Topological

Graph Theory

JONATHAN L. GROSS
Columbia University
New York, New York

THOMAS W. TUCKER
Colgate University
H ami/ton, New York

A Wiley-Interscience Publication
JOHN WILEY & SONS
New York • Chichester • Brisbane • Toronto • Singapore
Copyright© 1987 by Jonathan L. Gross and Thomas W. Tucker
All rights reserved. Published simultaneously in Canada.
Reproduction or translation of any part of this work
beyond that permitted by Section 107 or 108 of the
1976 United States Copyright Act without the permission
of the copyright owner is unlawful. Requests for
permission or further information should be addressed to
the Permissions Department, John Wiley & Sons, Inc.

Library of Congress Cataloging in Publication Data:

Gross, Jonathan L.
Topological graph theory.
(Wiley interscience series in discrete mathematics and optimization)
"A Wiley-Interscience publication."
Bibliography: p.
1. Graph theory. 2. Topology. I. Tucker, Thomas.
II. Title. III. Series.
QA166.G76 1987 511'.5 87-6221
ISBN 0-471-04926-3
Printed in the United States of America
10 9 8 7 6 5 4 3 2
To Ralph H. Fox
To my father, Albert W Tucker
Preface

The primitive objective of topological graph theory is to draw a graph on a

surface so that no two edges cross, an intuitive geometric problem that can be
enriched by specifying symmetries or combinatorial side-conditions. To solve
the problem and its variants, techniques are adapted from a broad range of
mathematics, especially from algebraic topology and group theory, and more
recently from enumerative combinatorics and the analysis of algorithms.
Topological graph theory remains a young specialty, in which two decades
ago there were but a scant handful of active contributors. Subsequently,
however, there have been well over a thousand research papers partially or
exclusively relevant to graph placement problems, written by hundreds of
different contributors, many with simultaneous mathematical interests else-
where.
Unmistakably, the turning point was the Ringel-Youngs solution, com-
pleted in 1968, of the classical Heawood map-coloring problem. Numerous
follow-up investigations analyzed and generalized the methods devised for the
Ringel-Youngs theorem, and the resulting techniques were applied to many
other graph imbedding problems. Substantial credit is also due to Arthur
White, whose lucid study, Graphs, Groups and Surfaces (1973), made topologi-
cal graph theory accessible to newcomers, and the revised edition (1984)
continues to serve that purpose well.
The present monograph is intended as a historically sensitive, comprehen-
sive introduction to the foundations and central concerns of topological graph
theory, with bridges to frontier topics and to other areas of mathematics,
wherever they fit in naturally. It is sufficiently self-contained that it can serve
as a first-year graduate text or a self-guided approach for ambitious persons
having a background in undergraduate discrete mathematics, in particular,
some acquaintance with graphs and with elementary facts about groups and
permutation groups. No point-set topology is necessary, beyond what one
might see in an honors calculus course.
Although the focus is concentrated on the exposition and application of
general topological methods to graph imbeddings, a small amount of effort is
invested in glimpses at the frontiers and in interesting digressions. Nonethe-
less, for the most part we have steered away both from heavy involvement in
non-topological methods-as in the proof of the four-color theorem or in
solutions to most crossing-number problems-and from complicated topologi-
cal methods whose cost of mastery has not yet been demonstrated to com-
viii Preface

pensate adequately in imbeddings (or in proofs of non-existence of imbed-

dings) of interesting graphs. Moreover, an extensive survey of the frontiers
would have been incompatible with our desire to stay close to the core, as
much as we regret omitting explicit mention of numerous important results,
including some of our own.
The first four chapters are basic, beginning with a review of graph theory
that stresses whatever is of topological importance. Discussion of imbeddings
into surfaces is blended with a complete proof of the classification of closed
surfaces. Voltage graphs, which are a combinatorial form of covering spaces,
are developed in full generality, acknowledging the role of covering-space
constructions as the single most important tools used in the derivation of
genus formulas. No prior experience in algebraic topology is needed for our
treatment of coverings and group actions on surfaces, and indeed, the special-
ized approach here might well serve as a preliminary to a general topological
treatment.
The foundations thus established are used in a straightforward explanation
of the very lengthy proof of the Ringel-Youngs theorem, which includes
sufficient detail to enable someone to comprehend the original proof in its
entirety. They are also used in an examination of the main results on the genus
of a group, in which one studies imbeddings of "Cayley graphs", which are
pictorial representations of a group, optimized over all possibilities for the
group.
In early 1987, the frontiers of topological graph theory are advancing in
numerous different directions, each pursued by a host of researchers, including
particularly active schools in France, Italy, the United Kingdom, the United
States, and Yugoslavia. An alphabetized partial listing of frontier topics would
include algorithms for imbedding problems, covering-space constructions (in-
cluding wrapped coverings and flows), enumerative analysis of imbedding
distributions (sometimes using representation theory of the symmetric group),
forbidden subgraph characterizations and their relationship to graph minors,
genus of groups, map-theoretic connections, neoclassical imbedding questions,
representation of higher-dimensional manifolds by graphs, and VLSI layouts.
Moreover, there are some significant one-person and two-person approaches
that fall into none of the foregoing categories. A single volume could not hope
to treat all of the major developments in detail, even though we are still able to
bring the reader within striking distance of most of them.
We are indebted to many colleagues and students who have read parts of
our many preliminary drafts, particularly to Arthur White and Ward Klein,
whose critical suggestions have greatly helped. Special thanks are due to Mary
Forry for typing the final draft.
JONATHAN L. GROSS
THOMAS W. TUCKER
New York, New York
Hamilton, New York
May 1987
Contents

1. INTRODUCfiON 1

1.1. Representation of graphs 1

1.1.1. Drawings 2
1.1.2. Incidence matrix 3
1.1.3. Euler's theorem on valence sum 3
1.1.4. Adjacency matrix 4
1.1.5. Directions 4
1.1.6. Graphs, maps, isomorphisms 5
1.1.7. Automorphisms 6
1.1.8. Exercises 7
1.2. Some important classes of graphs 7
1.2.1. Walks, paths, and cycles; connectedness 8
1.2.2. Trees 8
1.2.3. Complete graphs 10
1.2.4. Cayley graphs 10
1.2.5. Bipartite graphs 14
1.2.6. Bouquets of circles 15
1.2.7. Exercises 15
1.3. New graphs from old 16
1.3.1. Subgraphs 16
1.3.2. Topological representations, subdivisions,
graph homeomorphisms 17
1.3.3. Cartesian products 19
1.3.4. Edge-complements 19
1.3.5. Suspensions 20
1.3.6. Amalgamations 20
1.3.7. Regular quotients 21
1.3.8. Regular coverings 22
1.3.9. Exercises 23
ix
X Contents

1.4. Surfaces and imbeddings 24

1.4.1. Orientable surfaces 24
1.4.2. Nonorientable surfaces 25
1.4.3. 1mbeddings 26
1.4.4. Euler's equation for the sphere 27
1.4.5. Kuratowski's graphs 28
1.4.6. Genus of surfaces and graphs 29
1.4.7. The torus 30
1.4.8. Duality 31
1.4.9. Exercises 32
1.5. More graph-theoretic background 33
1.5.1. Traversability 34
1.5.2. Factors 35
1.5.3. Distance, neighborhoods 36
1.5.4. Graphs colorings and map colorings 37
1.5.5. Edge operations 38
1.5.6. Algorithms 40
1.5.7. Connectivity 40
1.5.8. Exercises 41
1.6. Planarity 42
1.6.1. A nearly complete sketch of the proof 42
1.6.2. Connectivity and region boundaries 45
1.6.3. Edge contraction and connectivity 46
1.6.4. Planarity theorems for 3-connected graphs 48
1.6.5. Graphs that are not 3-connected 49
1.6.6. Algorithms 51
1.6.7. Kuratowski graphs for higher genus 53
1.6.8. Other planarity criteria 53
1.6.9. Exercises 54

2. VOLTAGE GRAPHS AND COVERING SPACES 56

2.1. Ordinary voltages 57

2.1.1. Drawings of voltage graphs 57
2.1.2. Fibers and the natural projection 60
2.1.3. The net voltage on a walk 61
2.1.4. Unique walk lifting 62
Contents xi

2.1.5. Preimages of cycles 63

2.1.6. Exercises 64
2.2. Which graphs are derivable with ordinary voltages? 66
2.2.1. The natural action of the voltage group 66
2.2.2. Fixed-point free automorphisms 67
2.2.3. Cayley graphs revisited 69
2.2.4. Automorphism groups of graphs 70
2.2.5. Exercises 71
2.3. Irregular covering graphs 72
2.3.1. Schreier graphs 73
2.3.2. Relative voltages 74
2.3.3. Combinatorial coverings 75
2.3.4. Most regular graphs are Schreier graphs 78
2.3.5. Exercises 79
2.4. Permutation voltage graphs 81
2.4.1. Constructing covering spaces with permutations 81
2.4.2. Preimages of walks and cycles 82
2.4.3. Which graphs are derivable
by permutation voltages? 84
2.4.4. Identifying relative voltages
with permutation voltages 85
2.4.5. Exercises 86
2.5. Subgroups of the voltage group 86
2.5.1. The fundamental semigroup of closed walks 87
2.5.2. Counting components of ordinary derived graphs 89
2.5.3. The fundamental group of a graph 92
2.5.4. Contracting derived graphs onto Cayley graphs 92
2.5.5. Exercises 93

3. SURFACES AND GRAPH IMBEDDINGS 95

3.1. Surfaces and simplicial complexes 95

3.1.1. Geometric simplicial complexes 96
3.1.2. Abstract simplicial complexes 97
3.1.3. Triangulations 98
3.1.4. Cellular imbeddings 100
3.1.5. Representing surfaces by polygons 102
3.1.6. Pseudosurfaces and block designs 104
xii Contents

3.1.7. Orientations 106

3.1.8. Stars, links, and local properties 106
3.1.9. Exercises 107
3.2. Band decompositions and graph imbeddings 109
3.2.1. Band decomposition for surfaces 109
3.2.2. Orientability 110
3.2.3. Rotation systems 112
3.2.4. Pure rotation systems and orientable surfaces 113
3.2.5. Drawings of rotation systems 113
3.2.6. Tracing faces 114
3.2.7. Duality 116
3.2.8. Which 2-complexes are planar? 117
3.2.9. Exercises 118
3.3. The classification of surfaces 119
3.3.1. Euler characteristic relative to an imbedded graph 121
3.3.2. lnvariance of Euler characteristic 121
3.3.3. Edge-deletion surgery and edge sliding 124
3.3.4. Completeness of the set of orientable models 126
3.3.5. Completeness of the set of nonorientable models 128
3.3.6. Exercises 130
3.4. The imbedding distribution of a graph 132
3.4.1. The absence of gaps in the genus range 133
3.4.2. The absence of gaps in the crosscap range 134
3.4.3. A genus-related upper bound on the
crosscap number 136
3.4.4. The genus and crosscap number of the
complete graph K 7 137
3.4.5. Some graphs of crosscap number 1 but
arbitrarily large genus 140
3.4.6. Maximum genus 142
3.4.7. Distribution of genus and face sizes 146
3.4.8. Exercises 147
3.5. Algorithms and formulas for minimum imbeddings 149
3.5.1. Rotation-system algorithms 149
3.5.2. Genus of an amalgamation 150
3.5.3. Crosscap number of an amalgamation 154
3.5.4. The White-Pisanski imbedding of a
cartesian product 155
3.5.5. Genus and crosscap number of cartesian products 158
3.5.6. Exercises 160
Contents xiii

4. IMBEDDED VOLTAGE GRAPHS AND CURRENT GRAPHS 162

4.1. The derived imbedding 162

4.1.1. Lifting rotation systems 162
4.1.2. Lifting faces 163
4.1.3. The Kirchhoff Voltage Law 166
4.1.4. Imbedded permutation voltage graphs 166
4.1.5. Orientability 167
4.1.6. An orientability test for derived surfaces 170
4.1.7. Exercises 172
4.2. Branched coverings of surfaces 174
4.2.1. Riemann surfaces 174
4.2.2. Extension of the natural covering projection 176
4.2.3. Which branch coverings come from
voltage graphs? 177
4.2.4. The Riemann-Hurwitz equation 179
4.2.5. Alexander's theorem 179
4.2.6. Exercises 181
4.3. Regular branched coverings and group actions 182
4.3.1. Groups acting on surfaces 182
4.3.2. Graph automorphisms and rotation systems 184
4.3.3. Regular branched coverings and ordinary
imbedded voltage graphs 186
4.3.4. Which regular branched coverings come from
voltage graphs? 187
4.3.5. Applications to group actions on the surface S2 189
4.3.6. Exercises 190
4.4. Current graphs 191
4.4.1. Ringel's generating rows for Heffter's schemes 191
4.4.2. Gustin's combinatorial current graphs 193
4.4.3. Orientable topological current graphs 194
4.4.4. Faces of the derived graph 196
4.4.5. Nonorientable current graphs 198
4.4.6. Exercises 201
4.5. Voltage-current duality 202
4.5.1. Dual directions 202
4.5.2. The voltage graph dual to a current graph 204
4.5.3. The dual derived graph 206
4.5.4. The genus of the complete bipartite graph Km, n 210
4.5.5. Exercises 212
xiv Contents

5. MAP COLORINGS 215

5.1. The Heawood upper bound 216

5.1.1. Average valence 216
5.1.2. Chromatically critical graphs 217
5.1.3. The five-color theorem 219
5.1.4. The complete-graph imbedding problem 220
5.1.5. Triangulations of surfaces by complete graphs 223
5.1.6. Exercises 224
5.2. Quotients of complete-graph imbeddings and
some variations 224
5.2.1. A base imbedding for orientable case 7 225
5.2.2. Using a coil to assign voltages 226
5.2.3. A current-graph perspective on case 7 229
5.2.4. Orientable case 4: doubling 1-factors 230
5.2.5. About orientable cases 3 and 0 233
5.2.6. Exercises 235
5.3. The regular nonorientable cases 236
5.3.1. Some additional tactics 236
5.3.2. Nonorientable current graphs 237
5.3.3. Nonorientable cases 3 and 7 238
5.3.4. Nonorientable case 0 239
5.3.5. Nonorientable case 4 240
5.3.6. About nonorientable cases 1, 6, 9, and 10 240
5.3.7. Exercises 240
5.4. Additional adjacencies for irregular cases 241
5.4.1. Orientable case 5 241
5.4.2. Orientable case 10 242
5.4.3. About the other orientable cases 245
5.4.4. Nonorientable case 5 246
5.4.5. About nonorientable cases 11, 8, and 2 247
5.4.6. Exercises 247

6. TilE GENUS OF A GROUP 249

6.1. The genus of abelian groups 249

6.1.1. Recovering a Cayley graph from any
of its quotients 250
Contents XV

6.1.2. A lower bound for the genus of most

abelian groups 254
6.1.3. Constructing quadrilateral imbeddings for most
abelian groups 255
6.1.4. Exercises 263
6.2. lbe symmetric genus 264
6.2.1. Rotation systems and symmetry 265
6.2.2. Reflections 268
6.2.3. Quotient group actions on quotient surfaces 270
6.2.4. Alternative Cayley graphs revisited 271
6.2.5. Group actions and imbeddings 273
6.2.6. Are genus and symmetric genus the same? 275
6.2.7. Euclidean space groups and the torus 276
6.2.8. Triangle groups 279
6.2.9. Exercises 282
6.3. Groups of small symmetric genus 283
6.3.1. The Riemann-Hurwitz equation revisited 284
6.3.2. Strong symmetric genus 0 285
6.3.3. Symmetric genus 1 291
6.3.4. The geometry and algebra of groups of
symmetric genus 1 295
6.3.5. Hurwitz's theorem 296
6.3.6. Exercises 298
6.4. Groups of small genus 300
6.4.1. An example 300
6.4.2. A face-size inequality 302
6.4.3. Statement of main theorem 304
6.4.4. Proof of Theorem 6.4.2: valence d = 4 306
6.4.5. Proof of Theorem 6.4.2: valence d = 3 308
6.4.6. Remarks about Theorem 6.4.2 312
6.4.7. Exercises 317

References 319

Bibliography 333

Table of Notations 341

Subject Index 347

1
Introduction

In beginning the study of topological graph theory, one needs to prepare both
for the development of topological operations and for discussion of the
relationship between these operations and the other concepts of graph theory.
Most of the important topics in graph theory have some connection with the
topological viewpoint. Accordingly, this first chapter reads almost like a survey
of graph theory, encompassing trees, automorphisms, traversability, factoriza-
tion, colorings, algorithms, and more.

1.1. REPRESENTATION OF GRAPHS

In topological graph theory, the intuitive viewpoint is that a graph is a

network of nodes and curved arcs from some nodes to others or to themselves.
Nonetheless, it is a formal convenience to adopt a combinatorial dt:finition,
rather than a topological definition. Therefore, we say that a "graph" G
consists of a set Vc of vertices and a set Ec of edges. Each edge e has an
endpoint set Vc( e) containing either one or two elements of the vertex set Vc.
In a formal setting, we call the set Ic = { Vc( e )le E Ec} the "incidence
structure". For simplicity, when G is the only graph under consideration, one
omits the subscripts and writes V and E instead of Vc and Ec.
If both the vertex set and the edge set are finite, then the graph G is said to
be "finite". With very few exceptions, the graphs discussed here are implicitly
considered to be finite.
A graph is called "simplicial" if it has no self-loops or multiple edges,
because, in the standard sense of topology, such a graph is a simplicial
1-complex. (In nontopological contexts, the word "simple" 1 is sometimes used
in place of "simplicial".)

1 To a topologist, the word "simple" has pre-emptive meaning. In particular, a "simple" loop is
the continuous nonsingular image of a circle, that is, a loop without self-intersections or other
singularities. Our general rule is to avoid ambiguity or such conflicts of usage.
2 Introduction

Figure 1.1. The graph of Example 1.1.1.

1.1.1. Drawings
Any finite graph can be geometrically represented by a drawing obtained in
the following manner. First, draw a dot for each vertex. Then, for every edge
e E E with two endpoints, draw a line between the dots representing the
vertices of V( e), and for every edge with only one vertex, draw a line from
the dot representing that vertex to that dot itself. Two examples now illustrate
the representation of a graph by a drawing.

Example 1.1.1. Let V = { v 1 , v2 , v3 , v4 }, and let E = {e1 , e 2 , e 3 , e 4 }. Also,

let V(e 1 ) = {vp v2 }, V(e 2 ) = {v 1 , v3 }, V(e 3 ) = {v 2 , v3 }, and V(e 4 ) =
{ v 3 , v4 }. A drawing of this graph is shown in Figure 1.1. For the sake of clarity,
the dots and lines are labeled by the names of the vertices and edges they
represent.

Example 1.1.2. Let V = { v 1 , v2 , v3 }, and let E = {e 1 , e 2 , e 3 , e 4 }. Let V( e 1 )

= {vd, V(e 2 ) = {v 1 , v2 }, V(e 3 ) = {v2 , v3 }, and V(e 4 ) = {v 2 , v3 }. Figure
1.2 shows two different drawings of this second graph.

Certain graphs cannot be drawn on a planar surface without line-crossings.

Moreover, for the sake of illustration or other convenience, a drawing of a
graph sometimes contains more line-crossings than necessary. Fortunately, it is
never necessary for the interior of any line to touch any dot. In a drawing of a
graph, vertices are represented only by dots, not by line-crossings.
Readers already familiar with graph-theoretic literature may note dif-
ferences in the present terminology and notation from that appearing elsewhere.
For instance, if a graph has an edge with only one endpoint, then Harary
(1969) would call it a "pseudograph". Harary's distinction has been widely
adopted, because for many purposes, it permits more efficient discussions.

Figure 1.2. Two drawings of the graph of Example 1.1.2, one without line-crossings, the other
with crossings.
Representation of Graphs 3

Certain nonsimplicial graphs, however, such as bouquets of circles (see Section

1.2), are so essential to topological graph theory that it is more efficient to
include them here as graphs. In general, the authors have tried to preserve
standard terminology and notations except where there is substantive reason
in topological graph theory to introduce something new. For the convenience
of the reader, a list of all notation used here appears at the end of this book.

1.1.2. Incidence Matrix

The "incidence matrix" for a graph G has #V rows and #E columns, where
#S denotes the cardinality of a set S. The entry in row i and column j is 1 if
v; E V(e) and #V(e) = 2, 2 if { v;} = V(e), and 0 if v, ft. V(e). For
example, the graphs of Examples 1.1.1 and 1.1.2 have the following incidence
matrices, respectively:

e2 e3

n
1 0
0 1
1 1
0 0

Communicating a graph to a computer by its incidence matrix rather than by

its incidence structure may involve a loss of space efficiency, because the zeros
occupy storage space not needed for the incidence structure. However, since
certain information is recovered more easily from a stored incidence matrix
than from a stored incidence structure, there are times when sacrificing the
space yields worthwhile gains in operational speed.

1.1.3. Euler's Theorem on Valence Sum

From a geometric viewpoint, the "valence" 2 of a vertex v is the minimum
number of line-crossings that occur while one traverses a small circle around
the dot representing v in a drawing of the graph. For instance, the valences of
v1 , v2 , and v3 in Example 1.1.2 are 3, 3, and 2, respectively. Roughly speaking,
it is sometimes said that the valence of v is the number of edges incident on v,
but to get the correct number, one must take care to count an edge twice if v is
its only endpoint. Alternatively, one may give a strictly combinatorial defini-
tion of the valence of v as the sum of the entries in its row in the incidence
matrix.

2 The synonym "degree" appears to be more frequently used by graph theorists than "valence".
However, since "valence" is a self-explanatory abstraction from the terminology of chemistry, and
since "degree" is used for polynomials and as the name for a topological invariant of a continuous
function, it seems appropriate to use "valence" here for the graph-theoretic concept.
4 Introduction

Theorem 1.1.1 (Euler). The sum of the valences of the vertices of a graph equals
twice the number of edges.

Proof Every edge contributes exactly 2 to the valence sum. Alternatively,

it may be observed that the row sums in the incidence matrix are the valences
and that every column sum is 2, corresponding to the contribution of each
edge. Of course, the total of the row sums equals the total of the column sums.
D

1.1.4. Adjacency Matrix

Two vertices of a graph are called "adjacent" if they are the endpoints of the
same edge. They are called "multiply adjacent" if there are two or more edges
having them both as endpoints. The "multiplicity of their adjacency" is
defined to be the number of edges having them both as endpoints, possibly
zero.
A vertex is called "self-adjacent" if there is an edge for which it is the only
endpoint. The "multiplicity of its self-adjacency" is defined to be the number
of edges for which it is the only endpoint. An edge with only one endpoint is
called a "loop". If an edge is not a loop, then it is called a "proper edge".
The "adjacency matrix" for a graph G is square, with #V rows and
columns. The entry in row i and column j is the multiplicity of the adjacency
between the vertices v; and v1 . It follows from this definition that an adjacency
matrix is symmetric about the main diagonal. A loop contributes only 1 to a
main diagonal entry, so a row or column sum is not necessarily the valence of
a vertex. For example, the graphs of Examples 1.1.1 and 1.1.2 have the
following adjacency matrices, respectively:

n
vi v2 v3 v4 vi v2 v3

u n
vi 1 1 vi 1
v2
v3
v4
0
0
0
1
1
1
v2
v3 [i 0
2

1.1.5. Directions
A "direction" for an edge e is an onto function {BEGIN, END} ~ V(e). The
images of BEGIN and END are called the "initial point" and the "terminal
point", respectively. One says that a directed edge goes "from" its initial point
"to" its terminal point. In the incidence structure for a list format of a graph,
the direction of a proper edge may be indicated by putting the initial point
first in its endpoint list.
In topological graph theory, one frequently considers each edge e (even a
loop) to have two directions, arbitrarily distinguished as the "plus direction"
Representation of Graphs 5

Figure 1.3. Using arrowheads to indicate directions on

edges of the graph of Example 1.1.2.

e + and the "minus direction" e-. In a drawing of a graph, an arrowhead is

used to show the plus direction, with the initial point behind the arrowhead
and the terminal point in front of it. Figure 1.3 shows how the arrowheads are
placed on the left-hand drawing of Figure 1.2 if the direction e 2 + is from v2 to
v1 and the directions e3 + and e4 + are both from v3 to v2 •
An edge e together with either its plus direction or its minus direction is
called a "directed edge," and is denoted e+ or e-, respectively, thereby
slightly abusing the notation.

1.1.6. Graphs, Maps, Isomorphisms

Let G and G' be graphs. A "graph map" f: G ~ G' consists of a vertex
function VG ~ VG' and an edge function EG ~ EG' such that incidence is
preserved, in the sense that for every edge e E EG' the vertex function maps
the endpoints of e [i.e., the set Va(e)] onto the endpoints of the image e. It is
customary to denote both the vertex function and the edge function by the
same symbol as the graph map, namely, f. Under a graph map, a loop may be
the image of a proper edge, but a proper edge may never be the image of a
loop.
A graph map G ~ G' is called an "isomorphism" if both its vertex function
and its edge function are one-to-one and onto. Two graphs are called "isomor-
phic" if there exists an isomorphism from one to the other.

Example 1.1.3. Figure 1.4 shows two graphs, G and G'. Suppose the function f
has the values

f(vl)=vl' f(v2) = f(v3) = v2'

f(el) =J(e2) = e1' /( e3) = e2'

G G'
Figure 1.4. The domain G and the range G' of a graph map f: G .... G'.
6 Introduction

u§

H H'
Figure 1.5. Two isomorphic graphs.

The endpoints of the edge e1 are the vertices v 1 and v2 , which are mapped onto
v 1' and v2', respectively, the endpoints of the edge e 1', which is the image of e1 •
The endpoints v 1 and v3 of the edge e2 are mapped onto the endpoints v 1' and v2'
of the edge e 1', which is the image of e 2 • Both the endpoints v2 and v3 of the edge
e 3 are mapped onto the single endpoint v2' of its image e{ Thus the vertex map f
and the edge map f together form a graph map f: G ~ G'. However, f is not an
isomorphism.

Example 1.1.4. Figure 1.5 shows two isomorphic graphs, H and H'. One
isomorphism g: H ~ H' is given by the rule

g(u;)=u;' fori=l, ... ,4

g(d;) = d;' fori= 1, ... ,5

1.1.7. Automorphisms
An isomorphism from a graph to itself is called an "automorphism". Under
the operation of composition, the family of all automorphisms of a graph
forms a group, called "the automorphism group of the graph" and denoted
Aut( G).

Example 1.1.5. Consider a graph map h: H ~ H on the left-hand graph H of

Figure 1.5, in which h has the values

h(u 1 ) = u 3 h(u 2 ) = u 2 h(u 3 ) = u 1 h(u 4 ) = u 4

h(dl) = d2 h(d2) = dl h(d3) = d4 h(d4) = d3
Alternatively, using disjoint permutation-cycle notation, the graph map h may be
written as follows:
vertex map h = (u 1 u3 )(u 2 )(u 4 )
edge map h = (d 1 d 2 )(d 3 d 4 )(d 5 )
Then h is an automorphism.
Some Important Oasses of Graphs 7

1.1.8. Exercises
1. Draw the graph with vertices v1 , v2 , and v 3 , with edges e 1 , e 2 , e 3 , e4 ,
and e 5 , and with endpoint sets V(e 1 ) = {v 1 , v2 }, V(e 2 ) = {v 2 , v3 },
V(e 3 ) = {v 1 , v3 }, V(e 4 ) = {v 1 , v3 }, and V(e 5 ) = {v 3 }.
2. Write the incidence stru~tures for the graphs of Figure 1.4.
3. What is the incidence matrix for the graph of Exercise 1?
4. What is the adjacency matrix for the graph of Exercise 1?
5. The "eigenvalues" of a graph are defined to be the eigenvalues of its
adjacency matrix. What are the eigenvalues of the graph in Figure 1.1?
6. The "valence sequence" of a graph is the list of all the valences of its
vertices (usually in increasing order), so that if there are n vertices, there
are n numbers in the valence sequence. Construct two graphs that have
the same valence sequence, but that are not isomorphic.
7. (a) Prove that there is no graph map from the graph G' of Figure 1.4 to
the graph G of that figure. (b) Prove that G and G' are not isomorphic.
8. Draw two nonisomorphic graphs with four vertices and four edges each
and no loops or multiple adjacencies.
9. Prove that if the graph map f: G ~ G' is an isomorphism, then the
r
inverse map 1: G' ~ G is also an isomorphism.
10. Find another isomorphism H ~ H' for the graphs of Figure 1.5 besides
the one given in Example 1.1.4.
11. In permutation notation, list all the automorphisms of the graph H of
Figure 1.5.
12. Draw a graph with two vertices and no automorphisms except for the
identity.
13. Draw a graph with six vertices, no loops, and no automorphisms except
the identity.
14. Identify the automorphism group of the graph of Figure 1.2.
15. Draw a graph whose automorphism group is isomorphic to the cyclic
group ~3 •
16. Prove that every graph has an even number of vertices of odd valence.
17. Two graphs are called "co-spectral" if they have the same eigenvalues
(see Exercise 5) with the same multiplicities. Draw two co-spectral,
connected, 6-vertex, 7-edge graphs.

1.2. SOME IMPORTANT CLASSES OF GRAPHS

Numerous classes of graphs, arising in algebra or geometry as well as in

combinatorics or topology, are of special interest in topological graph theory.
In this section we examine a few such classes.
8 Introduction

1.2.1. Walks, Paths, and Cycles; Connectedness

A "walk" in a graph is the combinatorial analog of a continuous image of a
closed line segment, which may cross itself or retrace upon itself, backwards or
forwards, arbitrarily often. Precisely, if u and v are vertices of a graph G, then
a "walk of length n from u to v" is an alternating sequence of vertices and
directed edges,

whose initial vertex v0 is u, whose final vertex vn is v, and fori = 1, ... , n, the
directed edge e 1°1 runs from the vertex D;_ 1 to the vertex v,. If u-:/= v, then W
is called an "open walk". If u = v, then W is called a "closed walk".
An open walk is called a "path" if its vertices are distinct. Thus, a path is
the combinatorial analog of a homeomorphic image of a closed line segment.
The standard path with n vertices is called the "n-path" and is denoted Pn. A
closed walk is called a "cycle" if every pair of vertices except its starting and
stopping vertex are distinct. Thus, a cycle is the combinatorial counterpart to
the homeomorphic image of a circle. The standard cycle with n vertices is
called the "n-cycle" and is denoted en.
In order to relieve the cumbersomeness of the notation for a walk, path, or
cycle, one commonly omits the vertices from the sequence specified in the
definition, since there is no difficulty in inferring them from the resulting
sequence of directed edges. Moreover, when for i = 1, ... , n - 1 the edges e;
and ei+l have only one endpoint in common, one usually also omits the signs,
and writes

There are many different notations for a walk, path, or cycle, and the sensible
practice in a particular instance is to use the least cumbersome notation that
makes sense, even if it is not the same as the notation used in a previous
instance. Thus, fluency has priority over consistency.
A graph is called "connected" if for every pair of vertices u and v, there is a
path from u to v. Specifying plus directions on some of the edges does not
affect the connectedness of a graph, since the direction on an edge in a path
(or walk) is permitted to be the opposite of the plus direction. Thus, con-
nectedness in the present combinatorial sense agrees exactly with the underly-
ing topological sense.

1.2.2. Trees
A "tree" is a connected graph with no cycles, as illustrated in Figure 1.6. It is
one of the most important kinds of graphs in both applications and theory.

Theorem 1.2.1. Let u and v be any vertices of a tree T. Then there is a unique
path in T from u to v.
Some Important Gasses of Graphs 9

•
I
Figure 1.6. Four trees.

Proof Since the tree T is connected, there exists a path P from u to v.

Suppose that P' is another path from u to v. Then let w be the first vertex on
P and P', as we proceed from u to v, whose successor in the path P is not its
successor in the path P'. Let w' be the next vertex on P after w that also lies
on the path P'. Then the path segment from w to w' in P combines with the
path segment from w to w' in P' to form a cycle in T, which contradicts the
fact that trees have no cycles. D

Theorem 1.2.2. Let T be a tree. Then #Vr = #Er + 1.

Proof If #Vr = 1, then #Er = 0; otherwise there would be a loop, which

yields a cycle. By way of induction, assume the proposition is true for trees
with p or fewer vertices, and let T have p + 1 vertices. It follows from
Theorem 1.2.1 that removing an edge e from T disconnects it into two smaller
trees T' and T", so that #Vr = #VT' + #VT" and #Er = #ET' + #ET" + 1.
By the induction hypothesis, #VT' = #ET' + 1 and #VT" = #VT' + #ET" + 1.
It follows that #Vr = #Er + 1. D

Theorem 1.2.3. Every nontrivial tree T has at least two vertices of valence 1.

Proof By Theorem 1.1.1, for any graph, the sum of the valences is twice
the number of edges. By Theorem 1.2.2, 2#Er = 2#Vr - 2, since T is a tree.
By combining these two facts, one obtains the equation

L valence{ v) = 2#Vr- 2
ve Vr

Since every vertex v of the tree has valence at least 1, there must be at least
two vertices of valence exactly 1. D
 10 Introduction

Figure 1.7. Some complete graphs.

1.2.3. Complete Graphs

A simplicial graph is called "complete" if every pair of vertices is adjacent. If
two complete graphs have the same number of vertices, then they are isomor-
phic. The standard model for a complete graph on n vertices is denoted K n.
Figure 1.7 shows the complete graphs Kn, for n = 1, ... ,6. The graph K 1 is
sometimes called the "trivial graph".

1.2.4. Cayley Graphs

Cayley (1878) used graphs to draw pictures of groups, an inspiration that was
subsequently refined and used to algebraic advantage by Dehn (1911), Schreier
(1927), and others. As preliminary examples, consider the two drawings in
Figure 1.8 of the group ~5 of integers modulo 5.
In both diagrams of Figure 1.8, the elements of the group ~5 are the
vertices. Figure 1.8a shows ~5 as generated by the single residue class 1, so
from every vertex v there is a single solid edge to the vertex v + 1. Figure 1.8b
shows ~5 as generated by the two residue classes 1 and 2, so from each vertex
v there is a solid edge to the vertex v + 1 and also a dashed edge to the vertex
v + 2.
Rather than solid lines, dashed lines, and other such devices to distinguish
one class of edges from another, Cayley suggested a different color for each
generator. From this arose the name "Cayley color graph". A common abuse

0 0

4 1 4 1
--~--2

(a) (b)
Figure 1.8. Two Cayley color graphs for .2'5 : (a) with generating set {1}; (b) with generating set
{1, 2 }.
Some Important Gasses of Graphs 11

of terminology, adopted here, is to refer to whatever distinguishing features

are used for edges in the drawings as "colors", and also to refer to the
generators themselves (when they are regarded as generators) as "colors".
Although the construction now defined is more general than Cayley's, his
name is retained. As his generating set, Cayley used all the nonidentity
elements.
Let d be a group, and let X= { x 1, ... , xn} be a generating set for d.
The "(right) Cayley color graph" C(d, X) has as its vertex set the elements of
group d and as its edge set the cartesian product X X d. The edge (x, a) has
as its endpoints the vertices a and ax, with its plus direction from a to ax.
Sometimes it is convenient to write (x, a) in the subscript notation X 0 •
The designation of a plus direction and a color for every edge is an intrinsic
part of a Cayley color graph C(d, X). If these designations are suppressed,
the result is a graph C(d, X) 0 , called the "Cayley graph" for the group d
and the generating set X. Figure 1.9 shows a Cayley color graph, with its edge
names given in subscript notation, and the associated Cayley graph. The
names of vertices (e.g., a) and edges [e.g., (x, a) or xal are the same in a
Cayley graph as in the Cayley color graph from which it is derived. Since the
name of an edge of a Cayley color graph is readily determined from its color
and its initial point, names of edges are hereafter omitted from drawings,
thereby decreasing the visual clutter.
A graph map C(d, X)~ C(d', X') between Cayley color graphs is called
"color-consistent" if for every pair of edges with the same color in C(d, X),
their images have the same color in C(d', X'). It is called "direction-preserv-
ing" if the initial point of every edge in C(d, X) is mapped onto the initial
point of the image edge in C(d', X'). It is called "identity-preserving" if it
maps the group-identity vertex of C(d, X) to the group-identity vertex of

1 1

3 3 5

(a) (b)
Figure 1.9. (a) The Cayley color graph for the group .2"6 with the generators 2 and 3. (b) The
associated Cayley graph C(.2"6 , {2, 3 }) 0 .
12 Introduction

C( .J/1', X'). The following theorem shows the connection between Cayley color
graphs and group homomorphisms.

Theorem 1.2.4. Let f: C(.J/1, X)~ C(.J/1', X') be a color-consistent, direc-

tion-preserving, identity-preserving graph map. Then its vertex function coincides
with a group homomorphism .J/1 ~ .J/1'. Conversely, any group homomorphism
h: .J/1 ~ .J/1' such that h( X) c X' coincides with the vertex function of a color-
consistent, direction-preserving, identity-preserving graph map C( .J/1, X) ~
C( .J/1 I' X').

Proof Assume that the first hypothesis holds. To show that f is a group
homomorphism, it suffices to show that f(ax") = f(a)f(x)" for all a E .J/1,
x E X, and o = ± 1. Since there is an edge in the Cayley graph C(.J/1, X) from
a to ax with the color "x", it follows that in the Cayley graph C(.J/1', X')
there is an edge from /(a) to f( ax). Moreover, this edge has the same color as
the edge from 1..,, to f(x), since f is color-consistent, direction-preserving,
and identity-prese~ng. The edge from 1..,, to f(x) must be colored "f(x)", so
the edge from f(a) to f(ax) is also colored "f(x)". By the definition of
Cayley color graph, it follows that f(ax) = f(a)f(x). To show that f(ax- 1 )
= f(a)f(x)- 1, we now simply observe that f(a) = f(ax- 1x) = f(ax- 1 )/(x).
Conversely, let h: .J/1~ .J/1' be a group homomorphism such that h(X) c X'.
Define the graph map f: C(.J/1, X)~ C(.J/1', X') as follows. If a is a vertex of
C(.J/1, X), let f(a) = h(a). If e is an edge from a to ax (colored x), let f(e)
be the edge from f(a) to /(ax)= f(a)f(x) [colored /(x)]. It is easily verified
that f has the desired properties. D

Remark. In regard to Theorem 1.2.4, it may be observed that the vertex

function of a color-consistent, direction-preserving, identity-preserving graph map
from one Cayley color graph to another coincides with a group isomorphism if and
only if the graph map is a graph isomorphism.

An alternative definition of a Cayley graph is like the one already given,

except that when a generator x has order 2 in the group .J/1, for every element
a E .J/1 the two edges

xa (from a to ax) and xax (from ax to a)

of the Cayley color graph C(.J/1, X) collapse onto a single edge, denoted either
xa or xax· The "alternative Cayley graph" is denoted C(.J/1, X) 1. Both
C(.J/1, X) 0 and C(.J/1, X) 1 are called "Cayley graphs". Figure 1.10 shows a
Cayley color graph for a group with a generator of order 2 and the corre-
sponding alternative Cayley graph.
It may be observed that the Cayley color graphs in Figures 1.9 and 1.10
look quite alike, since their corresponding Cayley graphs are isomorphic, as
are their corresponding alternative Cayley graphs, even though the groups .2'6
Some Important Oasses of Graphs 13

(1 3) (1 3)

(1 2) (1 2) (2 3)

(a) (b)

(12 3) }
(12) - - - - - -
Figure 1.10. (a) The Cayley color graph for the symmetric group .9'3 with the generators (12 3)
and (1 2). (b) The alternative Cayley graph C( .9'3 , { (123), (12)} )1 .

and .51'3 are not isomorphic. Moreover, it is easy to construct a color-con-

sistent, identity-preserving graph isomorphism between those two Cayley color
graphs, but it does not preserve directions, of course, because of Theorem
1.2.4.
A graph is called "n-regular" (or just "regular") if every vertex has valence
n. If a group is presented with q generators, then the corresponding Cayley
graph is 2q-regular. The alternative Cayley graph is also regular, but the
valence depends on the number of generators of order 2. Regularity of a graph
is best understood as a local symmetry condition.
A graph G is called "vertex-transitive" if for every pair of vertices u and v,
there is an automorphism of G that carries u to v. Vertex-transitivity is best
regarded as a global symmetry condition.
As preparation for proving that a Cayley graph C(.J/1, X) 0 or C(.J/1, X) 1 is
vertex-transitive, define for every a E .J/1 the graph automorphism i a, accord-
ing to the rules

for every vertex u E .J/1

i a {( x, u)) = ( x, au) for every edge ( x, u) E X X .J/1

The graph automorphism ia is called a "left translation" of the Cayley graph.

Theorem 1.2.5. Any Cayley graph is vertex-transitive.

Proof For any pair of vertices u and v, let w = vu- 1• Then the left
translation i w takes u to v. D
14 Introduction

The converse of Theorem 1.2.5 is not true; for example, the Petersen graph
(see Example 1.3.13 and Exercise 2.2.13) is not a Cayley graph, yet it is
vertex-transitive. On the other hand, a left translation ia of a Cayley graph
also has the property that it leaves no vertex fixed, unless of course the group
element a is the identity. If the automorphism group of a graph G has a
subgroup .91 that is vertex-transitive, and if every nonidentity element of .91
leaves no vertex fixed, then indeed the graph G is a Cayley graph for the group
.91 (see the corollary to Theorem 2.2.3 and see Theorem 6.2.3).
The Cayley graph associated with a given group and generating set reveals
much about the group. Every walk in the graph corresponds to a "word" in
the generators; every closed walk corresponds to a "relator", namely a product
of generators that is equal to the identity. This relationship has proved to be a
powerful tool in combinatorial group theory, which is the study of presenta-
tions of groups in terms of generators and relations. It was Dehn (1911) who
first focused on this connection. For example, one of Dehn's three famous
decision problems for a presentation of a group is to find an algorithm that
will determine whether two given words in the generators represent the same
element; this is equivalent to finding an algorithm that constructs the Cayley
graph corresponding to the given presentation of the group. [Boone (1954) and
Novikov (1955) showed this "word problem" was unsolvable for some group
presentations.]
Dehn's influence is so great that Cayley graphs are sometimes called "Dehn
Gruppenbilder", although Cayley's priority is unquestioned [Burnside's fa-
mous Theory of Groups (1897, 1911) refers to "Cayley's colour groups" and
even has a four-color illustration of a Cayley graph as the frontispiece.] More
recent examples of the importance of graphical representations of groups can
be found in the work of Stallings (1971) and Serre and Bass (1977) (see also
Tretkoff, 1975, 1980). Cannon (1984) carries the word problem one step
further by studying presentations for infinite groups that have linear-time
algorithms to draw the associated Cayley graph. General monographs for
combinatorial group theory have been compiled by Magnus, et al. (1966), by
Lyndon and Schupp (1977), and by Coxeter and Moser (1957).

1.2.5. Bipartite Graphs

A graph is called "bipartite" if its vertex set can be partitioned into two
subsets U and W such that the vertices in U are mutually nonadjacent and the
vertices in W are mutually nonadjacent. If every point of U is adjacent to

Figure 1.11. The complete bipartite graph

K3.4·
Some Important Classes of Graphs 15

Figure 1.12. A bouquet of three circles.

every point of W, then the graph is called "complete bipartite" on the sets U
and W. The standard model for a complete bipartite graph on sets of m
vertices and n vertices is denoted Km, n. Figure 1.11 illustrates the complete
bipartite graph K 3, 4 •

1.2.6. Bouquets of Circles

For topological graph theory, one other class of graphs should be (and is)
introduced at the outset. A "bouquet of n circles" is a graph with one vertex
and n loops. The standard model is denoted Bn. The bouquet B3 is illustrated
in Figure 1.12.

1.2.7. Exercises
1. In the graph of Figure 1.10, find a path with six vertices and a cycle of
length five. Also find a closed walk with four edge-steps that is not a
cycle.
2. Prove that if .511 and .511' are groups of the same order, they have
respective generating sets X and X' such that the Cayley graphs
C(.Jil, X) 0 and C(.Jil', X') 0 are isomorphic. (Hint: Consider relatively
large generating sets.)
3. Show that a graph is connected if and only if for every pair of vertices u
and v, there is a walk from u to v.
4. Extend Theorem 1.2.4 to the case where the group .511 is infinite.
5. Draw the Cayley color graph for the group fr2 X fr2 X fr2 with genera-
tors (1, 0, 0), (0, 1, 0), and {0, 0, 1).
6. Draw the Cayley graph for the group fr8 with generating set {1, 2, 3}.
7. Draw the alternative Cayley graph for the group :Z6 with generating set
{1, 2, 3}.
8. Draw the Cayley color graph for the quaternions with generators i and j
and relations i 2 = / = {ij) 2 •
9. Draw the Cayley color graph for the alternating group .5114 with genera-
tors (1 2 3) and (2 3 4).
10. Draw a 4-regular simplicial graph on 11 vertices that is not a Cayley
graph. (Hint: Use two copies of K 5 .)
16 Introduction

11. Decide whether the complete bipartite graph K 3 3 is a Cayley graph.

(Hint: There are only two groups of order 6.) '
12. Prove that every complete graph Kn is a Cayley graph. (Hint: Use the
group ~n· When n is even, an alternative Cayley graph construction is
required.)
13. Prove that every tree is a bipartite graph.
14. Which n-cycles are bipartite?
15. What is the automorphism group of the complete graph Kn?
16. Verify that a left translation of a Cayley graph is a graph automorphism.
17. Prove that every Cayley graph is connected.

1.3. NEW GRAPHS FROM OLD

There are numerous ways to extract parts, make combinations, or do other

operations on graphs so as to obtain new graphs.

1.3.1. Subgraphs
If G and G' are graphs, then G' is called a "subgraph" of G if and only if VG'•
EG'• and JG, are subsets of VG, EG, and IG, respectively. If in addition,
VG' = VG, then one says that the subgraph G' "spans" the graph G. Figure
1.13 illustrates a graph at the left and two of its subgraphs.
If a subgraph is a tree and if it spans the graph, it is called a "spanning
tree". For instance, the graph in the middle of Figure 1.13 is a spanning tree
for the graph at the left. The subgraph at the right does not span, since it does
not include the vertex v4 • Nor is it a tree, since it is not connected.

Example 1.3.1. Every n-vertex simplicial graph is isomorphic to a subgraph of

the complete graph K n. Also, the n-path Pn is isomorphic to a subgraph of the
(n + k)-path Pn+k• for any nonnegative integer k. However, the m-cyc/e is not
isomorphic to a subgraph of the n-cyc/e unless m = n.

u2 u2 u2

ul
u3
e6
u4 ul u3
e6
u4 u
A 3

u5 u5 •
u5
Figure 1.13. A graph and two of its subgraphs.
New Graphs from Old 17

G Gl
Figure 1.14. A graph and two of its subgraphs.
=
For any graph G, let V' c Vc. The "induced subgraph" on the vertex
subset V' is the subgraph whose vertex set is V' and whose edge set consists of
every edge e of G whose endpoint set V(e) lies in V'. For instance, in Figure
1.14, the subgraph G 1 is an induced subgraph of G, but the subgraph G2 is
not.
There is a big difference between assertions about subgraphs and about
induced subgraphs. For example, since the complete graph Kn is, in fact, a
Cayley graph (Exercise 1.2.12), it follows that every finite graph is a subgraph
of some Cayley graph. The claim, however, that every finite graph is an
induced subgraph of some Cayley graph, seems quite surprising. This would
imply that although Cayley graphs are globally symmetric, locally they can be
arbitrarily complicated. Nevertheless, Babai (1976) has shown this claim to be
true. In fact, Godsil and Imrich (1985) prove that given any graph G on n
vertices and any group .91 of order greater than (24/3)n 3 there is a Cayley
graph for .91 containing G as an induced subgraph.
A "component" of a graph is a maximal connected subgraph. Thus, if a
graph is connected, it has only one component, itself.

1.3.2. Topological Representations, Subdivisions, Graph Homeomorphisms

Any graph G can be represented by a topological space in the following sense.
Each vertex is represented by a distinct point and each edge by a distinct arc,
homeomorphic to the closed interval [0, 1]. Naturally, the boundary points of
an arc represent the endpoints of the corresponding edge. (Of course, the
interiors of the arcs are mutually disjoint and do not meet the points
representing vertices.) Such a space is called a "topological representation"
of G.
Subdivision is an operation that does not change the homeomorphism type
of a topological representation of a graph. For simplicity, one first considers
the effect of the operation on a single edge e of a graph G, with endpoint set
V( e) = { v 1 , v2 }. From the informal viewpoint of drawing pictures, one subdi-
vides the edge e into two new edges e' and e" by putting a new vertex v
anywhere in its interior, as shown in Figure 1.15.
18 Introduction
e u

u4 u4
Figure 1.15. The graph at the right is obtained from the graph at the left by subdividing the
edge e.

From a formal combinatorial viewpoint, the result of "subdividing the

edge" e with endpoints v1 and v2 is a new graph whose vertex set is the union
of V and { v }, whose edge set is (£ - { e}) u {e', e"}, and whose incidence
structure is (I - V( e)) U { V( e'), V( e")}, where V( e') = { v1 , v} and V( e") =
{ v2 , v }. In general, a "subdivision of a graph is obtained by a finite sequence
of subdivisions of edges.
Among other uses, subdivision can make a graph become simplicial. The
first step is to subdivide every loop. The next and last step is to subdivide
every edge but one in each multiple adjacency.
The graphs G and H are called "homeomorphic" if they have respective
subdivisions G' and H' such that G' and H' are isomorphic graphs. It may be
observed that under this combinatorial definition, two graphs are homeomor-
phic if and only if their topological representations are homeomorphic as
topological spaces. Figure 1.16 shows two homeomorphic graphs such that
neither is a subdivision of the other. Homeomorphism is actually a point-set
topological concept, yet the "Hauptvermutung" [see Papakyriakopoulos (1943),
Moise (1952), Bing (1954), and Brown (1969)] enables us to define it combina-
torially for low-dimensional complexes. Milnor (1961) proved that Haupt-
vermutung does not hold in higher dimensions.

Example 1.3.2. For any two positive integers m and n, the m-cyc/e em and the
n-cyc/e en
are homeomorphic graphs.

Example 1.3.3. For any two integers m, n :;:: 2, them-path Pm and then-path
Pn are homeomorphic graphs.

Figure 1.16. Two homeomorphic graphs.

New Graphs from Old 19

ul

,, [ X
uJ.
eJ.
u2
e2
u§

u2
(ul, ul) (ul, u2) (ul, u 3)
(u 1, ei) (ul, e2)

(el, u2)

(u2 , ei)

(u2, ul) (u2, u2) (u2, u§)

Figure 1.17. The cartesian product of two graphs.

1.3.3. Cartesian Products

The "cartesian produc(' of the graphs G and G' is denoted G X G' and
defined to be the graph with vertex set Vc X Vc· and edge set ( Ec X Vc·) u
(Vc X Ec-). If the edge (e, v') E Ec X VG'• and if the endpoints of the edge e
are v1 and v2 then the endpoints of the edge (e, v') are the vertices ( v1 , v') and
(v 2 , v'). If (v, e') E Vc X Ec·· and if the endpoints of the edge e' are v1' and
v2', then the endpoints of the edge (v, e') are (v, v1') and (v, v2'). Figure 1.17
shows the cartesian product of the 2-path and the 3-path.

Example 1.3.4. The Cayley graph C(.2'6 , {2, 3}) 0 illustrated in Figure 1.9 is
isomorphic to the cartesian product of the 2-cyc/e c2 and the 3-cyc/e c3 .

Example 1.3.5. The alternative Cayley graph C( .51'3 , { (123), (12)} ) 1 illustrated
in Figure 1.10 is isomorphic to a cartesian product of the 3-cyc/e C3 and the
complete graph K 2 •

A graph consisting of the vertices and the edges of a polyhedron, with

incidence as on the polyhedron, is called the "!-skeleton" of the polyhedron. It
may be observed that the cartesian product of two graphs is the !-skeleton of
their topological product as complexes.

Example 1.3.6. The "n-cube graph" Qn is defined to be the !-skeleton of the

n-dimensional cube. Equivalently, it may be defined recursively using the oper-
ation of cartesian product, with Q 1 = K 2 , and for n ~ 1, Qn+l = K 2 X Qn.

1.3.4. Edge-Complements
The "edge-complement" of a simplicial graph G is denoted Gc and is defined
to be the graph with the same vertex set as G but such that two vertices are
adjacent if and only if they are not adjacent in G.
20 Introduction

Figure 1.18. The suspension of a 3-path from a 2-path.

Example 1.3.7. The edge-complement Knc of the complete graph has no edges at
all but has n vertices.

Example 1.3.9. The "n-couple cocktail-party graph" is defined to be the

edge-complement of the union of n disjoint copies of K 2 • The explanation of this
graph's name, due to Alan Hoffman, is that each of the 2n persons at the
cocktail party talks to each other person except for his or her spouse, and that two
vertices are adjacent if the persons they represent talk to each other at the party.

1.3.5. Suspensions
The "suspension" (elsewhere called the "join") of a graph G from a graph G'
is obtained by adjoining every vertex of G to every vertex of G', and is
denoted G + G'. Thus, its vertex set is Vc u Vc, and its edge set is Ec u Ec,
u (Vc X Vc,). The endpoints of an edge of G + G' that arises from Ec or
from Ec, are exactly as in the graphs G or G', respectively. The endpoints of
an edge ( v, v') that arises from Vc X Vc, are the vertices v and v'. Figure 1.18
shows the suspension of a 3-path from a 2-path.

Example 1.3.10. The complete graph Kn+ 1 is isomorphic to the suspension of

the complete graph K n from the graph K 1.

Example 1.3.11. The "n-octahedron graph" On is defined recursively, with

0 1 = K 2c, and for n ~ 1, On+ 1 = On+ K/. It may be observed that the
n-octahedron graph is isomorphic to the 1-skeleton of the n-dimensional oc-
tahedron and is also isomorphic to then-couple cocktail-party graph.

1.3.6. Amalgamations
Let G and G' be graphs, and let f: H ~ H' be an isomorphism from a
subgraph H of G to a subgraph H' of G'. The "amalgamation" G • 1 G' is
obtained from the union of G and G' by identifying the subgraphs H and H'
according to the isomorphism. Figure 1.19 shows the amalgamation of a
3-cycle and a 4-cycle along the thickened edge.

Example 1.3.12. The bouquet Bn+k is an amalgamation of the bouquets Bn and

B k at a vertex.
New Graphs from Old 21

f(v2)
Figure 1.19. The amalgamation of two graphs.

1.3.7. Regular Quotients

Let G be a graph, and let !!J be a group such that for each element b E !!J,
there is a graph automorphism cph: G ~ G and such that the following two
conditions hold:

i. If 1 is the group identity, then cp 1 : G ~ G is the identity automor-

phism.
ii. For all b, c E !!J, cpbo <Pc = <Pbc·

Then the group !!J is said to "act on (the left of) the graph" G. If, moreover,
the additional condition

iii. For every group element b * 1, there is no vertex v E VG such that

<Pb(v) = v and no edge e E EG such that <Pb(e) =e.

holds, then !!J is said to "act without fixed points" (or sometimes "act freely")
on G.
For any vertex v of the graph G, the "orbit" [v] (or sometimes [v]&r) is
defined to be the set {<Ph( v ): b E !!J} of all images of v under the group
action. Similarly, for any edge e, the orbit [e) (or sometimes [e]&r) is defined to
be the set {<Pb(e):bE!!J}. The sets of vertex orbits and edge orbits are
denoted Vj!!J and Ej!!J, respectively. It may be observed that the orbits
partition the vertex set and the edge set of G.

Example 1.3.13. Figure 1.20 contains a drawing of the" Petersen graph" in the
plane. ForiE fr5 , let <P, be the rotation of 27Tij5 radians. Under this action of
fr5 , there are two vertex orbits, one containing the outer five vertices and the
other containing the inner five. There are three edge orbits, one containing the
five edges of the outer pentagon, one containing the five edges of the inner star,
and one containing the five edges that run between the star and the pentagon.

Most important, if .511 is a group, if !!J is a subgroup, and if X is a

generating set, then !!J acts freely on the Cayley graph C(.Jil, X) 0 according to
the rules <Ph( a)= ba for vertices and <Pb(xa) = xha for edges.
22 Introduction

Figure 1.20. The Petersen graph.

Example 1.3.14. The n-cycle Cn is isomorphic to the Cayley graph C( .2'n, {1 }) 0 ,

so the group .2'n acts freely on it. There is only one vertex orbit and only one edge
orbit. It may be observed that although the dihedral group ~n is isomorphic to
Aut(Cn), it does not act freely, since each reflection fixes either a vertex or an
edge.

The "regular quotient" G1114 is the graph with vertex set VI 114 and edge set
E 1114 such that the vertex orbit [ v 1is an endpoint of the edge orbit [e 1if any
vertex in [ v 1 is an endpoint of any edge in [e 1. It is easy to verify that each
edge orbit has one or two vertex orbits as endpoints.

Example 1.3.13 Revisited. Figure 1.21 shows the quotient of the Petersen graph
under the prescribed action of .2'5 •

Example 1.3.14 Revisited. The quotient of the n-cycle Cn under the action of the
cyclic group .2'n is the bouquet B 1 of one circle. In general, when a group acts on
a Cayley graph for itself, the quotient is a bouquet of circles, one corresponding
to each generator.

1.3.8. Regular Coverings

Associated with a graph quotient there is a vertex function v ~ [ v1 and an
edge function e ~ [e1 that together are called the "quotient map" q~: G ~
GIPA.

Figure 1.21. A regular quotient of the Petersen graph.

New Graphs from Old 23

A graph map p: G ~ G is called a "regular covering projection" if it is

equivalent to a quotient map in the following sense: there exists a group !!4
that acts freely on the domain G, and there exists an isomorphism i: Gj!!A ~ G
such that i o q 8 = p, where q!ll: G ~ Gj!!A. The range graph G is called the
"base" (or "base space"), and the domain graph G is called a "regular
covering" (or "regular covering space").

Example 1.3.15. For n ~ 1 and r ~ 1, there is a regular covering projection

C(~rn' {1}) 0 ~ C(~n' {1}) 0 defined by the vertex function imodulorn ~
i modulo n and the edge function ( i, j modulo rn) ~ ( i, j modulo n ). The effect of
this regular covering projection is to wrap an rn-cycle around an n-cycle r times.
The group ~r acts freely on the domain Cayley graph.

For any vertex v in a graph G, the "star" of v, denoted star(v), is the

subgraph consisting of the vertex v and all the vertices adjacent to it and of all
the edges joining those other vertices to v. If the graph G is simplicial and if
the vertex v has valence k, then star(v) is isomorphic to the bipartite graph
Kl,k·
A graph map f: G' ~ G is called a "local isomorphism" if for every vertex
v in G', the restriction off to star(v) is an isomorphism onto star(/(v)). It
may be observed that the covering projection of Example 1.3.15 is a local
isomorphism for n ~ 3 but not for n = 1 or 2 and r ~ 2. One exercise for this
section is to show that if its base space is simplicial, then a covering projection
is a local isomorphism. To emphasize that it is more than a local isomorphism,
a graph isomorphism is sometimes called a "global isomorphism".

1.3.9. Exercises
1. How many different spanning trees are there in the graph at the left of
Figure 1.13? How many different isomorphism types of spanning trees
are there? How many isomorphism types of subgraphs are there?
2. Prove that every graph is homeomorphic to a bipartite graph. (Hint: Use
the operation of subdivision.)
3. How many vertices does the n-cube graph Qn have? How many edges?
4. How many vertices and edges does the n-octahedron graph On have?
5. A suspension K 1 + Cn is called a "wheel". Why?
6. Draw the suspension K/ + C4c, the suspension K/ + C6 , and the carte-
sian product K 2 X K 1, 3 .
7. Describe a free action of the cyclic group ~3 on the complete bipartite
graph K 3 , 3 and draw the quotient.
8. Describe a free action of the cycle group ~2 on the 3-cube graph Q 3
whose quotient is isomorphic to the complete graph K 4 •
9. Prove that the cyclic group ~7 cannot have a free action on the Petersen
graph.
24 Introduction

10. Prove that the cyclic group .2'6 cannot have a free action on the complete
bipartite graph K 3, 3 •
11. Prove that the even-order cyclic group .2'2 n cannot have a free action on
the complete graph K 2 n.
12. Prove that the odd-order cyclic group .2'2 n+ 1 has a free action on the
complete graph K 2 n + 1 •
13. Construct a simplicial graph whose automorphism group acts freely on it
and is a nontrivial group.
14. Prove that a local isomorphism from a connected graph onto a tree is a
global isomorphism.
15. Prove that every automorphism on a tree has either a fixed vertex or a
fixed edge. (Hint: Those familiar with the "center" of a tree may find
that concept helpful.)
16. Prove that a covering projection p: G ~ G with a simplicial base graph
G is a local isomorphism.
17. Let G and G' be graphs, and let v E Vc and v' E Vc'. Prove that the
induced subgraphs on the vertex sets Vc X { v'} and { v} x Vc' are
isomorphic toG and G', respectively, in the product graph G X G'.

1.4. SURFACES AND IMBEDDINGS

The central concern of topological graph theory is the placement of graphs on

surfaces. Ordinarily, such surfaces are "closed", that is, compact and without
boundary. The plane is not closed, of course but since it differs from the
sphere by only a single point, it follows that a given graph can be imbedded in
the plane if and only if it can be imbedded in the sphere. Accordingly, nothing
is lost by considering imbeddings in the sphere, as is the present practice,
rather than in the plane.

1.4.1. Orientable Surfaces

There are two kinds of closed surfaces, orientable and nonorientable. The
sphere, the torus, the double torus, the triple torus, and so on, as illustrated in
Figure 1.22, are orientable. They are commonly denoted S0 , S1 , S2 , S3 , •..•
Indeed, it is proved in Chapter 3 that every closed connected orientable
surface is homeomorphic to one of them.

08888:8:8 ... Figure 1.22. The closed orientable surfaces.

Surfaces and Imbeddings 25
e'
e' e'

e e
C> C>
e
fffiL
Figure 1.23. Adding a handle to eliminate an edge-crossing.

Another characterization of the closed orientable surfaces is that each one

can be obtained by adding some handles to a sphere in 3-space. Adding one
handle yields S 1, adding two yields S2 , and so on. From this characterization,
it is easily seen that every (finite) graph can be drawn without edge-crossings
on some closed surfaces, as follows. First, draw the graph on the sphere,
possibly with crossings. Now suppose that the edge e crosses the edge e'. As a
surgical operation, cut a hole in the surface on each side of e', near the
crossing by e, small enough so that it touches no edge. Next attach a handle
from one small hole to the other. Then reroute edge e so that it traverses the
handle instead of crossing edge e'. This procedure, illustrated in Figure 1.23,
can be repeated until all crossings are eliminated.

1.4.2. Nonorientable Surfaces

The "Mobius band" is a surface that is neither closed nor orientable. To
obtain a physical copy of a Mobius band, begin with a rectangular piece of
paper. Then, as illustrated in Figure 1.24, give the strip a half-twist, so as to
interchange the top and bottom at one side, and finish by pasting the right side
to the left. [See Mobius (1865).]
Although a Mobius band is compact, it has a boundary, which is homeo-
morphic to the circle, so it is not closed. What makes it nonorientable is that if
a 2 X 2 coordinate system specifying a forward direction and a right direction
is translated in the forward direction once around the center of the band, then
the orientation of the right direction is reversed. It seems from a local
viewpoint that a physical copy of a Mobius band is two-sided. However, one
can color "both" sides in a single hue without lifting the crayon or crossing
the boundary.

.______,X..__ _.
Figure 1.24. Constructing a Mobius band.
26 Introduction

The global one-sidedness of the Mobius band in 3-space is a result of the

imbedding in 3-space, not an intrinsic property of the Mobius band. For
instance, in the imbedding space Mb X [0, 1], where Mb is the Mobius band
and [0, 1] is the unit interval, the cross-section Mb X {1/2} is two-sided.
If one cuts a hole in a sphere, the resulting boundary is homeomorphic to
the circle, just as is the boundary of a Mobius band. The surface obtained by
attaching a Mobius band along its boundary to the hole in the sphere, thereby
closing off the hole, is called a "projective plane". As the reader may confirm
by experimentation with physical models, it is impossible to complete this
boundary matching in 3-space.
The "Klein bottle", another nonorientable closed surface, can be obtained
by cutting two holes in a sphere and then closing each of them off with a
Mobius band. Indeed, as it is proved in Chapter 3, every closed, connected
nonorientable surface can be obtained by cutting holes in a sphere and then
closing them off with Mobius bands. For k = 0, 1, ... , the surface obtained by
cutting k holes and closing them off with k Mobius bands is denoted Nk. In
this context a Mobius ba~d is often called a "crosscap". Finally, from the
discussion here, one may infer that a closed surface is nonorientable if and
only if it has a subspace homeomorphic to a Mobius band. See Klein (1911).

1.4.3. Imbeddings
To formalize the notion of a drawing without crossings, we define an "imbed-
ding" of a graph in a surface to be a continuous one-to-one function from a
topological representation of the graph into the surface. For most purposes, it
is natural to abuse the terminology by referring to the image of the topological
representation as "the graph".
If a connected graph is imbedded in a sphere, then the complement of its
image is a family of "regions" (or "faces"), each homeomorphic to an open
disk. In more complicated surfaces, the regions need not be open disks. If it
happens that they are all open disks, then the imbedding is called a "2-cell (or
cellular) imbedding". If the boundary circuit of an open disk region has one or
more repeated vertices, then the closure of the region is not a closed disk.
Nonetheless, whether the imbedding is a 2-cell imbedding depends only on
whether all the regions are open disks, not on whether the closures of the
regions are closed disks.
A convenient notation for the set of regions of a graph imbedding i: G --+ S
is Fe, where the letter F reminds one that the regions are something like the
faces of a polyhedron. If more than one imbedding of G is under considera-
tion, then the name of the imbedding should appear somewhere in the
notation for the set of regions. An even greater convenience, if G is the only
graph in the immediate discussion, is to use the notation F instead of Fe.
The "number of sides" (or "size") of a region f is defined to be the number
of edge-sides one encounters while traversing a simple circuit just inside the
boundary of the region, and is denoted s1 . As illustrated by the exterior region
Surfaces and Imbeddings 27

Figure 1.25. A spherical imbedding.

in Figure 1.25, the number of sides of a region need not equal the number of
edges in its boundary. Although there are seven sides to the exterior region,
there are only five edges in its boundary.
It may be observed that the imbedding in Figure 1.25 has a one-sided
region, which is called a "monogon". It also has a two-sided region, which is
called a "digon". (Some persons use the words "unigon" and "bigon", which
are mixtures of Latin and Greek. Fortunately, these mixtures have not become
as firmly entrenched as "hexadecimal".)
If every region of a graph imbedding is three-sided, then one says the
imbedding is "triangular". In a "quadrilateral imbedding", every region is
four-sided.

1.4.4. Euler's Equation for the Sphere

In a letter to Goldbach, Euler (1750) wrote that he had found a new formula
relating the numbers of vertices, edges, and faces of a (spherical) polyhedron:

#V- #E + #F = 2.
Cauchy (1813) gave basically a graph-theoretic proof of this "Euler equation",
but it was Lhuilier (1811) who sorted out the apparent exceptions (faces must
be 2-cells and the surface cannot have holes) and generalized the Euler
equation to all closed orientable surfaces. Listing (1861) later used the Euler
equation to study surfaces in detail; this work is commonly considered the
beginning of topology.
The left side of the Euler equation is called the "Euler formula". The right
side is called the "Euler characteristic" of the surface. Thus, the number 2 is
the Euler characteristic of the sphere. Before proving the Euler equation for
the sphere, it may be instructive to evaluate the Euler formula for the two
examples depicted in Figure 1.26.
The graph G in Figure 1.26 has five vertices and nine edges. Each of the
regions, including the "exterior" region, is three-sided. There are six regions in
all, since the exterior region is always counted in computations of the Euler
formula. Thus, #V - #E + #F = 5 - 9 + 6 = 2. The graph H in Figure
1.26 has four vertices and five edges. The imbedding has three regions. Thus,
#V- #E + #F = 4 - 5 + 3 = 2.

Theorem 1.4.1. Let i: G ~ S0 be an imbedding of a connected graph in the

sphere. Then #V- #E + #F = 2.
28 Introduction

G H

Figure 1.26. Two graph imbeddings in the sphere.

Proof This proof proceeds by induction on the number #F of regions.

First, observe that if #Fe = 1, then G must be a tree, since the Jordan curve
theorem implies that any cycle would separate the sphere. Thus, by Theorem
1.2.2, #Ve - #Ee = 1, from wnich it follows that #Ve - #Ee + #Fe = 2.
Now suppose that the Euler equation holds when the number of regions is
at most n, and suppose that #Fe= n + 1. Then some edge e lies in the
boundary circuit of two distinct regions. Since the two regions are distinct, the
subgraph G' obtained by removing the edge e is connected. Then #Fe'= #Fe
- 1 = n, so by induction, #Ve,- #Ee, + #Fe,= 2. Since #Ve, = #Ve,
#Ee, = #Ee - 1, and #Fe, = #Fe - 1, it follows that #Vc - #Ee + #Fe
= 2. D

1.4.5. Kuratowski's Graphs

The Euler equation is often used in conjunction with a relationship between
the numbers of edges and regions to prove that certain graphs cannot be
imbedded in the sphere. This relationship, called the "edge-region inequality",
is established by the following theorem.

Theorem 1.4.2. Let i: G ~ S be an imbedding of a connected, simplicial graph

with at least three vertices into any surface. Then 2#E ~ 3#F.

Proof The sum 'Lf E F sf of the numbers of sides of the regions counts
every edge exactly twice. Thus, 2#£ = '[,fEFsf. Since there are no loops or
multiple edges in the simplicial graph G, there are no monogons or digons in
the imbedding. Thus, for every region f, sf~ 3. It follows that 2#£ ~ 3#F.
D

The complete graph K 5 and the complete bipartite graph K 3, 3 are called
"Kuratowski' s graphs" (or the "Kuratowski graphs") because Kuratowski
(1930) proved that they are a complete set of obstructions to imbedding
graphs in the sphere, in the following sense:

Kuratowski's Theorem. The graph G has an imbedding in the sphere if and only
if it contains no homeomorph of K 5 or of K 3 , 3 •
Surfaces and Imbeddings 29

To see that K 5 is nonspherical, observe that it has five vertices and 10

edges. By the Euler equation #V- #E + #F = 2, there would have to be
seven regions to any spherical imbedding. However, this would violate the
edge-region inequality 2#£ ~ 3#F, since 2 · 10 < 3 · 7.
A proof that K 3, 3 is nonspherical begins in the same way. Since K 3• 3 has
six vertices and nine edges, the Euler equation implies that a spherical
imbedding would have five regions. However, the edge-region inequality is
satisfied, because 2 · 9 ~ 3 · 5. This proof is continued after a strengthening of
the edge-region inequality, which involves the notion of girth.
The "girth" of a graph is the number of edges in its shortest cycle. If a
graph has a loop, then its girth is 1. Otherwise, it is at least 2. If it also has no
multiple edges, then its girth is at least 3. The girth of a tree is infinite.
The key fact in establishing the edge-region inequality for simplicial graphs
is that the girth of a simplicial graph is at least 3. Since a bipartite graph has
no odd cycles, its girth cannot be 3. Thus, the girth of a bipartite simplicial
graph is at least 4. In particular, the girth of K 3 3 is 4, from which it follows
that 2#£ ~ 4#F. However, since a spherical ini.bedding of K 3, 3 would have
five regions, this stronger inequality would be violated, because 2 · 9 < 4 · 5.
For future use, it is convenient to isolate the "strong edge-region inequality",
as follows.

Theorem 1.4.3. Let G be a connected graph that is not a tree, and let i: G ~ S
be an imbedding. Then 2#£ ~ girth(G) · #F.

One direction of Kuratowski's theorem is now easily proved. Since K 5 and

K 3 , 3 are nonspherical, a graph G with an imbedding in the sphere cannot
possibly contain a homeomorph of either of them. A proof of the converse
appears in Section 1.6.

1.4.6. Genus of Surfaces and Graphs

If a surface is orientable, its "genus" is defined to be the number of handles
one must add to the sphere to obtain its homeomorphism type. That is, the
genus of the orientable surface Sg is g. If a surface is nonorientable, then its
"crosscap number" is defined to be the number of Mobius bands one must
attach to the sphere to obtain its homeomorphism type. Thus, the crosscap
number of the nonorientable surface Nk is k for k ~ 1. The notation N 0
means the 2-sphere, even though the 2-sphere is orientabl~.
The "(orientable) genus" of a graph G, denoted y(G) or simply y if G is
the only graph in context, is defined to be the smallest number g such that the
graph G imbeds in the orientable surface Sg. The "crosscap number" of the
graph G, denoted y(G) or y, is the smallest number k such that G imbeds in
the surface Nk. Sometimes one calls y(G) the "nonorientable genus" of G.
Since neither K 5 nor K 3 3 imbeds in the sphere, they both have positive
genus and positive crosscap ~umber. Figure 1.27 establishes that both K 5 and
30 Introduction

1 0 1
I I

·! iI
.<(!>' ~ t
3
I
4
a
I

l! '0), it
4 3
I
I
a

I I I I
I

Figure 1.27. lmbeddings of K 5 and K 3, 3 in the Mobius band.

K 3 3 have imbeddings in the Mobius band, and consequently in the projective

plane N1• Thus they both have crosscap number 1. A calculation of their
orientable genera appears later in this section.
The imbedding of K 5 in Figure 1.27 shows the vertices and all but two of
the edges in the "center" of the Mobius band. The edge from vertex 2 to
vertex 4 runs through the middle of the pasting edge, and the edge from vertex
1 to vertex 3 also passes through the pasting edge. It is precisely the half-twist
before pasting that makes the imbedding possible. The imbedding of K 3 , 3
depends on the same trick.
The method described earlier in this section for using handles to eliminate
crossings does not give a very good upper bound on the orientable genus,
because it is usually highly inefficient to send only one edge across a handle.
For instance, it is impossible to draw K 6 in the sphere with fewer than three
crossings, but only one handle is needed to eliminate them all (see Exercise
1.4.3).
During the 1960s and 1970s there was an intensity of interest in finding
minimum imbeddings for graphs, largely due to the outstanding work of
Gerhard Ringel, who calculated the genus of the n-cube graphs (see Ringel,
1955), the genus of the complete bipartite graphs (see Ringel, 1965), and the
nonorientable genus of the complete graphs (see Ringel, 1959). After calculat-
ing the orientable genus of one-third of the complete graphs, he teamed with
J. W. T. Youngs to finish them off, thereby solving the Heawood map-coloring
problem. An exposition of this solution, noting the contributions of other
persons, is given in Chapter 5.

1.4.7. The Torus

Just as the Mobius band can be cut open and flattened into a rectangle, so can
the torus, thereby permitting one to draw imbeddings on a flat piece of paper.
To retrieve the torus from a rectangle, first paste the top of the rectangle to the
bottom, in order to obtain a tube. Then paste the right end of the tube to the
left, which yields the torus, as illustrated in Figure 1.28.
In order to show an imbedding of a graph on a torus, one may draw some
of the edges of the graph "through" the sides of a rectangle, taking care in
each instance to continue at the corresponding position on the opposite side.
For instance, Figure 1.29 shows imbeddings of K 5 and K 3, 3 on the torus.
Surfaces and lmbeddings 31

- - (}-------j
-®
Figure 1.28. Constructing a torus from a rectangle.

It has already been proved that neither K 5 nor K 3 , 3 imbeds in the sphere.
From Figure 1.29 it now follows that both have orientable genus 1. Although
K 5 has the same nonorientable genus as orientable genus, as does K 3• 3 , for
most graphs the orientable genus and the nonorientable genus differ.

1.4.8. Duality
Given a connected graph G, a closed surface S, and a 2-cell imbedding
i: G -+ S, there is an idea due to Poincare for constructing what are called a
"dual graph" and a "dual imbedding". First, for each region f of the
imbedding i: G -+ S, place a vertex f* in its interior. Then, for each edge e of
the graph G, draw an edge e* between the vertices just placed in the interiors
of the regions containing e. (If both sides of the edge e lie in the same region
j, then the dual edge e* is a loop based at the dual vertex f*.) The resulting
graph with vertices f* and edges e* is called the "dual graph" for the
imbedding i: G-+ S and is denoted G*', or simply G*, if i is the only
imbedding under consideration. The resulting imbedding of the graph G* in
the surface S is called the "dual imbedding". In this context, the original
graph G and the original imbedding i are called the "primal graph" and the
"primal imbedding". Figure 1.30 shows how two different imbeddings of the
same graph in the same surface can yield different dual graphs.
In Figure 1.30 the primal graph is drawn with solid dots for vertices and
solid lines for edges, whereas the dual graphs are drawn with open dots for
vertices and dashed lines for edges. Since the dual graph at the left has a
vertex of valence 7, and the maximum valence in the dual graph at the right is

-- b
- b

Figure 1.29. lmbeddings of K 5 and K3, 3 on the torus.

32 Introduction

...._ ___ _--

Figure 1.30. Two different duals of the same graph in the sphere.

r-r----- - - - - - - .....71
1 I ,..... I
I I ,........... I
I I ,........... I
I I ,........... I
.....
I
I
I
--------1
_______ j
Figure 1.31. A graph in the torus and its dual.

6, the two duals are not isomorphic. It may be observed that for any region f
of the primal imbedding, the valence of the dual vertex f* equals the number
of sides of f.
Figure 1.31 shows the duality construction on a torus. The primal graph has
two vertices and three edges between them, and there is only one primal
region. The dual graph has only one vertex, at which three loops are based.
It should be noted that if one were to construct the dual of the dual
imbedding, then the primal imbedding of the primal graph would be restored.
This fact is what justifies the use of the word "dual".

1.4.9. Exercises
1. Cut out a rectangular strip of paper about 3 em wide and about 30 em
long. On both sides, draw a lengthwise center line. Next, paste the strip
into a Mobius band. Then cut the strip along the center line. How many
pieces are obtained? Next, cut each of the bands obtained in the previous
step along their center lines. What happens?
2. Draw K 5 in the sphere so that there is only one edge-crossing. Do the
same for K 3 , 3 •
3. Draw an imbedding of K 6 on the torus.
4. Draw K 6 on the sphere with only three edge-crossings.
More Graph-Theoretic Background 33

5. Prove that no simplicial graph with seven vertices and 16 edges can be
imbedded in the sphere.
6. Suppose that a simplicial graph has p vertices. What is the maximum
number of edges it can have and still have an imbedding in the sphere?
7. Suppose that a simplicial bipartite graph has p vertices. What is the
maximum number of edges it can have and still have an imbedding in the
sphere?
8. Prove that there are exactly five different simplicial graphs with six
vertices and 12 edges. Decide which ones can be imbedded in the sphere.
(Give proofs, of course.)
9. Draw three cellular imbeddings of different graphs in the torus, and then
calculate the Euler formula in each case. What number is its value? This
number is the Euler characteristic of the torus.
10. Find a subgraph of the Petersen graph (recall Figure 1.20) that is a
subdivision of K 3 3 , thereby proving that the Petersen graph is non-
spherical. '
11. Using the girth of the Petersen graph, prove it is nonspherical.
12. Draw the dual for the imbedding of Figure 1.25.
13. Draw 2-cell imbeddings for K 5 - e (i.e., the result of removing one edge
from K 5 ) in the sphere and in the torus. Then draw their duals.
14. Show that the 3-cube graph Q 3 is isomorphic to a dual of the 3-octahedron
graph 0 3 in the sphere.
15. Prove that there are exactly four isomorphism types of connected graphs
with three vertices and three edges, and that exactly two of them have
triangular imbeddings in the sphere.
16. Let G be a graph. The "line graph" L(G) has vertex set Ec, the edges of
G. Two members of Ec are adjacent in Lc if and only if they have a
common endpoint in G. Give an example of a graph G such that

y(L( G))> y( G)
Also, prove that if the maximum valence of a graph G is less than or
equal to 3, then
y(L(G)) 5; y(G)

17. Prove that a nonconnected graph cannot have a 2-cell imbedding in a

connected surface.

1.5. MORE GRAPH-THEORETIC BACKGROUND

This section gives some explicit attention to a few of the standard topics in
graph theory that arise most frequently in topological developments.
34 Introduction

Figure 1.32. The seven bridges of Konigsberg.

1.5.1. Traversability
In the earliest known paper on graph theory, Euler (1736) proved that it was
impossible to cross each of the seven bridges of Konigsberg once and only
once on a walk through the town. Figure 1.32 shows the original Konigsberg
problem, with two land areas on the opposite sides of the Pregel River and two
islands in the river, and also its graph-theoretic abstraction, in which the four
land areas are represented by vertices and the seven bridges by edges.
From the Konigsberg bridge problem arises the definition that a walk in a
connected graph is "eulerian" if it traverses each edge exactly once. A closed
eulerian walk is called an "eulerian circuit". A graph containing an eulerian
circuit is called an "eulerian graph".

Theorem 1.5.1. A connected graph G is eulerian if and only if every vertex has
even valence.

Proof First suppose that the closed walk W is an eulerian circuit in G that
starts and stops at the vertex v. Then each occurrence in W of a vertex of G,
other than v, contributes an addend of 2 to the valence of that vertex. It
follows that the valence of any vertex of G except v equals twice the number
of times it occurs on W. The valence of v is 2 for the beginning and end of W
plus twice the number of times that v occurs elsewhere in W.
Conversely, suppose that every vertex of the graph G has even valence. By
way of induction, assume that G is the smallest such graph that has no
eulerian circuit. The graph G is not K 1 , which is eulerian. Moreover, G has no
vertices of valence 1. Thus, from Theorem 1.2.3, the graph G is not a tree; so it
must contain a cycle C. Since every vertex of the graph G - C obtained by
removing the edges of the cycle C from the graph G has even valence, and
since each component of G - C is smaller than G, it follows that each
component of G - C has an eulerian circuit.
Let G1 , •.• , Gn be the components of G- C, indexed in the order in which
they are encountered on a traversal of the cycle C. An eulerian circuit for the
graph G is constructed as follows. Start traversing the cycle C. As soon as the
More Graph-Theoretic Background 35

Figure 1.33. The dodecahedron graph.

component G1 is encountered, interrupt the traversal of C to do a complete

traversal of the eulerian circuit for G1 . Then resume the traversal of the cycle
C until G2 is encountered. Once again interrupt the traversal of C, this time to
traverse completely the eulerian circuit for G2 , and so on. D

Theorem 1.5.1 does not quite solve the Konigsberg bridge problem, as
originally stated. Making the necessary modifications to solve that problem is
left as an exercise at the end of this section.
A more difficult problem than finding an eulerian circuit has grown out of a
puzzle invented by Sir William Hamilton (see Biggs et al., 1976). The
"dodecahedron" is the unique 12-sided regular three-dimensional polyhedron.
Each of its sides is a pentagon. The !-skeleton of this polyhedron, illustrated
in Figure 1.33, is called the "dodecahedron graph". From our perspective,
Hamilton's puzzle was to find a cycle containing every vertex of the dodeca-
hedron graph.
In general, a cycle that contains every vertex of a graph is called a
"hamiltonian cycle". A graph that has a hamiltonian cycle is called a "ham-
iltonian graph". Whereas there are very fast ways to construct an eulerian
circuit in a given graph or to determine that none exists, it is usually much
more difficult to construct a hamiltonian cycle or to determine that none exists
(see Karp, 1972).

1.5.2. Factors
The edge set of a spanning subgraph for a graph G is called an "n-factor" for
G if that spanning subgraph is regular of valence n. If the edges of a graph G
can be partitioned into n-factors for G, then G is called "n-factorable". For
instance, Figure 1.34 shows a !-factorization of the complete graph K 4 and a
2-factorization of the complete graph K 5 •
In general, the edges of a 2-factor form a family of mutually disjoint cycles
that includes every vertex. Petersen (1891) observed that every 4-regular graph
is 2-factorable, as demonstrated by the following argument. Since every vertex
36 Introduction

Figure 1.34. A !-factorization of K 4 and a 2-factorization of K 5 .

has even valence, a 4-regular graph G has an eulerian circuit W. As one

traverses the circuit W, one may color the edges along it alternately red and
blue. Except for the starting point of the circuit, it is obvious that each vertex
will thus obtain red valence 2 and blue valence 2. By Euler's theorem on
valence sum (Theorem 1.1.1), the number of edges of G is even, which implies
that the final color assigned along the eulerian circuit is different from the
starting color, so that the starting vertex also has red valence 2 and blue
valence 2. Thus the red edges form a 2-factor, as do the blue edges, thereby
yielding a 2-factorization. By reiteration of this edge-coloring argument, one
may prove that every 2n-regular graph is 2-factorable. Exercises 18 and 19
outline Petersen's method to extend this result to all regular graphs of even
valence.

lbeorem 1.5.2 (Petersen, 1891). Every regular graph of even valence is 2-

factorable.

1.5.3. Distance, Neighborhoods

The "length" of a walk is defined to be the number of edges it contains (when
considered as a sequence). Thus, the length of the path Pn is n - 1, and the
length of the n-cycle en is n.
The "distance" between two vertices u and v of a graph G is defined to be
the length of the shortest path between them. If u and v are in separate
components, then the distance is infinite.
A "neighbor" of a vertex v is any vertex at distance 1 from v. The
"neighborhood" of v is the induced subgraph on the neighbors of v. For
instance, in the complete graph K n, every vertex has a neighborhood isomor-
phic to K n_1 • In the n-cycle en, if n ~ 4, then every vertex has a neighbor-
hood isomorphic to K 2c. In a regular graph, every verte~ has the same number
of neighbors, but unlike the case in a vertex-transitive graph, the neighbor-
hoods in a regular graph need not be isomorphic.

Example 1.5.1. Let G be the complement in K 7 of a 2-factor consisting of a

3-cycle and a 4-cycle. The neighborhood of a vertex on the missing 3-cycle is
isomorphic to the disjoint union of two copies of K 2 , and the neighborhood of a
vertex on the missing 4-cycle is isomorphic to K 1, 3 •
More Graph-Theoretic Background 37

1.5.4. Graph Colorings and Map Colorings

An "n-coloring" for a graph G is a map from the vertices of G to a set of n
distinct objects called "colors", such that no two adjacent vertices are assigned
the same color. (Graphs with loops are noncolorable.) The "chromatic num-
ber" chr( G) is defined to be the smallest number n such that G has an
n-coloring. A graph G with chromatic number n is called "n-chromatic".
Obviously, a graph is 1-colorable if and only if it has no edges. Also, a
graph is 2-colorable if and only if it is bipartite. The simplest example of a
graph that requires at least three colors is a cycle of odd length. As indicated
by the following theorem, this example is the fundamental obstruction to
2-colorings.

Theorem 1.5.3. A graph is 2-colorable if and only if it has no odd-length cycles.

Proof In any partition of the vertices of an odd-length cycle into two

classes, at least one class must contain two adjacent vertices. Thus, an
odd-length cycle requires more than two colors, and consequently, no graph
containing an odd-length cycle is 2-colorable.
Conversely, let G be a graph with no odd-length cycles. Since the compo-
nents of a graph may be colored independently of each other, it suffices to
assume that G is connected. Select any vertex v1 of G, and let V1 be the
subset of Vc consisting of all vertices at even distance from v1 . Let V2 = Vc -
V1 . If two vertices v and v' in V1 are adjacent, then let P and P' be respective
minimum length paths to them from v1 . Let u be the last vertex of Gin which
P and P' meet. Then the segment of P from u to v plus the segment of P'
from u to v' plus the edge between v and v' forms an odd-length cycle, a
contradiction. Similarly, no two vertices of V2 are adjacent. Thus, the graph G
is 2-colorable, with color classes V1 and V2 • D

Example 1.5.2. A suspension of a cycle from K 1 is called a "wheel". For any

graph G, the chromatic number chr(K 1 + G) of the suspension of G from K 1 is
chr( G) + 1. Thus, the chromatic number of a wheel formed from an odd-length
cycle is four.

Let i: G-+ S be a graph imbedding. A "map coloring" for the imbedding i

is an assignment of colors to the regions such that no two regions with a
common edge have the same color. Thus, a map coloring for the imbedding i
is equivalent to a graph coloring for the dual graph G *;.
The "chromatic number chr(S) of the surface S" is defined to be the
minimum number of colors needed to color any map corresponding to a 2-cell
imbedding of a graph into S. Equivalently, it is defined to be the maximum of
the set of chromatic numbers of graphs that can be imbedded in S. (This
equivalence is realized through the duality construction.) Since multiple edges
between two vertices can be collapsed onto a single edge without changing the
38 Introduction

chromatic number of a graph, one need consider only the chromatic numbers
of simplicial graphs imbeddable in S. By the Ringel-Youngs (1968) solution
to the Heawood map-coloring problem,

chr () l
sg =
7 + f1 + 48g
2
J for every genus g ~ 1

One may observe that for g = 0, the value 3 of the right-hand side of this
equation is 4. The equation chr( S0 ) = 4 was verified separately by Appel and
Haken (1976), who thereby solved the four-color problem. Map-coloring
problems are discussed at length in Chapter 5.

1.5.5. Edge Operations

If H is either a subgraph of G or a subset of edges of G, then the notation
G - H means the graph obtained by deleting from G all the edges of H. The
graph G- His called the "complement of H in the graph G". One may recall
that if G is a complete graph, one also uses the notation He.
Two subgraphs H 1 and H 2 of a graph G are called "equivalent" if there is
an automorphism of G that carries H 1 onto H 2 • If the vertices and edges of a
graph H do not actually lie in G, but if all isomorphic copies of H in G are
equivalent, then the notation G- His used to denote the complementinG of
any one of them.

Example 1.5.3. The result of deleting an edge from K 5 is often denoted

Ks- K2.

Example 1.5.4. The graph of Example 1.5.1 might be denoted K 7 - (C3 U C4 ).

To any graph G, one may associate a maximal simplicial spanning subgraph

csimp, obtained by deleting all the loops of G and by removing all but one
edge of each multiple adjacency. It is easy enough to extend any imbedding
csimp --+ S to an imbedding G --+ S, by drawing each loop close to the vertex
at which it is based, in any region incident on that vertex, and by drawing the
removed edges from each multiedge of G close alongside the one that
remained in csimp. Figure 1.35 illustrates such an extension of imbeddings.
It follows from this principle for extending imbeddings of csimp to imbed-
dings of G that y(Gsimp) = y(G). However, as explained in Chapter 4,
sometimes one uses nonminimum imbeddings of a nonsimplicial graph G to
obtain a minimum (or other) imbedding of a (possibly simplicial) covering
3 The self-explanatory Iverson notations [xj ("floor" of x) and [ x] ("ceiling" of x) are used in
preference to the arbitrary notations [ x] and { x}, respectively, to indicate the greatest integer less
than or equal to the number x and the least integer greater than or equal to x.
More Graph-Theoretic Background 39

Figure 1.35. The extension of an imbedding of Gsimp to an imbedding of G.

Figure 1.36. The contraction of K 4 along the darkened edge.

space of G. Thus, this imbedding extension principle does not eliminate

nonsimplicial graphs from imbedding theorems.
A "contraction" of a graph G along the edge e is the result of deleting the
edge e from G and then identifying its endpoints. We image that the edge e is
topologically shrunk ("contracted") to a point, and we denote the result by
G1e. Figure 1.36 illustrates the contraction of K 4 along an edge. The word
"contraction" refers to the operation as well as to the resulting graph.
Contraction of an edge can introduce multiple adjacencies. In the context of
simplicial graphs, extra adjacencies are often discarded; that is, ( G1e )simp is
sometimes considered, in an abuse of terminology, to be the "contracted
graph".
If a graph G is imbedded in a surface, then the contraction of G along any
edge may be realized in the surface. Accordingly, the genus of a graph is at
least as large as the genus of any graph obtained by a sequence of contractions
applied to it.

Example 1.5.5. Since Petersen's graph can be contracted to K 5 , by a sequence

of contractions along the edges of the !-factor that runs between the inner star and
the outer pentagon, its genus must be at least 1, that is, at least y(K 5 ).

Warning 1.5.1. One of the standard mistakes in topological graph theory is to

proceed as if the inverse of a contraction does not increase genus. It is easy to
draw examples in which the obvious inverse contractions do not increase genus,
and to overlook temporarily the fact that every connected graph contracts to a
bouquet of circles, or, in the case of simplicial contraction, to K 1 .
40 Introduction

Hadwiger (1943) formulated a conjecture to help characterize n-chromatic

graphs. It is known to be true for n .:::;; 5 and for nearly all graphs. (See
Bollobas and Catlin, 1980.)

Hadwiger Conjecture. Every connected n-chromatic graph G contracts to K n,

or to a copy of Kn with some multiple adjacencies.

1.5.6. Algorithms
There are numerous known algorithms to decide whether a graph is spherical,
and there are several to calculate the genus. A naive sphericity algorithm is
obtained from Kuratowski's theorem, with the aid of a definition that permits
modular algorithm structure.
Let u be a vertex of valence 2 in a graph G, and let e 1 and e 2 be the edges
incident on u, with V(e 1 ) = {u, v} and V(e 2 ) = {u, w}. The graph obtained
by "smoothing at u" has for its vertex set VG - { u} and for its edge set
EG- { e 1 , e 2 } plus a new edge between v and w. Subdividing this new edge
would invert the operation of smoothing at u.

Naive Sphericity Algorithm. Let G be a graph. For each subset of EG whose

removal/eaves only one nontrivial component H, construct the graph Hsmooth by
successively smoothing over every vertex of valence 2 in H, except possibly for a
last vertex if His a cycle. If some Hsmooth is isomorphic to K 5 or to K 3 3 then the
graph is nonspherical, but otherwise, by Kuratowski 's theorem, G is' spherical.

Since there are 2#£ subsets of edges to consider, an execution of this naive
algorithm could be quite lengthy. In Section 1.6, we derive a polynomial-time
algorithm for sphericity. A general (exponential-time) algorithm for genus is
given in Chapter 3.

1.5.7. Connectivity
A "cutpoint" of a graph G is a vertex whose removal would increase the
number of components, where it is understood that removing a vertex from G
means also removing every edge incident on that vertex, so that the result of
the operation is a graph. A maximal subgraph for G with no cutpoints is
called a "block".
The "connectivity" (G) of a graph G is defined to be the minimum number
of points whose removal results in a disconnected graph, a trivial graph, or a
bouquet of circles. It may be observed that G has the same connectivity as
csimp.

A graph whose connectivity is k or greater is said to be "k-connected". For

instance, by a 3-connected graph, we mean graphs of connectivity 3, 4, 5, and
so on.
More Graph-Theoretic Background 41

1.5.8. Exercises
1. Prove that a connected graph G has an open eulerian walk if and only if
exactly two of the vertices have valence 1.
2. Use Theorem 1.5.1 and Exercise 1 to solve the Konigsberg bridge
problem.
3. Draw a hamiltonian circuit in the dodecahedron graph.
4. How many vertices and edges does the dodecahedron graph have?
Restriction on method: Do not count the vertices or the edges. Using the
fact that a dodecahedron is a 12-sided polyhedron, each side a pentagon,
it is possible to calculate #V and #E.
5. The "icosahedron" is the unique 20-sided regular three-dimensional
polyhedron. Each side is a triangle. The "icosahedron graph" is its
!-skeleton. It is known that the icosahedron is dual to the dodecahedron.
Draw the icosahedron graph as the dual to a spherical imbedding of the
dodecahedron graph. How many vertices and edges does the icosahedron
graph have? Restriction on method: Do not count the vertices or the
edges of the icosahedron graph.
6. Prove that the Petersen graph is not !-factorable.
7. Find a 2-factorization of the complete graph K 9 •
8. Let G and H be vertex-transitive graphs such that the neighborhood of
any vertex of G is isomorphic to the neighborhood of any vertex in H.
Are G and H necessarily isomorphic graphs? Explain.
9. Prove that any imbedding of a wheel in the sphere has for its dual graph
an isomorphic wheel.
10. What is the chromatic number of Petersen's graph?
11. Draw a map on the Klein bottle that requires six colors.
12. Draw a map on the torus that requires seven colors.
13. The Haj6s conjecture is that every n-chromatic graph contains a subdivi-
sion of Kn. Prove the Haj6s conjecture for n = 1, 2, and 3. (The Haj6s
conjecture is false for nearly all graphs. See Catlin (1979) and Erdos and
Fajtlowicz (1981).
14. Prove the Haj6s conjecture for n = 4.
15. Prove that a contraction of a tree is a tree.
16. Prove the Hadwiger conjecture for n = 2.
17. Give an example to show that a contraction can increase the chromatic
number of a graph.
18. Let e 1 and e 2 be proper edges with a common endpoint w in a graph G,
let u be the other endpoint of e 1 , and let v be the other endpoint of e 2 •
Let G' denote the graph obtained from G by deleting edges e 1 and e 2
and then adding a loop at the vertex w and a new edge between the
42 Introduction

vertices u and v (i.e., another new loop if u = v). Use Petersen's theorem
for 4-regular graphs to prove that if G' is 2-factorable, then so is G.
19. Using Exercise 18 and an induction on the number of proper edges,
prove Theorem 1.5.2.

1.6. PLANARITY

The study of planar imbeddings has a long and rich history that intertwines
with chromatic graph theory, algorithmic analysis, enumeration, and much
else. However, unlike higher-genus surfaces, the plane and sphere are homo-
logically trivial, which is to say, overly simple in a topological sense. Accord-
ingly, the methods used to study planar imbeddings are less topological than
the ones of primary interest to us here. Nonetheless, Kuratowski's theorem
might well be the most famous theorem in all of graph theory, and it might be
remiss to offer no proof of it.
Of the many published proofs, we have selected one whose organization is
due to Thomassen (1980). It is not the shortest, and it is not the easiest.
However, it is illuminating, in that it also yields other results about planar
imbeddings. One is Fary's theorem that the edges of a planar imbedding of a
simplicial graph can be chosen to be straight lines. Another is Tutte's theorem
that the regions of a planar, 3-connected graph imbedding can be chosen to be
convex. Still another is Whitney's theorem that a 3-connected simplicial planar
graph has only one planar imbedding. Moreover, Thomassen's proof is readily
converted into a planarity algorithm that is polynomial in the number of
vertices, a great improvement over the exponential Naive Algorithm given in
Section 1.5.

1.6.1. A Nearly Complete Sketch of the Proof

Thomassen's proof of Kuratowski's theorem depends on the use of simplicial
edge contractions. As noted in Warning 1.5.1, edge contraction can be a
dangerous tool, since a single edge contraction can lower the genus of a graph
dramatically. On the other hand, for planar graphs, where the genus is already
as low as it can go, edge contraction is safe and natural, because an edge
contraction of a planar graph yields a planar graph.
We have already proved in Section 1.4 that a graph containing a homeo-
morph of either the complete graph K 5 or the complete bipartite graph K 3 3
cannot be planar. Therefore, Kuratowski's theorem is proved if we establi~h
for every graph G that the following statement is true:

( KT) If the graph G contains no homeomorph of K 5 or of K 3 • 3 ,

then G is planar.

The proof is by induction on the number of vertices in the graph G.

Planarity 43

Basis step. The statement (KT) is vacuously true of all graphs with four or
fewer vertices.

Induction Hypothesis. We assume that the statement (KT) is true of all graphs
with fewer than n vertices, where n is any number greater than or equal to 5.

Induction Step. Let G be a graph with n vertices, and let G contain no

homeomorph of K 5 or K 3, 3 • We shall prove that G is planar.

Let e be any edge in the graph G, and suppose that its endpoints are the
vertices u and v. Let G' be the graph obtained from G by a simplicial
contraction of edge e, and let us give the name v' to the vertex to which e is
contracted. We momentarily interrupt the proof of Kuratowski's theorem to
establish a lemma.

Lemma. Let G be a graph that contains no homeomorph of K 5 or K 3 3 , and let

e be any edge in G. Then the result of simplicially contracting the g~aph G on
edge e is a graph G' that contains no homeomorph of K 5 or of K 3 3 •

Proof of Lemma. If we already knew that Kuratowski's theorem is true, we

would imbed G in the plane, then simplicially contract e within the imbed-
ding, and thereby obtain an imbedding of the contracted graph G'. However,
our task is to prove this lemma without using Kuratowski's theorem.
Suppose that the contracted graph G' contained a homeomorph of K 5 or of
K 3 , 3 • Then G would contain a subgraph H such that the contraction of a
single edge e of H produces a homeomorph of K 5 or of K 3 3 • Using reversal
of contraction, we see that there are only three essentially different possibilities
for H, up to homeomorphism; they are illustrated in Figure 1.37. In each case,
the subgraph H itself contains a homeomorph of K 3 3 or of K 5 , contradicting
the hypothesis for the original graph G. On the left, the subgraph His already
homeomorphic to K 3, 3 , and in the center H is homeomorphic to K 5 . On the
right, simply delete the two horizontal edges to obtain a homeomorph of K 3 3 ,
in which the vertices v, x, y form one half of the bipartition. D '

X y
Figure 1.37. The subgraph H obtained by reversing the contraction of edge e.
44 Introduction

Figure 1.38. Splitting vertex v' back into vertices u and v.

Having established the lemma, we resume the proof of Kuratowski's

theorem, now certain that the graph G' contains no homeomorph of K 5 or of
K 3, 3 • By the induction hypothesis, we know that there is a planar imbedding
of G'.
A key step in Thomassen's approach is assertion and proof that there is a
planar imbedding for the contracted graph G' that, near the vertex v', looks
like a cycle C with "spokes" to the vertex v'. (See the wheel on the left of Fig.
1.38.) We should like to split the vertex v' back into the original vertices u and
v and attach them to the appropriate vertices on the cycle C, all the while
maintaining a planar imbedding. (See the right of Fig. 1.38.)
We postpone the proof of this key assertion until our completion of the
sketch. For the time being, assume that the key assertion has been proved.
There are three types of vertices on the cycle C: those adjacent to vertex u
in the original graph G, those adjacent to vertex v, and those adjacent to both
u and v. As Figure 1.38 indicates, we should like there to be at most two
vertices adjacent to both u and v. If there were three such vertices, however,
then G would contain the homeomorph of K 5 shown on the left of Figure 1.39
(note the five vertices of valence 4). We should also like all the vertices
adjacent to u to occur contiguously around the cycle C, and the same for all
the vertices adjacent to v. If strings of u-neighbors and v-neighbors alternated,
however, then the graph G would contain the homeomorph of K 3 , 3 shown on
the right of Figure 1.39, in which vertices s, t, u form one half of the
bipartition and vertices v, x, y form the other half. Since by hypothesis, the

Figure 1.39. Obstacles at the vertex v'.

Planarity 45

graph G contains no homeomorph of K 5 or of K 3 3 , the desired planar

imbedding for G can be constructed as in Figure 1.38. This concludes our
sketch of the proof of Kuratowski's theorem. The proof will be entirely
complete once we have verified the key assertion.
This proof is particularly insightful because the roles played by K 5 and
K 3, 3 are so clear. The two Kuratowski graphs are exactly the two obstructions
to inverting an edge contraction.
The key assertion is a sticky point. Just how do we know that the
imbedding of the contracted graph G' near v' looks like Figure 1.38? In order
for this to be true, it would have to be the case that when the vertex v' and its
incident edges are deleted from the imbedding of G', the resulting region has a
cycle in G' as its boundary. Unfortunately, if the graph obtained from G' by
deleting the vertex v' is not 2-connected, then the resulting region might very
well not have a simple cycle as its boundary, in which case the whole proof
collapses. Thus, to preserve the sketch, we shall need G' and hence G to be at
least 3-connected. As a result, the !-connected and 2-connected cases must be
dealt with separately, and, what is more important, a delicate question about
3-connectivity and edge contraction must be considered. In what follows, we
first investigate the relationship between connectivity, region boundaries, and
edge contractions. Next we prove Kuratowski's theorem for 3-connected
graphs. The !-connected and 2-connected cases are treated last.

1.6.2. Connectivity and Region Boundaries

When picturing the regions of a planar imbedding, one tends to imagine
"round" (i.e., convex) polygons whose boundaries are cycles of the imbedded
graph, without any repeated vertices. Of course, the unbounded region is not
convex, but one still pictures the boundary of this region as a simple cycle.
This picture of region boundaries as simple cycles is not wholly accurate. For
planar imbeddings, the problem is easily pinpointed by the following theorem.

Theorem 1.6.1. Every region of a planar imbedding of a graph G has a simple

cycle for its boundary if and only if G is 2-connected.

Proof Suppose that the boundary of some region r is not a cycle, but
some other kind of closed walk, so that the vertex v occurs twice, as on the left
of Figure 1.40. Then there is a simple closed path in the plane that leaves from
vertex v between two edges of the boundary of r, stays the whole time within
the region r, and later comes back to vertex v between a different pair of
edges, as shown on the right-hand side of Figure 1.40. This closed path
separates the plane into two pieces, both of which contain parts of the graph
G. Since the path intersects the graph G only at vertex v, it follows that v is a
cutpoint of the graph G. Therefore, G is not 2-connected.
Conversely, suppose that G has a cutpoint. Then G may be viewed as the
amalgamation of two graphs H and K at the vertex v. In any imbedding of G,
46 Introduction

r
Figure 1.40. The region r (shaded) and vertex v.

on a traversal of a small circle centered around the vertex v, one must

encounter somewhere an edge d of H (with endpoint v) followed immediately
by an edge e of K (with endpoint v ). Consider the region r with the corner at
v between edges d and e. Since the edge d lies in subgraph H and since v is a
cutpoint, no edges of subgraph K are encountered until the boundary returns
to v. Since the edge e from K must occur somewhere on that boundary
traversal, and since that closed walk must return to vertex v, we conclude that
the boundary of region r is not a simple cycle. D

The situation for higher-genus surfaces is more complicated. Half of

Theorem 1.6.1 generalizes; that is, if G has a cutpoint, then every imbedding
of G in any surface has a region whose boundary is not a cycle, since the proof
given above applies verbatim. On the other hand, it is easy to give an
imbedding of a 2-connected graph in the torus having a region whose
boundary is not a cycle (Exercise 1.6.1).
Surprisingly enough, it is not even known whether every 2-connected graph
has an imbedding in some surface such that every face boundary is a cycle.
The "double cover" conjecture that such an imbedding always exists may have
had its origins in a paper by Tutte (1949) related to !-factorizations of
3-regular graphs. It is mentioned at the end of a paper of Haggard (1980).
Archdeacon (1984) proves the conjecture true for 4-connected graphs. One
might guess that a minimal genus imbedding of a 2-connected graph should
have every face boundary a cycle, but it is not hard to give a counterexample.

1.6.3. Edge Contraction and Connectivity

Suppose that the graph G' is obtained from the graph G by contracting the
edge e with endpoints u and v to the single vertex v', as in the sketch of the
proof of Kuratowski's theorem. If G is k-connected, then G' is at least
(k - I)-connected, since any set of m vertices disconnecting G' can be turned
into a set of at most m + 1 vertices disconnecting G by splitting the vertex v'
back into the vertices u and v. On the other hand, it is possible that G' is not
k-connected. In fact, if the identified vertices u and v are in a set of k vertices
that disconnects G, then the same set of vertices, with u and v replaced by v',
forms a set of k - 1 vertices that disconnects G'. Figure 1.41 illustrates an
Planarity 47
u

v v'
Figure 1.41. An edge contraction of a 3-connected graph to a 2-connected graph.

example in which the original graph G is 3-connected and the contracted

graph is 2-connected.
In order for Thomassen's proof to work for 3-connected graphs, it is
sufficient that the graph G' obtained by contracting an edge e of the original
graph G be still 3-connected. The previous discussion shows that this might
not be the case if the edge e were chosen at random. Thus the following
theorem is an essential ingredient in Thomassen's proof.

Theorem 1.6.2. Let G be a 3-connected graph with five or more vertices. Then
there is some edge e of G such that the graph G/ e obtained by contracting e is
also 3-connected.

Proof (Thomassen). Suppose, by way of contradiction that, for every edge

e, the contracted graph Gje has a set of two vertices that disconnects it. One
of those two vertices must be the vertex obtained by identifying the two
endpoints of the edge e; otherwise, the same set of two vertices would also
disconnect G, thereby contradicting the 3-connectivity of G. Thus for every
edge e, the endpoints u and v of e together with some third vertex w
disconnect G. Accordingly, let us choose an edge e and a vertex w such that
the largest component H of the graph G - { u, v, w} is the largest possible, for
any disconnecting set consisting of three vertices, two of which are adjacent.
Let x be a vertex adjacent to w such that x lies in a component of
G- { u, v, w} other than the maximum component H. Since vertices w and x
are the endpoints of an edge of G, it follows that G has a disconnecting set of
the form {w, x, y}.
We claim that some component of G - { w, x, y} is larger than H, a
contradiction. To see this, let H' be the subgraph of G induced by the vertices
of H together with u and v. Since both u and v are adjacent to vertices of H
(otherwise G would not be 3-connected), the subgraph H' is connected. On
one hand, perhaps the vertex y is not in H'. Since w and x are not in H'
either, it follows that H' is contained in a component of G - { w, x, y },
contradicting the maximality of H. On the other hand, perhaps y is in H'. If
H' - y is connected, then there is again a contradiction of the maximality of
H, since H' - y has one more vertex than H. If H' - y were not connected,
48 Introduction

then one component of H' - y would contain both the vertices u and v, since
u is adjacent to v; and, therefore, all the other components of H' - y are
connected to the rest of the graph G through the vertices y and w. This would
imply that { y, w} disconnects G, contradicting the 3-connectivity of G. We
conclude that for some edge e, the contracted graph Gje is 3-connected. 0

1.6.4. Planarity lbeorems for 3-Connected Graphs

We are now ready to prove Kuratowski's theorem for 3-connected graphs.

lbeorem 1.6.3. If G is a 3-connected graph containing no homeomorph of K 5

or K 3 , 3 , then G is planar.

Proof (Thomassen). The proof is by induction on the number of vertices of

G just as in the sketch given at the beginning of this section. By Theorem
1.6.2, the contracted edge can be chosen so that the new graph G' is still
3-connected. This implies that the 3-connected part of the induction hypothe-
sis holds for G'. It also means the graph obtained from G' by deleting the
vertex v' is 2-connected. It follows from Theorem 1.6.1 that the resulting face
containing vertex v' is bounded by a cycle, and hence the rest of the original
proof does apply. The other part of the induction hypothesis, that G' contains
no homeomorph of K 5 or K 3, 3 , holds because of the lemma given early in the
sketch. Thus, Kuratowski's theorem holds for all 3-connected graphs. o

The proof of Kuratowski's theorem in the 3-connected case can be refined

so as to retain geometric properties of imbeddings. Suppose that the bounded
regions of the planar imbedding for the contracted graph G' are all convex and
that the edges are all straight line segments. Then when vertex v' is split back
into vertices u and v, if u and v are chosen close enough to v', the resulting
imbedding for G can also be made to have convex regions and straight-line
edges. Thus, with this minor adjustment, Thomassen's proof of Kuratowski's
theorem also yields inductive proofs of the following theorems, in which the
basis step for the induction is provided by the obvious planar imbedding of
K4.

lbeorem 1.6.4 (Tutte, 1960). Any planar 3-connected graph has a planar
imbedding such that every bounded region is convex.

lbeorem 1.6.5 (Fary, 1948). Any planar 3-connected graph has a planar
imbedding such that every edge is a straight line segment.

It is obvious from Theorem 1.6.1 that Tutte's theorem does not hold for
!-connected planar graphs. Moreover, it does not hold even for 2-connected
planar graphs (Exercise 1.6.2). On the other hand, Fary's theorem holds for all
planar graphs, as we shall show later in this section.
Planarity 49

There is another observation worth making about the proof for Kuratowski's
theorem in the 3-connected case. When the vertex v' is split back into vertices
u and v, there really is no choice of how the resulting imbedding looks at u
and v. Thus, there is essentially only one way to imbed G in the plane. By
"essentially one way", we mean up to a homeomorphism of pairs. Further
discussion of this concept is deferred to Chapter 3. Anyone familiar both with
Thomassen's proof and that concept of uniqueness will recognize that
Thomassen's proof of Kuratowski's theorem can be adapted to prove the
following result of Whitney, by once again using an induction based on K 4 •

Theorem 1.6.6 (Whitney, 1933). There is only one way to imbed a 3-connected
planar graph in the plane.

Proof Omitted. D

1.6.5. Graphs That Are Not 3-Connected

The following theorem is a basic tool for deriving results about planar graphs
that are not 3-connected. Although it might appear intuitively obvious, the
proof does have details that need to be checked.

Theorem 1.6.7. Let H and K be planar graphs. Then the graph obtained by
amalgamating H and K either at a single vertex v or along a single edge e is
planar.

Proof If the graphs H and K are imbedded disjointly in the plane as in

Figure 1.42, so that both copies of the vertex v or of the edge e lie on the
"exterior" region, it would be easy to construct a planar imbedding of
the amalgamation. For a vertex amalgamation, one simply pulls each copy of
the vertex v out away from the rest of the graph H or K, along a path not
touching the rest of its respective graph, while "dragging" along the edges
incident on v. For an edge amalgamation, there is a preliminary step of
topologically shrinking each copy of e until it is scarcely larger than a vertex,
after which one does the same pulling away as with a vertex.

Figure 1.42. Amalgamating two planar imbeddings.

50 Introduction

To get the respective copies of v or of e to the exterior region, imagine that

H and K are first imbedded in separate copies of the 2-sphere, and select in
each copy a region that contains v or e on its boundary. Next, delete the
interior of a closed disk from both selected regions. Then paste the two
resulting surfaces together by identifying the boundaries of the removed disks,
thereby forming a new 2-sphere, in which the graphs H and K are imbedded
so that both copies of v or e lie on the boundary of the same region. (A fully
rigorous topological proof is readily formulated along this plan, with the aid of
the Jordan curve theorem.) D

Theorem 1.6.7 now leads directly to a proof of Kuratowski's theorem for all
graphs, 3-connected or not, starting from Theorem 1.6.3, the version for
3-connected graphs. However, we should like to prove Fary's theorem as well,
and the proof of Theorem 1.6.7 wreaks havoc with straight line segments.
Accordingly, we take a different tack, as follows. If a graph contains no
homeomorph of K 5 or K 3 3 , and it is not 3-connected, then we add as many
edges to the graph as possible without creating K 5 or K 3 3 • If the extra edges
make the resulting graph 3-connected, then Kuratowski's theorem and Fary's
theorem will hold for the original graph, because it is a subgraph of a
3-connected graph. The following theorem is what is needed.

Theorem 1.6.8. Let G be a graph containing no homeomorph of K 5 or K 3 3 and

such that the addition of any edge to G creates such a homeomorph. The~ G is
3-connected.

Proof. The proof is by induction on the number n of vertices of G. As a

basic step, we observe that the theorem is true for n = 5, since K 5 - e is
3-connected. Now assume that the theorem is true for all graphs with fewer
than n vertices, where n > 5, and further assume that the graph G has n
vertices.
Suppose, by way of contradiction, that G is only 2-connected, so that it is
the amalgamation of graphs Hand K at two vertices u and v. We assert that
the vertices u and v are adjacent in the graph G. Suppose not. Then the graph
H' obtained from H by adding an edge e from u to v still would not contain
a homeomorph of K 5 or K 3 , 3 , since any such homeomorph would be con-
tained in G (with a path through K playing the role of edge e). Also, the
addition of any edge to H' would create a homeograph of K 5 or K 3 3 in
the original graph G; since that homeomorph cannot be disconnected by the
removal of vertices u and v, that homeomorph must be contained in H' as
well. Thus by induction, H' is 3-connected and hence planar. Similarly, the
graph K' obtained by adding an edge e from u to v is 3-connected and
planar. The amalgamation of H' and K' along the edge e is then planar by
Theorem 1.6.7. In particular, that amalgamation contains no homeomorph of
K 5 or K 3 , 3 , by the arguments of Section 1.4. This contradicts the maximality
Planarity 51

of the original graph G, since the amalgamation of H' and K' along the edge e
is simply G with the edge e added.
We therefore assume that G is the amalgamation of graphs Hand K along
the edge e and that Hand K share the same maximality property as G with
respect to homeomorphs of K 5 and K 3, 3 • By the induction hypothesis, the
graphs H and K are 3-connected, and hence planar. By Theorem 1.6.7, the
graph G is planar. It follows that in any planar imbedding of G, H, or K,
there can be no region with four or more sides. Otherwise a "diagonal" edge
across the region could be added, creating a planar graph with one more edge,
contradicting the maximality of the given graph (there is a slight problem if
the vertices we want to join by a diagonal are already joined by an edge
elsewhere in the imbedding; see Exercises 3 and 4). Since the graphs in
question have no loops or multiple edges, every region must have three sides.
By Euler's equation it follows that #E = 3#V- 6 for the edge and vertex
sets of any of the graphs G, H, and K (see Exercise 5). But since G is the
amalgamation of Hand K along a single edge, #V(G) = #V(H) + #V(K)
- 2 and #E(G) = #E(H) + #E(K)- 1. Therefore

#E(G) = #E(H) + #E(K)- 1

= 3#V(H) + 3#V(K)- 13
= 3(#V(G) + 2) -13
-:F 3#V(G)- 6

a contradiction. We conclude that G cannot be the amalgamation of two

graphs Hand K at two vertices. A similar but easier argument applies if G is
!-connected but not 2-connected. D

Corollary. Kuratowski's theorem and Fary's theorem hold for graphs that are
not 3-connected.

1.6.6. Algorithms
The number of steps required by the Naive Sphericity (or Planarity) Algorithm
suggested in Section 1.5 is an exponential function of the number n of vertices
of a graph G. We shall see that Thomassen's design for a proof of the
Kuratowski theorem leads to an algorithm whose execution time is a poly-
nomial function of n, thereby a great improvement.
We first assume that the graph G to be tested is 3-connected. In that case,
according to Theorem 1.6.2, there exists an edge e such that contraction of G
on e yields a 3-connected graph. This step is iterated on that new 3-connected
graph, and reiterated until a 4-vertex graph is obtained.
Execution time analysis: It takes n - 4 edge contractions to convert an
n-vertex graph into a 4-vertex graph. Let us suppose that prior to each edge
contraction, we find it necessary to make an exhaustive search over all edges
52 Introduction

before we encounter one such that contraction will preserve 3-connectedness.

Since the original graph G has n vertices, its number of edges cannot exceed
(n 2 - n)/2. Indeed if G has more than 3n- 6 edges, then G is not planar (see
Exercise 5).
Perhaps, for each edge e in a graph created by the contraction sequence, we
shall need to do a preliminary contraction to see if the subsequent resulting
graph would be 3-connected. The number of steps needed to construct the
graph that would result from the contraction is, at worst, of the same order of
magnitude as the size of the graph itself. Naively, one might think of removing
every possible 2-vertex subset and checking whether the result is connected.
However, the only 2-vertex subsets that must be checked are the ones that
contain the new vertex to which the edge e was contracted. The time needed
to check whether the result of removing such a 2-vertex subset is connected is
at most linear in n, if you use a depth-first search to check connectedness.
Thus, the total number of steps needed to contract iteratively the original
graph G down to a 4-vertex graph is, at worst, of the order of magnitude of the
polynomial

In other words, the number of steps needed for the edge-contracting part of
the algorithm is at most quartic in the number of vertices.
Of course, since the complete graph K4 is the only 3-connected 4-vertex
graph, it must be the graph we ultimately obtain from the sequence of
contractions. At this point, we imbed K 4 in the plane, and iteratively reverse
each of the contractions, either until some reversal yields a copy of a
Kuratowski graph, or until we have a planar imbedding of the original graph G
in the plane.
At each reverse contraction, when a vertex v' is split back into two vertices
u and v, the cycle C of vertices adjacent to v' is checked for the following
patterns:

1. three vertices in cycle C adjacent to both u and v

2. four vertices in cycle C arranged so that when C is traversed, the ones
adjacent to u alternate with the ones adjacent to v

Patterns 1 and 2 lead to homeomorphs of K 5 and K 3 3 , respectively. If neither

pattern is encountered, then splitting v back into u imd v leads at once to a
planar imbedding of the resulting graph.
Execution-time analysis: Even the most naive algorithm to check for
patterns 1 and 2 requires at most quartic time in the number of vertices on
cycle C.
Far more efficient algorithms for planarity exist. For instance, Demoucron
et al. (1964) gave one of order n 2 • Later, Hopcroft and Tarjan (1974), in one
of the most famous early papers on computational complexity, reduced this to
Planarity 53

order n. Filotti et al. (1979) have found polynomial-time algorithms to de-

termine if a given graph can be imbedded in a surface of given genus g.
However, the degree of the polynomial is an increasing function of g, so it is
still not known at present whether there is a polynomial-time algorithm for
computing the genus of a given graph.

1.6.7. Kuratowski Graphs for Higher Genus

It is natural to ask whether there is an analog to Kuratowski's theorem for
surfaces other than the sphere. Given a closed surfaceS, let K(S) be the set of
all graphs G such that G has no 2-valent vertices and such that G is not
imbeddable in S but G - e is, for any edge e. Thus K(S) is the minimal set
of obstructions to imbedding a graph in the surface S. Kuratowski's theorem
asserts that K(S0 ) = {K5 , K 3 , 3 }. Must the obstruction set K(S) be finite for
an arbitrary, closed surface S?
Glover et al. (1979) and Archdeacon (1981) were the first to prove K(S)
finite for a surface other than the sphere, when they showed that K(N1 ) has
103 elements for the projective plane N1 • Unpublished work of Haggard,
Decker, Glover, and Huneke recorded more than 300 members of K(S 1).
Archdeacon and Huneke (1985) showed that K(N) contains only finitely
many 3-regular graphs for any nonorientable surface N. Finally, Robertson
and Seymour, in a series of papers deriving highly general results about
families of graphs, proved that K(S) is indeed finite for all surfaces. Although
the size of K(S) is in general enormous, there seems to be some structure.
Glover has conjectured that every element of K(Sg) is the union of at most
2g + 1 copies of the original Kuratowski graphs K 5 and K 3, 3 •

1.6.8. Other Planarity Criteria

There are many other planarity criteria besides Kuratowski's. We give two
more here that are due to mathematicians most famous for their work in
topology and algebra, Whitney and MacLane. Although these criteria are not
especially useful, they do involve elegant applications of duality and the cycle
structure of a graph. Tutte's proofs of the theorems of Whitney and MacLane
are given in Exercises 6-11.
Consider the dual graph G* for a planar imbedding of the graph G. Every
cycle C of the graph G separates the regions of the imbedding into two
collections: those inside the cycle and those outside the cycle. In terms of the
dual graph, the collection C* of edges dual to those in cycle C, when deleted
from the graph G*, separate G* into two disjoint parts such that every edge in
C* goes from one part to the other. In general, such a collection of edges is
called a "bond". Neither of the two parts is required to be connected. It is a
little harder to see, but still true that, conversely, every bond for G* corre-
sponds to a cycle in G. What is most surprising is that this relationship
between a graph and one of its planar duals characterizes planarity.
54 Introduction

Theorem 1.6.9 (Whitney, 1932). A 2-connected graph G is planar if and only if

there is a graph G* whose edges are in one-to-one correspondence with those of G
such that cycles in G correspond to bonds in G*.

Proof See Exercises 11 and 12. 0

MacLane's planarity condition depends on the cycle structure of a graph as

well. Given a finite set X of size n, let S(X) be the set of all formal sums
a 1x 1 + · · · +anxn, where each a; is 0 or 1. Add two formal sums by adding
corresponding coefficients of two elements. It is easiest to think of each formal
sum in S( X) as a subset of X, namely, those elements of X with coefficient 1.
If s corresponds to subset A and t to subset B, then s + t corresponds to the
symmetric difference A u B- (A n B).
Given a graph G with edge set E, the "cycle space" of G, C(G), is the
subspace of S(E) spanned by the cycles of G. Each element of C(G) is either
a cycle or a union of edge-disjoint cycles. If G is a 2-connected graph
imbedded in the plane, then every region boundary is a cycle of G. Moreover,
the outside (unbounded) region boundary cycle is the sum of all the inside
region boundary cycles (all interior edges occur in two cycles and hence cancel
them out in the sum). In fact, any cycle in the graph G is the sum of the region
boundary cycles contained in the interior of the cycle. Thus the set of inside
region boundary cycles form a basis for the cycle space C(G). Moreover, this
basis has the property that every edge of G is contained in at most two
elements of the basis. In general, call a collection B of cycles in a graph H a
"2-basis" if B is a basis for the cycle space of H and if every edge of H is
contained in at most two elements of B. The discussion so far shows that every
2-connected planar graph has a 2-basis. MacLane proved that the converse is
true as well:

Theorem 1.6.10 (MacLane, 1937). A 2-connected graph is planar if and only if

it has a 2-basis.

Proof See Exercises 6-10. 0

1.6.9. Exercises
1. Give an imbedding of a 2-connected graph in the torus such that some
region has a boundary that is not a cycle.
2. Find a planar 2-connected graph that has no planar imbedding with
every bounded region convex.
3. Prove that in any planar imbedding of the complete graph K 4 , every
region is 3-sided. Conclude that if a simple graph G has a planar
imbedding with four or more vertices all lying on the same region
boundary, then an edge can be added to G (without creating multiple
edges) and still maintain planarity.
Planarity 55

4. Use Exercise 3 to prove that if G is planar and the addition of any edge
to G creates a nonplanar graph, then every region in any planar imbed-
ding of G has size 3. (Caution: The number of distinct vertices in a
region boundary can be less than the number of sides of the region.)
5. Prove, using Euler's equation, that any planar imbedding for a connected
simplicial graph with edge set E such that #E ~ 2 and vertex set V
satisfies #E ::::;; 3#V- 6, with equality if and only if every region is
3-sided.
6. Given a graph G with vertex set V = { v1 , ••. , vn} and edge set E, define
the linear transformation d: S(E) ~ S(V) by d(e) = u + v, where u
and v are the endpoints of the edge e. Show that the cycle space C(E) is
the kernel (nullspace) of d, and that if P is a path of edges from vertex u
to vertex v, then d(P) = u + v. Show that if G is connected, then
v1 + v2 , v1 + v3 , ••• , v1 + vn are all in the range of d, but v1 is not.
Conclude that if G is connected, the range of d has dimension n - 1 and
therefore the dimension of the cycle space C(E) is #E- #V + 1.
7. Prove that K 3 3 has no 2-basis. (Hint: The dimension of C(K3 3 ) is 4, by
Exercise 6. Show that at least seven edges of K 3 3 must be each' contained
in two elements of a 2-basis for K 3, 3 and the~ consider the sum of all
elements of the 2-basis.)
8. Prove that K 5 has no 2-basis (see Exercise 7).
9. Prove that if the graph G has a 2-basis, then so does G - e where e is
any edge of G.
10. Use Exercises 7-9 and Kuratowski's theorem to prove MacLane's theo-
rem.
11. Given a 2-connected graph G with edge set E, let the bond space B(E)
be the subspace of S(E) spanned by the bonds of G. Let star(v) be the
subgraph induced by collection of all edges incident to the vertex v.
Prove that { E(star(V)) 1 v E V} is a spanning set for the bond space
B(E). [If A c E and G- A has two components H and K, consider
l: E(star(V)) where the sum is taken over all vertices of H.]
12. Use Exercise 11 and MacLane's theorem to prove Whitney's theorem.
13. Find an algorithm that determines in at most kn steps, k a constant,
whether a given n-vertex graph G with at most 3n edges is connected.
(Assume you are given for each vertex a list of the adjacent vertices.)
2
Voltage Graphs and
Covering Spaces

There is an obvious problem in trying to represent an imbedding of an

arbitrarily large graph in a surface, namely, to avoid inundating the reader in
details. If all graphs deserved equal interest, then the prospects for obtaining a
satisfactory representation would be dim. Fortunately, as a matter of nature
and esthetics, the graphs of greatest interest have symmetries that can be
exploited in developing an economical description.
One mathematical structure that permits an economical description of
graphs and their imbeddings is a covering space, which is a topological
generalization of a Riemann surface. As we demonstrate in this chapter,
complete graphs, Cayley graphs, n-cube graphs, and many other kinds of
graphs are derivable as covering spaces of graphs with only one vertex-that
is, of bouquets of circles-or are derivable by deleting a 1-factor from a
covering of such a bouquet.
The details for constructing a covering space of a graph are efficiently
encoded in what is called a "voltage graph". In an electrical network, there is
a voltage difference (perhaps zero) experienced in traversing any line. Al-
though the mathematical voltages here are elements of a finite group, and not
real numbers, the assignment of a voltage to every directed edge of a graph
provides a clear analogy to electrical networks, with an important exception:
two directed paths in parallel, that is, with the same initial and terminal
vertices, need not be assigned the same mathematical voltage.
As we have stated previously, the central concern of topological graph
theory is the placement of graphs on surfaces. Given the problem of placing a
large graph on a surface, the value of the voltage graph construction is that it
may enable one to reduce the problem to placing a quotient of the given graph
in some kind of quotient of the given surface and assigning appropriate
voltages. Since the quotient graph is a fraction of the size of the given graph,
designing a suitable placement of the quotient might be easier. Thus, one may
well wish to know which graphs have quotients (other than themselves) and
how to reconstruct a graph from one of its quotients.

56
Ordinary Voltages 57

2.1. ORDINARY VOLTAGES

Let G be a graph whose edges have all been given plus and minus directions,
and let .5!1 be a group, assumed to be finite unless explicitly stated to be
infinite. A set function a from the plus-directed edges of G into the group .5!1
is called an "ordinary voltage assignment" on G, and the pair (G, a) is called
an "ordinary voltage graph". The values of a are called "voltages", and .5!1 is
called the "voltage group". The ordinary voltage-graph construction was first
suggested by Gross (1974) and immediately improved by Gross and Tucker
(1974). Its advantage over various formalistic "covering graph" constructions,
all essentially equivalent, is largely its visual suggestiveness.

2.1.1. Drawings of Voltage Graphs

Voltage graphs are usually given by pictures, rather than by combinatorial
descriptions. For instance, suppose the graph G has vertex set V = { u, v} and
edge set E = {d, e }, with d a u-based loop and and e a proper edge from u
to v. Suppose that voltages a( d) = 1 and a( e) = 0 are assigned in the cyclic
group fr3 • Then the voltage graph ( G, a) is presented as shown at the left of
Figure 2.1.
The purpose of assigning voltages to the graph G, called the "base graph"
or simply the "base", is to obtain an object called the "(right) derived graph",
denoted Ga, which is illustrated at the right of Figure 2.1. The vertex set of Ga
is the cartesian product VG X .5!1, and the edge set of Ga is the cartesian
product EG X .5!1. Although the pair notations (v, a) and (e, a) are sometimes
used for the vertices and edges of G, more frequently the subscripted notations
va and ea are used instead. If the directed edge e+ of the base graph G runs
from the vertex u to the vertex v, and if v is the voltage assigned to e +, then

uo d2
u2

·~ (a)
el

ul

(b)
Figure 2.1. (a) A graph with voltages assigned m the cychc group ~3 ((G, a)) and (b) the
as;ociated derived graph (Gn).
58 Voltage Graphs and Covering Spaces

the directed edge ea + of the derived graph Go. runs from the vertex ua to the
vertex vab·

Remark. In the definition of the derived graph, note that whereas ea + runs
from ua to vab' the reverse edge ea- runs from vab to ua. In particular, the
subscript on the minus-directed edge e a- is the product of the subscript on its
initial vertex vab and the group element b- 1 that is inverse of the voltage on e+.
Thus, whereas the subscript of a plus-directed edge of the derived graph always
agrees with the subscript on its initial vertex, the subscript on a minus-directed
edge always agrees with the subscript on its terminal vertex.

Example 2.1.1. Since the base graph in Figure 2.1 has two vertices, u and v,
and two edges, d and e, and since the voltage group fl'3 has three elements, 0, 1,
and 2, it follows that the derived graph has six vertices, u0 , u1 , u2 , v0 , v1 , and
v2 , andsixedges, d 0 , di> d 2 , e0 , e 1 , ande 2 • Sincetheu-basedloopd+ ofthe
base graph G is assigned voltage 1, it follows that for i = 0, 1, 2, the edge d; + of
the derived graph G runs from the vertex u; to the vertex u; + 1 • Since the edge e +
is assigned the voltage 0, it follows that fori = 0, 1, 2, the edge e; + of the derived
graph runs from the vertex u; to the vertex V;.

Example 2.1.2. Figure 2.2 shows how the Petersen graph might be derived by
an assignment of voltages in the cyclic group fl'5 to a particular base graph with
two vertices and three edges, which is sometimes called a "dumbbell" graph. As
indicated, the directions on the edges of the derived graph are induced by the
directions of the corresponding edges in the base graph.

1 c
0--7-D o
d
u
e 2

(a)

c2

Figure 2.2. A derivation (b) of the Petersen graph from a voltage assignment in .2'5 to a
dumbbell graph (a).
Ordinary Voltages 59

Since the dumbbell graph in Figure 2.2 has two vertices, u and v, and since
the voltage group is f!E5 , the vertices of the derived graph are u 0 , u1 , u 2 , u 3 ,
v4 , v0 , v1 , v2 , v3 , and v4 • Since the dumbbell graph has edges c, d, and e, the
derived graph has edges c0 , c1 , c 2 , c3 , c4 , d 0 , d1 , d 2 , d 3 , d 4 , e 0 , e 1 , e 2 , e 3 ,
and e 4 • The u-based loop c+ has voltage 1 in the base graph. Thus, for
i E f!E5 , the edge c/ runs from the vertex u; to the vertex u; +1 , as shown on
the outer pentagon of the derived Petersen graph. The v-based loop e+ in the
base graph has voltage 2. Thus, for i E f!E5 , the edge e; + runs from the vertex
v; to the vertex v; + 2 , as shown on the inner star. The proper edge d+ has
initial vertex u and terminal vertex v in the dumbbell and is assigned voltage
0. Thus, for i E f!E5 , the edge d; + in the derived graph runs from the vertex u,
to the vertex v;.
If one uses a different cyclic group !!En as the voltage group and assigns
voltages 1 and k to the loops c and e, then one obtains the generalized
Petersen graph G(n, k) (Frucht et al., 1971).

Example 2.1.3. Figure 2.3 shows a 3-runged "Mobius ladder" as a base graph,
with voltages in f!E2 • Since this base has six vertices and nine edges, and since the
voltage group has order 2, the derived "circular ladder" C6 X P2 has 12 vertices
and 18 edges.

To avoid cluttering the diagram, the names of the edges have been omitted
from Figure 3.3. This does not affect the construction of the derived graph.
Quite often, the genus of the derived graph is higher than the genus of the
base graph, as illustrated by Example 2.1.2. However, the genus of the derived
graph need not be larger. Indeed, as Example 2.1.3 proves, it may be smaller
than the genus of the base graph, for we recognize that the 3-runged Mobius
ladder is isomorphic to the Kuratowski graph K 3 3 , whereas the 6-runged
circular ladder is visibly planar. '

0
Figure 2.3. A circular ladder graph derived by a voltage assignment in .2"2 to a 3-runged Mobius
ladder.
60 Voltage Graphs and Covering Spaces

Figure 2.4. The 3-octahedron graph 0 3 derived by assigning voltages in .2'6 to the bouquet 8 2 .

2.1.2. Fibers and the Natural Projection

If voltages are assigned in a group .511 to a base graph G, then for every vertex
v of G, the set of vertices va in the derived graph is called the "fiber" over v.
Also, for every edge e of G, the set of edges e a in the derived graph is called
the " fiber" over e.
If e is a proper edge running from vertex u to vertex v in the graph G and
if e+ is assigned voltage b, then each edge ea in the fiber over e runs from the
vertex ua in the fiber over the initial point u to the vertex vab in the fiber over
the terminal point v. It follows from elementary group theory that the edge
fiber over e matches the vertices in the vertex fiber over u one-to-one onto the
vertices of the vertex fiber over v. Thus, the fiber over a proper edge is
isomorphic to the disjoint union of #.511 copies of K 2 • For instance, in
Example 2.1.2, the fiber over the proper edge d is the 1-factor between the
outer pentagon and the inner star of the graph. However, as illustrated by
Figure 2.4, the fiber over a loop always forms a set of cycles if the voltage
group is finite.

Example 2.1.4. The fr6-voltage 1 on the loop d of the base bouquet in Figure 2.4
has order 6, and the edge fiber over d forms one 6-cyc/e. The fr6-voltage 2 on the
loop e of the bouquet has order 3, and the edge fiber over e forms two 3-cyc/es.

In general, if the voltage b on a v-based loop e + has order n in the group

.511, then each cycle in the edge fiber over e must have length n, and there must
be #.Jiljn such cycles. For instance, starting at the vertex va there is the
n-cycle

Since only the edge e c + of the fiber over the loop e + starts at the vertex vc of
Ordinary Voltages 61

the fiber over the vertex v, only one of these n-cycles passes through vc, which
implies that these n-cycles are mutually disjoint.
The "natural projection" from the derived graph to the base graph is the
graph map p: Go.~ G that maps every vertex in the fiber over v to the vertex
v, for all v E Vc, and every edge in the fiber over e to the edge e, for all
e E Ec. In terms of the notation, the natural projection wipes out the
subscripts.

2.1.3. The Net Voltage on a Walk

The " voltage on a minus-directed edge" e- is understood to be the group
inverse of the voltage on its reverse edge e+, that is, a(e-) = (a(e+))- 1. This
enables us to define the "net voltage" on a walk

to be the product a( e 1a,)0

a( ena.) of the voltages on the edges of w in the
0 0

order and direction of that walk. For instance, in Example 2.1.2, the net
voltage on the walk e+, d-, c-, d+ is

A "lift" of a walk W = e 1a 1, ••• , ena" in the base graph G is a walk

in the derived graph Go. such that for i = 1, ... , n the edge e; is in the fiber
over the edge e; .

Example 2.1.5. In Figure 2.5, the voltage group is the cyclic group :Z3 • There
are three lifts of the walk c-, e+ from the vertex u to the vertex v of the base
graph, namely, the following walks of the derived graph:

c1 -,e 1 + fromu 0 tov 2

c2 -, e 2 + from u1 to v0
c0 -, e0 + from u 2 to v 1

One important observation about Example 2.1.5 is that although the walk
c-, e- of the base graph includes the same edges as the walk c-, e + and
although it has the same endpoints, its lifts are different. The lifts of the walk
c-, e are
and c0 -,e 2 -.

Continuity is the reason that the lift of the walk c-, e+ starting at the vertex
62 Voltage Graphs and Covering Spaces

Figure 2.5. Another voltage graph and its derived graph. The voltage group is .2"3 •

U; in the derived graph is the walk c;+ 1 -, e;+ 1 +, whereas the lift of c-, e-
starting at u; is c;+ 1 -, e; -. Since the voltage on the directed edge c- is 1, the
next vertex on either lifted walk is V;+ 1 • When e+ is the next edge of the base
walk, the next lifted edge must be the unique edge in the fiber over e whose
plus direction originates at v,+ 1 (i.e., e;+ 1 ). When e- is the next edge of the
base walk, the next lifted edge must be the unique edge in the fiber over e
whose minus direction originates at V;+ 1 (i.e., e,). This is one reason that in
topological graph theory, the direction in which one traverses a loop cannot be
ignored.

2.1.4. Unique Walk Lifting

Another observation about Example 2.1.5 is that starting at each of the three
vertices in the fiber over the initial vertex u of the base walk c-, e+, there is a
unique lifted walk. That is, one lift starts at u0 , one at u1 , and the other at u 2 •
The following theorem indicates the generality of this observation.

Theorein 2.1.1. Let W be a walk in an ordinary voltage graph (G, a) such that
the initial vertex of W is u. Then for each vertex ua in the fiber over u, there is a
unique lift of W that starts at ua.

Proof Consider the first directed edge of the walk W, say e + or e-. If it is
e+, then, since only one plus-directed edge of the fiber over e starts at the
vertex ua (i.e., the edge ea +), that edge must be the first edge in the lift of W
starting at u a. If it is e-, then, since only one minus-directed edge of the fiber
over e starts at ua (i.e., the edge eaa(e-> -),it follows that the edge must be the
first edge in the lift of W starting at ua. Similarly, there is only one possible
choice for a second edge of the lift of W, since the initial point of that second
edge must be the terminal point of the first edge, and since that second edge of
Ordinary Voltages 63

the lift must lie in the fiber over the second edge of the base walk W. This
uniqueness holds, of course, for all the remaining edges as well. D

Notation. According to Theorem 2.1.1 it makes sense to designate the lift of a

walk w starting at the vertex ua by wa.

Theorem 2.1.2. Let W be a walk from u to v in an ordinary voltage graph

(G, a), and let b be the net voltage on W. Then the lift Wa starting at ua
terminates at the vertex vab·

Proof Let b 1, ••• , bn be the successive voltages encountered on a traversal

of the walk W. Then the subscripts of the successive vertices of the lift Wa are
a, ab 1 , ab 1b 2 , ••• , ab 1 ••• bn = ab. Since Wa terminates in the fiber over v, its
final vertex is Dab· D

Example 2.1.5 Revisited. The net voltage on the walk W = c-, e+ from u to v
is (-2) + 1 = 1 + 1 = 2. The walk W0 (= c1 -, e 1 +)ends at v0 + 2 = v2 , the
walk W 1 ( = c 2 -, e 2 +)ends at v 1 + 2 = v0 , and the walk W2 ( = c0 -, e 0 -) ends at
v2+2 = V1.

2.1.5. Preimages of Cycles

The "net voltage on a directed cycle C", based at a vertex u, is defined to be
the product of the voltages encountered on a traversal of C, starting with the
edge that originates at u and then proceeding in the preferred direction. If
the group is abelian, then the net voltage on a directed cycle is independent of
the choice of a base point, so we usually do not bother to designate one. If the
group is nonabelian, then the different net voltages on a directed cycle
corresponding to different choices of a base point are all conjugate group
elements, so they all have the same order in the group. Moreover, the net
voltage on a directed cycle has the same order as the net voltage on the reverse
cycle, because those net voltages are group inverses of each other. Since the
main concern is nearly always the order and not the net voltage itself, we are
usually relieved of the burden of specifying either a basepoint or a direction on
a cycle.

Theorem 2.1.3. Let C be a k-cycle in the base space of an ordinary voltage

graph (G, a) such that the net voltage on C has order min the voltage group .91.
Then each component of the preimage p- 1(C) is a km-cycle, and there are
#s#jm such components.

Proof Let the cycle C be represented by a closed u-based walk W, let b

be the net voltage on W, and let ua be a vertex in the fiber over u. Then the
component of p- 1(C) containing ua is formed by the edges in the walks
Wa, Wab• ... , Wabm-1, which attach end-to-end to form a km-cycle. For each of
64 Voltage Graphs and Covering Spaces

u0 (e5, 0)

Figure 2.6. The 3-cube graph Q3 derived by assigning .2'2-voltages to the complete graph K 4 •

the #.9.1/m left cosets of the cyclic group generated by the net voltage b, there
is a unique component of p - 1( C). 0

Example 2.1.6. Figure 2.6 shows how the 3-cube graph Q3 may be derived
using ~2-voltages on the complete graph K 4 • One observes that the sum (i.e., the
product operation for the group ~2 ) of the voltages on the 4-cycle e 1 , e 2 , e 3 , e 4
is 0 modulo 2, which has order 1, and that the preimage of that 4-cycle is the
union of the two 4-cycles ( e 1 , 0), ( e 2 , 1), ( e 3 , 0), ( e 4 , 1) and ( e 1 , 1),
(e 2 ,0),(e 3 , l),(e 4 ,0). On the other hand, one notes that the 3-cycle e 1 , e 2 , e 5 ,
in the base graph has voltage sum 1 modulo 2, which has order 2, and thus the
preimage in the derived graph of that 3-cycle is the 6-cycle ( e 1 , 0),
( e 2 , 1), ( e 5 , 0), ( e 1 , 1), ( e 2 , 0), ( e 5 , 1).

Example 2.1.5 Revisited Again. The net voltage on the 2-cycle c+, d+ in Figure
2.5 is 1 modulo 3. According to Theorem 2.1.3, the preimage of that 2-cycle
should be one 6-cycle. In fact, it is the 6-cycle c0 +, d 2 +, c 1 +, d 0 +, c 2 +, d 1 +.
Also, the net voltage on the !-cycle e- is 2 modulo 3, and its preimage should be
one 3-cycle. In fact, it is the 3-cycle e 2 -, e 1 -, e 0 -.

2.1.6. Exercises
Figure 2.7 shows four voltage graphs, whose derived graphs are to be con-
structed in Exercises 1-4.
1. Draw the derived graph corresponding to the voltage graph in Figure
2.7a.
2. Draw the derived graph corresponding to the voltage graph in Figure
2.7b.
3. Draw the derived graph corresponding to the voltage graph in Figure
2.7c.
Ordinary Voltages 65
0 u
0

1 0 0 1
u 0 w
(a)
(b)
(1,0,0)

1CD3
(c) (d)

Figure 2.7. Four voltage graphs. (a) In .2'3 , (b) in .2'2 , (c) in .2'2 X .2'2 X .2'2 , (d) in .2'5 .

4. Draw the derived graph corresponding to the voltage graph in Figure

2.7d.
5. Draw copies of all the quotient graphs of the 3-cube graph Q 3 •
6. Assign voltages to each of the graphs obtained in Exercise 5 so that the
derived graph in each case is Q3 •
7. Draw copies of all the quotient graphs of the complete graph K 6 .
8. Assign voltages to each of the graphs obtained in Exercise 7 so that the
derived graph in each case is K 6 •
9. Draw a representative of each of the isomorphism types of graph that
can be derived by assigning voltages fr3 to the dumbbell graph of Figure
2.2.
10. What is the net voltage on the walk e 1 +, e 6 -, e 3 - in Figure 2.6?
Exercises 11-15 all pertain to the voltage graphs of Figure 2.8.
11. What is the net voltage on the walk e 2 +, e 3 -, e 6 - in Figure 2.8a?
Construct the unique lift that begins at the vertex u 4 •
12. How many components does the preimage of the cycle e 4 +, e 5 +, e 6 - in
Figure 2.8a have? What is the length of each?
13. What is the net voltage on the walk e 1 +, e 5 +, e 3 -, e 2 + in Figure 2.8b?
Construct the unique lift that begins at the vertex t< 12>.
14. The preimage of the cycle et, e4 +, e 2 - in Figure 2.8b has how many
components? What is the length of each?
15. The preimage of the cycle e 1 +, e 5 +, e 3 - in Figure 2.8b has how many
components? What is the length of each? Construct the one that contains
the vertex w <234>.
66 Voltage Graphs and Covering Spaces

(134) u

"4 (123)

ll (1 3)(2 4) ill

(a) (h)

Figure 2.8. Two voltage graphs. (a) In .2"6 , (b) in Y4.

2.2. WHICH GRAPHS ARE DERIVABLE WITII ORDINARY VOLTAGES?

We have seen in Section 2.1 that the Petersen graph, the octahedral graph 0 3 ,
the 3-cube graph Q 3 , and several other graphs can all be obtained as the
derived graph of an ordinary voltage-graph construction. The unifying char-
acteristic of these derivable graphs is a certain kind of global symmetry,
precisely described as the existence of a group of automorphisms that acts
freely. Immediately below, we prove that a given graph is derivable with
ordinary voltages from a proposed base graph if and only if the proposed base
graph is a regular quotient of the given graph. This facilitates proof that some
graphs, such as trees, are derivable only from themselves, whereas other
graphs, such as Cayley graphs, are derivable from small base graphs.

2.2.1. The Natural Action of the Voltage Group

Let (G, a) be a voltage graph. For every element c of the voltage group .J/1, let
IJ>c: Ga -+ Ga denote the graph automorphism defined by the rules 1/>i va) = vca
on vertices and 1/>c(ea) = eca on edges. Obviously, lj> 1 : Ga-+ Ga is the identity
automorphism, and no other IJ>c fixes any vertex or edge. Moreover, it is a
straightforward matter to verify that the composition IJ>c o lj>d is the same graph
map as the automorphism 1/>cd· Thus, the voltage group .J/1 acts freely on the
left of the derived graph Ga, and this is called the "natural (left) action" of the
voltage group.

Theorem 2.2.1. The vertex orbits of the natural action of the voltage group on
an ordinary derived graph are the vertex fibers, and the edge orbits are the edge
fibers.

Proof For any vertex va of the derived graph Ga, the orbit of va is defined
to be the set { 1/>c( va) I c E A}, which is the same as the set { vca IcE A} =
Which Graphs Are Derivable with Ordinary Voltages? 67

E A}. That is, the orbit of va is the fiber over v. Similarly, the orbit of
{ vc 1 c
the edge e a of the derived graph is the fiber over the edge e. D

Corollary. The natural projection p: Go.~ G of the derived graph onto the base
graph is a covering projection.

Proof The composition of the quotient map qA: Go.~ Ga/A with the
graph isomorphism Go./A ~ G, which takes the fiber over a vertex or edge to
that vertex or edge, is equal to the natural projection p: co. ~ G. D

2.2.2. Fixed-Point Free Automorphisms

An automorphism of a graph G is called "fixed-point free" if it maps no
vertex or edge of G to itself. Thus, a free group action on a graph may be
identified with a subgroup of Aut( G) in which every automorphism except the
identity is fixed-point free. In particular, every regular quotient arises from
such a subgroup, and if the only such subgroup is trivial, then the only regular
quotient is the given graph itself.

Example 2.2.1. The only nontrivial automorphism of the n-path Pn fixes the
central vertex if n is odd, or the central edge if n is even. Thus then-path has no
regular quotient except itself.

Example 2.2.2. The automorphism group of the complete graph K 4 is isomor-

phic to the symmetric group ~. The only fixed-point free automorphisms of K 4
are 4-cyc/es on both the vertices and the edges. For instance, the vertex map
(abc d) corresponds to a fixed-point free automorphism. How ever, any sub-
group containing such an automorphism would also contain its square, which fixes
two edges. For instance, the square of (a b c d) is (a c)( b d ), whose correspond-
ing edge map would fix the edge between the vertices a and c and the edge
between the vertices band d. Thus, no nontrivial subgroup of Aut( K 4 ) acts freely
on K 4 , and accordingly, the complete graph K 4 has no regular quotients except
itself.

On the other hand, the subgroup of automorphisms on the Petersen graph

in Figure 2.2 generated by the rotation of 2'1T /5 radians acts freely, and the
corresponding quotient is isomorphic to the dumbbell graph. One observes
that the voltage group f!E5 used to derive the Petersen graph from the
dumbbell graph is isomorphic to this subgroup of automorphisms. Moreover,
the free action of the voltage group on the derived graph is consistent with the
free action of this subgroup. Theorem 2.2.2 is a generalization of these
observations.

Theorem 2.2.2 (Gross and Tucker, 1974). Let .91 be a group acting freely on a
graph G, and let G be the resulting quotient graph. Then there is an assignment a
68 Voltage Graphs and Covering Spaces

of ordinary voltages in .5!1 to the quotient graph C and a labeling of the vertices of
G by the elements of VG X .5!1 such that G = ca and that the given action of .5!1
on 6 is the natural left action of .5!1 on ca.

Proof First, choose positive directions for the edges of the graphs C and
G so that the quotient map q~: G -+ C is direction-preserving and that the
action of .5!1 on(; preserves directions. Next, for each vertex v of C, label one
vertex of the orbit q,.,- 1(v) in G as v1 , and for every group element a-:/= 1,
label the vertex cf>A v1 ) as va. Now one may label edges. If the edge e of C
runs from v to w, one assigns the label ea to the edge of the oribt a_,.,- 1(e)
that originates at the vertex va. Since the group .5!1 acts freely on C, there are
#.Jil edges in the orbit q_,.,- 1( e), one originating at each of the #.Jil vertices in
the vertex orbit q_,.,- 1( v ). Thus, the choice of an edge to be labeled e a is
unique. Finally, if the terminal vertex of the edge e 1 is wb, one assigns voltage
b to the edge e of the quotient graph C; that is, a(e+) =b. To show that this
labeling of edges in the orbit q~- 1( e) and the choice of a voltage b for the
edge e really yields an isomorphism G-+ ca, one must show for all a E .5!1
that the edge ea terminates at the vertex w ab· However, since ea = cf>Ae1), the
terminal vertex of the edge ea must be the terminal vertex of the edge cf>a(e 1),
which is

Under this labeling process, the isomorphism G-+ ca identifies orbits in G

with fibers of ca. Moreover, it is defined precisely so that the action of an
automorphism in A on G is consistent with its natural action on the derived
graph ca. D

As a result of Theorem 2.2.2, one knows that a given graph can be derived
from a smaller graph using ordinary voltages if and only if there is a nontrivial
free action on the given graph. To decide whether such an action exists, one
observes that the vertex orbits and the edge orbits under such a free action
would have to be of the same size as the group. Equivalently, in a derived
graph, all the vertex fibers and edge fibers have the same cardinality as the
voltage group. Thus, the order of a possible voltage group to derive a given
graph C must be a common divisor of #VG and #EG. However, this
common-divisor criterion is not always sufficient.

Example 2.2.3. A tree cannot be derived from voltages on some smaller graph,
because the number of vertices and the number of edges of a tree are relatively
prime. A more difficult proof of the fact that trees have no nontrivial regular
quotients might be obtained using Exercise 15 of Section 1.3.

Example 2.2.4. Since the 3-cube graph Q 3 has eight vertices and 12 edges, it
might have one or more regular quotients with four vertices and six edges or with
Which Graphs Are Derivable with Ordinary Voltages? 69

two vertices and three edges, but there are no other possibilities. See Exercise 5
of Section 2.1.

Example 2.2.2 Revisited. The complete graph K 4 has four vertices and six
edges. Although the number 2 is a common divisor of 4 and 6, there is no
quotient of K 4 with two vertices and three edges, because K 4 has no regular
quotients itself, as proved in the original discussion of Example 2.2.2.

Remark. One additional fact that may be used to find the quotients of a graph is
that all vertices in a fiber have the same valence, because the voltage group acts
transitively on the set of neighborhoods of all the vertices in the fiber over any
vertex of the base graph.

2.2.3. Cayley Graphs Revisited

To counterbalance the possibly disappointing news that not all interesting
graphs are derivable with ordinary voltages, we now consider the spectacularly
encouraging example of Cayley graphs. Theorem 2.2.3 identifies Cayley graphs
with regular coverings of bouquets Bn- This suggests that, in developing a
strategy of solving a graph-placement problem by first se~king placements of
its quotients, one might find Cayley graphs of special interest.

Theorem 2.2.3. Let d be a group, and let X= { x 1 , •.. , xn} be a generating

set for d. Then the Cayley graph C(d, X) 0 is a regular covering of the bouquet
Bn. Conversely, if the graph G is a regular covering of the bouquet Bn and if dis
the associated group that acts freely on G, then there is an assignment a of
voltages in d to the bouquet Bn such that the voltages assigned generate d and
that the derived graph Bn" is isomorphic to G.

Proof Let v denote the vertex of the bouquet Bn and e 1 , ••• , en the edges.
Given-the group d and the generating set X, assign the generator x, as the
voltage on the loop e/, for i = 1, ... , n. Then the derived graph Bn" is
isomorphic to the Cayley color graph C(d, X) under the vertex map va-+ a
and the edge map (e;, a)-+ (x;, a). The converse is a special case of Theorem
2.2.2. D

Corollary (Sabidussi, 1958). A connected graph G is isomorphic to a Cayley

graph C(d, X) 0 if and only if some subgroup of Aut( G) acts freely on G and
transitively on VG.

Proof Let G be isomorphic to a Cayley graph C(d, X) 0 • Then the group

of left translations on C(d, X) 0 composes with the isomorphism G -+
C( .!#, X) 0 to act freely on G and transitively on VG. Conversely, if some
group of automorphisms acts freely on G and transitively on VG, then VG is the
only vertex orbit and the resulting quotient is a bouquet of circles, from which
it follows that G is isomorphic to a Cayley graph. D
70 Voltage Graphs and Covering Spaces

Example 2.2.5. The complete graph K 2 n + 1 is isomorphic to the Cayley graph

C(.2'2 n+l, {1, ... , n }) 0 • Equivalently, if one assigns the voltages 1, ... , n
modulo 2n + 1 to the loops of the bouquet Bn, then the derived graph is
isomorphic to K 2 n + 1 •

Example 2.2.6. The octahedron graph On is isomorphic to the Cayley graph

C(.2'2 n,{1, ... , n -1}) 0 • Equivalently, if one assigns the voltages 1, ... ,
n - 1 modulo 2n to the loops of the bouquet Bn_ 1 , then the derived graph is
isomorphic to on.

One might recall that the alternative Cayley graph C(.s#, X) 1 is obtained
from the Cayley graph C(.J/1, X) 0 by deleting a 1-factor for each generator of
order 2.

Example 2.2.7. The complete graph K 2 n is isomorphic to the alternative Cayley

graph C(.2'2 n, {1, ... , n })1•

Example 2.2.8. The n-cube graph Qn is isomorphic to the alternative Cayley

graph C( .2'2n, X) 1, where X is the canonical set of generators {(1, 0, ... , 0),
(0, 1, 0, ... '0), ... '(0, ... '0, 1) }.

2.2.4. Automorphism Groups of Graphs

In general, a Cayley graph C(.s#, X) 0 has additional automorphisms besides
the group of left translations {cpa: a E A}. For instance, the Cayley graph
C(-2'3 , {1}) 0 is a 3-cycle, and although the group of left translations includes
all the rotations of that 3-cycle, the full automorphism group is the dihedral
group !7)3 , which also includes the reflections. Thus, it is not obvious which
finite groups can serve as the automorphism groups of graphs.
A classical theorem of Frucht (1938) is that every finite group is isomorphic
to the automorphism group of some graph. The method in Frucht's proof is to
make a modification on each edge of a Cayley graph for a given group .91 so
that the automorphisms of the modified graph are forced to respect the
directions and the colors on the edges of the Cayley color graph C( .91, X). For
instance, Figure 2.9 shows how to do this for the group fr3 •
All the edges of the Cayley color graph C( !Z3 , { 1}) have the same color. An
"arrow" is placed on each, in effect, by attaching a 2-path near the terminal
vertex and a 3-path near the initial vertex. If there were a second color for
some edges, those edges would be marked with a 4-path and a 5-path, so they
would be distinguished from edges of the first color, and so forth.

Frucht's Theorem (1938). Let .91 be a finite group. Then there is a group G
such that Aut( G) is isomorphic to .Jil.
Which Graphs Are Derivable with Ordinary Voltages? 71

(a)

(b)
Figure 2.9. (a) A Cayley graph for the group .2"3 • (b) A graph whose automorphism group is .2"3 •

Proof Let X= { x 1 , ... , xn} be a generating set for the group .91. It is
immediately clear that every left translation <Pa on the Cayley color graph
C( .91, X) is color-consistent and direction-preserving. It is first to be proved
that every color-consistent, direction-preserving automorphism on C{s#, X) is
a left translation. For this purpose, suppose that the automorphism I is
color-consistent and direction-preserving, and suppose that the image 1(1~) of
the identity vertex is the vertex b. It is to be proved that I= q,b.
r
Since the graph automorphism q,b o 1 on C(s#, X) is color-consistent,
direction-preserving, and identity-preserving, Theorem 1.2.4 implies that its
vertex function coincides with a group homomorphism g: A ~A. Since
f/>b 0 r l is color-preserving on edges, it follows that g(X;) =X; for every
generator X; E X. Therefore, g is the identity group automorphism, for which
it follows that I= <f>b·
The next step is to modify the graph C(s#, X) 0 into a graph G so that the
features of color and direction from C(s#, X) are somehow present in G. To
do this, first insert two new vertices on each edge of C( .91, X) 0 • Then for
i = 1, ... , n and for each edge in C(s#, X) 0 corresponding to color x;, attach
a 2i-path at the new vertex near the terminal end and a (2i + I)-path at the
new vertex near the initial end. The automorphism group of the resulting
graph G is isomorphic to the group of left translations on C(s#, X), which is,
of course, isomorphic to the group .91. D

2.2.5. Exercises
1. Prove that the "wheel" K 1 + Cn has no nontrivial quotient.
2. Let G and H be graphs without cutpoints. Prove that the result of
amalgamating G and H at a single vertex has no nontrivial regular
quotients.
72 Voltage Graphs and Covering Spaces

3. Prove that K 8 has no nontrivial quotients.

4. Prove that the complement of a tree with three or more vertices has no
nontrivial quotients.
5. Prove that K 3• 4 has no nontrivial quotients.
6. Prove that the valence of a vertex in the base graph equals the valence of
any vertex in the fiber over it.
7. Assign voltages in fl'2 to the Kuratowski graph K 5 so that the resulting
derived graph is planar.
8. Assign voltages in fl'2 to the Kuratowski graph K 3 , 3 so that the resulting
derived graph is planar.
9. Prove that for every positive integer n, the regular bipartite graph Kn, n is
a Cayley graph.
10. Prove that if the graph G and H are both Cayley graphs, then their
product G X H is also a Cayley graph.
11. Prove or disprove: The edge-complement of a Cayley graph is a Cayley
graph.
12. Prove or disprove: A regular covering space of a Cayley graph is a Cayley
graph.
13. Prove that the Petersen graph is not a Cayley graph. (Hint: There are
only two groups of order 10.)
14. Construct a graph whose automorphism group is isomorphic to fl'2 X fl'2 •
15. Construct a graph whose automorphism group is isomorphic to fl'3 x fl'3 •
16. One observes that the graph G constructed in the proof of Frucht's
theorem has many more vertices than the order of its automorphism
group .J/1. Moreover, the group .Ji1 does not act transitively on the vertices
of the graph G. A graph H is called a "graphical regular representation"
for the group .Ji1 if #VH = #.J/1 and if .Ji1 acts transitively on VH. For
instance, the complete graph K 2 is a graphical regular representation for
the cyclic group fl'2 . Prove that the dihedral group !!)3 has no graphical
regular representation.

2.3. IRREGULAR COVERING GRAPHS

In topology, a continuous function f: X~ Y is called a "covering projection"

if every point of the topological space Y has a neighborhood NY that is evenly
r
covered, that is, such that every component of 1(Ny) is mapped homeomor-
phically by f onto NY. For compact spaces, such as the topological representa-
tions of finite graphs, it is equivalent to stipulate that the function f be a local
homeomorphism onto the range graph. The covering projections one obtains
this way that are not regular are called "irregular", of course. By a straightfor-
ward generalization of ordinary voltage graphs, we now obtain the complete
Irregular Covering Graphs 73

combinatorial counterpart to topological covering projections, including all

the irregular ones. A graph-theoretic motivation for making such a generaliza-
tion lies in Schreier graphs and, more generally, in "relative" voltage-graph
constructions.

2.3.1. Schreier Graphs

Let ?A be a subgroup of a group .J/1, and let X= { x 1 , ..• , xn} be a generating
set for .J/1. The "(right) Schreier color graph" S( .J/1: ?A, X) has for vertices the
right cosets of ?A in .J/1, the set of which is denoted, .J/1: ?A. The edge (x, ?Aa)
has as its endpoints the vertices ?A a and ?Aax, with its plus direction from ?A a
to ?Aax. The subscript x~a is sometimes used for the edge (x, ?Aa).
If the choice of plus directions and the edge colors are ignored in
S( .J/1: ?A, X), the result is the "Schreier graph" S( .J/1: ?A, X) 0 , also called a
"Schreier coset graph". Schreier graphs are an obvious generalization of
Cayley graphs, in the sense that if ?A= {1..., }, then the Schreier color graph
S(.J/1: ?A, X) is isomorphic to the Cayley color graph C(.J/1, X). Moreover, it is
easily verified (see Exercise 13) that if ?A is any normal subgroup of .J/1, then
the Schreier graph S(.J/1: ?A, X) is isomorphic to the Cayley graph
C( .Jilj?A, Xj!IB), where .Jilj?A denotes the quotient group and X/?A denotes
the set of images of generators under the quotient homomorphism.

Example 2.3.1. The main interest in Schreier graphs is for the case in which the
subgroup is not normal. Consider the subgroup ?A= {1, (2 3)} of the dihedral
group !5)3 , with generating set X= {(123),(12)}. Then S(!5)3 : ?A, X) has as
vertices the right cosets

?A= {1,(23)}

?A(12) = {(12),(123)}
?A(13) = {(13),(132)}

and edges as illustrated in Figure 2.10.

It is obvious that if #X=q, then the Schreier graph S(.Jii:?A,X) 0 is

2q-regular. However, it is also obvious from Figure 2.10 that a Schreier graph
need not be vertex-transitive. In particular, no automorphism can map the
vertex ?A(1 3) onto any other vertex, because ?A(1 3) is the only vertex at which
a loop is based.
The "alternative Schreier graph" S(.J/1: ?A, X) 1 is the obvious generaliza-
tion of the alternative Cayley graph. That is, if the generator x has order 2,
then for every right coset ?Aa, the two edges x~a and x~ax collapse into a
single edge.
74 Voltage Graphs and Covering Spaces
.'fd .':kl
....
'
''
\\
\
\
I
~(13)
/ er--------~~(12)

l I
'-""
(1 2 3) } edge dihedral group ~3
(12) - - - - colors subgroup !14= (1 ~ 3 • (2 3)}

Figure 2.10. (left) The Schreier color graph for Example 2.3.1. (right) The associated Schreier
graph.

2.3.2. Relative Voltages

Since the definition of a Schreier graph is so similar to that of a Cayley graph,
one might think that a Schreier graph also covers a bouquet of circles. In the
topological sense mentioned at the beginning of this section, it does; but since
a Schreier graph need not be vertex-transitive, it follows that such a cover need
not be regular. Thus, to construct all Schreier graphs from bouquets, we need
something more general than ordinary voltage assignments. The more general
object is called a "relative voltage assignment", and it was introduced by
Gross and Tucker (1977).
Let G be a graph whose edges have all been given plus and minus
directions, let .J/1 be a group, and let !!J be a subgroup. A " voltage assign-
ment" for G in .J/1 "relative" to !!J is a set function a from the set of
plus-directed edges of G to the group .J/1. The pair (G, aj!!J) is called a
"relative voltage graph".
The "(right) derived graph" Gaf!ll has as its vertex set the cartesian product
Vc X (.J/1: f!B) and for its edge set the cartesian product Ec X (.J/1: !!J). If the
directed edge e+ of the base graph G runs from vertex u to vertex v, and if
a(e+) = cis the voltage assigned toe+, then the directed e!lla + runs from the
vertex u!!Ia to the vertex v!llac·

Example 2.3.2. Let !!J be the subgroup {1Ei13, (2 3)} of the dihedral group P}3 •
Figure 2.11 shows a voltage assignment on a dumbbell graph in the group !i)3
relative to the subgroup !!J, and the resulting derived graph.

Example 2.3.3. Let !!J be a subgroup of a group .J/1, and let X = {x 1 , ... , x n }
be a generating set for .J/1. Assign directions to the edges of the bouquet Bn and
Irregular Covering Graphs 75

u.Slf(13) y .'.d(l3) u.\11(13)

X,\11(13)

Z.\11(13) ZSlf(12)

U.Slf(12) Y.\11(12)

VSlf(12)

X .'.d(12) x.\lf
Y.Slt
z.\11
u.'.d V,'.d

u u
y

Figure 2.11. A relative voltage assignment on a dumbbell graph and the resulting derived graph.

assign relative voltages x 1 , ••• , xn to the respective plus-directed edges. Then the
resulting derived graph is isomorphic to the Schreier graphS(~: 114, X) 0 .

Given a relative voltage graph (G, ajl14) with voltages in the group ~.
fibers in the relative derived graph are defined as for the ordinary voltage-graph
construction. That is, for any vertex v of the base graph G, the "fiber" over v
is the vertex set { v} X ( ~: !14) in the derived graph. Similarly, for any edge e
of G, the "fiber" over e is the edge set { e} X (~: 114). The "natural
projection" of the relative derived graph onto the base graph is the graph map
that takes the fiber over every vertex onto that vertex and the fiber over every
edge onto that edge. If the subgroup 114 is not normal, then the relative derived
graph Gaf&W does not admit a natural fiber-consistent free group action, so that
the natural projection is not a regular covering. However, it may be called an
irregular covering, in a sense we can now make precise.

2.3.3. Combinatorial Coverings

Let f: G-+ H be a direction-preserving graph map, with H connected. Then f
is called a "(combinatorial) covering projection" if for every vertex v of VG
the following conditions hold: f maps the set of edges originating at v
one-to-one onto the set of edges originating at /( v ), and f maps the set of
edges terminating at v one-to-one onto the set of edges terminating at f( v ). In
the event that plus and minus edge-directions have not been chosen for the
graphs G and H, we call f a "covering projection" if there exists a choice of
edge directions for G and H such that f is direction-preserving and that the
76 Voltage Graphs and Covering Spaces

condition given above holds. It is a straightforward matter to verify that if that

condition holds for any one choice of edge-directions for G and H such that f
is direction-preserving, then it holds for all such choices of domain and range
edge-directions.
A "topological representation" of a graph map f: G ~ H is a continuous
function from a topological representation of G to a topological representa-
tion of H whose restriction to the interior of any edge of G is a homeomor-
phism and is consistent with the graph map f. Persons familiar with topology
will recognize that a graph map f is a combinatorial covering projection if and
only if a topological representation of f is a topological covering projection.

Example 2.3.4. Let p: Ga ~ G be the natural projection for an ordinary voltage

graph (G, a). If e 1 , ••• , em are the edges originating at the vertex u of the base
graph G, then for any element .91 of the voltage group, the edges originating at
the vertex ua of the fiber over u are (e 1 , a), ... ,(em, a). Thus, the natural
projection maps the set of edges originating at ua bijectively to the set of edges
originating at u. Moreover, if d 1 , ••• , dn are the edges terminating at the vertex
v, and if b 1 , ••• , bn are the respective voltages on those edges, then
( d 1 , ab 1 - 1 ), ••. , ( d n, abn - 1 ) are the edges terminating at va in the derived
graph. The natural projection p: Ga ~ G also maps the edges terminating at va
one-to-one onto the edges terminating at v. Thus, the natural projection p is a
covering projection. Since, according to Theorem 2.2.2, every regular covering
projection is equivalent to the natural projection of an ordinary derived graph onto
its base graph, it follows that the topological representation of every regular
covering projection of graphs is a covering projection, as one might expect from
the terminology.

Warning 2.3.1. Naive attempts to simplify the definition of a covering projection

lead to different constructions that fail to have some important properties of
covering projections. For instance, as Example 2.3.5 illustrates, it is not
sufficient to specify that the graph map f preserves valence, or even (see Exercise
15) that the inverse images of every vertex and edge of the range have the same
cardinality.

Example 2.3.5. The graph map f illustrated in Figure 2.12 that is defined by the
dropping of subscripts preserves valence, since all the vertices have valence 3.
However, the two edges d 1 and d 2 originating at u 1 are both mapped onto the
same edge d, originating at f(u 1 ) = u. Thus, f is not a covering projection,
because its restriction to edges originating at u 1 is not one-to-one. Moreover,
changing edge-directions on the graphs will not help at all.

Example 2.3.6. Let p: Gaf!ll ~ G be the natural projection for a relative

derived graph. Then p is a covering projection, by the same reasoning as in
Example 2.3.4.
Irregular Covering Graphs 77

u 1 ~--------...._ u2
el

Figure 2.12. A valence-preserving graph map that is not a covering projection.

Example 2.3.3 Revisited. As a special case of"Example 2.3.6, observe that a

Schreier graph S(s#: 114, X) 0 is a covering space of the bouquet Bn, where n is
the cardinality of the generating set X.

Example 2.3.7. Let G denote the edge-complement of the disjoint union of a

3-cycle and a 4-cycle. Then although G is 4-regular, it is not vertex-transitive.
Thus, the graph G is not a Cayley graph, and it cannot be derived by assigning
ordinary voltages to the bouquet of circles B 2 • However, the graph map il-
lustrated in Figure 2.13, which takes solid edges to solid edges and dashed edges
to dashed edges and preserves directions, is a covering projection.

The domain of a covering projection f: G ~ H of graphs is called the

"covering graph" (or "covering space") and the range is called the "base
graph" (or "base space"). Any covering projection that is not regular may be
called "irregular", and its domain may be called an "irregular covering space"
(or "irregular covering graph") of the base space (or base graph). For instance,
Example 2.3.7 shows that the graph (C3 u C4 )c is an irregular covering space

Uj

u6 u2

u3
-- 0'~)
..........

u4

Figure 2.13. An irregular covering projection of the 4-regular graph (C3 U C4 )c onto the bou-
quet 8 2 •
78 Voltage Graphs and Covering Spaces

of the bouquet B 2 • It is possible for the same graph to be both a regular

covering space and an irregular covering space of a given base graph (by two
different covering projections; see Exercise 2.4.8). However, the graph ( C3 u
C4 )c cannot, under any graph map, be a regular covering of the bouquet B 2 ,
because it is not vertex-transitive.

2.3.4. Most Regular Graphs Are Schreier Graphs

In examining Example 2.3.7 and the accompanying Figure 2.13, one might
notice that the solid edges in the domain graph form a hamiltonian cycle and
the dashed edges also form a hamiltonian cycle. With the solid-edge cycle we
associate the permutation 'IT1 = (1 2 7 3 4 56), corresponding to the order of the
indices of vertices encountered as one traverses the hamiltonian cycle in its
order of orientation. With the dashed-edge cycle we may in like fashion
associate the permutation 'IT2 = (1 3 2 4 6 7 5).
Now let .91 be the group of permutations generated by 'IT1 and '172 , and let
f1B be the subgroup of .91 that fixes the symbol 1. For i = 1, ... , 7, let /; be a
permutation in .91 such that /;(1) = i. For instance, it is possible to obtain /;
as a power either of 'IT1 or of '172 • Then, for i = 1, ... , 7, the right coset f!B/;
contains all the elements of .91 that permute 1 onto i. Rename the vertices of
the graph in Figure 2.13 so that, for i = 1, ... , 7, the right coset ?A/; replaces
the label v;. Also, every edge of the solid cycle is labeled 'ITp and every edge
of the dashed cycle is labeled 'IT2 • Then the resulting graph is a Schreier graph
S( .91: f!B, X) 0 , where x = {'IT1 , 'IT2 }. This observation leads to the following
result.

Theorem 2.3.1 (Gross, 1977). Let G be a connected regular graph of even

valence. Then G is isomorphic to a Schreier coset graph.

Proof Suppose that G is 2n-regular and that its vertices are v1 , ... , vr By
Theorem 1.5.2 (Petersen's theorem) the edges of G can be partitioned into n
2-factors, L 1 , ... , Ln. Assign an arbitrary orientation to each cycle of each
2-factor (thereby inducing plus directions on all the edges of G). If a 2-factor
L; is a hamiltonian cycle, then we could associate a cyclic permutation with it
exactly as in the preceding discussion of Figure 2.10. If the 2-factor L; has
several component cycles, cl' ... ' em, then with each of these edge cycles we
may associate a cyclic permutation corresponding to the order of the indices
encountered as one makes a traversal in the order of orientation, and the
permutation 'IT; to be associated with the 2-factor L; is the product of the
disjoint cycle permutations corresponding to its component cycles.
Let .91 be the group generated by the permutations 'IT1 , ••• , 'ITn correspond-
ing to the respective 2-factors L 1 , ••• , Ln, and let f1B be the subgroup of .91
consisting of all elements of .91 that fix the symbol 1. Then the graph G is
isomorphic to the Schreier graph S(.91: f!B, X) 0 , where X= { '171, ... , 'ITn}· To
construct the isomorphism, first assign each edge in the 2-factor L; the color
Irregular Covering Graphs 79

'IT,, for i = 1, ... , n. Then replace each vertex label vi by the right coset of !!4
that permutes 1 onto j, for j = 1, ... , p. The subgroup !!4 has such cosets
because G is connected. In particular, the product of the "signed colors" one
traverses on an edge path from v1 to vi is a permutation ~ such that /;(1) = j,
so that !14~ is the right coset that permutes 1 onto j. D

Corollary. Let B be a connected regular graph of odd valence such that G has a
!-factor. Then G is isomorphic to an alternative Schreier graph.

Proof The edge-complement in G of the 1-factor is a regular graph of

even valence. Thus, by Theorem 1.5.2, the edge set of G can be partitioned
into a 1-factor plus some 2-factors. By an argument similar to the proof of
Theorem 2.3.1, it is possible to construct an isomorphism of G to an
alternative Schreier graph, where the 1-factor is colored by a permutation of
order 2 and the respective 2-factors are each colored by other permutations.
D

Unfortunately, not all regular graphs of odd valence have a 1-factor (see
Exercise 5). Biggs eta!. (1976) give a historical account of 1-factors, including
Tait's "theorem" that 2-connected, 3-valent graphs are !-factorable and
Petersen's counterexample, the graph that now bears his name. Akiyama and
Kano (1985) survey more recent results on graph factorization. Tutte (1947)
gives a necessary and sufficient condition for the existence of a 1-factor, from
which he obtains the following theorem.

Theorem 2.3.2 (Tutte, 1947). Let G be an (n - I)-connected regular graph of

valence n with an even number of vertices. Then G has a !-factor.

Proof Omitted. D

Corollary. Let G be a 2r-connected regular graph of valence 2r + 1. Then G is

isomorphic to an alternative Schrier graph.

Proof From Theorem 1.1.1 (Euler's theorem on valence sum), it follows

that G has an even number of vertices. From Theorem 2.3.2, we infer that G
has a 1-factor. The conclusion then follows from the corollary to Theorem
2.3.1. D

2.3.5. Exercises
I. Let !!4 be the subgroup generated by the element (1 2 4) of the alternating
group .J/14 , and let X be the generating set {(1 2 3), (2 3 4)}. Draw the
Schreier graph S( .J/14 : !!4, X) 0 •
80 Voltage Graphs and Covering Spaces

Figure 2.14. A 4-regular graph.

2. Let !11 be the cyclic subgroup generated by the element (1 2 3 4) of the

symmetric group~. and let X be the generating set {(1243),(12)}.
Draw the alternative Schreier graphS(~: ?A, X) 1•
3. Let ?A be the subgroup generated by the element (1, 1, 1) of the group
~/.and let X be the canonical generating set {(1,0,0),(0, 1,0),(0,0, 1)}.
Draw the alternative Schreier graph S(~2 3 : ?A, X) 1•
4. Label the graph in Figure 2.14 as a Schreier coset graph.
5. Construct a 3-regular graph that has no 1-factor.
6. Prove that every regular bipartite graph of positive valence has a 1-factor
(Konig, 1916). (Hint: Consider the result of doubling every edge.)
7. Use Exercise 6 to prove that every regular bipartite graph is 1-factorable.
8. Prove that every graph has a bipartite double covering space.
9. Use Exercises 6 and 8 to prove that every connected regular graph has a
double covering that is a Schreier graph (or an alternative Schreier
graph).
10. Prove that any two Schreier graphs of the same valence have a common
covering space. (Hint: Consider the product group.)
11. Prove the theorem of Angluin and Gardiner (1980) that any two con-
nected n-regular graphs have a common covering. (Hint: Use Exercises 9
and 10.)
12. Does every graph of maximum valence 3 or less have a planar covering
space? From Example 2.1.3, we recall that the Kuratowski graph K 3, 3
has a planar double covering.
13. Let !11 be a normal subgroup of a group .91 and X a generating set for
.91. Prove that the Schreier graph S( .91: ?A, X) is isomorphic to the
Cayley graph C( .91/PA, X/PA), where Xj?A denotes the set of images of
generators under the quotient homomorphism.
Pennutation Voltage Graphs 81

14. Is every covering projection a local isomorphism? Is every local isomor-

phism a covering projection?
15. Give an example of a graph map f: G-+ H, with G and H connected,
such that the inverse image of every vertex or edge of H has the same
number of elements but that f is not a covering projection.

2.4. PERMUTATION VOLTAGE GRAPHS

The relative-voltage-graph construction is a natural extension of the ordinary

voltage-graph construction, but the subgroup makes it somewhat awkward to
use. In this section we introduce another computational device, the permuta-
tion voltage graph, which is equivalent to a relative voltage graph but is more
convenient to apply. The permutation voltage-graph construction is also due
to Gross and Tucker (1977).

2.4.1. Constructing Covering Spaces with Permutations

Let G be a graph whose edges have all been assigned plus and minus
directions. A "permutation voltage assignment" for G is a function a from the
plus-directed edges of G into a symmetric group Y,. The elements of the
image Ga are called "permutation voltages", and the subscripted pair (G, a)n
is called a "permutation voltage graph". The graph G is called the "base
graph" or "base space".
To each such permutation voltage graph there is associated a "permutation
derived graph" Ga, whose vertex set is the cartesian product Vc X {1, ... , n}
and whose edge set is the cartesian product Ec X {1, ... , n }. As for ordinary
derived graphs, one uses either the pair notations, (v, i) and (e, i), or the
subscripted notations, v, and e,, whichever alternative is more convenient in
the context at hand. If the edge e of the base graph G runs from the vertex u
to the vertex v, and if the voltage on e+ is the permutation TT, then for
i = 1, ... , n, the edge e; of the derived graph Ga runs from the vertex U; to the
vertex v,<;J.
The natural projection p: Ga-+ G for permutation voltage graph (G, a)n is
the graph map that takes any vertex V; or edge e; of the derived graph to the
vertex v or edge e of the base graph. The set of vertices { v; 1 i = 1, ... , n } is
called the "fiber" over v and the set of edges { e,l i = 1, ... , n} is called the
"fiber" over e.

Warning 2.4.1. When a assigns permutation voltages to a graph in a symmetric

group Y,, it is absolutely necessary to distinguish (G, a)n from the ordinary
voltage graph (G, a), because the respective derived graphs are quite different.
Whereas the cardinality of each fiber in the ordinary derived graph would ben!,
the order of the group Y,, the cardinality of each fiber in the permutation derived
graph would ben, the number of permuted objects.
82 Voltage Graphs and Covering Spaces

(2 3)
u v
(1 2)
y
Figure 2.15. A permutation voltage assignment and the derived graph.

Example 2.4.1. Figure 2.15 shows a permutation voltage assignment to the

dumbbell graph in the symmetric group Y'3 and the associated permutation
derived graph.

2.4.2. Preimages of Walks and Cycles

The permutation voltage on a minus-directed edge e- is understood to be the
inverse permutation of the voltage on e+. This enables us to define the "net
permutation voltage on a walk" as the product of the voltages encountered in
a traversal of that walk, exactly as for ordinary voltages. For instance, the net
permutation voltage on the walk y-, x-, y+ of Figure 2.15 is (2 3)(13 2)(2 3)
= (1 2 3). A lift of a walk W in the base graph G is a walk W in the
permutation derived graph such that the natural projection p maps the edges
of W onto the edges of W precisely in the order of traversal. It is implicit that
direction is preserved on every edge of a lift, which is an important require-
ment if loops occur.

Theorem 2.4.1. Let W be a walk in the base space of a permutation voltage

graph (G, a)n such that the initial vertex of W is u. Then for each vertex U; in
the fiber over u, there is a unique lift of W that starts at u,.

Proof This theorem is the analog of Theorem 2.1.1, and its proof is
exactly analogous to the proof of that theorem. D
Permutation Voltage Graphs 83

Notation. According to Theorem 2.4.1, it is appropriate to designate the lift of

a walk W starting at the vertex u, as W,.

Example 2.4.1 Revisited. The walk W = y-, x-, y + in the base graph of
Figure 2.15 has three lifts, namely, the walks WI = YI -, x3 -, Y3 +; = w2
Y3 -, x2 -, Y2 +; and w3 = Y2 -, xl -, YI +.

Theorem 2.4.2. Let W be a walk in the base space of a permutation voltage

graph (G, a)n, and let., be the net voltage on W. Then the lift W; starting at the
vertex u; terminates at the vertex v,<;>.

Proof Let 'TT1 , ••• , .,, be the successive permutation voltages encountered
on a traversal of the walk W. Then the subscripts of the successive vertices of
the lift W, are

Since W; terminates in the fiber over v, its final vertex is v,<;>· D

The "net voltage on a permutation directed cycle" C, based at a vertex u, is

defined to be the product of the voltages encountered on a traversal of C,
starting with the edge that originates at u and then proceeding in the given
direction. We recall from Theorem 2.1.3 that in an ordinary voltage graph, a
k-cycle with net voltage of order m in the voltage group .J/1 lifts to #.Jiljm
disjoint km-cycles. In a permutation voltage graph, although the preimage of a
cycle in the base graph is still a union of disjoint cycles in the derived graph,
the cycles in the preimage are not generally all of the same length. This is one
of the most important differences between permutation voltage graphs and
ordinary voltage graphs. In a permutation voltage graph, the cycle structure of
the net permutation voltage on the base cycle is what determines the number
and lengths of the component cycles of the preimage in the derived graph.
Let ., be a permutation in the symmetric group Y,. One recalls that any
permutation has an essentially unique decomposition as a product of disjoint
CY.clic permutations. For j = 1, ... , n, let c1 denote the number of j-cycles in
the decomposition of .,, Then the n-tuple ( c1, ••• , en) is called the "cycle
structure" of .,,

Theorem 2.4.3. Let C be a k-cycle in the base space of a permutation voltage

graph ( G, a )n with net voltage '1T, and let ( c 1 , ..• , cn) be the cycle structure of '1T.
Then the preimage of c in the derived graph G has c1 + · · · +en components,
including for j = 1, ... , n exactly c1 kj-cycles.

Proof Choose any vertex u of the cycle C and regard C as a walk from u
to itself. From Theorem 2.4.2, it follows that for i = 1, ... , n the lift of C
starting at u; terminates at u,<;>· Thus, the component of p- 1(C) that
84 Voltage Graphs and Covering Spaces

contains u; also contains uw(i)• uw2(i)• ••• , u.,,-'<i>• where j is the length of the
cyclic permutation containing the object i in the decomposition of 'IT. There-
fore, that component is a kj-cycle. The conclusion now follows readily. D

Example 2.4.2. Suppose that C is a u-based k-cyc/e with net permutation

voltage (1 4 5)(2)(3 9)(6 7 8) in .51'9 • Then the preimage p - 1( C) has four compo-
nents, namely, a k-cyc/e containing the vertex u 2 , a 2k-cyc/e containing the
vertices u 3 and u 9 , and two 3k-cyc/es, one containing the vertices u1 , u 4 , and
u 5 , the other containing the vertices u 6 , u 7 , and u 8 •

2.4.3. Which Graphs Are Derivable by Permutation Voltages?

The next two theorems indicate that the set of connected graphs that can be
obtained from a permutation voltage assignment of a given graph G is
precisely the set of covering graphs of G.

Theorem 2.4.4. The natural projection p: Ga -+ G associated with any permuta-

tion voltage graph (G, a)n is a covering projection on each component of its
domain.

Proof This is a straightforward exercise in applying the definitions of

permutation derived graph and covering projection, so details are omitted. D

Theorem 2.4.5 (Gross and Tucker, 1977). Let the graph map q: G-+ G be a
covering projection. Then there is an assignment a of permutation voltages to the
base graph G such that the derived graph Ga is isomorphic to G.

Proof Choose a spanning tree Tin the base graph G, pick a root vertex u,
and assign plus directions to the edges of T so that it is possible to travel from
the root u to any other vertex of G entirely along plus-directed edges of T.
Assign plus directions to the other edges of G arbitrarily. Then assign
directions to the edges of G so that the map q is direction-preserving.
Suppose that the preimage q- 1(u) has n vertices. Label them u1 , ••• , un
arbitrarily. Then choose an edge e of T originating at u and terminating, say,
at v. By the definition of covering projection, the preimage q- 1(e) consists of
n edges, one originating at each of the vertices u1 , ••• , un. Moreover, exactly
one of these terminates at each vertex of q - 1( v ), so that there are also n
vertices in q- 1(v). Label the edge of q- 1(e) that originates at u, as e;, for
i = 1, ... , n. Also, label then vertices of q- 1(v) as v1 , ••• , vn in any way. If
we match q- 1(u) to q- 1(v), the edges of q- 1(e) define a permutation 'IT in the
symmetric group Y',. That is, if u; is matched to v1 then 'TT(i) = j. The
permutation voltage 'IT is assigned to the edge e of the base graph G. (If the
vertices v1 , ••• , vn are labeled so that for i = 1, ... , n, the edge e, terminates
at v;, then the voltage assigned to the edge e is the identity permutation.)
Permutation Voltage Graphs 85

Continue this procedure until all edges of T have been assigned permuta-
tion voltages, always selecting as the next edge of Tone whose initial point lies
on a path from the root vertex u whose every edge already has a permutation
voltage assigned. (By the parenthetic remark at the end of the preceding
paragraph, it is possible to assign the identity permutation to every edge of the
maximal tree T.) The last step is to assign permutation voltages to the edges of
G not in T. Suppose that d is such an edge, running from a vertex s to a
vertex t. Then q- 1(d) matches s 1 , ••• ,sn to t 1 , ••• ,tn, so we may assign
names dp ... , dn to the edges in q- 1(d) in agreement with the subscripts on
their initial points, and we may assign a permutation voltage to the edge d
according to the matching. D

Corollary I. Let the graph map p: G-+ G be a covering projection. Let u and v
be any two vertices of the base graph G and e any edge of G, not necessarily
incident either on u or on v. Then #p- 1(u) = #p- 1(v) = #p- 1(e). That is,
there is a number n such that all vertex fibers and all edge fibers have cardinality
n. D

Corollary 2. Let the graph map p: G -+ G be a covering projection on every

component of its domain, and let the inverse images of all vertices of G have the
same cardinality n. Then there is a permutation voltage assignment a in Y, such
that the derived graph Ga is isomorphic to G. D

2.4.4. Identifying Relative Voltages with Permutation Voltages

Since every component of a relative derived voltage graph covers the base
graph by the natural projection, it follows from Theorem 2.4.5 that it could
also be derived by permutation voltages. An explicit way to convert a relative
voltage graph (G, aj?A) to a permutation voltage graph begins with the
observation that the voltage group .J/1 acts as a permutation group on the right
cosets of ?A. Thus, if n = #(.91: ?A), then the permutation action of .J/1 on
.91: ?A imbeds in the symmetric group action on {1, ... , n }. As a special case,
we observe that an ordinary voltage graph converts to a permutation voltage
graph because the action of a group on its own elements imbeds in a
symmetric group action.
Conversely, if (G, a)n is a permutation voltage graph, then let .Jil= Y,, and
let ?A be the subgroup that fixes the symbol 1. The right cosets

?A, ?A(12), ?A(13), ... , ?A(1 n)

behave under right multiplication by an element of '1T of .91 exactly as the

symbols 1, 2, 3, ... , n behave under the permutation action of 'TT. Thus, the
relative voltage graph (G, aj?A) is equivalent to the permutation voltage
graph (G, a).
86 Voltage Graphs and Covering Spaces

(1 2)

(1 2)

1y 2
Figure 2.16. A graph with permutation voltages in 92.

2.4.5. Exercises
1. Construct the derived graph corresponding to the permutation voltage
graph in Figure 2.16.
2. Prove that the composition of two covering projections is a covering
projection.
3. Prove that the composition of two regular covering projections can
sometimes be an irregular covering projection. (Hint: Use Exercise 1.)
4. Let G be a graph obtained from the Petersen graph by doubling the
1-factor that extends from the star to the pentagon. (That is, for each
edge e in the 1-factor, attach another edge e' with the same endpoints.)
Assign permutation voltages in .51'10 to the bouquet B2 so that the
derived graph is isomorphic to G.
5. Use Exercise 4 to prove that the Petersen graph is an alternative Schreier
graph, by explicitly exhibiting the group, the subgroup, and the gener-
ating set.
6. Prove that any two graphs that cover the same graph have a common
covering space.
7. Let C be a hamiltonian cycle in the base space of a permutation voltage
graph. Under what circumstance is the preimage of C (under the natural
projection) a hamiltonian cycle in the derived graph?
8. Show that there exists both a regular covering projection and an irregular
covering projection of the complete graph K 7 onto the bouquet B3 •
9. Let F be an n-factor in the base space of a covering projection
p: G-+ G. Prove that the preimage p- 1(F) is ann-factor in the covering
space.
10. Prove or disprove: If the base graph of a (finite) covering projection is
n-connected, then so is each component of the covering graph.
11. Prove that any 2-sheeted covering projection is a regular covering projec-
tion.

2.5. SUBGROUPS OF THE VOLTAGE GROUP

To decide whether two vertices in an ordinary derived graph are in the same
component or to count the number of components, one considers a form of
Subgroups of the Voltage Group 87

isotropy group called a local voltage group, which is a subgroup of the voltage
group itself. The considerations involved here also lead to the proof of Babai's
theorem, that given a subgroup !14 of a group .511 and a Cayley graph G for .511,
there is a Cayley graph H for !14 such that G contracts onto H.

2.5.1. The Fundamental Semigroup of Closed Walks

Let W and W' be walks in a graph G such that the initial vertex of W' is the
terminal vertex of W. Then the product walk WW' is constructed by extending
the sequence of edges in W by the sequence of edges in W'. If Wand W' are
both u-based closed walks, then so is WW'. Under this product operation, the
set of all u-based closed walks forms a semigroup, called the "fundamental
semigroup of u-based closed walks".
If u is any vertex of the base graph of either an ordinary voltage graph or a
permutation voltage graph, then the set of net voltages occurring on u-based
closed walks forms a subgroup of the voltage group, as originally observed (in
a current group formulation-see Chapter 4) by Alpert and Gross (1976,
p. 297). Following Stahl and White (1976), who first reinterpreted it for volt-
age assignments, we adopt the name "local (voltage) group" at u, and we de-
note it .511 ( u ).

Example 2.5.1. Since every u-based closed walk in the base graph of Figure
2.17 traverses the edge between u and v an even number of times, it follows that
the local group at u is the subgroup {0, 2, 4} of the voltage group ~6 • Similarly,
the local group at v is also {0, 2, 4 }.

Theorem 2.5.1 (Alpert and Gross, 1976). Let u be a vertex in the base space of
an ordinary voltage graph (G, a) with voltages in the group .511. Then the vertices
ua and ub are in the same component of the derived graph Ga. if and only if the
voltage group element a- 1b lies in the local group .Jil(u).

Proof If the vertices ua and ub are in the same component of the derived
graph Ga., then there is a path from ua to ub. The image of that path under the

Figure 2.17. A connected ordinary voltage graph whose derived graph is not connected. The
voltages are in .2'6 •
88 Voltage Graphs and Covering Spaces

natural projection is a u-based closed walk in the base graph G, whose net
voltage lies in the local group .Jil(u). The value of that net voltage must be
a- 1b, by Theorem 2.1.2. Reversing these steps yields the converse conclusion.
0

Corollary 1. Let u and v be the vertices in the base graph G and W a walk from
u to v with net ordinary voltage c. Then the vertices ua and vh of the derived
graph lie in the same component if and only if the voltage group element a-lb lies
in the right coset .Jil(u)c of the local group at u. 0

Corollary 2. The number of components of an ordinary derived graph Ga is

#(..#: .Jil(u)), where u is any vertex of a connected base graph G. 0

Since the voltage group ..# acts transitively on an ordinary derived graph
ca, it follows that the components of Ga are mutually isomorphic. Example
2.5.2 illustrates that the components of a permutation derived graph need not
be isomorphic. Example 2.5.2 also illustrates another fact. If c is the ordinary
or permutation net voltage on a walk from a vertex u to a vertex v in the base
graph, then the local group at u is conjugate to the local group at v. In
particular, ..#( u) = c- 1..#( u)c.

Example 2.5.2. In the permutation voltage graph of Figure 2.18, the local group
at u is the direct product of the subgroup of all permutations on {1, 2} with the
subgroup of all permutations on {3, 4, 5}. The local group at v is the direct
product of the subgroup of all permutations on {1, 3} with the subgroup of all
permutations on {2, 4, 5}.

For the most part, there are permutation voltage analogs to the theorems
about ordinary voltage graphs. For instance, Theorem 2.5.2 and its corollaries
are analogs to Theorem 2.5.1 and its corollaries.

Theorem 2.5.2 (Ezell, 1979). Let u be a vertex in the base space of the
permutation voltage graph (G, a)n· Then the vertices U; and u1 are in the same

Figure 2.18. A connected permutation voltage graph whose derived graph is not connected. The
voltages are in 5"5 .
Subgroups of the Voltage Group 89

component of the derived graph G if and only if the symbols i and j are in the
same orbit of the action of the local group at u on {1, ... , n }.

Proof The vertices u; and u1 are in the same component of the derived
graph if and only if there is a path from u; to u1 . The image of such a path
under the natural projection is a u-based closed walk in the base graph, whose
net voltage lies in the local group at u. The value of that net voltage must be a
permutation taking i onto j, by Theorem 2.4.2. D

Corollary 1. Let u and v be vertices in the base graph G and W a walk from u
to v with net permutation voltage '1T in Y,. Then the vertices u; and v1 of the
derived graph lie in the same component if the right coset Y,(u)'TT contains a
permutation that maps i onto j. D

Corollary 2. Let (G, a)n be a connected permutation voltage graph, and let u
be a vertex of the base graph. The number of components of the derived graph
equals the number of orbits induced by the local group at u on the set {1, ... , n }.
D

2.5.2. Counting Components of Ordinary Derived Graphs

If the base graph G of an ordinary voltage graph is a bouquet of circles, then
the local group at the vertex is simply the subgroup generated by the voltages
on all the loops. If the base graph is not a bouquet of circles, then a
preliminary procedure is required to isolate the generators of the local group
before one has any hope of counting the components. For instance, even
though the proper edge in the base graph of Example 2.5.1 has voltage
1 modulo 6, the local group does not include 1 modulo 6. Thus, the local group
is not necessarily generated by the image of the voltage assignment.
The preliminary procedure is, in effect, to convert the base graph G into a
bouquet of circles. In the discussion that follows, we suppose that the base
graph is connected, for otherwise we would simply iterate the procedure for
every component of G. The first step is to select an arbitrary spanning tree T
and to regard any vertex for which one wants the local group as the root.
Although the case illustrated is for ordinary voltages, the same procedure also
applies to permutation voltages.

Example 2.5.3. Figure 2.19a shows a voltage graph (G, a) with voltage group
.2"12 • Figure 2.19b shows the choice of a root u and a spanning tree T.

For every vertex v in the base graph G, there is a unique path in the tree T
from the root u to v. Define the T-potential a( v, T) of the vertex v to be the
net voltage along that path. The second step is to compute the T-potential of
every vertex of G. Figure 2.20 shows the result of this computation for the
voltage graph of Example 2.5.3.
90 Voltage Graphs and Covering Spaces

(a)

u 8 w u w
Figure 2.19. (a) A graph with voltages in the cyclic group .2"12 • (b) The choice of a root and a
spanning tree.

11 8 7 Figure 2.20. T-potentials for the vertices of G.

If the plus direction of the edge e in the base graph G has initial vertex v
and terminal vertex w, then the "T-voltage" aT( e) is defined to be the product

a(v, T)a(e)a(w, T}- 1

The third step is to compute the T-voltage of every edge in G. It should be

noted that the T-voltage of every edge in the spanning tree T is the group
identity. Figure 2.21 shows the T-voltages for the edges of the voltage graph of
Example 2.5.3.

u 0 w Figure 2.21. T-voltages for the edges of G.

Subgroups of the Voltage Group 91

If there is an edge between two vertices whose voltage is the identity, then
those two vertices have the same local group, by Corollary 1 to Theorem 2.5.1
(or by Corollary 1 to Theorem 2.5.2, for permutation voltages). If both of
those vertices are identified with the new vertex obtained as a result of
contracting the base graph along that edge, then it follows that the local group
at every vertex is preserved in the resulting voltage graph. Thus, the local
group at the root vertex u with respect to the T-voltages could be obtained
from the bouquet that results from contracting the base graph G along the
entire spanning tree T. That is, the voltages on that bouquet generate the local
group at u with respect to the T-voltage assignment. However, these genera-
tors are simply the set of T-voltages on those edges of the base graph G that
are not in the spanning tree T, so one need not actually carry out the
contraction.
For Example 2.5.3, the local group at u with respect to the T-voltages is
now calculated to be the subgroup {0, 3, 6, 9}. The purpose of this calculation
is to show that the local group at u with respect to the original voltage
assignment a is the same subgroup, as proved by the following theorem.

lbeorem 2.5.3 (Gross and Tucker, 1979a). Let G be a graph, let a be either an
ordinary or a permutation voltage assignment for G, and let u be any vertex of G.
Then the local group at u with respect to the T-voltages, for any choice of a
spanning tree T, is identical to the local group at u with respect to a.

Proof The net voltage on any u-based closed walk is the same with
respect to T-voltages as with respect to a. D

lbeorem 2.5.4. Let G be a graph, let a be either an ordinary or a permutation

voltage assignment for G, and let T be a spanning tree for G with root vertex u.
Then the derived graph Gar corresponding to the T-voltages is isomorphic to the
derived graph Ga corresponding to the original assignment.

Proof A relabeling procedure is implicit in the parenthetic comments in

the proof of Theorem 2.4.5. We now make it explicit. For each vertex v of G,
relabel the vertices in the fiber over v according to the rule i ~ ic- 1 , where c
is the net ordinary voltage on the unique path from the root u to the vertex v,
or according to the rule i ~ w- 1(i), if that path has net permutation voltage
w. If e is an edge originating at v, then also change the subscripts of edges e;
in the fiber over e so that they agree with the new subscripts of their initial
points. This relabeling defines an isomorphism Ga ~ Gar. D

Corollary. The number of components of the derived graph Ga equals the index
in the voltage group of the subgroup generated by the T-voltages. D
92 Voltage Graphs and Covering Spaces

2.5.3. The Fundamental Group of a Graph

Let W be a u-based closed walk in a connected graph G. If no directed edge in
W is followed by its reverse, we say that W is a "reduced walk".
Two u-based closed walks W and W' are called "equivalent" if there is a
sequence

W= wco>, w(l>, ... , wcr) = W'

such that for k = 1, ... ' r, the walk w(k) differs from its predecessor w(k-l)
by either the insertion or the removal of a directed edge followed by its
reverse. By induction, one readily verifies that every u-based closed walk is
equivalent to a unique reduced walk.
The equivalence classes formed in the fundamental semigroup of u-based
closed walks form a group, in which the inverse of the equivalence class of a
u-based closed walk is the equivalence class of the reverse walk. This group is
called the "fundamental group of the graph G based at u ", and is denoted
'1T 1(G, u). For a connected graph G, the isomorphism type of its fundamental
group is independent of the choice of the base vertex u.
There is an important correspondence between voltage assignments for a
graph G and representations of its fundamental group. Since the net voltage
on any two equivalent u-based closed walks is the same, any voltage assign-
ment for G in a group .J/1 induces a homomorphism '1T1(G, u) -+ .J/1. That is,
every element of '1T1(G, u) is mapped onto the net voltage assigned to any
u-based closed walk representing that element.
Conversely, every homomorphism '1T 1(G, u) -+ S/1 is induced by some volt-
age assignment in S/1 on the graph G. Given such a homomorphism, first
choose a spanning tree T and assign to every edge of T the identity voltage.
For any edge of G not in T, suppose that its initial point is v and that its
terminal vertex is w. Let W be the u-based closed walk that begins with the
unique path in T from u to v, next traverses the edge e, and then concludes
with the unique path in T that runs from w to u. Assign as a voltage to the
edge e the image in the group S/1 of the equivalence class in '1T 1(G, u) of the
walk W.
From the correspondence just described, one may obtain for graphs all of
the standard topological theorems on the relationship between fundamental
groups and covering spaces.

2.5.4. Contracting Derived Graphs onto Cayley Graphs

The technique of selecting a spanning tree in a voltage graph and using it to
obtain an equivalent voltage assignment is useful for other purposes, in
addition to counting components of derived graph. The following theorem is
equivalent to Lemma 3 of Babai (1977).
Subgroup<; of the Voltage Group 93

Theorem 2.5.5. Let G be a graph with a voltage assignment a in a group d

such that the derived graph Ga is connected. Then Ga contracts onto a Cayley
graph ford.

Proof Choose a spanning tree T for the base graph G. By Theorem 2.5.4,
the derived graph GaT corresponding to the T-voltages is isomorphic to Ga, so
it is sufficient to prove that GaT contracts onto a Cayley graph for d. Since
ar(T) = 0, by Theorem 2.1.3, the preimage p- 1(T) of T under the natural
projection has #d components, each mapped isomorphically by p onto the
spanning tree T. In particular, for each element d of the voltage group d,
one component of p -I(T) contains all of the vertices of GaT with the
subscript a. If each component of p- 1(T) is contracted onto a single vertex,
which is identified with the common subscript of all the vertices in that
component, then the resulting graph is a Cayley graph for d. One might
observe that one can also contract the base graph G along the spanning tree T,
thereby obtaining a bouquet of circles. The voltages on these circles generate
d, because GaT is connected. Then the contracted derived graph is precisely
the derived graph for the contracted voltage graph. 0

Corollary 1 (Babai, 1977). Let !!4 be a subgroup of a group d, and let G be a

Cayley graph for d. Then G contracts onto some Cayley graph for !!4.

Proof Since the group d acts freely on the Cayley graph G, so does its
subgroup !!4. It follows that G may be derived by assigning voltages in !!4 to
the quotient of G under the action of !!4. By the theorem, it follows that G
contracts onto a Cayley graph for !!4. o

The "genus of a group" is defined to be the least genus occurring among all
its Cayley graphs. White (1973, p. 80) asked whether a finite group d can have
a subgroup !!4 whose genus exceeds the genus of d. It is an immediate
consequence of Corollary 1 that it cannot.

Corollary 2 (Babai, 1977). The genus of a finite group is greater than or equal
to the genus of each of its subgroups. 0

2.5.5. Exercises
Exercises 1-4 concern the voltage graphs in Figure 2.22.
1. Calculate the local group at the vertex u of the voltage graph in Figure
2.22a. How many components are there in the derived graph?
2. Calculate the local group at the vertex u of the voltage graph in Figure
2.22b. How many components are there in the derived graph?
3. Calculate the local groups at vertices u and v of the voltage graph in
Figure 2.22c. How many components are there in the derived graph?
94 Voltage Graphs and Covering Spaces

9 u

7 (2,3)~(4,2)
(b)

w
2

(a)

(12)

(d)

(c)

Figure 2.22. Four voltage graphs. (a) Ordinary voltages in .;2"18 • (b) ordinary voltages in
(c) ordinary voltages in 94. (d) permutation voltages in .5/'7.
.;2"6 X .;2"6 •

4. Calculate a generating set for the local group at the vertex u of the voltage
graph in Figure 2.22d. How many components does the derived graph
have?
5. Prove that any graph map f: G -+ H induces a homomorphism
J.: '11'1(G, u)-+ (H, f(u)), defined as follows: If [w] is the equivalence
class of any closed u-based walkinG, then J.([w]) = [f(w)].
6. Prove that if p: G-+ H is a covering projection, then p.: '11'1(G, u)-+
'11'1(H, p(u)) is one-to-one.
7. Let !!4 be a subgroup of finite index in '11'1(G, u). Prove that there is a
covering p: G-+ G and a vertex u E p- 1(u) such that p.('11'1(G, u)) = !!4.
Prove that G is unique up to graph isomorphism. (Thus, G can be called
the covering space of G corresponding to the subgroup !14.)
3
Surfaces and Graph lmbeddings
It is a classical theorem of Rado (1925) that any compact surface can be
dissected into a finite number of triangular regions. The incidence structure of
these regions determines a combinatorial object called a "complex", which is a
higher-dimensional analog of a graph. Whereas a graph has only zero-dimen-
sional "cells" called vertices and one-dimensional "cells" called edges. a
complex may include pieces of any dimension. In particular. a 2-complex also
contains some two-dimensional cells. Often these cells correspond to triangu-
lar regions, but more generally they may be arbitrary polygons, with any
number of sides. Knowing that any closed surface can be represented by such
a finite combinatorial object enables us to classify the closed surfaces, that is.
to construct a countable list that includes them all and contains no repetitions.
The combinatorial viewpoint also enables us to decide which graphs can be
imbedded in which surfaces. Even more significantly, it enables us to construct
all of the essentially different possibilities for imbedding an arbitrary graph in
an arbitrary surface.

3.1. SURFACES AND SIMPUOAL COMPLEXES

A topological space M is called an "n-manifold" if M is Hausdorff and is

covered by countably many open sets, each one of which is homeomorphic
either to the n-dimensional open ball

or the n-dimensional half-ball

Thus every point in an n-manifold appears, at least locally, to be sitting in the

middle of n-space or at the edge of half of n-space. [The need for the
Hausdorff condition is explained in Munkres (1975).] The "boundary" of an
n-manifold M is the collection of all points in M that do not have neighbor-
hoods homeomorphic to the n-dimensional open ball. For example, the

95
96 Surface~ and Graph lmbeddinll"

annulus {(x. y)ll ~ x 2 + y 2 ~ 4} is a 2-manifold whose boundary is the pair

of circles of radius 1 and 2 centered at the origin. A manifold is "dosed" if it
is compact and its boundary is empty. By a "surface" we usually mean a
closed, connected 2-manifold. such as the sphere, the torus, or the Klein
bottle. When we are considering a surface with boundary. such as the Mobius
band. or a noncompact surface such as the plane. it is dear from the context
or explicitly stated.

3.1.1. Geometric simplicial complexes

A "(geometric) k-simplex" is the convex hull of k + 1 affinely independent
points or "vertices" in Euclidean n-space R 11 • For example, the triangular
region determined by three noncollinear points in R 2 or R J is a 2-simplex and
the line segment determined by two distinct points in R 11 is a 1-simplex. Let
( v0 • .••• v A) denote the k-simplex determined by the vertices v0 • •••• vA. Then
the simplex determined by any subset of { l'o• ...• vA} is called a "face" of the
simplex (v0, ...• vA)· Thus the faces of the 1-simplex (v 0, v1) are the empty
set, the endpoints (v0 ) and (v 1). and the 1-simplex (vu• v1) itself.
A (geometric) "simplicial complex" K is a finite collection of simplexes in
R 11 satisfying these two conditions:

1. Every face of every simplex in K is a simplex in K.

2. The intersection of any two simplexes in K is a simplex in K.

The point set

IKI= US
.\'E/\"

is called the "carrier" of the simplicial complex K. It is possible that the same
point set in R 11 can be the carrier of more than one simplicial complex, as
illustrated in Example 3.1.1. If m is the largest integer such that K contains an
m-simplex, then K is an "m-complex". The collection of all k-simplexes of K
for k ~ r is called the "r-skeleton" of K and is denoted Kul.

Example 3.1.1. Figure 3.1 shows three simplicial 2-complexes in R ·1• The
complex on the left has one 2-simplex: (v. w, x): nine 1-simplexes: (u. t•).
(u. w). (u. x). (v. w). (v. x). (v. y). (w. x). (w. y). (x. y): and .ftt•e
0-simplexes: (u), (v), (w), (x). (y). The center 2-complex has two 2-sim-
plexes: (u. v, w), (u. w, x); nine }-simplexes: (u. v). (u. w). (u. x). (u. y).
(v. w). (v. z). (w. x), (w, y), (x, z); and six 0-simplexes: (u). (v). (w).
(x). (y). (z). Note that the right-hand 2-complex has the same carrier as the
one on the left. but this time the 2-simplex ( v, w. x) is replaced by the three
2-simplexes (v. w, z). (w. x, z). (x. v, z) and their faces. In each instance. the
1-skeleton is the graph that remains when the 2-simplexes are discarded. (One
discards only the 2-simplexes, not their various faces.)
Surf:tl'l'' mul Simplid:tl ( 'ompll'U'' 97

II II

II'

z
\' y

3. 1.2. Abstract Simplicial Complexes

An ohvious comhinatorial structure can he refined from a simplicial complex.
Define an "ahstract simplicial complex·· ( V. C) to he a finite set V together
with a collection C of suhsets of V whose union is V and such that every
suhsct of a mcmhcr of C is itself a mcmher of C. Then with any (geometric)
simplicial complex K. one associates the ahstract simplicial complex ( K 0 • C)
where a suhset of the 0-skeleton K 0 is in the collection C if and only if the
suhset forms the vertex set of a simplex in K. In fact. every ahstract simplicial
complex ( V. C) can he associated in this way with some (geometric) simplicial
complex K: if the set V has 11 elements. simply let K 0 he 11 aflinely
independent points in IR" 1• One may ohscrve that an ahstract simplicial
complex is a special kind of" hypergraph".
It should not he surprising that the topological type of the carrier of a
simplicial complex is determined hy the associated ahstract simplicial com-
plex. To he precise. define an isomorphism of ahstract simplicial complexes
(V.C) and (V',C') to he a hijection f: V--+ V' such that /(s) E C' if and
only if s E C. Then any isomorphism f of the ahstract complexes associated
with the geometric complexes K and K' extends to a homeomorphism of their
respective carriers IK I and IK 'I- (The homeomorphism is defined one simplex
at a time. starting with the 0-skeleton. then extending to the ]-skeleton. next
the 2-skeleton. and so on. The crucial ohservation is that any homeomorphism
from the houndary of one 11-simplex to the houndary of another extends. say
radially. to a homeomorphism of the interiors.)
Is the converse true? Namely. if the carriers IKI and IK'I are homeomorphic
as topological spaces. must the associated ahstract simplicial complexes he
isomorphic? The answer is clearly no. since one can always suhdivide a
simplex into smaller simplices as in Example 3.1.1. But suppose one is allowed
to suhdivide K and K': if IKI and IK'I are homeomorphic. are there
suhdivisions I. and L' of K and K'. respectively. such that L and L' arc
isomorphic? This fundamental question is called the "Hauptvermutung". It
says. in cflcct. that if a topological space has the structure of a simplicial
98 Surface' and Graph Imbedding'

complex, then that structure is unique up to subdivision. The subdivision in

question can be restricted to be barycentric (see the next section) and can be
interpreted in purely combinatorial terms; thus the topology of a simplicial
complex can be faithfully combinatorialized, if the Hauptvermutung holds.
Unfortunately, it does not hold for all complexes: Milnor (1961) gives pairs of
n-dimensional complexes for each n ~ 6, that are homeomorphic but combi-
natorially distinct. On the other hand, for low-dimensional complexes, the
Hauptvermutung does hold. Papakyriakopoulos (1943) proves it for 2-com-
plexes, and Brown (1969) for 3-complexes.

3. 1.3. Triangulations
A "triangulation" of a topological space X is a homeomorphism h from the
carrier of some simplicial complex K to the space X. The image of a simplex
of K under h is called a simplex of the triangulation. The aforementioned
theorem of Rado states that every compact surface has a triangulation.

Example 3. 1.2. On the left of Figure 3.2 is a triangulation of a torus, where the
sides of the rectangle are to he identified in pairs, as in Chapter 1, so that vertices
nwtch up. Enen though all its faces are three-sided, the right side of Figure 3.2 is
not a triangulation (~(a torus since no two different simplexes of a triangulation
ha11e the exact same vertices. A /so an n-simplex of a triangulation must have
n + 1 distinct vertices, and this "triangulation" has only one vertex.

The vertices and edges of a graph in Ill 3 are not the 0-simplexes and
]-simplexes of a triangulation if the graph has any multiple edges or loops.
This can always be rectified, of course, by a subdivision of the graph inserting
one vertex on each multiple edge and two vertices on each loop.
The process of subdividing self-loops and multiple edges to make a nonsim-
plicial graph simplicial extends to higher dimensions. We might think of a
loop as an "inadequately subdivided" edge, and we might also think of a
nontriangular polygon (or a polygon with repeated vertices or edges) as

u 11 w u
11 u

X X

y y

11 11

u II w u
Figure 3.2. A triangulation and a non-triangulation or a torus.
Surfaces and Simplicial Complexes 99
1\
w w

u e u .
u
,.
e v
Figure 3.3. The first barycentric subdivision of a 2-simplex.

inadequately subdivided. The higher-dimensional analog of a nonsimplicial

complex is called a "cell complex", and just as the operation of subdivision
can make a nonsimplicial graph simplicial, it can also do the same for cell
complexes of arbitrarily high dimension. We shall first define a particular kind
of subdivision operation as it applies to simplicial complexes and then extend
the concept.
Let s = (v0 , •.• , vk) be a k-simplex in 111n. The point s = (v0 + v1
+ · · · +vd/k is the "barycenter" of s (addition here is usual vector addition
in Ill n ). The "first barycentric subdivision" of a simplicial complex K is the
collection of all simplexes of the form (s 0 , s 1 , ... , sm) where each s; is a
simplex of K and s; is a face of s;+I· Figure 3.3 shows the first barycentric
subdivision of the 2-complex consisting of a 2-simplex and its faces.
An s-sided polygonal region, where s ;;:::; 4, is an example of a nonsimplicial
two-dimensional cell. If the region is parametrized by a convex s-gon, then
there is a "center of gravity". We can make a barycentric subdivision by first
subdividing every edge on the boundary of the s-gon and then drawing an
edge from the center of gravity to each vertex, old and new, on the boundary.
Even a one-sided or two-sided region can be convex, if its sides are curved, so
there is little problem in extending the concept of barycentric subdivision to
all 2-complexes. The same process is also applicable to polyhedra of arbitrarily
high dimension.
The operation of barycentric subdivision may be iterated. We state, without
proof, the standard fact of combinatorial topology that the second barycentric
subdivsion of any cell complex is a simplicial complex. One subdivision might
not be enough, even in dimension 1, since the first barycentric subdivision of a
self-loop yields two edges between the same two vertices.
The definitions given here can be generalized to allow infinitely many
vertices in a simplicial complex. However, the topological considerations are
more subtle than the reader might expect, even if one demands that every
vertex lie on finitely many simplexes (see Exercise 6). As our interest is now in
finite graphs, we shall not pursue the matter further.
100 Surfaces and Graph lmbeddings

3.1.4. Cellular imbeddings

Although a graph is defined as a purely combinatorial object in Chapter 1, it is
now helpful to augment the definition by including a fixed topological repre-
sentation of the graph. For a simplicial graph, this might be the carrier of a
geometric 1-complex. This permits a more fluent discussion. Moreover, we
often abuse the terminology altogether by referring to a topological repre-
sentative as the "graph", particularly when we discuss an imbedding of the
graph.
An "imbedding" of a graph G in the surface S is a continuous one-to-one
function i: G -+ S. In nearly all cases, the graph G is assumed to be a subset of
the surface S, and the function i: G -+ S is the inclusion map. The imbedding
is then denoted simply G -+ S. It can be shown as a generalization of the
Schoenfliess theorem that for any imbedding G -+ S, the graph G is contained
in the 1-skeleton of a triangulation of the surfaceS. Thus graph imbeddings in
surfaces are "tame" and can be analyzed by combinatorial methods. [The
Schoenfliess theorem, a strengthening of the Jordan curve theorem, states that
any simple closed curve in the plane can be carried onto the unit circle by
some homeomorphism of the plane; for a proof, see Newman (1954). The
Alexander horned sphere (1924) shows that the Schoenfliess theorem does not
hold for spheres in three-dimensional space.]
Given an imbedding G -+ S, the components of the complement S - G are
called "regions". If each region is homeomorphic to an open disk, the
imbedding is said to be a "2-cell (or cellular) imbedding" and the regions are
also called "faces" of the imbedding. The closure in the surface S of a region
in the 2-cell imbedding G -+ S need not be homeomorphic to the closed disk.

Example 3.1.3. The imbedding of a bouquet of two circles in a torus on the left
of Figure 3.4 is not cellular, since one region is homeomorphic to an open
annulus. The imbedding on the right is cellular. Observe that the closure in the
surface of the one and only face is not a closed disk, but instead, the whole torus.

Unless we explicitly say otherwise, any imbedding in this book is assumed

to be a 2-cell imbedding. In particular, we shall not usually imbed a discon-
nected graph in a connected surface or imbed a tree in any surface other than
the sphere. Also, an imbedding of a nonempty finite graph in the plane cannot

Figure 3.4. Two imbeddings of the bouquet 8 2 on a torus.

Surfaces and Simplicial Complexes 101

be cellular. Thus, when an imbedding in the sphere is drawn in the plane, the
reader is expected to remember the necessary point at infinity.
Given a 2-cell imbedding G ~ S, the collection of all edges and vertices
lying in the closure of a region can be organized into a closed walk in the
graph G by traversing a simple closed curve just inside the region. This closed
walk is unique up to the choice of initial vertex and direction, and is called the
"boundary walk" of the region. A face of a 2-cell imbedding can usually be
specified unambiguously simply by giving its boundary walk (see Exercise 17).
For this reason, we shall frequently abuse the language with phrases such as
"the face abc d " where abc d is in fact the boundary walk of the face. An
occurrence of an edge in the boundary walk of a region is called a "side" of
the region. The "size" of a region is the length of its boundary walk. Since
individual edges may occur twice in a boundary walk, the size of a region can
be greater than the number of edges that are sides of the region. An n-sided
face in a 2-cell imbedding is called an "n-gon", although 3-gons and 4-gons
are also called triangles and quadrilaterals, respectively.

Example 3.1.4. The imbedding in the sphere given in Figure 3.5 has three
regions. There is a quadrilateral abc d, a triangle e e f, and a 9-gon
abggcdhfh. The 9-gon has only seven different edges as its sides.

Two imbeddings of a given graph i: G ~ S and j: G ~ T are called

"equivalent" if there is an orientation-preserving homeomorphism h: S ~ T
such that hi = j. It is important to observe that by this definition, it does not
suffice simply to find a homeomorphism from S to T that takes the image
i(G) to the image j(G); for the imbeddings to be equivalent, the homeomor-
phism must also respect a labeling of the vertices and edges of G.

Example 3.1.5. Four different imbeddings of the same tree in the sphere
(plane) are given in Figure 3.6. A rotation by 90° gives an equivalence between
the first and second imbeddings. A reflection in a horizontal line through the
vertex y yields an equivalence of the second and third imbeddings. However, the
fourth imbedding is not equivalent to any of the previous three.

Two imbeddings i: G ~ S and j: G ~ T are called "weakly equivalent" if

there is a homeomorphism h: S ~ T such that h(i(G)) = j(G). For instance,

Figure 3.5. An imbedding in the sphere.

102 Surfaces and Graph lmbeddings

·xx xx
X W W

Figure 3.6.
U X U

Four imbeddings of the same tree.

X W

the fourth imbedding in Figure 3.6 is weakly equivalent to the other three. It
will become apparent in Section 3.2 that equivalence is easier to analyze than
weak equivalence, although it may seem at first less natural. In any case, it
must be remembered that any discussion of equivalence here presumes that the
vertices and edges of the imbedded graph have been labeled. "Congruent" is a
possible synonym for "weakly equivalent".
Given a 2-cell imbedding G ~ S, the "dual imbedding" G* ~Sis defined
as in Chapter 1: the vertices of G* correspond to regions of the imbedding
G ~ S and vice versa. The dual graph gives an efficient description of the
incidence relationship among the regions of an imbedding, for which purpose
it is used extensively in Section 3.3.

3.1.5. Representing Surfaces by Polygons

One way to present a surface with a given triangulation is simply to list (or
draw) every triangle with each of its sides directed and labeled. The surface is
then assembled by identifying the two sides that have the same label so that
the directions agree. The homeomorphism doing the identifying for a particu-
lar edge does not matter, since any two triangulated surfaces obtained from
the same labels and directions have the same underlying abstract simplicial
complex and hence are homeomorphic. If desired, one may "preassemble"
some of the triangles into larger polygons whose labeled sides are to be
identified later to obtain the given surface. In fact, such preassembly permits
one to represent any compact connected surface by a single polygon with
labeled and directed sides (similarly to what was done in Chapter 1 for the
torus). A label that appears only once corresponds to an edge in the boundary
of the surface.

Example 3.1.6. In Chapter 1 the projective plane was defined to be the surface
obtained by identifying the boundary of a disk to the boundary of a Mobius band.
In Figure 3.7, polygon 2 becomes a Mobius band when the h-edges are
identified. Thus the hexagon formed by polygons 1, 2, and 3 represents the
projective plane. If the consecutive sides a, b, care considered to be just one edge,
we get the simpler representation on the right of Figure 3.7. In particular, the
projective plane can be viewed as the closed unit disk with antipodal points on the
unit circle identified in pairs. Thus, the definition given here for a projective plane
Surfaces and Simplicial Complexes 103

b 2 b e

Figure 3.7. Two representations of the projective plane.

corresponds to that given in geometry as the space of lines through the origin in
R 3 (represented by points on the northern hemisphere of the unit sphere with
antipodal points on the equator identified).

Example 3.1.7. The Klein bottle was defined in Chapter 1 as the surface
obtained by identifying the boundaries of two Mobius bands to the boundary of a
sphere with two holes. Since a sphere with two holes is homeomorphic to an
annulus and identifying one boundary component of an annulus to a Mobius band
still/eaves a Mobius band (with a "collar" on its boundary), the Klein bottle can
also be viewed as the surface obtained by identifying the boundaries of two Mobius
bands alone. In Figure 3.8, polygon 2 becomes a Mobius band when the h-edges
are identified. Polygons 1 and 3 also form a single Mobius band when their edges
are identified (first identify the d-edges to get a rectangle like polygon 2). Thus
the polygons I, 2, and 3 together represent the Klein bottle. On the right in
Figure 3.8 is a simpler representation.

If the direction of one copy of edge e in Figure 3.7 is reversed, the resulting
surface is the sphere. Similarly, if the direction of one copy of edge e in Figure
3.8 is reversed, the resulting surface is the torus.

Example 3.1.8. If all the identifications are performed on the two polygons in
Figure 3.9 except on the sides labeled e, one obtains two tori each with a
puncture. When thee-sides are then identified, the result is a surface of genus 2.

:
a c d
1

c
2

3
d
b

a ·I d
I·
Figure3.8. Two representations of the Klein bottle.
104 Surfaces and Graph Imbeddings

(l·D -e e - 88 Figure 3.9. How to obtain a genus two surface from polygons.

a b c d a b c d

g 3 4 g
f
f f f
1 2

-
e
e e
d a
b c d a b c
Figure 3.10. Three views of the same toroidal imbedding of K 4 , 4 minus a 1-factor.

If the e-sides were identified first, a representation of a genus-2 surface by a

single eight-sided polygon would be obtained.

A picture of a graph imbedding can be drawn on a polygon with identified

sides, as in Chapter 1. When a graph edge hits a side it continues at the other
identified side. The sides of the so-called outer polygon are sometimes allowed
to be edges of the graph also.

Example 3.1.9. An imbedding in a torus of K 4, 4 with a 1-factor removed is

given on the left in Figure 3.10. The sides of the "outer" rectangle are not edges
of the graph. The four six-sided regions of this imbedding (labeled 1, 2, 3, 4) can
be reassembled to give the middle imbedding, in which the outside edges are to be
identified in pairs as usual. This time the "outer" polygon's sides are edges of the
graph. On the right is the same imbedding with the path e f g straightened.

3.1.6. Pseudosurfaces and Block Designs

A "pseudosurface" is the carrier of a 2-complex in which every 1-simplex is a
face of exactly two 2-simplexes. Every point in a pseudosurface except for the
underlying points of some vertices has a neighborhood homeomorphic to the
interior of the unit disk. A neighborhood of the underlying point of a vertex,
however, is in general homeomorphic to a finite number m of disks all of
whose centers have been identified to the single vertex. If m > 1 then the point
is called a "singular point of the pseudosurface". Thus a pseudosurface can be
Surfaces and Simplicial Complexes lOS

Figure 3.11. A pseudosurface with its singular points darkened.

viewed as a surface in which finite sets of points are identified. Figure 3.11
shows a typical pseudosurface.
Alpert (1975) showed that covering space techniques could be useful in
constructing triangulated pseudosurfaces that represent "twofold triple sys-
tems", thereby reviving interest in the connection between block design and
topological graph theory explored earlier by Heffler (1897) and Emch (1929).
Subsequent advances in exploiting this connection were achieved by White
(1978) and Garman (1979). The basic ideas are easy enough to explain.
A "block design" consists of a set V of objects and a set B of subsets of V,
called "blocks", such that (i) each object appears in the same number of
blocks, called the "replication number", and (ii) each block has the same
number of objects, called the "blocksize". If every pair of objects appears in
the same number of blocks, called the "balancing index", then the design is
called "balanced". A trivial kind of balanced block design has every object in
every block, and is called a "complete" block design. If the blocksize is less
than # V then the design is called "incomplete".
Balanced incomplete block designs, abbreviated BIBDs, are important to
statisticians because they facilitate the design of experiments on samples of
limited size. They are also of considerable interest to algebraists and combina-
torial analysts.
A twofold triple system is a BIBD with blocksize equal to 3 and balancing
index equal to 2. Alpert proved that there is a bijective correspondence
between the twofold triple systems and the (equivalence classes of) triangula-
tion of closed pseudosurfaces by complete graphs. This topological insight
facilitates an easy proof that some classical twofold triple systems of
Bhattacharya (1943) and Hanani (1961) are distinct and that they are also
distinct from new twofold triple systems that Alpert developed from the
Ringel-Youngs constructions in cases 1, 6, 9, and 10 of the Heawood
map-coloring problem.
White and Garman both explored the extension to "partially" balanced
incomplete block designs with two association classes of objects. These associ-
ation classes partition the objects so that pairs of objects from one class have
one balancing index while pairs from the other class have another balancing
index. To obtain his results, Garman formally extended the imbedded voltage
graph construction described in Chapter 4 to voltage graphs imbedded in
pseudosurfaces.
106 Surfaces and Graph Imbeddings

Figure 3.12. An orientation for an annulus and an attempt at an orientation for a Mobius band.

3.1.7. Orientations
Each face of an imbedding G ~ S has two possible directions for its boundary
walk. A face is assigned an "orientation" by choosing one of these two
directions. An "orientation for the imbedding" G ~ S is an assignment of
orientations to all the faces so that adjacent regions induce opposite directions
on every common edge. If the graph G is the 1-skeleton of a triangulation of
the surface S, then an orientation for the imbedding G ~ S is called an
"orientation of the triangulation". A surface is "orientable" if it has an
imbedding with an orientation.
It is a simple matter to test whether a given connected imbedding G ~ S
has an orientation. First, assign an arbitrary orientation to a particular face.
This forces an orientation of each face that shares an edge with the original
face. Since the surface is connected, the process can be continued until the
orientation of each face has been forced. Either the result is an orientation for
the imbedding or else the given imbedding has no orientation.

Example 3.1.10. An orientation for a triangulation of an annulus is given on the

left in Figure 3.12. On the right is an attempt at orienting a Mobius band. Since
the attempt fails, this triangulation has no orientation. Indeed, it is clear that no
sequence of adjacent 2-simplexes along the central horizontal line of a Mobius
band can be oriented to achieve an orientation for the surface. Thus a Mobius
band is not orientable.

Example 3.1.10 is, in fact, quite general. If a surface contains a subset

homeomorphic to a Mobius band, it cannot be orientable. Conversely, suppose
a triangulation of a surface has no orientation. Then starting from an
orientation of a fixed 2-simplex, a chain of adjacent 2-simplexes can be formed
that is homeomorphic to a Mobius band.

3.1.8. Stars, Links, and Local Properties

Two important types of subcomplexes enable us to define "local" properties of
a graph. Let us suppose that K is a simplicial complex of arbitrary dimension.
The "star" of a vertex v is defined to be minimal subcomplex of K that
contains every simplex of K that is incident on v. In particular, if K is a
graph, then for any vertex v, star( v) is the subgraph consisting of all the edges
Surfaces and Simplicial Complexes 107

incident on v and the endpoints of those edges. Such a graph does actually
look something like a heavenly star, which is the inspiration for the terminol-
ogy. A topological space X is a closed surface if and only if there is a
triangulation of X in which the carrier of the star of each vertex is homeomor-
phic to a closed disk.
The "link" of a vertex v is defined to be the maximal subcomplex of star( v)
that does not contain v itself. If K is a graph, then link( v) is the totally
disconnected graph comprising the neighbors of v and no edges. If the carrier
of K is a surface, then the link of every vertex is a cycle through its neighbors.
If the carrier of K is a pseudosurface, then the link of a singular vertex is the
disjoint union of two or more cycles.
The standard definition of a "local property" of a complex K is to say it is
a property that is true of the star or the link of every vertex. In the case of
graphs, we sometimes associate it with the subgraph induced by the link of a
vertex.
For instance, we might say that a graph is "locally hamiltonian" if for every
vertex v, the subgraph induced by link(v) is hamiltonian. The principal of this
concept is that a graph cannot triangulate a surface unless it is locally
hamiltonian.

3.1.9. Exercises
1. Which points in the given set X fail to have a neighborhood homeomor-
phic to the unit disk?
(a) X= {(x, y, z) I xyz = 0}.
(b) X= {(x, y, z) lx 2 + y 2 = z 2 }.
2. Prove that the r-skeleton of a simplicial complex is a simplicial complex.
3. Figure 3.3 shows the first barycentric subdivision of the 2-complex
consisting of a 2-simplex and its faces. Prove that the first barycentric
subdivision is itself a simplicial complex.
4. Draw imbeddings in the plane of the 1-skeleton of the first two 2-com-
plexes in Figure 3.1. Show that the 1-skeleton of the first barycentric
subdivisions of these 2-complexes do not have planar imbeddings. (Hint:
Show that the first contains K 3, 3 and that the second contains K 5 .)
5. Find a 2-complex that cannot be imbedded in the sphere (or for that
matter in any surface) but whose first barycentric subdivision has a
1-skeleton that is imbeddable in the plane.
6. Let vn = (cos2'1Tjn, sin2'1T/n) E 111 2 for n = 1,2, .... Consider the col-
lection K of all simplexes of the form (vn, vn+l) or (vn) for n = 1, 2, ....
If the restrictions of finiteness were deleted from the definition of
simplicial complex, would K be a simplicial complex? Could K triangu-
late the unit circle?
108 Surfaces and Graph Imbeddings

7. Draw a triangulation of the projective plane, trying to use as few

triangles as possible.
8. The directions and edge labels for a polygonal representation of a surface
can be coded simply by giving a closed walk around the boundary of
each polygon. Thus the "outer" polygon on the left of Figure 3.8 could
be coded as d c b a d- 1 c b a. The projective plane polygon on the right of
Figure 3.7 would be ee. Thus to any list of "words" using edge labels
and their inverses, such that each edge label is used exactly twice, we can
associate a surface. What surface does the list { a b a c- \ b c} represent?
What surface does { a b a - 1 c - 1, b c} represent?
9. It should be clear that cyclically permuting or inverting any word coding
a polygon (see Exercise 8) corresponds only to choosing a different initial
edge or direction for the closed walk around the polygon and hence does
not change the surface being represented. Explain why the following
operations do not change the surface either.
(a) Replace the two words A e, e- 1 B by the single word A B (a picture
will suffice).
(b) Replace the word ABA' C by the word x B x• C where t: = ± 1.
(c) Replace A ee- 1 B by A B.
10. Use the operations described in Exercise 9 to find out what surface the
list {a b a- 1 cd, c- 1 er\ b- 1 e- 1 fd- 1 } represents.
11. How does one recognize when a surface represented by one word
(Exercise 8) is nonorientable? Suggest an algorithm for determining
whether a connected surface represented by a list of more than one word
is nonorientable.
12. Prove that equivalence and free equivalence of imbeddings are equiv-
alence relations.
13. Find an imbedding of the tree in Example 3.1.5 that is not equivalent to
any of the imbeddings shown in Figure 3.6.
14. Prove that any two imbeddings of a tree having largest valence 3 are
weakly equivalent. Give an example of a tree having two imbeddings that
are not freely equivalent.
15. Draw an imbedding of the complete graph K 5 on the torus as a polygon
whose sides are edges of K 5 (see Example 3.1.9). Do the same for K 6 •
16. Draw a 2-cell imbedding of K 5 in a polygonal representation of a surface
of genus 2 (see Example 3.1.8). Be sure the imbedding is cellular. How
many faces does it have?
17. Prove that if two regions of an imbedding G ~ S have the same
boundary walk, then the graph G is a cycle.
18. Give an imbedding of K 5 in the pseudosurface obtained by identifying
two points on the sphere. Prove that K 6 cannot be imbedded in this
pseudosurface.
Band Decompositions and Graph Imbeddings 109

3.2. BAND DECOMPOSITIONS AND GRAPH IMBEDDINGS

When a surface is formed as in Section 3.1 by pasting polygons to each other,

the vertices and edges of the polygons fit together to form a graph imbedded
in that surface. Alternatively, one might begin with a graph and attach
polygons to it so that a surface is formed. This can be achieved, as we see in
this section, without explicit mention of the polygons. Instead, we begin with a
graph and add some combinatorial structure that implicitly describes the
polygons, the surface to be obtained, and the imbedding.

3.2.1. Band Decomposition for Surfaces

Let G -+ S be a 2-cell imbedding of a graph in a surface. One can surround
each vertex of the graph G by a small disk in the surface S and each edge of G
by a thin band so that the union of all disks and bands is a neighborhood of G
in S whose shape preserves that of the graph itself. The complement in S of
this neighborhood consists of a family of disks, one just inside each face of the
imbedding. This decomposition of the surface into bands and (two kinds of)
disks is now described in detail.
Define a "1-band" to be a topological space b together with a homeo-
morphism h: I X I-+ b, where I denotes the unit interval [0, 1). The arcs
h(I X {j}) for j = 0, 1 are called the "ends" of the band b, and the arcs
h( {j} X /), for j = 0, 1, are called the "sides" of b. A "0-band" is simply a
homeomorph of the unit disk, as is a "2-band".
A "band decomposition" of the surface S is a collection B of 0-bands,
1-bands, and 2-bands satisfying these conditions:

1. Different bands intersect only along arcs in their boundaries.

2. The union of all the bands is S.
3. Each end of each 1-band is contained in a 0-band.
4. Each side of each 1-band is contained in a 2-band.
5. The 0-bands are pairwise disjoint and the 2-bands are pairwise disjoint.

The corresponding "reduced band decomposition" B omits the 2-bands.

Example 3.2.1. Figure 3.13 shows a graph imbedding and the associated
reduced band decomposition. The 0-bands assume the names of their respective
vertices and the 1-bands assume the names of their respective edges.

A band decomposition of a surface is the two-dimensional version of the

topological construction known as a "handle decomposition" of an n-mani-
fold. The manipulation of bands as used in this chapter to classify the surfaces
has higher-dimensional analogs that were instrumental in Smale's proof (1961)
of the Poincare conjecture in higher dimensions [see also Milnor (1965a)]. Use
110 Surfaces and Graph lmbeddings

- a

r----- -----...,
I I

bl I I t•
- a
Figure 3.13. (left) An imbedding of a graph with two vertices and three edges in the Klein bottle.
(right) The corresponding reduced band decomposition.

of band decompositions in the study of surfaces can be found implicitly in

Griffiths (1976). For the more traditional approach of sewing together labeled
polygons, see Blackett (1967).

3.2.2. Orientability
A band decomposition is called "locally oriented" if each 0-band is assigned
an orientation. Then a 1-band is called "orientation-preserving" if the direc-
tions induced on its ends by adjoining 0-bands are the same as those induced
by one of the two possible orientations of the 1-band; otherwise the 1-band is
called "orientation-reversing". These two possibilities are illustrated in Figure
3.14.
An edge e in the graph imbedding associated with a locally oriented band
decomposition is said to have "(orientation) type 0" if its corresponding
1-band is orientation preserving, and "(orientation) type 1" otherwise. A walk
in the associated graph has type 1 if it has an odd number of type-1 edges and
has type 0 otherwise.

Remark. The designation of orientation reversing for a 1-band depends only on

the orientations chosen for its adjoining 0-bands and does not necessarily imply
nonorientabi/ity of the associated surface. Rather, the surface is nonorientab/e if

Figure 3.14. Two orientation preserving bands (left) and two orientation reversing bands (right).
Band Decompositions and Graph Imbeddings Ill

and only if there exists a type-1 cycle in the associated graph, which implies
immediately that the surface contains a Mobius band. For instance, the cycle
e 2 , e 3 in Figure 3.13 has type 1, no matter what the choice of local orientations.

In searching for type-1 cycles, it is often advantageous first to reverse the

orientation of some 0-bands. This obviously does not change the surface.
However, when the orientation of a 0-band is reversed, this changes the
orientation type of every proper edge incident on the corresponding vertex
(the orientation type of a loop does not change).

Example 3.2.2. Consider the reduced band decomposition shown on the left in
Figure 3.15. First, reverse the orientation of the 0-band labeled u. Then reverse
the orientation of the 0-band v. The resulting decomposition has no orientation
reversing 1-bands, so the underlying surface is orientable, despite all the twisted
!-bands in the leftmost drawing.

Orientability Algorithm. To determine whether the surface S corresponding to a

given band decomposition is orientable or not, first choose a spanning tree T for
the associated graph G. Next choose a root vertex u for T and an orientation for
the 0-band of u. Then for each vertex v adjacent to u in the tree T, choose the
orientation for the 0-band of v so that the edge of T from u to v has type 0.
Continue this process on the tree T until every 0-band has an orientation. That is,
if v and w are adjacent in T and if the orientation at v has already been
determined, then choose whichever orientation at w makes the edge from v to w
have type 0. No conflicts in choice of orientation arise since Tis a tree. The result
is a band decomposition such that every edge in the tree T has type 0. Any type-!
edge in G- T forms a closed walk with a path in T joining its endpoints. Since
every walk in T has type 0, such a closed walk would have type 1. Thus the
surface S is nonorientable if and only if some edge in the complement G - T has
type 1.

Figure 3.15. Reversing the orientation of the O·bands u and v.

112 Surfaces and Graph Imbeddings

3.2.3. Rotation Systems

We wish to develop a convenient method of describing a band decomposition
or, equivalently, a graph imbedding G ~ S. One first observes that the
2-bands are not really needed to define the surface S, since the union of the
0-bands and 1-bands is a surface with boundary, and since there is essentially
only one way to fill in the faces to complete the closed surface. The second
part of this explanation follows from Lemma 3.2.1, whose truth may seem
intuitively obvious, but whose proof is actually rather delicate and involves
more topology than we care to consider here.

Lemma 3.2.1. Let S and T be surfaces with the same number of boundary
components. Let S' and T' be closed surfaces obtained by identifying the
boundaries of disks to the boundary components of S and T, respectively. Then
the surfaces S and T are homeomorphic if and only if the surfaces S' and T' are
homeomorphic. 0

Observe that Lemma 3.2.1 validates the omission of 2-bands as in Figures

3.13 and 3.15. Since the surface defined by the 0-bands and 1-bands can be
drawn in 3-space, while the whole surface sometimes cannot, the boon to
visual intuition is considerable.
Thus, to describe a graph imbedding G ~ S or its band decomposition, one
need specify only how the ends of the 1-bands are to be attached to the
0-bands. This is done by labeling a disjoint collection of arcs in the boundary
of each 0-band by the names of edges incident on the vertex surrounded by
that 0-band. Each arc of 1-band attachment is directed by the assigned
orientation of the given 0-band. The ends of each 1-band are then directed and
identified to the arcs labeled by the corresponding edge. The lengths of the
arcs do not matter. As complicated as this sounds, it means that one need only
provide for each vertex an ordered list, unique up to a cyclic permutation, of
the edges incident on that vertex and a designation of orientation preserving
or reversing to each 1-band (to tell which way the ends are pasted), or
equivalently, whether the corresponding edge is type 0 or type 1.

Example 3.2.3. For instance, if we orient the 0-bands of Figure 3.13 both to be
clockwise on the page, we have the following combinatorial description, where the
superscript "1" designates a type-1 edge and the absence of a superscript a type-0
edge:

v1• e1 e2 e/
v2 • e/ e 1 e 2
To make this method of combinatorial description precise, we define a
"rotation" at a vertex v of a graph to be an ordered list, unique up to a cyclic
permutation, of the edges incident on that vertex. Naturally, every v-based
Band Decompositions and Graph lmbeddings 113

loop is mentioned twice in the rotation at v. Let a "rotation system" on a

graph be an assignment of a rotation to each vertex and a designation of
orientation type for each edge. Then the preceding discussion can be sum-
marized by the following theorem, used extensively by Ringel in the 1950s.
The first formal proof was published by Stahl (1978).

Theorem 3.2.2. Every rotation system on a graph G defines (up to equivalence

of imbeddings) a unique locally oriented graph imbedding G ~ S. Conversely,
every locally oriented graph imbedding G ~ S defines a rotation system for G.
0

3.2.4. Pure Rotation Systems and Orientable Surfaces

Let the graph G have an imbedding in an oriented surfaceS. If the 0-bands of
the associated band decomposition are given local orientations consistent with
the orientation of the surfaceS, then every 1-band will be orientation-preserv-
ing. Conversely, any rotation system for the graph G such that every edge has
type 0 induces an imbedding in an orientable surface (obtained by supplying
the 2-bands) and a specific orientation of that surface. Theorem 3.2.3 sum-
marizes this discussion. For the sake of brevity we define a "pure rotation
system" for a graph to be one in which every edge has type 0.

Theorem 3.2.3. Every pure rotation system for a graph G induces (up to
orientation-preserving equivalence of imbeddings) a unique imbedding of G into
an oriented surface. Conversely, every imbedding of a graph G into an oriented
surface induces a unique pure rotation system for G. 0

Like Ringel (1974), we think it appropriate to ascribe credit for Theorem

3.2.3 jointly to Heffter (1891) and Edmonds (1960). See the Historical Note at
the end of the section.

3.2.5. Drawings of Rotation Systems

There is a particularly simple way to incorporate into a drawing of a graph a
rotation system for the graph: just be sure the clockwise order of the edges
incident on a vertex in the drawing agrees with the assigned rotation at the
vertex. The easiest way to do this is to draw first a dot for each vertex with
spokes radiating from the dot labeled in clockwise order according to the
rotation at the vertex. Then curves are drawn joining spokes with the same
label. Finally, all type-1 edges are marked with a cross. The resulting drawing
is called a "projection" of the given rotation. For instance, Figure 3.16 shows
a projection of the pure rotation system defined by the toroidal imbedding of
K 4•4 minus a 1-factor given in Figure 3.10. For this projection a judicious
choice of the "under" edge at each crossing helps make the surface " visible".
114 Surfaces and Graph lmbeddings

Figure 3.16. A rotation projection with no type-1 edges.

w
Figure 3.17. A rotation projection for K 4 and the corresponding reduced band decomposition.

If a given graph is simplicial, then the list format of a rotation system may
give adjacent vertices instead of edges. This is called the "vertex form of a
rotation system". In this context, it is also natural to list only the sequence of
vertices in a boundary walk.

Example 3.2.4. The rotation system for K 4 whose projection is illustrated on the
left in Figure 3.17 can be given a list format in either the edge form or the vertex
form.

u. c1 ba u. x 1 wv
v. fad v. xuw
w. dbe w. DUX
X. efcl X. wvu 1

The corresponding reduced band decomposition is given at the right. By tracing

along the boundary of this reduced band decomposition surface, one easily verifies
that the imbedding has two faces u v x w u x w v x and u v w in vertex form, or
afebcedfc and a db in edge form.

3.2.6. Tracing Faces

Given a rotation system for a graph, one frequently needs to obtain a listing or
enumeration of the boundary walks of the reduced faces. If the rotation
system is given by a projection, then the procedure is well illustrated in
Example 3.2.4. First, thicken each edge into a 1-band, and give the band a
twist if the edge is type 1. Then simply trace out, say with a pencil, the
Band Decompositions and Graph lmbeddings 115

boundary components of the resulting surface. On the other hand, if the

rotation system is given in list format, for example, as computer input, then
this geometric method is impossible. Some thought about the geometric
method should convince the reader that the following Face Tracing Algorithm
is correct. We first introduce some helpful terminology. If the rotation at
vertex v is ... de ... (up to a cyclic permutation), then we say that d is "the
edge before e at v", that e is "the edge after d at v", and that the edge pair
(d, e) is a "corner at v with second edge e".

Face Tracing Algorithm. Assume that the given graph G has no 2-valent
vertices. Choose an initial vertex v0 of G and a first edge e1 incident on v0 • Let v 1
be the other endpoint of e 1 . The second edge e 2 in the boundary walk is the edge
after (resp., before) e 1 at v 1 if e 1 is type 0 (resp., type I). If the edge e 1 is a
loop, then e 2 is the edge after (resp., before) the other occurrence of e1 at v 1 • In
general, if the walk traced so far ends with edge e; at vertex D;, then the next
edge ei+l is the edge after (resp., before) e; at D; if the walk so far is type 0
(resp., type I). The boundary walk is finished at edge en if the next two edges in
the walk would be e 1 and e 2 again. To start a different boundary walk, begin at
the second edge of any corner that does not appear in any previously traced faces.
If there are no unused corners, then all faces have been traced.

Observe that the walk does not necessarily stop when the first edge e 1 is
encountered a second time; we might not be on the same side of e 1 as at the
beginning. The followup by the edge e 2 is what confirms that we are on the
original side of e 1, assuming of course that the vertex v 1 does not have valence
2 (see Exercise I4).

Example 3.2.5. Consider the rotation system

u. a 1 fbd 1 a 1 e 1 be
v. cfg

w. e1 d 1 g

Begin the first face at vertex u with the edge a at the corner ( c, a). The next edge
in the boundary walk is d, the edge before (since the walk a is type I) the other
occurrence of the edge a at the vertex u. The next edge is g, the edge after d
(since the walk ad is type 0) at the vertex w. The next edge is c. Since the
following two edges would be a and d again, the walk terminates with c, yielding
the face adgc. Since the corner (a, f) has not yet appeared, we can begin a
second boundary walk with edge f at vertex u. This time the face f g e a is
obtained. The third and fourth faces b c f and deb are obtained in a similar
fashion by starting with edge b from corner (/, b) and edge d from corner ( b, d).
Figure 3.I8 shows these four faces, first separately, then assembled together.
116 Surfaces and Graph Imbeddings
u a u

di<' , ~,
b

g a
u u
w ' g I' v w e u
a a

fA g'
u u
u u

I b
u c v w e u
Figure 3.18. The faces from Example 3.2.5 assembled into a Klein bottle, so that the local
orientations at vertices u, v, and w are correctly completed.

Example 3.2.6. The rotation system

t. d 1 be 1
u. fd 1 a 1
v. c1 ba 1
w. cl elf

has as its faces the polygons dfe, bad, ecb, and fca. When these faces are
assembled, they yield a 2-sphere. Thus, as we have mentioned, the existence of
orientation reversing edges need not mean that the surface is nonorientable,
provided that no cycle contains an odd number of orientation reversing edges.

3.2.7. Duality
Let B be a band decomposition of the surface S. Suppose each 2-band has
been given a specific orientation. The dual band decomposition of B, denoted
B *, is defined by letting the i-bands of B be the (2 - i)-bands of B*, for
i = 0, 1, 2; the ends of each 1-band of B become the sides of the corre-
sponding 1-band of B* and vice versa. If G ~ S is the graph imbedding
associated with the decomposition B, then the imbedding associated with B*
is naturally the dual imbedding G* ~ S*.
Let e be an edge of the imbedding G ~ S associated with the decomposi-
tion B. The orientation type of the dual edge e* depends on the choice of
orientations for the 2-bands of the primal imbedding or, equivalently, on the
direction given to the closed walk around the boundary of each primal face.
The edge e appears twice in the course of listing all directed face boundaries.
If the two appearances have opposite directions, then the dual edge e* is type
0; otherwise it is type 1. Thus, in Example 3.2.4, if the directed boundaries are
u vx wux wvx and uvw, then (ux)*, (v w)*, and (wx)* are type 0, whereas
Band Decompositions and Graph Imbeddings 117

( u v )*, ( u w )*, and ( v x )* are type 1. If the direction u w v replaces direction

u v w for the second face, then ( u v )* and ( u w )* become type 0; the dual edge
( v w )* becomes type 1, and the loops ( u x )*, ( w x )*, and ( u v )* remain the
same.
If the surface S is orientable and all face orientations are chosen to agree
with one orientation for S, then all the dual edges are type 0.

3.2.8. Which 2-Complexes Are Planar?

One possible way to view the 2-cells in a 2-complex K (simplicial or not) is as
restrictions on the possible rotation systems for its !-skeleton. That is, if two
edges are consecutive on the boundary of any 2-cell of K, then they must be
consecutive in every rotation at their common vertex, or else the rotation
system would not extend the 2-complex K to a surface.
Using rotation systems, it is not difficult to prove that if the star of every
vertex of the 2-complex K is planar-in which case we say that K is "locally
planar"- then K has an imbedding in some closed surface. Harary and Rosen
(1976) give a complete proof of the folk theorem that characterizes the absence
of local planarity in terms of a single obstruction. A "thumbtack" is a
topological space homeomorphic to the following subset of IR 3 :

{(x, y,O)Ix 2 + y 2 ~ 1} u {(0,0, z)IO ~ z ~ 1}

The characterization is that a 2-complex is not locally planar if and only if it

contains a thumbtack.
One might guess naively that a locally planar 2-complex is planar if and
only if its !-skeleton is planar. However, Figure 3.1 contains a counterexample
to that assertion. In particular, the 2-complex on the left has a planar
!-skeleton, isomorphic to K 5 - e. However, the 2-complex on the right has the
same topological space as its carrier, and its !-skeleton contains K 3 3 •
Gross and Rosen (1981) prove that a simplicial 2-complex is imb~ddable in
the sphere if and only if it is locally planar and the !-skeleton of its first
barycentric subdivision is planar. If the 2-complex in question is not simpli-
cial, it can first be subdivided into a simplicial complex before applying this
test. In a sequel, Gross and Rosen (1979) prove that reducing the planarity
question for an arbitrary 2-complex to testing the planarity of a graph can be
accomplished in linear time in the number of vertices of the 2-complex.

Historical Note. The earliest form of Theorem 3.2.3 was implicitly given in a
dualized vertex form by Heffter (1891) and used extensively by Ringel in the
1950s. Evidently unaware of the existence of this work, Edmonds (1960)
invented the primalized vertex form, and its details were subsequently made
accessible by Youngs (1963). The full generalization to all graphs, simplicial or
not, requires the switch from the vertex form to the directed-edge form, and it
was developed by Gross and Alpert (1974).
118 Surfaces and Graph Imbeddings

3.2.9. Exercises
1. Consider the pure rotation system

u. cbd x. abg
v. ajh y. idf
w. ecfe z. hijg

Draw a projection of this rotation system and list the faces.

2. Consider the pure rotation projection given in Figure 3.19. Write out the
rotation system and list the faces.
3. Suppose that the edges a, d, e, and h in Exercise 1 were type 1. Draw
the 0-bands and 1-bands of the corresponding band decomposition. Then
list the faces.
4. Suppose that the edges a, d, h, and i in Exercise 2 were type 1. List the
faces of the resulting imbedding.
5. Draw a rotation projection for an imbedding of K 6 on a torus.
6. Suggest a way of describing a rotation system using an incidence matrix.
7. Suggest a way of describing a rotation system for a simplicial graph using
an adjacency matrix.
8. How many inequivalent vertex-labeled imbeddings in orientable surfaces
does K 7 have?
9. Show that there are at most 5!(3!) 6 labeled imbeddings of K 7 in
orientable surfaces having six three-sided faces meeting at vertex 1.
Conclude that the fraction of labeled orientable imbeddings of K 7 that
have all face~ triangular is less than 1.6 X 10- 8•

Figure 3.19. A rotation projection.

The Classification of Surfaces 119

Figure 3.20. A rotation system for a cubic graph. Hollow dots have counterclockwise rotation.

Figure 3.21. Two rotation proJections.

10. Let G --+ S be an imbedding. Suppose that / 1 and / 2 are two distinct
faces of the imbedding whose boundaries share the edge e. Let G' --+ S'
be the imbedding obtained from the rotation system for G --+ S by
"giving the edge e an extra twist" (that is, by changing the orientation
type of e). Show that the imbedding G --+ S' has one less face than the
imbedding G--+ S.
ll. Use Exercise 10 to show that every connected graph has an imbedding
with only one face (Edmonds, 1965).
12. Let G be a regular graph of valence 3. Each vertex has exactly two
possible rotations. Thus a rotation system for G can be given from a
drawing of G in the plane by, say, a hollow dot for a vertex whose
rotation is counterclockwise in the drawing and a solid dot if clockwise.
How many faces are there in the rotation system given by this method in
Figure 3.20?
13. Two rotation projections are given in Figure 3.21. Choose a spanning tree
in each graph and reverse some vertex orientations so that each edge in
the spanning tree is type 0. Then determine which of the two surfaces is
orientable.
14. Show by an example that the procedure given in the Face Tracing
Algorithm for terminating a boundary walk may stop the walk too soon
if the vertex v1 is 2-valent. Suggest a way of handling 2-valent vertices.

3.3. TilE CLASSIFICATION OF SURFACES

It is a remarkable fact of low-dimensional topology that the surfaces

S0 , S1 , S2 , ••• form a complete set of representatives of the homeomorphism
120 Surfaces and Graph Imbeddings

types of closed, connected, orientable surfaces, and moreover, that they are
distinguishable by a single integer-valued invariant, the Euler characteristic.
Similarly, the surfaces N1, N2 , N3 , ••• form a complete set of representatives of
the homeomorphism types of closed, connected, nonorientable surfaces, and
they too are distinguishable by Euler characteristic.
The proof of the classification theorem is facilitated by a designation of
convenient models of the surfaces. As a model of Sg we take the surface of an
unknotted g-holed solid doughnut in 3-space.
A "meridian" in this model is a topological loop on Sg that bounds a disk
in the g-holed solid that does not separate the solid. The topological closure of
the complement of the g-holed solid is also a g-holed solid doughnut. A
"longitude" on Sg is a topological loop that bounds a disk in the complemen-
tary solid that does not separate the complementary solid. Figure 3.22 shows a
three-holed solid doughnut and standard sets of meridians and longitudes.
As a model of Nk we take the surface obtained from a sphere by first
deleting the interiors of k disjoint disks and then closing each resulting
boundary component with a Mobius band. No closed nonorientable surface is
imbeddable in 3-space.
The classifications are derived in three steps. The first step is to calculate an
Euler characteristic-relative to a particular cellular imbedding of a particular
graph-for each of the standard surfaces. Second, we show that the Euler
characteristic is an invariant of every standard surface, that is, it is indepen-
dent of the choice of a graph and of the choice of an imbedding (provided that
the imbedding is cellular).
Since the standard orientable surfaces have different Euler characteristics, it
follows from the second step that no two of them are homeomorphic.
Similarly, the nonorientable standard surfaces are pairwise distinct. In the
third step, we introduce the concept of "surgery" to prove that every closed
surface is homeomorphic to one of the standard surfaces.

longitudes

meridians
Figure 3.22. An unknotted 3-holed solid doughnut in 3-space, with standard meridians and
longitudes.
The Classification of Surfaces 121

3.3.1. Euler Characteristic Relative to an Imbedded Graph

The "(relative) Euler characteristic of a cellular imbedding" G ~ S of a
connected graph into a closed, connected surface is the value of the Euler
formula #V- #E + #F, and it is denoted x(G--+ S). In the course of
proving the classification theorems we will prove that this number depends
only on the homeomorphism type of the surface S and not on the choice of a
graph G or its imbedding.

Theorem 3.3.1. For each orientable surface Sg (g = 0, 1, 2, ... ) there exists a

connected graph G and a cellular imbedding G ~ Sg whose Euler characteristic
satisfies the equation x(G ~ Sg) = 2- 2g.

Proof Theorem 1.4.1 implies this result for g = 0. For g > 0, let the
vertices of the graph G be the g intersection points of a standard set of g
meridian-longitude pairs on the surface Sg. The edges of G include the g
meridians, the g longitudes, and a set of g - 1 "bridges" that run directly
between consecutive intersection points. Thus,'the graph G has g vertices and
3g - 1 edges. One verifies by an induction on g that there is only one face
and that it is a 2-cell. It follows that x(G--+ Sg) = g- (3g- 1) + 1 = 2 - 2g.
D

Theorem 3.3.2. For each nonorientable surface Nk (k = 0, 1, 2, ... ) there is a

graph G and a cellular imbedding of G into the surface Nk such that x(G--+ Nk)
= 2- k.

Proof For k = 0, this is true by Theorem 1.4.1. For k > 0, choose a point
on the center loop of each of the k crosscaps as a vertex of G. The edges of G
are the k center loops plus k - 1 bridges to link the vertices to each other.
Thus, there are k vertices and 2k - 1 edges. An induction argument on the
number k proves that there is only one face and that it is a 2-cell. Therefore,

x( G ~ Nk) = k- (2k- 1) + 1 = 2- k. D

3.3.2. Invariance of Euler Characteristic

The choice in Theorem 3.3.1 and in Theorem 3.3.2 of particular cellular
imbeddings in the standard surfaces was designed to simplify the calculation
of their Euler characteristics. We shall now see that for every one of the
standard surfaces, the value of the Euler formula #V- #E + #F is indepen-
dent of the selections of a graph and of a cellular imbedding.
It follows immediately from the invariance of the Euler characteristic of the
standard surfaces that they are mutually distinct, because if two orientable
surfaces differ in genus or if two nonorientable surfaces differ in crosscap
number, then they differ in Euler characteristic. Establishing the invariance of
122 Surfaces and Graph Imbeddings

the Euler characteristic, and thereby, the distinctness of the standard surfaces,
is the second step in the classification of surfaces.

Theorem 3.3.3 (The Invariance of Euler Characteristic of Orientable Surfaces).

Let G-+ Sg be a cellular imbedding, for any g = 0, 1,2, .... Then x(G-+ Sg)
= 2- 2g.

Proof Theorem 1.4.1 implies that the conclusion holds for the sphere S0 •
As an inductive hypothesis, assume that it holds for the surface Sg, and
suppose that the graph G is cellularly imbedded in Sg+ 1 •
First, draw a meridian on Sg+ 1 so that it meets the graph G in finitely many
points, each in the interior of some edge of G and each a proper intersection,
not a tangential intersection. If necessary, we subdivide edges of G so that no
edge of G crosses the meridian more than once, which does not change the
relative Euler characteristic of the imbedding into Sg+ 1 . Next thicken the
meridian to an annular region R, as illustrated in Figure 3.23.
A homeomorphic copy of Sg can be obtained by excising from Sg+ 1 the
interior of region R and by then capping off the two holes with disks. We now
construct a cellular imbedding G' -+ Sg. The vertex set of the graph G' is the
union of the vertex set VG and the intersection set of G with the boundary of
R. If the graph G intersects the meridian in p places and if the region R is
selected to be adequately narrow, then #VG' = #VG + 2p.
The edge set EG' contains all the #EG- p edges of G that do not cross the
meridian. Furthermore, it contains the arc segments on the boundary compo-
nents of the region R that connect adjacent points of intersection of
boundary( R) and G. Since G intersects the meridian in p places, there are p
such arcs on each boundary component, for an additional subtotal of 2p. The
edge set EG' also contains the segments of edges of G that run from vertices of
G to intersections of G and boundary(R). There are 2p such segments in all.
Thus,

#EG' = ( #EG- p) + 2p + 2p = #EG + 3p

Figure 3.23. Thickening the meridian yields the shaded region R.

The Classification of Surfaces 123

Each face of the imbedding G -+ Sg+l is cellular. It follows from the Jordan
curve theorem that each of the p subarcs of the meridian subtended by G,
considered in succession, cuts a region into two parts, thereby yielding p
additional faces in the imbedding G' -+ Sg, it follows that

Therefore, we may compute that

X( G-+ Sg) = #VG'- #EG' + #FG'

= ( #VG + 2p)- ( #EG + 3p) + ( #FG + p + 2)
= #VG- #EG + #FG + 2
= x(G-+ sg+l) + 2

According to the induction hypothesis,

We immediately infer that

x( G-+ sg+l) = -2g = 2- 2(g + 1) o

Corollary. Let i andj be distinct nonnegative integers. Then the surfaces S; and
sj are not homeomorphic.

Proof A homeomorphism f: S;-+ S1 would carry a cellular imbedding

G -+ S; of relative Euler characteristic x(G-+ S;) = 2 - 2i to a cellular
imbedding G-+ s1 of relative Euler characteristic x(G-+ S) = 2- 2i, in
violation of Theorem 3.3.3, which implies that x(G-+ S) = 2- 2j. o

Theorem 3.3.4 (The Invariance of Euler Characteristic of Nonorientable

Surfaces). Let G-+ Nk be a cellular imbedding, for any k = 0, 1, 2, .... Then
x(G-+ Nk) = 2- k.

Proof The conclusion holds for the 2-sphere N0 by Theorem 1.4.1. By

way of an induction, assume that it holds for the surface Nk, and suppose that
the graph G is cellularly imbedded in Nk+l·
The induction step here is analogous to the one for the orientable case. This
time draw a center loop on one of the crosscaps of Nk+l so that it meets the
graph G in finitely many points, and subdivide some edges if needed, so that
no edge of G crosses that center loop more than once. Such subdivision does
not alter the relative Euler characteristic of the imbedding. When the center
loop is thickened, the result is a Mobius band B, rather than an annulus.
124 Surfaces and Graph Imbeddings

A homeomorphic copy of Nk can be obtained by excising from Nk+ 1 the

interior of the Mobius band B and then capping off the resulting hole with a
disk. Continuing the analogy, we construct an imbedding G' ~ Nk. If we
assume that the graph G intersects the center loop in p points, then the vertex
set Ve· contains #Ve vertices from G plus 2p points of intersection of G with
boundary( R ).
The edge set Ee' contains #Ee + 3p edges, as before. However, since only
one hole results from excising the interior of the Mobius band B, the face set
Fe· contains #Fe + p + 1 faces. It follows that

x( G ~ Nk) = #Ve·- #Ee· + #Fe·

= ( #Ve + 2p)- ( #Ee + 3p) + (#Fe+ p + 1)
= #Ve - #Ee + #Fe + 1
= x( G ~ Nk+ 1 ) + 1

The induction hypothesis implies that x(G ~ Nk) = 2- k, which enables us

to conclude that

x( G ~ Nk+l) = 1 - k = 2- (k + 1) D

Corollary. Let i andj be distinct nonnegative integers. Then the surfaces N; and
~are not homeomorphic. D

3.3.3. Edge-Deletion Surgery and Edge Sliding

What remains to complete the classification of closed surfaces is the third step,
which is to prove that every closed surface is homeomorphic to one of the
standard surfaces. To accomplish this, we introduce some topological tech-
niques called "surgery". This is the most difficult step in the classification.
Let G --+ S be a cellular imbedding of a graph in a surface, and let e be an
edge of G. To perform "edge-deletion surgery" at e, one first deletes from a
band decomposition (not reduced) for G ~ S the 2-band or 2-bands that meet
the e-band; one next deletes the e-band itself; and finally one closes the hole
or holes with one or two new 2-bands, as needed.

Remark 3.3.1. Suppose that the imbedding G ~ S induces the rotation system
R and that the imbedding G' ~ S' results from edge-deletion surgery at edge e.
Then the rotation system R' induced by G' ~ S' can be obtained simply by
deleting both occurrences of edge e from R. For a rotation projection, this means
erasing the edge e. For a reduced band decomposition, it means deleting the
e-band.
The Classification of Surfaces 125

There are three different possible occurrences when the edge e is deleted,
which are called cases i, ii, and iii. The effect of the surgery depends on the
case.

i. The two sides of the edge e lie in different faces, / 1 and / 2 . Then
deleting the fcband, the / 2-band, and the e-band leaves one hole,
which is closed off with one new 2-band.
ii. Face f is pasted to itself along edge e without a twist, so that deleting
the /-band and the e-band leaves two distinct holes (and possibly
disconnects the surface). These holes are closed off with two new
2-bands.
iii. Face f is pasted to itself with a twist along edge e, so that the union of
the /-band and the e-band is a Mobius band. Then deleting the /-band
and the e-band leaves only one hole, which is closed off with one new
2-band.

The following theorem follows immediately from the preceding description.

lbeorem 3.3.5. Let G ~ S be a cellular graph imbedding, and let e be an edge

of the graph. Let F be the set of faces for G ~ S, and let F' be the set of faces of
the imbedding obtained by edge-deletion surgery at e. Then in case i, #F' =
#F- 1, and the resulting surface is homeomorphic to S; in case ii, #F' =
#F + 1; and in case iii, #F' = #F. D

Another variety of imbedding modification is simpler, since it changes only

the imbedded graph, not the surface. Let f be a face of an imbedding G ~ S,
let d and e be consecutive edges in the boundary walk of f with common
vertex v, and let u and w be the other endpoints of d and e, respectively. To
"slide d along e ", one adds a new edge d' from u to w and places its image
across face f so that it is adjacent to d in the rotation at u and adjacent to e
in the rotation at w, and one then removes edge d from the graph. This is
illustrated in Figure 3.24.

-
Figure 3.24. Sliding edged along edge e.
126 Surfaces and Graph lmbeddings

Remark 3.3.2. Suppose that the edge-sliding operation is performed on a locally

oriented surface. If edged is type 0 and edge e is type 1, then the new edged'
that results from sliding d along e is type 1.

3.3.4. Completeness of the Set of Orientable Models

The following lemma is a crucial ingredient of the proof that the surfaces
S0 , S 1,... represent all the orientable types. We hope that its proof is
intuitively obvious.

Lemma 3.3.6. Let T be an orientable surface with two boundary components.

Let T1 be an orientable surface obtained from T by identifying one end of a
1-band I X I to an arc in one boundary component and identifying the other end
to an arc in the other boundary component. Suppose that T2 is another orientable
surface obtained from Tin a similar way. Then T1 and T2 are homeomorphic.

Proof Suppose that a; and b; are the arcs in boundary(T) to which the
band for 1'; is attached, for i = 1, 2. Then there is a homeomorphism of T to
itself taking a 1 to a 2 and b1 to b2 • This homeomorphism can be extended to
the attached bands by the product structure of I X I, since neither band is
"twisted" (the surfaces T, T1 , and T2 are all orientable by assumption). The
result is a homeomorphism from T1 to T2 • D

Corollary. Let G-+ S and H-+ T be two different one-face imbeddings in

orientable surfaces, and let d and e be arbitrary edges of the graphs G and H,
respectively. If edge-deletion surgery at d and edge-deletion surgery at e yield
homeomorphic surfaces S' and T', respectively, then the original surfaces Sand
Tare homeomorphic.

Proof Surgery on an edge of a one-face imbedding in an orientable

surface must be of type ii. Thus, it follows from theorem 3.3.5 that the
imbeddings (G- d) -+ S' and (H- e) -+ T' both have two faces, so that
their reduced band decompositions have two boundary components. By Re-
mark 3.3.1 on edge-deletion surgery, one can obtain reduced band decomposi-
tions for the imbeddings G -+ S and H -+ T, respectively, simply by restoring
to each a band from one boundary component to the other. By Lemma 3.3.6,
the reduced band decompositions for G -+ S and H -+ T are homeomorphic
surfaces. It follows from Lemma 3.2.1 that S and T are homeomorphic
surfaces. D

Example 3.3.1. It is not visually obvious that the two reduced band decomposi-
tion surfaces in Figure 3.25 are homeomorphic. However, one can easily verify
by tracing along the boundary walks that both surfaces have one boundary
component. Moreover, deleting band c from both yields two surfaces that have
two boundary components each and are obviously homeomorphic. It follows from
The Oassification of Surfaces 127

a b c d a b c d

~ Figure 3.25.
~
Homeomorphic reduced band decomposition surfaces.

Lemma 3.3.6 that the two surfaces illustrated are also homeomorphic. The
corollary implies that the imbedding surfaces induced by the rotation systems

v. abacbdcd and v. ababcdcd

are homeomorphic.

Theorem 3.3.7. Let G ~ S and H ~ T be cellular graph imbeddings into

closed, connected, orientable surfaces, such that x(G ~ S) = x(H ~ T). Then
the surfaces SandT are homeomorphic. Moreover, x(G ~ S) and x(H ~ T)
are even numbers less than or equal to 2.

Proof As the basis for an induction, suppose that #Ec + #EH = 0. Then
the graphs G and H have only one vertex each, and the imbeddings G ~ S
and H ~ T have one face each. Hence, the surfaces S and T are both
2-spheres, and x(G ~ S) = x(H ~ T) = 2.
Assume that the inclusion holds when #Ec + #EH = n, and suppose that
G ~ S and H ~ Tare imbeddings such that #Ec + #EH = n + 1. If the
imbedding G ~ S had more than one face, then the graph G would have an
edge e common to two faces. By case i of Theorem 3.3.5, the imbedding
( G - e) ~ S obtained by surgery at e would have Euler characteristic equal
to x(G ~ S). Thus, the conclusion would follow by induction. Therefore, one
may assume that the imbedding G ~ S has only one face. Likewise, one may
assume that the imbedding H ~ T also has only one face. Furthermore, the
same argument applied to the dual imbeddings permits one to assume that the
graphs G and H have but one vertex each.
Now let d and e be arbitrary edges of the graphs G and H, respectively,
and let G' ~ S' and H' ~ T' be the imbeddings that result from surgery at d
and e, respectively. Case ii of Theorem 3.3.5 implies that x(G' ~ S') =
x(G ~ S) + 2 and that x(H' ~ T') = x(H ~ T) + 2. Thus, x(G' ~ S') =
x(H' ~ T'). Since G and H are both one-vertex graphs, it follows that G',
H', S', and T' are each connected. Therefore, by induction, x(G' ~ S') and
x(H' ~ T') are even numbers less than 2, from which it follows that x(G ~ S)
and x(H ~ T) are even numbers less than 2. By induction, it follows that the
surfaces S' and T' are homeomorphic. By the corollary to Lemma 3.3.6, we
infer that the surfaces S and T are homeomorphic. D
128 Surfaces and Graph Imbeddings

Corollary. Every closed, connected, orientable surface is homeomorphic to one

of the standard surfaces S0 , S 1, ••••

Proof Let S be a closed, connected, orientable surface. By Theorem 3.3.7,

the surface S has a relative Euler characteristic x(G ~ S) that is even and
less than or equal to 2. (Indeed, Theorem 3.3.7 states that all the relative Euler
characteristics for S are even and less than or equal to 2.) Let g =
1 - 1/2 · x(G ~ S). By Theorem 3.3.1, the surface Sg has a relative Euler
characteristic x(H ~ Sg) = 2- 2g = x(G--+ S). It follows from Theorem
3.3.7 that S is homeomorphic to Sg. D

3.3.5. Completeness of the Set of Nonorientable Models

The proof of the completeness of the set N1 , N2 ,... is parallel to the
orientable case. It begins with an analogy to Lemma 3.3.6.

Lemma 3.3.8. Let T be a surface with one boundary component, and let T1 be a
surface also with one boundary component, obtained from T by identifying the
ends of a 1-band to disjoint arcs in the boundary. Let T2 be another surface with
one boundary component obtained from Tin the same way. Then T1 and T2 are
homeomorphic.

Proof The proof is essentially the same as that for Lemma 3.3.6, but this
time one may observe that both attached bands must have a "twist" so that
the surfaces T1 and T2 still have only one boundary component. D

Corollary. Let G ~ Sand H ~ T be two different one-face imbeddings, and

let d and e be arbitrary edges of the graphs G and H, respectively. If edge-dele-
tion surgery at d and edge-deletion surgery at e yield one-face imbeddings into
homeomorphic surfaces S' and T', respectively, then the original surfaces Sand
Tare homeomorphic.

Proof Since the surfaces S' and T' are homeomorphic, and since the
imbeddings (G- d)~ S' and (H- e) ~ T' are both one-faced, the reduced
band decomposition surfaces for (G- d) ~ S' and for (H- e)~ T' are
homeomorphic surfaces with one boundary component. By Remark 3.3.1 and
Lemma 3.3.8, it follows that the reduced band decomposition surfaces for
G --+ S and H ~ T are homeomorphic surfaces with one boundary compo-
nent each. Lemma 3.2.1 now implies that Sand Tare homeomorphic. D

Example 3.3.2. Both the surfaces in Figure 3.26 have one boundary component.
Moreover, when the b-band is deleted from both, the resulting surfaces are
obviously homeomorphic. It follows from Lemma 3.3.8 that the two surfaces
illustrated are homeomorphic, even though this might not be immediately obvious.
The Classification of Surfaces 129

a
a b

Figure 3.26. Homeomorphic nonorientable reduced band decomposition surfaces.

The Corollary implies that the imbedding surfaces induced by the rotation systems

are homeomorphic.

Theorem 3.3.9. Let G ~ S and H ~ T be cellular, locally oriented graph

imbeddings into closed, connected, nonorientable surfaces, such that x(G ~ S)
= x(H ~ T). Then the surfaces S and T are homeomorphic. Moreover,
x(G ~ S) and x(H ~ T) are less than or equal to 1.

Proof As for orientable surfaces, the proof is by induction on the sum

#EG + #EH. The induction begins with #EG + #EH = 2, since by hypothe-
sis, the imbeddings G ~ S and H ~ T must have at least one type-1 edge
apiece. If the sum is 2, then in both graphs G and H, the single edge must be
an orientation-reversing loop. Thus, the reduced band decomposition surfaces
for the imbeddings G ~ S and H ~ T are both Mobius strips, from which it
follows that S and T are both homeomorphic to the projective plane N1 . It
also follows that x(G ~ S) = x(H ~ T) = 1 - 1 + 1 = 1.
Now suppose that the conclusion holds whenever #EG + #EH.:::;; n for
some n ~ 2. Furthermore, suppose that G ~ S and H ~ T are imbeddings
satisfying the hypotheses of the theorem and the condition #EG + #EH = n
+ 1. As in Theorem 3.3.7 a type-i edge-deletion surgery argument enables us
to assume that both G and H have one vertex and that both imbeddings
G ~ S and H ~ T have one face.
Since #EG + #EH > 2 and since x(G ~ S) = x(H ~ T), both the
bouquets G and H have more than one loop. Since S and T are nonori-
entable, both imbeddings have at least one orientation-reversing loop. By
Remark 3.3.2 on edge sliding, it may be assumed that all the edges of G and H
are orientation-reversing loops.
Now let d and e be edges of the graphs G and H, respectively, whose duals
d* and e* are orientation-reversing loops in the respective dual imbeddings.
(Since the surfaces S and T are nonorientable, there must exist at least one
orientation-reversing loop in both of the respective dual imbeddings.) Next let
G' ~ S' and H' ~ T' be the imbeddings that result from edge-deletion
surgery at d and e, respectively. Since the dual loops d* and e* are both
orientation-reversing, it follows that both surgeries are type iii, which implies,
130 Surfaces and Graph Imbeddings

by Theorem 3.3.5, that the resulting imbeddings G' ~ S' and H' ~ T' have
one face each.
The surfaces S' and T' are both nonorientable, since the original im-
beddings G ~ S and H ~ T had at least one orientation-reversing loop each
besides the deleted loops d and e, respectively. Since x(G' ~ S') =
x(G ~ S) + 1 and x(H' ~ T') = x(G ~ T) + 1, it follows that x(G' ~ S')
= x(H' ~ T'). Moreover,
#Ec, + #Ew = ( #Ec- 1) + ( #EH- 1) = #Ec + #EH- 2 = n- 1 < n

By the induction hypothesis, the surfaces S' and T' are homeomorphic, and
x(G' ~ S') = x(H' ~ T') ~ 1

The corollary to Lemma 3.3.8 now implies that the surfaces S and T are
homeomorphic. Moreover,
X ( G ~ S) = X ( H ~ T) = X ( G' ~ S') - 1 ~ 0 ~ 1. D

Corollary. Every closed, connected, nonorientable surface is homeomorphic to

one of the standard surfaces N 1, N2 , .••.

Proof Let S be a closed, connected, nonorientable surface. By Theorem

3.3.2 it has a relative Euler characteristic x(G ~ S) less than or equal to 1.
Let k = 2 - x(G ~ S). By Theorem 3.3.2, the surface Nk has a relative Euler
characteristic x(H ~ Nk) = 2 - k = x(G ~ S). Theorem 3.3.7 implies that
S is homeomorphic to Nk. D

Theorems 3.3.3 and 3.3.4 and the corollaries to Theorems 3.3.7 and 3.3.9
justify the definition of the "Euler characteristic" x(S) of a surface as the
value of the formula #V- #E + #F for any cellular imbedding. In particu-
lar, x(Sg) = 2 - 2g and x(Nk) = 2- k.

3.3.6. Exercises
1. For each of the three surfaces in Figure 3.27, decide whether it is
orientable, calculate its Euler characteristic, and state to which of the
standard surfaces S0 , S 1, ••. or N1, N2 , ... it is homeomorphic.
2. List the faces of the imbedding obtained by edge-deletion surgery on the
edge uv in the imbedding of Figure 3.17. List the faces obtained when
the surgery is on edge vw instead.
3. Let e be an edge of the imbedding G ~ S whose dual edge e* is not a
loop. Let the imbedding H ~ T be the result of surgery on e. Show that
the dual imbedding H* ~ T* can be obtained from the imbedding
G* ~ S* by contracting the edge e*.
The Oassification of Surfaces 131

- b
--
a
~
a

·tOt· ct
tJ ~c o~
- b ';;'.
b
~
/a
Figure 3.27. Three surfaces.
~~
/c
a

4. Let e be an edge of a connected graph G such that the graph G - e is

disconnected. Show that for every imbedding of G into a surface, the
dual edge e* is a loop. (Hint: Use Exercise 3 and connectivity in primal
and dual imbeddings.)
5. Give an example of a one-face imbedding G ~ S with an edge e such
that surgery on e disconnects G and hence S. Why is it important in the
proof of Theorem 3.3.7 that we may assume the graphs G and H have
only one vertex each?
6. Give an example of a one-face imbedding G ~ N3 with an edge e such
that surgery on e leaves an orientable surface. Give an example of a
one-face imbedding H ~ N3 such that every surface obtainable by
edge-deletion surgery is nonorientable.
7. Give an example of a one-face, one-vertex imbedding G ~ S all of whose
edges are type 1 such that surgery on some edge of G results in a
two-face imbedding. (Hint: It is easiest to start with the two-face
imbedding and then to attach a 1-band.)
8. Find cellular imbeddings of the bouquet of three circles in the surfaces
N1 , N2 , and N3 • Can imbeddings be found in each case that have all
edges type 1?
9. Consider the partial rotation system

u. . .. de .. .
v. . .. e .. .

Give the rotation at vertices u and v after sliding the edge d along the
edge e, first when e is type 0, then when e is type 1.
10. Use your answer to Exercise 9 to give a sequence of edge slides that turns
the one-vertex imbedding with rotation ababc 1c1 into the one-vertex
imbedding with rotation a 1a 1b1b1c1c1• (For example, sliding b along c
gives ab 1ac1b1c1 and then sliding c along a gives ab1ac 1c1b1.)
11. The "disk sum" of two connected, closed surfaces S and T, denoted
So T, is the surface obtained by removing the interior of a disk from
132 Surfaces and Graph Imbeddings

Figure 3.28. The disk sum of two surfaces.

both S and T and then identifying the two resulting boundary compo-
nents (see Figure 3.28). Prove that x(S o T) = x(S) + x(T) - 2.
12. Prove that the disk sum defined in Exercise 11 is an associative operation
on closed, connected surfaces. What is its unit element? Is the operation
commutative?
13. Define the connected surface S to be prime if S is not the sphere and if
S = RoT implies that either R or T is a sphere. What are the prime
surfaces? Give two different factorizations of N3 into primes. Do orienta-
hie surfaces have unique factorizations?
14. Prove that if the surface S contains k disjoint Mobius bands, then
x(S) ~ 2- k. (Hint: Use the fact that if the surface S contains a
Mobius band, it can be factored into S' o N1.) Conclude that for any
nonorientable surface S, x(S) = 2 - k if and only if S contains k
disjoint Mobius bands but no more.
15. Explain how to imbed a graph in the surface N2 k+l so that there exists
an edge e on which surgery yields the orientable surface Sk. (Hint: See
Exercise 13.) What is the minimum number of edge-deletion surgery
steps required to make the surface N2k orientable, for k ~ 1?

3.4. TilE IMBEDDING DISTRIBUTION OF A GRAPH

The "genus range" of a graph G, denoted GR(G), is defined to be the set of

numbers g such that the graph G can be cellularly imbedded in the surface Sg.
Of course, the minimum number in GR(G) is the genus y(G) of the graph.
Sometimes we call y(G) the "minimum genus", to create direct contrast with
the maximum number in GR(G), which is called the "maximum genus" of G,
and is denoted 'YM(G). Whereas the study of minimum genus dates back into
the 19th century, interest in maximum genus began with Nordhaus, Ringeisen,
Stewart, and White (1971, 1972).
Suppose that a certain graph G can be cellularly imbedded in both the
surfaces Sm and Sn, and that r is an integer such that m < r < n. The
"interpolation theorem" -to be proved here as Theorem 3.4.1-for orientable
surfaces states that then there exists a cellular imbedding G ~ S,.
For any two integers m and n such that m :;;; n, we define the "interval of
integers" [m, n] to be the set of all integers r such that m ~ r:;;; n. Thus,
The Imbedding Distribution of a Graph 133

according to the interpolation theorem, the genus range GR(G) is the interval
[y(G), YM(G)].
Analogously, the "crosscap range" CR(G) is defined to be the set of
numbers k such that the graph G can be cellularly imbedded in the surface
Nk. The crosscap number y(G) is the minimum value in this range. The
maximum number in CR(G) is called the "maximum crosscap number" of G,
and is denoted YM(G). There is also an interpolation theorem for nonorienta-
ble surfaces, from which we infer that CR(G) = [y(G), YM(G)].
Because of the interpolation theorems, computations of the limiting values
of the imbedding ranges have received great emphasis. In general, the genus
and the crosscap number of a graph are difficult to compute, although they
have been determined for various special classes of graphs. By way of contrast,
the maximum genus is thoroughly tractable, mainly because of its character-
ization by Xuong (1979). The time needed for computation of the maximum
genus is bounded by a polynomial function of the number of vertices, as
proved by Furst, Gross, and McGeoch (1985a).
In this section we are concerned primarily with establishing the interpola-
tion theorems, with some relationships between genus and crosscap number,
and with computing the maximum crosscap number and maximum genus.
Some general statistical questions about the set of all imbeddings will also be
raised.

3.4.1. The Absence of Gaps in the Genus Range

One calls the graph imbeddings G ~ S and G ~ T "adjacent" if there is an
edge e in G such that the two imbeddings ( G - e) ~ S' and ( G - e) ~ T'
obtained by edge-deletion surgery on e are equivalent. Under such cir-
cumstances, one also calls the rotation systems for G ~ S and G ~ T adjac-
ent. In accordance with Remark 3.3.1, this means that two rotation systems
are adjacent if and only if the result of erasing some edge symbol e in both of
them yields identical rotation systems for the graph G - e.
Theorem 3.3.5 implies that the numbers of faces in adjacent imbeddings
G ~ S and G ~ T both differ by at most one from the common number of
faces in the equivalent imbeddings ( G - e) ~ S' and ( G - e) ~ T'. Thus, the
numbers of faces in G ~ S and G ~ T differ by at most 2. It follows that
the Euler characteristics x(S) and x(T) differ by at most 2. Thus, if the
imbedding surfaces S and Tare orientable, their genera y(S) and y(T) differ
by at most 1.

Example 3.4.1. The three pure rotation projections illustrated in Figure 3.29
correspond to mutually adjacent imbeddings, since edge-deletion surgery on the
edges in the "nine o'clock" position would make them all equivalent. The
left and middle imbedding surfaces visibly have genus 1. The third surface has
genus 2.
134 Surfaces and Graph Imbeddings

Figure 3.29. Projections of three mutually adjacent rotation systems.

Theorem 3.4.1 )Orientable Interpolation Theorem (Duke, 1966)).

Let G be a
connected graph. Then the genus range GR( G) is an unbroken interval of
integers.

Proof Let L 1 and L 2 be list formats for any two rotation systems of G.
Of course, L 2 may be obtained from L 1 by permuting the entries in the
various rows. Any such permutation may be accomplished by a sequence of
steps that moves one edge symbol at a time. The consecutive list formats in
such a sequence represent adjacent imbeddings, by Remark 3.3.1. Thus, there
exists a sequence of adjacent cellular imbeddings of the graph G starting with
one in the surface of genus y(G) and ending with one in the surface of
maximum genus YM(G). Since adjacent imbedding surfaces differ by at most
one in genus, the conclusion follows. D

3.4.2. The Absence of Gaps in the Crosscap Range

In order to obtain the nonorientable interpolation theorem, it is helpful to
introduce another form of surgery on an imbedding G ~ S. To perform
"edge-twisting surgery" at the edge e, one first deletes from a band de-
composition for G ~ S the 2-band or 2-bands that meet the e-band and gives
the e-band an extra twist. One then closes the resulting surface with one or
two 2-bands, as required.

Remark 3.4.1. Suppose that edge-twisting surgery at edge e of the imbedding

G ~ S yields the imbedding G ~ S' and that the corresponding rotation systems
are Rand R', respectively. Then rotation system R' can be obtained from rotation
system R simply by changing the orientation type of edge e, that is, either from
type 0 to type 1 or from type 1 to type 0.

Edge-twisting surgery on e falls into only two cases, illustrated in Figure

3.30.

i. The two sides of e lie in separate faces f 1 and f 2 • Then deleting the
fcband and the f 2-band leaves two holes, which are merged into a
single hole by twisting the e-band. Since the union of the e-band and
The Imbedding Distribution of a Graph 135

Figure 3.30. Case i of edge-twisting surgery goes left to right, whereas case ii goes right to left.
The 2-band boundaries are thickened.

the new 2-band used to close the hole is a Mobius band, the resulting
surface is nonorientable, regardless of whether the starting surface is
orientable.
ii. Both sides of e lie on the same face f. Then deleting the /-band leaves
one hole, which is changed to two holes by twisting the e-band.

Theorem 3.4.2 follows at once from this description.

Theorem 3.4.2. Let G--+ S be a cellular graph imbedding, and let e be an edge
of the graph. Let F be the set of faces of G --+ S, and let F' be the set of faces of
the imbedding G --+ S' obtained by edge-twisting surgery at e. Then in case i,
#F' = #F- 1, and x(S') = x(S) - 1; in case ii, #F' = #F + 1, and
x(S') = x(S) + 1. D

Edge-twisting surgery yields not only a proof that there are no gaps in the
crosscap range but also an easy formula for the maximum crosscap number.
The "Betti number" (also called the "cycle rank") of a connected graph G,
denoted {J(G), is defined by the equation {J(G) = 1 - #V + #E, and it
equals the number of edges in the complement of a spanning tree for G.

Theorem 3.4.3 (Edmonds, 1965). Let G be a connected graph. Then YM( G) =

{J(G).

Proof If the graph G is a tree, then YM(G) = 0, because its maximum

crosscap number is achieved by an imbedding in the sphere. By Theorem 1.2.2,
the Betti number of any tree is 0. If G is not a tree, then G has an imbedding
G --+ S with two or more faces. One might construct such an imbedding, for
instance, by fitting a polygon boundary to a cycle and then extending the
induced partial rotation system in an arbitrary manner to a complete rotation
system. For such an imbedding, one may choose an edge e that lies on two
different faces and perform edge-twisting surgery, so that the result is a
136 Surfaces and Graph Imbeddings

nonorientable manifold, and so that the number of faces is reduced by 1.

Iteration of edge-twisting surgery yields a one-face cellular imbedding in a
nonorientable surface N, so that x(N) = #V- #E + 1. Consequently,
Y(N) = 2- x(N) = 1 - #V + #E = {3(G). D

lbeorem 3.4.4 (Nonorientable Interpolation lbeorem (Stahl 1978)). Let G be

a connected graph. Then the crosscap range CR(G) is an unbroken interval of
integers.

Proof If G is a tree, then CR( G) is-the degenerate interval [0, 0]. Otherwise,
starting with an imbedding G -+ N into a surface that realizes the crosscap
number, one performs a succession of edge-twisting surgeries. At each stage,
one selects an edge that lies in two faces of the imbedding for that stage. Thus,
in YM(G)- y(G) steps, one has obtained a cellular imbedding into every
surface N, such that r lies in the interval [y(G), YM(G)]. D
The Betti number of a graph also gives an upper bound for the maximum
genus of a graph, namely, YM(G).::;; l/3(G)/2J with equality occurring only if
the graph G has an orientable imbedding having one or two faces. However,
unlike the maximum crosscap number, this upper bound for the maximum
genus is not always achieved. One might hope that the Betti number also
provides a useful upper bound on the minimum genus. Indeed, Duke (1971)
once conjectured that y( G) .::;; l f3{ G) 14J. Many counterexamples were found
to Duke's conjecture. Milgram and Ungar (1977) found perhaps the most
discouraging result: given any £ > 0, there are infinitely many graphs G such
that y(G)//3(G) > ! -£.In other words, we cannot obtain an upper bound
for y(G) of the form cf3(G) that is any better than {3(G)/2.

3.4.3. A Genus-Related Upper Bound on the Crosscap Number

Edge-twisting surgery now enables us to establish an upper bound on the
crosscap number of a graph, based on its genus. It follows that the crosscap
range is always nearly twice as large as the genus range of a graph or larger.

lbeorem 3.4.5. Let G be a connected graph. Then y(G).::;; 2y(G) + 1.

Proof If G is a tree, then y(G) = y(G) = 0. Otherwise, let e be an edge

that lies on two faces of an imbedding G -+ S into an orientable surface of
genus y(G). Then edge-twisting surgery one yields an imbedding of G into a
nonorientable surface of crosscap number 2y(G) + 1, by Thet>rem 3.4.2. D

Corollary. Let G be a connected graph. Then

#CR{G) ~ 2#GR(G)- 2
The Imbedding Distribution of a Graph 137

Proof If the maximum genus of G is realized by an imbedding in the

surface S, then

2- 2yM(G) = x(S) = #V- #E + #F~ #V- #E +I

from which it follows that YM(G) = {3(G) ~ 2yM(G). Thus,

# CR( G) = yM (G) - Y( G) + I
~ 2yM(G)- {2y{G) +I)+ 1
= 2(yM(G)- y(G) + 1)- 2
= 2#GR(G)- 2 o

3.4.4. The Genus and Crosscap Number of the Complete Graph K 7

The upper bound on y(G) given by Theorem 3.4.5 is essentially unimprovable.
What this means is that sometimes a graph can be imbedded in an orientable
surface S8 , of Euler characteristic 2 - 2g, but ~annot be imbedded in the
nonorientable surface N28 of the same characteristic. In particular, Heffter
(I891) proved that the complete graph K 7 can be imbedded in the torus S1 ,
while Franklin (I934) proved that K 7 cannot be imbedded in the Klein bottle
N2 • We combine these two results in the following theorem.

Theorem 3.4.6. The genus of the complete graph K 1 is I, and the crosscap
number of K 7 is 3.

Proof By Theorem 1.4.2, no imbedding of K 7 has more than (2/3)#£ =

(2/3)(2I) = 14 faces. Thus, the Euler characteristic of an imbedding surface
for K 7 is at most 7 - 2I + I4 = 0. Hence, y(K 7 ) ~ 1 and y(K 7 ) ~ 2. Figure
3.3I shows an imbedding K 7 ~ S 1 , which completes the proof that y(K 7 ) = 1.
Figure 5.2 contains a more elegant depiction of a weakly equivalent im-
bedding.

1 2
Figure 3.31. An imbedding of the complete graph K 7 in the torus.
138 Surfaces and Graph lmbeddings

It follows from Theorem 3.4.5 that Y(K 7 ) ~ 3. Thus, to complete the proof
that Y( K 7 ) = 3, it is sufficient to show that K 7 has no imbedding in N2 • By
the invariance of Euler characteristic, if K 7 did have an imbedding in N2 ,
there would have to be 14 faces. Since K 1 has no 1-cycles or 2-cycles, each of
the 14 faces would have to be three-sided. Therefore, suppose that K 7 ~ S is
an imbedding such that every face is three-sided. It is now to be proved that
the surface S is orientable, thereby establishing that K 7 is not imbeddable in
the Klein bottle N2 •
Regarding K 7 as a Cayley graph for the cyclic group f!E7 , we label its
vertices 0, 1, ... , 6. It may be assumed, without loss of generality, that the
rotation at vertex 0 is
0. 123456
The edges incident on vertex 0 form a spanning tree, and we locally orient the
vertices so that all six of these edges are type 0. The left side of Figure 3.32
shows a clockwise partial rotation projection "centered" at vertex 0.
Since each of the six faces incident on vertex 0 is three-sided, we may infer
from the given rotation at vertex 0 and the Face Tracing Algorithm the
following partial rotation system (in vertex format):

0. 123456
1. 206 00 0

2. 301 0 00

3. 402 000

4. 503 000

5. 604 000

6. 105 0 0 0

If the edge 12 had type 1, then the sequence 10, 02,21 in an execution of the
Face Tracing Algorithm would be followed by some edge 1x, x -:/= 0, thereby
creating a face that is not three-sided. Thus, the edge 12 must have type 0.

3 or4

6 or 1
Figure 3.32. Partial rotation projections for putative triangular imbeddings of K 1 .
The Imbedding Distribution of a Graph 139

Similarly, each of the edges 23, 34, 45, 56, and 61 must have type 0, as
indicated. We shall see that there are but two ways to extend this partial
system to a rotation system that has only three-sided faces, and that all of the
other edges in the complete system also have type 0, thereby proving that the
surface is orientable.
The edge 16 must lie in a second triangle besides 016. The third vertex of
this triangle must be 3 or 4, since the edges 56 and 12 are already in use. By
similar reasoning applied to the five edges 12, 23, 34, 45, and 56, we arrive at
the partial rotation projection shown on the right of Figure 3.32, where the
vertex 0 and its incident edges are now omitted, to simplify the picture.
Suppose that 3 is the third vertex of the second triangle containing the edge
16. Then the third vertex of the second triangle containing the edge 56 must be
2. Continuing in this manner around the figure, we obtain the partial rotation
projection shown at the left of Figure 3.33. The remaining free edge at each
vertex is now determined, thereby yielding the following complete rotation
system, whose projection is illustrated at the right of Figure 3.33:

0. 123456
1. 206354
2. 301465
3. 402516
4. 503621
5. 604132
6. 105243
If the edge 13 had type 1, then the successive edges 61 and 13 would be
followed by the edge 35 in the Face Tracing Algorithm execution, which
blocks the completion of a three-sided face. Thus, the edge 13 must have type
0. Similarly, the edges 24, 35, 46, 51, and 62 must have type 0. The only

3 3

6 6
Figure 3.33. Extensions of the partial rotation proJection of Figure 3.32.
140 Surfaces and Graph lmbeddings

remaining edges of undetermined type are now 36, 25, and 14. However, in
order for the face tracing sequence 61, 13, 36 to continue with 61, rather than
64, it is necessary that the edge 36 have type 0. Similarly, the edges 25 and 14
have type 0.
If the third vertex of the second triangle containing the edge 16 were 4,
instead of 3, then the same arguments would again yield a pure rotation
system. It follows that the complete graph K 7 has no imbedding in the Klein
bottle, and that Y( K 7 ) = 3. D

3.4.5. Some Graphs of Crosscap Number 1 but Arbitrarily Large Genus

In view of Theorem 3.4.5, one might expect to find a general upper bound on
the genus of a graph, based on its crosscap number. However, an example due
to Auslander et al. (1963) shows that no such upper bound exists. The key to
understanding this example is the idea of cutting a surface apart along a
nonseparating cycle of an imbedded graph, which we have already encoun-
tered in the proof of the invariance of Euler characteristic.
Let G-+ S be a graph imbedding, let B be a (nonreduced) band de-
composition for G -+ S, and let C be the cycle

in G. Then the bands for the vertices and edges of C form a topological
neighborhood N( C) that is homeomorphic to an annulus, if C is orientation
preserving, or to a Mobius band, if C is orientation reversing. In either case,
the cycle C is a topological loop that runs around the middle of the
neighborhood.
Suppose that we cut the surface S along the cycle C, that we regard the
0-band for each vertex v; as having been split into a v;-band and v;' -band, and
that we regard the 1-band for each edge e1 as having been split into an e)-band
and an ej'-band. This is illustrated in Figure 3.34.
Suppose further that we cap off the hole or holes created by cutting on C
with one or two new 2-bands, as required. The graph imbedding G' -+ S'
consistent with the resulting band decomposition is said to have been obtained
by "splitting" the imbedding G -+ S "along the cycle C ". An easy counting

e; ~
u \. .-...;_~(
--(0 \-s--- Vf+ I

Figure 3.34. Splitting an imbedding along a cycle.

The Imbedding Distribution of a Graph 141

to 5

Figure 3.35. The graph A. (for n = 2).

argument on vertices, edges, and faces, as given in the proofs of Theorems

3.3.7 and 3.3.9, yields the following theorem.

Theorem 3.4.7. Let G ~ S be a cellular graph imbedding, let C be a cycle of

G, and let G' ~ S' be the imbedding obtained by splitting G ~ S along C. If C
is orientation preserving, then x(S') = x(S) + 2. If Cis orientation reversing,
then x(S') = x(S) + 1.

Example 3.4.2 (Auslander et al. 1963). For each positive integer n, the graph
A. is formed from n + 1 concentric cycles, each of length 4n, plus some
additional edges, as shown in Figure 3.35. There are 4n 2 "inner" edges that
connect the n + 1 cycles to each other and 2 n "outer" edges that adjoin
antipodal vertices on the outermost cycle.

Suppose that the n + 1 concentric cycles of the graph A. and the inner
edges are imbedded on a disk so that the vertices of the outermost cycle
appear equally spaced on the boundary of the disk. Suppose also that a
rectangular sheet with 2n lines from top to bottom has its left and right sides
identified so that it becomes a Mobius band. If the boundary of the Mobius
band is identified to the boundary of the disk so that the 2n lines become
images of the outer edges of A., then the result is an imbedding A.~ N1 •
Figure 3.36 shows that the graph A 1 contains a subdivision of K 3 3 • Since A.
contains a subdivision of A1 , for all n ~ 1, it follows that y(A.) ~ 1.
It is clear from Figure 3.36 that y(A 1) = 1. Suppose, by way of induction,
that y(A._ 1 ) ~ n - 1. Then let C be the innermost cycle of the graph A., as
it appears in Figure 3.35, and let A. ~ S be an orientable cellular imbedding.
First, assume that the cycle does not separate the surface S. Then let the
imbedding A~~ S' be the result of splitting the imbedding A.~ S along the
cycle C. By Theorem 3.4.7, y(S) = y(S') + 1. Since the graph A._ 1 is a
subdivision of a subgraph of the graph A., it follows from the induction
hypothesis that y(S') ~ n - 1. Thus y(S) ~ n.
On the other hand, one might assume that the cycle C does separate the
surface S. Since the graph A.- C is connected, the cycle C must bound a
142 Surfaces and Graph lmbeddings

Figure 3.36. A subdivision of the Kuratowski graph K3• 3

in the graph A1 .

to 5

to 7

Figure 3.37. A rotation proJection for the imbedding Yn ..... T.

face of the cellular imbedding An~ S. Now suppose that images of all the
edges on the n other concentric cycles of the graph An are deleted, and that
every resulting vertex of valence 2 is smoothed over. The result is a possibly
noncellular imbedding Yn ~ S. If every noncellular face of Yn ~ Sis replaced
by a 2-cell, then one obtains the imbedding Yn ~ T whose rotation projection
is shown in Figure 3.37. Obviously, y(T) 5: y(S).
An elementary face-tracing argument shows that the imbedding Yn ~ T
has only two faces, so that x(T) = 4n - 6n + 2 = 2- 2n, which implies that
y(T) = n. It follows that y(S) ~ n.
We conclude from this two-pronged inductive argument that y(An) ~ n.
Since there is only one outer face in the imbedding of An given by the rotation
projection of Figure 3.35, it follows from a calculation of Euler characteristic
that the imbedding surface has genus n. Thus y(An) = n.

3.4.6. Maximum Genus

The simple formula YM(G) = {J(G) for maximum crosscap number is estab-
lished by Theorem 3.4.3. The problem of calculating the maximum genus
YM(G) is not so easily solved, but N.H. Xuong has demonstrated that it is still
much easier than computing the genus y( G). The immediate goal is to
determine which graphs have orientable one-face imbeddings. A survey of
results on maximum genus up to the time of Xuong's characterization is given
by Ringeisen (1978).
The Imbedding Distribution of a Graph 143

Two edges are called "adjacent" if they have a common endpoint.

Lemma 3.4.8. Let d and e be adjacent edges in a connected graph G such that
G - d - e is a connected graph having an orientable one-face imbedding. Then
the graph G has a one-face orientable imbedding.

Proof Let V(d) = {u, v}, and let V(e) = {v, w}. First, extend the one-
face imbedding (G- d- e)-+ S to a two-face imbedding (G- e) -+ S by
placing the image of d across the single face. Of course, the vertex v lies on
both faces. Thus, if one attaches a handle from one face of ( G - e) -+ S to
the other, one may then place the image of edge e so that it runs across the
handle, thereby creating a one-face imbedding G-+ S'. 0

Example 3.4.3. Consider the complete graph K 5 as a Cayley graph with vertices
0, 1, 2, 3, 4. Then the spanning tree T with edges 01, 12, 23, and 34 has a
one-face imbedding in the sphere. By Lemma 3.4.8, the graph T' = T + 04 + 42
also has a one-face imbedding, as does the graph T" = T' + 20 + 03, as does
K 5 = T" + 31 + 14. Thus YM(K 5 ) = 3.

Lemma 3.4.9. Let G be a connected graph such that every vertex has valence at
least 3, and let G have a one-face orientable imbedding G -+ S. Then there exist
adjacent edges d and e in G such that G - d - e has a one-face orientable
imbedding.

Proof Let d be an edge of G whose two occurrences in the single

boundary walk of the imbedding G-+ S are the closest together, among all
edges of G. Then the boundary walk can be written in the form dAd- B, where
no edge appears twice in the subwalk A. Since the graph G has no vertex of
valence 1, the subwalk A is nonempty, so that it has a first edge e. By case ii
of Theorem 3.3.5, edge-deletion surgery on edge d in the imbedding G-+ S
yields a two-face imbedding (G- d) -+ S'. The boundary walks of the two
faces are A and B, and the edge e appears in both A and B. Thus, by case i of
Theorem 3.3.5, the result of edge-deletion surgery on e in the imbedding
(G- d) -+ S' is a one-face imbedding of G- d-e. 0

The "deficiency ~(G, T) of a spanning tree" T for a connected graph G is

defined to be the number of components of G - T that have an odd number
of edges. The "deficiency ~(G) of the graph" G is defined to be the minimum
of ~(G, T) over all spanning trees T.

Example 3.4.4. On the left in Figure 3.38 is a spanning tree T for a graph G
such that ~(G, T) = 3. On the right is a spanning tree T' for the same graph G
such that ~( G, T') = 1. Since the edge complement of any spanning tree for this
graph has 1 - 7 + 11 edges, an odd number, the deficiency of any spanning tree
must be at least one. Therefore ~(G)= ~(G, T') = 1.
144 Surfaces and Graph Imbeddings

Figure 3.38. Two spanning trees for the same graph, one of deficiency 3, the other of de-
ficiency 1.

Lemma 3.4.10. Let T be a spanning tree for a graph G, and let d and e be a
pair of adjacent edges in G- T. If ~(G- d- e, T) = 0, then ~(G, T) = 0.

Proof Every component of G - d - e - T that meets either of the edges

d ore has an even number of edges, since ~(G- d- e, T) = 0. The number
of edges in the component of G - T that contains the edges d and e is two
plus the sum of these even numbers. All the other components of G - T have
evenly many edges, as in G- d-e- T. Thus, ~(G, T) = 0. o

Lemma 3.4.11. Let G be a graph other than a tree, and letT be a spanning tree
such that ~(G, T) = 0. Then there are adjacent edges d and e in G- T such that
~(G- d-e, T) = 0.

Proof Let H be a nontrivial component of G - T. Since H is connected

and has at least two edges, there are adjacent edges d and e in H such that
H- d- e has at most one nontrivial component. (See Exercise 16.) Thus,
~(G- d-e, T) = 0. o

lbeorem 3.4.12 (Xuong, 1979). Let G be a connected graph. Then G has a

one-face orientable imbedding if and only if ~(G)= 0.

Proof As the basis for an induction, one observes that if #EG = 0, then
both clauses of the conclusion are trivially true. Next, assume that the
conclusion holds for any graph with n or fewer edges, and let G be a graph
with n + 1 edges.
As a preliminary, suppose that G has a vertex v of valence 1 or 2, and let
G' be the graph obtained by contracting an edge incident on vertex v. Then
obviously, the graph G has a one-face orientable imbedding if and only if the
graph G' does. Also, ~(G) = 0 if and only if ~(G') = 0. Since tfie graph G' has
one edge less than the graph G, it follows from the induction hypothesis that
G' has a one-face orientable imbedding if and only if ~(G') = 0. The conclu-
sion follows immediately.
In the main case, every vertex of G has valence 3 or more. Suppose first
that G has a one-face orientable imbedding. By Lemma 3.4.9, there exist
adjacent edges d and e in G such that G- d-e has a one-face orientable
imbedding. It follows from the induction hypothesis that ~( G - d - e) = 0,
The Imbedding Distribution of a Graph 145

o--o
Figure 3.39. Two graphs that have deficiency 2.

so there is a spanning tree Tin G- d- e such that ~(G- d-e, T) = 0. Of

course, the tree T also spans G. By Lemma 3.4.10, it follows that ~(G, T) = 0,
which implies that ~(G) = 0.
Conversely in the main case, one may suppose that ~(G)= 0, so that there
is a spanning tree T such that ~(G, T) = 0. Since G has no vertex of valence
1, it is not a tree. By Lemma 3.4.11, it follows that there are adjacent edges d
and e in G- T such that ~(G- d-e, T) = 0. Thus, ~(G- d-e)= 0. By
the induction hypothesis, the graph G - d - e has a one-face orientable
imbedding. By Lemma 3.4.8, so does the graph G. D

Example 3.4.5. The dumbbell graph G on the left of Figure 3.39 has only one
spanning tree, from which it follows that ~(G)= 2. It is easily verified that the
removal of any pair of adjacent edges from the graph G' on the right of Figure
3.39, followed by a sequence of edge contractions to eliminate vertices of valence
1 or 2, yields the dumbbell graph. Therefore by Lemma 3.4.9, it follows that
{3(G') = 2 also.

A graph is "k-edge-connected" if the removal of fewer than k edges from G

still leaves a connected graph. The t.wo graphs in Example 3.4.5 can be
generalized to obtain l-edge-connected and 2-edge-connected planar graphs
with arbitrarily large deficiency. On the other hand, Kundu (1974) and Jaeger
(1976) have shown that every 4-edge-connected graph G has a spanning tree T
whose edge complement G - T is connected. Thus if G is 4-edge-connected,
then ~(G) = 0 or 1, according to whether {3(G) is even or odd.

lbeorem 3.4.13 (Xuong, 1979). Let G be a connected graph. Then the mini-
mum number of faces in any orientable imbedding of G is exactly ~(G) + 1.

Proof An equivalence to the conclusion is the statement that the graph G

has an orientable imbedding with n + 1 or fewer faces if and only if ~(G) ~ n.
The proof of this equivalent statement is by induction on the number n. It
holds for n = 0, by Theorem 3.4.12, and we now assume that it holds for all
values of k less than n, where n > 0.
First, we suppose that G --+ S is an orientable imbedding with #FG = n + 1.
Then perform edge-deletion surgery on an edge e common to two faces of the
146 Surfaces and Graph Imbeddings

imbedding G--+ S. By case i of Theorem 3.3.5, the resulting imbedding

( G - e) --+ S' has n faces. From the induction hypothesis, it follows that
~(G- e)~ n- 1. Therefore, ~(G)~ n.
Conversely, one may suppose that ~(G)= n, so that there is a spanning
tree Tin G such that ~(G, T) = n. Let H be a component of G- T with an
odd number of edges. Certainly the subgraph H has an edge e that does not
disconnect H or such that one endpoint of e has valence 1 in H. Accordingly,
HG- e, T) = n - 1. By the induction hypothesis, the graph G- e has an
orientable imbedding with at most n faces. Therefore the graph G has an
orientable imbedding with at most n + 1 faces. 0

Corollary (Xuong, 1979). Let g be a connected graph. Then YM(G) =

!(~(G)- ~(G)).

Proof Let g = YM(G). Then 2- 2g = #V- #E +(~(G)+ 1), by The-

orem 3.4.13. It follows that g = !(~(G)- ~(G)). 0

Example 3.4.6. Let T be the tree in the complete graph K n consisting of all
edges incident on a particular vertex. Then

1"f ( n -2 I) 1s. even

if ( n ; 1 ) is odd

or equivalently,

if n =1 or 2 modulo4
if n =0 or 3 modulo 4

The obvious computational problem presented by Theorem 3.4.13 is to

calculate the deficiency of a graph. The number of spanning trees is exponen-
tial, and no polynomial-time algorithm was found in the immediate years after
Xuong published his characterization. Ultimately, Furst, Gross, and McGeoch
(1985a) developed a polynomial-time algorithm involving a reduction to
matroid parity.

3.4.7. Distribution of Genus and Face Sizes

Suppose that a graph G has vertices v1, .•. , vn of respective valences d 1, ... , dn.
Then the total number of orientable imbeddings is
n
n (d;- I)!
i=l
The Imbedding Distribution of a Graph 147

Gross and Furst (1985) raise the problem of determining exactly how many of
these imbeddings lie in each surface of the genus range. From such a
determination, one would immediately know the minimum genus and the
maximum genus. Furst, Gross, and Statman (1985b) combined topological
methods and enumerative techniques to obtain a genus distribution formula
for two infinite classes of graphs with expanding genus ranges. Gross, Robbins,
and Tucker (1986) calculated the genus distributions of bouquets of circles,
with the aid of a formula of Jackson (1986).
Another problem raised by Gross and Furst (1985) is to compute the
distribution of face sizes over all imbeddings of a graph, which appears to be
harder than computing genus distributions. They introduce an extensive
hierarchy of invariants of the imbedding distribution of a graph, in which the
genus distribution and the face-size distribution lie near the bottom. Higher
invariants keep track of the distribution of face sizes within each imbedding
and the relationships between the faces.
The set of imbeddings of a graph has various statistical properties of
interest. For instance, all the genus distributions mentioned above are strongly
unimodal. Also, Gross and Klein (1986) classified the graphs of average genus
less than or equal to 1.
Little else is known about the number of imbeddings of a given graph in a
given surface, except when the imbedding surface has low genus. The proof of
Theorem 3.4.6 shows that the complete graph K 7 has 2 · 5! different im-
beddings in the torus (once a rotation at vertex number 1 is fixed, there are
only two choices for the imbedding). By Whitney's theorem (Theorem 1.6.6),
3-connected planar graphs have "only one" equivalence class of imbeddings in
the plane. Negami (1983, 1985) has investigated the uniqueness of imbeddings
in the torus and the projective plane.

3.4.8. Exercises
1. Show that the sequence of rotation projections for the complete graph K 5
given in Figure 3.40 are adjacent in pairs from left to right and that the
associated surfaces have genus 1, 2, and 3 (again from left to right).

Figure 3.40. Three rotation projections of the complete graph K 5 •

148 Surfaces and Graph lmbeddings

2. Draw a sequence of rotation projections for K 5 in surfaces of crosscap

number 1, 2, 3, 4, 5, 6 such that each imbedding is obtained from the
previous imbedding by edge-twisting surgery.
3. Draw a sequence of rotation projections for adjacent imbeddings of K 6
in orientable surfaces of genus 1, 2, 3, 4, 5.
4. Write a sequence of rotation systems in list format for adjacent im-
beddings of the octahedral graph 0 3 in surfaces of genus 0, 1, 2, 3.
S. The "Euler characteristic range" ER(G) of a connected graph G is
defined to be the set of numbers c such that G has a cellular imbedding
in a surface of Euler characteristic c. Prove that ER(G) is an interval.
6. Prove that ER(G) = [c, c] if and only if {3(G) 5: 1. Conclude that if
ER(G) = [c, c], then c = 2.
7. Give a genus-1 imbedding of the graph Yn shown in Figure 3.37.
8. Let G be connected and not a tree, and let k be the smallest integer such
that G imbeds, cellularly or not, in the nonorientable surface Nk. Prove
that G has a cellular imbedding in Nk.
9. Prove that the maximum genus of the complete bipartite graph K m n is
l/3(Km,n)/2J. (Ringeisen, 1972) '
10. Find the maximum genus of the octahedral graph On, for all n > 0.
11. Let e be an edge of the graph G such that G- e has two components G1
and G2 • Prove that y(G) = y(G 1 ) + y(G 2 ) and that YM(G) = YM(G 1 ) +
YM(G 2 ). (Hint: Use Exercise 3.3.4 and Theorem 3.3.3.)
12. Find a graph that has genus range [0, 1]. Find a graph that has genus
range [1, 2]. Use Exercise 11 to prove that for any m ~ 0 and any
n ~ 2m, there is a graph with genus range [m, n].
13. Prove that GR(G) = [m, m], m ~ 0, if and only if no vertex of the graph
G lies on two distinct cycles. (Hint: Exercise 7 might help.)
14. Use Exercise 13 to prove the result of Nordhaus et al. (1972) that if
GR(G) = [m, m], then m = 0.
15. By Exercise 12, there is a graph with genus range [2, n ], for any n ~ 4. By
Exercise 14, there is no graph with genus range [2, 2]. Is there a graph
with genus range [2, 3]? In general, find a graph of genus range [m, n] for
some m > 0, m < n <2m.
16. Prove that if His a connected graph with more than one edge, then there
are adjacent edges d and e such that H - d - e has only one nontrivial
component.
17. Calculate the genus distribution of the following graphs:
(a) K4•
(b) KJ,J·
(c) K 2 XK 3 •
Algorithms and Fonnulas for Minimum Imbeddings 149

18. Calculate the face-size distribution of the following graphs:

(a) K 4 •
(b) K3,3.
(c) K 2 X K 3 •

3.5. ALGORITHMS AND FORMULAS FOR MINIMUM IMBEDDINGS

Historically, most of the early emphasis in topological graph theory was to

determine the genus and crosscap number of a graph. Since this problem is
inherently finite, one might look for algorithms to compute them. One possibil-
ity, which is incorporated into the Heffter-Edmonds algorithm, requires an
examination of all the rotation systems, an unfeasibly lengthy search.
The most frequently adopted approach has been to calculate genus and
crosscap number formulas for interesting families of graphs and for the effect
of various graph operations. Most of the derivations of such formulas have
used covering space constructions, either in explicit topological form or in an
equivalent combinatorial guise. However, before extending voltage graph
techniques to include a lift of the imbedding-the main theme of Chapter
4-it is instructive to study two operations whose imbeddings have not
required covering space constructions: cartesian products and one-vertex
amalgamations.

3.5.1. Rotation-System Algorithms

In its least complicated form, the Heffter-Edmonds algorithm for minimum
genus is as follows: First, list all the pure rotation systems of a graph. Next,
compute the number of faces for each rotation system. Then, choose one
having the most faces. The obstruction to usefulness is that the number of
rotation systems is exponential in the number of vertices. For instance, a
regular (r + I)-valent graph with n vertices has (r!)n pure rotation systems.
Sampling from the set of all pure rotation systems in order to approximate
the genus of a given graph seems a plausible approach. However, not much is
known about the distribution of genera within a rotation system, except for a
few special cases. It seems that relatively few imbeddings are minimum genus
imbeddings. For example, the genus range of the complete graph K 7 is [1, 7],
and by details in the proof of Theorem 3.4.6, we see that only 2 · 5! of (S!f
possible pure rotation systems give a genus-1 imbedding, or about one in 10 12 •
A heuristic explanation of this phenomenon is given in Exercises 2-4.
A less exhaustive use of rotation systems might be to hill-climb by moving
from one imbedding of the given graph to an adjacent imbedding (as defined
in Section 3.4), while always trying to increase the number of faces at each
step. Unfortunately, as the following theorem shows, "local genus minima"
abound in the collection of orientable imbeddings of a graph.
150 Surfaces and Graph Imbeddings

Figure 3.41. Two imbeddings of a graph.

lbeorem 3.5.1. Let G-+ S be an imbedding such that every boundary walk is a
cycle, and such that the only boundary walks passing through both endpoints of
an edge are the two boundary walks containing that edge. Then every aqjacent
imbedding has larger genus.

Proof Let e be any edge of the graph G, and let G' -+ S' be the
imbedding obtained by deletion surgery on edge e. This imbedding has one
less face than the original imbedding, since the two faces containing edge e
have been joined into one. Moreover, no face of the imbedding G' -+ S' passes
through both endpoints of edge e, except the joined face. Therefore, when the
edge e is reattached in a different way, it must form a bridge joining two more
faces. Hence, any adjacent imbedding has fewer faces and larger genus. D

Example 3.5.1 (Gross and Tucker, 1979b). It is easy to verify that the toroidal
imbedding on the left of Figure 3.41 satisfies the hypothesis of Theorem 3.5.1.
Thus, every adjacent imbedding has larger genus. However, the given graph has
the planar imbedding given on the right. It follows that the imbedding at the left
is a local minimum that is not a global minimum.

One might try to salvage an algorithm by defining a different kind of

adjacency for graph imbeddings. One obvious candidate is" vertex-adjacency":
two imbeddings are vertex-adjacent if their rotation systems agree at every
vertex except one. However, a regular graph of valence 3, such as that in
Example 3.5.1, has only two rotations at any vertex, so that vertex-adjacent
imbeddings of these graphs would also be adjacent, presenting us with the
same problem of local genus minima. In general, algorithms based on rotation
systems have not been sufficiently fast.

3.5.2. Genus of an Amalgamation

Suppose that H and K are graphs imbeddable in surfaces Sh and Sk,
respectively, and that G is a graph obtained by amalgamating Hand K at a
vertex. Then there is no difficulty in constructing an imbedding of G into a
surface of genus h + k. First, one chooses a disk on Sh, whose interior is
disjoint from the imbedded graph H and whose boundary intersects H only at
Algorithms and Formulas for Minimum Imbeddings 151

Figure 3.42. Any amalgamation of two graphs is imbeddable into a disk sum of any two of their
respective imbedding surfaces.

the amalgamation vertex. Next, one chooses an analogous disk on the surface
Sk. Then, one deletes the interiors of both disks and identifies the resulting
surface boundaries so that the graphs H and K are amalgamated, as il-
lustrated in Figure 3.42. As we have mentioned in Exercise 3.3.11, this
composition of surfaces is called a "disk sum".
For this simplified type of graph amalgamation, we often use the notation
H *v K, where v is the vertex of amalgamation, to represent the resulting
graph. From the above discussion, it is obvious that

The so-called BHKY theorem (for Battle, Harary, Kodama, and Youngs)
asserts the nonobvious fact that

y(H *v K) = y(H) + y(K)

By way of contrast, the analogous assertion, is not true for maximum genus
(see Exercise 5), and it is not true for crosscap number (see Theorem 3.5.5).
The straightforward way to try to prove the BHKY theorem is to attempt
to show that an amalgamated graph H • v K must have a minimum imbedding
that permits reversal of the combined disk sum and graph amalgamation
operation. That is, one hopes for an imbedded circle in the surface that
intersects H *v K only at the amalgamation vertex v and that separates the
imbedding surface into two parts, one of which contains H and the other of
which contains K. Necessarily, all the H-edges would have to be grouped
contiguously in the rotation at v (as would the K-edges). If one started with a
minimum imbedding of H *v K in which the H-edges were not continuously
grouped, then one would wish to regroup, without increasing the genus of the
imbedding surface. The following lemma, whose application is not restricted
to imbeddings of amalgamated graphs, is what facilitates regrouping.

Lemma 3.5.2. Let

v. ab .. · cd
be the rotation at a vertex v of an orientable imbedding G --+ S such that the
corners ab and cd lie on the same face. Let G --+ S' be the adjacent imbedding
152 Surfaces and Graph Imbeddings

such that the edge a is inserted between edges c and d, so that vertex v has the
resulting rotation

v. b ... cad

and all other rotations remain the same. Then y(S') 5: y(S).

Proof Let G -+ S" be the imbedding obtained from G -+ S by deletion

surgery on edge a. We shall use F, F', and F" to denote the face sets of the
respective imbeddings G-+ S, G-+ S', and G-+ S". If the edge a appears
twice on the same face boundary of the imbedding G-+ S, then #F" = #F
+ 1, from which it follows that #F' ~ #F" - 1 = #F. Otherwise, the dele-
tion surgery on edge a in the imbedding G -+ S unites the face having the
corners ab and cd with another face, and the resulting "new" face in G -+ S"
contains the corner cd plus another occurrence of v across the "new" face to
the corner cd, thereby resplitting the "new" face. Thus, #F' = #F. D

Theorem 3.5.3 )The BHKY Theorem (Battle, Harary, Kodama, and Youngs,
1963)). y(H•vK) = y(H) + y(K).

Proof In any imbedding of H *v K, the rotation at vertex v contains

sequences of edges from the graph H interspersed with sequences from K.
Maximal sequences from Hand K are called "H-segments" and "K-segments",
respectively, relative to that imbedding. Let H *v K-+ S be a minimum-genus
imbedding that maximizes the sum of the squares of the segment lengths at v,
taken over all imbeddings of H *v K into the surfaceS. We assert that for this
imbedding, there is but one H-segment and one K-segment.
In order to verify the assertion, we observe that in any imbedding of
H *v K, there exists an edge d from H and an edge e from K such that d
immediately precedes e in the rotation at v. Moreover, whatever face boundary
contains the corner de must also contain a corner e'd' where e' is from K and
d' from H. Since edges of Hand K meet only at v, it follows that the rotation
v has the form

v. de ... e'd' ...

We claim that the edges d and d' must lie in the same H-segment. If not, then
suppose that the lengths of the H-segments containing d and d' are n and n',
respectively. Without loss of generality, we shall assume that n 5: n'. Apply
Lemma 3.5.2, with edges d, e, e', and d' filling the roles of a, b, ~' and d,
respectively. Evidently, the adjacent imbedding with the modified rotation

v. e ... e'dd'

(and no other changes) would have the same imbedding surface-lower genus
is impossible, since S is a minimum-genus imbedding surface-but a larger
Algorithms and Fonnulas for Minimum Imbeddings 153

sum of squares of segment lengths, since

because n ~ n'. Thus, the edges d and d' are in the same H-segment.
Similarly, the edges e and e' are in the same K-segment. Therefore, we have
established an assertion that there is only one H-segment and only one
K-segment.
We must still prove that a closed arc C through vertex v from corner de to
corner e'd' in the interior of their common face must separate the imbedding
surface. However, if not, then closed arc C would lie on a handle of surfaceS.
If the surface S were cut open along C and reclosed with two disks, the result
would be an imbedding of the disjoint union of the graphs H and K on a
surfaceS' of genus y(S) - 1. Such an imbedding would be noncellular, since
any region that contained both a closed H-walk and a closed K-walk in its
boundary would fail to be simply connected. By the construction of S', no
such region could arise. It follows that the closed arc C separates the surface
S. Since graph H lies on one side of the separation, that side must have genus
at least y(H). Similarly, the other side must have genus at least y(K). It
follows that

Since we have previously established the opposite direction of this inequality,

the BHKY theorem is proved. 0

A maximal 2-connected subgraph of a graph G is called a "block" of the

graph G. For example, the graph in Figure 3.43 has four blocks. A slight
generalization of Theorem 3.5.3 is the following.

lbeorem 3.5.4. If the blocks of graph G are G1, ... ,Gn, then

Proof One uses a simple induction on the number of vertices of the graph
G, while applying the fact that a subgraph His a block graph G if and only if
G = H *v K for some vertex v and some subgraph K (of course, K itself may
have several blocks). 0

Figure 3.43. A graph having four blocks.

154 Surfaces and Graph Imbeddings

Figure 3.44. A rotation projection of K 5 *• K 5 .

More general types of amalgamations have also been studied. The situation
here is much more complicated, as the following example illustrates.

Example 3.5.2. The rotation projection shown in Figure 3.44 gives an im-
bedding in a torus of an amalgamation of K 5 with K 5 along an edge. Thus the
genus of this graph is I, not the seemingly obvious value of 2.

The results known for other amalgamations are all of the form y( H • 1 K)
= y(H) + y(K) + B, where the parameter B is given a possible range of
values, and each of these values does occur for some pair of graphs Hand K.
Alpert (1973) shows that when the pasting function f identifies an edge of H
with an edge of K, then B = 0 or -1. Stahl (1980) and Decker et al. (1981)
prove that B = 1, 0, or -1 when f identifies copies of K/. Stahl also shows in
the same paper that B = 2, 1, 0, -1, or - 2, when the pasting function f
amalgamates two copies of K 3c.
In general amalgamations, the range of B values is much greater and
determining when B takes on a particular value is quite complicated [see, for
example, Decker et al. (1985)]. The best general theorem would be a bound on
B that depends only on the number k of vertices being identified by the
amalgamation. An upper bound for B if k = 1 is easily obtained and is
achieved for some amalgamations when k = 2. Archdeacon (1986) has shown,
however, that there is no lower bound for B even when k = 3; he constructs a
sequence of n-vertex graphs Gn and 3-vertex amalgamations Hn of Gn with
itself such that y(Hn) = 2y(Gn) + n.

3.5.3. Crosscap Number of an Amalgamation

Computation of the crosscap number of a one-vertex amalgamation of two
graphs is complicated by the fact that a nonorientable surface can be rep-
resented as the disk sum of a nonorientable surface and an orientable surface.

Example 3.5.3. The crosscap number of ·the complete graph K 7 is 3, by

Theorem 3.4.6. However, since K 7 is imbeddable in a torus, the amalgamation
Algorithms and Fonnulas for Minimum Imbeddings 155

K 7 • v K 7 has an imbedding in the nonorientab/e surface N5 obtained from a disk

sum of imbeddings of K 1 in a torus and in the surface N 3 • Thus, in general,
y(H *v K) -4= y(H) + y(K).

Theorem 3.5.5 (Stahl and Reineke, 1977). y( H • v K) = y( H) + y( K) + 8,

where 8 = -1 if either y(H) > 2y(H) or Y(K) > 2y(K), and 8 = 0
otherwise.

Proof Let g denote the value of the right-hand side of the equality and let
G = H•vK. To see that y(G) ~ g, use disk sums and the general fact that
y(H) ~ 2y(H) + 1 (and similarly forK).
Conversely, to show that y(G) ~ g, it suffices to prove, as in Theorem 3.5.3,
that there is an imbedding of the graph G in the nonorientable surface of
minimum crosscap number such that the rotation at vertex v has one H-seg-
ment and one K-segment. The details of the proof of Theorem 3.5.3 can be
used again. The only difficulty lies in the nonorientable version of Lemma
3.5.2. The type of the edge d might have to be changed when it is reattached
at vertex v, in order not to decrease the number of faces. (Certainly one of the
two possible types for edge d will be wrong: we want the sides of edge d in
the new imbedding to belong to separate faces, yet an extra twist in an edge
whose sides belong to separate faces joins the faces.) One might fear that
changing the type of edge d could make the new imbedding surface orientable.
However, in the present context, the sides of the edge d belongs to distinct
faces in the original imbedding. As we observed in Section 3.4 in the
discussion of extra twists, changing the type of edge under such a circumstance
always yields a nonorientable surface. D

3.5.4. The White-Pisanski Imbedding of a Cartesian Product

The cartesian product of two graphs contains many cycles of length 4, because
each instance of the product K 2 X K 2 is a cycle of length 4. One might hope,
therefore, that a cartesian product G X H has imbeddings in which all the
faces are quadrilaterals. If the graphs G and H both have girth greater than 3,
then the resulting imbeddings would be minimal genus or crosscap number
imbeddings. Indeed, it might be much easier to compute the genus of a
product G X H than to compute the genus of either G or H. White (1970) was
the first to exploit this idea. Our presentation includes generalizations due to
Pisanski (1980).
Define a "patchwork" of an imbedding G ~ S to be a collection P of faces
such that the union of all the boundary walks of the faces in the collection P
forms a 2-factor of the graph G. A set of patchworks P 1, P2 , ••• , P, is
"disjoint" if they are pairwise as collections of faces, that is, no face in P; is
also a face in lj unless i = j. (However, the boundary walk of a face in P; may
share an edge with a boundary walk of a face in lj.) The faces of an
imbedding G ~ S that are not in a given set of patchworks are called
 -
156 Surfaces and Graph Imbeddings

1 2 1 2 1 2

4 3 4 3 4 3

f 1 2 1 2 1 2 t
3 4 3 4 3 4

Figure 3.45. -
Some disjoint patchworks in a toroidal imbedding of C4 X C6 .

"residual". A patchwork is called "even" if all of its faces have even size, and
it is called "quadrilateral" if all of its faces have size 4.

Example 3.5.4. Let P; be the collection of faces labeled i in the toroidal

imbedding given in Figure 3.45. Then P 1, .•. , P4 are four disjoint quadrilateral
patchworks; this set of patchworks has no residual faces. The set P 1, P2 of
patchworks has 12 residual faces.

An "edge-coloring" of a graph G is an assignment of a "color", usually

denoted by an integer, to each edge such that edges incident to the same vertex
have different colors. A color i used in an edge-coloring of the graph G is
"saturated" if there is an i-colored edge incident to every vertex of G. Thus,
the set of edges of a saturated color in a given edge-coloring of the graph G
form a 1-factor of the graph. A regular graph of valence p has an edge-color-
ing with p colors if and only if it is !-factorable.

Example 3.5.5. On the left of Figure 3.46 is an edge-coloring of the cycle graph
C6 using two colors, both of which are saturated. On the right is an edge-coloring
of the graph C5 using three colors, none of which is saturated.

Theorem 3.5.6 (White, 1970; Pisanski, 1982). Let G ~ S be an imbedding that

has r disjoint, even patchworks such that all residual faces are quadrilaterals and
such that at least q of the patchworks are quadrilateral. Let H be a graph that
has an edge-coloring with r colors, at least r- q of which are saturated. Then the
cartesian product G X H has an imbedding with at least 2r - q disjoint patch-
works, such that all faces of the imbedding are quadrilaterals. The imbedding is
orientable if and only if the imbedding G ~ S is orientable and the graph H is
bipartite.
Algorithms and Fonnulas for Minimum Imbeddings 157

2
Figure 3.46. Edge-colorings of the graphs C6 and C5 .

Proof Let 1, 2, ... , r denote the colors of the edge-coloring of the graph H
with the colors that are not saturated listed first. Let P 1, P2 , ••• , P, be the
disjoint even patchworks of the imbedding G ~ S with the quadrilateral
patchworks listed first. We construct the desired imbedding for the graph
G X H by giving its rotation at each vertex ( u, v ). Recall that edges incident
to vertex (u, v) are of the form (d, v) where d is an edge of the graph G
incident on vertex u or of the form ( u, e) where e is an edge of the graph H
incident on vertex v. For convenience, we call the former G-edges and the
latter H-edges. We call the edges d and e their projected edges. The G-edges
are to appear in the rotation at vertex ( u, v) in the same order as their
projected edges in the rotation at vertex u for the imbedding G ~ S. The
H-edges are then interspersed as follows: if the projection of an H-edge is
colored i in the edge-coloring of the graph H, then that edge is placed between
the two G-edges whose projections form the corner of the face from patchwork
P; at vertex u. The type of every G-edge is the same as its projected edge while
every H-edge is given type 1. It is best to view this imbedding as #V(H)
copies of the imbedding G ~ S placed at different levels with "tubes" joining
corresponding faces at different levels where needed for H-edges. Figure 3.47
shows how a pair of H-edges having the same projection fit into four corners
of corresponding faces from patchwork P;. A pair of corresponding faces of

Figure 3.47. Part of a rotation projection of the imbedding for G X H; on the left, before
H-edges are added and, on the right, after.
158 Surfaces and Graph Imbeddings

Figure 3.48. A pair of corresponding faces JOined by a tube.

the imbedding G ~ S is thereby transformed into a tube of quadrilaterals (see

Fig. 3.48).
If color i is not saturated, then the patchwork P,, which is quadrilateral by
the choice of listing of the patchworks, generates a quadrilateral patchwork
consisting of alternate quadrilaterals around each tube between levels joined
by H-edges colored i and faces of P; itself at levels where no H-edge colored i
is incident. If color i is saturated, then patchwork P; generates a pair of
disjoint quadrilateral patchworks formed by alternate quadrilaterals around
each tube joining a pair of corresponding faces of P;. Thus the resulting
imbedding has at least 2r- q disjoint patchworks, as claimed. It is clear that
every face of the imbedding is quadrilateral, by construction. Finally, the
imbedding surface is orientable if and only if the imbedding G ~ S is
orientable (since G-edges have the same type as their projections), and the
graph H has no cycle of odd length (since every H-edge has type 1). D

The important part of the conclusion of Theorem 3.5.6 is that the con-
structed imbedding has all faces quadrilateral. The number of disjoint patch-
works in the imbedding is useful for repeated cartesian products.

3.5.5. Genus and Crosscap Number of Cartesian Products

The White-Pisanski imbedding enables us on compute the genus and crosscap
number of a variety of cartesian products. In all the applications that follow,
the minimal genus or crosscap imbedding has all faces quadrilateral. Thus the
number of faces is half the number of edges. In particular, if the graph is
regular of valence t and has n vertices, the genus of the imbedding surface S,
if it is orientable, is

1 - x(S)/2 = 1 - (n- nt/2 + nt/4)/2 = 1 + n(t- 4)/8

If the imbedding surface is nonorientable, its crosscap number is twice this.

Algorithms and Formulas for Minimum lmbeddings 159

Theorem 3.5.7 (White, 1970). Let G = en, X en 2 X · · · X en, where r > 1

and n, > 3 for all i. Then

a. y(G) = 1 + #V(G)(r- 2)/4 if each n, is even.

b. Y(G) = 2 + #V(G)(r- 2)/2 if r > 2, n 1 and n 2 are even, and n 3 is
odd.

Proof The graph en, X en 2 has an imbedding in the torus that has all
faces quadrilateral and four disjoint patchworks (see Example 3.5.4), whenever
n 1 and n 2 are even. Repeated application of Theorem 3.5.6 gives an all-
quadrilateral imbedding of the graph G, orientable in case a and nonorientable
in case b. Since the girth of the graph G is greater than three, the equations
follow from Euler's equation. 0

Theorem 3.5.8 (Ringel, 1955). The genus of the n-cube graph Qn is 1 +

2n- 3(n- 4).

Proof The proof is like that for Theorem 3.5.7; use the quadrilateral
imbedding of Q2 = e4 to get started. 0

Theorem 3.5.9 (Pisanski, 1980). Let G and H be 1-factorable, regular r-ualent

graphs with girth greater than three. If both G and Hare bipartite, then

y(G X H)= 1 + mn(r- 2)/4

and if G and H are not both bipartite, then

y(G X H)= 2 + mn(r- 2)/2

where m = #V(G) and n = #V(H).

Proof First, we construct an imbedding G ~ S that has r disjoint even

partitions (and hence no residual faces). Choose any cyclic ordering of the
1-factors of a !-factorization of the graph G, and let the rotation at each vertex
be given by this single ordering. Let every edge have type 1. Then each face of
the resulting imbedding consists of edges alternating between two 1-factors.
Thus, the imbedding has r disjoint even patchworks. The imbedding surface is
orientable if and only if the graph G is bipartite. The theorem then follows
from Theorem 3.5.6, since the graph H has an edge-coloring with r colors all
of which are saturated. o

Theorem 3.5.10 (Jungerman, 1978a). y(Qn) = 2y(Qn) for n * 4, 5 and

y(Qn) = 2y(Qn) + 1 for n = 4, 5.
160 Surfaces and Graph Imbeddings

Figure 3.49. Rotation projections for two imbeddings of Q3 •

Proof Since y(Qn) = 0 for n < 4, there is nothing to be proved there.

Jungerman shows that neither Q4 nor Q5 has a nonorientable quadrilateral
imbedding, but they do have orientable ones, hence the special equation for
n = 4, 5. We shall show that Q4 has a nonorientable imbedding with two
disjoint patchworks and all residual faces quadrilateral. Theorem 3.5.6 then
provides a nonorientable, quadrilateral imbedding of Q4 X Q 2 = Q6 with four
disjoint patchworks, from which one easily obtains quadrilateral imbeddings
of Qn for all n > 6, again by Theorem 3.5.6.
Consider the two imbeddings of Q3 whose rotation projections are given in
Figure 3.49. Both imbeddings have three disjoint patchworks consisting each
of faces with the same label; the imbedding on the right has two faces of size
8, each of which forms a patchwork by itself. We have labeled the corners of
those faces. Now perform the White-Pisanski construction on Q3 X Q1 = Q4 ,
only using the left imbedding of Q3 on one level and the right imbedding on
the other level and putting Qcedges into the corners of faces labeled 1. Faces
labeled 2 then form a patchwork in the resulting imbedding, while those
labeled 3 form another. Residual faces are the quadrilaterals on the tubes
joining faces labeled 1. Finally, the reader can easily verify that the resulting
imbedding has a cycle with an odd number of type-1 edges. D

Clearly an edge-coloring of a graph with maximum valence r requires at

least r colors. The following converse has been established by Vizing (1964).

Theorem 3.5.11 (Vizing). Any graph with maximum valence r can be edge-col-
ored with r + 1 colors or less.

Proof Omitted. D

3.5.6. Exercises
1. Prove that given any integer m, there is a graph having an edge whose
contraction reduces the genus of the graph by m. (Hint: Any n-vertex
graph reduces to a one-vertex graph by a sequence of n - 1 edge
con tractions.)
Algorithms and Fonnulas for Minimum lmbeddings 161

2. Given a pure rotation system for a graph G, let p be the permutation of

the edge set DE(G) such that for each vertex v the set of directed edges
with initial vertex v, ordered by the given rotation system, forms a cycle
of p. Let t be the involution of DE(G) given by t(e 0 ) = e-a, o = ±.
Show that the cycles of p o t are the directed boundary walks of the
associated orientable imbedding.
3. For each permutation w E .9;,, define t( w) to be the permutation ob-
tained by erasing the symbol n in the cycle notation for w. For example,
if w = (1 3 4 6 2)(5) then t( w) = (1 3 4 2)(5). Use t to prove that the
expected number of cycles in an element of .9;, is 1 + 1/2 +
1/3 + · · · + 1/n [see Knuth (1969) for more on cycle structure of
random permutations].
4. Use Exercises 2 and 3 to argue that one should expect the number of
faces of a random orientable imbedding of an m-edge graph to be around
ln(2m ). Contrast this with the number of faces of a "good" (e.g.,
triangular) minimal-genus imbedding. What should one expect for the
genus of a random imbedding of K 7 ?
5. Prove that YM(K 7 *v K 7 ) = YM(K 7 ) + YM(K 7 ) + 1.
6. Find graphs Hand K such that YM(H *v K) = YM(H) + YM(K).
7. Prove that YM(H *v K) = YM(H) + YM(K) + 8, where 8 = 0, 1.
8. For each possible value of 8, give an example of a pair of graphs H and
K and a function f identifying copies of K 2 in the graphs such that
y(H *t K) = y(H) + y(K) + 8.
9. Prove that if G and H are regular bipartite graphs of the same valence
r having m and n vertices, respectively, then y( G X H) = 1 +
mn(r- 2)/4.
10. Let H be any bipartite graph with maximum valence r and girth greater
than 3. Use Vizing's theorem and Theorem 3.5.6 to compute the genus of
Qn X H for all n > r.
11. Prove that if His any graph with maximum valence r > 3, then Qn X H
has a nonorientable quadrilateral imbedding for all n > r + 2. (Hint:
The proof of Theorem 3.5.10 provides nonorientable quadrilateral imbed-
dings of Qn with many disjoint patchworks for n > 5.)
12. Compute y(K 3• 3 X Qn) for all n.
4
Imbedded Voltage Graphs and
Current Graphs
Suppose that a graph G is simultaneously given a voltage assignment a and an
imbedding in some surface S. Then there is a natural way to construct a
derived surface sa and an imbedding of the derived graph Ga into sa; the key
idea is to "insert a region" into every closed walk of Ga that covers a
boundary walk of the "base" imbedding G ~ S. Moreover, there is a surface
map sa ~ S that is topologically a "branched covering" and whose restriction
to the derived graph Ga is the natural covering projection p: Ga ~ G. This
form of the voltage graph construction makes it easy to prove various
theoretical results, including Alexander's theorem that every closed orientable
surface is a branched covering space over the sphere. Furthermore, imbedded
voltage graphs and their duals, which are known as "current graphs", are of
primary importance in many calculations of the genus of graphs, because they
provide a powerful technique for generating complicated imbeddings from
simple ones. Finally, imbedded voltage graphs can be used to understand
"surface symmetry", by which we mean group actions on surfaces.

4.1. THE DERIVED IMBEDDING

Let G -; S be an imbedding of a graph G in a surface S, and let a be a

voltage assignment in a group .91. We call the pair (G ~ S, a) an "imbedded
voltage graph". The surface S may be orientable or not, and the voltages may
be ordinary or permutation voltages, although one usually employs the sub-
scripted notation ( G ~ S, a)n to indicate permutation voltages in Y,. Then
there is a natural way to define a "derived surface" sa and a "derived
imbedding" Ga ~ sa, either in terms of a lifted rotation system or in terms of
lifted faces.

4.1.1. Lifting Rotation Systems

Suppose first that a is an ordinary voltage assignment in a group .91, and let
p: Ga ~ G be the natural projection of the derived graph onto the base graph.
We call the given imbedding G ~ S of the base graph the "base imbedding".
162
The Derived Imbedding 163

We shall initially assume that the base imbedding is specified by a rotation

system for the base graph G, and specify how to construct the "derived
rotation system".
Suppose that the vertex u of the base graph G has the rotation
u. cd ... e
Then a vertex u a in the fiber over u has the rotation

If some edge incident on u has type 1 in the base imbedding (i.e., if the local
orientations at its endpoints are inconsistent), then the corresponding edge in
the derived graph is also assigned type 1. The surface sa induced by this
rotation system on the derived graph is called the "derived surface", and the
imbedding oa --+ sa is called the "derived imbedding".
This definition of the derived imbedding translates immediately into a band
decomposition. Each 0-band of the base imbedding lifts to #.9.1 different
0-bands in the derived imbedding, each of which has the same instructions (up
to .1<1-coordinates) for attaching 1-bands. Moreover, each 1-band of the base
imbedding lifts to #.9.1 different 1-bands, each with the same "twist" as the
base 1-band.

Example 4.1.1. Consider the graph imbedding G--+ S corresponding to the

rotation system
u. a1 c d
v. b d c
w. b e1 a1 e1

with voltages in ~2 , as indicated by the labeled rotation projection at the bottom

of Figure 4.1. Then the derived rotation system is

Uo. a 01 co do
u 1• a 11 c1 d1
Do. b1 do c1
v 1• bo d1 Co
Wo. bo e 11 a 01 e01
w1. b1 e0 1 a 11 e 11

It is illustrated by the rotation projection at the top of Figure 4.1.

4.1.2. Lifting Faces

The base imbedding in Example 4.1.1 has three faces. The boundary walks of
these faces and their net voltages are as follows:
0 + 1 + 1- 1 = 1 mod2
0 - 1 + 1 - 0 = 0 mod 2
1-0=1mod2
164 Imbedded Voltage Graphs and Current Graphs

0 1
Figure 4.1. Rotation projections for an imbedded voltage graph (bottom) and its derived
imbedding (top).

The derived imbedding has four faces and these boundary walks:

ao eo bl cl al el ho co
bl- -
do el ao
dl b0 - eo al
co d1 - cl d0-

Obviously, this information might be obtained by means of the Face Tracing

Algorithm, an arduous task if the derived graph is large. Fortunately, however,
the number and sizes of the faces of the derived imbedding are always
computed using general properties of the voltage graph construction, proper-
ties which are apparent in the pattern of boundary walks just given for
Example 4.1.1.
Since rotations in the derived imbedding Ga ~ sa are the same as those in
the base imbedding G ~ S, up to subscripts in the voltage group d, the faces
of the derived imbedding are determined by the faces of the base imbedding.
In particular, the sequence of edges in a boundary walk of Ga ~ sa corre-
sponds exactly (after erasure of subscripts) to one or more iterations of the
edge sequence in a boundary walk of the base imbedding. The iterations
continue until the first edge would once again have the same voltage subscript
The Derived Imbedding 165

it started with. In terms of the natural projection p: Ga ~ G, the boundary

walks of the derived imbedding are precisely the components of lifts of the
boundary walks of the base imbedding. Thus, the following theorem is
essentially a corollary of Theorem 2.1.3.

Theorem 4.1.1 (Gross and Alpert, 1974; Gross, 1974). Let C be the boundary
walk of a face of size k in the imbedded voltage graph (G ~ S,a). If the net
voltage on the closed walk C has order n in the voltage group .J/1, then there are
#.Jiljn faces of the derived imbedding oa ~ sa corresponding to the region
bounded by C, each with kn sides. 0

In Example 4.1.1 the base region aebc- has length 4 and net boundary
voltage of order 2, so it lifts to one eight-sided face a 0 e 0 b0 c 1 - a 1 e 1 b 1 c0 - ,
whose boundary walk covers the boundary walk of a e b c- twice. The base
region db- e a- has length 4 and net boundary voltage of order 1, so it lifts to
two four-sided faces, d 0 b 1 - e 1 a 0 - and d 1 b0 - e 0 a!, each of whose boundary
walks covers the boundary walk of db- e a- once. The base digon c d- has
net boundary voltage of order 2, so it lifts to one four-sided face c0 d 1 - c1 d 0 - .
We emphasize the principle that faces of the derived imbedding are
computed in practice from the net voltages on boundary walks of the base
imbedding, not by the lengthy task of face tracing in the derived rotation
system. Another example further illustrates this point.

Example 4.1.2. Let (G ~ S 1 , a) be the imbedded voltage graph shown in

Figure 4.2, where voltages are in the cyclic group .2'6 • The four boundary walks
have net voltages (listed clockwise starting from the upper left) 0, 5, 4, and 3.
Therefore the derived imbedding has six faces of size 4, one face of size 24, two
faces of size 12, and three faces of size 8. Since the base graph has four vertices
and eight edges, the derived surface has Euler characteristic
x(sa) = 6 · 4- 6 · 8 + (6 + 1 + 2 + 3) = -12

1 1 1
1 2

1 4

1
~
Figure 4.2. A voltage graph imbedded in the torus, with voltages in the cyclic group .2"6 .
166 Imbedded Voltage Graphs and Current Graphs

,o,
v
Figure 4.3.
1 v
An imbedding 8 2 -+
I 1I r r
v2 v3 v4 v0 v1
i
v2
S1 with voltages in .2'5 and the derived imbedding K 5 -+ S1 .

4.1.3. The Kirchhoff Voltage Law

A face of an imbedded voltage graph is said to satisfy the Kirchhoff voltage
law (abbreviated KVL) if the net voltage on its boundary walk is the identity.
Such faces are especially important in constructing minimal imbeddings, since
the lifts of a face satisfying KVL all have the same number of sides as the face
itself. For example, a triangular base imbedding with all faces satisfying KVL
gives rise to a triangular derived imbedding. Computation of the Euler
characteristic of the derived surface is also simpler when faces satisfy KVL.

Theorem 4.1.2. Let (G ~ S, a) be an imbedded voltage graph such that every

face satisfies KVL. Then x(S"') = nx(S), where n is the order of the voltage
group.

Proof By Theorem 4.1.1, each face of the base imbedding lifts to n faces
of the derived imbedding. The derived graph always has n times as many
vertices and edges as the base graph. Thus

Example 4.1.3. Let G ~ S be the imbedding of a bouquet of two circles in the

torus with voltages in f!E5 as shown on the left in Figure 4.3. Since KVL holds for
the one and only face of the base imbedding, it follows that x(S"') = Sx(S) = 0.
Every edge of the derived imbedding has type 0, because every edge of the base
imbedding does. Therefore, the derived surface S"' is orientable, and from its
Euler characteristic, we infer that it must be the torus. The derived graph G"' is
the complete graph K 5 • Thus, we have found a quadrilateral imbedding of K 5 in
the torus. This imbedding is shown on the right of Figure 4.3.

4.1.4. Imbedded Permutation Voltage Graphs

Let (G ~ S, a)n be an imbedded permutation voltage graph. The "derived
rotation system" and the resulting "derived imbedding" G"' ~ S"' and "de-
rived surface" S"' are defined just as for ordinary voltages: rotations at
corresponding vertices are the same (up to erasure of subscripts), and corre-
sponding edges have the same type. Once again, it is easier to calculate faces
of the derived graph not from the derived rotation system, but by directly
lifting faces of the base graph.
The Derived Imbedding 167

(1 4 6)(2 7 3 5) (1 2 3)(4 56 7)

(1 7 5)(2 6 3 4)
Figure 4.4. A permutation voltage graph imbedded in the sphere.

Theorem 4.1.3. Let C be the boundary walk of a k-sided face in the imbedded
permutation voltage graph (G ~ S, a)n· Suppose that the net permutation
voltage on C has cycle structure ( c 1 , c 2 , • . . , en)· Then there are
cl + c2 + ... +en faces of the derived imbedding oa
~sa corresponding to the
face C, including for j = 1, 2, ... , n exactly cj faces with kj sides.

Proof This is an easy corollary of Theorem 2.4.3. D

Example 4.1.4. Consider the imbedded permutation voltage graph

(B3 ~ S0 , a) 7 shown in Figure 4.4. The derived graph has seven vertices, 21
edges, and no multiple edges or loops, which implies that it is the complete graph
K 7 • The voltage on each of the three monogons of the bouquet B 3 has cycle
structure (0, 0, 1, 1, 0, 0, 0), so it gives rise to one quadrilateral and one triangle in
the derived imbedding. The three-sided outside face of the base imbedding has a
net voltage of (15732)(46); thus it lifts to one 15-sided face and one 6-sided face
in the derived imbedding. The derived surface sa is orientahle of genus 4, since

x(sa) = 7- 21 + 8 = -6

4.1.5. Orientability
Let (G ~ S, a) be an imbedded voltage graph. If the surfaceS is orientable,
then the orientability algorithm of Section 3.2 enables us to construct an
equivalent imbedding in which every edge of G has type 0. Therefore, we may
assume that every edge of the derived imbedding also has type 0, from which it
follows that the derived surface is orientable. If the base surface S is
nonorientable, the situation is not so clear. On the one hand, the derived
surface in Example 4.1.1 is nonorientable, since the closed walk a 0 b0 c0 - is
type 1. On the other hand, the derived surface might be orientable, as in the
following example.

Example 4.1.5. Let (G ~ S, a) be the imbedded voltage graph with fr4-volt-

ages, as shown on the left in Figure 4.5. The imbedding has only one face, so we
know from its Euler characteristic that the base surface S is the projective plane.
168 Imbedded Voltage Graphs and Current Graphs

0
Figure 4.5. Rotation proJections of a nonorientable imbedded .2"4 -voltage graph and its orienta-
hie derived imbedding.

The derived imbedding, shown on the right, has two four-sided faces, and its only
cycle consists of four edges of type 1. Thus the derived surface is the sphere; in
particular, it is orientable. If the voltages shown were in f!E3 instead, then the
derived surface would have had a cycle with an odd number of type-1 edges and
would have been nonorientable.

From this example, one may correctly suspect that the derived surface of an
imbedded voltage graph (G ~ S, a) is nonorientable, if and only if some
type-1 closed walkinG has a net voltage of odd order. In fact, one only need
check for net voltages of order 1.

Theorem 4.1.4. Let (G ~ S, a) be an imbedded voltage graph. Then the

derived surface S is nonorientable if and only if the net voltage on some type-1
closed walk in the base imbedding G ~ S equals the identity.

Proof Let W' be a type-1 closed walk in the derived surface sa, and let
p:Ga ~ G be the natural projection. Then the image W = p(W') of the
sequence of edges in W is a type-1 closed walk, and, by Theorem 2.1.2, the net
voltage on the walk W is the identity. Conversely, let W be a type-1 closed
walk in the base imbedding G ~ S whose net voltage equals the identity.
Then the lift p- 1(W) consists of #..#copies of the closed walk W, each of
which is type 1. D

Clearly one cannot check the net voltage of every one of the infinite number
of type-1 closed walks to see if it equals the identity. Some finite collection of
"basic" cycles is needed, as in the orientability algorithm of Section 3.2, or as
in the method described in Section 2.5 for finding generators of the local group
.Jil(u). The difficulty is that the voltages assigned to basic type-0 cycles have
just as much influence over the orientability of the derived imbedding as the
voltages assigned to basic type-1 cycles. The globalness is illustrated by the
following example.

Example 4.1.6. Let B2 ~ S be any imbedding of the bouquet of two circles such
that one loop d of B 2 is type 1 and the other loop e is type 0. Let a be the
f!E2-voltage assignment given by a( d) = a( e) = 1. Since the only "basic" type-1
The Derived Imbedding 169

cycle d does not have a net voltage of 0, it might appear that the derived surface
sa is orientable. However, the closed walk de is also type I, and its net voltage is
0, since a( e) = 1. Thus the derived surface is, in fact, nonorientable.

Example 4.1.6 indicates that the derived surface is orientable when the net
voltages on type-1 closed walks are in some way "disjoint" from the net
voltages on type-0 closed walks. To make this idea more precise, a base vertex
u is needed so that the local voltage group .Ji.l(u) can be used. Let .9.1°(u)
denote the collection of net voltages on all type-0 closed walks based at vertex
u. Since the set of such walks is closed under the product of walks and the
reversing of walks, the collection .9.1°(u) forms a subgroup of the group .Ji.l(u);
this subgroup is called the "type-0 local (voltage) group".

Theorem 4.1.5. Let (G ~ S, a) be an imbedded voltage graph with voltage

group .9.1, and let u be any vertex of G. Then the type-0 local group .9.1°(u) has
index 1 or 2 in the local group .Ji.l(u). Moreover, the following are equivalent:

i. .9.1°(u) has index 2 in .Ji.l(u).

ii. No type-1, u-based closed walk has its net voltage in .9.1°(u).
iii. The derived surface sa is orientable.

Proof We will first prove that i and ii are equivalent. To this end, let e be
any type-1, u-based closed walk, and suppose its net voltage is b. We claim
that .Ji.l(u) = .9.1°(u) u b.9.1°(u). In support of this claim, suppose that a E
.9.1 ( u ), so that a is the net voltage on some u-based closed walk D. If D is type
0, then a E .9.1°( u) by definition. If D is type 1, then e- D is type 0, so its net
voltage b -la is in .9.1°( u ). Therefore, a E b.9.1°( u ), and the claim is proved. It
follows that .9.1°(u) has index 1 or 2 in .Ji.l(u). Since the type-1 closed walk e
is arbitrary, this also shows that .9.1°(u) has index 2 if and only if no type-1,
u-based closed walk has its net voltage in .9.1°(u). Thus, i and ii are equivalent.
Now we will show that ii and iii are equivalent. Suppose that the derived
surface sa is nonorientable. Then, by Theorem 4.1.4, there is a type-1 closed
walk e in the base graph G with net voltage equal to the identity. If e is not
already based at the vertex u, then the closed walk Den-, where D is a path
from u to the initial vertex of e, is based at u and still is type-1 with net
voltage equal to the identity. Of course, the identity is in .9.1°(u). Thus, we
have proved that ii implies iii. Conversely, suppose that some type-1, u-based
closed walk e has its net voltage in .9.1°(u). By the definition of the subgroup
.9.1°(u), some type-0, u-based closed walk D has the same net voltage as C.
Thus, en- is a type-I closed walk in the base graph, whose net voltage equals
the identity. It follows from Theorem 4.1.4 that the derived surface sa is
nonorientable. Thus iii implies ii. D

Example 4.1.7. Let B 3 ~ S be any imbedding of a bouquet of three loops c, d,

and e, such that each loop is type 1. Let a be the unique voltage assignment in
the group ~2 X ~2 such that a( c)= (1,0), a(d) = (0, 1), and a(e) = (1, 1).
170 Imbedded Voltage Graphs and Current Graphs

Since each of the three possible index-2 subgroups of ~2 X ~2 contains the net
voltage of a type-1loop, it follows from Theorem 4.1.5 that the derived surface is
nonorientable.

4.1.6. An Orientability Test for Derived Surfaces

We wish to develop a finite test for the orientability of a derived surface based
on Theorem 4.1.5, in which the only closed walks to be checked are in the
"basic" set of cycles given by a spanning tree. Unfortunately, the type-0 local
group might not be generated by the net voltages on the basic type-0 cycles.
For instance, in returning to Example 4.1.7, we see that there are no type-0
cycles at all, yet the type-0 local group is not trivial. (In fact, it is the whole
voltage group.) Thus, the test is not quite so direct as one might hope. Before
giving the test, we shall prove a useful algebraic lemma about index-2
subgroups.

Theorem 4.1.6 (An Algebraic Lemma). Let !II be a subgroup of index 2 in a

group .91. Then !II is the kernel of a homeomorphism <P: .91 ~ ~2 • In particular,
a product a 1 • • • an is in the subgroup !II if and only if the number of elements
a; (/; !II is even.

Proof It suffices to show that the subgroup !II is normal in the group .91.
This follows immediately from the observation that, if a rt; !II, then a!JB = ?A a
= .91- !II. D

Orientability Test for the Derived Surface. Let (G ~ S, a) be a connected,

nonorientable imbedded voltage graph with voltage group .91. First, select a
spanning tree T and a base vertex u for the graph G, and change the local
orientation where necessary (as in the orientability algorithm of Section 3.2) so
that each edge of the tree Tis type 0. Next, compute the T-voltage assignment aT
and the local group s#(u), as in Section 2.5. Then the derived surface sa is
orientable if and only if this condition holds:

There is a subgroup !II of index 2 in the local group

s#(u) such that aT(e) E !II when
e is a type-0 edge and aT( e) (/; !II when
e is a type-1 edge.

This algorithmizable test was devised by Gross and Tucker (1979a). It is

related to criteria given by Stahl and White (1976).

To verify that this orientability test works, we first recall that the derived
surfaces for voltage assignments a and aT are the same. If their common
derived surface is orientable, then by Theorem 4.1.5, the type-0 local group
s# 0 (u) has index 2 in the local group s#(u), and no type-1 closed walk has a
The Derived Imbedding 171

u 3 u u 3 u

6 6 6 6

3
Figure 4.6. A voltage graph imbedded in the Klein bottle (the original voltages are given on the
left, the T-voltages on the right).

net voltage ind' 0 (u). Because every edge of the tree T is type 0 and has
T-voltage equal to the identity, it follows that every edge e not in the tree T
corresponds to a unique u-based closed walk having the same type as e and
net voltage equal to aT(e). Therefore, d' 0 (u) is the desired subgroup !!J.
Conversely, suppose that the subgroup !!J exists. Then aT assigns every
type-1 edge a voltage not in !!J. By the algebraic lemma (Theorem 4.1.6), it
follows that every type-1 closed walk has a net voltage not in !!J. In particular,
no type-1 closed walk has net voltage equal to the identity. Therefore, by
Theorem 4.1.4, the derived surface is orientable.

Example 4.1.8. Let (G ~ S, a) be the imbedded voltage graph given on the left
in Figure 4.6 where voltages are in the cyclic group ~12 • The base imbedding
surface S is a Klein bottle. The graph G is the same as that of Example 2.5.3,
and the tree T and root vertex u selected are again the same as in Example 2.5.3.
Local orientations are indicated by arrows around vertices. On the right of Figure
4.6 is the same imbedding with the orientation at vertex v reversed so that each
edge in the tree T is type 0. In addition the T-voltages have been given as
computed in Example 2.5.3. The local group d'(u) is {0, 3, 6, 9}. The only
possible 2-index subgroup is {0,6}. It is easily verified that aT(e) E {0,6} if
and only if the edge e is type 0. We conclude that the derived surface sa is
orientable. We also observe that the derived surface sa has three components.
Finally, since two of the boundary walks of the imbedding G ~ S have net
voltage 0 and the other two have net voltage 6, it follows that the derived
imbedding has 36 faces. Therefore, x(Sa) = 36- 84 + 36 = -12, from which
it follows that the derived surface sa consists of three orientable components of
genus 3.

It should be noted that the existence of the index-2 subgroup !!J in the test
described above is equivalent, by Theorem 4.1.6, to the existence of a homeo-
morphism cp: d'(u) ~ ~2 such that cf>(aT(e)) = 0 when e is a type-0 edge
and cf>( aT( e)) = 1 when e is a type-1 edge. This viewpoint is especially
convenient if the local group d'(u) is given by a presentation with generating
set { aT( e) 1 e E EG - T}. There really is no choice for the desired homeomor-
172 Imbedded Voltage Graphs and Current Graphs

z yz
Figure 4.7. A rotation projection for an imbedded voltage graph and a corresponding set of
T-voltages.

phism cf>, since its values on the generating set are determined. It is necessary
to check only that this assignment of Os and 1s to the generators of d(u) does
define a homeomorphism; that is, each relator for the given presentation must
have the value 0.

Example 4.1.9. Let (G ~ S, a) be the imbedded voltage graph given on the left
in Figure 4.7. The voltage group d is a finite group with a presentation of the
form

(x, y, z: x 2 = y 2 = z 2 = 1,(xyt = (yz) 4 = (xz) 3 = 1, ... )

On the right in Figure 4.7 are the T-voltages for the spanning tree shown in
boldface. If we let s = xy and t = yz, then the local group d(u) has a
presentation of the form

( s, t: s 4 = t 4 = ( st) 3 = 1, ... )

Then the desired homeomorphism cf>: Jli'(u) ~ fl'2 must have cf>(s) = 0 and
cf>(t) = 1. Butthen cf>((st) 3 ) = 3cf>(s) + 3cf>(t) -:F 0. Therefore the derived surface
is nonorientable.

4.1.7. Exercises
1. Let (G ~ S, a) be the imbedded voltage graph whose rotation projection
appears in Figure 4.8a. Find the number of faces of each size in the
derived imbedding and identify the derived surface.
2. Verify that the derived imbedding of the imbedded permutation voltage
graph given in Figure 4.8b is an imbedding of the complete bipartite
graph K 4• 4 in which every face is a quadrilateral.
3. Construct an imbedding of the complete graph K 6 that has two triangles,
two 6-gons, and three quadrilaterals. (Hint: Begin with a bouquet of
three circles imbedded in the sphere. Multiple edges in the derived graph
must be identified.)
The Derived Imbedding 173

(1 2)(3 4)

(1)(2)(3)(4)

(a) (b)
Figure 4.8. Two rotation proJections of imbedded voltage graphs. (a) Ordinary voltages in .2"6 ,
(b) permutation voltages in 94.

4. Derive a one-face imbedding for K 5 , starting with a bouquet of two

circles imbedded in the Klein bottle.
5. Find an orientable imbedding of the octahedron graph 0 3 in which all
the faces are quadrilaterals.
6. Suppose that the abelian group .91 has a generating set X that contains
no elements of order 2. If #X= n is even, show that the Cayley graph
C( .91, X) has an orientable imbedding in which every face is a 2n-gon.
7. Find an imbedding in the torus of the 4-regular graph (C3 U C4 )c, which
was previously considered in Example 2.3.7.
8. Use Example 2.1.6 to construct an orientable imbedding of the 3-cube
graph Q3 in which all faces are 6-gons.
9. Let (G --+ S, a) be an imbedded voltage graph where the voltage group
.91 is nontrivial. Prove that if x(S) < 0, then x(S"') < x(S), and that if
x(S) = 0, then x(S."') :;;; o.
10. Let .91 be a nontrivial finite group with presentation (x, y: x 3 = y 5 =
(xy) 2 = 1, ... ). Prove that the Cayley graph C(.91,{x, y}) has an im-
bedding in a connected, orientable surface of Euler characteristic #.91/30,
and that therefore #.91 = 60. Conclude that the group .91 is simple.
(Hint: Show that any nontrivial quotient of .91 has a presentation of the
same form as that of .91 and hence has order 60.)
11. Let .91 be a finite group with presentation (x, y: x 3 = y 6 = (xy) 2 =
1, ... ). Prove that the genus of the Cayley graph C( .91, { x, y}) is at
most 1.
12. Use the orientability test for derived surfaces to show that the derived
surface in Example 4.1.1 is nonorientable.
13. Let T be a spanning tree and u a root vertex for a connected, imbedded
voltage graph (G--+ S, a). Let !11' be the subgroup of the local group
generated by {aT( e) 1 e is a type-0 edge}. Show by an example that the
derived surface S may be nonorientable even if aT( e) E !11' when the
edge e is type 0 and aT( e) ft !11' when the edge is type 1.
174 Imbedded Voltage Graphs and Current Graphs

(1, 2)

(0, 5) (0, 1)

(0, 4)
(1, 0)
Figure 4.9. A rotation projection of an imbedded voltage graph, with voltages in .2"2 X .2"8 .

14. Let (G -+ S, a) be the imbedded voltage graph given by the rotation

projection in Figure 4.9 where the voltage group is ~2 X ~8 • Use the
methods of Example 4.1.8 to determine the topological type of the
derived surface sa.
15. Let d be a finite group with a presentation of the form (x, y: xyx- 1y =
1, ... ). Prove that the subgroup !!J generated by x 2 and y is normal in d
and has index at most 2. Show that if !!J has index 2 in d, then the
Cayley graph C( d, { x, y}) has an imbedding in the torus. Conclude that
in any case, the genus of the group d is at most 1.

4.2. BRANCHED COVERINGS OF SURFACES

For any voltage graph (G, a), the natural graph projection p: Ga-+ G is a
covering projection in the topological sense; that is, every point y E G has a
neighborhood NY such that each component of the lift p- 1(Ny) is mapped
homeomorphically by p onto NY, as described in Section 2.3. For any
imbedded voltage graph (G-+ S, a) one might expect that the natural projec-
tion p: Ga -+ G extends to a map p: sa -+ S with properties similar to those
of p. This is indeed the case, for there is an extension of p that is "almost" a
covering map.

4.2.1. Riemann Surfaces

Let p: S-+ S be a continuous function between closed surfaces. Then
p: S-+ S is called a "branched covering" if there is a finite set of points Y in
S such that p restricted to the punctured surface S - p- 1( Y) is a covering
onto the punctured surface S - Y. Since Sand S are compact, it is equivalent
to specify that the function p is a local homeomorphism except possibly at the
points of p- 1(Y). Points in the exceptional set Yare called "branch points",
and points in its lift p- 1(Y) are called "prebranch points".
The "order" of a prebranch point x is r if the function p is r-to-1 in a
small neighborhood of x. If the surface S is connected, an elementary
Branched Coverings of Surfaces 175

topological argument shows that the fiber size # p - 1( z) is the same for all
nonbranch points z E S. This common value is called the "number of sheets"
of the branched covering. The "deficiency" of the branch point y, denoted
def( y ), is n - # p- 1( y ), where n is the number of sheets of the branched
covering.

Example 4.2.1. Let S and S both be the extended complex plane, that is, the
complex plane with a point at oo adjoined (the resulting surface is homeomorphic,
by stereographic projection, to the sphere). Let p: S ~ S be the function
p(z) = zn, where n ~ 1 is an integer. Since p simply multiplies the argument of
each complex number by n, the effect of pis to wrap the complex plane n times
around the origin. The restricted map p: S- {0, oo} ~ S- {0, oo} is clearly a
covering, sop: S ~ Sis a branched covering. The number of sheets is n, since
any complex number other than 0 and oo has n different nth roots. Both 0 and
oo are branch points of deficiency n - 1, and the corresponding prebranch points,
again 0 and oo, have order n.

Example 4.2.2. Let S be the extended complex plane, and let S be the set of
points {(z, w) E C X qw 2 = z(z- l)(z- 2)} with a point at oo adjoined. It
can be shown that every point of S has a neighborhood homeomorphic to the unit
disk, so that Sis a surface. Let p: S ~ S be the projection p(z, w) = z. Then p
fails to be a local homeomorphism only at the points (0, 0), {1, 0), {2, 0) and oo.
Therefore p is a branched covering with branch points at 0, 1, 2, and oo. We
observe that p is two-sheeted, since the equation w 2 = z(z- l)(z- 2) has two
solutions for w for each value of z other than 0, 1, 2, or oo.

Example 4.2.1 gives a typical view of a branched covering near a prebranch

point, and the surfaceS of Example 4.2.2 together with the map p: S ~ S is
the Riemann surface associated with the equation w 2 = z(z - l)(z - 2). The
theory of Riemann surfaces has had a long and powerful influence in low-
dimensional topology. A small amount of this theory, including some classical
results of the 19th and early 20th centuries, is developed here in graph-theoret-
ical terms. For more about Riemann surfaces from an elementary viewpoint,
see Springer (1957).
The idea of a branched covering generalizes to higher dimensions. A
simplicial map p: M ~ N, where M and N are n-dimensional complexes is a
branched covering if there is a subcomplex Y in N of dimension a~ most n - 2
such that p restricted toM- p- 1(Y) is a covering onto N- Y, that is, each
point in M- p- 1(Y) has an open neighborhood mapped homeomorphically
by p onto an open set of N- Y (see Fox, 1957). If instead the subcomplex Y
has dimension n - 1, then the map p is said to be a "folded covering". Such
maps were first studied by Tucker (1936), but they have received little
attention. Bouchet's (1982a) work on lifting triangulations is one of the few
uses of folded surface coverings in topological graph theory.
176 Imbedded Voltage Graphs and Current Graphs

4.2.2. Extension of the Natural Covering Projection

The source of the branched coverings that we study is imbedded voltage
graphs. The following is the fundamental theorem relating branched coverings
to voltage graphs.

Theorem 4.2.1 (Gross, 1974; Gross and Tucker, 1977). Let (G ~ S, a) be an

imbedded voltage graph, with either ordinary or permutation voltages. Then the
natural projection p: Ga ~ G on the graphs extends to a branched covering
p: sa ~ Son their imbedding surfaces. If a is a permutation assignment in Y,,
and if the net voltage on face f has cycle structure ( c1, ••• , en), then the projection
p has a branch point inside face f with exactly c1 prebranch points of order j, for
j = 1, ... , n. If a is an ordinary assignment in the group .91 and the net voltage
on face f has order r, then p has a branch point inside face f with #.9/jr
prebranch points, each of order r.

Proof The natural projection p: Ga ~ G is already known to be a local

homeomorphism, because it is a covering projection. We first extend p to the
0-bands of the imbedding Ga ~ sa so that p remains a local homeomor-
phism. This is possible because the rotation at any vertex V; of G is lifted from
the rotation at v = p( v;), that is, the rotation at the vertex v; is the same as
that at the vertex v up to erasure of subscripts. Next we extend p to the
1-bands, still keeping p a local homeomorphism. Since the type of any edge e;
in the edge-fiber p- 1(e) is defined to be the same as the base edge e, and since
p maps the interior of the edge e; homeomorphically onto the interior of the
edge e, it follows that there is a homeomorphism of the 1-band fore; onto the
1-band fore, which can be used to extend p as desired.
Finally, in order to extend the map p to the 2-bands, we suppose that f is a
face of the imbedding G ~ S, and that the net permutation voltage on the
boundary walk C has cycle structure ( c 1, ••• , cn ). Then, by Theorem 4.1.3,
there are c1 faces of the derived imbedding Ga ~ sa, each of whose boundary
walks is wrapped by p exactly j times about the boundary walk C. The map p
can then be extended inside the 2-band for each of these faces so that its
restriction to any one of these 2-bands looks like the map z ~ z 1 on the unit
disk in the complex plane. In particular, the branch point inside the face f will
have exactly c1 prebranch points of order j, for j = 1, ... , n. If a is an
ordinary voltage assignment, then essentially the same proof holds, only using
Theorem 4.1.1 instead of Theorem 4.1.3. D

Corollary. Let (G ~ S, a)n be an imbedded voltage graph satisfying KVL on

every face. Then a natural projection p: sa~ Sis an unbranched covering. D

We should like to call the branched covering constructed in Theorem 4.2.1

"the" natural projection of the derived surface sa onto the base surface s.
Certainly in terms of the combinatorial structure of vertices, edges, and faces
Branched Coverings of Surfaces 177

(1)(2 3)(4 5) ~ (12)(3 4 5)

d e
Figure 4.10. A rotation proJection for an imbedded permutation voltage graph.

for these surfaces, the extension of the natural graph projection p: Ga ~ G is

unique. However, it is not at all unique topologically. For example, let us try
to be more precise about how p extends to the 2-bands so that it "looks like"
the map z ~ zl. If B is a 2-band in the derived surface and B is the
corresponding 2-band in the base surface, we first need homeomorphisms
iz: iJ ~ D and h: B ~ D, where Dis the unit disk in the complex plane. Then
p is defined to be h- 1gh on iJ, where g(z) = zl. Thus the extension of p
depends upon the choice of homeomorphisms h and h. In fact, the extension
of p to the 0-bands and 1-bands also depends on choices of homeomorphisms.
We could try to get around this difficulty by defining "equivalent" branched
coverings just as we defined equivalent imbeddings in Section 3.1. Rather than
become any more deeply involved with topological considerations, however,
we simply define a "natural (surface) projection" for an imbedded voltage
graph (G ~ S, a) to be any branched covering q: sa~ S that extends the
natural graph projection p: Ga ~ G and has at most one branch point in each
face of the base imbedding.

Example 4.2.3. Let (G ~ S, a) 5 be the imbedded permutation voltage graph

given by the rotation projection in Figure 4.10. By Theorem 4.2.1, there is a
natural surface projection p: sa
~ s that is an extension of the natural graph
projection p: Ga ~ G. The face bounded by the loop d has a branch point of
deficiency 5 - 3 = 2 and prebranch points of orders 1, 2, and 2. The face
bounded by e has a branch point of deficiency 5 - 2 = 3 and prebranch points of
orders 2 and 3. By computing its net boundary voltage, we determine that the
outside face has a branch point of deficiency 3 and prebranch points of order 1
and 4. Note that some prebranch points might have order 1, in which case the
map pis still a local homeomorphism there.

4.2.3. Which Branched Coverings Come from Voltage Graphs?

It might appear that branched coverings arising from imbedded permutation
voltage graphs are special in some way, but the following theorem shows this
not to be the case. Just as every graph covering can be obtained by the
permutation voltage graph construction, every branched covering of surfaces
can be obtained by extending the natural projection of a permutation voltage
graph.

Theorem 4.2.2 (Gross and Tucker, 1977). Let q: S ~ S be an n-sheeted

branched covering. Let G ~ S be any cellular graph imbedding such that no
178 Imbedded Voltage Graphs and Current Graphs

branch point of q lies on G and at most one branch point lies in any face. Then
there is a permutation voltage assignment a in .9;, such that q is a natural
projection S" --+ S.

Proof By hypothesis, the restnctwn of the covering q to the graph

preimage q- 1(G) is a local homeomorphism, and hence a covering. By
Theorem 2.4.4, this restriction of q must be the natural projection of a
permutation voltage graph (G, a)n· In order to show that the rotation system
for the imbedding q- 1(G)--+ Sis the same as the derived rotation system for
the derived imbedding G" --+ S", let v be a vertex of the imbedding
q- 1(G)-+ S. Since the covering q is a local homeomorphism at vertex v, it
takes a small disk D around vertex v homeomorphically onto a small disk
q( D) around the vertex q( v) in the imbedding G -+ S. Thus, if the rotation at
vertex v-which is determined by the disk D-is e 1 • • • em, then the rotation
at vertex q(v)-which is determined by the disk q(D)-must be q(e 1 ) • • •
q( en). That is, the rotation at the vertex v in the covering space S is the
derived rotation. A similar argument shows that each edge e of the imbedding
q- 1( G) -+ S has the same type as its image edge q( e). D

Example 4.2.1 Revisited. The branched covering of the extended complex plane
given by p(z) = zn could be obtained by extending the natural projection of the
permutation voltage graph ( G -+ S, a )n where G consists of a single loop d
imbedded in the sphere and a( d) = (1 2 3 · · · n ).

Example 4.2.2 Revisited. The branched covering of the Riemann surface for
w 2 = z(z - l)(z - 2) could be obtained from the imbedded voltage graph
(B 3 -+ S0 , a) given in Figure 4.11. The voltage group is f!E2 •

Theorem 4.2.2 can be used to obtain nontrivial information about branched

coverings. As one example, we show that any connected covering of the sphere
p: S -+ S0 must have at least two branch points if it is nontrivial (more than
one sheet). Indeed, if p has one branch point or none, the graph G of Theorem
4.2.2 can be chosen to consist of a single loop in the sphere enclosing the
branch point of p if it has one. The voltage assigned to the loop cannot be the
identity since the covering space is connected and has more than one sheet.
However, a nonidentity voltage on the loop yields two branch points, one

Figure 4.11. An imbedded voltage graph for the Riemann surface of

1 1 Example 4.2.2.
Branched Coverings of Surfaces 179

inside the loop and another outside the loop. The case of two branch points in
the sphere is considered in Exercise 13.

4.2.4. The Riemann-Hurwitz Equation

Given a branched covering p: S ~ S, it is useful to express the Euler
characteristic x(S) in terms of x(S) and the deficiencies of the branch points
of p. The following theorem establishes the Riemann-Hurwitz equation,
which does exactly that.

Theorem 4.2.3. Let p: S ~ S be an n-sheeted covering and let Y be the set of

branch points of p. Then

x(S) = nx(S) - L def(y)

yeY

Proof By Theorem 4.2.2, the map p is a natural surface projection for an

imbedded permutation voltage graph (G ~ S, a)n· Since #V(Ga) = n#V(G)
and #E(Ga) = n#E(G), we see that all that is needed to compute the Euler
characteristic x(S) = x(Sa) of the derived surface is the number of faces of
the derived imbedding Ga ~ sa. Each face of the imbedding G ~ S lifts to n
faces if the face contains no branch point and ton- def(y) faces if the face
contains the branch point y. Therefore, #F( Ga) = n#F( G) - Lye y def(y ),
from which it follows that

x(sa) = n#V(G)- n#E(G) + n#F(G)- L def(y)

yeY

= nx(S) - L def(y) o
yeY

Example 4.2.4. The Riemann surface for the equation w 3 = z(z - 1)(z- 2)
has branch points at z = 0, 1, 2, oo, each of deficiency 2. By the Riemann-Hurwitz
equation, the Euler characteristic of the Riemann surface is 3x(S0 ) - 4 · 2 =
- 2, and we infer that the surface has genus 2.

4.2.5. Alexander's Theorem

One of the more startling results in the theory of manifolds is Alexander's
theorem: any closed, orientable, triangulated manifold of dimension n is a
branched covering of the sphere of dimension n. Although the surface version
of this theory was known before 1920, the importance of Alexander's general-
ization to higher dimensions, especially dimensions 3 and 4, was not fully
realized for another half-century. The proof given here for surfaces is an easy
application of voltage graphs.
180 Imbedded Voltage Graphs and Current Graphs

2 + 2g loops in all

•• ••

~ Figure 4.12. An imbedded voltage graph.

Theorem 4.2.4 (Alexander, 1920). Every closed orientable surface S is a

branched covering space over the sphere S0 • In fact, Sg is a two-sheeted covering
of S0 with 2 + 2g branch points.

Proof Imbed the bouquet B2 + 2 g in the sphere S0 as shown in Figure 4.12.

Let a assign to each loop the voltage 1 in the group fl'2 • Then the net voltage
on the exterior face is 0, and on each of the 2 + 2g monogons it is 1.
Therefore, a natural surface projection p: sa ~ S0 has 2 + 2g branch points,
each of deficiency 1. From the Riemann-Hurwitz equation,

x(sa) = 2x(S0 ) - (2 + 2g) = 2- 2g

it follows that the derived surface has genus g. D

The sharpened conclusion given here, that the branched covering can be
chosen to be two-sheeted, does not hold in higher dimensions. It is known,
however, that every closed, orientable 3-manifold is a three-sheeted branched
covering of the 3-sphere (Hilden, 1976; Montesinos, 1976). The following
nonorientable version of Alexander's theorem is also not known in higher
dimensions.

Theorem 4.2.5. Every closed nonorientab/e surface Nk is a branched covering

space over the projective plane N 1 • In fact, there is a branched covering
p: Nk ~ N 1 having two sheets and k branch points if k is even, or three sheets
and (k + 1)/2 branch points if k is odd.

Proof For an even crosscap number k, let Bk ~ N1 be the imbedding

obtained by giving a twist to one loop of the imbedding in Figure 4.12. Again
let a assign to each loop the voltage 1 in the group fl'2 • Then the base
imbedding has k faces, and each has net voltage 1, including the "exterior"
face. Thus, a natural projection from the derived surface has k branch points,
each of deficiency 1. By the Riemann-Hurwitz equation, the derived surface
has Euler characteristic 2x(N1 ) - k = 2 - k. Since an orientability test for
the derived surface shows it is nonorientable, the derived surface must be Nk.
For an odd crosscap number k, we imbed the bouquet B(k+l)/2 in the
projective plane in the same way as in the even case. This time, however, the
Branched Coverings of Surfaces 181

(12 4)

(1 3)(5 6) (1 3)(5 6)
(1 5)

(2 3 4 5) (2 3 4 5)

(12 4)
Figure 4.13. A permutation voltage graph imbedded in the torus.

type-0 loops are assigned the voltage 1 in fr3 and the single type-1 loop is
assigned a voltage that gives the exterior face a nonzero net voltage. Then
there are (k + 1)/2 branch points, each of deficiency 2. By the Riemann-
Hurwitz equation, the derived surface has Euler characteristic 3x(N1) -
2(k + 1)/2 = 2 - k. Since the number 2 - k is odd, the derived surface must
be nonorientable, from which we conclude that it is homeomorphic to Nk. D

It is natural to ask whether the anomalous use of a three-sheeted branched

covering for odd k is really necessary. In fact, no two-sheeted covering space,
branched or unbranched, can have odd Euler characteristic (see Exercises 11
and 12).

4.2.6. Exercises
1. Let (G ~ S, a) 6 be the permutation voltage graph imbedding in the
torus given by Figure 4.13. Describe the branch and prebranch points of
a natural projection p: sa ~ S. Compute x(Sa) first by the Riemann-
Hurwitz equation, then by simply counting the number of vertices, edges,
and faces of the derived imbedding.
2. Find the branch points and prebranch points of the Riemann surface for
the equation w 4 = z 3 - z. What is its genus?
3. Show that the torus is a two-sheeted unbranched covering space over the
Klein bottle.
4. Generalize Exercise 3 by proving that the orientable surface S8 is a
two-sheeted unbranched covering of the nonorientable surface N8 + 1 for
all g ~ 0. (Hint: Represent the surface N8 + 1 by a bouquet of g + 1
type-1 loops.)
5. Use Theorem 3.4.5 and Exercise 4 for g = 0 to show that there are
graphs of arbitrarily large genus having two-sheeted planar coverings.
6. Give an imbedding of the Petersen graph in the projective plane such
that every face has size 5. Conclude that the !-skeleton of the dodeca-
hedron is a two-sheeted covering of the Petersen graph.
182 Imbedded Voltage Graphs and Current Graphs

7. Prove that the only surfaces that are unbranched coverings of themselves
(more than one-sheeted) are the torus and Klein bottle. Prove that the
only surfaces that are coverings of themselves with branch points are the
sphere and projective plane.
8. Prove that if S is an orientable surface, then S is a branched covering
over the sphere with exactly three branch points.
9. Prove that the nonorientable surface Nk, k > 3, is an unbranched
covering over the nonorientable surface N3 •
10. Prove that the orientable surface Sg, g > 2, is an unbranched covering
over the orientable surface S2 •
11. Show that for any imbedded voltage graph with voltage group fr2 the
sum of the net boundary voltages over all faces is 0. Conclude that any
two-sheeted branched covering must have an even number of branch
points.
12. Use Exercise 11 to prove that no surface of odd Euler characteristic is a
two-sheeted branched covering space over any surface.
13. Let p: S ~ S0 be a branched covering with two branch points in the
sphere S0 . Prove that if S is connected, it must be the sphere as well.

4.3. REGULAR BRANCHED COVERINGS AND GROUP ACfiONS

The natural projection of an ordinary voltage graph is more than a covering; it

is also the quotient map of the free action of the voltage group on the derived
graph. In the same way, a natural surface projection of an imbedded ordinary
voltage graph is not just a branched covering, but also the quotient map of a
group action on a surface.

4.3.1. Groups Acting on Surfaces

An "action of a group .J/1 on a surface S" is given by associating to each group
element a E .J/1 a homeomorphism cpa on the surface S such that

1. cpab = cpa o cpb for all a, b E .J/1.

2. cpa is the identity homeomorphism if and only if a is the identity
element of the group .J/1.

A "fixed point" of the action is a point x E S such that cpa(x) = x for

some group element a =F 1. An action with no fixed points is called "free" (or
"fixed-point free") and an action with a finite number of fixed points is called
"pseudofree".

Example 4.3.1. When the surface S2 is imbedded in 3-space as shown in Figure

4.14a, there is an action of the Klein group fr2 X fr2 on S2 whose nonidentity
elements are 180° rotations around the x, y, or z axis (the composition of any
Regular Branched Coverings and Group Actions 183
z

(a) (b)
Figure 4.14. Two placements of a surface of genus 2 in 3-space.

two of these is the third). This action on S2 clearly is pseudofree, but not free; in
particular, a 180° rotation about they axis has six fixed points. A free action of
~2 on the surface s2 is generated by the antipodal homeomorphism h(x, y, z) =
(- x, - y, - z ). The action of the group ~2 X ~2 X ~2 generated by the three
reflections through the coordinate planes is neither free nor pseudofree, since it
has entire circles of fixed points; however, this action contains both the two
previously described actions. Finally, the 3-space imbedding of the surface S2
depicted in Figure 4.14b reveals an action of the cyclic group ~3 generated by a
120° rotation about the z axis.

Example 4.3.2. In the graph imbedding whose band decomposition is illustrated

in Figure 4.15, there are three faces: an inner 5-gon, an outer 5-gon, and a
20-gon in between. There is an obvious action of the cyclic group ~5 on that band
decomposition. Since this action takes each boundary walk to itself, the action
can be extended radially inside each 2-band to obtain an action of ~5 on the
closed surface. Each non identity element of the ~5 action has three fixed points
-one at the center of each face. Note that the closed surface is orientable (just
reverse the rotation at each vertex on one of the pentagons) and that its Euler

Figure 4.15. A band decomposition of a surface of genus 2.

184 Imbedded Voltage Graphs and Current Graphs

characteristic equals 10 - 15 + 3 = - 2. Thus, we have an action of fr5 on a

closed surface of genus 2.

We observe that the various group actions in Example 4.3.1 could each be
obtained from an action on a graph at the "core" of the solid in 3-space
bounded by the surface. For instance, the core of the double torus in Figure
4.14a is a bouquet of two circles that run inside the handles. This observation
can be generalized.

Theorem 4.3.1. Let .91 be a group that acts on the graph G. Then .91 also acts
on a surface S whose genus is the cycle rank {J(G) = #E- #V + 1. If the
action of .91 on the graph G is free, then so is the action of .91 on the surfaceS.

Proof Imbed the graph G in 3-space and thicken it to form a solid, and let
S be the boundary of that solid. Obviously, the action of the group .91 on the
graph G extends to an action on the solid, from which we can extract an action
on the surface S. The action on the solid and on its boundary S is free if the
action on the graph G is free. If the graph G is a tree, then the surface S is a
sphere bounding a misshapen 3-cell. If T is a spanning tree from G, then each
edge not in T adds a handle to the surface S. Therefore, the genus of S is
{J(G), the number of edges not in the tree T. D

4.3.2. Graph Automorphisms and Rotation Systems

In Example 4.3.2, an action of a group on a graph is extended to an imbedding
surface for the graph. In general, to extend an automorphism of a graph to a
homeomorphism of an imbedding surface, there must be a vertex-orbit con-
sistency within the rotation system corresponding to the imbedding, and an
edge-type consistency.
Suppose that the graph imbedding G ~ S has a rotation

at the vertex v, and that his an automorphism of the graph G. We say that the
automorphism h "respects the rotation at the vertex v" if the rotation at the
image vertex h ( v) is

We say that h "reverses the rotation at v" if the rotation at h ( v) is

We say that the graph automorphism h "preserves the type of edge e" if
the image edge h (e) has the same type as the edge e.
Regular Branched Coverings and Group Actions 185

Finally, we say that the graph automorphism h "respects the rotation

system" if h preserves the type of each edge and respects the rotation at each
vertex. Similarly, we say that h "reverses the rotation system" if h preserves
the type of each edge and reverses the rotation at each vertex.

Theorem 4.3.2. Let G ~ S be an imbedding, and let h: G ~ G be a graph

automorphism such that h either respects or reverses the rotation system. Then h
extends to a surface homeomorphism h: S ~ S. Any face of the imbedding
G ~ S whose boundary walk is invariant under the graph automorphism h
contains a fixed point of the extension h. Moreover, if the graph automorphism h
has no fixed points, then its extension h can be chosen to have at most one fixed
point in any face.

Proof Topologically, the condition that the graph automorphism h re-

spects or reverses the rotation system simply ensures that h can be extended to
the 0-bands and 1-bands of the imbedding G ~ S; the extension to the whole
surface S can then be easily accomplished, say radially, inside each 0-band.
Combinatorially, what that condition ensures is that each boundary walk of a
face is mapped by h to a boundary walk of a face, since boundary walks are
determined completely by the rotation system. Any boundary walk mapped to
itself by h must bound a face mapped to itself by the extension h. By the
Brouwer fixed-point theorem [e.g., see Milnor (1965b) for a concise, self-con-
tained proof), the continuous function h must have at least one fixed point in
that face. If the graph automorphism h has no fixed points, then no boundary
walk will have any fixed points, in which case a radial extension of the graph
automorphism h inside each face will have the property that there is at most
one fixed point for the extension h inside any face. 0

Corollary. Let G ~ S be a graph imbedding, and let .91 be a group of

automorphisms acting on the graph G. If every element of .91 respects or reverses
the rotation system, then the action of .91 can be extended to an action on the
imbedding surface S. Moreover, if .91 acts freely on the graph G, then the
extended action can be chosen to have at most one fixed point inside any face. 0

It should be noted that the extended homeomorphism of Theorem 4.3.2

might have fixed circles, as in the reflections in Example 4.3.1; this observation
will prove useful in Chapter 6. The currently important application of Theo-
rem 4.3.2 is for imbedded voltage graphs.

Theorem 4.3.3. Let (G ~ S, a) be an imbedded voltage graph with voltage

group .91. Then the natural action of .91 on the derived graph Go. extends to an
action of .91 on the derived surface sa, having at most one fixed point inside any
face.

Proof It follows immediately from the definitions of the derived rotation

system and the natural action of the voltage group that every element of the
186 Imbedded Voltage Graphs and Current Graphs

0
3 1 Figure 4.16. An imbedded voltage graph with volt-
ages in .2'8 .

natural action respects the derived rotation system and that the natural action
is free. By the corollary to Theorem 4.3.2, the conclusion of Theorem 4.3.3
holds. D

As with the natural projection, it would be convenient to be able to call the

extended action of the voltage group d on the derived surface sa "the"
natural (surface) action. However, the extension is not unique topologically,
even if it is unique combinatorially. Therefore we define a "natural (surface)
action" of the voltage group d on the derived surface sa to be any action of
d on sa that extends the natural action of d on the derived graph Ga and
that has at most one fixed point inside any face.

Example 4.3.3. Let ( G -+ S0 , a) be the imbedded voltage graph shown in

Figure 4.16 whose voltages are in ~8 • The net voltages on the boundary walks
are 3, 4, and I, from which it follows that surface sa has six faces, and that its
Euler characteristic x(Sa) = 16- 24 + 6 = -2. Since sa is evidently con-
nected and orientable, it must be the surface of genus 2. Therefore, by Theorem
4.3.3, we have established that the cyclic group ~8 acts on the surface of genus 2.
(In fact, it can be shown that the full automorphism group of the derived graph
Ga acts on sa, but we postpone discussion of this until Chapter 6.)

4.3.3. Regular Branched Coverings and Ordinary Imbedded Voltage Graphs

We have seen that a regular covering of a graph corresponds to a group action
and to ordinary voltages. Similarly, we shall now see that it is possible to
define a regular branched covering of a surface that corresponds to a group
action on that surface and to an imbedding of an ordinary voltage graph.
Let the group d act on a closed surfaceS. By the "orbit" of a point x E S,
denoted [x], we mean the set {«<>a(x): a Ed}.
Let Sjd denote the set of all orbits, and let p: S-+ Sjd denote the
natural projection p(x) = [x]. The set Sjd is given the strongest topology in
which the natural projection p is continuous. That is, a subset U c Sjd is
open if and only if its inverse image p- 1(U) is open. With such a specification
of open sets, the resulting topological space is called the "quotient space" of
the action of the group d on the surface S, and the map p is called the
"quotient map".
It can be shown that if the group d acts pseudofreely, then the quotient
space Sjd is, in fact, a surface. Moreover, if a point x E Sis not left fixed by
a nonidentity element of d, then some neighborhood of x is moved off itself
by every nonidentity element of d. This ensures that the quotient map
Regular Branched Coverings and Group Actions 187

p: S-+ Sjd is locally one-to-one, except at fixed points. Since the domain is
compact, the quotient map is a branched covering. A branched covering
obtained this way is called "regular", and the group d is called the group of
"covering transformations". Note that p o IJ>a = p for any a Ed.

Example 4.2.1 Revisited. Once again we consider the complex analytic function
p(z) = zn. If p(z 1 ) = p(z 2 ), then the points z 1 and z 2 have the same modulus,
and their arguments differ by an integer multiple of 2'1Tjn. Thus, they are in the
same orbit of the pseudofree action of the cyclic group !!En on the extended
complex plane generated by a rotation about the origin through an angle of
2'1Tjn. It follows that the function p can be viewed as the quotient map of this
action, and accordingly, p is a regular branched covering.

This example indicates a general procedure for showing that a given

branched covering p: S-+ S is regular. It suffices to find a group d acting
pseudofreely on the covering space S and satisfying this condition for every
pair of points x 1 and x 2 in S: p(x 1 ) = p(x 2 ) if and only if x 1 and x 2 are in
the same orbit of the action of d.
The following extension theorem for regularity in the branched covering is
a sharpening of Theorem 4.2.1 for ordinary voltages. We omit the proof, since
it is routinely related to the proofs of Theorems 4.2.1 and 4.3.3.

Theorem 4.3.4. Let (C-+ S, a) be an imbedded ordinary voltage graph. The

natural projection p: ca. -+ C extends to a regular branched covering p: sa -+ S
that has at most one branch point inside each face. If the net voltage on a face
has order r in the voltage group d, then that face has a branch point of deficiency
#d- #djr, each of whose prebranch points has order r. D

Corollary. Let (C -+ S, a) be an imbedded ordinary voltage graph satisfying

KVL on every face. Then the natural projection p: ca.-+ C extends to an
(unbranched) regular covering. D

4.3.4. Which Regular Branched Coverings Come from Voltage Graphs?

As the following theorem shows, regular surface coverings arising from ordinary
voltage graphs are not unusual. The proof is a routine sharpening of the proof
of Theorem 4.2.2 and is omitted.

Theorem 4.3.5. Let p: S-+ S be a regular branched covering of surfaces, and

let C -+ S be any cellular imbedding in the base surface having at most one
branch point in any face and no branch points in C. Then there is an ordinary
voltage assignment a in the group of covering transformations of p such that p is
equivalent to a natural surface projection sa -+ S. D

Theorem 4.3.5 is an important aid to understanding pseudofree group

actions and regular branched coverings for surfaces. It implies that any regular
188 Imbedded Voltage Graphs and Current Graphs

branched covering behaves around branch points the same way that a natural
projection of an imbedded voltage graph behaves on faces. Thus we have the
following information about branch points.

Corollary. Let p: S ~ S be a regular branched covering with covering transfor-

mation group .91. Then for each branch pointy of the base surface, there is an
element b E .91 such that, if r is the order of bin the group .91, then the number
of prebranch points of y is #.9/jr, and each has order r.

Proof By Theorem 4.3.5, the covering p is a natural surface projection

sa ~ S for some imbedded voltage graph (G ~ S, a). The prebranch points
of the branch point y correspond to the faces in the derived imbedding that
cover the base face containing y. Thus, the conclusion of the corollary follows
directly from an application of Theorem 4.1.1 in which the group element b
equals the net voltage on the face containing the branch point y. 0

This corollary provides most of the information we need for later use, but
much more can be said. For instance, the subgroup of the covering transfor-
mation group .91 that leaves fixed a particular prebranch point of y is a cyclic
group, and it is generated by a conjugate of the element b (see Exercise 13).
Many consequences of Theorem 4.3.5 result from the freedom allowed by
this theorem in the choice of the base imbedding G ~ S. Two such conse-
quences are now derived. The first is of general importance, especially for the
study of the genus of groups in Chapter 6. The second is narrower, but deeper,
and it draws upon a variety of results developed so far.

Theorem 4.3.6. If the group .91 acts pseudofreely on the connected surfaceS,
then there is a Cayley graph for .91 that is imbeddable in S.

Proof Just let a bouquet of circles assume the role of the graph G of
Theorem 4.3.5. By Theorem 2.2.3, the derived graph Ga is a Cayley graph for
.91. 0

Theorem 4.3.7. No finite group .91 acting on an orientable surface Scan have
exactly one fixed point.

Proof Since any element of the group .91 leaving a point fixed generates a
cyclic subgroup leaving that point fixed, it suffices to prove the theorem for the
case in which the group .91 is cyclic. Under that additional condition, let
S = S/.91 be the quotient surface, and let p: S ~ S be the quotient map. If
the action of .91 has only one fixed point, then the regular branched covering
p: S ~ S has exactly one prebranch point, and its order must be #.9/.
Suppose first that the quotient surface S is not the sphere. Then let G ~ S
be a one-face imbedding of a bouquet of circles containing the single branch
point of the quotient map p in its only face. The boundary walk of that face is
Regular Branched Coverings and Group Actions 189

of the form
a 1 b1 a 1 - 1 b1 - 1 a 2 b2 a 2 - 1 b2 - 1 ••• if Sis orientable
c1 c1 c2 c2 c3 c3 • • • if S is nonorientable
By Theorem 4.3.5, there is a voltage assignment a such that the quotient map
p is equivalent to a natural projection sa ~ s. If the quotient surface s is
orientable, then the net voltage on the only boundary walk is the identity,
since the group .91 is abelian. This would contradict the existence of a branch
point of p. On the other hand, if the quotient surface S is nonorientable, then
the net voltage is of the form 2a for some a E .91. It follows that #.91 cannot
be even, since then the order of the net voltage 2a and the associated
prebranch point would be at most #.91/2, that is, not equal to #.91. However
if #.91 is odd, then the derived surface would be nonorientable, by Theorem
4.1.5. Thus, in either case, there is a contradiction.
Finally, if the quotient surface S is a sphere, then let G ~ S be an
imbedding of a single loop containing the branch point of p in one of its two
faces. However, the other face has the same boundary walk and net voltage, so
it must also contain a branch point. D

4.3.5. Applications to Group Actions on the Surface S 2

The Riemann-Hurwitz equation is a surprisingly powerful tool for determin-
ing which groups act pseudofreely on which surfaces. As an example of its
power, we shall show how it can be used to prove that there is no pseudofree
action of the cyclic group !!En on the surface S2 of genus 2, for any prime
number n > 5. Since Examples 4.3.1 and 4.3.2 exhibit actions of !!En on the
surface S2 for n = 2, 3, and 5, this demonstration completes a classification of
the prime, cyclic, pseudofree actions on s2 .
Let n > 3 be prime, and suppose there is a pseudofree action of !!En on the
surface S2 • Then the quotient map p: S2 ~ S2/f!En is a regular n-sheeted
branched covering. By the corollary to Theorem 4.3.5, it follows that each
branch point has deficiency of the form n - njr, where r is the order of a
voltage. However, since n is prime, we infer that the only possible deficiency is
n - 1. Therefore, if x(S2/f!En) = k, and if the number of branch points is t, it
follows from the Riemann-Hurwitz equation that
- 2 = X ( S2 ) = nk - t ( n - 1)
Since n > 2, it is impossible that t = 0. Moreover, since n - 1 > 2, it follows
that k must be positive. Since the group order n is odd and n - 1 is even, it
also follows that k must be even. Since k is the Euler characteristic of a
surface, the only possibility is that k = 2. Thus
-2 = 2n- t(n- 1)
Solving this equation for n, we obtain n = 1 + 4/(t - 2), which is at most 5.
190 Imbedded Voltage Graphs and Current Graphs

Information about the number of fixed points in a group action can also be
obtained directly from the Riemann-Hurwitz equation. In the preceding
computations, if n = 5, then the number t of branch points equals 3, and the
quotient surface is the sphere. Since each branch point has deficiency 4, there
are also only three prebranch points in all. Prebranch points of a regular
covering are just the fixed points of the group of covering transformations.
Therefore, any pseudofree action of ~5 on S2 has exactly three fixed points,
as is the case in Example 4.3.2. A similar inspection of the preceding
computations shows that any pseudofree action of ~3 on S2 having the sphere
as quotient space has four fixed points, as is the case in Example 4.3.1 (the
possibility of the torus as a quotient is considered in Exercise 7). Deeper
applications of the Riemann-Hurwitz equation are deferred until Chapter 6.

4.3.6. Exercises
1. Show that every finite group acts freely on some surface.
2. Draw a band decomposition for the surface S3 that reveals an action of
~7•
3. Prove that the cyclic group ~n acts freely on the orientable surface S if
and only if n divides x(S).
4. Generalize Exercise 3 to arbitrary finite abelian groups.
5. Use the Riemann-Hurwitz equation to prove that any pseudofree action
on the projective plane has exactly one fixed point.
6. Prove that no group of even order acts pseudofreely on Nk for k odd.
(Hint: Use Exercise 4.2.12 and the fact that any group of even order has
an element of order 2.)
7. Show that every pseudofree action of ~3 on the surface S2 has exactly
four fixed points.
8. Let p: S-+ S be a regular branched covering such that the surface S is
orientable, and suppose that the covering transformation group is abelian.
Prove that p cannot have only one branch point. Show by examples that
p can have exactly one branch point if .91 is nonabelian or if S is
nonorientable. Does this contradict Theorem 4.3.7?
9. Prove that ~12 does not act pseudofreely on the surface S2 • (This is like
showing that ~16 does not act pseudofreely on S2 , but there are more
cases to consider.)
10. Complete the classification of cyclic pseudofree actions on S2 by showing
that ~n acts pseudofreely on S2 if and only if n = 2, 3, 4, 5, 6, 8, 10.
11. Prove for prime numbers n > 2 that ~n acts pseudofreely on the surface
Sg if and only if n is of the form (2g- 2 + m)/(m- k), where k is
even, k .:5; 2, and m > 1.
12. Use Exercise 11 to find for which primes n there is a pseudofree action of
~n on S5 •
Current Graphs 191

13. Prove in regard to the corollary to Theorem 4.3.5 that the subgroup of
the covering transformation group .511 that leaves fixed a particular
prebranch point of x must be cyclic and generated by a conjugate of the
element b.

4.4. CURRENT GRAPHS

The historical origins of topological graph theory lie largely in map-coloring

problems concerned with the relationships between regions on a surface. From
our present perspective, it should not be surprising that duality plays a central
role. In fact, the main method used in the original solution of the Heawood
map-coloring problem is a dual form of the voltage graph called a "current
graph". The precursors of this technique can be traced back through Ringel's
work to Heffter (1891).

4.4.1. Ringel's Generating Rows for Heffter's Schemes

In trying to establish that the Heawood number for the surface Sg is the
chromatic number, and not just an upper bound, Heffler (1891) sought the
minimal-genus surface consisting of n regions (faces), each of which is a
neighbor of all the others. Heffter did consider the problem simultaneously in
primal and dual forms, and he also wished to compute the genus of the
complete graph Kn. To describe the surface, he started with n (n- I)-sided
polygons, and he labeled them 1, 2, ... , n. The sides of polygon i were then
labeled by
1,2, ... ,i-l,i+ I, ... ,n

to indicate where the neighboring polygons should be attached. Figure 4.17,

which is taken from Heffter's paper, illustrates this procedure for a surface
with five regions. The identifications along the sides of polygon 1 have already
been made. When the remaining sides are pasted together, the segments AB
and DC are identified as are the segments AD and BC, thereby yielding a
torus.
For the most part, Heffter did not depend on pictures and, instead, gave a
"table" of n rows, in which row i lists in cyclic order the labels for the sides of
polygon i. The table he gave for the preceding example of five regions was

(1) 3 245
(2) 4 3 51
(3) 5412
(4) 15 2 3
(5) 2134
192 Imbedded Voltage Graphs and Current Graphs

A 5

2 3 4

1 B
2 3 4

4 5 1 5 1 2 1 2 3

D 3 4 5
1

3 4 5

2 c
Figure 4.17. Five labeled polygons.

We now recognize such a table as a rotation system for an imbedding of the

complete graph K 5 , to which the mapping of Figure 4.17 is a dual. Heffter
showed how the cycle of polygons meeting at a vertex can be recovered from
the table; that is, he gave a face tracing algorithm for the imbedding of K n.
Having thus realized the "purely arithmetical nature of the problem", he
constructed tables yielding minimal imbeddings of K n for all n :;;; 12. Al-
though Ringel was well aware of Heffter's methods, their present popularity
evidently rose because of the abstract of Edmonds (1960).
In addition to the use of duality and the introduction of rotation systems, it
is especially noteworthy that Heffter gave tables for K n, n = 5, 6, 7, in which
row i + 1 was obtained by adding 1 modulo n to the numbers of row i. Heffter
proved that such conveniently generated tables exist for general n > 6 only if
n is of the form 12s + 7, and he constructed them in the cases in which 4s + 3
is prime and 2k is not congruent to ± 1 modulo4s + 3 for 0 < k < 12s + 1.
(It is still not known whether there are infinitely many such s.) These tables
generated from a single row using the cyclic group frn have ultimately evolved
into voltage graphs. For example, the table given above for n = 5 is precisely
the derived rotation system for the voltage graph of Example 4.1.3.
When Ringel began his work on the genus of complete graphs, he also
chose to describe surfaces with tables like Heffter's, calling them "schemes".
In an important generalization of Heffter's work, he designed the strategy of
letting most of the rows of the scheme be generated from a small number of
rows. Unlike Heffter, however, he had spectacular success with this technique.
For instance, Ringel (1961) obtained the minimal imbedding of Kn for all n of
the form 12s + 7, by constructing a single generating row for the scheme that
Heffter had been unable to obtain without restricting s. Despite Ringel's
Current Graphs 193

2 1

Figure 4.18. A current graph with currents in .2'19 .

particular brilliance in constructing such generating rows, a general method to

find them was still lacking. Such a method was ultimately provided in
rudimentary form by Gustin (1963).

4.4.2. Gustin's Combinatorial Current Graphs

Gustin began with a planar drawing of a graph G. At each vertex he assigned a
direction, clockwise or counterclockwise. This led to what Gustin thought of as
flows within the graph, which we now recognize as face boundaries. Each edge
was assigned a direction and a "current", by which he meant an element of
some group .91.

Example 4.4.1. The current graph in Figure 4.18 leads to a single generating
row for a triangular imbedding of the complete graph K 19 in the surface S20 • To
get that row (in our present notation), write the vertex v0 at the left, which is to
show which vertex has the rotation specified by the row. Then proceed as follows.
Start at any directed edge, for instance, the one labeled with current
9 modulo 19. Write that current as the subscript on the first vertex in the row.
(Since the resulting graph will have no multiple edges, the entries in the rows of
the scheme can be vertices, rather than directed edges.)
At the end of that first directed edge is a filled vertex, indicating that the
second edge is to be obtained by moving in a clockwise direction. That is, the
directed edge with current 7 is next, so write v 1 next in the generating row.
Continue in this manner, going clockwise at every solid vertex (•) and counter-
clockwise at every hollow vertex ( o ), until every edge has been traversed in both
directions. When proceeding in the minus direction on an edge, one writes the
additive inverse modulo 19 of the current shown on the plus direction. The
resulting generating row is
v0 • v9 v 1 v8 v3 v13 v15 v14 v11 v18 v4 v 11 v10 v 16 v5 v 1 v 12 v2 v6

The next row is

In general, the subscript on the jth entry in row v, is obtained by adding

i modulo 19 to the subscript on the jth entry of row v0 •
194 Imbedded Voltage Graphs and Current Graphs

Gustin imposed a number of rules on the graph, on the assignment of

rotations, and on the assignment of currents, whose total effect is to guarantee
that the resulting scheme described an imbedding of a Cayley graph for the
current group. The Kirchhoff current law (KCL) is that the sum (or product,
for nonabelian currents) of the currents leaving each vertex is the group
identity. Gustin's requirement that the graph be regular and that every vertex
satisfy KCL has the effect of ensuring that every face in the resulting
imbedding has the same number of sides as the valence of every vertex in the
current graph.
Gustin's current graphs were regarded as computational tools ("nomo-
grams" in Youngs's terminology) to aid in the construction of generating rows
for schemes. Progressively more constructive power was obtained over the next
few years as Gustin's original rules were relaxed. Very importantly, Youngs
(1967) explained how to use current graphs with excess currents at some
vertices. Whereas Gustin (1963) discussed only the cases of one, two, or three
generating rows, Jacques (1969) gave a general exposition, still restricted to
Cayley graphs, but covering the possibility of arbitrarily many generating
rows. Finally, Gross and Alpert (1973, 1974) showed how a topological
viewpoint makes it possible to eliminate all of Gustin's restrictions, and they
unified all previously defined kinds of current graphs into the one now
described.

4.4.3. Orientable Topological Current Graphs

Let G -+ S be an imbedding in an oriented surface. An "(ordinary) current
assignment" is a function /3 from the set of directed edges of G into a group !14
such that {J(e-) = {J(e+)- 1 for every edge e. The values of f3 are called
"currents" and !14 is called the "current group". The pair (G -+ S, /3) is called
a current graph. If the clockwise and counterclockwise directions used by
Gustin to order the entries in a generating row are reinterpreted as vertex
rotations, the result is a legitimate imbedding. Hence, Gustin's combinatorial
current graphs are properly considered a special case of a topological covering
space construction.
The point of assigning currents to an imbedding is to obtain a "derived
graph Gp" and "derived imbedding Gp -+ Sp''. The vertex set of Gp is the
cartesian product Fe X !14, and the edge set is the product Ec X !14. Thus,
vertices of the derived graph correspond to faces of the current graph, as in
Gustin's construction. Endpoints and plus directions of derived edges are
given as follows. The plus direction of the derived edge (e, b) has initial vertex
(/,b) and terminal vertex (g, be), where the rotation at the initial vertex of
the directed edge e+ carries face f to face g and /3(e+) =c. Thus, the
directed edge ( e, b) - goes from the vertex ( g, be) to the vertex (/, b). It will
be notationally convenient to refer sometimes to the directed edge ( e, be-l)-,
which goes from vertex (g, b) to vertex (/, bc- 1), as (e-, b); also, we let
(e+, b) stand for (e, b)+. The rotation system for the derived imbedding is
Current Graphs 195

obtained by lifting boundary walks of the imbedding G ~ S, as follows. If the

directed boundary walk of face f is

e n•.

then the rotation at vertex ( J, b) is

(e1, b1)'' (e2, b2)'2 ... (en, bn)'•

where b; = b if f.;= +, and b; = b/3(e;-) if f.;= - . This rotation is more

easily denoted

Observe that the rotation at vertex (/, ab) is obtained from the rotation at
vertex (/, b) by multiplying all ?A coordinates on the left by a.

Example 4.4.1 Revisited. For j = 1, ... , 9, let e1 be the edge whose plus
direction carries the current j. Gustin's method of obtaining a generating row is
clearly no more than an application of the Face Tracing Algorithm. Thus the base
imbedding G ~ S has one face whose directed boundary walk is

Denote this face by v. For j = 1, ... , 9 and i = 1, ... , 19, the derived edge
(e1 , i)+ runs from vertex (v, i) to vertex (v, i + j). Thus, the derived graph is
the complete graph K 19 • The rotation at vertex ( v, 0) is

Since the derived graph is simplicial, this rotation can also be given in vertex
form. If we let V; = ( v, i), the result is

which is the generating row given by Gustin's construction. Moreover, the full
rotation system is also the same as Gustin's scheme, since the rotation at vertex v;
is obtained from that at vertex v0 by adding i modulo 19 to all ?A coordinates.

Example 4.4.2 (Jungerman, 1975). Consider the current graph (G ~ S, /3)

given by Figure 4.19, where the current group is .2"16 • An application of the Face
Tracing Algorithm shows that the base imbedding has two faces, which we denote
u and·v, and that vertices u 0 and v0 of the derived imbedding have the following
rotations:
uo. v3 u2 V1s u6 v7 Vs Vu vl3 u14 v1 uw v9
Do. u3 v14 u9 U1s Un v6 Un v2 Us Vw u7 u1
196 Imbedded Voltage Graphs and Current Graphs

3
7

6
2 6

1 7
3
Figure 4.19. A current graph with currents in .2"16 .

An inspection of these two rows reveals that the vertex u; is adjacent to the vertex
v1 if and only if j - i is odd, and that u; is adjacent to u1 (similarly V; and v) if
=
and only if j - i 2 mod 4. It follows that { u;: i even} U { v/ j odd} is the
vertex set for one component of the derived graph. In this component, every pair
=
of vertices is adjacent except u; and u1 , where j - i 0 mod 4 and V; and v1 ,
=
where j - i 0 mod 4. Thus, this component is isomorphic to the complete
multipartite graph K 4 •4 • 4 •4 • The other 16 vertices of the derived graph lie in a
second copy of K 4 4 4 4 .

4.4.4. Faces of the Derived Graph

Just as the vertices of a derived graph correspond to the faces of its current
graph, faces of a derived graph correspond to vertices of its current graph. The
proof of the following theorem is postponed to the next section, where it is
shown to be a simple consequence of the duality between current and voltage
graphs. Originally, the theorem was proved by an application of the Face
Tracing Algorithm to the derived imbedding.

Theorem 4.4.1 (Gross and Alpert, 1974). Let e 1• 1 ••• en•" be the rotation at
vertex v of the current graph (G ~ S, {J), and let c; be the current carried by the
direction of the edge e; that has v as its initial vertex. Let c = c1 • . • cn, and let r
be the order of c in the current group 114. Then the derived imbedding has #114/r
faces corresponding to vertex v, each of size rn, and each of the form
(el•',b),(e2• 2 ,bcl),(e3• 3 ,bC1C2) ... (en•n,bclc2 ... cn-1)
( e 1•', be), ( e 2•2 , bcc 1 ), ( e 3•3 , bcc1c 2) ...

(e 1•',bc')=(e 1•1 ,b) D

The product c1 ... en in Theorem 4.4.1 is called the excess current at vertex
v. The Kirchhoff current law (KCL) holds at vertex v if the excess current is
the identity.
Current Graphs 197

Figure 4.20. A current graph with currents in .2'6 .

Corollary. Let (G-+ S, /3) be a current graph such that KCL holds at every
vertex and every vertex has valence three. Then the derived imbedding is
triangular. D

Since KCL holds at every vertex in both Example 4.4.1 and Example 4.4.2,
both derived imbeddings are triangular. Thus the first gives a minimal-genus
imbedding of K 19 , whereas the second gives a minimal-genus imbedding of
K4.4.4.4·

Example 4.4.3. The current graph given in Figure 4.20 with currents in .2'6 has
one face, denoted v. By face tracing, we see that the rotation at vertex v0 of the
derived imbedding is

The order of the excess current at the bottom vertex of the current graph is 2. If
we excise the resulting three digons of the derived imbedding and reclose the
surface by identifying opposite sides, we thereby identify three pairs of multiple
edges from D; to D;+J• fori= 1, 2, 3. The final product is an imbedding of the
complete graph K 6 • The 3-valent vertex of the current graph satisfies KCL and it
generates six triangles, and the two remaining !-valent vertices together yield one
hexagon and two triangles. Thus, the Euler characteristic of the derived surface is
6 - 15 + 9 = 0. The scheme whose generating row is the given rotation for v0
with the extra v 3 deleted is precisely the table Heffter gives for his minimal-genus
imbedding of K 6 •

Example 4.4.4. Consider the current graph given in Figure 4.21 with currents in
the cyclic group .2'100 • There are two faces, which we denote u and v. In the
derived graph, there is an edge between U; and vi+l for all i, and there are no
other edges. Therefore the derived graph has two components; the vertex set of
one component is
{ u;: i even} u { v/ j odd}

Each vertex of the current graph has excess current ± 2 and valence 2, and
thereby generates a pair of 100-gons; each of these 100-gons passes through every
vertex in its component. Add a vertex inside each of the six 100-gons of one
198 Imbedded Voltage Graphs and Current Graphs

1 1

Figure 4.21. A current graph with currents in .2"100 .

component, and add edges joining that vertex to the 100 original vertices. Then
delete all the original edges. The resulting graph is the complete bipartite graph
K 6 , 100 , and every face of the resulting imbedding is a quadrilateral. Hence the
imbedding has minimal genus.

Clearly, Example 4.4.4 generalizes to arbitrary even integers m and n,

yielding quadrilateral and hence minimal-genus imbeddings of K m, n for these
cases. Exercise 11 shows how quadrilateral imbeddings of Km, n can be
=
obtained when m is odd and n 2 mod 4.

4.4.5. Nonorientable Current Graphs

It is possible to consider current graphs in nonorientable surfaces as well. The
difficulty is that vertex rotations are reversed by type-1 edges during the Face
Tracing Algorithm, so care must be taken in using vertex rotations to define
the derived graph. For instance, suppose that e is a type-1 edge lying between
faces f and g, as illustrated in Figure 4.22. Unlike the orientable case, the
rotation at either endpoint of edge e sends face f to face g. If e+ carries
current c and e- were to carry current c-I, then in the derived graph, there
would be edges corresponding to e running from vertex(/, b) to both of the
vertices (g, be) and (g, bc- 1 ). Therefore, both directions of a type-1 edge must
be assigned the same current. Also, in tracing faces of the derived graph, the
edge e is traversed in the same direction twice; it follows that in computing
the excess currents at the endpoints of edge e, the current c must be used
twice. Thus type-1 edges in nonorientable current graphs ("cascades" in
Youngs's terminology) appear in print with arrows in both directions.

e
E
0 Figure 4.22. A type-1 edge.
Current Graphs 199

We conclude that a current graph (G ~ S, /3) in a nonorientable surface

must satisfy the rule that

if e is type 0
if e is type 1

Note the implication that the vertex rotations and edge types of the imbedding
G ~ S must be specified before currents are assigned (even in the orientable
case, an orientation of the imbedding surface must be chosen to specify vertex
rotations). The derived graph Gp is then defined as before.
To define the derived imbedding Gp ~ Sp, a direction for each boundary
walk must also be specified, since the Face Tracing Algorithm does not
determine a unique direction in the nonorientable case. It is also necessary to
specify at which corners a vertex rotation is reversed while tracing a face; such
a corner is called "reversed". Now suppose that the directed boundary of face
f is e 1'' . . . en••. The rotation at vertex (/, b) in the derived imbedding is then

(e 1 ,b 1 )'' ••• (en,bJ'·

where b; = b if f.;= + and the corner e;_ 1e; is not reversed or if f.;= - and
the corner e,e;+I is reversed, and where b; = bf3(e+)- 1 otherwise. Edge types
of the derived imbedding are defined so that if both directions of edge e
appear in the directed boundary walks of the imbedding G ~ S, then edge
( e, b) is type 0 for all b E ?A, and otherwise edge ( e, b) is type 1 for all
bE PA.

Example 4.4.5 (Youngs, 1968a). Let (G ~ S, /3) be the current graph il-
lustrated in Figure 4.23 with currents in .2"15 • As before, a filled vertex indicates
a clockwise rotation and unfilled, counterclockwise. Youngs's convention of
drawing type-1 edges broken in the middle with arrows in both directions,
suggestive of the usual " X", is used. The imbedding has one face, which we shall
call v. If we let e; denote the edge carrying current i, then the directed boundary
of face vis

1 5

6 3

2 4

Figure 4.23. A nonorientable current graph with currents in .'2j 5 .

200 Imbedded Voltage Graphs and Current Graphs

Again following Youngs, reversed corners are indicated by dots. The rotation in
the derived imbedding at vertex ( v, 0), in edge form, is

where type-1 edges are given the superscript 1, as usual. The vertex form of the
rotation with ( v;, 0) = V; is

The rotation at vertex V; is obtained by adding i modulo 15 to the subscripts in the

rotation at v0 • The derived graph Gp is clearly the complete graph K 15 •

Faces of the derived graph are determined exactly as in the orientable case.
In Example 4.4.5 each of the 3-valent vertices satisfies KCL and, accordingly,
each generates 15 triangles in the derived graph. Each of the two 1-valent
vertices has an excess current of order 15 and induces in the derived im-
bedding a 15-gon whose boundary passes through each vertex. If an extra
vertex is placed in the center of each 15-gon and joined by edges to the
original vertices, the result is a triangular imbedding of the graph K 17 - K 2 •
An extra twist in an edge common to the two original15-gons allows the extra
two vertices to be joined by an edge, thereby yielding a minimal nonorientable
imbedding of K 17 •

Example 4.4.6. The current graph given by Figure 4.24 with currents in .2'13
has a single face v. Since there is an edge in the derived graph between vertices V;
and vi+J• for every i and for j = 1, 2, 3, 4, 5, 6, the derived graph is isomorphic to
the complete graph K 13 • Since KCL holds at every vertex, the derived imbedding
is triangular.

The derived imbedding of a nonorientable current graph can be orientable

(see Exercise 12). A criterion for deciding when the derived imbedding is
orientable is best understood in terms of dual voltage graphs and is considered

9 3 Figure 4.24. A nonorientable current graph with currents in

~3·
Current Graphs 201

in Section 4.5. For present purposes, it suffices to know that a current graph in
a nonorientable surface has a nonorientable derived graph if the current group
has odd order.

4.4.6. Exercises
I. Give a current graph whose derived imbedding is a quadrilateral im-
bedding of the complete graph K 5 •
2. Draw a current graph that yields a triangular imbedding of the complete
graph K 7 • Give the resulting rotation system (scheme) in vertex form.
3. Find the first two rows in vertex form, of the rotation system for the
derived imbedding of Example 4.4.6.
4. Describe the derived imbedding for the current graph given in Figure
4.25a; the current group is .2"21 •
5. Show how a triangular imbedding of the complete 4-partite graph K 5 5 55
can be obtained from the current graph in Figure 4.25b. The cuire~t
group is .2"10 X .2"2 (Jungerman, 1975).
6. Obtain a triangular imbedding of K 10 - K 3 by changing the current
group in Figure 4.20.
7. Change the current group in Exercise 2 to get a triangular imbedding of
the octahedral graph 0 4 •
8. Assign currents in .2"13 to the nonorientable imbedding given in Figure
4.26a so that the derived imbedding can be used to obtain a nonorienta-
ble, triangular imbedding of K 16 - K 3 •
9. Complete the current assignment in .2"25 for the nonorientable im-
bedding given in Figure 4.26b in order to yield a nonorientable triangular
imbedding for K 25 •
10. Find the derived graph for the current graph illustrated in Figure 4.27
with current group .2"15 • (Hint: Show that the current graph has three
faces u, v, and w such that one component of the derived graph has
=
vertex set {(u, i), (v, j), (w, k) I i Omod 3, j =
1 mod 3, k 2 mod 3.} =
(5, 1) (1, 0)

(4, 1)
(0, 1)

(5,0) (1, 1)

2
Figure 4.25. Two current graphs; the cur~ent groups are, on the left, .2"21 and, on the right,
.2"w X .2"2 ·
202 Imbedded Voltage Graphs and Current Graphs

(a) (b)
Figure 4.26. Two nonorientable current graphs.

7 5 4 2
6

7 3 4

1 7 4 2

Figure 4.27. A current graph with currents in .2"15 .

11. =
Suppose that m is odd and n 2 mod 4, where m > 2 and n > 2. Assign
currents in ~n to an m-cycle so that the derived graph can be used to
construct a quadrilateral imbedding of K m, n. (One current will be 3, and
all others will be 1.)
12. Give a nonorientable current graph whose derived graph is orientable.

4.5. VOLTAGE-CURRENT DUALITY

The voltage and current graph constructions have obvious similarities, with
faces of a current graph playing a role of vertices of an imbedded voltage
graph and vice versa. We now give details of the formal duality.

4.5.1. Dual Directions

Since both voltage and current assignments work with directed edges, 1t IS
necessary to have a convention to specify the plus direction of a dual edge,
given the plus direction of the corresponding primal edge. Let G -+ S be a
locally oriented graph imbedding with plus directions specified for all edges.
The "dual plus direction" of the edge e*, denoted e+*, is the one crossing the
Voltage-Current Duality 203

e+*

+
I
I ..e+ • u
I
I Figure 4.28. The dual plus direction.

primal edge e in the same direction as the orientation at the initial vertex of
e+ (see Fig. 4.28).
It is important to note that the dual plus direction of e* depends as much
on the choice of orientation at the initial vertex of e + as on the choice of plus
direction for the edge e. Indeed, changing the orientation at the initial vertex
of e+ always changes the dual plus direction, but changing the plus direction
of edge e does not change the dual plus direction if the edge is type 1. One
unfortunate consequence of this dependence on local orientations is that the
dual of the dual plus direction is not necessarily the primal plus direction. This
is partly because there is no natural way to orient dual vertices; the Face
Tracing Algorithm can give different directions to a boundary walk around a
primal face depending on the choice of initial vertex for the walk. In some
cases, however, there is no way at all of orienting dual vertices that avoids this
problem.

Example 4.5.1. Figure 4.29 shows part of a locally oriented imbedding with
dual edges indicated by broken lines. The edge e is type 1 and the edged is type
0. If a clockwise orientation is chosen for the dual vertex f*, then d+ ** = d-; if
a counterclockwise orientation for the face f is chosen instead, then e + ** = e-.

On the other hand, suppose that S is an orientable surface and the

imbedding G ~ S has been locally oriented by an orientation of the surface S.
Then the Face Tracing Algorithm gives every boundary walk a direction
consistent with the opposite orientation of the surface. Thus the natural local
orientation of the dual graph is given by the opposite orientation of the
surface. In this case, the dual of the dual plus direction is the primal plus
direction (see Fig. 4.30). Hence, for an orientable surface, dual directions
behave nicely as long as the primal graph is locally oriented by one orientation
of the surface and the dual graph by the opposite.
It is still necessary to specify a orientation of the surface for the primal
imbedding, but this is done automatically by a pure rotation system for the

Figure 4.29. A locally oriented imbedding and its dual.

204 Imbedded Voltage Graphs and Current Graphs

~
I e
I "'

Figure 4.30. A directed edge and its dual.

imbedded graph. If the imbedding is given by a rotation projection, the

convention that all rotations are clockwise specifies an orientation. Only if the
imbedding is given by drawing the graph upon a polygon with edges to be
identified must care be taken to indicate the chosen orientation (by a circular
arrow, for example).

Waming. Drawing a dual imbedding is confusing if the primal imbedding is

given by a rotation projection, but relatively easy if the primal imbedding is
drawn on a polygon with edges to be identified. It frequently pays to convert a
presentation of the former type to one of the latter type before attempting to find
the dual. An alternative is first to list the directed boundary walks of the primal
imbedding and then to use the resulting rotation system for the dual to draw a
picture of the dual. In fact, dual imbeddings in orientable surfaces can be
computed in terms of pure rotation systems simply by composing two permutations
(see Exercise 4); this method is obviously the best suited for computers.

4.5.2. The Voltage Graph Dual to a Current Graph

Let (G ~ S, /3) be a locally oriented current graph, and let G* ~ S be the
dual imbedding with dual plus directions. The "voltage assignment dual to
{3 ", denoted {3*, is the one satisfying the rule /3*( e+ *) = {3( e+). The resulting
imbedded voltage graph (G* ~ S, {3*) is called the "imbedded voltage graph
dual to current graph (G ~ S, /3)".

Example 4.4.1 Revisited. Paying heed to our previous warning, we first draw
the given current graph on a polygon with edges to be identified. It is readily
verified that the rotation at each vertex in Figure 4.31a is the same as that at the
corresponding vertex of Figure 4.18 (the specified orientation of the surface S2 is
clockwise). Therefore, the imbedding given in Figure 4.31a is equivalent to the
original imbedding. Its dual imbedded voltage graph is given in Figure 4.3lb.
The corners of the octagon represent a single vertex of the dual graph, and edges
of the polygon are also edges of the dual graph. It is instructive to check that the
rotation at the dual vertex, when read counterclockwise, agrees with the first row
of the rotation system for the derived graph given in Example 4.4.1, and that all
dual plus directions are correctly given.

The dual relationship between the current and voltage graph constructions
can now be given explicitly. In its original form, before the development of
Voltage-Current Duality 205

- c
5

2 2

5
----c-
Figure 4.31. The current graph of Example 4.4.1 and its dual imbedded voltage graph.

voltage graphs, the following theorem stated that the derived imbedding for
any current graph ( H ~ S, /3) is a branched covering of the dual of the
imbedding H ~ S. In its present form, the proof is a routine verification that
the definitions are satisfied, and it is omitted.

Theorem 4.5.1 (Gross and Alpert, 1974). Let ( H ~ S, /3) be a locally oriented
current graph and let (G ~ S, a) be its dual imbedded voltage graph. Then the
derived imbeddings Hp ~ Sp and Ga.~ sa are identical.

Theorem 4.5.1 explains many aspects of current graphs that were not
considered in the previous section, such as orientability and connectivity tests
for current-derived imbeddings, or were considered only briefly and without
justification, such as face tracing (Theorem 4.4.1) and excess current. These
aspects are much easier to study with voltage graphs.

Example 4.5.2 (Alpert). The nonorientable .2'28 current graph given in Figure
4.32 has one face. The derived graph is the Cayley graph for the cyclic group .2'28
with generating set {1, 2, 3, ... , 13}. Thus, it is isomorphic to the octahedral
graph 0 14 • All faces of the derived imbedding are triangles, by KCL, except for
two 28-gons corresponding to the two !-valent vertices. If an extra vertex is
placed at the center of each 28-gon and is joined by edges to each of the original
28 vertices, then the result is a triangular imbedding of 0 15 • But what is the

• 5 7 3 9 1
•
Figure 4.32. A nonorientable current graph with currents in .2'28 .
206 Imbedded Voltage Graphs and Current Graphs

orientability type of the surface? The edges carrying currents -10, 11, -12, and
13 are traversed twice in the same direction in tracing the face of the current
graph imbedding. Therefore, the corresponding loops in the dual imbedded
voltage graph are type 1. Since the voltages -12 and -10 are in the only
index-2 subgroup of .2'28 , the derived imbedding is nonorientable. This current
graph can be generalized to yield nonorientable triangular imbeddings for all
=
octahedral graphs On, n 0 mod 3, n > 3 (see Exercise 10).

A word of caution about local orientation is in order at this point.

Changing the local orientation at a vertex of a current graph reverses the plus
direction of some dual edges. In effect, this simply inverts the value of the dual
voltage assignment on some edges. It should not be surprising that, even in the
orientable case, the two possible orientations of a surface can induce noniso-
morphic derived graphs for the same current graph! This is because there are
voltage graphs such that inverting the voltage assignment to each edge changes
the isomorphism type of the derived graph (see Exercise 8). This phenomenon
can only happen, however, if the current group is nonabelian (Exercise 9).
Therefore, for most applications, it is not a matter of concern.

4.5.3. lbe Dual Derived Graph

Let (H--+ S, {J) be a locally oriented graph, and let (G--+ S, a) be its dual
imbedded voltage graph. There is another graph imbedding that arises natu-
rally in the current-graph construction and has not yet been discussed: the
dual Hp* --+ Sp of the derived imbedding Hp --+ Sp. It is best to think of this
imbedding as p- 1(H)--+ Sp, where p is a natural surface projection Sp--+ S,
since then p restricts to a graph map Hp * --+ H. The imbedding Hp * --+ Sp is
called the "dual derived imbedding", the graph Hp* is called the "dual derived
graph", and the map p: Hp* --+ H is called the "natural projection" for the
dual derived graph.

Example 4.5.3. At the bottom left of Figure 4.33 is a current graph in the
sphere with current group .2'4 • At the bottom right is the dual imbedded voltage
graph. At the upper right is the derived imbedding, the 3-cube. At the upper left
is the dual derived imbedding, the 3-octahedron. Both base graphs can be viewed
as quotient graphs under the .2'4 action generated by a 90° rotation about the
vertical axis. For the cube (derived graph) the action is free; for the octahedron
(dual derived graph) the action has fixed points at the top and bottom vertex.

Let (H--+ S, {J) be a current graph, and let (G--+ S, a) be its dual
imbedded voltage graph. Since the surface projection p: sa --+ S may wrap a
face of the derived imbedding many times around a face of the imbedded
voltage graph, as in Example 4.5.3, the natural projection p: Hp* --+ H of the
dual derived graph onto the current graph is not, in general, a covering map.
Indeed, if r is the order of the excess current at vertex v of the current graph
H, then p: Hp* --+His locally r-to-1 near each vertex of p- 1(v).
Voltage-Current Duality 207

dual derived derived graph

graph imbedding imbedding

duality

1 1

duality

current graph voltage graph

Figure 4.33. A current graph, a dual voltage graph, and derived graphs.

The graph map p: Hp* ~ H is not a branched covering, but rather a folded
covering, because the set of points where the map fails to be a local homeo-
morphism has dimension not 2, but 1, less than the complex H. Parsons et al.
(1980) and Jackson et al. (1981) [also Gross et al. (1982)] use the term
"wrapped quasi-covering" in their papers that analyze the construction, and
they develop it for direct use in graph-imbedding problems. We shall call it
simply a "wrapped covering".
Even if KCL holds at very vertex of a current graph so that the dual
derived graph does not cover the current graph, it is still not easy to determine
the structure of the dual derived graph. The following example illustrates the
problem in one case and presents an ad hoc solution.

Example 4.5.4 (White, 1974). Suppose that .J/1 is an abelian group and
X= { x 1 , ••• , x 2 m} is a set of distinct generators for .J/1 such that X; * x1- 1 for
all i and j. Let H ~ S be the imbedding of a bouquet of ~m circles in the surface
of genus m given by the rotation

Let fJ be the current assignment such that {J(e/) =X;, for i = 1, ... , 2m.
Figure 4.34 illustrates the case m = 2, where the edges of the octagon form the
dual voltage graph.
208 Imbedded Voltage Graphs and Current Graphs

Figure 4.34. A current graph.

Clearly the derived graph is the Cayley graph C( .91, X). We claim this
imbedding is self-dual; that is, the dual derived graph Hp * is also the Cayley
graph C(.Jil, X). Since KCL holds at the one vertex of the current graph H,
the graph Hp * is a regular covering of H, and accordingly, it must be a Cayley
graph for the group .91. However, it is not at all obvious that the generating set
for this Cayley graph is X. To show that it is, we give an explicit labeling of
vertices and coloring of edges that makes Hp * the desired Cayley graph.
By Theorem 4.4.1, for each a E .91, exactly one of the #.91 faces of the
derived imbedding, which we label a, has a directed boundary walk of the
form

This can be simplified, since the group .91 is abelian, to

Using the relation (e 1 -I, ax 1x 2 ) = (e 1 , ax 1x 2 x 1 - 1 )- = (e 1 , ax 2 )-, we further

simplify to

The directed edge ( e 1 , a)- then forms part of the boundary walk of the face
labeled ax 2 -I, and the edge (e 2 , a)+ lies on the directed boundary of the face
labeled ax 1 - 1. If the vertex dual to face a is also labeled a, then the dual edge
( e 1 , a)* has its plus direction from dual vertex a to dual vertex ax 2 , and the
plus direction of ( e 2 , a)* goes from dual vertex ax 1 - to dual vertex a. The
other pairs of edges e 2 ;_ 1 and e 2 ;, for i > 1, behave similarly. Thus, if the
plus direction of the dual edge (e2i_ 1 , a)* is colored x 2 , and the minus
direction of the dual edge (e 2 ;, a)* is colored x 2;_ 1 , then the result is a Cayley
color graph for the group .J/1 and generating set X.
If the abelian group .J/1 in Example 4.5.4 is the cyclic group .2'4 m + 1 and
X= {1, 2, ... , 2m}, then the result is a self-dual imbedding of the complete
graph K 4 m+ 1 • The graph Kn has a self-dual imbedding only if n 0,1 mod4 =
Voltage-Current Duality 209

(0, 0) Figure 4.35. A .2"n X .2"n-current graph.

(Exercise 3); Stahl (1979) and Pengelley (1975) have obtained such self-dual
imbeddings of K 4 m+l (see Exercise 11). If the group .511 is ~4 m+ 2 and
X= {1, 2, ... , 2m} again, then the result is a self-dual imbedding of the
octahedron graph 0 2 m+I· Gross (1978) uses a variation of this construction to
compute the least number of edge crossings of certain octahedron graphs
imbedded in certain surfaces (Exercises 12-14).
Although it is difficult to determine the structure of the dual derived graph,
it is possible to give a readily applied criterion to decide when a dual derived
graph of a simplicial current graph is simplicial. The proof of the following
theorem is a straightforward application of Theorem 4.4.1 and is omitted.

Theorem 4.5.2. Let (H ~ S, /3) be a current graph without multiple edges or

loops. Then the dual derived graph has a multiple edge or loop if and only if there
are adjacent vertices with excess currents b and c such that b' = c 1 -:/= 1 for some
sand t.

Example 4.5.5. Let ( H ~ S, fJ) be the current graph given in Figure 4.35 with
the current group ~n X ~n. The excess currents at vertices v 1, v2 , v3 , v4 are
(1, 0), (1, 1), (- 2, 1), and (0, - 2), respectively. As long as the number n is odd,
no multiple (consider the group operation to be addition) of (1, 1) or ( -2, 1) has
=
a zero coordinate. If s(1, 1) = t(- 2, 1), then - 2t t mod n, so 3t 0 mod n. =
=
Thus, if n 1 or 5 mod 6, then the adjacent vertices of H satisfy the condition of
Theorem 4.5.2. Therefore the dual derived graph is simplicial. Since each vertex
of the graph H has excess current of order n, there are n vertices of the dual
derived graph corresponding to each V;, i = 1, ... , 4. Since no vertex correspond-
ing to V; is adjacent to another corresponding to V; and each dual derived vertex
has valence 3n, the dual derived graph is the complete multipartite graph
Kn. n, n. n. The imbedding is triangular and, therefore, minimal, because the dual
voltage graph is regular of valence 3.

Example 4.5.5 can be generalized to any triangular imbedding H ~ S,

where the vertices of graph H can be covered by a disjoint union of paths of
length 3. The derived graph Hp has n vertices v1, ... ,vn corresponding to each
vertex v in H and edges from u; to v1 if and only if there is an edge from u to
210 Imbedded Voltage Graphs and Current Graphs

v in H. Such a graph is denoted H<n> (it is the "composition" of the graph H

by a set of n independent vertices). If His the complete graph Km, then H<n>
is the complete multipartite graph Kn,n, ... ,n = Km(n)· A number of authors
have sought minimal imbeddings for H<n> given a minimal imbedding, usually
triangular or quadrilateral, for H. Bouchet (1982b) has shown, through an
ingenious use of "nowhere-zero flows", that if H has a triangular imbedding,
so does H<n> for any n not divisible by 2, 3, or 5. Jackson et al. (1981) and
Parsons et al. (1980) use wrapped coverings to show that a triangular imbed-
ding for H lifts to one for H<n> for all n relatively prime to a number M that
depends only on the chromatic number of H. Bouchet (1982a) proves by
means of folded coverings that if H has a eulerian circuit, then any triangular
imbedding for H lifts to a triangular imbedding of H<n>, for all n. Bouchet
(1983) and Abu-Sbeih and Parsons (1983) obtain similar results for quadri-
lateral imbeddings of bipartite graphs.

4.5.4. The Genus of the Complete Bipartite Graph Km,,.

In Section 4.4, current graphs were used to construct orientable quadrilateral
imbeddings for the complete bipartite graph Km, n when m and n are both
=
even (Example 4.4.4) or when m is odd and n 2 mod 4 (Exercise 4.4.11). It
is instructive to reconsider these imbeddings in terms of both voltage graphs
and dual derived graphs.
=
Suppose that m is odd and that n 2 mod 4, where m > 2 and n > 2. The
left side of Figure 4.36 illustrates, for m = 7, the .2'n-current graph that was
suggested by Exercise 4.4.11. The vertex having excess current 2 generates an
n-gon in the derived graph, just as the other vertices do, because n/2 is odd.
Accordingly, a quadrilateral imbedding can be constructed from the derived
graph, by adding a vertex inside each face, as in Example 4.4.4. In the middle
of Figure 4.36 is the voltage graph dual to the current graph shown on the left.
Clearly, the derived graph has n vertices and m faces, each bounded by a
hamiltonian cycle.
Since the desired imbedding of Km n is obtained from the derived imbed-
ding by adding a vertex inside each face, one might suspect that a quadri-
lateral imbedding of Km, n can also be constructed directly as a dual derived

0 1

2
Figure 4.36. A voltage graph and two current graphs.
Voltage-Current Duality 211

(0, 1)

(0, 0) (1, 1)

(1, 0) Figure 4.37. A !rm 12 X 1r. 12 -current graph.

graph. On the right of Figure 4.36 is another current graph with currents in
f!En 12 , illustrated for the case m = 7. By Theorem 4.5.2, the dual derived graph
has no multiple edges or loops. Since the top and bottom vertices satisfy KCL,
they generate n/2 + n/2 vertices, each of valence m. Thus, they form one part
of the complete bipartite graph K m, n. The excess current at each of the
2-valent vertices has order n/2 in f!En 12 . Therefore, these vertices generate m
vertices of valence n. Since every face of the current graph imbedding is a
quadrilateral, so is every face of the derived imbedding.
A quadrilateral imbedding of Km n for m and n both even can also be
given as a dual derived graph. In this case, the construction is especially
elegant. The current group is f!Em 12 X f!En 12 , and the current graph is given in
Figure 4.37. It is left to the reader to verify that the dual derived graph is
Km,n·
Since K m, n has no cycles of length less than 4, it follows that

y(Km.n) ~ 1- (m + n- mn + mn/2)/2 = (m- 2)(n- 2)/4.

The last quantity is an integer if and only if m and n are both even, or one or
the other is congruent to 2 mod 4. Quadrilateral imbeddings have already been
constructed for these cases. To finish the computation of y(Km n) for all m, n,
an ad hoc construction is used for the nonquadrilateral imbeddings.

Theorem 4.5.3 (Ringel, 1965). y(Km,n) = f(m- 2)(n- 2)/41.

=
Proof Consider first the irregular case m 3 mod 4, n 1 mod 4. Then =
K m _ 1 n has a quadrilateral imbedding, which for n = 9 looks like Figure 4.38
(left) ~ear a vertex u in the ( m - 1) part of K m _ 1 n. Split the vertex u into
two vertices v and w and join each of these vertices to the nearest "hair' of
the surrounding n vertices as shown in the center of Figure 4.38. Extra edges
from v and w are needed to obtain Km n; this in turn will necessitate some
extra handles. The imbedding surface for Km-l,n has genus (m- 3)(n- 2)/4,
whereas the desired imbedding for Km n has genus f(m - 2)(n - 2)/41, which
= =
for m 3 mod4 and n 1 mod4 is' the same as ((m- 2)(n- 2) + 1)/4.
Thus, there is a leeway of (n- 1)/4 handles that can be used to put in the
extra edges at vertices v and w. The right side of Figure 4.38 shows how to
add the handles: each labeled face at vertex v is joined by a tube to the
212 Imbedded Voltage Graphs and Current Graphs

Figure 4.38. Splitting vertex u.

z w Figure 4.39. The insides of tube 1.

like-labeled face at vertex w. Four edges are then drawn on tube 1, two from v
and two from w, as illustrated in Figure 4.39. Three edges are drawn on tube
=
2, two from v and one from w. For general n 1 mod4, each tube carries
four edges except the bottom tube, which carries three.
=
Similar constructions work for m 3 mod 4 and n 0 or 3 mod 4; n/4 =
=
handles are added for n 3 mod 4. Again, the same construction works for
m = = =
1 mod 4, n 0 mod 4. Finally, for m 1 mod 4 and n 1 mod 4, a =
quadrilateral imbedding for K m _ 3• n is obtained first. Three vertices are split,
and (n - 1)/4 handles are added near each split vertex, as before. Since
f(m- 2)(n- 2)/41 = (m- S)(n- 2)/4 + 3(n- 1)/4 form= n 1 mod4, =
this yields the desired imbedding of K m n. All other combinations of con-
gruence classes mod 4 are covered by inte~changing m and n. D

4.5.5. Exercises
1. Draw the current graph whose dual voltage graph is given on the left in
Figure 4.6.
2. Draw the current graph whose dual voltage graph is given in Figure 4.1.
3. Prove that if Kn has a self-dual imbedding, then n 0 or 1 mod4. =
Voltage-Current Duality 213

Figure 4.40. A .2"5-current graph.

4. In Exercise 3.5.2 it was shown how to represent an orientable imbedding

G ~ S by a single permutation p of the directed edge set DE(G). Show
that the dual imbedding G* ~ S is represented by p o t, where t is the
involution of DE(G) given by t(e•) = e-•. Prove that the dual of the
dual imbedding is the primal imbedding, using this viewpoint.
5. Name the edges of the imbedded voltage graph given in Exercise 4.1.1.
Represent this imbedding with a permutation p of the directed edge set,
as in Exercise 4. Then use the dual permutation pot to draw the current
graph dual to the original imbedded voltage graph.
6. Obtain a triangular imbedding of Ks,s,s from the ~5 current graph
shown in Figure 4.40. Generalize this construction to get triangular
imbeddings of Kmn,n,n for all m, n > 0.
7. Obtain a triangular imbedding of Kn n n as a dual derived imbedding by
assigning currents in ~n X ~n to a tri~ngle.
8. Let .91 be a group, and let X= .91- {1_,., }. Prove that if there is a
direction-preserving automorphism of the Cayley color graph C( .91, X)
taking each edge labeled x to the edge labeled x- 1, then .91 is abelian.
Then use the trick of Frucht's theorem, Section 2.2, to show that if .91 is
nonabelian, then there is an ..lll-voltage graph (G, P) such that the
derived graphs Ga and GP are not isomorphic, where p is the voltage
assignment /J(e) = a(e)- 1. Conclude that for an orientable, nonabelian
current graph, it is possible to generate nonisomorphic derived graphs by
choosing different orientations of the current-graph surface.
9. Prove that if the group .91 is abelian, then the two possible derived
graphs of an orientable .91 current graph (G ~ S, a) corresponding to
the two orientations of S are isomorphic (see Exercise 8).
10. =
Generalize Example 4.5.2 to arbitrary n, n 0 mod 3, n > 3.
11. =
A self-dual imbedding of Kn, n Omod4, can be constructed as follows
(Pengelley, 1975). Suppose n = 2't, r > 1. Let .91 be the group
~2 X · · · X~2 X~~ having r factors of ~2 • Let X be the set of
elements of .91 having order greater than 2 and Y the set of elements of
214 Imbedded Voltage Graphs and Current Graphs

order 2. Assign half of the elements of X as voltages to a #X-sided

polygon, as in Example 4.5.4 (check first that #X= Omod4). Then
attach a bouquet of #Y circles at one vertex of the polygon and assign
the elements of Y as voltages (do not nest the circles). Prove that the
derived imbedding, after identifying the sides of digons, is a self-dual
imbedding of K n.
=
12. Let p be a prime, p 1 mod 4. The multiplicative group of non-
zero elements of f!EP is cyclic generated by some element q. Let G ~ S
be the imbedding of a bouquet of ( p - 1)/2 circles in the surface of
genus (p - 1)/4 such that the rotation at the only vertex is
e 1 e 2 e 3 ••• e 1 - e 2 - e3 - •••• Let a be the voltage assignment given by
a(e;) = q;- q;- 1• Prove that the derived imbedding is a self-dual imbed-
ding of the complete graph KP in the surface of genus (p- 1)(p- 4)/4
such that any face is a simple closed walk.
13. Let G be the graph constructed from the union of primal and dual
complete graphs K P in Exercise 12 by adding edges from each dual
vertex to each primal vertex on the face containing the dual vertex. Show
that the graph G is the octahedron graph OF"
14. Suppose that a graph G is drawn on the surface S8 so that there are c
points where edges cross. Prove that 2 - 2g ~ (#V + c) -
(#£ + 2c)j3. Use this inequality and Exercise 13 to conclude that the
minimum number of edge crossings in a drawing of the octahedron graph
OP on the surface S<p-l)(p- 4>;4 is p(p- 1)/2, where p is a prime,
p = 1 mod4 (Gross, 1978).
5
Map Colorings

The chromatic number of a surface S is equal to the maximum of the set of

chromatic numbers of simplicial graphs that can be imbedded in S, as we
recall from Section 1.5. Heawood (1890) showed that there is a finite maxi-
mum, even though there is no limit to the number of vertices of a graph that
can be imbedded in S. In particular, if the surface S has Euler characteristic
c 5: 1, then

chr(S) 5: l
7 + /49 - 24c
2
J

which is now known as the "Heawood inequality". The value of the expression
on the right-hand side is called the "Heawood number" of the surfaceS and is
denoted H(S).
The determination of the chromatic numbers of the surfaces other than the
sphere is called the "Heawood problem". Its solution, mainly by Ringel and
Youngs (1968), gave topological graph theory the critical momentum to
develop into an independent research area. The solution is that, except for the
Klein bottle, which has chromatic number 6, the chromatic number of every
surface equals the Heawood number. For example, the projective plane has
chromatic number 6 and the torus has chromatic number 7, exactly their
Heawood numbers. The idea of the proof is to imbed in each surface a
complete graph whose number of vertices equals the Heawood number of the
surface.
Since the sphere has Euler characteristic c = 2, its Heawood number is 4.
However, Heawood's argument for other surfaces cannot be used to establish
four as an upper bound for the chromatic number of the sphere. Although the
sphere is the least complicated closed surface, and although some of the most
distinguished mathematicians attempted to solve the problem, the chromatic
number of the sphere was the last to be known. This last case was resolved
when Appel and Haken (1976) established that chr(S0 ) = H(S0 ) = 4, which is
called the Four-Color Theorem. Since the proof of the Four-Color Theorem is
fully explained elsewhere, quite lengthy, and not topological in character, we
confine our attention to the Heawood problem.

215
216 Map Colorings

The original solution to the Heawood problem occupies about 300 journal
pages, spread over numerous separate articles, and it requires several different
kinds of current graphs, whose properties are individually derived. Ringel
(1974) has condensed this proof somewhat. Although Ringel considers every
case, it remains useful to refer to the original papers for complete details of
some of the more difficult cases (such as "orientable case 6").
The introduction of voltage graphs and topological current graphs unifies
and simplifies the geometric part of the solution. However, the construction of
appropriate assignments of voltages or currents remains at about the same
level of difficulty as originally. The present review of the Heawood problem
concentrates on a representative sample of the cases whose solutions are most
readily generalizable to other imbedding problems.

5.1. THE HEAWOOD UPPER BOUND

The first step in calculating the chromatic numbers of all the closed surfaces
except the sphere is to derive the Heawood inequality. The second step is to
use Heawood's inequality to reduce the Heawood problem to finding the genus
and the crosscap number of every complete graph. A computation of the genus
=
of each complete graph Kn such that n 7 modulo 12 illustrates the basic
approach to completing the second step.

5.1.1. Average Valence

In order to establish an upper bound for the possible chromatic numbers of
graphs that can be imbedded in a surface S, the main concept needed is
average valence. By using the Euler characteristic, it is possible to show that
the average valence is bounded.

Theorem 5.1.1. LetS be a closed surface of Euler characteristic c, and let G be

a simplicial graph imbedded in S. Then
6c
average valence( G) ~ 6 - # V

Proof Whether or not the imbedding is a 2-cell imbedding, we know that

#V- #E + #F~ c
For a 2-cell imbedding, we have equality. Otherwise, we observe that all the
nonsimply connected regions could be subdivided into cellular regions by
adding edges to the graph G, thereby increasing #E without increasing #F.
From the edge-region inequality 2#£ ~ 3#F, established in Theorem 1.4.2,
we obtain an upper bound
The Heawood Upper Bound 217

for #F, which we substitute into the previous inequality. This yields the
inequality

and its consequence

#E ~ 3#V- 3c

By Theorem 1.1.1, the sum of the valences is equal to 2#E. Thus, the average
valence is 2#Ej#V. Substituting the upper bound 3#V- 3c for #E, we
conclude that

6c
average valence( G) ~ 6- #V 0

5.1.2. Chromatically Critical Graphs

A graph G is called "chromatically critical" if, no matter what edge is
removed, the chromatic number is decreased. Given any graph imbedded in S
whose chromatic number is that of the surface S, one can successively delete
edges until a chromatically critical graph imbedded in S is obtained. Obvi-
ously, every chromatically critical graph is simplicial and connected.

Theorem 5.1.2. LetS be a closed surface, and let G be a chromatically critical

graph such that chr(G) = chr(S). Then for every vertex v of G, chr(S)- 1 ~
valence ( v ).

Proof Suppose that v is a vertex of G with fewer than chr(S) - 1

neighbors. Since G is chromatically critical, its subgraph G - v can be colored
with chr(S) - 1 colors. At most chr(S)- 2 of these are assigned to neighbors
of v. This leaves one of those chr(S) - 1 colors available to color v, thereby
contradicting the fact that chr(G) = chr(S). 0

Example 5.1.1. Since the projective plane N1 has Euler characteristic c = 1, it

follows from Theorem 5.1.1 that any graph G imbedded in N1 has a vertex of
valence five or less. From Theorem 5.1.2 it then follows that chr(N1 ) ~ 6. Figure
5.1 shows an imbedding of the complete graph K 6 in the projective plane. Since
chr(K6 ) = 6, it follows that chr(N1 ) = 6, as first proved by Tietze (1910).

Example 5.1.2. The torus S1 has Euler characteristic c = 0. Thus, by Theorem

5.1.1, the average valence of a simplicial graph G imbeddable in S 1 is less than or
equal to 6. By Theorem 5.1.2, it follows that chr(S1 ) ~ 7. Figure 5.2 shows an
imbedding of the complete graph ·K1 in S 1 • Since chr(K 7 ) = 7, it follows that
chr(S 1 ) = 7, which was first proved by He.ffter (1891).
218 Map Colorings

Figure 5.1. An imbedding of the complete graph K 6 in the

projective plane N1 •
0

2 3 4 5 6 0 2
Figure 5.2. An imbedding K 7 -+ S1 is obtained by pasting the left side to the right side and then
pasting the top to the bottom with a 2/7 twist.

Example 5.1.3. Like the torus, the Klein bottle N2 has Euler characteristic
c = 0. By the same argument as for the torus it is proved that chr(N2 ) 5: 7.
Unlike the torus, however, the Klein bottle has chromatic number 6, one less
than its Heawood number, because-as we established in Theorem 3.4.6-the
crosscap number of K 1 is 3. This anomaly was discovered by Franklin (1934).

Theorem 5.1.3 (Heawood, 1890). Let S be a closed surface with Euler char-
acteristic c 5: 1. Then

chr(S) 5: l 7 + /49 - 24c

2
J

Proof If c = 1, then the expression on the right-hand side of the in-

equality has the value 6, and the surface S is a projective plane, so that
chr(S) = 6, as we saw in Example 5.1.1. Thus the inequality holds.
If c 5: 0, our immediate objective is to prove that the quadratic expression

chr(S) 2 - 7 chr(S) + 6c
is nonpos1ttve. To this end, let G be a graph imbedded in S such that
chr(G) = chr(S), and such that G is chromatically critical. From Theorem
5.1.1, it follows that
6c
average valence( G) 5: 6 - # V
The Heawood Upper Bound 219

From Theorem 5.1.2, it follows that

average valence( G) ~ chr( S) - 1

Combining these two inequalities, we obtain the inequality

6c
chr( S) - 1 < 6 - -
- #V

Since c:;;; 0, we have -6cj#V:;;; -6cjchr(S), because #V ~ chr(S). Thus,

we infer
6c
chr(S) - 1 :;;; 6- chr(S)

from which we immediately produce the inequality

chr(S) 2 - 7chr(S) + 6c:;;; 0

By factoring the quadratic expression on the left-hand side we obtain

( chr(S) -
1 - v49 - 24c ) ( 1 + v49 - 24c )
2 chr(S) - 2 :;;; 0

For c :;;; 0, the value of the expression 7 - v49 - 24c is less than or equal to
0. Thus, the value of the first factor is positive. It follows that the value of the
second factor is nonpositive. Since chr(S) is an integer, the conclusion follows.
D

5.1.3. The Five-Color Theorem

If one substitutes c = 2 into the Heawood inequality, one obtains the result
chr(S0 ):;;; 4. Even though Appel and Haken have subsequently proved this, it
does not follow from Heawood's argument. Indeed, one of the purposes of
Heawood's paper (1890) was to show the error in a purported proof by Kempe
(1879) that chr(S0 ):;;; 4. What Heawood was able to prove about the sphere is
the following theorem.

Theorem 5.1.4 (Heawood, 1890). The chromatic number of the sphere S0 is at

most 5.

Proof Let G be a graph imbedded in S0 such that chr(G) = chr(S0 ) and

that G is chromatically critical. By Theorem 5.1.1, the average valence of G is
less than 6, so that G must have a vertex v of valence less than or equal to 5.
By Theorem 5.1.2, the chromatic number of G is at most 6. Since G is
220 Map Colorings

Figure 5.3. The five neighbors of the vertex v.

chromatically critical, we can color the graph G - u with five colors, say the
colors e1, ••. , es. If any two of the five neighbors of u had the same color, then
some color e; would not be applied to any neighbor of u. Accordingly we
might extend the coloring of G - u to a coloring of G by assigning colore; to
u. Thus, we suppose that the five neighbors u1, ... , Us of u are assigned the
colors e 1, ••• , es, respectively, corresponding to the order in which they are
encountered in a clockwise traversal around u. This is illustrated in Figure 5.3
For any two colors e; and e1 , consider the subgraph of G- u containing
every vertex that is assigned color e; or e1 . Every component of this subgraph
is called a "ere1 Kempe chain".
First, suppose that the vertex u3 is not contained in the ece 3 Kempe chain
that contains the vertex u1 • One might recolor the vertices of this ece 3 Kempe
chain by reversing the roles of e1 and e3 , that is, by recoloring every eccolored
vertex e3 and recoloring every e3-colored vertex e1 . This results in a proper
coloring of the graph G - u such that u1 and u3 are both assigned color e1 ,
while vertices u2 , u4 , and us are still assigned colors e2 , e4 , and es respec-
tively. We can extend this revised coloring of G - u to a coloring of G by
assigning color e3 to u.
Alternatively, if the vertex u3 is in the same ece3 Kempe chain as u1 , then
u2 and u4 cannot be in the same e2-e 4 Kempe chain, because of the Jordan
curve theorem. Thus, we might revise the coloring of G- u by reversing the
assignment of e 2 and e4 to the vertices in the e2-e 4 Kempe chain that contains
u4 • Then color e4 would be assigned to no neighbor of u, and so it could be
assigned to u, yielding a 5-coloring of G. D

5.1.4. The Complete-Graph Imbedding Problem

The Euler characteristic of the orientable surface Sg is 2 - 2g. Thus,

H( sg) = l7 + /49 - ~4(2 - 2g) j = l7 + .;12+ 48g J

The Euler characteristic of the nonorientable surface Nk is 2 - k. Thus,

H(Nk) = l7+ /49 - 224(2 - k > j = l7 + .;12+ 24k J

The Heawood Upper Bound 221

Table 5.1 Heawood number of S11 , for g = 1, ... , 23

genusg123 4 56 7 8 91011121314151617181920212223
ll(S8 ) 7 8 9 10 11 12 12 13 13 14 15 15 16 16 16 17 17 18 18 19 19 19 20

Table 5.2 Heawood number of Nk, for k = 1, ... , 23

crosscap
numberk 1 2 3 4 56 7 8 910111213 1415 16 171819 20 21 22 23
ll(lVk) 6 7 7 8 9 9 10 10 10 11 11 12 12 12 13 13 13 13 14 14 14 15 15

Tables 5.1 and 5.2 show the calculated values of the Heawood number for
some orientable surfaces and for some nonorientable surfaces, respectively.
From the formula for the Heawood number, one sees that it must increase
as the genus or crosscap number increases, but at a rate proportional to the
square root of the genus or crosscap number. Tables 5.1 and 5.2 confirm this.
Thus, every sufficiently large integer is the chromatic number of at least one
orientable surface and of at least one nonorientable surface. Moreover, the
same chromatic number may serve for several different surfaces of consecutive
genera or of consecutive crosscap numbers, and the number of consecutive
surfaces that share the same chromatic number tends to increase along with
the chromatic number.
One reads from Table 5.1 that H(S20 ) = H(S21 ) = H(S22 ) = 19. Now
suppose that there were a graph G imbeddable in S20 such that chr(G) = 19.
It would follow that chr(S20 ) ~ 19. Moreover, since any such graph that is
imbeddable in s20 is also imbeddable in s21 and s22' it would also follow that
chr(S21 ) ~ 19 and chr(S22 ) ~ 19. By Theorem 5.1.3, the chromatic number of
each of these three surfaces is also at most 19, their common Heawood
number. Thus, they would all have chromatic number exactly 19. In fact, this
is true, because the complete graph K 19 is imbeddable in S20 , which is proved
in Example 4.4.1. In general, complete graphs are used in this manner to show
that the chromatic number of a surface is equal to its Heawood number. To
formalize the process, we define the numbers

I(n) = r
(n- 3)(n- 4)
12
l and
-
l(n) =
r(n-3)(n-4)l
6

Theorem 5.1.5. Assume that n ~ 7 and that y(Kn) ~ I(n). Then for any
orientable surface Sg such that I(n) ~ g < I(n + 1), the equation chr(Sg) =
H(Sg) = n holds.
222 Map Colorings

Proof Substitution of /(n) and l(n + 1) into the formula for the Heawood
number establishes by direct calculation that H(S1<n>) = n, that H(S1(n+l)) =
n + 1, and that for a surface Sg such that /(n) 5: g < l(n + 1), H(Sg) = n.
By Theorem 5.1.3, chr(Sg) 5: H(Sg) = n. Under the stated assumption, on the
other hand, the complete graph Kn is imbeddable in every such surface Sg, so
that chr(Sg) :;:: n = H(Sg). 0

Theorem 5.1.6. Assume that n > 8 and that y(Kn) 5: J(n). Then for
any nonorientable surface Nk such that I(n) 5: k < I(n + 1), the equation
chr(Nk) = H(Nk) = n holds.

The proof of Theorem 5.1.6 is simplistically analogous to the proof of

Theorem 5.1.5, and so it is omitted. These two theorems together reduce the
calculation of chromatic numbers to proving that the complete graph K n can
be imbedded in the surfaces S1<n> and Nl<n>· In fact, no better imbeddings
exist for Kn, as it is now proved.

Theorem 5.1.7 (The Complete-Graph Orientable-Imbedding Inequality).

y(Kn) :;:: r(n - 3)( n - 4)

12
l
Proof We begin with the Euler equation for the closed orientable surface
of genus y(Kn), that is,
#V- #E + #F= 2- 2y(Kn)
From the edge-region inequality 2#£:;:: 3#F, we infer that #F 5: f#E, and
substitute for #F into the equation above to obtain

from which we routinely calculate that

n2 - 1n + 12 (n- 3)(n- 4)
12
Since y(Kn) is an integer, the conclusion follows. 0

Theorem 5.1.8 (The Complete-Graph Nonorientable-Imbedding Inequality).

-y(Kn) :;:: r(n- 3)(n-

6
4)1
Proof Proceeding exactly as in the proof of Theorem 5.1.7, we obtain the
inequality
1 n(n- 1)
n- - > 2- y-(K )
3 2 - n
1be Heawood Upper Bound 223

from which we immediately derive

The conclusion follows, since y(Kn) is an integer. D

5.1.5. Triangulations of Surfaces by Complete Graphs

The main part of the solution to the Heawood problem is now before us: to
construct minimum imbeddings of the complete graphs in both orientable and
nonorientable surfaces. The general form of these imbeddings seems to depend
strongly on the residue class of n modulo 12.
=
First, if n 0, 3, 4, or 7 modulo 12, then the value of the expression
( n - 3)( n - 4)/12 is an integer. The value of the expression is not an integer
for other values of n. It follows that

I(n) -
- r (n- 3)(n- 4)
12
1-
-
(n- 3)(n- 4)
12
for n =0, 3, 4, or 7 modulo 12
A review of the proof of Theorem 5.1.7 reveals that this equality is obtained if
and only if #F = t#E, which happens, of course, if and only if the graph Kn
triangulates the surface of genus I ( n ).
In general, the quantity /(n)- (n- 3)(n- 4)/12 measures by how much
an orientable imbedding Kn ~ S1<n> would fail to be a triangulation, in the
following sense. For each positive integer i, let F; denote the set of i-sided
faces of the imbedding. Then

l(n)- (n- 3)(n- 4)/12 = ~( L #F;- 3#F) = ~(2#£- 3#F)

6 i 6

For instance, if n = 1, 6, 9, or 10 modulo 12, then /(n) - (n - 3)(n - 4)/12

= 3/6. The number of extra edges one would need to add to such an
imbedding to obtain a triangulation is 3. Table 5.3 gives this quantity for every
residue class.
=
Similarly if n 0, 1, 3, 4, 6, 7, 9, or 10 modulo 12, then (n - 3)(n - 4)/6 is
an integer. It is not an integer for other values of n. Table 5.4 shows the values

Table 5.3 The values of I(n) - (n- 3)(n- 4) /12

n modulo12 0 1 2 3 4 5 6 7 8 9 10 11
.J.
~ .J. l .J. .J. l
_s_
l(n)- (n- 3)(n- 4)/12 0 6 0 0 6 6 0 6 6 6 6
224 Map Colorings

Table 5.4 The values of i(n)- (n- 3)(n- 4) j 6

n modulo12 0 1 2 3 4 5 6 7 8 9 10 11
l(n)- (n- 3)(n- 4)/6 0 0 t 0 0 2.
3 0 0 1 0 0 2.
3

of i(n)- (n- 3)(n- 4)/6, which indicate by how much a nonorientable

imbedding Kn ~ Ni<n> would fail to be a triangulation.
Tables 5.3 and 5.4 suggest immediately that there might be at least four
cases to consider for orientable imbeddings and at least two cases to consider
for nonorientable imbeddings of K n. Only upon deeper investigation does one
discover the properties that lead to a natural split into 12 orientable cases and
12 nonorientable cases. These are commonly known as orientable cases
0, ... , 11 and nonorientable cases 0, ... , 11, corresponding to the residue of
n modulo 12. Orientable cases 0, 3, 4, and 7 and nonorientable cases 0, 1, 3, 4,
6, 7, 9, and 10 are called "regular cases", because the imbedded complete
graph actually triangulates the surface in those cases. All other cases are called
"irregular cases".

5.1.6. Exercises
1. Let G be a 16-vertex graph with four vertices each of valences 4, 5, 6, and
7. Prove that the graph G is not planar.
2. Construct a 5-regular planar graph with as few vertices as possible.
3. Suppose that G is a connected chromatically critical graph such that
chr( G) = 3. Is G necessarily an odd cycle?
4. Construct a chromatically critical graph G such that y(G) = 1 and
chr(G) = 5.
5. Prove that the removal of an edge from K 7 yields a graph of crosscap
number 2.
6. Construct an imbedding of K 6 in a Mobius band.
7. Let G be the edge-complement in K 8 of a 1-factor. What is the genus
of G?
8. What is the chromatic number of the graph G of Exercise 7?
9. Suppose that y(G) = 1 and that chr(G) = 7. Prove that G contains a
sub graph isomorphic to K 7 •

5.2. QUOTIENTS OF COMPLETE-GRAPH IMBEDDINGS AND

SOME VARIATIONS

It would be very convenient if minimum orientable and nonorientable imbed-

dings of all the complete graphs could be immediately derived from imbedded
voltage graphs. That would have greatly simplified the solution to the Heawood
Quotients of Complete-Graph Imbeddings and Some Variations 225

problem. However, only orientable case 7 admits such a direct construction,

and even in that case, it requires considerable ingenuity to assign the voltages
correctly to the imbedded base graph. Each of the other three regular orient-
able cases requires a few additional tactics beyond the basic strategy.

5.2.1. A Base Imbedding for Orientable Case 7

Suppose that n = 12s + 7. Then the complete graph Kn has 12s + 7 vertices
and (12s + 7)(6s ·+ 3) edges. Our objective is to construct an imbedding with
(12s + 7)(4s + 2) faces, every one a triangle. The greatest common divisor of
these three numbers is 12s + 7, so the obvious approach is to consider a base
graph imbedding such that

#V = 1 #E = 6s + 3 and #F = 4s + 2

that is, an imbedding of the bouquet B6.+ 3 in the surface Ss+l·

There is little trouble in constructing candidates for a base imbedding, if
one generalizes a construction for surfaces already considered in Figure 4.31.
In that figure, opposite sides of an octagon are pasted to each other to produce
the surface S2 • In general, the opposite sides of a 4g-sided polygon can be
pasted together to yield the surface Sg. To prove that this works, one considers
the corners of the polygon to be vertices of a graph and the sides of the
polygon to be edges. The result of the pasting leaves 2g edges and one face,
obviously. It is not difficult to verify that it also leaves only one vertex. (See
Exercises 1-3.) Thus the result is the bouquet B 2 g imbedded in the orientable
surface of Euler characteristic 1 - 2g + 1, that is, the surface Sg.
To obtain an imbedding of the bouquet B6 .+ 3 in the surface s.+ 1 , one
begins with a (4s + 4)-sided polygon. When each side is pasted to the opposite
side, the result is an imbedding B2 • + 2 ~ s. + 1 • If one now draws a line from
one corner of the (4s + 4)-sided face to each of the other 4s + 1 corners to
which it is not adjacent, one obtains an imbedding B6.+ 3 ~ Ss+l· The
scallop-shell appearance of this imbedding is illustrated in Figure 5.4.
Ordinarily one tries to choose the voltage group to be cyclic, since cyclic
groups are the least complicated. In this case, since the covering space to be
constructed is (12s + 7)-sheeted, the obvious candidate for a voltage group is
fr12 s+?· Moreover, since numbers of the form 12s + 7 can be prime, the only
hope to lift some of the scallop-shell base imbeddings to triangular imbeddings
of complete graphs is with cyclic voltage groups.
Any one-to-one assignment of the voltages 1, ... , 6s + 3 modulo 12s + 7 to
the edges of the bouquet B6 s+J yields K 12.+ 7 as the derived graph. However,
they must be assigned so that the Kirchhoff voltage law holds globally, else the
derived imbedding would not be a triangulation. There is no known general
reason to suppose that a particular problem of this kind can be solved at all.
226 Map Colorings

Figure 5.4. Imbeddings 8 9 --+ ~ and 8 15 --+ S3 . These are base graphs for imbeddings K19 --+ S20
and K31 --+ S64 .

Alternatively, one may observe that the global Kirchhoff voltage law corre-
sponds to a system of 4s + 2 linear equations in 6s + 3 unknown edge-voltages,
which would seem to have many solutions. However, the requirement that
each of the voltages 1, ... , 6s + 3 be used exactly once corresponds to an
inequality ·

n(x;- xi)* 0
I, j

of degree ( 65 ; 3 ). Accordingly, one does not expect to find a solution by the

routine application of general methods.

5.2.2. Using a Coil to Assign Voltages

To obtain a satisfactory voltage assignment for the scallop shell base imbed-
ding, the voltages 1, ... , 6s + 3 modulo 12s + 7 are first partitioned into three
intervals. Voltages 1, ... , 2s in the lower range are distributed along the rim of
the shell in consecutive clockwise order, with plus edge directions assigned so
they alternate. The voltage 2s + 1 is assigned to the middle interior edge.
Voltages 2s + 2, ... , 4s + 2 in the middle range are assigned to the right half
of the shell, and voltages 4s + 3, ... , 6s + 3 are assigned to the left half.
Figure 5.5 shows the way that they are assigned.
The relationship between the voltages 1, ... , 2s in the lower range and the
voltages 2s + 2, ... , 4s + 2 in the middle range is called a "coil". Note first
that the voltage 3s + 2 in the precise middle of the middle range is assigned to
the top edge of the scallop shell. The first outer voltage to follow 3s + 2 is the
voltage 1, and the plus direction of its edge is opposite to that of the top edge.
Thus, the top triangle will satisfy KVL if and only if 3s + 1 is assigned to its
Quotients of Complete-Graph Imbeddings and Some Variations 227

3s+2
Figure 5.5. This assignment of voltages in .2"12,+ 7 to the imbedded bouquet B6,+ 3 -+ Ss+l
satisfies KVL and yields as a derived graph a triangular imbedding K12 ,+ 7 -+ S1<12 s+?)·

other edge. Since the voltage 2 is assigned to the outer edge of the second
triangle from the top, in the same direction as the voltage 3s + 1, that triangle
will satisfy KVL if its other edge is assigned the voltage 3s + 3. In this manner
the voltages 1, ... , 2s are used to coil from the middle voltage 3s + 2 in the
middle range outward to the limiting voltages 2s + 2 and 4s + 2. Figure 5.6
shows a combinatorial abstraction of the coil.

2s

4s +2

2s-1
Figure 5.6. The voltages 1, ... ,2s are used to coil outward from 3s + 2 toward 2s + 1 and
4s + 2.
228 Map Colorings

5 voltages
in~l9

3LSJ3 2
voltages
in~7
2 2

(a)
5

(b)

voltages
in;2"31

3 3

8
(c)
Figure 5.7. Imbedded voltage graphs that yield triangular imbeddings of (a) K7 , (b) K19 , and
(c) K3,.

Because of the way the scallop shell is to be pasted together to form a

closed surface, the lower voltages 1, ... , 2s also appear on the left rim. There
they coil among the upper voltages 4s + 3, ... , 6s + 3, starting from Ss + 3
on the vertical interior edge and working outward. The result is that KVL also
holds on the triangles on the left side of the shell.
The voltage 2s + 1 on the middle interior edge welds the two coils together,
so that KVL is also satisfied on the two triangular faces in which it lies. Thus,
KVL holds globally, and the Heawood problem is solved for orientable case 7.
Figure 5.7 shows the voltage assignments for imbeddings B 3 ~ S 1 , B9 ~ S2 ,
and B 15 ~ S3 that have derived triangular imbeddings K 7 ~ S 1 , K 19 ~ S20 ,
and K 31 ~ S63 , respectively.
Quotients of Complete-Graph Imbeddings and Some Variations 229

• clockwise
11 14 12 o counterclockwise

4 3 2

6 9 7

5
Figure 5.8. A Gustin current graph for K31 .

5.2.3. A Current-Graph Perspective on Case 7

If one now draws the dual of the imbedded voltage graph in Figure 5.7b, one
obtains precisely the topological current graph already illustrated in Figure
4.31. The underlying graph is planar and 3-regular, so that it is also possible to
draw an equivalent Gustin current graph in the plane, with no edge-crossings.
Such a Gustin current graph appears in Figure 4.18. Figure 5.8 shows the
Gustin current graph that is equivalent to the imbedded voltage graph in
Figure 5.7c.
The coiling relationship among the currents is highly visible in this ladder-
like Gustin current graph. It is also easy to see that KCL is satisfied at every
vertex. This ease of visibility makes Gustin current graphs very useful in
computations. On the other hand, while the base graph imbedding for the
voltage graph solution is designed to have one vertex, it requires a certain
amount of work to check the corresponding face in the Gustin current graph.

• clockwise

11 14 o counterclockwise

6 9

5
Figure 5.9. Pictorial representation of the Face Tracing Algorithm.
230 Map Colorings

• clockwise
o counterclockwise
4s+3 6s+2 5s+5 5s+1 5s+4 5s+2

2s 2s-1 4 3 2

. . . ---¢----c>-----o--r--4
2s+2 4s+ 1 3s+4 3s 3s+3 3s+1

2s + 1
Figure 5.10. A general current graph for orientable case 7 of the Heawood problem.

Fortunately, there is a pictorial way to apply the Face Tracing Algorithm,

illustrated in Figure 5.9.
One may start at any edge. In this example, we have started with the
horizontal edge at the upper left, which carries the current 11. Since the
terminal vertex of that edge is solid, the next edge we traverse is the one
carrying the current 14. When this single edge orbit closes on itself, every edge
has been traversed in both directions. Thus, the corresponding imbedding has
only one face. The thin line that runs along the edges in Figure 5.9 traces out
the boundary circuit of that face.
The general voltage graph for orientable case 7, shown in Figure 5.5, also
has a dual that may be represented as a ladderlike Gustin current graph. This
is shown in Figure 5.10. Ringel (1961) solved this case with purely combina-
torial methods. The generating row he devised anticipates the technique of
coiling.

5.2.4. Orientable Case 4: Doubling 1-Factors

If n = 12s + 4, then the complete graph Kn has 12s + 4 vertices and
(6s + 2)(8s + 2) triangular faces. Since the greatest common divisor of these
three numbers is 6s + 2, the obvious properties one might want in an
imbedded voltage graph are 2 vertices, 12s + 3 edges, and 8s + 2 triangular
faces. Unfortunately, the complete graph K12.+4 cannot have such a quotient.
Theorem 5.2.1 (Gross and Tucker, 1974). Let p: Kn ~ G be a covering
projection. Then the cardinality of the fibers is an odd number.
Proof Let v be a vertex of the base graph G. Suppose that there are 2n
vertices in the fiber over v. Then the subgraph of G spanned by these 2n
vertices is isomorphic to K 2 n and contains n(2n - 1) edges. These n(2n - 1)
edges must be the union of the fibers of v-based loops. However, every edge
fiber contains 2n edges, and 2n does not divide n(2n - 1), a contradiction.
D
Quotients of Complete-Graph Imbeddings and Some Variations 231

It would not make sense to depend on the largest odd factor of 12s + 4,
since 12s + 4 might be a power of 2, for example if s = 5 or 85. However,
there is a tactic that does help, namely, to double some 1-factors of K 12 s+4·
"Doubling" a set of edges means to insert an additional adjacency along each
edge in the set, that is, so that it has the same two endpoints.
If one 1-factor of K 12 .+ 4 were doubled, then the resulting graph would
have 12s + 4 vertices and (6s + 2)(12s + 3) + (6s + 2) edges, that is,
(6s + 2)(12s + 4) edges. This looks promising, since assigning the voltages
1, ... , 6s + 2 modulo (12s + 4) to the edges of the bouquet B6.+ 2 yields as a
derived graph K 12.+ 4 with a doubled 1-factor. The edge with voltage 6s + 2
leads to the double edges in the derived graph, because it is a loop with a
voltage of order 2.
One does not want a triangular imbedding of K 12 s+4 with a doubled
1-factor. Instead, the idea is to derive an imbedding so that each pair of
doubled edges bounds a digon, but that all other faces are three-sided. Then
each digon may be excised and its two sides sewn together, that is, identified.
The result would be a triangular imbedding K 12.+4 ~ S1<12.+ 4 >. The number
of digons in the derived imbedding would be 6s + 2, and the number of
triangular faces would be (6s + 2)(8s + 2).
According to Theorem 4.1.1, such a derived imbedding could be obtained
from an imbedded voltage graph with one vertex, 6s + 2 edges, and 4s + 2
faces. Of these faces, 4s + 1 should be 3-sided and satisfy KVL. The other
face should be a monogon whose net voltage has order two. An easy calcula-
tion reveals that the base surface would have Euler characteristic 1 - 2s, an
odd number, so the base surface would be nonorientable. Figure 5.11 shows
such an imbedded voltage graph for K 4 , when s = 0.
Some experimentation reveals that it is not easy to generalize Figure 5.11.
One of the problems that would arise is the necessity of proving that the
derived surface is orientable. Of course, this is obvious for Figure 5.11,
because the derived surface has Euler characteristic 2. However, every non-
positive even Euler characteristic is shared by an orientable surface and a
nonorientable surface.
These problems are completely circumvented if one doubles three 1-factors
of K 12 s+4• thereby obtaining a total of (6s + 3)(12s + 4) edges. This time the
derived imbedding has 18s + 6 digons, to be excised, and (6s + 2)(8s + 2)
triangular faces. A plausible base imbedding for this configuration might have
one vertex, 6s + 3 edges, and 4s + 4 faces. Three of these faces could be

Figure 5.11. The bouquet 8 2 imbedded in the proJective plane N1 •

Voltages are in the cyclic group .2"4 •
232 Map Colorings

(3, 0)

(3, 1) (3, 1)

(3, 0)
Figure 5.12. Voltages in .2"8 X .2"2 for an imbedding 8 9 ..... S1 . The derived graph is K16 with
three doubled 1-factors, and the imbedding surface is S13 .

monogons and the other 4s + 1 could be three-sided. The Euler characteristic

of the base surface would then be 2 - 2s, so that it might be orientable.
However, this raises a new problem. The three monogons all need voltages
of order 2 if they are to lift to digons, and these voltages must differ from each
other if the derived graph is to be the one that is wanted. Therefore, the cyclic
group .2"12 s+ 4 cannot be the voltage group for this kind of base imbedding.
The obvious voltage group to try is .2"6 s+ 2 X .2"2 , because it has the right
order and it has three elements of order 2. For instance, one would try to
obtain an imbedding of K 16 ~ S 13 by an assignment of voltages in .2"8 X .2"2
to the imbedded bouquet B 9 ~ S 1 • Figure 5.12 shows an appropriate voltage
assignment.
The three loops that are monogon boundaries get the voltages (4, 0), (4, 1),
and (0, 1) that have order 2. Thus, the derived imbedding has 24 digons, each
of whose boundaries is a double adjacency. Each of these digons is excised and
its two boundary edges are identified to eliminate a multiple adjacency. Since
KVL holds on the boundary circuit of every base face except the monogons,
and since these other faces are three-sided, all the faces in the derived
imbedding except the digons are three-sided. Thus, excising the digons and
closing up yields an imbedding K 16 ~ S 13 •
The configuration illustrated in Figure 5.12 can be generalized. Figure 5.13
shows an imbedded bouquet B21 ~ S3 with voltages in .2"20 X .2"2 so that the
derived imbedding is of K 40 with three doubled 1-factors into S 148 . Excising
the digons and closing up would yield a triangular imbedding K 40 ~ S 148 •
Once again, coiling is used to assign voltages so that KVL holds everywhere
but on the three monogon boundaries. This time however, the voltages (3, 0),
(3, 1), (6, 0), (6, 1), (9, 0) and (9, 1) are used on the rim of the polygon. The
Quotients of Complete-Graph Imbeddings and Some Variations 233

(6, 1)

(10, 1)

(6,1)
Figure 5.13. An imbedding bouquet 8 21 -+ S3 with voltages in .2'20 X .2'2.

voltages (3, 0), (6, 0) and (9, 0) coil among the voltages whose first coordinate is
congruent to 1 modulo 3, while the voltages (3, 1) and (6, 1) coil among the
voltages whose first coordinate is congruent of 2 modulo 3. The voltage (9, 1) is
used to patch the coils together. The current graph in Figure 5.14 shows the
coiling and patching clearly. The general solution to orientable case 4 was first
achieved by Terry et al. (1970), and is left as an exercise. (See Exercise 11.)

5.2.5. About Orientable Cases 3 and 0

The two other regular orientable cases are much more difficult to solve than
cases 7 and 4. What makes case 3 difficult is that the voltage graph for the
solution has three vertices, rather than one. Nonabelian voltages are the
complicating factor in case 0. Ringel (1974) gives a complete explanation of
their solutions.

(8, 1) (5,0) (4, 1) (7, 1) (1, 1) (10, 1)

(0, 1)
(7 (3, 1) (3, 0) (9, 0)

(~
(6, 1) (9, 1) (6, 0)

(8,0) (5, 1) (4,0) (7,0) (1, 0) (10, 0)

Figure 5.14. A current graph corresponding to the voltage graph in Figure 5.13.
234 Map Colorings

6s

Figure 5.15. This base graph has 2s loops at each of its three vertices and 4s + 1 edges between
each pair of vertices. The derived graph consists of three isomorphic copies of K 12 , + 3 •

A triangular imbedding of K 12 s+J must have 12s + 3 vertices,

(6s + 1)(12s + 3) edges, and (6s + 1)(8s + 2) faces. In general, the largest
common factor of these three numbers is 4s + 1. Thus the smallest quotient
graph must have at least three vertices and 18s + 3 edges. Figure 5.15 shows
an unimbedded voltage graph that has three isomorphic copies of K 12 s+J as
its derived graph. The solution to orientable case 3 is achieved by imbedding
this voltage graph in the surface S3s so that every face is three-sided and that
KVL holds globally. The first solution to orientable case 3 was by Ringel
(1961).
Case 3 perfectly illustrates the mathematical principle that a primal object
and its dual can both be useful, without one entirely supplanting the other.
Since the derived graph imbedding directly covers an imbedded voltage graph,
theoretical investigations tend to concern themselves with voltage graphs,
rather than with current graphs. On the other hand, for small-face imbeddings
of high-valence graphs, it is often easier to assign appropriate currents than to
assign voltages-when the work is by hand, but not necessarily for computer
algorithms.
In particular, imagine assigning appropriate rotations to the voltage graph
of Figure 5.15 so that the derived imbedding would be a genus imbedding for
K ns + 3 • It might be a tedious task to traverse face boundaries in order to
verify KVL. On the other hand, Figure 5.16 contains a Gustin current graph
that generalizes into a solution for all of orientable case 3. As soon as the
combinatorial patterns of current assignments on various classes of edges are
identified, it becomes clear that KCL holds. Note the coiled currents on the
horizontal edges and on the globular rungs. An induction on the ladder length
Quotients of Complete-Graph Imbeddings and Some Variations 235

1 13

13 3 16

13

l<'igure 5.16. A three-orbit .2"27 -current graph that yields a triangular imbedding K 27 -+ S46 .

is used to show that the current graph has three edge orbits, each of which
contains each nonzero current exactly once.
Orientable case 0 was first solved by Terry et al. (1967). The complication
here is that if the number 12s has no large odd factors, then by Theorem 5.2.1,
the complete graph K12 , has no small quotients. The solution here involves
the doubling of many 1-factors and the use of nonabelian voltages or currents.
In particular, if 2k is the largest power of two that divides 12s, then 2k - 1
1-factors are doubled.

5.2.6. Exercises
1. Number the vertices of an eight-sided polygon in consecutive cyclic order
from 0 to 7. For each pair of vertices that is identified in pasting opposite
edges together, write a transposition. Note that fori = 1, ... , 7, there is a
transposition (i, i + 3 modulo 8). How many vertex orbits are there in the
permutation group generated by these transpositions?
2. Suppose that opposite sides of a 4n-sided polygon are pasted together,
and let the image of the polygon rim be regarded as a graph imbedded in
the resulting surface. Prove that the graph has only one vertex. (Hint: See
Exercise 1.)
3. Suppose that opposite sides of a 4n + 2 sided polygon are pasted
together as in Exercise 2. How many vertices does the graph have?
4. Try to complete the partial voltage assignment shown so that the derived
imbedding is K 19 ~ S20 • Paste opposite sides together.

voltages
in.2"19
236 Map Colorings

voltages
9 1
in~I9

5. Complete the partial voltage assignment shown so that the derived

imbedding is K 19 ~ S20 • The voltages shown imply the pasting rules.
6. The coil in Figure 5.6 might have been 3s + 2, 3s + 3, 3s + 1, ... ,
4s + 2, 2s + 2. Show how to use this in another solution to orientable
case 4.
7. Draw the Gustin current graph for the imbedding K 43 ~ S 130 .
8. Trace the face boundary in the current graph of Exercise 7.
9. Prove that if d is an odd divisor of n, then the complete graph Kn has a
unique d-fold quotient (Gross and Tucker, 1974).
10. Classify the quotients of the complete bipartite graph Km, n (Sit, 1976).
11. Draw a general current graph solution for orientable case 4, modeled
after Figure 5.14.

5.3. THE REGULAR NONORIENTABLE CASES

Perhaps surprisingly, the nonorientable cases of the Heawood problem seem to

admit easier solutions than the orientable cases. Moreover, Ringel (1959)
solved them all before the solution was known to any orientable case except
case 5.

5.3.1. Some Additional Tactics

To solve the nonorientable cases of the Heawood problem, several interesting
tactics are employed beyond the fundamental branched-covering-space
strategy. Most of them are illustrated in Figure 5.17, which is the key to
imbedding the complete graph K 15 in the nonorientable surface Nl(IS).
Two of the regions in Figure 5.17 are special digons, marked by a cross in a
circle. The interpretation of such a region is that the digon is to be excised and
its two boundary edges matched together so that the plus directions coincide.
This operation adds a crosscap to the surface. It is just like making the usual
boundary-matching operation that converts a disk into a projective plane, but
now viewed "inside-out". Thus the length of the boundary walk for the region
The Regular Nonorientable Cases 237

voltages
3 ln.2'i4

Figure 5.17. A voltage graph in N2 that ultimately yields a triangular imbedding K 15 -+ N22 •

on the other side of the digon boundary remains the same as before the
matching.
The two crosscaps in Figure 5.17 convert the apparently spherical imbed-
ding surface in which the figure is drawn into the Klein bottle N 2 • The edge
identifications convert the graph into the bouquet B 1 . Four of the regions are
three-sided and satisfy KVL in the group ~14 • In particular, 4 + 5 + 5 = 14
and 2 + 6 + 6 = 14. There are also two monogons, one with boundary
voltage 7, the other with boundary voltage 1.
The derived graph is K 14 with a doubled 1-factor, and the derived imbed-
ding has 56 = 14 · 4 triangular faces, seven digons, and one 14-sided face. The
Euler characteristic of the derived surface is 14 - 98 + 64 = -20. Since the
even numbers 0, 2, 4, 6, 8, 10,12 form the only subgroup of index 2 in ~14 , and
since the base edge carrying voltage 6 is orientation-reversing, the derived
surface is nonorientable, by Theorem 4.1.4.
Excising the seven digQns and closing up the holes converts the derived
imbedding into an imbedding K 14 ~ N22 such that every face is triangular,
except for one 14-sided face whose boundary circuit is a Hamiltonian cycle.
Insert a 15th vertex into the interior of that face and adjoin it to each of the
other 14 vertices. The result is a triangular imbedding K 15 ~ N22 •

5.3.2. Nonorientable Current Graphs

In Section 4.4 we observed that assigning currents to a nonorientably imbed-
ded graph does not automatically yield a well-defined derived imbedding.
However, Youngs (1968a, b) established some combinatorial conventions that
make it well-defined for certain restricted types. Youngs used these to give a
current-graph explanation of Ringel's solution (1959) to the nonorientable
cases. Figure 5.18 shows the Youngs nonorientable current graph that corre-
sponds to the voltage graph in Figure 5.17.
The vertices of valence 2 in a Youngs nonorientable current graph are not
topologically dual to a digon in the voltage graph. One of the edges with such
an endpoint must have valence one at its other endpoint. That edge crosses the
238 Map Colorings

5 6

4 2

7 1
Figure 5.18. The Youngs nonorientable current graph to produce the imbedding K 15 --+ N22 •

"warp" in the surface indicated bY. an encircled cross in a drawing of a

nonorientable voltage graph imbedding. The "warp" is where a digon was
excised and the hole closed up with a crosscap matching.
Another convention used by Gustin, Ringel, and Youngs occurs in both
orientable and nonorientable current graphs. If an edge has valence 1 at one
endpoint and current of order 2, then the dot representing that endpoint is
omitted. Under the topological interpretation of this combinatorial notation,
the absent dot reminds the reader to excise all the digons that lie over the
monogon dual to that endpoint, and to close up the holes in the surface by
matching the two sides together.

5.3.3. Nonorientable Cases 3 and 7

The example in Figure 5.17 is easily generalized to a system of voltage graphs
that solves nonorientable case 3. The triangular imbedding K 12 s+J ~ Nl(l 2 s+J)
is always obtained by first imbedding K 12 .+ 2 into Nl(l2s+J) so that the
boundary circuit of one face is a hamiltonian cycle and so that the other faces
are all three-sided. The imbedding K 12 .+ 2 ~ Nl(l2s+J) is obtained by excising
digons from an imbedding of K 12.+ 2 with a doubled 1-factor in Nl(l 2 s+J)·
The imbedding of K 12 .+ 2 with a doubled 1-factor is derived from a voltage
assignment in .2"12 .+ 2 to the bouquet B6 s+l· A monogon boundary loop
carries the voltage 6s + 1 of order 2, so that the monogon will lift to 6s + 1
digons. Another monogon boundary carries the voltage 1, so that that mono-
goo will lift to a (12s + 2)-sided face whose boundary circuit is hamiltonian.
All of the other 4s faces are three-sided. The base surface is N2 •• Figure 5.19
shows the general Youngs current graph for nonorientable case 3.
The 2s vertices of valence 2 correspond to 2s crosscaps in the base surface.
One observes that the odd currents 1, 3, ... , 4s - 1 are assigned by coiling
outward from 2s + 1, from left to right in the row of horizontal edges.
The details of nonorientable case 7 are remarkably similar to the ones for
nonorientable case 3. Figure 5.20 shows the general Youngs current graph. It
cannot be used on the anomolous graph K 7 , since y(K 7 ) = 3 > 1(7) = 2.
One difference in the solution to case 7 is that the vertical edge at the
extreme left carries the current 4s + 2, which has order 3 in .2"12 • + 6 • Thus, the
monogon dual to its endpoint of valence 1 lifts to 4s + 2 three-sided faces.
The Regular Nonorientable Cases 239

4s+ 1 6s 6s-1 6s-2 4s+3 4s+2

4s 2 4 6 4s-4 4s-2

6s + 1 2s + 1 2s - 1 2s + 3 2s - 3
••• 3 4s-1 1
Figure 5.19. The Youngs nonorientable current graph that yields K12 ,+ 3 ..... Ni(l 2 s+J). The
currents are in -2'"12,+2.

6s+2 6s-1 6s 4s+4 4s+3

4s+2 2 4 6 4s-2 4s

6s+3 2s+1 2s+3 2s-1 2s+5

•••
3 4s+1 1
Figure 5.20. The Youngs current graph for nonorientable case 7. The current group is .2'"12 ,+ 6 .

5.3.4. Nonorientable Case 0

To obtain an imbedding K12 • ~ Nl(l2s)' one first derives an imbedding
K l2s-l ~ N1(l2s) such that one face is (12s - I)-sided and the others are
three-sided. In that sense, nonorientable case 0 resembles nonorientable cases
3 and 7.
For instance, to imbed K 24 into N70 , we start with K 23 , which has 23
vertices and 23 · 11 edges. The kind of imbedding we seek in N70 , has one
23-sided face, whose boundary is a hamiltonian cycle, and 23 · 7 three-sided
faces. Thus, we hope to derive it from a voltage assignment in ~23 on the
bouquet B 11 , imbedded in the nonorientable surface N4 • Figure 5.21 shows
that this is possible.

6 6

3
Figure 5.21. An imbedding 8 11 ..... N4 with voltages in .2'"23 .
240 Map Colorings
2 4 6s-2 7 3s-2 3s+4

::::~·I ~~,,.. :. ..
6s-3 5 6s-1 8 3s-1
6

3s+5
3

Figure 5.22. The Youngs current graph for nonorientable case 0. The current group is .2'12 s _ 1 •

The rectangular sheet in Figure 5.21 is made into a torus in the usual
manner, by matching opposite sides together. Then two crosscaps are inserted,
to obtain N4 • The Kirchhoff voltage law holds on every face except the
monogon.
In general, to derive the imbedding K 12 • ~ Ni(l2s)• the base graph is the
bouquet B6 • _ 1 , and the base surface is N2 ., formed by inserting two crosscaps
into the orientable surface s._ 1 • The base imbedding has one monogon and
4s - 1 triangular faces. The voltages 3, 6, ... , 6s - 6 coil among the voltages
4, 7, 10, ... , 6s - 2 and also among the voltages 5, 8, 11, ... , 6s - 1. The
Youngs current graph is shown in Figure 5.22.

5.3.5. Nonorientable Case 4

Nonorientable case 4 is slightly different from nonorientable cases 0, 3, and 7,
because the base surface has an odd number of crosscaps. Starting with the
complete graph K 12 s+4• one adds a 1-factor to create the possibility
of a quotient graph with one vertex. With the 1-factor added, there are
(6s + 2)(12s + 4) edges, and the derived imbedding has 6s + 2 digons and
(6s + 2)(8s + 2) triangular faces. Taking a (12s + 4)-fold quotient, the base
graph has one vertex, 6s + 2 edges, one monogon, and 4s + 1 triangular faces,
for an Euler characteristic of 1 - 2s. Thus, the base surface is N2.+ 1 • Figure
5.23 shows the voltage graph for the case s = 2.

5.3.6. About Nonorientable Cases 1, 6, 9, and 10

We omit details of the solutions for nonorientable cases 1, 6, 9, and 10. For
nonorientable case 9, Ringel (1974) uses a three-orbit current graph, which
means that, as in orientable case 3, the underlying voltage graph has three
vertices. For nonorientable cases 1, 6, and 10, Ringel (1974) uses inductive
constructions. Cases 6 and 10 are symbiotically intertwined.

5.3. 7. Exercises
1. Using Figure 5.19, draw a voltage graph from which a triangular imbed-
ding K 27 --+ N92 can be obtained.
Additional Adjacencies for Irregular Cases 241

Figure 5.23. An imbedding 8 14 -+ N5 with voltages in .2"28 . The derived graph yields an
imbedding K 28 -+ N1 00 .

2. For the current graph of Figure 5.20, calculate exactly how many faces of
each size occur in the derived graph.

5.4. ADDITIONAL ADJACENCIFS FOR IRREGULAR CASFS

In orientable cases 1, 2, 5, 6, 8, 9, 10, and 11 and in nonorientable cases 2, 5, 8,

and 11, the minimum imbedding cannot be a triangulation, as we recall from
Tables 5.3 and 5.4. The solution strategy is the same for nearly all these cases.
First, a slightly smaller graph than Kn is imbedded in a slightly less com-
plicated surface than S1<n> or Si<n>. Then additional surface complications are
added in order to permit the slightly smaller graph to be built up into K n.

5.4.1. Orientable Case 5

The simplest example of an additional-adjacency solution occurs in case 5. As
a preliminary, we observe that the complete graph K 5 has one too many edges
to be imbedded in the sphere S 0 • That is, the graph K 5 - e already triangu-
lates S0 , and there is no room for an additional edge between the two
nonadjacent vertices. However, by adding a handle that runs from any face
whose boundary contains one of those vertices to any face whose boundary
contains the other, and by then running the missing edge across that handle,
we achieve a genus imbedding K 5 ~ S 1 .
In general, we shall discover that the graph K 12 s + 5 - e has a triangular
imbedding in the surface S1<12 s+ 5>_ 1 • By consideration of fiber sizes and orbit
sizes, we see immediately that the graph K 12 s+ 5 - e has no quotients (see
Exercise 1). It follows that we cannot hope to construct an imbedding

Kl2s+5- e~ Sl(l2s+5)-l

as a branched covering. However, consider what we would obtain if we deleted

242 Map Colorings

Figure 5.24. A three-orbit .2'27 -current graph that ultimately yields a triangular imbedding
Kz9- Kz ..... Ss4·

the two (12s + 3)-valent vertices from such an imbedding-an imbedding of

the complete graph K 12.+ 3 into S 1<12 s+S>- 1 such that two face-boundaries are
Hamiltonian circuits and all other faces are three-sided.
More specifically, such an imbedding

K12s+3 ~ SI(l2s+s)-1

would be for a graph with 12s + 3 vertices and (12s + 3)(6s + 1) edges. There
would be two (12s + 3)-gons and (12s + 3) · 4s triangles. One might hope
that it covers an imbedded voltage graph with one vertex, 6s + 1 edges, 4s
triangles, and two monogons. Unfortunately, no such 1-vertex voltage graph
or corresponding 1-orbit current graph for case 5 is known. The imbedding is
achieved, instead, with a 3-orbit current graph closely related to case 3, as
illustrated for K 20 in Figure 5.24, in which x and y denote the vertices with
excess current.
Once we have the imbedding K l2s + 3 ~ S 1<12 • + S) _ 1 , we install a vertex in
the interior of the two complete (12s + 3)-gons and adjoin it to each of
the 12s + 3 vertices on the region boundary. This yields an imbedding of
K 12 s+s - e into S 1<12 .+ 5>_ 1 • An additional handle is added for the missing
edge, thereby achieving our goal, an imbedding Kns+s ~ S 1<12 s+S>·

5.4.2. Orientable Case 10

To solve orientable case 10 with additional adjacencies, we must go a little
further out on a limb than in case 5, before we can crawl back. This time, as a
preliminary, we notice that K 10 has three too many edges to be imbedded in
S 1 <1o> = S 4 . If we could imbed K 10 - K 3 into S 3 and then restore the three
missing edges with only one handle, we would be done. Restoring all three
edges of a missing 3-cycle with only one handle is not always possible, so
special attention will be required.
The graph K 10 - K 3 has 10 vertices and 42 edges. If it triangulated its
imbedding surface, there would be 28 faces, from which it follows that the
Additional Adjacencies for Irregular Cases 243

0 0 0

2 5 4 3 6

1
Figure 5.25. A .2"Tvoltage graph with base imbedding B3 -+ So and derived imbedding K 7 -+ S3 •
The three complete 7-gons of the derived imbedding are shown at the right.

surface would have Euler characteristic 10 - 42 + 28 = -4. If the imbedding

surface were orientable, it would be S3 • Since the graph K 10 - K 3 has no
quotients (see Exercise 1), we delete the three 7-valent vertices from a
hypothetical imbedding K 10 - K 3 ~ S3 and thereby obtain an imbedding
K 7 ~ S3 with three complete 7-gons and seven 3-gons. (A "complete" face
boundary has one occurrence of every vertex.)
Such an imbedding of K 1 into S3 can be derived from a voltage graph with
one vertex, three edges, one 3-gon, and three 1-gons, as illustrated in Figure
5.25.
New vertices, labeled x, y, and z, are inserted into the interiors of the
7-gons covering the monogons with voltages 1, 2, and 3, respectively, and
adjoined to the old vertices on their 7-gon boundaries. The result is a
triangular imbedding of the graph K 10 - K 3 into the surface S3 • Figure 5.26
illustrates the neighborhood of the vertex 0 in this imbedding.
Our objective is to add the 3-cycle x, y, z with the aid of only one
additional handle. The handle itself is implemented by identifying the boundary
of a hole in the 3-gon 01x to the boundary of a hole in the 3-gon 054. Next,
the edges 04 and 05 are deleted, so that the vicinity of the vertex 0 becomes as
illustrated in Figure 5.27. Figure 5.28 shows how to install the five edges xy,
xz, yz, 04, and 05 so that the result is an imbedding K 10 ~ S4 •
In order to keep track of the four new edges that cross the hole boundary,
we have marked them with the symbols
·--and-

2 3

Figure 5.26. The vicinity of the vertex 0 in the imbedding

5 Kw- K3 ..... S3.
244 Map Colorings
X

2 3

Figure 5.27. An imbedding of (K10 - K 3 ) - {04,05} into

the surface S4 , in the vicinity of the vertex 0.

2 3

Figure 5.28. An imbedding K 10 ..... S4 , in the vicinity of

the vertex 0.
5 4

Whereas this sequence of symbols appears in clockwise order on the lower

hole in Figure 5.28, it appears in counterclockwise order on the upper hole.
This corresponds to an identification of hole boundaries that ensures orien-
tability of the resulting closed surface.
The general solution for orientable case 10 is very much like the solution for
K 10 itself. The strategy is first to imbed K 12 ,+ 1o- K 3 into S1<12 s+lOJ-l
triangularly, and then to use a single new handle to create room to insert the
edges of the missing 3-cycle. The derived imbedding associated with the
current graph in Figure 5.29 is the graph K 12 ,+ 7 in the surface S1<12 s+lO) with
three complete (12s + 7)-gons and all other faces three-sided. We see that this
current graph is closely related to the one used for orientable case 7.

5s+3 5s+4 5s+2 6s+3.

:: I~ I,
)I Jl Jl

• • • 2s+ 1
X

.... • -4
3s+2 3s+3 3s+ 1 4s+2
Figure 5.29. A .2'12 s+ 7-current graph used in the general solution to orientable case 10. The
vertices with excess currents are labeled x, y, and z.
Additional Adjacencies for Irregular Cases 245

5.4.3. About the Other Orientable Cases

Although the original solution to case 9 was quite complicated, Jungerman
(1974) ultimately designed an easier approach, which Ringel (1974) has
reproduced. First, the complete graph K 12 ,+ 8 is imbedded in the surface
S1<12 ,+ 9>_ 1 so that there are two (6s + 4)-gons and all other faces are three-
sided. The boundary circuits of the two (6s + 4)-gons together form a 2-factor.
Next, a handle is installed so that it runs from the interior of one (6s + 4)-gon
to the interior of the other, and a (12s + 9)th vertex is placed on the handle. It
is possible to adjoin the new vertex to every vertex of K 12 , + 8 , which thereby
creates an imbedding K 12 ,+9 ~ S1<12,+ 9l.
Orientable case 1 is solved by combining the methods already used in cases
4 and 10. Using currents in f!E2 X f!E6,_ 1 , an imbedding of K 125 _ 2 into
S 1<l2s+l)-l is obtained, in which there are three (12s- 2)-gons and all other
faces are 3-sided. Edges are added to create an imbedding

and then a handle is added to create room for the missing 3-cycle.
The rotation graph for orientable case 6 is a fascinating variation on the
one for case 3, and the assignment of currents presents interesting new
difficulties. Analogous to case 10, the near-final objective is a triangular
imbedding

Orientable cases 11, 2, and 8 require special ingenuity. In case 11, the
branched covering space is a mostly triangular imbedding of K 12 s+6• which
has five exceptional faces, each a complete (12s + 6)-gon. Figure 5.30 shows
the current graph used when s = 2 and 12s + 11 = 35. When new vertices are
inserted into the (12s + 6)-gons and adjoined to the vertices on the boundary
cycles, the result is a triangular imbedding

Kl2s+ll- Ks ~ SI(l2s+llJ-2

13 11 7 10

:~ >I >I :,f ~

1 z 4 8 5
14[),--1..;;.5_
Figure 5.30. A .2'30-current graph used to construct the imbedding K35
The vertices with excess currents are labelled x, y, and z.
-+ S1(3 5J from case 11.
246 Map Colorings

12

7 7

Figure 5.31. A .2"15 -voltage graph imbedding 8 7 -+ N2 with

derived imbedding K 15 -+ Nl(l?)-l· The two monogons are
covered by complete 15-gons.
12

J sing two handles to create room to add the rmssmg ten edges of K 5 is
ubstantially more difficult than the analogous problem of using one handle in
ase 10 to add K 3 • Cases 2 and 8 are like case 11 in some important ways, but
1ey include complicated additional irregularities.

5.4.4. Nonorientable Case 5

Nonorientable case 5 provides a simple illustration of a nonorientable ad-
ditional adjacency problem. Analogous to orientable case 5, the voltage graph
problem is to derive an imbedding

K12s+3 ~ Nl(l2s+5)-1

with two complete (12s + 3)-gons and all other faces three-sided. Figure 5.31
contains the voltage graph to be used when s = 1 and K 12 s+s = K 11 •
A new vertex is inserted into the interior of each of the two complete
(12s + 3)-gons and adjoined to each vertex of the (12s + 3)-gon boundary.
The result is a triangular imbedding

Figure 5.32 illustrates the vicinity of the vertex 0, where we have assigned x
as the label on the new vertex installed into the 15-gon covering the monogon

Figure 5.32. The vicinity of the vertex 0 in the imbedding

6 K17- K2-+ Nl(l7)-l·
Additional Adjacencies for Irregular Cases 247

6 6
Figure 5.33. Since the vertices x and y have vertex 0 as a common neighbor, there is a way to
adJoin them if only one more crosscap is added.

with voltage 1, and y as the label on the other new vertex, for the special case
of K 11 •
We shall now demonstrate that only one additional crosscap is needed to
create room for the missing edge xy. Figure 5.33 illustrates the appropriate
surgical procedure. First, split the vertex 0 into the two vertices o+ and o-, so
that the vertex x and all the vertices on one side of the path yOx are adjoined
to 0-, whereas the vertex y and all the vertices on the other side of the path
yOx are adjoined too+.
Next, delete the interior of a disk whose boundary passes through both o+
and 0- and parametrize that boundary by the unit circle so that 0 + and 0-
are antipodal points. Also, draw edges from x and y, respectively, to another
pair of antipodal points. Finally, reclose the surface by identifying all pairs of
antipodal points on the parametrized boundary. The antipodal identification
adds a crosscap, it completes an edge between x and y, and it reunites vertices
0 + and 0- into a single vertex 0 adjacent to every other vertex. Thus, the
result is an imbedding K 17 ~ Ni(I?). This technique easily generalizes to all of
nonorientable case 5.

5.4.5. About Nonorientable Cases 11, 8, and 2

Nonorientable case 11 is quite similar to nonorientable case 5. Nonorientable
case 8 is slightly more difficult, in that the derived imbedding contains two
(n/2)-gons and one n-gon, rather than two n-gons. Nonorientable case 2 is
solved inductively, rather like regular nonorientable cases 1, 6, and 10.

5.4.6. Exercises
1. Prove that graphs of the form K 12s+s- e or of the form K 12 s+l0- K 3
have no nontrivial quotients.
248 Map Colorings

2. Trace the edge orbits in Figure 5.24, in order to demonstrate that there are
three orbits and that each nonzero current occurs exactly once on each
orbit.
3. Use the pattern of Figure 5.24 to develop a current graph for the general
solution of orientable case 5.
4. Draw the imbedded voltage graph used for K 22 •
5. Use Figures 5.28 and 5.29 to construct a general solution to the ad-
ditional-adjacency problem for case 10.
6. Describe the face-size distribution for the voltage graph dual to Figure
5.30 and the order of the net voltage on each face.
6
The Genus of a Group

In calculating the genus of the complete graph Kn, we represented Kn as a

Cayley graph for the cyclic group ~n. By way of generalization, one might
hope that similar methods can be used to compute the genus of Cayley graphs
for other groups. As we saw in Chapter 5, the difficulties encountered even for
cyclic groups were numerous. Thus, it is natural to suspect that calculating the
genus of Cayley graphs for noncyclic groups might be an intractable problem,
unless the choice of generating set is somehow restricted. Accordingly, in this
chapter, we restrict our study to the case in which the generating set X is
"irredundant", by which we mean that no proper subset of X also generates
the group. In fact, to diminish the role of generating set even more, we shall
focus on determining the minimum genus over all Cayley graphs for a given
groups#, which White (1972) has called the genus of the group, and denoted
by y(s#).
One consequence of this concentration on the group itself rather than the
generating set is that one is led inevitably to "highly symmetric" imbeddings,
for which the group action on the Cayley graph extends to the imbedded
surface. Surprising and intricate relationships are revealed between the genus
of a group and the minimum genus surface on which the group acts. The result
is a complex interplay of topology, combinatorics, and algebra.
Levinson (1970) has proved that the genus of an infinite group is either 0 or
infinite (see Exercises 13 and 14). Accordingly, the emphasis is on the genus of
finite groups.

6.1. THE GENUS OF ABELIAN GROUPS

Abelian groups are an obvious first target in a general attack on the genus of a
group. Indeed, White (1972) initiated the modem study of the genus of a
group by computing the genus of certain abelian groups, using methods
discussed in Section 3.5. Jungerman and White (1980), whose work is now
described, subsequently succeeded in determining the genus of "most" other
abelian groups.
Before we proceed, a comment about generating sets is in order. A
generating set X for a group s# can be irredundant without being the
249
250 The Genus of a Group

O ---J)
....__.,.,-"'---/ /
,-
______
..,...
\ I
Figure 6.1. The Cayley graph C(.9'3 ,
"-
X) and two quotients. Edges corresponding to generator
(1 2 3) are solid, and generator (1 2), dashed.

minimum in size over all generating sets for .91. For example, consider the
group .91= ~6 and the irredundant generating set {2, 3}. If the generating
set X has "minimum" size, then the number #X is called the "rank" of the
group .91.
In computing the genus of a group, one need consider only irredundant
generating sets, because discarding extra generators serves only to delete edges
from the associated Cayley graph. Obviously, deleting edges cannot increase
the genus of a graph. On the other hand, as we shall see, it is possible that the
genus of a group is not attainable by using any of its minimum-size generating
sets. (See Exercise 6.4.16.) This was first observed by Sit (1980), in response to
a problem of Gross and Harary (1980).

6.1.1. Recovering a Cayley Graph from Any of Its Quotients

For many applications, it is helpful to know how to assign voltages to a given
quotient of a given Cayley graph, so as to recover the Cayley graph. For
example, consider the full symmetric group .51'3 and let X be the generating
set consisting of the permutations (1 2 3) and (1 2). The Cayley graph C( .51'3 , X)
is depicted in the center of Figure 6.1. On the left is the quotient graph G
under the cyclic action on C( .51'3 , X) generated by (1 2 3). On the right is the
quotient graph H under the cyclic action on C( .51'3 , X) generated by (12).
Observe that since the subgroup of .51'3 generated by (1 2 3) is normal, the
graph G is itself a Cayley graph for the group ~2 , with generating set cf>( X),
where cf>: .51'3 ~ ~2 is the canonical quotient map. On the other hand, since
the subgroup in .51'3 generated by (1 2) is not normal, the graph H is not a
Cayley graph, but a Schreier coset graph. In either case, the problem is to
assign voltages to the quotient so as to recover the original Cayley graph
C(sP3 , X).
The Genus of Abelian Groups 251

It is tempting simply to assign the voltage x to each edge in the quotient

graph that is the image under the quotient map of an edge colored x in the
Cayley graph C(s#, X). In the present example, the derived graph for such a
voltage assignment on base graph C would have 12 vertices, and on base graph
H, 18 vertices. If this assignment is to work, it is necessary that the derived
graph consist of n copies of the original Cayley graph, where n is the number
of vertices in the quotient graph. The following theorem establishes such a
property. By the "natural" voltage assignment on the Schreier coset graph
S(s#: !!J, X), we mean the map into the groups# that assigns the voltage x to
each directed edge colored x for every generator x E X.

Theorem 6.1.1. The natural voltage assignment a on the Schreier coset graph
C = S(s#: !!J, X) yields a derived graph with #[.5#: !!J] components, each
isomorphic to the Cayley graph C(s#, X).

Proof Suppose that !!J1 = !!J, !!J2 , ••• , !!Jn are the right cosets of !!J. Then
the n#s# vertices of the derived graph ca are the pairs (!!J;, a), where a E s#
and 1 5: i 5: n. There is an edge from the vertex ( !!J; , a) to the vertex ( !!J1 , b)
if and only if there is a generator x E X such that !!J1 = !!J;x and b = ax. We
define the vertex part of a graph map f from the Cayley graph C( s#, X) to the
derived graph ca by the rule f(a) =([a], a), where [a] denotes the right coset
of !!J that contains a. Also, let f take the edge of C(s#, X) from a to ax to
the edge in co. from ([a], a) to ([a]x, ax). Since [a]x = [ax], we infer that the
resulting map is a graph map. It clearly takes the set of directed edges
originating at the vertex a of C(s#, X) one-to-one onto the set of directed
edges originating at the vertex ([a], a) of ca. Therefore, f is a covering map
from C( s#, X) onto a component of ca. Since f is also one-to-one on the
whole vertex set of C(s#, X), it follows that f is a graph isomorphism onto
that component. The theorem follows, since the components of the derived
graph of an ordinary voltage graph are mutually isomorphic. 0

Corollary 1. Let X be a generating set for a groups#, and q, a homeomorphism

from s# onto the group !2. Let a be the voltage assignment for the Cayley graph
C(!l, <f>(X)) that gives a directed edge colored <f>(x) the voltage x for each
x E X. Then the derived graph consists of #!2 copies of the Cayley graph
C(s#, X).

Proof Apply Theorem 6.1.1 with the kernel of q, as the subgroup !!J, so
that the Schreier coset graph S( .5#: !!J, X) is isomorphic to the Cayley graph
C(fl, q,(X)). o

Corollary 2. Let X be a generating set for the group d. Suppose that

(C--+ S, a) is an oriented current graph with current group s# satisfying these
252 The Genus of a Group

conditions:

i. In each directed face boundary there is exactly one edge assigned the
current x and one the current x-I for every generator x E X, and all
currents or their inverses are in X.
ii. There is a subgroup ?A contained in .J/1 and a one-to-one correspondence
between the faces of the imbedding G --+ S and the cosets of ?A, such that
if the edge e lies in faces corresponding to cosets ?A 1 and ?A 2 , then
?A 1a( e •) = ?A2 , where e • is the direction of e given by the directed
boundry of the face for ?A 1 •

Then each component of the derived graph Ga. is isomorphic to the Cayley graph
C(.J/1, X).

Proof Conditions i and ii imply that the dual graph for the imbedding
G --+ S is the Schreier coset graph S(.J/1: ?A, X) and that the dual voltage
assignment is the natural one. The corollary then follows from Theorem 6.1.1.
D

The following two examples further develop the quotients of Figure 6.1 in
the context of Theorem 6.1.1 and its corollaries. The first example concerns the
left quotient, which is regular, and the second the right quotient, which is not.

Example 6.1.1. On the left of Figure 6.2 is a voltage graph imbedded in the
sphere with voltages in the symmetric group .51'3 generated by X= { x, y }, where
x = (1 2 3) and y = (1 2). The given voltage graph is the Schreier coset graph
S( .9'3 : ?A, X) where ?A is the subgroup generated by (12 3). Alternatively, since
?A is normal in .51'3 with quotient map <P: .9'3 --+ f!E2 such that <P(x) = 0 and
<P(y) = 1, the voltage graph may be viewed as the Cayley graph C(f!E2 , {0, 1}).
The vertices have been labeled by the cosets ?A and PAy (rather than 0 and 1) to
maintain the Schreier coset viewpoint. By Theorem 6.1.1, the derived imbedding
for this voltage graph consists of two copies of the Cayley graph C( .9'3 , X). If one
prefers to view the voltage graph as a Cayley graph, use the first corollary to
Theorem 6.1.1 instead. On the right of Figure 6.2 is the dual current graph. The

-- '
O _!_n
I / ...... \
I :dy \
x I I

'---'~
• .. i 'x • :y
X
y 1Y I
~ I

' .... _....

X= (12 3) \ I
y= (12) /

Figure 6.2. An imbedded voltage graph and its dual current graph.
The Genus of Abelian Groups 253

I
ly

l
y y

l -,'l
X
I 1,. X= (12 3)
1
X X
r
I
I
I
X \X

-- y
/
y = (12)

Figure 6.3. A current graph and its dual voltage graph.

reader should verify that the edge and face labels satisfy conditions i and ii of the
second corollary.

Example 6.1.2. On the left of Figure 6.3 is a current graph imbedded in the
torus with currents in the symmetric group 9"3 generated by X= { x, y }, where
x = (1 2 3) andy = (1 2). Clearly condition i of the second corollary to Theorem
6.1.1 is satisfied. If the faces are labeled as illustrated, by cosets of the subgroup
!!4 generated by the permutation (1 2), then condition ii is also satisfied. Thus the
derived graph for this current graph consists of two copies of the Cayley graph
C( 9"3 , X). On the right of Figure 6.3 is the dual voltage-graph imbedding, which
the reader should recognize as the right quotient in Figure 6.1, the Schreier coset
graph S(Y'3 : !!4, X).

The information contained in an imbedded voltage graph or current graph

like those in Examples 6.1.1 and 6.1.2 can be coded in tabular form. Since the
voltages uniquely identify each directed edge incident on a given vertex, it
follows that, to give the rotation system for such an imbedded voltage graph, it
suffices to list the cyclic ordering of voltages at each vertex, where each vertex
is identified by a coset of the subgroup !!4. For example, the imbedded voltage
graph of Example 6.1.2 is given by the following table, in which we adopt the
notations x = (1 2 3) and y = (1 2) of the figures:

!!4 x y x-1 y-1

!!Ax x y x- 1 y- 1
!J4x2 x y-1 x-1 y

In terms of the dual current graph, this table gives the currents encountered in
a trip around each face boundary. In the early literature, these are sometimes
called the "logs of the circuits". In effect, the table of logs is just a condensed
rotation system for the derived imbedding of the Cayley graph C( s#, X). That
is, edges incident on a vertex are identified by their "color" x or x- 1, x E X,
and the rotation at vertex a, where a is in coset !!4,, is simply the log for
circuit !!4,.
254 The Genus of a Group

6.1.2. A Lower Bound for the Genus of Most Abelian Groups

The classification theorem for finite abelian groups (e.g. see MacLane and
Birkhoff (1967)) states that any finite abelian group .J/1 of rank r has a unique
"canonical form"

such that m; divides mi+l fori= 1, ... , r- 1, and m 1 > 1. Each element of
.J/1 is written as an r-tuple whose coordinates correspond to the canonical
product form. The "canonical generating set" for .J/1 consists of the r elements
that have 1 as one coordinate and 0 for all others.
Cyclic groups !!En and abelian groups of the canonical form f!E2 X !IEm
have planar Cayley graphs, and groups with canonical form f!Em 1 X f!Em 2 ,
m 1 > 2, have genus 1 (see Exercise 6). Also, if the canonical form of an
abelian group has any f!E2-factors, then its genus is easily computed using
methods of Section 3.5. Therefore, in the rest of this section, we consider only
abelian groups of rank r > 2 having no f!E2-factors in their canonical form.
The following theorem establishes a lower bound on the genus of such an
abelian group that has no f!E3-factors in its canonical form either. The bound
given is simply the genus of an imbedding of C( .J/1, X), where X is minimum
size and all faces of the imbedding are quadrilaterals. At first glance this
bound appears to be obvious, since any cycle of length 2 or 3 in a Cayley
graph for an irredundant generating set must come from a generator of order 2
or 3, and the canonical generating set has no elements of order 2 or 3 under
the given restrictions. However, there are other irredundant generating sets,
and they can have elements of order 2 or 3. The argument needed to eliminate
such generating sets is actually rather delicate.

Theorem 6.1.2 (Jungerman and White, 1980). Let the abelian group .J/1 have
the canonical form !!Em X · · · X !!Em such that r > 1 and m; > 3 for all i. Then
I '
y(.J/1);;::: 1 + #.J/1· (r- 2)/4.

Proof Let X be a generating set for .J/1 such that y(C(.J/1, X))= y(.J/1).
As we have already observed, it may be assumed that X is irredundant. It
cannot be assumed that X is canonical or even that #X= r.
Let s be the number of generators in X of order at least 4, let t 2 be the
number of order 2, and let t 3 be the number of order 3. We first claim that
s + t 2 ;;::: r and s + t 3 ;;::: r. To show that s + t 3 ;;::: r, let !14 be the subgroup of
.J/1 consisting of all elements of order 2. Then b E !14 if and only if the ith
coordinate of b is 0, when m; is odd, and either 0 or m;/2, when m; is even.
Let <f>: .J/1-+ !2 be the natural quotient homomorphism whose kernel is !14.
Clearly, the group !2 is isomorphic to !!En X · · · X !!En , where n; = m; if m; is
I '
odd and n; = m;/2 if m; is even. In particular, the group !2 has rank r, since
m; > 2 for all i. Therefore, the set </>(X) must contain at least r nonidentity
The Genus of Abelian Groups 255

elements, because <P( X) generates fl. Since every element of order 2 is sent by
<P to the identity, it follows that s + t 3 ~ r as claimed. A similar argument
using the subgroup of elements of order three establishes the other inequality
s + t 2 ~ r.
Now suppose that C( .J/1, X) ~ S is an orientable imbedding. Let /; be the
number of i-sided faces. Then

because digons and triangles arise only from elements of order 2 or 3 in the
irredundant generating set X. Since '£if;= 2#£, it follows that

Therefore

Hence

x(S) = #V- #E + #F
~ #V- #E/2 + #.Jil(t 2 + t 3/3)/4
= #.91- #.9/(s + t 2 + t 3 )/2 + #.Jil(t 2 + t 3/3)/4
~ #.91(1 - (2s + t 2 + t 3 )/4)

From the inequalities s + t 2 ~ r and s + t3 ~ r, we obtain

x(S) ~ #.91(1- r/2)

and consequently

y(S) ~ 1 + #.9/(r- 2)/4 D

6.1.3. Constructing Quadrilateral Imbeddings for Most Abelian Groups

The lower bound for y(J/1) in Theorem 6.1.2 could be realized by a quadri-
lateral, orientable imbedding of a Cayley graph C(.J/1, X) with #X= r. Since
commutator relations in an abelian group provide an abundance of quadri-
laterals, one expects that y(J/1) = 1 + #.9/(r- 2)/4 whenever this number is
an integer and the canonical form for .91 contains no .2'3-factors. This is
precisely what Jungerman and White have shown, except when the group .91
has rank 3 and the first factor in the canonical form for .Ji1 is odd.

Theorem 6.1.3 (Jungerman and White, 1980). Let the abelian group .91 have
the canonical form .2'm X • · · X .2'm , where r > 1 and every m; > 3, and such
1 '
that either every m; is even or r > 3. Then y(J/1) = 1 + #.9/(r- 2)/4,
whenever the number on the right-hand side of this equation is an integer.
256 The Genus of a Group

a
Figure 6.4. A voltage graph and its dual current graph.

Proof The proof is broken down into various overlapping cases. In each
case, a generating set X is given such that #X= r and a quadrilateral,
orientable imbedding for the Cayley graph C( d, X) is constructed. In all but
Case 1, the imbedding is obtained by a voltage-current-graph construction,
using a quotient graph of C( d, X) corresponding to a certain subgroup 114.

Case 1. Every m; is even.

Historically, this was the first case handled (White, 1972). It is a conse-
quence of Theorem 3.5.7, where we take X to be the canonical generating set
for the group d. We observe that the Cayley graph C(d, X) is isomorphic to
the cartesian-product graph em X · · · X em . Although the proof of Theorem
I '
3.5.7 does not involve voltage graphs, it is also possible to obtain this
imbedding by a voltage-graph construction that has for its base graph a
Cayley graph for the rank-r abelian 2-group ~2 X · · · X ~2 (see Exercise 1).

Case 2. r even, m, even.

Let X; be the element of the group d having its rth and ith coordinates
equal to 1 and all others 0. Let 114 be the subgroup consisting of all elements
with even rth coordinate. Then the quotient group d/114 is isomorphic to ~2 ,
and the Schreier coset graph S( d: 114, X) has two vertices and no loops.
Consider for r = 4 the imbedded voltage graph given on the left in Figure 6.4.
The base graph is S( d: 114, X), the base surface is S2 , and the voltage
assignment is the natural one. On the right is the dual current graph (the
reader should check that conditions i and ii of the second corollary to
Theorem 6.1.1 both hold). The pattern for arbitrary even r should be obvious.
It is easily verified that every face of the imbedded voltage graph is a
quadrilateral with boundary walk of the form x;x,+l•x;- 1 X;+ 1 -•, t: = ±1.
Dually, KeL holds at every vertex of the current graph. Thus the derived
graph is C(d, X), and every face of the derived imbedding is a quadrilateral.

=
Case 3. r 2 mod 4, m 1 odd.
Let X be the canonical generating set for d. Let 114 be the subgroup
containing every element of d whose sum of coordinates is congruent to
0 mod m 1 . Then the quotient group dj 114 is isomorphic to ~m 1 , and each
directed edge of the Schreier coset graph S( d: 114, X) originating at vertex j
The Genus of Abelian Groups 257

--n TI: Q
r 1 2 3 r- 2
-1
m-1
9d 1 r r-1 r-2

9dm-}, 1 -1
• r
2. r-2
•• •• ••
• •
r-1

•••

•••

Figure 6.5. The current graph for Case 3. Hollow vertices have counterclockwise rotation and
solid vertices clockwise rotation.

terminates at the vertex j + 1 mod m1 . Consider the current graph given in

Figure 6.5. To avoid a blizzard of subscripts, we denote the canonical
generator X; simply by i, fori = 1, ... , r. An unfinished edge on the left of the
diagram is to be identified with the unfinished edge on the right carrying the
same current; similarly for unfinished edges at the top and bottom. To avoid
double subscripting, let m = m 1 • The correspondence between faces and
cosets of !14 is given by the capital letters !140 , .•. , !14m_ 1 , where every element
in !141 has a sum of coordinates congruent to j mod m.
The voltage graph dual to the current graph given in Figure 6.5 is more
difficult to draw, because the valence, 2r, of each vertex is too large. Instead,
we give a rotation system for the imbedding of S(.91: !14, X). That is, a table of
logs for the current graph of Figure 6.5. As in the current graph, we simply
denote generator X; by i, and xj 1 by i. The vertices of S(.91: !14, X) are the
cosets !140 , •.• , !14m _ 1 , where again elements of !14; have sum of coordinates
equal to i mod m. There are four kinds of rotation (logs): one for vertex
!14m_ 2 , one for vertex !Jim-1> one for all other vertices !141 with j even, and one
for all other vertices !141 , with j odd, as given here for r = 6:

!141 , j even : 123456123456

654321654321
161654543232
161654543232
...
~

.c?itu,
2 1 2 3 4 3 4 5 6 5 6 7 8 r-1 r-2

••• xx~ ~
1
2 1 2 3 4 3 4 5 6 5 6 7 8 r-1 r-2 -
!JdoJ, .c?itwi

u~u,
2 1 r 2 3 4 3 4 5 6
-1
••• -1
r-1 r-2 '.,.,..c?itw,
"'j+ I

1 4
n~u,
2 r 2 3 r 4 3 5
6
1
• • • 1

.:> ~ 6 r-1 r-2 ~-:c16'1o~

'-'
" 1 4
~

8ito11 2 3

8ito1 1 " -
2
, c.. ..., ... , '+ ,J 6
·nt
•••
r-1
1
r-2 ~-!Jdw~

Figure 6.6. The current graph for Case 4. Rotations at hollow and solid vertices as in Figure 6.5.
The Genus of Abelian Groups 259

It can be verified from this rotation scheme that all faces are quadrilaterals
satisfying KVL. For example, start with edge 1 (i.e., x 1 ) at vertex !?.4m_ 3 • That
edge terminates at vertex !14m_ 2 , where we obtain the next edge, 6. That edge
takes us to vertex !JBm-t· There we obtain edge I, which returns us to vertex
!?.4m_ 2 , where we find edge 6 awaiting to return us to !?.4m_ 3 • The next edge is 1
again, so the face is complete (since m - 3 is even, the rotation for !?.4m_ 3
is given by the first row). The face is a quadrilateral with net voltage
x 1 + x 6 - x 1 - x 6 = 0.

Case 4. r odd, r ~ 5, m,_ 1 and m, even, m 1 odd.

If i is even, let X; have (r- 1)th, rth, and ith coordinate equal to 1 and all
others equal to 0; if i is odd and i =F 1, let X; have rth and ith coordinate 1
and all others 0. Finally, x 1 has first coordinate 1 and all other coordinates 0.
Then X= {X; 11 :;;; i :;;; r} is a generating set for the group .91. Let !14 be the
subgroup of all elements having first coordinate 0, and (r- 1)th and rth
coordinates even. Then .91/!14 = fl'm I X fl'2 X fl'2 • The Schreier coset graph
S(.91: !14, X) is a Cayley graph for .91/!14 having edges labeled X; from vertex
{j, m, n) to vertex {j, m + 1, n + 1) fori even, from {i, m, n) to {i, m, n + 1)
for i odd and i =F 1, and from {j, m, n) to {j + 1, m, n) for i = 1. The desired
current graph is given in Figure 6.6. Unfinished edges on the left and right
match up as in Case 3. Unfinished edges disappearing into the hole a1 (resp.,
{31 , Y) reappear in hole a1+ 1 (resp., {31 + 1 , y1+ 1 ). As in Case 3, we abbreviate i
for x,.
The corresponding directed rotation system for the dual voltage graph is as
follows for r = 7.
j even, j =F m - 1:
(j,O,O). 71671654543232
(j,0,1). 71671654543232
(j,1,0). 61761723234545
(j, 1, 1). 61761723234545
j odd, j =F m - 2:
(j,O,O). 61761723234545
- - -
(j,O, 1). 61761723234545
(j,1,0). 71671654543232
(j, 1, 1). 71671654543232
j = m- 2:
(j,O,O). 76176514543232
(j,0,1). 76176514543232
(j,1,0). 15671672323454
(j,1,1). 15671672323454
Ill

Figure 6.7. The current graph for Case 5. Rotations at hollow and solid vertices as in Figure 6.5.

260
The Genus of Abelian Groups 261

j = m- 1:

(j,O,O). 76176541543232
(),0,1). 76176541543232
(),1,0). 45145671672323
(),1,1). 45145671672323

For arbitrary r, the generators 1, r - 3, r - 2, r - 1, r behave like

1, 4, 5, 6, 7, respectively, whereas all other generators follow, in pairs, the
pattern of 2, 3. For example, for r = 9, the second row under j = m- 1
would read:
--- --
(),0,1). 981987617654543232

As in previous cases, it can be verified from the scheme above that every
face is a quadrilateral satisfying KVL. For example, starting with edge 1 at
vertex (m- 3,0, 1), one obtains next the edge 7 at vertex (m- 3, 1, 1), then
edge I at vertex (m - 2, 1, 1), then edge 7 at vertex (m - 3, 1, 1), then edge 1
at vertex ( m - 3, 0, 1) again. We notice that m - 3 is even, so that rotations
come from the first and third groups.

Case 5. r odd, r ~ 5, m, = Omod4, m 1 odd.

Let x have rth coordinate 2, ith coordinate 1, and all other coordinates 0,
if i is even; let x, have rth coordinate 1 and all other coordinates 0, if i is odd
and i -:/= 1. Finally, x 1 has first coordinate 1 and all other coordinates 0. Then
X= { x, 11 :;;; i :;;; r} is a generating set for the group d. Let ?A be the
subgroup of all elements having even (r- 1) and rth coordinates and first
coordinate 0. Then s#j?A = fl'm X fl'4 , where again m = m 1 . The Schreier
coset graph S(s#: ~.X) is a Cayley graph for the quotient group s#j?A that
has edges labeled X; from vertex (j, m) to vertex (j, m + 2) for even i, from
vertex (j, m) to vertex (j, m + 1) for odd i and i-:/= 1, and from (j, m) to
(j + 1, m) fori= 1. The desired current graph is given in Figure 6.7. Again,
broken edges are identified by the same conventions as in Case 4.
The corresponding directed rotation system for the dual voltage graph is as
follows for r = 7:
j even, j -:/= m - 1:

(j,O). 76171632325454
(),1). 61716745452323
(),2). 61716745452323
(),3). 76171632325454
262 The Genus of a Group

j odd, j *m- 3:

(j,O). 61716745452323
(J, 1). 76 I 11 63 2 3 254 54
(},2). 76171632325454
(},3). 61716745452323

j = m- 2:

(j,O). 54514767163232
(},1). 67167451452323
(J' 2). 61 76 7 4 1 5 4 5 2 323
(},3). 54154761763232

j = m- 1:

(j,O). 54514761763232
(},1). 61767415452323
(},2). 67167451452323
(},3). 54514767163232

As in Case 4, for arbitrary r, generators 1, r - 3, r - 2, r - 1, r behave like

1, 4, 5, 6, 7, whereas all other generators behave in pairs like 2, 3. For example,
the vertex (j, 2) for j = m - 1 has the following rotation for r = 9:

(},2). 891897817823234545

As in the previous cases, it can be verified that every face is a quadrilateral

satisfying KVL. D
The smallest abelian group whose genus is not ascertainable by application
of Theorem 6.1.3 is fr3 X fr3 X fr3 • For this group, we have the following
special result, stated here without proof.

Theorem 6.1.4 (Mohar et al., 1985; Brio and Squier, 1986). The genus of the
group fr3 X !Z3 X fr3 is 7. D

It is not difficult (see Exercises 10 and 11) to prove that

and for about a decade, it was commonly suspected that the genus was 10,
based on the reasoning that a lower genus could not be realized with a
The Genus of Abelian Groups 263

symmetric imbedding. After Mohar et al. (1985) produced a chaotic imbed-

ding in the surface S 7 , Brin and Squier (1986) responded with an ingenious
case-by-case demonstration that no surface of smaller genus would be ade-
quate.
Just to illustrate the difficulties posed by .2"3-factors, although it is not hard
to prove that y(.2"3 X .2"3 ) = 1 (see Exercise 6.4.14), there are many different
toroidal imbeddings of the Cayley graph for .2"3 X .2"3 , each corresponding to
an irredundant generating set. Beyond the obvious quadrilateral imbedding,
there is an imbedding having six triangular faces and three hexagonal faces
(Exercise 7), and there is another imbedding having six triangles, two quadri-
laterals, and one face of size 10 (Exercise 8). In the same way, there are a
variety of bizarre imbeddings of Cayley graphs for .2"3 X .2"3 X .2"3 in the
surface of genus 10 (see Exercises 9 and 10).

6.1.4. Exercises
1. Construct a voltage-graph imbedding (or current graph) for Case 1 of
Theorem 6.1.3 using the quotient group .2"2 X · • · X .2"2 of rank r.
2. Verify that conditions i and ii hold for the current graph in Case 2 of
Theorem 6.1.3. Give the table of logs for this current graph.
3. Check that KCL holds at every vertex of the current graph given for
Case 3 of Theorem 6.1.3. Verify at least three rows in the associated table
of logs given in the text.
4. Same as Exercise 3, but for Case 4 of Theorem 6.1.3.
5. Same as Exercise 3, but for Case 5 of Theorem 6.1.3.
6. Prove that y(.2"m X .2"n) = 1, where m divides n and m > 2 (see also
Exercise 6.4.14 for the case m = n = 3).
7. Use a one-vertex voltage graph imbedded in the sphere to obtain a
toroidal imbedding of a Cayley graph for .2"3 X .2"3 •
8. Let X = {x, y} be an irredundant generating set for .2"3 X .2"3 • Give a
rotation scheme for an imbedding of C( .2"3 X .2"3 , X) that has six
triangles and at least one quadrilateral. (Hint: Draw a rectangle xyx- 1y- 1
bordered by a pair of x 3 triangles and a pair of y 3 triangles. This gives
you the full rotation at the four vertices of the rectangle and part of the
rotation at four more vertices. Now complete the rotation system so that
the remaining x 3 and y 3 circuits are faces.) Given that y(.2"3 X .2"3 ) = 1,
the imbedding constructed in this way must be toroidal. Why? What are
the sizes of the faces?
9. Use one-vertex voltage graphs to construct three different imbeddings of
a Cayley graph for .2"3 X .2"3 X .2"3 in the surface S 10 : one with 9
triangles, one with 18, and one with 27.
10. Let X = {x, y, z} be an irredundant generating set for .2"3 X .2"3 X .2"3 •
Let !14 be the subgroup generated by z. The Schreier coset graph
264 The Genus of a Group

S(.s#: !14, X) is just a Cayley graph for ~3 X ~3 , having x and y as

generators, with an additional loop for z attached at each vertex. Start
with the usual quadrilateral imbedding of a Cayley graph for ~3 X ~3
and then place z-loops inside some of the faces so that KVL holds
around every face not enclosed by a z-loop. (The directions of the
z-loops matter; one of the original quadrilateral faces contains three
z-loops.) Show that the derived imbedding of C(~3 X ~3 X ~3 • X)
given by the natural voltage assignment has genus 10. Find the sizes of
all faces of the derived imbedding.
11. Prove that y(~3 X ~3 X ~3 ) ~ 5.
12. Construct an imbedding of a Cayley graph for ~3 X ~3 X ~3 in the
nonorientable surface of Euler characteristic x = -15. Notice that
x(S 10 ) = -18. (A three-vertex voltage graph imbedded in the projective
plane can be used. See also Exercise 6.4.15.)
13. Let .s# be an infinite group and .r any finite subset of .91. First prove
that there exists an element b in .s# such that bYn .r is empty. Next,
suppose that a Cayley graph for the group .91 has a subgraph homeomor-
phic to K 5 or to K 3 3 • Prove that .s# has infinitely many disjoint
subgraphs homeomorphic to K 5 or to K 3 , 3 •
14. Assume the infinite version of Kuratowski's theorem, due to Dirac and
Schuster (1954), that an infinite graph can be imbedded in the plane if
and only if it contains no subgraph homeomorphic to K 5 or to K 3 3 • Use
Exercise 13 to prove Levinson's theorem (1970) that the genus of an
infinite group is either 0 or infinite.

6.2. THE SYMMETRIC GENUS

In attempting to determine the genus of a given Cayley graph C(.Jil, X), one
strategy is to suppose that there exists a minimum-genus imbedding that
somehow reflects the symmetry of the graph. In particular, the natural action
of the group .91 on the Cayley graph C( .s#, X) might extend to an action of .91
on the imbedding surface. If the natural action does so extend, then we call the
imbedding "symmetric". Furthermore, if the natural action is orientation
preserving, then we call the imbedding "strongly symmetric" and, otherwise,
"weakly symmetric". The "symmetric genus" a(.s#) (resp., "strong symmetric
genus" a 0 ( .91 )) is then defined to be the smallest number n such that the
surface sn contains a symmetric (resp., strong symmetric) imbedding of a
Cayley graph for the group .91. A main result of this section is that if the group
.91 acts on an orientable surface S, then there is a symmetric imbedding in S
of a Cayley graph for .s#. It follows from this result that the symmetric genus
could be defined equivalently without any reference at all to generating sets or
Cayley graphs.
The Symmetric Genus 265

The concept of strong symmetric genus is actually the oldest version of a

genus for groups. A form of it, where the quotient surface of the group action
is restricted to be the sphere, was studied by Dyck (1882), in the context of
automorphism of maps on surfaces. Burnside (1911) devotes Chapters 18 and
19 to this "graphical representation of a group" by map automorphisms and
determines all groups of "genus" 0 and 1. To those working on conformal
automorphisms of Riemann surfaces, "genus" means smallest genus surface,
other than the sphere or torus, in which the group has a strongly symmetric
imbedding. Maclachlin (1965) computes this genus for abelian groups. On the
other hand, Tucker (1983) seems to have been first to introduce the symmetric
genus, although Levinson and Maskit (1975) do discuss a related concept
called the weak point-symmetric genus.

6.2.1. Rotation Systems and Symmetry

Suppose C(.J/1, X)~ S is an imbedding of a Cayley graph in an oriented
surface. Since the edges of C(.J/1, X) are directed and labeled by elements of
the generating set X, a rotation at a particular vertex can be represented as a
cyclic ordering of the elements of X and their inverses. Thus, it makes sense to
compare rotations at different vertices.

Example 6.2.1. The dihedral group !i)4 is given by the presentation (x, y:
x 4 = y 2 = 1, yxy- 1 = x- 1 ), with generating set X= {x, y}. Consider the
planar imbedding of the Cayley graph C(!i)4 , X) given on the left in Figure 6.8.
Edges corresponding to x are solid, and edges corresponding toy are dashed.
Every vertex has the (clockwise) rotation xx- 1y- 1y.

Is the imbedding of Example 6.2.1 symmetric, or even strongly symmetric?

Perhaps it is not exactly clear geometrically how the natural action of f)4 on

'~"'...
, 'I
1:,4/ I
I
I
I

~
i..,~ ~~,
'
Figure 6.8. An imbedding of a Cayley graph in the sphere.
266 Tile Genus of a Group

C(~ 4 , X) might extend to the sphere. On the right in Figure 6.8 is an

equivalent imbedding, depicted as a cube. In the right-hand figure, the vertices
of C(~4 , X) have been labeled by elements of ~4 • Left multiplication by x
clearly rotates the bottom quadrilateral by 90° about the vertical line through
its center, counterclockwise when looking from above. Left multiplication by x
has the same effect on the top quadrilateral, since xy = yx- 1• Thus the natural
action on C( .J/1, X) given by left multiplication by x extends to a 90° rotation
of the cube. Similarly, it is easily verified that left multiplication by y is
equivalent to a 180° rotation about the line through the centers of the left
front and right rear digons. Thus, the imbedding is symmetric in our precise
sense. In fact, since the action on the sphere involves only rotations, and since
rotations preserve orientation, the imbedding is strongly symmetric.
Although geometric visualization is important, one might also want a purely
combinatorial way to show that a given imbedding is strongly symmetric. To
discover such a way, we first observe that the natural action of a group .Ji1 on
the Cayley graph C( .J/1, X) respects edge labels and directions. Accordingly, if
the action of .Ji1 extends to an imbedding surface, it must also respect the
rotation system, in the sense of Section 4.3. That is, if the imbedding
C( .Ji1, X) --+ S is strongly symmetric, then the rotation at every vertex is the
same, and if the imbedding is symmetric then the rotation at every vertex is
the same or reversed. For strongly symmetric imbeddings, the converse is also
true. If the rotation at every vertex is the same for the imbedding C(.J/1, X) --+
S, then by Theorem 4.3.3, the natural action of .Ji1 extends to an orientation-
preserving action on the surface. Summarizing, we have the following result.

Theorem 6.2.1. An imbedding C(.J/1, X) --+ Sis strongly symmetric if and only
if the rotation at every vertex is the same. If an imbedding C( .J/1, X) --+ S is
symmetric, then the rotation at every vertex is the same or reversed.

It might be helpful to restate Theorem 6.2.1 in terms of voltage graphs, at

least for strongly symmetric imbeddings. Since the extended action can be
chosen to have at most one fixed point in each face of the imbedding, the
natural projection to the quotient surface is a regular branched covering, and
hence by Theorem 4.3.5, it can be obtained by a regular voltage graph
construction.

Corollary. An imbedding C(.J/1, X)--+ Sis strongly symmetric if and only if it

it can be obtained as the derived imbedding for a one-vertex, regular voltage
graph. 0

In this context, the Dyck-Burnside genus of a group .Ji1 is simply the

minimum genus obtained by a one-vertex voltage graph "in the sphere".
The Symmetric Genus 21i7

Maclachlin's genus is the minimum genus "greater than 1" obtained by a

one-vertex voltage graph in any orientable surface.
The converse to the second statement of Theorem 6.2.1 is false, as readily
illustrated by some nonsymmetric imbeddings for 2-generator presentations of
odd-order cyclic groups. Since an orientation-reversing homeomorphism must
reverse every rotation and the action of a group on its Cayley graph has no
fixed points, exactly half the vertices in a weakly symmetric imbedding have
one rotation and the other half have the reverse. In particular, no group of odd
order can have a weakly symmetric Cayley graph imbedding. Necessary and
sufficient conditions for a rotation system to yield a weakly symmetric
imbedding are considered in Exercise 1.
One consequence of Theorem 6.2.1 should be mentioned: the number of
symmetric imbeddings of a given Cayley graph is very small compared with
the total number of imbeddings. For example, if .91 = ~n X ~n X ~n, and if
X is the canonical generating set for .91, then C(.91, X) has (5!)n 3 different
imbeddings but only 5! different strongly symmetric imbeddings.
The situation for weakly symmetric imbeddings is greatly complicated by
involutions. A variation on Example 6.2.1 will illuminate the problem. The
group ~2 X ~4 has the presentation (x, y: x 4 = y 2 = 1, yxy- 1 = x), for
which we let X denote the generating set { x, y }. Figure 6.8 also gives an
imbedding of C( ~2 X ~4 , X) in the sphere if the directions of the x-edges in
the inside quadrilateral, on the left (or the top quadrilateral on the right) are
all reversed. Although the underlying graph is isomorphic to the previously
examined Cayley graph for the dihedral group f) 4 , the group action on the
graph has changed. Left multiplication by x is still accomplished by a 90°
rotation of the cube. However, left multiplication by y now takes the vertices
of the top quadrilateral to the vertices immediately below them. The only way
this can be accomplished is by a reflection about the horizontal plane through
the center of the cube. The trouble is that such a reflection takes each y-edge
to itself, whereas left multiplication by y takes each edge labeled y to its
"twin" edge.
Thus the imbedding of C(~2 X ~4 , X) is not even weakly symmetric, in
our precise sense, because the natural action of ~2 X ~4 on the Cayley graph
cannot be extended to the imbedding. This could also be discovered from the
rotation system of the imbedding. Every rotation is either xx- 1y- 1y or
x- 1xy- 1y, but these rotations are not the reverse of each other.
The difficulty here is really an artifact of representing involutions by
doubled edges. Suppose we avoid double edges by using the alternative Cayley
graph C 1 (~2 X ~4 • X), and we tolerate the fact that the "natural action" on
C 1( ~2 X ~4 , X) fixes some edges. That is, left multiplication by y now
simply reverses the undoubted y-edges in the same way as reflection in the
horizontal plane through the cube's center. In terms of rotation systems, every
rotation is either xx- 1y or x- 1xy, which are respective reverses of each other.
Throughout the rest of this chapter we shall use only alternative Cayley
graphs. This provides a minor variation on the voltage-graph viewpoint, since
268 The Genus of a Group

an alternative Cayley graph C 1(.91, X) with an involution in its generating set

is not the derived graph of a bouquet of circles.

6.2.2. Reflections
If the group .91 acts pseudofreely on the surface S, then the methods of
Chapter 4 involving branched coverings and voltage graphs suffice to describe
the action of .91 on S. However, some group actions on surfaces are not
pseudofree. Under the most general conditions, the fixed point set of a surface
homeomorphism can be very complicated, and one might fear that nonfree
actions are completely intractable. Fortunately, the behavior of a surface
homeomorphism of finite order is fairly simple, which is reassuring since our
interest is in finite groups of homeomorphisms. For example, the main
theorem of this section asserts that an orientation-preserving surface homeo-
morphism of finite order greater than 1 has only a finite number of fixed
points.
We have already called attention, in passing, to some group actions that are
not pseudofree, such as the action discussed just above of fl'2 X fl'4 on the
sphere, which involves reflection across the equator. For the sphere, reflection
in the equator is, in effect, the only finite-order homeomorphism having an
infinite number of fixed points. For a surface of higher genus, it is easy to
visualize a similar reflection using a plane in 3-space to split the surface in
"half". There are, however, other "reflections" that are not so easily visual-
ized.

Example 6.2.2. Let S be the torus given by the rectangle with sides to be
identified, as in Figure 6.9. Note that after the top and bottom are identified, to
form a cylinder, the left and right sides are identified by a half-twist. Let f be a
reflection of the rectangle about the center broken line. Then f induces a
homeomorphism h on the torus S. The broken line becomes a circle C left
pointwise fixed by h, whereas the line v u v becomes a circle C' that is taken to
itself by h but is not left pointwise fixed-it is turned 180°, taking u to v and v to
u. The circles C and C' split the torus S into halves that are interchanged by h.
On the right of Figure 6.9 is the torus with edges identified. The symbols "F"

u u
G1~

l I
I
u I u
I
I

u
F : 1u
Figure 6.9. A reflection of the torus.
The Symmetric Genus 269

and "G" are given, together with their images under the reflection, to show how
the homeomorphism h acts on the torus.

Up to this point the term "reflection", which is well defined in Euclidean

geometry, has been used loosely to apply to certain types of surface homeo-
morphisms. With Example 6.2.1 in mind, we can now be more precise. A
homeomorphism h: S ~ Son an orientable surface S is called a "reflection"
if

i. h is an orientation-reversing involution;
ii. S is the union of two connected surfaces with boundary S0 and S 1 such
that h(S0 ) = S1 and S0 n S1 is a collection of disjoint circles im-
bedded in S, at least one of which is left pointwise fixed by h.

The surfaces S0 and S1 are called the "halves" of the reflection h, and the
components of S0 n S1 are called the "dividing circles".
It can be shown (Exercise 3) that x(S) = 2x(S0 ). Since there are only
finitely many surfaces with boundary having a given Euler characteristic, the
choice of halves for reflections on a given surface is limited. For a torus, there
are only two kinds of reflections: the one given in Example 6.2.2 and the
obvious reflection that has the same halves as Example 6.2.2 but leaves both
dividing circles pointwise fixed. Reflections for a higher-genus surface are
considered in Exercise 4. The main result of this section can now be stated.

Theorem 6.2.2. Any finite-order homeomorphism of an orientab/e surface either

has a finite number of fixed points or is a reflection.

Proof The proof in the general topological case, due mostly to Kerekjarto
(1921), with later corrections by Eilenberg (1934), is very difficult and is not
considered here. If one assumes that the given homeomorphism is an automor-
phism of a graph imbedding that behaves "piecewise linearly" on the faces,
then the proof follows from easily verified facts about linear transformations
(see Exercises 5-10). D

Given a group d acting on an orientable surface S, the collection of

orientation-preserving elements of d forms a subgroup of index at most 2. It
is a subgroup because the composition of two orientation-preserving homeo-
morphisms is orientation-preserving. The index is at most 2, because the
composition of two orientation-reversing homeomorphisms is orientation-
preserving. This subgroup is called the "orientation-preserving subgroup" and
is denoted d 0 • A given group can have many index-2 subgroups, and which
one of these is the subgroup d 0 depends, of course, on the given action.

Corollary. Given an action of the group d on the orientable surface S, the

orientation-preserving subgroup d 0 acts pseudofreely.
270 Tile Genus of a Group

6.2.3. Quotient Group Actions on Quotient Surfaces

If the group .5!1 acts pseudofreely on the surface S, then by Theorem 4.3.6
there is a Cayley graph for .5!1 imbedded in S. We wish to prove that this
result holds even if the action of .5!1 is not pseudofree. At the same time we
should like to use voltage graphs and regular coverings as in Section 4.3 to
obtain information about presentations for the group .Jil. Since the orienta-
tion-preserving subgroup .5!1° does act pseudofreely, the main difficulty is to
understand the relation between the subgroup .5!1° and the reflections in .Jil.
This is accomplished by studying the natural action of the quotient group
.Jilj.Ji1° on the quotient surface S j.Ji1°.
Let .5!1 act on the surface S, and let ?A be a normal subgroup of .5!1 that
acts pseudofreely on S. Let p: S ~ Sj?A be the natural projection of the
regular branched covering associated with the action of ?A. Given an element
a in .5!1, let ii denote the corresponding element of the quotient group .Jilj?A.
Then an action of the group .Jilj?A on the surface S/PA can be given by
associating to ii the function 1/la: Sj?A ~ S/PA defined by 1/la(p(x)) =
p(cpa(x)), where cpa is the homeomorphism of S for the element a given by the
action of .5!1. This is called the "natural action" of the quotient group .Jilj?A on
the surface Sj?A.
Of course it must be shown that 1/la is well defined, that 1/la is a homeomor-
phism, and that the association ii ~ 1/; a does define an action, that is,
1/la"b = 1/lat/l"b· First, suppose that c = ii and p(y) = p(x). We must show that
1/lc(p(y)) = 1/la(p(x)). By definition 1/lcp(y) = pcpc(Y) and 1/la(p(x)) =
pcpa(x). Thus it suffices to show the orbit of cPc(y) equals the orbit of cpa(x)
under the action of ?A. Since we are given p(y) = p(x), it follows that
y = cpb(x) for some bE ?A. Hence cpc(Y) = cpccpb(x) = cpcb(x). Since cb= cb =
ii, cb = b'a for some b' E ?A. Thus cpcb(x) = cpb,cpa(x), which implies that
cl>c(Y) = cl>cb(x) has the same orbit as cpa(x), as desired. To show that 1/la is a
homeomorphism, we first observe that 1/laP = Pcl>a· Therefore, since p is a
(possibly branched) covering and cpa is a homeomorphism, it follows that 1/la is
a (possibly branched) covering. Moreover, since the number of sheets of the
branched coverings 1/laP and Pcl>a must be equal, we infer that 1/la is 1-sheeted,
because cpa is. Thus 1/la is a homeomorphism. The proof that 1/la"b = 1/lat/l"b is
left to the reader as Exercise 11.

Example 6.2.3. Let S be the surface of genus 2, imbedded in 3-space as shown

in Figure 6.10. Consider the action of the group .5!1= ~2 X ~2 X ~2 generated
by reflections in the three coordinate planes:

a: reflection in the yz-plane,

b: reflection in the xz-plane,
c: reflection in the xy-plane.
The Symmetric Genus 271

z z

y r

X
Figure 6.10. A surface of genus 2 and a quotient surface.

The orientation-preserving subgroup .91° consists of the identity and the 180°
rotations ab, ac, be about the z, y, x axes, respectively.

To visualize the quotient surface Sjs# 0 and the quotient action of s#js# 0 ,
we proceed as follows. The surface S is divided into eight pieces by the three
coordinate planes. We keep two adjacent pieces to form a surface S' (with
boundary) and discard the rest of S. On the right of Figure 6.10 is shown the
case in which S' consists of all points of S satisfying x .::;; 0 and y .::;; 0. The
restriction of the quotient map p: S ~ S/.91° to S' still maps onto Sjs# 0 ,
since the subsurface S' contains a point from every orbit of the action of .91°.
Moreover, the restriction of the covering projection p to the interior of S' is a
homeomorphism, because that part of S' contains only one point from each
orbit. On the other hand, the quotient map identifies pieces of the boundary of
S'. Rotation ac about the y axis identifies the top and bottom halves of the
circle in the yz-plane on the left and the semicircle in the yz-plane on the
right, as indicated by double arrows. Rotation be about the x axis identifies
the top and bottom halves of the semicircle in the xz-plane, as indicated by
single arrows.
Thus the quotient map closes off the holes in S' to form a sphere for the
quotient surface S/.91°. The branch points of the covering p: S ~ Sjs# 0 are
indicated by heavy dots: r, s, t correspond to fixed points of rotation ac about
the y axis, u to fixed points of rotation ab about the z axis, and v to fixed
points of rotation be about the x axis. The sphere Sjs# 0 has an equator given
by the arcs rs (semicircles identified), st (half of a dividing circle for reflection
in the xy-plane), tu (quarter-circles identified), uv (quarter-circles identified),
and vr (half of a dividing circle for reflection in the xy-plane). The action of
the quotient group s#js# 0 on the quotient surface Sjs# 0 is simply reflection
in this equator.

6.2.4. Alternative Cayley Graphs Revisited

As we have already indicated at the beginning of this section, reflections
necessitate the use of alternative Cayley graphs when studying symmetric
imbeddings. The difficulty with alternative Cayley graphs is that the natural
272 1be Genus of a Group

action of a group on an alternative Cayley graph for the group is not

necessarily free; if the given generating set contains an involution z, then left
multiplication by z will "flip" the edge leading from the identity vertex to the
vertex z, leaving the midpoint of the edge fixed. This implies that an
alternative Cayley graph for a generating set with elements of order 2 is not a
regular covering of a bouquet of circles.
There is, however, a characterization of alternative Cayley graphs that is
analogous to the corollary to Theorem 2.2.3. Although the action of a group
on an alternative Cayley graph for the group is not necessarily free, it is
fixed-point free on the vertex set. The converse is also true:

Theorem 6.2.3 (Sabidussi, 1958). Let the group .91 act on the graph G so that
the action on the vertex set is free and transitive. Then G is a Cayley graph for
.J/1, in which edges in the same orbit of the action correspond to the same
generator of .J/1. Moreover, any edge left invariant by a nonidentity element of .91
corresponds to an involution in the generating set.

Proof Choose any vertex v, and label it by the group identity 1. For each
a E.!#, label the vertex <l>a(v) by a. If <Pu(v) = <Pb(v), then <Pab_,(v) = v,
which implies that ab - 1 = 1, since .J/1 acts freely on the vertex set. Thus, there
are no conflicts in the labeling. Moreover, every vertex receives a label, since
the action of .91 is transitive on vertices.
Next, choose an edge e incident on the vertex v, and suppose that its other
endpoint u is labeled with the group element x. Then assign the label x to
every edge in the orbit of e. In addition, if <Px(e) * e, then for each a E.!#,
direct the edge <Pa( e) so that its initial vertex is <Pa< v ); if <PA e) = e, do not
assign any directions to the edges labeled x. Observe that the edge <Pa( e),
which is labeled x, leads from vertex <Pa(v), which is labeled a, to the vertex
<Pa(u) = <Pa(<PAv)) = <PaAv), which is labeled ax. The only way a conflict in
assigning directions could arise would be if there were a group element b such
that <Pu{e) = <Pb(e) and <Pa(v) * <Pb(v). But then <Pab_,(e) = e, which implies
that <Pab_, ( v) = u, which in tum implies that ab - 1 = x; and therefore, we
have <PA e) = e, which implies that no direction should be assigned, thereby
ending the conflict. After edges in the orbit of the edge e are labeled and
directed, choose another edge incident on v, and then label and direct the
edges in its orbit. Continue in this manner until all edges incident on v have
been labeled. Since the action of group .91 is transitive on vertices, it follows
that all edges of the graph G have been labeled. As has already been observed,
this labeling transforms G into a Cayley color graph for .J/1.
The given labeling certainly assigns the same generator to every edge in the
same orbit. We still must show that if e is labeled x and if <Pa(e) = e for some
nonidentity element a, then x is an involution. So suppose that the initial
endpoint of e is labeled b. We may infer that ab = bx and that a(bx) = b,
since <Pa switches the endpoints of e. It follows that b = a(bx) = (ab)x =
bx x, and therefore that x 2 = 1. 0
The Symmetric Genus 273

6.2.5. Group Actions and Imbeddings

The following theorem is the fundamental result relating group actions on
surfaces and imbeddings of Cayley graphs. The pseudofree case has already
been considered in Theorem 4.3.6.

Theorem 6.2.4 (Tucker, 1983). Let the finite group .511 act on an orientable
surface S. Then there is a Cayley graph for .511 cel/ularly imbedded in S so that the
natural action of .511 on the Cayley graph extends to the given action of .511 on the
surfaceS.

Proof If there are no reflections in the action of .511 on S, then, by

Theorem 6.2.2, the action is pseudofree, and according to Theorem 4.3.6, the
conclusion holds. Assume, therefore, that the action of .511 on S contains
reflections. Consider the action of the quotient group .Jilj.Ji1° on the quotient
surface Sj.Ji1°, where, as usual, .511° denotes the index-2, orientation-preserv-
ing subgroup of the action .511. The only nonidentity element of the action of
.Jilj.Ji1° on Sj.Ji1° must be a reflection, because it leaves fixed the image in
S/.511° of any circle left fixed by a reflection of the action of .511 on S. Let f be
this reflection, and let p: S ~ Sj.Ji1° be the natural projection. Wr! wish to
construct a 2-vertex graph G cellularly imbedded in S/.511°, such that G
contains no branch points of the covering map p and no fixed points off,
such that each face of the imbedding contains at most one branch point, and
such that /(G)= G. Then the preimage graph p- 1(G) is invariant under the
action of .511 on S, and the action of .511 on the vertex set of p- 1(G) is
transitive and free. It follows from Theorem 6.2.3 that p- 1(G) is the natural
action of a group on a Cayley graph. The imbedding of p -l( G) in S is
cellular, since the imbedding of G in Sj.Ji1° is cellular and has at most one
branch point in each face.
To construct the imbedded graph, we first construct half of G and then
reflect it to get the other half. Let T be one-half of the reflection f. Suppose
that T is a surface of genus g with r boundary components corresponding to
the r dividing circles C1, ••• , C, of the reflection f. The case for g = 1 and
r = 2 is illustrated in Figure 6.11. Begin the construction of G by imbedding a
bouquet of 2g circles at a vertex v in the surface T so that the complement is
an open disk with r holes (see the left of Figure 6.11). Next, for each branch
point in the interior of T, imbed a loop at vertex v that encloses that branch
point but no others (see the middle of Figure 6.11). Finally, for each dividing

f(w)

w=f(w)

Figure 6.11. The imbedding of half of Gin T (black dots are branch points).
274 The Genus of a Group

circle C;, choose a finite set W; of points inC;, such that each component of
C; - W; contains exactly one branch point; if C; contains no branch points, let
W; consist of a single point in C;. Then, for each point w in W;, imbed an arc
in T running from v tow, and, if f(w) * w, then imbed another arc from v to
f(w) (see the right of Figure 6.11). Let G0 be the resulting collection of loops
and arcs in T, and let G = G0 U /( G0 ). Then G is a graph with two vertices v
and f(v) (the points in ware now just midpoints of edges). By construction, it
follows that f(G) = G and that the imbedding of G is cellular with at most
one branch point inside each face. D

Corollary. Let .91 be a finite group. Then the symmetric genus o(.91) is the
minimum genus over all surfaces on which .91 acts, and the strong symmetric
genus o 0 (.91) is the minimum genus over all surfaces on which .91 acts preserving
orientation. D

The imbedding G -+ S/.91° constructed in the proof of Theorem 6.2.4 is

more important to later sections of this chapter than the theorem itself is. Let
C(.91, X) be the Cayley graph p- 1(G). Then G is the quotient of C(.91, X) by
the index-2 subgroup .91°. This implies that the imbedding C(.91, X)-+ Sand
the covering p: S -+ S/.91° can be obtained by a voltage-graph construction.
Moreover, by Theorem 6.1.1, the voltages can be assigned in .91, if desired,
instead of d 0 : the loops in G correspond to elements of the generating set X
that are contained in .91°, while the arcs from v to f(v) correspond to
orientation-reversing elements of the generating set X. (Note that if the loop e
at vertex v corresponds to the generator x, then so does the reflected loop
f(e), because the edges of p- 1(e) and p- 1(/(e)) all belong to the same orbit
of the action of .91.) The voltage-graph construction can be used to read off
information about the presentation for .91 with generating set X. One example
suffices here.

Example 6.2.4. Let .91 be a finite group acting with reflections on the surface S
such that the quotient surface Sj.91° is the sphere and such that the covering
p: S-+ Sjd 0 has three branch points in S/.91°, one of which lies on the
dividing circle of the reflection f of the quotient action of .91j.91° on S j.91°. Since
f takes branch points to branch points, it follows that the remaining two branch
points must be paired by f. In particular, they have the same order. The
appearance of graph G is then as illustrated in Figure 6.12 The generator z
corresponding to the edge e through the dividing circle off must be an involution,
by Theorem 6.2.3, because at least one edge of its preimage p - 1( e) passes
through a circle left pointwise-fixed by a reflection of .91. Since the order of the
net voltage on each face equals the order of the branch point inside that face, the
group .91 can be presented
(x, z: z 2 = 1, xq = 1, (xzx- 1z )' = 1, ... ),
where r is the order of the branch point on the dividing circle and where q is the
order of the other paired branch points.
The Symmetric Genus 275

Figure 6.12. The imbedding G-+ Sj.JJ/ 0 with voltages in

the group JJI.

X
v u

y y

s
X

Figure 6.13. A quadrilateral in an imbedding of a Cayley graph C( JJI, X), where JJI is abelian.

6.2.6. Are Genus and Symmetric Genus the Same?

By definitions, y(.#) ~ o(d) for any group d. We shall show for most
abelian groups ..rat', the inequality is strict, that is, y( d) < o( d).
Let ..rat'= .2'10 X .2'10 X .2'10 , and let X= { x, y, z} be the canonical gener-
ating set. By Section 6.1, the Cayley graph C(.#, X) has a quadrilateral
imbedding, and the genus of this imbedding is y( d). We claim, however, that
no symmetric imbedding of C( ..rat', X) can have any quadrilaterals at all. The
only quadrilaterals inC(.#, X) correspond to commutators in the generators.
Suppose that the relator xyx- 1y- 1 bounds a face in a symmetric imbedding of
C(.#, X), as illustrated in Figure 6.13. Then the rotations at the vertices of
this quadrilateral must have the form (starting at the lower left vertex):
s. ...xy ...
t. .. . yx-1 .. .
u. . .. x-1Y-1 .. .
v. .. . y-1x .. .

Since the imbedding is symmetric, each of these rotations must be the same
or reversed. Since the reverse of ... xy . .. is ... yx . .. , the rotation at t is the
same as the rotation at s. By similar reasoning, the rotations at u and v must
be the same as those at t and u. From the information given at s, t, and u,
this rotation must be ... xyx- 1y- 1 • • • • The information given at vertex v
implies that x immediately follows y- 1 in this same rotation, but that is
impossible because z and z- 1 must appear somewhere in the rotation. We
276 The Genus of a Group

conclude that no symmetric imbedding of C(s#, X) can have a quadrilateral

xyx- 1y -1, and hence C( s#, X) has no symmetric, quadrilateral imbedding.
Since y( s#) is strictly less than the genus of any Cayley graph for s#
corresponding to any noncanonical generating set, it follows that
y(s#) < o(s#).
Of course, it should not be surprising that minimal-genus Cayley graph
imbeddings can be highly nonsymmetric. Jungerman and White's construc-
tions of quadrilateral imbeddings given in Section 6.1 are quite complicated
partly because so many different rotations are needed. Even the more easily
pictured tubing construction of Section 3.5, which suffices to obtain a quadri-
lateral imbedding for .!#= ~10 X ~10 X ~10 , is complicated if viewed in
terms of rotations; in fact, eight different rotations are used (Exercise 15; see
also Exercise 6.1.1).
The inequality o(s#):;;; o 0 (s#) can also be strict. In particular, the example
with .!#= ~2 X ~4 at the beginning of this section satisfies o(s#) = 0 and
o 0 (s#) = 1 (Exercise 14).

6.2.7. Euclidean Space Groups and the Torus

Group actions on a surface do not simply appear from nowhere; they arise
from some geometric structure of the surface. To understand a particular
group action, one often must search for some geometry that the group action
respects. Perhaps the most revealing example is the relationship between the
torus and plane Euclidean geometry.

Example 6.2.5. Consider the pattern of bricks shown in Figure 6.14, which the
reader has probably seen in a courtyard or patio. Imagine the pattern extending
infinitely in all directions in the plane. What rigid motions, that is, Euclidean
isometries of the plane, take this pattern to itself? First, one can simply
"translate" the whole pattern, so that, for instance, the set of four outlined
bricks around point u displaces those around point v. Second, one can "rotate"
the pattern 90° about the point u (note that this rotation takes the entire global
pattern onto itself, not just the four local bricks). Third, one can "reflect" the
pattern about a horizontal or vertical line such as k or I. Finally, one can
translate half the distance from u to v, at the same time reflecting about the line
through u and v (this last type of motion is called a "glide"). The whole pattern
can be generated as the orbit under rigid motions of a single half-brick like that
shaded near the point w in Figure 6.14.

The collection of all Euclidean isometries leaving the pattern in Figure 6.14
invariant forms a group ~. called the "symmetry group" of the pattern. This
group has two important properties:

1. It contains translations in linearly independent (i.e., different) direc-

tions in the plane.
2. It is discrete; that is, for any point u, the orbit { c( u) 1 c E ~} has no
accumulation points.
The Symmetric Genus 277

k-- 1- - - k

u ~w

Figure 6.14. A pattern in the Euclidean plane.

Another way of stating property 2 is that for any point u, there is a minimum
distance d such that every isometry in the group C(l either moves u at least the
distance d or does not move u at all. Any group of isometries of the Euclidean
plane satisfying properties 1 and 2 is called a (two-dimensional) "Euclidean
space group" (or alternatively a Euclidean "crystallographic group"). Every
pattern in the Euclidean plane having translational symmetry in more than
one direction and having no infinitesimal symmetry (N .B.: a pattern of vertical
stripes has infinitesimally small vertical translations as symmetries) has a
Euclidean space group as its symmetry group. Conversely, every Euclidean
space group has patterns for which it is the symmetry group: one first finds a
"fundamental domain" like the half-brick in Figure 6.14, next draws any motif
on this domain, and then replicates the motif throughout the plane using the
given space group.
The history of space groups, from the art of the Alhambra and of M. C.
Escher to the crystallography of Federov (1891) to the mathematics of
Bieberbach (1911), is too rich to elaborate here. The reader is urged to read
Weyl (1952) or Schattschneider (1978) for an informal account. Lyndon (1985)
and Coxeter and Moser (1957) provide the mathematical details. The im-
portant facts we need about Euclidean space groups are summarized in the
following theorem.

Theorem 6.2.5. There are exactly 17 Euclidean space groups (up to group
isomorphism), 5 of which contain only orientation-preserving isometries. The
collection !T of all translations contained in a given Euclidean space group C(l is a
278 The Genus of a Group

normal subgroup of C(l, is isomorphic to !ZX :Z, and has as quotient C(ljY either
the trivial group or one of the cyclic groups fr2 , fr3 , fr4 , fr6 or one of the
dihedral groups !!)2 , !!)3 , !!)4, !!)6 •

Proof Omitted. D

The first statement of Theorem 6.2.5 is a lovely contribution to mathe-

matics. Although it may well have been known to artists for centuries, it was
not proved until Polya (1924) and Niggli (1924). Most of the facts about the
subgroup Y are easily verified. The composition of two translations is clearly
a translation, making Y a subgroup. That Y is normal follows from the
classification of isometries as either translations, rotations, or glides (the
conjugate of a translation preserves orientation and has no fixed points and,
hence, must be another translation). That Y is isomorphic to :rx fr follows
from the obvious commutativity of translations and the discreteness of C(l (see
Exercise 17). That C(ljY must be on the given list of groups follows from the
"crystallographic restriction" that any rotation in a Euclidean space group
must have order 2, 3, 4, or 6 (see Exercise 18).
The present reason we are interested in Euclidean space groups is that they
give rise to group actions on the torus. For example, if we simply "roll up" the
pattern in Figure 6.14 in both the vertical and horizontal directions, identi-
fying in the process the top and bottom edges and the right and left edges of
Figure 6.14, we obtain a similar pattern on the torus. Let t be the vertical
translation from u to v, and let s be the horizontal translation from u to w.
Then the "rolling up" is accomplished by taking the quotient of the Euclidean
plane by the group..#' generated by s 3 and t 2 • In particular, the finite quotient
group C(/j..ff acts on the torus as the symmetry group of the rolled-up pattern.
The following theorem gives the general situation.

Theorem 6.2.6. Let C(l be a Euclidean space group and ..#' any normal subgroup
of finite index in C(l. Then the quotient group C(/j..ff acts on the torus or on the
sphere.

Proof Let Y be the subgroup of translations in C(l. Let C(l 0 be the

orientation-preserving subgroup of C(l, and let ..#' 0 = C(l 0 n ..#'. Since ..#' has
finite index in C(l, the normal subgroup ..#' 0 n Y must contain translations in
two directions. Thus the quotient surface S' = Ej% 0 n Y of the Euclidean
plane E by the action of ..#' 0 n Y is a torus (Exercise 17 shows that ..#' 0 n Y
is generated by two translations that roll up the plane in two directions). The
natural action of the quotient group ..ff 0 j..ff 0 n Y on S' has quotient
S = Ej% 0 . Since % 0 preserves orientation, it follows that the surface S is
orientable and that the projection S' ~ S is a regular branched covering. By
the Riemann-Hurwitz equation, we know that x(S') 5: x(S). Since S' is the
torus, it follows that S must be the torus or the sphere.
In order to complete the proof, we first observe that the group C(/j% 0 acts
naturally on S. Thus, if ..#' 0 = ..#', we already have the desired action.
The Symmetric Genus 279

Alternatively, if ..#' 0 * ..#', then ..#'~ 0 = ~. since the subgroup ~ 0 has index
2 in ~. By the fundamental isomorphism theorem for quotient groups, we
know that ~ 0/..#' 0 is isomorphic to ..#'~ 0/..ff. Since ~ 0/..#' 0 acts on Sand
..#'~ 0 = ~. it follows that ~/..#' acts on the torus or on the sphere. D

Corollary. If .J/1 is a finite quotient group of a Euclidean space group, then

a(.J/1).::;; 1.

One can also define "spherical space groups" using spherical geometry and
"hyperbolic space groups" using hyperbolic geometry. In both cases we
require the group to be discrete. For the sphere, which is a compact surface,
this automatically means the group is finite. For the hyperbolic plane, we
require the quotient space under the group action to be compact, in analogy to
condition 1. In either case, there is no subgroup analogous to the translation
subgroup !T for Euclidean space groups.

6.2.8. Triangle Groups

Triangle groups are a particularly important kind of space group, especially
for non-Euclidean geometry. We give here a sketchy account of the salient
features of these groups. For more details, pictures, history, and rigor, we
highly recommend Magnus (1974). For geometry, both Euclidean and non-
Euclidean, see Coxeter (1961) or Lyndon (1985).

Example 6.2.6. We begin with a "fundamental" isosceles right triangle in the

Euclidean plane, such as the one outlined in the pattern in Figure 6.15. Let x and
y be the reflections about the lines along the legs of the triangle, and let z be
reflection about the line along the hypotenuse. The group ~ of isometries
generated by x, y, and z flops the triangle all over the plane, creating the pattern
shown in Figure 6.15. Half the triangles in the pattern have the same orientation

Figure 6.15. The full (2, 4, 4) triangle group.

280 The Genus of a Group

as the fundamental triangle and half the opposite; the forward and backward
letter Fs tell which is which. The group ~ is the symmetry group of the pattern
and hence is a Euclidean space group.

This construction applies to certain other triangles. Since the triangles of

the pattern must fit together around each vertex in a circle of paired "forward"
and "backward" triangles, each angle a of the fundamental triangle must be
of the form 'IT jp, where p is an integer (so that 2a fits into 2'1T a whole number
of times). A "(p, q, r) triangle" is one with angles 'TTjp, 'TTjq, 'TTjr, where p, q,
and r are positive integers. Because the sum of the angles of a Euclidean
triangle is 'IT, the only Euclidean ( p, q, r) triangles are (2, 4, 4), (2, 3, 6), and
(3, 3, 3). There are, however, other kinds of geometry. In spherical geometry,
where lines are great circles, triangles have angle sum greater than 'IT. The
possible spherical (p, q, r) triangles are (1, r, r), (2, 2, r), for any r ~ 2, and
(2, 3, 3), (2, 3, 4), (2, 3, 5); although (1, r, r) is a degenerate triangle, it still
can be used. In hyperbolic geometry, which can be modeled by the interior
of the unit disk, where "lines" are circles orthogonal to the boundary and
"reflections" are inversions in those circles, triangles have angle sum
less than 'IT. There are hyperbolic ( p, q, r) triangles for any p, q, r such that
ljp + ljq + 1/r < 1.
Given any ( p, q, r) triangle, the group of isometries, either Euclidean or
spherical or hyperbolic, generated by reflections x, y, z in the three sides of
the triangle is called the "full (p, q, r) triangle group". The symmetry group
of Figure 6.15 is the full (2, 4, 4) triangle group. The number of elements of a
triangle group equals the number of copies of the original triangle in the
pattern, because there is a unique isometry in the group taking the fundamen-
tal triangle to any other triangle (isometries are determined by what they do to
a single triangle and no isometry in the group takes the fundamental triangle
to itself, other than the identity). Thus, Euclidean and hyperbolic triangle
groups are infinite, because the Euclidean and hyperbolic plane are un-
bounded, but spherical triangle groups are finite. In fact, since the area of a
triangle on the unit sphere is equal to the excess of its angle sum compared
with 'IT, the orders of the full triangle groups for the sphere are easily
computed. For example, the area of a (2, 3, 5) triangle on the unit sphere is
'TT/2 + 'TT/3 + 'TT/5- 'IT, or 'TT/30. Since the total area of the unit sphere is 4'1T,
the order of the full (2, 3, 5) triangle group must be 120.
A presentation in terms of generators and relations for the full triangle
group (p, q, r) is not difficult to obtain. The three generating reflections x, y,
and z clearly have order 2. The product of two reflections in lines meeting at
an angle a is a rotation by 2a about the intersection of the two lines. Thus xy,
yz, and xz are rotations by 2'1Tjp, 2'1Tjq, and 2'1Tjr and hence have orders p,
q, and r. So far, we have
The Symmetric Genus 281

These are in fact the only relators needed. We sketch a proof. First, note that
the dual graph to the triangular pattern for a ( p' q' r) triangle group re is a
Cayley graph for re, since re acts on the dual graph transitively, fixing no
vertex (i.e., fixing no triangle in the original pattern). Any relator in the
generators corresponds to a closed walk in this Cayley graph. Since the Cayley
graph is imbedded in the plane, every cycle, and hence very closed walk, can
be expressed in terms of the face boundary cycles, which simply correspond to
the relators (xy)P = (yz)q = (xz)' = 1. Thus, every relator can be expressed
in terms of the already given relators. (We are being vague about what we
mean by "expressed", but this can be made rigorous.)
Given a full (p, q, r) triangle group re, the subgroup ?A generated by the
rotations xy, yz, and xz at the three vertices of the fundamental triangle is
called the "ordinary" (p, q, r) triangle group. Since (xy)(yz) = xy 2z = xz,
this group is generated by u = xy and v = yz alone. It has the presentation

(u, v: uP= vq = (uv)' = 1)

Moreover, this subgroup has index 2 in the full triangle group [the map
f:re--+ ?!'2 given by f(x) = f(y) = f(z) = 1 has kernel ?A], and thus ?A
coincides with the orientation-preserving subgroup re 0 .
We summarize these facts in the following theorem.

Theorem 6.2.7. A full ( p, q, r) triangle group re has the presentation

( x, y, z : x 2 = y2 = z2 = 1, ( xy) P = ( yz) q = ( xz )' = 1)

Its orientation-preserving subgroup reo is the ordinary triangle group generated

by the rotations u = xy, v = yz and has the presentation

(u, v: uP= vq = (uv )' = 1)

If ( p, q, r) is (1, r, r ), (2, 2, r ), (2, 3, 3), (2, 3, 4), or (2, 3, 5), then the triangle is
spherical and the group re is finite of order 4j(ljp + ljq + 1/r- 1). If
( p, q, r) is (2, 3, 6), (2, 4, 4), or (3, 3, 3), then the triangle is Euclidean and re is
an infinite Euclidean space group. Otherwise, re is a hyperbolic full triangle
group.

An especially interesting application of triangle groups to graph imbeddings

can be found in the work of Jones and Singerman (1978) and Bryant and
Singerman (1985). By a clever combinatorial definition of graph imbedding or
"map", they show that every map is the quotient of part of some (2, q, r)
triangle pattern in the sphere, Euclidean plane, or hyperbolic plane. Map
automorphisms are then interpreted as quotients of subgroups of the ordinary
(2, q, r) triangle group in the orientable case and the full (2, q, r) triangle
group in the nonorientable case.
282 The Genus of a Group

6.2.9. Exercises
1. Prove that an imbedding of a Cayley graph C(~. X) is weakly symmet-
ric if and only if there is a subgroup !11 of index 2 in the group ~ such
that every vertex of the Cayley graph labeled by an element of !11 has the
same rotation and all other vertices have the reversed rotation.
2. Let S be a surface with r boundary components. Then x(S) is defined
to be x(S) - r, where S is the closed surface obtained by spanning each
boundary component with a disk. Find all orientable surfaces S with
boundary such that x(S) = -5.
3. Let S0 be one-half of a reflection of the surface S. Prove that x(S) =
2x(So)·
4. Find all reflections of the surface of genus 2.
5. Given simplexes CJ in R m and T in R n' a map f: CJ ~ T is said to be
"linear" iff takes vertices of a to vertices of -r, and iff is the restric-
tion to a of an affine map F: Rm ~ Rn [i.e. F(x) = L(x) + b, where
L: R m ~ R n is a linear transformation and b E R n]. Find a linear map
from the simplex a in R 2 having vertices (0,0), (1,0), (0, 1), to the
simplex -r in R 3 having vertices (1, 2, 1), (3, 5, 2), (0, 2, 8). Prove that a
linear map between simplexes is uniquely determined by what it does to
vertices.
6. Given simplicial complexes K and L, a continuous map f: IKI ~ILl is
"piecewise linear" if there are subdivisions K', L' of K, L such that f
restricted to each simplex of K' is linear ( K' is a subdivision of K if
IK'I = IKI and every vertex of K is a vertex of K'). Let K be the
simplicial complex in R 2 consisting of the 2-simplex with vertices (0, 0),
(0, 1), and (1, 0) (and all its faces). Find a piecewise-linear map
f: IKI ~ IKI that leaves fixed the two legs of the right triangle IKI and
takes the point (1/3, 2/3) to the point (2/3, Ij3).
7. The set of fixed points of a piecewise-linear homeomorphism can be quite
complicated. Let L be the two-dimensional simplicial complex in R 2
formed by four right triangles with vertices at (0, 0), ( ± 1, 0), (0, ± 1). Use
the function f of Exercise 6 to construct a piecewise-linear homeomor-
phism g: ILl~ ILl that leaves fixed all points of ILl on the x axis, on the
y axis, or in the second quadrant. Sketch ILl and its set of fixed points.
Does g have finite order?
8. Prove that if h: R 2 ~ R 2 is a linear transformation of finite order leaving
the x axis fixed, then either h is the identity or h is a reflection across the
x axis.
9. Let g be an orientation-preserving, piecewise-linear homeomorphism
g: ILl ~ ILl, where ILl is a surface. Use Exercise 8 to prove that if g has
finite order and leaves some 1-simplex of L fixed, then g is the identity.
10. Use Exercise 9 to prove Theorem 6.2.2 for piecewise-linear homeomor-
phisms.
Groups of Small Symmetric Genus 283

Figure 6.16. A quotient surface with branch points and

quotient reflection.

11. Prove that t¥a"b = t¥at/lt,, where t/1 defines the natural quotient action of
the quotient group dj!!l on the quotient surface Sj!!l.
12. Suppose that the group .s;l acts on a surfaceS with reflections so that the
quotient surface Sj.s;/ 0 is a surface of genus 2. Suppose that the quotient
action of dj.s;/ 0 on Sjd 0 is a reflection with a single dividing circle
and that the branch points of p: s--+ Sj.s;/ 0 have orders and locations
as given by Figure 6.16. Find a partial presentation for the group .s;l by
constructing a 2-vertex voltage graph as in the proof of Theorem 6.2.4.
13. Prove that a Cayley graph imbedding C( .s;l, X) --+ S is strongly symmet-
ric if and only if it is the derived imbedding of a 1-vertex imbedded
voltage graph.
14. Prove that a 0 Ul'2 X .2'4 ) = 1 (see also Theorem 6.3.1).
15. List all the rotations used in the quadrilateral imbedding given in Section
3.5 for the Cayley graph C( d, X), where .s;l = .2'10 X .2'10 X .2'10 and
X= { x, y, z }, a canonical generating set for d.
16. Give an example of an involution on the torus that is not a reflection.
17. Let .'T be a discrete group of translations of the Euclidean plane
containing translations in independent directions. Choose any point P.
Let Q be the closest point to P in the orbit of P, and let R be the closest
point to P in the orbit of P but not on the line through P and Q. Prove
that the translations by vectors PQ and PR generate .'T.
18. Let ct' be a Euclidean space group, and let r E ct' be a rotation about
some point P. Let Q be the closest point to P in the orbit of P. Prove
that if r is a rotation through an angle of less than '1T j3, then r( Q) is
closer to Q than Q is to P. Why does this contradict the choice of Q?
Use a similar argument to show that r is not a rotation of order 5 either.

63. GROUPS OF SMALL SYMMETRIC GENUS

A first project in studying the genus of groups is to identify all the groups of
genus 0, then maybe all the groups of genus 1. We shall see that the analogous
problem for symmetric genus happens to be a quite tractable problem, and
that preliminary consideration of the symmetric genus gives remarkably accu-
rate insight into the general case, whose derivation is otherwise rather lengthy.
Study of the symmetric genus is additionally rewarding, in that it also reveals
rich connections with Euclidean and non-Euclidean geometry. Moreover, the
moderately easy success in determining all the groups of symmetric genus 0 or
284 The Genus of a Group

1 encourages the study of groups of given symmetric genus greater than 1,

where we encounter Hurwitz's theorem that establishes a bound on the order
of any group acting on a surface of given genus.
The geometric group actions introduced at the end of the previous section
provide an abundance of groups acting on the torus and the sphere. An
optimist might hope that all actions arise geometrically, that every finite group
acting on the sphere is isomorphic to a spherical space group, and that every
finite group acting on the torus is a quotient of a Euclidean space group. The
main theorems of this section show that such optimism is justified. They also
classify all such groups in terms of lists of partial presentations.

6.3.1. lbe Riemann-Hurwitz Equation Revisited

The main tool used to study symmetric genus is the Riemann-Hurwitz
equation, which has appeared here as Theorem 4.2.3. Because the equation
was first introduced in the context of general (not necessarily regular) branched
coverings, it will be helpful to restate the equation in terms of group actions
and regular branched coverings.
Let the group .511 act pseudofreely on the surfaceS, and let p: S -+ S/.511 be
the associated regular branched covering map. Let Y c Sj.Jil be the branch
point set of p. Recall that, by Theorems 4.3.4 and 4.3.5, for any given branch
point y E Y, the order of every prebranch point in the covering surface S is
the same. It follows that the deficiency of y is given by the equation

def{ y) = #.511- #.Jill ry

where rv is the order of the branch point y. The Riemann-Hurwitz equation

then assumes the second form

x(S) = #.Jil[x{S/.511)- I: (1- 1/ry)]

yEY

The key observation about the order of a branch point is that if the regular
branched covering p: S -+ Sj.Jil corresponds to an imbedded voltage graph,
then the order of each branch point is the order of the net voltage on the face
containing that branch point.

Example 6.2.4 Revisited. Suppose that beyond the original circumstances, we

further specify that x(S) = 0. Then, since x(S/.511°) = 2, the second form of
the Riemann-Hurwitz equation implies that

0 = #.511°{2- (1- 1/q)- {1- 1jq)- {1- 1/r))

Groups of Small Symmetric Genus 285

We conclude that 2/q + ljr = 1. It is easily verified by simple arithmetic that

the only possibilities are q = 4, r = 2, and q = 3, r = 3. It follows that there are
only two classes of groups that act on the torus in the manner described in the
example itself, namely, the groups with these presentations:

(x, z: z 2 = 1, x 4 = 1, (xzx- 1z ) 2 = 1, ... )

(x, z: z 2 = 1, x 3 = 1, (xzx- 1z ) 3 = 1, ... )

6.3.2. Strong Symmetric Genus 0

Because dihedral groups are encountered frequently in classifying the groups
that act on the sphere, a word or two about presentations for dihedral groups
is in order. The dihedral group f)n is defined classically to be the group of
symmetries of a regular n-sided polygon. Its order is 2n, and it has two
obvious presentations. If the generator x is rotation of the n-gon by 360° jn,
and if the generator y is any reflection of the n-gon, then we have the
presentation

f!Jn = (x, y: xn = y 2 = 1, yxyx = 1)

If x and y are reflections of the n-sided polygon through different diameters

making an angle of 180° jn with each other, then we have

In both presentations, it is easily verified that the permutations x and y

generate the group f!Jn and that the given relations hold. No more relations are
needed to define f!Jn because it can be shown that the order of the group
presented is the same as the order of f)n, namely, 2n (see Exercise 1). The
second presentation given is particularly useful since it characterizes the
dihedral groups as the groups generated by a pair of involutions.
We divide the classification of groups acting on the sphere into two cases.
The first case is the subclass of orientation-preserving actions, the groups with
strong symmetric genus 0.

Theorem 6.3.1. For a finite group .J/1, the strong symmetric genus a 0 ( .J/1) is 0
if and only if either the group .J/1 is isomorphic to one of the cyclic groups fl'n or
to one of the dihedral groups f)n, or else the group .J/1 has one of the following
presentations and the given order:

(a) (x, y: x 3 = y 2 = (xy ) 3 = 1, ... ), #.J/1= 12

(b) (x, y: x 3 = y 2 = (xy ) 4 = 1, ... ), #.Jil= 24
(c) (x, y: x 3 = y 2 = (xy ) 5 = 1, ... ), #.Jil= 60
286 11te Genus of a Group

Proof If a 0 ( ..l11) = 0, then the group J1l acts on the sphere S preserving
orientation. By the corollary to Theorem 6.2.2, the action is pseudofree. Let
p: S --+ S j..l11 be the associated regular branched covering, with branch set B.
From the second form of the Riemann-Hurwitz equation, we observe that

2 = #..l11[x(S/..l11)- L: (1- ljry)]

yEB

We infer from this equation that the characteristic x(S/..l11) must be

positive. Since the quotient surface Sj..l11 is orientable, the only possibility is
that Sj..l11 is itself the sphere and that x(S/..l11) = 2. Moreover, because
t
1 - 1/rv ~ for each branch point y, it follows that there are at most three
branch points. If there are fewer than two, then

from which it follows that J1l is the trivial group. We conclude that there are
either two or three branch points.
Suppose, on one hand, that there are two branch points. Then imbed the
1-vertex, 1-loop graph B 1 in the sphere Sj..l11 so that the loop separates the
branch points. By Theorem 4.3.5, the covering p: S --+ Sj..l11 is the derived
covering for some voltage assignments on the imbedded bouquet B 1 into the
group ..l11. Since B 1 has only one edge, the group J1l must be generated by a
single element, that is, the group J1l is cyclic.
Now suppose, on the other hand, that there are three branch points. Let
q 5: r 5: s be the orders of those branch points. The only values for the triple
(q, r, s) that leave the right-hand side of the Riemann-Hurwitz equation
positive are (2, 2, n ), (2, 3, 3), (2, 3, 4), (2, 3, 5). Imbed the bouquet B2 into the
quotient surface Sj..l11 as in Figure 6.17 (in which heavy dots represent the
branch points of the given orders). Again it follows from Theorem 4.3.5 that
the covering p: S --+ Sj..l11 is obtainable by assigning the two loops voltages
x, y E J11, as shown in Figure 6.17. Since the order of the net voltage on each
face equals the order of the enclosed branch point, we infer that the group J1l
may be presented

J11= (x, y: x' = yq = (xy)s = 1, ... )

If (q, r, s) = (2,2, n), then J1l is the dihedral group f!)n· The other three
possibilities for (q, r, s) yield the three presentations given in the theorem.

s
•
Figure 6.17. An imbedded voltage graph.
Groups of Small Symmetric Genus 287

Moreover, in each of these cases the order of .J/1 is determined exactly by the
Riemann-Hurwitz equation. For example, if (q, r, s) = (2, 3, 5), then

2 = #.J/1(2 - 1/2 - 2/3 - 4/5) = #.Jilj30

so #.J/1 = 60.
We have shown that if o 0 (.Jil) = 0, then .J/1 is as described in the theorem.
The converse, that if .J/1 is isomorphic to !!En or f!)n or has presentations (a),
(b), or (c), then .J/1 acts on the sphere, follows directly from the voltage-graph
constructions already given in this proof. D

The reader should recognize immediately the presentation (x, y: xq = y' =

(xy)' = 1, ... ) as an ordinary spherical (q, r, s) triangle group with perhaps
some extra relations. In particular, by Theorem 6.2.7, the given presentation
yields a group of the required order without any extra relations. [For those
uncomfortable with the geometry of the sphere, or maybe just to make the
reader appreciate geometric arguments, Exercises 2-4 show how to use alge-
braic methods to establish that presentations (a)-(c) are automatically finite
even without any extra relations.] Thus presentations (a)-(c) are precisely the
ordinary (2, 3, 3), (2, 3, 4), and (2, 3, 5) ordinary triangle groups.
To identify these spherical triangle groups as more familiar groups, all we
need to do is find a group having two generators of the right order whose
product has the right order. The alternating group .J/14 is generated by
x = (12 3), y = (1 2)(3 4). Since xy = (1)(2 4 3) has order 3, the group pre-
sented by (a) must be .J/14 • Similarly, presentation (b) is the symmetric group
~ generated by x = (12 3) and y = (3 4), and presentation (c) is the alternat-
ing group .J/15 generated by x = (1 2 3) and y = (2 4)(3 5).

Corollary (Maschke, 1896). The finite group .J/1 has strong symmetric genus 0
if and only if it is isomorphic to one of the groups, !!En, !!)n, .J/14 , ~, or .J/15 •
These groups correspond respectively to the ordinary (1, n, n ), (2, 2, n ), (2, 3, 3),
(2, 3, 4), (2, 3, 5) triangle groups.

We have credited Maschke with the classification of the groups of symmet-

ries and strong symmetric genus 0, because his 1896 paper implies such a
classification. However, Maschke was actually computing groups of genus (i.e.,
White genus) 0. In truth, the classification of spherical space groups is more a
matter of folklore, and the general topological problem of findin~ which finite
groups act on the sphere can be traced through a variety of papers, including
Kerekjarto (1919), Scherrer (1929), and Babai et al. (1972).
The case of orientation-reversing actions on the sphere is also handled by
the Riemann-Hurwitz equation. It is complicated, however, by reflections.

Theorem 6.3.2. Suppose that the finite group .J/1 acts on the sphere, reversing
orientation. The action has no reflections if and only if .J/1 is isomorphic to ~2 n
288 The Genus of a Group

for some n > 1. The action has reflections if and only if either d is isomorphic to
f) n , ~2 X ~n, ~2 X f) n or d has one of the following presentations and the
given order:

(a) (x, y: x 3 = y 2 = (xyx- 1y ) 2 = 1, ... ), #d= 24

(b) (x, y, z: x 2 = y 2 = z 2 = (xy ) 2 = (yz ) 3 = (xz ) 3 = 1, ... ), #d= 24

(c) (x, y, z: x 2 = y 2 = z 2 = (xy) 2 = (yz) 3 = (xzt = 1, ... ), #d= 48
(d) ( x, y, z: x 2 = y 2 = z 2 = ( xy ) 2 = ( yz) 3 = ( xz) 5 = 1, ... ) , #d = 120

Moreover, in (a), the orientation-preserving subgroup d 0 is generated by x and

yxy, and in (b)-(d), d 0 is generated by xy and yz. In particular, these
subgroups have index 2 in d.

Proof Suppose d acts without reflections on the sphere S. Then, by

Theorem 6.2.2, the group d acts pseudofreely with nonorientable quotient. By
the Riemann-Hurwitz equation, the quotient surface must still have positive
Euler characteristic. Hence, x(S/d) = 1 and the quotient surface Sjd is the
projective plane. The Riemann-Hurwitz equation now reads

and therefore we conclude there is at most one branch point, lest the right side
become negative. If the order of the branch point is n, then #d= 2n.
Moreover, with a single branch point in the projective plane, the covering
p: S --+ Sjd can be obtained from an imbedded voltage group with a single
type-1 loop. Therefore the group d is cyclic.
Suppose d acts with reflections. The orientation-preserving subgroup d 0 ,
however, still acts pseudofreely. As in Theorem 6.3.1, the quotient surface
SId 0 is again the sphere, and the regular covering p: S --+ S 1d 0 has two or
three branch points. Suppose there are two branch points. As in Theorem
6.3.1, they must have the same order, call it n. The quotient group djd 0 acts
on Sjd 0 as a single reflection. Either both branch points lie on the dividing
circle of this reflection or one lies on one side and one lies on the other side.
The regular covering p: S--+ Sjd 0 can then be obtained by a voltage graph
as illustrated on the left of Figure 6.18 for the first case, and the right of
Figure 6.18 for the second case. In both cases, edges crossing the reflection line
correspond to reflections in the action of d on the sphere S, and hence the
voltages on those edges have order 2. Presentations for the group d are given
by setting the order of the net voltage on each face equal to the order of the
enclosed branch point. In the first case, we get the presentation

(x, y: x 2 = y 2 = (xyr = 1, ... )

Groups of Small Symmetric Genus 289

I I
I

¢ + 0+-0
X I X

I I
I I
I I
Figure 6.18. Voltage graphs in the case of two branch points.

I
I
X I X

0-f-0 I
I
I
Figure 6.19. Voltage graphs in the case of three branch points.

and thus .91 is the dihedral group p)n. In the second case, we get

(x, y: xn = y 2 = yxyx- 1 = 1, ... )

and thus .91 is the group ~2 X ~n.

Suppose instead there are three branch points. Let their orders be q ~ r ~ s.
As in Theorem 6.3.1, possible values for ( q, r, s) are (2, 2, n ), (2, 3, 3), (2, 3, 4),
(2, 3, 5). There are again two ways that the branch points are located with
respect to the dividing circle of the reflection for the action of s#/.91° on
Sjs# 0 • All three branch points may be located on the dividing circle, as on the
left of Figure 6.19. Or, if two of the branch points have the same order,
those two may be located on either side of the dividing circle, with the third
branch point lying on the circle, as on the right of Figure 6.19. The covering
p: S-+ Sjs# 0 can then be obtained from an imbedded voltage graph, as in
Figure 6.19. If (q, r, s) is (2,2, n), we obtain the presentations for .91:

(X, y, Z: X2 = y 2 = Z2 = ( XJ ) 2 = ( yz) 2 = ( XZ r = 1, ... )

(x, y: x2 = y2 = (xy)2n = 1, ... )

The second presentation is, of course, the dihedral group p)n. In the first
presentation, since y commutes with both x and z, and since the subgroup
generated by x and z is the dihedral group f)n, the group .91 must be ~2 X f)n
290 The Genus of a Group

(unless the subgroups (y) and (x, z) have a nontrivial intersection, in which
case y is a redundant generator and .J/1 is simply the dihedral group Edn)·
For the other possible values of (q, r, s), we obtain the presentations
(a)-(d) given in the statement of the theorem. Presentations (a) and (b)
correspond to the two ways of locating branch points when (q, r, s) = (2, 3, 3),
whereas (c) and (d) correspond to the only possible branch-point locations
from (2, 3, 4) and (2, 3, 5). The subgroup that must be of index two in each
presentation is just the orientation-preserving subgroup, which is also the
isotropy group for the given voltage graph. The order of .J/1 is determined as
given in each case, because #.Jil= 2#.91° and #.91° is determined by
Theorem 6.3.1.
Since all the voltage graph constructions given in this proof depend only on
.J/1 having the desired presentation, the conditions .J/1 must satisfy are not only
necessary but also sufficient in order that .J/1 act on the sphere as described.
D

It can be shown (Exercise 6) that the subgroup (x, yxy) in presentation (a)
and the subgroup (xy, yz) in presentations (b)-(d) must have index at most
two for purely algebraic reasons. Since these subgroups are just the groups
(a)-(c) of Theorem 6.3.1, they are automatically finite without any extra
relations. Therefore the presentations (a)-(b) of Theorem 6.3.2 are also
automatically finite without any extra relations. This means again that only
one group is defined by each presentation with the given subgroup restriction.
In Exercise 5 the reader is asked to find the appropriate generating sets to
present ~2 X .J/14 by (a), ~ by (b), ~2 X~ by (c), and ~2 X .915 by (d).
Again, the reader should recognize the full triangle groups appearing in the
proof of the theorem and in presentations (b)-(d). On the other hand, the
pre sen ta tion

(x, y: xm = y 2 = [x, yr = 1, ... )

which yields either ~2 X ~m' for n = 1, or presentation (a), for m = n = 3,
appears to be new. In fact, it can be obtained from an ordinary (m, m, n)
triangle group by adjoining a reflection in the perpendicular bisector to the
side opposite the 'TT/n angle, or as the index-2 subgroup of the full (2, m,2n)
triangle group generated by rotation of order m at the 'TT jm vertex and
reflection in the side opposite the 'TT/m vertex (see Exercise 20). Call this
group a "hybrid ( m, m, n) triangle" group, since there appears to be no
standard terminology for this group. Thus we can summarize Theorem 6.3.2 as
follows. ·

Corollary (Maschke, 1896). o(J/1) = 0 and o 0 (.91) =F 0 if and only if .91 is

isomorphic to one of the groups fl'2 X fl'2 n, fl'2 X Edn, fl'2 X .9/4 , fl'2 X~. fl'2 X
.915 • These groups correspond, respectively, to the hybrid (2, 2, n ), the full
Groups of Small Symmetric Genus 291

(2, 2, n ), the hybrid (2, 3, 3), the full (2, 3, 4), and the full (2, 3, 5) triangle
groups.

6.3.3. Symmetric Genus 1

The classification of groups acting on the torus again uses the Riemann-
Hurwitz equation, but not surprisingly, there are now more combinations of
quotient surfaces, branch points, and reflections. The list of presentations
looks bewildering at first, but the reader should know what to expect. There
should be at least 17 partial presentations, none of which limits the order of
the group without the addition of extra relations. These correspond to quo-
tients of the 17 Euclidean space groups. If there are exactly 17 partial
presentations, we might have struck gold-a quick check with the list of
presentations given in Coxeter and Moser (1957) for the Euclidean space
groups will settle the issue. The reader should also find full and ordinary
(2, 3, 6), (2, 4, 4), and (3, 3, 3) triangle groups and hybrid (2, 4, 4) and (3, 3, 3)
triangle groups. One last hint before counting presentations in the statement
of Theorem 6.3.3: presentations with the same letter but different subscripts
are different presentations for the same underlying group. The only reason
both presentations are listed is for convenience in later computations.

Theorem 6.3.3. The group d acts on the torus and preserves orientation if and
only if d has one of the following presentations

(a) (x,y:[x,y] = 1, ... )

(b 1 ) (x, y, z: x 2 = y 2 = z 2 = (xyz) 2 = 1, ... )

(b 2) (x, y, z: y 2 = z 2 = (xy) 2 = (xz) 2 = 1, ... )

(c) (x,y:x 3 =y 3 =(xy) 3 =1, ... )
(d) (x,y:x 4 =y 2 =(xy) 4 =1, ... )
(e) (x,y:x 3 =y 2 =(xy) 6 =1, ... )

The group d acts on the torus and reverses orientation, but without reflec-
tions, if and only if d has one of the following presentations, with the given
subgroup !14 having index 2 in d:

(f 1 ) (x, y: xyx- 1y = 1, ... ), !14 = (x 2 , y)

(f 2) (x, y: x 2y 2 = 1, ... ), !14 = (x 2, xy)

(g 1) (x, y: y 2 = (x 2y ) 2 = 1, ... ), !14 = (x 2, y, xyx)

(g2) (x, y: (xy )2 = (x-Iy )2 = 1, ... ), !14 = (x2, xy, y2)
292 The Genus of a Group

The group .91 acts on the torus with reflections if and only if .91 has one of the
following presentations, with the given subgroup !14 having index 2 in .91:

(h) (x, y, z: y 2 = z 2 = [x, y] = [x, z] = 1, ... ), !14 = (x, yz)

(i) (x, y: y 2 = [x 2 , y] = 1, ... ), !14= (x 2 , xy)
(i) (x, y, z, w: x 2 = y 2 = z 2 = w2 = (xy) 2 = (yz) 2
=(zw) 2 =(wx) 2 =1, ... ), !JI=(xy,yz,zw)
(k) (x, y, z: x 2 = y 2 = z 2 = (yz) 2 = (xyxz) 2 = 1, ... ),
!14 = (x, yz, yxy)
{1 1 ) (x, y, z: x 2 = y 2 = z 2 = xyzyxz = 1, ... ), !14 = (x, y, zxz)
(1 2 ) (x, y, z: y 2 = z 2 = (xy ) 2 = [x, z] = 1, ... ), !14 = (x, y, zyz)
(m) (x, y, z: x 2 = y 2 = z 2 = (xy) 3 = (yz) 3 = (xz) 3 = 1, ... ),
!14 = (xy, yz)
(n) (x, y: x 3 = y 2 = [x, y ] 3 = 1, ... ), !14 = (x, yxy)
(o) (x, y, z: x 2 = y 2 = z 2 = (xy) 2 = (yz) 4 = (xzt = 1, ... ),
!14 = (xy, yz)
(p) (x, y: x 4 = y 2 = [x, y ] 2 = 1, ... ), !14 = (x, yxy)
(q) (x, y, z: x 2 = y 2 = z 2 = (xy) 2 = (yz) 3 = (xz) 6 = 1, ... ),
!14 = (xy, yz)

Proof Suppose .91 acts on the torus S preserving orientation. Then .91
acts pseudofreely. By the Riemann-Hurwitz equation

o = x(s) = #d(x(S/.91)- 1:(1- ljry))

It follows that x(S/.91) ~ 0 and that S/.91 is the torus or the sphere. If S/.91
is the torus, then there can be no branch points. Presentation (a) is thus
obtained from a voltage assignment for a bouquet of two circles cellularly
imbedded as usual in the torus Sjs#. If Sjs# is the sphere, then, by the
Riemann-Hurwitz equation, there are either four branch points all of order 2
or three branch points. Figure 6.20 illustrates two ways to imbed a bouquet of
three circles in the sphere so that each face of the imbedding encloses one
branch point of order two. The imbedded voltage graph on the left yields
presentation (b 1 ), and the one on the right, (b 2 ). If there are three branch
points of order q:;;; r:;;; s, then, from the Riemann-Hurwitz equation, the
possible values for ( q, r, s) are (3, 3, 3), (2, 4, 4), (2, 3, 6). Presentations (c), (d),
Groups of Small Symmetric Genus 293
z
X
• •

Figure 6.20. Two imbedded voltage graphs for the case of four branch points.

and (e), respectively, are obtained from these possibilities by an imbedded

voltage graph of a bouquet of two circles, as in Theorem 6.3.1.
In all of the remaining cases, there is a subgroup !14 that is required to be of
index 2. Given the action of the group d on the torus, the subgroup !14 is
necessarily of index 2 because it corresponds to the orientation-preserving
subgroup d 0 • Conversely, given a group d having one of the presentations
(f 1 )-(q), the index of the subgroup !14 must be 2 in order to construct the
necessary voltage graph. In cases (f 1 )-(g 2 ), the subgroup !14 must have index 2
so that Theorem 4.1.5 can be invoked to make the derived surface orientable;
in cases (h)-(q), the subgroup !14 must have index 2 because the necessary
voltage graph is a 2-vertex quotient graph. It should be noted that the
generators given for the subgroup !14 are chosen to be irredundant. For
example, from the voltage graph for presentation (1 1 ), one would expect to see
zyz as one of the generators for !14. However, zyz = zyxz · zxz, and since
xyzyxz = 1 implies that zyxz = yx, we have zyz = yx · zxz E ( x, y, zxz) = !14.
Suppose that d acts on the torus, reversing orientation but without
reflections. Then d acts pseudofreely, with nonorientable quotient surface
Sjd. Since x(S/d):;:: 0, from the Riemann-Hurwitz equation, Sjd is
either the Klein bottle or the projective plane. If S jd is the Klein bottle,
then, by the Riemann-Hurwitz equation, there are no branch points. Two
ways of imbedding a bouquet of two circles in the Klein bottle are illustrated
on the upper left and lower left of Figure 6.21. The corresponding voltage
graphs yield presentations (f 1 ) and (f 2 ), respectively. If Sjd is the projective
plane, then there must be two branch points of order 2, by the
Riemann-Hurwitz equation. Two ways of imbedding a bouquet of two circles
enclosing the branch points are illustrated on the upper and lower right of
Figure 6.21. The corresponding voltage graphs yield presentations (g 1 ) and
(g 2 ), respectively.
Suppose that d acts on the torus with reflections. The orientation-preserv-
ing subgroup d 0 acts pseudofreely. As with the presentations (a)-(e) already
considered, the quotient surface Sjd 0 is either the torus or the sphere.
Suppose it is the torus. By the Riemann-Hurwitz equation, the covering
p: S ~ Sjd 0 has no branch points. As usual, the quotient group djd 0 acts
on S jd 0 by a single reflection. There are two possibilities for this reflection.
One is the "obvious" reflection both of whose dividing circles are left point-
294 The Genus of a Group

I/

"IE
X '
\
I • • I
\ y I
'' ,. /

y X

y
Figure 6.21. Some imbedded voltage graphs.

u.,.....___:=:;::::~-... u
X :
i
X

z
Iy
i u--==:::;:::-....,. u
Figure 6.22. Two voltage graphs imbedded in the torus.

wise fixed; it is illustrated on the left in Figure 6.22. The other possible
reflection is the one discussed in Example 6.2.2 and is illustrated on the right
in Figure 6.22 (the dotted line down the middle of the rectangle is the dividing
circle left pointwise fixed by the reflection, whereas the left and right sides of
the rectangle form the dividing circle that is taken to itself with a half-rota-
tion). In both cases, the figure shows a 2-vertex voltage graph imbedded in the
torus, invariant under the given reflection, as in the proof of Theorem 6.2.4.
Voltages on edges crossing a dividing circle left pointwise fixed are involu-
tions. These two cases yield presentations (h) and (i).
Finally, suppose that the quotient surface S/.91° is the sphere. As in the
orientation-preserving case, the covering p: S --+ Sjs# 0 has either four branch
points of order 2 or else three branch points. With four branch points of order
2, there are three ways the points can be located with respect to the dividing
circle of the quotient reflection: all four on the circle, two on and two off, all
four off. Figure 6.23 illustrates the corresponding imbedded voltage graphs;
two possible voltage graphs are shown for the case of no branch points on the
dividing circle. These imbedded voltage graphs yield presentations (j), (k), (1 1)
and (1 2 ), respectively. The last five presentations, (m)-(q), correspond to the
case of three branch points, as in presentations (a)-(d) of Theorem 6.3.2, only
here the branch-point orders are (3, 3, 3), (2, 4, 4), or (2, 3, 6).
The converse, that any group .91 having one of the presentations (a)-(q)
with the given subgroup restrictions must act on the torus as described,
Groups of Small Symmetric Genus 295

X X

X X

y y
Figure 6.23. Imbedded voltage graphs for four branch points of order 2.

follows directly from the voltage-graph construction applied to the imbedded

voltage graphs already listed. 0

The number of different classes of groups given in Theorem 6.3.3 is 17. In

the following corollary, p1, p2, and so on are the standard crystallographic
notations for the Euclidean space groups.

Corollary. If the finite group .91 acts on the torus, then .91 is a quotient of a
Euclidean space group. The presentations a-q of Theorem 6.3.3 correspond
respectively to the groups p1, p2, p3, p4, p6, pg, pgg, pm, em, pmm, cmm,
pmg, p3m1, p31m, p4m, p4g, p6m.

Proof See Coxeter and Moser (1957) and Exercise 21. 0

6.3.4. The Geometry and Algebra of Groups of Symmetric Genus 1

Some further comments on Theorem 6.3.3 are necessary. The subgroup !14
corresponds to the quotient of the orientation-preserving subgroup of the
underlying Euclidean space group. Thus !14 has index at most 2, for geometric
reasons. Alternatively, one can also show this by algebraic methods. For
example, in presentation g 2 the subgroup !14 is normal, since xx 2x- 1 = x 2 ,
x(xy)x- 1 = x 2y 2 (xy)- 1 E !14, and xy 2 x- 1 = (xy)y 2 (xy)- 1 E !14; similarly
y!JI y - 1 = !14. Given that !14 is normal, it is clear that !14 has index at most 2,
296 The Genus of a Group

because the quotient group s#j!!J has the presentation (x, y: x 2 = y 2 =

xy = 1). The other presentations, (/1) - (q), are handled by similar arguments
(Exercises 6-8).
What if !!J does not have index 2 in .91, that is, what if !!J = .91? This can
happen if .91 is a space group quotient ~/.ff where the normal subgroup .ff
contains orientation reversing isometries. Since !!J is the quotient of ~ 0 , we
expect that .91 must then have one of the orientation-preserving presentations
(a)-(e). This can be derived algebraically. For example, in presentation (g 2 )
again, let u = xy, v = y- 1x, w = x- 1y- 1• Then x 2 = uv and y 2 = (uv)- 1,
from which it follows that u, v, w generate .91 = !!J. Moreover, from the given
relations (xy) 2 = (x- 1y) 2 = 1, it follows that u 2 = v 2 = w 2 = 1 and (uvw) 2
= (xy- 1 ) 2 = 1, thus yielding presentation (b 1 ). The other presentations,
( /1 ) - ( q ), are considered in Exercises 9-11.
Finally, by Theorem 6.2.5, we know that every one of the presentations of
Theorem 6.3.3 has a normal, 2-generator, abelian group !T corresponding to
the translation subgroup of the underlying Euclidean space group ~.such that
s#j!T is trivial or fl'2, fl'3, fl'4, fl'6, I!JJ2, I!JJ 3, I!lJ4, I!JJ6. Again this can be derived
by purely algebraic methods (see Exercises 12-15). The following theorem
summarizes this discussion.

Theorem 6.3.4. If the group .91 has one of the presentations (f 1 )-(q), then the
given subgroup !!J has index at most 2, and if !!J = .91, then .91 also has one of
the presentations (a)-(e). Thus, if .91 has any of the presentations (a)-(q), with
or without subgroup restrictions, then o( .91) :;;; 1. Finally, if .91 has one of the
presentations of Theorem 6.3.3, then .91 has a normal, 2-generator, abelian
subgroup !T such that s#j!T is trivial or one of the following groups:
fl'2(b, f, h, i) I!JJ2 (g, j, k, I)
fl'3( c) EP 3 (m, n)
fl'4(d) I!lJi o, p)
fl'6 (e) I!JJ6(q)

6.3.5. Hurwitz's Theorem

The problem of determining all groups of a given symmetric genus g > 1
would appear to be hopelessly complicated, in light of the long list of
presentations just given for groups of symmetric genus 1. In one way, however,
the situation for g > 1 is easier, or at least quite different, from the case
g = 1: the number of groups of given symmetric genus g > 1 is finite. The
following theorem, first proved by Hurwitz (1892), for automorphisms of a
Riemann surface, provides a bound upon the order of any group acting on a
given surface of genus g > 1.

Theorem 6.3.5 (Hurwitz). Suppose the group .91 acts on the surface of genus
g > 1. Then #.91:;;; 168(g- 1).
Groups of Small Symmetric Genus 297

Proof Let p: S--+ Sjd 0 be the regular covering associated with the
pseudofree action of the orientation-preserving subgroup d 0 • By the
Riemann-Hurwitz equation,

If x(Sjd 0 ) < 0, then x(Sjd 0 ) 5: -2. Therefore 2- 2g 5: #d 0 (-2), and

hence #d 5: 2#d 0 5: 2(g- 1). If x(Sjd 0 ) = 0, then there must be branch
points; otherwise 2 - 2g = #d 0 (0), contradicting the assumption that g > 1.
Since rY ~ 2 for any branch point y, [(1 - ljry) 5: ! . Thus 2 - 2g 5:
#d 0 ( - !), and hence #d 5: 2#d 0 5: 8(g - 1).
If x(Sjd 0 ) > 0, then x(S/d 0 ) = 2. Because 2- 2g is negative, so is
2 - E(1 - 1/ry). In particular, there must be more than two branch points. If
there are five or more branch points, then 2- E(1 - ljry) 5: !, and hence
#d 5: 8( g - 1) again. Suppose there are four branch points. The negative
value for 2 - E(1 - ljry) nearest to 0 is - i, obtained with three branch
points of order 2 and one of order 3. Thus 2- 2g 5: #d 0 ( - i), and hence
#d 5: 24( g - 1). Finally, suppose there are three branch points of orders
q 5: r 5: s. Then 2 - E(1 - 1/ry) = ljq + 1/r + 1/s - 1, which has - -;&
for its negative value nearest to 0, occurring when (q, r, s) = (2, 3, 7). Thus, in
this case, 2 - 2g 5: #d 0 ( - -;&),from which it follows that #d 5: 168(g- 1).
D

The value of 1/q + 1/r + ljs- 1 next nearest to 0 is - f4, occurring for
(q, r, s) = (2, 3, 8). Thus the proof of Theorem 6.3.5 leads to the following
corollaries.

Corollary. Suppose o 0 (d) > 1. Then #d> 48(o 0 (d)- 1) if and only if
#d= 84(o 0 (d)- 1) and d has the presentation

(x, y: x 2 = y3 = (xy) 7 = 1, ... )

Corollary. Suppose o(d) > 1. Then #d> 96(o(d)- 1) if and only if

#d= 168(o(d)- 1) and d has the presentation

(x, y, z: x 2 = y2 = z2 = 1, (xy ) 2 = (yz ) 3 = (xz f = 1, ... )

Conder (1980) has shown that the symmetric group Y, has a presentation
of the form given in the second corollary for all n > 167. Thus o(Y,) =
n!j168 + 1, n > 167. He has also shown that the alternating group dn has a
presentation of the form given in the first corollary for all n > 167. Thus
o(dn) = o 0 (dn) = n!j168 + 1, n > 167. Conder (1984) has also computed
the symmetric genus of all the other symmetric and alternating groups. Glover
and Sjerve (1986) have computed the strong symmetric genus of the simple
groups PSL(2, p ), where p is a prime.
298 The Genus of a Group

A careful inspection of the proof of Hurwitz's theorem leads to the

following refinement. We leave it to the reader to verify the details. First, we
define the group .91 to be a "proper" quotient of the full triangle group rl if
the image of r/ 0 is a proper subgroup of .91. That is, if
.91= (x, y, z: x 2 =y 2 = z 2 = 1,(xy)q = (yz)' = (xz)' = 1, ... )
then the subgroup generated by xy and yz is required to be a proper
subgroup. Similarly, the presentation
.91= (x, y: xq = y 2 = [x, y r = 1, ... >
makes .91 a "proper" quotient of the hybrid (q, q, s) triangle group if the
group generated by x and yxy is a proper subgroup of .91.

Theorem 6.3.6 (Refined Hurwitz Theorem). Suppose o(.91) > 1 and #.91>
24( o ( .91) - 1). Then one of the following holds:

a. .91 is an ordinary (q, r, s) triangle group quotient where (q, r, s) is either

(2, 4, 5) or (2, 3, s) with 7 5: s 5: 11.
b. .91 is either a proper quotient of a full (q, r, s) triangle group quotient,
where (q, r, s) is (2, r, s) with 3 5: r 5: 5 or (3,3, s) with s = 4,5, or.91
is a proper quotient of a hybrid (3, 3, 4), (3, 3, 5), or (5, 5, 2) triangle
group.

Moreover, if #.91> 48(o(.91)- 1), then .91 is either a quotient of the ordinary
(2, 3, 7) triangle group or a proper quotient of the full (q, r, s) triangle group,
where (q, r, s) is (2, 4, 5), or (2, 3, s) with 7 5: s 5: 11.

6.3.6. Exercises
1. Prove that the group (x, y: y 2 = xn = yxyx = 1) has order 2n by pro-
ving that the subgroup ( x) is normal of index 2. Prove that the group
(x,y: x 2 = y 2 = (xyy = 1) has order 2n by proving that the subgroup
( xy) is normal of index 2.
2. Prove that the group (x, y: x 3 = y 2 = (xy) 3 = 1) has order 12 by show-
ing that ?A, PAy, ?Ayx- 1 are all of the right cosets of the subgroup
!II= (x). (Hint: Use yxy = x- 1yx- 1 and yx- 1y = xyx.)
3. Prove that the group (x, y: x 4 = y 2 = (xy) 3 = 1) has order 24 by show-
ing that !II, !!Byx', !!Byx 2y, r = 0, 1, 2, 3, are all of the right cosets of the
subgroup !II= (x).
4. Prove that the group (x, y: x 5 = y 2 = (xy) 3 = 1) has order 60 by show-
ing that !II, !!Byx', ?Ayx 2yx', ?Ayx 2yx- 2y, r = 0, 1, 2, 3, 4, are all of the
right cosets of the subgroup ?A= (x). (For example, !!Byx- 2y is in this
list of cosets because !!Byx 2yx is in the list, and ?Ayx 2yx = ?Axyxxyx =
!!Byx-1yyx-1y = ?Ayx-2y.)
Groups of Small Symmetric Genus 299

5. Find generators for fr2 X .914 , 94, fr2 X 94, fr2 X .915 that give, respec-
tively, presentations (a)-(d) of Theorem 6.3.2.
6. Prove that the subgroup !14 = (xy, yz) has index at most 2 in the group
.91= (x, y, z: x 2 = y 2 = z 2 = 1, ... ). Do the same for the subgroup
!14 = (x, yxy) in the group .91= (x, y: y 2 = 1, ... ).
7. Prove for presentations (f 1 )-(g 1 ) of Theorem 6.3.3 that the given sub-
group !14 is of index at most 2.
8. Same as Exercise 7 for presentations (h)-(1 2 ) of Theorem 6.3.3.
9. Prove that if !14 = .91 in presentations (f 1 ), (f 2 ), (h), or (i) of Theorem
6.3.3, then .91 has presentation (a).
10. Prove that in Theorem 6.3.3 if !14 = .91 in presentations (g 1 ), (g 2 ), (j), (k),
(1 1 ), then .91 has presentation (b 1 ), and if !14 = .91 in presentation (1 2 ),
then .91 has presentation (b 2 ).
11. Show that presentations (m)-(q) of Theorem 6.3.3 correspond, respec-
tively, to presentations (c), (c), (d), (d), (e) when !14 = .91.
12. Verify that the normal subgroup !T of Theorem 6.3.4 is given by
(x- 1y, xy- 1 ) for presentation (c), (xyx, x 2 y) for (d), and
(xyx- 1y, x- 1yxy) for (e).
13. Verify that the normal subgroup !T of Theorem 6.3.4 is !14 itself in
presentations (f 1 ), (h), (i).
14. Verify that the normal subgroup !T of Theorem 6.3.4 is given by
(xyx- 1y-I, x- 1yxy) for presentation (g 2 ), (xz, yw) for (j), (xyxy, yxz)
for (k), and (x, yz) for (1 2 ).
15. Verify that the normal subgroup !T of Theorem 6.3.4 is given by
(yxyz, xyzy) for presentations (m) and (o), (x- 1yx- 1y, xyxy) for (n)
and (p), and ([yz, xy], [zy, xy]) for (q).
16. Determine o(.P) and o- 0 (2), where 2 = (i, j: i 4 = 1, i 2 = ) 2 ), the
quaternion group of order 8.
17. Find o( .91) and o 0 ( .91) for all groups of order #.91 < 16.
18. * Determine o 0 ( .91) for all groups .91 such that o( .91) = 0.
19. * Compute o( .91) and o 0 ( .91) as for many groups as you can in Coxeter
and Moser's (1957) list of all groups of order #.91 < 32.
20. Prove that the group obtained by adjoining to the ordinary (m, m, n)
triangle group the reflection in the perpendicular bisector to the side
opposite the wjn angle, has presentation

(x, y: xm = y 2 = [x, y r = 1, ... >

(Hint: x is the rotation of order m at one wjm vertex, and y is the
reflection in the bisector.)
300 The Genus of a Group

21. Prove the corollary to Theorem 6.3.3 by comparing the presentations

given in the theorem to the presentations in Table 3 of Coxeter and
Moser (1957).
* denotes a problem whose solution is lengthy.

6.4. GROUPS OF SMALL GENUS

Now that the lists of all groups of symmetric genus 0 or 1 have been compiled
in Section 6.3, it is time to ask about genus rather than symmetric genus. Is it
possible that every group of genus 0 or 1 appears in the lists of Section 6.3?
Since y(..l11).::;; a(J11) for any finite group ..l11, every group in these lists has
genus at most 1. On the other hand, from the example at the end of Section
6.2, the inequality y( ..l11) .::;; a( J11) is strict for most abelian groups, so one
would not expect these lists to be complete. In this section, perhaps surpris-
ingly, it will be shown that if y(..l11).::;; 1, then y(..l11) = a(J11), with three
exceptions. Thus, with the addition of three extra presentations, the lists of
Section 6.3 classify all groups of genus 0 or 1. At the same time, it is shown
that Hurwitz's theorem applies to genus as well as to symmetric genus, that is,
#J11.::;; 168[y(..l11)- 1] if y(..l11) > 1. It follows that the number of groups of a
given genus g > 1 is finite.
The classification of groups of genus 0 or 1 begins with Maschke (1896),
who shows that if y( ..l11) = 0, then J1l is isomorphic to one of the groups !!En,
!!) n, J114 , ~, or J115 or to the direct product of f!E2 with any of these groups,
from which it follows that a(..l11) = y(..l11) when y(..l11) = 0. Baker (1931)
extended some of Maschke's work to the torus, but it was Proulx (1977, 1978)
who did all the massive work needed to classify groups of genus 1. Tucker
(1983, 1984a) clarified the connections of Proulx's results to group actions on
surface and space groups. He used those connections to prove that three of the
four group presentations not fitting into Proulx's list of general classes of
Euclidean space group quotients are truly exceptional. The imbedded-Cayley-
graph version of Hurwitz's theorem was proved first by Tucker (1980) and was
later refined by Tucker (1984b). This last paper includes also a proof of
Proulx's classification. Finally, this last paper opens the way to the determina-
tion by Tucker (1984c) of the only group of genus 2.

6.4.1. An Example
The basic idea behind the results of this section is fairly simple. Suppose that
the group J1l has a Cayley graph imbedded in a surface whose genus is small
compared with #..l11 (in particular, genus 0 or 1). Then this imbedding must
have many faces. If it has many faces, there must be some faces of small size.
A face of small size gives a short closed walk in the Cayley graph, which in
turn corresponds to a short relator in the group ..l11. Upon inspection, the
group presentations appearing in the lists in Section 6.3 all seem to have few
generators and short relators. If the imbedding forces enough short relators in
Groups of Small Genus 301

.J/1, then .J/1 will have one of the listed presentations. As example will illustrate
the details of this argument in a particular case.

Example 6.4.1. Let X= { x, y} be an irredundant generating set for the group

.J/1, such that both x andy have order greater than 3. Suppose there is an
imbedding C( .J/1, X) --+ S such that #.Jil> - 20x(S). The number of vertices of
the graph C(.J/1, X) is #.J/1, and the number of edges is 2#.91. Consider the
faces of the imbedding. A boundary of a two-sided or three-sided face gives a
relation such as x 2 = 1 or xyx = 1; clearly any such relation contradicts
irredundancy or the assumed orders of x andy. Thus every face has four or more
sides. Let F' be the set of four-sided faces and F" the set of faces with more than
four sides. By the Euler polyhedral equation,
x(S) = #.J/1- #2.91+ #F' + #F"
Also, since the sum of the face sizes taken over all faces is twice the number of
edges, it follows that
4#F' + 5#F" .::;; 4#.Jil
Suppose that #F' = 0. Then
x(S) = - #.91+ #F' + #F".::;; - #.J/1+ 4#.Jilj5,
and hence #.J/1.::;; - 5x(S), contradicting our assumption that #.Jil> - 20x(S).
Thus there must be a four-sided face and, consequently, a relator of length 4. The
relation xyx- 1y- 1 = 1 gives presentation (a) of Theorem 6.3.3. The relation
xyx- 1y = 1 yields presentation (f 1 ) of Theorem 6.3.3, and x 2y 2 = 1 yields
presentation (f 2 ) of Theorem 6.3.3. The only other relators of length 4 not
contradicting irredundancy are (xy) 2 , (xy- 1 ) 2 , x\ y 4 • Suppose four four-sided
faces all meet at one vertex v. Assuming the only relators of length 4 are the four
just given, then there are tw9 cases depending on whether or not x and y edges
alternate in the rotation at vertex v. If they alternate, as depicted on the left of
Figure 6.24, then (xy) 2 and (xy- 1 ) 2 are relators yielding presentation (g 2 ) of
Theorem 6.3.3. If they do not alternate, as on the right of Figure 6.24, then
x 4 , y 4 and (xy ± 1 ) 2 are relators, which gives presentation (d) of Theorem 6.3.3
(replace y by y- 1 if necessary).
Suppose instead that there is no vertex at which four four-sided faces meet.
Then counting vertices by counting four-sided faces and multiplying by 4, one
obtains 4#F'.::;; 3#.91. Then since #F".::;; 4(#.Jil- #F')/5,
x(S) .::;; - #.91+ #F' + 4( #.J/1- #F')/5 = - #.Jil/5 + #F'/5
.::;; - #.Jil/5 + 3 #.s#/20
Therefore, #.J/1.::;; - 20x(S), again contradicting the assumption.

We conclude that if #.91> -20x(S), in particular if x(S) ~ 0, then the

group .J/1 has a presentation in the lists in Section 6.3. Of course, there are
many other cases to consider, depending on the number of elements in the
given generating set and their orders. It is the consideration of these cases that
occupies most of this section.
302 The Genus of a Group

y y

X V X y V X

y X

Figure 6.24. The rotation at vertex v.

6.4.2. A Face-Size Inequality

The determination of groups of low genus depends, as in Example 6.4.1, on
some delicate computations for the number of faces of various sizes. The
following inequality is the main tool in these computations.

Theorem 6.4.1. Let G ~ S be a cellular imbedding, where G is a regular graph

of valence d. Let f; be the number of i-sided faces. Then for any integer r > 1,

rd ) r-1
( - - r- d #V + rx(S) ~ L (r- i)f;
2 i=1

Proof Since "£'f'_ 1i/; = 2#£, it follows that

which can be rewritten

oo r-i
rL /; ~ 2#£ + L (r- i)/;
i=1 i=1

Multiplying the Euler polyhedral equation by r and rearranging, one obtains

00

r#E - r#V + rx ( S) = r L /,
i=l

Therefore,
r-1
r#E- r#V + rx(S) ~ 2#£ + L (r- i)/;
i= 1

The substitution #E = d#V12 now yields the desired inequality. D

We shall put this inequality to immediate use. The following corollaries

show that it is necessary to consider only Cayley graphs of valence 4 or less
when studying groups of low genus.
Groups of Small Genus 303

Corollary I. Let G ~ S be an imbedding of a regular simple graph of valence

d > 6. Then #V 5: -6x(S). In particular, there are only finitely many regular
simple graphs of given valence d > 6 and genus g > 1, and there are no regular
simple graphs of valence d > 6 and genus g 5: 1.

Proof By the inequality of Theorem 6.4.1, with r = 3,

(d/2- 3)#V + 3x(S) 5: 0
since / 1 = / 2 = 0. If d > 6, then d/2- 3 ~ 1/2 so #V 5: -6x(S),
as claimed. o

Corollary 2. Let C 1(.9.1, X)~ S be an imbedding of an alternative Cayley

graph of valence d > 4, where X is an irredundant generating set for the group
.9.1. Then #.9.15: -12x(S). In particular, if d(.Ji.l) denotes the minimal valence
over all Cayley graphs for .9.1, then there are only finitely many groups .9.1 with
given genus and given d( .9.1) > 4, and there are no groups .9.1 with d( .9.1) > 4
and y(.Ji.l) 5: 1.

Proof By the inequality of Theorem 6.4.1, with r = 4,

(d- 4)#.9.1+ 4x{s) 5:/3
since / 1 = / 2 = 0 by irredundancy. Also by irredundancy, the only three-sided
faces correspond to a relation of the form x 3 = 1, x E X. Thus there are at
most ld/2J three-sided faces meeting at any vertex. Thus 3/3 5: ld/2J#V. By
the previous corollary, we need consider only the cases d = 5 and d = 6. For
d = 6, / 3 5: #V, and hence
2#V + 4x{S) 5:/3 5: #V
which gives #V 5: -4x(S). Ford= 5, / 3 5: 2#Vj3 and hence
#V + 4x{S) 5: f#V
which gives #V 5: -12x(S). o
The second corollary leaves open the question of whether there are in-
finitely many regular graphs of valence d 5: 6 and given genus g > 1. In fact,
there are infinitely many such graphs (see Exercises 1-3). These corollaries
also raise other questions. Is the hypothesis of irredundancy necessary? The
following example illustrates the difficulties here.

Example 6.4.2 (N. Wormald). Let

.9.1= (x y z w· x 2 n = y 2 = z 2 = w 2 = 1 z = xn w = xy)
' ' ' 0 ' ' ,
n > 2
Clearly z and ware redundant, and .9.1 is actually the dihedral group of order 4n.
Consider the Cayley graph C(.Ji.l, X), where X= {x, y, z, w}. The subgraph of
C( .9.1, X) of edges colored x or z forms two disjoint 2n-cycles with "diameters"
(see left of Figure 6.25). In particular for n ~ 3, the Cayley graph C( .9.1, X)
contains disjoint copies of K 3 3 and therefore y(C(.Ji.l, X))> 1. On the other
304 Tile Genus of a Group

y z y z y
X X

X
<Dx
z y
X ®-;
z
X

X X X
y
X X
X X
y

<D ®
Figure 6.25. A subgraph of the Cayley graph C(d', X) and a partial imbedding of C(.JJI, X).

hand, if a pair of antipodal x-edges is removed from each x-cycle, the resulting
graph can be imbedded in the plane; in effect, the crossing of the z-diameters
can be undone with a twist at the expense of two x-edges crossing (see the
right of Figure 6.25). The four remaining x-edges can be reattached using
two handles (hitting the plane at spots marked 1 and 2 in Figure 6.25). Thus
y(C(.J<I, X))= 2.

This example provides infinitely many Cayley graphs of valence 5 and

genus 2. Most evidence, however, indicates that this phenomenon is limited to
genus 2. A reasonable conjecture is that the second corollary is true without
the restriction of irredundancy, if g > 1 is replaced by g > 2. L. Babai has
made the stronger conjecture that there are only finitely many graphs of given
genus g > 2 whose automorphism groups act transitively on vertices.

6.4.3. Statement of Main Theorem

As was noted at the beginning of this chapter, one needs to consider only
irredundant generating sets to compute the genus of a group. Since the
condition of irredundancy is so frequently used, call a Cayley graph C( .!<1, X)
"irredundant" if the generating set X is irredundant.

Theorem 6.4.2 (Proulx, 1978; Tucker, 1980, 1984a). Let C(.!<l, X) --+ S be an
imbedding of an irredundant Cayley graph in a surfaceS. If #.!<1> -84x(S),
then .!<1 has one of the presentations listed in Theorems 6.3.1-6.3.3 or one of
these presentations:

(a) (x, y: x 3 = y 3 = xyxy- 1x- 1y- 1 = 1), #.J<Jf= 24

(b) (x, y: x 3 = y 2 = (xy ) 3 (x- 1y ) 3 = 1), #.J<Jf= 48
(c) (x, y: x 3 = y 2 = (xyxyx- 1y ) 2 = 1), #.J<Jf= 48

To give some idea of the power of this theorem, we list three immediate
corollaries.
Groups of Small Genus 305

Corollary I. If y(d) > 1, then #d:5. 168(y(d)- 1). Thus there are only
finitely many groups of a given genus g > 1.

Proof If #d> 168(y(d)- 1), then an irredundant Cayley graph ford

can be imbedded in a surface S such that #d> - 84x(S). But then, by
Theorem 6.4.2, we have o( d) :5. 1 and hence y( d) :5. 1. D

Analogously to the inequality for symmetric genus in Hurwitz's theorem,

the inequality of Corollary 1 becomes an equality if and only if the group d is
a proper quotient of the full (2, 3, 7) triangle group. In particular, it follows
from the work of Conder (1980) that
y(Y,) = o(Y,) = 1 + n!j168
for all n > 167.

Corollary 2. If y(d) :5. 1, then either dis one of the three groups (a), (b), (c)
of Theorem 6.4.2 or else o(d) :5. 1. In particular, the lists of Section 6.3
together with (a),(b),(c) classify all groups of genus 0 or 1.

Proof If y(d) :5. 1, then an irredundant Cayley graph ford imbeds in a

surface S of Euler characteristic x(S) ~ 0. In particular, #d> - 84x(S).
D

Corollary 3. If the group d has a Cayley graph imbeddable in the Klein bottle,
then it has a Cayley graph imbeddable in the torus.

Proof The surface S in Theorem 6.4.2 is allowed to be nonorientable. D

This last corollary does not say that any Cayley graph imbeddable in the
Klein bottle is imbeddable in the torus. In fact, for redundant Cayley graphs
this is not true.

Example 6.4.2 Revisited. The two cycles with diameters formed by x-z edges
imbed in a pair of projective planes so that the x-cycles bound faces (view the
projective plane as a disk with antipodal points identified). If the two projective
planes are then joined by a tube to form a Klein bottle, the remaining y- and
w-edges can be added around the tube. Thus the Cayley graph C(d, X) imbeds
in the Klein bottle but not in the torus.

In general, Theorem 6.4.2 and its corollaries apply to the genus of a group,
not the genus of a Cayley graph. Part of the difficulty is that in the proof of
Theorem 6.4.2, the given generating set X occasionally must be changed to
obtain one of the presentations in the Section-6.3 lists; thus one obtains a
Cayley graph for d imbeddable in the torus, but it might not be the original
Cayley graph C( d, X). In fact, this difficulty can be overcome, but it involves
some intricate group theory that would take us far afield (Tucker, 1984b).
Thus, in Theorem 6.4.2, without extra hypotheses, it can be shown that the
306 The Genus of a Group

given Cayley graph C(.Jl/, X) imbeds in the torus or sphere whenever #.9/>
- 84x(S). It also follows that there are only finitely many irredundant Cayley
graphs of a given genus g > 1, that an irredundant Cayley graph imbeddable
in the Klein bottle is imbeddable in the torus, and that the lists of Section 6.3
together with presentations (a)-(c) of Theorem 6.4.2 classify all generating sets
and groups for irredundant Cayley graphs of genus 0 or 1.
The other difficulty in applying the results of this section to the genus of a
Cayley graph, rather than the genus of a group, is the restriction of irre-
dundancy. Example 6.4.2 shows that this is a serious problem. It is conjectured
that this example is, in effect, the only example that makes redundant Cayley
graphs different.
The proof of Theorem 6.4.2 occupies the remainder of this chapter. By the
second corollary to Theorem 6.4.1, only Cayley graphs of valence 3 or 4 need
be considered. The proof is broken into two main arguments: valence 4 and
valence 3. These are, in turn, broken into cases and subcases depending on the
nature of the given generating set X.
The proof of Theorem 6.4.2 involves much manipulation of generators and
relators. The study of groups given by generators and relators (usually called
combinatorial group theory) has many famous logical pitfalls, but here the
obvious intuitive approach is fine. Some agreement on terminology is, never-
theless, in order. A "word" in a generating set X for the group .9/ is a finite
sequence of elements of X and inverses of elements of X. The "length" of the
word is the number of terms in the sequence. A "relator" is a word the
product of whose terms is the identity; by abuse of language, we shall refer to
the product itself as a relator. Any conjugate or inverse of a relator is a relator,
as is any cyclic permutation. A relator is "reduced" if no cyclic permutation of
it can be shortened by using another relator to replace one subword by
another word. For example, if xyxy is a relator, then yzxzyx would not be a
reduced relator, since zxzyxy = zxzx- 1• A relator of length k is called a
k-relator.

6.4.4. Proof of Theorem 6.4.2: Valence d =4

The inequality of Theorem 6.4.1 will be used repeatedly. It is given again here
in the specific case d = 4 for easy reference:
r-1
(r- 4)#.9/+ rx(S)::;; L: (r- i)/;
i=l
Case Ia. X= { x, y }, both x and y have order greater than 3.
This case was handled in Example 6.4.1.

Case lb. X = { x, y }, x 3 = 1, and y has order greater than 3.

By the inequality of Theorem 6.4.1, with r = 5,
#.9/+ Sx(S)::;; 2/3 + /4
The only 3-relator, by irredundancy, is x 3 , and hence / 3 ::;; #.9//3. Thus f 4 > 0
if #.9/> -15x(S). The only reduced 4-relators that do not immediately give
Groups of Small Genus 307

a Section-6.3 presentation are y 4 or (xy ± 1 ) 2 • If y 4 is the only 4-relator, then

/4 ~ #J/1'/4, which implies #J/1'~ -60x(S). Therefore we assume (xy) 2 is a
relator (if both (xy) 2 and (xy- 1 ) 2 are relators, we have presentation (f 1 ) of
Theorem 6.3.3). We claim that if #Jii'> -42x(S), then y has order at most
6, in which case we have presentation (a) or (b) of Theorem 6.3.1 or (c) of
Theorem 6.3.3. Suppose every k-relator, k ~ 6, involves x. The total number
of sides of edges labeled x available for building faces is 2#Jii'. By irre-
dundancy, any relator using x once must use it twice. Thus by counting
occurrences of x, we obtain

Combining this inequality with Theorem 6.4.1, when r = 7, we get

Since (xy) 2 is the only 4-relator, there can be at most two faces of size 4
meeting at any vertex, and hence 4/4 ~ 2#Jii'. Since / 3 ~ #J/1'/3, it follows
that #J/1'> -42x(S). Thus if #J/1'> -42x(S), there must be a k-relator,
k ~ 6, not involving x, that is, y has order at most 6, as claimed.

Case Ic. X= {x, y}, x 3 = y3 = 1.

By Theorem 6.4.1, with r = 7,

The only 3-relators are x 3 and y 3 , so 3/3 ~ 2#Jii'. Therefore if #J/1'>

-21x(S), at least one of / 4 , / 5 , or / 6 is nonzero. Any 4-relator yields one of
the presentations of Section 6.3. Since x 2 = x- 1 and y 2 = y- 1, there is no
reduced 5-relator. The only 6-relators, up to cyclic permutation, interchange of
x and y, or replacement of x by x- 1 or y by y-I, are (xy)3, xyxyxy-I,
xyxyx- 1y-I, or xyx- 1yxy- 1• The relator (xy) 3 gives presentation (c) of
Theorem 6.3.3. The relator xyxyxy- 1 implies that (xy) 2 = yx- 1• Thus yx- 1 is
in the subgroup (xy). So is yx- 1 • xy = y 2 = y-I, and therefore so is x. Thus
J/1' is cyclic generated by xy, and obviously o(J/1') = 0. The relator xyxyx- 1y- 1
implies xyx = yxy- 1• Therefore xyx has order 3, which yields (xyx) 3 =
xyx- 1yx- 1yx = 1. Cyclically permuting this last relator gives us (x- 1y) 3 ; this
is just presentation (c) of Theorem 6.3.3 again when x is replaced by x- 1.
Finally, the relator xyx- 1yxy- 1 becomes xyxy- 1x- 1y- 1 after replacing y by
y- 1 and cyclically permuting; this gives the first exceptional presentation (a)
of this theorem. We conclude that if #J/1'> - 21x(S), then the group J/1' has
one of the desired presentations.

Case Ila. X = {x, y, z }, y 2 = z 2 = 1, x has order greater than 3.

By irredundancy, / 3 = 0. If there is no vertex at which four faces of size 4
meet, then 4/4 ~ 3#.91. Thus by Theorem 6.4.1, with r = 5,

#J/1'+ 5x(s) ~ 2/3 + / 4 ~ 3#Jii'/4

308 The Genus of a Group

and therefore #.9.1::::;, - 20x(S). On the other hand, if four faces of size 4 meet
at a vertex, then it is easily verified that, no matter what the rotation at that
vertex, the group .9.1 must have presentation (b 1 ), (h), or (1 2 ) of Theorem 6.3.3.

Case lib. X= {x, y, z}, y 2 = z 2 = 1, x 3 = 1.

As in Case lb, if #.9.1> -15x(S), there must be a 4-relator. If xyx ± 1y and
xzx ± 1z are both relators, then one of the presentations (b 1 ), (h), or (1 2 ) of
Theorem 6.3.3 is obtained, as in Case Ila. Therefore assume that xzx ± 1z is not
a relator. The only possible 4-relators are then xyx ± 1y and yzyz. Let lx be
the number of faces of size 4 corresponding to xyx ± 1y and fz the number of
faces of size 4 corresponding to yzyz. We claim that 13 + lx : : ;, #.9.1/2. It
would then follow from Theorem 6.4.1, with r = 6, that
2#.9.1+ 6x(S)::::;, 313 + 2(/x + lz) +Is ::::;.13 + #.9.1+ lz +Is
Since the cycles yzyz give a 2-factor of the Cayley graph C(.Ji.l, X), lz::::;, #.9.1/4.
Any reduced 5-relator contradicts irredundancy, so Is = 0. Since 13 ::::;, #.9.1/3,
the above inequality then gives us #.9.1::::;, -72x(S)/5::::;, -15x(S).
To establish the claim that 13 + lx : : ;, #.9.1/2, consider the subgraph of the
Cayley graph generated by x- and y-edges. With the relators x 3 and xyx ± 1y,
the subgroup ( x, y) is just 1E2 X 1E3 or the dihedral group f) 3. Thus a
component of the x-y subgraph is simply a 6-vertex prism with a triangular
top and bottom (corresponding to x 3) and three quadrilateral sides (corre-
sponding to xyx ± 1y ). As soon as a single z-edge is inserted at a vertex in the
prism, one of the five faces of the prism is destroyed (as a face). At least two of
the prism faces are destroyed when all six z-edges are inserted. Thus each
prism can contribute at most three faces to the imbedding. Since there are
#.9.1/6 prisms, it follows that 13 + lx : : ;, 3#.9.1/6, as claimed.

Case III. X= { x, y, z, w }, x 2 = y 2 = z 2 = w 2 = 1.
As in Case Ia, if #.9.1> - 20x(S) there must be a vertex at which four
faces of sizes 4 meet. By irredundancy, each reduced 4-relator involves two
generators. Thus there is some cyclic ordering of the generators such that
(xy) 2 ,(yz) 2 ,(zw) 2 , and (wx) 2 are all relators, which gives us presentation (j)
of Theorem 6.3.3.

6.4.5. Proof of Theorem 6.4.2: Valence d =3

There are two main cases to consider here: two generators, one of order 2, and
three generators, all of order 2. The form of Theorem 6.4.1 used here is
r-1
(r/2- 3)#.9.1+ rx(S)::::;, L: (r- i)/;
i=l
Case Ia. X= { x, y }, y 2 = 1, x has order greater than 5.
There can be no 3-relator, by irredundancy. The only possible reduced
4-relators are xyx ± 1y; this gives presentation (a) or (f 1 ) of Theorem 6.3.3 since
y = y- 1• The only reduced 5-relators are yxyx± 2 • The relator yxyx- 2 implies
yxy = x 2 • Thus, x = yx 2y = yxy · yxy = x 2 • x 2 , which means x has order 3,
Groups of Small Genus 309

contradicting the assumed order of x (in any case x 2 = x- 1 so that the relator
yxy = x 2 makes .5!1 the dihedral group ~3 ). A similar argument handles
yxyx 2 • We conclude that / 3 = f 4 = f 5 = 0. By Theorem 6.4.1, with r = 7,
#.Jil/2 + 7x ( s) 5;, / 6

Therefore if #.Jil> -14x(S), there must be a 6-relator.

Possible reduced 6-relators are, up to cyclic permutations and replacement
of x by x-I,
x 2yx± 2y, yxyx± 3 , xyxyx- 1y, (xy) 3 , x6
The relator x 2yx 2y gives presentation (g 1) and x 2yx- 2y gives (i) of Theorem
6.3.3. If yxyx- 3 is a relator, then yxyx = x 4 • Thus (yx) 2 commutes with x.
Since (yx) 2 also commutes with yx, it must commute with y. Let u = yx.
Then u and y generate .Jil, and u 2yu- 2y is a relator yielding presentation (g 1 )
of Theorem 6.3.3 again. The same argument applies to the relator yxyx 3 • The
relator xyxyx- 1y implies x = yxyx- 1y, which means x is conjugate toy and
hence x has order 2, a contradiction. Finally, the remaining relators, (xy) 3
and x 6 , together give us presentation (e) of Theorem 6.3.3 with generators
y and u = xy. If three six-sided faces meet at a single vertex, then both (xy) 3
and x 6 are relators, or one of the other relators is present; in either case, we
have a desired presentation. If there is no vertex at which three six-sided faces
meet, then 6f 6 5;, 2 #.Jil and, by Theorem 6.4.1 again, #.5!15;. - 42 x(S ).

Case lb. X= {x, y}, y 2 = 1, x 5 = 1.

As in Case Ia, / 3 = / 4 = 0. If / 6 = 0, then, by Theorem 6.4.1, with r = 7,
#.Jil/2 + 7x(S) 5;, 2/5 5;, 2#.5!1/5
and hence #.5!15;. -70x(S). If / 6 -:/= 0, then the previous analysis of 6-relators
leads to a presentation from Section 6.3.

Case Ic. X= {x, :Y}, y 2 = 1, x 4 = 1.

As in Case Ia, / 3 = f 5 = 0. We can assume / 6 = 0, by the analysis of
6-relators in Case Ia. Since x 3 = x- 1, there are no reduced 7-relators involving
two y 's. Thus the only possible reduced 7-relators are of the form x 2yxyx ± 1y.
If x 2yxyxy is a relator, then xyxyxy = x- 1; thus x E (xy), which makes .5!1
cyclic, and a(.Jil) = 0. If x 2yxyx- 1y is a relator, then xyxyx- 1 = x- 1y. Thus
x - 1y is conjugate to x, and hence x - 1y has order 4. This gives us presentation
(e) of Theorem 6.3.3, generated by x and u = x- 1y. We assume therefore that
f 7 = 0. Theorem 6.4.1, with r = 9, then yields
3#.5!1/2 + 9x(S) 5;. 5!4 + fs
Since / 4 5;, #.5!1/4, it follows that / 8 > 0 when #.Jil> - 36x(S).
Consider then the possible reduced 8-relators. Since x 3 =x-I, there are no
reduced 8-relators involving two y 's. The only 8-relators involving three y 's
are of the form yx 2yx 2yx ± 1, but then yx 2yx 2y = x ±1, which implies that
x ± 1 is conjugate to y and has order 2, a contradiction. Therefore we need only
consider 8-relators involving four y 's. Up to cyclic permutation and replace-
310 The Genus of a Group

ment of x by x- \ the possibilities are

(xy )4, (xy )3x-1y, (xy )2(x-1y )2, (xyx-1y )2
The relator (xy) 4 gives us presentation (e) of Theorem 6.3.3 again. The relator
(xy) 3x- 1y implies that xyx = yx- 1yxy, that is, xyx is conjugate to y and
therefore has order 2. Hence xyxxyx is a relator, yielding presentation (g 1) of
Theorem 6.3.3, as in Case Ia. The relator (xy) 2 (x- 1y) 2 implies that (xy) 2 =
(yx) 2 • Thus y(xy) 2y = y(yx) 2y = (xy) 2 , and this gives us presentation (i) of
Theorem 6.3.3, generated by y and u = xy. Finally, the relator (xyx- 1y) 2
together with x 4 yields presentation (p) of Theorem 6.3.3.

Case Id. X= {x, y}, y 2 = 1, x 3 = 1.

Since x 2 = x - 1, there are no reduced relators of odd length. By the analysis
of the previous cases, we can assume that / 4 = / 6 = / 8 = 0. Up to cyclic
permutation and replacement of x by x-I, the possible 10-relators are
(xy )s, (xy )4x-1y, (xy )3(x-1y )2, (xy )2x-1yxyx-1y
The relator (xy) 5 gives us presentation (c) of Theorem 6.3.1. Since xy · yx =
x 2 = x- \ the elements xy and yx together generate the group .91. Therefore
the relator (xy) 4x- 1y, which makes yx = (xy) 4 E (xy), implies that .91 is
cyclic. The relator (xy) 3 (x- 1y) 2 implies that xyxyxyx- 1yx- 1 = y, but then y
is conjugate to x, contradicting the assumed orders of x and y. The relator
(xy) 2x- 1yxyx- 1y implies that yxyx- 1yxyx- 1y =x-I, and again x- 1 is con-
jugate toy, a contradiction. Therefore we can assume that / 10 = 0. Then by
Theorem 6.4.1, with r = 14 (recall that /n = / 13 = 0, since there are no
odd-length relators),
4#.91+ 14x(S) 5: 11/3 + 2/12
Since / 3 5: #.91/3, it follows that / 12 * 0 when #.91> -42x(S).
Up to cyclic permutation and replacement of x by x- \ the possible
6-relators are
(xy ) 6 , (xy ) 5 x- 1y, (xy )4(x- 1y ) 2 , (xy )3x- 1yxyx- 1y, (xyxyx- 1y ) 2
(xy )3(x-1y )3, (xy )2x-1yxy(x-1y )2, (xyx-1y )3
The relator (xy) 6 gives us presentation (e) of Theorem 6.3.3. The relator
(xy) 5x- 1y implies that .91 is cyclic, as before, for 10-relators. The relator
(xy) 4 (x- 1y) 2 implies that yxyxyx- 1yx- 1y = x- 1yx- 1; that is, x- 1yx- 1 is
conjugate to y. Therefore (x- 1yx- 1 ) 2 is a relator, which in tum gives the
relator (xy) 2 , making .91 the dihedral group !i)3 • Similarly, the relator
(xy) 3x- 1yxyx- 1y implies that yxyx- 1yxyx- 1y = x- 1yx-I, from which it fol-
lows again that x- 1yx- 1 is conjugate to y and .91= f)3 • The relators
(xyxyx- 1y) 2 and (xy) 3(x- 1y) 3 give us, respectively, the exceptional presenta-
tions (c) and (b). The relator (xy) 2x- 1yxy(x- 1y) 2 can be rewritten
(xy) 2x- 1yx- 1 · x- 1y(x- 1y) 2, from which we obtain xyxyx- 1yx- 1 = (yx) 3 •
Then (yx) 3 is conjugate to y, which means that x has order 6 and that .91
Groups of Small Genus 311

has presentation (e) of Theorem 6.3.3 again. Finally, the relator (xyx- 1y) 3
gives us presentation (n) of Theorem 6.3.3.

Case Ila. X= { x, y, z }, x 2 = y 2 = z 2 = 1, xy, yz, xz all have order greater

than 2.
By irredundancy, there are no reduced relators of length 3 or 5, and the
only relators of length 4 are of the form (xy) 2 , contradicting the assumed
orders of xy, yz, and xz. Therefore / 3 = / 4 = / 5 = 0. As in Case Ia, if
#.91> -14x(S), then / 6 * 0. Up to cyclic permutation and interchanging the
names x, y, z, the possible 6-relators are
(xyz) 2 , xyzyxz, (xy) 3 , (yz) 3 , (xz) 3
Relators (xyz) 2 and xyzyxz yield, respectively, presentations (b 1 ) and (1 1 ) of
Theorem 6.3.3. The remaining three relators together give presentation (m) of
Theorem 6.3.3. Thus if three faces of size 6 meet at a vertex, then one of the
desired presentations is obtained. As in Case Ia, if #.91> -42x(S), then
three faces of size 6 meet at a vertex.

Case lib. X= { x, y, z }, x 2 = y 2 = z 2 = 1, (xy) 2 = 1, yz and xz have order

greater than 3.
Observe that if, say, (xy) 2 = (yz) 2 = 1, then y commutes with both x and
z, which means that .91= f!E2 X f)n and o(Sil) = 0. Therefore at most one of
xy, yz, and xz can have order 2. By the analysis in Case Ila, we can assume
that / 3 = / 5 = / 6 = 0. The only possible reduced 7-relator, up to cyclic permu-
tation or interchange of x and y, is zxzyzyx (recall that x and y commute,
since (xy) 2 = 1). Thus zyzy = xzx and xzxz = yzy, which imply (zy) 2 and
(xz) 2 have order 2. This gives us presentation (o) of Theorem 6.3.3. Therefore
we assume that / 7 = 0. As in Case Ic, / 8 * 0 if #.91> - 36x(S).
Up to cyclic permutation or interchange of x and y, the possible reduced
8-relators are
zxzxyzxy, (zxzy) 2 , (zx) 2 (zy) 2 , (xzt, (yz) 4
The relator zxzxyzxy implies that xzxyzx = zy, which means that zy is
conjugate to xy and has order 2, a contradiction of the assumed order of yz.
The relator (zxzy) 2 gives us presentation (k) of Theorem 6.3.3 (interchange x
and z). Given the relator (zx) 2 (zy) 2 , let u = zxz and v = xy. Then (uv) 2 =
(zxzxy) 2 = (zx) 2y(zx) 2y = (zx) 2y(yz) 2y = (zx) 2 (zy) 2 = 1, and (zuzv) 2 =
(xxy) 2 = 1. Thus we have presentation (k) of Theorem 6.3.3. again, generated
by u, v, and z. Finally, if both (xz) 4 and (yz) 4 are relators, then we have
presentation (o) of Theorem 6.3.3. Thus if two 8-sided faces meet at a vertex,
we have one of the desired presentations. On the other hand, if there is at most
one 8-sided face at any vertex, then by Theorem 6.4.1, with r = 9,
3 #.91/2 = 9x ( s) .s: 5/4 + / 8 .s: 5 #.91/4 + #.91/8
and hence #Sil.s: -72x(S).

Case He. X= {x, y, z}, x 2 = y 2 = z 2 = 1,(xy) 2 = (yz) 3 = 1.

312 The Genus of a Group

We claim that if#..#> -84x(S), then xz must have order 6 or less, and
hence ..# has presentation (b), (c), or (d) of Theorem 6.3.2 or (q) of Theorem
6.3.3. The claim is equivalent to showing that there is a relator of length 13 or
less not involving y. Indeed, suppose that every relator of length 13 or less
involved y. By the analysis of previous cases, we can assume that 13 =Is=
1 7 =Is= 0 and that the only 4-relator is (xy) 2 and only 6-relator is (yz) 3•
Suppose we could show that 19 = 0, that every 10-relator uses at least four y 's,
and that every 11-relator uses at least three y's (by irredundancy, any relator
using y must use it twice). There are only #..# sides of edges labeled y
available for building faces. Counting occurrences of y, we would have

214 + 316 + 4110 + 3ln + 2112 + 2113 5: #..#

Then by Theorem 6.4.1, with r = 14,

4#..# + 14x ( S) 5: 1014 + 816 + 4110 + 31n + 2112 + 113 5: 814 + 516 + #..#

Since 14 5: #..fll/4 and 16 5: #..#/6, it follows that #..#5: - 84x(S).

Therefore, to establish our claim, we must eliminate all 9-relators, all
10-relators using three y's or fewer, and all 11-relators using two y's. In the
following computation, it should be recalled that because (xy) 2 = (yz) 3 = 1,
x and y commute and yzy = zyz. There is no reduced 9-relator using only two
z's or five or more z's. The only reduced 9-relator using three z's is (zxy) 3• If
we let u = xy, then (uy) 2 , (yz) 3 , and (zu) 3 are relators, thus giving us
presentation (b) of Theorem 6.3.2, generated by u, y, and z. Any reduced
9-relator using four z 's must contain an occurrence of zyz, and hence can be
turned into a 9-relator with three z 's, since zyz = yzy. It is easily verified,
using the relations xy = yx and yzy = zyz, that the only 10-relator involving
two y's that needs to be considered is (yxzxz) 2 • Let u = yx; then (uzxz) 2 and
(ux) 2 are relators, yielding presentation (k) of Theorem 6.3.3, generated by u,
x, and z. In any reduced 10-relator involving three y 's, there must occur the
word yxzy, yzyx, or yzy. In each case, the relations xy = yx and yzy = zyz
can be used to make a 10-relator involving two y 's. Finally, by similar
arguments, the only 11-relators involving two y 's that one needs to consider
are of the form yzxzy(zx) 3 and yzxzxzy(zx) 2 • The first implies that (zx) 3 is
conjugate to x and that hence zx has order 6, giving us presentation (q) of
Theorem 6.3.3. The second implies that (zx) 2 is conjugate to x and that hence
zx has order 4, giving us presentation (c) of Theorem 6.3.2. D

6.4.6. Remarks about Theorem 6.4.2

We have not yet addressed the exceptional groups (a)-(c) of Theorem 6.4.2.
These groups do have genus 1; toroidal imbeddings are given in Figures
6.26-6.28. The groups are related; group (a) is contained in both (b) and (c) as
the normal index-2 subgroup corresponding to the normal closure of x,
Groups of Small Genus 313

X-
Y-..,-

1 2 3 4 5
Figure 6.26. A toroidal imbedding of group (a): (x, y: x 3 = y 3 = xyxy-lx-ly- 1 = 1). Vertices
with the same number are to be identified.

namely, (x, yxy) (see Exercises 4 and 5). The orders of the groups (a)-(c) are
then easily computed from the order of any one of them; the order of (c) is
computed in Exercise 6. None of these groups has symmetric genus 1;
Theorem 6.3.4 is used in Exercise 7 to show that the group (a) does not act on
the torus, and therefore (b) and (c) cannot either, since (a) is contained in both
(b) and (c). Also, as a perhaps curious coincidence, group (c) is the classical
group GL(2, 3) of 2 X 2 invertible matrices over the field with three elements,
which means that group (a) must be the subgroup SL(2, 3) of matrices with
determinant 1 (see Exercise 8). In summary, we have the following result.

Theorem 6.4.3 (Tucker, 1984a). The groups (a)-(c) of Theorem 6.4.2 are the
only groups .J/1 such that y( .J/1) = 1 and a( .J/1) > 1. In addition, group (a) is a
subgroup of both (b) and (c), and (c) is the group GL(2, 3).

As it stands, Theorem 6.4.2 does not quite classify groups of genus 0.

However, a reconsideration of the proof of Theorem 6.4.2 in Exercises 9-13
leads to the following theorem that, in fact, does imply a( .J/1) = 0 whenever
y( .J/1) = 0. An interesting consequence of this theorem is that, although
y(-2'2 X .J/15 ) = 0, there is no two-element generating set X for the group
~2 X .J/15 such that y( C( ~2 X .J/15 , X)) = 0 (see Exercise 16).
314 The Genus of a Group

11 12
Figure 6.27. A toroidal imbedding of group (b): (x, y: x 3 = y2 = (xy) 3 (x- 1y) 2 = 1). Vertices
with the same number are to be identified.

Theorem 6.4.4 (Maschke, 1896). If the irredundant Cayley graph C(.s#, X)

imbeds in a surface S such that x(S) > 0, then either .!#= ~3 X ~3 or
o(.s#) = 0. Moreover, y(~3 X ~3 ) = 1, although there is a projective-plane
imbedding of a Cayley graph for ~3 X ~3 .

The actual proof of Theorem 6.4.2 is probably as important as the final

result, at least for the computation of groups of low genus. Many of the
bounds on #.!# that appear in the proof can be improved (Exercises 17-20),
and additional information about the structure of the given group .91 can be
determined when #.!#is near one of these bounds. The computation of y(.9'5 )
in Exercises 21-23 gives one example of how the proof of Theorem 6.4.2 can
be refined in a specific case. As another example, we state without proof the
following Cayley-graph version of the refined Hurwitz theorem (Theorem
6.3.6) given at the end of the previous section.

Theorem 6.4.5 (Tucker, 1984b). Suppose that y( .!#) > 1 and that
#..r#> 24(y(..r#)- 1). Then either the group ..r# is as described in Theorem
6.3.6, with the word "proper" omitted throughout, or else ..r# has a presentation
Groups of Small Genus 315

10

v-~13
,,.;-~ p~,
:( }>--<{ p~,
::}>--<{ }>--<{ ;::
::( )>--<{ p.-~!0
14'--_A k_/. 7
15 ~
17 \._
18."---
19
A2
-;N' 1
y. 3
5

Figure 6.28. A toroidal imbedding of group (c): (x, y: x 3 = y2 =

x-
y ---

(xyxyx- 1y) 2 = 1). Vertices

with the same number are to be identified.

of the form

(x, y: x 3 = y 2 = 1, w = 1, ... )

where w is a word in x, x-I, and y of length 14, 16, 18, 20, or 22. If
#..# > 48( y( ..#) - 1), then ..# is as described in Theorem 6.3.6, without the
word "proper".

Proof Omitted. 0

Conder (1987) has considered all possible words w. For w = (xy) 4 (x- 1y)\
Conder (1986) has shown that #..#> 24(y(..#)- 1) and ..# is not described
in Theorem 6.3.6. Thus the exceptional presentations of this type really do
provide examples of conflicts between y(..#) and o(..#), even when the order
of ..# is large compared with y( ..#). The role of the word "proper" in all of
this is not understood. It is known that improper quotients have Cayley-graph
imbeddings in the surface of "correct" Euler characteristic, but that surface
must be nonorientable. The genus of such groups is unknown; the alternating
groups .Jiln, n > 167, are an example of such groups, due to Conder (1980).
316 The Genus of a Group

Nevertheless, Theorem 6.4.5 greatly restricts the possibilities for a group of

small genus greater than 1. The reader is encouraged to try to use it to
complete the proof of the following theorem.

lbeorem 6.4.6 (Tucker, 1984c). There is exactly one group of genus 2. It has
order 96, and it has the presentation

(x, y, z: x 2 = y2 = z2 = (xy) 2 = (yz) 3 = (xz) 8 = 1, [y, (xzt] = 1)

Proof That the given group has an imbedding in the surface of genus 2 (a
symmetric imbedding at that!) follows from the given presentation. The reader
can verify that the subgroup generated by y' = xy and x' = yz is the
exceptional group of genus 1 with presentation (c) in Theorem 6.4.2. In
particular, since this group does not act on the torus, neither does the given
group. Since the given group does not have order 24 or order 48, it is not
exceptional either. Thus the given group must have genus 2. An exhaustive
analysis of groups of order between 24 and 168, using the restrictions of
Theorem 6.4.5, shows that there are no other groups of genus 2. D

The group given in Theorem 6.4.6 probably does not look familiar. Coxeter
(1982) provides a full account of its origins in geometry and combinatorics; it
is, for example, the automorphism group of the generalized Petersen graph
G(8, 3) (see Frucht et al., 1971).
Finally, it is possible to adapt the methods used in the proof of Theorem
6.4.2 to give a partial answer to another fundamental question about the genus
of a group. Recall Babai's theorem that if !!4 is a subgroup of d, then
y(!!A).::;; y(d). One would expect that the same holds true for quotients: to
wit, that if 9. is a quotient of d, then y(9.).::;; y(d). This is clearly true for
symmetric and strong symmetric genus, by the natural action of the quotient
group on the quotient surface. Unfortunately, taking the quotient of a Cayley
graph can increase its genus. For instance, in Example 6.4.2 (due to Wormald),
the genus-2 Cayley graph for the presentation

(X, y, Z: X 2n = y 2 = Z2= 1, [X, y] = [X, Z] = 1, Z = X n)

is a quotient of the genus-1 Cayley graph for

(x, y, z: x 2 n = y2 = z2 = 1, [x, y] = [x, z] = 1).

Of course, the first presentation is redundant, and in fact the underlying group
~2 X ~2 n has genus 0, but this example does indicate that some subtle
questions about generating sets are involved, which is not the case for the
corresponding subgroup problem.
Groups of Small Genus 317

6.4.7. Exercises
1. Given graphs G and Hand edges d in G and e in H, define Gd • He to
be the graph obtained by deleting edges d and e and adding a pair of
edges joining the endpoints of d to the endpoints of e. Note that if G
and H are regular of valence n, then so is G d • He . Prove that if
y(G- d)= y(G) and y(H- e)= y(H), then y(Gd *He)= y(G) +
y(H).
2. Use Exercise 1 to construct infinitely many regular graphs of valence n,
3 5: n 5: 5, and genus g ~ 0 (Proulx, 1983).
3. Construct infinitely many regular graphs of valence 6 and genus g > 0
using Exercise 1 (Proulx, 1983). (Hint: First construct infinitely many
valence-6 triangulations of the torus.)
4. Prove that the subgroup (x, yxy) of group (b) of Theorem 6.4.2 is
isomorphic to group (a).
5. Prove that the subgroup (x, yxy) of group (c) of Theorem 6.4.2
is isomorphic to group (a) (let u = x, v = yx- 1y, and consider
uvuv - 1 u- 1v- 1 ).
6. In group (c) of Theorem 6.4.2, prove that (xy) 4y(xy) 4y = 1. Show then
that xy has order 8 and (xy) 4 is central. Conclude that group (c) has
order 48, by considering the quotient by the normal subgroup ((xy) 4 ).
7. Let .511 be the group (a) of Theorem 6.4.2 and suppose that a( .511) = 1.
Use Theorem 6.3.4 to prove that then .511 must have a normal subgroup
JV with quotient !Z3 and that both xy- 1 and y- 1x are in JV. Then
show that [xy-1, y- 1x] = 1 implies that x- 1y, has order 2 and that
therefore .511 has order 12, a contradiction.
8. Show that

x=(_~ -U
and

generate GL(2, 3) and satisfy x 3 = y 2 = (xyxyx- 1y ) 2 = 1. Then compute

the order of GL(2, 3) and compare with the order of group (c) to
conclude that (c) is isomorphic to GL(2, 3).
9. Prove that if x(S) > 0 in Theorem 6.4.2, then / 3 * 0 for valence 4 and
f 6 * 0 for valence 3, which implies that Cases Ia, Ila, III for valence 4
and Ia, Ila for valence 3 do not occur.
10. Show by counting x-edges, that if x(S) > 0 in Theorem 6.4.2, then y
has order at most 5 in Case lb for valence 4, and yz has order 2
in Case lib for valence 4. Conclude that if .511 has a valence-4,
318 The Genus of a Group

irredundant Cayley graph, imbeddable in a surfaceS, x(S) > 0, then .91

is one of the groups fr3 X fr3 , fr 1s, fr12 , fr2 X fr6 , fr2 X !7)3 or else
.!#= (x, y: x 3 = y' = (xy) 2 = 1), r = 3,4,5.
11. Prove that if x(S) > 0 in Theorem 6.4.2, then in Case lb and Ic for
valence 3, xy has order 3. (For Case Ic you will need to show that
/7 = 0.)
12. Prove that if x(S) > 0 in Theorem 6.4.2, then in Case Id for valence 3,
either (xy) 4 or [x, y] 2 or (xy)s is a relator.
13. Prove that if x(S) > 0 in Theorem 6.4.2, then Case lib for valence 3
cannot occur, and in Case lie, the element xz has order 3 or 4.
14. Use Theorem 6.4.1, with r = 7, to prove that y(fr3 X fr3 ) * 0.
15. Give a projective-plane imbedding of a Cayley graph for fr3 X fr3 .
16. Show that the group .91 = fr2 X .s#s has a two-element generating set.
Prove, however, that although o(.s#) = 0, y(C(.s#, X))> 0 for any two-
element generating set X. Thus for a given group.!#, it is not necessarily
true that y( C( .!#, X)) is minimized by choosing a generating set of
minimal size (Sit, 1980).
17. By considering possible 5-relators in Case lb, valence 4, of Theorem
6.4.2, prove that if#.!#> -12x(S) (rather than -60x(S)), then xy has
order 2. Thus, in this case, if #.!#> -42(S), then y(.s#).::;; 1.
18. By considering possible 7-relators in Case lb, valence 3, show that
y(.s#).::;; 1 if#.!#> -20x(S) (rather than -70x(S)).
19. In Case lib, valence 3, of Theorem 6.4.2, if yz has order 4, prove by
counting y-edges and by using Theorem 6.4.1, that if #.!# > - 36x( S),
then xz also has order 4 and y( .91) .::;; 1.
20. Use Exercises 17-19 to prove that if #.!#> -42x(S) in Theorem 6.4.2,
then either y(.s#).::;; 1 or .!#= (x, y, z: x 2 = y 2 = z 2 = 1, (xy) 2 =
(yz)3 = 1. .. ).
21. Prove that the symmetric group Sf's has no normal subgroup as in
Theorem 6.3.4 and hence that y( Sf's) > 1. Then show that therefore Sf's
has no presentation of the form (x, y: x' = y• = (xyf = 1, ... ) or
(x, y, z: x 2 = y 2 = z 2 = (xy)' = (yzy = (xz) 1, ••• ) such that {2, 3} c
{ r, s, t }. (Hint: Every element of Sf's has order at most 6.)
22. Suppose that y(SI's) .::;; 3. Use Exercises 17-20 to show that all valence-4
cases and Case lb, Id, lie for valence 3 of Theorem 6.4.2 do not occur.
Eliminate the other valence-3 cases using Theorem 6.4.1 and #Sf's ~
- 30x(S). (Difficult cases are Ic, where you must show / 8 = / 9 = 0, IIa,
where 7-relators must be considered, and lib, where you must show that
yz has order 4 and xz order 5, which is impossible for Sf's.)
23. Prove that y(SI's) = 4 (Proulx, 1981).
24. Let .91 be the K-metacyclic group (x, y: xP = yP- 1 = yxy- 1x' = 1), p
a prime, r * ± 1. Prove that y(.s#).::;; 1 if and only if p = 5 or 7 (Gross
and Lomonaco, 1980).
References

Additional works of interest are listed in the Supplementary Bibliography.

Numbers in angle brackets indicate the page citations of the references.

Mohammed Abu-Sbeih
-and T. D. Parsons (1983). Embeddings of bipartite graphs, J. Graph Theory 7, 325-334.
(210)

Jin Akiyama
- and M. Kano (1985). Factors and factorization of graphs-a survey, J. Graph Theory 9,
1-42. (This entire issue of the JOurnal is devoted to factorizations of graphs.) (79)

James W. Alexander
(1920). Note on Riemann spaces, Bull. Amer. Math. Soc. 26, 370-372. (180)
(1924). An example of a simply connected surface bounding a region which is not simply
connected, Proc. Nat. Acad. Sci. USA 10, 8-10. (100)

Seth R. Alpert
(1973). The genera of amalgamations of graphs, Trans. A mer. Math. Soc. 178, 1-39. (154)
(1975). Two-fold triple systems and graph imbeddings, J. Combin. Theory Ser. A 18, 101-107.
(105)
-and J. L. Gross (1976). Components of branched coverings of current graphs, J. Combin.
Theory Ser. B 20, 283-303. (87)
see also J. L. Gross
Dana Angluin
-and A. Gardiner (1981). Finite common coverings of pairs of regular graphs, J. Combin.
Theory Ser. B 30, 184-187. (80)
Kenneth Appel
- and W. Haken (1976). Every planar map is four-colorable, Bull. Amer. Math. Soc. 82,
771-712. (38, 215)
Dan Archdeacon
(1981). A Kuratowski theorem for the projective plane, J. Graph Theory 5, 243-246. (53)
(1984). Face colorings of embedded graphs, J. Graph Theory 8, 387-398. (46)
(1986). The orientable genus is nonadditive, J. Graph Theory 10, 385-401. (154)
-and P. Huneke (1985). On cubic graphs which are irreducible for non-orientable surfaces,
J. Combin. Theory Ser. B 39, 233-264. (53)

Louis Auslander
-with T. A. Brown and J. W. T. Youngs (1963). The imbedding of graphs in manifolds,
J. Math. and Mech. 12, 629-634. (140, 141)
319
320 References

Laszlo Babai
(1976). Embedding graphs in Cayley graphs, Probl. Comb. Theorie des Graphes, Proc. Conf.
Paris-Orsay (J.-C. Bermond et al., eds.), CNRS Paris, 13-15. (17)
(1977). Some applications of graph contractions, J. Graph Theory I, 125-130. (92ft')
- with W. Imrich and L. Lovasz (1972). Finite homeomorphism groups of the 2-sphere,
Colloq. Math. Janos Bolyai 8, Topics in Topology, Keszthely (Hungary). (287)
R. P. Baker
(1931). Cayley diagrams on the anchor ring, Amer. J. Math. 53, 645-669. (300)
Hyman Bass
see J.-P. Serre
Joseph Battle
- with F. Harary, Y. Kodama, and J. W. T. Youngs (1962). Additivity of the genus of a
graph, Bull. Amer. Math. Soc. 68, 565-568. (152)
Lowell W. Beineke
seeS. Stahl
K. N. Bhattacharya
(1943). A note on two-fold triple systems, Sankhya 6, 313-314. (105)
Ludwig Bieberbach
(1911). Uber die Bewegungsgruppen der euklidische Raume I, Math. Ann. 70, 297-336; II, 72
(1912), 400-412. (277)
Norman L. Biggs
-with E. K. Lloyd and R. J. Wilson (1976). Graph Theory, I736-I936, Clarendon Press,
Oxford. (35, 79)
R H Bing
(1954). Locally tame sets are tame, Ann. of Math. 59, 145-158. (18)
Garrett Birkhoff
see S. MacLane
Donald W. Blackett
(1967). Elementary Topology: A Combinatorial and Algebraic Approach, Academic Press,
Orlando. (110)
Bela Bollobas
- and P. Catlin (1980). Hadwiger's conjecture is true for almost every graph, European J.
Combin. I, 195-199. (40)
W. W. Boone
(1954). Certain simple unsolvable problems of group theory, Nederl. Akad. Wetensch. Indag.
Math. I6, 231-237 and 492-497; I7, 252-256 and 571-577; I9, 22-27 and 227-232. (14)
Andre Bouchet
(1982a). Constructions of covering triangulations with folds, J. Graph Theory 6, 57-74.
(175, 210)
(1982b). Constructing a covering triangulation by means of a nowhere-zero dual flow,
J. Combin. Theory Ser. B 32, 316-325. (210)
(1983). A construction of a covering map with faces of even lengths, Ann. Discrete Math. I7,
119-130. (210)
References 321

Matthew G. Brin
-and C. C. Squier (1986). On the genus of Z3 X Z3 X Z3 , Europ. J. Combin., to appear.
(262)
Edward M. Brown
(1969). The Hauptvermutung for 3-complexes, Trans. A mer. Math. Soc. 144, 173-196. (18, 98)
T. A. Brown
see L. Auslander
Robin P. Bryant
-and D. Singerman (1985). Foundations of the theory of maps on surfaces with boundary,
Quart. J. Math. Oxford Ser. (2) 36, 17-41. (281)
W. Burnside
(1911). Theory of Groups of Finite Order (1st ed., 1897). Cambridge Univ. Press, Cambridge.
(14, 265)
James W. Cannon
(1984). The combinatorial structure of cocompact discrete hyperbolic groups, Geom. Dedicata
16, 123-148. (14)

Paul Catlin
(1979). HaJOS graph-coloring conjecture: Variations and counterexamples, J. Combin. Theory
Ser. B 26, 268-274. (41)
see also B. Bollobas
Augustin-Louis Cauchy
(1813). Recherches sur les polyedres-premier memoire, J. de /'Ecole Poly. 9 (Cah. 16), 68-86.
[Extract in Biggs, Lloyd, and Wilson (1976).] (27)
Arthur Cayley
(1878). The theory of groups: Graphical representations, A mer. J. Math. I, 174-176. (10)
Marston D. E. Conder
(1980). Generators of the alternating and symmetric groups, J. London Math. Soc. (2) 22,
75-86. (305, 315)
(1984). Some results on quotients of triangle groups, Bull. Austral. Math. Soc. 30, 73-90. (297)
(1986). A note on Cayley graphs, J. Combin. Theory Ser. B 40, 362-368. (315)
(1987). Three-relator quotients of the modular group, Quart. J. Math. Oxford, to appear. (315)
Harold Scott MacDonald Coxeter
(1961). Introduction to Geometry, Wiley, New York. (279)
(1982). The group of genus two, Rend. Sem. Mat. Brescia 7, 219-248. (316)
- and W. 0. J. Moser (1957). Generators and Relations for Discrete Groups (4th ed., 1980).
Springer-Verlag, Berlin and New York. (14, 277, 291, 295)
Richard W. Decker
-with H. H. Glover and J.P. Huneke (1981). The genus of the 2-amalgamations of graphs,
J. Graph Theory 5, 95-102. (154)
-with H. H. Glover and J.P. Huneke (1985). Computing the genus of the 2-amalgamations
of graphs, Combinatorica 5, 271-282. (154)
Max Dehn
(1911). Uber unendliche diskontinuierliche Gruppen, Math. Ann. 72, 413-421. (10, 14)
322 References

G. Demoucron
- - with Y. Malgrange and R. Pertuiset (1964). Graphs planaires: Reconnaissance et
construction de representations planaires topologiques, Rev. Fram;aise Recherche Operation-
neUe 8, 33-47. (52)
Gabriel A. Dirac
--and S. Schuster (1954). A theorem of Kuratowsk.i, Nederl. Akad. Wetensch. lndag. Math.
16, 343-348. (264)
Richard A. Duke
(1966). The genus, regional number, and Betti number of a graph, Canad. J. Math 18,
817-822. (134)
(1971). How is a graph's Betti number related to its genus?, A mer. Math. Monthly 78, 386-388.
(136)
W. Dyck
(1882). Gruppentheoretische Studien, Math. Ann. 20, 1-45. (265)
Jack R. Edmonds
(1960). A combinatorial representation for polyhedral surfaces (abstract), Notices A mer. Math.
Soc. 7, 646. (113, 117, 192)
(1965). On the surface duality of linear graphs, J. Res. Nat. Bur. Standards Sect. 8, 121-123.
(119, 135)
Samuel Eilenberg
(1934). Sur les transformations periodique de Ia surface de sphere, Fund. Math. 22, 28-41.
(269)
Arnold Emch
(1929). Triple and multiple systems, their geometric configurations and groups, Trans. Amer.
Math. Soc. 31, 25-42. (105)
Paul Erdos
--and S. Fajtlowicz (1981). On the conjecture of Hajos, Combinatorica 1, 141-143. (41)
Leonhard Euler
(1736). Solution problematis ad geometriam situs pertinentis, Comment. Acad. Sci. /. Petro-
politanae 8, 128-140. (34)
(1750). Letter to C. Goldbach [excerpted in Biggs, Lloyd, and Wilson (1976)). (27)
Cloyd L. Ezell
(1979). Observations on the construction of covers using permutation voltage assignments,
Discrete Math. 28, 7-20. (88)
Siemion Fajtlowicz
seeP. Erdos
Istvan Fary
(1948). On straight line representation of planar graphs, Acta Sci. Math. (Szeged) 11, 229-233.
(48)
E. S. Federov
(1891). Zapiski Mineralogischeskogo lmperatorskogo S. Petersburgskogo Obstrechestva (2) 28,
345-390. (277)
Ion S. Filotti
--with G. Miller and J. Reif (1979). On determining the genus of a graph in 0( v0 <G>) steps,
Proc. 11th Annual ACM Symp. Theory of Computing, 27-37. (53)
References 323

Ralph H. Fox
(1957). Covering spaces with Singularities, Algebraic Geometry and Topology: A Symposium in
Honor of S. Lefschetz, Princeton University Press, Princeton, 243-257. (175)
Philip Franklin
(1934). A six color problem, J. Math. Phys. 16, 363-369. (137, 218)
Roberto Frucht
(1938). Herstellung von Graphen mit vorgegebener abstrakten Gruppe, Compositio Math. 6,
239-250. (70)
- with J. E. Graver and M. Watkins (1971). The groups of generalized Petersen graphs,
Proc. Camb. Philos. Soc. 70, 211-218. (59, 316)
Merrick L. Furst
-with J. L. Gross and L. McGeoch (1985a). Finding a maximum-genus graph imbedding,
J. Assoc. Comp. Mach., to appear. (133, 146)
-with J. L. Gross and R. Statman (1985b). Genus distributions for two classes of graphs,
J. Combin. Theory Ser. B, to appear. (147)
see also J. L. Gross
A. Gardiner
see D. Angluin
Brian L. Garman
(1979). Voltage graph embeddings and the associated block designs, J. Graph Theory 2,
181-187. (105)
Henry H. Glover
- with P. Huneke and C. Wang (1979). 103 graphs that are irreducible for the projective
plane, J. Combin. Theory Ser. B 27, 332-370. (53)
- and D. Sjerve (1985). Representing PSL 2 (p) on a Riemann surface of least genus,
Enseignement Math. 31, 305-325. (297)
see also R. W. Decker
Chris Godsil
-and W. lmrich (1985). Embedding graphs in Cayley graphs, Graphs and Combinatorics, to
appear. (17)
Jack E. Graver
see R. Frucht
Hubert Brian Griffiths
(1976). Surfaces (2nd ed., 1981). Cambridge University Press, Cambridge. (110)
Jonathan L. Gross
(1974). Voltage graphs, Discrete Math. 9, 239-246. (57, 165, 176)
(1977). Every connected regular graph of even degree is a Schreier coset graph, J. Combin.
Theory Ser. B 22, 227-232. (78)
(1978). An infinite family of octahedral crossing numbers, J. Graph Theory 2, 171-178.
(209, 214)
-and S. R. Alpert (1973). Branched coverings of graph imbeddings, Bull. A mer. Math. Soc.
79, 942-945. (194)
- and S. R. Alpert (1974). The topological theory of current graphs, J. Combin. Theory Ser.
B 17,218-233. (117,165,194, 196, 205)
324 References

and M. Furst (1987). Hierarchy for imbedding-distribution invariants of a graph,

J. Graph Theory II, to appear. (147)
- and F. Harary (1980). Some problems in topological graph theory, J. Graph Theory 4,
253-263. (250)
- with B. Jackson, T. D. Parsons, and T. Pisanski (1982). Wrapped coverings and their
realizations, Congress. Numer. 35, 333-347. (207)
-and E. W. Klein (1986). A Kuratowski-type theorem for the average genus of a graph, in
preparation. (147)
- and S. J. Lomonaco (1980). A determination of the toroidal k-metacyclic groups,
J. Graph Theory 4, 165-172. (318)
-with D. P. Robbins and T. W. Tucker (1986). Genus distributions for bouquets of circles,
preprint. (147)
- and R. H. Rosen (1979). A linear-time planarity algorithm for 2-complexes, J. Assoc.
Comput. Mach. 20,611-617. (117)
-and R. H. Rosen (1981). A combinatorial characterization of planar 2-complexes, Colloq.
Math. 44, 241-247. (117)
- and T. W. Tucker (1974). Quotients of complete graphs: Revisiting the Heawood
map-coloring problem, Pacific J. Math. 55, 391-402. (57, 67, 230)
- and T. W. Tucker (1977). Generating all graph coverings by permutation voltage
assignments, Discrete Math. I8, 273-283. (74, 81, 84, 176, 177)
- and T. W. Tucker (1979a). Fast computations in voltage graph theory, Ann. New York
Acad. Sci. 3I9, 254-257. (91, 170)
- and T. W. Tucker (1979b). Local maxima in graded graphs of imbeddings, Ann. New
York A cad. Sci. 3I9, 254-257. (150)
see also S. R. Alpert and M. L. Furst
William Gustin
(1963). Orientable embedding of Cayley graphs, Bull. Amer. Math. Soc. 69, 272-275. (19311')
H. Hadwiger
(1943). Uber eine Klassification der Streckenkomplexe, Vierteljschr. Naturforsch. Ges. Zurich
88, 133-142. (40)
Gary Haggard
(1980). Loops in duals, A mer. Math. Monthly 87, 654-656. (46)
Wolfgang R. Haken
seeK. Appel
Haim Hanani
(1961). The existence and construction of balanced incomplete block designs, Ann. Math. Stat.
32, 361-386. (105)
Frank Harary
(1969). Graph Theory, Addison-Wesley, Reading, Massachusetts. (2)
- and R. H. Rosen (1976). On the planarity of 2-complexes, Colloq. Math. 36, 101-108.
(117)
see also J. Battle and J. L. Gross
Philip J. Heawood
(1890). Map-colour theorem, Quart. J. Math. 24, 332-338. (215tf)
Lothar Heffler
(1891). Uber das Problem der Nachbargebiete, Math. Ann. 38, 477-580. (113, 117, 137, 191,
217tf)
References 325
(1897). Uber Triplesysteme, Math. Ann. 49, 101-112. (105)
Hugh M. Hilden
(1976). Three-fold branched coverings of S 3 , Amer. J. Math. 98, 989-997. (180)
John E. Hopcroft
-and R. E. Trujan (1974). Efficient planarity testing, J. Assoc. Comput. Mach. 21, 271-283.
(52)
John Philip Huneke
see D. Archdeacon, R. W. Decker, and H. H. Glover
A. Hurwitz
(1892). Uber algebraische Gebilde mit eindeutigen Transformationen in sich, Math. Ann. 41,
403-442. (296)
Wilfried Imrich
see L. Babai and C. Godsil
Bradley W. Jackson
- with T. D. Parsons and T. Pisanski (1981). A duality theorem for graph embeddings,
J. Graph Theory 5, 55-77. (207, 210)
see also J. L. Gross and T. D. Parsons
David M. Jackson
(1987). Counting cycles in permutations by group characters, with an application to a
topological problem, Trans. Amer. Math. Soc. 299, 785-801. (147)
Alain Jacques
(1969). Constellations et Proprietes Algebrigues des Graphs, Ph.D. thesis, University of Paris.
(194)
Francoise Jaeger
(1976). Thesis, Universite de Grenoble. (145)
Gareth A. Jones
- and D. Singerman (1978). Theory of maps on orientable surfaces, Proc. London Math.
Soc. (3) 37, 273-307. (281)"
Mark Jungerman
(1974). Ph.D. thesis, University of California, Santa Cruz. (245)
(1975). The genus of the symmetric quadripartite graph, J. Combin. Theory Ser. B 19,
181-187. (195, 201)
(1978). The non-orientable genus of the n-cube, Pacific J. Math. 76, 443-451. (159)
- and A. T. White (1980). On the genus of finite abelian groups, European J. Combin. I,
243-251. (249ft')
Mikio Kano
see J. Akiyama
Richard M. Karp
(1972). Reducibility among combinatorial problems, in Complexity of Computer Computation
(R. Miller and J. Thatcher, eds.), Plenum, New York. (35)
Abraham Karrass
see W. Magnus
A. B. Kempe
(1879). On the geographical problem of four colors, Amer. J. Math. 2, 193-204. (219)
326 References

B. Kerekjarto
(1919). Uber die endlichen topologische Gruppen der Kugelflache, Amsterdam Akad. Verst. 28,
555-556. (287)
(1921). Uber die periodische Transformationen der Kreischeibe und der Kugelflache, Math.
Ann. 80, 36-38. (269)
E. Ward Klein
see J. L. Gross
Felix Klein
(1911). Elementar-Mathematik vom hoheren Standpunkte aus, Vol. 2, Teubner, Leipzig. (26)
Donald Erwin Knuth
(1969). The Art of Computer Programming, Vol. 2, Addison-Wesley, Reading, Massachusetts.
(161)
Yukihiro Kodama
see J. Battle
Denes Konig
(1916). Uber graphen und ihre Anwendung auf Determinantheorie und Mengenlehre, Math.
Ann. 77, 453-465. (80)
S. Kundu
(1974). Bounds on the number of disjoint spanning trees, J. Combin. Theory Ser. B 17,
199-203. (145)
Kasimierz Kuratowski
(1930). Sur le probleme des courbes gauches en topologie, Fund. Math. 15, 271-283. (28)
Henry M. Levinson
(1970). On the genera of graphs of group presentations, Ann. New York Acad. Sci. 175,
277-284. (249, 264)
-and B. Maskit (1975). Special embeddings of Cayley diagrams, J. Combin. Theory Ser. B
18, 12-17. (265)
S.-A.-J. Lhuilier
(1811). Memoire sur Ia polyedrometrie, Ann. de Mathematiques 3, 169-189. [Extract in Biggs,
Lloyd, and Wilson (1976).) (27)
J. B. Listing
(1861). Der Census raumlicher Complexe oder Verallgemeinerung des Euler'schen Satzes von
den Polyedern, Abh. K. Ges. Wiss. Gottingen Math. CliO, 97-182. (27)
E. Keith Lloyd
see N. L. Biggs
Samuel J. Lomonaco
see J. L. Gross
Laszlo Lovasz
see L. Babai
Roger C. Lyndon
(1985). Groups and Geometry (London Math. Soc. Lecture Note Series, No. 101), London
Math. Soc., Cambridge. (277, 279)
- - and P. E. Schupp (1977). Combinatorial Group Theory, Springer-Verlag, Berlin and New
York. (14)
References 327

C. Maclachlin
(1965). Abelian groups of automorphisms of compact Riemann surfaces, Proc. London Math.
Soc. 15, 699-712. (265)
Saunders MacLane
(1937). A structural characterization of planar combinatorial graphs, Duke Math. J. 3,
460-472. (54)
-and G. Birkhoff (1967), Algebra, Macmillan, New York. (254)
Wilhelm Magnus
(1974). Non-Euclidean Tessalations and Their Groups, Academic Press, Orlando. (279)
--with A. Karrass and D. Solitar (1966). Combinatorial Group Theory, Wiley-Interscience,
New York. (14)
Y. Malgrange
see G. Demoucron
H. Maschke
(1896). The representation of finite groups, A mer. J. Math. 18, 156-194. (287, 290, 300, 314)
Bernard Maskit
see H. Levinson
Lyle McGeoch
see M. L. Furst
Martin Milgram
--and P. Ungar (1977). Bounds for the genus of graphs with given Betti number, J. Combin.
Theory Ser. B 23, 227-233. (136)
Gary Miller
see I. S. Filotti
John Milnor
(1961). Two complexes which are homeomorphic but combinatorially distinct, Ann. of Math.
74, 575-590. (18, 98)
(1965a). Lectures on the h-Cobordism Theorem (Princeton Mathematical Notes), Princeton
University Press, Princeton. (109)
(1965b). Topology from the Differentiable Viewpoint, University Press of Virginia, Charlottes-
ville. (185)
August F. Mobius
(1865). Uber die Bestimmung des Inhaltes eines Polyeders, Ber. K. Sachs. Ges. Wiss. Leipzig
Math.-Phys. Cl17. (25)
Bojan Mohar
- - with T. Pisanski, M. Skoviera, and A. T. White (1985). The cartesian product of three
triangles can be imbedded into a surface of genus 7, Discrete Math. 56, 87-89. (262)
Edwin E. Moise
(1952). Affine structures in 3-manifolds, V, Ann. of Math. 56, 96-114. (18)
J. Montesinos
(1976). Three-manifolds as 3-fold branched covers of S 3 , Quart. J. Math. Oxford 27, 85-94.
(180)
William 0. J. Moser
see H. S. M. Coxeter
328 References

James R. Munkres
(1975). Topology: A First Course, Prentice-Hall, Englewood Cliffs, New Jersey. (95)

Seiya Negami
(1983). Uniqueness and faithfulness of embedding of toroidal graphs, Discrete Math. 44,
161-180. (147)
(1985). Unique and faithful embeddings of projective-planar graphs, J. Graph Theory 9,
235-243. (147)
M. H. A. Newman
(1954). Elements of the Topology of Plane Sets of Points, Cambridge Univ. Press, Cambridge.
(100)
P. Niggli
(1924). Die Flachensymmetrien homogener Diskontinuen, Z. Kristallogr. Mineralog. Petrog.
Abt. A 60, 283-298. (278)
Edward A. Nordhaus
- with B. M. Stewart and A. T. White (1971). On the maximum genus of a graph,
J. Combin. Theory Ser. B 11, 258-267. (132)
-with R. D. Ringeisen, B. M. Stewart, and A. T. White (1972). A Kuratowski-type theorem
for the maximum genus of a graph, J. Combin. Theory Ser. B 12,260-267. (132, 148)
P. S. Novikov
(1955). On the algorithmic unsolvability of the word problem in groups (in Russian), Trudy
Mat. lnst. Steklov 44, lzdat. Akad. Nauk. SSSR, Moscow. (14)

Christos D. Papakyriakopoulos
(1943). A new proof of the invariance of the homology groups of a complex, Bull. Soc. Math.
Grece 22, 1-154. (18, 98)
Torrence D. Parsons
-with T. Pisanski and B. Jackson (1980). Dual imbeddings and wrapped quasi-coverings of
graphs, Discrete Math. 31, 43-52. (207, 210)
see also M. Abu-Sbeih, J. L. Gross, and B. W. Jackson
David Pengelley
(1975). Self-dual orientable embeddings of K., J. Combin. Theory Ser. B 18, 46-52. (209)
R. Pertuisset
see G. Demoucron
Julius Petersen
(1891). Die Theorie der regularen Graphen, Acta Math. 15, 193-220. (35, 36)
Tomaz Pisanski
(1980). Genus of cartesian products of regular bipartite graphs, J. Graph Theory 4, 31-42.
(155, 159)
(1982). Nonorientable genus of cartesian products of regular graphs, J. Graph Theory 6,
31-42. (156)
see also J. L. Gross, B. W. Jackson, B. Mohar, and T. D. Parsons

Georg Polya
(1924). Uber die Analogie der Kristallsymmetrie in der Ebene, Z.Kristallogr. Mineralog.
Petrog. Abt. A 60, 278-282. (278)
References 329
Viera Krnanova Proulx
(1978). Classification of the toroidal groups, J. Graph Theory 2, 269-273. [Extraction of results
from Ph.D. thesis (1977), Columbia University.] (300, 304)
(1981). On the genus of symmetric groups, Trans. A mer. Math. Soc. 266, 531-538. (318)
(1983). On the genus of five- and six-regular graphs, J. Graph Theory 7, 231-234. (317)

Tibor Rado
(1925). Uber den Begriff der Riemannschen Flache, Acta Litt. Sci. Szeged 2, 101-121. (95)

John Reif
see I. S. Filotti

Richard D. Ringeisen
(1972). Determining all compact orientable 2-manifolds upon which Km. • has 2-cell imbed-
dings, J. Combin. Theory Ser. B 12, 101-104. (148)
(1978). Survey of results on the maximum genus of a graph, J. Graph Theory 3, 1-13. (142)
see also E. A. Nordhaus

Gerhard Ringel
(1955). Uber drei kombinatorische Probleme am n-dimensionalen Wurfel und Wurfelgitter,
Abh. Math. Sem. Univ. Hamburg 20, 10-19. (30, 159)
(1959). Farbungsprobleme auf Flachen und Graphen, Springer-Verlag, Berlin and New York.
(30, 236ff)
(1961). Uber das Problem der Nachbargebiete auf orientierbaren Flachen, Abh. Math. Sem.
Univ. Hamburg 25, 105-127. (192, 230, 234)
(1965). Das Geschlecht des vollstandigen paaren Graphen, Abh. Math. Sem. Univ. Hamburg
28, 139-150. (30)
(1974). Map Color Theorem, Springer-Verlag, Berlin and New York. (113, 216, 240, 245)
-and J. W. T. Youngs (1968). Solution of the Heawood map-coloring problem, Proc. Nat.
A cad. Sci. USA 60, 438-445. (38, 215)

Neil Robertson
-and P. D. Seymour (1983). Graph minors, I: excluding a forest, J. Combin. Theory Ser. B
35, 39-61. (53)
-and P. D. Seymour (1984). Graph minors, III: planar tree-width, J. Combin. Theory Ser.
B 36, 49·-64. (53)
- and P. D. Seymour (1986a). Graph minors, V: excluding a planar graph, J. Combin.
Theory Ser. B 41,92-114. (53)
-and P. D. Seymour (1986b). Graph minors, VI: disjoint paths across a disc, J. Combin.
Theory Ser. B 41, 115-138. (53)
- and P. D. Seymour (1987). Graph minors, II: algorithmic aspects of tree width,
J. Algorithms, to appear. (53)

Ronald H. Rosen
see J. L. Gross and F. Harary

Gert Sabidussi
(1958). On a class of fixed-point free graphs, Proc. Amer. Math. Soc. 9, 800-804. (69, 272)

Doris Schattschneider
(1978). The plane symmetry groups: their recognition and notation. A mer. Math. Monthly 85,
439-450. (277)
330 References

W. Scherrer
(1929). Zur Theorie der endlichen Gruppen topologischer Abbildungen von geschlossenen
Flachen in sich, Comment. Math. Helv. 1, 69-119. (287)
Otto Schreier
(1927). Die Untergruppen der freien Gruppen, Abh. Math. Sem. Univ. Hamburg 5, 161-183.
(10)
Paul E. Schupp
see R. C. Lyndon
Seymour Schuster
see G. A. Dirac
Jean-Pierre Serre
--with H. Bass (1977). Groups discretes: arbres, amalgams et SL2 , Cours redige avec Ia
collaboration of H. Bass, Asterisque 46. (14)
Paul D. Seymour
seeN. Robertson
David Singerman
see R. P. Bryant and G. A. Jones
Cho Wei Sit
(1976). Quotients of complete multipartite graphs, Pacific J. Math. 63, 531-538. (236)
(1980). Ph.D. thesis, Columbia University. (250, 318)
Denis Sjerve
see H. H. Glover
Martin Skoviera
see B. Mohar
Stephen Smale
(1961). Generalized Poincare's conjecture in dimensions greater than four, Ann. of Math. 74,
391-406. (109)
Donald Solitar
see W. Magnus
Craig C. Squier
seeM. G. Brin
Saul Stahl
(1978). Generalized embedding schemes, J. Graph Theory 2, 41-52. (113, 136)
(1979). Self-dual embeddings of Cayley graphs, J. Combin. Theory Ser. B 27, 92-107. (209)
(1980). Permutation-partition pairs: a combinatorial generalization of graph embeddings,
Trans. Amer. Math. Soc. 259, 129-145. (154)
(1981). Permutation-partition pairs II. Bounds on the genus of amalgamations of graphs,
Trans. A mer. Math. Soc. 271, 175-182.
- - and L. W. Beineke (1977). Blocks and the nonorientable genus of graphs, J. Graph
Theory 1, 75-78. (155)
--and A. T. White (1976). Genus embeddings for some complete tripartite graphs, Discrete
Math. 14, 279-296. (87, 170)
References 331

John Stallings
(1971). Group Theory and Three-Dimensional Manifolds, Yale University Press, New Haven.
(14)
Richard Statman
see M. L. Furst
Bonnie M. Stewart
see E. A. N ordhaus
Robert E. Trujan
see J. E. Hopcroft
Charles M. Terry
-with L. R. Welch and J. W. T. Youngs (1967). The genus of K 12., J. Combin. Theory 2,
43-60. (235)
-with L. R. Welch and J. W. T. Youngs (1970). Solution of the Heawood map-coloring
problem-case 4, J. Combin. Theory 8, 170-174. (233)
Carsten Thomassen
(1980). Planarity and duality of finite and infinite graphs, J. Combin. Theory Ser. B 29,
244-271. (42)
H. Tietze
(1910). Einige Bemerkerungen uber das Problem des Kartenfarbens auf einseitigen Flachen,
Jahresber. Deutsch. Math.-Verein. 19, 155-159. (217)
Marvin D. Tretkoff
(1975). Covering space proofs in combinatorial group theory, Comm. Algebra 3, 429-457. (14)
(1980). A topological approach to the theory of groups acting on trees, J. Pure Appl. Algebra
16, 323-333. (14)
Albert W. Tucker
(1936). Branched and folded coverings, Bull. Amer. Math. Soc. 42, 859-862. (175)
Thomas W. Tucker
(1980). The number of groups of a given genus, Trans. Amer. Math. Soc. 258, 167-179.
(300, 304)
(1983). Finite groups acting on surfaces and the genus of a group, J. Combin. Theory Ser. B 34,
82-98. (265, 273, 300)
(1984a) On Proulx's four exceptional toroidal groups, J. Graph Theory 8, 29-33. (300, 304,
313)
(1984b). A refined Hurwitz theorem for imbeddings of irredundant Cayley graphs, J. Combin.
Theory Ser. B 36, 244-268. (300, 305, 314)
(1984c). There is one group of genus two, J. Combin. Theory B 36, 269-275. (300, 316)
see also J. L. Gross
William Tutte
(1947). The factorizations of linear graphs, J. London Math. Soc. 22, 107-111. (79)
(1949). On the imbedding of linear graphs in surfaces, Proc. London Math. Soc. (2) 51,
474-483. (46)
(1960). Convex representation of graphs, Proc. London Math. Soc. 10, 304-320. (48)
Peter Ungar
see M. Milgram
332 References

V. G. Vizing
(1964). On an estimate of the chromatic class of a p-graph (in Russian), Diskret. Analiz 3,
25-30. (160)
Chin San Wang
see H. H. Glover
Mark E. Watkins
see R. Frucht
Lloyd R. Welch
see C. M. Terry
Hermann Weyl
(1952). Symmetry, Princeton University Press, Princeton. (277)
Arthur T. White
(1970). The genus of repeated cartesian products of bipartite graphs, Trans. Amer. Math. Soc.
151, 393-404. (15511')
(1972). On the genus of a group, Trans. Amer. Math. Soc. 173, 203-214. (249, 256)
(1974). Orientable imbeddings of Cayley graphs, Duke. Math. J. 41, 353-371. (207)
(1978). Block designs and graph imbeddings, J. Combin. Theory Ser. B 25, 166-183. (105)
(1973). Graphs, Groups and Surfaces (rev. ed., 1984), North-Holland, Amsterdam. (93)
see also M. Jungerman, B. Mohar, E. A. Nordhaus, and S. Stahl
Hassler Whitney
(1932). Non-separable and planar graphs, Trans. Amer. Math. Soc. 34, 339-362. (54)
(1933). A set of topological invariants for graphs, Amer. J. Math. 55, 231-235. (49)
Robin J. Wilson
see N. L. Biggs
Nguyen Huy Xuong
(1979). How to determine the maximum genus of a graph, J. Combin. Theory Ser. B 26,
217-225. (133, 14411')
J. W. T. Youngs
(1963). Minimal imbeddings and the genus of a graph, J. Math. and Mech. 12, 303-315. (117)
(1967). The Heawood map coloring conJecture, Graph Theory and Theoretical Physics
(F. Harary, ed.), Academic Press, Orlando, 313-354. (194)
(1968a). Remarks on the Heawood conjecture (nonorientable case), Bull. Amer. Math. Soc. 74,
347-353. (199, 237)
(1968b). The nonorientable genus of Kn, Bull. Amer. Math. Soc. 74, 354-358. (237)
see also L. Auslander, G. Ringel, and C. M. Terry
Bibliography
Pointers are given to further results on topics discussed in the text and also to
several major research areas mentioned scarcely or not at all. A few useful
monographs are cited. We have favored papers with topological content and
tended to exclude papers not yet published.

Michael Albertson
-and J. P. Hutchinson (1977). The independence ratio and genus of a graph, Trans. A mer.
Math. Soc. 226, 161-173.
-and J. P. Hutchinson (1978). On the independence ratio of a graph, J. Graph Theory 2,
1-8.
-and J.P. Hutchinson (1980). On six-chromatic toroidal graphs, Proc. London Math. Soc.
( 3) 41, 533-556.
- and J. P. Hutchinson (1980). Hadwiger's conJecture for graphs on the Klein bottle,
Discrete Math. 29, 1-11.
V. B. Alekseev
-.and V. S. Gonchakov (1976). Thickness of arbitrary complete graphs, Mat. Sb. 101,
212-230.
Amos Altschuler
(1973). Construction and enumeration of regular maps on the torus, Discrete Math. 4,
201-217.
I. Anderson
-and A. T. White (1978). Current graphs and bi-embeddings, J. Graph Theory 2, 231-239.
Dan Archdeacon
(1986). The non-orientable genus is additive, J. Graph Theory 10, 363-383.
Konhei Asano
(1986). The crossing number of K1• 3• • and K 2• 3. "' J. Graph Theory 10, 1-8.
M. Behzad
- with G. Chartrand and L. Lesniak-Foster. Graphs and Digraphs, Prindle, Weber, and
Schmidt, Boston.
Lowell W. Beineke
- and F. Harary (1969). The thickness of the complete graph, Canad. J. Math. 21,
850-859.
-with F. Harary and J. W. Moon (1964). On the thickness of the complete bipartite graph,
Proc. Cambridge Philos. Soc. 60, 1-5.

333
334 Bibliography

and R. D. Ringeisen (1978). On crossing numbers of certain products of graphs, J.

Combin. Theory. Ser. B 24, 134-136.
--and R. D. Ringeisen (1980). On the crossing numbers of products of cycles and graphs of
order four, J. Graph Theory 4, 145-155.
see also R. K. Guy
D. Benard
(1978). Orientable imbedding of line graphs, J. Combin. Theory Ser. B. 24, 34-43.
Edward A. Bender
- - and E. R. Canfield (1986). The asymptotic number of rooted maps on a surface, J.
Combin. Theory Ser. A 43, 244-257.
- - and L. B. Richmond (1986). A survey of the asymptotic behavior of maps, J. Combin.
Theory Ser. B 40, 297-329.
Claude Berge
(1973). Graphs and Hypergraphs, North-Holland, Amsterdam.
Frank Bernhart
- a n d P. C. Kainen (1979). The book thickness of a graph, J. Combin. Theory Ser. B 27,
320-331.
Norman L. Biggs
(1971 ). Automorphisms of imbedded graphs. J. Combin. Theory II, 132-138.
(1974). Algebraic Graph Theory, Cambridge University Press, Cambridge.
- - and A. T. White (1979). Permutation Groups and Combinatorial Structures, Cambridge
University Press, Cambridge.
Gary S. Bloom
- - see E. F. Schmeichel
J. Ch. Boland
(1967). Embedding of graphs into orientable surfaces, Nederl. Akad. Wetensch. lndag. Math.
29, 33-44.
J. Adrian Bondy
- - and U. S. R. Murty (1976). Graph Theory with Applications, American Elsevier, New
York.
0. V. Borodin
- - and L. S. Melnikov (1974). The chromatic number of a pseudosurface, Diskret. Analiz
24, 8-20.
Morton Brown
-and R. Connelly (1975). On graphs with a constant link-11, Discrete Math. II, 199-232.
William G. Brown
(1966). On the enumeration of non-planar maps, Mem. Amer. Math. Soc. 65.
E. Rodney Canfield
see E. A. Bender
Gary Chartrand
see M. Behzad
F. W. Clarke
- - with A. D. Thomas and D. A. Waller (1980). Embeddings of covering proJections of
graphs, J. Combin. Theory. Ser. B. 28, 10-17.
Bibliography 335

David S. Cochran
-and R. H. Crowell (1970). H2 (G') for tamely embedded graphs, Quart. J. Math. Oxford
Ser. (2) 21, 25-27.
H. Coldeway
see H. Zieschang
R. Connelly
seeM. Brown
John H. Conway
- and C. MeA. Gordon (1983). Knots and links in spatial graphs, J. Graph Theory 7,
445-453.
Richard H. Crowell
see D. S. Cochran
A. K. Dewdney
(1972). The chromatic number of a class of pseudo-manifolds, Manuscripta Math. 6, 311-320.
Gabriel A. Dirac
(1953). The coloring of maps, J. London Math. Soc. 28, 476-480.
(1957). Short proof of a map colour theorem, Canad. J. Math. 9, 225-226.
Dragomir Z. Djokovic
(1974). Automorphisms of graphs and coverings, J. Combin. Theory Ser. B 16, 243-247.
W. Dorfler
(1978). Double covers of hypergraphs and their properties, Ars Combin. 6, 293-313.
Allan L. Edmonds
- with J. H. Ewing and R. S. Kulkarni (1982). Regular tesselations of surfaces and
( P, Q, 2) triangle groups, Ann. of Math. 116, 113-132.
John H. Ewing
see A. L. Edmonds
M. Farzan
-and D. A. Waller (1977). Antipodal embeddings of graphs, J. London Math. Soc. (2) 15,
377-383.
Massimo Ferri
(1979). Crystallisation of 2-fold branched coverings of S 3 , Proc. Amer. Math. Soc. 73,
271-276.
-with C. Gagliardi and L. Graselli (1986). A graph-theoretical representation of PL-mani-
folds-a survey on crystallizations, A equationes Math. 31, 121-141.
Steve Fisk
(1977). Geometric coloring theory, Adv. ,in Math. 24, 298-340.
Carlo Gagliardi
(1981). Extending the concept of genus to dimension n, Proc. A mer. Math. Soc 81,473-481.
see also M. Ferri
V. S. Gonchakov
see V. B. Alekseev
Cameron MeA. Gordon
see J. H. Conway
336 Bibliography

L. Grasselli
seeM. Ferri
Jack Graver
- - and M. Watkins (1977). Combinatorics with Emphasis on the Theory of Graphs, Springer-
Verlag, New York.
Jonathan L. Gross
(1975). The genus of nearly complete graphs-case 6, Aequationes Math. 13,243-249.
Branko Grunbaum
(1970). Polytopes, graphs, and complexes, Bull. Amer. Math. Soc. 76, 1131-1201.
Richard K. Guy
(1969). The decline and fall of Zarankiewicz's theorem, Proof Techniques in Graph Theory
(F. Harary, ed.), Academic Press, Orlando.
(1972). Crossing number of graphs, Graph Theory and Applications, Springer-Verlag, New
York, 111-124.
- - and L. W. Beineke (1966). The coarseness of the complete graph, Canad. J. Math. 20,
888-894.
Frank Harary
- - and W. Tutte (1965). A dual form of Kuratowski's theorem, Canad. Math. Bull. 8,
17-20.
see also L. W. Beineke
Nora Hartsfield
(1986). The toroidal splitting number of the complete graph Kn, Discrete Math. 62, 35-47.
Peter Hoffman
- - and B. Richter (1984). Embedding graphs in surfaces, J. Combin. Theory Ser. B 36,
65-84.
Joan P. Hutchinson
--and George F. McNulty (1983). Connected graphs of genus g with complementary orbits,
Discrete Math. 45, 255-275.
see also M. 0. Albertson
T. Ito
(1981). On a graph of O'Keefe and Wong, J. Graph Theory 5, 87-94.
Bradley W. Jackson
- and G. Ringel (1984). Coloring island maps, Bull. Austral. Math. Soc. 29, 151-165.
- - and G. Ringel (1984). The splitting number of complete bipartite graphs, Arch. Math.
42,178-184.
--and G. Ringel (1984). Heawood's empire problem on the plane, J. Reine Angew. Math.
347, 148-153.
- - and G. Ringel (1985). Heawood's empire problem, J. Combin. Theory Ser. B 38,
168-178.
Lynne D. James
- - and G. A. Jones (1985). Regular orientable imbeddings of complete graphs, J. Combin.
Theory Ser. B 39, 353-367.
Gareth A. Jones
see L. D. James
Bibliography 337

Paul C. Kainen
- and A. T. White (1978). On stable crossing numbers, J. Graph Theory 2, 181-187.
see also F. Bernhart
Daniel J. Kleitman
(1970). The crossing number of K 5•• , J. Combin. Theory 9, 315-323.
Hudson V. Kronk
(1969). An analogue to the Heawood map-coloring problem, J. London Math. Soc. (2) I,
550-552.
Ravi S. Kulkarni
see A. L. Edmonds
Gerard Lallament
--and D. Perrin (1981). A graph covering construction of all finite complete biprefix codes,
Discrete Math. 36, 261-272.
Frank Thomson Leighton
(1982). Finite common coverings, J. Combin. Theory Ser. B 33, 231-238.
Linda Lesniak-Foster
see M. Behzad
Sostenes Lins
-and A. Mandel (1985). Graph-encoded 3-manifolds, Discrete Math. 57, 261-284.
Bernard Liouville
(1978). Sur le genre d'une somme cartesienne de graphes, Problemes combinatoires et theorie
des • * h • graphes, Colloques Internal. CNRS 260, Paris, 275-277.
Richard J. Lipton
-and R. E. Trujan (1979). A separator theorem for planar graphs, SIAM J. Appl. Math.
36, 177-189.
--and R. E. TaiJan (1980). Applications of a planar separator theorem, SIAM J. Comput.
9, 615-627.
Heidi Mahnke
(1972). The necessity of non-abelian groups in the case 0 of the Heawood map-coloring
problem, J. Combin. Theory B I3, 263-265.
Joseph Malkevitch
(1970). Properties of planar graphs with uniform vertex and face structure, Mem. Amer.
Math. Soc. 99.
Arnaldo Mandel
seeS. Lins
George F. McNulty
see J. P. Hutchinson
L. S. Melnikov
see 0. V. Borodin
Bojan Mohar
(1986). A common cover of graphs and 2-cell embeddings, J. Combin. Theory Ser. B 40,
94-106.
John W. Moon
see L. W. Beineke
338 Bibliography

Bruce P. Mull
- - and R. G. Rieper and A. T. White (1987). Enumerating 2-cell imbeddings of complete
graphs, Proc. A mer. Math. Soc., to appear.
Ronald C. Mullin
(1965). On counting rooted triangular maps, Canad. J. Math. 17, 373-382.
L. Nebesky
(1981). Every connected, locally connected graph is upper embeddable, J. Graph Theory 5,
197-199.
Seiya Negarni
(1987). Enumeration of proJective-planar embeddings of graphs, Discrete Math., to appear.
Lee P. Neuwirth
(1966). Imbedding in low dimension, lllinois J. Math. 10, 470-478.
Torrence D. Parsons
(1980). Circulant graph imbeddings, J. Combin. Theory SerB 29, 310-320.
Dominique Perrin
see G. Lallament
Giustina Pica
- - with T. Pisanski and A. G. S. Ventre. The genera of amalgamations of cube graphs,
Glasnik Mat. 19, 21-26.
Tomaz Pisanski
- with J. Shawe-Taylor and J. Vrabec (1983). Edge-colorability of graph bundles, J.
Combin. Theory Ser. B 35, 12-19.
see also G. Pica
L. Bruce Richmond
see E. A. Bender
Bruce Richter
- and H. Shank (1984). The cycle space of an embedded graph, J. Graph Theory 8,
365-369.
see also P. Hoffman
Robert G. Rieper
see B. P. Mull
Richard D. Ringeisen
see L. W. Beineke
Gerhard Ringel
see B. Jackson
Thomas L. Saaty
(1967). Two theorems on the minimum number of intersections for complete graphs, J.
Combin. Theory 2, 571-584.
Edward F. Schmeichel
- - and G. S. Bloom (1979). Connectivity, genus, and the number of components in
vertex-deleted subgraphs, J. Combin. Theory Ser. B 27, 198-201.
Herbert Shank
see B. Richter
Bibliography 339

John Shawe-Taylor
see T. Pisanski
Jozef Siran
- - and M. Skoviera (1985). Quotients of connected regular graphs of even degree, J.
Combin. Theory Ser. B 38, 214-225.
Martin Skoviera
(1986). A contribution to the theory of voltage graphs, Discrete Math. 61, 281-292.
see also J. Siran
Z. Skupien
(1966). Locally Hamiltonian and planar graphs, Fund. Math. 58, 193-200.
Saul Stahl
(1983). The average genus of classes of graph embeddings, Congr. Number. 40, 375-388.
Richard P. Stanley
(1981). Factorization of permutations into n-cycles, Discrete Math. 37, 255-262.
H. Joseph Straight
(1979). Cochromatic number and the genus of a graph, J. Graph Theory 3, 43-51.
Robert E. T3.1Jan
see R. J. Lipton
William T. Tutte
(1973). What is a map? New Directions in the Theory of Graphs (F. Harary, ed.), Academic
Press, Orlando, 309-325.
(1986). From topology to algebra, J. Graph Theory 10, 331-338.
see also F. Harary
Brian Ummel
(1978). The product of nonplanar complexes does not imbed in 4-space, Trans. A mer. Math.
Soc. 242, 319-328.
Aldo G. S. Ventre
see G. Pica
Andrew Vince
(1983). Combinatorial maps, J. Combin. Theory Ser. B 34, 1-21.
(1983). Regular combinatorial maps, J. Combin. Theory Ser. B 35, 256-277.
E. Vogt
see H. Zeischang
Joze Vrabec
see T. Pisanski
Derek A. Waller
(1976). Double covers of graphs, Bull. Austral. Math. Soc. 14, 233-248.
see also F. W. Clarke and M. Farzan
Mark Watkins
see J. C. Graver
Arthur T. White
(1983). Ringing the changes, Math. Proc. Camb. Philos. Soc. 94, 203-215.
see also I. Anderson and B. G. Mull
340 Bibliography

Nick C. Wormald
(1981). Counting unrooted planar maps, Discrete Math. 36, 205-225.
Joseph Zaks
(1974). The maximum genus of cartesian products of graphs, Canad. J. Math. 26, 1025-1035.
Heiner Zieschang
--with E. Vogt and H. Coldeway (1980). Surfaces and Planar Discontinuous Groups (Lecture
Notes in Mathematics, Vol. 835), Springer-Verlag, New York.
Table of Notations

GRAPHS

G,H, ... arbitrary graphs, 1

Ec,E the edge set of graph G, 1
Vc, V the vertex set of graph G, 1
d, e, . . . edges, 1
u, v, .. . vertices, 2
Vc(e), V(e) the endpoint set of edge e, 1
e+,e- the plus, minus directions of edge e, 5
G-e the graph G with edge e deleted, 33
Gje the graph G with edge e contracted, 39
GC the edge complement of graph G, 19
GxH the cartesian product of graphs G and H, 19
G+H the suspension (join) of graphs G and H, 20
G •1 H the amalgamation of graphs G and H by map f, 20
G •v
H the amalgamation of graphs G and H at vertex v,
151
Gsimp the maximal simplicial subgraph of G, 38
Gsmooth the smoothing of graph G, 40
C(.91, X) the Cayley color graph for group .91 and generating
set X, 11
C(.91, X)o the Cayley graph for group .91 and generating set X,
11
C(.91, X)l the alternative Cayley graph for group .91 and gener-
ating set X, 12
S(.91: !II, X) the Schreier color graph for the cosets of subgroup !II
and generating set X, 73
S(.91: !II, X) 0 the Schreier graph for the cosets of subgroup !II and
generating set X, 73
S( .91: !II, X) 1 the alternative Schreier graph for the cosets of sub-
group !II and generating set X, 73
a tree, 8
a walk, 8
the cycle of length n, 8

341
342 Table of Notations

pn the path with n vertices (and n - 1 edges), 8

Kn the complete graph on n vertices, 10
Km,n the complete bipartite graph with partsizes m, n, 15
Km,n,r, ... the complete multipartite graph with partsizes
m, n, r, ... , 196
Bn the bouquet of n circles, 15
on the n-octahedron graph, 20
Qn the n-cube graph, 19
G(n,k) a generalized Petersen graph, 59
H(n) the composition of graph H by a set of n indepen-
dent vertices, 210
K(S) the set of Kuratowski graphs for the surface S, 53

GROUPS

.J/1' !II' .. . arbitrary groups, 11

a, b, .. . group elements, 11
X a generating set, 11
X, y, ... generators, 11
1_,..,, 1 the identity element of group .J/1, 12
(x, ... : R = 1, ... ) the group with generators x, . . . and relators R, . .. ,
172
(x,y, ... ) the subgroup generated by x, y, ... , 290
[x, y] the commutator xyx- 1y-I, 290
.J/1:!!1 the collection of cosets of the subgroup !II in the
group .J/1, 73
the quotient of group .J/1 by the normal subgroup ?A,
80
.JiiX!JB the direct product of groups .J/1 and !II, 15
;?Zn the cyclic group of order n, 10
!!)n the dihedral group of order 2n, 22
g;, the full symmetric group on n symbols, 13
.Jiln the alternating group on n symbols, 15
'TT,p permutations, 83
(cl, ... ,cn) the cycle structure of a permutation having c; cycles
of length i, 83
GL(2, p) the general linear group of 2 X 2 matrices over the
field of p elements, 313
SL(2, p) the special linear group of 2 X 2 matrices of determi-
nant 1, 313
PSL(2, p) the projective special linear group, 297
Aut( G) the automorphism group of graph G, 6
'TT1 (G, u) the fundamental group of graph G based at vertex u,
92
Imbeddings and Rotations 343

SURFACES AND COMPLEXES

the orientable surface of genus g, 24

the nonorientable surface with k crosscaps, 26
a (simplicial) complex, 96
CJ a simplex, 96
(vo, ... ,vk) the k-simplex having vertices v0, ... , vk, 96
IKI the carrier of complex K, 96
K(r) the r-skeleton of complex K, 96
(V, C) an abstract complex with vertex set V, 97
star(u) the star of a vertex, 106
link(v) the link of a vertex, 107
SoT the disk sum of surfaces S and T, 131

GROUP AcnONS AND COVERINGS

graph automorphisms or surface homeomorphisms

corresponding to group element a, 21, 182
[v], [e], [x] the orbit of a vertex, edge, or point under a group
action, 21, 186
Gj.Jii,Sj.Jil the graph, surface quotient under the action of group
.Jil, 22, 186
p: G ~ Gj.Jil the quotient map for the action of group .Jil, 22, 186
p: S ~ Sj.Jil
p:G~ G a covering projection for graphs, 75
p:s~s a covering projection, possibly branched, for surfaces,
174
y the branchpoint set of a branched covering, 174
Silo the orientation-preserving subgroup of a surface
action of group .Jil, 269
def(y) the deficiency of branchpoint y, 175
the order of the branchpoint y for a regular cover-
ing, 284
a symmetry group, 276
the translation subgroup of a Euclidean space group,
277

IMBEDDINGS AND ROTATIONS

i: G ~ S,G ~ S an imbedding of graph G in surface S, 26

FG, F the set of faces for an imbedding of graph G, 26
/; the number of faces of size i, 302
G* the graph dual to an imbedding of G, 31
344 Table of Notations

e*, v*, f* dual edges, vertices, faces, 31

e+* the dual plus direction of edge e, 202
v.e 1 ••• en a rotation at vertex v, 112
el a type-1 edge, 112
p a rotation as a permutation of the directed edge set,
161
... 23 ... a rotation ... x 2 x] 1 ... for a Cayley graph with
generators x 2 , x 3 , 257

VOLTAGE AND CURRENT GRAPHS

(G, a) an (ordinary) voltage graph with voltage assignment

a, 57
(G, ajPA) a relative voltage graph, 74
(G, a)n a permutation voltage graph with assignment in the
symmetric group Y,, 81
the derived graph for voltage assignment a, 57
the natural projection of the derived graph onto the
base graph, 61
(G--+ S, a) an imbedded (ordinary) voltage graph, 162
(G--+ S, a)n an imbedded permutation voltage graph, 162
sa the derived surface for an imbedded voltage graph,
162
the derived imbedding for an imbedded voltage graph,
162
p: sa --+ s a natural projection of the derived surface onto the
base surface, 176
(G--+S,/3) a current graph with current assignment /3, 194
Gp --+ Sp the derived imbedding for current assignment /3, 194
(e, a), ea edges of a derived graph, 57
(v, a), va vertices of a derived graph, 57
(e, b)-,(e-, b) minus directions of derived edges for a current graph,
194
(G* --+ S, /3*) the voltage graph dual to the current graph
(G--+ S, /3), 204
the dual derived imbedding for the current graph
(H--+ S,/3), 206
p: Hp*--+ H the natural projection of the dual derived graph onto
the base current graph, 207
d(u) the local (voltage) group at vertex u, 87
d 0 (u) the type-0 local (voltage) group, 169
a(v, T) the T-potential of vertex v, 89
aT( e) the T-voltage for edge e, 90
General Notation 345

GENUS AND RELATED PARAMETERS

y(G) the (minimum) genus of graph G, 29

YM(G) the maximum genus of graph G, 132
y(G) the (minimum) crosscap number of graph G, 29
YM(G) the maximum crosscap number of graph G, 133
{J(G) the Betti number (cycle rank) of graph G, 135
~(G, T) the deficiency of graph G with respect to the span-
ning tree T, 143
~(G) the deficiency of graph G, 143
girth( G) the girth of graph G, 29
chr(G) the chromatic number of graph G, 37
GR(G) the genus range of graph G, 132
CR(G) the crosscap range of graph G, 133
ER(G) the Euler characteristic range of graph G, 148
y(Jaf) the genus of group Jaf, 249
a(Jaf) the symmetric genus of group Jaf, 264
ao( Jaf) the strong symmetric genus of group Jaf, 264
x(G--+ S) the Euler characteristic of the imbedding G --+ S, 121
x(S) the Euler characteristic of surface S, 130
H(S) the Heawood number of surface S, 215
chr(S) the chromatic number of surfaceS, 37
!_(n) the integer r<n - 3)(n - 4)/121, 221
l(n) the integer r<n - 3)(n - 4)/61, 221

GENERAL NOTATION

{xiP} the set of all x such that P holds, 1

vEV,vf/.V v is, is not, an element of set V, 1
AcB set A is contained in set B, 12
AUB the union of sets A and B, 18
A-B the set of elements in A but not in B, 18
AXB the cartesian product of sets A and B, 11
#V the number of elements in the set V, 3
<,~ strictly less than, less than or equal, 29
>,~ strictly greater than, greater than or equal, 19
a= b mod n the integers a and b are congruent modulo n, 146
[xj the "floor of x ", i.e., the greatest integer less than or
equal to x, 38
rx1 the "ceiling of x ", i.e., the least integer greater than
or equal to x, 38
346 Table of Notations

NOTATION USED IN FIGURES

the plus direction of edge e, 5

a type-1 edge, 114

• )I ....4;E--..
• a type-1 edge in a current graph, 199

• 0 clockwise, counterclockwise rotations, 193

t .____I____.I t identify sides of the polygon according to the

arrows, 30

identify sides of the polygon according to the arrows,

104
if E

identify sides of the digon, non-orientably, 237

0 a local orientation, 106

Subject Index

Abelian group: 1-band, 109

canonical form, 254 ends of, 109
canonical generating set, 254 orientation- preserving, reversing, 110
genus of, 251 If locally oriented, 11 0
Action of group: orientability algorithm, 111
fixed point of, 182 reduced, 109
free, 21, 182 2-Basis for graph, 54
on graph, 21 Battle, Harary, Kodama, Youngs (BHKY)
orbits of, 21, 186 theorem, 151
quotient: Betti number, 135
map, 21, 186 Block, 40
space, 21, 186 Block design, 105
on surface, 186 balanced, 105
natural action of quotient group, 270 balanced incomplete (BIBD), 105
natural action of voltage group, 186 Bouquet of n circles, 15
orientation-preserving subgroup, 269 Branched covering, 174
pseudofree, 186 branch points, 174
Adjacency matrix, 4 deficiency of, 175
Adjacent imbeddings, 133 number of sheets, 175
Adjacent vertices, 4 prebranch points, 174
Alexander's horned sphere, 100 regular, 186
Alexander's Theorem, 179 covering transformations of, 187
Algorithm:
face- tracing, 115 Cartesian product of graphs, 19
genus, 53, 149 genus of, 155 If
maximum genus, 146 Cayley color graph, 10
orientability, 111 Cayley graph, 10
planarity, 52 alternative, 12, 271
Alternating group, 287 imbedding of:
Alternative Cayley graph, 12, 271 self-dual, 208
Amalgamation of graphs, 20 strongly symmetric, 264ft'
genus of, 150ft' symmetric, 264ft'
Appel and Haken, 38 weakly symmetric, 264ft'
Automorphism of graph, 6 irredundant, 304
fixed-point free, 67 left translation, 13
preserving edge type, 184 Cell complex, 99
respecting rotation, 184 Cellular ( 2- cell) imbedding, 100
reversing rotation, 184 Chromatic number:
of graph, 37
Band decomposition of surface, 109 of surface, 37, 215ft'
0-, 2-bands, 109 Chromatically critical, 217

347
348 Subject Index

Coi1,226 Cycle:
Coloring: rank, 135
of graph, 37 space, 54
of map, 37 structure of permutation, 83
Commutator of group elements, 29, 275
Complete bipartite graph, 14 Deficiency:
genus of, 210ff of branch point, 175
maximum genus of, 148 of graph, 143
Complete graph, 10 of spanning tree, 143
crosscap number of, 223tT Dihedral group, 285
genus of, 223ff presentation for, 285
maximum genus of, 146 Disk sum of two surfaces, 131
Complex: Distance between vertices, 36
cell, 99 Dodecahedron, 19
planar 2-complex, 117 Double-cover conjecture, 46
simplicial, 96 Dual:
abstract, 97 graph, 31,116
Composition of graph, 210 imbedding, 31, 116
triangular imbedding for, 210 plus-direction of edge, 202
Connected, 8 voltage graph, 204tT
component, 18 Duality, 31, 116
k-connected, 40 current- voltage, 202tT
k-edge-connected, 145
Co- spectral graphs, 7 Edge:
Covering, 23, 72 coloring of graph, 156
branched, 182tT saturated color, 156
folded, 175 contraction, 39
irregular, 72 tT -deletion surgery, 124
projection, 23, 72 direction (plus, minus), 4
regular, 23, 75, 187 initial, terminal vertices of, 4
space, 23, 75 loop, 4
wrapped, 207 proper, 4
Crosscap, 26 sliding, 125
number of graph, 13 3 tT -twisting surgery, 134
relation to genus, 137tT Euclidean:
range, 13 3 tT space group, 276tT
Cube ( n- cube), 19 triangle group, 280
crosscap number of, 159 Euler:
genus of, 159 characteristic:
Current assignment, 194 of cellular imbedding (relative), 121 tT
Current graph (ordinary), 194ff of surface, 2 7, 130
derived graph of, 194 equation, 27
derived imbedding of, 194 formula, 27
dual, 206 Eulerian:
derived graph of, 206 circuit, 34
derived imbedding of, 206 graph, 34
natural projection for, 206
nonorientable, 198tT Face, 100
edge types of, 199 Tracing Algorithm, 115
reversed comer of, 199 Factor ( n- factor), 35
orientable, 194 Factorable (n-factorable), 35
Current group, 194 Five-color theorem, 219
Cutpoint, 40 Folded covering, 175
Cycle (n-cycle), 8 Four-Color theorem, 38, 215
Subject Index 349
Frucht's theorem, 70 Gustin's current graphs, 193
Fundamental:
group of graph, 92 Hadwiger conjecture, 40
semigroup of graph, 92 Hajos conjecture, 41
Hamiltonian cycle, 35
Genus: Hamiltonian graph, 35
of graph, 29, 132 Handle decomposition, 109
average, 147 Hauptvermutung, 98
distribution, 14 7 Heawood:
maximum, 132 inequality, 215 ff
minimum, 132 number, 215fT
nonorientable ( crosscap number), 29, 133 problem:
range, 132 orientable, 215
of group, 93, 249fT regular cases, 224fT
abelian, 249fT irregular cases, 224fT
infinite, 249 nonorientable, 215
strongly symmetric, 264 regular cases, 224fT
symmetric, 264 irregular cases, 224fT
of surface, 29 Heffier, Lothar, 191
nonorientable ( crosscap number), 29 Hill-climb, 149
Girth, 29 Hurwitz's theorem, 296
Glide, 276 refined, 298
n-Gon, 27 Hypergraph, 97
Graph, l
adjacency matrix, 4 Icosahedron, 41
amalgamation, 20 Imbedding:
automorphism, 6 adjacent, 13 3
group, 6 cellular (2-cell), 100
bipartite, 14 distribution, 13 2
cartesian product, 19 crosscap range, 132
edge-complement, 19 Euler characteristic range, 148
edge set, l genus range, 132
homeomorphic, 18 dual, 31, 116
incidence matrix, 3 equivalent, lO l
isomorphism, 5 face (region) of, 26, 100
map, 5 boundary walk of, l 0 l
regular ( n- regular), 13 side of, 26, lO l
simplicial (simple), l size of, 26, 101
subdivision, 18 quadrilateral, 27
subgraph, 16 triangular, 2 7
induced, 17 type (0, l) of edge, II 0
spanning, 16 type (0, l) of walk, 110
suspension (join), 20 weakly equivalent (congruent), l 02
topological representation, l 7 Interpolation theorem:
vertex set, l nonorientable, 135
Graphical regular representation for group, orientable, 134
72
Group: Jordan curve theorem, 100
abelian, 254
Cayley graph for, lOfT Kempe chain, 220
crystallographic, 2 77 Kirchoff current law ( KCL), 196
current, l 94 ff Kirchoff voltage law (KVL), 166
of small genus, 283fT Klein bottle, 26, 103
voltage, 57fT K-metacyclic group, 318
350 Subject Index
Konigsberg bridge problem, 34 Petersen graph, 21, 58, 72, 86, 181
Kuratowski graphs, 28 generalized, 59, 3 16
for higher genus, 53 Petersen's theorem, 36
Kuratowski's theorem, 28, 42fT Piecewise linear, 282
Planarity, 42fT
Ladder: algorithms, 51 ff
circular, 59 Fary's theorem, 48
Mobius, 59 Kuratowski's theorem, 42fT
Lift: Mac Lane's theorem, 42fT
offace, 163 Tutte' s theorem, 48
of rotation, 162 Whitney's Duality theorem, 54
of walk, 62 Whitney's Uniqueness theorem, 49
Linear map of simplexes, 282 Projective plane, 26, 102
Line graph, 3 3 Proulx's classification of groups of genus one,
Link of vertex, 107 300fT
Local genus minimum, 149 Pseudofree, 182
Local voltage group, 87, 169 Pseudosurface, 104
Log of circuits, 253 two-fold triple systems, 105
Longitude, 120
Loop, 3 Quatemion group, 299
Quotient:
Manifold (n-manifold), 95 graph, 22
boundary of, 95 natural group action, 270
closed, 96 space, 186
Maschke's classification of genus zero groups, surface, 186, 270
290, 314
Meridian, 120 Rank of group, 250
Mobius band, 25 Reflection, 269
dividing circles of, 269
Octahedron (cocktail-party) graph, 20, 206 halves of, 269
crossing number for, 214 Regular covering, 23, 186
Ordinary voltages, 57 graph, 13
Orientability: Relator, 306
algorithm for surface, Ill reduced, 306
of surface, 6, 10 Riemann- Hurwitz equation, 179, 284
test for derived surface, 170 Riemann surface, 174
type of edge or walk, II 0 Ringel, Gerhard, 30, 191
Orientation: Ringel- Youngs Theorem, 30, 215fT
of imbedding, l 06 Rotation:
of surface, l 06 system, 113
of triangulation, l 06 as permutation, 161
projection of, 113
Patchwork, 155 pure, 113
disjoint, 15 5 vertex form, 114
even, 156 at vertex, 112
quadrilateral, 156
residual faces of, 156 Schemes, 192
Path (n-path), 8 Schoenfliess theorem, l 00
Permutation voltage graph, 81 Schreier color graph, 73
base graph, 81 Schreier (coset) graph, 7 3
derived graph, 81 alternative, 73
imbedded, 166 Self- adjacent, 4
derived imbedding, 166 Self-dual imbeddings:
face lifting, 16 7 of Cayley graphs, 208
natural projection, 81 of complete graphs, 208, 213
Subject Index 351
Simplex (geometric k-simplex), 96ff Twofold triple system, 105
barycenter of, 99 Twisting an edge, 134
face of, 96 Type-0, -1:
Simplicial complex, 96 of edge, 110
abstract, 97 of walk, 110
isomorphism of, 97
carrier of, 96 Valence, 3
m- complex, 96 average, 216
r- skeleton, 96 sequence, 7
subdivision, 97 Voltage assignment, 57ff
first barycentric, 99, 107 natural, 251
Slide of one edge along another, 125 net:
Space group, 277 on cycle, 61
Euclidean, 277 on path, 61
classification, 2 77 ordinary, 57ff
hyperbolic, 279 permutation, 81 ff
spherical, 279 relative, 74ff
Star of vertex, 24, 106 Voltage graph:
Surface, 24ff, 96ff imbedded, 162ff
classification, 119ff derived imbedding, 163
closed, 24, 96 derived rotation system, 163
crosscap number of, 29 dual to current graph, 204
disk sum, 131 face lifting, 16 3
Euler characteristic of, 130 natural action of voltage group, 186
genus of, 29 natural projection, 177
nonorientable, 25 orientability test, 170
orientable, 25 type-0 local group, 169
Suspension (join), 20 ordinary, 57ff
Symmetric, 13 base graph, 57
genus, 264ff derived graph, 57
strong, 264ff fiber, 60
group on n symbols, 13 natural (left) action of voltage group,
imbedding, 264ff 66
strong, 264ff natural projection, 60
weak, 264ff permutation, 81
Symmetry group, 276 base graph, 81
derived graph, 81
Torus, 30, 276 fiber, 81
Tree, 8 Voltage group, 57
T- potential for, 8 9 local, 87
T- voltage for, 90 natural action of, 66
Triangle group, 279ff
Euclidean, 279 Walk, 8
full (p, q, r), 280 boundary, 101
hybrid (m, m, n), 290 closed, 8
hyperbolic, 280 length of, 8
ordinary~ q, r), 281 open, 8
presentation of, 281 Wheel graph, 23
proper quotient of, 298 White-Pisanski imbedding, 155ff
spherical, 280 Word:
Triangulation: in group, 14, 306
orientation for, 106 problem, 14
of surface, 98 Wrapped covering, 207
of topological space, 98
Tubes, 157 Youngs, J. W. T., 30