0% found this document useful (0 votes)

36 views320 pages

Cambridge Books Online © Cambridge University Press, 2010

Uploaded by

Renan Fernandes

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

36 views320 pages

Cambridge Books Online © Cambridge University Press, 2010

Uploaded by

Renan Fernandes

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 320

LONDON MATHEMATICAL SOCIETY STUDENT TEXTS

Managing editor: Professor E.B. Davies, Department of Mathematics,

King's College, Strand, London WC2R 2LS

1 Introduction to combinators and X-calculus, J.R. HINDLEY &

J.P. SELDIN
2 Building models by games, WILFRID HODGES
3 Local fields, J.W.S. CASSELS
4 An introduction to twistor theory, S.A. HUGGETT & K.P. TOD
5 Introduction to general relativity, L. HUGHSTON & K.P. TOD
6 Lectures on stochastic analysis: diffusion theory, DANIEL W. STROOCK
7 The theory of evolution and dynamical systems, J. HOFB AUER &
K. SIGMUND
8 Summing and nuclear norms in Banach space theory, G.J.O. JAMESON
9 Automorphisms of surfaces after Nielsen and Thurston, A.CASSON &
S. BLEILER
10 Nonstandard analysis and its applications, N.CUTLAND (ed)
11 Spacetime and singularities, G. NABER
12 Undergraduate algebraic geometry, MILES REID
13 An introduction to Hankel operators, J.R. PARTINGTON
14 Combinatorial group theory: a topological approach, DANIEL E. COHEN
16 An introduction to noncommutative noetherian rings, K.R.GOODEARL &
R.B.WARFIELDJR.
17 Aspects of quantum field theory in curved space-time, S.A. FULLING

Downloaded from https://www.cambridge.org/core. Tufts Univ, on 11 Jul 2018 at 01:30:01, subject to the Cambridge Core terms of use, available at
Cambridge
https://www.cambridge.org/core/terms. Books Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878
Downloaded from https://www.cambridge.org/core. Tufts Univ, on 11 Jul 2018 at 01:30:01, subject to the Cambridge Core terms of use, available at
Cambridge
https://www.cambridge.org/core/terms. Books Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878
London Mathematical Society Student Texts. 14

Combinatorial Group Theory:

a topological approach
DANIEL E. COHEN
Reader in Pure Mathematics, Queen Mary College, University of London

The right of the

University of Cambridge
to print and sell
all manner of books
was granted by
Henry VIII in 1534.
The University has printed
and published continuously
since 1584.

CAMBRIDGE UNIVERSITY PRESS

Cambridge
New York Port Chester Melbourne Sydney

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/9780521341332
© Cambridge University Press 1989

This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.

First published 1989

Re-issued in this digitally printed version 2008

A catalogue record for this publication is available from the British Library

ISBN 978-0-521-34133-2 hardback

ISBN 978-0-521-34936-9 paperback

To the memories of Ralph Fox, from whom I first learned of the

connections between group theory and topology, and of Hanna Neumann and
Roger Lyndon, whose works extended my interest in group theory.

Combinatorial group theory can be regarded as that branch of group

theory which considers groups given by generators and relations. Some of its
basic results involve manipulation with words; that is, products of the generators
and their inverses. It is this aspect to which the word "combinatorial" refers; it
is not connected with that branch of mathematics known as combinatorics.
From its earliest stages this theory has been closely connected with
topology. To any topological space there is an associated group, called the
fundamental group of the space. In trying to investigate properties of certain
spaces we are led to problems in combinatorial group theory. Conversely, some
problems in combinatorial group theory are best solved by geometric and
topological discussions of suitable fundamental groups.
It is this interplay between group theory and topology which is the
theme of this book. The texts on combinatorial group theory by Magnus, Karrass,
and Solitar (1966) and by Lyndon and Schupp (1977) go much deeper into the
group theory, but have little to say about the topology, while the texts by
Massey (1967) and Stilwell (1980) concentrate on the topology rather than the
group theory.
Chapter 1 contains the main constructions of combinatorial group
theory; free groups, presentations, free products, amalgamated free products, and
the HNN extension. The various Normal Form Theorems are proved, with several
different proofs, and applications of the constructions are made (for instance, to
show the existence of a finitely generated infinite simple group).
Chapter 2 has some topological preliminaries; ways of building new
spaces from old ones, and a discussion of paths in spaces.
From looking at paths in spaces, we are led to consider objects
very like groups, but for which the multiplication is not defined everywhere.
These objects are called groupoids, and their theory is developed in Chapter 3.
Chapter 4 contains the main properties of fundamental groups,
including van Kampen's theorem which gives the fundamental group of the union

of two spaces in terms of the fundamental groups of the spaces and their
intersection. The proof of this theorem is simplified by the use of the
fundamental groupoid. There follows a brief account of covering spaces, the
deeper properties being left to a later chapter. It is shown how these properties
can be used to obtain results about the complex plane; in particular, to give a
proof of the Fundamental Theorem of Algebra. The chapter concludes with
a discussion of some surprising situations which can occur when the spaces
considered are not nice enough.
In Chapter 5 we look: at graphs and complexes. Part of their theory
parallels that of Chapter 4, but without the topological complications that may
occur; occasionally, however, there is some cost in avoiding these technicalities.
Among the material in this chapter is a discussion of the fixed subgroup of an
automorphism of a free group. This uses the very easy approach of Goldstein
and Turner, and its recent extension by M. Cohen and Lustig.
In Chapter 6 we discuss coverings of spaces and complexes. This is
put to use in group theory in Chapter 7. In that chapter we prove Schreier's
theorem on subgroups of a free group, and Kurosh's theorem on subgroups of a
free product. These theorems are proved both using coverings and by purely
algebraic means, and the connection between these two approaches is
investigated in detail. The material in this chapter is perhaps the heart of the
interplay between topology and group theory.
The long Chapter 8 is devoted to the theory of groups acting on
trees, due to Bass and Serre. To some extent, this can be described as a
one-dimensional rewriting of the theory of coverings, and this one-dimensional
aspect makes it easier to work with than the other approaches, once the main
theorems have been proved. We begin with a discussion of free actions of groups
on trees, which leads to yet another proof of Schreier's theorem. We then look at
Nielsen's approach to subgroups of free groups. After that we develop the general
results of Bass-Serre theory. These results are then applied, in the first place to
give a simple proof of the Kurosh Subgroup Theorem and related theorems about
the subgroups of amalgamated free products and HNN extensions. Further
applications are to Grushko's theorem on free products, and to the theorems of
M. Hall and of Howson about finitely generated subgroups of free groups. These
theorems find their best versions, including various generalisations, in this
context. The chapter concludes with a discussion of general ways of constructing
trees, including a statement of the important new theorem due to Dicks and
Dunwoody, showing how this theorem can be applied to give the structure of
groups with a free subgroup of finite index.

Chapter 9 is about decision problems for groups. It begins with a

discussion about the nature of decision problems. Some easy decision problems
are then looked at, together with an introduction to the word problem. We then
show that there is a group with unsolvable word problem, and we prove
Higman's Embedding Theorem, in both cases using the method of modular
machines, which provides short proofs of these results. Various other related
problems are then looked at, including"the unsolvabiiity of the isomorphism
problem. The chapter concludes with a modern proof of Magnus's Freiheitssatz
for one-relator groups and his solution of the word problem for these groups.
Chapter 10 is a quick account, without proofs, of some other topics
in the theory, including small cancellation theory and Whitehead's results on
automorphisms.
This book is an expanded version of a book produced in the Queen
Mary College Lecture Notes series some ten years ago. The latter was an account
of a two-term course of postgraduate lectures given in 1976-7.
In my original two-term course I covered most of the material in
Chapters 1 to 7. For a one-semester course the best choice is to concentrate
either on the group theory or the topology. 1 am currently giving a one-semester
course, in which I cover Chapter 1, the first three sections of Chapter 5, and the
part of Chapter 7 dealing with Schreier's Theorem using graphs but not
coverings. Alternatively, Chapters 2,4, possibly parts of 5, 6, and some of 7
could form a course on the fundamental group, while Chapter 2 and parts of
Chapter 4 (especially the early part and the section on the circle and the
complex plane) could even form part of an undergraduate topology course.
Readers will notice that I have used the phrase "it is easy to see
that..." very often. I believe that, whenever I have used this phrase, all the
results are genuinely easy to prove. Readers should regard such results as
additional exercises.
I am grateful to Ian Chiswell for reading the early parts of this
book and for pointing out obscurities and inaccuracies, which I have endeavoured
to correct.
The typist of a complex mathematical work like this deserves
credit, and I am happy to acknowledge all my hard work in typing this book.
I would also like to thank Franz Schmerbeck, the creator of the Signum word
processor (running on the Atari ST), without whose valuable work the production
of this book would have been much more difficult.

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:27:53, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.001
Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:27:53, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.001
TABLE OF CONTENTS

Page
CHAPTER 1. COMBINATORIAL GROUP THEORY 1
1.1 Free groups 1
1.2 Generators and relators 17
1.3 Free products 24
1.4 Pushouts and amalgamated free products 29
1.5 HNN extensions 34

CHAPTER 2. SPACES AND THEIR PATHS 49

2.1 Some point-set topology 49
2.2 Paths and homotopies 54

CHAPTER 3. GROUPOIDS 62
3.1 Groupoids 62
3.2 Direct limits 70

CHAPTER 4. THE FUNDAMENTAL GROUPOID AND THE FUNDAMENTAL

GROUP 74
4.1 The fundamental groupoid and the fundamental
group 74
4.2 Van Kampen's theorem 83
4.3 Covering spaces 96
4.4 The circle and the complex plane 101
4.5 Joins and weak joins 107

CHAPTER 5. COMPLEXES 113

5.1 Graphs 113
5.2 Complexes and their fundamental groups 121
5.3 Free groups and their automorphisms 132
5.4 Coverings of complexes 139
5.5 Subdivisions 146
5.6 Geometric realisations 149

CHAPTER 6. COVERINGS OF SPACES AND COMPLEXES 151

Downloaded from https://www.cambridge.org/core. Tufts Univ, on 11 Jul 2018 at 01:30:08, subject to the Cambridge Core terms of use, available at
Cambridge
https://www.cambridge.org/core/terms. Books Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878
CHAPTER 7. COVERINGS AND GROUP THEORY 167

CHAPTER 8. BASS-SERRE THEORY 182

8.1 Trees and free groups 182
8.2 Nielsen's method 188
8.3 Graphs of groups 198
8.4 The structure theorems 204
8.5 Applications of the structure theorems 211
8.6 Construction of trees 233

CHAPTER 9. DECISION PROBLEMS 243

9.1 Decision problems in general 243
9.2 Some easy decision problems in groups 251
9.3 The word problem 254
9.4 Modular machines and unsolvabie word
problems 265
9.5 Some other unsolvabie problems 268
9.6 Higman's embedding theorem 274
9.7 Groups with one relator 281

CHAPTER 10. FURTHER TOPICS 286

10.1 Small cancellation theory 286
10.2 Other topics 289

NOTES AND REFERENCES 291

BIBLIOGRAPHY 297

INDEX 306

1.1 FREE GROUPS

Let A' be a generating subset of a group G. Certain products of
members of X and their inverses will be 1 whatever X and G are; for instance,
xyyz" zy y~ x~ . Other products, such as xyz or JCJC, will be 1 for some choices
of X and G but not for other choices. Those pairs G and X for which a product
of elements in XuX~ is 1 only when the properties holding in all groups
require it to be 1 are obviously of interest.
They are called free groups; a more formal definition will be given
later. If G is such a group, any function / from AT to a group H can be extended
uniquely to a homomorphism from G to //. For any gtG can be written as
x • £ l . . . x. Sn
where s - ±1 and x - e X for r - 1 , . . . ,n. Now suppose that g can
l
also be written as x • . . . x • m, where 8s - ±1 and x . e X for s - 1 , . . . ,tn. Then
J\ Jm Js

and our assumption on G and A' then tells us we must have

U . / ) . . . U . / ) ( j c . / ) . . . ( J C . / ) 1.
z z
l /z Jm J\

Hence the element of H given by {x. ffl ...{x. ffn depends only on g and not
on how g is written as a product of elements of X u X . I t follows that we can
define a function <p:G^> H by requiring g<p to be this element. It is easy to check
that 9 is a homomorphism and that x<p - xf for all x e X. Since every element of
G is a product of elements of X and their inverses, there can only be one
homomorphism with specified values on X. Conversely, suppose that A' is a
subset of a group G such that any function from X into an arbitrary group H
extends uniquely to a homomorphism from G to H. Let x - £ l . . . x. E/z equal 1 in
l
'1 n
G, and let / be a function from X into a group //. Since / extends to a
homomorphism from G to //, it follows that {x • ffl .. .(x. ffn equals 1 in H.
l l
l n

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:28:06, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.002
Free groups 2

Since this holds for all H and /, this amounts to saying that the product of
elements of X u X~ equals 1 in G only if the corresponding products in all
groups are also 1. It can be shown that this holds iff we can derive the fact that
the product equals 1 from the group laws. These observations lead to the formal
definition of free groups.

Definition Let AT be a set, G a group, and i:X->G a function. The

pair (G,i) is called free on X if for every group H and function f:X •* H there is
a unique homomorphism <p:G^ H such that / - i<p.

In particular, the trivial group is free on the empty set, and the
infinite cyclic group Z is free on any one element set ix), where xf is the
integer 1.
The following three questions are natural ones to ask. Do free
groups on an arbitrary set exist? Are free groups on a given set unique? If (G,i)
is free on X is / injective? We leave the first question till later. The answer to
the second question and a partial answer to the third are of an abstract nature,
applying in many situations in abstract algebra and category theory.
It will be useful at times to use the language of category theory,
which frequently sheds light on general situations. (Barry Mitchell once wrote
"The purpose of category theory is to show that what is trivial is trivially
trivial.") For instance, in asking whether free groups on an arbitrary set exist we
are asking whether the forgetful functor from groups to sets has an adjoint, and
the fact that free groups on a given set are essentially unique is just a
uniqueness property of adjoints (or of initial objects in categories). Readers who
have no knowledge of category theory can safely ignore any remarks about it;
such remarks are made only to give extra insight to those with the relevant
knowledge.
Plainly, if (G9i) is free on X and cp:(7 -» H is an isomorphism then
(//,/cp) is also free on X. Our first proposition is the converse of this.

be
Proposition 1 Let (Gj,/j) and (G^,^) free on
X. Then there is
an isomorphism <p:G« •* GL such that /«<p - i~ •

Proof Since (Gj ,/,) is free on X, there is, by definition, a

homomorphism (p:G. •*• GL such that /.cp - /« • Similarly, there is a homomorphism
IJJ.-G^-* G^ such that ^ " ^ • W e t n e n n a v e '"i<P"'i " ' ^ i » where Ij is the identity
function on G^ . The uniqueness property in the definition now requires that

(ptjj-lj, and, similarly, tpcp-I2, the identity function on 6^. Hence 9 is an

isomorphism.//

Proposition 2 Let (F,i) be free on X.

(/) / / there is a group G with an injective function from X to G
then i is injective.
(ii) There is a group into which X maps injectively; for instance,
the set Z of all functions from X to Z, where a + p is defined by
jc(a + P) - jca + jcp for all x.
(Hi) i is injective.

Proof (i) Let f.X-*G be an injection. Then there is a

homomorphism <p:F-* G such that / - /cp. It then follows that / is injective. Notice
that this part of the argument is of an abstract nature, holding in many
situations, whereas (ii) requires a specific example.
y
(ii) Plainly Z is a group. If we define a^ by jca^ - 1 and yoLx - 0
for y *x then the function sending x to a^ is plainly injective. Part (iii) is now
immediate.//

We now proceed to construct a group free on X, beginning with an

auxiliary construction. Let X be any set, and let M(X) denote the set of all
finite sequences (JC. , . . . , J C ; ) of elements of X, where n ^ 0 (the case nm0
l l
l n
corresponds to the empty sequence). Define a multiplication on M(X) by

'l '« J\ Jm *1 '« ^1 Jm

This multiplication is obviously associative, with an identity element which we

call 1 (namely, the empty sequence). Also, JC •* (JC) is obviously one-one, and, if
we identify JC with (JC), every element of M(X) can be uniquely written as a
product JC. . . . J C . for some n; we shall always use this notation. We call M(X)
the free monoid on X.
By a segment of JC . . . . JC . we mean an element JC; JC; . . . JC - ,
l l l l
*1 n r r+l s
w h e r e l^r^s^n; t h i s i s c a l l e d a n initial segment if r-l, a n d a final segment
(or terminal segment) if s - n, a n d it is a proper segment u n l e s s rm\ a n d s - n.
One technical point has been glossed over in this construction.
Since X can be any set, it is possible that some element of X is itself a finite
sequence of other elements of X. We would then want to distinguish between
this sequence as an element of X and as an element of M(X). The simplest way
of dealing with this problem is to replace X by the set X' which is defined to

be the set of all ix) for x*X, and to define the free monoid on X to be M(X').
However, we shall ignore this technicality in future.
We now proceed to construct the free group on X. Take a set A'
bijective with X under a bijection which sends x to x~, and such that Xn X - 0.
We usually denote x by x~ , and we may also write x instead of x. The
elements of M(X v X ) are called words on X. Let w be the word x- £ l . . . x. £/z.
l l
\ n
Then « is called the length of w, written |w| or /(w), and we call the elements
x- £r the letters of w.
The word vu is called reduced if, for 1^ r^n-\y either /' , # / or
/^j - /j. but e^j * -E^; the empty word is also called reduced. Suppose that w is
not reduced, and choose r such that i^ - if and e^j - -e^.. Let wf be the word
obtained from w by deleting the adjacent pair of letters x • Sr and x • e r + 1 . We
say that wf comes from w by an elementary reduction. If w" is obtained from w
by a sequence of elementary reductions we say that wH comes from w by
reduction. It is usually convenient to allow w" to be w in this definition
(corresponding to the empty sequence of elementary reductions); readers will be
left to decide for themselves on each occasion whether this case is permitted or
not.

Examples Let w be the word zxx zy y. Then both zzy y and

zxx z come from w by elementary reductions, and zz comes from w by
reduction.
Let w be zxx' xy. Then zxy comes from w by elementary reduction.
Two elementary reductions may be applied to w, both giving the same result; the
first deletes the pair of letters xx , while the second deletes the pair x" x.

Write, for the moment, w^vv* iff either w is identical to w' or there
is a sequence of words w^,... ,w, for some k such that w^ is w and w, is W and,
for each j<k, one of w-+« and w- comes from the other by elementary reduction.
Plainly ^ is an equivalence relation. We denote the set of equivalence classes by
F(X). Whenever we have an equivalence relation on a set, the equivalence class
of the element w will be denoted by [wl This notation will be used without
further comment (but occasionally the notation will have a different meaning; for
instance, la,bl is used for a closed interval of real numbers).
It is easy to see that if u> v, w, and w' are words such that w % W
then uwv *uw'v. It follows, since ^ is an equivalence relation, that if u'^u' and
w *w* then uw ^u'w' (as both are equivalent to uw'). This enables us to define a
multiplication on F(X), by requirng [wJ[w] to be luw]. This multiplication is

plainly associative, and has an identity (the class of the empty sequence) which
we denote by 1 (just as the empty sequence itself is also denoted by 1). When w
is the word x . El . . . x • £n we let w" denote the word x • ~ £ *... x . ~£l. Then
l l l
i 'l n n \

ww «* 1, as we see easily. Hence F(X) is a group. Define a function i:X->F{X)

by xi - [JCL Plainly F(X) is generated by XL

Theorem 3 (F\X),i) is free on X.

Proof Let G be a group and f:X->G a function. Then / extends to

a function from M{X u X ) to G sending the word JC - £ l . . . JC. £ " to the element
E _ *i *«
l
( JC f) . . .(x f) n. It is immediate that if wf comes from w by an elementary
l l
\ n
reduction then w and w' have the same image. It follows that if w^vv" then w
and vvM have the same image. Hence we may define a function cp from F(X) to G
by requiring \.w]<p to be the image of w. It is obvious that 9 is a homomorphism
and that xf - xi<p. Further, since Xi generates F(X), there can be at most one
homomorphism from F(X) to G with specified values on XL//
Clearly, because elementary reduction decreases length, when we
start with any word and apply elementary reductions one after another in an
arbitrary way until no more can be applied, we will ultimately reach a reduced
word. In particular, every equivalence class contains at least one reduced word.
Theorem 4 (Normal form theorem for free groups) There is exactly
one reduced word in each equivalence class.

Remarks As a special case of this theorem, we find that there is

exactly one reduced word which can be obtained by reduction from a given
word.
Because this theorem is so important, we give several proofs.
Similar proofs apply in many situations.
We have seen that every equivalence class contains at least one
reduced word. The difficult property is that each class contains at most one
reduced word, and we now give proofs of this fact.

Proofs (I) Canonical reduction.

Define a function \:M(X u X) •* M(X u X ) inductively by

IX - 1, xe\ - x e , (uxe)\ - (u\)xs if u\ does not end in x"e, and

(uxe)\-v if u\ is vx'£.

It is easy to check, by induction on the number of letters in w, that wX is

reduced for any word w, that wX is w if w is reduced, and that wX comes from
w by reduction. Also, because u\ is reduced, {ux£x~£)\ - u\ for any u. It then
£ £
follows, by induction on the length of v, that (ux x' v)\ - {uv)\ for all v; that
is, wX-w'X if vvf comes from w by elementary reduction. From this we see that
wX-wMX if w '^wM . In particular, if w and w" are reduced words with w *w"
then w - w\ - w"X - w", as required.

(II)The Diamond Lemma.

We first show that if both w' and w" come from w by elementary
reduction then either w' is the same as w" or there is a word w* which comes
from both w' and w" by elementary reductions.
If w' and w" come from w by different elementary reductions then
there are two possibilities (interchanging vvf and w" if necessary), both of which
were illustrated in the examples earlier. The first is that w is uxsx~£x£v for
some (possibly empty) words u and v, and that w* is obtained by deleting x£x'£
£ £
and w" is obtained by deleting x~ x . In this case w" is the same as w\ The
£ £
second possibility is that w is ux x" ty y v for some (possibly empty) words t,
uy and v, and that w1 is uty y * v and w" is wx e x" e rv. In this case the word utv is
obtained by elementary reduction from both wf and w".
Now take two equivalent words w and w' and consider a sequence
w^y...ywn such that w^ is w, w^ is wf, and, for each /, one of w-+^ and w.
comes by elementary reduction from the other. Suppose that there is some r such
that v^.+1 comes by elementary reduction from w. for all i<r and w. comes from
w-+j by elementary reduction for all i^r. Then w comes from both w and w* by
reduction.
If there is no such r then there must be some k such that both w. ,
and vv, + | come from w, by elementary reduction. In this case, the argument
above tells us that we may obtain a new sequence which also shows the
equivalence of w and w\ either by deleting w^ and w^+^ or by replacing w^ by a
word w* such that both w, , and vv^+, reduce to w*. Since the sum of the
lengths of the members of the sequence decreases when we make this change, it
follows by induction that there must be some word which comes from both w
and vuf by reduction.
In particular, if w and wf are equivalent reduced words then they
must be the same.
A slightly more complicated version of this proof is given in the
exercises, which has the advantage that it can be applied in more general
situations.

(III) van der Waerden's method.

This method is one we shall use in several later theorems.
Let S be the set of all reduced words, and let G be the group of
all permutations of S. We shall define a homomorphism tp:F(X)-> G such that
[w]<p acting on the empty sequence ( ) give the sequence w whenever w is a
reduced word. In particular, if w and w' are reduced words with w *w' then
[vv]-Cwf] and so w and w\ being the results of acting on ( ) by Cw]G. We define xf as the permutation sending w to
wx if w does not end in x and sending w to u if w is ux \ It is easy to
check that this is a permutation of S, whose inverse sends w to wx if w does
not end in x and sends w to v if w is VJC. This holds because a reduced word
cannot end in xx~ or JC JC.
If w is the reduced word JC. E l . . . JC . B/t then Ew]<p is the product of
the permutations {x;f)r and, inductively, the reuslt of acting on ( ) by Cwlcp is
l
r
just w itself, as needed.

(IV) For the final method, we show that if w is a reduced word

distinct from 1 then there is a homomorphism 9 from F(X) to a finite group
such that Cw]cp is not the identity. Since Cll<p is the identity, we cannot have
w **1.
Suppose that this has been shown, and let u and v be reduced
words such that u *v; then uv * 1. The word uv need not be reduced, but if
u and v are different it is easy to see that by reduction from uv we will obtain
a reduced word w different from 1. Further, w * uv % 1, which is prohibited by
the previous paragraph.
So let w be the reduced word x - £ l . . . JC - 6/l . We will define a
homomorphism 9 into the group Sn+i of permutations of { 1 , . . . , « +1} such that
Cw]<p sends 1 into /i+l, whence Cw]cp is not the identity. To do this we require
sucn
a function f.X-* $n+i that (JC^ ) Er sends r into r+1 for all r *n.
So we need xf to send r to r+1 if JC. is JC and e. - 1 and to send
l
r '
r+1 to r if JC. is x and e r - -1. If these conditions define a one-one function
from some subset of (1,...,/z + l) to another subset we can extend this function to
a permutation of ( 1 , . . . ,/*+l), and then define xf to be this permutation.
Can this definition require us to send r to two different numbers?
This
6 could only occur if JC were JC with e - 1 and JC. were also JC with
r-l" "** ^k*s cannot happen; since w is a reduced word. Also this function could
ir r ir.x

only map two different numbers to s if x • were x and e e , - 1 and x- were

l s l l
s-l ~ s
also x with es - -1, which is again impossible as w is reduced.//

We can now obtain a new proof of Proposition 2(iii).

Corollary i:X-*F(X) is injective.

Proof if x and y are distinct elements of X then they are distinct

reduced words. Hence they lie in different equivalence classes.//

Exercise 1 Prove the claimed results about canonical reduction. In

particular, show that wX is always reduced, that wX is w if w is reduced, and
that wX comes from w by reduction. Show also that {uxBx 8)X is the same as t/X,
and, by induction on the number of letters in v, that {ux£x~£v)\ is the same as
(nv)X.
Exercise 2 Let w be uxEx %. If ux£ is reduced we say that uv
comes from w by leftmost reduction. Show that wX comes from w by repeating
leftmost reduction until we get a reduced word.

We may prove the Normal Form Theorem, starting from the first
result in the Diamond Lemma approach, by methods which apply to very general
reduction relations where the previously given induction on length cannot be
used. This is done in the exercises below.
Exercise 3 Let u and v be obtained by reduction from w. Show that
there is a word w* which is obtained from both u and v by reduction. (An
indication of the argument is given in Figure 1. Here the top diamond exists by
the Diamond Lemma approach, the left-hand diamond exists inductively, and
then the right-hand diamond also exists inductively.)
Exercise 4 Let w *>w\ and let w,,...,w be a sequence with w,
being w and w being w' such that, for all /, one of w.+| and w- is an
elementary reduction of the other. If vv^j and w^+j are elementary reductions of
w^we say k is a peak. Use Exercise 3 to show that if the sequence has a peak
then there is another sequence joining w to vvf with fewer peaks. (Figure 2
provides a sketch of the argument.) Deduce the Normal Form Theorem.
Exercise 5 It is possible to give a proof of the Normal Form
Theorem by the Diamond Lemma approach without using induction on the
number of peaks. Consider the relation between words in which u is related to
v iff there is a word w such that both u and v reduce to w. Using Exercise 3,
show that this is an equivalence relation. Deduce that this relation coincides
with the relation *.

Figure 1

5+1

Figure 2

Exercise 6 Show that the function from the set of reduced words to
itself which sends w to wx if w does not end in x and which sends vx~ to v
is a permutation of the set of reduced words.

We usually regard AT as a subset of F(X) with / being inclusion;

consequently we shall usually omit mention of / in future. We frequently
identify elements of F{X) with the corresponding reduced words. At times we
need to regard words as elements of M{X u X) and at other times we want to

regard them as giving elements of F{X). We write w=w' when w and w' are the
same word, while we write w-vv* if they define the same element of F{X) (that
is, if they are equivalent words; we no longer use the notation *).
It is easy to check, without using the normal form theorem, that if
u and v are reduced words then there is only one sequence of elementary
reductions which can be applied to uv to obtain a reduced word. Let u be
JC£I...JC. e
" and let v be x- l
...x- m
. Take 5 ^ 0 as large as possible such
l
'l n J\ Jm
n
that for all r £ 5 we h a v e / - -y and £ „ . , - - e . T h e n t h e pairs x. +i~rx.r
l +
n+i-r r n+i-r r n i-r Jr
m
for r 1, . . . ,5 are successively deleted in reducing uv. The reduced word we
obtain is therefore x£l...x- £n s
' x • hs+l ... x • S « . It follows that it would be
l l
\ n-s Js+l Jm
possible to define F{X) to be the set of ail reduced words with the product of u
and v being given by the above formula. Some authors use this approach, but it
is not really satisfactory. There is a practical problem, in that the proof of
associativity is surprisingly messy with this definition (see Exercise 7) There is a
more important theoretical objection to this approach, as it confuses the question
of the existence of a free group and the question of the nature of normal forms
for the elements. We frequently find in algebraic situations that it is easy to
prove the existence of a free object but extremely difficult (sometimes
impossible) to find normal forms for the elements.
Exercise 7 Suppose that we define F(X) to be the set of reduced
words, with the product of u and v being the unique reduced word obtained from
uv by reduction. (We have seen that this word is unique without using the
Normal Form Theorem.) Show that this product is associative.

A totally different approach may be used to show the existence of

free groups. This approach works in other algebraic situations. It is of a
categorical nature, amounting to the fact that we are looking for an adjoint to
the forgetful functor. Unfortunately the proof runs into some set-theoretical
difficulties. To make the proof easier, I first give the argument in the (incorrect!)
form ignoring these difficulties, and then show how to resolve them.
Consider all pairs {f,GA where G, is a group and / is a function
from X to Gr. Take the cartesian product over all / of the groups 6y, which we
call K. The set-theoretic problem arises here, since in formal set theory this
product cannot be constructed because there are too many functions /; this
problem will be ignored for the moment. There is a function i:X^> K such that
the component of xi in the factor G* is xf. Let F be the subgroup of K
generated by Xi. Let / be any function from X to a group. Then this group is

Gy, and the projection of K on this factor provides the needed homomorphism cp
such that / - i<p. Because F is generated by Xi, there cannot be more than one
such homomorphism.
We now deal with the problem. To begin with, observe that we
need only consider those / such that Xf generates Gy.. For let / be a function
that does not satisfy this property, and let H* be the subgroup of Gy. generated
by Xf. As above, we have the corresponding homomorphism from F to H*> and
this can be regarded as a homomorphism into Gy..
It follows that we need only look at those pairs (/,Gy) for which Gy.
has cardinality at most maxOA'lJINI). This holds because Gf, being generated by
-1
Xf, has cardinality at most that of the set of finite sequences on (Xf)u{Xf) ,
and hence at most that of the set of finite sequences on X u X , which has the
stated cardinality (using the Axiom of Choice). However the set-theoretic
difficulties have not yet been resolved.
Next we note that if we have two pairs (/j,G^) and C/^.G^ a n d a n
isomorphism y:Gj-» G^ such that f\Jmf2 t ^ ien t n e r e i s n o n e e c l t o include both
groups in the product (though we may do so if we wish). For if Gj is included
then we have <PyF->G^ with /j - /<pj and we need only define y^.F-^G^ by
Now take a fixed set S of cardinality max(IXI,IINI). Let (/,G) be a
pair such that G has at most this cardinality. Then there is a bijection % from
G to a subset H of S. We can then make H into a group by defining the product
of Aj and h^ to be {h^r .hr£r )%, and we may use the pair (/$,//) instead of
(/,G). It follows that in our cartesian product we need only consider functions
into those groups whose underlying set is a subset of S. Those readers with
some experience of formal set theory should have no difficulty in showing
that the collection of such / forms a set, and consequently the cartesian product
exists; readers with no knowledge of formal set theory will have to be content
with my assurance that the originally suggested construction cannot be made,
while this modified construction is permitted.

Proposition 5 Free groups are residually finite. That is, if F is

free and 1 ^ w e f then there is a normal subgroup N of F with wiN and FlN
finite.

Proof Let w be given by a non-trivial reduced word. Method (IV)

for proving Theorem 4 then gives a homomorphism cp from F into the finite
group Sn+i for some n such that wep is not the identity.//

Corollary Finitely generated free groups are hopfian. That is, any
onto endomorphism is an automorphism.

Proof This holds for any finitely generated residually finite group
a
Let a be an endomorphism onto G, and let N< G with G/N finite.
Then GIN<xn is isomorphic to G/N for any n. Now there are only finitely many
homomorphisms from G to G/N, since any homomorphism is determined by its
values on the finite generating set of G and there are only finitely many
elements of G/N to which a given generator can map.
Since NoTn is the kernel of such a homomorphism for every n, it
follows that there must be m and-n with m > n such that NOL~n « Not'm'. Hence
N- N<x^mn' , and so N 2 kera. Since G is residually finite, the intersection of
all such N is {1}, and so {U^kera, as required.//

If X is infinite, F{X) cannot be hopfian. For there will be a

function / from a proper subset Y of X onto A", and we can define an
endomorphism of F(X) onto itself with non-trivial kernel as the homomorphism
obtained from the function sending y € Y to yf and x e X - Y to 1.

Proposition 6 F{X) is isomorphic to F(Y) iff \X\* \Y\.

Proof Let / be a bijection from X to Y. Then / extends to a

homomorphism 
from F{Y) to F{X). Now both 9^ and the identity function on F{X) extend the
identity function on X, so the uniqueness in the definition of a free group
requires that <pt|j is the identity function. Similarly <\><p is the identity function,
and so cp is an isomorphism.
Now suppose that F{X) is isomorphic to F{Y). The number of
homomorphisms from F(X) to Z~ , the cyclic group of order 2, is the same as
I vi I yi 1 Y\
the number of functions from X to Z 2 , and so equals 2 . Hence 2 - 2 . It
follows at once that if one of X and Y is finite then so is the other and they
are bijective.
If both X and Y are infinite we use a different approach requiring
the Axiom of Choice. (To obtain the result from the previous paragraph would
require some such set-theoretic principle as the Generalised Continuuum
Hypothesis, and this implies the Axiom of Choice.) It can be shown, using the
Axiom of Choice, that \M(X u X)\- \X u X\ = \X\, and as F{X) is a set of

equivalence classes of M{XvX) it follows that \F(X)\z\X\. Since X embeds in

F(X\ we see that \F(X)\~\X\. Hence, as required, IATI - IKI if F(X) is isomorphic
to F(Y). Note that, without using the Axiom of Choice, we can show that F(X)
is countable iff X is either finite or countably infinite.//

The same argument shows that if A generates F{X) then \A\>\X\.

A group G is called a free group if it is isomorphic to F{X) for

some X. If i:F{X)->G is an isomorphism then Xi is called a basis of G\ we also
say that G is free on Xi. Plainly, if a is a basis of G and p is any
automorphism of G then A$ is also a basis of G. Conversely, if both A and B are
bases of G then 141-151, by the previous proposition, and then any bijection from
A to B extends to an automorphism of G. The cardinality of a basis of a free
group is called its rank.
Further, if a is any automorphism of F{X) then <xz is also an
isomorphism from F{X) to G3 so that Xoni is also a basis of G. In particular, XOL
is a basis of F{X). Any automorphism a comes from a function f:X -> F{X), and
any such function / gives rise to an endomorphism a of F{X). This will be an
automorphism iff there is another function f':X-*F(X), and corresponding
endomorphism a* such that both /a' and / ' a are the identity on X. By the
corollary to Proposition 5, when X is finite a will be an automorphism if it
maps onto F(X). This holds if for every JC in X there is a word w such that

Examples Let F be F{xyy). Then [x~*,x2y) is a basis of F. So also

are [xy,xy } and {xyyyxyxy). In each case, we have an endomorphism of F
sending x and y to the two corresponding elements. This endomorphism is onto,
since it is easy to find an element which maps onto y, and, using this, to find
another element which maps onto JC.
The set iyx,xy } is not a basis of F. For if it were then there
would be a reduced word w such that wa - JC, where a is the endomorphism
2
sending JC to yx and y to xy . But wa-vv', where w' is obtained from w by
replacing x by yx, y by xy , x" by JC" y , and y by y JC" . It is easy to
check that wf is reduced, because w is reduced, and also easy to see, because of
the way w* is obtained from w by substitution, that w' has length ^ 2 unless w
is the identity. However, the next proposition shows that the subgroup of F
2
generated by iyx,xy ) has this pair of elements as a basis, because w1 is reduced
whenever w is reduced.

2 3
Also [xy, x y } is not a basis of F. For if it were then the set
[xyfx y{xy)' } would also be a basis. That is, {xy, x y x~ ) would be a basis. If
we conjugate by x we would get another basis. But the result of conjugation by
x is [yxyxy }, which we know is not a basis.
Let X be any set, Yc X, and let JC e X - Y. Then the map sending y
to yx for y t Y and z to £ for ziY induces an automorphism of F{X), as does
the map sending y to x~ y and z to ?. Any permutation of X also induces an
automorphism of F(X), as does the map sending y to y for .y c Y and fixing the
other elements of X. When X is finite, the latter two types of automorphism,
together with the first type with Y consisting of a single element, generate
Aut F(X); this will be proved in section 8.2.

Proposition 8 Let X be a subset of the group G. Then the

following are equivalent. (/) G is free with basis X, (ii) every element of G can
be uniquely written as x • £l ... x • £fl for some n*0, x- t X, e - ±1, where
L l l r
\ n r
E^. + J * -sr if i ^ - ir, (Hi) X generates G, and 1 is not equal to any product
x • e * . . . x • £/z with n > 0, x . e X, e - ±1, and e . • * - E / / / ^, - / .
i-y in if r r^~\ r r +i r

Proof If (ii) holds then (iii) obviously holds. If (iii) holds then any
element of G can be written as in (ii), since X generates G Also, if an element
could be written in two different ways then, multiplying one by the inverse of
the other and performing all possible cancellations, we would obtain a way of
writing 1 as a non-trivial product of the kind considered in (iii).
If G is free on X then G has these properties, since F(X) does. If
G has properties (ii) and (iii) consider the homomorphism from F{X) to G which
is induced by the identity map on X. It maps onto G, since G is generated by
X. It is one-one, since (iii) tells us that a non-trivial reduced word on X does
not map to 1 in G.II

Corollary 1 Let X generate G, and let <p:G-> H be a homomorphism

which is one-one on X and such that Op is free with basis X<p. Then G is free
with basis X.

Proof Condition (iii) holds for X and G because it holds for ATcp
and Gp.//

Corolllary 2 Let G be free with basis X, and let Y c X. Then the

subgroup <Y> of G is free with basis Y.

Proof Condition (iii) obviously holds for <Y> and Y because it

holds for G and XJI
Corollary 3 Let F be free with basis {x,y}. Let cp:F-*Z send x to
1 and y to 0. Then ker<p is free with basis {x'lyxl, all i € Z).

Proof Plainly all these elements are in ker<p. Also it is easy to

check that any element of F can be written as uxn for some n and some u in
the stated subgroup. Since we must have n - 0 for an element of the kernel, we
see that ker<p is the subgroup generated by these elements .
Now let x- be x'Lyx\ and consider any reduced word
x *l- £ l . . . x i• En. This word equals xJlyElxj2y*2 .. .y£nx^n+l, where j rm i r-\1 - / .r with
n

'0 - 0" z 'fl+r Such a product is not 1, as F is free on {x,y}.//

Let g be an element of F(X). We regard it as a reduced word As

such, it has a length, which we write as \g\. If h is another element of F{X),
which we also regard as a reduced word, the word gh need not be reduced (but
the product gh in F(X) is the equivalence class of this word). If the word gh is
reduced we say the product gh is reduced as written. More generally, if
gy . . . , g^ are reduced words and the word g^ .. .g^ is a reduced word, we say
the product is reduced as written.
It is easy to check the following properties, (i) \gh\ < \g\ + IAI, with
equality iff gh is reduced as written, (ii) Either the first letter of gh is the first
letter of g or g cancels completely in the product gh; this latter holds iff h = g k
for some reduced word k. Similarly, either gh ends with the last letter of h or h
cancels completely in the product gh; this latter holds iff g = kh for some
reduced word k.

Definition Let g = x- l ... x • n be a reduced word. Then g is

cyclically reduced if either in^i^ or inmi^ but zn*-ty Also 1 is cyclically
reduced.

Clearly g is cyclically reduced iff gg is reduced as written. If g is

reduced then, inductively, gn is cyclically reduced as written, and
cyclically rec
\gn\-n\g\.
If g=u~ vu, reduced as written, and v is cyclically reduced, then
gn~u vnu, reduced as written.

El £n
By a cyclic permutation of x. . . . x. is meant any word
l
'1 r-\
Proposition 9 (/) .4/y/ element of F(X) is conjugate to a cyclically
reduced word.
(ii) Any cyclic permutation of a cyclically reduced word is cyclically
reduced.
(Hi) Two cyclically reduced words are conjugate iff they are cyclic
permutations of each other.

Proof (i) Let g - x • E l ... x - E * be reduced but not c y c l i c a l l y

ll n
P P P P
l l 2 n l
reduced. Then g=x. g'x. , where g' is the word x . ... x- ~ . The
l l l
*i \ 2 n-\
result follows by induction.
(ii) The word JC E*JC . £ l . . . x. £/I 1
" is cyclically reduced since, by
l l l
n \ n-\
assumption, either in *i^ or zn 2 -e^. The result follows inductively.
(iii) Any cyclic permutation of the cyclically reduced word g is a
conjugate of g.
Conversely, take any conjugate u gu of g. If u gu is reduced as
written, with u ^ l , then it is not cyclically reduced, since it ends with the last
letter of u and begins with its inverse, the first letter of u . If u gu is not
reduced as written then the first letter of u is either x • El or x. ~E/2. In either
ll l
-1 -1 "
case u gu - v /zv, where lvl<lwl and h is a c y c l i c permutation of g. The result
follows by induction on the length.//
Proposition 10 Free groups are tors ion-free.

Proof Let ge F> where F is free and g^l. Taking conjugates if

necessary, we may assume that g is cyclically reduced. Then as already remarked,
I** l-/il*l*0 f so gn*\J/

Proposition 11 Let F be free, and g and h in F. If gk~hk for some

k*Q then g'h.

Proof We may assume that k * 0 . Taking conjugates if necessary,

we may assume that g is cyclically reduced. Then g is cyclically reduced as
k
written. If h is not cyclically reduced then h is not cyclically reduced, and so
g 2h . If h is cyclically reduced then h is cyclically reduced as written. Hence
g consists of g repeated k times, and similarly for h , so g - h requires g - h.11

Exercise 8 Let u and v be elements of a free group, and let m and

n be non-zero integers. Prove the following results, (i) If uvn - vnu then uv - vu.
(ii) If umvn-vnum then wv-vw.

Proposition 12 Lef F be free, and take g and h in F. If gh - hg

then <g,h> is cyclic; that is, 3u e F and integers r and s with gm u and hm us.

Proof By induction on \g\ + l/zl, the case \g\ + \h\ £ 2 being obvious.
Taking conjugates if necessary, we can assume (without increasing Igl + \h\ ) that
g is cyclically reduced.
If gh is reduced as written then \gh\ - \g\ + l/zl, and so \hg\ - \h\ + \g\.
Hence hg is also reduced as written. Then g is an initial segment of gh and h is
an initial segment of hg. As hg is the same as g/z, one of £ and h must be an
initial segment of the other. That is, interchanging g and h if necessary, we have
g - /zw, reduced as written. We then have hw-wh, and the result follows by
induction.
If gh is not reduced as written we have g- vxs and h~ x~£t, both
reduced as written. Now hg begins with the first letter of h, which is x e , unless
h cancels completely in the product. Also gh begins with the first letter of
g unless g cancels completely in the product. Since g is cyclically reduced and
ends in x£ it cannot begin with x E. So gh and hg are different unless g cancels
completely in gh or h cancels completely in hg. Hence either h - g~ w or g - h w,
reduced as written in either case. The result now follows by induction.//

1.2 GENERATORS AND RELATORS

Proposition 13 Any group G is a quotient of some free group.

Proof The identity map from G to G extends to a homomorphism

from F{G) to Gy which is plainly onto G.I I

Definitions Let G be a group, X a set, and <p:F{X) -» G an

epimorphism. Then X is called a set of generating symbols for G {under <p), and
the family {JC<P;JC € X) is called a family of generators of G. (Plainly G - <<3p>.
Since we can have jccp-^<p with x different from y, we refer to a family rather
than a set.) We also refer to X simply as a set of generators of G; the reader
then has to decide on each occasion whether or not the generators are to be
regarded as elements of G.

We call kercp the set of relators of G (under 9). If u = x. E l . . . x. *n

lft
S S -1
l m
and v = x • . . . x • are (not necessarily reduced) words with uv representing
an element of ker<p, and x z-<p is az•, we say that az • l ... al • n - a • l . . . a • m is
l n J\ Jm
a relation in C?. In particular, if u represents an element of ker(p then the
corresponding relation in G is a- £ l . . . a- £/z - 1.
1
" //
For any subset S of a group //, the normal closure <5"> in H of
the subgroup <5> is called the set of consequences of S in //, or simply the
normal subgroup of H generated by S.
If kenp is the set of consequences of some subset R of F{X) we
call R a set of defining relators of G {under 9). We have a corresponding set of
defining relations. We shall also say that a relation u- v is a consequence of a
set of defining relators (or defining relations) if the corresponding relator uv is
a consequence of the defining relators.
A presentation <X;R>^ of G consists of a set X, an epimorphism 9
from F\X) to G, and a set ^ of defining relators of G under 9. Frequently we
omit mention of 9, especially when 9 is the natural map from F{X) to
F(X)/<R>F(v X)' or when 9 is one-one on X (in which case we regard I as a
subset of G). We write (7- <X\R>^ when <X;R>^ is a presentation of G At times
it is more convenient to replace each relator r by the corresponding relation /"-I,
or, more generally, by a relation M- v where r is wv~ (or v" u). If convenient, we
can mix relators and relations. For instance, if we know that a group G
generated by a set X is abelian, it is easier to write relations xy yx, where x
and y are in X, rather than the corresponding relator x~ y xy, whether or not we
replace the other relators by relations. Examples of presentations of specific
groups will have to wait until after the next theorem.
If both X and R are finite, we refer to a finite presentation and to
a finitely presented group.
Examples The free group on X has a presentation <^;^> with R
empty.
The presentation <x,y\xy *y x, yx * x y> presents the trivial
group. To show this we have to obtain a considerable number of consequences of
the defining relations.
First observe that xy x~ - (xy x~ ) my , and so
x y x~ - xy x~l = {xy2x~l)3-y9. It follows that x2y4x~2 -yx2y4x~2y~l -
2 A 2 6
9 - 1 4 - 9 - 1 7 4 - 1 4 4 6 9 4 - 9 4
yx^y y yx Ly - x*y x J, and so xy -y*x. But then y - xLy x -y . Hence
2 2 3
y - 1. From this and xy my x we deduce that y - 1, and the second relation
gives x - 1.

It would be significantly harder to show in F(x,y) that x and y are

in the subgroup of consequences of the corresponding two relators. In general,
the advantage of using relations rather than relators is that we are working in
F( X)
v
the quotient group F(X)/<R> ' rather than looking in F{X) at congruence
F( X)
modulo the normasl subgroup <R> K '.
Theorem 14 (von Dyck's Theorem) Let G be <X;R>tS?, let f:X •* H
be a function into some group H, and let &:F(X) -* H be the corresponding
homomorphism. Then there is a homomorphismty:G-» H such that xfm xyty for
all x c X if r&- 1 for all rt R. Further, <p is an epimorphism if Xf generates H.

Proof We are told that ^c.ker$, and, since kercp is, by definition,
the normal subgroup generated by Ry we see that kercp£ ker$. It follows, as is
well-known, that the required homomorphism ty may be defined by requiring gty
to be w& for any w such that w, and so Gp = H if Xf
generates H.I I

As a particular case of the theorem, the inclusion of X in X u Y

induces a homomorphism from <X;R> to <X uY\ R vS> for any subset S of
F{XuY). This homomorphism will be used in future without explicitly stating
how it is constructed.
By the same argument as in the theorem, we can prove the
following result, which we will also refer to as von Dyck's Theorem (more
strictly, it is a case of one of the homomorphism theorems). Let R be a subset
of a group A, and let $:A^> H be a homomorphism such that R&- (1). Then there
is a homomorphism ty:A/<R>• -> H such that 9-- 7iip, where n is the
A
homomorphism from A onto AKK> .

When we are looking for a presentation for a group H the following

procedure is often useful. Begin by finding a set X of generators of H. Then
find a set R of relators of H with generators Xy chosen so that one has reason to
believe (or just guesses) that R is a set of defining relators. By von Dyck's
theorem, there is an epimorphism from G - <X;R> to H. It may be that the
relators R enable us to write the elements of G in a simple form from which we
can see that the map is one-one. Alternatively, if H is finite we may be able to
show that G is also finite and that 1(71 si//I, which will again show that the map
is one-one. There is a procedure due to Todd and Coxeter (computer
implementations of which are efficient in practice, though not in theory.) which.

given any presentation <X;R> with X and R finite, will give a bound to the
order of the group (and will even give the multiplication table if required)
provided the group is finite; however the procedure will continue for ever if the
group is infinite, and so cannot be used to tell whether or not the group is
finite.

Examples The cyclic group Z^ of order n has presentation <x; xn>,

since the group with this presentation certainly has an epimorphism to Z^ and is
easily seen to have at most n elements.
Similarly, the dihedral group D of symmetries of a regular «-gon
has presentation <x,y; x n,yi l , (xy) >. For it is clear that Dn is generated by a
rotation through 2%/n and a reflection, and that the stated relators hold between
these generators. We can easily see that Dn has exactly 2n elements. On the
other hand, the group G with the stated presentation is easily seen to have at
most 2n elements, namely the elements xl and xly where 0±i<n. It then follows
that the homomorphism from G onto Dn must be one-one, and that G has exactly
In elements. It was not immediately obvious from the presentation for G that no
equalities between the elements xl and xly for lzizn follow from the defining
relators.
Further, the group D of distance-preserving bijections of E has
2 2 °°
presentation <ix,y; y ,(xy) >. For D is seen easily to be generated by the
functions n ->n + \ and n -»-«, which satisfy these relators. The group with the
stated presentation has an epimorphism to D , and it is easily checked that its
elements are either of the form xl or of the form xLy for some integer /. Since
the groups are infinite, we cannot use a counting argument to show that the map
is one-one. However, we can easily find the images in D of these elements,
and check that only the identity element maps to the identity in D .
Any group G has a multiplication table presentation, which is finite
when G is finite. The generators of this presentation form a set X bijective with
G\ the element corresponding to g is written x . The relators are the elements
-1
x XftXfc for all g, h, and k with gh-k. By von Dyck's Theorem, the group H
relators,
with thisany element ofhasH aequals
presentation in H someonto
homomorphism element
G. Asx a , consequence
and so the of the
relators, any element of H
homomorphism is one-one.
Suppose that G- <X;R>V and that H-<Y;S>^, where XnY-0. Let K
be a group for which there is an epimorphism n:K -> H with kernel G. Define
&:F(XuY) -» K by requiring that JC<{}- xcp for all JC € X and, for all yiY, choosing

yd- to be some element of K such that y$ii~yty. Then, for all s e 5", we have
5-8-TI-5^-1, so that sd- e kenr - G. Consequently we can find an element u of
S
F ( X ) s u c h t h a t 5 # - u 9 - u -9-. F u r t h e r , a s G< K, f o r a n y x * X a n d y * Y w e h a v e
b
-1
(y xy)$tG, and so we can find elements v and v* in F(X) such that
(y lxy)$- v 9 - v $ and (yxy l )#« v* 9 - v* •&.
We
1 can then show
l that
l AT hasl thel presentation
<X,Y; R , i s u b' . s*S). i y ' x y v xy
~ , j > x y ' v * ~xy . for all x € X and yiY)>^.
First observe that for any k*. K we can find we F()0 such that
kit - W4J - WOTT. Hence &(w#) is in ker?r, and so equals w.O for some w« € F(X).
Hence K is generated by ( I u W - The stated relators plainly hold in K.
Let M be the group with the stated presentation, so that $ can be
regarded as an epimorphism from M to K. Take wzF(XvY) with w8--l. Because
of the relations y xy v and yxy - v* there are words wt a F(X) and
vi^ e F(Y) such that vv» ^ ^ 2 in M. Then 1 - W&TT - ^^971.^2^- W2^. Hence, in F(Y),
^2 equals the product of conjugates of 5 or 5 for 5 e S. Now 5 - u in A/, and
any conjugate of u$ equals in M some element of F{X) (because us is in F{X)),
so there must be an element w^ in F(X) such that M^ " ^ in A/. So w-w.w^ in
A/. Then 1 - w$~ (v^w^-fr- (w1w^)9. This shows that vu^w^ is, in F(X), a product of
conjugates of elements of R and their inverses, and so WjW^-l in A/, as required.
This presentation of K can be simplified. Since X$ generates G, so
does (v xy9}, for each ytY, since conjugation by ;# preserves G. Hence, for each
x e X and y * Y , there is some product of the elements v 9 for z * X and their
inverses which equals x<p in G. Let v* in ^(X) denote the product of the
xy
corresponding e l e m e n t s z and their inverses, and let w d e n o t e the product of t h e
corresponding v and their inverses.. T h e n the relation y v* y-w is a
zy 1 xy
consequence of the relations y zy*v . Also w<pmx<p, so that w- x is a
-i
consequence of the relators R. Hence the relation y v* ym x is a consequence
-1 -1
of the other relations; equivalently, the relators yxy (v* ) for all y are
consequences of the other relators. As remarked below, it follows that these
relators can be deleted from the presentation.
Plainly a group G can have many presentations, even for given X
and 9. We now look at how different presentations of the same group compare
with each other.
Let <X;R>^ be a presentation of G. Then so is <J\T;^u5>F^X\ We say that <X,Ru S>9 comes from <X;R>(p by a
general Tietze transformation of type I, and that <X;R>^ comes from <X;R u S><f>
by a general Tietze transformation of type I'. If 151-1, we refer to simple Tietze
transform a tions.

Let Y be a set such that X n K - 0, and let w be an element of

F(X) for each y*Y. Then <X u K; Rv{yu~ , all j>< K}>^ also presents G, where
52 w r et
xty - jeep and JA]; 9- P ° ^ ^ be the normal subgroup of F(XuY) generated by
_1 y
Ruiyu }. Then ty induces an epimorphism 7t: F( X u Y)l N -» G, since
'>.
y - li .
(/?uO>w })I[J-1. But, by von Dyck's Theorem, there is also an epimorphism
$:G -» F\XuY)/N with (jccp)^-JCM Plainly 9-TT is the identity map. Also, since
~ u 4 # - u <p# - 1/ N= yN, ii§ is the identity.
We say that <X u K; / ? u O ^ ~ \ all ^}>* comes from <X;R>9 by a
general Tietze transformation of type II, and that <iX;R^ comes from
<X uY; Ruiyu ' , all .y}>^ by a general Tietze transformation of type II'. If IKI-1
we refer to simple Tietze transformations.
Plainly a general Tietze transformation with 151 or IKI finite can be
obtained by performing a finite number of simple Tietze transformations.

Theorem 15 Any two presentations of the same group can be

obtained from each other by a sequence of general Tietze transformations. If
both presentations are finite then each can be obtained from the other by a
sequence of simple Tietze transformations.

Proof Let <X;R>^ both present G. First assume that
XnY-Q. For each yeY choose u t F{X) with yty - u <p, and for each JC e X choose

Defining & by x&m x<$ and y§m u 9, we get a presentation

<XuY; Ru{yu " , all ;>}> of G, obtained from the presentation <X;R>^ by a
general Tietze transformation of type II. Now y§mu<$y and this equals yty by
definition. It follows that w9*-vwjj for any we F(Y). In particular, 5-9-- sty-1 and
V
JC' 8 " V JC^' wnicn
> b y definition, equals xcp, which in turn is x&.
We find from this that the presentation
-1 -1 9"
<XuY; R,S, {yu },{JCV }> comes from the previous presentation by a general
y x
Tietze transformation of type I. By symmetry, this presentation also comes from
* by a general Tietze transformation of type II followed by one of type I.
Hence <Y\S>* comes from this presentation by a general Tietze transformation of
type I' followed by one of type II'.
Now suppose that X meets Y. We can find a set X* bijective with
X and not meeting X or Y. Plainly G is presented by <A>;/?*>(p*, where R*
comes from R by replacing each x by the corresponding x*, and similarly for
<p*. The previous discussion shows that this presentation comes from eacli of the
other two by a sequence of general Tietze transformations, as required.

Finally, if A\ K, R, and S are all finite each general Tietze

transformation used can be replaced by a finite sequence of simple Tietze
transformations.//

Proposition 16 Let G be finitely generated, and let <Y;S>^ be a

presentation of G. Then there is a finite subset of Y which generates G.

Proof Let X be a finite subset of G with G- <X>. Each x t X is the

product of a finite number of elements of Yty and their inverses. Hence there is
a finite subset Y^ of Y such that X c<Y^>. It follows that Y^ generates G.II
Notice, however, that the set of defining relators for this generating
set is usually larger than S

Proposition 17 Let G have presentations <X;R>9 and <Y;S>^. If X,

R, and Y are all finite then there is a finite subset S^ of S such that G is
<K;5,>*.

Proof We use the notation of Theorem 15. Then G has the

presentation <XuY; T, Uv }> , where T-Ruiyu ~1}. Let t* come from / by
x y
replacing each element x by the corresponding v , and let T* be {/*; all f . J ) .
Let N be <xx>x l>F(XuY\ a n d let n:F(XvY) •* F{XuY)/N be the natural map.
Since v TC - JC7i, we see that t*n - tit. Thus t* is a consequence of t and the
-1
relators JCV . It follows that a Tietze transformation of type I gives us a
_ i a.

presentation <XuY; T, 7**, (JCV " }> , and then a Tietze transformation of type I'
gives the presentation <X uY; T*y (xvx~ }> . Now vx is in F(Y), and so
T*cF(Y). So we may apply a Tietze transformation of type II' to get the
presentation <K;7>>*.
Since X, Yy and R are all finite, T* is finite. Since T*£<S>F^,
each element of T* is the product of finitely many conjugates of elements of 5"
and their inverses, and we choose one such expression for each t*. Let S^ consist
of those elements of S which occur in some such expression. Then S^ is finite,
and as ken[; equals <5*> ' ^ and also equals <T*> ' ' it must also equal
C S 1 ^ ^ , as required.//
3 2
Exercise 9 Show that Z^ can be presented as <y9z\ y , Z • yzm zy>-
Exercise 10 In F{x,y), write x as a product of conjugates of
xy x~ y and yx y x~ and their inverses.

Exercise 11 Show that the following are presentations of the trivial

group.
(i) <aM abnaX - bn*\banb'1 - an+1>.
(ii) <ayb\ an-bn*1, aba-bab>.
(iii) <a,b,c; aba'1 - b2, bcb~l - c 2 , cac'1 - a2>.
(iv) The group with generators xty, and z, with defining relators
-1 -2 -1 -2 -1 -2 -1 -2 2
(JC^X j ; ){yzy Z )(xyx y )m(yzy Z ) and the two other relators obtained by
permuting JC, y, and z cyclically.
Exercise 12 Find presentations of the symmetric group SV and the
n
alternating group An for n < 5.

1.3 FREE PRODUCTS

Definition Let {G^} be a family of groups, G a group, and let
: e
'<x ^a"*^ ^ homomorphisms. Then (G, U^}) is called a /ree product of the
groups (7a if for every group H and homomorphisms / a : ^ a "^ # there is a unique
homomorphism /:<?-» // such that / a » / ^ fo all a.

As with free groups, we ask if free products exist, if they are

unique, and if the maps i'a are monomorphisms. The proofs of Propositions 1
and 2(i) generalise immediately. Because of the uniqueness, we usually refer to
H
the free product** of the groups C?a, rather than to **a free product**.

Proposition 18 / / (Q{i a » and (H,Ua)) are both free products of the

groups <?a then there is a (unique) isomorphism f:G-> H such that igf'Jq, f°r a^
a.//

Proof There is a homomorphism f:G-> H such that Z^/"- y'a for all a,
and also a homomorphism /*://-» G such that Jof*m *a for all a. Since i^ff m *'a
for all a, the uniqueness of homomorphisms in the definition ensures that ff is
the identity map on G9 and similarly /*/ is the identity on H.I I

Proposition 19 Let (GM^) be the free product of the groups G^.

Then (i) i is a monomorphism iff there is some group H and homomorphisms
fo-Ga -* H for all P such that / a is a monomorphism; (ii) z*a is a monomorphism
for all a.

Proof (ii) follows from (i) by taking H to be G^ with / being the

identity function and all other /„ being trivial, while (i) holds because there is a
homomorphism / with / a - i^fll

Theorem 20 Any family of groups G^ has a free product.

Proof Let G^ have presentation <X^R^y^*. We may assume

(replacing each A" by a set bijective with it, if necessary) that X n Xa - 0 for
F(\\y \ « P
rv
a*p. Let G be the group F(U A^KUtf a > <*', so that G has presentation
<U*a; UJ?^*, where 9 is the natural map. By von Dyck's Theorem, the inclusion
of X in U X^ induces a homomorphism *a from G^ to G. We shall show that
(C7,{i'a}) is the free product.
So let / a : ^ a " ^ ^ b e homomorphisms. Now / induces a
homomorphism I|J from ^(A a ) to // such that ^ a ^ a - l . There is then a
homomorphism ip from FiUX^) to // such that t|) is 4>a on ATa. Since (Uâ)4> - 1, ty
induces a homomorphism f:G+ H with cp/-<Jj. Also, for every ^ Q J 6 ^ *
(xacpa)/of/'- * a <p/- ^ a ^, by definition of / and / a , and *a<J>" ââ" XQÔ/OL9 b y
definition of ip and ^ a - Since ^ a <P a generates 6^, we see that / - if. Also / is
unique, since G is generated by UX^î^J/

The free product of the family of groups G^ is denoted by * G^,

while the free product of two groups is usually written as G^*Gj\ more
generally, the free product of n groups G^ is written G^*... *G^ If we regard
the free product as being given by the presentation of the previous theorem, we
see that Gj*G^ is the same as GfG^, while GfGj*G^ is the same as (GfG^)*G^
and as 6^ * {G^G^. These identifications, and other more complicated ones, will
be made without further comment. Further, since each z'a is a monomorphism, we
usually regard G^ as a subgroup of *GU with z'a being inclusion, so that we do
not need to mention it.

Examples The free group F(X) is the free product of the groups
Cx for all JC e AT, where Cx is infinite cyclic with generator JC.
The free product of two cyclic groups of order 2 has presentation
<a,b; a2,b2 >. Applying Tietze transformations, this group may also be written as
<afb,c; a2,b2,c~lab>, as <a,c; a2, {cla)'2\ as <a9d; a29{da)29 or <a,d; a2f{ad)2>9
or as <a9c; a , a cac>, which shows that this group is the infinite dihedral
group.

Let {G^} and {H^} be collections of groups, and let 9 ^ ^ " * ^ &e
homomorphisms. Let G be *<?a, and let H be *// a - Regarding H as a subgroup
of #, the homomorphisms <pa from 6^ to // give rise to a unique
homomorphism from G to //, which we denote by *<pa.

Theorem 21 (Normal Form Theorem) Let (G.ti^))) be the free

product of the groups G^. Then (/) each z*a is a monomorphism; (//) regarding i
as inclusion, any element of G can be uniquely written as gA ...gm, where
l n
S^O^ for some a^, g. * 1, and a^. * a^j for r < n.

Proof Temporarily denote £ *a by £ a for Sa€<2a- We prove both

parts (the first part has already been proved by another method) by showing
that any u c G can be uniquely written as g j . . . g , with n> 0, g^e G^ , ^. J* 1,
and
Since our construction of G shows that U(? a / a generates G, any «
can certainly be written as g^...gn , with n*0, g^G^t g^ * 1, where a r " a f + j
is permitted. If 0Lrm°Lr+i anc^ ^ f + i * £ r t n e n w e m a y write u as
^•••*M*/+2'••*/! ' W h e r e X *h*8rgr+\*G0L; w h i l e if a r " a ^ l a n d ^ ^ r ^ " 1
we may write u as g j . . &r-\8r+2 '' '&n' ^ y induction on n, there is at least one
way of writing u in the required form.
To prove uniqueness we follow van der Waerden's method (method
III of Theorem 4). Let 5 be the set of all sequences (g^...9gn) with n>0,
O 9 1, and oCj. / a f + p *n particular, the empty sequence () is in S.
Take g e G - {1}. A map from 5 to S is defined by mapping
(gv • • • ,gn) to (gv . . . ,gn,goL) if a * o^,
if a
(8V • •. ,*„) to (gv ... , ^ . 1 ' W " % a n d V a ^ X'
to telf . . . . ^ j ) if a - *n and g^ - 1.
This defines a function <pa from C?a to the set of maps from S to
itself (defining l<pa to be the identity). It is easy to check that cpa preserves
multiplication. It follows that <pa maps into Sym5, the group of permutations of
5, since £atpa has inverse £a~ 9 a .
Let <p:G -> SymS be the homomorphism such that cpa» i^y for all a.
wit
Take acGand write it as ^---^ h h. * 1, A . E ^ , and a^. * a ^ . It is
easy to see (by induction) that u<p applied to the empty sequence () gives the
sequence (Aj,... ,A ), and so hy . . . ,hm are uniquely determined by u.ll

An alternative proof of uniqueness may be given along the lines of

method IV of Theorem 4, and this enables us to obtain results about residual

finiteness of free products. Take some particular u expressed as in the

previous paragraph. Let Sm consist of those elements of S with n £ m. We can
define a map 4>a from 6^ to the set of functions from Sm to itself by using
the previous definition except in the case where n - m and a * a^; in this case
g a is required to send (gy...,gm) to ( g j , . . . ,gm). As before, we get a
homomorphism <J> from G to SymSm such that u<>
| is not the identity.
The advantage of this method is that S is finite if there are only
finitely many 6^ all of which are finite. Thus we have proved the following
result.

Proposition 22 The free product of finitely many finite groups is

residually finite.//

In fact a stronger result is true. If we take the free product G of an

arbitrary collection of residually finite groups 6^, and an element utG-{\}, there
is a homomorphism <p from G onto the free product of finite groups G^, all but
finitely many of which are trivial, such that u<p 2 1 (see Exercise 17). Applying
the previous argument to u<p, we get the following.

Corollary The free product of residually finite groups is residually

finite.//

We have an analogue of Proposition 8.

Proposition 23 Let G^ be subgroups of a group G. Then the
following are equivalent, (i) G is the free product of the subgroups G^; (ii)
every element of G can be uniquely written as g^ — -8n, with n ^ 0, g^tG^ ,
gj * 1 , and a- * a/+i»* ( " 0 & / 5 generated by the subgroups G^ and 1 cannot be
written as a product g*...g with n > 0, gtG , g. & \, and a- # <*+i-

Proof (ii) and (iii) are equivalent, as in Proposition 8. If (i) holds

so does (ii), by the Normal Form Theorem. Suppose that (iii) holds. We know
that there is a homomorphism from #C?a to G which is the inclusion on each G^,
and (iii) tells us that this homomorphism is onto and has trivial kernel.//

Let u in %G^ be g^...gn, with g^G^, g. * 1, and a^. * a ^ j . We

say that u has length n, and that the product g^...gn is reduced. If v is
written as a reduced product h^... hm, we say that uv is reduced as written if
gn and Aj are not in the same G^. If uv is not reduced as written, we say there

is cancellation if gnh^ - 1 , while we say there is coalescence (or amalgamation;

but this latter phrase is in danger of being confused with another concept) if
g ft* £ 1. Plainly luvl-lul +Ivl if uv is reduced as written, while \uv\ - I til + Ivl- 1
if there is coalescence. If there is cancellation then luvl< lul + lvl.
With u written as above, we say u is cyclically reduced if n - 1 or
oin * <Xj. As with free groups, every element is conjugate to a cyclically reduced
element. If v is cyclically reduced and Ivl > 1 then Iv^l - n\v\. It follows that any
element of finite order in a free product must be in a conjugate of one of the
free factors.
The exercises indicate an alternative way of constructing free
products, similar to that used for free groups.

Exercise 13 Show that the function <pa defined in the proof of the
Normal Form Theorem, and the function 4>a used in proving Proposition 22
preserve multiplication.
Exercise 14 Let {G^} be a collection of groups such that C7an<jU-l for
a * p. Let M be the free monoid A/(UC?a). The word wf is said to come from w
by elementary reduction if w is 8\---8n anc * either vvf is g^... £;_i^S;+2 • • • &n
where g^ and g^ are in the same factor G^ and h is 8fi^ or else wf is
mg w g g ein
Si • • •S-i£/+2* * n **k i i+l ^ ^ * (note that the second type of reduction is
obtained by two applications of the first, except for the reduction of the
sequence 1 to the empty sequence). Obtain an equivalence relation from this as
in the construction of free groups. Show that the set of equivalence classes is
the free product.
Exercise 15 Using the construction of the previous exercise, prove
the Normal Form Theorem both by canonical reduction and by the Diamond
Lemma.
Exercise 16 Let G^ be the free product of the groups G^ for a € A,
and let G^ be the free product of the groups G^ for a € B, where B c A. Show
that Gg is a subgroup of 6 ^ , and that there is a homomorphism from G^ to Gg
which is the identity on Gg.
Exercise 17 Let G be a residually finite group, and let S be a finite
subset of G. Show that there is a normal subgroup N of finite index in G such
that Nr\Sm0. Use this to show that if the groups G^ are residually finite then
for any u * 1 in their free product there is a homomorphism 9 from the free
product to a free product of finitely many finite groups such that w<p^l. Deduce
the corollary to Proposition 22.

Exercise 18 Let G- A* B, where A{\)B. Let gtG be such that

Ang~ Agtil). . SShow that geA. Deduce that A contains no non-trivial normal
subgroup of G.

1.4 PUSH-OUTS AND AMALGAMATED FREE PRODUCTS

Definition Let G^9 G^ and C?2 be groups, and let iyG^-* G^ and
'*2:^b"* ^2 ^ e homomorphisms. Let G be a group, and let j\:CL •* G and y'o-'^o ~*
be homomorphisms. We call (Gyj^j^) the push-out of (*p ^ ^
(1) /jy'j - / y 2 '
(2) for any group H and homomorphisms <Pr:Gr-* H (rml,2) with
e r e IS a u n ue m
'l^l"'2^2 ^ i Q homomorphism <p:<7-» / / such that <Pr J<p (r-1,2).
We then say that we have a push-out square

Gn + G
2
h
As usual, the push-out is unique up to isomorphism.

Theorem 24 Any pair (iyi^) has a push-out.

Proof For r -1,2, let Gf have presentation <Xr;R^r9 where

* j nA^-G. Let K generate GQ, and choose, for all yeY, w e F(X J such that
yifmw &r. Let C7 be the group with presentation O ^ u A ^ ; ^ j , /?2» ^wy\ w
y2^'
Thus we have the natural maps j from 6y into G , induced by the inclusions of
Xf into X^ u * 2 , and C? is generated by G^/j u G^2 • Hence there can be at most
one homomorphism from G with specified values on GJ^ u
Suppose that we have homomorphisms <pf from Gf to a group / /
such that /j<pj - /*2<p2 • Then <pr defines a homomorphism ^ from F(X^ to / /
which is trivial on Rf. By von Dyck's Theorem, the homomorphism from
F(X* u X^) which is tyf on Xf defines a homomorphism 9 from G to H, which,
by construction, satisifes 7,9-9,..//

Let 6^ be trivial. By using the above proof, or by arguing directly

from the definition, we find that the pushout is the quotient group

The maps j^ and j ^ need not be one-one, even if one of /j and i^

is one-one. For instance, suppose that G^ is simple, and that L is one-one and
z*2 is onto but not one-one. Take VVCC/Q with w / 1 but ^ " l - Then w / j * l but
onto an
vv/jy'j - 1. As C7j is simple, this ensures that y'j is trivial. Since z'2 is d
s tr v al ol ws tnat s triv
'2^2'Vl * * i » ** f l ° 7*2 * ial- Hence the push-out is trivial.
are
When both i^ and ^ one-one, the pushout G is called the
amalgamated free product of Gj and C?2 w i ^ &Q amalgamated. In this case we
usually regard GQ as a subgroup of Gy and 6^, and regard /j and i^ as
inclusions. The usual notation for this situation is G* *u 6U. Sometimes it is
mor convenient to use the notation G^ Q*fj G^ where GQ^CGJ, HQ£G2> /J is
an
the inclusion, and ^ is isomorphism between GQ and HQ. For more precision,
we could mention the specific isomorphism from GQ to / / Q .
For an amalgamated free product we shall see that j^ and j \ are
one-one, and we regard them as inclusions. We shall also see that, with this
convention, G ^ n G ^ " ^b'
Let G be A*c B. Let 5" and T be left transversals of C in ^ and 5
with \iS(\T (that is, 5 contains one member of each coset aC). We shall obtain
a normal form theorem using S and T.
As in Propositions 2 and 19, in order to show that the maps j ^ and
jg from A and 5 to G are monomorphisms it is enough to find some group H
and monomorphisms <p:A-* H and ty:B~>H such that 9 - ^ on C. We can take H
to be the group of permutations of S*T*C. Let a4, and so a bijection of SxT*C with rxy4,
under which a similarly. It is easy to check that both 9 and ty are
one-one, and that 9 - ty on C, as required.
If we take B to be the same as A then, provided C2A, we obtain
two distinct homomorphisms from A to H which agree on C. Also H will be
finite if A is finite.
In an arbitrary category a morphism Ty.X ->Y is defined to be an
epimorphism if, for any Z and morphisms 9 and 4> from Y to Z, the condition
7)9 - T|ip implies that 9-4). In the category of rings with identity, the inclusion of
Z in Q is easily seen to be an epimorphism which is not an onto map. From the
preceding paragraph, it is easy to check that in the category of groups (or of
finite groups) epimorphisms are the same as onto homomorphisms.
Similarly the morphism TJ is a monomorphism if for any morphisms
9 and t|> from Z to X such that <pr\ - cpr) we have 9 - <!>. It is rather easier to

prove that monomorphisms are the same as one-one homomorphisms in the

category of groups (or the category of finite groups). We need only observe
that if H<G then the trivial map and the inclusion map from H to G give the
same result when composed with the map from G to G/H.

Theorem 25 (Normal Form Theorem) Let G be A*CB. With the

above notation
(/) jj^ and jg are monomorphisms,
(ii)AJAnBJB-CJA-CJB,
(Hi) regarding j * and j« as inclusions, any element of G can be
uniquely written as u^...unc, where n>0, aC, and u^,...,un come alternately
from S-{\) and T-{\).

Remark By the definition of an amalgamated free product, we know

that j . - Jg on C. Also, regarding J* and Jg as inclusions, (ii) can be written as
AnB-C

Proof (ii) follows at once from uniqueness in (iii).

We will temporarily denote aj^ and bjg by a" and b . As in the
normal form theorem for free products, to prove both (i) and (iii) it is enough
to show that for any g*A*gB there are unique u^9...9u 9 and c with
c e C, and u^9...9un alternately from S - {1} and T- (1) such that gm u^ ... u nc.
Now any g can be written as 5 j . . . ^ ? where g.^AuB, for some k.
If gy and g. + j are both in A or both in B we may write g as
wnere n
#1 • • • £/-i^$"/+2 * * •£#» ~ gig/+\ • Continuing like this, we see that g can
be written either as c~ or as g^... gn> where g^,... ,gn are alternately from A - C
and B- C,
Inductively, write g " j . . . g"n_j as u^... u~n ^c, where u^(S uT)- {1}
and ttj€gjC and u- e Cg-C for each i > 1, so that u- comes alternately from
S - {1} and T- (1). Then

m i ' * * ^n-l'fn " 5 ' - Vl*"'m where h' cgn>

which equals u....u
l n-l*u nd, where h u nd with numiSvT.

Since und - cg^ , we have un € Cfc C, as required. Also un * 1, since gn 4 C. So

we have shown that every element of G can be written in the required form.
We prove uniqueness by van der Waerden's method (another proof
is given in the next section). Let X be the set of all sequences (Uy...9un,c)

with n * 0, ceC, and u^9 ...,un alternately in 5 - { l ) and T- (1). We define a

homomorphism <p from G to SymA" such that, when g is written as u^ . . . iT^tf
of the required form, the action of g<p on the sequence (1) is the sequence
(«j, ...,1/^,0). This will prove uniqueness of the representation of g.
To construct 9 it is enough, by definition of the amalgamated free
product, to define 9 on A u B so that it maps into the set of functions from X
to X and is multiplication-preserving on A and on B (then it automatically maps
A into SymAT, since 0 9 will be the inverse of a<p). We define 9 on A by
requiring the action of a<p on (u^9 . . . ,ti^,c) to be

(u1v . . . ,t/ ,v,tf) if H *<4, camvd with v e ^ - C l } and rfeC,

(«i,
i ••• n>u.d) n if a*.<4 and ca-d with tfeC(that is, if aeQ
(uu . . . ,wM i,v,tf) if "M€^4 and « M c a - v ^ with vzS-il) and J e C ,
(u+9 . . . ,uM *,d) if a €,4 and wMcam d with <i€C.

Readers are left to check for themselves that (aa')<p equals

(a9)(tf f 9). We define 9 on B by similar formulae. It is easy to check that the
definitions of 9 on A and on B agree on C.I I

There is an alternative version of the Normal Form Theorem, using

right transversals instead of left ones, which may be proved symmetrically.
Because of the inconvenience of working with transversals (and the fact that
they can be arbitrary transversals) the weaker theorem below is often more
useful than the Normal Form Theorem.

Theorem 26 (Reduced Form Theorem) With the above notation, and

regarding j * and j« as inclusions,
(/) any wiG-C can be written as g* ... g , where n*l, and the g-
are alternately from A-C and B- C,
(ii) if we can also write w as h^ ... hm with the h. alternately
from A- C and B- C then mm n and h^ cg^C, hn€ Cgn, and h^€ Cg-C for all other
i,
(Hi) if n > 1 then wiAvB,
(iv) such a product cannot be in C.

Proof In the proof of the Normal Form Theorem we proved (i). We

also showed that if w is expressed as in (i) then its normal form is u^ . . . unc
u c and
where UjcgjC, n* 8n> u.*Cg.C for all other /. Then (ii) follows at once

from the uniqueness of the normal form. The Normal Form Theorem also tells
us that such a product cannot be in C, and can only be in AuB if / i - l . / /

It is easy to see that the Normal Form Theorem could be deduced

from the Reduced Form Theorem (and the fact that j * and y « are one-one), if
we had an alternative proof of the Reduced Form Theorem. Readers may wonder
if one of these theorems (and the residual finiteness of suitable amalgamated
free products) can be proved along the lines of Proposition 22 . Such proofs do
exist, but the only one I know is notationally complicated. We shall give
another approach in the exercises to the next section.
If w is written as g^ . . . gn with the g. alternately from A - C and
B- C we refer to this expression as a reduced form of w. Then the integer n
(which we have just seen depends only on w and not on how it is written as
such a product) is called the length of w. Thus w has length 1 iff it is in
(A- C)\)(B- C). An element of C is said to have length 0. We have results
similar to those for free groups and free products about the length of products.
Also any element not in C is easily seen to be conjugate to an element
gj . . . gn in reduced form with g^ and gn in different factors if n * 1. It follows
that w has finite order only if it is conjugate to an element of A u B.

Proposition 27 Let A and B be subgroups of a group G, and let C

be An B. Then Gm A*gB iff every element of G- C can be written as a product
8 g w tn tne g a ternate rom
\"' n * i ^ ^y f A- C and B- C, and no such product is 1.
Proof We know that A*gB has this property. Let G have the
property. The inclusions of A and B into G give rise to a homomorphism from
A*gB into G. The conditions tell us that this homomorphism is both one-one
and onto.//

Proposition 28 Let G be A* CB. Let A^cA and B.cB be such that

v4j n C - /?j n C - C j , say. Then the subgroup <A^ ,B^> of G is A^ * g B^. Also
C4j ,By>nA-A^ and <A^ fB^>nB- B^.

Proof The first part is immediate, by applying Proposition 27 to

O4j 9B^>. The second part then follows from part (iii) of the Reduced Form
Theorem, as this tells us that the product of elements alternately from A^ - Cj
and 2?j - Cj cannot be in A if its length is >1.

Proposition 29 Let G be A* CB. If G and C are finitely generated

then A and B are finitely generated.

Proof Write each element of G as the product of finitely many

elements of A u B. Let A^ be the subgroup of A generated by C and the finitely
many elements of A which occur in the chosen expressions for each generator of
G, and similarly for B^. Then Gm <A^ tB^>, by construction. Thus Proposition 28
tells us that Am A^ ,which gives the result.//

Remark More generally, given groups G^ and monomorphisms

i^.C-^G^ we can define the amalgamated free product *r,Gfa» with similar
properties. In particular, if C-(1), this group is just the free product of the G^.
The general case is perhaps best regarded as a special case of the tree products
of section 8.3

Exercise 19 Suppose that we have a pushout square with 6^ trivial.

G
Show directly from the definition that G is

Exercise 20 Suppose that we have three groups G^, 6 ^ , and 6 ^ ,

and subgroups H of Gf for rml9 2, 3, together with monomorphisms / from H
into Gy+j (where G^ is C7j) such that H^n Hr^m {!) for all r. Give an example
to show that if we take the quotient of C7j * G^ * G^ by the normal subgroup
generated by {(hr<$>^h~ , all r, all hfiH} then the resulting group may be trivial
(one of the earlier exercises gives an example of this situation). Show that if we
make the corresponding construction using k>3 groups then the natural map
from each Gf to the quotient group is a monomorphism.

1.5 HNN EXTENSIONS

In 1949 Higman, Neumann, and Neumann studied a construction
related to amalgamated free products. This construction is therefore now called
an HNN extension. Later Britton (1963) used this construction in his proof of the
unsolvability of the word problem for finitely presented groups (see Chapter 9).
In recent years it has become clear that HNN extensions (originally obtained as
subgroups of certain amalgamated free products) should be treated in their own
right as one of the basic constructions of combinatorial group theory.
This time we shall not use a universal property in the definition,
but will obtain this property later. Let G and A be groups, and let / and /- be

monomorphisms from A to G. Let P be an infinite cyclic group with generator

p. Let N be the normal subgroup of G*P generated by the set {p (ai)p(aL) )
where a runs over A (or just over a set of generators of A, since this provides
the same subgroup N). Let H denote (G* P)/N Then H is called the HNN
extension of the base group G with stable letter p and associated subgroups
Aiq and Aiy We usually take A as a subgroup of G with /Q being inclusion, and
we write H as « / , p ; p~ Ap - Z?>, where /? is Ai^. This notation, however, does
not make ij explicit, and we may want to be more precise by writing H as
<G,p; p'lAp-Aix>.
Let G have presentation <AT;/?>(p with piX, and let K be a subset of
/"(A") and let % be a one-one map from Y into F(A> Let Y. Conversely, a
presentation of this form is an HNN extension provided that & induces an
isomorphism from <K to <Y&.
Let gQ and g^ be in C? and define j ^ and y'j from A to C/ by
aJrm(air)8r for f-0,1. Then the HNN extensions <G9p; (ai^)p- ai^> and
m
<G,q; (OJ'Q)** aj\> are easily seen to be isomorphic, with the isomorphism
sending g to g and p to ^ ^ "
More generally, we can take a family of groups A^ and
monomorphisms Uuoc and /,l o t from A ex to C7. Let P be free on {p
oc }, and let the
1 1
normal subgroup of C7*/* generated by {pa~ ( a **oa^a^ "la' ' ^ a ' a ^ aoL£j^<x)
a a/ a

be TV*. Then H-(G* P)/N is called the HNN extension of the base group G with
stable letters (p a ) and associated pairs of subgroups A^i^ and ^ a * l a - The
notations and remarks of the previous two paragraphs can be extended to this
more general case. It is occasionally useful to consider the same construction
when the maps i^ and i^ are not monomorphisms; in this case we refer to a
pseudo-HNN extension.

Examples The free group with basis (p a ) is the HNN extension of

the trivial group with stable letters p a .
Let G be any group, andf let A be a subgroup of G. Then we can
form the HNN extension <G,k; k akm a for all a*A>.
-1 2
The group with presentation <a9b\ a bam b > is the HNN extension
of the infinite cyclic group <£> with stable letter a and associated subgroups 
and <b2>.

The inclusion of G in G* P induces a homomorphism j;G -» H, which

we shall see later is a monomorphism. The universal property of HNN
extensions is given in the proposition below.

Proposition 30 Let 9 be a homomorphism from G to a group K.

Suppose that K has elements k^ such that k^ ( ^ O a ^ o c " ' V l a ^ f°r al1 a and
all aa € i4a. Let H be the HNN extension of G with stable letters p a and
associated subgroups ^ a ' Q a <*nd ^ a ' j a -Then there is a unique homomorphism
ty:H •* K such that jty - <p and pjty - &a for all a.

Proof Since H is generated by Gj and {pa> there can be at most one

such homomorphism. There is certainly a homomorphism from G* P to K which
is 9 on G and which sends p to kn, and this will send the elements
-1 -i
P~ (a~if\~)P~(a~h~) to
1- Since the normal subgroup of G* P generated by
CX CX U v X OC Ot IOC

these elements is the kernel of the natural map from G* P to //, this
homomorphism will induce the required <|>.//
As usual, there can be only one group (up to isomorphism) with
the property of this proposition; hence, if we construct a group with this
property we know it must be the HNN extension.
In particular, there is a homomorphism from H to P sending each
P a to itself and mapping G trivially. It follows that the subgroup of H
generated by the image ^-P^N) of (p a ) is free with these elements as basis.
Because of this, we regard P as a subgroup of H.

Let H be the HNN extension <QPa; Pa ^ a ^ a ' ^ - a ^ » w ^ ^ a n

isomorphism i^ from A^ to A^. Let S^ and S_a be left transversals of A^ and
j4_a in G, all containing 1. We have the following Normal Form and Reduced
Form theorems.

Theorem 31 (Normal Form Theorem) (0 The homomorphism j is a

monomorphism. ( / / ) Regarding G as a subgroup of H, any htH can be uniquely
written as

where n * 0, e.-±l (for nm0 the expression is just g^) and

g. € 5 a if e / - 1 and g( € S if £; - -1,
/ / oc^j - o^. and e^ - -e;- then g. * 1.

We say that an element gnp°g . . . g» P» Zft~l gm of G P , where

/z ^ 0 , ei. -i ± l , geG has a /?/«cA if there is some r such that a r - a r~v, and either
e f - - e f _ j - l and 8r*Ar or *r" ~£r-\m ~l anc ^ 8r*A_r. When G has presentation
<Ar;^>, it is also convenient to say that an element of F(ATu{p a , all a)) has a
pinch if the corresponding element of G* P has a pinch.

Theorem 32 (Reduced Form Theorem, or Britton's Lemma)

(/) Regarding G as a subgroup of H, any h c H can be written as

where n * 0 , e«- ±1, g^G and the corresponding element of G* P has no pinch,
(ii) If h can also be written as

with similar conditions on the v. then mm n and, for all i, a.- P^. and t-- r\..
(///) / / h has an expression as above with no pinch and n>0 then
hiG.
(iv) / / the product Ztf>^&x... V l p a ^ ' l g « ' with 8i'G' e / " ±1
'
and n > 0, equals 1 rAe« rAe corresponding element of G* P has a pinch.
(v) / / a product SQP^SI • • • ^-i^a^"1^' w/r/r
^/ € ^ e
/ " ±l> and

n > 0, //^5 //i C? Me« Me corresponding element of G* P has a pinch.

Remark In both theorems, the case « - 0 gives an element g Q of C7;

the product g^p^0 corresponds to n - 1 with gj - 1.

Proofs Parts (iii) and (iv) of Theorem 32 are immediate from (ii).
Part (v) follows from part (iv) since go...gnmg can be written as
h
We use van der Waerden's method to prove that j is a
monomorphism and to prove that an element of H has at most one representation
of the form in Theorem 31(ii). So we consider the set W of all sequences
0,s1,...,£^1,<x/I_1,E/r_1,s/z) satisfying the conditions of Theorem 31(ii).

We can obviously define a monomorphism cp:<3-> SymW by

(g0,... .E^J , gn).g<p - U o , . . . , 6 ^ , g^g). We can also define a function n^W* W
by

unless a^.|"«, e^^j - -1, and 8n^Aa, where we write g^ as sa for s c S a and
<2€^4a and £"tf*ia> while

when n >1, a ^ j - a, e^ j - - 1 , Sn^AQL and ^"S^^f and, finally,

e
(^O»a,-l,^j)7r a -£Q6 when # i ^ a and £ " # i ' i a - (Note that the case « - 0 is
covered by the first situation).
It is easy to check that 7i a is a permutation of W and that

<*«*)*«• V V l a * } f ° r aU
°* in A
*
Hence, by Proposition 30, there is a homomorphism
(JJ://-» SymW such that yty-9 and and also Pa4>"rca. Since 9 is a monomorphism
so is y, proving (i) of the Normal Form Theorem.
Now suppose that htH can be written as in (ii) of the Normal
Form Theorem. It is easy to check, inductively, that /HJJ acting on ( 1 ) gives the
element (gQ,otQ,£Q,... >£n-i>8n) of W. It follows that the expression for h is
unique, since it is determined by the action of hty on (1).
Take any h e H. It can certainly be written as

where n * 0 , e^-±l, and g-^G. Suppose that a /_i" a ;> e /"^"" e /-l» anc
*
be the image of g.. Then h can also be written as

with a similar expression if a / _ i " a ; » e /"~"~ e /-r an<^ ^ / € ^ - a

It follows (by induction on n, for instance) that any A c / / can be
written as in (i) of Britton's Lemma.
Now let h be written in this form. It is easy to see that, fot any
r<n, we have another expression for h in which gy is replaced by gya and
is replaced by bgr+]> the remainder of the expression being unchanged, where a

is in A if e - 1 and in A_ if e - -1, and b is the corresponding element of

A^ or A^ , respectively. In particular, we can replace gf by an element of the
relevant transversal, altering g +« but no other term of the product.
Applying this change successively to r - 0 , 1 , . . . gives an expression
for h in normal form.
Now suppose that we have two reduced forms for h. Applying the
above process to each of these gives two normal forms for A, which we have
shown must be the same. Since the process of obtaining the normal form from
a reduced form does not affect the />a, part (ii) of Theorem 32 follows at once.//

Conversely, uniqueness of normal forms, and also part (ii) of

Theorem 32, follows from part (iv) of Theorem 32 together with the fact that the
natural map from G to H is a monomorphism, For suppose h has both

< i • • • wCr1**and V
as reduced forms. Then

Because the two expression for h are in reduced form, if Theorem 32(iv) holds
then <x o -B o , e o - n o , and v ^ 1 ^ is in A^ if eQ - 1 and in A_^ if e o - - l .
If both expressions for h are normal forms this requires that
V Then
both
O " *0' ^i^ a £ l • • • a n d vi^R Tl1 • • • a r e norr*ial forms for Pa£°8Q A.
Inductively we see that the two normal forms must coincide.
When we simply have reduced forms we have already seen (because
of the condition on VQ" g^) that the second reduced form can be replaced by a
third reduced form, changing VQ to £ Q , changing Vj to another element, and not
changing any other term. We will then get two shorter reduced forms for
£
P~a °gn K and Theorem 32(ii) follows inductively.
o "
We can give yet another proof of Britton's Lemma, using methods
similar to method (IV) for Theorem 4 and Proposition 22, which will also give a
result on residual finiteness. We have already seen that the key part of Britton's
Lemma is part (iv), together with the fact that G embeds in //, the remainder of
the theorem following from this.
First we reduce to the case where there are only finitely many
stable letters. For in the general case if a product is 1 in H the corresponding
product in G* P is the product of conjugates of elements /? (a / n )p (a L )
(X OC U O t OC OC IOC

and their inverses. The set Q of those p« which occur in these conjugates is
finite, and our given product will then also be I in the HNN extension
<G, pap; Pap (ani o)pap{aRp
p rkup i,aip
), all 3 with pap in Q>. This argument is a version of
the direct limit argument of the examples in section 3.2.
Thus we may assume that there are n stable letters for some
finite ny and use induction on n. Let H be the HNN extension of O with stable
letters pr and associated subgroups Af and A_r for r<-ny and let K be the HNN
extension of G with stable letters pf for r<n, with the same associated
subgroups as before. Inductively, we can regard G as a subgroup of K, and so
we can form the HNN extension <AT, pn\ p~ ^nPnm A_n>. A comparison of
presentations shows that this HNN extension is just H. As K embeds in H (by
the initial case of the induction, which we prove later), we see that G embeds in
H.
Now look at a product of elements of G and P| > • • • ,P^ -
Regarding H as an HNN extension of K just amounts to bracketing together
those terms between two successive occurrences of Pn • By the case for a
single stable letter, such a product can only be 1 if it contains a portion
Pn kpn with k in An or a portion pnkp^ with k in A_n. If k does not involve
any pr with r<n we immediately have the situation we want. If k does involve
some pr then, inductively by Theorem 32(iii), the product k must have a portion
p~ gpr with g in Af or a portion PfiP' with g in A_f. Thus we have the
situation we require. We still have to prove the result when there is only one
stable letter.
So we now look at an HNN extension with a single stable letter,
say / / - <G,p\ p~lCp- D>. Let G/C denote the set of left cosets of C For the
moment we will suppose that \G/C\ - \G/D\. This certainly holds if G is finite,
since ICI - \D\, but need not be true if G\s infinite (for instance, if G is an
infinite dimensional vector space over 2^ a n d ^ a n d D a r e subspaces whose
complements have different finite dimensions).
As we still have to prove that j is a monomorphism, we shall
temporarily denote by g the image in H of g. We have seen that any HtH is
either g or has a reduced form £Q/> 8 ° . . . P£n~lgn> In the latter case we will obtain
a homomorphism <> | from H to Sym((7x A/), where N- { 0 , . . . ,/z}, such that /K|J*1.
Also ty will be such that g<l>*l for all giG-{\), showing that G embeds in H.
Let <p:G-*Sym{G* N) be given by (a,r).g<p - (ag,r) for all xtX, a and
g in (7, and r^n. If TC is any permutation of G* N satisfying the condition

(1) for any c*Cy g*G, and r*n, if {g,r)it - (g\s) then (gc,r)7i - (g'd,s),
where d is the image in D of c under the given isomorphism /:C-» A

then, by Proposition 30, there is a homomorphism ty:H->Sym(GxN) sending p to

n and with jty - 9.
Plainly any bijection from G/CxN to GIDxN can be lifted to a
permutation of Gx N satisfying (1). Further, if for each r and each coset U of C
we choose an element uf in U we can require (ur,r)n to be the element (v,$),
for any v such that the bijection sends (U,r) to (vD,s).
We shall define a relation between GICxN and G/DxN, and will
show that it defines a one-one function from a finite subset of G/CxN to
G/DxN. As we are assuming that IG/CI - 1(7/£M, this function can be extended
to a bijection from G/C x N to G/DxN.
We define the relation as follows. If 6^.-1 then ( £Q . . . grC,r) is
related to (g$.. .grDtr+l), while if £,"-1 then (g$.. .grC,r+l) is related to
( £Q • • • gjPJ)- Now ( # C » occurs as the first member of a pair of this relation
iff either E^-1 and gCm g^.. .gfC or e^j - -1 and gCm gQ.. .g^C. Since our
1 8 we
word is reduced, if e^" "" ^] cannot have g^C. Hence (gC,r) occurs as
the first member of at most one pair, so the relation is a function.
Similarly (gD9r) occurs as the second member of some pair iff
either e^- -1 and gDm g^... gfD or E^J - 1 and gDm g^... g^D- Again, at most
one of these can hold since we have a reduced word. Hence the function is
one-one.
As remarked, this function can be extended to a bijection of the
whole sets, which can then be lifted to a permutation 7c of Gx N. Further, since
the relevant members of GICxN and G/DxN correspond, it can be chosen so
that if e^-1 then ( s Q . . . gr,r)n - ( g Q . . . gr,r+l) while if e ^ - l then
( S o • • • gr,r+l)n ' ( £ 0 • • • gr,r).
By the construction of ^» hty sends (1,0) to (g^... g 9n). It follows
that /np^l, as required.
We still have to consider the case where \G/C\*\G/D\. We take
any set X such that \Xx G/C\ - \Xx GID\. (Properties of cardinal numbers using
the Axiom of Choice enable us to take X to be G. Readers who feel that the
Axiom of Choice should not be needed are reminded that in the proof of
Theorem 31 we used transversals, and the existence of a transversal in general
requires the Axiom of Choice. For a proof not involving the Axiom of Choice,
see Exercise 29). Fix some element x^iX.
Now let <p:G-> Sym(X x Gx N) be given by (x9atr)g://-» Sym(ATx Gx N) sending p to n and such that y<p-<l>. As
before, n can be obtained by taking a bijection of Xx GICxN to XxGIDx N and

lifting it, and this bijection can be obtained by extending a one-one function
from a finite subset of X*G/C*N into X*G/D*N This function is obtained
from the earlier function by modifying it to map a subset of {JCQ} X G/CX N to
{JCQ} XG/D* N. We now find that A^^l, because hty maps (JCQ,1,O) to
(JCQ,gQ- •• gn,n). The details are left to the reader.
If G is finite so is Sym(Gx N). So the discussion above, in addition
to proving Britton's Lemma, proves the following result.

Proposition 33 / / G is finite then the HNN extension

<G,p; p'^Cp- D> is residually finite.//

We can also obtain Britton's Lemma from the Reduced Form

Theorem for amalgamated free products. Let Y be free with basis {y } and let Z
-1
be free with basis iz^)- Let L be the subgroup ^O^y^ A^y^ of G*Y. Then L is
the free product of G and the groups y Av , as can easily be seen using the
Normal Form Theorem for free products. Hence L is isomorphic to the subgroup
M of G*Z generated by G and the subgroups z A_ z - So we may form the
amalgamated free product K~(G*Y)*^mM (G* Z). Let p a be y^z^ , and let H be
the subgroup of K generated by G and (p a ). An easy sequence of Tietze
transformations gives a presentation of K which shows that AT- H* Z and that H
is the HNN extension <Gp a ; pa~ ^ a P a " ^-a>* 1* *s n o w easv to
obtain Britton's
Lemma. By a similar construction we can obtain the Reduced Form Theorem for
amalgamated free products from Britton's Lemma and the Normal Form Theorem
for free products..
If hzH is written as in Theorems 31 or 32 we call n the length of
h. As before, an element h of length n > 0 has infinite order if no conjugate of h
has shorter length than h. In particular, h has finite order only if it is conjugate
to an element of G.
We have results similar to Propositions 28 and 29.

Proposition 34 Let H be an HNN extension

m
AP A^^> where <p^ is a monomorphism from Ay to G, and X is
in some index set A. Let M be a subset of A, and let B be a subgroup of G
such that ( ^ n ^ u ) 9 u " &"(* 9n) f°r al1
H€M. Then, the subgroup <B,p (ueM)>
is the HNN extension of B with stable letters p and associated subgroups
BcsA and BnA n O B.

Proof Let K be the HNN extension of B with stable letters q and

associated subgroups BnA and BnA of H. The conditions ensure that, in the corresponding
map from B*Q to C7*/\ where Q is free on {q }, an element with no pinch
maps to an element with no pinch. By Britton's Lemma, this tells us both that
the homomorphism from K to H has trivial kernel and that an element of K not
in B maps to an element of H not in C7, which proves the final part.//

Proposition 35 Let H be the HNN extension

<Gtp-^;pjT A^py - A y>. Let H be finitely generated. Then there are only finitely
many X. //, in addition, each A^ is finitely generated then G is finitely
generated.

Proof Since the free group on ipy) is an image of //, there can
only be finitely many X.
Write each of the finitely many generators of H in reduced form,
and let S be the set of elements of G which occur. Let B be the subgroup
<S, Ax, A_x (all X)>. Then H-<B, p x (all X)>, so Proposition 34 tells us that
Gm B, giving the result.//

Theorem 36 Any countable group G can be embedded in a

2-generator group.

Proof Let G be generated by [gn, all «elN}, where £ Q - 1 (with

repetitions if G is finite). It is sometimes most convenient to take ign, all n) to
be the whole of G, but at other times the use of a generating set is best.
Let F be free with basis ia9b). By Corollaries 3 and 1 to
Proposition 8, in G* F the subgroups <b'nabn> and <gna'nban> are both free with
the given elements as basis.
Hence H-<G*F,p; p~Xbnabnp-gna~nban> is an HNN extension of
G* F, and so G embeds in H. But gn*<a,b,p> and p apm b, so that Hm<a,p>.//
We note for use later that in the above construction F embeds in
//, so that both a and p have infinite order; also, any element of H with finite
order is conjugate to an element of G. Note also that any subgroup of a
2-generator group must be countable.

Proposition 37 There are exactly 2 ° non-isomorphic 2-generator

groups.

Proof Since each 2-generator group is a quotient of the countable

group F(a,b) which has only 2 ° subsets, there cannot be more than 2 °
2-generator groups.
It is enough to show that for any set S of primes there is a
2-generator group Hg such that the prime p is the order of an element of Hg iff
ptS, as then there will be 2 ° non-isomorphic groups Hg.
Let Gcm Z Z , the direct sum of cyclic groups. Let / / o be
obtained from Gg by the construction of the previous theorem. Since Hg is an
HNN extension of Gg*F, any element of finite order in Hg is a conjugate of
an element of Ggf and the result follows.//

Proposition 38 Any (countable) group G can be embedded in a

(countable) group H such that any two elements of H with the same order are
conjugate.

Proof Let U^a»^a)) be the set of all ordered pairs of elements of G

with the same order. Then <aa> and <b^> are isomorphic and so the group

is an HNN extensions of G, whence G embeds in <T'.

By construction, any two elements of G with the same order are
conjugate in O*, but two elements of <T with the same order need not be
conjugate (but they will be conjugate when their order is finite). We define
groups G for all n by <7Q- G and 6^ +1 - G^#. Let H-UGn . Then H is a group
which contains G. Any two elements of H will both be in Gn for some n, and if
they have the same order they will be conjugate in Gn + j , and so will be
conjugate in H.I I

In the above construction, any element of H with finite order is

conjugate to an
an element of G. This holds because any element of Gn with finite
order is conjugate to an element of

Theorem 39 Any countable group G can be embedded in a

countable simple group.

Proof The group G* <x\ where x has infinite order, can be

embedded in a 2-generator group K whose generators have infinite order.
Further, K can be embedded in a countable group S in which any two elements
of the same order are conjugate. Also any element of S with finite order is
conjugate to an element of G.
Let N< 5, with N * 1 . Take any w e N with w * 1 . If w has finite
order, it has a conjugate g in G, and g is in N Then JC~ g xg*N9 and this
element of <7* <x> has infinite order. Since N has an element of infinite order,
any element of infinite order in 5* is in N9 being conjugate to an element of N
In particular, N contains the generators of AT, and so N^K. Then any element
of S either has infinite order, and so is conjugate to an element of N9 or has
finite order and then it is conjugate to an element of G, which is contained in
N. As N is normal, we see that Nm S, so that S is simple.//

Our construction shows that for any set Q of primes there is a

countable simple group such that p is the order of an element of the group iff
piQ. We see that there must be 2 ° countable simple groups.

Proposition 40 There is a finitely generated infinite simple group.

Proof Let G be the group

<a,bpc,d; b~lab-a2, c~lbc-b2, d'lcd-c2, a~lda-d2>.

Let H~<a,b; bXab-a2>y and let K-<a,b,c; blab~a2, c~lbc-b2>.

Then H is an HNN extension of the infinite cyclic group <a> with
stable letter b; hence b has infinite order. It then follows that K is an HNN
extension of H with stable letter c. By Britton's Lemma, <a>n<&>-(l} in H. By
Britton's Lemma again (and Proposition 34) the subgroup <a,c> of K is free with
basis {atc). Now G can be regarded as the amalgamated free product of two
copies of K, so that G is infinite. Precisely, G is the amalgamated free product of
K and <x,c -1 2 - 1 2
K and <x,d,y; d xdm x , y dym d >, the amalgamated subgroups being <<2,c>
and <y,jc>.
G is very far from being simple. In fact it is possible to show that
G is SQ-universal; that is, any countable group is a subgroup of some quotient
of G
But there is (by Zorn's Lemma) a normal subgroup Af maximal with
respect to Nn<a>-{\). Then G/N is finitely generated, and it is infinite since

<<2> is infinite and intersects N trivially. Because of our choice of N, in any

proper quotient of GIN the image of a will have finite order. We will show that
such a quotient is trivial, so that GIN is simple.
Let 5 be a quotient of G in which the image a of a has finite
order m. Let 3, Y. and 5 be the images of b, c, and d. Then of mhixm - $?m 9 so
that 8 has finite order s which divides 2m-\. Similarly y and p have finite
orders r and n where r, n, and m divide 2 5 - l , 2^-1, and 2n-\, respectively. If
m, n, r> and s are all 1, then G is trivial. Otherwise let p be the smallest prime
factor of mnrs. We may assume that p divides m. Then p divides 2n - 1 . Also,
D -1 k.
by Fermat's Theorem, p divides 2y - 1. It follows easily that p divides 2 - 1 ,
where k is the greatest common divisor of n and p-\. In particular, k * 1, so
that /: has a prime factor smaller than p, contradicting the definition of p. This
contradiction proves the result.//
The group of this proposition is finitely generated but not finitely
presented. Examples of finitely presented infinite simple groups are known, but
their construction is difficult.

Exercise 21 Fill in the details in the proof of the Normal Form

Theorem for HNN extensions.

Exercise 22 Show in detail that an element of finite order in an

HNN extension must be conjugate to an element of the base group.

Exercise 23 In the construction following Proposition 33, give a

sequence of Tietze transformations which shows that the amalgamted free product
constructed is the free product of an HNN extension and a free group. Hence
obtain Britton's Lemma from the Reduced Form Theorem for amalgamated free
products.
Exercise 24 Let A and B be groups, and let C and D be isomorphic
subgroups of A and B respectively. Let K be the subgroup <p~ Ap,B> of the
HNN extension <A* B9p\ p Cpm D>. Show, without using any reduced form
theorems, that K is isomorphic to the amalgamated free product A* gmpB. (Use
Proposition 30 to show K satisfies the definition of the amalgamated free
product.) Hence obtain the Reduced Form Theorem for amalgamated free products
from Britton's Lemma and the Normal Form Theorem for free products.
Exercise 25 Let A and B be groups, and let C and D be isomorphic
subgroups of A and B respectively. Let G be any group containing (copies of) A

and B and such that AnBm{\). Show, using Proposition 27, that the subgroup
 of the HNN extension <G,p;p Cpm D> is isomorphic to the
amalgamated free product A* gmpB. (Note that Proposition 27 uses the Reduced
Form Theorem, so that the previous exercise cannot be deduced from this one.)

Exercise 26 Let C and D be isomorphic finite subgroups of the

residually finite group G. Show that the HNN extension <G,p; p Cpm D> is
residually finite.
Exercise 27 Let C and D be isomorphic finite subgroups of the
residually finite groups A and B. Show that the amalgamated free product
A* gmpB is residually finite. (Use the previous two exercises with Gm A* B. We
could also take G to be A * B9 but that would require the corollary to Proposition
22, whereas the suggested approach allows us to deduce that result by taking C
to be {1}.)
Exercise 28 Show that an HNN extension of a residually finite
group need not be residually finite when the associated subgroups are infinite.
(One possible approach is to argue that if all such HNN extensions were
residually finite then the group of Proposition 40 would be residually finite.)

Exercise 29 In this exercise we shall obtain a proof of Britton's

Lemma not using the Axiom of Choice.
Let C and D be isomorphic subgcups of the group G. Let M be
the set of all finite sequences (#Q,p e °,gj,... ,gn_\,P*n~l,gn). Show that M is a
monoid whose identity is (1) under the multiplication
( % • • • *P n~l>gn)(ho>P °»• • • >hm)" (So, • • • ,P n~l>8nhQ>P °»• • • >hm)>
Define a relation ~ on M by
(g(\*P °»• • • tP rt"*»£/) ** (h(\*P °»• •• tP /Il»^#») i^ there are c u , . . . *&„* > br\» • • • >bn-\
in G such that AQ - g§a§\ h^ - b^g^a^ \ ..., h m t>n+gn> where a- € C if e •- 1 and
a-^D if £;"-!> and bj is the element in D or C respectively which corresponds
to a.. Show that ~ is an equivalence relation. Let u - (gQ,pB°,... tg ) . Show that
if u^v where v is ( . . . ,// l ,l,p~ T l ,...) iff for some / either £. - 1 - - E ^ J and g- € C or
E / --l--E / ._ 1 and g;t D.
Let A/j be the set of equivalence classes. Define a relation -» on
A/j by a •* 3 if there is u - (gQ,pB°f... ,gn) in a with E^- - E ^ J and g-m 1 and 0
contains the element (g^t... fp£i~29g^g-+^p£i+2s • • • ,gn)- Let <x = p hold iff there
are a i 9 . . . , a . with a, - a, a, - 0, and, for all z, either a- « - » a . or a- - * a . ,. Show
1 iC 1 /C Z~l Z Z Z""l

that 2 is an equivalence relation. Show that the set of equivalence classes is a

group, and, using the comment following Proposition 30, that this group is the
HNN extension <G,p; p'^Cp- D>.

Show that if a -» 3j and a •* ^ then there is y such that &j -» y

and fL "* Y- Deduce, as in Exercise 4, that if a is equivalent to the identity then
there is some p such that a -» p. Hence obtain Britton's Lemma.

2 SPACES AND THEIR PATHS

2.1 SOME POINT-SET TOPOLOGY

We begin with a lemma that will be frequently used, often without
explicit mention.
Lemma 1 (Glueing Lemma) (/) Let X and Y be sets, let X^ be
subsets of X such that X-UX^, and let f^-X -^Y be functions such that
fJX^nXa" fJXnXa for all a and g. Then there is a unique function f.X+Y
(X Ot p p CX p
such that / I ^ a " / a for all a.
(ii) Let the conditions of (i) hold, and let X and Y be topological
spaces. Suppose that each f is continuous (when X is given the subspace
topology). Suppose either that there are only finitely many sets X^ each of
which is a closed subspace of X or that each X^ is an open subspace of X.
Then f is continuous.

RemarkWhen f is a function from a set A' to a set Y and A is a

subset of X the notation f\A means the restriction of / to A; that is, the function
from A to Y whose value on a € A is fa.

Proof (i) Let ScXxY be i(x,y); there is a such that xtX^ and
j>-x/ a ). Since X-UX^, for every x there is at least one y with (x,y)tS.
Suppose that (x,y)cS and (x,z)*S. Then there are a and 3 with xtX^, y~ * / a ,
and JC c XQ , zm xf» . Since / a - /« on Xa n Xn , by hypothesis, it follows that
y z.
This means that we can define a function f.X -» Y by y~ xf iff
(x,y)zS. By construction, flX^-f^, as required. Since X-UX^ there can be only
one such function.
(ii) Suppose that the first condition holds. Let C be a closed subset
of Y. Then C/ - UCZ . Since each / is continuous, Cy " is a closed subset
of Jf . Since X is closed in A", this implies that C/ is a closed subset of X.

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:28:17, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.003
Some point-set topology 50

As there are only finitely many sets X^, it follows that Cf is a closed subset
of X. Hence / is continuous.
The case where each X^ is open is similar.//
We shall often be faced with the situation of a set X which is the
union of subsets X^ each of which has a topology £ a , and we will want to
define a topology on X. Let £ be M; A n Xa e £ a for all a}. Then £ is a
topology on X, called the weak topology induced by ^ a )»
In particular, if ^ 0 ^ 0 - 0 for ocs*p we call X the disjoint union of
the spaces X^, written UATa. Also, if there is a point p such that X^ n Xn - ip)
for all a ?*3 then we call X the join of the spaces X (at p), which we write as
V
*cf
It is immediate from the definition that a set C is closed for the
topology £ iff Cc\X^ is closed for £^ for all a. In particular, if each £^ is a
Tj topology (that is, each point is a closed set) then so is £.
Let Y be any space, and let f.X ^ Y be a function. Then / is
continuous (using topology X on X) if fiX is continuous for £^ for all a (the
converse is obvious). We need only take an open subset V of Y and observe that
VflnXa-Vlf\XQL)'1 is open for £ a for all a.
Let £ be the weak topology induced by the topologies £^. Then
ATa has two topologies, namely the given topology £Q[ and the topology as a
subspace of the topological space X. We would like to ensure that these two
topologies are the same. This plainly requires that £ and £* induce the same
topology on X^ n Xn for all a and p (since the corresponding property holds for
the subspace topologies). In general this necessary condition is not sufficient.
Suppose that this condition holds and that X n Xn is a closed
subset of X (with the topology £ a ) for all a and p (or that X^nXn is open in
X^ for all a and P). Then the two topologies are the same.
For let CcX^ be closed for £a. Then CnX* is closed for the
topology on X n XQ induced by £ a . By hypothesis this is also the topology
induced by £Q . Since X^ n Xn is, by hypothesis, a closed subset of Xn with the
topology £a, it follows that CnXa is a closed subset of Xa with the topology
P p P
£n. Then C is closed for £, since this holds for ail (3. The other case is similar.
Proposition 2 Let X have the weak topology £ induced by i£ah
where each £^ is a Tj topology. Suppose that X has a subset A such that
X^n XQ-A for a^p. Then any compact subset C of X is contained in the union
of finitely many X^.

Proof For each a such that C meets X^- A take some

'p eCn(X - A). Let P be the set of all these / ? a , and let Q be any subset of P.
Since p*p
Q4X
a for B^a, we see that, for each a, QnX a consists of at most one
point, and so is closed for X^. Hence Q is closed for X.
As this holds for every subset Q of P, we see that P is discrete.
Also P is compact, being a closed subset of the compact set C. It follows that P
is finite, being compact and discrete. This gives the result.//

Let p be a function from a set X onto a set Y. Let £ be a topology

on X. Then we can define a topology XY on Y by requiring a subset B of Y to
_i *
be open iff Bp is open. This is called the identification or quotient topology
on Y {induced by p). We are frequently given an equivalence relation = on X,
with Y being the set of equivalence classes and p being the natural map. In this
case we write X/= for Y with the identification topology. We may simply refer
to this space as being obtained from X by identifying the points of each
equivalence class. More generally, given any relation R on X, there is a smallest
equivalence relation containing R, and we refer to the corresponding
identification space as being obtained by identifying pairs of elements in R.

Proposition 3 Let = be an equivalence relation on X, let Z be any

set and let f.X + Z be any function. Let Y denote X/= and let p:X-*Y be the
projection.
(i) There is a function g:Y+ Z such that f~pg iff x*f- x~f
whenever x+ = x~ •
(//) Let X and Z be topological spaces, and let Y have the
identification topology. Let g be as in (/). Then g is continuous iff f is
continuous.

Proof (i) is obvious.

From the definition of the identification topology, p is continuous.
Hence / is continuous if g is.
Conversely, let / be continuous. Let W be an open subset of Z.
Then Wfl is open in X. Since tyfl-(Wg"l)p'1, the definition of the
identification topology shows that Wg" is open in Y, as required.//

Let AT be a compact space and Y a Hausdorff space and let / be

a continuous map from X onto Y. Then Y has the identification topology
induced by /. For we may define an equivalence relation on X by JC, = x~ iff

jCj/- JC2/. By the proposition, / induces a continuous bijection from X/= to Y.

Since X/= is compact and Y is Hausdorff, this must be a homeomorphism.

Example Let X be any space, and let / be the unit interval [0,11
On X*I, define (JCpfj)^.^.^) ^ either ^ " O " ^ (Wl^ x
\ anc
* X2 af
bitrary) or
both JCj - JC2 and /j - t^ more simply expressed, CAT is obtained from X x / by
identifying (JCpO) and (.^O) for a ^ *i an
d *2* ^ e
identification space is called
the co/ie cvi AT, written CX, and the equivalence class consisting of all (JC.O) is
called the vertex of CX. It is easy to check that the subspace ATx{l} of CX is
homeomorphic to X, so we usually regard X as a subspace of CX in this way.
Let if be {(*,,...,jcJeR*; I x . 2 * l ) , and let Sn~l be
« 2
{(jCj,... ,JC^) € R ; Txr-'l); these are called the n-disk and (/*- \y sphere.. Then
CS''1" is homeomorphic to zA For the continuous map from Sn~ x / to if1
sending (JCJ, . . . ,jcn,f) to (tx^9... ,txn) obviously defines a bijection from CSn
to z/ , and this is a homeomorphism since Sn~ * I is compact.
1

Similarly, if X is any compact subset of R^ then CX is

homeomorphic to the subset of R7* which consists of all (tx*t... ttx ,l-t) with
( j t j , . . . ,JC^) € X. This explains the terminology, as this set is what would be
called the cone in geometric situations.
The functor C is not as well behaved as one might expect. For
instance, if X is the closed interval [0,1 J and A is the open interval (0,1) then
the topology of CA as the cone on A is not the same as the topology of this set
regarded as a subspace of CX.
Real projective n-space, Rf^, is the space obtained from Sn by
identifying ( x , , . . . , * .«) with {-x*,... ,-x ,,). We may map the disk if1 into
n I 2 2
S by sending (JC^ , . . . 9xn) to (jCj,...,jc / | ,y(l-JCj - ...-xn) ). This provides a
map from D* onto R/^. It follows that R / ^ can also be obtained from E/1 by
2
identifying a point (x^ , . . . ,JC^) such that IJC^. - 1 with the point ( - X j , . . . ,-JC^).
It is easy to see from this that Rf^ is homeomorphic to S\ although, of course,
the identification from S to R/^ is not even one-one.

Let X and Y be spaces, and let f:A -> Y be a continuous map from
a subspace A of AT. Then we may form the disjoint union XdY and then take
the space obtained by identifying a and af for all a* A. The resulting space is
written Yu^X, and it is called the space obtained by adjoining (or attaching) X
to Y by f. In particular, X may be the disjoint union of spaces X-^, in which
case / consists of a family of continuous maps fy}A^ •* Y, and we refer to the
space obtained by adjoining the spaces Xy^ to Y by the maps / ^ . Note that the

join of the spaces X-^ may be regarded as the space obtained from one of them,
X say, by adjoining all the other spaces X^ by the maps sending the point p^
of X-^ to the point p of X^.
In particular, RP' 2 is obtained from RP'2 by adjoining i/1 by the
identification map from Sn~^ to RP' 1 " 1 .

Proposition 4 Let f:X-*Y be an identification. Let Z be a locally

compact space. Then the function F:X * Z-+Y* Z given by (x,z)Fm (xf,z) is also
an identification.

Proof Let W be a subset of Y*Z such that WF~l is open We need

to show that W is open. So take CVQ ,ZQ) * W. It is enough to find open subsets U
and K of r and Z with iyo,zo)^UxV£W.
Choose some XQ with *(/')>§- Since Z is locally compact, there is
an open neighbourhood V of ZQ such that V is compact and
V^ £.(z;(xQ9z)tWF }, as this latter set is an open subset of Z. Let
h
Then AfxVcW, and so 4 / 7 l x]T c WF"1. Hence Affl-A. Also, as
is openn and V~ is compact, it is well-known (and easy to check) that A is
open. Hence Af is the required set U.I I

Proposition 5 Let X be a compact metric space with metric d, and

let U be an open covering of X. Then ] 8 > 0 such that any set of diameter
less than 5 is contained in some member of U.
Remark We call such a 8 a Lebesgue number of U.

Proof Since X is compact, we may assume that U is finite, say

Define f.:X^R by xf^ inf{d(x,y); y^U^. Then f. is continuous,

and xfj^O for JCC£A, since U- is open. Hence the function max^- f. is continuous
and is always positive, since U is a covering of X. As X is compact, this
function has a positive minimum 5.
Let S be a non-empty set of diameter less than 8. Take some x c 5*.
There is some / with xf. * 8. Then y^U. if d(xy) < 8, so that ScU^ as
required.//

Exercise 1 Let X have the weak topology by the family (* a »£ a )

where £ a and X* induce the same topology on X^n^. If there are only two

sets Ar(X show that Za is the subspace topology on X^. Show that this need not
hold if there are three or more subsets. (There is an example with X having
three eler
elements; a slight variation of this will give an example where each X is
infinite.)
Exercise 2 Let X and (* a »£ a ) b e a s in exercise 1. Suppose that
there is a topology X on X such that $L is the topology induced on X by £
for all a. Show that £ a is also the topology induced by the weak topology.
Exercise 3 Show that the cone on (0,1) is not a subspace of the
cone on [0,11
Exercise 4 Let AT be a space, and let f.X^Y and g:Y+Z be onto
maps. Let Y be given the identification topology induced by /, and let Z be
then given the identification topology induced by g. Show that the topology on
Z is the identification topology induced by fg.
Exercise 5 Show that the space Y is a subspace of the space Yu,X.
Show that X- A is a subspace of Yu^X if A is a closed subset of X.

2.2 PATHS AND HOMOTOPIES

We denote the non-negative real numbers by IR .

Definition A path in a space X is a continuous map /:CO,r] •» X for

some /•€ IR .If /•« 1 we call the path a standard path, while if r - 0 we say the
path is trivial. If 0 / - a and rfm b, we say the path is a path from a to b (or a
path starting at a and ending at b, or other similar phrases). If am b we say that
the path is a loop based at a. Finally, if f£0,r] -» A' is a path then the
standardisation of / is the path f^.I-* X defined by tfo~{tr)f.

If the space is one of which we can make an intuitive picture then

we can picture a path as a piece of string or elastic. It is then quite reasonable
to talk of deforming a path, by stretching, pulling, pushing, and so on, but
keeping the end-points fixed. The reader may find it easy to believe that any
2
two paths in the plane IR between the same end-points can be deformed into
each other in such a way. On the other hand, if we take the upper and lower
2
halves of the unit circle in IR -(0,0) then these two paths cannot be deformed
into each other (we could think of a spike coming out of the plane at (0,0), and
any attempt to deform one semicircle into the other would get caught up on the
spike). This chapter and Chapter 4 are devoted to making these ideas precise.
The intuitive idea of deformation could be formalised by referring
to a continuous family of paths / . But we would need to define a topology on

the set of paths to make this precise. This approach is useful in some advanced
work, but an easier approach is possible.
Observe that a family of functions / can be regarded as a function
F of two variables by defining (t,u)F to be tfu; conversely, any F gives rise to
a family of functions / . We shall replace the idea of a continuous family of
functions by continuity of the corresponding F using the definition below. For
formal proofs we need to use F and the definition below, but to get an intuitive
idea of what is happening it is often useful to think of the family of paths f u.

Figure 1, showing some of the paths fu

Definition Let /:CO,/\I -» X and g:CO,s] •» X be paths from a to b.

Suppose that there is a continuous function <x:/-*IR with Ooc-r and l a - 5 and
2
a c o n t i n u o u s f u n c t i o n F : A - > X, w h e r e A c R i s i ( t . u ) ; O^t^uoL, O^u^l) such
that (t,O)F- tf for t*r, ( U ) F « tg for t^s, (O,a)F- a and (UOL,U)F- b for all u.
Then we say that / is variable-length homotopic to g9 and the map F is called a
variable-length homotopy between / and g. If / and g are standard paths and
UOL - 1 for all u we refer to a standard homotopy.
Lemma 6 Let f be variable-length homotopic to g. Then there is a
variable-length homotopy F whose corresponding function a is linear; that is,
uoim us + (1- u)r for all u.

Proof Let G be a variable-length homotopy defined on the set B

with corresponding function 3. If rmsm0, we define a by n a - 0 for all u and we
define F by (Q,u)Fm a for all u, giving the required homotopy.
If r*0*s9 let A be {(t.u); O^r^wa, Ozu^l), where a is given by

HOC - us + (1 - u)r. Then the function j:A-* B given by (t,u)jm ((u£>)t/uoL, u) is

obviously continuous. The required homotopy F is just jG.
Suppose that r~Q*s (the remaining case is similar). Define a and A
as in the previous paragraph. The previous definition of j makes sense except
when u - 0 (that is, except at (0,0) ) and is continuous except at (0,0). If we
define (0,0)y to be (0,0) it is easy to check that j is also continuous at (0,0)
because p is continuous. As before, we can now define F to be jGJ7

The following corollary is immediate.

Corollary / / two standard paths are variable-length homotopic then
they are standard homotopic.//

H
Because of this corollary, we shall omit the words variable-length"
and "standard" and will simply refer to "homotopies". We shall also write f*g
to mean that / is homotopic to g.

Lemma 7 Any path is homotopic to its standardisation.

Proof Let f:lO,r] -> X be a path. Define oc:/-» R+ by UOL - u + (1 - u)r,

and let A be i(t,u); O^r^t/a, O^u^l). If r - 0 let F:A-> X be given by (tfu)F-0f
for all t and u. If r*0 let F be given by (t,u)F- (tr/uat)f. Then F is the
required homotopy.//

Proposition 8 Homotopy is an equivalence relation.

Proof f^f by the homotopy F given by (t,u)F- tf, the

corresponding a being constant.
Suppose that f<*g by the homotopy F and corresponding a. Then
g*f by the homotopy G and corresponding p given by (r,w)G-(r,l- u)F and
u0 - (1 - K)ot.
Suppose that f*g by the homotopy F and corresponding a, and that
g*h by the homotopy G and corresponding &. Then / ^ A by the homotopy H
and corresponding Y given by

(t,)H-(t,2u)F and «Y-(2t/)a for

(r,u)//-(r,2«-l)C/and KY-(2H-1)P for u* 1/2.

The Glueing Lemma assures us that H and y are continuous, and the other
properties needed are obvious.//

Figure 2, showing some of the maps fu and gu

This result is most easy to see pictorially. If we have a continuous

family of paths fu joining / and g9 and another continuous family gu joining g
and A, then we get a continuous family joining / and h by following the family
fu by the family gu> In order to do this in total time 1 and not in total time 2
we need to change u to 2u. Formalising this gives the definition of H above.
Lemma 7 and the corollary to Lemma 6 tell us that every
variable-length homotopy class of paths contains exactly one standard homotopy
class of standard paths. It follows that, when considering homotopy classes, we
can restrict ourselves to standard homotopy of standard paths whenever this is
convenient.
Let f:LO,rl + X and g:[O,s] -» X be paths such that Og-rf. Then we
may define a path f.g from CO,/* +s] to A' by

t{f-g) - tf for
t{f.g) - (/- r)g for t*r.

We call f.g the product of / and g. We write f.g rather than fg to avoid
confusion with the composition of functions.
It is clear that the product of paths is associative; that is, if one of
{f.g).h and f{g.h) is defined then so is the other and they are equal; also these

products are defined if both f.g and g.h are defined. Also, if / is a path from a
to b and / and i^ are the trivial paths at a and b then we have '^ • / " / " / • ' £ •
If we restricted attention to standard paths (when we would also
have to replace f.g by its standardisation) these results would not hold, although
they would hold up to homotopy. The main reason I made the definition of
paths used here (rather than looking only at standard paths, which is more usual)
is to have an associative product of paths. The details of the situation for
standard paths are left to the exercises.
Let flO,r\ •> X be a path from a to b. We define a path / from b
to a, called the inverse of / , by tfm (r- t)f for Qztzr. If g is another path and
m
f.g is defined then {f.g) g.f, obviously. Plainly, J - / .

Lemma 9 Let f be a path from a to b. Then f.f • / and f.f. / ^ .

Proof Define F by

(t,u)F- tf for O
(t9u)F- (2ru - t)f for ru * t< 2ru.

Then F is a homotopy from iQ to / . / . The second part follows from the first,
replacing / by / .//

If we think in terms of a family of paths gu then gu follows / for

a proportion u of the total time taken by / and then follows the same path in
reverse.

Lemma 10 Let f^g and f <*g\ and let f (and g') start where f
(and g) end. Then ff'^g.g' and f ^W .

Proof Let F and F' be the homotopies, with corresponding

functions a and a'. Then the homotopy between / . / ' and g.g* is F.F' given by

(tyu)F.F'-(tyu)F for r*noc,

(t,u) F.F*- (t- UOL,U)F' for UOL* t^ua* u<x\

Also the homotopy between / and g~ is F given by (t,u)F ~ (UQL- ttu)FJ/

In terms of a family of paths, given families fu and f'u we obtain
families fu-f'u and fu .

Figure 3, showing some of the maps fu and f'u

Lemma 11 Let f and g be paths from a to b. Then f^g iff

f-g~in'

Proof If f^g then f.g <*g.g. As g.g ^iQ9 we find that f.g °*iQ-
Conversely, if f.g °*ia we have fmfi^f.g.g^ia.gmg.ll

It is now easy to check that the set of homotopy classes of loops

based at the point a is a group. This group will be our main link between
topology and group theory. The set of all homotopy classes of all paths is an
unfamiliar algebraic object which is closely related to a group, but differs from a
group in that the product of two elements is not always defined. We shall study
these algebraic objects in Chapter 3, and return to their connections with
topology in Chapter 4.

Let / be a standard path from a to b> and let g be a standard path

from b to c. Let fog denote the standardisation of f.g.
Exercise 6 Write down the formula for fog.
Exercise 7 Let A be a standard path from c to d. Write down the
formulas for (fog)oh and for fo{goh). Show that these paths are different but are
homotopic. Obtain a formula for a homotopy between these paths.
Exercise 8 Let j Q be the standard path given by tj m a for O ^ r ^ l .
Show that jof^f, and obtain a formula for a homotopy between these two
paths.

2.3 PATH-COMPONENTS
We say that the space X is path-connected (or pathwise connected)
if any two points can be joined by a path. Because of the properties of products
and inverses of paths, X is path-connected if there is some a c X such that a can
be joined to any point of A" by a path. Plainly, if <p is a continuous map from
X onto Y then Y is path-connected if X is path-connected.
A subset A of Rn such that whenever x and y are in ^4 then so is
the whole line-segment joining x and y is called convex. Evidently, any convex
set is path-connected. It is easy to construct many examples of convex sets. For
instance, disks are convex, as are the products of intervals.
More generally, for any space X we may say that a is equivalent to
b if a and b can be joined by a path; the properties of products and inverses
show that this is an equivalence relation. The equivalence classes are called the
path-components of X; X is path-connected iff it has only one path-component.
Some authors use the phrase "arcwise connected*' rather than
"pathwise connected". Strictly speaking, this should mean that any two points
can be joined by an arc; that is, by a homeomorphic image of an interval, rather
than just by a continuous image. It is a deep theorem of point-set topology that
these two notions are in fact the same; fortunately, we shall not need this.
Since the continuous image of an interval is connected, any path-
connected space is connected. The converse is false, as is shown by the example
below.
Example Let A be {(xfy)eR2; y- sin(l/jc), JC >0}. Then A is
connected, being the continuous image of an interval, and hence so is its closure
B. It is easy to check that B - A u {(Oj;); \y\ £ 1). Also A is path-connected.
Now let / be a path in B with 0 / - (0,0), and let tf be (r/, ,r/ 0 ). By
-1
continuity, 0/j is closed, and it is plainly non-empty. We shall show that it is
also open. This will show that tf^ - 0 for all t, and, in particular, there is no
path in B from (0,0) to (l/7t,0).
So take c with c/j - 0, and suppose that c/^ * 1. Then there is 8 > 0
such that t/2*l for lr-cl<8. It follows that, for lf-cl<8, r/j * \/(2n+\/2)n for
any n. Now take t with \t- cl<8, and assume, if possible, that tf^ *0. Then
there will be some n with l/(2« + 1/2)TU < r/j, and then, by the Intermediate Value
Theorem, some u between c and t such that uf^ - l/(2/i + 1/2)TT. Since this cannot
happen, we must have tf^-O for all t with \t-c\<h. If c/ 2 - 1 we take 8>0 such
that c/ 2 # -1 for \t-c\<h and argue similarly.

Figure 4, showing part of B

3 GROUPOIDS

3.1 GROUPOIDS
At the end of the previous chapter we had a brief indication that it
might be useful to consider objects similar to groups, but where the product of
two elements is not always defined. In this section we consider these objects in
detail.
A partial multiplication on a set G is a function from some subset
X of Gx G to G. If (x,y)*X we denote the value of the function on (xfy) by xy
or by x.y. We say that xy is defined to mean that {xyy)*X.
The element e is an identity for a partial multiplication if exmx
whenever ex is defined and also yemy whenever ye is defined. There may be
many identities, but it is clear from the definition that if e and / are identities
with ef defined then emf.

Definition A groupoid is a set G with a partial multiplication such

that:
(1) (associative law) if one of (ab)c and a(bc) is defined then so is
the other and they are equal; also, if both ab and be are defined then (ab)c is
defined,
(2) (existence of identities) for any a, there are identities e and /
with ea and af defined,
(3) (existence of inverses) for any a, and e and / as in (2), there is
an element a such that aa - e and a am f.
We nearly always want our groupoids to be non-empty. I leave it
to the reader to decide which properties stated should have the empty groupoid
given as an exception. It is nonetheless convenient to permit the empty
groupoid; for instance, the intersection of subgroupoids of a groupoid (unlike the
intersection of subgroups of a group) can be empty.

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:29:50, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.004
Groupoids 63

Examples Plainly any group is a groupoid.

Let A be a set, and, for each X € A let X-^ be a set. Let G^ be the
set of all bijections from X-^ to X (this may be empty), and let G be UG^ .
Then G is a groupoid under composition. This generalises the group of
permutations of a set, which corresponds to the case when A has only one
member.
Let C be a small category (that is, a category in which the objects
form a set, in which case the morphisms also form a set). Let G be the set of
all isomorphisms of C. Then G is a groupoid. The previous example is a special
case of this one.
Let H be a subgroup of the group G, and let {H^} be the set of
right cosets of H. For g*G let (g,<x) denote the bijection from H^ to H^g
obtained by multiplication by g. Then i(g,0L); all g and a) is a groupoid.

The simplest properties of groupoids are very similar to those of

groups, with modifications needed because multiplication is not always defined.
We note first that for each a there can only be one identity e such that ea is
defined. For suppose that e^ is also an identity and eâ is also defined. Then
eâ~ a, since e^ is an identity. Hence e(eâ) is defined. By the associative law,
ee. is also defined, and we then know that e~ev
-1
The equation axm b has a solution only if a b is defined, and then
its only solution is a b. First suppose that a b is defined. We know that ad
is defined and equals the identity e. Hence a(d b) is defined and equals
(ad )bmebmb, since e is an identity. Conversely, suppose that there is some JC
m
such that ax b. Since da is defined and equals the identity / , we know that
a (ax) is defined and equals (a a)xm fxm x; that is, a b is defined and equals
JC. Similarly, the equation yam b has a solution only if ba is defined, and then
the unique solution is ba' .
As a particular case of this, the inverse of an element a is unique;
in fact there is only one element x such that ax - e, where e is the unique
identity with ea defined.
2
Also, any element e such that e - e is an identity. For there is an
identity / such that efm e, and then ee - ef shows us that e - / .
Notice that ab will be defined iff there is an identity / such that
both af and fb are defined. For if this happens then (af)b will be defined, and
this is just ab. Conversely, if ab is defined let / be the identity such that, af is
defined. Since afm a, we see that (af)b is defined, and hence fb is defined. In
particular, if G has only one identity then ab is always defined, and so G is a
group.

We shall frequently refer to the identities of a groupoid G as the

vertices of G. If atG has identities e and / with ea and af defined, we then say
that a is an arrow or edge from e to /, and we say that e is the start
(or initial vertex) of a, while / is the end (or ///ia/ vertex or terminal vertex)
of a. We denote the set of vertices of G by K(6).

It is often more convenient to require the set of vertices of G to

be a set bijective with the set of identities of G. When we are given a groupoid
G we require for definiteness that V(G) is the set of identities of G. However,
when we are constructing a groupoid there is often a natural choice for the
vertices. For instance, in the example where H is a subgroup of the group G, it
is natural to take the cosets of H as vertices, while in the example of bijections
between sets we would naturally take the sets X-^ (or the elements of A) to be
the vertices.
We have seen that a small category in which all the morphisms
are isomorphisms is a groupoid. Conversely, any groupoid is a small category in
which all the morphisms are isomorphisms; the objects of this category are the
vertices of the groupoid, and its morphisms are the arrows.

Examples Let V be any set. Let D(V) be the groupoid whose vertex
set is V and such that, for all v*V and all weV with w#v, there is exactly
one arrow from v to v and there are no arrows from v to w. Let T(V) be the
groupoid with vertex set V and with exactly one arrow from v to w for any v
and w in V. In both cases the definition of the multiplication is obvious. We
call D(V) the discrete groupoid on V, and we call T(V) the tree groupoid on V.

A subgroupoid of a groupoid G is a subset H of G such that if a

and b are in H then ab is in H (if ab is defined) and a' is in H. It then
follows that if a*H and e and / are the identities such that ea and af are
defined then e and / are in H. This holds because e- ad and / - da.

Examples Let G be a group. Then a non-empty subgroupoid of G is

the same as a subgroup of C7.
The intersection of any collection of subgroupoids of a groupoid is
a (possibly empty) subgroupoid.
For any set V, the discrete groupoid on V is a subgroupoid of the
tree groupoid on V.

Let G be a groupoid, and let 5 be a subset of V(G). Define G~ to

consist of all arrows which begin and end in S. Then Gg is a subgroupoid of G
which we call the full subgroupoid of G with vertex set S. If Sm ie), we
write G& rather than O,ey Since G0 has e as its only identity, Ge is a group,
which we call the vertex group of G at e.

Let e and / be vertices of a groupoid G. Writing e=f iff there is

an arrow from e to /, it is easy to check that • is an equivalence relation on
V(G). The equivalence classes are called components of C7, and G is called
connected if there is only one component. For instance, T(V) is connected and
D(V) is totally disconnected (that is, each component consists of only one
vertex).
Let a be an arrow from e to /. Then there is a function a* from
Gg to (jy defined by ga* - a go. It is easily checked that a* is a homomorphism,
and that if b is an arrow starting at / then (ab)* - a#b*; also e* is the identity.
It follows that a* is an isomorphism, since it has (a )# as its inverse. Notice
that if c is an arrow in GQ then c# is just conjugation by c. It follows that if a
and b are both arrows from e to / then the isomorphisms <*# and b* differ by a
conjugation. Because a* is an isomorphism, we obtain the following proposition.

Proposition I If e and f are vertices in the same component then

G re isomorphic;
and Gy are isomorphic; in particular, the vertex groups of a connected
groupoid are isomorphic.lI

Let 9 be a function from the groupoid G to the groupoid //. We

call 9 a homomorphism if whenever ab is defined then (a<p)(b<p) is defined and
(a<p)(b<p)-(ab)<p.
Let 9 be a homomorphism, and let e be an identity of G. Then
ee - e, and so (e<p)(e<p) - e<p. Hence e (because e<p is an
identity such that (e<p){a<p) is defined), and then a <p - (a<p) . If we regard
groupoids as categories then a homomorphism is the same as a covariant functor.
If V and W are sets then any function from V to W induces
obvious homomorphisms from D{V) to D(W) and from T(V) to T(W). These are
functors from the category of sets to the category of groupoids.
Let G be a groupoid and let If be a set. Then any map <p:V(G)-> W
extends uniquely to a homomorphism $:C7-» T{W). For if a is an arrow from e

to / we must define a$ to be the unique arrow in T(W) from e Groupoids
are adjoint.)
Also, any mapty:W~*V(G) extends uniquely to a homomorphism
ViDiW)-* G, but this property is much less important than the previous one.
Let G and H be groupoids, and let S be any subset of V(G). Then
any homomorphism <p:Gg-* H extends (non-uniquely) to a homomorphism
ty:G^> H. First we choose for each vcK(6) whose component meets S an arrow jcy
which ends at v and starts in S\ if vtS we take jcy to be an identity. We also
choose some identity i in H. Let g be an arrow of G from v to w. If the
component of v (which is also the component of w) meets S then both JC and
xw are defined. Hence * v £ * w is defined and is in Gg . In this case we define
gty to be (xvgxw )<p. If the component of v does not meet S we define gty to be
/. It is clear that IJJ extends 9, and it is easy to check that 4> is a
homomorphism.
It is possible to define normal subgroupoids of a groupoid, and to
prove isomorphism theorems. We can then develop a theory of generators and
relators. These topics are important in the general theory of groupoids, but we
will not need them for the applications that concern us. Consequently we shall
not develop them further. For the same reason, though we shall need to know
something of the behaviour of pushouts of groupoids, we shall not need to
show that pushouts always exist.
We may define a partial multiplication on the set G x H by the rule
(£» h){g^, h^) - (ggj, /zAj). This plainly makes G x H into a groupoid, which we call
the direct product of G and H. The element (eff) is an identity iff e is an
identity of G and / is an identity of H. When this happens, the vertex group
(G x / / ) , A is the direct product of the vertex groups G. and H*-

The definition of a pushout of groupoids (or of sets) is similar to

that for a pushout of groups. Thus we take groupoids GQ , G^, and 6^, with
homomorphisms i^.G^^ G^ and i^-G^ G^- Suppose that we also have a groupoid
G and homomorphisms JyO^ •* G and j'2'^2 "* ^ Then (G9j^ ^j^) is the pushout of
(/j, iy) if zy'j - I2J2 a n d f° r a n v groupoid H and homomorphisms 9yO^ -» H and
t$~\Gry -> H such that z\<p| - '9^2 ^ e r e *s a u n i < l u e homomorphism <p:C7-» H such
that <p|"y\<p and ^2'J2^' ^G^acin& groupoids and homomorphisms by sets and
maps, we get the corresponding definition of a pushout of sets.

When GQ is empty, the pushout is just the disjoint union of C/j and
G
2-
If we have a pushout of groupoids in which each groupoid is a
group then it is immediate that we have a pushout of groups. Conversely, any
pushout of groups is also a pushout of groupoids. To see this, we must
observe that a groupoid homomorphism <p from a group G to a groupoid H is
just a group homomorphism from G to HQj where a is l<p.
All examples of pushouts that are of importance to us have the
following extra property; V(GQ)'V(Gl)nV(G2) and V(G)'V(Gl)uV(G2)9 with the
relevant maps being inclusions on the sets of vertices. Pushouts with this
property will be called pointed pushouts.

Example 1 Let G b e a connected groupoid with V{G) - ia,b). Then

there is a groupoid H such that D(a,b) •* T(a,b)

G -> H
is a pushout square.
Further, H is connected, V(H)m ia9b), the map G -» H is a
monomorphism, which we regard as inclusion, and the vertex group HQm GQ*<t\
where t has infinite order. We show this by constructing a groupoid H with
these properties, and proving that it is the pushout. Note that these properties
define H up to isomorphism.
Let x € T{a,b) be the arrow from a to b, and let u be an arrow in G
from a to b. We define a groupoid H as follows: it has iatb) as vertex set, the
arrows in H from a to a form the group Ga*<t>, where t has infinite order,
and we denote this group by H ; the arrows in H from a to b form a set
bijective with HQ, and we denote the element corresponding to h by hu;
similarly the arrows in H from b to a form a set bijective with Z/^, the element
corresponding to h being written as u h, while the arrows from b to b are also
bijective with HQ, the element corresponding to h being written a" hu. The
multiplication in // is given by the obvious definition, and there is an obvious
inclusion of G into H. The required homomorphism from T(a,b) to H is given
by requiring it to send x to tu. We have to show that //, constructed in this
way, is the pushout.
Take any groupoid AT, and homomorphisms <p:G-> K andty:T{a,b)->K
such that 9 - ij> on D(a,b). Hence a<p - aty and £<p - bty. Let £ be JC^, which is an
arrow from a<p to b<p. Define $://-» K as follows. First define %• on 7/fl to be the
homomorphism from the group HQ to the group K which is cp on GQ and

which sends t to k(u<p)~*. Now define $ on all of H by requiring that, for all
h*HQ w e have (hu)&~ (h&)(u<p), (u~lh)&- {u<p)'l(h&), and (u~lhu)&- (u<p)~l(h&)(u<p).
Then $ is easily checked to be a homomorphism with the required properties,
and to be the only such homomorphism.

In particular, if G is also T(a,b) then HQ is infinite cyclic,

generated by xu" .

Example 2 We can generalise the previous example, and this

generalisation will be useful in Chapter 4. Let G and H be groupoids with
V(G)' ia,b)m V(H), and let H be connected. Let K be the disconnected groupoid
with V(K)~{a,b) and with KQm HQ and Kb~ H^ Given any homomorphism from
K to G there is a groupoid L such that we have a pushout
K-> G

H •* L .
The homomorphism from AT to G can be regarded as two
homomorphisms 0L:Ha-*Ga and fcH^G^. Since / / is connected, we can choose
an arrow u*H from a to £. We define P't/Z^-^G^ by Apf- (u~ hu)§. If (7 is
connected we also choose an arrow viG from a to b, and we define fT:// •* G
by APH - v(Apf)v .
We shall construct a groupoid L and prove that it is the pushout.
We require L to be connected with ia,b} as its set of vertices. Hence L is
determined by its vertex group LQ. We specify an arrow x from a to b in L.
If C7 is disconnected we define L to be the pushout of the
homomorphisms a and p\ If G is connected we define LQ to be the pseudo-HNN
extension corresponding to a and pH.
Suppose that we have homomorphisms from G and H to a groupoid
M such that the homomorphisms K~*G-*M and K-* H-*M are the same. First
suppose that G is disconnected. Then the homomorphism from G to M consists of
two homomorphissm V-Oa^MQ and ty-.G^M^. The homomorphism &:H->M
coincides with oc<p on H and with pd* on / / . . Let w be nd. For Ac// we have
a a
-1 -1 -1 * -1
hPty-(u lhu)$ty-(u lhu)&-w l(hoc<p)w. It follows that a<p-pf<Kw )#. Since LQ is
the pushout of a and p\ it follows that there is a homomorphism \y.L -* A/ such
that Xu-<p and X'u-<KW~ )*, where X and Xf are the homomorphisms from GQ and
Gb to Z,^. We have a homomorphism from (2 to L which is X on GQ and which
is X'JC* on Gp Similarly we have a homomorphism from H to L which is ocX on
HQ and which sends u to x. Then Z, with these two homomorphisms is the

required pushout, because we can extend u to a homomorphism from L to M by

requiring it to send JC to w, and this is the only possible homomorphism
satisfying the relevant conditions.
Now let G be connected, and choose an arrow vc(/ from a to b. A
homomorphism $ from G to M is given by a homomorphism 9'O -^ M and an
m
element y v$ of M. Let $://-» A/ be a homomorphism such that K-+G-+M is
the same as K-+H-+M. Then $-ot9 on // f l . Also, for Ac//^, we have
(u'lhu)^'(u'lhu)^'Lv'l(h^)vl(p'h^<py^ We see that A<x<p - ApXjmT1)*. The
remainder of the proof that the stated groupoid L is the required pushout will
be left to the reader.

Example 3 Suppose that we have a pointed pushout

i2 * * j\
Go -f G-
2
J2 3
and let S~ be a subset of V(G~ ). For r-0,1,2, let Sr be S^nV(G^9 and suppose
that Sf meets every component of Gf. Let //^ be (6>-)^ for rm 0,1,2,3. Then we
have a pushout square
Ho+ H
4, 4,

In particular, suppose that Gf is connected for f-0,1,2 (which implies that G^ is

also connected), and let QIV(GQ). Then we have a pushout square of groups

To see this, take a groupoid / / and homomorphisms y^H^H for

e
/•-0,1,2 such that on //Q we have 'i^j " 9 Q " ' 2 ^ 2 * ^ ^ a v e a l r e a ^ y shown how
to extend <pr to a homomorphism iJyGy-* //. We would like to do this so that

'1*1 "*0"'2*2*
In the construction we needed to take for each vertex v c Gf an
arrow X^BG^ from a vertex of 5^ to v, and the only condition in the general
construction is that xrv must be an identity if vcS^. Having made this choice,
4) is uniquely defined because Sr meets every component of Gf. We begin by

choosing suitable x^v, and then we make the additional requirements that for
any V*V{GQ) the arrow x must be x^ir for r-1,2. This ensures that
'iW'2^2-
It follows that there is a homomorphism tyyG^^ H such that
^1^3*^1 * 2^3mt^2'
an( J
^ e t ^3 ^ e ^ e r e s t r ^ c t i ° n °f ^3 t o Hy Then, for r-1,2, on
a s neec ec
//j, we have J/P^'^r* * * t o s n o w t n a t ^ 3 *s t n e pushout wanted.
We still have to show that <p^ *s unique. It is enough to show that
any <f>g can be extended to a corresponding 4^ with y 4^ • <\>r on G. For ^3 will
then be unique, since G^ is a pushout, and so <p^ w * ^ a ^ s o ^ e unique. Because
our pushout is pointed, we can find uniquely for each v e V(G) an arrow x^v
mx
such that JC3 rvfr ^ o r vcK(Gy) for r-1,2. Using these arrows to define ^3
w e see tnat nas tne
from <p^» ^3 required properties.

Exercise 1 Let Sj and 5*2 be sets. Show that we have a pushout

square of sets

u.
Exercise 2 Let the groupoid G^ be the pushout of GQ ->• C7j and
. Show that the set V(G^) is the pushout of K(GU)-» K(Gj) and
).
Exercise 3 Fill in the missing details in Example 2.

3.2 DIRECT LIMITS

We are used to limiting processes in analysis, but they are less
familiar in algebra. We shall define a procedure which will enable us to obtain
properties of infinite objects (for instance, a group which is not finitely
presented) from related finite objects.

Definition A directed set is a set A together with a relation £ on A

such that:
(1) if X £ u and \L * v then X £ v,
(2) X * X for all X,
(3) if X z u and u * X then X - u,
(4) for all X and n there is some v wih X £ v and u £ v.

Examples The natural numbers IN form a directed set with the usual
meaning of £. The positive integers also form a directed set when X*u means
that X divides u.
Let X be any set, and let A be the set of all subsets of X. Then A
is a directed set, with £ being inclusion.
Similarly, the set of all subgroups of a group G (or of all finitely
generated subgroups of G) is a directed set under inclusion.

Definition Let A be a directed set. For all XtA let G^ be a group.

For all Xand u in A with X^u let 9\.,:£*v "* O be a homomorphism such that
9 ^ is the identity map for all X and such that ^Xu^uv " ^Xv w ^ e n e v e r ^ u ^ v .
Then the collection of groups 6^ and homomorphisms <p, *s c a ^ e c l a direct
system of groups (and homomorphisms) indexed by A.
Replacing groups by groupoids gives the notion of a direct system
of groupoids. Replacing groups and homomorphisms by sets and maps, we also
have the notion of a direct system of sets.

Examples Let A be the set of all finitely generated subgroups of a

group G. Then we have a direct system of groups over A, obtained by assigning
to each member of A (which is, by definition, a subgroup regarded as an index)
the subgroup itself, and to each X * u the corresponding inclusion. It is
notationally confusing to use the same symbol for a subgroup regarded as a set
and for the same subgroup regarded as an index. Consequently, in this and other
examples, we shall refer to an index set A with subgroup G^ corresponding to X.

Definition Suppose that we have a direct system of groups over A,

and let G be a group with homomorphisms tyyG^^> G for all X. Then (C7,{<|K)) is
called the direct limit of the direct system if 4^ " ^XtAi ^ or a ^ ^ * ^ anc* ^ ^ or
any group H and homomorphisms /^:G^-»// such that f\mV\\J ^ or a ^ ^ ^
thre is a unique homomorphism f:G-± H such that A - I|K / for all X. We write
Gm limG^, omitting mention of the 4^- Similarly we can define the direct limit
of a direct system of groupoids or of sets.

As usual the direct limit is unique up to isomorphism.

Proposition 2 Any direct system of groups has a direct limit.

Proof Let K be %G^, and let N be the normal subgroup of K

generated by ((Sx^Xu^X > a ^ A i ^ and all g^G^}. The inclusion of G^ in AT
induces a homomorphism tyyG-^ ~* K/N. Suppose that we have homomorphisms
fy\G-^-*H such that / \ " ^Xu/ii ^or a ^ A ^ - Then we have a homomorphism
&.K+H, which is / x on Gx. Now ( ( ^ X ^ X ^ " ^ X ^ x / ^ x A ^ 1 " lf a n d s o
$ induces a homomorphism f.K/N-* H. It is now clear that KIN has the required
properties.//

Proposition 3 Let O limG^. Then G-UG^x- Also *x satisifes

*X^X m^ *ff there is some u* X such that ^x^Xu" *'
Corollary 1 / / each <p. is a monomorphism then each <J^ is a
monomorphism.
Corollary 2 Let H be a group, and let &yiCy*H be
homomorphisms. Suppose that ^"x " ^Xu^u f°r A i l i ' tnat ^ " ^ X ^ X an<*
g-^-^ " 1 iff there is some u* X such that S^9^' 1. 7^e/i / / - //w(7x .

Proof Corollary 1 is obvious from the proposition.

Under the hypotheses of Corollary 2, there is a homomorphism
H which is seen by the proposition to be an isomorphism.
We may take G to be (%G^IN', where N is the normal subgroup
generated by all (^x^Xu^X* * T h e n any g G
* ls a
P r o c l u c t (8\ 4>\ )—(^ <\>^ ) where
m
g* e Gv . Since A is directed, 3 u with u^X- for all /. Then g\ ty\ g\ <9\ 4^,,
A/ Az- f A/ A£* A/ A/(l (I
by the definition of A/, and so g* O \\) . as required for the first part of the
proposition.
Now take g^e G-^ such that gy^-^ " 1- Then, in the free product *C?a,
we have g* mU(u" v-w), where w. is arbitrary and v . - ( g . <p. )g. " for some
A I I I I * A/ A / U/ A/
X . a n d u . w i t h X . ^ ^ . a n d g^ e 6 \ . S o ] v w i t h v ^ X a n d v ^ j i . f o r a l l /. T h e r e i s
II I I A.\ A. j I

a homomorphism from *<3a to G^ which is (pav for all a £ v and trivial for all
other a. Under this homomorphism g^ maps to £^<PAV and v^. maps to 1. Hence
^ as
Examples Let C? be any group. We have seen that the finitely
generated subgroups of G form a direct system of groups. By Corollary 2, G is
the direct limit of this direct system. More generally, if m is any infinite
cardinal, the collection of subgroups of G with cardinality less than m is a direct
system with direct limit G.
Let (7- <X;R>. Consider pairs (Y,S) where Y is a finite subset of X
and S is a finite subset of RnF(Y). We obtain a directed set by defining

(Y^S^) if Y cY^ and S cS^. As usual, it is more convenient to denote

this set by A, and to denote the corresponding pair by (Y^,S^). Let 6^ be the
group <Yyj;S^>. There are obvious homomorphisms from 6^ to Gif X £ n, and
from 6^ to G. With these maps, G is the direct limit of the direct system (O-^).
This can be proved using Corollary 2 above. However, it is simpler
to prove the result directly. By Proposition 2 the direct limit has a presentation
whose set of generators is the disjoint union of the sets h\ and whose set of
relators is the disjoint union of the sets S^ together with extra relators which
equate an element of y\ with the corresponding element of Y, for \*\L An
easy use of Tietze transformations shows that this group also has the presentation
<Ur x ;U5 x >; that is, it is G
There are many variations on this example. For instance, let G be
<X;R>. Take X^cX and ^ c ^ n F f ^ Q ) . Then, by essentially the same argument,
G is the direct limit of the direct system of groups <y\ ;5\ >, where each J\ is
the union of XQ and a finite subset of X, while ^ c ^ c ^ n F f ^ ) (and S-^-RQ
finite, if we wish) for ail X. We gave, after the proof of Britton's Lemma, an
alternative proof in which it was necessary to reduce from the general case to
the case of finitely many stable letters. That reduction could be regarded as
applying Proposition 3 to this example.

Exercise 4 Let {5^} be a direct system of sets. If each <p^ is an

inclusion show that U5"x is the direct limit.
Exercise 5 Let the groupoid G be the direct limit of the direct
system of groupoids iO^). Show that the set V(G) is the direct limit of the
sets iV(Gy)).
Exercise 6 Fill in the details of the proof of Proposition 2.

4 THE FUNDAMENTAL GROUPOID AND THE FUNDAMENTAL

GROUP

4.1 THE FUNDAMENTAL GROUPOID AND THE FUNDAMENTAL

GROUP
We have seen that the set of paths in a space X has a partial
multiplication, and that this multiplication is associative and has identities. It
does not have inverses, but to each path / there is a path / such that both / . /
and / . / are homotopic to identities. Further, homotopy is preserved by talcing
products. It follows that there is an induced partial multiplication on the set of
homotopy classes y(X), and that y(X) is a groupoid under this multiplication.
We call y{X) the fundamental groupoid of X.
The vertex set of y(X) is just X itself (more precisely, it is a set
bijective with A" in a natural way; however, it is more convenient to regard it as
being X). The vertex group of y(X) at the vertex a is called the fundamental
group of X with base-point a; it is denoted by n^(x,a). It consists, by definition,
of the homotopy classes of loops based at a. For both y(X) and n^(X,a) we may
restrict attention to standard paths (or loops) and standard homotopies, and this
will usually be done without further comment.
The components of y{X) are the path-components of X. In
particular, X is path-connected iff "x(X) is connected. In this case n^(X,a) is
isomorphic to n^(X,b) for any a and b in X, and we usually denote the group
by nAX) unless there is a need to look at the base-point. The isomorphism is
X#, where X is the homotopy class of some path from a to b; if X is the class
of / we may denote the isomorphism by / # . If we take a different class y. of
paths from a to b it is easy to see that \L# is X# followed by an inner
automorphism; in particular, if n^{X) is abelian we obtain a unique isomorphism
from nJX,a) to nAX9b).
Let <p:X^Y be a continuous map (we will usually omit the word
"continuous", and just refer to a "map"). If / and g are homotopic paths in X

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:29:52, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.005
Fundamental groupoids and groups 75

then / y(Y). It is easy to see that if Z is another continuous map then (940* " 9*<1>* (so that we have a covariant
functor). Plainly, if a<p-b then we also have a homomorphism 9* from n^(Xfa)
to it^Y.b).
Readers are warned that when / is the inclusion of a subspace A
into X the corresponding /* need not be one-one. We will see examples of this
later. For the moment we just observe that if /* were one-one then any loop in
A which is homotopic to the trivial loop in X would have to be homotopic to
the trivial loop in A, and there is no reason to expect this.
If X is homeomorphic to Y then y{X) is isomorphic to y(Y). For we
have maps <p:X->Y and ty:Y-> X such that 9^ is the identity on X and 4>9 is the
identity on Y. Hence <p*ty* is the identity on y(X) and tp*<p* is the identity on
y(Y). We will give later on a weaker relation between X and Y which makes
Y(Af) isomorphic to

Proposition 1 y(X x Y) is isomorphic to y(X) xy(10 and

Y, {x,y)) is isomorphic to n^(X,x) X

Proof Let p and q be the projections of X x Y onto X and Y. Then

p*x q* is a homomorphism from y(X x Y) to y{X) x y(Y). Let / and g be standard
paths in X and Y, respectively. Define h by th'(tfftg). Then A is a standard path
m
in XxY such that A p - / a n d hq g. It follows that p**q* is onto.
To show that this map is one-one, it is enough to show that if h
and K are standard loops in X x Y such that hp and hq are standard homotopic to
h*p and h'q respectively then h is standard homotopic to h\ Let F and G be the
standard homotopies from hp to Kp and hq to Kq respectively. Then the standard
homotopy from h to h" is evidently //, where (t,u)Hm {(t,u)F,(t,u)G).
The result for the fundamental group follows at once.//

The map from / to the unit circle S sending t to (cos27rt, sin27i/)

is continuous, is one-one on CO,1), and sends 0 and 1 to P Q - ( 1 , 0 ) . It is an
identification, being a map from a compact space to a Hausdorff space.
Consequently, a standard loop in X based at x can be regarded as the same
thing as a map from S to X sending PQ to x.

The map from S x / to the unit disk Er which sends (p,t) to

tpQ+(\-t)p is continuous, is one-one on ( 5 - {p^}) x [0,1), sends 5 x 1 to p Q and
(p,0) to p. For the same reason as before it is an identification. The map from /
to S induces a map from / x / to S x/, which will also be an identification.
The composite of these two is an identification from / x / to Er. It then follows
that a standard loop based at x is standard homotopic to the constant standard
loop at x iff the corresponding map from S to X can be extended to a map
from Er to X. Also, two maps f,g:S -» X will give rise to homotopic standard
loops iff there is some map F:S *I+X such that (PQ,t)-x for all t, and
(pfi)F-pfi (p9l)F'pg for all p.

Let / and g be standard loops in X, based at a and b respectively.

We say that / is freely homotopic to g if there is F : / x / - » X such that (t,Q)F- tf
and (f,l)F-f£ for O^r^l, and {0,u)F- (l,u)F for O^u^l. In terms of a family of
paths, this amounts to saying that there is a continuous family fu of standard
loops (whose base-point depends on u) such that / n is / and / , is g.
1 ~2
We have another identification map from S x / to XT', this time
sending (p,f) to tp\ again, this leads to an identification map from / x / to Er.
Using this identification, it is easy to check that a map from 5* to X gives rise
to a standard loop which is freely homotopic to a constant standard loop iff it
extends to a map from Er to X. It follows that a standard loop based at x
which is freely homotopic to a constant standard loop is actually homotopic to
the constant standard loop based at x. The corollary to the next lemma deals
with the case of arbitrary standard loops which are freely homotopic.

Lemma 2 Let F be a map from / x / to X. Let / Q , fy gQ, and gj

be the standard paths given by tfo~ (t,0)F, tf^-{t,\)F, tgo-(O,t)F, and
tgx - (l,t)F. Then fQ.gx *gQ.fx.

Proof It is enough to show that iac*fQ-Si-fx>i~Q, where <2-(0,0)F.

In terms of a family of paths, at time u we apply F to the boundary of a square
of side u. Formally, the homotopy G is defined by

(r,a)O (f,0)F for 0 * t< u,

{t,u)G-{u,t-u)F for u<t±2u,
(t,u)G'{3u-t,u)F for 2u^t^3u,
{t,u)G- (0,4t/-t)F for 3u*t*4u. II

Figure 1 g i v e s an indication of what is happening. T h e map sending

t to (t,u)G is, for i i - l , 2 / 3 , and 1/3, the image by F of the paths in / x / w h i c h
are the boundaries (starting at A) of the squares ABCD, APQR, and ALMN
respectively.

N M

A L P B

Figure 1

Corollary Let F be a free homotopy between the standard loops f

and g. Let the classes of f and g be a in %^(X,a) and P in n^(X,b). Let X be
the class of the path h from a to b which sends t to (O,t)F. Then p - a X # .

Proof By the lemma, fh<*h.g, so that g^h.f.h. This gives the

result, by the definition of X#.//

Let 9 and ty be maps from X to Y. We say that 9 is homotopic to

<J> if there is a map $:XxI->Y such that (JC,O)- x.
Let <pQ*<pyX -> Y and let $ be the homotopy, and let ty:Y->Z and
X be maps. Then &9Q<|>S*&9I<|>> with the homotopy sending (w,r) to
since
|>. Further, if ^^^IY-^Z then (PQ^O * ^ l ' *0*0 * *0*l * * l * r
Let C be a subset of X. If there is a homotopy $ from 9 to i|> such
that (cfu)<b~c
are homotopic relative to C, and we write 9^4* rel C. In this notation, a
standard homotopy of standard paths is a homotopy rel {0,1} of maps from / to
X.
Let <p:X->Y be any map. We write 9:(AT,^/)-> (K,^), where i is in
some index set and A.£X and B.£Y, to mean that ApcBj for all 1 (and we use
similar simpler notations where relevant). We say that 9 and ty are homotopic as
maps from (XtA^) to (YfB-)9 and write 9 **ty:(X,A-)->(Y,Bj) to mean that there is
a homotopy * such that (A-x I)$£B. for all 1. Note that we may have maps

<p,ty:(X9A)-> {Y,B) such that <p*ty:X -+Y but for which we do not have
9 <*i\r.{X,A)^ (Y,B), since the latter imposes an extra condition on the homotopy.
Let <p*ty:(X,a)->(Y,b). Let / be a standard loop based at a. Then /<p
and /ty are homotopic standard loops based at b. Hence the homomorphisms 9*
and tjj* from n^X.a) to n^{Y,b) are the same.
More generally, let <p^^:Ar-»K, and let a<pmb and atymc. Let $ be
the homotopy from cp to (p. Then there is a free homotopy from / 9 to /ij> which
sends (tfu) to (tf,u)$. It follows, by the corollary to Lemma 2, that the
homomorphisms 9*:7i|(Ar,tf)-»7tj(yr,&) and <Jj*:7T|(A'.<2)-»7r|(K,c) are related by
(Jj*-<p*X*, where X is the class of the path which sends u to (a,u)$. We shall
not investigate the connection between 9* and i|>* as homomorphisms from y(X)
to Y(K), as this is a little messy.

Let \j£ and 1^ be the identity maps on X and Y. A map 9:^-* K is

called a homotopy equivalence if there is a map ty:Y-*X such that <p4>M^ and
4><P^ly. We then say that X is homotopy equivalent to Y, and that <J) is a
homotopy inverse to 9. It is easy to see that homotopy equivalence is an
equivalence relation. Plainly homeomorphic spaces are homotopy equivalent. We shall
give later examples of many different spaces all of which are homotopy equivalent
to a point (and hence to each other).
Suppose that X is homotopy equivalent to Y and that X is path-
connected. Then Y is also path-connected. For the homotopy between ty9, and ^ 9 is path-connected
because X is.

Proposition 3 Let <p:X -» Y be a homotopy equivalence, and let

a<pm b. Then <$*:iiAX,a)-*nAY,b) is an isomorphism.
Proof We have to be careful because 9 need not be a homotopy
equivalence from {X,a) to (Y,b).
So let bty~c, and let c<p-d. We have fp*:n^(X,a)-^nAY9b) and
^*:n^(Ytb) -»7tj(AT,c); we also have a homomorphism from n^(Xtc) to n^(Yfd)
induced by 9, and this will be denoted by 9 ^ . By the previous remarks, there is
a homotopy class X of paths in X from a to c and a homotopy class \i of paths
in Y from b to d such that 9*^*- X# and <J)*9^ - (x#. Since X* and u# are
isomorphisms, it follows first that ty* is an isomorphism, and then that 9* is an
isomorphism.//

A space X is called simply connected if it is path-connected and

iz^(X) is trivial. Plainly a space consisting of a single point is simply connected;
we will give some other examples of simply connected spaces shortly. Examples
of spaces which are path-connected but not simply connected will be given in
the next section.
Let X be simply connected. By Lemma 2.11, since any loop is
homotopic to the trivial loop, we see that any two paths from a point a to a
point b are homotopic. Hence there is exactly one homotopy class of paths from
a to b, for all a and b. In other words, y(X) is the tree groupoid on the vertex
set X.

We say that the space X is contractible to the point a if there is a

map F:XxI-*X such that (JC,O)F« x and (x,l)F- a for all xtX. Equivalently, X
is contractible to a if the inclusion {a} •* X and the trivial map X -*{a} are
homotopy inverses, and so X is homotopy equivalent to {a}.. It follows that a
contractible space X is simply connected, using Proposition 3 and the remark
before it. It also follows that X is contractible to each of its points. We refer to
the map F a s a contraction.
We say that X is contractible relative to a (or rel a) if there is a
map F:XxI->X such that (JC,O)F- x and (x,\)F- a for all x c X and (a,t)F~ a
for all t. We shall see that a space may be contractible relative to one of its
points but not relative to another.

Example 1 A subset X of Un is called star-shaped from at X if for

every JC e X the line-segment joining x to a lies in X. Examples of star-shaped
sets include R^ itself, the closed disk //*, the open disk which is the interior of
[f1, the /z-dimensional cube /*, and many others. It is easy to see that a set is
convex iff it is star-shaped from each of its points. Figure 2 shows an example
of a star-shaped space (it is star-shaped from V) which explains the name. This
space is not convex, since the line-segment joining A to B does not lie entirely
in the set. In a star-shaped art gallery there is a place from which all the
pictures can be seen!
The line-segment joining x to a is {ta + (1 - t)x; O^t^l). It follows
that if X is star-shaped from a then X is contractible, the contraction being
given by (x9t)F- ta + (l - t)x.

Figure 2

Example 2 Let X be any space, and CX the cone on X. Then CX

is contractible relative to its vertex v. We have a map F:(X * I) x /-» X x / given
by (jc,f,i/)F-(x,fi/). This induces a function G from CA'x/ to C*. Now G is
plainly continuous at all points other than the points (v,w), and G will be the
required contraction if it is continuous at these points also.
By definition of CX as an identification space, a subset of CX
containing v is open iff it is the image of an open set U of X*I with X*0cU.
Given such a set U, consider the set Vm {(x,t)\ x xlQ,t]cU} which contains X*0.
If (x,t) c V then the compactness of CO,f] ensures that there is an open set W of X
contining x such that W*l.Q9ficU. It is now easy to see that V is open. Further,
the image of V x / in CX x / plainly maps by G into the image of U. Hence G is
continuous at the points (v,u), as required.
Example 3 It is easy to check that Sn- {(0,... ,0,1)} is
homeomorphic to R^, and that, for any p € Sn, Sn - {p) is homeomorphic to
Sn- (0,...,0,1)}. It follows that Sn-{p) is simply connected for any p.
Then Sn is path-connected if /*/0. If / is a standard loop such that
n
If*S it follows from the previous paragraph that / is homotopic to the trivial
loop. Unfortunately, even when n > 1, it is possible to find loops / with / / - Sn
(these are the space-filling curves due to Peano and others). However, the
argument at the start of the proof of van Kampen's Theorem in the next section
shows that any loop / may be written as a product /j /^ where each / . is
n n
either a path in S - {(0,... ,0,1)} or in S - {(0,... ,0,-1)}. Since each of these
spaces is simply connected, each ft is homotopic to a path g> whose underlying
set is a circular arc. Then f*g, where Ig is the union of a finite number of
circular arcs. Hence Ig?Sn when n>\. Then g is homotopic to a trivial loop,
and hence so is /. We deduce that Sn is simply connected for n>\.
A little geometric intuition should persuade the reader that Sn
cannot be contractible. A proof of this would require consideration of the higher

homotopy groups TCn(Sn) or the homology groups H£Sn), neither of which will
be discussed in this book.
2
Example 4 The Fox comb is the subspace X of R given by
X - (0 x I) u (7x1) u U((l//i,r); 0 * r * l , all positive integers n).

Figure 3, showing (part of) the Fox comb.

It is easy to check that X is contractible relative to (1,0), and even

that it is contractible relative to any point not in 0 x [0,1).
However X is not contractible relative to (0,0), and (by a similar
argument) is not contractible relative to any point in Ox[0,1). For let F:X*I->X
be a map such that (p,0)F-p for all p and (0,0,t)Fm (0,0) for all t.
Let U be {(JCJOCR2; \y\<\/2). Since ((0,0) x I)Fc U, there is some
8 > 0 such that {p x I)Fc U if \\p\\ < 8 (for instance, using uniform continuity of the
second coordinate of F and the fact that this coordinate has an attained
maximum on (0,0) x 7 which is less than 1/2). In particular, if n is large enough,
((l/«,0) x I)F must lie in that path-component of X n U which contains (l//i,0).
This path-component is easily seen to be {(l//z,f); 0^r<l/2). Since this does not
contain (0,0), we cannot have (l/nt09l)Fm (0,0), and so F is not a contraction.
Let -X be {(JCJ>); (-x,-y) € X). Then -X is also a Fox comb. Let
m
Y Xu-X, which is the union of two Fox combs with (0,0) as their only
common point; that is, Y is the join of two copies of the Fox comb.. Then Y is
not contractible.

Figure 4, showing (part of) the join of two Fox combs

First note that the function r.Y+X given by prm p, (-/>)/•-(0,0) for
all p*X is continuous, by the Glueing Lemma. It follows that if we had a
contraction F of Y relative to (0,0) then Fr, restricted to X x /, would give a
contraction of X relative to (0,0), which we know is impossible.
Suppose that we could have a function F:Yx/->Y such that
(p,Q)F-p for all p*Y. We shall show that we must have (0,0,r)F-(0,0) for all r,
and the previous paragraph then tells us that F is not a contraction.
The set of all c such that (0,0,r)F-(0.0) for all t^c is plainly a
closed subset of / containing 0. If we show that it is also open it must be the
whole of /, which is what we want to prove.
Suppose that c has the above property. Then we can find d > c such
that (0,0,/)eU for t±d. Then, as before, we can find 8>0 such that (p,t)iU if
llpll<8 and t<d. As before, we find that, for n large enough, ((l/«,0) x [Q,d1)F is
contained in {(\/n,t)\ 0<t<\/2). In particular, for any fcCO,^], continuity of F
tells us that (0,0,r)F- (0j>) for some y^0. Similarly, looking at (-l//z,0), we see
by continuity that, for any fc[O,tf], (0,0,r)F- (0,z) for some z<0. Thus (0,0,r)F
must be (0,0) for all t±d, which gives the result.

Exercise 1 With X and U as in example 4, show that the

path-component of (l//i,0) is {(l//z,f);O£f*l/2).
Exercise 2 Find a homeomorphism from OR* to Sn- ((0,... ,0,1)}.

A B

Figure 5, showing a homeomorphism from R to S - {(0,1)}.

A and B are mapped to a and b, respectively.

Exercise 3 Show that there is a homeomorphism from Sn-ip) to

Sn- {(0,.. .,0,1)} for any pzSn. (You might want to show first that there is a
homeomorphism from Sn-ip) to Sn-{q) for some suitable q which has one more
coordinate zero than p has. Alternatively, use the fact that any unit vector is
part of an orthonormal basis.)
Exercise 4 Show that homotopy equivalence between spaces is an
equivalence relation.

Exercise 5 Can you think of any reason why we might expect

71,(5" ) to be non-trivial? Why we might expect it to be cyclic? To be infinite
cyclic? Don't try to find a convincing argument; just see if you can see anything
which suggests what kind of answer we should look for.

4.2 VAN KAMPEN'S THEOREM

So far we have not been able to calculate any non-trivial
fundamental groups. In this section we shall find a way of doing this, and we
will show that any group occurs as the fundamental group of some space.

Theorem 4 (van Kampen's Theorem) Let X^ and X2 be subspaces

of a space X, and let X be the union of the interiors of X^ and X2 - Then the
inclusions induce a pushout diagram of groupoids

i2*

Proof Let Un be the interior of Xn for n- 1,2. Let filO,rl + X be a

path. Let 8 be the Lebesgue number (see Proposition 2.5) of the covering
{t/j/1,^/1} of
tO»1 Choose a partition $ of CO,/-] of mesh less than 8; that is,
choose points u - 0 < t* < u < . . . < t, - r for some k such that t. -1. , < 8 for all /.
Then / is the product of paths / * , . . . J^ such that each / . is either a path in X+
or a path in X2, where fi is defined on LO,t.- t.^l by tf^ (t +1^f. This
property was used in Example 3 of the previous section.
Let (pyy(X^) -» G and <p2:y{X2)-> G be homomorphisms into a
groupoid G such that /]*<p^ - ^2*^2 * W e c a n ^ ^ e a c ^ ^/i ( /I '^»2) to a
multiplicative map (Ji^ from the paths of Xn to Cr, and these two will agree on
the paths of X^ n X2. It follows that to each partition ^> as in the previous
paragraph there corresponds an element of G, namely the product of the elements
ffin. (where / . is a path in X ). We need to be sure that this product is
defined. This follows because the trivial path at the point t^f. is a right identity
for / . and a left identity for /. + * , and so its image in G will be a right and a
left identity for the elements of G corresponding to / . and /-+« , and the product
is therefore defined.
Because each tyn is multiplicative, it is immediate that adding one
extra division point to the partition ^) does not change the resulting element of

G. Inductively, adding any number of new division points does not change the
element. Now let $ and ft be any partitions, and let SK be the partition which
has the division points of V together with those of ft Then the element of G
corresponding to $ is the same as that corresponding to 9t, which is in turn the
same as that corresponding to ft It follows that the element of G depends only
on the path / and not on the partition 9> chosen (provided that V satisfies the
required condition). We denote this element of G by /i|>. In particular, if / is a
path in X then ftym ftyn. To make the notation easier to read, we shall write
t|> instead of ty in future. Also, 4> is plainly multiplicative.
If we can show that <J> is constant on each homotopy class it will
an
induce a homomorphism 9 from y(X) to G. By construction, 9^ " y^* d 9 is
plainly the only homomorphism with this property. This will show that we have
a pushout, as required.
We first observe that if / is the product of the paths / « , . . . , / , as
above then its standardisation fQ is the product of paths gj, . . . , g £ . Here each g.
is obtained from the corresponding f. by a change of scale, and so £ z ^/ z -. Since
i|> (on each subspace X ) is induced from a function defined on homotopy
classes, we see that gjtymffi Hence fotymfty. It follows that we need only show
that cp is constant on standard homotopy classes.
Let / and / ' be standard paths, and let F be a standard homotopy
between them. Let 8 be the Lebesgue number of the covering {U^F \lJ2F } of
/x/. Take an integer k with / 2 / £ < 8 . Then F maps any square of side XI k into
one of X^ and X^ •
Let Z>.. be the square of side \lk whose lower left vertex is
(i/kj/k). Let / . . and g.* be the paths defined on [0,1//:] given by
tf.m(t+i/kJ/k)F a n d t g . . - ( i / k , t+j/k)F r e s p e c t i v e l y . L e t e.. a n d e*.. b e t h e
horizontal and vertical edges of £>•• beginning at {i/kj/k), which we may regard
as paths of length \/k. Then / . . and g.. are the composites of these paths with
F. In Figure 6, the square ZX, j and the edges ^ 1 > ^2 2 ' e*2 1' and e
3 1 are

indicated.

e*U ^2,1 '

3,1

"2,1

Figure 6

By Lemma 2 (with a change of scale) applied to Z>.., we know that

g
^ i/i j+\y ^ e
homotopy taking place entirely in X^ or entirely in A^.
Since is constant on homotopy classes of paths in X^ or X^ , we see that
m
l*ift*fij+lW f o r O*iJ<k.-
An easy induction shows that, for r<k, we have

We use this with rm k-l, and a further easy induction shows that, for s<k, we
have

•JMi s i - k-\
0

Now / is the product of the / ^ , while / ' is the product of the / ^ .

Also each g^- is the constant path at the point a, while each g,- is the constant
path at the point b, and so their homotopy classes are identities of either the
groupoid y(X^) or the groupoid Y ( ^ 2 ^ ^ e n c e e a c ^ SQJ^ and g^to is an identity
in the groupoid G. Applying the formula above with smk-\ now shows that
/ty-/\|>. This completes the proof of the theorem.//

The following corollaries are immediate, using Example 3 on

pushouts of groupoids.
Corollary 1 Let X^ and X^ be subspaces of X such that X is the
union of the interiors of Afj and X~ . Let V be a subset of X such that
VnX*,VnX2, and Vn X+c\ X~ meet every path-component of X., X~
X^ n A*2 respectively. Then the inclusions induce a push-out square

Corollary 2 Let X* and A^ be path-connected subspaces of X

such that X is the union of the interiors of Aj and A 2 • Let X^n X^ be
path-connected, and take some point a in ATj n A^ . Then the inclusions induce a
push-out square of groups

Most authors do not develop the theory of groupoids, and

consequently they discuss the fundamental group but not the fundamental
groupoid. As.a result, Corollary 2 is usually referred to as van Kampen's
Theorem. Our version of the theorem is slightly stronger than this one. For
instance, we will be able to find Ttj(A) even when Aj n A^ is not
path-connected. Our version, though complicated to prove, is nonetheless easier
to prove than the more usual version. If we do not wish to use groupoids at all,
it is neccesary to have various auxiliary paths in order to replace our paths /-•
and g.. by loops. Our proof does not get rid of these auxiliary paths entirely,
but it separates out the topological and algebraic aspects of the proof; we used
such auxiliary paths when we discussed Example 3 on pushouts of groupoids.
It was primarily for their use in van Kampen's Theorem that we
introduced groupoids. Most of our results in future will be expressed in terms of
the fundamental group. When X is path-connected, the fundamental groupoid of
X is determined by the fundamental group of X together with X itself (as a set).

Example Let X be 5 1 , let a be (1,0), and let b be (-1,0). Let Xx be

S - {(0,-1)}, and let X2 be S1 - {(0,1)}. We cannot apply corollary 2, since
l

X^ n X2 is not path-connected. However, Aj and A^ are contractible, while

A* n X2 consists of two contractible components, one containing a and the other
containing b. It follows that Y(^j){a ^ and ^(^2\a b) a r e ^ o t ^ t r e e sroupoids on
ia,b), while y(X^ n ^ ^ { a b) ls ^ e ^^ s c r e t e groupoid on [a,b). Corollary 1 can
now be used to give a push-out square whose structure has already been analysed
(see the special case of Example 1 on pushouts of groupoids).

We find from this that n^(S \a) is infinite cyclic, and its generator
is the class of f.g, where / is a path in X^ from a to b and g is a path in Xy
from b to a. In particular, the generator is the class of the standard path which
sends t to (COS2TU/, sinlnt). We will obtain another proof of this result, and
related stronger properties, after we have developed the theory of covering
spaces.
Notice that the inclusion i:S •» Lr does not have /* one-one, since
KAU) is trivial but nAS ) is not trivial.

The subspace A of the space X is called a retract of X if there is a

map r:X->A such that arm a for all at A; the map r is then called a retraction.

Proposition 5 Let A be a retract of X, and let i:A •> X be the

inclusion. Then i*nzAA)-> it AX) is one-one.

Proof Since ir is the identity on A, we see that ir is the identity

on n^(A). II

Theorem 6 (No Retraction Theorem) 5"1 is not a retract of Er. II

Theorem 7 (Browser's Fixed-point Theorem) Let f.Lr -» D1 be a

continuous function. Then there is some x e Lr such that xfmx.

Proof Suppose not. Define g:Lr -» S by requiring JC# to be the

point where the line segment from xf to JC, when extended, meets 5 . Because

Figure 7

we are assuming that xf is always distinct from JC, the function g is defined for
all j c e / r . Plainly xg~ x if x*Sl. It is easy to check that g is continuous (for
instance, by finding a formula for g). This contradicts the previous theorem. //

More generally, S is not a retract of //*, and any continuous

function from tf1 to itself has a fixed point. These are proved by almost
identical methods, but using the higher homotopy or homology groups.
When A> A^ u A^ but X is not the union of the interiors of X^ and
A^, the conclusion of van Kampen's Theorem need not apply. For instance, we
shall obtain in the last section two contractible spaces whose join is not even
simply connected.
We frequently need to use van Kampen's Theorem when X - X+ u X~
with AT|, X2, and A^ n X^ all path-connected but X is not the union of the
interiors of X* and X^ • This can often be done by enlarging the spaces X. and
X2 • More precisely, suppose that we have open subspaces Y^ and Y2 of X such
that Xns^Yn for / i - l , 2, with Y^, Y^* and Y^ n Y^ all path-connected, and such
that the inclusion of A^ into Y^, of X^ into ^ , and of X^ n A" 2 into Y^ n y^ all
induce isomorphisms of the fundamental groups. Then we can apply Corollary 2
to Kj and Y2, and then use these isomorphisms to get the corresponding result
for A'j and A^. In particular, this holds if we can find Y^ and Y^ such that A^
is a deformation retract of Ym (defined below) for /z-1, 2, and XAnX~ a
deformation retract of Y* n ^ •

Definition The subspace .<4 of the space X is a deformation retract

of X if there is a retraction r.X-> A such that ^ ^ l j ^ » where i is the inclusion
and ljj^. is the identity on X. If the homotopy is relative to A we say that A is
a strong deformation retract of X.

Examples X is contractible to p iff p is a deformation retract of X,

and AT is contractible rel p iff p is a strong deformation retract of X.
For any space X, X x 1 is a strong deformation retract of A" x / and
of ATx(O,lJ.
Lemma 8 £er ^4 />e a strong deformation retract of A^, tf/ztf let
f\A + ATj &e a map. r/re« A^ is a strong deformation retract of the space

Proof Let <p:A^ x /-* A^ be a map such that (jc,O)<p-x, (jc,l)<p€v4 for
>
all JC€A 2, and (tf,f)<p - a for all <2€,4 and nl.
Since AT^ ^yA^ is obtained from the disjoint union of A", and A^ by
identifying a and af for all a €,4, it follows from Proposition 2.4 that
(A'j u*X^) x / is obtained from the space obtained from the disjoint union of
A'| x / and ^ x ^ by identifying (a9t) and (af9t) for all a*A and Ul.

Then, by Proposition 2.3, we obtain a continuous function on

(A"j "fX^) x / by defining continuous functions on X^ x / and A^ x / which are
compatible with this identification. Thus we get a homotopy proving the result
by taking the functions which send (JCJ ,r) to JCj and (JC2 ,t) to (JC2 ,^9
respectively.//

Example Let X be the join of the two path-connected spaces A^

and A^ with the common point p. It follows from the definition of the join that
each of X^ and X^ is a retract of X. In particular, the inclusions induce
monomorphisms, which we treat as inclusions, of 7tj(A"j,p) and ^(A!^,/*) into

If AT2 is contractible rel p we find that X^ is a strong deformation

retract of X. Further, if both A^ and A^ are contractible rel p, it follows from
this that X is contractible rel p.
Now suppose that A^ and A^ have open sets U^ and £/« containing
anc u are open sets
p which are contractible rel p. Then A^ u L^ * ^2 ^1 °* X
whose union is X, and they have X^ and A^ respectively as strong deformation
retracts. Also their intersection is (/, u ^ , which is contractible. Applying van
Kampen's Theorem to these sets, and using the relevant isomorphisms, we find
that n^(X, p) is the free product of n^(X^3p) and 7Tj(A^,p). We will give another
condition later for this to hold, and we will show that it does not hold in
general.
This example can be generalised. Let X be the join of n
path-connected spaces A"., the common point being p, and suppose that for all 1
there is an open set of X-t which is contractible rel p. It is easy to check that X
is the join of X* v . . . v Xn_^ and X , so that we can prove results by induction.
In particular, we find that there is, for all r, an open subset of A^ v . . . v Xf
which is contractible rel p. We also see that n^(X) is the free product of the
n^(X-). We shall look at infinite joins later in this section.
As a special case of this, the join of n circles is free of rank n.
As a further example, let Y be the join of a path-connected space X
and the interval /, where UI is identified with some point of X. Then AT is a
strong deformation retract of Y, and so n^(Y) is isomorphic to n^(X). The reason
for looking at Y rather than X is that the point 0 in Y has (arbitrarily small)
open neighbourhoods which are contractibie rel 0, while there may be no points
with a similar property in X.

Definition Let S be a subset of a group G. Then the group G/<S>

is said to be obtained from G by cancelling S. (I do not want to say that it is
obtained by factoring out S, as this phrase is best used only when S itself is a
normal subgroup of G. The process is often referred to as "killing S M, but there
is enough violence in the world already without introducing it into
mathematics).

Proposition 9 Let A and X be path-connected spaces, and let

f:(A,aQ)->(X,XQ) be a map. Let Y be X v^CA, where A is regarded as the
subset A*{\} of the cone CA. Then TZJY,XQ) is %JX,XQ) with nAAfa^)f*
cancelled.

Proof Let Y^ be the open subset X uJA x(0,l]), and let Y2 be the
open subset consisting of all (a,t) with t<\. Let gM-^Kny^ be the map given
by ag' (a,l/2), and let b^ be aQg.
Then g is a homotopy equivalence (in fact Y+ n Y~ is homeomorphic
to A x (0,1) ), and y^ is contractible. Hence van Kampen's Theorem tells us that
7Uj(]K, Z?Q) is obtained from ^^X\» ^Q) by cancelling nÂ, a^)g*.
Now let X be the class of the path which sends t to (<3Q,r+ 1/2).
Then X* is an isomorphism from ^1(^1 » ^Q) to 7ti(K. , <2Q). Also, since (<z,l) is
identified with af9 we see that g^f, where the homotopy sends (a9u) to
(a, (w+l)/2). By the corollary to Lemma 2 it follows that ^*X#-/*. Hence
n^{Y, JCQ) is obtained from 7TJ(^,JCQ) by cancelling nÂ, a^)f*.
Finally, Lemma 8 tells us that X is a strong deformation retract of
Kj, and so the inclusion induces an isomorphism from n^(Xf x^) to î(î » XQ)>
giving the result.//

Remark Let a and b be points of the path-connected space X. Now

5°-{-1.1), so we may regard CS° as the interval C-1,11. Let fisP + X be given by
(-1)/- a and 1/- b. A similar proof at the groupoid level shows that
nAXufCS ,a) - 7r1(AT,<3)*<a>, where a has infinite order and is the class of the
^ 0
path g.hy where g is the obvious path from a to b in CS , while h is some path
from b to a in X.

Proposition 10 Let X be a space. Let A be a directed sett and for

each XeA let X^ be a subspace of X such that X^c.X if \<[L. Suppose that
for every compact subset of X there is some X such that X^ contains this
subset. Then y(X) is the direct limit of the groupoids -y(Xy), the

homomorphisms being induced by inclusions. Further, if there is a point JCQ such

that *Q*X\ for all X then n^X, XQ) is the direct limit of the groups n^(X^f JCQ).

Proof The result for the fundamental groups is identical to that for
fundamental groupoids, replacing arbitrary paths by loops based at JCQ.
The proof is similar to the proof of van Kampenfs Theorem. Let P
be the set of paths of X, and let P^ be the set of paths of X^. If / e P then
[0,13/ is compact, and so, by asssumption, it lies in some Xy Hence PmUP^,
and / \ £P if X < \JL. For the same reason, any homotopy between two standard
A [I
paths takes place entirely in some X-^.
Let G be a groupoid and let <p^:y(X-^)-> G be homomorphisms such
that 9 \ " ZX * ^ ^ A<
& where *\ is the inclusion of X^ in X .
Then <PA lifts to a multiplicative map ty^:P^-> G, and, for X<jx, t|^
is the restriction of I|J to A . For any X and u there is some v with v ^ X and
U A

v > u. On P^ n P we then see that ^ - <b , since both equal (J^. It follows that
we may define uniquely a map ty:P-*G by requiring (p to be ij^ on / ^ .
Let / and g be paths in X such that /.£ is defined. We take X and
u with / and g in / \ and P , respectively. We can find v with v * X and v ^ u,
A {I

and we can then regard / and g as elements of P^. Since 4>v is multiplicative, it
follows that If <JJ induces a function cpryCA') -» C? it will be the required
homomorphism (and is easily seen to be the only such homomorphism). So we
need only show that ty is constant on homotopy classes. Let / and g be
homotopic paths in X. Then there is some X such that / and g are homotopic in
A^ . Then /ty-/4>A and gip- gty^ , and /ty^ - giJ>A , by definition of 4>A , because /
and g are homotopic in X-^ .11

Remark Let X be the union of subspaces X^ for a in some index

set A. Let A be the set of all finite subsets of A. When X is ( a , , . . . , a } let X^
in A
be A"a u...uXa . Then A is directed and X^cX if X<^t. Proposition 2.2 can
often be used to show that every compact subset of X is contained in some X^.
In particular, let X be obtained by adjoining a collection of spaces A^ to a
space B, and let X^ be obtained by adjoining A^ to B. Then X has the weak
topology defined by the subspaces X^, and Proposition 2.2 applies.
Example Let X be the join of the spaces X^, the common point
being p. Suppose that each X is a 7Vspace and that each X has an open
subset contractible to p. With the above notation, when X is taj,... ,<x^}, we

have already seen that the group n^(X-^,p) is the free product of the groups
nJX , p) for i - l , . . . ,/z. Also nj(A\p) is the direct limit of the groups
n^(X^9p). It follows that %^(Xtp) is the free product of the groups n^X^p).
In particular, suppose that each X^ is a circle. In this case we call
X a bunch of circles. The discussion above shows that n^(X) is free of rank \A\,
where A is the index set.
2
Warning The subspace of 1R consisting of those (x,y) such that
x +y - 2x1 n for some positive integer n is the union of countably many circles;
namely, the circles with centres (l/«,0) and radius \/n. However its topology is
not that of the join of countably many circles, and its fundamental group is
complicated; it will be discussed further in the last section of this chapter.

Proposition 11 Let X be a path-connected space, and let the

spaces A- {for i in some index set) be path-connected. Let f.:(A.,aj)-> (X, x^)
be maps. Let Y be obtained from X by adjoining the spaces CA- by the maps
fj. Suppose either that the index set is finite or that each A^ is a T^-space.
Then n^(Y, x^) is obtained from n^(X, x^) by cancelling all the subgroups

Proof If the index set has only one element the result holds by
Proposition 9. The result for a finite index set follows by induction. The general
case holds using Proposition 10 and the remark following it.//

The result extends to the case where we do not require that

a
i^i" 0' * ^ a t c a s e w e m u s t u s e t^ie isomorphism induced by a path to map
x n

7Tj(y4j., <*•)/•* into 7Cj(A\ XQ). A different path will change this homomorphism by
an inner automorphism, and this will make no difference when we cancel.
As a particular case, we can take each A- to be S , so that CA; is
Er. The map / . can be regarded as a loop whose homotopy class will be
' i

denoted by a.. Since it AS ) is infinite cyclic, we see that TtJAJf-* is the

subgroup generated by a-. So the following corollary is immediate.
Corollary Let X be a path-connected space, and let
f.:(S , a)-»(X, XQ), where a is (1,0), be maps representing the elements a. of
7ij(AT, JCQ). Adjoin copies of the disk Er to X using the maps f^, and let Y be
the resulting space. Then 7Cj(K, XQ) is obtained from n^X, x^) by cancelling all
the elements OL..//

When we wish to calculate homotopy groups, Proposition 11 and its

corollary are often used in preference to using van Kampen's Theorem.
For instance, we may regard / x / as sa square with vertices A, B, C,
and D, and we may obtain spaces by identifying edges of the square. If we
identify the edges AB and DC, and the edges AD and BC, we obtain the torus,
while if we identify AB and CD, and AD and BC, we obtain the Klein bottle.
These are shown in the figure below. The torus is easily seen to be
homeomorphic to S x S , so we know, by Proposition 1, that its fundamental
group is ZxZ. But we may also use the discussion above. The space obtained
from the edges of / x / after the identifications is just the join of two circles, one
coming from AB and the other from AD. The fundamental group of this is free
of rank 2, with generators x and y, say. The torus is obtained from this by
adjoining (a homeomorph of) Er, and the element which is to be cancelled is
plainly xyx y . Thus the fundamental group of the torus has presentation, as
we would expect, <x,y, xyx' y~ >. The Klein bottle is left as an exercise.

D D

B A B
torus Klein bottle
Figure 8

We can now see that groups and their homomorphisms can be

represented by spaces and their maps.

Theorem 12 Any group is the fundamental group of some T^-space.

Proof Let G have presentation <A;R>. Let X^ be a bunch of circles,

one for each element of A. Then n^(X^) is (isomorphic to) the free group F(A).
Now obtain X by adjoining to X^ copies of Er using maps from S
to AT| representing the elements of R. It is easy to check that AT is a 7\-space.
The corollary above shows that nJX) has presentation <A;R>.//

Theorem 13 Let Y be a path-connected space, and let G be a

group. Let <pm^(Y,yQ)^ G be a homomorphism. Then there is a space X
containing Y, which is a T^-space if Y is a 7^-space, such that nJX,yQ) is G
and the inclusion of Y in X induces <p.

Proof Take a T^space XQ with ^(ATQ) being G. Let Kj be the join

of Y and / with y^ being identified with 1, and let X^ be the join of XQ and /.
Let Z be the join of X^ and Y^ at the point 0 of /. We have seen that Y^ and
AT| have the same fundamental groups as Y and XQ , and that TT^Z) is the free
product of 7ij(K) and G (since G is 7TJ(ATQ) ). Let B be a set of generators of
7ij(yr,j>Q). By adding discs to Z we obtain a space A" (which is 7^ if K is T^)
such that TTJCA'^Q)) is obtained from TCJ(K,^Q)* C7 by cancelling the elements
) for all &€#. Thus TC^X,^) is just G, and the homomorphism induced by
the inclusion of Z in X sends b to &<p. Hence the inclusion of Y in X induces
9.//

Corollary I Let G and H'. (for i in some index set) be groups, and
let <p:H-*G be homomorphisms. Then there is a T+-space X and subspaces Y-
with y . n y . " CJCQ) for i*j and such that Tiy(Xyx^) is G, Wj(K., JCQ) is Hi$ and
the inclusion of Y'. in X induces <p..

Proof L e t H> h a v e p r e s e n t a t i o n < A . ; R . > , w h e r e A n A * Q f o r /*/.

Then ^H- has presentation <Uy4y ;U^.>. The space K we constructed with
fundamental group %H^ plainly has subspaces Y- with fundamental group H- and
such that Y.nY.- {XQ) for i&j, because A- is disjoint from A.. Now embed Y in
a space X so that the inclusion induces ¥9..//

Corollary 2 £ef G be a group. Let Y. be spaces such that there is

y. € Y. with an open set which is contractible re I y^. If the index set is infinite
suppose that each Y- is a 7\-space. Let 9 .:TT,(K. ,y.) -» G be homomorphisms.
Then there is a space X (which is T^ if the Y. are T^) containing subspaces
(homeomorphic to) Y'. such that the union of the Y'. is their join and such that
n^(X) is G and the inclusion of Y'. in X induces 9^.

Proof Let Y be the join of the Y>, the common point being y. in
Yj. The conditions ensure that n^(Y) is the free product of the n^(Y^). We then
have the homomorphism $9^. from iz^Y) to G. The result now follows from the
theorem.//

Corollary 3 Let G be a group. Let Y. be spaces with base-points

y^, and let 9z-:7Cj(l^.,^.)-> G be homomorphisms. Then there is a space X (which
is 7*| // the Y- are T^) with base-point JCQ containing disjoint subspaces
(homeomorphic to) Y- such that it AX, XQ) is G and such that 9. is induced by

inclusion followed by f* for some suitably chosen path f- from y^ to JCQ.

Proof Take the spaces Z- which are the joins of Y. and the interval
/, and apply Corollary 2 to these.//

Let AQ and A* be path-connected subspaces of the space X such

that there is a homeomorphism 9'.AQ -» A^. Let Y be obtained from X by adding
a handle from AQ to A^; that is, Ym X UAAQ X I) where f:A^ x {0,1} -> AT is given
by (a,0)/« a and ( a , l ) / - <2<p. Let Z be obtained from * by identifying a with 09
for all azA^. There is an obvious map from Y to Z which is known to be a
homotopy equivalence subject to some fairly mild restrictions on the position of
AQ and A^ in X.

I / \ \

O^o

Figure 9, showing two cases of attaching a handle

We can apply van Kampen's Theorem (in the groupoid form, and
using the corollary with a pair of vertices (<ZQ,1/3) and ( £ Q , 2 / 3 ) ) to the subspaces
AQX(Q,\) and Y- AQ X {1/2}. This situation was considered in Example 2 of
pushouts of groupoids. Since A!" is a deformation retract of Y- A^y-il/2) we
obtain the following results.
If X is path-connected then nJY) is the pseudo-HNN extension of
7Tj(AT) with associated subgroups the images of TTJ(^Q) and nÂ^). Note here that
the images are only determined up to conjugacy (since we have to use the
isomorphism induced by a path to refer all the groups to the same base-point,
and this isomorphism is only known up to conjugacy); but we know that this
makes no difference to the pseudo-HNN extension.
Suppose that AQ and A* are in different path-components XQ and
X^ of X. In this case 7Tj(K, a^) is the pushout of the homomorphisms
•KÂ^ + TZ^XQ) and TIÂQ) -> nÂ^ •* n^X^.
Theorems 12 and 13 ensure that any pseudo-HNN extension or
pushout may be obtained in this way from suitable spaces. Thus the notion of
adding a handle is the topological analogue of these group-theoretic concepts.

Let W be obtained from X by adjoining ^IQX[0,1/2J and A^x[1/2,11

Then Y is obtained from W by identifying A^ x {1/2} and A^ x {1/2}. We could
also refer to W as obtained from Y by cutting along A^ x {1/2}.
Conversely, suppose that we are given a space Y with a subspace A
such that there is an open set U with (U,A) homeomorphic to (A x (-l,l),A). With
slight restrictions on A, we can cut Y along ^4 to obtain a space W We can also
regard Y as being obtained from Y- A x (-1/2,1/2) by adding a handle from
>4x{-l/2} to A*{\/2). Such a space K will then have its fundamental group either
a pseudo-HNN extension or a pushout (but the pushout could be the trivial kind
in which all the homomorphisms are the identity).

Exercise 6 Prove the remark after Proposition 9.

Exercise 7 Find %^ of the Klein bottle.

4.3 COVERING SPACES

Definition Let p-.tf-tX be a map of non-empty path-connected
spaces. Then p is called a covering map and X is called a covering space of X
(both abbreviated to covering) if every x e A" has an open neighbourhood U such
that the components of Up are 'open and are mapped homeomorphically onto U
by p. Such a neighbourhood £/ is called an elementary neighbourhood of x.

If U is an elementary neighbourhood then Up p must be U unless

£/p - 0 (the definition permits this, as then Up" has no components). Hence
Uslfp if t / n j f p / 0 . It follows that J?p is both open and closed, and so
X - l?p, since A" is connected. In particular, we cannot have Up - 0.
Note that each point of xp has an open neighbourhood which
maps homeomorphically onto U, and so does not contain any other point of
xp' . Hence xp is discrete. In particular (this is a special case of Lemma 14
below), if / is a map from a connected space to Jt such that fp is constant then
/ itself is constant.
We shall give later a simple topological condition which ensures
that every component of Up is open.

Examples Regard S as the set of complex numbers iz\ \z\ - 1). Then
R is a covering of S , with the covering map sending t to e . We only need
two elementary neighbourhoods, namely S - {1} and S -{-!}. This example,

which is one of the simplest and most important, will be referred to as

unwrapping the circle.
The map of S to itself which sends z to zn is also a covering
map, with the same elementary neighbourhoods.
2
Let X be the join of two circles. The subset W of R consisting of
({xj); either JC or y is an integer} is a covering of X, the horizontal lines
mapping to one circle and the vertical lines to the other. (See Figure 10)

W
Figure 10

Let J? be obtained from X by unwrapping one circle; that is, it

consists of R with a circle attached at each integer point (see Figure 11). We
may obtain a covering Y of J? by unwrapping one circle, and another covering
Z of J? by unwrapping every circle. Both Y and Z are also coverings of X.
They are illustrated in Figure 11 (the picture of Z is approximate, as it is
difficult to draw in such a way that intervals remain disjoint).

Figure 11

Let p:R-> S be the unwrapping. Let / and g be maps from a

connected space Y to R such that fpmgp. Then / - g is a continuous map which
is always an integer. Hence / - g is constant, and so / - g if there is some point
^Q with y^f- y^g. This is a general property of coverings, as is shown in

Lemma 14. The proof for the special case is so simple that I feel it was
worthwhile giving it separately.
Lemma 14 Let p\£ -*X be a covering. Let f and g be maps from
a connected space Y to it such that fpm gp and y^f-y^g for some y^. Then
f-g.
Proof Since Y is connected, it is enough to show that {y; yfmyg) is
both open and closed.
Take y with yf-yg. Let U be an elementary neighbourhood of yfp
(mygp)- Let lT be an open neighbourhood of yf which p maps homeomorphically
to U. Then there is an open neighbourhood V of y such that Vfc C and Vgc [7.
Since fpm gp, it follows that / - g on K, and so iz; zfm zg) is open.
Now take y with yf^yg. Let U and tf be as before, and let lT* be
an open neighbourhood of yg which maps homeomorpically to U. Then there will
be an open neighbourhood V of y such that Vfc (I and Vgc £?1 . Now [7 and I?.
-1
are components of Up which must be distinct, since they contain distinct
points with the same image. They are therefore disjoint, from which we see that
Vc{z; zf*zg), and so this set is also open.//
The next result is one of the key properties of coverings.
Proposition 15 Let p:Jt-^X be a covering. Let /:CO,r] •* X be a
path, and let a be a point of ]t such that ap - Of. Then there is a unqiue path
/:[0,f] -» ]? such that 0 / - a and ?p- f.

Remark The path / will be called a lift of /.

Proof Uniqueness is immediate from Lemma 14.
Let 8 be the Lebesgue number of iUf" ; U is an elementary
neighbourhood), which is an open cover of CO,/0 by the definition of covering
spaces. Take n such that r<nh. Then / maps each interval Ur/n, (i+l)r/ni into an
elementary neighbourhood. Suppose that we have lifted / on the interval [Q,kr/ri]
to a path Tj,,. Let U be an elementary neighbourhood such that
Lkr/n,(k+\)r/n]fcUy and let b be (kr/n)?^. Then b has an open neighbourhood
which p maps homeomorphically onto U. It follows that there is a map
g-lkr/n, {k+\)r/n\ -» X such that (kr/n)g- b and gpmf on this interval. By the
Glueing Lemma, the function which is J^ on lO,kr/n\ and g on ikr/n, (k+l)r/n\ is
continuous. Inductively, we obtain the required lift of /.//

Plainly the standardisation of a lift of / is a lift of the

standardisation of /.

Proposition 16 Let p\Jt ^ X be a covering. Let f and g be

homotopic paths in X. Let J and g be lifts of f and g beginning at the same
point. Then / and g are homotopic.
Corollary With the above notation, f and g end at the same point.

Proof J and g will be homotopic if their standardisations are

homotopic. Since / and g are homotopic, the standardisations of / and g will
also be homotopic. Hence it is enough to prove the result for standard maps, and
we may take a standard homotopy F from / to g.
We now take 8 to be the Lebesgue number of the open cover of
/ x / consisting of {UF ; U is an elementary neighbourhood]. Take an integer n
with y 2 < nh, so that any square with side \/n is mapped by F into an
2
elementary neighbourhood. Divide / x / into n squares of side \ln. Let D-. be
the square whose bottom left corner is (i/n>j/n), and let C- be the union of the
squares which are either lower than D- or on the same level as it but to its left.

D
1,1

The shaded area is C<

1,1"
Figure 12

Let / and g start at a and end at b, and let / and g start at a.

Suppose that we have found a map F*. .:C..-» A*' such that F*--pm F on its
domain, and with {0,0)F-- a. We will obtain a map H\D..-*7? such that
F on its domain and such that H agrees with /*• • where they are both
defined.
Because F maps D- into an elementary neighbourhood, it is
immediate, as in the previous proposition, that there is a map H\D-->X with
Hpm F on D- and with (i/n9j/n)Hm(i/n,j/n)F^... Suppose that / and j are greater
than 0 (the other cases are simpler). Then C- nD •• is the union of the horizontal
line-segment Uin,{i+\)/n\x{jfn) and the vertical line-segment {i/n} x \_j/n, {j+\)/n\.
This set is connected, and Hpm F-p on it, and there is one point of it where H
and fi'.j are the same. Hence H agrees with /^. where they are both defined, by

Lemma 14. Then the Glueing Lemma gives a continuous function defined on
C..uZ)... Inductively we will obtain a map F:I*I'*]F such that (0,0)f - a and
Pp- F. Now (0,u)Pp-a for all u and (0,0)/* - <J\ Hence (0,1/)/* must be <? for
all u. Similarly, (l,u)F* must also be constant. Since (r,0)//>- (r,0)F- f/- r/p and
( 0 , 0 ) / * - 0 / , we find that (f,O)/*-f/, by Lemma 14. For the same reasons, since
we know that (0,1)/*- a, we find that U l ) / * - tg for all t. Hence /* is a
homotopy from / to g.
The corollary is now immediate.//

In our definition of a covering space we required the components of

Up to be open. The lemma below gives a condition under which the
components of open sets are open.
Lemma 17 The following properties of a space X are equivalent.
(/) the path-components of any open set are open;
(ii) for any open U and x <L U there is a path-connected open V with xiVcU;
(Hi) for any open U and x E U there is an open V with x € V c U such that any
two points of V can be joined by a path in U.

Proof If (i) holds so does (ii), since we need only take V to be the
path-component of U containing x. If (ii) holds then (iii) obviously holds.
Suppose that (iii) holds, and let C be a path-component of the open
set U. Take x e C. Then the open set V of (iii) lies completely in C. Hence C is
open.//
A space with the properties of this lemma is called locally path-
connected. In particular, the path-components of a locally path-connected space
are open. Also, if U is an open set in a locally path-connected space then the
path-components of U are disjoint open connected sets; they must therefore be
the components of U. As a special case of this, a connected and locally
path-connected space must be path-connected; thus our example of a space which
is connected but not path connected provides an example of a space which is
not locally path-connected.
Evidently, if A^ is a covering of X then ^ is locally
path-connected iff X is. Also, in this case, if U is an elementary neighbourhood
and V is any open set then the components of UnV arc also elementary
neighbourhoods. Further, for any open neighbourhood W of a point v' € it\ that
component of V? n Up (where U is an elementary neighbourhood of vp) which
contains v will be mapped homeomorphically by p onto an elementary
neighbourhood in X.

Exercise 8 Show that the identification map from S to R/^ is a

covering.
Exercise 9 Give an example of a space which is locally
path-connected but not path-connected.
Exercise 10 Give an example of a space which is path-connected but
not locally path-connected. (You could remove a suitable infinite set of
2
line-segments from IR .)

4.4 THE CIRCLE AND THE COMPLEX PLANE

In this section we shall investigate the properties of n^(S ) further,
giving a new proof that it is infinite cyclic. We shall use these properties to
obtain results about the complex numbers; in particular, we will prove the
Fundamental Theorem of Algebra. Because we are working with the complex
numbers C, we will regard S as iz; lzl-1) and Er as (zjlzl^l).
Let pM^S be given by tpm e . In the previous section we saw
that to any path / in 5 and any a with a p - 0/ there is a unique path 9 in IR
such that / - <pp and O9 - a. We refer to 9 as a lift of /. Note that two different
choices of a differ by an integer, and hence two different lifts of / have their
difference a constant integer function. We also saw that if we lift homotopic
paths in S to paths in IR with the same start then these lifts are homotopic; in
particular, they have the same end-point.
The proofs of these results were very similar to the proof of van
Kampen's Theorem. However, van Kampen's Theorem applied to S would enable
us to show that every path was homotopic to a path which has a lift, but would
not be enough to tell us that every path lifts. Several of the results can be
proved using the discussion of n^S ) after van Kampen's Theorem if the reader
prefers; the proofs would be similar to those given here, but would need a little
more care.
Let f:LO,r] -» 5*1 be a loop. Since / / - 0/, np and O9 must differ by an
integer. This integer does not depened on the choice of 9. We call it the degree
of/.

Theorem 18 For any a*S \ the degree is an isomorphism from

7iAS ,a) to Z.

Proof Let / and g be homotopic loops. Let 9 and ij; be lifts of /

and g, chosen so that O9 - 0iJ>. Then we know that 9 and ty have the same

endpoint. Hence / and g have the same degree, and so the degree can be
regarded as a function from n^S ) to Z.
Now let /CO,/-] -» 5 be a loop with degree n, and let g-10,s1 •* Sl be
a loop with degree m, both loops being based at a. Let 9 be a lift of / , with
Ocp-oc, and hence with rep - a + n. Since 0g-a-0f, we may take a lift ip of g with
Oijj-a + /z. As g has degree m, we find that sty- <x + n + m. Now <p.{J> is continuous,
by the Glueing Lemma, and it plainly lifts fg. Hence f.g has degree n +m,
showing that the degree defines a homomorphism.
This homomorphism is onto, since if 9 is the standard path sending
t to nt then epp has degree n. Also, if / has degree 0 then cp is a loop in R, and
so / represents an element of TC.(IR.(X)/?*. As TT,(IR) is trivial, this tells us that a
1
loop of degree 0 represents the identity element of n^(S ).//

Lemma 19 Let f be a path in S from a to bt and let g be a path

in S from b to a. Then the loops f.g and g.f have the same degree.

Proof Let 9 be a lift of / starting at a and ending at (3, and let ^

be a lift of g starting at £. Then 4> ends at <x + /z for some n. As cp.ip is a lift of
f.g, the loop f.g has degree n. Let cp' be given by ftp' - t(p + n. Then g.f lifts to
<Jj.<p\ which shows that g.f also has degree n.lI

Corollary 1 Let f be a loop in S based at a, and let g be a path

in S from a to b. Then g^.f.g has the same degree as f.

Proof By the lemma, i~./-£ has the same degree as fg.g^, and this
is homotopic to / , hence it has the same degree as / . An easy direct proof can
also be given.//

Corollary 2 Let f and g be freely homotopic loops in 5*. Then f

and g have the same degree.

Proof By Lemma 2 and its corollary there is a path h such that / is

homotopic to h .g.h. The result is now immediate.//

Lemma 20 Let f and g be maps from l0tr\ to S . Define h by

th~{tf){tg). Then degreeh - degreef + degreeg.

Proof Let cp and ip be lifts of / and g respectively. Then 9 + 4> is a

lift of h, and the result follows.//

We can extend the above results to C - {0}. For C - (0) is

homeomorphic to 5* xR + , the homeomorphism sending z to (z/l?l, Izl). Since R+
is contractible, S is a deformation retract of C-{0}, the retraction sending z to
z/\z\- Consequently we can define the degree of a loop in C - {0} as the degree
of the corresponding loop in S , and obtain similar results immediately.
Thus 7Tj(C-{0}) is infinite cyclic, with an isomorphism given by the
degree. Freely homotopic loops have the same degree, and two loops with the
same basepoint and the same degree are homotopic. Also, if / and g are two
loops in C-{0} with the same domain, and h is given by thm(tf)(tg), then
degree h - degree/ + degreeg.
The degree of a loop / in C - {0} is also called the winding number
of f about 0. More generally, if / is a loop in C not going through the point a,
then the winding number of / about a is the degree of the loop g given by
tg-tf-a.
Any map / from S to any space X gives rise to the standard loop
which is the restriction of pf to /, and any standard loop is of this form. It
follows that we can extend our definitions to give the degree of a map / from
S to C-{0}, and similar results hold. The basepoint of this loop may be taken
as 1/ for definiteness, but Corollary 1 to Lemma 19 tells us that the choice of
basepoint does not affect the degree.
As an example, the map of S to itself which sends z to zn is
easily seen to have degree n.

Theorem 21 (Rouche's Theorem) Let f and g-£0,r) -» C be loops such

that \tg\<\tf\ for all t. Then f and f+g have the same degree.

Remark The conditions ensure that / and f+g both map into C - (0).
Proof f and / + g are freely homotopic in C, the homotopy F being
given by {t,u)F- tf+(tg)u. It is easy to check that F is never 0, so F is a
homotopy in C-{0}, giving the result.//

Theorem 22 (Fundamental Theorem of Algebra) Every non-constant

polynomial with complex coefficients has a complex root.

Proof Let / be the polynomial given by zf'^a^z, with an*0.

For any real number r>0, let fr'-S -> C be given by zfrm(rz)f Then, for any ry
f is freely homotopic in C to / Q , the homotopy sending {z,u) to (urz)f If /
were never zero then this would be a free homotopy in C - {0}, and so ff and / Q
would have the same degree.
Now / Q has degree 0, since it is constant. Suppose that, for some R,
II?" 1 a,wk\ <\a v/*l for all w with \w\-R. Then, by Rouche^s Theorem, fD will
have the same degree as the function on S sending z to a R zn, which is n.
The theorem would then be proved.
Now \1LQ~1 akwk\±Y.Q~l la^llwl*, which is at most (Z^"1 la ;
provided Ivvl^l. So we have the property needed provided that
R>

The degree of the loop / in C - {0} can be described as 1/2TT times

the change in the argument of z as z moves round the image of / . If / is
sufficiently well-behaved, this can also be described as Mini times the integral
of \/z round / . The degree (or the winding number, which is essentially the
same) plays an important part in complex variable theory.

Lemma 23 Let f be a standard loop in Sl such that (r+1/2)/- -tf

for O^tzl/2. Then f has odd degree.

Remark In terms of maps from S to itself, the lemma tells us that

a map g:5-» S has odd degree if (-z)gm~(zg) for all zeS .

Proof Let <p be a lift of / on [0,1/2]. Since (1/2)/- -0/, there is an

odd integer m such that (l/2)cp - 0<p + mil. Since (f+ 1/2)/- -tf for 0 ^ r ^ l / 2 , / l i f t s
on Cl/2,1] to the function ty such that (r+1/2)4)-rep + w/2. Then cp.ip is a lift of /
on [0,11 As lip- (1/2)9 + mil - 0<p + m, the loop / h a s odd degree m.l I
Theorem 24 (Borsuk-Ulam Theorem, £r Ham Sandwich Theorem)
Let f.S -»R be a map such that (-p)fm-{pf) for all p. Then there is some q
2 2

such that qf~ (0,0).

Proof Suppose not. Then the retraction of IR - {(0,0)} onto S gives

2 1
a map g:S -> S such that (-p)gm -(pg) for all p. We can then define a map
by (x,y)h-(x,y,J(\- x2-y2))g, and we find that (-p)h--(ph) for all
p e S . By Lemma 23 and the remark following it, the map h restricted to S v»
then have odd degree. But this loop is freely homotopic to the constant loop,

since h is defined on Er, and so its degree is 0. This contradiction shows that /
must take on the value (0,0) somewhere.//

The corresponding result for maps from Sn to R'1 also holds. For
n - 1 it is just a version of the Intermediate Value Theorem; for general n a
significantly more difficult proof (for instance, involving cohomology rings) is
needed.
I now explain why the name "Ham Sandwich Theorem" is used.
First consider a nice set A in the plane. (The precise definition of a "nice" set
would require some knowledge of measure theory; what we need is that some
intuitive properties of area and volume are valid for the sets we look at.)
Perpendicular to any direction there is a line which divides A into two parts of
equal area. To see this, consider the difference between the areas of the portion
of A above the line x cosa +>>sina- c and the portion below it. This is a
continuous function of c, which is negative for large positive c and positive for
large negative c. Thus it must be zero somewhere. For nice A there will be only
one such line, which we call /(a).
This gives c as a function of a. This function can be shown to be
continuous (since, if we take a disc containing A, the lines corresponding to a
and p must meet inside this disc).
Now take a second nice set B in the plane. Consider the area of
the portion of B above /(a) minus the area of the portion of B below it. This
will be a continuous function of a. If we replace a by OC + TT the value of this
function will be multiplied by -1, since /(OC + TT) is the same line as /(a), but
"above" and "below" get interchanged. By the Intermediate Value Theorem, this
function is zero somewhere. This means that there is a line which simultaneously
bisects A and B.
We now extend these results to sets in (R . Let A be a nice set in
3 2
IR . We can regard a direction u as a point of S . As before, perpendicular to
any direction there is a unique plane, which we call p{u), dividing A into two
pieces of equal volume, and this plane varies continuously with the direction.
Let B and C be two further nice subsets of IR . We get a
continuous function f\S 2 •* IR2 whose components are the volume of B and C
respectively above p(u) minus the volume of the same set below it. Since p{-u)
is the same plane as p(u), but with above and below interchanged, (-u)fm-{uf).
Hence the theorem tells us that there is some direction perpendicular to which
there is a plane dividing A,B> and C simultaneously into two equal halves. Now
take the sets to be the two pieces of bread and the filling in a ham sandwich!

Exercise 11 Let / and g be loops in S of the same degree, but with

different basepoints. Show that / is freely homotopic to g.
Exercise 12 Let / and g be maps of Sl to itself. Show that the
degree of their composition is the product of their degrees. (You may want to
take maps freely homotopic to / and g for which the result is easy.)
Exercise 13 Show that a homeomorphism of S has degree ±1.

Exercise 14 Let U be an open neighbourhood of ZQ in C, and let /

be a map from U- [ZQ) to C-{0}. Define ff on S by zfrm (ZQ + n)f for positive
real r. Show that the degree of / is constant for r small enough. We will call
this the order of f at ZQ-
Exercise 15 Let g be a map from an open neighbourhood of ZQ to C
such that Zftg * 0. Let / be given by zfm(z- ZQ) (zg) for some ktZ. Show that
the order of / at ZQ is k.
Exercise 16 Let ^ , . . . ,zn be points of Er - S . Let r be a small
enough positive real number. Let c^ be the loop given by zc^'iz^rz) for
z e S , and let CQ be the loop given by zc^ - z. Show that c^ is homotopic in
Er - {z-y,... ,zn) to the product n^.c^.g^ for some paths g^. (You may want to
break Er - U j , . . . ,zj into pieces and use induction with the help of
homeomorphisms. Or you could break D minus some small discs into the union
of star-shaped sets. The proof may be rather messy.)
Exercise 17 Let ^ , . . . ,zn be points of Er - s \ and let / be a map
from Er - iz< , . . . ,z ) to C- (0). Suppose that / has an order at each ZK • Show
that the degree of the restriction of / to S is the sum of the orders of / at the
points £ j , . . . ,zn- (Use the previous exercise.)

Exercise 18 Let A be a set which has two binary operations x and

*, each of which has a two-sided identity. Suppose that
{w* x)*(y* z)m (w*y)*{x* z) for all w, JC, y, and z. Show that x and * coincide,
and that the operation is commutative and associative. (Begin by proving that
the identities for the two operations must be the same.)
Exercise 19 Use the preceding exercise to show, for standard loops /
and g in S and their product fg defined by r(fg)m {rf)(rg) for r e / , that
deg/+ degg.

4.5 JOINS AND WEAK JOINS

In this section we consider some matters of topological interest
which are not relevant to the group-theoretical applications. We begin by
looking at a space closely related to the join of spaces and discuss aspects of
its fundamental group, which is very complicated. From this we are able to
obtain two contractible spaces whose join is not even simply connected. We then
obtain some new conditions under which the fundamental group of a join is the
free product of the fundamental groups of the spaces.
Let X be the union of the spaces X-, with X-n X •-{p} for i*j9 and
let {/>} be closed in each X-. The join topology on X can be given the
following description. If x*p there is exactly one / with x*X., and the open
neighbourhoods of JC in X. provide a basis for the neighbourhoods of x in X. A
basis for the neighbourhoods of p in X is given by the sets UU- where the U-
are open neighbourhoods of p in X-.
When the index set is infinite, there is another interesting topology
on X9 called the weak join of the X-. We take the same neighbourhood basis as
before for points other than p. But the basis for the neighbourhoods of p is now
to consist of those UU- with U- an open neighbourhood of p in X- and, in
addition, UmX- for all but finitely many /.
The subspace of IR2 which is the union of the circles S , where
S m i(xj)\ (x-\/n)
2+y 2ml/n 2}, is often called the Hawaian earring (look at the
picture!). It is easily seen to be the weak join of countably many circles.

Figure 13, showing (part of) the Hawaian earring

Let X be the weak join of countably many spaces X.. Then nJX)
is complicated. It is easy to see that it contains standard loops / such that
(1-1/n)f~ p for all n, with / mapping 11-1/n, 1-1/(^+1)] into Xn. We could regard
such an element as amounting to an infinite product otjO^... with a snAX , P)
for all n. But n^(X) would also contain products of the form . . . a^a* , which are
inverses of the previous infinite products, as well as more complicated elements
obtained by multiplication from these.

We will now proceed to give some properties of K^X) when X is

the weak join of reasonably nice spaces. We shall assume that we have only
countably many spaces Xn, each of which has non-trivial fundamental group. We
assume that ip) is closed in each Xn, and that the spaces are such that the
fundamental group of the join of any finite number of them is the free product
of the fundamental groups of the corresponding spaces.
Let X+ be the weak join of those Xn with n even, and let X_ be
the weak join of those X with n odd. Then X is the join of X+ and X . Let
Y+ be the cone on X+ and let Y be the cone on X . Let Y be the join at p of
K+ and Y_. Then K+ and Y_ are contractible. Applying Proposition 9 twice, we
find that nAY) is obtained from n^{X) by cancelling n^{X+) and nJX). We
shall show that this group is uncountable.
Let G^-TCjtAf^), which we are assuming non-trivial. Let H , Hn + ,
and H _ be the free products, respectively, of all G> with i*n, of those G. with
i*n and / even, and of those Gy with i*-n and / odd.
There is a homomorphism from Hn^ onto Hn which is trivial on
G ++ and which maps H identically. Let H be the group consisting of all
infinite sequences ( A j , / ^ , . . . ) such that, for all n9 hn*Hn and hn is the image of
hn+i in the homomorphism of the previous sentence, with the multiplication
being componentwise. This group is known as the inverse limit of the H . The
subgroup H+ consists of those sequences for which every hn is in # / f + , and
similarly for //_. The subset (not subgroup) L consists of those sequences for
which hnm g^2 • • • Sn with g. € G. for all /.
Now X has an obvious retraction to X^ v . . . v X for all n. This
provides, by our assumption on the spaces, a homomorphism from n^(X) to Hn
for all n. The sequence of images of an element of it^X) under these
homomorphisms is evidently an element of //, so we obtain a homomorphism
from Tij(A') to H. This homomorphism plainly sends n^{X+) into H+ and n^XJ
into //_. An element of L gives rise to loops a^ e Xn for all /z, and the infinite
product a ^ 2 • • • discussed above is an element of n^(X) which maps to the given
element of L.
Hence our description of K^Y) shows that it is uncountable if we
can find uncountably many elements of L, no two of which are in the same
coset of the subgroup H+ //_[//,//]. The proof of this will take some time.
Fix n for the moment. Consider an element u of H . Then u is a
n
product JCj... x^ with each x^ in some G., x.*\, and x- and x.^ in different
factors. If Xj*Gp and x^^G we say that (/?,<?) occurs at i. Let UOL be the

number of occurrences of (/>,<?)> and let B be (Xpa'0Lap- Finally, let u\ be

Let hzH have component hn^Hn. We shall show that h^K is a

bounded function on n if /ze // + / / _ [ / / , / / ] . We shall find an uncountable subset A/
of Z, such that hnX is not a bounded function of rt if \*h*M~ M. This will
prove the result. We begin by obtaining some properties of a and £.
Plainly ul*pq~™qp* and so u
'\qm '%q-
Given u^,... ,u in Hn let y (which is a function of
Uy,.... ,u ,but we shall omit these arguments) be given by

^•(»l'"»5»wEi "iW Thetl We WiU ShOW that

WW* 6 **" 1 *'
This follows by an easy induction from the case 5 - 2 , which we
now prove. So let u - x^... JCJ (note the order) and let v - y^... ym. There is
some j so that Xj... Xj - 0 ^ . . ..Vy)"1 but Xy+1 ^ y + f 1 - Let Xj+l € C7fl, jcy.e Cr^ (and
so j> • e ^^j), and j ; . + ^ e 6?c. Then a*b*c; we may or may not have a - c. Let a' be
x
k---xj*V »~'*j---*\> V - J ' p . - J ' y . and V - > y . + 1 . . . ^ m .
It is easy to check: that uoiab - "'a^^ + " " a a ^ + 1» while
wa - w'a + wMa for all other pairs (/>,<?)• Similarly for v, the special pair
being (b,c). Also, if a*c we have (uv)<*ac~ " ' a a c + v'foc(3c+ ^' w n i l e for all pairs
(A<?), except (a,c) when a / c , we have ( " v ) ^ " " ' ^ / ^ ^ - As w" - v'~ , we
have " X j ' ^ j n f ° r a 1 1 (P»^)- Combining these gives the result.
Now suppose that h~ xyl!^ [^.,vv^], where x € / / + , ytH , and ^. and
vv. are in //. Then, with A , JC , etc, being the components of h, x. etc, in H ,
we have

Plainly a zr-i,zr
0 , * « and a zr,zr-i
o , O i . , are both zero on // n+
. and on //n- , so the same
e
holds for p2/--l 2r' ^ ^ o r m u ^ a ^or P o n an
inverse, and the formula relating g
on a product to 3 on the individual terms and to y» now tells us that
A
/ ! P 2 / - 1 2 / - ' Y 2 / - 1 2 / " W h e r e I A ^ ' Y P ^ ' ^ 6 ( 4 5 + 1 ) ' H e n c e A ^ ^ 6 ( 4 5 + 1 ) for all n.
Choose an to be a non-identity element of Gn for all «. An element
of L is given by a sequence {^} where gn*Gn for all /2. Let S be an infinite set
of positive integers. Let m^ be the element of L whose entry in G^ is ^ if
«€5* and is 1 if # • 5 ' while its entry in ^ / z - l is ^2«-l if n*S and is 1 if n*S.
Let hmm.g nij. , where 7" is another infinite set. If k is the smallest integer
which is in exactly one of 5" and T, it is easy to check that in computing hn
cancellation cannot reach beyond ^2A-2* ^ e f i n d e a s i l v t n a t ^rPlr-\ 2r*®
whenever 2r^/z and r is in exactly one of S and T. Hence A X is an unbounded
n

function of n if the set of elements in exactly one of S and T is infinite.

So we want to find uncountably many subsets of the positive
integers, any two of which are either equal or have finite intersection. The
easiest way to show this is to prove the same result for the rational numbers,
and then use the bijection between the positive integers and the rationals. To
each real number r we take an (increasing) sequence of rationals which tends to
/•, and let S be the set of members of this sequence. Then the sets S for all r
have the required property.
It is possible to show that the homomorphism from izAX) to H is
one-one. Also the image of this homomorphism consists of those h for which, for
every r, the number of occurrences of an element of G^ in h is a bounded
function of n. We shall not prove these results, which are difficult.

We now investigate circumstances under which the fundamental

group of a join of spaces is the free product of the fundamental groups of the
spaces
The space X is called almost simply connected at p if X - {p) is the
union of disjoint open sets U. (for / in some index set) such that each U-uip) is
simply connected. Such a space is plainly pathwise connected. Also {p} must be
closed. We call X locally almost simply connected at p if {p) is closed and each
open neighbourhood of p contains an almost simply connected at p open
neighbourhood of p. Notice that the join at p of two spaces both (locally)
almost simply connected at p is itself (locally) almost simply connected at p. As
remarked above, in contrast, the join of two simply connected spaces need not
be simply connected.
We say that X is first countable at p if there are open
neighbourhoods U of p (for n a positive integer) such that every open
neighbourhood of p contains Un for some n. The join of two spaces both first
countable at p is itself first countable at p.
Let f£0,r1 -» X be a loop based at p, where {/?} is closed. Then
{X - (p))f is an open subset of (0,/-). It will then be the union of countably
many disjoint open intervals. This holds because the components of an open set
must be disjoint connected open sets; that is, they must be disjoint open
intervals. There can only be countably many of them, since each interval
contains a rational number, and no rational number is in two of the intervals.
Note that the endpoints of these intervals must be in pf . If there are only'
finitely many such intervals we say that / is geometrically finite.

Lemma 25 Let X be a space which is first countable at p and

locally almost simply connected at p. Suppose that {p} is the intersection of
open sets. Then every loop based at p is homotopic to a geometrically finite
loop. Also X will be simply connected if it is almost simply connected.

Proof We may as well assume that / is a standard loop. Let

(X - [p})f be the union of the disjoint open intervals /..
We first show that for any open neighbourhood U of p the set Uf
contains all but finitely many of the J-. Suppose not. Then we can find, for all
ny distinct an<cn<*>n w i t n Cr/*U a n d ân^n^ b e * n g o n e °^ tîe intervals J-7 so
that a^m p. Taking subsequences if necessary, we may assume that the sequence
a converges; let a be its limit. Since the intervals do not overlap, /Jib - a )^1,
so that b -a -» 0. Hence the sequence c also converges to a. But this
n n ^ n °
m
contradicts continuity of / at a, since cif p and {p} is closed, while cf is in
the closed set X - U which does not contain p.
The hypotheses ensure that there are open neighbourhoods U of p,
such that each U is almost simply connected, every neighbourhood of p
contains U for some n, and such that Un+Ûn and f)U'n~ {p}. Let £/Q be X.
Take an interval J-. Then J-f is contained in £/Q but is not
contained in C\Un . There must then be some n- i
such that i/ • / is contained in Un
iff n^n.. In particular, the result we have just shown tells us that n. - 0 for only
finitely many /. If n-^0 then Un~{p} is t n e union of disjoint open sets each of
which has its union with ip) simply connected. Now J.f9 being connected, lies in
one of these sets, W say, and / / s £/'u (/?}, so that the restriction of / to T. is
homotopic in U , relative to the endpoints, to the constant map. If X itself is
almost simply connected this holds for n.-0 also. Let F. be such a homotopy.
Now define F o n / x / as follows. If tf-p then (t,u)F~p for all u.
If r c T with n-^0 then (t,u)F* (ttu)F-; when X is almost simply connected this
definition is used even when n. - 0. When X is not almost simply connected and
tiJ. with nm0 then (t,u)Fm tf for all u. This does define a function, since when
t is an endpoint of /- we have tfmp and (t,u)F- p. If we can show that F is
continuous it will provide a homotopy between / and the loop sending t to
{t,\)F. By definition, (t,l)Fm p unless tzJ- with n.-0 (and everywhere when X is
almost simply connected). Hence this loop is geometrically finite, as required.
Now F is plainly continuous at any (t,u) with te J. for some /,
since F- and / a r e continuous. So let tfmp, so that (t,u)F-p. Let V be any
neighbourhood of p. Take n with U cV. Then for all but finitely many / we
have J-f£Un, and hence n.^n. In particular, there is some 8 such that (f-8,f+$)

meets only those / . with n- ^ n and those (at most two) .A with t as an
end-point. Then, for t'i(t- h,t+h) and any u\ either ffmp and so (t\u')Fm p, or
t'tJj with n-^n, or (for at most two values of /) both t and t' are in J. . In the
second case, (t\u')Fm (t\u')F-t Un £UncV. In the last case, the continuity of F-
on T't shows that (t\ u')F\V if (t\u%) is close enough to (t,u). This shows F is
continuous.//

Proposition 26 Lef ^ be the join at p of the spaces X.. Let each

X- be locally almost simply connected at p and first countable at p, and let ip)
be the intersection of open sets in each X-. If the index set is finite, or if
each X. is a T*-space then nAX,p) is the free product of the itJX.,p).

Remark If a space is 7^ then ip) is the intersection of all the open

sets containing p.
Proof The usual direct limit argument (Proposition 10)
gives the result for infinitely many spaces when we have proved it for finitely
many. Since the join of two spaces with the stated properties also has these
properties, the result holds by induction for any finite number of spaces once we
have proved it for two spaces.
So let X be X^ v A^, with the conditions holding for X^ and A^,
and so holding for X. We already know that the inclusion induces
monomorphisms of n^(X^9p) and rc^A^, p) into n^X, p). By Lemma 25, any
element of nJX, p) can be represented by a geometrically finite loop. Such a
loop is the product of loops in X^ and loops in X^ • Hence K^X) is generated
by the images of n^(X^) and ^ ( A ^ ) .
In particular, if A^ is simply connected (which, by Lemma 25,
holds if it is almost simply connected) then the inclusion induces an
isomorphism from 7TJ(A"J) to n^X).
In the general case, there are almost simply connected
neighbourhoods U^ and U^ of p in A^ and A^ respectively. So the previous
paragraph tells us that the inclusions induce isomorphisms of 7tj(A'j) in
K^X^vUy) and of %^X^) in -K^JJ^M X^ while U^U^ is simply connected. The
result now follows by van Kampen's Theorem.//

5 COMPLEXES

5.1 GRAPHS
A graph V consists of two disjoint sets, the set V(T) of vertices of T
and the set E(T) of edges of T, together with two functions o:E(T)-> V(T) and
~:£(O-» E(T) such that, for all ee£(O, e*e~ and e-<T. We define another
function T:E{F)-> V(T) by ex- e"o. We call eo the starf (or beginning) of e and ex
the end (or finish) of e, referring to them both as end-points of e. We call e~
the inverse of e. We call e a circle if ex - eo (such an edge is often called a
"loop", but this name could lead to confusion later). We simply refer to E and
K, rather than E(T) and K(T), where no confusion is likely. It is sometimes
convenient to allow a graph to be empty, and sometimes inconvenient; I will
leave it to the reader to decide which statements require the graphs considered to
be non-empty.
An alternative definition is often used, in which edges do not come
in mutually inverse pairs. This approach is sometimes easier, but at some stage
we would need to introduce some version of the inverse of an edge. It is really
a matter of personal preference which definition to use.
Our definition does differ from other definitions of a graph in two
further ways. We permit edges to have both end-points the same. We also permit
more than one edge with a given starting point and finishing point. We will find
it very useful to allow this generality.
A path of length n > 0 in T is a finite sequence e*,... ,e of edges
such that e-x'e-^o for all i < n. This path starts at e^a and ends at e^x, and the
vertices of the path are e-a for all i±n and also ej. A path of length 0 is just
a vertex v, and v is the start and end of this path. A loop is a path which ends
at the point at which it starts. A path is simple if its vertices are all distinct. A
simple loop or circuit is a loop of positive length for which the n vertices ev
are all distinct, and which is not of the form e,e~. A single edge is a circuit iff
it is a circle. If e+,... ,e is a circuit or a loop then so are all its cyclic
permutations £,-»tfj+j, • • • >en>e\> • • • >ei-\ • W n e n lt iS necessary to distinguish

Downloaded from https:/www.cambridge.org/core. New York University, on 30 Jun 2017 at 09:33:05, subject to the Cambridge Core terms of use
Cambridge Books
, available at https:/www.cambridge.org/core/terms. Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.006
Graphs 114

these paths and loops from the topological ones we shall call them edge-paths
and edge-loops.
Let / be the path e^... yen and let g be the path e f j , . . . , e * . When'
e'jO - enx we define the product f.g of / and g to be the path
ei,...,en,e'i,...,e'm. When e^a is v and ej is w, we also define v./and f.w to
be /. The inverse f of / is e"^,... ,e"j; if the path is a vertex v, then the inverse
of / is also v.
A subgraph Tj of T consists of subsets Kj of V and E* of E such
that, for any e € E^, we have e~ c £ j and eo € V^; evidently ex is also in Kj, and
T| is a graph. Evidently the intersection and the union of arbitrary subgraphs of
T are themselves subgraphs of T. When V^ is a subset of K, the full subgraph on
V* is the subgraph whose vertex set is Kj and whose edge-set E. is given by
£ j - [e; ea and e\ are in V^i.
We say that T is connected if given any two vertices of T there is
a path joining them. More generally, for any T we can define a relation on V by
requiring v to be related to w iff there is a path starting at v and ending at w.
This is easily seen to be an equivalence relation. The full subgraphs whose
vertex sets are the equivalence classes are called the components of T.
Note that the union of connected graphs T- will be connected
provided that C\T.^0. We need only take some VQeOF^. and observe that every
by a path in some I 1 ., so that the component
vertex of u r . can be joined to v^ b
of VQ in Ur^. contains every vertex.

Lemma 1 Let T be connected. Let S be a non-empty subset of V

such that, for any edge e, if eo is in S then ei is also in S. Then SmV.

Proof Take v t S. Let w be any vertex. Since T is connected, there

is a path e^f... ,en from v to w. Since e^a - v is in 5, an easy induction shows
that the end-points of all e. are in S. In particular, w- en% is in S. As w was an
arbitrary vertex, we see that SmV.II

A path is called reducible if there is some / such that ^/+i"^/J

otherwise it is called irreducible. Evidently a reducible path is not simple. If
n>2 and the path is reducible then e^... ,e-_^ ,^ /+ 2 »• • • >en ls a ^ s o a Pa*h. I n
particular, if two vertices can be joined by a path (equivalently, if they are in
the same component) then they can be joined by an irreducible path.
A forest is a graph all of whose loops of positive length are
reducible. A tree is a connected forest. Plainly any subgraph of a forest is a
forest. It is also clear that a graph is a forest iff all its components are trees.

Lemma 2 A graph is a forest iff it contains no circuits.

Proof P l a i n l y a forest c o n t a i n s n o c i r c u i t s . S u p p o s e that T i s n o t a

forest. T h e n it c o n t a i n s a n i r r e d u c i b l e l o o p o f p o s i t i v e l e n g t h e^9...,e . Take r
and s w i t h s-r as s m a l l a s p o s s i b l e subject t o l^r^s^n a n d e o-e T. ( S u c h r
r s
and s exist, since e*omejz.) Then em,ew,.^,...fec. is a circuit, since it is a loop
1 n T T "^l S

of positive length by definition of r and s, it is irreducible because the original

loop is irreducible, and the beginnings of its edges are distinct by the choice of
r and s.ll
Lemma 3 A graph is a forest iff, for all vertices v and w, there is
at most one irreducible path from v to w.
Proof If the graph is not a forest it contains a circuit e+,... ,en.
e e an< are tn
Then \>- • • > n-\ * ^ b° irreducible paths from e^a to ena (when nm\
the first of these paths is a single vertex).
Now suppose that the graph contains two distinct irreducible paths
joining some pair of vertices. Take irreducible paths / and g which start at the
same point and end at the same point and with lengths n and m such that m + n
is as small as possible. Let / b e e* , . . . ,em and let g be e\ , . . . ,e' . If m - 0 and
in v m
nm\ then / is a circle, and hence / is a circuit. If m- 0 and n> 1, then / is itself
an irreducible loop.
If we had enm e"m then the paths e^,... , e n ^ and e^ , . . . ,e{'m_i
would be distinct irreducible paths from e^a to e^p. This would contradict the
minimality of m + n, and so eM#e' . The path f.g~ is therefore an irreducible
minima n rn
loop.//

Lemma 4 If T is a tree then any two vertices of T can be joined

by exactly one irreducible path. Conversely, if T has a vertex VQ such that v^
can be joined to every vertex by exactly one irreducible path then T is a tree.

Proof Let r be a tree. As T is connected, any two vertices may be

joined by an irreducible path. As T is a forest, there cannot be more than one
irreducible path joining two given vertices.
Conversely, suppose that the vertex VQ can be joined to every
vertex by exactly one irreducible path. Then T is connected, since every vertex
is in the component of v^. Suppose that T is not a tree, so that it contains a
circuit # j , . . . ,en. Let / be the shortest path from VQ to a vertex of this circuit.
Since cyclic permutations of circuits are circuits, we may assume / ends at e^o.

Also, the last edge of / cannot be en, else we could omit this edge from / and
get a path from VQ to e^o which is shorter than /. For the same reason, the last
edge of / is not e~^. Let g be the path e^,... ,^n^ • Then fg and fe~n are two
distinct irreducible paths from VQ to eno.//

Lemma 5 Let T be a tree. Let T. be subtrees of T (for i in some

index set) such that 07^*0. Then both UT. and C\T. are trees.

Proof Since each T. is connected and 07^*0, we know that UT- is

connected. As a subgraph of a tree, UT. must be a forest. Hence it is a tree.
As a subgraph of a tree, C\T- must be a forest. We must show that
it is connected. Take two vertices v and w of ftT.. For each / there is an
irreducible path f. from v to w in T-. Since T is a tree, there is only one
irreducible path in T from v to w. Hence the / . for differing / are all the same,
and they are therefore a path in DT- from v to w.ll

Lemma 6 Let T• be trees such that there is a vertex v with

T^T.'iv} for i*j. Then UTi is a tree.

Proof As O7T.^0, UT- is connected. Suppose that it contains a

circuit. Taking a cyclic permutation if necessary, we may assume that the circuit
is e^9...9en and that ep±v except possibly for r~ 1. For r<n, if er*T- and
e
r+\*T- we must have / - / . For otherwise er^o-ex would be in T.nT.y which
is {v}. It then follows that the circuit must lie completely in one T-. This is
impossible, as T^ is a tree.//

A maximal forest in a graph T is a subgraph F of T such that F is

a forest and no subgraph of V which properly contains F is a forest.

Proposition 7 Let the forest F be a subgraph of T. Then there is a

maximal forest of T which contains F. In particular (take F empty) every graph
contains a maximal forest.

Proof Suppose that we have a collection of forests F- such that for

all / and j either F.cF- or F-cF-. Then U/^. is a forest. Otherwise the union
contains a circuit. This circuit would be in the union of finitely many of the
Fj, and the assumed property ensures that any finite union equals one of them.

Thus some F- would contain a circuit, which is impossible.

The result now follows by Zorn's Lemma.//

Proposition 8 Let F be a maximal forest in F, and let Fj be a

component of F. Then F contains every vertex of T. Also FnT^ is a maximal
forest of F| and is a tree.
Proof If there were a vertex v of F not in F then FvM would be a
larger forest.
FnTj is a forest in F j , being a subgraph of F. If it were not a
maximal forest in I\ it would be properly contained in a maximal forest F* of
Tp Then FuF^, being the disjoint union of F^ and ^ " T p would be a forest of
T properly containing F.
We have to show that FnF^ is connected. By Lemma 1, it is
enough to show that any edge of Fj has both its vertices in the same component
of F. Suppose not. Take an edge e such that ea and ex are in different
components.
We shall show that F\j{eye~) is a forest. It is certainly a graph, since
F contains all the vertices of F. If it contained a circuit, we could assume
(taking cyclic permutations, if necessary) that the circuit is e^,... ,en where er is
not e or e~ if r> 1. If e. is not e or e~ we would have a circuit in F. If e* is e
or e~ then the path e^, • • • >en would show that ea and ex are in the same
component of F, contrary to assumption.
Since F is a maximal forest, we have a contradiction, and so Fc^T*
must be connected.//
In particular, a maximal forest in a connected graph is a tree (which
we refer to as a maximal tree). It is remarkable that a maximal forest contains
only one tree! The following corollary is immediate.

Corollary Any forest in a connected graph is contained in a

maximal tree.

Lemma 9 Let F be a graph, and let T be a tree in F which

contains every vertex. Then T is a maximal tree.

Proof Let A be a subgraph of F wich strictly contains T. Since

every vertex is in 7\ there must be an edge e with e e A and e < T. Then A
contains two distinct irreducible paths from ea to ex, one consisting of the edge
e only, the other being a path in T. Hence A is not a tree.//

Lemma 10 Let T be a finite tree with more than one vertex. Then
T has {at least) two vertices each of which is the start of exactly one edge
{and the end of exactly one edge).

Proof Since T is finite, there is a bound on the lengths of paths all

of whose vertices are distinct. Let e^9...9e be the longest such irreducible path.
Suppose that there is an edge e**^ with eo - e j . . If ex - ep for some o we
would have an irreducible loop er,... 9en,e. This is impossible as T is a tree. It
follows that e^,... ,en,e is also an irreducible path with distinct vertices,
contradicting our choice of n. Hence en and e~n are the only edges of which e^
is a vertex, so that e x is one of the required vertices. Similarly e^a is the other
vertex required.//

Proposition 11 Let F be a finite connected graph with n vertices

and m edge-pairs. Then F is a tree iff nmm + l.

Proof Let F be a tree. If nm\ then mm0 (since any edge would be
a circle). Otherwise let v be a vertex which is the start of exactly one edge,
which we call e. Let A be the graph obtained from F by deleting the edges e
and e~ and the vertex v. (This is a graph, since no other edge has v as a vertex.)
Then A is a forest. If it is a tree, the result will follow by induction.
Take any vertices u and w of A. They can be joined by an
irreducible path e* , . . . ,e in F. Suppose that there is some r with e x - v. Then
m
er^o v. Since e is the only edge starting at v, we must have ef + j - e - e~f. The
path is then reducible, contrary to hypothesis. Hence the path does not go
through the vertex v, and so it is a path in A, as required.
Now let F be a finite connected graph with m- n- 1. Let T be a
maximal tree in F. Since T contains every vertex of F, it has n vertices. By what
we have just proved, it must contain n-\ edge-pairs. By hypothesis, these will be
all the edge-pairs of F, and so J T - F . / /

Let 7\ , for X in the index set A, be subtrees of the graph F such

that 7 \ n T - 0 for \^[i. Assume that every vertex is in some 7 \ (if this did
A JX A
not hold we would just add further trees, each consisting of one vertex).
Construct a graph A as follows. The vertex set of A is to be A, and the edge set
of A is to consist of those edges of F which are not in U 7 ^ . The map "• on F
evidently provides a map ~ on A. When e is an edge of A whose start, regarded

as an edge in F, lies in 7^ we define the start of e as an edge of A to be X.

We say that A is obtained from F by contracting the trees T^.

Lemma 12 Let A be obtained from F by contracting the subtrees

T^ . Then A is connected iff F is connected, and A is a forest iff F is a forest.

Proof Let e« , . . . , ? be a path in F. Suppose that, for some r<s, we

have er<\JT^ and es<UT^ but e^UT^ for r<i<s. If e.tT-^ and e / + 1 e r then
e .T - ez-+|0 is in r ^ n T . Hence X - JI, by our requirements on the subtrees. It
follows that there is some v such that T contains e T and e o. We see from
this that when we omit from the path e^ ... ,en all the edges lying in U7^ the
result is a path in A.
Suppose that the path in F is irreducible, and keep the same
notation. We cannot have e - e~ with s - r + 1 , since the path in F is irreducible.
s r
And we cannot have esm^r with s>r+\, as then er+\>-• • >es.\ would be an
irreducible loop in 7*. We deduce that the path in A is also irreducible.
Now let us start with an irreducible path e+ , . . . ,e in A. Then for
each / there is some u with e-i and e ^ o , as vertices of F, both lying in T .
Hence there is an irreducible path p- in T from e >x to e + i o. Then the path
qm eyPye* p _*.e is an irreducible path in F. More generally, let e*o and
e nT in A be a and P, and let v and w be vertices of F in T oc and Tpa,
respectively. Then there are paths p^ and pn in T^ and 7V such that p^.q.pn is
an irreducible path in F from v to w.
Let F be connected, and let a and P be vertices of A. Take vertices
v and w of F in T^ and TQ. Then a path in F from v to w gives rise, by the
above, to a path in A from a to p. Hence A is connected.
Let A be connected, and let v and w be vertices of F. Take a and 3
with v e 7*a and w e TQ . Then there is a path in A from a to P, which gives rise,
by the above, to a path in F from v to w.
If F is not a forest it contains an irreducible loop. The
corresponding path in A will be an irreducible loop, and so A will not be a
forest.
If A is not a forest there are vertices a and p such that there are
two distinct irreducible paths in A from a to p. Take vertices v and w of F in
7*a and TQ. These paths give rise to two distinct irreducible paths in F from v
to w, and so F is not a forest.//

The remaining property of trees will be used only in the proof of

Grushko's Theorem in Chapter 8. It is the beginning of homology theory.
Let F be a graph. Let CQ be the free abelian group with basis the
vertices of F. Let Cj be obtained from the free abelian group with basis the
edges of F by factoring out the subgroup generated by all e +e~. This may also
be described as the free abelian group with one basis element for each edge-pair.
Let E:CQ-»Z be the homomorphism given by ve-1 for each vertex v.
Let d:Cj-» CQ be the homomorphism given by ed - ex - eo for each edge e (it is
easy to check that d is defined on Cj). Plainly de - 0, and so Co

Lemma 13 T is connected iff C^d • kers, and F is a forest iff

kerb - 0.

Proof Let F be connected, and let v and w be vertices. Take a path

e^9...9en from v to w. Then e^ + . . . + en is an element of Cj which maps by d
to w- v. It is easy to check that kere is generated by the elements w- v of CQ for
all vertices v and w. Hence kere - C^d,
Let F be disconnected, and let A be a component of F. It is easy to
check that dt' - 0, where e' is the homomorphism given by vef - 1 if v € A and
vef - 0 otherwise. So w- u is not in C+d if u is in A and w is not in A.
Suppose that F is not a forest. Then F contains a circuit e*9...9e .
By the definition of a circuit, we cannot have ^-"C-- Hence e^ +... + en is a
non-zero element of Cj which maps by d to 0.
Suppose that there is a non-zero element x of Cj with xd - 0. Write
x as I.mê-9 with each w.eZ. We may assume the e- are all distinct and that
each m. ? 0. We may also assume that we never have e-m e~.. For suppose that
^.-e~. . We need only rewrite mê^m-e. as {m.-m.)e. (note that the definition
of a graph ensures that / ^ y), and similarly for other pairs, to get rid of this
possibility. Finally, we can assume that each m. is positive, replacing m-e- by
(-m.)ej if necessary.
Suppose this done. Let A be the finite subgraph of F consisting of
all e. and e~.and
andtheir
theirvertices.
vertices.Because
Becausexdxd- 0,
- 0,it itisiseasy
easytotosee,
see,using
usingthe
the
itions w
conditions we have required on x, that for each / there must be j and k such
mm
that e-o-ejz and eê^ eô.
eô. By Lemma 10, this tells us that A is not a forest.
Hence F is not a forest.//

Exercise 1 Let e be an edge in a connected graph F. Show that

r - ieye~) is connected if e is in some circuit and consists of exactly two
components if e is not in any circuit.

5.2 COMPLEXES AND THEIR FUNDAMENTAL GROUPS

Let / be a loop in a graph. The set F consisting of / and all its
cyclic permutations is called a cyclic loop. The cyclic loop /*" consisting of f
and all its cyclic permutations is called the inverse of F.
A complex K consists of three disjoint sets, K, E> and C (the
vertices, edges, and ce//s of K) with the following properties;
(i) the vertices and edges form a graph;
(ii) there is a map ~ from C to C such that, for all c*C, c~2c and
zr - c;
(iii) to each ceC there is a cyclic loop cc) such that <Tdmcd .
We will refer to an edge of cd or cB as an edge of c, and to a
vertex of such an edge as a vertex of c. For convenience of notation, we will
usually write cd to denote some loop belonging to the cyclic loop, rather than
the cyclic loop itself. We call cd the boundary of c.
As with graphs, readers are left to decide for themselves in each
situation whether the empty complex is permitted or not.
The graph consisting of the vertices and edges of the complex K is
called the \-skeleton K of AT. By a path (loop, circuit, and so on) in K we will
mean a path in K .
A complex as defined above is the combinatorial analogue of the
topologicai (and partly combinatorial) notion of a 2-dimensional CW-complex.
(We have no need of this concept and shall not discuss it further, but readers
with some familiarity with the notion will find that some of the concepts of this
chapter are suggested by results in that theory.) This seems to me to be the most
natural setting in which to work. Fortunately we do not require analogues of
CW-complexes of higher dimension, which would be messy to define. Any
2-dimensional simplicial complex gives a complex in the above sense (once we
have added inverses of edges and cells).
A graph is just a complex with no cells.
As already remarked about graphs, the use of the map "" on cells
and edges is just a technical convenience. When giving a picture of a complex
we usually only refer to one edge or cell in each pair.
A subcomplex AT, of a complex AT consists of sets V* c. V, E+£ E9
and Cj.cC such that if c € Cj then c~ € Cj and every edge of c is in E^ and if
e € £j then e~ e E^ and eo e Kj. It then follows that K^ is a complex, the
operations being those of AT. The intersection and union of subcomplexes of AT is
obviously a subcomplex of K.

The full subcomplex of K corresponding to the subset V^ of V is

the subcomplex whose vertex set is V^, whose edge set consists of all edges
whose vertices are in V* , and whose cell set consists of all cells whose vertices-
are in Kj.
We say that the complex K is connected if any two vertices can be
joined by a path. Thus K is connected iff K is connected. More generally, we
can obtain an equivalence relation on the vertices by making v related to w iff
there is a path from v to w. Then K is connected iff there is only one
equivalence class. A full subcomplex whose vertices form an equivalence class is
a component of K. If £ is a component of K then L is a component of K .

A map or morphism 9 from a graph F to a graph A consists of two

maps, both denoted by 9, one from V(F) to K(A) and the other from E(T) to £(A)
such that, for all e * E(T), e"cp - ey~ and e(po-eo<p. It immediately follows that
e<pj - exep.
Let / be the path e^,... ,en in T. It is easy to check that
£j<p,... ,en<p is a path in A. We denote this path by /cp. If / is a loop then /<p is
a loop. If F is a cyclic loop containing the loop / then the cyclic loop
containing f<p depends only on F, and not on the choice of /. We denote it by
F<p.
A map or morphism <p from the complex K to the complex L
consists of three maps, all denoted by 9, one from V(K) to K(L), one from E{K)
to E(L)y and one from C(K) to C{L) such that, for e * e{K) and c e C(K), e~<p-elp
and ecpo-eocp, and c~cp - c K such that cptp a ^d ^9 are identity maps.
If V}£V, E^cE, and C^cC then (V^9E^,C^) is a subcompiex of K
iff the inclusion is a map of complexes.
Evidently, the identity map on a complex K is a map of complexes
in the above sense. Also, if cp:AT-» L and ty:L -> M are maps of complexes then 9^
is a map of complexes. In category language, we have a category of complexes
and maps (and also a category of graphs and maps).
Let <p:K^ £ be a map of complexes, and let K, and L. be
-1
subcomplexes of K and L. Then K^$> and L^9 are easily seen to be
subcomplexes of L and K, respectively.
Let (p:K^ L be a map of complexes, and let K be connected. Then
it is easy to check that Aip is also connected.
Note that the maps in our category are not the analogue of cellular
maps in CW-complexes; they correspond to a special kind of cellular map.

Let / be t h e p a t h e^9...9e . W e say that t h e p a t h g is o b t a i n e d

from / by elementary reduction if
either (i) for some /, e .^ - e~. and g is e^,... ,e-^ ,e -^ »• • • *en
or (ii) for some / and j with i^j and some cell c, the path
e.y...,e- is a loop belonging to the cyclic loop cd, and g is the path

It is easy to check that the deletion of the stated edges does leave a path in
either case if not all edges have been deleted. If the deletion would remove
every edge we regard g as being the path consisting of the vertex e*o.
We call the path / homotopic to the path g, and we write f^g, if
there is a sequence of paths / for X^r^m (some m) such that /^ is / , / is g,
and, for all r<m, either / +« is an elementary reduction of / or / is an
elementary reduction of / ^ j . T h i s is obviously an equivalence relation. We shall
use the phrase edge-homotopic when we need to distinguish this from the
topological notion of homotopy.
It is immediate that if / is a path from v to w and f^g then g is
also a path from v to w, and f^g". Further, for any path / ' starting at w, if /
is a sequence of paths which show that f^g then the sequence fr.f shows that
ff'^g.f. Similarly, if / " is a path ending at v then f".f^f".g .It follows that
/ ' is a path starting at w and f'^g' then / . / ' ^g.f °*g.g'.
The homotopy classes of paths therefore form a groupoid, which we
call the fundamental groupoid (or edge-groupoid) of AT, written Y(^0- This
groupoid has vertex set V{K). The vertex group of the groupoid at the vertex v
(that is, the set of homotopy classes of loops starting at v) is called the
fundamental group of K with base-point v, written 7tj(AT,v). When K is connected
this does not depend on v, and we often just write nAK). Any path from v to w
induces an isomorphism from iz.(K,v) to nJK,w).
-1
We will frequently find it convenient to write / to denote the
inverse of the path / , rather than / , to emphasise that its class is the inverse of
the class of /.
Let <p:K-> L be a map of complexes. Let the path g be obtained
from the path / of K by elementary reduction. Then gcp is obtained from ftp by
elementary reduction (for instance, if we remove cd from / to get g then we need
only remove cc)(p from /<p to get g<p, and cd<p - c<pd is the boundary of ccp). It
follows that if f^g then f<p^g<p. Hence cp gives rise to a homomorphism
<p*:y{K) •* y{L), and also to a homomorphism 9* from 7TJ(AT,V) to 7ij(L,vcp).
If ty:L -» M is a map of complexes then (<p(JO* * 9*^*- If 9 is the
identity map on K then 9* is the identity homomorphism on "x(K). Hence, if 9 is

an isomorphism of complexes then cp* is an isomorphism of group(oid)s. In

category language, we have a covariant functor from the category of complexes
to the category of group(oid)s.

Examples Let K be the complex with two vertices v and w and one
edge-pair ie,e~) with ea - v and ei - w. Then K is connected and y(K) is the tree
groupoid on (v,w), so that 7ij(AT,v) is trivial. For any non-trivial path must have
its edges alternately e and e~, and so it must be homotopic to one of v, w, e, or
e~, and no two of these can be homotopic, since no two have the same start and
finish.
Let AT be a graph -with one vertex v and one edge-pair {eye~)\ we
must have e o - v - e x Then n^(K9v) is infinite cyclic, generated by the class of e.
For any path is a loop, and any path is plainly homotopic either to v or to en
or to e~n for some n > 0. If a path has r occurrences of e and s occurrences of e~
then r - 5 in unchanged by elementary reduction. Hence r-s is constant on
homotopy classes, which shows that no two of the above loops are homotopic.
The theorems on fundamental group(oids) of spaces have analogues
for the fundamental group(oid)s of complexes, which are usually much easier to
prove.
Theorem 14 (van Kampen's Theorem) Let K^ and K^ be
subcomplexes of the complex K, and let ATQ be AT, n AT« . Then the inclusions
induce a pushout of groupoids

Y(AT2) -*

Corollary 1 Suppose further that S is a subset of V{K) such that,

for / - 0,1,2, S meets every component of K-, and let S. be SnV{K-). Then the
inclusions induce a pushout of groupoids

are
Corollary 2 Suppose further that KQ, K^, and K2 connected
(and then K mil also be connected). Let v € V(K^). Then the inclusions induce a
pushout of groups

Corollary 3 Let K^ and K^ be connected, and let K^n K^m M. Then

Proof Corollary 3 is a special case of Corollary 2, which in turn is

a special case of Corollary 1. Corollary 1 follows from the theorem and Example
3 on pushouts of groupoids.
The proof of the theorem is similar to the topological case, but
simpler.
We take homomorphisms <p- from y{K) to a groupoid G for / - 1 , 2 ,
such that the two homomorphisms from Y(^Q) to G are the same.
We may extend the <p^ to multiplicative maps 4>; from the set of
paths of K. into G. Our assumption tells us that 4>j" 4>2 o n P a t n s °f ^Q* *n
particular, since each edge or vertex of K may be regarded as a path and lies in
either A^ or K2, these maps give a map ijj from the edges and vertices of K into
G. Further, for any vertex v, vcp will be an identity in G> since v is an identity
on the paths of K.
Let e and e% be edges such that e\ - efo. If e and e% are both in K^
(or in Kj) we know that (ety)(e'ty) is defined (and equals (e.e')ty ), since i|>i is
multiplicative. If one of them is in K^ and the other in A^ then (e<W(e'<Jj) is still
defined, because {e\)ty is a right identity for ety and a left identity for e ty.
Now let / be any path, say the path e^,... ,e . By the previous
paragraph, the product (e^)... ( e ^ ) exists in G. We define /ty to be this
product. Evidently cp is a multiplicative map which extends ^ and (J^ •
Let / ' be obtained from / by elementary reduction. If we have
e tnen e e
r+\~^r r+1$ ~ ( ft) f a n d s o f'tymfty> Otherwise, / ' comes from / by
deleting the sequence of edges er,...,es which form the boundary of a cell c.
Now c is in ATj or in K^- Hence the path er,...,es is homotopic in K^ (or in
K2) to a trivial path. Since ^- arises from <p., which is defined on homotopy
classes, it follows that (e i|>)... (e A>) is an identity in G. Again we find that

Since ty is unchanged by an elementary reduction, it is constant on

homotopy classes. Hence it gives rise to a homomorphism <p:y(K) •» G, as required.
There is only one possibility for such a map <p, since its value on the class of
an edge is given to us, and any element is the product of such classes.//

By a similar simplification of the topological case, we obtain the

following results on direct limits.
Theorem 15 Let A be a directed set, and let K^ (for X e A) be
subcomplexes of the complex K such that AT-UA!\ and K^£K for \<{L. Then
A A ^1

y(K) is the direct limit of the T ( ^ ) {the maps being induced by the inclusions.
Suppose further that each K-^ is connected, and that v e OA^ . Then
7Cj(AT,v) is the direct limit of the n^Ky).//
Corollary 1 For any complex K, y(K)m limy(L), where L runs over
all finite subcomplexes of K.I I
Corollary 2 Let K be a connected complex. Then
7TJ(/L,V)- limn^(L,v)f where L runs over all connected subcomplexes of K which
contain v.lI

Example Let r be a graph with only one vertex v and a collection

of edge-pairs {e.,e~.}. Then 7ij(O is free with basis ie.}. For we have seen that
this holds when there is only one edge-pair. Inductively, using van Kampen's
Theorem, it holds when there are only finitely many edge-pairs. It holds in
general by Corollary 2.
We now obtain a presentation for nAK), where AT is a connected
complex. It turns out to be inconvenient to use van Kampen's Theorem, because
the boundary of a cell need not be a circuit.
Any path in K can be regarded as an element of F(E), the free
group with basis the set E of edges of K, regarding the path e ^ . . . ,en as the
product e*... e in F(E). We shall use the same symbol for a path and the
corresponding element of F(E).
Let RcF{E) be {ee~; all etE)v{cd; ail aC). We call R the set of
relators of the first kind.
Choose a vertex a of K. For each vertex v choose a path / y from a
to v, and let fQ be the trivial path. Let ScF(E) be {/y; all v € V). We call S the
set of relators of the second kind.

Theorem 16 <E; RuS> is a presentation of ii^(Kfa). In this

presentation the edge e from v to w maps to the homotopy class of the path

Remarks We can replace the loop cd by another loop giving the

same cyclic loop. This just replaces an element of R by a conjugate. We plainly
need only take cd for one cell c from each cell-pair. We could take only one
generator e from each edge-pair, replacing the corresponding e~ by e in the
relators; we will always do this when we have specific examples.

Proof Let G be <E; RuS>. Let N be <RuS>F^E\ so that wN is the

element of G corresponding to weF(E). Let If] be the homotopy class of the
path /.
Take the homomorphism from F(E) to izJK,a) which sends e to
_i *
Lfv.e.fw ], where e starts at v and ends at w. It is easy to check (inductively)
that, for any path / from v to w, this sends / t o [/ . / . / ].
In particular, this homomorphism sends ee~ to the identity. Also, if
the loop cd starts at v, it sends cd to If . cd.f~ ], which is the identity.
-1
Finally, it sends / y to [/ - / v . / v X which is also the identity. Hence it induces
a homomorphism <p:G-* nJK,a).
To each path / we have a corresponding element fN of G. This is
plainly a multiplicative map. Also ee~ and cd go to the identity. Hence this map
is constant on homotopy classes. It therefore defines a homomorphism 4> from
y(K) to C7, which restricts to a homomorphism i|):7rj(AT,a) -> G.
In particular, / maps to 1. Hence, when e begins at v and ends at
1
w, ^
Also, for any edge e from v to w, we have
- (eN)<p-lf .e.f'h. Inductively, for any path / f r o m v to w,
y w
- r/y././ 1 In particular, since / is trivial, if / is a loop based at a
then C/lcJxp - [/].
It follows that the homornorphisms <p and ip are inverses of each
other, and so <p is an isomorphism.//

This theorem will rarely be used, as the special case of it given

below is more useful. We will refer to the presentation below as the presentation
using the maximal tree T.

Theorem 17 Let T be a maximal tree in K. Then nAK) is

presented by <E; R u {e; e e T}>.

Proof For each vertex v, let / be the unique irreducible path in T

from a to v. Evidently the elements / y of F{E) are consequences of te; eiT).
Conversely, let e be an edge in T from v to w. If / does not end
in e~ then / y . e will be an irreducible path in T from a to w\ that is, fv-emfw- If
/ y does end in e" then removal of this edge gives an irreducible path in T from
a to w\ that is, / - / .e~. In either case, e is a consequence of {/ ; all
v € V) u tee"; all e € E). Hence both presentations give the same group.//

Corollary 1 Let Y be a connected graph, and let T be any maximal

tree in Y. Then KAY) is free, with one basis element for each edge-pair not in
T. Further, the basis element corresponding to a chosen edge e is the class of
fv.e.fw~ , where e starts at v and ends at w, and / y is the unique irreducible
path in T from the base-point to v.l I

Corollary 2 A connected graph is a tree iff its fundamental group

is trivial.//

Corollary 3 Let T be a connected graph with n vertices and m

edge-pairs. Then the rank of n^{Y) is m-n+l, provided n is finite.

Proof We know there is one basis element for each edge-pair not in
a chosen maximal tree. Now the maximal tree has exactly n vertices, so by
Proposition 11 it has n-\ edge-pairs. Since there are m edge-pairs altogether, we
see that the number of basis elements is m-n+l.//

Corollary 4 Let K be a connected complex, and let L be K\J{C,C~),

where c is a cell not in K and cd is a cyclic loop in K. Then nJL) is obtained
from nJK) by cancelling the class of the loop cd.//

Remark More precisely, when we specify base-points, nAL9a) is

L
_i

obtained from n^(Kfa) by cancelling the class of f.cd.f , where / is any path
from a to the start of cd.
We can extend Corollary 4 inductively to the case where L is
obtained from K by adding a finite number of cells. Then Theorem 15 allows us
to extend to the case where we add an arbitrary number of cells. Thus we have
the following result.

Corollary 5 Let K be a connected complex, and let L be

cu ,c~^ (XeA)), where c^ is a cell not in K and c-^6 is a cyclic loop in K.
Then nAL) is obtained from Ky{K) by cancelling the classes of the loops Cyd.

Theorem 18 For any group G there is a complex K such that n^{K)

is {isomorphic to) G.

Proof Let <X;R> be a presentation of G. Let K be the graph with

only one vertex and with one edge-pair for each element of X. Then nAK ) is
the free group on X.
Hence any element of R can be represented as a loop in K . We
obtain a complex K from K by adding ceils. Let K have one cell-pair (cr><y} for
each nRy with cfd being the cyclic loop containing the loop corresponding to r.
Evidently n^{K) is isomorphic to G.ll

Theorem 19 Let L be a connected complex, and let <p be a

homomorphism from nJL) to the group G. Then there is a complex K containing
L, such that n^{K) is G and the inclusion of L in K induces 9.

Proof Let M be a complex with 7tj(A/)- G, and let N be the union

of L and M with one common vertex. Then n^{N) - n^{L)* Gt by Corollary 3 to
Theorem 14.
We obtain K from N by adding cells. Let Y generate n^L). For
each y £ Y take a loop f in N whose class is (jxp)y . Let K be obtained from N
by adding cell-pairs (cy><^J f° r e a c n J ^ ^ where c 6 is the cyclic loop
containing / . Then, by Corollary 5 to Theorem 17, 71,(^0 is obtained from
-1
7Tj(A0 by cancelling all the elements 0><pb> . This gives the result.//
Corollary Let L- be connected complexes, let G be a group, and
let (p-:n^{L.)^ G be homomorphisms. (/) There is a complex K containing {copies
of) the L- and a vertex VQ such that L-nL-a {VQ} for i£j and such that n^{K) is
G and the inclusion of L- in K induces <p-. (//) There is a complex M
containing {copies of) the L- such that L.nLm0 for i*j and such that n^{M) is
G and the hhomomorphism induced by inclusion of L- in K is 9 • followed by a
conjugation.

Remark In (ii) we have to use different base-points for the different

-L, and in order to refer each to a fixed base-point for K we must use the

isomorphism induced by a path. A different path would provide another

isomorphism, related to the first by conjugation. In fact our construction will
give a preferred path to use.

Proof (i) Let L be obtained from the disjoint union of the L. by

identifying one vertex from each, so that L-OL.-CVQ} for i&j. Either from
Theorem 17 and Lemma 6 or from van Kampen's Theorem and the direct limit
argument, we see that TTJ(L) is tn^L.). Now apply the previous theorem to L
and the homomorphism i<p-.
(ii) We may assume that the L- are disjoint. Choose a vertex v. of
L-. Let M- be obtained from L- by adding a new vertex w (the same for each /)
and an edge-pair {e-,e~-} with eamw and eimv-. Then nJM-,w) is isomorphic
to n^LjtVj), and we need only apply part (i) to the M- with common vertex w.//

Let AT be a connected complex, and let ATQ and K^ be connected

subcomplexes of AT such that ATQO AJ - {v} and such that there is an isomorphism
&:KQ-> K^ with V&- v. .We may obtain a complex L from K by identifying
vertices, edges, and cells of ATQ with the corresponding elements of K^ under #.
Let /Q and /« be the inclusions of ATQ and A!* into AT.
Proposition 20 WrA this notation, nJL,v) is obtained from TZAK,V)
-1
by cancelling the elements (giQ*&*) (gi^*) for all gt 7TJ(ATQ,V).

Proof To begin with we show that £ is a tree if AT is a tree.

Let AT be a tree. Then AQ and ATj are both trees, and so is KQ U ATJ ,
since ATQ n ATJ is a vertex. By Lemma 12, the graph obtained from K by
contracting KQ U ATJ is also a tree. This graph is plainly also the graph obtained
from L by contracting the tree A Q . By Lemma 12 again, L is a tree..
We now return to the general case. Let TQ be a maximal tree in
K~, and let T* be TQ&, which is a maximal tree in K* . Then TQ U 7\ is a tree,
and so (by the corollary to Proposition 8) it is contained in a maximal tree T of
AT. By the previous paragraph, the image S in L of 7* is a tree, which is a
maximal tree, since it contains every vertex of L.
Now look at the presentations of 7Tj(AT,v) and 7ij(Z,,v) given by
Theorem 17, using the maximal trees T and 5\ It is clear that the presentation of
7Tj(L,v) comes from the presentation of 7tj(AT,v) by first adding new relations
e - e%- for all edges e of ATQ not in T^ and then using Tietze transformations to
eliminate the generators e§.

Since (e; ei^^e^T^) is a set of generators for TI^{KQ9V), we obtain

the result.//

Now suppose that AT is a connected complex, that KQ and ATj are

connected subcomplexes with ATQOATJ-0, and that &:KQ-> K^ is an isomorphism.
Let L be obtained from K by identifying KQ with K^ using $, as before. Let /Q
and i* be the inclusions of KQ and AT, in K. Let VQ be a vertex of ATQ and let v«
be VQO. Let / be a path in K from Vj to VQ .

Proposition 21 With this notation, iz^L) is the pseudo-HNN

extension Kn^K.v^t; (TT^A^ ,v o )i Q *) r - TT^ATJ ^ / J * / * ^

Remark By (ii) of the corollary above, every pseudo-HNN extension

can be obtained in this way.

Proof Let M be obtained from K by identifying VQ with v^. Then L

comes from M by identifying the images of ATQ and AT, . So we may apply the
previous proposition, once we have determined 7TJA/,VQ) and the homomorphisms
induced by the inclusions J'Q and y'j of KQ and ATj in M.
Let Nm ATu ie,e~}, where e is a new edge with ea-v~ and ex - v^.
Then M is obtained from N by contracting the edge e. Take the presentation of
7TJ(7V,VQ) corresponding to a maximal tree containing e and the presentation of
7TJ(A/,VQ) corresponding to the image of this tree (which will be a maximal tree
in M). We see that the two groups are isomorphic, since the only difference
between the presentations is that the former has an extra generator e and also an
extra relator e.
By the groupoid version of van Kampen's Theorem and Example 1
on pushouts of groupoids, TTJ(MVQ)- TTJ(AT,VQ)* <r>, where t is the class of the
path e.f. If we regard TTJ(A/,VQ) as the same as TTJ(A/,VQ) then /Q* is plainly the
same as /Q*. Also j \ * will be i**e~#. Since t is the class of e.f, it follows that,
for gen^K^ ,Vj), g/j* - tigi^ *f#)fl. The result now follows.//

Exercise 2 Let (p:K-> L be a map. Let K, and L, be subcomplexes

-1
of K and L. Shows that K+<$ and Z,|9 are subcomplexes of L and AT.
Exercise 3 Prove Theorem 15.

5.3 FREE GROUPS AND THEIR AUTOMORPHISMS

In Chapter 7 we show that any subgroup of a free group is free.
Three proofs are given in that chapter, all very closely related, and another
related proof is given in section 8.1. The first proof in Chapter 7 requires the
theory of covering complexes, which is developed in Chapter 6. The other
proofs, however, could be read at this point. In this section we look ar
automorphisms of free groups.
Let a be an automorphism of a finitely generated free group F, and
let Fixa denote the subgroup igtF; got-g) For some years it was an open
question whether Fixa was always finitely generated, and some interesting
partial results were obtained. Then, almost simultaneously, several different
prooofs appeared. We shall follow the proof due to Goldstein and Turner (1986),
which is remarkably simple, and which shows that Fixr) is finitely generated for
any endomorphism rj of F.
We begin with a lemma about graphs. An orientation of a graph T
is a subset A of E such that An A' 0 and A u A - E. The outdegree vw of the
m
vertex v (with respect to A) is \{e c A; ea v}\.

Lemma 22 Let Y be a connected graph. Then nAY) has finite rank

iff there is an orientation A such that vw is finite for ail v and vo> * 1 for all
but finitely many v.

Proof Suppose that TT^T) has finite rank. Let T be a maximal tree
in T. By Corollary 1 to Theorem 17, there are only finitely many edges not in T.
Take any vertex v Q . Let AQ consist of those edges of T such that
the irreducible path in T from eo to v^ begins with e. It is easy to see that AQ
is an orientation of T such that VQW - 0 and vw-1 for v^v^ (this orientation,
called the orientation of T inwards towards v Q , will also be used in section 8.5).
Enlarge A^ to an orientation A of Y by orienting the edges not in T arbitrarily.
Clearly A satisfies the required conditions.
Now suppose that there is an orientation A of r satisfying the
conditions. Removing finitely many edge-pairs, we obtain a graph A such that
vo)£l for every vertex v (in the orientation Anb of A). Since Y is connected, and
we remove only finitely many edge-pairs, it follows from Exercise 1 that A has
only finitely many components.
It is enough to show that each component of A has finitely
generated fundamental group. For we may then take a maximal tree in each
component which, by Corollary 1 to Theorem 17, contains all but finitely many

edges in that component. The union of these is a forest in A containing all but
finitely many edges of A. This forest can be enlarged to a maximal tree of T,
which will contain all but finitely many edges of r. Then 7Tj(O will be finitely
generated, using Corollary 1 to Theorem 17 again.
So let Aj be a component of A. If Aj is not a tree it contains a
circuit e.,...,e . We may assume that en* A (if not, replace the circuit by the
circuit F i,... 3e^ ye~n). Since eno has outdegree at most 1, we see that e~n^4 A9

and so e _-.€A By induction on n-r, we find that e^A for all r.

Now suppose that Aj has a vertex v such that vo - 0. Let w be any
vertex, and let JC^,...,JC be the irreducible path from w to v. Since vw-O, we
cannot have xm e A, and so xm e A. As in the previous paragraph, we find that
x- *A for all /. We also find, since no vertex has outdegree > 1, that the only
irreducible paths starting at w in which all edges are in A are the paths
JCJ , . . . , * £ for 0<k<m; in particular, there is a bound to the lengths of such
paths.
It follows that if Aj contains a vertex of outdegree 0 then it cannot
contain any circuit, since a circuit e«9...9e with all e.eA gives rise to
arbitrarily long irreducible paths starting at e^a and with all edges in A; namely,
the paths (e^,... ,en)p for all integers p.
Thus if Aj contains a vertex of outdegree 0 then it is a tree. Also,
if Aj contains a circuit, then we may remove an edge-pair occurring in the
circuit, and we get a connected graph with a vertex whose outdegree is now 0;
that is, A. is obtained from a tree by adding just one edge-pair, and so 7ij(Aj) is
infinite cyclic in this case.//

Let F be a finitely generated free group with basis S9 and let T) be

an endomorphism of F. We construct a graph T, together with a function from
E(T) to SuS~l (called the labelling of the edges) as follows. V(T) is to be F.
There is to be an edge from v to w with label xtSuS iff w- x~ V(JCTJ). When
this happens, we also have an edge labelled x from w to v (since
v - (JC ) W(JC T)) ), and this is to be the inverse of the first edge. T need not be
connected, and we let P.. be the component of 1.

Lemma 23 With the above notation, Fixy\ is isomorphic to 7^(1^).

Proof First observe that there is a homomorphism from the

fundamental groupoid of Tj into F which sends each edge into its label. This
induces a homomorphism from ^1(^1 >1) into F.

Notice that there is only one edge starting (or ending) at a given
vertex and with a given label. Also, if e has label x then e~ has label JC . It
follows that the image of an irreducible path is an irreducible word of the same
length. In particular, the homomorphism from rc^I^ ,1) to F is a monomorphism.
To any JCj,... ,JC^ there is a unique path starting at 1 whose edges
have labels x^9...,x respectively. This path ends at v, where v-w (wr\) and
w-JCJ .. .JC^. In particular, this path is a loop iff W-WTJ. Thus the image of
TCjCrpl) is FixT).//

Theorem 24 Let i\ be an endomorphism of a finitely generated free

group F. Then Fixr\ is finitely generated.

Proof With the above notation, it is enough to find an orientation

A of r satisfying the hypotheses of Lemma 22.
Let e be an edge, and let e start at v, end at w, and have label x,
so that w x v(xr\). Suppose that v does not cancel completely in forming the
product V(JCT)) and that w does not cancel completely in forming the product
w(x~ 7)). Then we put e in A iff v begins with JC. For the moment we do not
say whether or not e is in A if the conditions on cancellation are not satisfied.
It is clear that there is at most one edge in A beginning at a given
vertex v. Since w(x~ r\)m x" v, and we are assuming that w does not cancel
completely in forming w(x r\), the first letter of w will be the first letter of
JC" v. That is, it will be JC" iff v does not begin with JC. Hence, when the
conditions on cancellation are satisfied, exactly one of e and e~ is in A.
Now, to each JC, there are only finitely many choices of v such that
v cancels completely in V(JCT)), namely those v such that v is an initial segment
of JCTJ. Similarly, there are only finitely many choices of w such that w cancels
completely in w(x~ TJ). Since F is finitely generated, it follows that there are
finitely many edge-pairs not previously considered. We make an arbitrary choice
of which edge of the pair to put in A for these pairs.
Plainly, if v is not one of the finitely many vertices where the
cancellation condition is not satisfied then vw £ 1, as required.//

We can extend this theorem, but we need a result from Chapter 8.

Let H be a (finitely generated) subgroup of a free group F. All
mention of words, cancellation, reduction, and so on, wil be with respect to a
fixed basis of F, which we do not need to name. It is shown in section 8.2 that
H is free, with a basis S satisfying the following condition. To any JC e SuS

there are words X(JC), [L(X), and p(jc) such that \L(X) has length 1, X(JC )-p(jc)~ ,
\L(X~ ) - \JL(X)~ , and p(x ) - \{x)~ , and x - \{x) {L{X) p{x), reduced as written, and
such that, for any y*SvS~ with y*x, when cancellation is performed on y x it
does not reach u(x) or u(j;)~ .

Proposition 25 Let H be a finitely generated subgroup of a free

group F. Let cp://-» F be a homomorphism. Then (we H; w<p - w) is finitely
generated.
Corollary Let F be a finitely generated free group, let $ and r\ be
endomorphisms of F such that u is one-one. Then (we F; w&m wr\] is finitely
generated.

Proof The corollary follows from applying the proposition to the

homomorphism r\ & on Fr\.
To prove the proposition, we proceed much as before. This time we
take S to be a basis of H with the special property described above. As before,
we take a graph T whose vertex set is F, and whose edges have labels in
S u S~ , there being an edge from v to w with label x iff w- JC" v(jc<p). As before,
{w; wtp - w] is isomorphic to njr*), and we look for an orientation of T
satisfying the conditions of Lemma 22.
If v cancels completely in forming JC" v(jc<p) then there are words p
and q such that v - pq, reduced as written, with p being a final segment of JC
and a being an initial segment of jeep. Thus there are only finitely many such
v.
Let e be an edge from v to w with label JC. Suppose that v does
not cancel completely in JC v(jccp) and w does not cancel completely in
xw(x (p). Then we put e in A iff in JC v at least (IX(JC) X(JC)) cancels from
JC" , or, equivalently, if v begins with X(JC)U(JC).
If J>*JC, by our choice of *S\ cancellation in y x does not reach
[i(y) or ti(jc). Since both JC and v begin with X(JC)U(JC), it follows that
cancellation in y v does not reach [i{y) . We see that at most one edge starting
at v can lie in A.
We complete the proof as before, once we have shown that, when v
and w satisfy the conditions on cancellation, exactly one of e and e~ is in A.
Suppose that ezA, so that v- X(JC)U(JC) q, reduced as written, for
some word q. Since v does not cancel completely in JC v(jccp), it follows that q
cannot cancel completely in v(jccp), and so the reduced form of JCW- v(jc<p) must
begin with X(JC)U(JC), and the reduced form of X(JC) JCW must begin with U(JC). If

we could write w as p(jc) U(JC) U, reduced as written, then u could not begin
with \±(x). Since u is the reduced form of \{x) xw, we get a contradiction. Thus
w does not begin with p(x)~ [i(x)~ , and so e~ is not in A.
Suppose that e 4 A. Then the reduced form of x~ v will begin with
p(jc)" {L(X)~ (since the cancellation cannot go further than X(JC)~ ). Since v does
not cancel completely in x~ v(xcp), it follows that the reduced form of
w-JC" v(x(p) also begins with p(x)~ u(x)~ , which tells us that e~ * A.I I

It has recently been proved by Bestvina (unpublished at present)

that if a is an automorphism of a free group F of finite rank then Fixa has rank
at most the rank of F. We shall not prove this. But we will give a bound for
the rank of Fixi) which can be calculated easily from the endomorphism rj, and
which sometimes gives a better bound.
We take an arbitrary set B of edges of a graph F. As before, we
define the outdegree vo of a vertex v to be \{e € B\ eam v}l. We call an edge-pair
{e,e~) repelling if neither e nor e~ is in B, ordinary if exactly one of e and e~ is
in B, and attracting if both e and e~ are in B. (To explain the wording, consider
a picture for F, in which as usual one edge in the picture corresponds to an
edge-pair. Place an arrow near the vertex eo of the edge, pointing away from eo
if e e B and towards eo if e 4 B. There will be another arrow near ex - e~o. Then
an attracting edge has its arrows •* «-, an ordinary edge has arrows -* *• or
-« «- , and a repelling edge has its arrows -« »— .) A ray is defined to
be an infinite sequence of edges e^ ^ » • • • all *n B such that e-x - e^o for all /,
and e-o^ea for /^y. Two rays are called equivalent if they become the same
after deleting a finite number of edges from each.
Let r and a denote the number of repelling and attracting
edge-pairs, let c be the number of components, and let n be the number of
equivalence classes of rays. Let dmZd-, where d- is the rank of n^(F-)9 where
the F. are the components of P. Notice that if vo is finite for all v and vw^l for
all but finitely many v then the sum Z(va)-l) is well-defined (It may be -co, but
cannot be oo).

Lemma 26 Suppose that vw is finite for all v and vo £ 1 for all but
finitely many v. Suppose also that r and c are finite. Then
r+ c + Z(vd) -1) - d+a + n.
In particular, if T is connected and vw £ 1 for alt v and Vjd) - 0 for some Vj then
d^r. More generally, if vw^l for all v and Vj(o-0 for some Vj, and T^ is the
component of v, then ^(Fj) has rank at most r.

Proof The particular case is immediate from the formula, since we

then have c-1 and I(vo-1)^-1. The more general case then follows, since we
^ ( i p will have rank at most the number of repelling edge-pairs in H. , which is
certainly at most r.
Let us remove an edge-pair {e,e~) from T. Then Z(vo)-l) is
decreased by 0, 1, or 2 when the edge-pair is repelling, ordinary, or attracting,
respectively. Also n remains unchanged.
If the edge-pair lies in a circuit of a component A of T, then, by
Exercise 1, A - {e,e~} is connected, so c remains unchanged. Any maximal tree of
h-{e,e~) will be a maximal tree of A, and Corollary 1 to Theorem 17 shows us
that d is decreased by 1. If the edge-pair is in A but does not lie in a circuit of
A, then Exercise 1 tells us that A - ie,e~) has two components A* and A", and so c
increases by 1. If 7*'and 7™ are maximal trees in A* and A" then T'vT"u{e,e~} is
a maximal tree in A (using Lemma 6), and Corollary 1 to Theorem 17 shows us
that d remains unchanged.
It follows that removal of an edge-pair does not alter the difference
between the two sides. We may therefore remove the finitely many repelling
pairs, and the finitely many edges at those vertices v with vo > 1, and prove the
result for the graph obtained. That is, we may assume that V has no repelling
edges, and that vw^l for all v.
We now show that under these circumstances there can be at most
one attracting edge-pair in each component. For if there were two we could take
an irreducible path e^,... ,en joining a vertex of one attracting edge-pair to a
vertex of the other. If e+ were not in the attracting edge-pair, we could add one
edge of the pair at the beginning of this path, and similarly for the end. Thus
we may assum that e* and e are attracting. Then e is the only edge in B
starting at eô, since every vertex has outdegree ^ 1. Since there are no repelling
edges and e~n î B, we find that en j£ B. Inductively, we see that e . e i? for all
/. In particular, e~ *• B. We get a contradiction, since both e~ and ej" are in B
and both begin at eô>
It follows that there are only finitely many attracting edge-pairs. As
before, we can remove them all, so we may assume that all the edge-pairs are
ordinary.
It is now enough to show that, when all edge-pairs are ordinary and
VG) £ 1 for all v, if r is connected then there is either exactly one circuit or
exactly one vertex of outdegree 0 or exactly one equivalence class of rays, and
that no two of these can hold simultaneously. The proof of Lemma 22 shows
that if there is a vertex of outdegree 0 then there can be no circuits and no
rays.

If there is a circuit, the proof of Lemma 22 shows that we may

remove an edge-pair, and reduce to the case where there is a vertex of outdegree
0.
If v is a vertex of outdegree 0, the proof of Lemma 22 shows that,
for any vertex w, the irreducible path from w to v has all its edges in B. In
particular, if w^v then w has outdegree * 1. So there is at most one vertex of
outdegree 0.
Now suppose that there are no circuits, and no vertices of outdegree
0, and so all vertices have outdegree 1. It follows immediately that any path
£ j , . . . ,en all of whose edges are in B can be extended by a further edge e + ^.

Since there are no circuits, this gives rise to a ray, e^e^,... . Let w be any
vertex, and let x^,... ,JC be the shortest irreducible path from w to a point of
the ray. Let this path end at ena. Since en is the only edge in B beginning at
eno, and all edges are ordinary, we see that xm* B, and then, inductively, that
jt.ei? for all j . It is then clear, since all vertices have outdegree 1, that the only
ray beginning at w is x^,...,xm ^n^n +j»• • • » anc * so a
^ ravs are
equivalent.//

We take the same graph as in Theorem 24. Let e be an edge from

v to w with label x. Then e is to be in B iff v begins with JC. Thus 1 has
outdegree 0, and all other vertices have outdegree 1. So the lemma applies, and
we have only to find a way of calculating the number of repelling edge-pairs.
Take any edge-pair. Let the edge in the pair whose label is s e S
s l)) start iat v and end at w. If v begins with s then the edge-pair is
(rather than s~
attracting or ordinary.
Suppose that v does not begin with s. Then w - s V(ST)). Since v
does not begin with 5, either w begins with 5" , and so the edge-pair is ordinary,
or else 5 v cancels completely in reducing 5" v{sr\). In the latter case we find
that 5T) - v 5w, reduced as written, and so w does not begin with s . In this case
the edge-pair is repelling. Conversely, if there are words v and w such that
sr\m v 5vv, reduced as written, then there is a repelling edge-pair joining v and w.
Thus the number of repelling edge-pairs, which is an upper bound
for the rank of Fixr), equals the total number of occurrences for all S£ S of 5
(not s" ) in sr\.

Examples In F{a,b), let ai\ - ba and bi\ - bba. Then a occurs once in
<2T), and b occurs twice in br\. Thus Fixr) has rank at most 1 + 2.
In F(a,b,c)t let a<p - baca b ac , b(p - a ba,
c*(p - baca b~ aca bac a~ b . Then a occurs twice in a<p, b occurs once in btp,

and c occurs twice in ccp. Thus Fix<p has rank at most 2 + 1 + 2. We shall analyse
this example more closely in a moment, and we will show that Fixep-Cl}.

We do not need to count all the repelling edge-pairs in F, but only

those in I\ . These correspond to the expressions of sr\ as v sw with v € r 1 . As
-1
we have already seen, veF« iff there is some u such that vmu (ur\).
We cannot expect in general to find out whether or not a given v
satisfies this condition. But it may well happen that we know enough about r\ to
find a necessary condition for this to occur.
For instance, r\ always induces an endomorphism of the free abelian
group FIF\ If this induced endomorphism is the identity then v-u~\ur\) only if
v - 1 in F/F\ which we can check. In particular, in our second example cp does
induce the identity on FlF\ and none of the v which begin a repelling edge are
in T j . Hence Fix<p is trivial.

Exercise 4 Show that the attracting edge-pairs correspond to

occurences of s in sr\, and that if srj- ps~ a, reduced as written, then the
corresponding edge-pair is in Fj iff sp e Fj.

5.4
COVERINGS OF COMPLEXES
The notion of a covering of a complex is remarkably tricky to
define. The definition in several books is wrong; that is, the authors claim that
certain properties follow from their definitions, but there are examples showing
that their claims are false. Even knowing this, the definition 1 gave in an earlier
version of this book was also wrong. I think that I have dealt with all the
problems in the definitions I give below.
The difficulties arise because the boundary of a cell is a loop
which may be a proper power of some other loop. Because of this, we define
some special classes of complexes.
The complex K is called powerless if there is no cell c such that
k.
cd can be written as a proper power (e^,... ,e^ with k>\. We say the complex
K has simple boundaries if, for every cell c, the boundary cd is a circuit.
The complex AT is a simplicial complex if it satisfies the following
conditions: (i) no edge is a circle, (ii) for any two vertices v and w there is at
most one edge e with eom v and ex - w, (iii) the boundary of each cell is a loop
£j ,^2 > e2 w i t h t n r e e distinct edges, and (iv) for any three distinct edges e^, ^ >
and e^ such that e^ e^ ,e^ is a loop there is at most one cell with this loop as
boundary.

Plainly a simplicial complex has simple boundaries, and a complex

with simple boundaries is powerless.
Let cp:AT-» L be a map. Then K will be powerless if L is powerless,
k k
since if cd were (e^,... ,^1 with A > 1 we would have (ctp)d - cdcp - (JCJ , . . . ,jcp ,
where JC^. is e-<p. Also AT will have simple boundaries if L has simple boundaries.
If L is a simplicial complex then K has no circles, and each cd is of the
required form, but the uniqueness properties required for a simpliciai complex
need not hold for K.
In Theorem 18 we constructed for each presentation of a group G a
complex whose fundamental group is G. This complex will be powerless iff no
relator is a proper power. To any presentation <A";/?> of G there is another
presentation which has this property, namely <X,y ; [yr for all nR)yy>, Hence, for
any group G there is a powerless complex whose fundamental group is G.
Further, if L is a powerless complex then the complex containing L which is
constructed in Theorem 19 can also be taken to be powerless.
For any group G there is a complex with simple boundary, and
even a simplicial complex, whose fundamental group is G. However, this result is
harder than the previous ones. It will be shown in the next section that there is
a process of subdivision of a complex which yields a simplicial complex with
the same fundamental group.
We can now define coverings. We start with an easy case, and
become progressively more general.

Definition Let K and £ be connected graphs, and let p:£ -> K be a

map. Then p is called a covering map and R is called a covering graph of K if,
for every vertex v' of ft, p induces a bijection from the set of edges of £
starting at v' to the set of edges of K starting at vp.

Examples Let K be the complex with one vertex v and one

edge-pair {e,e~). Let £ have one vertex vn and one edge-pair ^nfe~n) for each
n € Z, with e o - vn and e^ - en + j . The obvious map is a covering map. Note that
in K both e and e~ start at v, while in & the edges starting at v are e and
F
n-V
Another covering L, of K is given by the complex with vertices v
and edge-pairs {en,e~n) for l^n^k. Here e^m v^ for all n, while e^ - v^ + j for
n < k and e^x - Vj.

K K L6
Figure 1

Lemma 27 Let p\f?•* K be a covering of graphs. Then £p- K.

Proof The condition ensures that p maps E(f?) onto E(K) provided
that it maps V(£) onto V(K). By Lemma 1 it is enough to show that if V£V(£)p
and e is an edge with eom v then eiiV{£)p. We need only take some v* with
vp- v, and then take an edge e with eo - v' and epm eJI

Warning The examples above show that a covering p need not be

one-one on the set of edges having v* as a vertex. This situation can occur when
K has a circle with vertex vp.

Definition Let K and R be connected complexes with simple

boundaries. Then the map p :^-> K is a covering map and fit is a covering
complex of K if, for each vertex v' of XT, p induces a bijection from the set of
edges of F? starting at v" to the set of edges of K starting at vp and also
induces a bijection from the set of cells of £ with v' as a vertex to the set of
cells of K with vp as a vertex.

Plainly, if AT is a graph then this definition coincides with the previous

one. It is also easy to check that if AT is a simplicial complex then any covering
complex is a simplicial complex. Lemma 27 and the hypothesis on cells
immediately tell us that a covering map is onto.
When K does not have simple boundaries we need to look at a slightly
more delicate notion.
Let c be any cell of a complex AT, and choose a specific loop
# j , . . . ,en of the cyclic loop cd. With the cell c we associate n objects which we
call indexed cells, which we denote by (c,l),..., (c,/z). The boundary of the
indexed ceil (c,/) is to be the loop (not the cyclic loop) e.9.. .fe fe^t... 9e^9
and we define the starting vertex of (c,z) to be e^a. The inverse of the indexed

cell (c,/) is to be the indexed cell (c,« + 2 - / ) , where we regard n+l as 1. When
we take the boundary of c~ to be the obvious choice, namely e~n,...ye~^y the
boundary of this indexed cell is the inverse of the boundary of (c,i). Note that
the indexing of cells is not unique, since we begin by a choice of loop in each
cyclic loop cd. However, we shall always assume that some choice has been
made in all the complexes we consider.
When K has simple boundaries the indexed cell (c,i) is determined by
specifying the cell c and the vertex which is the starting vertex of the indexed
cell. When K does not have simple boundaries there will always be a cell c and
two distinct indexed cells (c,z) and (c,y) with the same starting vertex. When K
is powerless the indexed cell (c,z) is determined by specifying the cell c and the
boundary of (c,0, but this is not enough in general. For let us write cd as
k.
{e^y...,er) with k as large as possible. Then it is easy to check that the loops
e
i>- -">el-\ a n c l e • , . . . , e - _ p which are the boundaries of (c,i) and (cj), are the
same iff j - i is a multiple of r.
Let <p:L •* K be a map of powerless complexes. Let c be a cell of L,
and let the indexed cell (c,l) have boundary e^,...,en. Then the loop
£ j 9 , . . . ,en<p of K is the boundary of some indexed cell corresponding to c<p.
Suppose this indexed cell is (c, j). Then we have an induced map of indexed
cells sending (c,/) to (c<p,i+J-l) (where a number greater than n is to be reduced
mod «), with the boundary of the image being the image of the boundary. This
induced map will send an indexed cell and its inverse to an indexed cell and its
inverse.

Definition Let p:t?-* K be a map of connected powerless complexes.

Then p is a covering map and R is a covering complex of K if, for each vertex
v of R', p induces a bijection from the set of edges of F? starting at v to the
set of edges of K starting at vp and also induces a bijection from the set of
indexed cells of I? starting at v' to the set of indexed cells of K starting at vp.

Again, coverings are onto. Also, this notion coincides with the previous
one when the complex K has simple boundary, in which case we know that K
also has simple boundary.

Example Let K have one vertex v, two edge-pairs {*, ST} and {y9 y~}9
and one cell-pair {c,c~} with cd being JC,J;. Thus K has two indexed cells
associated with c, namely (c,l) with boundary x,y and (c,2) with boundary yy x;
both these cells have starting vertex v.

Let have vertices v , edge-pairs ^xn>xn) and {yn,yn), and cell-pairs

icn,cn) for all /icZ. Let V 3 " W and y - v f l + f V and let c^d be J C ^ .
Let p be the obvious map. Then p does not give a bijection of cells with vertex
vn to cells with vertex v. However, the indexed cells corresponding to cn are
with starting vertex vn and (^,2) with starting vertex . Thus p maps
the indexed cells (cn>\) and ), which both start at v^, onto the indexed
cells (c,l) and (c,2). We see that p is a covering map.

(part of) K
Figure 2
Finally, we must look at general complexes. We see that a map
<p:L -» K does not determine a unique map on indexed cells. Suppose that the
boundary of c\, and that the image
under 9 of the boundary of the indexed cell (c,l) is the boundary of the indexed
cell (cj). Then we can map the indexed cells (c,/) to the indexed cells
(ccp,y +/-1), as before. But we are also permitted to take any integer m, and then
map (c,z) to (C9,mr + / + /-1).
Consequently, we have to make a choice, which we call 9,, of maps of
indexed cells corresponding to cp. We may make the choice so that 9, sends an
indexed cell and its inverse to an indexed cell and its inverse. Note that if
ty:M->L is another map then ^f9t iS a possible choice for (^9) t ; because of the
non-uniqueness, we cannot just say that (4><p)f is t|>,<p, • However, if our complexes
are powerless, so that only one choice is possible (once we have specified the
indexing of the ceils) , then we do know that 4jf9t

Definition Let p:A^-» K be a map of connected complexes. Then p is

a covering map and XT is a covering complex of K if there is a choice of pt
such that, for every vertex v of ^ , p induces a bijection of the set of edges of
AT starting at v* to the set of edges of K starting at vp and pf induces a
bijection of the set of indexed cells of £ starting at v* to the set of indexed
cells of K starting at vp.

This coincides with the previous definition when K has simple

boundaries, in which case f? also has simple boundaries.
As before, covering maps are onto.
It would be more precise to refer to the pair (p, p t ) as a covering map.
The example below shows that one choice of p, may satisfy the required
property, while another choice may not. However, we shall not need to pay
much attention to p f , so the simpler notation is preferable.

Example Let K have one vertex v, one edge-pair ie,F), and one
two
cell-pair {c,c~} with cd being e, e. Let F? have two vertices v^ and v^»
anci t w o c e airs c anc C ; et
edge-pairs {e^ ,e~^) and l ^ , ^ ^ ^ P ^ j »^j) * ^ 2 »^V *
and e l V e anci let c and C d bot be e
e^o- Vj - e^t \" 2 " 2 °' l^ 2 ^ \ *e2 ' ^ e t ^ b e t ie
^
obvious map.
Define the indexed cells of R so that (c^,l) and (C2»2) have
boundary e^, ^2» while (Cj ,2) and (C2»l) have boundary ^ »^j • Let Pj map (c^ ,1)
to (c,l). We may choose pf either to map (C2 ,1) to (c,l) or to map (c^ ,1) to.
(c,2); in the latter case it maps ^ ,2) to (c,l). It is easy to check that the first
choice of pf gives a covering map, while the second does not.
Let L be a complex with two vertices Wj and w^, two edge-pairs
{jCj ,JC~J} and ( J C J , * ^ ) , and one cell-pair {J,J"}; let JtjO - Wj - .X^T and
jCjT - W2 - JC2O, and let dd be Xj ,x^ • With AT and XT as above, there are two
obvious maps from L to A*', the first, 9, sending d to Cj and the second, ty,
sending d to c^ • Plainly cpp-ijjp. However, with (tf,l) being the indexed cell
starting at Wj , we must define <p, to map (dy\) to (c^ ,1), This explains the
conditions in the next proposition. It is examples like this one which need the
extreme care in our definitions, and which show the inadequacies of many other
suggested definitions.

Proposition 28 Let p.K'-» K be a covering. Let L be a connected

complex, and let <P,IJJ:Z,-»^ be maps such that <ppmtyp, and such that there is
a vertex v with v; (///) /'/ we can choose 9, and ipf so that
<PIPI " ty\P\ then 9, - (J>t, and so 9 - ty.

Proof (i) We first show that 9 - 4> on V(L). By Lemma 1, it is

enough to show that if e is an edge of L such that ea, and they have the same image under p . By the definition of a

covering, this implies that e<p - ety. It follows that ei<p - eity (because et<p - ecpx,
etc.), as required. The same argument now tells us that eyety for any edge e.
(ii) This can be proved directly, but we may as well use part (iii).
Because our complexes are powerless, we know that <p,p, - (<pp)j - (^p)| - il^p,.
(iii) Let (c,i) be an indexed cell of L, and let its starting vertex be
v. Then (c,/)<p, and (c,/)ify have the same image under />,. Because cp - ty on the
vertices, they both have the starting vertex v<p. The definition of a covering now
tells us that (c,/)<p, - (c,/)<J>t, as required.//

Proposition 29 Let p:f?'-» K be a covering. Let f be a path of K

starting at v. Then, for any vertex v of R such that vp - v, there is a unique
path J of f? starting at v and such that Jp - / .

Proof Let / b e e^,... ,e n. Suppose that for some r we have found

uniquely edges e^,... ,er of R such that e^,... ,er is a path of £ starting at v"
and such that e^p is e^ for i ^r. Then e^xp-e^T, and so there is a unique edge
e +| starting 2X e \ and with e .pm e +« . The result follows by induction.//

Proposition 30 Let p:ft'-» AT be a covering. Let / and g be paths

of R starting at the same point and such that Jp *• gp. Then J <* g.

Proof It is enough to show that if gp comes from fp by an

elementary reduction then g comes from / by elementary reduction. Let / be
e ' j , . . . ,e , and let e be e p for r^n.
Suppose that e-+^m e~-, and that gp comes from Jp by deletion of
e e
j* i+l- Then ^ / + j ' ^ / » since both start at e *L and they have the same images
under p. Then removal of ^ » ^ - + | ls a n elementary reduction, and the path
obtained must be g, by the previous proposition, since both it and g start at the
start of / , and they evidently have the same image under p.
Suppose that gp comes from Jp by removal of the loop e.,... ,e -,
where this loop is the boundary of a cell c. For convenience, assume that c is
indexed so that this loop is the boundary of the indexed cell (c,l). Then there is
a cell c of £ one of whose indexed cells, say {c, k), starts at e'.a and maps to
(c,l). T h e n { c \ k ) d p - { c , k ) p } d - ( c , \ ) d is the loop e • , . . . , e - . Since the path
e•,...,e- also starts at e'-o and maps to this loop by p, the previous proposition
tells us that (c , k)d is e.,... ,e .. Hence deletion of these edges is an elementary
reduction of / . The result of this deletion must be gt since they both start at
the start of J and they have the same image under p.11

5.5 SUBDIVISIONS
In this section we will show how, given a complex K, we can
construct a simplicial complex related to K with the same fundamental group as
K. In the course of the construction we will obtain other complexes with the
same fundamental group as AT, one of which is powerless and the other of which
has simple boundaries. This material is not needed for the later development of
the theory. But it shows that we may use the simpler approaches to coverings of
complexes if we wish, without running into any problems in the applications.
Let AT be a connected complex. Its edge subdivision fif is the
complex defined as follows. The vertices of K? are the vertices of K together
with one new vertex for each edge-pair {e.e") of K. This new vertex will be
denoted by v or by v - . There are two edges of K6 to each edge e of K. These
are denoted by e_ and e+. The edge e_ starts at ea and ends at v e , while the
edge e+ starts at ve and ends at ex. We define e~^ and ~e~ to be e~+ and e~_ ,
resepctively. The cells of K? are the cells of K. The boundary in K? of a cell is
obtained from its boundary in K by replacing each edge e by the two edges

The cell-subdivision K° of K is defined as follows. The vertices of

K° are the vertices of K together with one new vertex for each cell-pair. It is
most convenient to take a set C^ of cells which contains exactly one cell from
each pair, and to let the new vertices be v, for all CECQ. The edges of K? are
the edges of K together with new edges. Let c e CQ have as boundary the loop
# j , . . . yen. Then there are to be n new edges fc- for \<i± tf-with fc-o - vQ and
/ ci .T-<?.O,
i
together with the inverses f ci. of these. The cells of K° are all to be
new cells. Corresponding to the cell c e CQ there are n new cell-pairs ^ c / » ^ c / ^
for \<^i^n. The boundary of d^ is to be fc-, e-> f c [+\-

Example In Figure 3 below, K has four vertices, four edge-pairs,

and one cell-pair. Then K? has eight vertices, eight edge-pairs, and one cell-pair,
while K° has five vertices, eight edge-pairs, and four cell-pairs. Finally
has nine vertices, sixteen edge-pairs, and eight cell-pairs.

K
Figure 3.

It is easy to check the following facts. A^ has no circles. If K has

no circles then K6 has at most one edge with a given start and finish. If K is
powerless then tf is powerless. If K has simple boundaries then tf has simple
boundaries.
Also, if K has no circles then K° has no circles. If K has no circles
then K? has simple boundaries. If K has at most one edge with a given start and
finish and K has simple boundaries then K° has at most one edge with a given
start and finish.

The barycentric subdivision AT* of K is defined to be (A^)c. Then

K* has simple boundaries, and K" is a simplicial complex.
Proposition 31 7Tj(^) is isomorphic to TT^/O for any connected K.

Proof Let T be a maximal tree in K, and let EQ£ E consist of

exactly one edge from each edge-pair not in T. Let S be the subgraph of K?
consisting of all vertices of A^, and whose edges are e_ and e+ for all e e T and
e and ~ for all e e EQ . Note that if e e T then e~ e T, and so ~ and e^~, being
e~+ and e~_ respectively, are in S.
We shall show that S is a tree. Choose some vertex a of AT. Let v
be any vertex of K. Then there is a path in T from a to v. Replacing each edge
e by the pair of edges e_, e + in S, we get a path in S from a to v. Now take a
vertex vQ of A^. We may assume thate* E^uT. Then there is a path in S from
a to v^ which first goes from a to eo, and follows this with the edge e_ . Hence
S is connected.
In an irreducible path in A^, the edge (if any) following an edge e_
must be the corresponding e+. It follows that two irreducible paths in S from a
to v must be the same, since they both come from irreducible paths in T by
splitting edges into two, and these two paths in T are the same, as T is a tree.
Now consider two irreducible paths / and g in S from a to the vertex v . If
they both end in e_ or both end in e~_ then, deleting the last edge, we get two
irreducible paths in 5" from a to a vertex of K. We have just seen that these
must be the same. Suppose that / ends in e and g ends in e~_. Then if we
remove e~_ from g we get an irreducible path in 5* from a to ex which does not
end in e+ (as e~ - i^"). If we add e+ to / we get an irreducible path in S from
a to ex which does end in e+. Thus we get two distinct irreducible paths in S
from a to ex. Hence this case cannot happen.
It follows that S is a tree in K?t and it must be a maximal tree,
since it contains every vertex.

Take the presentation of 7ij(/^,a) using the maximal tree S and the
presentation of n^(Kfa) using the maximal tree T. In both presentations, the
relations say that certain generators are trivial. If we delete these generators, the
resulting presentations of the two groups become identical.//

Proposition 32 n^K0) is isomorphic to n^{K) for any connected K.

Proof Let T be a maximal tree of K. Let S be the subgraph of K?

conisting of all the vertices of K, all the edges of T, and all the edges / c j and
f^ for CECQ. Because each vertex v^ is a vertex of exactly one edge-pair in S,
it is easy to check that S is a tree. As S contains every vertex, it is a maximal
tree in K°.
The presentation of K^K0) using this tree is the following. It has
g e n e r a t o r s e f o r a l l ezE t o g e t h e r w i t h t h e f c- for C I C Q ( w e d o n o t need f .f
m
w h i c h w e j u s t r e g a r d a s / „ , - ) . T h e r e l a t i o n s a r e ee~-l f o r a l l e*E and e \ for
ci
all e € T, together with further relations for each ciC^. If cd is e^,... ,e then
there are n+\ relations corresponding to this c. They are / c j " 1, fciei~?c /+1 ^ o r
0<i<n, and fcnenmfc\- Tietze transformations replace these with / c j " 1.
/ .- e* . . . e- , for 1 < / < n, and A M e " A i , and a further Tietze transformation
Cl 1 Z~l CrZ At Ci

lets us replace the last of these with the relation e^... enm\. Now Tietze
transformations let us delete every fc-. The result is just the presentation of
71^ using the tree T.I I
Let tp:AT-» L be a map. We can define a map <$e\K? •* Le such that
<p^ corresponds to <p under the isomorphism of Proposition 31. We can also
define a map <pc:A^'-» L c such that <pc corresponds to 9 under the isomorphism of
proposition 32. While <pe is uniquely given by 9, there are choices for <pc, but it
is uniquely given by 9 and 9, .The details are left to the reader.
The constructions can easily be modified so that we do not
subdivide all edges or cells, but just choose some of them to subdivide. In
particular, suppose that L is a subcomplex of K, and that L is a simpliciai
complex. Then, by subdividing the remaining edges and cells (twice) we obtain
a simpliciai complex M which contains L (rather than containing a subdivision
of L) and such that the inclusion of L in M and the inclusion of L in K induce
the same homomorphisms of fundamental groups.

5.6 GEOMETRIC REALISATIONS

In this section, we will associate a space IAH with each complex K
in such a way that TTÎAH) is isomorphic to TTJ(AT). When K is the complex
associated with the group G by Theorem 18 then IAH will be the space associated
with G by Theorem 4.12. The results of this section are not needed elsewhere.
But I feel it is worth noting that the analogy between the theory of fundamental
groups of complexes and the theory of fundamental groups of spaces is more
than merely formal.
We begin with a graph F, and take a set EQ of edges which
contains exactly one member of each edge-pair. We can regard the set V of
vertices as a discrete space. For each e e EQ take a copy / of the unit interval
/, these copies being disjoint. We attach U/ to V by the map which sends 0
and 1 in / to ea and ex in V. The resulting space is defined to be the
geometric realisation IFI of F.
Let / be the edge-path e^9... ,en of F. We associate with it a
standard path 1/1 in IFI as follows. If e-eE^ then 1/1 maps the interval
L(i-l)/n, i/ri] linearly onto the image of / in IFI, sending (i-\)/n to 0 and i/n to
1. If eÊ^ then e-Ê^; In this case I/I maps C(/-l)//i, i/nl linearly to the image
of / - . , this time sending (i-\)/n to 1 and i/n to 0. If / is an edge-loop
then I/I is a standard loop. In this case we can, as usual, regard 1/1 as a map
from 5 1 to F.
Now let AT be a complex. Take a set CQ of cells, containing exactly
one cell-from each pair. We have already constructed the space \K I
corresponding to the 1-skeleton K of K. For any cell ciC^ choose an edge-loop
(not just a cyclic loop) as its boundary. As we have seen, this gives rise to a
map &c from Sl to I AT11. Take a copy DQ of the disc D2 for each ceCQ. We
now define I AH, the geometric realisation of AT, to be the space obtained from
\K I by attaching UDC using the maps &c. A different choice of the edge loop in
the cyclic loop cd would provide a homeomorphic space.
Let K be the complex associated with the group G by Theorem 18.
It is easy to check that \K\ is the space associated with G by Theorem 4.12.

Proposition 33 TTj(lATI) is isomorphic to n^K) for any connected

complex K.

Proof The direct limit arguments of Theorem 15 and Proposition

4.10 tell us that the result is true in general if it is true when A" is a finite
complex.

Let K be a finite complex. Let c be a cell of AT, and let L be the

complex K- {c,c~}. Then Corollary 4 to Theorem 17 and the corollary to
Proposition 4.11 ensure that TT ^ CI ATI is isomorphic to n^K) provided that 7ij(IZ,l) is
isomorphic to Tij(L). Thus, inductively, it is enough to prove the result when K
is a finite graph.
Let T be a finite graph, and let T be a maximal tree in F. Suppose
F has an edge e not in T. Let A be the graph T-{e,e~). Then A is also a
connected graph. By Corollary 1 to Theorem 17 and the remark after Proposition
4.9, Tij(iri) is isomorphic to TTJ(F) provided that TTJ(IAI) is isomorphic to 7tj(A).
Thus, inductively, it is enough to prove the result when AT is a finite tree T.
We know from Lemma 10 that any finite tree T has a vertex v
belonging to exactly one edge-pair ie,e~}. Let 5 be obtained from T by deleting
the edge-pair {e,e~} and the vertex v. Then 5 is a tree. It is easy to check that
IT! is obtained from 15! by attaching an interval at one end-point. So, by the
example preceding Proposition 4.9, 151 is a strong deformation retract of i n . It
follows that i n is contractible if 151 is contractible. Inductively, we find that i n
is contractible, and so n^(\T\) is trivial. Since we know that iz^T) is trivial, the
result is proved.//

Given a map <p:K-> L we can obtain a map lcpl:IAH -» ILL This map is
not unique, but is uniquely defined by cp and <pt. The homomorphism induced by
9 and l<pl correspond under the isomorphisms of the proposition. Proofs are left
to the reader.
It needs some care to establish the topological conditions needed to
show that Ipl is a covering of spaces when p is a covering of complexes, but
this can be done. In our examples this fact will usually be clear.
For instance, let F be the graph with one vertex and one edge-pair.
Then IF! is 5 . Our two examples of coverings of F give rise to the covering of
5 by IR and the covering of 5 by itself in which the map sends the complex
k
number z to z .
Let K be the complex with one vertex, one edge-pair and one
cell-pair, with c6 being e,e, and let p:ft •* K be the covering discussed. Then
•y 1 2
IA I is 5 , and \K\ is the real projective space RP .

6 COVERINGS OF SPACES AND COMPLEXES

In this chapter we will develop the theory of covering spaces and

covering complexes, and in the next chapter we will use this theory to obtain
properties of groups.
Most of our results will be true both for spaces and for complexes.
The proofs for the two cases will be closely related; the proof for complexes is
often just a simplification of the proof for spaces, and I omit the proof for
complexes when this happens. When proving results for complexes, 1 will give
the general case using the map p,. For all the results that we need, however, it
would be sufficient to consider powerless complexes only. Thus we need never
consider p,, and readers may make this simplification if they wish.
All spaces we consider will be assumed path-connected and locally
path-connected, and all complexes will be assumed connected, unless otherwise
stated. The symbols jt and £ will always denote coverings of the space X and
the complex K, and p will always denote a covering map (from / to X or from
I? to K). Also V* will denote a point of Jt (or vertex of £) such that vp- v,
where v is a point of X (or vertex of K). We will also say that v* lies above v.
Recall (Propositions 4.15 and 5.29) that given any path / (in X or
K) starting at v and some v' lying above v then there is a unique path / in
^ ( o r £) starting at v" such that / p - / . We call / the lift of / starting at v'. We
also showed (Propositions 4.16 and 5.30) that homotopic paths starting at v have
homotopic lifts starting at v"; in particular, these lifts end at the same point.
We showed further (Lemma 4.14) that if two maps p typ then cp — ip. The same holds (by Proposition
5.28) for maps from (L,w) to (£,v) when the complexes are powerless. For
general complexes we need to assume that cp,p, - ifyp,, and we can then deduce
that <pf - <J>f, and so <p - ip.
-1
We will find it convenient to denote the inverse of a path / by / ,
rather than / . Plainly, if / is a lift of / then 7 ' 1 is a lift of f~l.

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:31:58, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.007
Coverings 152

Lemma 1 A loop at v lifts to a loop at v iff it represents an

element of n^()ffv^)p* (or of 7rj(^,vOp* for the case of complexes).

Proof If / - / p , where / is a loop based at v* , then plainly /

represents an element of it AX .v )p*.
Conversely, let / represent at element of Ti^(XJ,vJ)p*. Then there is a
loop g based at v' such that f^gp. Since gp lifts to g, which begins and ends
at v, the lift / of / starting at v' must also end at v .//

Corollary 1 Let f and g be paths from v to wf with lifts / and g

starting at v. Then J and g end at the same point iff f.g represents an
element of KA£,V)p* (or nJI?,v^)p* for the case of complexes).

Proof Suppose that / and g end at the same point. Then we can
form J. g , and this is a loop. The loop /. g is the image of this loop under p,
and so f.g represents an element of n^(^\v)p*.
Conversely, let f.g represent an element of n^(]F ,v)p*. By the
lemma, /. g"1 lifts to a loop h based at v\ Then (f.g).g will lift to h.g. In
particular, the lift K. g of (f. g ). g starting at v* has the same end-point as g.
Since f^(f.g ). g, we know that the end-points of / and h.g are the same.//

Corollary 2 The cardinality of xp is independent of x, where x

is a point of X (a vertex of K).

Remark This cardinality is called the number of sheets of the

covering.

Proof Take a path / from v to x. To each v' above v there is a

unique lift / of / starting at v . The end-point of / will be a point above x.
Thus / defines a function from vp to xp . Similarly, the path / defines a
function from xp to vp" . Let / start at v* and end at x. Then / is a lift of
f which starts at JC and ends at v . Thus the composite of the two maps (in
either order) is the identity (on vp or xp , respectively).//

Corollary 3 There is a bijection from vp to the set of cosets of

Proof To each point w in vp there is at least one path from v* to

w, and the image of this path under p represents an element of KJX,V). TWO
different paths from v" to w give rise to elements of n^(X9v) in the same coset
of n^(JtJtvJ)p*9 by Corollary 1. Conversely, any element of n^(X9v) corresponds to
some w€vp in this way, and two elements in the same coset correspond to the
same point.//

Theorem 2 p* is a monomorphism for any covering p.

Proof Let / be a loop at v' which represents an element a with ap*

the identity element. Then fp is homotopic to the trivial loop at v in X (or K).
Hence / is homotopic to the trivial loop at v9 and so a is the identity.//

Proposition 3 Let vj be another point in vp" . Then n^jt,v^)p* is

conjugate in -K^(X,V) to -n^(]t ,v). Further, any conjugate of 7Uj(^f,v")p* is of the
form -KA]£\V+)P* for a suitable choice of v'j in vp' . The same holds for
complexes.

Proof Let J be a path in )? from v' to vl , and let fm?p. Then /

is a loop at v, which represents an element a of n^{X9v). We know that
nl(X',vl)-'Kl(XJ,v)?*. Hence n^jt.v^p* - (nl(X',v)p*)(?p)*9 and this is the
J
conjugate of n^(jttv )p* by a.
Conversely, given a in n^{X9v), we represent it by a loop / at v.
This loop lifts to a path in if starting at v . The end of this path will be a
point vj in vp' . The argument of the previous paragraph then shows that
7C|(^T,v^)/7* is the conjugate of n^(XJfvJ)p* by a.//

We say that the covering p:}?-* X (of spaces or complexes) belongs

to the subgroup n^(jT,v')p*. Evidently, the subgroup of 7 T J ( A » to which a given
covering belongs is determined only up to conjugacy. We shall see later that
there is essentially at most one covering belonging to a given subgroup. Also,
for sufficiently complicated spaces, some subgroups have no coverings belonging
to them, but there is always a covering of a complex belonging to a given
subgroup. We now investigate under what conditions there is a covering of a
space X belonging to the subgroup H of n^(X,v).
We call H sufficiently large if each x e X has a path-connected
open neighbourhood U such that, for any path / from v to x and any loop g in
U based at JC, the loop f.g.f represents an element of H. Such a neighbourhood

U will be called a basic neighbourhood of JC.

Let U be a basic neighbourhood of JC and let y e £/. Then U is also
a basic neighbourhood of y. For we may choose a path h in U from ^ to JC.
Then, given any path / in X from v to y, and any loop g in U based at j>, the
loop {f.h).(h.g.h).(f.h)~ represents, by definition of a basic neighbourhood of JC,
an element of H. Since this loop is homotopic to f.g.f , we see that U satisfies
the required condition with respect to y. Also, any connected open
neighbourhood of JC contained in U will itself be a basic neighbourhood.
Plainly the trivial subgroup is sufficiently large (and hence every
subgroup is sufficiently large) iff each JC e X has an open neighbourhood U such
that 7TJ(£/,JC)/* is trivial, where i:U->X is the inclusion. A space with this
property is s o m e t i m e s called semi-locally simply connected (if nJU,x) were
itself trivial, we would refer to local simple connectedness). In particular, X is
semi-locally simply connected if it is locally contractible; that is, if each point
has a contractible neighbourhood.
The geometric realisation I ATI of a complex K is locally contractible.
This follows from the next lemma, going from the discrete space V{K) to the
space IAT1!, and then to \K\.

Lemma 4 Let X be a locally contractible space, and let Y be

obtained from X by attaching disks of various dimensions. Then Y is locally
contractible.

Proof Let the disks D. be attached by maps f.. Any point of Y- X

lies in the interior of some D^ and so has a contractible neighbourhood.
Let JC e X, and let U be a contractible neighbourhood of JC in X.
For each /, the set Uf[ is an open subset of the boundary (rt-l)-sphere S. of
the «-disk: Dr
We can regard D- as consisting of all points ta, where atS- and
1
re [0,11 Let V. be Ua\ re (0,1] and azUf-' }. Then V. has a strong deformation
1
_i t i

retraction onto Uf^ . Now the set W- UvU{V-; all /} is an open neighbourhood
of JC in Y. By Proposition 2.4 W* I is an identification space of the disjoint
union of U*I and the sets V.xl. Thus the maps from V-*I to Uf^ which
show the strong deformation retraction combine to give a continuous function
from Wx I to U. Then W is contractible, since it has a strong deformation
retraction onto the contractible set U.I I

Corollary With the same hypotheses, X is locally path-connected.11

Theorem 5 A path-connected and locally path-connected space X

has a covering belonging to the subgroup H of n^(X,v) iff H is sufficiently
large.

Theorem 6 A complex K has a covering belonging to any subgroup

Proof of Theorem 5 Let p\iC + X be a covering, and let H be

nJjF,v)p* . Take JC € AT, and let U be an elementary neighbourhood of JC. Let /
be a path from v to JC. Then / lifts to a path / starting at v. Let x be the end
of / . Then x has a neighbourhood if which p maps homeomorphically onto U.
Let g be any loop in U based at x.
Then g lifts to a loop g in (7 based at x . So the loop f.g.f is
the image under p of the loop J.g.J' , and therefore represents an element of
H. Hence H is sufficiently large.
Now let H be sufficiently large. If there is a covering belonging to
H then we know that every path starting at v determines a point of the covering
space (the end of the lift starting at v" of the path), that every point of the
covering space comes from such a path, and we know (by Corollary 1 to Lemma
1) when two paths determine the same point.
This leads to the following construction. We consider all paths in
X starting at v, and we write f=g if fg represents an element of H. This is
easily seen to be an equivalence relation, and we denote the equivalence class of
/ by [/]. We define if to be the set of equivalence classes, with v' being the
class of the trivial path. We define p:)£ •* X by requiring lf\p to be the
end-point of /.
Take x e ]?, and let x be xp. Choose / with xm lf\. Let U be a
basic neighbourhood of x. Define (x,U) to be the subset of )? consisting of all
points C/.g] with g a path in U starting at x. Notice that this set does not
depend on the choice of / , since f'.g = fg if / ' s / . We shall show that the sets
(xllf) form a basis for a topology.
We first show that if y *(x,U) then {y,U)~(x,U). For there must be
a path g in U from x to y such that y mLfgl Thus (y,U)m {[f.g.hl; h is a path in
U starting at y). Hence G*U)c(xtU). Also, x-[f.g.g\ and so x i{y,U). For the
same reason, this gives (x9U)c{ytU).
Now let z *{x,U)c\(yy). Let W be the component of UnV
containing z- Then W is also a basic neighbourhood (it is open because X is
locally path-connected). We see that (z,W)c{z,U)n(z,V)m (x,U)n{y,V). It follows

that the sets (x,U) form a basis for a topology on Jt', which is the topology we
use.
We now show that p is continuous. It is enough to show that Up
is open for every basic neighbourhood U. Since Up is the union, over all
x € xp and all x € U, of the sets (x,U), this is immediate.
The next step is to show that p is a bijection from (x,U) onto U.
Take any ytU. Since basic neighbourhoods are path-connected, there is a path g
in U from x to y. Then [/.#] is in (x,U) and C/.£]p-.y, and so p is onto. Now
suppose that Lf.gl and If.hi are elements of (JT,£/) with the same image under p.
Then g and A are both paths in U from JC to some point y, and so g.h is a
loop in £/. Since H is sufficiently large, the loop (f.g).{f.h) represents an
element of H. So, by definition, If.gi - C/./r], and therefore p is one-one.
Since p maps each member {x,U) of a basis for ^T onto an open
set U, p maps every open set of X* to an open set of X. In particular, since p is
a bijection on (x,U), p is a homeomorphism from (oT,LO onto £/. It follows
further that if is locally path-connected, since each point JC° has a
neighbourhood (JC'.LO which is locally path-connected, being homeomorphic to U.
The next step is to show that Up is the disjoint union of the sets
(x,U) for all x ixp~l. First take any y\Up~X. Then (y,U)pmU, so that we can
find xl e xp with x^z{y,U). We have already seen that this gives (x^,U)m (y,U)>
and so y ^(^,U). That is, Up' is the union of the sets (x,U).
Now suppose that (x^,U) ni^yU)*^, where x^ p- x - JC^A Let jT be
in this intersection. Then (JC^,6/)- 6 s £ / ) - (JC^*^)- Smce p is a bijection on (y,U),
we see that jcl - JC^ » anc * s o ^P ls
* n e disjoint union of the sets (x,U).
Since these sets are open and connected, p will be a covering
provided that X is path-connected.
Let x e X be [/I, where /:[O,rl ->X is a path starting at v. Let / be
the path defined on CO,f] such that sft~sf for O^s^t. Let / be the function
from from CO,/-] to if defined by tf - [/p. If / is continuous, it is plainly the lift
of / starting at v, and it ends at x'. Hence X is path-connected if / is
continuous.
Take s e [0,/\l. Let U be a basic neighbourhood of sf. Choose 8 > 0
such that JfcU, where J-[O,r\n{?,\t-s\<h). Let iT be (C/,W). If r e / t h e n either
ft'fc-'S or femft.g (depending on whether t>s or t^s), where g is a path in U.
In either case, we see that 1/JelT. Hence jf £(?. Since ?pmf and p is a
homeomorphism from I? to U, we see that / is continuous at s.
We have now shown that p is a covering, but we still have to
identify the subgroup it belongs to. So we take a loop / based at v. This lifts to

a path / in )T starting at v\ The discussion above shows that / ends at [/I

Thus / is a loop iff If] - v'; that is, iff / represents an element of H. Hence, by
Lemma 1, TI^A^V")/?*«//, as required.//

Proof of Theorem 6 Now let AT be a complex, and let H be a

subgroup of 7T|(AT,v). Much as before, we define an equivalence relation on the
edge-paths starting at v by f=g iff / and g have the same end-point and f.g
represents an element of //, and we let v" be the path of zero length at v . This
time the set of equivalence classes forms the vertex set V(ft). As before, we
define Lflp to be the end-point of / .
We define the set of edges E(ft) to consist of all pairs (xye) where
x tV{Jt), e z E(K), and xp~ eo. Let ea be x and let ex be y, and let x be [/].
We define {x,e)o to be x, and (x,e) to be (y,e~), where y is If.el Since
If.e.e"! - [/], we see that ~~ satisifes the required conditions for the vertices and
edges of it to be a graph. We define (xfe)p to be e. Then p is clearly a map of
graphs, which induces a bijection from the set of edges of ft starting at x to
the set of edges of K starting at x.
Let g be the path e^ en, and let e-o be x^. Let JC^ e x^p be
e anc et e ls easv to cnec
C/l Define x'- to be Ife^ > i-\*' * * */ ^ (*/»*/)• ^ ^ that
e+ e is the unique path in ft which starts at xj and is a lift of g. In
particular, if g starts at v then this path starts at v and ends at [#]. Hence K is
connected.
Take the lift / starting at v* of a loop g based at v. From the
previous paragraph, it is easy to see that g ends at [g]. In particular, g is a loop
iff g is equivalent to the trivial loop; that is, iff g represents an element of H.
Hence, by Lemma 1, once we have defined the cells of £ so as to make it a
covering of K, the covering belongs to the subgroup H.
We see also that, for an arbitrary path g, the lift of g starting at Ifl
ends at C/.gl In particular, when g is the boundary cd of a cell c then the
lift of g starting at lf\ is a loop, since in this case f.g.f °*ff which is
homotopic to the trivial loop.
We now proceed to define the cells of £. It is most convenient to
choose a set C+(K) consisting of one cell c from each cell-pair ic,c~} of K. We
shall assume that an indexing of the cells of K has been made. Let C+if?)
consist of the pairs {xyc) for each chosen cell c and vertex x € xp~ , where JC is
the vertex of the indexed cell (c,l). Let C_(R) be bijective with C + (^) and
disjoint from it, with the bijection being denoted by . The set C(£) of cells of
it is to be the union of these two sets. When c is (x,c) we define cd to be the

lift starting at x of the boundary (c9l)d of the indexed cell (c,l). We saw in the
previous paragraph that this lift is a loop. We define p by {x,c)p~c and
m
{x ,c)p c~. Plainly p is a map of complexes.
Let c be (x^,c). We define the corresponding indexed cells so that
{c,\) has vertex JC' and boundary the lift e j , . . . ,e'n of the boundary e^,... ,e of
(c,l)c). We then define py to send ( c , / ) to (c,y), and we extend the definition to
the indexed cells corresponding to the cells of C {£) so that p t sends an indexed
cell and its inverse to an indexed cell and its inverse.
Let the indexed cell (c,i) with ctC+(K) have vertex y. Let y be in
yp , and take a path h with y - C/zl Let g be the path e^ e-^ from x to y.
Let JC' be [A.g ], and let g be the lift of g starting at x . Then the end-point of
g is ih.g .gl-lhl-y. By definition of the indexing of cells, it follows that y is
the vertex of the indexed cell (cj), where c is (x,c). Hence (c,/) is in the
image under pt of the set of indexed cells with vertex y.
Now suppose that (c,i) and {d,j) are two indexed cells with vertex
y with the same image under p, and both in C + ( A ) . Let c be (x,c) and let a
be (z,d). Plainly we must have c- d, and then JC - z, as both are the vertex of
(c,l). Also imj, since (c,/) - (c',/)p, - (<^,y)Pj - (cj). It then follows, with g as
before, that the lifts of g starting at x' and z both end at y. Hence x - z, both
being the end of that lift of g' 1 which starts at y. Thus (c,i) - (ct,j).
Now a cell in C_(f?) cannot have the same image under p as a cell
in C + (^), since the latter image is in C+{K) while the former is not. -It follows
easily from this and the previous two paragraphs that pt is a bijection from the
set of indexed cells with vertex y to the set of indexed cells with vertex y.
Hence p (with the associated pt) is a covering.//

Theorem 7 (/) Let <p:(K, J>Q) -*(X,V) be a map such that

7r«(K, .>/Q)<P*.C 7ij(^,v')p*. Then there is a unique map ^'\Y,y^) -^{)ttv) such that
<ppmtp. (//) Under the similar conditions for complexes, and with a given choice
of (p,, there is a unique <$ such that <pp - 9 and <9\P\' <pj .

Remark We call cpT a lift of 9.

Proof We already know there can be at most one such map.

(i) Take any point yt Y. Then there is a path / in Y from ^Q to y.
The path /cp in X lifts to a path in J? starting at v\ We show first that the end
of this path depends only on y, and not on the choice of / .

So we take another path g from y^ to y. Then f.g" is a loop at yQ.

By hypothesis, the path (/<p).(£fp)~ - {f.g )<p represents an element of TI^(X\v)p*.
It follows, by Corollary 1 to Lemma 1, that the lifts of /cp and gcp end at the
same point.
Thus we have defined a function y:Y-*X which clearly satisfies
9P«cp and y^ip - v\ We still have to show that <p' is continuous.
Now let w be yip, so that wp - ;><p and let W be a neighbourhood
of w. Let W be W^p. We may assume (replacing W by a smaller set if necessary,
using the remarks about elementary neighbourhoods following Lemma 4.17) that
W is an elementary neighbourhood of wp.
Since Y is assumed locally pathwise-connected, there is a pathwise
connected open neighbourhood U of y such that UQ to ^. Given z* U, we may take a path £ in U from ;> to e. Now yip is
the end of that lift h of f<p which begins at v\ while z<p is the end of that lift
of (f.g)<p which begins at v\ Equivalently, zip is the end of that lift of g<p which
begins at vv\ Now g<p is a path in W, and p is a homeomorphism from W onto
W. Thus the lift of g<p starting at w lies completely in $/. In particular z*. Let y be any vertex of L. Then there is a path f in L
from J>Q to j/. The path /<p in K lifts to a path in R starting at v . As in (i), the
end of this path does not depend on the choice of / , and we denote it by yip.
Plainly, yippmy<p and y ^ - v\
We next have to define ip on the edges. Let e be an edge of L
from y to z. Then ecp is an edge of K starting at .yep. Since p is a covering, there
is a unique edge of £ starting at yip whose image under p is e<p. Let eip be this
edge. By construction eipo - eocp. With / as before, we may construct zip using
the path f.e. Thus zip is the end of that lift of (/<p).(e<p) which starts at v. Since
the lift of f(p ends at jap*, it follows that zip is the end of that edge above e<p
which starts at yip. This edge is e<p\ and so ecpr - e\<p. Further, e~ip - e<p' , since
both edges start at zip and both have e~<p as their image under p.
Finally we have to define <pT and <p"f on the cells and indexed cells.
Let {c,c~} be a cell-pair in L. Let the indexed cell (c,l) have vertex ^ and
boundary e^9...9e . Let (c,l)<pf be (c<p,y). This indexed cell has vertex .y<p, and
there is a unique indexed cell (</,£} with vertex y^ such that (</, £}pt - (ccp,/).
We define ccp to be <i, with (c^ljtp', being defined as (d, k). We then define c"tp
to be i'.

Now e^<$,... ,e ^ is a loop in R starting at y<$ which maps by p

to the loop Sjtp,... 9en<p. This loop is (c<p,j)d, and so it is the image under p of
the loop (<^, k)d, which also starts at j><p\ Hence these two loops are the same;
that is, (c,\)d<pJ - (c,l)qTfc), as required. We can now define cjT, on the remaining
indexed cells corresponding to c, and on the indexed cells corresponding to c\
in exactly one consistent way.//

Definition Let PyX^^X and p^.it^ X be coverings of spaces. A

map 9 : ^ -» X~ is a homomorphism if 9P2 " P\» and it *s an
isomorphism if it is
a bijection and both 2! " P\\ •

Proposition 8 Let Py^ -* X and P 2 : ^ 2 ~* X be coverin s

8 (°f
spaces or complexes), and let v, and v ~ be points such that v iPi " v o^2 * Then
there is a homomorphism from )C'^ to X'^ sending v'j to v^ iff
and an
^1(^1 *^-i)Pv*£.^^(^2 ' ^ 2 ^ 2 * ' isomorPhism from (XJ, v^) to ( A ^ * ^ ^ ^

Proof The result for homomorphisms is immediate from Theorem 4.

Suppose that ^ ( X ^ ,v^)P|* - n\(^2 '^2^2*' ^ e n there will be
homomorphisms 9:(A^ ,v^)-^ ( ^ '^V anc ^ ^ ' ^ 2 ' ^ 2 ^ ^ 1 '^P* ^ ^ e n ^ ^ l " ^l
(and, for complexes, ^jfyPu " P\\ also). Thus uniqueness of liftings shows that 9^
( and, similarly, ^9) is the identity; in the case of complexes, the extra
condition needed also holds.//

Corollary The coverings A^ and A^ °f X are isomorphic iff

7Tj(A^ ,v^)pj* and ^ ( A ^ ' ^ 2 ^ 2 * a/>e conjugate in n^(X,v), where v^Pj - v - v^P2 •
7*^/5 holds for spaces or complexes.

Proof Let 9 be an isomorphism. Then ^(A"^ ,v^)pj* - T T ^ ^ »VJ9)P2*

s
and this is conjugate to T C ^ A ^ * ^ ^ * ' * n c e ^ 1 ^ 2 " ^2^2 *
Conversely, if the groups are conjugate we can find a point w of
sucn
J^2 that w/>2 " v and TTJC^J ,V^)PJ*" ^1(^2 »^^2*' ^ y the
P r °P° s ition, this
gives us an isomorphism from (J?l,vl) to (X^2,w).//

By Proposition 8 and its corollary, the (conjugacy class of the)

subgroup to which a covering belongs determines the covering up to

isomorphism.

Proposition 9 Let 9 : ^ -> A^ be a homomorphism of coverings of

X. Then 9 is a covering. The same result holds for complexes.

Proof Let U* and U* be elementary neighbourhoods of x € X for p-,

and /?2 • Then £/, the component of U^ n 6^ containing JC, is an elementary
neighbourhood for both p^ and p^ •
Take any point JC^ of %2 J a n d l e t * ^ e ^ 2 ^ 2 ' ^ e t ^ ^ e t n e
component of UP2 containing JC^ . Then Up^ is the disjoint union of (T and
another open set (since the components of open sets are open). Hence Up<> 9
is the disjoint union of (7<p~ and another open set. It follows that each
component of (7<p~ is a component of Up^ 9" " Up^ • Now />2 ls a
homeomorphism from (7 onto U, while p. is a homeomorphism from each
l
.1

component of Up* onto U. Hence 9 is a homeomorphism from each component

of (?(p onto (?, as required.
The result for complexes is left as an exercise.//

With a change of notation, this result says that if p:lf -» X and

q\X -» X are maps such that p and qp are coverings then q is a covering.
For complexes, it is easy to check that if both p and q are
coverings then qp is also a covering, with (<?p)t being defined as <?tpt. The
following example shows that in the case of spaces qp need not be a covering.
The proposition below shows that qp will be a covering if it possibly can be, in
a sense made precise in the proposition.

Example Let X be the Hawaiian earring. That is, A^R is UX ,

where A^ is the circle {(x,y); x 2+y2 - 2JC//Z} with centre (1/^,0) and radius l/n (for
every positive integer n). Let if be the covering of X obtained by unwrapping
the first circle; that is, Jt is obtained from IR by attaching at each integer point
a copy of those Xn with n > 1. We can obtain a covering ^ of / by
unwrapping, for all n > 1, the copy of X which is attached at the integer point
n. It is easy to check that this is a covering.
Now any neighbourhood U of (0,0) contains X for n large enough.
Consequently, some component of the counter-image of U in X* will contain an
unwrapping of Ar/J, aand so will not map homeomorphically onto U. Thus X is
not a covering of X.

It is clear that the trivial subgroup is not sufficiently large.

Additionally, any sufficiently large subgroup must contain, for all large enough
n, an element which maps to the generator of n^{X ) under the homomorphisrn
induced by the retraction of X to X . Further, the discussion of the Hawaian
earring in section 4.5 shows that any sufficiently large subgroup must be
uncountable.

Proposition 10 Let p:lt -» Xand q:Jt -^ it be coverings of spaces.

Then qp is a covering, iff X has a covering belonging to the subgroup

Proof The condition is clearly necessary.

Suppose that the condition holds, and let r:Y-> X be the covering
belonging to this subgroup. By Theorem 7 there is a map (p:Y-> it such that
/•-9P, and 9 is a covering by Proposition 9. Since p* is a monomorphism and
n^X' )q*p*- n^(Y)<p*p*9 we see that n^J? )<?* - n^Y)?*. By Proposition 8, the
coverings it and Y of it are isomorphic. Under this isomorphism qp corresponds
to (pp. As the latter is a covering, so is the former.//

From Proposition 8 we see at once that if X is simply connected

then (up to isomorphism) the only covering of X is the identity.
Suppose that X has a covering it such that )t is simply
connected. We then call X' the universal covering of X. By Theorem 7, for any
covering Y of X there is a homomorphism from it to Y, and by Proposition 9
this homomorphism is a covering. Thus it is a covering of every covering of X,
which explains its name. These properties hold both for spaces and complexes.

Example Let K+ be the cone on the Hawaiian earring. Let Y_ be

another copy of Y+9 and let Y be the join of Y+ and Y , the common point v
being (0,0) in the earrings. It was shown in section 4.5 that TZ^Y) is uncountable.
We will show that Y has (up to isomorphism) no covering other than the
identity. Thus, for complicated spaces, there may be a covering which is not
simply connected but which still covers every covering.
Take a covering p\{¥,v) -»(K,v). Let ¥+ and ¥ be the components
of y\p and Y p containing v'. By Proposition 12 below, p\1F^->Y\ is a
covering. Since Y+ is simply connected, it follows that p is a homeomorphism
from Kl to y\ and also from ¥ to Y .

This implies that p is a homeomorphism from ?+ u f onto Y. For

let (T be any open subset of this union. Then iTp is the union of the sets
(iT nl?+)p and ((7 n f ) A which are open in V+ and K_ respectively, and if v is
in one of these two sets it is also in the other. Hence iTp is open in Y.
Now take any point y of 1?'. It is the end of some path starting at
v". This path is the unique lift of some path in Y starting at v. Since, by the
previous paragraph, every path in Y starting at v lifts to a path in ?+ u ?'_
starting at v\ we see that ¥ must be 1?+ u ¥_ .

Proposition 11 Let p\{]?,v) ->{X,v) be a covering of spaces {or

complexes). Let A be a path-connected subset {a connected subcomplex) of X
with v e A. Let H be the subgroup TiAXJ,v)p* of TZAX,v), and let i:A -» X be the
-1
inclusion. Then Ap is a path-connected space (a connected complex) iff
n^{Afv)i* meets every right coset of H.

Proof Let Ap be path-connected. Let <xeiiAX,v) be represented by

the loop / at v. Lift / to a path / starting at v\ and let / end at w. Then
w e vp cAp~ . Hence there is a path g in Ap' from v" to w. Now gp is a loop
in A, so it represents in n^(X,v) an element 3 of n^(A,v)i*. Also fta~ is
represented by (gp).f - (g>7~ )p- Since g.?~ is a loop in X', it follows from
Lemma 1 that Pa e H. Hence nJA,v)i* meets the coset Hoc.
Conversely, suppose that n^{Atv)i* meets every coset of H. Take
x f Ap' . There is a path in A from xp to v, and this path lifts to a path in
-i ^ ^ -i -i
Ap from x to some point w e vp . So we need only find a path in Ap from
v to w.
Take any path / in ^T from v~ to vv. Let / be Jp. Then / is a loop
in X based at v. By the hypothesis, there is a loop g in A based at v such that
-1 /v. -1

f.g represents an element of H. Then g will lift to a path g in Ap starting at

v\ By Corollary 1 to Lemma 1, the paths f and g have the same end-point, as
required.//
Proposition 12 Let p\{X,'v) ->{X,v) be a covering (of spaces or
complexes) belonging to the subgroup H of nAX,v). Let A be a path-connected
and locally path-connected subspace of the space X {or a connected
subcomplex of the complex X) with v e A. Let A be the component of Ap
containing v. Then p\{A,v) -* (A,v) is a covering belonging to the subgroup
Hi* , where i:A -> X is the inclusion.

Proof First suppose that X is a space. Take any a e A. Let U be an

elementary neighbourhood of a. Since ^4 is locally path-connected, the component
of A n U containing a is an open subset of A. Hence it is A n K, for some open
subset V of X. Replacing U by the component of UnV containing a> it follows
that we can assume that the elementary neighbourhood U of a is such that An U
is connected.
Take any a e A* with ap~ a, and let (7 be the component of Up
containing a. Then the component of (A n U)p containing a is A n Z?', and it
maps homeomorphically onto An U by p. Hence A is a covering space of A.
Now suppose that X is a complex. Take any vertex a of ^4, and any
vertex a of A with op- a. Then any edge or cell of ]? which has a as a
vertex and maps into A by p must itself lie in A. It is then easy to check from
the definition that the restriction to A of p (together with the corresponding />,)
is a covering.
In either case, let K be it^(A,v)p*9 and let f:A ~* J? be the
inclusion. Then Ki* - n^A\v')p*i* - n^A,V')T*p*cH. Conversely, let a € Hi* .
Then a is represented by a loop f in A starting at v and such that / lifts to a
loop / in J? starting at v\ On the other hand, since A is a covering of A, f
lifts to a path in A starting at v . Uniqueness of lifts now tells us that / must
lie in A', and so a is in AT, being represented by /.//
Let ^T be a covering of X. An isomorphism of ^T with itself is
called a covering transformation of )t. They are clearly of importance when
looking at a covering.

Proposition 13 Let X" be a covering of X belonging to the subgroup

H. Then the group of covering transformations of )T is isomorphic to N(H)/H,
where N(H) is the normaliser of H in nAX,v).

Proof Let <p be a covering transformation. By Corollary 3 to Lemma

^' -1 e vp-1 . One element of this coset is
1, there is a coset of H corresponding to vtp
the homotopy class a of / p , where / is a path from v' to vV .
By Proposition 3, Tt^A^v^p" )p*-<x~ HOL. Since cp is an isomorphism,
l
7T1(^T, v<p~ )p*-H. Hence a is in N{H).
Conversely, given <xeN(H) let J be a path starting at v such that
fp represents a. Then the end-point w of / lies in vp" , and, by Corollary 3 to
Lemma 1, it depends only on the coset HOL. Since n^(jt,w)p*" a" //ex- //, by
Proposition 8 there is an isomorphism 9 of X sending w to v . Thus we have
found to every covering transformation a coset of H in N(H) and to every coset

a covering transformation. It is easy to check: that these functions from the group
of covering transformations to N(H)/H and vice versa are inverses of each other.
That is, we have a bijection from the group of covering transformations to
N(H)/H.
Let 4> be another isomorphism, and let g be a path from v' to ify" .
Let P be the class of gp. Since g*p is a path from vcp~ to VM|T (g<P~ ) is a path from v" to ^(9^)" • The image of this path under p represents
aft, and so the image of 9^ is the coset of aft. Hence the bijection is a
homomorphism.//

A covering is called regular if for every w c vp' there is an

isomorphism sending \T to w. The proof of Proposition 13 shows that this holds
iff N(H)"K^(Xfv); that is, iff H is normal in n^(X,v).
If a covering is regular then any loop in X which lifts to a loop
starting at v* will also lift to a loop starting at w for every w c vp (by
applying to the loop starting at v* the isomorphism sending v* to w).
Conversely, suppose this condition is satisfied. Then every element of H lies in
n^XJw)p* for every w € vp" . By Proposition 3, this tells us that H£(x." Hot for
every a in nAX,v). Hence H is normal in n*(X,v).
Let Y be a space, and let G be a group of homeomorphisms of Y.
We say that the action of G is properly discontinuous if each y*Y has a
neighbourhood if such that UgnUh - 0 for g*h.
Let G act properly discontinuously on Y, and let X be the quotient
space YIG (whose points are the orbits of G, and whose topology is the
identification topology). It is easy to see that Y is a regular covering space of
X, whose group of covering transformations is G. Conversely, let /f be a
covering space of Xy and let G be the group of covering transformations. Then G
acts properly discontinuously on X\ the relevant neighbourhoods being the
components of Up" , where U is an elementary neighbourhood.
When Y is simply connected it will be the universal covering of X.
In this case the group of covering transformations is n^X), by Proposition 13.
Hence n^(X)mG. In general, Proposition 13 tells us that n^(X) is an extension of
nx(Y) by G.
Similar results hold when G is a group of automorphisms of a
complex Ky provided that G acts freely and without inversions; that is, provided
that if g *1 then g fixes no vertex, edge, or cell, and eg^e" and cg*c~ for every
edge e and cell c.

These results can be extended to situations where (/ has fixed

points, and the extended results can be derived from similar results about
fundamental groupoids.

Exercise 1 Prove Proposition 9 for complexes.

Exercise 2 Let p\L-* K and q\M^> L be coverings of complexes.
Show that qp is a covering.
Exercise 3 In Proposition 12 for complexes, check that the
restriction of p to A is a covering.
Exercise 4 Find ^ ( I R / ^ ) , both by writing it as the quotient of a
simply connected space by a properly discontinuous group action, and by the
method used to obtain TT, of the torus and Klein bottle.
Exercise 5 With the notation of Proposition 11, show that there is a
bijection between the components of Ap~ and the double cosets of (//, n^{A,v)i*).
Use this to deduce Proposition 11.

7 COVERINGS AND GROUP THEORY

Now that we have developed the theory of covering complexes and

covering spaces we will use this theory to obtain major results in combinatorial
group theory. We will give three proofs of Schreier's theorem which states that
all subgroups of free groups are themselves free. The first proof uses covering
complexes, while the third is an algebraic formulation of the first. The second is
essentially the same as the first, but is redesigned so that no explicit mention of
coverings is needed. Yet another proof, which is also essentially the same as the
others, is given in section 8.1, while section 8.2 contains the very different proof
due to Nielsen.

Theorem 1 (weak form of Schreier's Theorem) A subgroup of a free

group is free. Further, if H is a subgroup of finite index r in a free group of
rank n then H has rank r{n-\)+\.

Proof Let F be free with basis X. Take a graph F with one vertex
v and one edge-pair for each element of X. By Corollary 1 to Theorem 5.17,
7ij(r) is isomorphic to F. There is a covering f of T belonging to the subgroup
H. Then H is isomorphic to rc^f), which is free, by Corollary 1 to Theorem
5.17 again, since f is a graph.
Since T has only one vertex, the set of vertices of f is vp~ , and
this is bijective with the set of cosets of H. That is, f has r vertices. Since T
has 2n edges, all of which start at v, f has In edges starting at each vertex of
f, by the definition of a covering. Hence f has 2rn edges in total, and so it
has rn edge-pairs. By Corollary 3 to Theorem 5.17, when r is finite the rank of
Tr^f') is rn-r+\.//

This version of the theorem does not tell us a basis for //, which is
the reason it is called the weak form of Schreier's Theorem. Theorem 4 gives the
more precise version which does tell us a basis.

Downloaded from https:/www.cambridge.org/core. New York University, on 30 Jun 2017 at 09:32:36, subject to the Cambridge Core terms of use
Cambridge Books
, available at https:/www.cambridge.org/core/terms. Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.008
Coverings and group theory 168

When r is infinite nothing can be said about the rank of H. For

instance, the free group of rank 2 contains a free subgroup of infinite rank which
has infinite index. Contained in this subgroup there are free subgroups of every
finite rank, all of which have infinite index.

We now give a variant of the above proof which does not require
knowledge of coverings.
We construct a graph A whose vertices are the cosets Hg of H in F.
In particular, the coset H is a vertex which we take as our base-point. There is
to be a function X from the set of edges of A into X u X~ , which we call the
labelling of the edges. Starting at each vertex there is to be exactly one edge
with a given label. The edge with label xs starting at Hg is to end at Hgxe,
and the inverse of the edge with label x from Hg to Hgx is to be the edge
with label x from Hgx to Hg.
Using Corollary 3 to Lemma 6.1, we can show that A is isomorphic
to f, so that 7ij(A) is isomorphic to H. We shall show directly, without using
coverings, that this is true. Then, as above but using A instead of f, we see that
H is free, and we find its rank when its index is finite. The strong form of
Schreier's Theorem given in Theorem 4 can also be proved using A instead of f.
The map X obviously extends to a multiplicative map from paths of
A to words on X. If the path / ' comes from the path / by elementary reduction
then the word corresponding to / ' comes from that corresponding to / by
elementary reduction. It follows that X induces a homomorphism from Y(A), and
also from 7r.(A,//), to F.
In particular, an irreducible path maps to a reduced word of the
same length, and so the homomorphism from 7ij(A, H) to F has trivial kernel.
An easy induction on the length of the word shows that for any
word w and coset Hg there is exactly one path in A starting at Hg whose
corresponding word is w, and this path ends at Hgw. Thus an element w of F is
in the image of TTJA, H) iff //w- H. That is, the image is H.

The next proposition can be regarded as a partial converse to the

numerical part of Theorem 1. It can be proved for subnormal subgroups, not just
for normal subgroups.

Proposition 2 Let H be a non-trivial finitely generated normal

subgroup of a free group F. Then H has finite index and F has finite rank.

Proof There is a finitely generated free factor of F which contains

H. This is the subgroup of F generated by those elements of the basis X of F
which occur in a chosen finite set of generators of H. Since (Exercise 1.18) a
proper free factor of a group does not contain any non-trivial normal subgroup,
it follows that F must be finitely generated.
Take F and F as before. Since H is non-trivial, there is a
non-trivial irreducible loop in F based at v'. Let this loop have length k.
Because H is a normal subgroup of F, the covering is regular. Hence there is an
isomorphism of F taking v' onto any vertex. In particular, every vertex of F lies
in some irreducible loop of length k.
Let T be a maximal tree in F. Because H is finitely generated,
there are only finitely many edges of F not in T. Each irreducible loop must
contain one of these edges. Hence F has a finite subset 5* of vertices such that
every vertex can be joined to a vertex of S by a path of length less than k.
Since F is finitely generated, there are only finitely many edges of F starting at
a given vertex, and so only finitely many paths of a given length with a given
start.
It follows that F has only finitely many vertices, and so H has
finite index.//

The set of edge-pairs of F is bijective with the basis X of F.

Hence we may label the edges of F by the elements of XuX . We may then
give the edges of F the same label as the corresponding edge of F. When this is
done, the paths in F starting at v' (and all paths in F, which are all loops) may
be identified with the word5 on X u X' , and the irreducible paths may be
identified with the reduced words; that is, with elements of F(X). The vertices
of F may be regarded as cosets of //, and a collection of irreducible paths
starting at v with exactly one ending at each vertex may be regarded as a
transversal of H in F(X).
In particular, let T be a maximal tree in F9 and let S be the
transversal corresponding to the set of irreducible paths in T starting at v\ If we
delete the last edge from such a path the result is another such path. It follows
that if x . E l . . . x . £ « f 5 then JC - 6 l . . . x-£n ~l f 5 (we shall assume unless
l l l l
l n \ n i

otherwise stated that x- eX and £^-±1 fo r<>n, and that e f + i ^ " e / . f° r r<n). Such
a transversal is called a Schreier transversal of H in the free group F with basis
X (or, more simply, of H in F{X); but note that the concept depends on the
basis X and not just on the group). Plainly we have 1 € S for any Schreier
transversal S.

Conversely, let S be a Schreier transversal. We regard 5, as already

discussed, as giving a set of irreducible paths in f\ Let T be the graph whose
vertices are the vertices of these paths, and whose edges are the edges in
these paths and their inverses. The lemma below tells us that T is a tree, and
that the irreducible paths in T are exactly the paths we started with. In
particular, S is the Schreier transversal associated with the tree T.

Lemma 3 Let A be a graph. Let P be a set of irreducible paths

starting at v such that there is exactly one path in P ending at each vertex of
A. Suppose that whenever a path is in P then so is the path obtained by
deleting its last edge. Let T be the subgraph whose vertices are the vertices
of the paths in P and whose edges are the edges of the paths in P and their
inverses. Then T is a maximal tree in A. Also the paths in P are exactly the
irreducible paths in T starting at v.

Proof It is enough to show that each irreducible path in T starting

at v is in P. For our hypotheses then tell us that there is only one irreducible
path in T from v to a given vertex. Hence T is a forest. Since each path in P is
plainly an irreducible path in T, we see that T is a tree. As T contains every
vertex, it will be a maximal tree, and the lemma will be proved.
The hypotheses ensure that the trivial path is in P. We will show
by induction on n that an irreducible path e« , . . . ,e in T starting at v belongs
to P.
By the definition of T, either en or e~n belongs to a path in P. By
the hypotheses, this edge will be the last edge of a path belonging to P. By
induction, the path e\y — >en.\ belongs to P, and it is then the only path in P
ending° at e n -1 * ^e~n but en-\iX"C.nT» n o path in P ends with e~
«T. Since e n -1 n.
Hence there is a path in P whose last edge is e . If we delete this
edge we get another path in P. This path ends at e ^ j T , and so must be
6 > j , . . . , e ^ _ j . Hence the path in P with last edge en must be e^9...,en, as
required.//

Theorem 4 (Schreier's Subgroup Theorem) Let H be a subgroup of

a free group with basis X. Let S be a Schreier transversal for H. For all s e 5*
and xtX, let ysx be sxt f where t is the element of S such that Ht~ Hsx.
Then H is free with basis conisting of those y, which are non-trivial.
&x

Proof Take Y and f as before, and let T be the maximal tree in f

corresponding to S. Theorem 5.17 gives a presentation of 7Tj(f), and tells us
explicitly a set of loops which represent the generators. Under the isomorphism
of 7Tj(f) with //, we see that H is free with the stated basis.//
We can look at this theorem and Theorem 1 somewhat differently.
Let F be the universal covering of F. Then it is a tree, being a connected graph
with trivial fundamental group. The group F acts freely on F^ without inversions
(that is, it fixes no vertex or edge, and sends no edge to its inverse), being the
group of covering transformations. Hence the subgroup H also acts freely without
inversions, f is also a covering of f. The group of covering transformations of
this covering is //, and T - f*7//.
More generally, let T be a tree on which a group H acts freely
without inversions. The remarks after Proposition 6.13 then tell us the following
facts. We may form the quotient graph 77H (whose vertices and edges are the
orbits of H in T), and the natural map from T to 77// is a covering. The group
of covering transformations is //. Also, T is simply connected, so the group of
covering transformations must be KAT/H). Hence H is isomorphic to nAT/H),
and this group is free, being the fundamental group of a graph.
Thus we have proved the following theorem, from which the weak
form of Schreier's Theorem follows at once. This theorem will be obtained again
from another viewpoint in Chapter 8.

Theorem 5 A group is free iff it acts freely and without inversions

on some tree.11

We now give a purely algebraic proof of Theorem 4. This proof,

like most good proofs of similar results, is closely modelled on the covering
complex approach, and we look at their relationship.

Lemma 6 Let G be any group and let X generate G. Let <p:G-> G be

a homomorphism such that cp - <p. Then
(/) G- {ker<p){im<p),
{ii) ker,
(zv) ker.

Proof Note that g'^ is defined to be {g )<p - (£cp)~ . Now

g~ (gg^Hg*?)' Since . Plainly iV.cker<p. Since JC = x, we see easily that g = g<p mod N for every g€(7. So if
g€ker<p we have gtN.//

Let G be free with basis X, and let H be a subgroup of G. Let 1

be a transversal for H with le 7*. Denote the cosets of / / b y / / . (where / runs
over some index set /), and let Tf~- denote that member of T which is in H^.
Let Y be the set of symbols y. , where / € / and x € X, and le
be the free group with basis Y. Define a homomorphism &:F(Y)-> H by

We define inductively for each coset H- a function from W, the set

of not necessarily reduced words on XuX , to F(Y); the image of w by the
function corresponding to H- will be denoted by w '. With 1 being the empty
word, and noting that H w and H x are themselves cosets, the inductive
definition (simultaneously for all i) is the following:

l ^ ' - l , xH'*yix, ( J C " 1 ) / / ' - ( J C / / ' J C " 1 ) " 1 , and

( w x £ ) H i - wH'(xs)HiW for a l l w a n d e - ± l .

It is easy to see by induction on the length of v that, for any

words u and v, (uv)Hi-uHi HiU
, and that {v~l)H< - (vHiv~l)~l. It then follows
_ __ rr rr

that (ux x v) im(uv) l. Hence H- defines a mapping, which we still denote by

/ / . , from F{X) to F{Y).
Let 4* be the restriction to H of the map which sends u to u .
Plainly ty is a homomorphism.
By induction on length we see that, for any word w,
rr _____ _______ _1

(w *)&m H-w(H-w) . In particular, for any / u / / , hty§mh, so that ty is one-one

and $ is onto. It follows that F(Y)$ty - //ip, and this is isomorphic to //.
Consider &ty on F{Y). Since (^)(-9>(J;) - ^((J;^)^ • ^ , Lemma 6 tells us
that ^ " ^ y ^ - y " ^ix^^'
For any utF(X), w ^ e k e r ^ iff w^ekerfr iff TTuTTu"1 - 1; that is,
since 7 T - 1 , iff we r .
Further, letting um H • and vm H x we see that

1 H
Hence y. {y- &ty) is a consequence of iu \u±T). It follows that H has a
presentation <y-x ; u for all utT>.
This presentation is more general than the one in Schreier's
Theorem, since T is an arbitrary transversal. It is not clear from this presentation
that H is free. However, we shall not bother to obtain the previous presentation
from this one when T is a Schreier transversal, since that can be obtained from
the graph-theoretic interpretation below of the current approach; also a related
simplification is obtained in the algebraic proof of Kurosh's Theorem.
Let T and P be the graphs previously considered. Let \T. be the
vertex corresponding to the coset H., and let e-x be the edge of P starting at
vj. which lies above the edge of F with label JC. The element ST" of T> regarded
as an element of TT^F), lifts to a path p- in P from v* to vj..
We know, by Theorem 5.16, that nJp) has a presentation with
generators Y, where y. maps to the class of the path p.e . p . , where v. is
**ixT ' a l s o t ie r e
^ ^ a t o r s correspond to the edge-paths p.. If we identify n^(P)
with H using the map induced by the projection from P to F then the map from
F{Y) to 7ij(f') becomes the map $ considered above.
Paths in F and elements of W can be regarded as the same. To each
word w and each coset H. we can find a unique path in P starting at vj. and
lying above w. This path defines, by replacing each e- by the corresponding
yix, an element of F(Y) which is just wHi. The property (uv)H' - uH* vH'u then
corresponds to an obvious formula about the lift of a product of two paths. Also,
rj

when we7\ the element w of F{Y) corresponds to the path p., where H- is Hw.
Thus the algebraic construction of a presentation for H is just
obtained by interpreting the graph-theoretic construction of the presentation
obtained in Theorem 5.16. We obtained in Theorem 5.17 a slightly simpler
presentation than the one in Theorem 5.16 by talcing a maximal tree and using
paths in it rather than arbitrary paths. Since Schreier transversals correspond to
maximal trees, this presentation is the one in Schreier's Theorem, and it can be
obtained from the presentation just constructed algebraically in exactly the same
way that Theorem 5.17 is obtained from Theorem 5.16..
A Schreier transversal can be obtained algebraically as follows.
Define the length of a coset to be the shortest length of any element of that
coset. Let TT - 1, and suppose that we have defined representatives for all cosets
of length < r, so as to satisfy the Schreier condition. Let Hw be a coset of length
r, and choose some w of length r in the coset. Write w as uxs, where u has
length r-\. Then Hu has a representative v already chosen, since Hu has length
at most r-\. Now vxe € Hw, and VJC€ has length at most r. Hence its length must

be exactly /•, and it must be reduced as written. We can then take VJC£ as the
representative of //w, and the Schreier condition still holds. Thus we obtain a
Schreier transversal inductively.

We can extend Schreier's Theorem to obtain a presentation for a

subgroup of an arbitrary group. Let (/be a group with presentation <X)R>^. Let
H be a subgroup of G, and let M be //cp . Let T be a transversal for M in
F{X). We use the previous notation and results for A/. We obtain the following
theorem, which can also be obtained by applying Theorem 5.16 to the covering
C corresponding to H of the complex L which is obtained from F by adding
cell-pairs corresponding to the relators.

Theorem 7 (Reidemeister-Schreier Theorem) H has the presentation

<yix; uM (u<T), rMi (r<R, all i)>**.

Proof Let A/- kercp. Then N is the normal closure in F\X) of R.

This is easily seen to be the normal closure in M of [W- rM~. }. Now rtNcM,
and N< F(X). Hence, for any w, Mwr- M(wrw )w- Mw, and so W-r - JT-. It
follows that W- r(M~.) ~* - (r^')^. We get the stated presentation for //, since M
has the presentation ^y^x\ u (u*T)> .11
Corollary / / T is chosen to be a Schreier transversal then H has
the presentation <Y; Y* , /^>^ ( p , where Y^ - [y-x such that M -x - Wi x)JI

In order to apply the theorem and its corollary, we need to compute

the elements rMi . This element may also be written as (MrM~*)M. Thus, what
H ii
we need to do is compute u , where H is a subgroup of the free group F and
uiH. This can be done inductively from the definition. An alternative approach,
which is essentially the same but with a different notation, is to use the
following way of expressing u in terms of the generators of H given from our
chosen Schreier transversal T.
Let u - x. E l . . . JC - e«. Let tr f T be such that x.l e *... xl 6 '-i t Htr ,
'l '/z \ r-\
and let ucHtnm^; in particular, ^ + i " l if w e / / , and /Q - 1 always. T h e n
u-LU(t x-Srt 1~
1
)Jr n . If 6 - 1 then t x- ^ r + 1 " 1 is one of the generators of H
g i v e n by Schreier's T h e o r e m ( but it may be 1), w h i l e if £ , " - 1 then t^x- ^ . j "
is the inverse of a generator. T h u s w e h a v e written a n y u e H in terms of t h e
generators of H.

Examples Let F be free on (a,ft}.Let A be <a>F. Then A has

{ft', all /} as a Schreier transversal. Hence A has basis [blab~l).
The commutator subgroup f has { a V , all / and j) as a Schreier
transversal. The representatives of the cosets containing albf b and alb^a are
a V + 1 and a^Xbf, respectively. Hence F f has basis ( f l V ^ a ' ^ ^ } .
Let N be the kernel of the homomorphism from F onto the
symmetric group of degree 3 which sends a and b to the permutations (1 2) and
(1 2 3). Then N has the set [\yb,b\a,ba,b a) as a Schreier transversal. Now a ,
3 2 2
ft , and (aft) are in N, and so babiNa and ft abiNba. It follows that N has
basis (ft3, a 2 , ba2b~l, b2a2b2, aba'Xb~2, baba~X, b2aba'Xb'X). Another Schreier
transversal is {l,ft,ft ,a,abyab }, and the corresponding basis is
{ft3, aftV 1 , bab~2a~l, b2abXa~X, a2, abab'2, ab2ab'X).
We now apply similar, but more complicated, techniques to
determine the structure of subgroups of a free product. A rather simpler (but
fundamentally similar) proof of the result is given in the next chapter.
Theorem 8 (Kurosh Subgroup Theorem) Let H be a subgroup of the
m
free product %G
a . Then H F*%{uGa u n H), where F is free and there is one
factor uG u n H for each a and each u in a suitably chosen set of
representatives of the (^»^a) double cosets.

Proof {using covering spaces) Take disjoint connected complexes L

with a vertex v^ in L^ such that ^ j ( ^ a >va) = ^ a • Take a new vertex v and new
edge-pairs {eoc,e~a} such that eex
o - v and eaT - Va for all a, and let Ka be
L a u(v,e a ,<f a }. Let K be UK a , so that T T ^ / Q - %G^.
Let p:£ -* K be the covering corresponding to H. Let the
components of L^p be L^- (where / runs over some index set depending on a),
and let v* . be a vertex of vp~ in C^.. It is easy to check that the components
of K p will be complexes £ • , where ft • is the union of L • with certain
r
Or OLl (XL OLL

edge-pairs in {e ,e~ )p a n d t h e vertices of these; further, each such e d g e - p a i r

has e x a c t l y o n e vertex in C^ and n o t w o h a v e a vertex in c o m m o n .
Let T • be a m a x i m a l tree in L' •. T h e n U7* • is a forest. Let Y b e
a; on on
the subgraph of f? w h o s e edges are t h e e d g e s of U7 1 • t o g e t h e r w i t h t h e edges of
[e
(x^(x]p'X X f o r a11 a* T h e n VW"V(£)> s i n c e V{ T
< oii)mV^oLi)' N o w a n y t w o
v
vertices of Y may be joined by a path in A , each edge of which is either in
{e a ,e a )p for some a or is in L^- for some a and /. In the latter case there is
a path in T^ beginning and ending at the ends of this edge. Replacing each

such edge by the corresponding path, we obtain a path in Y. It follows that Y is

connected, and so U7" - lies in a maximal tree T of Y.
The intersection of any two of the connected complexes Z^-uT is
T. Also Kn(Z* -U7 1 )-T, and g-YvUL^. Take a vertex v\vp~l. Using the
presentations of the various fundamental groups given by Theorem 5.17 and the
maximal tree T, we see that TT-JCAT.V') - TZ^{Y,V) * *7Tj(La^-u 7>').
Since p* maps 7Tj(A^,v) isomorphically to //, and n^{Y) is free, the
theorem follows once we have shown that 7C|(£'a/-u 7\v')p* is of the required
form.
Any edge with one vertex in L^ either lies in L^ or is e^ or e~a.
It follows that any edge with one vertex in C . either lies in £ . or has its
i
other vertex in vp . We then see that a path in T from v* which ends in £' .
and has minimal length must be of the form f .e , where J ends in a vertex of
-1 ~ y
vp and eQLPmeQL' Let X e 7ij(AT,v) - (7 be represented by / p . (Strictly, we should
refer to paths ?^ and edges e V , but this complicates the notation too much.)
Since L^^-nT* T^-, we see by van Kampen's Theorem (since 7^(7")
is trivial) that njC .U7\V' .)"TTI(£' .,v'-.). By Proposition 6.12 this is mapped
X VA it vA *• JL \A • VA 4
isomorphically by p * t o the group nAL ,v )n-Kj/?,v',)p* This image equals

>vot) "

which in turn equals (G^rni^f?,v)p*\ )e a # , which, since X# is just conjugation

H)**
by X, can be written as (XC?aX" n H)\*e
But we may also write

We deduce that ^(Z,^. u T,v )/?*- XC7aX n //, and so 7r^(Z.a/- u T,v ) is isomorphic
to XG^X"1 n //.
This is almost what we want. We still have to show that the
elements X for a given a form a set of double coset representatives. This is not
difficult, as the subcomplexes ft^. can easily be shown to correspond to double
cosets. Thuis will be left as an exercise, as the algebraic proof below gives more
specific information, and a simplification of the above proof is given in the next
chapter.//

Proof (using purely algebraic methods) Let G- *<?a, where a runs

over some index set A. To begin with we shall assume that each G is free and

has basis X^. We shall also choose one element of A, which we denote by 1.
We shall obtain a presentation of H (which is, of course, a free group in this
case) from which the result for arbitrary G^ can be deduced.
Let Hc_G, and let Ht (for / in some index set /) be the cosets of
//, with //j being //. For each a take an element a77T in the coset H-; this
a
element will be called an a-representative of //..We require that / / - l for all a;
no other conditions are required at present, though they will be imposed later.
Let K be a set with elements y. for all / and all x*UX . Let Z
be a set with elements z-^ for all i*\ and all a ^ l . We define a homomorphism
§:F(Y)* F(Z)-> H by

y- d- - ^TT- x ( a / / -x ) where x e X^, and z. 9 - a7TT( T/T ) .

For convenience, let z- be 1 if / - 1 or a - 1; the formula for z- 9- will hold in

this case also.
We define, for each / / - , a function, whose value on w will be
H
denoted by w ', from the set of all (not necessarily reduced) words on
UArau(UAra)"1 to F{Y)*F(Z). These functions, are defined inductively by

l / 7 ' - l , xHi*y. ix
, (x~l)Hi-(xHJ)~l where H.-H-x,
i j
and

{ux£)Hi - uHi zko Zk'J(x£)Hk, where e - ±1, H, - H-u,, xeX^,,

and the last letter of u is in G* (when u has normal form g* . . . g the elements
gj and gn are called the first and last letters of u, and when u - 1 we say that
the first and last letters of u are in G for all a).
As in the algebraic proof of Schreier's Theorem, we can easily show
rj LJ -1/7
im
inductively that {uv) u ' z o z~ v J for all u and v, where H-H^u, and the
last letter of u is in <7« and the first letter of v is in G . In particular,
£ -E H H
{ux x v) f»(wv) '", and so these functions induce functions from G to
F(Y)*F(Z) with the same properties.
(inductively) Ifthat
the(u^'W-^H
first letter .ofu{$H
u is-uin)'*.
GMItand the last
follows that isthe
in restriction
GU we find<J> to H
l
H
homomorphis such that (JJ-9- is the identity. As
of the map sending u to u is a homomorphism
before, this shows that H has the presentation

<yu (all i and JC), zioL (all i*\ and

We now proceed to simplify this by a good choice of

representatives (called a Kurosh system of representatives) satisfying the
following conditions:
(i) One coset in the double coset HgG has an a-representative
which is either 1 or has its last syllable not in G ; this element will be called
an a-representative in the large. When H^HgG^ and g is the ex-representative in
the large, then a77~ e gG^.
(ii) If g is an a-representative in the large whose last syllable is in
6U then ^Tfg is g.
Our construction will provide a Kurosh system for a subgroup of an
arbitrary free product, not just for a free product of free groups. This will be of
importance later.
We construct these representatives by induction on the shortest
length of elements of HgG^ (length in the expression of G as the free product
*O , not as a free group). Once we have chosen an a-representative in the large
for such a double coset it is easy to choose oc-representatives of each coset in
this double coset so as to satisfy the condition (i).
Take g of shortest length, r say, in the double coset Hg^- If r - O
then g-l, and we take 1 to be the a-representative in the large. If r / 0 we may
write g as ugn where l&gneGn and u has length r-\. We cannot have ft" ex, as
then u would be a shorter element in the same double coset. Now HgG* contains
u of length /--I, so, by our inductive assumption, we already have a
^-representative in the large of HgGa • Let this element be v. Then v has length
at most r-\. As remarked, we may also assume that a 3*representative of form vw
with we C7Q has been chosen for the coset Hg. We then choose vw to be the
a-representative in the large of HgG . Plainly condition (ii) is satisfied for this
element, and we have completed the inductive definition.
If a^€ker$ij> then u *. kerft, and so u{^7Tu)~l - 1, where the last
letter of u is in G~ (since all representatives of H are 1, it does not matter where
the first letter of u lies). Thus u is a representative. Conversely, if u is a
3-representative whose last syllable is in G then either 3 " Y» so that u- ^ Hu,
or, because of (i) and (ii), u is the 3-representative in the large, in which case
we again have w^Hu. Hence u e kerfr^ for any representative u.
Let u be an a-representative in the large whose last syllable is in
Go, and let Hi denote Hu. Since a 7/T - w - ^ 7 7 ~ , we see that zioLzfa f- ker&J/. We
call the elements Z^Z^'n relators of the first kind. Note that if u is any
a-representative, and the last syllable of u is in GU, then either a - 3 and so
-1 -1
Z- z-a "1 o r e l s e Z- Z-o is a relator of the first kind.
10L Ip 10L Zp

Take any / and a, and any xiX^, and write um<x7J~ and v-aH-x
and H.- Hv~ Hux. Let the last syllable of u be in GU, the last syllable of ux in
£ , and the last syllable of v in 6^. Then y - a unless the last syllable of u is
JC , and in this case ux is the a-representative in the large, by the properties (i)
and (ii). Thus z- Z- is either 1 or a relator of the first kind.
Observe that 1 - (vv" 1 ) 7 7 - vH(v~l)Hv, so that {v'l)Hux « {vH)~l. Hence

y.xW - (uxv-1)* - uH *-p ^ ^ ^ ^ (v")- 1 ,

where z- Z -c , being (z- z- ) (<:• z •* ), is a consequence of relators of the first

kind.
H -1 H -1
Similarly, we find that e- &ty~u Z:nZiyr(v ) for some
representatives u and v, where Z-^z-'n is either 1 or a relator of the first kind, as
is ziyzn~l. Recall that zn - 1.
Thus the defining relators of the presentation are consequences of
LJ

the relators of the first kind and the relators [u ; all representatives u). We
proceed to simplify the second type of relator further.
Take any a and any a-representative in the large u. Let K^a be
[y- : JC eX and H.£HuG }. It is easy to check (inductively on its length as a
word in X u X^) that for any g^O^ the element g is in F(Y ).
Let v be any a-representative, and write v as ug, where u is the
a-representative in the large and giG . Then v - u Zn Z • g , where H is
Hu and the last letter of u is in GU. It follows that this element g is a
relator. It also follows, by induction on the length of a relator, that every relator
is a consequence of relators of the first kind and those relators which are of
form gHu.
We have therefore shown that H has a presentation on the
generators z- and j>. in which each defining relator is either of the first kind
or is in some F{Yu(x). Hence H'(F{Z)S)* *(F(YU(X)&). Also F(Z)& is free, since
the relators of the first kind simply identify certain elements of Z.
Let u be an a-representative in the large, and let H-^HuG and let
x c ArQt. Then the representatives a7TT and a / / ; j c are ug and ug' for some g and
g'*G . Thus y. &- (ug)x(ug')~ iuG' u~ , and, of course, y- 8-e//. Conversely, let
hdlnuG' u\ Write h as K£H \ Then {ug)n ~ u" lgMU , where 5 is a relator of
the first kind, and (hu)n - {to\>)un, since /ze//.As remarked, s " is in F(K ).
n n
Since [hu) - {ug) , we see that A- A<l#- g # t F{Y )§. It follows that
F{Yuo()§ - Hc\ uGau . This proves Kurosh's Theorem when each G is free.

Now let G- *^ a , where the groups G^ are arbitrary. Take a

presentation <X^ ;^a><Poc of <?a, and let <p: *F(X ) -» %G^ be induced by the <p .
For each £€<?a take an element uiF(X^) with utp^-g, where w\ if g - 1 . Then,
for each element of G. say g^... gn where g. and g ^ are in different factors, we
have a corresponding element of *F(Xa) mapping to it under <p, namely the
element u^... u where u^ is the element chosen to map to g..
Let H be a subgroup of G, and let K be //<p . Take a Kurosh
system for H in G. It is easy to check that the corresponding elements of %F(X^
form a Kurosh system for K in ^ ( A ^ ) .
The presentation already obtained for K gives rise to a presentation
for H. Now H is obtained from K by factoring out by the normal subgroup
generated by the elements vr^v for all a, all >"a€^a> and all a-representatives
v of cosets of K. These elements may be written as (/• v)0. Since the
presentation of K is obtained using &, we see that H is obtained from K by
adding the extra relators r^ v. Now let w be the a-representative in the large
such that v*KwF{X ). It is not difficult to check (by induction on the length in
the generators X^) that for any u^FiX^) we have u^ € F(Y^). Thus the
expression of AT as a free product gives rise to an expression of H as a free
product. This is
is what we require, noting also that {KnwF(X(x)w~)<p-HngG(Xg\
where gm wcp.//

We note also that any subgroup HngG^g is conjugate in H to one

of the free factors. To see this, we need only write g as /zc#a, where h e // and c
is an a-representative in the large of H.

The above algebraic construction has a significance similar to that

in Schreier's Theorem. Take complexes L^ such that T T ^ L ^ - G ^ , and such that
L a has only one vertex and has one edge-pair corresponding to each member of
Ara. Let L-UL^, the vertices being identified, and let p:C -» L be the covering
corresponding to H. Choose a vertex v' of C, and the vertices then correspond
to cosets of H. For each component of L p choose some path from v' to a
1 i
vertex in L^p , and then for each vertex v^. of L^p take a path / ^ (trivial if
v^.- vO from v' to v^- which is the product of the chosen path and a path in
Then by Theorem 5.16 H-n^C) is generated by the classes of the loops
•^/l'^oc'^/1^ ' w n e r e *^a is a n e c l S e in L p from \T. to v..
v*.. Thus H is generate
generated
by the classes of the loops / - a . e^ •/ -a~ together with the classes of the loops
/ . . / - / . These correspond to v- and ziri respectively. Given
edge-loop u in L whose first edge is in L^ and whose last is in

(equivalently, any element of G whose first syllable is in G^ and whose last is

in GQ) and given any vertex v' of C, we may lift u to a path u starting at v.,
and we denote its last vertex by v'. Then w ' is the expression for the loop
-Q
/ .zcx .u ./ yp in terms of the generators. Note that this construction is not quite
the same as the one used in the covering space argument. The complexes used
here can be obtained from the ones previously used by contracting subtrees.

2 2
Exercise 1 Let G be <a^,... ,an ; a^ ... a >. Let (p,^:Cr-> ^ be the
homomorphisms given by tfj<p»l, a.(p - 0 for / / I , and <^4"1 for all /. Use the
Reidemeister-Schreier Theorem to find presentations of kerG. Show
that (a', 0 ^ / ^ 7} is a Schreier transversal for N. Show that /V has a presentation
<xyy,z,w\ xyzw wzyx>, where x, v, z, and w are among the generators of N
coming from the stated Schreier transversal. Find a presentation for N in terms
of the generators x, y, u, and v, where wyxz and v~wz.
2 3
Exercise 3 Let G be <a, b\ a ,b >. Find a normal subgroup of index
6, for which the Reidemeister-Schreier method (and some Tietze transformations)
shows that the subgroup is free. Find another proof that the subgroup is free.

8 BASS-SERRE THEORY

In 1968-69 Serre, in a course of lectures at the College de France,

introduced a new technique into combinatorial group theory. Some of the more
general results of the theory were due to Bass, so that the material is known as
Bass-Serre theory. The most convenient source for the original material is in
the book: by Serre (1980).
We shall see that the theory is to a great extent just a disguised
form of the theory of covering complexes. However, it permits us to work with
graphs only, rather than arbitrary complexes, and with an analogue of the
universal covering rather than an arbitrary covering. These simplifications permit
us to avoid the notational complications of the earlier theory, and so make
possible powerful techniques for obtaining a range of results.

8.1 TREES AND FREE GROUPS

An action of a group G on a set X is a homomorphism <p from
G to SymA', the group of permutations of X. We let Sym.Y act on the left (so as
to be consistent with other writers on the subject), and we write gx instead of
{g<p)x, where gaG and x e X. When G acts on X we say that X is a G-set.
A graph V is called a G-graph if both V(Y) and E(V) are Osets, and
ge - ge~ a n d {ge)T - g(ei) for a l l g t G a n d e *• E{T).

Example Let G be any group, and let S be any subset of G. Let

V{O,S) be the graph with vertex set G and with edge set C7x S x {-1,1}, where
{g,s,l)o~g and {g,s,\)x - gst and with (£,s,e) - (£,s,-e) for E - ±1. Plainly V{G,S) is a
graph on which G acts. We call V{G.S) the Cay ley graph of O with respect to S.

Lemma 1 (/) T(G,S) is connected iff S generates G. {ii)r{G,S) is a

forest iff S freely generates <S>. {iii)r(G,S) is a tree iff G is free with basis
S.

Downloaded from https:/www.cambridge.org/core. New York University, on 30 Jun 2017 at 09:32:37, subject to the Cambridge Core terms of use
Cambridge Books
, available at https:/www.cambridge.org/core/terms. Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.009
Trees and free groups 183

Proof These results follow immediately if we show that there is a

bijection between ways of writing g h (for arbitrary g and h in 6^ as s^1... s^n
with n arbitrary, 5- e S, and s - ±1, and paths in l\GyS) from g to /z,
with the bijection making reduced words correspond to irreducible paths. The
required bijection maps s ^ 1 . . . s*n to the path e^,... ,en, where er is ( g ^ s ^ e p ,
{ r l 1
with gr being gs^ ... sr_^ ~ if E ^ - I and gr being gs^ ... s*r if £ ^ - - 1 . / /

Note that when G is free with basis 5, the tree T(G,S) is the graph
obtained in Chapter 5 as the universal cover of the graph with only one vertex
and with one edge-pair corresponding to each element of S.
Let the group G act on the graph X. If there is some gzG and some
edge e such that ge - e~ we say that G acts with inversions; otherwise we say that
G acts without inversions. Plainly G acts on V{G,S) without inversions.
We discussed in section 5.5 how to construct from any complex K
another complex K\ which we called its barycentric subdivision. When K is a
graph, which is the only case that will concern us, AT' is just obtained by
subdivision of edges (for a general complex we have to subdivide edges, and
then subdivide cells). It is easy to see that if G acts on a graph X then it also
acts on its subdivision X\ Also, G always acts on X' without inversions.
Because of this, in future when we are given an action of a group
on a graph we shall always assume that the group acts without inversions.

Let G act on X without inversions. We form a graph called the

quotient graph of X by G, and written G\X this notation is preferable to writing
X/G, which suggests that G acts on the right), as follows. On the vertex set
V(X) we have an equivalence relation given by w=v if w= gv for some g, and
we have the similar equivalence relation on the edge set E(X). We denote the
equivalence class of v or e by [vJ or CeJ. Then the vertex set of G\X is defined
to be the set of equivalence classes lv.1 for veK(A'), and the edge set of G\X is
defined to be the set of equivalence classes LeJ for e n E(X). We define [e] to be
le~], and leh and lelo to be iei\ and Leo]. It is clear that these only depend on
le], and not on the choice of e. We need G to act without inversions to ensure
that we have

The next three lemmas could be extracted from material in Chapters

5 and 6, but I prefer to give an explicit proof. Only the first lemma is needed
immediately. We begin with two definitions.

Let p:X -> Y be a map of graphs. We call p locally surjective if for

every vertex v of X and every edge y of Y such that yo - vp there is an edge x
of X with x o - v and xp~y. We call p locally injective if for any two distinct
are
edges x* and x^ of A' such that x*o - JC«O the edges JC^/? and ^ p °f ^
distinct. We call p locally bijective if it is both locally surjective and locally
injective.

Lemma 2 Let p:X^Y be locally surjective. Let T be a tree and

let f:T->Y be a map. Let a be a vertex of T, and let v be a vertex of X such
that vp- af. Then there is a map <$>:T-> X such that 9 / ? - / and a<p~ v.

Proof Consider pairs (7\ ,cp.) such that 7\ is a subtree of T with

tf€ 7\ and 9, is a map from 7\ to X such that #<p| - v and cp.p equals / on 7\ .
Such pairs exist; for instance, 7\ could consists solely of the vertex a. We may
and
give a partial ordering to these pairs, writing (Tj ,cpj) < (7\, ,92) ^ ^ 1 - ^ 2 ^1
to
is the restriction of ^ ^1 • By Zorn's Lemma, there is a maximal pair
(7*0,<PQ)> and we need only show that T^m T.
Since r is a tree, either T^~T or some vertex of T is not in TQ.
In the latter case there is an edge e with eo f TQ and ex*' TQ (this must hold for
some edge in a path starting in 7'Q and ending outside TQ). By hypothesis, there
is an edge x of X starting at eocp^ such that xp-ef We can then enlarge TQ to
the tree 7\ - 7'Q U {e,e~,ex}, and we can extend <PQ to a map 9, on 7\ such that
<PjP"/on r^ by defining ecpj to be JC, e~<p^ to be i~, and ^xcpj to be JCX. This
contradicts the maximality of T^JI

Lemma 3 Let p:X •> Y be locally injective. Let 9 and i[> be maps
from a connected graph Z to X such that (pp'typ. If there is some vertex z
such that Z9 - zty then 9 - ty.

Proof Let S be {ve V{Z)\ V9 - VIJJL Take an edge e with eo*S. Then
29 and ety both begin at ^09 ( - eoty, by hypothesis) and have the same image
under p. As p is locally injective, this implies that 29 - ety. In particular,
ei(p - #9X - eijn - ex4>, and so ex c *S\ Since Z is connected, and 5*9*0 by hypothesis,
Lemma 5.1 tells us that S - K(Z). This, combined with what we have just proved
about edges, tells us that 9 - ^ . / /

Lemma 4 Let p:S^Y and q:T->Y be local bijections, and let S

and T be trees. Let v and w be vertices of S and T respectively such that
vp-wq. Then there is a unique isomorphism <p:S-> T such that v<p-w and <pq-p.

Proof Since S is a tree and q is locally surjective, by Lemma 2

there is a map cp:5 -» T such that vcp • w and <pqmp. Since <? is locally injective, cp
is unique, by Lemma 3. Reversing the roles of S and T, there is also a map
ty-.T^-S such that wj;»v and typm q.
Then vtptf - v - vl^, and <ptypm p-l^p, where 1^, is the identity on S.
Since p is locally injective, Lemma 3 tells us that <pi|;-I~. Similarly, 4*9 "IT**
and so cp is an isomorphism.//

Lemma 5 Let f.X^Y be a map of graphs. (/) / / / is locally

surjective and Y is connected then f is surjective. (//) / / / is locally injective,
X is connected, and Y is a tree then f is injective. (Hi) If f is locally bijective
and X and Y are trees then f is bijective.

Proof (i) Let e be an edge of Y such that eoeK(A')/. Take a vertex

v of A' such that v/- eo. Since f is ocally surjective, there is an edge x of X
such that JCO-V and xfme. It follows that ei*V(X)f9 and the result follows by
Lemma 5.1.
(ii) Since / is locally injective, it is injective on edges if it is
injective on vertices.
Suppose, if possible, that there are distinct vertices v and w such
that v/« wf. Since X is connected, there is an irreducible path e* , . . . ,e from v
to w. Then e^fy...,ef1f will be a loop in Y. Since Y is a tree, there must be
some / such that e.^f^eT-f Because / is locally injective, this gives £/+i"^/»
contradicting the irreducibility of the path e^,... ,en.//

Let AT be a G-graph, and let p be the natural map from X to G\X.

Let T be a maximal tree in G\X. We see from Lemma 2 that there is a map
j:T-*X such that jp is the inclusion of T in G\X. In particular, p defines an
isomorphism from Tj to T. Hence Tj is a tree, which we call a representative
tree for the action of G on X.
We say that G acts freely on X if for any ge G - {1} and any vertex
v of X we have gv*v. For instance, for any group G and any subset S of Gy we
see that G acts freely on V(G9S).
An orientation of a graph is a set of edges containing exactly one
edge from each pair. Let G act on a graph X. Let B be an orientation of G\X,
and let A be 5/? . Then A is an orientation of X and 6^ - A. This property is
needed below.

Theorem 6 Let G act freely on a tree X. Let T be a representative

tree for the action of G, and let A be an orientation of V such that GA - A. Let
S be [gtG-il); le e A with eotT and eitgT). Then G is free with basis S.

Proof Observe first that gTnhT-0 for g^h. Otherwise we can

choose vertices v and w in T such that gv-hw. Since T is a representative tree it
contains only one vertex from each orbit, and so v-w. As G acts freely, this
shows that gm h, as required. Further, every vertex is in gT for some g, since T
is a representative tree.
Let Y be obtained from X by contracting each of the trees gT. By
Lema 5.12, Y is a tree. By Lemma 2, it is enough to show that F(G9S) is
isomorphic to Y.
Map T(G,S) to Y as follows. The vertices of Y can be identified
with the trees gT, and we map the vertex g of V{G,S) to the vertex gT of Y.
To each s e S there is an edge e of X in A with eo * T and ex e sT.
There can only be one such edge, as two such edges would give two distinct
edges in Y with the same start and finish. The edge ge of X gives an edge of
Y from gT to gsT. We now map the edge (g,s,l) of T{G,S) to this edge of Y, and
the inverse (g,s,-l) of (g,s,l) to the inverse of this edge.
This function is clearly a map of graphs which is bijective on the
vertex set. The edges (£,s,l) and {hys\-\) have different images, since the image
of the first is in GA = A, while the image of the second is in GA- A . The
images of two distinct edges (g,s,\) and (/z,s',l) have different starting points gT
and hT if g±K while if g= h they have different finishing points gsT and gs'T.
Thus the map is injective on edges.
Now take any edge of Y. By definition of K, it is the image of an
edge e with eo t gT and exzhT for some g and h in G with g^h. Replacing the
edge by its inverse if necessary, we may assume that e e A. Then g eeGA-Ay
and g htS, since g e is an edge starting in T and finishing in g hT. Then
{g,g h,\) is an edge of Y{GyS) whose image in Y is the given edge. Hence our
map is also surjective on edges, as required.//

Combining Theorem 6 and Lemma 1, we see that a group is free

iff it acts freely on a tree. When a group acts freely on a tree, any subgroup
also acts freely on the same tree, and we find once more that any subgroup of a
free group is free.
By going into a little more detail, we may obtain a Schreier basis
for a subgroup of a free group. Let F be free with basis U, and let 6" be a

subgroup of F. Let X be r(F,U), with the orientation consisting of the edges

(/,w,l). Let T be a representative tree for the action of G on X such that 1 e T.
Since K(*)« F9 we see that ^(JT) must be a transversal of G Let
/ e 7*, and let / - hus, reduced as written. Then the unique irreducible path in X
from 1 to / goes through h. This path must lie in the tree T, and so hzT,
showing that V{T) is a Schreier transversal.
Conversely, take any Schreier transversal. As a subset of F, we can
regard it as a set of vertices of X, and we let T be the full subgraph spanned
by these vertices. Because we have a Schreier transversal, the irreducible path in
X from 1 to any vertex of T lies entirely in T. Hence T is connected, and so
T is a tree because X is a tree. Plainly T is a representative tree, since V(T)
contains exactly one vertex in each orbit of G Thus Schreier transversals and
representative trees containing 1 are essentially the same things.
The basis S for G consists of those g / 1 such that there is some
te T and u € U such that {t,u,l) ends in gT. This means that tuzgT, and so there
must be some t^tT with tu-gt^. Thus gm tut^ , where t^ is the element of T in
the coset Gtu. This is exactly the basis we have previously obtained.
This proof is essentially just a translation of the previous proofs.
The tree X is the universal covering of the graph with one vertex only and with
one edge-pair for each element of U. The graph G\X is just the covering of this
graph corresponding to the subgroup G. The projection of X onto G\X maps T
bijectively to its image, by definition of T, so this image is just a maximal tree
of G\X, as used before.
The current proof is slightly simpler than the previous ones,
mainly because we work only with the universal covering rather than with
arbitrary coverings, and we do not have to talk about fundamental groups at all.
Also, we gain some notational simplifications by obtaining a basis for G by
considering an arbitrary tree on which it acts freely, and only later specialising
to the tree F(F,U).
We shall develop in later sections the theory of groups which act
non-freely on trees. We shall obtain subgroup theorems which generalise the
Kurosh subgroup theorem, and which include subgroup theorems for amalgamated
free products and HNN extensions.

8.2 NIELSEN'S METHOD

We have seen that every subgroup of a free group is free. In order
to find a basis for a subgroup C7, the approach we have considered requires us to
find a transversal for G. We are quite likely in explicit examples to be given
a subset S which generates G, and in such contexts reference to a transversal is
not very natural. We would prefer a method which gives a sufficient condition
for S to be a basis of G, and which gives a procedure for replacing any 5 by
another generating set which satisfies such a condition. Nielsen's approach and
its generalisations provide a method for doing this.
We call the subset S of the free group F{X) Nielsen reduced if it
satisfies the following three -conditions: (NO) W 5, (Nl) IvwUlvl for all v and w
in SuS~* such that vw*l, (N2) \uvw\ > \u\ + \w\ - \v\ for all u, v, and w in S u S~^
such that uv^l^vw. Note that condition (Nl) includes the condition
Iw v I > Iw I, which may be rewritten as lvwl>lwl.
Condition (Nl) holds iff in reducing vw at most half of w (and v) is
cancelled. In the presence of (Nl), condition (N2) holds iff either the left half of
v does not cancel in reducing uv or the right half of v does not cancel in
reducing vw.
For vtSuS' , let X(v) be the longest initial segment of v which
cancels in reducing some uv with u <• S u S~ and zvv/1, and let p(v) be the
longest final segment of v which cancels in reducing some vw with w^SuS
and vw^l. Plainly p)v)-X(v~ )" . Note that X(v) and p(v) may be trivial.
From the remarks already made, we see that if S is Nielsen reduced
then for each vy c« 5S uu5S~ there
there isis aa non-trivial segment {i(v) of v such that
v - X(v)^(v)p(v), reduced as written.

Lemma 7 Let S be Nielsen reduced, and take u^ u^S^S'

such that u-^^u- for i < n. When the product u^...u is reduced, the
cancellations do not affect any [i(u.). Also \u^ ... u^\ * \u^ ... u^ and
\ux...un\±\uv..unA\.
Corollary 1 With the above notation, \u^.. .u^*n, and
lu, . . . u I ^ \u-\ for all i.
Corollary 2 When S is Nielsen reduced it freely generates <S>.

Proof Corollary 2 and the first part of Corollary 1 are obvious

consequences of the lemma. The second part of Corollary 1 comes by an obvious
induction from the lemma.
We prove the lemma by induction. Suppose that the cancellations

required to reduce the product u^ •.. un do not affect any [L{U-) for / - 2 , . . . ,/z.
Since no part of ^(z^) sets cancelled, the product u~...u must begin with
W y
X(«2W 2^- ^ definition of p(ttj) and \{u^), in the product tt^ cancellation
tnat we
cannot affect [L{U^) or ^(z^). (Note require S to be Nielsen reduced in
order that the segments \L(U-) are defined.) Since both u~...u and u~ begin with
\(u2)^(1^2), it follows that in the product a ^ * *" un cance
^ a t i ° n cannot affect
u(Wj) or (1(2^2)- Since it does not affect u(ii2), it cannot affect any of the
segments \±(u.) for i>2, as these occur to the right of lUt^)-
Since 2lp(z^)l ^ Iwjl, we see that we also have I t t j . . . u^ * Iz^ • • • "^'«
Applying this to ( u « . . . u ) , we also obtain IK« . . . a I ^ l i / | . . . 1/ , 1 . / /

We will show that every subgroup of F(X) has a Nielsen reduced

basis. We will need to define a fairly complicated ordering of the reduced words.
We begin with an arbitrary well-ordering < of X. Let
£l
w- x • . . . x • Bn and v~ x • l . . . x m
be reduced words. Write w << v if either
n<m or n m and there is some k such that ifmJr
m a
nd s - 8 for r<k and either
:
x• x- or /, - j , and e ^ " l and 5, - - 1 . It is easy to see that if w<<v then
uw<<uv provided that uw and uv are reduced as written and wu«vu^ provided
that \u\ ^ \uA and wu and vu* are reduced as written. The relation << is a
well-ordering, since given any set S of words we can look at those members of
S of smallest length and among them look at those with smallest first letter,
then the subset of these with smallest second letter and so on.
We define the left half h{w) of a word w of length 2n to be its
initial segment of length n9 and the left half of a word of length 2/z + l to be its
initial segment of length « + l . Finally, we define w<v if Iwl < Ivl or Iwl-lvl and
either h(w)« h{v) or h{w)~h(v) and h(w~ )<< /z(v" ). Plainly < is a well-ordering.

Proposition 8 Let G be a subgroup of F{X). For a*G, let GQ be

<g(G; g<a>. Let A be [at G; ai G }. Then A is a Nielsen-reduced set of
generators for G.
Theorem 9 (Nielsen's Theorem) Any subgroup of a free group is
free.

Proof The theorem is immediate from the proposition and Lemma 7.

We first show, by transfinite induction, that G~ <A>. Take any a t C7,
and suppose that gt<A> for all g<a. Then Ga£<A>, so ae<A> if azG , while
aeA by definition if ai G .
We now show that A is Nielsen reduced. Since 1 e <0>, we see that

1 / A. Suppose that we had distinct x and y in A such that \x£y I < \y\ for some
choice of E and 8 as +1. Since x r <xEy ,y> and yt <x^y ,JC>, this contradicts the
assumption that x * A if y< x and the assumption that ycA if x <y. Thus
condition (Nl) holds for A.
Since (Nl) holds, (A/2) will also hold unless we have elements JC, yy
and z in A with x ?*.y ^ e such that xE - up , y ~ pq , and z^ " qv> all reduced
as written, for some choice of e, 5, and r\ as ±1, where there is no cancellation
between p and v or between u and <? , and \p\-\q\ and Ipl^lal and \p\<-\v\.
Taking inverses if necessary, we may assume that 8 - 1 . Since yzA, we have
y<y , and so p<< q.
If T]=l w e h a v e y<z a n d yzmpv<Z, c o n t r a d i c t i n g z * A. If T J - - 1 a n d
m
\y\ - \z\ w e h a v e y z a n d zy v p ' v q -<:, with the same contradiction.
S u p p o s e that T J - - 1 a n d \y\ - 1^1, a n d s o \p\ - Ivl. If p< <v we again have y z
a n d zy < z, w h i l e if v" < < p w e h a v e zKy a n d zy <y* t h i s t i m e c o n t r a d i c t i n g
yiAJI

We shall later simplify this argument, obtaining a basis for G which

satisfies a condition weaker than being Nielsen reduced. While we have now
proved that every subgroup of a free group has a Nielsen reduced basis, the
current argument is not entirely satisfactory, since it does not give a
straightforward method of finding such a basis when G is <iS> for some given
finite set S. We now show how to do this.
It is often convenient to regard a finite set S as having its
elements given in some specified order, and we shall assume this done without
further mention.
Let S be the ordered set ( S j , . . . ,5^}. The elementary Nielsen
transformations on S are the following: (Tl) replace some s- by s-~ ,
(T2) replace some s- by s-s, for some k*i, (T3) delete s. provided that 5 . - 1 ; in
I I K. L L
all three cases 5- remains unchanged if j&i. A product of elementary Nielsen
transformations (on different sets) is called a Nielsen transformation, and a
Nielsen transformation is called regular if it does not involve any transformation
of type (T3).
It is easy to check that each transformation of type (Tl) or (T2) has
an inverse of the same type (applied to a different set!). It is also not difficult
to show that every permutation of 5* is a Nielsen transformation, as are the
transformations fixing 5- for j£i and replacing 5^- by any of s-£s^ or js^ s-Ey
where e and 8 are ±1. We shall call any product of the latter transformations for
fixed / (but varying k) a regular Nielsen transformation on S fixing S-(s^).

If T is obtained from S by a Nielsen transformation then <7> - <S>,

and if T is obtained from S by a regular Nielsen transformation then T is a
basis for <S> iff 5* is a basis for <S>. These results are obvious for an
elementary Nielsen transformation, and follow by induction for the general case.

Theorem 10 For any finite subset S of F(X) there is a Nielsen

reduced subset T which comes from S by a Nielsen transformation.

Remark Our proof will show how to find the elementary Nielsen
transformations which obtain T from S.

Proof If 1 € S9 delete it. If there are JC and y in S such that

\x*y I < \y\ (or < IJCI, similarly), then there is a Nielsen transformation which
replaces y by xBy .
Suppose that neither of the first two possibilities occurs. Replacing
elements of S by their inverses if necessary, we may assume that s<s for all
s € S. Suppose that this has been done, and that there are JC, y and z in S such
that x*y~ *z with x£ - up> , y m
pq , and z^mqv, all reduced as written, for
some choice of e, 8, and TJ as ±1, where there is no cancellation between p and v
or between u and q , and \p\ - \q\ and \p\^\u\ and Ipl^lvl. Taking inverses if
necessary, we may assume that 8 - 1 . Since y<y , this gives p< <q. If TJ-1 we
replace z by yz, noting that yz<Z. If T)--1 we replace zmv q by zy - v" p" ,
noting that zy < Z-
By construction, if none of these operations can be performed then
S is Nielsen reduced. Hence it is enough to show that by repeatedly performing
these operations we ultimately reach a set to which none of the operations can
be applied.
Deleting an occurrence of 1 reduces the number of elements of the
set. The second kind of operation reduces the sum of the lengths of the members
of the set. The third and fourth operations do not alter either of these two
numbers. Hence the first two operations can only be performed a finite number
of times in all. So we need only show that between any two operations of the
first and second kinds there can only be a finite number of operations of the
third and fourth kinds.
We may define an ordering < on ^-tuples of elements by
( 5 j , . . . , 5 / I ) < ( r 1 , . . . , r / z ) iff there is some k with s^ t- for i<k and sk<t^. Since
< is a well-ordering of elements, the ordering < on ^-tuples is also a
well-ordering. Now operations of the third and fourth kinds by construction

replace an /z-tuple by another /z-tuple which is smaller in this ordering. Hence

only a finite number of them can be applied before no further such operation
can be used.//

Lemma 11 Let Y be a Nielsen reduced basis of F{X). Then

Y-AuB, where AcX and B~l - X-A.

Proof Any element x of X can be written as the product of

elements of YvY . B y Lemma 7, such a product (assumed not involving any
yEy'E) must have length 1. That is, XcYuY' , from which the result is
immediate.//

Any regular elementary Nielsen transformation on X extends to an

automorphism of F(X), which we call an elementary Nielsen X-automorphism.

Proposition 12 Let X be finite. Then the elementary Nielsen

X-automorphisms generate the group of automorphisms of F{X).

Proof Let a be an automorphism of F{X)y and let Y be X<x. By

Theorem 10 and Lemma 11, using the length function given by the basis Y rather
than that given by X, X can be reduced to a permutation of Y by applying
elementary Nielsen transformations. These must all be regular, since otherwise
F(X)m F(Y) could be generated by fewer than \X\ elements.
At first glance this seems to prove the result. However, the Nielsen
transformations applied are not Nielsen transformations on X, so we have to be
more careful.
There are ordered subsets X - X^,..., X , Y of F(X) such that, for
l * / < « , X-^ is obtained from X- by applying a regular elementary Nielsen
transformation and Y is a permutation of X .
Now any permutation is the product of transpositions. We show that
any transposition on a set U can be obtained as the product of regular
elementary Nielsen transformations.
Consider the sequence of pairs

There are elementary Nielsen transformations which send each pair onto the next
one and which fix u,K for i^kitj. The product of these will transpose u-i and u..
j

It follows that we can find Xn+^9... 9Xm'Y such that X / + 1 comes

from AT. by a regular elementary Nielsen transformation for all i<m.
We show, by induction on r9 that Xf- XOLF for some automorphism
a which is the product of elementary Nielsen ^-automorphisms. In particular,
X*m-Y-Xoif and so a"« m > since ocâ is an automorphism of F(X) which
fixes X.
Let Xf be iu^9...9u^ and let ATf+1 be ( v j , . . . ,v^}, the order being
taken into account. Suppose that v^.- UM. and v^- u^ for k*i (the other case is
similar). Let p.. be the Nielsen AT-automorphism which maps x^ to x^x. and
fixes the remaining elements of X. By assumption, u^ - .xâ,. for all k. Hence, for
k*i9 vk- xk<xr- xkf>;f*r, while v^.- (xa^XjOiJ- (JCÔ^- xfiij*r. It follows that
we may define a f + j to be (J.-a^.//

We should investigate the relationship between Nielsen's method

and Schreier's method. It turns out (Lyndon 1963) that Nielsen's method can be
applied to any group G which has an abstract length function satisfying certain
conditions (in particular, when G is a subgroup of a free group F with the
length function on G induced from that on F), and that given such a length
function on G we can always embed G in a free group so that the given length
function is the induced one. Chiswell (1976a) also showed how, given such a
length function on G9 we can obtain a tree on which G acts freely. This would
again show that a group with such a length function is free, and the basis
obtained from Nielsen's approach is that obtained from a suitable representative
tree; equivalently, that it is obtained as in Schreier's method. The details of this
interesting material would take us too far afield.
Reducing Schreier's method to Nielsen's is easier, as is shown by
the following result.

Proposition 13 Let G be a subgroup of F(X), and let T be a

Schreier transversal for G such that each representative is of shortest length in
its coset. Then the set of non-trivial generators of G obtained from T is
Nielsen reduced.

Remark We know that such a Schreier transversal can be obtained.

Proof Let t e T and x € X. Let u represent the coset Gtx. We first
show that txu is reduced as written if it is not the identity.
If x cancels into t then t is VJC , reduced as written. Since T is a
Schreier transversal, v e T, and so we have u - v since v € Gtx. Similarly, if JC

cancels into u then u is WJC, reduced as written, and then we7\ which gives
w-t since tzGux: . In both these cases rjcw" - 1.
Now we have Iwl^ 1/1 + 1, since /\x and u are in the same coset, and
u is the shortest element of its coset, and, for a similar reason, we also have
l/i£lal + l. It follows that in txu , the element x is the middle element when
txu has odd length and is one of the two middle elements when txu has
even length.
Notice that the inverse of a generator is of the form px q \ where
p and q are in T and qzGpx .
Now suppose that we have two non-identity elements tx£u and
py q which are generators or inverses of generators. It is enough to show that
in the product txsu~ py q cancellation cannot reach the factors JCE and j ; unless
the two elements are inverses of each other. For this property, together with the
fact that JC£ and y are (one of the two) middle elements is equivalent to
Nielsen's conditions.
Suppose, if possible, that this does not hold, and assume that
cancellation reaches JC£ first. We cannot have u - p, as y cannot then cancel
against xs unless our elements are inverses of each other. So, by assumption,
ux~s must be an initial segment of p. Since T is a Schreier transversal, this
means that ux £ e T. Since t* Gux e and t* T, we find /- ux £, contary to our
assumption that tx£u~ ±\.ll

We now look: at a condition on subsets of F\X) which is weaker

than being Nielsen reduced, but which enables us to prove similar results.
Definition Let U be a subset of F(X) with \4V. We call U weakly
reduced if for every n and every u^,...un in UuU with u-^^u-' for ail /
we have \uUr.u^1 . . . un I ^ \u*
1
... n
u I. Notice that, looking at inverses, we must also
have I«Q. .. H I .* IMQ. . . H^_JI. It follows inductively that \u^... u^'^\u-\ for all /.

By Lemma 7, any Nielsen reduced set is weakly reduced. In the free

group F{a,b,Cid) the subset {ab ybc ,cd ) is not Nielsen reduced, but later
results will show that it is weakly reduced. Plainly, any weakly reduced set is a
basis of the subgroup it generates.
For any v and w in F{X) we define the distance d{v,w) between v
and w to be the length of the maximum common initial segment of v and w.
This equals the amount of cancellation between v and w , and so equals
(Ivl + lwl- lvw"1l)/2. Plainly,

if d(u,v)>m and d(v,w)>m then d(u,w)^m. (1)

Since d(u,v ) is the length of that portion of v which cancels in

reducing uv and d(v,w~ ) is the length of that portion of v which cancels in
reducing vw, we see that

if d(uyl) + d(v.w~l)±\v\ then lavwU \u\ + Ivi/I - Ivl, (2)

since the whole of v must cancel. Also,

if d{uyX) + d{v,w~X)< Ivl then ^ ( w v . w ' V ^ v . w " 1 ) , (3)

since in this case the whole of v does not cancel in reducing uvw. Further,

if M - l v l and d(u,v) + d(u ,v~ ) ^ \u\ then u~v (4)

Lemma 14 L e r b^, ... ,bn be in F{X), and write a.~ b^... b. ^ for
0 < i z n, and c - - b.- ... b for 0 ^ i < n. Suppose that, for 0 < i < n, we have
l<3 la l<3 an< lc lc Su se
/+l' ^ /'' /+l' ^ ' ^ / ' ' * /-l' ^ /'* PP° father that \b^ . . . b^ < \b^ . . . b^\.
T h e n , for 0 < i < n, \ a . + J ~ \ a - \ and I c ^ j i - Ic-I.

Proof T h e r e is n o t h i n g to prove if AZM, SO take n>\. By h y p o t h e s i s ,

for Q<i<ny w e have \a.+^\ + \c-_^\^\a.\ + \c-\ w i t h e q u a l i t y iff l a . + j l - l a . l and
\c• J - Ic-I. We shall show i n d u c t i v e l y that w e have e q u a l i t y for all /.
So w e suppose that w e have e q u a l i t y for i < r. W e will a s s u m e that
we have strict i n e q u a l i t y for imr and derive a c o n t r a d i c t i o n . Since a + , - a b and
t ^ j - brcf, w e have

{\a\ + 1^1 - l a ^ l ) + (1^1 + Ic^l - \bc\) < 2\b\,

by the assumed inequality. That is, d{ar,b~ ) + d{br,c~ )<\br\. Consequently,

d(arbr,c~l)m d{br,c~l); that is, \arbf\ + lcfl - \abrc\ - 1^1 + Ic^l - l ^ c H . Since we are
given that l a ^ U 1^1, w e see that \abrc\^\brc\. N o w a ^ c ^ - 6Q . . . 6 , while
/JC u" ! . Since our c h o i c e of /• tells us that \c J - lc n l and c A = ^>, . . . b this
contradicts the h y p o t h e s e s . / /

Lemma 15 Let U be weakly reduced, let v satisfy Ivl^lwl for all

utU, and suppose that Uviv} is not weakly reduced. Then there is a regular
Nielsen transformation on UuM which fixes U and replaces v by an element of
shorter length.

Proof Since UuM is not weakly reduced, we can find n and

bo,...9bn in UvU~l v[v,v~l), with b^tbf1 for all /, such that
I £uA . . . bn\ <\b*l ... bn I. Take n minimal subject to this condition. Then no subword
or its inverse satisfies the corresponding inequality. In particular, the hypotheses
of Lemma 14 hold. Hence ICQI-IC^I- . . . " • c /2 .^» where c-m b-^ . . . b .
Now b.t{v,v } for some /, since U is weakly reduced. If there is
only one such /, then we have a regular Nielsen transformation on UuM which
fixes U and sends b- to b^...bn (this works whether b- is v or v ). By
hypothesis, Ivl ^ I£H. Now bnmcn.\> anc
* CQ " ^\ • • • t>n- Since I c^_ ^ I - ICQI, our
Nielsen transformation replaces v by a shorter element.
So it is enough to show that there is at most one i with b-t[v,v }.
Suppose that there are at least two such values, which we denote by r and s
where r< s.
Since \b \^\b-\ for all /, we see that, for 0<i<n,
d(ci_lici)'(\ci_l\ + \ci\-\bi\)/2^(2\c()\'\br\)/2. It then follows that for any / and j
we also have d(c,c)^{2\c^\-\b\)/2, by (1). Also
d(c[X,bt) - (\Cj\ + 1^.1 - lc M l)/2 - 1^-1/2.
There are a number of cases to consider. First suppose that br~bs
and s<n. Applying the final result of the last paragraph both for i:- r and for
i-s, and then using (1), we see that d(c ~ \ c " 1 ) ^ \b 1/2. Thus
- 1 - 1
d(cr,cs) + d(cr ,cs ) ^ ICQI - \cr\ - \cg\. By (4), this gives cr'Cg, and so
b
r+\" ' bs~^' T h i s c o r i t r a c l i c t s t n e minimality of n.
Next suppose that bf'bn. Since c/z_i*^/7» w e n a v e
1 1
^ ( c ^ , ^ ^ ) - ^ ^ " , ^ ) - 1 ^ 1 / 2 , while J(c f ,c / z . 1 )^(2lc ( ) l-l^l)/2-
( 2 1 ^ ^ ! - \b\)l2~ \b\l2. By (1), this gives d{cr,c~l) > \b\l2. Since br-bn, by
c
assumption, and ^ * ^ _ p by definition, and lc^_jl * Ic^l, this simplifies to
lcf I ^ IcM. This can only hold if c ^ - l , which again contradicts the minimality of
n.
F i n a l l y , s u p p o s e that bf~b~ . We must then h a v e r<s-\. Now
^ c ^ , " 1 ^ ~l)~{\cA + \b\- \c \)/2±\b \/2= \b 1/2 (here, if 5 - n we define c t o be
1). A l s o I^QCQI < ICQI by h y p o t h e s i s , w h i l e l^-cJ - lc^_jl - IcJ for 0 < / < « . Hence
^ ( c / . " 1 , ^ r ) - ( l c f l + 1 ^ 1 - 1 ^ 1 ) 7 2 >-\b\l2. By (1), this gives d(c~X%cs^X)'±\b}l2. Also,
- l ^ l ) / 2 . Hence d{cf ,0^) + d(cr~\cs_{1) * \c\ - lcfl = I c ^ j l . By (4)

this gives c - c_ , , leading to the same contradiction as before.//

T S" 1

We now have a simpler and stronger version of Proposition 8.

Theorem 16 Let G be a subgroup of F(X). Take any well-ordering
of G such that g<h whenever \g\<\h\. For any atGJet GQ be <biG; b<d> and
let S be {atG; a 4 G }. Then S is a basis for G.

Proof As in Proposition 8, S generates G. It is enough to show that

S is weakly reduced.
Suppose not. Then there is a least element v c A such that the set
{atA'9azv} is not weakly reduced. In particular, letting U be {a* A\a<v), we see
that any finite subset of U will be weakly reduced, and so U itself is weakly
reduced. Also, by our choice of well-ordering, Ivl * \u\ for all utA. Then Lemma
15 tells us that there are u and w in <£/> such that \uvw\ < Ivl. Since U£Gv, this
gives v € Gv, contradicting v e A.I I

Lemma 17 Let U be a finite subset of F(X) with m elements. Then

U is weakly reduced iff we cannot find b^9...,bn in UuU with bj+i*bf
\bn\jb*i ...b\<\b*...
n i
b\n and n<2m.

Remark By definition, U is weakly reduced iff there is no n for

which there are b^,... ,bn with these properties. The point of the lemma is that
we have an explicit bound for «, and so we can write down all sequences
—i #»t

£ Q , . . . , £ ^ of elements of UuU~ with n<2 and check whether or not any of

them satisfy the conditions of the lemma. Consequently we have a finite process
for checking whether or not U is weakly reduced. In particular, this could be
used to show that the subset iab"\bc~\cd" ) of F(a,b,ctd) is weakly reduced.

Proof If U is weakly reduced then there is no n for which there are

bf, bM in UKJ If1 with b.. 1 tb-'1 and \bn... bj <\bA... bj, so there is
m
certainly no n<2 for which such elements exist.
Suppose that U is not weakly reduced. As before, we take n
minimal subject to Ib^b*...
\j i
bj<\b.
n l
... bj.
n
We need to show that n<2m.
The proof of Lemma 15 shows that w e cannot have r and 5 w i t h
r<s and £ > € { £ , A " 1 } such that \b:\*\b\ for r<i<s.
s r r i r ..
Let Wy be that member of [b- 9b- } w h i c h is in U. It w i l l be
enough to show that if n^2m then there are r and s such that 0<r<s, and
u - a and for r<i<s either u.mu or \u.\<\u I. This is in fact a property of
r s i r i r

arbitrary partial ordered sets (the partial ordering being g i v e n by wv if lul -• Ivl).
Let a be an element of U of m a x i m a l length. If a occurs twice in
the sequence u.9... ,u then we have the required r and s. If a occurs only once
(or not at all) take k such that u^m a. If k>2m~ then u^,... ,u^_^ is a sequence
of at least 2m elements of U-ia), while if k-~2m~ then u ^ % . . . ,un is again a
m
sequence of at least 2 ~ elements of U-{a). Inductively, such a sequence will
contain a suitable pair of elements uf and u^.//

8.3 GRAPHS OF GROUPS

Definition A graph of groups & consists of
(i) a connected graph A^,
(ii) for each vertex v of X a group C7y, and for each edge e of X a
group G such that C7—- C7, for all ^, and
(iii) for each edge e of A' a monomorphism from G to (7 . If we
have (i) and (ii), but (iii) is replaced by
(iii') for each edge e a homomorphism from C7 to G'
6 el

we refer to a generalised graph of groups. The groups G and 6' are called
vertex groups and edge groups of @, respectively. Note that, because G-~Ge
and e~T - ea, we also have a monomorphism (or homomorphism, in the
generalised case) from G> to Gy . We shall use T and o to denote the
homomorphisms from G to G^ and 6 ^ o . If we wish to draw attention to the
graph X we will say that & is a graph of groups on X, or that the pair (&,X) is
a graph of groups.
Let @ be a generalised graph of groups. Let E be the free group
with basis FAX), the set of edges of X. Let F(&) be the group
where N is the normal closure of the subset {e (ao)e{ai) ; ec
a^. Ge) u {ee~\e t E{X)}. Note that when @ is a graph of groups (rather than a
generalised one) F{®) is an HNN extension of *C?v with one stable letter for
each edge pair. Notice that we can regard any path in X as an element of E,
and hence as an element of F(&).
Let T be a maximal tree in X. We define n(®9X,T) to be F{&)/My
where M is the normal closure of [e\e t T)\ that is, n{&,X,T) is obtained from
F{&) by adding relations e=\ for all <? e T.
Let v Q 6K(A r ). We define 7r(@,Ar,v()) to be the subset of F{&)
consisting of all elements which can be written in the form £o e 1^1^2 ' " ' e 8
nn
such that < ? j , . . . ,en is a loop based at VQ and £Q e 6^ and g. e GQ T for 1 % i ± n.
We allow « - 0 , in which case we just have an element of G . This subset is
v
o
plainly a subgroup.

The next result is a generalisation of Theorem 5.17.

in V{X), and let T be a maximal tree of X. Then the natural map from F((&) to
TI(®,X,T) induces an isomorphism from TZ(®,X,V0) to n(®,X,T)

Proof For each veV(X) let p y be the irreducible path in T from VQ

to v. In particular, p v is the trivial path, and so, when it is regarded as an
o
element of E, is the identity. We may define a homomorphism from E* *G^ to
itself which sends g*Gv to Pvgp~ and which sends e to PeoePex~ • By von
Dyck's Theorem, this induces a homomorphism from F((B) to itself, which in fact
maps into ^ ( © ^ V Q ) . If e e T then either Pex~ Peae o r Pea~pei^ ^as Pat^s» o r a s
elements of E). It follows that our homomorphism sends any e in T to 1, and so
it induces a homomorphism %• from iz((&,X,T) to 7T(@,A',VQ). It is easy to check,
as in Theorem 5.16, that # is the inverse of the natural map from TT(@,AT,VQ) to

It follows that, up to isomorphism, n{&,X,T) is independent of T,

and IZ(@9X,VQ) is independent of VQ . If we are only interested in these groups up
to isomorphism we denote them by it(@9X) or by TT(@), calling them the
fundamental group of ©. Notice that n^{X) is a quotient of n{&).

Examples Let all the vertex groups and edge groups be trivial.
Then TT(@) is just n^X).
Let all the edge groups be trivial. Then TT(@) is the free product of
7i.{X) and all the vertex groups. In particular, suppose that we have groups G-
for / in some index set /. Let 0 be a symbol not in /. Let X be the graph
whose vertex set is /u{0}, with one edge pair (e-,e~-) for each /, where efi-Q
and e-i - /. Let & be the graph of groups on X in which all the edge groups are
trivial, G. is the given group for / € /, and GU is trivial. Then n(&) is the free
product of the G-.
Alternatively, choose some /Q C /. Let Y be the graph with vertex set
/, and one edge pair {>.,e~.} for all i^ir\, where eo^i^ and e-imi. Let ^ be the
graph of groups on Y with all edge groups trivial and with G- being the given
group for each vertex /. Then TT(^) is also the free product of the G-. The
advantage of the first formulation is that all vertices play the same role, whereas
in the second formulation we have to treat one vertex in a special way (but we
need no additional vertices).

one edge pair. Let the edge e have two distinct vertices v and w. Then 7t(@) is
the amalgamated free product of C/y and Ow with the subgroups GQQ and G x
being amalgamated. If the edge e has its two vertices the same, both being v,
then 7r(@) is the HNN extension of G with associated subgroups G o and G x.
When we have a generalised graph of groups whose graph has only one edge
pair, the corresponding fundamental groups are the pushout and the pseudo-HNN
extension.
Let © be a generalised graph of groups on X, and let Y be a
connected subgraph of X. Then <&\Y denotes the generalised graph of groups
whose underlying graph is K, and whose vertex groups, edge groups, and
homomorphisms are the same as those for @.
Let @ be a generalised graph of groups on X, and let Y be a
connected subgraph of X. Let VQCK(K), let S be a maximal tree of K, and let T
be a maximal tree of X which contains S (such a tree exists, by the corollary to
Proposition 5.8). It is easy to check: that there are natural homomorphisms from
7r(@ir,r,v0) to 7r(@,*,v0) and from n{®\Y,YfS) to TZ{®,X,T). It is also easy to
check that these homomorphisms transform to each other when the isomorphism
of Proposition 18 is applied. So we may regard these homomorphisms as going
from it{®\Y) to TC(@).

Lemma 19 Let 0) be a graph of groups on X, and let Y be a

connected subgraph of X. With the above notations, the homomorphism from
~K{®\Y) to 7T(@) is a monomorphism. In particular, for any vertex v of X, the
natural homomorphism from G to TT(@) is a monomorphism.

Proof Suppose that we have shown that for any graph of groups £
on a tree T and any subtree S of T the map from n{^\S,S,S) to n(iQ,T,T) is a
monomorphism. In particular it will follow that the map from H to 7r(£,7\7") is
a monomorphism.
Now n(®,X,T) is defined to be a pseudo-HNN extension of
n{<8\T,T,T), and this will be an HNN extension if each Gy embeds in it(®\T,T9T).
For the same reason, iz(Q)\Y9Y,S) will be an HNN extension of %{®\S,S9S). If we
also know that n(®\S,S,S) embeds in 7r(@l7\7\n it will follow from Proposition
1.34 that n(®\Y,Y,S) embeds in iz(®,X9T), as required.
So we start with a graph of groups ^ on some tree T9 and a
subtree S of T. We first consider the case when T is finite, and use induction
on the number of vertices of T.

If S - T there is nothing to prove, so assume S * T. There must be

an edge e with eo € S and ex 4 S (some edge on a path starting in S and ending
outside S has this property). By Exercise 5.1, T- {efe"} is the disjoint union of
two trees 5j and 5 2 one of which, say 5 j , contains S. Now the definition tells
us that n(%,T9T) is the pushout of Hg •* H •* it($l5j ,5\ ,5|) and
H
e "* ^ex * * W 1 ^ '^2 *S2)m I n d u c t i v e I v » Heo a n d ^ T embed in
and ^ ( ^ l ^ ,^ 2 »^2^ r e s P e c t i v e l v » a n d a l s o n(fyS,S9S) embeds in ^ j j
Thus the pushout is an amlagamated free product, and so the factors embed in
7t(^,r,r), and n($)\S,S,S) also embeds in n(%,T,T)f as required.
Next let T be arbitrary and let 5" be finite. Then 7c($,7\7*) is the
direct limit of the groups 7T(#I7*J ,7^ ,7^) for all finite subtrees 7\ of T which
contain S. Since we have just seen that ^(^l^,^,^) embeds in 7c(//l7\ ,7\ ,7\) for
any such 7^, Proposition 3.3 tells us that n($\S,StS) also embeds in
Now let T and 5 be arbitrary. Choose a vertex vc5. Then
n{$\S,S,S) is the direct limit of the groups 7r(#l5\ ,S* ,S«) over all finite
subtrees S^ of 5" which contain v. We saw in the last paragraph that
,5^) embeds in n(JQ,T,T), and so n($\S,S,S) itself must embed in
), by Proposition 3.3.//

In the course of proving the lemma, we have seen that when X is

a finite tree the group 7i(®) is obtained (inductively) by a process of repeating
amalgamated free products. This can be done in various ways, since the removal
of any edge-pair gives two disjoint trees smaller than X9 and 7i(@) is the
amalgamated free product of the fundamental groups of the restrictions of @ to
these trees. In particular, by Lemma 5.10, we can remove an edge-pair in such a
way that one of the trees consists of a single vertex.
We have also seen that when X is an infinite tree then 7t(@) is the
direct limit of groups 7t(@iy), where Y is a finite subtree of X. When A" is a
tree, we often refer to n(@) as a tree product of the vertex groups C7y.
For arbitrary X, we know that n((&) is an HNN extension of n(&\T),
where T is a maximal tree of X. Hence TT(@) is an HNN extension of a tree
product, and we often refer to it as a treed HNN group.

Lemma 20 Let & be a graph of groups., and let H be a group. /

Suppose that there are homomorphisms cp :G -» H for each vertex v, and
elements QLeO of H for each edge e such that OL— -1 and, for all g^G^,
^ e -ea e
<*e~ (^ O( P eo ) a e " ^)^ei - Then there is a homomorphism from 7i(@) to H whose
restriction to Gy is a conjugate of cpy.

Proof By von Dyck's Theorem, if a^ - 1 for e e T then there is, as

required, a homomorphism from n(@,X,T) to H which is <py on G^ (and which
sends the generator e to a ).
In the general case, we will show how to find elements 3^ of H
and homomorphisms ty^.G^ H such that 4>v is a conjugate of <py and
3 (£oi[; )3 " grip for g<-GQ, and with 3 " I for e e 71. By the previous
paragraph, this will prove the result,
Let p be the path e^,... %en. Let a be a^ . . . a^ . Choose some
vertex v^, and, for each vertex v, let a be a where p is the irreducible path in
-1
T from v0 to v. Let 3 be a a a , where ^o - v and ex- w. Plainly, 3 " 1 for
e € 7'. We then define tpv by g^v - a v (^9 v )a y for gzG^J/

An isomorphism from the generalised graph of groups 0 on X to

the generalised graph of groups $ on Y consists of an isomorphism cp from X to
K, for each vertex v of X an isomorphism (still called cp) from G to H , and
for each edge e of X an isomorphism from G& to // such that the
isomorphisms from G^ to H and from G- to H— are the same and the
homomorphisms
F Ge -* Gex ~* H
exq>
and eG -* t?cp
H -*e^cpi
// are the same. Plainly
J an
isomorphism from @ to ,^ induces an isomorphism from TT(®) to 7i(^).
The generalised graph of groups $ is conjugate to the generalised
graph of groups & if they have the same underlying graph, the same vertex
groups and the same edge groups, and the homomorphism from H& to Hex is the
homomorphism from G? to Gf followed by conjugation by an element of Gf .
It is easy to check that conjugacy is an equivalence relation between generalised
graphs of groups. When fa is isomorphic to a conjugate of @ we say that $ is
conjugate isomorphic to ©.

Lemma 21 Let & and $ be conjugate generalised graphs of groups.

Let T be a maximal tree of the underlying graph X. Then there is an
isomorphism $:n{(8,X,T)^ n{$fr,X,T) such that for each vertex v of X there is
an element c of n{$,X,T) with JC^-C^ xc^ for all x e G . Further, given a
vertex v^ of X, we may take c to be I.
o
Remark Strictly speaking, we should write xi&-cv~ xjcv, where /
and j are the natural maps from C/y into 7i(@,A',r) and 7r(.^,Ar,7'O. However, / and
j are inclusions when we have graphs of groups rather than generalised graphs of
groups. Even in the general situation, it is more convenient not to mention i
and j explicitly, though the reader should be aware that strictly speaking they
are necessary.

Proof The previous lemma applies, where each cp is the inclusion

of Gv into 7i(£,AT,r), to give a homomorphism $:7r(@,A\7*) -» n($),X,T). Similarly,
we may construct a homomorphism from n(Sfr,X,T) to n(®,X,T). These two
homomorphisms are plainly inverses of each other.//
Let ® be a (generalised) graph of groups on X, and let X' be the
barycentric subdivision of X. Thus X' has as vertices the vertices of X together
v v
with new vertices e ' e~ f° r eacn
edge-pair {e,e~), and has two edges e_ and e^
for each edge e of X, where e_ starts at eo and ends at v and e+ starts at v
and ends at ex. We define a (generalised) graph of groups &' on X' as follows.
The vertex group of @' at the vertex v is Gv and its vertex group at vQ is G .
The edge groups of © ' at the edges e_ and e+ are both G . The homomorphisms
from an edge group of @f to the vertex group at v ? is the identity, while the
homomorphisms from an edge group to the vertex group at a vertex of X are the
corresponding homomorphisms for ©. It is easy to check that 7r(@f) is isomorphic
to 7r(@). The purpose of using @' is that X' contains no circles, so the
proposition below can be used.
We now show how generalised graphs of groups can be associated
with certain complexes. Let AT be a connected complex which is the union of
connected subcomplexes ATy for v in some index set V, such that K n K^n Kwm&
for distinct u,v, and w. Let AT be a graph with V(X)m V and with one edge-pair
joining v and w for each component of K n K . Thus X has no circles, and X is
connected because K is connected. We denote by K& the component of K^ n Kw
which corresponds to the edge e. Let @ be the generalised graph of groups on X
such that G mnJK ) and G -7Uj(A'>) for each vertex v and edge e of Xs with
the homomorphism from G to G being induced by the inclusion of K in
K , followed if necessary by a map induced by change of base-point. Thus & is
determined up to conjugate isomorphism, since the map induced by change of
base-point is determined up to conjugacy.

Proposition 22 Any generalised graph of groups whose graph has no

circles is conjugate isomorphic to a graph obtained from a complex by the
procedure above.

Proof For each e t. E{X) take a complex Ke with ^ ( A ^ ) isomorphic

to Ge, and such that K- is isomorphic to Ke for all e. For each vertex v there
is, by the corollary to Theorem 5.19, a complex ATy with ^ ( A ^ ) isomorphic to Cry
such that Kv contains disjoint copies of those complexes Ke for which ex - v and
also such that the homomorphism from G to G is obtained by inclusion of K

in Kv followed by a map induced by change of base-point.

Let K be obtained from the disjoint union of the K by identifying
ATe— with Keo for all e. Then KvytnKw is the disjoint union of those Ke for which
e joins v and w, and so K n Kv n Kw - 0 for distinct u9 v, and w. It is easy to
check that the generalised graph of groups constructed from K is conjugate
isomorphic to ©.//

Proposition 23 Let @ 2>e Me generalised graph of groups

associated with a complex K by the procedure above. Then n(&) is isomorphic
to n^K). In particular, X is a tree if K is simply connected.

Proof Since 7Tj(AT) is a quotient of 7i(@), the second part follows at

once from the first part.
Let T be a maximal tree of X. Let L be the complex obtained from
the disjoint union of the complexes Ky by identifying KQ and AT— for all e € T.
Then K is obtained from L by identifying Ke with K- for all e 4T.
Notice that if e and / are distinct edges and f?e~ then AT^.nAT^-0.
For either / and e have three or four distinct end-points, in which case the
result holds because Ku n ATy n /fw - 0, or else / and e both have v and w as
end-points, in which case K~ and K are distinct components of ATy n AT .
By van Kampen's Theorem, and the usual direct limit argument,
7tj(Z,) is isomorphic to n(®\T,TtT). It then follows by Proposition 5.21 and a
direct limit argument that nAK) is isomorphic to n{@).//

8.4 THE STRUCTURE THEOREMS

Let the group G act on the connected graph X (we will primarily
be interested in the case when A' is a tree, but the construction applies
generally). Let Y be G\X, p the projection, T a maximal tree of K, and A be
orientation of Y. Let j:T-*X be a map such that jp is the identity on T (we
have called Tj a representative tree for the action).
We extend j to a function from Y to X. This function, which we
still denote by y, will not be a map of graphs, and Yj will not be a subgraph of
X. Since T is a maximal tree, j is defined already on all the vertices. Let
e c A - r. Since p is locally surjective, there is at least one edge x of X such
that xp~ e and JCO- eay. We define ej to be such an edge, and we define e~j to
be e j . Notice that j is still an injection.
For etA we have ejipm ejpx - ex - exjp. It follows that there is some

Y e G such that eji - yAey). For e z AnT we have ejx - eiy, and we choose y to
e
* _i *

be 1 in this case. We also define y — to be y e •

The stabiliser stabjc of a vertex or edge x of X is defined to be
the subgroup {& gx - JC).
We can now define a graph of groups & on K as follows. For any
vertex or edge y of K, we define G to be stabOv). We need to define
monomorphisms from G to G and to G for each edge e € A.
For any e e/I we have eja-eaj. Hence stab(£/)£Stab(eay); that is,
G c G , and the requiredM monomorphism is this inclusion. Also
e — eo* .
stab(ey) £Stab( eyi) - ye{s\zbe\j)y ~ . We define the monomorphism from Gg to GQ^
to send g to y^ 87e •
One of our main theorems, to be proved later, is that G is
isomorphic to TI(($) when I is a tree.

and let T be a maximal tree of Y, and let A be an orientation of Y. Let C7 be

We proceed to define a graph )^, which we call the universal cover

of the graph of groups &. We consider pairs (g,v) where geG and v is a vertex
of Y. We say that (/z,w) is equivalent to (g,v) if w- v and htgG . This is plainly
an equivalence relation, and the vertices of f are defined to be the equivalence
classes lg,vJ. Similarly, we consider pairs {g,y) and {g, y~), where g t G ' a n d yzA.
We say that (/z,z) is equivalent to {g,y) and {h,z) is equivalent to (g,jT) if z my
and hig(Go). This is an equivalence, and the edges of Y are defined to be the
equivalence classes Ig, y] and Ig, jTL We leave the definitions of o, T, and ~ on
V^ for the moment.
We define p:Y'->Y and y.Y^Y by [gj;]p-j; and yj*L\,y] for any
vertex or edge y of V. We make G act on V in the obvious manner; that is,
gih,y] ~ lgh,y] for any vertex or edge y of Y. This is easily seen to be
well-defined, and to be an action without inversions. Also we have Lg,y} - g{yj).
We define Lg,y] to be CgJT]. For j>e^4 we define igjio to be L^,^oJ
and lg,yh to be [©vyr! (here y is being used both for an edge of Y and for the
corresponding element of G). It is easy to check that, because y \G o)y G T,
these definitions depend only on [gj>] and not on the particular choice of g.
Thus they make ¥ into a graph, and p into a map of graphs. Also, because the
group element y is 1 for yi T, j:T-*¥ is a map of graphs, though j:Y-*¥ is not
a map of graphs.
It is convenient to have an expression for lg,ylo and lg,yh that can

be used whether or not ye A. Recalling that the group element y equals y~ , it is

easy to check that we always have [&j>]o - hL\,yo~\ and [gjdi - AyCljrr], where hm g
if ye A and h-gy~* if yiA.
Evidently, for any vertex v, g[l,v] - Cl,v] iff g*Gv, while for ye A we
have gtlji-lljf] iff geGo. Since o:G -» C7 is a monomorphism, it is easy to
check that the graph of groups associated with the action of G on f is
isomorphic to @ itself.

We now proceed to show that f is a tree. The proof will require a

considerable amount of discussion. We begin by showing that ? is connected.
Let Z be the subgraph of f whose edges are yj and y~j for all ye A
and whose vertices are the endpoints of all such edges. Then TjsZ, and every
edge of Z has at least one vertex in Tj. Thus Z is connected. Plainly ¥ - GZ.
Let g*Gv for some vertex v. Then gCl,v] - Cl,v], and so ZngZ*Q>
since ll,v1eTJ£Z. Also, for any ye A we have Clodt- lyjn] -ytljnl This shows
l
thatZnj>Z*0, and so also Zny~ Z*0.
It follows by induction on n, using the remark before Lemma 5.1,
that for any n and any s^,... ,sn in U((7V; all vertices v}u{y, y ; all j>€y4}, the
subgraph Zu SjZu s^s^Zv . . . u 5 j . . . $^Z is connected. Hence the union of all
these subgraphs is itself connected. Since G is generated by the subgroups Gv and
the elements y9 this subgraph is just GZ, as required.
The fact that f is a tree, and not an arbitrary connected graph, is
equivalent to a reduced form theorem for G. There is no theoretical difficulty in
this, since we already have reduced form theorems for amalgamated free
products and HNN extensions, and G is built up out of these. However, the
reader will easily see that such a reduced form theorem is likely to be
notationally complicated. Hence it is better to find a proof not using such a
result (which can then be used to obtain the reduced form theorem if one
wishes).
The easiest method is to use the results of Propositions 22 and 23.
First note that we may assume that Y has no circles. For it is not difficult to
see that the graph Z associated with the graph of groups @f is just the
subdivision of ?, and hence f is a tree if Z is a tree.
Let Y have no circles. Let K be the complex associated with & by
Proposition 22. Let p:£'-* K be the universal cover of AT. Then f? is the union
of the components of the complexes K^p , and we may obtain the graph of
groups (&,W0 associated with this expression for £. By Proposition 23, we see

that W is a tree. It is easy to check, using Exercise 6.5, that W is isomorphic to

?\ showing that V is a tree.
While this is probably the shortest proof, it is not in the spirit of
this chapter, where we want to talk about graphs and not about arbitrary
complexes. Consequently I give another proof, which may be longer than proofs
using the reduced form theorem but which avoids to some extent the notational
complexities of such proofs, though it is still quite complex.

To begin with, we look at the case where @ is the graph of trivial

groups on Y. In this case, we just refer to Y as the universal cover of Y. (It is
the same as the universal cover discussed in Chapter 7, but we do not need the
results of that chapter here.)
Let ig, ,e*\,. .. ,[g ,£? ] be a loop in Y based at Cl,vl. It is enough
to show that there is some / such that e;*me~-. For then, because the vertex
groups are trivial, it is easy to see that, because lg.+^te-+y]omlg.,e.]Tt we must
also have g * ~g -, and so the path will be reducible.
Let h . ~ g . if e-*.A a n d h - ~ g - e . if e . * A . T h e n [g.fe-h'h.e.\.\,e.T\
and \.g, ,e Jo - h\ \,e oJ. Thus we have, since the vertex groups are trivial,
/z-,|
/+l - h-e
/ /• for /• n, and he
n n - hx\ . That is, e« . . . e?in - 1 in O. Since O is free with
basis the set of edges of A not in 7\ we have three possibilities. Either there is
some r with e^ = e~r or there are r and s with /•+ 1 < s such that esm er and e.t T
for r< i s or e-( T for all /. In the latter two cases we have a non-trivial loop
in 7\ and so, as this loop is reducible, there must be some / with e-+.-e~-9 as
required.
Thus the universal cover of any connected graph is a tree. Let Z be
the universal cover of K, and let p:Z-> Y be the projection. Because the vertex
groups are trivial, p is locally injective. Let a be an automorphism of Y, and let
z and z' be vertices of Z such that z'pm Zpoi. Then Lemma 4 may be applied to
p and POL to give a unique automorphism P of Z such that z$ - z' and (3/?= pa.
We call (3 an extension of a.
We now return to the general situation. Let W be the universal
cover of V\ and let q be the projection. There is a map m:Tj-» W such that mq
is the identity on Tj. For ytA, we also define yjm to be the unique edge of
W starting at yojm and mapping onto yj by q. We also define y~jm to be yjm .
Denote jm by k. Although k is not a map of graphs, we have ykomyok for all
y*Ay by construction.
Let v be a vertex of Yy and let g^G^. Then g gives an
automorphism of f fixing vj, and this extends to an automorphism of W fixing

vk. We denote this automorphism by <p and we will write it on the left. If h is
also in Gv we see that both 9 <P^ and <p ^ fix v/: and extend gh. By uniqueness,

Now take any j;« .4. The element j> of G gives an automorphism of
f which sends yij to yji. This extends uniquely to an automorphism 9 of W
such that cp sends yr& to ykx.
Let j ; be in A and g in Go. Then (cp (yk))o - 9 Ayko) - 9 (yak) - yok
(by the definition of cp ) which equals j;X:o. Also (cp (yk))q- g{yj) - yj, because
o o
gtGo, and this equals {yk)q. Since 4 is locally injective, cp fixes yk.
y o
Let y be in A and A in O' Then cp^. extends hi and fixes yzk,
since hi ( G . We have seen that <p, fixes ,y&, and consequently fixes ^ x . Thus
9 <P^ 9 fixes yik (remember that these automorphisms act on the left!) and it
also extends y (ho)y, which equals hi in G. Then uniqueness of extensions shows
that cp <P^09 * 9/JX • Also, if y*T we have yij'yji and yzk-ykt, and so cp is
the identity.
Thus we have homomorphisms from each G into AutPV, and
elements cp^ of AutW. From the facts above and Lemma 20, these give a
homomorphism from G to AutVK; that is, they give an action of G on W. Any
element of f (whether vertex or edge) can be written as g{yj) for some unique
vertex or edge y of Y. The construction of t and the remarks made about
elements of W fixed by certain elements of G then show that the element g{yk)
of W depends only on the chosen element of Y and not on the choice of g.
Thus we have a function s:f'-» W. It is easy to check that 5 is a map of graphs
such that sq is the identity on Y. Thus Y' maps injectively into the tree W.
Since we already know that Y is connected, it follows that Y is a tree. That is,
we have proved the following theorem.

Theorem 24 (First Structure Theorem) The universal cover of a

graph of groups is a tree. 11

We have now made two constructions. The first starts with a group
acting on a connected graph and constructs a graph of groups, while the
second starts with a graph of groups and constructs a tree on which a group
acts. We need to elucidate the relationship between these two constructions.
As already remarked, if we start with a graph of groups, construct
from it a group acting on a tree, and then construct from that a new graph of
groups, then the two graphs of groups are isomorphic.

Conversely, let us start with a group G acting on a connected graph

X. Let Y be CAA\ and let p:X-*Y be the projection. Let T be a maximal tree in
K, and let /I be an orientation of Y. Let j:Y-+ X be as before. Let @ be the
graph of groups on Y constructed using 7\ A, and y". Let n denote 7r(@,K,r), and
let ¥ be the universal cover of @. Thus f is a tree on which n acts.
Plainly, X cannot be isomorphic to Y unless X is a tree. We shall
show that if X is a tree then X is isomorphic to ¥, that (7 is isomorphic to TT,
and that the isomorphisms respect the actions.
Each vertex group C7y of & embeds in G To each edge e e A we
have associated an clement y of (7, with y - 1 for e € T. Also, for e € A and
gf(7 , we have homomorphisms a:G -* G and v.G -» @ such that gamg and
£t * Ye £T^ • ^ follows that there is a homomorphism 9:7c -* G which is inclusion
on each C7 and sends e to y , for e e A.
The vertices and edges of ¥ are classes [A,j>], where /ZETT and ^ is
a vertex or edge of Y. When y is a vertex of Y or an edge in A we have
lh,y]mlk,z] iff .v* z and kihG , while if .y is an edge not in A we have
f/z,_vl - lk,zl iff r/z, jT 1 - \k,z 1. Also, for y an edge in A or a vertex we have
G - stabOy), and the restriction of 9 to G is inclusion in C7. It follows that we
may define a function (J> from f to A' by C/z, ylty - (h(p)(yj).
For £? c /4 we have L/z,eJo * IKeol and L/z,e]x - [/ze,ei] Also eyo - eay,
while ej\myeey\ and ecp - y ^ . Using these, we see easily that 4^ is a map of
graphs.
Since X - CjKy), it is immediate that ^ maps onto A' provided that 9
maps onto G. We shall show later that 9 always maps onto G.
We next show that <[> is locally injective. Take any y and z in /I,
and any h and & in TT. We cannot have L/z, y\§ - [&,^~]I)J, since ^9^^". Suppose that
[ A , ^ - [ / : , e J 4 ; and LA,^]o - [£,?]o. Then j ; - CA,^]i|;p- tk,ztypm Z and ti^ktG
G and
also (/z" ^)9 «• stabOy). Since 9 maps G isomorphically onto stab(^oy), and the
subgroup G of G is mapped identically to the subgroup stabOy) of stabO>ay),
we see that h k*G , and so lh,y\ - lk,zl-
Similarly, suppose that C/z,y~ Ity - [£,^~]I|J and C/z,3^ lo » [£,£"]o. Then
l/z,,yJ^ - ik,z\ty and C/z,^JT - [.k,z1i- As before, j ; - z and (/z /:)<p e stabCyy).Further,
(Ay) (ky)zG . Now, by construction, 6" is just stab(vy), and G is given as
y^ y •% y
a s u b g r o u p of G , o n w h i c h 9 is j u s t i n c l u s i o n . T h u s {h k)<p e G 9 . N o w 9 is
-1 -1
o n e - o n e on G a n d s o is a l s o o n e - o n e o n yGv . T h i s s u b g r o u p c o n t a i n s h ky
a n d a l s o c o n t a i n s t h e s u b g r o u p yG xy~ 'Go (as a s u b g r o u p of TT, u s i n g t h e
d e f i n i t i o n of 7r), w h i c h is j u s t G , b y t h e d e f i n i t i o n of 0. T h u s w e h a v e h keG ,
w h i c h e n s u r e s t h a t [A, jT ]-[.&,£"], as r e q u i r e d .

Lemma 25 Let H be a subgroup of G. Let X^ and X2 be

subgraphs of X such that X2cX{, GX{ - X, V{XX) c HV(X2), and
gV(X2) n K(* 2 ) -<d for gi H. Then H- G.

Proof We show first that HX^ - A'. Since A' is connected, it is

enough, by Lemma 5.1, to show that if e is an edge with eo < HX^ then etHX^.
Plainly, we may assume that eotX*. There exist £e(7and xtE(X*) with e - gx.
There are h and k in H such that A(eo) and k{xo) are in A^ . Then
hgkl(k{xo)) « hg(xo) - h(ea) f /zs£~V(A'2) n V(X2), so /zg/f1< //. Thus ge //, and
so e ' gx e //A^!.
In particular, //^(ATj) - K(*). Since ^(A' 1 )£ //^(A: 2 ), we see that
V(X)-HV(X2).
Take gt(/. Choose any vertex v of X2 . Then there will be some
he H and wf.V(X2) such that gv - Aw. Thus w / z ^' ^ v ^ l ^1 j n A " 1 ^ ^ ) . Hence
/z g(//, and so geH, as required.//

Corollary 9 maps n onto G.

Proof Let AT0 be 7y, let X* be the subgraph consisting of all edges
in Yj and all endpoints of these edges, and let H be 7rcp. Since K- CAAf, we have
X~GX*. Every vertex of Y is a vertex of T. Thus K(A'1)» K(,V2) u (eyi; e L A}.
For «?€i4 we have ejT~ye(eij)( HV(X2), since yff f //. Also, if gV(X2) nV(X2)*0
m
there will be vertices y and £ of Y such that g(yj) zj. By definition of y, this
requires that y- z. We will then have gcstabOV)- C/t , and so gt //. Thus the
conditions of the lemma are satisfied.//

We can now prove the main theorem, using the notations above.

Theorem 26 (Second Structure Theorem, or Main Theorem of

Bass-Serre Theory) The following are equivalent. (/) X is a tree, {ii) ip is an
isomorphism of graphs, {Hi) <p is an isomorphism of groups.

Proof We have seen that 9 maps onto G and consequently that ty

maps onto X. We have also seen that ip is locally injective. Thus ¥ty-±G(YJ)m X-
That is, (J> maps onto X.
It then follows from Lemma 5 that i|; is an isomorphism if X is a
tree. Conversely, by Theorem 24, A' is a tree if ty is an isomorphism.
Suppose that <p is an isomorphism. Now if. being locally injective,

will be an isomorphism if it is injective on V(f). So let CA,v]^- [&,w]i|;, where h

and k are in n and v and w are vertices of Y. Then (/t<p)(vy)- (&<p)(wy). Applying
A we see that v-w. Also (h k) is an isomorphism. Let h be in kenp. Take any
vertex v of Y. Then Lh,vty~ (A<p)(vy)- vy- Cl,vDip. Hence [A,v]-Cl,v], and so
A e C?v- stab(vy).Since we know that cp restricted to (7y is just inclusion, this gives
A- 1, as needed.//

Exercise 1 Let @ be a graph of groups on K, where Y has only one

edge (it may have one or two vertices). Show explicitly, using the Reduced
Form Theorems for HNN extensions and amalgamated free products, that 1? is
a tree.
Exercise 2 A sketch argument using covering complexes was given
to show that 1? is a tree for any graph of groups on any graph Y. Fill in the
details of this proof.

8.5 APPLICATIONS OF THE STRUCTURE THEOREMS

Evidently, if a group acts on a tree then every subgroup also acts
on a tree. By the Structure Theorems, we see that any subgroup of the
fundamental group of a graph of groups is itself the fundamental group of some
graph of groups. However, in this generality the result is vacuous, since every
group is the fundamental group of a graph of groups; we need only take the
graph to consist of one vertex and no edges, with the vertex group being the
given group. To obtain a significant result, we need to be more precise about
the relationship between the original graph of groups and that corresponding to
the subgroup.

Theorem 27 Let (<8,X) be a graph of groups, and let H be a

subgroup of 7i(@). Then //-7r(&), where the vertex groups of fy are HngGvg
for all vertices v of X, where g runs over a suitable set of (H,G ) double coset v
_i
representatives, and the edge groups are HngGeg for all edges e of X. where
g runs over a suitable set of {HfGQ) double coset representatives.

Proof We know that 7i(@) acts on the tree X\ and that the vertices
<>, -i
of X are [g,v] with stabilisers gG g , where g is given up to its coset gG . The
//-stabiliser of Cg,v] is HngGvg -1, and Cg.v] and C^ ,v] are in the same //-orbit iff
g and £j are in the same {H,GJ double coset. We know that //-7i(^), where the

underlying graph of $ is H\Jt, and that the vertex group at a vertex of H\Jt is
the stabiliser of a corresponding vertex of Jt. A similar result holds for edges.
This proves the theorem.//

The following corollaries are obvious.

Corollary 1 Let H be a subgroup of the fundamental group of a
graph of groups such that H meets every conjugate of a vertex group in the
trivial group. Then H is free.11
Corollary 2 Let @ be a graph of groups. Suppose that there is a
group H and a homomorphism <p:7i(@) -» H such that 9 is one-one on each vertex
group. Then ker<p is free.11

This theorem includes the Schreier Subgroup Theorem, taking & to

have its vertex and edge groups trivial, and it includes the Kurosh Subgroup
Theorem, talcing the edge groups of & to be trivial.

Exercise 3 Let// be a subgroup of finite index m in the free product

G]*.. .*Gn, and let m- be the number of (HyG-) double cosets for \±i±n. Show
that the free part of H in the expression of H as a free product according to the
Kurosh Subgroup Theorem has rank 1 + m(n-1)- Zm..

What groups can act on trees? Plainly, any group can act trivially
on any tree. More generally, the exercises show that for any group there are
many trees on which the group acts non-trivially but stabilising some vertex.
Thus the interesting question is what groups can act on trees with the action
stabilising no vertex.
Any HNN extension and any proper amalgamated free product can
be written as the fundamental group of a graph of groups with only one
edge-pair. The universal cover of this graph of groups is then a tree on which
the group acts without stabilising a vertex.
Let G be the union of a strictly increasing sequence of subgroups
G• . Let X be the graph with one edge-pair te ,^} for each n, and one vertex vn
m m
for each n, where eja
n vn and ejt n vMn+.,i . Then G is the fundamental group
« * * - of
the graph of groups on X with the groups at the vertex vn and at the edge en
both being G . Again, the universal cover of this graph of groups is a tree on
which G acts without stabilising a vertex.
We shall see in Theorem 29 below that these are the only possible
groups which can act on a tree without stabilising a vertex. However, the trees

involved may correspond to more complicated expressions of the group than the
ones above.
Let T be a (7-tree. We say that g translates the edge e if there is
an irreducible path e^9...fen with e^{eye~\ such that g{e*o)mej and ge^e~n.

Proposition 28 Let T be a G-tree, and let g be in G. Then the

following are equivalent. (/) g stabilises no vertex, (ii) g translates some edge,
m e ana
{Hi) there are edges e. {all /eZ) such that e^a e-x and i+\*&i * a> number
n such that ge • - e + for all i.

Proof If (iii) holds, the path e^,... >e shows that g translates e^.
Conversely, if g translates an edge, then from the irreducible path e , , . . . tem
i n
whose existence is assumed we get a seqeunce of edges satisfying (iii) by
defining e^^ to be g e. for all ktZ. and l
Suppose that g stabilises no vertex. C h o o s e the vertex v s o that the
irreducible path from v to gv is as short as possible. Let this path be e j , . . . ^ .
Then g(e*o)m gv ej, and (ii) holds provided that ge^^e^. Since G acts without
m
inversions, w e cannot have ge* - e~ if n \. If n>\ then ge* m e~ would tell us
e e startec at e x anc enc
that the path 2'"' /z-l * \ * *ed at g(e*T), contradicting the
choice of v.
Finally, suppose that (iii) holds. Take any vertex v. Take / s o that
the irreducible path from v to ep is as short as possible. Let this path be
jct,...,jc . Evidently, w e have e- *x *e~.. It follows that the path
l m i-\% m i
x x e e x an
\ >• • •» ~,> ;•> — -> ;±» i»& ~~,»• • •»£-^i is irreducible path from v to gv, and
so vtgv./l

Theorem 29 Let T be a G-tree. Then exactly one of the following

occurs: {i) G stabilises a vertex, {ii) there is an infinite sequence of edges
e« ,?2 t • • • w/rA en+i°m e J- and en+\ ^~n f°r a^ n and sucn tnat

stable o)cstab{e +*o) for all n {from which we see that stable o)m stab e ) and
G'Ustab{en) but G*stab{en), {Hi) some element of G translates some edge e.
In case {iii) G is either a proper amalgamated free product with GQ as
amalgamated subgroup or is an HNN extension with G as an associated
subgroup.

Proof In case (ii), every element of G stabilises some vertex. Hence,

by Proposition 28, if (iii) holds neither (i) nor (ii) can hold.
Suppose that (ii) holds, and let v be any vertex. Take m so that the

irreducible path from v to e^o is as short as possible. It is then easy to check:

that, for all n> m, the irreducible path from v to e a contains e . In particular,
n nt
any g which stabilises both v and ena will (since T is a tree) also stabilise e .
It is now easy to see that (i) cannot hold. Thus at most one of the three
possibilites can hold.
Suppose that (i) and (iii) do not hold. Take any vertex v. Since (i)
does not hold, (7/stabv. Take any g* stabv. Since (iii) does not hold, it follows
from Proposition 28 that g stabilises some vertex. Take w so that gestabw and
such that the irreducible path from v to w is as short as possible subject to this.
Let this path begin with the edge e and end with the edge x.
Because w is chosen so that the path is as short as possible, we cannot have
gx - x. It follows that we have an irreducible path e,..., x, gx~,..., ge~ whose
first edge is e and whose last edge is ge'.
We next show that the edge e does not depend on the choice of g
or w. Let us choose another gj i stabv (we permit gj - g to allow for a different
choice of w) and the corresponding edge e^ Then there is an irreducible path
starting with e^ and ending with 8 ^ . If eê then the path
g*e~, îî»• • • »î » e, — -9gF is irreduicble. Thus g" gj translates F, contradicting
our assumption that no element of G translates any edge.
Thus e depends only on v, and we have a C?-map <p:K-»K, where
vcp - ex. Since 9 is a Omap, stabv£Stab(vcp) for all v. By construction, for any
gi stabv, any w which is as close as possible to v subject to the condition gw-w
has vcp on the irreducible path from v to w, and so does not have v on the
irreducible path from vcp to w. Now take gtf stab(vcp), and we see from this that
(v<p)<p/v. It then follows that, letting en be the edge corresponding to v<pn~ , we
have e xme + ,o and e
n+i*^n> which is part of what we need.
It is also clear from the construction and the remarks of the
previous paragraph, that when w is chosen as close as possible to v subject to
being stabilised by g then w is also a vertex as close as possible to vcp subject
to being stabilised by g. It then follows, inductively, that w- v<pn for some n. In
particular, g, which was any element not in stabv, lies in statyvep72) for some n.
Thus Gm Ustab(v<p/2), as needed. Finally, G cannot stabilise any edge, since it
would then have to stabilise some vertex.
We still have to obtain the decomposition in case (iii). Now
T-(Geu Ge~) is a C7-forest. Contracting each of the trees in this forest, we
obtain, by Lemma 5.12, a C7-tree S with exactly one orbit of edge-pairs.- Since e
is translated in T by some element of G, e will also be translated in S by this
element. Thus, by Proposition 28, G stabilises no vertex of S. The Second
Structure Theorem now gives the result.//

Exercise 4 Let G be any group, and let T be any tree and v any
vertex of T. Take disjoint copies T of T for ail g*G. Take the union of these
with the vertices v for all g being identified to a single vertex. Show that this
is a Octree such that G stabilises a vertex, and such that no #*1 stabilises any
other vertex.
Exercise 5 Let G be any group, and let T be any tree and v any
vertex of T. Show that there is a graph of groups @ on 7* whose fundamental
group is G. What other graphs of groups can you find whose fundamental group
is Gl (There are some easy modifications of the graph of groups already found
which apply for all G. There are also other examples, depending on the subgroup
structure of G.)

We know that a finite subgroup of an amalgamated free product lies

in a conjugate of one of the factors, with a similar result for HNN extensions.
We now prove a stronger result.
Let T be a tree. A subset S of V is bounded if there is a number k
such that, for any two vertices v and w in S, the irreducible path from v to w
has length at most k; the minimum such k will be called the diameter of S. In
particular, if T is a (7-tree, it is easy to check that if one orbit of vertices is
bounded then every orbit of vertices is bounded. We then say that G has
bounded action on T. More generally, we can refer to a subgroup of G as having
bounded action on T. Plainly any finite subgroup has bounded action.

Proposition 30 Let G have bounded action on a tree T. Then G

stabilises some vertex of T.
Corollary Let H be a finite subgroup of a fundamental group of a
graph of groups. Then H is contained in a conjugate of some vertex group.

Proof The corollary is immediate from the proposition, using the

First Structure Theorem and the remarks already made.
Let G have bounded action on T, and let v be a vertex of T. Let X
be the subtree of T consisting of all vertices and edges lying on some
irreducible path joining two vertices of Gv. Plainly, A" is a (2-tree.
By hypothesis, there is some k such that the irreducible path joining
two vertices of Gv has length at most k. By definition of X, it follows that to
any vertex JC of X there is a vertex of Gv such that the irreducible path from x
to it has length at most k/2. We then see that any two vertices of X have an
irreducible path joining them of length at most k/2 + k + k/2; that is, X is
bounded.

We shall show that to any bounded (7-subtree of T with diameter

> 1 there is a (7-subtree of smaller diameter. Since an action of G on a tree of
diameter 0 or 1 plainly stabilises the vertices, the result follows by induction.
We showed in Lemma 5.10 that in any non-trivial finite tree there
is a vertex of degree 1 (that is, a vertex which is the endpoint of only one
edge-pair). It is easy to see that the proof of Lemma 5.10 also shows that the
same result holds for a bounded tree. In particular, let Y be the set of vertices of
X of degree 1. Then Y is a non-empty G-set. Suppose that X has diameter > 1,
so that X does not consist of only one vertex, nor of two vertices joined by an
edge. Let X^ be the graph obtained from X by deleting the vertices of Y and
deleing those edges that have a member of Y as an endpoint. Then X^*to.
Plainly X^ is a Osubset. As in Lemma 5.11, X^ is also a tree, and it will have
smaller diameter than X, because two vertices of X joined by an irreducible path
of length the diameter of X must lie in Y.ll

We now look at finitely generated groups acting on trees. We need

some general definitions about trees.
Let v be a vertex of a tree T. We say that the edge e points
towards v if the irreducible path from eo to v begins with e, and points away
from v otherwise. It is easy to check that the set of edges pointing towards v
forms an orientation of T, and that if we replace v by another vertex then we
change the orientation of only finitely many edges (namely, those on the
irreducible path between the two vertices).
Evidently there is no edge starting at v which points towards v, but
for every other vertex w there is exactly one edge starting at w and pointing
towards v; we denote this edge by ew. This orientation has also been considered
in section 5.3.
For any irreducible path e^,...,en in T there will be some r such
that e- points towards v for i±r and away from v for i>r (possibly rm0). For
otherwise there will be some k such that e, points away from v and e, +* points
towards v. Then both e". and e^+i point towards v and both start at e,i. As
already remarked, this would make ^jcmejc^.\y contradicting the irreducibility of
the path.
Let 5 be a subtree of T with v e S. Then for any vertex w of S we
have ^ w € ^» since the irreducible path in S from w to v must begin with e . It
follows that if we have an irreducible path in T all of whose edges point
towards v and the starting vertex of the path is in S then all vertices and edges
of the path are in S.

Now let A" be a G-tree, and let Y, T, and j have the same meaning
as in the previous section. Orient X towards some vertex a of Tj. We say that g
is negative for e if ge and e are differently oriented. We say that a (7-orbit of
edges is reversing if for some (and then for every) edge e in the orbit there is
some g such that g is negative for e.

Lemma 31 There are finitely many reversing orbits iff there are
finitely many edges of Y not in T and also there are only finitely many vertices
v of Tj such that stab(v)*stab(ev).

Proof In an orbit above an edge of Y not in T there will be an

edge e such that eotTj but eiiTj. There will be some g such that g(ei)*Tj\
and then g(eo)4Tj. As already remarked, this tells us that both e and ge" point
away from v, and so the orbit is reversing. In particular, if there are only finitely
many reversing orbits then there can be only finitely many edges of Y-T.
Let v be a vertex of Tj such that stabv*stabe y . Since ev is the
only edge starting at v and pointing towards a, it is immediate that any element
of stabv-stabe y is negative for e y . Hence ev is in a reversing orbit. Also, if w
is another vertex of Tj then ev and ew are in different orbits, since all vertices
of ev and e are in Tj and distinct vertices of Tj are in distinct orbits. Thus
there are only finitely many vertices of Tj such that stabv*stabe y if there are
only finitely many reversing orbits.
Suppose that there are only finitely many edges of Y- T and only
finitely many vertices v of Tj such that stabv/stabe y . Let Z be a finite subtree
of X containing a, all vertices v of Tj such that stabv^stabe y , and all vertices
yj with y a vertex of an edge of Y-T. We will prove the result by showing that
any reversing orbit either lies above an edge of Y-T ot else has an edge in Z.
Take a reversing orbit lying above an edge of T. Take an edge e of
Tj in this orbit, and let g be negative for e. Take an irreducible path from some
vertex of e to some vertex of ge. If this path starts at ea but does not begin
with e we may add the edge e~ at the beginning of the path. Similarly, if it
starts at ei but does not begin with e~ we may add the edge e at the beginning.
We may make a similar addition at the end. Finally, replacing e with e~ if
necessary, we may assume that we have an irreducible path e*,...,e such that
ej - e and en € igey ge~}. Write vQ - e^o7 and v. - e-x for 1 ^ / £ n.
As we have seen, if en points towards a then every edge e- points
towards a. Since e and ge have different orientations, we would then have
en - ge~, and so v - gv^; of course, vn * VQ , as n > 0 and X has no

irreducible loops. Since VQ e Tj, we know that every vertex of the path, and so
partic
vn in particular, lies in Tj. This is impossible, as Tj contains only one vertex in
each orbit.
Suppose that there is some r with 0<r<n such that e- points
towards a for i^r and away from a for i> r. In particular, e^ and e have
different orientations. As g is negative for e, we see that en* ge. If VQ*Z, we
know that e-tZ for all i<r, since e- points towards a; in particular, e e Z. If
VQ • Z, then any edge with vertex VQ lies above an edge in T. In particular, the
edge g~ en_i which ends at g~ (en<*)m VQ is hx for some h*G and JC e Tj. Since
xizTj and /Z(JCT)-VQ, and 77 contains only one element in each orbit, we must
a n d let e
have xx - VQ, and so hv^ - VQ . Let v_^ be hg vn_2> o ê ^ « - l *
Consider the path e^,e*,... ,e , .Since e points towards a and v^4 Z, we know
we
that stabvQ- stabe. Thus he^ - e, , and since ^ i ^ ^ . i * have e+2e~Q, so that
the path is irreducible and e~^ points away from a. By induction on n-r, we see
that there must be an irreducible path whose first r+1 edges point towards a,
with e among these r+l edges, and whose first vertex is in Z. As already
remarked, it then follows that each of these r + l edges, and so e in particular, is
in Z.
We are left with the case when e -t points away from a for all /. We
m v m v
must have e ge~, and so n S r\- Since v ^VQ and VQC77, we have v 4 Tj.
Take r so that v.e7y' for i^r but v^- 4 Tj; r>0, by our assumption on e. If e
lies above an edge not in T then v ^ Z , by definition of Z. We then find that
e € Z, by our remarks about the orientation. If ef lies above an edge in T, ^h
with he^Tj. Then hv^Tj, and so hvf~vr. Now take the path
e j , . . . , er, her+i,..., hen. If this path is irreducible, then, since e^ points away
from a, all edges point away from a, and e is in Z, by induction on n-r. If this
path is reducible we must have her+^m e~r, which points towards a, and so e y is
her^. As er+\*e~ry we
have êr+\êr+\- Thus, by definition, v^cZ, and, as
before, e e Z.I I

Lemma 32 Let G be finitely generated. Then G has finitely many

reversing orbits. If, in addition, each edge has finitely generated stabiliser then
each vertex has finitely generated stabiliser. Conversely, if G has finitely many
reversing orbits and each vertex has finitely generated stabiliser then G is
finitely generated.

Proof G has as homomorphic image the free group n^(Y), which has
one basis element y for each edge-pair iy,y~) not in T. Hence there are only

finitely many edges not in T if G is finitely generated.

The finitely many generators of G will involve the y and elements
from the stabilisers of finitely many vertices. Thus we may take a finite subtree
Z of Tj containing a, all these vertices, and all vertices of Tj above a vertex of
an edge not in T. It follows that G, which is the HNN extension with stable
letters {y ) of the tree product over Tj of the groups stabv, is equal to its
subgroup in which the base group is the tree product over Z. This requires, by
Proposition 1.34, that the tree products over Tj and over Z are the same.
Obviously, then, the tree products over Z and over any finite tree Zj with
ZcZ^cTj will be the same. By an easy induction over the length of the
irreducible path from v to a, we see that for v 4 Z we must have stabv - stabe y ,
since otherwise adding the vertex v and the edge-pair {ev,e~v) will enlarge the
group. The result now follows by Lemma 31.
Now suppose that every edge has finitely generated stabiliser. Any
vertex is in the orbit of some vertex v of 7y, and so has stabiliser conjugate to
stabv. If viZ then stabv- stabey is finitely generated. Now G is an HNN
extension with finitely many stable letters of the tree product of the groups stabv
for v€Z, and so, by Proposition 1.35 and Proposition 1.29 and induction, each
stabv for v e Z is finitely generated.
Conversely, suppose that G has finitely many reversing orbits. By
Lemma 31, the expression of G given by the Structure Theorems is an HNN
extension with finitely many stable letters of a tree product of vertex stabilisers,
and this tree product can be regarded as a product over a finite tree. The result
follows.//

A group is said to have the Howson property (or finitely generated

intersection property) if the intersection of any two finitely generated subgroups
is finitely generated. Plainly, any finite group has the Howson property, as does
the infinite cyclic group. Howson (1954) proved that free groups have the
Howson property. This is immediate from the stronger result below.

Theorem 33 Let ® be a graph of groups in which each edge group

is finite and each vertex group has the Howson property. Then 7i(@) has the
Howson property.

Proof Let H and K be finitely generated subgroups of TZ(&). Let X

be the universal cover of &. Then both H and K act on X.
By Lemma 32, /here are only finitely many reversing //-orbits, and

//nstabv is finitely generated for each vertex v, and the same for K. Since stabv
has the Howson property for all v, by hypothesis, we know that HnKn stabv is
finitely generated for alt v. Using Lemma 32 again, the theorem follows if we
can show that there are only finitely many reversing (Hn K)-orbits.
Now any reversing (//n/Q-orbit is plainly in the intersection of a
reversing //-orbit and a reversing AT-orbit. Since there are only finitely many of
each of these, it is enough to prove that the intersection of an //-orbit and a
AT-orbit contains only finitely many (//n K)-orbits.
Let e be an edge with stabiliser A. Then ge c He n Ke iff
g e HA n KA. Now HA n KA is the union of the sets Ha n Kb for a and b in A.
Each of these sets is either empty or a coset of HnK. Since A is finite, HA n KA
consists of the union of finitely many cosets of HnK. In particular, it is the
union of finitely many double cosets (Hn/Qg^A, and so ge is in the (//n/Q-orbit
of one of finitely many edges g^e.lI

A more complicated analysis on generally similar lines would

enable us to prove the following. Let H be a finitely generated subgroup of the
fundamental group G of a graph of groups with finite edge groups. If H
contains an infinite subnormal subgroup of G then H has finite index in G.

Howson's argument for free groups was very different from the
above proof, and gives a bound on the rank of HnK in terms of the ranks of
H and K. We give a graphical version of Howson's proof.
Let F be any connected graph, and let T be a maximal tree in F.
Let B be the set of edges not in T. Let F* be the subgraph of F consisting of B
together with all vertices and edges on the irreducible paths in T joining two
vertices of edges in B. We say that an irreducible loop e^9...9en is cyclically
irreducible if e+
i * e*
n .

Lemma 34 (/) Any edge of T* lies in a cyclically irreducible loop,

(ii) Any edge in a cyclically irreducible loop lies in IX (Hi) T*nT is connected
(and hence is a maximal tree of T*).

Proof Any edge of B lies in a circuit, and so lies in a cyclically

irreducible loop. If e is an edge of F* not in B then there is an irreducible path
e
\>- - - >en *n ^ c o n t a i n i n S e a n d beginning and ending at vertices of edges in B.
Thus there are edges e Q and en ^ of B with e^x - e^a and en +1o - e^. If
*«+l" e 0 then eQfe^9...fen is a cyclically irreducible loop. Otherwise there is

also a path JC« , . . . >xm in T from e + jt to e^o, and

A , e « , . . . , e , e M+
eu i n n J *,i JC,l , . . . , x^
/7z is a cyclically irreducible loop.
Conversely, let e be in a cyclically irreducible loop e^t...9e . We
may assume e 4 B. Thus we may assume, taking a cyclic permutation, that e^ « B,
and that e - e f for some r>\. Then we may find p<r<q^n + l so that e and e
are in B (where &n+i is £j) and e ^4 B for p<i<q. Then the irreducible path
e I , . . . ,e _ | shows that e e F*.
Now take any two vertices of F*. There is a path in F joining
them, and this path will, by definition of F*, consist of edges either in B or in
F* n T. Each of the edges in B can be replaced by an irreducible path in T with
the same endpoints, and this path lies in F*, by definition. Thus we obtain a
path in F* n T, as required.//

From (i) and (ii), F* can be described as the set of edges lying in
some cyclically irreducible loop, together with their vertices. In particular, F*
does not depend on the choice of T. We call F* the core of F.
The inclusion of F* in F induces an isomorphism of 7tj(F*) to 7ij(F),
since by (iii) and Corollary 1 to Theorem 5.17 both have the same basis.
If TT^F) is finitely generated then B is finite, and so F* is finite.
Let F be a finite connected graph with m vertices and n edges. Let
d. (or ^ ( F ) , if we need to consider several graphs) be the number of vertices
with / edges starting at the vertex. Then mmTd^ while 2nmI.id-9 since there are
2/z edges in all, and for each / there are d- vertices at which exactly / edges
start. By Corollary 1 to Theorem 5.17, rc^F) has rank n-m + \. Thus
2(rank7rj(F)-1)-Z(/-2)^.. In particular, if F has vertices of degrees 2, 3, and 4
only then 2(rank7T|(F) -1) - d~ + 2d^.

Proposition 35 Let H and K be finitely generated subgroups of a

free group. Then HnK is finitely generated and
rank(HnK) ~ 1 * 2(rankH~ \){rankK-1).

Proof Since H and K are finitely generated, there is a finitely

generated free group containing them both (for instance, the group on those basis
elements of the given free group which are involved in the generators of H and
K). Since any finitely generated free group is contained in a free group of rank
2, we may assume that H and K are subgroups of a group F with basis {JC,^}.
Then F acts freely on a tree T. We may label the edges of T with
the symbols x9 y, JC , and y~ , and starting at each vertex there is exactly one
edge with a given label.

We know that H\T and K\T have H and K respectively as their

fundamental groups. Let F and A be the cores of H\T and K\T, Since H and K
are finitely generated, both F and A are finite. By the definition of a core, no
vertex of either F or A can have degree 1, while by construction of T no vertex
can have degree greater than 4.
Let TxA be the graph such that K(Fx A)- K(F)x K(A), there being an
edge of FxA from (a,b) to (/>,<?) iff there is an edge in F from a to p and an
edge in A from b to q> with both edges having the same label. Plainly
c/4(Fx A)£rf4(D<tf4(A), while ^(Fx A) * ^3(F)^3(A) + ^3(F)^4(A) + ^4(F)^3(A). From this
it follows that ^3(Fx A) + 2<*4(Fx A) * C^3(F) • 2^4(F)]C^3(A) + 2</4(A)L From the
previous discussion, it follows that for any connected graph Z which has a
one-one map into FxA we have rankTTj(Z)-1 ^2(rank//- l)(rankAT- 1).
Thus it is enough to show that there is a one-one map from the
core of (HnK)\T into FxA. We plainly have a map from (HnK)\T into each of
H\T and K\T. Further, it is easy to see that a loop in any of these three graphs
is cyclically reduced iff the element of F corresponding to it (when we replace
each edge by its label) is cyclically reduced. Thus the core Z of (//n K)\T maps
into the cores F and A of H\T and K\T. This plainly gives a map of Z into Fx A.
Now the vertices of T can be regarded as the elements of F, and the vertices of
H\T will be the cosets of //, and similarly for K and Hr\K. Since the
intersection of a coset of H and a coset of K is either empty or consists of
exactly one coset of HnK9 the map of Z into FxA will be one-one on vertices.
Because the edges are labelled, it will then also be one-one on edges.//

Exercise 6 Let H and K be finitely generated subgroups of a free

group F, and suppose that H has finite index in F. Show that
rank(//n/Q- U(rank//- l)(rank/:-1).
Exercise 7 Find an example of a map from a graph Fto a graph A
which does not map F* into A*.

We now prove an important result about free products. We will

need to begin with some results about graphs.
Let AT be a Cr-graph (since G can be trivial, the results below hold
for arbitrary graphs). Let = be an equivalence relation on E such that ge^ = ge
and e j s e " if e^ = e and such that we never have e~ s ge. We may define an
equivalence relation on K, which we still denote by =, to be the smallest
equivalence relation such that e^z s ex whenever e^ = e (we will also have
e+o e ea). Thus gv^ = gv if Vj = v. It is clear that we have a <7-graph, which we

denote by X/=t whose vertex set is the set of equivalence classes of V and
whose edge set is the set of equivalence classes of E, and that we have a
C?-map p from X to X/=. We need e~ not to be equivalent to ge in order to
ensure that G acts without inversions on X/= (and to ensure that "~ does not fix
any edge of Xl=).
Let J be a G-graph. A fold consists of a pair of edges e^ and e^
anc
such that e j G - ^ o * e2 * ^\ u (
^ l * When f.X + Z is a map of (7-graphs and
m
we also have e*f e~f, we refer to a fold along f
Let X have a fold. Let = be the smallest equivalence relation on E
such that ^2 s e l anc
* suc
^ ^at ^ JfS
J ; t ^ ien x~=if anc
* S-x=©>. We construct the
Cr-graph XI s as above, and we denote A7= by K The condition that e^ 4 Ge~^ is
needed to ensure that G acts without inversions on Y (and to ensure that ~~ fixes
i s use< t o ensu
no edge of Y) , while the condition e^ * ^ i * r e that if no edge of
X is fixed by a non-trivial element of G then the same holds for Y. Plainly, if
we have a fold along the map / then there is a map f:Y-> Z such that fmpf'.
We will call X edge-free if every edge has trivial stabiliser.

Lemma 36 The graph obtained by folding a G-tree is a G-tree.

Proof We use Lemma 5.13. Thus we define CQ(X) to be the free

abeLian group on E(X)9 CAX) to be the quotient of the free abelian group on
E(X) by the subgroup generated by all e + e~ , £^:CQ{X)^Z to be given by
v e ^ - 1 for all vertexes v, and dj,:CJX)-> C^(X) to be given by ed^m ei- eo for
all edges e of X, and we make similar definitions for Y.
Since * is a C7-graph, both C^(X) and C^(X) are {/-modules, and
6j£ and Eg are homomorphisms of <7-modules (where Z has the trivial module
structure), and similarly for Y. It is easy to check that C^(Y) is the quotient of
C^(X) by the submodule generated by e^ - e^, and that CQ(Y) is the quotient of
CQ(X) by the submodule generated by e^ 1 " e\x- Since e^ 0 " e\a* we nave

X
~ e\x' ^ e s e e ^ a t ^Y a n d £Y a r e * n d u c e c l ^ r o m ^x anc^ £X'
By Lemma 5.13, since X is a tree, we have k e r c ^ - 0 and
.. It follows easily that kerdy-0 and kerEy- imdy, which, using
Lemma 5.13 again, shows that Y is a tree.//

Lemma 37 Ler f.X + Z be a map of G-trees, and let X be

edge-free. Then there is an edge-free G-tree X\ such that X' is XI ^ for some
equivalence relation for which xfmyf whenever x=y, and such that X' has no
fold along f\ where f':X' -» Z is the map such that / - pf\

Proof We will need to look at various equivalences on E. We shall

therefore regard an equivalence as a subset R of E* E, and we denote the
quotient graph by XIR. All our equivalences will be such that if x=y then
xfmyA gx = gy, and JT^JT, and such that JT is never equivalent
to gx. We further assume that X/R is an edge-free tree.
Suppose that there is an equivalence R which is maximal with
respect to having these properties, and let X* be X/R. Then X* cannot have a
fold along f\ For if it did we could lift the equivalence on E\ which exists
because there is a fold, to an equivalence S on E9 which would strictly contain
R, and which would satisfy all the relevant properties (X/S is the same as the
result of folding X\ and so it is a tree, by Lemma 36.).
Hence it is enough to show that there is such a maximal R, and we
use Zorn's Lemma to prove this. So we take a collection of equivalences / ? a , all
satisfying the relevant conditions, such that for all a and p either Ra£Ra or
R~cR . Then U^ is an equivalence, and all the relevant conditions obviously
hold for U/? a , except for the requirement that X/UR^ is a tree. We show that
this condition also holds.
Since X is connected and maps onto X/UR , it is immediate that
X/UR^ is also connected. Suppose that we have a loop in X/UR^. Then we have
edges # j , . . . 9en of X such that, for 1 £ / £ n, e /c is equivalent to e / + j0 (where
e n^ means e^) in the equivalence relation on V induced by U/? a . This means
e
that we have edges e-- (l^j^m^) such that e^T-eft, imx' e
i+\a' and e
if
equivalent to e- +jT under UK . Then there will be o^. • such that e. -x is

equivalent to e- +*t under ^ a . By the assumption on the family / ? a , there will

be some p such that Ra R ffor all / and j . It then follows that
cR~ e^,...9en
gives a loop in X/R&.. Since X/R* is a tree, this loop is reducible in X/RQ,
X/RQ, and
so our original loop, being
bei the image of this in the map from X/R& to
will also be reducible.//

Lemma 38 Let f:X+ Y be a map of G-trees. Let X be edge-free

and let there be no fold along f Then either the induced map from G\X to
G\Y is injective or else there is a vertex v of X, an edge e of Y, and some
ge G such that eom vf and g fixes v and e but does not fix all of Y.

Proof Suppose that the conclusions do not hold. We will obtain a

fold along /.
Suppose first that there are vertices Vj and v^ °f X such that
V
2 * ^ 1 k ut V
2^"^ v i-^ f° r some
8- Replacing v. by gv., we may assume that

v 2 / - Vj/. Choose Vj and v 2 with Vj/- v 2 / such that the irreducible path p (which
must have positive length) from Vj to v 2 has minimal length. Then pf is a
closed path in Y. Since Y is a tree, it follows that we must be able to write p
w
as p, e* e2P2 ^ e r e e i anc* ^2 a r e 2 1
If we had e2 € Ge^, we would get e^/c 6(£j/). This cannot happen
as G acts without inversions on Y. Since e 2 a - eô we have a fold along / unless
e2tGev
So we may assume that there is some g such that e^mge2. Thus
g(e2f)" ^ 2 / and g{e2a) - e 2 a. Since the second of the two conclusions was
supposed false, it follows that g fixes all of Y; in particular, it fixes v 2 /. Since
e"j - ge2, the path gp2 starts where Pj finishes, and so we have a path <7 which
consists of p« followed by gp2. This path starts at Vj and finishes at gv2, which
are not in the same <7-orbit. Also we have v ^ - v 2 / - g ( v 2 / ) - ( g v 2 ) / . Since q is
shorter than p, this contradicts our original choice of Vj and v 2 .
Now suppose that there are edges e^ and e2 such that e2 4 Ge* but
e2f* CJê^f). As in the first part, we may assume that e^ and e2 are chosen so
that e2f~ e^f and so that the irreducible path from eô to e2o has minimal
length. We get the same contradiction as before if this path has positive length.
If this path has length 0, so that eâ - e 2 o, we have a fold along / , since
e^f- e2ft e2 4 Ge^, and e2 4 Ge^, for the same reason as before.//

We define a homomorphism from the graph of groups {®,X) to the

graph of groups {$fr,Y) to consist of a map of graphs <p:X-*Y, and
homomorphisms from Gx to H for each vertex or edge JC of X such that, for
every edge e of X, the homomorphisms GQ -> Gex -» He and GQ -» H -» // e<pT
are the same.

Lemma 39 Let <p be a homomorphism from the graph of groups

(®,X) to the graph of groups ($,Y). Suppose that <p maps a maximal tree T of
X into a maximal tree S of Y. Then <p induces a homomorphism from TI((8,X,T)
to n{$),Y,S) whose restriction to Gv is the given homomorphism from Gv to H
for every vertex v of X. Further, let B be an orientation of Y, and let A be
B<p~ (so that A is an orientation of X). Then cp induces a -K(G,X,T)-map from
the universal cover Jt of (@,X) to the universal cover f of Y.

Proof That <p induces a suitable homomorphism, which we still

denote by <p, from n((8,X,T) to n{&Y,S) is immediate from Lemma 20.

Let ]t and f be the universal covers of (®,X) and ($,Y)

corresponding to the maximal tree T and orientation A of X, and the maximal
tree S and orientation B of K, respectively. The homomorphism from n{&,X,T)
to n($,Y,S) makes ? into a n(®,X,T)-tree.
It is easy to check, first, that there is a well-defined function from
/ to f sending Cg,x] to [£<p,x<p] for all g* n(@,X,T) and every vertex or edge x
of X, and, second, that this function is a map of graphs which is a
?r(@,;if,7>map.//

Theorem 40 Let G^ and H^ be groups, and let tp be a

homomorphism from G^ onto H^ (for a in some index set). Let $ be the
induced homomorphism from $6^ to *// a . Let A be a subgroup of *C?a such that
A$ - * / / a . Then there are subgroups A^ of A such that A - tA^, and A^ - H
for all a.
Warning In general, A^An H^.

Proof Let Y be the graph with one vertex v a for each a, one other
vertex v, and one edge-pair {e^ 7e~Q(] joining v to each v a . Let @ and # be the
graphs of groups on Y with G (or H^) being the vertex group at v , the vertex
group at v and all the edge groups being trivial in both cases. Let X be the
universal cover of (®,Y), and let f be the universal cover of ($,Y). Applying
Lemma 39, and restrcting to the subgroup A, we have an ^4-map from X to f.
Because A$m%H a , we find that AX?mY. Since the action of %Ha on f is
edge-free, an element of A which stabilises an edge of f must fix all of f.
By Lemma 37, there is an edge-free ^4-tree Z and an ^4-map f.Z+Y
such that / has no folds. Then Lemma 38, together with the above remarks about
the action of A on V, tell us that the induced map from A\Z to Y must be an
injection. This map must also be a surjection, since / comes from a surjective
map from X to f by factoring out an equivalence relation. Thus we may as
well identify A\Z with Y.
Let j:Y-*¥ be the function considered in the definition of the
universal cover. Suppose that we can find a function kiY+Z such that kf~j.
Then the Second Structure Theorem tells us that Am%A^y where A^- stab(va&).
Also A $£Stab(v kf)m H . Since $ maps A onto %H , it must map A onto

Now there will be a vertex u of Z such that ufm Cl,v], since there is
a vertex of X which maps to Cl,vl There will be an edge z^ of Z starting at u
whose image in Y is e^. Then z^f will be ^-h,e^\ for some httH^. We have

Cl,v] - uf~ Zqff' Zqfo - CA,ea]o - CA,v3. Since the vertex group of ^ at v is trivial,
this gives hm\. We may now define k by vkm u, e km z , and, necessarily,
and v
*"<**" *a <x*'V7/

Corollary 1 (Orushko's Theorem) Let 9 be a homomorphism from a

free group F onto a free product * / / a . Then F is the free product of subgroups
F^ such that F^ - H^ for all a.

Proof Let G - #a let cpa be the restriction of cp to C?a, and let
<& be the induced map from %G^ to *// a . Because <p is onto, we have G^y - H^,
and so $ is onto. Let fr*^ "* F be the homomorphism which is inclusion on
each C7a, so that $ - $ 9 . Since each G^ contains kercp, we see that im^2ker<p.
From this it follows easily that # maps onto F. Since F is free, there is a
homomorphism ty:F->*G such that ^ is the identity homomorphism on F. Then
4> embeds F in *G , and the result follows by applying the theorem to Fty./I

Corollary 2 Let A and B be finitely generated groups, whose

minimum number of generators are m and n respectively. Then the minimum
number of generators of A*B is m + n.

Proof Plainly A*B can be generated by /« + « elements.

Let r be the minimum number of generators of A*B. Then there is a
homomorphism from a free group F of rank r onto A*B. By Corollary 1, we can
write F as ^1*^0 » w^GTe ^1 anc* ^9 m a p o n t o ^ a n d ^ respectively. Thus
rankFj^/w and r a n k / ^ f l . Since rank/ 7 - rankF^ •»• rankF^» we see that r*m + n, as
required.//

We can generalise Corollary 1.

Corollary 3 Let 9 be a homomorphism from a free product
Am F**A-> (for X in some index set A) onto Hm %H such that for each X there
A -1
are h\*H <*nd an index ou such that Ayy^h-.H A> . Then A has subgroups
<7a such that A~%G and G 9 - H .

Proof Let G^ be (A^//a h-^~ )<p . Let 0 be a symbol not in A, and

let (3aQ be # a 9 • Let G1 be the free product of the groups G^ for all a and all
li in Au(O}. As in corollary 1, we have homomorphisms &:G-> A and $:G-> %H
such that on each factor $ is inclusion and <f> is 9. As in Corollary 1, it is
enough to define a homomorphism ty.A -» G such that i|>$ is the identity on A. On

the factor A^ we need only require ty to be the inclusion of A-^ in G^ ^. Since

F is free, we can define a suitable i|> on F if # is onto. As before, this is
immediate once we have seen that $ is onto. But $ certainly maps the groups
G Q onto the corresponding / / a (this is why these extra groups were introduced),
so $ is onto.//

There are extended versions of these results, of which an example

follows.
Proposition 41 Suppose that we have homomorphisms from G^ to H^
and from G^ to H^ such that the induced homomorphism <p from G^G^ to
H
l*H0H2is onto. Let A be c2. subgroup of G^G^ such that A<p - //j
• •''0
Then
, where either / U<P£. //j and A, H^ or else A^ <pn//0#{]1} and A. 9 /$
contained in one of H^ and 1V

Proof Let Y be t;he graph with two vertices v 1 and v^ and with one
edge-pair (e,el joining v l and V2. Let @ be the graph of groups on Y with (
being the vertex group at v. for / - 1,2, with the edge group being trivial. Let £
be the graph of groups on Y with H- being the vertex group at v. and the edge
group being HQ. Let X be the universal cover of (<&,Y), and let Y be the
universal cover of ($&,Y). As in Theorem 40, there is an A-map from X onto f9
which, factoring by an equivalence relation, gives rise to an edge-free -4-tree Z
and an .4-map f.Z-*Y.
As before, A\f - Y, and / induces a map from A\Z onto Y. If this
map is injective then A\Z is isomorphic to Y. The Second Structure Theorem
then shows, as before, that the first case in the proposition holds. If the map
from A\Z is not injective then Lemma 38 shows that the second case of the
proposition holds.//

The following corollary can be deduced from the proposition in a

manner similar to the proof of Corollary 1 of Theorem 40 from the theorem. The
details, and the formulation and proof of similar results for HNN extensions, will
be left to the reader.

Corollary Let <p be a homomorphism from a free group F onto

H]*jf # 2 • Then F is the free product of subgroups F^ and Z7^ such that either
<p maps F j and F^ into ^ <**d # 2 respectively or F^<pn HQ^U) and F^<p is
contained in one of H^ and H^ -II

Marshall Hall (1949b) proved that any finitely generated subgroup

of a free group F is contained in a subgroup of F whose index is finite. His
proof used Schreier transversals. Although he did not explicitly say so, his method
proved a stronger result, which we specify below.
Definition A group G is called an M.Hall group if, given any
finitely generated subgroup H of G and any finite number of elements a*,... ,a
of G not in //, there is a subgroup K of G whose index is finite and such that H
is a free factor of K and a+,... ,a 4 K.
i n
Then the stronger form of Hall's theorem is that any free group is
an M.Hall group. This follows at once from the next theorem, which was first
proved by Tretkoff (1975).
Plainly any finite group is an M. Hall group (we may take K to be
H). Also the infinite cyclic group is an M. Hall group, since if H is non-trivial
we may take K to be //, while if H is trivial we need only take K to be
generated by a sufficiently large power of the generator of G.
We need to observe that if H is a finitely generated subgroup of an
M. Hall group G and b^,... 9b are elements of G no two of which lie in the
same coset of H then there is a subgroup K of finite index containing H as a
free factor and such that no two of b^,...,bn lie in the same coset of K. We
need only apply the definition to H and the elements b.b" for /#/.
Theorem 42 The free product of M. Hall groups is an M. Hall
group.

Proof Suppose first that G is the free product of an infinite number

of M. Hall groups. Let H be a finitely generated subgroup of G, and a.,... ,an
elements of G - H. Then we may write G - P*Q, where P is the free product of
finitely many M. Hall groups, and P contains the subgroup H and the elements
av...,an.
Suppose that there is a subgroup K of P which has finite index, has
H as a free factor, and contains none of a^,... >an. Let cp be the homomorphism
from G to P which is the identity on P and is trivial on Q. Then K<p has
finite index in G. Since K<p nPmK, we see that K<p contains none of
a* ,...,a . Also the Kurosh Subgroup Theorem tells us that AT is a free factor of
-1 -1
K<p , so that H is a free factor of K<p .
It follows that it is enough to show that the free product of finitely
many M. Hall groups is an M. Hall group. Then, inductively, it is enough to
prove the result when G - GQ*G^ , where GQ and G^ are M. Hall groups.

In this case there is an edge-free (7-tree X (namely, the universal

cover of the graph of groups which has one edge-pair with trivial edge group,
and two distinct vertices whose vertex groups are GQ and G^) with the following
properties. X has an edge e and vertices VQ and Vj such that VQ and Vj are in
different orbits, and have stabilisers GQ and G^ respectively, and such that
eo'v^y ex-Vj, and every edge is in Ge u Ge~ (and every vertex is in GVQUGVJ).
Let Y be H\X, and let T be a maximal tree of Y. Let j\Y*X be
the usual function. When convenient, we denote a vertex of Y by Hgv^ or Hgv.
and an edge by Hge or Hge". Give Y the orientation A consisting of all edges
Hge.
Then / / - n($,Y,T), where the edge groups of £> are trivial and the
vertex group at w is //n stabwy. Since H is finitely generated, and is the free
product of 7t|(]0 and the vertex groups Hw, it follows that Hwm{\) for all but
finitely many w and that there are only finitely many edges of Y not in T.
Let S be the finite set consisting of the vertices Hav*, the vertices
of the edges of Y- T, and all vertices w such that Hw*il}. Let ZQ be the finite
subgraph of Y consisting of the edges of Y- T and all those vertices and edges
and their inverses which lie on the irreducible paths from HVQ to the vertices of
S. Notice that TKZQ is connected, and so it is a maximal tree of Z Q .
Now take any vertex w of Y, and assume that w is of form Hgv^.
Consider {yeA;yomw}, and let y^ be an element of this set. Then yjomwjmy^fa. It
follows that there is a bijection from the set of cosets of Hw in stabwy onto this
set of edges, sending the coset H^j? onto that y such that yj'e Hpiy^j). Similarly,
if w is of form Hgv* then there is a bijection from the set of cosets of H in
stabwy onto {yzA;yxm w}, where the coset H^p maps to that y such that
Y (yJ)^-Hp^ CV(y), where y is the element of//-7i($) corresponding to y
Take a vertex w of ZQ of form Hgv^. It follows that any edge
starting at w is in A. Since ZQ is connected, there is at least one edge starting
at w. Choosing one such as the reference edge, there are finitely many elements
c, - 1,... ,c of stabwy in different cosets of H such that, in the bijection above,
the cosets H^c^ correspond to the edges of ZQ starting at w.
Now stabwy is an M. Hall group, being isomorphic to GU, and H
is finitely generated, being a free factor of H. Thus there is a subgroup Kw of
stabwy, containing Hw as a free factor, of finite index in stabwy, and with
C j , . . . tcr in different cosets of K^.
Let cr+i>"' *cn &e s u c n t n a t c\»• • • >cn form a transversal of Kw in
stabwy. Add to ZQ those edges of Y starting at w which correspond to the cosets
HyjC^ for r<i^n, together with their inverses and their other end-points. Do this

for all such vertices, and perform the similar construction for the vertices of form

The result is a finite graph Z such that Z^cZcY. Let SL be the

graph of groups on Z whose edge groups are trivial, whose group at a vertex w
of ZQ is Kw, and whose group at a new vertex w is stabwy. Then TnZ is a
maximal tree in Z, and / / - 7T($IZQ , ZQ,TO ZQ) is a free factor of 7t(&, Z,TnZ),
since / / w is a free factor of Kw for all vertices w of ZQ and each edge of Z - T
is in Z Q .
By Lemma 20, there is a homomorphism <p:?r(&, Z,7*n Z)-» C/ which
is the inclusion on each Kw and which sends each z to the element y of (the
subgroup H of) G The restriction of <p to / / is just the inclusion of H in (7.
Let 2f be the universal cover of the graph of groups 51. Because 9
maps Kw into stabwy, there is a well-defined function / from V(£) to V{X)
sending Uc,wl to (k<p){wj). We may also define f.E{Z)+ E(X) by tk,zVm (k<p)(zj)
and \.k,z~}fm(k<p){z~j) for all ?£i4nZ. Then / is easily seen to be a map of trees.
We next show that the theorem holds provided that / is a bijection.
For then, we see first that k acts trivially on % if k<p - 1, and so 9 is one-one.
It then follows, letting K be the image of 9, that H is a free factor of K.
Second, letting w^ be the vertex //VQ of Y and w- the vertex Ha^v^, we see that
w. and WQ are different vertices of Z. Hence the vertices C1,WQ] and Cl,wJ of ^
are in different orbits, and so w^j and vty are in different AT-orbits. Since w^/- VQ
and viy'e Ha-v^, we see that a^i K. Finally, there are only finitely many orbits
of edges in 2,', since Z is finite, and so there will be only finitely many
AT-orbits of edges of X. It follows that K has finite index in G.
To show that / is a bijection, it is enough, by Lemma 5, to show
that it is locally bijective. It is plainly enough to show that it is locally
bijective at all vertices of the form Cl,wl First let w be a vertex of ZQ of form
#£VQ . We have a transversal ( C j , . . . ,c^} of Kw in stabwy, and the edges of Z
starting at w are j ' j , . . . , ^ where we may assume that j y • c-(y^j) (multiplying
each c. by an element of Hw, if necessary). Then the edges of % starting at
Cl,w] are all lk9y-l with k*Kw (these being all distinct). Now
lkty-lfmk{y^)m(kc.)(y^j). Since the elements kc^ are exactly the distinct elements
of stabwy, we see that / provides a bijection from the edges of £ starting at
Cl,w] to the edges of X starting at wj\ as required. When w is a vertex of ZQ of
form //gvj a similar argument applies (but we need to take account of elements
y j . When w is a new vertex similar but simpler arguments apply, since then
Kwm stabwy and there is only one edge of Z starting at w.//

Theorem 43 Let G be a free product of non-trivial groups, let H

be a finitely generated subgroup of G whose index is infinite, and let
a* , ...,a be elements of G not in H. Then there is a subgroup K of G not
1 n
containing any of a+,... ,an such that H is a proper free factor of K.

Proof The proof is almost identical to that of the previous theorem.

We may write G as a free product G^*G^. We construct ZQ as
before. We obtain Z by adding all edges (and their inverses and their endpoints)
of Y which have one vertex in ZQ . We let Kw be Hw for any vertex w of ZQ ,
while Kw is to be stabvvy for any new vertex w (stabwy is, of course, a conjugate
of 6U or &). The previous argument shows that we have a subgroup AT of G of
which H is a free factor and not containing any of a^9...9a . Since H has
infinite index, Y will be infinite. As ZQ is finite, there will be at least one
vertex of Z not in ZQ. This shows that K is strictly larger than H.I I

Proposition 44 Let H be a finitely generated subgroup of a free

product G of non-trivial groups. If H contains a non-trivial subnormal subgroup
of G then H has finite index in G.

Proof Suppose that H has infinite index. By Theorem 43, there is a

subgroup K of G such that K' H*L where L*{\).
Let S be a subnormal subgroup of G contained in H. Then 5 is a
subnormal subgroup of K. Hence there is a sequence £ j - Kt..., S - S of
subgroups of K such that S-^ <S- for i<n. There will be some r such that <S'f+j
is contained in H but S is not contained in H. Take kiS - H and siS ,+ -{\).
-1 -1
Then k- sktS +•, since S ^< S , but k ski //, since K has H as a proper free
factor. This contradiction proves the result.//

We have already remarked that a different method of proof enables

us to prove the following generalisation of Proposition 44. / / G is the
fundamental group of a graph of groups with finite edge groups then any finitely
generated subgroup containing an infinite subnormal subgroup will have finite
index.
This would follow, as in the proof of the proposition, if we proved
the following generalisation of Theorem 43. Let G be the fundamental group of a
graph of groups whose edge groups are finite. Let H be a finitely generated
subgroup of O, and let a^9...fan be elements not in H. Then there is a subgroup
of G not containing any of a^,...,an such that K is the fundamental group

of a graph of groups, one vertex group being H, and all edge groups being
finite.
The proof of this is a routine, but messy, generalisation of the
proof of Theorem 43. It would be possible to generalise the definition of M.Hall
groups so as to obtain a similar generalisation of Theorem 42.

8.6 CONSTRUCTION OF TREES

In order to apply the structure theorems, we need to be able to
construct trees. Sometimes this can be done geometrically. For instance, Serre
shows by such methods that certain linear groups are amalgamated free products.
It will obviously be of interest to know if a given set can be
regarded as the set of vertices of a tree (or as the set of edges of a tree). It is
easy to see that to any set A there is another set B such that A is the set of
vertices of a tree and B is the set of edges, and also a set C which is the set of
vertices of a tree for which A is the set of edges of an orientation.
The situation is very different when we are given a C7-set A and
wish to find a C7-tree. For instance, it is easy to see that if \A\ > 2 then A cannot
be the set of vertices of a C7-tree if G is doubly transitive, and (using Lemma
5.10) if A is finite then it cannot be the set of vertices of a C/-tree if G is
transitive. We give next some useful conditions which will ensure that A is the
set of vertices (or edges) of a Cr-tree.
The next proposition is obvious.

Proposition 45 Let T be a tree, and let VQIV(T). For each

vzV(T), let vX be the length of the irreducible path from VQ to v. Then
X:K(71-^IN has the following properties: (i) vX - 0 iff v- VQ , (//) // v and w are
adjacent (that is, there is an edge joining v and w) then vX - wX - ±1, (///) //
VS*VQ then there is exactly one w with w\ - vX - 1 . / /

The converse of this is not difficult to prove.

Theorem 46 Let V be a set with a symmetric relation called
adjacency. Let X:K-»!N be a function satisfying (/), (//), and (Hi) of the previous
proposition. Then there is a tree T such that V(T)m V, with an edge joining v
and w iff they are adjacent, and such that the function of the previous
proposition is the given function X.

Remark Let X be a function satisfying (i) and (iii). If we are not

given an adjacency relation, we could define v and w to be adjacent iff

wX - vX - ±1, and (ii) would automatically hold. But it is more useful to permit
arbitrary adjacency relations; for instance, if V is a C7-set then T will be a
(7-tree iff the adjacency relation is preserved under the action of G.

Proof There is a graph T whose vertex set is V and such that there
is one edge-pair joining v and w for any two adjacent vertices v and w, and no
edge-pair joining non-adjacent vertices.
By induction on vX, there is for each vertex v a path joining v~ to

Let e j , . . . , en be a loop based at VQ, with n>0. Then « * 1 , by (ii).

Let v^-e/c, and choose r so that v^X - maxvX Then, by (ii) and the choice of r,
v iX - v X -1 - v -X. So, by (iii), v , - v +«. Since there is at most one
edge-pair joining any two vertices, this shows that er^me~r. That is, all loops
based at VQ are reducible, and so T is a tree.
Suppose that vX > 1, and let u be the vertex adjacent to v such that
uK - vX - 1 . Inductively, we may assume that the irreducible path from VQ to u has
length u\, and that, when " ^ V Q , that the vertex w adjacent to u on this path
has wX - u\ -1. Then there is a path from VQ to v of length vX, on which the
vertex adjacent to v is u, and this path must be irreducible since v*w.//

It is slightly inconvenient to have one vertex playing a special role,

and the next results avoid this.
Proposition 47 Let T be a tree. For u and v in V, let (utv)d be
the length of the irreducible path from u to v. Then d is an integer-valued
metric on V (which is G-invariant if T is a G-tree). For any integer r<{u,v)d
there is exactly one vertex p such that both (u,p)dmr and (u,v)d- (u,p)d + (p,v)d.
If (u,v)dm\ then, for any vertex w, (u,w)d^(vfw)d. If u,v, and w are vertices
with (u,v)d<(u,w)d + (w,v)d then there is a vertex q such that (u,q)d<{u,w)d and
{q,v)d<(w,v)d.

Proof Suppose that there is an irreducible path x* , . . . , x of length

n from u to w, and an irreducible path J^ >•••».% °f length m from w to v. Then
*l »• • • >xn>y\»• • • >ym iS a P at h fr°m " t o v- If this P at h ls reducible, the
irreducible path from u to v is obtained by reducing this path. Hence
(ufv)d z(u,w)d + (w,v)d, which shows that d is a metric.
S u p p o s e that (u,v)d < (u,w)d + (w,v)d. T h e n t h e p a t h x\* — ->ym
m
be reducible. This can only happen if y^ JC^, and the required vertex q will then
be yxx.

Now suppose that (u,v)dml and that (u,w)d*(v,w)d + 1. By the

previous paragraph there will be a vertex q such that (q9w)d < (v,w)d and
(u.q)d<{u,v)d. Since {u,v)dml, this requires that qmu, and so (u9w)d < (v,w)d, as
needed.
For any r<(u,v)d there is, plainly, a unique vertex p on the
irreducible path from u to v such that (u9p)dmr and (u,v)dm (u,p)d + (p,v)d. Also
the arguments of the first paragraph show that if w is a vertex such that
(u,v)d- {u,w)d + (w,v)d then w lies on the irreducible path from u to vJI

Theorem 48 Let d be an integer-valued metric on a set V.

Suppose that, for all u*v there is a unique p such that (u,p)dml and
(p,v)d- (u,v)d - 1 . Suppose also either that if(u,v)dml then, for all w,
(u,w)d*(v,w)d or that for all utvf and w such that (u,v)d < (u,w)d + (w,v)d there is
some q such that {u,q)d<(u,w)d and (q,v)d<{w,v)d. Then V is the set of
vertices of a tree in which there is an edge-pair joining u and v iff (u,v)dm 1,
and, for all u and v, (u,v)d is the length of the irreducible path from u to v.
Also, if V is a G-set and (gu,gv) - (u,v)d for all gtG and all u,viV then this
tree is a G-tree.
Proof As in the proof of the previous proposition, the last condition
on d implies the next-to-last condtion. Since d is an integer-valued metric, this
in turn implies that if (u,v)d-\ then (u,w)d- (v,w)d- ±1.
Take any vertex VQ, and define vX to be {v^yv)d. Also define u and
v to be adjacent iff (u,v)d~\. Then the conditions of Theorem 46 are satisfied.
Thus the result follows, except for identifying d with the length of the relevant
irreducible path. However, Theorem 46 tells us that the length of the irreducible
path from v^ to v is vX, which equals {v^yv)d. Since VQ is an arbitrary vertex,
the proof is complete.//

We now look at edge-sets of trees.

Proposition 49 Let T be a tree. For x and y in E write x zy iff
the irreducible path from xa to yi begins with x and ends with y. Then £ is a
partial order {which is preserved by G when T is a G-tree) satisfying the
following conditions: (/) if x*y then there are only finitely many edges z such
that x^z^y, (H) for any x and y, at least one of x<y, x^y", x~zy, x ±y~
holds, (Hi) if x^y then y~ £ x, (/v) we never have x^y and x<y~, (v) we never
have x^y and x~£y. Further, for y&x&jT , xxmyr iff JC^JT and there is no
edge z such that x^z^-F and

Proof Suppose that x^y and y^z. Then there is an irreducible path
beginning with x and ending with y, and an irreducible path beginning with y
and ending with z. These paths fit together to give an irreducible path beginning
with JC and ending with z, so that x * z> Further, if x<y and y^x we must have
x - y, since otherwise the paths fit together to give an irreducible path of length
> 1 beginning and ending with x, which is impossible as T is a tree. Thus * is
a partial order.
Suppose that x^y. Then the inverse of the irreducible path
beginning with JC and ending with y is an irreducible path beginning with y~ and
ending with JC\ so that >T <: x~. Also, the unique irreducible path from xa to
yTx-yo is obtained from the original path by deleting the edge y. In particular,
its last edge cannot be y~, so that we do not have JC* jr. Similarly we cannot
have x±y. (Alternatively, (v) is an immediate consequence of (iv) and (iii).)
If the irreducible path from xa to y% does not end in y we ccan add
the edge y~ at the end and still get an irreducible path. Similarly, if the path
does not begin with x we can add the edge x at the beginning and still get an
irreducible path. This proves (ii).
Suppose that x^z^y- Then, as already seen, there is an irreducible
path beginning with JC, ending with y9 and containing z. Since there is only one
irreducible path starting with JC and ending with y, z must be one of the edges
on this path, proving (i).
If jcx-yi and x*y then JC, jT is an irreducible path, and so JC^JT.
Further, by the previous paragraph, there can be no edge z such that x^z^y'
and jc*z*jT.Conversely, if JC *y and there is no edge z such that x^z^y~ and
m m
then either x y~ or JC,JT is an irreducible path, so that JCT- y"a yx.//

Theorem 50 Let E be a set with an involution without fixed points

(that is, with a map "~ such that e#e~ and e~*e). Let £ be a partial order on E
satisfying (i)-(iv) of Proposition 49. Then there is a tree T such that E(T)- E
with the given partial order on E being the same as the partial order obtained
in Proposition 49. Furtherf when E is a G-set and z and ~~ are preserved by G
then T is a G-tree.

Proof Write JC<^ if jc^.y and JC^^. Write x<y ifjc<;> but there is
no z such that jc<?<.y. It is clear from (iii) that if x<y then y" < x~.
Write x*y iff JC-J> or x<jT. We will show that s is an equivalence
relation. When this has been shown, we obtain a graph T with E(T)m E such
that V(T) is the set of equivalence classes, with JCT being the equivalence class
of JC (and JCO being defined as JC~T).

Plainly s is reflexive, and it is symmetric. Suppose that x=y and

y=z. We may assume that x*y*z (since otherwise it is clear that x = z), so that
x < y~ and y < z . Then we have z 4 y~> Now z * x requires z - x9 since otherwise
we cannot have z<y~ and JC < jr. Also z * x~ gives x* z, which again leads to
Zm x. We cannot have z *• x, since this would lead to z ^y" (because x *y~) as
well as the known z^y~> which contradicts (v) ( as already remarked, (v) follows
from (iii) and (iv».
By (ii), the remaining possibility is that z<x9 and so x*z~. We
want to show that JC < z, so let w be such that x^w^z~. Now w*y gives x*y9
which contradicts (iv) since we are given JC £jT , while w^y gives w~*z~ (as
y^Z^)9 which contradicts (v), since w^z^. If w^y" then y^w, and ymw is
impossible, as before. Thus y<wf which, with y< z and w^z~9 gives w- z.
Similarly, if w^jT then we cannot have w-y~ (since then jTzz', contradicting (v),
as we are given y^z~)9 while w<y~ together with x^w and JC <jT leads to w-JC.
By (ii) these are the only possibilities for w, and so JC < z, as required.
Thus we have constructed a graph T, which will be a Cr-graph if E
is a G-set with £ and ~ being preserved by G.
Let JC ?*jT. Then JC<.>; iff JC =y~. By definition of x this holds iff
JCT-JTT, which, by definition of o, holds iff xxmyoy which holds iff JCJ> is a path.
Suppose that x^y. Then either JC-J> or JC<>> or there is an edge z
such that J C < ^ < ^ . By (i) there are only finitely many edges between JC and y. By
induction on the number of these edges we see that if JC<J> then there are edges
Z},...,zn such that z^ - x, znm y, and z^ < z^ for / < n. Thus z^,... ,zn is a path
in T, which is irreducible since there is (by (iv) ) no edge w such that w<w".
Conversely, if there is an irreducible path whose first edge is JC and whose last
edge is yy then there are edges z^,. -. ,Zn such that z^ " xt znmy, and z-t 4 Z^% for
i<n9 and so
Now every vertex is the end of some edge. By (ii) and the
preceding paragraph, it follows that T is connected.
As already remarked, we cannot have w<w, and, by definition,
w*w. Thus w is never equivalent to w\ Hence WT*W"T, and so wx^wo, and there
are no circles. Also there can be no irreducible loops j c t , . . . ,JC for n>\9 as then
l n
we would have JC^ < JC2 < . . . < Jc^< JCj, which is absurd. Hence T is a tree.//

We now look at another important way of constructing trees, and

some of its consequences.
Let E be any set, and let A be a non-empty set. The set of all
functions from E to A will be denoted by LE9A1. When E is a (7-set we make
IE9A\ into a C7-set by defining g<p by e(g<p) - (g e)<p.

Let cp and ty be in lEyA\. We say that 9 is almost equal to i|> if

{e; e<p*ety) is finite. It is easy to see that almost equality is an equivalence
relation; the equivalence classes are called almost equality classes. When E is a
(7-set, G also acts on the set of almost equality classes. A class will be called
G-stable if it is fixed by G.
The following theorem, due to Dicks and Dunwoody (1989), is long
and difficult to prove, and I will not give the proof here. However, it is fairly
easy to use, and we will obtain an important application.

Theorem 51 (Dicks-Dunwoody Almost Stability Theorem) Let E be

a G-set such that each element of E has finite stabiliser, and let A be a
nonm empty set. Let V be a G-stable almost equality class of IE,AL Then there
is a G-tree with vertex set V whose edges have finite stabilisers.//

We can identify {E,Z*} w i t n ^ e s e t °f subsets of E, identifying 9

with ie; e<p-U. Thus we may refer to two subsets of E as being almost equal.
For cp and tp in IEJ*J\ we may define then the subset corresponding to 9 + <JJ
is the symmetric difference of P and Q (that is, the set of elements belonging to
exactly one of P and Q).
Let H be a subgroup of the group G> and let E be an //-set. On
G*E there is an equivalence relation given by ( g p ^ ) 5 ^ » ^ l^ ~$ht H such
that e~ • he* and g^ " S\n • The class of (g,e) is denoted by g * e, and the set of
equivalence classes by GxÊ; it has a (2-action defined by g(g' * e)m {gg') x e.
We can regard £ a s a subset of G* Ê by regarding e as 1 x e.
Lemma 52 Let H be a subgroup of G with l<7://l<oo, and let E be
an H-set. To each H-stable almost equality class V of C/^Z^J there is a
wn
G-stable almost equality class W of \.G* Ê^j} ose restriction to E is V.
Proof Let g* - 1,... ,g be a transversal of H in G. Then Gx Ê can
be regarded as the disjoint union of the sets gÊ for / - 1,...,«. Let v be an
element of the class V\ we regard v as a subset of E. Let w be g^vu... u gnv.
Then w is a subset of GXJJE, which we regard as an element of IGxÊ^^X
and let W be the almost equality class of w. By construction, the restriction of W
to E is V. We need to show that W is ^-stable; that is, we need to show that
gw is almost equal to w for all gcG.
Now there exists a permutation y'j,... ,y of 1,... ,/i, and elements
h*,...,h
1 M n
of H such that gg-'g:h-.
1 j£ 1
Thus gw- Ugg1f-v- Ug.j 1h-v.
1
Since h1fv is almost
equal to v, this union almost equals Ug.y-w, as required.//

Let T be a tree, v a vertex of T, and e an edge of T. We say that

e points towards v if the irreducible path from eo to v begins with e. To each v
we take the subset of those edges pointing towards v. Thus we may regard v as
an element of lEJL^l. Let w be another vertex of T. Unless e or e" is in the
irreducible path from v to w, e either points to both of v and w or to neither of
them. Hence v and w are almost equal elements of IE9Z^ Thus V(T) lies in one
almost equality class of S.EJ.jl, which we denote by 95. When T is a (7-tree it is
clear that IB is a (7-stable almost equality class.
There are elements of 95 which do not lie in V{T). We proceed to
construct a (7-map from 95 to V{T) which is the identity on V(T).
Fix some vertex v of T. Take any ae95. Then, since u and v are
almost equal, the element u + v of \.EJL~]9 regarded as a subset of E, is finite.
Let it be {e^,... .e^LTake gtstabu. Since gu - u, we have u + gu - 0, and so
v + gv- (u + v) + (gu + gv) contains only edges in ( e ^ , . . . , en, ge^y... ,ge J. We have
already remarked that an edge e occurs in v + gv iff either e or e~ is on the
irreducible path from v to gv. It follows that the distance from v to gv is at
most In. Consequently the orbit (staba)v is bounded. It follows from Proposition
30 that stabt/ fixes some vertex of T. Now choose one element u from each
(7-orbit of 95, and a corresponding element of V(T) fixed by stabw, taking the
element to be u itself when u e V. We may now extend this function to a G-map.

Lemma 53 Let H be a subgroup of G with \G:H\ <oo, and let T be

an H-tree. Then there is a G-stable almost equality class W of [C/x^i^Z^]
such that there is an H-map from W to V(T).
Proof Let OS be the almost equality class of C E , ^ containing V{T).
By lemma 52 we have a G-stable almost equality class W of LG x^EXy\ whose
restriction to E is OS. In particular, there is an //-map from W to 05- By the
previous remarks there is also an //-map from OS to V{T), and so we get an
//-map from W to V(T)J/

We can now prove some interesting theorems.

Theorem 54 Let H be a subgroup of G with \G\H\<<x>. Suppose that
there is an H-tree T whose edges have finite stabilisers. Then there is a
G-tree f whose edges have finite stabilisers and such that there is an H-map
from V{f) to V{T).

Proof All elements of the C7-set Gx^E(T^) have finite stabilisers,

since g x e - 1 x e iff g is in H and stabilises e. By the previous lemma, there is

a C-stable almost equality class W of IGx^Ei^T),^] with an //-map from W to

V(T). By the Almost Stability Theorem, there is a Cr-tree f whose edges have
finite stabilisers and whose vertex set is W.I I

Corollary Let H be a subgroup of G with \G:H\<a>. If there is an

H-tree all of whose vertices have finite stabilisers then there is a G-tree all of
whose vertices have finite stabilisers. Also, if the vertex stabilisers of the
H-tree all have order < k then the vertex stabilisers of the G-tree all have
order <k\G:H\.
Proof Let T be the given //-tree. Plainly every edge of T has finite
stabiliser. Let f be the tree obtained in the theorem. Let v be a vertex of f.
Then Istabv: //nstabvl < \G:H\. Since we have an //-map from V(f) to V(T),
//nstabv lies in the stabiliser of a vertex of T, and the result follows.//

Theorem 55 The following are equivalent : (/) G acts on a tree with

the vertex stabilisers being finite and all of order < k for some k, {ii) G is the
fundamental group of a graph of groups whose vertex groups are finite and all
of order <k for some k, (Hi) G has a free subgroup of finite index.
Corollary A torsion-free group with a free subgroup of finite index
is free.//

Proof The equivalence of (i) and (ii) is immediate from the

Structure Theorems.
Suppose that G has a free subgroup H of finite index. Then H acts
on a tree with all vertex stabilisers being trivial. Thus (i) holds by the corollary
to the previous theorem.
The next proposition shows that (iii) follows from (ii).//

We now look at free actions of a group G on a set; that is, actions

for which all elements of the set have trivial stabilisers.
Let S be any set. Then G acts on C7x5 by g(g^ 9s) - (gg^ ,s), and this
is plainly a free action. Hence any set X bijective with Gx S can be given a free
Cr- act ion.
Conversely, let G act freely on X. Let S be a transversal for the
action; that is, 5* is a subset of X with exactly one element in each Cr-orbit.
Then the function from Gx S to X which sends (g,s) to gs is easily seen to be a
C?-map which is a bijection.
It follows that G can act freely on X iff either X and G are finite

and \G\ divides 1*1 or (using known properties of infinite sets) if * is infinite
and 101*1*1.
Suppose that we also have another free (7-action on X, which we
denote by g9x. Let 5 j be a transversal for this action. Then the map sending
(g,s.) to g*s* is also a Cr-map which is a bijection. If 151 - I5J we have a
bijection from 5* to 5^ which will extend to a G-map from (7x5 to (7x5^ which
is a bijection. It follows that there will be a bijection a : * - » * such that
(gx)(x - £O(JCOC) for all x and g. We may regard the (7-actions as homomorphisms 9
and cp1 from G to Sym*, the group of permutations of X. We then have
a c Sym*, and g<pj - a (g<p)a for all g.
When X is infinite and 1(71-1*1 we may not have 15*1 - ISjI.
However, we do have 151- 15jI when * is finite and \G I divides 1*1 and when
* is infinite and \G\<\X\ (using properties of infinite sets).
We now apply these facts to the study of graphs of groups.

Proposition 56 Let & be a graph of groups. Let X be a set such

that either X is infinite and 1*1 > IGM for all vertices v or else X is finite and
\G I divides 1*1 for all vertices v. Then there is a homomorphism from n(&) to
SymX such that the maps Gv •* n(&) -» SymX are one-one for all vertices v. The
kernel of the homomorphism from 7t(@) to SymX will be free.

Remarks For any & there exists a set * such that 1*1 > l(7yl for all
v (by properties of infinite sets). Thus the proposition provides another proof of
the fact that the maps from (7y to 7i(@) are one-one.
There is plainly a finite set * such that l(7yl divides 1*1 for all v
iff the groups (7y are finite and there is a bound on their order. This completes
the proof of Theorem 55.

Proof The conditions ensure that * has a free (7y-action for all v.
Thus we have a one-one homomorphism from G to S y m * for all v.
Let e be an edge. Then we have two free actions of GQ on * , the
first coming from embedding G in Geo, and the second coming from embedding
G in G . Since \G I divides \G I when G is finite, we see that, for any edge
e, we have 16^1 < 1*1 when * is infinite and \G£\ divides 1*1 when * is finite.
By the discussion above, it follows that the two maps from G to
S y m * are conjugate. Lemma 20 now shows the existence of a homomorphism
from 7r(@) to Sym X such that Gv -» n{&) -> S y m * is one-one for all v. Further,
Corollary 2 to Theorem 27 tells us that the kernel of the map from Tt(@) to
S y m * is free.//

For those readers who know about cohomology of groups, I mention

that the structure of groups of cohomological dimension one can be obtained
from the Almost Stability Theorem by similar techniques; a group has
cohomological dimension one over a ring R iff it is the fundamental group of a
graph of groups in which all vertex groups are finite and the order of every
vertex group is invertible in R.
We say that the group G has more than one end if there is a subset
S of G, infinite and with infinite complement, such that gS is almost equal to S
for all g*G. (This notion has a geometric significance. It is possible to define
the number of ends of G in a way which essentially measures "the number of
ways of going to infinity in G"\ this number is always 0 (which holds iff G is
finite), 1, 2, or oo.) Another similar proof yields the following result. The
following are equivalent, (i) G has more than one end, (ii) there is a G-tree
whose edges have finite stabilisers and such that G stabilises no vertex, (ill) G
is either an amalgamated free product A*nB with C finite and A^C^B or an
c
_i
HNN extension <A,t; t Ct- D> with C finite or a locally finite countably
infinite group.

9 DECISION PROBLEMS

9.1 DECISION PROBLEMS IN GENERAL

Can we tell whether or not an element of a free group is in the
commutator subgroup? Whether or not it is a commutator? Given a finite
presentation, can we tell whether or not an element of the free group is 1 in the
group presented? Can we tell whether or not two finite presentations present
isomorphic groups?
These questions, and other similar ones, are obviously of interest.
We shall see later that the first two questions have the answer "Yes", while the
other two have the answer "No". The techniques used are those of combinatorial
group theory, but have no specific connection with the topological approach. I
include this chapter because I find the material particularly interesting.
In order to make these questions precise, we need to be clearer
about what is meant by "We can tell . . . ". This should mean that we can tell by
some purely mechanical process, not requiring thought. In other words, we want
to be able to feed the data to a computer and have the computer arrive at the
answer as to whether or not the required property holds for the given data.
Readers might think that a computer's ability to arrive at the answer
for given data would depend very much on the computer. This turns out not to
be so, though the speed and efficiency of the computer's answer will depend on
the computer.
More precisely, there is a class of functions (called partial recursive
functions) with the following properties. In many different models of computation
the computable functions turn out to be exactly the partial recursive functions,
while for other models of computation the computable functions are a subset of
the partial recursive functions. No-one has suggested a model of computation in
which there are computable functions which are not partial recursive. The
techniques used in proving that in certain models the computable functions are
partial recursive are sufficiently general that it appears likely that any model
suggested in future can be treated by similar methods so as to show that the

Downloaded from https:/www.cambridge.org/core. New York University, on 30 Jun 2017 at 09:32:40, subject to the Cambridge Core terms of use
Cambridge Books
, available at https:/www.cambridge.org/core/terms. Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.010
Decision problems in general 244

functions computable in that model are partial recursive. Finally, the partial
recursive functions do appear to be intuitively computable. (Here I say "do
appear to be" rather than "are" because I cannot claim to know what someone
else's intuition is; however, anyone who accepts that certain fairly simple
processes applied to intuitively computable functions yield intuitively
computable functions will find that the partial recursive functions are intuitively
computable.)
This identification of the intuitively computable functions with the
partial recursive functions is known as Church's Thesis. For the most part, we
will work with intuitively computable functions, and I hope that the reader will
find various claims made to be intuitively acceptable. It would be possible to
take a formal approach, but often the technicalities would obscure the ideas. At
a few crucial points we will have to look at some of the precise definitions. For
the general background and the detailed formalism, the reader should look at
books on the theory of computation, for instance Cohen (1987).
Certain specialised classes of computers, such as the finite-state
automata and the push-down automata, do not have the full generality we need.
They compensate for this, however, by being simpler and faster. A new theory of
automata and groups has been developed in the past few years, and is discussed
in a book-length paper in preparation by Cannon, Epstein, and others; a little
more will be said about this in Chapter 10.
We think of our computers as operating in discrete steps. For
convenience, to begin with we think of them as working with elements of IN, the
natural numbers, though in the later development it may be convenient to think
of them as working with elements of some other countable set, for instance a
free monoid M(X), where X is finite or countable. On some inputs the computer
will stop after a finite number of steps, giving us an output. On other inputs the
computer may continue computing for ever. We do not permit our computers to
stop without giving an output. (Real computers may stop giving an error
message, but we could think of this as the computer continuing for ever saying
to itself "Something's wrong".) Thus our computers enable us to compute
functions from IN to IN or, more generally, from INk to IN for some k. However it
should be noted that these functions are partial functions. That is, they need not
be functions from IN to IN but rather from a subset of IN to IN, since on some
inputs the computer runs for ever and gives no output. It might be thought that
this situation could be avoided by a careful choice of our computer programs,
but it turns out that it is an unavoidable feature of any theory of computation.
When we are looking at partial functions we will generally use the word 'partial';

a function in the ordinary sense will sometimes be called a total function for
emphasis (in Cohen (1987) and other books on computability, though, the word
'function* usually means 'partial function').
We shall need to refer to algorithms. An algorithm is a description,
in language which may be at various levels of formality, of a precise
mathematical procedure to be followed. The euclidean algorithm for finding the
greatest common divisor of two integers is a typical example. A computer
program is an example of an algorithm; it is written in a programming language
suitable for the computer being considered. As before, our algorithms operate one
step at a time, although this step may consist of performing a number of simpler
processes. For instance, we might have an algorithm one step of which involved
finding the greatest common divisor of two integers or finding the maximum
common part of two words. As before, on a given input, an algorithm may either
halt after a finite number of steps, giving an output, or it may continue for ever,
in which case it gives no output. We sometimes use the phrase 'partial
algorithm' to empasise that for certain inputs the process may not halt.

Definition Let A be a subset of IN. Then A is called recursive if

there is an algorithm which on input n has output "Yes" if n *A and has output
"No" if ni A. Recursive sets are sometimes called decidabe.
Definition Let A be a subset of IN. Then A is called recursively
enumerable (usually abbreviated to r.e.) if there is a partial algorithm which on
input n has output "Yes" iff ne A. For ni A the partial algorithm may or may not
halt.

When A is r.e. we can always modify the partial algorithm so that

if n 4 A then it does not halt. All that is needed is to add further instructions
such as "If the algorithm halts and its output is not "Yes" then go through all
the members of IN in order starting at 0."
Note that any recursive set is r.e., since the algorithm for deciding
membership of A is a suitable partial algorithm. Also, if A is recursive so is
IN - A, simply interchanging "Yes" and "No".
Any finite set is recursive. If the set is in^,... ,n^ then a suitable
algorithm is "Check whether nm n^. If so answer "Yes". If not check whether
n - «2 • • • • Check whether n - n^. If so answer "Yes". If none of these checks
have given answer "Yes" then answer "No"." It should be reasonably clear that
such sets as the set of all even numbers, the set of all squares, the set of all
powers of 3, and the set of all primes, are recursive.

It is not immediately obvious that there are sets which are r.e. but
not recursive. The proof of this fact, which we shall need, may be found in
Chapter 2 of Cohen (1987) and elsewhere. It is not difficult, but the details
would take us too far afield.
Suppose that we are given a set A, and we have a computable total
function /:IN-»IN such that A - IN/. We can regard / as enabling us to find a list
of the elements of A, namely 0/, 1/, 2/, . . . . Consequently sets of this kind are
sometimes called listable; we also regard the empty set as listable. We shall see
that listable sets are the same as r.e. sets.
One might at first sight believe that a list of the elements of A
would enable us to decide membership of A. This is not so, because we did not
require the list to give the elements of A in any sensible order. If the integer 1
has not appeared in the first 193674 terms of the list, for instance, this may be
because 1 is not on the list at all, but it could be that its first appearance on
the list is at the 193675 th term.
There is an important case where a list of members of A enables us
to decide membership of A. This occurs when we can tell for any n how far
along the list we need to check before n will occur if it is ever going to occur.
More formally, we have A - IN/, where / is computable and total, and we also
have a computable total function g such that if niA then n* mf for some m^ng.
Then, given /z, we have the answer "No" if n does not occur among 0/, 1/, . . . ,
(ng)f, and "Yes" if it does appear among these elements.
In particular, if the list is given in strictly increasing order then A
is recursive. For we then have mf> n for all m > n, so we may take ng to be n.
If the list is given in increasing order which is not strictly
increasing, it is still true that A is recursive. Here, though, we run into a subtle
point which it is easy to get confused about; similar points will also occur later.
When we claim that the set A is recursive we are simply saying that we can
show that there must be some algorithm that will work. We do not need to
exhibit the algorithm explicitly. A similar point in a less confusing form often
occurs in analytical and topological situations; for instance, Brouwer's Fixed
Point Theorem (Theorem 4.7) tells us that any continuous map from a disk to
itself must have a fixed point, but it does not tell us any way to find the fixed
point.
So let A - IN/, where / is an increasing computable total function.
Then A must be either finite or infinite, though we may not know whfch. If A is
finite then it is recursive. Let A be infinite. Given ny compute 0/, 1/,... until
(which must happen at some point) we find m with mf^n. Then n*A iff it

occurs among 0 / , . . . ,mf. However this algorithm will not decide membership of
A when A is finite, since it continues for ever when n is larger than all elements
of A. So we know that some algorithm must work to enable us to decide
membership of A, but we do not know what algorithm works.

Proposition 1 Let AcN. Then the following are equivalent-, (i) A is

r.e., {ii) A is listable, (iii) A - IN/, where f is a computable partial function, (iv)
The function which is 0 on A and undefined elsewhere is computable, (v) A is
the domain of some computable partial function.

Proof Suppose that (ii) holds. Since 0 is r.e. (the partial algorithm
which never answers "Yes" shows this), we may assume that A - IN/, where / is a
computable total function. Then a partial algorithm for membership of A is
"Given n, test 0/, 1/, and so on until, if ever, we find m with nm mf."
Suppose that (i) holds. Then a computable partial function / is
given by the following (partial) algorithm. "Given n, apply the partial algorithm
for membership of A to n. If this partial algorithm halts, then output n and halt."
Plainly IN/ is A, which proves (iii). Replacing "output n and halt" by "output 0
and halt", we obtain a partial algorithm for the function in (iv), so this function
is computable. This also shows that (v) holds.
Suppose that (v) holds. Take a partial algorithm for computing this
function /. Then a partial algorithm for membership of A is given by "On input
n, run the partial algorithm for computing /, and output "Yes" if this halts."
Suppose that (iii) holds. If A - 0 then A is listable. So we may
assume A&&. Hence there is some a^tA. Here we again encounter the subtlety
previously mentioned. We do not know a^, but we are sure that such an a^
exists. In fact, the problem really arises at an earlier stage. Given a partial
algorithm for /, we cannot tell whether A is empty or non-empty, but we know
it must be one or the other, and we use different arguments to show that A is
listable in the two cases. It is quite easy (by a variant of the technique below)
to obtain from a partial algorithm for / a partial algorithm which never halts if
A is empty and outputs an element of A if A is not empty, and we could then
use the element obtained as our choice of a~.
From a partial algorithm for / we obtain, when A&&, by the
2
procedure below a total algorithm for a function Cr:IN •*IN, so that G is a
computable total function.

On input (n,k) run the partial algorithm for / with input n for (at
most) k steps. Output the output of the algorithm for / if it has stopped in this
time; otherwise output a^. Denote the output by (n,k)G.
Then (n,k)G is either nf or a^, and, if nf is defined then there is
some k such that (n,k)G- nf. Thus IN 2 O A.
2
Now we know that IN is bijective with IN. It is easy to choose a
specific bijection / for which the formula shows that J is computable. We can
also find an algorithm for computing / . Thus the function J G is computable
and total, and its range is A> which shows that (ii) holds.//

Proposition 2 (/) Let A and B be recursive subsets of IN, and let

/:IN-»IN be a computable total function. Then AuB, AnB, and Af are
recursive.
(ii) Let C and D be r.e. subsets of IN, and let #IN-»IN be a
computable partial function. Then Cv D, Cn D, Cg" , and Cg are r.e.

Remark Note that Af is r.e. by (ii), since recursive sets are r.e.
However, Af need not be recursive, since Proposition 1 tells us that any
non -empty r.e. set can be written as IN/ for some computable total /.

Proof (i) We have total algorithms for deciding membership of A

and of B. On input /z, apply first one algorithm and then the other. Then
ntAuB iff at least one algorithm answers "Yes", while ntAnB iff both
algorithms answer "Yes".
The algorithm for deciding membership of Af" is "Apply an
algorithm which computes / to the input n. When this stops, take the output and
apply the algorithm for A to this."
(ii) Take a partial algorithm for C which does not halt on input n
if ni C , and similarly for D. The partial algorithms for CnD and Cg are
obtained in the same way as the corresponding ones in (i), but they will now
halt for some values of n but not for others.
We cannot use an algorithm similar to (i) for Cu D. For this
algorithm will not halt if niC and /zc/?, whereas we want the answer "Yes" in
this case. However, instead of running the two algorithms consecutively, we may
run them simultaneously. More precisely, we take alternately a step of one
algorithm and then a step of the other (in terms of computers, we could feed the
same input n to both computers). As soon as one algorithm halts on input nt we
output the answer "Yes". This is the required partial algorithm for CvD.

Alternatively, we could (unless C or D is empty, when the result is

obvious) find computable total functions 9 and ip such that C-INcp and Z?-INi|>.
Define d by (2/z)$-/i<p and (2« + 1)$- zap. It is easy to see that 0 is computable,
total, and with range CuD.
Finally, Cg - IN<pg. Plainly, <pg is a computable partial function, so
(ii) of Proposition 1 shows that Cg is r.e.
We could also use the following partial algorithm for Cg. "On input
iu search sytematicaily through all pairs (m,k) until, if ever, we find a pair such
that the partial algorithm for computing g, when given input m produces output
n in at most k steps. Then output "Yes"."
To complete this argument, we have to make precise the notion of
'searching sytematically'.
1 This can be done by successively testing 0/" , 1 7 ,
and so on.//

Notice that this notion of a systematic search of the elements of IN

k k
can be extended to N , using the bijection J to obtain bijections from N to N.

While a list of the elements of A does not make A recursive, two

lists, one of the elements of A and the other of the elements of N-^4, combine
to show that A is recursive. Essentially, given n we check the two lists
alternately until we see which list n occurs in. This is made precise by the
proposition below.

Proposition 3 A is recursive iff both A and N - A are r.e.

Proof If A is recursive then so is N - A. Hence both A and N - A are

r.e., since all recursive sets are r.e.
Now suppose that A and IN - A are r.e. Take partial algorithms for
the two sets. On any input one of these two must halt with answer "Yes". Hence
the following is an algorithm for deciding membership of A. "Run both
algorithms simultaneously on the input n until one of them stops with answer
"Yes". If it is the partial algorithm for A which has given answer "Yes" then
output "Yes"; otherwise output "No"."
Alternatively, take total computable 9 and ty such that A - IN9 and
N-^4-N^. Define a computable total function 0 by (2m)$mm<p and (2m +1)$- mty.
Plainly IN&-IN, and k&*A iff k is even. Thus we have the following algorithm
for deciding membership of A. "On input n, compute 0$, 10-, . . . until (which
must happen sooner or later) we find k with n-k&. Output "Yes" if k is even,
and "No" if k is odd."//

I have already mentioned that the next result is true, but will not
be proved here. It is important enough to be stated as a proposition.

Proposition 4 There is a set which is r.e. but not recursive.//

For applications to group theory we need to extend these concepts

to countable sets other than IN. This is easy to do. We simply take a bijection
9 from S to IN, and then we call a subset T of S recursive (or r.e.) if T<p is
recursive (or r.e.). All the propositions above lead at once to corresponding
results for this situation.
Thus the notion of recursive (or r.e.) subsets of S depends on the
bijection <p chosen. In practice this does not matter. There will usually be a
natural choice of one or more bijections, all of which give the same recursive
and r.e. subsets.
It is possible to be more general, which is often useful. A dodel
numbering of a countable set S is a one-one function <p from S onto a recursive
subset of IN. It turns out (see Chapter 8 of Cohen (1987)) that we may
satisfactorily define a subset T of S to be recursive (or r.e.) if its image TS^9 we may define / to be computable if the
partial function 9 / 9 | from IN to IN is computable.
Let X be a countable set. Choosing a bijection from X to IN, we
may denote the elements of X by xn for all /z e IN. This leads to the following
Godel numbering 9 of the free monoid M(X).
Let weM(X) be x. . . . x. . We define w<p to be 2kTlr~£'pJr, where
l l
\ k TIT

Pn are the prime numbers in increasing order (so p^ - 2, p^~ 3, p^ m 5, and so on).
It is easy to check that 9 is one-one and has recursive image.
When X is finite we may choose a one-one map from X into IN,
which enables us to denote the elements of X by xn for some finite set of
values of n. Then the previous formula for 9 still gives a Godel numbering for
M(X).. Also when X is countable we may, if we wish, use the same formula but
starting from a Godel numbering of X rather than from a bijection of X with IN.
When X is finite all choices of one-one maps from X into IN turn
out to leave the recursive and r.e. subsets of M(X) unchanged. This will not be
true in general when X is infinite, but they will be left unchanged if the map is
changed in a sufficiently straightforward way. Also, when X is finite we may
obtain a Go'del numbering of M(X) by taking any bijection of X with (1,...,«},

with x-i corresponding to /, say, and mapping x. £j . . . x. ij£ to X//z

r
, or, if
preferred, to Y.ifn \ all these give the same recursive and r.e. subsets of M(X)
as the previous Godel numberings.
Any homomorphism from M(X) to M(Y) is determined by its values
on X. For finite X we find that any such homomorphism is computable. For
infinite X the homomorphism will be computable provided that the map from X
to M(Y) is conputable. Both of these results are fairly easy to see, though their
precise proof requires some messy technical details.
When considering the free group F(X) we need to look at the free
monoid M(X u X~ ), where X is a set bijective with X and disjoint from it.
Thus we need a one-one map a from X u X~ into IN with recursive image.
Given a one-one map from A" to IN with recursive image, and denoting the
element (if any) which maps to n by x , we may define a by x win and
x " <x-2«+l. The details of these maps and the corresponding Godel numberings
-1
of M(X) and M(X v X ) are not important, but we need to know that such maps
exist and have the expected properties.
When we have a set which consists of all elements with some
property, it is often convenient to say "We can decide whether or not an element
has the property" when we mean that the set is recursive. Note that the words
"whether or not" are used; if we simply said "We can decide whether an element
has the property" we might only mean that the set is r.e.
Exercise 1 Find a bijection fron IN to IN which is computable and
for which (the coordinate functions of) its inverse are also computable. (One
2
possibility is to arrange the elements of IN in an infinite square array and
obtain a bijection by moving succesively along the finite diagonals. This gives
the bijection J defined by (m,n)Jm (m+n)(m+n+\)/2+m.)
Exercise 2 Suppose that we have an algorithm for a partial function
/:N-»IN. Show how to obtain an algorithm (on no inputs!) which runs for ever if
/ is not defined anywhere and which halts after a finite amount of time with
output an element of IN/ if / is defined somewhere.

9.2 SOME EASY DECISION PROBLEMS IN GROUPS

When we look at free monoids or free groups, we shall always
assume that they are countable; that is, that the basis X is finite or countable.
This is because the notion of a (partial) algorithm requires us to be working in a
countable set. We can sometimes work in uncountable groups by restricting
attention to countable parts of them.

In the monoid M(X u X ) we can decide whether or not a word is

reduced. We need only look through the word to see whether or not some x and
the corresponding x~ are next to each other or not. It should be clear that this
is an algorithm.
The function X sending each word to its reduced form is
computable. This can be shown using the canonical reduction approach to
Theorem 1.4. This method lets us compute (inductively on the length of u) the
reduced words u\ for all the initial segments u of w until we finally obtain wX
itself. Alternatively, we may observe that leftmost reduction (as discussed in
Exercise 1.2) of a word which is not reduced is a computable partial function,
and wX can be obtained by repeatedly performing leftmost reduction, checking at
eeach stage before we perform the reduction whether or not we have a reduced
word, until we reach a reduced word.
We can then decide whether or not two words give the same
element of F(X), since we need only compute their reduced forms and see
whether or not these are the same.
Let S be any set of reduced words, and let 5" be the set of those
words whose reduced form is in S. By Proposition 2 S is r.e. iff S' is r.e. and
5" is recursive if S is recursive. Also S is recursive if 5 f is recursive, since S
is the intersection of 5" with the the recursive set of all reduced words.
It follows that, when we want to ask whether certain subsets of the
free group F(X) are recursive (or r.e.), it does not matter whether we consider all
words giving members of this subset or if we only consider reduced words. We
shall use whichever form is most convenient.
The cyclic reduction of a given (reduced) word is computable. For
we need only reduce the word, and then successively remove pairs of elements
xs from the beginning and JC"£ from the end, until this can no longer be done.
It follows that we can decide whether or not two elements of F(X)
are conjugate. By Proposition 1.9, we need only take the cyclic reductions of the
two words, and then check whether or not one can be obtained from the other
by a cyclic permutation.

Proposition 5 The commutator subgroup of a free group is a

recursive subset.

Proof Let w be the word x • £ l . . . x- £/1

. We define the
L l
\ n
exponent-sum over x, of w, written wo, > to be Ee , the sum being taken over
those r such that / - k. (Sometimes it is more convenient to denote x, by a

single letter; if y is x^ we then write o rather than o^.) Plainly o^ is a

computable function.
It is now clear that w is in the commutator subgroup iff w o , - 0 for
every k such that x^ occurs in w. We can plainly check whether or not this
condition holds.//

Proposition 6 In a free group the set of commutators is recursive.

Remark The set of all words (not necessarily reduced) of form

uvu" v~ is plainly recursive. The set of commutators is the image of this set
under the map which sends each word to its reduced form. But this argument
only shows that the set of commutators is r.e.

Proof It is enough to show that the reduced word w is a

commutator iff w- abca' b c , where the right-hand side is reduced as written
(we permit one of a,b, and c to be 1). For suppose we have shown this. Then w
is not a commutator if its length is odd. If w has even length, there are only
finitely many possibilities for words a, b, and c such that abc is reduced as
written and is the first half of w. We can find all these possibilities, and for
each possibility check whether or not w is abca' b e , reduced as written.
Now abca~lb~lc~l - {ab){ca~l)(ab)~l(ca~l)~l, and so it is a
commutator.
Conversely, let w be a commutator. Then we can certainly write w
as abca b c \ where the right -hand side is not necessarily reduced (for
instance, by taking c to be 1). Among all ways of writing w in this form, choose
one for which \a\ + \b\ + Icl is as small as possible. We shall show this expression
must be reduced as written.
Suppose not. Then (other cases are possible, but they are dealt with
similarly) we may write a as uxz and b as JC~£V. Thus w- uvcx~£u v x£c .
This is of the same form as before, this time using the three elements u, v, and
d, where d is the reduced form of cx"B. Since \u\ + \v\ + \d\<\a\ + \b\ + \c\9 this
contradicts our choice of a,b, and c.lI

We can decide whether or not a finite subset of a free group is

Nielsen reduced (see section 8.2). If not, we can perform an elementary Nielsen
reduction. Continuing this process for as long as necessary, given a finite subset
A of a free group, we can find a finite subset B such that B is Nielsen reduced
and <A>-.

There is a slight problem in this. If a set is not Nielsen reduced,

there is usually more than one way of reducing it. We can either allow our
algorithms to have choices, so that the final result of the computation is just
one possible Nielsen reduced set. Alternatively, we can reorganise the algorithm
so that a specific choice is always made. For instance, we could insist that the
elements of our set are given in some specific order, and we could always make
the first Nielsen reduction possible (having specified how the reductions are
ordered).
It is not clear from the definition that we can tell whether or not a
finite subset S of F{X) is weakly reduced, since the definition requires us to
look at all products of elements of S u S" . However, Lemma 8.17 tells us that
we need only look at products of at most 2 elements, and so it is possible to
tell whether or not S is weakly reduced.

Proposition 7 Given an element w of a free group F{X) and a

finite subset A of the group, we can decide whether or not wi<A>.

Proof Begin by computing a Nielsen reduced set B such that

Let w have length n in the basis X. By Corollary 1 to Lemma 8.7,

we know that if we<Z?> then w is the product of at most n elements of BuB .
Hence, to decide whether or not w £ <A>, we need only construct the finite set of
products of at most n elements of BvB , and see whether or not w is in this
set.//

9.3 THE WORD PROBLEM

Let <X;R> be a presentation of a group G. Let N be <R>Fi<X\ We
say the presentation <X;R> has solvable word problem if N is recursive. We also
say that <X;R> has solvable conjugacy problem if i(u,v)\ u, v z F{X), and Nu and
Nv are conjugate in Gi is recursive. If <X;R> has solvable conjugacy problem
then it also has solvable word problem, since u is in N iff in G the element Nu
is conjugate to A/1. In particular, the presentation <X;0> of F{X) has solvable
word and conjugacy problem. Also, as already remarked, when considering the
word and conjugacy problems for C7, we can look at reduced words only or at all
words, whichever is most convenient.
We shall see later that when G is finitely generated and one finitely
generated presentation of G has solvable word problem then every finite
presentation of G has solvable word problem. We then say that the group G has
solvable word problem. The same applies to the conjugacy problem.

This does not hold for groups which are not finitely generated. For
instance, let A be a set which is r.e. but not recursive. Let X be (x^; /2elN}, and
let R be {JC ; neA). Then xn<L N iff mA. Since A is not recursive, it follows
from Proposition 2 that N cannot be recursive. However, G is just a free group
of countable rank, so it has a presentation <Y;Q) whose word problem is solvable.
We can write Y as {y ; «elN}, and we may take a function /:IN-»IN with
IN/-IN-A This gives rise to a function from Y into X, and so to a
homomorphism from F{Y) to F{X). However, these functions cannot be
computable. For if one of them were computable then they would all be
computable. But, by Proposition 3, IN-/* is not r.e. (as A is r.e. but not
recursive) so it cannot be the image of a computable function.

Lemma 8 Let <Y;S> be a presentation of a group H. Suppose that

there is a computable homomorphism QL:F(Y)-> F{X) which induces a
monomorphism from H to G. Then <Y;S> has solvable word problem if <X;R>has
solvable word problem. Also <S>FiY}
v
' is r.e. if N is r.e.

Remark We have already seen that a will be computable if Y is

finite, and also if Y is countable and a comes from a computable map from Y to
F(X).

Proof Let w be in F(Y). Because a induces a monomorphism, we

see that w*<S> ' ' iff WOLZN. The result follows from Proposition 2.11

Letting a induce the identity (or inclusion), the following corolllary

is immediate.
Corollary Let the presentation <X;R> of G have solvable word
problem, and let X be finite. Then any finitely generated presentation of G has
solvable word problem. More generally, any finitely generated presentation of a
subgroup of G has solvable word problem.!I

Lemma 9 Let R be recursive. Then N is r.e.

Proof We may assume that R- R~ , since, if not, we may replace R

by RuR~ , which is also recursive, by Proposition 2.
Let N* be the set of all (not necessarily reduced) words of the form
-1
u-i r^u* ... u n-1ru^ for some n, some r*R, and some words u-,i' where no
1 1 1 n n i
reductions have been made in this product. We can plainly determine all the

ways of writing a word w in the form u^ r^u^... u~ rnun for some n (this would
not be true if we were looking at the reduced form of such a word); then, since
R is recursive, for each such way we can decide whether or not every r- is in R.
Hence M is recursive.
As N is the image of N^ under the map which assigns to each
word the corresponding reduced word, we see from Proposition 2 that N is r.e.//
Note that, in particular, N is r.e. if R is finite.

Lemma 10 Let R be r.e. Then G has a presentation <X^ ;R^> such

that R« is recursive, and X is a recursive subset of X^ .

Proof Write the elements of F(X) as (w^; /zelNL By Proposition 1,

there is a computable total function /:IN-»N such that R is (vv .; /zelN). (This
holds unless R - 0, and the result is immediate in this case).
Let X^ be X u [y}9 where y is some new symbol. Thus A' is a
recursive subset of X^.
Let R. be iy) u {w -yn\ /z e IN). Plainly <X^;R^> presents G. Also,
given any element of F(X^), we can decide whether or not it is of the form wyn
for some n and some wzF{X). If so, we can decide whether or not w is w -,
since / is computable and n is known. Hence R^ is recursive.//
This procedure, and similar procedures in logic and elsewhere, is
known as Craig's Trick.

When R is recursive, we naturally refer to <Ar;/?> as a recursive

presentation of G. A group with a recursive presentation is called recursively
presentable. We have just seen that a group with a presentation in which R is
r.e. is recursively presentable.
Further, G is recursively presentable iff there is, for some X, a
homomorphism from F{X) onto G whose kernel is r.e. For Lemma 9 tells us that
the kernel is r.e. if the set of defining relators is recursive. Conversely, if the
kernel is r.e. then we have a presentation whose set of defining relators is r.e.
(namely, take every element of the kernel to be a defining relator) and Lemma
10 then tells us that there is another presentation whose set of defining relators
is recursive.
Let G be finitely generated and recursively presentable. Then
Lemmas 8, 9, and 10 tell us that for every finitely generated presentation <Y;S>
of G the group <S>F^ is r.e.

This does not hold for infinitely generated groups. For instance,
take G to be free of countable rank, and let Y be [y ; «€lN}, and let S be
iy ; ntA), where A is a set which is not r.e. However Lemma 8 tells us that,
F(Y\
v
even for infinitely generated groups, the subgroup <S> ' will be r.e. if the
map from Y to F{X) is computable.
Lemma 8 also tells us that any finitely generated subgroup of a
finitely presented group is recursively presentable (and that the same holds for
any finitely generated subgroup of any recursively presentable group). Higman's
Embedding Theorem, which will be proved later, is the converse of this; that is,
any finitely generated recursively presentable group is a subgroup of some
fintely presented group.
Lemma 11 Let R be r.e. Then N is r.e.

Proof There are two possible proofs. With the notation of Lemma
10, the inclusion of X in X-, is computable, because A' is a recursive subset of
X^ (the algorithm for computing the inclusion is "Take an element Jtj of X* .
Determine whether or not it is in X. If it is, output x* ; if not, continue
computing for ever.")
l
As R^ is recursive, <Ry> ' ' is r.e. Now apply Lemma 8.
Alternatively, let F(X) be (w^; «elN}, and let R be (w -; «e(N},
where /:IN-»IN is a computable total function. We may assume that R-R~ , since
otherwise we may replace R by RuR~ , which is also r.e. by Proposition 2. We
have the following partial algorithm for membership of N.
"At stage k, consider all words u^ r^u^...u~ rnun satisfying the
following conditions, (i) n^k, (ii) ui*{*>m\ m^k) for all i^n, (iii) r^{w
for all i^n (note that all such words can be expiictly constructed). Take the
reduced forms of the finitely many such words. Then the word w is in N iff it
appears among the reduced words obtained at some stage." //

We now look at some kinds of finitely presented groups whose

word problem is solvable. The corollary to Lemma 8 tells us that it is enough to
find one presentation which has solvable word problem.
Any finitely generated free group has solvable word problem. This
is clear, taking a free presentation of the group.
Any finite group G has solvable word problem. We take a
multiplication table presentation. This may be written (using a bijection of G
with ( 1 , . . . ,/z} for suitable n) with generators Xm ix^,... ,JC^} and defining

relations xx-mx,\ here there is one relation for each pair /,/, and, for each
I J K
fixed y, the function sending x- to the corresponding x, is a permutation of X
I K,

which is the identity when y - 1 . It is then easy to compute for each wzF(X) an
integer r such that wtNxr. In particular, w*N iff /--I, and we can check
whether or not this holds.
Any finitely generated abelian group has solvable word problem.
For it is known, from the theory of abelian groups, that any finitely generated
abelian group G has a presentation with generators x. , . . . ,x for some n, and
defining relators [x.,x.] for all / and j together with x. ' for i^k for some k,
where the d- are positive integers. Letting wo- be the exponent sum of x- in w,
we see that weAf iff wo.*0 for i> k and d. divides wo. for i<k. Since o. is a
computable function, we can decide whether or not these conditions hold.
Suppose that we have a different finite presentation of the same
group G, with generators y\>--">ym anc* defining relators including all the
commutators ly.,y.]. The general theory of this section tells us that this
presentation must have a solvable word problem. In principle, we know how to
solve the word problem for this presentation, using the homomorphism from F(Y)
to F(X) which induces the identity on G. This is not a very practical method
until we know this homomorphism. However, the general theory of abelian
groups enables us to compute explicitly from this presentation a presentation of
the form in the previous paragraph, and it also lets us obtain functions from Y
into F(X) which induce the identity on G. Then, to determine whether or not an
element of F{Y) is 1 in 6" we need only take its image in F(X) and check
whether or not this is in N.
Note that there is no reason to suppose that we can decide of a
finite presentation <X;R> whether or not the group it presents is abelian (or
finite, or free). We shall see later that these problems are not decidable.
However, there is plainly a class of finite presentations of an abelian group from
which we can tell that the group is abelian, namely, those presentations for
which every commutator of two distinct generators is one of the defining
relators.
Similarly, there is at least one finite presentation of a finite group
from which we can tell that it is finite, and at least one presentation of a free
group from which we can tell that the group is free.
By contrast, in the next two results, not only are we unable to
decide of a finite presentation whether or not the group has the relevant
property, but there is no known class of presentations such that we can tell from
the presentation that the group has the property.

Proposition 12 Finitely presentable residually finite groups have

solvable word problem.

Proof Let the residually finite group G have finite presentation

<X;R>. We know that N is r.e. Hence, by Proposition 3, it is enough to prove
that {w, wi N) is also r.e.
If w i N then w and 1 are different in C7, and so there is a
homomorphism from G to a finite group such that the image of w is not 1. Such
a homomorphism can be regarded as a homomorphism from F(X) which sends R
to 1, and homomorphisms from F(X) to a group can be identified with functions
from X to the group.
We cannot consider all finite groups, since there are too many of
them (and, in formal set theory, there is no set whose members are exactly the
finite groups). But we need only consider one group in each isomorphism class.
Now a finite group H of order n is bijective with ( 1 , . . . ,n). Using this bijection
to transform the multiplication on H into a multiplication on ( 1 , . . . ,/*}, we have
an isomorphism from H to some group whose underlying set is {1,... n).
This gives the following partial algorithm for membership of
F(X)-N.
"Given wtF{X), at stage n perform the following construction. First
take all the (finitely many) multiplication tables on {I,...,/*} and determine which
of them make ( 1 , . . . ,n) a group. For each such group, take all the (finitely
many) functions from X to {1,... ,«}. Each of these can be regarded as a
homomorphism from F{X), and for each one we determine whether or not R is
mapped to the identity. For those maps with this property, we determine whether
or not w is mapped to the identity. If there is one such map for which w is not
mapped to the identity, we answer "YesW/

Proposition 13 Countably generated recursively presented simple

groups have solvable word problem.

Proof If G is trivial, it certainly has solvable word problem. Hence

we need only consider the case when G is non-trivial. (In general, we cannot
tell from the presentation whether G is trivial or not. Hence we cannot give an
algorithm for solving the word problem; however, we show that such an
algorithm must exist.)
As before, it is enough to show that (w, wi N) is r.e.

Since G is not trivial, there is some word a not in N. (As usual,

the fact that we may not be able to find such an element a only means that we
cannot explicitly find an algorithm, but the existence of an algorithm follows
from the existence of the word a.) If wtf N then w is not 1 in G. Thus we get
the trivial group if we cancel the element w; that is, <X;R u (w)> presents the
trivial group. In particular, a€<Ruiw)>F^X\ If w*N then <Rv{w)>F^X^ is just
Nt and so, by our choice of a, we have a*<R u [w)> ' '.
Also, since we are assuming that R is recursive, the subgroup
<Ru[w}> ' ' is r.e., and, by Proposition 1, we can make a list of its elements.
Thus we have the following partial algorithm for membership of F(X)-N.
"Given wtF(X), write down the list of elements of <Ru{w)>F^X\ If
H
we find the word a in the list then answer YesV
There is an alternative algorithm which applies to all non-trivial
finitely generated simple groups (but we still cannot tell whether or not the
group is trivial). This algorithm for membership of F(X)-N is
"Given w*F(X), write down the list of elements of <Ru{w)>F^X\ If
at some stage we have found all the members of X in the list then answer
"Yes"."//

Proposition 14 Let X be finite and let R be recursive. Suppose

that every w*N- {1} contains more than half of some member of R. Then the
presentation <X;R> has solvable word problem.

Remark The hypothesis is that there are words a, b, u> and v such
that \u\ > Ivl such that w-aub, reduced as written, and uv is reduced as written
and is in R.
Certain of the small cancellation groups discussed in Chapter 10
satisfy this condition.

Proof Let U* iu; u is more than half of some member of R). Then
U is recursive. For, given u, we need only write down all the finitely many
words v with lvl<lwl and check for each of them whether or not uv is reduced as
written and lies in R.
We also have a computable function a with domain U such that UOL
is some element v of the kind considered. For instance, we could choose, among
the finitely many possibilities for v given u, that one with the smallest Godel
number.

Given a word w, we can obtain all the finitely many ways of

writing it as aub, reduced as written, and see whether or not any of these ways
has u in U. Thus the set S of all words containing more than half of a member
of R is recursive. Also the only reduced word in N but not in 5* is 1.
We can define a computable function p on S by letting wP be the
reduced form of a(uoi) b. For definiteness, we may require that among all
possibilities for u we choose that one with the smallest Godel number, and if
this u occurs more than once we take its leftmost occurrence. It is clear that, for
w e S, we have w e N iff wp e N. Also w(3 is shorter than w, by the choice of u and
UOL.

This gives the following algorithm for deciding whether or not a

reduced word is in N.
"Given w, determine whether or not w is in 5, and, if so, replace w
by wp. Now determine whether or not wP is in 5, and, if so, replace wp by (wp)p.
Repeat this process as often as necessary until we reach a reduced word wf which
is not in S. Then w is in N iff wf is l. H //

Plainly, when the above condition holds, if Iwl - n then after at most
n replacements we reach a word w' which is not in S. Thus we can say that the
solution to the word problem is linear (in terms of how many steps it takes to
make the decision, as function of the length of the word). Book (1988) shows
that the solution is linear in a more precise sense.

Proposition 15 Let G and H have solvable word problems. Then

their free product G* H also has solvable word problem.

Proof Let X generate Gy and let Y generate H, where XnYmQ, so

that X u Y generates G* H.
Any we F(X uY) can be written as u\v\- • - unvn> where u-tF(X)f
VjtF(Y), and u^ and vn may be trivial, but no other syllable u- or v. is trivial.
Let S be the set of such words for which some non-trivial u.
equals 1 in G or else some non-trivial v^. equals 1 in H. Since G and H have
solvable word problems, the set S is recursive. Further, a word w equals 1 in
G* H only if it is trivial or lies in 5, by the Normal Form Theorem for free
products.
For w e 5 , let wot be the word obtained by deleting the first syllable
which equals 1 in G or H. Then a is a computable partial function, and woe is
shorter than w. Also w - 1 in G* H iff wot - 1 in G* H.

This give rise to the following algorithm for solving the word
problem in G* H.
"Given a word wzF(XuY) decide whther or not it is in S. If so,
compute wa. Decide whether or not wa is in 5, and if so compute (wa)oc.
Continue this process until it stops. We will end with a word wf which is not in
S. Answer "Yes" if w' is trivial and "No" otherwise."//

If we want to extend this result to amalgamated free products or

HNN extensions we have to consider the positions of the amalgamated or
associated subgroups in the larger groups. This leads to the membership problem
discussed below.
Let G be <X;R>9. Let Y be a subset of F(X), and let A be <Y.
We say that the membership problem for A in G is solvable if we can decide
whether or not an element of F(X) maps by and the subset Y\ for finite X and Y the solvability or otherwise of the
problem depends only on G and A, but in general it may depend on the
generators chosen.)
Suppose that the word problem for <X;R> is solvable, the
membership problem for A in G is solvable, and that Y is an r.e. subset of F(X).
Then, for any we F(X) such that w such
that w, since Y is r.e. For each
such v we can decide whether or not wv equals 1 in G, since G has a solvable
word problem. As our element w was chosen such that w<p e A (and, by
hypothesis, for any w we can decide whether or not this holds) there must be
some element of the list with this property, and we just take the first such
element. In fact, this procedure gives an explicit expression for w<p as a product
of elements of (YvY~l)cp.
Let Y be a subset of a countable set X. It is easy to see that the
membership problem for F(Y) in F{X) is solvable iff Y is a recursive subset of
X.

Lemma 16 Let Kc HcG. If the membership problems for H in G

and for K in H are solvable then the membership problem for K in G is
solvable.

Proof Let G be <X;R> and let H be <Y;S>. Take a word w*F(X).

Then w represents an element of K only if it represents an element of H. By

hypothesis, we can decide whether or not this holds. If it does, we have seen
above that we can compute an element v of F(Y) such that w equals in G the
element of H represented by v. By hypothesis, we can tell whether or not v
represents an element of K.I I

Lemma 17 Let G and G^ be groups with solvable word problems.

Let H and H* be subgroups of G and G* respectively whose membership
problems are solvable. Then the membership problem for H*H^ in G* G^ is
solvable.

Proof Let X and X* generate G and Gj , where X n X* - 0. An

element w of F(XvX^) may be written as u\v\- - • unvn f° r some n, where
u-eF(X), v^F(X^), and u^ and vn may be trivial but all other u- and v^. are
non-trivial. Since the word problems for G and G^ are solvable we can determine
whether any non-trivial syllable represents 1 in G or G, . If so, we can obtain a
shorter word by omitting such a syllable. Thus we can compute from w another
word vvf whose non-trivial syllables do not represent 1 in G or G+ . We may as
well assume that w itself has this property.
It then follows that w represents an element of H*H* iff each u-
represents an element of H and each v. represents an element of //, . By
hypothesis, we can decide whether or not these hold.//

Proposition 18 Let G' <X;R> have solvable word problem. Let A ,

and A^ be subgroups of G whose membership problem is solvable. Suppose that
there is an isomorphism from A-, to A , which is induced by a computable
bijection from a subset Y* of F(X) to a subset Y * . Then the HNN extension
H-<G,p; p -1A^pmA_j> has solvable word problem.

Remark As Kj is the domain of a computable function it is r.e.

Then the inverse bijection from K_j to Y* is also a computable partial function.
To compute the element of Y+ corresponding to y.. we need only list the
elements of Y* , computing their images as we list the elements, and see which
one has image y_^ .
Further, there is an induced computable map from <^> to <Y_j>,
and conversely. For given any word w, we can list the products of elements of
Y^ u Y^~ until the first time, if ever, we find a product which, when reduced,
equals w, and then take the corresponding product of elements of Y^uY_^~ .

Proof H is generated by Xu{p}. Any wtF(Xv{p)) can be written as

MQP^MJP61 . . . u n lu
n.\P* ~ n where n>Oy e^.-tl, and u^F(X).
Let *S be the set of those elements for which, for some /, either
e^ - 1 - - £J._J and u. represents an element of A^ or else e ^ - 1 - - s- and u-
represents an element of A * . By Britton's Lemma any w which equals 1 in H
must either lie in S or else have /z-0. Further, if /z-0 then w-1 in / / iff w-1
in C7, and we can decide whether or not this holds, since the word problem for G
is solvable.
It is enough to show that S is recursive and that there is a
computable function a with domain S such that woe contains fewer occurrences of
p than w. For then the required algorithm for solving the word problem for H is
a very slight modification of the algorithm of Proposition 15.
Since the membership problems for both A^ and A _j are solvable, it
is easy to see that 5" is recursive.
For wtS, take the smallest possible / satisfying the relevant
conditions. For convenience, assume e^-1, so that u- represents an element of
Ay (the other case is similar). As we have seen, since the word problem for G is
solvable, we can explicitly find a product of elements of Y^ u Y^ which equals
u- (when reduced). Let v. be the reduced form of the product of the
corresponding elements of K , u K «-1 . Let woe be obtained from w by replacing
-1
p u-p by v. . Using earlier remarks, we see easily that a is a computable
function with the required properties.//
A similar result holds for amalgamated free products.
The result extends at once to a finite number of stable letters.
Suppose that we have a countable number of stable letters and
associated subgroups A- and A_^ with generating sets Y. and Y .. Then it is
not enough for the relevant conditions to hold for each /, as the algorithms for
varying i might have no connection with one another.
Instead, we need algorithms which are uniform in /. More precisely,
we need an algorithm which decides for any word w and number / whether or
not w represents an element of A-t (and similarly for A-). We also need a
computable partial function / of two variables w and / such that (w,/)/ is defined
iff W€^-, and such that, for any fixed /, the function sending w to (w,/)/ is a
bijection from Y- onto Y-. If these conditions are satisfied and the word
problem for G is solvable, the argument extends at once to show that the HNN
extension also has solvable word problem.

A group is called computable if it has a Gt3del numbering for

which there is a computable partial function n:IN •* IN such that, when m and n
are the Gttdel numbers of g and h respectively, {m,n)y. is the Gttdel number of gh.
These groups were discussed by Rabin (1960). Some of their properties are given
in the exercises below.

Exercise 3 Let G be a computable group. Show that there is a

partial computable function i:IN-»IN such that ni is the Godel number of g
when n is the Gbdel number of g.
Exercise 4 Let 9 and 4> be Godel numberings of G. Show that if
<pi|f is a computable partial function then so is ^9" .
Exercise 5 Let 9 and ip be Godel numberings of G such that 9^ is
computable. Show that if G is computable when the CJodel numbering 9 is used
then it is also computable when the Godel numbering ^ is used.
Exercise 6 Show that the free group F(X) is computable, when the
Godel numbering of F(X) comes from a Godel numbering of X.
Exercise 7 Let G be a computable group. Let AT be a recursive set
of generators of G. Show that G has solvable word problem in these generators.
(In particular, if G is finitely generated, we find that G has solvable word
problem).
Exercise 8 Let M be a subgroup of F(X) whose membership
problem is solvable (with respect to the generators X), and let 9 be the Godel
numbering of F(X) corresponding to some Godel numbering of X. Let woe be the
smallest member of iu(p; ueMw). Show that a is computable.
Exercise 9 Using the previous exercise, show that if G has solvable
word problem in some set of generators X then G is computable for a Godel
numbering such that the map from X to G is computable.
Exercise 10 Formulate and prove a result about amalgamated free
products similar to Proposition 18.

9.4 MODULAR MACHINES AND UNSOLVABLE WORD PROBLEMS

A modular machine M consists of the following data: an integer
m>\, and a number of quadruples of form (a,b,c,R) and {afb,c,L) where a, b, and
9
c are integers with 0 ^ a < m, 0 ^ b < m, 0 £ c < m , and R and L are two special
symbols (they stand for 'right' and 'left'). At most one quadruple is to begin with
a given pair a9b. (A further integer n is required if we want to define the
function computed by M> but this will not concern us.)

The modular machine M operates on IN as follows. Take any

2
(<x,P) e IN , and write a - um + a, p - vm + b, where 0 £ a < m and 0 £ b < m. If no
quadruple begins with a,b, we say that (oc,P) is terminal. If there is a quadruple
2
(a,£,c,^) then (a,P) yields (urn + c,v), while if there is a quadruple (a,b,c,L) then
2
(a,P) yields (u,vm + c).
The computation of M starting with (CXQ ,£Q) is the finite or infinite
sequence (0^,0^) of elements of IN such that (a/_i >3/_i) yields (a^P-) for all
/ > 0, and such that the last element is terminal when the sequence is finite.
When (0,0) is terminal for M we define H^{M) to be {(<x,g); the computation of
M starting at (oc.P) is finite and ends with (0,0) ).
This definition is not at all intuitive. When we discuss Higman's
Embedding Theorem in a later section, we will look at a class of machines
called Turing machines. These will be a fairly intuitive class of machines, and
they are capable of computing any computable function. We shall then be able
to indicate that modular machines are essentially a way of obtaining a numerical
coding for Turing machines, in a fashion which is particularly convenient for
group theory and some aspects of the theory of computable functions. Modular
machines are discussed in detail in Cohen (1987).
Let A be a set which is r.e. but not recursive. Then the function
which is 0 on A and not defined outside A is computable. We can then find a
modular machine which computes this function. It turns out that H^{M) cannot
be recursive, because A is not recursive.
This indicates (but we cannot give a proper proof here) that the
following result holds.

Proposition 19 There is a modular machine M such that HQ{M) is

not recursive.//

From this machine we shall be able to construct a finitely presented

group with unsolvable word problem. Another construction of such a group can
be made using Higman's Embedding Theorem. All such constructions use in
some form or another the fact that there is a set which is r.e. but not recursive.
We shall begin with the group KmKx,y, t\ xymyx>. This group can
be regarded as the free product <ix,y; xymyx>* <f> of a free abelian group of
rank two and an infinite cyclic group. It can also be regarded as the HNN
extension of the free group <jc,r> of rank two with the stable letter y and both
associated subgroups being <JC>.

Let t(r,s) be y~sx~rtxrys, and let T be the subgroup <t(r,s); all r and
s> of K. Then T is easily seen to be the normal closure <f> of <r> in K. Also
T is free with basis it(rys)}. This can be proved using the Kurosh subgroup
theorem, but it is more easily proved directly, by taking a reduced word in the
letters t(r,s) and expanding it out.
For any /, y, m, and n, the subgroup T n </(/,/), xm, yn> has basis
{/(/•,$); r=i mod m, s=J mod n). For these elements are plainly in this subgroup.
Also any element of <t(ij), xm, yn> is easily seen to be of form uxpmy<in for
some integers p and q and some element u of </(/•,£); r=i mod m, s = j mod «>,
from which the result follows.
The subgroups </(/,y), x m , j/*> for any / and j and any non-zero m
and /z are all isomorphic. For such a subgroup is conjugate to the subgroup
<r, xm, yn>, and this is the free product of <t> and the subgroup <xm, yn>. Since
<x,y> is free abelian of rank two, we see that our subgroup is isomorphic to K.
Let /, y, z'|, and y'j be arbitrary integers, and let m, n, m^, and n^ be
arbitrary non-zero integers. It follows from the paragraphs above that the map
sending t(i,j) to f(/j,y'j), xm to x™1, and ^ to j/* 1 induces an isomorphism of
<t{ij), xmy yn> with <r(/py'1), x'"1, ^ X It also induces an isomorphism of
m n
p 1 m
m n m /* iyvn + j) to
Tn<t{i,j\ x , y > with TnKHi^J^), x K /**> which sends

Now take a modular machine A/ with quadruples (a^b^c^R) and

(a>,b>,c-9L) for / in some index set / and j in some index set /. Let K^ be the
group with generators x, y, t, r- (all / €/), and /- (all ye J) and defining relations

xyyx, r7Xt(ai,bi)ri-t(ci,0), r[X xm r{-xmZ, rf^r.-y, ljlKaj,bj)lj- KO,cj),

i:Xxmlj-x, i:lymlj-ym2 (for / c / and ; « A

Then AT^ is just the HNN extension of K with stable letters r. and /. , the
subgroups <t{a.,b.), xm, ym> and </(c;-,0), JCW , j;> being associated with r-9 and
similarly for /..
We define T^ to be the subgroup of T generated by
{/(<x,P); (oc,3) € // 0 (A/)}. Because r has basis {/{/,/), all / and / } , we see that
/(«,« € r ^ iff (a>P) e // 0 (A/).
Take a quadruple (a,b,c,R) with corresponding stable letter r, and
take a pair (<x,0) such that a s a mod w and 3 = d mod m. Then the isomorphism
of associated subgroups sends /(<x,(3) to /(a',3 f ), where (a,3) yields (a',3f)- In
particular, this isomorphism maps T\, n </(<3,^), x m , ^ m > onto

7"w n </(c,0), xm , ^>. A similar result holds for a quadruple (a,b,c,L).

It now follows from Proposition 1.34 that
Kn<T 9 r ( a l 1 UI)
M i > lj ( a l 1 JfLj)>mTM'
Finally, we show that <TM, r^ /. (all /, /)>- <r/.,/.>.We need only
prove that /(a,P)c <tfr-9l> for all (oc,p) e HQ(M). We use induction on the length of
the computation. If the computation has length 0 then we must have a - 0 - p,
since (a,P) is then terminal and in //Q(A/), SO the result is immediate. Otherwise,
let (a,P) yield (a\p f ). We may suppose, without loss of generality, that the
quadruple (a,b,c,R) is the one which applies to (<x,p), and that the corresponding
stable letter is r. Then r~*t{oL,$)r - r(a\P'). Inductively, we have /(<x\P') e <tJri,/ >,
and so we also have /(<x,P) €</,/>.,/.>.

Proposition 20 Lef M be a modular machine. Let GM be

<KM, k; kf tk, kr^ r^k, kir.- l-k (all i and j)>. Then G is a finitely presented
grouk such that (oc,p)e H0(M) iff A/(a,p)- /(a,p)£. In particular, //Q(A/) w
recursive if G^ has solvable word problem.

Proof Plainly G^ is finitely presented. Also, G^ is an HNN

extension of ATw, the stable letter being k and both associated subgroups being
<t,r.,l>. Hence &/(a,P) - /(a,p)/: iff /(a,P) e <r,^.,/.>. We have seen that this holds
iff (<x,P)€//0(M).
-1 -1
Since the function sending (<x,P) to k f(<x,P)&/{a,P) is plainly
computable, it follows from Proposition 2 that HQ(A4) is recursive if C7»^ has
solvable word problem.//

Combining the purely group-theoretical Proposition 20 with the

machine-theoretical Proposition 19, the following theorem is immediate.
Theorem 21 There is a modular machine M such that the
corresponding finitely presented group G^ has unsolvable word problem. 11

9.5 SOME OTHER UNSOLVABLE PROBLEMS

By arguments similar to those in the previous section, but more
complicated, we can find a group whose word problem is solvable but whose
conjugacy problem is unsolvable.
It is possible to classify unsolvable problems according to their
degrees of difficulty. It is also possible to classify an interesting subclass of
solvable problems according to how hard it is to solve them. In both cases it is
possible to find a finitely presented group whose word problem is of a specified

degree of difficulty and whose conjugacy problem is of any harder degree of

difficulty. We shall not pursue these aspects further, as they would take us too
far into the theory of computable functions, and they are also rather technical
genralisations of the construction we made previously.

From the unsolvability of the word problem we can show that other
interesting questions about elements of a group are also unsolvable.
Proposition 22 There is a finite presentation <X;R> of a group G
such that it is not possible to decide whether or not an element of F(X)
represents in G an element of any of the following kinds: (i) an element of the
centre of G, {ii) an nth power, where n>\, (Hi) an element with only finitely
many conjugates, (iv) a commutator, (v) an element of finite order (not 1). Also
there is gtG such that we cannot decide whether or not an element of F(X)
represents an element of the centraliser of g.

Proof We start with a finitely presented group H~<Y;S> with

unsolvable word problem.. Let A be the group <a,b: b >. The group A* H will
be the required group. It has presentation <iX;R>, where XmYu{a9b) and
R-Suib3}.
Take any wzF(Y). Let W be the commutator Ca,w] in F(X). In G, W
represents a non-trivial commutator of elements of A and H if w * l in //, while
XV- 1 in G if w - 1 in H.
Since a non-trivial free product has trivial centre, W represents an
element in the centre of G iff w - 1 in H. Also, the element represented by W
commutes with the image of b iff w - 1 in //, and it lies in A iff w - 1 in H,
since otherwise it has length (in the free product sense) four.
Further, since the element represented is not in A when w^l in //,
it cannot commute with any element of A, and so it has infinitely many
conjugates (namely, the conjugates by the powers of a). Since the element has
length 4, it could only be a proper power if it were (ch)2 for some a A and
h* H. We would then have in G both c - a and c - a, which is impossible, as a
has infinite order.
Thus we could decide whether or not w - 1 in H if we could decide
of W whether or not the element it represented was in the centre of G, commuted
with the image of b, lay in A, had only finitely many conjugates, or was a
proper power. Since the word problem for H is unsolvable, none of these
problems can be decided.

Similarly, bw represents an element of order 3 if vv-1 in H and an

element of infinite order (since its length is 2) if w/1 in H. Also, La,blw
represents a commutator if w- 1 in //, while if vv^l in H then its length in G is
2. However, as in Proposition 6, we can show that any commutator in a free
product which is not in a conjugate of one or the factors may be written as
pqrp q r , where there is no cancellation between adjacent pairs, though there
may be coalescence. Analysing the possibilities for coalescence it is not hard to
see that commutators in a free product but not in a conjugate of one of the
factors have length at least 4.//

By contrast, for any finitely presented group Gm <X;R> we can

decide whether or not an element of F{X) represents an element of the
commutator subgroup G. For this holds iff it represents 1 in G/G\ and this
latter group has solvable word problem, being finitely generated abelian.
We now look at decision problems for properties of subgroups of a
given group, and of presentations of groups. We need to have some idea of
Godel numberings for these objects if we are to look at such problems. For
definiteness, I give an indication of this below, although we shall never need to
use the details.
Suppose that we have a Godel numbering of a set S> so that the
elements of S can be written as sn for certain n. We know how to find a Godel
numbering for the finite sequences of elements of S (or, equivalently, for the
elements of the free monoid M{S) ). We can regard this as giving a Godel
numbering of the finite subsets of S, regarding a subset as being a sequence
si{ ,...,sik where ^ < * 2 < . . . <ik.
If we want to look at finite presentations we cannot possibly
consider all <X,R> where X is an arbitrary finite set, as there are set-theoretical
difficulties. But, of course, we need only consider presentations <X;R> where X
is a finite subset of some fixed countable set 3B. Such a presentation gives rise
to a pair of numbers m and n, where m is the number corresponding to the finite
subset X of 3E, while n is the number corresponding to the finite subset R of
F{£). We let J(m,n) correspond to <X;R>f where / is a computable bijection from
IN to IN. Given m and n, we can tell whether or not they correspond to some X
and R, and, if so, we can tell whether or not R is a subset of F{X) and not just
a subset of FC3L). Thus we get a Godel numbering of finite presentations.

Proposition 23 The set of pairs <X;R> and <Y;S> of finite

presentations such that the corresponding groups are isomorphic forms an r.e.
set.

Proof First, replacing Y by another subset of 3E bijective with it if

necessary, we may assume that X n K - 0.
Applying Tietze transformations of type I, the group with
presentation <X\R> also has presentation <X\JY;R^>9 where R^- Ruiyw , all ytY)y
where w can be any element of F{X). Applying Tietze transformations of type
II to this, we get a presentation <X\JY\T>, where T comes from R^ by adding
certain products of conjugates of elements of R^ and their inverses.
It was shown in Theorem 1.15 that <Ar;/?>and <K;5f> present
isomorphic groups iff it is possible to find T such that the presentation
<XuY;T> also comes from <Y;S> by the same two processes.
This leads to the following partial algorithm for determining
whether the two presentations are isomorphic.
"At stage n, do the following. First take ail R^ of the form
Rviyw , all yiY) for which the elements w in F(X) have length ±n (there are
only finitely many of them). For each such R^, take all the finitely many T
which are obtained from R^ by adding some products of conjugates of elements
of R^ and their inverses, where each product has at most n factors, and each
conjugating element in F(XuY) has length < n. Perform the similar construction
on <Y;S>, and see whether or not some presentation occurs in both constructions
at this stage. If so, answer "Yes", and if not go on to the next stage."//

We now look at properties of groups, and decision problems

connected with them. The only properties we consider are those such that
whenever a group has the property then so does any isomorphic group. Thus we
do not consider the property of being a subgroup of a group of matrices, but we
could consider the property of having a monomorphism into a group of matrices.
We shall always assume that there are some finitely presented groups having the
property and other finitely presented groups not having the property.

Proposition 24 Suppose that there is an integer k such that no free

group of rank ^k has the property ^>. Then there is a finitely presented group
GL- <Y;S> such that we cannot decide of a subset Z of F{Y) whether or not the
subgroup of GL generated by Z has the property ^>.

Remark Among the properties included in the proposition are those

of being trivial, finite, cyclic, abelian, and simple, while, looking at the
complement of ^> rather than ^> itself, the property of being free is also covered.
Readers can find many other such properties for themselves.

Proof Let P be a finitely presented group having the property ^).

We may assume that P has k generators P j , . . . ,p^, since we are permitted to
increase both k (from its definition) and the number of generators of P (adding
suitable relators).
Let G- <X;R> be the group considered in Proposition 22, and use
the notation of that proposition. Define GL to be P*G.
For any w*F(Y), let u be layw1 and let v be tb,wl For i<k, let z^
l l
be p.v' uv , and consider the subgroup of GL generated by iz* , . . . , ^ ) . If vv-1 in
H then z^mPz-, and so this subgroup is P, which has the property $. We shall
show that this subgroup is free of rank k when w*l in H. Consequently the
subgroup has ty iff w-1 in H. Since the word problem for H is unsolvable, we
cannot decide whether or not this subgroup has ^>.
It is enough, by Corollary 1 to Proposition 1.8, to show that the
elements v~luv\ regarded as elements of G, are a free basis for the subgroup they
generate. For this, it is enough to show that u and v are a free basis for the
subgroup of G which they generate.
Since G- A* H9 by the Kurosh Subgroup Theorem if the subgroup
<u,v> were not free it would have a non-trivial intersection with some conjugate
of A or H. This is not possible, as we see by mapping G onto one factor. We
are left with the possibility that this subbgroup, though free, only has rank 1.
But that would require both u and v to be powers of the same element. Since
they both have length 4, and they are not equal or inverses of each other, this is
plainly impossible.//

Theorem 25 Suppose that there is a finitely presented group Q

which is not a subgroup of any finitely presented group having the property $.
Then we cannot decide of a finite presentation whether or not it presents a
group having $•

Remark The easiest way to ensure this condition is to assume that

V is hereditary; that is, that if a group has ^> then all its subgroups have $ . The
slightly weaker condition that whenever a finitely presented group has $ then all
its finitely presented subgroups have V is also enough.
In particular, ty could be any of the following properties: being
trivial, finite, abelian, cyclic, free, having a faithful representation in some
matrix group GL^K), or being residuaily finite.
By Lemma 8 V could also be the property of having a solvable
word problem.

Also, by Lemma 8 and Proposition 13, a finitely presented group

with unsolvable word problem cannot be embedded in a finitely presented simple
group. Thus ^> could also be the property of being embeddable in a finitely
presented simple group.
It follows from this theorem that the isomorphism problem is
unsolvable; that is, it is impossible to decide of two finite presentations
whether or not they present isomorphic groups. Indeed, we have the stronger
result that it is impossible to decide whether or not a finite presentation presents
the same group as the trivial presentation <0;0>, which presents the trivial group.

Proof Let P be a finitely presented group having the property $.

Let U be a finitely presented group with unsolvable word problem. Let A be the
finitely presented group U*Q, and let <X;R>9 where Xm ix^,... ,JC }, be a
presentation of A. Then A has unsolvable word problem, since its subgroup U
has unsolvable word problem.
In the free product A * the elements xfi have infinite order.
Hence the group B with presentation

is an HNN extension of A * with stable letters . ,c ^

In particular, { c ^ , . . . ,c /J+ j) is a basis for a free subgroup of B. It
follows that there is an isomorphism of <c\ , . . . ,c +«> to <c, , . . . ,c +. > sending
c. to c- for all /.
Hence we can form the HNN extension
C-<Byd; d Cjd-Cj, all /£/z + l> of B, and this will also contain A*<£>, and
hence also Q, as a subgroup.
-1 2
The group Ym<y, z\ z yzmy > is an HNN extension of the infinite
cyclic group <y> with stable letter z. In particular, <z> is an infinite cyclic
-1 2
subgroup of Y, and we may form the HNN extension T-<Y,f, t ztm z >.
In K, non-trivial powers of y and z cannot be equal, by Britton's
Lemma. It follows, using Britton's Lemma again, that {y,t) is a basis for the
subgroup of T which it generates.
Now take any word w*F{X). If w does not equal 1 in A then lw,bl
is not 1 in A*. Further, using Britton's Lemma, we then see that lw,bl is not
equal in B to any word in ic^9... 9cn+^}. By Britton's Lemma again, it follows
that {[w,£], d) is a basis for the subgroup of C which it generates.

Let Hw be obtained from the free product C*T by adding the

relations lw,bl - t and d*y. Let Gw be Hw* P.
Thus we have obtained for each wtF(X) a finite presentation of a
group Gw, and this construction is computable. Also, if w does not equal 1 in A
then Hw is just an amalgamated free product of C and T. Thus, if w*l in A
then tf contains Q, and so, by the definition of Q, the group G does not have
the property $•
On the other hand, if w- 1 in Ay it is easy to check that the
relations of Hw tell us that all the generators of Hw equal 1 in H^. Hence Gw
has the property ^ if w- 1 in ^4, since it is then just P.
That is, w-1 in A iff the presentation constructed presents a group
with the property ^>.
Since we cannot decide whether or nor vv-1 in A, it follows that
we cannot decide whether or not a finitely presented group has the property ty./S

9.6 HIGMAN'S EMBEDDING THEOREM

We have proved in Lemmas 8 and 9 that any finitely generated
subgroup of a finitely presented group is recursively presentable (and in Lemma
10 that a group with an r.e. presentation also has a recursive presentation). In
this section we prove the converse, and deduce some important consequences.

Theorem 26 (Higman's Embedding Theorem) A finitely generated

group can be embedded in a finitely presented group iff it is recursively
presentable.

Before proving the theorem, we shall obtain some propositions that

follow from it.

Corollary Let G be a countably generated recursively presentable

group. Then G can be embedded in a finitely presented group.

Proof Let G have presentation <X;R>, where X is countable and R

is an r.e. subset of F{X).
We showed in Proposition 1.36 that G can be embedded in a group
/ / with two generators. The defining relators of H are obtained in a computable
way from the members of Ry and so they form an r.e. set, being the image of an
r.e. set by a computable mapping. By the theorem, H can be embedded in a
finitely presented group K.I I

Proposition 27 Let H be a subgroup of the finitely presented group

<X;R>. Then H is the image of an r.e. subset of F(X) iff there is a fintely
presented group K containing G and a finitely generated subgroup L of K such
that G nL-H.

Remark Let H be the image of the r.e. subset S of F(X). Then H

is also the image of <5>, which is an r.e. subgroup of F(X), and the
counter-image of H in F(X) is <S u R ^ h, which is again r.e. Thus there are
several different but equivalent ways of expressing the property considered.

Proof Suppose there are such groups K and L. Let K be <yr;7*>^ and
let L be <Z>\ where Y, Z, and T are finite. There is a homomorphism from
F(X) into F{Y) which induces the inclusion of G into K. The counter-image of
H in F{X) is the counter-image under this map, which is a recursive function,
of the subgroup <Z> T F(Y)>, which is r.e. Hence the counter-image of H is
itself r.e.
Conversely, suppose that H is the image of an r.e. subset S. Let (/.
be a group isomorphic to G, with H^ being the subgroup corresponding to //,
and take the amalgamated free product A/- £*//»// 6^. Let G have the finite
presentation <X;R>. Then M is finitely generated by XvXv and has as defining
relators Ru R^uis s^, all 5 € 5), which is an r.e. set. Thus, by the Embedding
Theorem, M can be embedded in a finitely presented group K. Since G nG^m H,
by construction, the required subgroup L is just G+ ,11

Let ^ be a subset of IN which is r.e. but not recursive. Then the

m
group G <JC , all n\ x , n e A> has an r.e. presentation and an unsolvable word
problem (for this presentation, though, being free, it has a solvable word problem
when a suitable different generating set is taken) . By the corollary above, this
group can be embedded in a finitely presented group K (with generators Z), by
first embedding it in a finitely generated group H (with generators Y). Now, by
the construction of //, there is a computable function sending n to an element of
F{Y) which is equal in H to x . Since Y and Z are finite, we saw towards the
end of section 1 that any map from F(Y) to F(Z) is computable. Thus we have a
computable function sending n to an element of F(Z) which is equal in K to
xn. By Lemma 8, K has unsolvable word problem, since it contains G. This
provides another proof that there is a finitely presented group with unsolvable
word problem.

Proposition 28 There is a finitely presented group which contains a

copy of every recursively presentable group.

Proof We may restrict attention to presentations with a countable

generating set and an r.e. set of relators, since we may always enlarge a finite
set of generators to a countable one and enlarge the defining relators by making
each new generator a relator. Thus we need only look at presentations <X;R>
where A" is a fixed countable set (x^, n e IN}, and R is an r.e. subset of F(X).
Letting the elements of F(X) be {w^, n e IN}, which is more
convenient than taking a general Godel numbering, we know that there is
a computable partial function /:IN -» IN such that R is (w^, n € IN/}, and that any
such / gives rise to an r.e. presentation.
Now it may be shown that there is a computable partial function
$:1N -» IN such that for any partial computable /:IN -> IN there is some k with
nfm (n,k)$ for all n. The proof of this requires detailed knowledge of
computable functions, but an indication can be given here. The heart of the
argument is that computer programs are just words in some language, and we can
tell whether or not a given word is a program. Thus it makes sense to define 4>
by requiring (n,k)$ to be the result of applying the kth program to the input n.
We then find that $ is computable and satisfies the required property. Note also
that (by a version of Cantor's diagonal argument which is used in showing that
IR is uncountable) it is not hard to prove that in; (n,n)$ is defined} is r.e. but
not recursive.
We may then replace X by other countable sets, and we find that
for each fixed k the presentation <JC , , n e IN; w , , where nm (m9k)& for some
m> is an r.e. presentation (where w . comes from w by replacing each x; by the
rlK rl I

corresponding x-jj, and that each group with an r.e. presentation has a
presentation of this kind.
It follows that the group G with generators x^, all n and k, and
with defining relators w ^ for all n and k with n - im,k)$ for some m contains
copies of every recursively presentable group. Since $ is a computable function,
this presentation of G is itself an r.e. presentation. By the corollary to the
Embedding Theorem, G embeds in a finitely presented group, as required.//

Theorem 29 A finitely generated group has solvable word problem

iff it can be embedded in a simple subgroup of a finitely presented group.

Proof Suppose that G £ P£ H, where P is simple and H has finite

presentation <Y;S>. Let G have presentation <X;R> with X finite. Let

<p:F(X)+ F{Y) induce the inclusion of G in P. Let N be <R>F(X\

If P is trivial then so is G, and so G has solvable word problem.
Otherwise let u e F(Y) represent in H a non-trivial element of P.
As H is finitely presented, by Lemmas 8 and 9 N is r.e. Hence we
need only show that F(X)-N is also r.e.
T a k e w e F(X). If w*N t h e n wcp € <S>F^Y^9 and, as u 4 <S>F^Y)
by definition, we have u 4 <S u {w<p}> ' '.
If w 4 N, then wcp represents a non-identity element of P. Since P
is simple, we see that u e <5 u {w<p}> ' .
Thus a partial algorithm for membership of F{X) - N is to list the
v
members of <S u {w<p}> ' and answer "Yes" if u occurs in this list.
Note that this argument is almost the same as that in Proposition
13. Note also that X can be infinite provided that the map <p from X to F(Y) is
computable.
To prove the converse, take a countably generated group G whose
word problem with respect to some countable set of generators X is solvable.
We shall assume that we have a Godel numbering of X which is not necessarily
a bijection. It is not enough to assume that G is finitely generated, because our
construction below will lead to countably generated groups.
The Godel numbering of X leads to a Godel numbering of F{X), as
usual, and so also to a Godel numbering of all pairs of elements (w,v) of F(X).
Because G has a solvable word problem, there is a recursive subset A of IN such
that the pairs {u-tv-) for it A are all the pairs (u,v) such that u*l*v in G.
We may now form the HNN extension Gj of G*<y> with stable
letters t. for / e A, with t.~ u.y~ u.yt. - v.y u.y. This is an HNN extension
because the relevant elements of G* <y> are cyclically reduced of length 4, and
so both have infinite order.
Evidently we have a Godel numbering of the generators of G+ .
Also, given a word w in F{X v iy)) and a number /, we can first decide whether
or not / e A, and, if so, whether or not w equals in G*<y> a power of u-y^u-y
-1 -1 -1
(or of v.y u-y). For this holds iff w freely equals w+y w^y-.-w^ _*y W2^ (OT
the inverse of such an element) for some n, where each w^. equals u^ in G.
Because G has solvable word problem, we can decide whether or not this holds.
It follows from Proposition 18 and the comments following it that
C7j also has solvable word problem.
Since G^ is recursively generated and has solvable word problem, we
may apply the same construction to embed G^ in a group GU • We may continue
this process, getting groups G, for all k, and we let P be U(/, .

Let p and q be non-identity elements of P. They are both in G^ for

some k. By construction, there is a stable letter t and a generator y, such that
-1 - 1 - 1 P
t Py^ pyiJ " Wi, Py^ in ^£+i • Hence q e </?> , which shows that P is simple.
By the corollary to the Embedding Theorem, it is enough to show
that P is recursively presented. Let us assume that the original Godel numbering
2
of X maps into the even integers. Let /:IN -» IN be a computable bijection. In
going from G^ to G^^ we add a new generator y^ and new generators t-^ where
/ is in some recursive set A^. We get a C?odel numbering of the generators of P
by mapping y, to 2(0,£)/+ 1 and t.^ to 2(i,k)J+ 1. It is not difficult to check that
the set of defining relations is r.e., as required.//
Before proving Higman's theorem, we need to look: at the computers
called Turing machines, and their relationship with modular machines. The
discussion of Turing machines will be informal, using terms from the everyday
world.
Consider a tape, infinite in both directions, divided into squares. On
each square is written a letter from some alphabet. One letter, called "blank",
plays a special role, in that all but finitely many squares are required to have
the blank letter written on them; we simply say that such a square is blank.
In order to convey information to and from the tape, we need a
read-write head, which at any instant is positioned on some square of the tape
(consider a cassette recorder). The machine may be in any one of a finite
number of states (for instance, position of switches, whether or not a capacitor is
charged, etc.). At any instant what the machine does will depend on what state
it is in and on what symbol it is scanning; it may do one or all of changing the
symbol on the scanned square, changing its state, or moving the head one square
right or left. Instructions detailing which actions are to be performed are given
to us. Thus, started in a configuration which consists of some symbols on the
tape, a particular square being scanned, and a particular state, the instructions
will successively cause it to pass through various configurations. It may compute
for ever, or it may reach a configuration for which no instructions are given
about what to do when in that state scanning that symbol. If this happens the
machine stops.
There is only a notational change if we assume that the alphabet
consists of the numbers 0,1,... yn for some n, with 0 being the blank, and that
the states are also numbers (at most one of which can be a letter, else technical
problems arise). Take m greater than any letter or state. Then any configuration
can be described by four numbers; the state q, the letter a scanned, the number u
whose m-ary representation is the sequence of letters to the left of the scanned

square (taken from left to right, so the letter to the left of the scanned square is
the coefficient of m ), and the number v whose m-zsy representation is the
sequence of letters to the right of the scanned square (taken from right to left,
so the letter to the right of the scanned square is the coefficient of m ).
These four numbers can be conveniently combined into two, namely
either (um + a,vm + q) or (um + q,vm + a), whichever turns out to be more
convenient (sometimes one is better, sometimes the other). If we analyse what an
element of IN corresponding to a configuration changes to when we perform the
next step in the computation, we find that the change is exactly that of some
modular machine. It is this that leads to the definition of modular machines. For
more about this topic, see the book by Cohen (1987) and the papers by Aanderaa
and Cohen (1980a,b).
Turing machines can be used to compute numerical functions, and,
in addition, because of their definition, they can be conveniently used to discuss
questions of recursiveness for subsets of a free monoid without using a Godel
numbering. In particular, let S be an r.e. subset of the free monoid on the finite
set X. Then there is a Turing machine T which, started in a special starting
state with an element w of the monoid written on the tape, the scanned square
being the last letter of w, will halt iff wtS, and if wtS will halt on a blank
tape with its state being the starting state.
In particular, let C be a recursively presented group. Thus C is
<c\ , . . . ,c ; S>, where S is an r.e. subset of the free group < C j , . . . ,c^>. We can
regard a word as being an element of the free monoid on { C j , . . . yC^^y where
c
n+ici ^
corresponding modular machine A/, and the integer m which is part of the
definition of M.
To every word w there is a number a whose m-ary representation is
w (regarding the element c- for / < In as the integer /). Let / be the set of those
a € IN which correspond to some word. For a € /, let vt>a(c) be the corresponding
word. In particular, WQ(C)-1, and w.{c)-c. for /^2/z, while w a m + / ( c ) " w a ( c ) c / f° r
all a. We shall also write wa(£) and w^ibc) for the words obtained from wa(c) by
replacing each c (for j^n) by b. and {b -c.) respectively, where the b- are
new letters.
From the discussion above, when all the details are filled in, we
find that, for a e /, w^c) c S iff (a,0) € // Q (A/). It is this fact that we shall use in
proving Higman's theorem.
Let Kjy be the group contructed in the discussion of Proposition
20. Let U be the subset of K** consisting of t and all the /^symbols and

/-symbols. We will write t^ instead of t{oL,0).

Let / / j - KM*(C*<bv...,bn»*<d>, and let 6^ + / - bfl.
We know that [t^; a e /} is a basis for the subgroup it generates. We
may map H* to K*, by mapping K*, identically and mapping the other
generators to 1. Then, by Corollary 1 to Proposition 1.8, it follows that
i\ a € /} is also a basis for the subgroup it generates.
Hence we may form the HNN extension
2 j p t^ y^ ^
reference that one of the relations of H^ is p tpm td.
Let A be the subgroup <t, x, d, b• (1 <jzn),p> and, for 1 £ z £2/z, let
A- be the subgroup <^, x m , fytf, b. {\<j±n)9p>. Then A is just the HNN
extension of the free group F with basis it, x, d, b • (l^j^n)} with stable letter p,
the relations being /> <aP" r a w a ( ^ ) ^ f° r a ^ a€
^-
We know that <^ f x m >n<f a (all a€/)> is just <t* (all p c / with £=/
mod m>. Mapping F onto the free group with basis it, x) by sending d and £ - to
1, we see that <ti>xm,d9bj (W*rt)>n<r a (all ae/)> is also . The group <tifxm,d,bj (Uyi«)>n<r a w a (ft)rf (all ae/)> is, for the
same reason, <fQWo(!>)<tf (all p e / with p=/ mod /«)>. In particular, Proposition 1.34
tells us that A- is the HNN extension of the free group <t,xm, bd, b> (\£j£n)>
r w r e
with stable letter p, the HNN relations being p Qp"^D g(^)^ f° all P ^ with
ps/ mod m. Since W a / W + / W ' w a W^/» ^ *s e a s y t o s e e from this that there is an
isomorphism from A to ^ . which maps each of the stated generators of A to the
corresponding stated generator of A-.
Now let A+ be the subgroup <U,d,b. (\<j<n),p> and let A_ be the
subgroup <XJ9d,bjC. (\^j^n\p>. We will show that there is an isomorphism of
v4+ onto A_ under which the stated generators correspond.
There is a homomorphism of H~ to itself which maps each c • to 1
and maps every other generator to itself. This induces a homomorphism from A_
to A+ for which the relevant generators correspond.
We know that nK- </(a,p), all (oc,B) € H0(M)\ so that
<X0>n<fa (all a € /)> - </(a,3), all (<x,p) e //Q(A^> n O a , all oc€/>-
<r , all a c/with (a,0) € //Q(A/)>. As above, it follows that
<U, d, bj ( l * / * « » n < f a , all a € / > - < r a , all occ/ with (oc,0)€ HQ{M)>9 and that
<U,d,bj (l^y^/2)>n<r a w a (^)^, all a * f> - Kt^w^tyd, all a c / with (a,0) € HQ{Af}>.
In particular, by Proposition 1.34, v4+ is just the HNN extension of the free
product of U and the free group with basis id, b(l <j:<n)) with stable letter pt
the HNN relations being p~l t^p - t^w^ityd for all a t / with (oc,0) e // 0 (A/). Thus,
to show that the given correspondence of generators provides a homomorphism of

A+ to A_ (which will then be an isomorphism, since we have found a

homomorphism which is its inverse), it is enough, by von Dyck's theorem, to
show that p t P- tw(bc)d for all a e / with (a,0) e HJJA). Because each b-
commutes with each c., we have w^bc) - w a (£)w a (c) for all a. By the definition
of A/, we have w a (c)- 1 in C when (a,0) e HQ(M). Since p'1^" r
a
w
a^^ for al1

a, the required relation does hold.

It follows that there is an HNN extension H^ of ^ whose stable
letters are a . , l £ / £ 2 / z , and £, the subgroups associated with a^ being A and ^4-,
and the subgroups associated with k being A+ and ^4_.
Then H^ contains C, and it is plainly finitely generated. We will
show that it is finitely related, which will complete the proof of Higman's
theorem.
The relations of H^ consist of the following: (I) the finitely many
relations of AT^, the relations b.c-mcJ). for lî,j^n9 the relation p tp~ tdt and
the finitely many relations involving the stable letters of //~, (II) the relations
P t Pm t w (b)d for all a e / with aÔ, and (III) the relations corresponding to
the relators in S, that is to say the relations w a ( c ) - l for all oc € / with
(a,O)e// o (A/).
We know that one consequence of the relations of KKA is that if
-1
(<x,0) € HQ(M) then t^ can be written as a product of elements of Uulf . Then,
as a consequence of the relations involving k, we have k t k~ t if
- 1 - 1 a a
m
(a,0) € HQ(M). Since k pk- p and k dk d, we have, as consequences of the
relations (I) and (II), the relations k w a (£)&- wîb) for all a with a e / and
(<x,0) e HQ(M). However, the relations in (I) give k w^b)km w^bc)- w a (£)w a (c).
Thus the relations (III) are consequences of (I) and (II).
Regard a word w^b) as a word involving only positive powers of
b- for l £ / £ 2 / i (recall that b . is b. ). Let vv (a) be obtained from the word
w
n(b) by replacing each b- by a- for lî^2n. Ita is easy to check, by induction
-1 -1
on the length of the word, that w (a) tw (a)-t and w (a) dw (a)~ w (b)d are
m
consequences of the relations in (I). Since the relations p tp td and a- pa- p
for all / are in (I), we see that all the relations in (II) are consequences of the
relations in (I). Since (I) is finite, w e have proved Higman's Embedding
Theorem.

9.7GROUPS WITH ONE RELATOR

Groups with only one defining relator have nice properties. In
particular, their word problem is solvable. This class of groups includes the
fundamental groups of compact 2-manifolds, which are either

9
<a. ,b*9... 9a ,b ; Illa.,b-1> or < C j , . . . 9c ; Ucr'l. The word problem for these
was originally solved by Dehn (1912a,b) using geometrical methods.

Theorem 30 (Freiheitssatz) Let G be <X;r>, where r is cyclically

reduced. Let Y c X be such that r involves a generator in X-Y. Then the
subgroup of G generated by Y is free with basis Y.

Proof We use induction on the length of r. Notice first that the

result is obvious when r involves only one generator (no matter what the length
of r is), as in this case G is just the free product of a finite cyclic group and a
free group.
So we assume that r involves at least two generators. First suppose
that some generator t occurs in r with exponent sum 0. Let a,b, . . . be the
remaining generators, where a occurs in r. Taking a cyclic conjugate of r if
necessary, we may assume that r begins with a . Note that a could be any
generator occurring in r other than t.
Write a- for t lat19 etc. Because t occurs with exponent sum 0 in /•,
we can write r as a word 5 in a-, b-,... , and 5 will have shorter length than r.
2-13 2 3
For instance, if r is a t b t then s is a^ b* . Notice that our assumption on r
tells us that s begins with a^ .
Let m be the smallest index such that a- occurs in 5, and let M be
the largest index such that a- occurs in s. Thus m^O^M. It is easy to check,
using Tietze transformations, that G has the presentation

<t, a. {m^i^M), b- (all / ) , . . . ; s, t~ a.tm a.+^ {m £ i < M- 1), f b^t ~ b.+. (all
/),... X

Here we may have m - A/, in which case the relators involving a. are omitted.
Let H be the group Ka- (m ± i ^ M), b- (all / ) , . . . ; sX Since s is
shorter than r, we may assume (inductively) that the subgroups
<a- {m and <<3^+1 (m are
free with the given generators as bases, since they do not involve all the
generators occurring in 5. Note that if m - M we omit the generators a- from
these subgroups. The presentation we have obtained for G then shows that G is
an HNN extension of H.
Now let Y be a subset of X which does not contain every generator
occurring in r. Suppose first that t 4 Y. Then Y is just a subset of //, regarding
a, b,... as <2Q ,&Q , . . . . Since t does occur in r, some generator of H which is

not in Y must occur in s. The result now follows by induction.

Since a could be any element of X occurring in r other than t> if
fe Y we may assume that a 4 Y. Let we F(X) be a non-trivial reduced word
not involving a. If the exponent sum of t in w is not 0 then w cannot equal 1
in G. If the exponent sum of t in w is 0 then we can write w as a non-trivial
reduced word in b.y... . Inductively, such a word, as it does not involve the
generator <?Q which occurs in s, cannot be 1 in //, and so cannot be 1 in G.
Now t in the above argument could have been any generator which
occurs with exponent sum 0 in r. Hence the result is proved unless every
generator occurring in r does so with non-zero exponent sum. In this case we
choose t to be any generator occurring in r but not in Y, and we let a be any
other generator occurring in r.
Let the exponent sums of t and a in r be T and a (both non-zero,
by assumption). Let G^ be the group <Xfp; r, pxd>. This is a one-relator group,
since we can simply omit the generator a and the second relator. Further, we
may define q to be ///*, and we have a presentation of G* generated by
X* - ipjQ) u (X - it,a}), and with a single relator r* . It is easy to check that with
this generating set the exponent sum of p in r^ is 0, by our choice of p and q.
Thus the construction above gives G+ as an HNN extension of a group H* with
one relator Sj.
Now fj may be longer than r, because various powers of p have
been inserted. But the length of s^ is the total number of occurrences of
generators other than p in r* , and this is the same as the number of occurrences
of generators other than a in r. Hence s^ is shorter than r. It follows by the
argument of the first case that the subgroup of 6^ generated by pu(Y-ia}) is
freely generated by this set.
Now let w be a non-trivial reduced word of F{X) which only
involves the generators in Y. Let w' be obtained from w by replacing a by pT.
Then w - w' in G^. Since w' is a non-trivial reduced word in ip) u (Y- {a}), we
see that w V l in 65. , and so vv/1 in G, as required.//

When no generator occurs in r with exponent sum 0, the group G^

is just the amalgamated free product of G and the infinite cyclic group </>>, with
a being identified with p1. This is immediate, once we have see (by the
Freiheitssatz) that a has infinite order in G.
We give a similar inductive proof that the word problem for
one-relator groups is solvable. In order to give an inductive proof, we need to
prove a stronger result.

Theorem 31 Let G be <X;r>, where X is finite or countable. Let Y

be a recursive subset of X. Then the membership problem for the subgroup
<Y> of G is solvable. In particular, the word problem for G is solvable.

Proof We first note that it is enough to prove that the result holds
when Y is obtained from X by deleting one generator which occurs in r.
For suppose we have shown this. If we are given a recursive set Z
which omits at least one such generator, then we can regard Z as a recursive
subset of such a Y. Now, by the Freiheitssatz, the subgroup <y> of G is free
with basis Y. Thus the membership problem for <Z> in <Y> is solvable, and as
we are assuming that the membership problem for <Y> in G is solvable, Lemma
16 tells us that the membership problem for <Z> in G is solvable.
As a particular case, taking Z to be empty, we see that the word
problem for G is solvable.
On the other hand, suppose that Z contains all the generators
occurring in r. We can write G as the free product of a group G' with a finite
set X' of generators and one relator r and a free group with basis X", say. Now
the word problem for G' will be solvable (by the argument above). Since
<Z> - G' *<Zn X M>, we see that <Z> has solvable membership problem in G by
Lemma 17.
So we are left to prove the result when Y is obtained from X by
deleting a generator occurring in r. If R only involves one generator then G is
the free product of a finite cyclic group and a free group, and the result is then
obvious.
Suppose first that some generator t occurs in r with exponent sum
0. We use the notation of the Freiheitssatz. Thus G is an HNN extension of a
one-relator group H whose relator is shorter than r. Inductively, the membership
problems for the two associated subgroups of H are solvable. Now we may
regard any reduced word w in F(X) as a word in t, <2U,&Q,... .AS in Proposition
18 we may compute from this another word v in these latter generators such that
v has no pinches and w - v in G.
Let Y be X- {t}. Then <Y> - <rQ> c //, where YQ - {aQ ,bQ,...}.
Then w represents in G an element of <Y> iff, first, t does not occur in v and, if
so, v represents in H an element of <Yry>, and this is decidable, by induction.
Now let Y be X - {a). Let the exponent sum of t in w be x. We may
assume that T - 0, since w represents an element of <Y> iff wt~* represents an
element of <Y>. Then the exponent sum of t in the corresponding word v is also
0. Also, if w - u in G9 where u e <X - {a}>, then t has exponent sum 0 in u
(since t has exponent sum 0 in r and wu is a consequence of r).

It follows that u can be written as a word in b^ (all /) etc. (not

involving t or any a). Since v has no pinches and / does not occur in u, we
can only have vu~ - 1 in G if t does not occur in v and v - u in H.
Thus, to decide whether or not w represents in G an element of
<X - ia)> we first find v, and when t does not occur in v decide (which is
possible by induction) whether or not v represents in H an element of < £ , . . . > .
We are left with the case when no generator occurs with exponent
sum 0 in r. We use the same notation as before, and we assume that Y is X - it}.
We have seen that G^ is the amalgamated free product of G and an infinite
cyclic group. Consequently it is enough to decide whether or not a word
w € F(X) represents in G^ an element of <K>. Now we may compute an element
w* in the generators X^ of G^ such that w - w* in 6^ (simply replacing t and a
by <7P~a and pT). Thus it is enough to show that the membership problem for
<y> in <7j is solvable.
As in the proof of the Freiheitssatz, we may write G* as an HNN
extension of a one-relator group whose relator is shorter than r. The first part of
the current proof now tells us that the membership problem for the subgroup
<Kj> of (7j, where Y^ - {p} u {Y- {a))y is solvable. Also <Kj> is free with basis
Kj. The membership problem for <Y> in <Y^> is then also solvable, by Lemma
17, since Y is just (p T ,£,...}. The result follows by Lemma 16.//

Exercise 11 Let G be <X;sn\ where s is not a proper power. Show

that the only elements of finite order in G are the conjugates of powers of s.
Exercise 12 Let G be <a,b; a'lb2a - b \ and let c - b2. Show that
G is generated by (<2,c), and find a presentation in terms of these generators.
Show that there is an endomorphism of G onto itself which sends a to a and b
to c. Show, using Britton's Lemma, that this endomorphism is not one-one.

10 FURTHER TOPICS

In this chapter we will consider a few other topics in combinatorial

group theory where the methods have a topological flavour.

10.1 SMALL CANCELLATION THEORY

Small cancellation theory is one of the major aspects of
combinatorial group theory. The methods are somewhat more geometrical than
topological (insofar as it is possible to make such a distinction). I have
something of a blind spot in this area, so I only summarise the results. For
details see Lyndon and Schupp (1977). The paper by Greendlinger and
Greendlinger (1984) simplifies one of the proofs given there.
Suppose that, in F(X), we have w Ul.^u.r.u.~ , where w and each
u. and r. are reduced. Then it is possible to make a diagram in the plane,
composed of regions, edges, and vertices, with each edge being given a label
from XuX~ , in such a way that the boundary of the whole diagram is a
sequence of edges whose label (up to cyclic permutation) is w, while there are
m regions, whose boundaries consist of sequences of edges whose labels are (up
to cyclic permutation) the r..
Let R be a subset of F- F(X), and let N-<R>F. We will assume
that R is symmetrised; that is, that if rzR then all cyclic permutations of r and
of r~ are in R.
We define a piece (of R) to be an element u of F such that there
are distinct r^ and r^ in R with r^ - uv^ and r^ " uv^» b o t n reduced as written (/^
is permitted to be /•/ or a cyclic permutation of r, or ry ). For instance, if r+
3 2 2
is a b then /^ could be a ba, so that a b is a piece.
We say that R satisfies the condition C(n) if no element of R is the
product of fewer than n pieces, and we say that R satisfies the condition C\\) if
for any nR and any piece p of r we have lpl<XIH.
Another condition T{n)> n * 3, is also useful. We say that R satisfies
T(n) if whenever we have r^,... ,r^ e R for some k < n then either at least one of

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:32:29, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.011
Small cancellation theory 287

the products r^, r2r3' — • >rk-lrk9 rkr\ ls r e c i u c e c i a s written or at least one of

them is the identity.
The conditions C(n) and T{n) give conditions on the number of
boundary edges of a region and the number of edges at a vertex, respectively.
There is a well-known formula due to Euler which relates the number of
vertices, edges, and regions of a map (this formula is used in showing that there
are only five regular polyhedra). A detailed analysis of this formula, and related
properties, gives a proof of the following propositions.

Proposition 1 Let R be a symmetrised set satisfying either C\\/6)

or both C'(l/4) and T(4). Then any non-trivial element of N contains more than
half of some element of RJ/

As already remarked in Proposition 9.14, this allows us to solve the

word problem for the group <,X\R> when R is finite and satisfies the above
condition. It is also clear that the solution of the word problem is linear, in the
sense that if w*N then w is the product of at most Iwl conjugates of elements of
Rf and we can show that each conjugating element has length at most Iwl. As
remarked after Proposition 9.14, the solution is linear in another sense.

Proposition 2 There is a number K such that for any symmetrised R

satisfying either C(6), or O(4) and 7\4), or C(3) and T(6) the minimal diagram
for an element weN has at most K\w\ regions.//

Thus, under these conditions, w is the product of m conjugates of

2
elements of R> where m<K\w\ . Also, when R is finite, there is a bound (in
terms of m and Iwl, and hence in terms of Iwl) for the lengths of the conjugating
elements. Thus we can solve the word problem for such finitely presented groups
by simply checking whether or not w is one of the finitely many products of
conjugates of elements of R allowed by these conditions.
It is also possible to give similar conditions under which the
conjugacy problem can be solved.
These results have a strong geometric flavour. Proposition 2 is
closely related to the isoperimetric inequality, which states that any set in IR
2
whose boundary has a given length L has area <L /4TT, the maximum occurring
only for a disk. This result was known from classical Greek times (with a legend
relating to the foundation of the city of Carthage), but a full proof did not come
until modern times.

2
Versions of the isoperimetric inequality hold not just in IR but,
more generally, in surfaces of non-positive curvature (and in surfaces of strictly
negative curvature, there is a linear inequality, rather than a quadratic one;
Proposition 1 is connected with this fact). The various conditions imposed on R
are essentially ways of regarding the diagrams as surfaces of non-positive
curvature.
The theory can be extended without essential difficulty (but with
some complication in the details) to the situation where R is a subset of a free
product, rather than of a free group. Length will, of course, now refer to length
in the free product. We now define a piece to be a reduced word u in the free
product such that there are reduced words v« and v« and elements r« and /u of R
such that >i"Nvi and f2"wv2» wnere
these products are without cancellation but
may have coalescence. The following is one interesting application due to
Schupp (1971).

Proposition 3 Let X and Y be non-trivial groups, not both of order

2. Then every countable group is isomorphic to a subgroup of some quotient
group of X*YJ/

As remarked, the theory of small cancellation groups is closely

related to the theory of spaces of non-positive curvature. The various conditions
given in Proposition 2 are enough to show that a certain space has non-positive
curvature. Juhasz (1986, 1987a,b) has given variants of these conditions, which are
not uniform (in the sense that different conditions apply for different regions and
vertices of the diagram, while in Proposition 2 there is no distinction between
one region or vertex and another), from which similar conclusions can be drawn.
He has recently (unpublished, even as preprint, at present) shown that any
one-relator group satisfies one of a number of such conditions, from which he
has shown that the conjugacy problem for one-relator groups is solvable.
Gromov has developed a theory of hyperbolic groups, closely related
to small cancellation theory. Similar isoperimetric inequalities hold for these
groups, and so the word problem can be solved for them. A survey of this theory
is given in an article by Gromov in the book edited by Gersten (1987).
As already mentioned, there is a theory of automatic groups
developed in a long paper (of which preprints are available from Epstein) by Cannon
Epstein, Holt, Paterson and Thurston. If this theory only told of how and when
we could compute quickly in groups, it would be of practical importance but
perhaps not of great theoretical importance. However, it turns out that small

cancellation groups, hyperbolic groups, and other important groups such as some
knot groups, are all automatic groups. The theory of automatic groups is moving
very rapidly at present, and the generalisations of small cancellation theory and
related results are very interesting.

10.2 OTHER TOPICS

Whitehead (1936a,b) used topological methods to prove some results
about automorphisms of free groups. His results were given an algebraic form by
Higgins and Lyndon (1974), and Hoare (1979) gave a proof using some aspects of
graph theory.
Let F be free on the finite set X. A Whitehead automorphism a is
an automorphism which is either a permutation of XuX or is such that, for
some fixed atXvX~ , for all x x<x is one of JC, xa, a' x, or a~ xa, and aoc- a.
Plainly there are only finitely many Whitehead automorphisms.
We shall consider cyclic words; that is, cyclically reduced words,
with a cyclic permutation of a word being identified with the word itself. Two
cyclic words are equivalent if there is an automorphism which sends one to the
other. A cyclic word w is minimal if Iwl £ Iwal for all automorphisms a.

Theorem 4 Let u and v be equivalent cyclic words, and let v be

minimal. Then there are Whitehead automorphisms otj, . . . ,a^ (for some n) such
that v - wccj . . . oin and \UOL^ ... a-l £ \u\ for all i. Further, if u is not minimal then
these inequalities are all strict.//

It follows at once that we can determine for all u an equivalent

minimal cyclic word. For either u is itself minimal or its length can be reduced
by applying one of the finite set of Whitehead automorphisms. After performing
the latter operation a number of times, we will reach (and recognise) a minimal
cyclic word equivalent to u.
We can also decide for all pairs of cyclic words (and therefore for
all pair of words, since two words which give the same cyclic word come from
each other by an inner automorphism) whether or not they are equivalent. Given
any two cyclic words u and v, we can find minimal cyclic word u^ and VQ
which are equivalent to u and v, resepctively. Then u and v are equivalent iff UQ
and VQ are equivalent. If this happens, there is a sequence of distinct cyclic
words WQ,...,W for some n, all of the same length, with w~m UQ and wn~vr\>
such that w- + | comes from w- by a Whitehead automorphism. Since the words
are distinct and all of the same length, there is a computable upper bound for n,

and we simply have to see whether or not VQ can be obtained from u^ by

applying a sequence of at most this number of Whitehead automorphisms.
More detailed analysis enabled McCool (1975b) to prove the
following.

Proposition 5 The subgroup of the automorphism group of F which

fixes a finite set of words (or cyclic words) is finitely presentable; further,
there is a procedure to determine a finite presentation of this subgroup.

McCool (1974) was also able to obtain explicitly a presentation of

AutF, and later (1975a) to obtain by Tietze transformations another presentation
which was originally obtained using difficult methods by Nielsen.
Generalisations of the theory were obtained by Gersten (1984) who
was able to prove that we can decide, given two subgroups H and K, whether or
not there is an automorphism of F which sends H to K.
Collins (in an article in the book by Gersten and Stallings (1987)
and elsewhere) has extended the theory of Whitehead automorphissm to free
products.

Howie (1981) uses CW-complexes (the topological form of the

complexes we have discussed) to obtain some results on equations over groups.
This is just one of many places where deeper aspects of topology than those we
have looked at enable us to prove results in group theory.

According to Propositions 8.47 and 8.48 a set V is the set of

vertices of a tree iff it is a metric space with an integer-valued metric d such
that for any a and b in V the function sending x to (a,x)d is a bijection from
{x; (a,x)d + (x,b)d- (a,b)d) to the interval l.Q,(a,b)d]. If we removed the
condition that the metric is integer-valued we would get a space which is called
an R-tree. More generally, we could require the function d to map into any
ordered group A, and we would obtain a A-tree. The theory of these parallels
some topological results in an important way, in which not only does the
topology influence the group theory, but the group theory feeds back into the
topology. For more about this subject see an article by Shalen called
"Dendrology of groups" in the book edited by Gersten (1988).

NOTES AND REFERENCES

These notes and references are highly selective. For some topics 1
have listed all or most relevant papers, while for others I have only listed those
I have actually used. In particular, the work of Bernard Neumann and of Roger
Lyndon is so influential in combinatorial group theory that, rather than making
references to their work on almost every page, I have cited their papers only
occasionally.

CHAPTER 1
Van der Waerden's method for proving the Normal Form Theorem
for free groups appears in van der Waerden (1948). The Diamond Lemma
approach, in the form given in the exercises, appears in Newman (1942). See also
Bergman (1978) for more about this interesting topic, which comes up in many
situations requiring normal forms.
Von Dyck's theorem is one of the earliest in combinatorial group
theory, dating from 1882.
Britton's Lemma was first proved in Britton (1963). The argument
that proves both Britton's Lemma and the results on residual finiteness of
HNN extensions appears in Cohen (1977) and in Baumslag and Tretkoff (1978).
The general theory of HNN extensions was begun by Higman,
Neumann, and Neumann (1949), in which theorems 36-39 occurr. The finitely
generated infinite simple group was obtained by Higman (1951), while Schupp
(1971) gives a proof that any countable group embeds in a quotient of the group
G of Theorem 40.
Miller and Schupp (1973) give a proof of Britton's Lemma using
small cancellation theory (see Chapter 10), and this does not require the Axiom
of Choice.
Further general results on combinatorial group theory can be found
in the books by Magnus, Karrass, and Solitar (1966) and by Lyndon and Schupp
(1977), and the historical account of Chandler and Magnus (1982). For more

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:32:32, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.012
Notes and references 292

about the topological connections see the long articles by Collins and Zieschang
(to appear) and by Scott and Wail (1979), as well as some of the books cited.

CHAPTER 2
For more about spaces, paths, etc., see books on topology, in
particular Brown (1968, 1989) and Massey (1967).

CHAPTER 3
The book by Higgins (1971) gives a full account of the theory of
groupoids and its applications (many of which are algebraic forms of our later
geometric theorems).

CHAPTER 4
Van Kampen's theorem appears in van Kampen (1933). A very
similar result had been obtained earlier by Seifert (1931), and the theorem is
sometimes referred to as the Seifert-van-Kampen theorem. The proof using
groupoids is due to Brown (1967).
The example of two contractible spaces whose join is not simply
connected comes from Griffiths (1954), but has been modified so that the
calculation of the fundamental group of the Hawaian earring in Griffiths (1956) is
not needed. Morgan and Morrison (1986) show that Griffiths' proof (1956) is
wrong, and they provide a correct proof of his result. Lemma 25 and Proposition
26 come from their 1986 paper, with the proofs changed slightly to tidy up some
minor difficulties in their argument.

CHAPTER 5
The result that Fixa is finitely generated when a is an
automorphism of a finitely generated free group has proofs by Culler and
Vogtmann (1986), Cooper (1987), and Gersten (1987), as well as the proof by
Goldstein and Turner (1986) which I have followed. The results on the rank of
Fixa come from M. Cohen and Lustig (1988). Bestvina and Handel (1989) show
that Fixa has rank at most rankF.

CHAPTER 1
The Schreier and Reidemeister-Schreier theorems were proved in
Schreier's paper (1927) and Reidemeister's book (1932, 1950). The Kurosh
subgroup theorem appears first in Kurosh (1934).

The covering space argument for the Kurosh subgroup theorem

appears in Baer and Levi ((1936). The algebraic proof of the Reidemeister-Schreier
and Kurosh theorems is given in the form due to Weir (1956).
Further applications of covering spaces to group theory may be
found in papers by myself (1963) and by Tretlcoff (1973a, 1973b, 1975).

CHAPTER 8
The general results of Bass-Serre theory may be found in the book:
by Serre and its translation (1977, 1980).
Theorem 6 was proved by Serre; essentially the same proof, but
without the details, can be found in the book by Reidemeister (1932, 1950).
Nielsen (1921, more easily available in the 1986 collection of his
papers) proved his theorem for finitely generated subgroups. The general case is
due to Federer and Jonsson (1950).
The discussion of weakly reduced sets comes from Hoare (1976, and
also 1981).
The theory of abstract length functions on groups was started by
Lyndon (1963). This has been developed in many ways, including the cited
papers by Hoare and papers by Chiswell (1976a, 1979a, 1981).
In the proof of the First Structure Theorem I have followed the
approach of Chiswell (1979b).
A detailed proof of the subgroup theorems for amalgamated free
products and HNN extensions using Bass-Serre theory was first given by Cohen
(1974). Earlier purely algebraic proofs were given by Karrass and Solitar
(1970, 1971), improved by Karrass, Pietrowski, and Solitar (1974). Their
method was extended to arbitrary tree products by Fischer (1975).
The version of the Kurosh Subgroup Theorem obtained from
Theorem 27 is slightly weaker than that in Chapter 7, but is almost always
enough for any uses we need to make of it. To obtain the precise form we must
see how a Kurosh system gives rise to a representative tree.
Let G- *C7a. Then G is the fundamental group of a graph of groups
(®,X), where X has vertices x^ at which the vertex group is G^, and another
vertex JCQ which has trivial vertex group, with one edge pair joining JCQ and x^.
The universal cover of @ is a C7-tree K, which has a vertex y^ with trivial
stabiliser, vertices y with stabiliser G , an edge-pair ie ,e~ } joining y^ and y ,
with every vertex and edge being in the (7-orbit of one of these.
Suppose that we are given a Kurosh system. Let S be the subgraph
of Y with vertices uy^ and uy^ for all a-representatives in the large and all a,

together with the edges joining uy^ and uy^ and the edges joining uy^. and uy&
whenever u ends in GU. The conditions on a Kurosh system ensure that S is
connected.
Choose some index, which we denote by 1. Obtain T from S by
adding those vertices gy^ and the edges joining gy^ to gy^ for all g such that g
is a 1-representative for which Hgy^n S * 0.
Unless there are representatives in the large u and v with Hu - Hv,
it is easy to check that T is a representative tree for //, and the detailed
application of the Second Structure Theorem will give the Kurosh Subgroup
Theorem in the precise form obtained in Chapter 7.
It is part of the algebraic proof of the theorem that we must have
Hu^Hv for distinct representatives in the large u and v. This should be easy to
prove directly, but I am not sure how. It is easy, however, to find a Kurosh
system which obviously satisfies this condition. We need only construct the
Kurosh system as in Chapter 7 by induction on the length of the shortest
element of the double coset, and, for the double cosets of a given length,
construct the representatives in the large in some order (for instance, using a
well-ordering of the index set), ensuring that at each step we choose our
representative among the representatives already chosen, if this is possible.
The results on finitely generated subgroups are due to Cohen (1974,
1976), based on algebraic arguments of Burns (1972, 1973). The proof of
Grushko's theorem is based on that of Stailings (1987). The original proof is due
to Grushko (1940), with additions and extensions from Burns and Chau (1984),
Chiswell (1976b), Higgins (1966), Neumann (1943), Stailings (1965), and Wagner
(1957). M. Hail's theorem first appears in Hall (1949b); see also Burns (1971b).
Tretkoff (1975) first proved that the free product of M. Hall groups is an M. Hall
group. The proof given here is new, but is essentially no more than a translation
into Bass-Serre theory of the proofs of Tretkoff and of Imrich (1977). Howson's
theorem was proved in Howson (1954); the graphical proof was first published by
Imrich (1977). An improvement of the bound was given in an algebraic proof by
Burns (1971a), which has been given a graphical form by Nickolas (1985). These
improvements are still of the order of magnitude 2(rank-4- l)(rank/?-1) rather
than the conjectured bound (which we know could not be improved) of
(rank.4-l)(rank£-1) + 1.
The methods of Imrich (1977) are graphical but rather different from
those of Bass-Serre theory, and are an interesting alternative.
The Almost Stability Theorem appears in the book by Dicks and
Dunwoody (1989), which gives an account of Bass-Serre theory and some other of

its applications with a very different flavour from the approach here. The results
on ends of groups first appear (for finitely generated groups) in Stallings (1968),
which also proves that groups of cohomological dimension 1 over Z are free.
Swan (1969) extended this result on cohomological dimension to arbitrary groups;
The book: by Cohen (1972) gives a simpler account of these results. The structure
of groups with a free subgroup of finite index was found by Karrass, Pietrowski,
and Solitar (1973), Cohen (1973), and Scott (1974). Readers who know this
material should not be surprised that they are proved easily from the Almost
Stability Theorem, as the long proof of that theorem uses most of the ideas from
the earlier proofs. The great advantage of the Almost Stability Theorem is that it
offers the possibility of obtaining new results using that theorem, rather than
having each time to give a new proof with only slight differences from the old
ones (as happened, for instance, in extending the Stallings-Swan theorems to
groups with a free subgroup of finite index).
Theorem 8.50 first appears in Dunwoody (1979), where he uses it to
obtain Stallings' result on finitely generated groups with more than one end.
Dunwoody regards this theorem as superseded by the Almost Stability Theorem,
but I do not agree, as I have a fondness for this theorem. I had been
endeavouring to find a proof of Stallings' theorem by obtaining a suitable tree
on which the group acted. It was clear what the edges of the tree should be, but
I could not discover what the vertices were to be. Dunwoody's insight was that
it was unnecessary to find the vertex set explicitly; this kind of insight is an
aspect of mathematics which I find very pleasing.

CHAPTER 9
For more about computability in free groups, see Boydron and
Truffault (1974), Hoare (198), and Petresco (1962), among others. For instance,
given finite subsets A and B of a free group, we know by Howson's Theorem
8.33 or 8.35 that there is a finite set C such that <A>n-<O; it is possible to
compute C, given A and B. There are also significant questions about the
difficulty of solving problems in a free group; certain problems, for instance, are
solvable in polynomial time, but are, in a sense which can be made precise, the
most difficult problems solvable in polynomial time, as is shown by Avenhaus
and Madlener (1981, 1982, 1984a,b).
Theorem 9.29 is due to Boone and Higman (1974).
The first account of the word problem is in Dehn (1912a,b;
translated in the book Dehn (1980) )

The first proof that there was a group with unsolvabie word
problem is due to Novikov (1955). This proof was simplifies by Boone (1959) and
Britton (1958, 1963). The proof given here, taken from Aanderaa and Cohen
(1980a) is the shortest.
The Embedding Theorem comes from Higman (1961). A proof by
Valiev is given in Lyndon and Schupp's book. Other proofs occur in the books
by Rotman (1973) and Shoenfield (1967). The proof here comes from Aanderaa
and Cohen (1980b), which is again shorter than the other proofs, though it
possibly gives less insight into why the theorem holds than do some other proofs
(especially that in Shoenfield's book). The traditional proofs for several
unsolvability results uses Turing machines. It seems that for group-theoretic
decision problems the use of modular machines frequently simplifies the proofs
considerably.
The various other unsolvability results proved comes from Baumslag,
Boone, and Neumann (1959) and Rabin (1958); similar results were proved by
Adian (1955, 1957, 1958).
Stilwell (1982) gives a survey of the word problem and related
results, with further properties and more historical detail.
The results on one-relator groups were originally proved by Magnus
(1930, 1932). Here we follow the approach of McCool and Schupp (1973).

BIBLIOGRAPHY

BOOKS
Adian, S.I., Boone, W.W., and Higman, G., eds. (1980) Word
Problems II: the Oxford Book (North-Holland, Amsterdam)
Appel, K.I., Ratcliffe, J.G., and Schupp, P.E., eds. (1984)
Contributions to Group Theory (Contemporary Mathematics 33, American
Mathematical Society, Providence, Rhode Island)
Brown, R. (1968) Elements of Modern Topology (McGraw-Hill, New
York; new edition, Ellis Horwood, Chichester, 1989)
Chandler, B. and Magnus, W. (1982) The History of Combinatorial
Group Theory: a case study in the history of ideas (Springer, New York)
Cohen, D.E. (1972) Groups of Cohomological Dimension One
(Lecture Notes in Mathematics 245, Springer, Berlin)
Cohen, D. E. (1987) Computability and Logic (Ellis Horwood,
Chichester)
Dehn, M. (1987) Papers on Group Theory and Topology (translated
and introduced by J. Stilwell) (Springer, New York)
Dicks, W. (1980) Groups, Trees and Projective Modules (Lecture
Notes in Mathematics 790, Springer, Berlin)
Dicks, W. and Dunwoody, M.J. (1989) Groups Acting on Graphs
(Cambridge University Press, Cambridge)
Gersten, S.M., ed. (1987) Essays in Group Theory (Springer, New
York)
Gersten, S.M. and Stallings, J.R., eds. (1987) Combinatorial Group
Theory and Topology (Annals of Mathematics Studies 111, Princeton University
Press, Princeton)
Higgins, P.J. (1971) Notes on Categories and Groupoids (van
Nostrand, London)
Johnson, D.L. (1976) Presentations of Groups (London Mathematical
Society Lecture Notes 22, Cambridge University Press, Cambridge)
Lyndon, R.C. and Schupp, P.E. (1977) Combinatorial Group Theory
(Springer, Berlin)
Magnus, W., Karrass, A., and Solitar, D. (1966) Combinatorial Group
Theory (Wiley, New York)
Massey, W.S. (1967) Algebraic Topology: an introduction (Harcourt,
Brace, and World, Inc., New York)
Nielsen, J. (1986) Collected Mathematical Papers (Birkhauser, Boston)

Downloaded from https:/www.cambridge.org/core. Cornell University Library, on 30 Jun 2017 at 09:34:46, subject to the Cambridge Core terms
Cambridge Books Online
of use, available at https:/www.cambridge.org/core/terms. © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878.013
Bibliography 298

Reidemeister, K. (1932) Einfuhrung in die Kombinatorische

Topologie (Vieweg, Braunschweig; reprinted, Chelsea, New York, 1950)
Rotman, JJ. (1973) The Theory of Groups (Aiiyn and Bacon, Boston)
Serre, J.-P. (1977) Arbres, Amalgames, SL2 (Soc. Math, de France,
series Asterisque 46)
Serre, J.-P. (1980) Trees (Springer, Berlin)
Shoenfield, J. (1967) Mathematical Logic (Addison-Wesley, Reading.
Massachussets)
Staiiings, J. R. (1971) Group Theory and Three-Dimensional
Manifolds (Yale University Press, New Haven)
Stilwell, J. (1980) Classical Topology and Combinatorial Group
Theory (Springer, New York)

PAPERS
Aanderaa, S. and Cohen, D.E. (1980a) Modular machines, the word
problem for finitely presented groups, and Collins* Theorem, in Word Problems
II: the Oxford Book (North-Holland, Amsterdam), 1-16
Aanderaa, S. and Cohen, D.E. (1980b) Modular machines and the
Higman-Clapham-Valiev embedding theorem, in Word Problems II: the Oxford
Book (North-Holland, Amsterdam), 17-28
Adian, S.I. (1955) Algorithmic unsolvability of problems of
recognition of certain properties of groups, Dokl. Akad. Nauk SSSR (103), 533-535
Adian, S.I. (1957) The unsolvability of certain algorthmic problems
in the theory of groups, Trudy Moskov. Mat. Obsc. (6), 231-298
Adian, S.I. (1958) On algorithmic problems in effectively complete
classes of groups, Dokl. Akad. Nauk SSSR (123), 13-16
Althoen, S.C. (1974) A geometric realisation of a construction of
Bass and Serre, J. Pure Appl. Alg. (5), 233-237
Avenhaus, J. and Madlener, K. (1981) P-complete problems in free
groups, in Theoretical Computer Science: 5th GI conference, Karlsruhe (Springer
Lecture Notes in Computer Science 104), 42-51
Avenhaus, J. and Madlener, K. (1982) The Nielsen reduction as key
problem to polynomial reductions in free groups, in Computer Algebra:
EUROCAM 82 (Springer Lecture Notes in Computer Science 144), 49-56
Avenhaus, J. and Madlener, K. (1984a) The Nielsen reduction and
P-complete problems in free groups, Theoretical Computer Science (32), 61-76

Avenhaus, J. and Madlener, K. (1984b) On the complexity of

intersection and conjugacy problems in free groups, Theoretical Computer Science
(32), 279-295
Baer, R. and Levi, F. (1936) Freie Produkte und ihre Untergruppen,
Compositio Math. (3), 391-398
Baumslag, B. and Tretkoff, M. (1978) Residually finite HNN
extensions, Comm. Alg. (6), 179-194
Baumslag, G. (1963) On the residual finiteness of generalised free
products of niipotent groups, Trans. Amer. Math. Soc. (106), 193-209
Baumslag, G., Boone, W.W., and Neumann, B.H. (1959) Some
unsolvable problems about elements and subgroups of groups, Math. Scand. (7),
191-201
Bergman, G.M. (1978) The diamond lemma for ring theory, Advances
in Math. (29), 178-218
Bestvina, M. and Handel, M. (1989) Train tracks and automorphisms
of free groups (preprint)
Book, R.V. (1988) Dehn's algorithm and the complexity of word
problems, American Math. Monthly (95), 919-925
Boone, W.W. (1959) The word problem, Ann. of Math. (70), 207-265
Boone, W.W. and Higman, G. (1974) An algebraic characterisation of
the solvability of the word problem, J. Austral. Math. Soc. (18), 41-53
Boydron, Y. and Truffault, B. (1974) Problemes de Tordre generalisee
pour les groupes libres, C.R. Acad. Sc. Paris (279), 843-845
Britton, J.L. (1958) The word problem for groups, Proc. London
Math. Soc. (third series, 8), 493-506
Britton, J.L. (1963) The word problem, Ann. of Math. (77), 16-32
Brown, R. (1967) Groupoids and van Kampen's theorem, Proc.
London Math. Soc. (third series, 8), 385-401
Burns, R.G. (1969) A note on free groups, Proc. Amer. Math. Soc.
(23), 14-17
Burns, R.G. (1971a) On the intersection of finitely generated
subgroups of a free group, Math. Zeit. (118), 121-130
Burns, R.G. (1971b) On finitely generated subgroups of free products,
J. Austral. Math. Soc. (12), 358-364
Burns, R.G. (1972) On the finitely generated subgroups of an
amalgamated product of two groups, Trans. Amer. Math. Soc. (169), 293-306
Burns, R.G. (1973) Finitely generated subgroups of HNN groups,
Canadian. J. Math. (25), 1103-1112

Burns, R.G. and Chau, T.C. (1984) Another proof of the

Grushko-Neumann-Wagner theorem for free products, in Contributions to Group
Theory (op. cit.), 116-133
Chiswell, I.M. (1976a) Abstract length functions in groups, Math.
Proc. Camb. Phil Soc. (80), 451-463
Chiswell, I.M. (1976b) The Grushko-Neumann theorem, Proc. London
Math. Soc. (third series, 33), 385-400
Chiswell, I.M. (1979a) Embedding theorems for groups with an
integer-valued length function, Math. Proc. Camb. Phil. Soc. (85), 417-429
Chiswell, I.M. (1979b) The Bass-Serre theorem revisited, J. Pure
Appl. Alg. (15), 117-123
Chiswell, I.M. (1981) Length functions and free products with
amalgamations of groups, Proc. London Math. Soc. (third series, 42), 42-58
Cohen, D.E. (1963) A topological proof in group theory, Proc.
Camb. Phil. Soc. (59), 277-282
Cohen, D.E. (1973) Groups with free subgroups of finite index, in
Conference on Group Theory (Springer Lecture Notes in Mathematics 319,
Springer, Berlin)
Cohen, D.E. (1974) Subgroups of HNN groups, J. Austral. Math. Soc.
(17), 394-405
Cohen, D.E. (1976) Finitely generated subgroups of amalgamated
free products and HNN groups, J. Austral. Math. Soc. (22), 274-281
Cohen, D.E. (1977) Residual finiteness and Britton's Lemma,
J. London Math. Soc. (second series, 16), 232-234
Cohen, M.M. and Lustig, M. (1988) On the dynamics and the fixed
subgroup of a free group automorphism (preprint)
Collins, D.J. and Zieschang, H. (to appear) Combinatorial group
theory and fundamental groups (Encyclopaedia of Mathematics, USSR)
Cooper, D. (1987) Automorphisms of free groups have finitely generated
fixed point set, J. Alg. (Ill), 453-456
Culler, M. and Vogtmann, K. (1986) Moduli of graphs and outer
automorphisms of free groups, Invent. Math. (84), 91-119
Dehn, M. (1912a) Transformation der Kurve auf zweiseitigen
Flachen, Math. Ann. (72), 413-421
Dehn, M. (1912b) Uber unendliche diskontinuierliche Gruppen, Math.
Ann. (71), 116-144
Dey, I.M.S. (1965) Schreier systems in free products, Proc. Glasgow
Math. Assoc. (7), 61-79

Dunwoody, M.J. (1979) Accessibility and groups of cohomologicai

dimension one, Proc. London Math. Soc. (third series, 38), 193-215
Federer, H. and Jonsson, B. (1950) Some properties of free groups,
Trans. Amer. Math. Soc. (68), 1-27
Fischer, J. (1975) The subgroups of a tree product of groups, Trans.
Amer. Math. Soc. (210), 27-50
Gersten, S.M. (1984) On Whitehead's algorithm, Bull. Amer. Math.
Soc. (10), 281-284
Gersten, S.M. (1987) Fixed points of automorphisms of free groups,
Advances in Math. (64), 51-85
Goldstein, R.Z. and Turner, E.C. (1984) Automorphisms of free
groups and their fixed points, Invent. Math. (78), 1-12
Goldstein, R.Z. and Turner, E.C. (1986) Fixed subgroups of
homomorphisms of free groups, Bull. London Math. Soc. (18), 468-470
Greendlinger, L. and Greendlinger, M.D. (1984) On three of Lyndon's
results about maps, in Contributions to Group Theory (op. cit.), 212-213
Greendlinger, M.D. (1960) Dehn's algorithm for the word problem,
Comm. Pure Appl. Math. (13), 67-83
Griffiths, H.B. (1954) The fundamental group of two spaces with a
common point, Quart. J. Math. (4), 175-190
Griffiths, H.B. (1956) Infinite products of semigroups and local
connectivity, Proc. London Math. Soc. (third series, 6), 455-485
Grushko, I.A. (1940) Uber die Basen einen freien Produktes von Grupper'
Sbornik (8), 169-182
Hall, M. Jr. (1949a) Coset representation in free groups, Trans.
Amer. Math. Soc. (67), 421-432
Hall, M. Jr. (1949b) Subgroups of finite index in free groups,
Canadian J. Math. (1), 187-190
Hall, M. Jr. and Rado, T. (1948) On Schreier systems in free groups,
Trans. Amer. Math. Soc. (64), 386-408
Herriich, F. (1988) Graphs of groups with isomorphic fundamental
groups, Arch. Math. (51), 232-237
Higgins, P.J. (1964) Presentations of groupoids, with applications to
groups, Proc. Camb. Phil. Soc. (60), 7-20
Higgins, P.J. (1966) Grushko's theorem, J. Alg. (4), 365-372
Higgins, P.J. and Lyndon, R.C. (1974) Equivalence of elements
under automorphisms of a free group, J. London Math. Soc. (Second series, 8),
254-258

Higman, G. (1951) A finitely generated infinite simple group, J.

London Math. Soc. (26), 61-64
Higman, G. (1961) Subgroups of finitely presented groups, Proc.
Royal Soc. London, Series A (262), 455-475
Higman, G., Neumann, B.H., and Neumann, H. (1949) Embedding
theorems for groups, J. London Math. Soc. (26), 247-254
Hoare, A.H.M. (1976) On length functions and Nielsen methods in
free groups, J. London Math. Soc. (second series, 14), 188-192
Hoare, A.H.M. (1979) Coinitial graphs and Whitehead
automorphisms, Canadian J. Math. (31), 112-123
Hoare, A.H.M. (1981) Nielsen methods in groups with a length
function, Math. Scand. (48), 153-164
Hoare, A.H.M. (1984) Length functions and algorithms in free
groups, in Contributions to Group Theory (op. cit.), 247-264
Holt, D.F. (1981) Uncountable locally finite groups have one end,
Bull. London Math. Soc. (13), 557-560
Horadam, K.J. (1981) The word problem and related results for graph
product group, Proc. Amer. Math. Soc. (820, 157-164
Howie, J. (1981) On pairs of 2-complexes and systems of equations
over groups, J. Reine Angew. Math. (324), 165-174
Howson, A. G. (1954) On the intersection of finitely generated free
groups, J. London Math. Soc. (29), 428-434
Imrich, W. (1977) Subgroup theorems and graphs, in Combinatorial
Mathematics V (Lecture Notes in Mathematics 622, Springer, Berlin)
Juhasz, A. (1986) Small cancellation theory with a weakened small
cancellation condition: (I) the basic theory, Israel J. Math. (55), 65-93
Juhasz, A. (1987a) Small cancellation theory with a weakened small
cancellation condition: (II) the word problem. Israel J. Math. (58), 19-36
Juhasz, A. (1987b) Small cancellation theory with a weakened small
cancellation condition: (III) the conjugacy problem, Israel J. Math. (58), 37-53
Karrass, A. and Solitar, D. (1970) The subgroups of a free product
of two groups with an amalgamated subgroup, Trans. Amer. Math. Soc. (149),
227-255
Karrass, A. and Solitar, D. (1971) Subgroups of HNN groups and
groups with one defining relation, Canadian J. Math. (23), 527-543
Karrass, A., Pietrowski, A., and Solitar, D. (1973) Finite and infinite
cyclic extensions of free groups, J. Austral. Math. Soc. (16),

Karrass, A., Pietrowski, A., and Solitar, D. (1974) An improved

subgroup theorem for HNN groups with some applications, Canadian J. Math.
(26), 214-224
Kurosh, A.G. (1934) Die Untergruppen der freien Produkte von
beliebigen Gruppen, Math. Ann. (109), 647-660
Lyndon, R.C. (1963) Length functions in groups, Math. Scand. (12),
209-234
Lyndon, R.C. (1965) Grushko's Theorem, Proc. Amer. Math. Soc. (16),
822-826
MacLane, S. (1958) A proof of the subgroup theorem for free
products, Mathematika (5), 13-19
Magnus, W. (1930) Uber diskontinuierliche Gruppen mit einer
definierenden Relation (Der Freiheitssatz), J. Reine Angew. Math. (163), 141-165
Magnus, W. (1932) Das Identitatsproblem fur Gruppen mit einer
definierenden Relation, Math. Ann. (106), 295-307
McCool, J. (1974) A presentation for the automorphism group of a
free group, J. London Math. Soc. (second series, 8) 259-266
McCool, J. (1975a) On Nielsen's presentation of the automorphism
group of a free group, J. London Math. Soc. (second series, 10), 265-270
McCool, J. (1975b) Some finitely generated subgroups of the
automorphism group of a free group, J. Algebra (35), 205-213
McCool, J. and Schupp, P.E. (1973) On one relator groups and HNN
extensions, J. Austral. Math. Soc. (16), 249-256
Miller, C.F. Ill and Schupp, P.E. (1973) The geometry of HNN
extensions, Comm. Pure Appl. Math. 26 (787-802)
Morgan, J.W. and Morrison, I. (1986) A van Kampen theorem for
weak joins, Proc. London Math. Soc. (third series, 53), 562-576
Neumann, B.H. (1943) On the number of generators of a free
product, J. London Math. Soc. (18), 12-20
Newman, M.H.A. (1942) On theories with a combinatorial definition
of equivalence, Ann. of Math. (43), 223-243
Nickolas, P. (1985) Intersections of finitely generated free groups,
Bull. Austral. Math. Soc. (31), 339-348
Nielsen, J. (1921) Om Regning med ikke kommutative Faktorer og
dens Anvendlese i Gruppeteorien, Mat. Tidskrift B , 77-94
Novikov, P.S. (1955) On the algorithmic unsolvability of the word
problem in group theory, Trudy Mat. Inst. Steklov 44 (translated in American
Math. Soc. Translations (9) )

Perraud, J. (1978) Sur les conditions de petite simplification qui

permettent d'utiliser l'algorithme de Dehn, Math. Zeit. (163), 133-143
Petresco, J. (1962) Algorithmes de decision et de construction dans
les groupes libres, Math. Zeit. (79), 32-43
Rabin, M.O. (1958) Recursive unsolvability of group theoretic
problems,, Ann. of Math. (67), 172-194
Rabin, M.O. (1960) Computable algebra, general theory and theory
of computable fields, Trans. Amer. Math. Soc. (95), 341-360
Schupp. P.E. (1971) Small cancellation theory over free products
with amalgamations, Math. Ann. (193), 255-264
Schupp, P.E. (1984) Some aspects of Lyndon's work on group theory,
in Contributions to Group Theory (op. cit.), 33-43
Scott, G.P. (1974) An embedding theorem for groups with a free
subgroup of finite index, Bull. London Math. Soc. (6), 304-306
Scott, G.P. and Wall, C.T.C. (1979) Topological methods in group
theory, in Homological Group Theory (London Mathematical Society Lecture
Notes 36, Cambridge University Press, Cambridge)
Seifert, H. (1931) Konstruktion dreidimensionale geschlossener
Raume, Ber. Sachs. Akad. Wiss. (83), 26-66
Stallings, J. R. (1965) A topological proof of Grushko's theorem on
free products, Math. Zeit. (90), 1-8
Stallings, J.R. (1968) On torsion-free groups with infinitely many
ends, Ann. of Math. (88) 312-334
Stallings, J.R. (1987) Foldings of Otrees (preprint)
Swan, R.G. (1969) Groups of cohomological dimension one, J. Alg.
(12) 585-610
Stilwell, J. (1982) The word problem and the isomorphism problem
for groups, Bull. Amer. Math. Soc. (new series, 6), 33-56
Tretkoff, M. (1973a) A topological proof of the residual finiteness of
certain amalgamated free products, Comm. Pure Appl. Math. (26), 855-859
Tretkoff, M. (1973b) The residual finiteness of certain amalgamated
free products, Math. Zeit. (132), 179-182
Tretkoff, M. (1975) Covering space proofs in combinatorial group
theory, Comm. Alg. (3), 429-457
Tretkoff, M. (1980) A topological approach to the theory of groups
acting on trees, J. Pure Appl. Alg. (16), 323-333
van der Waerden, B.L. (1948) Free products of groups, Amer. J.
Math. (70), 527-528

van Kampen, E.R. (1933) On the connection between the

fundamental groups of some related spaces, Amer. J. Math. (55), 261-267
Wagner, D.H. (1957) On free products of groups, Trans. Amer. Math.
Soc. (84), 352-378
Weir, A.J. (1956) The Reidemeister-Schreier and Kurosh subgroup
theorems, Mathematika (3), 47-55
Whitehead, J.H.C. (1936a) On certain sets of elements in a free
group, Proc. London Math. Soc. (second series, 41), 48-56
Whitehead, J.H.C. (1936b) On equivalent sets of elements in a free
group, Ann. of Math. (37), 782-800

INDEX
Action of a group 182 Cell, indexed 141
Action with (without) inversions 183 Cell subdivision 146
Action, bounded 215 Church's thesis 244
Action, free 185 Circle 113
Action, properly discontinuous 165 Circuit 113
Adjoining a space 52 Coalescence 28
Algorithm 245 Cohomological dimension one 242
Algorithm, partial 245 Complex 121
Almost Stability Theorem 238 Complex, powerless 139
Amalgamated free product 30 Complex, simplicial 139
Amalgamated subgroup 30 Complex with simple boundaries 139
Arrow 64 Component 65, 114, 122
Associated subgroups 35 Computable group 265
Attaching a space 52 Cone 52, 80
Automatic group 288 Conjugacy problem 254
Automorphisms of free groups 132 Conjugate isomorphic 202
Automorphisms, Whitehead 289 Connected 65, 114, 122
Consequence 18
Barycentric subdivision 147 Contractible, locally 154
Base group 35 Contractible space 79
Base-point 74, 123 Contracting of trees 119
Basic neighbourhood 154 Contraction 79
Basis 13 Convex set 60
Bass-Serre theory 182 Core 221
Borsuk-Ulam theorem 104 Covering 96
Boundary 121 Covering belonging to a subgroup 153
Bounded action 215 Covering complex 141-3
Bounded set 215 Covering graph 140
Britton's Lemma 37, 46, 47 Covering map 96, 140-3
Brouwer's Fixed Point Theorem 87 Covering, regular 164
Bunch of circles 92 Covering space 96
Covering transformation 164
Cancellation 28 Covering, universal 162
Cancelling a subset 90 Cyclic loop
Canonical reduction 5, 8, 252 Cyclic permutation 16
Cayiey graph 182 Cyclically irreducible loop 220
Cell 121 Cyclically reduced 15, 28

Downloaded from https://www.cambridge.org/core. Tufts Univ, on 11 Jul 2018 at 01:30:13, subject to the Cambridge Core terms of use, available at
Cambridge
https://www.cambridge.org/core/terms. Books Online © Cambridge University Press, 2010
https://doi.org/10.1017/CBO9780511565878
307

Defining relation 18 Fold 223

Defining relator 18 Forest 114
Deformation retract 88 Forest, maximal 116
Deformation retract, strong 88 Fox comb 81
Degree 101 Free action 185
Diameter 215 Free groups 1, 2, 13
Diamond Lemma 6, 8 Free homotopy 76
Dihedral group 20 Free monoid 3
Direct limit 71 Free on X 2
Direct product of groupoids 66 Free product 24
Direct system 71 Free product, amalgamated 30
Directed set 70 Freiheitssatz 282
Discrete groupoid 64 Full subcomplex 122
Disjoint union 50 Full subgraph 114
Disk 52 Full subgroupoid 65
Fundamental group 74, 123, 199
Edge group 198 Fundamental groupoid 74, 123
Edge 64, 113, 121 Fundamental Theorem of Algebra 103
Edge subdivision 146
Edge-free graph 223 G-graph 182
Edge-groupoid 123 G-set 182
Edge-homotopic 123 Generating symbols 17
Edge-loop 114 Generators 17
Edge-path 114 Geometrical realsiation 149
Elementary neighbourhood 96 Glueing lemma 49
Elementary reduction of a path 123 Godel numbering 250
Elementary reduction of a word 4, 28 Graph 113
End of a path 54, 113 Graph, connected 114
End of an edge 113 Graph of groups 198
Endomorphism of a free group 134, 135 Graph of groups, generalised 198
Ends of a group 242 Group, fundamental 74, 123, 199
Exponent-sum 252 Groupoid 62
Groupoid, connected 65
Final vertex 64 Groupoid, discrete 64
Finite presentation 18 Groupoid, fundamental 74
Finitely generated intersection Groupoid, tree 64
property 219 Grushko's theorem 227
Finitely presented group 18

H0(M) 266 Klein bottle 93

Hall group 229 Kurosh subgroup theorem 175, 212, 293
Ham Sandwich Theorem 104 Kurosh system 178
Handle 95
Hawaian earring 107 Labelling 133, 168
Higman's Embedding Theorem 274 Lebesgue number 53
HNN extension 34, 35 Leftmost reduction 252
Homomorphism of coverings 160 Length 4, 27, 33, 42
Homomorphism of groupoids 65 Letter 4
Homotopy (in complexes) 123 Lies above 151
Homotopy class 57 Lift of a map 151, 158
Homotopy equivalence 78 Lift of a path 98
Homotopy inverse 78 Listable set 246
Homotopy of loops, free 76 Locally bijective 184
Homotopy of maps 77 Locally contractible 154
Homotopy of paths, standard 55 Locally injective 184
Homotopy of paths, variable-length 55 Locally path-connected 100
Hopfian 12 Locally surjective 184
Howson property 219 Loop 54, 113
Hyperbolic group 288 Loop, cyclic 121
Loop, cyclically irreducible 220
Identification space 51 Loop, simple 113
Identification topology 51
Identity for partial multiplication 62 M. Hall group 229
Indexed cell 141 Machine, modular 265, 279
Initial vertex 64 Machine, Turing 278
Inverse of a cyclic loop 121 Map 74, 112
Inverse of a path 58, 114 Maximal forest 116
Inverse of an edge 113 Maximal tree 117
Inversions 183 Membership problem 262
Irreducible path 114 Modular machine 265, 279
Isomorphism of complexes 122 Morphism 122
Isomorphism of coverings 160 Multiplication, partial 62
Isomorphism of graphs of groups 202
Isoperimetric inequality 287 Negative for an edge 217
Neighbourhood, basic 154
Join 50 Neighbourhood, elementary 96
Join, weak 107 Nielsen reduced 188, 253

Nielsen transformation 190 Quotient graph 183

Nielsen's theorem 189 Quotient topology 51
No Retraction theorem 87
Normal Form theorem 5, 26, 31, 36 Rank 13
Number of sheets 142 Real projective space 52
Realisation, geometric 149
Orbit, reversing 217 Recursive set 245
Order of / at Z.Q 106 Recursively enumerable (r.e.) set 245
Orientation 132, 185 Reduced as written 15, 27
Orientation inwards 132 Reduced form 33, 46
Outdegree 132 Reduced Form theorem 32, 37
Reduced set, Nielsen 188
Partial algorithm 245 Reduced set, Nielsen 253
Partial function 244 Reduced set, weakly 194
Partial multiplication 62 Reduced word 4, 27
Partial recursive function 243 Reduced word, cyclically 15, 28
Path 54, 113 Reduced, weakly 254
Path, irreducible 114 Reducible path 114
Path, reducible 114 Reduction (of a word) 4
Path, simple 113 Reduction (of path), elementary 123
Path, standard 54 Reduction (of word), elementary 4, 28
Path, trivial 54 Reduction, canonical 5, 8
Path-component 60 Reduction, leftmost 8
Path-connected space 60 Regular covering 164
Path-connected, locally 100 Reidemeister- Schreier theorem 174
Piece 286 Relations 18
Pinch 37 Relations, defining 18
Points away from 216 Relators 18
Points towards 216, 239 Relators, defining 18
Powerless complex 139 Relators of the first kind 126
Presentation 18 Relators of the second kind 126
Presentation, finite 18 Representative in the large 178
Presentation using a maximal tree 127 Representative tree 185
Product of paths 57 Residually finite 11, 27, 28, 42, 47
Properly discontinuous action 165 Retract 87
Pseudo-HNN extension 35 Retract, deformation 88
Pushout 29, 66, 67 Retraction 87
Reversing orbit 217
Rouche's theorem 103

Schreier's subgroup theorem 170, 212 7^ topology 50

Schreier's subgroup theorem, weak form 167Tietze transformation 21, 22
Schreier transversal 169, 173, 187 Torus 93
Segment 3 Total function 244
Semi-locally simply connected 154 Translation of an edge 213
Sheets, number of 142 Tree 114
Simple group 44, 45, 259, 276 Tree, maximal 117
Simple path 113 Tree, representative 185
Simplicial complex 139 Tree groupoid 64
Simply connected 79 Tree product 201
Skeleton 121 Treed HNN group 201
Small cancellation theory 286 Trivial path 54
Sphere 52 Turing machine 278
SQ-universal 45
Stabiliser 205 Universal cover (graph of groups) 205
Stable letter 35 Universal covering 162
Standard path 54 Unwrapping the circle 97
Standardisation 54
Star-shaped set 79 Van der Waerden 7
Start (of a path) 54, 113 Van Kampen's theorem 83, 124
Start (of an edge) 113 Variable-length homotopy.55
Starting vertex 141 Vertex group 65, 198
Strong deformation retract 88 Vertex, starting 141
Structure Theorem, First 208 Vertices 64, 113, 121
Structure Theorem, Second 210 Von Dyck's theorem 19
Subcomplex 121
Subcomplex, full 122 Weak join 107
Subdivision 146 Weak topology 50
Subdivision, barycentric 147 Weakly reduced 194
Subdivision, cell 146 Weakly reduced set 254
Subdivision, edge 146 Whitehead automorphisms 289
Subgraph 114 Winding number 103
Subgraph, full 114 Word 4
Subgroupoid 64 Word problem 254
Subgroupoid, full 65
Sufficiently large subgroup 153

Topology & Group Theory Basics
No ratings yet
Topology & Group Theory Basics
74 pages
TG 050908
No ratings yet
TG 050908
74 pages
Topology
No ratings yet
Topology
74 pages
Algebra Seven Combinatorial Group Theory. Applications To Geometry (Parshin A. N. (Ed), Shafarevich I. R. (Ed) )
No ratings yet
Algebra Seven Combinatorial Group Theory. Applications To Geometry (Parshin A. N. (Ed), Shafarevich I. R. (Ed) )
124 pages
(London Mathematical Society Student Texts) John Meier-Groups, Graphs and Trees - An Introduction To The Geometry of Infinite Groups-Cambridge University Press (2008)
80% (5)
(London Mathematical Society Student Texts) John Meier-Groups, Graphs and Trees - An Introduction To The Geometry of Infinite Groups-Cambridge University Press (2008)
245 pages
Groups Graphs and Trees An Introduction To The Geometry of Infinite Groups
100% (1)
Groups Graphs and Trees An Introduction To The Geometry of Infinite Groups
245 pages
Diecktop PDF
100% (2)
Diecktop PDF
580 pages
Algebraic Topology EMS Textbooks in Mathematics
100% (1)
Algebraic Topology EMS Textbooks in Mathematics
580 pages
Notes On Group Theory: Mark Reeder
No ratings yet
Notes On Group Theory: Mark Reeder
113 pages
Groups PDF
No ratings yet
Groups PDF
126 pages
Math 806 Notes On Group Theory: Mark Reeder
No ratings yet
Math 806 Notes On Group Theory: Mark Reeder
111 pages
Group Theory Sumary
No ratings yet
Group Theory Sumary
3 pages
חבורות קונבינטריות
No ratings yet
חבורות קונבינטריות
99 pages
Group Theory for Undergrads
No ratings yet
Group Theory for Undergrads
286 pages
Algebraic Topology: Or257/teaching/notes/at PDF
No ratings yet
Algebraic Topology: Or257/teaching/notes/at PDF
80 pages
Shaum Group Theory
100% (2)
Shaum Group Theory
290 pages
Topology and Grupoids
100% (1)
Topology and Grupoids
538 pages
Milne - Group Theory
No ratings yet
Milne - Group Theory
123 pages
Notes On Group Theory: Mark Reeder
No ratings yet
Notes On Group Theory: Mark Reeder
127 pages
Geometric Group Theory Notes
No ratings yet
Geometric Group Theory Notes
58 pages
Master04 22 2025
No ratings yet
Master04 22 2025
144 pages
Group Theory
100% (2)
Group Theory
131 pages
Group Theory - J.S. Milne PDF
No ratings yet
Group Theory - J.S. Milne PDF
131 pages
Groups Def
No ratings yet
Groups Def
16 pages
GT
No ratings yet
GT
135 pages
On The Geometry and Topology of Discrete Groups
No ratings yet
On The Geometry and Topology of Discrete Groups
10 pages
Probability and Geometry On Groups Lecture Notes For A Graduate Course
No ratings yet
Probability and Geometry On Groups Lecture Notes For A Graduate Course
209 pages
(London Mathematical Society Student Texts 88) Derek F. Holt, Sarah Rees, Claas E. Röver-Groups, Languages and Automata-Cambridge University Press (2017)
No ratings yet
(London Mathematical Society Student Texts 88) Derek F. Holt, Sarah Rees, Claas E. Röver-Groups, Languages and Automata-Cambridge University Press (2017)
307 pages
Q. HO-KIM - Group Theory: A Physicist's Primer
100% (3)
Q. HO-KIM - Group Theory: A Physicist's Primer
320 pages
A. G. Kurosh - Theory of Groups, Volume Two-Chelsea Pub. Co. (1960) PDF
100% (2)
A. G. Kurosh - Theory of Groups, Volume Two-Chelsea Pub. Co. (1960) PDF
309 pages
Rial Group Theory 2nd Ed TQW - Darksiderg
100% (5)
Rial Group Theory 2nd Ed TQW - Darksiderg
456 pages
Groups, Rings, and Fields
No ratings yet
Groups, Rings, and Fields
137 pages
Semester Project Report - Topics in Finite Group Theory: December 2015
No ratings yet
Semester Project Report - Topics in Finite Group Theory: December 2015
31 pages
Part IA - Groups: Based On Lectures by J. Goedecke
No ratings yet
Part IA - Groups: Based On Lectures by J. Goedecke
56 pages
GTe 6
No ratings yet
GTe 6
270 pages
Group Theory: Higher Math Handout
No ratings yet
Group Theory: Higher Math Handout
31 pages
Workbook in Higher Algebra
No ratings yet
Workbook in Higher Algebra
194 pages
Milne Group Theory
No ratings yet
Milne Group Theory
135 pages
MA3K4 Complete Notes
No ratings yet
MA3K4 Complete Notes
79 pages
Algebraic Groups
No ratings yet
Algebraic Groups
422 pages
Introduction to Homology Theory
No ratings yet
Introduction to Homology Theory
10 pages
Abs Alg
No ratings yet
Abs Alg
401 pages
Finite Group Theory Project Report
No ratings yet
Finite Group Theory Project Report
31 pages
Symm 1-60
No ratings yet
Symm 1-60
60 pages
Expansion in Finite Simple Groups of Lie Type Terence Tao PDF Version
100% (2)
Expansion in Finite Simple Groups of Lie Type Terence Tao PDF Version
139 pages
Spice and Wolf Vol 4 Isuna Hasekura Jyuu Ayakura Instant Download
No ratings yet
Spice and Wolf Vol 4 Isuna Hasekura Jyuu Ayakura Instant Download
82 pages
Stuckenbruck AngelsGiantsGenesis 2000
No ratings yet
Stuckenbruck AngelsGiantsGenesis 2000
25 pages
Canonical Authors: That Exposed The Ills of The Society.)
No ratings yet
Canonical Authors: That Exposed The Ills of The Society.)
3 pages
The Breathing Permit of Hor
No ratings yet
The Breathing Permit of Hor
26 pages
Lesson Slides Willy Wonka and The Writing Factory
No ratings yet
Lesson Slides Willy Wonka and The Writing Factory
40 pages
Ecryme Starter Kit
100% (1)
Ecryme Starter Kit
32 pages
Grade 5 Unit 7 Condensed Revised 6 20 14test - Prep
No ratings yet
Grade 5 Unit 7 Condensed Revised 6 20 14test - Prep
62 pages
Extraordinary Lives-Anne Frank Reading Comprehension Text
No ratings yet
Extraordinary Lives-Anne Frank Reading Comprehension Text
4 pages
Heart of Darkness by Joseph Conrad
No ratings yet
Heart of Darkness by Joseph Conrad
12 pages
4 Logan (Samantha Whiskey)
No ratings yet
4 Logan (Samantha Whiskey)
163 pages
Tight Corners
No ratings yet
Tight Corners
2 pages
Group 1 Answer Sheet
No ratings yet
Group 1 Answer Sheet
42 pages
SpeedPro Aqua FC6729-FC6728 Guide
No ratings yet
SpeedPro Aqua FC6729-FC6728 Guide
20 pages
Vintage Furniture Invoice Details
No ratings yet
Vintage Furniture Invoice Details
3 pages
Victorian Duality and Secrets
No ratings yet
Victorian Duality and Secrets
2 pages
FRP Guidelines 2020-21
No ratings yet
FRP Guidelines 2020-21
13 pages
Guy de Maupassant
No ratings yet
Guy de Maupassant
8 pages
Three Questions - Leo Tolstoy
0% (1)
Three Questions - Leo Tolstoy
2 pages
Module in Contemporary Popular and Emergent Literature
No ratings yet
Module in Contemporary Popular and Emergent Literature
176 pages
Let Us Review In: English!
No ratings yet
Let Us Review In: English!
17 pages
The Journey (Rida)
No ratings yet
The Journey (Rida)
4 pages
Instant Access To 103 Hikes in Southwestern British Columbia 6th Ed Edition Bryceland Ebook Full Chapters
100% (18)
Instant Access To 103 Hikes in Southwestern British Columbia 6th Ed Edition Bryceland Ebook Full Chapters
85 pages
Creative Haven 13 - Entangled Landscapes Coloring Books
100% (7)
Creative Haven 13 - Entangled Landscapes Coloring Books
68 pages
A Woman Cracked by Multiple Migrations: Search For Identity in Meena Alexander's Fault Lines
No ratings yet
A Woman Cracked by Multiple Migrations: Search For Identity in Meena Alexander's Fault Lines
4 pages
TV Session 1 - Big Train (S1 E1)
No ratings yet
TV Session 1 - Big Train (S1 E1)
3 pages
Love Pascal Francine Download
No ratings yet
Love Pascal Francine Download
38 pages
Wholeissue 48 7
No ratings yet
Wholeissue 48 7
75 pages
12 - 17 - Act III
No ratings yet
12 - 17 - Act III
8 pages
Chapter Four
No ratings yet
Chapter Four
11 pages
PTP IELTS 3 Audioscript Test 6
No ratings yet
PTP IELTS 3 Audioscript Test 6
8 pages