N. D. Gilbert
Department of Mathematical Sciences,
University of Durham.
T. Porter
School of Mathematics,
University College of North Wales, Bangor
OXFORD NEW YORK TOKYO
OXFORD UNIVERSITY PRESS
1994
Oxford University Press, Walton Street, O:iford OX2 6DP
Oxford New York
Athens Auckland Bangkok Bombay
Calcutta Cap Town Dar es Salamm Delhi
Florence Hong Kong Istanbul Karachi
Kuala Lumpur 111adras Madrid Melbourne
Mexico City Nairobi Paris Singapore
Taipei Tokyo Toronto
and associated companies in
Berlin Ibadan
Oxford is a trade mark of Oxford University Press
Published in the United States
hy Oxford University Press Inc., New York
e s. D. Gilbert and T. Porter, 1994
All rights resened. No part of this publication may be
reproduced, stored in a retrieval system, or transmitted, in any
form or by any means, without the prior permission in writing of Oxford
University Press. Within the UK, exceptions are allowed in respect of any
fair dealing for the purpose of research or private study, or criticism or
review, as permitted under the Copyright, Designs and Patents Act, 1988, or
in the case of reprographic reproduction in accordance with the terms of
licences issued by the Copyright Licensing Agency. Enquiries concerning
reproduction outside those terms and in other countries should be sent to
the Rights Department, Oxford University Press, at the address above.
This book is sold subject to the condition that it shall not,
by way of trade or otherwise, be lent, resold, hired out, or otherwise
circulated without the publisher's prior consent in any form of binding
or cover other than that in which it is published and without a similar
condition including this condition being imposed on the subsequent purchaser.
A catalogue record for this book is available from the British Library
Library of Congress Cataloging in Publication Data
Gilbert, N.D.
Knots and surfaces I N.D. Gilbert, T. Porter.
1. Knot theory. 2. Surfaces. I. Porter, T. (Timothy), 1947
II. Title.
QA612.2.G55 1994 5J4'.224dc20 945209
ISBN 0 19 853397 7
Typeset by Pure Tech Corporation, Pondicherry, India
Printed in Great Britain by St Edmundsbury Press,
Bury St Edmunds, Suffolk
Preface
The fascination exerted by interlaced patterns and knotted forms is
evident in their use in decorative and symbolic art across the
centuries. Knotwork is a distinguishing feature of Celtic art, and the
intricate beauty of decorated stonework and illuminated manuscripts
such as the Lindisfarne Gospels, the Book of Durrow, and the Book
of Kells manifests the technical and creative mastery over knotted
forms achieved by Celtic artists before AD 800.
The mathematical theory of knots has a comparatively short
history, properly beginning with some remarks of C.F. Gauss and
being advanced by Victorian pioneers inspired by Lord Kelvin's
vortex theory of atomic structure. Whilst the vortex theory fell by
the wayside, knot theory became a flourishing branch of pure
mathematics. Recent developments have revealed unexpected con
nections with the theoretical physics of quantum field theory, re
uniting knot theory with physics. One of the aims of this book is to
provide the background material so as to make accessible the ideas
and discoveries of this challenging and expanding area of mathemat
ical research.
Our main theme is of course the mathematical theory of knots, and
especially its interaction with, and enrichment by, the theory of
surfaces and of group presentations. We begin with a study of knots
based on pictures of interlacing, and discuss two complementary
problems. Firstly, what is the precise nature of the interlacing that
we wish to describe? More precisely, when should different pictures
be regarded as showing the same interlacing? Secondly, how can we
classify the different interlaced forms represented by the pictures?
Indeed, the fundamental problem of classifying knots, of organizing
them according to their properties and characteristics, is a guiding
theme of the entire book. There is a wealth of raw material for this
process: the knots themselves! A few will be familiar from the
practical problems of tying shoelaces, parcel string, mooring and
rigging lines, safety ropes, and so on. Most interlacing patterns do
not produce knots of practical use, but their decorative appeal
remains, and their mathematical interest does not depend on their
practical value. We have plenty of knots to work with: for example,
ce
it is a fact that there are precisely 12965 different knots that can be
represented by interlacing patterns in which the string crosses itself
at most 13 times.
The pictures of interlacing represent spatial relationships in three
dimensional space, and to pursue this aspect of knot theory we use
topology, the study of shape and continuity. Topological methods
provide new tools to aid our understanding of knotted forms and
connect the study of knots with the second major subject of the
book: the study and classification of surfaces. The immediate vicinity
of any point of a surface looks just like twodimensional space, and
we want to understand what the possibilities might be for the shape
of the entire surface. Here we have a second topological classification
problem, and one with a complete solution whose essentials were
known in the nineteenth century. The classification theorem is a
masterpiece of pure mathematics in action. A subtle problem is
broken down into intermediate steps, the essential information is
recorded by simple invariants, and the invariants can then be
organized to produce a list of all the possible objects under scrutiny.
Knots, surfaces, and the connections between them lead us natu
rally into other subject areas, principally to graphs and to group
presentations. These subjects have their own motivating questions,
their own deep problems, and their own fascinations and appeal. We
try to draw out the connecting strands, the exchange of ideas,
concepts, and methods. We hope that the reader will become happily
entangled in the web that is woven by the theory of knots and
surfaces.
Durham
Bangor
Oct 1993
N.D.G
T.P.
Acknowledgements
This book is based on a thirdyear undergraduate course given at the
University College of North Wales, Bangor. We are grateful to our
colleagues who have been involved with the course at various times,
and in particular to Ronnie Brown, Peter Cromwell, Hugh Morton,
and Tom Thickstun for discussions, advice, and the provision of
some material. We are especially grateful to the large number of
students that have responded to our enthusiasm for the subject by
electing to take knot theory as an option course at Bangor, and have
criticized, complained, and commended at various points. The first
draft of this book was based on lecture notes of improbable neatness
taken by Ms Caren Williams in the academic year 19878.
The first author carried out most of his share of the preparation
of this book whilst a lecturer in mathematics at HeriotWatt Univer
sity, and thanks his former colleagues for the congenial working
environment that they created. He is also most grateful to Robin
Nicholls for her assistance in preparing the typescript for the pub
lisher.
Contents
1. Knots. links. and diagrams 1
1.1 Knot and link diagrams 1
1.2 Isotopy and the Reidemeister moves 5
1.3 3colouring 8
1.4 Numerical invariants II
1.5 Chiral and invertible knots 15
2. Knot and link polynomials 19
2.1 State models and the Jones polynomial 19
2.2 The Jones polynomial 28
2.3 Applications of the Kauffman polynomial 30
2.4 The oriented or Homfly polynomial 34
2.5 The Alexander polynomial 38
3. Topological spaces 42
3.1 Topological spaces 42
3.2 Continuity 45
3.3 Connectedness 48
3.4 Identification spaces 49
3.5 Knots and isotopy 51
4. Surfaces 55
4.1 Combinatorial surfaces 57
4.2 Cutting and pasting 68
4.3 Invariants of surfaces: orientability
and boundary curves 80
4.4 Invariants of surfaces: Euler characteristic
and genus 84
4.5 Knots and surfaces 90
5. The arithmetic of knots 96
5.1 The sum of oriented knots 96
5.2 Genus of K + L 100
nts
5.3 Bridge number of K + L 103
5.4 Factorizations, prime knots, and the
factorization theorem 105
5.5 ncolourability 107
5.6 Polynomials and the sum of knots 108
6. Presentations of groups 111
6.1 Examples of presentations III
6.2 Dihedral groups 115
6.3 Cayley quivers 118
6.4 Free groups 122
6.5 Presentations 129
6.6 Tietze transformations 131
6.7 Quotient groups 138
6.8 Abelianization 139
6.9 Finitely generated Abelian groups 140
6.10 Pushouts of groups 144
6.11 Knot groups 150
7. Graphs and trees 163
7.1 Graphs, quivers, and trees 163
7.2 Examples of graphs 169
7.3 Planar graphs 171
7.4 Embedding graphs in surfaces 173
8. Alexander matrices and Alexander polynomials 178
8.1 Group rings 178
8.2 Derivatives 182
8.3 The Alexander matrix 184
8.4 Elementary ideals 187
8.5 Alexander polynomials via the
Wirtinger presentation 192
9. The fundamental group 197
9.1 Paths 197
9.2 Homotopy 199
9.3 Calculations of the fundamental group 203
9.4 Homotopy equivalence 208
10. Van Kampen's theorem 212
11. Applications of the Van Kampen theorem
11.1 Calculations of fundamental groups
11.2 Surface groups
11.3 Spaces with finite cyclic fundamental group
11.4 Spaces with given fundamental group
11.5 The group of a knot
11.6 Torus knots
11.7 Analysis of the knot group G
m
, n
12. Covering spaces
12.1 Covering maps and covering spaces
12.2 Deck transformations and universal covers
12.3 Covering graphs
Bibliography
Index
Contents xi
224
224
233
236
237
239
243
246
250
250
258
259
264
266
1
Knots, links, and diagrams
Interlaced or knotted patterns and forms are encountered in many
settings. Their occurrence may be essential, as in the practical
applications of knot tying, or inconvenient, as with a knotted
shoelace or inextricably entangled garden hose. Interlacing motifs
are the basis for much Celtic and Islamic art, and language is
enriched by many figures of speech referring to knots. Our subject is
how mathematics gets into knots and our aim is to knit a coherent
theory of knots from the ravelled strands of our cultural and
practical fascination.
t and link diagrams
Instructions on the skills of knotting are commonly given by means
of pictures of the kind shown in Fig. 1.1.
l
+
II III
Fig. 1.1
Some attempt may be made to render the appearance of the
strands ropelike, as shown in Fig. 1.2, but the significant informa
tion conveyed by the picture is the pattern of interlacing of the
links, and diagrams
Fig. 1.2
strands: a break in a strand indicates that the broken strand passes
under the unbroken one. Such pictures will be the object of study in
this chapter and we shall see to what extent we can develop a pictorial
theory of knottedness, for both knots, in which a single strand is
interlaced, and for links, in which more than one strand is involved.
We call the pictures knot or link diagrams; no formal definition will
be given and none seems necessary at this point. Figure 1.3 contains
some examples, together with the names commonly given to the
knots and links represented.
There are a number of observations to be made, and some
terminology to be established. In tying a knot in a piece of string, we
Trefoil Figureeight
GO
Hopf link Borromean rings
Fig. 1.3
Knot and link diagrams 3
are left with two loose ends. In a knot diagram (Fig. 1.4) these loose
ends are shown spliced together, and this removes any need to treat
the loose ends in a special way. The splicing also fixes the knotted
ness in the diagram; there is no chance of the loose ends moving so
as to undo the knot! To sum up then, a link diagram encodes a way
of interlacing one or more circles in three dimensions.
Fig. 1.4
A continuous segment III a diagram IS an are, and arcs meet at
crossings. The unbroken arc at a crossing is the overpass, and the
remaining two arcs form the underpass. All knot and link diagrams
will have only finitely many crossings, and arcs may only meet at
crossings. The arcs in a diagram are divided into a number of separate
classes, called components, corresponding intuitively to the different
pieces of string used in the physical knot or link being modelled. The
unknot, the trefoil and the figureeight have one component, and so
are knots, whilst the Hopf link has two components, and the Borro
mean rings have three. By definition, a knot diagram has exactly one
component. A link diagram may have any finite number of compon
ents, and we shall use link as the general term, using knot only when
we wish to specialize to the case of one component. An orientation
of a link diagram is a choice of direction on each component,
indicated on the diagram by arrows as shown in Fig. 1.5.
In an oriented diagram (Fig. 1.6) we distinguish two types of
crossing and every crossing is of one type or the other. One method
of remembering the allocation of signs is to imagine an approach to
the crossing along the underpass in the direction of the orientation:
if the overpass orientation runs from left to right, the crossing is
positive, and if from right to left then the crossing is negative. The
writhe of an oriented diagram is the sum of the signs of all its
crossings: we denote the writhe of a diagram D by weD).
Knots, links, and diagrams
mples
00
Fig. 1.5
/ /
/
/
Positive crossing Negative crossing
Fig. 1.6
A variation of the writhe is useful for the study of links. Given an
oriented link diagram D, with components CJ,"" C
m
define the
linking number of C
i
with 0, where C
i
and 0 are distinct components
of D, to be onehalf the sum of the signs of the crossings of C
i
with
C
j
, denoted Ik(C
b
0). The linking number of D is then the sum of the
linking numbers of all pairs of components:
Ik(D) =
l ~ i ~ j ~ / n
Isotopy and the Reidemeister moves 5
Writhe = +3 Writhe = 0
Writhe = 2 Writhe = 0
Fig. 1.7
opy and the Reidemeister moves
We now introduce the idea of isotopy, an equivalence relation
between diagrams that relates their study more closely to our
intuitive notions of interlacing, and, thanks to a deep theorem of
Reidemeister, also to the study of knotted curves in threedimensional
space. We shall meet Reidemeister's theorem in Chapter 3, in which
we look at the topological approach to knot theory.
The Reidemeister moves are the changes shown in Fig. 1.8 to link
diagrams. It is assumed that the diagram is otherwise unchanged,
and that no other part of the diagram appears at the crossings
changed by the Reidemeister moves. We shall use a similar conven
tion for the manipulation of diagrams in Chapter 2, and only draw
that part of a diagram at which the manipulation occurS. Reidemeis
ter moves first occur in the book by Reidemeister (1932).
We also consider a class of diagram changes, which for conveni
ence we label RO, though they are not Reidemeister moves. Two
diagrams D and D' differ by a move of type RO if there is a
continuous deformation of the plane that carries D to D' without
changing any crossings. Moves of type RO correspond to pulling and
stretching a knot without affecting its interlacing (Fig. 1.9).
links, and diagrams
R1
XJ
R2 Q _)(_Q
R3
I
I 
Fig. 1.8
I
/
Link diagrams D and D' are isotopic if D can be transformed into
D' by some succession of moves RO, RI, R2, and R3, and regularly
isotopic if the transformation can be achieved without use of the
move Rl. It is clear that isotopy is an equivalence relation on link
diagrams, and the same is true of regular isotopy. An equivalence
@
Fig. 1.9
Isotopy and the Reidemeister moves 7
class of diagrams modulo isotopy is an isotopy class, and evidently
diagrams in the same class represent the same essential pattern of
interlacing. An isotopy class of link diagrams will be our model for
an interlaced form in three dimensions; the correctness of this model
is in fact assured by Reidemeister's theorem mentioned already, that
deformations of interlaced circles in threedimensional space can be
mirrored exactly by successions of Reidemeister moves on diagrams.
We shall call isotopy classes of link diagrams by the name custom
arily given to the associated link.
What are the problems that pictorial link theory must face?
Ultimately we might hope for a classification of all possible knots
and links, in the form of a list of isotopy classes and a means of
deciding into which class a given link diagram belongs. A more
fundamental problem is to decide if two given link diagrams are
isotopic. How can we decide if a given link diagram is in fact isotopic
to an mcomponent unlink? What sort of properties should we seek,
in order to establish some means of comparison and differentiation
of diagrams? Plainly, a property useful for such a purpose must be
isotopy invariant: if it holds for a diagram D then it must hold for all
diagrams isotopic to D. We can verify isotopy in variance by examin
ing the effects of the Reidemeister moves (and the plane deforma
tions RO).
The obvious first classification of links is by the number of
components: none of the moves RO, Rt, R2, R3 alters the number of
components of a diagram. Hence the number of components of a
link diagram is an isotopy invariant. We deduce that the Hopf link
is not isotopic to the Borromean rings! ~
To go further, let us consider linking numbers. A Reidemeister
move that involves at least two components of a link (and must
therefore be an example of R2 or R3) leaves all linking numbers of
an oriented diagram unchanged. A move R2 removes or introduces
two crossings of opposite sign, whilst a move R3 leaves the number
of crossings and their signs unaltered. Since any two components
of an unlink have zero linking number, the Hopf link is not iso
topic to the twocomponent unlink, for any assignment of orienta
tions to the components of the Hopf link results in a nonzero
linking number. Check that this is so! However, for the Borromean
rings, the linking number of any two components is zero. Further
more, linking number cannot help us to study the isotopy of knots,
and the writhe of a diagram is not an isotopy invariant. Indeed, we
can alter the writhe at will by moves of type RI. We now go on to
consider some isotopy invariants particularly adapted to the study
of knots.
Knots, links, and diagrams
3colouring
mples
A knot diagram is 3colourable if we can assign colours to its arcs
such that
3Cl: each arc is assigned one colour;
3C2: exactly three colours are used in the assignment;
3C3: at each crossing, either all the arcs have the same colour, or arcs
of all three colours meet.
{ .......
I \
Fig. 1.10
Theorem. 3colourability is an invariant of isotopy type.
Proof. A deformation of type RO changes no crossing, so cannot
change the property of 3colourability. For the Reidemeister moves,
we show that the colouring at any crossing changed by the move can
be amended if necessary so that the resulting diagram is again
3coloured. In Figs 1.11 and 1.12, we show only that part of the knot
diagram changed by the Reidemeister moves.
Fig. 1.11
 ')
/
Fig. 1.12
f
I
\
\
"
,
,
,
3colouring 9
There are many cases to check: try some! You should bear in mind
the following facts: to give a complete proof for the case R3, you
need to be exhaustive; that is, to have checked all the possibilities.
However, permutation of the colours will not affect the problem, but
you are not allowed to change the colours on arcs that leave the
diagram, for such changes might effect 3coloured configurations at
other crossings. •
Since the unknot is not 3colourable (why not?) the 3coloured
diagrams above are not isotopic to the unknot; so knots exist! We
4
**
Fig. 1.13
ots, links, and diagrams
can also show that certain knots are different, that is not isotopic to
one another. For example, the figureeight knot in Fig. 1.13 is not
3colourable.
If arc I were coloured red, and arc 2 red as well, then arc 3 would
have to be red to satisfy 3C3, and arc 4 is then also forced to be red.
Then the whole diagram is red, which contravenes 3C2. If instead we
try arc I red and arc 2 white, then arc 3 must be blue (at *) and arc
4 red (at **), but this forces the other two crossings to contravene
3C2.
es 1.3
I. Which of the knots in Fig. 1.14 are 3colourable?
(a) (b)
(c)
(d)
Fig. 1.14
2. Can 3colourability be used to study links with m ~ 2 compon
ents?
3. The notion of 3colourability generalizes to ncolourability. Let
lL
n
denote the integers modulo n. An ncolouring of a knot
diagram is an assignment to each arc of an element of lL
n
in such
a way that, at each crossing, the sum of the values assigned on
the underpass arcs is twice that on the overpass, and such that
at least two elements of lL
n
are assigned. Is this notion really a
generalization of 3colourability? Investigate the two meanings of
the term 3colourable and see that they do coincide. Find several
examples of 5colourable knots. Prove that ncolourability is an
isotopy invariant.
Numerical invariants 11
1.4 Numerical invariants
Properties of knot diagrams that are defined solely in terms of the
isotopy class to which the diagram belongs will certainly be isotopy
invariant. In considering such properties, remember that we shall call
an isotopy class of diagrams simply a knot, and if we wish to
particularize, to use names of knots such as the trefoil or figureeight
for the isotopy class containing the named diagram. This matches
our intention to use isotopy classes as the models of interlaced form.
Crossing number
Examples
Let K be a knot. The crossing number c(K) of K is the mlllimum
number of crossings in a diagram in the isotopy class K. The crossing
number is therefore the number of crossings in the simplest picture
of a knot. A diagram of a knot K with c(K) crossings is a minimal
diagram.
c(unknot) = 0; c(trefoil) = 3; c(figureeight) = 4.
The lists of knots given in Burde and Zieschang (1985) and
Kauffman (l987a) are arranged according to crossing number: a
minimal diagram is given for each knot with crossing number at most
9. Choices have to be made of one mirror image or the other, a topic
to which we shall return shortly. The crossing number provides a
means of arranging data on knots in a systematic form according to
an intuitively reasonable measure of complexity. However, given
some arbitrary diagram, the crossing number of the knot that it
represents may be hard to determine.
Unknotting number
A crossing change in a diagram exchanges the overpass and under
pass at a crossing as shown in Fig. 1.15.
Crossing changes will alter the isotopy type of a diagram.
xx
Fig. 1.15
Knots, links, and diagrams
Lemma. Let D be a diagram with c crossings. Then changing at most el2
crossings of D produces a diagram of the unknot.
Note. You will probably find it useful, before reading the following
proof, to tryout a few examples of using crossing changes to produce
unknot diagrams. Try to find a proced ure for doing this
systematically, and then compare your procedure with the one
suggested in the proof. You should certainly test the proof on some
examples of your own devising.
Proof. Travel once around the diagram D from any starting point:
at each first encounter with a crossing, note whether you are on the
overpass or the underpass. Count the final tally of overpasses and
underpasses and change all the crossings contributing to the smaller
count: this can involve at most c/2 changes. By construction, at each
first encounter with a crossing in the new diagram, you arrive
consistently on the overpass or on the underpass, so that no
interlacing ever occurs and the new diagram is an unknot. •
The unknotting number u(K)of a knot K is the smallest number of
crossing changes required to obtain the unknot from some diagram
of the knot. The un knotting number is hard to calculate, and is still
unknown for some knots with crossing number 9. The essential
difficulty is illustrated by the following.
Theorem. (Bleiler 1984) The unknotting number of a knot does not neces
sarily occur in a minimal diagram.
Proof. Let K be the knot with the diagram shown in Fig. 1.16.
It can be shown that this diagram is minimal and is unknotted by
three crossing changes. However, no two crossing changes will
unknot it. Now the diagram in Fig. 1.17 is isotopic to the previous
Fig. 1.16
Numerical invariants 13
Fig. 1.17
one. It has 14 crossings, so is not minimal, but is unknotted by two
crossing changes. •
Note. In fact u(K) = 2. See Bleiler's (1984) paper for more details.
Bridge number
Example
Example
A bridge in a knot diagram is an arc that is the overpass in at least
one crossing.
Fig. 1.18
The bridge number b(K) of a knot K is the minimum number of
bridges occurring in a diagram of the knot. By convention the
unknot has bridge number 1.
The standard diagram of the trefoil has three bridges as shown in
Fig. 1.19.
Knots, links, and diagrams
Fig. 1.19
However, the diagram of the trefoil (Fig. 1.20) has only two
bridges, so b(trefoil) ~ 2.
Fig. 1.20
Proposition. A knot K has bridge number 1 if and only if it is the unknot.
Proof. It is our convention that the unknot has bridge number 1.
So suppose that K is a knot with b(K) = I and let D be a diagram of
K with one bridge and the smallest possible number of crossings, m
say. If m ~ 2 then K is the unknot, so we may assume that m ~ 3.
No arc of the diagram, except the bridge, is an overpass at any
crossing. Therefore, there must occur two adjacent crossings at
which the same arc is the underpass, and a Reidemeister move R2 is
applicable, which reduces the number of crossings by two. This
produces either a diagram of the unknot, in which case the original
diagram was the unknot, or else a Ibridge diagram with a number
of crossings smaller than m, which contradicts the choice of m. •
It now follows that the trefoil has bridge number 2 (why?). The
2bridge knots have been classified by H. Schubert: see his original
paper (Schubert 1956) or the account in Burde and Zieschang (1985).
Chiral and invertible knots 15
es 1.4
1. Can you improve the bound on the number of crossing changes
needed to unknot a diagram? Try some examples.
2. Construct an isotopy between the two diagrams in the proof of
Theorem 1.3.
3. Calculate the unknotting number of the knots shown in Fig. 1.21.
I
b
lQ7J
Fig. 1.21
4. What is the unknotting number of the knot with the diagram
shown in Fig. 1.22.
Fig. 1.22
5. Construct an isotopy between the two trefoil diagrams in Fig.
1.19 and Fig. 1.20.
iral and invertible knots
Need the mirror image D! of a knot diagram D be isotopic to D?
Consideration of the trefoil in Fig. 1.23 indicates that the answer is no.
No amount of persuasive manipulation of a string tied as on the
left seems to transform it to being tied as on the right; it appears that
s, links, and diagrams
Fig. 1.23
the diagrams lie in distinct isotopy classes. This is in fact the case,
though we do not have the means to prove it as yet. However, the
following manipulations show that the figureeight and its mirror
image are isotopic (Fig. 1.24).
@®
Fig. 1.24
The effect of a mirror reflection on a single crossing is to perform
a crossing change as shown in Fig. 1.25.
To obtain D! from D we can simply change all the crossings in D,
for a mirror image of D is obtained by changing all the crossings and
then rotating the entire diagram through 180
0
about the line repres
enting the mirror. This last operation does not affect the interlaced
form, so does not change the isotopy class of the diagram (Fig. 1.26).
Chiral and invertible knots 17
x x
Fig. 1.25
A knot that is isotopic to its mirror image is said to be aehiral, so
that the figureeight is achiral. Otherwise, the knot is ehiral and we
shall see in the next chapter that the trefoil is indeed chiral; the
polynomial invariants of Chapter 2 give a ready means of deciding
chirality for many knots.
Fig. 1.26
If D is an oriented diagram, let 15 be the diagram with the
orientation reversed. Are D and 15 necessarily isotopic? Here we
require the isotopy to match up the orientations. An oriented knot
that is isotopic to itself with the orientation reversed is invertible. The
trefoil is invertible (Fig. 1.27).
Fig. 1.27
Noninvertible knots were not known to exist until 1964, when
H.F. Trotter found an infinite family of them (Trotter 1964). One is
shown in Fig. 1.28.
ots, links, and diagrams
Q
~
~
§
~
§
~
§
n
Fig. 1.28
We close this chapter with a definition that will be crucial in the
next. A knot diagram is alternating if you meet crossings alternately
at overpasses and underpasses as you travel around the diagram. Not
every knot possesses an alternating diagram: the first examples that
do not have crossing number 8. See the list in Kauffman (l987a)
es 1.5
1. Find a sequence of Reidemeister moves carrying the figureeight
to its mirror image.
2. Show that the figureeight and its mirror image are regularly
isotopic.
3. Simply flipping over shows that the two diagrams in Fig. 1.27.
are indeed isotopic. Give a sequence of Reidemeister moves that
effects the isotopy.
2
Knot and link polynomials
Polynomial invariants of knots and links have played an important
part in knot theory since the work of J.W. Alexander (1928). The
appearance in 1984 of a new polynomial invariant, found by V.F.R.
Jones, led to an upsurge in research activity and to a number of
striking advances iOn our knowledge of knots and links and their
representation by diagrams, as well as to new interactions between
knot theory, geometry, and quantum physics. In recognition of his
achievements, Jones was awarded a Fields Medal (the equivalent in
mathematics of a Nobel Prize) at the 1990 International Congress of
Mathematicians. Jones discovered his polynomial as an offshoot of
work in a quite different areathe theory of operator algebras. He
tells the story in Jones (1985). Subsequently, Alexander's and Jones'
polynomials were generalized in a twovariable polynomial; this
generalization was accomplished virtually simultaneously by several
groups of mathematicians, using a variety of approaches.
Jones' polynomial may also be derived from some results due to
L.H. Kauffman. Some of this work is phrased in the terminology of
statistical physics, which provides a relevant theoretical parallel; see
Kauffman (1991). The place of Jones' polynomial and its generaliza
tions within the framework of preexisting knot theory remains a
matter of much interest and no little mystery.
We begin this chapter with Kauffman's work and examine his state
model polynomial, and the Jones polynomial that is derived from it.
We then move on to the twovariable generalization, and finally look
at Alexander's original treatment of his polynomial.
e models and the Jones polynomial
The title of this section is also that of Kauffman's paper explaining
his construction (Kauffman 1987b).
Let D and D' be (unoriented) link diagrams. Recall from Chapter 1
the notion of regular isotopy: D and D' are regularly isotopic if they
and link polynomials
differ by a sequence of Reidemeister moves of types R2 and R3; the
move RI that introduces Or removes a loop is not allowed.
We use U to denote a diagram of the unknot with no crossings,
and if D is a link diagram, then we will use D U to denote a diagram
with one extra unknotted component that introduces no extra cross
ings. The bracket polynomial of a diagram D is the Laurent poly
nomial (D) in one variable A defined by means of the following rules:
BPI: (U)=I
BP2: (DU) =  (A2 + A
2
)(D)
BP3: (X> =A (X> +A
1
()C>
In rule BP3 we use the convention from Chapter I of drawing only
that part of a knot diagram that is changed in the intended manip
ulation. BP3 is a shorthand for the following computational proce
dure. Take any diagram D and choose any crossing of D: delete the
crossing and consider the now separated strands (Fig. 2.1). Imagine
travelling along each of the two strands in the remains of the
overpass towards the former crossing point: join these strands to
those from the underpass that lie to your left, to produce a new
diagram L. If instead you join the strands from the overpass to those
from the underpass that lie on your right, you obtain a new diagram
R. The rule BP3 now says that (D) = A(L) + A1(R). The removal of
crossings and the rejoining of strands are not Reidemeister moves; a
little experimentation now, and a wealth of experience to come in the
next few pages, will certainly convince you that the isotopy type of
diagrams will be changed by these manipulations.
x
D L R
Fig. 2.1
Note that BP3 then implies the rule
. /X\ 1)(> l/U\
BPJ'. \ \ I = A \ +A \ f\ )
State models and the Jones polynomial 21
Fig. 2.2
Some care is needed in applying these rules to the computation of
bracket polynomials. It is evident that BPI and BP2 together
determine the bracket polynomial for any diagram of an unlink with
no crossings, whilst BP3 enables the calculation of (D) for any
diagram D by successive reduction of the number of crossings.
Starting with a diagram D we can produce a polynomial (D). To do
this we have to change the diagram by successive reduction of the
number of crossings. For example, with the figureeight knot we
could reduce as shown in Fig. 2.2 or as in Fig. 2.3.
Fig. 2.3
This raises the problem of deciding if the end result might depend
on the route taken. We shall see that (D) does not depend on the
order in which the crossings in D are removed; for the moment, we
assume this fact and go on to derive some of the properties of the
bracket polynomial that underlie its theoretical significance and
facilitate its computation.
rna. The bracket polynomial is an invariant of regular isotopy; that is,
applications of Reidemeister moves R2 and R3 to a diagram leave its
bracket polynomial unchanged.
Proof. For move R2 we proceed as follows, making full use of the
device of showing only that portion of a diagram at which changes
are taking place.
(:s::;?)=A( Q)+A1(Y) (BP3)
(BP3')
Knot and link polynomials
2 2 I ~ ) /} ( ') / ~ )
=(A +A ) \ ..... + \ + \..... . u
(BP2)
Invariance under move R2 now implies invariance under move R3:
\X) = A \'0) +A I \)()
=A( :: ) +A
1
\}()
= \  ~ )
by symmetry.
(BP3)
(R2)
•
Lemma. The bracket polynomial is not an isotopy invariant. Reidemeister
move Rl produces the following changes to bracket polynomials:
Proof. We have
\ p) = A \) 0) +A
1
\ ~ )
= _A(A
2
+A
2
) \ } )+A
1
( } )
= _A
3
\ ) )
The other case is left to the reader.
0 c§J
<c2J> = A<cQ»+A
1
<c€»
= _A
4
< 0 >_A
4
< 0 >
=(A 4+ A
4
)
Fig. 2.4 (a)
(BP3)
(BP2)
•
State models and the Jones polynomial 23
D=
<db> A <£>+A'<c0>
A
7
 A
3
 A
5
8
7 3 5
<0b> A A A
Fig. 2.4 (b)
We now proceed to some i11ustrative computations of bracket
polynomials in Figs 24(a)(c).
<00> =A <cY0> +A
1
<00>
=A(A
7
A
3
_A
5
)_A
4
<6>
=A
8
A
4
A 4_ A
4
(A
4
+A
4
)
=A
8
A
4
+ 1A
4
+ A8
Fig. 2.4 (e)
We now turn to the effect on the computation of a bracket
polynomial of a change in the ordering of the crossings and their
removal. We use Kauffman's idea of a state for a knot diagram.
For each crossing in a diagram D there are two ways to eliminate
it, as we discussed in relation to the rule BP3 above: either as shown
in Fig. 2.5(a) or as shown in Fig. 2.5(b).
Having made such a choice, we attach to each crossing a splitting
marker that connects the two regions of the diagram that will be
and link polynomials
x
Fig. 2.5 (a)
x
)(
Fig. 2.5 (b)
joined by removal of the crossing: either as shown in Fig. 2.6.(a) or
as shown in Fig. 2.6(b).
Fig. 2.6 (a)
Once a splitting marker has been attached to every crossing in D
we have a state of D; thus a diagram with c crossings has a total of
2
c
possible states. If we now remove all the crossings in accordance
with the splitting markers in a given state, we obtain a diagram of
an unlink with the number of components in the unlink depending
on the state. The bracket of the unlink diagram is
determined by BPI and BP2: it is (_(A
2
+ A
2
»k1, where k is the
number of components.
State models and the Jones polynomial 25
Fig. 2.6 (b)
Now consider the computation of (D) for a diagram D. Repeated
use of BP3 results in an expression for (D) as a sum of terms
±Am(U ... U), where U ... U is a diagram of an unlink, with exactly one
such term for each state of D. This expression is uniquely determined
by D, and (D) is therefore independent of the route chosen for its
computation. We can be a little more precise and obtain a formula
for (D) as a sum over its states.
Consider the four regions adjacent to a crossing in D (Fig. 2.7): the
two regions that lie between the overpass and an underpass, in that
order, when circling the crossing anticlockwise, we shall call positive;
the other two regions we sha11 call negative. Observe that in the
earlier discussion of the elimination of crossings, in order to apply
BP3 the diagram L was obtained from D by rejoining the separated
strands so as to connect the two positive regions, and the diagram R
was obtained by connecting the two negative regions.
x
~
/
Fig. 2.7
A splitting marker connects regions of the same sign, and in BP3
the bracket (L) is multiplied by A and the bracket (R) by A I. Let S
be a state of the diagram D and let p(S), n(S) be respectively the
number of splitting markers in S joining positive and negative
regions. Now define (DIS) = AP(s)n(s). We write lSI for the number
of components in the diagram of unlinks obtained from D by
removing all crossings in accordance with the state S: this diagram
Knot and link polynomials
receives coefficient (D I S) in the expression for D. We can summarize
the above remarks in a formula for D.
Proposition. Let D be a link diagram. Then
(D) = L (D I S) [_(A
2
+ A
2
)]I
S
I I
s
where the sum is taken over all states S of the diagram D.
•
The formula just given does not provide a practical method for the
evaluation of (D) but it is of theoretical interest because of its
resemblance to the form of expressions found in statistical physics.
It is a pleasing and surprising nexus between two strands of mathe
matical thought, and an excellent place to seek further information
is Kauffman's book (Kauffman 1991).
At first sightand certainly according to the spirit of Chapter Ithe
bracket polynomial is not much use in the study of knots and links,
since it is not an isotopy invariant. Happily, this deficiency is easily
remedied to produce an isotopy invariant for oriented links.
Let L be an oriented link and let D be an oriented diagram of L
with writhe w(D). Recall from Chapter I that the writhe is not an
isotopy invariant; however, it is an invariant of regular isotopy, as
may easily be checked by examining the effect of Reidemeister moves
R2 and R3. Define the Laurent polynomial f[L] by
f[L] = ( A) 3 w(D) (D)
where (D) is the bracket polynomial of the diagram obtained from D
by simply forgetting its orientation. The notation already suggests
that f[L] is determined by L; that is, it is an isotopy invariant.
Theorem. The polynomial f[L] is an isotopy invariant.
Proof. Since w( D) and (D) are invariants of regular isotopy, so is
f[L]. Therefore, we need only to check the effect of Reidemeister
move RI. One application of this move either introduces or removes
a crossing and so changes w(D) to w(D) ± 1; the term ( A)3W(L)in
f[L] contributes an additional factor  A± 3 which is exactly cancelled
by the factor contributed by (D) and given by Lemma 2.2. •
Combining our calculations of bracket polynomials and writhes,
we have the following computations of Kauffman's polynomial f[L]:
State models and the Jones polynomial 27
(1) L = lefthand Hopf link (Fig. 2.8)
f[L] = _A
6
(A
4
+A
4
)
= _A
10
+ A2
Fig. 2.8
(2) L = lefthand trefoil (Fig. 2.9)
f[L]
Fig. 2.9
(3) L = righthand trefoil (Fig. 2.10)
f[L]
Fig. 2.10
(4) L = figureeight knot (Fig. 2.11)
f[L] =<D>
= Af!.. A4 + 1 A
4
+ Aa
Fig. 2.11
and link polynomials
Observe that the Kauffman polynomial for the righthand trefoil
is obtained from that for the lefthand trefoil by interchanging A and
AI. This observation prompts the question: is it always the case that
the Kauffman polynomial of a mirror image is obtained by inter
changing A and A I? If the answer is yes then achiral knots must
have Kauffman polynomials that are symmetric in A and AI, and
this is the case for the figureeight knot. It is therefore reasonable to
propose, as our first formal property of the Kauffman polynomial,
the following result.
osition. Let L be a link diagram and let L'be its mirror image. Then
f[L!] is obtained from f[L] by interchanging A and AI. Moreover,
(L!) is also obtained from (L) in this way.
Proof. We know from Chapter I that a diagram of L! is obtained
from one of L by reversing all the crossings, and a computation of
(L') from this diagram proceeds exactly as for (L) except that the
roles of A and AI are interchanged. Since it is also the case that
w(L!) =  w(L), the result now follows. •
2.1
1. Show that the two diagrams in Fig. 2. P are regularly isotopic.
Fig. 2.12
2. Show that the two diagrams in Fig. 2.13 are not regularly isotopic.
3. Compute (D) directly from the formula in Proposition 2.3 for the
diagrams D in Fig. 2.14.
4. Compute f[L] for L equal to the links in Fig. 2.15.
Jones polynom ial
The Jones polynomial and the Kauffman polynomial are really the
same, apart from a change of variable, so that the Jones polynomial
The Jones polynomial 29
Fig. 2.13
cg
&
Fig. 2.14
® ~
c@
Fig. 2.15
is an isotopy invariant of oriented links. The Jones polynomial of an
oriented link L is denoted V
L
and the customary choice of variable
is t. The formula connecting V
L
and I[L] is then
VL(t) = f[L](tli4).
The change of variable implies that the Jones polynomial can be
defined by the following two rules:
JPl: Vu(t) = 1, where U is the oriented unknot;
tIV ;f (t)tV (t)=(t
I12
_t
1I2
)V ~ (t).
JP2: )<...;< ~
Knot and link polynomials
In JP2 we show only that part of the link at which changes occur:
so JP2 relates the Jones polynomials of three links, identical except
at one crossing, at which we find the configurations shown in
Fig. 2.16.
x
x
Fig. 2.16
cises 2.2
1. Show that VL(f) = f[L](r I!4) does satisfy JPl and JP2.
2. Use JPl and JP2 to compute the Jones polynomial for the links
in Exercise 2.1.4.
Applications of the Kauffman polynomial
Kauffman's approach to the Jones polynomial enables relatively
simple proofs to be given of some of its formal properties; one such
property is the reversing result. It tackles the question: how is the
Kauffman or Jones polynomial of an oriented link affected if the
orientations of some of the components are reversed?
cises 2.3
(Exercises are placed at strategic points throughout this section, and
are numbered concurrently within it.)
1. If K is an oriented knot, and K is the reverse of K, show that
f[K] =f[K].
2. If L is a twocomponent oriented link, and the orientation of just
one component is reversed to form a new oriented link L', how
is f[L] related to f[L']? Experiment with some twocomponent
links and conjecture a formula relating f[L] and f[L 1. Try to
prove that your conjecture is correct. [Hint. Your formula should
involve the linking number of the two components.]
3. Consider the general setting of a ccomponent oriented link L
and a link L' obtained from L by reversing the orientation of r
components of L. How are f[L] and f[L1 related?
Applications of the Kauffman polynomial 31
One of the most striking applications of the Kauffman polynomial
concerns its relationship to the crossing number of a knot. The result
may be summed up as follows: any two reduced alternating diagrams
of a link L must have the same number of crossings. The meaning of
the term reduced will be explained in due course. Not every link has
an alternating diagram, but the result asserts that for those that do,
the crossing number in a reduced diagram is an isotopy invariant.
Proofs of the result have been given in Kauffman (1987b), Murasugi
(1987), and Thistlethwaite (1987). We shall follow Kauffman, as we
have done so far, but strongly recommend that the reader examine
the alternatives. We have need of one result whose proof is deferred
until Chapter 7, since it is most naturally seen as a special case of
Euler's formula for planar graphs; see Section 7.2.
Firstly we shall define reduced diagrams. An isthmus in a link
diagram is a crossing at which only three regions of the diagram
meet, as in the exam pIes in Fig. 2.17.
Fig. 2.17
A diagram is reduced if it does not contain an isthmus, and is
connected if it is not the disjoint union of two smaller diagrams. The
special case of Euler's formula that we need is the following.
rna. Let D be a connected link diagram with c crossings. Then D divides
the plane into c + 2 regions. .'
We need one more piece of terminology before proceeding to the
results. The span of a bracket polynomial (D) is the difference
between the maximum and minimum powers of A occurring in it;
recall that (D) may contain both negative and positive powers of A.
As an example,  A
5
 A
3
+ A
7
has span 5  ( 7) = 12. Clearly the
definition of span extends to any Laurent polynomial in one vari
able.
2.3
4. If D is a diagram of link L, what is the relationship between the
span of (D) and the span of f[L]?
Knot and link polynomials
We now state the main theorem on crossing numbers and the
Kauffman polynomial, and follow it by a more technical proposition
that contains the core of the argument. We then show how the theorem
is derived from the proposition, and give the proof of the proposition
assisted by a series of exercises.
Theorem. If a link L possesses a connected, reduced, alternating diagram D
with c crossings then the span of (D) is 4c. Hence any two connected,
reduced, alternating diagrams of a given link have the same number of
crossings.
The proposition uses the notion of positive and negative regions in a
link diagram, introduced earlier in the chapter. The sign of a region is
determined at a crossing to which it is adjacent, and in general the
allocation of signs will not be consistent for the same region at different
crossings. However, in an alternating diagram, the allocation of signs is
consistent, so that every region can be given an unambiguous sign.
Proposition. Let D be a connected, reduced, alternating link diagram with
c crossings, p positive regions, and n negative regions. Then the term
in (D) of maximum degree is ( l)n  lAc + 2n  2 and the term in (D) of
minimum degree is ( J)PI A
C

2
p+2.
Proof of the theorem. From the proposition we see that the span of
(D) is c + 2n  2  ( c  2p + 2) = 2c + 2(n + p)  4. Now n + p is the
total number of regions in D, which is also equal to c + 2. Therefore
the span of (D) is 2c + 2(c + 2)  4 = 4c. •
Proof of the proposition. Let S be the state of D with all the splitting
markers joining positive regions, so that peS) = c and n(S) = O. Then
I S I = nand (D I S) = A C and we see that S contributes a term
(l)n
1
A
c
(A
2
+ A
2
t
1
to the expansion of (D) as a sum over its
states, and so a term ( J)nI A
c
+
2n

2
to (D). We wish to show that
all other states contribute terms of smaller degree.
rcise 2.3
5 Consider a state SI obtained from S by changing exactly one
splitting marker. Show that I SII < 1 S I. [Hint. Use the fact that
D is reduced.]
Now consider an arbitrary state Sf. There exists a sequence of states
S= So, S], S2,"" SmI, Sm = Sf
Applications of the Kauffman polynomial 33
such that Si + I is obtained from Si by changing exactly one splitting
marker. Therefore P(Si+ I) = p(SJ  1 and n(Si+ I) = n(Si) + 1, and it
follows that (DISi+l) = A2(DISi).
Exercises 2.3
6. Show that I Si+ II = I S;j ± 1, and deduce that the largest degree
of a term in (D) contributed by Si + I is less than or equal to the
largest degree of a term contributed by Si.
7. Combine Exercises 2.3.5 and 2.3.6 to show that S' contributes
terms to (D) of degree smaller than c + 2n  2.
This completes the proof of the proposition for the term of
maximum degree.
8. Prove the statement in the proposition for the term of minimum
degree. •
We now consider a generalization of Theorem 2.7 observed inde
pendently by Murasugi and by Thistlethwaite; however, we continue
to follow Kauffman's account of the result, which we state at once.
2.9 Theorem. If a link L possesses a connected diagram D with c crossings then
the span of the bracket polynomial (D) is at most 4c.
This theorem follows from a result that Kauffman calls the dual state
lemma. Given a state S of a connected diagram D, the dual state S
is obtained from S by changing all the splitting markers.
2.10 Dual state lemma. Let S be a state of a connected diagram D and let S
be the dual state. Suppose that D has r regions. Then I S I + I S I ~ r.
Proof. Once again, some parts of the proof are left to you.
Exercise 2.3
9. Verify the dual state lemma for diagrams with at most two
crossings.
Now suppose that D has c crossings: select one crossing and split D
at that crossing to obtain diagrams Land R.
Exercise 2.3
10. Use the connectivity of D to show that at least one of Land R
is connected.
We may assume that it is L that is connected. The splitting to form
L is done according to a splitting marker in one of the states S or
and link polynomials
S; by interchanging these states if necessary, we may assume that the
splitting marker is in S. The other splitting markers in S now give a
state S' of L. Note that 1 S' 1 = 1 S I·
2.3
11. Show that L has r  1 regions. Use induction on c to deduce
that 1 S' [ + [S I ~ r  1.
12. Show that I S' I + 1 ~ I S I·
13. Complete the proof of the dual state lemma. •
Proof of Theorem 2.9. Again we consider the state S of D with all
the splitting markers joining positive regions, so that p(S) = c and
n(S) = O.
The proof of Proposition 2.8 shows that S contributes a term
A('[ (A2 + A
2
)] lsi I to (D), and so a term of maximum degree
c + 21 S 1  2, and that no other state can contribute a term of higher
degree to (D). Similarly S contributes the term of minimum degree
 c  2[ S [ + 2 Therefore the span of (D) is 2c + 2( 1 S I + IS [ )  4
and so by the dual state lemma, the span of (D) is at most
2c + 2r  4, where D has r regions. Now the fact tha t r = c + 2
completes the proof. •
oriented or Homfly polynomial
The appearance of the Jones polynomial revealed an unexpected
interplay between different fields of mathematics, and opened up the
possibility of advances in knot theory using the new results and
techniques. There were formal similarities between the defining
formulae for the Jones polynomial and that given for the Alexander
polynomial by J .H. Conway in his work on the enumeration of knots
and links (Conway 1969). These similarities indicated that the Alex
ander polynomial and the Jones polynomial were special cases of a
more general construction. A new polynomial invariant for knots
and links, generalizing both the Alexander and the Jones polynomial,
was found virtually simultaneously by four sets of mathematicians:
P.J. Freyd and D.N. Yetter, J. Hoste, W.B.R. Lickorish, and K.C.
Millett, and A. Ocneanu. Their work was published as a coauthored
paper (Freyd et al. 1985) and the polynomial was dubbed the
H.O.M.F.L.Y.
Let L be an oriented link. The Homfly polynomial for L, denoted
p(L)(e, m), is the unique twovariable Laurent polynomial defined
by
The oriented or Homfly polynomial 35
HPl: peL) is an isotopy invariant;
HP2: P(V) = 1;
HP3: if L+, L_, and Lo are three oriented links identical except at one
crossing, at which we see the configurations as shown in Fig. 2.18,
then fP(L+) + e1p(L_) + mP(L
o
) = O.
x x
L
Fig. 2.18
HPI allows the computation of peL) from any diagram of L, whilst
HP2 and HP3 give a practical procedure for carrying out such a
computation, by crossing changes and the elimination of crossings in
a diagram. Let us consider some examples. Observe that the orienta
tion of L is crucial, for it tells us how to eliminate crossings in an
application of HP3.
( 1) The ccomponent unlink (Fig. 2.19)
00
·0
Fig. 2.19
We denote the ccomponent unlink by V
C
• Since Vi = V, HPI tells
us that P( U
1
) = 1. Now consider V
2
• We begin with a diagram of
V
2
with no crossings, and take it as our Lo (Fig. 2.20). Then we
introduce a crossing to give L+ and L_.
000000
Fig. 2.20
and link polynomials
DOO DOO
L
Fig. 2.21
Each of L+, L_ is now an unknot, and so an application of HP3
gives f + e
I
+ mP( U
2
) = 0, and it follows at once that P( U
2
) =
 mI(f + e\ What happens for U
3
? We follow the same procedure,
taking as Lo a threecomponent unlink with no crossings, and
introducing a crossing to form L+ and L_ (Fig. 2.21); these are
twocomponent unlinks.
We have P(L+) =  mI(f + e
I
) = P(L_) and we want to calculate
P(L
o
) HP3 gives f[ mI(f + e
I
)] + e
I
[ mI(f + fI)] + mP(L
o
) = °
and we see that P(U
3
)=P(Lo)=[mI(f+e
I
)]2. Write p=m
I
(f + fI). We leave it to you as an exercise to complete an inductive
proof that P( U{) = pCI.
CD
Fig. 2.22
(2) The lefthand Hopf link (Fig. 2.22)
QD QD
Fig. 2.23
We operate on the top crossing, which has negative sign (do you
agree?), so that we take the given diagram as L_ to obtain the triple
shown in Fig. 2.23.
We know that P(L+) = P and that P(L
o
) = 1. An application of HP3
gives the equation fp + e
I
P(L_) + m = 0, from which it follows that
P(L_) =  f(m + fp) =  fm + f3
m
1 + fmI.
The oriented or Homfly polynomial 37
Some formal properties of the Homfly polynomial follow easily
from the definition. As for the Kauffman polynomial, we may easily
compute the Homfly polynomials of reverses and mirror images.
position. Let L be an oriented link, with reverse L and mirror image L
Then p(L)(e, m) = p(L)(e, m) = P(L!)(e
1
, m).
Proof. Reversing orientation on all components of L leaves the
signs of crossings unchanged, so that a computation of p(L ) is
identical to one of P(L). Taking a mirror image changes the sign of
every crossing, so that a computation of P(L!) is the same as one of
P(L) with e and e
I
interchanged. •
Computerassisted calculations of Homfly polynomials, the most
comprehensive of which cover all the 12965 knots with crossing
number at most 13, invite speculation on two key questions about
the Homfly polynomial:
1. Can we characterize the twovariable Laurent polynomials that
arise as Homfly polynomials of links?
2. Is there a nontrivial link L with P(L) = I?
We have yet to establish that the computational procedure implicit
in HPI, HP2, and HP3 does give a unique result. The details of this
fact can be found in Lickorish and Millett (1987). Their proof is by
induction on the number of crossings in a link diagram; the argu
ment is necessarily more subtle than for the bracket polynomial,
since the use of HP3 does not guarantee a reduction in the number
of crossings of diagrams. The inductive hypothesis is as follows.
Let : z ~ be the set of all link diagrams with at most n crossings, with
an ordering chosen for the components and a starting point selected
on each component. We assume that for every K EX
n
_ I' a poly
nomial p(K)(e, m) has been defined, such that p(K)(e, m) = P(K')
(e, m) if K and K' differ only in a choice of ordering of components
and of starting points on components, that P(K) is unchanged by
Reidemeister moves carried out withinZ
n
_ I> that it satisfies HP2
and HP3, and that in addition, if U
C
E5{;n _ I is a projection of a
ccomponent unlink, then p(UC)(e, m) =J.LCI.
This is a bit of a mouthful, so as an aid to digestion it is
recommended that you check that the induction starts successfully
on 20
0
, Note that P(U
C
) is then defined by the inductive hypothesis.
The definition of P is now extended to L E56
n
by applying crossing
changes to the components of D in order, beginning on each
component at the specified starting point, and carried out so that the
first encounter with each crossing is made along an underpass, as in
and link polynomials
the proof of Lemma 1.2. The result is an unlink aL with c compon
ents, and P(aL)(e, m) is defined as f.1CI. Now at each crossing
change, the inductive hypothesis is used to compute p(Ko)(e, m) for
each Ko that arises, and HP3 permits the calculation of a polynomial
defined as p(L)(e, m). With this detailed inductive definition of
peL) to hand, it is now a matter of showing that it satisfies the
requirements of the inductive hypothesis on .Y5
n
. This verification is
accomplished in Lickorish and Millett (1987), to which we refer the
reader for further details.
2.4
1. Verify that if L is the righthand Hopf link (Fig. 2.24) then
peL) =  eIm + e
3
m
1
+ eIm
l
.
m
Fig. 2.24
2. Use the triple of diagrams in Fig. 2.25 to show that the Homfly
polynomial of the lefthand trefoil is em
2
 2e
2
 e
4
•
Fig. 2.25
3. Guess the Homfly polynomial of the righthand trefoil, and
check your guess with a computation.
Alexander polynomial
l.W. Alexander introduced the polynomial that now bears his name
(Alexander 1928) as a simple and effective invariant for distinguish
ing knot types. The Alexander polynomial may now be obtained
from the Kauffman or Homfly polynomials by a change of variable,
The Alexander polynomial 39
so is not as effective as these parvenus, but it retains a great deal of
intrinsic interest because of its connection with knot groups, which
we shall see in Chapter 8, and with covering spaces in Chapter 12.
Here we discuss Alexander's original treatment of his polynomial,
which is based on an ingenious calculation involving determinants,
and uses a labelling of knot diagrams that is a forerunner of the idea
of state used by Kauffman.
Let D be an oriented link diagram with c crossings. Then by
Lemma 2.6 D divides the plane into c + 2 regions. We construct a
matrix M(D) whose rows are indexed by the crossings of D and
whose columns are indexed by the regions, so that M(D) has size
eX (c + 2).
Fig. 2.26
Suppose that regions P, Q, R, and S meet at the ith crossing of D
(Fig. 2.26). Then in row i of M(D) we enter t, t, 1, 1 in columns P,
Q, R, and S and zero in all other columns. Note that according to
our sign convention for regions, Q and S are positive. At an isthmus,
with only three regions adjacent to a crossing, the column indexing
the repeated region receives the sum of the symbols allocated to that
region (Fig. 2.27).
Fig. 2.27
For example, the figureeight knot with crossings and regions
labelled as shown in Fig. 2.28 has the associated matrix
and link polynomials
p
Q
R S T U
r
0  1 1
tJ
M{D) = 2 0 t  I t o .
3 t t 0 0 I  I
4 t  1 0 1 0  t
Q
Fig. 2.28
Now delete from M(D) any two columns indexed by a region
which have a common boundary arc in D; call the resulting (c x c)
matrix Mo{D). The Alexander polynomial AD of D is defined to be the
determinant of Mo{D): AD = det{Mo{D)). It is clear that AD is a
polynomial in t with integer coefficients. If, in the above example, we
delete columns T and U to obtain
p
Q
R S
I 0 0  1 t
Mo{D) = 2 0 t  1 I
3 t 0 0
4 t  1 0 I
and expand the determinant by the first row, we find
..1
D
{t) =  1( t + 2t
2
)  t(t  t
2
) = t
3
 3t
2
+ t.
If instead we delete columns Rand S to obtain
p
Q T U
I 0 0
J
t
Mo{D) = 2 0  t 0
3  t  I
4 t  1 0  t
The Alexander polynomial 41
and again expand by the first row, we find
LlD(t) = t
2
 t
3
+ t( t
2
 t
2
+ t
3
) = t
4
 3t
3
+ t
2
.
You will observe that this is a different answer. However, it differs
from the first answer by a factor of t. This is a typical outcome, as
explained by a theorem proved by Alexander.
orem. Given an oriented link diagram D, the determinant det(Mo(D»
is independent of the choice of columns deleted from the matrix M(D),
up to a factor of ± t
m
for some integer m. Furthermore, det(Mo(D» is
then an isotopy invariant. •
We shall not prove the theorem here, but recommend the reader to
enjoy the discursive style of Alexander (1928). Allowing for adjust
ments of sign and powers of t, we now have an isotopy invariant for
a link L computable from any oriented diagram, namely the Alexan
der polynomial of L, denoted LlL(t). The convention is to adjust
det(Mo(D» by a term ± t
m
to give a positive constant term and a
positive highest power of t.
2.5
1. Choose your favourite knot or link from the book so far, and
compute its Alexander polynomial.
3
Topological spaces
The purpose of this chapter is to provide the theoretical background
for a rigorous discussion of knots and surfaces, against which the
vaguely expressed ideas lying behind our previous discussion can be
made precise. The central notion is that of a topological space. The
precision that topological spaces bring to the ideas of continuity and
nearness is, of course, essential to any proper treatment of knots and
surfaces. Moreover, topological ideas will prove crucial in the devel
opment of more sophisticated invariants in later chapters. Since our
interest in topological spaces stems from specific applications that we
already have in mind, we make no attempt to give a comprehensive
account of even the basic theory of topological spaces here. There are
several textbooks that do the job well, and we mention Armstrong
(1983) and Brown (1988) as providing suitable further reading.
ological spaces
Topological spaces provide an abstract framework for the notion of
continuity, and as is usually the case with abstraction, there is a great
gain in the clarity of the concept, even if some of the examples do
not accord with our intuitive expectations. However, the definition
of continuity in terms of e and {j familiar from introductory analysis
is now seen to arise from a particular topology chosen for the real
numbers. This topology is the one appropriate for the analytic
viewpoint, and we call it the usual topology on ~ n to emphasize its
significance.
What are the concepts that we need to make precise in order to put
our study of knots and surfaces on a firm base? A glance back at our
informal definition of a knot in Chapter I shows immediately that
we need to define continuous maps from the circle Sl to ~ 3 in order
to describe a knot as a continuous copy of S, and for links we shall
need to define a continuous map from the union of several copies of
Sl into ~ 3 . In the treatment of surfaces in Chapter 4, we shall have
5
Topological spaces 43
need of a notion of a subset of IR
n
'looking like' 1R2 'near' any point.
What do we mean by 'looking like' and 'near'? In IR, 'near' often
means 'within EO of'; that is, we fit a small, open interval around the
given point. What might 'open' mean in more generality? With these
questions in mind, let us proceed to the definition of a topological
space without more ado.
Let X be a nonempty set. A topology on X is a collection of subsets
of X, called open sets, satisfying
OSI: the empty set and X itself are open sets;
OS2: an arbitrary union of open sets is an open set;
OS3: the intersection of finitely many open sets is an open set.
The set X, together with the collection of its special subsets that we
have called open, is a topological space. We shall usually denote a
topology by a script symbol such as ':!T, so that a space is a pair
(X, ':!T). However, it is customary to suppress explicit mention of ':!T
and refer to X as a topological space, keeping the topology in the
back of our minds. Note that a given set can have many different
topologies; there is nothing intrinsically special about the open sets,
other than that we have selected them to playa certain distinguished
role. However, in many circumstances, there will be good reasons to
select one topology above all others.
1. Given any set X, we may define every subset of X to be open.
This is the discrete topology on X.
2. Again on any set X, define a subset A ~ X to be open if either
A = 0 or A has finite complement in X, that is I X\A I < 00. This
is the finite complement topology on X. What do we get if X is a
finite set?
3. The usual topology on, [Rn, (n ~ 1).
For x, Y E IR
n
, define
[
112
IIx  yll ~ ~ (x,  y,)' J
This is the customary notion of the distance between two points in
IRn. The open ball with centre x and radius r, denoted by B(x; r), is the
subset {y E IRk I IIx  yll < r}; here r E IR and r > O. If k = 1, we see
that an open ball in IR is an open interval:
ropological spaces
r r
x
whilst if k = 2, an open ball in ~ 2 is an open disc. The usual topology
on ~ n is now defined in terms of open balls in the following fashion.
We say that a subset U ~ ~ n is open if, for every x E U, there exists
r E ~ with r> 0 such that B(x; r) ~ U. Note that r depends on x.
Let X be a topological space and let A ~ X be a subset. We say
that A is closed if its complement X\A is open. If X is a topological
space, any subset A ~ X can be made into a topological space by
defining the open sets in A to be all sets A n U where U is an open
set in X. This is the subspace topology on A and we say that A is a
subspace of X.
The following result will be of use to prove the important gluing
lemma later in this chapter.
emma. If A is a closed subspace of a topological space X, and if C ~ A is
closed in A, then C is closed in X.
ises
Proof By definition, A\C is open in A, whence A\C = A n U for
some open set U ~ X. Therefore,
X\C = (X\A) u (A\C) =(X\A) u (A n U) = (X\A) u U.
Now X\A and U are open in X and so by OS2 X\C is open in x.
Therefore C is closed in X. •
3.1
1.
2.
3.
4.
5.
6.
Let X have the discrete topology. Which subsets of X are closed?
What if X has the finite complement topology?
In ~ n with its usual topology, show that the closed ball with
centre x and radius r, defined as B(x; r) = {y E ~ n I IIx  yll ::::; r}
is indeed a closed subset of ~ n .
Show that the subspace topology is a topology on a subset A of
a topological space X.
Show that any subspace of a topological space with the discrete
topology also has the discrete topology. Prove that, as a subspace
of ~ with its usual topology, 7L has the discrete topology.
Show that the union of finitely many closed subsets of a topo
logical space is again a closed subset.
(a) Let X be a topological space. We call a subset N ~ X a
neighbourhood of x E X if there exists an open set U ~ X such that
x E U ~ N. Let N(x) denote the collection of all neighbourhoods
of x E x. Show that N(x) is nonempty and satisfies the following
four properties:
NHDl: if N E N(x) then x E N;
NHD2: if N E N(x) and AI ~ N then M E N(x);
NHD3: if N, N' E N(x) then N n N' E N(x);
Continuity 45
NHD4: if N E N(x) and N° = {y E NIN E N(y)} then ~ E N(x).
(b) Suppose now that we are given a set Y and for each y E Y a
nonempty collection N(y) of subsets of Y satisfying NHDl, ... ,
NHD4. Define U ~ Y to be open if, for every u E U, U E N(u). Show
that this is a topology on Y.
(c) Part (b) opens up the embarrassing possibility of going round
in everdecreasing circles, repeatedly defining 'open sets' and 'neigh
bourhoods' in terms of one another. Suppose we start with a
topological space X, with neighbourhoods as in (a), and define open
sets as in (b). Are these new open sets the same as the open sets of
X that give the topology? What if we start with a set Y as in (b),
define a topology on it as in (b), and then define neighbourhoods as
in (a)? Are these neighbourhoods the same as those in the given
collections N(y) for y E Y?
7. Find all the possible topologies on a set with precisely three
elements. How many different topologies are there for a set with
precisely four elements?
ntinuity
Let X and Y be topological spaces. (Note that we suppress any
mention of the topology on the underlying sets, but you should keep
in mind that the topologies may be very different.) A function
f:X + Y is continuous if, for every open set V in Y, its preimage
f!(V) is open in X. (Recall that the preimagef1(V) is the subset
{x E X If(x) E V}; we do not assume that an inverse f
'
to f exists.)
Since continuity is defined in terms of open sets, it depends on the
choice of topology for each of the sets X and Y. For example, if 2J
and 2J' are different topologies on X then the identity map x H x
need not be a continuous function (X, 2J) + (X, 2J'). However, the
familiar notion of 'e  8' continuity is identical to the notion of
continuity with respect to the usual topologies on [Rm and [Rn.
position. Let [Rm and [Rn have their usual topologies. Then f: Rnl + [Rn
is continuous if and only if, given any e> 0, there exists 0> 0 such that
IIx  yll < 0 implies that IIf(x)  f(y)1I < e.
ological spaces
Proof. Suppose thatfis continuous and let e> 0 be given. The open
ball B(f(x); e) is an open set in [Rn and therefore f
I
(B(f(x); e» is
an open set in [Rm: since x EfI(B(f(x);e» there exists 8> 0 such
that B(x; 8) r;;;;.fI(B(f(x); e», that is IIx  YII < 8 implies that
IIf(x)  f(y)1I < e. The proof of the converse we leave as an exercise
for the reader. •
mao Let X, Y, and Z be topological spaces. If f: X ~ Y and
g: Y ~ Z are continuous functions, then gf: X ~ Z is continuous.
Proof. Let W be an open set in Z. By the continuity of g, we know
that V= gI(W) is open in Y, and by the continuity offwe know that
fI(V) is open in X. Thus (gi)I(W) =fI(gI(W» is open in X and
gf is continuous. •
gluing lemma. Given a function f : X ~ Y, suppose that there are
closed sets A and B such that X = A u B and such that the restrictions
f I A: A ~ Y and fl B: B ~ Yare continuous. Then f is continuous.
Proof. Let V be an open set in Y. We want to show that f
I
(V) is
open in X. Continuity of the restricted functions implies that
f\V) n A is open in A and that fI(V) n B is open in B; thus
A\fI(V) is closed in A and B\fI(V) is closed in B. By Lemma 3.1,
each of A\fI(V) and B\f
1
(V) is closed in X. Using Exercise 3.1.5
we see that X\fI(V) = (A\f
1
(V» u (B\fl(V» is closed in X and
so fI( V) is open in X. •
Let X and Y be topological spaces. A function f: X ~ Y which is
bijective, continuous, and has a continuous inverse f
I
: Y ~ X is
called a homeomorphism. A function f: X ~ Y which is injective,
continuous, and such that the bijection f: X ~ f(X) has a continu
ous inverse is called an embedding. Simple examples of embeddings
include [R2 ~ [R3 given by (x, y) H (x, y, 0) and Sl ~ [R3 given by
(x, y) H (x, y, I) where x
2
+ y = I.
If X and Yare topological spaces, is there a natural topology on
the set X x Y = {(x, y) I x E X, Y E Y}? The first question to answer is
this: what do we mean by 'natura!'? Clearly the topology will have
to be defined in terms of the given topologies for X and Y. Furthermore,
there are the canonical projection functions p: X x Y ~ X given by
(x, y) H x and q: X x Y ~ Y given by (x, y) H Y and these should
surely be continuous. So let us try to construct a topology meeting
these requirements. Let U r;;;;. X be open: then pI (U) is U x Y so that
U x Y needs to be one of our open sets in X x Y; similarly we
need X x V for any open set V in Y. However, (U x Y) n (X x V) =
Continuity 47
u x V and so if we are to satisfy OS3 we require that all sub
sets U x V, where U is open in X and Y is open in Y, are open in
X x Y. Then to satisfy OS2 we require arbitrary unions of such
subsets to be open, and this is enough. If we define A ~ X x Y to be
open if A is a union of subsets U x V, where U is open in X and V
is open in Y, then we do obtain a topology on X x Y: it is called the
product topology on X x Y.
3.2
I. Let X be a topological space. Prove that every function
f: X ~ Y, where Y is any topological space, is continuous if and
only if X has the discrete topology.
2. Let (X, fJ) and (X, fJ') be topological spaces. Find a condition on
the topologies fJ and fJ' that ensures that the identity map
x H x is a continuous function (X, fJ) ~ (X, fJ').
3. Given sets X and Y, their disjoint union X II Y is the set obtained
by taking the union of disjoint copies of X and Y, so that in
X II Y we have two copies of any element in X n Y. If
X n Y = 0 then X II Y is the usual union Xu Y. How might we
define a topology on the disjoint union X II Y of topological
spaces X and Y? The simplest idea is to take a subset U ~ X II Y
to be open if U = V II W where V is open in X and W is open in
Y. Show that this does define a topology, and makes the inclu
sions of X and Y into their disjoint union continuous functions.
4. Find an example of a continuous bijection whose inverse is not
continuous. [Hint. You have already seen examples if you have
done the exercises so far.]
5. Let X be a topological space and let U ~ X be a subspace. Show
that the inclusion U ~ X is an embedding. If X is a topological
space and U is a subset of X with a topology such that the
inclusion is an embedding, need the topology on U be the
subspace topology?
6. Verify that the product topology is indeed a topology on X x Y.
7. Describe the neighbourhoods of a point (x, y) E X X Y in the
product topology.
8. A topological space X is called Hausdorff if, given any distinct
elements x, y E X, there exist open sets U containing x and V
containing y with U and V disjoint; that is, we have x E U and
y E V with U n V = 0. (The concept is named in honour of F.
Hausdorff (1 8681942) and a punning aidembnoire is to say that
the elements x and yare housed off by the opens sets U and V).
The diagonal subspace of X x X is L1(X) = {(x, x) I x EX}. Show
that X is Hausdorff if and only if the diagonal subspace is closed
Topological spaces
in X x X. Suppose that the finite complement topology on a set
X makes X a Hausdorff space. What must be true of X? Find
some nonHausdorff topologies.
9. Let U ~ Xbe a continuous injection between topological spaces.
Show that if X is Hausdorff then so is U.
Connectedness
In addition to providing a framework for continuity, topological
notions also allow us to consider connectedness. Intuitively we want
to make precise the idea of a topological space being in one piece,
and we do this by considering what we might be able to do with a
space that was not in one piece. We surely expect to be able to divide
such a space into at least two subspaces that do not intersect, and
this leads to the following definition.
A disconnection of a topological space X is a pair (A,B) of
nonempty closed subspaces of X such that X = A u Band
A n B = 0. A space X is connected if no disconnection of it exists.
A key property of connected spaces is that the only subsets which
are both open and closed are the empty set and the whole space. We
shall make use of this fact in Chapter 12. We shall show here that
this scarcity of subsets that are both open and closed is equivalent to
connectedness, so that we could have chosen it as the definition.
Would this have been a sensible thing to do?
Proposition. A topological space X is connected if and only if the only subsets
of X that are both open and closed are the empty set and X itself.
Proof. We first suppose that X is connected, and let A be a subset
of X that is both open and closed. Let B be the complement of A in
X, so that B is also both open and closed. If both A and Bare
nonempty, we see that (A, B) is a disconnection of X; since X is
connected, no disconnection exists. Therefore one of A and B is
empty, and the other is then X itself.
Conversely, suppose that the only subsets of X that are open and
closed are 0 and X but that X has a disconnection (A, B). Then A is
closed, so that its complement B is open; but B is also closed and
since B is nonempty we must have B = X. But then A = 0, which is
a contradiction. •
Given a continuous functionf; X ~ Ywith X connected, might we
expect Y to be connected? A little thought should convince you that
there is no reason to expect such a result for Y itself, but that it does
Connectedness 49
seem reasonable to suppose that the image of f will be connected.
This reasonable supposition turns out to be true.
Preservation of connectedness
Let X be a connected space and let f: X ~ Y be a continuous function.
Then the image f(X) off is connected.
Proof. Suppose that A is a subset of f(X) which is both closed and
open. Since f is continuous, f1(A) is both closed and open. By
connectedness of X, f
I
(A) is either X or the empty set, and it
follows that A is either f(X) or the empty set. Therefore f(X) is
connected, by Proposition 3.5. •
rcises 3.3
1. Show that ~ (with the usual topology) is connected, but that the
subspace (() is not connected.
2. Given any topological space X, define a relation  on X as
follows. For any x, y E X we say x  y if and only if x and y
belong to some connected subspace of X. Show that  is an
equivalence relation on X. The equivalence classes are called the
connected components of X. A space in which every connected
component contains just one element is said to be totally discon
nected. Show that any space with the discrete topology is totally
disconnected. Is the converse true?
Identification spaces
Let X be a topological space, and let f: X ~ Y be a function to a set
Y. Can we equip Y with a topology that makes f continuous? Of
course we can! We can give Y the indiscrete topology for which the
only open sets are Y itself and the empty set. Evidently this is
something of a swindle, because the topology has nothing to do with
the function f and has made Y into a dull sort of space that accepts
every function into it as continuous. Perhaps we should reformulate
the problem. It is better to have as rich a structure of open sets as
we can on Y subject to the condition that f be continuous, and this
idea motivates the following definition.
Given a topological space X, a set Y, and a function f: X ~ Y, the
identification topology on Yis obtained by defining U ~ Yto be open
if and only ifits inverse imagefI(U) is open in X. It is immediately
obvious that this definition makes f continuous.
Topological spaces
mples
An important application of the identification topology is to the
quotient set of an equivalence relation. Let X be a topological space,
and suppose that  is an equivalence relation on X. Is there a natural
topology on the set of equivalence classes X/? As in our discussion
of the product topology, we require the topology on X/ to be
defined in terms of the given topology on X. Further, there is a
function that we shall require to be continuous: namely, the quotient
function X ~ X/ that takes an element x E X to its class [xJ E X/.
We achieve all that we require by giving X/ the identification
topology defined by the quotient function.
We can now take any equivalence relation on a topological space
X and construct a new space, namely the quotient set of the
equivalence relation equipped with the identification topology. In
particular, we can identify certain elements of a space with one
another by an equivalence relation; in picturesque terminology, we
think of gluing one part of a space to another part. Here are some
examples of this process in action; it is tempting to recommend the
acquisition of paper and a pot of glue before reading further!
(a) Take a rectangle, and identify two opposite ends after perfor
ming a halftwist. The result is a Mobius band (or strip).
Formally we might begin with a rectangle in [R!2, say
R = {(x, y) E [R!21 o:s.:: x :s.:: 2, 0 :s.:: y :s.:: I}. where [R!2 has its usual
topology and R has the subspace topology, and take the equi
valence relation  on R generated by equality and the additional
relations (0, y)  (2, 1  y) for all y satisfying 0 :s.:: y :s.:: 1. Then
R/ with the identification topology is a mathematical model for
a Mobius band.
(b) Begin again with a rectangle, and glue together corresponding
points of two opposite sides. The result is an openended cylin
der. Now bend the cylinder so as to bring together the boundary
circles of the two open ends, and glue these circles together. The
result is a hollow, doughnutshaped surface called a torus.
More formally, we take R as in (a) and impose the equivalence
relation generated by equality and the further relations
(0, y)  (2, y) and (x, 0)  (x, I).
(c) What happens if we try to combine (a) and (b) by introducing a
twist into our construction of the torus? We impose on R the
equivalence relation generated by equality and the further relations
(x, 0)  (x, 1) and (0, y)  (2, 1  y). This appears to be harmless
enough, but if you try and envisage what happens to a rectangle
Knots and isotopy 51
Fig. 3.1
under the corresponding bending and gluing, a difficulty at once
emerges. The nicest picture is obtained by imagining the cylin
der formed by one gluing to pass through itself before the
boundary circles of the cylinder are identified (Fig. 3.1).
The resulting surface is known as the Klein bottle. Figure 3.1
does not truly represent the Klein bottle because of the self
intersection, but no model of it can be made in three dimensions
(nor any picture drawn in two dimensions!) without such self
intersections. The Klein bottle is an example of a nonorientable
surface, and as such will reappear in Chapter 4.
(d) Let f: X ~ Y be a continuous function between topological
spaces, and let J denote the unit interval [0, 1] with the subspace
topology in ~ . Form the space X x J with the product topology,
and then the disjoint union (X x J).u Y with the topology given
in Exercise 3 of Section 3.2. Now impose the equivalence relation
generated by equality and the further relations (x, 1) ~ f(x). The
resulting identification space is called the mapping cylinder of f
and it is a useful construction in the further theory of topologi
cal spaces; see (Brown 1988, chapter 7).
3.4
1. Given topological spaces X and Y and a continuous function
f: X ~ Y, let (Y, f) denote the set Y given the identification
topology. Show that the identity map (Y, f) ~ Y is continuous.
2. Find an embedding of the torus into ~ 3 .
s and isotopy
Hereafter, ~ n will always have the usual topology, and we denote by
Sl the unit circle {(x, y) I x
2
+ y = I} ~ ~ 2 with the subspace topology.
logical spaces
We now have the necessary theoretical background to give our
precise definition of a knot. Here it is: a knot is an embedding
k: SI ~ ~ 3 .
Our definition thus defines a knot as a function k from SI that has
certain nice properties. Firstly, k is injective; this corresponds to our
informal insistence in Chapter I that the string of a knot does not
intersect itself. Secondly, k is continuous and there is a continuous
function back from its image to SI. This takes care of our require
ment that the string forms an unbroken, albeit entangled, copy of the
unit circle.
A link of c components is an embedding
SI
11
SI
11
... 11SI ~ ~ 3
of c disjoint copies of SI into ~ 3 . The topology on SI 11 Sill ... 11 SI
is of course that for a disjoint union given in Exercise 3 of Section
3.2.
When should we consider two knots k and 1 to be the same? We
want to capture the notion of their knottedness or interlacing being
the same, and in ~ 3 we might think of moving their images around
in a continuous fashion until one image coincides with the other, just
as we might deform a physical piece of string from one interlaced
form into an equivalent one. The idea that we move the images in
~ 3 continuously leads to the concept of ambientisotopy in which
deformations of the images are effected by applying homeomorph
isms of the ambient space ~ 3 , and in the following definition it may
be helpful to think of the parameter t as representing time, so that
the definition provides a continuous sequence of homeomorphisms
of ~ 3 from time t = 0 to time t = I. Now to the definition: two knots
k and 1 are ambientisotopic if there exists a continuous function
H: ~ 3 x [0, I J ~ ~ 3 such that
(a) ho = H(, 0) is the identity ~ 3 ~ ~ 3
(b) for all t E [0, I], hI = H(, t) is a homeomorphism ~ 3 ~ ~ 3
(c) if hi = H(, I), then hJk = I.
If k and 1 are ambientisotopic, then the knots htk (t E [0, 1]) form
a continuously varying family deforming k into I. There is a corres
ponding definition of ambientisotopy for links, which we leave the
reader to formulate.
Ambientisotopy of knots and links is precisely the topological
analogue of the isotopy of knots that we considered in Chapter I.
We state this crucial fact as a theorem below, but we shall not give
a proof. The theorem is due to Reidemeister, who has given a proof
Knots and isotopy 53
in his book (Reidemeister 1932). A more conceptual account is to be
found in Burde and Zieschang (1985).
heorem. Two knots are ambientisotopic if and only if they possess isotopic
diagrams. •
There are other equivalence relations between functions S ~ ~ 3
that might be used to define a notion of two knots being essentially
the same. We can consider a homeomorphism ~ 3 ~ ~ 3 as analogous
to a change of coordinates, and so defme two knots k: Sl ~ ~ 3 and
I : Sl ~ ~ 3 to be equivalent if there exists a homeomorphism
h: ~ 3 ~ ~ 3 such that 1= hk. An equivalence class of knots under the
equivalence relation of equivalence is commonly called a knot type.
It is immediate that ambientisotopic knots are then equivalent, and
it should come as no surprise that equivalence is a much weaker
notion than ambientisotopy. For example, it is obvious that the
righthand and lefthand trefoils are equivalent.
A knot k is piecewise linear if its image in ~ 3 is a union of a finite
number of line segments. A piecewise linear knot can be represented
by a diagram whose arcs are straight lines. A knot is said to be tame
if it is equivalent to a piecewise linear knot. Tameness may be
considered as a technical restriction on knots to ensure that each has
only finitely many crossings, and the theory developed in this book
will treat exclusively tame knots. All other knots are wild; the theory
of wild knots encounters different problems to that of tame knots
and we propose to say no more about it.
ises 3.5
1. Why is it obvious that the righthand and lefthand trefoils are
equivalent?
The final three exercises in this chapter propose, and demolish, a
third possible equivalence relation for knots.
2. Call two knots k: SI ~ 1R3 and I: SI ~ ~ 3 comparable if there
exists a continuous function H: S x [0, 1] ~ ~ 3 such that
(a) ho = H(, 0) = k
(b) hi = H(, 1) = I.
Show that comparability is an equivalence relation on knots.
3. Prove that equivalent knots are comparable.
4. Prove that any two knots are comparable [Hint. Show that any
given knot is comparable with the unknot, by the wellknown
process of pulling the knot tight.]
4
Surfaces
Knots are onedimensional objects living in threedimensional space.
Surfaces are twodimensional objects. Some of the ideas that help in
understanding knots can also be adapted to help study surfaces. The
flow of ideas is two way and we shall see later on how surfaces help
us to study knots. The study of surfaces, however, has no need of
knot theory for motivation. There are important applications of the
classification theorem for surfaces in other areas of mathematics,
perhaps most notably in complex analysis, but we will not be going
in that direction here.
Sphere
Fig. 4.1
Torus
Fig. 4.2
Surfaces 55
What is a surface? In the last chapter, we saw some surfaces
constructed by gluing. We have examples such as those shown in Figs
4.14.3.
Double torus or pretzel
Fig. 4.3
Here we also want to allow for holes or edges in surfaces, as in
Figs 4.44.7.
Sphere
with
one hole
Fig. 4.4
Fig. 4.5
Recall from Chapter 3 that the Mobius band (Fig. 4.8) is obtained
by twisting and gluing a rectangle (Fig. 4.9).
Here we have seen some of the simpler examples of surfaces. Can
we list all of them? This question needs tightening up before it really
faces
Sphere
with
two holes
Fig. 4.6
Fig. 4.7
Fig. 4.8
will make sense. There are infinitely many surfaces; for instance,
consider a multiple torus with n 'handles', as in Fig. 4.10.
For different values of n, it seems likely that the corresponding
multiple tori will be different. Likewise a sphere with n holes is not
likely to be the same as a sphere with m holes (unless m = n, of
j
I
Fia.4.9
Combinatorial surfaces 57
Fig. 4.10
course!). Thus thinking of the problem a bit more we might try to
classify surfaces according to certain invariants such as the number
of holes or the number of tori in the above examples. Then, given
any surface, we would hope to be able to calculate its invariants and
assign it some standard name such as an ntorus with m holes in it.
This basic idea in fact works quite well but of course our view of it
here is a bit primitive. For instance, we have used 'same as' several
times without worrying about what that means. To get that point out
of the way, we will be more precise and say that we want to classify
surfaces 'up to homeomorphism' and so we may tacitly use phrases
such as 'SI is the same as S2' as a more relaxed way of saying 'There
is a homeomorphism between SI and S2'.
mbinatorial surfaces
We want to use geometric and combinatorial arguments wherever
possible. To do this we are first of all going to consider combinator
ial surfaces.
A combinatorial surface, S, is a finite union of (solid) triangles
satisfying the following rules:
S.l: Any two triangles meet in one edge, one vertex, or not at all.
S.2: Any edge is an edge of one or two triangles.
S.3: For any vertex v, we write St v for the union of all triangles
having v as a vertex; then St v is homeomorphic to a disc (St v is
called the star of v).
S.4: Given any two vertices in S, then there is a path along edges of
triangles joining them.
We now give some comments and further explanation of the rules.
S.l: Two of the triangles in S can meet as in Fig. 4.11, or as in Fig.
4.12, or they may not meet at all, as in Fig. 4.13.
Two triangles cannot meet as in Fig. 4.14, or as in Fig. 4.15, or as
in Fig. 4.16.
58 Surfaces
Fig. 4.11
Fig. 4.12
Fig. 4.13
S.2: An edge, e, might only be in one triangle. This happe
on the boundary of some hole (Fig. 4.17).
Fig. 4.14
Combinatorial surfaces 59
Fig. 4.15
Fig. 4.16
Fig. 4.17
Many edges will be in two triangles, but none may be in three or
more.
S.3: The star of a vertex, v, will normally look as shown in Fig. 4.18.
If the vertex is in a boundary edge, its star may be different, as in
Fig. 4.19, but it cannot be as shown in Fig. 4.20.
faces
Fig. 4.18
Fig. 4.19
Fig. 4.20
A path in a topological space X is a continuous map a: [0, 1] ~ X.
We say a joins the points a(O) and a(l) and t h ~ t X is path connected
if any two points can be joined by a path; that is, given x, y E X,
there is a path a in X such that a(O) = x, a(l) = y. We will be looking
Combinatorial surfaces 61
a b
c
d d
e
e
f f
a b
c
Fig. 4.21
at paths much more closely in Chapter 9. For the moment, we merely
note that S.4 implies that S is path connected.
As we saw in Chapter 3, we can indicate the construction of surfaces
by diagrams giving gluing instructions, for example as in Fig. 4.21.
In such a diagram the edges with the same label are to be identified
with the arrows matching. Forgetting for the moment the edges
inside the rectangle, we can imagine performing these identifications
within 1R(3 (in this case) and the result is a torus (Fig. 4.22).
f
•
t
6)
+ ()
)
+
Fig. 4.22
e b
c
f
d
e e
d f
a b c
Fig. 4.23
aces
t
+ ('L...) _ ~ )
Fig. 4.24
Gf)
How can we
glue this?
The reason why we had to put 'in this case' is that a different
diagram such as in Fig. 4.23 still yields a surface, the Klein bottle
(Fig. 4.24), but the identification cannot be done in /R
3
•
To obtain some idea of what this space looks like it is usual to
draw it with a selfintersection as in Fig. 4.25.
Fig. 4.25
We will see some of the strange properties of the Klein bottle later.
To see the restrictions that the rules place on the way the triangles
fit together, let us try to triangulate the cylinder (represented as a
rectangle with identification in Fig. 4.26).
}
Fig. 4.26
Attempt 1 (Fig. 4.27)
This simple decomposition contravenes S.l as the two triangles meet
in two edges, so it does not give a triangulation.
Combinatorial surfaces 63
Fig. 4.27
Attempt 2 (Fig. 4.28)
Triangles I and 2 meet in three vertices, so this also contravenes S.l.
Fig. 4.28
Attempt 3 (Fig. 4.29)
Triangles I and 3 meet in two vertices, so again this is disallowed.
Fig. 4.29
Attempt 4 (Fig. 4.30)
Again the intersection of two of the triangles (2 and 4 this time)
consists of just two vertices.
Fig. 4.30
aces
Attempt 5 (Fig. 4.31)
Fig. 4.31
S.l is now satisfied as is S.2. Path connectedness is clear, so we are
left to check S.3, the condition on stars of vertices. We leave you to
check this in detail but it is clear that St VI is the union of triangles
4 and 5, so is a triangle and thus homeomorphic to a disc, whilst
St V
2
consists of triangles I, 2, 7, and 8 and so similarly is a rectangle
and hence homeomorphic to a disc. This leaves six other vertices for
you to check.
So we managed to find a triangulation of the cylinder. It was not
as obvious as it may first have seemed. We do not claim to have found
the smallest number of triangles. Can you improve on our number of
eight? What is the smallest number of triangles needed to get a
triangulation of the cylinder? How might you attack this question?
Our main aim will be to classify surfaces up to homeomorphism.
To do this we will assume that we have triangulated our surface, but
we are interested in the surface, not primarily in the triangulation,
and so we can and will change the triangulation as needed, perhaps
chopping up some of the triangles into smaller triangles (subdivi
sion). Our search will be for invariants of the surface, that is
invariants that do not depend on which triangulation we are using.
We will also see certain moves on combinatorial surfaces which
change the triangulation. This is much like the situation we saw
earlier where moves on knot diagrams allowed us to check if certain
quantities were invariants.
The first invariant property we will look at is orientability. There
are two basic classes of surfaces: those which are orientable and
those which are nonorientable. For instance, the sphere, torus, and
multiple tori are orientable whilst the Mobius band and the Klein
bottle are nonorientable. We met the Mobius band in Chapter 3. It
has a strange singlesidedness property. You can draw a line from
any point on the surface to any other without going over an edge.
The same is true of the Klein bottle. These strange properties are due
to the nonorientability of these surfaces.
Combinatorial surfaces 65
Another important nonorientable surface is the projective plane.
This first arose in projective geometry. The set 1]J>2 is the set of lines
through 0 in !R
3
• This set can be thought of in several other ways. We
can consider the set of nonzero points in !R
3
and put on them an
equivalence relation defined by
(x, y, z) ~ (x', y', z')
if there is some nonzero A E !R such that AX = x', AY = y, Az = z', that
is (x, y, z) and (x', y', z') are on the line through O. The set of
equivalence classes is 1]J>2. Alternatively since each line through the
origin in !R
3
cuts the unit sphere, 52, in two points, (x, y, z) and
( x,  y,  z) say (where x
2
+ y + Z2 = 1), we can consider an equi
valence relation on 52 generated by
(x, y, z) ~ ( x,  y,  z).
Again the set of equivalence classes form 1]J>2. This gives us a clue as
to how to give 1]J>2 a topology. If (x, y, z) is in 52, write [x, y, z] for
the corresponding point of 1]J>2. This gives a function f: 52 ~ 1]J>2
defined by f(x, y, z) = [x, y, z]. If U is a set in 1]J>2 and [x, y, z] E U,
then f1(U) contains both (x, y, z) and ( x,  y,  z). If f1(U) is
open in 52, then we will say it is open in 1]J>2. It is a good exercise to
check that this does give a topology on 1fJl2.
A final way of looking at 1]J>2 is to note that the set
Di = {(x, y, z) I (x, y, z) E 52, Z ~ O}
is a disc. Each equivalence class [x, y, z] with z '# 0 meets D ~ in one
point, but all the [x, y, 0] meet D ~ in two diametrically opposite
points. Thus 1]J>2 can be obtained from a disc by identifying accord
ing to the rule shown in Fig. 4.32. This surface cannot be embedded
in !R
3
•
a
a
Fia.4.32
aces
We must now examine what orientability means. As our surfaces
are assumed to be triangulated we will first consider what orientation
of a triangle means.
The intuitive answer must be that there are two choices of orienta
tion, as shown in Figs. 4.33 and 4.34.
Fig. 4.33
Fig. 4.34
An orientation on a triangle induces one on its edges, as in Fig.
4.35. (Thinking of the triangle as being full of a swirling liquid may
help!)
Fig. 4.35
Two triangles are adjacent if they have a common edge. Looking
at the diagrams in Figs 4.36 and 4.37 we see that adjacent triangles
with similar orientations induce different orientations on their com
mon edge, whilst if the induced directions are the same then the
triangles are oppositely oriented.
A surface S is orientable if one can choose orientations for all
triangles in S such that adjacent triangles are similarly oriented.
Otherwise the surface is nonorientable.
Suppose someone presents us with a combinatorial surface; how
are we to find out if it is orientable or not? We clearly cannot try all
the possible combinations of orientations on the different triangles.
Combinatorial surfaces 67
Fig. 4.36
Fig. 4.37
The way to do it is to propagate an orientation starting at some
point.
Pick any triangle in S and orient it in either sense. Now go to each of
the adjacent triangles and orient them similarly (if you can). Go on to
any new triangles adjacent to those already oriented and orient them
similarly if possible and so on. Either eventually you will have given an
orientation to all the triangles and end up with an orientation of S or
at some stage you end up with a situation like that in Fig. 4.38.
There cannot exist an orientation of the new triangle similar to
Fig. 4.38
both those already chosen, so the surface must be nonorientable.
These cases are mutually exclusive.
As a simple example consider the triangulation of the Mobius band
in Fig. 4.39.
faces
Fig. 4.39
Starting with an anticlockwise orientation of 1, we propagate.
(Why is 2 oriented in a clockwise way?) It is impossible to choose an
orientation for 3 which is similar to that on adjacent triangles.
We will return later on to look in more detail at orientations.
ting and pasting
The method we will use for analysing surfaces further is called cutting
and pasting. The idea is simple. Label both sides of each edge and
then cut along the edge so as to get a collection of triangles. The fact
that the edges are labelled means that we can reconstruct the surface
by gluing along edges having the same label. The usefulness of this
is that by choosing a neat way of regluing step by step we can obtain
a convenient representation of the surface as a polygon with identi
fied edges. To be slightly more formal about this process we will give
a stepbystep description of it.
Step 1
Label and orient each edge of all triangles in S with a letter and an
arrow. Distinct edges must be labelled with distinct letters. We let
Fig. 4.40
Cutting and pasting 69
x
and
E T(S)
Fig. 4.41
T(S) be the set of all labelled triangles in S. For example, Fig. 4.40
gives Fig. 4.41.
Clearly we can reconstruct S from T(S).
Step 2
Pick any triangle To E T(S). Find a TI E T(S) which has an edge label
in common with 10. Stick 10 and TI together, matching the common
edge by letter and arrow to form a disc, PI' Find 'Xi E T(S)\{ 10, T
I
}
with one or two edge labels in common with adjacent edge labels in
the boundary of PI' Stick 'Xi to PI to get a disc P2' Continue in this
way matching 'Zi E T(S) to one or two adjacent label1ed edges of Pi to
form a disc Pi+ I until T(S) is exhausted. The final result is a polygon
with labelled oriented edges. The process can be illustrated as shown
in Figs 4.424.48.
Fig. 4.42
x
Fig. 4.43
As we always glue along a proper connected segment of the
boundary we can never get a hole. (The proof that this always gives
a disc is too technical to be included here.)
A typical polygon one might obtain would be something like that
in Fig. 4.49.
70 Surfaces
e
Fig. 4.44
Fig. 4.45
e
Fig. 4.46
Fig. 4.47
Cutting and pasting 71
Fig. 4.48
Labels can occur once or twice but never more since each label
corresponds to an edge of one or two triangles, never more.
The information stored in such a diagram tells us how to identify
edges and hence to reconstruct S. However, the key information is
contained in the letters a, b, etc., of the boundary edges of the
polygon and the order and direction in which they go; this is encoded
in the surface symbol of S.
Fig. 4.49
To work out the surface symbol of a polygon, pick any vertex (we
will use I in the diagram in Fig. 4.49); then, reading off clockwise,
write down the edge label if the arrow is pointing clockwise, or write
down that label to the power minus one if the arrow is pointing
anticlockwise. Thus in our diagram, we get
b
I
cda
I
dCI ea.
The choice of start vertex is ours to make so we might have got (by
starting at 2)
and we must therefore consider these two symbols to be 'the same',
that is equivalent. We will play with these symbols much as we would
aces
play with relations in a group presentation (see Chapter 6) or with a
knot diagram using Reidemeister moves. We must, of course, specify
the rules of the game we shall play.
Step 3: operations
(We will write capital letters for words in small letters and their
inverses.)
Operation 1 (Fig 4.50)
cut
A B C ~ AxCand Bx
1
paste
A A
Fig. 4.50
The usefulness of this operation is that one can cut along one edge
and paste back along another one, provided of course that labels and
arrows permit this. We will see this in action shortly.
Operation 2 (Fig. 4.51)
paste
A X X  1 B ~ AB
cut
Fig. 4.51
Cutting and pasting 73
Operation 3
x ~ u v
where u, v occur nowhere else in the symbol and the replacement is
at every occurrence of x.
The inverse of this is
u v ~ x
and can be done if u and v always occur together in the symbol.
(Here the order is very important: uv or VIU
I
is all right but vu or
u
I
VI does not count.)
This corresponds to splitting an edge (Fig. 4.52).
x
u v
c : ~
A A
Fig. 4.52
Remark on the operations
You may be wondering why, in operation 1, we can assume that only
a single edge is needed to join the two ends of B. We mentioned
earlier the idea of chopping or subdividing the triangulation further.
By doing this if necessary we can find a path along edges inside the
polygon, joining the vertices I and 2. Cutting along this path we have
a series of labels which always occur in the same order and so we can
apply operation 3 to them to replace them by a single edge label. This
may alter our triangulation slightly but this does not affect the
arguments or ideas we will be using later on.
Composite operations
By combining the above elementary operations we can build up some
very useful composite operations. The validity of these moves is
indicated by diagrams.
Operation 4
Replace ABxCxD by AyCB1yD, where y is a new symbol not in
A, B, C, or D (Fig. 4.53).
faces
~ ~ 0 •. 3
A L J ~ : 3 A cr::fL__ ; ...... o__ li_:....JI By
D D D
Fig. 4.53
Operation 5
Replace ABxCx
1
D by AyCy1 BD, where y is a new symbol not in
A, B, C, or D.
The proof is much as above, so try and draw the sequence of
diagrams for yourselves.
Operation 6
Replace AxBCxD by AyCyB
I
D.
Operation 7
Replace AxBCx
1
D by AyCBy1 D.
Again for these two composite operations, we leave it up to you to
draw the diagrams which check the validity of the operation. This is
an extremely useful exercise to do, especially if you do not see
immediately why one of the operations is valid.
Step 4
We call this step assembling the crosscaps. A crosscap is a projective
plane contained in S. The projective plane has symbol aa and in any
surface symbol the occurrence of a repeated letter ... a ... a ... or
aI ... aI . .. indicates the presence of a projective plane and hence
of nonorientability (more on this later). Our first task now that we
have these composite operations at hand is to get all the crosscaps
to the start of the symbol.
We note that if we have ABxCxD in a symbol, we can use
operation 4 to replace it by AyCB1yD and then operation 6 to
obtain AzzBC
1
D. (In operation 6 we take C to be empty and B to
be CB
t
.)
Now suppose we can find some recurring letter, say s, in our symbol.
Taking A to be empty, B the block of letters before the first s, and so
on, we can use the above pair of operations to replace the symbol with
one of the form ftE', that is one which starts with a crosscap.
Cutting and pasting 75
Next examine B' to see if some other letter recurs. If it does (e.g.
we might find B' = Bu CuD, for new meanings of B, C, and D), we
can apply the pair of operations, this time with A = tt to replace the
symbol with ttvv BC
I
D. Repeating this process as many times as
proves necessary, we will eventually get the symbol in the form
where B has only edges that occur singly or as ... x ... XI • .•.
Step 5
A torus has, as we have seen, a description as a rectangle with
opposite sides identified. Thus a surface symbol for the torus will be
abaI b
I
. If a surface symbol contains the pattern
... a ... b ... a_
I
... b
I
it means that there is an attached torus hidden in the surface. As
an attached torus looks like a handle, we say that the surface
contains a handle if such a pattern is contained within the symbol
(Fig. 4.54)
Fig. 4.54
The fifth step is to bring any handles together and then collect
them up just behind any crosscaps. To do this we use the following
series of operations
ABaCbDaIEbIF
Op.5
  ~ AxCbDxIBEb1F.
~ ~ AxCyBEDx1yIF
~ ~ AuBEDCyu1yIF
~ AuVUIVIBEDCF
aces
So taking A to be the string of crosscaps, B to be the block before
the first letter of the chosen handle, and so on, we can manipulate
the symbol so that it has the form
Of course B' may still contain handles, so we repeat as many times
as is necessary to get the symbol in the form
[crosscaps] [handles] [the remainder].
Step 6
You might expect us to attack 'the remainder' next; however, there
is a curious phenomenon that occurs if both crosscaps and handles
are together in a symbol. The crosscaps influence the handles,
changing each handle to a pair of crosscaps, that is
crosscap + handle = 3 crosscaps.
The sequence of moves necessary to check this is
aahch I c I reverse) hdhc I dc I
ofOp.8
followed by the collection process of the crosscaps as described
previously. It is a valuable exercise to write down all the details to
check that aabcb
I
(,I can be replaced by xxyyzz.
Using this as many times as need be, we can reduce any symbol
either to
alai' ... a"a"H as soon as any ... a ... a ... occurs
or otherwise
b
l
C
I
b"ll e"ll ... hmcmb;,/ c;',' H if no ... a ... a ... occurs.
In both cases H contains no crosscaps and no handles. (It may be
empty in which case we stop.)
Step 7: assembling the cuffs
We now need to look at H which we shall assume is nonempty. One
obvious form we might expect to find in H is a cuff corresponding
to CdCI (Fig. 4.55).
Is this the only form within H? To find out we suppose x occurs
twice in H; thus we have either ... x . .. xlor ... XI ... x ....
Choose such an x with as few letters between x and x I as possible.
Then the letters occurring between x and x I can occur nowhere else.
Cutting and pasting 77
b'
Fig. 4.55
(Why? Remember there are no crosscaps or handles in H.) We can
therefore replace (using operation 3) this string of letters from x to
x' by a single letter
... x ... XI ... ... xbx' ....
Then H becomes Axbx
'
B which in turn can be changed to yzyl AB
(by which operations?).
Continuing in this way we can replace H by
y, ZI Yi' Y2Z2y;1 .... I W.
Now W may be empty in which case we stop and breathe a sigh of relief.
If, however, W is not empty, it must consist of single Occurrences of
letters. These can be amalgamated using operation 3, being replaced by
a single letter, w say, and will eventually give us another cuff. At the
moment, however, it is more a hole than a cuff. To make another cuff
we use operation 2 to add u 1 u at the end of the symbol; thus we now
have A H wu I u which, since cyclic permutations of the letters make no
difference (corresponding to a change of starting vertex), we can replace
by uAHwu
l
. Now A consists of crosscaps or handles, whilst H consists
of cuffs. If a
l
a
l
is the first crosscap in A, we get
, 1 ' n "
(using which operations?). Repeating as many times as is necessary
we can pass u through the crosscaps. If A consists of handles, we note
that ubcb
I
c
I
can be replaced by fgflgI U (again we leave it up to
you to work out why), so that we may pass u through all the handles.
Finally we use that uyzy
I
can be replaced by vzvlu to show that
u can be passed through all the cuffs. Thus from uAHwu
1
we obtain
AHuwu
'
, that is another cuff as promised. We have proved the
following proposition.
Surfaces
Proposition. Any surface symbol can be transformed by the operations I, 2,
and 3 to one of the form AH where A is a string of crosscaps or handles
(not both) and H is a string of cuffs. (This form is called the normal
form of the surface symbol.)
If any two surfaces S) and S2 yield surface symbols with the same
normal form then as their polygons are homeomorphic by a homeo
morphism which recognizes the labels and orientations of each edge
(i.e. the instructions for gluing the polygon to get the surface), the
two surfaces must themselves be homeomorphic. As the converse is
clearly surfaces have symbols with the same
normal get the following first form of the classification
theorem.
The classification theorem for surfaces. Two surfaces are homeomor
phic if and only if the normal forms of any symbols representing them
have the same number of crosscaps/or handles and also the same
number of cuffs.
mples of normal form
(a) 0 the empty normal form. In this form it is hard to visualize,
but using operation 2 we can replace it by aa
I
• This is a sphere.
a
z
V
Fig. 4.56
Cone
Fig. 4.57
Sphere with
one hole
Cutting and pasting 79
Fig. 4.58
(c) aba'b'. As we have already seen, this is a torus (Fig. 4.58).
(d) aa is the projective plane, jp>2
(e) a,a,a
2
a2 two crosscaps = Klein bottle.
This is not obvious, but look at Fig. 4.59.
a
a
0:;;;;::;: •
a
b b
Fig. 4.59
4.2
1. Classify the symbol b'cda'dc'ea.
a
2. The Mobius band can be represented as shown in Fig. 4.60. Find
the normal form of its surface symbol abac.
b
a
c
Fig. 4.60
3. Find the normal form of the following symbols:
(a) a,a2a3aj'a2Ia31
a
Surfaces
(c)
(d)
When you have done a few of these calculations and you can
imagine the tedium of performing all the operations step by step on
some long surface symbol, you will probably find yourself asking: is
there a simpler way of doing this? Happily the answer is yes. The use
of invariants of surface symbols allows the normal form to be found
very quickly and easily, so let us get on with this straight away.
Invariants of surfaces: orientability and boundary curves
If we choose a triangulation, we can get a surface symbol and thus
eventually a normal form. We can use the triangulation or the
surface symbol to obtain certain information on the surface, but any
information we thus obtain should not depend on the choices we
have made. For instance, the number of letters that occur in a
surface symbol is clearly no good as a measure of the geometric
properties of the surface since if we have a disc, we can subdivide it
in many different ways to get different numbers of edges, but
topologically it is still a disc.
ntability revisited
The first invariant property is one we have already considered
briefly, namely orientability.
Proposition. A surface S is nonorientable if and only if the normal form of
a surface symbol for S contains a crosscap. This occurs if and only if
any symbol for S contains a crosscap.
Proof Suppose in a symbol for S we have a crosscap, so somewhere
in S we have the pattern ... x ... x .... Choose a path within the
polygon between these two occurrences of x, passing through
the edges of triangles and through each triangle at most once. Orient
successive triangles along this path similarly; one finds that the two
triangles that share the edge x are oppositely oriented (Fig. 4.61), so
the surface must be nonorientable.
If, on the other hand, no such repeated x occurs, then as the
operations we have used can neither create nor destroy orientability
we find no crosscaps in the normal form. However, we can easily
construct an orientation for any surface whose normal form has only
Invariants of surfaces: orientability and boundary curves 81
Fig. 4.61
handles or cuffs. (Try some simple examples such as a torus and a
torus with one cuff etc., to see how you might find an orientation in
general.)
ber of boundary curves
In the normal form it is clear that the number of cuffs is an
important source of information. This number is precisely the number
of edges in the normal form that occur exactly once. Unfortunately
because operation 3 allows us to subdivide an edge, the number of
edges that occur exactly once in a symbol may not be the same as
the number of cuffs in the final normal form. For instance, if the
symbol is abaI b
I
cdec
I
, the number of edges that occur exactly once
is two, but we clearly manufactured this by using operation 3 on the
normal form abaI b
I
cfc
1
• With this example it was easy to guess
what the normal form was, but if the symbol had been all mixed up
using others of the operations this might not have been so easy. How
can we attack this in general?
We start by finding out how many distinct vertices there are
around the boundary of the polygon. For instance, in the torus
symbol aba
1
b1, all the edges start and end at the same vertex, but
in aba I b
1
cf c
1
one needs two vertices as we shall see. Looking at
the polygon of the symbol, we start by choosing one edge and
labelling one end, say
a
)
Q
If the label a occurs twice on the boundary of the polygon, we
insert Q as a label at the appropriate end of the second edge. The
next edge to a, say b, now starts or finishes at Q, so we must insert
Q again, that is
aces
Q
I
b
•
or
a
•
Q
I
b
..
wherever b reoccurs (if indeed it does). Often this will insert Q at at
least one new position allowing us to repeat the process until no
further insertions of Q are called for. Now label any unlabelled end
of an edge with a new label R, say, and insert R wherever required.
Continue until all vertices on the boundary of the polygon are
labelled.
As a simple example look at a surface with two crosscaps and a
cuff. We will not draw the polygon, merely work with the symbol.
We will show intermediate stages:
a a b b CdCI
Q
Put Q as the label of the end of a; this occurs twice
a a
b b c d c
1
Q Q
so a starts as well as finishes at Q, b starts at Q, and c
1
ends at Q
(since the symbol as a whole must start and finish at the same
vertex). This means c starts at Q and at this stage we get
a a b b
Q Q
Q
Q Q Q.
No more Q can be fitted in, so we insert a label R as the end of c
(and thus the start of c
I
). This gives
a a b b c d
Q Q Q Q Q
R R
Q
so there are two distinct vertices on the boundary of the polygon.
An edge whose label occurs just once in a symbol is called a
boundary edge. Earlier we looked at the example of a polygon with
symbol b
I
cda
I
dc
1
ea; this has two boundary edges, band e. If we
work out the vertex labels on this example we find
b
I
C d
aI
d
c
1
e a
Q R Q Q Q Q R Q Q
so we find that b and e share vertices. They fit together to gIve a
boundary curve:
b e
..
•
Q
R Q
Invariants of surfaces: orientability and boundary curves 83
In general, we join up each string of boundary edges to make
boundary curves, but there may be more than one of these. As we
hinted at earlier, the number of the boundary curves does not depend
on the triangulation we chose.
Proposition. The number of boundary curves is the same as the number of
cuffs in the normal form for the surface symbol. /t is therefore an
invariant of the surface.
Proof. We will show that the number of boundary curves is
unchanged by the processes that the symbol goes through as it is
transformed into its normal form. As the number of boundary curves
in the normal form is easily checked to be the number of cuffs (do
check this), this will prove the result.
Firstly we note that the choice of starting vertex and direction in
the reading of the surface symbol does not affect the boundary edges.
Now consider what happens under each of the elementary operations
in turn.
Operation 1
AxCand Bx
I
•
The new label x occurs twice, so this does not alter the number of
boundary edges or curves.
Operation 2
Again it is clear that as x occurs twice, no change occurs in the
number of boundary curves.
Operation 3
If x is not a boundary edge then no change occurs. If, however, x is
a boundary edge, the number of boundary edges will increase by one,
but the number of boundary curves will stay the same.
cises 4.3
1. Find the number of distinct vertices for a polygon with symbol
having g crosscaps or handles and r cuffs.
2. How many vertices are needed for the symbol
abc
l
adbc
1
efd
l
?
Surfaces
Invariants of surfaces: Euler characteristic and genus
Suppose S is a (combinatorial) surface with a
o
distinct vertices, a]
distinct edges, and a
2
triangles. The Euler characteristic X(S) is given
by
The usefulness of this number in various contexts and guises will be
one of the themes of the coming chapters. Its usefulness here is that
it does not depend on the triangulation used and is an invariant of
the surface. This is perhaps a bit of a surprise, until you see why it
happens.
Proposition. Let P be a polygon representing a surface S. Suppose that the
surface symbol of S gives v distinct vertices and e distinct edges on the
boundary of P; then X(S) = v  e + 1.
Proof. We will prove this by induction on the number, k, of
triangles making up P. If k = 1, then either there are no
identifications, so S = P, a
o
= v, a] = e, and OQ. = 1, or we have a cuff
and again a
o
= v, a] = e, and a
2
= 1. In this case it is clearly true.
Now suppose k > 1 and that we know the result as long as P has
fewer than k triangles in it. Recall that P was constructed by adding
triangles in one of two ways to the disc already constructed. The two
ways were as shown in Figs 4.62 and 4.63.
y
P'
Fig. 4.62
We handle the two cases separately and each will lead to some
'subcases' .
(A) We first look at what happens if a and b occur nowhere else in
the symbol; then the vertex between a and b is a new vertex. On
recording the change in v and e from P' to P, we find v = v' + 1, one
Invariants of surfaces: Euler characteristic and genus 85
Fig. 4.63
new vertex, and e = e' + I since although we have gained a and b we
have lost y from the boundary. Thus v  e + 1 = v'  e' + 1 and so
stays unchanged. Now P' represents some surface, say S', and we
have, by hypothesis, that since P' has only k  I triangles,
x(S') = a
o
 ~ + a; = v'  e' + 1.
We can, however, work out a
o
, aj, Gl2 from do, ~ , a;, the corres
ponding numbers for S'.
We had one new vertex, so a
o
= a
o
+ 1; two new edges, a and b, so
a] = ~ + 2; and one new triangle, so Gl2 = a; + 1. Thus
x(S) = (a
o
+ I)  ( ~ + 2) + (a; + 1)
that is, X(S) = X(S') = v'  e' + 1 = v  e + I, as required.
The other subcases are similar. We will use the same notation but
will explain less along the way, leaving the details for you to check.
Suppose now that a is not a label on the boundary of P' but b is;
then v = v', e = e' (a is in, y is out!), so v  e + 1 = v'  e' + 1 as
before. Also we have a
o
= a
o
, a] = ~ + 1, Gl2 = a; + 1, so XeS) = xeS')
and we complete the argument as before.
If a is a label on the boundary of P' and b is not, then the same
argument applies.
Finally if both a and b occur on the boundary of P', then v = v',
e = e'  1, a
o
= do , a] = a; , a
2
= a; + 1.
We now look at case (B). Here there are only two subcases: either
a is in the boundary of P' or it is not. If it is, then we get a] = a; ,
a2 = a; + 1, whilst v = v'  1, e = e'  2. The equations balance as
before: since by hypothesis v'  e' + 1 = ao  ~ + a2, one gets
v  e + I = ao  a] + a2' If a is a new edge, not in the boundary of
P', then we get ao=ao, a,=a;+I, Gl2=a;+I, whilst v=v'l,
e = e'  I which again balances the equations. •
Surfaces
The next stage in checking that this Euler characteristic is an
invariant is to see what happens under the operations.
Lemma. The transition to normal form does not change the values of the
expression v  e + I determined by vertex and edge labels.
Proof. (As the detailed proof is a bit like that of the invariance of
the number of boundary curves, we will leave the details up to you
(but do write them down!). We will provide just a sketch.)
Both operations 2 and 3 give us one new vertex label and one new
edge label, so v  e + 1 stays the same.
Any application of operation I is followed by one of the type of its
inverse. This reglues an edge, reassembling the polygon in a new way,
but overall the number of vertex and edge labels stays constant. •
We have shown that X(S) is independent of the triangulation and
can be calculated from the polygon in its normal form. We now need
only to work out X(S) for the various normal forms to have complete
information.
Type: Nonorientable, g crosscaps, r cuffs.
Normal form: alai'" agagylzly;l . .. Yrzry;l.
From the exercise earlier in this section, we have:
number of vertex labels
number of edge labels
hence Euler characteristic =
r + I
2r + g
2  g  r.
Type: Orientable, g handles, r cuffs.
Normal form: a1b1ailb;I ... agbgaglbglYIZIYi! ... Yrzry;l.
We get (again from the exercise):
number of vertex labels
number of edge labels
hence Eu ler characteristic =
r + I
2r + 2g
2  2g  r.1
We have seen that (i) orientability, (ii) number of boundary curves,
and (iii) Euler characteristic are not only invariants of the surface,
but also able to be calculated from any symbol for the surface. We
can therefore find rand X(S) and by choosing either the orientable
or nonorientable formula from the above, calculate g, the number
Invariants of surfaces: Euler characteristic and genus 87
of crosscaps or handles. This number, g, is cal1ed the genus of the
surface. We have:
if Sis nonorientable, g = 2  r  X
if S is orientable, g = 4 (2  r  X)·
We summarize the theoretical aspect of this as follows.
classification theorem for surfaces. Two surfaces are homeomorphic
if and only if they are both orientable or both nonorientable. have the
same number of boundary curves and the same genus.
The practical aspect of the result is clear: it saves us work! The
method of cutting and pasting does lead eventually to the normal
form but the above ideas allow us to classify surfaces quickly and
painlessly. As an example, we will look again at our symbol
b1cda1dclea.
Orientability: Nonorientable because of .. d . .. d ..
Vertex labels:
c d d e a
Q
p
Q Q Q Q
p
Q Q
Boundary edges: band e giving one boundary curve, so r = 1.
Genus: First calculate XeS) = v  e + 1 = 2  5 + 1 =  2,
so g(S) = 3. Thus S has normal
form alala2a2a]a3yzyl.
s 4.4
I. Classify the following surfaces giving genus, Euler characteristic,
orientability, etc.
(a) ablcdad1ceb
(b) eabc1adbclef
(c) ab1cdald1c1eb
2. Given two surfaces Sl and S2, their connected sum, denoted
Sl # Sb is formed by cutting a small circular hole in each surface
and then gluing the two surfaces together along the boundary of
the holes (Fig. 4.64).
If Al is a symbol for SI and A2 one for Sz. one can obtain a symbol
for S = Sl # S2 by eliminating y from the pair of symbols Axyx 1 and
AzyzI, (Fig. 4.65) that is by replacing the 'holes' by cuffs.
aces
Holes cut out
J. Glue
Fig. 4.64
Glue
Fig. 4.65
(a) Find XeS) in terms of X(Sl) and X(S2)'
(b) Find g(S) in terms of g(SJ) and g(S2).
(Note you need to take care if one surface is nonorientable whilst
the other is orientable.)
(c) Prove that !p
2
# torus is the same as !p
2
# !p
2
# !P
2
•
The classification theorem can be rephrased in terms of connected
sums since a genus g surface with r cuffs is either the connected sum
of g tori (if orientable) or g projective planes (if nonorientable)
together with r cuffs.
3. Let S be a surface without boundary curves (so no cuffs appear
in the normal form). Suppose we have a triangulation of S using
ao vertices, a
J
edges, and a
2
triangular faces. Show that
Invariants of surfaces: Euler characteristic and genus 89
(b) al = 3(ao  XeS»)
Now by considering the maximal possible number of edges joining
a
o
vertices, show that
(c) ao ~ t (7 + ...J49  24X(S) )
These calculations give a lower bound for the number of vertices
needed to triangulate S. Find a lower bound for the number of
triangles in a triangulation of (i) the projective plane, (ii) the Klein
bottle, (iii) the sphere. Can you find a triangulation with exactly this
number of triangles?
Fig. 4.66
4. Let SI and S2 be two homeomorphic orient able surfaces of genus
g and with r > 0 boundary curves. Construct a new surface, D,
by identifying the boundary edges of S] and S2; this is the double
of SI (or S2). You might visualize it as being constructed by
replacing each triangle (Fig. 4.66) in a surface by two copies near
to each other (Fig. 4.67) except if an edge is a boundary edge
(Fig. 4.68).
Fig. 4.67
Fig. 4.68
Surfaces
Show that D is orientable, has genus 2g + r  1, Euler charac
teristic 4  4g  2r, and no boundary curves.
Knots and surfaces
So far the only connections between knots and surfaces that we have
met in this chapter have been ones of methodology: ideas such as the
use of invariants, the reduction of classification problems to moves, and
so on. In the next few pages we will see how these two classes of object
interact at a deeper level and hence that some parts of the classification
result for surfaces can be used to produce results in knot theory. In the
next chapter these will be used in a study of the 'arithmetic' of knots.
Our main aim here is to prove the following result of Seifert.
Theorem. Given any knot K (in 1R
3
) there is an orientable surface S with
one boundary curve embedded in 1R3 in such a way that the boundary
of S is the knot K.
Proof. Take a projection of the knot, K. Orient the diagram then
at each crossing, add two extra directed arcs bypassing the crossing,
but compatible with the orientation, as shown in Fig. 4.69.
Fig. 4.69
Now delete the crossing leaving a set of oriented circles, called
Seifert circles. (In our crossing we throwaway the parts of the knot
between XI and Y2 and between YI and X2')
As an illustration, consider the process outlined above performed
on the figureeight in its usual projection, as in Fig. 4.70.
The Seifert circles mayor may not be nested. Starting from any
innermost circle and working outwards fill in (the word 'span' is
often used) each circle with a disc, the discs to be disjoint.
Knots and surfaces 91
.. 0 ..
, /
\ ...
Remove the
:>
()
crossings
Three Seifert circles
Fig. 4.70
x ,
Y2
I
I
Y 1.
I
I
of
x
2 Y ,
Fig. 4.71
The two arcs and the parts of the crossing removed make up a
rectangle as in Fig. 4.71.
Fill this in with a solid rectangle. Do this at each crossing. Thus
the crossings now look as in Fig. 4.72, each one filled with a twisted
Fig. 4.72
rfaces
Fig. 4.73
rectangle. Each spanning disc of each Seifert circle has a natural
induced orientation coming from the orientation of the knot. As each
twisted rectangle (Fig. 4.73) can be oriented coherently with these
orientations, the surface obtained by gluing the twisted rectangles
onto the Seifert circles will be orientable.
position. The surface, S, constructed from K by the above method, has
genus
g = ~ (d  s + 1)
where d is the number of crossings in K and s the number of Seifert
circles arising from K.
Proof. A disc has Euler characteristic 1. If we add a rectangle
between two discs (Fig. 4.74) then the number of vertices, ao, does
not change since there are no new vertices; the number of edges, ai'
increases by three; and the number of triangles, lX:l, increases by two.
Disc
Disc
Fig. 4.74
We start with s discs, each contributing 1 to %(S), and then add d
rectangles, each reducing a
o
 a
l
+ ~ by 1; hence %(S) = s  d and as
S is orientable g(S) = 112 (d  s + 1). •
You might expect this number to be an invariant of the knot;
however, a different projection might give a different surface and
hence a different g. To avoid these problems we take the easy way
Knots and surfaces 93
out and make the following definition: the genus g(K) of a knot K is
the minimal genus of all orientable surfaces which span K.
The genus thus defined gives an invariant of the knot, but the form
of definition makes it difficult to calculate. Since the Seifert construc
tion does not necessarily give a minimal genus surface, we can only
say that
g(K) ~ t (d  s + I).
A theoretical algorithm was developed by W. Haken in 1961 but
at the time it was too unwieldy to be useful in practice. With more
powerful computing methods and with the renewed interest in knot
theory due both to the development of the new polynomials and the
perception of potential for applications, the question of developing
methods for calculating the genus of knots is again receiving a lot of
attention.
It is useful to note the following:
I. It is quite easy to span certain knots with nonorientable sur
faces. For example, the trefoil can be spanned by a Mobius band
(Fig. 4.75), so BEWARE!
Fig. 4.75
2. If g(K) = 0, then K is the trivial knot.
3. If K is known to be nontrivial and one finds some orientable
surface of genus 1 spanning the knot then g(K) = 1. (In our
example, we looked at the figureeight knot having four crossings
and three Seifert circles, so the surface that resulted did have
minimum genus 1.)
4. If g(K)::t= 0, 1, it can be extremely difficult to calculate the
genus. Can one be certain that no smaller genus surface spans
the knot?
rfaces
5. There is a strange link with the Alexander polynomial. If K is an
alternating knot (i.e. there is some projection in which, as one
goes along the knot, overpasses alternate with underpasses), then
g(K) is half the degree of the Alexander polynomial of K. In
general, the degree of .d(t) is less than or equal to 2g(K) (see
Burde and Zieschang 1985).
es 4.5
Find bounds for the genus of the following knots:
(a) trefoil (Fig. 4.76)
Fig. 4.76
(b) cinquefoil (Fig. 4.77)
(c) any (2, n)torus knot (Fig. 4.78)
(d) any generalized figureeight knot (Fig. 4.79)
Fig. 4.77
Knots and surfaces 95
n crossings
n odd
Fig. 4.78
Fig. 4.79
5
The arithmetic of knots
Analogy is a very powerful logical tool. When two different types of
situation seem to be running parallel from certain viewpoints, one
can try to see if the similar behaviour extends further to give new
insight into the different structures. Of course it is as important to
notice where the analogy fails as to follow those situations where it
works. A basic analogy used in knot theory is that the sum of
oriented knots behaves very much like the product of natural
numbers; hence the title of this chapter.
sum of oriented knots
Suppose that we have two oriented knots K and L. Then we can
form their 'sum' in an obvious way following the natural intuition of
first tying K and then L in a piece of string. The details are as
follows.
Cut each knot as shown in Fig. 5.1.
,
K
L '
Fig. 5.1
s
The sum of oriented knots 97
Fig. 5.2
Join them together preserving orientation, as in Fig. 5.2.
We wiH write K + L for the result.
(Figs 5.3 and 5.4)
granny knot = trefoil + mirror trefoil
Granny knot
Fig. 5.3
How does this operation + act on the class of all knots? Does
it respect equivalence for instance? Does it depend on where the cuts
are made? Is + an appropriate choice of symbol? Does + behave like
the addition of integers? The following goes some way to answer
ing these questions. We shaH denote equivalence of knots by the
symbol .
The arithmetic of knots
reef knot = trefoil + trefoil
Reef knot
Fig. 5.4
Proposition
(a) If K  K' then for any L, K + L  K' + L
(b) K + L  L + K
(c) (K + L) + H  K + (L + H)
(d) K + trivial knot  K.
The full proofs of (a) and (c) are long. For (a), you have to use the
homeomorphism sending K onto K' to construct that sending
K + L onto K' + L. This uses details of the behaviour of homeo
morphisms that are deeper than we have yet gone, so we will omit
the proof. For (c), the fact itself is unsurprising if not obvious, but
again one has to construct things with great care so as to be able
explicitly to produce the homeomorphisms. Again we omit the proof.
For (d), a sufficient reason to believe the result is as shown in Fig.
5.5.
Fig. 5.5
Finally, for (b), the result seems a lot less obvious than any of the
others; however, the reason for this property is very pretty and easy
to seeonce it is pointed out to you. A full proof can be based on
this 'sketch' proof, Or rather 'proof by sketches'. We illustrate the
idea in Fig. 5.6.
Make K smaller and slide it along L as in Figs 5.7 and 5.8.
Now pull K back to the size it was to start with (Fig. 5.9).
The sum of oriented knots 99
K L
K+L &1)
Fig. 5.6
Fig. 5.7
Fig. 5.8
= L + K
Fig. 5.9
The arithmetic of knots
These results then show that the operation + on knot types is well
defined, commutative, associative, and has an identity which we
write 0, a very suitable symbol when one remembers that the identity
is the type of the unknot.
So far we have developed this knot sum showing that it behaves
much like addition of natural numbers. It is necessary to pause here
a moment and consider a couple of points. Firstly that there are
many other systems which behave as above. (These structures are
called commutative monoids.) As this is the case we must not jump
to the conclusion that this 'analogy' is in any way complete. For
instance, we do not yet know if it is possible to subtract knots; that
is, if we have K + L is there some L' such that K + L + L'  K? What
about cancellation: if K + L  K' + L, is K  K'?
Although we have as yet few invariants for knots (bridge number,
genus, colouring properties, various polynomials) we can hope
that investigation of how the value of an invariant on K + L depends
on its values on K and L will at the very least shed some light on
these problems and perhaps indicate new properties to investigate
further.
ienus of K + L
Suppose we have surfaces SK and SL spanning K and L respect
ively and such that g(SK) = g(K) and g(SL) = g(L) that is they are
minimal genus spanning surfaces. If we place SK and SL in different
halfspaces, we can 'pipe' them with a rectangle as shown in
Fig. 5.10.
1t
L
I
'Piping'
rectangle
Fia.5.10
Genus of K + L 101
The result will be an orientable surface S with a single knotted
boundary component. That boundary component forms a copy of
K + L, that is S spans K + L and so
g(S) ~ g(K + L).
We can calculate g(S) in terms of g(K) and geL). We leave the
justification to you; the equation to verify is
so
g(K + L) ~ g(K) + geL).
This simple argument thus suggests that maybe g(K + L) is
g(K) + geL). We will investigate this relationship more closely, trying
to see if that is the case or if sO,me as yet unknown feature comes in
to block equality.
We suppose K and L as before are on different' sides of some plane
1r with the joining arcs crossing cleanly, that is each intersecting 1r in
a single point, as in Fig. 5.11, so 1r n (K + L) = {A, B}, say.
A
B
It
Fig. 5.11
We have some oriented surface S spanning K + L and having genus
exactly g(K + L) (Fig. 5.12).
Shifting S slightly if necessary, we can assume that it also cuts 1r
cleanly, that is that S n 1r consists of a finite number of disjoint
circles and an arc joining A to B (Fig. 5.13). (The proof that this can
be done requires a study of transversality. If you like this type of
argument and want to know more, we suggest you look up Lang
(1972).)
Some of these circles may be nested. Choose a circle that does not
loop around the arc from A to B and, moreover, does not surround
any smaller loop (Fig. 5.14).
e arithmetic of knots
,
' ~  +    = 
Fig. 5.12
A
(
B
o
o
@)
Fig. 5.13
s
Put a cap on this circle and pull it away from 1C to get a new surface
Snew still spanning K + L. This process may have disconnected the
surface; if so throwaway the part not containing K + L in its
boundary. The new surface has the same genus as our original S
s
:JC
1t
Fig. 5.14
Genus of K + L 103
since Snew still spans K + L, Snew has genus at most that of Snew (since
our 'surgery' above can at best have chopped through a handle or
have resulted in some handles in a thrownaway piece), but S has
minimal genus among spanning surfaces of K + L, that IS
g(Snew) = g(S) = g(K + L).
We repeat this process until we are left only with circles which loop
around A and B. Starting with the largest such circle attach discs
around the backs of K and L. Continuing in this way we get a
minimal genus surface S' spanning K + L such that S' n 1r is just the
arc joining A to B. Cut along this arc to obtain two surfaces SI and
S2 spanning K and L respectively. We have
Combining this with the result of our earlier work gives a proof of
the following theorem.
rem (sum theorem for genus). If K and L are oriented knots, then
g(K + L) = g(K) + geL).
ge number of K + L
The above discussion on g(K + L) leant heavily on the minimality of
the genus of the chosen spanning surface. Bridge number was also
defined using a minimum over all possible positions. Can a similar
type of argument to that used in Section 5.2 give us a similar result
for b(K + L) in terms of b(K) and beL)?
Our first step is to use guesswork: is b(K + L) equal to b(K) + b(L)?
The obvious first reaction is 'probably', but if both K and L are the
unknot then b(K) = beL) = b(K + L) = I, so this first guess is not
right. The reason is not hard to see.
Suppose we take a diagram of K and a diagram of L, each having
the minimal number of bridges for the respective knot. Now compose
the knots as in Fig. 5.15.
The result would seem to have b(K) + beL) bridges; but is this the
smallest possible number?
By choosing the points where K and L are cut we can reduce the
number of bridges by one. Pick the cutting point A in K to
be between the first underpass and the first overpass and in L to be
between the last overpass and the second underpass, as in Fig. 5. 16;
then the second arc thus created contains no overpass and hence
no. of bridges in diagram = b(K) + beL)  1.
arithmetic of knots
Fig. 5.15
Fig. 5.16
It might be possible to reduce the number of bridges yet more by
some 'cunning trick', so we get
b(K + L) ~ b(K) + b(L)  1.
Our experience with the genus sum formula suggests that perhaps
b(K + L) = b(K) + b(L)  1.
We note that if L is trivial, since b(L) = I, then this would be
consistent (giving b(K) = b(K) + I  I). This formula is correct but
we will not attempt to prove the result here.
Factorizations, prime knots, and the factorization theorem 105
rem (sum formulae for bridge number). If K and L are oriented
knots, then
b(K + L) = b(K) + b(L)  1,
If you can read German, you could look up the original proof. It is
in Schubert (1956).
orizations, prime knots, and the factorization theorem
We saw how the sum of knots was well behaved, much like the sum
of numbers. Is there a 'subtraction' of knots? Let us suppose we
could find a ' K' so that K + ( K) = 0; then
g(K) + g(  K) = 0
and g(K) =  g(K). That is strange: how could a knot have negative
genus? So the only knot K for which  K exists in this sense is
K = 0 itself.
We started out this chapter talking about analogy. Our attempt to
make an analogy between the sum of knots and the sum of numbers
only goes so far. There is no 'additive inverse', just as there is no
multiplicative inverse in the natural numbers. Perhaps the better
analogy would be between the sum of knots and multiplication (not
addition) on N. Both are commutative, associative, and have a
neutral element (l for N, 0 for knots) and we have just seen that +
on knots does not have an inverse operation.
This suggests that an analogy between the factorization of knots
and the factorization of numbers would be worth investigation. If we
could fracture a complicated knot into simpler factors then presum
ably knowledge of these simpler factors should tell us everything
about the original knot. This would be the case provided certain
conditions on this factorization process were satisfied. Ideally we
could hope that the fracturing process would terminate just as
18 = 3 x 6 = 2 x 3 x 3 cannot go further without introducing trivial
factors. The reasons that such termination is important are clear. It
is not much good if one can go on and on and never know if one
more step would give a complete answer. We can also hope that a
complete answer is unique. The halfway stage of our fracturing of
18 was given as 3 x 6 but we could equally well have given 2 x 9. The
uniqueness of the 'prime factorization" 2 x 3 x 3, is what makes it
useful. What evidence can we get for 'primes' among knots?
Suppose K is a knot with genus I; can we fracture it, say as
K
J
+ K2? We would have
e arithmetic of knots
so as any genus is nonnegative, one of the g(K
j
} must be zero, so
any genus I knot must be 'prime'. We can therefore list quite a few
'prime' knots starting with the trefoil and the figureeight.
We could equally well use the bridge number formula. We know
that b(K) = 1 means that K  O. What can we say about knots with
bridge number equal to 2? Are these prime?
If b(K) = 2 and K = K
J
+ K
2
, then
2 = b(K) = b(K
j
) + b(K2)  I
so b(Kil + b(K2) = 3. The only way this can happen is if one of
b(K
J
) and b(K
2
) is 1, that is the corresponding K
j
 O. We therefore
have that all 2bridge knots are prime. Either argument shows that
the trefoil is prime.
It is about time that we gave a formal definition of prime knot as
we have been happily (?) discussing them for a while now.
A knot K is said to be prime if K is not trivial and whenever there
are knots K
j
and K2 so that K  K
j
+ K2 one at least of K
j
and K2 is
the unknot.
In studying decompositions of knots we earlier mentioned termina
tion. The first result guarantees this.
rna. A knot cannot be indefinitely factorized. More precisely, given a
knot K, there is a natural number N such that if K  K
j
+ ... + Knfor
some n > N then some of the K
j
are trivial. (In fact N = g(K) will do,
but need not be 'best possible'.)
Proof. Suppose K  K ~ + K;, both K ~ and K; nontrivial; then as
g(K;) ~ I for i= 1,2
g(K) > g(K;), i = 1,2.
The lemma follows since if K  K
J
+ ... + Kn and all g(K
j
) ~ 1 then,
of course, g(K) ~ n. •
The second result we note is a 'uniqueness' result.
rna. If K is prime then KI L + M implies KI Lor KI AI (where we write
K I L etc. to mean K is a factor of L.)
We only sketch the proof giving an idea of why the result is true.
There is a plane n such that n u (L + M) consists of two points only,
L lies on one side of n, and M on the other. The 'knotted part' of K
ncolourability 107
must also lie totally on one side or the other of 1C since otherwise it
would be simple to decompose K as a sum in a nontrivial way. Thus
K is a factor of either L or M as required.
rem (prime factorization of knots). If K is a knot, then there is a
decomposition
where each Pi is a prime knot. Such a decomposition is unique up to
order.
Proof (instructions only). Really this should be left up to you to do;
however, here are instructions for writing out a proof. (Do follow
them.)
First look up some proofs of the prime factorization of integers.
Most will proceed by saying that if n is not prime then there are
nb n
2
such that n = n
1
n
2
and then will go on splitting n
1
and n
2
•
Termination and uniqueness use ideas similar to Lemmas 5.4 and 5.5
above. Find a proof which is simple and adapt it to handle knots.
(The moral of the story would seem to be something like: if your
analogy is a good one, the underlying logic of the situations will be
the same. Certain parts of mathematics study exactly the logic of such
situations. If you enjoy a challenge, see if you can formulate a system
or structure which contains just enough to be able to prove a prime
factorization theorem. The idea of the lattice of divisors of a number
may set you on the right lines. Do the 'divisors' or 'factors' of a knot
also form a lattice? Are these lattices special in any way? The theory
is known, but you should be able to make some headway towards
finding a solution. Have fun! If you have not met the notion of a
lattice, it is a special type of ordered set. Try to find the definition.)
lourability
We have seen that the genus and bridge number are 'well behaved'
with regard to the sum of knots. What about other invariants,
especially the 'elementary' ones? The easiest to handle is ncolouring.
If K is ncolourable then suppose in forming a sum with L, we cut
K on an arc coloured by the colour a. How is one to colour K + L?
Simply use the colour a for all parts of K + L that were originally in
L and check this works.
In some ways, this shows that ncolourability is limited in its
usefulness as a property. It can be useful for prime knots, but for
composite knots ncolourability is rather a weak property.
The arithmetic of knots
What if K is mcolourable and L ncolourable? Is it possible to
combine the two indices of colourability in some way? The answer is
known but we will not give it here. We suggest that you try some
simple examples, put forward a conjecture, and then attempt to
prove it. Of course your conjecture may be false!
Polynomials and the sum of knots
We earlier met three polynomial invariants. How do these react with
respect to the sum of knots?
ander polynomials
We take K and L and label the diagrams in the usual way. We take
the same face label, a, for the unbounded region of the diagrams in
Fig. 5.17.
a
Fig. 5.17
When we have drawn up the m x (m + 2) matrix for K we delete the
columns corresponding to the adjacent regions, x and a, thus getting
a square m x m matrix M
K
• The Alexander polynomial is det(M
K
)
after normalization. We do the similar process for L, deleting the
columns corresponding to Yand a, with Lidt) = M
L
• The diagram for
K + L will look like that in Fig. 5.18.
This time in the (m + n + 2) x (m + n) matrix we delete columns
corresponding to z and a. It should be clear that
Fig. 5.18
Polynomials and the sum of knots 109
since no crossing originally in K can contribute nontrivially to a
column in the Lpart of MK+L since the only possibility of such a
contribution would have been via column a or z and these have been
deleted.
Of course det( MK+ L) = det( M K)· (detML). that is
d
K
+ dt) = dK(t) • dL(t).
In many ways the Alexander polynomial serves as a testing ground
for properties of the more recently defined polynomials, so we may
conjecture that these react similarly.
fly polynomial
To prove that the Homfly polynomial of a sum of knots is the
product of those of the constituent parts we need to prove one or
two preliminary results.
If K and L are two links, we let K.ll L be their distant (i.e. disjoint,
unlinked) union; then
P(K.llL) =  e + e
1
• P(K). peL).
m
Fig. 5.19
If L is the unknot this comes from the cases shown in Fig. 5.19 since
K + L+  K  K + L_, whilst K.ll L  Lo. Feeding these into the defin
ing equation for P gives
so
e P(K) + e I P(K) + mP(K.ll L) = 0
P(K.llL) = _ e + rl . P(K)
m
arithmetic of knots
If L is more complicated, then use induction on the crossing number
of anyone of its diagrams; try to set up this induction formally.
To use the Homfly equation we will need to extend the notion of
oriented sum of knots to links. One simply chooses one pair of
components on which to perform the usual knot sum.
We now can consider the three diagrams (Fig. 5.20), so
£P(K + L) + e
1
P(K + L) + mP(Kll L) = O.
Fig. 5.20
This gives
(£ + e1)P(K + L) =  mP(KllL)
= (£ + e1)P(K) . P(L)
and so as expected
P(K + L) = P(K) . peL).
Given the link between P and ,,1, we could retrieve our earlier result
on ,,1 from this one.
We leave you to investigate the Kauffman polynomial.
6
Presentations of groups
We need group theory so as to be able to handle some of the deeper
invariants of knots, in particular the group of a knot. We might
expect the type of group theory needed would be fairly limited and
somewhat specialized. However, the overall process translating geo
metric problems into algebraic ones is very much two wayalgebra
pervades the topology and vice versaso quite a lot of group theory
turns out to have implications in topology, whilst the topology can
provide new methods and insights in group theory. Attempts to
compartmentalize mathematics fail; there are too many important
links joining apparently distinct branches. There is an underlying and
allpervading logical structure that is difficult to pin down and to see
clearly, but which is none the less there behind all of mathematics.
ark to the reader
The contents of this section will be needed at different times and how
you use them depends on various factors. You may have a course of
group theory in which some or all of this material is handled; if so
this section merely provides a view of that material that will be useful
in this book. If such a course of group theory is not available then
it will be necessary to treat some, at least, of the topics to a minimum
depth depending on what material from the rest of the book is used
in the course. Our intention in this chapter is thus to provide a
central source for this material and a delineation of the group theory
that will be needed. As explained earlier, we consider that group
theory and lowdimensional topology interact so beautifully that it
would be a pity not to include some vision of this interaction in any
course involving knots, surfaces, etc.
Examples of presentations
We hope you will learn some useful and appealing group theory from
this chapter, but we must remember it is not primarily a group theory
sentations of groups
textbook, and so we will take a fairly informal approach. We shall
assume that you are familiar with the contents of a first basic course
in group theory and have met subgroups, normal subgroups, quo
tient groups, homeomorphisms, and the 'isomorphism theorems'. It
will also help if you have some simple examples of groups at your
disposal, such as the cyclic groups, the dihedral groups, and the
symmetric groups. We will start by examining the dihedral group,
D
3
, of order 6. This is the symmetry group of an equilateral triangle
(Fig. 6.1)
A
c B
Fig. 6.1
Let R be the clockwise rotation through 120°, so that R permutes
the vertices according to the rule A ,) B, B ,) C, C ,) A. Let 8 be
reflection in the axis of symmetry (shown) that goes through A and
the midpoint of BC. so that 8 permutes the vertices according to the
rule
A ,) A, B ,) C, C ,) B.
One easily checks that any symmetry of the triangle can be written
(in more than one way) as a composite of Rand 8. For instance,
rotation through 240° (clockwise) is clearly R2 = RR, whilst reflection
in the axis through B (which is given by the permutation A ,) C,
B ,) B, C ,) A can be written as R
2
8R (= 'do R twice, reflect, finish
with R'). The identity symmetry can be written as R3 or R!' or ](3 etc.
or as 8
2
, 8
4
, etc.
H you have not yourself 'played' with D
3
and it is not enough to
have watched someone else doing soit would be a good idea to
experiment now. For instance, give a description of each of the
elements in the form Ri 8
j
. Write each of them in the form 8
k
R as
wellis it possible? Remember that RO = 8° = I, the identity sym
metry.
As we mentioned above, every element of D3 can be written as
a 'word' in the elements Rand 8, that is it can be expressed as a
product of these two elements. We say that Rand 8 generate the
Examples of presentations 113
group. This generating set {R. S} is certainly not unique. If some set
X of elements generates D3 (so any element of D3 can be written as a
product of elements of X and their inverses in some order and
possibly with repeats), then any Y containing X clearly must also
generate the group. In particular, we could take Y = D
3
, so the set of
elements of D3 can act as a set of generators for the group D3!
Usually, of course, we hope to find as small a set of generators as
possible for a group, but we also want to try to get as simple an
expression as possible for all the elements in terms of the generators.
Of course the meaning of simple tends to vary somewhat from
context to context. For instance, a set of generators for a group
arising in a geometric context may be considered to be a good one
because it allows the elements of the group to be expressed in a
geometrically significant way, but whilst a geometer may prefer those
generators, the significance of the same group to an algebraist may
be such that a completely different set of generators may be con
sidered much simpler. This suggests that we must develop ways of
moving from one set of generators to another, a means of translating
words in one 'alphabet' to words in another.
Knowing the generators is not enough. The symmetry group D4 of
a square has order 8; it is also generated by two elements. All
symmetric groups are generated by two elements. The cyclic group
C
6
or order 6 can be generated by two elements and so on. What
distinguishes one group from another are the relationships or equa
tions satisfied by the generators. For instance. in D3 we saw that R3
is the identity as is S2. so R has order 3 whilst S has order 2. but in
D4 the two corresponding generators have orders 4 and 2. As an
experiment let us try and generate words from Rand S knowing only
tha t R3 = 1 = S2:
Thinking about this, it is clear that any word in R, R2, and Scan
occur and there is no reason why any two of these should coincide.
Thus we have not yet found a defining set of relations for D
3
.
Using the same notation as before, RS is the composite
A ~ B ~ B A
B ~ C ~ A = B
C ~ A ~ C C
so is the reflection in the axis through C and has order 2, that is
(RS)(RS) = 1. This gives us
RS = (RS)l = Sl R
1
= SR
2
Presentations of groups
(remember S2 = 1 and R3 = I). We add (RS)2 = I into our list of
relations and note that we now get at most six words:
(For instance, if one tries to form SR we find
RSRS= 1 =} SR= K1S
1
= R
2
S
so it is already in our list.)
This suggests that D3 can be specified by naming two generators R,
S satisfying
This is the case but we have not proved it yet! It might happen that
(RSi = I forces others of our list of six words to be equal. How can
we know?
ise 6.1
pies
In each of the following cases we have specified two generators and
two relations. One extra relation is needed to obtain the desired
groupfind such a relation
(a) C
6
: the cyclic group of order 6
generators x, y such that x
3
= y2 = 1
plus another relation.
(b) D
4
: the symmetry group of the square = dihedral group of order 8
generators x, y such that X4 = y2 = 1
plus another relation.
(In both cases, the answer will be given later on in this chapter.)
Thus we have arrived at an informal notion of a presentation of a
group. We write
G = (X: R)
where X is a set of generators and R a set of relations.
(a) D3 = (x, y: x
3
= y2 = (xy)2 = 1)
(b) C
6
= (a: a
6
= 1).
Dihedral groups 115
edral groups
The dihedral group Dn of order 2n is defined as being the symmetry
group of a regular ngon. This group has a presentation
(x, y : xn = y = (xy)2 = 1).
We will need later on to handle presentations 'effortlessly'. We will
also find certain ideas from group theory, such as commutators and
conjugacy, useful. We will use this presentation of Dn to illustrate
these concepts, thus also indicating how one plays with presenta
tions.
Firstly we will show that any word in x and y can be reduced to
one of the form xiyl with 0 i < n,} = 0 or 1. The idea is simple. As
xn = y2 = 1, any occurrence of xm for m> n can be reduced to one
with x
m

kn
where 0 m  kn < n. Similarly, any power of yother
than 0 or I can be reduced to either yO = I or y itself. Now if
anywhere in our word we have yx we can use xyxy = I to get
yx = xIyI = xn Iy to move each x to the left of y and then to repeat
the reduction as before. (For instance, if n = 5, xyr = x(yx)x
= xx4yx = x
5
yx = yx = x
4
y.) This really needs to be set out more
formally as a proof by induction on the number of letters in the
word. If w is a word in x and y, the length of w, denoted few), is
the number of letters in w counted with multiplicities, for example
the length of x
2
yx
3
is 6. (Note: x
3
= XI. x J • XI counts as + 3
not  3.)
If few) = 1 then w is one of x, xl, y, or yl. Using the relations
these reduce to x, xnl, y, and y respectively, so if few) = 1, every
thing is fine.
Assume that provided f(lV) < m then the result holds. Now look at
w with few) = m. The word w can be written in one of four forms
w'x, w'x
I
, w'y or w'yI where w' consists of all the letters except the
last one, but few') = m  I and so w' reduces to xiyi by assumption.
The proof now proceeds to check everyone of the different cases that
can occur. (We will not do them all.)
I. If w = w'x then w reduces to XiyJX. If j = 0 then we have Xi+ I
which either is of the required form or if i = n  I reduces further
to xO. If j = I, xiyx reduces to Xi + n  Iy which in turn red uces to
Xi  1 or xn  Iy depending on whether i > 0 or i = O.
2. If w = w'x
1
then xiyixI 'reduces' to This then reduces
to xi+nlyJxn2 if} = I, and so on. We can safely leave the details
to you. Try to find a formula for xiy.i[ I where all the x to the
right of y have been removed.
sentations of groups
The last cases, 3 and 4, are easier. Try them yourself.
Given two elements gl and g2 of some group G, we say gl and g2
are conjugate if there is some g E G such that g2 = gglgl. Being
conjugate is an equivalence relation on G and the equivalence classes
of conjugate elements are called conjugacy classes. Thus g] and g2 are
in the same conjugacy class exactly when they conjugate. Note that
a subgroup N of G is normal if it is selfconjugate, that is if
N = gNg I for all g E G.
Our next aim is to work out what the conjugacy classes in Dn look
like. First we note that (xy)2 = I and y = 1 together have another
consequence:
(xy)Z = I ~ xy=yIxI =yx
I
.
We consider conjugating xiyj by x and by y:
(1) j = 0: x(Xi)X
1
= Xi
j = 1: x(XiY)X
1
= Xi+ lyx
I
= xi+ 2y
(which may need reducing if i ~ n  2)
(2) j = 0: y(xi)y I = yxiy = xn  i
j= l:y(xiY)y=yxi=xniy.
Now if we try to conjugate xiyj by xky we have
(Xky)(xiy1)(yIx k)
and so we can find the conjugacy class of any element by repeated
application of conjugation by the y and the x.
Find the conjugacy class containing x:
x conjugated by any power of itself is x.
x conjugated by y is xn  I.
xn  I conjugated by any power of x is xn  I.
xn  ] conjugated by y is x (why does this return on itself?). Thus
the conjugacy class is {x, xn  ]}.
If G is a group, and gb gz E G, then the commutator [gl> g2] is the
element
[gl> gzl = glg2 g"llg21.
Thus [g], g2] measures how much gl and g2 do not commute. The
commutators in G generate a subgroup [G,G] of G called the
commutator subgroup of G. This subgroup is normal since a con
jugate of a commutator is again a commutator. (Check it yourself.
Note: g(ab)gI can be rewritten (gagI)(gbg]).)
Dihedral groups 117
The corresponding quotient group Gab is called the Abelianization
of G. It is an Abelian group and the quotient homeomorphism
qJ : G 7 Gab
has the pleasing and useful property that given any Abelian group H
and a homeomorphism
IfF: G 7 H
we can factorize lfFuniquely via qJ, that is we can find a homeomorphism
IfF' : Gab 7 H
such that IfF = IfF' qJ and IfF' is the only homeomorphism with this
property.
What is the Abelianization of Dn?
Remarks
Note that commutators obey some simply checked rules such as
[a, b]l = [b, a]
and
lab, c] = alb, c]a
I
[a, c].
In fact, if G = (X: R) then [G, G] is the smallest normal subgroup
containing the set {[x, x1 : x, x' E X}, so the Abelianization Gab has
presentation
Gab = (X: R u ([X, x']: X, X' EX}).
(Again you should try to prove these statements now even though
you will be given a proof later. In any case you will need the result
several times when working with knot groups later.)
6.2
1. Prove that a subgroup N of G is normal if and only if it is a union
of conjugacy classes.
2. Prove that [Dm Dn] = (x
2
), the subgroup generated by x
2
• What is
(Dn)ab if n is odd, and if n is even?
3. Find the conjugacy class of y in Dn Repeat for x
2
and then for
xy. Is there a pattern building up? Try to list all the conjugacy
classes in Dn (Note that when n is even there is a conjugacy class
consisting just of a particular power of x.)
4. We have worked in some detail through the various properties of
the dihedral groups, Dm in order to get you thinking and working
sentations of groups
with presentations. To check that you can handle group presen
tations in this informal way, we suggest you work out the
corresponding facts about the dicyclic or generalized quaternion
groups. The nth of these groups, denoted Q2m has presentation
(x, y: xn = y, yI
xy
= XI).
To start with, prove that x2n = 1, so y4 = 1. (In fact, Q2n has order 4n.)
Next work out the division of the elements of Q2n into conjugacy
classes. Using the fact that any normal subgroup must be a union
of complete conjugacy classes, try to find all normal subgroups of
Q2n (and hence all quotient groups) and finally work out [Q2m Q2,J
and (Qzn)ab'
ey quivers
Given a group G and some set of generators X of G, how is one to
find a complete set of relations? 'Complete' clearly means that no
more are needed to get G. 'Trial and error' may work but is haphazard.
For a start, if one has a set of generators X and some set R of relations,
how is one to know even if (X : R) has the same order as G? The group
G may, for instance, be finite, yet (X: R) is infinite because we have
not yet found all the relations necessary to 'capture' G.
If G is fairly small one can draw a Cayley quiver (also called a
Cayley colour graph). This is a neat way of displaying the needed
information and is also of great theoretical importance, as we will
see later on.
We assume that G is known and X is known to be a set of
generators. We construct a graph having the elements of G as
vertices. More precisely, we set
V= {Vg: g E G}
to be a set of vertices in o n e ~ o n e correspondence with G (and thus
labelled by the elements of G). If Vg is a vertex and x E X a generator,
we draw an edge (labelled with x) from the vertex Vg to the vertex
v
gx
• We do this for each g E G and x E X. The resulting graph is both
directed (the edge goes 'from Vg to vgx') and hence is a quiver (see
Chapter 7) and labelled. The labelling is often thought of as a
colouring of the graph, hence the term 'Cayley colour graph'.
r. C
s
generated by x (Fig. 6.2)
C
s
= {I, x, x
2
, x
3
, X4} X = {x}
There is only one label.
Cayley quivers 119
x
Fig. 6.2
2. Of more interest is D3 (or S3) generated by Rand S (Fig. 6.3).
R
R
Fig. 6.3
The problem of drawing the Cayley quiver of a group G with
respect to generating set X is that it rarely turns out to look as nice
as those which have been given to you in a book! Knowing the Cay
ley quiver does, however, yield the relations. One can 'see' that
R3 = 1, S2 = I and RSRS = 1. To evaluate a word in Rand S. we need
only follow the path through the quiver specified by that word. For
instance, consider
SRRS1RSR
What is it? Reading from left to right and starting at vertex 1, go
along the S labelled edge if you meet an S and along the R labelled
resentations of groups
L  ~ ~   ~       ~ R ~           ~ ~   R
Fig. 6.4
one if you meet an R. The word corresponds to the path as shown
in bold in Fig. 6.4. So we end up at the vertex labelled RS.
Using this idea gives the relations listed, R3 = 1, S2 = 1, and
(RSi = 1, and <clearly' all other loops at 1 can be considered as
consequences of these relations. We only intend this as a vague idea
here. We need more machinery before making it precise.
5 6.3
1. Draw up a Cayley quiver for C
2
x C
3
(== C
6
) generated by a of
order 2 and b of order 3. (If you prefer, C
6
represented as
then a and b are as shown.) Compare this with the Cayley quiver
of D
3
•
2. (a) Draw up a Cayley quiver for
D4 = (x, y: X4 = y2 = (xy)2 = 1)
with x and y as the generators.
(b) What will be the structure of the Cayley quiver of
Dn = (x, t : xn = y2 = (xyf = 1)
again with x and y as the generators?
3. Investigate the Cayley quiver of Q2n with the presentation given
earlier. In particular draw that of Q4 which has order 8.
Cayley quivers 121
4. An automorphism of a labelled quiver is a bijective mapping qJ
which assigns vertices to vertices and edges to edges respecting
labels (so that if edge e is labelled x then it can only be sent to
another edge labelled x) and respecting incidence, that is if e is
an edge from a to b, qJ(e) will be an edge from qJ(a) to qJ(b).
For example, taking r to be the Cayley quiver of D3 with gener
ators Rand S, a rotation of the graph through 120° is an automorph
ism of r as is a transformation that interchanges the inner and outer
triangles reversing directions. List all automorphisms of this quiver.
You can compose automorphisms by 'do one and then the other' so
that they form a group Aut(r). Find the structure of Aut(r) when
r is the Cayley quiver of D3 as above. Repeat with other Cayley
quivers. This should enable you to conjecture a result. Try to prove
what you have conjectured.
Remarks
1. We have assumed that the idea of a quiver is fairly intuitive in
this section. Before we investigate quivers in greater detail and in
particular Cayley quivers, we will take more care with the definitions.
Which comes first, examples or definitions? Historically it is usually
the examples first, and then definitions followed by more examples.
You do not think of a precise definition of a concept until some
examples are available, but the precise concept then brings to light
other examples.
2. Given a presentation (X : R), it can be impossible to construct
the Cayley quiver of the group. The construction is equivalent to
solving the 'word problem' for that presentation. The 'word problem'
is of deciding when two words in the generators yield the same
element of the group. This has been proved in general to be
unsolvable!
If the group is known to be finite, then there is a method known
as coset enumeration which not only gives the exact size of the
group, but also yields a table which can help in drawing a Cayley
quiver. Unfortunately the ToddCoxeter algorithm used in this
method will not tell you if a group is not finite, so this method of
obtaining a Cayley quiver cannot be extended to infinite groups. This
is annoying as most of the groups we will be using later on (knot
groups and surface groups) are infinite groups. (For a detailed
treatment of the ToddCoxeter method of coset enumeration, the
reader can consult Johnson (1990) ).
Notwithstanding this comment about the difficulty of drawing
Cayley quivers of infinite groups in general, Cayley quivers will be
useful later on both for giving examples of certain topological
Presentations of groups
constructions and for 'spinoff" back in group theory of those same
constructions.
Free groups
tion
There is a problem of an almost philosophical nature that needs to
be resolved when dealing with presentations. If we start with a group
G and elements x, y E G, we can easily talk about how to generate
words from the 'letters' x, y and their inverses, and then look for
relations amongst these words. However, we sometimes do not know
anything about G except that x and yare in it and possibly that there
are some relations. What are these 'relations'? We can think of them
as equations satisfied by x and y: the use of the word 'equation'
implies that two things are the same. Where are these two 'things' to
live? The logical problem involved is important but there is also a
practical side. Unless we have a precise meaning for 'word' and
'relation', it will be difficult to avoid circular arguments; we will not
know what to believe or what methods are valid. The idea that solves
the problem is a very elegant one, namely that of a free group. We
start with a set, X, and generate new elements by 'formally' multi
plying the elements of X together to get 'words' in the elements of X,
continuing until we have constructed a group F(X), the free group
on X. For instance, if X = {x} then the free group on X must contain
all powers of x, both positive and negative, and that is all. It is 'free'
since no two of these powers are the same, so there are no constraints
imposed on it. We will return to the formalities of the construction
later. We will start with a formal definition in terms of the most
powerful property that free groups have.
A group F is said to be free with basis X ~ F if for any group H, any
mapping t/J: X ~ H can be uniquely extended to a group homeomorph
ism ¢l: F ~ H (i.e. if x E X ~ F, t/J'(x) = t/J(x); the diagram
Free groups 123
commutes, q/ i = ¢ as mappings).
Remarks
1. This situation can be thought of as being a nonlinear analogue
of the fact in linear algebra that in order to define a linear trans
formation
T : V ~ V '
between two vector spaces, it is sufficient to choose a basis
{ej: i E /} for Vand to specify all the vectors, Tei E V'. Since {ei} is
a basis for V we have, if U E V, U = Lief Utet for some unique set of
coefficients {Ui: i E I} and as T has to be linear, the only possible
value for T(v) is LujT(ei). Thus, specifying the images of the basis
specifies a unique T defined on all of V.
2. If we suppose X consists of a single element x, then ¢ : X ~ H
is determined by the element ¢(x). We suggested that the free group
on {x} consisted of all positive and negative (formal) powers of x,
that is
F({x}) = {x
n
: n E Z}.
This is an infinite cyclic group. We can see that it is free on X since
the fact that ¢': F({x}) ~ Hhas to be a homeomorphism ensures that
¢'(x
n
) = ¢'(xt and this is ¢(xt E H.
Hopefully you can guess at least some of the construction of F(X)
in general. We will call the set X the alphabet; the elements of X will
be letters. We define a symbol an to be a syllable, so cf1, a \ S2, alOl,
etc., are all syllables. A word is a finite ordered sequence of syllables,
for example
where a, b, e EX.
There is one special word, namely the empty word, which has no
syllables. We denote it by 1.
The set of all possible words on the alphabet X will be denoted
W(X), and we can map X into W(X) via the mapping a ~ aI, We
define the product of words by juxtaposition; for instance, for
aZblel and b
l
a
3
d
z
E W(X)
then we take
(azbIe
l
) • (b
l
a
3
d
z
) = dblelbla3d2 E W(X),
We note (i) that multiplication of words is closed: if WI' W2 E W(X)
then WI' W
2
E W(X); (ii) that multiplication is associative: if
4 Presentations of groups
WI' Wz, W3 E W(X) then (WI W
2
)W
3
= WI (W2W3); and (iii) that for any
WE W(X), I.w = w.1 = w, so the empty word, I, acts as an identity.
Since we have treated words as formal objects, W(X) with this
multiplication is not a group. Even though a
l
and aI are in W(X),
we do not have that a
l
• aI is empty. It is clear that if a group is to
be constructed from W(X) then a
l
. aI and I must be identified, that
is we must form the quotient of W(X) by some equivalence relation
so that the set of equivalence classes is a group. The equivalence
relation must clearly include the relations a
l
• a I  1, a I • a
l
1 and
must be compatible with multiplication. We construct the equi
valence relation in a series of steps.
finition of the equivalence relation on W(X)
1. Ifw = Wlaow2, w' = WI w
2
, where WI> W
2
E W(X) and a E X, then we
say that the change, w goes to w', is an elementary contraction of
type l. Its reverse, w' goes to w, is an elementary expansion of type
l.
2. Ifw = wlaPa'lw20 w" = w
l
a
P
+
Q
w2 then we say that the change from
w to w" is an elementary contraction of type II etc. (This just
means we have introduced the usual laws of indices.)
3. lfu, v are in W(X), we will write u  v ifone can be obtainedfrom
the other by a finite sequence of the above elementary operations.
It should be clear that  is an equivalence relation; however, it is
worth repeating that 'it is clear' should be a signal for the reader to
check that it is indeed clear and why. We therefore suggest that you
sketch out a proof that  is an equivalence relation. We will denote
by [v] the equivalence class containing v E W(X). We then note the
following.
Lemma. Ifu, v, WE W(X) and u  v then uw  vw and wu  wv.
We outline a proof, again leaving it up to you to fill in the details.
The first step is to note that if u = U
1
~ U 2 ' V = UI U
2
then uw and vw are
linked by an elementary contraction of type I. (Why?) The same holds
for wu and wv. The next point is to note a similar result for contractions
of type II. Finally we have that there are U2, ... , Un for some n, u = u
l
,
V = Un. and each Ui linked to Ui + I by elementary operations. We can thus
use U
I
w, ... , Unw to link uw and vw in a similar way. A repeat of this
argument with WUj, WU
n
completes the proof.
As a result of this lemma, we have that the operation
[U][v] = [UV]
Free groups 125
is well defined, that is it does not depend on the representatives
u E [u] and v E [v]. If we write F(X) for the set of equivalence classes,
we can note the following properties of this operation:
(1) [I] acts as an identity element since lu = ul = u implies that
[1][u] = [u][l] = [u] for all u E W(X),
(2) the operation is associative as, if [u], [v], [w] E F(X), we have
u(vw) = (uv)w and so [u]([v][w]) = [u][vw] = [u(vw)] = [(uv)w] =
... , etc.
(3) each element of F(X) has an inverse. To see this, first note that
for any syllable a
P
, we have a
P
E W(X) and
[ap][a
P
] = [aJ>+(P)] = [ a ~ = [1].
Any word u E W(X) can be written as a finite string of syllables
and hence can be expressed in the form u = af
1
af2 ... a/:", where the
ai E X and the Pi E lL, the integers. Of course in general the list of
letters aI' ... , an may contain many repeats, they may, in fact, all be
equal, and some of the Pi may be zero. This does not matter as will
be clear, we hope. Define a 'reverse' of u to be
that is the string in reverse order with the sign of each power
changed. An induction on n shows that
tiT) u 1 and uli
r
)  1.
(As usual we expect you to check that you can do this basic type of
argument satisfactorily.) Now we have
[tlYl][u] = 1 and [uJ[d
r
)] = 1
so [u] has an inverse. One might ask if [d
rl
] depended on the choice
of u within the equivalence class, [u]. This question can be tackled in
two different ways. As they are again usual proof forms we expect
you to complete the details. The first involves checking that if u and
v are linked by an elementary operation then tl
r
) and v(r) are linked
by the same sort of elementary operation. (The argument is quite
simple, do not try to make it difficult!) Then assume u and v are
linked by a finite sequence Uj, ... , Un as before and prove d
r
) and
v(r) are linked by U)rJ, ... , li:l.
The other proof is neater but uses the above in a hidden form, that
is by using results we have already noted. We assume u  v so
[u] = [v] and wish to compare [tin] and [v(rJ]. We use that multiplica
tion is associative, that [u<rl][u] = [1] and [v][v(r
1
] = [1], together with
the identity property of [1]:
sentations of groups
[lP)] = [lP)] [1] = [1fT)] ([V] [V(rJ]) = ([ drJ] [V J) [V(r)]
= ([dr)][u])[V<r)] = [l][v(rj] = [V(r)]
Summing up, we see that F(X) is a group. Is it a free group on X?
We can think of X as included in F(X) via the mapping a r? [d], so
suppose we have a group H and a mapping t/J as in the definition of
freeness. We have to prove that (i) t/J extends to give some homeo
morphism ¢f: F(X) ~ Hand (ii) t/J' is unique with this property (i.e.
if If/: F(X) ~ H is any homeomorphism satisfying If/(x) = t/J(x) for
each x E X then If/ = t/J').
We first look at the existence of ¢f. We have to decide what t/J'[uJ is
for an arbitrary [uJ E F(X). There is an obvious way of defining a
mapping t/J1: W(X) ~ H which extends t/J, is completely determined
by it, and is as near to being a homeomorphism as makes sense
(remember W(X) is not a group). For instance, if a E X there is a
syllable an E W(X). We clearly would like t/J1(a
n
) = t/J(at since this
would have to be true if W(X) was a group and t/JI was a homeo
morphism. An arbitrary word w E W(X) is a string of syllables
and if t/J1 was a homeomorphism we would get
which makes sense since it is an equation between elements of H.
Since this is what we hope will happen, why not explore the
consequences of taking it as a definition of t/JI? The set W(X) has a
multiplication and it is clear that t/J1(W
1
w
2
) = t/J1(W
I
)t/JI(W2). At this point
we must pausewhat is t/J1(l)? We have not dotted our i's and crossed
our t's. The empty word, I, is in W(X) but is an empty string (so is
as above with r = 0). What, therefore, is t/JI (l)? Clearly it must be the
product of an empty family of elements of H, but the exact meaning
of that is not clear. However, in H, the usual laws of indices etc. hold
so we always have, for instance,
If we have a string of elements in H, and multiply it by the empty
string (in H) then we must leave the original string unchanged, so
the only sensible interpretation for t/J1 (1) is that it is the identity
elemen t of H.
So far we have used t/J to construct a nicely behaved
t/JI: W(X) ~ H. Is t/JI compatible with the equivalence relation we
have defined on W(X)?
Free groups 127
rna. 1. Ifw = Wlaow2 and w' = W
I
W
2
then =
2. Ifw = w
l
a
P
dlw
2
and w' = w
l
a
P
+
Q
w
2
then =
We leave the proof to you. It just uses the fact that the laws of
indices hold in H. As an immediate consequence of this lemma, we
get that if u, v are in W(X) and u  v then = This in turn
implies that we can (without fear of getting into difficulties) define
= (u).
This will not depend on which representing u we take for [u]. (It is
well defined.)
We have taken as operation on these equivalence classes
[u][v] = [uv]
and we can check that = for all [u], [v] E F(X)
(including the class of the empty string, 1). This completes our
checking that does extend to a suitable It also shows that if
w = ... a:.'r then = ...
How about uniqueness? Suppose 'II". F(X) H extends ¢. This
means that lfIIa] = If [w] E F(X), then if w = ai'l ... a;'r we have
[w] = [ad
n1
••• [arr
r
, and as I/f is a homeomorphism
I/f[W] = (l/f[al])n
1
••• (I/f[ar])n
r
= ...
= ¢'[w]
by our calculation earlier.
We have proved that ptX) is indeed free with X as basis. We note
that we have never actually checked that the mapping a [a] from
X to F(X) was oneone but if [al] = [a21 for some ai' a2 E X with
a
l
'* a2> then we could pick H = C
b
the cyclic group of order 2, and
a mapping X C
2
such that '* ¢(a2)' Now to extend to the
unique F(X) C
20
we must have = (since [al] = [a21!)
but ¢'[ad = and f[a21 = .... We clearly have to abandon
any idea that we can have both a
l
'* a
2
and [al] = [a21.
If two groups FI and F2 are free with bases XI and X
2
respectively,
then any functionf: XI X
2
extends uniquely to a homeomorphism
F(f): FI F
2
. We use the 'universal' defining property of freeness to
check this.
Suppose we write i
l
: XI FI for the inclusion, ;1 (a) = [a] similarly
for i
l
: X
2
F
2
• The function f: XI X
2
gives us a mapping
= i2f: XI F2 which extends uniquely by the defining property to
a homeomorphism FI F2 which satisfies = i2f. It is this
that we take as F(j). Thus if XI has the same cardinality as Xl> the
Presentations of groups
corresponding free groups FI and F2 are isomorphic. (Check this.) In
fact the converse holds: if F; is free on Xi for i = I, 2 and FI and F2
are isomorphic, then XI and X
2
have the same cardinality. This
cardinality will be called the rank of the free group.
We refer the interested reader to Johnson (1990) where the theory
of free groups is explored in more detail. We need free groups
primarily for their use within the theory of group presentations. We
introduced them earlier as a formal solution to the problem of where
the words, that were the 'relations' in a presentation, could live. In
detail the way they are used is a result of the following proposition.
oposition
(a) A set X G generates G if and only if the homeomorphism from
F(X) to G, associated to the inclusion, is an epimorphism.
(b) Every group is a homeomorphic image of a free group.
Proof. First a notational solution to our earlier 'philosophical'
problem: if x E X, we will write x if we are thinking of it as a
generator of F(X) and x, without a bar if as an element of G. This
is necessary since we have, in F(X), elements such as [xlJ[xzl and if.
as might happen, X
I
X2 E X, we have the possible confusion between
[xd[xzl and [X
I
X2], which are different elements of F(X). (With this
notation we have = [xlxzl whilst the other element is [X
I
X2']
The homeomorphism qJ associated to the inclusion of X into G is
then given by removing the bar above each generator and each
square bracket; for instance,
E F(X)
will be sent to XIX; I X
3
X
2
E G.
Now assume X generates G. Recall that formally this means that
any element of G can be written as a product of elements of
X u XI where XI = {XI: x E X} is the set of inverses of elements
in X. Thus given any element g E G, we find an expression of the
form
where t; = ± 1 and all Xi E X. The word
W = [x
l
e
, ... . "'n"n] E F(X)
is sent to g by qJ, so qJ is onto. The converse is simple.
To see (b) we just have to note that any group has a set of
generators, for instance one can take X to be the set G itself. _
Prese ntations 129
entations
Given a set of generators X for a group G we obtain the correspond
ing epimorphism
We can now give a precise sense to the term 'relation' and in
particular 'defining relation'. The above epimorphism gives us an
isomorphism between the quotient group F(X)/Ker q> and G (remem
ber the first isomorphism theorem of group theory), so if we knew
how to construct this normal subgroup Ker q>, we could construct G
from our set of generators which is what we hope to be able to do.
Any word w in Ker q> is, of course, 'killed off' by q>. Let us examine
this in our example of D3 with X = {R, S}. (For the moment we will
keep our cumbersome notation R, 5 etc. to denote the elements of X
when they are thought of as basic elements of F(X), but be warned
that we shall soon drop the use of the bar and you will have to do a
bit of 'doublethink', when looking at presentations.) In F(X) we
have, for example, the word
w = 5
1
RR5R
1
which is 'reduced', that is it cannot be made shorter using elementary
contractions or expansions. This word w is in the kernel of q>,
however, since in D3 we have
(so SI = S)
(so R
1
= R2)
and (RS)2 = 1 (so RS = SI R
1
= SR
2
). Hence
q>(w) = SI R
2
SR
1
= SR
2
SR
2
= RSRS
= 1,
as stated. Thus our word w expresses a relationship between the
chosen generators of D
3
; it is a relation.
Given a set of generators X of a group G with q>: F(X) G the
associated epimorphism, an element w E Kerq> is called a relation
(between the given generators) and Ker q> the relation subgroup.
Except when it is trivial, Ker q> will always be an infinite group, so
it is clearly out of the question to specify all relations; however, we
could specify a generating set of relations. Let us return to our
example of D3 with generators Rand S. We have R
3
, S2 and (RS)2 are
relations; do they generate the relation subgroup? Suppose we
sentations of groups
conjugate R3 by S to get SR
3
SI; this is in Ker q> but it is not clear
whether or not it can be written as a product of R
3
, S2 and (RS)2.
How are we to tell? As usual the answer iswith difficulty. You
might try all sorts of combinations of copies of R
3
, S2 and (RS)2 in
various orders for several days, and yet not hit on one which gave
the result, but clearly that will not do as a proof that it cannot be
done. The fact is that Ker q> in this case is a free group of rank 7
(why this is true will be explained later) and because of this it cannot
be generated by only three elements (again this must wait for
justification later). However, it is generated by these three elements
together with their conjugates. This shows a point of great import
ance: we need only specify a set of 'normal generators' for Ker q> as
it is a normal subgroup of F(X). We need to examine this in slightly
more detail. First some observations.
1. Given any group G and a collection of subgroups SA c G, A E A,
some indexing set, then n {SA: A E A} is also a subgroup of G.
(Check this for yourselves if you have not met the result before.)
2. If Xc G and X generates a subgroup, H, say, of G (i.e. the
elements of H are exactly the products of elements from X and
XI) then
H = n {S: S a subgroup of G, X ~ S}.
3. If all the subgroups SA in observation 1 are normal in G, so IS
their intersection.
4. If X c G, then
H = n{S: S a normal subgroup of G, X ~ S}
is the smallest normal subgroup of G that contains X. (It is often
called the normal closure of X in G.)
If we denote by XI the set of inverses of elements of X, so
XI = {XI: x E X}, then the subgroup generated by X, denoted (X),
consists of all products of elements from X and XI. It is the image
of the natural homeomorphism from F(X) to G corresponding to the
inclusion of X into G. Can we give a similar description of the
normal closure of X in G?
Let us denote this by «X». Clearly since «X» is normal in G, it
must contain not only each x E X but all the conjugates gxg I of x
by elements of G, together with all the inverses and products of all
these elements. If g E G, x E X, we will write g x for gxg I. A con
sequence of X will be an element of G that can be written as a product
of elements of the form gx and their inverses, so y is a consequence
of X if and only if
n
5
Tietze transformations 131
y = {gIXI)£I{g2X2)£2 ... (gnxn)£n
for some gl' ... ,gn E G, XI , ... , Xn E X and where each Ej is ± 1. The
consequences of X in G form a normal subgroup. (Prove it!) In fact the
normal subgroup is the normal closure of X in G. (Again prove it.)
(a) Given a group G and a set X of generators, a subset R of F{X) is
called a set of defining relations for G if «R» =
Ker (cp: F(X) ~ G).
(b) A presentation of a group G consists of a set X and a subset R of
F( X) so that F(X)/«R» is isomorphic to G. We write
G = (X: R).
Let X = {x, y}, R = {X3, y, (xyi}; then (X: R) is the presentation of
D3 that we have been looking to for 'guidance'. Usually we write
things slightly less formally and it is quite usual to write alternative
versions such as
(x, y : X3 = y2 = (xy)2 = 1)
or possibly, using that (xy){xy) = 1 is the same as xy = y I XI,
(x, y: x
3
= ,V = 1, xy = yIXI).
The point is that now we have set up the formal structure of
presentations, we can play around with them with more confidence.
This 'playing around' often involves substituting some consequence
of one relation. For example, consider the presentation
(x, y, z : x
3
= y2 = 1, xyzJ = 1, Z2 = 1).
The third relation tells us that z = xy, so we do not need z as a
generatorit is already there in terms of x and y. We can substitute
xy for every occurrence of z in other relations and delete z from the
list of generators without changing the group. This sort of process is
so useful that it will be used time and time again later on in the
discussion of presentations that arise from knots. Because of this, it
is important to have a list of acceptable 'transformations' of presen
tations that do not change the group.
ze transformations
As usual G = (X: R). The following transformations do not change
the group G:
sentations of groups
Tl: Adding a superfluous relation
(X: R) becomes (X: R') where R' = R u {r}
and r E «R», that is a consequence of R.
T2: Removing a superfluous relation
(X: R) becomes (X: R') where R' = R\{r}
and r is a consequence of R'.
T3: Adding a superfluous generator
(X: R) becomes (X' : R') where X' = Xu {g}
(g is a new symbol not in X) and R' = R u {wgl} where w is a word
in the other generators, that is w is in the image of the inclusion of
F(X) into F(X').
T4: Removing a superfluous generator
(X: R) becomes (X' : R') where X' = X\{g}
and R' = R\{wgl}, with WE F(X') and wg
l
E R and no other mem
ber of R' involves g.
These four operations are called Tietze transformations, after H.
Tietze, who did pioneering work on group presentations (see Tietze
1908; Magnus and Chandler 1982).
We normally carry out more than one change at a time and it is
advisable to work things out informally first before launching into
formal manipulations.
To show (x, y : xl = ym = (xyy = 1) is isomorphic to (a, b : an = b
m
=
(ab)l = 1).
Formally, start:
(x, y : xl = ym = (xyt = 1).
Apply T3 twice to introduce a and b:
(x, y, a, b : xl = ym = (xyr = (xya
1
) = y1b
1
= 1).
Apply Tl three times to introduce superfluous relations:
Tietze transformations 133
Really we should justify that an = 1, b
m
= 1 and (ab)' = 1 are conse
quences of the other relations. This step is often not easy to do. Here
we can write
an = II
i= I
so an is a product of conjugates of the other relations, similarly for
b
m
and (ab)'. The formula above may be represented in a Van
Kampen diagram. For n = 3, the appropriate Van Kampen diagram
is as shown in Fig. 6.5.
a a
a
Fig. 6.5
A circuit around the inner hexagon gives (xy)3 whilst the three
sided faces all have circuit Note that a circuit around the
outside of the diagram gives one a
3
• The 'secret' is to break up this
outer circuit into a 'sum' of the various faces it 'encircles'. A full
account of the theory underlying the use of these diagrams may be
found in Johnson (1990).
Returning to the presentation we can treat J, J J, an, b
m
and (ab)' as 'basic relations' now and prove that x', ym, and (xy)n
are superfluous relations, so three uses of T2 reduce our presenta
tion to
(x, y, a, b: = = an = b
m
= (ab)' = 1).
Next we use and or rather its conjugate to deduce
and hence and its conjugate Now we have
sentations of groups
The relation is superfluous and can be deleted; gives us
b1yl, so with we can apply T4 twice to get the desired
presentation
(a, b : an = b
m
= (ab)' = 1).
This formal method is usually very lengthy, and hence it is usual that
people use the informal 'substitution' method described earlier. What
is important is to realize that the formality is possible and if any step
in the informal sequence of transformations is doubtful, the formal
method is there to act as final arbiter. Informally, we start with
{x, y : xl = ym = (xyt = 1). We try to define a and b in terms of x and
y to obtain an = b
m
= (ab)' = 1. The obvious choice is a = xy; then
an = (xy)n = l.
Next, since we want (ab)' = 1, it seems a good idea to pick b so that
ab = x, but this means b = yI. We try this out to see if it works. The
only relation left to be checked (as we have used two already) is
b
m
= 1, but b
m
= m = 1.
Working the other way: given (a, b : an = b
m
= (ab)' = 1), we must
define x and y in terms of a and b. Trying x = ab and y = I clearly
works.
This process of transforming a set of relations to get a simpler
equivalent one is very similar to the process we have seen in use on
surface symbols. In fact the transformation of surface symbols by
operations is completely analogous to the Tietze transformation
process, and formally these processes are the same. The intuitive
resemblance can be made rigorous via the fundamental group of the
surface. This group, which we will meet in detail in Chapter 11, has
generators the edge letters and a single relation A = 1 where A is the
symbol. Transforming the symbol corresponds to applying a Tietze
transformation to the presentation.
Sometimes the result of Tietze transformations can be quite sur
prising. Consider
(a, b, e, d: ab = e, be = d, cd = a, da = b).
We can try to build a Van Kampen diagram by sticking copies of the
triangles (Fig. 6.6) or merely by 'playing' around with the relations
to see what results:
ab = e b = aIe } => d =
ed= a Impbes d= = b
I
We can, therefore, reduce the presentation to the following:
(a, b, e: ab = e, be = = b).
Tietze transformations 135
6666
abc d
Fig. 6.6
The second and third of these yield c = b
2
, a = b2, so ab = c implies
b
3
= b
2
and b
S
= 1. Elimination of superfluous generators then gives
that the group is isomorphic to C
s
.
The success of Tietze transformations is complete. We note the
following theorem, proved by Tietze (1908).
rem. Given two finite presentations of the same group, one can be
obtained from the other by a finite sequence of Tietze transformations.
Proof. This proof is developed in a structured sequence of exercises.
We begin with some discussion. Presentations are a combinatorial
way of coding up information on groups. When we study groups, we
constantly use homeomorphisms as a way of comparing them.
Natural curiosity should make us ask if there is a notion of mapping
of presentation that will reflect information about homeomorphisms.
Suppose we have
(1) a presentation P = (X: R) of a group G, yielding a homeo
morphism
qJp: F(X) ~ G with «R» = Ker qJp;
(2) a presentation Q = (Y : S) of a group H, yielding a homeo
morphism
qJQ: F( Y) ~ H with «S» = Ker qJQ;
(3) a homeomorphism 8: G ~ H.
Prove there is a homeomorphism f: F(X) ~ F( Y) so that in the
diagram
(*)
G .H
e
Presentations of groups
we have that the two composites 8qJp and qJQfare equal. (The square
'commutes'.) If you want a hint, think of the universal property of
free groups. What does f do to elements in K! (Remember
«R» = Ker qJ.)
Whenever we show that something exists in mathematics, we
should have learnt the reflex of asking if it is unique with the
required property. If it is not unique, a variant of the question is to
ask how close to unique it is. How close to unique is f for a given
8? If.iJ and h both fit the bill, how can we compare.iJ and h?
A mapping f: (X: R) ~ (Y : S) of presentations is a homeomorphism
f: F(X) ~ F( Y)
such that for each r E R, f(r) E «S».
If you have completed the above exercise, you will have proved
that given a homeomorphism 8: G ~ H and presentations P of G, Q
of H, then there is a presentation mapping, f: P ~ Q, which mirrors
the homeomorphism 8 in as much as the square (*) above commutes
(qJQf = 8qJp).
ses 6.6
1. Let P = (X: R) be a presentation of G and Q = (X: R u S) a
presentation of H, say, with f: P ~ Q given by
f = id: F(X) ~ F(X). Show that f is a presentation mapping.
What is the corresponding homeomorphism 8 from G to H?
When is 8 an isomorphism?
2. Let Y be a set, y a given element of Y, and set X = Y\{ y}.
Suppose P = (X: R) and Q = (Y: S) where, if we write
f: F(X) ~ F( Y) for the homeomorphism induced by the inclu
sion of X into Y, feR) c;:; S. (This homeomorphism maps the
generator, x E X, of F(X) to the corresponding generator, x E Y,
of F( Y). It is the unique homeomorphism making the diagram
Inclusion
• Y
x                       ~
1
f
1
.. F(Y) F(X)           ~
commute. The conditions we have imposed on S make f: P ~ Q into
a presentation mapping.) Describe the corresponding group homeo
morphism, 8.
Tietze transformations 137
Now suppose S = feR) u I} where w is some word in the
image of f. Prove that the homeomorphism e is an isomorphism.
(This is of course the basis for the type T3 and T4 Tietze transfor
mations, so do not use those in your proof. If you find it hard to get
started, try the following hint: define g: F( Y) F(X) by
g(x) = x if x E Xc Y
g(y) = w.
Remember that it is only necessary to say where the generators of a
free group go to determine where everything else goes. Now prove
that g is a presentation mapping from Q to P and that the corres
ponding group homeomorphism is an inverse for e.)
3. Let P = (X: R) be a presentation of a group G and let
rp:
be the epimorphism corresponding to the presentation. Suppose Z is
a finite set disjoint from X (X n Z = 0) and pick a homeomorphism
e: F(Xu F(X)
so that e(x) = x for all x E X. Try to describe a set of normal
generators for Ker rpe. Of course, ReKer rpO, but what other
elemen ts do we need? The case when Z = {y} is very close to the
previous example/exercise, so is a good place to start.
4. Now we are ready to approach the task of proving Tietze's
theorem. Suppose we have two finite presentations P = (X: R)
Q = (Y: S) of a group G. We have to show that we can get from P
to Q using only Tietze transformations. First we lift the identity map
on G to give presentation mappings f: P Q and g: Q P. We
have to find a sequence of Tietze transformations giving f. The plan
of action is to assume first of all that X and Y do not overlap, that
is they are disjoint. As we want all of the Ygenerators in, we can do
this one by one using Tietze transformations. At this stage we have
a presentation of the form (X u Y: ?). We then reverse the process
and go to (Y: S) by getting rid of the Xgenerators. That much is
clear, but how can we do it in detail?
Using Exercise 6.6.3 above we can look at
F(X) ( ) F(Xu Y)
and describe a presentation of G of the form (X u Y: R u A) say.
Similarly, again using Exercise 6.6.3 we get
F(Y) ( F(Xu Y)
esentations of groups
and a presentation (X u Y: SuB). Clearly (but check it), the
normal closures of R u A and SuB are the same.
We now look at the obvious presentation mappings:
(X u Y: R u A) ( ) (X u Y: R u SuA u B) ( ) (X u U: SuB),
and show that (i) they induce isomorphisms, and (ii) they can be
realized by Tietze transformations (of types Tl and T2). You can
now look at the various mappings to see if you can get f as the
composite (by choosing the 'retractions' from F(X u Y) to F(X) etc.
in some cunning way)!
You now have the building blocks and the plan. Write out a proof
up to this point, with the assumption of X Il Y = 0.
If we are unfortunate and X Il Y"# 0, we replace X by Xl> Y by
Y
j
so that X and Xl are in oneone correspondence, Yand Y
j
are in
oneone correspondence, and Xl Il Y
1
= 0, X Il Xl = 0, Y Il Y
j
= 0.
Now the isomorphism F(X) F(Xj), induces a presentation
mapping (X: R) (XI: R
j
) and so on. Can you complete the proof
of Tietze's theorem? You may need to introduce some notation for
the various mappings. Choose it carefully.
Proofs of the theorem can be found in Crowell and Fox (1967) and
Johnson (1990).
otient groups
Suppose we have G = (X : R) and we know H is a quotient group of
G, say H == G! N. Is there a presentation of H easily derived from the
given one for G? The first point to make is to note why this might
be expected. We have formed G as a quotient F(X)/«R» and so have
an epimorphism
F(X) G.
The normal subgroup N <J G corresponds to some normal subgroup
M, say, of F(X), that is, the kernel of the composite epimorphism
so H == F(X)/M. We know that «R» eM, so it is reasonable to
expect that H has a presentation of the form (X: R u S) where
rp(S) normally generates N in G.
The above argument can be made rigorous so as to prove the
following theorem. We leave the proof as an exercise as it is quite
simple.
Abelianization 139
Von Dyck's theorem. If Rand S are subsets of the free group F on a set
X and R ~ S then there is an epimorphism
8: (X: R ) ~ ( X : S).
The kernel of 8 is just the normal closure of S\R as a subset of (X: R).
As an example, consider a group G with presentation
(a, b : a
l
= b
n
) for n odd and greater than 2. (This is the knot group
of a (2, n)torus knot.) For which values of n does there exist an
epimorphism from G to S3? The group S3 has a presentation
(x, y : Xl = y3 = (xy)2 = 1), so it is clear that we can find an epimorph
ism from G to S3 if n is divisible by 3. In fact adding the relations
a
2
= 1, b
3
= 1 and (ab)2 = 1 into the relations for the given presentation
of G yields a presentation equivalent to that we have given for S3' This
condition 31n is also necessary, but we leave this to you to check. The
method is to assume the existence of an epimorphism rp: G ~ S3 and
then to see what possibilities there are .for rp(a) and qJ(b).
rcise 6.7
Prove Von Dyck's theorem.
Abelianization
We saw earlier in this chapter that there was an epimorphism from any
group G to its Abelianization Gab' Recall that Gab was the quotient of
G by the commutator subgroup [G. G) of G. We also saw that
[ab, c) = a[b, c)a
I
[a, c]
and [a, br
l
= [b, a). These facts together with Von Dyck's theorem
(6.5) enable us to prove the following theorem.
Theorem If G has a presentation (X: R) then Gab has a presentation
(X: R u {[x, x'] : x, x' E X}).
We know that the kernel of rp: G ~ Gab is the commutator sub
group and is thus generated by all [g, g'), g, g' E G. We also have a
commutative diagram
esentations of groups
and «R» is a subgroup of «8». Moreover, the kernel of qJ is
generated by the images of all [w, w1, w, w' E F[X] (since any com
mutator in G is the image of a commutator in F(X», so we could
take 8 = R u [F(X), F(X)]. What we have to do is to show that the
smaller set of relations
R u {[x, x'] : x, x' E X} not,;tion R u [X, X]
has the same normal closure as R u [F(X), F(X)]. Clearly R u [X, X]
c R u [F(X), F(X)] and so all we need to do is to show any
[w, w'], w, w' E F(X) can be written as a product of conjugates of
elements from R u [X, X1 (In fact R is not needed here as we really
only use [X, X].) As the proof is an induction on the lengths of wand
w', we will leave the formalities to you and merely sketch the
inductive step.
Suppose we know [w, w'] can be written as a product of conjugates of
the various [x, x'], provided t'(w), t'(w') ",,; n  1, and we look at
[w, w1 where t'(w), t'(w') ",,; n. We assume t'(w) = n; if not we can go
to the next step. Then w = WI X or w = WI x I for some WI of length
n  1 and some x EX. Now use
[w, 11/] = [WI x, w1 = WI [x, w1 wI I • [WI' w1
(similarly for w[x
I
) to reduce one level in the first variable.
If t'(w') ",,; n  1, this ends the inductive step; if not write w' = w ~ y (or
w' = w ~ y  I ) and use the expression for [a, be] to simplify. We have not
given all the cases: what about WIX
I
? This gives us a term [xI, w1,
not one of the form [x', w] or [w, x']  how can we get around this?
We leave it to you to finish this proof.
This result gives us a powerful tool for calculating Abelianizations.
We used it earlier to look at [Dm Dn] and thus at (Dn)ab' This tool
becomes even more powerful once we have the results on finitely
generated Abelian groups that follow.
itely generated Abelian groups
A group G is said to be finitely generated (f.g.) ifit has a presentation
of the form (X: R) with X a finite set. Most of the groups that we
will meet are finitely generated. If G is f.g., so is any quotient group
of G, and so in particular if G is finitely generated then so is Gab.
Finitely generated Abelian groups are easy to classify up to iso
morphism, and so the comparison of Abelianizations of f.g. groups
is a useful way of checking for nonisomorphism; that is if G and H
Finitely generated Abelian groups 141
are f.g. groups and we can show that Gab and Hab are not isomorphic,
then G and H are not isomorphic.
Presentations of Abelian groups are often written additively, and
so when considered as a group C
2
has presentation (x: x
2
) but
traditionally when considered as an Abelian group one writes
(x : 2x = 0). A more complicated example is
Dn = (x, y : xn = y2 = (xyf = I)
(Dn)ab = (x, y : xn = y2 = (xy)2 = [x, y] = 1)
as a group but as an Abelian group we might replace
xn = I by nx = 0
y2 = 1 by 2y = 0
(xy)2 = I by 2x + 2y = 0
and of course, since the group is Abelian,
[x,y] by no relation.
It is now simple to see what (Dn)ab is. We have 2y = 0 and 2x + 2y = 0,
so 2x = 0; as nx = 0, if n is odd we obtain x = 0, so (Dn)ab == C
2
generated by y, but if n is even nx = 0 is implied by 2x = 0, so we
have C
2
ffi C
2
in this case (ffi indicates direct sum, see below).
Often manipulation of the (additive) relations using common sense
suffices to find the Abelian group up to isomorphism. However, the
theory of additive presentations is quite easy, so it pays to have the
additional tools provided by the method at your disposal. (Proofs
will be omitted as they are readily available in many group theory
texts.) We start by making precise what we mean by direct sum.
If AI and A2 are subgroups of an Abelian group A, then we can
form a new subgroup
AI + A2 = {a
l
+ a2 : al E AJ, a2 E A
2
};
AI + A2 is called the sum of AI and A
2
• You should check to your
satisfaction that this is a subgroup. It is easy to extend this to any
finite family of subgroups AI> ... , An of A. (We will not need to
extend it to infinite families here.)
We say A has a direct sum decomposition if there are subgroups
AI. ... , As such that
(1) A = AI + ... + As;
(2) for each i, 1 !!S;: i !!S;: s, Ai n L . Aj = O.
'''J
(In (2) we have written Li;<j Aj for the sum of all the subgroups except
Ad We write A = AI ffi A2 ffi ... ffi A ~ .
Presentations of groups
For example, look at C
6
= {O, 1,2,3,4, 5} with operation 'addi
tion modulo 6'. Let AI = {O, 2, 4} which is a subgroup isomorphic to
C
3
= {O, I, 2} (addition being mod 3) and let A2 = {O, 3} which is
isomorphic to C
2
(i.e. {O,I} with addition mod 2). Check these
statements yourselfdo not just take our word for them! We claim
C
6
== Al EB A
2
• Clearly A1 () A2 = {O} so (2) is satisfied and to prove (l),
that is C
6
= A1 + A
2
, involves the extra hard calculation: 5 = 2 + 3
mod 6 and I = 4 + 3 mod 6. Note that A1 u A2 is not a subgroup as
it is {O, 2, 3, 4} and so is not closed under addition (mod 6).
This shows that C
6
is an internal direct sum of a copy of C
2
and one
of C
3
• It is also useful to be able to construct external direct sums
starting with a family of groups and forming a big group containing
copies of the members of the family. For instance, if the family
consists of two Abelian groups A and B, we can form
A EB B = {(a, b) : a E A, b E B}
with addition performed componentwise. This group A EB B has a
direct sum decomposition involving its subgroups:
A' = {(a, 0): a E A} isomorphic to A,
B' = {(O, b): b E B} isomorphic to B.
One can interchange the two forms of direct sum as they are really
equivalent notions. One usually drops the 'internal' and 'external'
as these are clear from the context. One writes, for instance,
'C
6
== C
2
EB C/ and 'C
6
is the direct sum of subgroups {O, 3} and
{O, 2, 4}'.
In fact, if m, n are coprime then C
mn
== C
m
EB Cn. The proof uses
the fact that there are a, b E ~ such that I = am + bn. We leave it as
an exercise.
This decomposition as a direct sum of simpler parts is the idea
behind the two main theorems on finitely generated Abelian
groups.
imary decomposition theorem. Let A be a finitely generated Abelian
group; then A has a direct sum decomposition
where (i) Bi is a nontrivial cyclic group of prime power order pfi for
i = I, ... ,s (the Pi are not necessarily distinct), and (ii) Bio for
i = s + I, ... , s + t, is an infinite cyclic group.
The integer t is called the torsion free rank of A and the prime
powers pfi are called the primary invariants of A. These are uniquely
Finitely generated Abelian groups 143
determined up to order by A and themselves determine A up to
isomorphism.
c decomposition theorem. Let A be a finitely generated Abelian
group; then A has a direct sum decomposition
A = AI (£I A2 (£I ... (£I Ar (£I Ar+ I (£I ... (£I Ar+t
where (i) Ai is a nontrivial finite cyclic group of order ni, for
i = 1, ... , r, (ii) Ai is an infinite cyclic group for i = r + 1, ... , r + t,
and (iii) nlln21 ... Inr (the vertical line is 'divides').
The integer t is the torsion free rank (as in the previous theorem)
and n" ... , n" the torsion invariants, are uniquely determined by A
and determine A up to isomorphism.
These two results are usually illustrated by lists of isomorphism
classes of Abelian groups of some given finite order. More useful to
us is an example showing how to derive torsion or primary invariants
and the torsion free rank of an Abelian group A given by an
(additive) presentation. Suppose A has torsion free rank t and
torsion invariants n], ... , n
r
; then it has a presentation
(remember additive presentation). A general presentation of A will
have the form
(x], ... , Xn :}; = ... = fs = 0)
where Jj = L f ~ I rijxi is some linear expression with integer coefficients.
The relations are thus specified by a matrix and Tietze transforma
tions correspond to integer row and column operations on this
matrix. The aim of the transformations is to diagonalize this matrix.
A = (a, b, c, d: 2a + 5b + 8c + 7d = 5a  4b  7c + 4d = 8a + 5b + 2c +
7d= 0)
has matrix
The resulting diagonal matrix must satisfy n
1
I n21 n3 ... , so the method
must aim to get the highest common factor of the entries of R into
the top left corner. In this example the h.c.f is 1:
resentations of groups
5
4
7
4
j = [  I  I j Rl ( ;
2 ___ 8 7 2 ) 8
7 I 11 5 I
 14
33
105
3
llj [I
30 C2 + 14Cl 0
C3+11Cl
90 ) 0
6 0
o
33
105
3
 14  llj
5 8
7 2
II 5
OJ
30
90 .
6
Next look for the h.c.f. of the remaining submatrix. As this is 3, we
will do a row interchange and multiply by  1 to get
1 _I ;n ) U
U
o
3
o
o
J R4 3R3
120
o
3
o
o
o J R4 [1
o followed by 0
36 R4  3R3 0
12 ) 0
o
3
o
o
 J
 36
So A == C
3
E9 C
l2
E9 SInce In the new generators a', b', c', d' we
have a presentation
(a', b', c', d' : a' = 0, 3a' = 0, 12c' = 0).
If one wants to find a', b', c', and d' in terms of the original
generators, one applies the row operations used in reverse order to a
unit matrix of size s x s where s is the number of original generators
(in the example s = 4). We leave you to do this in the example.
To find the primary invariants, note that 3 is prime whilst
12 = 22 x 3, so C
l2
== C
4
E9 C
3
and the primary decomposition is
A == C
4
E9 C
3
E9 C
3
E9 C
w
(Readers who have not studied this theory
in detail can find treatments in Johnson (1990) or Rotman (1984).)
ushouts of groups
There is a configuration that will arise many times in various forms
in the later parts of this book. The idea is simple. We have two
groups G and H and want to form a group that contains both G and
H. We saw that even in as simple a case as C
2
and C
3
considered as
subgroups of C
6
, the setwise union C
2
u C
3
was not a group. In that
Pushouts of groups 145
example we had a bigger group in which we could work. Now if we
are given G and H, how should we multiply g E G and hE H? We
have seen something like this before. When developing the ideas that
led to free groups, we formed 'words' in the elements of the set of
generators. Here we could clearly form words in the elements of
G u H (as sets) and then 'remember' the multiplication in G and in
H. We can try this out: for instance, if we take the words glhg
2
and
g3h' g4h" and 'multiply' them together we get
(gl hg
2
)(g3h' g4h") = gl h(g2g3)h' g4h";
that is, put the words next to each other (remove the brackets) and
multiply the end elements together if this is possible. This is vague.
It can be made more exact, but as we have a limited need for these
ideas and we have the machinery of presentations available to us, we
will take a short cut.
We first note a distinctive property of the union of two sets that is
linked to the gluing lemma of Chapter 3. Suppose X, Y, and Z are
sets with f: X ~ Z, g: Y ~ Z two functions that agree on X n Y,
that is if x E X n Y then lex) = g(x), then there is a unique function
h: Xu Y ~ Z such that if x EX, hex) =/(x), and if y E Y,
hey) = g(y). This is so obvious as to be almost unnecessary to state.
The role of X n Y is important here; if / and g did not agree on
X n Y then no joint extension h could exist.
This sort of situation arises in numerous contexts and is abstracted
in the notion of a pushout diagram. Suppose we have a diagram
a'
of sets and functions. It is called a pushout diagram (of sets) if it is
commutative (i.e. if the two composites /3' a and f3d are equal) and
it has the universal property that, given functions
f: X ~ Z, g: Y ~ Z satisfying fa = g/3, there is a unique function
h: T ~ Z such that h[3' = / and ha' = g.
(The uniqueness of h with this property is very important
and should be thought of in the example we looked at where
W = X n Yand a and /3 were inclusions. Note in general that a and
/3 are not assumed to be inclusions.)
Presentations of groups
We leave it to you to show that given any pair of func
tions a: W X, /3: W Y there is a pushout diagram as above
containing them. (You form T by taking a disjoint union of X and
Y and dividing out by a certain equivalence relation. The disjoint
union of two sets can be obtained by taking two new sets X" Y
2
consisting of pairs (x, 1), x E X, or (y, 2), y E Y and then forming
Xl u Y
2
• Even if X n Y is nonempty, Xl n Y
2
is empty, so X
1
and
Y
2
are disjoint copies of X and Y; we write X 11 Y for this disjoint
union.)
We are not really going to use push outs of sets but a similar notion
can be obtained for groups by asking X, Y, W, T, Z to be groups
and a, {3, a', {3', f, and g to be homeomorphisms. We could also use
another analogous situation with all the sets replaced by topological
spaces and all the functions by continuous maps.
If we are given a 'top lefthand corner' of groups
K __ G
H
can we complete it to a pushout diagram
K __ G
jw
H T
a'
of groups? The case K = {I} is the simplest case and is that with
which we started this section.
oposition. IfG=(X: R), H=(Y: S) then the group T=(Xu Y: RuS)
with the obvious inclusions forms a pushout diagram
Push outs of groups 147
{1}
a
.. G
~ l lw
H .. T
a'
and T is uniquely determined up to isomorphism by this property.
Proof Suppose we have a group Z and homeomorphisms};: G ~ Z
h: H ~ Z. Of course the commutativity of the square is automatic,
since {I} only contains the identity element.
We note that a'(y) = y, {3'(x) = x (with some abuse of notation
here!). We define f: T ~ Z by defining f on generators: if
a E Xu Y then a E X or a E Y but not in both; if a E X, then
f(a) = };(a), and if a E Y then f(a) = Ji(a). We have to check that if
w E R u S then f(w) = I, but if w E R u S then W E R or W E S, but
not both, so we can check if W ERe F(X) then W only involves the
elements of X and f(w) = };(w) = 1, similarly if WE S. (We will look
at the uniqueness of T in the general case later.) •
This group T will usually be written G * H and it is called the free
product of G and H. Even if G and H are finite groups, then G * H
is infinite unless one of G or H is the trivial group.
The general case is only slightly more difficult to describe.
fJ a
position. Ifin a corner H ~ K ~ G, G = (X: R), H = (Y: S) and K is
generated by a set Z then T = (X u Y: R u S, {a(z) {3(z) I : z E Z}),
with the obvious homeomorphisms, forms a pushout square
K __ =:a __ .. ~ G
~ I lw
H .. T
a'
and T is uniquely determined up to isomorphism by this property.
Proof. Firstly notice that the addition of a(z) {3(Z)1 = I for all
z E Z into the relations guarantees that the square commutes.
Again we assume that given };: G ~ Z, h: H ~ Z such that
.r. a = hf3 and define f: T ~ Z by the same method as before. Again
sentations of groups
we have to check that f of any relation is trivial but if lV E R u S we
can use the argument we used before and if w is of the form
a(z)f3(zr
l
, then few) =fa(z)ff3(zr
l
=j;a(z)hf3(zr I = 1. (We have
systematically abused notation in the above. The idea of the proof is
simple but to make the distinction between words in F(X) and
F(X u Y) say, which strictly speaking needs to be done, would seem
to complicate and obscure that simple idea.)
We now turn to uniqueness. Firstly we should explain what IS
meant here. Suppose someone presents us with another square
G
jy
H T,
with the same universal property: given any j;: G Z and
h: H Z such that fi a = hf3 then there is a unique f: TI Z
satisfying fi = fy, h = f 15. If T is a pushout as claimed in the proposi
tion then T == T
j
• To see this, we look at the two diagrams
a a
K .. G K
•
G
jw
(U) j jy
H .T
...
... , ....
o T,
a' T
In (i), taking j; = y: G TI> h = 15: H T
J
gives us a unique
tjJ: T T
j
satisfying 13' = y, a' = 15. In (ii), takingfi = /3',h = a' gives us
a unique IfF: T
j
T satisfying IfFY= /3' and IfFI5= a. Now we note that
IfFtjJ/3' = IfFY= /3' and IfFtjJa' = lfFa = a
whilst
Push outs of groups 149
t/Hfl= tjJf3' = r
and tjJ IfFO = tjJ a' = 0.
Looking at the diagram
ex
K .. G
~ j
j ~ '
H ..
where J; is taken to be 13' and h to be a' gives a unique map
f: T ~ T satisfying fa' = a', ff3' = 13'. We have two candidates for f,
the identity map on T and IfFtjJ. Both satisfy the equations and so they
must be equal since we know there is a unique map satisfying these
equations. A similar argument shows that tjJlfF is the identity on T
1
;
hence tjJ and IfF are isomorphisms.
(We have given this argument in quite a lot of detail as it illustrates
the power of definitions involving universal properties. A similar
result is true for the universal property defining free groups.)
To complete the proof we have to return to our original T and
f:T ~ Z. Why is f unique with the property that ff3' = J;, fa' = h?
Any homeomorphism h: T ~ Z is determined by what it does to
generators and any generator is really either an a'(y) or a f3'(x), so
ha'(y) = h(y) whilst hf3'(x) = J;(x), that is h is the same as f! (Here,
to some extent, we must avoid our earlier deliberate abuse and
simplification of notation. We wanted previously to think of F(X) as
sitting inside F(X u Y) and we used the letter x equally for a
generator in X or in the Xpart of Xu Y; now we need to make a
distinction between them, so we call x to X by the name f3' (x) if we
are considering it as an element of Xu Y.) •
When K = {I} we of course get back exactly to our earlier descrip
tion of the free product. If a and 13 are monomorphisms so that G
and H contain K as a subgroup, then the group T is called the free
product of G and H with K as amalgamated subgroup. Its construc
tion can be thought of as being obtained from G * H by gluing
together the two copies of K, one in G and one in H.
sentations of groups
The above proof, although an abstract one, contains all the
information to enable us to describe the elements of this free product
with amalgamated subgroup. Suppose g E G, h E H, k E K; then
ga(k) E G, f3(k)h E H, and we can form the words (ga(k»h and
g(f3(k)h) within T. As k E K, it can be written as a word in the
generators, Z, and for each Z E Z, a(z) = f3(z) within T; as a result we
find that
(ga(k))h = g(f3(k)h).
This example shows what the elements of T look like. They are
strings of the g and the h as in G * H, but now any element in
a(K) can be identified with the corresponding one in f3(K).
These pushouts of groups are very useful. They also include other,
often used constructions.
6.10
Suppose in a corner
a
H
that K is a normal subgroup of G with a the inclusion and that
H = {I}. Prove that the group T that completes the pushout square
is isomorphic to G/ K with 13' the usual quotient map. What happens
if K is not assumed to be normal in G?
t groups
One of our main purposes for spending a fair amount of time on
group presentations is that they are extremely useful when working
with knot groups. The full geometric significance of the definitions
and constructions we will give will not be clear until later on, but the
combinatorial methods used make calculations very easy to do. We
will look at two alternative methods of obtaining a presentation of a
group from a knot diagram, and then examine the independence or
otherwise of these groups with regard to Reidemeister moves before
looking at various links with other ideas we have met.
Knot groups 151
presentations
The naIve idea behind the Wirtinger presentation is that if you want
to explore a knot, just going for a walk along the knot itself is not
much help as it is only a circle, but examining the space around the
knot will tell you a lot more. You can think of that space as being a
threedimensional labyrinth. Taking the advice of Ariadne, we follow
Theseus in taking with us in our exploration a ball of string, letting
it out as we go so as to know where we have passed already.
Examination of the possible paths yields some generating paths and
the crossings give us relations.
(Before we proceed further, it may help if the reader takes a piece
of moderately stiff wire and bends it into a trefoil knot, joining the
ends together to complete the 'circle'. A piece of strong cord will also
be useful!)
Take a picture of the knot, orient it, and label each arc. These
labels will be the generators of the knot group, so use sensible
distinct labels, possibly labelling around the knot in some order. We
will illustrate using the cinquefoil in Fig. 6.7.
Fig. 6.7
The knot group will have a presentation
G(K) = (XI> ... , xn : r j , ••• , r m)
where n is the number of arcs in the picture and m is the number of
crossings. (It helps to number the crossings so as to avoid missing
any out of the calculation.)
The rule for writing it down is: starting at some point near the
crossing move anticlockwise around a small rectangle; at each arc write
down the corresponding generator with exponent + I if the arc is
sentations of groups
Fig. 6.8
entering the crossing and  1 if it is leaving it. Hence starting at the
top gives Xc XiX" I xi I. Clearly starting at a different position produces
a conjugate of this relation.
Let us assume that at the first crossing we had the form shown in
Fig. 6.8. This corresponds to the relation r
l
: Xcxixfc I xi I.
There is a second possible type of crossing of the form shown in
Fig. 6.9 which gives XcXiIXfclX;= I.
Fig. 6.9
For our example of the cinquefoil knot, we have five generators
and the following five relations:
r
l
: XSXIX.;lxjl = I
r2 : XIX4X21X4 = 1
r3 : X4X2X;1 X2
1
= I
r4 : X2X3X:sIXil = 1
rs : x3xsxilX:s1 = I
or x4=xllxsxl
or X2=X.;IXIX4
or X3 = X21X4X2
or Xs = x; I X2X3
or XI = X:sl X3XS'
There are several things to notice here. Firstly each relation
expresses a generator (corresponding to the arc leaving from under
the overpass) in terms of the 'incoming' generator conjugated by the
Knot groups 153
'overpass' generator or its inverse. This observation allows one to set
up inductive machines to calculate the generators leaving a series of
twists in terms of the incoming arcs (Fig. 6.10).
a
 .... are
_,' ... ....../'...../"' these?
b
n twists
Fig. 6.10
Secondly the cinquefoil is extremely symmetric and this is reflected
in the relations. In fact one can write them in a 'string'
This can be used to take short cuts to simplify the presentation. You
are advised to use special structures such as symmetry to help in
calculations. At the same time be warned that as not all knots are
symmetric, we need basic simplification techniques that work regard
less of any special features. We will apply simple Tietze transforma
tions to the Wirtinger presentation so as to simplify it and will not
use the symmetry explicitly.
The presentation is
G(K) = (XI' X2, X
3
, X4' Xs: XI = XS!X3
X
S' X2 = X:;IXIX4' X3 = X:;IX4X2' X
4
=
XII XSXI' Xs = xi! X
2
X
3
).
We eliminate (any) one generator, say Xs, by substituting for Xs the
expression XiI X2X3:
XI=XS!X3 X S
X
2
= X:; I X!X
4
X3 = X2! X
4
X
2
X
4
= XI! X
3
X
I
becomes XI = XiI X:; I X
3
X
2
X
3
does not change
does not change
becomes X
4
= Xj
l
xi!X
2
X
3
X!
and we delete Xs from the list of generators. Here strategic thought
has to come in. Substitutions of complicated expressions are more
likely to lead to slips than are substitutions of simple ones. It
therefore is better to proceed by substituting for X2 or X
3
, not XI or
X
4
• We choose to substitute for X3 next. We will write down the
results:
G(K) == (Xl> X2, X4 : Xl = X2! X:;! X
2
X
4
X2
1
X
4
X
2
' X
2
= X:; ! X! X
4
,
X
4
= Xj!X2IX4IX2X4X2X!),
Finally, using the expression for Xl gives
resentations of groups
G(K) == (x], x
4
: XI = X4IX.IX;;IX.IX4XIX4XIX4'
x
4
= Xli X
4
1
XlI x
4
I XI X
4
XI X
4
X
I
).
We note that the two relations are essentially the same:
In Wirtinger presentations, anyone relation is a consequence of the
others. (The reason is that each of the relations corresponds to a path
or a loop of string that can be pulled off the knot. Any single such
loop can be pulled until it involves not that chosen crossing but all
the others.) Keeping all the relations acts as a check on your
calculations, so do not just throw a relation away. As a final step we
simplify one stage further.
Write a = XI X
4
, b = XIX4XIX4XI; then a few more Tietze transforma
tions give
G(K) == (a, b : as = b
2
).
The cinquefoil is a (2, 5)torus knot. Here are the 2 and the 5 coming
out from the calculation of G(K). We will be studying torus knots
and their groups more closely later on in Chapter II.
esentations
You may recall the way in which the Alexander polynomial was
originally introduced by Alexander. That approach also leads to a
group presentation introduced by Dehn.
Orient the knot as before but now label the faces of the resulting
picture. These facelabels will be the generators. As in the Wirtinger
presentation we get a relation for each crossing. A typical crossing
might be as shown in Fig. 6.11 and corresponding to this we write
down the relation XiX] I XkX,1 Finally we add a relation Xm = I for
Xi
t
Xl
Xj
t
X
k
Fig. 6.11
Knot groups 155
g
;;
Fig. 6.12
some generator/facelabel Xm  anyone will do. This last relation
effectively deletes Xm from the presentation.
Again we will examine the cinquefoil (Fig. 6.12).
The generators will be a, b, e, d, e, f, and g. The relators are
bg1af
1
, eg1bf1, dg1ef
1
, eg1df
l
, ag1efl
and we will choose to 'kill off' g as this seems to cause a lot of
simplification to the presentation. This gives
(a, b, e, d, e,f: baf
I
= ebf
I
= defl = edfI = aef
I
= 1).
Eliminating f (since f = ba = eb = de = ed = ae) gives us
(a, b, e, d, e : ba = cd = de = ed = ae).
Glancing back at the symmetric relations we had for the Wirtinger
presentation suggests that this is the same group and so should have
a presentation of the form (x, y : X5 = y2).
Knowing where we are going shows a short cut. This common
element f that we have actually got rid of, if repeated, gives an
interesting pattern
f5 = (ed)(eb)(ae)(de)(ba) = (edebaf
so we take y = edeba and, using a few Tietze transformations, get
(f, y : f 5 = y2).
We leave it to you to try to prove, using the knowledge that you
have at this point in the book, that the Dehn presentation is a
presentation of the same group as the Wirtinger presentation. We
also suggest that you repeat the exercises above for the Dehn method
sentations of groups
e
and also examine why it does not matter which of the original
facelabel generators you kill off.
To be useful these knot groups, obtained either by the Dehn or
Wirtinger method, should be invariants, that is should not depend
on which picture we use to calculate them. To put this more formally
we hope that if K and L are equivalent knots then G(K) and G(L)
will be isomorphic groups.
Later on we will show that G(K) is the fundamental group of the
knot complement ~ 3 \ K . This provides a neat use of the fundamental
group calculations that we will meet in Chapter 11 and proves
invariance in an elegant way. However, there is much to be said for
a barehands approach. We know equivalences can be broken down
into Reidemeister moves, so it would seem a good idea to see what
happens to the knot group if we change a knot by a Reidemeister
move. We look at Wirtinger presentations, and leave the Deh]
versions as an exercise .
. Move I (Fig. 6.13)
x
KOld Knew
Fig. 6.13
The presentation for Knew will have one more generator, x', and some
of the relations involving x in G(Kold> will be replaced by relations
involving x'. In addition we have a relation for the new crossing,
namely XXXI x' I = 1, which, of course, gives x' = x. This in turn
means that we can substitute x for x' in all relations, eliminating x'
and reducing our presentation to that of G(Kold). Thus in this case
the two groups are isomorphic.
Move II (Fig. 6.14)
Again the change produces new generators and some changes in
other relations. It also gives relations
Fig. 6.14
a t the two new crossings. The first gives
z = x'y,x
Knot groups 157
meaning that we can eliminate z from all other relations, but of
course it can only occur in the second of these relations. Substituting
this in the second relation gives Y2 = YI, so we can again reduce to the
old presentation.
We should also look at the move with one changed orientation
(Fig. 6.15) but this causes no problems. Check that you can do the
adaptation of the argument above to this case.
Fig. 6.15
Move III
This causes more bother and again we will only look at one of the
numerous different subcases caused by relative positions and orienta
tions.
In Fig. 6.16 the relations are a'ebe\ e'b'db, and iIeee',
whilst in Fig. 6.17 the relations are aI ebe', g I ede
I
, and gI
aI gao (We can safely leave all labels other than e and g
sentations of groups
/
•
Fig. 6.16
Fig. 6.17
unchanged as we are only changing the knot in the vicinity of this
triple crossing.)
Simplifying the old presentation eliminating e we get
a! cbc! and j !a
l
cdc! a.
Comparing these we find that the old one has cb
1
where the new one
has aI c, but the other unchanging relation gives a t c = cb
l
, so the
two presentations are equivalent and the two knot groups G(K
o1d
)
and G(Knew) are isomorphic.
6.11.1
1. Calculate the Wirtinger presentation of (i) the unknot, (ii) the
trefoil, and (iii) the figureeight knot.
2. The (2, n)torus knots can all be drawn in the form shown in Fig.
6.18 where n, the number of crossings, is odd. Using the labelling
Knot groups 159
T2.n
Fig. 6.18
suggested, work out a formula expressing the generator Xk of the
Wirtinger presentation in terms of the generators x I and X2.
Using the last two crossings (the furthest to the right in the
diagram), write XI and X2 in terms of themselves and hence write
down a presentation of the group of this knot. Simplify to show
the knot group is isomorphic to
(a, b : an = b
2
).
3. There is another family of knots that might be called the
generalized figureeight knots; their general form is as shown in
Fig. 6.19 with a total of n + 2 crossings for the nth knot in the
series. For n = 1, we have a trefoil and for n = 2, a figureeight.
Using similar methods to the preceding exercise, find a presenta
tion for the knot group of the nth knot in the family. Investigate
if it can always be presented with just two generators. Is there an
Fig. 6.19
Presentations of groups
'elegant' way of presenting this group in any way analogous to
the an = b
2
relation form of G( T
2
. n)?
4. Prove that G(T2. 3) is not isomorphic to G(unknot). (You may not
be able to complete the proof now but try to understand what
the problem is.)
5. If K is any knot, show that G(K)ab, the Abelianized knot group,
is infinite cyclic.
6. Examine how the Dehn presentation changes under Reidemeister
moves and check that the results are equivalent presentations,
that is giving isomorphic groups.
groups and the sum of knots
What is G(K + L) in terms of G(K) and G(L)? For a change we will
examine this for the Dehn presentation but suggest you examine the
case of Wirtinger presentations for yourselves.
a
Fig. 6.20
Orient the two knots and label the faces as in Fig. 6.20. Use the
same label for the unbounded region, a say, and in both presenta
tions use a as the generator to kill off. (If you have not checked that
the choice of which generator to kill is of no consequence to the end
result, now would be a good time to pause and think about the
problem.)
Now look at K + L; this has picture shown in Fig. 6.21 where z
is assumed to be a facelabel not used in K or L. It should be clear
that if
a
Fig. 6.21
Knot groups 161
G(K) = (X: R) X a set of generators containing x
and
G(L) = (y: S) Ya set of generators containing y
with X n Y= 0, then
G(K + L) (X, Y: R, S, x = y)
Thus the rule for forming G(K + L) is that of forming the pushout
a.
(z: $) G(L)
j
G(K)
where a(z) = y, {3(z) = x.
ups and 3colourability
Suppose we have a 3colourable knot, K, coloured by the colours a,
b, c. We will construct a group G with generators a, b, e and the
relations so that colouring K, that is attaching a colour to each arc
and hence to each generator of G(K), defines a map on generators
from G(K) to G compatible with the relations, thus giving an
epimorphism from G(K) onto G.
The idea is that if Xi and Xj are generators both coloured a then we
add in the relation Xi = Xi and call the amalgamated generator thus
formed a, and so on.
A typical relation in a Wirtinger presentation is of form xyxIz
I
or xIyxz
I
. If this occurs at a crossing where x, y, and z are all
coloured the same, the relation will vanish in G, that is we need not
worry about compatibility. If x, y, z are coloured a, b, e in some
order then mapping arcs to colours maps this relation to one of the
following pairs of relations:
(i)
abaIeI,
(ii)
babIeI,
(iii)
eaeIb
I
,
aIbae
I
bIabc
I
eIaeb
I
(iv)
aeaIb
I
,
(v)
bebIa
I
,
(iv)
ebeIa
I
aIeab
I
bIeba
I
c1beb
l
•
Half of these are consequences of others, for instance if abaIe I = 1
then aba
I
= e, so b = alea, that is aIeab
l
= 1. The six relations in
sentations of groups
pairs (iv) to (vi) can therefore be omitted. This leaves us to look for
a group with three generators and relations (i), (ii), and (iii). We
therefore take
G = (a, b, c : aba1c
1
= a Ibac
I
= bab1c
1
= b1abc
1
= cac1b
1
= c1acb
1
= 1)
and we have a universal quotient of all groups of 3colourable knots.
This group G is rather unfriendly. It is infinite and its structure is not
immediately recognizable. but it does have a wellknown quotient
group.
Thinking of the transpositions (12), (13), (23) in S3 we note that
(12)(13)(12)1 = (23) etc.; this is similar to the type of relation we have
above except that the generators in S3 are of order 2. If we add the
relations a
2
= b
2
= c
2
= 1 into the presentation of G we get a quotient
group
G
1
= (a, b, c : aba = c, bcb = a, cae = b. a
2
= b
2
= c
2
= I)
and G
1
mimics much of the behaviour of S3' We do not need to work
out whether or not G
1
is or is not S3, since by mapping a to (12), b
to (13), and c to (23), we can construct a homeomorphism from G
1
onto S3' What we have done is to use the colouring of K to map G(K)
onto S3, that is we have shown that the geometric and combinatorial
notion of 3colouring translates to some extent to the existence of an
epimorphism from G(K) onto S3' The interest of S3 rather than G or
G
1
is that it is a weBknown group and so such a homeomorphism
helps our understanding of G(K) itself. The converse of this result
holds: if G(K) maps epimorphically to S3 then K is 3colourable.
6.11.2
I. Prove that if G(K) maps epimorphically to S3 then K IS 3
colourable. (This is quite hard.)
2. Prove, using this method, that the (2, n)torus knot [2. n is 3
colourable if and only if n is divisible by 3. (Some of the parts
necessary to this are handled earlier in this chapter.)
3. Investigate the possibility of an analogous result for ncolourability.
(This has been done by Crowell but we will not tell you which
target group he used.)
7
Graphs and trees
The theory of graphs impinges on the study of knots and surfaces in
four significant ways. We shall see how knots can be studied by
graph theory applied to their diagrams, and how graphs can be
studied by embedding them in surfaces. We have already met (in
Chapter 6) the Cayley graph of a group presentation, which produces
a geometric representation of a group from sets of generators and
defining relations, and later on we shall meet covering graphs that
give a proof, via Van Kampen's theorem, of a fundamental result
about free groups. It is not our aim to compress a complete
introduction to graph theory into this single chapter; rather, we
select the essential facts and leave the reader to pursue the wider
subject in texts intended for that purpose. We mention Wilson (1985)
and White (1984) as being particularly suitable.
phs, quivers, and trees
An initial difficulty in graph theory is that of terminology. No
accepted standard has arisen to settle the question of what a graph
should be. We shall stretch the term to its fullest extent, but the
reader must be warned to consult other authors' definitions when
referring to other sources.
A choice presents itself as to whether the edges in a graph come
with a preferred direction or not. Our main interest happens to be in
directed graphs, but many of the important concepts are most easily
formulated for graphs without a preferred direction for edges. We
shall develop the ideas for directed and undirected graphs side by
side, thereby avoiding a choice altogether. An undirected graph we
call simply a graph; a directed graph could be called just that, or a
digraph, but we prefer the more picturesque term quiver.
A graph r consists of a set V r of vertices and a collection Er of
pairs of elements of V r. A pair {u, v} is called an edge of r, we allow
pairs of the form {u, u} and for pairs to be repeated in Er. It is
phs and trees
extremely useful to represent a graph r (at least when V r is small)
by a diagram in which the elements of V r are represented by points
and an edge e = {u, u} is represented by a line joining the points u
and u. For example, the graph with V r= {I, 2, 3, 4, 5, 6, 7} and
Er= {{I, l}, {t, 2}, {I, 3}, {l, 4}, {3, 4}, {3, 4}, {3, 4},
{3, 5}, {5, 4}, {5, 4}, {6, 7}}
could be represented by the diagram in Fig. 7.1.
7
I
2 3 6
Fig. 7.1
In drawing such a diagram, the essential requirement is correctly
to represent the disposition of edges joining vertices; the actual
placement of the points, and whether the lines are straight, curved,
or have to cross one another, is irrelevant. So the two diagrams
below both represent the same graph, as in Fig. 7.2.
Fig. 7.2
Of course, the fact that we can draw a diagram of a graph r
without edges intersecting is a significant property of r, and we shall
consider it in detail later in the chapter.
Different choices of labelling for the vertices and edges of a graph
will not affect its graph theoretic properties, so we say that two
graphs rand r' are isomorphic if there exists a bijection
f: v r ~ vr' such that mapping {u, u} E Erto {f(u),f(u)} E Er' is
a bijection E r ~ Er'. The bijectionfis then an isomorphism between
rand r'.
It is evident that the notion of a graph embodies a basic idea of
everyday life: that two things are connected in some way. Examples
Graphs, quivers, and trees 165
include the stations on a railway network, the components of an
electrical circuit, students who attend a lecture together, or tennis
players who meet in a tournament. Sometimes the direction of
connection is important, and sometimes it is irrelevant.
A quiver Q consists of a set VQ of vertices and a collection EQ of
ordered pairs of elements of VQ. A pair (u, u) E VQ is an edge from
u to u. We allow pairs (u, u) and for pairs to be repeated in EQ.
Evidently we can represent a quiver by the same sort of diagram as
for graphs, with an arrow on the line representing (u, u) pointing
from u to u (Fig. 7.3).
7
I
2 3 6
Fig. 7.3
Since every edge in a quiver has a welldefined initial vertex, or
source, and a terminal vertex, or target, a neat definition of a quiver
with some conceptual advantages is as follows. A quiver Q consists
of two sets VQ and EQ and two functions s, t: EQ ~ V Q. That is all
we have to say. An element of EQ is an edge of Q and an element of
VQ is a vertex of Q. The functions sand t give the source and target
of an edge. Given two quivers Q and Q' with source and target
functions s, t and s', t' a map of quivers f: Q ~ Q' is a pair of
functions fo: V Q ~ V Q' and j;: EQ ~ EQ' such that s'J; = fo sand
,'J; = fo t. If each of fo and J; is a bijection then f is an isomorphism
and Q and Q' are isomorphic. The equivalence of the two definitions
of quiver should be clear: an edge e = (u, v) has see) = u and tee) = u.
We shall generally prefer the second definition, but the first makes
plain the fact that a quiver is a graph with extra structure.
Any quiver has a graph associated to it, obtained by forgetting
about the ordering of pairs in EQ, or by ignoring the arrows in a
diagram. We call this associated graph the underlying graph of Q and
denote it by 1 QI. It follows at once that Vi QI = VQ and that
EI Q I = {{s(e), tee)} 1 e E EQ}.
The notion of a subgraph of a given graph is straightforward: in
terms of diagrams, a subgraph is represented by some subset of the
vertices and some subset of the edges joining these vertices. More
precisely, if Tis a graph, then a subgraph of Tconsists of a subset JV
of V r and a set F of pairs of elements of W such that F ~ E r. I t is
phs and trees
often. convenient to regard a graph r' as a subgraph of r merely if
r' is isomorphic to a sub graph of r in the sense just described.
Let rbe a graph. An edge e = {u, v} E Eris said to be incident with
u and v. A walk in r is a finite sequence of vertices and edges
(Vi E V r, ei E Er) such that e; is incident with V; and V; _ j. Successive
edges are therefore incident with a common vertex. The vertex Vo is
the initial vertex of the walk and Vj is its final vertex. A walk is closed
if its initial and final vertices are equal. The graph r is connected if,
given a, bE V r, there exists a walk in rwith initial vertex a and final
vertex b.
A walk V
o
, ej, Vj, e
2
, ... ,Vl j, e/, VI is a path if it contains no
subsequence u, e, v, e or e,v, e, u with u, V E V r, u:;r. v, and e E Er.
Such a subsequence introduces backtracking into the walk. It is
convenient to regard two walks as related if they differ only by such
backtracking, and so we call two walks equivalent if one can be
obtained from the other from some finite succession of insertions or
deletions of backtracking subsequences u, e, v, e or e, v, e, u with
u, v E V rand u :;r. v. Clearly any walk is equivalent to a path, and if
a walk is closed then so is any other walk equivalent to it. This idea
of equivalence will be generalized to the setting of topological spaces
in Chapter 9, to give the important notion of homotopy, and we shall
return to it in Chapter 12 in order to apply graphs to the study of
free groups. A walk in a quiver Q is defined to be a walk in the
underlying graph 1 Q I. This definition implies that a walk in a quiver
may include an edge that is traversed in the 'wrong' direction; that
is, a walk may include a subsequence of the form tee), e, see). If an
edge is traversed one way and then immediately the other way, then
the walk contains a subsequence see), e, tCe), e or tee), e, see), e which
may be deleted in an equivalence.
Fig. 7.4
Graphs, quivers, and trees 167
A closed path in a graph Tis called a circuit. The graph on the left
in Fig. 7.4. contains many circuits, whilst that on the right contains
none.
A graph is a tree if it is connected and contains no circuits.
7.1
1. Find which of the three graphs in Fig. 7.5 are mutually isomorphic.
Fig. 7.5
2. Show that the graphs in Fig. 7.6 are not isomorphic.
Fig. 7.6
3. Sort out the graphs in Fig. 7.7 into families of isomorphic
graphs.
4. A simple graph has at most one edge joining a given pair of
vertices. How many nonisomorphic simple graphs are there
with at most four vertices?
5. Translate the definition of isomorphism of quivers from the
terms of the second definition of quiver to those of the first.
6. Show that the two quivers in Fig. 7.8 are not isomorphic.
7. Let V be a set and ~ a relation on V. Define a quiver Q as
follows: the vertex set VQ is V and EQ = {(a, b) I a, b E Q and
a ~ b}. Hence, given any a, b E VQ, there exists at most one
edge in EQ from a to b. A quiver with this property is called
relational. What properties does Q possess if ~ is (i) reflexive,
(ii) symmetric, (iii) transitive, (iv) antisymmetric?
8. Formulate the definition of a subquiver of a quiver.
phs and trees
ABCDEFGHI
J'KLMNOPOR
STUVWXYZ
Fig. 7.7
Fig. 7.8
(a) (b)
(c)
Fig. 7.9
9. Which of the graphs in Fig. 7.9 are trees?
10. Show that a tree is a simple graph. Prove that a graph is a tree
if and only if there exists a unique path between any two
vertices.
II. How many nonisomorphic trees are there with at most four
vertices?
Examples of graphs 169
mples of graphs
(a) The nleafed rose Xn
Xn is the graph (unique up to isomorphism) with one vertex and n
edges (Fig. 7.10).
o
8
Xl
Fig. 7.10
(b) The complete graph Kn
The complete graph Kn has n vertices and n(n  1)/2 edges, with
exactly one edge joining any two vertices (Fig. 7.11).
Fig. 7.11
(c) The complete bipartite graph Km. n
A graph r is bipartite if V r is the union of two nonempty disjoint
subsets VI and V
2
such that every edge of ris incident with one vertex
of VI and one vertex of V
2
(Fig. 7.12).
If every vertex of VI is joined to every vertex of V
2
by exactly one
edge then r is a complete bipartite graph, and is determined up to
Fig. 7.12
Graphs and trees
K
3
•
3
Fig. 7.13
isomorphism by the sizes of the sets VI and V
2
• If I VI I = m and
! V
2
! = n we denote the complete bipartite graph with vertex set
V= VI U V
2
by K
m
.
n
(Fig. 7.13).
(d) Cayley quivers
The Cayley quiver of a group presentation appeared in Chapter 6.
Cayley quivers have the feature that their edge sets are partitioned
into subsets, with one subset for each generator in the presentation.
The next result explains a method of encoding the partition of edges,
and their direction, in a graph (and not a quiver!).
emma. To each Cayley quiver C there corresponds a graph rCfrom which
C may be reconstructed.
Proof. Fix an ordering of the generators in the presentation, and
for each K 1 do the following. If x is the kth generator, replace
each edge e in C labelled by x by a path of three edges and two new
vertices e' and e":
e' e"
s(e) t(e)
e' e"
s(e)
Planar graphs 171
Add a path of length 2k  2, with initial vertex e', but otherwise
consisting of new edges and vertices, and a similar path of length
2k  I with initial vertex en.
The new graph is rc, and it is easy to see that it has the property
claimed in the lemma. •
ar graphs
A graph r is planar if it has a diagram that may be drawn in the
plane without edges intersecting (except at the vertices!). We have
seen several examples of planar graphs, and some that you may
suspect are not planar. The basic fact about planar graphs is that
they satisfy Euler's formula, which dates from 1750 and relates the
number of vertices and edges in the graph to the number of regions
or faces into which a diagram of the graph divides the plane.
For the remainder of this chapter we shall assume that graphs are
finite; that is, for any r we assume that vr and Er are finite sets.
Recall that a graph is connected if there exists a walk between any
two vertices. Now any graph consists of a collection of connected
graphs called its connected components, so a connected graph has
just one connected component.
's formula for planar graphs. Let r be a planar graph with v ver
tices, e edges, and having c connected components. Suppose that a
diagram of r divides the plane into f faces. Then
ve+f=c+1.
Proof. We prove the formula by induction on e. If e = 0 then rhas
as many components as vertices so that v = c, and clearly f= 1.
Therefore the formula holds.
Now suppose that rhas e> 0 edges, and let r' be obtained from
r by removing one edge. Certainly r' is planar, and has v vertices
and e  1 edges. Further, r' has c' connected components, where
either c' = c or c' = c + 1. From a diagram of r drawn in the plane
we obtain a diagram of r' dividing the plane into f' faces, so that by
inductive assumption, v  e + I + J' = c' + 1, that is v  e + f' = c'.
First suppose that c' = c; that is, the edge removed from r does not
separate a connected component of r into two components of r'.
This edge either joins two distinct vertices of r, or else is an edge of
the form {u, u}, and in either case its removal from r reduces the
number of faces in the plane by one; that is, f' = f  I and
v  e + f = C + I as required. If the edge removed does separa te
Graphs and trees
pies
a component of r into two components of r', so that c' = c + I,
then this edge cannot be on the boundary of a finite face in the
diagram of r and so occurs on the boundary of the infinite face. Its
deletion does not then change the number of faces, whence/, =Jand
again v  e + J = c + 1. •
We may verify the application of Euler's formula to the two graphs
shown in Fig. 7.14.
(b)
Fig. 7.14
In Fig. 7.14 (a), v = 5, e = 7,J= 5, and c = 1. In Fig. 7.14. (b),
v = 7, e = II,J= 7, and c = 2.
orollary. Let r be J connected simple graph with v vertices and e edges,
where v ~ 3 and is finite. If r is planar then e ~ 3v  6.
Proof. In any diagram of r, every face has at least three edges in
its boundary, and every edge occurs in the boundary of at most two
faces. Therefore 3J ~ 2e and this inequality combines with Euler's
formula to give the result. •
ises 7.3
1. The most important case of Euler's formula is that for c = 1.
Show that a graph is planar if and only if each of its connected
components is planar.
2. Prove that K5 is not planar.
3. Prove that K3,3 is not planar.
4. The most celebrated result about the planarity of graphs is
Kuratowski's theorem (Kuratowski 1930). We say that two graphs
rand r' are isomorphic modulo vertices oj degree 2 if r is
isomorphic to a graph r" obtained from r' by the addition or
deletion of vertices with just two edges incident:
o ~     ~ o ~     ~ o o ~             ~ o
Embedding graphs in surfaces 173
Give a precise formulation of this notion in terms of V r', V r",
Er' and Er". Kuratowski's theorem is as follows.
Theorem. Let r be a finite graph. Then r is planar if and only if it contains
no subgraph isomorphic modulo vertices of degree 2 to K5 or K
3
.
3
•
We shall not prove this theorem here, but you are urged to consult
an extended discussion, such as that in Wilson (1985).
Embedding graphs in surfaces
The embedding of a graph in the plane is just one part of the more
general procedure of embedding graphs in surfaces, as the following
observation shows.
Proposition A graph r is planar if and only if it can be embedded in the
sphere S2.
Proof. Stereographic projection (see Fig. 7.15) carries plane
embeddings to embeddings in the sphere, and vice versa.
"
Fig. 7.15
It is now a natural step to consider the embeddings of graphs in
other surfaces. If a graph r is not planar, how complicated a surface
must we take in order to be able to embed r? That is, how large must
the genus be? Clearly, the larger the genus required, the more
complicated is the combinatorial structure of r. Note that for the
moment it suffices to discuss only simple graphs, for if we are given
a rthat is not simple, then we may construct To by replacing multiple
edges with single edges, so that ro is simple, and embeds in a surface
S if and only if r does.
Denote by Sg an orientable surface without boundary and of genus
g. We may picture Sg as a sphere with g handles attached. A graph
phs and trees
rwith g edges then embeds in Sg; draw one edge around each handle.
So K5 embeds in SIO. This is a rather extravagant waste of genus, for
we shall see that Ks may be embedded in Sb the torus.
A couple of questions now arise, answers to which would shape the
theory of graphs embedded in surfaces quite nicely:
1. Which graphs embed in Sg for fixed g?
2. Given a graph r, what is the smallest g for which Tembeds in Sg?
If r embeds in Sg then it certainly embeds in Sh for any h ;?: g, so
that an answer to question 2 seems more fundamental. An ideal
answer to question I would be a version of Kuratowski's theorem
(see Exercise 4 of Section 7.3) that characterized the graphs embed
dable in Sg for each g ;?: o. Yet far from having ideal answers, we
have only partial ones, continually improved and updated by re
search work. The books by Gross and Tucker (1987) and White
(1984) are excellent places to look for more detailed information.
Question 2 does have a theoretical solution, as a consequence of a
general method for constructing an embedding of a graph in a
surface. This method is known as the Edmonds algorithm, named
after J. Edmonds who described it in 1960.
Let Tbe a connected simple graph with finitely many vertices. For
notational convenience we shall suppose that VT= {l, 2, ... , n}.
NowforeachiE Vr,set V(i)={jE VTI{i,j}EET};thatis, V(l} is
the set of vertices connected to vertex i. Let I Veil I = ni and for each
i choose an nrcycle Pi permuting the set Vel). For example if
V(1) = {4, 6, 7}, we could take PI = (6 4 7). Now let D = {(i, j) I
{i ,j} E E r}; hence D is the set of all ordered pairs of vertices
connected by edges in r. The choices of the permutations Pb ... ,Pn
determine a permutation P of D defined by P(i, J) = (j, Pj(i). If
(i, j) E D then i E V(j) and so pii) E V(j). Therefore {j, p;(l)} E ET
and the orbits of P consist of edges of rthat fit together into circuits
in r. We represent such a circuit by drawing it around a polygon; a
vertex of such a polygon corresponds to a vertex of T and so can be
labelled with an element of {I, 2, ... , n}; the edge from vertexj to
vertex p;(i) is labelled with the ordered pair (j, Pj(i». Having labelled
all the polygons in this way, we can paste them together by matching
an edge (a, b) with an edge (b, a) to form a surface in which Tis
embedded.
We show that K3.3 may be embedded on the torus SI (Fig. 7.16).
Embedding graphs in surfaces 175
2
4 5
Fig. 7.16
V(Ku) = {I, 2, 3, 4,5, 6}
V(l) = V(2) = V(3) = {4, 5, 6}
V(4) = V(5) = V(6) = {l, 2, 3}.
We will take the permutations
PI = P2 = P3 = (4 5 6)
P4 = Ps = P6 = (l 2 3).
3
6
Thus P(1, 4) = (4, 2), P( 4, 2) = (2, 5) and continuing the calculation,
we find three orbits:
(1,4) + (4, 2) + (2, 5) + (5, 3) + (3,6) + (6, 1).
(1, 5) + (5, 2) + (2, 6) + (6, 3) + (3, 4) + (4, 1)
(1,6) + (6,2) + (2, 4) + (4, 3) + (3,5) + (5, 1).
This gives us three hexagons that can be pasted together to form a
surface embedding of K
3
• 3' At this stage we can work out the genus
of the surface, for the embedding of K
3
• 3 subdivides the surface into
the three regions given by the orbits, with nine edges and six vertices.
Therefore the surface has Euler characteristic
X=69+3=0
and genus g = 1, and so it is indeed a torus. We can fit the hexagons
together to make the usual gluing instructions for the torus, with
K
3
.
3
embedded:
4 2
5 __ l
6 3
raphs and trees
Any embedding of a graph r in a surface S constructed by the
Edmonds algorithm has the property that S\r is a union of disjoint
regions each homeomorphic to an open disc; these regions arise from
the interiors of the polygons around which the circuits are drawn.
We call an embedding with this property a disc embedding. Now
given a disc embedding of r in S, the boundaries of the regions of
S\r are circuits in r and the succession of the edges in any given
circuit can be read as a set of ordered pairs of vertices, that is as a
subset of D. Since S is assumed to be orientable, neighbouring
regions can be oriented coherently and their boundaries read accord
ing to the orientations, partition the set D. The parts of the partition
are orbits for a map P*: D ~ D from which permutations Pi for each
i E V r can be found. These considerations show that every disc
embedding of a graph rin an orientable surface S can be constructed
by the Edmonds algorithm.
If a graph r can be embedded in an orientable surface of genus g,
but not in one of genus h < g, we say that r has genus g. An
embedding of r into a surface of genus g is then called a minimal
embedding. We have seen that Ks has genus I, as does K
3
,3' The
planar graphs are precisely those of genus O.
position. A minimal embedding of a graph r in an orientable surface S
is a disc embedding.
Proof. Suppose that some circuit in r bounds a region in S not
homeomorphic to an open disc. We may assume that the region of S
contains no edges of r (or else we take a circuit containing fewer
edges until we find such a region), and we cut S along the circuit. If
it happens that S is thereby cut into components, discard the
component containing no edges of r and attach a disc around the
circuit in the other component. This creates a new surface S' of
smaller genus than S in which r is embedded. Hence the original
embedding was not minimal, and this is a contradiction. If S is not
cut into two components by cutting along the chosen circuit, then
attach disc around each hole caused by the cut, again to form a new
surface with smaller genus and with r still embedded, and so arrive
at the same contradiction. •
Given a graph r we can theoretically use the Edmonds algorithm
to construct all the disc embeddings of rinto orientable surfaces, and
if the smallest genus that arises is g then we know from Proposition
7.6 that r has genus g. The practical difficulty is that there a huge
number of embeddings constructed by the Edmonds algorithm,
corresponding to the choices of permutations to be made, so that it
Embedding graphs in surfaces 177
is infeasible to calculate the genus of a graph by this method.
However, advanced techniques have been applied to calculate the
genus for some special classes of graphs, and we quote two cases
here. The symbol I X l denotes the smallest integer greater than or
equal to X.
heorem. The complete graph Kn (n ~ 3) has genus I(n  3)(n  4)/121, and
the complete bipartite graph Km. n has genus I(n  2)(m  2)/4l. •
These results are due to G. Ringel and J.W.T. Youngs; see White
(1984) for a fuller discussion.
ises 7.4
1. Attempt to embed Ks in 51; that is, try to draw a diagram of Ks
on the torus without edges crossing. Represent the torus by the
gluing instructions.
2. Try a different choice of permutations in the example given
above for embedding K
3
,3 on the torus. For example, take
PI = (4 5 6), P2 = (4 6 5) = P3, P4 = (1 2 3) and Ps = (1 3 2) = P6'
3. Use the Edmonds algorithm to construct an embedding of Ks on
a torus.
4. What happens if you apply the Edmonds algorithm to a planar
graph? Does an embedding in the sphere always result? Try
K
4
, K
2
,2 and K
2
,3'
5. Prove the following generalization of Euler's formula. Let Tbe a
connected graph with u vertices, e edges (where u and e are finite).
Suppose that a disc embedding into a surface 5 of genus g is such
that 5\F is a union off regions. Then u  e + f = 2  2g.
8
Alexander matrices and Alexander
polynomials
Although the method that we have used to calculate the Alexander
polynomial of a knot only uses the theory of determinants, the
trouble with it is to know why it works. It is easy to check that the
result is invariant under the Reidemeister moves but it would be nice
to link it in with other invariants. 'It would be nice' sounds a weak
reason, but often when one can do something by a different method,
it is possible to see how to 'squeeze' the method a bit to get out
more detailed information. What we will do is give a second method
of calculating Alexander polynomials, this time using presentations of
groups. The payoffs are (i) methods that can be applied to other group
presentations which we will use later in Chapter II; (ii) intermediate
stages and related constructions that show more of the fine detail of
the situation even in the case of knots; and (iii) new invariants that
have important geometric significance although in this book we will
not be able to do justice to this aspect.
To do all this will require new algebraic ideas. These in turn have
exciting consequences within algebra but we will not be able to
explore these fully.
up rings
For any group, G, consider the set TLG (where, as usual, 7L denotes
the ring of integers) of all formal sums of the form
where each ng E 7L and only finitely many of the ng are nonzero.
Examples of such expressions are gl  2g
2
, 4g
3
+ 3g
4
 6g
s
, and so on.
The set TLG has a natural structure of an Abelian group where
addition is given by
Group rings 179
(As only finitely many of the ng and mg are nonzero, the same is
true of the (ng + m
g
), so the righthand side is an element of 7LG.) If
we think of the set I G I of elements of G, that is we forget
the multiplication of G, then TLG is the free Abelian group on this
set, I GI·
Remembering the multiplication of G, we get a multiplication on
7LG, which makes it into a ring. Explicitly we have
The result is not quite in the right form but if we write for kEG,
Pk = L g G k' then we can partition up the sum on the right
hand side to get Lh G Pkk. We leave it to you to check that Pk will be
nontrivial only for finitely many k in G. This ring TLG is called the
group ring of G.
One of the most useful examples is when G = the infinite cyclic
group generated by (, say. A typical element of is therefore (' and
a formal sum can be thought of as a polynomial in t and r I or as a
Laurent polynomial
where j:%; k are integers. For instance, r I  1 + t is an element
of Another example is when G = C
2
= {I, a} with a
2
= 1; then
7LG = {nl + n2a : nj, n
2
E 7L} where, for instance, (4a)(1  a) =
4a  4a
2
=  4 + 4a. It is clear that TLG is, in general, a commutative
ring if and only if G is an Abelian group.
The group G can be thought of as sitting inside 7L G via the function
g H 19, that is the element Lnhh where ng = 1 and nh = 0 if h =#= g. Any
function fjJ: G A, with A an Abelian group, extends uniquely to an
Abelian group homeomorphism from TLG to A. If fjJ: G A is such
that A is a ring and fjJ preserves multiplication then the extension to
7LG is a ring homeomorphism. Using this we can see that any group
homeomorphism G H composed with the inclusion of H into
7LH extends uniquely to a ring homeomorphism from TLG to 7LH.
This last result will be useful in two cases in particular:
(1) s'l: G Gab, the Abelianization homeomorphism;
xander matrices and Alexander polynomials
(2) £: G ~ {I}, the trivial homeomorphism. The resulting ring
homeomorphism sends Lngg to Lng, and is called the augmenta
tion homeomorphism.
There is a whole class of problems which involve the transfer of
properties from G to 7LG. For instance, we might ask for what groups
G is TLG an integral domain or a unique factorization domain? (If
you have not met integral domains or unique factorization domains,
now would be a good time to look up the definitions. You will find
them in almost all the usual algebra textbooks.) We will review
briefly some ideas of ring theory for convenience:
Let R be a ring with an identity element, 1.
A unit in R is an element u E R with an inverse, that is there is
some v E R with vu = uv = 1. For instance, in a group ring the
element 1 is the formal sum nIl = Lngg where ng = 1 E 7L if g = 1 E G
and ng = 0 otherwise, so any element g E G gives 1 g E 7L G and
(lg) . (1 gI) = 1. Thus any element of G considered as an element of
the ring 7L G is a unit. Can there be units other than those of the form
g or ( l)g in 7LG?
We call two elements a, b in a commutative ring R associated if a
divides b (written a I b), so there is some C E R with b = ac, and
b divides a. If R is an integral domain (so if ab = 0, one of a and h
must itself be zero), then if a and b are associated, there is a unit
u with a = ub. (We leave the proof to you.)
An element d of a commutative ring R is called a greatest common
divisor (g.c.d.) of a finite set of elements a
h
... , an E R if d divides
each ai, i = 1, ... , n, and if c I ai for i = 1, ... , n then c I d. Any two
g.c.d.s of a given set are therefore associated and so if R is an
integral domain, the g.c.d. of a set of numbers is determined up
to multiplication by units. For instance, both 2 and  2 are g.c.d.s of
{ 6, 4, 24,  26},  2 = ( 1) X 2, and  1 is a unit in 7L. We say R is
a g.c.d. domain if it is an integral domain in which any finite set of
elements has a g.c.d.
Recall 'R is a unique factorization domain' means that it is an
integral domain with 1 in which every x E R, x :t= 0 or 1, has a de
composition as a product of prime elements, that is unique up to
choice of order. Any unique factorization domain is a g.c.d. domain
as we can use essentially the same method as we might use for finding
g.c.d.s in 7L given prime decompositions (e.g. g.c.d. (201, 33) = 3 since
201 = 3 x 67 and 3 = 3 x 11, so 3 is the only prime power common to
both).
If Q is a subring of R, and there is a mapping p: R ~ Q preserving
multiplication and such that a and pea) are associated for any a E R,
then
(l) if Q is an integral domain, then so is R;
(2) if Q is a g.c.d. domain, then so is R.
Group rings 181
The proofs are quite easy. In each case, to verify a property in R
map the test elements back down into Q, verify the property for the
images, and then complete since always a and p(a) are associated.
(Write out a complete proof yourself.)
Before returning to group rings we need to import two results from
ring theory, namely (i) that if R is an integral domain, so is the
polynomial ring R[t] and the only units in R[t] are those within the
subring R (considered a subring via the usual embedding as constant
polynomials), and (ii) if R is a unique factorization domain, so is
R[t}. (Proofs of these results can be found in many standard books
on ring theory if you wish to check them.)
The appearance of R[t] in a section on group rings should give a
hint as to which group ring is going to be in the limelight for the
next few pages. We have already met :EC= as a ring of Laurent poly
nomials and clearly :E[t] is a subring of :E C
w
Are we in the situation
above? Is there a multiplication preserving mapping p: :E C= ~ :E[t}
such that a and p(a) are associated for each a E :E C=? Any such a has
the form
n=oo
where only finitely many of the an are nonzero. Define, for a *" 0,
p(a) to be the smallest n such that an *" 0, so that for example
p(l) = 0 and p(t
2
 21 + r 24) =  24. If a = 0 then it will be conveni
ent and consistent with our usage of p to define p(O) = 00. We can
easily check (do it) that if a, b E :E C=,
p(ab) = pea) + pCb).
(The case where a or b is zero needs care but is fairly obvious. The
other cases are quite simple. Write out a proof.)
Now define pea) = at
J1
(a). This gives a normalization of a, changing
it into a polynomial. Of course, a and p(a) are associated, since
r
J1
(a) is a unit unless pea) = 00. We make the convention that r= = O.
You should check that p preserves multiplication.
A consequence of this is the following result.
The group ring :E C= of the infinite cyclic group is a g.c.d. domain.
We can, in fact, say more. The units in :EC= are all the elements of
the form ± tn, n E :E. This will be of use when working out Alexander
polynomials.
182 Alexander matrices and Alexander polynomials
You may wonder why we introduced a general ring R in our earlier
discussion. The reason is that a similar argument would have worked
for the group ring of any finitely generated free Abelian group. We
will not need this in detail, but note the result. You may like to try
to investigate this situation and to prove that ~ ( C ~ E9 ... E9 C ~ ) is
always a g.c.d. domain whose only units are group elements and their
negatives.
8.2 Derivatives
R. H. Fox, one of the 'founding fathers' of modern knot theory,
developed a theory of 'derivatives' (also called derivations) which is
a powerful tool in the study of group presentations.
A derivative on a group G is a mapping D: G ~ ~ G such that
D(gh) = D(g) + gD(h).
Example and Exercise
The augmentation homeomorphism
e: ~ G ~ ~
sends Lngg to Lng. The kernel of e contains such elements as g  1.
Prove that if G is finite,
(a) Ker e is a free Abelian group with basis
{g  1: g € G}
(b) the mapping from G to ~ G sending g to g  1 is a derivative.
These two properties are very important in the study of the
homology groups of a group. In this, one tries to apply methods based
on topological ideas to study algebraic objects. At several times in the
following chapters we will touch on material related to this area.
Given any derivative D: G ~ ~ G , there is an extension to a
mapping D: ~ G ~ ~ G given by D(Lngg) = LngD(g). This is additive
D(g + h) = D(g) + D(h)
but is not a ring homeomorphism.
8.1 Lemma.
(a)
(b)
If D: G ~ ~ G is a derivative, then
gn _ 1
D(gn) = D(g) for all n € ~ .
g  I
Derivatives 183
The proofs are by easy algebraic manipUlation using the definition
D(gh) = D(g) + gD(h) and are left as an exercise.
The case when G is a free group, F = F(X), is especially useful. In
this case, to each free generator x of F there corresponds a unique
derivative written
defined by
a.
ax' 7LF
ay _ {I if x = y
ax  0 if x =1= y.
This defines a/ax on generators. You should check that there is no
difficulty in extending it to the whole of F. You can do this by
remembering that elements of F are equivalence classes of words and
then checking that the 'obvious' extension to aw/ax works in such a
way that equivalent words have the same image.
We will take X = {x, y, z}, F = F( X), the free group on X, and will
look at aw/ax where w = xyl X
2
ZX
1
,
= + x
2
zx') = 1 + + xy' (x
2
zx').
ax ax ax ax ax
We note that ax
2
/ax = 1 + x, so this expression equals
I + xyl ((1 + x) + x
2
aa
x
zx,)
= 1 + xyl ((1 + x) + x
2
}xz + x
2
z aa
xx

1
)
= 1 +xyl (1 +x+x
2
Z(X'»)
which can now be simplified. (Caution: since we are in TLF and F is
free, X
2
ZX
1
is not the same as xz since the generators do not
commute.)
8.2
1. Working in F(x, y), the free group on {x, y}, calculate a/ax and
a/ay of the following words:
(a) xyxy' x1yl
(b) (xytx(xYrn, for n E N
exander matrices and Alexander polynomials
(c) [xm, yjP, for m, n, p € N.
2. Investigate the following 'short cut' and its generalizations.
Suppose we want to calculate dw/dx and we can decompose w as
W
1
X
n
W
2
where neither Wj nor W
2
involves the generator x; then
dx x  1
3. Prove that if F is a free group on a set X and D is any derivative,
then D is completely determined by the values D(x) for x € X that
it takes on the generators. Knowing these values a formula of the
form
D(w) = L ax<w)D(x), aAw) € 7l.F
XEX
can be found. What are these coefficients axCw)?
4. We suggested earlier that you proved that defining d/dx on the
generators of F was sufficient to define it on all of F. This is a
bit like the proof that any mapping defined on the generators
extends to a homeomorphism of F. Using the earlier exercise on
Ker E, and the fact that it is generated freely by the set of x  1,
X € X, it is possible to find an elegant proof of the 'extension of
derivatives' result using just the freeness of F and of Ker E. Try
it out, and see what you can do.
Alexander matrix
Suppose we have a group presentation (X : R) of a group
G == F/«R». We will only need the case when X and R are finite and
so will assume this is true throughout. We write X = {xI> ... , x
n
},
R = {r
j
, ••• , rm}. Furthermore we will be a bit careful and will
suppose all our relations are of the form r = I (i.e. we rewrite a
relation r = s as rs
j
= 1). We will write
for the quotient map and r: 7l.F 7l.G also for its 'extension' to the
corresponding group rings. Similarly we write .sil: G Gab for the
natural epimorphism Abelianizing G.
We form a 'Jacobian matrix'
J= '
(
dr.)
dXj
5
The Alexander matrix 185
with coefficients in 7LF. In many ways TLF is like a Laurent polyno
mial ring in noncommuting variables X" •.• , Xw Partially to make
life simpler (and there, as usual in mathematics, we run the risk of
simplifying the interesting detail out of existence) and partially
because we are really more interested in G or Gab rather than in F,
we use yand :A to take these coefficients in J down to 7L Gab' As this
is a commutative ring, we can work more easily here.
We set
This is the Alexander matrix of the presentation. The definition
perhaps looks slightly technical to you, but the various steps are
easily followed through as you will see when we do some examples.
The significance of yis that it allows us to simplify using the relations
in Rand .'il allows us to simplify even further.
1. G = S3 = (a, b : a
3
= b
2
= (abf = 1).
We take F= F(a, b), the free group on {a, b}, and relations
rl = a
3
, r
2
= b
2
, r3 = (abi,
d I 2
rl = +a+a
da
d
r2 =0
da
d d
da r3 = [1 + (ab)] da (ab)
= (l + ab)
so our 'Jacobian matrix' is
d
r
l
=0
db
d
dbr2 = 1 + b
d d
db r3 = [1 + (ab)] db (ab)
= (1 + ab)a
(
1 + a
O
+ d 0 J
I + b .
I + ab a + aba
Applying y to the coefficients means that a
3
, b
2
, and (abf become 1;
then applying the Abelianization and writing a for .'il(a) we find that
abab = 1 implies a
2
b
2
= 1; hence a
2
= 1 and a
3
= 1, so a = 1, that
is (S3)ab == c 2 = (t : t
2
) with generator t corresponding, under the
isomorphism, to b. Thus we obtain an Alexander matrix
xander matrices and Alexander polynomials
A =(
I + t
since s1(a) = 1.
Before considering another example, we will look at a computational
hint which saves time and diminishes the chance of errors, especially
when relations are long.
Consider a relation of the form r = s; this normalizes to give
rs
I
= 1 and if x is a generator, we have
(rs
I
) = ar + r
ax ax ax
= ar _
ax ax
However, when we calculate the Alexander matrix of the presenta
tion we first apply y: F and in y(rsI) = I, so
y (rs ) = y    .
(
a_I) (ar as)
ax ax ax
This implies that we can compute the derivatives of rs
I
by using
r  s in the group ring 7LF. If rand s are of moderate length, rs
I
is
long and it is easy to make an error in calculating a(rsI)/ax. For
instance, in the cinquefoil knot group presentation, we get
(x, y : xyxyx = yxyxy)
or
(x, y: xyxyxyIxIyIxlyI = I).
In fact, here the derivatives of this single relation can be handled
in several different ways, but one of the best is to calculate the
derivatives of xyxyx  yxyxy in 7LF(x, y).
2. G = (x, y : x
2
= I, y3
xy
2
X
I).
The Abelianization Gab will have y = I and we will put s1 y(x) = t
(t
2
= I in Gab, which is isomorphic to C
2
).
Put r = x
2
so
a
axr=l+x
Put s = yxy2xl; we use the calculation via y3 x  xy in 7LF(x, y)
to get
Similarly
ses 8.3
Elementary ideals 187
.sd r (dd
X
S) = .sd r (dd
X
(y3 x)  dd
X
(x y) )
= .sdr(y3  I) = o.
s:tr(aayS J = s:tr((l + y + y2)  X(l + y»)
= 3  21.
I. If you have not already done this, check that if G is a knot group,
Gab is an infinite cyclic group with one generator, t. (Although
you can use either the Dehn or Wirtinger presentations to obtain
this, the Wirtinger presentation has the advantage that all the
generators map to t under the composite .sd y.)
2. Check that (a, b : a
P
= bq)ab is infinite cyclic if the highest com
mon factor of p and q is 1. (This exercise is a repeat.)
3. Calculate the Alexander matrices of the following presentations:
(a) G(trefoil) = (x, y, z : z = xyxI, x = yzyI, Y = zxz
I
).
(b) G(cinquefoil) = (x, y : (xyi x = y(xy/).
(c) If n = 2m + 1, and t
2
,n is the (2, n)torus knot,
G(t
2
, n) = (x, y : (xy)mx = y(xy)m).
(d) Again as in (c), G(t
2
, n) = (a, b : an = b
2
).
(Beware, you need to do the preceding exercise first.)
lementary ideals
The Alexander matrix A of a presentation is not invariant under
Tietze transformations since even its size can change if a generator
is added or removed. What sorts of quantities could we expect to
construct from A that might be invariant? One way to guess the sort
of answer to expect to such a question is to see how an Alexander
matrix of a given presentation changes as we apply Tietze trans
formations. (We suggest that you do this investigation yourself. That
way you can adjust the complexity of the example you tackle to your
present understanding of the question. If you feel you understand
quite well, it would be a good idea to consider the completely general
question of how the Alexander matrix of (XI' ... , Xm : rl> ... , r2)
Alexander matrices and Alexander polynomials
pies
changes under the various transformations. Whichever way you go
about this, compare the transformations that result with those you
know from linear algebra.)
Alexander matrices are matrices over group rings, 7lG
ab
• You are
most used to matrices over IR! or C and may also have met, for
instance when handling presentations of finitely generated Abelian
groups, matrices over 7l. To keep things fairly general for the
moment, we will start by working over a general commutative ring
R. If S is a subset of R, the ideal E generated by S is the set of all
finite sums Larsi' ai € R, Si € S. Formally an ideal behaves like the
normal subgroup generated by a subset of a group.
Let A be an m x n matrix over R. For k, a nonnegative integer,
the kth elementary ideal Ek(A) of A is:
(1) if 0 < n  k ~ m, the ideal generated by all (n  k) x (n  k)
minors of A.
(2) if (n  k) > m, Ek(A) = O.
(3) if 0 ~ n  k, Ek(A) = R, the whole ring.
Remarks
(a) In case you have not met the term (or have need to recall it) an
r x r minor of A is the determinant of an r x r submatrix of A.
(b) The slight awkwardness in the definition is needed because if we
merely said that Ek(A) is generated by the (n  k) x (n  k)
minors of A then what could be the meaning of, say, a
 2 x  2 minor or a lOx 10 minor in a 4 x 5 matrix? Because
of this, we had to assign meanings to Ek(A) when r  k is small
or larger than m.
(c) Since any (r + 1) x (r + 1) minor can be expanded as a linear
combination (over R) of r x r minors of A, these ideals are nested:
o ~ Eo(A) ~ El(A) ~ ... ~ En(A) = R.
1.
A = [ ~
1 + t
1 ~ tl'
I + t
Eo(A) is generated by 2 x 2 minors:
I ~ l ~ t l ' I l ~ t ~ : ; I ' 11:'t
I + tl
1 + t '
Elementary ideals 189
hence by 3(1 + I),  (1 + 1)2, 3(1 + I).
We have not stated over which ring we are working but A is, in
fact, the form of the Alexander matrix of the usual presentation of
S3 and (S3)ab is cyclic of order 2, so we will assume R = 7L C
2
generated
by 1 with t
2
= 1. This gives us that Eo(A) is generated by 3(1 + t) and
 2(1 + t) (we write Eo(A) = (3(1 + t),  2(1 + t».
Since (l + t) = 3(1 + t)  2(1 + t), we get
Eo(A) = (1 + t).
El (A) is generated by 1 x 1 minors and hence by the entries of A. We
have
El (A) = (3, 1 + t) over 7L C
2
•
We know
Are both of the inclusions proper or is one an equality? If
Eo(A) = E( (A) then 3 E (l + t). This would mean there was some
A., fl E 7L such that (A. + flt)(l + t) = 3. (We must remember that we are
working in 7L C
2
, so all elements have the form a + pt for a, p E 7L,
and on multiplying, we put t
2
= 1.) Multiplying out our conjectured
eq uation gives
(A. + fl) + (A. + fl)t = 3
which is silly since A. + fl cannot be both 3 and O. This shows that
3 e Eo(A) and the first inclusion is proper. Similarly one knows that
1 E 7L C
2
and we can easily adapt the above to show that 1 e EI (A).
2. G = (x, y : r = 1, y3 X = X y2).
As we saw Gab == C
2
and the Alexander matrix is
over 7LC
2
as before.
Eo(A) is generated by 2 x 2 minors, that is by the determinant of A:
(l + 1)(3  2/) = 3 + 3 t  21  21
= 1 + I.
El (A) is generated by 1 x 1 minors 1 + 1 and 3  2/, but as
(1 + t) = (l + t)(3  2t), EI (A) = (3  2t).
We get a chain
o k (1 + t) <;;;;; (3  2 t) <;;;;; 7L C
2
•
xander matrices and Alexander polynomials
This is a different chain of ideals than that in Example 1 since
3 ~ (3  2/). (Check for yourself.)
We can now see how we might use these chains of elementary
ideals. Let us suppose we can prove that the chain of elemen
tary ideals is an invariant of the group G and does not depend on
the presentation used. We have above presentations
(a, b : a
3
= b
2
= (ab)2 = 1)
of S3' and
(x, y : x
2
= 1, ~ x = X y2)
of our group G of example (2). Is G == S3? We know that Gab ==
C
2
== (S3)ab. SO Abelianization does not help, but we also know that
if G == S3, then we could find Tietze transformations linking the two
presentations. This in its turn would imply that the two presentations
would have the same chains of elementary ideals. This last statement
is false, so G ?/= S3'
In this case, it would be possible to prove that G was not S3 in
many other different (and probably quicker) ways. However, later on
we will need this technique to prove that certain groups arising from
surfaces are not free groups even though their Abelianization is free
Abelian. It is important to note that, as usual, the logic is: the invari
ants are different, and therefore the groups are nonisomorphic; we
cannot conclude that two groups are isomorphic if their elementary
ideals are the same.
Our next preoccupation must be to prove invariance of the element
ary ideals of a presentation.
rem. Any two (finite) presentations of a group G have the same chains
of elementary ideals.
Proof Since by Tietze's theorem, we can link the two presentations
by Tietze transformations, the proof reduces to checking that Tietze
transformations leave the elementary ideals unchanged.
Clearly we need only look at Tl (Adding a superfluous relation)
and T3 (Adding a superfluous generator).
Tl: Replace (X: R) by (X: R u {r}) where r E «R». Thus there are
elements ri E R, Ui E F(X) (possibly with repeats), such that
Then if x EX,
p
r = II u;rjui'·
i= ,
Elementary ideals 191
However, y(ri) = 1 for i = 1, ... ,p, so
Next we calculate these p individual terms:
Again we map this down by y to get
Thus the Alexander matrix of (X: R u {r}) is like that of (X : R)
except that it has an extra row which is a linear combination of other
rows. Hence for any k, the corresponding minors generate the same
ideal. (If you do not see this last step immediately try a simple
example and then try to prove that the statement is true.)
T3: Replace (X: R) by (Xu {g} : R u {wgI}). We denote the two
(isomorphic) groups by G and G', and the two Alexander matrices
by A and A' respectively. We assume R = {rio . .. , r
n
}; then as g is a
symbol not in X, dr/dg = 0 for all i. Of course the last relation
s = wg
I
will involve some at least of the x E X. The inclusion of X
into Xu {g} induces a homeomorphism
a: F(X) ~ F(X u {g}).
Dividing out by relations, this induces
a': G ~ G' = F(Xu {g})/«R u {wgI}»
(which is an isomorphism) and hence an isomorphism
It is clear that the (i, j)entry of A' is a" (aij) except if the entry is in
the last row or column. In fact our calculations show that
Alexander matrices and Alexander polynomials
A' = (;"(A) ~ )
for some row vector a'. We can use a column operation to eliminate
a' and get an equivalent matrix with zeros in those positions of the
bottom row:
A'  (a'ri
A
) ~ ) = A", say.
Now it follows from elementary arguments of linear algebra that
the «n + 1)  k) x «n + 1)  k) minors of An are of the form
where M
n

k
is an (n  k) x (n  k) minor of A, or are an image
a" (M"(n + I) k) of some «n + 1)  k) x «n + 1)  k) minor, or have
determinant zero. All possible M
n

k
can arise in this way, so
Ek(A") = Ek(A') = Ek(A). •
se 8.4
Show, using elementary ideals, that
G = (a, b : a
3
b = ba, b
2
a
2
= db)
is not isomorphic to
H = (x, y: r = 1, y3x = xy2).
lexander polynomials via the Wirtinger presentation
Given a group G and a presentation (X : R), we have found a matrix
A over lLG
ab
and a chain {Ek(A)} of ideals that are invariants of G.
As we saw, the ideals Ek(A) need not always be singly generated. If,
however, Gab is a free (finitely generated) Abelian group, then lL Gab
is a g.c.d. domain, so if Ek(A) is generated by elements
aj, ... , an E lLG
ab
, letting d be a greatest common divisor of
a), ... , am then as each of its generators is in the principal ideal (d),
Ek(A) ~ (d) and (d) is the smallest principal ideal with this property
(prove it!). The various d for the different Ek(A) form an important
and calculable set of invariants of G. Of course this holds in
particular if G = G(K), a knot group, as G(K)ab == C=.
Let A be the Alexander matrix of the Wirtinger presentation of the
knot group G(K) of a knot K. The kth knot polynomial .1
k
(t) is the
Alexander polynomials via the Wirtinger presentation 193
greatest common divisor of all (n  k) x (n  k) minors of A. For
k = I, Al (t) is the Alexander polynomial of K.
Earlier you should have found that the Alexander matrix for the
Wirtinger presentation of the trefoil is
 I
t  I
This has 2 x 2 minors (t  1)2 + t = t
2
 t + I and  [2 + t  I and
again ,2  t + I; thus AI (t) = t
2
 t + I.
Remark
We see that since g.c.d.s can only be determined up to multiplication
by units, the same is true for these polynomials Ak(t). This in part
explains why the combinatorial derivation of the Alexander polyno
mial had to be 'normalized' to be well defined. It also means that
'the' in the definition of Ak(t) is rather too strong as 4(t) is not
completely determined by the definition.
It often pays to simplify a Wirtinger presentation before working out
A and hence the Ak(t). An eightcrossing knot with seven generators
will have a 7 x 8 Alexander matrix if the presentation is not reduced
in size first; then AI (t) will involve evaluating all 7 x 7 minors of A!
The example we will look at in detail only involves five crossings and
five generators (Fig. 8.1), but even that would be bad enough.
a
b
Fig. 8.1
xander matrices and Alexander polynomials
Here G(K) = (a, b, e, d, e : d = alea, b = dIad. a = bIeb,
e = e I be, e = e
l
de). Simplifying the presentation, we obtain
(a, b, e, d: d = alea, b = dIad, a = bIeIdeb, e = eIdIebeIde) =
(a, e, d: d = aI ea, a = dIa
I
de1ded1ad, e = d
I
edIadeId)
= (a, e: a = eIaIeaeIaIeaeaIelae, e = aIelaeaIeIaeaeIaIea).
Examination of the two relations shows that they are conjugate as
one would expect, as we know that anyone relation in a Wirtinger
presentation is a consequence of the others. (This is a useful check
that we have made no errors.) Thus
(We have evened up the two sides of the relation to help with the
calculation.)
Take
so
Similarly
ar
ae =  e
I
+ eIa
I
 eIaIeae + eIa1eaeIa
1
Applying r and then .sil (in effect substituting a = e = t) we get
J=  2r2 + 3rl  2
whilst
2r2  3rl + 2.
Our Alexander matrix is
( 2r2 + 3rl  2 2r2  3t + 2).
We will 'normalize' Al (t) to have no negative powers of t and positive
constant term, so
Alexander polynomials via the Wirtinger presentation 195
5 of .:1,(t)
We list the following without proof. They can be useful in checking
for errors:
1. For any knot K, LiIO) = ± l.
2. For any knot K, there is an integer n such that
Lil(t) = tnLiI(tI).
Thus if Li
l
(t) = cnt
n
+ ... + Co, the coefficients pair up Ci = C
n
_ i for
i = 0, ... ,n. (Proofs are given in Crowell and Fox (1967, chap
ter IX).
3. If the knot is alternating, and the genus of K can be calculated,
it is useful to note that
genus of K = 1 degree of Li
l
(t).
8.5
1. The group G(tp, q) with presentation
(x, y : x
p
= yq)
is the group of a (P. q)torus knot where p. q are coprime positive
integers. Calculate Lil(t) for this knot and show that
2. Suppose K is a knot with crossing number c. Investigate the
relationships between C and the degree of the normalized Alex
ander polynomial.
3. Investigate what happens if one applies the Fox calculus, Alexan
der matrix, elementary ideal analysis to a Dehn presentation of a
knot group. Examine in particular the relationship with Alexan
der's original method of defining the Alexander polynomial.
4. Investigate, using results and exercises from earlier chapters, how
the Alexander polynomial reacts to knot sum. (In fact we showed
that it changed knot sum to product of polynomials using the
'original' method. Try to see why this is true from the point of
view of Wirtinger presentations, Alexander polynomials, etc.
xander matrices and Alexander polynomials
This may seem to you more complicated but it does provide a
new type of insight.)
5. Calculate lots of Alexander polynomials. (You choose the knots
from anywhere in the book!)
6. Can the Alexander polynomial construction be extended to be
\
\
Fig. 8.2
applicable to links with more than one component? See how far
you can get. Many of the necessary changes in the theory are
hinted at in this chapter. If you are successful in doing such a
generalization, then try to adapt Alexander's original method!
7. Using the preceding exercise, find how the Alexander polyno
mials of knots related by the sequence of changes in Fig. 8.2 are
themselves related. There is a simple equation rather like the
Homfly defining equation. This should help you to find a
relationship between the Alexander polynomial and the Homfly
polynomial. Use either formulation of the Alexander polynomial.
s
9
The fundamental group
The fundamental group is an algebraic label attached to a topolo
gical space. Naturally enough, it is a group and is constructed using
paths in the space, continuous analogues of the sequences of edges
constituting paths in a graph. The more convoluted is the shape of
the space, the more variety there will be among the paths in that
space and the larger will be the fundamental group. The funda
mental group first arose in Poincare's work on the threebody
problem. His idea was to classify the possible evolutions of a system
of three bodies moving under mutual gravitational forces as paths
in the phase space with coordinates given by the bodies' positions
and velocities, but to regard nearby paths as the same. This led to
the notion of deformation of one path into another, now formalized
as homotopy of paths. The fundamental group consists of homotopy
classes of paths in a space. The notion of homotopy is also crucial
in complex analysis: Cauchy's theorem on contour integration may
be phrased as saying that the integrals of a holomorphic function
around two homotopic contours are the same (see e.g. Priestley
1990). Further development of the extensive theory of the funda
mental group, beyond that covered here, may be found in Armstrong
(1983) and Brown (1988). Our principal application will be to
knots: any knot yields a group, namely the fundamental group of
its complement in [R:3, and algebraic properties of the group reflect
properties of the knot.
Let X be a topological space. A path (0 in X is a continuous map
(0: [0, r] ~ X, where r ~ O. The real number r is the duration of the
path; if r = I we say that the path is normalized. Given any path (0
in X, there is a normalized path (01 with the same image in X; we
simply define COJ(t) = (O(rt) for 0 ~ t ~ 1. We can therefore restrict
e fundamental group
attention to normalized paths if we choose, but it is often more
convenient to allow paths of arbitrary duration.
If w: [0, r] ~ X and m: [0, r'] ~ X are paths in X, such that
w(r) = w'(O), then wand w' have an obvious product, the path
obtained by first going along the image of wand then along
the image of W'. We write this product as W· w'; then we have
W· w': [0, r+ r'] ~ X defined by W· w'(t) = w(t) if ° ~ t ~ rand
W· w'(t) = w'(t  r) if r ~ t ~ r + r'. The duration of W· w' is r + r'
and clearly W· w'(O) = w(O), W· w'(r + r') = w'(r'). There is a pro
duct of normalized paths as well. If w: [0, 1] ~ X and w': [0, 1] ~ X
are normalized, then their normalized product is the normalized
path a obtained from W· w': [0, 2] ~ X. Since a(t) = W· w'(2t), we
have aCt) = w(2t) if 0 ~ t ~ 112 and aCt) = w'(2t  1) if l/2 ~ t ~ 1.
Thus a is obtained by going along wand w' in turn, but twice as
quickly.
It is important to note that the product of paths is not defined
for arbitrary paths wand w': it must be that wends at the same point
at which w' begins. In particular, the product W· w' may be defined
whilst w' . w is not.
rna. Let w: [0, r] ~ X, w': [0, r'] ~ X and w": [0, r"] ~ X be paths in X
with w(r) = w'(O) and w'(r') = w"(O). Then the products W· (w' • w")
and (w· w') . w" are both defined and are equal.
We leave the proof of Lemma 9.1 as an exercise. Note that it says
that the product of paths is an associative operation. This is not true
of the normalized product: why not?
For each point x E X there is a constant path 0,: {O} ~ X of
duration 0 defined by OAO) = x. If w: [0, r] ~ X is a path then
W· OW(r) = wand OW(O) • w = w, so that the paths of duration 0 provide
a set of left and right identities for the product operation on paths.
However, there are no inverses; for if w has duration r> 0, then for
any w' the path W· w' (if it is defined) has duration at least r and so
cannot be a path 0,. The best approximation that we have to an
inverse for w: [0, r] ~ X is the reverse path OJ: [0, r] ~ X defined by
OJ(t) = w(r  t). Both W· OJ and OJ. ware defined and are paths of
duration 2r.
We see that the set of paths in X is endowed with quite a reasonable
structure, but one which is none the less deficient in certain respects,
and is plainly not that of a group. The circumstances are similar to
those encountered in the construction of free groups in Chapter 6,
when it proved necessary to consider equivalence classes of words on
an alphabet. The next step towards the fundamental group is to put
an equivalence relation on the set of paths in the space X. It turns
Homotopy 199
out that, with this equivalence relation, it will be sufficient to
consider only normalized paths.
motopy
Let X and Y be topological spaces and suppose that f: Y ~ X and
g: y ~ X are continuous maps. A homotopy between I and g is a
continuous map h: Y x I ~ X such that, for all y E Y, hey, 0) = I(y)
and hey, I) = g(y). The set of functions {h(, t) I 0 ~ t ~ I} can be
thought of as a continuous deformation from I to g. If a homotopy
between I and g exists, then f and g are homotopic and we write
f=g·
position. Homotopy is an equivalence relation on the set of all maps
Y ~ x .
There are variations on the notion of homotopy obtained either by
considering only a limited class of functions, or by adding extra
conditions on the homotopy. Pointed homotopy of closed paths leads to
the fundamental group. The properties of paths with which this chapter
began are more accurately reflected by fixed end point homotopy of
paths, and this leads to the fundamental groupoid. For a full account of
groupoids, see Brown (1988), and for the first steps in the theory of the
fundamental groupoid, see Exercises 9.3.27 below.
As already remarked, it will suffice to consider only normalized
paths in our route to the fundamental group. Let X be a topological
space and Xo E X any point of X. A closed path in X at Xo is a
continuous map m: [0, 1] ~ Xwith m(O) = Xo = m(1). Observe that the
normalized product of any two closed paths at Xo is defined and is
again a closed path at Xo' Suppose that a and m are closed paths in
X at Xo' A pointed homotopy from a to m is a continuous map
h: [0, 1] x [0, 1] ~ X such that, for all s E [0, 1). h(s, 0) = a(s) and
h(s, 1) = m(s), and further such that for all t, h(O, t) = Xo = h(l, t). In
0)
x
0:
The fundamental group
a pointed homotopy, we have a set of closed paths at Xo that give a
continuous deformation from a to (0. A simple but useful picture to
bear in mind is the unit square [0, 1] x [0, 1] adorned with labels to
show how it is mapped into X by h.
Any point on the sides of the square is mapped to Xo: the bottom
of the square is mapped into X by a, the top by OJ, and horizontal
slices give the intermediate closed paths h(, t), 0 ~ t ~ 1.
Pointed homotopy is an equivalence relation on the set of all
closed paths in X at x
o
, and we will use [a] to denote the equivalence
class of a closed path a. The set of all the equivalence classes of
closed paths in X at Xo is denoted 1TI (X, x
o
). Paths are composed by
juxtaposition, just as were the words in a free group, and this
analogy with the construction of free groups leads us to hope that
1T] (X, x
o
) is a group, with multiplication of equivalence classes of
paths obtained by taking the product of representative paths.
heorem. For any [aJ, [a1 E 1Tl (X, x
o
), the binary operation [aHa1 =
[a. a1 is well defined and makes 1TI(X, x
o
) into a group.
Proof. For ease of digestion the proof is divided into four steps.
Step 1
We shall show that the operation is well defined; that is, it is independent
of any choices made of one closed path in an equivalence class. Suppose
that [a] = [OJ] and that [a'] = [OJ']. There are pointed homotopies h from
a to OJ and h' from a' to OJ'. The following picture indicates the
(0
(0'
a a'
definition of a pointed homotopy h" from a· a' to OJ· OJ'.
Define h" as follows. We have, for all t, h"(s, t) = h(2s, t) if
o ~ s ~ 112, and h"(s, t) = h'(2s  1, t) if 112 < s ~ 1. Then h" is
indeed a pointed homotopy from a· a' to OJ· OJ', and therefore
[a· a'] = [OJ' OJ']'
Homotopy 201
Step 2
The operation is associative. Given closed paths a, a', aN at x
o
, we
need to construct a homotopy h from (a· a') . an to a' (a' • an).
Remember that, because we are dealing with normalized paths,
(a· a') • a" and a· (a' . aN) need not be equal paths. They are
homotopic, and again a picture indicates what we should do.
a a'
a"
Define h(s, t) = a(4s/(t + I)) if 0 ~ s ~ (t + 1)/4, h(s, t) = a'(4s  t
 1) if (t + 1)/4 ~ s ~ (t + 2)/4, and h(s, t) = a"«4s  t  2)/(2  t) if
(t + 2)/4 < s ~ I, and this has the properties we need.
Step 3
Let e: [0, 1] ~ X be the constant path at Xo. Then [e] is an identity
element in Jr
j
(X, x
o
); for given any closed path a in X at Xo we have
a· e = a and so [a][e] = [a). A homotopy from a· e to a is indicated
by the picture
a
and given by the explicit formulae h(s, t) = a(2s/(t + I)) if
o ~ s ~ (t + 1)/2 and h(s, t) = Xo otherwise.
The fundamental group
Step 4
If a is a closed path in X at x
o
, its reverse a (which you will recall is
defined by a(s) = a(l  s» is also a closed path in X at Xo and [a]
is the inverse of [a]; that is, [a][a] = [e]. A homotopy from a· a to e is
indicated by the picture
a a
and we leave it as an exercise for the reader to write out an explicit
formula for the homotopy. •
The group HI (X, xo) is the fundamental group of the space X at the
point Xo. What happens if we choose another point XI instead of Xo
at which to base all our closed paths?
roposition. If X is path connected and X
o
, XI are two points of X, then there
is an isomorphism between HI (X, xo) and HI (X, xd.
Proof. Let w: [0, 1] ~ X be a path in X with w(O) = Xo and
w(l) = XI' Suppose that a is a closed path in X at Xo; then W· a· iij is
a closed path in X at XI' and mapping [a] H [w • a' iij] gives an
isomorphism HI (X, xo) ~ HI (X, xd. •
ises 9.2
1. The definition of homotopy will have reminded the reader of the
definition of isotopy of knots. If k, k': S' ~ [R3 are isotopic, then
are they homotopic? Why isn't homotopy a useful equivalence
relation for knots?
2. Prove Proposition 9.2. [Hint. The only difficult part is transitiv
ity. The gluing lemma should help.]
3. The proof of Proposition 9.4 is rather sketchy. Supply the
details.
4. Is it reasonable to expect an isomorphism between HI (X, xo) and
HI (X, XI) if Xo and XI lie in different path components of X?
Calculations of the fundamental group 203
Calculations of the fundamental group
The computation of fundamental groups will be the main theme of
the remainder of the book, so that examples will be in plentiful
supply. We pause here to consider some spaces whose fundamental
group is trivial.
A topological space is simp(v connected if it is path connected and
if trl (X, x
o
) = I for some (and hence all) Xo E X.
Proposition. If X ~ [Rn is a convex set then X is simply connected.
Proof. Recall that X ~ IR
n
is convex if for all x, y E X, and all t
with ° ~ t ~ I, we have tx + (t  I)y E X. Let m: [0, I] ~ X be a
closed path at Xo E X, and define h: [0, IJ x [0, 1] ~ X by
h(s, t) = tx
o
+ (l  t)m(s). Then h is a homotopy from m to the
constant path at Xo. •
Our first example of a nontrivial fundamental group is Sl, the
circle. It is convenient to regard 1R2 as the complex plane and so
identify Sl as {z Eel I z 1 = I}; we take lEe as basepoint.
Theorem. Thefundamental group tr1(St, 1) is infinite cyclic, and is generated
by the homotopy class of the closed path m: t H exp (2trit).
Proof. Before we embark on the detailed proof. it is worth
considering the matter in a vaguer fashion. A path in Sl may wander
around in a complicated way, but it is forced to remain within the
onedimensional bounds of the circle. Its wanderings are thus no
more than changes of direction, or backtracking, and these portions
of the path can be smoothed out by some suitable homotopies. Any
path will be homotopic to one without any backtracking, and will
wind around the centre of the circle a certain number of times. Our
claim is that this winding number completely determines the homotopy
class of the path; that is, that two paths in Sl are homotopic if and
only if they have the same winding number. Since composition of
paths induces addition of winding numbers, we will have an
isomorphism rp: trl(St, I) ~ ~ as claimed.
Our first problem is to give a definition of the winding number.
You might like to think of ways in which this might be done, and
see if you can use your definition to prove the theorem. The solution
we shall adopt here may initially seem strange: we consider paths in
[R instead of paths in Sl. The path m: [0, 1] ~ SI given by
m: t H exp (2trit) extends to a function IR ~ Sl given by the same
formula: you may picture m as winding the real line IR into a spiral
and then flattening the spiral onto the unit circle. The crucial fact
fundamental group
about OJ for our purposes is this: given any path a: [0, lJ ~ SI there
exists a unique path 0: [0, 1] ~ IR with OJ 0 0 = a. The path 0 is a
lifting of a and is unwound along IR; furthermore, since 0(1) = 1 we
must have o(l) E 7L. We can therefore define the winding number of
ato be 0(1). The homeomorphism ¢: JrI(SI, I) ~ 7L that we seek will
then be defined by ¢([aD = o(l), and it will only remain to show that
¢ is bijective.
Is the previous sentence true? In fact it disguises the problems that
we must still face: it may be difficult to show that ¢ is bijective (what
do you think at this stage?), but more fundamentally (pun intended),
is our definition of ¢ a workable one? Firstly we have yet to prove
that OJ does possess the crucial pathlifting property, so that given a
we can get a unique 0. Secondly, in the definition of ¢, we use the
homotopy class of a and not a itself. Does it matter which path in
the homotopy class we lift to IR? Of course, we would not have come
this far unless everything was going to be all right in the end, but it
is extremely important to see exactly what it is we still have to do to
make the idea of the proof into a watertight argument. We carry out
this waterproofing process in four steps:
Step 1: we show that given any path a: [0, 1J ~ SI there exists a
unique path 0: [0, 1] ~ IR with OJo 0= a.
Step 2: we show that ¢: JrI(SI, 1) ~ 7L is well defined by
¢([a]) = 0(1); that the answer we get does not depend on the
choice of a path in the homotopy class [a].
Step 3: we show that ¢ is a homeomorphism.
Step 4: we show that ¢ is bijective.
Step 1
lifting lemma. If a: [0, I] ~ SI is a path with a(O) = 1 there exists a
unique path 0: [0, 1] ~ IR such that OJ· 0 = a and 0(0) = o.
To prove the pathlifting lemma we set U= S\{ l} and V= SI\{1};
then U and V are open and Uu V= SI. Now OJI(U) = u
nE
7L
(n  112, n + 1/2) and for each n E 7L, OJI (n  l!2, n + 112) is a homeo
morphism onto U. Let an: U ~ (n  1/2, n + 1/2) be its inverse.
Similarly, we consider OJ I (V) = Un E ~ ( n , n + 1), and we let f3n: V ~
(n, n + 1) be the inverse of OJI(n, n + 1). Each of OJI{U) and OJI{V)
is open in [0, I] and [0, 1] = OJI (U) U OJI (V). It follows from Lebes
gue's lemma in the theory of metric spaces that there exist points
0= to < tl < ... < tn = I such that OJ([tj , tj+ d) lies in U or in V. We
shall not prove Lebesgue's lemma in this book, but it is easy to find
in standard texts on metric spaces, such as Sutherland (1981), or in
Calculations of the fundamental group 205
Armstrong (1983). Since (0(0) E U we have (0([0, tID ~ U and we can
define a: [0, til ~ ~ by a(t) = ao 0 o(t). If we now assume that we have
defmed a on the interval [0, td, we show that we can extend the definition
to [0, tk+ I]' If 0 ([th tk+ I)) ~ U then a(tk) E (n  112, n + 112) for
some nEZ, and we may define CT: [th tk + I] ~ ~ by CT(t) = an 0 o(t).
Similarly, if o([tb tk+ I]) ~ V then a(tk) E (n, n + 1) for some n E Z
and we may define CT: [tb tk + I] ~ ~ by CT(t) = Pn 0 o(t). Repeating this
procedure, we eventually define CT ([0, I]). Notice that we have no
choice for the definition of a at each step if we are to ensure that
(0 0 CT = 0 and CT(O) = 0, and so CT is unique.
Step 2
This is also accomplished by lifting the problem to ~ . The result we
need is as follows.
ylifting lemma. lj h: [0, I] x [0, IJ ~ SI is a homotopy between the
closed paths oand r, (so that h(O, t) = I = h(l, t)for all t E [0, I)) there
is a unique homotopy h: [0, I] x [0, 1] ~ ~ such that (00 h = hand
h(O, t) = 0 for all t.
We shall not give a detailed proof of this lemma because the ideas
needed are the same as for the pathlifting lemma. The unit square
[0, 1] x [0, 1] must be divided into smaller squares, again by use of
Lebesgue's lemma, each of which maps into U or into V, and h then
defined square by square.
N ow suppose that 0 and 1: are closed paths in SI and are homotopic
by means of a homotopy h. The homotopy h given by the homotopy
lifting lemma has the property that (00 h(s, 0) = h(s, 0) = o(s). Fur
thermore, h(O, 0) = O. It follows by the pathlifting lemma that
h(, 0) = CT and so q>(o) = CT(1) = h(1, 0). Similarly, h(, I) = r and so
q>(tJ = T(l) = h(l, 1). Now for all t E [0, IJ we have h(l, t) E (01(1) =
Z; but the image of the connected set {(1, t) It E [0, I]} under the
continuous map h must be connected, and so is a single integer, n
say, and now q>(o) = n = q>(1:).
Step 3
This is nice and easy. Given two closed paths 0 and 1: in SI, with
corresponding paths CT and T, define a path IJF: [0, 1] ~ ~ by
lJF(t) = CT(2t), 0 ~ t ~ 1/2
= T(2t  1) + CT(1), 1/2 ~ t ~
Then (0. IJF = 0 + 1: and IJF(O) = CT(O) = O. By uniqueness, IJF = (0 + 1:)
and 1JF(l) = T(I) + a(1).
The fundamental group
Step 4
It is easy to see that cp is surjective: given n E If, let an: [0, 1] ~ SJ be
defined by an (t) = exp (2n Jri t). We leave it to the reader to check that
on(1) = n so that cp(a
n
) = n.
It remains to show that cp is injective; all we need to prove is that
if cp([aD = 0 then a is homotopic to the constant path at l. If
cp([aD = 0 then 0(1) = 0, so that 0 is a closed path in IR. Define a
homotopy H: [0, 1] x [0, 1] ~ !R by H(s, t) = (I  t)o(s). Then H is
a homotopy from 0 to the constant path at O. Set h = (t) 0 H. Then
h(s, 0) = OJ 0 o(s) = a(s) and h(s, I) = OJ(O) = l. We see that h is a
homotopy from a to the constant path at 1, and finally the proof is
complete. •
The complexities of the proof should have convinced you that the
computation of fundamental groups from first principles is likely to
be a formidable task. Even though we began with a very plausible,
and rather simple, candidate for ll"J (S\ 1) we still encountered some
subtle problems in establishing that our candidate was indeed cor
rect. How are we to compute the fundamental group of a space
when no likely candidate presents itself to our imaginations? Fortu
nately, a powerful tool exists: Van Kampen's theorem. This will be
the subject of the following two chapters. For the remainder of this
chapter, we develop some more basic theory of the fundamental
group.
How are operations on spaces reflected in the calculation of the
fundamental group? For products of spaces (with the product topo
logy) the answer is very satisfactory.
roposition. Let X and Y be topological spaces with Xo E X and Yo E Y.
Then ll"J (X x Y, (xo, Yo» is isomorphic to the direct product
ll"J (X, xo) X ll"J (Y, Yo).
Proof. Let a be a closed path in X x Y at (xo, Yo). There is an
obvious candidate for an isomorphism:
we take cp([a}) = ([pa}, [qa}) where p: X x Y ~ X and q: X x Y ~ Y
are the projection maps. The proof of the proposition consists of a
verification of the credentials of cpo
Is cp well defined? If a = OJ then do we have pa = pOJ and
q a = q OJ? Let h be a homotopy from a to OJ. Then ph and qh are
homotopies from pa to pOJ and from qa to qOJ respectively. Check
the details of this assertion.
Calculations of the fundamental group 207
Is qJ a homeomorphism? If a and f3 are closed paths at (xo, Yo),
check that p (a· {3) = pa· p{3 as closed paths in X at Xo and that
q( a . {3) = q a • q {3 as closed paths in Y at yo.
Is qJ injective? Let e, e' be the constant paths at Xo and Yo
respectively and suppose that hand h' are homotopies from pa to e
and from qa to e. Define H: [0, 1] x [0, 1] ~ X x Y by H(s, t) =
(h(s, t), h'(s, t)). Show that H is a homotopy from a to the constant
path at (xo, Yo). Explain why this shows qJ to be injective.
Complete the proof by showing that qJ is surjective. •
Before continuing with the theory of the fundamental group, we
return to the properties of paths, and in the following exercises
we consider a homotopy notion more closely related to path pro
ducts than based homotopy.
9.3
I. X ~ [Rn is starlike if it is path connected and if there exists
Xo E X such that for any y E X and for all t with 0 ~ t ~ I, we
have txo + (1  t)y E X. Show that starlike subsets of [Rn are
simply connected.
2. Let a and w be paths in X with a(O) = Xo = w(O) and a(l) = Xl =
w(I). A fixedendpoint homotopy from a to w is a continuous
function h: [0, I} x [0, l} ~ X such that, for all s E [0, 1], h(s, 0) =
a(s) and h(s, 1) = w(s) , and further such that for all t E [0, IJ,
h(O, t) = Xo and h(l, t) = Xl' If there is a fixedendpoint homo
topy from a to w we write a = w (f.e.p.). Show that fixed
endpoint homotopy is an equivalence relation on the set of all
paths a in X with fixed end points a(O) = Xo and a(l) = Xl'
3. Show that if a, {3, r are normalized paths such that the nor
malized product a· ({3. r) is defined, then (a· {3) • r is defined
and a· ({3 • r) = (a· {3) • r (f.e.p.).
4. Show that W· iii = 0",(0) (Le.p.) and that iii· w = 0"'(1) (f.e.p.).
5. Let a, w, a', w' be paths in X such that a = w (f.e.p.), a' = w'
(f.e.p.), and such that a· a' is defined. Show that W· w' is
defined and that a' a' = W· w' (f.e.p.).
6. Let nx denote the set of fixedendpoint homotopy classes of
paths in X (so we take all possible pairs of end points, but within
each class the end points are fixed). Show that the product of
homotopy classes [w][w'] = [w· w'], defined when w(l) = w'(O), is
well defined.
7. Let X be a path connected space. Show that X is simply
connected if and only if given any two paths a and w in X with
a(O) = w(O) and act) = w(l) then a = w (f.e.p.).
The fundamental group
Homotopy equivalence
The fundamental group construction attaches an algebraic label to a
topological space that records some information about the shape of
the space. A task that we might hope to have made easier by this
labelling is that of distinguishing one space from another: spaces
with different shapes should carry different labels. On the other
hand, homeomorphic spaces might be expected to have the same
label, that is to have isomorphic fundamental groups. This is indeed
the case, and we arrive at the result by extending the construction of
the fundamental group to continuous maps between spaces: a con
tinuous map becomes a homeomorphism of groups.
Let X and Y be topological spaces. To discuss fundamental groups,
we need a pair of basepoints, Xo E X and Yo E Y. The space X and
the selected point Xo together form the pointed space (X, xo). Given the
pointed spaces (X, xo) and (Y, Yo), a continuous function f: Y ~ X
is pointed if Xo = f(yo)' Iff is pointed we write f: (Y, Yo) ~ (X, xo). If
a: [0, I] ~ Y is a closed path in Y at Yo and f: (Y, Yo) ~ (X, xo) is a
pointed continuous function then fa is a closed path in X at Xo.
[Exercise: exactly what has just been asserted? Why is it true?] If h is
a pointed homotopy from a to win Y then fh is a pointed homotopy
from fa to fw, and so given raj E lli(Y, Yo) the class [fa] is well
defined in 1f
J
(X, x
o
). From f: (Y, Yo) ~ (X, x
o
) we have obtained a
function f*: 1f
J
(Y, Yo) ~ 1fj (X, x
o
) defined by [aJ H [f aJ. Now if a, f3
are closed paths in Yat Yo, thenf(a· f3) = (fa) • (ff3) [Exercise: check
this!] and therefore f* is a homeomorphism between the fundamental
groups 1fj (Y, Yo) ~ 1fj (X, x
o
).
Since homotopy is clearly a very good idea, we should attempt to
perceive its effects wherever possible. What is the effect on the
homeomorphismf* of a homotopy fromfto another pointed map g?
The answer is that there is no effect at alL Let f,
g: (Y, Yo) ~ (X, x
o
) be pointed maps. A pointed homotopy fromfto g
is a homotopy H: Yx [0, I] ~ X fromfto g such that H(yo, t) = Xo
for all t, with ° ~ t ~ 1. If a: [0, 1] ~ Y is a closed path in Y at Yo
then h: [0, 1] x [0, 1] ~ X defined by h(s, t) = H(a(s), t) is a pointed
homotopy fromfa to ga and so in 1fj (X, x
o
) we have [faJ = [gal; that
is, for all raj E 1f
J
(Y, Yo), }:[a] = g*[a] and so}: = g*. Our conclusions
are summarized in the next result.
Theorem. A pointed map f: (Y, Yo) ~ (X, x
o
) induces a homeomorphism
f*: 1fj (Y, Yo) ~ 1f) (X, x
o
) between fundamental groups. Iff is homotopic
to g: (Y, Yo) ~ (X, x
o
) by a pointed homotopy then h = g*; that is,
pointed homotopic maps induce the same homeomorphism on funda
mental groups.
Homotopy equivalence 209
•
We say that a pointed map f: (Y, Yo) ~ (X, xo) is a homotopy
equivalence if there exists a pointed map g: (X, x
o
) ~ (Y, Yo) and
pointed homotopies between gf and id y and betweenfg and id
x
. The
map g is a homotopy inverse for f. If fis a homeomorphism then it is
certainly a homotopy equivalence, and f I is a homotopy inverse for f.
Proposition. If f: (Y, Yo) ~ (X, x
o
) is a homotopy equivalence then
f*: H
I
( Y, Yo) ~ HI (X, x
o
) is an isomorphism.
Proof. Let g be a homotopy inverse for f. We leave it to you to
show that g* is an inverse for f.,.
Especially interesting is the setting in which there exists a homo
topy equivalence between a space and a subspace. The subspace is
then a simpler version of the space that, with our vision blurred by
homotopy, is indistinguishable. Let A be a subspace of X and choose
a basepoint Xo E A. Let i: (A, xo) ~ (X, x
o
) be the inclusion. A retrac
tion of X to A is a continuous map r: (X, x
o
) ~ (A, x
o
) such that
riCa) = a for all a E A. If there exists a retraction of X to A we say
that A is a retract of X.
Proposition. If A is a retract of X then the retraction r: (X, x
o
) ~ (A, x
o
)
induces a surjection r*: HI (X, x
o
) ~ HI(A, x
o
) and r*i* is the identity on
H,(A, x
o
).
Proof. Exercise!
•
As an application we can prove a famous result of L. E. J.
Brouwer.
The Brouwer fixed pOint theorem. Let D be the unit disc in 1R2, that is
D = {x E 1R2111XII ~ l}. Then any continuous mapj: D ~ D has afixed
point; there is some xED such that f(x) = x.
Proof. Suppose that no fixed point exists. Then for each XED we
may define rex) to be the point on the boundary SI of D crossed by
the line segment fromf(x) to x (Fig. 9.1).
Then r is a retraction of D to S', and we leave you to check this.
Therefore r* is surjective from HI(D, (1,0» to H,(S', (1,0»; now find
a contradiction by combining Proposition 9.5 with Theorem 9.6. •
The subspace A of the space X is a deformation retract of X if there
exists a retraction r: (X, x
o
) ~ (A, x
o
) and a pointed homotopy h
from ir to id
x
. The inclusion i: (A, x
o
) ~ (X, x
o
) is then a homotopy
equivalence.
The fundamental group
pies
f(x)
Fig. 9.1
(a) Let and A={(x,y)lx
2
+y2=1}.
Then rex, y) = (x, + y2) is a retraction and
I  t
h«x, Y),t) = t(x, y) + y)
X +y
is a homotopy ir = id x.
(b) In Fig. 9.2, the subspace A (a line segment joining two circles)
is a deformation retract of the space X: visualize the retraction!
Fig. 9.2
(c) See Fig. 9.3.
ises 9.4
1. If g: (Z, zo) (Y, Yo) and j: (Y, Yo) (X, x
o
) are pointed maps,
show that (jg)* andj*g* are equal.
2. Letj, g: Y Xbe homotopic maps (with no assumption that the
homotopy is pointed) and let Yo E Y. Show that there exists
an isomorphism (J: lli (X, j(yo» "I (X, g(yo» such that
g*: "I (Y, Yo) "1 (X, g(yo» is equal to the composite
Homotopy equivalence 211
x
Fig. 9.3
3. Let f: Y ~ X be a continuous map. We say that f is a homotopy
equivalence if there exists a map g: X ~ Y such that fg = id
x
and
gf = id yo If X, Yare path connected spaces and f: Y ~ X is a
homotopy equivalence, show that for any Yo E Y the fundamental
groups HI (Y, Yo) and HI (X, f(yo}) are isomorphic.
10
Van Kampen's theorem
In this chapter we will examine the statement and proof of the Van
Kampen theorem. If you glance forward to where this theorem is
stated in full and in detail, you may wonder why such an apparently
technical theorem is needed. Complicated theorems occur at numer
ous places within mathematics courses at this level and students often
ask themselves why such theorems are deemed necessary. Again it
is often the case that the lecturers have used the results explicitly
or implicitly so many times in research projects that it becomes self
evident to them that these theorems are important and hence it can
be extremely difficult to explain their importance to someone meet
ing them for the first time. With the Van Kampen theorem, we are
lucky. We have been developing the theory of the fundamental group
and if we pause to take stock of our present knowledge (and
ignorance), it should become clearer why theorems of the Van
Kampen type are needed. An even closer look will suggest what it
should say and will also give us a sketch proof. Writing out a formal
proof is another matter and will be dealt with later on; the first
priorities are to guess what the result will say (before we meet its
detailed statement), to see why the result is true and why certain
technical restrictions are necessary; the details of the formal proof
are (initially) somewhat less important. Once you understand the
idea of the proof the details can be seen merely as the 'justification
of the obvious'. The fact that these details are technically nontrivial
merely shows that the 'obvious' often does need justification!
Why do we need a Van Kampen theorem? The state of our art at
present is that we have calculated a large class of fundamental groups
with great ease only to find they are all trivial! Working considerably
harder we were able to prove that 1l't of a circle is an infinite cyclic
group. We proved a result on 1l
J
of a product of spaces and so can
show 1l'J(torus) == C ~ ~ C ~ , but we cannot handle 1l'J of an nIeafed
rose, a simple enough space you might think, nor 1l'J (genus g
orientable surface) if g > 1. These spaces are good examples as we
can understand them well. We can build geometric models of them,
Van Kampen's theorem 213
we can picture them, and so should be able, one would hope, to
compute their fundamental groups. One important thing they have
in common is that they can be built up from simpler 'building block'
spaces such as lines and discs. The nleafed rose is got from n circles.
The genus g orientable surface can be built up from a 2gleafed rose
and a disc.
This raises the question: can one calculate HI of a space if one
knows H) of the pieces it is built up from? (If the answer is no, it is
not at all clear how to go further in calculating fundamental groups!)
To answer this question we will need to examine how loops in a space
are related to loops in parts of a space.
Suppose we have a path connected space X divided up into two
nonempty parts U and V. We are working with pointed spaces since
we want to calculate HI (X); it is therefore best to assume that x
o
, the
basepoint of X, is in U 11 V. (If U 11 V were empty then X would not
be pathwise connected.) Let us consider a picture (Fig. 10.1) that will
help the discussion (but always beware of arguments that rely too
much on pictures!).
loop w
Fig. 10.1
Suppose w: [0, 1] ~ X is a normalized closed path at Xo. We hope
to decompose w, up to homotopy, into parts, each of which is either
wholly in U or wholly in V. In the picture, it is clear that one could
go around w until one arrives in U 11 V, and then continue in V until
one is back in U 11 V and so on. This idea would seem to decompose
the path w into paths totally within U or totally within V. However,
this decomposition would be taking place not within H] eX) but within
the fundamental groupoid of X, as we get paths in U and V that need
not be closed.
Let us be a bit more precise (Fig. 10.2). We will assume that we
can find points I], Ib ... , tn _) E. [0, 1 J so that (putting 10 = 0, tn = 1)
each wi [th ti+]] is a path wholly within U or wholly within V.
How are we to ensure this is possible? For instance, if we take
n Kampen's theorem
Fig. 10.2
W: [0, lJ ~ ~ 2 defined by w(t) = (t, t sin(l/t», U= {(x, y): y ~ O},
V = {(x, y): y ~ O} then we will be in difficulties, since we would
need to chop [0,1 J into infinitely many subintervals to get the sort of
property we want. We cannot, however, compose an infinite family
of paths in the fundamental groupoid! Our problem is that V and V
in this example are not open. If we increased V and V slightly in
size, say V
E
= {(x, y): y ~  oS} and V
E
= {(x, y): y ~ oS} for any small
oS> 0, then we would be able to get away with finitely many subpaths.
Why?
To handle the general case we need to recall the idea of compact
ness. First the definition: a space X is compact if, given any family
U = {U
A
: A E A} of open sets such that X = u {U
A
: A E A}, there is a
finite subfamily {U
AI
' •• . , VAn} of U so that X = Uf= 1 VA;" (Every
open covering of a compact space has an open subcovering.) Now if
U, V are open sets, so are w  1 (U) and w  t (V). The open subsets of
[0, IJ are unions of open intervals and [0, IJ is compact. (The
HeineBorel theorem from real analysis can be thought of as saying:
X k ~ n is compact if and only if it is closed and bounded. Thus both
1 = [0, lJ and 12 = 1 x I, which is the closed unit square in ~ \ are
compact spaces.) We need a further piece of information about compact
spaces or, more precisely, compact metric spaces, that is spaces where
the open sets are defined by a distance function or metric. If Y is any
compact metric space and U is any open covering, there is a number
oS> ° (often called the Lebesgue numbering of the covering) such that
any subset of Y of diameter less than oS is wholly contained in some
set of the covering. (Try to prove this yourself using techniques from
analysis, or look it up in a book on metric spaces or on point set
topology.)
Applying these ideas to our situation, we find that there is an
oS> ° such that any interval of length less than oS is contained wholly
within w I (U) or w I (V). We can thus pick points to = 0, t], . .. ,
tn _., tn = I in 1 such that each ti  ti _ 1 < oS and hence w[t; _], tiJ is either
within U or within V. Now clearly if we write Wi for w i[t; _ I> ta
Van Kampen's theorem 215
composed with the linear homeomorphism between [0, ti  ti_ J and
[Ii _ 1, til, then we have
and so we have written W as a composite of paths wholly within U
or within V. This takes place in the fundamental groupoid of X. The
fundamental groupoid is useful, but for the moment we are searching
for a result on fundamental groups and the paths Wi need not be
closed. To get around this difficulty we proceed as follows: pick for
each i = 0, I, ... , n, a path Vi joining Xo to W(ti); these paths are to
satisfy
(1) Vo and Vn are the trivial path at Xo;
(2) if Wi and Wi+ 1 are both paths in U, Vi is wholly within U;
similarly if they are both in V, Vi is wholly within V;
(3) if one of Wi and Wi+ I is in U and the other in V, then Vi is a path
mUnV.
This last condition overlaps with condition (2) since if i'Vi and
Wi+ I are both in Un V then Vi' under both conditions, has to be
within Un V. In any case we must have U, V, and Un V path
connected if we are to be able to fulfil these conditions. (This
requirement could have been avoided to some extent if we had
worked with the fundamental groupoid rather than the fundamental
group as our principal tool; we will briefly examine this possibility
later.)
We can rewrite the decomposition of w, up to homotopy this time,
as
as each Vi Vi is homotopic to the trivial path. (At this point, it is
useful to think of working with unnormalized paths as otherwise we
would need to bother with the fact that composition of normalized
paths is not associative. In fact, we can disregard this technical
problem as we know that if we work with unnormalized paths and
then normalize at the end of a calculation, the end result is homo
topic to the end result obtained by consistently normalizing and taking
into account the lack of associativity that results.) We can group this
decomposition as
within X, where Ui = Vi _ I WiVi . Each Ui is a loop either in V or in V.
We will use the convention that Ui is in VU) where Vel) = Vor V. This
Kampen's theorem
allows us to introduce a neat piece of notation: if w is a loop in X
we write [w]x for the corresponding equivalence class in 1t'j (X);
similarly with X replaced by V, V or V n V. This means that the
induced homeomorphism
is defined by idwJu = [wJx. We note that in the decomposition above
[wJx = [u1Jx[u:Jx' .. [un]x
= iU(l) [uj]U(l) ... iU(n) [un]u(n)'
This gives us a result that we can summarize as follows.
rna. The images Of1t'l(V) and 1t'j(V) in 1t'j(X) generate that group.
An alternative and slightly more formal way of putting this same
result is as follows. The induced homeomorphisms
iu: 1t'](X)
iv: 1t'j(X)
in their turn induce a homeomorphism
where as before * indicates free product and the lemma states that
this homeomorphism is onto 1t'J (X). This is already a positive result.
For instance, if we have presentations (x : r) of 1t'J(V) and (y : s) of
1t'] (V) then it says that 1t'J (X) is a quotient of the group with
presentation (x u y: r us). This means that generators for 1t'J (V)
and 1t'J (V) together generate 1t'j (X) and that the relations between
these generators still hold in 1t'1 (X). What it fails to say is what addi
tional relations are needed, or equivalently what the kernel of the
epimorphism looks like. If we try to interpret this geometrically, we
can see it as the uncertainty about uniqueness or otherwise of the list
[u]JU(l)' ... ,[unJU(n) due to the choices that had to be made in its
formation, and perhaps to questions of invariance of the list under
homotopies of w. The general rule should always be that if we
make choices we should see how much the end result depends on
the choices.
One of the subtlest choices we made was deciding in which open
set a given Wi was. This may seem a strange place to start, but it will
point out the necessity of thinking clearly. Suppose w was a closed
path in V n V; we would not need to decompose it at all, but if we
write iu: 1t'] (V n V) 1t'J (V) and iv: 1t'J (V n V) 1t'] (V), then it is
Van Kampen's theorem 217
clear that the two elements [w]u =ju[w]un v and [w]v =jV[w]Un v give
the same element of HI (X). In other words this subtle choice is
really basic: any closed path in U rl V 'decomposes' trivially in two
different ways giving the same element of HI (X). This first observa
tion suggests the following points:
1. The square
1t1(Un V )        l . ~ 1t1(U)
1t
1
( V)         l . ~ 1t1(X)
is commutative, so there is an induced homeomorphism, (J say, from
the pushout
to HI (X).
2. As iulwJu = iv[w]v for any [w]un V E HI(U rl V), we should 'kill
off' within HI (U) * HI (V) those elements of the form [W]ul W]v 1. We
do this by dividing out by the normal subgroup generated by them.
An efficient way to do this is to pick a set of generators z, say, for
HI (U rl V) and to kill off the words juz • jvz I for z E Z within the
presentation (x u y: r us) of HI(U) * H
1
(V), giving a group G with
presentation (x u y : r U s U {(Ju(z» UV(Z»I}). Here we have writ
ten ju(z) for any element of F(x u y) that maps down to ju(z) under
the natural composite epimorphism
F(x uy) ~ F(x) ~ (x: r);
similarly for jv(z).
Of course, we saw earlier that G is precisely the pushout
HI (U) It] (u*n V) H1 (V)
so really the two points are the same, that is there is an epimorphism
from G onto HI (X) induced by the inclusions of the subspaces. This
group G is often much smaller than the free product, so this is real
progress to cutting HI (X) 'down to size'.
n Kampen's theorem
We next turn to the other choices that were made in determining
the decomposition of w. We might expect that, each time, we would
find a new geometrically inspired set of relations to add in, further
reducing the size of the group we can construct and also that of the
kernel of the resulting epimorphism. What actually happens is differ
ent from this, and is slightly surprising, making one realize that the
end result is a very elegant and pretty theorem indeed (as well as
being an extremely useful tool.)
The first step is to investigate what happens when we add another
point in the decomposition. The reasoning behind this is the follow
ing. Suppose we decompose w using to, t10 ••• , tn and also using
t ~ , ... , t;"'; we would seem to get two elements g and g, say, in G. Are
these elements the same? Putting the two subdivisions of [0,1]
together gives yet another subdivision. If the resulting element g" of
G can be shown to be equal to g and to g', then of course g = g'. The
advantage of g" over g' is that the subdivision used can be obtained
by adding in finitely many extra points into that used for g and hence
if we can show that one extra point added into the subdivision
changes nothing we can show that g = g', and also g' = g" as hoped
for, simply using proof by induction. You may recognize the basic
method here. To prove invariance under choices, prove invariance
under basic elementary changes of choice (here adding in or taking
out a point, and previously by performing one of the Reidemeister
moves or Tietze transformations).
Suppose we take t' E [t
f
_ 10 til and without loss of generality we
assume that w maps [ti], til into U. In the decomposition before
adding 1', we have
[U;]u E 1l'1(X)
and after adding t' we choose a v from Xo to w(t') and find closed paths
Vi_I' wi [ti h 1'] • v = u, say v . wi [t
l
, t
2
] • vi = U', giving [u]u' [w]u in
place of [uau. However, [u]u' [u']u = [u • u']u = lUi], so nothing has
changed. If the section w, [t', ta is in Un V and v is chosen to be
within Un V then u' will be a closed path in Un V. Given that we
have already within the presentation of G allowed for the double
identity of such closed paths (within both 1l'J (U) and 1l'J (V» we can,
if need be, now consider u' as being a closed path in V and delete ti
from the subdivision (provided ti"* I of course). Thus the inde
pendence of the element defined by the decomposition on the choice
of subdivision is clear and the mechanism set up to handle the
problem of the closed paths with a 'split personality' allows one to
pass parts of a closed path from U to V and vice versa.
If we do not change the subdivision but merely the choice of the
Vi then again the same idea works (Fig. 10.3).
Van Kampen's theorem 219
Fig. 10.3
Writing Uj == Vil WiVi, u; == Vi_ IlVivj, Ui+) == ViWi+ tVi+ I, u;+ 1 ==
ViWi+ jVi+ j, we find that Ui' Uitj  Ui' (Vi' v;) . U;+ 1  u; • U;+ I within
X. If both Uj and Ui + I are in one only of U and V, this is unremark
able as the homotopies happen equally well within U and V, but if,
say, Ui is in U and Ui+ 1 is in V, then Vi and vi are in Un V (by the
way we chose them) and we obtain within the group G,
iU[Ui]U iV[Ui + d v = iu[uJu iV[Vi • u;Jv iv[u; + a v
= iu[uJu lviv[Vi' uJUl' v iv[u;; I]V
= iu[u;]u ivJulvi • vi]un v ivlu;+ dv
= iu[u;]u iv[U;+ I]V,
so the algebraic structure in G faithfully reproduces the geometric
structure within 7l'j (X). Note that we have used the 'iu(a) = iv(a)'
condition in the form luiu = lviv.
The final thing we have to check is that if Wo  WI then the resulting
elements of G are equal. We suppose given
h: [0, I] x [0, I] 7 X
such that h: Wo = Wt. Using the compactness of [0, I] x [0, I] we find
an £ > ° such that any set of diameter less than £ is mapped by h
either wholly into U or wholly into V. Next we subdivide the square
[0, I] x [0, 1] into rectangles small enough that the diagonal length is
(t,. Sj) L__________ "
Kampen's theorem
less than e, that is we pick numbers to = 0, tl> t
2
, ••• , tn _ b tn = 1,
So = 0, sj, ... , Sm _ I> sm = I so that each rectangle is mapped within
one of U or V. Using to = 0, I» ... , tn = 1 as our subdivision of (0, 1],
we get decompositions of Wo = hi [0, 1] x {O}. hi [0, 1] x {sJ,
11 1[0, 1] x {S2}, etc. As we hope to show that the elements go
and gl E G corresponding to Wo and w, are equal, it will be enough
to prove that the elements corresponding to any two successive
restrictions of 11, say 111[0, I] x {si} and 111[0,1] x {sj+d, agree as
iteration will do the rest. Because of this we will assume m = 1, so
our picture is
2 3
n 1 n
o
tn 1
As h is a fixedendpoint homotopy the ends of the square are both
mapped to Xo. We assume as always that paths Vi from Xo to wet;)
have been chosen. We have seen that the element of G that we obtain
is independent of the choice of the paths, so when we decompose
w' we will use paths v; obtained by composing Vi with the path
wi {ti} x (0, I] = (Xi (Fig. 10.4).
Fig. 10.4
Although we will write W' in terms of loops in U or V, it will as
usual be better to do the calculation in the fundamental groupoids
of X, U, V, and Un Vas follows:
" ,
w'  U
o
• U
1
• • . . • Un
where
, , ,,
Ui = Vi • Wi • Vi + b i=L ... ,n1
Van Kampen's theorem 221
within U or within V since the homotopy is just a reindexation via a
linear homeomorphism of [0, 1]. The path ai 0 w; 0 ai+ 1 is obtained
by mapping three sides of the (i + 1)st rectangle using h:
wi
I
~
I
a't
i+ 1
r"
~
Wj
This rectangle is mapped (by h) wholly within U, or wholly within V
(possibly within U (\ V). Thus the path ai 0 w; 0 iii + I determines
in the fundamental groupoid of U (resp. of V, resp. of U (\ V)
the same element as Wi' Composing either side with the Vi takes this
back into the fundamental groups, so the elements go and gl are
equal.
Thus we have investigated all the indeterminacy linking the ele
ment g to [w] E HI (X). In fact there was none. Given [w], g was well
defined and 8(g) = [w]. We have proved more since if we compose [w]
and [w'l in HI (X) it is clear that by picking wand w' in these classes
and composing one can pick decompositions already used for wand
w' to get one for wow', but any decomposition defines the same
element as none of the choices makes any difference. That is, the
assignment of g to [w] gives a homeomorphism
that 'splits' the epimorphism 8; in other words, 8a is the identity
homeomorphism on HI (X). The way in which a is defined makes it
clear that a8 is the identity on G, so HI (X) == G, the isomorphism
being induced by the inclusions iu, iv.
We have arrived informally at the result known as Van Kampen's
theorem. We state it formally in two different versions.
Kampen theorem (first form). Let X = U (\ V where U, V, and
U (\ V are nonempty, open, and pathwise connected. Pick
Xo E U (\ V. Then the following diagram is a pushout square of groups.
Van Kampen's theorem
        . ~ 1tl(X)
Van Kampen theorem (second form). Let X = Un V as before. Pick
Xo E U n V and suppose
ises
Jr
l
(U, x
o
) has a presentation (x : r),
Jr
l
(V, xo) has a presentation (y : s),
Jr
l
(U n V, xo) has a presentation (z : t).
Then Jr
l
(X, xo) has a presentation
(x u y : r U s u {Uu*(z» Uv*(z» ~ I : Z E z})
where Ju*(z) is a word in F(x) representing Ju*(z) and.iv·(z) is a word in
F(y) representing Jv*(z).
We have given above all the necessary building blocks for con
structing a formal proof of the Van Kampen theorem. We will, on
purpose, not give one ourselves, but suggest the following exercises.
I. Construct for yourself a formal proof of either form of Van
Kampen's theorem. You can choose whether or not to follow the
outline we have given. For instance, you may wish to verify the
pushout property of the square (*) directly rather than showing
that Jrl (X) is isomorphic to the algebraically constructed pu
shout.
2. Having attempted the previous exercise, go to your library and
search out books which contain a formal proof of the Van
Kampen theorem. Compare their methods with yours. Criticize
both your proof and theirs. Try to improve your proof with
ideas, phrases, or notation from their version. Do not think their
version will necessarily be fundamentally better than yours, but
Van Kampen's theorem 223
as the authors tend to have had years' more experience in the
game of writing proofs, you should expect to gain something, at
least in style.
3. Find a reference to Van Kampen's original paper. Try and read
his proof.
4. Several times we have worked within the fundamental groupoid.
Try to formulate and prove a groupoid version of the theorem
(then compare with the paper by Brown (1967».
11
Applications of the Van Kampen
theorem
Having got the Van Kampen theorem, how are we to apply it? As
with any result, before we try to apply it to a given example, we must
check that the conditions necessary for the theorem to hold are
satisfied in that example. In order to apply the Van Kampen
theorem, the parts into which we divide our space must be open and
path connected. As we will see, this sometimes requires an adjust
ment to the obvious choice of decomposition. Initially we will give
nearly all the details of this necessary adjustment, but later on we
will leave such details increasingly up to you.
The other thing that we will need is a starting point. The Van
Kampen theorem relates the fundamental group of a space to those
of its parts. If we want to calculate Jil of a space, X, we have to find
an open decomposition in terms of simpler spaces. Just as in a proof
by induction, we need to start somewhere and in this situation our
starting point is the proof that Ji
1
of the circle, Sl, is infinite cyclic,
that is free on one generator, namely the class of a path going once
around the circle.
We start with simple examples and get progressively to handle
more complicated spaces. As we do this, our calculations will more
often than not refer to earlier examples.
culations of fundamental groups
fed rose
So as not to be too ambitious to start with, we will first examine the
fundamental group of a one and a twoleafed rose.
n = 1. A oneleafed rose is just a circle, so Ji
l
(oneleafed rose) is free
on one generator.
Calculations of fundamental groups 225
x
Fig. 11.1
n = 2. Already here the main ideas for the general case occur. A
twoleafed rose (Fig. 11.1), has an 'obvious' decomposition (Fig. 11.2).
Fig. 11.2
We see that U
b
U
2
are path connected, as is U
I
n V
2
= {xo}. The
space X = VI U V
2
, but VI and V
2
are not open subspaces of X, so we
cannot apply the Van Kampen theorem to this present decomposi
tion. The necessary adjustment is easily made. We let N be a small
open neighbourhood of Xo in X itself. Thus N looks as shown in
Fig. 11.3.
Fig. 11.3
Now let VI = U
I
uN, V
2
= V
2
U N. These two subs paces are open
and path connected; we note that X = VI U V
2
and N = VI n V
2
• The
final criterion is our ability to calculate HI (VI)' HI (V
2
), HI (N), and the
induced homeomorphisms between them.
As VI and V
2
are clearly homeomorphic, we need only look at one
of them, say VI (Fig. 11.4).
This space contains VI and by retracting the two 'antennae', we
find that VI is a deformation retract of Vb so that HI (VI) == HI (U
I
)
and is freely generated by the class of the loop ai_ We will abuse
notation and write
plications of the Van Kampen theorem
Antennae
V=
Fig. 11.4
Similarly
Now N itself contains {xo} as a deformation retract, so Tel (N) is the
trivial group. This neatly absolves us from the task of calculating the
induced homeomorphisms
TeI(V
I
n V
2
) ~ TeI(V
i
), i= 1,2
since there is a unique homeomorphism from the trivial group to any
group.
We have now collected up all the information so as to calculate
Tel (X). The Van Kampen theorem gives a presentation
Tel (X) == (ah a2 : 0),
that is a free group on two generators corresponding to loops around
the two 'leaves' of the rose. This is as we might have guessed since a
typical loop in X will clearly wander around the two leaves one after
the other, perhaps going around ai' say n
l
times, then around a
2
112
times, then back around ai' and so on, corresponding to a word
starting a7'dl
2
a7' . .... Thus the Van Kampen theorem provides a
proof of the geometrically obvious result.
What about the general case of an 111eafed rose? The sort of
intuitive reasoning just used would suggest that if Xn is an llIeafed
rose with loops around the 'leaves' labelled ai' ... ,am then TeI(X
n
)
will be a free group on n generators corresponding to these loops.
The proof by induction of this is not too hard, but as it again
involves the addition of a small open neighbourhood of the basepoint,
we will include a sketch of it.
We note that an nleafed rose Xn can be written as the onepoint
union of Xn _ I with a circle (Fig. 11.5). This gives Xn = Xn _ I V Y with
Y a circle. This is not an open decomposition, so again let N be a
small open neighbourhood of the basepoint Xo in Xn (How small is
small? What were the necessary facts about the small neigh bourhood
N that was used in the twoleafed case? How do we know that such
a neighbourhood exists? We will return to these points shortly, but
will continue with the proof for the moment.)
Calculations of fundamental groups 227
v
o
Fig. 11.5
Set VI = XII _ 1 U N, V ~ = Y u N; then XII = VI U V
2
• VI n V
2
= N.
The basepoint is a deformation retract of N (if N has been chosen
small enough!), X" _ I is a deformation retract of VI. whilst the circle.
Y. is a deformation retract of V ~ . (Draw diagrams to illustrate this
and to convince yourself it is so.) Collecting up facts, we get
1. "1 (VI) == "I (X
II
_ I)  (a
1
• ...• all _ I : 0) by our (unstated) induc
tion hypothesis.
2. "1 ( V ~ ) == "I (Y) == (a" : 0) since Y is a circle.
3. "I (V1 n V
2
) == (0 : 0). the trivial group.
The formula given by Van Kampen now gives
"1 (XII) == (alo ...• an : 0)
as expected.
What about 'small'? Clearly picking an arbitrary neighbourhood
of Xo in Xu might result in one of the leaves being within N. How can
we be certain that this can be avoided? We have a lot of latitude in
our way of modelling the nIeafed rose as the property that we are
examining only depends on homotopy type. Hence we may assume
that ~ I is chosen to be a subspace of /R' with Xo the origin and each
leaf a vertical unit circle (Fig. 11.6).
Now let N = {x E XII: IIxll < O.I}. Since we have an explicit descrip
tion of XII and of N, our requirements can be checked explicitly if so
e.g.
x
Fig. 11.6
pplications of the Van Kampen theorem
desired. Care is required: as an example of what could go wrong
consider the spaces
C
n
= {x E ~ 2 1 1 1 x  (lin, 0) II = lin}
so that C
n
is a circle through the origin centred on the point (lIn, 0)
with radius lin. If m:t= n, C
n
n C
m
= {O}. Now let X= Un d', Cn. as a
subspace of the plane. If one looks at any neighbourhood, N, of 0 in
X, it must contain an open set of the form B(O; e) n X since it is an
open neighbourhood of 0, but there is some N£ E N depending on e
such that for n > N, 21n < e, that is C
n
c B(O; e) n Xc Nand N
contains infinitely many loops of X. The question of the structure of
HI (X) would lead us too far astray. Its structure is known, but if you
feel like trying your hand at it, bewareit is not that simple.
connected graph
If r is a finite graph with a
o
vertices and a, edges, than r is
homotopy equivalent to an nIeafed rose, where n = a
l
 a
o
+ I is the
number of edges not in a maximal tree. Using the previous example,
we find that HI (r) is free on n generators. How shall we interpret
these generators?
We suppose the basepoint of r is chosen to be a vertex, v. Pick a
maximal tree T in r, orient, and label the edges not in T, say
a" ... ,an' Now see what happens to these edges when we shrink
T to a point. They become a set of generating loops of HI (Xn).
Retracing the process back into r, we find that a set of free
generators for HI (T) can be constructed using the path classes
[aJ, i = 1, ... , n, where ai is a path that goes from the basepoint, v,
out in the maximal tree to the initial vertex of the oriented edge, ai;
it then crosses ai and goes back to v in the maximal tree (Fig. 11.7).
Here a
o
= 8, a
l
= 11, and n = 4. Now al = ebalc1e', a2 =
eba2cIeI, G3 = a3gf
I
, G
4
=fha
4
g
l
f
l
, and HI(T) == (ai' a2, G3'
a4 : 0) is the free group on four generators.
Fig. 11.7
b
Calculations of fundamental groups 229
Again, so as to gain some intuition of wha t to expect, we start with
almost the simplest example possible, a torus. We have already noted
that the torus can be realized as a product Sl x Sl, so by our earlier
results, we know that ni(torus) is the product of two infinite cyclic
groups C= x C=; it is therefore free Abelian on two generators. This
will enable us to check our calculation independently.
The torus can be obtained by gluing a rectangle as we have seen in
Chapter 4 (Fig. 11.8).
Fig. 11.8
Note that the boundary of the rectangle gives an embedded
twoleafed rose. We have to look for a suitable open decomposition
satisfying the conditions of the Van Kampen theorem. Such a
decomposition is more easily visualized from the rectangle than
from the other picture. The idea is to take one of the sets in the
decomposition to be the union of the two loops a, b, and the other
to be the interior of the rectangle. Of course, this does not work, as
the intersection is empty and the first set is not open, but by
'thickening' the first set up to an open neighbourhood we will get
around these difficulties. We do not, this time, need to be that careful
with the neighbourhood.
We first pick a basepoint, the choice of which is immaterial as the
space is path connected. However, a good choice of basepoint will
a
p
a
r(.x) p
~ t c
'b
r
b , , b= U
1
•
J / ~
'"
"'"'"
a
p
a
p
plications of the Van Kampen theorem
make the presentation simpler. so before choosing a basepoint, let us
see how the argument will go.
We let VI = X  centre of the rectangle, U
I
= boundary of the
rectangle; then VI radially projects onto VI and U
I
is a deformation
retract of VI: (see also Fig. 11.9).
a b
Fig. 11.9
After the identification of edges, the boundary, VI' of the rectangle
is a twoieafed rose embedded in the torus. We thus have
Now for V ~ : for this we take the interior of the rectangle, namely
V
2
= X  VI_ This is an open disc. and so 1£1 (V
2
) == (0 : 0), the trivial
group.
What about VI (1 V,? This is an open (rectangular) annulus
(Fig. 11.10).
c
Fig. 11.10
Thus 1£1 (VI (1 V
2
) == 1£1 (SI). an infinite cyclic group generated by a
loop around the hole.
Looking again at the pushout diagram coming from the Van
Kampen theorem. we have
Calculations of fundamental groups 231
; 1*
1t,(V}
          J ~ ~ 1t, (X)
We do not need to worry about j2* since ]fl (V
2
) is trivial; we do,
however, need to describe jl *. As ]fl (VI n V
2
) is infinite cyclic gener
ated by c, it is sufficient to calculate jl (c) in terms of a and h. (As
usual, we are using the same letter for a loop and the class it
determines.) The radial projection r pushes the loop jl(c) onto the
composite path abah, that is jl *(c) = ahaI b
'
.
At this stage, it pays to note that this neat expression for ji *(c)
occurs only if r(xo) is the starting vertex of the boundary surface
symbol. The structure of the group ]fl(torus) does not depend on the
choice of Xu, so Xu is for us to choose. Choosing this basepoint so that
r sends jl (c) onto ahah, rather than ahah or haha or something more
complex still, pays off in the simplicity of the resulting presentation.
The moral of the story is 'When a calculation does not depend on a
choice but cannot be made without that choice, choose wellit will
payoff later'.
We can now finish the calculation. Van Kampen's theorem gives
us a presentation
that is (a, h : [a, h]). This accords with our previous calculation since
if F(a,b) is the free group on generators a and h, what we have here
is its Abelianization which will be the free Abelian group on two
generators, that is C ~ x C ~ . Alternatively by considering the presen
tation (a, b : [a, h]) as a presentation of an Abelian group, we write
everything additively to get (a, h : 0)ah, that is C ~ EB C ~ (where by
tradition we replace the product, x, by the direct sum, EB, when
writing things additively).
A general surface (without boundary curves) is not very different.
We leave the details to the reader, but will give a sketch. The surface
can be represented by gluing a polygon using the corresponding
standard surface symbol, A, which is a word in edges ai' ... , an, for
some n (Fig. 11.11).
plications of the Van Kampen theorem
Fig. 11.11
Take V
j
= boundary of polygon (as a graph embedded in the surface)
and V
j
= X  centre. Choose a basepoint Xo that projects radially onto
the start vertex of A. Take V
2
= X  V
j
• Repeating the same argument
as before (details should be written down by you!), we get
Jrl (X) == (ab ... , an : A).
Thus if X is an orientable surface, T
g
, of genus g (with no boundary)
then n = 2g and A = alb1allbll ... agbgagjbg, so
Jrj(Tg) == (aj , ... , a
g
, b j , ... , b
g
: raj, b j] ... lag, bgD.
If X is a nonorientable surface, Sg, of genus g (again with no
boundary curves) then n = g and A = a i ~ , ... , a:, so
Jrj (Sg) == (aj , ••• , a
g
: aT ... a:).
In particular, for g = 1, Sj = IfJll, the projective plane, and
Jr
l
(1fJl2) == (aj : a ~ ) == C
2
, the cyclic group of order 2. Thus if a surface
has genus 1 then its fundamental group is Abelian.
If we allow boundary curves and thus cuffs, the picture is slightly
more complex. We will illustrate this with a simple example showing
how to adapt the previous argument to this case.
torus with two cuffs
The surface symbol is, of course, A = aba
j
b
j
C
j
d j cil c
2
d
2
c;j. Every
thing proceeds as before until we identify Jrj (VI) == Xj (Vj) where V j is
the embedded graph coming from the boundary of the polygon.
Instead of being simply an nIeafed rose for some n, Vj is a slightly
more complex graph (Fig. 11.12).
Surface groups 233
b
Fig. 11.12
Thus, although of course still free, ll'1 (VI) is not free on the set of
edges of the polygon. We will write df = CidiCi I, i = 1, 2, for the
generator of ll'1 (VI) corresponding to the two cuffs. The calculation
now continues more or less as before, giving us
ll'1(torus with two cuffs) == (a, b, d ~ , d ~ : [a, b] d ~ d ~ )
which is a free group of rank 3, since the relation allows us to write
d; in terms of the other generators. (Which Tietze transformations
are we using here?)
We will denote by Tg,r (resp. Sg.r) an orientable (resp. nonorientable)
surface of genus g having r cuffs.
11.1
1. Prove that if ll'1 (Tg) or ll') (Sg) is Abelian, then g = 1.
2. Show that ll'1 (Tg,r) is a free group on 2g + r  1 generators while
1X'1 (Sg,r) is free on g + r  1 generators.
3. Write out the proof that ll'1 (1fll2) is cyclic of order 2. (Although this
may seem a trivial exercise, it is important to gain an idea of why
this 2torsion is there as later we will be examining how to
construct a space whose fundamental group is a cyclic group of
given order n.)
face groups
We are now in a position to ask, and better still to answer complete
ly, the obvious question: can ll'1 distinguish between different sur
faces? In other words, if X and Yare surfaces, and ll'1 (X) and ll't (Y)
are known to be isomorphic, does this mean that X and Yare
plications of the Van Kampen theorem
s
homeomorphic? If we allow surfaces with boundary the answer is
clearly 'no', since both Tg,r and S2g,r have fundamental groups free of
rank 2g + r  I, yet we know that orientability is an invariant for
surfaces, so Tg,r and S21V are not homeomorphic. This suggests that
we ask the question in a stronger form by requiring the surfaces to
be without boundary curves. To attack this we will launch into an
analysis of the surface groups Jr
l
(Tg) and Jr
l
(Sg)'
We first look at their Abelianizations. Recall from Chapter 6 that
if a group G is given by a presentation (X: R), the Abelianization
G
ah
of G has a presentation (X: R u {[x, y]: x, Y EX}).
1. Jr
l
(Tg)ah == (ai' b
l
, ... , ag' bg : [a
b
bJ ... [ag, bg], [a
b
a2],
... , rag, bg]).
Thus the product of commutators [a
h
b
l
] ... rag, b
g
] is already a
consequence of the other relators and so may be deleted. (You
should check the justification of this using Tietze transformations.)
We have, therefore, that Jr
l
(Tg)ah is a free Abelian group of rank 2g.
2. Jr
l
(Sg)ah == (ai' ... , ag : aT ... a ~ , ([ai, aj]: i, j = I, 2, ... , g}).
Here it is simpler to use the alternative method of handling Abelian
izations of presentations, namely by writing everything additively.
This gives
You can now choose to attack this presentation formally with the
methods from the theory of finitely generated Abelian groups or you
can use additive Tietze transformations as follows.
Introduce a new generator b and set it equal to a
l
+ ... + a
g
; then
substitute and rearrange to get
(a
b
. •• , ag, b : 2b = 0, a
g
= b  a
l
 .••  a
g
_ I)'
Finally, eliminating a
g
gives an Abelian group of torsion free rank
g  I and a single generator of order 2, that is
g  lcopies
Thus given a surface X, we can tell if X is orientable or not by
looking for an element of order 2 in Jrl (X). Once we have determined
orientability or lack of it, we can find out which surface X is from
the genus which can be read off from our calculations of the torsion
free rank of Jr1 (X)ab.
Surface groups 235
Hopefully your curiosity has been aroused and you will have asked
whether Jrr can detect cuffs; that is, if we know X is a surface and
have some presentation (not necessarily the nice one we have given
above) of Jr
r
(X), can we tell if X has cuffs or not? If Jr
l
(X)"h has an
element of order 2, then X must have been a nonorientable surface
without cuffs, but suppose Jr
l
(X)"h is free Abelian of rank n, n even;
we could have obtained this answer from a T" with n = 2g or from a
T
2g
,r with n = 2g + r  I or from an Sg,r with n = g + r  1. Thus our
problem is to decide whether Jr
l
(Tg) is free or not. At present we only
have information about its Abelianization. Might not some cunning
subtle sequence of Tietze transformations applied to (ai' h" ... ,
ag, hg : [ai' hi] ... [ag, h
g
]) produce a way of eliminating the relator,
thus giving us a free group? Perhaps it seems unlikely, but can we be
sure? There are several different methods we could use, but the
easiest one for us is to use elementary ideals and other ideas from
Chapter 8.
Let r = a,h,ai'hi
'
... aghga;'h;'. We need to calculate dr/da/.; and
dr/dh/.; for each k. This gives
and
:;r = [ai' hi] ... [ak _ .. hk _ 1](1  akhkak I)
uak
The Abelianization of G = Jr
l
(Tg) was earlier seen to be free of rank
2g. To avoid problems with the multiple use of symbols, we set s[ and
ti to be the images in (C=)2
g
of at and hi respectively with y: F(a , . hi .
. . . , ag' h
g
) G and sd: G G
ah
, as in Chapter 8, the homeomorph
ism obtained by 'killing off' rand Abelianizing G respectively.
The images of the differentials are
sdy(k) = 1  t
dak k
and
sdY(:;k) = Sk  1
for k = 1, ... , g. This gives the following sequence of elementary
ideals in
Ei = 0 for i < 2g  1,
E
2g

1
=({Si1. 1  ti: i= 1, ... ,g}),
plications of the Van Kampen theorem
the ideal generated by the images of the differentials, and
E
2g
= Z(c=f
g
•
If G had been a free group of rank 2g (and given our knowledge
of its Abelianization, no other rank would be possible), then the
elementary ideals would be
E
j
= 0, i < 2g,
E
2g
= Z(C=)2
g
•
Thus, as seemed likely, 7r] (Tg) is not free and its lack of freeness is
detectable using the machinery we have at our disposal.
aces with finite cyclic fundamental group
We saw earlier that
7r](1FD2) ~ C
2
•
Can we find a space X which has 7r1 (X) ~ C
3
? Clearly X cannot be a
surface since our calculations showed that jp2 was the only surface
with nontrivial finite fundamental group.
A means of attacking this question is to rerun rapidly the ideas
that lead to the 2torsion in 7r] (1FD2). (If you have not written out this
calculation in detail yet, at least write it out in 'sketch' form now.)
The a
2
= I comes from the image of the loop in the annulus
VI 11 V
2
when projected onto Uj, the boundary of the polygon:
7r1 (VI 11 V
2
) ~ 7r1 (VI) ~ 7r] (U]).
This suggests that we take a triangle and identify the three sides as
if we had a symbol aaa (Fig. 11.13).
Now rerun the proof that 7r1 (1FD2) ~ (a : a
2
), but with 1FD2 replaced by
X and of course with the obvious modifications.
p
p
a
p
Fig. 11.13
Spaces with given fundamental group 237
No prizes will be given for instructions on how to construct a space
with HI a cyclic group of order 4. Here they are (Fig. 11.14).
There are no obstacles to using a similar construction to get a
space X= X(C
n
) with HI (X) == Cn, a cyclic group of order n. Try to
write down the details with a proof, using the Van Kampen theorem,
that your construction does in fact give what is required.
a
p
a
a
a
p
Fig. 11.14
aces with given fundamental group
The previous set of examples raises an interesting problem. Given a
group, G, can you construct a pointed space X with HI (X) == G?
To gain some insight as to how we might do this, at least for
groups G with a finite presentation (Xl> ... , Xn : r
l
, •.. , r
s
), we will
take another look at the construction of the space with Jr
l
(X) == C
3
that we examined in the previous section. We can think of this
construction as starting with a oneleafed rose into which we glue a
disc D2 via a suitable map f defined on its boundary circle, S. Recall
that this is done by taking VI ~ ~ and then gluing, using an
equivalence relation, , defined by: if X E Sl C D2, X  f(x). To get
X, we use a map f that sends a generating loop C E HI (Sl) into
a
3
E Jr
l
(VI)'
In general, if we have a space Y and an element r E Jr
l
(Y), then we
can choose a closed path a = ar:S
I
~ Y so that [a] = r and can use
a to glue a disc onto Y to form a new space X = Yu D2. One can
show, using the Van Kampen theorem as before, that if HI (Y) ==
(XI' •.• , Xn : r
l
, •.. ,r
s
) then HI (X) has a presentation HI (X) == (XI'
... , xn : r
l
, ••• , r
s
, r) in which r has been added to the relations in
the presentation. Using an inductive construction it is now easy to
construct a space having any (finitely presented) group G as its
fundamental group. (The problems of handling nonfinitely presented
groups are merely ones of technique; the same method works but the
justification that it works is technically slightly more demanding as
plications of the Van Kampen theorem
it uses a more powerful form of the Van Kampen theorem than we
have introduced above.)
As an example, we will see how to construct a space X having
Jrl (X) == D
3
, the group of symmetries of a triangle. This group has a
nice presentation, namely
D3 == (a, b : a
3
= b
2
= (ab)2 = I).
We set r
l
= a
3
, r2 = b
2
, r3 = (ab)2. Let r be a twoleafed rose
(Fig. 11.15).
b
Fig. 11.15
The relator a
3
E Jrl (r) has, as <t representing loop, a map a
l
from
51 into the lefthand circle of r. Gluing in a disc via this map gives
a space Y
1
with Jrl(Y1) == (a, b: a
3
= 1) (Fig. 11.16).
Y, a
p
Fig. 11.16
The element b
2
in this group is represented by a loop a2 going
around the righthand loop twice.
We use tX:2 to glue in a disc to get Y
z
(Fig. 11.17).
(If we write as before X(C
3
) for the space we constructed earlier
having Jrl(X(C3» == C
3
, then Y
2
= X(C
3
) V !fD2). The fundamental
group of Jr
1
(Y
2
) is C
3
* C
z
== (a, b : a
3
= b
2
= 1). The element r3 = abab
is represented by a loop lX:J going along a then along b then along a
and again along b. The space X = Y
2
D2 then has Jrl (X) == D
3
• (At this
point the artistic powers of the authors come up against insurmount
The group of a knot 239
a
p
p
Fig. 11.17
able difficulties, so we have not attempted to draw this space even
schematically. )
e group of a knot
In Chapter 6 we introduced the group of a knot and gave two
different ways of obtaining a presentation of that group from a
diagram of the knot. As there, we will write G(K) for the group that
results from a diagram of the knot, K. This notation presupposes
that G(K) is an invariant of the knot and not just something
concocted from one particular diagram. We saw, however, that the
Reidemeister moves left the group unchanged even though they
changed the presentation, so the notation is justified. We also saw
that the Alexander polynomial of K could be obtained from the
Wirtinger presentation of G(K) using the Fox derivatives and a sort
of Jacobian matrix, the Alexander matrix of the presentation. So
what is the geometric interpretation of this mysterious group G(K)?
Why is it invariant? Can we see this directly? In this section we prove
that G(K) is, in fact, the fundamental group of the complement of
K, that is 1f1 (1R3  10. Moreover, the Wirtinger presentation comes
from a fairly simple decomposition of 1R3  K after one application
of the Van Kampen theorem. The decomposition will need some care
in construction, but this will help later on when we come to analyse
the various fundamental groups involved and the homomorphisms
between them.
Suppose we are given a knot K which we think of. as usual, as
being a subset of 1R3. Orient the knot and label each arc of the knot
in the usual way. We shall use aI' a2' ... as labels following on
around the knot with the orientation. Thus a typical crossing would
look like that shown in Fig. 11.18 with, for some n, an+l being a
l
.
We next deform the knot to replace each such crossing by one
which looks as shown in Fig. 11.19, with most of the knot within the
plications of the Van Kampen theorem
Fig. 11.18
Fig. 11.19
z = I plane, but at each crossing we drop a vertical to the z = 0 plane
followed by a segment labelled f1j in that plane from the foot of the
vertical line ending at to the foot of that starting ai + I
Next we remove from ~ 3 a 'tunnel' neighbourhood N of K_ This is
best thought of as follows_
Imagine that we have a closed cube of side £ fixed to the knot but
able to slide along it. When it comes to a crossing section as above,
it slides down the first vertical with its faces parallel to the axes
in ~ 3 and then it goes along /3i and back up again to the z = I part
of at + \_ By taking £ small enough we can ensure that ~ 3  N is a
The group of a knot 241
Fig. 11.20
deformation retract of 1R3  K so that HI (1R3  N) IS isomorphic to
HI (1R3  K) which we want to calculate.
That part of the tunnel containing aj will be labelled Ai whilst that
containing /3.i will be called B
j
(Fig. 11.20).
We are now ready to describe a decomposition of 1R3  N as the
union of two open sets. We take
VI = {z > O}  N,
V
2
= {z < e/2}  N.
Now VI is halfspace with some tubeshape tunnels removed. It has
as a deformation retract an nIeafed rose with loops a
j
, ••• , an where
Fia.11.21
plications of the Van Kampen theorem
ai is a loop passing under the tunnel Ai that contains at (Fig. 11.21).
Hence 1l"1( VI) = (aI' ... , an : 0).
On the other hand VI is an open halfspace with trenches cut in it
(Fig. 11.22). It is therefore contractible and 1l"1(V
2
) = (0: 0).
Fig. 11.22
Finally VI n V
2
= {O < z < e/2}  N is an infinite plate with m
holes in it, where m is the number of crossings (Fig. 11.23).
Fig. 11.23
Thus 1l"1 (VI n V
2
) = free group of rank m. A typical generator will
correspond to a circuit around a hole and so its image in 1l"1 (VI) has
the form ajai1 lata; if the crossing is as shown in Fig. 11.24, or
aj I ai 11 ajai if the crossing is as shown in Fig. 11.25.
Fig. 11.24
Torus knots 243
Fig. 11.25
It is a good idea to try to draw diagrams to illustrate thisa
diagram drawn by you will be more use to you than one drawn by us!
Now the Van Kampen theorem can be applied and we find a
presentation
where a typical rk is as above. This is precisely the Wirtinger
presentation of the group G(K).
We can now see geometrically (or perhaps globally might be a
better term) why G(K) is an invariant of the knot K. Suppose K and
L are equivalent knots. Recall that this means there is a homeomorph
ism rp: 1R;3 ~ 1R;3 such that rpK = L. This homeomorphism therefore
restricts to one from 1R;3  K to 1R;3  L. This induces an isomorphism
between Jl'1 (1R;3  K) (= G(K» and Jl'1 (1R;3  L) (= G(L» as required.
The above argument also gets around the possible objection that
we picked a very special representative of the knot so as to be able
to use Van Kampen's theorem. If we had wanted to we could have
picked a homeomorphic image of the open decomposition {VI' V
2
}
of the complement of our original knot and could have applied
exactly the same argument to obtain Jl'1 of that complement. Of
course the description would have been a lot harder.
us knots
A knot K is said to be a torus knot if it can be embedded on the
surface of a torus, that is we have K: S ~ 1R;3 factors via some torus
Tl sitting in 1R;3:
plications of the Van Kampen theorem
SI ~ TI C ) 1R3.
Since Jr
l
(Sl) == C ~ , and Jr
1
(T
1
) == C ~ EB C ~ a torus knot K determines
a pair of integers (m, n) such that writing Jr1(SI) = (a : 0) and
Jr1(T
1
) = (b, c : 0) we have K*(a) = (b
m
, c"). The knot K thus goes
around one of the two generators m times and around the other n
times. As an example, it is easy to see that the trefoil is a (2,3)torus
knot as we have already mentioned (Fig. 11.26).
Fig. 11.26
As K is a knot one finds that m and n can have no common factor,
that is they are coprime. In this case it is simple to realize K in a nice,
regular way.
Consider a solid cylinder C with m line segments on its curved face
equally spaced and parallel to the axis. If the ends of C are identified
after a twist of 2Jr(nlm) then we obtain a single curve 1m n on the
surface of a solid torus T. This Im,n is an (m, n)torus knot:
We introduced these informally earlier and saw that the cinquefoil
or pentoil was a (2, 5)torus knot. In fact we examined many of the
invariants of these (2, n)torus knots as they formed the family able
to be drawn in the form shown in Fig. 11.27.
When we calculated the knot group of these we found
This is typical of torus knots. In fact
as we will see.
Torus knots 245
Fig. 11.27
The advantage of torus knots (and similar things hold for vari
ous others of the families of knots, e.g. pretzel knots) is that the
embeddability of these knots on a torus allows a direct application
of Van Kampen's theorem to 1R3  tm n' We consider tm n embedding
as above on the surface of a solid T sitting in 1R3. Picking a
small s, we carve out an sneighbourhood N of tm n in 1R3. Of course
1l'1 (1R3  N) 1l'j (1R3  t
m
•
n
). The complement of N i'n the surface of T
is an annulus, Lm,n, which results from the m parallel strips on the
cylinder being joined after a twist of 21l'(nlm) (Fig. 11.28).
Thus 1l'1 (Lm,n) is infinite cyclic, a typical generator being the class
of the centre line em n of Lm n'
The space T  N a solid'torus with a groove gouged in its surface
and so it deforms onto a central circle (marked a in Fig. 11.28), thus
1l'j(T N) (a: 0).
Similarly (1R3  T)  N is infinite cyclic generated by a loop b
through the 'hole' in T. (In fact 1R3  T is homeomorphic to a solid
torus with one point removedtry to see why.)
Since em, n makes m circuits of T, the image of em, n in 1l'1 (T  N) is
am. Similarly its image in 1l'j(1R3  T)  N) is b
n
.
We must now adjust our decomposition to be open so as to allow
ourselves to apply the Van Kampen theorem. We expand T  Nand
(1R3  T)  N slightly to get open sets Vj, V
2
with VI n V
2
a neigh
bourhood of L
m
, n Vb deforming to T  N, V
2
deforming to
(1R3  T)  N, etc. Applying Van Kampen's theorem then gives a
knot group, G(tm n) 1l'1 (1R3  tm n) (a, b : am = b
n
), as promised.
These groups much of'the information about the knot. As
with the surface groups, a group theoretic analysis is worth doing.
We leave it to you to calculate Alexander polynomials and will
instead use other methods in our analysis. (The Alexander polyno
mials are difficult to work with except when the smaller of m and n
is 2. It is, however, easy to work out what they are.)
plications of the Van Kampen theorem
//+1 m.n
Fig. 11.28
alysis of the knot group Gm,n
(This section uses slightly more group theory than the rest and can
be omitted.) We will write G
m
• n for G(t
m
, n).
We start by looking at the centre of Gm,n. Recall that if G is a
group, its centre Z(G) is the subgroup formed from the elements that
commute with everything, that is
Z(G) = {c E G: cg = gc for all g E G}.
Now since am = b
n
, am commutes with both a and b, so
am E Z(G
m
n) and of course, for any k, a
km
E Z(G
m
n) as a con
sequence. Let us write N = {a
km
: kEd"} for the s u b g r ~ u p generated
by am. This subgroup N is thus in Z{G
m
n) and so is normal, and the
corresponding quotient group Gm,nl N clearly has presentation
and so is isomorphic to C
m
* C
m
the free product of two cyclic
groups. Taking care over the notation we will write a = Na, b = Nb
for the elements of Gm,nl N.
At this point, we should explain the strategy of our attack, and
why we are looking at this quotient. If we are trying to write down
Z(G) for some group G, and we can find a nontrivial normal
subgroup N in Z{G), any element x in Z(G) which is not in N will
give a nontrivial element x = Nx in Z(GI N). So we can hope to find
out more about Z(G) by examining Z(GIN). If this is trivial then
Z{G) = N; if it is nontrivial, we can try to repeat the process. The
Analysis of the knot group Gm, n 247
group GIN is in some sense smaller and simpler (even where G
and GIN are infinite), so at each stage we have a better chance
of obtaining complete information on the centre of the resulting
quotient group.
In our case there are no problems since Z(G
m
nl N) is trivial, so
Z(Gm,n) = N. This statement seems intuitively in the
presentation of em * en the generators a and b are controlled only
by relations am = 1, b
n
= 1, each of which involves only one generator;
we have no interaction between a and b, so it seems unlikely that
there are any elements in em * en which are central. This is not
sufficient to be a proof, however, so we will have to probe a bit
deeper.
In em * en any element has a unique reduced form
where
o =:;:: Xl < m, 0 < Xi < m for i >
o =:;:: Yp < n, 0 < Yi < n for i < P
and all Xi, Yj E lL. Such an expression commutes with a only if it has
a at both ends and commutes with b only if it has b at both ends!
Hence a nontrivial element cannot commute with both a and b.
This completes the proof that Z(G
m
n) = (d
n
), the subgroup generated
'
We next look at elements of finite order in this group, em * en.
Suppose g E em * en G
m
nlZ(G
m
n)) has normal form aX! ••• bYp
and that it has order r < 00. The wo'rd (normal form of gt cannot be
shortened unless the normal form starts and ends with the same
letter, that is we must have one, but not both, of Xl and Yp equal to
zero.
If p = I this must mean that g is a power of a or b. Any conjugate
of a power of a or b will also have finite order. Are these all of the
elements of finite order? We will try an induction on p. We have
already seen that for p = I all elements of finite order are powers of
a or b. If p is arbitrary, assume that any g of finite order having length
less than p is a conjugate of a power of a or b. If Yp = 0, we have
(*)
is of finite order and has length less than p. By the induction
hypothesis this element is a conjugate of a power of a or b but then
g will also be such a conjugate. If XI = 0 a similar argument applies,
so we have that any element of finite order in em * en is conjugate
Applications of the Van Kampen theorem
to a power ofa or to a power ofb. In particular the elements of finite
order in G
m
nlZ(G
m
n) cannot have order greater than max(m,n).
If we make em * 'e
n
Abelian, we get em $ en which is of order mn.
As min(m,n) = mnlmax(m, n) we can theoretically determine (m, n),
up to order, if we are given G
m
n' Thus if {m, n} with m < nand
(m', n') with m' < n' are given and are distinct then G
m
n ~ G
m
, n' and
tm nand tm' n' are distinct knot types. "
By combining geometric arguments with this sort of analysis, these
results can be pushed further. We will not do this here but will note that
Neuwirth (1961) proves the following pretty results on torus knots:
1. If K is an alternating knot and G(K) has a nontrivial centre,
then K is equivalent to t
2
,2r + 1 for some r.
2. The torus knots tp,q with p > 2 and q > 2 have no alternating
projections.
(If you think a bit you will see that result I implies result 2, but
result I itself is much harder to prove.)
ses 11.7
1. (a) A space Y is constructed from a torus by gluing in a disc
along one of the 'vertical' sections as shown in Fig. 11.29.
Use Van Kampen's theorem to calculate the fundamental group
of Y.
Fig. 11.29
(b) A space Zg is constructed (rather as in (a» by gluing a disc
into a vertical circle in an orientable surface, T
g
, of genus g
having no boundary (thus Y = ZI)' Calculate 1I"j (Zg) , 1I"1(Zg)ab,
and the elementary ideals of 1I"j (Zg).
2. Find the fundamental group of 1hI;3  X when (i) X is the unlinked
union of two circles, (ii) X is the linked union of two circles
(Hopf link) as shown in Fig. 11.30.
Analysis of the knot group Gm, n 249
00
Ii} Iii}
Fig. 11.30 (i)
3. Let SI and S2 be two surfaces and S = SI # S2 their connected
sum. If A is a surface symbol for SI in edges aI' ... , am and B
one for S2 in edges b
l
, ••• , b
n
, find a presentation for
1rl (S1 # S2)'
12
Covering spaces
The fundamental group HI (X, x
o
) of a topological space X at a base
point Xo describes the variety of possible closed paths in X at Xo. Does
the algebraic structure of HI (X, x
o
) reflect further features of the
topological nature of X? In this chapter we shall see that the collection
of subgroups of HI (X, x
o
) corresponds to a collection of topological
spaces that map to X: the covering spaces of X. A covering space
preserves the local features of X but its overall topological structure is
generally more simple. Questions about X can often be transferred into
more tractable questions about a covering space. Our applications will
concentrate on the covering spaces of graphs, which we use to prove
two theorems about free groups: the NielsenSchreier theorem, which
says that a subgroup of a free group is itself free; and the Schreier index
formula for the number of free generators of a subgroup of a free
group. Further, we show how to discover free generators for the relation
subgroup of a group presentation from its Cayley quiver.
overing maps and covering spaces
es
Let X be a topological space. In this chapter, all spaces are assumed
to be path connected and locally path connected. A map p: Y ~ X
is called a covering map, and Ya covering space of X, if for each point
x E X there is a neighbourhood U of x such that pI(U) is the
disjoint union of open sets Vi of Y, so that p maps each Vi homeo
morphically to U. Informally, each x E X has a neighbourhood U
and Y contains a number of copies of U mapped to U by the covering
map p (Fig. 12.1).
1. The identity map X ~ X is a covering map. More generally, if F is
any space with the discrete topology, the projection map
F x X ~ X is a covering map. Given a neighbourhood U of
x E X, we have
Covering maps and covering spaces 251
Fig. 12.1
pI(U) = U {f} x u.
fEF
2. The map ~ ~ SI given by t ~ exp(it) is a covering map. A useful
picture of this map is to imagine the real line twisted into a spiral
(Fig. 12.2), as we did in Chapter 9 for the computation of
HI (SI, 1).
IR
Fig. 12.2
Covering spaces
3. The map P: S' Sl given by exp(iO) H exp(niO) is a covering
map. A small neighbourhood U of x E Sl has preimage pl (U)
consisting of n disjoint copies, illustrated in Fig. 12.3 here for
n = 3.
u
Fig. 12.3
4. is a covering space of the torus. Divide into squares, with
the points of 7l. x 7l. at the corners, and map each square to the
torus according to the usual gluing instructions.
5. Let X consist of two copies of Sl joined at a single point. Then
the Cayley quiver of the presentation (x, y : 0) of the free group
with basis {x, y} is a covering space of X. Edges coloured x are
mapped around one copy of Sl, and edges coloured y around the
other.
Covering spaces allow the solution of some maplifting problems.
Given maps f: Z X and p: Y X, a lift of f over p is a map
1 X Y such that pol = f. Lifts need not exist nor, if they exist,
need they be unique. However, if p: Y X is a covering map, many
maplifting problems admit unique solutions.
Uniqueness lemma. Let p: Y X be a covering map and Z a path
connected space. Then two lifts of a map f: Z X over p either are
equal or differ at every point of Z.
Proof. The details of this proof are suggested as exercises. Let us
suppose that f has two lifts.h and hover p. We want to examine
the points at which.h and h agree or differ, so let us set
ZO={ZE Z
We shall show that Zo is both open and closed in Z.
Exercise
Exercise
Exercise
Covering maps and covering spaces 253
1. Why will this suffice to prove the uniqueness lemma? Remember
that Z is path connected.
Let Z E Z\Zo. Since 1. (z) * h.(z), but poi. (z) = p 0 h.(z), it fol
lows that 1. (z) and h.(z) lie in distinct components of pI (U),
where U is some neighbourhood of fez) in X.
2. Use the continuity of f, .k and h to deduce that there exists
a neighbourhood of z contained in Z\Zo' Deduce that Zo is
closed.
Now suppose that z E Zo nfI(U); then 1.(z) and h(z) lie in
the same component of fI(U).
3. Show that z has a neighbourhood contained in Zo nfl(U) and
deduce that Zo is open. •
12.2 Pathlifting lemma. Let X be a topological space with a chosen basepoint
Xo. and let p: Y ~ X be a covering map. Given any path
CT: [0, 1] ~ X with CT(O) = Xo and Yo E Y with p(Yo) = x
o
, there exists a
unique path 0: [0, 1] ~ Y with 0(0) = Yo and p 0 0 = CT.
Exercises
Proof The uniqueness lemma (12.1) tells us that there is at most
one lift 0 of CT with 0(0) = Yo. Since p is a covering map, X is a union
of open sets U
i
such that pI (U
i
) is a disjoint union of open sets V ij
such that pi Vij: Vij ~ U
i
is a homeomorphism. If the image of CT is
contained in some single such U
b
then we define 0 by choosing some
Vij ~ pI (U
i
) and setting 0 = (p I Vijr 1 0 CT. In general, the image of CT
need not be contained in a single Ui, but we can nevertheless write
CT as a product of paths CT = CT
1
••••• CT
m
with OJ: [ti' ti + a ~ X (where
tl = 0 and tm + 1 = I) with the image of CT; contained in a single open
set as reg uired.
4. Show that such a decomposition CT= CT
I
••••• CT
m
is indeed poss
ible. To do this in detail, you will need some metric space properties
of the interval [0, 1]. Try to generalize step one of the proof of
Theorem 9.6.
5. Conclude the proof of Theorem 12.2 by showing that a lift oi can
be chosen for each CTi such that CT = 0
1
••••• om is a lift of CT. •
Covering spaces
Homotopylifting lemma. Let X, XO and Y, Yo be as in the pathlifting
lemma. If H:' [0. 1] x [0, 1] ~ X is a map such that H(O, t) = Xo =
H(l, t) for all t E [0, 1], then there exists a unique map
H: [0, 1] x [0, 1] ~ Y such that p 0 H = Hand H(O, t) = Yo for all
t E [0, 1].
Proof. This result is a twodimensional version of the pathlifting
lemma, and to prove it we can use the same ideas. We content
ourselves with a sketch. The unit square [0, 1] x [0, 1] is a compact
metric space, so may be subdivided, if necessary, into smaller squares
each of which is mapped by H into an open set of X on which a lift
to Y is easy to define. Definitions of these lifts can be made so as to
agree on common edges of the smaller squares, and so H may be
constructed. •
The path and homotopylifting lemmas indicate how the local
similarities between X and a covering space Y allow processes occur
ring in X to be mimicked in Y. An important consequence of this
idea now follows.
Theorem. Let X be a space with basepoint Xo and p: Y ~ X a covering
map, with basepoint Yo E Y chosen so that p(Yo) = Xo. Then the homeo
morphism p.: 1fj(Y, Yo) ~ 1fj(X, xo) induced by p is injective.
Proof. Let a be a closed path in Yat Yo such that p 0 a is homotopic
in X to the constant path at Xo: that is, [a] lies in the kernel of p*. Let
H: [0, 1] x [0, 1] ~ X be a homotopy from po a to the constant path
at xo, and let H be its unique lift over p. Consider how if maps the
unit square into Y. The lefthand side maps to Yo, and the top and
righthand side into p j (x
o
), but by the preservation of connectedness
(Theorem 3.6) we see that on all these three sides if must be constant
at Yo. The remaining side gives a lift of the path p 0 a over p and by
the pathlifting lemma this must be a itself. Thus H is a homotopy
from a to the constant path at Yo, and the class of [a] is trivial in
1fj (Y, Yo). •
Theorem 12.4 is the key that unlocks the structure of the collection
of covering spaces of a space X, for we see that to any covering space
there is associated a subgroup of 1fj (X, x
o
). This association is not
haphazard; it provides a neat algebraic description of the covering
spaces of X as part of a profound interaction between group theory
and topology. As our next step, we consider the implications of the
choice of a particular basepoint Yo E Y. Suppose we pick Yj E Y with
p(yj) = xo. Since Y is path connected, there is a path a: [0, 1] ~ Y
with a(O) = Yo and a(1) = yj. Clearly pea is a closed path in X at Xo
and conjugation by the class [p e oj in 1fj(X, x
o
) carries P*(1fj(Y, yj»
Covering maps and covering spaces 255
to P*(Jr1(Y, Yo)). Conversely, suppose that H JrI(X, xo) is conjugate
to P*(Jr
I
(Y, Yo»); that is, suppose that there exists a class [a] such that
H = [ar Ip*(Jr
1
(Y, yo»[a]. By the pathlifting lemma, there exists a
unique path a in Y with £l(O) = Yo and p 0 a = a. Set Yl = £l(1); then
H = P*(JrI (Y, Yl))' This argument establishes the following refinement
of Theorem 12.4.
Theorem. Let (X, x
o
) be a pointed space and p: Y X a covering map.
Then the subgroups {P*(JrI(Y, y))ly E pI(XO)} Jorm a complete con
jugacy class oj subgroups oj Jr
l
(X, xo). •
The pathlifting and homotopylifting lemmas were proved by
exploiting a particular feature of the unit interval and unit square,
namely compactness. What might we hope for as a more general
maplifting result for covering spaces, where no special features of
the domain can be called on to help? Suppose that p: Y X is a
covering map and that J: Z X has a lift 1 Z Y; we assume
basepoints Xo E x, Yo E Y, and Zo E Z are fixed such that p(Yo) = Xo,
J(zo) = X
o
, and l(zo) = Yo· Then the maps p*, .t: and .t. fit into a
commutative diagram of fundamental groups
from which it follows thatf.,(Jr1(Z, zo)) It is one of the
beauties of covering space theory that this necessary algebraic
condition for the existence of a lift of J over p is also sufficient.
Maplifting theorem. Let p: Y X be a covering map, with basepoints
Yo E Y and Xo E X chosen so that p(Yo) = Xo. Given a map J: Z X
carrying the basepoint Zo E Z to X
o
, there exists a lift 1 oj J over p if
and only if f.,(JrI(Z, zo» P*( Y, Yo))·
Proof. We have seen that the condition is necessary. To establish
sufficiency, we shall define 1 directly. Consider an arbitrary point
venng spaces
y
Fig. 12.4
Z E Z, and let G: [0, 1] ~ Z be a path with a(O) = Zo and a(1) = z; G
exists because of our assumption that all spaces are path connected.
Now foG is a path in X with f 0 a(O) = xo: by the pathlifting lemma
there exists a unique path a: [0, 1] ~ Y such that G(O) = Yo and
p 0 a = foG. We define 1 by setting I(z) = a(l) (Fig. 12.4).
We now have a candidate for I: at present we have a function
Z ~ X that satisfies p ol=f and depends on the choice of a path
in Z for each z E Z. We first show that the definition is in fact
independent of the choice of path, and then that 1 is continuous.
Given Z E Z, suppose that G and r are two paths from Zo to z,
and that r lifts to a unique path /3: [0, 1] ~ Y with /3(0) = Yo and
p 0 /3 = for. If "i' is the reverse of r then G'"i' is a closed path at
Zo E Z and soJ:([a· "i']) E P*(Jr
1
(Y, Yo». Let rbe a closed path in Yat Yo
with [f 0 (a· "i')] = [p 0 r] and lift the homotopy f 0 (G' "i') = P 0 r to
( Y, Yo). The lifted homotopy maps the unit square into Y according
to the following scheme:
y
Yo
and again by the preservation of connectedness, the righthand side
must be mapped to Yo. Therefore f 0 (G' "i') lifts to a closed path in
Covering maps and covering spaces 257
(Y, Yo) at Yo and so a(t) = 13(1). This establishes that the definition of
j is independent of the choice of path.
It remains to show that j is continuous. Let z E Z and suppose
that N is a neighbourhood of j(z) in Y. Let U be a neighbourhood
off(z) in X and Van open set in Y containingj(z) such that p maps
V homeomorphically to U. Thenfl(p(N n V)) is a neighbourhood
of z in Z and so contains a path connected neighbourhood W of z in
Z. We shall show that j( W) ~ N and this will complete the proof.
6. Why will the fact thatj(W) ~ V complete the proof?
Let z' E Wand now let r: [0, 1] ~ W be a path with reO) = z and
r(l) = Z'. Lift f 0 (0". r) to a path in Y beginning at Yo. Its end point
is j(z'). Now for has its image in peN n V) and its lift starts at
the end point of the lift of f 0 0", which is j(z). But pi (N n V) is a
homeomorphism, so that the end point of this lift lies in N n V. •
It is now possible to produce a hierarchical structure for the
coverings of a given space, based upon the interrelationships of
their fundamental groups. Consider two covering maps p: Y ~ X
and p': Y' ~ X and choose basepoints y E Y and y' E Y' with
p(y) = Xo = p(y'). Let us set H = P*(H
1
(Y, y)) and H' = p: (HI ( Y', y,)).
position. If H ~ H' then there is a covering map h: Y ~ Y' sending y
to y' and such that p 0 h = p'.
Proof. Apply the maplifting theorem and lift p over p' to obtain
h: Y ~ Y' with p 0 h = p'. We leave it to you as an exercise to verify
that the equation p 0 h = P', together with the fact that both p and p'
are covering maps, suffice to ensure that h is also a covering map .•
Of course, if H = H' then Proposition 12.7 works in both directions
to give covering maps g: Y' ~ Yand h: Y ~ Y' with p' 0 g = p and
p 0 h = p'. It follows that p 0 hog = p so that both hog and the
identity map id lift p to a map Y ~ Y, and so by the uniqueness
lemma we must have that hog = id. Similarly g 0 h = id and so
h: Y ~ Y' is a homeomorphism.
12.1
1. We call two covering maps p: Y ~ X and p': Y' ~ X equivalent
if there exists a homeomorphism h: Y ~ Y' such that p 0 h = p'.
We saw in Theorem 12.5 that a covering map determines a
complete conjugacy class of subgroups of H1 (X, xo). Show that
Covering spaces
two covering maps are equivalent if and only if they determine
the same conjugacy class of subgroups.
Deck transformations and universal covers
Let p: Y ~ X be a covering map. A deck transformation is a homeo
morphism h: Y ~ Y such that p Q h = p. A deck transformation is
therefore a homeomorphism of the covering space that preserves the
fibres of p: that is the subsets pI(X) for each x E X.
cises 12.2
1. Show that the set of all deck transformations for a fixed covering
map p: Y ~ X forms a group under the operation of composi
tion of ma ps.
2. Show that if h is a deck transformation and is not the identity
map. then for all y E Y we have hey) =I: y.
Is the group in Exercise 12.2.1 related to the fundamental group of
X, or to its subgroups? Note that h is a lift of h over itself, so that
by the uniqueness lemma, distinct deck transformations must differ
at every point of Y. We may translate the conclusion of exercises
12.2.1 and 12.2.2 so as to say that the deck transformations of a
covering map p: Y ~ X form a group that acts freely on Y. This idea
allows us to understand the deck transformations of certain covering
maps p: Y ~ X in terms of the fundamental group of X.
Theorem. Let p: Y ~ X be a covering map and let Q be the group of deck
transformations. If for some y E pI(XO), H = P* (Jrt ( Y, Yo») is a normal
subgroup of Jr
t
(X, xo) then X is homeomorphic to the orbit space Y/ Q
and Q is isomorphic to the quotient group Jr, (X, xo)/ H.
Proof. Given y E Y, let [y] denote its equivalence class in Y/ Q.
Since p Q hey) = p(y) for all hE Q, the function 8: Y IQ ~ X given
by [y] ~ p(y) is well defined, and is plainly surjective. For each
x E X and for y, y' E p'(x), we have that P*(Jrt(Y, y» = P*(Jr,(Y ,Y'»
since they are both conjugate, by Theorem 12.5, to P*(Jr,(Y, Yo»
which is normal. By Proposition 12.7 there is a deck transformation
h with hey) = y' and it follows that 8 is a bijection. We leave it to you
as an exercise to complete the verification that eis a homeomorphism.
Given a closed path a at Xo, lift a to a closed path a beginning at
Yo Then p 0 a(l) = Xo so that there exists a unique deck transforma
tion ka E Q such that ka(yo) = &,(1). Then the map [a] H ka is a
Covering graphs 259
surjective homeomorphism HI (Y, xo) ~ Q with kernel H. Again we
leave the verification of the details of this assertion to you. •
A covering map p: Y ~ X for which P*(HI(Y,Yo» is a normal
subgroup of HI (X, xo) is called regular. A covering map p: Y ~ X for
which HI (Y, Yo) = 1 is unique up to homeomorphism (by Proposition
12.7) and is a regular covering space of any covering space of X. For
these reasons, any such space is called the universal covering space of
X. Not every space possesses a universal covering space; see Arm
strong (1982) for a further discussion. For those spaces that do
possess universal covering spaces we can establish an exact corres
pondence between subgroups of the fundamental group HI(X, x
o
) and
covering spaces of X. This correspondence is a refinement of Theo
rem 12.8.
Theorem. Let X be a space with a universal covering space X. Then
1l'1 (X, xo) is the group of deck transformations of X, and if
H ~ HI (X, x
o
) then XI H is a covering space of X with H as fundamental
group. Thus if X is a space with a universal covering space, then there
exists a covering space corresponding to any subgroup of HI (X, xo). •
Covering graphs
We have seen how successfully the theory of covering spaces con
nects the topological nature of a space with the subgroup structure
of its fundamental group. We shall now concentrate on a particular
aspect of this theory: that of covering spaces of graphs. We have
seen in Chapter 10 that the fundamental group of a graph is always
a free group, so that graphs provide a topological model of free
groups, and we can use the covering space theory of graphs to
establish facts about free groups and their subgroups. This is the aim
of the remainder of the chapter.
A choice has to be made about the type of approach that we
follow. In our previous discussions, we have treated graphs essen
tially as combinatorial objects, their structure determined by the in
cidence between edges and vertices. Meanwhile our topological theory
has been developed in a very general setting. Now comes the crunch.
We could repeat the topological theory in a combinatorial setting,
and keep everything as close to our previous discussion of graphs as
possible. While the ideas may be more intuitive here, there are still
subtleties to be confronted, and for a careful account of covering
spaces from a combinatorial viewpoint we refer to Cohen (1989). We
shall make use of the topological theory, and this choice requires us
to set up a rigorous topological theory of graphs. This end could be
en ng spaces
achie:ved by modelling combinatorial graphs as subspaces of some
IR
n
, taking n as large as we need, but we face problems of ensuring
that this approach is consistent and independent of the embeddings
chosen. Instead we opt for a more abstract setting, and redefine a
graph. The new definition will define a graph as a onedimensional
CW complex. CW complexes are a class of topological spaces so
constructed that their homotopy theory avoids some pathologies that
arise in the more general theory. Happily, most spaces encountered
in everyday life can be given the structure of a CW complex. The
transition from combinatorial graph to CW complex and back again
will be fairly evident, and readers who prefer to keep their minds on
the combinatorial side of things should try to rephrase the definitions
and results according to their preference.
Let us proceed to the definition of a graph. A graph is a Hausdorff
space r together with a subspace rO such that the following condi
tions hold:
Gl: The subspace topology on rO is the discrete topology, and rO is
closed;
G2: r\ro is a disjoint union of open subsets e;t each of which is
homeomorphic to the open interval (O, 1) ~ IR;
G3: the closure e;t of e;t meets rO in either one or two points: in the
first case there is a homeomorphism/;t: [0, 1] ~ e;t mapping (0,1) to
e;t, whilst in the second case there is a homeomorphism f;t: Sl ~ e;t
mapping S\{ I} to e;t;
G4: a subset A ~ r is closed if and only if A Il e;t is closed for all A..
Properties GI, G2, and G3 have straightforward combinatorial
reinterpretations; G4 looks at first sight more mysterious, and indeed
is redundant if there are only finitely many e;t. Its purpose is to set
up the topology on r when there are infinitely many ek It is of course
convenient to retain the obvious terminology for graphs under the
new definition, so we call points of ro vertices, and each e;t is an edge.
12.3
1. The Hawaiian earring. Consider the following subset r of 1R2:
r = G {(x, y) I (x  (lIn)i + y2 = (lln)2}.
n= 1
We see that r consists of an infinite number of circles, each tangent
to the yaxis at the origin. Regarding the origin as rO, and each circle
as an edge, show that Gl, G2, and G3 hold, but that G4 fails. The
Covering graphs 261
Hawaiian earring is also infamous for not possessing a universal
covering space.
The first result in the covering space theory of graphs is that we
do not need to consider any other type of space. See Massey (1967)
for a proof of the following proposition.
Proposition.
a graph.
Let r be a graph and p: Y ~ r a covering map. Then Y is
•
The second result of importance is that any graph possesses a
universal covering space. Since this will be a graph with trivial
fundamental group, it must be a tree.
The first theorem that we seek on free groups now follows easily. It
was first proved by J. Nielsen in 1927 for finitely generated sub
groups of free groups, and then generalized by O. Schreier.
The NielsenSchreier theorem. Every subgroup of a free group is
free.
Proof. Given a subgroup G of the free group F, we construct a graph
r such that F ~ HI r, the universal cover r, and a covering space
Y = rl G of r with HI Y ~ G. Then G must be a free group because it
is isomorphic to the fundamental group of a graph. •
For subgroups of finite index in finitely generated free groups, we
can be more precise and compute the number of free generators
required for the subgroup. This result is another due to Schreier.
The Schreier index formula. Let G be a subgroup of finite index i in the
free group F of finite rank r. Then G is a free group of rank
i(r  1) + l.
Proof. We may construct a graph r with one vertex and with r
edges whose fundamental group is F. The covering space corres
ponding to G will have i vertices and ir edges, and since a maximal
tree must contain iI edges, there will be ir  i + I edges not in a
maximal tree, and this number of edges equals the number of free
generators of G. •
Finally we show how to obtain free generators for relation sub
groups in group presentations. Recall that the relation subgroup of
a group presentation (X : R) is the normal subgroup of the free
group with basis X generated by the elements R. If there are n
generators and the presentation is of a finite group G then the
ermg spaces
Schreier index formula tells us that the relation subgroup is a free
group of rank I GI (n  1) + 1. Can we produce free generators? The
Cayley quiver gives an easy method, for it is a covering graph of
the rose with n leaves, and its fundamental group is isomorphic to the
relation subgroup. All we have to do is find generators for the funda
mental group of the Cayley quiver, and see what are their images in
the fundamental group of the nIeafed rose. A detailed example will
illustrate the method involved.
Consider the presentation (a, b : a
3
, if, (ab)2) of the symmetric group
of order 6. Its Cayley quiver is shown in Fig. 12.5, with 1 marked as
basepoint and a maximal tree indicated.
a
Fig. 12.5
For each edge of the quiver not in the maximal tree, we write down
the label of the path that takes us from 1 to the source of the edge
in the maximal tree, then along the edge, and finally back to I from
the target vertex. In our current example there are seven such edges,
and their labels are
(1) baba
(2)
a'ba'b'
(3) ba
2
ba
2
(4)
a
2
ba
2
b'
(5) a
3
(6) ba
3
b'
(7) b
2
Covering graphs 263
These seven elements generate the relation subgroup. Each can be
written as a product of conjugates of the relators, since the relation
subgroup is the smallest normal subgroup of the free group with
basis {a, b} containing the relators. Note that (5), (6), and (7) are
conjugates of a single relator as they stand.
12.3 continued
2. Write the elements (1), (2), (3), and (4) in the example just given
as products of conjugates of the relators a
3
, b
2
, and (ab)2.
3. Use the Schreier index formula to compute the rank of the
relation subgroup for the following presentations:
(a) (x, y : x1yxy, yl xyX) == Q8, the quaternion group of order 8;
(b) (x, y, z : xyzl, yzx
1
, zxyl) == Q8;
(c) (x: x
60
), presenting the cyclic group of order 60;
(d) (x, y, z : x
3
. y4, Z5, [x, y], [y, z], [z, x]), which also presents the cy
clic group of order 60;
(e) (x, y : x
2
, y3, (xy)5) == A
5
, the alternating group of degree 5,
which again has order 60.
4. Draw Cayley quivers for examples (a) and (b) in Exercise 12.3.3,
and find sets of free generators for the relation subgroups in each
case.
Bibliography
Alexander, J. W. (1928). Topological invariants of knots and links. Trans.
Am. Math. Soc., 30, 275306.
Armstrong, M.A. (1983). Basic topology. SpringerVerlag, New York.
Heidelberg, Berlin.
Bleiler, S. A. (1984). A note on unknotting number. Math. Proc. Cambridge
Philos. Soc., 96, 46971.
Brown, R. (1967). Groupoids and Van Kampen's Theorem. Proc. London
Math. Soc. (3),17,385401. .
Brown, R. (1988). Topology. Ellis Horwood, Chichester.
Burde, G. and Zieschang, H. (1985). Knots. De Gruyter, Berlin, New York.
Cohen, D. E. (1989). Combinatorial group theory, London Math. Soc.
Student Text 14. Cambridge University Press.
Conway, J. H. (1969). An enumeration of knots and links. In Computational
problems in abstract algebra (ed. J. Leech). Pergamon, Oxford.
Crowell, R. H. and Fox, R. H. (1967). Introduction to knot theory. Springer
Verlag, New York, Heidelberg, Berlin.
Edmonds, D. (1960). A combinatorial representation for polyhedral surfaces.
Not. Am. Math. Soc., 7, 646.
Freyd, P., Yetter, D., Hoste, J., Lickorish, W. B. R., Millett, K. C. and
Ocneanu, A. (1985). A new polynomial invariant of knots and links. Bull.
Am. Math. Soc., 12, 23946.
Gross, J. L. and Tucker, T. W. (1987). Topological graph theory, Wiley
Interscience, New York.
Johnson, D. L. (1990). Presentations of groups, London Math. Soc. Student
Text 15. Cambridge University Press.
Jones, V. F. R. (1985). A polYI1omial invariant for knots via Von Neumann
algebras. Bull. Am. Math. Soc., 12, 10311.
Kauffman, L. H. (1987a). On knots, Ann. Math. Studies 115. Princeton
University Press.
Kauffman, L. H. (l987b). State models and the Jones polynomial. Topology,
26, 395407.
Kauffman, L. H. (1991). Knots and physics. World Scientific Publishing,
Singapore.
Kuratowski, K. (1930). Sur Ie probleme des courbes gauches en topologie.
Fund. Math., 15, 27183.
Lang, S. (1972). Differential manifolds. AddisonWesley, Reading, Massa
chusetts.
Lickorish, W. B. R. and Millett, K. C. (1987). A polynomial invariant of
oriented links. Topology, 26, 10741.
Bibliography 265
Magnus, W. and Chandler, B. (1982). The history of combinatorial r:roup
theor.v. SpringerVerlag, New York, Heidelberg, Berlin.
Massey, W. S. (1967). Algebraic topology. SpringerVerlag. 'e\\ York.
Heidelberg, Berlin.
Murasugi, K. (1987). Jones polynomials and classical conjectures in knot
theory. Topology, 26, 18794.
Neuwirth, L. (1961). A note on torus knots and links determined by their
groups. Duke Math. J., 28, 54551.
Priestley, H. A. (1990). Introduction to complex analysis. Oxford University
Press.
Reidemeister, K. (1932). Knotentheorie, Ergebnisse der Mathematik 1.
SpringerVerlag, Berlin.
Rotman, J. (1984). The theory of groups. W. C. Brown.
Schubert, H. (1956). Knoten mit zwei Briicken. Math. z., 65, 13370.
Sutherland, W. A. (1981). Introduction to metric and topological spaces.
Oxford University Press.
Thistlethwaite, M. B. (1987). A spanning tree expansion of the Jones
polynomial. Topology, 26, 297309.
Tietze, H. (1908). Uber die topologischen Invarianten mehrdimensionaler
Mannigfaltigkeiten. Monatsschr. Math. Phys., 19, 1118.
Trotter, H. F. (1964). Noninvertible knots exist. Topology, 2, 27580.
White, A. T. (1984). Graphs, groups and surfaces, NorthHolland Mathemat
ical Studies 8. Elsevier, Amsterdam.
Wilson, R. J. (1985). An introduction to graph theory. Longman, Harlow.
Index
Abelianization, see groups
Alexander matrix 1847
arc 3
augmentation 180
ball
closed 44
open 43
boundary curve 82
boundary edge 82
bridge 13
bridge number 13, 1035
Brouwer Fixed Point Theorem 209
commutator 116
commutator subgroup 116
components
of link diagram 3
of topological space 49
conjugacy 116
consequence 130
continuity 458
crosscap 748
crossing 3
change 11
number 11
cuff 76
decomposition, of abelian group 14J.
143
derivative 1827,235
diagram
connected 31
of knot or link 23
oriented 3
reduced 31
splitting marker in 23
state of 246
Van Kampen 133
writhe of 3
disconnection 48
domain, g.c.d. 1802
edge, of graph or quiver 163, 165
Edmonds algorithm 174
embedding 46
of graph in surface 1737
disc 176
minimal 176
Euler characteristic 8490
fibres 258
function
continuous 45
mapping cylinder of 51
pointed 208
genus
of a graph 176
of a knot 935, 1003
of a sum of knots 103
of a surface 8790, 92
graphs 16377
bipartite 169
circuit in 167
complete 169
connected 166
coverings of 2613
directed 163
embedding of 1737
fundamental group of 228
isomorphism of 164
planar 1713
Euler's formula for 171
rose 169,224
simple 167
subgraph of 165
topological 260
tree 1678
walk in 166
see also quivers
groupoid, fundamental 199, 215, 223
group ring 17882
groups
Abelianization of 117, 13940,234
centre of 246
dinclic 118
dihedral 1l218
direct sum of 142
finitely generated 140
fundamental 156, 202, 2369, 254
calculations of 203, 206, 22433
free 1228
rank of 128
see also word
free product of 147,216
with amalgamated subgroup 149
generalized quatemion 118
generators of 113, 128
knot 15062, 23943, 2469
pushout of 1469,217,221
surface 234
see also presentation
handle 758
homeomorphism 46
homotopy
equivalence 209
fixedendpoint 199, 207
inverse 209
lifting of 205, 254
of maps 199
pointed 199, 208
ideals
elementary 18892, 2356
invariance of 190
generated by subset 188
isotopy
ambient 52
class 6
of diagrams 6
invariant of 7
regular 6, 1920
isthmus 31
knots
achiral 17, 28
alternating 18, 94
bridge number of 13
chiral 17
cinquefoil 151, 155
colouring of 810, 1078, 1612
crossing number of II
diagram 2, 52
equivalence of 53
figureeight 2, 16, 27
generalized 159
granny 97
invertible 17
mirror images of II, 1517,28,37
Index 267
polynomial in,aflants c':_
polynomial
prime 1067
reef 97
torus 155, 1589. "
trefoil 2, 14. 27
knot type 53
link 52
Borromean rings
component of 3
diagram of 3
Hopf link 2.27. 36
linking number 
moves, Reidemelster :;. "I
1568
neighbourhood
orientation
of link diagram
of surface 66
overpass 3
path
closed 199
duration of 19
in a graph 166
homotopy 199
lifting of
product of 191<
reverse 198
in a space 199. :L; .::
polynomial
Alexander 3841. :""
bracket 206
HomOy 348. 109 I" :
Jones 2830
Kauffman 26
span of 31
presentation I"' I.
of Abelian group
of Abelianization 1"''1
Alexander matn" c.f : ... .!. 
consequence of rddth'::' 
Dehn 1546
mapping of
of quotient group
relation in 129
Tietze transformatlc.n .':
1902
Tietze's Theorem
Wirtinger 1514. 1:'6" '
Index
quivers 16571
Cayley 11822,1701,2623
isomorphism of 165
map of 165
relational 167
relation 129
retract, deformation 209
space, topological 438
connected 489
compact 214
convex 203
covering 2502
deck transformation of 2589
regular 259
universal 259
with given fundamental group 2369
Hausdorff 47
Hawaiian earring 260
pointed 208
simply connected 203
subspace 44
closed 44
diagonal 47
open 435
topology induced on 44
sum
direct, of Abelian groups 1424
of knots 96110, 1601
of surfaces 87
surfaces
boundary of 82
classification of 78
combinatorial 5764
double of 89
Euler characteristic of 8490
fundamental groups of 22933
Klein bottle 62
Mobius band 56
nonorient able 66, 80
orientable 66
pretzel 55
projective plane 65
Seifert 91
sphere 54
symbol 7190
normal form of 7890
torus 54
transformation, Tietze, see presentat
tree 1678
topology 435
discrete 434
finite complement 43
Hausdorff 47
identification 4951
indiscrete 49
product 47
usual 43
underpass 3
unknotting number 12
Van Kampen theorem 2212
applications of 22633, 2379, 243
245,248
winding number 2034
word 115, 1238
alphabet for 123
contraction of 124
expansion of 124
length of 115
letter of 123
syllable in 123
writhe 3,26
Knots and Surfaces
N. D. Gilbert
Department of Mathematical Sciences, University of Durham.
T. Porter
School of Mathematics, University College of North Wales, Bangor
OXFORD
NEW YORK
TOKYO
OXFORD UNIVERSITY PRESS 1994
Oxford University Press, Walton Street, O:iford OX2 Oxford New York Athens Auckland Bangkok Bombay Calcutta Cap Town Dar es Salamm Delhi Florence Hong Kong Istanbul Karachi Kuala Lumpur 111adras Madrid Melbourne Mexico City Nairobi Paris Singapore Taipei Tokyo Toronto and associated companies in Berlin Ibadan
6DP
Oxford is a trade mark of Oxford University Press Published in the United States
hy Oxford University Press Inc., New York
e s. D. Gilbert and T. Porter, 1994
All rights resened. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press. Within the UK, exceptions are allowed in respect of any fair dealing for the purpose of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act, 1988, or in the case of reprographic reproduction in accordance with the terms of licences issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms and in other countries should be sent to the Rights Department, Oxford University Press, at the address above. This book is sold subject to the condition that it shall not, by way of trade or otherwise, be lent, resold, hired out, or otherwise circulated without the publisher's prior consent in any form of binding or cover other than that in which it is published and without a similar condition including this condition being imposed on the subsequent purchaser. A catalogue record for this book is available from the British Library Library of Congress Cataloging in Publication Data Gilbert, N.D. Knots and surfaces I N.D. Gilbert, T. Porter. 1. Knot theory. 2. Surfaces. I. Porter, T. (Timothy), 1947II. Title. QA612.2.G55 1994 5J4'.224dc20 945209 ISBN 0 19 853397 7 Typeset by Pure Tech Corporation, Pondicherry, India Printed in Great Britain by St Edmundsbury Press, Bury St Edmunds, Suffolk
Preface
The fascination exerted by interlaced patterns and knotted forms is evident in their use in decorative and symbolic art across the centuries. Knotwork is a distinguishing feature of Celtic art, and the intricate beauty of decorated stonework and illuminated manuscripts such as the Lindisfarne Gospels, the Book of Durrow, and the Book of Kells manifests the technical and creative mastery over knotted forms achieved by Celtic artists before AD 800. The mathematical theory of knots has a comparatively short history, properly beginning with some remarks of C.F. Gauss and being advanced by Victorian pioneers inspired by Lord Kelvin's vortex theory of atomic structure. Whilst the vortex theory fell by the wayside, knot theory became a flourishing branch of pure mathematics. Recent developments have revealed unexpected connections with the theoretical physics of quantum field theory, reuniting knot theory with physics. One of the aims of this book is to provide the background material so as to make accessible the ideas and discoveries of this challenging and expanding area of mathematical research. Our main theme is of course the mathematical theory of knots, and especially its interaction with, and enrichment by, the theory of surfaces and of group presentations. We begin with a study of knots based on pictures of interlacing, and discuss two complementary problems. Firstly, what is the precise nature of the interlacing that we wish to describe? More precisely, when should different pictures be regarded as showing the same interlacing? Secondly, how can we classify the different interlaced forms represented by the pictures? Indeed, the fundamental problem of classifying knots, of organizing them according to their properties and characteristics, is a guiding theme of the entire book. There is a wealth of raw material for this process: the knots themselves! A few will be familiar from the practical problems of tying shoelaces, parcel string, mooring and rigging lines, safety ropes, and so on. Most interlacing patterns do not produce knots of practical use, but their decorative appeal remains, and their mathematical interest does not depend on their practical value. We have plenty of knots to work with: for example,
ce
it is a fact that there are precisely 12965 different knots that can be represented by interlacing patterns in which the string crosses itself at most 13 times. The pictures of interlacing represent spatial relationships in threedimensional space, and to pursue this aspect of knot theory we use topology, the study of shape and continuity. Topological methods provide new tools to aid our understanding of knotted forms and connect the study of knots with the second major subject of the book: the study and classification of surfaces. The immediate vicinity of any point of a surface looks just like twodimensional space, and we want to understand what the possibilities might be for the shape of the entire surface. Here we have a second topological classification problem, and one with a complete solution whose essentials were known in the nineteenth century. The classification theorem is a masterpiece of pure mathematics in action. A subtle problem is broken down into intermediate steps, the essential information is recorded by simple invariants, and the invariants can then be organized to produce a list of all the possible objects under scrutiny. Knots, surfaces, and the connections between them lead us naturally into other subject areas, principally to graphs and to group presentations. These subjects have their own motivating questions, their own deep problems, and their own fascinations and appeal. We try to draw out the connecting strands, the exchange of ideas, concepts, and methods. We hope that the reader will become happily entangled in the web that is woven by the theory of knots and surfaces. Durham Bangor Oct 1993
N.D.G T.P.
The first author carried out most of his share of the preparation of this book whilst a lecturer in mathematics at HeriotWatt University. We are especially grateful to the large number of students that have responded to our enthusiasm for the subject by electing to take knot theory as an option course at Bangor. Bangor. advice. Hugh Morton. He is also most grateful to Robin Nicholls for her assistance in preparing the typescript for the publisher. and in particular to Ronnie Brown. and commended at various points. complained. . and have criticized.Acknowledgements This book is based on a thirdyear undergraduate course given at the University College of North Wales. and thanks his former colleagues for the congenial working environment that they created. Peter Cromwell. The first draft of this book was based on lecture notes of improbable neatness taken by Ms Caren Williams in the academic year 19878. and Tom Thickstun for discussions. and the provision of some material. We are grateful to our colleagues who have been involved with the course at various times.
.
2. Knots.2 Isotopy and the Reidemeister moves 1.3 3colouring 1. and diagrams 1.2 Cutting and pasting 4.1 2.2 2.1 3. State models and the Jones polynomial The Jones polynomial Applications of the Kauffman polynomial The oriented or Homfly polynomial The Alexander polynomial 19 28 30 34 38 42 Topological spaces 3.4 2.4 Numerical invariants 1.1 The sum of oriented knots 5.5 Knots and surfaces 5.2 Genus of K + L 96 100 . The arithmetic of knots 80 84 90 96 5.1 Knot and link diagrams 1.5 3.Contents 1.1 Combinatorial surfaces 4.5 Chiral and invertible knots Knot and link polynomials 1 1 5 8 II 15 19 2.4 Invariants of surfaces: Euler characteristic and genus 4. links.3 Invariants of surfaces: orientability and boundary curves 4.3 3.4 3.5 4.2 3. Topological spaces Continuity Connectedness Identification spaces Knots and isotopy 42 45 48 49 51 55 57 68 Surfaces 4.3 2.
and the factorization theorem ncolourability Polynomials and the sum of knots 103 105 107 108 111 III 6.nts 5.9 Finitely generated Abelian groups 6.4 Free groups 6.5 5.11 Knot groups 115 118 122 129 131 138 139 140 144 150 163 7.6 Tietze transformations 6.3 Calculations of the fundamental group 9. and trees Examples of graphs Planar graphs Embedding graphs in surfaces 163 169 171 173 178 8.1 7. quivers.3 5.5 Alexander polynomials via the Wirtinger presentation 9.3 The Alexander matrix 8.4 Elementary ideals 8. Graphs and trees 7. Alexander matrices and Alexander polynomials 8.4 5.2 7.6 Bridge number of K + L Factorizations. prime knots. The fundamental group 9.4 Graphs.8 Abelianization 6.1 Group rings 8.3 7.3 Cayley quivers 6.5 Presentations 6.4 Homotopy equivalence 10.2 Dihedral groups 6. Van Kampen's theorem 178 182 184 187 192 197 197 199 203 208 212 .1 6.2 Derivatives 8.1 Paths 9.7 Quotient groups 6.2 Homotopy 9. Presentations of groups Examples of presentations 6.10 Pushouts of groups 6.
4 11.2 Deck transformations and universal covers 12.3 Covering graphs Bibliography Index 264 266 .7 Calculations of fundamental groups Surface groups Spaces with finite cyclic fundamental group Spaces with given fundamental group The group of a knot Torus knots Analysis of the knot group Gm .6 11.1 Covering maps and covering spaces 12.2 11. Applications of the Van Kampen theorem 224 224 233 236 237 239 243 246 11.3 11.5 11. n 12.1 11. Covering spaces 250 250 258 259 12.Contents xi 11.
.
1.2. and language is enriched by many figures of speech referring to knots. t and link diagrams Instructions on the skills of knotting are commonly given by means of pictures of the kind shown in Fig. l + II III Fig. Our subject is how mathematics gets into knots and our aim is to knit a coherent theory of knots from the ravelled strands of our cultural and practical fascination. or inconvenient.1 Knots. and diagrams Interlaced or knotted patterns and forms are encountered in many settings. 1.1 Some attempt may be made to render the appearance of the strands ropelike. 1. as with a knotted shoelace or inextricably entangled garden hose. 1. as shown in Fig. links. Interlacing motifs are the basis for much Celtic and Islamic art. but the significant information conveyed by the picture is the pattern of interlacing of the . as in the practical applications of knot tying. Their occurrence may be essential.
Figure 1. no formal definition will be given and none seems necessary at this point. Such pictures will be the object of study in this chapter and we shall see to what extent we can develop a pictorial theory of knottedness. in which a single strand is interlaced.links. and diagrams Fig. in which more than one strand is involved. In tying a knot in a piece of string. We call the pictures knot or link diagrams. There are a number of observations to be made.2 strands: a break in a strand indicates that the broken strand passes under the unbroken one.3 . together with the names commonly given to the knots and links represented. and some terminology to be established. 1.3 contains some examples. 1. we Trefoil Figureeight GO Hopf link Borromean rings Fig. and for links. for both knots.
4) these loose ends are shown spliced together. . In a knot diagram (Fig. using knot only when we wish to specialize to the case of one component. In an oriented diagram (Fig. The unknot.5. Fig. One method of remembering the allocation of signs is to imagine an approach to the crossing along the underpass in the direction of the orientation: if the overpass orientation runs from left to right.6) we distinguish two types of crossing and every crossing is of one type or the other.Knot and link diagrams 3 are left with two loose ends. whilst the Hopf link has two components. indicated on the diagram by arrows as shown in Fig. The unbroken arc at a crossing is the overpass. 1. and the Borromean rings have three. and the remaining two arcs form the underpass. a knot diagram has exactly one component. The writhe of an oriented diagram is the sum of the signs of all its crossings: we denote the writhe of a diagram D by weD). A link diagram may have any finite number of components. the trefoil and the figureeight have one component. 1. called components. the crossing is positive.4 A continuous segment III a diagram IS an are. and arcs meet at crossings. All knot and link diagrams will have only finitely many crossings. corresponding intuitively to the different pieces of string used in the physical knot or link being modelled. An orientation of a link diagram is a choice of direction on each component. there is no chance of the loose ends moving so as to undo the knot! To sum up then. and arcs may only meet at crossings. and so are knots. The splicing also fixes the knottedness in the diagram. and if from right to left then the crossing is negative. a link diagram encodes a way of interlacing one or more circles in three dimensions. By definition. The arcs in a diagram are divided into a number of separate classes. 1. and we shall use link as the general term. and this removes any need to treat the loose ends in a special way. 1.
links. and diagrams 00 Fig.6 / / Positive crossing Negative crossing A variation of the writhe is useful for the study of links. 1. with components CJ.Knots."" C m define the linking number of C i with 0. denoted Ik(Cb 0). The linking number of D is then the sum of the linking numbers of all pairs of components: Ik(D) = l~i~j~/n . to be onehalf the sum of the signs of the crossings of C i with Cj . 1.5 / mples / Fig. Given an oriented link diagram D. where C i and 0 are distinct components of D.
and only draw that part of a diagram at which the manipulation occurS. . The Reidemeister moves are the changes shown in Fig. Reidemeister moves first occur in the book by Reidemeister (1932). and that no other part of the diagram appears at the crossings changed by the Reidemeister moves. in which we look at the topological approach to knot theory. Two diagrams D and D' differ by a move of type RO if there is a continuous deformation of the plane that carries D to D' without changing any crossings. 1.8 to link diagrams. which for convenience we label RO. It is assumed that the diagram is otherwise unchanged. thanks to a deep theorem of Reidemeister.7 opy and the Reidemeister moves We now introduce the idea of isotopy.9). We shall meet Reidemeister's theorem in Chapter 3. We shall use a similar convention for the manipulation of diagrams in Chapter 2. and. an equivalence relation between diagrams that relates their study more closely to our intuitive notions of interlacing.Isotopy and the Reidemeister moves 5 Writhe = +3 Writhe = 0 Writhe = 2 Writhe = 0 Fig. We also consider a class of diagram changes. also to the study of knotted curves in threedimensional space. Moves of type RO correspond to pulling and stretching a knot without affecting its interlacing (Fig. though they are not Reidemeister moves. 1. 1.
9 . RI.links./ Fig. 1. and R3. and regularly isotopic if the transformation can be achieved without use of the move Rl. An equivalence @Fig.8 Link diagrams D and D' are isotopic if D can be transformed into D' by some succession of moves RO. It is clear that isotopy is an equivalence relation on link diagrams. and diagrams R1 XJ R2 Q _)(_Q I R3 I I . 1. R2. and the same is true of regular isotopy.
R3 alters the number of components of a diagram. Indeed. How can we decide if a given link diagram is in fact isotopic to an mcomponent unlink? What sort of properties should we seek. A more fundamental problem is to decide if two given link diagrams are isotopic. Hence the number of components of a link diagram is an isotopy invariant. for the Borromean rings. the linking number of any two components is zero. and the writhe of a diagram is not an isotopy invariant. Rt. An isotopy class of link diagrams will be our model for an interlaced form in three dimensions. . We deduce that the Hopf link is not isotopic to the Borromean rings! ~ To go further. for any assignment of orientations to the components of the Hopf link results in a nonzero linking number. let us consider linking numbers. we can alter the writhe at will by moves of type RI. We now go on to consider some isotopy invariants particularly adapted to the study of knots. Furthermore. linking number cannot help us to study the isotopy of knots. whilst a move R3 leaves the number of crossings and their signs unaltered. A move R2 removes or introduces two crossings of opposite sign. Since any two components of an unlink have zero linking number.Isotopy and the Reidemeister moves 7 class of diagrams modulo isotopy is an isotopy class. What are the problems that pictorial link theory must face? Ultimately we might hope for a classification of all possible knots and links. We shall call isotopy classes of link diagrams by the name customarily given to the associated link. and evidently diagrams in the same class represent the same essential pattern of interlacing. The obvious first classification of links is by the number of components: none of the moves RO. that deformations of interlaced circles in threedimensional space can be mirrored exactly by successions of Reidemeister moves on diagrams. in order to establish some means of comparison and differentiation of diagrams? Plainly. R2. a property useful for such a purpose must be isotopy invariant: if it holds for a diagram D then it must hold for all diagrams isotopic to D. the correctness of this model is in fact assured by Reidemeister's theorem mentioned already. the Hopf link is not isotopic to the twocomponent unlink. We can verify isotopy in variance by examining the effects of the Reidemeister moves (and the plane deformations RO). A Reidemeister move that involves at least two components of a link (and must therefore be an example of R2 or R3) leaves all linking numbers of an oriented diagram unchanged. in the form of a list of isotopy classes and a means of deciding into which class a given link diagram belongs. Check that this is so! However.
1. either all the arcs have the same colour. and diagrams 3colouring A knot diagram is 3colourable if we can assign colours to its arcs such that 3Cl: each arc is assigned one colour. 3C2: exactly three colours are used in the assignment. 1.... we show that the colouring at any crossing changed by the move can be amended if necessary so that the resulting diagram is again 3coloured. In Figs 1. 3colourability is an invariant of isotopy type.11 . Proof. mples { .. I \ Fig.10 Theorem.. we show only that part of the knot diagram changed by the Reidemeister moves. or arcs of all three colours meet.11 and 1. A deformation of type RO changes no crossing. For the Reidemeister moves. links..Knots.12. 3C3: at each crossing. so cannot change the property of 3colourability. Fig.
permutation of the colours will not affect the problem.. so knots exist! We 4 ** Fig. • Since the unknot is not 3colourable (why not?) the 3coloured diagrams above are not isotopic to the unknot. that is. for such changes might effect 3coloured configurations at other crossings. but you are not allowed to change the colours on arcs that leave the diagram.13 .') / ." f I \ \ .3colouring 9  .12 There are many cases to check: try some! You should bear in mind the following facts: to give a complete proof for the case R3. you need to be exhaustive. However. Fig. to have checked all the possibilities. 1. 1.
es 1. and arc 4 is then also forced to be red. If arc I were coloured red. An ncolouring of a knot diagram is an assignment to each arc of an element of lL n in such a way that. the figureeight knot in Fig.13 is not 3colourable. Find several examples of 5colourable knots. but this forces the other two crossings to contravene 3C2. Is this notion really a generalization of 3colourability? Investigate the two meanings of the term 3colourable and see that they do coincide. that is not isotopic to one another.14 are 3colourable? (a) (b) (c) (d) Fig. Let lL n denote the integers modulo n. then arc 3 would have to be red to satisfy 3C3. Then the whole diagram is red. at each crossing. which contravenes 3C2.14 2. and such that at least two elements of lL n are assigned.ots. For example. then arc 3 must be blue (at *) and arc 4 red (at **). 1. Prove that ncolourability is an isotopy invariant. the sum of the values assigned on the underpass arcs is twice that on the overpass. If instead we try arc I red and arc 2 white. links. 1. . and diagrams can also show that certain knots are different.3 I. 1. and arc 2 red as well. Can 3colourability be used to study links with m ~ 2 components? The notion of 3colourability generalizes to ncolourability. 3. Which of the knots in Fig.
This matches our intention to use isotopy classes as the models of interlaced form. and if we wish to particularize. Examples c(unknot) = 0. In considering such properties. xx Fig. Crossing changes will alter the isotopy type of a diagram. The crossing number c(K) of K is the mlllimum number of crossings in a diagram in the isotopy class K. given some arbitrary diagram.Numerical invariants 11 1. c(trefoil) = 3. The crossing number provides a means of arranging data on knots in a systematic form according to an intuitively reasonable measure of complexity. a topic to which we shall return shortly. The lists of knots given in Burde and Zieschang (1985) and Kauffman (l987a) are arranged according to crossing number: a minimal diagram is given for each knot with crossing number at most 9. the crossing number of the knot that it represents may be hard to determine. However. 1. A diagram of a knot K with c(K) crossings is a minimal diagram.15. c(figureeight) = 4. 1.15 . Choices have to be made of one mirror image or the other. remember that we shall call an isotopy class of diagrams simply a knot. Unknotting number A crossing change in a diagram exchanges the overpass and underpass at a crossing as shown in Fig. to use names of knots such as the trefoil or figureeight for the isotopy class containing the named diagram. Crossing number Let K be a knot. The crossing number is therefore the number of crossings in the simplest picture of a knot.4 Numerical invariants Properties of knot diagrams that are defined solely in terms of the isotopy class to which the diagram belongs will certainly be isotopy invariant.
Proof. to tryout a few examples of using crossing changes to produce unknot diagrams. at each first encounter with a crossing in the new diagram. before reading the following proof. Then changing at most el2 crossings of D produces a diagram of the unknot. However. The un knotting number is hard to calculate. You should certainly test the proof on some examples of your own devising. By construction. and is still unknown for some knots with crossing number 9. Let D be a diagram with c crossings. Proof.17 is isotopic to the previous Fig. Travel once around the diagram D from any starting point: at each first encounter with a crossing. links. Now the diagram in Fig. and then compare your procedure with the one suggested in the proof. (Bleiler 1984) The unknotting number of a knot does not necessarily occur in a minimal diagram. note whether you are on the overpass or the underpass. Try to find a proced ure for doing this systematically. Count the final tally of overpasses and underpasses and change all the crossings contributing to the smaller count: this can involve at most c/2 changes. no two crossing changes will unknot it.16. • The unknotting number u(K)of a knot K is the smallest number of crossing changes required to obtain the unknot from some diagram of the knot. You will probably find it useful. The essential difficulty is illustrated by the following. Let K be the knot with the diagram shown in Fig. 1. 1. you arrive consistently on the overpass or on the underpass.Knots. It can be shown that this diagram is minimal and is unknotted by three crossing changes. Note.16 . and diagrams Lemma. so that no interlacing ever occurs and the new diagram is an unknot. 1. Theorem.
• Note. 1. but is unknotted by two crossing changes. 1. Bridge number In fact u(K) = 2. See Bleiler's (1984) paper for more details.17 one. so is not minimal. By convention the unknot has bridge number 1. It has 14 crossings. A bridge in a knot diagram is an arc that is the overpass in at least one crossing. Example Fig. Example The standard diagram of the trefoil has three bridges as shown in Fig.18 The bridge number b(K) of a knot K is the minimum number of bridges occurring in a diagram of the knot.Numerical invariants 13 Fig. 1. .19.
1.20 Proposition. the diagram of the trefoil (Fig. so b(trefoil) ~ 2. This produces either a diagram of the unknot.20) has only two bridges. and a Reidemeister move R2 is applicable. Fig. Schubert: see his original paper (Schubert 1956) or the account in Burde and Zieschang (1985). links. is an overpass at any crossing. It is our convention that the unknot has bridge number 1. 1. there must occur two adjacent crossings at which the same arc is the underpass.19 However. So suppose that K is a knot with b(K) = I and let D be a diagram of K with one bridge and the smallest possible number of crossings. in which case the original diagram was the unknot. Therefore. 1. If m ~ 2 then K is the unknot.Knots. except the bridge. Proof. and diagrams Fig. which contradicts the choice of m. No arc of the diagram. m say. A knot K has bridge number 1 if and only if it is the unknot. so we may assume that m ~ 3. • It now follows that the trefoil has bridge number 2 (why?). The 2bridge knots have been classified by H. or else a Ibridge diagram with a number of crossings smaller than m. which reduces the number of crossings by two. .
Chiral and invertible knots 15 es 1. 3. 1. Construct an isotopy between the two diagrams in the proof of Theorem 1. What is the unknotting number of the knot with the diagram shown in Fig.23 indicates that the answer is no. 1.21.4 1. iral and invertible knots Need the mirror image D! of a knot diagram D be isotopic to D? Consideration of the trefoil in Fig. it appears that . 1. Construct an isotopy between the two trefoil diagrams in Fig. 1. Can you improve the bound on the number of crossing changes needed to unknot a diagram? Try some examples.21 4. 2. 1.22. 1.20. 1.22 5. Calculate the unknotting number of the knots shown in Fig. Fig. No amount of persuasive manipulation of a string tied as on the left seems to transform it to being tied as on the right.19 and Fig.3. IlQ7J b Fig.
1.25. However.24). and diagrams Fig.s. so does not change the isotopy class of the diagram (Fig. This is in fact the case. This last operation does not affect the interlaced form. links. 1. the following manipulations show that the figureeight and its mirror image are isotopic (Fig.24 The effect of a mirror reflection on a single crossing is to perform a crossing change as shown in Fig. 1. for a mirror image of D is obtained by changing all the crossings and then rotating the entire diagram through 180 about the line representing the mirror. though we do not have the means to prove it as yet.23 the diagrams lie in distinct isotopy classes.26). @®Fig. 1. 0 . To obtain D! from D we can simply change all the crossings in D. 1.
the knot is ehiral and we shall see in the next chapter that the trefoil is indeed chiral.27). Fig. 1. 1. 1.26 If D is an oriented diagram. Trotter found an infinite family of them (Trotter 1964).28. so that the figureeight is achiral. An oriented knot that is isotopic to itself with the orientation reversed is invertible. . Are D and 15 necessarily isotopic? Here we require the isotopy to match up the orientations. 1.F. Otherwise.27 Noninvertible knots were not known to exist until 1964. One is shown in Fig. 1.25 A knot that is isotopic to its mirror image is said to be aehiral. when H. the polynomial invariants of Chapter 2 give a ready means of deciding chirality for many knots. let 15 be the diagram with the orientation reversed.Chiral and invertible knots 17 x x Fig. Fig. The trefoil is invertible (Fig.
28 We close this chapter with a definition that will be crucial in the next. Not every knot possesses an alternating diagram: the first examples that do not have crossing number 8. Show that the figureeight and its mirror image are regularly isotopic. are indeed isotopic.27. links.ots. A knot diagram is alternating if you meet crossings alternately at overpasses and underpasses as you travel around the diagram. and diagrams Q ~ § ~ ~ ~ § n § Fig. 1. See the list in Kauffman (l987a) es 1. Simply flipping over shows that the two diagrams in Fig. 3. . 1. 2. Give a sequence of Reidemeister moves that effects the isotopy. Find a sequence of Reidemeister moves carrying the figureeight to its mirror image.5 1.
geometry. this generalization was accomplished virtually simultaneously by several groups of mathematicians. as well as to new interactions between knot theory. using a variety of approaches. and quantum physics. which provides a relevant theoretical parallel. Alexander (1928). The appearance in 1984 of a new polynomial invariant. Jones was awarded a Fields Medal (the equivalent in mathematics of a Nobel Prize) at the 1990 International Congress of Mathematicians.F. Jones. Subsequently.H. see Kauffman (1991). led to an upsurge in research activity and to a number of striking advances iOn our knowledge of knots and links and their representation by diagrams. Jones' polynomial may also be derived from some results due to L. Some of this work is phrased in the terminology of statistical physics. found by V. Kauffman. Let D and D' be (unoriented) link diagrams. e models and the Jones polynomial The title of this section is also that of Kauffman's paper explaining his construction (Kauffman 1987b). Jones discovered his polynomial as an offshoot of work in a quite different areathe theory of operator algebras.W. He tells the story in Jones (1985).2 Knot and link polynomials Polynomial invariants of knots and links have played an important part in knot theory since the work of J. The place of Jones' polynomial and its generalizations within the framework of preexisting knot theory remains a matter of much interest and no little mystery. Alexander's and Jones' polynomials were generalized in a twovariable polynomial. In recognition of his achievements. and finally look at Alexander's original treatment of his polynomial. and the Jones polynomial that is derived from it. We begin this chapter with Kauffman's work and examine his state model polynomial. Recall from Chapter 1 the notion of regular isotopy: D and D' are regularly isotopic if they . We then move on to the twovariable generalization.R.
BP3 is a shorthand for the following computational procedure. and a wealth of experience to come in the next few pages.1 Note that BP3 then implies the rule BPJ'. you obtain a new diagram R. \/X\ \ I = A 1)(> +A l/U\ \ \ f\ ) . will certainly convince you that the isotopy type of diagrams will be changed by these manipulations. 2. a little experimentation now. Take any diagram D and choose any crossing of D: delete the crossing and consider the now separated strands (Fig. x D L R Fig. The rule BP3 now says that (D) = A(L) + A1(R). Imagine travelling along each of the two strands in the remains of the overpass towards the former crossing point: join these strands to those from the underpass that lie to your left. If instead you join the strands from the overpass to those from the underpass that lie on your right. to produce a new diagram L. the move RI that introduces Or removes a loop is not allowed. and if D is a link diagram.1). The bracket polynomial of a diagram D is the Laurent polynomial (D) in one variable A defined by means of the following rules: BPI: BP2: BP3: (U)=I (DU) = . then we will use D U to denote a diagram with one extra unknotted component that introduces no extra crossings. We use U to denote a diagram of the unknot with no crossings.(A2 + A. 2.and link polynomials differ by a sequence of Reidemeister moves of types R2 and R3. The removal of crossings and the rejoining of strands are not Reidemeister moves.2 )(D) (X> =A (X> +A 1 ()C> In rule BP3 we use the convention from Chapter I of drawing only that part of a knot diagram that is changed in the intended manipulation. .
we assume this fact and go on to derive some of the properties of the bracket polynomial that underlie its theoretical significance and facilitate its computation.3.State models and the Jones polynomial 21 Fig. rna. making full use of the device of showing only that portion of a diagram at which changes are taking place. whilst BP3 enables the calculation of (D) for any diagram D by successive reduction of the number of crossings. For example. The bracket polynomial is an invariant of regular isotopy. To do this we have to change the diagram by successive reduction of the number of crossings. Fig.2 Some care is needed in applying these rules to the computation of bracket polynomials. 2. Proof. It is evident that BPI and BP2 together determine the bracket polynomial for any diagram of an unlink with no crossings. for the moment. with the figureeight knot we could reduce as shown in Fig. that is. Starting with a diagram D we can produce a polynomial (D).3 This raises the problem of deciding if the end result might depend on the route taken. applications of Reidemeister moves R2 and R3 to a diagram leave its bracket polynomial unchanged. 2. For move R2 we proceed as follows. 2. We shall see that (D) does not depend on the order in which the crossings in D are removed.2 or as in Fig. 2. (:s::.?)=A( Q)+A1(Y) (BP3) (BP3') .
.Knot and link polynomials /~ =(A 2 +A2 ) I . • Lemma..4 (a) ..1( } ) (BP3) (BP2) } = _A 3 \ ) ) The other case is left to the reader. • 0 c§J = <c2J> A<cQ»+A1 <c€» = _A4 <0 >_A.. . The bracket polynomial is not an isotopy invariant. 2.4 ) Fig.. + \ ( ') + \.. Reidemeister move Rl produces the following changes to bracket polynomials: Proof. We have 1 \ \ p) = A \) 0) +A= _A(A 2 +A2 ) \ ~) )+A.4 <0 > =(A 4+ A.. u ) \ ~) /} (BP2) Invariance under move R2 now implies invariance under move R3: \X) = A \'0) +A = I \)() (BP3) (R2) =A( : : ) +A 1 \}() \~) by symmetry..
4 + A8 Fig. <00> =A <cY0> +A. 2. 2.4 <6> =A 8 A 4 A 4_ A.State models and the Jones polynomial 23 D= <db> A <£>+A'<c0> A 7  A3  A. 2.5(a) or as shown in Fig. For each crossing in a diagram D there are two ways to eliminate it. We use Kauffman's idea of a state for a knot diagram.4 =A 8 A 4 + 1A. 2.5 )_A.1 <00> =A(A 7 A 3 _A.4 (b) We now proceed to some i11ustrative computations of bracket polynomials in Figs 24(a)(c).4 (e) ) We now turn to the effect on the computation of a bracket polynomial of a change in the ordering of the crossings and their removal.4 (A 4 +A. Having made such a choice. we attach to each crossing a splitting marker that connects the two regions of the diagram that will be .5(b). as we discussed in relation to the rule BP3 above: either as shown in Fig.5 8 <0b> A 7 A3 A 5 Fig.
2»k1. where k is the number of components. The bracket polyn~mial of the unlink diagram is determined by BPI and BP2: it is (_(A 2 + A.6. thus a diagram with c crossings has a total of 2 c possible states. .and link polynomials x Fig. we obtain a diagram of an unlink with the number of components in the unlink depending on the state.5 (b) Fig. 2. 2. 2.(a) or as shown in Fig.6 (a) )( joined by removal of the crossing: either as shown in Fig. 2. 2. Once a splitting marker has been attached to every crossing in D we have a state of D.5 (a) x Fig. If we now remove all the crossings in accordance with the splitting markers in a given state.6(b).
in that order.7 A splitting marker connects regions of the same sign. and (D) is therefore independent of the route chosen for its computation. We write lSI for the number of components in the diagram of unlinks obtained from D by removing all crossings in accordance with the state S: this diagram . the other two regions we sha11 call negative. Consider the four regions adjacent to a crossing in D (Fig... Let S be a state of the diagram D and let p(S). 2. Now define (DIS) = AP(s)n(s). 2. and in BP3 the bracket (L) is multiplied by A and the bracket (R) by A. with exactly one such term for each state of D. U is a diagram of an unlink.7): the two regions that lie between the overpass and an underpass.. /x ~ Fig. We can be a little more precise and obtain a formula for (D) as a sum over its states. Observe that in the earlier discussion of the elimination of crossings. U). n(S) be respectively the number of splitting markers in S joining positive and negative regions.I.6 (b) Now consider the computation of (D) for a diagram D. where U . 2.. when circling the crossing anticlockwise. This expression is uniquely determined by D. and the diagram R was obtained by connecting the two negative regions. in order to apply BP3 the diagram L was obtained from D by rejoining the separated strands so as to connect the two positive regions. Repeated use of BP3 results in an expression for (D) as a sum of terms ±Am(U.State models and the Jones polynomial 25 Fig. we shall call positive.
Proof. however. Therefore. Theorem. the term (. At first sightand certainly according to the spirit of Chapter Ithe bracket polynomial is not much use in the study of knots and links.Knot and link polynomials receives coefficient (D I S) in the expression for D. so is f[L]. • The formula just given does not provide a practical method for the evaluation of (D) but it is of theoretical interest because of its resemblance to the form of expressions found in statistical physics.2 )]I I . The polynomial f[L] is an isotopy invariant. that is. it is an isotopy invariant. since it is not an isotopy invariant. this deficiency is easily remedied to produce an isotopy invariant for oriented links. Since w( D) and (D) are invariants of regular isotopy. We can summarize the above remarks in a formula for D. Define the Laurent polynomial f[L] by f[L] = (.A). we need only to check the effect of Reidemeister move RI. Let L be an oriented link and let D be an oriented diagram of L with writhe w(D). Then (D) =L s (D IS) [_(A 2 + A.3 w(D) (D) where (D) is the bracket polynomial of the diagram obtained from D by simply forgetting its orientation. • Combining our calculations of bracket polynomials and writhes. it is an invariant of regular isotopy. Recall from Chapter I that the writhe is not an isotopy invariant.I S where the sum is taken over all states S of the diagram D. Happily.A)3W(L)in f[L] contributes an additional factor . Proposition. we have the following computations of Kauffman's polynomial f[L]: . and an excellent place to seek further information is Kauffman's book (Kauffman 1991). The notation already suggests that f[L] is determined by L. Let D be a link diagram. It is a pleasing and surprising nexus between two strands of mathematical thought. One application of this move either introduces or removes a crossing and so changes w(D) to w(D) ± 1. as may easily be checked by examining the effect of Reidemeister moves R2 and R3.2.A± 3 which is exactly cancelled by the factor contributed by (D) and given by Lemma 2.
10) f[L] Fig.8) f[L] = = _A6 (A +A. 2. 2.9 (3) L = righthand trefoil (Fig. 2.State models and the Jones polynomial 27 (1) L = lefthand Hopf link (Fig.4 + Aa Fig. 2. 2.11) f[L] =<D> = Af!. A4 + 1.9) f[L] Fig.. 2.A.10 (4) L = figureeight knot (Fig.) 4 4 _A 10+ A2 Fig.8 (2) L = lefthand trefoil (Fig.11 . 2. 2.
2. 2. Compute (D) directly from the formula in Proposition 2. Moreover. 2. as our first formal property of the Kauffman polynomial. It is therefore reasonable to propose.I? If the answer is yes then achiral knots must have Kauffman polynomials that are symmetric in A and AI. 2. Since it is also the case that w(L!) = .15. 2. Compute f[L] for L equal to the links in Fig.3 for the diagrams D in Fig. the following result.13 are not regularly isotopic. This observation prompts the question: is it always the case that the Kauffman polynomial of a mirror image is obtained by interchanging A and A. 3. We know from Chapter I that a diagram of L! is obtained from one of L by reversing all the crossings.w(L). (L!) is also obtained from (L) in this way. apart from a change of variable.1 1. • 2. osition.14. Jones polynom ial The Jones polynomial and the Kauffman polynomial are really the same.and link polynomials Observe that the Kauffman polynomial for the righthand trefoil is obtained from that for the lefthand trefoil by interchanging A and AI. Show that the two diagrams in Fig. 4. so that the Jones polynomial . Proof. and a computation of (L') from this diagram proceeds exactly as for (L) except that the roles of A and AI are interchanged.12 2. Fig. P are regularly isotopic. Let L be a link diagram and let L'be its mirror image. and this is the case for the figureeight knot. Show that the two diagrams in Fig. Then f[L!] is obtained from f[L] by interchanging A and AI. the result now follows.
14 ®~ c@ Fig. The change of variable implies that the Jones polynomial can be defined by the following two rules: JPl: Vu(t) = 1. ~ .13 cg & Fig.f (t)tV (t)=(t I12 _t 1I2 )V ~ (t). The formula connecting V L and I[L] is then VL(t) = f[L](tli4). where U is the oriented unknot.15 is an isotopy invariant of oriented links.. The Jones polynomial of an oriented link L is denoted V L and the customary choice of variable is t.. 2.< . JP2: tIV )<.The Jones polynomial 29 Fig. 2. 2..
It tackles the question: how is the Kauffman or Jones polynomial of an oriented link affected if the orientations of some of the components are reversed? cises 2.3 (Exercises are placed at strategic points throughout this section.4. and f[K] =f[K].] Consider the general setting of a ccomponent oriented link L and a link L' obtained from L by reversing the orientation of r components of L. Show that VL(f) = f[L](r I!4) does satisfy JPl and JP2. If K is an oriented knot. How are f[L] and f[L1 related? . K is the reverse of K.2 Fig. how is f[L] related to f[L']? Experiment with some twocomponent links and conjecture a formula relating f[L] and f[L 1. 2. identical except at one crossing. 2.1. and are numbered concurrently within it. If L is a twocomponent oriented link. Your formula should involve the linking number of the two components.) 1. 2. [Hint. at which we find the configurations shown in Fig.16.Knot and link polynomials In JP2 we show only that part of the link at which changes occur: so JP2 relates the Jones polynomials of three links. and the orientation of just one component is reversed to form a new oriented link L'. Use JPl and JP2 to compute the Jones polynomial for the links in Exercise 2. Try to prove that your conjecture is correct. Applications of the Kauffman polynomial Kauffman's approach to the Jones polynomial enables relatively simple proofs to be given of some of its formal properties.16 1. x x cises 2. show that 3. one such property is the reversing result. 2.
rna.A 5 .17. the crossing number in a reduced diagram is an isotopy invariant.A. Not every link has an alternating diagram. see Section 7. what is the relationship between the span of (D) and the span of f[L]? . Then D divides the plane into c + 2 regions. Firstly we shall define reduced diagrams. The special case of Euler's formula that we need is the following.Applications of the Kauffman polynomial 31 One of the most striking applications of the Kauffman polynomial concerns its relationship to the crossing number of a knot.7 has span 5 . recall that (D) may contain both negative and positive powers of A.(.7) = 12. as we have done so far. If D is a diagram of link L. as in the exam pIes in Fig. but strongly recommend that the reader examine the alternatives.3 4. . An isthmus in a link diagram is a crossing at which only three regions of the diagram meet. 2.3 + A. Proofs of the result have been given in Kauffman (1987b). Fig. The result may be summed up as follows: any two reduced alternating diagrams of a link L must have the same number of crossings. and is connected if it is not the disjoint union of two smaller diagrams.2.' We need one more piece of terminology before proceeding to the results. 2. since it is most naturally seen as a special case of Euler's formula for planar graphs. Clearly the definition of span extends to any Laurent polynomial in one variable. As an example.17 A diagram is reduced if it does not contain an isthmus. 2. . Murasugi (1987). We shall follow Kauffman. The meaning of the term reduced will be explained in due course. The span of a bracket polynomial (D) is the difference between the maximum and minimum powers of A occurring in it. and Thistlethwaite (1987). We have need of one result whose proof is deferred until Chapter 7. but the result asserts that for those that do. Let D be a connected link diagram with c crossings.
c . Sm = Sf . in an alternating diagram.2 . The sign of a region is determined at a crossing to which it is adjacent.2 and the term in (D) of minimum degree is (. Proposition. so that every region can be given an unambiguous sign. Now n + p is the total number of regions in D. Then IS I = nand (D IS) = A C and we see that S contributes a term (l)n.2 to (D). We then show how the theorem is derived from the proposition.2t. reduced. Let D be a connected. Show that I SII < 1 I.J)PI A. and in general the allocation of signs will not be consistent for the same region at different crossings. The proposition uses the notion of positive and negative regions in a link diagram.J)nI A c + 2n . Proof of the theorem. [Hint. reduced. alternating diagram D with c crossings then the span of (D) is 4c. so that peS) = c and n(S) = O. Hence any two connected.1 to the expansion of (D) as a sum over its states. S]. Then the term in (D) of maximum degree is (. and follow it by a more technical proposition that contains the core of the argument. Theorem. alternating diagrams of a given link have the same number of crossings."" SmI.] Now consider an arbitrary state Sf.Knot and link polynomials We now state the main theorem on crossing numbers and the Kauffman polynomial. Therefore the span of (D) is 2c + 2(c + 2) .lAc + 2n . • Proof of the proposition.4. S2.C . the allocation of signs is consistent. Use the fact that S D is reduced. However.1 A c (A 2 + A.2p + 2) = 2c + 2(n + p) . which is also equal to c + 2. alternating link diagram with c crossings. and so a term (.4 = 4c. reduced. and give the proof of the proposition assisted by a series of exercises.3 5 Consider a state SI obtained from S by changing exactly one splitting marker. and n negative regions.(. There exists a sequence of states S= So. Let S be the state of D with all the splitting markers joining positive regions. From the proposition we see that the span of (D) is c + 2n .2p+2. p positive regions.l)n . introduced earlier in the chapter. If a link L possesses a connected. rcise 2. We wish to show that all other states contribute terms of smaller degree.
6 to show that S' contributes terms to (D) of degree smaller than c + 2n . 9. If a link L possesses a connected diagram D with c crossings then the span of the bracket polynomial (D) is at most 4c. Let S be a state of a connected diagram D and let S be the dual state. Then I S I + I S I ~ r. Exercise 2.j ± 1. Use the connectivity of D to show that at least one of Land R is connected. Proof.5 and 2.3. Prove the statement in the proposition for the term of minimum degree. 8. 2. This theorem follows from a result that Kauffman calls the dual state lemma. some parts of the proof are left to you. which we state at once. Suppose that D has r regions.1 and n(Si+ I) = n(Si) + 1.3. • We now consider a generalization of Theorem 2. This completes the proof of the proposition for the term of maximum degree.3 10. Given a state S of a connected diagram D. the dual state S is obtained from S by changing all the splitting markers. 7.3 6. The splitting to form L is done according to a splitting marker in one of the states S or . Combine Exercises 2. Now suppose that D has c crossings: select one crossing and split D at that crossing to obtain diagrams Land R. Show that I Si+ II = I S. Verify the dual state lemma for diagrams with at most two crossings. Therefore P(Si+ I) = p(SJ . and deduce that the largest degree of a term in (D) contributed by Si + I is less than or equal to the largest degree of a term contributed by Si. We may assume that it is L that is connected.7 observed independently by Murasugi and by Thistlethwaite. however.2.10 Dual state lemma. 2.3 Once again.Applications of the Kauffman polynomial 33 such that Si + I is obtained from Si by changing exactly one splitting marker. Exercise 2. Exercises 2. we continue to follow Kauffman's account of the result. and it follows that (DISi+l) = A2(DISi).9 Theorem.
and K.L. 12. Show that I S' I + 1 ~ IS I· Complete the proof of the dual state lemma. and so a term of maximum degree c + 21 S 1 . 13. Use induction on c to deduce that 1 S' [ + [S I ~ r . Conway in his work on the enumeration of knots and links (Conway 1969). so that p(S) = c and n(S) = O. Let L be an oriented link.1 regions.C. where D has r regions.(A2 + A. by interchanging these states if necessary.I to (D). J. 1985) and the polynomial was dubbed the H.c .8 shows that S contributes a term A('[.3 11. Now the fact tha t r = c + 2 completes the proof. Similarly S contributes the term of minimum degree . Hoste. Millett. W. The proof of Proposition 2. Lickorish.M. Note that 1S' 1 = 1S I· 2.R. These similarities indicated that the Alexander polynomial and the Jones polynomial were special cases of a more general construction. Ocneanu.2)] lsi . The Homfly polynomial for L.J.and link polynomials S. Yetter.2[ S[ + 2 Therefore the span of (D) is 2c + 2( 1 I + IS [ ) .Y. Their work was published as a coauthored paper (Freyd et al. is the unique twovariable Laurent polynomial defined by . Show that L has r . Freyd and D.9. • Proof of Theorem 2. we may assume that the splitting marker is in S.4 S and so by the dual state lemma.F. generalizing both the Alexander and the Jones polynomial.1. m).H. The other splitting markers in S now give a state S' of L. and opened up the possibility of advances in knot theory using the new results and techniques. • oriented or Homfly polynomial The appearance of the Jones polynomial revealed an unexpected interplay between different fields of mathematics. denoted p(L)(e.B. Again we consider the state S of D with all the splitting markers joining positive regions. the span of (D) is at most 2c + 2r . and A.4.2. There were formal similarities between the defining formulae for the Jones polynomial and that given for the Alexander polynomial by J . was found virtually simultaneously by four sets of mathematicians: P. and that no other state can contribute a term of higher degree to (D).N. A new polynomial invariant for knots and links.O.
The oriented or Homfly polynomial
35
HPl: peL) is an isotopy invariant; HP2: P(V) = 1;
HP3: if L+, L_, and Lo are three oriented links identical except at one crossing, at which we see the configurations as shown in Fig. 2.18, then fP(L+) + e1p(L_) + mP(Lo) = O.
x x
L
Fig. 2.18
HPI allows the computation of peL) from any diagram of L, whilst HP2 and HP3 give a practical procedure for carrying out such a computation, by crossing changes and the elimination of crossings in a diagram. Let us consider some examples. Observe that the orientation of L is crucial, for it tells us how to eliminate crossings in an application of HP3.
( 1) The ccomponent unlink (Fig. 2.19)
00
Fig. 2.19
·0
We denote the ccomponent unlink by V C • Since Vi = V, HPI tells us that P( U 1) = 1. Now consider V 2 • We begin with a diagram of V 2 with no crossings, and take it as our Lo (Fig. 2.20). Then we introduce a crossing to give L+ and L_.
000000
Fig. 2.20
and link polynomials
DOO DOO
L
Fig. 2.21
Each of L+, L_ is now an unknot, and so an application of HP3 I gives f + e + mP( U 2 ) = 0, and it follows at once that P( U 2 ) =  mI(f + e\ What happens for U 3? We follow the same procedure, taking as Lo a threecomponent unlink with no crossings, and introducing a crossing to form L+ and L_ (Fig. 2.21); these are twocomponent unlinks. We have P(L+) =  mI(f + e I ) = P(L_) and we want to calculate I P(Lo) HP3 gives f[ mI(f + e )] + e I [ mI(f + fI)] + mP(Lo) = 3)=P(Lo)=[mI(f+e I )]2. Write p=m I and we see that P(U (f + fI). We leave it to you as an exercise to complete an inductive proof that P( U{) = pCI.
°
CD
Fig. 2.22
(2) The lefthand Hopf link (Fig. 2.22)
QD
Fig. 2.23
P(L_)
QD
f(m + fp)
We operate on the top crossing, which has negative sign (do you agree?), so that we take the given diagram as L_ to obtain the triple shown in Fig. 2.23. We know that P(L+) = P and that P(L o) = 1. An application of HP3 gives the equation fp + e I P(L_) + m = 0, from which it follows that
=
=
fm + f3 m 1 + fmI.
The oriented or Homfly polynomial
37
Some formal properties of the Homfly polynomial follow easily from the definition. As for the Kauffman polynomial, we may easily compute the Homfly polynomials of reverses and mirror images.
position. Let L be an oriented link, with reverse L and mirror image L Then p(L)(e, m) = p(L)(e, m) = P(L!)(e 1, m).
Proof. Reversing orientation on all components of L leaves the signs of crossings unchanged, so that a computation of p(L ) is identical to one of P(L). Taking a mirror image changes the sign of every crossing, so that a computation of P(L!) is the same as one of P(L) with e and e I interchanged. •
Computerassisted calculations of Homfly polynomials, the most comprehensive of which cover all the 12965 knots with crossing number at most 13, invite speculation on two key questions about the Homfly polynomial: 1. 2. Can we characterize the twovariable Laurent polynomials that arise as Homfly polynomials of links? Is there a nontrivial link L with P(L) = I?
We have yet to establish that the computational procedure implicit in HPI, HP2, and HP3 does give a unique result. The details of this fact can be found in Lickorish and Millett (1987). Their proof is by induction on the number of crossings in a link diagram; the argument is necessarily more subtle than for the bracket polynomial, since the use of HP3 does not guarantee a reduction in the number of crossings of diagrams. The inductive hypothesis is as follows. Let :z~ be the set of all link diagrams with at most n crossings, with an ordering chosen for the components and a starting point selected on each component. We assume that for every K EXn _ I' a polynomial p(K)(e, m) has been defined, such that p(K)(e, m) = P(K') (e, m) if K and K' differ only in a choice of ordering of components and of starting points on components, that P(K) is unchanged by Reidemeister moves carried out withinZn _ I> that it satisfies HP2 and HP3, and that in addition, if U C E5{;n _ I is a projection of a ccomponent unlink, then p(UC)(e, m) =J.LCI. This is a bit of a mouthful, so as an aid to digestion it is recommended that you check that the induction starts successfully on 200 , Note that P(U C ) is then defined by the inductive hypothesis. The definition of P is now extended to L E56n by applying crossing changes to the components of D in order, beginning on each component at the specified starting point, and carried out so that the first encounter with each crossing is made along an underpass, as in
and link polynomials
the proof of Lemma 1.2. The result is an unlink aL with c components, and P(aL)(e, m) is defined as f.1CI. Now at each crossing change, the inductive hypothesis is used to compute p(Ko)(e, m) for each Ko that arises, and HP3 permits the calculation of a polynomial defined as p(L)(e, m). With this detailed inductive definition of peL) to hand, it is now a matter of showing that it satisfies the requirements of the inductive hypothesis on .Y5n . This verification is accomplished in Lickorish and Millett (1987), to which we refer the reader for further details.
2.4
1. Verify that if L is the righthand Hopf link (Fig. 2.24) then peL) =  eIm + e 3m 1 + eIm l .
m
2. 3.
Fig. 2.24
Use the triple of diagrams in Fig. 2.25 to show that the Homfly 2 2 4 polynomial of the lefthand trefoil is  2e  e •
em
Fig. 2.25
Guess the Homfly polynomial of the righthand trefoil, and check your guess with a computation.
Alexander polynomial
l.W. Alexander introduced the polynomial that now bears his name (Alexander 1928) as a simple and effective invariant for distinguishing knot types. The Alexander polynomial may now be obtained from the Kauffman or Homfly polynomials by a change of variable,
The Alexander polynomial
39
so is not as effective as these parvenus, but it retains a great deal of intrinsic interest because of its connection with knot groups, which we shall see in Chapter 8, and with covering spaces in Chapter 12. Here we discuss Alexander's original treatment of his polynomial, which is based on an ingenious calculation involving determinants, and uses a labelling of knot diagrams that is a forerunner of the idea of state used by Kauffman. Let D be an oriented link diagram with c crossings. Then by Lemma 2.6 D divides the plane into c + 2 regions. We construct a matrix M(D) whose rows are indexed by the crossings of D and whose columns are indexed by the regions, so that M(D) has size
eX (c
+ 2).
Fig. 2.26
Suppose that regions P, Q, R, and S meet at the ith crossing of D (Fig. 2.26). Then in row i of M(D) we enter t, t, 1, 1 in columns P, Q, R, and S and zero in all other columns. Note that according to our sign convention for regions, Q and S are positive. At an isthmus, with only three regions adjacent to a crossing, the column indexing the repeated region receives the sum of the symbols allocated to that region (Fig. 2.27).
Fig. 2.27
For example, the figureeight knot with crossings and regions labelled as shown in Fig. 2.28 has the associated matrix
1 .t t T U t 0 . in the above example. It is clear that AD is a polynomial in t with integer coefficients. call the resulting (c x c) matrix Mo{D). we find .t J 0 . 2.28 Now delete from M(D) any two columns indexed by a region which have a common boundary arc in D.1 0 0 S t I 0 I t .1 . we delete columns T and U to obtain I Mo{D) =2 3 4 p 0 0 t Q 0 t R . If. t 0 Q Fig.t 2 ) = t3  3t 2 + t.1 and expand the determinant by the first row.t + 2t 2 )  t(t . The Alexander polynomial AD of D is defined to be the determinant of Mo{D): AD = det{Mo{D)).1 .and link polynomials M{D) =2 3 r 4 0 t t p Q 0 t t R S T U .I .I  o tJ .1 D {t) = 1(. If instead we delete columns Rand S to obtain I Mo{D) =2 3 4 p 0 0 t Q 0 .1 ..I 0 0 1 t 0 1 I .
t 2 . det(Mo(D» is then an isotopy invariant. • We shall not prove the theorem here. This is a typical outcome. namely the Alexander polynomial of L. You will observe that this is a different answer. the determinant det(Mo(D» is independent of the choice of columns deleted from the matrix M(D). it differs from the first answer by a factor of t. Given an oriented link diagram D. as explained by a theorem proved by Alexander. we now have an isotopy invariant for a link L computable from any oriented diagram. but recommend the reader to enjoy the discursive style of Alexander (1928). 2. The convention is to adjust det(Mo(D» by a term ± t m to give a positive constant term and a positive highest power of t. Choose your favourite knot or link from the book so far. orem.5 1. . up to a factor of ± t m for some integer m. denoted LlL(t).t 2 + t 3) = t4  3t 3 + t 2. we find LlD(t) = t2  t 3 + t(. and compute its Alexander polynomial. Allowing for adjustments of sign and powers of t.The Alexander polynomial 41 and again expand by the first row. Furthermore. However.
This topology is the one appropriate for the analytic viewpoint. the definition of continuity in terms of e and {j familiar from introductory analysis is now seen to arise from a particular topology chosen for the real numbers. Since our interest in topological spaces stems from specific applications that we already have in mind. In the treatment of surfaces in Chapter 4. against which the vaguely expressed ideas lying behind our previous discussion can be made precise. even if some of the examples do not accord with our intuitive expectations. topological ideas will prove crucial in the development of more sophisticated invariants in later chapters. there is a great gain in the clarity of the concept. and we call it the usual topology on ~n to emphasize its significance. and as is usually the case with abstraction. essential to any proper treatment of knots and surfaces. The precision that topological spaces bring to the ideas of continuity and nearness is. Moreover. There are several textbooks that do the job well. we shall have . ological spaces Topological spaces provide an abstract framework for the notion of continuity. However.3 Topological spaces The purpose of this chapter is to provide the theoretical background for a rigorous discussion of knots and surfaces. we make no attempt to give a comprehensive account of even the basic theory of topological spaces here. and for links we shall need to define a continuous map from the union of several copies of Sl into ~3. What are the concepts that we need to make precise in order to put our study of knots and surfaces on a firm base? A glance back at our informal definition of a knot in Chapter I shows immediately that we need to define continuous maps from the circle Sl to ~3 in order to describe a knot as a continuous copy of S. and we mention Armstrong (1983) and Brown (1988) as providing suitable further reading. of course. The central notion is that of a topological space.
This is the finite complement topology on X. 2. What do we mean by 'looking like' and 'near'? In IR. ':!T).Topological spaces 43 need of a notion of a subset of IR n 'looking like' 1R2 'near' any point.y. [Rn. We shall usually denote a topology by a script symbol such as ':!T. there will be good reasons to select one topology above all others. satisfying OSI: the empty set and X itself are open sets. open interval around the given point. is the subset {y E IRk I IIx . here r E IR and r > O. What might 'open' mean in more generality? With these questions in mind. This is the discrete topology on X. we may define every subset of X to be open. 'near' often means 'within EO of'. OS3: the intersection of finitely many open sets is an open set. we see that an open ball in IR is an open interval: . other than that we have selected them to playa certain distinguished role. there is nothing intrinsically special about the open sets. However.yll ~~ [ 112 (x. so that a space is a pair (X. let us proceed to the definition of a topological space without more ado. Again on any set X. OS2: an arbitrary union of open sets is an open set. Y IR n. A = E 3. in many circumstances. If k = 1.yll < r}. . What do we get if X is a finite set? The usual topology on. For x. called open sets. Let X be a nonempty set. is a topological space. The open ball with centre x and radius r. (n ~ 1). that is. Given any set X. keeping the topology in the back of our minds. The set X. However. define IIx . that is I X\A I < 00. we fit a small. r). A topology on X is a collection of subsets of X.)' J This is the customary notion of the distance between two points in IRn. it is customary to suppress explicit mention of ':!T and refer to X as a topological space. define a subset A ~ X to be open if either 0 or A has finite complement in X. Note that a given set can have many different topologies. together with the collection of its special subsets that we have called open. denoted by B(x. 5 1.
ropological spaces
r r
x
whilst if k = 2, an open ball in ~2 is an open disc. The usual topology on ~n is now defined in terms of open balls in the following fashion. We say that a subset U ~ ~n is open if, for every x E U, there exists r E ~ with r> 0 such that B(x; r) ~ U. Note that r depends on x. Let X be a topological space and let A ~ X be a subset. We say that A is closed if its complement X\A is open. If X is a topological space, any subset A ~ X can be made into a topological space by defining the open sets in A to be all sets A n U where U is an open set in X. This is the subspace topology on A and we say that A is a subspace of X. The following result will be of use to prove the important gluing lemma later in this chapter.
emma. If A is a closed subspace of a topological space X, and closed in A, then C is closed in X.
if C
~
A is
Proof By definition, A\C is open in A, whence A\C = A n U for some open set U ~ X. Therefore,
X\C = (X\A) u (A\C) =(X\A) u (A n U) = (X\A) u U.
Now X\A and U are open in X and so by OS2 X\C is open in x. Therefore C is closed in X. •
ises 3.1
1. Let X have the discrete topology. Which subsets of X are closed? What if X has the finite complement topology? In ~n with its usual topology, show that the closed ball with centre x and radius r, defined as B(x; r) = {y E ~n I IIx  yll ::::; r} is indeed a closed subset of ~n. Show that the subspace topology is a topology on a subset A of a topological space X. Show that any subspace of a topological space with the discrete topology also has the discrete topology. Prove that, as a subspace of ~ with its usual topology, 7L has the discrete topology. Show that the union of finitely many closed subsets of a topological space is again a closed subset. (a) Let X be a topological space. We call a subset N ~ X a neighbourhood of x E X if there exists an open set U ~ X such that x E U ~ N. Let N(x) denote the collection of all neighbourhoods of x E x. Show that N(x) is nonempty and satisfies the following four properties:
2.
3.
4.
5.
6.
Continuity
45
NHDl: if N NHD2: if N
E
N(x) then x N(x) and AI
E
E
N;
E
~
N then M
E
E
N(x);
NHD3: if N, N' NHD4: if N
E
N(x) then N n N'
E
N(x);
E
N(x) and N° = {y
NIN
N(y)} then ~
E
N(x).
(b) Suppose now that we are given a set Y and for each y E Y a nonempty collection N(y) of subsets of Y satisfying NHDl, ... , NHD4. Define U ~ Y to be open if, for every u E U, U E N(u). Show that this is a topology on Y. (c) Part (b) opens up the embarrassing possibility of going round in everdecreasing circles, repeatedly defining 'open sets' and 'neighbourhoods' in terms of one another. Suppose we start with a topological space X, with neighbourhoods as in (a), and define open sets as in (b). Are these new open sets the same as the open sets of X that give the topology? What if we start with a set Y as in (b), define a topology on it as in (b), and then define neighbourhoods as in (a)? Are these neighbourhoods the same as those in the given collections N(y) for y E Y? 7. Find all the possible topologies on a set with precisely three elements. How many different topologies are there for a set with precisely four elements?
ntinuity
Let X and Y be topological spaces. (Note that we suppress any mention of the topology on the underlying sets, but you should keep in mind that the topologies may be very different.) A function f:X + Y is continuous if, for every open set V in Y, its preimage f!(V) is open in X. (Recall that the preimagef1(V) is the subset {x E X If(x) E V}; we do not assume that an inverse f ' to f exists.) Since continuity is defined in terms of open sets, it depends on the choice of topology for each of the sets X and Y. For example, if 2J and 2J' are different topologies on X then the identity map x H x need not be a continuous function (X, 2J) + (X, 2J'). However, the familiar notion of 'e  8' continuity is identical to the notion of continuity with respect to the usual topologies on [Rm and [Rn.
position. Let [Rm and [Rn have their usual topologies. Then f: Rnl + [Rn is continuous if and only if, given any e> 0, there exists 0> 0 such that IIx  yll < 0 implies that IIf(x)  f(y)1I < e.
ological spaces Proof. Suppose thatfis continuous and let e> 0 be given. The open ball B(f(x); e) is an open set in [Rn and therefore f I(B(f(x); e» is an open set in [Rm: since x EfI(B(f(x);e» there exists 8> 0 such that B(x; 8) r;;;;.fI(B(f(x); e», that is IIx  YII < 8 implies that IIf(x)  f(y)1I < e. The proof of the converse we leave as an exercise for the reader. •
mao Let X, Y, and Z be topological spaces. If f: X ~ Y and g: Y ~ Z are continuous functions, then gf: X ~ Z is continuous.
Proof. Let W be an open set in Z. By the continuity of g, we know that V= gI(W) is open in Y, and by the continuity offwe know that fI(V) is open in X. Thus (gi)I(W) =fI(gI(W» is open in X and gf is continuous. •
gluing lemma. Given a function f : X ~ Y, suppose that there are closed sets A and B such that X = A u B and such that the restrictions f IA: A ~ Y and fl B: B ~ Yare continuous. Then f is continuous.
Proof. Let V be an open set in Y. We want to show that f I (V) is open in X. Continuity of the restricted functions implies that f\V) n A is open in A and that fI(V) n B is open in B; thus A\fI(V) is closed in A and B\fI(V) is closed in B. By Lemma 3.1, each of A\fI(V) and B\f 1 (V) is closed in X. Using Exercise 3.1.5 we see that X\fI(V) = (A\f 1 (V» u (B\fl(V» is closed in X and so fI( V) is open in X. • Let X and Y be topological spaces. A function f: X ~ Y which is bijective, continuous, and has a continuous inverse f I : Y ~ X is called a homeomorphism. A function f: X ~ Y which is injective, continuous, and such that the bijection f: X ~ f(X) has a continuous inverse is called an embedding. Simple examples of embeddings include [R2 ~ [R3 given by (x, y) H (x, y, 0) and Sl ~ [R3 given by (x, y) H (x, y, I) where x 2 + y = I. If X and Yare topological spaces, is there a natural topology on the set X x Y = {(x, y) I x E X, Y E Y}? The first question to answer is this: what do we mean by 'natura!'? Clearly the topology will have to be defined in terms of the given topologies for X and Y. Furthermore, there are the canonical projection functions p: X x Y ~ X given by (x, y) H x and q: X x Y ~ Y given by (x, y) H Y and these should surely be continuous. So let us try to construct a topology meeting these requirements. Let U r;;;;. X be open: then pI (U) is U x Y so that U x Y needs to be one of our open sets in X x Y; similarly we need X x V for any open set V in Y. However, (U x Y) n (X x V) =
Continuity
47
u x V and so if we are to satisfy OS3 we require that all subsets U x V, where U is open in X and Y is open in Y, are open in X x Y. Then to satisfy OS2 we require arbitrary unions of such subsets to be open, and this is enough. If we define A ~ X x Y to be open if A is a union of subsets U x V, where U is open in X and V is open in Y, then we do obtain a topology on X x Y: it is called the
product topology on X x Y.
3.2
I. Let X
f: X
~
be a
topological space.
Prove that every function
Y, where Y is any topological space, is continuous if and
only if X has the discrete topology. Let (X, fJ) and (X, fJ') be topological spaces. Find a condition on the topologies fJ and fJ' that ensures that the identity map x H x is a continuous function (X, fJ) ~ (X, fJ'). 3. Given sets X and Y, their disjoint union X I I Y is the set obtained by taking the union of disjoint copies of X and Y, so that in X I I Y we have two copies of any element in X n Y. If X n Y = 0 then X I I Y is the usual union Xu Y. How might we define a topology on the disjoint union X I I Y of topological spaces X and Y? The simplest idea is to take a subset U ~ X I I Y to be open if U = V I I W where V is open in X and W is open in Y. Show that this does define a topology, and makes the inclusions of X and Y into their disjoint union continuous functions. 4. Find an example of a continuous bijection whose inverse is not continuous. [Hint. You have already seen examples if you have done the exercises so far.] 5. Let X be a topological space and let U ~ X be a subspace. Show that the inclusion U ~ X is an embedding. If X is a topological space and U is a subset of X with a topology such that the inclusion is an embedding, need the topology on U be the subspace topology? 6. Verify that the product topology is indeed a topology on X x Y. 7. Describe the neighbourhoods of a point (x, y) E X X Y in the product topology. 8. A topological space X is called Hausdorff if, given any distinct elements x, y E X, there exist open sets U containing x and V containing y with U and V disjoint; that is, we have x E U and y E V with U n V = 0. (The concept is named in honour of F. Hausdorff (1 8681942) and a punning aidembnoire is to say that the elements x and yare housed off by the opens sets U and V). The diagonal subspace of X x X is L1(X) = {(x, x) Ix EX}. Show that X is Hausdorff if and only if the diagonal subspace is closed 2.
Topological spaces
9.
in X x X. Suppose that the finite complement topology on a set X makes X a Hausdorff space. What must be true of X? Find some nonHausdorff topologies. Let U ~ Xbe a continuous injection between topological spaces. Show that if X is Hausdorff then so is U.
Connectedness
In addition to providing a framework for continuity, topological notions also allow us to consider connectedness. Intuitively we want to make precise the idea of a topological space being in one piece, and we do this by considering what we might be able to do with a space that was not in one piece. We surely expect to be able to divide such a space into at least two subspaces that do not intersect, and this leads to the following definition. A disconnection of a topological space X is a pair (A,B) of nonempty closed subspaces of X such that X = A u Band A n B = 0. A space X is connected if no disconnection of it exists. A key property of connected spaces is that the only subsets which are both open and closed are the empty set and the whole space. We shall make use of this fact in Chapter 12. We shall show here that this scarcity of subsets that are both open and closed is equivalent to connectedness, so that we could have chosen it as the definition. Would this have been a sensible thing to do?
Proposition. A topological space X is connected if and only if the only subsets of X that are both open and closed are the empty set and X itself.
Proof. We first suppose that X is connected, and let A be a subset of X that is both open and closed. Let B be the complement of A in X, so that B is also both open and closed. If both A and Bare nonempty, we see that (A, B) is a disconnection of X; since X is connected, no disconnection exists. Therefore one of A and B is empty, and the other is then X itself. Conversely, suppose that the only subsets of X that are open and closed are 0 and X but that X has a disconnection (A, B). Then A is closed, so that its complement B is open; but B is also closed and since B is nonempty we must have B = X. But then A = 0, which is • a contradiction.
Given a continuous functionf; X ~ Ywith X connected, might we expect Y to be connected? A little thought should convince you that there is no reason to expect such a result for Y itself, but that it does
define a relation . but that the subspace (() is not connected.Connectedness 49 seem reasonable to suppose that the image of f will be connected. • rcises 3. Show that any space with the discrete topology is totally disconnected. It is better to have as rich a structure of open sets as we can on Y subject to the condition that f be continuous. Therefore f(X) is connected. Evidently this is something of a swindle. and it follows that A is either f(X) or the empty set. a set Y. Show that ~ (with the usual topology) is connected. the identification topology on Yis obtained by defining U ~ Yto be open if and only ifits inverse imagefI(U) is open in X. Show that . because the topology has nothing to do with the function f and has made Y into a dull sort of space that accepts every function into it as continuous. It is immediately obvious that this definition makes f continuous. Preservation of connectedness Let X be a connected space and let f: X Then the image f(X) off is connected.is an equivalence relation on X. 2. Suppose that A is a subset of f(X) which is both closed and open.5. and this idea motivates the following definition. and a function f: X ~ Y. y E X we say x . Since f is continuous.on X as follows. Perhaps we should reformulate the problem. ~ Y be a continuous function.y if and only if x and y belong to some connected subspace of X. The equivalence classes are called the connected components of X. and let f: X ~ Y be a function to a set Y. By connectedness of X. Given a topological space X. f. For any x.3 1. . A space in which every connected component contains just one element is said to be totally disconnected.I (A) is either X or the empty set. Is the converse true? Identification spaces Let X be a topological space. f1(A) is both closed and open. Can we equip Y with a topology that makes f continuous? Of course we can! We can give Y the indiscrete topology for which the only open sets are Y itself and the empty set. by Proposition 3. Proof. This reasonable supposition turns out to be true. Given any topological space X.
(x. Now bend the cylinder so as to bring together the boundary circles of the two open ends. Formally we might begin with a rectangle in [R!2. say R = {(x. we take R as in (a) and impose the equivalence relation generated by equality and the further relations (0. y) and (x. where [R!2 has its usual topology and R has the subspace topology.to be defined in terms of the given topology on X. Further. there is a function that we shall require to be continuous: namely.:: y :s. The result is a Mobius band (or strip). but if you try and envisage what happens to a rectangle (b) (c) .Topological spaces An important application of the identification topology is to the quotient set of an equivalence relation.on R generated by equality and the additional relations (0. Begin again with a rectangle. and suppose that .:: 1.the identification topology defined by the quotient function. The result is an openended cylinder. y) . 0 :s. Let X be a topological space. in picturesque terminology. We achieve all that we require by giving X/. the quotient function X ~ X/.:: y :s. 1) and (0.y). I). and glue these circles together. 0) .is an equivalence relation on X.:: x :s. y) . and identify two opposite ends after performing a halftwist. 1 . What happens if we try to combine (a) and (b) by introducing a twist into our construction of the torus? We impose on R the equivalence relation generated by equality and the further relations (x.(2. y) . and take the equivalence relation .(x. it is tempting to recommend the acquisition of paper and a pot of glue before reading further! mples (a) Take a rectangle. Then R/. The result is a hollow.:: 2. 0) .with the identification topology is a mathematical model for a Mobius band. we require the topology on X/.(2. namely the quotient set of the equivalence relation equipped with the identification topology. we think of gluing one part of a space to another part. In particular. 1 . Is there a natural topology on the set of equivalence classes X/? As in our discussion of the product topology. we can identify certain elements of a space with one another by an equivalence relation.:: I}. y) E [R!21 o:s. We can now take any equivalence relation on a topological space X and construct a new space. and glue together corresponding points of two opposite sides.that takes an element x E X to its class [xJ E X/. Here are some examples of this process in action. This appears to be harmless enough.y) for all y satisfying 0 :s. More formally. doughnutshaped surface called a torus.(2.
and as such will reappear in Chapter 4.Knots and isotopy 51 Fig. 3. Now impose the equivalence relation generated by equality and the further relations (x. and then the disjoint union (X x J). a difficulty at once emerges.1). y) I x 2 + = I} ~ ~2 with the subspace topology. Form the space X x J with the product topology. but no model of it can be made in three dimensions (nor any picture drawn in two dimensions!) without such selfintersections. let (Y. 1) ~ f(x). see (Brown 1988. ~n will always have the usual topology. The Klein bottle is an example of a nonorientable surface. Show that the identity map (Y. (d) Let f: X ~ Y be a continuous function between topological spaces. f) ~ Y is continuous. and we denote by Sl the unit circle {(x. Given topological spaces X and Y and a continuous function f: X ~ Y. The resulting surface is known as the Klein bottle.1 under the corresponding bending and gluing. f) denote the set Y given the identification topology. and let J denote the unit interval [0. Find an embedding of the torus into ~3. s and isotopy Hereafter.u Y with the topology given in Exercise 3 of Section 3. Figure 3.1 does not truly represent the Klein bottle because of the selfintersection. 1] with the subspace topology in ~. y . 3. The nicest picture is obtained by imagining the cylinder formed by one gluing to pass through itself before the boundary circles of the cylinder are identified (Fig. 3. 2.2. chapter 7).4 1. The resulting identification space is called the mapping cylinder of f and it is a useful construction in the further theory of topological spaces.
k is injective. so that the definition provides a continuous sequence of homeomorphisms of ~3 from time t = 0 to time t = I. just as we might deform a physical piece of string from one interlaced form into an equivalent one. Firstly. k is continuous and there is a continuous function back from its image to SI. Ambientisotopy of knots and links is precisely the topological analogue of the isotopy of knots that we considered in Chapter I. but we shall not give a proof. If k and 1 are ambientisotopic. this corresponds to our informal insistence in Chapter I that the string of a knot does not intersect itself. The theorem is due to Reidemeister. and in ~3 we might think of moving their images around in a continuous fashion until one image coincides with the other. When should we consider two knots k and 1 to be the same? We want to capture the notion of their knottedness or interlacing being the same. The topology on SI 11 Sill . 0) is the identity ~3 ~ ~3 for all t E [0. hI = H(. which we leave the reader to formulate. 1]) form a continuously varying family deforming k into I. 11SI ~ ~3 of c disjoint copies of SI into ~3. The idea that we move the images in ~3 continuously leads to the concept of ambientisotopy in which deformations of the images are effected by applying homeomorphisms of the ambient space ~3. . Here it is: a knot is an embedding k: SI ~ ~3. Secondly. and in the following definition it may be helpful to think of the parameter t as representing time. . copy of the unit circle. then hJk = I.logical spaces We now have the necessary theoretical background to give our precise definition of a knot. Our definition thus defines a knot as a function k from SI that has certain nice properties. A link of c components is an embedding SI 11 SI 11 . This takes care of our requirement that the string forms an unbroken. . 11 SI is of course that for a disjoint union given in Exercise 3 of Section 3. who has given a proof . albeit entangled. t) is a homeomorphism ~3 ~ ~3 if hi = H(. Now to the definition: two knots k and 1 are ambientisotopic if there exists a continuous function H: ~3 x [0. .2. I J ~ ~3 such that (a) (b) (c) ho = H(. There is a corresponding definition of ambientisotopy for links. I]. then the knots htk (t E [0. We state this crucial fact as a theorem below. I).
ises 3. A piecewise linear knot can be represented by a diagram whose arcs are straight lines. 2. by the wellknown process of pulling the knot tight. a third possible equivalence relation for knots. and demolish.Knots and isotopy 53 in his book (Reidemeister 1932). if and only if they possess isotopic • There are other equivalence relations between functions S ~ ~3 that might be used to define a notion of two knots being essentially the same. 1) = I. Tameness may be considered as a technical restriction on knots to ensure that each has only finitely many crossings. We can consider a homeomorphism ~3 ~ ~3 as analogous to a change of coordinates. 0) = k hi = H(. and the theory developed in this book will treat exclusively tame knots. 3. 4. heorem.5 Why is it obvious that the righthand and lefthand trefoils are equivalent? The final three exercises in this chapter propose. An equivalence class of knots under the equivalence relation of equivalence is commonly called a knot type. and it should come as no surprise that equivalence is a much weaker notion than ambientisotopy. the theory of wild knots encounters different problems to that of tame knots and we propose to say no more about it. Two knots are ambientisotopic diagrams. ho = H(.] . Prove that equivalent knots are comparable. it is obvious that the righthand and lefthand trefoils are equivalent. 1] ~ ~3 such that (a) (b) 1. Prove that any two knots are comparable [Hint. A knot is said to be tame if it is equivalent to a piecewise linear knot. A knot k is piecewise linear if its image in ~3 is a union of a finite number of line segments. Call two knots k: SI ~ 1R3 and I: SI ~ ~3 comparable if there exists a continuous function H: S x [0. and so defme two knots k: Sl ~ ~3 and I : Sl ~ ~3 to be equivalent if there exists a homeomorphism h: ~3 ~ ~3 such that 1= hk. A more conceptual account is to be found in Burde and Zieschang (1985). Show that any given knot is comparable with the unknot. It is immediate that ambientisotopic knots are then equivalent. For example. Show that comparability is an equivalence relation on knots. All other knots are wild.
4 Surfaces Knots are onedimensional objects living in threedimensional space. perhaps most notably in complex analysis. 4. The study of surfaces. however.1 Torus Fig. The flow of ideas is two way and we shall see later on how surfaces help us to study knots.2 . 4. Some of the ideas that help in understanding knots can also be adapted to help study surfaces. There are important applications of the classification theorem for surfaces in other areas of mathematics. has no need of knot theory for motivation. Surfaces are twodimensional objects. Sphere Fig. but we will not be going in that direction here.
3. Here we have seen some of the simpler examples of surfaces.9). 4. Fig.7. 4. 4. as in Figs 4. we saw some surfaces constructed by gluing.14.Surfaces 55 What is a surface? In the last chapter.44. 4.4 Sphere with one hole Fig.5 Recall from Chapter 3 that the Mobius band (Fig.3 Here we also want to allow for holes or edges in surfaces. 4. We have examples such as those shown in Figs 4.8) is obtained by twisting and gluing a rectangle (Fig. Can we list all of them? This question needs tightening up before it really . Double torus or pretzel Fig.
as in Fig. There are infinitely many surfaces. 4. Likewise a sphere with n holes is not likely to be the same as a sphere with m holes (unless m = n.faces Fig. 4. for instance.6 Sphere with two holes Fig. 4. 4. consider a multiple torus with n 'handles'.4. it seems likely that the corresponding multiple tori will be different.8 will make sense.10.9 I . of j Fia.7 Fig. For different values of n.
l: Two of the triangles in S can meet as in Fig. 4. 4. Then.16.14. A combinatorial surface. or they may not meet at all. 4. or as in Fig. .13.Combinatorial surfaces 57 Fig. given any surface. S.12. we have used 'same as' several times without worrying about what that means. as in Fig. or as in Fig.10 course!).11.2: Any edge is an edge of one or two triangles. S.3: For any vertex v. we will be more precise and say that we want to classify surfaces 'up to homeomorphism' and so we may tacitly use phrases such as 'SI is the same as S2' as a more relaxed way of saying 'There is a homeomorphism between SI and S2'.4: Given any two vertices in S. To get that point out of the way. 4. one vertex. We now give some comments and further explanation of the rules. or as in Fig.15. This basic idea in fact works quite well but of course our view of it here is a bit primitive. 4. then St v is homeomorphic to a disc (St v is called the star of v). S. or not at all. For instance. mbinatorial surfaces We want to use geometric and combinatorial arguments wherever possible. To do this we are first of all going to consider combinatorial surfaces. Thus thinking of the problem a bit more we might try to classify surfaces according to certain invariants such as the number of holes or the number of tori in the above examples. is a finite union of (solid) triangles satisfying the following rules: S. S. we write St v for the union of all triangles having v as a vertex. S. we would hope to be able to calculate its invariants and assign it some standard name such as an ntorus with m holes in it.l: Any two triangles meet in one edge. then there is a path along edges of triangles joining them. 4. Two triangles cannot meet as in Fig. 4.
This happe on the boundary of some hole (Fig.14 . e.12 Fig. 4.11 Fig.58 Surfaces Fig. 4. 4.2: An edge.13 S.17). 4. might only be in one triangle. 4. Fig.
If the vertex is in a boundary edge. as in Fig. 4.20. v. .15 Fig. will normally look as shown in Fig. 4.Combinatorial surfaces 59 Fig. 4.3: The star of a vertex. 4. S. 4.18. its star may be different.16 Fig. but it cannot be as shown in Fig.19. 4.17 Many edges will be in two triangles. but none may be in three or more.
there is a path a in X such that a(O) = x.19 Fig. We say a joins the points a(O) and a(l) and th~t X is path connected if any two points can be joined by a path. 4. 4.20 A path in a topological space X is a continuous map a: [0. 1] ~ X. that is. given x. a(l) = y. We will be looking .faces Fig. y E X. 4.18 Fig.
As we saw in Chapter 3. we can imagine performing these identifications within 1R(3 (in this case) and the result is a torus (Fig.21 at paths much more closely in Chapter 9. • f t + () Fig.4 implies that S is path connected.Combinatorial surfaces 61 a b c d d e e f f a b c Fig. 4.23 . 4. 4. 4. for example as in Fig.22 b ) + 6) d e f c e e d f a b c Fig.21. Forgetting for the moment the edges inside the rectangle. we merely note that S. 4.22). we can indicate the construction of surfaces by diagrams giving gluing instructions. In such a diagram the edges with the same label are to be identified with the arrows matching. For the moment.
4. let us try to triangulate the cylinder (represented as a rectangle with identification in Fig. } Fig..24). 4.23 still yields a surface. To see the restrictions that the rules place on the way the triangles fit together.24 The reason why we had to put 'in this case' is that a different diagram such as in Fig. 4.aces t + ('L. . 4.25. 4. Fig..l as the two triangles meet in two edges.27) This simple decomposition contravenes S. so it does not give a triangulation.25 We will see some of the strange properties of the Klein bottle later. the Klein bottle (Fig. 4.26). but the identification cannot be done in /R 3 • To obtain some idea of what this space looks like it is usual to draw it with a selfintersection as in Fig.26 Attempt 1 (Fig. 4.) _~) Gf) How can we glue this? Fig. 4.
so again this is disallowed. 4.27 Attempt 2 (Fig.l. 4. 4.29 Attempt 4 (Fig. 4. 4. 4.30 .28) Triangles I and 2 meet in three vertices. 4.28 Attempt 3 (Fig. so this also contravenes S. Fig.29) Triangles I and 3 meet in two vertices. Fig.Combinatorial surfaces 63 Fig.30) Again the intersection of two of the triangles (2 and 4 this time) consists of just two vertices. Fig.
and 8 and so similarly is a rectangle and hence homeomorphic to a disc.aces Attempt 5 (Fig.31) Fig. 7. Path connectedness is clear.31 S. These strange properties are due to the nonorientability of these surfaces. and so we can and will change the triangulation as needed. perhaps chopping up some of the triangles into smaller triangles (subdivision). not primarily in the triangulation. This leaves six other vertices for you to check. We do not claim to have found the smallest number of triangles.l is now satisfied as is S. It was not as obvious as it may first have seemed. To do this we will assume that we have triangulated our surface. 4. Our search will be for invariants of the surface.3. but we are interested in the surface. the sphere. torus. We will also see certain moves on combinatorial surfaces which change the triangulation. 4. It has a strange singlesidedness property. We leave you to check this in detail but it is clear that St VI is the union of triangles 4 and 5. and multiple tori are orientable whilst the Mobius band and the Klein bottle are nonorientable. Can you improve on our number of eight? What is the smallest number of triangles needed to get a triangulation of the cylinder? How might you attack this question? Our main aim will be to classify surfaces up to homeomorphism. . so is a triangle and thus homeomorphic to a disc. the condition on stars of vertices. so we are left to check S. This is much like the situation we saw earlier where moves on knot diagrams allowed us to check if certain quantities were invariants. The same is true of the Klein bottle. We met the Mobius band in Chapter 3. The first invariant property we will look at is orientability. whilst St V 2 consists of triangles I. 2. You can draw a line from any point on the surface to any other without going over an edge. So we managed to find a triangulation of the cylinder. that is invariants that do not depend on which triangulation we are using. There are two basic classes of surfaces: those which are orientable and those which are nonorientable. For instance.2.
y.Combinatorial surfaces 65 Another important nonorientable surface is the projective plane. y.z) say (where x 2 + y + Z2 = 1). z') if there is some nonzero A E !R such that AX = x'. 4. y. z) and (. y.y. Alternatively since each line through the origin in !R 3 cuts the unit sphere. y'. AY = y. z] E U. z) and (x'. Thus 1]J>2 can be obtained from a disc by identifying according to the rule shown in Fig.x.y.32 . The set 1]J>2 is the set of lines through 0 in !R 3 • This set can be thought of in several other ways. then f1(U) contains both (x. y.32. Each equivalence class [x.x. z] for the corresponding point of 1]J>2. . . If (x. y. z] with z '# 0 meets D~ in one point. If f1(U) is open in 52. y. . y'. 0] meet D~ in two diametrically opposite points. y. z) = [x. .z). This gives us a clue as to how to give 1]J>2 a topology. z) ~ (x'. 52. we can consider an equivalence relation on 52 generated by (x. A final way of looking at 1]J>2 is to note that the set Di = {(x. then we will say it is open in 1]J>2. in two points. z) I (x. Z ~ O} is a disc. It is a good exercise to check that this does give a topology on 1fJl2. The set of equivalence classes is 1]J>2. This surface cannot be embedded in !R 3 • a a Fia. y. z) and (.4. We can consider the set of nonzero points in !R 3 and put on them an equivalence relation defined by (x. z) is in 52.z). . z') are on the line through O. z) E 52. Again the set of equivalence classes form 1]J>2. (x. y. If U is a set in 1]J>2 and [x. z]. Az = z'. but all the [x. y.y. This first arose in projective geometry.x. write [x. that is (x. y. z) ~ (. This gives a function f: 52 ~ 1]J>2 defined by f(x. . y. y.
4.aces We must now examine what orientability means.33 and 4. as shown in Figs. whilst if the induced directions are the same then the triangles are oppositely oriented. 4.33 Fig. how are we to find out if it is orientable or not? We clearly cannot try all the possible combinations of orientations on the different triangles. Otherwise the surface is nonorientable.34.35 Two triangles are adjacent if they have a common edge. as in Fig. 4. As our surfaces are assumed to be triangulated we will first consider what orientation of a triangle means. . 4. The intuitive answer must be that there are two choices of orientation.37 we see that adjacent triangles with similar orientations induce different orientations on their common edge. A surface S is orientable if one can choose orientations for all triangles in S such that adjacent triangles are similarly oriented.34 An orientation on a triangle induces one on its edges. (Thinking of the triangle as being full of a swirling liquid may help!) Fig. Suppose someone presents us with a combinatorial surface. Looking at the diagrams in Figs 4. 4. Fig.36 and 4.35.
As a simple example consider the triangulation of the Mobius band in Fig.39. 4. Go on to any new triangles adjacent to those already oriented and orient them similarly if possible and so on.Combinatorial surfaces 67 Fig. These cases are mutually exclusive. Pick any triangle in S and orient it in either sense. Either eventually you will have given an orientation to all the triangles and end up with an orientation of S or at some stage you end up with a situation like that in Fig. 4.36 Fig. so the surface must be nonorientable.38.38 both those already chosen. Now go to each of the adjacent triangles and orient them similarly (if you can).37 The way to do it is to propagate an orientation starting at some point. There cannot exist an orientation of the new triangle similar to Fig. 4. 4. 4. .
The idea is simple. Step 1 Label and orient each edge of all triangles in S with a letter and an arrow. We let Fig. Label both sides of each edge and then cut along the edge so as to get a collection of triangles. 4. 4. we propagate. The fact that the edges are labelled means that we can reconstruct the surface by gluing along edges having the same label. The usefulness of this is that by choosing a neat way of regluing step by step we can obtain a convenient representation of the surface as a polygon with identified edges. ting and pasting The method we will use for analysing surfaces further is called cutting and pasting. (Why is 2 oriented in a clockwise way?) It is impossible to choose an orientation for 3 which is similar to that on adjacent triangles. Distinct edges must be labelled with distinct letters.faces Fig. We will return later on to look in more detail at orientations.39 Starting with an anticlockwise orientation of 1.40 . To be slightly more formal about this process we will give a stepbystep description of it.
4.49. The process can be illustrated as shown in Figs 4.424. 4. The final result is a polygon with labelled oriented edges. .40 gives Fig.43 As we always glue along a proper connected segment of the boundary we can never get a hole.Cutting and pasting 69 x and E T(S) Fig. 4. PI' Find 'Xi E T(S)\{ 10. Clearly we can reconstruct S from T(S). matching the common edge by letter and arrow to form a disc. Fig. Fig. Step 2 Pick any triangle To E T(S). For example.) A typical polygon one might obtain would be something like that in Fig.41. Find a TI E T(S) which has an edge label in common with 10. Stick 10 and TI together.48. 4. (The proof that this always gives a disc is too technical to be included here.41 T(S) be the set of all labelled triangles in S. 4.42 x Fig. 4. T I } with one or two edge labels in common with adjacent edge labels in the boundary of PI' Stick 'Xi to PI to get a disc P2' Continue in this way matching 'Zi E T(S) to one or two adjacent label1ed edges of Pi to form a disc Pi+ I until T(S) is exhausted.
4. 4. 4.70 Surfaces e Fig.46 Fig. 4.44 Fig.45 e Fig.47 .
b. we get b. never more.I cda.Cutting and pasting 71 Fig.48 Labels can occur once or twice but never more since each label corresponds to an edge of one or two triangles. reading off clockwise. Thus in our diagram. of the boundary edges of the polygon and the order and direction in which they go. However. The choice of start vertex is ours to make so we might have got (by starting at 2) and we must therefore consider these two symbols to be 'the same'. Fig. write down the edge label if the arrow is pointing clockwise.I dCI ea. 4. or write down that label to the power minus one if the arrow is pointing anticlockwise. The information stored in such a diagram tells us how to identify edges and hence to reconstruct S. pick any vertex (we will use I in the diagram in Fig. 4.49). etc. that is equivalent.49 To work out the surface symbol of a polygon. 4.. this is encoded in the surface symbol of S. then. the key information is contained in the letters a. We will play with these symbols much as we would .
Operation 2 (Fig. 4.50 The usefulness of this operation is that one can cut along one edge and paste back along another one.51 .51) AXX1B~ AB cut paste Fig. 4. We will see this in action shortly. of course. 4. We must. Step 3: operations (We will write capital letters for words in small letters and their inverses. provided of course that labels and arrows permit this. specify the rules of the game we shall play.aces play with relations in a group presentation (see Chapter 6) or with a knot diagram using Reidemeister moves.50) ABC~ AxCand Bx.1 paste cut A A Fig.) Operation 1 (Fig 4.
53).) This corresponds to splitting an edge (Fig. we can assume that only a single edge is needed to join the two ends of B. (Here the order is very important: uv or VIU.I is all right but vu or u.Cutting and pasting 73 Operation 3 x~uv where u. Composite operations By combining the above elementary operations we can build up some very useful composite operations. The inverse of this is uv~x and can be done if u and v always occur together in the symbol.I VI does not count. 4. where y is a new symbol not in A. We mentioned earlier the idea of chopping or subdividing the triangulation further. 4. in operation 1. Cutting along this path we have a series of labels which always occur in the same order and so we can apply operation 3 to them to replace them by a single edge label.52 Remark on the operations You may be wondering why. This may alter our triangulation slightly but this does not affect the arguments or ideas we will be using later on. B. The validity of these moves is indicated by diagrams. joining the vertices I and 2. . 4.52). x u v A c:~ A Fig. C. v occur nowhere else in the symbol and the replacement is at every occurrence of x. Operation 4 Replace ABxCxD by AyCB1yD. By doing this if necessary we can find a path along edges inside the polygon. or D (Fig.
where y is a new symbol not in A.I D. or D. (In operation 6 we take C to be empty and B to be CB. a .1D by AyCy1 BD.) Now suppose we can find some recurring letter. Operation 7 Replace AxBCx. Operation 6 Replace AxBCxD by AyCyB. aI ..53 Operation 5 Replace ABxCx.faces ALJ~:3 A cr::fL__. 3 ~ D ~ D o_ _ li_:. 0 •.JI By D Fig. B the block of letters before the first s. a . The proof is much as above.. ..1D by AyCBy1 D. . Step 4 We call this step assembling the crosscaps. A crosscap is a projective plane contained in S. C.. we can use the above pair of operations to replace the symbol with one of the form ftE'.. we can use operation 4 to replace it by AyCB1yD and then operation 6 to obtain AzzBC.. Again for these two composite operations. so try and draw the sequence of diagrams for yourselves. We note that if we have ABxCxD in a symbol. we leave it up to you to draw the diagrams which check the validity of the operation. This is an extremely useful exercise to do.. and so on... Taking A to be empty.. especially if you do not see immediately why one of the operations is valid. Our first task now that we have these composite operations at hand is to get all the crosscaps to the start of the symbol. that is one which starts with a crosscap. say s.1 D. . 4. The projective plane has symbol aa and in any surface symbol the occurrence of a repeated letter .. or aI .t . B... indicates the presence of a projective plane and hence of nonorientability (more on this later). in our symbol.
AxCyBEDx1yIF AuBEDCyu1yIF AuVUIVIBEDCF ~~ ~~ ~ . To do this we use the following series of operations ABaCbDaIEbIF ~ Op.Cutting and pasting 75 Next examine B' to see if some other letter recurs. b . If a surface symbol contains the pattern .. C.I D.. • .5 AxCbDxIBEb1F.I it means that there is an attached torus hidden in the surface. b.54) Fig. and D). for new meanings of B... 4. we say that the surface contains a handle if such a pattern is contained within the symbol (Fig...g.. Repeating this process as many times as proves necessary. a_I . we will eventually get the symbol in the form where B has only edges that occur singly or as . a . as we have seen...I . we might find B' = Bu CuD. As an attached torus looks like a handle. Thus a surface symbol for the torus will be abaI b. If it does (e. Step 5 XI • . x . we can apply the pair of operations.. 4.. A torus has.. a description as a rectangle with opposite sides identified.54 The fifth step is to bring any handles together and then collect them up just behind any crosscaps. this time with A = tt to replace the symbol with ttvv BC.
x .I can be replaced by xxyyzz. Then the letters occurring between x and x. B to be the block before the first letter of the chosen handle.I c.I dc. The crosscaps influence the handles.. . One obvious form we might expect to find in H is a cuff corresponding to CdCI (Fig. a . a ..I (. XI . occurs./ c.. .. changing each handle to a pair of crosscaps..I reverse) ofOp.' H if no ..'. we can reduce any symbol either to alai' . a . thus we have either .I can occur nowhere else.. however.I followed by the collection process of the crosscaps as described previously. Is this the only form within H? To find out we suppose x occurs twice in H..aces So taking A to be the string of crosscaps. a ... so we repeat as many times as is necessary to get the symbol in the form [crosscaps] [handles] [the remainder]. and so on. as soon as any . 4...55).. we can manipulate the symbol so that it has the form Of course B' may still contain handles. Choose such an x with as few letters between x and x.... Using this as many times as need be... . hmcmb.. a"a"H or otherwise b l C I b"ll e"ll .. (It may be empty in which case we stop.. there is a curious phenomenon that occurs if both crosscaps and handles are together in a symbol. occurs In both cases H contains no crosscaps and no handles.I as possible. It is a valuable exercise to write down all the details to check that aabcb.) Step 7: assembling the cuffs We now need to look at H which we shall assume is nonempty.8 hdhc.. xlor . that is crosscap + handle = 3 crosscaps.. The sequence of moves necessary to check this is aahch. x .. .. Step 6 You might expect us to attack 'the remainder' next.
however.. At the moment. ~ . . Then H becomes Axbx. These can be amalgamated using operation 3.1 u at the end of the symbol. XI . Now W may be empty in which case we stop and breathe a sigh of relief. Continuing in this way we can replace H by y. it must consist of single Occurrences of letters. If A consists of handles. we get . To make another cuff we use operation 2 to add u. since cyclic permutations of the letters make no difference (corresponding to a change of starting vertex). that is another cuff as promised. Repeating as many times as is necessary we can pass u through the crosscaps.) We can therefore replace (using operation 3) this string of letters from x to x' by a single letter . thus we now have A H wu.u (using which operations?). Now A consists of crosscaps or handles.... If. we can replace by uAHwu.1 we obtain AHuwu. xbx' . ZI Yi' Y2Z2y.1 . 1 ' n " ualal~alu al~ala.I u which. w say.Cutting and pasting 77 b' Fig.' B which in turn can be changed to yzyl AB (by which operations?). Finally we use that uyzy.. . Thus from uAHwu..Ic.. x . 4..55 (Why? Remember there are no crosscaps or handles in H. W is not empty.. we note that ubcb. We have proved the following proposition. and will eventually give us another cuff. If a l a l is the first crosscap in A. so that we may pass u through all the handles. . it is more a hole than a cuff.. Yszsy~ I W.' .. whilst H consists of cuffs. being replaced by a single letter.l .I can be replaced by fgflgI U (again we leave it up to you to work out why).I can be replaced by vzvlu to show that u can be passed through all the cuffs.. however.
the instructions for gluing the polygon to get the surface). but using operation 2 we can replace it by aa. The classification theorem for surfaces.56 z V Cone Sphere with one hole Fig. (This form is called the normal form of the surface symbol. Two surfaces are homeomorphic if and only if the normal forms of any symbols representing them have the same number of crosscaps/or handles and also the same number of cuffs.I • This is a sphere. and 3 to one of the form AH where A is a string of crosscaps or handles (not both) and H is a string of cuffs. mples of normal form (a) 0 the empty normal form. 4. As the converse is clearly true~homeomorphic surfaces have symbols with the same normal form~we get the following first form of the classification theorem. Any surface symbol can be transformed by the operations I.57 . the two surfaces must themselves be homeomorphic. In this form it is hard to visualize.) If any two surfaces S) and S2 yield surface symbols with the same normal form then as their polygons are homeomorphic by a homeomorphism which recognizes the labels and orientations of each edge (i. 4. a Fig. 2.Surfaces Proposition.e.
4.58).58 (c) aba'b'.2 1. 4. this is a torus (Fig. jp>2 (e) a. a 0:..a2a2 two crosscaps = Klein bottle.: • a a b b a Fig. 4. This is not obvious. 4.60. Find the normal form of its surface symbol abac.a. (d) aa is the projective plane.::. Classify the symbol b'cda'dc'ea.59 4. but look at Fig.a2a3aj'a2Ia31 . (a) Find the normal form of the following symbols: a. 4.59.Cutting and pasting 79 Fig. As we have already seen. The Mobius band can be represented as shown in Fig. 2.60 3.. 4. b a a c Fig..
we can easily construct an orientation for any surface whose normal form has only . 4. A surface S is nonorientable if and only if the normal form of a surface symbol for S contains a crosscap.. one finds that the two triangles that share the edge x are oppositely oriented (Fig. x . so the surface must be nonorientable. no such repeated x occurs. on the other hand.. Choose a path within the polygon between these two occurrences of x. then as the operations we have used can neither create nor destroy orientability we find no crosscaps in the normal form. ntability revisited The first invariant property is one we have already considered briefly. we can get a surface symbol and thus eventually a normal form.. We can use the triangulation or the surface symbol to obtain certain information on the surface.. but any information we thus obtain should not depend on the choices we have made. For instance. However.61). so somewhere in S we have the pattern . This occurs if and only if any symbol for S contains a crosscap.Surfaces (c) (d) When you have done a few of these calculations and you can imagine the tedium of performing all the operations step by step on some long surface symbol. you will probably find yourself asking: is there a simpler way of doing this? Happily the answer is yes. the number of letters that occur in a surface symbol is clearly no good as a measure of the geometric properties of the surface since if we have a disc.. so let us get on with this straight away. Proposition. The use of invariants of surface symbols allows the normal form to be found very quickly and easily. but topologically it is still a disc. Invariants of surfaces: orientability and boundary curves If we choose a triangulation. Proof Suppose in a symbol for S we have a crosscap. we can subdivide it in many different ways to get different numbers of edges. passing through the edges of triangles and through each triangle at most once. namely orientability. Orient successive triangles along this path similarly.. x . If..
Unfortunately because operation 3 allows us to subdivide an edge. we start by choosing one edge and labelling one end. say a ) Q If the label a occurs twice on the boundary of the polygon. but in aba. 4.I b. The next edge to a. to see how you might find an orientation in general.Invariants of surfaces: orientability and boundary curves 81 Fig. Looking at the polygon of the symbol. all the edges start and end at the same vertex. but we clearly manufactured this by using operation 3 on the normal form abaI b.1 • With this example it was easy to guess what the normal form was. For instance. the number of edges that occur exactly once is two. say b. we insert Q as a label at the appropriate end of the second edge. now starts or finishes at Q. so we must insert Q again.1 b1.61 handles or cuffs. the number of edges that occur exactly once in a symbol may not be the same as the number of cuffs in the final normal form. For instance.I cfc. in the torus symbol aba.1 one needs two vertices as we shall see. This number is precisely the number of edges in the normal form that occur exactly once. How can we attack this in general? We start by finding out how many distinct vertices there are around the boundary of the polygon. if the symbol is abaI b.I .) ber of boundary curves In the normal form it is clear that the number of cuffs is an important source of information.I cdec.1 cfc. but if the symbol had been all mixed up using others of the operations this might not have been so easy. (Try some simple examples such as a torus and a torus with one cuff etc. that is ..
b Q R • e Q . This means c starts at Q and at this stage we get a a b Q b Q Q Q Q Q. this occurs twice a a b b c d c.I ). this has two boundary edges. Continue until all vertices on the boundary of the polygon are labelled. As a simple example look at a surface with two crosscaps and a cuff.1 ends at Q (since the symbol as a whole must start and finish at the same vertex).aces Q I • b a or • Q I . merely work with the symbol. Earlier we looked at the example of a polygon with symbol b... They fit together to gIve a boundary curve: . This gives a a b b c d Q Q Q Q Q R R Q so there are two distinct vertices on the boundary of the polygon. Now label any unlabelled end of an edge with a new label R.1 e R a R Q Q so we find that b and e share vertices.1 Q Q so a starts as well as finishes at Q.I Q C d Q Q aI d Q Q c. No more Q can be fitted in. and insert R wherever required. Often this will insert Q at at least one new position allowing us to repeat the process until no further insertions of Q are called for. We will not draw the polygon.I dc. so we insert a label R as the end of c (and thus the start of c.1 ea. We will show intermediate stages: a a b b CdCI Q Put Q as the label of the end of a. An edge whose label occurs just once in a symbol is called a boundary edge. b starts at Q. b wherever b reoccurs (if indeed it does). If we work out the vertex labels on this example we find b. band e.I cda. and c. say.
If. cises 4.Invariants of surfaces: orientability and boundary curves 83 In general.3 1. so this does not alter the number of boundary edges or curves. If x is not a boundary edge then no change occurs. Again it is clear that as x occurs twice. Operation 1 ABC~ AxCand BxI • The new label x occurs twice. We will show that the number of boundary curves is unchanged by the processes that the symbol goes through as it is transformed into its normal form.l adbc. Operation 3 x~~uu. Operation 2 AXXIB~AB. but there may be more than one of these.1efd. The number of boundary curves is the same as the number of cuffs in the normal form for the surface symbol. no change occurs in the number of boundary curves. Firstly we note that the choice of starting vertex and direction in the reading of the surface symbol does not affect the boundary edges. /t is therefore an invariant of the surface. How many vertices are needed for the symbol abc. Proposition. this will prove the result. Find the number of distinct vertices for a polygon with symbol having g crosscaps or handles and r cuffs. the number of the boundary curves does not depend on the triangulation we chose. we join up each string of boundary edges to make boundary curves. x is a boundary edge. As the number of boundary curves in the normal form is easily checked to be the number of cuffs (do check this). Now consider what happens under each of the elementary operations in turn.l ? . but the number of boundary curves will stay the same. the number of boundary edges will increase by one. however. Proof. 2. As we hinted at earlier.
4. In this case it is clearly true. or we have a cuff and again a o = v. we find v = v' + 1.62 We handle the two cases separately and each will lead to some 'subcases' . then X(S) = v . The Euler characteristic X(S) is given by The usefulness of this number in various contexts and guises will be one of the themes of the coming chapters.e + 1. and a 2 = 1. This is perhaps a bit of a surprise. Its usefulness here is that it does not depend on the triangulation used and is an invariant of the surface. and a 2 triangles. a] distinct edges. If k = 1. and OQ. until you see why it happens. = 1. k. Suppose that the surface symbol of S gives v distinct vertices and e distinct edges on the boundary of P.Surfaces Invariants of surfaces: Euler characteristic and genus Suppose S is a (combinatorial) surface with a o distinct vertices. so S = P. then either there are no identifications. Recall that P was constructed by adding triangles in one of two ways to the disc already constructed. y P' Fig. of triangles making up P. a] = e. We will prove this by induction on the number. then the vertex between a and b is a new vertex. The two ways were as shown in Figs 4. one . Proof. Let P be a polygon representing a surface S.62 and 4. Proposition. On recording the change in v and e from P' to P. Now suppose k > 1 and that we know the result as long as P has fewer than k triangles in it. (A) We first look at what happens if a and b occur nowhere else in the symbol.63. a o = v. a] = e.
by hypothesis. and e = e' + I since although we have gained a and b we have lost y from the boundary. + 1. Gl2 from do. The other subcases are similar. x(S') = ao  ~ + a.e' + 1 and so stays unchanged.(~ + 2) + (a. We had one new vertex.I which again balances the equations. so Gl2 = a. ~. e = e' . + 1) that is.e + 1 = v' .1. then we get ao=ao. Here there are only two subcases: either a is in the boundary of P' or it is not. a2 = a.+I. If a is a label on the boundary of P' and b is not. The equations balance as before: since by hypothesis v' . say S'. e = e' . + 1. as required. e = e' (a is in. so v .e + I = ao . Gl2=a. then v = v'. Now P' represents some surface. a and b. and we have.a] + a2' If a is a new edge. + 1. two new edges. Also we have a o = a o. then v = v'. leaving the details for you to check. so a] = ~ + 2.1.e + 1 = v' . a] = ~ + 1. not in the boundary of P'. a o = do .I triangles. and one new triangle. whilst v = v' . Thus a x(S) = (ao + I) . Finally if both a and b occur on the boundary of P'. whilst v=v'l. the corresponding numbers for S'.e + I. that since P' has only k .63 new vertex. a. .~ + one gets v . 4.Invariants of surfaces: Euler characteristic and genus 85 Fig. = v'  e' + 1.e' + 1 as before. then the same argument applies. y is out!).e' + 1 = ao . • a2. . however. e = e' . so a o = o + 1. Thus v . a] = a. We can. X(S) = X(S') = v' . aj.+I. We will use the same notation but will explain less along the way. a 2 = a.. then we get a] = a.e' + 1 = v . If it is.2. a. Suppose now that a is not a label on the boundary of P' but b is. so XeS) = xeS') and we complete the argument as before. Gl2 = a. We now look at case (B). + 1. .=a. work out a o.
and (iii) Euler characteristic are not only invariants of the surface. the number . we have: number of vertex labels number of edge labels hence Euler characteristic Type: Normal form: r = + I 2r + g 2 . We will provide just a sketch.. agbgaglbglYIZIYi! .l. but overall the number of vertex and edge labels stays constant. we will leave the details up to you (but do write them down!).l . calculate g. (ii) number of boundary curves.2g  r.. Type: Normal form: Nonorientable. From the exercise earlier in this section.l. r cuffs. so v . The transition to normal form does not change the values of the expression v . We now need only to work out X(S) for the various normal forms to have complete information. Yrzry.. Orientable. r cuffs.g  r. .1 We have seen that (i) orientability..e + 1 stays the same. reassembling the polygon in a new way. alai'" agagylzly.I .. We can therefore find rand X(S) and by choosing either the orientable or nonorientable formula from the above.e + I determined by vertex and edge labels. Proof. (As the detailed proof is a bit like that of the invariance of the number of boundary curves. Lemma. g crosscaps.Surfaces The next stage in checking that this Euler characteristic is an invariant is to see what happens under the operations. but also able to be calculated from any symbol for the surface. This reglues an edge. • We have shown that X(S) is independent of the triangulation and can be calculated from the polygon in its normal form. We get (again from the exercise): number of vertex labels number of edge labels hence Eu ler characteristic r + I 2r + 2g = 2 . a1b1ailb.) Both operations 2 and 3 give us one new vertex label and one new edge label. Any application of operation I is followed by one of the type of its inverse. Yrzry. g handles.
d . etc. we will look again at our symbol b1cda1dclea.. If Al is a symbol for SI and A2 one for Sz. orientability.X if S is orientable.. Given two surfaces Sl and S2. This number.2. Classify the following surfaces giving genus. (a) ablcdad1ceb eabc1adbclef ab1cdald1c1eb (b) (c) 2. . d d e a Q Boundary edges: Genus: p Q Q Q Q p Q Q band e giving one boundary curve. denoted Sl # Sb is formed by cutting a small circular hole in each surface and then gluing the two surfaces together along the boundary of the holes (Fig. 4. Two surfaces are homeomorphic if and only if they are both orientable or both nonorientable. 4.1 and AzyzI. g. s 4. Thus S has normal form alala2a2a]a3yzyl.. Orientability: Vertex labels: Nonorientable because of c .4 I.e + 1 = 2 .64). First calculate XeS) = v . have the same number of boundary curves and the same genus. g = 2 .65) that is by replacing the 'holes' by cuffs.r . As an example. g = 4(2  r . Euler characteristic. (Fig. d .X)· We summarize the theoretical aspect of this as follows. classification theorem for surfaces. We have: if Sis nonorientable. one can obtain a symbol for S = Sl # S2 by eliminating y from the pair of symbols Axyx. their connected sum. so r = 1.5 + 1 = .Invariants of surfaces: Euler characteristic and genus 87 of crosscaps or handles. The practical aspect of the result is clear: it saves us work! The method of cutting and pasting does lead eventually to the normal form but the above ideas allow us to classify surfaces quickly and painlessly. is cal1ed the genus of the surface. . so g(S) = 3.
(Note you need to take care if one surface is nonorientable whilst the other is orientable. 3.) (c) Prove that !p 2 # torus is the same as !p 2 # !p 2 # !P 2 • The classification theorem can be rephrased in terms of connected sums since a genus g surface with r cuffs is either the connected sum of g tori (if orientable) or g projective planes (if nonorientable) together with r cuffs. Glue Fig.65 (a) Find XeS) in terms of X(Sl) and X(S2)' (b) Find g(S) in terms of g(SJ) and g(S2). Show that . Suppose we have a triangulation of S using ao vertices. a J edges. and a 2 triangular faces. 4.aces Holes cut out J. Let S be a surface without boundary curves (so no cuffs appear in the normal form).64 Glue Fig. 4.
66 4.68 .67 Fig. (ii) the Klein bottle.XeS») 89 Now by considering the maximal possible number of edges joining a o vertices. this is the double of SI (or S2). by identifying the boundary edges of S] and S2.68).67) except if an edge is a boundary edge (Fig. 4.. Can you find a triangulation with exactly this number of triangles? Fig. You might visualize it as being constructed by replacing each triangle (Fig. (iii) the sphere. 4.J49  24X(S) ) These calculations give a lower bound for the number of vertices needed to triangulate S.Invariants of surfaces: Euler characteristic and genus (b) al = 3(ao . D. 4. 4. 4. Find a lower bound for the number of triangles in a triangulation of (i) the projective plane.. Construct a new surface. show that (c) ao ~ t (7 + . 4.66) in a surface by two copies near to each other (Fig. Fig. Let SI and S2 be two homeomorphic orient able surfaces of genus g and with r > 0 boundary curves.
has genus 2g + r . In the next chapter these will be used in a study of the 'arithmetic' of knots.70. Euler characteristic 4 . Given any knot K (in 1R 3 ) there is an orientable surface S with one boundary curve embedded in 1R3 in such a way that the boundary of S is the knot K. (In our crossing we throwaway the parts of the knot between XI and Y2 and between YI and X2') As an illustration. Theorem. the discs to be disjoint.69. K. 4.69 Now delete the crossing leaving a set of oriented circles. Orient the diagram then at each crossing.Surfaces Show that D is orientable. and so on. 4. The Seifert circles mayor may not be nested. and no boundary curves. . as in Fig.4g .2r. Fig. In the next few pages we will see how these two classes of object interact at a deeper level and hence that some parts of the classification result for surfaces can be used to produce results in knot theory.1. but compatible with the orientation. called Seifert circles. the reduction of classification problems to moves. Knots and surfaces So far the only connections between knots and surfaces that we have met in this chapter have been ones of methodology: ideas such as the use of invariants. Take a projection of the knot. consider the process outlined above performed on the figureeight in its usual projection. Starting from any innermost circle and working outwards fill in (the word 'span' is often used) each circle with a disc. as shown in Fig. Proof. 4. add two extra directed arcs bypassing the crossing. Our main aim here is to prove the following result of Seifert.
Knots and surfaces 91 .0 /. 4. I Y x2 I of 1. 4.. Fig.72 .72.70 x.71. 4. 4. :> crossings () Three Seifert circles Y2 I Fig. Thus the crossings now look as in Fig. I Y. Do this at each crossing. each one filled with a twisted Fig.71 The two arcs and the parts of the crossing removed make up a rectangle as in Fig.. . 4. Fill this in with a solid rectangle. Remove the \ .
increases by two. 4. has g = ~ (d . 4.74 We start with s discs. As each twisted rectangle (Fig. does not change since there are no new vertices. Proof.rfaces Fig.a l + ~ by 1.73 rectangle. 4. the number of edges. each contributing 1 to %(S). S.d and as S is orientable g(S) = 112 (d .74) then the number of vertices. the surface obtained by gluing the twisted rectangles onto the Seifert circles will be orientable. a different projection might give a different surface and hence a different g. position. ao. and then add d rectangles. If we add a rectangle between two discs (Fig. however.s + 1). ai' increases by three. 4. • You might expect this number to be an invariant of the knot. Disc Disc Fig. each reducing a o . A disc has Euler characteristic 1. constructed from K by the above method. lX:l. Each spanning disc of each Seifert circle has a natural induced orientation coming from the orientation of the knot. To avoid these problems we take the easy way . genus The surface. and the number of triangles.73) can be oriented coherently with these orientations.s + 1) where d is the number of crossings in K and s the number of Seifert circles arising from K. hence %(S) = s .
so the surface that resulted did have minimum genus 1.75 2. With more powerful computing methods and with the renewed interest in knot theory due both to the development of the new polynomials and the perception of potential for applications. but the form of definition makes it difficult to calculate. the question of developing methods for calculating the genus of knots is again receiving a lot of attention. It is useful to note the following: I. For example. we can only say that g(K) ~ t (d  s + I).75). If g(K)::t= 0. If g(K) = 0. If K is known to be nontrivial and one finds some orientable surface of genus 1 spanning the knot then g(K) = 1. The genus thus defined gives an invariant of the knot. then K is the trivial knot. (In our example. the trefoil can be spanned by a Mobius band (Fig. so BEWARE! Fig. Since the Seifert construction does not necessarily give a minimal genus surface.Knots and surfaces 93 out and make the following definition: the genus g(K) of a knot K is the minimal genus of all orientable surfaces which span K. 4. 4.) 4. we looked at the figureeight knot having four crossings and three Seifert circles. Haken in 1961 but at the time it was too unwieldy to be useful in practice. it can be extremely difficult to calculate the genus. A theoretical algorithm was developed by W. 3. 1. It is quite easy to span certain knots with nonorientable sur faces. Can one be certain that no smaller genus surface spans the knot? .
77 . 4. In general.d(t) is less than or equal to 2g(K) (see Burde and Zieschang 1985).e. 4. the degree of .77) (c) any (2.79) Fig. 4. there is some projection in which. 4.5 Find bounds for the genus of the following knots: (a) trefoil (Fig.rfaces 5. 4. n)torus knot (Fig. es 4. If K is an alternating knot (i. as one goes along the knot.78) (d) any generalized figureeight knot (Fig. There is a strange link with the Alexander polynomial. overpasses alternate with underpasses). then g(K) is half the degree of the Alexander polynomial of K.76) Fig. 4.76 (b) cinquefoil (Fig.
79 . 4.Knots and surfaces 95 n crossings n odd Fig.78 Fig. 4.
hence the title of this chapter. A basic analogy used in knot theory is that the sum of oriented knots behaves very much like the product of natural numbers. When two different types of situation seem to be running parallel from certain viewpoints.1 . The details are as follows. 5. K L ' Fig. Of course it is as important to notice where the analogy fails as to follow those situations where it works. 5. one can try to see if the similar behaviour extends further to give new insight into the different structures.1.5 The arithmetic of knots Analogy is a very powerful logical tool. Then we can form their 'sum' in an obvious way following the natural intuition of first tying K and then L in a piece of string. sum of oriented knots Suppose that we have two oriented knots K and L. Cut each knot as shown in Fig. .
as in Fig.2. 5. 5. . 5.3 and 5.The sum of oriented knots 97 Fig.3 How does this operation + act on the class of all knots? Does it respect equivalence for instance? Does it depend on where the cuts are made? Is + an appropriate choice of symbol? Does + behave like the addition of integers? The following goes some way to answering these questions. s (Figs 5.2 Join them together preserving orientation. We wiH write K + L for the result. We shaH denote equivalence of knots by the symbol .4) granny knot = trefoil + mirror trefoil Granny knot Fig.
K + (L + H) (d) K + trivial knot . Or rather 'proof by sketches'. Fig. the result seems a lot less obvious than any of the others. Now pull K back to the size it was to start with (Fig. Again we omit the proof.6. K + L . you have to use the homeomorphism sending K onto K' to construct that sending K + L onto K' + L. but again one has to construct things with great care so as to be able explicitly to produce the homeomorphisms. We illustrate the idea in Fig.7 and 5.4 Proposition (a) If K .L + K (K + L) + H . for (b).The arithmetic of knots reef knot = trefoil + trefoil Reef knot Fig.K' + L (b) (c) K+ L . This uses details of the behaviour of homeomorphisms that are deeper than we have yet gone.K. a sufficient reason to believe the result is as shown in Fig. The full proofs of (a) and (c) are long.K' then for any L. the fact itself is unsurprising if not obvious. 5. For (c). 5. Make K smaller and slide it along L as in Figs 5.5 Finally. 5. 5. however.8. For (d). .5. 5. so we will omit the proof. A full proof can be based on this 'sketch' proof. For (a). the reason for this property is very pretty and easy to seeonce it is pointed out to you.9).
9 . 5.The sum of oriented knots 99 K+L &1) Fig.6 Fig. 5. 5. 5.8 = L + K Fig.7 K L Fig.
genus. 5. (These structures are called commutative monoids.5. Firstly that there are many other systems which behave as above. So far we have developed this knot sum showing that it behaves much like addition of natural numbers.K' + L. 1t L 'Piping' rectangle I Fia. various polynomials) we can hope that investigation of how the value of an invariant on K + L depends on its values on K and L will at the very least shed some light on these problems and perhaps indicate new properties to investigate further. If we place SK and SL in different halfspaces. colouring properties. we can 'pipe' them with a rectangle as shown in Fig. associative. is K . that is. if we have K + L is there some L' such that K + L + L' .) As this is the case we must not jump to the conclusion that this 'analogy' is in any way complete. commutative. a very suitable symbol when one remembers that the identity is the type of the unknot.K? What about cancellation: if K + L . ienus of K + L Suppose we have surfaces SK and SL spanning K and L respectively and such that g(SK) = g(K) and g(SL) = g(L) that is they are minimal genus spanning surfaces. we do not yet know if it is possible to subtract knots.10. and has an identity which we write 0. For instance.10 .The arithmetic of knots These results then show that the operation + on knot types is well defined.K'? Although we have as yet few invariants for knots (bridge number. It is necessary to pause here a moment and consider a couple of points.
does not surround any smaller loop (Fig. 5. we suggest you look up Lang (1972). that is that S n 1r consists of a finite number of disjoint circles and an arc joining A to B (Fig. the equation to verify is so g(K + L) ~ g(K) + geL). We suppose K and L as before are on different' sides of some plane 1r with the joining arcs crossing cleanly. trying to see if that is the case or if sO. 5. 5. . 5. that is each intersecting 1r in a single point. say. Shifting S slightly if necessary.me as yet unknown feature comes in to block equality. We leave the justification to you. g(K) This simple argument thus suggests that maybe g(K + L) is + geL).11. (The proof that this can be done requires a study of transversality. We will investigate this relationship more closely. 5.) Some of these circles may be nested. We can calculate g(S) in terms of g(K) and geL). that is S spans K + L and so g(S) ~ g(K + L). we can assume that it also cuts 1r cleanly.14). as in Fig. B}. so 1r n (K + L) = {A.11 We have some oriented surface S spanning K + L and having genus exactly g(K + L) (Fig. moreover. If you like this type of argument and want to know more. That boundary component forms a copy of K + L.Genus of K + L 101 The result will be an orientable surface S with a single knotted boundary component. A B It Fig. Choose a circle that does not loop around the arc from A to B and.12).13).
5. if so throwaway the part not containing K + L in its boundary. '~+= s Fig. The new surface has the same genus as our original S Snew s :JC 1t Fig. 5.13 Put a cap on this circle and pull it away from 1C to get a new surface still spanning K + L.14 . This process may have disconnected the surface. 5.e arithmetic of knots .12 A ( o@) B o Fig.
so this first guess is not right. then the second arc thus created contains no overpass and hence no.2 give us a similar result for b(K + L) in terms of b(K) and beL)? Our first step is to use guesswork: is b(K + L) equal to b(K) + b(L)? The obvious first reaction is 'probably'. Now compose the knots as in Fig. Pick the cutting point A in K to be between the first underpass and the first overpass and in L to be between the last overpass and the second underpass.15. We have Combining this with the result of our earlier work gives a proof of the following theorem. Cut along this arc to obtain two surfaces SI and S2 spanning K and L respectively. as in Fig. If K and L are oriented knots. Suppose we take a diagram of K and a diagram of L. The reason is not hard to see. but is this the smallest possible number? By choosing the points where K and L are cut we can reduce the number of bridges by one. . The result would seem to have b(K) + beL) bridges.1.Genus of K + L 103 since Snew still spans K + L. that IS g(Snew) = g(S) = g(K + L). ge number of K + L The above discussion on g(K + L) leant heavily on the minimality of the genus of the chosen spanning surface. Continuing in this way we get a minimal genus surface S' spanning K + L such that S' n 1r is just the arc joining A to B. rem (sum theorem for genus). but S has minimal genus among spanning surfaces of K + L. then g(K + L) = g(K) + geL). of bridges in diagram = b(K) + beL) . 16. Can a similar type of argument to that used in Section 5. 5. 5. We repeat this process until we are left only with circles which loop around A and B. Snew has genus at most that of Snew (since our 'surgery' above can at best have chopped through a handle or have resulted in some handles in a thrownaway piece). Bridge number was also defined using a minimum over all possible positions. each having the minimal number of bridges for the respective knot. but if both K and L are the unknot then b(K) = beL) = b(K + L) = I. Starting with the largest such circle attach discs around the backs of K and L.
arithmetic of knots Fig.1.I).15 Fig. 5. since b(L) = I. This formula is correct but we will not attempt to prove the result here. then this would be consistent (giving b(K) = b(K) + I . We note that if L is trivial.16 It might be possible to reduce the number of bridges yet more by some 'cunning trick'. Our experience with the genus sum formula suggests that perhaps b(K + L) = b(K) + b(L) . 5. so we get b(K + L) ~ b(K) + b(L) . .1.
can we fracture it.K) = 0. The reasons that such termination is important are clear. then b(K + L) = b(K) + b(L) . associative. orizations. and the factorization theorem 105 rem (sum formulae for bridge number). We can also hope that a complete answer is unique. If K and L are oriented knots. is what makes it useful. The uniqueness of the 'prime factorization" 2 x 3 x 3. The halfway stage of our fracturing of 18 was given as 3 x 6 but we could equally well have given 2 x 9. 0 for knots) and we have just seen that + on knots does not have an inverse operation. This suggests that an analogy between the factorization of knots and the factorization of numbers would be worth investigation. Is there a 'subtraction' of knots? Let us suppose we could find a '. What evidence can we get for 'primes' among knots? Suppose K is a knot with genus I. Both are commutative.1.K' so that K + (. and the factorization theorem We saw how the sum of knots was well behaved. you could look up the original proof. We started out this chapter talking about analogy. It is in Schubert (1956). There is no 'additive inverse'.g(K). It is not much good if one can go on and on and never know if one more step would give a complete answer. Perhaps the better analogy would be between the sum of knots and multiplication (not addition) on N. say as K J + K2? We would have . Our attempt to make an analogy between the sum of knots and the sum of numbers only goes so far. If you can read German. prime knots.K exists in this sense is K = 0 itself.Factorizations. and have a neutral element (l for N. just as there is no multiplicative inverse in the natural numbers. This would be the case provided certain conditions on this factorization process were satisfied. prime knots. If we could fracture a complicated knot into simpler factors then presumably knowledge of these simpler factors should tell us everything about the original knot. then g(K) + g(  K) = 0 and g(K) = . Ideally we could hope that the fracturing process would terminate just as 18 = 3 x 6 = 2 x 3 x 3 cannot go further without introducing trivial factors. much like the sum of numbers. That is strange: how could a knot have negative genus? So the only knot K for which .
More precisely. The only way this can happen is if one of b(K J ) and b(K2) is 1. both K~ and K.... We know that b(K) = 1 means that K .K j + K2 one at least of K j and K2 is the unknot. + Knfor some n > N then some of the K j are trivial. rna. nontrivial. • KI Lor KI AI (where we write We only sketch the proof giving an idea of why the result is true. Suppose K . i = 1.) Proof. We could equally well use the bridge number formula.) 1 then. + Kn and all g(K j ) of course. What can we say about knots with bridge number equal to 2? Are these prime? If b(K) = 2 and K = K J + K 2. given a knot K.. The first result guarantees this.O. L lies on one side of n. It is about time that we gave a formal definition of prime knot as we have been happily (?) discussing them for a while now.I so b(Kil + b(K2) = 3.K j + .2 g(K) > g(K.2.K~ + K. g(K) ~ n. that is the corresponding K j . There is a plane n such that n u (L + M) consists of two points only.. Either argument shows that the trefoil is prime. ~ The lemma follows since if K . but need not be 'best possible'.K J + . A knot K is said to be prime if K is not trivial and whenever there are knots K j and K2 so that K .) ~ I for i= 1. (In fact N = g(K) will do. In studying decompositions of knots we earlier mentioned termination. rna. The 'knotted part' of K .). then 2 = b(K) = b(K j ) + b(K2) . We can therefore list quite a few 'prime' knots starting with the trefoil and the figureeight. to mean K is a factor of L. If K is prime then KI L + M implies K I L etc.e arithmetic of knots so as any genus is nonnegative.O. A knot cannot be indefinitely factorized. there is a natural number N such that if K . one of the g(K j } must be zero. and M on the other. The second result we note is a 'uniqueness' result. so any genus I knot must be 'prime'. We therefore have that all 2bridge knots are prime. then as g(K.
but you should be able to make some headway towards finding a solution. it is a special type of ordered set.) First look up some proofs of the prime factorization of integers.+Pm where each Pi is a prime knot. rem (prime factorization of knots). If K is a knot. but for composite knots ncolourability is rather a weak property. If K is ncolourable then suppose in forming a sum with L. (The moral of the story would seem to be something like: if your analogy is a good one. The idea of the lattice of divisors of a number may set you on the right lines. Proof (instructions only). Really this should be left up to you to do. Do the 'divisors' or 'factors' of a knot also form a lattice? Are these lattices special in any way? The theory is known. Thus K is a factor of either L or M as required. What about other invariants. we cut K on an arc coloured by the colour a. Such a decomposition is unique up to order. . the underlying logic of the situations will be the same. It can be useful for prime knots. Most will proceed by saying that if n is not prime then there are nb n 2 such that n = n 1n 2 and then will go on splitting n 1 and n 2 • Termination and uniqueness use ideas similar to Lemmas 5. this shows that ncolourability is limited in its usefulness as a property. Find a proof which is simple and adapt it to handle knots.ncolourability 107 must also lie totally on one side or the other of 1C since otherwise it would be simple to decompose K as a sum in a nontrivial way. here are instructions for writing out a proof. then there is a decomposition K~Pl+". Certain parts of mathematics study exactly the logic of such situations.) lourability We have seen that the genus and bridge number are 'well behaved' with regard to the sum of knots. If you enjoy a challenge. see if you can formulate a system or structure which contains just enough to be able to prove a prime factorization theorem. Try to find the definition. (Do follow them. In some ways. Have fun! If you have not met the notion of a lattice. How is one to colour K + L? Simply use the colour a for all parts of K + L that were originally in L and check this works. especially the 'elementary' ones? The easiest to handle is ncolouring.5 above.4 and 5. however.
18 . We suggest that you try some simple examples. 5.18. 5.The arithmetic of knots What if K is mcolourable and L ncolourable? Is it possible to combine the two indices of colourability in some way? The answer is known but we will not give it here.17. Of course your conjecture may be false! Polynomials and the sum of knots We earlier met three polynomial invariants. 5. It should be clear that Fig. deleting the columns corresponding to Yand a. for the unbounded region of the diagrams in Fig. thus getting a square m x m matrix M K • The Alexander polynomial is det(M K ) after normalization. We do the similar process for L. a Fig. a. with Lidt) = M L • The diagram for K + L will look like that in Fig. and then attempt to prove it. put forward a conjecture. This time in the (m + n + 2) x (m + n) matrix we delete columns corresponding to z and a. 5. How do these react with respect to the sum of knots? ander polynomials We take K and L and label the diagrams in the usual way. x and a. We take the same face label.17 When we have drawn up the m x (m + 2) matrix for K we delete the columns corresponding to the adjacent regions.
Polynomials and the sum of knots
109
since no crossing originally in K can contribute nontrivially to a column in the Lpart of MK+L since the only possibility of such a contribution would have been via column a or z and these have been deleted. Of course det( MK+ L) = det(M K)· (detML). that is
dK
+
dt) = dK(t) • dL(t).
In many ways the Alexander polynomial serves as a testing ground for properties of the more recently defined polynomials, so we may conjecture that these react similarly.
fly polynomial
To prove that the Homfly polynomial of a sum of knots is the product of those of the constituent parts we need to prove one or two preliminary results. If K and L are two links, we let K.ll L be their distant (i.e. disjoint, unlinked) union; then
P(K.llL)
=  e + em
1 •
P(K). peL).
Fig. 5.19 If L is the unknot this comes from the cases shown in Fig. 5.19 since K + L+  K  K + L_, whilst K.ll L  Lo. Feeding these into the defining equation for P gives
eP(K) + eso
I
P(K) + mP(K.ll L) = 0
P(K.llL) = _
e + rl . P(K)
m
arithmetic of knots
If L is more complicated, then use induction on the crossing number
of anyone of its diagrams; try to set up this induction formally. To use the Homfly equation we will need to extend the notion of oriented sum of knots to links. One simply chooses one pair of components on which to perform the usual knot sum. We now can consider the three diagrams (Fig. 5.20), so
£P(K + L)
+ e 1 P(K + L) + mP(Kll L)
= O.
Fig. 5.20 This gives
(£
+
e1)P(K + L) =  mP(KllL)
= (£
+ e1)P(K) . P(L)
and so as expected
P(K + L) = P(K) . peL).
Given the link between P and ,,1, we could retrieve our earlier result on ,,1 from this one. We leave you to investigate the Kauffman polynomial.
6
Presentations of groups
We need group theory so as to be able to handle some of the deeper invariants of knots, in particular the group of a knot. We might expect the type of group theory needed would be fairly limited and somewhat specialized. However, the overall process translating geometric problems into algebraic ones is very much two wayalgebra pervades the topology and vice versaso quite a lot of group theory turns out to have implications in topology, whilst the topology can provide new methods and insights in group theory. Attempts to compartmentalize mathematics fail; there are too many important links joining apparently distinct branches. There is an underlying and allpervading logical structure that is difficult to pin down and to see clearly, but which is none the less there behind all of mathematics. ark to the reader The contents of this section will be needed at different times and how you use them depends on various factors. You may have a course of group theory in which some or all of this material is handled; if so this section merely provides a view of that material that will be useful in this book. If such a course of group theory is not available then it will be necessary to treat some, at least, of the topics to a minimum depth depending on what material from the rest of the book is used in the course. Our intention in this chapter is thus to provide a central source for this material and a delineation of the group theory that will be needed. As explained earlier, we consider that group theory and lowdimensional topology interact so beautifully that it would be a pity not to include some vision of this interaction in any course involving knots, surfaces, etc.
Examples of presentations
We hope you will learn some useful and appealing group theory from this chapter, but we must remember it is not primarily a group theory
sentations of groups
textbook, and so we will take a fairly informal approach. We shall assume that you are familiar with the contents of a first basic course in group theory and have met subgroups, normal subgroups, quotient groups, homeomorphisms, and the 'isomorphism theorems'. It will also help if you have some simple examples of groups at your disposal, such as the cyclic groups, the dihedral groups, and the symmetric groups. We will start by examining the dihedral group, D 3 , of order 6. This is the symmetry group of an equilateral triangle (Fig. 6.1)
A
c
Fig. 6.1
B
Let R be the clockwise rotation through 120°, so that R permutes the vertices according to the rule A ,) B, B ,) C, C ,) A. Let 8 be reflection in the axis of symmetry (shown) that goes through A and the midpoint of BC. so that 8 permutes the vertices according to the rule
A ,) A, B ,) C,
C ,) B.
One easily checks that any symmetry of the triangle can be written (in more than one way) as a composite of Rand 8. For instance, rotation through 240° (clockwise) is clearly R2 = RR, whilst reflection in the axis through B (which is given by the permutation A ,) C, B ,) B, C ,) A can be written as R 28R (= 'do R twice, reflect, finish with R'). The identity symmetry can be written as R3 or R!' or ](3 etc. or as 8 2 , 8 4 , etc. H you have not yourself 'played' with D 3and it is not enough to have watched someone else doing soit would be a good idea to experiment now. For instance, give a description of each of the elements in the form Ri 8 j . Write each of them in the form 8 k R as wellis it possible? Remember that RO = 8° = I, the identity symmetry. As we mentioned above, every element of D3 can be written as a 'word' in the elements Rand 8, that is it can be expressed as a product of these two elements. We say that Rand 8 generate the
Examples of presentations
113
group. This generating set {R. S} is certainly not unique. If some set X of elements generates D3 (so any element of D3 can be written as a product of elements of X and their inverses in some order and possibly with repeats), then any Y containing X clearly must also generate the group. In particular, we could take Y = D 3 , so the set of elements of D3 can act as a set of generators for the group D3! Usually, of course, we hope to find as small a set of generators as possible for a group, but we also want to try to get as simple an expression as possible for all the elements in terms of the generators. Of course the meaning of simple tends to vary somewhat from context to context. For instance, a set of generators for a group arising in a geometric context may be considered to be a good one because it allows the elements of the group to be expressed in a geometrically significant way, but whilst a geometer may prefer those generators, the significance of the same group to an algebraist may be such that a completely different set of generators may be considered much simpler. This suggests that we must develop ways of moving from one set of generators to another, a means of translating words in one 'alphabet' to words in another. Knowing the generators is not enough. The symmetry group D4 of a square has order 8; it is also generated by two elements. All symmetric groups are generated by two elements. The cyclic group C 6 or order 6 can be generated by two elements and so on. What distinguishes one group from another are the relationships or equations satisfied by the generators. For instance. in D3 we saw that R3 is the identity as is S2. so R has order 3 whilst S has order 2. but in D4 the two corresponding generators have orders 4 and 2. As an experiment let us try and generate words from Rand S knowing only tha t R3 = 1 = S2:
Thinking about this, it is clear that any word in R, R2, and Scan occur and there is no reason why any two of these should coincide. Thus we have not yet found a defining set of relations for D 3 . Using the same notation as before, RS is the composite
A
C
~B
~B
A
C
B~C~A=B
~A ~C
so is the reflection in the axis through C and has order 2, that is
(RS)(RS) = 1. This gives us RS
= (RS)l = Sl R 1 = SR2
S satisfying This is the case but we have not proved it yet! It might happen that (RSi = I forces others of our list of six words to be equal.1 = R 2 S so it is already in our list. y such that x 3 = y2 = 1 plus another relation. How can we know? ise 6. We add (RS)2 = I into our list of relations and note that we now get at most six words: (For instance.) Thus we have arrived at an informal notion of a presentation of a group. the answer will be given later on in this chapter. (b) D 4 : the symmetry group of the square = dihedral group of order 8 generators x. if one tries to form SR we find RSRS= 1 =} SR= K1S. pies (a) (b) D3 = (x. y such that X4 = y2 = 1 plus another relation. (In both cases.1 In each of the following cases we have specified two generators and two relations. We write G = (X: R) where X is a set of generators and R a set of relations. One extra relation is needed to obtain the desired groupfind such a relation (a) C 6 : the cyclic group of order 6 generators x. y: x 3 = y2 = (xy)2 = 1) C 6 = (a: a 6 = 1).) This suggests that D3 can be specified by naming two generators R. .Presentations of groups (remember S2 = 1 and R3 = I).
Iy to move each x to the left of y and then to repeat the reduction as before. y : xn =y = (xy)2 = 1). If w = w'x. This then reduces to xi+nlyJxn2 if} = I. Try to find a formula for xiy. and y respectively.kn < n. If j = 0 then we have Xi+ I which either is of the required form or if i = n .I . x. If w is a word in x and y. Firstly we will show that any word in x and y can be reduced to one of the form xiyl with 0 ~ i < n. We will need later on to handle presentations 'effortlessly'.Iy which in turn red uces to Xi . 2. This group has a presentation (x. such as commutators and conjugacy. Now look at w with few) = m. Now if anywhere in our word we have yx we can use xyxy = I to get yx = xIyI = xn. The proof now proceeds to check everyone of the different cases that can occur. w'x. The word w can be written in one of four forms w'x.3 is 6.1 or xn . the length of w. Using the relations these reduce to x. or yl.3 = XI. . everything is fine.1 then xiyixI 'reduces' to xiyJ~. w'y or w'yI where w' consists of all the letters except the last one.I and so w' reduces to xiyi by assumption. any power of yother than 0 or I can be reduced to either yO = I or y itself.kn where 0 ~ m .Iy depending on whether i > 0 or i = O. Similarly. As xn = y2 = 1. if n = 5.I reduces further to xO. xiyx reduces to Xi + n . If j = I. any occurrence of xm for m> n can be reduced to one with x m .Dihedral groups 115 edral groups The dihedral group Dn of order 2n is defined as being the symmetry group of a regular ngon. for example the length of x 2yx. is the number of letters in w counted with multiplicities. and so on.) If few) = 1 then w is one of x.<nl.3. xnl. xyr = x(yx)x = xx4yx = x 5 yx = yx = x 4y.i[ I where all the x to the right of y have been removed.J • XI counts as + 3 not . (Note: x. Assume that provided f(lV) < m then the result holds. y. We will use this presentation of Dn to illustrate these concepts. We will also find certain ideas from group theory.) I. (We will not do them all. We can safely leave the details to you. useful. (For instance. y.) This really needs to be set out more formally as a proof by induction on the number of letters in the word. so if few) = 1. If w = w'x then w reduces to XiyJX. thus also indicating how one plays with presentations. but few') = m . denoted few). xl.} = 0 or 1. The idea is simple.
Note: g(ab)gI can be rewritten (gagI)(gbg]). that is if N = gNg. The commutators in G generate a subgroup [G. x conjugated by y is xn .I .i j= l:y(xiY)y=yxi=xniy. we say gl and g2 are conjugate if there is some g E G such that g2 = gglgl. Now if we try to conjugate xiyj by xky we have (Xky)(xiy1)(yIx. Thus the conjugacy class is {x. Our next aim is to work out what the conjugacy classes in Dn look like.I. Being conjugate is an equivalence relation on G and the equivalence classes of conjugate elements are called conjugacy classes. Thus [g]. Find the conjugacy class containing x: x conjugated by any power of itself is x.I. First we note that (xy)2 = I and y = 1 together have another consequence: (xy)Z = I ~ xy=yIxI =yx.I conjugated by any power of x is xn . g2] measures how much gl and g2 do not commute. xn . then the commutator [gl> g2] is the element [gl> gzl = glg2 g"llg21.k) and so we can find the conjugacy class of any element by repeated application of conjugation by the y and the x.1 = Xi+ lyx.G] of G called the commutator subgroup of G.I for all g E G. Try them yourself. and gb gz E G. This subgroup is normal since a conjugate of a commutator is again a commutator.I = xi+ 2y (which may need reducing if i ~ n . Thus g] and g2 are in the same conjugacy class exactly when they conjugate. are easier.]}. xn . (Check it yourself.) . Given two elements gl and g2 of some group G. Note that a subgroup N of G is normal if it is selfconjugate.] conjugated by y is x (why does this return on itself?). We consider conjugating xiyj by x and by y: (1) j = 0: x(Xi)X.2) (2) j = 0: y(xi)y I = yxiy = xn . xn .sentations of groups The last cases. 3 and 4. If G is a group.1 = Xi j = 1: x(XiY)X.
b]l = [b. if G = (X: R) then [G.(Note that when n is even there is a conjugacy class consisting just of a particular power of x. so the Abelianization Gab has presentation Gab = (X: R u ([X.) We have worked in some detail through the various properties of the dihedral groups. x' E X}. Prove that a subgroup N of G is normal if and only if it is a union of conjugacy classes. . a] and lab. In any case you will need the result several times when working with knot groups later. What is the Abelianization of Dn? Remarks Note that commutators obey some simply checked rules such as [a. 2. 3. Dm in order to get you thinking and working 4. c]. x1 : x.Repeat for x 2 and then for xy. Is there a pattern building up? Try to list all the conjugacy classes in Dn. and if n is even? Find the conjugacy class of y in Dn. the subgroup generated by x 2 • What is (Dn)ab if n is odd.) 6. In fact. X' EX}). x']: X. G] is the smallest normal subgroup containing the set {[x.Dihedral groups 117 The corresponding quotient group Gab is called the Abelianization of G. Prove that [Dm Dn] = (x 2 ).2 1. c] = alb.I [a. It is an Abelian group and the quotient homeomorphism qJ : G 7 Gab has the pleasing and useful property that given any Abelian group H and a homeomorphism IfF: G 7 H we can factorize lfFuniquely via qJ. that is we can find a homeomorphism 7 IfF' : Gab H such that IfF = IfF' qJ and IfF' is the only homeomorphism with this property. (Again you should try to prove these statements now even though you will be given a proof later. c]a.
'Trial and error' may work but is haphazard. Using the fact that any normal subgroup must be a union of complete conjugacy classes. 6. we set V= {Vg: g E G} to be a set of vertices in one~one correspondence with G (and thus labelled by the elements of G). Q2n has order 4n. x 2.sentations of groups with presentations. x 3. r. If Vg is a vertex and x E X a generator. If G is fairly small one can draw a Cayley quiver (also called a Cayley colour graph). be finite. This is a neat way of displaying the needed information and is also of great theoretical importance. The labelling is often thought of as a colouring of the graph.J and (Qzn)ab' ey quivers Given a group G and some set of generators X of G. try to find all normal subgroups of Q2n (and hence all quotient groups) and finally work out [Q2m Q2. denoted Q2m has presentation (x. . as we will see later on. We assume that G is known and X is known to be a set of generators.) Next work out the division of the elements of Q2n into conjugacy classes. (In fact. so y4 = 1. X4} X = {x} There is only one label. We construct a graph having the elements of G as vertices. for instance. To start with. if one has a set of generators X and some set R of relations. we suggest you work out the corresponding facts about the dicyclic or generalized quaternion groups. yet (X: R) is infinite because we have not yet found all the relations necessary to 'capture' G. The nth of these groups. C s generated by x (Fig. yI xy = XI). To check that you can handle group presentations in this informal way. hence the term 'Cayley colour graph'. y: xn = y. For a start.2) C s = {I. prove that x2n = 1. More precisely. The resulting graph is both directed (the edge goes 'from Vg to vgx') and hence is a quiver (see Chapter 7) and labelled. how is one to know even if (X : R) has the same order as G? The group G may. we draw an edge (labelled with x) from the vertex Vg to the vertex v gx • We do this for each g E G and x E X. how is one to find a complete set of relations? 'Complete' clearly means that no more are needed to get G. x.
6. One can 'see' that R3 = 1. To evaluate a word in Rand S. however.Cayley quivers x 119 Fig. consider SRRS1RSR What is it? Reading from left to right and starting at vertex 1. we need only follow the path through the quiver specified by that word. 6. S2 = I and RSRS = 1. For instance.3 The problem of drawing the Cayley quiver of a group G with respect to generating set X is that it rarely turns out to look as nice as those which have been given to you in a book! Knowing the Cayley quiver does. yield the relations. go along the S labelled edge if you meet an S and along the R labelled . Of more interest is D3 (or S3) generated by Rand S (Fig.3). 6. R R Fig.2 2.
then a and b are as shown. S2 = 1. and (RSi = 1. In particular draw that of Q4 which has order 8. (b) What will be the structure of the Cayley quiver of Dn = (x. .4. C 6 represented as 2. Using this idea gives the relations listed. We need more machinery before making it precise. t : xn = y2 = (xyf = 1) 3. and <clearly' all other loops at 1 can be considered as consequences of these relations. again with x and y as the generators? Investigate the Cayley quiver of Q2n with the presentation given earlier.) Compare this with the Cayley quiver of D 3 • (a) Draw up a Cayley quiver for D4 = (x. y: X4 = y2 = (xy)2 = 1) with x and y as the generators. So we end up at the vertex labelled RS. 6.resentations of groups L~~~~R~~~R Fig. R3 = 1. Draw up a Cayley quiver for C 2 x C 3 (== C 6 ) generated by a of order 2 and b of order 3.3 1. The word corresponds to the path as shown in bold in Fig. 5 6.4 one if you meet an R. We only intend this as a vague idea here. 6. (If you prefer.
We have assumed that the idea of a quiver is fairly intuitive in this section. The 'word problem' is of deciding when two words in the generators yield the same element of the group. Repeat with other Cayley quivers. Find the structure of Aut(r) when r is the Cayley quiver of D3 as above. the reader can consult Johnson (1990) ). (For a detailed treatment of the ToddCoxeter method of coset enumeration. but also yields a table which can help in drawing a Cayley quiver. and then definitions followed by more examples. Notwithstanding this comment about the difficulty of drawing Cayley quivers of infinite groups in general. Remarks 4. Which comes first. This has been proved in general to be unsolvable! If the group is known to be finite. This is annoying as most of the groups we will be using later on (knot groups and surface groups) are infinite groups. Cayley quivers will be useful later on both for giving examples of certain topological . qJ(e) will be an edge from qJ(a) to qJ(b). The construction is equivalent to solving the 'word problem' for that presentation. a rotation of the graph through 120° is an automorphism of r as is a transformation that interchanges the inner and outer triangles reversing directions. that is if e is an edge from a to b. 2. but the precise concept then brings to light other examples. it can be impossible to construct the Cayley quiver of the group. Unfortunately the ToddCoxeter algorithm used in this method will not tell you if a group is not finite. taking r to be the Cayley quiver of D3 with generators Rand S. You can compose automorphisms by 'do one and then the other' so that they form a group Aut(r). Try to prove what you have conjectured. 1. This should enable you to conjecture a result. so this method of obtaining a Cayley quiver cannot be extended to infinite groups. You do not think of a precise definition of a concept until some examples are available. then there is a method known as coset enumeration which not only gives the exact size of the group. examples or definitions? Historically it is usually the examples first. Given a presentation (X : R).Cayley quivers 121 An automorphism of a labelled quiver is a bijective mapping qJ which assigns vertices to vertices and edges to edges respecting labels (so that if edge e is labelled x then it can only be sent to another edge labelled x) and respecting incidence. we will take more care with the definitions. Before we investigate quivers in greater detail and in particular Cayley quivers. List all automorphisms of this quiver. For example.
Presentations of groups constructions and for 'spinoff" back in group theory of those same constructions. t/J'(x) = t/J(x). What are these 'relations'? We can think of them as equations satisfied by x and y: the use of the word 'equation' implies that two things are the same. we can easily talk about how to generate words from the 'letters' x. the free group on X. and then look for relations amongst these words. It is 'free' since no two of these powers are the same. any mapping t/J: X ~ H can be uniquely extended to a group homeomorphism ¢l: F ~ H (i. X. tion A group F is said to be free with basis X ~ F if for any group H. y and their inverses. it will be difficult to avoid circular arguments. both positive and negative. Free groups There is a problem of an almost philosophical nature that needs to be resolved when dealing with presentations. if x E X ~ F. so there are no constraints imposed on it. and generate new elements by 'formally' multiplying the elements of X together to get 'words' in the elements of X. However. For instance. if X = {x} then the free group on X must contain all powers of x. y E G. If we start with a group G and elements x. we will not know what to believe or what methods are valid. and that is all. Unless we have a precise meaning for 'word' and 'relation'.e. continuing until we have constructed a group F(X). namely that of a free group. we sometimes do not know anything about G except that x and yare in it and possibly that there are some relations. We will start with a formal definition in terms of the most powerful property that free groups have. The idea that solves the problem is a very elegant one. the diagram . We will return to the formalities of the construction later. Where are these two 'things' to live? The logical problem involved is important but there is also a practical side. We start with a set.
so cf1. We define the product of words by juxtaposition. Thus.. This is an infinite cyclic group. Hopefully you can guess at least some of the construction of F(X) in general. which has no syllables. We define a symbol an to be a syllable. This situation can be thought of as being a nonlinear analogue of the fact in linear algebra that in order to define a linear transformation T:V~V' between two vector spaces. We note (i) that multiplication of words is closed: if WI' W2 E W(X) then WI' W 2 E W(X). A word is a finite ordered sequence of syllables. Remarks 1.\ S2. are all syllables. if U E V. The set of all possible words on the alphabet X will be denoted W(X). the elements of X will be letters. the only possible value for T(v) is LujT(ei). and we can map X into W(X) via the mapping a ~ aI. If we suppose X consists of a single element x. There is one special word. U = Lief Utet for some unique set of coefficients {Ui: i E I} and as T has to be linear. for instance. specifying the images of the basis specifies a unique T defined on all of V. b. that is F({x}) = {x n : n E Z}. Since {ei} is a basis for V we have. etc.z E W(X) then we take (azbIe l ) • (b la 3 d. (ii) that multiplication is associative: if . alOl. 2. We will call the set X the alphabet. then ¢ : X ~ H is determined by the element ¢(x). it is sufficient to choose a basis {ej: i E /} for Vand to specify all the vectors. We denote it by 1. for example where a. q/ i = ¢ as mappings). e EX.z) = dblelbla3d2 E W(X). namely the empty word. Tei E V'. a.Free groups 123 commutes. for aZblel and b l a 3 d. We can see that it is free on X since the fact that ¢': F({x}) ~ Hhas to be a homeomorphism ensures that ¢'(x n ) = ¢'(xt and this is ¢(xt E H. We suggested that the free group on {x} consisted of all positive and negative (formal) powers of x.
The equivalence relation must clearly include the relations a l • a. It is clear that if a group is to be constructed from W(X) then a l . and (iii) that for any W(X). Lemma.wv. Its reverse. WUn completes the proof.) lfu. . The first step is to note that if u = U1 ~U2' V = UI U2 then uw and vw are linked by an elementary contraction of type I.. then we say that the change. We outline a proof. we will write u . We can thus use UI w. I. Un for some n.w = w. Finally we have that there are U2. 2.v then uw . We construct the equivalence relation in a series of steps. is an elementary expansion of type l. Unw to link uw and vw in a similar way.is an equivalence relation.I . Ifu. We then note the following. aI and I must be identified. we do not have that a l • aI is empty. it is worth repeating that 'it is clear' should be a signal for the reader to check that it is indeed clear and why. finition of the equivalence relation on W(X) 1.. that is we must form the quotient of W(X) by some equivalence relation WI' Wz. We therefore suggest that you sketch out a proof that . W3 WE so that the set of equivalence classes is a group. where WI> W2 E W(X) and a E X. acts as an identity. V = Un. w goes to w'. As a result of this lemma.1 = w. a. Ifw = wlaPa'lw20 w" = w l a P + Q w2 then we say that the change from w to w" is an elementary contraction of type II etc. (This just means we have introduced the usual laws of indices. We will denote by [v] the equivalence class containing v E W(X). It should be clear that . Even though a l and aI are in W(X). Since we have treated words as formal objects. w' = WI w 2. A repeat of this argument with WUj. we have that the operation [U][v] = [UV] ..vw and wu .4 Presentations of groups E W(X) then (WI W2)W3 = WI (W2W3). (Why?) The same holds for wu and wv. however. and each Ui linked to Ui + I by elementary operations.is an equivalence relation. v. v are in W(X). so the empty word. is an elementary contraction of type l. u = u l . ..1. The next point is to note a similar result for contractions of type II. W(X) with this multiplication is not a group. . Ifw = Wlaow2.v ifone can be obtainedfrom the other by a finite sequence of the above elementary operations. WE W(X) and u . . 3.I • a l 1 and must be compatible with multiplication. w' goes to w. again leaving it up to you to fill in the details. I.
This question can be tackled in two different ways. As they are again usual proof forms we expect you to complete the details. [u].. . Any word u E W(X) can be written as a finite string of syllables and hence can be expressed in the form u = af1af2 . An induction on n shows that tiT) u 1 and uli r )  1. (As usual we expect you to check that you can do this basic type of argument satisfactorily. (3) each element of F(X) has an inverse. Of course in general the list of letters aI' ... (2) u(vw) = (uv)w the operation is associative as. Un as before and prove d r ) and v(r) are linked by U)rJ. The other proof is neater but uses the above in a hidden form. that is by using results we have already noted. [v]. .. an may contain many repeats. they may. . This does not matter as will be clear.P E W(X) and [ap][a. if [u]. do not try to make it difficult!) Then assume u and v are linked by a finite sequence Uj. we have a. To see this. Define a 'reverse' of u to be that is the string in reverse order with the sign of each power changed. together with ] the identity property of [1]: .. etc.v so [u] = [v] and wish to compare [tin] and [v(rJ]. that is it does not depend on the representatives E [u] and v E [v]. a/:". We assume u . One might ask if [d rl] depended on the choice of u within the equivalence class.Free groups 125 u is well defined. the integers. The first involves checking that if u and v are linked by an elementary operation then tl r ) and v(r) are linked by the same sort of elementary operation. first note that for any syllable a P . in fact. If we write F(X) for the set of equivalence classes. [w] E F(X).. .. we can note the following properties of this operation: [1][u] (1) [I] acts as an identity element since lu = ul = u implies that = [u][l] = [u] for all u E W(X). that [u<rl][u] = [1] and [v][v(r 1 = [1]. . We use that multiplication is associative.. we have and so [u]([v][w]) = [u][vw] = [u(vw)] = [(uv)w] = . all be equal. where the ai E X and the Pi E lL. we hope. li:l. . and some of the Pi may be zero. .) Now we have [tlYl][u] = 1 and [uJ[d r )] =1 so [u] has an inverse.P] = [aJ>+(P)] = [a~ = [1]. . (The argument is quite simple.
sentations of groups
[lP)] = [lP)] [1] = [1fT)] ([V] [V(rJ]) = ([ drJ] [V J) [V(r)]
= ([dr)][u])[V<r)] =
[l][v(rj] = [V(r)]
Summing up, we see that F(X) is a group. Is it a free group on X? We can think of X as included in F(X) via the mapping a r? [d], so suppose we have a group H and a mapping t/J as in the definition of freeness. We have to prove that (i) t/J extends to give some homeomorphism ¢f: F(X) ~ Hand (ii) t/J' is unique with this property (i.e. if If/: F(X) ~ H is any homeomorphism satisfying If/(x) = t/J(x) for each x E X then If/ = t/J'). We first look at the existence of ¢f. We have to decide what t/J'[uJ is for an arbitrary [uJ E F(X). There is an obvious way of defining a mapping t/J1: W(X) ~ H which extends t/J, is completely determined by it, and is as near to being a homeomorphism as makes sense (remember W(X) is not a group). For instance, if a E X there is a syllable an E W(X). We clearly would like t/J1(a n) = t/J(at since this would have to be true if W(X) was a group and t/JI was a homeomorphism. An arbitrary word w E W(X) is a string of syllables
and if t/J1 was a homeomorphism we would get
which makes sense since it is an equation between elements of H. Since this is what we hope will happen, why not explore the consequences of taking it as a definition of t/JI? The set W(X) has a multiplication and it is clear that t/J1(W 1w 2) = t/J1(W I )t/JI(W2). At this point we must pausewhat is t/J1(l)? We have not dotted our i's and crossed our t's. The empty word, I, is in W(X) but is an empty string (so is as above with r = 0). What, therefore, is t/JI (l)? Clearly it must be the product of an empty family of elements of H, but the exact meaning of that is not clear. However, in H, the usual laws of indices etc. hold so we always have, for instance,
If we have a string of elements in H, and multiply it by the empty string (in H) then we must leave the original string unchanged, so the only sensible interpretation for t/J1 (1) is that it is the identity elemen t of H. So far we have used t/J to construct a nicely behaved t/JI: W(X) ~ H. Is t/JI compatible with the equivalence relation we have defined on W(X)?
Free groups
127
rna.
1. Ifw
2.
= Wlaow2 and w' = WI W2 then ~I(W) = ~I(W'). Ifw = w l a Pdlw2 and w' = w l a P + w 2 then ~I(W) = ~I(w').
Q
We leave the proof to you. It just uses the fact that the laws of indices hold in H. As an immediate consequence of this lemma, we get that if u, v are in W(X) and u  v then ~I(U) = ~I(V). This in turn implies that we can (without fear of getting into difficulties) define
~'[u]
=
~I (u).
This will not depend on which representing u we take for [u]. (It is well defined.) We have taken as operation on these equivalence classes
[u][v]
= [uv]
and we can check that ~'([u][v]) = ~([u])f([v]) for all [u], [v] E F(X) (including the class of the empty string, 1). This completes our checking that ~ does extend to a suitable ~I' It also shows that if w = a~l ... a:.'r then ~I[W] = (~(al)tl ... (~(ar»nr. How about uniqueness? Suppose 'II". F(X) ~ H extends ¢. This means that lfIIa] = ~(a). If [w] E F(X), then if w = ai'l ... a;'r we have [w] = [ad n1 • • • [arr r , and as I/f is a homeomorphism
I/f[W] = (l/f[al])n 1
•••
(I/f[ar])nr
=
~(al)nl ... ~(ar)nr
= ¢'[w]
by our calculation earlier. We have proved that ptX) is indeed free with X as basis. We note that we have never actually checked that the mapping a ~ [a] from X to F(X) was oneone but if [al] = [a21 for some ai' a2 E X with al a2> then we could pick H = C b the cyclic group of order 2, and a mapping ~: X ~ C 2 such that ~(al) ¢(a2)' Now to extend to the unique ~: F(X) ~ C 20 we must have ~'[al] = ~'[a21 (since [al] = [a21!) but ¢'[ad = ~al) and f[a21 = ~(a2)' .... We clearly have to abandon any idea that we can have both a l a 2 and [al] = [a21. If two groups FI and F2 are free with bases XI and X 2 respectively, then any functionf: XI ~ X 2 extends uniquely to a homeomorphism F(f): FI ~ F 2. We use the 'universal' defining property of freeness to check this. Suppose we write i l : XI ~ FI for the inclusion, ;1 (a) = [a] similarly for i l : X 2 ~ F 2 • The function f: XI ~ X 2 gives us a mapping ~ = i2f: XI ~ F2 which extends uniquely by the defining property to a homeomorphism ~I: FI ~ F2 which satisfies ~Iil = i2f. It is this ~I that we take as F(j). Thus if XI has the same cardinality as Xl> the
'*
'*
'*
Presentations of groups
corresponding free groups FI and F2 are isomorphic. (Check this.) In fact the converse holds: if F; is free on Xi for i = I, 2 and FI and F2 are isomorphic, then XI and X 2 have the same cardinality. This cardinality will be called the rank of the free group. We refer the interested reader to Johnson (1990) where the theory of free groups is explored in more detail. We need free groups primarily for their use within the theory of group presentations. We introduced them earlier as a formal solution to the problem of where the words, that were the 'relations' in a presentation, could live. In detail the way they are used is a result of the following proposition.
oposition (a) (b)
A set X ~ G generates G if and only if the homeomorphism from F(X) to G, associated to the inclusion, is an epimorphism. Every group is a homeomorphic image of a free group.
Proof. First a notational solution to our earlier 'philosophical' problem: if x E X, we will write x if we are thinking of it as a generator of F(X) and x, without a bar if as an element of G. This is necessary since we have, in F(X), elements such as [xlJ[xzl and if. as might happen, X I X2 E X, we have the possible confusion between [xd[xzl and [X I X2], which are different elements of F(X). (With this notation we have [.~d['~2] = [xlxzl whilst the other element is [X IX2'] The homeomorphism qJ associated to the inclusion of X into G is then given by removing the bar above each generator and each square bracket; for instance,
[.XIX21'~3Xzl E
F(X)
will be sent to XIX; I X 3 X 2 E G. Now assume X generates G. Recall that formally this means that any element of G can be written as a product of elements of X u XI where XI = {XI: x E X} is the set of inverses of elements in X. Thus given any element g E G, we find an expression of the form
where t; = ± 1 and all
Xi E
W
X. The word
E
= [x le, ... ."'n"n]
F(X)
is sent to g by qJ, so qJ is onto. The converse is simple. To see (b) we just have to note that any group has a set of generators, for instance one can take X to be the set G itself. _
Prese ntations
129
entations
Given a set of generators X for a group G we obtain the corresponding epimorphism
q>:F(X)~G.
We can now give a precise sense to the term 'relation' and in particular 'defining relation'. The above epimorphism gives us an isomorphism between the quotient group F(X)/Ker q> and G (remember the first isomorphism theorem of group theory), so if we knew how to construct this normal subgroup Ker q>, we could construct G from our set of generators which is what we hope to be able to do. Any word w in Ker q> is, of course, 'killed off' by q>. Let us examine this in our example of D3 with X = {R, S}. (For the moment we will keep our cumbersome notation R, 5 etc. to denote the elements of X when they are thought of as basic elements of F(X), but be warned that we shall soon drop the use of the bar and you will have to do a bit of 'doublethink', when looking at presentations.) In F(X) we have, for example, the word
w
=
5 1 RR5R 1
which is 'reduced', that is it cannot be made shorter using elementary contractions or expansions. This word w is in the kernel of q>, however, since in D3 we have (so SI = S) (so R 1 = R2) and (RS)2 = 1 (so RS = SI R 1 = SR2 ). Hence
q>(w) = SI R 2 SR 1
= SR2 SR2 = RSRS
= 1,
as stated. Thus our word w expresses a relationship between the chosen generators of D 3 ; it is a relation. Given a set of generators X of a group G with q>: F(X) ~ G the associated epimorphism, an element w E Kerq> is called a relation (between the given generators) and Ker q> the relation subgroup. Except when it is trivial, Ker q> will always be an infinite group, so it is clearly out of the question to specify all relations; however, we could specify a generating set of relations. Let us return to our example of D3 with generators Rand S. We have R 3 , S2 and (RS)2 are relations; do they generate the relation subgroup? Suppose we
sentations of groups
conjugate R3 by S to get SR 3SI; this is in Ker q> but it is not clear whether or not it can be written as a product of R 3 , S2 and (RS)2. How are we to tell? As usual the answer iswith difficulty. You might try all sorts of combinations of copies of R 3, S2 and (RS)2 in various orders for several days, and yet not hit on one which gave the result, but clearly that will not do as a proof that it cannot be done. The fact is that Ker q> in this case is a free group of rank 7 (why this is true will be explained later) and because of this it cannot be generated by only three elements (again this must wait for justification later). However, it is generated by these three elements together with their conjugates. This shows a point of great importance: we need only specify a set of 'normal generators' for Ker q> as it is a normal subgroup of F(X). We need to examine this in slightly more detail. First some observations. 1. Given any group G and a collection of subgroups SA c G, A E A, some indexing set, then n {SA: A E A} is also a subgroup of G. (Check this for yourselves if you have not met the result before.)
If Xc G and X generates a subgroup, H, say, of G (i.e. the elements of H are exactly the products of elements from X and XI) then
2.
H = n {S: S a subgroup of G, X
~
S}.
IS
3. 4.
If all the subgroups SA in observation 1 are normal in G, so their intersection. If X c G, then
H = n{S: S a normal subgroup of G, X
~
S}
is the smallest normal subgroup of G that contains X. (It is often called the normal closure of X in G.)
If we denote by XI the set of inverses of elements of X, so XI = {XI: x E X}, then the subgroup generated by X, denoted (X), consists of all products of elements from X and XI. It is the image of the natural homeomorphism from F(X) to G corresponding to the
inclusion of X into G. Can we give a similar description of the normal closure of X in G? Let us denote this by Clearly since is normal in G, it must contain not only each x E X but all the conjugates gxg I of x by elements of G, together with all the inverses and products of all these elements. If g E G, x E X, we will write g x for gxg I. A consequence of X will be an element of G that can be written as a product of elements of the form gx and their inverses, so y is a consequence of X if and only if
«X».
«X»
This 'playing around' often involves substituting some consequence of one relation. XI . R = {X3. so we do not need z as a generatorit is already there in terms of x and y. We write G (b) = (X: R). A presentation of a group G consists of a set X and a subset R of F( X) so that F(X)/«R» is isomorphic to G. (Prove it!) In fact the normal subgroup is the normal closure of X in G. xy 1) or possibly.. We can substitute xy for every occurrence of z in other relations and delete z from the list of generators without changing the group. y. .) n (a) Given a group G and a set X of generators. . (x. using that (xy){xy) = 1 is the same as xy = y I XI. For example. y}. y : X3 = y2 = (xy)2 = = 1.gn E G.V = yIXI).Tietze transformations 131 y = {gIXI)£I{g2X2)£2 . This sort of process is so useful that it will be used time and time again later on in the discussion of presentations that arise from knots. . . Usually we write things slightly less formally and it is quite usual to write alternative versions such as (x. xyzJ = 1. Because of this. y: x 3 = . 5 Let X = {x. a subset R of F{X) is called a set of defining relations for G if «R» = Ker (cp: F(X) ~ G). The point is that now we have set up the formal structure of presentations.. then (X: R) is the presentation of D3 that we have been looking to for 'guidance'. z : x 3 = y2 = 1.. (Again prove it. y. (xyi}. . The consequences of X in G form a normal subgroup. (gnxn)£n for some gl' . ze transformations As usual G = (X: R). The third relation tells us that z = xy. The following transformations do not change the group G: . Z2 = 1). consider the presentation (x. Xn E X and where each Ej is ± 1. it is important to have a list of acceptable 'transformations' of presentations that do not change the group.. we can play around with them with more confidence.
that is w is in the image of the inclusion of F(X) into F(X'). y : xl (ab)l = 1). (X: R) becomes (X: R') where R' T2: Removing a superfluous relation = R\{r} and r is a consequence of R'. Tietze.1 = 1). b : xl a and b: = ym = (xyr = (xya. b : an = bm = Formally. These four operations are called Tietze transformations. that is a consequence of R. T3: Adding a superfluous generator (X: R) becomes (X' : R') where X' = Xu {g} (g is a new symbol not in X) and R' = R u {wgl} where w is a word in the other generators. start: (x. who did pioneering work on group presentations (see Tietze 1908. with WE F(X') and wg. = y1b. We normally carry out more than one change at a time and it is advisable to work things out informally first before launching into formal manipulations. after H. Magnus and Chandler 1982). y. y : xl = ym = (xyt = 1). T4: Removing a superfluous generator (X: R) becomes (X' : R') where X' = X\{g} and R' = R\{wgl}.sentations of groups Tl: Adding a superfluous relation (X: R) becomes (X: R') where R' = R u {r} and r E «R». Apply T3 twice to introduce (x.l E R and no other member of R' involves g.1) Apply Tl three times to introduce superfluous relations: . To show (x. a. = ym = (xyy = 1) is isomorphic to (a.
Tietze transformations 133 Really we should justify that an = 1.5.5 A circuit around the inner hexagon gives (xy)3 whilst the threesided faces all have circuit xya~l. Note that a circuit around the outside of the diagram gives one a 3 • The 'secret' is to break up this outer circuit into a 'sum' of the various faces it 'encircles'. b: xya~l = y~Jb~1 = an = b m = (ab)' = 1). the appropriate Van Kampen diagram is as shown in Fig. Now we have . similarly for b m and (ab)'. y~ J b~ J. This step is often not easy to do. 6. Here we can write n~l an = (xya~l)~l II i= I (xy). b~la~lx Next we use y~Jb~J and xya~1 or rather its conjugate ya~lx to deduce and hence x~Jab and its conjugate abx~l. For n = 3. 6. so three uses of T2 reduce our presentation to (x.(xya~l)~l(xy)~i(xy)n so an is a product of conjugates of the other relations. ym. and (xy)n are superfluous relations. b m and (ab)' as 'basic relations' now and prove that x'. Returning to the presentation we can treat xya~ J. y. a a a Fig. a. A full account of the theory underlying the use of these diagrams may be found in Johnson (1990). b m = 1 and (ab)' = 1 are consequences of the other relations. The formula above may be represented in a Van Kampen diagram. an.
b~la = b). b : an = b m = (ab)' = 1). We can try to build a Van Kampen diagram by sticking copies of the triangles (Fig.6) or merely by 'playing' around with the relations to see what results: ab = e ~mpl~es b = aIe } ed= a Impbes d= e~la = b. Transforming the symbol corresponds to applying a Tietze transformation to the presentation. da = b). so with abx~l we can apply T4 twice to get the desired presentation (a. the formal method is there to act as final arbiter. What is important is to realize that the formality is possible and if any step in the informal sequence of transformations is doubtful. We try this out to see if it works. it seems a good idea to pick b so that ab = x. and formally these processes are the same. but this means b = yI. 6. cd = a. This formal method is usually very lengthy. which we will meet in detail in Chapter 11. In fact the transformation of surface symbols by operations is completely analogous to the Tietze transformation process. Working the other way: given (a. The obvious choice is a = xy. This process of transforming a set of relations to get a simpler equivalent one is very similar to the process we have seen in use on surface symbols. Trying x = ab and y = b~ I clearly works. Next. b. The intuitive resemblance can be made rigorous via the fundamental group of the surface.I => d = b~1 We can.sentations of groups The relation xya~l is superfluous and can be deleted. since we want (ab)' = 1. b. we start with {x. therefore. but b m = y~ m = 1. d: ab = e. we must define x and y in terms of a and b. The only relation left to be checked (as we have used two already) is b m = 1. ylb~1 gives us b1yl. be = d. reduce the presentation to the following: (a. then an = (xy)n = l. e. b : an = b m = (ab)' = 1). e: ab = e. has generators the edge letters and a single relation A = 1 where A is the symbol. y : xl = ym = (xyt = 1). be = b~l. and hence it is usual that people use the informal 'substitution' method described earlier. We try to define a and b in terms of x and y to obtain an = bm = (ab)' = 1. . This group. Consider (a. Sometimes the result of Tietze transformations can be quite surprising. Informally.
Elimination of superfluous generators then gives that the group is isomorphic to C s. Natural curiosity should make us ask if there is a notion of mapping of presentation that will reflect information about homeomorphisms. one can be obtained from the other by a finite sequence of Tietze transformations. rem.H . so ab = c implies b 3 = b. We begin with some discussion. a = b2. Presentations are a combinatorial way of coding up information on groups.Tietze transformations 135 6666 a b c d Fig.2 and b S = 1.6 The second and third of these yield c = b. We note the following theorem. F(X) F( Y) so that in the Prove there is a homeomorphism diagram f: (*) G e . Proof. When we study groups.2 . The success of Tietze transformations is complete. yielding a homeomorphism qJQ: F( Y) ~ H with «S» ~ = Ker ~ qJQ. Given two finite presentations of the same group. (3) a homeomorphism 8: G H. 6. Suppose we have (1) a presentation P = (X: R) of a group G. we constantly use homeomorphisms as a way of comparing them. (2) a presentation Q = (Y : S) of a group H. proved by Tietze (1908). This proof is developed in a structured sequence of exercises. yielding a homeomorphism qJp: F(X) ~ G with «R» = Ker qJp.
ses 6. Suppose P = (X: R) and Q = (Y: S) where. It is the unique homeomorphism making the diagram 1. we should have learnt the reflex of asking if it is unique with the required property. say.:. of F( Y). How close to unique is f for a given 8? If. x E X. x E Y.iJ and h? A mapping f: (X: R) ~ (Y : S) of presentations is a homeomorphism f: F(X) ~ F( Y) such that for each r E R. you will have proved that given a homeomorphism 8: G ~ H and presentations P of G.) If you want a hint. f: P ~ Q. how can we compare. of F(X) to the corresponding generator. F(Y) 1 commute. f(r) E «S». What does f do to elements in K! (Remember «R» = Ker qJ. with f: P ~ Q given by f = id: F(X) ~ F(X). a variant of the question is to ask how close to unique it is.. S. (The square 'commutes'. (This homeomorphism maps the generator.) Whenever we show that something exists in mathematics. and set X = Y\{ y}. Show that f is a presentation mapping. y a given element of Y.) Describe the corresponding group homeomorphism. If it is not unique. The conditions we have imposed on S make f: P ~ Q into a presentation mapping. feR) c. What is the corresponding homeomorphism 8 from G to H? When is 8 an isomorphism? 2. 8. think of the universal property of free groups. Q of H. which mirrors the homeomorphism 8 in as much as the square (*) above commutes (qJQf = 8qJp). then there is a presentation mapping. If you have completed the above exercise. .iJ and h both fit the bill. Inclusion • x~ Y F(X) ~ 1 f . Let Y be a set.Presentations of groups we have that the two composites 8qJp and qJQfare equal.6 Let P = (X: R) be a presentation of G and Q = (X: R u S) a presentation of H. if we write f: F(X) ~ F( Y) for the homeomorphism induced by the inclusion of X into Y.
Try to describe a set of normal generators for Ker rpe. Let P = (X: R) be a presentation of a group G and let rp: F(X)~G be the epimorphism corresponding to the presentation.6. if x E Xc Y Remember that it is only necessary to say where the generators of a free group go to determine where everything else goes. Suppose Z is a finite set disjoint from X (X n Z = 0) and pick a homeomorphism e: F(Xu Z)~ F(X) so that e(x) = x for all x E X.3 we get F(Y) ( ~ F(Xu Y) . ReKer rpO. Now prove that g is a presentation mapping from Q to P and that the corresponding group homeomorphism is an inverse for e.6. As we want all of the Ygenerators in. so do not use those in your proof. but what other elemen ts do we need? The case when Z = {y} is very close to the previous example/exercise. We have to find a sequence of Tietze transformations giving f. That much is clear.) 3. so is a good place to start. We have to show that we can get from P to Q using only Tietze transformations.Tietze transformations 137 Now suppose S = feR) u {yw~ I} where w is some word in the image of f. that is they are disjoint. The plan of action is to assume first of all that X and Y do not overlap. At this stage we have a presentation of the form (X u Y: ?). try the following hint: define g: F( Y) ~ F(X) by g(x) = x g(y) = w. Suppose we have two finite presentations P = (X: R) Q = (Y: S) of a group G. (This is of course the basis for the type T3 and T4 Tietze transformations. we can do this one by one using Tietze transformations. Prove that the homeomorphism e is an isomorphism. but how can we do it in detail? Using Exercise 6. First we lift the identity map on G to give presentation mappings f: P ~ Q and g: Q ~ P. If you find it hard to get started. again using Exercise 6. Similarly. We then reverse the process and go to (Y: S) by getting rid of the Xgenerators. Now we are ready to approach the task of proving Tietze's theorem. 4.3 above we can look at F(X) ( ) F(Xu Y) and describe a presentation of G of the form (X u Y: R u A) say. Of course.
the normal closures of R u A and SuB are the same. Choose it carefully. Yand Yj are in oneone correspondence. If we are unfortunate and X I l Y"# 0. in some cunning way)! You now have the building blocks and the plan. that is. and Xl I l Y 1 = 0. X I l Xl = 0. We now look at the obvious presentation mappings: (X u Y: R u A) ( ) (X u Y: R u SuA u B) ( ) (X u U: SuB). . say. otient groups Suppose we have G = (X : R) and we know H is a quotient group of G. We have formed G as a quotient F(X)/«R» and so have an epimorphism F(X) <p~ G. induces a presentation mapping (X: R) ~ (XI: R j ) and so on. with the assumption of X I l Y = 0. of F(X). The normal subgroup N <J G corresponds to some normal subgroup M.esentations of groups and a presentation (X u Y: SuB). Proofs of the theorem can be found in Crowell and Fox (1967) and Johnson (1990). Now the isomorphism F(X) ==~ F(Xj). You can now look at the various mappings to see if you can get f as the composite (by choosing the 'retractions' from F(X u Y) to F(X) etc. so it is reasonable to expect that H has a presentation of the form (X: R u S) where rp(S) normally generates N in G. we replace X by Xl> Y by Yj so that X and Xl are in oneone correspondence. and show that (i) they induce isomorphisms. We leave the proof as an exercise as it is quite simple. the kernel of the composite epimorphism F(X)~G ~G/N==H so H == F(X)/M. The above argument can be made rigorous so as to prove the following theorem. Is there a presentation of H easily derived from the given one for G? The first point to make is to note why this might be expected. and (ii) they can be realized by Tietze transformations (of types Tl and T2). say H == G! N. Y I l Yj = 0. Write out a proof up to this point. We know that «R» eM. Can you complete the proof of Tietze's theorem? You may need to introduce some notation for the various mappings. Clearly (but check it).
so it is clear that we can find an epimorphism from G to S3 if n is divisible by 3. G) of G. g. but we leave this to you to check. g' E G. rcise 6.) For which values of n does there exist an epimorphism from G to S3? The group S3 has a presentation (x. y : Xl = y3 = (xy)2 = 1). x'] : x. (X: R u {[x. c) = a[b. As an example.I [a.for rp(a) and qJ(b). We also have a commutative diagram . We also saw that [ab. The method is to assume the existence of an epimorphism rp: G ~ S3 and then to see what possibilities there are . Abelianization We saw earlier in this chapter that there was an epimorphism from any group G to its Abelianization Gab' Recall that Gab was the quotient of G by the commutator subgroup [G. (This is the knot group of a (2. In fact adding the relations a 2 = 1. br = [b. These facts together with Von Dyck's theorem (6. x' We know that the kernel of rp: G ~ Gab is the commutator subgroup and is thus generated by all [g.7 Prove Von Dyck's theorem. b 3 = 1 and (ab)2 = 1 into the relations for the given presentation of G yields a presentation equivalent to that we have given for S3' This condition 31n is also necessary. a). consider a group G with presentation (a. c] and [a.Abelianization 139 Von Dyck's theorem.5) enable us to prove the following theorem. n)torus knot. The kernel of 8 is just the normal closure of S\R as a subset of (X: R). If Rand S are subsets of the free group F on a set X and R ~ S then there is an epimorphism 8: (X: R)~(X: S). c)a. Theorem l If G has a presentation (X: R) then Gab E has a presentation X}). g'). b : a l = bn ) for n odd and greater than 2.
if not write w' = w~y (or w' = w~yI) and use the expression for [a. Then w = WI X or w = WI x. the kernel of qJ is generated by the images of all [w. What we have to do is to show that the smaller set of relations R u {[x. We have not given all the cases: what about WIX. x']. 11/] = [WI x. x'] : x.1 and some x EX. this ends the inductive step. n .(In fact R is not needed here as we really only use [X. X] has the same normal closure as R u [F(X).how can we get around this? We leave it to you to finish this proof. w] or [w. n.I ? This gives us a term [xI.. w1.I ) to reduce one level in the first variable. We used it earlier to look at [Dm Dn] and thus at (Dn)ab' This tool becomes even more powerful once we have the results on finitely generated Abelian groups that follow. Now use [w. and we look at [w. w. w. Clearly R u [X.1. w'] can be written as a product of conjugates of the various [x. w1 = WI [x.) As the proof is an induction on the lengths of wand w'. F(X)].) ifit has a presentation of the form (X: R) with X a finite set. w1 wI I • [WI' w1 (similarly for w[x. and so in particular if G is finitely generated then so is Gab. so we could take 8 = R u [F(X).tion R u [X. F(X)] and so all we need to do is to show any [w. F(X)]. If G is f. If t'(w') ". w' E F[X] (since any commutator in G is the image of a commutator in F(X». t'(w') ". so is any quotient group of G. Most of the groups that we will meet are finitely generated. w1 where t'(w)..g.. not one of the form [x'.I for some WI of length n . x'] .g. we will leave the formalities to you and merely sketch the inductive step. We assume t'(w) = n. X]. Finitely generated Abelian groups are easy to classify up to isomorphism.g..esentations of groups and is a subgroup of «8». if not we can go to the next step.1.. n . X] c R u [F(X).. itely generated Abelian groups A group G is said to be finitely generated (f. This result gives us a powerful tool for calculating Abelianizations. w' E F(X) can be written as a product of conjugates of elements from R u [X. groups is a useful way of checking for nonisomorphism. be] to simplify. Moreover. Suppose we know [w. provided t'(w).. and so the comparison of Abelianizations of f. w1. that is if G and H . x' E «R» X} not. t'(w') ".. w']. X1.
. but if n is even nx = 0 is implied by 2x = 0. We have 2y = 0 and 2x + 2y = 0. However. if n is odd we obtain x = 0. 1 !!S. It is now simple to see what (Dn)ab is.Aj = '''J O. a2 E A 2 }. It is easy to extend this to any finite family of subgroups AI> .. [x. Often manipulation of the (additive) relations using common sense suffices to find the Abelian group up to isomorphism. so (Dn)ab == C 2 generated by y. the theory of additive presentations is quite easy.. (Proofs will be omitted as they are readily available in many group theory texts. since the group is Abelian. . Ai n L .. . + As. then G and H are not isomorphic. .. as nx = 0. and so when considered as a group C 2 has presentation (x: x 2 ) but traditionally when considered as an Abelian group one writes (x : 2x = 0).. y : xn y : xn = y2 = (xyf = I) = y2 = (xy)2 = [x. AI + A2 is called the sum of AI and A 2 • You should check to your satisfaction that this is a subgroup. so we have C 2 ffi C 2 in this case (ffi indicates direct sum.g. If AI and A2 are subgroups of an Abelian group A.) We start by making precise what we mean by direct sum. . (We will not need to extend it to infinite families here. y] = 1) as a group but as an Abelian group we might replace xn = I by nx = 0 y2 = 1 by 2y =0 (xy)2 = I by 2x + 2y = 0 and of course. then we can form a new subgroup AI + A2 = {a l + a2 : al E AJ. A more complicated example is Dn (Dn)ab = (x.Finitely generated Abelian groups 141 are f. groups and we can show that Gab and Hab are not isomorphic. As such that (1) A = AI + .) We say A has a direct sum decomposition if there are subgroups AI.y] by no relation. (2) for each i. so it pays to have the additional tools provided by the method at your disposal. An of A.. ffi A~. = (x.: s. see below). so 2x = 0. (In (2) we have written Li.: i !!S. Presentations of Abelian groups are often written additively.<j Aj for the sum of all the subgroups except Ad We write A = AI ffi A2 ffi ..
One can interchange the two forms of direct sum as they are really equivalent notions.Presentations of groups For example. Note that A1 u A2 is not a subgroup as it is {O. b): b E E A} B} isomorphic to A. 1. if m. b) : a E A. {O. we can form A EB B = {(a. . 3} which is isomorphic to C 2 (i. One usually drops the 'internal' and 'external' as these are clear from the context. I. These are uniquely . 4} and so is not closed under addition (mod 6). where (i) Bi is a nontrivial cyclic group of prime power order pfi for and (ii) Bio for The integer t is called the torsion free rank of A and the prime powers pfi are called the primary invariants of A. i = s + I. Let AI = {O. 4} which is a subgroup isomorphic to C 3 = {O.I} with addition mod 2). 2} (addition being mod 3) and let A2 = {O. look at C 6 = {O. b E B} with addition performed componentwise. then A has a direct sum decomposition i = I. 0): a B' = {(O. 3. We leave it as an exercise. is an infinite cyclic group. 'C6 == C 2 EB C/ and 'C6 is the direct sum of subgroups {O. s + t. . 2. 2.. that is C 6 = A1 + A 2. This decomposition as a direct sum of simpler parts is the idea behind the two main theorems on finitely generated Abelian groups. This group A EB B has a direct sum decomposition involving its subgroups: A' = {(a. imary decomposition theorem. Let A be a finitely generated Abelian group.2. 4}'.. The proof uses the fact that there are a.e. for instance. if the family consists of two Abelian groups A and B. For instance.4. isomorphic to B. . This shows that C 6 is an internal direct sum of a copy of C 2 and one of C 3 • It is also useful to be able to construct external direct sums starting with a family of groups and forming a big group containing copies of the members of the family. . 5} with operation 'addition modulo 6'... Check these statements yourselfdo not just take our word for them! We claim C 6 == Al EB A 2 • Clearly A1 () A2 = {O} so (2) is satisfied and to prove (l).s (the Pi are not necessarily distinct). In fact. One writes. b E ~ such that I = am + bn. 2. involves the extra hard calculation: 5 = 2 + 3 mod 6 and I = 4 + 3 mod 6. 3} and {O.3. n are coprime then C mn == C m EB Cn.
A = (a. for i = 1..4b .Finitely generated Abelian groups 143 determined up to order by A and themselves determine A up to isomorphism. then A has a direct sum decomposition A = AI (£I A2 (£I . c. . r + t. . . The relations are thus specified by a matrix and Tietze transformations correspond to integer row and column operations on this matrix.. r. More useful to us is an example showing how to derive torsion or primary invariants and the torsion free rank of an Abelian group A given by an (additive) presentation. . Inr (the vertical line is 'divides'). (ii) Ai is an infinite cyclic group for i = r + 1.. d: 2a + 5b + 8c + 7d = 5a . . (£I Ar+t where (i) Ai is a nontrivial finite cyclic group of order ni. In this example the h. . n" the torsion invariants. These two results are usually illustrated by lists of isomorphism classes of Abelian groups of some given finite order. The aim of the transformations is to diagonalize this matrix. are uniquely determined by A and determine A up to isomorphism... Xn :}. Let A be a finitely generated Abelian group. .. = fs = 0) where Jj = Lf~ I rijxi is some linear expression with integer coefficients.. ...7c + 4d = 8a + 5b + 2c + 7d= 0) has matrix The resulting diagonal matrix must satisfy n 1 I n21 n3 ... and (iii) nlln21 . = .. A general presentation of A will have the form (x]. The integer t is the torsion free rank (as in the previous theorem) and n" . . (£I Ar (£I Ar+ I (£I . .c. Suppose A has torsion free rank t and torsion invariants n].f is 1: . then it has a presentation (remember additive presentation).. so the method must aim to get the highest common factor of the entries of R into the top left corner. .. n r .. b. . .. . c decomposition theorem.
one applies the row operations used in reverse order to a unit matrix of size s x s where s is the number of original generators (in the example s = 4). b'. If one wants to find a'. c'. the setwise union C 2 u C 3 was not a group. c'.14 33 105 llj 30 90 C2 + 14Cl C3+11Cl ) [I 0 0 o 33 105 30 90 . b'.resentations of groups 5 4 7 2 _ _ _~) 4 7 ~ ~~ =~~l j [ 8 I ~ I ~ . note that 3 is prime whilst 12 = 22 x 3.14 .1 to get [~ ~~ ~n 1 [~ ~ o R4.f.) w ushouts of groups There is a configuration that will arise many times in various forms in the later parts of this book.n )U~ o R3~ R4 o J followed by R4 . The idea is simple. We saw that even in as simple a case as C 2 and C 3 considered as subgroups of C6 . 3a' = 0. we will do a row interchange and multiply by .llj 5 8 7 II 2 5 .3R3  . so C l2 == C 4 E9 C 3 and the primary decomposition is A == C 4 E9 C 3 E9 C 3 E9 C (Readers who have not studied this theory in detail can find treatments in Johnson (1990) or Rotman (1984).3R3 _I .36 12~ J o U o o 3 120 3~ J o o 3 36 12 [1 ) o o o 3 0 0 0 So A == C 3 E9 C l2 E9 have a presentation C~ SInce In the new generators a'. and d' in terms of the original generators. . b'. d' : a' = 0. We have two groups G and H and want to form a group that contains both G and H. 12c' = 0). As this is 3. of the remaining submatrix. In that . We leave you to do this in the example. OJ 3 6 0 3 6 Next look for the h. To find the primary invariants.I ~ ~R2  7 11 2 5 j Rl ) ( 8 I . d' we (a'. c'.c.
g: Y ~ Z two functions that agree on X n Y. Here we could clearly form words in the elements of G u H (as sets) and then 'remember' the multiplication in G and in H. if the two composites /3' a and f3d are equal) and it has the universal property that.) . The role of X n Y is important here. Suppose X. Y. if we take the words glhg2 and g3h' g4h" and 'multiply' them together we get (gl hg2)(g3 h' g4 h") = gl h(g2g3)h' g4h". Suppose we have a diagram a' of sets and functions.Pushouts of groups 145 example we had a bigger group in which we could work. how should we multiply g E G and hE H? We have seen something like this before. g: Y ~ Z satisfying fa = g/3. (The uniqueness of h with this property is very important and should be thought of in the example we looked at where W = X n Yand a and /3 were inclusions. It is called a pushout diagram (of sets) if it is commutative (i. This is vague. and Z are sets with f: X ~ Z. given functions f: X ~ Z. if / and g did not agree on X n Y then no joint extension h could exist. and if y E Y.e. put the words next to each other (remove the brackets) and multiply the end elements together if this is possible. We can try this out: for instance. we formed 'words' in the elements of the set of generators. This is so obvious as to be almost unnecessary to state. When developing the ideas that led to free groups. that is. We first note a distinctive property of the union of two sets that is linked to the gluing lemma of Chapter 3. there is a unique function h: T ~ Z such that h[3' = / and ha' = g. we will take a short cut. then there is a unique function h: Xu Y ~ Z such that if x EX. hey) = g(y). that is if x E X n Y then lex) = g(x). hex) =/(x). but as we have a limited need for these ideas and we have the machinery of presentations available to us. It can be made more exact. Note in general that a and /3 are not assumed to be inclusions. This sort of situation arises in numerous contexts and is abstracted in the notion of a pushout diagram. Now if we are given G and H.
We could also use another analogous situation with all the sets replaced by topological spaces and all the functions by continuous maps. {3. 2). we write X 11 Y for this disjoint union. a'. T. (You form T by taking a disjoint union of X and Y and dividing out by a certain equivalence relation. and g to be homeomorphisms. The disjoint union of two sets can be obtained by taking two new sets X" Y 2 consisting of pairs (x. 1). oposition. IfG=(X: R). {3'. W. y E Y and then forming Xl u Y 2 • Even if X n Y is nonempty. If we are given a 'top lefthand corner' of groups K _ _=:a"_J~~ G ~j H can we complete it to a pushout diagram K _ _=:a'_J~~ G ~j H jw a' ~ T of groups? The case K = {I} is the simplest case and is that with which we started this section. x E X. or (y. f. so X 1 and Y 2 are disjoint copies of X and Y. H=(Y: S) then the group T=(Xu Y: RuS) with the obvious inclusions forms a pushout diagram . /3: W ~ Y there is a pushout diagram as above containing them. Y. Z to be groups and a.) We are not really going to use push outs of sets but a similar notion can be obtained for groups by asking X. Xl n Y 2 is empty.Presentations of groups We leave it to you to show that given any pair of functions a: W ~ X.
{a(z) {3(z). similarly if WE S.: G ~ Z Z.: G ~ Z.) • This group T will usually be written G * H and it is called the free product of G and H. . {3'(x) = x (with some abuse of notation here!). Ifin a corner H ~ K ~ G. The general case is only slightly more difficult to describe. but if w E R u S then W E R or W E S.. then G * H is infinite unless one of G or H is the trivial group. H = (Y: S) and K is generated by a set Z then T = (X u Y: R u S. G ~ fJ a ~I H lw . then f(a) = }. Proof. and if a E Y then f(a) = Ji(a).. We note that a'(y) = y. G lw . but not both. We define f: T ~ Z by defining f on generators: if a E Xu Y then a E X or a E Y but not in both.(a).. with the obvious homeomorphisms.(w) = 1.Push outs of groups {1} 147 a ~l H . (We will look at the uniqueness of T in the general case later. We have to check that if w E R u S then f(w) = I.. Of course the commutativity of the square is automatic. Proof H ~ position. Even if G and H are finite groups. h: Suppose we have a group Z and homeomorphisms}. T a' and T is uniquely determined up to isomorphism by this property. forms a pushout square K _ _=:a _ _. Again . h: H ~ Z such that a = hf3 and define f: T ~ Z by the same method as before. T a' and T is uniquely determined up to isomorphism by this property.I : z E Z}). if a E X. Firstly notice that the addition of a(z) {3(Z)1 = I for all z E Z into the relations guarantees that the square commutes. G = (X: R).r. so we can check if W ERe F(X) then W only involves the elements of X and f(w) = }. Again we assume that given }. since {I} only contains the identity element.
h = f 15. a' = 15.)~j H .sentations of groups we have to check that f of any relation is trivial but if lV E R u S we can use the argument we used before and if w is of the form a(z)f3(zr l . If T is a pushout as claimed in the proposition then T == T j • To see this. G K a jw o ~ . we look at the two diagrams the H ~ Z K a (. with h: same universal property: given any j. (We have systematically abused notation in the above. taking j. which strictly speaking needs to be done..~ T..~ G ~j ~ jy H 1.. T. = /3'. ..: G ~ Z and such that fi a =hf3 then there is a unique f: TI ~ Z satisfying fi = fy. Firstly we should explain what meant here. The idea of the proof is simple but to make the distinction between words in F(X) and F(X u Y) say. would seem to complicate and obscure that simple idea.. In (i). Suppose someone presents us with another square IS KJ. then few) =fa(z)ff3(zr l =j.) We now turn to uniqueness. (U) ~j a' • G jy T .... In (ii).T ~ H~ .h unique gives us Now we note that = a' IfFY= /3' and IfFtjJa' = lfFa = a whilst .a(z)hf3(zr I = 1. takingfi a unique IfF: T j ~ T satisfying IfFY= /3' and IfFI5= IfFtjJ/3' = y: G = = ~ TI> h = 15: H ~ T J gives us a a.. tjJ: T ~ T j satisfying 13' y.
A similar argument shows that tjJlfF is the identity on T 1. G . the identity map on T and IfFtjJ. we must avoid our earlier deliberate abuse and simplification of notation.) To complete the proof we have to return to our original T and f:T ~ Z. is taken to be 13' and h to be a' gives a unique map T ~ T satisfying fa' = a'. so we call x to X by the name f3' (x) if we are considering it as an element of Xu Y. ff3' = 13'. (We have given this argument in quite a lot of detail as it illustrates the power of definitions involving universal properties. then the group T is called the free product of G and H with K as amalgamated subgroup.. Its construction can be thought of as being obtained from G * H by gluing together the two copies of K. We wanted previously to think of F(X) as sitting inside F(X u Y) and we used the letter x equally for a generator in X or in the Xpart of Xu Y. Why is f unique with the property that ff3' = J. hence tjJ and IfF are isomorphisms. j~' where J. f: . so ha'(y) = h(y) whilst hf3'(x) = J. If a and 13 are monomorphisms so that G and H contain K as a subgroup.. to some extent.) • When K = {I} we of course get back exactly to our earlier description of the free product. now we need to make a distinction between them. A similar result is true for the universal property defining free groups.Push outs of groups 149 t/Hfl= tjJf3' = r ex and tjJ IfFO = tjJ a' = 0. that is h is the same as f! (Here. Looking at the diagram K ~j H .(x). We have two candidates for f.. one in G and one in H. fa' = h? Any homeomorphism h: T ~ Z is determined by what it does to generators and any generator is really either an a'(y) or a f3'(x). Both satisfy the equations and so they must be equal since we know there is a unique map satisfying these equations.
10 the as the are elements of T look like. Suppose in a corner a H that K is a normal subgroup of G with a the inclusion and that H = {I}. then ga(k) E G. The full geometric significance of the definitions and constructions we will give will not be clear until later on. f3(k)h E H. 6. Z. We will look at two alternative methods of obtaining a presentation of a group from a knot diagram. .sentations of groups The above proof. contains all the information to enable us to describe the elements of this free product with amalgamated subgroup. h E H. as a result we find that (ga(k))h = g(f3(k)h). but the combinatorial methods used make calculations very easy to do. What happens if K is not assumed to be normal in G? t groups One of our main purposes for spending a fair amount of time on group presentations is that they are extremely useful when working with knot groups. although an abstract one. They are in G * H. it can be written as a word in the generators. Suppose g E G. very useful. but now any element in corresponding one in f3(K). Prove that the group T that completes the pushout square is isomorphic to G/ K with 13' the usual quotient map. and for each Z E Z. and then examine the independence or otherwise of these groups with regard to Reidemeister moves before looking at various links with other ideas we have met. a(z) = f3(z) within T. This example shows what strings of the g and the h a(K) can be identified with These pushouts of groups often used constructions. and we can form the words (ga(k»h and g(f3(k)h) within T. k E K. As k E K. They also include other.
Knot groups 151 presentations The naIve idea behind the Wirtinger presentation is that if you want to explore a knot. and label each arc. A piece of strong cord will also be useful!) Take a picture of the knot. 6. so use sensible distinct labels. orient it. . . (Before we proceed further. possibly labelling around the knot in some order. letting it out as we go so as to know where we have passed already. You can think of that space as being a threedimensional labyrinth. r m) where n is the number of arcs in the picture and m is the number of crossings. we follow Theseus in taking with us in our exploration a ball of string.7 The knot group will have a presentation G(K) = (XI> . ••• . xn : r j . it may help if the reader takes a piece of moderately stiff wire and bends it into a trefoil knot. but examining the space around the knot will tell you a lot more. at each arc write down the corresponding generator with exponent + I if the arc is . We will illustrate using the cinquefoil in Fig. Examination of the possible paths yields some generating paths and the crossings give us relations. (It helps to number the crossings so as to avoid missing any out of the calculation. .7. just going for a walk along the knot itself is not much help as it is only a circle. These labels will be the generators of the knot group. 6. joining the ends together to complete the 'circle'. Taking the advice of Ariadne.) The rule for writing it down is: starting at some point near the crossing move anticlockwise around a small rectangle. Fig.
6.9 which gives XcXiIXfclX.lxjl = I = 1 1 X4X2X. 6.= I.. 6. This corresponds to the relation r l : Xcxixfc I xi I. Clearly starting at a different position produces a conjugate of this relation. we have five generators and the following five relations: rl : r2 : XIX4X21X4 r3 : r4 : rs : XSXIX.8 entering the crossing and .sentations of groups Fig.9 For our example of the cinquefoil knot.. Let us assume that at the first crossing we had the form shown in Fig.8. I X2X3 or XI = X:sl X3XS' There are several things to notice here. Hence starting at the top gives Xc XiX" I xi I. Firstly each relation expresses a generator (corresponding to the arc leaving from under the overpass) in terms of the 'incoming' generator conjugated by the . 6. Fig.1 if it is leaving it.IXIX4 or X3 = X21X4X2 or Xs = x. There is a second possible type of crossing of the form shown in Fig.1 X2 = I X2X3X:sIXil = 1 x3xsxilX:s1 = I or x4=xllxsxl or X2=X.
IXIX4' X3 = X:. ! X! X 4 . ..10 Secondly the cinquefoil is extremely symmetric and this is reflected in the relations. This observation allows one to set up inductive machines to calculate the generators leaving a series of twists in terms of the incoming arcs (Fig. I X!X 4 X2! X 4 X 2 XI! X 3 X I becomes does not change does not change becomes XI = XiI X:. Xs.. using the expression for gives ./"' these? n twists Fig. X4' Xs: XI = XS!X3 X S' X2 = X:. Xl Finally.. It therefore is better to proceed by substituting for X2 or X 3. 6. say expression XiI X2X3: XI=XS!X3 X S X2 X3 X4 by substituting for the = = = X:.IX4X2' Xs X4 = XII XSXI' Xs = xi! X 2 X 3). At the same time be warned that as not all knots are symmetric. X 4 = Xj!X2IX4IX2X4X2X!).. . I X 3 X 2 X 3 X4 = Xj l xi!X 2 X 3 X! and we delete Xs from the list of generators..... a ~~:~<.~~~>what are _./'.. we need basic simplification techniques that work regardless of any special features. b . X 3 . We eliminate (any) one generator. We will apply simple Tietze transformations to the Wirtinger presentation so as to simplify it and will not use the symmetry explicitly.' .! X 2 X 4 X2 1X 4 X 2' X 2 = X:. You are advised to use special structures such as symmetry to help in calculations... The presentation is G(K) = (XI' X2.. X4 : Xl = X2! X:.10)... 6. We will write down the results: G(K) == (Xl> X2. Substitutions of complicated expressions are more likely to lead to slips than are substitutions of simple ones.Knot groups 153 'overpass' generator or its inverse. In fact one can write them in a 'string' This can be used to take short cuts to simplify the presentation.. Here strategic thought has to come in. not XI or X 4 • We choose to substitute for X3 next.
5)torus knot. That approach also leads to a group presentation introduced by Dehn. Orient the knot as before but now label the faces of the resulting picture. A typical crossing might be as shown in Fig. These facelabels will be the generators. Any single such loop can be pulled until it involves not that chosen crossing but all the others. 6. so do not just throw a relation away. The cinquefoil is a (2.IX4XIX4XIX4' 1 Xli X 4 XlI x 4 I XI X 4 XI X 4 X I )..11 and corresponding to this we write down the relation XiX] I XkX.11 Xk .IX. As a final step we simplify one stage further.) Keeping all the relations acts as a check on your calculations. b : as = b 2 ). As in the Wirtinger presentation we get a relation for each crossing. We note that the two relations are essentially the same: In Wirtinger presentations. (The reason is that each of the relations corresponds to a path or a loop of string that can be pulled off the knot. Here are the 2 and the 5 coming out from the calculation of G(K).resentations of groups G(K) == (x].1 Finally we add a relation Xm = I for Xi t Xl Xj t Fig. 6. then a few more Tietze transformations give G(K) == (a. esentations You may recall the way in which the Alexander polynomial was originally introduced by Alexander. b = XIX4XIX4XI. We will be studying torus knots and their groups more closely later on in Chapter II. x 4 : XI = x4 = X4IX.IX. Write a = XI X 4. anyone relation is a consequence of the others.
y : f 5 = y2). gives an f5 = (ed)(eb)(ae)(de)(ba) = (edebaf so we take y = edeba and.12).I = 1).. dg1ef. Fig.I = ebf. y : X5 = y2).1. This gives (a. The generators will be a. e.Knot groups 155 g . This common of. We also suggest that you repeat the exercises above for the Dehn method . Glancing back at the symmetric relations presentation suggests that this is the same a presentation of the form (x. Again we will examine the cinquefoil (Fig. Knowing where we are going shows a element f that we have actually got rid interesting pattern we had for the Wirtinger group and so should have short cut. 6. ag1efl and we will choose to 'kill off' g as this seems to cause a lot of simplification to the presentation.I = defl = edfI = aef. Eliminating f (since f = ba = eb = de = ed = ae) gives us (a. using the knowledge that you have at this point in the book. e. The relators are bg1af. We leave it to you to try to prove. d.anyone will do. if repeated. eg1bf1. and g. e : ba = cd = de = ed = ae). d. using a few Tietze transformations. eg1df. b. that the Dehn presentation is a presentation of the same group as the Wirtinger presentation. This last relation effectively deletes Xm from the presentation. 6. b. e. e.f: baf. e. f. b.12 some generator/facelabel Xm .1. d. get (f.l .
eliminating x' and reducing our presentation to that of G(Kold). However.13 The presentation for Knew will have one more generator.13) x KOld Knew Fig. We know equivalences can be broken down into Reidemeister moves. that is should not depend on which picture we use to calculate them. We look at Wirtinger presentations. and leave the Deh] versions as an exercise . Move II (Fig. In addition we have a relation for the new crossing.14) Again the change produces new generators and some changes in other relations. . Thus in this case the two groups are isomorphic.I = 1. 6. of course. It also gives relations . and some of the relations involving x in G(Kold> will be replaced by relations involving x'. gives x' = x. This provides a neat use of the fundamental group calculations that we will meet in Chapter 11 and proves invariance in an elegant way. which. obtained either by the Dehn or Wirtinger method. This in turn means that we can substitute x for x' in all relations. there is much to be said for a barehands approach. so it would seem a good idea to see what happens to the knot group if we change a knot by a Reidemeister move. namely XXXI x'. x'. 6.sentations of groups and also examine why it does not matter which of the original facelabel generators you kill off. should be invariants. 6. To put this more formally we hope that if K and L are equivalent knots then G(K) and G(L) will be isomorphic groups. Later on we will show that G(K) is the fundamental group of the knot complement ~3\K. e To be useful these knot groups.Move I (Fig.
g. Substituting this in the second relation gives Y2 = YI. We should also look at the move with one changed orientation (Fig.15 Move III This causes more bother and again we will only look at one of the numerous different subcases caused by relative positions and orientations. In Fig. 6. whilst in Fig. and iIeee'.14 a t the two new crossings. 6. so we can again reduce to the old presentation.x meaning that we can eliminate z from all other relations.I. 6. 6.16 the relations are a'ebe\ e'b'db. Check that you can do the adaptation of the argument above to this case.15) but this causes no problems.17 the relations are aI ebe'. The first gives z = x'y.I ede. but of course it can only occur in the second of these relations.Knot groups 157 Fig. and gI aI gao (We can safely leave all labels other than e and g . 6. Fig.
n)torus knots can all be drawn in the form shown in Fig. Comparing these we find that the old one has cb. Using the labelling . but the other unchanging relation gives a.t c = cb. 6.17 unchanged as we are only changing the knot in the vicinity of this triple crossing. is odd. (ii) the trefoil.sentations of groups / Fig. so the two presentations are equivalent and the two knot groups G(Ko1d ) and G(Knew) are isomorphic.) Simplifying the old presentation eliminating e we get a! cbc! and j. The (2.18 where n.1 where the new one has aI c.l cdc! a. 6.1 1. 6.l .11.16 • Fig. 6. and (iii) the figureeight knot.!a. Calculate the Wirtinger presentation of (i) the unknot. 2. the number of crossings.
6. Investigate if it can always be presented with just two generators.19 .n Fig. There is another family of knots that might be called the generalized figureeight knots. a figureeight. Using the last two crossings (the furthest to the right in the diagram). Using similar methods to the preceding exercise. write XI and X2 in terms of themselves and hence write down a presentation of the group of this knot. 3. 6. work out a formula expressing the generator Xk of the Wirtinger presentation in terms of the generators x I and X2. Is there an Fig. find a presentation for the knot group of the nth knot in the family. Simplify to show the knot group is isomorphic to (a.Knot groups 159 T2. b : an = b 2 ). we have a trefoil and for n = 2. their general form is as shown in Fig. 6.19 with a total of n + 2 crossings for the nth knot in the series. For n = 1.18 suggested.
) If K is any knot. (If you have not checked that the choice of which generator to kill is of no consequence to the end result. n)? Prove that G(T2. is infinite cyclic.20 Orient the two knots and label the faces as in Fig.Presentations of groups 4. 'elegant' way of presenting this group in any way analogous to the an = b 2 relation form of G( T 2 . the Abelianized knot group. a Fig. (You may not be able to complete the proof now but try to understand what the problem is. a say. show that G(K)ab. It should be clear that if a Fig. 3) is not isomorphic to G(unknot).21 . Examine how the Dehn presentation changes under Reidemeister moves and check that the results are equivalent presentations.21 where z is assumed to be a facelabel not used in K or L.) Now look at K + L. 6. 6.20. 6. and in both presentations use a as the generator to kill off. that is giving isomorphic groups. now would be a good time to pause and think about the problem. 6. Use the same label for the unbounded region. groups and the sum of knots What is G(K + L) in terms of G(K) and G(L)? For a change we will examine this for the Dehn presentation but suggest you examine the case of Wirtinger presentations for yourselves. 5. 6. this has picture shown in Fig.
S. aIeab. A typical relation in a Wirtinger presentation is of form xyxIz. b.Knot groups 161 G(K) = (X: R) X a set of generators containing x and G(L) = (y: S) Ya set of generators containing y with X n Y= 0. bIabc. {3(z) = x.I ebeIa. c. so b = alea.~G(K+L) where a(z) = y. (z: $) i.I .I (iii) (iv) eaeIb. z are coloured a. b.I = 1 then aba.l • Half of these are consequences of others. e in some order then mapping arcs to colours maps this relation to one of the following pairs of relations: (i) (iv) abaIeI.I bebIa. x = y) Thus the rule for forming G(K + L) is that of forming the pushout a. The six relations in .I or xIyxz. for instance if abaIe. bIeba. eIaeb. If this occurs at a crossing where x. b. y. thus giving an epimorphism from G(K) onto G. and z are all coloured the same. e and the relations so that colouring K.I . that is we need not worry about compatibility. The idea is that if Xi and Xj are generators both coloured a then we add in the relation Xi = Xi and call the amalgamated generator thus formed a. Y: R. defines a map on generators from G(K) to G compatible with the relations. the relation will vanish in G. and so on.~ G(L) ~j G(K) j I. that is attaching a colour to each arc and hence to each generator of G(K). If x. y. coloured by the colours a. We will construct a group G with generators a. then G(K + L) ~ (X.I . K.I aeaIb.l = 1. that is aIeab.I = e.I (ii) (v) babIeI.I c1beb. aIbae. ups and 3colourability Suppose we have a 3colourable knot.I .
Ibac. (This is quite hard.1 = a. (13). b. 2.) . Thinking of the transpositions (12). This leaves us to look for a group with three generators and relations (i). b to (13). The converse of this result holds: if G(K) maps epimorphically to S3 then K is 3colourable. this is similar to the type of relation we have above except that the generators in S3 are of order 2. but it does have a wellknown quotient group..11. (This has been done by Crowell but we will not tell you which target group he used.1 = cac1b.1 = b1abc. bcb = a. n is 3colourable if and only if n is divisible by 3.1 = 1) and we have a universal quotient of all groups of 3colourable knots. since by mapping a to (12). a2 = b 2 = c2 = I) and G 1 mimics much of the behaviour of S3' We do not need to work out whether or not G 1 is or is not S3. (ii). that the (2. and (iii).sentations of groups pairs (iv) to (vi) can therefore be omitted. It is infinite and its structure is not immediately recognizable.I = bab1c. that is we have shown that the geometric and combinatorial notion of 3colouring translates to some extent to the existence of an epimorphism from G(K) onto S3' The interest of S3 rather than G or G 1 is that it is a weBknown group and so such a homeomorphism helps our understanding of G(K) itself. c : aba1c. (23) in S3 we note that (12)(13)(12)1 = (23) etc. cae = b.2 I. n)torus knot [2. Prove that if G(K) maps epimorphically to S3 then K IS 3colourable.1 = c1acb. 3. We therefore take G = (a. This group G is rather unfriendly. (Some of the parts necessary to this are handled earlier in this chapter. using this method.) Prove. If we add the relations a 2 = b 2 = c 2 = 1 into the presentation of G we get a quotient group G 1 = (a. c : aba = c. b.) Investigate the possibility of an analogous result for ncolourability. 6. we can construct a homeomorphism from G 1 onto S3' What we have done is to use the colouring of K to map G(K) onto S3. and c to (23).
of a fundamental result about free groups. We shall develop the ideas for directed and undirected graphs side by side. v} is called an edge of r. but the reader must be warned to consult other authors' definitions when referring to other sources. via Van Kampen's theorem. phs. No accepted standard has arisen to settle the question of what a graph should be. Our main interest happens to be in directed graphs. A pair {u. and later on we shall meet covering graphs that give a proof. We have already met (in Chapter 6) the Cayley graph of a group presentation. u} and for pairs to be repeated in Er. we allow pairs of the form {u.7 Graphs and trees The theory of graphs impinges on the study of knots and surfaces in four significant ways. or a digraph. we select the essential facts and leave the reader to pursue the wider subject in texts intended for that purpose. rather. and trees An initial difficulty in graph theory is that of terminology. thereby avoiding a choice altogether. and how graphs can be studied by embedding them in surfaces. but many of the important concepts are most easily formulated for graphs without a preferred direction for edges. which produces a geometric representation of a group from sets of generators and defining relations. A choice presents itself as to whether the edges in a graph come with a preferred direction or not. We shall stretch the term to its fullest extent. but we prefer the more picturesque term quiver. We mention Wilson (1985) and White (1984) as being particularly suitable. We shall see how knots can be studied by graph theory applied to their diagrams. A graph r consists of a set V r of vertices and a collection Er of pairs of elements of V r. quivers. It is . An undirected graph we call simply a graph. It is not our aim to compress a complete introduction to graph theory into this single chapter. a directed graph could be called just that.
The bijectionfis then an isomorphism between rand r'. so we say that two graphs rand r' are isomorphic if there exists a bijection f: vr~ vr' such that mapping {u. the actual placement of the points. curved. 4}. 4}. l}. Examples . 6. {5. 4}. 4}. u} is represented by a line joining the points u and u. {3. or have to cross one another.f(u)} E Er' is a bijection Er~ Er'. 7. the graph with V r= {I. 3. {3.2 Of course. 7} and Er= {{I. 4}. 4.phs and trees extremely useful to represent a graph r (at least when V r is small) by a diagram in which the elements of V r are represented by points and an edge e = {u. {l. Different choices of labelling for the vertices and edges of a graph will not affect its graph theoretic properties. 7}} could be represented by the diagram in Fig.2. as in Fig. 7. 3}. u} E Erto {f(u). 7. and whether the lines are straight.1 In drawing such a diagram. {6. {t. 2. {3. 2}. {I.1. Fig. is irrelevant. 5}. 7. It is evident that the notion of a graph embodies a basic idea of everyday life: that two things are connected in some way. the essential requirement is correctly to represent the disposition of edges joining vertices. and we shall consider it in detail later in the chapter. 5. {3. {5. For example. So the two diagrams below both represent the same graph. the fact that we can draw a diagram of a graph r without edges intersecting is a significant property of r. 7 2 3 I 6 Fig. 4}.
and a terminal vertex. quivers. u) E VQ is an edge from u to u. or tennis players who meet in a tournament. 7. a subgraph is represented by some subset of the vertices and some subset of the edges joining these vertices. I t is . u) and for pairs to be repeated in EQ. obtained by forgetting about the ordering of pairs in EQ.Graphs. v) has see) = u and tee) = u. = fo sand . a neat definition of a quiver with some conceptual advantages is as follows. The functions sand t give the source and target of an edge. A pair (u. or target. A quiver Q consists of two sets VQ and EQ and two functions s.'J. and trees 165 include the stations on a railway network. We shall generally prefer the second definition. That is all we have to say. An element of EQ is an edge of Q and an element of VQ is a vertex of Q. or by ignoring the arrows in a diagram. and sometimes it is irrelevant. the components of an electrical circuit. Given two quivers Q and Q' with source and target functions s.3). Sometimes the direction of connection is important. is a bijection then f is an isomorphism and Q and Q' are isomorphic. 7. but the first makes plain the fact that a quiver is a graph with extra structure. A quiver Q consists of a set VQ of vertices and a collection EQ of ordered pairs of elements of VQ. If each of fo and J. u) pointing from u to u (Fig. students who attend a lecture together. More precisely. The notion of a subgraph of a given graph is straightforward: in terms of diagrams. or source. 7 2 3 I 6 Fig. = fo t. t' a map of quivers f: Q ~ Q' is a pair of functions fo: V Q ~ V Q' and j. The equivalence of the two definitions of quiver should be clear: an edge e = (u. It follows at once that Vi QI = VQ and that EI Q I = {{s(e). We allow pairs (u. tee)} 1 e E EQ}. Evidently we can represent a quiver by the same sort of diagram as for graphs. t: EQ ~ V Q.3 Since every edge in a quiver has a welldefined initial vertex. then a subgraph of Tconsists of a subset JV of V r and a set F of pairs of elements of W such that F ~ E r. if Tis a graph. t and s'. with an arrow on the line representing (u. We call this associated graph the underlying graph of Q and denote it by 1 QI.: EQ ~ EQ' such that s'J. Any quiver has a graph associated to it.
Vj. A walk in a quiver Q is defined to be a walk in the underlying graph 1 I. e 2. Successive edges are therefore incident with a common vertex.phs and trees often. An edge e = {u.j . u:. (Vi E Fig. there exists a walk in rwith initial vertex a and final vertex b. and if a walk is closed then so is any other walk equivalent to it. u with u. Clearly any walk is equivalent to a path. . given a. . is incident with V. e which may be deleted in an equivalence. v E V rand u :. v.Vl. see). . e. v. tCe). VI is a path if it contains no subsequence u. and so we call two walks equivalent if one can be obtained from the other from some finite succession of insertions or deletions of backtracking subsequences u. v. and we shall return to it in Chapter 12 in order to apply graphs to the study of free groups. e/. to give the important notion of homotopy. A walk in r is a finite sequence of vertices and edges V r. that is. This idea of equivalence will be generalized to the setting of topological spaces in Chapter 9. e. This definition implies that a walk in a quiver Q may include an edge that is traversed in the 'wrong' direction. ej. If an edge is traversed one way and then immediately the other way.4 . A walk is closed if its initial and final vertices are equal. e. e. ei E Er) such that e. then the walk contains a subsequence see). V E V r. convenient to regard a graph r' as a subgraph of r merely if r' is isomorphic to a sub graph of r in the sense just described. A walk V o. and V. e or e. Let rbe a graph. . u with u. The vertex Vo is the initial vertex of the walk and Vj is its final vertex. bE V r. Such a subsequence introduces backtracking into the walk.v. v. It is convenient to regard two walks as related if they differ only by such backtracking.r.r. a walk may include a subsequence of the form tee). and e E Er. e. v} E Eris said to be incident with u and v. e. e. 7. e or e. e or tee). see). The graph r is connected if. v. _ j.
Hence. (iii) transitive. 7. Define a quiver Q as follows: the vertex set VQ is V and EQ = {(a. b E VQ. (iv) antisymmetric? Formulate the definition of a subquiver of a quiver. b) I a.5 are mutually isomorphic.6 3. quivers. 7. b E Q and a ~ b}. The graph on the left in Fig.7 into families of isomorphic graphs. 6. 7. 5. 7.5 2. Show that the two quivers in Fig. Let V be a set and ~ a relation on V. 7. contains many circuits. 7. A graph is a tree if it is connected and contains no circuits. whilst that on the right contains none. given any a. 8. Fig. 7. . How many nonisomorphic simple graphs are there with at most four vertices? Translate the definition of isomorphism of quivers from the terms of the second definition of quiver to those of the first.Graphs. A simple graph has at most one edge joining a given pair of vertices. Find which of the three graphs in Fig. 4.6 are not isomorphic. and trees 167 A closed path in a graph Tis called a circuit.8 are not isomorphic. What properties does Q possess if ~ is (i) reflexive. Sort out the graphs in Fig. Show that the graphs in Fig. A quiver with this property is called relational. there exists at most one edge in EQ from a to b.1 1.4. 7. 7. (ii) symmetric. Fig.
7.7 Fig.phs and trees ABCDEFGHI J'KLMNOPOR STUVWXYZ Fig. 10. How many nonisomorphic vertices? 7. Prove that a graph is a tree a unique path between any two trees are there with at most four II.8 (a) (b) (c) Fig. 7. . Which of the graphs in Fig.9 9.9 are trees? graph. Show that a tree is a simple if and only if there exists vertices. 7.
11).12). 7. 7.1)/2 edges.12 . If every vertex of VI is joined to every vertex of V2 by exactly one edge then r is a complete bipartite graph. and is determined up to Fig. with exactly one edge joining any two vertices (Fig. Fig.10).10 (b) The complete graph Kn The complete graph Kn has n vertices and n(n . n A graph r is bipartite if V r is the union of two nonempty disjoint subsets VI and V2 such that every edge of ris incident with one vertex of VI and one vertex of V2 (Fig. 7. 7.11 (c) The complete bipartite graph Km. o Xl 8 Fig. 7. 7.Examples of graphs 169 mples of graphs (a) The nleafed rose Xn Xn is the graph (unique up to isomorphism) with one vertex and n edges (Fig.
with one subset for each generator in the presentation. Cayley quivers have the feature that their edge sets are partitioned into subsets. and for each K ~ 1 do the following. emma. The next result explains a method of encoding the partition of edges. To each Cayley quiver C there corresponds a graph rCfrom which C may be reconstructed. in a graph (and not a quiver!). Proof.Graphs and trees K 3 •3 Fig. Fix an ordering of the generators in the presentation. 7. (d) Cayley quivers The Cayley quiver of a group presentation appeared in Chapter 6. n (Fig. If x is the kth generator. 7.13).13 isomorphism by the sizes of the sets VI and V 2 • If I VI I = m and ! V2 ! = n we denote the complete bipartite graph with vertex set V= VI U V 2 by K m . replace each edge e in C labelled by x by a path of three edges and two new vertices e' and e": e' s(e) s(e) e' e" t(e) e" o<o_<o)~o oo~>__ot{e) . and their direction.
and a similar path of length 2k . Now suppose that rhas e> 0 edges.e + f = C + I as required. This edge either joins two distinct vertices of r. Further. and having c connected components. that is. that is v . 's formula for planar graphs. v .2. r' has c' connected components. so that by inductive assumption. the edge removed from r does not separate a connected component of r into two components of r'. Then ve+f=c+1.I and v . We have seen several examples of planar graphs. that is. The new graph is rc. where either c' = c or c' = c + 1. and let r' be obtained from r by removing one edge.e + I + J' = c' + 1. Now any graph consists of a collection of connected graphs called its connected components. for any r we assume that vr and Er are finite sets. If e = 0 then rhas as many components as vertices so that v = c. The basic fact about planar graphs is that they satisfy Euler's formula. and clearly f= 1.Planar graphs 171 Add a path of length 2k . Proof. or else is an edge of the form {u. and in either case its removal from r reduces the number of faces in the plane by one. u}. From a diagram of r drawn in the plane we obtain a diagram of r' dividing the plane into f' faces. Certainly r' is planar. Let r be a planar graph with v vertices. but otherwise consisting of new edges and vertices. which dates from 1750 and relates the number of vertices and edges in the graph to the number of regions or faces into which a diagram of the graph divides the plane. We prove the formula by induction on e. and has v vertices and e . with initial vertex e'. Recall that a graph is connected if there exists a walk between any two vertices. and it is easy to see that it has the property claimed in the lemma. and some that you may suspect are not planar.e + f' = c'. First suppose that c' = c. f' = f .I with initial vertex en.1 edges. Therefore the formula holds. e edges. that is. • ar graphs A graph r is planar if it has a diagram that may be drawn in the plane without edges intersecting (except at the vertices!). For the remainder of this chapter we shall assume that graphs are finite. so a connected graph has just one connected component. Suppose that a diagram of r divides the plane into f faces. If the edge removed does separa te .
e = II. Proof.Graphs and trees a component of r into two components of r'.14 (a).3 is not planar.14. v = 5. Prove that K3. e = 7. 4. In Fig. 3. 7. and every edge occurs in the boundary of at most two faces. Its deletion does not then change the number of faces. and c = 2. 2. If r is planar then e ~ 3v . Therefore 3J ~ 2e and this inequality combines with Euler's formula to give the result. • pies We may verify the application of Euler's formula to the two graphs shown in Fig. Let r be J connected simple graph with v vertices and e edges. The most celebrated result about the planarity of graphs is Kuratowski's theorem (Kuratowski 1930). (b). orollary. =Jand again v .e + J = c + 1. Show that a graph is planar if and only if each of its connected components is planar.6. 7. v = 7. whence/.14.J= 5.J= 7. 7. In any diagram of r.3 1.14 In Fig. every face has at least three edges in its boundary. We say that two graphs rand r' are isomorphic modulo vertices oj degree 2 if r is isomorphic to a graph r" obtained from r' by the addition or deletion of vertices with just two edges incident: o~~o~~o o~~o . so that c' = c + I. and c = 1. where v ~ 3 and is finite. The most important case of Euler's formula is that for c = 1. then this edge cannot be on the boundary of a finite face in the diagram of r and so occurs on the boundary of the infinite face. (b) Fig. Prove that K5 is not planar. • ises 7. 7.
Let r be a finite graph.15) carries embeddings to embeddings in the sphere. the larger the genus required. 7. the more complicated is the combinatorial structure of r. " Fig. for if we are given a rthat is not simple. Stereographic projection (see Fig. Proposition A graph sphere S2. as the following observation shows. Embedding graphs in surfaces The embedding of a graph in the plane is just one part of the more general procedure of embedding graphs in surfaces. 7. V r". how complicated a surface must we take in order to be able to embed r? That is. Theorem. and vice versa. Kuratowski's theorem is as follows. A graph .15 It is now a natural step to consider the embeddings of graphs in other surfaces. so that ro is simple. such as that in Wilson (1985). Er' and Er". 3 • We shall not prove this theorem here. If a graph r is not planar. Then r is planar if and only if it contains no subgraph isomorphic modulo vertices of degree 2 to K5 or K 3 . r is planar if and only if it can be embedded in the plane Proof. and embeds in a surface S if and only if r does. Note that for the moment it suffices to discuss only simple graphs. Denote by Sg an orientable surface without boundary and of genus g.Embedding graphs in surfaces 173 Give a precise formulation of this notion in terms of V r'. We may picture Sg as a sphere with g handles attached. then we may construct To by replacing multiple edges with single edges. but you are urged to consult an extended discussion. how large must the genus be? Clearly.
j} E E r}. We represent such a circuit by drawing it around a polygon.3) that characterized the graphs embeddable in Sg for each g . . Now let D = {(i. 2. what is the smallest g for which Tembeds in Sg? If r embeds in Sg then it certainly embeds in Sh for any h . .. Let Tbe a connected simple graph with finitely many vertices. The choices of the permutations Pb . V(l} is the set of vertices connected to vertex i.. Given a graph r. n}. answers to which would shape the theory of graphs embedded in surfaces quite nicely: 1. we can paste them together by matching an edge (a.?: g. a) to form a surface in which Tis embedded. as a consequence of a general method for constructing an embedding of a graph in a surface. For notational convenience we shall suppose that VT= {l. a vertex of such a polygon corresponds to a vertex of T and so can be labelled with an element of {I. for we shall see that Ks may be embedded in Sb the torus. Pj(i». We show that K3. we have only partial ones.16).set V(i)={jE VTI{i. j) E D then i E V(j) and so pii) E V(j).phs and trees rwith g edges then embeds in Sg.. If (i. J) = (j. 7. So K5 embeds in SIO.3 may be embedded on the torus SI (Fig..?: o. Having labelled all the polygons in this way. p.Pn determine a permutation P of D defined by P(i. j) I {i . Yet far from having ideal answers.(l)} E ET and the orbits of P consist of edges of rthat fit together into circuits in r. Question 2 does have a theoretical solution. Edmonds who described it in 1960.(i) is labelled with the ordered pair (j. 2. the edge from vertexj to vertex p.. Let I Veil I = ni and for each i choose an nrcycle Pi permuting the set Vel). Therefore {j. Which graphs embed in Sg for fixed g? 2. This is a rather extravagant waste of genus. A couple of questions now arise. . An ideal answer to question I would be a version of Kuratowski's theorem (see Exercise 4 of Section 7. Pj(i). The books by Gross and Tucker (1987) and White (1984) are excellent places to look for more detailed information. continually improved and updated by research work. named after J. draw one edge around each handle. This method is known as the Edmonds algorithm. 6. . . NowforeachiE Vr..j}EET}. n}.thatis. b) with an edge (b. so that an answer to question 2 seems more fundamental. 7}. . we could take PI = (6 4 7). For example if V(1) = {4. hence D is the set of all ordered pairs of vertices connected by edges in r.
2. 6} V(4) = V(5) = V(6) = {l.4) + (4. (1. for the embedding of K 3 • 3 subdivides the surface into the three regions given by the orbits. 2) + (2. We can fit the hexagons together to make the usual gluing instructions for the torus. 6) + (6. 5. This gives us three hexagons that can be pasted together to form a surface embedding of K 3 • 3' At this stage we can work out the genus of the surface. 4) = (4. 4) + (4. 5) + (5. 4) + (4. 5) + (5. we find three orbits: = P2 = P3 = (4 5 6) = Ps = P6 = (l 2 3). We will take the permutations PI P4 Thus P(1. 2) = (2. with K 3 . 1) (1.Embedding graphs in surfaces 2 3 175 4 5 6 Fig. 6} V(l) = V(2) = V(3) = {4. 4. 3. 3) + (3. (1.5) + (5. 3) + (3. 1). 3}.6) + (6.6) + (6.5. 3) + (3. P( 4. 7.16 V(Ku) = {I. with nine edges and six vertices. 1).2) + (2. 2. and so it is indeed a torus. 2). 5) and continuing the calculation. Therefore the surface has Euler characteristic X=69+3=0 and genus g = 1. 2) + (2. 3 embedded: 4 2 5 __l 6 3 .
• Given a graph r we can theoretically use the Edmonds algorithm to construct all the disc embeddings of rinto orientable surfaces. and if the smallest genus that arises is g then we know from Proposition 7. These considerations show that every disc embedding of a graph rin an orientable surface S can be constructed by the Edmonds algorithm. we say that r has genus g. and so arrive at the same contradiction. as does K 3 . If S is not cut into two components by cutting along the chosen circuit. If a graph r can be embedded in an orientable surface of genus g. these regions arise from the interiors of the polygons around which the circuits are drawn.6 that r has genus g. discard the component containing no edges of r and attach a disc around the circuit in the other component. r in an orientable surface S Proof. The practical difficulty is that there a huge number of embeddings constructed by the Edmonds algorithm. We may assume that the region of S contains no edges of r (or else we take a circuit containing fewer edges until we find such a region). We have seen that Ks has genus I.3' The planar graphs are precisely those of genus O. Hence the original embedding was not minimal. If it happens that S is thereby cut into components. again to form a new surface with smaller genus and with r still embedded. but not in one of genus h < g. corresponding to the choices of permutations to be made. that is as a subset of D. The parts of the partition are orbits for a map P*: D ~ D from which permutations Pi for each i E V r can be found. We call an embedding with this property a disc embedding. the boundaries of the regions of S\r are circuits in r and the succession of the edges in any given circuit can be read as a set of ordered pairs of vertices. position. Since S is assumed to be orientable. Suppose that some circuit in r bounds a region in S not homeomorphic to an open disc. and we cut S along the circuit. neighbouring regions can be oriented coherently and their boundaries read according to the orientations. A minimal embedding of a graph is a disc embedding.raphs and trees Any embedding of a graph r in a surface S constructed by the Edmonds algorithm has the property that S\r is a union of disjoint regions each homeomorphic to an open disc. Now given a disc embedding of r in S. then attach disc around each hole caused by the cut. An embedding of r into a surface of genus g is then called a minimal embedding. This creates a new surface S' of smaller genus than S in which r is embedded. partition the set D. and this is a contradiction. so that it .
take = (4 5 6). Youngs.Embedding graphs in surfaces 177 is infeasible to calculate the genus of a graph by this method. see White (1984) for a fuller discussion. The symbol I X l denotes the smallest integer greater than or equal to X.2 and K 2 . that is. Use the Edmonds algorithm to construct an embedding of Ks on a torus. Then u .3 on the torus.2g. What happens if you apply the Edmonds algorithm to a planar graph? Does an embedding in the sphere always result? Try K 4 . The complete graph Kn (n ~ 3) has genus I(n . . heorem.e + f = 2 . Let Tbe a connected graph with u vertices.2)/4l. K 2 . advanced techniques have been applied to calculate the genus for some special classes of graphs. • These results are due to G. 5.T. Suppose that a disc embedding into a surface 5 of genus g is such that 5\F is a union off regions. Attempt to embed Ks in 51. For example. 4.3' Prove the following generalization of Euler's formula. ises 7.4 1.4)/121. P4 = (1 2 3) and Ps = (1 3 2) = P6' 3.3)(n . 2. Represent the torus by the gluing instructions. Ringel and J. P2 = (4 6 5) = P3. try to draw a diagram of Ks on the torus without edges crossing. and we quote two cases here.2)(m . However. and the complete bipartite graph Km. PI Try a different choice of permutations in the example given above for embedding K 3 . e edges (where u and e are finite).W. n has genus I(n .
Examples of such expressions are gl . 7L denotes the ring of integers) of all formal sums of the form where each ng E 7L and only finitely many of the ng are nonzero. the trouble with it is to know why it works. but often when one can do something by a different method.6g s. The set TLG has a natural structure of an Abelian group where addition is given by . What we will do is give a second method of calculating Alexander polynomials. it is possible to see how to 'squeeze' the method a bit to get out more detailed information. The payoffs are (i) methods that can be applied to other group presentations which we will use later in Chapter II.8 Alexander matrices and Alexander polynomials Although the method that we have used to calculate the Alexander polynomial of a knot only uses the theory of determinants. 'It would be nice' sounds a weak reason. and so on. as usual. (ii) intermediate stages and related constructions that show more of the fine detail of the situation even in the case of knots. G. It is easy to check that the result is invariant under the Reidemeister moves but it would be nice to link it in with other invariants. up rings For any group.2g2 . consider the set TLG (where. this time using presentations of groups. and (iii) new invariants that have important geometric significance although in this book we will not be able to do justice to this aspect. 4g 3 + 3g4 . These in turn have exciting consequences within algebra but we will not be able to explore these fully. To do all this will require new algebraic ideas.
say.4 + 4a. The group G can be thought of as sitting inside 7L G via the function g H 19.) If we think of the set I G I of elements of G.a) = 4a . the same is true of the (ng + m g). Another example is when G = C 2 = {I. k are integers. Using this we can see that any group homeomorphism G ~ H composed with the inclusion of H into 7LH extends uniquely to a ring homeomorphism from TLG to 7LH. This last result will be useful in two cases in particular: (1) s'l: G ~ Gab. extends uniquely to an Abelian group homeomorphism from TLG to A.1 + t is an element of 7LC~. the Abelianization homeomorphism. Explicitly we have The result is not quite in the right form but if we write for kEG. It is clear that TLG is. Pk = L g ~ G ngmg~' k' then we can partition up the sum on the righthand side to get Lh G Pkk. so the righthand side is an element of 7LG. For instance. Any function fjJ: G ~ A. we get a multiplication on 7LG. that is we forget the multiplication of G. If fjJ: G ~ A is such that A is a ring and fjJ preserves multiplication then the extension to 7LG is a ring homeomorphism. in general.4a 2 = . n 2 E 7L} where. One of the most useful examples is when G = C~. with A an Abelian group. I GI· Remembering the multiplication of G. This ring TLG is called the group ring of G. A typical element of C~ is therefore (' and a formal sum can be thought of as a polynomial in t and r I or as a Laurent polynomial where j:%. then TLG is the free Abelian group on this set. . a} with a 2 = 1. then 7LG = {nl + n2a : nj. We leave it to you to check that Pk will be nontrivial only for finitely many k in G.Group rings 179 (As only finitely many of the ng and mg are nonzero. (4a)(1 . which makes it into a ring. the infinite cyclic group generated by (. a commutative ring if and only if G is an Abelian group. for instance. that is the element Lnhh where ng = 1 and nh = 0 if h =#= g. r I .
26}.g. 24.d.l)g in 7LG? We call two elements a.) of a finite set of elements a h . there is a unit u with a = ub. both 2 and . .d. and b divides a.c. and .c. so 3 is the only prime power common to both). g. Thus any element of G considered as an element of the ring 7L G is a unit. . . domain as we can use essentially the same method as we might use for finding g. then if a and b are associated. . the trivial homeomorphism. For instance. n then c Id. (We leave the proof to you. that is unique up to choice of order. The resulting ring homeomorphism sends Lngg to Lng. now would be a good time to look up the definitions. i = 1. .c.d. n. .. 1. b in a commutative ring R associated if a divides b (written a I b).1) X 2. . has a decomposition as a product of prime elements. (1 gI) = 1.) We will review briefly some ideas of ring theory for convenience: Let R be a ring with an identity element. the g. we might ask for what groups G is TLG an integral domain or a unique factorization domain? (If you have not met integral domains or unique factorization domains.2 = (.d..s in 7L given prime decompositions (e. We say R is a g. For instance.s of a given set are therefore associated and so if R is an integral domain.d.2 are g. an E R if d divides each ai.) An element d of a commutative ring R is called a greatest common divisor (g.xander matrices and Alexander polynomials (2) £: G ~ {I}.c. If Q is a subring of R. 4.d.c. in a group ring the element 1 is the formal sum nIl = Lngg where ng = 1 E 7L if g = 1 E G and ng = 0 otherwise. so there is some C E R with b = ac. of a set of numbers is determined up to multiplication by units.c. one of a and h must itself be zero). Any unique factorization domain is a g. and if c Iai for i = 1. A unit in R is an element u E R with an inverse. (201. then .. For instance.c. x :t= 0 or 1.6.1 is a unit in 7L. and is called the augmentation homeomorphism. domain if it is an integral domain in which any finite set of elements has a g. There is a whole class of problems which involve the transfer of properties from G to 7LG. . Can there be units other than those of the form g or (.c.d. You will find them in almost all the usual algebra textbooks.c.s of {. and there is a mapping p: R ~ Q preserving multiplication and such that a and pea) are associated for any a E R.d. that is there is some v E R with vu = uv = 1. . so any element g E G gives 1g E 7L G and (lg) . Recall 'R is a unique factorization domain' means that it is an integral domain with 1 in which every x E R.. Any two g.d. 33) = 3 since 201 = 3 x 67 and 3 = 3 x 11. If R is an integral domain (so if ab = 0.
Write out a proof. . so is R[t}. say more. The units in :EC= are all the elements of the form ± tn. b E :E C=. In each case.24. You should check that p preserves multiplication.J1 (a). We have already met :EC= as a ring of Laurent polynomials and clearly :E[t] is a subring of :E Cw Are we in the situation above? Is there a multiplication preserving mapping p: :E C= ~ :E[t} such that a and p(a) are associated for each a E :E C=? Any such a has the form n=oo where only finitely many of the an are nonzero. This gives a normalization of a. and then complete since always a and p(a) are associated. verify the property for the images. (The case where a or b is zero needs care but is fairly obvious. Of course.d.21 + r 24) = . so is the polynomial ring R[t] and the only units in R[t] are those within the subring R (considered a subring via the usual embedding as constant polynomials). If a = 0 then it will be convenient and consistent with our usage of p to define p(O) = 00. A consequence of this is the following result. (Proofs of these results can be found in many standard books on ring theory if you wish to check them. in fact.) Before returning to group rings we need to import two results from ring theory. Define. then so is R. and (ii) if R is a unique factorization domain. (Write out a complete proof yourself. The other cases are quite simple. p(a) to be the smallest n such that an 0. changing it into a polynomial. *" *" p(ab) = pea) + pCb). The group ring :E C= of the infinite cyclic group is a g.c.d. so that for example p(l) = 0 and p(t 2 . We can. then so is R.Group rings 181 (l) if Q is an integral domain. a and p(a) are associated.) The appearance of R[t] in a section on group rings should give a hint as to which group ring is going to be in the limelight for the next few pages. domain. to verify a property in R map the test elements back down into Q. if Q is a g. namely (i) that if R is an integral domain. n E :E. (2) The proofs are quite easy. domain. 00. We can easily check (do it) that if a.c. This will be of use when working out Alexander polynomials. since r J1 (a) is a unit unless pea) = We make the convention that r= = O.) Now define pea) = at. for a 0.
(a) Ker e is a free Abelian group with basis {g . one of the 'founding fathers' of modern knot theory. At several times in the following chapters we will touch on material related to this area. The kernel of e contains such elements as g . This is additive D(g + h) = D(g) + D(h) but is not a ring homeomorphism. but note the result.1. These two properties are very important in the study of the homology groups of a group. developed a theory of 'derivatives' (also called derivations) which is a powerful tool in the study of group presentations. The reason is that a similar argument would have worked for the group ring of any finitely generated free Abelian group.c. then (b) D(gn) = . one tries to apply methods based on topological ideas to study algebraic objects. In this. 8. We will not need this in detail. A derivative on a group G is a mapping D: G ~ ~G such that D(gh) = D(g) + gD(h). there is an extension to a mapping D: ~G ~ ~G given by D(Lngg) = LngD(g).1 Lemma.182 Alexander matrices and Alexander polynomials You may wonder why we introduced a general ring R in our earlier discussion. Example and Exercise The augmentation homeomorphism e: ~G ~ ~ sends Lngg to Lng.1: g € G} (b) the mapping from G to ~G sending g to g . You may like to try to investigate this situation and to prove that ~(C~ E9 . Fox.1 is a derivative.d. domain whose only units are group elements and their negatives.D(g) for g  gn _ 1 I all n € ~. 8. Prove that if G is finite. Given any derivative D: G ~ ~G. E9 C~) is always a g...2 Derivatives R. (a) If D: G ~ ~G is a derivative. .. H.
ax We note that ax /ax = ax 2 ~w = ~x + x~(y' x 2zx') = 1 + x~y' + ax ax xy' ~ ax (x 2zx').) 2 }xz + x 2 z aa . the free group on {x.) 8. so this expression equals 2 = = ((1 + x) + x 1 + xyl ((1 + x) + x I + xyl aax zx.1 is not the same as xz since the generators do not commute.1) xx 1+xyl (1 +x+x 2 Z(X'») which can now be simplified. (Caution: since we are in TLF and F is free. F = F( X). defined by ay _ ax  {I 0 if x = y if x =1= y. We will take X = {x. z}. You can do this by remembering that elements of F are equivalence classes of words and then checking that the 'obvious' extension to aw/ax works in such a way that equivalent words have the same image. y}. the free group on X. This defines a/ax on generators. y). You should check that there is no difficulty in extending it to the whole of F. is especially useful.1. X 2 ZX. In this case. y. calculate a/ay of the following words: xyxy' x1yl (xytx(xYrn.2 1. and will look at aw/ax where w = xyl X 2 ZX.Derivatives 183 The proofs are by easy algebraic manipUlation using the definition D(gh) = D(g) + gD(h) and are left as an exercise. 1 + x. to each free generator x of F there corresponds a unique derivative written ax' F~ 7LF a. F = F(X). (a) Working in F(x. for n E a/ax and (b) N . The case when G is a free group.
aAw) € 7l. yjP.sil: G ~ Gab for the natural epimorphism Abelianizing G. This is a bit like the proof that any mapping defined on the generators extends to a homeomorphism of F. Try it out. and see what you can do. Similarly we write . Investigate the following 'short cut' and its generalizations. We write X = {xI> .' dXj (dr. Suppose we want to calculate dw/dx and we can decompose w as W1Xn W where neither Wj nor W involves the generator x. x n }. Knowing these values a formula of the form D(w) = ax<w)D(x).F ~ 7l. for m. rm}.j = 1).1. Furthermore we will be a bit careful and will suppose all our relations are of the form r = I (i. then D is completely determined by the values D(x) for x € X that it takes on the generators. • • • . it is possible to find an elegant proof of the 'extension of derivatives' result using just the freeness of F and of Ker E. n. R = {r j . What are these coefficients axCw)? We suggested earlier that you proved that defining d/dx on the generators of F was sufficient to define it on all of F.. can be found. then 2 2 ~W=Wj(~)' dx x . X € X.) . Alexander matrix Suppose we have a group presentation (X : R) of a group G == F/«R».1 3. . we rewrite a relation r = s as rs. Prove that if F is a free group on a set X and D is any derivative. We form a 'Jacobian matrix' J= .G also for its 'extension' to the corresponding group rings. [xm.. Using the earlier exercise on Ker E. p € N. We will only need the case when X and R are finite and so will assume this is true throughout.exander matrices and Alexander polynomials (c) 2.e. and the fact that it is generated freely by the set of x . We will write r:F~G for the quotient map and r: 7l.F L XEX 4.
under the isomorphism. b). and relations rl = a 3. The significance of yis that it allows us to simplify using the relations in Rand . G = S3 = (a.2 b. that is (S3)ab == c 2 = (t : t 2 ) with generator t corresponding. the free group on {a. The definition perhaps looks slightly technical to you. b : a 3 = b 2 = (abf = 1). Thus we obtain an Alexander matrix . r3 = (abi. Xw Partially to make life simpler (and there. hence a. so a = 1. then applying the Abelianization and writing a for .'il(a) we find that abab = 1 implies a. a + aba J Applying y to the coefficients means that a 3. we use yand :A to take these coefficients in J down to 7L Gab' As this is a commutative ring. and (abf become 1. r 2 = b 2. We set This is the Alexander matrix of the presentation.'il allows us to simplify even further.2 = 1. as usual in mathematics. d rl da da =I +a+a2 d .r l =0 db dbr2 = 1 + b d r2 =0 da r3 d d = [1 + (ab)] da (ab) d d d db r3 = [1 + (ab)] db (ab) = (l + ab) = (1 + ab)a so our 'Jacobian matrix' is 1+ a + ( O I + ab d 0 I +b . We take F= F(a. b}. we run the risk of simplifying the interesting detail out of existence) and partially because we are really more interested in G or Gab rather than in F.3 = 1. In many ways TLF is like a Laurent polynomial ring in noncommuting variables X" •. but the various steps are easily followed through as you will see when we do some examples.The Alexander matrix 185 with coefficients in 7LF. we can work more easily here.2 = 1 and a. to b. 5 1.• . b 2 .
I is long and it is easy to make an error in calculating a(rsI)/ax. Consider a relation of the form r = s. However. 2. y(rsI) = I.I) ax = ar + r ~(sI) ax ax ax ax = ar _ rsI~.yxyxy in 7LF(x. especially when relations are long. we use the calculation via y3 x . Before considering another example. Put r = x 2 so axr=l+x a Put s = yxy2 x l. y).s in the group ring 7LF. we have ~ (rs.xander matrices and Alexander polynomials A =( ~ I + t since s1(a) = 1. this normalizes to give rs. when we calculate the Alexander matrix of the presentation we first apply y: F ~ F/~R~ and in F/~R~. ( ax ax ax This implies that we can compute the derivatives of rs. In fact. rs.xy in 7LF(x. The Abelianization Gab will have y = I and we will put s1 y(x) = t (t 2 = I in Gab. If rand s are of moderate length.I = 1 and if x is a generator. but one of the best is to calculate the derivatives of xyxyx . we will look at a computational hint which saves time and diminishes the chance of errors.I by using r . we get (x. y) to get . so a_ ) y (rs I ) = y (ar . y : x 2 = I. y: xyxyxyIxIyIxlyI = I). y : xyxyx = yxyxy) or (x. in the cinquefoil knot group presentation. y3 xy2 X I). here the derivatives of this single relation can be handled in several different ways. which is isomorphic to C 2).as) . G = (x. For instance.
21. x = yzyI. Y = zxz. What sorts of quantities could we expect to construct from A that might be invariant? One way to guess the sort of answer to expect to such a question is to see how an Alexander matrix of a given presentation changes as we apply Tietze transformations. If you feel you understand quite well. (a) (b) (c) you can use either the Dehn or Wirtinger presentations to obtain this.. the Wirtinger presentation has the advantage that all the generators map to t under the composite . (Beware.) Check that (a. (Although 2. b : a P = bq)ab is infinite cyclic if the highest common factor of p and q is 1. . G(t2 .I) = o. If n = 2m + 1. G(t2.sd y. 3. G(cinquefoil) = (x. and t 2. n)torus knot. y : (xyi x = y(xy/). J ses 8.) Calculate the Alexander matrices of the following presentations: G(trefoil) = (x.3 I. n) = (x.. you need to do the preceding exercise first. . Gab is an infinite cyclic group with one generator. (This exercise is a repeat. z : z = xyxI. y.ddX (x y) ) = .Elementary ideals 187 .) lementary ideals The Alexander matrix A of a presentation is not invariant under Tietze transformations since even its size can change if a generator is added or removed.X(l + y») = 3 .n is the (2. r2) .sdr(y3 . t.sd r (ddX (y3 x) .I). Similarly s:tr(aayS = s:tr((l + y + y2) . Xm : rl> . (d) Again as in (c). it would be a good idea to consider the completely general question of how the Alexander matrix of (XI' . b : an = b 2 ). If you have not already done this...sd r (ddX S) = . y : (xy)mx = y(xy)m). (We suggest that you do this investigation yourself. That way you can adjust the complexity of the example you tackle to your present understanding of the question. n) = (a. check that if G is a knot group.
the ideal generated by all (n . we had to assign meanings to Ek(A) when r .k) x (n . Formally an ideal behaves like the normal subgroup generated by a subset of a group. To keep things fairly general for the moment. Remarks (a) (b) In case you have not met the term (or have need to recall it) an r x r minor of A is the determinant of an r x r submatrix of A.k) minors of A. Since any (r + 1) x (r + 1) minor can be expanded as a linear combination (over R) of r x r minors of A. Si € S. matrices over 7l. If S is a subset of R. A = [ 1+ t Eo(A) is generated by 2 x 2 minors: ~ 1+ t I ~ tl' I + 1+ I~ l~tl' Il~t ~:.I' 11:'t tl t ' . Whichever way you go about this.2 x . ~ En(A) = R. Ek(A) = O. the whole ring.k ~ m. we will start by working over a general commutative ring R. a nonnegative integer. a .) Alexander matrices are matrices over group rings. for instance when handling presentations of finitely generated Abelian groups. The slight awkwardness in the definition is needed because if we merely said that Ek(A) is generated by the (n . For k. the kth elementary ideal Ek(A) of A is: (1) (2) if 0 < n .2 minor or a lOx 10 minor in a 4 x 5 matrix? Because of this..k. if (n . Eo(A) ~ El(A) ~ . Let A be an m x n matrix over R.Alexander matrices and Alexander polynomials changes under the various transformations. Ek(A) = R.k is small or larger than m. say.. the ideal E generated by S is the set of all finite sums Larsi' ai € R. these ideals are nested: (c) o~ pies 1.k) minors of A then what could be the meaning of.k) x (n . 7lG ab • You are most used to matrices over IR! or C and may also have met.k) > m. compare the transformations that result with those you know from linear algebra. if 0 ~ (3) n .
y : r == = 1.. that is by the determinant of A: (l + 1)(3 . G = (x. y3 X = X y2). This gives us that Eo(A) is generated by 3(1 + t) and . so all elements have the form a + pt for a...2(1 + t». Eo(A) is generated by 2 x 2 minors.2/.2t). + flt)(l + t) = 3. but as We get a chain ok (1 + t) <. 7L C 2 • . As we saw Gab C 2 and the Alexander matrix is over 7LC2 as before. El (A) is generated by 1 x 1 minors (1 + t) = (l + t)(3 . (3 .... This would mean there was some A.2 t) <.2/) = 3 + 3 t . .. We have El (A) = (3. p E 7L. the form of the Alexander matrix of the usual presentation of S3 and (S3)ab is cyclic of order 2. 3(1 + I).2t). + fl) + (A. Since (l + t) = 3(1 + t) . Similarly one knows that 1 E 7L C 2 and we can easily adapt the above to show that 1 e EI (A). we get Eo(A) = (1 + t). + fl)t = 3 which is silly since A.) Multiplying out our conjectured eq uation gives (A. + fl cannot be both 3 and O. fl E 7L such that (A.. El (A) is generated by 1 x 1 minors and hence by the entries of A.(1 + 1)2.2(1 + t). 2.Elementary ideals 189 hence by 3(1 + I). . We have not stated over which ring we are working but A is. 1 + 1 and 3 . 1 + t) over 7L C 2 • We know Are both of the inclusions proper or is one an equality? If Eo(A) = E( (A) then 3 E (l + t). we put t 2 = 1. in fact. and on multiplying.21 .. EI (A) = (3 . (We must remember that we are working in 7L C 2 . so we will assume R = 7L C 2 generated by 1 with t 2 = 1.21 = 1 + I.2(1 + t) (we write Eo(A) = (3(1 + t). This shows that 3 e Eo(A) and the first inclusion is proper.
Tl: Replace (X: R) by (X: R u {r}) where r E «R». Any two (finite) presentations of a group G have the same chains of elementary ideals. we can link the two presentations by Tietze transformations. the proof reduces to checking that Tietze transformations leave the elementary ideals unchanged. Let us suppose we can prove that the chain of elementary ideals is an invariant of the group G and does not depend on the presentation used. . (Check for yourself. Our next preoccupation must be to prove invariance of the elementary ideals of a presentation. We have above presentations (a. later on we will need this technique to prove that certain groups arising from surfaces are not free groups even though their Abelianization is free Abelian. SO Abelianization does not help. y : x 2 = 1. Clearly we need only look at Tl (Adding a superfluous relation) and T3 (Adding a superfluous generator). Is G know that Gab == C 2 == (S3)ab. but we also know that if G == S3. such that r = II u. we cannot conclude that two groups are isomorphic if their elementary ideals are the same. the logic is: the invariants are different. = X y2) == S3? We Proof Since by Tietze's theorem. Ui E F(X) (possibly with repeats).) We can now see how we might use these chains of elementary ideals. It is important to note that. p Then if x EX. as usual. However. ~ x of our group G of example (2).rjui'· i= . rem.2/). and therefore the groups are nonisomorphic. it would be possible to prove that G was not S3 in many other different (and probably quicker) ways. then we could find Tietze transformations linking the two presentations.xander matrices and Alexander polynomials This is a different chain of ideals than that in Example 1 since 3 ~ (3 . so G ?/= S3' In this case. Thus there are elements ri E R. b : a 3 = b 2 = (ab)2 = 1) of S3' and (x. This in its turn would imply that the two presentations would have the same chains of elementary ideals. This last statement is false.
In fact our calculations show that .I will involve some at least of the x E X. j)entry of A' is a" (aij) except if the entry is in the last row or column. Of course the last relation s = wg. dr/dg = 0 for all i. y(ri) = 1 for i = 1. ..) T3: Replace (X: R) by (Xu {g} : R u {wgI}).. rn }. We denote the two (isomorphic) groups by G and G'. Hence for any k.p. . .. this induces a': G~ G' = F(Xu {g})/«R u {wgI}» (which is an isomorphism) and hence an isomorphism It is clear that the (i. and the two Alexander matrices by A and A' respectively. We assume R = {rio . The inclusion of X into Xu {g} induces a homeomorphism a: F(X) ~ F(X u {g}). Dividing out by relations. so Next we calculate these p individual terms: Again we map this down by y to get Thus the Alexander matrix of (X: R u {r}) is like that of (X : R) except that it has an extra row which is a linear combination of other rows. (If you do not see this last step immediately try a simple example and then try to prove that the statement is true. the corresponding minors generate the same ideal. then as g is a symbol not in X.Elementary ideals 191 However. .
an E lLGab .. using elementary ideals. b 2 a 2 = db) = (x. y3 x H = xy2). The kth knot polynomial ."(A) = ~) for some row vector a'. Let A be the Alexander matrix of the Wirtinger presentation of the knot group G(K) of a knot K.k) x (n . . then lL Gab is a g. . say. letting d be a greatest common divisor of a). or have determinant zero.k can arise in this way. am then as each of its generators is in the principal ideal (d). As we saw.1k (t) is the . Ek(A) ~ (d) and (d) is the smallest principal ideal with this property (prove it!).k) x + 1) . the ideals Ek(A) need not always be singly generated.k) minor of A. y: r = 1. If. a knot group. .k) x «n + 1) . we have found a matrix A over lLGab and a chain {Ek(A)} of ideals that are invariants of G. domain. lexander polynomials via the Wirtinger presentation Given a group G and a presentation (X : R).d. All possible M n . • where M n  a" (M"(n + I) se 8.k) minors of An are of the form «n «n k is an (n .. The various d for the different Ek(A) form an important and calculable set of invariants of G. Now it follows from elementary arguments of linear algebra that the + 1) . so if Ek(A) is generated by elements aj. Gab is a free (finitely generated) Abelian group. that G is not isomorphic to = (a..k) minor. ..c. however. as G(K)ab == C=. so Ek(A") = Ek(A') = Ek(A).4 Show. Of course this holds in particular if G = G(K). or are an image k) of some «n + 1) . b : a 3 b = ba.Alexander matrices and Alexander polynomials A' (. We can use a column operation to eliminate a' and get an equivalent matrix with zeros in those positions of the bottom row: A'  (a'ri A ) ~) = A".
s can only be determined up to multiplication by units.I This has 2 x 2 minors (t .[2 + t .I t . An eightcrossing knot with seven generators will have a 7 x 8 Alexander matrix if the presentation is not reduced in size first.1). the same is true for these polynomials Ak(t). Earlier you should have found that the Alexander matrix for the Wirtinger presentation of the trefoil is . a b Fig. then AI (t) will involve evaluating all 7 x 7 minors of A! The example we will look at in detail only involves five crossings and five generators (Fig. 8. It often pays to simplify a Wirtinger presentation before working out A and hence the Ak(t).k) minors of A.c.I and Remark We see that since g. It also means that 'the' in the definition of Ak(t) is rather too strong as 4(t) is not completely determined by the definition.2   t + I and .Alexander polynomials via the Wirtinger presentation 193 greatest common divisor of all (n .1 .k) x (n .1)2 + t = t 2 again t + I. Al (t) is the Alexander polynomial of K. . but even that would be bad enough.t + I. thus AI (t) = t 2 . For k = I. This in part explains why the combinatorial derivation of the Alexander polynomial had to be 'normalized' to be well defined.d. 8.
l de).) Thus (We have evened up the two sides of the relation to help with the calculation. d.I de1ded1ad. Simplifying the presentation. e.silr(~: J= .sil (in effect substituting .xander matrices and Alexander polynomials G(K) = (a. a Examination of the two relations shows that they are conjugate as one would expect. (a. (. b. e : d = alea. e = d. b = dIad. e = aIelaeaIeIaeaeIaIea). e = e.I + eIa.2 2r2 . b = dIad. (This is a useful check that we have made no errors.) Take so Similarly ae ar = e.I edIadeId) = eIaIeaeIaIeaeaIelae. e. a = bIeIdeb.2r2 + 3rl 2 a = e = t) we get whilst . so .eIaIeae + eIa1eaeIa. e. b.I be.3t + 2). d: d = (a. = aI ea. e: a = dIa.1 Applying rand then . e = eIdIebeIde) = Here e = e. a = bIeb.silr(~:)= 2r2 Our Alexander matrix is 3rl + 2. We will 'normalize' Al (t) to have no negative powers of t and positive constant term. we obtain (a.2r2 + 3rl .I . d: d = alea. as we know that anyone relation in a Wirtinger presentation is a consequence of the others.
4.:1. Investigate what happens if one applies the Fox calculus. y : x p = yq) is the group of a (P. chapter IX). Examine in particular the relationship with Alexander's original method of defining the Alexander polynomial. . q are coprime positive integers. using results and exercises from earlier chapters. They can be useful in checking for errors: 1. If the knot is alternating. For any knot K. (Proofs are given in Crowell and Fox (1967. etc. . Suppose K is a knot with crossing number c. l 8. elementary ideal analysis to a Dehn presentation of a knot group. and the genus of K can be calculated. Thus if Li l (t) = cnt + . Alexander polynomials. (In fact we showed that it changed knot sum to product of polynomials using the 'original' method. Try to see why this is true from the point of view of Wirtinger presentations.. The group G(tp. + Co. For any knot K. 3. 2.(t) We list the following without proof. . q)torus knot where p. q) with presentation (x..5 1. 3. how the Alexander polynomial reacts to knot sum. Alexander matrix. there is an integer n such that Lil(t) = tnLiI(tI).. LiIO) =± l.n. Investigate. Investigate the relationships between C and the degree of the normalized Alexander polynomial.Alexander polynomials via the Wirtinger presentation 195 5 of . Calculate Lil(t) for this knot and show that 2. the coefficients pair up Ci = Cn _ i for i = 0. it is useful to note that genus of K = n 1degree of Li (t)..
then try to adapt Alexander's original method! Using the preceding exercise.xander matrices and Alexander polynomials 5.) Calculate lots of Alexander polynomials. If you are successful in doing such a generalization.2 are themselves related. There is a simple equation rather like the Homfly defining equation. (You choose the knots from anywhere in the book!) Can the Alexander polynomial construction be extended to be \ \ Fig. 8. Many of the necessary changes in the theory are hinted at in this chapter. 7. find how the Alexander polynomials of knots related by the sequence of changes in Fig. This should help you to find a relationship between the Alexander polynomial and the Homfly polynomial. 8. 6. . Use either formulation of the Alexander polynomial. This may seem to you more complicated but it does provide a new type of insight.2 applicable to links with more than one component? See how far you can get.
9 The fundamental group The fundamental group is an algebraic label attached to a topological space. We can therefore restrict (0: . we simply define COJ(t) = (O(rt) for 0 ~ t ~ 1. where r ~ O. The real number r is the duration of the path. Our principal application will be to knots: any knot yields a group. beyond that covered here. the more variety there will be among the paths in that space and the larger will be the fundamental group. The fundamental group consists of homotopy classes of paths in a space. The notion of homotopy is also crucial in complex analysis: Cauchy's theorem on contour integration may be phrased as saying that the integrals of a holomorphic function around two homotopic contours are the same (see e. and algebraic properties of the group reflect properties of the knot. A path (0 in X is a continuous map [0. s Let X be a topological space. Naturally enough. may be found in Armstrong (1983) and Brown (1988). The more convoluted is the shape of the space. it is a group and is constructed using paths in the space. there is a normalized path (01 with the same image in X. Further development of the extensive theory of the fundamental group. This led to the notion of deformation of one path into another. continuous analogues of the sequences of edges constituting paths in a graph. His idea was to classify the possible evolutions of a system of three bodies moving under mutual gravitational forces as paths in the phase space with coordinates given by the bodies' positions and velocities. now formalized as homotopy of paths. if r = I we say that the path is normalized.g. namely the fundamental group of its complement in [R:3. but to regard nearby paths as the same. The fundamental group first arose in Poincare's work on the threebody problem. r] ~ X. Given any path (0 in X. Priestley 1990).
. r"] ~ X be paths in X with w(r) = w'(O) and w'(r') = w"(O). w is not.t). the product W· w' may be defined whilst w' . This is not true of the normalized product: why not? For each point x E X there is a constant path 0. and is plainly not that of a group. If w: [0. The next step towards the fundamental group is to put an equivalence relation on the set of paths in the space X. but twice as quickly. We write this product as W· w'. then wand w' have an obvious product. when it proved necessary to consider equivalence classes of words on an alphabet. r'] ~ X and w": [0. then for any w' the path W· w' (if it is defined) has duration at least r and so cannot be a path 0. We leave the proof of Lemma 9. If w: [0. r] ~ X is a path then W· OW(r) = wand OW(O) • w = w. It turns .: {O} ~ X of duration 0 defined by OAO) = x. r+ r'] ~ X defined by W· w'(t) = w(t) if ~ t ~ rand W· w'(t) = w'(t . ware defined and are paths of duration 2r. Then the products W· (w' • w") and (w· w') . W· w'(r + r') = w'(r'). Since a(t) = W· w'(2t). Thus a is obtained by going along wand w' in turn. Both W· OJ and OJ. w" are both defined and are equal. w': [0. The best approximation that we have to an inverse for w: [0. we have aCt) = w(2t) if 0 ~ t ~ 112 and aCt) = w'(2t .1 as an exercise. In particular. However. r] ~ X defined by OJ(t) = w(r . If w: [0. We see that the set of paths in X is endowed with quite a reasonable structure. such that w(r) = w'(O). r] ~ X is the reverse path OJ: [0.e fundamental group attention to normalized paths if we choose. The duration of W· w' is r + r' and clearly W· w'(O) = w(O). for if w has duration r> 0. so that the paths of duration 0 provide a set of left and right identities for the product operation on paths. r] ~ X and m: [0. There is a product of normalized paths as well. but one which is none the less deficient in certain respects. ° rna.1) if l/2 ~ t ~ 1. It is important to note that the product of paths is not defined for arbitrary paths wand w': it must be that wends at the same point at which w' begins. the path obtained by first going along the image of wand then along the image of W'. r] ~ X. Note that it says that the product of paths is an associative operation. but it is often more convenient to allow paths of arbitrary duration.r) if r ~ t ~ r + r'. 1] ~ X and w': [0. then we have W· w': [0. Let w: [0. 1] ~ X are normalized. 2] ~ X. r'] ~ X are paths in X. then their normalized product is the normalized path a obtained from W· w': [0. The circumstances are similar to those encountered in the construction of free groups in Chapter 6. there are no inverses.
3. with this equivalence relation. As already remarked. see Exercises 9. Observe that the normalized product of any two closed paths at Xo is defined and is again a closed path at Xo' Suppose that a and m are closed paths in X at Xo' A pointed homotopy from a to m is a continuous map h: [0.27 below. and this leads to the fundamental groupoid. Let X be a topological space and Xo E X any point of X. then f and g are homotopic and we write f=g· position. it will be sufficient to consider only normalized paths. t) = Xo = h(l. For a full account of groupoids. 1] ~ Xwith m(O) = Xo = m(1). h(s. I) = g(y). Y~x. t). 1] x [0. A homotopy between I and g is a continuous map h: Y x I ~ X such that. 0) = a(s) and h(s. and further such that for all t. t) I0 ~ t ~ I} can be thought of as a continuous deformation from I to g. it will suffice to consider only normalized paths in our route to the fundamental group. If a homotopy between I and g exists. In 0) x 0: . motopy Let X and Y be topological spaces and suppose that f: Y ~ X and g: y ~ X are continuous maps. or by adding extra conditions on the homotopy. The set of functions {h(.Homotopy 199 out that. see Brown (1988). for all s E [0. The properties of paths with which this chapter began are more accurately reflected by fixed end point homotopy of paths. hey. A closed path in X at Xo is a continuous map m: [0. and for the first steps in the theory of the fundamental groupoid. h(O. 1). 1) = m(s). 1] ~ X such that. Pointed homotopy of closed paths leads to the fundamental group. Homotopy is an equivalence relation on the set of all maps There are variations on the notion of homotopy obtained either by considering only a limited class of functions. for all y E Y. 0) = I(y) and hey.
heorem. x o) into a group. We have. and h"(s. Then h" is indeed a pointed homotopy from a· a' to OJ· OJ'. and horizontal slices give the intermediate closed paths h(. t) = h(2s. the binary operation [aHa1 = [a. The following picture indicates the (0 (0' a a' definition of a pointed homotopy h" from a· a' to OJ· OJ'. x o). For any [aJ. Define h" as follows. and this analogy with the construction of free groups leads us to hope that 1T] (X. x o). 1] x [0. the top by OJ. Paths are composed by juxtaposition. There are pointed homotopies h from a to OJ and h' from a' to OJ'. Suppose that [a] = [OJ] and that [a'] = [OJ']. with multiplication of equivalence classes of paths obtained by taking the product of representative paths. A simple but useful picture to bear in mind is the unit square [0. we have a set of closed paths at Xo that give a continuous deformation from a to (0. 1] adorned with labels to show how it is mapped into X by h. t) if o ~ s ~ 112. Proof. t) if 112 < s ~ 1. that is. x o) is a group. it is independent of any choices made of one closed path in an equivalence class. t). a1 is well defined and makes 1TI(X. 0 ~ t ~ 1. and we will use [a] to denote the equivalence class of a closed path a. just as were the words in a free group. for all t.The fundamental group a pointed homotopy. and therefore [a· a'] = [OJ' OJ']' . Pointed homotopy is an equivalence relation on the set of all closed paths in X at x o.1. Any point on the sides of the square is mapped to Xo: the bottom of the square is mapped into X by a. h"(s. [a1 E 1Tl (X. Step 1 For ease of digestion the proof is divided into four steps. We shall show that the operation is well defined. The set of all the equivalence classes of closed paths in X at Xo is denoted 1TI (X. t) = h'(2s .
They are homotopic. t) = Xo formulae h(s. t) = a'(4s . Remember that. t) = a(2s/(t + I)) otherwise.t) if (t + 2)/4 < s ~ I.t . t) = a"«4s . we need to construct a homotopy h from (a· a') . a a' a" Define h(s. an to a' (a' • an). and this has the properties we need. h(s. Step 3 Let e: [0. A homotopy from a· e to a is indicated by the picture a and given s ~ (t by the explicit o~ + 1)/2 and h(s. Then [e] is an identity element in Jrj (X. x o).t . aN) need not be equal paths. 1] ~ X be the constant path at Xo.Homotopy 201 Step 2 The operation is associative.1) if (t + 1)/4 ~ s ~ (t + 2)/4. aN at x o. for given any closed path a in X at Xo we have a· e = a and so [a][e] = [a). because we are dealing with normalized paths. (a· a') • a" and a· (a' .2)/(2 . and again a picture indicates what we should do. and h(s. if . Given closed paths a. t) = a(4s/(t + I)) if 0 ~ s ~ (t + 1)/4. a'.
4 is rather sketchy. xo) and HI (X. The gluing lemma should help. XI) if Xo and XI lie in different path components of X? 2. xo) is the fundamental group of the space X at the point Xo.2 1. its reverse a (which you will recall is defined by a(s) = a(l . xo) and HI (X. If X is path connected and X o. 1] ~ X be a path in X with w(O) = Xo and Suppose that a is a closed path in X at Xo. xd.The fundamental group Step 4 If a is a closed path in X at x o. Supply the details.] The proof of Proposition 9. . What happens if we choose another point XI instead of Xo at which to base all our closed paths? roposition. • The group HI (X. xd. Let w: [0. The only difficult part is transitivity. xo) ~ HI (X. w(l) = XI' ises 9. that is. A homotopy from a· a to e is indicated by the picture a a and we leave it as an exercise for the reader to write out an explicit formula for the homotopy. • Proof. Is it reasonable to expect an isomorphism between HI (X. The definition of homotopy will have reminded the reader of the definition of isotopy of knots. [Hint. then are they homotopic? Why isn't homotopy a useful equivalence relation for knots? Prove Proposition 9. 4. k': S' ~ [R3 are isotopic. If k.s» is also a closed path in X at Xo and [a] is the inverse of [a]. then there is an isomorphism between HI (X. [a][a] = [e]. 3.2. then W· a· iij is a closed path in X at XI' and mapping [a] H [w • a' iij] gives an isomorphism HI (X. XI are two points of X.
t)m(s). and these portions of the path can be smoothed out by some suitable homotopies. and will wind around the centre of the circle a certain number of times. and is generated by the homotopy class of the closed path m: t H exp (2trit). The solution we shall adopt here may initially seem strange: we consider paths in [R instead of paths in Sl. Its wanderings are thus no more than changes of direction. A topological space is simp(v connected if it is path connected and if trl (X. Thefundamental group tr1(St. Our claim is that this winding number completely determines the homotopy class of the path. Proof. and see if you can use your definition to prove the theorem. If X ~ [Rn is a convex set then X is simply connected. IJ x [0. It is convenient to regard 1R2 as the complex plane and so identify Sl as {z Eel I z 1 = I}. Since composition of paths induces addition of winding numbers. t) = txo + (l . Then h is a homotopy from m to the constant path at Xo. y E X. and all t with ~ t ~ I. 1] ~ SI given by m: t H exp (2trit) extends to a function IR ~ Sl given by the same formula: you may picture m as winding the real line IR into a spiral and then flattening the spiral onto the unit circle. 1) is infinite cyclic. we will have an isomorphism rp: trl(St. and define h: [0.Calculations of the fundamental group 203 Calculations of the fundamental group The computation of fundamental groups will be the main theme of the remainder of the book. x o) = I for some (and hence all) Xo E X. that is. the circle. The crucial fact . A path in Sl may wander around in a complicated way. • Our first example of a nontrivial fundamental group is Sl. ° Theorem. it is worth considering the matter in a vaguer fashion. Before we embark on the detailed proof. Any path will be homotopic to one without any backtracking. Our first problem is to give a definition of the winding number. The path m: [0. I] ~ X be a closed path at Xo E X. Let m: [0. You might like to think of ways in which this might be done. or backtracking. I) ~ ~ as claimed. we take lEe as basepoint. We pause here to consider some spaces whose fundamental group is trivial.I)y E X. that two paths in Sl are homotopic if and only if they have the same winding number. Proposition. but it is forced to remain within the onedimensional bounds of the circle. so that examples will be in plentiful supply. 1] ~ X by h(s. Proof. Recall that X ~ IR n is convex if for all x. we have tx + (t .
= a. lJ ~ SI there exists a unique path 0: [0. then U and V are open and Uu V= SI. is our definition of ¢ a workable one? Firstly we have yet to prove that OJ does possess the crucial pathlifting property. n + 112) is a homeomorphism onto U.= a and 0(0) = o.l} and V= SI\{1}. The homeomorphism ¢: JrI(SI. Let an: U ~ (n .1/2. and we let f3n: V ~ (n. n + 1) be the inverse of OJI(n.l!2. 1] = OJI (U) U OJI (V). n + 1/2) be its inverse. but it is extremely important to see exactly what it is we still have to do to make the idea of the proof into a watertight argument. we consider OJ. we would not have come this far unless everything was going to be all right in the end. 1] ~ IR such that OJ· 0.. in the definition of ¢. tj+ d) lies in U or in V.I (V) = Un E ~(n. that the answer we get does not depend on the choice of a path in the homotopy class [a]. we ¢([a]) Step 3: Step 4: Step 1 we show that ¢ is a homeomorphism. Similarly.112. < tn = I such that OJ([t j . To prove the pathlifting lemma we set U= S\{. n + 1/2) and for each n E 7L. OJI (n . we use the homotopy class of a and not a itself. Each of OJI{U) and OJI{V) is open in [0. Now OJI(U) = u nE 7L (n . We can therefore define the winding number of ato be 0(1).. such as Sutherland (1981). 1] ~ IR with OJ 0. 1) ~ 7L is well defined by = 0(1). n + 1).fundamental group about OJ for our purposes is this: given any path a: [0. since 0(1) = 1 we must have o(l) E 7L. I) ~ 7L that we seek will then be defined by ¢([aD = o(l). We carry out this waterproofing process in four steps: 0 Step 1: Step 2: we show that given any path a: [0. n + 1). 1J ~ SI there exists a unique path 0: [0. but more fundamentally (pun intended).is a lifting of a and is unwound along IR. The path 0. so that given a we can get a unique 0. Is the previous sentence true? In fact it disguises the problems that we must still face: it may be difficult to show that ¢ is bijective (what do you think at this stage?). It follows from Lebesgue's lemma in the theory of metric spaces that there exist points 0= to < tl < . we show that ¢ is bijective. or in . If a: [0. 1] ~ IR with OJo 0= a. I] ~ SI is a path with a(O) = 1 there exists a unique path 0: [0. and it will only remain to show that ¢ is bijective. We shall not prove Lebesgue's lemma in this book. but it is easy to find in standard texts on metric spaces. Does it matter which path in the homotopy class we lift to IR? Of course. Secondly. furthermore. show that ¢: JrI(SI. lifting lemma. I] and [0.
The unit square [0. The result we ylifting lemma.and 1: in SI. lj h: [0. The homotopy h given by the homotopylifting lemma has the property that (00 h(s. tk+ I]' If 0. Similarly. n + 1) for some n E Z and we may define CT: [tb tk + I] ~ ~ by CT(t) = Pn o(t).+ 1: and IJF(O) and 1JF(l) = T(I) + a(1). t)for all t E [0. IJ we have h(l. = CT(O) = O. Since (0(0) E U we have (0([0. 1/2 ~ t ~ Then ( 0 . n + 112) for some nEZ. define a path IJF: [0. again by use of Lebesgue's lemma. n say. If we now assume that we have defmed on the interval [0. Now for all t E [0.and CT(O) = 0. I]). 1] ~ ~ such that (00 h = hand h(O. Notice that we have no choice for the definition of at each step if we are to ensure that (0 CT = 0. we eventually define CT ([0. I] x [0. t) = 0 for all t. Repeating this procedure. h(. 0 ~ t ~ 1/2 = T(2t . t) It E [0. I] x [0. I) = r and so q>(tJ = T(l) = h(l. 0) = O. 1). 0) = h(s. Similarly. t) = I = h(l. til ~ ~ by a(t) = ao o(t). IJ ~ SI is a homotopy between the closed paths oand r. and now q>(o) = n = q>(1:). 0).([th tk+ I)) ~ U then a(tk) E (n . and we may define CT: [th tk + I] ~ ~ by CT(t) = an o(t).112. Furthermore. but the image of the connected set {(1. I]} under the continuous map h must be connected. 0) = o(s). By uniqueness. a 0 0 0 a 0 Step 2 This is also accomplished by lifting the problem to need is as follows. we show that we can extend the definition to [0. Given two closed paths 0. if o([tb tk+ I]) ~ V then a(tk) E (n. 1] ~ ~ by lJF(t) = CT(2t). tID ~ U and we can define a: [0. It follows by the pathlifting lemma that h(. and so is a single integer. Step 3 This is nice and easy. We shall not give a detailed proof of this lemma because the ideas needed are the same as for the pathlifting lemma. IJF = (0. 1] must be divided into smaller squares. each of which maps into U or into V. 1] x [0. IJF = 0.Calculations of the fundamental group 205 Armstrong (1983). td. h(O. (so that h(O.and 1: are closed paths in SI and are homotopic by means of a homotopy h. 0) = CT and so q>(o) = CT(1) = h(1. ~.1) + CT(1). and so CT is unique.+ 1:) . t) E (01(1) = Z. N ow suppose that 0. and h then defined square by square. I)) there is a unique homotopy h: [0. with corresponding paths CT and T.
• The complexities of the proof should have convinced you that the computation of fundamental groups from first principles is likely to be a formidable task. candidate for ll"J (S\ 1) we still encountered some subtle problems in establishing that our candidate was indeed correct.is a closed path in IR. 1] ~ SJ be defined by an (t) = exp (2n Jri t). This will be the subject of the following two chapters. It remains to show that cp is injective. 0) = OJ o(s) = a(s) and h(s. so that 0. we develop some more basic theory of the fundamental group. Set h = (t) H. Let a be a closed path in X x Y at (xo. Then h(s. Yo). 1] ~ !R by H(s. a powerful tool exists: Van Kampen's theorem. and rather simple. Check the details of this assertion. Proof. How are operations on spaces reflected in the calculation of the fundamental group? For products of spaces (with the product topology) the answer is very satisfactory. We leave it to the reader to check that on(1) = n so that cp(an ) = n. If cp([aD = 0 then 0(1) = 0. [qa}) where p: X x Y ~ X and q: X x Y ~ Y are the projection maps. Even though we began with a very plausible. How are we to compute the fundamental group of a space when no likely candidate presents itself to our imaginations? Fortunately. xo) (Y. The proof of the proposition consists of a verification of the credentials of cpo Is cp well defined? If a = OJ then do we have pa = pOJ and q a = q OJ? Let h be a homotopy from a to OJ. t) = (I . We see that h is a 0 0 homotopy from a to the constant path at 1. Then ll"J ll"J Let X and Y be topological spaces with Xo E X and Yo E Y. Yo» is isomorphic to the direct product X ll"J (X.to the constant path at O. I) = OJ(O) = l. Then H is a homotopy from 0. let an: [0. and finally the proof is complete. Define a homotopy H: [0. Yo).t)o(s). (xo.The fundamental group Step 4 It is easy to see that cp is surjective: given n E If. There is an obvious candidate for an isomorphism: we take cp([a}) = ([pa}. all we need to prove is that if cp([aD = 0 then a is homotopic to the constant path at l. roposition. 1] x [0. Then ph and qh are homotopies from pa to pOJ and from qa to qOJ respectively. For the remainder of this chapter. . (X x Y.
e. h'(s. but within each class the end points are fixed). Let nx denote the set of fixedendpoint homotopy classes of paths in X (so we take all possible pairs of end points. e' be the constant paths at Xo and Yo respectively and suppose that hand h' are homotopies from pa to e and from qa to e. t) = Xl' If there is a fixedendpoint homotopy from a to w we write a = w (f. and further such that for all t E [0. {3) = q a • q {3 as closed paths in Y at yo. simply connected.) and that iii· w = 0"'(1) (f. Show that W· iii = 0". defined when w(l) = w'(O).).). Show that H is a homotopy from a to the constant path at (xo. w' be paths in X such that a = w (f.p.e. Complete the proof by showing that qJ is surjective.e.).p. r) is defined. X ~ [Rn is starlike if it is path connected and if there exists Xo E X such that for any y E X and for all t with 0 ~ t ~ I. r are normalized paths such that the normalized product a· ({3. 0) = a(s) and h(s. Show that the product of homotopy classes [w][w'] = [w· w']. t) = (h(s. Let a. {3. Explain why this shows qJ to be injective. and such that a· a' is defined.e. 9.p. Define H: [0.(0) (Le.p. check that p (a· {3) = pa· p{3 as closed paths in X at Xo and that q( a . t) = Xo and h(l. l} ~ X such that. 3. h(s. for all s E [0. A fixedendpoint homotopy from a to w is a continuous function h: [0.3 I.). 5. 1) = w(s) .). • Before continuing with the theory of the fundamental group. Show that X is simply connected if and only if given any two paths a and w in X with a(O) = w(O) and act) = w(l) then a = w (f. 6.t)y E X. Yo).p. 7.p.). we return to the properties of paths. 1]. Show that W· w' is defined and that a' a' = W· w' (f.p. Let X be a path connected space. Show that starlike subsets of [Rn are 2.e.). 1] ~ X x Y by H(s. 1] x [0.p. w. Is qJ injective? Let e. t)). Yo). Show that fixedendpoint homotopy is an equivalence relation on the set of all paths a in X with fixed end points a(O) = Xo and a(l) = Xl' Show that if a.e. IJ. a' = w' (f. h(O. then (a· {3) • r is defined and a· ({3 • r) = (a· {3) • r (f. 4. Let a and w be paths in X with a(O) = Xo = w(O) and a(l) = Xl = w(I). is well defined. I} x [0. a'. we have txo + (1 . t). and in the following exercises we consider a homotopy notion more closely related to path products than based homotopy.Calculations of the fundamental group 207 Is qJ a homeomorphism? If a and f3 are closed paths at (xo. .e.
° Theorem. Since homotopy is clearly a very good idea. Iff is homotopic to g: (Y. that is. x o) induces a homeomorphism f*: 1fj (Y. Yo). that is to have isomorphic fundamental groups. A pointed homotopy fromfto g is a homotopy H: Yx [0. What is the effect on the homeomorphismf* of a homotopy fromfto another pointed map g? The answer is that there is no effect at alL Let f. I] ~ X fromfto g such that H(yo. x o) between fundamental groups. x o) we have obtained a function f*: 1fJ (Y. thenf(a· f3) = (fa) • (ff3) [Exercise: check this!] and therefore f* is a homeomorphism between the fundamental groups 1fj (Y. x o) we have [faJ = [gal. and we arrive at the result by extending the construction of the fundamental group to continuous maps between spaces: a continuous map becomes a homeomorphism of groups. pointed homotopic maps induce the same homeomorphism on fundamental groups. x o). and so given raj E lli(Y. Yo) the class [fa] is well defined in 1fJ (X. Yo) ~ 1fj (X. Yo) ~ (X. Yo) ~ (X. If a: [0. for all raj E 1fJ (Y. If a: [0. that is. xo) is a pointed continuous function then fa is a closed path in X at Xo. x o) defined by [aJ H [f aJ. xo). 1] ~ X defined by h(s. On the other hand. x o) by a pointed homotopy then h = g*. t) = H(a(s). This is indeed the case. we should attempt to perceive its effects wherever possible. xo) and (Y. Yo) ~ 1f) (X. Our conclusions are summarized in the next result. Yo) ~ (X. x o). A pointed map f: (Y. Now if a. A task that we might hope to have made easier by this labelling is that of distinguishing one space from another: spaces with different shapes should carry different labels. with ~ t ~ 1. Yo) ~ (X. Yo) ~ (X. I] ~ Y is a closed path in Y at Yo and f: (Y. 1] ~ Y is a closed path in Y at Yo then h: [0. Let X and Y be topological spaces.The fundamental group Homotopy equivalence The fundamental group construction attaches an algebraic label to a topological space that records some information about the shape of the space. g: (Y. . xo). Yo) ~ 1fj (X. homeomorphic spaces might be expected to have the same label. Xo E X and Yo E Y. x o) be pointed maps. we need a pair of basepoints. t) = Xo for all t. }:[a] = g*[a] and so}: = g*. 1] x [0. To discuss fundamental groups. Yo). From f: (Y. [Exercise: exactly what has just been asserted? Why is it true?] If h is a pointed homotopy from a to win Y then fh is a pointed homotopy from fa to fw. Given the pointed spaces (X. f3 are closed paths in Yat Yo. a continuous function f: Y ~ X is pointed if Xo = f(yo)' Iff is pointed we write f: (Y. Yo) ~ (X. t) is a pointed homotopy fromfa to ga and so in 1fj (X. The space X and the selected point Xo together form the pointed space (X.
If there exists a retraction of X to A we say that A is a retract of X.. E. and f . If fis a homeomorphism then it is certainly a homotopy equivalence.Homotopy equivalence 209 We say that a pointed map f: (Y. Proposition. • The subspace A of the space X is a deformation retract of X if there exists a retraction r: (X. Then for each XED we may define rex) to be the point on the boundary SI of D crossed by the line segment fromf(x) to x (Fig. x o) ~ (X. Therefore r* is surjective from HI(D. . x o) ~ HI(A.1). Yo) ~ HI (X. J.(A. is indistinguishable. x o) ~ (A. x o) ~ (A. now find a contradiction by combining Proposition 9. x o) such that riCa) = a for all a E A. A retraction of X to A is a continuous map r: (X. 9. Let D be the unit disc in 1R2. and we leave you to check this. x o). x o) and r*i* is the identity on H. xo) is a homotopy equivalence if there exists a pointed map g: (X. x o) ~ (A. x o) be the inclusion. Let i: (A. Let g be a homotopy inverse for show that g* is an inverse for f. x o) is a homotopy f*: H I ( Y. (1. The Brouwer fixed pOint theorem. xo) ~ (X. x o) is an isomorphism. We leave it to you to Especially interesting is the setting in which there exists a homotopy equivalence between a space and a subspace. Brouwer.. If A is a retract of X then the retraction r: (X. Proposition. x o) ~ (Y. (1.0». If f: (Y.I is a homotopy inverse for f. x o) and a pointed homotopy h from ir to id x.(S'. x o) induces a surjection r*: HI (X. Then r is a retraction of D to S'. Proof. with our vision blurred by homotopy. Suppose that no fixed point exists. that is D = {x E 1R2111XII ~ l}.6.5 with Theorem 9. Yo) ~ (X. Yo) and pointed homotopies between gf and id y and betweenfg and id x. there is some xED such that f(x) = x. The subspace is then a simpler version of the space that. Proof. Proof. Yo) ~ (X. Let A be a subspace of X and choose a basepoint Xo E A. x o) is then a homotopy equivalence. Then any continuous mapj: D ~ D has afixed point. Exercise! • As an application we can prove a famous result of L. The map g is a homotopy inverse for f.0» to H. • equivalence then f. The inclusion i: (A.
the subspace A (a line segment joining two circles) is a deformation retract of the space X: visualize the retraction! Fig. Yo) ~ (X. 2. 9. g(yo» is equal to the composite "I "1 .t is a homotopy ir = id x.2. Y). 9.3.The fundamental group f(x) Fig. y)/(~ + y2) is a retraction and h«x.4 See Fig. y) X +y I . Letj. zo) ~ (Y. j(yo» ~ "I (X. 9. show that (jg)* andj*g* are equal. g(yo» such that g*: (Y. If g: (Z. (b) In Fig. Yo) and j: (Y. 9.y)lx2 +y2=1}.t) = t(x.y)ll~x2+Y~4} and A={(x.2 (c) ises 9. g: Y ~ Xbe homotopic maps (with no assumption that the homotopy is pointed) and let Yo E Y. Then rex. Show that there exists an isomorphism (J: lli (X.1 pies (a) Let X={(x. Yo) ~ (X. 1. y) + ~(x. x o) are pointed maps. y) = (x.
Yo) and HI (X. show that for any Yo E Y the fundamental groups HI (Y. . Yare path connected spaces and f: Y ~ X is a homotopy equivalence.3 3. 9. We say that f is a homotopy equivalence if there exists a map g: X ~ Y such that fg = id x and gf = id yo If X.Homotopy equivalence 211 x Fig. f(yo}) are isomorphic. Let f: Y ~ X be a continuous map.
We proved a result on 1lJ of a product of spaces and so can show 1l'J(torus) == C~ ~ C~. An even closer look will suggest what it should say and will also give us a sketch proof. the details of the formal proof are (initially) somewhat less important. If you glance forward to where this theorem is stated in full and in detail. Again it is often the case that the lecturers have used the results explicitly or implicitly so many times in research projects that it becomes selfevident to them that these theorems are important and hence it can be extremely difficult to explain their importance to someone meeting them for the first time. These spaces are good examples as we can understand them well. we are lucky. With the Van Kampen theorem. Complicated theorems occur at numerous places within mathematics courses at this level and students often ask themselves why such theorems are deemed necessary. . the first priorities are to guess what the result will say (before we meet its detailed statement). We can build geometric models of them. a simple enough space you might think. The fact that these details are technically nontrivial merely shows that the 'obvious' often does need justification! Why do we need a Van Kampen theorem? The state of our art at present is that we have calculated a large class of fundamental groups with great ease only to find they are all trivial! Working considerably harder we were able to prove that 1l't of a circle is an infinite cyclic group. it should become clearer why theorems of the Van Kampen type are needed. you may wonder why such an apparently technical theorem is needed. Once you understand the idea of the proof the details can be seen merely as the 'justification of the obvious'. but we cannot handle 1l'J of an nIeafed rose. We have been developing the theory of the fundamental group and if we pause to take stock of our present knowledge (and ignorance).10 Van Kampen's theorem In this chapter we will examine the statement and proof of the Van Kampen theorem. Writing out a formal proof is another matter and will be dealt with later on. nor 1l'J (genus g orientable surface) if g > 1. to see why the result is true and why certain technical restrictions are necessary.
into parts. The nleafed rose is got from n circles. Suppose we have a path connected space X divided up into two nonempty parts U and V. One important thing they have in common is that they can be built up from simpler 'building block' spaces such as lines and discs. We hope to decompose w. tn _) E. the basepoint of X. Let us be a bit more precise (Fig.2). as we get paths in U and V that need not be closed.1 Suppose w: [0.) Let us consider a picture (Fig. it is clear that one could go around w until one arrives in U 11 V. 1J so that (putting 10 = 0. one would hope. and then continue in V until one is back in U 11 V and so on. and so should be able. tn = 1) each wi [th ti+]] is a path wholly within U or wholly within V. (If U 11 V were empty then X would not be pathwise connected. . . loop w Fig. This idea would seem to decompose the path w into paths totally within U or totally within V. Ib . How are we to ensure this is possible? For instance. In the picture. up to homotopy. it is not at all clear how to go further in calculating fundamental groups!) To answer this question we will need to examine how loops in a space are related to loops in parts of a space. 10. if we take . 10. The genus g orientable surface can be built up from a 2gleafed rose and a disc.Van Kampen's theorem 213 we can picture them.1) that will help the discussion (but always beware of arguments that rely too much on pictures!). We are working with pointed spaces since we want to calculate HI (X). However. this decomposition would be taking place not within H] eX) but within the fundamental groupoid of X. it is therefore best to assume that x o. We will assume that we can find points I]. each of which is either wholly in U or wholly in V. . 10. to compute their fundamental groups. [0. 1] ~ X is a normalized closed path at Xo. This raises the question: can one calculate HI of a space if one knows H) of the pieces it is built up from? (If the answer is no. is in U 11 V.
y): y ~ O} then we will be in difficulties. there is a number oS> (often called the Lebesgue numbering of the covering) such that any subset of Y of diameter less than oS is wholly contained in some set of the covering. which is the closed unit square in ~\ are compact spaces.1 J into infinitely many subintervals to get the sort of property we want. We can thus pick points to = 0. We cannot. Why? To handle the general case we need to recall the idea of compactness.t (V)." (Every open covering of a compact space has an open subcovering. compose an infinite family of paths in the fundamental groupoid! Our problem is that V and V in this example are not open. there is a finite subfamily {U AI ' •• . tiJ is either within U or within V. more precisely.. . _ I> ta W: ° ° . that is spaces where the open sets are defined by a distance function or metric. _]..) We need a further piece of information about compact spaces or. V are open sets. First the definition: a space X is compact if. (Try to prove this yourself using techniques from analysis. tn _.) Applying these ideas to our situation. 10. say V E = {(x. . we find that there is an oS> such that any interval of length less than oS is contained wholly within w .oS} and V E = {(x. however. then we would be able to get away with finitely many subpaths. given any family U = {U A : A E A} of open sets such that X = u {U A : A E A}. y): y ~ O}. . VAn} of U so that X = Uf= 1 VA. If we increased V and V slightly in size. lJ and 12 = 1 x I. y): y ~ .1 (U) and w .I (V). U= {(x. or look it up in a book on metric spaces or on point set topology. compact metric spaces.2 [0. IJ is compact. The open subsets of [0. since we would need to chop [0. t sin(l/t». tn = I in 1 such that each ti . Thus both 1 = [0. IJ are unions of open intervals and [0. so are w . lJ ~ ~2 defined by w(t) = (t.ti _ 1 < oS and hence w[t. Now clearly if we write Wi for w i[t.) Now if U. .I (U) or w. t].n Kampen's theorem Fig. If Y is any compact metric space and U is any open covering. (The HeineBorel theorem from real analysis can be thought of as saying: X k ~n is compact if and only if it is closed and bounded. V = {(x. y): y ~ oS} for any small oS> 0.
. [Ii _ 1. . . In any case we must have U. In fact.. (This requirement could have been avoided to some extent if we had worked with the fundamental groupoid rather than the fundamental group as our principal tool. (2) if Wi and Wi+ 1 are both paths in U. (At this point.) We can rewrite the decomposition of w. has to be within Un V. but for the moment we are searching for a result on fundamental groups and the paths Wi need not be closed. I.Van Kampen's theorem composed with the linear homeomorphism between [0. Vi is wholly within V. if one of Wi and mUnV. and Un V path connected if we are to be able to fulfil these conditions. we can disregard this technical problem as we know that if we work with unnormalized paths and then normalize at the end of a calculation. as Wi+ as each Vi Vi is homotopic to the trivial path. a path Vi joining Xo to W(ti). til. it is useful to think of working with unnormalized paths as otherwise we would need to bother with the fact that composition of normalized paths is not associative. then Vi is a path This last condition overlaps with condition (2) since if i'Vi and I are both in Un V then Vi' under both conditions. Each Ui is a loop either in V or in V. This . similarly if they are both in V. The fundamental groupoid is useful. We will use the convention that Ui is in VU) where Vel) = Vor V. V. where Ui = Vi _ I WiVi . Wi+ I (3) is in U and the other in V. This takes place in the fundamental groupoid of X. To get around this difficulty we proceed as follows: pick for each i = 0. Vi is wholly within U. up to homotopy this time. then we have ti ti_ 215 J and and so we have written W as a composite of paths wholly within U or within V. we will briefly examine this possibility later. n.) We can group this decomposition as within X. these paths are to satisfy (1) Vo and Vn are the trivial path at Xo. the end result is homotopic to the end result obtained by consistently normalizing and taking into account the lack of associativity that results.
. iU(n) [un]u(n)' This gives us a result that we can summarize as follows. The induced homeomorphisms iu: iv: 1t'J(V)~ 1t'](V)~ 1t'](X) 1t'j(X) in their turn induce a homeomorphism where as before * indicates free product and the lemma states that this homeomorphism is onto 1t'J (X). [un]x = iU(l) [uj]U(l) . The images Of1t'l(V) and 1t'j(V) in 1t'j(X) generate that group. we would not need to decompose it at all. or equivalently what the kernel of the epimorphism looks like. and perhaps to questions of invariance of the list under homotopies of w. if we have presentations (x : r) of 1t'J(V) and (y : s) of 1t'] (V) then it says that 1t'J (X) is a quotient of the group with presentation (x u y: r us). then it is ... We note that in the decomposition above [wJx = [u1Jx[u:Jx' . If we try to interpret this geometrically. This may seem a strange place to start. similarly with X replaced by V. The general rule should always be that if we make choices we should see how much the end result depends on the choices. For instance. This means that generators for 1t'J (V) and 1t'J (V) together generate 1t'j (X) and that the relations between these generators still hold in 1t'1 (X). An alternative and slightly more formal way of putting this same result is as follows. we can see it as the uncertainty about uniqueness or otherwise of the list [u]JU(l)' . but if we write iu: 1t'] (V n V) ~ 1t'J (V) and iv: 1t'J (V n V) ~ 1t'] (V).Kampen's theorem allows us to introduce a neat piece of notation: if w is a loop in X we write [w]x for the corresponding equivalence class in 1t'j (X).. rna. This means that the induced homeomorphism is defined by idwJu = [wJx. Suppose w was a closed path in V n V.. This is already a positive result. V or V n V..[unJU(n) due to the choices that had to be made in its formation. What it fails to say is what additional relations are needed. but it will point out the necessity of thinking clearly. One of the subtlest choices we made was deciding in which open set a given Wi was.
we saw earlier that G is precisely the pushout HI (U) It] (u*n V) H1 (V) so really the two points are the same. similarly for jv(z).~ 1t1(U) 1t 1( V) l. we should 'kill off' within HI (U) * HI (V) those elements of the form [W]ul W ]v 1. The square 1t1(Un V)l. so this is real progress to cutting HI (X) 'down to size'. We do this by dividing out by the normal subgroup generated by them. In other words this subtle choice is really basic: any closed path in U r l V 'decomposes' trivially in two different ways giving the same element of HI (X). so there is an induced homeomorphism.~ 1t1(X) is commutative. This group G is often much smaller than the free product. An efficient way to do this is to pick a set of generators z. for HI (U r l V) and to kill off the words juz • jvz .I for z E Z within the presentation (x u y: r us) of HI(U) * H1(V). giving a group G with presentation (x u y : r U s U {(Ju(z» UV(Z»I}). that is there is an epimorphism from G onto HI (X) induced by the inclusions of the subspaces. 2. from to HI (X). say. Of course. Here we have written ju(z) for any element of F(x u y) that maps down to ju(z) under the natural composite epimorphism F(x uy) ~ F(x) ~ (x: r). As iulwJu = iv[w]v for any [w]un V E HI(U r l V). . This first observation suggests the following points: 1.Van Kampen's theorem 217 clear that the two elements [w]u =ju[w]un v and [w]v =jV[w]Un v give the same element of HI (X). the pushout (J say.
However. simply using proof by induction..1] together gives yet another subdivision..) The first step is to investigate what happens when we add another point in the decomposition. if need be. You may recognize the basic method here."'. Suppose we decompose w using to. 10. We might expect that. further reducing the size of the group we can construct and also that of the kernel of the resulting epimorphism. then of course g = g'. If the resulting element g" of G can be shown to be equal to g and to g'. now consider u' as being a closed path in V and delete ti from the subdivision (provided ti"* I of course). tn and also using t~. ta . To prove invariance under choices. t 2] • vi = U'.]u E 1l'1(X) and after adding t' we choose a v from Xo to w(t') and find closed paths wi [ti. we would seem to get two elements g and g. in G. we would find a new geometrically inspired set of relations to add in. Thus the independence of the element defined by the decomposition on the choice of subdivision is clear and the mechanism set up to handle the problem of the closed paths with a 'split personality' allows one to pass parts of a closed path from U to V and vice versa. is in Un V and v is chosen to be within Un V then u' will be a closed path in Un V. [u]u' [u']u = [u • u']u = lUi]. Are these elements the same? Putting the two subdivisions of [0. say. wi [t l . each time. The advantage of g" over g' is that the subdivision used can be obtained by adding in finitely many extra points into that used for g and hence if we can show that one extra point added into the subdivision changes nothing we can show that g = g'. Given that we have already within the presentation of G allowed for the double identity of such closed paths (within both 1l'J (U) and 1l'J (V» we can. . giving [u]u' [w]u in place of [uau. If the section [t'. What actually happens is different from this. Suppose we take t' E [t f _ 10 til and without loss of generality we assume that w maps [ti]. If we do not change the subdivision but merely the choice of the Vi then again the same idea works (Fig.3). . Vi_I' w. t.h 1'] • v = u.n Kampen's theorem We next turn to the other choices that were made in determining the decomposition of w. The reasoning behind this is the following. so nothing has changed. say v . In the decomposition before adding 1'. and is slightly surprising. and also g' = g" as hoped for. prove invariance under basic elementary changes of choice (here adding in or taking out a point. making one realize that the end result is a very elegant and pretty theorem indeed (as well as being an extremely useful tool. we have [U. til into U. and previously by performing one of the Reidemeister moves or Tietze transformations). t10 ••• .
Next we subdivide the square [0. U.Jv iv[u. we find that Ui' Uitj . Using the compactness of [0. I] 7 X such that h: Wo = Wt.+ 1 == way we chose them) and we obtain within the group G. Note that we have used the 'iu(a) = iv(a)' condition in the form luiu = lviv. I] x [0. so the algebraic structure in G faithfully reproduces the geometric structure within 7l'j (X). Sj) L _ _ _ _ _ _ _ _ _ _" .. say. Ui is in U and Ui+ 1 is in V. then Vi and vi are in Un V (by the Writing Uj == Vil WiVi.u. Ui+) == ViWi+ tVi+ I. • U.Van Kampen's theorem 219 Fig. u. We suppose given h: [0. iU[Ui]U iV[Ui + dv = iu[uJu iV[Vi • u. I] we find an £ > such that any set of diameter less than £ is mapped by h either wholly into U or wholly into V. I] x [0. u.+ 1 . 1] into rectangles small enough that the diagonal length is ° (t.WI then the resulting elements of G are equal.. + av dv = iu[uJu lviv[Vi' uJUl' v iv[u.+ I within X. this is unremarkable as the homotopies happen equally well within U and V.3 ViWi+ jVi+ j. == Vi_ IlVivj. I] x [0. but if.+ = iu[u..+ I]V.Ui' (Vi' v. The final thing we have to check is that if Wo .]u ivJulvi • vi]un v ivlu.]u iv[U.I]V = iu[u. 10. If both Uj and Ui + I are in one only of U and V.) .
n1 . • Un . tl> t 2 ... 1] x {S2}. V. I] = (Xi (Fig. agree as iteration will do the rest. tn = 1 as our subdivision of (0. etc. and Un Vas follows: w' where Uo "• U 1 • • .4 Although we will write W' in terms of loops in U or V.1 tn1 n o As h is a fixedendpoint homotopy the ends of the square are both mapped to Xo. 10.. . 1]. I» . Ui = Vi • Wi • Vi + . As we hope to show that the elements go and gl E G corresponding to Wo and w.. 11 1[0. ..4). sj. so our picture is 2 3 n. . Fig. hi [0. say 111[0. I] x {si} and 111[0. We assume as always that paths Vi from Xo to wet. . are equal. U. so when we decompose w' we will use paths v. = 0. .Kampen's theorem So less than e. that is we pick numbers to = 0. Because of this we will assume m = 1. 10. b i=L . We have seen that the element of G that we obtain is independent of the choice of the paths. ••• . .1] x {sj+d.. 1] x {sJ. Using to = 0. it will as usual be better to do the calculation in the fundamental groupoids of X. obtained by composing Vi with the path wi {ti} x (0.) have been chosen. Sm _ I> sm = I so that each rectangle is mapped within one of U or V. .. 1] x {O}. we get decompositions of Wo = hi [0. tn _ b tn = 1. it will be enough to prove that the elements corresponding to any two successive restrictions of 11.
open. or wholly within V (possibly within U (\ V). 8a is the identity homeomorphism on HI (X). and U (\ V are nonempty. the assignment of g to [w] gives a homeomorphism w.Van Kampen's theorem 221 within U or within V since the homotopy is just a reindexation via a linear homeomorphism of [0. and pathwise connected. but any decomposition defines the same element as none of the choices makes any difference. g was well defined and 8(g) = [w]. Let X = U (\ V where U. Thus the path ai iii + I determines in the fundamental groupoid of U (resp. the isomorphism being induced by the inclusions iu. Thus we have investigated all the indeterminacy linking the element g to [w] E HI (X). . The way in which a is defined makes it clear that a8 is the identity on G. Kampen theorem (first form). V. We state it formally in two different versions. Pick Xo E U (\ V. 0 that 'splits' the epimorphism 8. of U (\ V) the same element as Wi' Composing either side with the Vi takes this back into the fundamental groups. resp. so HI (X) == G. That is. in other words. We have proved more since if we compose [w] and [w'l in HI (X) it is clear that by picking wand w' in these classes and composing one can pick decompositions already used for wand w' to get one for wow'. 0 wi ~ a't I i+ 1 ~ Wj r" 0 I This rectangle is mapped (by h) wholly within U. We have arrived informally at the result known as Van Kampen's theorem. In fact there was none. Then the following diagram is a pushout square of groups. so the elements go and gl are equal. The path ai ai+ 1 is obtained by mapping three sides of the (i + 1)st rectangle using h: 0 w. iv. of V. Given [w]. 1].
Construct for yourself a formal proof of either form of Van Kampen's theorem.iv·(z) is a word in F(y) representing Jv*(z). Xo E U n V and suppose Jrl (U. (V. ises I. xo) has a presentation (z : t). For instance. You can choose whether or not to follow the outline we have given. on purpose. xo) has a presentation (y : s). Do not think their version will necessarily be fundamentally better than yours. go to your library and search out books which contain a formal proof of the Van Kampen theorem. Pick x o) has a presentation (x : r). not give one ourselves. phrases. but suggest the following exercises. Having attempted the previous exercise. V. or notation from their version. but . 2. We will. Try to improve your proof with ideas. Compare their methods with yours. Jrl Jrl Let X = Un V as before.Van Kampen's theorem . xo) has a presentation (x u y : r U s u {Uu*(z» Uv*(z» ~ I : Z E z}) where Ju*(z) is a word in F(x) representing Ju*(z) and. (U n Then Jrl (X. We have given above all the necessary building blocks for constructing a formal proof of the Van Kampen theorem.~ 1tl(X) Van Kampen theorem (second form). Criticize both your proof and theirs. you may wish to verify the pushout property of the square (*) directly rather than showing that Jrl (X) is isomorphic to the algebraically constructed pushout.
Find a reference to Van Kampen's original paper. 4.Van Kampen's theorem 223 3. Try and read his proof. . you should expect to gain something. Several times we have worked within the fundamental groupoid. Try to formulate and prove a groupoid version of the theorem (then compare with the paper by Brown (1967». as the authors tend to have had years' more experience in the game of writing proofs. at least in style.
our calculations will more often than not refer to earlier examples. the parts into which we divide our space must be open and path connected. how are we to apply it? As with any result. but later on we will leave such details increasingly up to you. n = 1. As we will see. If we want to calculate Jil of a space. Just as in a proof by induction. A oneleafed rose is just a circle. Initially we will give nearly all the details of this necessary adjustment. . this sometimes requires an adjustment to the obvious choice of decomposition. namely the class of a path going once around the circle. is infinite cyclic. we must check that the conditions necessary for the theorem to hold are satisfied in that example. X. As we do this. The Van Kampen theorem relates the fundamental group of a space to those of its parts. We start with simple examples and get progressively to handle more complicated spaces. we need to start somewhere and in this situation our starting point is the proof that Ji1 of the circle. we will first examine the fundamental group of a one. The other thing that we will need is a starting point. In order to apply the Van Kampen theorem. we have to find an open decomposition in terms of simpler spaces. that is free on one generator. culations of fundamental groups fed rose So as not to be too ambitious to start with. before we try to apply it to a given example. so Jil (oneleafed rose) is free on one generator.11 Applications of the Van Kampen theorem Having got the Van Kampen theorem.and a twoleafed rose. Sl.
HI (N).1 n = 2.1). V2 = V 2 U N. Thus N looks as shown in Fig. A twoleafed rose (Fig. 11. 11. we find that VI is a deformation retract of Vb so that HI (VI) == HI (U I ) and is freely generated by the class of the loop ai_ We will abuse notation and write . We let N be a small open neighbourhood of Xo in X itself.Calculations of fundamental groups 225 x Fig. The space X = VI U V 2 . we note that X = VI U V 2 and N = VI n V 2 • The final criterion is our ability to calculate HI (VI)' HI (V2 ). so we cannot apply the Van Kampen theorem to this present decomposition. Fig.4). 11. we need only look at one of them. 11. 11. Already here the main ideas for the general case occur. These two subs paces are open and path connected.2). as is U I n V 2 = {xo}. As VI and V 2 are clearly homeomorphic. 11. 11. say VI (Fig. The necessary adjustment is easily made. Fig.3 Now let VI = U I uN. and the induced homeomorphisms between them. but VI and V 2 are not open subspaces of X. has an 'obvious' decomposition (Fig.2 We see that U b U 2 are path connected.3. This space contains VI and by retracting the two 'antennae'.
. . Thus the Van Kampen theorem provides a proof of the geometrically obvious result. This is not an open decomposition.. 11.. but will continue with the proof for the moment. i= 1. then back around ai' and so on. This is as we might have guessed since a typical loop in X will clearly wander around the two leaves one after the other.am then TeI(Xn ) will be a free group on n generators corresponding to these loops. then around a 2 112 times. 11. but as it again involves the addition of a small open neighbourhood of the basepoint. so again let N be a small open neighbourhood of the basepoint Xo in Xn. This neatly absolves us from the task of calculating the induced homeomorphisms TeI(V I n V2 ) ~ TeI(Vi ).5).4 Similarly Now N itself contains {xo} as a deformation retract. corresponding to a word starting a7'dl 2 a7' .plications of the Van Kampen theorem V= Antennae Fig. . .. This gives Xn = Xn _ I V Y with Y a circle. What about the general case of an 111eafed rose? The sort of intuitive reasoning just used would suggest that if Xn is an llIeafed rose with loops around the 'leaves' labelled ai' . perhaps going around ai' say n l times. that is a free group on two generators corresponding to loops around the two 'leaves' of the rose. The proof by induction of this is not too hard.) . so Tel (N) is the trivial group.2 since there is a unique homeomorphism from the trivial group to any group. we will include a sketch of it. We have now collected up all the information so as to calculate Tel (X). We note that an nleafed rose Xn can be written as the onepoint union of Xn _ I with a circle (Fig.(How small is small? What were the necessary facts about the small neigh bourhood N that was used in the twoleafed case? How do we know that such a neighbourhood exists? We will return to these points shortly. The Van Kampen theorem gives a presentation Tel (X) == (ah a2 : 0).
2. 11.Calculations of fundamental groups 227 v o Fig.. then XII = VI U V 2• VI n V 2 = N.• an : 0) as expected. The basepoint is a deformation retract of N (if N has been chosen small enough!).• all _ I : 0) by our (unstated) induc tion hypothesis. we get 1. X" _ I is a deformation retract of VI.6). x Fig.. 3. V~ = Y u N. is a deformation retract of V~.5 Set VI = XII _ 1 U N. How can we be certain that this can be avoided? We have a lot of latitude in our way of modelling the nIeafed rose as the property that we are examining only depends on homotopy type. What about 'small'? Clearly picking an arbitrary neighbourhood of Xo in Xu might result in one of the leaves being within N. 11.. Hence we may assume that ~I is chosen to be a subspace of /R' with Xo the origin and each leaf a vertical unit circle (Fig. Since we have an explicit description of XII and of N. Y. "1 (V~) == "I (Y) == (a" : 0) since Y is a circle.I}. (Draw diagrams to illustrate this and to convince yourself it is so. "1 (VI) == "I (XII _ I)  (a 1• . whilst the circle. Now let N = {x E XII: IIxll < O.) Collecting up facts. the trivial group. 11..6 .g. 2) The formula given by Van Kampen now gives "1 (XII) == (alo . our requirements can be checked explicitly if so e. "I (V1 n V == (0 : 0).
0) II = lin} so that C n is a circle through the origin centred on the point (lIn. . C n n C m = {O}. connected graph If r is a finite graph with a o vertices and a.pplications of the Van Kampen theorem desired. v. orient. that is C n c B(O. a l = 11. Pick a maximal tree T in r.. 11. we find that a set of free generators for HI (T) can be constructed using the path classes [aJ. a2 = eba2cIeI. it then crosses ai and goes back to v in the maximal tree (Fig.a o + I is the number of edges not in a maximal tree. it must contain an open set of the form B(O. say a" . and label the edges not in T. N. v. where n = a l . ai. where ai is a path that goes from the basepoint. bewareit is not that simple. out in the maximal tree to the initial vertex of the oriented edge. How shall we interpret these generators? We suppose the basepoint of r is chosen to be a vertex. Its structure is known...lf. Retracing the process back into r. i = 1. Using the previous example. If one looks at any neighbourhood. Care is required: as an example of what could go wrong consider the spaces C n = {x E ~2111x  (lin. . e) n Xc Nand N contains infinitely many loops of X. of 0 in X. e) n X since it is an open neighbourhood of 0. n.7). Here a o = 8. The question of the structure of HI (X) would lead us too far astray.. and HI(T) == (ai' a2. G4 =fha4g. G3 = a3gf. They become a set of generating loops of HI (Xn). Now let X= Un d'. but there is some N£ E N depending on e such that for n > N. edges. 21n < e. 0) with radius lin. Fig. than r is homotopy equivalent to an nIeafed rose. 11. as a subspace of the plane. and n = 4.l . . Now al = ebalc1e'.7 .I. G3' a4 : 0) is the free group on four generators. we find that HI (r) is free on n generators.an' Now see what happens to these edges when we shrink T to a point. Cn. but if you feel like trying your hand at it. If m:t= n.
so by our earlier results. We have already noted that the torus can be realized as a product Sl x Sl. b= U 1 "'"'" p a p . this does not work. Fig. We do not. need to be that careful with the neighbourhood. we know that ni(torus) is the product of two infinite cyclic groups C= x C=. as the intersection is empty and the first set is not open.8). a good choice of basepoint will a p a r(.x) p b ~tc J/~ '" a r 'b • b . and the other to be the interior of the rectangle. so as to gain some intuition of wha t to expect. this time.Calculations of fundamental groups 229 Again. the choice of which is immaterial as the space is path connected. We first pick a basepoint. Such a decomposition is more easily visualized from the rectangle than from the other picture. We have to look for a suitable open decomposition satisfying the conditions of the Van Kampen theorem. but by 'thickening' the first set up to an open neighbourhood we will get around these difficulties. b. This will enable us to check our calculation independently. However. it is therefore free Abelian on two generators. we start with almost the simplest example possible. The torus can be obtained by gluing a rectangle as we have seen in Chapter 4 (Fig. Of course. a torus. 11. . The idea is to take one of the sets in the decomposition to be the union of the two loops a.8 Note that the boundary of the rectangle gives an embedded twoleafed rose. 11.
the trivial group. Looking again at the pushout diagram coming from the Van Kampen theorem.9 After the identification of edges. so before choosing a basepoint. 11.plications of the Van Kampen theorem make the presentation simpler. What about (Fig. 11.10). then VI radially projects onto VI and U I is a deformation retract of VI: (see also Fig. We thus have Now for V2 = X  V~: VI_ This is an open disc.9). an infinite cyclic group generated by a loop around the hole. 11.? This is an open (rectangular) annulus c Fig. the boundary. V. and so VI (1 for this we take the interior of the rectangle. a b Fig. 11. VI' of the rectangle is a twoieafed rose embedded in the torus. we have .10 Thus 1£1 (VI (1 V 2 ) == 1£1 (SI). We let VI = X . let us see how the argument will go. namely 1£1 (V 2 ) == (0 : 0). U I = boundary of the rectangle.centre of the rectangle.
h]) as a presentation of an Abelian group. we are using the same letter for a loop and the class it determines. by the direct sum. Choosing this basepoint so that r sends jl (c) onto ahah. for some n (Fig. when writing things additively). however. A. EB. choose wellit will payoff later'.b) is the free group on generators a and h. Alternatively by considering the presentation (a. The structure of the group ]fl(torus) does not depend on the choice of Xu. (X) We do not need to worry about j2* since ]fl (V2) is trivial. At this stage. it pays to note that this neat expression for ji *(c) occurs only if r(xo) is the starting vertex of the boundary surface symbol. an.. The surface can be represented by gluing a polygon using the corresponding standard surface symbol. but will give a sketch. A general surface (without boundary curves) is not very different. pays off in the simplicity of the resulting presentation. which is a word in edges ai' .11). We can now finish the calculation. that is C~ x C~. that is jl *(c) = ahaI b.(V} J~~ 1t. .' . (As usual.) The radial projection r pushes the loop jl(c) onto the composite path abah. We leave the details to the reader.Calculations of fundamental groups 231 . h]). . that is C~ EB C~ (where by tradition we replace the product. so Xu is for us to choose. 11. h : 0)ah. h : [a. we do. Van Kampen's theorem gives us a presentation that is (a. it is sufficient to calculate jl (c) in terms of a and h. b : [a. what we have here is its Abelianization which will be the free Abelian group on two generators. 1* 1t. need to describe jl *. rather than ahah or haha or something more complex still. As ]fl (VI n V 2 ) is infinite cyclic generated by c. This accords with our previous calculation since if F(a. The moral of the story is 'When a calculation does not depend on a choice but cannot be made without that choice. we write everything additively to get (a.. x.
j.V j • Repeating the same argument as before (details should be written down by you!).. In particular.11 Take V j = boundary of polygon (as a graph embedded in the surface) and V j = X .12)... 11... of genus g (with no boundary) then n = 2g and A = alb1allbll . We will illustrate this with a simple example showing how to adapt the previous argument to this case.. lag.. . the projective plane. for g = 1. Instead of being simply an nIeafed rose for some n. b g : raj. Sj = IfJll. . ••• . a:). j . Thus if a surface has genus 1 then its fundamental group is Abelian.... If we allow boundary curves and thus cuffs. A = aba. and == C 2 . of genus g (again with no boundary curves) then n = g and A = ai~.j Cj d j cil c 2 d 2 c. we get Jrl (X) == (ab . b j .b. . of course. . the picture is slightly more complex.. . Everything proceeds as before until we identify Jrj (VI) == Xj (Vj) where V j is the embedded graph coming from the boundary of the polygon.centre. a g. Jrl (1fJl2) == (aj : a~) torus with two cuffs The surface symbol is. Sg. so a:..plications of the Van Kampen theorem Fig.. Jrj (Sg) == (a j . If X is a nonorientable surface. 11. Vj is a slightly more complex graph (Fig. Thus if X is an orientable surface. b j] . an : A). the cyclic group of order 2. . . Take V 2 = X . so Jrj(Tg) == (aj . Choose a basepoint Xo that projects radially onto the start vertex of A. agbgagjbg. a g : aT . T g .. bgD.
r) is a free group on 2g + r . We will write df = CidiCi I. d~. 2. 11. Prove that if ll'1 (Tg) or ll') (Sg) is Abelian.r) is free on g + r . Write out the proof that ll'1 (1fll2) is cyclic of order 2. i = 1.) face groups We are now in a position to ask.r) an orientable (resp. (Although this may seem a trivial exercise. does this mean that X and Yare . 2. the obvious question: can ll'1 distinguish between different surfaces? In other words. giving us ll'1(torus with two cuffs) == (a. 3. since the relation allows us to write in terms of the other generators. 11. if X and Yare surfaces.Surface groups 233 b Fig. then g = 1. ll'1 (VI) is not free on the set of edges of the polygon.1 1. it is important to gain an idea of why this 2torsion is there as later we will be examining how to construct a space whose fundamental group is a cyclic group of given order n.r (resp. which is a free group of rank 3. nonorientable) surface of genus g having r cuffs. Sg. b. The calculation now continues more or less as before. Show that ll'1 (Tg.1 generators.1 generators while 1X'1 (Sg. b] d~d~) d. and better still to answer completely. although of course still free. d~ : [a. (Which Tietze transformations are we using here?) We will denote by Tg.12 Thus. for the generator of ll'1 (VI) corresponding to the two cuffs. and ll'1 (X) and ll't (Y) are known to be isomorphic.
ag' b g : [a b bJ .. s 1. Recall from Chapter 6 that if a group G is given by a presentation (X: R). Once we have determined orientability or lack of it. rag. bg] is already a consequence of the other relators and so may be deleted. . then substitute and rearrange to get (a b g . 2.. Jrl (Tg)ah == (ai' b l . .I. . that is g . . Here it is simpler to use the alternative method of handling Abelianizations of presentations. ag =b .r and S2g. Introduce a new generator b and set it equal to a l + . a g : aT . ag. Y EX}).r and S21V are not homeomorphic. 2. [a b a2]. so Tg... therefore. yet we know that orientability is an invariant for surfaces. . . •• . that Jrl (Tg)ah is a free Abelian group of rank 2g. the Abelianization G ah of G has a presentation (X: R u {[x. namely by writing everything additively. we can find out which surface X is from the genus which can be read off from our calculations of the torsion free rank of Jr1 (X)ab.) We have.. . a~.. g}). To attack this we will launch into an analysis of the surface groups Jr l (Tg) and Jrl (Sg)' We first look at their Abelianizations. b : 2b = 0. This gives You can now choose to attack this presentation formally with the methods from the theory of finitely generated Abelian groups or you can use additive Tietze transformations as follows. j = I. y]: x. [ag. This suggests that we ask the question in a stronger form by requiring the surfaces to be without boundary curves. . rag.. Jrl (Sg)ah == (ai' . bg]). + a g . (You should check the justification of this using Tietze transformations. . . aj]: i.plications of the Van Kampen theorem homeomorphic? If we allow surfaces with boundary the answer is clearly 'no'... eliminating ag gives an Abelian group of torsion free rank I and a single generator of order 2. bg].••  a g _ I)' Finally.al  .lcopies Thus given a surface X. ([ai. since both Tg.. .r have fundamental groups free of rank 2g + r . Thus the product of commutators [a h b l ] . we can tell if X is orientable or not by looking for an element of order 2 in Jrl (X)... .
g.t k sdY(:.akhkak I) uak and The Abelianization of G = Jrl (Tg) was earlier seen to be free of rank 2g. as in Chapter 8.1.Surface groups 235 Hopefully your curiosity has been aroused and you will have asked whether Jrr can detect cuffs.k) Ei = 0 for = Sk  1 for k = 1.h.. We need to calculate dr/da/.r = [ai' hi] ... for each k. E 2g . At present we only have information about its Abelianization.. ..I or from an Sg. .g})... can we tell if X has cuffs or not? If Jrl (X)"h has an element of order 2. h g]) produce a way of eliminating the relator. .1 =({Si1. .. Thus our problem is to decide whether Jrl (Tg) is free or not. ag.. . . but the easiest one for us is to use elementary ideals and other ideas from Chapter 8.ai'hi ' . [ag. To avoid problems with the multiple use of symbols.'h..'. . thus giving us a free group? Perhaps it seems unlikely. 1 .r with n = g + r . ag' h g) ~ G and G ~ G ah. .ti: i= 1. hg : [ai' hi] . hi . . but suppose Jrl (X)"h is free Abelian of rank n. then X must have been a nonorientable surface without cuffs. and dr/dh/. the homeomorphism obtained by 'killing off' rand Abelianizing G respectively. This gives the following sequence of elementary ideals in Z(C~)2g: i < 2g . The images of the differentials are sd: sdy(k) dak and = 1. n even. we set s[ and ti to be the images in (C=)2 g of at and hi respectively with y: F(a ... but can we be sure? There are several different methods we could use. Let r = a.1. aghga.. that is. Might not some cunning subtle sequence of Tietze transformations applied to (ai' h" . if we know X is a surface and have some presentation (not necessarily the nice one we have given above) of Jrr (X). hk _ 1](1 . [ak _ .. This gives :. . we could have obtained this answer from a T" with n = 2g or from a T 2g ..r with n = 2g + r . .
7r] (Tg) is not free and its lack of freeness is detectable using the machinery we have at our disposal. no other rank would be possible). and E 2g = Z(c=fg • If G had been a free group of rank 2g (and given our knowledge of its Abelianization. then the elementary ideals would be E j = 0. as seemed likely. 11. the boundary of the polygon: 7r1 (VI 11 V 2 ) ~ 7r1 (VI) ~ 7r] (U]). i < 2g. at least write it out in 'sketch' form now. aces with finite cyclic fundamental group We saw earlier that 7r](1FD2) ~ C2• Can we find a space X which has 7r1 (X) ~ C 3 ? Clearly X cannot be a surface since our calculations showed that jp2 was the only surface with nontrivial finite fundamental group. (If you have not written out this calculation in detail yet. A means of attacking this question is to rerun rapidly the ideas that lead to the 2torsion in 7r] (1FD2). This suggests that we take a triangle and identify the three sides as if we had a symbol aaa (Fig.plications of the Van Kampen theorem the ideal generated by the images of the differentials. 11.13). Now rerun the proof that 7r1 (1FD2) ~ (a : a 2 ).13 . p p a p Fig. but with 1FD2 replaced by X and of course with the obvious modifications. E 2g = Z(C=)2 g • Thus.) The a 2 = I comes from the image of the loop in the annulus VI 11 V 2 when projected onto Uj.
r s . Xn : r l . . To get X. X . Using an inductive construction it is now easy to construct a space having any (finitely presented) group G as its fundamental group. • . using the Van Kampen theorem as before. if we have a space Y and an element r E Jrl (Y). rs ).. G. defined by: if X E Sl C D2. a p a a a p Fig. r) in which r has been added to the relations in the presentation.f(x). There are no obstacles to using a similar construction to get a space X= X(Cn) with HI (X) == Cn. Try to write down the details with a proof. using the Van Kampen theorem.. . the same method works but the justification that it works is technically slightly more demanding as . xn : r l . S. Here they are (Fig.Spaces with given fundamental group 237 No prizes will be given for instructions on how to construct a space with HI a cyclic group of order 4. . we use a map f that sends a generating loop C E HI (Sl) into a 3 E Jrl (VI)' In general. Recall that this is done by taking VI ~ ~ and then gluing. Given a group. . We can think of this construction as starting with a oneleafed rose into which we glue a disc D2 via a suitable map f defined on its boundary circle. . (The problems of handling nonfinitely presented groups are merely ones of technique. . . that your construction does in fact give what is required. 11. then we can choose a closed path a = ar:S I ~ Y so that [a] = r and can use a to glue a disc onto Y to form a new space X = Yu D2. Xn : r l . • . at least for groups G with a finite presentation (Xl> . we will take another look at the construction of the space with Jrl (X) == C 3 that we examined in the previous section. • • • . using an equivalence relation. a cyclic group of order n. can you construct a pointed space X with HI (X) == G? To gain some insight as to how we might do this.14). . 11. .rs ) then HI (X) has a presentation HI (X) == (XI' . that if HI (Y) == (XI' • . One can show. • .14 aces with given fundamental group The previous set of examples raises an interesting problem.
plications of the Van Kampen theorem it uses a more powerful form of the Van Kampen theorem than we have introduced above. (a.16 The element b 2 in this group is represented by a loop a2 going around the righthand loop twice. a p Fig. 3 . Y.16).) As an example. 11. We set r l = a (Fig. We use tX:2 to glue in a disc to get Y z (Fig. the group of symmetries of a triangle. 11. This group has a nice presentation. b: a 3 = 1) (Fig. r2 =b = (ab)2. namely D3 == 2 . b : a 3 = b 2 = 1). as <t representing loop.15). b : a 3 r3 = b 2 = (ab)2 = I). then Y 2 = X(C3 ) V !fD2). Let r b be a twoleafed rose Fig. (If we write as before X(C3 ) for the space we constructed earlier having Jrl(X(C3» == C 3 . 11. The space X = Y 2 D2 then has Jrl (X) == D 3 • (At this point the artistic powers of the authors come up against insurmount .17). The fundamental group of Jr1 (Y2 ) is C 3 * C z == (a. a map a l from 51 into the lefthand circle of r. Gluing in a disc via this map gives a space Y 1 with Jrl(Y1) == (a.15 The relator a 3 E Jrl (r) has. 11. 11. we will see how to construct a space X having Jrl (X) == D 3 . The element r3 = abab is represented by a loop lX:J going along a then along b then along a and again along b.
11.K after one application of the Van Kampen theorem.. so the notation is justified.. K. 11.17 able difficulties. for some n. an+l being a l . so we have not attempted to draw this space even schematically. that the Reidemeister moves left the group unchanged even though they changed the presentation. but this will help later on when we come to analyse the various fundamental groups involved and the homomorphisms between them. as being a subset of 1R3. We shall use aI' a2' . the Alexander matrix of the presentation.10. We saw. ) e group of a knot In Chapter 6 we introduced the group of a knot and gave two different ways of obtaining a presentation of that group from a diagram of the knot. Suppose we are given a knot K which we think of. The decomposition will need some care in construction. Thus a typical crossing would look like that shown in Fig. we will write G(K) for the group that results from a diagram of the knot. however. 11. Moreover. So what is the geometric interpretation of this mysterious group G(K)? Why is it invariant? Can we see this directly? In this section we prove that G(K) is. We also saw that the Alexander polynomial of K could be obtained from the Wirtinger presentation of G(K) using the Fox derivatives and a sort of Jacobian matrix.19. As there. We next deform the knot to replace each such crossing by one which looks as shown in Fig.18 with. as labels following on around the knot with the orientation. This notation presupposes that G(K) is an invariant of the knot and not just something concocted from one particular diagram.The group of a knot 239 a p p Fig. as usual. Orient the knot and label each arc of the knot in the usual way. in fact. that is 1f1 (1R3 . the fundamental group of the complement of K. with most of the knot within the . the Wirtinger presentation comes from a fairly simple decomposition of 1R3 .
N is a . it slides down the first vertical with its faces parallel to the axes in ~3 and then it goes along /3i and back up again to the z = I part of at + \_ By taking £ small enough we can ensure that ~3 .plications of the Van Kampen theorem Fig. 11. 11. but at each crossing we drop a vertical to the z = 0 plane followed by a segment labelled f1j in that plane from the foot of the vertical line ending at to the foot of that starting ai + INext we remove from ~3 a 'tunnel' neighbourhood N of K_ This is best thought of as follows_ Imagine that we have a closed cube of side £ fixed to the knot but able to slide along it. When it comes to a crossing section as above.19 z = I plane.18 Fig.
N as the union of two open sets.K so that HI (1R3 .N. 11.11. That part of the tunnel containing aj will be labelled Ai whilst that containing /3. V 2 = {z < e/2} .20 deformation retract of 1R3 . Now VI is halfspace with some tubeshape tunnels removed.20). an where Fia.N.21 . It has as a deformation retract an nIeafed rose with loops a j .i will be called Bj (Fig.N) IS isomorphic to HI (1R3 .The group of a knot 241 Fig. We take VI = {z > O} . We are now ready to describe a decomposition of 1R3 . 11. • • • .K) which we want to calculate.
24. On the other hand VI is an open halfspace with trenches cut in it (Fig.23 Thus 1l"1 (VI n V 2 ) = free group of rank m. It is therefore contractible and 1l"1(V2 ) = (0: 0).24 . 11. if the crossing is as shown in Fig.22 Finally VI n V 2 = {O < z < e/2} . Hence 1l"1( VI) = (aI' . Fig. 11.23). where m is the number of crossings (Fig. Fig. 11.25.22). . or aj I ai 11 ajai if the crossing is as shown in Fig. 11. 11.plications of the Van Kampen theorem ai is a loop passing under the tunnel Ai that contains at (Fig.. an : 0). A typical generator will correspond to a circuit around a hole and so its image in 1l"1 (VI) has the form ajai1 lata. 11. Fig..21). 11.N is an infinite plate with m holes in it. 11.
This is precisely the Wirtinger presentation of the group G(K).L) (= G(L» as required.L.K to 1R.3 factors via some torus Tl sitting in 1R. We can now see geometrically (or perhaps globally might be a better term) why G(K) is an invariant of the knot K. 11. us knots A knot K is said to be a torus knot if it can be embedded on the surface of a torus.3 . If we had wanted to we could have picked a homeomorphic image of the open decomposition {VI' V2 } of the complement of our original knot and could have applied exactly the same argument to obtain Jl'1 of that complement.25 It is a good idea to try to draw diagrams to illustrate thisa diagram drawn by you will be more use to you than one drawn by us! Now the Van Kampen theorem can be applied and we find a presentation where a typical rk is as above.3 .3 ~ 1R.3: . Suppose K and L are equivalent knots.3 .3 such that rpK = L. that is we have K: S ~ 1R. This homeomorphism therefore restricts to one from 1R. This induces an isomorphism between Jl'1 (1R. Recall that this means there is a homeomorphism rp: 1R.Torus knots 243 Fig.3 . The above argument also gets around the possible objection that we picked a very special representative of the knot so as to be able to use Van Kampen's theorem. Of course the description would have been a lot harder.K) (= G(K» and Jl'1 (1R.
When we calculated the knot group of these we found This is typical of torus knots. In this case it is simple to realize K in a nice. Consider a solid cylinder C with m line segments on its curved face equally spaced and parallel to the axis. Fig. c"). In fact we examined many of the invariants of these (2.26 As K is a knot one finds that m and n can have no common factor. 5)torus knot. 11. .3)torus knot as we have already mentioned (Fig. c : 0) we have K*(a) = (b m . This Im.27. 11. 11. If the ends of C are identified after a twist of 2Jr(nlm) then we obtain a single curve 1m n on the surface of a solid torus T. and Jr1(T1) == C~ EB C~ a torus knot K determines a pair of integers (m. regular way. n)torus knots as they formed the family able to be drawn in the form shown in Fig. it is easy to see that the trefoil is a (2. In fact as we will see. n)torus knot: We introduced these informally earlier and saw that the cinquefoil or pentoil was a (2. As an example. that is they are coprime. The knot K thus goes around one of the two generators m times and around the other n times. Since Jrl (Sl) == C~.plications of the Van Kampen theorem SI ~ TI C ) 1R3. n) such that writing Jr1(SI) = (a : 0) and Jr1(T1) = (b.26).n is an (m.
n Vb deforming to T .t m •n ).Nand (1R3 .) Since em.g. G(tm n) ~ 1l'1 (1R3 . Similarly its image in 1l'j(1R3 . We leave it to you to calculate Alexander polynomials and will instead use other methods in our analysis. a group theoretic analysis is worth doing.T is homeomorphic to a solid torus with one point removedtry to see why. a typical generator being the class of the centre line em n of Lm n' The space T .Torus knots 245 Fig.N) is am.N slightly to get open sets Vj. which results from the m parallel strips on the cylinder being joined after a twist of 21l'(nlm) (Fig. as promised. thus is infinite cyclic generated by a loop b through the 'hole' in T. Similarly (1R3 . easy to work out what they are. etc. the image of em. Picking a small s.N i~ a solid'torus with a groove gouged in its surface and so it deforms onto a central circle (marked a in Fig. The complement of N i'n the surface of T is an annulus.N) ~ (a: 0). (The Alexander polynomials are difficult to work with except when the smaller of m and n is 2. e.T) .N) ~ 1l'j (1R3 .T) . pretzel knots) is that the embeddability of these knots on a torus allows a direct application of Van Kampen's theorem to 1R3 . b : am = b n).T) .28). Lm. we carve out an sneighbourhood N of tm n in 1R3.T) . It is.27 The advantage of torus knots (and similar things hold for various others of the families of knots. 11.N.n) is infinite cyclic.tm n' We consider tm n embedding as above on the surface of a solid t~rus T sitting in 1R3. (In fact 1R3 .N . Applying Van Kampen's theorem then gives a knot group. We expand T . These groups c~:mtain much of'the information about the knot. V 2 deforming to (1R3 . V 2 with VI n V2 a neighbourhood of L m . n in 1l'1 (T .N. Of course 1l'1 (1R3 . 11. Thus 1l'1 (Lm. We must now adjust our decomposition to be open so as to allow ourselves to apply the Van Kampen theorem. however.tm n) ~ (a. As with the surface groups.n.28).) 1l'j(T.N) is b n. n makes m circuits of T. 11.
28 alysis of the knot group Gm. So we can hope to find out more about Z(G) by examining Z(GIN). a km E Z(Gm n) as a consequence. am E Z(Gm n) and so is isomorphic to C m * C m the free product of two cyclic groups. At this point. if it is nontrivial. Recall that if G is a group. This subgroup N is thus in Z{Gm n) and so is normal.n. for any k. any element x in Z(G) which is not in N will give a nontrivial element x = Nx in Z(GI N). n). We start by looking at the centre of Gm. and we can find a nontrivial normal subgroup N in Z{G). Let us write N = {a km : kEd"} for the subgr~up generated by am. b = Nb for the elements of Gm. Taking care over the notation we will write a = Na.n (This section uses slightly more group theory than the rest and can be omitted.) We will write G m • n for G(tm . we should explain the strategy of our attack. am commutes with g E G}. we can try to repeat the process. 11. that is Z(G) am both a and b. If this is trivial then Z{G) = N. so and of course. The .n Fig. and why we are looking at this quotient. If we are trying to write down Z(G) for some group G. its centre Z(G) is the subgroup formed from the elements that commute with everything. and the corresponding quotient group Gm.plications of the Van Kampen theorem //+1 m.nl N clearly has presentation Now since = {c E G: cg = gc for all = b n .nl N.
Such an expression commutes with a only if it has at both ends and commutes with b only if it has b at both ends! Hence a nontrivial element cannot commute with both a and b. This statement seems intuitively ~easonable. we have a (*) is of finite order and has length less than p. This completes the proof that Z(G m n) = (d n ). If p is arbitrary. b n = 1. so at each stage we have a better chance of obtaining complete information on the centre of the resulting quotient group.Analysis of the knot group Gm. however. so Z(Gm. we have no interaction between a and b. In em * en any element has a unique reduced form where o =:. of Xl and Yp equal to zero. the subgroup generated by~. that is we must have one.n) = N. so it seems unlikely that there are any elements in em * en which are central. each of which involves only one generator. If XI = 0 a similar argument applies.:: o =:. Are these all of the elements of finite order? We will try an induction on p. so we will have to probe a bit deeper. 0 < Yi < n and all Xi. assume that any g of finite order having length less than p is a conjugate of a power of a or b. in the presentation of em * en the generators a and b are controlled only by relations am = 1. Any conjugate of a power of a or b will also have finite order. If Yp = 0. Yj E lL. By the induction hypothesis this element is a conjugate of a power of a or b but then g will also be such a conjugate.:: Yp Xl < m. em * en. ' We next look at elements of finite order in this group. 0 < Xi < m for for i > i < P < n. This is not sufficient to be a proof. Suppose g E em * en (~ G m nlZ(Gm n)) has normal form aX! • • • bYp and that it has order r < 00. In our case there are no problems since Z(Gm nl N) is trivial. but not both. so we have that any element of finite order in em * en is conjugate . We have already seen that for p = I all elements of finite order are powers of a or b. If p = I this must mean that g is a power of a or b. The wo'rd (normal form of gt cannot be shortened unless the normal form starts and ends with the same letter. n 247 group GIN is in some sense smaller and simpler (even where G and GIN are infinite).
tp. n} with m < nand (m'. If K is an alternating knot and G(K) has a nontrivial centre. (b) A space Zg is constructed (rather as in (a» by gluing a disc into a vertical circle in an orientable surface. 1I"1(Zg)ab. Use Van Kampen's theorem to calculate the fundamental group of Y.Applications of the Van Kampen theorem to a power ofa or to a power ofb. 11.29. then K is equivalent to t 2 .3 . T g .30. As min(m. n) we can theoretically determine (m. " By combining geometric arguments with this sort of analysis. of genus g having no boundary (thus Y = ZI)' Calculate 1I"j (Zg) . n') with m' < n' are given and are distinct then G m n ~ G m . Fig.) ses 11. If we make em * 'en Abelian. (a) A space Y is constructed from a torus by gluing in a disc along one of the 'vertical' sections as shown in Fig. we get em $ en which is of order mn. n). if we are given G m n' Thus if {m.29 2. The torus knots projections.n). (ii) X is the linked union of two circles (Hopf link) as shown in Fig.7 1. . and the elementary ideals of 1I"j (Zg). 2.q with p > 2 and q > 2 have no alternating (If you think a bit you will see that result I implies result 2. these results can be pushed further. We will not do this here but will note that Neuwirth (1961) proves the following pretty results on torus knots: 1.X when (i) X is the unlinked union of two circles. 11. up to order. n' and tm nand tm' n' are distinct knot types. but result I itself is much harder to prove.n) = mnlmax(m. 11.2r + 1 for some r. In particular the elements of finite order in G m nlZ(Gm n) cannot have order greater than max(m. Find the fundamental group of 1hI.
30 (i) 3.. . am and B one for S2 in edges b l . b n .. Let SI and S2 be two surfaces and S = SI # S2 their connected sum. If A is a surface symbol for SI in edges aI' .Analysis of the knot group Gm. find a presentation for 1rl (S1 # S2)' . n 249 00 Ii} Iii} Fig. • • • . 11.
Given a neighbourhood U of x E X. Does the algebraic structure of HI (X. if for each point x E X there is a neighbourhood U of x such that pI(U) is the disjoint union of open sets Vi of Y. Questions about X can often be transferred into more tractable questions about a covering space. A covering space preserves the local features of X but its overall topological structure is generally more simple. which says that a subgroup of a free group is itself free. 12. so that p maps each Vi homeomorphically to U. and the Schreier index formula for the number of free generators of a subgroup of a free group. es 1. Further. overing maps and covering spaces Let X be a topological space. A map p: Y ~ X is called a covering map.12 Covering spaces The fundamental group HI (X. Our applications will concentrate on the covering spaces of graphs. which we use to prove two theorems about free groups: the NielsenSchreier theorem. we have .1). x o) of a topological space X at a basepoint Xo describes the variety of possible closed paths in X at Xo. each x E X has a neighbourhood U and Y contains a number of copies of U mapped to U by the covering map p (Fig. Informally. x o) corresponds to a collection of topological spaces that map to X: the covering spaces of X. the projection map F x X ~ X is a covering map. we show how to discover free generators for the relation subgroup of a group presentation from its Cayley quiver. and Ya covering space of X. The identity map X ~ X is a covering map. all spaces are assumed to be path connected and locally path connected. if F is any space with the discrete topology. More generally. In this chapter. x o) reflect further features of the topological nature of X? In this chapter we shall see that the collection of subgroups of HI (X.
2). 12. 1). The map ~ ~ SI given by t ~ exp(it) is a covering map.Covering maps and covering spaces 251 Fig. 2. A useful picture of this map is to imagine the real line twisted into a spiral (Fig. as we did in Chapter 9 for the computation of HI (SI. 12. 12.1 pI(U) = U fEF {f} x u. IR Fig.2 .
so let us set ZO={ZE Z I~(z)=h(z)}.h and hover p. Covering spaces allow the solution of some maplifting problems. with the points of 7l. We shall show that Zo is both open and closed in Z.Covering spaces 3. Divide ~2 into squares.3 here for n = 3. Uniqueness lemma. Edges coloured x are mapped around one copy of Sl. Then the Cayley quiver of the presentation (x. . Let p: Y ~ X be a covering map and Z a path connected space. if they exist.h and h agree or differ. Let X consist of two copies of Sl joined at a single point. and edges coloured y around the other. y} is a covering space of X. at the corners. Let us suppose that f has two lifts. A small neighbourhood U of x E Sl has preimage pl (U) consisting of n disjoint copies. Proof. Lifts need not exist nor. u Fig.3 4. y : 0) of the free group with basis {x. and map each square to the torus according to the usual gluing instructions. a lift of f over p is a map 1 X ~ Y such that pol = f. Then two lifts of a map f: Z ~ X over p either are equal or differ at every point of Z. We want to examine the points at which. 5. The details of this proof are suggested as exercises. 12. 12. many maplifting problems admit unique solutions. if p: Y ~ X is a covering map. illustrated in Fig. Given maps f: Z ~ X and p: Y ~ X. The map P: S' ~ Sl given by exp(iO) H exp(niO) is a covering map. x 7l. However. need they be unique. ~2 is a covering space of the torus.
of CT with 0(0) = Yo. 1] ~ Y with 0(0) = Yo and p 0. Show that such a decomposition CT= CTI • • • • • CTm is indeed possible. To do this in detail. * 0 Exercise 2. contained in a single open set as reg uired.2 Pathlifting lemma. 0 a Exercises 4. the image of CT need not be contained in a single Ui. Let X be a topological space with a chosen basepoint Xo. but we can nevertheless write CT as a product of paths CT = CT1 ••••• CTm with OJ: [ti' ti + ~ X (where tl = 0 and tm + 1 = I) with the image of CT. Given any path CT: [0. it follows that 1. Try to generalize step one of the proof of Theorem 9. but poi.(z). and let p: Y ~ X be a covering map. Conclude the proof of Theorem 12. Since (z) h.(z). Now suppose that z E Zo nfI(U). If the image of CT is contained in some single such U b then we define 0. there exists a unique path 0: [0. Let Z E Z\Zo. • 12.by choosing some Vij ~ pI (Ui ) and setting 0. 1] ~ X with CT(O) = Xo and Yo E Y with p(Yo) = x o.(z) and h(z) lie in the same component of fI(U). (z) = p h.1) tells us that there is at most one lift 0. • . then 1. where U is some neighbourhood of fez) in X. Why will this suffice to prove the uniqueness lemma? Remember that Z is path connected. Use the continuity of f. In general. .Covering maps and covering spaces 253 Exercise 1.k and to deduce that there exists a neighbourhood of z contained in Z\Zo' Deduce that Zo is closed.2 by showing that a lift oi can be chosen for each CTi such that CT = 01 • • • • • om is a lift of CT. Since p is a covering map. (z) and h. Show that z has a neighbourhood contained in Zo nfl(U) and deduce that Zo is open.6.= CT.(z) lie in distinct components of pI (U). 1. X is a union of open sets U i such that pI (Ui ) is a disjoint union of open sets V ij such that pi Vij: Vij ~ U i is a homeomorphism.= (p I Vijr 1 CT. 0 Proof The uniqueness lemma (12. you will need some metric space properties of the interval [0. 5. h Exercise 3. 1].
j (x o). so may be subdivided. 1].Covering spaces Homotopylifting lemma. 1] ~ Y with a(O) = Yo and a(1) = yj. and to prove it we can use the same ideas. 1] ~ Y such that p H = Hand H(O. Suppose we pick Yj E Y with p(yj) = xo. 1] x [0. and the top and righthand side into p. Let a be a closed path in Yat Yo such that p a is homotopic in X to the constant path at Xo: that is. The unit square [0. x o). there is a path a: [0. it provides a neat algebraic description of the covering spaces of X as part of a profound interaction between group theory and topology. XO and Y. for we see that to any covering space there is associated a subgroup of 1fj (X.and homotopylifting lemmas indicate how the local similarities between X and a covering space Y allow processes occurring in X to be mimicked in Y. if necessary. As our next step. xo) induced by p is injective. and the class of [a] is trivial in 1fj (Y. Definitions of these lifts can be made so as to agree on common edges of the smaller squares. x o) carries P*(1fj(Y. • The path. we consider the implications of the choice of a particular basepoint Yo E Y. then there exists a unique map H: [0. 1] x [0. Let X. Then the homeomorphism p. Yo) ~ 1fj(X. into smaller squares each of which is mapped by H into an open set of X on which a lift to Y is easy to define. If H:' [0. 1] ~ X is a map such that H(O. and so H may be constructed. 1] ~ X be a homotopy from po a to the constant path at xo. We content ourselves with a sketch. t) for all t E [0. An important consequence of this idea now follows. 1] x [0. 1]. [a] lies in the kernel of p*. with basepoint Yo E Y chosen so that p(Yo) = Xo. but by the preservation of connectedness (Theorem 3.6) we see that on all these three sides if must be constant at Yo. Since Y is path connected. t) = Xo = H(l. Yo). 1] x [0. This result is a twodimensional version of the pathlifting lemma. Theorem. Thus H is a homotopy from a to the constant path at Yo. Let H: [0. This association is not haphazard. Let X be a space with basepoint Xo and p: Y ~ X a covering map. Clearly pea is a closed path in X at Xo and conjugation by the class [p e oj in 1fj(X. Consider how if maps the unit square into Y. 1] is a compact metric space. 0 Proof. Yo be as in the pathlifting lemma. Proof.: 1fj (Y. The lefthand side maps to Yo.4 is the key that unlocks the structure of the collection of covering spaces of a space X. t) = Yo for all t E [0. and let H be its unique lift over p. yj» 0 0 . The remaining side gives a lift of the path p a over p and by the pathlifting lemma this must be a itself. • Theorem 12.
.4.t: and fit into a commutative diagram of fundamental groups 1 . namely compactness. yo»[a].t. Yo))· 1 Proof. then H = P*(JrI (Y. xo). y))ly E pI(XO)} Jorm a complete conjugacy class oj subgroups oj Jrl (X. suppose that H ~ JrI(X. Given a map J: Z ~ X carrying the basepoint Zo E Z to X o. zo» ~ P*( Y. To establish sufficiency. Let (X. and Zo E Z are fixed such that p(Yo) = Xo. Let p: Y ~ X be a covering map. Yl))' This argument establishes the following refinement of Theorem 12. xo) is conjugate to P*(JrI (Y. We have seen that the condition is necessary. By the pathlifting lemma. Set Yl = £l(1). there exists a unique path a in Y with £l(O) = Yo and p a = a. . J(zo) = X o. Maplifting theorem. x o be a pointed space and p: Y ~ X a covering map. Yo E Y.Covering maps and covering spaces 255 to P*(Jr1(Y.(JrI(Z. with basepoints Yo E Y and Xo E X chosen so that p(Yo) = Xo. and l(zo) = Yo· Then the maps p*. What might we hope for as a more general maplifting result for covering spaces. ) Then the subgroups {P*(JrI(Y. there exists a lift oj J over p if and only if f. we shall define directly. • The pathlifting and homotopylifting lemmas were proved by exploiting a particular feature of the unit interval and unit square. Yo»). from which it follows thatf.(Jr1(Z. Consider an arbitrary point 1 . 0 Theorem. where no special features of the domain can be called on to help? Suppose that p: Y ~ X is a covering map and that J: Z ~ X has a lift Z ~ Y. zo)) ~P*(Jrl(Y'YO))' It is one of the beauties of covering space theory that this necessary algebraic condition for the existence of a lift of J over p is also sufficient. Conversely. suppose that there exists a class [a] such that H = [ar Ip*(Jr1(Y. we assume basepoints Xo E x.. Yo)). that is.
If "i' is the reverse of r then G'"i' is a closed path at Zo E Z and soJ:([a· "i']) E P*(Jr1 (Y. 1] ~ Z be a path with a(O) = Zo and a(1) = z.4). the righthand side must be mapped to Yo. 12. and then that 1 is continuous. Yo». and that r lifts to a unique path /3: [0. Let rbe a closed path in Yat Yo with [f (a· "i')] = [p r] and lift the homotopy f (G' "i') = P r to ( Y. Now foG is a path in X with f a(O) = xo: by the pathlifting lemma there exists a unique path a: [0. Therefore f (G' "i') lifts to a closed path in 0 . 1] ~ Y with /3(0) = Yo and p /3 = for. G exists because of our assumption that all spaces are path connected. 12. We define 1 by setting I(z) = a(l) (Fig. We first show that the definition is in fact independent of the choice of path. and let G: [0. 1] ~ Y such that G(O) = Yo and p a = foG. The lifted homotopy maps the unit square into Y according to the following scheme: 0 0 I: 0 0 0 0 0 y Yo and again by the preservation of connectedness.4 Z E Z. Yo). suppose that G and r are two paths from Zo to z.venng spaces y Fig. Given Z E Z. We now have a candidate for at present we have a function Z ~ X that satisfies p ol=f and depends on the choice of a path in Z for each z E Z.
Covering maps and covering spaces 257 (Y. 6. Let us set H = P*(H1 (Y. We leave it to you as an exercise to verify that the equation p h = P'. If H ~ H' then there is a covering map h: Y ~ Y' sending y to y' and such that p h = p'. This establishes that the definition of j is independent of the choice of path. based upon the interrelationships of their fundamental groups. y. Its end point is j(z'). • 0 0 Of course. Now for has its image in peN n V) and its lift starts at the end point of the lift of f 0". It follows that p hog = p so that both hog and the identity map id lift p to a map Y ~ Y. But pi (N n V) is a homeomorphism. together with the fact that both p and p' are covering maps. if H = H' then Proposition 12. 0 position. suffice to ensure that h is also a covering map . so that the end point of this lift lies in N n V. Thenfl(p(N n V)) is a neighbourhood of z in Z and so contains a path connected neighbourhood W of z in Z. Lift f (0". It remains to show that j is continuous. Apply the maplifting theorem and lift p over p' to obtain h: Y ~ Y' with p h = p'. Similarly g h = id and so h: Y ~ Y' is a homeomorphism. Why will the fact thatj(W) ~ V complete the proof? 0 Let z' E Wand now let r: [0.)). We shall show that j( W) ~ N and this will complete the proof.1 1. 0 Proof.7 works in both directions to give covering maps g: Y' ~ Yand h: Y ~ Y' with p' g = p and p h = p'. Let z E Z and suppose that N is a neighbourhood of j(z) in Y. We saw in Theorem 12. Consider two covering maps p: Y ~ X and p': Y' ~ X and choose basepoints y E Y and y' E Y' with p(y) = Xo = p(y'). which is j(z). 0 0 0 0 12.5 that a covering map determines a complete conjugacy class of subgroups of H1 (X. 1] ~ W be a path with reO) = z and r(l) = Z'. We call two covering maps p: Y ~ X and p': Y' ~ X equivalent if there exists a homeomorphism h: Y ~ Y' such that p h = p'. xo). Yo) at Yo and so a(t) = 13(1). r) to a path in Y beginning at Yo. • It is now possible to produce a hierarchical structure for the coverings of a given space. y)) and H' = p: (HI ( Y'. Let U be a neighbourhood off(z) in X and Van open set in Y containingj(z) such that p maps V homeomorphically to U. Show that 0 . and so by the uniqueness lemma we must have that hog = id.
then for all y E Y we have hey) =I: y. and is plainly surjective. we have that P*(Jrt(Y.(Y. Show that if h is a deck transformation and is not the identity map. the function 8: Y IQ ~ X given by [y] ~ p(y) is well defined. If for some y E pI(XO). We leave it to you as an exercise to complete the verification that eis a homeomorphism. y' E p'(x).Y'» since they are both conjugate.(1). H = P* (Jrt ( Y. or to its subgroups? Note that h is a lift of h over itself. Yo») is a normal subgroup of Jrt (X.2. Given a closed path a at Xo.7 there is a deck transformation h with hey) = y' and it follows that 8 is a bijection. For each x E X and for y. so that by the uniqueness lemma. xo) then X is homeomorphic to the orbit space Y/ Q and Q is isomorphic to the quotient group Jr. 2. (X. We may translate the conclusion of exercises 12. Then the map [a] H ka is a Q 0 .Then p a(l) = Xo so that there exists a unique deck transformation ka E Q such that ka(yo) = &. Deck transformations and universal covers Let p: Y ~ X be a covering map. let [y] denote its equivalence class in Y/ Q. distinct deck transformations must differ at every point of Y.2 1.1 related to the fundamental group of X. Proof. Given y E Y. Show that the set of all deck transformations for a fixed covering map p: Y ~ X forms a group under the operation of composition of ma ps.2.2. A deck transformation is a homeomorphism h: Y ~ Y such that p h = p. Let p: Y ~ X be a covering map and let Q be the group of deck transformations. y» = P*(Jr.1 and 12. By Proposition 12. This idea allows us to understand the deck transformations of certain covering maps p: Y ~ X in terms of the fundamental group of X. Yo» which is normal. to P*(Jr.(Y . Theorem.Covering spaces two covering maps are equivalent if and only if they determine the same conjugacy class of subgroups. A deck transformation is therefore a homeomorphism of the covering space that preserves the fibres of p: that is the subsets pI(X) for each x E X. Q cises 12. lift a to a closed path a beginning at Yo. by Theorem 12. xo)/ H. Since p hey) = p(y) for all hE Q.5. Is the group in Exercise 12.2 so as to say that the deck transformations of a covering map p: Y ~ X form a group that acts freely on Y.
Thus if X is a space with a universal covering space. and for a careful account of covering spaces from a combinatorial viewpoint we refer to Cohen (1989). any such space is called the universal covering space of X. We shall now concentrate on a particular aspect of this theory: that of covering spaces of graphs. there are still subtleties to be confronted. Again we leave the verification of the details of this assertion to you. so that graphs provide a topological model of free groups. xo) is called regular. For those spaces that do possess universal covering spaces we can establish an exact correspondence between subgroups of the fundamental group HI(X. then there exists a covering space corresponding to any subgroup of HI (X. We could repeat the topological theory in a combinatorial setting.Yo» is a normal subgroup of HI (X. x o) then XI H is a covering space of X with H as fundamental group. A covering map p: Y ~ X for which HI (Y. Meanwhile our topological theory has been developed in a very general setting.Covering graphs 259 surjective homeomorphism HI (Y. xo). This end could be . • A covering map p: Y ~ X for which P*(HI(Y. xo) ~ Q with kernel H. xo) is the group of deck transformations of X. We have seen in Chapter 10 that the fundamental group of a graph is always a free group. see Armstrong (1982) for a further discussion. In our previous discussions. Theorem.8. We shall make use of the topological theory. and keep everything as close to our previous discussion of graphs as possible. Let X be a space with a universal covering space X. While the ideas may be more intuitive here. Not every space possesses a universal covering space. and if H ~ HI (X. Yo) = 1 is unique up to homeomorphism (by Proposition 12. For these reasons. Then 1l'1 (X. and we can use the covering space theory of graphs to establish facts about free groups and their subgroups. Now comes the crunch. and this choice requires us to set up a rigorous topological theory of graphs. x o) and covering spaces of X. This correspondence is a refinement of Theorem 12. A choice has to be made about the type of approach that we follow.7) and is a regular covering space of any covering space of X. their structure determined by the incidence between edges and vertices. This is the aim of the remainder of the chapter. • Covering graphs We have seen how successfully the theory of covering spaces connects the topological nature of a space with the subgroup structure of its fundamental group. we have treated graphs essentially as combinatorial objects.
t mapping (0. and each e. G2. The new definition will define a graph as a onedimensional CW complex.t mapping S\{ I} to e. G3: the closure e. Regarding the origin as rO. Consider the following subset r = r of 1R2: n= G {(x. and readers who prefer to keep their minds on the combinatorial side of things should try to rephrase the definitions and results according to their preference.1) to e. y) I(x 1 (lIn)i + y2 = (lln)2}. 12.t of e.t is an edge.t: Sl ~ e. and G3 hold.en ng spaces achie:ved by modelling combinatorial graphs as subspaces of some IR n . Its purpose is to set up the topology on r when there are infinitely many ek It is of course convenient to retain the obvious terminology for graphs under the new definition. rO is the discrete topology. The . G4 looks at first sight more mysterious. Happily. CW complexes are a class of topological spaces so constructed that their homotopy theory avoids some pathologies that arise in the more general theory. Let us proceed to the definition of a graph. taking n as large as we need. The transition from combinatorial graph to CW complex and back again will be fairly evident. and redefine a graph.t meets rO in either one or two points: in the first case there is a homeomorphism/. We see that r consists of an infinite number of circles.t. and rO is G2: r\ro is a disjoint union of open subsets e..t is closed for all A. but we face problems of ensuring that this approach is consistent and independent of the embeddings chosen. The Hawaiian earring. 1] ~ e.3 1.t. G4: a subset A ~ r is closed if and only if A Il e. most spaces encountered in everyday life can be given the structure of a CW complex. and indeed is redundant if there are only finitely many e. Instead we opt for a more abstract setting. show that Gl. but that G4 fails. A graph is a Hausdorff space r together with a subspace rO such that the following conditions hold: Gl: The subspace topology on closed. Properties GI. each tangent to the yaxis at the origin. G2.t.t: [0. 1) ~ IR. and G3 have straightforward combinatorial reinterpretations. whilst in the second case there is a homeomorphism f.t each of which is homeomorphic to the open interval (O. and each circle as an edge. so we call points of ro vertices.
If there are n generators and the presentation is of a finite group G then the . Proof. The covering space corresponding to G will have i vertices and ir edges. Let r be a graph and p: Y ~ r a covering map. there will be ir .1) + l. and this number of edges equals the number of free generators of G. We may construct a graph r with one vertex and with r edges whose fundamental group is F. we construct a graph r. it must be a tree. and then generalized by O. The first result in the covering space theory of graphs is that we do not need to consider any other type of space. See Massey (1967) for a proof of the following proposition. a graph.Covering graphs 261 Hawaiian earring is also infamous for not possessing a universal covering space. Then G is a free group of rank i(r . The first theorem that we seek on free groups now follows easily. The NielsenSchreier theorem. the universal cover and a covering space Y = rl G of r with HI Y ~ G.i + I edges not in a maximal tree. and since a maximal tree must contain i . Then G must be a free group because it is isomorphic to the fundamental group of a graph. Let G be a subgroup of finite index i in the free group F of finite rank r. Given a subgroup G of the free group F. Schreier. Since this will be a graph with trivial fundamental group. • For subgroups of finite index in finitely generated free groups. The Schreier index formula. Proposition. Then Y is • The second result of importance is that any graph possesses a universal covering space. Nielsen in 1927 for finitely generated subgroups of free groups. This result is another due to Schreier. free. It was first proved by J. • Finally we show how to obtain free generators for relation subgroups in group presentations. Recall that the relation subgroup of a group presentation (X : R) is the normal subgroup of the free group with basis X generated by the elements R. we can be more precise and compute the number of free generators required for the subgroup. Proof. Every subgroup of a free group is r such that F ~ HI r.I edges.
2 ba.ermg spaces Schreier index formula tells us that the relation subgroup is a free group of rank I GI (n . Its Cayley quiver is shown in Fig. Can we produce free generators? The Cayley quiver gives an easy method. then along the edge. and see what are their images in the fundamental group of the nIeafed rose. we write down the label of the path that takes us from 1 to the source of the edge in the maximal tree. In our current example there are seven such edges.5. A detailed example will illustrate the method involved.2 b' a3 ba3b' b2 (3) (4) (5) (6) (7) . for it is a covering graph of the rose with n leaves. Consider the presentation (a. if. and its fundamental group is isomorphic to the relation subgroup. and their labels are (1) (2) baba a'ba'b' ba 2ba 2 a. 12. with 1 marked as basepoint and a maximal tree indicated.1) + 1. All we have to do is find generators for the fundamental group of the Cayley quiver. (ab)2) of the symmetric group of order 6. 12. and finally back to I from the target vertex. a Fig. b : a 3 .5 For each edge of the quiver not in the maximal tree.
(x. z]. Each can be written as a product of conjugates of the relators. (2). Draw Cayley quivers for examples (a) and (b) in Exercise 12. which again has order 60. b} containing the relators.3. the quaternion group of order 8. [y. y. 3. since the relation subgroup is the smallest normal subgroup of the free group with basis {a. y]. and (ab)2. y4. (c) (d) (e) 4. . b 2.3. y. yl xyX) == Q8. 12. (3). and (7) are conjugates of a single relator as they stand. z : xyzl. [z. (xy)5) == A 5. the alternating group of degree 5. x]). Use the Schreier index formula to compute the rank of the relation subgroup for the following presentations: (x. (x. .3 continued 2. and find sets of free generators for the relation subgroups in each case. presenting the cyclic group of order 60. [x.Covering graphs 263 These seven elements generate the relation subgroup. y : x1yxy. (6). y : x 2. Z5.1 zxyl) (x: x 60 ). (x. yzx. z : x 3. (a) (b) Write the elements (1). == Q8. Note that (5). y3. and (4) in the example just given as products of conjugates of the relators a 3 . which also presents the cy clic group of order 60.
.. New York. A combinatorial representation for polyhedral surfaces. New York. and Millett. (1988).A. R. R. K. G. R. J. E. C. W. 395407. J. A. (1960). J. R. In Computational problems in abstract algebra (ed. Bleiler. Studies 115. and Ocneanu. H. Topological graph theory. Soc. W. P. 12. Math. A note on unknotting number. Burde. Student Text 15. Johnson. Knots. R. Cambridge Philos. Soc. Soc. Brown. and Tucker. Fund. 26. SpringerVerlag. (1987a). A. Math. Ann. Trans. Math. 96. Pergamon.. AddisonWesley. 27183. Bull. Berlin. D. (1967). (1987). Millett. Math. Yetter. Soc. New York.. H.. D. 30. (1983). F. Math. Princeton University Press. and Zieschang. Student Text 14. Introduction to knot theory. Oxford. Soc. Singapore. D. 46971. Ellis Horwood. 23946. . (1987). (1984). New York. (1972). (1989). J. 15. W. L. 26. (1969). L. An enumeration of knots and links. Topology. L. Combinatorial group theory. Berlin. 12. (1985). Am. 7. Soc. State models and the Jones polynomial. and Fox. Lickorish. Kauffman.. 275306. Kauffman.. De Gruyter. Heidelberg. M. Not. H. K. London Math. London Math. Freyd. London Math. L. Topology. (l987b). Gross. . B. Berlin. On knots. Topological invariants of knots and links. Wiley Interscience. Knots and physics. H. D. (1928). (1930). Sur Ie probleme des courbes gauches en topologie. H. Bull. J. 10311. SpringerVerlag. Leech). Proc. C. K. Kauffman. Crowell. Soc. Massachusetts. Chichester. Hoste. Topology. H. Math. Proc. Armstrong.. A new polynomial invariant of knots and links. World Scientific Publishing. Basic topology. Differential manifolds. Cambridge University Press. A polynomial invariant of oriented links.385401. B. Am. (1990). R. 646. Lickorish. (1985). (1985). A polYI1omial invariant for knots via Von Neumann algebras. (3). Presentations of groups. Conway. Edmonds.. S.17. Jones. S. Lang. Am. Am. H. Math. 10741. L. R. Soc. Kuratowski. (1991). W. Reading. Brown. Cohen.Bibliography Alexander. Groupoids and Van Kampen's Theorem. V. T. (1967). Heidelberg.. Cambridge University Press.
2. (1964). Introduction to metric and topological spaces. R. S. Wilson. 28. NorthHolland Mathematical Studies 8. 13370. Phys. Monatsschr. Graphs. Topology. 19. J. An introduction to graph theory. Sutherland. (1984). C. A. 1118. Schubert. L. Brown. K.. z. (1961). Massey. Heidelberg. SpringerVerlag. . (1985). M. Oxford University Press. 65. 26. 'e\\ York. Duke Math. Rotman. 54551. W. B. Heidelberg. (1981). Knotentheorie. A. (1932). H. White. B. (1984). 297309. and Chandler. Introduction to complex analysis. Noninvertible knots exist. Neuwirth. The history of combinatorial r:roup theor. W. Topology. 26. Knoten mit zwei Briicken. K. Uber die topologischen Invarianten mehrdimensionaler Mannigfaltigkeiten. Ergebnisse der Mathematik 1. (1990). Oxford University Press. J.. Berlin. Math. A spanning tree expansion of the Jones polynomial. H. Harlow. (1982). (1908). Tietze. The theory of groups. A. Algebraic topology. Topology. Math. Elsevier. Thistlethwaite. (1987). 27580. (1956). Amsterdam. Longman. T. H. Priestley. H. SpringerVerlag. A note on torus knots and links determined by their groups.Bibliography 265 Magnus. W.v.. Murasugi. New York. Trotter. 18794. Reidemeister. J. groups and surfaces. F. (1967). SpringerVerlag. Jones polynomials and classical conjectures in knot theory. W. (1987). Berlin. Berlin.
see groups Alexander matrix 1847 arc 3 augmentation 180 ball closed 44 open 43 boundary curve 82 boundary edge 82 bridge 13 bridge number 13.c. 1003 of a sum of knots 103 of a surface 8790. 1802 genus of a graph 176 of a knot 935.224 simple 167 subgraph of 165 topological 260 tree 1678 walk in 166 see also quivers groupoid.Index Abelianization.234 centre of 246 .d. 1035 Brouwer Fixed Point Theorem 209 edge.235 diagram connected 31 of knot or link 23 oriented 3 reduced 31 splitting marker in 23 state of 246 Van Kampen 133 writhe of 3 disconnection 48 domain. 215. 165 Edmonds algorithm 174 embedding 46 of graph in surface 1737 disc 176 minimal 176 Euler characteristic 8490 fibres 258 function continuous 45 mapping cylinder of 51 pointed 208 commutator 116 commutator subgroup 116 components of link diagram 3 of topological space 49 conjugacy 116 consequence 130 continuity 458 crosscap 748 crossing 3 change 11 number 11 cuff 76 decomposition. 143 derivative 1827. fundamental 199. of abelian group 14J. 92 graphs 16377 bipartite 169 circuit in 167 complete 169 connected 166 coverings of 2613 directed 163 embedding of 1737 fundamental group of 228 isomorphism of 164 planar 1713 Euler's formula for 171 rose 169. 223 group ring 17882 groups Abelianization of 117. 13940. of graph or quiver 163. g.
~".': 1902 Tietze's Theorem f. 94 bridge number of 13 chiral 17 cinquefoil 151. 254 calculations of 203. 254 of maps 199 pointed 199. 1612 crossing number of II diagram 2. 27 generalized 159 granny 97 invertible 17 mirror images of II.221 surface 234 see also presentation polynomial in.3~ span of 31 presentation II~. 2469 pushout of 1469. 207 inverse 209 lifting of 205. 208 neighbourhood ~ orientation of link diagram of surface 66 overpass 3 path closed 199 duration of 19in a graph 166homotopy 199 lifting of 2~.!. 27 knot type 53 link 52 Borromean rings ~ component of 3 diagram of 3 Hopf link 2..'~ Wirtinger 1514.28. 28 alternating 18. 1589. 202..Index 267 dinclic 118 dihedral 1l218 direct sum of 142 finitely generated 140 fundamental 156. of Abelian group I~l ~ of Abelianization 1"''1 Alexander matn" c.aflants c':_ polynomial prime 1067 reef 97 torus 155. moves.:: polynomial Alexander 3841.37 . 1568 "I handle 758 homeomorphism 46 homotopy equivalence 209 fixedendpoint 199. consequence of rddth'::' Dehn 1546 mapping of 135~ of quotient group I_~'" relation in 129 Tietze transformatlc. 128 knot 15062. 2369. . 1078.n . 2356 invariance of 190 generated by subset 188 isotopy ambient 52 class 6 of diagrams 6 invariant of 7 regular 6. 109 I" :Jones 2830 Kauffman 26. 9~. 52 equivalence of 53 figureeight 2. Reidemelster :. 22433 free 1228 rank of 128 see also word free product of 147. 1920 isthmus 31 knots achiral 17. product of 191< reverse 198 in a space 199. ~~_"' " trefoil 2. 16. :L.. 1517. 1:'6" ' ideals elementary 18892. . I"' I. 14. 206. :"" bracket 206 HomOy 348. 155 colouring of 810. 23943.216 with amalgamated subgroup 149 generalized quatemion 118 generators of 113.27.f : . 36 linking number ~.217.
deformation 209 space.Index quivers 16571 Cayley 11822. of Abelian groups 1424 of knots 96110. topological 438 connected 489 compact 214 convex 203 covering 2502 deck transformation of 2589 regular 259 universal 259 with given fundamental group 2369 Hausdorff 47 Hawaiian earring 260 pointed 208 simply connected 203 subspace 44 closed 44 diagonal 47 open 435 topology induced on 44 sum direct.2623 isomorphism of 165 map of 165 relational 167 relation 129 retract. 243 245. 2379. 80 orientable 66 pretzel 55 projective plane 65 Seifert 91 sphere 54 symbol 7190 normal form of 7890 torus 54 transformation. see presentat tree 1678 topology 435 discrete 434 finite complement 43 Hausdorff 47 identification 4951 indiscrete 49 product 47 usual 43 underpass 3 unknotting number 12 Van Kampen theorem 2212 applications of 22633.1701. 1238 alphabet for 123 contraction of 124 expansion of 124 length of 115 letter of 123 syllable in 123 writhe 3. Tietze.26 . 1601 of surfaces 87 surfaces boundary of 82 classification of 78 combinatorial 5764 double of 89 Euler characteristic of 8490 fundamental groups of 22933 Klein bottle 62 Mobius band 56 nonorient able 66.248 winding number 2034 word 115.