This page intentionally left blank

L O N D O N M AT H E M AT I C A L S O C I E T Y S T U D E N T T E X T S

Managing editor: Professor D. Benson,

Department of Mathematics, University of Aberdeen, UK
20 Communication theory, C. M. GOLDIE & R. G. E. PINCH
21 Representations of finite groups of Lie type, FRANCOIS DIGNE & JEAN MICHEL
22 Designs, graphs, codes, and their links, P. J. CAMERON & J. H. VAN LINT
23 Complex algebraic curves, FRANCES KIRWAN
24 Lectures on elliptic curves, J. W. S. CASSELS
26 An introduction to the theory of L-functions and Eisenstein series, H. HIDA
27 Hilbert Space; compact operators and the trace theorem, J. R. RETHERFORD
28 Potential theory in the complex plane, T. RANSFORD
29 Undergraduate commutative algebra, M. REID
31 The Laplacian on a Riemannian manifold, S. ROSENBERG
32 Lectures on Lie groups and Lie algebras, R. CARTER, G. SEGAL & I. MACDONALD
33 A primer of algebraic D-modules, S. C. COUNTINHO
34 Complex algebraic surfaces, A. BEAUVILLE
35 Young tableaux, W. FULTON
37 A mathematical introduction to wavelets, P. WOJTASZCZYK
38 Harmonic maps, loop groups, and integrable systems, M. GUEST
39 Set theory for the working mathematician, K. CIESIELSKI
40 Ergodic theory and dynamical systems, M. POLLICOTT & M. YURI
41 The algorithmic resolution of diophantine equations, N. P. SMART
42 Equilibrium states in ergodic theory, G. KELLER
43 Fourier analysis on finite groups and applications, AUDREY TERRAS
44 Classical invariant theory, PETER J. OLVER
45 Permutation groups, P. J. CAMERON
46 Riemann surfaces: A primer, A. BEARDON
47 Introductory lectures on rings and modules, J. BEACHY
48 Set theory, A HAJNAL & P. HAMBURGER
49 K-theory for C*-algebras, M. RORDAM, F. LARSEN & N. LAUSTSEN
50 A brief guide to algebraic number theory, H. P. F. SWINNERTON-DYER
51 Steps in commutative algebra: Second edition, R. Y. SHARP
52 Finite Markov chains and algorithmic applications, O. HÄGGSTRÖM
53 The prime number theorem, G. J. O. JAMESON
54 Topics in graph automorphisms and reconstruction, J. LAURI & R. SCAPELLATO
55 Elementary number theory, group theory, and Ramanujan graphs, G. DAVIDOFF,
P. SARNAK & A. VALETTE
56 Logic, induction and sets, T. FORSTER
57 Introduction to Banach algebras and harmonic analysis, H. G. DALES et al.
58 Computational algebraic geometry, HAL SCHENCK
59 Frobenius algebras and 2-D topological quantum field theories, J. KOCK
60 Linear operators and linear systems, J. R. PARTINGTON
61 An introduction to noncommutative Noetherian rings, K. R. GOODEARL & R. B. WARFIELD
62 Topics from one dimensional dynamics, K. M. BRUCKS & H. BRUIN
63 Singularities of plane curves, C. T. C. WALL
64 A short course on Banach space theory, N. L. CAROTHERS
65 Elements of the representation theory of associative algebras Volume I, I. ASSEM,
A. SKOWROŃSKI & D. SIMSON
66 An introduction to sieve methods and their applications, A. C. COJOCARU & M. R. MURTY
67 Elliptic functions, V. ARMITAGE & W. F. EBERLEIN
68 Hyperbolic geometry from a local viewpoint, L. KEEN & N. LAKIC
69 Lectures on Kähler geometry, A. MOROIANU
70 Dependence logic, J. VÄÄNÄNEN
71 Elements of the representation theory of associative algebras Volume 2, D. SIMSON & A. SKOWROŃSKI
72 Elements of the representation theory of associative algebras Volume 3, D. SIMSON & A. SKOWROŃSKI
73 Groups, graphs and trees, J. MEIER
 L O N D O N M AT H E M AT I C A L S O C I E T Y S T U D E N T
TEXTS 73

Groups, Graphs and Trees

An Introduction to the Geometry of Infinite Groups

JOHN MEIER
Lafayette College
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521895453

© J. Meier 2008

This publication is in copyright. Subject to statutory exception and to the provision of

relevant collective licensing agreements, no reproduction of any part may take place
without the written permission of Cambridge University Press.
First published in print format 2008

ISBN-13 978-0-511-42345-1 eBook (EBL)

ISBN-13 978-0-521-89545-3 hardback

ISBN-13 978-0-521-71977-3 paperback

Cambridge University Press has no responsibility for the persistence or accuracy of urls
for external or third-party internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.
For my driver, Piotr
 Contents

Preface page ix
1 Cayley’s Theorems 1
1.1 Cayley’s Basic Theorem 1
1.2 Graphs 6
1.3 Symmetry Groups of Graphs 10
1.4 Orbits and Stabilizers 15
1.5 Generating Sets and Cayley Graphs 17
1.6 More Cayley Graphs 22
1.7 Symmetries of Cayley Graphs 29
1.8 Fundamental Domains and Generating Sets 30
1.9 Words and Paths 37
2 Groups Generated by Reflections 44
3 Groups Acting on Trees 54
3.1 Free Groups 54
3.2 F3 is a Subgroup of F2 65
3.3 Free Group Homomorphisms and
Group Presentations 67
3.4 Free Groups and Actions on Trees 70
3.5 The Group Z3 ∗ Z4 73
3.6 Free Products of Groups 79
3.7 Free Products of Finite Groups are Virtually Free 83
3.8 A Geometric View of Theorem 3.35 87
3.9 Finite Groups Acting on Trees 89
3.10 Serre’s Property FA and Infinite Groups 90
4 Baumslag–Solitar Groups 100

vii
viii Contents

5 Words and Dehn’s Word Problem 105

5.1 Normal Forms 105
5.2 Dehn’s Word Problem 109
5.3 The Word Problem and Cayley Graphs 111
5.4 The Cayley Graph of BS(1,2) 115
6 A Finitely Generated, Infinite Torsion Group 120
7 Regular Languages and Normal Forms 130
7.1 Regular Languages and Automata 130
7.2 Not All Languages are Regular 136
7.3 Regular Word Problem? 140
7.4 A Return to Normal Forms 141
7.5 Finitely Generated Subgroups of Free Groups 143
8 The Lamplighter Group 151
9 The Geometry of Infinite Groups 162
9.1 Gromov’s Corollary, aka The Word Metric 162
9.2 The Growth of Groups, I 168
9.3 Growth and Regular Languages 172
9.4 Cannon Pairs 175
9.5 Cannon’s Almost Convexity 179
10 Thompson’s Group 187
11 The Large-Scale Geometry of Groups 198
11.1 Changing Generators 198
11.2 The Growth of Groups, II 202
11.3 The Growth of Thompson’s Group 205
11.4 The Ends of Groups 208
11.5 The Freudenthal–Hopf Theorem 211
11.6 Two-Ended Groups 212
11.7 Commensurable Groups and Quasi-Isometry 217
Bibliography 227
Index 230
 Preface

Groups are algebraic objects, consisting of a set with a binary operation

that satisfies a short list of required properties: the binary operation
must be associative; there is an identity element; and every element has
an inverse. Presenting groups in this formal, abstract algebraic manner is
both useful and powerful. Yet it avoids a wonderful geometric perspective
on group theory that is also useful and powerful, particularly in the study
of infinite groups. This perspective is hinted at in the combinatorial
approach to finite groups that is often seen in a first course in abstract
algebra. It is my intention to bring the geometric perspective forward,
to establish some elementary results that indicate the utility of this
perspective, and to highlight some interesting examples of particular
infinite groups along the way. My own bias is that these groups are just
as interesting as the theorems.
The topics covered in this book fit inside of “geometric group the-
ory,” a field that sits in the impressively large intersection of abstract
algebra, geometry, topology, formal language theory, and many other
fields. I hope that this book will provide an introduction to geometric
group theory at a broadly accessible level, requiring nothing more than
a single-semester exposure to groups and a naive familiarity with the
combinatorial theory of graphs.
The chapters alternate between those devoted to general techniques
and theorems (odd numbers) and brief chapters introducing some of
the standard examples of infinite groups (even numbers). Chapter 2
presents a few groups generated by reflections; Chapter 4 presents the
Baumslag–Solitar group BS(1, 2) in terms of linear functions; Chapter 6
is the Gupta–Sidki variant of Grigorchuk’s group; the Lamplighter group
is discussed in Chapter 8; and Thompson’s group F is the subject of
Chapter 10. When I taught this material at Lafayette College I referred

ix
x Preface

to the material in these even numbered chapters as “field trips to the

Zoo of Infinite Groups.”
The first chapter should be relatively easy to work through, as it
reviews material on groups (mainly finite groups), group actions, and
the combinatorial theory of graphs. It establishes quite a bit of notation
and introduces the construction of Cayley graphs. While some material
in this chapter may be new to the reader, most of it should seem to be
a repackaging of ideas that she or he has previously encountered.
Chapter 3 is an introduction to free groups and free products of
groups. Chapters 5 and 7 are devoted to connections between finitely
generated groups and formal language theory. Chapters 9 and 11 deal
with the geometry of infinite groups, with Chapter 9 focusing on what
might be called the “fine geometry” of Cayley graphs, while Chapter 11
treats what is called the large-scale geometry of groups.
While no background beyond elementary group theory is necessary for
this book, a broader undergraduate exposure to mathematics is certainly
helpful. My experience in the classroom indicates that the material in
Chapter 7 is demanding for people who have not previously encountered
formal language theory. Similarly Chapter 11 is easier for people who
have had a course in real analysis. Because hyperbolic geometry is not
a standard undergraduate topic, Gromov’s theory of hyperbolic groups
does not appear in this book. Similarly, because algebraic topology is
not a standard undergraduate topic, I have avoided fundamental groups
and covering spaces.
There are two forms of exercises in this book. A few exercises are
embedded within the chapters. These should be done, at least at the
level of the reader convincing themselves they know how to do them,
while reading through the material. There are also end-of-chapter ex-
ercises that are arranged in the order that material is presented in the
chapter. Some of these end-of-chapter exercises are challenging but most
are reasonably accessible.
Groups, Graphs and Trees was developed from notes used in two under-
graduate course offerings at Lafayette College, and it can certainly serve
as a primary text for an advanced undergraduate course. It should also
be useful as a text for a reading course and as a gentle introduction to
geometric group theory for mathematicians with a broader background
than this. An undergraduate course that attempted to cover this text,
omitting no details, in one semester, would have to move at a rather
brisk pace. The critical background information is contained in the first
five chapters and those should not be trimmed. With a bit of forethought
 Preface xi

an instructor can cover much of the rest of this book, if for example the
material in Chapter 7 or Chapter 11 is presented more as a colloquium
than as course material. My own hope is that various classes will find
the space in their semester to pursue tangents of interest to them, and
then let me know the results of their exploration.
I have many people to thank. My wife Trisha and son Robert were
unreasonably supportive of this project. Many students provided import-
ant feedback as I fumbled through the process of presenting this bit of
advanced mathematics at an elementary level: George Armagh, Kari
Barkley, Jenna Bratz, Jacob Carson, Joellen Cope, Joe Dudek, Josh
Goldstein, Ekaterina Jager, Brian Kronenthal, Rob McEwen, and
Zachary Reiter. I also benefited from extensive feedback given by my
colleagues Ethan Berkove and Jon McCammond. Finally, a number of
anonymous referees provided comments on various draft chapters. I was
impressed by the fact that there was no intersection between the com-
ments provided by students, the comments provided by colleagues, and
the anonymous referees!
 1
Cayley’s Theorems

As for everything else, so for a mathematical theory: beauty can be perceived

but not explained.
. –Arthur Cayley

An introduction to group theory often begins with a number of examples

of finite groups (symmetric, alternating, dihedral, ...) and constructions
for combining groups into larger groups (direct products, for exam-
ple). Then one encounters Cayley’s Theorem, claiming that every finite
group can be viewed as a subgroup of a symmetric group. This chapter
begins by recalling Cayley’s Theorem, then establishes notation, termino-
logy, and background material, and concludes with the construction and
elementary exploration of Cayley graphs. This is the foundation we use
throughout the rest of the text where we present a series of variations on
Cayley’s original insight that are particularly appropriate for the study
of infinite groups.
Relative to the rest of the text, this chapter is gentle, and should
contain material that is somewhat familiar to the reader. A reader who
has not previously studied groups and encountered graphs will find the
treatment presented here “brisk.”

1.1 Cayley’s Basic Theorem

You probably already have good intuition for what it means for a group
to act on a set or geometric object. For example:

• The cyclic group of order n – denoted Zn – acts by rotations on a

regular n-sided polygon.

1
2 Cayley’s Theorems

• The dihedral group of order 2n – denoted Dn – also acts on the reg-

ular n-sided polygon, where the elements either rotate or reflect the
polygon.
• We use Symn to denote the symmetric group of all permutations
of [n] = {1, 2, . . . , n}. (More common notations are Sn and Σn .)
By its definition, Symn acts on this set of numbers, as does its
index 2 subgroup, the alternating group An , consisting of the even
permutations.
• Matrix groups, such GLn (R) (the group of invertible n-by-n matrices
with real number entries), act on vector spaces.

Because the general theme of this book is to study groups via actions,
we need a bit of notation and a formal definition.

Convention 1.1. If X is a mathematical object (such as a regular poly-

gon or a set of numbers), then we use Sym(X) to denote all bijections
from X to X that preserve the indicated mathematical structure. For
example, if X is a set, then Sym(X) is simply the group of permutations
of the elements of X. In fact, if n = |X| then Sym(X) ≈ Symn . More-
over, if X and X
have the same cardinality, then Sym(X) ≈ Sym(X
).
If X is a regular polygon, then angles and lengths are important, and
Sym(X) will be composed of rotations and reflections (and it will in fact
be a dihedral group). Similarly, if X is a vector space, then Sym(X) will
consist of bijective linear transformations.
What we are referring to as “Sym(X)” does have a number of different
names in different contexts within mathematics. For example, if G is a
group, then the collection of its symmetries is referred to as Aut(G), the
group of automorphisms. If we are working with the Euclidean plane, R2 ,
and are considering functions that preserve the distance between points,
then we are looking at Isom(R2 ), the group of isometries of the plane.
It is quite useful to have individual names for these groups, as their
names highlight what mathematical structures are being preserved. Our
convention of lumping these various groups all together under the name
“Sym” is vague, but we believe that in context it will be clear what
is intended, and we like the fact that this uniform terminology empha-
sizes that these various situations where groups arise are not all that
different.1 One egregious example, which highlights the need to be care-
1 In his book, Symmetry, Hermann Weyl wrote: “[W]hat has indeed become a guid-
ing principle in modern mathematics is this lesson: Whenever you have to deal
with a structure-endowed entity Σ try to determine its group of automorphisms,
the group of those element-wise transformations which leave all structural relations
 1.1 Cayley’s Basic Theorem 3

ful in using our convention, comes from the integers. If the integers are
thought of as simply a set, containing infinitely many elements, then
Sym(Z) is an infinite permutation group, which contains Symn for any
n. On the other hand, if Z denotes the group of integers under addition,
then Sym(Z) ≈ Z2 . (The only non-trivial automorphism of the group of
integers sends n to −n for all n ∈ Z.)
Definition 1.2. An action of a group G on a mathematical object X
is a group homomorphism from G to Sym(X). Equivalently, it is a map
from G × X → X such that

1. e · x = x, for all x ∈ X; and

2. (gh) · x = g · (h · x), for all g, h ∈ G and x ∈ X.
We denote “G acts on X” by G
X.
If one has a group action G
X, then the associated homomorphism
is a representation of G. The representation is faithful if the map is
injective. In other words, it is faithful if, given any non-identity element
g ∈ G, there is some x ∈ X such that g · x = x.
Example 1.3. The dihedral group Dn is the symmetry group of a
regular n-gon. As such, it also permutes the vertices of the n-gon, hence
there is a representation Dn → Symn . As every non-identity element of
Dn moves at least (n − 2) vertices, this representation is faithful.
Remark 1.4 (left vs. right). In terms of avoiding confusion, this is
perhaps the most important remark in this book. Because not all groups
are abelian, it is very important to keep left and right straight. All of
our actions will be left actions (as described above). We have chosen to
work with left actions since it matches function notation and because
left actions are standard in geometric group theory and topology.
Groups arise in a number of different contexts, most commonly as
symmetries of any one of a number of possible mathematical objects X.
In these situations, one can often understand the group directly from
our understanding of X. The dihedral and symmetric groups are two
examples of this. However, groups are abstract objects, being merely
a set with a binary operation that satisfies a certain minimal list of
requirements. Cayley’s Theorem shows that the abstract notion of a
group and the notion of a group of permutations are one and the same.
undisturbed. You can expect to gain a deep insight into the constitution of Σ in
this way.” Our use of Sym(Σ) instead of Aut(Σ) is a small notational deviation
from Weyl’s recommendation.
4 Cayley’s Theorems

Theorem 1.5 (Cayley’s Basic Theorem). Every group can be faithfully

represented as a group of permutations.

Proof. The objects that G permutes are the elements of G. In this proof
we use “SymG ” to denote Sym(G), to emphasize that “G” denotes the
underlying set of elements, not the group. The permutation associated
to g ∈ G is defined by left multiplication by g. That is, g → πg ∈ SymG
where πg (h) = g · h for all h ∈ G. This is a permutation of the elements
of G, since if g · h = g · h
, then by left cancellation, h = h
. Denote the
map taking the element g to the permutation πg by π : G → SymG .
To check that π is a group homomorphism we need to verify that
π(gh) = π(g) · π(h). In other words, we need to show that πgh = πg · πh .
We do this by evaluating what each side does to an arbitrary element of
G. We denote the arbitrary element by “x”, thinking of it as a variable.
The permutation πgh takes x → (gh) · x, and successively applying πh
then πg sends x → h · x → g · (h · x). Thus checking that φ is a homo-
morphism amounts to verifying the associative law: (gh) · x = g · (h · x).
As this is part of the definition of a group, the equation holds.
In order to see that the map is faithful it suffices to show that no
non-identity element is mapped to the trivial permutation. One can do
this by simply noting that if g ∈ G \ {e}, then g · e = g, hence πg (e) = g,
and so πg is not the identity (or trivial) permutation.

The proof of Cayley’s Basic Theorem constructs a representation

of G as a group of permutations of itself. Before moving on we should
examine what these permutations look like in some concrete situa-
tions. We first consider Sym3 , the group of all permutations of three
objects.

Notation 1.6 (cycle notation). In describing elements of Symn we use

cycle notation, and multiply (that is, compose permutations) right to
left. This matches with our intuition from functions where f ◦g(x) means
that you first apply g then apply f , and it is consistent with our use of
left actions. Here is a concrete example: (12)(35) ∈ Sym5 is the ele-
ment that transposes 1 and 2, as well as 3 and 5; the element (234)
sends 2 to 3, 3 to 4 and 4 to 2; the product (12)(35) · (234) = (12534).
(The product is not (13542), which is the result of multiplying left
to right.)
 1.1 Cayley’s Basic Theorem 5

Example 1.7. The group Sym3 has six elements, shown as disjoint
vertices in Figure 1.1. The permutations described by Cayley’s Basic
Theorem – for the elements (12) and (123) – are also shown.

(12) (12)

I I

(132) (123) (132) (123)

(13) (23) (13) (23)

Fig. 1.1. The permutation of Sym3 induced by (12) is shown on the left, and
the permutation induced by (123) is shown on the right.

(–1,1) (0,1) (1,1) (2,1)

(–1,0) (0,0) (1,0) (2,0)

(–1,–1) (0,–1) (1,–1) (2,–1)

Fig. 1.2. The action of (2, 1) on Z ⊕ Z.

Example 1.8. In most introductions to group theory, Cayley’s Basic

Theorem is stated for finite groups. But we made no such assumption in
our statement and the same proof as is given for finite groups works for
infinite groups. Consider for example the direct product of two copies
of the group of integers, G = Z ⊕ Z. Here elements are represented by
pairs of integers, and the binary operation is coordinatewise addition:
(a, b) + (c, d) = (a + c, b + d). In Figure 1.2 we have arranged the vertices
corresponding to elements of G as the integral lattice in the plane. The
arrows indicate the permutation of the elements of Z ⊕ Z induced by the
element (2, 1).
6 Cayley’s Theorems

1.2 Graphs
One of the key insights into the study of groups is that they can be
viewed as symmetry groups of graphs. We refer to this as “Cayley’s
Better Theorem,” which we prove in Section 1.5.2. In this section we
establish some terminology from graph theory, and in the following
section we discuss groups acting on graphs.

Definition 1.9. A graph Γ consists of a set V (Γ) of vertices and a

set E(Γ) of edges, each edge being associated to an unordered pair of
vertices by a function “Ends”: Ends(e) = {v, w} where v, w ∈ V . In
this case we call v and w the ends of the edge e and we also say v and
w are adjacent.
We allow the possibility that there are multiple edges with the same
associated pair of vertices. Thus for two distinct edges e and e
it can be
the case that Ends(e) = Ends(e
). We also allow loops, that is, edges
whose associated vertices are the same. Graphs without loops or multiple
edges are simple graphs.

Graphs are often visualized by making the vertices points on paper and
edges arcs connecting the appropriate vertices. Two simple graphs are
shown in Figure 1.3; a graph which is not simple is shown in Figure 1.4.

Fig. 1.3. The complete graph on five vertices, K5 , and the complete bipartite
graph K3,4 .

There are a number of families of graphs that arise in mathematics.

The complete graph on n vertices has exactly one edge joining each pair
of distinct vertices, and is denoted Kn . At the opposite extreme are the
null graphs, which have no edges.
A graph is bipartite if its vertices can be partitioned into two subsets –
by convention these subsets are referred to as the “black” and “white”
 1.2 Graphs 7

vertices – such that, for every e ∈ E(Γ), Ends(e) contains one black
vertex and one white vertex. The complete bipartite graphs are simple
graphs whose vertex sets have been partitioned into two collections, V◦
and V• , with edges joining each vertex in V◦ with each vertex in V• . If
|V◦ | = n and |V• | = m then the corresponding complete bipartite graph
is denoted Kn,m .
The valence or degree of a vertex is the number of edges that contain
it. For example, the valence of any vertex in Kn is n − 1. If a vertex
v is the vertex for a loop, that is an edge e where Ends(e) = {v, v},
then this loop contributes twice to the computation of the valence of v.
For example, the valence of the leftmost vertex in the graph shown in
Figure 1.4 is six.
A graph is locally finite if each vertex is contained in a finite number
of edges, that is, if the valence of every vertex is finite.
An edge path, or more simply a path, in a graph consists of an alter-
nating sequence of vertices and edges, {v0 , e1 , v1 , . . . , vn−1 , en , vn } where
Ends(ei ) = {vi−1 , vi } (for each i). A graph is connected if any two ver-
tices can be joined by an edge path. In Figure 1.4 we have indicated an

b
c i
a
1 d 2 3
h j
e 4 f 5
g

Fig. 1.4. On top is a graph which is not simple, with its vertices labelled by
numbers and its edges labelled by letters. Below is the set of edges traversed in
an edge path, joining the vertex labelled 1 to the vertex labelled 3, is indicated.

edge path from the leftmost vertex to the rightmost vertex. If vi is the
vertex labelled i and eα is the edge labelled α, then this path is:
{v1 , ea , v1 , ee , v4 , eg , v5 , eh , v2 , eb , v1 , ed , v2 , ei , v3 }
8 Cayley’s Theorems

This notation is obviously a bit cumbersome and is only feasible when

the graphs are small and the paths are short. If the graph is a simple
graph, then one really only needs to list the sequence of vertices, which
is a small economy of notation. In general, we will not need this level of
specificity in dealing with edge paths in graphs.
A backtrack is a path of the form {v, e, w, e, v} where one has trav-
elled along the edge e and then immediately returned along e. A path
is reduced if it contains no backtracks.
A cycle or circuit is a non-trivial edge path whose first and last vertices
are the same, but no other vertex is repeated. The following paths in
the graph shown in Figure 1.4 are all cycles:
1. {v1 , ee , v4 , ef , v2 , ec , v1 },
2. {v1 , eb , v2 , ed , v1 },
3. {v1 , ea , v1 }.
In Chapter 3 we study various groups that are closely connected to
trees. A tree is a connected graph with no cycles. If you have not encoun-
tered trees in a previous course, working through the following exercise
will help you gain some intuition for trees.
Exercise 1.10. Prove that the following conditions on a connected
graph Γ are equivalent.
1. Γ is a tree.
2. Given any two vertices v and w in Γ, there is a unique reduced
edge path from v to w.
3. For every edge e ∈ E(Γ), removing e from Γ disconnects the
graph. (Note: Removing e does not remove its associated ver-
tices.)
4. If Γ is finite then #V (Γ) = #E(Γ) + 1.
While there are a number of interesting results about finite trees, in
this book we shall be mainly interested in infinite trees. In particular,
in Chapter 3 we explore groups that act on certain infinite, symmetric
trees.
Definition 1.11 (regular and biregular trees). A regular m-tree is a tree
where every vertex has fixed valence m. For a given value of m there is
only one regular m-tree, which we denote Tm . Notice that, since every
vertex has valence m, the tree Tm is infinite when m ≥ 2.
A graph is biregular if it is bipartite, and all the vertices in one class
have fixed valence m and all the vertices in the other class have fixed
 1.2 Graphs 9

valence n. Thus, for example, the complete bipartite graphs are all bireg-
ular. Given the valences, there is a unique biregular tree, which we denote
by Tm,n . You can see an example in Figure 1.5.

Fig. 1.5. The biregular tree T2,3 .

It is often convenient to think of graphs as geometric objects where

each edge is identified with the unit interval [0, 1], where 0 corresponds
to one of the associated vertices and 1 to the other. This convention
allows us to refer to midpoints of edges, for example. The graphs shown
in Figures 1.3, 1.4, and 1.5 have distorted this metric, meaning that if
you measured the lengths of the edges in a given figure, those lengths will
not all be the same. (This will occur in almost all of the graphs drawn
in this book.) In addition to allowing us to specify points on edges, this
geometric perspective allows us to think of paths as parametrized curves.
Our convention will be that, given an edge path
ω = {v0 , e1 , v1 , e2 , . . . , en , vn }
in a graph Γ, there is an associated function pω : [0, 1] → Γ where
pω (i/n) = vi and pω is linear when restricted to each of the subintervals
[i/n, (i + 1)/n].
Remark 1.12. Our view of graphs is more topological than combina-
torial. The reader who is a bit uneasy about our thinking of graphs as
geometric objects, where edges have lengths, might want to look up the
10 Cayley’s Theorems

definition of CW complexes, as we are viewing graphs as 1-dimensional

CW complexes. We have not introduced this terminology or explicitly
used the associated definition as it requires an understanding of topo-
logical spaces and quotient topologies.
There is one final variation on graphs that we will encounter in this
text:
Definition 1.13. A directed graph consists of a vertex set V and an edge
set E of ordered pairs of vertices. Thus each edge has an initial vertex
and a terminal vertex. Graphically this direction is often indicated via
an arrow on the edge. In thinking of directed graphs geometrically we
assume that the initial vertex is identified with 0 ∈ [0, 1] and the terminal
vertex with 1 ∈ [0, 1].
If we say a directed graph is connected we mean the underlying undi-
rected graph is connected. (One can study “directed-connectedness” but
that will not be relevant for us.)
In addition to directions on the edges, there are other sorts of dec-
orations one can add to a graph. For example, one can have a set of
labels L and a function V : V (Γ) → L that provides a labelling of the
vertices. Or one could label the edges via E : E(Γ) → L, where the set
of labels might be the same or different than the labels for the vertices.
As an example, in Figure 1.4 we have shown a labelled graph where the
vertices have been labelled with numbers and the edges with lower case
letters.

1.3 Symmetry Groups of Graphs

Many important finite groups arise as symmetry groups of geometric
objects. The dihedral groups are the symmetry groups of regular n-gons;
the symmetric group Symn is isomorphic to the symmetry group of the
regular (n − 1)-dimensional simplex, for example the convex hull of
{(1, 0, 0, . . . , 0), (0, 1, 0, . . . , 0), . . . , (0, . . . , 0, 1)} ⊂ Rn ;
the alternating subgroup An is isomorphic to the subgroup of symmetries
of the regular (n − 1)-dimensional simplex consisting of rotations in Rn .
In this section we explore a similar theme, namely, we explore sym-
metry groups of graphs.
Definition 1.14. A symmetry of a graph Γ is a bijection α taking ver-
tices to vertices and edges to edges such that if Ends(e) = {v, w}, then
 1.3 Symmetry Groups of Graphs 11

Ends(α(e)) = {α(v), α(w)}. The symmetry group of Γ is the collection

of all its symmetries. We denote this group by Sym(Γ).2
Example 1.15. The symmetry group of the complete graph Kn is iso-
morphic to Symn .

Fig. 1.6. The graph for Example 1.16.

Example 1.16. The symmetry group of the graph Γ shown in Figure 1.6
is isomorphic to Sym3 ⊕ Z2 (or Sym3 ⊕ Sym2 if you prefer the symmetry
in notation). The symmetries come from permuting the multiple edges
joining the leftmost vertices of Γ (which gives the Sym3 factor) and from
a reflection through a horizontal axis (which gives the Sym2 factor). The
fact that these two types of symmetries commute with each other leads
to the direct product structure.
Definition 1.17. Let G be a subgroup of Sym(Γ). Then G is vertex
transitive if, given any two vertices, v and v
, there is an α ∈ G where
α(v) = v
. The symmetry groups of the complete graphs are all vertex
transitive, while the symmetry group of a complete bipartite graph Kn,m
is not vertex transitive when n = m. (Why?)
A group G < Sym(Γ) is edge transitive if, given any two edges, e and
e
, there is a symmetry α ∈ G where α(e) = e
.
A flag consists of a pair (v, e), where v ∈ Ends(e). The group G is flag
transitive if, given any two flags, (v, e) and (v
, e
), there is a symmetry
α ∈ G where α(v) = v
and α(e) = e
.
The symmetry group G acts simply transitively on the vertices (edges,
flags, etc.) if it is vertex transitive, and given any two vertices v and v

2 Graph theorists usually refer to automorphisms of a graph, and call the collection
of all automorphisms the automorphism group of the graph.
12 Cayley’s Theorems

there is a unique α ∈ G such that α(v) = v
(respectively, α(e) =

e
, etc.)

Fig. 1.7. The graph for Example 1.18.

Example 1.18. Let Γ be the vertices and edges of an octahedron, as in

Figure 1.7. The vertices of Γ can be partitioned into three pairs of op-
posite vertices. These pairs are: the top–bottom pair (containing the
vertex labelled u in Figure 1.7); the front–back pair (containing v);
and the left–right pair (containing w). Using reflections one can
exchange the two vertices in the top–bottom, front–back, or left–right
pair, leaving the other four vertices fixed. Direct inspection shows that
any two such reflections commute, so
3
(Z2 ) = Z2 ⊕ Z2 ⊕ Z2 < Sym(Γ).

The action of this subgroup is not transitive on the vertices of Γ. How-

ever, given any two sets of opposing vertices, there is a reflection that
exchanges those two sets, which fixes the remaining pair of vertices.
These reflections generate a subgroup that permutes the pairs of oppo-
site vertices, giving a subgroup isomorphic to Sym3 . Let the subgroup
of Sym(Γ) generated by this copy of Sym3 and the copy of (Z2 )3 be
denoted H. (H is the collection of all possible products of elements of
Sym3 and (Z2 )3 .) This subgroup is vertex transitive. To move, for ex-
ample, the vertex labelled u, to vertex v, first exchange the top–bottom
pair of vertices with the front–back pair of vertices. Then, if necessary,
 1.3 Symmetry Groups of Graphs 13

one can exchange the front–back pair in order to have the image of u
be v.
The subgroup H described above is in fact the entire group of sym-
metries. To show this, let α ∈ Sym(Γ) be an arbitrary symmetry. Once
we know where α moves the vertices u, v, and w, we have determined α.
Since the action of H is vertex transitive, there is an a ∈ H that takes
u to α(u). Thus a−1 α is a symmetry that fixes u. If a−1 α exchanges the
front–back pair of vertices with the left–right pair of vertices, then there
is a b ∈ H that fixes the top–bottom pair of vertices and also exchanges
the front–back pair of vertices with the left–right pair of vertices. Taking
b to be the identity if a−1 α did not make this exchange, one sees that
b−1 a−1 α fixes u and takes the front–back pair of vertices to itself, and
similarly with the left–right pair. Thus there is a c ∈ (Z2 )3 < H such
that c−1 b−1 a−1 α fixes u, v, and w. But then c−1 b−1 a−1 α = e, hence
α = abc ∈ H.3
Once one has this description of Sym(Γ), it is not difficult to argue
that not only is Sym(Γ) vertex transitive, it is flag transitive.

Fig. 1.8. This is the Petersen graph. A 4-vertex subgraph, a null graph, has
been highlighted.

Example 1.19. Let Γ be the Petersen graph, shown in Figure 1.8.

Exploiting the symmetry present in this particular picture, one sees that
D5 < Sym(Γ). There is, however, much more symmetry in this graph
than is immediately apparent from the figure.
3 If you have already encountered semi-direct products of groups, then you should be
able to show Sym(Γ) ≈ (Z2 )3
Sym3 . We define semi-direct products in Chapter 8.
14 Cayley’s Theorems

Exercise 1.20. The following questions explore the symmetries of the

Petersen graph.
a. Prove that the symmetry group of Γ is vertex transitive by finding
a symmetry that exchanges the set of “inner” vertices with the
“outer” vertices.
b. Prove that Sym(Γ) is edge transitive as well. Is it also flag tran-
sitive?
c. A 4-vertex subgraph, a null graph, is highlighted in Figure 1.8.
There are five such subgraphs. Find them all. (Hint: Each one
is determined by picking a pair of non-adjacent vertices on the
“outer” ring.)
d. Use the null graphs to construct a homomorphism
φ : Sym(Γ) → Sym5
e. Show that φ is onto.
f. Show that the kernel of φ is trivial, hence, by the First Isomor-
phism Theorem, Sym(Γ) ≈ Sym5 .
We now turn to a discussion of particular subgroups of symmetry
groups of graphs.
Lemma 1.21. If Γ is a directed graph, the collection of all symmetries
of Γ that preserve every edge’s direction forms a subgroup of Sym(Γ).
Proof. In order to prove a subset of a group is a subgroup, one only
needs to show that the subset is closed under products and inverses.
It is clear that if g and h preserve edge directions then so does their
product and so do their inverses.
Essentially the same comments establish the following result.
Lemma 1.22. If the edges and/or vertices of Γ are labelled, the collec-
tion of all symmetries that preserve the edge and/or vertex labels forms
a subgroup of Sym(Γ).
Definition 1.23. If Γ comes with certain declared decorations – such
as directed edges, labelled or colored vertices and/or edges, etc. – let
Sym+ (Γ) denote the subgroup of Sym(Γ) that preserves all of the de-
clared decorations.
Exercise 1.24. Consider the collection of complete bipartite graphs,
Kn,m . Let Sym+ (Kn,m ) be the group of symmetries of Kn,m that pre-
serve the bipartite structure. That is, if g ∈ Sym+ (Kn,m ), then g takes
 1.4 Orbits and Stabilizers 15

white vertices to white vertices and black vertices to black vertices. (This
is described more symbolically by writing g · V◦ = V◦ and g · V• = V• .)
Prove that Sym+ (Kn,m ) = Symn ⊕ Symm . What can you say about the
full symmetry group Sym(Kn,m )?
(Hint: It matters whether or not n = m.)

1.4 Orbits and Stabilizers

In this section we introduce terminology that gets used quite frequently
when studying group actions.
Definition 1.25. Let X be some mathematical object (perhaps X is a
set, a graph, a regular polygon, ...), and let G act on X. If x ∈ X then
the stabilizer of x is
Stab(x) = {g ∈ G | g · x = x}.
Since composing two symmetries that fix x yields a symmetry that
still fixes x, and similarly the inverse of a symmetry fixing x must fix x,
we have:
Lemma 1.26. For any G
X, and any x ∈ X, Stab(x) is a subgroup
of G.

Fig. 1.9. The dihedral group D6 acts on the regular hexagon. The stabilizer
of any grey vertex is isomorphic to Z2 . The stabilizer of the center point is all
of D6 . Any point which is not indicated and which is not on one of the dotted
lines, has a trivial stabilizer.

Example 1.27. The dihedral group Dn acts on a regular n-gon. If x is a

vertex of this n-gon, or the exact middle of any edge, then Stab(x) ≈ Z2 .
16 Cayley’s Theorems

Notice, however, that if x and x
are two vertices that are not directly
opposite each other, then Stab(x) = Stab(x
). If x is the center of the
n-gon then Stab(x) ≈ Dn . If x is any point that is not a vertex, the
center of an edge, or on a line connecting the center of the polygon to a
vertex or the center of an edge, then Stab(x) = {e}.

Example 1.28. Let Symn
[n] = {1, 2, . . . , n}. Then Stab(i) ≈

Symn−1 for any i ∈ [n].

Definition 1.29. Let G
X. If for some x ∈ X Stab(x) = {e}, then x

is said to be moved freely by the action of G. The action G
X is free
if for every x ∈ X, Stab(x) = {e}.

Example 1.30. Let R be the subgroup of rotations inside of Dn . If P

is a regular n-gon, let ∂P be its boundary, consisting of n vertices and
n edges. The action of R on ∂P is free, while the action of R on P is
not free.

Definition 1.31. Let G
X and let x ∈ X. Then the orbit of x is

Orb(x) = {g · x | g ∈ G}.

Example 1.32. The orbit of any vertex under the action of Dn on a

regular n-gon P is the set of all vertices. Similarly if x is the midpoint
of any edge in P then the orbit of x is the set of all midpoints of edges.

Fix an action of a group G on some X, and a particular x ∈ X.

Then g · x = h · x if and only if g −1 h · x = x, that is, if and only if
g −1 h ∈ Stab(x). This implies that g · x = h · x if and only if the left
cosets g · Stab(x) and h · Stab(x) are the same. This short argument
establishes:

Theorem 1.33. Given an action G
X and any x ∈ X, there is a bi-

jective correspondence between the set Orb(x) and left cosets of Stab(x),
given by g · x ↔ g · Stab(x).

A corollary of this result, which is often useful in computing the orders

of finite groups, is the Orbit-Stabilizer Theorem:

Corollary 1.34 (Orbit-Stabilizer Theorem). Let G be a finite group

acting on X. Then, for any x ∈ X,

|G| = |Stab(x)| · |Orb(x)| .

 1.5 Generating Sets and Cayley Graphs 17

Proof. The order of G is the product of the order of Stab(x) and the
index of Stab(x) in G. But the index of Stab(x) is the number of left
cosets of Stab(x), which by Theorem 1.33 is the size of the orbit of x.

Example 1.35. Consider the graph Γ shown in Figure 1.7. It is not

difficult to argue that the action of Sym(Γ) is vertex transitive, hence for
any vertex v, Orb(v) consists of all six vertices. It is also not too difficult
to argue that Stab(v) ≈ D4 for any vertex v. Thus |Sym(Γ)| = 8·6 = 48.

Another insight provided by Theorem 1.33 is the following method of

identifying elements in a group with the elements in the orbit of an x
where Stab(x) is trivial.

Corollary 1.36. Let G
X. For any x ∈ X, where Stab(x) = e,

there is a one-to-one correspondence between the elements of G and the
elements of Orb(x).

1.5 Generating Sets and Cayley Graphs

We are now in a position to present a useful extension of Cayley’s Basic
Theorem that applies to finitely generated groups.

1.5.1 Generators
Definition 1.37. If G is a group and S is a subset of elements, then
S generates G if every element of G can be expressed as a product of
elements from S and inverses of elements of S. A group G is finitely
generated if it has a finite generating set.

Example 1.38. Every finite group is finitely generated; take S = G

and you satisfy the definition. However, it is sometimes useful to know a
natural, proper subset of the elements that generate the whole group. For
example, it is often demonstrated in an introductory group theory course
that the adjacent transpositions S = {(12), (23), . . . , (n−1 n)} generate
Symn . Another well-known generating set consists of a transposition and
an n-cycle: S = {(12), (12 · · · n)}.

Example 1.39. One may be tempted to think of generating sets as

analogous to bases of a vector space. This intuition is not horrible, but
it is not perfect either. For example, the group of integers is generated
by the single element 1 ∈ Z. One can form a larger and redundant
set such as S = {1, 2}, which also forms a generating set. Of course
18 Cayley’s Theorems

this set can be “trimmed” to form a minimal generating set, just as a

set of vectors which spans a vector space can be “trimmed” to a basis.
However, there are other sets such as {2, 3} that generate Z and that can
not be made smaller by simply removing a single element. Similarly one
can show that {6, 10, 15} is a generating set for Z, but no proper subset
generates Z.

Example 1.40. Any element of the dihedral group of order 2n, Dn ,

is determined by where it takes two adjacent vertices of the associated
regular n-gon. We use this fact to establish that Dn is generated by a
rotation and a reflection.
Let • and ◦ denote adjacent vertices of the n-gon (as in the hexagon
of Figure 1.10) and let g be any element of Dn . The element g takes • to
g(•) and ◦ to g(◦). Let ρ be a rotation – clockwise or counterclockwise,
it is your choice – through an angle of 2π/n. Let φ be the reflection that
fixes •. Then there is a number m such that ρm (•) = g(•). If it happens
that ρm (◦) = g(◦), we are done, since in this case, g = ρm . Said another
way, since g −1 ρm fixes • and ◦, g −1 ρm = e ⇒ ρm = g. Otherwise
the element ρm · φ takes • → g(•) and ◦ → g(◦). Thus g −1 ρm φ = e
and therefore ρm φ = g. Since g was arbitrary, we have shown that any
element of Dn can be expressed as a product of ρ’s and φ’s.
One can also generate Dn using two adjacent reflections. Let φ be
as above and let ψ be a reflection of the n-gon that forms an angle of
π/n with φ. Then φ · ψ is a rotation through an angle of 2π/n. By the
previous argument, φ and the product φ · ψ generate Dn , hence φ and
ψ generate Dn .

Fig. 1.10. A regular hexagon with two distinguished adjacent vertices.

 1.5 Generating Sets and Cayley Graphs 19

As you can see in the examples above, a given group can have dif-
ferent, interesting, and fairly distinct generating sets. Although it is a
guideline that is frequently broken, it is best to not refer to a particular
set of generators as the generators of a group. Authors often circumvent
this issue using phrases such as “the natural generating set” or “the
generating set we are considering.”
We end this list of examples with a group where it is easy to show
that the group admits no finite set of generators.

Example 1.41. The group of rational numbers under addition is not

finitely generated.

 assume to the contrary that Q is generated
To see this,
n1 n2 nk
by S = , ,..., . If d is the least common multiple of the
d1 d2 dk
denominators, that is d = LCM[d1 , d2 , . . . , dk ], then every element that
can be written as a sum of the supposed generators and their inverses
n
can be expressed as for some n ∈ Z. But not every element in Q has
d
such an expression.

1.5.2 Cayley’s Better Theorem

The improvement of Cayley’s Basic Theorem is:

Theorem 1.42 (Cayley’s Better Theorem). Every finitely generated

group can be faithfully represented as a symmetry group of a connected,
directed, locally finite graph.

Just as in Cayley’s Basic Theorem it is hard, initially, to see how one

might prove Cayley’s Better Theorem. One has a group G and some finite
generating set S, and somehow out of whole cloth one must produce both
a graph and an action of G on that graph.

Proof of Cayley’s Better Theorem. We construct a graph ΓG,S using the

group G and the generating set S. The vertices of ΓG,S are the elements
of G. For each g ∈ G and each s ∈ S form a directed edge with initial
vertex g and terminal vertex g ·s. Thus edges of ΓG,S correspond to right
multiplication by elements of the generating set S.
To show that the graph is connected it suffices to show that one can
get from the vertex corresponding to the identity e to any other ver-
tex g. (Explain to yourself why this is sufficient!) Since S is a genera-
ting set, we can write g as a product of generators and their inverses
g = s1 s2 · · · sn , where each si is in S or S −1 . In the graph ΓG,S there is
20 Cayley’s Theorems

an associated path:
1s 2 3s n s s
• −→• −→ • −→ · · · −→ •
e s1 s1 s2 s1 s2 ···sn =g

Here we have been sloppy, and have not bothered to take orientation into
account. If any of the letters si in the expression g = s1 s2 · · · sn is the
inverse of an element of S, then one would be travelling opposite to the
orientation of the directed edge joining s1 s2 · · · si−1 to s1 s2 · · · si−1 si .
To avoid confusion between elements of G and vertices in ΓG,S , let
the vertex in ΓG,S associated to g be denoted vg .
The graph ΓG,S is locally finite because the set S is finite. In fact,
each and every vertex will be incident with exactly 2|S| edges. Then
vg is the initial vertex of exactly one directed edge labelled s for each
s ∈ S (going from vg to vg·s ) and it is the terminal vertex of exactly one
directed edge labelled s for each s ∈ S (going from vg·s−1 to vg ).
The proof of Cayley’s Basic Theorem describes the left action of G
on the vertices of this graph: the element g ∈ G sends the vertex vh
to vg·h .
Does this action on vertices extend to an action on the edges of ΓG,S ?
Notice that the vertex vh is joined to vh·s by a directed edge labelled s.
The element g ∈ G takes the vertex vh to vg·h and the vertex vh·s to
vg·h·s . Thus we may define the action of g ∈ G on the edges by stating
the edge labelled s joining vh to vhs is sent to the edge labelled s joining
vgh to vghs . (See Figure 1.11.)

hs ghs

s s

h gh

Fig. 1.11. The action of an element g ∈ G on the vertices of ΓG,S extends to

an action on the edges.

We note that it is precisely because the group action is on the left,

while the edges are defined in terms of right multiplication, that the
action of G on the vertices can be extended to the edges of ΓG,S . In
other words: if s is a generator, then there is an edge labelled s joining
vh to vhs ; if we left multiply by g ∈ G then vh and vhs go to vgh and vghs ,
 1.5 Generating Sets and Cayley Graphs 21

respectively; but then by definition there is an edge labelled s joining

vgh to vghs . This is illustrated in Figure 1.11.

Definition 1.43. Given a group G and a finite generating set S, call

the directed graph ΓG,S the Cayley graph of G with respect to S, or just
the Cayley graph when the group G and generating set S are sufficiently
clear from context.

We note that the construction of the action of G on the Cayley graph

ΓG,S produces an action that preserves the orientation of the directed
edges and the labellings on those edges. Thus G < Sym+ (ΓG,S ).

Example 1.44. In Figure 1.12 we show the Cayley graph of Sym3 with
respect to the generating set {(12), (123)}. The version on the left follows
the definition precisely. However, we notice that when a generator is an
involution (an element of order 2) there is little need to have a “parallel”
pair of directed edges. It is a standard convention to replace the two
directed edges with a single undirected edge, as is done in the figure on
the right. While this is not what is described by the definition of a Cayley
graph, it is often more aesthetically pleasing and easier to understand.
When this convention is employed, the claim that every vertex is of
degree 2|S| is no longer true. Further, the graph is then a strange hybrid
of directed and undirected edges.

(12) (12)

e e

(132) (123) (132) (123)

(23) (13) (23) (13)

Fig. 1.12. The Cayley graph of Sym3 with respect to {(12), (123)}, drawn
using two different conventions on how to treat generators of order 2.

Exercise 1.45. You may have noticed that Figures 1.1 and 1.12 are
quite similar. However, the vertices corresponding to group elements
have been slightly shifted. Figure out what is similar and what is different
22 Cayley’s Theorems

in these two figures. Along the way, determine what the action of Sym3
on its Cayley graph looks like.

Remark 1.46. Arthur Cayley introduced Cayley graphs in a paper pub-

lished in 1878 [Ca78]. In that paper he mentions that when there are two
or more generators “we require different colours, the [edges] belonging
to any one substitution [i.e. generator] being of the same colour.” It is
often quite helpful, if you own enough colored pens or pencils, to draw
Cayley graphs using this sort of color convention.

1.6 More Cayley Graphs

The best way to understand a new definition is to look at a few non-
trivial examples.

1.6.1 Dihedral Groups

Let Dn be the dihedral group of order 2n. As was discussed in Exam-
ple 1.40, there are two standard generating sets for Dn : the set consisting
of a reflection φ and a minimal rotation ρ or a set consisting of two reflec-
tions {φ, ψ} where the angle between the fixed lines of these reflections
forms an angle of π/n. The corresponding Cayley graphs, in the case
where n = 4, are shown in Figure 1.13.

Fig. 1.13. Two Cayley graphs of D4 . On the left is the graph with respect
to a generating set consisting of a rotation and a reflection; on the right the
generating set consists of two adjacent reflections.

One can construct these Cayley graphs directly from the action of
Dn on a regular n-gon. In Figure 1.14, on the left, we have chosen a
 1.6 More Cayley Graphs 23

point whose stabilizer is trivial, so that the elements in the orbit of this
point correspond to the six elements of D3 (Corollary 1.36). We have
also indicated two particular reflections to be used as the generating
set of D3 . One can then connect the points in the orbit to form the
Cayley graph with respect to the indicated generators. We start with
our original point, the only non-solid point in the figure, and apply the
two generators to it, in order to get the intermediate figure. If ◦ denotes
our original point, then the top-right point is a · ◦. By definition, in the
Cayley graph this point is joined by a dashed edge to the vertex ab · ◦.
Noticing that ab is a clockwise rotation through an angle of 120◦ , one
can begin to verify that the picture shown on the right in Figure 1.14 is
indeed accurate.

Fig. 1.14. The dihedral group D3 is generated by the reflections a and b. The
associated Cayley graph can be constructed by connecting points in an orbit
of a point in “general position.”

This particular example, where n = 3, is easily generalized to the case

where one is using two adjacent reflections to generate Dn .
Exercise 1.47. Follow a procedure similar to the one outlined above to
find an embedding of the Cayley graph of Dn , with respect to a reflection
and a rotation through an angle of 2π/n, inside a regular n-gon. You
will want to make a judicious choice for the initial point whose orbit
corresponds with the elements of Dn .

1.6.2 Symmetric Groups

At the end of the previous section we produced the Cayley graph of
Sym3 with respect to the generating set {(12), (123)}. Here we will ex-
amine the Cayley graph of Sym4 with respect to the generating set
{(12), (23), (34)}. This Cayley graph has 4! = 24 vertices, where each ver-
tex is incident to three edges, using the convention that involutions are
24 Cayley’s Theorems

represented by unoriented edges. The products (12)(23) = (123) and

(23)(34) = (234) are both of order 3, while the transpositions (12) and
(34) commute. It follows that at each vertex of Γ the edges correspond-
ing to (12) and (23) will determine a 6-cycle, the edges corresponding to
(23) and (34) give another 6-cycle, while the edges corresponding to (12)
and (34) determine a 4-cycle. Thus at each vertex of this Cayley graph
one sees two hexagons and a square. This is enough data to determine
that this Cayley graph has the form indicated in Figure 1.15.

Fig. 1.15. The Cayley graph of the symmetric group Sym4 with respect to the
generating set S = {(12), (23), (34)}.

If you have studied the Archimedean solids you might notice that this
Cayley graph looks like the edges of a truncated octahedron. This is no
accident. The following argument – similar to the one that embedded
Cayley graphs of dihedral groups into the interior of regular n-gons –
explains why.
First, represent Sym4 as the group of symmetries of the regular tetra-
hedron T. To do this, identify the four vertices of T with the numbers
[4] = {1, 2, 3, 4}. Any symmetry of T results in a permutation of [4],
which is how one constructs a homomorphism φ : Sym(T) → Sym4 .
One can prove that φ is onto by showing that the three transpositions
(12), (23) and (34) are in the image. The reflection across the plane
spanned by the edge joining vertex 3 to vertex 4, and the midpoint
of the edge joining 1 to 2, is taken by φ to the transposition (12). (See
 1.6 More Cayley Graphs 25
1 1

4 4

3 3

2 2

Fig. 1.16. The group of symmetries of a regular tetrahedron is isomorphic

to Sym4 . On the left, the shaded portion indicates the fixed points for the
reflection of T corresponding to the transposition (12). On the right, portions
of the fixed point set for reflections corresponding to (12), (23) and (34) are
indicated by dashed lines.

Figure 1.16.) Similar reflections yield the generating set for Sym4 .
Because the two groups both have order 24, and φ is onto, it must
be the case that φ is an isomorphism.
To draw the Cayley graph with respect to S = {(12), (23), (34)} we
pick a point in T whose stabilizer is trivial. By drawing in some of the
fixed sets for the reflections corresponding to these transpositions, we
see a triangular region on the exterior of the tetrahedron, shown on the
right in Figure 1.16, any of whose points fit the bill. Picking a point, and
applying the reflections corresponding to (12) and (23) we see a hexag-
onal portion of this Cayley graph. (This subgraph illustrates the fact
that this subgroup is isomorphic to D3 .) Using the same point and the
reflections corresponding to (12) and (34) we see a rectangular portion
of the Cayley graph. These are highlighted in Figure 1.17. The hexagon
corresponding to the generators (23) and (34) is not shown in this figure,
but it sits just a bit below the vertex labelled 1. From these facts it is
not hard to establish that the Cayley graph we are constructing has its
underlying structure coming from the edges of an Archimedean solid.

Exercise 1.48 (hard!). Give a description of the Cayley graph of Symn ,

with respect to the generating set S = {(12), (23), . . . (n−1 n)}, for all
n ≥ 5.

Remark 1.49 (Drawing Trick). In the previous examples we used a

common technique for discovering the structure of – also known as
26 Cayley’s Theorems
1 1

4 4

3 3

2 2

Fig. 1.17. Parts of the Cayley graph of Sym4 shown embedded in the tetra-
hedron.

“figuring out how to draw” – a Cayley graph. We started with an action

of the group we were interested in on a geometric object (Dn acting on
a regular n-gon and Sym4 acting on the tetrahedron), then we found a
point in that space that is moved freely by the group action. By Corol-
lary 1.36, there is a bijection between the orbit of this point under the
action of G and the elements of G. Thus we may use the orbit of the point
as the vertex set of the Cayley graph. Our point was sufficiently well-
chosen that the edges ended up embedding nicely in the space. While
this method is not guaranteed to be useful in all situations, it is quite
often helpful for visualizing Cayley graphs.

1.6.3 The Symmetry Group of a Cube

Let
denote a cube, and let G = Sym(
) be the group of all symmetries
of
. In order to be very concrete, let
be thought of as the convex hull
of the eight points in R3 whose coordinates are of the form (±1, ±1, ±1),
and let Sym(
) consist of all reflections and rotations that take
back
to itself.
The order of G can be computed using the Orbit-Stabilizer Theorem
(Corollary 1.34). Let v be the vertex at (1, 1, 1). Then the orbit of v
consists of the eight vertices of
. Since any element in the stabilizer of
v permutes the three edges incident with v, and no non-trivial symmetry
fixes these three edges, we have Stab(v) ≈ Sym3 . Thus

|G| = |Orb(v)| · |Stab(v)| = 8 · 6 = 48.

 1.6 More Cayley Graphs 27

The insight behind the Orbit-Stabilizer Theorem also gives us a

method for finding a nice set of generators for G. Reflections in the
coordinate planes are all that are needed to move v to any other vertex
of
. Reflections in the appropriate planes of the form x = ±y, x = ±z,
and y = ±z generate Stab(v) for any vertex v. We can put all of these
reflecting planes into a picture of the cube, as in Figure 1.18. This shows
the surface of the cube has been dissected into 48 triangles. Fix any one
of these triangles. It can be checked by hand that the three reflections in
the sides of the chosen triangle generate G. (This is an idea we discuss
a bit more formally in Section 1.8.)

Fig. 1.18. A cube with the fixed planes for reflections indicated.

To find the Cayley graph, we use the Drawing Trick, using any point
inside our chosen triangle. The reader should check the following claims:
The portion of the Cayley graph around each vertex of
forms a
hexagon; each edge of
has a corresponding square in the Cayley graph;
and each face of
has an octagon corresponding to the D4 stabilizer of
that face. From this one sees that the Cayley graph is the edge set of
the great rhombicuboctahedron (see Figure 1.19).

1.6.4 Free Abelian Groups

Introductions to group theory rarely feature extended field trips to the
Zoo of Infinite Groups. The few times that they do venture into the Zoo,
28 Cayley’s Theorems

Fig. 1.19. A Cayley graph of the symmetry group of a cube.

they stay well away from the more exotic exhibits. We will encounter
some of these later in the text but, for now, let’s just look at one of the
basic beasts.

Fig. 1.20. Two Cayley graphs of Z. The top is with respect to S = {1}, the
bottom with respect to S
= {2, 3}.

The Cayley graph of Z with respect to the single cyclic generator {1}
is a combinatorial version of the real line. If one uses a generating set
such as {2, 3} the Cayley graph is a bit more complicated, but somehow
it still seems to keep the basic “long and skinny” shape of the real line.
(See Figure 1.20.)
If you generate the free abelian group Zn using “the standard basis
vectors”

S = {(1, 0, 0, . . . , 0), (0, 1, 0, . . . , 0), . . . , (0, 0, 0, . . . , 1)}

the resulting Cayley graph consists of the edges of a cubic lattice in Rn .

The vertices correspond to points whose coordinates are all integers; the
 1.7 Symmetries of Cayley Graphs 29

Fig. 1.21. The Cayley graph of Z ⊕ Z with respect to the generators

{(1, 0), (0, 1)}.

edges connect points that are exactly a distance 1 apart. The case where
n = 2 is shown in Figure 1.21.
Remark 1.50. Cayley restricted his attention to the Cayley graphs
of finite groups, and this restriction also occurs in many elementary
presentations of Cayley graphs. In 1912 Max Dehn demonstrated the
utility of Cayley graphs, and a related construction called Gruppenbilder,
in the study of finitely generated infinite groups [De12]. This work of
Max Dehn is often cited as the start of the combinatorial and geometric
study of finitely generated infinite groups.

1.7 Symmetries of Cayley Graphs

Having defined symmetry groups of graphs, one might wonder if there is
any nice description of the groups that arise as the symmetry groups of
graphs. The following theorem, first proved by Frucht in 1938, indicates
that all finitely generated groups can be realized as label and orientation
preserving symmetries of locally finite, directed graphs. In Exercise 20,
at the end of this chapter, you are asked to show even more: Every
finitely generated group is realizable as the full symmetry group of a
locally finite graph. Thus knowing that a group G acts on a graph Γ
30 Cayley’s Theorems

does not immediately imply anything about the group G. On the other
hand, this result also shows that viewing groups via their actions on
graphs is a very flexible approach to the study of group theory.

Theorem 1.51. Let ΓG,S be the Cayley graph of a group G with respect
to a finite generating set S. Consider ΓG,S to be decorated with directions
on its edges and labellings of its edges, corresponding to the generating
set S. Then Sym+ (ΓG,S ) ≈ G.

Proof. The left action of G on ΓG,S given by Cayley’s Better Theorem

(1.42) shows that G → Sym(ΓG,S ). Since this is a left action, it does
not affect the directions or the labellings on the edges of ΓG,S , which
are defined in terms of right multiplication. Thus G → Sym+ (ΓG,S ).
To show that this map is surjective, we consider an arbitrary element
γ ∈ Sym+ (ΓG,S ). For any element g ∈ G let •g be the vertex in ΓG,S cor-
responding to g. Then there is a g such that γ(•e ) = •g . Considering g as
a symmetry of ΓG,S , the product γ · g −1 ∈ Sym+ (ΓG,S ). This symmetry
takes •e to •e . Further, it fixes all the edges arriving at or leaving from
e, as elements of Sym+ (ΓG,S ) preserve direction and labelling. Hence
γ · g −1 fixes all of the vertices adjacent to •e . But then, again, it will fix
edges incident to these vertices and so on. So the symmetry γ · g −1 is
the identity, which is to say, γ = g (as elements in Sym+ (ΓG,S )). Since
γ was an arbitrary symmetry of the Cayley graph, our homomorphism
is a bijection, and G ≈ Sym+ (ΓG,S ).

1.8 Fundamental Domains and Generating Sets

Let G be a group acting on a connected graph Γ. In this section we show
how one can form a fundamental domain for the action, and then given
a fundamental domain, how one can find a generating set for G. This
technique is one example of how an understanding of a group action on
some geometric object can result in information about the group itself.
In the lemma below we have used the adjective “closed.” This term
does have a precise mathematical meaning, which you may have encoun-
tered in a course on real analysis or topology. If you have not previously
encountered this word, you may take “closed” to mean “containing its
end points” as in the interval [0, 1] is a closed subset of R, but [0, 1)
and (0, 1) are not. Having a deep understanding of closed sets is not
necessary to follow the arguments given here.
 1.8 Fundamental Domains and Generating Sets 31

Lemma 1.52. If a group G acts on a connected graph Γ then there is

a subset F ⊂ Γ such that:

1. F is closed;
2. the set {g · F | g ∈ G} covers the graph Γ; and
3. no subset of F satisfies properties (1) and (2).

The subset F described in Lemma 1.52 is called a fundamental domain

for the action G
Γ. Two examples of fundamental domains are shown
in Figure 1.22. We establish Lemma 1.52 by building up a subset of Γ,
starting with a single vertex, and then adding to it until it is sufficiently
large so that its image under the action of G covers the entire graph Γ.
The actual proof of Lemma 1.52 is long, and is probably the most chal-
lenging argument presented in this first chapter. The reader is strongly
encouraged to use all of his or her favorite techniques for understanding
complicated arguments as they proceed through the details.

Fig. 1.22. A wheel graph Γ and two examples of fundamental domains for the
action of Z6 on Γ, the action being by rotations. Any fundamental domain
for this action must include the central vertex but, after that, there are many
choices as to how to construct a fundamental domain.

Proof of Lemma 1.52. To construct a fundamental domain, start with a

fixed vertex v ∈ Γ, and consider connected subgraphs C ⊂ Γ where

1. v ∈ C; and
2. if x and y are distinct vertices in C then there is no element g ∈ G
such that g · x = y.

The vertex v by itself satisfies these two conditions, so there are such
subgraphs in Γ. Let C0 ⊂ C1 ⊂ C2 ⊂ · · · be a sequence of subgraphs
satisfying our conditions, one properly contained in the next. If there
are only finitely many orbits of vertices under the action of G, and the
graph Γ is locally finite, then at some point one of the Ci ’s in the sequence
above will be a maximal subgraph satisfying the conditions above. In this
32 Cayley’s Theorems

case, call this subgraph the Core. Otherwise form a maximal subgraph
satisfying our conditions by defining

Core = Ci

where the sequence of Ci ’s is as large as it can possibly be.4 It is perhaps

less clear that Core satisfies our conditions (1) and (2) in this case. But
notice that:

1. v ∈ C0 ⊂ Core; and
2. if x and y are distinct vertices in Core then there must be some
Ci that contains them. Thus there is no g ∈ G taking x to y.

The subgraph Core does not necessarily form a fundamental domain

for the action of G on Γ, but it does contain all the vertices needed.

Claim: The image of Core under the action of G contains all the vertices
of Γ.
Assume to the contrary that there is a vertex v ∈ Γ that is not con-
tained in some g · Core, g ∈ G. Let {v = v0 , v1 , . . . , vn } be the vertices
in a minimal-length edge path joining v to a vertex in G · Core. It must
be the case that vn−1 is not in G · Core (since otherwise there would
be a shorter path to G · Core), which implies that vn−1 is a vertex
outside of G · Core which is joined to G · Core by a single edge. So
we may assume that v ∈ Γ \ G · Core and there is an edge e joining
v to some vertex in g · Core for some g ∈ G. (In other words, we can
now assume “vn−1 ” is “v.”) But then we could add g −1 · v and g −1 · e to
Core to create a larger subgraph satisfying the properties listed above,
contradicting the assumption that Core is a maximal subgraph with
these properties. Thus if v ∈ Γ then v ∈ g · Core for some g ∈ G.

To complete the proof of Lemma 1.52, we need to (possibly) expand

the Core, so that the image of Core under the action of G covers every
edge. An example where the Core cannot be a fundamental domain by
itself comes from Cayley graphs. The action of a group G on its Cayley
graph Γ is vertex transitive. From this it follows that Core consists
of the single vertex v, and therefore G · Core is the vertex set of Γ,
hence G · Core contains no edges. If there are edges in Γ that are not
in G · Core, we form a fundamental domain F by adding half-edges
to Core. Assume then that we have an edge e ⊂ G · Core where
4 Here we are running into some dicey logical/set-theoretic issues, directly connected
to the axiom of choice.
 1.8 Fundamental Domains and Generating Sets 33

Ends(e) ∩ Core = ∅. If Ends(e) ⊂ Core, then we could have added e

to Core and still satisfied our conditions above. So we may assume that
Ends(e) = {v, w} with v ∈ Core and w ∈ Core. Let he be the closed
half of e that contains v ∈ Core. Define F to be the union of Core and
these half-edges. In the case of a group acting on its Cayley graph, F is
a “star of half-edges,” consisting of a single vertex v and every half-edge
he where Ends(e) contains v.
Since F ⊃ Core, V(Γ) ⊂ G · F. It remains only to show that the
edges of Γ are also in G · F. To that end, let e be an edge in Γ where
Ends(e) = {v, w}. If e ∈ G · Core then there is a g and ĝ ∈ G such that
v ∈ g · Core and w ∈ ĝ · Core. Thus e will be covered by the g and ĝ
images of the half-edges added to Core in creating F.

Example 1.53. Let Γ = Kn,m with n = m. Let G = Sym(Γ) and note

that G acts transitively on V◦ and on V• . Thus Core can be taken to
be any edge of Γ, and one does not need to add half-edges to Core in
order to form a fundamental domain.
On the other hand, if n = m then there is a symmetry that exchanges
V◦ and V• . In this case Core consists of a single vertex, and a fun-
damental domain is formed by adding a half-edge that contains this
vertex.

Example 1.54. The symmetry group of the graph Γ shown in Fig-

ure 1.23 is isomorphic to D3 . This group acts transitively on the set of
4-valent vertices and on the set of 2-valent vertices, but no symmetry of
Γ takes a vertex of valence 2 to a vertex of valence 4. Thus Core can be
taken to be any edge joining a vertex of valence 2 to a vertex of valence
4. However, the image of such an edge under the action of Sym(Γ) is
just the hexagon on the outside of the graph; it does not include the
interior edges. Thus a half-edge needs to be added to Core in order to
form a fundamental domain. This is indicated in the picture on the right
of Figure 1.23.

Theorem 1.55. Let G act on a connected graph Γ with fundamental

domain F. Then the set of elements

S = {g ∈ G | g = e and g · F ∩ F = ∅}

is a generating set for G.

Proof. Let g ∈ G be an arbitrary element and let v be a vertex in

the fundamental domain F. Choose a path p joining v to g · v and
34 Cayley’s Theorems

Fig. 1.23. A bipartite graph Γ where there are two orbits of vertices under
the action of Sym(Γ). A fundamental domain for the action Sym(Γ)
Γ is
shown on the right.

let {F, g1 F, g2 F, . . . , gn F = gF} be a finite sequence of images of the

fundamental domain such that

1. the entire path p is contained in gi F, and
2. gi F ∩ gi+1 F = ∅, where g0 is defined to be the identity.

Since F ∩ g1 F = ∅, g1 ∈ S by definition. Similarly g1 F ∩ g2 F = ∅ implies

F ∩ g1−1 g2 F = ∅, thus g1−1 g2 ∈ S and so g2 is a product of elements in
S. Continuing in this manner we see that gn = g must also be a product
of elements of S.

Exercise 1.56. Let Γ be the graph shown in Figure 1.23 and let F be
the indicated fundamental domain for the action of Sym(Γ) ≈ D3 . Show
that the generating set given by Theorem 1.55 consists of two adjacent
reflections.

Example 1.57. If one applies Theorem 1.55 to the action Sym5
K5 ,

the resulting set of generators is quite large. Label the vertices of a
complete graph K5 by the numbers 1 through 5 (to make the isomor-
phism Sym5 ≈ Sym(K5 ) explicit) and take the fundamental domain
to be the one indicated in Figure 1.24. Let Sym{1,2,3,4} be the copy of
the symmetric group Sym4 that acts on K5 , fixing the vertex labelled
5. The image of the chosen fundamental domain under the action of
any element of Sym{1,2,3,4} intersects the fundamental domain at least
at the vertex labelled 5, so this entire subgroup (except the identity
element) is contained in S. Continuing to abuse notation we note that
 1.8 Fundamental Domains and Generating Sets 35
5

1 4

2 3

Fig. 1.24. A fundamental domain for the action of Sym5 on the complete
graph with five vertices.

the only other elements in S can be expressed as a transposition (45)

combined with some (possibly trivial) permutation from Sym3 . There
are six such symmetries, so |S| = 23 + 6 = 29.
The argument above applies to the action of Symn on Kn , giving
a huge generating set. However, an induction argument based on these
actions shows that the subset {(12), (23), . . . , (n−1, n)} is also a generat-
ing set for Symn . The base case consists of checking that (12) generates
Sym2 ; the induction step assumes {(12), (23), . . . , (n − 1, n)} generates
Symn , and then notes that the generating set given by Theorem 1.55
for the action Σn+1
Kn+1 consists of the non-identity elements in
Symn and the product of the transposition (n, n + 1) with elements of
Symn−1 . Thus adding (n, n + 1) to the generating set for Symn is all
that is needed to generate Symn+1 .

In addition to finding generating sets, fundamental domains can be

useful in computing the index of subgroups.

Theorem 1.58. Let G
Γ with fundamental domain F. Further as-

sume that if g · F = F then g = e ∈ G. If H < G and a fundamental
domain for the induced action H
Γ is a union of n copies of F (where
n ∈ N ∪ {∞}) then the index of H in G is n.

Proof. This is essentially an application of Theorem 1.33. Since no non-

identity element of G fixes F, there is a point x in F that is moved freely
under the action of G. Thus there is a bijection between the elements of
G and points in Orb(x). Express the fundamental domain for H
Γ as
36 Cayley’s Theorems

an explicit union of copies of F:


n
FH = gi F
i=1

where the gi are distinct elements of G and n ∈ N ∪ {∞}.

Since FH is a fundamental domain for the action of H, if g ∈ G then
n
g · x = gi h · x for one of the gi above. Thus G = i=1 gi H. Further, if
gi H ∩ gj H = ∅, then gi gj−1 ∈ H, which contradicts the fact that FH is
a fundamental domain. Thus the gi form a set of coset representatives
for H in G and [G : H] = n.
Example 1.59. Let C8 be an 8-cycle, that is, a connected graph with
eight vertices joined in a single cycle. The symmetry group of C8 is
D8 , and a fundamental domain for D8
C8 consists of a single vertex
and half-edge. There is a cyclic subgroup Z4 < Sym(C8 ), which can be
visualized as rotations through an angle of 90n◦ . A fundamental domain
for Z4
C8 consists of a combinatorial arc consisting of three vertices
joined by two edges. (See Figure 1.8.) As the fundamental domain for
Z4 is four times larger than the fundamental domain for Sym(C8 ) ≈ D8 ,
we have [D8 : Z4 ] = 4. Of course, this index can simply be computed
by dividing |D8 | by 4, but that approach will not work when the groups
involved are infinite.

Fig. 1.25. A fundamental domain for the action of D8 is highlighted on the

left side of the C8 graph, while a fundamental domain for the action of Z4 on
C8 is highlighted on the right.

The idea of finding fundamental domains for group actions, and using
them to find generating sets, is much more general than what we have
 1.9 Words and Paths 37

presented here. For example, instead of working with groups acting on

graphs one could look at groups acting on the Euclidean plane. A simple
example would be Z⊕Z where (1, 0) is a horizontal translation and (0, 1)
is a vertical translation, both of unit length. The square S formed by
taking the convex hull of the four vertices (±1/2, ±1/2) is a fundamental
domain for this action. The set
S = {g ∈ Z ⊕ Z | g = (0, 0) and g · S ∩ S = ∅}
is just {(±1, 0), (0, ±1), (±1, ±1)}. While this is larger than is necessary,
it is indeed a generating set for Z ⊕ Z.
We note that we used a variation of this technique when considering
the symmetry group of a cube in Section 1.6.3. There we noted that a
triangle formed a fundamental domain for the action of the symmetry
group, and we took as our generating set the reflections in the sides of
this triangle.

1.9 Words and Paths

The following set of definitions is our first entry into formal language
theory. At the moment we are just establishing convenient terminology,
but in Chapter 5 we explore some deeper connections between formal
languages and infinite groups.
Definition 1.60. Given a set S, a finite sequence of elements from S,
possibly with repetition, is called a word. The original set S is the alpha-
bet and the collection of all words, including the empty word consisting
of a sequence of zero elements, is denoted S ∗ . This is the free monoid on
S and, while we will use this terminology, we shall not need to explore
exactly what is implied by this term.
We let S −1 denote a set of formal inverses to the elements of S. For
example, if S = {a, b} then S −1 = {a−1 , b−1 }. We can then form the free
monoid {S ∪ S −1 }∗ whose elements are finite lists of elements from S
and their formal inverses. We emphasize that these are formal inverses.
If S = {a, b} then b, aba−1 , ab, and aba−1 a are all distinct elements in
the free monoid {S ∪ S −1 }∗ .
We add two conventions: if a ∈ S −1 , so a = x−1 for some x ∈ S, then
−1
we −1a be an alternate notation for x. In other words, we declare
 −1let
x to be x. We also allow ourselves to take formal inverses of
words, not just elements. If
ω = x1 x2 · · · xk−1 xk ∈ {S ∪ S −1 }∗
38 Cayley’s Theorems

then ω −1 stands for the word

ω −1 = x−1 −1 −1 −1
k xk−1 · · · x2 x1 ∈ {S ∪ S
−1 ∗
} .

Given any word ω ∈ {S ∪ S −1 }∗ , there is an associated edge path

in the Cayley graph ΓG,S . The path starts at the vertex corresponding
to the identity and then traverses edges of ΓG,S as dictated by ω. For
example, consider Z × Z generated by x = (1, 0) and y = (0, 1). Then
the word ω = xxy −1 x−1 yyyxxx describes the edge path illustrated in
Figure 1.26.

End of path

Identity

Fig. 1.26. The edge path pω associated to the word ω = xxy −1 x−1 yyyxxx.

We will frequently refer to the edge path described by a word ω, and

we let pω denote this path. Thus given a word ω the path pω : [0, 1] →
ΓG,S will join the vertex corresponding to the identity to the vertex
corresponding to the group element gotten by computing the product of
ω’s terms in G.
Conversely, every finite edge path in a Cayley graph ΓG,S describes
a word in the generators and their inverses: one reads off the labels of
the edges being traversed, and adds an inverse if they are travelling
in the opposite direction of the orientation of the edge. Thus one sees
that given a group G and a finite generating set S there is a one-to-one
correspondence between

1. finite edge paths in the Cayley graph ΓG,S which begin at the
vertex corresponding to the identity and
2. words in the free monoid {S ∪ S −1 }∗ .
 Exercises 39

Because of this connection between words and paths, the Cayley graph
of a group G is often thought of as a calculator for G. Let g and h be
elements of G and let ωg and ωh be words in {S ∪ S −1 }∗ representing
these elements. To compute the product g · h, one can follow the edge
path pg described by ωg and then, starting at the vertex corresponding to
g instead of the vertex corresponding to the identity, trace out the edge
path described by the word ωh . Continuing with the Z × Z example, if
g = y 2 x3 and h = y −3 x then g·h = y 2 x3 y −3 x, as is shown in Figure 1.27.
From this picture it is clear that g · h can be expressed as x4 y −1 .

Identity

gh

Fig. 1.27. Visualizing the product of g = y 2 x3 and h = y −3 x.

Exercises
(1) Enumerate all of the cycles in the graph of Figure 1.4.
(2) Construct infinitely many distinct biregular graphs where the white
vertices have degree 2 and the black vertices have degree 3.
(3) Construct a graph Γ where Sym(Γ) is vertex transitive but not
edge transitive.
(4) Construct a graph whose symmetry group is the cyclic group Zn
for each n ≥ 3.
(5) There are regular graphs whose symmetry groups are edge transi-
tive but not vertex transitive. (The smallest example is the “Folk-
man graph,” pictures of which are easily found on the Internet.)
Prove that any such graph must be bipartite. (A graph is regular
if the valence of every vertex is the same.)
40 Cayley’s Theorems

(6) Let Pn be the graph formed by the edges of an n-prism (for n ≥ 3).
The graphs P3 , P4 and P5 are shown in Figure 1.28. Prove that
Sym(Pn ) ≈ Dn ⊕ Z2 when n = 4. What goes wrong when n = 4?

Fig. 1.28. The prism graphs referred to in Exercise 6.

(7) Draw the Cayley graph of G = Z5 ⊕ Z2 with respect to the gener-

ating set {(1, 0), (0, 1)}. What is the difference between this Cayley
graph and the Cayley graph of D5 (with respect to {ρ, f })?
(8) Let ΓG,S be the Cayley graph of a group G.
a. Show by example that Sym(ΓG,S ) can be a strictly larger
group than Sym+ (ΓG,S ) ≈ G.
b. Show by example that the size of the gap between Sym(ΓG,S )
and Sym+ (ΓG,S ) can be infinite. That is, find a group G
such that the index [Sym(ΓG,S ) : Sym+ (ΓG,S )] is infinite.
(9) Let Γ be the Petersen graph. You may assume that Sym(Γ) ≈
Sym5 . Prove that if v is a vertex of Γ then |Stab(v)| = 12. Can
you determine which group of order 12 this is?
(10) Prove that a set of elements S generates a group G if and only if
the only subgroup of G that contains S is G.
(11) Prove that every finitely generated group is countable.
(12) The first published Cayley graph, appearing in Cayley’s original
paper [Ca78], was the Cayley graph for A4 with respect to the
generating set {(123), (234)}. Draw this Cayley graph.
(13) Let G be the symmetry group of the regular octahedron.
a. Compute the order of G.
b. Find a set of generators S for G and explain how you know
these form a set of generators. (Don’t cheat and pick S =
G!)
c. Draw or explain in detail the shape and structure of the
Cayley graph ΓG,S .
(14) What is the order of the symmetry group of a dodecahedron? Find
a set of generators and describe the associated Cayley graph.
 Exercises 41

(15) Let
n be the n-cube formed by taking the convex hull of the
points in Rn whose coordinates are (±1, . . . , ±1). Let Sym(
n ) be
the associated symmetry group.
a. What is the order of Sym(
n )?
b. Find a set of generators for Sym(
n ).
c. Draw the Cayley graph of Sym(
3 ) with respect to your
chosen set of generators.
(16) The alternating group A4 consists of the twelve even permutations
of {1, 2, 3, 4}. The permutations (123) and (12)(34) generate A4 .
Draw the associated Cayley graph.
(17) Let G = Sym(K3,3 ).
a. Show that G is vertex transitive.
b. Show that G is flag transitive.
c. Show that G is 2-transitive. That is, elements of G can take
any embedded arc of combinatorial length 2 to any other
such arc.
d. Show that G is also 3-transitive.
e. Is G 4-transitive?
(18) Let Γ be a graph and let G
Γ.
a. If the action of G on Γ is vertex transitive, show that, for
any v and w in V (Γ), Stab(v) ≈ Stab(w).
b. If you didn’t already show it above, show that Stab(v) is
conjugate to Stab(w).
(19) Given an edge e in a graph Γ there are two reasonable ways to
define the stabilizer of e. One is to say

Stab(e) = {g ∈ G | g · e = e}

and the other is to say

Stab
(e) = {g ∈ G | g fixes e pointwise}.

a. Let Γ be the wheel graph from Figure 1.22 and let G =

Sym(Γ). Show that if e is one of the edges attached to the
central vertex, then

Stab
(e) = Stab(e) ≈ Z2 .

b. Let e be one of the exterior edges of the wheel graph Γ.

Show that Stab
(e) is trivial while Stab(e) ≈ Z2 .
42 Cayley’s Theorems

Fig. 1.29. The Doyle graph, which is vertex and edge transitive, but not flag
transitive.

c. Let G
Γ be an arbitrary action of a group on a graph.

Let e be any edge of Γ. Prove that Stab
(e) is a subgroup
of index at most 2 in Stab(e).

(20) Prove the following Corollary to Theorem 1.51: For every finitely
generated group G, there is a graph Γ, neither directed nor edge
labelled, where Sym(Γ) ≈ G.
(21) Let H be the subset of permutations of the set of integers, SymZ ,
such that h ∈ H if and only if there is a finite set of integers
C ⊂ Z and a number k such that h(n) = n + k if n ∈ C. In other
words, the elements of H all look like translations outside of a finite
set of numbers. (Note: the subset C and the number k are both
dependent on which element h you are examining.)

a. Show that H is a subgroup of SymZ .

b. Show that H is finitely generated.

(The group H is one example of an interesting family of groups

called the Houghton groups.)
(22) Let Γ be the Petersen graph, shown in Figure 1.8. The group of
symmetries of Γ is vertex transitive, just as is the case for a Cayley
graph. Is there a group G and a generating set S such that the
 Exercises 43

Petersen graph is the underlying graph of the Cayley graph ΓG,S ?

(Hint: The answer is no.)
(23) The graph Γ shown in Figure 1.29 was discovered by Doyle in 1976
as part of his senior thesis (see [Ho81]). Prove that Sym(Γ) is vertex
transitive and edge transitive but it is not flag transitive.
 2
Groups Generated by Reflections

So, my interest in symmetry has not been misplaced.

. –H. S. M. Coxeter (upon learning that his brain
. displayed a high degree of bilateral symmetry)

Let a and b be reflections in parallel lines in the Euclidean plane (as in

Figure 2.1). If we think of the line of reflection for a as being x = 0
and the line of reflection for b as being x = 1, then we may express a as
the function a[(x, y)] = (−x, y) and b as b[(x, y)] = (2 − x, y). It follows
that ab[(x, y)] = (x − 2, y) and ba[(x, y)] = (x + 2, y). Moreover, for any
n ∈ N, (ab)n is a horizontal translation to the left through a distance of
2n and (ba)n is a horizontal translation to the right through a distance
of 2n. Thus the reflections a and b generate an infinite group.

a b

Fig. 2.1. The reflections a (across the solid line) and b (across the dashed line)
generate an infinite group.

If the two lines of reflection actually met, at an angle of π/n, then

the group generated by a and b would be the dihedral group of order
2n, and the product ab would be a rotation through an angle of 2π/n.
Thus the group we are considering is something like a dihedral group,

44
 Groups Generated by Reflections 45

except that the two reflections generate a translation, not a rotation. As

the group generated by two parallel reflections is infinite, it is referred
to as the infinite dihedral group, denoted D∞ .
Because ab moves the line x = 2 to the line x = 0, which is fixed by
a, and its inverse ba returns the line x = 0 to the line x = 2, the reader
can (and should!) verify that ba · a · ab = bab is the reflection in the line
x = 2. Similarly aba gives a reflection in the line x = −1. Starting from
these computations one can build an induction argument that proves
that every vertical line x = n is the axis of reflection for some element
of D∞ . In fact, these computations form the base cases of an induction
argument (which we ask you to complete as Exercise 1) that establishes
the following description of the elements of D∞ :

Proposition 2.1. Let a be the reflection that fixes x = 0 and b the

reflection fixing x = 1, and let D∞ be the group generated by a and b.
Then every element of D∞ can be expressed as an alternating product
of a’s and b’s. Further, if n ∈ N, then the following hold.

1. The element
abab ·

· · abab = (ab)n is the horizontal translation
2n letters
described by (x, y) → (x − 2n, y).
2. The element baba

·
· · baba = (ba)n is the horizontal translation
2n letters
described by (x, y) → (x + 2n, y).
3. The element baba

· · abab = (ba)n−1 b is the reflection that fixes
·
2n − 1 letters
x = n.
4. The element abab

· · baba = (ab)n a is the reflection that fixes
·
2n + 1 letters
x = −n.

The point in the middle of Figure 2.1 is not on an axis of reflection,

hence, by the description given in Proposition 2.1, it is moved freely
by D∞ . This means we may apply the Drawing Trick (Remark 1.49)
to construct the Cayley graph Γ of D∞ , with respect to {a, b}, using
the orbits of this point as the vertices. Appealing to the convention
of dropping double arrows when generators are of order 2, we get the
Cayley graph embedded in R2 as in Figure 2.2. Looking at this Cayley
graph one realizes that having D∞ act on the Euclidean plane is a bit
of overkill: D∞ acts on the real line. The generator a can be thought of
as the reflection that fixes 0 ∈ R and b as the reflection fixing 1 ∈ R.
46 Groups Generated by Reflections

a b

Fig. 2.2. The Cayley graph of the infinite dihedral group

The group D∞ is a non-abelian, infinite group that exhibits some

striking behavior.

Proposition 2.2. The subgroup H of D∞ generated by {a, bab} is a

subgroup of index 2 that is isomorphic to D∞ .

Proof. Let α = a and β = bab. There is a homomorphism φ : H → D∞

defined by declaring φ(α) = a, φ(β) = b and then for an arbitrary
element φ(x1 x2 · · · xn ) = φ(x1 )φ(x2 ) · · · φ(xn ), where xi ∈ {α, β}. This
map is well defined, as it follows from the description in Proposition 2.1
that an element of D∞ has only one such expression as a product of
a’s and b’s. That this map satisfies the condition φ(xy) = φ(x)φ(y),
is immediate from its description. Similarly it is immediate that the
homomorphism is surjective, as a and b are in the image, and D∞ is
generated by a and b.
To see that φ is injective, consider the possibility that some non-
identity element in H is taken to the identity in D∞ . This element is
either of even or odd length, and either it begins with α or β. But then
the image under φ has the same form and, by Proposition 2.1, it is not
the identity.
To see that H is of index 2, we note that the elements of H are simply
the elements of D∞ that have an even number of b’s in their expression.
Thus D∞ = H ∪ bH.

(The reader who finds Proposition 2.2 interesting should work through
Exercise 3 at the end of this chapter.)

Having considered groups generated by two reflections, we now con-

sider a few groups generated by three reflections. The Euclidean plane
can be tiled by equilateral triangles. Fix a triangle T inside of such a
 Groups Generated by Reflections 47

tiling, and consider the group generated by the three reflections whose
axes of reflection are the (extended) sides of T. In Figure 2.3 we have

g b

Fig. 2.3. A tiling of the Euclidean plane where the edges in the tiling have
one of three types associated to them.

highlighted one such triangle; we refer to the reflections across the sides
of this triangle as r, g and b, as is indicated. As a quick exercise, the
reader should verify that r, rgr and rbr are the reflections in the sides
of the triangle immediately above T in this tiling. For example, in rgr
the rightmost “r” flips the triangle above T down to T’s position; the
middle “g” flips T to the left; then the leftmost r flips g · T up to the
correct position.

Think of the edges of T as being wet with three types of paint, r, b

and g. When we apply the reflection r to T, the edges painted b and g
are flipped upward, and they leave traces of their color on these edges
in the tiling. If we continue to flip T across its edges, it will continue to
leave traces of wet paint on the edges of the tiling. In this manner every
edge in the tiling can be labelled as either r, b or g. We have indicated
this in Figure 2.3 with differing line styles. This labelling of the edges in
the tiling makes it quite easy to determine the action of any particular
element of G.
Let T be any triangle in the tiling. Then T can be joined to T by a
corridor of triangles, that is, a sequence of triangles

{T = T0 , T1 , T2 , · · · , Tn = T }
48 Groups Generated by Reflections

ρ1 ρ3
ρ2
r

g b

Fig. 2.4. A corridor joining two triangles in the tiling of the Euclidean plane.

where Ti intersects Ti+1 in an edge (when 0 ≤ i < n). An example is

shown in Figure 2.4. In this figure, if ρi denotes a reflection across the
associated edge, then the element ρ3 ρ2 ρ1 r moves T to T . As was noted
previously, ρ1 is a conjugate of b, namely ρ1 = rbr. Thus we can rewrite
ρ1 r as rbr · r = rb. Since ρ1 r(= rb) takes T to the third triangle in the
indicated corridor, it follows that ρ2 = ρ1 r · r · rρ1 = rbrbr. Hence
ρ2 ρ1 b = (rbrbr)(rbr)r = rbr.
Continuing in this fashion we get
ρ3 ρ2 ρ1 b = rbrg
since ρ3 = ρ2 ρ1 r · g · rρ1 ρ2 = rbr · g · rbr.
Lemma 2.3. Let T be joined to T by a corridor
{T = T0 , T1 , T2 , · · · , Tn = T }.
Let i ∈ {r, g, b} be the label of the edge Ti−1 ∩ Ti . Then the element
1 2 · · · n takes T to T .
Proof. The base case follows from the definition of r, g and b.
Assuming our result is true for corridors of length n, we consider a
corridor
{T = T0 , T1 , T2 , · · · , Tn , Tn+1 = T }.
Because ρ1 ρ2 · · · ρn takes T to Tn , we know ρn ρn−1 · · · ρ1 takes Tn+1 to
the triangle that is adjacent to T along the edge labelled ρn+1 . Thus
ρn+1 ρn ρn−1 · · · ρ1 takes Tn+1 to T, hence ρ1 ρ2 · · · ρn ρn+1 takes T to
Tn+1 .
 Groups Generated by Reflections 49

Remark 2.4. Lemma 2.3 is an example of what we refer to as the

“two wrongs make a right” principle. As we have frequently mentioned,
our actions are always left actions. In particular, if we want to know
where the element brg takes T, we start by applying g, then r, then b.
And yet, Lemma 2.3 tells us we can work left-to-right, if we make one
further mistake. Namely, after we apply b, we then apply “r” thinking
of this as the reflection across the edge labelled r in b · T. We then apply
“g”, repeating the same mistake, by reflecting br · T across its edge
labelled g.
This principle will be discussed again in Section 3.5.

"r"
b "g"

Fig. 2.5. The element brg moves T to the triangle highlighted on the right.

Lemma 2.3 implies that G acts transitively on the triangles of this

tiling. That is, given any two triangles T and T , there is an element
h ∈ G such that h · T = T . Further, let h = ρ1 ρ2 ρ3 · · · ρn where each
ρi ∈ {r, g, b}. Then there is a corridor {T = T0 , T1 , T2 , · · · , Tn } where
the label of the edge Ti−1 ∩ Ti is ρi . By Lemma 2.3, h · T = Tn .
There is a theorem of Euclidean geometry that says any isometry of
the plane is determined by where it sends three non-collinear points.
Using this one can see that the action of G on the triangles of this tiling
is simply transitive. If there is an h ∈ G with h · T = T then, by the
theorem stated above, h is the identity. Since G acts simply transitively
on the triangles, G also acts simply transitively on the barycenters of the
triangles in the tiling. We may then apply the Drawing Trick using the
barycenter, x, of our chosen triangle T. The barycenters of the triangles
form the orbit of x under the actions of G, hence they correspond to
50 Groups Generated by Reflections

the vertices of the Cayley graph. When the edges of the Cayley graph
are added, one sees that the Cayley graph is the vertices and edges of
the regular hexagonal tiling of plane that is dual to the original tiling
by equilateral triangles. This is shown in Figure 2.6.

Fig. 2.6. The Cayley graph of the reflection group W333 .

There are two other tilings of the Euclidean plane that lead naturally
to infinite groups generated by three reflections. One can tile the plane
by triangles with interior angles {π/2, π/4, π/4} and one can also tile
the plane by triangles with interior angles {π/2, π/3, π/6}. The groups
generated by reflecting in the sides of such triangles are similar to the
one we have discussed above, but they have their individual quirks.
At this point we have a reasonably long list of “groups generated by
reflections.” There are the finite dihedral groups as well as D∞ and the
groups generated by reflections in the sides of certain Euclidean triangles.
The symmetry groups of all five Platonic solids are also groups generated
by reflections, as are the symmetry groups of cubes in all dimensions.
There is a well-developed theory that contains these groups and many
 Exercises 51

others; the general class of groups being considered was identified by

Jacques Tits who called them Coxeter groups in honor of Coxeter’s work
on symmetry groups of regular polygons. Within the general theory of
Coxeter groups, the groups we have been considering in this chapter
are called “Coxeter groups of type ”: the infinite dihedral group D∞
is the Coxeter group of type A 1 ; the group generated by reflections in
the sides of an equilateral triangle is of type A2 ; the group generated
2 ; and
in the sides of a triangle with angles {π/2, π/4, π/4} is of type B
the group generated by reflections in the sides of a triangle with angles
{π/2, π/3, π/6} has type G 2 .
This notation makes sense in the context of the classification of finite
Coxeter groups and Coxeter groups generated by reflections in Euclidean
spaces (of all dimensions), but it is not of much use to us. We will
continue to refer to D∞ as D∞ and we will refer to the group generated
by reflections in the sides of a Euclidean triangle with interior angles
{π/p, π/q, π/r} as Wpqr . (The letter W is commonly used to denote a
Coxeter group.)

Example 2.5. Theorem 1.58 was written in terms of fundamental

domains for actions on graphs, but it did not need to be so restric-
tive. In Figure 2.7 we show the axes of reflection for elements of W333
and have highlighted three such axes that form an equilaterial triangle,
which is four times bigger than the fundamental domain for W333
R2 .
If H is the subgroup generated by these three reflections, then H ≈ W333
is a group generated by reflections in the sides of an equilateral triangle.
Using the argument of Theorem 1.58 in this context we find that H has
index 4 in the original W333 .

The reader who is interested in Coxeter groups can learn the basics by
reading [Hu90]. An excellent book introducing the deep combinatorial
information associated to Coxeter groups should read [BjBr05]. A reader
with a stronger background in mathematics than we are assuming in this
text, who is interested in the geometry and topology of Coxeter groups,
should read [Da08].

Exercises
(1) Prove Proposition 2.1. This is perhaps easiest to do if you view
a as the function a(x) = −x and b as the function b(x) = 2 − x,
remembering that the group operation is function composition.
52 Groups Generated by Reflections

Fig. 2.7. The group W333 contains an index 4 subgroup that is isomorphic
to W333 .

(2) Recall that T2,2 is the regular 2-valent, bipartite tree. Prove that
D∞ ≈ Sym+ (T2,2 ).
(3) Let the two reflections generating D∞ be denoted a and b, and use
the same conventions for the action D∞
R as in Proposition 2.1.
a. Prove that the subgroup a, babab is isomorphic to D∞ and
is of index 3 in the original group.
b. Prove that for any n ≥ 1 the group D∞ contains a subgroup
that is isomorphic to D∞ and of index n.
c. Prove that if α is any automorphism of D∞ then α takes a
and b to reflections fixing x = n and x = n + 1, for some
n ∈ Z.
d. Prove that the center of D∞ = {e} and conclude that D∞ ≈
Inn(D∞ ) < Aut(D∞ ).
e. Show that the index of Inn(D∞ ) in Aut(D∞ ) is 2. (Hint:
There is an automorphism given by φ(a) = b and φ(b) = a,
which is not an inner automorphism.)
f. Prove that Aut(D∞ ) ≈ D∞ .
(4) Determine how the element rgb acts on the tiling of the Euclidean
plane by regular triangles. What about gbr? or brg?
(5) Tile the Euclidean plane by squares of side length 1. Let W be
the group generated by the four reflections in the (extended) sides
of any one square. Draw the Cayley graph of W and prove that
W ≈ D∞ ⊕ D∞ .
(6) Prove that Z ⊕ Z is a subgroup of W333 .
(7) Draw the Cayley graphs of W244 and W236 .
 Exercises 53

(8) Show that W236 contains a finite index subgroup isomorphic

to W333 .
(9) Show that each of the groups W236 , W244 , and W333 contains a
proper subgroup isomorphic to itself.
Bonus: For which values of n is W333 a subgroup of index n in
itself?
 3
Groups Acting on Trees

Eine abstrakte Theorie der diskontinuierlichen Gruppen, die von vornherein

darauf abzielt, unendliche Gruppen einzubegreifen, wird in dem Studium der
freien Gruppen irhren natürlichen Ausgangspunkt finden.
. –Jakob Nielsen

3.1 Free Groups

Free groups are central to the study of infinite groups and provide rela-
tively accessible examples that highlight the power of mixing algebraic
and geometric approaches. Free groups can be defined in a number of
ways. The standard method, which we give in Section 3.1.1, is formal
and algebraic. Another common method is to use actions on trees, which
is the perspective we develop in Section 3.4. While the formal definition
lacks the intuition building power of the geometric perspective, it is of-
ten the easiest definition to use when trying to prove that a given group
is a free group.

3.1.1 Free Groups of Rank n

Definition 3.1. Let S = {x1 , . . . , xn } be a set of elements in a group, G.
A word ω ∈ {S ∪ S −1 }∗ is said to be freely reduced if it does not contain
a subword consisting of an element adjacent to its formal inverse. For
example, the word ω = xyx−1 y −1 is freely reduced while ω  = xy −1 yxy
is not. The group G is a free group with basis S if S is a set of generators
for G and no freely reduced word in the xi and their inverses represents
the identity.
The rank of a free group with basis S is the number of elements in S.
We denote a free group of rank n by Fn .

54
 3.1 Free Groups 55

In the definition above we explicitly assume that S is a finite set, but

one can discuss free groups of infinite rank, where the generating set S is
not finite. Before encountering such groups, we establish some important
results about free groups of finite rank, not the least of which is that
they exist!

Theorem 3.2. Given any n ∈ N, there is a free group of rank n.

We only sketch the argument. A very similar argument is given in

Section 3.6, where the situation is a bit more complicated.

Outline of proof. Let S = {x1 , . . . , xn } be n distinct symbols. Consider

the equivalence relation on all words, in {S ∪ S −1 }∗ induced by

a1 · · · ai−1 ai a−1
i ai+1 · · · ak ∼ a1 · · · ai−1 ai+1 · · · ak

where we have cancelled an adjacent pair of elements because they are

inverses of each other.
The elements of the group Fn are the equivalence classes of words in
{S ∪ S −1 }∗ under the equivalence relation ∼. If ω is a given word, we
denote its equivalence class as [ω] and claim that the following binary
operation defines a group structure on this set: [α][β] = [αβ] where
αβ is just the concatenation of the two words α and β. The identity
element is then [ε], where ε represents the empty string, and the inverse
of [ω] is [ω −1 ] where ω −1 is the formal inverse of the word ω (as described
in Section 1.9).

One should think of the elements of Fn as corresponding to freely

reduced words, and the following exercise justifies this intuition.

Exercise 3.3. Show that every equivalence class contains exactly one
freely reduced word from {x1 , . . . , xn , x−1 −1 ∗
1 , . . . , xn } .

The reason for introducing the equivalence relation, and not simply
stating that freely reduced words are the elements of Fn , is that the con-
catenation of two freely reduced words is not necessarily freely reduced.

3.1.2 F2 as a Group of Tree Symmetries

As will become more clear as this chapter progresses, there is a direct
connection between free groups and free actions on trees. One way to
produce a free action of a given group G on a graph is via Cayley’s Better
Theorem (Theorem 1.42), which we exploit in the following proposition.
56 Groups Acting on Trees

y
x x

y
y y
x x v x x

y y
y

x x

Fig. 3.1. The Cayley graph of F2 .

Proposition 3.4. The Cayley graph of F2 with respect to {x, y} is an

oriented version of T4 , the uniformly 4-valent tree, shown in Figure 3.1.

Proof. As F2 is generated by two elements, the associated Cayley graph

will be uniformly 4-valent. Appealing to the correspondence between
words and paths, we note that because no freely reduced word represents
the identity, this Cayley graph of F2 contains no cycles. Thus the Cayley
graph is a connected, uniformly 4-valent tree, T4 .

Since every group acts on its Cayley graph (Theorem 1.51) we imme-
diately get:

Corollary 3.5. The free group F2 can be viewed as a group of symme-

tries of T4 .

The group F2 acts on T4 on the left. So, for example, the generator x
takes the vertex corresponding to a freely reduced word ω to the vertex
corresponding to the freely reduced word equivalent to x · ω. One can
 3.1 Free Groups 57

write a formula for this action. Let ω be a freely reduced word, and let
ω = α · ω  , where α is a single letter. Then the action of x is given by


xω if x−1 = α
x · ω = x · α · ω =
ω  if x−1 = α.
The action of x ∈ F2 on T4 is something of a “shift to the right.” To see
this, consider the subgraph of T4 induced by the vertices corresponding
to {xn | n ∈ Z}. Left multiplication by x shifts this combinatorial line
one unit to the right, thereby shifting the entire Cayley graph, as is
illustrated in Figure 3.2. The action of y is similar, the main difference
being that the combinatorial line induced by vertices corresponding to
{y n | n ∈ Z} is shifted one unit up.

Fig. 3.2. Left multiplying by x moves the vertical florets to the right, as in
the left-hand side of the picture; left multiplying by y moves the horizontal
florets up.

This is perhaps a good time for a reminder: visualizing the action

of a given element in a group G on its Cayley graph is not the same
as following paths from the identity to that element in a Cayley graph.
Consider the specific example shown in Figure 3.3. Here we have an edge
path between the vertex associated with the identity, v, and another
vertex w, in the Cayley graph of F2 . The edge path is labelled by the
word xy −1 x−1 y. One might guess that the edge path traces out the
motion of v under the action of these four transformations, but that is
not the case.
In Figure 3.4 we see the movement of the vertex v under the action
of xy −1 x−1 y. Remember that the action of F2 is on the left, so that the
first thing one does is apply y to v, moving it up. Then one applies x−1
58 Groups Acting on Trees

v
w

Fig. 3.3. The edge path connecting v and w in the Cayley graph of F2 .

x –1
y y
v v

x –1 x –1
y y
v v
y –1 y –1
x w

Fig. 3.4. Getting from v to w via the action of F2 .

 3.1 Free Groups 59

to the vertex y · v, as shown in the upper right-hand corner. Then y −1

moves x−1 y · v down, and x moves y −1 x−1 y · v to the vertex w. Thus
while the end result is the same, the route that the vertex v travels is
not along the path associated to the word xy −1 x−1 y. (This comment is
similar to the discussion around Lemma 2.3.)
Remark 3.6. The reader should keep in mind that free groups and free
abelian groups are different. Free abelian groups are the groups of the
form Z ⊕ Z ⊕ Z ⊕ · · · ; they are not the same as the free groups Fn .
In particular, free abelian groups are commutative, while free groups
are not.

3.1.3 Free Groups in Nature

The definition of free groups in terms of freely reduced words may ap-
pear to be quite formal, and in fact free groups may appear to have
been created out of whole cloth, but these groups do arise in a startling
number of mathematical contexts. (This is what mathematicians mean
when they say something arises “naturally” or “in nature.”) If you have
studied non-Euclidean geometry, you can find free groups inside of the
group of isometries of the hyperbolic plane (Exercise 5). Directly related
to that situation, consider the group SL2 (Z) consisting of two-by-two ma-
trices with integer entries and determinant 1. The subgroup of SL2 (Z)
generated by

   
1 0 1 2
S= ,
2 1 0 1
has index 12 in SL2 (Z). In fact, this subgroup is the kernel of the map
φ : SL2 (Z) → SL2 (Z2 ). (It is clear that S ⊂ ker(φ) but it is not obvious
that S generates this kernel. Proving this fact would be a challenging
exercise.)
Proposition 3.7. The subgroup of SL2 (Z) generated by

   
1 0 1 2
S= ,
2 1 0 1
is isomorphic to F2 .
   
1 0 1 2
Proof. Let l = and r = . We need to show that no
2 1 0 1
freely reduced word in {l, r, l−1 , r−1 }∗ represents the identity. We do
this using the action of SL2 (Z) on the Euclidean plane. We think of the
60 Groups Acting on Trees
 
a
point (a, b) as the tip of the vector , and allow the elements of our
b
subgroup to act on points in the plane via matrix multiplication. For
example, the generator l acts as follows:
     
1 0 a a
· = .
2 1 b 2a + b

Thus l · (a, b) = (a, 2a + b). A similar computation establishes that r ·

(a, b) = (a + 2b, b). Inverting the matrices gives l−1 · (a, b) = (a, −2a + b)
and r−1 · (a, b) = (a − 2b, b).

Fig. 3.5. Four subsets in the plane corresponding to the elements of

{l, r, l−1 , r−1 }.

In Figure 3.5 we show specific subsets of the plane, each affiliated

with a generator or an inverse of a generator. These subsets are unions
of open sectors, each of whose boundary is contained in the union of
the coordinate axes and the lines y = x and y = −x. For example, Xl
is the union of two open sectors, where the bounding edges are x = 0
and y = x. (The point (1, 3) ∈ Xl , while (1, 1) ∈ Xl , for example.) The
 3.1 Free Groups 61

subset Xl−1 is the reflection of Xl across the y-axis. Similarly, Xr is the

union of two open sectors, where the bounding edges are y = 0 and
y = x; Xr−1 is the reflection of Xr across the x-axis.
Because l is a linear transformation, where we know l · (0, 1) = (0, 1)
and l · (1, 1) = (1, 3), we see that l · Xl is properly contained inside of
Xl . Similarly, since l · (1, 0) = (1, 2) and l · (1, 1) = (1, 3), we see that
l·Xr ⊂ Xl . One more computation shows that l·Xr−1 is also contained in
Xl . In fact, this containment property holds for each w ∈ {l, r, l−1 , r−1 }.
That is, if w ∈ {l, r, l−1 , r−1 } then

w · Xy ⊂ Xw for y = w−1 ,

where the containment is always proper.

Let ω = w1 w2 · · · wn be a freely reduced word in {l, r, l−1 , r−1 }∗ . In
order to prove that {l, r} generates a free group of rank 2, we need to
show that ω is not the identity. In other words, we need to show that ω
does not act trivially on the plane. If |ω| = 1, then it is clear from the
definition that ω = e. Furthermore, because ω is freely reduced, we may
apply the containment property described above, one letter at a time:

ω · Xwn = w1 w2 · · · wn−2 wn−1 wn · Xwn

⊂ w1 w2 · · · wn−2 wn−1 · Xwn ⊂ w1 w2 · · · wn−2 · Xwn−1
⊂ · · · ⊂ w1 Xw2 ⊂ Xw1

where each of the containments is proper. Thus ω·Xwn is a proper subset

of Xw1 . It follows that ω · Xwn = Xwn and therefore ω = e.

Remark 3.8. The fact that SL2 (Z) contains a free subgroup is not an
isolated fact. There is a celebrated theorem of Jacques Tits [Ti72] that
says: Let G be a finitely generated group that can be faithfully represented
as a group of matrices.1 Then either G contains a finite index subgroup
that is solvable, or G contains a subgroup isomorphic to F2 . The reader
unfamiliar with solvable groups can interpret this result as saying: linear
groups contain free subgroups unless they obviously do not.

In the example above we demonstrated that F2 < SL2 (Z), and in

Remark 3.8 we indicated that free groups commonly arise as subgroups
of linear groups. Free groups also arise as subgroups in other important
classes of groups. For example, a continuous, one-to-one and onto map
φ : R → R, whose inverse is also continuous, is called a homeomorphism
1 A group that can be faithfully represented as a matrix group is called a linear
group.
62 Groups Acting on Trees

of R. Examples of homeomorphisms are f (x) = x3 , f (x) = 7 − x and

f (x) = x+sin(x). The collection of all such continuous maps forms a very
large group, Homeo(R), whose binary operation is function composition.
The identity element is the function f (x) = x, and inverses exist because
we have restricted ourselves to functions that have inverses.
A number of interesting infinite groups arise as subgroups of Homeo(R).
For example, the subgroup generated by f (x) = x3 is infinite cyclic. The
subgroup H < Homeo(R) generated by f (x) = x3 and g(x) = x5 is a
free abelian group of rank 2. To prove this, note that any h ∈ H has
m n
the form h(x) = x3 5 . Thus there is a map from H to Z ⊕ Z defined
m n
by φ : h(x) → (m, n) when h(x) = x3 5 . Checking that φ is an isomor-
phism is a straightforward exercise. The infinite dihedral group, D∞ , is
an example of a non-abelian, infinite subgroup of Homeo(R).
We construct an example of a free subgroup of Homeo(R) using two
explicit functions. Let f (x) be the function whose graph is shown in
Figure 3.6. This function is defined on [0, 1] by

1
ˆ
f (x) = 4 1
4x
 1
 if 01 ≤ x ≤ 5
5 + 4 x− 5 if 5 ≤ x ≤ 1

and is extended to the real line by f (x) = x

+ fˆ(x − x
). Define Xf
to be small closed intervals around the integers,
 1 1

Xf = i − ,i + .
5 5
i∈Z

This subset of R is indicated in both the domain and range in Fig-

ure 3.6. In this picture you can see that f takes (1/5, 4/5) into (4/5, 1).
In general,
  
1 4 4
f i + ,i + ⊂ i + , i + 1 ⊂ Xf ,
5 5 5
for all i ∈ Z. Thus f takes points that are outside of Xf , into Xf .
The argument above provides the base case for an induction argument
establishing that, for any positive integer n, if x ∈ Xf then f n (x) ∈ Xf .
An important fact to notice is that if x ∈ Xf then f (x) ∈ (i − 1/5, i)
for some i ∈ Z. But f [(i − 1/5, i)] ⊂ (i − 1/5, i), so f 2 (x) is also in
(i − 1/5, i). For the induction step we may assume that if x ∈ Xf then
f n−1 (x) ∈(i − 1/5, i) for some i ∈ Z, and we then simply note that
f n (x) = f f n−1 (x) must also be in (i − 1/5, i) ⊂ Xf .
The graph of the inverse of f (x) is just the reflection of the graph of
f (x) across the line y = x. Thus the graph of f −1 (x) is also contained
 3.1 Free Groups 63

Fig. 3.6. The graph of the function f , which takes points in the open intervals
{i + (1/5, 4/5) | i ∈ Z} to points in the open intervals {i + (4/5, 1) | i ∈ Z}.

in the grid formed by Xf in the domain and in the range. In this case,
if x ∈ (i − 4/5, i − 1/5) then f −1 (x) ∈ (i − 1, i − 4/5), so f −1 (x) ∈ Xf .
Further, since f −1 takes (i − 1, i − 4/5) into (i − 1, i − 4/5), f −n (x) ∈ Xf
for any n ∈ N and x ∈ Xf .
Our second function is just a shifted copy of f . Define g(x) to be
f (x − 1/2) + 1/2 and let

Xg = [i + 1/2 − 1/5, i + 1/2 + 1/5].
i∈Z

Then the same arguments as are given above show that, for any non-zero
integer n, positive or negative, g n (x) ∈ Xg if x ∈ Xg .
We can now establish our claim that the subgroup of Homeo(R) gen-
erated by f and g is a free group, using an argument that is similar to
the proof of Proposition 3.7.

Proposition 3.9. The group of homeomorphisms of the real line gen-

erated by f and g (defined above) is a free group of rank 2.

Proof. Let ω = w1 · · · wn be a freely reduced word in {f, g, f −1 , g −1 }∗ .

We show that ω = e by showing that ω does not act trivially on R. By
collecting terms we may express ω as w1n1 w2n2 · · · wknk where wi = wi+1
−1
or wi+1 .
64 Groups Acting on Trees

Notice that Xf ∩ Xg = ∅ and that 1/4 ∈ Xf ∪ Xg . Define

Xf if wi ∈ {f, f −1 }
Xi =
Xg if wi ∈ {g, g −1 }.
If we apply ω to the point x = 1/4 we see that
n
ω(1/4) = w1n1 w2n2 · · · wk−1
k−1
wknk (1/4)
n n
⊂ w1n1 w2n2 · · · wk−1
k−1
(Xk ) ⊂ w1n1 w2n2 · · · wk−2
k−2
(Xk−1 )
⊂ · · · ⊂ w1n1 w2n2 (X3 ) ⊂ w1n1 (X2 ) ⊂ X1 ⊂ Xf ∪ Xg .

But 1/4 ∈ Xf ∪ Xg so ω(1/4) = 1/4 and therefore ω = e.

The method used to establish Propositions 3.7 and 3.9 is often referred
to as the Ping Pong Lemma. The version of the Ping Pong Lemma that
was used to establish 3.9 is stated as:

Lemma 3.10 (Ping Pong Lemma). Let G act on X and let S = S ∪S −1

be a set of generators and their inverses. For each s ∈ S let Xs be a
subset of X, and let p be a point in X \ {∪s∈S Xs }. If

1. s(p) ∈ Xs for each s ∈ S; and

2. s(Xt ) is a proper subset of Xs , for each t ∈ S \ {s−1 },

then G is a free group with basis S.

As we have essentially proved the Ping Pong Lemma in two specific

cases, we leave the proof of the general result as an end of chapter
exercise.

Remark 3.11. Proposition 3.9 was originally presented in a paper by

Curtis Bennett [Be97]. The functions f and g used in Proposition 3.9
are somewhat complicated. There are, in fact, more elementary functions
inside of Homeo(R) that generate free subgroups. For example, if p is an
odd prime, then the functions f (x) = xp and g(x) = x + 1 generate a
free subgroup of Homeo(R). This seemingly accessible result is strikingly
difficult to prove (see [Wh88]).
Our examples above are by no means an exhaustive list of places
where free groups appear in mathematics. For example, free groups are
an integral part of the proof of the Banach–Tarski Paradox.2 If you are
2 In 1924, Banach and Tarski proved that a solid ball in R3 can be decomposed
into a finite number of pieces, which can then be reassembled to form two disjoint
balls, each of the same volume as the single ball you started with. See [Wa86] for
more information.
 3.2 F3 is a Subgroup of F2 65

interested in the diverse places where free groups arise, [Gl92] is a nice
expository article on this topic.

3.2 F3 is a Subgroup of F2
It is easy to see that a free group of rank 2 sits as a subgroup inside
a free group of rank 3 (or rank 4, 5, ...). If {x1 , x2 , . . . , xn } is a basis
for Fn , then no freely reduced word in {x1 , x2 , x−1 −1 ∗
1 , x2 } can equal the
identity. Thus the subgroup generated by {x1 , x2 } is a free group of rank
2. On the other hand, the following result is surprising.

Proposition 3.12. There is a finite index subgroup of F2 that is a free

group of rank 3.

In the remainder of this section we outline a proof of this proposition.

Two key steps are left as exercises that the reader should complete. Our
argument exploits the fact that elements of F2 can be assigned lengths.
Since every g ∈ F2 can be uniquely expressed as a freely reduced word
in the generators and their inverses, we define the length of g to be
the number of terms in this freely reduced expression. We denote this
number by |g|. As an example,

|x3 x−1 y 2 xy −5 y| = |x2 y 2 xy −4 | = 9.

This notion of length has the useful property that |g| = |g −1 |, since
the freely reduced expression for g −1 is the formal inverse of the freely
reduced expression for g.
Let H be the subset of F2 consisting of elements of even length:

H = {g ∈ F2 | |g| is even}.

Since |g| = |g −1 |, this set is closed under taking inverses. Because cancel-
lation in non-freely reduced words occurs in pairs, if |a| and |b| are both
even, |ab| will also be even, although it could be less than |a| + |b|. Thus
H is closed under products and inverses, and therefore it is a subgroup
of F2 . We call H the even subgroup of F2 . Finally, because the length of
every element in F2 is either even or odd, [F2 : H] = 2.

Exercise 3.13. Prove that the even subgroup of F2 is generated by

S = {x2 , xy, xy −1 }. (Hint: Start by proving that every freely reduced
word of length 2 can be expressed as a product of elements of S ∪ S −1 ,
and then proceed by induction.)
66 Groups Acting on Trees

Let a = x2 , b = xy and c = xy −1 , and let ω = w1 w2 · · · wn be a freely

reduced word in the generators {a, b, c} and their inverses. In order to
show that the even subgroup is free with basis {a, b, c}, we need to prove
that ω = e ∈ F2 .
Notice that if you rewrite ω = w1 w2 · · · wn in terms of the original
generators (x and y) there may be cancellation. For example, ω may
contain a−1 b, and expressing a−1 b in terms of the original generators we
get: a−1 b = x−1 x−1 xy = x−1 y. Similarly,



c−1 ac−1 b = (yx−1 )(xx)(yx−1 )(xy) = yx−1 xx yx−1 xy = yxy 2 .

Notice that in this example, while there is cancellation in c−1 a and b−1 a,
there is no cancellation when one rewrites ac−1 in terms of the original
generators.
Converting ω from a freely reduced word in {a, b, c, a−1 , b−1 , c−1 }∗ to
a word in {x, y, x−1 , y −1 }∗ is done by replacing each wi by αi βi where
αi , βi ∈ {x, y, x−1 , y −1 }. In the example above, with the word c−1 ac−1 b,
we had β1 = x−1 = α2−1 and β3 = x−1 = α4−1 . Because of this there
was a small amount of internal cancellation, but it ends up that such
constrained cancellations are the worst that can happen.

Exercise 3.14. Let ω = w1 w2 · · · wn be a freely reduced word in the

free monoid {a, b, c, a−1 , b−1 , c−1 }∗ and let αi βi be the expression for
−1
wi using {x, y, x−1 , y −1 }. By considering cases, show that if βi = αi+1 ,
then:

−1
1. αi = βi+1 ;
−1
2. βi+1 = αi+2 ; and
−1
3. βi−1 = αi .

We can now show that the even subgroup is a free group with basis
{a, b, c}. We first convert a given freely reduced word

ω ∈ {a, b, c, a−1 , b−1 , c−1 }∗

to a word α1 β1 · α2 β2 · · · αn βn ∈ {x, y, x−1 , y −1 }∗ . Given 3.14, we then

know that, when we reduce this word, at most one of the letters, αi or
βi , is cancelled from each pair. Thus

|α1 β1 · α2 β2 · · · αn βn | ≥ n,

hence w = e ∈ F2 .
 3.3 Group Presentations 67

3.3 Free Group Homomorphisms and Group Presentations

Given a group G we describe how one can define a group homomorphism
Fn → G simply by picking targets for a basis of Fn .

Theorem 3.15. Let G be any group and let {g1 , . . . , gn } be a list of

elements of G, which are not necessarily distinct or non-trivial. Let
S = {x1 , . . . , xn } be a basis for a free group Fn . Then there is a group
homomorphism φ : Fn → G where φ(xi ) = gi .

Proof. We know where φ takes the basis {x1 , . . . , xn } and if φ is to be a

homomorphism it follows that φ(x−1 −1
i ) must be gi . If ω is an arbitary
element of Fn then ω can be expressed as ω = w1 w2 · · · wk where each
wi ∈ {X ∪ X −1 }. Define φ(ω) to be

φ(ω) = φ(w1 )φ(w2 ) · · · φ(wn ) ∈ G.

The elements of Fn are equivalence classes of words, where the equiva-

lence relation is generated by

w1 · · · wi xx−1 wi+1 · · · wk ∼ w1 · · · wi wi+1 · · · wk

where x ∈ {X ∪ X −1 }. Since our definition of φ used a specific word to

represent ω, we need to make sure that the result would be the same if
we had chosen a different word to represent ω. However,

φ(ω) = φ(w1 w2 · · · wk ) = φ(w1 )φ(w2 ) · · · φ(wk )

= φ(w1 )φ(w2 ) · · · φ(wi )φ(x)φ(x)−1 φ(wi+1 ) · · · φ(wk )
= φ(w1 w2 · · · wi · x · x−1 · wi+1 · · · wn ).

Thus the element φ(ω) ∈ G does not depend on which word is used to
represent ω. It follows immediately from the definition that φ(ω −1 ) =
−1
[φ(ω)] and φ(ω · ω  ) = φ(ω)φ(ω  ). So the function φ is indeed a group
homomorphism.

It is important to notice that the technique for constructing homomor-

phisms used in Theorem 3.15 does not work in general. For example, let
a and b be two reflections that generate the infinite dihedral group D∞ .
There is no homomorphism φ : D∞ → Z where φ(s) = φ(t) = 1 ∈ Z. If
there were, then φ(s2 ) = 2, but s2 = e ∈ D∞ , so φ would not take the
identity element of D∞ to the identity element of Z. The fact that this
approach does work for free groups has a number of important conse-
quences. One is the following corollary, which indicates why people refer
to Fn as the free group of rank n (and not as a free group of rank n).
68 Groups Acting on Trees

Corollary 3.16. Any two free groups of rank n are isomorphic.

Proof. Let G be a free group with basis {x1 , . . . , xn } and let H be

free with basis {z1 , . . . , zn }. By Theorem 3.15 there is a homomorphism
φ : G → H where φ(xi ) = zi and a homomorphism ψ : H → G where
ψ(zi ) = xi . It follows that ψ ◦ φ is the identity automorphism of G
and φ ◦ ψ is the identity automorphism of H. Thus φ and ψ must be
bijections, and G ≈ H.

There is also the following result, which is the foundation for what are
called group presentations.

Corollary 3.17. If G is generated by n elements then G is a quotient

of Fn .

Proof. Theorem 3.15 establishes that there is a group homomorphism

φ : Fn → G that takes a basis of Fn to a generating set for G. Since
the image of φ contains the generators of G, the image must be all of G.
Thus, by the First Isomorphism Theorem, G ≈ Fn /ker(φ).

Group presentations are not something that we will explore in earnest

in this book but, as we are so close to the basic definition, we might as
well take a minor detour and introduce the topic.
Let G be a finitely generated group, with a chosen set of generators
S = {g1 , . . . , gn }. Then by Corollary 3.17 there is a homomorphism φ
from the free group Fn with basis X = {x1 , . . . , xn } onto G, induced by
requiring φ(xi ) = gi . If ω is a word that maps to the identity in G under
φ, then ω is called a relation. This is a natural phrase, as such words
give rise to equations of the form

“some product of generators and their inverses” = e ∈ G.

For example, if a and b generate Z ⊕ Z then aba−1 b−1 is a relation.

Remark 3.18. Because relations are words that represent the identity,
there is a one-to-one correspondence between relations and paths in the
Cayley graph that begin and end at the vertex associated to the identity.

Let G be a group and φ : Fn → G a surjective map as above. A subset

R ⊂ ker(φ) is a set of defining relations if the smallest normal subgroup
of Fn that contains R is ker(φ). Note that the smallest normal subgroup
containing R must contain both R and R−1 (because it is a subgroup)
as well as every conjugate grg −1 for g ∈ Fn and r ∈ R ∪ R−1 (because
it is normal). In fact, the set of all finite products of all conjugates of
 3.3 Group Presentations 69

the elements of R ∪ R−1 is the smallest normal subgroup containing R.

Thus if R is a set of defining relations then every element in the kernel
can be expressed as a finite product of conjugates of the elements of R
and their inverses.
A finitely generated group G is said to be finitely presented if there
is a finite set of defining relations, R. In this case one indicates these
generators and defining relations by

G = g1 , . . . , gn | ω1 , . . . , ωm 

where {g1 , . . . , gn } is the generating set and R = {ω1 , . . . , ωm }. In gen-

eral, every group has a presentation like that given above, except that
the list of generators and/or the list of relations may not be finite. On the
other hand, most finitely generated groups that you have encountered
to date are finitely presented. For example, Z ⊕ Z = a, b | aba−1 b−1 ,
which is often written a, b | ab = ba. A standard presentation of the
infinite dihedral group is

D∞ = a, b | a2 , b2 

and the finite dihedral groups have presentations

Dn = a, b | a2 , b2 , (ab)n .

We are certainly not claiming these presentations are obvious; it takes

some effort to establish that a given set of relations forms a set of defining
relations. A good introduction to the topic of group presentations is
Johnson’s Presentations of Groups [Jo97].
Group presentations are not just of interest in the study of infinite
groups. They are used, for example, in Conway and Lagarias’s proof of
the following combinatorial theorem:

Theorem 3.19 (See [CL90] and [Th90]). Let Tn be a triangle of hexagons,

with n hexagons on each side, as in Figure 3.7. Then

1. Tn can be tiled by T3 if and only if

n ≡ 0, 2, 9 or 11(mod 12);

2. Tn cannot be tiled by the line of hexagons L3 for any n.

70 Groups Acting on Trees

Fig. 3.7. Two triangles of hexagons are shown: On the left is T4 and on the
top right is T2 . On the bottom right is the line of hexagons L3 .

3.4 Free Groups and Actions on Trees

The following result characterizes free groups via actions on trees. It
gives insight into the nature of free groups and it provides a useful ap-
proach for proving results, notably the Nielsen–Schreier Theorem (Corol-
lary 3.23).

Theorem 3.20. A group G is free if and only if it acts freely on a tree.

In Section 1.8 we proved that if a group acts on a graph then there is a

fundamental domain for that action (see Lemma 1.52 in particular). We
briefly recall the construction of this fundamental domain in the context
of a group G acting freely on a tree T . Start with a fixed vertex v ∈ T ,
and consider subtrees C ⊂ T where

1. v ∈ C; and
2. if x and y are distinct vertices in C then there is no element g ∈ G
such that g · x = y.

There is at least one maximal subtree satisfying these conditions, and

we call it Core. Because the action of G is free, the subtree Core does
not form a fundamental domain; the image of Core under the action
of G contains every vertex of T , but it does not contain every edge. For
each edge e ⊂ Core with e ∩ Core = ∅, let he be the closed half of e
that is connected to Core. Define F to be the union of Core and these
half-edges. The proof of Lemma 1.52 shows:

Lemma 3.21. The subset F ⊂ T described above is a fundamental

domain for the action G
T, and F contains one vertex for each orbit
of vertices under the action of G.
 3.4 Free Groups and Actions on Trees 71

Proof of Theorem 3.20. One direction is easy: if G is a free group, then

it acts freely on its Cayley graph, which is a tree.
Assume then that G acts freely on a tree T and that F is a fundamen-
tal domain as described above. For each half-edge he ∈ F let ge ∈ G be
the element such that F ∩ ge · F is the midpoint of e. Denote the set of
all such ge by S. Since F ∩ ge · F is the midpoint of an edge, ge−1 · F ∩ F
is also the midpoint of an edge. Thus if ge ∈ S so is ge−1 , hence:
Claim: If ge ∈ S then there is a half-edge hê ∈ F such that ge−1 = gê .
The set of elements S is a set of generators of G by Theorem 1.55. As
each element in S can be paired with its inverse, we may take S to be
a maximal subset of S which does not contain both an element and its
inverse. Thus S = S ∪ S −1 .3
Fix a vertex v ∈ Core and for each ge ∈ S ∪ S −1 let Te be the subtree
of T induced by the set of vertices v  ∈ T \ Core such that the reduced
path from v to v  passes through the half-edge he . Equivalently one could
define Te to be the maximal subtree contained in T \ Core that shares a
vertex with the edge e. For example, take the action of F2 on its Cayley
graph. Then Core is a single vertex and the fundamental domain F
consists of this vertex plus four half-edges. The set S will consist of two
elements, and the four subtrees Te are the maximal subtrees contained
in the connected components formed by removing F (see Figure 3.8).
We may now apply the Ping Pong Lemma (Lemma 3.10) to prove
that G is a free group with basis S. Using the notation of the Ping Pong
Lemma:

• for each ge ∈ S, let Xs be the associated subtree Te ;

• take p = v ∈ Core to be a point outside of the union of the

subtrees Te .

Condition 1 of the Ping Pong Lemma follows from the definition of

the subtrees Te . Let ge ∈ S and let Te
be one of the subtrees, where
ge
= ge−1 . Let ê = ge−1 · e denote the edge connecting ge−1 (Core) to
Core. Given any vertex w ∈ Te
there is a minimal edge path connecting
ge−1 · v to w. Applying ge we get a minimal edge path from v to ge · w.
The original path passed through ê = ge−1 · e, hence the the image under
3 We have cheated in this paragraph. We have assumed that there is no g ∈ G such
that g = g −1 . In other words, we are assuming there is no element of order 2 in G.
However, Theorem 3.46 states that when a finite group acts on a tree, it must fix
some point in that tree. It follows that if G contains an element of order 2, then
G cannot act freely on a tree. The reader worried about circularity should work
through the proof of 3.46 before proceeding.
72 Groups Acting on Trees

Fig. 3.8. A fundamental domain for the action of F2 on its Cayley graph is a
cross of four half-edges. For each of these half-edges, there is a corresponding
subset Te ⊂ T . One such subtree is highlighted above.

ge will pass through e. Thus ge · w ∈ Te by construction. Since w was an

arbitrary vertex of Te
, ge (Te
) ⊂ Te , establishing Condition 2.

A careful reading of the proof of Theorem 3.20 shows that it is con-

structive; it provides a method for finding a basis for G.

Example 3.22. Consider the even subgroup H < F2 , where {x, y} is

a basis of F2 , discussed in Proposition 3.12. The elements of H consist
of all elements of F2 of even length. This subgroup acts on the Cayley
graph Γ of F2 and we may begin constructing a core by selecting the
vertex associated to the identity to be part of Core. As we also need a
vertex representing elements of odd length, we add the vertex associated
to x. The edge e joining these two vertices completes the construction
of Core. The fundamental domain F consists of this edge e and the six
half-edges incident with it.
Since there are six half-edges, we immediately know that our basis
will have three elements. To find a basis, we need to find all the g ∈ H
where g · F ∩ F = ∅. The complete list of such elements is

{x2 , x−2 , xy, y −1 x−1 , xy −1 , yx−1 }.

 3.5 The Group Z3 ∗ Z4 73

Fig. 3.9. A Core and fundamental domain for the action of the even subgroup
on the Cayley graph of F2 .

After removing inverses we see that {x2 , xy, xy −1 } is a basis for the even
subgroup of F2 .
We end this section with a striking result about free groups.
Corollary 3.23 (Nielsen–Schreier Theorem). Every subgroup of a free
group is free.
Proof. Let H be a subgroup of a free group F. The group F acts freely
on its Cayley graph, and therefore H acts freely on the Cayley graph of
F. Thus, by the characterization of free groups given by Theorem 3.20,
H is a free group.
This result was first proved by Jakob Nielsen for finitely generated
free groups in 1921. His result was extended to all free groups by Otto
Schreier a few years later.

3.5 The Group Z3 ∗ Z4

The infinite dihedral group, D∞ , can be thought of as a subgroup of
Sym+ (T2,2 ), where T2,2 is the bipartite tree where every vertex has
valence 2. (In fact, in Exercise 2 of Chapter 2 you were asked to prove
74 Groups Acting on Trees

Fig. 3.10. The element x2 can be chosen to be part of a basis for the even
subgroup of F2 since x2 · F intersects F at the rightmost point of F. Because
xy·F intersects F at the highest point highlighted above, and xy −1 ·F intersect
F at the lowest point highlighted above, we know that {x2 , xy, xy −1 } is a basis.

that D∞ ≈ Sym+ (T2,2 ).) In this section we explore a similar group,

defined in terms of symmetries of the biregular tree T3,4 . Let a be a
rotation about a vertex of valence 4 in T3,4 and let b be a rotation about
an adjacent vertex of valence 3. Let G be the subgroup of Sym(T3,4 )
generated by a and b.
In order to avoid confusion about what is meant by a “rotation” of
T3,4 , we introduce labels for the vertices. Every vertex in T3,4 can be
reached by a unique, reduced path starting at the vertex fixed by a.
Such a path starts by going North, East, South or West. Then one
chooses to go either Left or Right. After the Left/Right choice, there are
three directions available that do not involve backtracking. If we always
pretend that the edge we are on corresponds to East, keep the clockwise
orientation and use the pneumonic “Never Eat Sour Watermelons,” then
we can unambiguously move forward by going South, West or North.
Every vertex of T3,4 then has a unique label involving an alternating
sequence of cardinal directions and left/right choices. These coordinates
are illustrated in Figure 3.12. The only vertex that does not have a label
is the vertex fixed by a, which we will refer to as “the origin.”
 3.5 The Group Z3 ∗ Z4 75

a
e
b

Fig. 3.11. The “rotations” a and b in Sym(T3,4 ).

NLN NRS
NLW NRW
NL NR
N
WRW NLS NRN ELW
WRN ELS
WR EL
WRS ELN
W E
WLN ERS
WL ER
WLS ERN
WLW SRN SLS ERW
S
SR SL
SRW SLW
SRS SLN

Fig. 3.12. Labelling the vertices of T3,4 using left/right and the four cardinal
directions.
76 Groups Acting on Trees

The element a fixes the origin, and otherwise it applies the permuta-
tion N ⇒ E ⇒ S ⇒ W ⇒ N to the first letter labelling a vertex v. So,
as examples, under the action of a the vertex labelled NRW is taken to
the vertex labelled ERW and the vertex labelled WLS is taken to the
vertex labelled NLS.
The action of b is a bit more complicated to describe in terms of these
coordinates. In order to avoid repeatedly writing “the vertex labelled by
” we conflate vertices with their labels. The vertex E is fixed by b. The
origin is taken to EL; the vertex EL is taken to ER; and ER is taken to
the origin. The rules for determining the action of b on the remaining
vertices are:
1. If a vertex has “EL” as a prefix, it is taken to the vertex obtained
by replacing this prefix by “ER.”
2. If a vertex has “ER” as a prefix, strip off this prefix.
3. If v has any other prefix, then add EL as a prefix. (The vertex
WL is taken to ELWL, for example.)
To get a little bit of practice with this group action, you should show:
1. The element aba−1 is a rotation about the vertex labelled S in
Figure 3.12.
2. The element a2 ba−2 is a rotation about the vertex W.
3. The element a3 ba−3 is a rotation about the vertex N.

The group G, generated by the rotations a and b, is referred to as the

free product of Z3 and Z4 , and it is denoted Z3 ∗ Z4 . The Z3 refers to
the fact that b ∈ G generates a cyclic subgroup of order 3 and the Z4
refers to the fact that a ∈ G generates a cyclic subgroup of order 4. (We
discuss free products of groups in general in the next section.)
Because a and b generate Z3 ∗ Z4 , and a has order 4 and b has order
3, any g ∈ Z3 ∗ Z4 can be expressed as a product g = ai1 bj1 · · · ain bjn
where 1 ≤ ik ≤ 3 and 1 ≤ jk ≤ 2, except possibly i1 = 0 and/or jn = 0.
Define the norm 4 of such an expression to be the number of non-zero
exponents, and denote this by ||si1 tj1 · · · sin tjn ||. For example,
||st3 s2 tst3 s2 || = 7 .
Let e be the edge joining the vertex fixed by a to the vertex fixed by
b or, equivalently, the edge joining the origin and the vertex labelled E
4 What we are calling the “norm” is also frequently referred to as the “syllable
length” of a word.
 3.5 The Group Z3 ∗ Z4 77

in Figure 3.12. By reading a word ω = ai1 bj1 · · · ain bjn left-to-right one
can easily determine the image of e under the action of ω.
Assuming that i1 = 0, one rotates e about the white vertex fixed by
a as far as i1 dictates. Denote the resulting edge e . Rotate e around
its black vertex as far as is dictated by j1 . Denote the resulting edge
e , and then rotate e about its white vertex as far as is dictated by
i2 , and so on. Similarly, if i1 = 0, and so ω really should be expressed
as ω = bj1 ai2 · · · ain bjn , then one applies the same process, starting by
rotating about the black vertex of e as far as j1 dictates, then rotating
about the resulting white vertex as far as i2 dictates, and so on. For
example, if ω = a2 ba then the image of e under the action of ω is the
edge f shown in Figure 3.13.

e
b
a2
a
f

Fig. 3.13. The element a2 ba in Z3 ∗ Z4 takes the edge e to the edge f . The
arrows are labelled using the “two wrongs make a right” principle.

You should be very suspicious of the previous description. Mixing up

right and left can lead to errors when one is dealing with non-abelian
groups. Yet it appears that we have done just that and we have also
seemingly confused our definitions. In considering ω = a2 ba it appears
that we are letting “b” denote a rotation about the vertex labelled W.
78 Groups Acting on Trees

But b actually rotates about the vertex labelled E. These two errors
together give the right answer by the “two wrongs make a right” principle
that we previously encountered in our discussion of reflection groups
(Chapter 2). You can see why this works if you notice that ω can be
rewritten as a product of conjugates:
ω = a2 ba = (a2 bab−1 a−2 )(a2 ba−2 )a2
The rightmost a2 gives the rotation about the unlabelled vertex; the term
(a2 ba−2 ) is the rotation falsely labelled “b” in Figure 3.13; the final ex-
pression (a2 bab−1 a−2 ) is the rotation falsely labelled as “a”
Figure 3.13.
This “two wrongs make a right” principle is actually just an appli-
cation of a broader principle, namely, “conjugation is doing the same
thing somewhere else.” If g ∈ Z3 ∗ Z4 takes the unlabelled vertex fixed
by a to some vertex v, then gaj g −1 will be a rotation about v. Thus
gaj = (gaj g −1 )g will take the unlabelled vertex to v (via g) and then
rotate about v (via gaj g −1 ).
This description of how to determine where the edge e goes under
the action of ω = ai1 bj1 · · · ain bjn is the key insight for establishing the
following lemma.
Lemma 3.24. The group Z3 ∗ Z4 acts transitively on the edges of T3,4 .
Further, let e be the edge joining the vertex stablized by a to the
vertex stabilized by b, and let g = ai1 bj1 · · · ain bjn (as prescribed above)
with n = ||ai1 bj1 · · · aik bjk ||. Then the number of vertices between e and
g · e is n.
Proof. Given any edge f we can use the method described above to write
down an element of gf ∈ Z3 ∗ Z4 that will take e to f . Thus the product
gf · gf−1

will take an edge f  to f , so the action is edge-transitive. The
claim about the norm determining the distance is immediate from the
description of how one finds the image of e under the action of g.
Proposition 3.25. The action of Z3 ∗ Z4 on T3,4 is simply transitive
on the edges. Further, any g ∈ Z3 ∗ Z4 can be uniquely expressed as
a word σ s1 τ t1 · · · σ sn τ tn where s1 and/or tn can be zero, but otherwise
1 ≤ si ≤ 2 and 1 ≤ ti ≤ 3.
Proof. That the action is transitive is established by Lemma 3.24. To
establish that it is simply transitive, it suffices to show that no element
of Z3 ∗ Z4 can have two such expressions:
g = si1 tj1 · · · sin tjn = sk1 tl1 · · · skm tlm .
 3.6 Free Products of Groups 79

If there was such a pair of expressions, then

si1 tj1 · · · sin tjn · t−lm s−km · · · t−l1 s−k1 = e.

But then by Lemma 3.24 the norm of this product must be zero. This is
only possible if tjn · t−lm = e, that is, jn = lm , and then sin · s−km = e,
hence in = km , and so on until there has been complete cancellation.
This contradicts the assumption that g has two distinct expressions.

3.6 Free Products of Groups

A group G is the direct product of two subgroups, A and B, if both A
and B are normal, G = A · B, and A ∩ B = e. It makes sense to refer to
G as a product, since in essence it is produced directly from the groups
A and B. A distinguishing feature of the direct product is the fact that
elements of A commute with elements of B in G = A ⊕ B.
In this section we introduce the free product of two groups A and B,
denoted A∗B. The free product of A and B can be thought of as a group
generated by a copy of A and a copy of B where there is no interaction –
such as commuting – between the subgroups A and B.
To describe the elements of the free product A ∗ B we start with an
alphabet that is the union of the elements of A and B, {A ∪ B}. A word
ω = x1 x2 · · · xn ∈ {A ∪ B}∗ is a reduced word if

1. no xi = eA or eB (where eA is the identity element of A and eB

is the identity element of B); and
2. when xi ∈ A then xi+1 ∈ A, and similarly if xi ∈ B then
xi+1 ∈ B.

In other words, a reduced word is one that alternates between non-

identity elements of A and B.
We would like to take the set of reduced words to be the elements of
the free product A ∗ B, with the binary operation being concatenation.
But there is a minor difficulty, in that the concatenation of two reduced
words might not be reduced. Thus we introduce the following equivalence
relation. Let ∼ be the equivalence relation on {A ∪ B}∗ generated by

1. ωex ω  ∼ ωω  where ex is either eA or eB ;

2. ω · ab · ω  ∼ ω · c · ω  where ab = c holds in either the group A or
the group B.
80 Groups Acting on Trees

We define the elements of A ∗ B to be the equivalence classes [ω] of

words ω ∈ {A ∪ B}∗ . The binary operation is defined by taking the
equivalence class of the concatenation of two words: [ω1 ][ω2 ] = [ω1 ω2 ].
(The reader should check that this definition does not depend on the
choice of representatives of the equivalence classes.) The identity element
is the empty string, and if ω = x1 x2 · · · xn then the inverse of [ω] is the
−1
equivalence class of its formal inverse ω −1 = x−1n · · · x1 .

Theorem 3.26. Each equivalence class [ω] for ω ∈ {A ∪ B}∗ contains

exactly one reduced word.
Proof. We first prove that every word ω is equivalent to a reduced word
(ω). We do this by induction on the length of the word ω. If ω is the
empty word or is of length one, then ω is reduced and we may take
(ω) = ω. For any x ∈ A ∪ B and any reduced word ω = x1 · · · xn define
the reduced word (xω) by


 ω if x = eA or eB



 xω if x ∈ A \ {eA } and x1 ∈ B


or x ∈ B \ {eB } and x1 ∈ A
(xω) =

 yx2 · · · xn if x, x1 ∈ A and x · x1 = y = eA



 or x, x1 ∈ B and x · x1 = y = eB


x2 · · · xn if x = x−1
1 .

For an arbitrary word ω we may now define (ω) inductively. Write

ω = xω  , where x is the leading letter of ω, and set (ω) = (x (ω)).
By induction ω  ∼ (ω  ) and therefore xω  ∼ x (ω  ). Thus we have ω =
xω  ∼ (x (ω  )), and this last expression is by definition a reduced word.
The argument that there is a unique reduced word in each equiva-
lence class mimics the proof of Cayley’s Basic Theorem. Let ΩR be the
collection of all reduced words in {A ∪ B}∗ . Let x ∈ A ∪ B and let πx
be the permutation of ΩR described by πx : ω → (xω). For any word
ω = x1 · · · xn define the permutation πω as πω = πx1 · · · πxn . By con-
sidering the generating conditions (1) and (2) above, one can show that
if ω ∼ ω  then πω is the same permutation as πω
. Since the elements
of A ∗ B are the equivalence classes of words, it follows that we have
constructed a representation π : A ∗ B → Sym(ΩR ).
To see that no equivalence class contains two distinct reduced words,
it suffices to show that two distinct reduced words induce distinct per-
mutations. (In other words, we show that the representation above is
faithful.) But if ω is any reduced word, and ∅ represents the empty
word, then πω (∅) = ω. Thus if ω1 and ω2 are distinct reduced words,
 3.6 Free Products of Groups 81

then πω1 = πω2 , as these permutations do not agree on the reduced word
∅. So no two distinct, reduced words can represent the same element
of A ∗ B.
Corollary 3.27. The groups A and B inject into A ∗ B.
Proof. The elements of A and B form reduced words (of length 1).
Theorem 3.28. Every free product of groups A ∗ B can be realized as
a group of symmetries of a biregular tree T , and a fundamental domain
for this action consists of a single edge and its two vertices. Further, if
A and B are both finite, this tree is T|A|,|B| .
Proof. Define V (T ) to be the the left cosets gA and gB. In order to have
a bit of notation, we let vgA denote the vertex associated to the coset
gA and similarly vgB is the vertex associated to gB.
The edges of the tree T correspond to the elements of A ∗ B, and we
use eg to denote the edge associated to g ∈ A ∗ B. This does have the
unfortunate consequence that the edge associated to the identity is “ee .”
We define Ends(eg ) = {vgA , vgB }. The bipartite structure of T comes
from considering the vertices associated to cosets of A as one “part” and
the vertices associated to cosets of B as the other “part.”
Note that two edges eg and eh intersect if and only if g −1 h ∈ A or B.
For example, if eg ∩ eh = vkA then g ∈ kA and h ∈ kA, hence g −1 h ∈ A.
Conversely, if g −1 h ∈ A then gA = hA and so vgA = vhA . The same
argument holds mutatis mutandis for B.
The action of A ∗ B on T is induced by the standard left action of
a group on itself and on its left cosets. An element g ∈ A ∗ B moves
vertices by g : vhA → vghA and g : vhB → vghB ; the edges move by
g : eh → egh . As in the case of the action of a group on any of its Cayley
graphs, we get an action G
T primarily because the group is per-
muting labels by left multiplication, while incidence is defined by right
multiplication. It follows from the description of this action that the edge
ee , which joins vA to the vB , is a fundamental domain for the action of
A ∗ B on T .
To prove that the graph T is a tree we need to show that it is
connected and contains no cycles. If g ∈ A ∗ B then g = [ω] for a
reduced word ω = x1 x2 · · · xn . Assume for convenience that x1 ∈ A,
x2 ∈ B, and so on. Then ee ∩ ex1 = vA , ex1 ∩ ex1 x2 = vx1 B , and so
on. Hence {ee , ex1 , ex1 x2 , . . . , ex1 x2 ···xn } is an edge path from ee to eg . It
follows that any edge eg can be joined to ee by an edge path, hence T is
connected.
82 Groups Acting on Trees

We use Theorem 3.26 to show that T has no circuits. If there is a

circuit in T , then using the action A ∗ B
T we may move this
to a circuit that begins with the edge corresponding to the identity:
{ee , e1 , . . . , en , ee }. Then e1 = ex1 for some x1 ∈ A or B and, without
loss of generality, we will assume x1 ∈ A. Then e2 = ex1 x2 , for some
x2 ∈ B, and in general ek = ex1 x2 ···xk where x1 x2 · · · xk is a reduced
word in {A ∪ B}∗ . Since this edge path is assumed to be a circuit,
ex1 x2 ···xn ∩ ee = vA or vB . Hence x1 x2 · · · xn = x for x ∈ A ∪ B. But
this is impossible, as this would imply that we have two distinct reduced
words in an equivalence class.
To establish the final claim, notice that vgA ∈ Ends(eh ) if and only
if h ∈ gA. Thus the number of edges incident with vgA is the size of
the coset gA, which is the same as the order of A. The same argument
shows that the degree of vertices of the form vg B is |B|.

Let’s revisit the example of Z3 ∗ Z4 considered in the previous section.

The edges of the tree T associated to a free product A∗B (Theorem 3.28)
correspond to elements of A ∗ B, which in turn correspond to reduced
words. Our labelling of the vertices of T3,4 in Section 3.5 secretly ex-
ploited this fact. One can see this by making the substitutions

E → e, S → a, W → a2 , and N → a3

along with
L → b, and R → b2 .

The edge joining the origin to the vertex associated to E is the edge ee
from Theorem 3.28; the edge joining the origin to S is ea ; and the edge
joining W L to W LS is ea2 ba , as is shown in Figure 3.13. The action of a
and b on T3,4 , described in Section 3.5 in terms of adjusting labels, can
be reconstructed by simply applying the above translation scheme with
left multiplication.

Exercise 3.29. Pick two finite groups, A and B, that you are comfort-
able with. Construct the tree T , and the action A ∗ B
T , given by
Theorem 3.28. Be as concrete as possible.

Corollary 3.30. The stabilizers of the vertices of the tree associated to

a free product A ∗ B are conjugates of A and B.

Proof. Let h be an element of the subgroup gAg −1 . Then

h · vgA = gag −1 · vgA = ga · vA = g · vA = vgA .

 3.7 Free Products of Finite Groups 83

Thus gAg −1 is a subgroup of the stabilizer of vgA , Stab(vgA ). A similar

computation shows that g −1 Stab(vgA )g < A. Thus
A < g −1 Stab(vgA )g < A
hence Stab(vgA ) = gAg −1 .
The same argument shows that Stab(vgB ) = gBg −1 .
Exercise 3.31. Let e be any edge in the tree T associated to a free
product A ∗ B. Show that Stab(e) is trivial.
Remark 3.32. Theorem 3.26, which establishes a one-to-one correspon-
dence between elements of A ∗ B and reduced words, comes from work
of van der Waerden in the late 1940s. The idea of free products, how-
ever, dates at least to the 1920s, and was expressed quite clearly in a
foundational paper by Kurosch in 1933 [Ku33]:
Die Gruppe G heißt freies Produkt ihrer Untergruppen Hα ... wenn jedes
Element g von G auf eine und nur eine Weise als Product
g = h1 h2 · · · hn , hi = 1, hi ∈ Hαi , Hαi = Hαi+1
darstellbar ist; es gibt also in G keine nichtidentische Relationen, welche Ele-
mente von verschiedenen Hα verbinden.

3.7 Free Products of Finite Groups are Virtually Free

If P is a group-theoretic property, then a group G is said to be virtually
P if it has a finite-index subgroup H < G where H is a P-group. For
instance, a group G is virtually abelian if it contains an abelian subgroup
of finite-index. Example 3.36 shows there is a finite-index subgroup of
D∞ that is isomorphic to Z, hence the infinite dihedral group is virtually
cyclic. The goal of this section is to prove that every free product of
finite groups is virtually free, or, in other words, that such free products
contain finite-index subgroups which are free.
Corollary 3.27 says that A and B inject into the free product A ∗ B.
Let ιA : A → A ∗ B be the homomorphism sending a ∈ A to the element
of A ∗ B corresponding to the reduced word “a”; let ιB : B → A ∗ B
be the same sort of map for B. Exercise 19 asks you to provide a proof
of the following theorem, which is the analog of Theorem 3.15 in the
context of free products.
Theorem 3.33. If there are group homomorphisms φA : A → G and
φB : B → G then there is a map φ : A ∗ B → G such that φA = φ ◦ ιA
and φB = φ ◦ ιB .
84 Groups Acting on Trees

Example 3.34. Let x generate a group A of order 2 and let y gener-

ate another group B which is also of order 2. Then x and y (or more
accurately ιA (x) and ιB (y)) generate Z2 ∗ Z2 . Denote the generators of
the infinite dihedral group, D∞ , as a and b, just as in Chapter 2. By
Theorem 3.33 there is a homomorphism φ : Z2 ∗ Z2 → D∞ induced by
declaring φ(x) = a and φ(y) = b. Since a and b generate D∞ this homo-
morphism is surjective. The map φ is injective because every g ∈ Z2 ∗ Z2
can be represented by a reduced word, which in this case means a word
that alternates x’s and y’s. The image is then a word that alternates a’s
and b’s and no such word represents e ∈ D∞ by Proposition 2.1.

Given two groups A and B there are “coordinate inclusions” mapping

the groups A and B into A ⊕ B:

φA (a) = (a, eB ) and φB (b) = (eA , b).

Let φ : A∗B
A⊕B be the map induced by these coordinate inclusions.

As an example, if g = a1 b1 a2 b2 a3 ∈ A ∗ B then

φ(g) = (a1 a2 a3 , b1 b2 ) ∈ A ⊕ B.

This map cannot be injective (if A and B are non-trivial). Take any two
non-identity elements a ∈ Z \ {eA } and b ∈ B \ {eB }, and form their
commutator, which is defined as

[a, b] = a−1 b−1 ab.

The word a−1 b−1 ab is a reduced word, hence a−1 b−1 ab = e ∈ A ∗ B.

However, φ([a, b]) = (a−1 a, b−1 b) = e ∈ A ⊕ B. Thus the kernel of φ
contains more than the identity element.

Theorem 3.35. Let φ : A ∗ B
A ⊕ B be the map induced by the

coordinate inclusions of A and B into A⊕B. The kernel of φ is generated
by the set of commutators of non-identity elements of A and B

S = {[a, b] | a ∈ A \ {eA } and b ∈ B \ {eB }}

and this kernel is a free group with basis S.

Example 3.36. In the case of D∞ ≈ Z2 ∗ Z2 the map φ is onto Z2 ⊕ Z2

and by inspection one sees that g ∈ D∞ is in the kernel if and only
if the reduced expression for g has an even number of a’s and an even
number of b’s. From this, it is not too difficult to believe that the kernel
is generated by the element [a, b] = abab, and therefore the kernel is
isomorphic to Z.
 3.7 Free Products of Finite Groups 85

We now present a proof of Theorem 3.35 based on the definition of a

free group; a geometric argument is outlined in Section 3.8.

Lemma 3.37. Let K denote the subgroup of A∗B generated by S. Then

K is a normal subgroup.

Proof. Let α ∈ A and [a, b] be a commutator in S. Then

α[a, b]α−1 = αa−1 b−1 abα−1

= αa−1 b−1 (aα−1 b · b−1 αa−1 )abα−1
= (αa−1 b−1 aα−1 b)(b−1 αa−1 abα−1 )
= (αa−1 b−1 aα−1 b)(b−1 αbα−1 )
= [aα−1 , b][α−1 , b]−1 .

Thus conjugating a commutator by any α ∈ A results in a product of a

commutator and the inverse of a commutator, which is in K. A similar
string of equations shows that conjugating a commutator by β ∈ B
results in an element of K:

β[a, b]β −1 = [a, β −1 ]−1 [a, bβ −1 ].

Since A and B generate A ∗ B, an induction argument shows that

gKg −1 < K for any g ∈ A ∗ B.

Lemma 3.38. The kernel of φ is the subgroup K generated by S.

We leave the proof of this lemma as an exercise:

Exercise 3.39. Let A be the subgroup of A ∗ B/K consisting of cosets

of the form {aK | a ∈ A}, and let B = {bK | b ∈ B}.

a. Show that A · B = A ∗ B/K.

b. Show that A ∩ B = eK.

It follows from the definition of a direct sum that A ∗ B/K ≈ A ⊕ B.

You have established Lemma 3.38 once you have proved:

c. A ≈ A and B ≈ B.

Lemma 3.40. The subgroup K is a free group with basis S.

Proof. Our proof is by induction on the length of a reduced word in

{S ∪ S −1 }∗ . Let

ω = [a1 , b1 ]1 [a2 , b2 ]2 · · · [an , bn ]n

86 Groups Acting on Trees

where i = ±1 and no commutator is listed next to its inverse. Thought

of as an element in {A ∪ B}∗ , this word may not be reduced, but we can
show it is in the same equivalence class as a reduced word (ω) whose
length is at least n + 3. Thus ω = e ∈ A ∗ B and therefore K is free with
basis S.
Our proof establishes slightly more than is claimed above. Our induc-
tion hypothesis is: ω is equivalent to a reduced word of length ≥ n + 3
and the reduced expression (ω) ends with an bn if n = +1 and bn an if
n = −1. The base case is immediate.
We assume that any reduced word in {S ∪S −1 }∗ of length ≤ n satisfies
the induction hypothesis, and consider

ω = [a1 , b1 ]1 [a2 , b2 ]2 · · · [an , bn ]n [an+1 , bn+1 ]n+1 .

With only a small loss of generality, we assume that n = +1. Hence ω

is equivalent to the word

x1 · · · xk an bn [an+1 , bn+1 ]n+1

with x1 · · · xk an bn a reduced word of length ≥ n + 3.

If n+1 = +1, then we may take

(ω) = x1 · · · xk an bn a−1 −1
n+1 bn+1 an+1 bn+1 ,

which is reduced, of length ≥ n + 7, and ends with the correct letters. If

n+1 = −1 then

ω ∼ x1 · · · xk an bn b−1 −1
n+1 an+1 bn+1 an+1 .

If bn = bn+1 we can let b = bn b−1

n+1 ∈ B \ {eB } and then

(ω) = x1 · · · xk an ba−1
n+1 bn+1 an+1 ,

which is reduced, of length ≥ n + 6, and ends with the correct letters.

Finally, if bn = bn+1 , then an = an+1 , as ω was freely reduced when
thought of as a product of commutators. We may set

a = an a−1
n+1 ∈ A \ {eA }

and then
(ω) = x1 · · · xk abn+1 an+1 ,

which is reduced, of length ≥ n + 4 = (n + 1) + 3, and ends with the

correct sequence of letters.
 3.8 A Geometric View of Theorem 3.35 87

We can now establish the title of this section:

Corollary 3.41. If A and B are finite, then A ∗ B contains a normal
subgroup of index |A| × |B| that is free of rank (|A| − 1) × (|B| − 1).
Proof of Corollary 3.41. The index of the kernel is the size of the quo-
tient, which is |A| × |B|. The size of S is (|A| − 1) × (|B| − 1) by its
definition.
Example 3.42. The kernel of Zn ∗ Zm
Zn ⊕ Zm is free of rank
(n − 1)(m − 1).

3.8 A Geometric View of Theorem 3.35

The proof of Theorem 3.35 given in Section 3.7 is algebraic. There is
also a geometric argument which we outline here in the special case of
Z3 ∗ Z4 . Recall our notation for the generators: the generator a fixed a
vertex of valence 4 and b fixed a vertex of valence 3 in T3,4 . We will show:
Proposition 3.43. Let φ : Z3 ∗ Z4
Z3 ⊕ Z4 be the map induced by
the “coordinate” inclusions of Z3 and Z4 into Z3 ⊕ Z4 . The kernel of φ
is generated by the set of commutators of non-identity elements of Z3
and Z4
S = {[a, b] | a ∈ Z4 \ {0} and b ∈ Z3 \ {0}}
and this kernel is a free group with basis S.
The proof begins by noticing that K = Ker(φ) acts on T3,4 since it is
a subgroup of Z3 ∗ Z4 .
Lemma 3.44. The kernel K acts freely on T3,4 .
Proof. We need to show that no non-identity element fixes any vertex
or edge in T3,4 . If e ∈ E(T3,4 ) then Exercise 3.31 shows that Stab(e) is
trivial. Corollary 3.30 shows that the vertex stabilizers are conjugates of
Z3 and Z4 . However, φ(g · ai · g −1 ) = (0, i) ∈ Z3 ⊕ Z4 for any g ∈ A ∗ B.
Similarly φ(g · bj · g −1 ) = (j, 0) ∈ Z3 ⊕ Z4 . Thus the intersection of any
vertex stabilizer with the kernel K is just the identity element. Thus no
element of the kernel K stabilizes a vertex. It follows that the action
K
T3,4 is free.
Since K acts freely on a tree, Theorem 3.20 implies the following
result.
Corollary 3.45. The kernel K is a free group.
88 Groups Acting on Trees

Theorem 3.20 also gives a concrete way to find a basis for K using a
fundamental domain. In describing a fundamental domain for the action
K
T3,4 we use the notation introduced in Section 3.5. In particular, a
is the generator of order 4 and b is the generator of order 3. To construct
a Core for K
T3,4 we begin with the vertex we called “the origin” and
we add edges and vertices, until we can no longer add vertices coming
from distinct orbits.

Fig. 3.14. A Core and fundamental domain for the action of the kernel K on
T3,4 . The dotted line indicates two half-edges that are connected by the action
of K.

One possible Core is shown in Figure 3.14. At first glance, our Core
may appear to be too small. All of the vertices in Core are contained
in ai · ee or bj · ee , where ee is the edge joining the vertex fixed by a to
the vertex fixed by b. The vertex labelled NL in Figure 3.12 is an end of
a3 b · ee , and the image of a3 b in Z3 ⊕ Z4 is not the same as the image
of a single power of a or b. However, the vertex labelled NL is also an
end of a3 ba · ee , and the image of a3 ba is the same as the image of b in
the quotient. One further check that the Core, and the corresponding
fundamental domain formed by adding half-edges, is correct comes from
 3.9 Finite Groups Acting on Trees 89

counting the vertices of each type. Because the index of Z3 in Z3 ⊕ Z4 is

four, there must be exactly four vertices of valence 3 in Core. Similarly
there need to be exactly three vertices of valence 4.
To find the basis associated with our fundamental domain, we need
to see how the half-edges pair up. The half-edge connecting the fun-
damental domain to the vertex labelled NL comes from the element
a3 b ∈ Z3 ∗ Z4 . The half-edge connecting the fundamental domain to the
vertex labelled ELE comes from the element ba3 . But in the quotient
group, Z3 ⊕ Z4 , a3 b = ba3 , and so these half-edges are paired. The corre-
sponding element of the basis for the kernel K is the commutator [b, a3 ].
Continuing in this manner one gets the basis claimed in Proposition 3.43.

3.9 Finite Groups Acting on Trees

Given any group G, there is a tree T , and a non-trivial action of G on T .
One way to construct such an action on a tree is by having the vertex set
of T correspond to the elements of G, with an additional distinct vertex
“ .” For each g ∈ G there is an edge joining to the vertex associated
with g, and there are no other edges. The action is induced by declaring
that g ∈ G takes the vertex associated to h to the vertex associated
to gh, and it takes to . With this definition, we have constructed an
injective map from G to Sym(T ), hence the claim that the action G
T
is “non-trivial.” However, there is a sense in which this action is trivial:
there is a point that is never moved by the action of G: Stab( ) = G.
Jean-Pierre Serre defined a group G to have Property FA if, whenever
there is an action of G on a tree T , there is a point x ∈ T such that
Stab(x) = G. (The letter “F” corresponds to “fixed” while the “A”
comes from the French word for tree: arbre.)
Theorem 3.46. All finite groups have Serre’s Property FA.
Proof. Let G be finite and let G
T be an action of G on a tree. Let
v be any vertex of T and consider the orbit of v under the action of G:
O(v) = {g · v | g ∈ G} .
Let TO(v) be the subtree created by taking the union of all the vertices
and edges of T that occur in minimal-length paths connecting elements
of O(v). Since g · O(v) = O(v) for any g ∈ G, it follows that G acts
on TO(v) .
If we have been lucky, and O(v) = v, we have our fixed point.
Otherwise TO(v) contains vertices of valence 1. (Vertices of valence 1
90 Groups Acting on Trees

are called leaves of the tree.) The action of G must permute the leaves
of TO(v) , hence it also permutes all the non-leaf vertices of TO(v) . Let
T  be the tree formed by removing the leaves of TO(v) and the edges
incident to these leaves. We now have an action of G on T  . If T  is not
an edge or a single vertex, then we can remove the leaves of T  to form
a smaller subtree T  on which G acts.
Because G is finite, O(v) and TO(v) are finite as well. Each time we
trim off leaves, we create even smaller subtrees, and so by continuing in
this manner we eventually reduce our original subtree to either a single
vertex or a single edge. If it is an edge, then either all of the elements
of G fix the edge, or some elements of G invert the edge. In either case,
the center point of the edge is fixed.

Examining the proof of Theorem 3.46 we see that G either fixes a

vertex of T or there is an edge e whose midpoint is fixed by the action
of G. If T is bipartite and G preserves the bipartite structure, then this
second possibility cannot happen, and so we get:

Corollary 3.47. If T is a bipartite tree, and G is a finite group con-

tained in Sym+ (T ), then G fixes a vertex.

Corollary 3.48. If H is a finite subgroup of A ∗ B then H is conjugate

to a subgroup of A or B.

Proof. The vertex stabilizers are conjugates of A and B for the action
of A ∗ B on its associated tree, by Corollary 3.30. By Corollary 3.47, a
finite subgroup of A ∗ B must fix a vertex, hence H < gAg −1 or gBg −1
for some g ∈ A ∗ B.

3.10 Serre’s Property FA and Infinite Groups

It is not just the finite groups that have Serre’s Property FA. In fact,
a number of interesting infinite groups are also unable to act on a tree
without a global fixed point. Before we can present a few examples of
such groups, we need to establish a small amount of background mate-
rial.

Definition 3.49. Let G act on X. Then for any element g ∈ G define

the fixed subset to be

Fix(g) = {x ∈ X | g · x = x}.
 3.10 Serre’s Property FA and Infinite Groups 91

If H is any subgroup of G then set


Fix(H) = Fix(h).
h∈H

Serre’s definition of Property FA used the notion of fixed subsets,

although his notation was different from that given above. If G
X
then he used X G to denote Fix(G), as in the following definition from
[Se77]:

Nous nous intéressons aux groupes G ayant la propriété: (FA) – Quel que soit
l’arbre X sur lequel opère G, on a X G = ∅.

Lemma 3.50. Let G act on a tree T . Then either Fix(G) is empty, or

it is a subtree of T .

Proof. Assume Fix(G) is not empty. If v and w are two distinct fixed
points in Fix(G), then there is exactly one minimal-length edge path
containing v and w. Since this edge path is unique, it must also be fixed
(pointwise) by G. Thus Fix(G) is a connected subgraph of T , hence it
is a subtree.

If S is a subset of G we have used S to denote the subgroup of G

generated by the set S. In the following lemma we use this notation in the
context of two subgroups, H, K < G. Namely, H, K is the subgroup of
G generated by the elements in the set H ∪K. In particular, the elements
of H, K consist of all possible products of finitely many elements from
H and K.

Lemma 3.51. Let G act on a tree T and let H and K be two subgroups
of G. Then
Fix(H, K) = Fix(H) ∩ Fix(K).

Proof. If x ∈ Fix(H) ∩ Fix(K), then x is fixed by every h ∈ H and

k ∈ K, hence x is fixed by every element of H, K. Thus we have
Fix(H, K) ⊃ Fix(H) ∩ Fix(K). To establish the other containment,
notice that, because H and K are subgroups of H, K, it follows that
Fix(H, K) ⊂ Fix(H) and Fix(H, K) ⊂ Fix(K). Thus we have the
other containment: Fix(H, K) ⊂ Fix(H) ∩ Fix(K).

Proposition 3.52 (Triangle Condition). Let G be generated by three

subgroups, H1 , H2 and H3 . Define Hij to be Hi , Hk . If all of the Hi
and Hij are finite, then G has Serre’s Property FA.
92 Groups Acting on Trees

Proof. We begin by showing that

Fix(H1 ) ∩ Fix(H2 ) ∩ Fix(H3 ) = ∅.
Since each Hi is assumed to be finite, Fix(Hi ) = ∅ by Theorem 3.46.
Further, because Fix(Hi ) ∩ Fix(Hj ) = Fix(Hij ) (Lemma 3.51) and
Fix(Hij ) = ∅ (because Hij is finite), we have
Fix(Hi ) ∩ Fix(Hj ) = ∅.
We may then take v to be a point in Fix(H1 ) ∩ Fix(H2 ) and w to be a
point in Fix(H2 ) ∩ Fix(H3 ). Thus v and w, as well as the minimal-
length edge path containing v and w, are all contained in Fix(H2 ).
However, v and w are also in Fix(H1 ) ∪ Fix(H3 ). As this union is a
subtree, the minimal-length edge path containing v and w is also in
the union Fix(H1 ) ∪ Fix(H3 ). It follows that this edge path must inter-
sect Fix(H1 ) ∩ Fix(H3 ), implying that there is a point of intersection of
Fix(H2 ) with Fix(Hi ) ∩ Fix(Hj ).
Essentially the same argument as is given for Lemma 3.51 shows that
Fix(H1 , H2 , H3 ) = Fix(H1 ) ∩ Fix(H2 ) ∩ Fix(H3 ).
The argument of the previous paragraph shows the right side is non-
empty, while we have assumed that G = H1 , H2 , H3 . It follows that
Fix(G) = ∅.
Example 3.53. In Chapter 2 we discussed groups generated by
reflections. The groups W236 , W244 and W333 are all groups generated
in the lines formed by certain Euclidean triangles. These three groups
all have Property FA. To see why, consider W333 and let {r, g, b} be the
three generating reflections. Set H1 = r ≈ Z2 , H2 = g ≈ Z2 , and
H3 = b ≈ Z2 . Then each of the three groups Hij is isomorphic to D3 ,
the dihedral group of order 6. (The groups Hij are, in fact, the stabiliz-
ers of the vertices of the fundamental domain for the action of W333 on
R2 .) Since all of these groups are finite, and a generating set of W333 is
contained in H1 ∪ H2 ∪ H3 , it follows from Proposition 3.52 that W333
has Property FA.
The only change that occurs when considering W236 and W244 is that
the Hij can equal D2 , D3 , D4 or D6 , depending on the reflection group
under consideration and which pair of generators is chosen.

For the rest of this section we will explore the group of automorphisms
of a free group Fn , with the goal of establishing that Aut(Fn ), for n ≥ 3,
 3.10 Serre’s Property FA and Infinite Groups 93

has Serre’s property FA. In order not to take up too many pages, our
exposition will be fairly terse. Any reader who finds this discussion in-
teresting is encouraged to fill in the missing details, and to consult the
various references mentioned.
Let S = {x1 , . . . , xn } be a basis for Fn and let S = S∪S −1 be the union
of this basis and its inverses. There are a number of automorphisms of
Fn that come from permutations of S. For example, if σ ∈ Symn is a per-
mutation of {1, . . . , n} then one can define an associated automorphism
σ ∈ Aut(Fn ) induced by requiring

σ(xi ) = xσ(i)

(where we are abusing the use of the symbol “σ”). To verify that this
is indeed an automorphism, notice that σ must be surjective, since the
image of S is S. Further, it must be injective, since freely reduced words
are mapped to freely reduced words, hence no non-identity element is
mapped to the identity. The subgroup of Aut(Fn ) generated by these
permutation automorphisms is isomorphic to Symn .
Similarly one can exchange generators and their inverses. That is, for
each i ∈ {1, . . . , n} there is an automorphism ιi ∈ Aut(Fn ) induced by
requiring

−1
xj if i = j
ιi (xj ) =
xj otherwise.
The subgroup generated by these inversion automorphisms is isomorphic
to (Z2 )n ≈ Z2 ⊕ · · · ⊕ Z2 . The subgroup of Aut(Fn ) generated by the
permutation and inversion automorphisms injects into Sym(S), hence
it must be finite. It is, in fact, a group of order 2n · n!. We denote this
subgroup of Aut(Fn ) by Ωn .
Jakob Nielsen introduced another important collection of automor-
phisms, which we call Nielsen automorphisms. These are the automor-
phisms {ρij } induced by


xi xj if k = i
ρij (xk ) =
xk otherwise.
That this defines a homomorphism from Fn to Fn is an application of
Theorem 3.15. To see that it is a bijection, one needs to establish that the
image of S = {x1 , . . . , xn } under ρij is a basis for Fn . This is essentially
part (a) of Exercise 6.
The following theorem (first proved by Nielsen in 1924 [Ni24]) is not
too difficult, and the interested reader can find a good exposition of it
94 Groups Acting on Trees

in [MKS66]. The details, however, take too many pages for us to include
here.

Theorem 3.54. The elementary automorphisms generate Aut(Fn ).

Having a generating set allows us to begin an argument that estab-

lishes the Triangle Condition (3.52) for Aut(Fn ). Because the Triangle
Condition requires that the group is generated by finite subgroups, and
the generators ρij have infinite order, our initial task will be to transform
our generating set into one that does not use the ρij . The first step is to
notice that if σ is a permutation automorphism then σρij σ −1 = ρσ(i)σ(j) .
From this, one gets:

Corollary 3.55. The automorphism group of a free group, Aut(Fn ), is

generated by Ωn and ρ12 .

In order to remove ρ12 from our generating set, we introduce three

automorphisms of Fn , each of which has order 2. The first is θ, which
takes x1 to x1 x2 and x2 to x−12 , leaving all the other elements in the basis
fixed. This automorphism can be expressed as a product of elementary
automorphisms: θ = ρ12 ι2 . To see that the order of θ is 2, first notice
that θ2 (xi ) = xi for all i ≥ 2, and then compute

θ2 (x1 ) = θ(x1 x2 ) = θ(x1 )θ(x2 ) = x1 x2 · x−1

2 = x1 .

The remaining two automorphisms are particular elements of Ωn . De-

fine τ ∈ Aut(Fn ) by τ (x1 ) = x−1
1 , τ (x2 ) = x3 , τ (x3 ) = x2 and τ (xi ) = xi
for i ≥ 3. (So τ is a product of an inversion and a permutation auto-
morphism.) Finally, let η be defined by η(x1 ) = x−1 −1
2 , η(x2 ) = x1 , and
η(xi ) = xi for i ≥ 3.
Let Symn−2 be the subgroup of Aut(Fn ) generated by permutation
automorphisms where σ(1) = 1 and σ(2) = 2, and let H1 be the sub-
group generated by Symn−2 along with ιn and η. As all of these auto-
morphisms are contained in Ωn , H1 is a subgroup of Ωn , hence H1 is a
finite group. Define H2 to be the cyclic subgroup generated by θ, and
let H3 be the cyclic subgroup generated by τ .

Lemma 3.56. For n ≥ 3, the subgroups {H1 , H2 , H3 } of Aut(Fn ) satisfy

the Triangle Condition.

Proof. First, we have to show that {H1 ∪H2 ∪H3 } is a generating set for
Aut(Fn ). Let σij denote the permutation automorphism that exchanges
xi and xj . Conjugating the automorphism σ3n ∈ H1 by τ ∈ H3 we get
 3.10 Serre’s Property FA and Infinite Groups 95

σ2n , as can be seen by the following computations:

→ x−1 x−1
τ σ3n τ
x1 1 → 1 → x1
τ σ3n τ
x2 → x3 → xn → xn
τ σ3n τ
x3 → x2 → x2 → x3
τ σ3n τ
xn → xn → x3 → x2
(Every other xi is fixed by both τ and σ3n .) Conjugating ιn ∈ H1 by σ2n
results in ι2 ; conjugating σ3n by σ2n creates σ23 . We then also have ι1 ,
as ι1 = σ23 τ , and σ12 , which is ηι1 2 . Thus the subgroup generated by
the Hi contains all the “transposition automorphisms” σij and all the
inversions ιi , hence it contains Ωn . Since θ ∈ H2 , and ι2 ∈ Ωn , it follows
that ρ12 = θι2 ∈ H1 , H2 , H3 . Thus H1 , H2 , H3  contains Ωn and ρ12 .
Thus, by Corollary 3.55, H1 , H2 , H3  = Aut(Fn ).
As we already know each Hi is finite, it remains to verify that each
Hij = Hi , Hj  is finite. This we do, case by case.
The group H13 is generated by elements of Ωn , hence H13 is a subgroup
of Ωn , and is therefore finite.
The group H23 is generated by two elements of order 2, θ and τ . Their
product is the automorphism that takes x1 to x−1 −1
2 x1 , x2 to x3 and x3 to
x−1
2 . Direct computation then shows that θτ has order 4, and therefore
H23 ≈ D4 .
The subgroup generated by Symn−2 and ιn is Ωn−2 , that is, it consists
of all elements of Ωn that fix x1 and x2 . The subgroup H12 is generated
by this Ωn−2 and the elements θ and η. The elements of Ωn−2 commute
with both θ and η so H12 ≈ Ωn−2 ⊕ θ, η. Further, θ and η are both of
order 2, and their product θη has order 3. Thus θ, η ≈ D3 , and H12 is
therefore the finite group Ωn−2 ⊕ D3 .

Lemma 3.56 and Proposition 3.52 imply the following result:

Theorem 3.57. The automorphism group of a free group Fn , for n ≥ 3,

has Serre’s Property FA.

This result was first established by Bogopolski in [Bo87]; the argument

outlined here was given by Bridson in [Br08]. A third, also accessible,
argument appears in a paper by Culler and Vogtmann [CV96].
The group GLn (Z), the group of invertible n × n matrices with integer
entries, is a quotient of Aut(Fn ). To see this, note that the image of a
commutator under an automorphism α ∈ Aut(Fn ) is another commuta-
tor. Thus any automorphism of Fn induces an automorphism of Zn , and
GLn (Z) ≈ Aut(Zn ).
96 Groups Acting on Trees

Lemma 3.58. If G has Property FA, then any quotient of G also has
Property FA.

Proof. Let G/N ≈ H. If H acts on a tree T without fixed point, then

so does G. The action is given by g · T = (gN ) · T .

Corollary 3.59. For n ≥ 3, GLn (Z) has Serre’s Property FA.

J-P. Serre used a variation on the Triangle Condition to establish that

SLn (Z) has Property FA for all n ≥ 3. (A variation of the argument
above also establishes this fact.) This runs somewhat counter to the
intuition given by Proposition 3.7, which states that the matrix group
SL2 (Z) is virtually free. Given the fact that free groups are closely con-
nected to trees, it is not surprising that the group SL2 (Z) does not have
Property FA, and in fact it admits a nice action on T2,3 . Further, the
larger group GL2 (Z) also admits a nice action on T2,3 , hence the restric-
tion that n ≥ 3 cannot be removed from Corollary 3.59 or Theorem 3.57.
These actions are some of the more elementary examples of how a free
product with amalgamation acts on a tree. See Serre’s book ([Se77] or
the English translation [Se80]) for an excellent discussion of this and
related topics.

Exercises
(1) Define the distance between two vertices in a tree to be the number
of edges that occur in the unique reduced path between them.
a. Prove that the number of vertices a distance n from a fixed
vertex v in Tm is m · (m − 1)n−1 .
b. Find a formula for the number of vertices a distance n from
a fixed vertex v in T{m,n} . (You will actually need to find two
formulas, one for the black vertices and one for the white
vertices.)
(2) Prove that Zn is a subgroup of Homeo(R), for all n ∈ N.
(3) In Remark 3.11 it is stated that the subgroup of Homeo(R) gener-
ated by f (x) = x3 and g(x) = x + 1 is a free group.
a. What are f −1 and g −1 ?
b. Show that no freely reduced word in {f, g, f −1 , g −1 }∗ , of
length at most 3, represents the identity.
(4) Prove the Ping Pong Lemma (Lemma 3.10).
 Exercises 97

Fig. 3.15. The hyperbolic isometry whose axis is the horizontal line and the
hyperbolic isometry whose axis is the vertical line generate a free subgroup of
the isometry group of the hyperbolic plane.

(5) (This problem is for people who have studied hyperbolic geome-
try.) Consider the hyperbolic plane with six specified geodesics, as
shown in Figure 3.15. Let x be the hyperbolic isometry of the hy-
perbolic plane whose axis is the horizontal line, and which takes
the geodesic on the left to the geodesic on the right. Let y be the
hyperbolic isometry with axis the vertical line, and which takes the
lower geodesic to the upper geodesic. Show that {x, y} generate a
free group.
(6) Let F2 be the free group with basis {x, y}.

a. Show that {x, xy} is also a basis for F2 .

b. Show that {x, x2 y} is also a basis for F2 .
c. Show that there are infinitely many distinct bases of F2 .

(7) Let g ∈ F2 and express g as a word in the generators and their

inverses: g = a1 a2 · · · an . This word is said to be cyclically reduced
if it is freely reduced and a1 = a−1n . Let g
 be the element produced
by progressively removing such pairs of letters until the resulting
expression is cyclically reduced.

a. Show that g is conjugate to g.

b. Show that g and h are conjugate if and only if g and 
h are
cyclic rearrangements of each other.
98 Groups Acting on Trees

(8) Let F2 be a free group with basis {x, y}. For any natural number
n let βn = {y −i xy i | 1 ≤ i ≤ n}. Show that the subgroup of F2
generated by βn is free of rank n.
(9) Show that F2 has exactly three distinct subgroups of index 2.
(10) Let F2 be the free group with basis {x, y}. Define a map φ : F2

Z2 by φ(x) = (1, 0) and φ(y) = (0, 1). Prove that the kernel of φ is
a free group generated by
X = {[xk , y l ] | k, l ∈ Z and k · l = 0}.
Conclude that this kernel is a free group of infinite rank.
(11) Show that Fn ≈ Fm if n = m.
(12) Show that F2 ≈ Z ∗ Z.
(13) Show that A ∗ B is a free group if and only if A and B are free.
(14) Let F2 be a free group with basis {x, y}, and let K be the kernel
of the homomorphism φ : F2 → Z3 induced by φ(x) = φ(y) = 1.
a. Find a fundamental domain for the action of K on the
Cayley graph of F2 .
b. Find a basis for the kernel K.
(15) Let F2 be a free group of rank 2. Show that, for any n ≥ 2, F2
contains a finite-index subgroup isomorphic to Fn .
(16) Let N be an infinite-index normal subgroup of a free group Fn
(n ≥ 2). Show that N is not finitely generated.
(17) Let G = A ∗ B and let g = a1 b1 · · · an bn with the right-hand side
being a reduced expression. Prove that the order of g is infinite.
(18) Let A and B be non-trivial groups. Prove that the center of A ∗ B
is trivial.
(19) Prove Theorem 3.33. You may want to use the proof of Theo-
rem 3.15 as a model for your argument.
(20) The alternating group A4 is generated by (12)(34) and (123). Let
s generate a cyclic group of order 2, let t generate a cyclic group
of order 3, and let Z2 ∗ Z3 be their free product. Finally, let φ :
Z2 ∗ Z3 → A4 be the map induced by s → (12)(34) and t → (123).
a. Show that K = Ker(φ) is a free group.
b. Find a basis for K.
c. If you are ambitious, replace “3” by “n.”
(21) The centralizer of an element g ∈ G, C(g), is the subgroup of G
consisting of all elements that commute with g. Let g ∈ A ∗ B.
a. Show that if g is in a conjugate of A or B, then C(g) is a
subgroup of the same conjugate of A or B.
 Exercises 99

b. Show that if g is not in a conjugate of A or B, then C(g) is

infinite cyclic.
(22) Let G be the free product of two finite groups, G ≈ A ∗ B. Prove
that if H is a subgroup of G that contains no elements of finite
order (aside from the identity) then H is a free group.
(23) In discussing free products we concentrated exclusively on the free
product of two groups. One can also define free products of three
(or more) groups. Prove that (A ∗ B) ∗ C ≈ A ∗ (B ∗ C), hence we
may write A ∗ B ∗ C to unambiguously denote the free product of
three groups.
(24) Describe the Cayley graph of Z2 ∗ Z3 ∗ Z4 , with respect to a set of
cyclic generators of the factors.
(25) Let {x, y} be a basis for F2 and let ρ ∈ Aut(F2 ) be the Nielsen
automorphism induced by ρ(x) = xy and ρ(y) = y. Describe ρ−1 .
 4
Baumslag–Solitar Groups

The first time I saw the group
a, b | ab = b2 a was in a paper of Graham
Higman’s around 1950, in the Journal of the London Mathematical Society.
Actually the group may have been known to others. Anyway, the reference
to Graham Higman is the reason I often complain about this group being
attributed to Donald and myself.
. –Gilbert Baumslag

Many of the infinite groups we have considered so far can be viewed as

groups acting on the real line. The infinite dihedral group is one example
and in Section 3.1.3 we showed that free groups also act on the real line.
Let us now consider the subgroup of Homeo(R) generated by the linear
functions a and b, where a(x) = 2x and b(x) = x + 1, and where the
binary operation is function composition. The product ab is then the
function
b a
x → x + 1 → 2x + 2

hence ab(x) = 2x + 2. One can also see that ba(x) = 2x + 1 since

a b
x → 2x → 2x + 1 .

Thus ab = ba, so this group is not abelian. However, b2 a(x) = 2x + 2,

so ab = b2 a, and therefore this group is not a free group with basis
{a, b}. This group is referred to as a Baumslag–Solitar group.1 In slightly
greater generality, if a(x) = nx for some integer n ≥ 2, and b is again
b(x) = x + 1, then the group generated by {a, b} is the Baumslag–Solitar
group denoted BS(1, n).
1 This terminology comes from the fact that Gilbert Baumslag and Donald Solitar
wrote a key paper on the groups BS(1, n) in 1962 [BS62]. As the quotation at
the start of this chapter indicates, Baumslag has waged a vigorous, sustained, and
ultimately doomed campaign against referring to BS(1, 2) as a Baumslag–Solitar
group. The paper by Graham Higman is actually from 1951 [Hi51].

100
 Baumslag–Solitar Groups 101

The family of Baumslag–Solitar groups are most commonly defined

in terms of generators and relations (see Section 3.3). It is a fact that
the group BS(1, 2) is the group with presentation a, b | ab = b2 a. In
general, if m and n are positive integers, then the Baumslag–Solitar
group BS(m, n) is defined to be the group given by the presentation
a, b | abm = bn a.

The action of BS(1, 2) on R is different than most actions we have

previously considered. If X is a geometric or topological object, like the
real line, then an action of a finitely generated group G
X is discrete
if every orbit forms a discrete subset of X. That is, for each x ∈ X there
is a small neighborhood Nx  x such that
Orb(x) ∩ Nx = x.
In the case of a group G acting on the real line, one can rephrase this,
to say that G
R is a discrete action if for every x ∈ R there is an

> 0 such that
Orb(x) ∩ (x −
, x +
) = x.
The action D∞
R described in Chapter 2 is an example of a discrete
action on R. The closed unit interval, [0, 1], is a fundamental domain for
D∞
R, hence if x ∈ R then there is some r ∈ [0, 1] and some g ∈ D∞
such that g · r = x. It follows that the orbit of any x ∈ R has the form:
{2n ± r | n ∈ Z and r ∈ [0, 1]}. Thus, if x is an integer, then set
= 1/2;
otherwise set
= min{|n − x|/2 | n ∈ Z}. Then for any x ∈ R one has
Orb(x) ∩ (x −
, x +
) = x, and so this action is discrete.
A discrete action of a finitely generated group on a space X is proper
if, for any x ∈ X, Stab(x) is finite. Again, the action D∞
R is proper,
as, for any x ∈ R,


Z2 if x ∈ Z
Stab(x) ≈
e otherwise.
Unlike the action D∞
R, the action BS(1, 2)
R is neither discrete
nor proper. To see that the action is not proper, notice that the stabilizer
of 0 ∈ R contains the subgroup generated by a, and a ≈ Z. (In fact,
Exercise 3 asks you to prove that Stab(0) = a ≈ Z.)
To show that the action BS(1, 2)
R is not discrete, notice that if n
1
is any positive integer then a−n b(0) = n . Thus, given any
> 0, the
2
orbit of 0 ∈ R under the action of BS(1, 2) intersects (−
,
) in infinitely
many points.
102 Baumslag–Solitar Groups
1
The fact that the orbit of 0 includes is a hint that to understand
2n
BS(1, 2) we should focus on the dyadic rationals. The dyadic rationals
are the rational numbers that can be expressed as an integer over a
power of 2. This set is denoted Z[1/2], and just to be explicit we repeat:
m 
Z[1/2] = | m ∈ Z and n ∈ N .
2n
m
If m is any integer, then an bm (0) = n , which is to say that the orbit of
2
0 under the action of BS(1, 2) contains Z[1/2]. But each generator and
its inverse maps Z[1/2] to Z[1/2]:
m m
a n = n−1
,
2
m 2
m
a−1 n = n+1
,
2 2
m m m + 2n
b n = + 1 = ,
2 2n 2n
m m m−2 n
b−1 n = n
−1= .
2 2 2n

Thus the orbit of 0, or in fact any element of Z[1/2], is Z[1/2]. This fact
also follows from the following description of this group.

Proposition 4.1. The elements of BS(1, 2) are the linear functions

m
g : R → R that can be expressed in the form g(x) = 2n · x + k where
2
n, m and k are integers.

Outline of the proof. We ask the reader to verify the following steps as
a useful exercise.

1. Since BS(1, 2) is generated by linear functions, and the composi-

tion of linear functions is linear, every g ∈ BS(1, 2) has the form
g(x) = αx + β.
2. Since BS(1, 2) is defined as the group generated by a and b, every
element can be expressed as a product of a’s, b’s, a−1 ’s and b−1 ’s.
Arguing by induction on the length of these products one sees that
if g(x) = αx + β then α = 2n for some integer n and β ∈ Z[1/2].
3. The translation x
→ x+ 2mk can be expressed using our generators
a and b as a−k bm ak .
4. It follows that if
m
g(x) = 2n · x + k
2
 Baumslag–Solitar Groups 103

then
 
g(x) = a−k bm ak an .

Hence g(x) = a−k bm ak+n and therefore BS(1, 2) is composed

entirely of linear functions of the prescribed form.

By its construction we know that BS(1, 2) acts on the real line. But
this action is neither discrete nor proper, and the tools developed in pre-
vious chapters seem to rely on, or at least prefer, discrete and proper ac-
tions. For example, the Drawing Trick (Remark 1.49) provides a method
for constructing a Cayley graph of a group G, by using the orbit of a
point which is moved freely by G as the vertices of the Cayley graph.
This method worked nicely in Chapter 2, where we used it to draw Cay-
ley graphs of some groups generated by reflections, but it does not work
well for the action BS(1, 2)
R. Because of this, we delay discussing
the Cayley graph of BS(1, 2) with respect to {a, b} until Section 5.4.
Instead, we introduce an alternative set on which BS(1, 2) acts.

Corollary 4.2. The orbit of the closed interval [0, 1] under the action
BS(1, 2)
R is the set of closed intervals

I = {[x, x + 2i ] | x ∈ Z[1/2] and i ∈ Z}.

Further, if g([0, 1]) = [0, 1] then g = e ∈ BS(1, 2). Thus g ∼ g([0, 1]) is a
bijection between I and the elements of BS(1, 2).

Proof. Let [x, y] ∈ I. If g ∈ BS(1, 2), then by Proposition 4.1 we may

express g as g(x) = 2n · x + 2mk . Thus
m n m
g([x, x + 2i ]) = [2n · x + k
, 2 · x + k + 2n+i ].
2 2
The left endpoint, 2n · x + 2mk , is in Z[1/2] and the width of the interval
is 2n+i , hence the image of an interval in I is an interval in I.
The element a−k bm ak+i is the linear function g(x) = 2i · x + 2mk , which
takes the interval [0, 1] to the interval [ 2mk , 2mk + 2i ], proving that the
action of BS(1, 2) on I is transitive.
Finally, any linear function is determined by the image to two points.
If g([0, 1]) = [0, 1] then either g is the identity, or g(0) = 1 and g(1) = 0.
But this is not possible, since the coefficient of x in g(x) = 2n · x + 2mk is
positive.
104 Baumslag–Solitar Groups

The group BS(1, 2) will show up from time to time in the remainder
of this text. In order to give an example of the importance of Baumslag–
Solitar groups within geometric group theory, we state a result that we
will not prove:
Theorem 4.3 (See [Me72]). There is a surjective group homomorphism
φ : BS(2, 3) → BS(2, 3) (where φ(a) = a and φ(b) = b2 ) that is not an
isomorphism.

Exercises
(1) Fill in the outline of the proof of Proposition 4.1.
(2) Show that ab = bn a in BS(1, n).
(3) Show that Stab(0) = a for BS(1, 2)
R.
(4) Let x ∈ Z[1/2] be any dyadic rational. What is Stab(x) under the
action BS(1, 2)
R?
(5) State and prove the analog of Proposition 4.1 for BS(1, n).
(6) Prove that the cyclic subgroup of BS(1, 2) generated by the element
b is not a normal subgroup.
(7) Prove that the smallest normal subgroup of BS(1, 2) that contains
b is isomorphic to the dyadic rationals.
(8) Prove that
 BS(1, 2) is isomorphic
 to the group of matrices gener-
2 0 1 1
ated by and .
0 1 0 1
 5
Words and Dehn’s Word Problem

Irgend ein Element der Gruppe ist durch seine Zusammensetzung aus den
Erzeugenden gegeben. Man soll eine Methode angeben, um mit einer endlichen
Anzahl von Schritten zu entscheiden, ob dies Element der Identität gleich ist
oder nicht.
. –Max Dehn

5.1 Normal Forms

In this chapter we develop connections between the study of infinite
groups and the study of formal languages, a theme that was hinted at in
Section 1.9. We begin by establishing some notation and terminology.
Let S ⊂ G be a generating set for a group G, let “S” be a copy of
S, thought of as just a set of letters, and let S −1 be a set of formal
inverses for the elements of S. There is a function from the free monoid
{S ∪ S −1 }∗ to G:
π : {S ∪ S −1 }∗ → G.
This evaluation map takes each s ∈ S to the corresponding generator
of G, it takes each s−1 to the inverse of the corresponding generator of
G, and it takes a word ω ∈ {S ∪ S −1 }∗ to the corresponding product
of generators and their inverses in G. Since S is a generating set for
G, the map π is onto. We will often want to be clear about whether ω
stands for a string of letters or an element in G. Unless it is indicated
otherwise, we let ω be the string of letters and π(ω) the element of G.
So, for example, if we let a = (1, 0) ∈ Z ⊕ Z and b = (0, 1) ∈ Z ⊕ Z,
then the words ω1 = aba−1 bbba and ω2 = bbabb are distinct elements of
{a, b, a−1 , b−1 }∗ , but π(ω1 ) = π(ω2 ) = (1, 4) ∈ Z ⊕ Z.
The fact that elements in a group G have different “spellings” in terms
of the generators is sometimes a cause of confusion. Generally one prefers

105
106 Words and Dehn’s Word Problem

to have a prescribed method of expressing any given g ∈ G as a product

of the generators and their inverses. A normal form is precisely that.
That is, a normal form is a function

η : G → {S ∪ S −1 }∗

such that the composition π ◦ η : G → G is the identity. In other words,

it is a section of the evaluation map π. Often one thinks of a normal form
in terms of the image of G in the free monoid {S ∪S −1 }∗ . In this setting,
a normal form is simply a subset of {S ∪ S −1 }∗ that maps bijectively to
G under the evaluation map π.

Example 5.1. The set of freely reduced words in {S ∪S −1 }∗ is a normal

form for the free group with basis S.

Example 5.2. Consider the free abelian group Z ⊕ Z with generators

{a, b} as above. Since a and b commute, in any product of a’s, b’s, and
their inverses, one can move the a’s and a−1 ’s to the left and the b’s and
b−1 ’s to the right. This shows that

NF1 = {ai bj | i, j ∈ Z}

is a normal form for Z ⊕ Z. There is certainly no claim being made

that this is the only normal form. For example, one could push a’s and
a−1 ’s to the right (and b’s and b−1 ’s to the left) and get the resulting
normal form: NF2 = {bi aj | i, j ∈ Z}. As yet another example, one could
have a preference for alternating a’s and b’s. Thus instead of writing
a3 b4 one might prefer abababb = (ab)3 b. This leads to the normal form
NF3 = {(ab)i bj | i, j ∈ Z}, where for n ≥ 0 we let (ab)n be shorthand
−n
for the word
ab · ab
· · · ab and similarly we let (ab) (n > 0) denote the
n copies of ab
word (b−1 a−1 )n .
 ∗
Because of the correspondence between words in S ∪ S −1 and
paths in the corresponding
 ∗ Cayley graph (see Section 1.9), a normal
form η : G → S ∪ S −1 can also be thought of as a collection of paths
in the Cayley graph ΓG,S . If g ∈ G then the word η(g) has an associ-
ated edge path in ΓG,S that begins at the identity and traverses edges
as prescribed by η(g). For example, consider the normal forms given in
Example 5.2. If we think of the generator a ∈ Z ⊕ Z as corresponding
to the horizontal edges in Figure 5.1, and b ∈ Z ⊕ Z as corresponding to
the vertical edges, then the paths corresponding to the elements of NF1
move horizontally, then vertically. Similarly the paths corresponding to
 5.1 Normal Forms 107

Fig. 5.1. Three paths corresponding to the three normal forms NF1 , NF2
and NF3 . Each path joins the vertex associated to the identity to the vertex
associated to a3 b2 = b2 a3 = (ab)3 b−1 ∈ Z ⊕ Z.

words in NF2 move vertically then horizontally. Finally the paths corre-
sponding to the words in NF3 move along the diagonal, and then they
either move directly up or down.
Remark 5.3. When thought of as a collection of paths in a Cayley
graph, a normal form is occasionally referred to as a “combing,” al-
though that word is often restricted to sets of normal forms where the
associated paths satisfy certain geometric and/or language theoretic re-
strictions. The terminology is appropriate, at least for “nice” normal
forms. Consider for example G = Z ⊕ Z4 with generators x = (1, 0) and
y = (0, 1). There is a normal form
N (G) = {xi y j | i ∈ Z and 0 ≤ j ≤ 3} .
If you trace out the paths associated to this normal form, you get some-
thing that looks “combed” (Figure 5.2).

Fig. 5.2. A combing of the Cayley graph of Z ⊕ Z4 . Paths corresponding to

the combing are contained in the spanning tree indicated by darker edges.

For a survey of “hairdressing in groups,” see [Re98].

In the previous chapter we introduced the Baumslag–Solitar group
BS(1, 2). In proving various results about the action of this group on R,
we developed and then relied on particular expressions for the elements
108 Words and Dehn’s Word Problem

of BS(1, 2); we showed that every element of BS(1, 2) can be expressed

as a word of the form:
W = {a−k bm ak+n | k, m, n ∈ Z}.
This set of words is not, however, a normal form. The words aba−1
(k = −1, m = 1 and n = 0) and b2 (k = n = 0 and m = 2) are both
in W. But aba−1 = b2 ∈BS(1, 2). Thus the evaluation map restricted
∗
to W, π : a, b, a−1 , b−1 → BS(1, 2), is onto BS(1, 2), but it is not
one-to-one. However, we are not far from a normal form.
Proposition 5.4. The set
NF = {a−k b2m+1 ak+n | k, m, n ∈ Z} ∪ {an | n ∈ Z}
is a normal form for BS(1, 2).
Proof. We know that every g ∈ BS(1, 2) can be expressed as a linear
function of the form g(x) = 2n x + 2m̂k . By reducing to lowest terms we
may assume that m̂ is either odd or zero. Hence every g ∈ BS(1, 2) has
the form g(x) = 2n x + 2m+1
2k
or g(x) = 2n x. The words
{a−k b2m+1 ak+n | k, m, n ∈ Z}
describe the functions g(x) = 2n x + 2m+1
2k
and the words {an | n ∈ Z}
give g(x) = 2n x, so our set of words maps onto BS(1, 2) under π.
To establish uniqueness, assume to the contrary that
a−k b2m+1 ak+n = a−K b2M +1 aK+N
where at least one of the following inequalities holds: k = K, m = M or
n = N . The function on the left-hand side takes
 
m 2m + 1
[0, 1]
→ k , + 2n
2 2k
while the function on the right-hand side takes
 
M 2M + 1
[0, 1]
→ K , +2 .N
2 2K
It follows that
2m + 1 2M + 1
= and 2n = 2N .
2k 2K
Thus m = M , k = K and n = N , which is a contradiction. The same
sort of argument shows that no expression of the form a−k b2m+1 ak+n
yields a function of the form g(x) = 2n x, hence NF is a normal form for
BS(1, 2).
 5.2 Dehn’s Word Problem 109

5.2 Dehn’s Word Problem

There are a number of questions in the study of groups that can be

thought of as “linguistic” issues. The most famous example of this is
Dehn’s word problem, published in 1912 (see [De12]). Max Dehn
asked: Given a group G and a finite generating set S, is it possible to
decide which words in {S ∪S −1 }∗ represent the identity? In other words,
given an arbitrary word ω ∈ {S ∪ S −1 }∗ is there a way to decide if
π(ω) = e ∈ G?
Sometimes it is easy to solve Dehn’s word problem. If we generate
Z ⊕ Z by a = (1, 0) and b = (0, 1), then the words in {a, b, a−1 , b−1 }
that represent the identity are precisely the words where the sum of the
a-exponents and the sum of the b-exponents are both zero.
A slightly more complicated case would be the infinite dihedral group
D∞ generated by the reflections a and b (see Chapter 2). Since a and b
are their own inverses we do not need to add a set of formal inverses to
our generating set; we can simply consider words in {a, b}∗ . Given such
a word we can repeatedly apply the following rule:

D∞ -rule: If ω ∈ {a, b}∗ and ω contains a subword of the form aa or bb, then
remove this subword to make a new word ω  .

This new word represents the same element of D∞ as the word we started
with, so determining if π(ω  ) = e is sufficient for determining if the
original word represents the identity. There are, potentially, a number
of choices that could be made in applying this rule, as is illustrated in
Figure 5.2. Any word that results from such a process is equivalent to
w – in the sense that its image under the evaluation map is π(w) – and
the final word one gets is either empty or an alternating sequence of a’s
and b’s. But Proposition 2.1 shows that all alternating sequences of a’s
and b’s yield non-identity elements of D∞ . Thus the word w represents
the identity in D∞ if and only if w reduces to the empty word by repeated
application of the D∞ -rule.
There is a variant of Dehn’s word problem that asks not only if one
can determine when a word represents the identity, but: given a group
G, a finite generating set S, and two words ω, ω  ∈ {S ∪ S − }∗ , can one
determine if ω and ω  represent the same element of G? In other words,
can you determine if π(ω) = π(ω  )? This is referred to as the “equality
problem.”
110 Words and Dehn’s Word Problem
baaaababb

baababb baaaaba

bbabb baaba

abb bba

Fig. 5.3. There are a variety of choices that can be made in showing that
baaaababb = a ∈ D∞ . Thicker arrows with solid arrowheads indicate when
multiple choices result in the same word.

Proposition 5.5. Given a group G and a finite generating set S, the

equality problem can be solved if and only if Dehn’s word problem can be
solved.

Proof. Since the identity is represented by the empty word, if one can
solve the equality problem then one can determine when a given word ω
gives the same element as the empty word, hence one can solve Dehn’s
word problem.
Conversely, ω and ω  represent the same element in G if and only if
ω · ω −1 represents the identity element. Thus if one can solve Dehn’s


word problem, one can also solve the equality problem.

Dehn’s original motivation for formulating and studying the word

problem came from issues in algebraic topology. It was eventually discov-
ered that there are finitely generated (and, better yet, finitely presented)
groups where the word problem is unsolvable. This was established inde-
pendently by Novikov and Boone in the mid-1950s. In fact, these tech-
niques establish a quite striking result:
 5.3 The Word Problem and Cayley Graphs 111

Theorem 5.6 ([BBN59]). There are finitely presented groups G such

that there is no algorithm that can determine whether a given word in
the generators represents:

1. the identity;
2. an element in the center Z(G);
3. an element in the commutator subgroup;
4. an element of finite order.

We will not construct examples of such groups in this book, but they
are of great importance in a variety of fields and they have practical
applications to areas such as coding theory. For more information, see
the survey article [Mi92].

5.3 The Word Problem and Cayley Graphs

We commented in Section 1.9 that there is a one-to-one correspondence
between words in the generating set of a group and edge paths in the
Cayley graph based at the vertex associated to the identity. Given this
correspondence there are a number of ways to give a geometric interpre-
tation to questions that seem quite algebraic. Dehn’s word problem, for
example, can be translated as in the Proposition below.

Proposition 5.7. Given a group G and a finite generating set S, one

can solve Dehn’s word problem if and only if one can decide which words
ω correspond to closed paths pω in the Cayley graph ΓG,S .

Example 5.8. The word problem for Dn (with respect to the generating
set {f, ρ} where f is a reflection and ρ a rotation through an angle of
2π/n) can be solved. To determine if a word ω represents the identity,
simply trace out the corresponding edge path pω in the Cayley graph Γ.
The word ω represents the identity if and only if the end of the path pω
is at the vertex in Γ that corresponds to the identity.

This example indicates that the word problem can be solved in any
group where the Cayley graph can be constructed. Since our primary
focus is on infinite groups, we need an appropriate notion of what it
means to say a Cayley graph “can be constructed” if we hope to be able
to convert this example to a theorem. A reasonable answer is that one
can build arbitrarily large pieces of the Cayley graph, where the idea of
“arbitrarily large pieces” is codified by the notion of balls in a Cayley
graph.
112 Words and Dehn’s Word Problem

Definition 5.9. Let Γ be a connected graph. The distance from a vertex

v to a vertex w is the minimum number of edges in a combinatorial path
joining v to w. For example, if v and w are distinct, but form Ends(e)
for some edge e, then the distance from v to w is 1. We denote this
distance function by dΓ (v, w).
For a fixed vertex v ∈ V(Γ) and a fixed n ∈ N, the sphere of radius n
centered at v is the set of vertices

S(v, n) = {w ∈ V(Γ) | dΓ (v, w) = n}.

The ball of radius n centered at the vertex v is the subgraph formed as

the union of all paths in Γ of length ≤ n that start at the vertex v. We
denote this subgraph by B(v, n). The vertices of B(v, n) are the union
of the vertices in S(v, i) for i ≤ n.

Fig. 5.4. The sphere of radius 3 in the Cayley graph of the reflection group
W333 (see Chapter 2) consists of all the vertices with a dark halo; the ball of
radius 3 is indicated by thicker edges.

In the case of a Cayley graph Γ of a group G we abuse notation

and write S(g, n) and B(g, n) to denote the sphere and ball of radius n
centered at the vertex associated to g ∈ G. (An example of a ball and
sphere in a Cayley graph is shown in Figure 5.4.)
 5.3 The Word Problem and Cayley Graphs 113

A Cayley graph Γ is constructible if given n ∈ N one can construct

B(e, n) in a finite amount of time.

Fig. 5.5. The ball of radius 2 in the Cayley graph of Z ⊕ Z consists of all the
dark vertices and solid arrows.

Theorem 5.10. Let G be a group with finite generating set S. Then

the word problem is solvable if and only if the Cayley graph ΓG,S is
constructible.

Proof. (⇐) Let ω ∈ {S ∪ S −1 }∗ and let n = |ω|. Construct the subgraph

B(e, n) ⊂ ΓG,S , and trace out the path pω inside this ball. The terminal
vertex of this path is ve if and only if π(ω) = e ∈ G. Thus one can solve
the word problem.
(⇒) The proof that B(e, n) is constructible is by induction on n. The
base case is immediate: B(e, 0) consists of a single vertex corresponding
to the identity.
Assume that B(e, n) has been constructed, so that our goal is to con-
struct B(e, n + 1). Every vertex in ΓG,S is incident with exactly k edges
corresponding to the generators S ∪ S −1 . If v ∈ B(e, n − 1), then v is
already incident with k edges labelled by S ∪ S −1 . If dΓ (ve , v) = n, let
Labelsn (v) be the subset of S ∪ S −1 corresponding to edges that are
114 Words and Dehn’s Word Problem

attached to v in B(e, n). In forming B(e, n + 1)  we need  to add an (ori-

−1
ented) edge to v, labelled s, for each s ∈ S ∪ S \ Labelsn (v).
The trick is to discover what the other vertex is that bounds this edge.
Notice that none of the missing edges can join vi to a vertex in B(e, n−1).
If one did, then the edge in question would have been present in B(e, n),
and so its label would be in Labelsn (vi ).
For convenience, list the elements in the sphere of radius n in some
fixed order: S(e, n) = {v1 , v2 , . . . , vl }. To each vi associate a word ωi
whose corresponding path in B(e, n) joins the vertex associated to e to
vi . Starting with vertex v1 , list the generators in
 
S ∪ S −1 \ Labelsn (v1 ) = {s1 , s2 , . . . , sm }.
The edge of ΓG,S that is attached to v1 and is labelled si cannot join
v1 to a vertex in B(e, n − 1), so it must either join v1 to some other
vj ∈ S(e, n), or join v1 to a vertex in S(e, n + 1). This edge joins v1 to
vj ∈ S(e, n) if and only if the word ω1
si (ωj )−1 describes a circuit in ΓG,S ,
that is, if and only if π ω1 si (ωj )−1 = e ∈ G. Since we can
solve the
word problem, we can check this condition. If π ω1 si (ωj )−1 = e ∈ G
for each j, then the edge labelled si joins v1 to a new vertex, not seen
in B(e, n).
Let i > 1 and
 let sj be a missing label for the vertex vi , that is,
−1
sj ∈ S ∪ S \ Labelsn (vi ). The edge labelled sj , which is attached
to vi in ΓG,S , joins vi to one of three types of vertices:
1. another vk ∈ S(e, n);
2. a vertex in S(e, n+1) that is joined to vm ∈ S(e, n), where m < i;
or
3. a vertex in S(e, n + 1) that is not attached to any of the previous
vertices {v1 , . . . , vi−1 }.
(The second
case does −1

not occur when i = 1.) Case 1 occurs if and
only if π ωi sj (ωm ) = e ∈

G, as before. Similarly, case 2 occurs
 if
and only if π ωi sj (ωm sn )−1 = e ∈ G for some sn ∈ S ∪ S −1 , and
this can be checked since we can solve the word problem. If none of the
appropriate words represent the identity, then we know we are in Case
3, and a new vertex needs to be added in forming B(e, n + 1). As any
vertex in B(e, n + 1) is joined to a vertex in B(e, n), this process will
terminate in a finite number of steps.

Example 5.11. In Section 3.5 we explored the free product G = Z3 ∗Z4 .

This group is generated by an element a of order 4 (corresponding to
 5.4 The Cayley Graph of BS(1, 2) 115

Fig. 5.6. Part of the Cayley graph of Z3 ∗ Z4 with respect to the generating
set {a, b}.

the Z4 ) and an element b of order 3 (corresponding to the Z3 ). We can

construct a normal form for G by remembering that the elements of G
can be expressed as words that alternate between non-identity elements
of Z3 and Z4 . These non-identity elements can be uniquely expressed
by the words {a, a2 , a3 , b, b2 }. Knowing this, and applying the method
described in the proof of Theorem 5.10, one can verify that the Cayley
graph of G with respect to the generating set {a, b} looks like a “tree of
squares and triangles” as in Figure 5.6.

Remark 5.12. In discussing the distance between vertices in a Cayley

graph, and the notion of spheres and balls of a given radius, we have
touched on the geometric side of group theory. This topic is explored in
Chapter 9.

5.4 The Cayley Graph of BS(1,2)

It is easy to solve the word problem for BS(1, 2). Given any word ω ∈
{a, b, a−1 , b−1 }∗ one can produce the associated group element by
116 Words and Dehn’s Word Problem

composing the appropriate linear functions. In the end one has g(x) = x
if and only if π(ω) = e. For example, consider the word
ω = a2 ba−2 ba2 b−1 a−2 b−1 ,
and let g = π(ω). Then

b−1 a−2 1 1 b−1 1 5 a2

g : x −→ x − 1 −→ x − −→ x − −→ x − 5
4 4 4 4
b a−2 1 b 1 a2
−→ x − 4 −→ x − 1 −→ x −→ x
4 4
so g = π(ω) = e ∈ BS(1, 2).
Since we can solve the word problem, Theorem 5.10 implies that we
should be able to build the Cayley graph of BS(1, 2) relative to the
generators {a, b}. Because ab = b2 a, or equivalently aba−1 = b2 , the
Cayley graph must contain a 5-cycle corresponding to this relation. This
5-cycle is informally referred to as a brick in the Cayley graph. We can

Fig. 5.7. The “brick” corresponding to ab = b2 a, where horizontal arrows

correspond to the generator b, and vertical arrows to a.

begin to construct the Cayley graph by stacking these bricks together.

The resulting grid of bricks can be continued indefinitely, forming what
is referred to as a sheet in the Cayley graph of BS(1, 2), as is shown in
Figure 5.8.
The sheet shown in Figure 5.8 is not the complete Cayley graph of
BS(1, 2) with respect to {a, b}. One can see that there must be some
edges missing, as some of the vertices in this grid have only three edges
incident with them, not four. For example, if you assume the identity is
the vertex in the lower left corner of Figure 5.8, you cannot trace out the
final a in the word aba. Because we know aba is the unique representation
of an element of BS(1, 2) (in our normal form), and because every vertex
in this Cayley graph must be incident to four edges, there must be a
missing vertex and edge. One standard way to depict this is to have the
missing edges come out from the sheet, as in Figure 5.9.
 5.4 The Cayley Graph of BS(1, 2) 117

Fig. 5.8. A “sheet” in the Cayley graph of BS(1, 2).

Fig. 5.9. More of the Cayley graph of BS(1, 2).

Given any vertex in the Cayley graph, one can form a combinato-
rial line by right multiplying by positive and negative powers of b. Call
such a line a b-line in the Cayley graph of BS(1, 2). Each such line
will have one set of edges labelled a coming into it and two sets of
edges labelled a pointing away from it, corresponding to whether the
exponent of b is odd or even. Figure 5.9 illustrates the fact that two
distinct sheets come together at each b-line. These sheets merge below
the b-line (the portion corresponding to a-edges which point into the b-
line) but are distinct above the b-line. This is why, informally speaking,
a side view of the Cayley graph of BS(1, 2) looks like a trivalent tree
(Figure 5.10).
At this stage the reader is well-poised to think about our characteriza-
tion of BS(1, 2) given in Corollary 4.2. There it is shown that there is a
one-to-one correspondence between elements of BS(1, 2) and the images
of the unit interval [0, 1] under the action of BS(1, 2). Let φ : Γ → R be
the map induced by sending the vertex associated to bn to n ∈ R and
by requiring that the image of any edge labelled a is a single point, as
indicated in Figure 5.11. We leave the proof of the following insight to
the exercises.
118 Words and Dehn’s Word Problem

Fig. 5.10. The Cayley graph of BS(1, 2) projects onto a trivalent tree.

Proposition 5.13. Let g be an element of BS(1, 2). Let eg be the edge

in the Cayley graph Γ that is labelled b and has the vertex associated to
g as its initial vertex. Then φ(eg ) = g([0, 1]).

–2 –1 0 1 2

Fig. 5.11. Projecting the Cayley graph of BS(1, 2) onto R.

 Exercises 119

Exercises
(1) Three different normal forms for Z ⊕ Z are given in Example 5.2.
Draw the Cayley graph of Z ⊕ Z and the paths corresponding to
the words in NF1 , NF2 and NF3 for the elements:
a. (3, 5)
b. (−2, −1)
c. (1, 6)
(2) Find a normal form for F2 ⊕ F2 .
(3) Use colored pencils to draw spheres of radius n, for n = 1, 2 and 3,
in the Cayley graph of:
a. the cyclic group Z6 (with respect to a cyclic generator).
b. the symmetric group Sym4 (with respect to the adjacent
transpositions {(12), (23), (34)}).
c. the reflection group W333 from Chapter 2 (with respect to
the three reflections in the sides of the equilateral triangle).
d. BS(1, 2) (with respect to {a, b}).
(4) Prove that if you can solve the word problem for groups A and B
then you can solve it for G = A ⊕ B.
(5) Prove that if you can solve the word problem for groups A and B
then you can solve it for G = A ∗ B.
(6) Solve the word problem for the reflection group W333 from Chap-
ter 2.
(7) Show that π(a3 b2 a−3 ba3 b−2 a−3 b−1 ) = e ∈ BS(1, 2).
(8) Prove Proposition 5.13.
(9) Construct the Cayley graph of BS(1, 3).
(10) In addition to the word problem, Max Dehn also introduced the
following conjugacy problem in his 1912 paper.
Let S be a finite generating set for a group G. Let ω and ω  be words in
{S ∪ S −1 }∗ . Can one determine if π(ω) and π(ω  ) are conjugate?
Solve the conjugacy problem in the infinite dihedral group D∞ .
(11) Show that you can solve the conjugacy problem (see Exercise 10)
in A ∗ B if you can solve it in A and B.
 6
A Finitely Generated, Infinite Torsion
Group

In 1902, William Burnside asked the following natural question:

Does every finitely generated infinite group contain an element of infinite

order?

Today this is known as the Weak Burnside Problem and, while it appears
to be an elementary question, it took over sixty years to resolve. On the
face of it, one would probably believe the answer is “yes.” The infinite
groups we have encountered so far all have elements of infinite order.
This is true for Zn , Fn , A ∗ B where A and B are non-trivial groups, and
BS(1, 2). However, in this chapter we construct an example that shows
that the answer is actually “no.”

A group is torsion-free if no non-identity element has finite order;

examples include the free groups and BS(1, 2). A group is a torsion group
if every element has finite order; the only examples you are likely to have
encountered are finite groups. These two conditions are at opposite ends
of a spectrum. There are a number of groups that are neither torsion-free
nor torsion, such as Z2 ∗ Z3 . In this terminology, Burnside was asking if
there are finitely generated, infinite torsion groups.
There are easy ways to construct infinite torsion groups: one can form
a countable direct sum of finite cyclic groups, such as

Z2 ⊕ Z2 ⊕ Z2 ⊕ · · ·

and the factor group Q/Z is also an infinite torsion group. But in nei-
ther of these examples is the group finitely generated. (Establishing these
claims is your Exercise 1.) To construct our example of a finitely gener-
ated, infinite torsion group we use a variation on the theme of Chapter 3

120
 A Finitely Generated, Infinite Torsion Group 121

(groups acting on trees), and describe our group via symmetries of a

rooted tree.

Definition 6.1. A rooted tree consists of a tree T and a selected vertex

v, called the root. In contrast to common experience, rooted trees are
generally drawn with the root at the top, with edges and other vertices
descending downward (as in Figure 6.1). Using this convention one can
define the children of a vertex v ∈ T to be those vertices that are
immediately below v. They can also be defined as the vertices that are
adjacent to v which are not on the unique, reduced edge path joining
v to the root. Similarly, the descendants of v are those vertices that
are joined to v by a descending path, or more formally, a vertex w is
a descendant of v if the only point of intersection between the unique
reduced edge path joining w to v and the unique reduced edge path
joining v to the root, is the vertex v.
A rooted tree where each vertex which is not a leaf has n children is
referred to as a rooted n-ary tree.
The group Sym(T ), for a rooted tree T , consists of all symmetries of
T that fix the root.

0 1 2

00 01 02 10 11 12 20 21 22

Fig. 6.1. The rooted, infinite ternary tree T .

We let T denote the rooted ternary tree for the remainder of this chap-
ter. We label the vertices of this tree using finite sequences n1 n2 · · · nk
where each ni ∈ Z3 . The root corresponds to the empty sequence; the
vertex associated to n1 · · · nk nk+1 is a child of the vertex associated to
n1 · · · nk ; and the vertex associated to n1 · · · nk nk+1 · · · nl is a descendant
122 A Finitely Generated, Infinite Torsion Group

of the vertex associated to n1 · · · nk . Denote the vertex associated to the

finite sequence n1 · · · nk by vn1 ···nk , and define the vertices at level k in
T to be those vertices whose labels are sequences of length k. As an
example, {v0 , v1 , v2 } are the vertices at level 1. The first two levels of
T , with the vertices labelled, are illustrated in Figure 6.1. Notice that,
because the root is fixed, any symmetry of the rooted ternary tree T
must preserve the levels.
Let σ ∈ Sym(T ) be the symmetry of T described by
σ(vn1 n2 ···nk ) = v(n1 +1)n2 ···nk
where the addition in the first digit of the subscript is done in Z3 . Thus
σ(v0 ) = v1 , σ(v1 ) = v2 and σ(v2 ) = v0 . Notice that, following the
drawing conventions of Figure 6.1, the subtree hanging below v0 is taken
by σ to the subtree hanging below v1 without any twisting or reflecting.
One can think of σ is something like a shuffle of a deck of cards.
Any vertex vn1 ···nk ∈ T can serve as the root of the subtree of T
induced by vn1 ···nk and its descendants. Denote this subtree as Tn1 ···nk .
One can perform a version of the shuffle σ on Tn1 ···nk . This symmetry
will fix any vertex whose label does not begin with the sequence n1 · · · nk ,
and it will take
vn1 ···nk nk+1 nk+2 ···nl → vn1 ···nk (nk+1 +1)nk+2 ···nl .
Denote this symmetry by σn1 ···nk .
We can now define a more complicated symmetry of T as an infinite
product of shuffles:
1 −1 1
ω = σ01 σ1−1 σ20 −1
σ21 σ220 σ221 ···
The symmetry σ0 shuffles the subtrees below v0 ; the symmetry σ1−1
shuffles the subtrees below v1 (in the opposite direction). The vertices
v2 and its children v20 , v21 , and v22 are all fixed by ω. However, the sub-
trees below v20 are shuffled, as are the subtrees below v21 (in the opposite
direction). In general, ω will fix any vertex whose label is 222 · · · 2i, and
it will shuffle the subtrees below v222···20 and v222···21 (in the opposite
direction). We refer to this symmetry as a ωaterfall (Figure 6.2).
Theorem 6.2. Let U be the subgroup of Sym(T ) generated by σ and
ω. Then U is a finitely generated, infinite group where every g ∈ G has
finite order. Further, the order of any element is a power of 3.
Before turning to the proof of this result, we set up a bit of useful
notation. Just as σn1 ···nk denotes a shuffle based at the vertex vn1 ···nk ,
 A Finitely Generated, Infinite Torsion Group 123

Fig. 6.2. The “ωaterfall” symmetry ω.

let ωn1 ···nk denote a waterfall based at vn1 ···nk . More formally, ωn1 ···nk
can be written as an infinite product of shuffles, where one adds n1 · · · nk
as a prefix to all the indices used in describing ω. For example, ω21
would be
1 −1 1 −1 1 −1
ω21 = σ210 σ211 σ2120 σ2121 σ21220 σ21221 ···

Definition 6.3. There are a number of subgroups of Sym(T ) that

are isomorphic to U. Namely, if n1 n2 · · · nk is a label, let Un1 n2 ···nk be
the subgroup generated by σn1 n2 ···nk and ωn1 n2 ···nk . This subgroup of
Sym(T ) acts on the subtree Tn1 n2 ···nk exactly like U acts on T .

Using this notation we notice the following formulas, which the reader
can verify:

ω = σ0 σ1−1 ω2 ,
σωσ −1 = ω0 σ1 σ2−1 , (6.1)
2
σ ωσ −2
= σ0−1 ω1 σ2 .

The following is another set of formulas that are useful in the proofs
of Lemmas 6.4 and 6.5:

ωσ = σ(σ 2 ωσ −2 ),
(σωσ −1 )σ = σω, (6.2)
2 −2 −1
(σ ωσ )σ = σ(σωσ ).
124 A Finitely Generated, Infinite Torsion Group

Lemma 6.4. Each g ∈ U can be expressed as g = σ k x1 · · · xn where

each xi ∈ {ω, σωσ −1 , σ 2 ωσ −2 }.

Proof. Since U is the group generated by σ and ω, every g ∈ U can be

expressed as a word in {σ, σ −1 , ω, ω −1 }∗ . Further, since σ 2 = σ −1 and
ω 2 = ω −1 , every g ∈ U can be expressed as a word in {σ, ω}∗ . Our goal
is to convert such an expression into an expression of the form indicated
above.
Let g = σ j y1 y2 · · · yn where each yi ∈ {σ, ω} and y1 = ω. Let yk be the
leftmost σ in the suffix y1 y2 · · · yn . Then yk−1 = ω. By the first equation
given in Equations 6.2,

yk−1 yk = ωσ = σ(σ 2 ωσ −2 ).

Thus this σ can be moved to the left, converting ω’s to (σ 2 ωσ −2 )’s as

it goes. The next σ in y1 y2 · · · yn can also be moved to the left, convert-
ing ω’s to (σ 2 ωσ −2 )’s, and (σ 2 ωσ −2 )’s to (σωσ −1 )’s. Continuing in this
manner, we convert σ j y1 y2 · · · yn into the specified form.

We will need to refer to the process described above in proofs that oc-
cur below. To facilitate this, we define an operation Σ on words x1 · · · xn ∈
{ω, σωσ −1 , σ 2 ωσ −2 } by first defining Σ on the three basic elements

Σ(ω) = (σ 2 ωσ −2 ),
Σ(σωσ −1 ) = ω,
Σ(σ 2 ωσ −2 ) = (σωσ −1 ),

and then we extend Σ to words using Σ(x1 · · · xn ) = Σ(x1 ) · · · Σ(xn ).

As was exploited in the argument above, Equations 6.2 simply state
that xi σ = σΣ(xi ) for each xi ∈ {ω, σωσ −1 , σ 2 ωσ −2 }. Repeated appli-
cation of this fact shows that x1 · · · xn σ = σΣ(x1 · · · xn ) for any word
x1 · · · xn ∈ {ω, σωσ −1 , σ 2 ωσ −2 }∗ .
As an example, if we use this notation and apply the process described
in the proof of Lemma 6.4 to the word ωωσωσωσωωσ we get

ωωσωσωσωωσ → σΣ(ω)Σ(ω)ωσωσωωσ
→ σ 2 Σ2 (ω)Σ2 (ω)Σ(ω)ωσωωσ
→ σ 3 Σ3 (ω)Σ3 (ω)Σ2 (ω)Σ(ω)ωωσ
→ σ 4 Σ4 (ω)Σ4 (ω)Σ3 (ω)Σ2 (ω)Σ(ω)Σ(ω)
= σ(σ 2 ωσ −2 )2 (ω)(σωσ −1 )(σ 2 ωσ −2 )2
 A Finitely Generated, Infinite Torsion Group 125

where the final equality comes from applying Σ and using the fact that
the group element σ and the operation Σ both have order 3.

Lemma 6.5. There is a homomorphism φ : U → Z3 whose kernel K is

the subgroup generated by the elements {ω, σωσ −1 , σ 2 ωσ −2 }.

Proof. Since σ and ω send the vertices of level k to the vertices of level k
(for any k ≥ 0), the group U sends vertices at level k to vertices at level
k. Thus for a fixed k there is a map from U to Sym3k . In this proof we
focus on the action of U on the three vertices {v0 , v1 , v2 }. The symmetry
σ induces the permutation (123) (on the indices) while ω induces the
trivial permutation. Thus the only permutations that occur in the image
of φ are
Image of φ = {e, (123), (132)} = A3 ;

so φ : U → A3 ≈ Z3 . Since the subgroup generated by σ maps onto the

quotient, φ is surjective.
Because ω is in the kernel of φ, and kernels are normal subgroups, the
set {ω, σωσ −1 , σ 2 ωσ −2 } is contained in K. The reader can check that
the subgroup H, generated by the set {ω, σωσ −1 , σ 2 ωσ −2 }, is a normal
subgroup of U that is contained in K.
If g ∈ U then by Lemma 6.4 we can express g as σ k x1 · · · xn where
each xi ∈ {ω, σωσ −1 , σ 2 ωσ −2 }. Thus every g ∈ U can be expressed as
g = σ k h for k ∈ {0, 1, 2} and h ∈ H. Thus

U/H = {H, σH, σ 2 H} ≈ Z3 .

Therefore H is a normal subgroup of U, contained in K, with the same

quotient. So the kernel K is actually the subgroup we (temporarily)
referred to as H, which is generated by {ω, σωσ −1 , σ 2 ωσ −2 }.

As U is defined to be the group generated by σ and ω, the claim that

U is finitely generated is immediate. The claim that U is infinite follows
from Lemma 6.6 below, since in a finite group no proper subgroup can
map onto the larger group.

Lemma 6.6. The kernel K surjects onto U.

Proof. The subgroup K is the collection of elements in U that fix v0 ,

v1 and v2 . Each of the generators of K, {ω, σωσ −1 , σ 2 ωσ −2 }, can be
expressed as a product of elements of the form σi , σi−1 and ωi (i ∈ Z3 ), as
in Equations 6.1. Since the three subtrees T0 , T1 and T2 do not intersect,
each generator of K can be viewed as an element of U0 × U1 × U2 , where
126 A Finitely Generated, Infinite Torsion Group

each Ui is as in Definition 6.3. It follows that K < U0 × U1 × U2 . We may

then project K onto any one coordinate, giving a map πi : K → Ui . For
example, under π0 we have
ω → σ0 ,
−1
σωσ → ω0 ,
2
σ ωσ −2
→ σ0−1 .
The image of π0 is then the subgroup generated by σ0 and ω0 , hence the
map π0 : H → U0 is onto, and the result follows since U0 ≈ U.
We now turn to the proof of most difficult claim, that every element
has order a power of 3.
We may define a “length” for each g ∈ U by finding an expression
g = σ k x1 · · · xn , where x1 · · · xn ∈ {ω, σωσ −1 , σ 2 ωσ −2 }∗ and n is as
small as possible. We then define


n if k = 0
(g) =
n + 1 if k = 0.
Lemma 6.7. Every g ∈ U has order a power of 3.
Proof. The proof is by induction on (g). The bases case, where (g) = 1,
consists of showing that all of the elements σ, σ 2 , ω, σωσ −1 and σ 2 ωσ −2 ,
have order 3.
Assume for the induction step that (g) = n + 1. There are three cases
to consider:
Case 1: g = σx1 · · · xn .
Let n0 be the number of occurrences of ω among the xi , let n1 be
the number of occurrences of (σωσ −1 ) among the xi , and let n2 be the
number of occurrences of (σ 2 ωσ −2 ). Thus n = n0 + n1 + n2 .
Since we have assumed that g = σx1 · · · xn it follows that
g 3 = σx1 · · · xn · σx1 · · · xn · σx1 · · · xn
= Σ2 (x1 · · · xn )Σ(x1 · · · xn )(x1 · · · xn ) ∈ K,
where we are using the Σ notation and the “sliding σ left” technique
from the proof of Lemma 6.4. In the resulting expression ω occurs n0
times in x1 · · · xn , n1 times in Σ(x1 · · · xn ) and n2 times in Σ2 (x1 · · · xn ).
Thus in total ω occurs n = n0 + n1 + n2 times. Similarly (σωσ −1 ) occurs
n times and (σ 2 ωσ −2 ) occurs n times.
Viewing K as a subgroup of U0 ×U1 ×U2 , as in the proof of Lemma 6.6,
we see that g 3 = (g0 , g1 , g2 ), where each gi ∈ Ui . Applying the
 A Finitely Generated, Infinite Torsion Group 127

expressions in Equations 6.1 we see that each gi can be expressed as

a word in {ωi , σi , σi−1 }∗ of length 3n, with exactly n occurrences of each
element ωi , σi and σi−1 . Using Equations 6.2 we see that the σi and σi−1
can be moved to the left and cancelled, resulting in a word of length n
using the alphabet {ωi , σi ωi σi−1 , σi2 ωi σi−2 }. By induction, the order of
gi is then a power of 3, making the order of g 3 a power of 3, hence the
order of g is a power of 3.
Case 2: g = σ 2 x1 · · · xn .
This is quite similar to Case 1, and is left as an exercise.
Case 3: g = x1 · · · xn xn+1 .
Here, borrowing notation from Case 1, we have n0 + n1 + n2 = n + 1.
We may assume that two of n0 , n1 , and n2 are non-zero, as otherwise g
is a power of ω, σωσ −1 or σ 2 ωσ −2 in which case g would have order 1
or 3. Thus we may assume that each ni is less than or equal to n.
Since the xi ∈ K we know g ∈ K and therefore

g = (g0 , g1 , g2 ) ∈ U0 × U1 × U2 .

Using Equations 6.1 we see that each gi can be expressed as a word of

length n + 1 in {σi , σi−1 , ωi }∗ . In fact, we know exactly how many times
each of the elements of this alphabet shows up. For example, ω0 occurs
n1 times in the expression for g0 , ω1 occurs n2 times in the expression
for g1 , and ω2 occurs n0 times in the expression for g2 . We can replace
each σi−1 that occurs in the expression for gi by σi2 . Then, applying the
“sliding σ to the left” technique of Lemma 6.4, we can express each
gi as

gi = σij x1 x2 · · · xmi

where x1 x2 · · · xmi ∈ {ωi , σi ωi σi−1 , σi2 ωi σi−2 }∗ and mi = ni+1 . If j = 0

then gi ∈ Ui has (gi ) = ni+1 ≤ n and so we are done by induction. If
j = 0 then (gi ) = 1 + ni+1 ≤ n + 1. If (gi ) happens to be ≤ n then we
are done by induction. If (gi ) = n + 1 then we are done by the previous
two cases.

The construction given above was first presented by Gupta and Sidki
[GuSi83]; it is inspired by a construction of Grigorchuk [Gr80]. The first
examples of finitely generated, infinite torsion groups were found in the
1960s; a discussion of the first known example can be found in [Ol95].
The example presented here does not completely settle all questions that
128 A Finitely Generated, Infinite Torsion Group

can be derived from the problem Burnside posed in 1902. In fact, there
is a stronger version, often called the Burnside Problem, which asks:
If n is a fixed positive integer and G is a finitely generated group where every
element in G satisfies g n = e, is G finite?

Burnside noted that this problem is quite tractable for n = 2, where the
answer is “yes,” and established that the claim is also true for n = 3.
Burnside’s question can be further refined by asking if for given in-
tegers k and n (both ≥ 2) there is an infinite group with a generating
set of size k where every element satisfies g n = e. In fact, there is a
“free Burnside group” B(k, n) which is the quotient of the free group of
rank k, Fk , by the smallest normal subgroup containing the nth powers:
{xn | x ∈ Fk }. There is an infinite group with generating set of size k
where every element satisfies g n = e if and only if B(k, n) is infinite.
Burnside proved that B(k, 3) is finite, for all k, but did not determine
the order of this group. In 1933 Levi and van der Waerden proved that
the order of B(k, 3) is 3n where
   
k k
n=k+ + .
2 3
(Question: Why does the fact that the group B(2, 3) has order 33 not
contradict our claim that the group U < Sym(T ), generated by σ and
ω, is infinite?)
In 1940 Sanov showed that B(k, 4) is always finite, and in 1958 Mar-
shall Hall showed that the groups B(k, 6) are all finite and was able to
give a concrete formula for the size of B(k, 6), similar to the formula
above. It is still an open question if B(2, 5) is finite or infinite.
On the other hand, for large values of n there are finitely generated
infinite groups where every element satisfies g n = e. This was first shown
by Novikov and Adian in 1968, who showed that B(k, n) is infinite when
k > 2 and n is an odd number ≥ 665.
There is yet another refinement, formulated by Wilhelm Magnus in
the 1930s:
For a fixed choice of k and n, is there is a largest, finite k generator group in
which every element satisfies g n = e?

A rephrasing of this question asks if B(k, n) has only finitely many finite
quotients. In 1991 Efim Zelmanov showed that this so-called restricted
Burnside Problem is true; this is work that earned him a Fields Medal
in 1994.
 Exercises 129

Exercises
(1) Prove that the following groups are infinite torsion groups, but
neither is finitely generated.
a. A direct sum of a countable number of copies of Z2 :
Z2 ⊕ Z2 ⊕ Z2 ⊕ · · · .
The elements of this group consist of all N-tuples of elements
of Z2 , where only finitely many entries are non-zero.
b. The factor group Q/Z.
(2) Show that any finitely generated group G, where every element
satisfies g 2 = e, is a finite group.
(3) Prove Case 2 of Lemma 6.7.
(4) Show that ωσ has order 9.
(5) Let T be the rooted, infinite binary tree and let Sym(T ) be the
group of all symmetries of T .
a. Show that any α ∈ Sym(T ) sends the root to the root, even
if you just view T as an unrooted tree.
b. Show that Sym(T ) contains an index 2 subgroup isomorphic
to Sym(T ) ⊕ Sym(T ).
c. Show that Sym(T ) is uncountable. (Hence, by Exercise 11
in Chapter 1, Sym(T ) is not finitely generated.)
(6) Does Sym(T ) contain any element of infinite order?
(7) Let Tk be the regular, unrooted, k-valent tree, where k ≥ 3. Show
that Sym(Tk ) is not finitely generated. (Hint: See Exercise 5.)
(8) Let U be the group constructed in this chapter.
a. Prove that for any g ∈ U, g = e, there is a finite group H
and a homomorphism φ : U → H such that φ(g) = e ∈ H.
(Hint: The group U acts on the set of vertices at level k.)
b. Prove that the intersection of all finite index subgroups of
U is the trivial subgroup {e}.
 7
Regular Languages and Normal Forms

The concept of a finite-state automaton has emerged as significant in many

branches of human knowledge and understanding, including linguistics, com-
puter science, philosophy, biology and mathematics .... In our work, finite-state
automata are of fundamental importance: our objective is to use them to un-
derstand individual groups.
. –from Word Processing in Groups [E+92]

7.1 Regular Languages and Automata

We have explored the word problem about as far as is prudent with-
out being a bit more careful with words like “decide” and “construct.”
Here we introduce a standard mathematical model of computation. In
thinking about modelling computation, there are a few key aspects that
need to be preserved. One needs to have a way of inputing a sequence
of information, and then a machine needs to execute various commands
based on the input. One expects the total number of computations to be
finite and possibly that there is some provision for the machine to have
memory. The notion of finite-state automata, and their corresponding
languages, fits these expectations.

Definition 7.1. Given an alphabet S = {x1 , . . . , xn }, any subset of

words in the free monoid on S, L ⊂ S ∗ , is a language.

There are numerous examples of languages. If S = {a}, then the set

of all strings of even length, {a2n | n ∈ N}, forms a language. Similarly
if S consists of the standard letters in the English alphabet, then the
entries of any edition of the Oxford English Dictionary form a language.
The languages we are most interested in are those that can be pro-
duced by a somewhat simple computer.

130
 7.1 Regular Languages and Automata 131

Definition 7.2. An automaton consists of a directed graph M, an as-

sociated alphabet S, and a number of decorations:

1. There is a subset of vertices called the start states.

2. There is a subset A ⊂ V (M) called the accept states.
3. The directed edges are labelled by elements of the alphabet S.

The language accepted by an automaton M is the set of all words

ω ∈ S ∗ corresponding to directed paths pω that begin at a start state
and end at an accept state of M. This language is denoted L(M).
If M is finite then it is referred to as a finite-state automaton, which
is commonly abbreviated as “FSA.”

Example 7.3. The language {a3n | n ∈ N} is the language associated

to the automaton shown in Figure 7.1. As will be our convention in the
rest of this book, vertices with double circles are the accept states, and
vertices with an inscribed “S” are the start states.

S a a

Fig. 7.1. A finite-state automaton whose language consists of the words

{a3n | n ≥ 0}.

Example 7.4. Consider the finite-state automaton shown in Figure 7.2,

whose associated alphabet is {a, b}. The vertices represented by double
circles are the accept states, and the start state is indicated by the letter
S. The language for this FSA consists of all words in {a, b}∗ that contain
an even number of b’s.

Example 7.5. Every element in the infinite dihedral group D∞ can be

represented uniquely by a word in the generators {a, b} that alternates
between a and b. This set of words is the language accepted by either of
the finite-state automata illustrated in Figure 7.3.
The FSA on the left in Figure 7.3 contains the FSA on the right. The
additional vertex in the FSA on the left is called a fail state; once you
reach it, you cannot leave it, and it is not an accept state. The FSA
on the left also has a property that the FSA on the right does not.
Namely, every word in {a, b}∗ can be traced along the edges of the FSA
132 Regular Languages and Normal Forms
a

b b
a S a
b b

Fig. 7.2. A finite-state automaton whose language consists of all words in

{a, b}∗ that contain an even number of b’s.

a a a

S b a S b a

b b b

Fig. 7.3. Two automata whose languages are the reduced words describing
elements in D∞ .

on the left. In the FSA on the right, words containing aa and bb do not
correspond to directed paths.

Definition 7.6. A non-deterministic automaton is an automaton where

some edges are labelled by a new letter, ε, which is not in the original
alphabet of the automaton. The language of a non-deterministic au-
tomaton is the set of all words corresponding to paths that begin at a
start state and end at an accept state, with all of the ε’s stripped out.

Example 7.7. Let L be the language in {a, b, c}∗ consisting of words

starting with a non-empty sequence of a’s, followed by a possibly empty
string of b’s, followed by a possibly empty string of c’s. This language is
the language of the non-deterministic automaton shown in Figure 7.4.
 7.1 Regular Languages and Automata 133

b
a

a S ε

a
c

Fig. 7.4. A non-deterministic automaton whose language consists of all words

in {a, b, c}∗ that begin with some a’s, followed by b’s and then c’s, where there
may be no b’s and/or c’s.

Definition 7.8. A deterministic automaton, often abbreviated “DFA,”

is a finite-state automaton with the following additional requirements:

1. There is exactly one start state.

2. No two edges leaving a vertex have the same label.

A deterministic automaton is complete if for each vertex v, and each

letter a in the alphabet S, there is an edge leaving v labelled a.

The assumption that there is exactly one start state is not always part
of the definition of DFAs. As we will see in Theorem 7.12, in terms of
the languages accepted by DFAs, the two conventions are equivalent.
Similarly, the following result indicates that one can work with DFAs
or complete DFAs and, in terms of the resulting languages, there is no
difference. The proof of Lemma 7.9 is left as an end-of-chapter exercise.

Lemma 7.9. If L is the language associated to a deterministic autom-

aton, then L is the language associated to a complete, deterministic
automaton.

The following definition is sometimes stated as a theorem, the class

of regular languages having been defined in terms of certain properties.
However, for our purposes this definition is the most handy.

Definition 7.10. A regular language is any language that is accepted

by a deterministic automaton.

Theorem 7.11. Let K and L be regular languages with a common

alphabet S. Then the following languages are also regular languages.
134 Regular Languages and Normal Forms

1. The complementary language S ∗ − K

2. The union K ∪ L
3. The intersection K ∩ L
4. The concatenation KL consisting of all words of the form
KL = {ωK ωL | ωK ∈ K and ωL ∈ L}
5. K ∗ = K ∪ KK ∪ KKK ∪ · · ·
This theorem is most easily established using the following, somewhat
surprising, fact.
Theorem 7.12. The set of languages accepted by non-deterministic
finite-state automata is the same as the set of regular languages.
Theorem 7.12 is one of many results whose theme is “you may assume
your automaton has properties X, Y, and Z.” Such statements are of-
ten quite helpful in proving particular results about regular languages.
However, the arguments establishing these results are not critical for
understanding the applications to group theory.
Before starting the proof of Theorem 7.12, we consider how one would
test if a given word is in the language of a non-deterministic automaton.
It is clear how to test if a word is in the language of a deterministic
automaton. Every word corresponds to a unique path in the DFA, and
so you just follow the associated path and check if it ends in an accept
state. But in a non-deterministic automaton there are choices, and we
have to keep track of these choices. If you have a large stack of paper
available, then you could test a given word ω by making piles, one pile
per start state in the given automaton M. Then, reading off the first
letter of ω, you would consider each start state, and write down all of
the locations that you can arrive at from that start state, using only the
first letter of ω and any available ε-edges. You then consult your lists,
and make a new set of stacks, one stack of paper for each state you could
be in after reading the first letter of ω. Continue to follow this process
until you have read all of ω. If the resulting set of states contains an
accept state, then there is a path from some start state to some accept
state, which is labelled by the word ω (with possibly some ε’s tossed in
along the way). The proof of Theorem 7.12 follows this line of thought.
Proof of Theorem 7.12. It is clear that every deterministic FSA is a non-
deterministic FSA, so the set of regular languages is contained in the set
of languages accepted by non-deterministic FSAs. The trick is proving
the other containment.
 7.1 Regular Languages and Automata 135

Let M be a non-deterministic finite-state automaton. We first build

an automaton that accepts the same language as M and that does not
use the edge label ε. Let S be a subset of V(M). The ε-closure of S,
ε(S), is the set of all vertices of M that can be reached by a directed
path beginning in S. A subset S ⊂ V(M) is ε-closed if S = ε(S).
Let Mε be an automaton whose states are the ε-closed subsets of
V(M). If S ⊂ V(M), and a ∈ S, define Sa to be the set of all states
at the ends of edges labelled a that begin in S. The start states of Mε
are all ε-closed subsets that contain a start state; the accept states are
all ε-closed subsets that contain an accept state of M; and there is an
edge labelled a from S to ε(Sa) for each a ∈ A. By construction, in
Mε there are no edges labelled by ε, and the reader can verify that
L(M) = L(Mε ).
We can now convert Mε into a deterministic automaton, D. The states
of D consist of all the subsets of V (Mε ). The single start state of D is
the subset of V (Mε ) consisting of all the start states of Mε . The accept
states of D are the subsets of V (Mε ) that contain at least one accept
state of Mε . In D there is an edge from U to U  labelled by x if, for
each v ∈ U , there is an edge labelled x from v to some v  ∈ U  , and U 
is entirely composed of such vertices. That is, there is an edge labelled
x from U to the vertex corresponding to the set

U  = {v  ∈ V (Mε ) | v  is at the end of an edge

labelled x that begins at some v ∈ U }.

By its construction, D is a deterministic automaton. Further, if ω ∈

L(Mε ) then ω describes a directed edge path in D starting at the start
state. Further, the terminal vertex of the associated path pω corresponds
to a set of vertices of Mε , at least one of which is an accept state of
Mε , since ω ∈ L(Mε ). Thus L(D) = L(Mε ) = L(M).

Proof of Theorem 7.11. (1) By Lemma 7.9 we may assume that the lan-
guage K = L(M) where M is a complete deterministic automaton. The
language S ∗ − K is then accepted by the machine M◦ formed by taking
M and making all non-accept states into accept states, and making all
accept states non-accept states.
(2) Let MK and ML be deterministic automata whose languages are
K and L respectively. The disjoint union of MK and ML is a non-
deterministic automaton (with two start states) that accepts K ∪ L. By
Theorem 7.12, K ∪ L is then regular.
136 Regular Languages and Normal Forms

(3) Since K ∩ L = S ∗ − ((S ∗ − K) ∪ (S ∗ − L)), this follows from (1)

and (2).
(4) Construct a non-deterministic automaton by forming the disjoint
union of MK and ML , declaring that the only start state of the new
machine is the start state of MK , and adding edges labelled ε from each
accept state of MK to the start state of ML .
(5) From each accept state in MK , add a ε-labelled edge connecting
it to the start state of MK .
Example 7.13. The set of freely reduced words in {x, x−1 , y, y −1 }∗
forms a regular language. An automaton accepting this language is
shown in Figure 7.5. It has five states, one for each possible “last letter
read” as one considers a given word. Initially there is no “last letter
read” and so one starts at the central vertex, which is connected to each
of the other four states. If one is in the state corresponding to having
just read an x, then one can transition to any state except the state
corresponding to having just read an x−1 , as one would not read x−1
after an x in a freely reduced word. The rest of the structure of this FSA
is similar.
The elements of the free group F2 correspond to the freely reduced
words in {x, x−1 , y, y −1 }∗ . Thus these elements correspond to the words
in a regular language. The even subgroup of F2 , introduced in the proof
of Proposition 3.12, corresponds to all freely reduced words that are of
even length. It is easy to construct an FSA whose language consists of
all words in {x, x−1 , y, y −1 }∗ that are of even length. (Try it!) Thus by
part (3) of Theorem 7.11, the set of words corresponding to the even
subgroup is also a regular language.

7.2 Not All Languages are Regular

Not all languages are regular languages. In Example 7.13 a deterministic
automaton is constructed that can remember a finite amount of infor-
mation. Namely, the machine can remember the last letter that it has
read. One of the easiest ways to detect if a language is regular or not is
to consider if one would need to have unlimited memory in order to de-
termine if an arbitrary word is in the language. This instinct is codified
by the Pumping Lemma.
Theorem 7.14 (Pumping Lemma). Let L be a regular language. Then
there is an integer n ≥ 1 such that any word x ∈ L of length greater
than n can be expressed as x = uvw where v is a non-empty word, and
 7.2 Not All Languages are Regular 137

X x

y y y

X x
X S x

Y Y Y

x
X

Fig. 7.5. A finite-state automata whose language consists of all freely

reduced words in {x, x−1 , y, y −1 }∗ . Here we have used X to denote x−1 and Y
to denote y −1 .

1. |u| < n;
2. uv i w ∈ L for all i ≥ 0.

Proof. Let M be a complete DFA whose associated language is L. Set

n = |V (M)| and let x be a word of length greater than n. Because M
is a complete DFA, there is a unique directed edge path starting at the
start state whose sequence of labels matches the word x. Since the length
of x is greater than n, this path must visit some vertex z ∈ V (M) twice.
Thus the path associated to x contains a circuit. Let u be the (possibly
empty) prefix of x associated to the initial portion of the path from the
start state to z; let v be the subword corresponding to the subpath from
z back to z; and let w be the (possibly empty) suffix of x corresponding
to the part of the path from z to the accept state at the end of this path.
It then follows that the words uv i w (for i ≥ 0) all describe paths starting
at the start state and ending at an accept state. The only difference in
138 Regular Languages and Normal Forms

the paths is how many times they wind around the closed path from z
to z.

v
w
u

S Z

Fig. 7.6. The idea behind the Pumping Lemma.

Example 7.15. The Pumping Lemma can be used to prove that the
language L = {an bn | n ≥ 1} is not a regular language. Assume to
the contrary that L is regular and let n be the number of states of an
automaton M whose language is L. Then the path corresponding to
an+1 contains a closed subpath, of some length k. It follows that the
word an+1+k bn must also be accepted by M, contradicting the claim
that L = L(M).
Definition 7.16. Let L be a language in S ∗ and let ω ∈ S ∗ . The cone
type of ω is the set of all words ω  ∈ S ∗ such that ω · ω  ∈ L. Denote the
cone type of ω by Cone(ω). The collection
Cone(L) = {Cone(ω) | ω ∈ S ∗ }
is the cone type of the language L.
Example 7.17. Let S = {a} and let L be the language of all words of
even length. Then Cone(a) = Cone(a3 ) = Cone(a5 ) = · · · and they are
all equal to the set of words of odd length, while Cone() = Cone(a2 ) =
Cone(a4 ) = · · · are all the words of even length. Thus Cone(L) consists
of exactly two sets, the set of words of even length and the set of words
of odd length.
Theorem 7.18 (Myhill–Nerode Theorem). A language L is regular if
and only if its set of cone types is finite.
Proof. We first consider the forward direction. If L is regular then it is
accepted by a complete DFA M. For each vertex v ∈ V (M) let Mv be
 7.2 Not All Languages are Regular 139

the automaton formed by making v the only start state. The cone type
of a given word ω is then determined by the end vertex of the associated
path pω . If this vertex is denoted vω then Cone(ω) = L(Mvω ). Thus the
number of cone types is at most the number of vertices in M.
To establish the converse, assume that Cone(L) is finite. We construct
an automaton M starting with declaring the states of M to be the cone
types of L. Designate Cone(ε) to be the unique start state, where ε
represents the empty word. The accept states of M are the cone types
that contain ε.
We add an edge labelled x from the vertex given by Cone(ω) to
Cone(ωx) for each cone type and each x ∈ S. There is a possibility
that this definition depends on the word ω, not on the cone type of ω.
That is, perhaps Cone(ω) = Cone(
ω ), but Cone(ω · x) = Cone(

ω · x).
∗
This cannot happen, because if x ∈ S and ω ∈ S then

Cone(ωx) = {ω  ∈ S ∗ | xω  ∈ Cone(ω)} .

ω ) then
Thus if Cone(ω) = Cone(

Cone(ωx) = {ω  ∈ S ∗ | xω  ∈ Cone(ω)}
= {ω  ∈ S ∗ | xω  ∈ Cone(

ω )} = Cone(

ω x)

for each x ∈ S. The reader may check that this automaton M is a

complete DFA.
What is the language accepted by M? First, we can characterize L by
noticing that a word ω is in L if and only if the empty word ε ∈ Cone(ω).
Let ω be an arbitrary word in S ∗ . Since M is a complete DFA, there is
a unique path in M, starting at the start state, and traversing edges as
described by ω. By the construction of M, this path ends at the state
associated to Cone(ω). Hence ω ∈ L(M) if and only if Cone(ω) is an
accept state, which is if and only if ε ∈ Cone(ω). Thus L(M) = L by
the characterization of L given at the start of this paragraph.

Example 7.19. Let the alphabet be S = {a, b} and let

L = {ωω | ω ∈ S ∗ }

be the language of formal squares of words in S ∗ . Then bai b ∈ Cone(ai )

but bai b ∈ Cone(aj ) for any j = i. Thus the language L has infinitely
many distinct cone types, so, by the Myhill–Nerode Theorem, L is not
regular.
140 Regular Languages and Normal Forms

7.3 Regular Word Problem?

We began our discussion of regular languages motivated by a desire to
have a more formal understanding of what it means for something to be
computable. We now return to the group theory setting, ready to test
drive our new understanding.
Let G be a finitely generated group with generating set S. Define

WP(G, S) = {ω ∈ {S ∪ S −1 }∗ | π(ω) = e ∈ G}.

Thus WP(G, S) is the set of all words in the generators and their inverses
that evaluates to the identity in G. This set is often colloquially referred
to as “the word problem” for G, since being able to solve Dehn’s word
problem for G (and generating set S) amounts to being able to determine
if a given word ω ∈ {S ∪ S −1 }∗ is in WP(G, S).
Since WP(G, S) ⊂ {S ∪ S −1 }∗ , it is a language, and one can ask:
When is WP(G, S) a regular language? Surprisingly, it is not easy to
come up with examples where the word problem is a regular language.
For example, let G be an infinite cyclic group with generator x. Then a
word in {x, x−1 }∗ represents the identity if and only if the total exponent
sum is zero. But then x−i ∈ Cone(xi ) but x−i ∈ Cone(xj ) for any i and
j ∈ Z where j = i. So WP(G, x) has infinitely many cone types, and the
Myhill–Nerode Theorem (Theorem 7.18) implies that WP(G, x) is not a
regular language. Faced with this non-example, you might start looking
at free groups, free products, the Baumslag–Solitar group BS(1, 2), etc.,
but you would only be wasting your time.

Theorem 7.20. The word problem WP(G, S) is a regular language if

and only if G is a finite group.

Proof. If G is finite, the Cayley graph with respect to {S ∪ S −1 } forms a

complete DFA, with start and accept state the vertex corresponding to
the identity. The language accepted by this DFA is precisely WP(G, S).
Conversely, suppose WP(G, S) = L(M) for some finite-state automa-
ton M. One may assume that every vertex in M is connected by a
directed path to an accept state. Otherwise form a new automaton that
accepts the same language as M by simply removing that vertex and all
incident edges. Repeatedly applying this technique trims M to a finite
automaton where every vertex is connected by a directed path to at least
one accept state.
Let ω and ω  be two words whose corresponding paths both end at
z ∈ M. Let v be a word describing a path from z to an accept state of
 7.4 A Return to Normal Forms 141

M. Then ωv and ω  v are both in L(M). It follows that

π(ωv) = π(ω  v) = e ∈ G.
But in the group G we have π(ω)π(v) = π(ω  )π(v). Thus, by right
cancellation, π(ω) = π(ω  ). So if two words describe paths ending at the
same vertex, then they represent the same element of G. But then the
order of G can be no larger than the number of states of the finite-state
automaton of M.
Remark 7.21. Theorem 7.20 was first proved by Anisimov in 1971
[An71]. It is a striking result that seems to show that asking when
WP(G, S) is a regular language is not the right question, especially if
one is mainly interested in infinite groups!
Readers with a background in formal language theory might be
interested in exploring an analogous theorem of Muller and Schupp,
who showed that WP(G, S) is a context-free language if and only if G
contains a free subgroup of finite index [MuS83].

7.4 A Return to Normal Forms

Recall that, given a group G and generating set S, a normal form is a
subset of the free monoid {S ∪ S −1 }∗ that maps bijectively to G under
the evaluation map π : {S ∪ S −1 }∗ → G. (This idea was introduced in
Section 5.1.) Thus a normal form N F ⊂ {S ∪S −1 }∗ can be thought of as
a language, and one might hope that it forms a regular language. When
this happens, we say G has a regular normal form. It is also common to
say that G admits a regular normal form.
Example 7.22. The freely reduced words are a normal form for Fn . As
part of Example 7.13 we showed that this is a regular language, hence
Fn has a regular normal form.
Example 7.23. Let a and b be generators of the free abelian group
Z⊕Z. Then every element can be expressed as ai bj for some i and j ∈ Z.
(This is NF1 from Example 5.2.) This too is a regular language. Perhaps
this is most easily shown by noting that {ai | i ∈ Z} and {bj | j ∈ Z} are
regular languages, and then appeal to the fact that the concatenation of
two regular languages forms a regular language.
Proposition 7.24. Let S be a finite generating set for a group G. If G
has a regular normal form N (G) ⊂ {S ∪ S −1 }∗ then for any other finite
generating set T there is a regular normal form contained in {T ∪T −1 }∗ .
142 Regular Languages and Normal Forms

Proof. Since T is a set of generators, each s ∈ S (or S −1 ) can be rep-

resented as a word in {T ∪ T −1 }∗ , s = ωs . The language obtained by
replacing each generator s by ωs in each word in N (G) is then a regular
normal form in {T ∪ T −1 }∗ . The corresponding automaton is formed by
replacing each edge labelled s by an edge path labelled ωs , where none
of the new vertices are start or accept states.
Example 7.25. The infinite dihedral group D∞ is generated by two
reflections, a and b. It is also generated by a reflection, a, and a trans-
lation τ = ba. A regular normal form for D∞ with respect to {a, b}
consists of all words which alternate a’s and b’s, and an FSA accepting
this language is shown on the left in Figure 7.7. To construct a normal

a a τ

S b a S b a

b τ b a

Fig. 7.7. Converting the finite-state automaton giving a regular normal form
for D∞ with respect to {a, b} to one using {a, τ, τ −1 } as its generating set.

form using {a, τ, τ −1 } we note that a = a and b = τ a. Thus we should

replace all edges labelled b in the original FSA by edge paths labelled τ a.
We note that the resulting normal form is neither efficient nor natural
to the new set of generators. For example, instead of using the word τ −1
to represent a horizontal translation to the left, this new normal form
uses aτ a.
The following is an exercise in applying Theorem 7.11.
Proposition 7.26. If G and H have regular normal forms, then so does
G ⊕ H and G ∗ H.
Corollary 7.27. Any free product of two finite groups has a regular
normal form.
Proof. Every finite group admits a regular normal form. Simply take
S = G \ {e}, and let N (G) be the words of length ≤ 1. The result then
follows from Proposition 7.26.
 7.5 Finitely Generated Subgroups of Free Groups 143

Bob Gilman introduced the study of groups admitting regular normal

forms in 1987 [Gi87], where he established the following result:

Theorem 7.28. If G has a regular normal form, and G is not finite,

then G has an element of infinite order.

Proof. Let M be a complete automaton whose language is a regular

normal form for G, and let n be the number of vertices of M. Then
since G is infinite there is a word x ∈ L(M) of length greater than n. So
by the Pumping Lemma (7.14) we may write x = uvw where, for each
i ≥ 0, uv i w ∈ L(M). Since L(M) is a normal form, π(uv i w) = π(uv j w)
for i = j. But then

π(uv i w) = π(uv j w)
⇒ π(u)π(v i )π(w) = π(u)π(v j )π(w)
⇒ π(v)i = π(v)j

for i = j. It follows that π(v) is an element of infinite order in G.

A mild restatement of this result is:

Corollary 7.29. Finitely generated, infinite torsion groups, like the

group U considered in Chapter 6, do not have regular normal forms.

7.5 Finitely Generated Subgroups of Free Groups

If G is a finitely generated infinite group, it may be the case that some
subgroups of G are not finitely generated. For example, as a homework
problem in Chapter 3 you showed that the kernel of the map from F2
onto Z ⊕ Z is a free group with basis

S = {[xi , y j ] | i, j ∈ Z and i · j = 0}.

Here is another normal subgroup of F2 that is not finitely generated.

Lemma 7.30. Let F2 be generated by x and y and let φ : F2 → Z be the

map induced by sending both x and y to 1. The kernel of φ is free with
basis S = {xi y −i | i ∈ Z}.

Proof. The given set is in the kernal, and the quotient by the normal sub-
group containing these elements is isomorphic to Z. So it suffices to show
that the subgroup of F2 generated by S is normal. But the following com-
putations show that the conjugate of each element in S is in the subgroup
144 Regular Languages and Normal Forms

generated by S
x · xi y −i · x−1 = xi+1 y −(i+1) · yx−1 = xi+1 y −(i+1) · (xy −1 )−1 ,
x−1 · xi y −i · x = xi−1 y −(i−1) · y −1 x = xi−1 y −(i−1) · (x−1 y)−1 ,
y · xi y −i y −1 = yx−1 · xi+1 y −(i+1) = (xy −1 )−1 · xi+1 y −(i+1) ,
y −1 · xi y −i · y = y −1 x · xi−1 y −(i−1) = (x−1 y)−1 · xi−1 y −(i−1) .
So the subgroup generated by S is normal. Exercise 7 asks you to show
that S forms a basis.
The following example, combined with Theorem 7.32 below, indicates
that determining which subgroups of a given group G are finitely gen-
erated can be rather subtle.
Example 7.31. The intersection of two finitely generated subgroups of
a group G is not necessarily a finitely generated subgroup. Consider the
group F2 ⊕ Z with {x, y} being the basis for F2 and z a cyclic generator
of Z. Then the kernel K of the map ϕ : F2 ⊕ Z → Z induced by
ϕ(x) = ϕ(y) = ϕ(z) = 1 ∈ Z
is a free group generated by {x−1 z, y −1 z}. The proof of this claim is
similar to the proof of Lemma 7.30: show that the subgroup generated
by this set is normal and note that the quotient is indeed Z. Normality
is established by equations of the form:
y · x−1 z · y −1 = yz −1 · x−1 z · y −1 z.
Working out the details is left to Exercise 13.
The intersection of K with direct factor F2 is the kernel of the map
φ : F2 → Z considered in Lemma 7.30 above:
ker(ϕ) ∩ F2 = ker(φ).
The subgroups K = ker(ϕ) and F2 are both finitely generated but, as
we have seen, ker(φ) is not finitely generated.
Theorem 7.32 (Howson’s Theorem). The intersection of finitely gen-
erated subgroups of a free group is a finitely generated free group.
Our proof of Howson’s Theorem will exploit our understanding of
regular languages. If G is a group with finite generating set S, then a
subgroup H < G is the image of a regular language over S ∪ S −1 if
there is a regular language L ⊂ {S ∪ S −1 }∗ such that π(L) = H. This
condition is independent of one’s choice of finite generating set (see the
analogous argument in the proof of Proposition 7.24). Moreover, there
 7.5 Finitely Generated Subgroups of Free Groups 145

is a surprising characterization of which subgroups are the images of

regular languages:
Theorem 7.33. Let S be a finite set of generators for a group G. A
subgroup H of G is the image of a regular language L ⊂ {S ∪ S −1 }∗ if
and only if H is finitely generated.
Proof. If H is finitely generated, then fix a generating set T and for
each t ∈ T choose a word ωt ∈ {S ∪ S −1 }∗ such that π(ωt ) = t. The free
monoid on these words and their formal inverses is a regular language
that projects to H under π.
Conversely let L = L(M) be a regular language that projects to
a subgroup H < G. We may assume that M has a number of nice
properties:
1. There are no ε-edges.
2. There is a single start state that is not the terminal vertex of any
directed edge of M. This may or may not be an accept state.
(Exercise)
3. Aside from the possibility of the start state being an accept state,
there is a single accept state. This accept state is not the initial
vertex of any directed edge of M. (Exercise)
4. Every vertex v of M can be reached by a directed path beginning
at the start state (otherwise remove such vertices).
5. From every vertex v of M there is a directed path from v to the
accept state of M (otherwise remove such vertices).
Given all these conditions, there is a directed spanning tree T ⊂ M,
which is rooted at the start state of M. In other words, T is a sub-
directed graph that contains every vertex of M and, given any vertex,
there is a path in T from the single start state to that vertex. For each
edge e ∈ M \ T let α(e) denote the label of the path in T connecting
the start state to the initial vertex of e, and let β(e) be the label of the
path connecting the start state to the terminal vertex of e. Let z denote
the label of the path in T from the start state to the accept state.
If ω ∈ L then ω describes a path from the start state of M to the
accept state of M. We may then view ω as a sequence of subwords
ω = g0 h1 g1 h2 g2 · · · gn−1 hn gn
where each gi is the label of a path contained in the spanning tree
T and each hi is the label of an edge ei ∈ T . Note: while some of
the gi may be empty words, we may assume that none of the hi are
146 Regular Languages and Normal Forms

z(e)
b (e)
S e
a (e)

Fig. 7.8. The spanning tree and other featured elements of the automaton.

empty. It follows from the definitions that α(e1 ) = g0 and more generally
α(ei+1 ) = β(ei )gi . Thus we may write
π[ω] = π[α(e1 )h1 β(e1 )−1 · α(e2 )h2 β(e2 )−1 · · · α(en )hn β(en )−1 · z].
If we let
(e) denote the label of an edge e ∈ M then the image of L is
contained in the subgroup generated by
E = {π(z), π[α(e)
(e)β(e)−1 ] | e is an edge in M \ T }.
It then remains to be shown that each of these elements is contained in
H, for then E will be a finite set of generators for H. Since z ends in an
accept state, π(z) ∈ H. If e is an edge in M \ T let z(e) be a path from
the target vertex of e to the accept state. Then
π[α(e)eβ(e)−1 ] = π[α(e)ez(e) · (β(e)z(e))−1 ].
But π[α(e)·e·z(e)] and π[(β(e)·z(e))] are both in H, so π[α(e)·e·β(e)−1 ]
is in H.
The proof given above, when restricted to the case of subgroups of
free groups, establishes the following result:
Corollary 7.34. Finitely generated subgroups of free groups are free.
(This result is, of course, an immediate corollary of the Nielsen–Schreier
Theorem 3.23, but the formal language argument given below is useful
none-the-less.)
Proof. If H is a finitely generated subgroup of Fn then one direction of
Theorem 7.33 implies that H is the image of a regular language L. The
 7.5 Finitely Generated Subgroups of Free Groups 147

proof of Theorem 7.33 indicates that every h ∈ H can be expressed as a

product of the elements

E = {π(z), π[α(e)
(e)β(e)−1 ] | e is an edge in M \ T }

and their inverses. But any such word which is freely reduced when
written as a product of elements of E ∪E −1 cannot represent the identity
in Fn . Consider for example the product
 −1
α(e1 )
(e1 )β(e1 )−1 · α(e2 )
(e2 )β(e2 )−1
= α(e1 )
(e1 )β(e1 )−1 · β(e2 )
(e2 )−1 α(e2 ).

If β(e1 )−1 ·β(e2 ) cancels completely, that means that the initial vertex of
e2 is the initial vertex of e1 . But since this expression was freely reduced
when written in E ∪ E −1 , e1 = e2 . Thus no more cancellation occurs. An
induction argument along these lines establishes that a freely reduced
word written with n letters from E ∪E − is equivalent to a freely reduced
word of length at least n, when written in terms of the generators of Fn .
Thus

E = {π(z), π[α(e)
(e)β(e)−1 ] | e is an edge in M \ T }

forms a free basis for the image of L, and so H is a free subgroup.

Before finally proving Howson’s Theorem, we have one more lemma

that needs to be established.

Lemma 7.35. Let L be a regular language over {S ∪ S −1 } and let R be

the language formed by freely reducing the words in L. If L is regular,
then so is R.

Proof. Let L = L(M), for some finite automaton M. If for any s ∈ S

there is a path in M with label ss−1 or s−1 s (some edges of which may
be labelled by ε) from a vertex u to a vertex v, add a new edge from u
to v labelled ε. Call this new automaton M . If M accepted a not-freely
reduced word wss−1 u then the new automaton will accept this word and
the word wu. Thus M accepts both L and the reduced words formed
from words in L. Intersecting L(M ) with the regular language of freely
reduced words, and applying Theorem 7.11, establishes the lemma.

Proof of Howson’s Theorem. Let H and K be two finitely generated sub-

groups of Fn . Then H is the image of a regular language LH and K is
the image of a regular language LK . By Lemma 7.35 we may assume
both languages consist entirely of freely reduced words, which since Fn
148 Regular Languages and Normal Forms

is free means they consist of all freely reduced words representing ele-
ments of H and K. The intersection LH ∩ LK is then a regular language
consisting of all freely reduced words which represent elements in H and
in K, hence LH ∩ LK maps onto H ∩ K. So H ∩ K is finitely generated
by Theorem 7.33.
In the argument above, being able to claim that the languages consist
of freely reduced words is critical. In general, if LH and LK are two
regular languages whose images are H and K respectively, then LH ∩LK
maps to H ∩ K, but it does not have to map onto H ∩ K. There may be
elements in the intersection that have different words representing them
in LH and LK .
Howson proved Theorem 7.32 in 1954. The topological techniques he
used were quite different than the language-theoretic approach given
above. In addition to knowing that H ∩ K is finitely generated it would
be nice to have a method for bounding the rank of H ∩ K. In his paper
Howson established:
rk(H ∩ K) ≤ 2rk(H)rk(K) − rk(H) − rk(K) + 1.
In papers published a few years later Hanna Neumann improved How-
son’s bound, and conjectured that the rank of H ∩ K is at most
(rk(H) − 1)(rk(K) − 1) + 1.
This question is still open, and is referred to as the Hanna Neumann
Conjecture. Hanna Neumann’s son, Walter Neumann, has suggested a
strengthened version of the conjecture [Ne90], and this strengthened
version can be formulated as a purely graph theoretic conjecture [Di94].
Remark 7.36. The reader who would like to continue exploring con-
nections between formal language theory and the study of infinite groups
might begin by consulting Bob Gilman’s survey article [Gi05].

Exercises
(1) Show that the language {ak | k ≡ 0 or 2 (mod 5)} is a regular
language.
(2) Prove Lemma 7.9: If L is the language associated to a determi-
nistic automaton, then L is the language associated to a complete,
deterministic automaton.
(3) Let L be a language. Define its reverse language to be
Rev(L) = {xn · · · x1 | x1 · · · xn ∈ L}.
 Exercises 149

Show that if L is a regular language then its reverse is a regular

language.
(4) Let L be a regular language. Show that L is the language of a finite
(but not necessarily deterministic) automaton with:
• a single start state (that may be an accept state as well)
• at most one more accept state
• no ε-transitions.
Further, show that the start state is not at the end of any transition
and the accept state is not the initial vertex of any transition.
(5) Fix an alphabet S and a language L ⊂ S ∗ . For any two words
ω, ω  ∈ S ∗ define a relation

ω ∼ ω  if and only if Cone(ω) = Cone(ω  ).

Prove that this is an equivalence relation on S ∗ and rephrase the

Myhill–Nerode Theorem in terms of this equivalence relation.
(6) Determine if the following languages, contained in {a, b}∗ , are reg-
ular.
a. L = {ai bj | i is even and j is odd}
b. L = {ai | i is a Fibonnaci number}
c. L = {ai bj | j > i}
d. L consists of all words with more b’s than a’s.
(7) Show that the set {xi y −i | i ∈ Z\{0}} is a basis for a free subgroup
of F2 .
(8) Show that every finitely generated abelian group admits a regular
normal form.
(9) Show that BS(1, 2) admits a regular normal form.
(10) In Lemma 6.4 we showed that every element g in the finitely gen-
erated, infinite torsion group, B, can be expressed as:

g = σ k x1 x2 · · · xn

where k ∈ {0, 1, 2} and x1 x2 · · · xn ∈ {ω, σωσ −1 , σ 2 ωσ −2 }. Why

isn’t this a regular normal form?
(11) Prove Proposition 7.26: If G and H have regular normal forms,
then so does G ⊕ H and G ∗ H.
(12) Prove that the group generated by the reflections in the sides of
an equilaterial triangle, W333 , admits a regular normal form.
150 Regular Languages and Normal Forms

(13) Let G = F2 ⊕ Z with {x, y} being the basis for F2 and z a cyclic
generator of Z. Prove that the kernel K of the map φ : F2 ⊕Z → Z,
induced by φ(x) = φ(y) = φ(z) = 1, is a free group generated by
{x−1 z, y −1 z}.
(14) The even subgroup of F2 is a free group of rank 3. In Exercise 14
in Chapter 3 you showed that the kernel of φ : F2 → Z3 is a free
group of rank 4. What is the rank of their intersection?
 8
The Lamplighter Group

Brief explanation: a lamplighter lives in a town whose street map is C and

walks about the town lighting various lamps (elements of K, especially when
K = Z/2Z).
. –Walter Parry

A short drive from Lafayette College (in Easton, Pennsylvania) one can
find barns decorated in a classic Pennsylvania Dutch style with Hex
diagrams placed around on the sides of the barn. In Figure 8.1 there is
a simplified version of this, where copies of a single Hex diagram have
been placed at the corners of a square.

Fig. 8.1. An arrangement of simple Hex diagrams.

151
152 The Lamplighter Group

Let G be the color-preserving symmetry group of this pattern. If we

want to understand G – perhaps at the level of being able to draw the
Cayley graph of G – we need some basic information about its structure.
How many elements are there in G? Is there a natural generating set?
Given that set of generators, is there a normal form for the elements
of G?
The Hex designs only exhibit rotational symmetry, so their individ-
ual symmetry groups are cyclic of order 6. Because the Hex designs
are disjoint, the subgroup of G consisting of individual rotations with
no permuting of the locations of the Hexes is isomorphic to the direct
product (Z6 )4 = Z6 ⊕ Z6 ⊕ Z6 ⊕ Z6 . This gives a total of 64 = 1296 sym-
metries, but this does not account for all of the symmetries of the Hex
arrangement. Because the four Hex designs are identical, we can freely
permute them. Thus we also have Sym4 showing up as a subgroup of
the symmetry group of this Hex arrangement.
On the other hand, any symmetry of the Hex arrangement can be
described by permuting the Hex designs (without rotation), and then
rotating each individual design. Thus each symmetry can be described
by a pair consisting of an element from (Z6 )4 and a permutation from
Sym4 . It is easy to see that this description is unique, hence the order
of G is 64 · 4! = 31104. This is certainly too large for us to draw the
complete Cayley graph, but our analysis so far is quite helpful in giving
a general description of what the Cayley graph looks like.

Exercise 8.1. If A and B are normal subgroups of G, where G = A · B

and A ∩ B = e ∈ G, then G ≈ A ⊕ B. We have just argued that
there are two subgroups of the symmetry group of our Hex diagram,
(Z6 )4 and Sym4 , whose product is G. Further, the intersection of these
subgroups is the identity. Prove that G is not the direct product of these
two subgroups, by showing that the Sym4 subgroup is not a normal
subgroup of G.

The exercise above shows that G is not the direct product of (Z6 )4
and Sym4 ; G is, however, a “semi-direct product” of (Z6 )4 and Sym4 ,
denoted (Z6 )4
Sym4 . To see the distinction between direct and semi-
direct products, we should examine how the binary operation works. In
order to emphasize that we are not working with a direct product, we
denote elements of G using square brackets:

G = {[(a, b, c, d), σ] | (a, b, c, d) ∈ (Z6 )4 and σ ∈ Sym4 }.

 The Lamplighter Group 153

Now consider the product

[(1, 0, 0, 4), (34)] · [(0, 0, 2, 1), (23)].

The corresponding action on the Hex arrangement can then be expressed

in four steps:

1. The second Hex and third Hex are exchanged.

2. The Hex in the third position is rotated through an angle of 120◦
and the Hex in the fourth position is rotated through an angle of
60◦ .
3. The third Hex and fourth Hex are exchanged.
4. The first Hex is rotated 60◦ and the fourth Hex is rotated 240◦ .

So, in total, what happened?

1. The first Hex is never permuted, and it is rotated through an

angle of 60◦ .
2. The second Hex is sent to the third position, then it is rotated
120◦ ; after this it is moved into the fourth position and rotated
an additional 240◦ . Thus at the end of this process, the second
Hex ends up in the fourth position, where it is unrotated.
3. A similar analysis shows the third Hex ends up in the second
position, unrotated.
4. The fourth Hex ends up in the third position, rotated 60◦ .

Thus we have:

[(1, 0, 0, 4), (34)] · [(0, 0, 2, 1), (23)] = [(1, 0, 1, 0), (243)].

Notice that the permutations were unaffected by the presence of the

subgroup of individual rotations, but the rotations seem to have been
affected by the permutations. In fact, we can expand the equation above
to make this explicit:

[(1, 0, 0, 4), (34)] · [(0, 0, 2, 1), (23)]

= [(1, 0, 0, 4) + (0, 0, 1, 2), (34)(23)]
= [(1, 0, 1, 0), (243)].

The key insight is that, in moving the permutation (34) past the ro-
tations (0, 0, 2, 1), we actually apply the permutation to the 4-tuple of
rotations. This example, we hope, provides some insight into and moti-
vation for the following construction.
154 The Lamplighter Group

Definition 8.2. Let H be a group and let φ : K → Aut(H) be a group

homomorphism from K to the automorphism group of H. For notational
convenience below, let φk denote the automorphism φ(k) ∈ Aut(H).
The elements of the associated semi-direct product are ordered pairs of
elements [h, k]. The binary operation is:
[h1 , k1 ] · [h2 , k2 ] = [h1 φk1 (h2 ), k1 k2 ].
Since φk1 (h2 ) ∈ H, the product h1 φk1 (h2 ) is computed in H; the product
k1 k2 is computed in K.
It is a not-particularly-pleasant exercise to verify that this set of el-
ements and binary operation does indeed define a group. It is denoted
H
K, or H
φ K when one needs to be explicit about the homomor-
phism φ from K to Aut(H).
We leave the proof of the following result as an exercise.
Theorem 8.3. Let H be a group, let φ : K → Aut(H), and let G be the
semi-direct product H
K.
1. The map from H to G given by h → [h, eK ] is an injective ho-
momorphism.
2. The image of H in G (as above) is a normal subgroup.
3. The map from K to G given by k → [eH , k] is an injective homo-
morphism.
4. The quotient of G mod H is isomorphic to K, G/H ≈ K.
5. There is a conjugation relation that in shorthand says
khk −1 = φk (h),
or more accurately:
[eH , k] · [h, eK ] · [eH , k −1 ] = [φk (h), eK ].
Example 8.4. The most commonly encountered examples of semi-
direct products are the dihedral groups. Here we have Dn ≈ Zn
φ Z2 .
The subgroup Zn consists of the rotations and the indicated homomor-
phism φ : Z2 → Aut(Zn ) takes the non-identity element of Z2 to the
automorphism of Zn given by sending each element to its inverse. Item
5 of Theorem 8.3 is essentially just the familiar fact that f ρi f = ρ−i ,
where f is a chosen reflection, and ρ is a rotation.
Example 8.5. Let α be the automorphism of Z7 given by α(x) = 2x,
so that α3 (x) = 23 x ≡ x (mod 7). There is then a homomorphism
φ : Z3 → Aut(Z7 ) given by φ(1) = α. The induced semi-direct product
 The Lamplighter Group 155

Z7
Z3 is a non-abelian group of order 21. In fact, it is the unique

non-abelian group of order 21. The Cayley graph, with respect to {α, b}
where b = 1 ∈ Z7 , is shown in Figure 8.2.

Fig. 8.2. A Cayley graph of the non-abelian group of order 21.

Similarly, let φ be the automorphism of Z9 given by φ(x) = 4x, so that

φ is of order 3 in Aut(Z9 ). Then we can form a non-abelian group of order
27, Z9
Z3 . The Doyle graph, shown in Fig 1.29, is the underlying graph
of a Cayley graph of this group. If we let a be an element generating
Z9 and b an element generating Z3 , then the Doyle graph is the Cayley
graph for Z9
Z3 with respect to the generating set {a, c = ba−1 }.

Example 8.6. The group Z2 = Z ⊕ Z admits an automorphism α of

order 2, given by α [(a, b)] = (b, a). Thus we may form a semi-direct
product Z2
Z2 . We will return to this seemingly simple example in
Section 9.4.

Example 8.7. The semi-direct product H
Aut(H) is called the holo-

morph of H. In some vague sense, best not to be made precise, the holo-
morph of H is the largest meaningful semi-direct product one can make
starting from H. As an example, we note that Hol(Z2 ⊕ Z2 ) ≈ Sym4 .
(This isomorphism is not obvious.)
156 The Lamplighter Group

The “wreath product” of two groups G and H is defined using semi-

direct products. One starts by forming a direct sum of copies of G, one
copy for each h ∈ H. In fact, one often indexes the copies of G by the


elements of H, hence this direct sum can be expressed as h∈H G. The


wreath product – denoted G  H – is the semi-direct product of h∈H G
and H:
 

GH = G
H.
h∈H

The action of H on the direct sum defined is via the standard Cayley


action of H on itself. Thus, given g ∈ h∈H G, the element h ∈ H
permutes the entries of g by taking the entry in position h to the position
h · h (for every h ∈ H).
As an elementary example, consider Z2 Z3 . This group is a semi-direct
product (Z2 )3
Z3 where the associated automorphism
φ : Z3 → Aut(Z2 ⊕ Z2 ⊕ Z2 )
is given by declaring that φ(1) ∈ Aut(Z2 ⊕ Z2 ⊕ Z2 ) is the cyclic permu-
tation α(a, b, c) = (b, c, a).
Exercise 8.8. Verify the following equation:
[(0, 1, 1), 1] · [(1, 0, 0), 1] = [(0, 1, 0), 2] ∈ Z2  Z3
The remainder of this chapter is devoted to understanding a single
example of this wreath product construction, Z2  Z. This group is a
semi-direct product:
 

Z2  Z = Z2
Z = (· · · ⊕ Z2 ⊕ Z2 ⊕ Z2 ⊕ · · · )
Z.
i∈Z

The infinite direct sum has as its elements Z-tuples of elements of Z2 ,

only finitely many of which are allowed to be non-zero.1
This particular wreath product of groups is called the lamplighter
group, for reasons that will become clear, and it is denoted L2 . For


convenience, call the subgroup of L2 given by i∈Z Z2 “A.” Since an
element of A is determined by which (finite number) entries have non-
zero elements, we can represent the elements of A as finite subsets of
1 The reader may have encountered the phrases “direct product” and “direct sum”
in various contexts and might believe they are synonymous. They are not. In a
direct sum, often denoted ⊕, only finitely many entries are allowed to be non-
identity elements; in a direct product, often denoted ×, there is no restriction on
the number of non-identity elements.
 The Lamplighter Group 157

integers. The empty set would then correspond to the identity element,
0 = (. . . , 0, 0, 0, . . .) ∈ A. The binary operation in A is coordinate-wise
addition in Z2 . Viewing elements of A as subsets of Z, we see that this
corresponds to the symmetric difference
S T = {n ∈ Z | n ∈ S or n ∈ T but n
∈ S ∩ T }.
In particular, {−2, 0, 1} {−3, 0, 4} = {−3, −2, 1, 4}. Thus every element
in L2 can be represented by an ordered pair [S, n] where S is a subset
of integers, and n ∈ Z. The binary operation is given by
[S, n] · [T, m] = [S (T + n), n + m]
where T + n = {t + n | t ∈ T }. The identity element is [∅, 0].
As the subgroup A < L2 is an infinite direct sum, the following lemma
may at first seem a bit surprising.
Lemma 8.9. The lamplighter group L2 can be generated by two ele-
ments: one of order 2 and the other of infinite order.
Proof. Let t be the element corresponding to [∅, 1] ∈ L2 and let a be the
element [{0}, 0] ∈ L2 . The action of Z on A can be described by noticing
that φ(1) ∈ Aut(A) has the effect of adding 1 to each coordinate. Thus
ta = [∅, 1] · [{0}, 0] = [∅ {1}, 1 + 0] = [{1}, 1]
and more generally
tn a = [∅, n] · [{0}, 0] = [{n}, n]
for any n ∈ Z. Thus
tn at−n = [{n}, 0].
It follows that an arbitrary element of the subgroup A can be expressed
as:
[{n1 , n2 , . . . , nm }, 0] = tn1 at−n1 · tn2 at−n2 · · · tnm at−nm .
Further, an arbitrary element of L2 can be expressed as:
[{n1 , n2 , . . . , nm }, k] = tn1 at−n1 · tn2 at−n2 · · · tnm at−nm · tk .
Therefore, the set {a, t} is a generating set for the lamplighter group.
The perspective of this book is geometric, and it is this perspec-
tive that illuminates L2 and explains why it is called the lamplighter
group. We know that an arbitrary element of L2 can be expressed as
[{n1 , n2 , . . . , nm }, k]. To think of this geometrically, take the Cayley
158 The Lamplighter Group

Fig. 8.3. A geometric representation of an element of the lamplighter group.

The central vertex corresponds to the identity element in Z, so this element
of L2 is [{−2, 0, 1, 2}, −1].

graph of Z and color the vertices corresponding to {n1 , n2 , . . . , nm } yel-

low; color all of the other vertices black; and add a pointer pointing to
the vertex associated to k. (An example is shown in Figure 8.3.) We will
refer to such a decorated graph as the picture of an element. The pic-
ture of the identity element consists of all black vertices with the pointer
pointing at zero.
If we want to study the Cayley graph with respect to the generating
set {a, t}, then we need to determine the effect of right multiplying by
an a or t. If g = [S, k] is an arbitrary element of L2 , then
 k]
g · a = [S, k] · a = [S, k] · [{0}, 0] = [S {k}, k + 0] = [S,
where S either adds k to S (if the vertex at position k was colored black)
or it removes k from S (if the vertex at position k was colored yellow).
Thus, in terms of the pictures of elements, right multiplying an arbitrary
element by a has the effect of changing the color of the vertex that the
pointer is pointing at. Right multiplication by t gives:
g · t = [S, k] · t = [S, k] · [∅, 1] = [S, k + 1]
which simply moves the pointer one unit to the right. Similarly, right
multiplication by t−1 moves the pointer one unit to the left.
We are now in a position to explain the intuition behind these pictures
of the elements of L2 . We have “lit lamps” at all the positions where
there are non-zero entries in the infinite direct sum of copies of Z2 , all
other lamps are “unlit,” and our lamplighter is stationed at position k.
Right multiplication by a has the effect of lighting or dowsing the lamp
that the lamplighter is stationed at; right multiplication by t or t−1
moves the lamplighter to the right or left. (See Walter Parry’s quotation
at the beginning of this chapter.)
The use of these pictures makes it simple to express an arbitrary
element in terms of the generators. Take, for instance the element of L2
shown in Figure 8.4. To create the element corresponding to the picture
in Figure 8.4, starting from the identity, we could:
 Exercises 159

Fig. 8.4. An element of L2 , which can be expressed as t3 at−2 at or tat2 at−1 .

1. Move the lamplighter three units to the right.

2. Light the lamp at x = 3.
3. Move the lamplighter two units to the left.
4. Light the lamp at x = 1.
5. Move the lamplighter one unit to the right.

This procedure shows that this element can be expressed as t3 at−2 at. Of
course there are other ways one could express this particular element.
For example, one could follow a slightly different process:

1. Move the lamplighter one unit to the right.

2. Light the lamp at x = 1.
3. Move the lamplighter two units to the right.
4. Light the lamp at x = 3.
5. Move the lamplighter one unit to the left.

From this we see that the same element can be expressed as tat2 at−1 .
In other words, t3 at−2 at = tat2 at−1 in L2 .
In Figure 8.5 we show a cycle in the Cayley graph of L2 (with respect
to {a, t}) that corresponds to the relation atat−1 = tat−1 a. We have
denoted the vertex corresponding to g ∈ L2 by its picture.

Exercises
(1) Prove Theorem 8.3.
(2) Prove that the symmetry group of a cube (Section 1.6.3) is isomor-
phic to (Z2 ⊕ Z2 ⊕ Z2 )
Sym3 .
(3) A finite subgroup of Aut(Fn ), Ωn , was introduced in Section 3.10.
Prove that Ωn ≈ (Z2 ⊕ · · · ⊕ Z2 )
Symn . (What does this have to
 
n copies
do with Exercise 2?)
(4) Prove that Symn ≈ An
Z2 .
(5) Prove Hol(Z2 ⊕ Z2 ) ≈ Sym4 .
(6) Verify that Figure 8.2 is correct.
160 The Lamplighter Group

Fig. 8.5. A cycle in the Cayley graph of the lamplighter group, where
undirected edges correspond to the generator a and directed edges to the
generator t.

(7) There are five distinct groups of order 12. You have probably en-
countered four of them: Z12 ≈ Z3 ⊕Z4 , Z6 ⊕Z2 , D6 ≈ D3 ⊕Z2 , and
A4 . The fifth group of order 12 can be described using semi-direct
products. Let H = Z3 and let K = Z4 , where H is generated by
h and K is generated by k. Let φ : K → Aut(Z3 ) by φ(k) = α
where α(h) = h2 . Show that G = H
φ K is a group of order 12
that is not isomorphic to any of the four previously listed groups of
order 12.
(8) Prove that D4 ≈ Z2  Z2 .
(9) Let G and H be two finite groups. Show that the order of G  H is
|G||H| · |H|.
(10) In this exercise we examine the symmetries of rooted binary trees.
(See Chapter 6.)
a. Let T denote the finite, rooted binary tree with eight leaves
shown in Figure 8.6. Prove that Sym(T ) ≈ (Z2  Z2 )  Z2 .
b. Let T denote the infinite, rooted binary tree where every
vertex has two descendants. Prove that Sym(T ) ≈ Sym(T )
Z2 .
 Exercises 161

Fig. 8.6. A finite, rooted binary tree.

(11) Prove that the wreath product of two finitely generated groups is
finitely generated.
(12) Show that atn at−n = tn at−n a ∈ L2 , for any n ∈ N.
(13) The Cayley graph Γ of L2 with respect to {a, t} is fairly compli-
cated. You can gain some intuition for this by drawing the ball
of radius 3, B(e, 3) ⊂ Γ. You should find that this is a tree, but
you should also realize that B(e, 4) is not a tree, as is shown by
Figure 8.5.
 9
The Geometry of Infinite Groups

[E]ssentially all geometric constructs which are global in nature, such as paths
of shortest length, global manifestations of curvature, planes, half-spaces, rates
of growth, which are studied in differential geometry have manifestations in
combinatorial approximations to that geometry.
. –James Cannon

9.1 Gromov’s Corollary, aka the Word Metric

We have seen a number of examples of groups acting on the real line.
Sometimes these actions preserve the distance between points on the
line, such as the action of D∞ . Other actions we have considered do not
preserve distances, for example, the action BS(1, 2)
R presented in
Chapter 4. One of the most powerful insights in the study of finitely gen-
erated infinite groups is that they can always be viewed as groups acting
in a distance-preserving way on a geometric object. We refer to this in-
sight as Gromov’s Corollary to Cayley’s Better Theorem, in honor of a
groundbreaking paper that Mikhail Gromov wrote in the 1980s [Gr87],
which highlighted this perspective, introduced a number of questions
motivated by the geometry of infinite groups, and introduced powerful
tools that can be used to answer them.
In order to present Gromov’s Corollary, we need to introduce a rea-
sonably flexible notion of “geometric object” as well as formally define
what we mean by saying that a group action “preserves distances.”

Definition 9.1. A metric space consists of a set X and a distance

function d : X × X → R such that, for any x, y, and z ∈ X:

1. d(x, y) ≥ 0,
2. d(x, y) = 0 if and only if x = y,

162
 9.1 Gromov’s Corollary, aka the Word Metric 163

3. d(x, y) = d(y, x),

4. d(x, y) + d(y, z) ≥ d(x, z).
This final condition is referred to as the triangle inequality. Examples of
metric spaces include the real line as well as Rn with their usual notion
of distance. It also includes objects such as spheres, where the distance
between two points is declared to be the length of the shortest path
connecting the two points.
A function from one metric space to another, φ : X1 → X2 , is an
isometry if it is onto and, for all x, y ∈ X, d1 (x, y) = d2 (φ(x), φ(y)).
(In other words, the function preserves (“iso”) distances (“metry”).) If
G
X, then the action is by isometries, or is an isometric action, if for
all x, y ∈ X and g ∈ G, one has:
d(x, y) = d(g · x, g · y).
The three reflection groups W244 , W236 and W333 actions on R2 (dis-
cussed in Chapter 2) are isometric actions.
Theorem 9.2 (Gromov’s Corollary). Every finitely generated group can
be faithfully represented as a group of isometries of a metric space.
Similar to the proofs of Cayley’s Theorem and Cayley’s Better Theo-
rem, the metric space is built from the group G. Because the vertices of
a Cayley graph correspond to group elements, one can use the distance
between vertices on a Cayley graph to define a distance function on the
set of elements of the group. Further, one can phrase this entirely in
terms of group elements, with no direct mention of the Cayley graph. If
S is a finite set of generators for G, and g, h ∈ G, then set
dS (g, h) = the length of the shortest word representing g −1 h .
If ω is a word in {S ∪ S −1 }∗ representing g −1 h then
g −1 h = π(ω) ⇒ h = gπ(ω).
So ω labels a path connecting the vertex associated to g to the vertex
associated to h. Thus a minimal-length word describes a minimal-length
path between the associated vertices in the Cayley graph ΓG,S .
If S is a generating set for G, then we have previously discussed the
length of words in {S ∪ S −1 }∗ . We can extend this to the elements
of G.
Definition 9.3. Let G be a group with a fixed, finite generating set S.
Then the length of g ∈ G is the minimal-length of a word ω ∈ {S ∪S −1 }∗
164 The Geometry of Infinite Groups

where π(ω) = g. Denote this value by |g|. Notice, in particular, that |g|
represents the minimal length of a path from the vertex representing e
to the vertex representing g in the Cayley graph ΓG,S .
As a simple example, if G = Z⊕Z, S = {(1, 0), (0, 1)} is the generating
set, and g = (m, n) ∈ Z ⊕ Z, then the length of G is just |m| + |n|.
This metric on the group G is often referred to as the word metric
and it does depend on the choice of generators. This is the reason that
S appears as a subscript on the distance function. We reserve the right to
be sloppy and drop this subscript from time to time, when the fixed set
of generators is either clear or not of central importance to the argument.
Proof of Gromov’s Corollary. Conditions 1 and 2 have already been es-
tablished, so it only remains to show that the distance function is sym-
metric and that the triangle inequality holds.
The function dS is symmetric: if dS (g, h) = n then there is a word
ω of length n such that g −1 h = π(ω). But then h−1 g = π(ω −1 ) and
so dS (h, g) is at most n. But if there were a shorter word representing
h−1 g, then we could take its formal inverse and form a shorter word
representing g −1 h. Thus dS (h, g) = dS (g, h) for any g, h ∈ G.
Our function dS also satisfies the triangle inequality:
dS (g, k) + dS (k, h) ≥ dS (g, h)
for all g, h and k ∈ G. To see this, let ωgk and ωkh be minimal-length
words such that g · ωgk = k and k · ωkh = h. Thus g · π(ωgk ωkh ) = h,
hence
dS (g, h) ≤ |ωgk | + |ωkh | = dS (g, k) + dS (k, h).
The comments above show that the group G can be viewed as a metric
space. Cayley’s Theorem shows that left multiplication gives an action
of G on itself, and this action preserves distances. That is,
dS (h, k) = |h−1 k| = |h−1 g −1 gk| = dS (gh, gk)
for any g, h, k ∈ G.

Example 9.4. A standard Cayley graph of Z⊕Z is shown in Figure 9.1.

The distance between the lower left-hand vertex and the
upper
 right-
hand vertex is 7. This distance is realized by a total of 74 different
words of length 7. To see this, let the generator corresponding to hor-
izontal arrows be x and the generator corresponding to vertical arrows
be y. Then any word corresponding to a minimal-length path between
 9.1 Gromov’s Corollary, aka the Word Metric 165

Fig. 9.1. The distance between the lower left-hand vertex and the upper right-
hand vertex is 7.

these vertices consists of seven letters: four x’s and three y’s. Such a
word is determined by the placement of the x’s (or by the placement
of the y’s). For example, the edge path going across the bottom of the
figure and then the right edge corresponds to xxxxyyy while a diago-
nal path is given by xyxyxyx. In fact, the union of all edges that occur
in minimal-length paths between these two vertices forms the rectangle
that is the focus of this figure.

Definition 9.5. A minimal-length edge path joining two vertices v and

w is a geodesic path. Similarly, a word ω is a geodesic 1 if |ω| = |π(ω)|.
In other words, a word is a geodesic if the corresponding path in the
Cayley graph is a minimal-length path between its endpoints.
 In the
example above we have essentially pointed out there are 74 geodesic
words expressing the group element (4, 3) ∈ Z ⊕ Z.

Remark 9.6. The Word Metric is primarily studied in the context of

infinite groups, but it does show up from time to time in the context
of finite groups. For example, the diameter of any finite graph Γ is the
minimal integer D such that one can get between any two vertices in Γ
1 In geometry, a path is a geodesic if it is (locally) of minimal-length. Examples
include line segments in the Euclidean plane and arcs of great circles on spheres.
166 The Geometry of Infinite Groups

by an edge path of length ≤ D. If Γ happens to be a Cayley graph of a

group G, the diameter represents the maximum length that is necessary
to express any element in G in terms of the chosen generators.
There is a famous example where the diameter of a Cayley graph is of
popular interest. Let R be the Rubik’s Cube. The set of all symmetries
of R forms a very large, but still finite, group. The standard move in
working with a Rubik’s Cube is to rotate one of the six faces by 90◦ .
These standard moves generate all possible moves, and so they form
a set of generators for the Rubik’s Cube group. The diameter of the
Cayley graph of Sym(R) with respect to this generating set gives the
maximum number of steps necessary to get from an arbitrary position
to the solution position using the standard moves. Since the Rubik’s
Cube group has over 43 quintillion elements, it is not a surprise that
this diameter is not yet known. (Our most recent Google search shows
that the diameter is known to be ≤ 26.)

Finding geodesic words is not always as easy as it is in Z ⊕ Z. The

lamplighter group L2 gives us a more illuminating example of some of
the difficulties one can encounter. Before establishing a general formula
for the length of an arbitrary element, we should consider the element
g ∈ L2 corresponding the picture in Figure 8.3. In order to construct
such a configuration of lit lamps and pointer, the lamplighter must light
the four indicated lamps, and she must end her travels at x = −1.
The most efficient way to do this is to have the lamplighter travel to
x = 2, then move back to x = −2, and finish at x = −1, remembering
to turn on the appropriate lamps when in position. Thus any minimal-
length expression for g must involve 2 + 4 + 1 + 4 = 11 letters from
{a, t, t−1 }. This discussion indicates why we need to specify the following
two constants associated to an arbitrary element g = [S, k] ∈ L2 .

R = Max{S ∪ {0}},
L = min{S ∪ {0}}.

Proposition 9.7. Let g = [S, k] ∈ L2 , and let R and L be as above.

Then the length of g is given by

|g| = |S| + min {2R + |L| + |k − L|, 2|L| + R + |k − R|}.

In order to establish this formula, we introduce two normal forms

for L2 .
 9.1 Gromov’s Corollary, aka the Word Metric 167

The right-first normal form: Every element of L2 can be uniquely

expressed by first moving the lamplighter to the position of the rightmost
lit lamp, then moving to the leftmost lit lamp, lighting lamps each time
the lamplighter is first in that position. Finally, move the lamplighter
left or right, if necessary, to get the lamplighter into position indicated
by the pointer.
As an example, consider the element g = [{−2, 0, 1, 2}, −1] ∈ L2 ,
which is shown in Figure 8.3. The right-first normal form expression for
this element is:
g = atatat−4 at.

The left-first normal form: This is identical to the normal form

above, except that one first moves to the left, and so on. If we again
consider g = [{−2, 0, 1, 2}, −1] ∈ L2 we see that the left-first normal
form expression for this element is:
g = at−2 at3 atat−3 .
Since we already know |g| = 11, it follows that the right-first normal form
for g is a geodesic word, while the left-first normal form for g is not.

The descriptions above are not completely clear when applied to ele-
ments like g = [{1, 3}, 2] ∈ L2 , which is shown in Figure 8.4. The picture
for this element contains no lit lamps to the left of zero. The right-first
normal form for such an element simply moves the lamplighter right,
lighting lamps along the way, and then moves the pointer to the correct
ending position. The left-first normal form is the exact same.

Exercise 9.8. Find the right-first and left-first normal forms of the
element g = [{−3, 1, 2}, 4] ∈ L2 .

Proof of Proposition 9.7. We first establish connections between the right-

first and left-first normal forms for g ∈ L2 and the expressions in the
statement of the proposition.
Case 1: Assume there is a lit lamp at some m < 0, in which case
L < 0. The right-first normal form will have |S| occurrences of a (for
lighting the lamps). There will also be 2R + |L| + |k − L| occurrences of
t or t−1 , since the lamplighter will move from 0 to R, then back to L
and from there to k (the location of the pointer).
Case 2: Assume there are no lit lamps at negative integers, so L = 0.
The right-first normal form will again have |S| occurrences of a. The
lamplighter will need to take R steps to the right, and then will need to
168 The Geometry of Infinite Groups

move |k − R| steps to end at the location of the pointer. So the length

of the right-first normal form is
|S| + R + |k − R| = |S| + 2|L| + R + |k − R|
since L = 0.
Case 3: Assume there is a lit lamp at some n > 0, in which case
R > 0. Then, as in Case 1, the length of the left-first normal form is
|S| + 2|L| + R + |k − R|.
Case 4: Assume there is no lamp lit at a positive integer, so R = 0.
Then, as in Case 2, the length of the left-first normal form is
|S| + |L| + |k − L| = |S| + 2R + |L| + |k − L|.
Thus, by considering the right-first and left-first normal forms of g we
see that
|g| ≤ |S| + min {2R + |L| + |k − L|, 2|L| + R + |k − R|} .
It remains to be shown that this is actually an equality (which will imply
that for any g ∈ L2 , either its left-first or right-first normal form is a
geodesic). But if g = [S, k] then, in any expression for g, there must
be |S| occurrences of the generator a. Further, the lamplighter must
travel from 0 to every lit lamp, and then move to the final position of
the pointer. By considering cases similar to the ones above, it can be
shown that min {2R + |L| + |k − L|, 2|L| + R + |k − R|} is the minimal
number of steps that needed. As each step the lamplighter takes requires
an application of either t or t−1 , we get
|g| ≥ |S| + min {2R + |L| + |k − L|, 2|L| + R + |k − R|} .
Hence the equality stated in the theorem holds.

9.2 The Growth of Groups, I

In Section 5.3 we introduced the idea of spheres and balls of radius n
in Cayley graphs. For a fixed generating set S, the associated distance
function on the group G is the same as the edge-path distance between
vertices in the Cayley graph ΓG,S . Thus the following re-definition is
consistent with the earlier introduction of spheres.
Definition 9.9. Given a group element g ∈ G and a positive integer n,
the sphere of radius n is
S(g, n) = {h ∈ G | dS (g, h) = n}.
 9.2 The Growth of Groups, I 169

The ball of radius n is

B(g, n) = {h ∈ G | dS (g, h) ≤ n}.

Notice that there is a difference between balls of radius n in a Cayley

graph and balls of radius n in the group. The ball of radius n, based at
the vertex vg associated to g in the Cayley graph ΓG,S , has vertex set
{vh | h ∈ B(g, n)}. However, the ball of radius n in the Cayley graph
also contains edges connecting these vertices to vg .

The size of spheres and balls of radius n in a Cayley graph are of

considerable importance, as they give us a concrete way to talk about
the growth of a given infinite group.

Definition 9.10. The spherical growth function of a group G with re-

spect to a finite generating set S is the function σ : N → N defined by
σ(n) = |S(e, n)|. The associated growth series is the formal power series
 
S(z) = σ(n)z n = z |g| .
n≥0 g∈G

Example 9.11. Let G = Z with a single generator t. Then


1 n=0
σ(n) =
2 n>0
and so the associated growth series is

S(z) = 1 + 2z + 2z 2 + 2z 3 + · · · .

In studying formal power series, combinatorialists often ask if a given

series is rational, meaning it is the power series associated to a rational
function. In the situation above, it is elementary to exhibit a rational
function corresponding to the growth series:

S(z) = 1 + 2z + 2z 2 + 2z 3 + · · ·
1
= 2(1 + z + z 2 + z 3 + · · · ) − 1 = 2 −1
1−z
2 1−z 1+z
= − = .
1−z 1−z 1−z
If Z is generated by {2, 3} then one would expect a fairly different
growth function and associated series. (The Cayley graphs for these
two generating sets are shown in Figure 1.20.) Let S{2,3} (n) be the
sphere of radius n, with respect to {2, 3}, and let σ{2,3} (n) be the size
of S{2,3} (n). By inspection we find that σ{2,3} (0) = 1 and σ{2,3} (1) = 4;
170 The Geometry of Infinite Groups

a short computation shows S{2,3} (2) = {±1, ±4, ±5, ±6} and therefore
σ{2,3} (2) = 8. After this, the pattern becomes simple. Because 2 and 3
are relatively prime, any integer can be expressed as n = l·2+m·3 where
l and m are integers. But is this a geodesic expression? That is, is |l|+|m|
as small as it can be in order to produce n? Not if |l| ≥ 3, for then the
substitution 3 · 2 = 2 · 3 would reduce the length (where our generators
have been underlined). Thus a minimal-length word representing n ∈ Z
contains at most two instances of 2. So the sphere of radius n contains
at most ten elements: ±n · 3, ±(n − 1) · 3 ± 2 and ±(n − 2) · 3 ± 2 · 2.
However, for n ≥ 3, (n − 1) · 3 − 2 = (n − 3) · 3 + 2 · 2, so (n − 1) · 3 − 2 is
actually in S{2,3} (n − 1). Further, (n − 2) · 3 − 2 · 2 = (n − 4) · 3 + 2, so
(n − 2) · 3 − 2 · 2 ∈ S{2,3} (n − 1). So for n ≥ 3, the sphere of radius n is

S{2,3} (n) = {±n · 3, ± [(n − 1) · 3 + 2] , ± [(n − 2) · 3 + 2 · 2]}

and therefore


 1 n=0

4 n=1
σ{2,3} (n) =

 8 n=2

6 n ≥ 3.

A similar computation as before shows that the growth series with

respect to this generating set is also rational:
1 + 3z + 4z 2 − 2z 3
S{2,3} (z) = 1 + 4z + 8z 2 + 6z 3 + 6z 4 + · · · = .
1−z
We see that the growth series with respect to both generating sets are
rational and they have the same denominator. The following exercise
asks you to add a bit of evidence in support of the claim that, in the
case of an infinite cyclic group, the growth series with respect to any
generating set is rational with denominator 1 − z.

Exercise 9.12. Let S be a pair of relatively prime integers, so that S

forms a finite generatoring set for Z. Show that the associated growth
series is rational, with denominator 1 − z.

Example 9.13. Let W333 be the group generated by three reflections in

the sides of an equilateral triangle, discussed in Chapter 2. The Cayley
graph embeds in the plane as the edges of the triangulation by regular
hexagons. Thus it suffices to compute how many vertices are a distance
n from a fixed base vertex in this graph. Or one can count the number
of triangles separated from a base triangle by at most n edges in the
 9.2 The Growth of Groups, I 171

tiling by equilateral triangles. In either approach one quickly comes to

the conjecture that

1 n=0
σ(n) =
3 · n n > 0.

This formula is correct, but it takes some care to actually prove it. One
can also show that the corresponding growth series is the formal power
series associated to the rational function
1 + z + z2
S(z) = .
(1 − z)2

(Exercises 5 and 6 ask you to verify these claims.)

Below are two theorems that compute growth series of products of

groups, assuming you know the growth series of the individual factors.

Theorem 9.14. Let G and H be groups with finite generating sets

SG and SH , and corresponding growth series SG (z) and SH (z). Then
the set
SG⊕H = {(s, eH ) | s ∈ SG } ∪ {(eG , h) | h ∈ SH }

is a generating set for G ⊕ H and the corresponding growth series for

G ⊕ H is given by

SG⊕H (z) = SG (z) · SH (z).

Proof. The claim about generating sets is clear.

The length of (g, h) ∈ G ⊕ H is the sum of the lengths of g and h. So


n
σG⊕H (n) = σG (i)σH (n − i),
i=0

from which the statement follows.

An induction argument establishes the following result:

Corollary 9.15. The growth series for Zn with respect to a standard

generating set is

n
1+z
SZn (z) = .
1−z
172 The Geometry of Infinite Groups

Theorem 9.16. Let G and H be groups with finite generating sets SG

and SH , and corresponding growth series SG (z) and SH (z). Take as
generating set for G ∗ H the image of SG and SH in G ∗ H. Then
1 1 1
= + − 1.
SG∗H (z) SG (z) SH (z)
Proof. Every element in G ∗ H can be expressed as

g0 · h1 g1 h2 g2 · · · hn−1 gn−1 hn gn ·hn+1

  
none equal to the identity

where g0 and/or hn+1 may be equal to the identity. It follows that

 n
SG∗H (z) = SG (z) [(SH (z) − 1)(SG (z) − 1)] SH (z)
n≥0
SG (z) · SH (z)
=
1 − (SH (z) − 1)(SG (z) − 1)
where the last equality is established using the standard formula for
geometric series. A bit of algebra converts this formula into the one
stated in the theorem.

Example 9.17. Pick cyclic generators for Z3 and Z4 and the corre-
sponding growth “series” for these finite groups are 1+2z and 1+2z+z 2 .
Applying Theorem 9.16 we see that the growth series for Z3 ∗ Z4 is
(1 + 2z)(1 + 2z + z 2 ) 1 + 4z + 5z 2 + 2z 3
SZ3 ∗Z4 (z) = =
1 − (2z)(2z + z 2 ) 1 − 4z 2 − 2z 3
= 1 + 4z + 9z 2 + 20z 3 + 44z 4 + 98z 5 + 216z 6 + · · · .
Remark 9.18. Not all growth series of finitely generated groups are
rational. For example, the growth series of Z2 F2 , a “lamplighter” group
based on lighting lamps on the Cayley graph of F2 instead of Z, is
not rational. (See [Pa92] for the details on this and related examples.)
The finitely generated, infinite torsion group U, which we constructed in
Chapter 6, is another example. Perhaps a bit more worrisome is the fact
that the growth series can be rational with respect to one generating set
but not rational with respect to another [St96].

9.3 Growth and Regular Languages

There is a connection between the idea of growth series and groups which
admit regular normal forms.
 9.3 Growth and Regular Languages 173

Theorem 9.19. Let G be a group with S a finite set of generators. If

(G, S) admits a regular normal form that is composed of geodesic words,
then the growth series for G with respect to S is rational.

The proof of this theorem uses the “transfer-matrix method,” a tool

of broad utility in combinatorial arguments. In fact, this result really
has little to do with groups; it is a theorem about the growth series of
regular languages.
Let G be a finite directed graph. The adjacency matrix A has as its
(i, j)-entry the number of edges connecting vertex i to vertex j in G.
Let A(n) be the matrix whose (i, j)-entry is the number of directed edge
paths of length exactly n, which connect vertex i to vertex j.

Lemma 9.20. The matrix A(n) equals An .

Proof. This is what matrix multiplication is all about! (If that doesn’t
make you happy, prove it by induction.)

Lemma 9.21. Let Sij (z) = n≥0 Aij (n)z n . Then Sij (z) is the series
corresponding to the rational function
(−1)i+j det (I − zA : j, i)
Sij (z) =
det (I − zA)
where (I − zA : j, i) denotes the matrix formed by subtracting zA from
the identity matrix, and then removing the jth row and ith column.

Proof. The series Sij (z) is the (i, j)-entry of

  −1
An z n = (zA)n = (I − zA) .
n≥0 n≥0

Thus the formula stated is simply the standard formula for the inverse
of a matrix.

Proof of Theorem 9.19. Let M be a finite automaton, with no -edges,

whose language is the regular language of normal forms for G. Let A be
the associated adjacency matrix. Then, by the previous two lemmas, we
get that the rational function associated to the growth series is given by
−1
S · (I − zA) ·A

where: S is a 1 × |V (M)| matrix with 1s in each spot associated to a

start state, and zeros elsewhere; and A is a |V (M)| × 1 matrix with 1s
corresponding to accept states, and zeros elsewhere.
174 The Geometry of Infinite Groups

Example 9.22. The collection of freely reduced words is a normal

form for F2 . This language is also regular and geodesic. An automaton
whose language consists of the freely reduced words in {x, y, x−1 , y −1 }∗
is shown in Figure 7.5. This machine has five states, consisting of a start
state, and then four states corresponding to the last letter read. The
associated adjacency matrix, with the states organized as indicated, is:
 
Start state 0 1 1 1 1
Last read x  0 1 1 0 1 
 
Last read y  0 1 1 1 0 .
 
Last read x−1  0 0 1 1 1 
Last read y −1 0 1 0 1 1

The formula given in the proof of Theorem 9.19 is then

 −1  
1 −z −z −z −z 1
 0 1−z −z 0 −z   1 

   
1 0 0 0 0  0 −z 1−z −z 0   1 .
   
 0 0 −z 1−z −z   1 
0 −z 0 −z 1−z 1

This may look daunting, but a program such as Mathematica can do the
computation almost instantly, giving the answer:
1+z
= 1 + 4z + 12z 2 + 36z 3 + 108z 4 + · · · .
1 − 3z
Since σ(n) = 4 · 3n−1 for F2 with respect to a basis, the reader should
be able to independently verify that this formula is indeed correct.

The method described above involves one step that is computationally

intensive, namely, the inversion of a matrix with polynomial entries.
While Mathematica can handle the computation in the case considered
in Example 9.22, in more complicated situations such programs have
great difficulty computing inverses. One can, however, avoid this step.
The inverse of I − zA can be expressed as

−1 1
(I − zA) = M
det [I − zA]
where M is a matrix with polynomial entries. It follows that

−1 p(z)
S · (I − zA) ·A=
det [I − zA]
 9.4 Cannon Pairs 175

where p(z) is some polynomial. Further, by appealing to the formula in

Lemma 9.21, we see that the degree of p(z) must be no greater than
(n − 1) if A is an n-by-n matrix.
Revisiting Example 9.22, we see that the determinant of I −zA is (z −
1)2 (1 + z)(1 − 3z). Thus we know the growth series can be
expressed as

a + bz + cz 2 + dz 3 + ez 4
= a + (4a + b)z
(z − 1)2 (1 + z)(1 − 3z)
+ (14a + 4b + c)z 2 + (44a + 14b + 4c + d)z 3
+ (135a + 44b + 14c + 4d + e)z 4 + · · · ,

where we have used Mathematica to produce the expansion. Since we

know the constant term in the growth series is 1, we have a = 1. The
coefficient of z is just the number of generators (and their inverses), so
4a + b = 4, hence b = 0. Continuing in this manner we can quickly
determine the numerator, and hence the rational function.

Remark 9.23. It is known that the growth series for BS(1, 2), with
respect to the standard generators a and b, is rational, but where the
rational function has degree 14. That the growth series is rational was
established by Brazil in [Bz94] and independently by Collins, Edjvet
and Gill [CEG94].2 However, it has also been proved that BS(1, 2) does
not admit a regular normal form consisting of geodesic words [Gv96]. In
particular, it is possible to have a rational growth series without it being
computed via the method described in this section.
All three of the papers cited above are accessible to anyone who has
read this far in this book, and are recommended for further study.

Finally we should point out that connections between numerical se-

quences, formal power series and rational functions has been well-studied
by combinatorialists. A good introduction to the topic of “generating
functions” is [Wi06].

9.4 Cannon Pairs

Fix a group G and a finite generating set S. Let γ(n) be the number
of geodesic words of length n, and let γ(g) be the number of geodesic
2 The formula for the rational function mentioned above was not computed by Brazil
as it would be “complicated and tedious”; Collins, Edjvet and Gill arrived at the
result independently and worked through the details to arrive at a final formula.
176 The Geometry of Infinite Groups

expressions for the element g ∈ G. The geodesic growth series is

 
Γ(z) = γ(n)z n = γ(g)z |g| .
n≥0 g∈G

Example 9.24. Consider Z ⊕ Z with generators {(1, 0), (0, 1)}. If n and
m are two positive
integers,
 then the number of geodesics expressing the
element (m, n) is n+mn . The number of geodesics of length n that occur
in the first quadrant is then
n

n n n n n
= + + ··· + + = 2n .
i=0
i 0 1 n − 1 n

Adding the four quadrants, and subtracting 4 for the elements on the
axes that have been counted twice shows γ(n) = 4(2n − 1) for n ≥ 1.
Thus
Γ(z) = 1 + 4z + 12z 2 + 28z 3 + 60z 4 + · · ·
= 1 + [8z + 16z 2 + 32z 3 + · · · ] − [4z + 4z 2 + 4z 3 + · · · ]
   
2z z 1 + z + 2z 2
=1+4 −4 =
1 − 2z 1−z (1 − z)(1 − 2z)
This example indicates that the geodesic growth series of a group,
much like the ordinary growth series of a group, may be rational. An
argument similar to that given in Section 9.3, using the transfer-matrix
method, establishes:
Theorem 9.25. Let (G, S) be a group paired with a finite generating
set. If the language of all geodesic words is a regular language, then the
geodesic growth series is rational.
In 1984, James Cannon published a foundational paper for the study
of infinite groups from a geometric perspective [Ca84]. In particular,
he imported the idea of cone types, which we have encountered in our
discussion of formal languages, to the study of geodesics in the Cayley
graphs of infinite groups.
Definition 9.26. Fix a group G and a finite generating set S, and let
Γ = ΓG,S be the associated Cayley graph. Then, for a given g ∈ G,

Cone(g) = The subgraph of Γ induced by

{vh | vg is on a geodesic from ve to vh }.
The elements g and h ∈ G are said to have the same cone type if there
is an edge-label preserving isomorphism between the directed graphs
 9.4 Cannon Pairs 177

Cone(g) and Cone(h). This is an equivalence relation, and if there are

only finitely many equivalence classes then G is said to have finitely
many cone types.

Fig. 9.2. The lower-left vertex is the vertex associated to (0, 0) ∈ Z ⊕ Z while
the highlighted vertex is associated to (2, 1) ∈ Z ⊕ Z. The cone of this element
is the highlighted subgraph, which looks much like a quadrant in the Euclidean
plane.

Example 9.27. Consider Z ⊕ Z with the generators {(1, 0), (0, 1)}. The
cone of the element (2, 1) consists of the subgraph of the Cayley graph
that is induced by {(m, n) | m ≥ 2 and n ≥ 1}. This is illustrated in
Figure 9.2.

Theorem 9.28. Let G be a group with finite generating set S. Then

(G, S) has finitely many cone types if and only if the language of all
geodesic words is regular.

Proof. In Section 7.2 we defined “cone type” in the context of languages:

Let L be a language in S ∗ and let ω ∈ S ∗ . The cone type of ω is the set of all
words ω  ∈ S ∗ such that ω · ω  ∈ L. Denote the cone type of ω by Cone(ω).
The collection
Cone(L) = {Cone(ω) | ω ∈ S ∗ } .

is the cone type of the language L, denoted Cone(L).

178 The Geometry of Infinite Groups

To avoid confusing the different uses of this phrase, we refer to the cone
types of L as L-Cone types.
Let L be the language of all geodesic words in {S ∪ S −1 }∗ . If L is
regular, then the Myhill–Nerode Theorem (7.18) states that there are
finitely many L-cone types. Given an element g ∈ G, the subgraph of
the Cayley graph Cone(g) is just the union of paths, based at g, corre-
sponding to words in L-Cone(ω) for all geodesic words ω with π(ω) = g.
As there are finitely many sets of L-Cones, there can only be finitely
many graphs arising as Cone(g) for g ∈ G.
Conversely, assume the Cayley graph ΓG,S has finitely many Cones.
Let M be the automaton whose vertex set corresponds to the finitely
many cone types, where the cone type of the identity is the only start
state, and every vertex is an accept state. For each generator s ∈ S add
a directed edge labelled s from Cone(g) to Cone(gs) when the vertex
associated to gs is in Cone(g). The language accepted by this machine
is L, hence L is regular.

Definition 9.29. If the pair (G, S) has finitely many cone types, or
equivalently if the language of all geodesics is regular, then (G, S) is a
Cannon pair.

In the remainder of this section we outline an example due to Jim

Cannon. Let x and y be the standard generators of Z ⊕ Z and let α ∈
Aut(Z ⊕ Z) be the automorphism induced by α(x) = y and α(y) = x.
The automorphism α has order 2, hence we can use it to construct a
semi-direct product G = (Z ⊕ Z)  Z2 .
For the remainder of this example we will slightly abuse notation,
and let x = [x, e], y = [y, e] and α = [e, α]. The Cayley graph of G
with respect to the generating set {x, y, α} consists of parallel grids.
The subgraph induced by {x, y} is the standard Cayley graph of Z ⊕ Z.
There are (unoriented) edges labelled α joining [xm y n , e] to [xm y n , α] for
all m, n ∈ Z. So it only remains to understand how right multiplication
by x and y influences [xm y n , α]. This is relatively straightforward, since

[xm y n , α] · [x, e] = [xm y n α(x), α · e] = [xm y n+1 , α],

and similarly

[xm y n , α] · [y, e] = [xm y n α(y), α · e] = [xm+1 y n , α].

Thus the subgraph containing [e, α], whose edges are all labelled x and
y, is a copy of the Z ⊕ Z Cayley graph with the roles of x and y reversed.
 9.5 Cannon’s Almost Convexity 179

Exercise 13 asks you to verify that this Cayley graph has finitely many
cone types.
We now switch our generating set. Define d = xy and z = x2 and
consider the set of elements S = {x, d, z, α} ∈ G. Since x and d are in S, 
y = x−1 d ∈ {S∪ S−1 }∗ . Thus S is a generating set for G. Exercise 14 asks
you to prove that, in this generating set, the word αz n αz m is geodesic
if and only if m < n. Thus the language of geodesics is not regular with
respect to this generating set by an application of the Pumping Lemma
(Theorem 7.14).
Because of examples like this, it is important to remember that having
the language of all geodesics be regular is a property of a pair consisting
of a group and a fixed generating set. It is not a property of the group
independent of the choice of generators.

9.5 Cannon’s Almost Convexity

Jim Cannon introduced the following property in 1987.

Definition 9.30 (Almost convex). A Cayley graph is almost convex

if there is a uniform constant K ≥ 0 such that if v, w ∈ SΓ (n), and
dΓ (v, w) ≤ 2, then dB(n) (v, w) ≤ K. In other words, a Cayley graph Γ
is almost convex if whenever two vertices on the sphere of radius n can
be joined by a path of length ≤ 2 in Γ then they can be joined by a
path of length ≤ K that stays inside the ball of radius n. (As this is a
property of the Cayley graph, the “balls” under consideration are balls
in the Cayley graph.)
A group is almost convex if it has a Cayley graph that is almost
convex.

Example 9.31. Free abelian groups are almost convex as are free groups
and free products of finite groups. Each of these claims is true regard-
less of the set of generators that is chosen, but these claims are easiest
to verify using “standard” generators. Showing this is true for any set
of generators is actually somewhat difficult. There are, in fact, groups
whose Cayley graphs are almost convex with respect to one set of gen-
erators, but not with respect to another.

Note: In the ball of radius n you can always join two vertices by a path
of length ≤ 2n. You simply move from one vertex to the identity, and
then back out to the other vertex. The power of the almost convexity
condition is that there is a single, uniform constant that works for all n.
180 The Geometry of Infinite Groups

It should be noted that if there is such a constant for vertices a distance

at most 2 apart, then there is also a uniform constant when the vertices
are a distance 3 apart, or 4 apart, or ...

Proposition 9.32. If Γ is almost convex, then given any m ≥ 2 there

is a fixed constant K(m) such that if v, w ∈ SΓ (n), and dΓ (v, w) ≤ m,
then dB(n) (v, w) ≤ K(m).

Proof. Let K = K(2). We argue by induction, the base case being our
hypothesis that Γ is almost convex. Assume then that v, w ∈ SΓ (n), and
dΓ (v, w) = m, where m ≥ 3. Let M be the maximum distance from the
vertex associated to the identity to a vertex on the chosen edge path
joining v to w. (Thus the chosen path is contained in BΓ (M + 1).) Note
that M must be less than n +
m+1 2 .
Let z be a vertex on this path where the distance from z to the vertex
associated to the identity is M . The neighboring vertices on this path,
call them z1 and z2 , are then either closer to the identity, or at least one of
them is the same distance from the identity as z. If both z1 and z2 are in
SΓ (M − 1) then they may be joined by a path inside of BΓ (M − 1) whose
length is at most K. This removes a vertex in SΓ (M ). In the other case,
z and z1 (or z2 ) can be joined by a path inside of BΓ (M ). This removes
an edge that was outside of BΓ (M ). In a worse-than-is-possible scenario,
this process would have to be applied m times, rerouting the original
path from its original vertices or edges, resulting in a path joining v to
w of length mK. This new path is contained in BΓ (M ).
Repeating this process results in an edge path of length less than
(mK)K = mK 2 , which is contained in BΓ (M − 1). Continuing this
process we eventually convert the original path to one which is contained
in BΓ (n). Since (M + 1) − n ≤ 1 +
m+1 2 , if we set l = 1 +
2  then
m+1

this process stops in at most l steps. The length of the path at the end
can be no more than mK l , hence we may take K(m) = mK l .

One reason to be interested in almost convex groups is:

Theorem 9.33. If G is almost convex, then it has a solvable word prob-

lem.

Proof. We will prove that if G is almost convex then there is an algo-

rithm that allows you to construct its (almost convex) Cayley graph Γ.
The argument is, however, non-constructive. In other words, we prove
 9.5 Cannon’s Almost Convexity 181

that the algorithm exists, but we do not actually provide it! The argu-
ment begins by mentioning that BΓ (K +2) exists, where K is the almost
convexity constant. Let C be any circuit in Γ of length ≤ K + 2, and
let vg be a vertex in this circuit. Then by multiplying by g −1 we may
translate this circuit into BΓ (K + 2). Thus all circuits of length ≤ K + 2
in Γ can be realized as translates of circuits in BΓ (K + 2).
Assuming you have built BΓ (n), you can make a list of “missing edges,”
that is, edges which are attached to vertices in SΓ (n) but which are
not contained in BΓ (n). (See the proof of Theorem 5.10.) We need to
determine how to attach these missing edges in order to form BΓ (n + 1).
If a missing edge joins two vertices in SΓ (n) then those vertices must be
joined by a path of length ≤ K in BΓ (n). Similarly, if two missing edges
meet at a vertex in SΓ (n + 1) then their associated vertices in SΓ (n)
are joined by a path of length ≤ K in BΓ (n). In either event, there is
then a circuit formed by the missing edge, or pair of edges, and an edge
path of length ≤ K. If there is such a circuit, a copy of it can be found
in BΓ (K + 2). If no such circuit appears in BΓ (K + 2) then the missing
edges go to distinct vertices in SΓ (n + 1).
Since there is an algorithm which constructs the Cayley graph Γ, then
by Theorem 5.10 there is an algorithm that solves the word problem.

As we mention below, not all groups are almost convex. There are
a number of weaker versions of almost convexity that have attracted
interest. The weakest (useful) variation is minimal almost convexity. A
Cayley graph Γ is minimally almost convex if, given v, w ∈ SΓ (n), where
dΓ (v, w) ≤ 2, then dB(n) (v, w) ≤ 2n − 1. (Remember, dB(n) (v, w) is
always less than or equal to 2n.) A finitely generated group is then
minimally almost convex if it has a minimally almost convex Cayley
graph.

Example 9.34. It was shown in 1998 that BS(1, 2) is not almost convex
with respect to any finite generating set [MiS98]. (Exercise 16 gives a
hint as to how to prove that the Cayley graph of BS(1, 2), with respect
to the generators given in Chapter 4, is not almost convex.) However, in
2005 it was shown that BS(1, 2) is minimally almost convex [EH05].

There are groups that are not minimally almost convex. For example,
Cleary and Taback proved that the lamplighter group L2 is not mini-
mally almost convex [CT05]. Their approach is to consider the elements
[{−n, n}, 1] and [{−n, n}, −1]. By Proposition 9.7, the length of both of
these elements is 4n + 1. They are both adjacent to [{−n, n}, 0], hence
182 The Geometry of Infinite Groups

the distance between them is 2. However, the intermediate element,

[{−n, n}, 0], has length 4n + 2, hence it is not in the ball of
radius 4n + 1.
In order to get from [{−n, n}, −1] to [{−n, n}, 1] the location of the
lamplighter must at some point be 0. If both lamps at −n and n are
lit, then having the lamplighter pass though 0 requires that you move
outside of the ball of radius 4n + 1. Thus to get from [{−n, n}, −1] to
[{−n, n}, 1], staying inside the ball of radius 4n + 1, one must first dowse
the lamp at −n. This requires a minimum of n steps. In the end, the
lamp at −n will need to be lit, but it cannot be lit until the lamp at n is
dowsed. Thus a minimum of 2n + 1 more steps need to be taken in order
to turn off the lamp at n. Because the lamplighter ends up to the right
of the origin, we need to light the lamp at −n before lighting the lamp
at n. So we again need a minimum of 2n + 1 steps to return to −n and
light the lamp. Finally, in order to return to the right side and light the
lamp at n an additional 2n + 1 steps are needed, and moving from n to
1 requires n − 1 steps. Thus, no matter how clever you are in choosing
your path, any path from [{−n, n}, −1] to [{−n, n}, 1] that is contained
in the ball of radius 4n + 1 must have length at least 8n + 2 = 2(4n + 1).
Thus we see the following result:

Proposition 9.35. The Cayley graph of the lamplighter group with re-
spect to the generators t = [∅, 1] and a = [{0}, 0] is not minimally almost
convex.

Exercises
(1) Let x generate a cyclic group of order 5. Show that the length of x2
and the length of x3 (with respect to {x, x−1 } are both equal to 2.
(2) Characterize the set of elements in L2 for whom the right-first
normal form is geodesic.
(3) Let G be a group with generating set S = S ∪S −1 . The correspond-
ing notion of length induces a function L : G → N given by L(g) =
|g|. These length functions are surprisingly easy to axiomatize. Let
 : G → N be a function. Prove that  is the same as L if and only if
a. (e) = 0;
b. |(gs) − (g)| ≤ 1 for all g ∈ G and s ∈ S; and
c. if g ∈ G \ {e}, then there is at least one s ∈ S such that
(gs) < (g).
 Exercises 183

(4) Find a formula for the size of S(n) and B(n) for Z ⊕ Z with respect
to the standard generating set (1, 0) and (0, 1). Then find a formula
for the size of S(n) and B(n) with respect to the larger generating
set {(1, 0), (0, 1), (1, 1)}.
(5) Let W333 be the group generated by three reflections in the sides of
an equilateral triangle, discussed in Chapter 2. Prove that σ(n) =
|S(n)| is given by

1 n=0
σ(n) =
3 · n n > 0.

(6) Prove that the growth series for W333 is the formal power series
associated to the rational function
1 + z + z2
S(z) = .
(1 − z)2
(7) Theorem 9.16 gives a formula for the growth series of a free product
of two groups. Find a similar formula for G ≈ A ∗ B ∗ C.
(8) The growth series for finitely generated free groups, with respect
to a basis, are always rational. In Section 9.2 it is shown that the
growth series for F1 is the series associated to (1+z)/(1−z); in Sec-
tion 9.3 we saw that the growth series for F2 is the series associated
to (1 + z)/(1 − 3z). What is the rational function corresponding to
the growth series for Fn ?
(9) Let {a, b} be the generating set for BS(1, 2) described in Chapter 4.
Show that the length of bn (n ≥ 1) is obtained by writing n in bi-
nary: If n = bk bk−1 · · · b2 b1 b0 (in binary), with bk = 0, then |bn | is
k
2k + i=0 bi . For example, 14 = 8 + 4 + 2 so, written in binary, 14
is 1110. The claim is that |b14 | is 2 · 3 + 1 + 1 + 1 = 9.

(10) If σ(n) = |S(e, n)|, then the associated series S(z) = n≥0 σ(n)z n
is sometimes referred to as the spherical growth series. One can also
consider the function β(n) = |B(e, n)| which gives the number of
vertices inside the ball of radius n, and the associated ball growth

series B(z) = n≥0 β(n)z n . Show that the spherical growth series
is rational if and only if the ball growth series is rational.
(11) Given any regular language L over an alphabet A, one can form
the associated growth series z |w| . The proof of Theorem 9.19
w∈L
establishes that, if L is regular, this growth series is rational. Show
that the converse is false using the language L of all palindromes
over the alphabet A = {a, b}.
184 The Geometry of Infinite Groups

(12) Let G = Z ⊕ Z and let S = {(1, 0), (0, 1)} be the standard set of
generators. Show that (G, S) has exactly nine cone types.
(13) Let G = (Z ⊕ Z)  Z2 and let S be the set of generators
S = {[(1, 0), e], [(0, 1), e], [(0, 0), α]}
(as discussed in Section 9.4). Prove that (G, S) has finitely many
cone types, that is, that (G, S) is a Cannon pair.
(14) Let G be as in Exercise 13 and define d = xy and z = x2 . Show
that in this generating set the word αz n αz m is geodesic if and only
if m < n.
(15) Let G and H be almost convex groups. Show that G⊕H and G∗H
are also almost convex groups.

Fig. 9.3. Two vertices in the sphere of radius 10 in the Cayley graph of BS(1, 2)
are highlighted. They are a distance 2 apart in the Cayley graph but are much
further apart inside the ball of radius 10.

(16) This exercise indicates how to prove that the Cayley graph of
BS(1, 2), with respect to the generators {a, b}, is not almost con-
vex. The reader may want to refer to the view of the Cayley graph
of BS(1, 2) shown in Figure 9.3 throughout this problem.
 Exercises 185

a. Show that ab4 a−1 and bab4 are both in the sphere of radius
6, and that the distance between them in the Cayley graph
is 2.
b. What is the distance between ab4 a−1 and bab4 when re-
stricted to paths in the ball of radius 6?
c. Show that a2 b4 a−2 and ba2 b4 a−1 are both in the sphere of
radius 8, and that the distance between them in the Cayley
graph is 2.
d. What is the distance between a2 b4 a−2 and ba2 b4 a−1 when
restricted to paths in the ball of radius 8?
e. Show that a3 b4 a−3 and ba3 b4 a−2 are both in the sphere
of radius 10, and that the distance between them in the
Cayley graph is 2. (These are the vertices highlighted in
Figure 9.3.)
f. What is the distance between a3 b4 a−3 and ba3 b4 a−2 when
restricted to paths in the ball of radius 10?
g. Attempt to prove that the Cayley graph of BS(1, 2), with
respect to {a, b}, is not almost convex. If you get stuck, try
reading [MiS98] or [EH05].
(17) Given a group G and a finite generating set S, a dead end is an
element g such that |gs| ≤ |g| for all s ∈ {S ∪ S −1 }. Finite groups
always have dead end elements, but it is somewhat surprising that
infinite groups can have dead end elements as well.
Dead end elements are also sometimes referred to as pockets.
(If the Cayley graph Γ is thought of as a net, dangling down
from the vertex associated to the identity, then a dead end ele-
ment sits at the bottom of a “pocket.”) The depth of a dead end
element g is

Depth(g) = Min{d(g, h) | |h| > |g|}.

In other words, g has depth d if one needs to take d steps in the

Cayley graph in order to get from g to an element further from
the identity than g. The Cayley graph ΓG,S is said to have deep
pockets if there is n ∈ N such that n ≥ Depth(g) for all dead ends
g ∈ G. Dead ends in L2 and other wreath products are studied
in [CT05].
Let d = [{−1, 0, 1}, 0] ∈ L2 . Show that |d| = 7, but |d · t| =
|d · t−1 | = |d · a| = 6, and therefore d ∈ L2 is a dead end. Also show
that the depth of d is 2.
186 The Geometry of Infinite Groups

(18) Let dn ∈ L2 be the element whose picture has lit lamps at each
integer from −n to n, with the lamplighter positioned at the origin.
Show that dn is a dead end element of depth n + 1, and conclude
that the Cayley graph of L2 with respect to the generators {a, t}
has deep pockets.
 10
Thompson’s Group

Thompson’s group comes up in many different situations, it has interesting

properties, and almost every question you can ask about it is challenging.
. –Kenneth S. Brown

In this chapter we introduce another interesting infinite group described

in terms of functions from R to R, although in this case we only use
the closed interval [0, 1]. We begin by discussing dyadic1 divisions of
[0, 1]. A dyadic division of [0, 1] is constructed by first dividing [0, 1]
into [0, 1/2] and [1/2, 1], and then proceeding to “pick middles” of the
resulting pieces, a finite number of times. For example,

[0, 1] = [0, 1/2] ∪ [1/2, 5/8] ∪ [5/8, 3/4] ∪ [3/4, 1]

is a dyadic division. We refer to the points of subdivision as the chosen

middles, so in the example above the chosen middles are: 1/2, 3/4 and
5/8 (listed
in the order
 they are chosen). We refer to any interval of
m m+1
the form , (0 ≤ m ≤ 2n − 1) as a standard dyadic interval.
2n 2n
Exercise 1 asks you to show that any time you divide [0, 1] into standard
dyadic intervals, you necessarily have a dyadic division of [0, 1].
Dyadic divisions of [0, 1] can be encoded using finite, rooted binary
trees. Recall from Chapter 6 that a rooted binary tree is a rooted tree
where every non-leaf vertex has two descendants. Let T be the rooted
tree whose vertex set corresponds to the standard dyadic intervals, with
[0, 1] being the root. Define the left child of [a, b] to be [a, (a + b)/2] and
the right child to be [(a+b)/2, b]. The “top” of T is shown in Figure 10.1.
Every finite, rooted binary tree T can be viewed as a subtree of T , where
1 m
Recall that the dyadic rationals are all rational numbers of the form where m
2n
and n are integers. They are discussed in our presentation of BS(1, 2) in Chapter 4.

187
188 Thompson’s Group
[0,1]

[0,1/2] [1/2,1]

[0,1/4] [1/4,1/2][1/2,3/4] [3/4,1]

Fig. 10.1. The dyadic intervals generate a rooted binary tree.

the root of T is identified with [0, 1]. The standard dyadic intervals that
form the leaves of such a finite, rooted binary tree give a dyadic division
of [0, 1], and any dyadic subdivision generates a finite, rooted binary
tree. As we will frequently refer to “finite, rooted binary trees” in the
remainder of this chapter, we abbreviate our terminology to “frb-trees.”

Given an ordered pair of dyadic divisions of [0, 1], with the same
number of pieces, there is a corresponding piecewise linear function
f : [0, 1] → [0, 1]. If 0 < m1 < · · · < mk < 1 denote the first set of
chosen middles and 0 < µ1 < · · · < µk < 1 denote the second set of
chosen middles, then f is defined by requiring:

1. f (0) = 0 and f (1) = 1;

2. f (mi ) = µi for all i;
3. f is linear when restricted to [0, m1 ], [mi , mi+1 ] (for 1 ≤ i ≤ k)
and [mk , 1].
For example, if the two dyadic divisions are
[0, 1/2] ∪ [1/2, 3/4] ∪ [3/4, 1]
and
[0, 1/4] ∪ [1/4, 1/2] ∪ [1/2, 1],
then the graph of the associated function is the one shown on the left
in Figure 10.2. If the two divisions have chosen middles {1/2, 3/4, 7/8}
and {1/2, 5/8, 3/4} then the associated function is the one on the right
in Figure 10.2. We will refer to functions created in this manner as
Thompson functions. We show in Proposition 10.3 that the set of all
Thompson functions forms a group whose binary operation is function
composition. This group is called Thompson’s group F .
 Thompson’s Group 189

Fig. 10.2. Two elements of Thompson’s group F .

Thompson functions are continuous bijections of [0, 1], hence they can
be viewed as elements of the group Homeo([0, 1]), whose binary operation
is function composition. Thus, to show that this subset is a subgroup,
it suffices to show that it is closed under inverses and composition. To
do this, we introduce a bit of notation. Since dyadic divisions of [0, 1]
correspond with frb-trees, we may use an ordered pair of frb-trees to
describe elements of the group F . In particular, we will use [T2 ← T1 ]
to denote the Thompson function where the domain has been divided
according to T1 and the range according to T2 . The reader should verify
the following lemma, which exploits this notation.

Lemma 10.1. Let T1 , T2 and T3 be frb-trees with the same number of

leaves. Then
[T3 ← T2 ][T2 ← T1 ] = [T3 ← T1 ]
and
[T2 ← T1 ]−1 = [T1 ← T2 ].

Given an ordered pair of frb-trees (with the same number of leaves)

there is an associated Thompson function. However, many different or-
dered pairs of frb-trees define the same element of F . For example, the
identity corresponds to [T ← T ] for any frb-tree T . If T is a frb-tree,
let T ∧ i denote the frb-tree created by adding a “wedge” or “caret” to
T at bottom of the ith leaf, where the leaves are enumerated left-to-
right, with the leftmost leaf being numbered 0. An example is shown in
Figure 10.3. Notice that the functions associated to the pair [T2 ← T1 ]
and to [T2 ∧ i ← T1 ∧ i] are the same functions. Each has simply had
their (i + 1)st dyadic interval divided in half, which does not alter the
190 Thompson’s Group

Fig. 10.3. A finite, rooted binary tree T is shown above, in darker and thicker
lines. The lighter edges show the result of forming (T ∧ 2) ∧ 2.

associated functions. Let ∼ be the equivalence relation generated by

[T2 ← T1 ] ∼ [T2 ∧ i ← T1 ∧ i].
The functions defined by any two elements in an equivalence class are
the same.
Since equivalence relations are symmetric, not only are pairs of frb-
trees equivalent to pairs with carets added to their leaves, but sometimes
pairs of carets can be deleted. If the leaves numbered i and i + 1 form a
caret in T , then this is an exposed caret. This is equivalent to saying that
there is an frb-tree T
such that T = T
∧ i. If T1 and T2 have matched,
exposed carets, that is T1 = T1
∧ i and T2 = T2
∧ i, then
[T2
← T1
] ∼ [T2 ← T1 ].
If T1 and T2 do not have a pair of matched, exposed carets, then the
ordered pair [T2 ← T1 ] is reduced. Exercise 3 asks you to prove the
following helpful fact.
Lemma 10.2. Each equivalence class contains a unique reduced pair of
frb-trees.
We have now introduced enough terminology to establish that we are
indeed discussing a group.
Proposition 10.3. The set of Thompson functions forms a group under
function composition.
Proof. All that remains to be shown is that the composition of Thompson
functions is another Thompson function. Viewing frb-trees as subtrees
 Thompson’s Group 191

of T , you can form unions T1 ∪T2 , and the result is another frb-tree. This
process can be thought of in terms of overhead transparencies: take two
transparencies, one with T1 on it and the other containing T2 ; lay one
transparency over the other, lining up the roots, and you have produced
T1 ∪ T2 . As each Ti (i = 1 or 2) is a subtree of the union, each Ti can be
expanded to T1 ∪ T2 by adding carets.
Given two pairs of frb-trees, [T2 ← T1 ] and [T4 ← T3 ], there is a
set of equivalent pairs of frb-trees, [T2
← T1
] and [T4
← T3
], where
T2
= T3
= T2 ∪ T3 . Thus, by Lemma 10.1, the composition of the as-
sociated functions is the function associated to [T4
← T1
]. Thus the set
of Thompson functions is closed under composition; as we already knew
they were closed under inversion, the set of Thompson functions forms
a group.

Directly from the definition we can derive a surprising fact about F .

Define the support of f ∈ F to be

Supp(f ) = {x ∈ [0, 1] | f (x) = x}.

Call an element f ∈ F a left element if Supp(f ) ⊂ (0, 1/2). Similarly

f ∈ F is a right element if Supp(f ) ⊂ (1/2, 1). The set of left elements
forms a subgroup of F , which we denote Fl ; the set of right elements
forms a subgroup, Fr . Let l be the homomorphism F → F that takes
f ∈ F to the function

f (2x)/2 0 ≤ x ≤ 1/2
fl (x) =
x 1/2 ≤ x ≤ 1.
The graph of fl consists of a copy of the graph of f that has been shrunk
and tucked into [0, 1/2] × [0, 1/2], and then is extended by the identity
on the remainder of the domain. Similarly let r be the homomorphism
F → F that takes f ∈ F to the function

x 0 ≤ x ≤ 1/2
fr (x) =
1/2 + f (2x)/2 1/2 ≤ x ≤ 1.
The homomorphisms l and r show that both Fl and Fr are isomorphic
to F . Since the support of elements in Fl is disjoint from the support
of elements in Fr , elements of Fl commute with elements of Fr . Thus
the subgroup generated by Fl and Fr is Fl ⊕ Fr ≈ F ⊕ F . We have now
outlined an argument that shows the following:

Proposition 10.4. Thompson’s group F contains a subgroup isomor-

phic to F ⊕ F .
192 Thompson’s Group

In order to construct a generating set for F , we introduced two infinite

families of frb-trees. Let Tn be the frb-tree where T0 is a single caret
and Tn+1 = Tn ∧ (n + 1). Let Sn be the frb-tree where S0 = T0 and
Sn+1 = Tn ∧ n. (Examples are shown in Figure 10.4.)

Fig. 10.4. The frb-trees T3 and S3 .

An element f ∈ F is said to be positive if it corresponds to an ordered

pair of the form [T ← Tn ]; it is negative if it corresponds to an ordered
pair [Tn ← T ]. Since [Tn ← T ] = [T ← Tn ]−1 , the negative elements are
the inverses of the positive elements.

Lemma 10.5. Every element of Thompson’s group F can be expressed

as a product p·n where p is a positive element and n is a negative element
of F .

Proof. Let f be the Thompson function given by [S ← T ] where S and

T have n + 1 leaves. Then

f = [S ← T ] = [S ← Tn ][Tn ← T ]

expresses f as a product of a positive and a negative element.

Define xi to be the Thompson function described by the ordered pair

[Si+1 ← Ti+1 ]. The graphs of the elements x0 and x1 are shown in
Figure 10.2, with x0 on the left and x1 on the right. The reader should
note that x1 is essentially just a copy of the function x0 , tucked into
[1/2, 1] × [1/2, 1]. In fact, the element xn is a rescaled copy of x0 tucked
into [1 − (1/2)n , 1] × [1 − (1/2)n , 1].
This collection of elements satisfies a number of nice relations, two of
which we highlight in the following lemmas.

Lemma 10.6. If i < n then xn xi = xi xn+1 .

This relation can be rewritten as: x−1

i xn xi = xn+1 . As the function xi
 Thompson’s Group 193

takes the interval [1−(1/2)i+2 , 1] linearly to [1−(1/2)i+1 , 1], Lemma 10.6

is an application of the principle that “conjugation is doing the same
thing somewhere else.”
Lemma 10.7. Let i < n + 2. Then [T ← Tn ] · xi = [T ∧ i ← Tn+1 ].
Proof. By definition the element xi is [Si+1 ← Ti+1 ]. Since i < n + 2,
Tn ∪ Si+1 = Tn ∧ i. Thus we have
[T ← Tn ] · xi = [T ← Tn ] · [Si+1 ← Ti+1 ]
= [T ← Tn ] · [Tn ∧ i ← Tn+1 ]
= [T ∧ i ← Tn ∧ i] · [Tn ∧ i ← Tn+1 ]
= [T ∧ i ← Tn+1 ]

Theorem 10.8. Thompson’s group F is generated by the infinite set of

positive elements {x0 , x1 , x2 , . . .}.
Proof. As every element of F is a product of a positive and a negative
element (Lemma 10.5), and the negative elements are inverses of positive
elements, it suffices to show that every positive element of F is a product
of the xi ’s. Let [T ← Tn ] be a positive element. Note that T has a
maximal subtree of the form Tk , such that T = Tk ∧ i1 ∧ i2 ∧ · · · ∧ im ,
where one never adds a caret to the rightmost leaf. The formula given
in Lemma 10.7 then shows
[T ← Tn ] = [Tk ← Tk ] · xi1 · · · xim .
Since [Tk ← Tk ] is just the identity element of F , it follows that the
positive element [T ← Tn ] is xi1 · · · xim .
Corollary 10.9. Thompson’s group F is generated by two elements,
x0 and x1 .
Proof. Lemma 10.6 shows that x2 = x−1 0 x1 x0 . Thus an induction argu-
ment shows xn+1 = x−n0 x x
1 0
n
. Since every element of F can be expressed
as a product of the xi and their inverses, and every xi can be expressed
as a word in the alphabet {x0 , x1 , x−1 −1
0 , x1 }, the result follows.

We conclude by quoting a few known results about Thompson’s group

F . Proofs of the first two results, along with a number of other interesting
facts, can be found in the survey [CFP96].
Theorem 10.10. Every proper quotient of Thompson’s group F is
abelian.
194 Thompson’s Group

Theorem 10.11. Every non-abelian subgroup of F contains a copy of

the wreath product ZZ. In particular, since every subgroup of a free group
is free, Thompson’s group F does not contain a subgroup isomorphic to
the free group F2 .

The next set of results relate to topics introduced in Chapter 9.

Theorem 10.12 ([GuSa97]). Thompson’s group F admits a regular nor-

mal form.

The normal form constructed by Guba and Sapir in [GuSa97] is easy

to describe. It consists of all words in {x0 , x1 , x−1 −1 ∗
0 , x1 } which do not
contain one of these elements followed by its inverse, and which do not
contain any subword of the form

1. x1 xi0 x1 ,
2. x−1
1 x0 x1 ,
i
i+1 −1
3. x1 x0 x1 ,
4. x−1 i+1 −1
1 x0 x1 .

where i is any positive integer. Checking that this set of words is a regular
language is routine; proving that it is a normal form is not. It should
be noted that this normal form is not composed of geodesic words. This
was not due to a lack of insight on the part of the authors. Rather, it
was later shown by Cleary, Elder and Taback that it is impossible to
construct such a normal form.

Theorem 10.13 ([CET06]). The Cayley graph of Thompson’s group F

with respect to {x0 , x1 } has infinitely many cone types. In particular,
there is no regular normal form consisting of geodesic words.

This shows that one cannot hope to show that the growth series for
F is rational using the approach of Section 9.3. As of the time of this
writing, it is unknown if the growth series for Thompson’s group F is
rational or not.
Finally, the geometry of the Cayley graph of F , with respect to {x0 , x1 },
is quite complicated. In particular, Belk and Bux have shown:

Theorem 10.14 ([BeBu05]). The Cayley graph of F with respect to

{x0 , x1 } is not minimally almost convex.

This work exploits a method introduced by Blake Fordham in his PhD

thesis that gives a simple method of computing the length of elements
 Exercises 195

in F with respect to the generators {x0 , x1 }. An alternative, also elemen-

tary, method for computing length was later introduced by Victor Guba.
(See [BeBr05] and [Gu04] for reasonably accessible discussions of these
methods.) It is a rather annoying state of affairs that elementary and
elegant means of computing length are available, but many questions
about the geometry of the Cayley graph of F with respect to {x0 , x1 }
remain unresolved.
Remark 10.15. Thompson’s group F was introduced in 1965 by Richard
Thompson. His original treatment of the group was motivated by alge-
braic questions arising from formal logic. Since then this group has ap-
peared in many different contexts, from deep studies of the associative
law to topological considerations of idempotents.
In addition to F , there are also Thompson groups T and V that add
permutations into the mix. Similar groups where the slopes are not as
constricted, for instance allowing slopes of the form 2m 3n , have also been
studied. The previously mentioned survey [CFP96] is a well-written,
fairly elementary introduction to the groups F, T and V , as well as some
of the known variations.

Exercises
(1) Show that if you divide [0, 1] into dyadic intervals, then this sub-
division must be a dyadic subdivision.
(2) Prove that a continuous, bijective, piecewise linear function f :
[0, 1] → [0, 1] (with finitely many corners) is a Thompson function
if and only if:
a. all of the slopes (away from the corners) are powers of 2;
and
b. all of the corners have both coordinates in the dyadic
rationals.
(3) Prove Lemma 10.2.
(4)
F is torsion-free.
Prove that 
m m+1
(5) Let I = n , be a standard dyadic interval. Let FI be the
2 2n
subset of functions
FI = {f ∈ F | Supp(f ) ⊂ I}.
a. Prove that FI is a subgroup.
b. Prove that FI ≈ F .
196 Thompson’s Group

(6) Show that the set of positive elements of F is closed under com-
position.
(7) In this problem we discuss the standard normal form for elements
of Thompson’s group F using {x0 , x1 , x2 , . . .}.

a. Show that every element of F can be expressed as a word

of the form

xa0 0 xa1 1 · · · xann x−b

n
n
· · · x−b 1 −b0
1 x0

where the ai and bi are non-negative integers. We refer to

such expressions as up-and-down words.
b. For an frb-tree T , define the exponent of a leaf to be the
length of the maximal path consisting only of left edges,
that contains the leaf but does not reach the right side of
T . (An example is shown in Figure 10.5.) If {a0 , a1 , . . . , am }

2 0
0 1 0 0 0
0
3 0

Fig. 10.5. The exponents of the leaves in this frb-tree, T , are as indicated.
This implies the element [T ← T8 ] can be expressed as x20 x32 x6 .

are the exponents of the leaves of an frb-tree T , show that

the positive element [T ← Tm−1 ] is xa0 0 xa1 1 · · · xamm .
c. If f ∈ F corresponds to [S ← T ] describe how to express f
as an up-and-down word using the exponents of S and T .
d. Show that [S ← T ] is a reduced diagram if and only if its
associated up-and-down word xa0 0 · · · xann x−b
n
n
· · · x−b
0
0
satis-
fies:

1. exactly one of an and bn is non-zero; and

2. if ai and bi are both non-zero, then either ai+1 = 0
or bi+1 = 0.
 Exercises 197

Fig. 10.6. An element of Thompson’s group F .

(8) An element of F is shown in Figure 10.6. Express this element as

a word in {x0 , x1 , x−1 −1 ∗
0 , x1 } .
 11
The Large-Scale Geometry of Groups

What one really cares about are the inherent properties of the group, not the
artefacts of a particular presentation.
. –Martin Bridson

11.1 Changing Generators

There is a danger in working with a specific Cayley graph for a given
group G. If you focus on a particular generating set, the results you get
may not immediately translate into similar results when a different set
of generators is used. Even worse, sometimes interesting properties hold
in one Cayley graph but not in another, even though the group under
consideration has not changed.
In this chapter we introduce some of the ways geometry can be
imported into the study of infinite groups, which are independent of the
choice of finite generating set. These are properties that always hold or
always fail, no matter which finite set of generators one uses to construct
a Cayley graph. Of course, such properties cannot focus too closely on
a given Cayley graph, since changing generators changes the details, in
other words the local structure, of the graphs. In Figure 11.1 we high-
light this with two Cayley graphs for D3 . The location of the vertices
has been kept constant in both pictures, which makes the second picture
look a bit odd. When the generators change, so do the distances between
vertices, the number of cycles, and so on.
Because changing generators can have dramatic consequences for the
local structure of a Cayley graph, the properties we consider here are
referred to as large-scale properties; many authors also use the term
geometric properties. A property of a Cayley graph of a finitely generated
group is large-scale, or geometric, only if it is invariant under changes in

198
 11.1 Changing Generators 199

Fig. 11.1. The Cayley graph of D3 with respect to two different generating
sets. The Cayley graph on the left corresponds to using a reflection and a
rotation; the Cayley graph on the right corresponds to using two reflections.

generating set.1 With this convention, the property of having a rational

growth series is not a geometric property of a group (see Remark 9.18).
On the other hand, there is a notion of growth that is geometric, which
we discuss in Section 11.2.
In this section we introduce a standard technique for passing between
Cayley graphs of a fixed group G.

Definition 11.1. Let Γ and Λ be two graphs. A map from Γ to Λ is a

function φ taking vertices of Γ to vertices of Λ, and edges of Γ to edge
paths in Λ, such that if v and w are vertices attached to an edge e ∈ Γ,
then φ(e) joins φ(v) to φ(w).

As an example, consider a tree T consisting of three vertices and two

edges, along with the Cayley graph Γ of Z with respect to {2, 3}. Send
the vertices of T to the vertices corresponding to 0, 1 and 2 in Γ, and
map each edge of T to the edge path corresponding to {3, −2}. The
result is indicated in Figure 11.2.

Proposition 11.2. Let S and T be two finite generating sets for a group
G and let ΓS and ΓT be the corresponding Cayley graphs. Then there are
maps φT ←S : ΓS → ΓT and φS←T : ΓT → ΓS such that:

1. The compositions φS←T ◦ φT ←S and φT ←S ◦ φS←T induce the

identity on V (ΓS ) and V (ΓT ) respectively.
2. There is a constant K > 0 such that the image of any edge e ∈ ΓS
1 Actually, the term is used more broadly than this; it is not restricted to Cayley
graphs or to changes in generating sets. In particular, these terms are used in
the study of metric spaces where a property is geometric if it is invariant under
“quasi-isometry.” This is a term we introduce in Section 11.7.
200 The Large-Scale Geometry of Groups

Fig. 11.2. A map from a finite tree to the Cayley graph of Z with respect to
the generating set {2, 3}.

under φS←T ◦ φT ←S is contained in the ball B(v, K) ⊂ ΓS , where

v ∈ Ends(e). The similar statement also holds for edges of ΓT .

Proof. The map φT ←S is the identity map on the vertices. That is, if
vg denotes the vertex corresponding to g ∈ G in ΓS , and vg denotes
the vertex corresponding to g in ΓT , then φT ←S (vg ) = vg . The map
φS←T is defined in the same way on vertices, and the first claim is then
immediate.
For each generator s ∈ S choose a word ωs = t1 t2 · · · tk ∈ {T ∪ T −1 }∗
such that s = π(ws ) ∈ G. (One can always do this since T is a generating
set for G.) By the construction of Cayley graphs, if e is an edge of ΓS
then e is labelled by a generator s ∈ S and joints the vertex associated
to g to the vertex associated to g · s. The map φT ←S sends every such
edge to the edge path

g −→ g · t1 −→ gt1 · t2 −→ · · · −→ gt1 t2 · · · tk

in ΓT .
The map φS←T is defined similarly. For each generator t ∈ T choose
a word ωt ∈ {S ∪ S −1 }∗ representing t and use these words to describe
the images of the edges of ΓT .
Let k be the maximal length of the words ωt and ωs . It follows that
φT ←S (e) is an edge path of length ≤ k. The image φS←T of this path is
then an edge path of length ≤ k 2 . Hence the constant K in the second
claim of the theorem can be taken to be k 2 .

Corollary 11.3. Let G, S and T be as above. Then there is a constant

λ ≥ 1 such that for any g and h in G,
1
dS (g, h) ≤ dT (g, h) ≤ λdS (g, h).
λ
 11.1 Changing Generators 201

Proof. Let

Λ1 = Max{|ωt | | t ∈ T }

be the maximum length of the words in {S ∪ S −1 }∗ which were chosen

to represent the generators in T . If dT (g, h) = n, then there is an edge
path between vg and vh in ΓT of length n. This path maps under φS←T
to an edge path joining vg and vh in ΓS whose length is at most Λ1 · n.
Thus dS (g, h) ≤ Λ1 · dT (g, h). Repeating this argument, with the roles of
S and T reversed, shows that if Λ2 = Max{|ωs | | s ∈ S}, then dT (g, h) ≤
Λ2 · dS (g, h). Set λ = Max(Λ1 , Λ2 ), and the stated inequalities hold.

Example 11.4. In Chapter 2 we introduced the infinite dihedral group

D∞ , which is generated by two reflections a and b. The product τ = ab
is a translation and {a, τ } is an alternate generating set for D∞ . The
Cayley graph of D∞ with respect to S = {a, b} is shown in Figure 2.2
and the Cayley graph with respect to T = {a, τ } is shown below in
Figure 11.3. If ρ is a rotation in a finite dihedral group, and a is a
reflection, then aρ = ρ−1 a. Similarly in the infinite dihedral group we
have aτ = τ −1 a, as can be seen in the Cayley graph.

Fig. 11.3. The Cayley graph of D∞ with respect to the generating set {a, τ }.

We may express the original generators in terms of the new generators

by setting ωa = a and ωb = aτ . We can express the new generators in
terms of the old generators by setting ωa = a (again) and ωτ = ab.
The map φT ←S will send edges labelled b to edge paths labelled aτ , and
φS←T will send edges labelled τ to edge paths labelled ab. The constant
λ from Corollary 11.3 can be taken to be 2. Thus in passing from the
original generating set to the new generating set (or vice versa) the
distance between vertices is altered by at most a factor of 2.
In this example you can begin to see, in a very vague sense, that some
sort of geometric structure is preserved when you switch generating sets.
The Cayley graphs ΓS and ΓT are different. Yet, if one were to step well
202 The Large-Scale Geometry of Groups

back from their graphs, in both instances it would seem as if you are
looking at the real line.

Remark 11.5. In Chapter 5 we discussed Dehn’s word problem. There

are a number of geometric approaches to the word problem, but all of
them require some familiarity with fundamental groups and covering
spaces, so they will not be discussed in this book. The reader with an
interest in this topic, who has the appropriate background, should take
a look at [BRS07] and the references cited there.

11.2 The Growth of Groups, II

In this section we revisit the notion of growth in groups, first discussed
in Section 9.2. Part of our discussion applies more generally to growth in
metric spaces, and specifically the notion of growth functions. Define any
non-decreasing function f : [0, ∞) → [0, ∞) to be a growth function. For
example, we could take f (x) to be the area enclosed by a circle of radius
x in the Euclidean plane, in which case f (x) = πx2 .
Our interest is in measuring the growth of groups. Let G be a group
with finite generating set S, and let BS (n) denote the ball of radius
n about the identity in G. (This time we are thinking of BS (n) as a
subset of G, not a subgraph of ΓG,S .) The function βS (n) = |BS (n)|
gives the number of elements in G inside the ball of radius n, that is,
the number of elements of G that can be expressed as a product of at
most n generators, and their inverses, coming from S. We extend the
domain of such growth functions by defining β(x) to be β( x
), for all
x ∈ [0, ∞). (Although when we state explicit formulas we do so viewing
βS as a function from N to N.)
Different generating sets for a fixed group G do produce different
growth functions. For example, if we take two reflections as our gener-
ating set for D∞ , S = {a, b}, then βS (n) = 2n + 1. On the other hand,
we may generate D∞ by {a, τ } where τ = ab is a translation. The Cay-
ley graph with respect to this generating set is shown in Figure 11.3.
Using this figure, you will find that βT (n) = 4n (for n ≥ 1). While
βS and βT are not the same, both functions are linear, and it can be
proved that this fact is independent of the generators one chooses. This
is the idea we explore in this section: The rate at which a group expands
(linearly, quadratically, . . . , exponentially) is a geometric property of
the group.
 11.2 The Growth of Groups, II 203

Lemma 11.6. Let S and T be two finite generating sets for a fixed group
G. Then there is a constant λ ≥ 1 such that


1
βS n ≤ βT (n) ≤ βS (λn)
λ
for all n ∈ N.
Proof. Let |g|S denote the length of g ∈ G with respect to the generating
set S, and define |g|T similarly. By Corollary 11.3, there is a λ ≥ 1
such that |g|T ≤ λ · |g|S for all g ∈ G. Thus if g ∈ BT (n) then g ∈
BS (λn), which implies βT (n) ≤ βS (λn). Exchanging the roles of S and
T establishes the other inequality.
Lemma 11.6 is one motivation for the following definition; another
motivation arises in Section 11.7.
Definition 11.7. Define  to be the relation on growth functions de-
fined by f  g if there is a constant λ ≥ 1 such that
f (x) ≤ λg(λx + λ) + λ
for all x ∈ [0, ∞). If f  g then we say g dominates f . (Notice that the
right-hand side is the result of pre- and post-composing g(x) with the
linear function y = λx + λ.)
If f  g and g  f then g strictly dominates f . We denote this by
f ≺ g. On the other hand, if f  g and g  f then f and g are said to
be equivalent, denoted f ∼ g. Two equivalent growth functions are said
to grow at the same rate.
Exercise 2 at the end of this chapter asks you to prove that “” is
a reflexive and transitive relation, and that f ∼ g is an equivalence
relation.
Example 11.8. Consider the following growth functions: f (n) = n2 ,
g(n) = 15n2 + 10n + 1 and h(n) = 2n . We have f (n) < g(n) for all
n ∈ N, so f  g. On the other hand,
f (5n) + 5 = 25n2 + 5 = 15n2 + 10n2 + 5 > 15n2 + 10n + 1 = g(n)
so g  f , hence f ∼ g.
One expects that f  h, since exponential functions grow faster than
polynomial functions. There is a slight hitch in that f (3) = 9 > 8 = h(3)
but, for all other n ∈ N, f (n) ≤ h(n). Thus f (n) < h(2n) + 2 for all n
and so it is indeed the case that f  h. On the other hand,
λ(λn + λ)2 + λ
lim =0
n→∞ 2n
204 The Large-Scale Geometry of Groups

for any λ ≥ 1. Thus f  h and so h strictly dominates f .

Lemma 11.6 implies the following:

Corollary 11.9. If S and T are two finite generating sets for a group
G then the associated growth functions are equivalent: βS (n) ∼ βT (n).

The equivalence class of a growth function for a group G is then

a large-scale invariant of the group. For example, if β(n) is a growth
function for D∞ , then β(n) ∼ n. Hence one one can say that the infinite
dihedral group has linear growth.

Definition 11.10. Let G be some fixed group with β(n) an associated

growth function. Then G is said to have polynomial growth if there is
some d ∈ N such that β(n)  nd . The group G has polynomial growth
of degree d if β(n) ∼ nd .

It is easy to check that the growth function for Z is β(n) = 2n + 1,

with respect to a cyclic generator. Hence Z is another example of a group
with linear growth. You can also establish that β(n) = 2n2 + 2n + 1 for
Z ⊕ Z with respect to its standard generating set. Thus Z2 is said to
have quadratic growth. These examples should lead us to conjecture
that free abelian groups of rank d have polynomial growth of degree
d. One could verify this by explicitly computing formulas for each β(n),
with respect to the standard generators of Zd . But when looking at large-
scale issues one rarely needs such precise information. The elements in
Zd that are contained in the ball of radius n consist of those d-tuples of
integers (n1 , n2 , . . . , nd ) where |n1 | + |n2 | + · · · +|nd | ≤ n. A standard
n+d−1
combinatorial argument shows there n+d−1 are d such d-tuples whose
entries are non-negative. Thus d  β(n). Similarly, if we allow
both
n+d−1positive and negative entries, you can get the bound β(n)  2d ·
for Zd , with respect to the standard
d
d . Thus the growth function

generators, satisfies β(n) ∼ n+d−1 d ∼ n . We therefore have:

Proposition 11.11. The free abelian group Zd = Z ⊕ · · · ⊕ Z has poly-

  
d copies
nomial growth of degree d.

Not all growth functions of groups are polynomial. Consider the free
group of rank k, Fk , along with a fixed basis. The number of elements
in the sphere of radius n, S(n), is 2k · (2k − 1)n−1 (for n ≥ 1). Thus
(2k − 1)n < |S(n)| < |B(n)|. Hence the growth function of Fk satisfies
(2k − 1)n  β(n). In particular, as long as k ≥ 2, the growth is at least
 11.3 The Growth of Thompson’s Group 205

exponential. On the other hand,

n
β(n) = |S(n)| = 1 + 2k + 2k · (2k − 1) + · · · + 2k · (2k − 1)n−1
i=0

< 1 + 2k + (2k)2 + · · · (2k)n−1 < (2k)n .

So β(n)  (2k)n , hence the growth is at most exponential. But these
bounding exponential functions are equivalent.

Lemma 11.12. Let a and b be two integers, both greater than 1, and
let α(n) = an and β(n) = bn be the corresponding exponential functions.
Then α(n) ∼ β(n).

Proof. We will assume that a < b so that α(n)  β(n) trivially. Con-
versely, if λ ≥ loga (b), then β(n) ≤ α(λn) + λ, so β(n)  α(n).

A group G is said to have exponential growth if its associated growth

function satisfies β(n) ∼ 2n . Lemma 11.12 shows that the base 2 could
be replaced by any integer ≥ 2, and this combined with the previous
discussion of free groups shows:

Proposition 11.13. Every finitely generated free group, whose rank is

at least 2, has exponential growth. Further, if βn represents a growth
function for Fn , then βn ∼ βm for any integers m and n that are both
greater than or equal to 2.

Corollary 11.14. If G is a finitely generated group, then its growth is

dominated by 2n .

Proof. Assume G has a generating set with k elements. As every element

of G can be expressed as a freely reduced word in its generators and their
inverses, the associated growth function is bounded above by the growth
function of Fk . But the growth functions of free groups are either linear
(k = 1) or equivalent to 2n (k > 1).

11.3 The Growth of Thompson’s Group

In 1968 John Milnor asked if there are groups of “intermediate growth,”
which is to say: can the growth function of a finitely generated group G
strictly dominate nd for all d, and yet not
√
be equivalent to 2n ? There are
n
functions with this property, such as 2 , but can such functions arise
as the growth function of a finitely generated group? Since Thompson’s
group F contains a copy of F ⊕ F (Proposition 10.4), it also contains
206 The Large-Scale Geometry of Groups

a copy of F ⊕ F ⊕ F , and so on. Thus Thompson’s group F contains

a copy of Zd for all d. So it is reasonable guess that the growth of F
dominates nd for all d. Further, F does not contain a non-abelian free
group (Theorem 10.11), so perhaps the growth of F is not exponential.

Theorem 11.15. Thompson’s group F has exponential growth.

Proof. Recall that Thompson’s group F is generated by two elements,

x0 and x1 , whose graphs are shown in Figure 10.2. Let ω0 and ω1 be
two words in {x0 , x1 }∗ . (We are ignoring all inverses!) Our goal in this
proof is to show that no two such words represent the same element of
F , that is, π(ω0 ) = π(ω1 ), where π : {x0 , x1 , x−1 −1 ∗
0 , x1 } → F is the
evaluation map.
Assume to the contrary that π(ω0 ) = π(ω1 ) for some pair of words
in the generators x0 and x1 . Choose an example where the combined
length of ω0 and ω1 is as short as possible. Note that the last letters of
such a pair must be distinct. If they both end in the same letter, “x,”
this letter can be removed from both words, forming ω0 and ω1 . But

π(ω0 ) = π(ω0 · x−1 )

= π(ω0 )π(x−1 )
= π(ω1 )π(x−1 )
= π(ω1 ).

Thus the resulting pair would still satisfy π(ω0 ) = π(ω1 ), but their
combined length will have been reduced. Since |ω0 | + |ω1 | was minimal,
this is not possible, and so the two words end in distinct generators.
Without loss of generality we assume ω0 ends in x0 and ω1 ends in x1 .
If f is an element of F , then when restricted to some small interval
[0, ), f is linear with slope 2k . The function φ : F → Z that takes f
to this exponent k is a homomorphism. (If you compose a function with
slope 2k with a function of slope 2l then the result has slope 2k+l .) In
particular, φ(x0 ) = −1 and φ(x1 ) = 0. Since ω0 and ω1 represent the
same element of F , φ(ω0 ) = φ(ω1 ). Further, |φ(ω0 )| = |φ(ω1 )| is the
number of x0 ’s in these words. Call this number n. Note that n > 0
because otherwise ω0 and ω1 would simply be powers of x1 .
Thinking of x0 as a function from [0, 1] to [0, 1] we see x0 (3/4) = 1/2
and, more generally, xk0 (3/4) = 1/2k for any positive integer k. Since x1
is the identity when restricted to [0, 1/2], and ω0 ends in x0 , it follows
that π(ω0 ) takes 3/4 to 1/2n (where n is as defined above).
 11.3 The Growth of Thompson’s Group 207

Now consider the action of π(ω1 ) on [0, 1]. There is some positive k
such that ω1 ends with x0 xk1 . The function xk1 takes 3/4 to some number
strictly smaller than 3/4, hence x0 xk1 takes 3/4 to some number strictly
smaller than 1/2. Any additional x1 ’s that appear in ω1 have no effect on
[0, 1/2], but there are n−1 more x0 ’s in ω1 . Thus, in the end, π(ω1 ) takes
3/4 to some number which is strictly less than 1/2n . So it is impossible
for ω0 and ω1 to represent the same element of Thompson’s group F .
Because there are 2n words of length n in {x0 , x1 }∗ , it follows that
2n ≤ β(n), and so Thompson’s group F has exponential growth.

While it is interesting to see that Thompson’s group F has exponential

growth, and that this can be determined even though details about the
growth series of F are currently unknown, this example does not resolve
Milnor’s Question: are there groups of intermediate growth? In 1983,
Grigorchuck proved the following:

Theorem 11.16. There are finitely generated groups G with growth

function β where
nd ≺ β ≺ 2n

for all d.

The first known example of a group of intermediate growth was very

similar to the infinite torsion group that we constructed in Chapter 6.
A nice, accessible discussion of this result can be found in [GrPa07].
We close this section by quoting another important result, a full dis-
cussion of which is beyond the scope of this text. The notion of groups
of polynomial growth is a geometric notion, defined in terms of met-
ric properties of the group. Seemingly unrelated is the algebraic prop-
erty of being nilpotent. Let Z1 = Z(G) be the center of G and define
Zn+1 = {g ∈ G | for all h ∈ G, hgh−1 g −1 ∈ Zn }. Thus Z2 /Z1 is the
center of G/Z1 and in general Zn+1 /Zn is the center of G/Zn . A group
G is nilpotent if G = Zn for some n. Abelian groups are nilpotent since
G = Z(G) = Z1 . The Heisenberg group H consists of 3-by-3 integer
matrices of the form:
  
 1 a b 
 0 1 c  | a, b, c ∈ Z .
 
0 0 1

The Heisenberg group H is not abelian. However, Z(H) ≈ Z and H/Z(H) ≈

Z ⊕ Z, so H is nilpotent.
In 1981, Gromov published the following amazing result:
208 The Large-Scale Geometry of Groups

Theorem 11.17. A finitely generated group has polynomial growth if

and only if it is virtually nilpotent.
The original paper, [Gr81], not only established this result, it also
introduced advanced techniques that have proved to be quite useful in
studying the geometry of infinite groups.

11.4 The Ends of Groups

Let Γ be a connected, locally finite graph, and let B(n) be the ball
of radius n in Γ, based at some fixed vertex. Define ||Γ \ B(n)|| to be
the number of connected, unbounded components in the complement of
B(n).
Lemma 11.18. Let Γ be a locally finite graph and let m < n be two
positive integers. Then
||Γ \ B(m)|| ≤ ||Γ \ B(n)||.
Proof. Let C be an unbounded, connected component of Γ\B(m). Either
C remains connected when B(n) is removed, or it does not. Thus each
unbounded, connected component of Γ \ B(m) contributes at least one
unbounded, connected component to Γ \ B(n).
Definition 11.19 (Ends of a graph). Let Γ be a connected, locally
finite graph, and let B(n) be the ball of radius n about a fixed vertex
v ∈ V (Γ). The number of ends of Γ is
e(Γ) = lim ||Γ \ B(n)||.
n→∞

This limit exists, although it may equal ∞, since, by Lemma 11.18, the
sequence {||Γ \ B(n)||} is a non-decreasing sequence of integers.

Example 11.20. Let Combn be the graph formed by joining n “rays”

by a combinatorial path connecting their base vertices, as in Figure 11.4.
Let v be the vertex in the bottom left-hand corner. Then:
||Combn \ B(0)|| = 2,
||Combn \ B(1)|| = 3,
||Combn \ B(2)|| = 4,
and so on, until ||Combn \ B(n − 1)|| = n, after which
||Combn \ B(k)|| = n
 11.4 The Ends of Groups 209

Fig. 11.4. The number of ends of “Combn ” is n.

for all k ≥ n − 1. Thus e(Combn ) = n.

You may wonder why the adjective “unbounded” is included in the

definition of ||Γ \ B(n)||. If Γ \ B(n) could contain a finite connected
component, then such a component will eventually be contained in some
B(m) for m > n. Thus that finite component is not a part of the graph
that extends indefinitely. Further, such finite components can occur. At
the end of the exercises in Chapter 9 there is a discussion of dead ends
in Cayley graphs. This discussion and the following exercises show that
if Γ is the Cayley graph of the lamplighter group then Γ \ B(n) will have
finite, connected components.
The definition of the number of ends of a graph Γ uses a particular
choice of a base vertex, but this vertex does not show up in the notation.
This choice does not influence the value of e(Γ) and, to show this, we
extend an earlier definition. Let C be any finite subgraph of Γ and let
||Γ \ C|| denote the number of unbounded, connected components of the
complement of C. Let

ec (Γ) = Max{||Γ \ C|| | C is a finite subgraph of Γ}

be the largest value that occurs for all finite subgraphs C, or set ec (Γ) =
∞ if there is no n ∈ N such that n ≥ ||Γ \ C|| for all finite subgraphs
C ⊂ Γ. (If you have studied real analysis, feel free to replace “Max” with
“sup.”)
210 The Large-Scale Geometry of Groups

Lemma 11.21. Let Γ be a connected, locally finite graph and let e(Γ)
and ec (Γ) be as defined above. Then e(Γ) = ec (Γ).

Proof. Since metric balls B(n) are finite subgraphs in any locally finite
graph Γ, e(Γ) ≤ ec (Γ). Conversely, given any finite subgraph C, there is
an n ∈ N such that C ⊂ B(n). Thus ||Γ \ C|| ≤ ||Γ \ B(n)||, by the same
argument used to prove Lemma 11.18, and therefore ec (Γ) ≤ e(Γ).

It is natural to consider the number of ends of a Cayley graph of some

finitely generated group G. In keeping with the theme of this chapter,
we first establish that the resulting number is independent of the choice
of Cayley graph.

Lemma 11.22. Let S and T be two finite generating sets for G, and
let BS (n) and BT (n) be the balls of radius n in ΓS and ΓT , respectively.
Then there is a constant µ ≥ 1 such that if vg and vh are vertices in ΓS
that can be joined by an edge path outside of BS (µn + µ) then vg and vh
in ΓT are outside BT (n) and can be joined by a path that stays outside
of BT (n).

Proof. Let φT ←S be a map taking ΓS → ΓT as in Proposition 11.2. By

Corollary 11.3, if λ is at least the length of the longest expression chosen
to represent the generators from S in terms of words in {T ∪ T −1 }∗ then
1
dT (g, h) ≥ dS (g, h).
λ
Set µ = λ2 + 1.
Let {vg = v0 , v1 , v2 , . . . , vn = vh } be the vertices that occur on an
edge path from vg to vh that is contained in ΓS \ BS (µn + µ). Then
1
dT (ve , vi ) ≥ (µn + µ) > n + λ.
λ
Thus each vi is outside of BT (n + λ). If e is the edge joining vi to
vi+1 in ΓS then the length of φT ←S (e) is at most λ. It follows that
φT ←S (e) is outside of BT (n). Thus φT ←S takes the path joining vg to
vh in ΓS \ BS (µn + µ) to a path connecting vg to vh in ΓT \ BT (n).

Theorem 11.23. Let S and T be two finite generating sets for a group
G, and let ΓS and ΓT be the corresponding Cayley graphs. Then

e(ΓS ) = e(ΓT ).
 11.5 The Freudenthal–Hopf Theorem 211

Proof. By Lemma 11.22 if two vertices in ΓS can be connected in the

complement of BS (µn+µ) then they can be connected in the complement
of BT (n). Thus ||ΓS \ BS (µn + µ)|| ≥ ||ΓT \ BT (n)||. It follows that

lim ||ΓS \ BS (n)|| ≥ lim ||ΓT \ BT (n)||,

n→∞ n→∞

that is, e(ΓS ) ≥ e(ΓT ). The opposite inequality holds mutatis mutandis,
hence we are done.

As the number of ends of a Cayley graph for G does not depend on

the choice of Cayley graph, we may make the following definition.

Definition 11.24. Let G be a finitely generated group. The number

of ends of G is the number of ends of any of its Cayley graphs. By
Theorem 11.23 this number is a property of the group G; it does not
depend on the choice of Cayley graph. We denote the number of ends of
a finitely generated group G by e(G).

Exercise 11.25. Show that a finitely generated group G has zero ends
if and only if G is finite.

Example 11.26. Consider a finitely generated free group Fn where n is

at least 2. The Cayley graph, with respect to a basis, is a tree where every
vertex has valence 2n. Thus ||Γ \ B(k)|| = 2n(2n − 1)k , and therefore
e(Fn ) = ∞, for all n ≥ 2.

11.5 The Freudenthal–Hopf Theorem

At the end of Section 11.4 we pointed out that finite groups have zero
ends (Exercise 11.25) and we sketched an argument showing that free
groups have infinitely many ends (Example 11.26). Here are two more
examples that are easily verified: Z has two ends while Z ⊕ Z is one-
ended. What other values can e(G) take? This question was resolved in
the early 1930s, and the answer is referred to as the Freudenthal–Hopf
Theorem.

Theorem 11.27 (Freudenthal–Hopf Theorem). Every finitely generated

group has either zero, one, two or infinitely many ends.

Proof. Previous examples show that there are groups with zero, one, two
or infinitely many ends. Thus we need to establish that there is no group
G where e(G) is finite but e(G) > 2. Assume to the contrary that the
Cayley graph Γ of a finitely generated group G has k ≥ 3 ends. Note that
212 The Large-Scale Geometry of Groups

this implies that G must be an infinite group, and that there is a number
n ∈ N such that Γ \ B(n) has k unbounded, connected components.
Since G is infinite there is an element g ∈ G such that d(e, g) > 2n,
where vg is in an unbounded, connected component of Γ \ B(n). Then
g · B(n) ∩ B(n) = ∅. Further, since gB(n) is contained in an unbounded,
connected component of Γ \ B(n), it divides this component into at least
k connected pieces, at least (k − 1) of which are unbounded. Thus if
C = B(n) ∪ g · B(n) then C is a finite subgraph of Γ and ||Γ \ C|| ≥ 2k − 2
(see Figure 11.5). It follows that ec (Γ) – as defined in Lemma 11.21 – is
at least 2k − 2. Thus e(Γ) ≥ 2k − 2 > k, since k ≥ 3. But this contradicts
the claim that e(G) = k.

(n) g (n)

Fig. 11.5. If a group G has at least three ends, then it has at least four ends,
which means it has at least six ends, ...

Exercise 11.28. Copy and complete Table 11.1, filling in the values
of e(G). The formatting of this table is intended to help you guess the
correct value of e(G), even if you are not clear on how to actually prove
that your guess is correct.

11.6 Two-Ended Groups

Given that every finitely generated group has zero, one, two or infinitely
many ends, it is natural to ask which groups belong in a particular
collection. We know that a group has zero ends if and only if it is finite.
Looking over Table 11.1 it appears that the next simplest collection are
the groups with two ends. The examples of such groups that come to
 11.6 Two-Ended Groups 213

G e(G) G e(G)

Sym5 Z⊕Z

Z7
Z3 W244
(D397
Z736 )8

(D397
Z736 )8 BS(1, 2)

G e(G) G e(G)

Z F2
(D397
Z736 )8 (D397
Z736 )8

D∞ Z3 ∗ Z4

Z ⊕ Sym7 (Z ⊕ Z) ∗ Z2

Table 11.1. Exercise 11.28 asks you to fill in the values e(G) in this
table.

mind are Z, D∞ , and any direct sum of a two-ended group with a finite
group. In each case the group has a finite index subgroup ≈ Z, and the
goal of this section is to establish that this is indeed always the case.
Recall that the symmetric difference of two sets is defined to be
AB = (A ∪ B) \ (A ∩ B) .
(The symmetric difference was previously mentioned in our discussion of
the lamplighter group in Chapter 8.) Since our discussion of two-ended
groups will rely heavily on this operation, you should verify the following
two formulas.
214 The Large-Scale Geometry of Groups

Exercise 11.29. Let A, B and C be sets. Then

AB = (AC)  (CB) .
Denote the complement of a set X by X c . Then
AB = Ac B c .
For the remainder of this section, let G be a finitely generated group
with e(G) = 2, and let Γ be a Cayley graph of G. There is then a
finite subgraph C such that Γ \ C has exactly two connected, unbounded
components. By adding all the finite components of Γ \ C to C, we may
assume that Γ \ C consists of exactly two connected components, each
of which is unbounded. Choose one of these two complements, and let
E ⊂ G consist of those elements of G that correspond to the vertices in
that component. Notice that we may apply elements of G to E, forming
subsets
gE = {g · h | h ∈ E}.
Similarly we can left multiply E c :
g · E c = {g · k | k ∈ E}.
c
Notice that g · E c is the same set as (g · E) . Thus we may simply write
gE c , avoiding the use of parentheses.
The proof of our first lemma is left as an exercise (Exercise 12 at the
end of this chapter).
Lemma 11.30. Let G and E be as above, and let g ∈ G. Then, because
c
G is two-ended, either EgE is finite or (EgE) is finite.
Lemma 11.31. Let G and E be as above. Then the subset of elements
H = {g ∈ G | EgE is finite}
forms a subgroup of index at most 2.
Proof. We first establish that H is a subgroup. To show that it is closed
under inverses, one only needs to note that
|Eh−1 E| = |hEE|.
Hence if EhE is finite, so is Eh−1 E. To show that H is closed under
products, take two elements g, h ∈ H. Then
EghE = (EgE)  (gEghE)
= (EgE)  g (EhE) .
 11.6 Two-Ended Groups 215

Since we have assume EgE and EhE are both finite, so is EghE.
In order to establish that the index of H in G is at most 2, assume
that H = G and that g, h ∈ G \ H. Then
 
Egh−1 E = (EgE)  gEgh−1 E
 
= (EgE)  g Eh−1 E
 c
= (EgE)  g Eh−1 E .
c

 c
By Lemma 11.30, both (EgE) and Eh−1 E are finite, hence so
c

is Egh−1 E.
Let C be the finite subgraph that divided the Cayley graph Γ into
two unbounded, connected components, and let g ∈ H be an element
where gC ∩ C = ∅. Then either gC ⊂ E or gC ⊂ E c . From this and the
definition of H, one gets:
Lemma 11.32. Let G, H, C and E be as above, and let g ∈ H be an
element such that C ∩ gC = ∅. Then either
1. E ∩ gE c = ∅ and E c ∩ gE = ∅, or
2. E ∩ gE c = ∅ and E c ∩ gE = ∅.
We are now well-positioned to establish the main result of this section.
Theorem 11.33 below was first established by C. T. C. Wall in 1967,
using purely algebraic techniques. Our approach has been to build up
an intuitive notion of the “right side” of the Cayley graph Γ, which is
the part that contains E. On an intuitive level, an element g ∈ H moves
Γ “to the right” if gE ⊂ E. Further, g moves Γ “far to the right” if
E ∩ gE c is large. Surprisingly, this vague intuition can be made quite
precise.
Theorem 11.33. Let G and H be as above. Then there is a homomor-
phism φ : H → Z whose kernel is finite.
Proof. Since g ∈ H, EgE is finite, hence E ∩ gE c and E c ∩ gE are
also finite. We may then define a function φ : H → Z by
φ(g) = |E ∩ gE c | − |E c ∩ gE|.
Notice that E ∩ gE c is the disjoint union
E ∩ gE c = (E ∩ gE c ∩ ghE c ) ∪ (E ∩ gE c ∩ ghE)
and similarly
E c ∩ gE = (E c ∩ gE ∩ ghE c ) ∪ (E c ∩ gE ∩ ghE) .
216 The Large-Scale Geometry of Groups

Thus we may also define φ(g) to be

φ(g) =|E ∩ gE c ∩ ghE c | + |E ∩ gE c ∩ ghE|
− |E c ∩ gE ∩ ghE c | − |E c ∩ gE ∩ ghE|.
Now let h ∈ H. We have
φ(h) = |E ∩ hE c | − |E c ∩ hE| = |gE ∩ ghE c | − |gE c ∩ ghE|.
By dividing gE ∩ ghE c and gE c ∩ ghE into the parts that are in E and
the parts that are in E c , we have
φ(h) =|E ∩ gE ∩ ghE c | + |E c ∩ gE ∩ ghE c |
− |E ∩ gE c ∩ ghE| − |E c ∩ gE c ∩ ghE|.
Cancelling terms and applying the same trick as above, but in reverse,
we find
φ(g) + φ(h) = |E ∩ gE c ∩ ghE c | + |E ∩ gE ∩ ghE c |
= − |E c ∩ gE ∩ ghE| − |E c ∩ gE c ∩ ghE|
= |E ∩ ghE c | − |E c ∩ ghE|
= φ(gh).
Thus, not only is φ a function from H to Z, it is a group homomorphism.
We use Lemma 11.32 to see that the kernel of φ must be finite. As C
is a finite subgraph of a Cayley graph Γ, there are only finitely many
elements where C ∩gC = ∅. But if C ∩gC = ∅ then Lemma 11.32 implies
φ(g) = 0.
Corollary 11.34. A group G is two-ended if and only if G contains a
finite index subgroup isomorphic to Z.
Proof. Theorem 11.33 states that G has a finite index subgroup H where
there is a group homomorphism φ : H → Z with finite kernel. If h ∈ H
is any element where φ(h) = 0, then h ≈ Z and h has finite index in
H. Since either G = H or [G : H] = 2 it follows that h has finite index
in G.
Conversely, in Section 11.7 we show that if H is a finite index sub-
group of G, then e(H) = e(G). In particular, since e(Z) = 2, if G has a
subgroup ≈ Z, then e(G) = 2.
We now know that zero-ended groups are the same as finite groups, and
two-ended groups are the same as groups that contain Z as a finite index
subgroup. So in these two cases the connection between the structure of
the group and the number of ends is clear. In 1968 John Stallings proved
 11.7 Commensurable Groups and Quasi-Isometry 217

that given an infinitely ended group G, there is a tree T and an action

G
T . Further, this action satisfies a number of nice properties that
allow you to express G directly in terms of stabilizers of vertices and
edges of T . A full description of this theorem is beyond the aims of this
text, but there is a result that is simple to state.
Theorem 11.35. Let G be a finitely generated, torsion-free group. Then
G has infinitely many ends if and only if G is a free product, G ≈ H ∗K,
where neither H nor K is the trivial group.
The original proof of this result, under the stronger hypothesis that G is
finitely presented, can be found in [St68]. While some of the techniques
used in Stallings’ article are beyond a standard undergraduate experi-
ence, many parts of it are accessible. In fact, our discussion of two-ended
groups is based on portions of this paper.
By repeatedly applying Theorem 11.35 to a finitely generated group,
along with a few other powerful results, one eventually sees that every
finitely generated, torsion-free group can be expressed as a free product
of a free group with a finite number of one-ended groups (where the
free group may be trivial or it might be the entire group). Since free
groups are reasonably well understood, it is the one-ended groups that
are mysterious. Here, indeed, there is great variety. Because of this, the
development of more refined tools that can be used to explore what
happens “at infinity” in an infinite group is of active interest. See for
example [Ge08].

11.7 Commensurable Groups and Quasi-Isometry

When we introduced the Drawing Trick (Remark 1.49), we described it
as a means of constructing a Cayley graph of a group. If G
X where
X is some geometric object, and there is an x ∈ X that G moves freely,
then there is a bijective correspondence between points in the orbit of
x and the elements of G. The set Orb(x) is something like a collection
of polka dots decorating X, and often it provides an easy-to-visualize
collection of vertices for a Cayley graph. Consider the elementary ex-
ample of Z ⊕ Z acting on R2 with generators (1, 0) and (0, 1) being a
horizontal translation and a vertical translation, respectively. The orbit
of the origin under this action is the set of all points (m, n) whose coor-
dinates are integers. We refer to such points as integral points. We can
view the collection of integral points as the vertices of a Cayley graph,
Γ, embedded in R2 , where the edges are arcs of length 1.
218 The Large-Scale Geometry of Groups

Given two integral points we can measure the distance between them
in R2 using the standard metric:

dR2 = (m1 − m0 )2 + (n1 − n0 )2 .

We can also measure the distance between them in the Cayley graph of
Z ⊕ Z or, in other words, in the word metric associated to the generators
{(1, 0), (0, 1)}:
dZ2 = |m1 − m0 | + |n1 − n0 |.

How different are these two notions of distance? The distance between
two points in the plane is realized by a straight line segment and, given
any such line, there is an edge path in Γ ⊂ R2 that stays close to it (as in
Figure 11.6). Because straight lines in R2 can be roughly approximated

Fig. 11.6. Straight lines in the Euclidean plane can be roughly approximated
by geodesic edge paths in an embedded Cayley graph for Z ⊕ Z.

by edge paths in Γ, one expects that there is a way to move between

the geometry of the Euclidean plane and the geometry of this Cayley
graph.
√ The straight line from (0, 0) to (m, n) in R2 has length equal to
m2 + n2 , which is less than (or occasionally equal to) the length of an
edge path in Γ between these points. On the other hand, we can always
arrange it so that a straight line joining integral points is the hypotenuse
of a right triangle, the sides of which sit in the Cayley graph Γ. Thus
the distance
√ in Γ is bounded from above by twice the distance in R2
(or by 2 times the distance in R2 if you want to keep your constants
minimal). If we let ι : Z2 → R2 be the polka-dot embedding of Z ⊕ Z in
R2 that we have been discussing, then we have just established that for
any two elements x, y ∈ Z2 :
1
dZ2 (x, y) ≤ dR2 (ι(x), ι(y)) ≤ 2dZ2 (x, y)
2
where we have included an unnecessary “2” in the right-hand inequality
in order to draw a parallel with the inequalities given in Corollary 11.3.
 11.7 Commensurable Groups and Quasi-Isometry 219

We are now in a situation that is similar to our consideration of differ-

ent Cayley graphs of a fixed group. Both R2 and Γ provide a geometric
picture of Z ⊕ Z, one using the Euclidean metric and the other using
the word metric, and the distortion between them seems controllable.
There is, however, one important difference that shows up when we con-
struct a map R2 → Z2 . Define φ : R2 → Z2 to be a map sending (x, y)
to a closest integral point; when there is more than one closest inte-
gral point, such as when (x, y) = (0.5, 0.5), simply choose one of the
available options. Notice that φ will take some points which are close
together and move them further apart. For example, φ[(0.4, 0.4)] = (0, 0)
and φ[(0.6, 0.6)] = (1, 1). In fact, φ[(0.49 · · · 99, 0.49 · · · 99)] = (0, 0) and
φ[(0.50 · · · 01, 0.50 · · · 01)] = (1, 1). Thus the map φ will take points that
are arbitrarily close to each other in R2 to points that are a distance 2
apart in Z2 . Hence there is no real number λ ≥ 1 such that
dZ2 (φ(x), φ(y)) ≤ λdR2 (x, y)
for any x, y ∈ R2 . What is true is that the distortion introduced by φ
is also controllable, but one needs to allow for a small amount√of local
noise. Given any point x ∈ R2 , the point φ(x) will be at most 1/ 2 away
from x; we also know that dR2 (φ(x), φ(y)) ≤ dZ2 (φ(x), φ(y)); combining
this with the triangle inequality gives us
dR2 (x, y) ≤ dR2 (x, φ(x)) + dR2 (φ(x), φ(y)) + dR2 (φ(y), y)
≤ dR2 (φ(x), φ(y)) + 2
≤ dZ2 (φ(x), φ(y)) + 2.
We also know that dZ2 (φ(x),√φ(y)) ≤ 2dR2 (φ(x), φ(y)) as well as that √
dR2 (φ(x), φ(y)) ≤ dR2 (x, y) + 2. So dZ2 (φ(x), φ(y)) ≤ 2dR2 (x, y) + 2 2.
Combining these results we get:
√
dR2 (x, y) − 2 ≤ dZ2 (φ(x), φ(y)) ≤ 2dR2 (x, y) + 2 2.
There is little reason to attempt to make these constants as small as
possible, and we prefer a more symmetric statement than that given
above, hence we conclude by noting that the inequalities above imply:

3 dR (x, y) − 3 ≤ dZ2 (φ(x), φ(y)) ≤ 3dR2 (x, y) + 3.

1
2

Definition 11.36. Let X and Y be two metric spaces and let φ be a

function from X to Y . The function φ is a quasi-isometric embedding if
there is a constant λ ≥ 1 such that for all x, x ∈ X:
1
dX (x, x ) − λ ≤ dY (φ(x), φ(x )) ≤ λdX (x, x ) + λ.
λ
220 The Large-Scale Geometry of Groups

(See Definition 11.7, where similar inequalities were used to define equiv-
alent growth functions.)
The function φ is quasi-dense, which is also sometimes referred to as
quasi-onto, if there is a λ such that, for any point y ∈ Y , there is some
x ∈ X such that dY (φ(x), y) ≤ λ.
Finally, φ : X → Y is a quasi-isometry if it is both a quasi-isometric
embedding and is quasi-dense.
If φ is a quasi-isometry and λ is large enough to satisfy all the re-
quirements above, then λ is referred to as the quasi-isometry constant
of φ.
There are a number of elementary examples of quasi-isometric spaces.
Our discussion of the polka-dot embedding of Z2 in R2 shows that the
group Z2 with its standard word metric is quasi-isometric to the Eu-
clidean plane. The cyclic subgroup of Z2 generated by (1, 0) is quasi-
isometrically embedded in R2 , but it is not quasi-isometric to R2 , as it
is not quasi-dense. In addition to the specific case of Z2
R2 , we have
also encountered a large collection of quasi-isometric metric spaces. If
G is a finitely generated group, and S and T are two finite generating
sets, then the word metric on G induced by S and the word metric on
G induced by T yield quasi-isometric metric spaces; this follows imme-
diately from Corollary 11.3. As this is an important point, we record
it as:
Lemma 11.37. Let G be a finitely generated group and let S and T be
two finite generating sets. Then G with the word metric induced by S is
quasi-isometric to G with the word metric induced by T .
Let φ : X → Y be a quasi-isometry with associated constant λ. Define
a function ψ : Y → X by ψ(y) = x, where x ∈ X is chosen so that
dY (φ(x), y) ≤ λ. (Such an x always exists because φ is quasi-dense.) It
follows from the definition of ψ that
1. dY (φ ◦ ψ(y), y) ≤ λ for all y ∈ Y .
Because φ is a quasi-isometry, with quasi-isometry constant λ, we know
that
1
dX (x, x ) − λ ≤ dY (φ(x), φ(x )) ≤ λdX (x, x ) + λ.
λ
In particular, if x = ψ ◦ φ(x), then φ(x) = φ(x ), which implies that
dY (φ(x), φ(x )) = 0. This establishes two additional facts:
2. dX (x, ψ ◦ φ(x)) ≤ λ for all x ∈ X.
 11.7 Commensurable Groups and Quasi-Isometry 221

3. Given any x ∈ X there is a y ∈ Y such that dX (x, ψ(y)) ≤ λ.

(Take y = φ(x).)

We can rewrite the basic quasi-isometry inequalities for φ to read:

1
dY (φ(x), φ(x )) − λ ≤ dX (x, x ) ≤ λdY (φ(x), φ(x )) + λ.
λ
Thus, for any y and y  ∈ Y ,
1
dY (φ ◦ ψ(y), φ ◦ ψ(y  )) − λ
λ
≤ dX (ψ(y), ψ(y  ))
≤ λdY (φ ◦ ψ(y), φ ◦ ψ(y  )) + λ.

However, dY (φ ◦ ψ(y), y) ≤ λ. So by the triangle inequality we have

dY (φ ◦ ψ(y), φ ◦ ψ(y  ))
≤ dY (φ ◦ ψ(y), y) + dY (y, y  ) + dY (y  , φ ◦ ψ(y  ))
≤ dY (y, y  ) + 2λ.

By substitution we then get:

 
dX (ψ(y), ψ(y  )) ≤ λdY (y, y  ) + 2λ2 + λ .

Combining this with a similar argument applied to the other inequality

establishes:

4. There is a constant Λ ≥ 1 such that, for any y, y  ∈ Y ,

1
dY (y, y  ) − Λ ≤ dX (ψ(y), ψ(y  )) ≤ ΛdY (y, y  ) + Λ.
Λ
Thus we have the following result.

Lemma 11.38. If φ : X → Y is a quasi-isometry between metric spaces,

then there is a quasi-isometry ψ : Y → X. Further, there is a constant
k such that, for all x ∈ X and all y ∈ Y , dX (ψ ◦ φ(x), x) ≤ k and
dY (φ ◦ ψ(y), y) ≤ k.

If X and Y are metric spaces, we let X ∼QI Y denote the relation “X

is quasi-isometric to Y .”

Proposition 11.39. The relation X ∼QI Y is an equivalence relation.

Proof. The identity map is a quasi-isometry from X to X hence X ∼QI

X for all metric spaces X. If X ∼QI Y , then Lemma 11.38 tells us there
222 The Large-Scale Geometry of Groups

is a “quasi-inverse” that establishes Y ∼QI X. All that remains to be

done is to establish the transitive property.
Let φ1 : X → Y and φ2 : Y → Z be two quasi-isometries. Let λ
be the maximum of the two associated quasi-isometry constants. The
composition φ = φ2 ◦ φ1 is a function from X to Z. Given any two
points x, x ∈ X we know that
1
dX (x, x ) − λ ≤ dY (φ1 (x), φ1 (x )) ≤ λdX (x, x ) + λ.
λ
and that
1
dY (φ1 (x), φ1 (x )) − λ
λ
≤ dZ (φ2 ◦ φ1 (x), φ2 ◦ φ1 (x ))
≤ λdY (φ1 (x), φ1 (x )) + λ.
Thus we have
1
dX (x, x ) − (λ + 1) ≤ dZ (φ(x), φ(x )) ≤ λ2 dX (x), x ) + (λ2 + λ).
λ2
Setting Λ = λ2 + λ, we find
1
dX (x, x ) − Λ ≤ dZ (φ(x), φ(x )) ≤ ΛdX (x, x ) + Λ,
Λ
so the function φ : X → Z is a quasi-isometric embedding.
It remains to be shown that φ is quasi-dense. To do this, let z be a point
in Z. By hypothesis, there is some y ∈ Y such that
dZ (φ2 (y), z) ≤ λ. There is also an x ∈ X such that dY (φ1 (x), y) ≤ λ,
and therefore
dZ (φ(x), φ2 (y)) ≤ λ2 + λ.
The triangle inequality implies
dZ (φ(x), z) ≤ dZ (φ(x), φ2 (y)) + dZ (φ2 (y), z) ≤ λ2 + 2λ.
Making our previously chosen value of Λ a bit larger, Λ = λ2 + 2λ,
insures that the composition φ = φ2 ◦ φ1 is a quasi-isometry from X to
Z with quasi-isometry constant Λ. Thus if X ∼QI Y and Y ∼QI Z then
X ∼QI Z.
The group D∞ acts on the real line, as was discussed in Chapter 2.
Mimicking the discussion of Z2
R2 , it is not hard to establish that
D∞ ∼QI R. If a and b are generating reflections, then τ = ab is a
translation, and Z ≈ τ  < D∞ . Not surprisingly, Z ∼QI R. Because
∼QI is an equivalence relation, it follows that Z ∼QI D∞ , which implies
 11.7 Commensurable Groups and Quasi-Isometry 223

that a group can be quasi-isometric to one of its subgroups. In fact, this

is always the case, assuming that the subgroup is of finite index. In order
to establish this claim, we first need to prove the following result.
Theorem 11.40. Let H be a finite index subgroup of G. Then H is
finitely generated if and only if G is finitely generated.
Proof. First, assume that G is finitely generated, and let Γ be an
associated Cayley graph. Thus G
Γ, where Γ is a locally finite, con-
nected graph. Further, a fundamental domain for this action is a finite
subgraph. Since H is a subgroup of G, H also acts on Γ. A fundamen-
tal domain for the action of H on Γ consists of [G : H] copies of a
fundamental domain for G
Γ. Thus H
Γ also has a finite funda-
mental domain, since H has finite index in G. Using this fundamental
domain for H
Γ one can construct a finite generating set for H. (Use
Theorems 1.55 and 1.58 to fill in the details of this argument.)
Conversely, assume H has a finite generating set {h1 , . . . , hm }.
Because H has finite index in G, we can express G as a union of finitely
many H-cosets: G = H ∪ g1 H ∪ · · · ∪ gn H. Thus every g ∈ G can be
expressed as g = gi h for some h ∈ H, and therefore every g ∈ G can be
expressed as a word in {g1 , . . . gn , h1 , . . . , hm , h−1 −1 ∗
1 , . . . , hm } .

Theorem 11.40 is useful in many contexts in the study of infinite

groups. For us it is most useful in allowing us to view a finitely generated
group G and any finite index subgroup H < G as metric spaces, where
the distance function in each is given by a word metric. As any two word
metrics for a given group yield quasi-isometric metric spaces, we rarely
need to be specific about which word metric is being considered.
Proposition 11.41. Let H be a finite-index subgroup of G, both of
which are finitely generated. Then H ∼QI G.
Proof. Let Γ be the Cayley graph of G with respect to a generating set
S. Since H is a subgroup of G there is an action H
Γ. Since H has
finite index in G, there is a finite fundamental domain F for this action.
Let BR be the ball of radius R in Γ, centered at the vertex corresponding
to the identity, where R is chosen large enough so that F ⊂ BR . Since
H · F = Γ, we also know H · BR = Γ, hence the distance from any g ∈ G
to some h ∈ G is at most R. So the embedding of the subgroup H into
G is quasi-dense.
The same argument as is given for Theorem 1.55 shows that
T = {h ∈ H | hBR ∩ BR = ∅}
224 The Large-Scale Geometry of Groups

is a generating set for H. Further, if h ∈ T then the length of h, with

respect to the generators S ⊂ G, is at most 2R. Hence by replacing
each h ∈ T by a word of length ≤ 2R in {S ∪ S −1 }∗ we prove that
dH (h, h ) ≤ 2RdG (h, h ).
To establish the other inequality, notice that any finite edge path
in Γ is contained in a finite sequence {h0 BR , h1 BR , . . . , hk BR } where
hi BR ∩ hi+1 BR = ∅. Further, if the edge path contains n edges, then
k ≤ n. Thus dH (h, h ) ≤ dG (h, h ). It follows that
1
dH (h, h ) − 2R ≤ dG (h, h ) ≤ 2RdH (h, h ) + 2R
2R
and so the embedding of the subgroup H into G is a quasi-isometry.
Definition 11.42. Two groups G and H are commensurable if there is
a finite sequence {G = H0 , H1 , . . . , Hn = H} where either Hi contains
Hi+1 as a finite-index subgroup, or Hi is a finite-index subgroup in Hi+1 .
An induction argument, using Proposition 11.41, shows:
Corollary 11.43. Let G and H be commensurable groups. Then G is
finitely generated if and only if H is finitely generated. If they are finitely
generated, then G ∼QI H.
Example 11.44. Consider the group Z3 ∗ Z4 from Section 3.5 and
SL2 (Z). By Corollary 3.41, Z3 ∗Z4 has a finite-index subgroup isomorphic
to F6 . Proposition 3.7 shows that SL2 (Z) has a finite-index subgroup iso-
morphic to F2 . Proposition 3.12 shows that F2 contains a finite-index
subgroup isomorphic to F3 . A generalization of this is given as Exer-
cise 15 in Chapter 3, which asked you to show that, for any n ≥ 2, Fn
is a finite-index subgroup in F2 . Thus the sequence
{Z3 ∗ Z4 , F6 , F2 , SL2 (Z)}
satisfies the conditions of Definition 11.42, whence Z3 ∗ Z4 ∼QI SL2 (Z).
A similar argument establishes a more general result, which we leave
to the reader as Exercise 15.
Remark 11.45. While commensurable groups are quasi-isometric, the
converse is not true. There are many quasi-isometric groups that are
not commensurable. One way to construct such examples is via the
Švark–Milnor Theorem, a general result that encompasses much of our
discussion in this section. The theorem states that if X is a “reason-
able” metric space, and G
X is “reasonable,” then G ∼QI X. To
construct quasi-isometric groups that are not commensurable one then
 Exercises 225

just needs to find “reasonable” actions of two groups on the same “rea-
sonable” metric space, where the sizes of the associated fundamental
domains are not rational multiples of each other. (An account of the
Švark–Milnor Theorem and its consequences – including explanations of
“reasonable” – can be found in [BrHa99].)
We have previously established that the growth functions for a given
group G, with respect to two finite generating sets, are equivalent (Corol-
lary 11.9) and that the number of ends of a group is not dependent on
the choice of finite generating set (Theorem 11.23). These results also
hold for quasi-isometric groups. For example, assume φ : H → G is a
quasi-isometry with associated constant λ. If h ∈ H is in the ball of ra-
dius n about the identity, then φ(h) must be in the ball of radius λn + λ
about the identity (in G). Thus βH (n) ≤ βG (λn + λ), so βH  βG . The
quasi-inverse ψ : G → H shows that βG  βH , and therefore βG ∼ βH .
We leave the proof that the number of ends is a quasi-isometry invariant
as an exercise (Exercise 16 at the end of this chapter).
Theorem 11.46. Let G and H be finitely generated groups that are
quasi-isometric. Then
1. if βG and βH denote growth functions of G and H, then βG ∼ βH ;
2. e(G) = e(H).
Thus we know, for example, that Zn is not quasi-isometric to Zm (for
m = n) because they have inequivalent growth rates. Similarly, F2 and
the lamplighter group are not quasi-isometric because e(F2 ) = ∞ while
e(L2 ) = 1 (Exercise 9).

Exercises
(1) The dihedral group D3 can be generated by two reflections or a
reflection and a rotation, and the associated Cayley graphs are
shown in Figure 11.1. Construct a map from the Cayley graph on
the right to the Cayley graph on the left in this figure, along the
lines discussed in the proof of Proposition 11.2.
(2) Let  be the relation on growth functions defined by f  g if there
is a constant λ ≥ 1 such that
f (x) ≤ λg(λx + λ) + λ
for all x ∈ [0, ∞). Show that  is symmetric and transitive, and
that the relation f ∼ g – defined to mean both f  g and g  f is
an equivalence relation
226 The Large-Scale Geometry of Groups

(3) Prove that if√c < d, with c and d ∈ N, then nc  nd but nd  nc .

(4) Show that 2 n is a growth function and that
√
nd ≺ 2 n
≺ 2n
√
for any d ∈ N. (In other words, show that 2 n is an intermediate
growth function.)
(5) Let G and H be finitely generated groups of polynomial growth.
Show that G ⊕ H also has polynomial growth.
(6) Let H be a subgroup of G, both finitely generated. Let βH and βG
be growth functions for these groups. Show that βH  βG .
(7) Prove that a finitely generated group G has linear growth if and
only if it is virtually cyclic.
(8) Prove that if H is a finite index subgroup of G then e(H) = e(G).
(9) Prove that the lamplighter group is one-ended.
(10) Show that e(F2 ⊕ F2 ) = 1.
(11) Let G and H be finitely generated, infinite groups. Prove that
e(G ⊕ H) = 1. (Warning: The existence of “dead ends” in Cayley
graphs indicates that some care needs to be taken in establishing
this claim.)
(12) Prove Lemma 11.30.
(13) Prove that b is not a quasi-isometric embedding of Z into BS(1, 2).
(14) Prove that the Euclidean reflection groups W244 , W236 and W333
are commensurable with each other.
(15) Let G = G1 ∗ G2 and H = H1 ∗ H2 be two free products of non-
trivial finite groups, where at least one Gi and at least one Hi has
order greater than 2. Prove G ∼QI H.
(16) Prove that the number of ends is a quasi-isometry invariant. That
is, if G and H are quasi-isometric groups, then e(G) = e(H).
 Bibliography

[An71] A.V. Anı̄sı̄mov, The group languages, Kibernetika (Kiev) 4 (1971)

18–24.
[BBN59] G. Baumslag, W. W. Boone, and B. H. Neumann, Some unsolvable
problems about elements and subgroups of groups, Math. Scand. 7 (1959)
191–201.
[BS62] G. Baumslag and D. Solitar, Some two-generator one-relator non-
Hopfian groups, Bull. Am. Math. Soc. 68 (1962) 199–201.
[BeBr05] J. Belk and K. S. Brown, Forest diagrams for elements of Thompson’s
group F , Int. J. Algebra Comput. 15 (2005) 815–50.
[BeBu05] J. Belk and K-U. Bux, Thompson’s group F is maximally nonconvex,
in Geometric Methods in Group Theory, Cont. Math. 372 (2005) 131–46.
[Be97] C. Bennett, Explicit free subgroups of Aut(R, ≤), Proc. Am. Math. Soc.
125 (1997) 1305–1308.
[BjBr05] A. Björner and F. Brenti, Combinatorics of Coxeter Groups, Graduate
Texts in Mathematics 231, Springer, 2005.
[Bo87] O. V. Bogopolski, Arboreal decomposability of the group of automor-
phisms of a free group, Algebra i Logica 26 (1987) 131–49.
[BRS07] N. Brady, T. Riley, and H. Short, The Geometry of the Word Problem
for Finitely Generated Groups, Birkhäuser, 2007.
[Bz94] M. Brazil, Growth functions for some nonautomatic Baumslag-Solitar
groups, Trans. Am. Math. Soc. 342 (1994) 137–54.
[Br08] M. Bridson, Conditions that prevent groups from acting non-trivially
on trees, Geometry & Topology Monographs 14, Geometry & Topology
Publications, (2008) 129–133.
[BrHa99] M. Bridson and A. Haefliger, Metric Spaces of Non-Positive Curva-
ture, Springer, 1999.
[Ca84] J. W. Cannon, The combinatorial structure of cocompact discrete hy-
perbolic groups, Geom. Dedicata 16 (1984) 123–48.
[Ca87] J. W. Cannon, Almost convex groups, Geom. Dedicata 22 (1987)
197–210.
[CFP96] J. W. Cannon, W. J. Floyd, and W. R. Parry, Introductory notes on
Richard Thompson’s groups, Enseign. Math. (2) 42 (1996) 215–56.
[Ca78] A. Cayley, The theory of groups: Graphical representation, Desiderata
and Suggestions No. 2 Am. J. Math. 1 (1878) 174–6.
[CT05] S. Cleary and J. Taback, Dead end words in lamplighter groups and
other wreath products, Quarterly J. Math. 56 (2005) 165–78.

227
 228 Bibliography

[CET06] S. Cleary, M. Elder, and J. Taback, Cone types and geodesic languages
for lamplighter groups and Thompson’s group F , J. Algebra 303 (2006)
476–500.
[CEG94] D. J. Collins, M. Edjvet, and C. P. Gill, Growth series for the group
x, y|x−1 yx = y l , Arch. Math. (Basel) 62 (1994) 1–11.
[CL90] J. H. Conway and J. C. Lagarias, Tiling with polyminoes and combina-
torial group theory, J. Comb. Theory, A 53 (1990) 183–208.
[CV96] M. Culler and K. Vogtmann, A group-theoretic criterion for property
FA. Proc. Am. Math. Soc. 124 (1996) 677–83.
[Da08] M. W. Davis, The Geometry and Topology of Coxeter Groups, London
Mathematical Society Monographs, Princeton University Press, 2008.
[De12] M. Dehn, Über unendliche diskontinuierliche gruppen, Math. Ann. 71
(1912) 116–44.
[Di94] W. Dicks, Equivalence of the strengthened Hanna Neumann conjecture
and the amalgamated graph conjecture, Invent. Math. 117 (1994) 373–89.
[EH05] M. Elder and S. Hermiller, Minimal almost convexity, J. Group Theory
8 (2005) 239–66.
[E+92] D. Epstein, J. W. Cannon, D. Holt, S. V. F. Levy, M. S. Paterson, and
W. P. Thurston, Word Processing in Groups, Jones and Bartlett, Boston,
1992.
[Ge08] R. Geoghegan, Topological Methods in Group Theory, Springer, 2008.
[Gi87] R. Gilman, Groups with a rational cross-section, in S. M. Gersten and
J. R. Stallings (eds), Combinatorial Group Theory and Topology, Prince-
ton University Press, 1987, pp. 175–83.
[Gi05] R. Gilman, Formal languages and their application to combinatorial
group theory, in Groups, Languages, Algorithms, Cont. Math. 378, Amer-
ican Mathematical Society, 2005, pp. 1–36.
[Gl92] A. M. W. Glass, The ubiquity of free groups, Math. Intelligencer 14
(1992) 54–7.
[Gr80] R. I. Grigorchuk, On Burside’s problem on periodic groups, Funkt-
sional. Anal. i Prilozhen 14 (1980) 53–4.
[GrPa07] R. I. Grigorchuk and I. Pak, Groups of intermediate growth: an in-
troduction for beginners, to appear, L’Enseignement Mathématique.
[Gr81] M. Gromov, Groups of polynomial growth and expanding maps, Publ.
math. de l’ I.H.E.S., 53 (1981) 53–78.
[Gr87] M. Gromov, Hyperbolic groups, in S. M. Gersten (ed.), Essays in Group
Theory, Math. Sci. Res. Inst. Publ. 8, Springer, 1987, pp. 75–263.
[GrMa64] I. Grossman and W. Magnus, Groups and their Graphs, Mathemat-
ical Association of America, 1964.
[Gv96] J. R. J. Groves, Mimimal length normal forms for some soluble groups,
J. Pure Appl. Algebra 114 (1996) 51–8.
[Gu04] V. S. Guba, On the properties of the Cayley graph of Richard Thomp-
son’s group F , Int. J. Algebra Comput. 14 (2004) 677–702.
[GuSa97] V. S. Guba and M. V. Sapir, The Dehn function and a regular set of
normal forms for R. Thompson’s group F , J. Austral. Math. Soc. Ser. A
62 (1997) 315–28.
[GuSi83] N. Gupta and S. Sidki, On the Burside problem for periodic groups,
Math. Z. 182 (1983) 385–8.
[Hi51] G. Higman, A finitely related group with an isomorphic proper factor
group, J. Lond. Math. Soc. 26 (1951) 59–61.
 Bibliography 229

[Ho81] D. F. Holt, A graph which is edge transitive but not arc transitive,
J. Graph Theory 5 (1981) 201–4.
[HoMe04] P. Hotchkiss and J. Meier, The growth of trees, The College Mathe-
matics Journal 35 (2004) 143–51.
[Hu90] J. E. Humphreys, Reflection Groups and Coxeter Groups, Cambridge
Studies in Advanced Mathematics 29, Cambridge University Press, 1990.
[Jo97] D. L. Johnson, Presentations of Groups, 2nd edition, Cambridge
University Press, 1997.
[Ku33] A. Kurosch, Über freie Produkte von Gruppen, Math. Ann. 108 (1933)
26–36.
[MKS66] W. Magnus, A. Karrass, and D. Solitar, Combinatorial Group Theory,
John Wiley, 1966. (Reprinted by Dover)
[Me72] S. Meskin, Non-residually finite one-relator groups, Trans. Am. Math.
Soc. 64 (1972) 105–14.
[Mi92] C. F. Miller, Decision problems for groups – survey and reflections, in
G. Baumslag and C. F. Miller (eds), Algorithms and Classification in
Combinatorial Group Theory, Math. Sci. Res. Inst. Publ. 23, Springer,
1992 pp. 1–59.
[MiS98] C. F. Miller and M. Shapiro, Solvable Baumslag-Solitar groups are not
almost convex, Geom. Dedicata 72 (1998) 123–7.
[MuS83] D. Muller and P. Schupp, Groups, the theory of ends and context-free
languages, J. Computer and System Sciences 26 (1983) 295–310.
[Ne90] W. D. Neumann, On intersections of finitely generated subgroups of free
groups, in L. G. Kovacs (ed.), Groups: Canberra 1989, Springer, 1990,
pp. 161–70.
[Ni24] J. Nielsen, Die Isomorphismengruppe der freien Gruppen, Math. Ann.
91 169–209.
[Ol95] A.Yu. Ol’shanskii, A simplification of Golod’s example, in A. C. Kim
and D. L. Johnson (eds), Groups: Korea ’94 (Pusan), de Gruyter, 1995,
pp. 263–5.
[Pa92] W. Parry, Growth series of some wreath products, Trans. Am. Math.
Soc. 331 (1992) 751–9.
[Re98] S. Rees, Hairdressing in groups: a survey of combings and formal lan-
guages, in The Epstein Birthday Shrift, Geometry and Topology Mono-
graphs 1, 1998, pp. 493–509.
[Se77] J-P. Serre, Arbres, Amalgams, SL2 , Astérisque No. 46, Société
Mathématique de France, 1977.
[Se80] J-P. Serre Trees (trans. John Stilwell), Springer-Verlag, 1980.
[St68] J. Stallings, On torsion-free groups with infinitely many ends, Ann.
Math. 88 (1968) 312–34.
[St96] M. Stoll, Rational and transcendental growth series for the higher
Heisenberg groups, Invent. Math. 126 (1996) 85–109.
[Th90] W. P. Thurston, Conway’s tiling Groups, Am. Math. Monthly 97 (1990)
757–73.
[Ti72] J. Tits, Free subgroups in linear groups, J. Algebra 20 (1972) 250–270.
[Wa86] S. Wagon, The Banach-Tarski Paradox, Cambridge University Press,
1986.
[Wh88] S. White, The group generated by x → x + 1 and x → xp is free,
J. Algebra 118 (1988) 408–22.
[Wi06] H. S. Wilf, Generating Functionology, 3rd edition, A. K. Peters, 2006.
 Index

action, 3 Fix(g) and Fix(G), 90

discrete, 101 free group, 54
isometric, 163 free monoid, 37
proper, 101 free product, 80
transitive, 11 Freudenthal–Hopf Theorem, 211
almost convexity, 179 fundamental domain, 31
minimal, 181 and generators, 33
alphabet, 37 and index of a subgroup, 35
automaton, 131 Core, 32
complete, 133
deterministic, 133
finite-state, 131 generators, 17
non-deterministic, 132 geodesic, 165
graph, 6
ball (in a group), 169 bipartite, 6
Baumslag–Solitar group, 100 Cayley, 21
Burnside Problem, 128 complete, 6
restricted, 128 directed, 10
weak, 120 locally finite, 7
null, 6
Cannon pair, 178 Petersen, 13
Cayley graph, 21 simple, 6
Cayley’s Basic Theorem, 4 Gromov’s Corollary, 163
Cayley’s Better Theorem, 19 group (notation)
combing, 107 Aut(Fn ), 93
commensurable, 224 BS(1, n), 100
cone type Dn , 2
graphical, 176 D∞ , 45
linguistic, 138 GLn (R), 2
conjugacy problem, 119 Homeo(R), 62
SL2 (Z), 59
dead end, 185 Symn , 2
deep pockets, 185 Sym(X), 2
Drawing Trick, 25 Wpqr , 51
dyadic rationals, 102 Zn , 1
growth
ends rational, 169
of a graph, 208 series, 169
of a group, 211 spherical, 169

230
 Index 231

Hanna Neumann Conjecture, 148 quasi-isometry, 220

Howson’s Theorem, 144
hyperbolic groups, vii representation, 3
raithful, 3
isometry, 163
semi-direct product, 154
lamplighter group, 156 sphere (in a group), 168
language, 130 stabilizer, 15
regular, 133 of an edge, 41
leaf, 90
length
of a group element, 163 Thompson’s group F , 188
of a word, 65 torsion group, 120
torsion-free, 120
metric space, 162 transfer-matrix method, 173
Myhill–Nerode Theorem, 138 tree, 8
frb, 188
Nielsen–Schreier Theorem, 73 leaves, 90
normal form, 106 regular, 8
regular, 141 rooted, 121
Tm,n , 9
orbit, 16
Orbit-Stabilizer Theorem, 16 virtually abelian, cyclic, free ..., 83

path, 7 word, 37
reduced, 8 associated edge path, 38
Ping Pong Lemma, 64 freely reduced, 54
presentation, 68 word metric, 164
Property FA, 89 word problem, 109
Pumping Lemma, 136 wreath product, 156