 1-.v; k; r/ designs, you want to choose a design in C able to give as
much information as possible; that is, the most “efficient” design in C with respect
to one or more of the efficiency measures defined below. The reader who wants to
learn more about statistical design theory and the theory of optimal designs should
consult the excellent survey article [6], which was written for combinatorialists.
Other useful references for these topics include [3, 4, 11, 29].
Let  be a 1-.v; k; r/ design, with v  2. The concurrence matrix of  is the vv
matrix whose rows and columns are indexed by the points, and whose .˛; ˇ/-entry
is the number of blocks containing both points ˛ and ˇ. The scaled information
matrix of  is
F ./ WD Iv  .rk/1 ;
where Iv is the v  v identity matrix and  is the concurrence matrix of . The
matrix F ./ is real, symmetric and positive semi-definite, and is scaled so that its
(all real) eigenvalues lie in the interval Œ0; 1. Moreover, F ./ has constant row-sum
0, so the all-1 vector is an eigenvector with corresponding eigenvalue 0. It can be
shown that the remaining eigenvalues are all non-zero if and only if  is connected
(i.e. its point-block incidence graph is connected). Omitting the zero eigenvalue
corresponding to the all-1 vector, the eigenvalues

ı1  ı2      ıv1

of F ./ are called the canonical efficiency factors of . Note that the canonical
efficiency factors of two isomorphic 1-designs are the same.
If  is not connected, then we define A D D D E WD 0. Otherwise, we
define these efficiency measures by

X
v1
A WD .v  1/= 1=ıi ;
i D1
!1=.v1/
Y
v1
D WD ıi ;
i D1

E WD ı1 D minfı1 ; : : : ; ıv1 g:

Note that A (respectively D ) is the harmonic mean (respectively geometric mean)

of the canonical efficiency factors of .
If k D v (so each block contains every point) then each canonical efficiency
factor of  is equal to 1 and so is each of the efficiency measures above. If k < v, we
want to minimize the loss of “information” due to being forced to use “incomplete”
blocks, and want the above efficiency measures to be as close to 1 as possible.
86 L.H. Soicher

We now define optimality in a class of 1-designs with respect to a given efficiency

measure. The 1-.v; k; r/ design  is A-optimal in a class C of 1-.v; k; r/ designs
containing  if A  A for each 2 C . D-optimal and E-optimal are defined
similarly. We say  is Schur-optimal in a class C of 1-.v; k; r/ designs containing
 if for each design 2 C , with canonical efficiency factors 1      v1 , we
have
X̀ X̀
ıi  i;
i D1 i D1

for ` D 1; : : : ; v  1. A Schur-optimal design need not exist within a given class

C of 1-.v; k; r/ designs, but when it does, that design is optimal in C with respect
to a very wide range of statistical optimality criteria, including being A-, D- and
E-optimal [19].
It is not difficult to see that if  is a 2-.v; k; / design, then the canonical
efficiency factors of  are all equal (to v.k  1/=..v  1/k/), from which it
follows that  is Schur-optimal in the class of all 1-.v; k; .v  1/=.k  1// designs.
However, a 2-design may well not exist with the properties we are interested in.
The canonical efficiency factors of a 1-design  not equal to 1, and their
multiplicities, are the same as those of the dual block design  of  (see [6,
Sect. 3.1.1]). We thus obtain the following:
Theorem 3.1. A 1-.v; k; r/ design  is A-optimal (respectively D-optimal, E-
optimal, Schur-optimal) in a class C of 1-.v; k; r/ designs if and only if  is
A-optimal (respectively D-optimal, E-optimal, Schur-optimal) in the class consisting
of the dual block designs of the designs in C .

3.2.3 Permutation Groups

We now review some basic definitions and results in the theory of permutation
groups and group actions. An excellent reference for permutation groups is [10].
A permutation group G on a set V of points is a subgroup of the group Sym(V )
of all permutations of V , with the group operation being composition. The image of
˛ 2 V under g 2 G is denoted ˛g (our permutations act on the right). The degree
of G is the size of V . The symmetric group of degree n, denoted Sn , is the group
Sym(f1; : : : ; ng/ of all permutations of f1; : : : ; ng.
An action of a group G on a set V is a function W V  G ! V , with .˛; g/
denoted ˛ g , such that ˛ 1 D ˛ and .˛ g /h D ˛ gh , for all ˛ 2 V and all g; h 2 G.
Given an action of G on V , we say that G acts on V .
If G is a permutation group on V then G acts naturally on V , where ˛ g WD ˛g.
An action of G on V gives rise to a homomorphism W G ! Sym.V / defined by
˛.g / WD ˛ g for all ˛ 2 V and g 2 G.
3 Designs, Groups and Computing 87

Suppose G acts on V and ˛ 2 V . The G-orbit of ˛ is ˛ G WD f˛ g W g 2 Gg.

The set f˛ G W ˛ 2 V g of all G-orbits is a partition of V . The stabilizer in G of ˛ is
G˛ WD fg 2 G W ˛ g D ˛g. This stabilizer G˛ is a subgroup of G, and if G is finite
then jGj D jG˛ jj˛ G j.
Let G act on a finite set V . Then G has actions on many sets. Here are some:
• G acts on the set of all t-tuples of distinct elements of V , where .˛1 ; : : : ; ˛t /g WD
g g
.˛1 ; : : : ; ˛t /.
g g
• G acts on the set of all subsets of V , where f˛1 ; : : : ; ˛k gg WD f˛1 ; : : : ; ˛k g.
• G acts on the set of all finite multisets of subsets of V , where ŒA1 ; : : : ; Ab g WD
Œ.A1 /g ; : : : ; .Ab /g .
• G acts on the set of all block designs with point set V , where .V; B/g WD
.V; B g /.
A permutation group G on a non-empty set V is transitive if for every ˛; ˇ 2 V
there is a g 2 G with ˛g D ˇ (i.e. there is just one G-orbit in the natural action of
G on V ). More generally, a permutation group G on a set V of size at least t is
t-transitive if for every pair .˛1 ; : : : ; ˛t /, .ˇ1 ; : : : ; ˇt / of t-tuples of distinct
elements of V , there is a g 2 G with .˛1 g; : : : ; ˛t g/ D .ˇ1 ; : : : ; ˇt /. It is easy
to see that if G is a t-transitive permutation group on a finite set V of size v  t,
and B is a subset of V of size k  t, then the block design with point  V and
  set
block (multi)set the G-orbit B G is a t-.v; k; / design, with  D jB G j kt = vt .

3.2.4 Latin Squares

A Latin square of order n is an n  n array L, whose entries are elements of an n-set

˝, the set of symbols for L, such that each symbol occurs exactly once in each row
and exactly once in each column of L. For example, a completed Sudoku puzzle is
a special kind of Latin square of order 9.
Two Latin squares L1 and L2 of order n, with respective symbol sets ˝1 and ˝2 ,
are orthogonal if for every ˛1 2 ˝1 and ˛2 2 ˝2 , there is an .i; j / such that ˛1
is the .i; j /-entry in L1 and ˛2 is the .i; j /-entry in L2 . For example, here are two
orthogonal Latin squares of order 3:

1 2 3 4 5 6
3 1 2 5 6 4
2 3 1 6 4 5

Latin squares L1 ; : : : ; Lm of order n are said to be mutually orthogonal if they

are pairwise orthogonal, in which case fL1 ; : : : ; Lm g is called a set of mutually
orthogonal Latin squares or a set of MOLS.
88 L.H. Soicher

Let n > 1. A set of MOLS of order n has size at most n  1, and the existence
of a set of MOLS of order n having size n  1 (called a complete set of MOLS) is
equivalent to the existence of a projective plane of order n. A complete set of MOLS
of order n exists when n is a prime power, but it is a famous problem as to whether
a complete set of MOLS of order n exists for some non-prime-power n.

3.2.5 Semi-Latin Squares

An .n  n/=k semi-Latin square is an n  n array S , whose entries are k-subsets

of an nk-set ˝, the set of symbols for S , such that each symbol is in exactly one
entry in each row and exactly one entry in each column of S . The entry in row
i and column j is called the .i; j /-entry of S and is denoted by S.i; j /. To avoid
trivialities, we assume throughout that n > 1, k > 0. Note that an .n  n/=1 semi-
Latin square is (essentially) the same thing as a Latin square of order n. We consider
two .n  n/=k semi-Latin squares to be isomorphic if one can be obtained from
the other by applying one or more of: a row permutation, a column permutation,
transposing, and renaming symbols.
Semi-Latin squares have many applications, including the design of agricultural
experiments, consumer testing, and message authentication (see [1, 2, 5, 17, 26]).
Semi-Latin squares exist in profusion, and a good choice of semi-Latin square for a
given application can be very important.
Let s be a positive integer. An s-fold inflation of an .n  n/=k semi-Latin square
is obtained by replacing each symbol ˛ in the semi-Latin square by s symbols
˛;1 ; : : : ; ˛;s , such that ˛;i D ˇ;j if and only if ˛ D ˇ and i D j . The result
is an .n  n/=.ks/ semi-Latin square. For example, here is a .3  3/=2 semi-Latin
square formed by a twofold inflation of a Latin square of order 3:

14 25 36
36 14 25
25 36 14

The superposition of an .n  n/=k semi-Latin square with an .n  n/=` semi-

Latin square (with disjoint symbol sets) is performed by superimposing the first
square upon the second, resulting in an .n  n/=.k C `/ semi-Latin square. For
example, here is a .3  3/=2 semi-Latin square which is the superposition of two
(mutually orthogonal) Latin squares of order 3:

14 25 36
35 16 24 (3.1)
26 34 15
3 Designs, Groups and Computing 89

An .n  n/=k semi-Latin square in which any two distinct symbols occur

together in at most one block is called a SOMA.k; n/ (SOMA is an acronym for
“simple orthogonal multi-array” [25]). For example, the semi-Latin square (3.1) is
a SOMA.2; 3/, and more generally, the superposition of k MOLS of order n is a
SOMA.k; n/. However, a SOMA.k; n/ need not be a superposition of MOLS, and
may exist even when there do not exist k MOLS of order n. However, it is easy to
see that if a SOMA.k; n/ exists then k < n.
The dual S  of an .n  n/=k semi-Latin square S is a block design with point set
f1; : : : ; ng2 and nk blocks, one for each symbol of S , with the block for a symbol
˛ consisting precisely of the ordered pairs .i; j / such that ˛ 2 S.i; j /. Each block
of S  is of the form Œ.1; 1g /; : : : ; .n; ng / for some g 2 Sn , and S  is a 1-.n2 ; n; k/
design, which may have repeated blocks. Up to the naming of its symbols, the semi-
Latin square S can be recovered from S  , so S  really represents the class of semi-
Latin squares obtainable from S by renaming symbols. For example, let S be the
semi-Latin square (3.1). Then S  D .V; B/, with V D f1; 2; 3g2 and

[f(1,1),(2,2),(3,3)g, f(1,2),(2,3),(3,1)g,
BD f(1,3),(2,1),(3,2)g, f(1,1),(2,3),(3,2)g,
f(1,2),(2,1),(3,3)g, f(1,3),(2,2),(3,1)g].

3.3 Optimality Results for Semi-Latin Squares

Let S be an .n  n/=k semi-Latin square. If we ignore the row and column structure
of S , we obtain its underlying block design .S /, whose points are the symbols of
S and whose blocks are the entries of S . Note that .S / is a 1-.nk; k; n/ design, and
that the dual S  of S is isomorphic as a block design to the dual of .S /. However,
S  has a structured point-set, unlike .S / .
Following the analysis of Bailey [2], S is optimal with respect to a given
statistical optimality criterion if and only if .S / is optimal with respect to that
criterion in the class of underlying block designs of .n  n/=k semi-Latin squares.
For this reason, we say that S has canonical efficiency factors, or a given efficiency
measure, when .S / has those canonical efficiency factors, or that efficiency
measure.
Various optimality results for .n  n/=k semi-Latin squares are known. These
include:
• Cheng and Bailey [13] proved that a superposition of k MOLS of order n (with
pairwise disjoint symbol sets) is A-, D- and E-optimal.
• Bailey [2] proved that for all s  1, an s-fold inflation of a superposition of n  1
MOLS of order n is A-, D- and E-optimal.
• Bailey [2] classified the 3  3 semi-Latin squares and determined the ones that
are A-, D- and E-optimal.
90 L.H. Soicher

• Chigbu [14] determined the .4  4/=4 semi-Latin squares that are A-, D- and
E-optimal.
• Using the computational methods described in this chapter, Soicher [37] classi-
fied the .4  4/=k semi-Latin squares for k D 5; : : : ; 10, and determined those
that are A-, D- and E-optimal.
It is widely believed that when a SOMA.k; n/ exists, then one that is optimal in
the class of all SOMA.k; n/s is in fact optimal in the class of all .n  n/=k semi-
Latin squares. There are not even two MOLS of order 6, but there are SOMA.2; 6/s
and SOMA.3; 6/s. These have been classified and the best with respect to various
measures of efficiency have been determined (see [5, 8, 25, 27, 31, 37]). There is no
SOMA.k; 6/ with k > 3. It is not known whether there are three MOLS of order
10, but there are SOMA.3; 10/s and SOMA.4; 10/s (see [5, 30, 31]).

3.4 Uniform Semi-Latin Squares

We now introduce the concept of uniform semi-Latin squares, and some results
about such squares from [36]. Uniform semi-Latin squares provide Schur-optimal
semi-Latin squares of many sizes for which no optimal semi-Latin square was
previously known for any optimality criterion.
An .n  n/=k semi-Latin square S is uniform if every two entries of S , not in
the same row or column, intersect in a constant number  D .S / of points. For
example, the semi-Latin square (3.1) is uniform, with  D 1.
Note that if S is a uniform semi-Latin square then an s-fold inflation of S is also
uniform, and if S and T are both n  n uniform semi-Latin squares (with disjoint
symbol sets) then the superposition of S and T is also uniform.
Lemma 3.2 ([36]). If S is a uniform .n  n/=k semi-Latin square then

.S / D k=.n  1/;

and in particular, n  1 divides k.

Theorem 3.3 ([36]). An .n  n/=.n  1/ semi-Latin square S is uniform if and only
if S is a superposition of n  1 MOLS of order n.
Uniform semi-Latin squares can be seen as generalising the concept of complete
sets of MOLS. Since the -fold inflation of a uniform semi-Latin square is uniform,
we see that the existence of a uniform .n  n/=..n  1// semi-Latin square for all
integers  > 0 is equivalent to the existence of a complete set of MOLS of order
n, and such a complete set exists when n is a prime power. Although there are not
even two MOLS of order 6, the following is proved in [36].
Theorem 3.4 ([36]). There exist uniform .6  6/=.5/ semi-Latin squares for all
integers  > 1.
3 Designs, Groups and Computing 91

The statistical importance of uniform semi-Latin squares is due to the following:

Theorem 3.5 ([36]). Suppose that S is a uniform .n  n/=k semi-Latin square.
Then S is Schur-optimal; that is, .S / is Schur-optimal in the class of underlying
block designs of .n  n/=k semi-Latin squares.
Proof. We give an outline of the proof. Details can be found in [36].
• The dual S  of S is a partially balanced incomplete-block design with respect
to the L2 -type association scheme (see [3]), so we may easily determine the
eigenvalues of the concurrence matrix of S  (see [39]), and hence the canonical
efficiency factors of S  .
• These canonical efficiency factors are 1  1=.n  1/, with multiplicity .n  1/2 ,
and 1, with multiplicity 2.n  1/.
• The dual T  of any .nn/=k semi-Latin square T has at most .n1/2 canonical
efficiency factors not equal to 1.
• It follows that S  is Schur-optimal in the class of duals of .n  n/=k semi-Latin
squares, and so .S / is Schur-optimal in the class of underlying block designs
of .n  n/=k semi-Latin squares. t
u

3.5 Semi-Latin Squares from Transitive Permutation Groups

In [36], a simple construction is given which produces a semi-Latin square SLS.G/

from a transitive permutation group G, and in this section we discuss how properties
of G determine properties of SLS.G/.
Let G be a transitive permutation group on f1; : : : ; ng, with n > 1. For all i; j 2
f1; : : : ; ng, there are exactly jGj=n elements of G mapping i to j (the elements
mapping i to j are precisely those in the right coset Gi g, where Gi is the stabilizer
in G of i and g is any element of G with ig D j ). Thus G defines a semi-Latin
square, as follows. The .n  n/=k semi-Latin square SLS(G), with k WD jGj=n, has
symbol set G itself, and the symbol g is in the .i; j /-entry of SLS.G/ if and only if
ig D j . For example,

1 (23) (12) (123) (13) (132)

SLS.S3 / D (12) (132) 1 (13) (23) (123) :
(13) (123) (23) (132) 1 (12)

Theorem 3.6 ([36]). Let G be a transitive permutation group on f1; : : : ; ng, with
n > 1, and let S WD SLS.G/. Then:
• Let H be a transitive subgroup of G of index m. Then S is a superposition of m
semi-Latin squares, each isomorphic to SLS.H / (this comes from the partition of
92 L.H. Soicher

G into the right (or left) cosets of H ). In particular, if H has order n then S is a
superposition of jGj=n isomorphic Latin squares.
• G contains a non-identity element with exactly f fixed points if and only if there
are two distinct symbols of S which occur together in exactly f entries of S .
• G is a Frobenius group (that is, a transitive permutation group in which no non-
identity element fixes more than one point) if and only if S is a superposition of
MOLS.
• .S / is connected if and only if G has no normal subgroup N satisfying G1 
N 6D G.
• The automorphism group of S (i.e. the group of all isomorphisms from S to S )
has structure
.G  G/:..NSn .G/=G/  C2 /
(in ATLAS [15] notation).
• G is 2-transitive if and only if S is uniform.
Proof. See [36] for the proofs of these assertions. We only repeat the (easy) proof
of the important last statement. Also see [38].
Suppose G is 2-transitive. Then for every i; i 0 ; j; j 0 2 f1; : : : ; ng with i 6D i 0 and
j 6D j 0 , there are precisely  WD jGj=.n.n  1// elements g 2 G with ig D j and
i 0 g D j 0 . Thus, S.i; j / and S.i 0 ; j 0 / intersect in exactly these  elements, and so
S is uniform.
Conversely, suppose S is uniform. Then if i; i 0 ; j; j 0 2 f1; : : : ; ng with i 6D i 0
and j 6D j 0 , then S.i; j / and S.i 0 ; j 0 / intersect in  WD k=.n  1/ > 0 symbols
(recall that n > 1, k > 0), so there is an element of G mapping i to j and i 0 to j 0 .
Thus G is 2-transitive. t
u
Using the Classification of Finite Simple Groups, all the finite 2-transitive
permutation groups have been classified, and tables of these groups are given in
[10]. Each 2-transitive group G gives rise to a uniform semi-Latin square SLS.G/,
certain properties of which can be deduced from properties of G. For example,
consideration of the groups P GL2 .q/ and PSL2 .q/, of degree q C 1 where q is a
prime power, yields the following result.
Theorem 3.7 ([36]). Let q be a prime power. Then there exists a uniform, and
hence Schur-optimal, ..q C 1/  .q C 1//=.q.q  1// semi-Latin square S which
is a superposition of isomorphic Latin squares and in which every two distinct
symbols occur together in at most two entries. Moreover, if q is odd then S is also a
superposition of two isomorphic uniform ..qC1/.qC1//=.q.q1/=2/ semi-Latin
squares.

3.5.1 The Canonical Efficiency Factors of SLS.G /

Let G be a transitive permutation group on f1; : : : ; ng, with n > 1, and let  be
the concurrence matrix of the underlying block design of SLS.G/. Then  is a
3 Designs, Groups and Computing 93

jGj  jGj matrix whose rows and columns are indexed by the elements of G and
whose .g; h/-entry is the number of fixed points of g1 h, which is .g 1 h/, where
 is the permutation character of G. Applying this observation, Martin Liebeck
(at the Fifth de Brún Workshop itself) discovered and proved the theorem below.
The statement of the theorem and its proof use basic representation theory of finite
groups over the complex numbers, such as can be found in [20].
Theorem 3.8 (M. Liebeck). Let G be a transitive permutation group of degree
n > 1 with permutation character . Then the canonical efficiency factors of (the
underlying block design of) SLS.G/ are

1  h ; i= .1/;

repeated .1/2 times, where runs over the non-trivial complex irreducible
characters of G.
Proof. Suppose jGj D nk, and let GLnk .C/ be the group of all invertible nk  nk
matrices over the complex numbers, whose rows and columns are indexed by the
elements of G, and let W G ! GLnk .C/ be defined by .x/g;h D 1 if gx D h and
.x/g;h D 0 otherwise. In other words, is the right-regular matrix representation
of G, with natural G-module V WD Cnk .
Now let C0 D f1g; C1; : : : ; Cd be the conjugacy classes of G, with respective
representatives c0 ; : : : ; cd , let 0 ; : : : ; d be the complex irreducible characters of
G, with 0 the trivial character, and let
X
Ai WD .c/:
c2Ci

Then the matrices Ai commute pairwise, and V decomposes into a direct sum of
common eigenspaces V0 ; : : : ; Vd of A0 ; : : : ; Ad , with Vj being a G-submodule of
V isomorphic to the direct sum of j .1/ copies of the irreducible G-module with
character j . In particular Vj has dimension j .1/2 , and Vj is an eigenspace for Ai
with corresponding eigenvalue

jCi j j .ci /= j .1/:

Note that V0 is spanned by the all-1 vector.

Now Ai is a .0; 1/-matrix, with .g; h/-entry equal to 1 if and only if g 1 h 2 Ci ,
and so
Xd
D .ci /Ai :
i D0

Thus, for j D 0; : : : ; d ,  has eigenvalue

94 L.H. Soicher

X
d X
1
.ci /jCi j j .ci /= j .1/ D j .g/.g /= j .1/ D jGjh j ; i= j .1/;
i D0 g2G

repeated j .1/2 times, where h ; i denotes the standard inner product of characters
(so h j ; i is the multiplicity of the irreducible character j as a constituent of the
permutation character ).
The scaled information matrix of the underlying block design  of SLS.G/ is

F ./ WD Ink  .nk/1 ;

and so the canonical efficiency factors of SLS.G/ are

1h j ; i= j .1/;

2
repeated j .1/ times, for j D 1; : : : ; d . t
u
Thus, the canonical efficiency factors of SLS.G/, and its A-, D- and E-efficiency
measures can be determined from the degrees and multiplicities of the irreducible
constituents of the permutation character of G. This has been done by Eamonn
O’Brien (using MAGMA [9]) for all transitive permutation groups of degree  23,
and by Soicher (using GAP) for the primitive permutation groups of non-prime-
power degree n  500 and with order  n.n  1/.
As an illustrative example, let G be the group A5 in its primitive permutation
representation of degree 10. The permutation character of G (in ATLAS [15]
notation) decomposes as 1a C 4a C 5a, so the canonical efficiency factors of
SLS.G/ are 3/4 (with multiplicity 16), 4/5 (with multiplicity 25), and 1 (with
multiplicity 18). If six MOLS of order 10 exist, then their superposition S would
have canonical efficiency factors 5/6 (with multiplicity 54) and 1 (with multiplicity
5); see [2, Corollary 5.2]. The ratios of the A-, D- and E-efficiency measures of
SLS.G/ with those of S are respectively approximately 0:9889, 0:9943 and 0:9, so
in the (likely) absence of six MOLS of order 10, SLS.G/ provides a highly efficient
(and possibly optimal) .10  10/=6 semi-Latin square.

3.6 The DESIGN Package

GAP [18] is an internationally developed, freely available, Open Source system for
Computational Group Theory and related areas in algebra and combinatorics. The
DESIGN package [34] is a refereed and officially accepted GAP package which
provides functionality for constructing, classifying, partitioning and analysing block
designs.
In this section we focus on the DESIGN package functions BlockDesigns,
used to classify block designs, and BlockDesignEfficiency, used to
3 Designs, Groups and Computing 95

determine efficiency measures of 1-designs. There are many other functions in

the DESIGN package to construct and analyse block designs, including isomor-
phism testing and automorphism group computation, and functions to determine
information about t-designs from their parameters. The best reference for all this,
and for precise details on the parameters for the functions discussed here, is the
DESIGN package documentation, which includes many examples. The details of
the techniques used in the DESIGN package can be found in its documented Open
Source code.

3.6.1 The BlockDesigns Function

The most important DESIGN package function is BlockDesigns, which can

construct and classify block designs satisfying a wide range of user-specified
properties. The properties which must be specified are:
• The number v of points (the point set is then f1; : : : ; vg, although the points may
also be given names).
• The possible block sizes.
• For a given t, for each t-subset T of the points, the number of blocks containing
T (this number may depend on T ).
The properties which may optionally be additionally specified are:
• The maximum multiplicity of a block, for each possible block-size.
• The total number b of blocks.
• The block-size distribution.
• A replication number r (that is, specifying that every point is in exactly r blocks).
• The possible sizes of intersections of pairs of blocks of given sizes.
• A subgroup G of Sv , to specify that G-orbits of block designs are isomorphism
classes (default: G D Sv , giving the usual notion of isomorphism).
• A subgroup H of G, such that H is required to be a subgroup of the
automorphism group of each returned design (default: H D f1g, but specifying
a non-trivial H can be a very powerful constraint; see [16, 21, 32, 33]).
• Whether the user wants a single design with the specified properties (if one
exists), a list of G-orbit representatives of all such designs (i.e. isomorphism
class representatives as determined by G; this is the default), or a list of distinct
such designs containing at least one representative from each G-orbit.
The BlockDesigns function works by transforming the design classification
problem into a problem of classifying cliques with a given vertex-weight sum in a
certain graph whose vertices are “weighted” with non-zero vectors of non-negative
integers. Each vertex of this graph represents a possible H -orbit of blocks, each
with the same specified multiplicity, with two distinct vertices not joined by an
edge only when the totality of the blocks they represent cannot be a submultiset
of the blocks of a required design. Such a non-edge may be a result of user-
specified properties of the required designs, or may be determined by applying
96 L.H. Soicher

block intersection polynomials [12, 33]. The graph problem is then handled by
the GRAPE [35] function CompleteSubgraphsOfGivenSize, which uses
a complicated backtrack search. The reader may wish to consult the reference [21],
which gives detailed information on techniques used to classify block designs.
We now give some straightforward examples of the use of the BlockDesigns
function. A more complicated example will follow in Sect. 3.7. Note that first we
load the DESIGN package, which also loads the GRAPE package for graphs and
groups, which is heavily used by the DESIGN package.
gap> LoadPackage("design");
--------------------------------------------------------------
Loading GRAPE 4.5 (GRaph Algorithms using PErmutation groups)
by Leonard H. Soicher (http://www.maths.qmul.ac.uk/˜leonard/).
Homepage: http://www.maths.qmul.ac.uk/˜leonard/grape/
--------------------------------------------------------------
--------------------------------------------------------------
Loading DESIGN 1.6 (The Design Package for GAP)
by Leonard H. Soicher (http://www.maths.qmul.ac.uk/˜leonard/).
Homepage: http://www.designtheory.org/software/gap_design/
--------------------------------------------------------------
true

We now classify the 2-.7; 3; 1/ designs.

gap> designs:=BlockDesigns(rec( v:=7, blockSizes:=[3],
> tSubsetStructure:=rec(t:=2, lambdas:=[1] ) ) );
[ rec( autGroup := Group([ (1,2)(5,7), (1,2,3)(5,7,6),
(1,2,3)(4,7,5), (1,5,3)(2,4,7) ]),
blockNumbers := [ 7 ], blockSizes := [ 3 ],
blocks := [ [ 1, 2, 3 ], [ 1, 4, 5 ], [ 1, 6, 7 ],
[ 2, 4, 7 ], [ 2, 5, 6 ], [ 3, 4, 6 ], [ 3, 5, 7 ] ],
isBinary := true, isBlockDesign := true, isSimple := true,
r := 3, tSubsetStructure := rec( lambdas := [ 1 ], t := 2 ),
v := 7 ) ]

There is, as is very well known, just one such design up to isomorphism. Note
that GAP has printed the value assigned to the variable designs. This output is
a list containing exactly one block design, in DESIGN package format, stored as a
GAP record with properties stored as record components. This output could have
been suppressed by ending the assignment statement with “;;” instead of “;”.
We next classify (but do not display) the 1-.7; 3; 3/ designs having no repeated
block and invariant under the group generated by .1; 2/.3; 4/.
gap> onedesigns:=BlockDesigns(rec( v:=7, blockSizes:=[3],
> blockMaxMultiplicities:=[1],
> requiredAutSubgroup:=Group((1,2)(3,4)),
> tSubsetStructure:=rec(t:=1, lambdas:=[3] ) ) );;
gap> List(onedesigns,AllTDesignLambdas);
[ [ 7, 3 ], [ 7, 3 ], [ 7, 3 ], [ 7, 3 ], [ 7, 3 ],
[ 7, 3, 1 ] ]
gap> List(onedesigns,d->Size(AutomorphismGroup(d)));
[ 48, 4, 6, 4, 8, 168 ]
3 Designs, Groups and Computing 97

For a more serious calculation, used in [24], we classify the block designs having
11 points, such that each block has size 4 or 5, and every pair of distinct points is
contained in exactly two blocks. This calculation takes about 220 s of CPU-time on
a 3.1 GHz PC running Linux.
gap> designs:=BlockDesigns(rec(v:=11, blockSizes:=[4,5],
> tSubsetStructure:=rec(t:=2, lambdas:=[2] ) ) );;
gap> List(designs,BlockSizes);
[ [ 5 ], [ 4, 5 ], [ 4, 5 ], [ 4, 5 ], [ 4, 5 ] ]
gap> List(designs,BlockNumbers);
[ [ 11 ], [ 10, 5 ], [ 10, 5 ], [ 10, 5 ], [ 10, 5 ] ]
gap> List(designs,d->Size(AutomorphismGroup(d)));
[ 660, 6, 8, 12, 120 ]

More generally, BlockDesigns can construct subdesigns of a given block

design, such that the subdesigns each have the same user-specified properties. Here,
a subdesign of  means a block design with the same point set as  and whose
block multiset is a submultiset of the blocks of . In this case, the default G
determining isomorphism is Aut./. For example, given a block design  having
v points and each of whose blocks has size k, we classify the “parallel classes” of
 by classifying the subdesigns of  that are 1-.v; k; 1/ designs (up to the action of
Aut./).
A DESIGN package function closely related to BlockDesigns is the function
PartitionsIntoBlockDesigns, which classifies the partitions of (the block
multiset of) a given block design , such that the subdesigns of  whose block
multisets are the parts of this partition each have the same user-specified properties.
For example, given a block design  having v points and each of whose blocks has
size k, we classify the “resolutions” of  by classifying the partitions of  into
1-.v; k; 1/ subdesigns (up to the action of Aut./).

3.6.2 The BlockDesignEfficiency Function

To test conjectures and rank designs we have classified, we need to be able to

compare efficiency measures exactly. We can do this using algebraic computation
in GAP, as described in [37], and this functionality is included in the most recent
release [34] of the DESIGN package.
If delta is a 1-.v; k; r/ design (in DESIGN package format) with v > 1, and
eps is a positive rational number, then, in DESIGN 1.6, the function call
BlockDesignEfficiency(delta,eps)
returns a GAP record eff (say) having the following components. The component
eff.A contains the rational number which is the A-efficiency measure of delta,
eff.Dpowered contains the rational number which is the D-efficiency measure
of delta raised to the power v  1, and eff.Einterval is a list Œa; b of non-
negative rational numbers such that if E is the E-efficiency measure of delta then
a  E  b, b  a  eps, and if E is rational then a D E D b. In addition,
98 L.H. Soicher

the component eff.CEFpolynomial contains the monic polynomial over the

rationals whose zeros (counting multiplicities) are the canonical efficiency factors
of the design delta.
For example, we calculate the block design efficiency record for one of the
1-.7; 3; 3/ designs classified above.
gap> eps:=10ˆ(-8);;
gap> delta:=onedesigns[1];
rec( allTDesignLambdas := [ 7, 3 ],
autGroup := Group([ (1,2), (6,7), (4,5)(6,7),
(3,4,5)(6,7), (1,6,2,7)(4,5) ]),
blockNumbers := [ 7 ], blockSizes := [ 3 ],
blocks := [ [ 1, 2, 3 ], [ 1, 2, 4 ], [ 1, 2, 5 ],
[ 3, 4, 5 ], [ 3, 6, 7 ], [ 4, 6, 7 ], [ 5, 6, 7 ] ],
isBinary := true, isBlockDesign := true,
isSimple := true, r := 3,
tSubsetStructure := rec( lambdas := [ 3 ], t := 1 ),
v := 7 )
gap> eff:=BlockDesignEfficiency(delta,eps);;
gap> eff.A;
21/31
gap> eff.Dpowered;
343/2187
gap> eff.Einterval;
[ 1/3, 1/3 ]
gap> Factors(eff.CEFpolynomial);
[ x_1-1, x_1-1, x_1-7/9, x_1-7/9, x_1-7/9, x_1-1/3 ]

3.7 Classifying Semi-Latin Squares

We now describe how to classify semi-Latin squares via their duals. Our approach,
which is somewhat similar to that of Bailey and Chigbu [7], is implemented in the
DESIGN package function SemiLatinSquareDuals, but can be applied with
more flexibility using the function BlockDesigns.
The group Wn below will be used to define isomorphism of semi-Latin squares,
via their duals. We define

Wn WD hSn  Sn ;  j  2 D 1; .a; b/ D .b; a/ for all a; b 2 Sn i:

Thus Wn is isomorphic to the wreath product Sn o C2 . If g 2 Wn then g D .a; b/ or

g D .a; b/, for some a; b 2 Sn , and Wn acts on V WD f1; : : : ; ng2 as follows. For
.i; j / 2 V and a; b 2 Sn :

.i; j /.a;b/ WD .i a ; j b /I
.i; j /.a;b/ WD .j b ; i a /:
3 Designs, Groups and Computing 99

Now let S be an .n  n/=k semi-Latin square, let S  D .V; B/ be the dual of

S , and let g 2 Wn . Define .S  /g WD .V; B g / D .V; ŒB g W B 2 B/. Then .S  /g
is the dual of a semi-Latin square T isomorphic to S . Indeed, if g D .a; b/ then
T can be obtained from S by permuting its rows by a and its columns by b, and if
g D .a; b/ then T is obtained from S by permuting its rows by a, its columns by b,
and then transposing. Conversely, suppose S and T are isomorphic .n  n/=k semi-
Latin squares, with respective duals S  and T  . Then T can be obtained from S by
applying some row permutation a, some column permutation b, followed possibly
by transposing and/or renaming symbols. Then .S  /.a;b/ D T  if transposing does
not take place, and otherwise .S  /.a;b/ D T  . We thus have an action of Wn on
the set of duals of .n  n/=k semi-Latin squares, and the duals of two .n  n/=k
semi-Latin squares X and Y are in the same Wn -orbit if and only if X and Y are
isomorphic (as semi-Latin squares); see also [37].
The following GAP function, to be used later, returns a homomorphism from the
imprimitive wreath product Sn o C2 with block system ff1; : : : ; ng; fn C 1; : : : ; 2ngg
onto the group Wn as a permutation group in its action on V WD f1; : : : ; ng2 as
described above. However, the domain of a permutation group in GAP must be
a set of positive integers, so the image of the homomorphism is made to be a
permutation group on the set f1; : : : ; n2 g, with i representing the i -th element of
V in lexicographic order.
gap> L2ActionHomomorphism := function(n)
> local action,tau,W;
> if not IsPosInt(n) then
> Error("usage: L2ActionHomomorphism( <PosInt> )");
> fi;
> action := function(x,g)
> # the function which determines the image of
> # x (in {1,...,nˆ2}) under g.
> local i,j,ii,jj;
> i:=QuoInt(x-1,n)+1;
> j:=x-(i-1)*n+n;
> ii:=iˆg;
> jj:=jˆg;
> if ii<=n then
> return n*(ii-1)+(jj-n);
> else
> return n*(jj-1)+(ii-n);
> fi;
> end;
> tau:=Product(List([1..n],i->(i,i+n)));
> W:=Group(Concatenation(
> GeneratorsOfGroup(SymmetricGroup([1..n])),[tau]));
> # so W = Sn wr C2 in its imprimitive action on [1..2*n].
> return ActionHomomorphism(
> W, # group
> [1..nˆ2], # domain of action
> action); # action of group on domain
> end;;
100 L.H. Soicher

We now define a Wn -invariant block design Un;m D .V; Bn;m /, which contains
the dual of every .n  n/=k semi-Latin square as a subdesign, as long as this dual
has no block of multiplicity greater than m. As before, V WD f1; : : : ; ng2 . The block
multiset Bn;m consists of all the subsets of V of the form

f.1; 1g /; : : : ; .n; ng /g;

such that g 2 Sn , and with each such block having multiplicity m. (Note that Un;m
is the dual of a certain .n  n/=.m.n  1/Š/ semi-Latin square.) The GAP function
defined below returns this “universal semi-Latin square dual” Un;m in DESIGN
package format.
gap> UniversalSemiLatinSquareDual := function(n,m)
> local g,i,block,blocks,U;
> if n<=1 or m<=0 then
> Error("<n> must be > 1 and <m> must be > 0");
> fi;
> blocks:=[];
> for g in SymmetricGroup([1..n]) do
> block:=[];
> for i in [1..n] do
> Add(block,(i-1)*n+iˆg);
> od;
> for i in [1..m] do
> Add(blocks,block);
> od;
> od;
> U:=BlockDesign(nˆ2,blocks);
> U.pointNames:=Immutable(Cartesian([1..n],[1..n]));
> return U;
> end;;

Observe that a block design  D .V; B/ is the dual of an .n  n/=k semi-Latin

square if and only if  is a 1-.n2 ; n; k/ design as well as a subdesign of Un;k (i.e. 
and Un;k have the same point set and B is a submultiset of Bn;k ). We thus obtain
the following:
Theorem 3.9 ([37]). The isomorphism classes of the .n  n/=k semi-Latin squares
are in one-to-one correspondence with the Wn -orbits of 1-.n2 ; n; k/ subdesigns
of Un;k . Representatives of these orbits give the duals of isomorphism class
representatives of the .n  n/=k semi-Latin squares.
This theorem can be adapted to be able to apply the function BlockDesigns
in the DESIGN package to construct and classify semi-Latin squares whose duals
satisfy certain Wn -invariant properties, and/or whose duals are invariant under a
user-specified subgroup of Wn . In particular:
Theorem 3.10 ([37]). The isomorphism classes of the SOMA.k; n/s are in one-
to-one correspondence with the Wn -orbits of the 1-.n2 ; n; k/ subdesigns of Un;1
having the property that any two distinct blocks meet in at most one point.
3 Designs, Groups and Computing 101

Representatives of these orbits give the duals of isomorphism class representatives

of the SOMA.k; n/s.
As an application, we find that, up to isomorphism, there are 2799 SOMA.2; 6/s,
and just 4 SOMA.3; 6/s (see [31]).
Theorem 3.11 ([37]). Suppose n  1 divides k and let  WD k=.n  1/. The
isomorphism classes of the uniform .n  n/=k semi-Latin squares are in one-to-one
correspondence with the Wn -orbits of the subdesigns of Un; with the property that
any two points having no co-ordinate in common occur together in exactly  blocks.
Representatives of these orbits give the duals of isomorphism class representatives
of the uniform .n  n/=k semi-Latin squares.
As an application, we find that, up to isomorphism, there are just 10 uniform
.5  5/=8 semi-Latin squares, and just 277 uniform .5  5/=12 semi-Latin squares.
A further, more complicated application is given below. We use the function
BlockDesigns to classify, up to isomorphism, the uniform .6  6/=10 semi-
Latin squares with the property that any two symbols occur together in at most two
entries (equivalently, any two blocks in the dual of such a square meet in at most
two points). It turns out that there are exactly 98 such semi-Latin squares, and we
compute a list L of their duals. We then determine the sizes of the automorphism
groups of the elements of L. (The automorphism group of the dual S  of an
.nn/=k semi-Latin square S is defined to be the subgroup of Wn that preserves the
block multiset of S  . This is the intersection of Wn with the standard automorphism
group of S  regarded as a block design with no structure on the point set.) The total
time taken for this calculation is about 7 min on a 3.1 GHz PC running Linux.
gap> n:=6;;
gap> hom:=L2ActionHomomorphism(n);;
gap> W:=Image(hom);;
gap> mu:=2;;
gap> U:=UniversalSemiLatinSquareDual(n,mu);;
gap> rel1:=Set(Orbit(W,[1,2],OnSets));;
gap> rel2:=Difference(Combinations([1..nˆ2],2),rel1);;
gap> L:=BlockDesigns(rec(v:=nˆ2,
> blockDesign:=U, # we are looking at subdesigns of U
> blockSizes:=[n],
> tSubsetStructure:=rec(t:=2, partition:=[rel1,rel2],
> lambdas:=[0,mu]),
> blockIntersectionNumbers:=[[[0,1,2]]],
> isoGroup:=W) );;
gap> A:=List(L,x->Intersection(W,AutomorphismGroup(x)));;
gap> autsizes:=List(A,Size);;
gap> Collected(autsizes);
[ [ 4, 24 ], [ 8, 33 ], [ 12, 4 ], [ 16, 14 ], [ 24, 11 ],
[ 32, 2 ], [ 48, 2 ], [ 80, 1 ], [ 96, 2 ], [ 144, 1 ],
[ 192, 2 ], [ 576, 1 ], [ 14400, 1 ] ]

A further calculation, taking just 1s, shows that no design in L is resolvable;

equivalently, no semi-Latin square whose dual is in L is the superposition of Latin
squares.
102 L.H. Soicher

gap> R:=List(L,design->PartitionsIntoBlockDesigns(rec(v:=nˆ2,
> blockDesign:=design,
> blockSizes:=[n],
> tSubsetStructure:=rec(t:=1, lambdas:=[1]),
> isoLevel:=0) ) );;
gap> Collected(List(R,Length));
[ [ 0, 98 ] ]

3.8 Efficient Semi-Latin Squares as Subsquares of Uniform

Semi-Latin Squares

For n a prime power, Bailey [2] gives a construction which produces efficient
(although not necessarily optimal [37]) .n  n/=k semi-Latin squares for all k > 1.
Efficient (but not known to be optimal) .6  6/=k semi-Latin squares are known for
k D 2; 3; 4; 5; 6; see [5, 8, 31, 37]. Any uniform .6  6/=10 semi-Latin square, such
as SLS.PSL2 .5//, is Schur-optimal, and so is A-, D- and E-optimal. This leaves
open the problem of finding efficient .6  6/=k semi-Latin squares, for k D 7; 8; 9.
To do this, we look at “subsquares” of a uniform .6  6/=10 semi-Latin square.
We say that an n  n semi-Latin square S is a subsquare of an n  n semi-Latin
square T if S D T or T is the superposition of S and another n  n semi-Latin
square. Another way of looking at subsquares is via duals. We have S a subsquare
of T if and only if the symbol set of S is a subset of the symbol set of T and the
dual S  of S is a 1-.n2 ; n; r/ subdesign of T  , for some r > 0. In [37], subsquares
of uniform semi-Latin squares are investigated, and the following result is proved.
Theorem 3.12 ([37]). Let n  3 and let S be an .n  n/=k subsquare of a uniform
.n  n/=t semi-Latin square T , such that t  k < n  1. Then

ES D 1  t=.k.n  1// D 1  .T /=k:

Now let k 2 f7; 8; 9g. By Theorem 3.12, each .6  6/=k subsquare of a uniform
.6  6/=10 semi-Latin square has E-efficiency measure 1  2=k. We obtain an
efficient .6  6/=k semi-Latin square Yk by taking the most A-efficient subsquare
of that size of a certain uniform .6  6/=10 semi-Latin square Y10 . The square Y10
was chosen as follows. The list L above of the duals of the 98 uniform .6  6/=10
semi-Latin squares with the property that any two distinct symbols occur together in
at most two blocks includes the dual of SLS.PSL2 .5//, whose automorphism group
has size 14,400. However, SLS.PSL2 .5// has no .6  6/=9 subsquare (equivalently
it has no Latin square of order 6 as a subsquare). The next largest automorphism
group size amongst the 98 dual squares in L is 576, and the one square whose dual
in L has an automorphism group of size 576 is chosen as Y10 , which does indeed
3 Designs, Groups and Computing 103

have .6  6/=k subsquares, for k D 7; 8; 9. We have chosen a square whose dual

has a large automorphism group to facilitate the classification of subsquares.
gap> f:=First([1..Length(L)],i->Size(A[i])=576);;
gap> Y10star:=L[f];;
gap> autY10star:=A[f];;
gap> StructureDescription(autY10star);
"((A4 x A4) : C2) : C2"

We give the square Y10 columnwise below:

1 2 3 4 5 6 7 8 9 10
11 12 21 22 31 32 41 42 51 52
13 19 23 25 35 37 47 49 53 59
14 16 27 29 33 40 44 46 57 60
15 18 28 30 34 39 43 45 55 58
17 20 24 26 36 38 48 50 54 56

11 12 13 14 15 16 17 18 19 20
1 2 23 24 33 34 43 44 53 54
3 9 21 27 38 39 45 48 51 57
5 8 22 30 32 36 47 50 55 59
6 7 26 29 31 35 42 49 56 60
4 10 25 28 37 40 41 46 52 58

21 22 23 24 25 26 27 28 29 30
3 4 13 14 35 36 45 46 55 56
1 5 11 17 31 40 43 50 58 60
2 7 15 19 37 39 42 48 52 54
8 10 16 20 33 38 41 47 51 53
6 9 12 18 32 34 44 49 57 59

31 32 33 34 35 36 37 38 39 40
5 6 15 16 25 26 47 48 57 58
4 7 18 20 22 29 41 44 54 55
1 10 12 17 23 28 45 49 51 56
2 9 11 13 24 27 46 50 52 59
3 8 14 19 21 30 42 43 53 60
104 L.H. Soicher

41 42 43 44 45 46 47 48 49 50
7 8 17 18 27 28 37 38 59 60
6 10 14 15 24 30 32 33 52 56
4 9 11 20 21 26 34 35 53 58
1 3 12 19 22 25 36 40 54 57
2 5 13 16 23 29 31 39 51 55

51 52 53 54 55 56 57 58 59 60
9 10 19 20 29 30 39 40 49 50
2 8 12 16 26 28 34 36 42 46
3 6 13 18 24 25 31 38 41 43
4 5 14 17 21 23 32 37 44 48
1 7 11 15 22 27 33 35 45 47

We now classify the .6  6/=7 subsquares of Y10 by classifying the 1-.36; 6; 7/

 
subdesigns of Y10 , up to the action of the automorphism group of Y10 , as follows.
gap> k:=7;;
gap> subdesigns:=BlockDesigns(rec(v:=nˆ2,
> blockDesign:=Y10star,
> blockSizes:=[n],
> tSubsetStructure:=rec(t:=1, lambdas:=[k]),
> isoGroup:=autY10star) );;
gap> Length(subdesigns);
150

The determination of these 150 subdesigns takes about 7 min on a 3.1 GHz PC
running Linux.
We next determine the design(s) in the list subdesigns with the highest A-
efficiency measure. There is just one such subdesign, and it also has the highest
D-efficiency measure of those in the list. These calculations take about 16 s.
gap> eff:=List(subdesigns,BlockDesignEfficiency);;
gap> maxA:=Maximum(List(eff,x->x.A));
18972014997910099125/22524377910796536046
gap> pos:=Filtered([1..Length(eff)],j->eff[j].A=maxA);
[ 65 ]
gap> ForAll([1..Length(eff)],
> j->eff[pos[1]].Dpowered>=eff[j].Dpowered);
true

Let Y7 be the subsquare of Y10 whose dual Y7 is the design in the list sub-
designs with the highest A-efficiency measure. Then Y7 is the .6  6/=7 semi-
Latin square obtained from Y10 by removing the .66/=3 semi-Latin square induced
on the symbols

f4; 8; 9; 13; 15; 17; 22; 25; 26; 32; 33; 39; 43; 46; 49; 51; 54; 60g:
3 Designs, Groups and Computing 105

The A-efficiency measure of Y7 is

22224360426123258975=25776723339009695896 0:8622:

Similar calculations find a .6  6/=8 subsquare of Y10 with the highest A-

efficiency measure. It also has the highest D-efficiency measure. This subsquare
Y8 can be obtained from Y10 by removing the SOMA.2; 6/ induced on the symbols

f2; 6; 17; 19; 21; 29; 33; 36; 41; 45; 58; 59g:

The A-efficiency measure of Y8 is 3643863=4141988 0:8797.

Up to isomorphism, we find there is just one .6  6/=9 subsquare of Y10 . We may
take this to be the semi-Latin square Y9 obtained from Y10 by removing the Latin
square induced on the symbols

f8; 11; 25; 39; 44; 56g:

The A-efficiency measure of Y9 is 2968=3323 0:8932.

3.9 Some Open Problems

We conclude this chapter with some open problems.

When n is a prime power or n D 6, we know precisely the values of  for
which there exists a uniform .n  n/=..n  1// semi-Latin square, but we do not
know exactly which values of  have this property for any other n > 1. The first
unsettled case is n D 10. It is a celebrated computational result that there is no
projective plane of order 10 [22, 28], so there do not exist nine MOLS of order 10,
and so a uniform .10  10/=9 semi-Latin square does not exist. On the other hand,
SLS.PSL2 .9// and inflations of this square yield uniform .10  10/=.9/ semi-
Latin squares for  D 4; 8; 12; 16; : : :. Our first question is: do there exist uniform
.10  10/=18 or .10  10/=27 semi-Latin squares?
We have classified certain types of uniform .6  6/=10 semi-Latin squares, and
found none which is a superposition of Latin squares. Our second question is: does
there exist a uniform .6  6/=10 semi-Latin square which is a superposition of ten
Latin squares?
Finally, are there general constructions for optimal (say E-optimal) .n  n/=k
semi-Latin squares when there do not exist k MOLS of order n and there is no
uniform .n  n/=k semi-Latin square? For example, is every .n  n/=.k  1/
subsquare of a uniform .n  n/=k semi-Latin square E-optimal?

Acknowledgements I am grateful to Martin Liebeck for his result of Sect. 3.5.1 and for allowing
its inclusion in this chapter. I also thank Eamonn O’Brien for his calculations in M AGMA .
The hospitality of the de Brún Centre for Computational Algebra, NUI Galway, during the Fifth
de Brún Workshop is gratefully acknowledged.
106 L.H. Soicher

References

1. R.A. Bailey, Semi-Latin squares. J. Stat. Plan. Inference 18, 299–312 (1988)
2. R.A. Bailey, Efficient semi-Latin squares. Stat. Sinica 2, 413–437 (1992)
3. R.A. Bailey, Association Schemes. Designed Experiments, Algebra and Combinatorics (Cam-
bridge University Press, Cambridge, 2004)
4. R.A. Bailey, Design of Comparative Experiments (Cambridge University Press, Cambridge,
2008)
5. R.A. Bailey, Symmetric factorial designs in blocks. J. Stat. Theor. Pract. 5, 13–24 (2011)
6. R.A. Bailey, P.J. Cameron, Combinatorics of optimal designs, in Surveys in Combinatorics
2009, ed. by S. Huczynska et al. (Cambridge University Press, Cambridge, 2009), pp. 19–73
7. R.A. Bailey, P.E. Chigbu, Enumeration of semi-Latin squares. Discrete Math. 167–168, 73–84
(1997)
8. R.A. Bailey, G. Royle, Optimal semi-Latin squares with side six and block size two. Proc. Roy.
Soc. Lond. Ser. A 453, 1903–1914 (1997)
9. W. Bosma, J. Cannon, C. Playoust, The Magma algebra system. I. The user language. J.
Symbolic Comput. 24, 235–265 (1997)
10. P.J. Cameron, Permutation Groups (Cambridge University Press, Cambridge, 1999)
11. P.J. Cameron, P. Dobcsányi, J.P. Morgan, L.H. Soicher, The External Representation of Block
Designs, http://designtheory.org/library/extrep/extrep-1.1-html/
12. P.J. Cameron, L.H. Soicher, Block intersection polynomials. Bull. Lond. Math. Soc. 39, 559–
564 (2007)
13. C.-S. Cheng, R.A. Bailey, Optimality of some two-associate-class partially balanced
incomplete-block designs. Ann. Stat. 19, 1667–1671 (1991)
14. P.E. Chigbu, Semi-Latin Squares: Methods for Enumeration and Comparison, Ph.D. Thesis,
University of London, 1996
15. J.H. Conway, R.T. Curtis, S.P. Norton, R.A. Parker, R.A. Wilson, ATLAS of Finite Groups
(Clarendon Press, Oxford, 1985)
16. P. Dobcsányi, D.A. Preece, L.H. Soicher, On balanced incomplete-block designs with repeated
blocks. Eur. J. Comb. 28, 1955–1970 (2007)
17. R.N. Edmondson, Trojan square and incomplete Trojan square designs for crop research. J.
Agric. Sci. 131, 135–142 (1998)
18. The GAP Group, GAP — Groups, Algorithms, and Programming, Version 4.5.2(beta), 2011,
http://www.gap-system.org/
19. A. Giovagnoli, H.P. Wynn, Optimum continuous block designs. Proc. Roy. Soc. Lond. Ser. A
377, 405–416 (1981)
20. G. James, M. Liebeck, Representations and Characters of Groups, 2nd edn. (Cambridge
University Press, Cambridge, 2001)
21. P. Kaski, P.R.J. Östergård, Classification Algorithms for Codes and Designs (Springer, Berlin,
2006)
22. C.W.H. Lam, L. Thiel, S. Swiercz, The non-existence of finite projective planes of order 10.
Can. J. Math. 41, 1117–1123 (1989)
23. J.H. van Lint, R.M. Wilson, A Course in Combinatorics (Cambridge University Press,
Cambridge, 1992)
24. J.P. McSorley, L.H. Soicher, Constructing t -designs from t -wise balanced designs. Eur. J.
Comb. 28, 567–571 (2007)
25. N.C.K. Phillips, W.D. Wallis, All solutions to a tournament problem. Congr. Numerantium
114, 193–196 (1996)
26. D.A. Preece, G.H. Freeman, Semi-Latin squares and related designs. J. Roy. Stat. Soc. Ser. B
45, 267–277 (1983)
27. D.A. Preece, N.C.K. Phillips, Euler at the bowling green. Utilitas Math. 61, 129–165 (2002)
28. D.J. Roy, Confirmation of the Non-existence of a Projective Plane of Order 10, M.Sc. Thesis,
Carleton University, Ottawa, 2011
3 Designs, Groups and Computing 107

29. K.R. Shah, B.K. Sinha, Theory of Optimal Designs. Lecture Notes in Statistics, vol. 54
(Springer, Berlin, 1989)
30. L.H. Soicher, On the structure and classification of SOMAs: generalizations of mutually
orthogonal Latin squares. Electron. J. Comb. 6, R32 (1999), Printed version: J. Comb. 6, 427–
441 (1999)
31. L.H. Soicher, SOMA Update, http://www.maths.qmul.ac.uk/leonard/soma/
32. L.H. Soicher, Computational group theory problems arising from computational design theory.
Oberwolfach Rep. 3 (2006), Report 30/2006: Computational group theory, 1809–1811
33. L.H. Soicher, More on block intersection polynomials and new applications to graphs and
block designs. J. Comb. Theor. Ser. A 117, 799–809 (2010)
34. L.H. Soicher, The DESIGN package for GAP (2011), Version 1.6, http://designtheory.org/
software/gap design/
35. L.H. Soicher, The GRAPE package for GAP, Version 4.5, 2011, http://www.maths.qmul.ac.
uk/leonard/grape/
36. L.H. Soicher, Uniform semi-Latin squares and their Schur-optimality. J. Comb. Des. 20, 265–
277 (2012)
37. L.H. Soicher, Optimal and efficient semi-Latin squares. J. Stat. Plan. Inference 143, 573–582
(2013)
38. C.-Y. Suen, I.M. Chakravarti, Efficient two-factor balanced designs. J. Roy. Stat. Soc. Ser. B
48, 107–114 (1986)
39. M.N. Vartak, The non-existence of certain PBIB designs. Ann. Math. Stat. 30, 1051–1062
(1959)
LECTURE NOTES IN MATHEMATICS 123
Edited by J.-M. Morel, B. Teissier; P.K. Maini

Editorial Policy (for the publication of monographs)

1. Lecture Notes aim to report new developments in all areas of mathematics and their
applications - quickly, informally and at a high level. Mathematical texts analysing new
developments in modelling and numerical simulation are welcome.
Monograph manuscripts should be reasonably self-contained and rounded off. Thus they
may, and often will, present not only results of the author but also related work by other
people. They may be based on specialised lecture courses. Furthermore, the manuscripts
should provide sufficient motivation, examples and applications. This clearly distinguishes
Lecture Notes from journal articles or technical reports which normally are very concise.
Articles intended for a journal but too long to be accepted by most journals, usually do not
have this “lecture notes” character. For similar reasons it is unusual for doctoral theses to
be accepted for the Lecture Notes series, though habilitation theses may be appropriate.
2. Manuscripts should be submitted either online at www.editorialmanager.com/lnm to
Springer’s mathematics editorial in Heidelberg, or to one of the series editors. In general,
manuscripts will be sent out to 2 external referees for evaluation. If a decision cannot yet
be reached on the basis of the first 2 reports, further referees may be contacted: The author
will be informed of this. A final decision to publish can be made only on the basis of the
complete manuscript, however a refereeing process leading to a preliminary decision can
be based on a pre-final or incomplete manuscript. The strict minimum amount of material
that will be considered should include a detailed outline describing the planned contents
of each chapter, a bibliography and several sample chapters.
Authors should be aware that incomplete or insufficiently close to final manuscripts almost
always result in longer refereeing times and nevertheless unclear referees’ recommenda-
tions, making further refereeing of a final draft necessary.
Authors should also be aware that parallel submission of their manuscript to another
publisher while under consideration for LNM will in general lead to immediate rejection.
3. Manuscripts should in general be submitted in English. Final manuscripts should contain
at least 100 pages of mathematical text and should always include
– a table of contents;
– an informative introduction, with adequate motivation and perhaps some historical
remarks: it should be accessible to a reader not intimately familiar with the topic
treated;
– a subject index: as a rule this is genuinely helpful for the reader.
For evaluation purposes, manuscripts may be submitted in print or electronic form (print
form is still preferred by most referees), in the latter case preferably as pdf- or zipped
psfiles. Lecture Notes volumes are, as a rule, printed digitally from the authors’ files.
To ensure best results, authors are asked to use the LaTeX2e style files available from
Springer’s web-server at:
ftp://ftp.springer.de/pub/tex/latex/svmonot1/ (for monographs) and
ftp://ftp.springer.de/pub/tex/latex/svmultt1/ (for summer schools/tutorials).
 Additional technical instructions, if necessary, are available on request from lnm@springer.
com.
4. Careful preparation of the manuscripts will help keep production time short besides
ensuring satisfactory appearance of the finished book in print and online. After acceptance
of the manuscript authors will be asked to prepare the final LaTeX source files and
also the corresponding dvi-, pdf- or zipped ps-file. The LaTeX source files are essential
for producing the full-text online version of the book (see http://www.springerlink.com/
openurl.asp?genre=journal&issn=0075-8434 for the existing online volumes of LNM).
The actual production of a Lecture Notes volume takes approximately 12 weeks.
5. Authors receive a total of 50 free copies of their volume, but no royalties. They are entitled
to a discount of 33.3 % on the price of Springer books purchased for their personal use, if
ordering directly from Springer.
6. Commitment to publish is made by letter of intent rather than by signing a formal contract.
Springer-Verlag secures the copyright for each volume. Authors are free to reuse material
contained in their LNM volumes in later publications: a brief written (or e-mail) request
for formal permission is sufficient.

Addresses:
Professor J.-M. Morel, CMLA,
École Normale Supérieure de Cachan,
61 Avenue du Président Wilson, 94235 Cachan Cedex, France
E-mail: [email protected]
Professor B. Teissier, Institut Mathématique de Jussieu,
UMR 7586 du CNRS, Équipe “Géométrie et Dynamique”,
175 rue du Chevaleret
75013 Paris, France
E-mail: [email protected]
For the “Mathematical Biosciences Subseries” of LNM:
Professor P. K. Maini, Center for Mathematical Biology,
Mathematical Institute, 24-29 St Giles,
Oxford OX1 3LP, UK
E-mail : [email protected]
Springer, Mathematics Editorial, Tiergartenstr. 17,
69121 Heidelberg, Germany,
Tel.: +49 (6221) 4876-8259

Fax: +49 (6221) 4876-8259

E-mail: [email protected]