S T U D E N T M AT H E M AT I C A L L I B R A RY

Volume 60

Mostly Surfaces
Richard Evan Schwartz
S T U D E N T M AT H E M AT I C A L L I B R A RY
Volume 60

Mostly Surfaces

Richard Evan Schwartz

American Mathematical Society

Providence, Rhode Island
 Editorial Board
Gerald B. Folland Brad G. Osgood (Chair)
Robin Forman John Stillwell

2010 Mathematics Subject Classification. Primary 14Q10, 30F10, 32J15,

37E30, 53A05, 51M20, 32Q45.
The cover illustration was drawn by the author.

For additional information and updates on this book, visit

www.ams.org/bookpages/stml-60

Library of Congress Cataloging-in-Publication Data

Schwartz, Richard Evan.
Mostly surfaces / Richard Evan Schwartz.
p. cm. — (Student mathematical library ; v. 60)
Includes bibliographical references and index.
ISBN 978-0-8218-5368-9 (alk. paper)
1. Hypersurfaces. 2. Riemann surfaces. 3. Surfaces, Algebraic. I. Title.

QA571.S385 2011
516.352—dc22
2011005544

Copying and reprinting. Individual readers of this publication, and nonprofit

libraries acting for them, are permitted to make fair use of the material, such as to
copy a chapter for use in teaching or research. Permission is granted to quote brief
passages from this publication in reviews, provided the customary acknowledgment of
the source is given.
Republication, systematic copying, or multiple reproduction of any material in this
publication is permitted only under license from the American Mathematical Society.
Requests for such permission should be addressed to the Acquisitions Department,
American Mathematical Society, 201 Charles Street, Providence, Rhode Island 02904-
2294 USA. Requests can also be made by e-mail to [email protected].


c 2011 by the American Mathematical Society. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.

∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at http://www.ams.org/
10 9 8 7 6 5 4 3 2 1 16 15 14 13 12 11
Contents

Preface xi

Chapter 1. Book Overview 1

§1.1. Behold, the Torus! 1
§1.2. Gluing Polygons 3
§1.3. Drawing on a Surface 5
§1.4. Covering Spaces 8
§1.5. Hyperbolic Geometry and the Octagon 9
§1.6. Complex Analysis and Riemann Surfaces 11
§1.7. Cone Surfaces and Translation Surfaces 13
§1.8. The Modular Group and the Veech Group 14
§1.9. Moduli Space 16
§1.10. Dessert 17

Part 1. Surfaces and Topology

Chapter 2. Definition of a Surface 21

§2.1. A Word about Sets 21
§2.2. Metric Spaces 22
§2.3. Open and Closed Sets 23
§2.4. Continuous Maps 24

iii
iv Contents

§2.5. Homeomorphisms 25
§2.6. Compactness 26
§2.7. Surfaces 27
§2.8. Manifolds 28

Chapter 3. The Gluing Construction 31

§3.1. Gluing Spaces Together 31
§3.2. The Gluing Construction in Action 33
§3.3. The Classification of Surfaces 36
§3.4. The Euler Characteristic 37

Chapter 4. The Fundamental Group 43

§4.1. A Primer on Groups 43
§4.2. Homotopy Equivalence 45
§4.3. The Fundamental Group 46
§4.4. Changing the Basepoint 48
§4.5. Functoriality 49
§4.6. Some First Steps 51

Chapter 5. Examples of Fundamental Groups 53

§5.1. The Winding Number 53
§5.2. The Circle 56
§5.3. The Fundamental Theorem of Algebra 57
§5.4. The Torus 58
§5.5. The 2-Sphere 58
§5.6. The Projective Plane 59
§5.7. A Lens Space 59
§5.8. The Poincaré Homology Sphere 62

Chapter 6. Covering Spaces and the Deck Group 65

§6.1. Covering Spaces 65
§6.2. The Deck Group 66
§6.3. A Flat Torus 67
Contents v

§6.4. More Examples 69

§6.5. Simply Connected Spaces 70
§6.6. The Isomorphism Theorem 71
§6.7. The Bolzano–Weierstrass Theorem 72
§6.8. The Lifting Property 73
§6.9. Proof of the Isomorphism Theorem 74

Chapter 7. Existence of Universal Covers 79

§7.1. The Main Result 80
§7.2. The Covering Property 82
§7.3. Simple Connectivity 84

Part 2. Surfaces and Geometry

Chapter 8. Euclidean Geometry 87

§8.1. Euclidean Space 87
§8.2. The Pythagorean Theorem 90
§8.3. The X Theorem 91
§8.4. Pick’s Theorem 92
§8.5. The Polygon Dissection Theorem 96
§8.6. Line Integrals 98
§8.7. Green’s Theorem for Polygons 100

Chapter 9. Spherical Geometry 103

§9.1. Metrics, Tangent Planes, and Isometries 103
§9.2. Geodesics 105
§9.3. Geodesic Triangles 107
§9.4. Convexity 110
§9.5. Stereographic Projection 111
§9.6. The Hairy Ball Theorem 113

Chapter 10. Hyperbolic Geometry 115

§10.1. Linear Fractional Transformations 115
§10.2. Circle Preserving Property 116
vi Contents

§10.3. The Upper Half-Plane Model 118

§10.4. Another Point of View 121
§10.5. Symmetries 121
§10.6. Geodesics 123
§10.7. The Disk Model 125
§10.8. Geodesic Polygons 127
§10.9. Classification of Isometries 130

Chapter 11. Riemannian Metrics on Surfaces 133

§11.1. Curves in the Plane 133
§11.2. Riemannian Metrics on the Plane 134
§11.3. Diffeomorphisms and Isometries 135
§11.4. Atlases and Smooth Surfaces 136
§11.5. Smooth Curves and the Tangent Plane 137
§11.6. Riemannian Surfaces 139

Chapter 12. Hyperbolic Surfaces 143

§12.1. Definition 143
§12.2. Gluing Recipes 145
§12.3. Gluing Recipes Lead to Surfaces 147
§12.4. Some Examples 149
§12.5. Geodesic Triangulations 150
§12.6. Riemannian Covers 152
§12.7. Hadamard’s Theorem 154
§12.8. The Hyperbolic Cover 156

Part 3. Surfaces and Complex Analysis

Chapter 13. A Primer on Complex Analysis 163

§13.1. Basic Definitions 163
§13.2. Cauchy’s Theorem 165
§13.3. The Cauchy Integral Formula 167
§13.4. Differentiability 168
Contents vii

§13.5. The Maximum Principle 170

§13.6. Removable Singularities 171
§13.7. Power Series 172
§13.8. Taylor Series 174

Chapter 14. Disk and Plane Rigidity 177

§14.1. Disk Rigidity 177
§14.2. Liouville’s Theorem 179
§14.3. Stereographic Projection Revisited 181

Chapter 15. The Schwarz–Christoffel Transformation 183

§15.1. The Basic Construction 184
§15.2. The Inverse Function Theorem 185
§15.3. Proof of Theorem 15.1 186
§15.4. The Range of Possibilities 188
§15.5. Invariance of Domain 189
§15.6. The Existence Proof 190

Chapter 16. Riemann Surfaces and Uniformization 195

§16.1. Riemann Surfaces 195
§16.2. Maps Between Riemann Surfaces 197
§16.3. The Riemann Mapping Theorem 199
§16.4. The Uniformization Theorem 201
§16.5. The Small Picard Theorem 202
§16.6. Implications for Compact Surfaces 203

Part 4. Flat Cone Surfaces

Chapter 17. Flat Cone Surfaces 207

§17.1. Sectors and Euclidean Cones 207
§17.2. Euclidean Cone Surfaces 208
§17.3. The Gauss–Bonnet Theorem 209
§17.4. Translation Surfaces 211
§17.5. Billiards and Translation Surfaces 213
viii Contents

§17.6. Special Maps on a Translation Surface 217

§17.7. Existence of Periodic Billiard Paths 219

Chapter 18. Translation Surfaces and the Veech Group 221

§18.1. Affine Automorphisms 221
§18.2. The Diffential Representation 223
§18.3. Hyperbolic Group Actions 224
§18.4. Proof of Theorem 18.1 226
§18.5. Triangle Groups 228
§18.6. Linear and Hyperbolic Reflections 229
§18.7. Behold, The Double Octagon! 232

Part 5. The Totality of Surfaces

Chapter 19. Continued Fractions 239

§19.1. The Gauss Map 239
§19.2. Continued Fractions 241
§19.3. The Farey Graph 242
§19.4. Structure of the Modular Group 244
§19.5. Continued Fractions and the Farey Graph 245
§19.6. The Irrational Case 247

Chapter 20. Teichmüller Space and Moduli Space 251

§20.1. Parallelograms 251
§20.2. Flat Tori 252
§20.3. The Modular Group Again 254
§20.4. Moduli Space 256
§20.5. Teichmüller Space 258
§20.6. The Mapping Class Group 260

Chapter 21. Topology of Teichmüller Space 263

§21.1. Pairs of Pants 263
§21.2. Pants Decompositions 265
§21.3. Special Maps and Triples 267
Contents ix

§21.4. The End of the Proof 269

Part 6. Dessert

Chapter 22. The Banach–Tarski Theorem 275

§22.1. The Result 275
§22.2. The Schroeder–Bernstein Theorem 276
§22.3. The Doubling Theorem 278
§22.4. Depleted Balls 279
§22.5. The Depleted Ball Theorem 280
§22.6. The Injective Homomorphism 282

Chapter 23. Dehn’s Dissection Theorem 287

§23.1. The Result 287
§23.2. Dihedral Angles 288
§23.3. Irrationality Proof 289
§23.4. Rational Vector Spaces 290
§23.5. Dehn’s Invariant 291
§23.6. Clean Dissections 292
§23.7. The Proof 294

Chapter 24. The Cauchy Rigidity Theorem 295

§24.1. The Main Result 295
§24.2. The Dual Graph 296
§24.3. Outline of the Proof 297
§24.4. Proof of Lemma 24.3 298
§24.5. Proof of Lemma 24.2 301
§24.6. Euclidean Intuition Does Not Work 303
§24.7. Proof of Cauchy’s Arm Lemma 304

Bibliography 309

Index 311
Preface

This book is based on notes I wrote when teaching an undergraduate

seminar on surfaces at Brown University in 2005. Each week I wrote
up notes on a different topic. Basically, I told the students about
many of the great things I have learned about surfaces over the years.
I tried to do things in as direct a fashion as possible, favoring concrete
results over a buildup of theory. Originally, I had written 14 chapters,
but later I added 9 more chapters so as to make a more substantial
book.
Each chapter has its own set of exercises. The exercises are em-
bedded within the text. Most of the exercises are fairly routine, and
advance the arguments being developed, but I tried to put a few
challenging problems in each batch. If you are willing to accept some
results on faith, it should be possible for you to understand the mate-
rial without working the exercises. However, you will get much more
out of the book if you do the exercises.
The central object in the book is a surface. I discuss surfaces
from many points of view: as metric spaces, triangulated surfaces,
hyperbolic surfaces, and so on. The book has many classical results
about surfaces, both geometric and topological, and it also has some
extraneous stuff that I included because I like it. For instance, the
book contains proofs of the Pythagorean Theorem, Pick’s Theorem,

xi
xii Preface

Green’s Theorem, Dehn’s Dissection Theorem, the Cauchy Rigidity

Theorem, and the Fundamental Theorem of Algebra.
All the material in the book can be found in various textbooks,
though there probably isn’t one textbook that has it all. Whenever
possible, I will point out textbooks or other sources where you can
read more about what I am talking about. The various fields of math
surrounding the concept of a surface—geometry, topology, complex
analysis, combinatorics—are deeply intertwined and often related in
surprising ways. I hope to present this tapestry of ideas in a clear
and rigorous yet informal way.
My general view of mathematics is that most of the complicated
things we learn have their origins in very simple examples and phe-
nomena. A good way to master a body of mathematics is to first
understand all the sources that lead to it. In this book, the square
torus is one of the key simple examples. A great deal of the the-
ory of surfaces is a kind of elaboration of phenomena one encounters
when studying the square torus. In the first chapter of the book, I
will introduce the square torus and describe the various ways that
its structure can be modified and generalized. I hope that this first
chapter serves as a good guide to the rest of the book.
I aimed the class at fairly advanced undergraduates, but I tried
to cover each topic from scratch. My idea is that, with some effort,
you could learn the material for the whole course without knowing
too much advanced math. You should be perfectly well prepared for
the intended version of the class if you have had a semester each of
real analysis, abstract algebra, and complex analysis. If you have just
had the first 2 items, you should still be alright, because I embedded
a kind of mini-course on complex analysis in the middle of the book.
Following an introductory chapter, this book is divided into 6
parts. The first 5 parts have to do with different aspects of the theory
of surfaces. The 6th part is a collection of several topics, loosely
related to the rest of the book, which I included because I really like
them. Here is an outline of the book.
Preface xiii

Part 1: Surfaces and Topology. In this part, we define such

concepts as surface, Euler characteristic, fundamental group, deck
group, and covering space. We prove that the deck group of a surface
and its fundamental group are isomorphic. We also prove, under some
conditions, that a space has a universal cover.
Part 2: Surfaces and Geometry. The first 3 chapters in this
part introduce Euclidean, spherical, and hyperbolic geometry, respec-
tively. (In the Euclidean case, which is so well known, we concentrate
on nontrivial theorems.) Following this, we discuss the notion of a
Riemannian metric on a surface. In the final chapter, we discuss
hyperbolic surfaces, as special examples of Riemannian manifolds.
Part 3: Surfaces and Complex Analysis. In this part, we give a
rapid primer on the main points taught in the first semester of com-
plex analysis. Following this, we introduce the concept of a Riemann
surface and prove some results about complex analytic maps between
Riemann surfaces.
Part 4: Flat Cone Surfaces. In this part, we define what is meant
by a flat cone surface. As a special case, we consider the notion of
a translation surface. We show how the “affine symmetry group” of
a translation surface, known as the Veech group, leads right back to
complex analysis and hyperbolic geometry. We end this part with an
application to polygonal billiards.
Part 5: The Totality of Surfaces. In this part, we discuss the
basic objects one considers when studying the totality of all flat or
hyperbolic surfaces, namely moduli space, Teichmüller space, and the
mapping class group. As a warmup for the flat-surface case, we discuss
continued fractions and the modular group in detail.
Part 6: Dessert. In this part, we prove 3 classic results in geome-
try. The Banach–Tarski Theorem says that—assuming the Axiom of
Choice—you can cut up a ball of radius 1 into finitely many pieces
and rearrange those pieces into a (solid) ball of radius 2. Dehn’s
Theorem says that you cannot cut up a cube with planar cuts and re-
arrange it into a regular tetrahedron. The Cauchy Rigidity Theorem
says roughly that you cannot flex a convex polyhedron.
Chapter 1

Book Overview

1.1. Behold, the Torus!

The Euclidean plane, denoted R2 , is probably the simplest of all
surfaces. R2 consists of all points X = (x1 , x2 ) where x1 and x2 are
real numbers. One may similarly define Euclidean 3-space R3 . Even
though the Euclidean plane is very simple, it has the complicating
feature that you cannot really see it all at once: it is unbounded.
Perhaps the next simplest surface is the unit sphere. Anyone who
has played ball or blown a bubble knows what a sphere is. One way
to define the sphere mathematically is to say that it is the solution
set, in R3 , to the equation

x21 + x22 + x23 = 1.

The sphere is bounded and one can, so to speak, comprehend it all

at once. However, one complicating feature of the sphere is that it
is fundamentally curved. Also, its most basic definition involves a
higher-dimensional space, namely R3 .
The square torus is a kind of compromise between the plane and
the sphere. It is a surface that is bounded like the sphere yet flat
like the plane. The square torus is obtained by gluing together the
opposite sides of a square, in the manner shown in Figure 1.1.

1
2 1. Book Overview

Figure 1.1. The square torus

We will not yet say exactly what we mean by gluing, but we say
intuitively that a 2-dimensional being—call it a bug—that wanders
off the top of the square would reappear magically on the bottom, in
the same horizontal position. Likewise, a bug that wanders off the
right side of the square would magically reappear on the left side at
the same vertical position. We have drawn a continuous curve on the
flat torus to indicate what we are talking about. In §3.1 we give a
formal treatment of the gluing construction.
At first it appears that the square torus has an edge to it, but this
is an illusion. Certainly, points in the middle of the square look just
look like the Euclidean plane. A myopic bug sitting near the center
of the square would not be able to tell he was living in the torus.
Consider what the bug sees if he sits on one of the horizontal
edges. First of all, the bug actually sits simultaneously on both hori-
zontal edges, because these edges are glued together. Looking “down-
ward”, the bug sees a little half-disk. Looking “upward”, the bug sees
another little half-disk. These 2 half-disks are glued together and
make one full Euclidean disk. So, the bug would again think that he
was sitting in the middle of the Euclidean plane. The same argument
goes for any point on any of the edges.
The only tricky points are the corners. What if the bug sits
at one of the corners of the squares? Note first of all that the bug
actually sits simultaneously at all 4 corners, because these corners are
1.2. Gluing Polygons 3

all glued together. As the bug looks in various directions, he sees 4

little quarter-disks that glue together to form a single disk. Even at
the corner(s), the bug thinks that he is living in the Euclidean plane.
Modulo a ton of details, we have shown that the square torus has
no edges at all. At every point it “looks locally” like the Euclidean
plane. In particular, it is perfectly flat at every point. At the same
time, the square torus is bounded, like the sphere.
The torus is such a great example that it demands a careful and
rigorous treatment. The first question that comes to mind is What do
we mean by a surface? We will explain this in Chapter 2. Roughly
speaking, a surface is a space that “looks like” the Euclidean plane in
the vicinity of each point. We do not want to make the definition of
“looks like” too restrictive. For instance, a little patch on the sphere
does not look exactly like the Euclidean plane, but we still want the
sphere to count as a surface. We will make the definition of “looks
like” flexible enough so that the sphere and lots of other examples all
count.

1.2. Gluing Polygons

In Chapter 3 we give many examples of surfaces and their higher-
dimensional analogues, manifolds. One of the main tools we use is
the gluing construction. The square torus construction above is the
starting point for a whole zoo of related constructions.

3
2
1

3
2
Figure 1.2. Another torus
4 1. Book Overview

Imagine, for example, that we take the hexagon shown in Figure

1.2 and glue the sides in the pattern shown. What we mean is that
the 2 edges labelled 1 are glued together, according to the direction
given by the arrows, and likewise for the edges labelled 2 and 3. We
can think of Figure 1.2 as a distorted version of Figure 1.1. The
hexagon has a left side, a right side, a top, and a bottom. The top
is made from 2 sides and the bottom is made from 2 sides. The left
and right sides are glued together and the top is glued to the bottom.
The resulting surface retains some of the features of the flat torus:
a bug walking around on it would not detect an edge. On the other
hand, consider what happens when the bug sits at the point of the
surface corresponding to the white dots. Spinning around, the bug
would notice that he turns less than 360 degrees before returning
to his original position. What is going on is that the sum of the
interior angles at the white dots is less than 360 degrees. Similarly,
the bug would have to spin around by more than 360 degrees before
returning to his original position were he to sit at the point of the
surface corresponding to the black points. So, in general, the bug
would not really feel like he was living in the Euclidean plane. Our
general definition of surfaces and gluing will be such that the example
we gave still counts as a surface.
Figure 1.3 shows an example based on the regular octagon, in
which the opposite sides of the octagon are glued together.
1

4 2

3 3

2 4
1
Figure 1.3. Gluing an octagon together
1.3. Drawing on a Surface 5

This example is similar to the square torus, except that this time
8 corners, rather than 4, are glued together. A myopic bug sitting
anywhere on the surface except at the point corresponding the 8 cor-
ners might think that he was sitting in the Euclidean plane. However,
at the special point, the bug would have to turn around 720 degrees
(or 6π radians) before returning to his original position. We will an-
alyze this surface in great detail. One can view it as the next one in
the sequence that starts out sphere, torus, . . . . At least for this intro-
ductory chapter, we will call it the octagon surface. (It is commonly
called the genus 2 torus.) We can construct similar examples based
on regular 2n-gons, for each n = 5, 6, 7 . . . .

1.3. Drawing on a Surface

Once we have defined surfaces and given some examples, we want
to work with them to discover their properties. One natural thing
we can do is divide a surface up into smaller pieces and then count
them. Figure 1.4 shows 2 different subdivisions of the square torus
into polygons. We have left off the arrows in the diagram, but we
mean for the left/right and top/bottom sides to be glued together.

Figure 1.4. Dividing the torus into faces

In the first subdivision, there are 4 faces, 8 edges, and 4 vertices.

It first appears that there are more edges, but the edges around the
boundary are glued together in pairs. So each edge on the bound-
ary only counts for half an edge. A similar thing happens with the
6 1. Book Overview

vertices. We make the count

faces − edges + vertices = 4 − 8 + 4 = 0.

In the second example, we get the count

faces − edges + vertices = 8 − 14 + 6 = 0.

The same result holds for practically any subdivision of the square
torus into polygons. This result is known as the Euler formula for
the torus. We discuss this formula in more detail in §3.4.
You can probably imagine that you would get the same result for
a torus based on a rectangle rather than a square. Likewise, we get
the same result for the surface based on the hexagon gluing in Figure
1.2. All these surfaces have an Euler characteristic of 0.
Things turn out differently for the sphere. For instance, thinking
of the sphere as a puffed-out cube, we get the count

faces − edges + vertices = 6 − 12 + 8 = 2.

Thinking of the sphere as a puffed-out tetrahedron, we get the count

faces − edges + vertices = 4 − 6 + 4 = 2.

Thinking of the sphere as a puffed-out icosahedron, we get the count

faces − edges + vertices = 20 − 30 + 12 = 2.

The Euler formula for the sphere says that the result of this count
is always 2, under very mild restrictions. You can probably see that
we would get the same result for any of the “sphere-like” surfaces
mentioned above.
Were we to make the count for any reasonable subdivision of the
octagon surface, we would get an Euler characteristic of −2. Can
you guess the Euler characteristic, as a function of n, for the surface
obtained by gluing together the opposite sides of a regular 2n-gon?
Another thing we can do on a surface is draw loops—meaning
closed curves—and study how they move around. The left side of
Figure 1.5 shows 3 different loops on the square torus.
1.3. Drawing on a Surface 7

Figure 1.5. Loops on the torus

One of the loops, the one represented by the thick vertical line,
is different from the others. Imagine that these loops are made from
rubber bands, and are allowed to compress in a continuous way. The
first 2 loops can shrink continuously to points, whereas the third loop
is “stuck”. It can’t make itself any shorter no matter how it moves.
Such a loop is commonly called essential . There are many essential
loops on the torus. The right side of Figure 1.5 shows another essential
loop. In contrast, the sphere has no essential loops at all.
We will see in Chapter 4 that there is an algebraic object we can
associate to a surface (and many other kinds of spaces) called the fun-
damental group. The fundamental group organizes all the different
ways of drawing loops on the surface into one basic structure. The
nice thing about the fundamental group is that it links the theory
of surfaces to algebra, especially group theory. Beautifully, it turns
out that 2 (compact) surfaces have the same Euler characteristic if
and only if they have the same fundamental group. The Euler char-
acteristic and the fundamental group are 2 entry points into the vast
subject of algebraic topology.
For the most part, studying algebraic topology is beyond the
scope of this book, but we will study the fundamental group and
related constructions, in great detail. After defining the fundamental
group in Chapter 4, we will compute a number of examples in Chapter
5.
8 1. Book Overview

1.4. Covering Spaces

There is a nice way to unwrap the essential loops on a torus. The
idea is that we remember that the square torus is made from a square,
which we think of as the unit square with vertices (0, 0), (0, 1), (1, 0)
and (1, 1). We draw a line segment in the plane that starts out at
the same point as the loop and has the same length. We think of this
path starting at the point (0, 0). Figure 1.6 shows an example. In
this example, the unwrapped path joins (0, 0) to (3, 2).
The process can be reversed. Starting with a line segment that
joins (0, 0) to (m, n), a point with integer coordinates, we can wrap the
segment around the torus so that it makes an essential loop. In fact,
the essential loops that start at (0, 0) are, in the appropriate sense, in
one-to-one correspondence with the points of Z 2 , the integer grid in
the plane. The basic result is that any 2 essential loops L1 and L2 ,
corresponding to points (m1 , n1 ) and (m2 , n2 ), can be continuously
moved, one into the other, if and only if (m1 , n1 ) = (m2 , n2 ).
(3,2)

(0,0)
Figure 1.6: Unwrapping a loop on the torus

As we will explain in Chapter 6 and Chapter 7, this unwrapping

construction can be done for any surface. In the case of the torus, we
see that the (equivalence classes of) essential simple loops are in exact
correspondence with the points of the integer grid in the plane. One
might wonder if a similarly nice picture exists in general. The answer
is “yes”, and in fact the picture becomes more interesting when we
1.5. Hyperbolic Geometry and the Octagon 9

consider surfaces, such as the octagon surface. However, in order to

“see” the picture in these cases, you have to draw it in the possibly
unfamiliar world of hyperbolic geometry. The idea is that hyperbolic
geometry does for the octagon surface (and most other surfaces as
well) what Euclidean geometry does for the square torus and what
spherical geometry does for the sphere.
We will discuss Euclidean, spherical, and hyperbolic geometry in
Chapter 8, Chapter 9, and Chapter 10 respectively. Our main goal is
to understand how these geometries interact with surfaces, but we will
also take time out to prove some classical geometric theorems, such as
Pick’s Theorem (a relative of the Euler formula) and the angle-sum
formula for hyperbolic and spherical triangles.
The Euclidean, spherical, and hyperbolic geometries are the 3
most symmetrical examples of 2-dimensional Riemannian geometries.
To put the 3 special geometries into a general context, we will discuss
Riemannian geometry in Chapter 11.

1.5. Hyperbolic Geometry and the Octagon

Now let us return to the question of unwrapping essential loops on
the octagon surface. The octagon surface looks a bit less natural than
the square torus, thanks to the special point. However, it turns out
that the octagon surface “wears” hyperbolic geometry very much in
the same way that the square torus “wears” Euclidean geometry.
We already mentioned that we will study hyperbolic geometry in
detail in Chapter 10. Here we just give the barest of sketches, in order
to give you a taste of the beauty that lies in this direction. One of the
many models for the hyperbolic plane is the open unit disk. There is
a way to measure distances in the open unit disk so that the shortest
paths between points are circular arcs that meet the boundary at
right angles. These shortest paths are known as geodesics. The left-
hand side of Figure 1.7 shows some of the geodesics in the hyperbolic
plane. The boundary of the unit disk is not part of the hyperbolic
plane and the lengths of these geodesics are all infinite. A bug living
in the hyperbolic plane would see it as unbounded in all directions.
10 1. Book Overview

Figure 1.7. Gluing the octagon together

The hyperbolic plane shares many features with the Euclidean

plane. There is a unique geodesic joining any 2 distinct points, and
any 2 distinct geodesics meet in at most one point. Furthermore,
the hyperbolic plane is totally symmetric, in the sense that every
point and every direction looks exactly the same. A bug living in an
otherwise empty hyperbolic plane would not be able to tell where he
was.
On the other hand, the hyperbolic plane and the Euclidean plane
have some important differences. For instance, the sum of the angles
of a hyperbolic triangle, a shape bounded by 3 geodesic segments, is
always less than 180 degrees, or π radians. (When we discuss angles
in radians, we will often leave off the word “radians”.) Similarly,
the individual interior angles of a regular octagon can take on any
value less than 3π/8, which is the value in the Euclidean case. The
right hand side of Figure 1.4 shows a regular hyperbolic octagon. We
decrease the interior angles by making the octagon larger and we
increase the interior angles by making the octagon smaller.
In particular, we can adjust the size of the regular octagon so that
the interior angles are exactly π/8. We can then cut the resulting
octagon out of the hyperbolic plane and glue the sides together just
as in Figure 1.3. From the hyperbolic geometry point of view, the
resulting surface would be completely seamless: a myopic bug living
on the surface could not tell that he was not living in the hyperbolic
plane. With the chosen interior angles, the 8 corners fit together like
1.6. Complex Analysis and Riemann Surfaces 11

8 slices in a pizza to make a perfect hyperbolic disk. We will consider

this construction in detail in Chapter 12.
A similar construction can be made for the surfaces obtained
by gluing together the opposite sides of a regular 2n-gon, for each
n = 5, 6, 7 . . . . All these surfaces “wear” hyperbolic geometry in a
seamless way, just like the square torus “wears” Euclidean geometry.
Now, we can tile the Euclidean plane by copies of the unit square.
The vertices of this tiling are precisely the integer grid points. In the
same way, we can move our hyperbolic octagon around the hyperbolic
plane and tile the hyperbolic plane with copies of it. When drawn in
the disk model, the picture looks like the drawings in M. C. Escher’s
Circle Woodcut series. To our Euclidean eyes, the octagons appear
to get smaller as they move out toward the boundary of the disk.
However, in the hyperbolic world, the various octagons all have the
same size.
The vertices of this tiling are a kind of hyperbolic geometry ver-
sion of the integer grid. These points are in one-to-one correspondence
with the equivalence classes of essential loops on the octagon surface.
The same kind of thing works for the surfaces corresponding to the
(2n)-gons for n = 4, 5, 6 . . . . In fact, such a construction works for
all surfaces that have negative Euler characteristic: one always gets
a grid of points in the hyperbolic plane that names the different es-
sential loops on the surface. We will explore this in detail in Chapter
12.

1.6. Complex Analysis and Riemann Surfaces

It turns out that there is a single kind of geometry which unifies
Euclidean, spherical, and hyperbolic geometry. This geometry, called
Möbius or conformal geometry, takes place in the Riemann sphere.
The Riemann sphere is the set C ∪ ∞, where C is the complex plane
and ∞ is an extra point that is added. For starters,

• The Euclidean plane is identified with C.

• The hyperbolic plane is identified with the open disk {z ∈
C| z < 1}.
12 1. Book Overview

• The sphere is identified with all of C ∪ ∞, via stereographic

projection
   
x1 x2
(x1 , x2 , x3 ) → + i, (0, 0, 1) → ∞.
1 − x3 1 − x3

See §9.5 for details on stereograpic projection.

Once these identifications are made, the symmetries of the rele-

vant objects are all given by maps of the form

az + b
(1.1) z→ , a, b, c, d ∈ C, ad − bc = 1.
cz + d
We will discuss these maps in more detail in §10.1. The point ∞
is added so that when the expression in equation (1.1) looks like
“something over 0”, we define it to be ∞. Various conditions are
placed on the coefficients a, b, c, d to guarantee that the relevant set—
e.g., the unit disk—is preserved by the map.
These kinds of transformations are called linear fractional, or
Möbius, transformations. The Möbius transformations are prototyp-
ical examples of complex analytic functions. These are continuous
maps from C to C which have the additional property that their
matrix of partial derivatives, at each point, is a similarity—i.e., a
rotation followed by a dilation. This constraint on the partial deriva-
tives leads to a surprisingly rich family of functions and this is the
subject of complex analysis. In Chapter 13, we will give a rapid
overview of basic complex analysis, with a view towards its applica-
tion to surfaces. In Chapter 14 and Chapter 15 we will discuss some
special complex analytic functions in detail.
Going back to our polygon gluing construction, we can view sur-
faces as being made out of pieces of C that have been glued together.
This point of view leads to the notion of a Riemann surface, as we
explain in Chapter 16. One can think of a Riemann surface as a sur-
face that “wears” C in the same seamless way that the square torus
“wears” Euclidean geometry or the octagon surface “wears” hyper-
bolic geometry. Once we have the notion of a Riemann surface, we
can “do complex analysis on it” in much the same way that one can
do complex analysis in C or in C ∪ ∞.
1.7. Cone Surfaces and Translation Surfaces 13

The complex analysis point of view on a surface at first seems

rather remote from the geometric point of view discussed above, but in
fact they are quite similar. The close connection comes from the fact
that the Möbius transformations play a distinguished role amongst
the complex analytic functions. One example of this is the following
result, known as the Schwarz–Pick Theorem:
Theorem 1.1. Let f be a complex analytic function from the unit
disk to itself. If f is one-to-one and onto, then f is a Möbius trans-
formation (and hence a hyperbolic isometry).

Theorem 1.1 is part of a larger theorem, called the Poincaré

Uniformization Theorem. The Uniformization Theorem gives a com-
plete equivalence between the Euclidean/spherical/hyperbolic geom-
etry points of view of surfaces and the Riemann surface point of view.
The proof of this result is beyond the scope of our book, but in Chap-
ter 16 we will at least explain the result and its ramifications.

1.7. Cone Surfaces and Translation Surfaces

We have mentioned several times that the octagon surface does not
“wear” Euclidean geometry as well as the square torus does, and we
have taken some pains to explain how one can profitably view the
octagon surface with hyperbolic geometry eyes. However, in Chapter
17 we come full circle and consider the octagon surface and related
surfaces from the Euclidean geometry point of view.
Suppose, as in Figure 1.2 above, we glue together the sides of
a polygon in such a way that the sides in each pair of glued sides
have the same length. The resulting surface has the property that
it is locally indistinguishable from the Euclidean plane, except at
finitely many points. At these finitely many points, a bug living in the
surface would notice some problem related to spinning around, as we
discussed above. These special points are cone points. A Euclidean
cone surface is a surface that is flat except at finitely many cone
points.
When we discussed the “torus-like” surface defined in connection
with Figure 1.2, we mentioned the spinning-around problem a bug
would face when sitting at the 2 special points. At one of the special
14 1. Book Overview

points, the bug needs to spin more than 2π, say 2π + δ1 , before re-
turning to his original position. At the other special point, the bug
needs to spin less than 2π, say 2π −δ2 , before returning to his original
position. The numbers δ1 and −δ2 might be called the angle error at
the special points.
The numbers δ1 and δ2 depend on the hexagon in question. As
one can see by adding up the interior angles of a hexagon, we have
δ1 = δ2 . That is, the total angle error is 0. This result holds for any
Euclidean cone surface with Euler characteristic 0. More generally,
on a surface with Euler characteristic χ, the total angle error is 2πχ.
This result, known as the combinatorial Gauss–Bonnet Theorem is
one of the main results of Chapter 17.
Another topic in Chapter 17 is the application of Euclidean cone
surfaces to polygonal billiards. It turns out that the contemplation of
rolling a frictionless, infinitesimally small billiard ball around inside a
polygonal shaped billard table, whose angles are all rational multiples
of π, leads naturally to a certain Euclidean cone surface. One can
profitably study this surface to get information about how billiards
would work out in the polygon.
The Euclidean cone surfaces associated to polygonal billiards have
a special structure. They are called translation surfaces. A transla-
tion surface is a Euclidean cone surface, all of whose angle errors are
integer multiples of π. The square torus is the prototypical example
of a translation surface, but it is a bit too simple of an example in
this case. The octagon surface provides a better example. The oc-
tagon surface, considered from the Euclidean geometry perspective,
is a translation surface. This surface has a single cone point, and
the angle error there is 4π. Translation surfaces are nicer than gen-
eral Euclidean cone surfaces for a variety of reasons. One reason is
that, as it turns out, it is possible to speak about directions (such as
due north) on a translation surface without any ambiguity. We will
discuss these surfaces in detail in Chapter 18.

1.8. The Modular Group and the Veech Group

We have wandered away from hyperbolic geometry and complex anal-
ysis, but actually hyperbolic geometry and complex analysis are very
1.8. The Modular Group and the Veech Group 15

closely related to the subject of translation surfaces. Once again, let

us consider the square torus. A linear transformation of the form
(1.2)
T (x, y) = (ax + by, cx + dy), a, b, c, d ∈ Z, ad − bc = 1.
acts as transformation of the square torus, via the following 4-step
process:
(1) Start with a point p in the square torus.
(2) Choose a point (x, y) such that p represents the collection
of points glued to (x, y).
(3) Subtract off integer coordinates of T (x, y) until the result
(x , y  ) lies in the unit square.
(4) The image of the map is p , the point that names the col-
lection of points glued to (x , y  ).
Any ambiguity in the process that takes us from p to p is absorbed
by the gluing process.
So, any integer 2 × 2 matrix with determinant 1 gives rise to
a transformation of the square torus that, on small scales, is indis-
tinguishable from a linear transformation. The set of all such maps
forms a group known as modular group. The maps in equation (1.2)
have the same form as the Möbius transformations discussed above.
Interpreting the maps in equation (1.2) as Möbius transformations in-
stead of linear transformations, we can interpret the modular group
as a group of symmetries of the hyperbolic plane.
The modular group is an object of great significance in mathe-
matics, and we cannot resist exploring some of its properties that are
not, strictly speaking, directly related to surfaces. For instance, in
Chapter 19 we will discuss continued fractions and their connection
to the modular group and hyperbolic geometry. In Chapter 22 we will
see that the modular group is the main ingredient in the proof of the
Banach–Tarski Theorem. The Banach–Tarski Theorem says in par-
ticular that, assuming the axiom of choice, one can cut the unit ball
in R3 into finitely many pieces and rearrange these pieces so that they
make a solid ball of radius 100000. Though this result seems a bit far
removed from the theory of surfaces, it is quite beautiful and it shows
how objects such as the modular group pop up all over mathematics.
16 1. Book Overview

Getting back to translation surfaces, we will see in Chapter 18

that one can associate to any translation surface a group of symme-
tries of the hyperbolic plane. This group is known as the Veech group
of the translation surface. It often happens that the Veech group is
trivial, or very small, but for many special examples the Veech group
is large and beautiful. For instance, the Veech group associated to the
regular octagon surface is closely related to a tiling of the hyperbolic
plane by triangles having angles 0, 0, and π/8. One of the highlights
of Chapter 18 is a discussion of (essentially) this example.

1.9. Moduli Space

The square torus is not the only translation surface without any cone
points. In Chapter 20 we consider the family M unit area parallelo-
gram, in the same pattern as in Figure 1.1. Essentially the same anal-
ysis we made in connection with Figure 1.1 can be made in connection
with any surface in our family. All these surfaces are seamlessly flat
at each point. A myopic bug on any of these surfaces would not be
able to tell that he was not in the Euclidean plane.
On the other hand, these various surfaces are typically not the
same geometrically. For instance, a surface made from a long thin
rectangle obviously has diameter greater than the diameter of the
square torus. Similarly, such a surface has a very short essential loop
whereas all essential loops on the square torus have length at least 1.
We can consider the family M as a space in its own right. Each point
of M corresponds to a different flat torus. This space M is known as
the moduli space of flat tori. We will discuss M and related objects
in Chapter 20.
Amazingly, M turns out to be a surface in its own right, and
(with the exception of 2 special points) this surface is modelled on
hyperbolic geometry! Just to repeat: the space of all tori made by
gluing together unit area parallelograms turns out to be a surface that
naturally wears hyperbolic geometry (away from 2 special points).
One of the special points in M corresponds to the square torus, and
the other one corresponds to the surface obtained by gluing together
the opposite sides of a rhombus made from 2 equilateral triangles.
Referring to the discussion of covering spaces above, we can consider
1.10. Dessert 17

the grid in the hyperbolic plane associated to M. It turns out that

the modular group acts as a group of symmetries of this grid. So,
when we consider the moduli space M of unit area flat tori, we get
right back to the modular group.
We can play a similar game for the octagon surface. As we dis-
cussed above, we can create the octagon surface using a suitable cho-
sen regular octagon. However, we can also glue together other hyper-
bolic octagons to produce a surface that “looks hyperbolic” at each
point and has the same Euler characteristic. When we consider the
totality of such surfaces, we arrive at a higher-dimensional generaliza-
tion of M, also called moduli space. This higher-dimensional space
is not a surface, but it does share some features in common with a
hyperbolic surface.
In Chapter 20 we also discuss Teichmüller space, the space that
relates to the higher-dimensional version of M in the same way that
the hyperbolic plane relates to M. Teichmüller space shares some
features with the hyperbolic plane, but is much more mysterious and
somewhat less symmetric. We will discuss the group of symmetries
of Teichmüller space, called the mapping class group. The mapping
class groups relate to the surfaces of negative Euler characteristic in
the same way that the modular group relates to the square torus. We
will further explore Teichmüller space in Chapter 21.

1.10. Dessert
There are a few topics in this book that I simply threw in because
I like them. Chapter 22 has a proof of the Banach–Tarski Paradox.
One nice thing about the proof is that it involves the modular group
in an essential way. So, in a strange way, the Banach–Tarski Paradox
has some connection to hyperbolic geometry.
Chapter 23 has a proof of Dehn’s Dissection Theorem, which says
that one cannot cut a cube into finitely many pieces, using planar
cuts, and rearrange the result into a regular tetrahedron. This result
serves as a kind of foil for the decomposition methods we use to
prove the combinatorial Gauss–Bonnet Theorem and other results.
Polyhedral decomposition is quite robust in 2 dimensions, but not in
higher dimensions.
18 1. Book Overview

Chapter 24 has a proof of the Cauchy Rigidity Theorem. This

result says that there at most one way to snap together a given collec-
tion of convex polygons to produce a convex polyhedron. The proof
involves some spherical geometry and also the combinatorial Gauss–
Bonnet Theorem.
Part 1

Surfaces and Topology

Chapter 2

Definition of a Surface

We discussed surfaces informally in the previous chapter, and now

the time has come to give a formal definition of a surface. Here is the
main definition.

Definition 2.1. A surface is a metric space X such that every point

in X has a neighborhood which is homeomorphic to the plane.

Don’t worry if you don’t know what some of the words in the above
definition mean. The point of this chapter is to explain what they
mean. At the end of the chapter, we will say a few words about
higher-dimensional surfaces, called manifolds.

2.1. A Word about Sets

A set is an undefined notion for us. Informally, a set is a collection
of things, called elements. A book on set theory, such as [DEV], will
tell you all about sets. You should be familiar with such sets as

• Z, the integers.
• N = {1, 2, 3, . . . }, the natural numbers.
• R, the real numbers.

21
22 2. Definition of a Surface

A map between sets A and B is a rule, say f , which assigns to each

element a ∈ A, an element b = f (a) ∈ B. This is usually written
as f : A → B. The map f is one-to-one if f (a1 ) = f (a2 ) implies
that a1 = a2 . The map f is onto if the set {f (a)| a ∈ A} equals B.
The map f is a bijection if it is both one-to-one and onto. Two sets
are bijective if there is some bijection between them. All the sets we
consider will be bijective to either a finite set, or N, or R.
The product A × B of sets is the set of ordered pairs (a, b) with
a ∈ A and b ∈ B. In particular, R2 = R × R is the plane.

2.2. Metric Spaces

A metric space is a set X together with a map d : X × X → R
satisfying the following properties:
• Nondegeneracy. d(x, y) ≥ 0 for all x, y, with equality iff
x = y.
• Symmetry. d(x, y) = d(y, x) for all x, y.
• Triangle Inequality. d(x, z) ≤ d(x, y) + d(z, y) for all x, y, z.
d is called a metric on X. Note that the same set can have many
different metrics.
Here is the most boring example of a metric space. Given any set
X define d(x, y) = 0 if x = y and d(x, y) = 1 if x = y. This is called
the discrete metric on X.
Exercise 1. Let X = R2 , the plane. Define the dot product
V · W = v1 w1 + v2 w2 .
Here V = (v1 , v2 ) and W = (w1 , w2 ). Also define
√
V  = V · V .
Finally, define d(V, W ) = V − W . Prove that d is a metric on R2 .
The metric in this exercise is known as the Euclidean metric on R2 ,
or else the standard metric.

If X is a metric space and Y ⊂ X is a subset, then the metric on

X automatically defines a metric on Y , by restriction. For instance,
any subset of the plane automatically can be interpreted as a metric
2.3. Open and Closed Sets 23

space, using the metric from Exercise 1.

Exercise 2. On Z define d(m, n) = 2−k , where k is such that 2k

divides |m − n| but 2k+1 does not. Also define d(m, m) = 0. For
instance, d(3, 7) = 2−2 = 1/4 because 22 divides 4 but 23 = 8 does
not. Prove that d is a metric on Z. This metric is called the 2-adic
metric. It is quite different from the usual metric on the integers.

2.3. Open and Closed Sets

Let X be a metric space with metric d. An open ball in X is a subset
of the form
{x| d(x, c) < r}.
Here c is the center of the ball and r is the radius. Say that a sub-
set U ⊂ X is open if for every point x ∈ U there is some open ball
Bx such that x ∈ Bx and Bx ⊂ U . Note that open balls are open sets.

Exercise 3. Prove that the intersection of two open sets is open.

Prove also that the arbitrary union of open sets is open.

Here is some vocabulary, which will be familiar to you if you have

had a real analysis class:
• The notation X − A means the complement of A in X,
namely the set of points in X which are not in A.
• Given a point x ∈ X, a neighborhood of x is any open subset
U ⊂ X such that x ∈ U . For instance, the ball of radius r
about x is a perfectly good neighborhood of x.
• The interior of a set A ⊂ X is the union of all open subsets
of A. By Exercise 3, the interior of a set is open. Sometimes
the interior of A is denoted as Ao . Put another way Ao is
the largest open set contained in A.
• A set C ⊂ X is closed if X − C is open.
• The closure of a set A is the set

A = X − (X − A)o .
24 2. Definition of a Surface

Put another way, A is the smallest closed set which contains

A.
• The boundary of A is the set
∂A = A − Ao .
• A set A ⊂ X is dense if A = X. For instance, the set of
rational numbers is dense in the set of real numbers.

2.4. Continuous Maps

A map between metric spaces is just a map in the set theoretic sense.
There are two equivalent definitions of continuity for maps between
metric spaces. The first one is much cleaner but the second one is
probably more familiar.

Definition 2.2. The map f : X → Y is continuous if it has the

following property: For any open V ⊂ Y the set
U = f −1 (V ) := {x| f (x) ∈ V }
is an open set of X.

Definition 2.3. First, f is continuous at x ∈ X if, for any  > 0, there

is some δ > 0 such that dX (x, x ) < δ implies that dY (f (x), f (x )) < .
Here dX is the metric on X and dY is the metric on Y . Then f is
continuous on X if f is continuous at each point of X.

Exercise 4. Show that the two definitions of continuity coincide.

Now let X, Y, Z all be metric spaces. Let f : X → Y be a map,

and let g : Y → Z be map. The composition h = g ◦ f is defined as
h(x) = g(f (x)). So h is a map from X to Z.

Lemma 2.1. The composition of continuous maps is continuous.

Proof. Definition 2.2 works much better for this. Let W be an open
subset of Z. Our goal is to show that h−1 (W ) is open in X. Note
that h−1 (W ) = f −1 (V ), where V = g −1 (W ). Since g is continuous,
2.5. Homeomorphisms 25

V is open. Since f is continuous and V is open, U is open. This

works for any choice of open W , so we are done. 

Exercise 5. Give an example of metric spaces X and Y , and f :

X → Y such that
• f is a bijection.
• f is continuous.
• f −1 (the inverse map) is not continuous.
This is a classic problem.

2.5. Homeomorphisms
Let X and Y be two metric spaces. A map h : X → Y is a home-
omorphism if h is a bijection and both h and h−1 are continuous.
Compare Exercise 5. The spaces X and Y are said to be homeomor-
phic if there is some homeomorphism from X to Y . Intuitively, two
sets are homeomorphic if one can be “warped” into the other one.
Often we do not care exactly which metric we are using, but we just
bring in the metric to be able to talk about things like continuity and
open sets. Another way to “throw out the metric” is to introduce
the notion of a topological space. In some ways topological spaces are
easier to work with than metric spaces and more flexible, but they
are more abstract. If you are interested in this, check out a book on
point-set topology, such as [MUN].
Even though sets might look very different to the eye, they might
be homeomorphic. The next exercise gives some examples of this.

Exercise 6. Prove that the following subsets of the plane (with

the standard metric) are all homeomorphic to each other:
• An open ball.
• The interior of a (filled-in) triangle.
• The plane itself.
Exercise 7. We can give R the standard metric d(x, y) = |x − y|.
Prove that R is not homeomorphic to R2 , with its standard metric.
26 2. Definition of a Surface

Exercise 8 (Challenge). You can imitate the construction in Ex-

ercise 1 to put a metric on R3 , 3-dimensional space. Prove that R2
is not homeomorphic to R3 . As it turns out Rm and Rn are homeo-
morphic if and only if m = n. When you try to prove something like
this, you start getting into algebraic topology.

2.6. Compactness
We will sometimes use the notion of compactness. Say that an open
covering of a metric space X is a collection {Uα } of open sets in X
whose union is X. Say that a subcovering of a given covering is a sub-
collection that still covers X. Say that a finite subcover is a subcover
that only has finitely many elements in it.

Definition 2.4. A metric space X is compact if every covering of X

has a finite subcover.

The notion of compactness is easier to understand for subsets

of Euclidean space. When X is a subset of Euclidean space, X is
compact if and only if X is closed and contained in some ball. This
result is known as the Heine–Borel Theorem.
The original definition of compactness is perfectly adapted to the
notion of continuous maps. Suppose that X and Y are homeomorphic.
Then X is compact if and only if Y is compact. Here we prove one
result which indicates the power of the definition of compactness.

Lemma 2.2. Suppose that f : X → R is a continuous function. If

X is compact, then f is bounded.

Proof. Let Un = f −1 (−n, n). Since f is continuous, the set Un is

open. Evidently, the collection {Un } covers X. Since X is compact,
there is some finite list of these sets which also covers X. Letting
UN be the largest of these finitely many sets, we see that |f | ≤ N on
X. 
2.7. Surfaces 27

2.7. Surfaces
Now let’s go back to Definition 2.1. Let X be a surface. This means,
first of all, that X is a metric space. So, it makes sense to talk about
open and closed sets on X and also continuous functions from X
to other metric spaces. What makes X a surface is that each point
x ∈ X has an open neighborhood U such that U is homeomorphic to
R2 . You should picture U as a little open disk drawn around x. So X
has the property that, around every point, it “looks like” the plane.
This is how we make sense of the discussion at the end of §1.1.

Exercise 9. The unit sphere S 2 in R3 is the set {(x, y, z)| x2 +

y 2 + z 2 = 1}. This set inherits a metric from R3 . Prove that S is
a surface, according to our definition. So, for each point x ∈ S you
need to find an open subset Ux ⊂ S and also a map fx : Ux → R2
which is a homeomorphism. (Hint: Try to use symmetry to reduce
the problem to showing that just one point in S 2 has the desired
neighborhood.)

Exercise 10. Consider the following subset of R4 :

T 2 = {(x, y, z, w)| x2 + y 2 = 1; z 2 + w2 = 1}.

This set inherits a metric from R4 . You might recognize T 2 as the

product of two circles. Prove that T 2 is a surface. This surface is
known as a torus. (Hint: Again, try to use symmetry.) Once we
make sense of the gluing construction, we will see that T 2 is homeo-
morphic to the square torus discussed in the previous chapter.

Figure 2.1: A torus

In the coming chapters, we will construct many more examples

of surfaces besides the ones in Exercises 9 and 10.
28 2. Definition of a Surface

2.8. Manifolds
A manifold is essentially a higher-dimensional surface. Though this
book is about surfaces, I am including a section about manifolds in
case you are curious about them. If you just want to learn about
surfaces, you can safely skip this section.

Definition 2.5: An n-dimensional manifold is a metric space, such

that every point has a neighborhood which is homeomorphic to Rn .

Technical Comment. This definition of a manifold is slightly non-

standard. The usual definition replaces metric space with Hausdorff
topological space. However, in most cases the metric space definition
singles out the same objects as manifolds. The reason we are using
the metric space definition is that it is more concrete.

I will give a nice example of a manifold at the end of this section,

but first I will introduce a general tool for producing manifolds. The
tool is the Implicit Function Theorem, a classic result from multi-
variable calculus. The full Implicit Function Theorem and its proof
can be found in practically any book on advanced calculus; e.g., see
[SPI]. We will prove a special case below, a case that is fairly easy
to prove yet still produces nice examples.
Let f : Rn+1 → R be a continuous function. Assume also that
the partial derivatives of f exist and are continuous functions. This
means that the gradient ∇f exists and is continuous. Say that 0 is a
regular value for f if it never happens that both f (x1 , . . . , xn+1 ) = 0
and ∇f (x1 , . . . , xn+1 ) = 0 at the same point.

Theorem 2.3. If 0 is a regular value for f , then f −1 (0) is an n

dimensional manifold.

Proof. Let S = f −1 (0). First of all, S is a metric space: The distance

between any two points in S is defined to be their Euclidean distance
in Rn . It remains to check that every point in S has a neighborhood
that is homeomorphic to Rn .
2.8. Manifolds 29

Let p = (x1 , . . . , xn+1 ) ∈ S be an arbitrary point. We know that

∇f (p) is nonzero. If we rotate and scale space and replace f by a
constant multiple of f , we do not change S at all. So, without loss of
generality, we can assume that
p = (0, . . . , 0); ∇f (p) = (0, . . . , 0, 1).
Let P = Rn × {0}. We think of (0, . . . , 0, 1) as the vertical direction
and P as the horizontal directions. See Figure 2.2 below.
Let Q denote the open cube of diameter  centered at (0, . . . , 0).
We call a line segment special if it has one endpoint on the bottom
face of Q and one endpoint on the top face of Q. Since ∇f varies
continuously, we can choose  small enough so that f increases along
any special segment, if we move along it from the bottom to the top.
Let U = Q ∩ S. Then U is an open neighborhood of p in S. It suffices
to show that U is homeomorphic to an open cube in Rn , since an
open cube in Rn is homeomorphic to Rn itself.

Q
S
P

(0,0,0)

Figure 2.2. Putting a cube around S

Now, Q ∩ P is an open cube, and the map h(x1 , . . . , xn+1 ) =

(x1 , . . . , xn , 0) ss a map from U to Q ∩ P . We just have to show that
h is a homeomorphism. Here are the main points.
• h is a distance decreasing map so (using the  − δ definition
of continuity) h is continuous.
• Each vertical line intersects S at most once, because f in-
creases as we move upward along a vertical line. Hence, h
is one-to-one.
• We can connect any point on the top face of Q to (0, . . . , 0)
by “half” of a special segment. Since f (0, . . . , 0) = 0, and f
30 2. Definition of a Surface

increases along special segments, f is positive on the top face

of Q. Similarly, f is negative on the bottom face of Q. Since
f increases along vertical segments, we have f = 0 some-
where on each vertical segment, by the intermediate value
theorem. So, every vertical segment intersects S. Hence h
is onto.
• Suppose that X1 and X2 are two points in Q ∩ P that are
very close together. Consider the last coordinates z1 and z2
of h−1 (X1 ) and h−1 (X2 ), respectively. If z1 and z2 are too
far apart, then we can join the points (X1 , z1 ) and (X2 , z2 )
by part of a special segment. Since both these points lie in
S, we have a contradiction. This shows, a bit informally,
that h−1 is continuous.
We have succeeded in showing that an arbitrary point of S has a
neighborhood which is homeomorphic to Rn . 

Now we give a nice example of a 3-dimensional manifold. You

can think of the set of 2 × 2 (real valued) matrices as a copy of R4 .
There is a nice map from this space into R, namely the determinant
(minus 1):  
a b
f = ad − bc − 1.
c d

Exercise 11. Show that 0 is a regular value for f .

In the above example, f −1 (0) is usually denoted by SL2 (R).

Thus SL2 (R) is the set of unit determinant real 2 × 2 matrices. By
Theorem 2.3, the space SL2 (R) is a 3-dimensional manifold. A sim-
ilar argument shows that SLn (R), the set of unit determinant n × n
matrices, is a manifold of dimension n2 − 1. The space SLn (R) is
an example of a manifold which is also, and in a compatible way, a
group. Such objects are called Lie groups. The book [CHE] is a
classic reference on this subject; see also [TAP].
Chapter 3

The Gluing
Construction

The purpose of this chapter is to explain the gluing construction dis-

cussed informally in Chapter 1. This construction is usually done for
topological spaces, but it can be done for metric spaces as long as we
are a bit careful. The advantage to using topological spaces is that
the construction always works. The disadvantage to using topological
spaces is that it takes a long time to figure out what the construction
actually means. For metric spaces, things don’t always work out, but
whatever happens is more understandable. Also, for our purposes,
things always work out.

3.1. Gluing Spaces Together

Before we start, we need to recall the notion of the infimum from
real analysis. Let S ⊂ R be a set consisting entirely of nonnegative
numbers. Then x = inf S denotes the smallest member of the closure
of S. Such a number always exists and is unique. The existence (and
uniqueness) of the infimum is known as the completeness axiom for
the reals.

Let X be a set and let δ : X × X → R be a map which satisfies

the equation δ(x, y) = δ(y, x) ≥ 0. Note that δ need not satisfy the

31
32 3. The Gluing Construction

triangle inequality. The purpose of this section is to show how to

replace δ by a new function which sometimes remembers some of the
structure of δ and yet satisfies the triangle inequality.
Let x, y ∈ X be two points. Say that a chain from x to y is a
finite sequence of points x = x0 , x1 , . . . , xn = y. Let us call this chain
C. We define
δ(C) = δ(x0 , x1 ) + δ(x1 , x2 ) + · · · + δ(xn−1 , xn ).
Certainly, δ(C) ≥ 0 as long as x = y. Next, we define
d(x, y) = inf δ(C).
C

The infimum is taken over the set of all possible values δ(C), where
C is a chain from x to y.
This probably looks like an insane definition, but we will try to
make it intuitive. Think of δ(x, y) as the cost of flying from city x to
city y—let’s say from Providence to Tahiti. Now, you’re really desper-
ate to get to Tahiti, and have tons of free time but little money. So,
you look on the Internet and try to find all possible flights. You are
willing to take any conceivable chain of connecting flights, as long as
you start in Providence and end in Tahiti. After searching through all
the possiblities, you select the most economical flight. This is d(x, y).
The difference between this scenario and the idealized one we’re talk-
ing about is that X could be an infinite metric space. So, there could
be infinitely many chains, and you need to take the infimum rather
than just a minimum (which may not exist.) The function d is some-
times called the pathification of δ.

Exercise 1. Show that d satisfies the following axioms:

• d(x, y) ≥ 0.
• d(x, y) = d(y, x).
• d(x, y) ≤ d(x, z) + d(z, y).
So it looks like d is a metric. However, note that we are leaving off
the part that would say d(x, y) = 0 iff x = y. In fact give an example
of a δ on X = R2 which satisfies the first two axioms for a metric,
whose pathification is the zero map.
3.2. The Gluing Construction in Action 33

Again, let X be a set. An equivalence relation on X is a relation

of the form ∼, which satisfies three properties:
• x ∼ x for all x.
• x ∼ y iff y ∼ x.
• x ∼ y and y ∼ z imply x ∼ z.
An equivalence class is a subset S = {y ∈ X| y ∼ x}. So, S is the
set of all elements which are equivalent to x. Note that every two
equivalence classes are either disjoint or identical. Thus, it makes
sense to talk about the set of equivalence classes. This set is denoted
X/ ∼.
Now let’s see how ∼ interacts with a metric. Let d be a metric
on X. As above, let X/ ∼ denote the set of equivalence classes of X.
Let us define, for S1 , S2 ∈ [X], the function
δ(S1 , S2 ) = inf d (s1 , s2 ).
The infimum is taken over all possibilities where s1 ∈ S1 and s2 ∈
S2 . In other words the “distance” from S1 to S2 is the “minimum”
distance between a member of S1 and a member of S2 .
Let d be the pathification of δ. We call X/ ∼ a good quotient if d
is a metric on X/ ∼. We think of X/ ∼ as the result of gluing certain
points of X together. Moreover, if x ∼ y and x is near x and y  is
near y, then the pathification process forces x to be near y  . So, at
least in the case when we get a good quotient, the operation of gluing
two points together sort of drags the rest of X along, like a rubber
sheet. Before we give concrete examples, we point out that sometimes
the gluing process leads to a horrible mess.

Exercise 2. Let X = R and write x = y iff x − y is rational.

Show that R/ ∼ is not a good quotient.

3.2. The Gluing Construction in Action

In this section, which is mainly a series of exercises, we give you a
chance to work through lots of concrete examples of the abstract con-
struction given in the previous section.
34 3. The Gluing Construction

Exercise 3. Let X = X1 ∪ X2 , where X1 and X2 are each copies of

the unit disk, equipped with the standard metric, and d(p1 , p2 ) = 1 if
p1 ∈ X1 and p2 ∈ X2 . You should picture two disks hovering, one on
top of the other. Define p1 ∼ p2 if and only if either p1 = p2 or else
p1 and p2 are corresponding points in the boundaries of X1 and X2 .
Prove that the space X/ ∼ is a good quotient, and is homeomorphic
to the 2-sphere.

Exercise 4. The projective plane is the quotient of the sphere S 2

by the equivalence relation p ∼ −p. The points p and −p are called
antipodal points. Prove that the projective plane is a surface.

Exercise 5. Let X = S 1 × [0, 1] be a cylinder. Define an equiv-

alence relation by the rule that (x, 0) ∼ (x, 1) and also (x, y) ∼ (x, y).
Prove that X/ ∼ is a good quotient, and also a surface, and also
homeomorphic to the torus. See Figure 2.1.

Exercise 6. Let X be a metric space of the form

T × {1, 2, 3, 4, 5, 6, 7, 8}.

So, X is the disjoint union of 8 triangles. Define an equivalence rela-

tion on X so that the resulting space is a surface and homeomorphic
to a sphere.

Exercise 7. Describe how to glue a finite number of triangles to-

gether to make the octagon surface discussed in Chapter 1.

Exercise 8. We have already discussed the square torus in §1.1. Here

is another desciption of the same space. On R2 define the equivalence
relation (x1 , y1 ) ∼ (x2 , y2 ) iff x1 − x2 and y1 − y2 are both integers.
Prove that both quotients we have described are good quotients and
homeomorphic to each other. Prove also that the resulting space is
homeomorphic to the surface of a donut, as in Figure 2.1.

Now we mention the cylinder and the Möbius band. Technically,

these are surfaces with boundary. Both surfaces are obtained by
3.2. The Gluing Construction in Action 35

gluing together one pair of opposite sides of a rectangle, as shown in

Figure 3.1.

Figure 3.1 The cylinder and the Möbius band

The Möbius band has some amazing properties. First of all, it

only has one boundary component. Second of all, consider a bug
embedded in the Möbius band that travels from top to bottom. When
the bug gets back to its original position, its notions of left and right
are reversed.
We call a compact surface orientable if it does not contain a
Möbius band. Otherwise, we call the surface nonorientable. Figure
3.2 shows two prototypical examples of nonorientable surfaces, the
projective plane (left) and the Klein bottle (right). We have drawn
in Möbius band subsets in both cases.

Figure 3.2. The projective plane and the Klein bottle

36 3. The Gluing Construction

Exercise 9. Prove that the version of the projective plane shown

on the left-hand side of Figure 3.2 is homeomorphic to the version
described in Exercise 4.

3.3. The Classification of Surfaces

Suppose that S1 and S2 are two compact surfaces. Let D1 and D2
be small open disks in S1 and S2 . We assume that the lengths of the
boundaries of D1 and D2 are the same. We cut out D1 and D2 from
S1 and S2 to produce two new spaces. Finally, we glue the boundary
of S1 − D1 to the boundary of S2 − D2 by an isometric map. We call
the result S1 S2 . Technically, the result depends on the choice of D1
and D2 , but any choice of D1 and D2 leads to the same surface up
to homeomorphism. Figure 3.3 shows an example of the connect-sum
operation applied to two tori.

Figure 3.3. The connected sum

Exercise 10. Prove that S1 S2 is a surface. (Hint: The main diffi-
culty is finding coordinate charts along the “seam”.)

Letting T 2 stand for the torus, the surface

T2  · · · T2 ,

made from g connected sum operations, is called the surface of genus

g. In general, the genus of a compact surface is the integer g such
that χ(S) = 2 − 2g. A surface of genus g is sometimes denoted Σg .
We say “it”, because any g-fold connected sum of g tori gives rise
to the same surface up to homeomorphism. This fact is part of the
classification of surfaces.
3.4. The Euler Characteristic 37

Theorem 3.1 (Classification of Surfaces). Let X be a compact sur-

face. If X is orientable, then there is some g such that X is homeo-
morphic to Σg . If X is nonorientable, then there is some g such that
Σ is homeomorphic to Σg P . Here P is the projective plane.

If we assume that X is made from gluing together finitely many

triangles according to the construction above, then the proof of the
Theorem 3.1 is elementary. Roughly speaking, you cut X open into
one big polygon and analyze the way the sides are glued together.
The book [KIN] has a proof along these lines. The proof of the Euler
formula that we give in the next section is quite similar to the proof
of Theorem 3.1.
The proof for an arbitrary compact surface is to reduce to the
special case where X is built from triangles. In other words, one
shows that X is homeomorphic to a surface built from triangles. We
say in this case that X has a triangulation. It turns out that every
compact surface does have a triangulation, but the result is quite
difficult to prove.

3.4. The Euler Characteristic

We will establish Euler’s formula for orientable surfaces. Suppose
that Σg is decomposed into polygons. We will prove that
(3.1) χ(Σg ) := faces − edges + vertices = 2 − 2g.
The sum in equation (3.1) defines χ(Σg ), and the result is the formula
for χ(Σg ).
First of all, we reduce to the case when the decomposition has
just a single face. Suppose that Σg is decomposed into more than
one face. We can find faces F1 and F2 joined together along an edge
e. We can remove e and set F = F1 ∪e F2 . By this we mean that
we stick F1 and F2 together along e and call the union F . We have
created a new decomposition with one fewer face and one fewer edge.
In particular, we have not changed the Euler characteristic.
It remains to consider the case when there is just one face whose
boundary edges are paired in some way. We call this a gluing pattern
for Σg . We say that the gluing pattern has a cross if we can find two
38 3. The Gluing Construction

pairs of glued edges (e1 , e2 ) and (f1 , f2 ) such that any line segment
connecting e1 to e2 crosses any line segment connecting f1 to f2 , as
shown in Figure 3.4. In other words, the edges e1 , e2 separate the
edges f1 , f2 from each other on the boundary of P .

f2

e1 e2

f1
Figure 3.4. A crossing pattern of edges
Lemma 3.2. If the gluing pattern for Σg does not have a cross, then
it has a pair of consecutive edges that are glued together.

Proof. We will assume that the gluing pattern has neither a cross
nor a pair of consecutive edges and derive a contradiction. Say that
a special segment is a line segment in the interior of F that joins
the midpoints of a glued pair of edges. Let L1 be a special segment.
We rotate so that L1 is vertical. Since L1 does not join consecutive
segments, and there are no crosses, we can find a special segment L2
that lies to the left of L1 . Since L2 does not join consecutive segments,
we can find a special segment L3 , separated from L1 by L2 . Next, we
can find a special segment, L4 , separated from L1 and L2 by L3 . And
so on. In this way, we produce an infinite list L1 , L2 , . . . of distinct
special segments. This contradicts the fact that F only has finitely
many edges. 
Lemma 3.3. If the gluing pattern for Σg has a cross, then Σg is not
a sphere.

Proof. Let (e1 , e2 ) and (f1 , f2 ) be two pairs of edges participating in

a cross, as shown in Figure 3.5. Without loss of generality, it suffices
3.4. The Euler Characteristic 39

to consider the case when the e and f edges are contained in the edges
of the unit square. Figure 3.5 below shows the situation. The thick
segments between the e edges and the f edges each represent a finite
union of edges of F . Though we have not drawn things this way, one
more of these segments could be empty.

Figure 3.5. A torus with flaps.

Were we to glue the opposite sides of the unit square, we would

get a torus, as shown on the left-hand side of Figure 3.5. To obtain Σg
from this picture, we delete the white “flaps” from the torus, and then
glue together the edges of the corresponding boundary, according to
the original gluing pattern. The relevant boundary is drawn thickly.
It is convenient to draw the torus in a different way, this time
with the handle drawn on the inside. Rather than draw the handle,
we have just added the letter H, to denote that the drawn disk is
really a disk with a handle attached. The right-hand side of Figure
3.5 shows this. Were we to draw the “flaps”, they would be on the
outside of the shaded region.
The right-hand side of Figure 3.5 realizes Σg as the connected
sum of another oriented surface and a torus. In particular, Σg cannot
be a sphere. 

Lemma 3.4. Euler’s formula is true for a sphere.

Proof. Our proof goes by induction on the number of edges of F .

In case F just has 2 edges, the decomposition of Σ has 1 edge and 2
40 3. The Gluing Construction

vertices. We have
χ(Σ0 ) = 2 − 1 + 1 = 2.
This takes care of the base case. In general, some pair of consecutive
edges of F is glued together. Since F is orientable, these edges point
in opposite directions, as shown in Figure 3.6.

Figure 3.6. Gluing consecutive edges

In this case, we glue up the edges and erase the vertex between
them. The result is a gluing pattern for Σg that is based on a polygon
with 2 fewer edges. This is the induction step. 

Now we consider the general case. Our result goes by induction

on g. We have already taken care of Σ0 . In light of Lemma 3.4, the
converse of Lemma 3.3 is true: If Σg is not a sphere, then the gluing
pattern for Σg does have a cross. Otherwise, we could just “zip up”
Σg one pair of edges at a time and produce a sphere.
So, we start with a cross and reproduce the construction made in
the proof of Lemma 3.3. That is, we arrive at the picture in Figure
3.5. When we replace our disk-with-handle with a disk, we produce
a gluing diagram, based on a polygon F  , for the surface Σg−1 . Here
F  has 4 fewer edges than F does. At the same time, F and F  have
the same set of vertices, and they are glued together in the same way.
By induction, Euler’s formula holds for Σ . Hence
f  − e + f  = 2 − 2(g − 1).
Here f  = 1 is the number of faces in the decomposition and e is the
number of edges, and v  is the number of vertices. By construction
3.4. The Euler Characteristic 41

f  = f and e = e − 2 and v  = v. So, we get

f − e + v = 2 − 2g,
as desired.

Exercise 11. Prove Euler’s formula for nonorientable surfaces.

Chapter 4

The Fundamental Group

The purpose of this chapter is to define the fundamental group, an ob-

ject we discussed briefly in §1.3. In the next chapter, we will compute
some examples. As we mentioned in §1.3, the fundamental group is an
object that organizes all the different loops on a surface (or any topo-
logical space, for that matter). In this chapter, I will first talk about
groups in general, then groups will disappear from the discussion for
a while; then they will come back in a really surprising way. For a
more formal treatment of the fundamental group, see, e.g., [HAT].

4.1. A Primer on Groups

If you haven’t had any group theory, you can find a treatment in
any number of abstract algebra books; see, for instance, [HER]. A
group is a set G, together with an “operation” ∗, which satisfies the
following axioms:
• g1 ∗ g2 is defined and belongs to G for all g1 , g2 ∈ G.
• g1 ∗ (g2 ∗ g2 ) = (g1 ∗ g2 ) ∗ g3 for all g1 , g2 , g3 .
• There exists a (unique) e ∈ G such that e ∗ g = g ∗ e = g for
all g ∈ G.
• For each g ∈ G there is a (unique) element h such that
g ∗ h = h ∗ g = e. This element is called “g inverse” and is
usually written as h = g −1 .

43
44 4. The Fundamental Group

The group G is called Abelian if, additionally, g1 ∗ g2 and g2 ∗ g1 are

always equal. A subgroup of a group is a subset H ⊂ G which is closed
under the group law. So, if h ∈ H then h−1 ∈ H and if h1 , h2 ∈ H
then h1 ∗ h2 ∈ H.
Here are some examples of groups:

• Z, with the + operation, forms an Abelian group.

• If G1 and G2 are groups, then G1 × G2 can be made a group

using the law (g1 , g2 ) ∗ (h1 , h2 ) = (g1 ∗ h1 , g2 ∗ h2 ).

• The set SLn (Z) of n × n integer matrices with determi-

nant 1 forms a non-Abelian group. The group law is matrix
multiplication.

• Let A be a collection of n things, for instance A = {1, ..., n}.

Say that a permutation is a bijection f : A → A. There are
n! different permutations, and they form a finite group. The
∗ operation is composition of maps. This group is called Sn .

Let G1 and G2 be groups. A map f : G1 → G2 is a homomor-

phism if

f (a ∗ b) = f (a) ∗ f (b)

for all a, b ∈ G1 . Here the ∗ on the left-hand side is the rule for G1
and the ∗ on the right-hand side is the one for G2 . The map f is
called an isomorphism if f is a bijection and also a homomorphism.
Here is a nice example. Let G be a finite group, and let n be the
number of elements in G. We’re going to produce a homomorphism
from G into Sn , the permutation group on n things. The n things
are just the elements of G. So, given an element g ∈ G, how do we
permute the elements of G? We define the map fg : G → G using
the rule fg (h) = gh. It turns out that fg is a bijection, and fg1 = fg2
only if g1 = g2 . The map g → fg is a one-to-one homomorphism from
G into Sn . We have essentially given the proof of Cayley’s theorem:
every finite group is isomorphic to a subgroup of a permutation group.
4.2. Homotopy Equivalence 45

4.2. Homotopy Equivalence

Now we go back to metric spaces. Let X and Y be metric spaces. Let
I = [0, 1] be the unit interval. Two maps f0 , f1 : X → Y are said to
be homotopic if there is a continuous map F : X × I → Y such that

• F (x, 0) = f0 (x) for all x ∈ X.

• F (x, 1) = f1 (x) for all x ∈ X.

To explain the intuitive idea, it is useful to define ft : X → Y by the

formula ft (x) = F (x, t). Then the map ft interpolates between f0
and f1 , with ft being very close to f0 when t is near 0 and ft being
very close to f1 when t is near 1. The map F is called a homotopy
from f0 to f1 .
If is useful to write f0 ∼ f1 if these maps are homotopic. Let
C(X, Y ) denote the set of all continuous maps from X to Y . One can
think of ∼ as a relation on the set C(X, Y ).

Exercise 1. Prove that ∼ is an equivalence relation on C(X, Y ).

Exercise 2. Prove that every two elements of C(X, Rn ) are ho-

motopic. (Hint: Prove that any map f : X → Rn is homotopic to
the zero-map f0 defined by the property f0 (x) = 0 for all x. Then,
use the fact that ∼ is an equivalence relation.)

Exercise 3 (Challenge). Let P be a polynomial

P (x) = xn + an−1 xn−1 + · · · + a0 .

Let Q be the polynomial Q(x) = xn . So, P and Q have the same

leading term. We can think of P as a map from C to C. Here C
is the complex plane. For any R we can let X ⊂ C be the circle of
radius R centered at 0. That is

X = {z ∈ C| |z| = R}.

First of all, prove that 0 ∈ P (X) if R is sufficiently large. This means

that we can think of P and Q as maps from X to Y = C − {0}. Prove
that P, Q : X → Y are homotopic if R is sufficiently large.
46 4. The Fundamental Group

4.3. The Fundamental Group

From now on we are going to take X = I, the unit interval, and we
are going to study the space Y by looking at the maps from I into
Y . For this entire discussion we choose a special “reference point”
y0 ∈ Y , which we call the basepoint.
Say that a loop in Y is a continuous map f : I → Y such that

f (0) = f (1) = y0 .

The reason for the terminology should be pretty clear. Say that two
loops f0 and f1 are loop homotopic if there is a homotopy F from
f0 to f1 such that ft is a loop for all t ∈ [0, 1]. This is to say that
F (0, t) = F (1, t) = y0 for all t. We write f0 ∼ f1 in this case.
Figure 4.1 shows an example. Just as in Exercise 1, this relation is
an equivalence relation. Note that the equivalence relation here is
slightly different than the one in the previous section, because of the
added constraint that F (0, t) = F (1, t) = y0 for all t.

Figure 4.1. Homotopic loops

As a set, π1 (Y, y0 ) is the set of equivalence classes of loops. The

really interesting thing is that we can make π1 (Y, y0 ) into a group.
Here is the construction. Suppose that we have two elements [f ]
and [g] of π1 (Y, y0 ). We can let f and g be representatives of the
equivalence classes [f ] and [g], respectively. That is, f : [0, 1] → Y is
a loop and g[0, 1] → Y are both loops.
4.3. The Fundamental Group 47

f
g

Figure 4.2. Composing loops

We define the new loop h = f ∗ g by the following rule.

• If x ∈ [0, 1/2], we define h(x) = f (2x). That is, the first
half of h traces out all of f , but twice as fast.
• If x ∈ [1/2, 1], we let x = x − 1/2 and then we define
h(x) = g(2x ). That is, the second half of h traces out g,
but twice as fast.
We write h = f ∗ g. See Figure 4.2.

Exercise 4. Suppose that f and g are different representatives for

[f ] and [g]. That is, f and f are equivalent loops and g and g are
equivalent loops. Let h = f ∗ g. Prove that [
h] = [h]. In other words,

prove that h and h are equivalent loops. This exercise is pretty easy,
but quite tedious.

Given Exercise 4, we can define

(4.1) [f ] ∗ [g] = [f ∗ g],
and this definition is independent of the equivalence class representa-
tives we used to make the definition.

Exercise 5. Show, for any three loops, f, g, h, that (f ∗ g) ∗ h is

equivalent to f ∗(g ∗h). This means that ([f ]∗[g])∗[h] = [f ]∗([g]∗[h]).
This is the associative law for groups.

Exercise 6. Let e be the loop defined by the rule e(x) = y0 for

all x ∈ I. Show that [e] ∗ [g] = [g] ∗ [e] = [g] for all loops g. This
48 4. The Fundamental Group

means that [e] plays the role of the identity element in π1 (Y, y0 ).

Exercise 7. Let g be any loop. Define the loop g ∗ so that it satisfies

the equation g ∗ (x) = g(1 − x). In other words, g ∗ traces out the same
loop as g, but in the opposite direction. Prove the following result:
If g1 and g2 are equivalent, then g1∗ and g2∗ are equivalent. Finally,
prove that [g] ∗ [g ∗ ] = [e] and [g ∗ ] ∗ [g] = [e]. In other words, the
inverse of [g] is given by [g ∗ ].

Combining Exercises 5, 6, and 7, we see that π1 (Y, y0 ) is a group.

So, to each space Y we can pick a basepoint y0 and then define the
group π1 (Y, y0 ). This group is known as the fundamental group of Y .
(We will see below that the group you get does not really depend on
the basepoint.)

4.4. Changing the Basepoint

Say that two points y0 , y1 are connected by a path if there is a con-
tinuous map f : I → Y such that f (0) = y0 and f (1) = y1 . Say that
Y is path connected if every two points in Y can be connected by a
path. For instance Rn is path connected whereas Z is not.

Lemma 4.1. Suppose that y0 , y1 ∈ Y are connected by a path. Then

π1 (Y, y0 ) and π1 (Y, y1 ) are isomorphic groups. In particular, if Y is
path connected, then the (isomorphism type of the) group π1 (Y, y) is
independent of the choice of basepoint y and we can just write π1 (Y ).

Proof (Sketch). Let d be a path which joins y0 to y1 . Let d∗ be

the reverse path, which connects y1 to y0 . We want to use d and
d∗ to define a map from π1 (Y, y0 ) to π1 (Y, y1 ). Given any y0 -loop
f0 : I → X with f0 (0) = f0 (1) = y0 , we can form a y1 -loop by the
formula
f1 = d ∗ f ∗ d∗ .
In other words, the first part of f1 travels backward along d from y1
to y0 , the second part travels around f0 , and the third part travels
back to y1 . You should picture a lasso, as in Figure 4.3.
4.5. Functoriality 49

f
y d
0 y
1

Figure 4.3. A lasso

Using arguments similar to the ones for the exercises above, you
can show the following result: If f0 and f0 are equivalent, then f1
and f1 are equivalent. In other words, the map H, which sends
[f0 ] ∈ π1 (Y, y0 ) to [f1 ] ∈ π1 (Y, y1 ) is well defined independent of the
equivalence class representative used to define it. So, now we have
a well-defined map H : π1 (Y, y0 ) → π1 (Y, y1 ). After this, one shows
that H is a homomorphism. That is, H([f ] ∗ [g]) = H([f ]) ∗ H([g]).
This is not hard to do once you draw a picture of what is going on.
Rather than show that H is one-to-one and onto directly, one can
define a map H ∗ : π1 (Y, y1 ) → π1 (Y, y0 ) just by reversing the roles of
the two points. In other words, the loop f1 is mapped to
f0∗ = d∗ ∗ f1 ∗ d.
Note that f0∗ and f0 are not precisely the same loop. If you draw
pictures you will see that there is some extra slack in f0∗ . However,
it turns out that [f0∗ ] = [f0 ]. In other words, the two loops are loop
homotopic. Thus H and H ∗ are inverses of each other. Hence H is
an isomorphism.

4.5. Functoriality
The word functoriality refers to a situation where you are assign-
ing one kind of an object to another in a way which respects the
“natural” transformations between the two kinds of objects. This
notion is defined precisely in any book on category theory. In our
case, we are assigning a group π1 (Y, y0 ) to a pointed space (Y, y0 ).
(By pointed space we mean a space with a chosen basepoint.) The
natural transformations of pointed spaces are basepoint-preserving
50 4. The Fundamental Group

continuous maps and the natural transformations between groups are

homomorphisms.
We would like to see that our transformation (or functor ) from
spaces to groups respects these transformations. Lemmas 4.2 and 4.3
together contain this information.

Lemma 4.2. Let (Y, y0 ) and (Z, z0 ) be two pointed spaces, and let
f : Y → Z be a continuous map such that f (y0 ) = z0 . Then there is
a homomorphism f∗ : π1 (Y, y0 ) → π1 (Z, z0 ).

Proof. Let [a] ∈ π1 (Y, y0 ) be an equivalence class of loops, with

representative a. So, a : I → Y is a loop. The composition f ◦ a is
loop in Z. We define f∗ [a] = [f ◦ a]. If [a0 ] = [a1 ], then there is a
homotopy H from a0 to a1 . But then f ◦ H is a loop homotopy from
f ◦a0 to f ◦a1 . So, [f ◦a0 ] = [f ◦a1 ] and our map is well defined. Note
that f ◦(a∗b) = (f ◦a)∗(f ◦b). Hence f ∗([a]∗[b]) = (f∗ ([a]))∗(f∗ ([b])).
Hence f∗ is a homomorphism. 

Suppose that f : Y → Z is a continuous map and g : Z → W is

a continuous map. Let’s arrange so that f (y0 ) = z0 and f (z0 ) = w0 .
Then g ◦ f is a map from Y to W and (g ◦ f )∗ is a homomorphism
from π1 (Y, y0 ) to π1 (W, w0 ).

Lemma 4.3. (g ◦ f )∗ = g∗ ◦ f∗ .

Proof. Let [a] ∈ π1 (Y, y0 ). Then

(g ◦ f )∗ [a] = [(g ◦ f ) ◦ a] = [g ◦ (f ◦ a)] = g∗ [f ◦ a] = g∗ f∗ [a].
That is it. 

If f : Y → Y is the identity map, then f∗ is the identity map

on π1 (Y, y0 ). Also, if h : Y → Z is a homeomorphism, then we
have the inverse homeomorphism h−1 . But h ◦ h−1 is the identity.
Hence h∗ ◦ h−1 ∗ is the identity homomorphism. Likewise h
−1
◦ h∗ is
−1
the identity homomorphism. In short h∗ (and also h∗ ) is a group
isomorphism. So

Theorem 4.4. If π1 (Y, y0 ) and π1 (Z, z0 ) are not isomorphic groups,

then there is no homeomorphism from Y to Z which maps y0 to z0 .
4.6. Some First Steps 51

The above is slightly contrived because we don’t really care about

these basepoints. Recall that π1 (Y, y0 ) does not depend on the base-
point if Y is path connected. So

Theorem 4.5. Suppose Y and Z are path connected spaces. If π1 (Y )

and π1 (Z) are not isomorphic, then Y and Z are not homeomorphic.

What’s really great about this result is that we can use it to

tell the difference between spaces just by looking at these groups.
Of course, the question remains: How do we actually compute these
groups? In the next chapter, we will go into much more details about
this.

4.6. Some First Steps

Here we will just take some first steps in the computation of funda-
mental groups. Once we have more theory, these computations will
be easy. So, what fundamental groups can we compute? It is easy to
see (compare Exercise 2) that any two loops in Rn (based at 0) are
equivalent. Hence π1 (Rn , 0) is the trivial group.

Exercise 8A (Challenge). Prove that there is a loop in S 2 (the

2-sphere) whose image is all of S 2 . (Hint: If you know about the
Hilbert plane-filling curve from real analysis, you’re in good shape for
this problem.)

Exercise 8B (Challenge). Prove that π1 (S 2 , p) is the trivial group.

Here p ∈ S 2 is any point. (Hint: The intuitive idea is this: If the
loop misses some point q = p, you can just “slide” the loop “down to
p” by pushing it away from the missed point. However, you have to
deal with the loops which come from Exercise 8A.)

Exercise 9. If (Y, y0 ) and (Z, z0 ) are two pointed spaces, then the
product
(Y × Z, (y0 , z0 )
is again a pointed space. Prove that
π1 (Y × Z, (y0 , z0 )) = π1 (Y, y0 ) × π1 (Z, z0 ).
52 4. The Fundamental Group

Exercise 10 (Challenge). Prove that π1 (S 1 , p) is nontrivial. (Hint:

Think of S 1 as the unit circle in R2 and consider the loop
f (t) = (cos(2πt), sin(2πt)).
Show that this loop is inequivalent to the identity loop.)

Let T = S 1 × S 1 . Here T 2 is the torus. From Exercises 9 and 10

we know that π1 (T 2 ) is nontrivial. (We don’t worry about the base-
point because T is obviously path connected.) On the other hand, by
Exercise 8, π1 (S 2 ) is trivial. Hence S 2 and T 2 are not homeomorphic!
Chapter 5

Examples of
Fundamental Groups

The purpose of this chapter is to compute the fundamental group for

some familiar objects:

• the circle;
• the torus;
• the 2-sphere;
• the projective plane;
• lens spaces;
• the Poincaré homology sphere.

I will work out the first three in detail and then guide you through
the computation for the others. The last section is too advanced for
an undergraduate course but I couldn’t resist.

5.1. The Winding Number

Let S 1 be the circle. We think of S 1 as the set of unit complex
numbers in C. We choose 1 for our basepoint of S 1 . In this section
we will describe how to assign an integer to a continuous loop g :
[0, 1] → S 1 .

53
54 5. Examples of Fundamental Groups

First we will explain the idea intuitively and then we will get
to the formalities. Think of the loop g as describing a bug crawling
around the unit circle. Imagine that you are at the center of the
circle watching the bug. You always follow the bug with your eyes,
staring straight at it the whole time. (Your rubber neck allows you
to do this.) After the bug has completed his trip, you are looking in
the same direction as initially. However, your head has been twisted
around some number of times. The winding number is the integer,
positive for counterclockwise and negative for clockwise, which names
how many times your head is twisted around.
Now we come to the formalities. Let R denote the real numbers.
There is a natural map E : R → S 1 given by
E(t) = exp(2πit) = cos(2πt) + i sin(2πt).
This map is certainly onto and continuous, but it has some other
special properties. Say that an open special arc in S 1 is a set of the
form
C(z) = {w ∈ S 1 | d(z, w) < 1/100}.
Here d(z, w) = |z − w|, the usual Euclidean distance. The choice of
1/100 is convenient but fairly arbitrary. The point is just that open
special arcs are smaller than semicircles.

Exercise 1. Let C be an open special arc. Prove that E −1 (C)

consists of a countably infinite number of disjoint open intervals and
that the restriction of E to any of them is a homeomorphism from
the interval onto C.
Lemma 5.1. Let [a, b] ⊂ R be an interval. Suppose g : [a, b] → S 1 is
a map such that g([a, b]) is contained in a special arc. Suppose also
that there is a map g : {a} → R such that E ◦ g(a) = g(a). Then we
can define g : [a, b] → R such that E ◦ g = g on all of [a, b]. This
extension of g is unique.

Proof. If E had an inverse we could define g = E −1 ◦ g. Also, we

would be forced to make this definition and so the extension of g
to [a, b] would be unique. Unfortunately, E is not invertible. For-
tunately, we have Exercise 1, which shows that E is “invertible” in
some sense. Let C be the special arc which exists by hypothesis. Let
5.1. The Winding Number 55

C ⊂ E −1 (C) be the unique interval from Exercise 1 which contains

g(a). By Exercise 1, the map E : C → C is a homeomorphism. Let
F : C → C  be the inverse of (the restricted version of) E. Since
g[a, b] ⊂ C we can (and must) define G  = F ◦ g. 

Let 1 = E(Z) be the basepoint of S 1 . Let I = [0, 1]. Recall that

an element of π1 (S 1 , 1) is a map g : I → S 1 such that g(0) = g(1) = 1.

Exercise 2. Given the map g, prove that there exists some N with
the following property. If x, y ∈ [0, 1] and |x − y| < 1/N then the set
g([x, y]) is contained in a special arc. (Hint: You might want to use
the fact that every infinite sequence in [0, 1] has a convergent subse-
quence. This is basically the Bolzano–Weierstrass theorem.)

Here is an improved version of Lemma 5.1.

Lemma 5.2. Let g : [0, 1] → S 1 be a loop. Then there is a unique

 = G on all of [0, 1].
map g : [0, 1] → R such that g(0) = 0 and E ◦ G

Proof. From Exercise 2 we can find some N so that the points ti =

i/N have the following property. The image g([ti , ti+1 ]) is contained
in a special arc for i = 0, . . . , (N − 1). Now we go by induction. First
of all, by Lemma 5.1 we can define g uniquely on [t0 , t1 ]. But then by
Lemma 5.1 again, we can define g uniquely on [t1 , t2 ]. And so on. 

Definition 5.1. We define the winding number of g to be the value

of g(1) ∈ Z. We write this as w(g). Note that g(1) ∈ Z because
g(1) = E(g(1)) = 0.

We would like to see that the winding number only depends on

the homotopy class of the loop. Going back to our intuitive notion
of the winding number, suppose that two bugs are running around
the unit circle, and they stay pretty close to each other. Then you
will always be looking in about the same direction if you watch either
bug. So, your head will be turned around the same number of times
if you watch either bug. Now we give the formal proof.
56 5. Examples of Fundamental Groups

Lemma 5.3. Suppose that g0 and g1 are homotopic loops in S 1 .

Then w(g1 ) = w(g1 ).

Proof. Let G be the homotopy between g0 and g1 . Let gt (x) =

G(x, t). The same argument as in Exercise 2 proves that there is
some N with the following property: If s, t ∈ [0, 1] are any points
such that |s − t| < 1/N and x ∈ [0, 1] is fixed, then
|G(x, s) − G(x, t)| < 1/100.
Using the other notation, we have d(gs (x), gt (x)) < 1/100. But then
gs (x), 
d( gt (x)) is either less than 1/100 or greater than 1/2. By con-
tinuity, the alternative cannot change. Also
gs (0), 
d( gt (0)) = d(0, 0) = 0 < 1/100.
This shows that the first alternative always holds and gs (x) and gt (x)
are always within 1/100 of each other. But then w(gs ) = w(gt ),
because both are integers within 1/100 of each other. From here it is
easy to see that w(g0 ) = w(g1 ). 

5.2. The Circle

We will use the winding number to compute π1 (S 1 , 1), the fundamen-
tal group of the circle.
Given a loop g, representing an element of π1 (S 1 , 1), we define
w([g]) = w(g).
By Lemma 5.3, this gives us a well-defined map w : π1 (S 1 , 1) → Z.
Lemma 5.4. w is onto.

Proof. Let g(t) = exp(2πint). Then w(g) = n. 

Execise 3. Prove that w is a homomorphism.

Lemma 5.5. w is an isomorphism.

Proof. Since w is a homomorphism, it suffices to prove the following

statement. If w(g) = 0, then g is homotopic to the constant loop.
Now, if w(g) = 0, then g : [0, 1] → R is a loop. But π1 (R, 0) = 0.
Hence there is a loop homotopy G  from g to the constant loop g0 :
5.3. The Fundamental Theorem of Algebra 57

S 1 → R. But then E ◦ G is a loop homotopy from g to the constant

1
loop in S . This shows that w is an isomorphism. 

The last result shows that π1 (S 1 , 1) is isomorphic to Z.

Remark. The main property we used about the circle was the exis-
tence and special properties of the map E : R → S 1 . We also used
the property that π1 (R, 0) = 0. It turns out that this will be a general
method for us when we compute the fundamental groups. All the spe-
cial properties we established are summarized by the statement that
R is the universal cover of S 1 and E is the universal covering map.
In the next chapter I will develop these ideas in great generality.

5.3. The Fundamental Theorem of Algebra

The Fundamental Theorem of Algebra says that every complex poly-
nomial
P (z) = a0 + a1 z + · · · + an z n
has a root. This result has a nice proof based on the ideas we have
been developing. For convenience, we divide through so that an = 1.
We think of P as a continuous map from C to C. If P has no roots,
then P is a continuous map from C to C − {0}.
Let Cr denote the circle of radius r centered at the origin. Let
S 1 denote the unit complex numbers. Given r > 0, consider the map
γr : S 1 → S 1 given by
P (ru)
γr (u) = .
|P (ru)|
By construction γr is a continuous loop, and hence an element of
π1 (S1 ).
When r is small, P (Cr ) is just a tiny loop around f (0) = 0.
Hence [γr ] = 0 ∈ π1 (S1 ) for r small. But γr varies continuously with
r. Hence [γr ] = 0 for all r. On the other hand, when z ∈ Cr and r is
large, we have
P (z) = z n + f (z), |f (z)| < r |P (z)|.
Here r is some constant that tends to 0 as r → ∞. The point is that
the highest order term dominates the sum of the remaining terms.
58 5. Examples of Fundamental Groups

Our estimate tells us that γr converges to the loop z → z n as

r → ∞. Hence [γr ] = n for r large. This is a contradiction. The only
way out is that P is not a continuous map from C to C − {0}. But
then 0 must be in the image of P . That is, P has a root.

5.4. The Torus

Exercise 9 of Chapter 4 asked you to show that

π1 ((Y, y) × (Z, z)) = π1 (Y, y) × π1 (Z, z).

The torus T 2 is homeomorphic to S 1 × S 1 and also path connected.

Hence π1 (T 2 ) = Z × Z. Iterating, we get π1 (T n ) = Z n .

5.5. The 2-Sphere

Let I = [0, 1] as above. Let x ∈ S 2 be some basepoint. This section,
which consists mainly of exercises, will guide you through the proof
that π1 (S 2 , x) = 0.
Say that a loop g : I → S 2 (anchored at x) is bad if g(I) = S 2
and otherwise is good .

Exercise 4. Prove that any good loop is homotopic to a point.

Exercise 5. Let [a, b] be an interval, and let H be a hemisphere

in R2 . Let f : [a, b] → H be a continuous map. Prove that there is
homotopy F : [a, b] × [0, 1] → Δ such that
• F (a, t) and F (b, t) are independent of t.
• F (x, 0) = f (x) for all x.
• f1 : [a, b] → Δ is contained in a circular arc joining f (a) to
f (b).
Exercise 6. Let g be an arbitrary loop on S 2 . Prove that there is
a finite partition 0 = t0 < t1 < · · · < tn = 1 such that g maps each
interval [ti , ti+1 ] into a hemisphere. Now conclude from Exercise 5
that g is loop homotopic to a good loop.
5.7. A Lens Space 59

Since every loop in S 2 is loop homotopic to a good loop, and

every good loop is loop homotopic to a point, every loop in S 2 is
homotopic to a point. Therefore, π1 (S 2 , x) = 0. The same argument
works for S n , with n > 2.

5.6. The Projective Plane

As in §3.2, we think of P 2 , the projective plane, as the quotient S 2 / ∼,
where x ∈ S 2 is equivalent to itself and to the antipodal point −x.
There is a nice map E : S 2 → P 2 given by E(x) = [x]. As our
notation suggests, E plays the same role here that the same-named
map played above when we considered the circle.
Let x+ = (0, 0, 1), and let x− = (0, 0, −1). Clearly, we have
E(x+ ) = E(x− ).

Exercise 7. Suppose that g : [0, 1] → P 2 is a loop based at x+ .

Prove that there is a unique map g : [0, 1] → S 2 such that g(0) = x+
and E ◦ g = g. (Hint: Just imitate what was done for the circle.)

Note that either g(1) = x+ or g(1) = x− . We define w(g) = +1

if g(1) = x+ and w(g) = −1 if g(1) = x− .

Exercise 8. Prove that w([g]) is well defined independent of the

loop homotopy equivalence class of g. Prove also that w gives an
isomorphism from π1 (P 2 ) to Z/2.

In general we have P n = S n / ∼, where x ∼ −x. Thus there is

always this two-to-one map from S n to P n . An argument similar to
the one given above shows that π1 (P n ) = Z/2. Here P n is called
projective n-space.

5.7. A Lens Space

Before reading this section, you should probably know what a mani-
fold is; see §2.8 for details.
We think of S 3 as the set of the form
{(z, w)| |z|2 + |w|2 = 1} ⊂ C 2 = R4 .
60 5. Examples of Fundamental Groups

This is an exotic way of expressing the fact that S 3 , the 3-sphere, is

the unit sphere in R4 . The equality C 2 = R4 comes from the map

(x1 + iy1 , x2 + iy2 ) → (x1 , y1 , x2 , y2 ).

Here is a nice equivalence relation on S 3 . Let’s define

(z, w) ∼ (uz, u2 w)

if and only if u is some 5th root of unity. Each equivalence class on

S 3 / ∼ has 5 points. Let’s call this space L(2, 5). The 2 comes from
the u2 term, and the 5 comes from the fact that we are taking 5th
roots of unity. Obviously, you could make this construction for other
choices.
Here is a sketch of how to visualize L(2, 5). Any point in L(2, 5)
is equivalent to a point of the form (z, w), where the argument of z
lies in the interval (0, 2π/5). Let S ⊂ S 3 be this set. We can write


S= Sθ ,
θ∈[0,2π/5]

where Sθ consists of points of the form (z, w) where z = exp(iθ). The

whole sphere S 3 is tiled by 5 copies of S. For instance, one of the
adjacent copies consists of those sets Sθ , where θ ∈ [2π/5, 4π/5].
Now we are going to (partially) explain how to visualize S. The
“slice” Sθ is a disk, and the boundary of Sθ is the circle

C = {0} × {w : |w| = 1}.

All the slices share C as a common boundary, but otherwise they are
disjoint. So, S looks something like a circular pillow, or the solid
region between two contact lenses stuck boundary to boundary. The
left-hand side of Figure 5.1 shows a side view. The two dots represent
C. To get a better picture, you could revolve this planar figure about
the vertical axis.
5.7. A Lens Space 61

Figure 5.1. The domain S

The right-hand side of Figure 5.1 shows a top view of S. We

imagine that we are looking at S0 , and that the rest of S is under-
neath. The other boundary component is S2π/5 . We have drawn the
circle C as a pentagon, to suggest the what is going on. We observe
the following things.

• Each point in the interior of S is equivalent only to itself.

• Each point on the interior of the “front” of S, meaning the

set S0 − C, is equivalent to one point on S2π/5 − C.

• Each point on one of the edges of C is equivalent to the 4

other points at corresponding positions on the other edges.
The triangular subdivision is supposed to serve as a guide
to the gluings.

I have not described things completely, because I want to leave you

something to think about.

Exercise 9. Prove that L(2, 5) is a good quotient in the sense of

§3.1, and also a manifold.

There is an obvious map E : S 3 → L(2, 5). Using E we can show that

π1 (L(2, 5)) = Z/5. Generalizing this construction in an obvious way,
we see that we can produce a 3-manifold whose fundamental group is
Z/n. These spaces L(m, n) are called lens spaces.
62 5. Examples of Fundamental Groups

5.8. The Poincaré Homology Sphere

Before we give the last example, we need to make a detour and discuss
a different way to think about S 3 . Let SO(3) denote the group of
orientation preserving (i.e., physically possible) rotations of S 2 . It
turns out that there is an amazing map from S 3 to SO(3) which
is really the map from S 3 to P 3 in disguise. So, given an element
q ∈ S 3 , we need to produce a rotation Rq of S 2 .
Here is the construction. We think of S 3 as the unit quaternions.
That is, a point in S 3 can be thought of as a symbol of the form
a + bi + cj + dk, a, b, c, d ∈ R, a2 + b2 + c2 + d2 = 1.
The symbols i, j, k satisfy the following rules:
• i2 = j 2 = k2 = −1.
• ij = k and jk = i and ki = j.
Given these rules, you can multiply quaternions together in a way
which is similar to how you multiply complex numbers together.
Given any q ∈ S 3 as above, we define
q −1 = a − bi − cj − dk.
Then you can check that qq −1 = q −1 q = 1. In other words, the unit
quaternions form a group under multiplication!
We can identify R3 with the pure quaternions, namely those of
the form 0 + bi + cj + dj. The isomorphism to R3 is just given by
0 + ai + bj + ck → (a, b, c).
3
Thus our special R has the usual Euclidean metric on it, coming
from the identification with the usual R3 .
Given p ∈ R3 we define
Rq (p) = qpq −1 .
Exercise 10 (Challenge). Show that Rq preserves R3 (the pure
quaternions) and is an orientation-preserving rotation.

Multiplication turns out to be associative and so we have

Rq1 ◦ Rq2 (p) = q1 (q2 pq2−1 )q1−1 = Rq1 ◦ Rq2 (p).
5.8. The Poincaré Homology Sphere 63

This works for any p. Hence the map q → Rq is a homomorphism.

As you might expect, we define E(q) = Rq . Note that E(−q) = E(q).
It turns out that the kernel of E is precisely {1, −1}. So, E is both
a continuous surjection (with good local inverse properties) and a
two-to-one homomorphism from S 3 to SO(3).
Now for our last example. Given the quaternionic picture of S 3 ,
we can define a very interesting 3-dimensional manifold. If G ⊂ SO(3)
is a finite subgroup, then G  = E −1 (G) is a subgroup with twice the
number of elements. Now we can define an equivalence on S 3 by the
rule q1 ∼ q2 iff there exists some g ∈ G such that gq1 = q2 . If G has

N elements, then G has 2N elements and each equivalance class of
S 3 / ∼ has 2N elements. It turns out the quotient space is a manifold
with fundamental group G. 
As a special case, let G be the orientation-preserving symmetries
of the icosahedron, the most interesting finite subgroup of SO(3).
Then G  is an order 120 group known as the binary icosahedral group.
The quotient in this case is called the Poincaré homology sphere, and
its fundamental group is G. 

Figure 5.2. A dodecahedron

The Poincaré homology sphere is one of the great examples in ge-

ometry. In the lens space example, S 3 is tiled by 5 copies of a kind of
“double lens”. In the Poincaré homology sphere example, it turns out
that S 3 is tiled by 120 spherical dodecahedra. The spherical dodeca-
hedra look combinatorially the same as Euclidean dodecahedra, but
they are “puffed-out” much in the same way that spherical triangles
64 5. Examples of Fundamental Groups

are. Figure 5.2 shows a dodecahedron, drawn so that the thick lines
represent visible edges and the thin lines represent hidden edges.
Any point in S 3 is equivalent to a point in one of these dodecahe-
dra, and no two points in the interior of a dodecahedron are equivalent
to each other. Thus, analyzing the Poincaré homology sphere boils
down to understanding how points on the boundary of one of the do-
decahedra are glued together. What happens is that each face of the
dodecahedron is glued to the opposite face, with a 2π/5 twist.
Chapter 6

Covering Spaces and the

Deck Group

In §1.4 we discussed how the process of “unwrapping” the essential

loops on the square torus leads naturally to the integer grid in the
plane. We also mentioned that something similar can be done for the
octagon surface and its relatives. The purpose of this chapter and the
next one is to make the unwrapping process precise, and to consider
it in much greater generality. The central objects in this chapter are
covering spaces and the deck group, objects which play the role that
the plane and the integer grid, respectively, played in §1.4. Along
the way, we will relate covering spaces and the deck group to the
fundamental group.

6.1. Covering Spaces

Let X  and X be path connected metric spaces. Let E : X  → X be
a continuous map. An open set U ⊂ X is said to be evenly covered
if the preimage E −1 (U ) consists of a countable disjoint union of sets
U1 , U
2 , . . . such that the restriction E : U
j → U is a homeomorphism.

(This makes sense because Uj is a metric space in its own right.) It is
customary to require that U is path connected in this definition. The
sets U j are called components of the pre-image. The map E is said

65
66 6. Covering Spaces and the Deck Group

to be a covering map if every point in X has a neighborhood that is

 is said to be a covering space of X.
evenly covered. In this case, X
The “mother of all examples” is the map E : R → S 1 discussed
in §5.1. Here we will describe this map in another way. We still think
of the line as R, but now we think of the circle as the space X ob-
tained from [0, 1] by gluing 0 to 1. This time, our map E is given by
E(x) = [x−floor(x)]. Here floor(x) is the greatest integer less or equal
to x. So, x − floor(x) is the fractional part of x. Finally, E(x) is the
equivalence class of the fractional part of x. The map E is continuous
even though it does not appear to be so. If x1 is sligtly smaller than
an integer and x2 is slightly larger than the same integer, then E(x1 )
and E(x2 ) are on opposite sides of [0, 1]. However, the gluing brings
them close together in X.

Exercise 1. Verify that E : R → S 1 is indeed a covering map

in the example(s) given above. Reconcile the two examples and see
that essentially they are the same thing.

6.2. The Deck Group

We are going to give more examples of covering spaces below, but the
whole idea of a covering space is enhanced by another concept—the
deck group. So, we will bring up the deck group before talking more
about covering spaces. We have already associated one group to a
(pointed) metric space, namely the fundamental group. Now we are
 → X be a covering
going to assign a group in a second way. Let E : X
map as above. Say that a deck transformation is a homeomorphism
h:X  →X  such that E ◦ h = E.
As a mnemonic, think about how the deck group relates to shuf-
fling a deck of cards. There is a natural map E, from your deck of
cards to a single card. You can think of holding the deck of cards
directly above the single card and then E is vertical projection. If
you shuffle the cards and redo the map E there is no change. So, a
deck transformation in this case corresponds to shuffling the deck.
In general, you can think of X as a kind of deck of cards and X
as a single card. The analogy isn’t perfect because X  is connected,
6.3. A Flat Torus 67

but for an evenly covered neighborhood U ⊂ X, the set U = E −1 (U )

really is like a deck of cards. The deck transformation h somehow
permutes the disjoint components of U  like shuffling permutes the
cards.
If h is a deck transformation, so is h−1 . Likewise if h1 and h2 are
deck transformations, then so is h1 ◦ h2 . Thus, the set of deck trans-
formations forms a group under composition. This group is called the
deck group of (X, X, E).
Let’s revisit our covering space example considered in the last
section. In both examples, the transformation x → x + 1 is a cov-
ering transformation. In the first case, this follows from the identity
exp(2πi(x + 1)) = exp(2πix). In the second example, it is obvious
from the definition of E.

Exercise 2. Verify that the deck group in the above example is

Z. In other words, the maps x → x + n for n ∈ Z are the only
covering transformations.

Note the deck group of (R, S 1 , E) is Z, the same as π1 (S 1 ); that

is, the deck group and the fundamental group are isomorphic. Below
we will prove a result that gives general conditions under which this
is true.

6.3. A Flat Torus

The next really great example of a covering space is E : R2 → X,
where X is a flat torus. As discussed in §1.9, we can make a flat
torus X by gluing together the opposite sides of a parallelogram P0 ,
as shown in Figure 6.1.

Figure 6.1. The flat torus

68 6. Covering Spaces and the Deck Group

The resulting surface X is homeomorphic to S1 × S1 , and the

fundamental group is isomorphic to Z 2 . There is a nice covering map
from R2 to X. We can tile R2 with translates of P0 , as shown in
Figure 6.2. Given any point x ∈ R2 , we choose a parallelogram Px
such that x ∈ Px . There is a unique translation Tx : Px → P0 and we
define E(x) = [Tx (x)] ∈ X.

Figure 6.2. The parallelogram tiling

The beautiful thing about this map is that it is well defined even
when x lies on the interface between two or more parallelograms. For
example, suppose that x lies on a horizontal edge, as shown in Figure
6.2. Then we could take Px to be the parallelogram either above x
or below x. In the one case Tx (x) would like in the middle of the
top edge of P0 and in the other case Tx (x) would lie in the middle of
the bottom edge of P0 . However, these two points are identified on X.

Exercise 3. Prove that E : R2 → X is a covering map and that

the deck group in this case is precisely the group of translation sym-
metries of the tiling, namely Z 2 . Once again, the deck group and the
fundamental group are isomorphic.
6.4. More Examples 69

There are a few things about the flat torus example that do not
quite represent the general case. For instance, the deck group and
fundamental group are both Abelian, and this is rather a special
situation. However, in spite of the limitations of the torus example,
I would say that it accounts for 80 percent of my intuition about
covering spaces. In any case, it is a good example to learn well! The
example in Exercise 5 below accounts for another 19 percent of my
intuition, and then the last 1 percent comes from more complicated
examples.

6.4. More Examples

Here are two more examples of covering spaces and deck groups. In
the next example, the fundamental group and the deck group are
nontrivial finite groups.

Exercise 4. Let S 2 be the 2-sphere and let P 2 be the projective

plane, defined as the set of equivalence classes of antipodal points on
S 2 . Show that the obvious map S 2 → P 2 is a covering map. (Note:
In order to do this problem, you first have to recall the metric on
P 2 .) Show that the deck group in this example is Z/2. Once again,
the deck group and the fundamental group are isomorphic.

So far, all the examples we have seen have Abelian deck groups.
The next exercise shows an important example in the case when the
group is not Abelian.

Exercise 5. Let X be a space that is homeomorphic to an ∞ symbol,

as shown on the right-hand side of Figure 6.3 below. Let X  be the

4-valent infinite tree. Exhibit a map E : X → X which is a covering
map. (The tree in Figure 6.3 is only partially drawn. It is meant
to go on forever and have valence 4 at each vertex.) Prove that the
deck group for (X, X, E) is isomorphic to the free group on 2 gen-
erators. Once again, the deck group and the fundamental group are
isomorphic.
70 6. Covering Spaces and the Deck Group

Figure 6.3. The 4-valent tree and the figure 8

6.5. Simply Connected Spaces

Recall that a path connected space is one in which every two points
can be joined by a continuous path. Let X be a path connected
metric space. X is said to be simply connected if π1 (X) is trivial. This
definition does not depend on the basepoint, because the isomorphism
type of the fundamental group is independent of basepoint in path
connected spaces. The plane is simply connected and so is a tree.
Suppose that f0 , f1 : [0, 1] → X are two paths. Suppose also that
f0 (0) = f1 (0) and f1 (0) = f1 (1). In other words, the two paths have
the same beginning and the same ending. We say that f0 and f1 are
path homotopic if there is a homotopy F from f0 to f1 such that ft (0)
and ft (1) are independent of t. Here, as usual, ft (x) = F (x, t), where
F is a map on the unit square. Intuitively, a path homotopy slides
the one path to the other without moving the endpoints. In the case
where ft (0) = ft (1), the notion of a path homotopy coincides with
the notion of a loop homotopy.
The next exercise relates the idea of a path homotopy to the idea
of simple connectivity.
6.6. The Isomorphism Theorem 71

Exercise 6. Suppose that X is simply connected. Prove that any two

paths, which have the same endpoints as each other, are homotopic.
(Outline: Let x be the starting point of both loops. Consider the
loop g formed by first doing f0 forward and then doing f1 backward.
Then [g] ∈ π1 (X, x). Hence g is loop homotopic to the identity. Let
G be the loop homotopy. Try to modify G slightly so that G becomes
a path homotopy from f0 to f1 . Figure 6.4 shows what we hope is a
suggestive picture.)

Figure 6.4. Altering a homotopy

Exercise 7. Let {Bi } denote any countable union of disjoint closed

balls in R3 . Prove that R3 − Bi is simply connected.

6.6. The Isomorphism Theorem

Here is the main theorem in this chapter, and (in my opinion) one of
the best theorems in algebraic topology.
Theorem 6.1 (Isomorphism). Suppose that
• E:X → X is a covering map.
 are path connected.
• X and X
 is simply connected.
• X
 X, E).
Then π1 (X) is isomorphic to the deck group for (X,

The rest of the chapter is devoted to proving the Isomorphism

Theorem.
72 6. Covering Spaces and the Deck Group

6.7. The Bolzano–Weierstrass Theorem

A sequence of points {cj } in a metric space X is called Cauchy if,
for every  > 0, there is some N such that i, j > N implies that
d(ci , cj ) < . A convergent sequence is automatically Cauchy, and
one can ask about the converse. X is said to be complete if every
Cauchy sequence in X converges to a point in X.

Exercise 8. Prove that Q, the rationals, is not complete.

The basic axiom for R is that it is complete. You might ask how
one proves that R is complete. The usual way is to construct R from
Q in a way that builds in completeness. Here is the barest sketch of
the idea. Start with the set X of all Cauchy sequences in Q. Define
two Cauchy sequences {ai } and {bi } to be equivalent if the shuffled se-
quence a1 , b1 , a2 , b2 , a3 , b3 , . . . is also a Cauchy sequence. Intuitively,
equivalent sequences (were they to converge) have the same limit. R
is defined as the set of equivalence classes in X. Cauchy sequences
are added, subtracted, multiplied, and (when possible) divided term
by term, and you have to check that these operations respect the
equivalence relation.

Exercise 9. Using the completeness of R as an axiom, prove the

following result. Let Q1 ⊃ Q2 ⊃ Q3 · · · be a nested sequence of cubes
in Rn such that the diameter of Qn tends to 0 as n tends to ∞. Then
Qn is one point. (Hint: Look at the sequence of centers.)

Theorem 6.2 (Bolzano–Weierstrass). A sequence {cn } contained in

the unit cube Q0 has a convergent subsequence.

Proof. Note that Q0 is the union of 2n cubes having half the size as
Q0 . At least one of these subcubes, Q1 , must contain cj for infinitely
many indices. But Q1 is a union of 2n subcubes having half the size as
Q1 . At least one of these subcubes, Q2 , must contain cj for infinitely
many indices. And so on. The intersection Qn , a single point, by
Exercise 2, is the limit of some subsequence of {cj }. 
6.8. The Lifting Property 73

6.8. The Lifting Property

 → X is a covering map. Let Q be a cube, and
In this section, E : X
let f : Q → X be a continuous map. We say that a lift of f is a map
f : Q → X such that E ◦ f = f . This notion is just a generalization
of what we talked about in the previous chapter. The purpose of this
section is to prove the formal version of the result we talked about,
for some examples, in the previous chapter.
We begin with a technical result.

Lemma 6.3. There is some N with the following property. If Q ⊂ Q

is a subcube with side length less than 1/N , then f (Q) is contained
in an evenly covered neighborhood of X.

Proof. If this result is false, then we can find a sequence of subcubes

{Qn }, with diameter tending to 0 such that f (Qj ) is not contained in
an evenly covered neighborhood. Let {cj } be the center of Qj . This
sequence has a convergent subsequence, by the Bolzano–Weierstrass
Theorem. Tossing out everything but the cubes corresponding to this
subsequence, we can assume that {cj } converges to some x ∈ Q. Then
f (x) is contained in some evenly covered neighborhood U . But then
f (Qn ) ⊂ U for n large, by continuity. This is a contradiction. 

Lemma 6.4. Let Q be a cube, and let f : Q → X be a continuous

map. Let v be a vertex of Q, and let x  ∈ X be a point such that
E(x) = f (v). Suppose that f (Q) is contained in an evenly covered
neighborhood. Then there is a unique lift f : Q → X
 such that f(v) =
.
x

Proof. Let U ⊂ X be the evenly covered neighborhood such that

1 , U
f (Q) ⊂ U . Recall that E −1 (U ) is a disjoint union of sets U 2 , . . .
such that the restriction E : Uj → U is a homeomorphism. Let U k
be the component that contains x , and let F be the inverse of the
k . Then we can and must define f = F ◦ f . 
restriction of E to U

Just as we did in the previous chapter, we want to now remove the

hypothesis that f (Q) is contained in an evenly covered neighborhood.
74 6. Covering Spaces and the Deck Group

Theorem 6.5. Let Q be a cube and let f : Q → X be a continuous

∈X
map. Let v be a vertex of Q, and let x  be such that E( x) = f (v).
Then there is a unique lift f : Q → X
 such that f(v) = x
.

Proof. By Lemma 6.3, we can find some N such that any subcube of
diameter less than N is mapped into an evenly covered neighborhood
of f . Let’s partition Q into such cubes, say Q = Q1 , . . . , Qm . We
can order these cubes so that, for each k, the cube Qk shares at least
one vertex vk with some Qj for j < k. Also, we set things up so that
the initial vertex v = v1 is a vertex of Q1 . We define f on Q1 using
Lemma 6.4. This tells us the value of f on v2 and determines how
we define f on Q2 . The uniqueness guarantees that the definition
on Q2 is compatible with the definition on Q1 . The key point is that
Q1 ∩ Q2 is contained in an evenly covered neighborhood. We continue
like this, from cube to cube, until we have defined f in the only way
possible on all of Q. 

We will only need the above result for the case of the unit interval
[0, 1] and the unit square [0, 1]2 , but it is nice to know in general.

6.9. Proof of the Isomorphism Theorem

The proof comes in 4 steps:
(1) Define the isomorphism.
(2) Prove that it is a homomorphism.
(3) Prove that the homomorphism is injective.
(4) Prove that the homomorphism is surjective.

6.9.1. Define the Isomorphism. Since X is path connected, the

isomorphism type of π1 (X, x) is independent of the choice of base-
point. Let x ∈ X be a basepoint. Let G = π1 (X). Let D be the deck
∈X
transformation group. Let x  be some point such that E(x) = x.
We make this choice once and for all. Suppose that h ∈ D is a deck
transformation. Then y = h(
x) is some other point. Note that

y) = E ◦ h(
E( x) = E(
x) = x.
6.9. Proof of the Isomorphism Theorem 75

Since X is path connected, there is some path f : [0, 1] → X such

  
 and f (1) = y. Let f = E ◦ f . By construction, f is a
that f (0) = x
loop based at f . Define

(6.1) Φ(h) = [f ] ∈ G.

To see that Φ is well defined, suppose that f0 and f1 are two loops
connecting x to y. Since X  is simply connected, there is a path
 
homotopy F from f0 to f1 . But then F = E ◦ F is a loop homotopy

from f0 to f1 . Hence [f0 ] = [f1 ] and Φ is well defined. Φ is our map
from D to G.

6.9.2. Homomorphism. This step looks quite mysterious, but it

is fairly obvious if you draw pictures. Let h1 , h2 ∈ D be two deck
transformations. We want to prove that

Φ(h1 ◦ h2 ) = Φ(h1 )Φ(h2 ).

x) for j = 1, 2. Let fj be a path joining x

Let yj = hj (  to yj . Let

fj = E ◦ fj . Then Φ(hj ) = [fj ].
x). Note that h1 ◦ f2 is a path joining the points
Let z = h1 ◦ h2 (

x) = y1
h1 ( and y2 ) = h1 ◦ h2 (
h1 ( x).

Therefore, the concatenated path f1 ∗ (h1 ◦ f2 ) joins x

 to z. But then

Φ(h1 ◦ h2 ) = [E ◦ (f1 ∗ (h1 ◦ f2 ))] = [(E ◦ f1 ) ∗ (E ◦ h1 ◦ f2 )]

=∗ [(E ◦ f1 ) ∗ (E ◦ f2 )] = [f1 ∗ f2 ] = [f1 ][f2 ] = Φ(h1 )Φ(h2 ).

The starred equality comes from the fact that E ◦ h1 = E.

Exercise 10. Choose the example of the flat torus, given above,
and go through the above argument step by step, illustrating the
proof with pictures.

6.9.3. Injectivity. Since Φ is a homomorphism, we can show that

Φ is injective just by showing that Φ has a trivial kernel. So, suppose
that Φ(h) is the trivial element in π1 (X, x).

Lemma 6.6. h( .

x) = x
76 6. Covering Spaces and the Deck Group

Proof. Let y = h( x). We want to show that y = x. Let f be a path
which joins x  to y. It suffices to show that f is path homotopic to
the constant path. Let f = E ◦ f. Then Φ(h) = [f ]. By hypothesis,
there is a loop homotopy F from f to the trivial loop. Let Q be the
unit square. By construction, F : Q → X is a continuous map such
that f0 = f and f1 is the constant map. From the lifting theorem,
there is a lift F : Q → X such that F(0, 0) = x
 and E ◦ F = F . Here
are 3 properties of F:
• f0 is a lift of f0 = f . From the uniqueness of lifts, f0 = f.
• f1 is the constant path since f1 is the constant path.
• F (0, t) and F (1, t) are the basepoint in X, independent of
t. Hence F(0, t) and F(1, t) are constant maps. That is, the
endpoints of ft do not change with t.
From the first item, the endpoints of f0 are x
 and y. From the second
item, the endpoints of f1 are x
 and x. From the third item, we see
 = y.
that the two sets of endpoints coincide, forcing x 

The following lemma finishes our injectivity proof.

Lemma 6.7. If h is a deck transformation such that h( , then
x) = x
h is the identity.

Proof. Let y be some other point of X.  We want to show that

h( 
y ) = y. Let f be a path joining x
 to y. Let x = E(
x) and y = E(y).
Let f = E ◦ f. Then f : [0, 1] → X is a path which joins x to y.
The paths f and h ◦ f are both lifts of f which agree at 0. That

is, f (0) = x  and h ◦ f(0) = h(x) = x. By uniqueness of lifts, these
two lifts are the same. In particular, y = f(1) = h ◦ f(1) = h(
y). 

6.9.4. Surjectivity. Let [g] ∈ π1 (X, x) be some element. We want

to produce a deck transformation h such that Φ(h) = [g]. Let y ∈ X 
be any point. We need to define h(  
y ). So, let f be a path joining x
to y. Let f = E ◦ f. Then f is a path in X joining x to y = E( y).
Consider the concatenated path γ = g ∗ f . From the lifting property
we can find a lifted path γ which joins x  to some other point, which we
define as h(y ). Figure 6.5 illustrates the construction in case X = R2
2
and X = T , the torus.
6.9. Proof of the Isomorphism Theorem 77

y h(y)

f
g
x

Figure 6.5. Lifted paths

Exercise 11. Show that the definition of h(

y ) is independent of the
choices of f and g. (Hint: imitate the proof given in the previous
section.)

To compute Φ(h), we consider the case that y = x . Then we

can take f to be the trivial path. In this case γ
 is a path joining x

x) and E ◦ γ
to h(  differs from g = E ◦ g just by concatenating the
constant loop. Assuming that h is a deck transformation, we have
Φ(h) = [γ] = [g].
To finish the proof, we just have to show that h is a deck trans-
formation.

Lemma 6.8. E ◦ h = h.

Let’s compute E ◦ h(

y ). By construction, both γ and f connect
x to y. We have

E ◦ h( (1) = γ(1) = y = f (1) = E ◦ f(1) =2 E(

y) =1 E ◦ γ y).
78 6. Covering Spaces and the Deck Group

Equality 1 comes from the fact that γ(1) = h(y ) by definition. Equal-
ity 2 comes from the fact that f(1) = y, by definition.
Lemma 6.9. h is continuous.

Proof. Let y ∈ X  be a point. Let y = E( y). There is an evenly


covered neighborhood U ⊂ X of y. Let U − 1 be the component
of h−1 (U ) which contains y. Let U 2 = h(U 1 ). Then U2 is another
−1
component of h (U ) because E ◦ h = E. Let Fj be the inverse of the
restriction of E to U 1 . Being the composition
j . Then h = F2 ◦ E on U
of continuous maps, h is continuous. 

Were we to make the above construction for the element [g]−1 ,

we would produce the map h−1 . Hence h is invertible. The same
argument as above shows that h−1 is continuous. Hence h is a home-
omorphism. Now we know that h belongs to the deck group. This
completes our proof.
Chapter 7

Existence of Universal
Covers

In the previous chapter, we proved the Isomorphism Theorem, a result

 X, E) to the fundamental group π1 (X).
which relates the triple (X,
Here X  is a simply connected covering space of X and E : X  →X
is a covering map. X  is known as the universal cover of X. We use
the word “the” because, as it turns out, any two universal covering
spaces of X are homeomorphic to each other.
In this chapter we will prove the existence (but not uniqueness) of
a universal cover X under certain assumptions on X. The conditions
we place on X are somewhat contrived, but we want to streamline the
existence proof. Our main interest in this result is the case when X
is a compact surface, and any compact surface satisfies the conditions
we impose.
The reader interested in seeing the fully general existence and
uniqueness proof should consult an algebraic topology book such as
[HAT]. The exact condition on X that guarantees the existence of X 

is that X is semilocally simply connected , and in all such cases X is
unique.

79
80 7. Existence of Universal Covers

7.1. The Main Result

Given a metric space M and two continuous paths f0 , f1 : [0, 1] → M ,
we define
(7.1) D(f0 , f1 ) = sup d(f0 (t), f1 (t)).
t∈[0,1]

Let x ∈ M . We say that the pair (M, x) is conical if, for each
y ∈ M there is a continuous path γy : [0, 1] → M such that γy (0) = x
and γy (1) = y. We insist that γx is the trivial path, and also we make
the following continuity requirement. For any y ∈ M and any  > 0,
there is some δ > 0 such that d(y, z) < δ implies that D(γy , γz ) < .
The idea behind our definition is that you are making M into a
kind of cone, with x as the apex. The pair (Rn , 0) is a prototypical
example of a conical pair. The paths you can use in this example are
just line segments traced out at unit speed.

Exercise 1. Prove (M, x) is conical if M is homeomorphic to Rn .

Say that the path f0 in M is good if there is some  > 0 with

the following property: Suppose that D(f0 , f1 ) <  and f0 and f1
have the same endpoints. Then there is a homotopy F from f0 to
f1 which does not move the endpoints. That is, ft (0) and ft (1) are
independent of t.

Definition 7.1. A metric space X is good if every path in X is

good and every point x ∈ X is such that the ball B (x) is both sim-
ply connected and conical for some  > 0. The value of  is allowed
to vary with the point and the path.

Here is our main result.

Theorem 7.1. Any good metric space has a simply connected cover.

Exercise 2. Prove that a flat torus is good.

Exercise 3. Prove that any finite graph is good.

7.1. The Main Result 81

Exercise 4 (Challenge). Prove that any compact surface is good.

(Hint: In the proof of Theorem 12.10 we sketch the argument for
complete hyperbolic surfaces.)

Exercise 5. Give an example of a metric space that has no non-

trivial good paths. (Hint: Swiss cheese.)

Here is the construction of X and the map E : X  → X. Choose

a basepoint x ∈ X. We define X  to be the set of pairs (y, [f ]) where
y ∈ X is a point and f is a path which joins x to y. Here [f ] denotes
the path homotopy equivalence class of f . That is, [f1 ] = [f2 ] if and
only if there is a homotopy from f1 to f2 that does not move the
endpoints.
So far X is just a set. We define

(7.2) D([f0 ], [f1 ]) = inf D(f0 , f1 ).

The infimum is taken over all paths f0 which represent [f0 ] and all
paths f1 which represent [f1 ]. Finally, we define

(7.3)  0 , [f0 ]), (y1 , [f1 ])) = d(y0 , y1 ) + D([f0 ], [f1 ]).
d((y

Exercise 6. Prove that d is a metric on X.  (Hint: The only hard

 q) = 0 implies p = q. Here
part of this exercise is showing that d(p,
 This amounts to showing that D([f0 ], [f1 ]) = 0 implies that
p, q ∈ X.
[f0 ] = [f1 ]. Deduce this from the goodness of X.)

There is an obvious map E : X  → X, given by E(y, [f ]) = y.

There are a few things about E that we can see right away. Since E
does not increase distances, E is a continuous map. Also, E is onto
because X is path connected.

Exercise 7. Use the fact that X is path connected to prove that

 is also path connected.
X

It remains to prove that E is a covering map and that X  is

simply connected. We will prove these two statements in the next
two sections.
82 7. Existence of Universal Covers

7.2. The Covering Property

Let y ∈ X be a point, and let U be an -ball about y, chosen to be both
simply connected and conical. Let H denote the set of path homotopy
classes of curves joining x to y. We first produce a homeomorphism Ψ
from E −1 (U ) to U × H. This is a formal way of saying that E −1 (U )
is a disjoint union of copies of U .

x y

Figure 7.1. The path f ∗ γ(z, y)

For any z ∈ U , let γ(y, z) be the path joining y to z, as specified

by the definition of a conical metric space. Let γ(z, y) denote the
reverse path. Let (z, [f ]) ∈ E −1 (U ) be a point. We define
(7.4) Ψ((z, [f ])) = (z, [f ∗ γ(z, y)]).
See Figure 7.1. If f0 and f1 are both representatives of [f ], then a
path homotopy from f0 to f1 extends to a path homotopy from f0 ∗ γ
to f1 ∗ γ. Hence [f0 ∗ γ] = [f1 ∗ γ]. Hence, our map Ψ is well defined.

Lemma 7.2. Ψ is a bijection.

Proof. Suppose Ψ(z0 , [f0 ]) = Ψ(z1 , [f1 ]). Then z0 = z1 . We set

z = z0 = z1 . We know that [f0 ∗ γ(z, y)] = [f1 ∗ γ(z, y)]. Writing
γ = γ(z, y), we have [f0 ∗ γ] = [f1 ∗ γ] but then
[f0 ] = [f0 ∗ γ ∗ γ −1 ] = [f1 ∗ γ ∗ γ −1 ] = [f1 ].
This shows that Ψ is injective.
Now we show that Ψ is surjective. Given any pair (z, [g]) ∈ U ×H,
the path f = g ∗ γ(y, z) connects x to z. The two paths g and
f ∗ γ(z, y) = g ∗ γ(y, z) ∗ γ(z, y)
are clearly homotopic. Hence Ψ(z, [f ]) = (z, [g]). 
7.2. The Covering Property 83

We put a metric on U × H by declaring that points in different

components are 1 apart. Within a single component, U × {h}, we
just use the metric we already have on U .

Lemma 7.3. Ψ is a homeomorphism.

Proof. We already know that Ψ is a bijection. We just have to show

that Ψ and Ψ−1 are both continuous. We will consider Ψ. Suppose
that (z0 , [f0 ]) and (z1 , [f1 ]) are very close. Then f0 ∗ γ(z0 , y) and
f1 ∗ γ(z1 , y) are two very nearby paths, both having endpoints x and
y. Since X is good, we have
[f0 ∗ γ(z0 , y)] = [f1 ∗ γ(z1 , y)]
once these paths are sufficiently close. Also z0 and z1 are very close.
So, the second coordinates of Ψ(z0 , [f0 ]) and Ψ(z1 , [f1 ]) agree, and the
first coordinates are very close. This shows (a bit informally) that Ψ
is continuous.
Now we consider Ψ−1 . Using the notation from the proof of the
previous lemma, we have
Ψ−1 (z, [g]) = (z, [f ]),
where f = g ◦ γ −1 . If (z0 , [g0 ]) and (z1 , [g1 ]) are less than 1 apart,
then [g0 ] = [g1 ]. But then, we can use the same path g to represent
both [g0 ] and [g1 ]. But then f0 = g ∗ γ(z0 , y)−1 and f1 = g ∗ γ(z1 , y)−1
are also close. This shows that Ψ−1 is continuous. 

Now we know that Ψ is a homeomorhism from E −1 (U ) to U × H.

Let π : U × H → U be projection onto U . Then the restriction of
π to each component of U × H is clearly a homeomorphism. These
components are of the form U × {h}, where h ∈ H.
Finally, note that
(7.5) E = π ◦ Ψ.
For each component U  of E −1 (U ) there is some h ∈ H so that
Ψ(U ) = U × {h} and Ψ is a homeomorphism from U  to U × {h}.

But then the restriction to U of E = π ◦ Ψ is the composition of two
homeomorphisms, and hence a homeomorphism. This completes the
proof that E is a covering map.
84 7. Existence of Universal Covers

7.3. Simple Connectivity

We take the basepoint x  ∈ X  to be the pair (x, ∗) where ∗ is the
trivial loop connecting x to x. Suppose f : [0, 1] → X  is a loop. This
means that f (t) = (xt , [γt ]), where xt ∈ X and γt is a path connecting
x to xt . Both [γ0 ] and [γ1 ] are trivial elements of π1 (X).
Let β(s) = xt . Define βt : [0, 1] → X by the formula
(7.6) βt (s) = β(st).
Note that βt and γt are both paths which join x to xt .
Lemma 7.4. [βt ] = [γt ] for all t ∈ [0, 1].

Proof. Let J be the set of parameter values for which [βt ] = [γt ].
We have 0 ∈ J because β0 and γ0 are both trivial in π1 (X, x). We
show J = [0, 1] by showing that J is both closed and open.
Closed : Suppose that [βt ] = [γt ] for a sequence of t values con-
verging to s. Since β and f are both continuous,
(xs , [γs ]) = lim (xt , [γt ]) = lim (xt , [βt ]) = (xs , [βs ]).
t→s t→s

Therefore [βs ] = [γs ].

Open: Suppose [βt ] = [γt ]. Let βst denote the restriction of β to
[s, t]. For s close to t we can take γt ∗ βst as a representative for [γs ].
Here we are using the fact that E : X  → X is a covering map. But
then
[γs ] = [γt ∗ βst ] = [βt ∗ βst ] = [βs ].
The central equality comes from the fact that [βt ] = [γt ]. 

By Lemma 7.4 we have

(7.7) f (t) = (β(t), [βt ]).
 the point f (1) = (x, [β]) is just the basepoint
Since f is a loop in X,
 Hence [β] is the trivial element in π1 (X, x).
in X.
For any null loop β, we get the path f = fβ defined by equa-
tion (7.7). The loop fβ depends continuoutly on the loop β. As β
shrinks down to a point, fβ shrinks down to the constant map. This
shows that f is homotopic to a constant map, and hence X is simply
connected.
Part 2

Surfaces and Geometry

Chapter 8

Euclidean Geometry

This chapter begins the second part of the book. It is the first in a
series of 3 chapters in which we consider the classical 2 dimensional
geometries. In this chapter we will prove some results about Euclidean
geometry in the plane. Since Euclidean geometry is so familiar, we
will not spend too much time on the basics. Following an introductory
first section, we will concentrate on interesting theorems. Most of the
theorems revolve around the theme of cutting complicated polygons
into simpler ones.

8.1. Euclidean Space

The standard dot product on Rn is given by the formula

(8.1) (x1 , . . . , xn ) · (y1 , . . . , yn ) = x1 y1 + · · · + xn yn .

The norm of a vector X = (x1 , . . . , xn ) is given by

√
(8.2) X = X · X.

The dot product satisfies the fundamental Cauchy–Schwarz In-

equality. We will give two proofs of this inequality.

Lemma 8.1. For any vectors X and Y , we have

|X · Y | ≤ XY .

87
88 8. Euclidean Geometry

Assuming Y is nonzero, we get equality if and only if X is a multiple

of Y .

First Proof. To avoid trivialities, assume Y is nonzero. For any

choice of t, we have
X2 + t2 Y 2 + 2t(X · Y ) = X − tY  ≥ 0.
Plugging in t = (X · Y )/Y 2 , multiplying through by Y 2 , and
simplifying, we get the inequality. The only way to get equality is
that X − tY  = 0. But then X = tY .

The proof above is the standard proof. Now I will give a second
proof which, though more involved, makes the result look less myste-
rious.

Second Proof. If c and s are real numbers such that c2 + s2 = 1,

then the map
⎛ ⎞ ⎛ ⎞
x1 cx1 + sx2
⎜ x2 ⎟ ⎜−sx1 + cx2 ⎟
(8.3) R12 ⎜ ⎟ ⎜
⎝. . .⎠ = ⎝
⎟
⎠
...
xn xn
preserves the dot product. The map R12 changes coordinates 1 and
2 and leaves the rest alone. There is an analogous symmetry Rij
(depending on c and s) which changes coordinates i and j and leaves
the rest alone. Applying suitable choices of these symmetries, we can
reduce to the special case when Y = (x1 , 0, . . . , 0). In this case, the
inequality is obvious.

The Euclidean distance Rn is given by the formula

(8.4) d(X, Y ) = X − Y .
Lemma 8.2. d satisfies the triangle inequality.

Proof. For any vectors A and B, we have

A + B2 = (A + B) · (A + B)
(8.5) = A2 + 2(A · B) + B2
≤∗ A2 + 2AB + B2 ≤ (A + B)2 .
8.1. Euclidean Space 89

The starred inequality follows from the Cauchy–Schwarz inequality.

Hence
A + B ≤ A + B.
Setting A = X − Y and B = Y − Z, we see that
d(X, Y ) = X − Z = A + B ≤ A + B
≤ X − Y  + Y − Z = d(X, Y ) + d(Y, Z).
This holds for any triple X, Y, Z of vectors, and thereby completes
the proof. 

The angle θ between two vectors X and Y obeys the equation

X ·Y
(8.6) cos(θ) = .
XY 
To understand this equation, we consider the case X = Y  = 1.
We can use compositions of the isometries mentioned above to rotate
so that X = (1, 0, . . . , 0) and Y = (c, s, 0, . . . , 0), where c2 + s2 = 1.
Then, we have
(8.7) cos(θ) = X · Y = c.
This last equation matches our expectation that cos(θ) is the first
coordinate of a unit vector in the plane that makes an angle of θ with
the positive x-axis.
Now that we have defined distances and angles in Euclidean
space, we talk a bit about volumes of solids. Given n linearly in-
dependent vectors V1 , . . . , Vn in Rn , the parallelepiped spanned by
these vectors is defined as the set of all linear combinations

aj vj , a1 , . . . , an ∈ [0, 1].
The volume of this parallelepiped is given by
 
n
(8.8) det(V1 , . . . , Vn ) = (−1)|σ| Vi,σ(i) .
σ i=1

The sum takes place over all permutations σ. The quantity |σ| is 0 if
σ is an even permutation and 1 if σ is an odd permutation. Finally,
Vij is the jth component of Vi . If you have not seen the definition
of the determinant before, this book is not place to learn it. See any
book on linear algebra.
90 8. Euclidean Geometry

It would be nice if every solid body could be decomposed into

finitely many parallelepipeds. Then one could define the volume of
an arbitrary solid body by summing up the volumes of the pieces.
Unfortunately, this doesn’t work, and one must resort to some kind
of limiting process. For instance, you fill up a given solid, as best
as possible, with increasingly small cubes, and take a limit of the
corresponding sums. This is what is typically done in a calculus class.
This procedure suffices to give a satisfactory definition of volume for
household solids, such as spheres and ellipsoids.
Taking a measure-theoretic approach vastly broadens the number
of solid bodies whose volume one can define in a satisfactory way.
With the exception of Chapter 22, where we prove the Banach–Tarski
Theorem, we will always deal with very simple solids for which all
reasonable definitions of volume coincide.

8.2. The Pythagorean Theorem

Our definition of distance in R2 somewhat has the Pythagorean The-
orem built into it.√ The distance from the point (a, b) to (0, 0) is
defined to be c = a2 + b2 . So, we automatically have a2 + b2 + c2 .
Here a, b and c are the side lengths of the right triangle with vertices
(0, 0) and (a, 0) and (a, b). Note that this triangle is rather special:
two of its sides are parallel to the coordinate axes.
Here we will prove the Pythagorean Theorem for an arbitrary
right triangle in the plane. There are many, many proofs; I’ll present
my two favorites.

A B
A B
C
B C
A
C B
C
C
A
B A
Figure 8.1. Two views of the Pythagorean Theorem
8.3. The X Theorem 91

Referring to the left half of Figure 8.1, the outer square has area
(A + B)2 . At the same time, the outer square breaks into 4 right
triangles, each having area AB/2, and an inner square having area
C 2 . Hence (A + B)2 = 2AB + C 2 . Simplifying gives A2 + B 2 = C 2 .
That is the first proof.
Here is the second proof. For any right triangle, there is a con-
stant k such that the distance from the right-angled vertex to the
hypotenuse is k times the length of the hypotenuse. This constant k
only depends on the shape of the triangle, and not on its size. By the
base times height formula for area, the area of the triangle is kL2 ,
where L is the length of the hypotenuse. Again, the constant k only
depends on the shape of the triangle and not on its size. The three
triangles on the right-hand side of Figure 8.1 have the same shape.
The large one has area kC 2 , and the two small ones have area kA2
and kB 2 . Hence kC 2 = kA2 + kB 2 . Cancelling the k (a constant we
don’t care about) gives A2 + B 2 = C 2 .

8.3. The X Theorem

Here we prove a classic result from high school geometry. Let S 1 be
the unit circle in the plane and let A and B be two chords of S 1 , as
shown on the left-hand side of Figure 8.2. Let L(A, B) be the length
of the region R(A, B) ⊂ S 1 opposite the two acute angles of A ∩ B.
(In case A ⊥ B we choose arbitrarily.) Figure 8.2 shows R(A, B)
drawn thickly.

Figure 8.2. The chords A and B.

92 8. Euclidean Geometry

Theorem 8.3 (The X Theorem). L(A, B) only depends on the acute

angle θ(A, B) between A and B and not on the positions.

Proof. To see this, imagine that A and B are toothpicks that we

can roll to a new location. The right-hand side of Figure 8.2 shows
what happens when roll A is parallel to itself. By symmetry (about
the line perpendicular to the direction of motion) the same length of
arc is added to one side of R(A, B) as is subtracted from the other.
Hence, the sum of the lengths does not change. The same goes when
we roll B parallel to itself. At the same time, rotating the disk by
any amount changes neither the angle between A and B nor L(A, B).
Rotating and rolling as necessary, we can get to any position without
changing L(A, B). 

When A and B cross at the center of S 1 , we have L(A, B) =

2θ(A, B). By the X Theorem, this result holds in general.
As a limiting case, the X Theorem applies when A ∩ B ∈ S 1 . In
this case, we can reformulate the result. We fix two points x1 , x2 ∈ S 1
and consider the angle θ(y) between yx1 and yx2 as a function of
y ∈ S 1 . The X Theorem implies that θ(y) is independent of y.

8.4. Pick’s Theorem

During college I learned Pick’s Theorem from a friend and classmate
of mine, Sinai Robins. If you want to learn a whole lot about Pick’s
Theorem and its higher-dimensional generalizations, see the the book
[BRO] by Matthias Beck and Sinai Robins.

Figure 8.3. Some lattice polygons

8.4. Pick’s Theorem 93

Let Z 2 ⊂ R2 denote the ordinary lattice of integer points. Say

that a lattice polygon is a polygon in R2 whose vertices lie in Z 2 .
That is, the vertices have integer coordinates. Figure 8.3 shows some
examples. Let P be a lattice polygon. We let i(P ) denote the number
of vertices contained in the interior of the region bounded by P . We
let e(P ) denote the number of vertices contained on the edges of P .
(The vertices of P are included in the count for e(P ).)

Theorem 8.4 (Pick). The area of the region bounded by P is

e(P )
i(P ) + − 1.
2
For the examples in Figure 8.3, you can of course verify the for-
mula directly. During our proof, we will often use the phrase “the area
of P ”, when we really mean to say “the area of the region bounded
by P ”. We hope that this slight abuse of terminology does not cause
confusion.

Exercise 1. Let P be a parallelogram whose vertices have integer

coordinates. Prove that the area of P is an integer. (Hint: Work in
C and translate so that the vertices are 0 and V and W and V + W .
Then establish the formula area(P ) = Im(V W ).)

We say that a lattice parallelogram P is primitive if i(P ) = 0 and

e(P ) = 4.

Lemma 8.5. Pick’s Theorem holds for primitive parallelograms.

Proof. By Exercise 1, the parallelogram P has integer area. To finish

the proof, we just have to show that P has area at most 1.
Let X be the square torus obtained by identifying the opposite
sides of the unit square. Note that X has area 1. Let E : R2 → X
be the universal covering map. See §6.3. Let P o denote the interior
of the region bounded by the primitive parallelogram P .
We claim that E(P o ) is embedded in X. Otherwise, we can find
two points x1 , x2 ∈ P o such that e(x1 ) = e(x2 ). But then x1 − x2 ∈
Z 2 . Let V be the vector whose tail is x1 and whose head is x2 . This
is a vector with integer coordinates. Using the convexity of P , we can
94 8. Euclidean Geometry

find a vector W parallel to V whose tail is a vertex of P and whose

head lies either on the interior of an edge of P or in P0 . Figure 8.4
shows the situation.

P
V

Figure 8.4. Translating a vector

Since W ∈ Z 2 , and the vertices of P are in Z 2 , the head of W

lies in Z 2 . But then we either have i(P ) > 0 or e(P ) > 4, which is a
contradiction. Now we know that E(P ) is embedded. Since E(P ) is
embedded, we see that
area(P ) = area(E(P )) ≤ area(X) = 1.
This completes the proof. 

We say that a primitive triangle is a lattice triangle T such that

i(T ) = 0 and e(T ) = 3.

Exercise 2. Prove Pick’s Theorem for primitive triangles.

We say that P dissects into two lattice polygons P1 and P2 if

• P1 and P2 bound disjoint open regions, and P1 ∩ P2 is a
connected arc.
• The closed region bounded by P is the union of the closed
region bounded by P1 and the closed region bounded by P2 .

Lemma 8.6. Suppose that P dissects into P1 and P2 . If Pick’s The-

orem holds for P1 and P2 , then it also holds for P .

Proof. Let A = area(P ) and A1 = area(P1 ), etc. Obviously A =

A1 + A2 . Let n denote the number of vertices on P1 ∩ P2 . Let
8.4. Pick’s Theorem 95

i = i(P ) and i1 = i(P1 ), etc. We have

i = i1 + i2 + n − 2, e = e1 + e2 − 2n + 2.
Therefore,
i + e/2 − 1
= i1 + i2 + n − 2 + e1 /2 + e2 /2 − n + 1 − 1
= (i1 + e2 /2 − 1) + (i2 + e2 /2 − 1) =∗ A1 + A2 = A.
The starred equality comes from Pick’s Theorem applied to P1 and
P2 . 

P1

P
P2

Figure 8.5. Dissecting a polygon

Exercise 3. Suppose that P is a lattice polygon that is not a primi-

tive triangle. Prove that P can be dissected into two lattice polygons.

By Exercise 3, any lattice polygon can be written as the finite

union of primitive triangles, each of which have area 1/2. Hence, any
lattice polygon has area which is a half-integer. The rest of our proof
goes by induction on the area.

Lemma 8.7. If P is a lattice polygon with area at most 1/2 then P

is a primitive triangle. In particular, Pick’s Theorem holds for P .

Proof. Applying Exercise 3 iteratively, we see that any lattice poly-

gon can be divided into primitive triangles. If P is not a primitive
triangle, then P can be divided into at least 2 primitive triangles.
But each such triangle has area 1/2. This would force P to have area
at least 1. 

Now let P be a general lattice polygon. If P is not a primitive

triangle, we can dissect P into two lattice polygons P1 and P2 having
96 8. Euclidean Geometry

smaller area. By induction Pick’s Theorem holds for P1 and P2 . But

then Pick’s Theorem holds for P as well. This completes the proof.

8.5. The Polygon Dissection Theorem

We continue with the theme of polygon dissections. Here we prove
a classic result about polygon dissections. This result is called the
Bolyai–Gerwein Theorem, but the earliest attribution I have seen is
to a work by William Wallace from 1807; see [WAL]. A dissection of
a polygon P is a description of P as the union
P1 ∪ · · · ∪ P n
of smaller polygon, no two of which overlap. That is, the polygons
have disjoint interiors.
Two polygons P and P  are said to be dissection equivalent if
there are dissections

n n
P = Pi , P = Pi
i=1 i=1

such that Pi and Pi are isometric for all i = 1, . . . , n. In this case, we
write P ∼ P  .

Exercise 4. Prove that ∼ is an equivalence relation.

Figure 8.6 illustrates why a triangle is always equivalent to a

parallelogram.

Figure 8.6. Equivalence between a triangle and a parallelogram

Figure 8.7 illustrates why a parallelogram is always equivalent to

a rectangle.
8.5. The Polygon Dissection Theorem 97

Figure 8.7. Equivalence between a parallelogram and a rectangle

Combining the two facts we have just illustrated, we see that a

triangle is always equivalent to some rectangle. Let R(A, B) be a
rectangle with side lengths A and B. We take A < B.
Lemma 8.8. Let A ∈ (A, B). Then R(A, B) ∼ R(A , B  ). Here B 
is such that A B  = AB. In particular, any rectangle is equivalent to
a square.

Proof. Figure 8.8 shows a 2 step construction, based on a real pa-

rameter t ∈ (0, B). The first part of the figure shows that R ∼ S, and
the second part shows that S ∼ T . The two central figures are both
copies of S, but we have chosen to emphasize a different decomposi-
tion in each copy. The shape of the rectangle T varies continuously
with the parameter t! The construction works when t is small, and
continues to work until we reach some t0 so that the point x(t0 ) coin-
cides with a corner of T (t0 ). But, in this extreme case, T is a square.
As t varies in [0, t0 ], the rectangle T (t) interpolates between R(A, B)
and a square. 
R S

S T
Figure 8.8. Two part construction
98 8. Euclidean Geometry

Lemma 8.9. A triangle of area A is equivalent to a 1 × A rectangle.

Proof. First of all, our triangle is equivalent to some rectangle. By

the previous result, any two rectangles of the same area are equivalent.


Now we can finish the proof. It suffices to prove the result for
unit area polygons. Let P be a polygon of unit area. We first dissect
P into finitely many triangles T1 , . . . , Tm , having areas a1 , . . . , am .
Each Tk is equivalent to a rectangle R(1, ak ). But, when we stack
up all these rectangles, we get a rectangle having side lengths 1 and

ak = 1. That is, any unit area polygon is equivalent to the unit
square. The final result is immediate.
You might wonder whether the same result holds for polyhedra
in higher dimensions. This turns out to be false, and the result is
known as Dehn’s Dissection Theorem. We will give a proof of Dehn’s
Dissection Theorem in Chapter 23.

8.6. Line Integrals

We now discuss line integrals as a preparation for presenting and
proving Green’s Theorem. This material can be found in any book
on several variable calculus; see, for instance, [SPI].
A linear functional is a linear map from R2 to R. A 1-form on an
open subset U ⊂ R2 is a smooth choice p → ωp of a linear functional
at each point p ∈ U . We mention two special 1-forms, dx and dy.
These 1-forms are defined on every point of R2 , and
(8.9) dx(v1 , v2 ) = v1 , dy(v1 , v2 ) = v2 ,
for any tangent vector (v1 , v2 ) based at any point. One can write a
general 1-form ω as a pointwise varying linear combination of these
two special ones. That is,
(8.10) ω = f dx + gdy,
where f, g : U → R are smooth functions. At the point p, we have
(8.11) ωp (V ) = f (p)v1 + g(p)v2 .
Here V = (v1 , v2 ) is some vector based at p.
8.6. Line Integrals 99

Let γ : [0, 1] → R be a smooth curve, and let ω be a 1-form. We

define
  1
ω= ωγ(t) (γ  (t))dt.
γ 0

Exercise 5. Prove that

  
ω1 + ω2 = ω1 + ω2 .
γ γ γ

In other words, the integral is linear.

Exercise 6. Prove that

 
ω=− ω.
−γ γ

Here −γ is the curve obtained by reversing the direction of γ.

It turns out that the integral only depends on the image and
orientation of γ. If
s : [0, 1] → [0, 1]
is an orientation-preserving diffeomorphism, then setting β = γ ◦ s,
we have

Lemma 8.10.  
ω= (ω).
β γ

Proof. By Exercise 5, it suffices to consider the forms f dx and gdy.

The proof for gdy is the same as for f dx, so we will just consider the
case ω = f dx. In this case we set γ(t) = (u(t), v(t)) and note that
  1
ω= (f u ) dt.
γ 0

Here u = du/dt. At the same time

  1  1 
d(u ◦ s) ∗
ω= f ◦ s(t) dt = f u ◦ s(t) s (t)dt.
β 0 dt 0

The starred equality is the chain rule. The first integral equals the
last by the change-of-variables formula for integration. 
100 8. Euclidean Geometry

Here is an important observation. Since the line integral only

depend on the oriented image of γ, we can specify a line integral just
by specifying a curve in the plane and its orientation.
Line integrals can be more generally defined for piecewise smooth
curves. To say that γ is a piecewise smooth curve is to say that
γ = γ1 ∪ · · · ∪ γn , where each γj is a smooth curve, and consecutive
curves meet end to end. We define
 n 
ω= ω.
γ j=1 γj

In particular, line integrals make sense for polygonal arcs.

Exercise 7. This is a crucial exercise. Let P1 and P2 and P be

the polygons from Figure 8.4. Suppose that all these polygons are
oriented counterclockwise. Prove that
  
ω= ω+ ω.
P P1 P2

8.7. Green’s Theorem for Polygons

Let D be a polygon in the plane, and let γ = ∂D, the boundary of
D oriented counterclockwise. Let ω = f dx + gdy be a 1-form defined
in an open set that contains D in its interior. Green’s Theorem says
that
 
(8.12) ω= (gx − fy ) dxdy.
γ D

Here fy = ∂f /∂y and gx = ∂g/∂x. The integral on the right is a

double integral.
In our proof, it is convenient to let dω be the integrand on the
right hand side of equation (8.12). We will just use this piece of
notation to shorten our equations, but actually dω has a meaning as
the exterior derivative of ω. See [SPI] if you are curious about this.
We say that a special triangle is a right triangle whose sides are
parallel to the coordinate axes. The three white triangles in Figure
8.9 below are examples of special triangles.

Exercise 8. Let D be the special triangle with vertices (0, 0) and

8.7. Green’s Theorem for Polygons 101

(A, 0) and (0, B) with A and B positive. Let γ be the boundary of

D, oriented counterclockwise. Let ω = f dx. Prove that
  A
ω= (f (x, 0) − f (x, x ))dx,
γ 0

where x (as a function of x) is such that (x, x ) lies on the diagonal



of D.

Lemma 8.11. Green’s Theorem is true for special triangles.

Proof. Let D be a special triangle. We can translate the whole pic-

ture so that the vertices of D are as in Exercise 8. By the Fundamental
Theorem of Calculus, we get
   A   x 
dω = (−fy ) = (−fy )dy dx
D D x=0 y=0
 A 
= (f (x, 0) − f (x, x ))dx = ω.
0 γ
The last equality comes from Exercise 8. 

Our next result has an easy direct proof, but we will give a rather
long-winded proof to illustrate a crucial property of line integrals.

Lemma 8.12. Green’s Theorem is true for any rectangle whose sides
are parallel to the coordinate axes.

Proof. Let R be such a rectangle. We write R = T1 ∪ T2 , where

T1 and T2 are two special triangles meeting along a diagonal. We
certainly have   
dω = dω + dω.
R T1 T2
On the other hand, by Exercise 7, we have
  
ω= ω+ ω.
∂Rd ∂T1 ∂T2

Here ∂R denotes the boundary of R taken counterclockwise, and

likewise for the other expressions. Since Green’s Theorem holds for
special triangles, we can equate the right-hand sides of our last two
102 8. Euclidean Geometry

equations. But then we can equate the left-hand sides as well. Hence
Green’s Theorem holds for R. 
Lemma 8.13. Green’s Theorem is true for any triangle.

Proof. Figure 8.9 shows how we can realize an arbitrary triangle D

as a set of the form R − T1 − T2 − T3 , where R is a rectangle and Tk
is a special triangle for k = 1, 2, 3. We have
  
dω + dω = dω.
D Tk R
The same cancellation trick as in the previous lemma shows that
  
ω+ ω= ω.
∂D ∂Tk ∂R
Green’s Theorem, applied to cases we already know, allows us to
cancel off all terms, leaving just the one we don’t know. 

Figure 8.9. A union of triangles

Lemma 8.14. Green’s Theorem is true when the domain D is an
arbitrary polygon.

Proof. Partition D into triangles and apply the same cancellation

trick as above. 
Chapter 9

Spherical Geometry

The purpose of this chapter is to prove some results about spherical

geometry. As usual, S 2 denotes the unit sphere in R3 . Most of
the results in this chapter can be found in any book on differential
geometry; see, for instance, [BAL]. The one topological result, the
Hairy Ball Theorem, can be found in most topology books; see, for
instance, [GPO].

9.1. Metrics, Tangent Planes, and Isometries

S 2 has two natural metrics on it. The easiest one to define is the
chordal metric: the distance between p, q ∈ S 2 is p − q. This just
uses the Euclidean metric on R3 .
The other metric is often called the round metric. We define the
length of a curve on S 2 to be its length when considered a curve in
R3 . So, if γ : [a, b] → S 2 is a differentiable curve, we have
 b
(9.1) L(γ) = γ  (t) dt.
a

The distance between two points p and q in the round metric is the
infimum of the lengths of all paths on S 2 that join p to q. We will see
below that this infimum is realized by a path that is an arc of a great
circle. We will see in Chapter 11 that this way of defining a metric is
part of a general construction.

103
104 9. Spherical Geometry

In this chapter we will ignore the chordal metric and work with
the round metric. Fortunately, any isometry of the chordal metric is
an isometry of the round metric and vice versa. The point is that one
can give a formula for the one metric in terms of the other. This will
become more clear when we work out what the shortest paths are in
the round metric.
Later, when we study Riemannian surfaces, we will see that an
object called the tangent plane plays a fundamental role in the theory.
For the case of the sphere, the tangent plane has a very short and
simple definition. The tangent plane to S 2 at the point p ∈ S 2 is the
plane Tp (S 2 ) such that p ∈ Tp (S 2 ) and Tp (S 2 ) is perpendicular to the
vector pointing from 0 to p. The tangent plane has the following nice
property. Any curve γ : [a, b] → S 2 is such that the velocity γ  (t) lies
in the tangent plane Tγ(t) (S 2 ).
Any rotation of R3 gives rise to an isometry of S 2 . One such
rotation is given by the matrix

⎡ ⎤
cos(t) sin(t) 0
Mt = ⎣− sin(t) cos(t) 0⎦ .
0 0 1

This map rotates by t around the z axis and thus rotates S 2 about
the north and south poles. One can find similar maps that rotate
around the other two coordinate axes. We call a rotation about one
of the coordinate axes a basic rotation.
Just by composing the basic rotations, we can move any one point
of S 2 to any other point. Moreover, once we know that we can move
any point of S 2 to any other, we see that we can find an isometry of
S 2 that fixes any given point and rotates through an angle t about
that point. Indeed, if T : S 2 → S 2 is an isometry that carries (0, 0, 1)
to p, then T Mt T −1 is the desired rotation about p.
All the isometries we have described so far come from orientation-
preserving linear maps of R3 . The other “half” of the isometries come
from orientation-reversing linear maps of R3 . One such isometry is
given by the map (x, y, z) → (x, y, −z). This map interchanges the
north and south poles of S 2 and fixes the equator. More generally, if
9.2. Geodesics 105

v ∈ S 2 is any point, the map

(9.2) Tv (w) = −w + 2(v · w)v

is an orientation-reversing isometry. The point is that Tv is obviously
a linear transformation, and a short calculation shows that

Tv (w1 ) · Tv (w2 ) = w1 · w2 .

Note also that Tv (v) = −v, so that Tv swaps v and −v. We call the
maps in equation (9.2) basic reflections.

9.2. Geodesics
There are many equivalent definitions of a geodesic. To avoid a
buildup of terminology, we will give a definition that only relies on
what we have already presented. A geodesic on S 2 is a curve γ :
[a, b] → S 2 with the following properties:
• γ has constant speed.
• If t1 and t2 are any sufficiently close parameters in [a, b],
then the restriction of γ to [t1 , t2 ] is the shortest curve on
S 2 that joins γ(t1 ) to γ(t2 ). In other words, γ is locally a
length-minimizing curve.
We will see that a curve is a geodesic if and only if it has constant
speed and its image lies in a great circle. A great circle is the inter-
section of a plane through the origin with S 2 . The study of geodesics
on S 2 is a classical one. It is treated in essentially every book on
differential geometry. Here we just establish a few basic facts.

Lemma 9.1. The shortest differentiable path joining two points on

the sphere exists and is an arc of a great circle.

Proof. Let x and y be two points. We rotate so that x is the north

pole. For convenience, we assume that y is not the south pole, so
that there is a unique great circle C joining x and y. Let γ be any
differentiable curve that joins x to y.
There is a map φ : S 2 → C. The point φ(p) is the point of C that
lies on the same line of latitude. Geometrically, we think of rotating
106 9. Spherical Geometry

S 2 around the north and south poles, and watching p rotate around
until it sticks on C.
The differential dφp is a map from the tangent plane Tp (S 2 ) onto
the line tangent to C at φ(p). We note two properties of this map.
• If v is parallel to a line of longitude, then dφ(v) = v||.
• If w is parallel to a line of latitude, then dφ(v) = 0.
Note also that the lines of longitude and latitude are perpendicular
whenever they intersect.
These properties imply that dφ is a distance nonincreasing map.
Moreover, dφ strictly decreases the length of any tangent vector that
is not parallel to a line of longitude. Therefore, the length of φ(γ) is
strictly less than the length of γ unless γ traces out a line of longitude.
But then γ traces out C, because the lines of longitude are great circles
and only one great circle connects x to y (in our case). 

Recall that two points x, y ∈ S 2 are called antipodal if x = −y.

Two points are antipodal if and only if they lie on more than one
great circle. In case x and y are not antipodal, we define the geodesic
connecting x to y to be the shorter of the two great circular arcs
connecting x to y.
Now that we know about geodesics, we can prove a basic result
about isometries of S 2 .

Lemma 9.2. Any isometry of S 2 is a composition of basic reflections.

Proof. Note that a basic rotation, i.e. a rotation about one of the
coordinate axes, is a composition of two basic reflections. So, if we
can prove that every isometry is the composition of basic rotations
and basic reflections, then we have proved that every isometry is the
composition of basic reflections.
Let I be a mystery isometry of S 2 . Since we can move any point
2
of S to any other point using compositions of basic rotations, we can
compose I with basic rotations so that the result fixes (0, 0, 1). So,
without loss of generality, we can assume that I fixes (0, 0, 1).
The equator E on S 2 is the set of points of the form (x, y, 0).
The equator divides the sphere into the upper hemisphere and the
9.3. Geodesic Triangles 107

lower hemisphere. Any point on the lower hemisphere is farther from

(0, 0, 1) than is any point on the upper hemisphere. For this reason,
I(E) = E. Now, E is a circle and I is an isometry of E. So, I acts
on E either as a rotation or a reflection. Composing I with basic
reflections and/or rotations, we can assume that I fixes every point
on E.
Any point p ∈ E is connected to (0, 0, 1) by the arc γp , which is
one quarter of a great circle. Since I fixes the endpoints of γp , and γp
is the unique shortest path joining p to (0, 0, 1), we have I(γp ) = γp .
Moreover, I preserves distances along γp . Hence I fixes every point of
γp . Since p is an arbitary point of E, we see that I fixes every point
of the upper hemisphere. A similar argument shows that I fixes every
point of the lower hemisphere. Hence, I is the identity. 

9.3. Geodesic Triangles

Let d denote the distance on S 2 . If x and y are antipodal, then
d(x, y) = π. In general, d(x, y) = θ, where θ is the angle between the
vector pointing to x and the vector pointing to y. Familiar formulas
in linear algebra give

(9.3) cos(d(x, y)) = x · y, sin(d(x, y)) = x × y.

Here × is the cross product. What makes these formulas simple is

the fact that x = y = 1.
We measure angles on S 2 using the dot product on R3 . Suppose
that C1 and C2 are two geodesics connecting x to y1 and y2 . The
angle between C1 and C2 at x is just the angle between the tangent
vectors at x. This is the same as the dihedral angle between the
plane Π1 containing (0, x, y1 ) and the plane Π2 containing (0, x, y2 ).
As usual, there are two angles we can measure at x, and the sum of
these angles is π.
Let x1 , x2 , x3 be three points, all contained in the same hemi-
sphere. Then there is a unique geodesic Cj joining xj−1 and xj+1 ,
with the indices taken cyclically. The union of these geodesics is
called a spherical triangle. Let θj be the interior angle at xj , and let
Lj denote the length of Cj .
108 9. Spherical Geometry

There is a beautiful formula for the area of a spherical triangle,

known as Giraud’s Theorem. (Thomas Harriot discovered the result
in 1603 but did not publish it.) The area is given by
(9.4) θ1 + θ2 + θ3 − π.
This result is a special case of the general Gauss–Bonnet Theorem, a
result proved much later on. Here we sketch a proof of Giraud’s Theo-
rem. The case when the 3 points lie on the same great circle is trivial.
In all other cases, the whole triangle lies in an open hemisphere.
Say that a lune is a region bounded by two great semicircles. A
lune has two vertices. By symmetry, the interior angles at either end
of the lune are the same. Any two lunes having the same interior
angles are isometric to each other. Let A(θ) be the area of a lune
having angle θ.

Lemma 9.3. A(θ) = 2θ for all θ ∈ [0, π].

Proof. If θ = π, the lune is precisely a hemisphere. Hence

(9.5) A(π) = 2π.
Moreover, a lune having interior angle θ decomposes into n lunes
having interior angle θ/n. Hence
(9.6) A(θ) = nA(θ/n).
Combining equations (9.5) and (9.6) we see that A(θ) = 2θ whenever
θ is a rational multiple of π. But A is a continuous function of θ.
Hence A(θ) = 2θ for all θ. 

Now let T be a geodesic triangle, contained in a hemisphere,

having interior angles θ1 , θ2 , and θ3 . Extending the sides of T , we
can cover S 2 by 6 lunes.
Figure 9.1, which needs some interpretation, shows the situation.
In Figure 9.1, we have drawn T extremely small, and placed near
(say) the north pole. We are looking down on T . The sides of T
practically look straight because T is very small. We have extended
the sides of T and partially shown them. These sides continue all the
way around S 2 and join up again near the south pole, where they form
another copy T  of T . Our technical assumption about T lying in a
9.3. Geodesic Triangles 109

hemisphere guarantees that T and T  are disjoint. We have drawn the

boundary of T thickly, and we have shaded two of the lunes. These
two lunes meet at both vertices, the other vertex being near the south
pole.

Figure 9.1. Dissected sphere

By Lemma 9.3, the total area of the lunes is

(9.7) 4(θ1 + θ2 + θ3 ).

With the exception of some points on the edges of the lunes, every
point of S 2 − (T ∪ T  ) is covered once by a lune. At the same time,
every point of T ∪ T  is covered 3 times by the union of the lunes.
Letting A be the area of T (and T  ) we have

(9.8) 4(θ1 + θ2 + θ3 ) = (4π − 2A) + 6A = 4π + 4A.

Simplifying this equation gives A = θ1 + θ2 + θ3 − π, as desired.

Exercise 1. Try to draw a version of Figure 9.1 that shows the

entire sphere, as well as the 6 lunes.
110 9. Spherical Geometry

9.4. Convexity
It doesn’t really make much sense to talk about the convexity of a
general subset of S 2 , because some pairs of points on S 2 can be joined
by more than one shortest path. However, if X ⊂ S 2 is entirely
contained in an open hemisphere H, then any two points of X can be
joined by a unique arc of a great circle that has length less than π.
This arc remains inside H. We call this the geodesic segment joining
the points.
We call X convex if the geodesic segment joining any pair of
points in X remains in X. This definition appears to depend on H,
but it does not.

Exercise 2. Prove that the notion of convexity for X does not de-
pend on the hemisphere relative to which it is defined. That is, if
X is contained in the intersection H1 ∩ H2 of two open hemispheres,
then X is convex relative to either one of them.

If X ⊂ H ⊂ S 2 is an arbitrary set, we define the convex hull of

X to be the intersection of all the closed convex subsets of H that
contain X. We call this set Hull(X).

Exercise 3. Prove that Hull(X) is well defined, independent of the

open hemisphere that contains X. Prove also that Hull(X) is convex
relative to any open hemisphere that contains it.

The purpose of the next 2 exercises is to establish some back-

ground results needed for the Cauchy Rigidity Theorem, proved in
Chapter 24. We say that a convex spherical polygon is a simple closed
polygonal curve in S 2 made from arcs of great circles that is contained
in a hemisphere and bounds a convex set contained in that same
hemisphere. We insist that consecutive arcs make an angle which is
distinct from π. From Exercise 3, the definition of a convex spheri-
cal polygon does not depend the choice of hemisphere that contains it.

 be a great
Exercise 4. Let Γ be a convex spherical polygon. Let C
9.5. Stereographic Projection 111

circle that extends a side C of Γ. Then Γ − C is contained in one of

the two open hemispheres bounded by C. 

Exercise 5. Let Q and Q be convex spherical quadrilaterals. Let

the sides of Q be C1 , C2 , C3 , C4 . Let θj be the interior angle between
Cj and Cj+1 . Make the same definitions for Q . Suppose that Cj
and Cj have the same length for all j. Label a vertex of Q by a
(+) if θ > θ  at that vertex, and by a (−) if the opposite inequality
holds. Prove that the labels of the vertices of Q must have the form
(+, −, +, −) or (0, 0, 0, 0), up to cyclic ordering.

9.5. Stereographic Projection

Let C denote the complex numbers. We think of ∞ as an extra
point and consider C ∪ ∞. We want to think of C ∪ ∞ as a sphere.
To do this, we want to put a metric on C ∪ ∞ so that the result
is homeomorphic to a sphere. The metric we get on C ∪ ∞ is not
really so natural, but it does allow us to speak of continuous maps
from C ∪ ∞ to itself. This is something we will take up in the next
chapter.
One way to put a metric on C ∪ ∞ is to choose a map from S 2
to C ∪ ∞ which is a homeomorphism from C to S 2 minus a single
point, say (0, 0, 1). Then, we put a metric on C ∪ ∞ so that our map
is an isometry. One very nice map from S 2 to C ∪ ∞ is stereographic
projection.

(0,0,1)

Figure 9.2. Stereographic projection

112 9. Spherical Geometry

As Figure 9.2 (drawn 1 dimension down) illustrates, stereographic

projection has the following geometric description. We identify C
with the horizontal plane R2 × {0} in R3 . Half of S 2 lies above this
plane and half below. We map (0, 0, 1) to ∞. Given any other point
p ∈ S 2 , we define φ(p) ∈ C to be the point such that (0, 0, 1) and p
and φ(p) are collinear.
The formula is given by
   
x y
(9.9) φ(x, y, z) = + i.
1−z 1−z

The inverse map is given by

 
−1 2x 2y 2
φ (x + iy) = , ,1− .
1 + x2 + y 2 1 + x2 + y 2 1 + x2 + y 2

One can check easily that these maps are inverses of each other.

Exercise 6. Check that our formula for stereographic projection

matches the geometric description.

Exercise 7. Check that φ gives a homeomorphism from S 2 − (0, 0, 1)

to C.

One of the nice facts about stereograpic projection is the follow-

ing. If C ⊂ S 2 is a circle, then φ(C) is either a circle in C or else
a straight line (union ∞). When C contains the point (0, 0, 1), this
result is fairly obvious from the geometric description. The idea is
that any circle C ⊂ S 2 has the form ΠC ∩ S 2 for some plane ΠC .
When (0, 0, 1) ∈ ΠC , we see from the geometric description that

φ(ΠC ) = (C ∩ ΠC ) ∪ ∞.

A general geometric proof, based on conic sections, is given in [HCV].

In §14.3, we will give a proof based on complex analysis.

Exercise 8 (Challenge). Find your own proof that stereographic

projection maps circles in S 2 to either circles or straight lines in C.
9.6. The Hairy Ball Theorem 113

9.6. The Hairy Ball Theorem

Let me end the chapter with the Hairy Ball Theorem. This is really
a result about the topology of the sphere, and not its geometry, but
it is such a great result that I wanted to put it in.
A unit field on S 2 is a continuous choice of unit vector tangent
to S 2 at each point. The Hairy Ball Theorem says that a unit field
on S 2 does not exist. The name of the theorem (which has somewhat
fallen out of favor) comes from the following interpretation: If you
have a sphere that is completely covered in hair, you cannot comb
the hair so that it lies flat and varies continuously. There has to be
some kind of cowlick somewhere.
We will suppose that a unit field exists, and derive a contradic-
tion. Suppose we have a unit field U on S 2 . Let γ : [0, 1] → S 2 be the
a smooth loop, so that γ(0) = γ(1). For each t ∈ [0, 1), we let θ(t) de-
note the counterclockwise angle between the tangent vector γ  (t) and
our vector field at γ(t). We choose so that θ(t) varies continuously.
As t → 1, the value θ(t) necessarily tends to an integer multiple of
2π. We let
(9.10) N (U , γ) = lim θ(t) − θ(0).
t→1

Intuitively, you are walking along γ, turning your head according the
direction of U . Once you get back to where you start, you will be
looking in the same direction as when you started, except that your
head will be turned around N times counterclockwise! Compare this
discussion about the winding number given in §5.1,

Exercise 9. Prove that the quantity N (U , γ) is independent of

the smooth parametrization of γ, as long as the orientation does not
switch. Also prove that N (U , γ) = N (U , γ  ) when γ and γ  are ho-
motopic loops. (Hint: N (U , γ) is continuous and integer-valued.)

Consider the case when γ is a small loop that winds once around
the north pole. As we walk around γ, keeping our head aligned with
the unit field, our head always points in roughly the same direction.
So, when we make one complete circuit, our neck is twisted once
around. (Don’t try this at home.) That is N (U , γ) = ±1. If we
114 9. Spherical Geometry

replace γ by −γ, the loop that is oriented in the opposite direction,

the sign of N (U , γ) switches.

Figure 9.3. Two homotopies

We orient γ so that N (U , γ) = 1. There are two ways to slide γ

to a small loop around the south pole. On the one hand, we can push
γ around the side, following a single line of longitude and keeping γ
small. On the other hand, we can pull γ down over the whole sphere,
moving through the circles of latitude. Figure 9.1 shows a top-down
view of the two methods. One method leads to a small loop β about
the south pole and the other method leads to −β, the oppositely
oriented loop. By Exercise 1, we have
N (U , β) = N (U , γ) = N (U , −β) = −N (U , β) = 1.
This equation says that 1 = −1, which is a contradiction. This proves
the Hairy Ball Theorem.
Chapter 10

Hyperbolic Geometry

The purpose of this chapter is to give a bare bones introduction to

hyperbolic geometry. Most of material in this chapter can be found in
a variery of sources, for example [BE1], [KAT], [RAT], or [THU].
The first 2 sections of this chapter might not look like geometry at
all, but they turn out to be very important for the subject.

10.1. Linear Fractional Transformations

Now we take up the discussion started in §1.6. Suppose that
 
a b
A=
c d

is a 2 × 2 matrix with complex number entries and determinant 1.

The set of these matrices is denoted by SL2 (C). In fact, this set
forms a group under matrix multiplication.
The matrix A defines a complex linear fractional transformation
az + b
TA (z) = .
cz + d
Such maps are also called Möbius transformations. Note that the
denominator of TA (z) is nonzero as long as z = −d/c. It is convenient
to introduce an extra point ∞ and define TA (−d/c) = ∞. This

115
116 10. Hyperbolic Geometry

definition is a natural one because of the limit

lim |TA (z)| = ∞.
z→−d/c

The determinant condition guarantees that a(−d/c) + b = 0, which

explains why the above limit works. We define TA (∞) = a/c. This
makes sense because of the limit
lim TA (z) = a/c.
|z|→∞

Exercise 1. As in §9.5, we introduce a metric on C ∪ ∞ so that

C ∪ ∞ is homeomorphic to the unit sphere S 2 ⊂ R3 . Prove that
TA is continuous with respect to this metric. (Hint: Use the limit
formulas above to deal with the tricky points.)

Exercise 2. Establish the general formula

TAB = TA ◦ TB ,
where A, B ∈ SL2 (R). In particular (since A−1 exists) the inverse
map TA−1 exists. By Exercise 1, this map is also a continuous map of
C ∪ ∞. Conclude that TA is a homeomorphism of C ∪ ∞.

10.2. Circle Preserving Property

A generalized circle in C ∪∞ is either a circle in C or a set of the form
L ∪ ∞, where L is a straight line in C. Topologically, the generalized
circles are all homeomorphic to circles. In this section we will prove
the following well-known result.
Theorem 10.1. Let C be a generalized circle and let T be a linear
fractional transformation. Then T (C) is also a generalized circle.

One can prove this result by a direct (though tedious) calcula-

tion. The book [HCV] has a nice proof involving the geometry of
stereographic projection. For fun, I will give a rather unconventional
proof. I’ll prove 4 straightforward lemmas and then give the main
argument.
Lemma 10.2. Let C be any generalized circle in C. Then there
exists a linear fractional transformation T such that T (R ∪ ∞) = C.
10.2. Circle Preserving Property 117

Proof. If C is a straight line (union ∞), then a suitable translation

followed by rotation will work. So, consider the case when C is a
circle. The linear fractional transformation
z−i
T (z) =
z+i
maps R ∪ ∞ onto the unit circle C0 satisfying the equation |z| = 1.
The point is that every point z ∈ R is the same distance from i
and −i, so that |T (z)| = 1. Next, one can find a map of the form
S(z) = az + b that carries C0 to C. The composition S ◦ T does the
job. 
Lemma 10.3. Suppose that L is a closed loop in C ∪ ∞. Then there
exists a generalized circle C that intersects L in at least 3 points.

Proof. If L is contained in a straight line (union ∞) the result is

obvious. Otherwise, L has 3 noncollinear points and, like any 3 non-
collinear points, these lie on a common circle. 
Lemma 10.4. Let (z1 , z2 , z3 ) = (0, 1, ∞). Let a1 , a2 , a3 be a triple
of distinct points in R ∪ ∞. Then there exists a linear fractional
transformation that preserves R ∪ ∞ and maps ai to zi for i = 1, 2, 3.

Proof. The map T (z) = 1/(a3 − z) carries a3 to ∞, but does not

necessarily do the right thing on the points a1 and a2 . However, we
can compose T by a suitable map of the form z → rz + s to fix the
images of a1 and a2 . 
Lemma 10.5. Suppose T is a linear fractional transformation that
fixes 0 and 1 and ∞. Then T is the identity map.

Proof. Let
az + b
T (z) = .
cz + d
The condition T (0) = 0 gives b = 0. The condition T (∞) = ∞ gives
c = 0. The condition T (1) = 1 gives a = d. Hence T (z) = z. 

Now we can give the main argument. Suppose that there is a

linear fractional transformation T and a generalized circle C such
that T (C) is not a generalized circle. Composing T with the map
from Lemma 10.2, we can assume that C = R ∪ ∞. By Lemma 10.3
118 10. Hyperbolic Geometry

there is a generalized circle D such that D and T (R ∪ ∞) share at

least 3 points. Call these 3 points c1 , c2 , c3 .
Again by Lemma 10.2, there is a linear fractional transformation
S such that S(R ∪ ∞) = D. There are points a1 , a2 , a3 ∈ R ∪ ∞ such
that S(aj ) = cj for j = 1, 2, 3. Also, there are points b1 , b2 , b3 ∈ R∪∞
such that T (bj ) = cj for j = 1, 2, 3. By Lemma 10.4 we can find linear
fractional transformations A and B, both preserving R ∪ ∞ such that
A(aj ) = zj and B(bj ) = zj for j = 1, 2, 3. Here (z1 , z2 , z3 ) = (0, 1, ∞).
The two maps
T ◦ B −1 , S ◦ A−1
both map (0, 1, ∞) to the same 3 points, namely (c1 , c2 , c3 ). By
Lemma 10.5, these maps coincide. However, note that

T ◦ B −1 (R ∪ ∞) = T (R ∪ ∞)

is not a generalized circle and S ◦ A−1 (R ∪ ∞) = D is a generalized

circle. This is a contradiction.

10.3. The Upper Half-Plane Model

Now we turn to hyperbolic geometry. We are going to imitate the
procedure we used in §9.1 to define the round metric on the sphere.
Once we define the hyperbolic plane as a set of points, we will define
what we mean by the lengths of curves in the hyperbolic plane. Then,
we will proceed as in the case of the sphere.
Let U ⊂ C be the upper half-plane, consisting of points z with
Im(z) > 0. As a set, the hyperbolic plane is just U . However, we will
describe a funny way of measuring the lengths of curves in U . Were
we to use the ordinary method, we would just produce a subset of
the Euclidean plane. So, given a differentiable curve γ : [a, b] → U ,
we define
 b
|γ  (t)|
(10.1) L(γ) = dt.
a Im(γ(t))

In words, the hyperbolic speed of the curve is the ratio of its Euclidean
speed to its height above the real axis.
10.3. The Upper Half-Plane Model 119

Here is a simple example. Consider the curve γ : R → U defined

by
γ(t) = i exp(t).
Then the length of the portion of γ connecting γ(a) to γ(b), with
a < b, is given by
 b  b
exp(t)
dt = dt = b − a.
a exp(t) a

The image of γ is an open vertical ray, but our formula tells us that
this ray, measured hyperbolically, is infinite in both directions. More-
over, the formula tells us that γ is a unit speed curve: it accumulates
b − a units of length between time a and time b.
The hyperbolic distance between two points p, q ∈ U is defined
to be the infimum of the lengths of all piecewise differentiable curves
connecting p to q. Let us consider informally what these shortest
curves ought to look like. Suppose that p and q are very near the real
axis, say
p = 0 + i 10−100 , q = 1 + i 10−100 .
The most obvious way to connect these two points would be to use
the path
γ(t) = t + i 10−100 .
This curve traces out the bottom of the (Euclidean unit) square shown
in Figure 10.1. Our formula tells us that this curve has length 10100 .

Figure 10.1. Some paths in the hyperbolic plane

120 10. Hyperbolic Geometry

Another thing we could do is go around the other three sides of

the square. For the left vertical edge, we could use the path γ from
our first calculation. This edge has length

log(1) − log(10−100 ) = 100.

The top horizontal edge has height 1 and Euclidean length 1. So,
this leg of the path has length 1. Finally, by symmetry, the length
of the right vertical edge is 100. All in all, we have connected p to q
by a path of length 201. This length is obviously much shorter than
the first path. It pays to go upward because, so to speak, unit speed
hyperbolic curves cover more ground the farther up they are. Our
second path is much better than the first but certainly not the best.
For openers, we could save some distance by rounding off the corners.
We will show in §10.6 below that the shortest curves, or geodesics, in
the hyperbolic plane are either arcs of vertical rays or arcs of circles
that are centered on the real axis.
When U is equipped with the metric we have defined, we call
U the hyperbolic plane and denote it by H 2 . So far we have talked
about lengths of curves in H 2 , but we can also talk about angles.
The angle between two differentiable and regular (i.e., nonzero speed)
curves in H 2 is defined simply to be the ordinary Euclidean angle
between them. That is, the hyperbolic and Euclidean angle between
two intersecting curves is just the Euclidean angle between the two
tangent vectors at the point of intersection. So, in the upper half-
plane model of hyperbolic geometry, the distances are distorted (from
the Euclidean model) but the angles are not.
Now that we have talked about hyperbolic length and angles,
we discuss hyperbolic area. Given how hyperbolic length relates to
Euclidean length, it makes sense to say that the area of a small patch
of hyperbolic space is the ratio of its Euclidean area to its height
squared. Since the “height” of a patch varies throughout the patch,
we really have something infinitesimal in mind. Thus, precisely, we
define the hyperbolic area of a region D ⊂ H 2 to be the integral

dx dy
(10.2) .
D y2
10.5. Symmetries 121

10.4. Another Point of View

An inner product on a real vector space V is a map  ,  : V × V → R
which satisfies the following properties:
• av + w, x = av, x + w, x for all a ∈ R and v, w, x ∈ V .
• x, y = y, x.
• x, x) ≥ 0 and x, x) = 0 if and only if x = 0.
You can remember this by noting that an inner product satisfies the
same formal properties as the dot product.
For the moment, we care mainly about inner products on R2 . At
the point z = x + iy we introduce the inner product
1
(10.3) v, wz = (v · w).
y2
We mean to apply this to vectors v and w that are “based at” z. We
then define the hyperbolic norm to be

(10.4) vH 2 = v, vz .

With this definition, the length of γ : [a, b] → H 2 is given by

 b
(10.5) γ  (t)γ(t) dt.
a

With this formalism, the notion of hyperbolic length looks much closer
to the Euclidean notion. In Chapter 11 we will see that this way of
doing things is the beginning of Riemannian geometry.

10.5. Symmetries
The hyperbolic metric has more symmetries than you might think.
Say that a real linear transformation is a linear transformation TA
based on a matrix with real entries. In this case, TA (z) ∈ C provided
z ∈ C − R.

Exercise 3. Prove that z ∈ R implies that TA (z) ∈ R. Prove

also that TA maps H 2 into itself.
122 10. Hyperbolic Geometry

The element TA is a homeomorphism of C ∪ ∞ which preserves

H 2.

Exercise 4. We say that a real linear fractional transformation is

basic if it has one of three forms:
• T (z) = z + 1.
• T (z) = rz.
• T (z) = −1/z.
Prove that any real linear fractional transformation is the composi-
tion of basic ones.

It turns out that these maps are all hyperbolic isometries. This
is pretty obvious for the map T (z) = z + 1. The hyperbolic metric
is built so that the second map is a hyperbolic isometry, and in a
moment we will give two proofs of that fact. The really surprising
thing is that the third map turns out to be a hyperbolic isometry as
well.

Lemma 10.6. The map T (z) = rz is a hyperbolic isometry.

First Proof. If γ is any curve in H 2 , then the dilated curve T (γ)

moves r times as fast in the Euclidean sense but is r times farther
from the real axis. Hence T (γ) and γ move at the same hyperblic
speed at corresponding points. So, if we connect points p and q by
some curve γ we can connect the points T (p) and T (q) by the curve
T (γ), which has the same length—and vice versa. This shows that
the distance from p to q is the same as the distance from T (p) to
T (q).

Second Proof. Suppose that v and w are two vectors based at

z ∈ H 2 . Then we think of dT (v) = rv and dT (w) = rw as two
vectors based at T (z). Here dT is linear differential of T , i.e., the
matrix of first partial derivatives. Looking at the formula in equation
(10.3), we see that
1 1
dT (v), dT (w)T (z) = rv, rwrz = (rv·rw) = (v·w) = v, wz .
r2 y2 y2
10.6. Geodesics 123

So, T preserves the hyperbolic inner product at each point. Since the
hyperbolic metric is defined entirely in terms of this family of inner
products, T is an isometry.

Exercise 5. Prove that the map T (z) = −1/z is a hyperbolic isom-

etry.

Combining Exercises 4 and 5, we see that any real linear fractional

transformation is a hyperbolic isometry of H 2 . Recall that in §2.8
we proved SL2 (R) is a 3-dimensional manifold. So, H 2 has a 3-
dimensional group of symmetries!
Say that a generalized circular arc is an arc of a generalized cir-
cle. We already know that any linear fractional transformation maps
generalized circles to circles. Hence, any real linear transformation
maps generalized circular arcs to generalized circular arcs.

Exercise 6. Prove that a real linear fractional transformation T

has the following property: if a and b are two smooth curves in H 2
which intersect at a point x and make an angle of θ, then T (a) and
T (b) make the same angle θ at the point T (x). (Hint: If you don’t
feel like grinding out the calculation, you can assume the result is
false and then deduce that the differential dT fails to map circle to
circles. In any case, the result is obvious for all the basic maps except
z → −1/z, and so it suffices to consider this one.)

10.6. Geodesics
In this section we will describe the shortest curves connecting two
points in H 2 . We first consider the case of points p and q that lie on
the imaginary axis.
Lemma 10.7. The portion of the imaginary axis connecting p to q
is the unique shortest curve in H 2 that connects p to q.

Proof. Our proof is very similar to the proof we gave in Lemma 9.1
for the spherical case. Consider the map F defined by the equation
F (x + iy) = iy; see Figure 10.2. Looking at the definition of the
hyperbolic metric, we see that F is hyperbolic speed nonincreasing.
124 10. Hyperbolic Geometry

That is, if γ is a curve in H 2 , then the hyperbolic speed of F (γ) at

any point is at most the hyperbolic speed of γ at the corresponding
point. Moreover, if the velocity of γ has any x-component at all, then
F (γ) is slower at the corresponding point. The idea here is that F
does not change the y-component of the hyperbolic speed, but kills
the x-component. The total hyperbolic length of γ is the integral of
its hyperbolic speed. Thus the hyperbolic length of F (γ) is less than
the hyperbolic length of γ, unless γ travels vertically the whole time.
Our result follows immediately from this. 

Figure 10.2. The map F

It follows from symmetry that the vertical rays in H 2 are all

geodesics. A vertical ray is the unique shortest path in H 2 connect-
ing any pair of points on that ray.

Exercise 7. Let p and q be two arbitary points in H 2 . Prove

that there is a hyperbolic isometry—specifically, some linear frac-
tional transformation—that carries p and q to points that lie on the
same vertical ray.

Theorem 10.8. Any two distinct points in H 2 can be joined by a

unique shortest path. This path is either a vertical line segment or
else an arc of a circle that is centered on the real axis.

Proof. We have already proved this result for points that lie on the
same vertical ray. in light of Exercise 7, it suffices to prove, in general,
10.7. The Disk Model 125

that the image of a vertical ray under a linear fractional hyperbolic

isometry is one of the two kinds of curves described in the theorem.
Let ρ be a vertical ray, and let T be a linear fractional transfor-
mation that is also a hyperbolic isometry. From the work in §10.2
we know that T (ρ) is an arc of a circle. Since T preserves R ∪ ∞,
both endpoints of this circular arc lie on R ∪ ∞. Finally, since T pre-
serves angles, T (ρ) meets R at right angles at any point where T (ρ)
intersects R. If T (ρ) limits on ∞, then T (ρ) is another vertical ray.
Otherwise, T (ρ) is a semicircle, contained in a circle that is centered
on the real axis. 

10.7. The Disk Model

Now that we have defined geodesics in the hyperbolic plane, we can go
forward and define geodesics polygons. Before we do this, we would
like to have another model in which to draw pictures. This other
model is sometimes more convenient.
Let Δ be the open unit disk. There is a (complex) linear fractional
map M : H 2 → Δ given by
z−i
(10.6) M (z) = .
z+i
This map does the right thing because z ∈ H 2 is always closer to i
than to −i and so |M (z)| < 1. Since M maps circles to circles and
preserves angles, M maps geodesics in H 2 to circular arcs in Δ that
meet the unit circle at right angles.
Sometimes it is convenient to draw pictures of geodesics in the
unit disk rather than in the hyperbolic plane. So, when it comes time
to draw pictures, we will be drawing circular arcs that meet the unit
circle at right angles. The geodesics that go through the Euclidean
center of Δ are just unit line segments. The rest of them “bend in-
ward” toward the origin.

Exercise 8. Draw pictures of 10 geodesics in the disk model.

Rather than just think of Δ as a convenient place to draw pic-

tures, we can also think of Δ as another model of H 2 . The cheapest
126 10. Hyperbolic Geometry

way to do this is to say that the distance the two points p, q ∈ Δ is

defined to be the hyperbolic distance between the points M −1 (p) and
M −1 (q) in H 2 .
A more direct approach is to define a new inner product at each
point z ∈ Δ. The formula is given by

4v · w
(10.7) v, wz = .
(1 − |z|)2

Once we have this inner product, we can directly define lengths of

curves in Δ as in equation (10.5). Then we can define distances in Δ
as in the upper half-plane model. It turns out that this new method
produces the same result as the cheap method. The proof is a calcu-
lation similar to our second proof of Lemma 10.6. We just prove that
M is an isometry relative to the inner product on H 2 and the inner
product on Δ.

Exercise 9. Prove that the map M is an isometry from H 2 and

Δ, when lengths are defined in terms of the inner product in equa-
tion (10.7). That is, prove that

v, wz = dM (v), dM (w)M (z)

for any pair of vectors v and w based at z ∈ H 2 .

The disk Δ, equipped with its metric, is known as the Poincaré

disk model of the hyperbolic plane. When T is a real linear fractional
transformation, the map M ◦ T ◦ M −1 is an isometry of Δ. Since
M preserves angles, the hyperbolic angle between two curves in Δ is
the same as the Euclidean angle between them. Thus, in both our
models, Euclidean and hyperbolic angles coincide.
Before we continue, we mention one more piece of terminology.
The ideal boundary of H 2 is defined to be R ∪ ∞ in the upper half-
plane model and the unit circle in the disk model. Points on the ideal
boundary are called ideal points. The ideal points are not points in
H 2 . They are considered “limit points” of geodesics in H 2 .
10.8. Geodesic Polygons 127

10.8. Geodesic Polygons

Now that we have our two models of the hyperbolic plane, and we
know that the geodesics are, we are ready to consider geodesic poly-
gons in the hyperbolic plane. To save words, we will use the term H 2
rather loosely to refer to either of our two models of the hyperbolic
plane. Since there is an isometry, namely M , carrying one model to
the other, there doesn’t seem to be much harm in doing this.
Say that a geodesic polygon in H 2 is a simple closed path made
from geodesic segments. Here, “simple” means that the path does not
intersect itself. Say that a solid geodesic polygon is the region in H 2
bounded by a geodesic polygon. It is convenient to allow some of the
“vertices” of the polygon to be ideal points. We call such “vertices”
by the name ideal vertices. The interior angle of a polygon at an ideal
vertex is 0: the two geodesics both meet the ideal point perpendicular
to the ideal boundary.
We point out a special geodesic triangle, called an ideal triangle.
An ideal triangle is a geodesic triangle having 3 infinite geodesic sides
and 3 ideal vertices; see Figure 10.3 below. The main result in this
section, the Gauss–Bonnet formula for hyperbolic geodesic triangles,
is the hyperbolic analogue of the result in §9.3. The proof is very
similar, too.
Theorem 10.9. Let T be a geodesic triangle in the hyperbolic plane.
The area of T equals π minus the sum of the interior angles of π. In
particular, the sum of these interior angles is less than π.

We will give the same kind of proof that we gave for the analogous
result in §9.3.
Lemma 10.10. Theorem 10.9 holds for ideal triangles.

Proof. We are trying to prove that any ideal triangle has area π.
You can move any one ideal triangle to any other using an isometry
of H 2 . So, it suffices to prove this result for a single triangle. Let
us prove this for the triangle T , in the upper half-plane model, with
vertices −1 and 1 and ∞. We first observe that
 ∞
1
2
dy = 1/y0 .
y=y0 y
128 10. Hyperbolic Geometry

Now we compute our area, using equation (10.2). Integrating in the

y direction, we have
 1  ∞  1
1 1
area(T ) = √ dy = √ dx = π.
x=−1 y= 1−x2 y 2
−1 1 − x2
The last integral is most easily done making the trigonometric sub-
stitution x = sin(t) and dx = cos(t). 

Let T (θ) denote a geodesic triangle having two vertices on the

ideal boundary of H 2 and one interior vertex having interior angle θ.

Lemma 10.11. Theorem 10.9 holds for T (θ).

Proof. Any two such triangles are isometric to each other. We first
match up the interior vertices and then suitably rotate one trian-
gle so that the sides emanating from the common vertex match. In
particular, any incarnation of T (θ) has the same area. Let

f (θ) = π − area(T (θ)).

We want to show that f (θ) = θ for all θ ∈ [0, π). We already know
that f (0) = 0, by the previous result.

A B

C
O

Figure 10.3. Two dissections

10.8. Geodesic Polygons 129

To analyze the general situation, we work in the disk model and

choose T (θ) so that it has an interior vertex O at 0. Figure 10.3 shows
a dissection proof that
f (θ1 + θ2 ) = f (θ1 ) + f (θ2 ),
as long as θ1 + θ2 ≤ π. Just to make the picture clear, we point out
the following:
• The triangle T (θ1 ) has vertices O, A, B.
• The triangle T (θ2 ) has vertices O, B, C.
• The triangle T (θ1 + θ2 ) has vertices O, A, C.
• The triangle with vertices A, B, C is an ideal triangle.
To make this formula work even when θ1 + θ2 = π, we set f (π) = π.
The quadrilateral we have drawn can be dissected in two ways. One
way gives A1 + A2 . The other way gives A + π. Here Ak is the area
of T (θk ) and A is the area of T (θ1 + θ2 ).
Since f (π) = π, we can use our formula inductively to show
f (rπ) = rπ for any rational r ∈ (0, 1). But the function f is pretty
clearly continuous. Since f is the identity on a dense set, f is the
identity everywhere. 

Now we take an arbitrary geodesic triangle and extend the sides

so that they hit the ideal boundary of H 2 . Then we consider the
dissection of the ideal triangle defined by the (ideal) endpoints of
these sides, as shown in Figure 10.4.

Figure 10.4. A dissected ideal triangle

130 10. Hyperbolic Geometry

The ideal triangle and also the three outer triangles are of the
kind we have already considered. Theorem 10.9 holds true for these.
The ideal triangle has area π, and the three outer triangles have areas
α, β, and γ, the three interior angles of the inner triangle. Hence, the
inner triangle has area π − α − β − γ, as desired. This completes the
proof.
A solid geodesic polygon P is convex if it has the following prop-
ery: if p, q ∈ P are two points then the geodesic segment joining p
and q is also contained in P . It is easy to prove, inductively, that any
convex geodesic polygon can be decomposed into geodesic triangles.

Lemma 10.12. The area of a convex geodesic n-gon is (n − 2)π

minus the sum of the interior angles.

Proof. Just decompose into triangles and then apply the triangle
theorem multiple times. 

Exercise 10 (Challenge). Suppose that θ1 , θ2 , θ3 are three numbers

whose sum is less than π. Prove that there is a hyperbolic geodesic
triangle with angles θ1 , θ2 , θ3 .

Exercise 11 (Challenge). Say that a geodesic triangle is δ-thin

if every point in the interior of the (solid version of) triangle is within
δ of a point on the boundary. Note that there is no universal δ so that
all Euclidean triangles are δ-thin. Prove that all hyperbolic geodesic
triangles are 10-thin. (The value δ = 10 is far from optimal.)

10.9. Classification of Isometries

Let T be a real linear fractional transformation. If T (∞) = ∞, then
we have T (z) = az + b. If T (∞) = ∞, then the equation T (z) = z
leads to a quadratic equation az 2 + bz + c, with a, b, c ∈ R. If T is
not the identity, then there are 3 possibilities:
• T fixes one point in H 2 and no other points.
• T fixes no points in H 2 and one point in R ∪ ∞.
• T fixes no points in H 2 and two points in R ∪ ∞.
10.9. Classification of Isometries 131

T is called elliptic, parabolic, or hyperbolic, according to which pos-

sibility occurs. We will discuss these three cases in turn. Before we
start, we mention a helpful construction. Given isometries g and T ,
we call S = gT g −1 a conjugate of T . Note that g maps the fixed
points of T to the fixed points of S.
Suppose T is elliptic. Working in the disk model, we can con-
jugate T so that the result S fixes the origin. In this case, S maps
each geodesic through the origin to another geodesic through the ori-
gin. Moreover, S preserves the distances along these geodesics. From
here, we see that S must be a rotation. So, in the disk model, all the
elliptic isometries are conjugate to ordinary rotations.
Suppose that T is parabolic. Working in the upper half-plane
model, we can conjugate T so that the result S fixes ∞. In this case
S(z) = az + b. If a = 1, then S fixes an additional point in R.
Since this does not happen, a = 1. Hence S(z) = z + b. So, in the
upper half-plane model, all parabolic isometries are conjugate to a
translation.
Suppose that T is hyperbolic. Working in the upper half-plane
model, we can conjugate T so that the result S fixes 0 and ∞. But
then S(z) = rz for some r = 0. So, in the upper half-plane model, all
hyperbolic isometries are conjugate to dilations (or contractions).
Neither the parabolic elements nor the hyperbolic elements have
fixed points in H 2 , but they still behave in a qualitatively different
way. Considering the parabolic map S(z) = z + b, we see that there
is no  > 0 such that S moves all points of H 2 more than . For
example, the hyperbolic distance between iy and S(iy) tends to 0 as
y → ∞. On the other hand, if we examine the map S(z) = rz, we
see that there is some  > 0 such that S moves all points of H 2 by at
least . Indeed,  = | log(r)|.
Chapter 11

Riemannian Metrics on
Surfaces

The purpose of this chapter is explain what is meant by a smooth

surface with a Riemannian metric. The main construction generalizes
what we did for the sphere in §9.1 and also (especially) what we did
for the hyperbolic plane in §10.3. We will give the main definition of
a surface with a Riemannian metric at the end, after assembling all
the preliminary definitions.
A smooth surface with a Riemannian metric is a special case of
a smooth Riemannian manifold . Smooth Riemannian manifolds are
the subject of differential, or Riemannian, geometry. A book such
as [DOC] gives an excellent general account of smooth Riemannian
manifolds.

11.1. Curves in the Plane

A smooth curve in R2 is a smooth map f : (a, b) → R2 . Such a map
is typically given by equations
f (t) = (x(t), y(t))
such that x(t) and y(t) are smooth functions. This is to say that
 n 
dn f d x dn y
= ,
dtn dtn dtn

133
134 11. Riemannian Metrics on Surfaces

exists for all n. We will usually write f  (t) for df /dt.

The function f is regular if f  (t) = 0 for all t ∈ (a, b). As usual,

f (t) is known as the velocity of f at t. Sometimes it is useful to talk
about smooth curves defined on closed intervals. Thus, if we write
f : [a, b] → R2 , we really mean that f is defined on some larger open
interval (a−, b+) and is smooth there. In particular f : [0, 0] → R2
is defined in a neighborhood of 0. This is the usual treatment of the
problem with taking derivatives at the endpoints.

11.2. Riemannian Metrics on the Plane

We defined inner products at the top of §10.4. Let I denote the set
of inner products on R2 . Let U ⊂ R2 be an open set. A Riemannian
metric on U is a smooth map Ψ : U → I. In other words, a Rie-
mannian metric on U is a choice Gp of inner product for each p ∈ U .
This choice gives rise to the functions gij (p), via the formula
(11.1) gij (p) = Gp (ei , ej ).
Here e1 = (1, 0) and e2 = (0, 1). We require that the functions gij
are smooth functions on U . So, you can specify a Riemannian metric
on U by specifying 4 smooth functions gij : U → R subject to the
following constraints:
• g12 (p) = g21 (p) for all p ∈ U .
• For all p ∈ U , the matrix {gij (p)} is positive definite. That
is, the matrix has positive eigenvalues.
A curve in U is just a curve in R2 which happens to lie entirely
in U . We can measure the length of a curve in U relative to the given
Riemannian metric, as follows: Let f : [a, b] → U be a smooth curve.
We define
 b
(11.2) Riemannian length(f ) = Gf (t) (f  (t), f  (t)) dt.
a

The integrand above is called the Riemannian speed of f at t. So,

we are computing the Riemannian length of f by integrating its Rie-
mannian speed. Of course, these quantities depend on the choice of
Riemannian metric. If we choose the standard Riemannian metric,
11.3. Diffeomorphisms and Isometries 135

which is to say the ordinary dot product at each point, then we re-
cover the ordinary notions of speed and length.

Exercise 1: Using the material in the previous chapter, describe

the Riemannian metric on the upper half-plane which gives rise to
the hyperbolic plane.

Exercise 2: Come up with a sensible definition of the Riemann-

ian area of a subset of R2 , assuming that R2 has been equipped with
a Riemannian metric.

Exercise 3. Give an example of a Riemannian metric, defined on all

of R2 , which has the following property. Any two points in R2 can
be joined by a smooth curve whose Riemannian length is less than 1.

Exercise 4. Let G be a Riemannian metric on the plane and let

p, q be two distinct points. Prove that there is some  > 0 such that
any curve connecting p to q has length at least  relative to G. Of
course,  depends on the metric. (Hint: Use the fact that a positive
continuous function on a compact set has a positive infimum.)

11.3. Diffeomorphisms and Isometries

Let U and V be two open subsets of R2 . A diffeomorphism from U
to V is a homeomorphism f : U → V with the following additional
properties:
• f is smooth, that is, all orders of partial derivatives of f
exist.
• For each p ∈ U , the matrix df (p) of first partial derivatives is
nonsingular. That is, df defines a vector space isomorphism
at each point. We abbreviate this by saying that f is regular .
• f −1 is smooth and regular.
Actually, the third condition follows from the other two and the In-
verse Function Theorem.
Note that dfp maps a tangent vector based at p to a tangent
vector based at f (p). Suppose that U and V are given Riemannian
136 11. Riemannian Metrics on Surfaces

metrics. We say that a diffeomorphism f : U → V is a Riemannian

isometry if
Hf (p) (dfp (v), dfp (w)) = Gp (v, w), ∀v, w, p.
Here v and w are vectors and p ∈ U . Also G is the Riemannian metric
defined on U , and H is the Riemannian metric defined on V . We have
already encountered this concept in our second proof of Lemma 10.6.
Here is another point of view on Riemannian isometries. A Rie-
mannian metric on U ⊂ R2 turns U into a metric space in the fol-
lowing way. Given p, q ∈ U we define S(p, q) to be the set of smooth
curves in U which join p to q. We define d(p, q) to be the infimum of
the lengths of curves in S(p, q). This is exactly what we did both for
the sphere and for the hyperbolic plane in the preceding chapters. A
smooth map f : U → V is a Riemannian isometry if and only if it is
a metric isometry relative to the two metric space structures.

Exercise 5. Prove that d really is a metric on U . Prove also that a

Riemannian isometry between U and V gives rise to a metric space
isometry.

Exercise 6 (Challenge). Prove that there is a Riemannian metric

on the plane which makes it isometric to the upper hemisphere of S 2 ,
relative to the round metric. (This part is not so hard.) Now, prove
that there is no Riemannian metric on the plane which makes it iso-
metric to the upper hemisphere of S 2 relative to the chordal metric.
See §9.1 for definitions.

11.4. Atlases and Smooth Surfaces

Recall that a surface is a metric space S such that every point has
a neighborhood which is homeomorphic to R2 . We say that a col-
lection of such neighborhoods is called an atlas. The neighborhoods
themselves are called coordinate charts. So, each element of the atlas
is a pair (U, h), where U is an open subset of Σ and h : U → R2 is
a homeomorphism. We require that the union of all the coordinate
charts in the atlas is the entire surface. In other words, each point in
the surface is contained in at least one coordinate chart.
11.5. Smooth Curves and the Tangent Plane 137

Suppose now that (U1 , h1 ) and (U2 , h2 ) are 2 coordinate charts,

and it happens that V = U1 ∩ U2 is not empty. We define V1 = h1 (V )
and V2 = h2 (V ). Being the intersection of two open sets, V is an open
subset of both U1 and U2 . Since h1 and h2 are homeomorphisms, V1
and V2 are open subsets of R2 . On V1 the map
h12 = h2 ◦ h−1
1

is well defined. We have h12 (V1 ) = V2 . The map

h21 = h1 ◦ h−1
2

is defined on V2 and evidently h21 (V2 ) = V1 . The two maps h12

and h21 are inverses of each other. Also, both maps are continuous,
since they are the composition of continuous maps. In summary h12 :
V1 → V2 is a homeomorphism and h21 : V2 → V1 is the inverse
homeomorphism. These two functions are called overlap functions
because they are defined on the overlaps between coordinate charts.
Our atlas on Σ is said to be a smooth structure if all its overlap
functions are smooth diffeomorphisms. In other words, every time we
can produce an overlap function h12 : V1 → V2 , it turns out to be a
diffeomorphism. We say that a smooth surface is a surface equipped
with a smooth structure.
Here is an annoying technical point. Let (U, h) be a pair such
that U is an open subset of Σ and h : U → R2 is a homeomorphism.
If (U, h) is not part of our atlas, then we can enlarge our atlas by
including (U, h) in it. This will produce possibly some new overlap
functions. If all the new overlap functions are diffeomorphisms, then
we say that (U, h) is compatible with our atlas. We say that our atlas
is maximal if it already contains all compatible coordinate charts. It is
conventional for us to require that our atlases be maximal. However,
this point never actually comes up in practice.

11.5. Smooth Curves and the Tangent Plane

We have already discussed what we mean by a smooth curve in the
plane. Now we will generalize the idea, and speak about smooth
curves on a smooth surface. If we happen to have a smooth surface
embedded in Euclidean space, such as the sphere embedded in R3 ,
138 11. Riemannian Metrics on Surfaces

then it is easy to talk about smooth curves. For instance, we could

say that a smooth curve f : (a, b) → S 2 is smooth if each of the
3 coordinate functions is smooth. When we deal with an abstract
smooth surface, the situation is a bit trickier. We always need to
refer back to the coordinate charts defining the surface.
Let Σ be a smooth surface. Say that a map f : (a, b) → Σ is
smooth at t if there is some  > 0 such that the following holds.

• (t − , t + ) ∈ (a, b).
• f ((t − , t + )) is contained in a coordinate chart (U, h) in
our atlas.
• The curve h ◦ f : (t − , t + ) → R2 is a smooth curve.

The fact that our overlap functions are all diffeomorphisms means
that the notion of smoothness does not depend on which coordinate
chart we use. In other words, if f (t − , t + ) ⊂ U1 ∩ U2 and (U1 , h1 )
and (U2 , h2 ) are both coordinate charts, then

h2 ◦ f = h12 ◦ (h1 ◦ f ).

Since h12 is smooth, the curve h1 ◦ f is smooth if and only if the curve
h2 ◦ f is smooth. Here we are using the fact that the composition of
smooth maps is again smooth. This fact is in turn a consequence of
the chain rule.
We say that f : (a, b) → Σ is smooth if f is smooth at each
t ∈ (a, b). We say that f : [a, b] → Σ is smooth if f is defined and
smooth on a larger interval (a − , b + ).
Let p ∈ Σ be a point. Suppose that

f1 , f2 : [0, 0] → Σ

are two curves such that f1 (0) = f2 (0) = p. We write f1 ∼ f2 if there

is a coordinate chart (U, h) such that p ∈ U and h ◦ f1 and h ◦ f2 have
the same velocity at 0. In other words, (h ◦ f1 ) (0) = (h ◦ f2 ) (0).

Exercise 7. Prove that ∼ is well defined, independent of the co-

ordinate chart we use. Prove also that ∼ is an equivalence relation.
11.6. Riemannian Surfaces 139

We define Tp (Σ) to be the set of equivalence classes of curves

f : [0, 0] → Σ such that f (0) = p. We can make Tp (Σ) into a vector
space as follows. If [f1 ] and [f2 ] are two equivalence classes of curves,
we define [f1 ] + [f2 ] to be the equivalence class of the curve g such
that the velocity of h ◦ g is the velocity of h ◦ f1 plus the velocity of
h ◦ f2 . That is,

(h ◦ g) (0) = (h ◦ f1 ) (0) + (h ◦ f2 ) (0).

Exercise 8. Prove that this notion of addition is well defined. In

other words, if we made this definition relative to two different coordi-
nate charts (U1 , h1 ) and (U2 , h2 ), then we could get the same answer.
(Hint: Use the fact that

h2 ◦ g = h12 ◦ (h1 ◦ g)

(and likewise for f1 and f2 ) and the fact that dh12 is a linear trans-
formation at each point. Now use the chain rule.)

We can also define scaling on Tp (Σ). We define r[f ] to be the

equivalence class of the curve which has r times the velocity of f at 0,
measured in any coordinate chart. Again, this is well defined because
the overlap functions are diffeomorphisms.
All in all, Tp (Σ) is a vector space for each p ∈ Σ.

Exercise 9. Prove that Tp (Σ) is isomorphic to R2 .

11.6. Riemannian Surfaces

Suppose that Σ is a smooth surface. This means that we have a (max-
imal) atlas on Σ whose overlap functions are smooth diffeomorphisms.
Suppose, for each coordinate chart (U, h), we choose a Riemannian
metric on R2 . We say that our choice is consistent if all the overlap
functions are Riemannian isometries relative to the choices. Thus,
the overlap function h12 considered above is a Riemannian isometry
from V1 to V2 , when V1 is equipped with the Riemannian metric as-
sociated to (U1 , h1 ) and V2 is equipped with the Riemannian metric
associated to (U2 , h2 ).
140 11. Riemannian Metrics on Surfaces

A Riemannian metric on Σ is a consistent choice of Riemannian

metrics on R2 , one per coordinate chart. This definition is pretty
abstract, so I will give a second definition at the end of this section
which is perhaps more concrete.
Let f : [a, b] → Σ be a smooth curve. We can define the Rie-
mannian length of f as follows: First of all, we can find a partition
a = t0 < · · · < tn = b such that f ([ti , ti+1 ]) is contained in a coor-
dinate chart (Ui , hi ). Next, we can define Li to be the Riemannian
length of
hi ◦ f ([ti , ti+1 ]).
Finally, we define the length of f to be L0 + · · · + Ln . In other words,
we compute the lengths of a bunch of little pieces of f and then add
them together.
Lemma 11.1. The Riemannian length of f is well defined, indepen-
dent of the choices made in its definition.

Proof. Suppose first of all that we keep the partition the same but
use new coordinate charts (Ui , hi ) such that f ([ti , ti+1 ]) ⊂ Ui . Then,
on [ti , ti+1 ] we have
hi ◦ f = (hi ◦ hi ) ◦ (hi ◦ f ).
But the map hi ◦ hi is an overlap function and is an isometry relative
to the two Riemannian metrics. Thus Li = Li . This shows that the
Riemannian length of f does not change if we use different coordinate
charts from our atlas.
Suppose now that a = s0 < · · · < sm = b is another partition,
and we are using a different sequence {(Ui , hi )} of coordinate charts
to calculate the length. Then by considering all the si and also all
the tj (from our original partition) we can find a refinement a = u0 <
· · · < ul = b which contains all the si and also all the tj . (Basically,
we just take the collection of all the numbers and then resort them.)
We can use the charts (Ui , hi ) to compute the length relative to
the u-partition, and we will get the same answer as if we used the
t-partition. The point here is just that integration is additive:
 ti+1  uk+1  ti+1
= +···+ .
ti t1 uk+h−1
11.6. Riemannian Surfaces 141

Here ti = uk < · · · < uk+h = ti+1 . Likewise, we can use the charts
(Ui , hi ) to compute the length relative to the u-partition, and we will
get the same answer as if we used the s-partition. Thus, we reduce
to the previously considered case where the partition is the same but
the charts change. 

Here is another point of view. The object Tp (Σ) is a 2-dimensional

real vector space for each point p ∈ Σ. We could define a Riemannian
metric on Σ to be a smoothly varying choice of inner product Gp on
the vector space Tp (Σ) for each point p ∈ Σ. We just have to make
sense of the notion of smoothness. If we fix a coordinate chart (U, h),
then a Riemannian metric G on Σ gives rise to a Riemannian met-
ric H on R2 as follows. Suppose we have a point q ∈ R2 and two
vectors v, w. Let p = h−1 (q) ∈ U and [f1 ], [f2 ] ∈ Tp (Σ) be the two
classes so that (h ◦ f1 ) (0) = v and (h ◦ f2 ) (0) = w. Then we define
Hq (v, w) = Gp ([f1 ], [f2 ]). To say that our Riemannian metric on Σ
varies smoothly is to say that H is a smooth Riemannian metric on
R2 for any choice of coordinate chart. This other definition is com-
pletely equivalent to the one I gave above.

Exercise 10. Make up a plausible definition for what a smooth

Riemannian n-manifold ought to be, and develop the theory as far as
we have done here for surfaces.
Chapter 12

Hyperbolic Surfaces

In this chapter we will take up the informal discussion from §1.5. We

will first explain what a hyperbolic surface is, and then we will show
how the gluing construction discussed informally in §1.5 leads to a
hyperbolic surface; see [RAT] for a much more general treatment.
In fact, we will present a general method of constructing hyperbolic
surfaces out of convex geodesic hyperbolic polygons. At the end, we
will prove that every complete hyperbolic surface is covered by the
hyperbolic plane.

12.1. Definition
We will give two definitions of a hyperbolic surface. The first defini-
tion requires the material in the last chapter while the second defini-
tion does not.

Definition 12.1. A hyperbolic surface is a smooth surface with a

Riemannian metric, such that each point on the surface has a neigh-
borhood that is isometric to an open disk in the hyperbolic plane.

Our second definition is more elementary and does not require the
material on Riemannian manifolds discussed in the previous chapter.

143
144 12. Hyperbolic Surfaces

On the other hand, the second definition requires a few preliminar-

ies of its own. Let U and V be two open subsets of H 2 . Say that
a disk-like set is a subset of the plane that is homeomorphic to an
open disk. Say that a map f : U → V is a local hyperbolic isometry
if the restriction of f to each open component of U agrees with the
restriction of a hyperbolic isometry. The easiest case to think about
is when U and V are both connected. Then f : U → V is a local
isometry iff f is the restriction of a hyperbolic isometry to U .

Definition 12.2. A hyperbolic structure on Σ is an atlas of co-

ordinate charts on Σ such that the following holds:
• The image of every coordinate chart is a disk-like subset of
H 2.
• The overlap functions are local hyperbolic isometries.
• The atlas is maximal.
Now we reconcile the two definitions. Suppose that Σ is a hyper-
bolic surface according to Definition 12.1. Then the local isometries
mentioned in Definition 12.1 give rise to an atlas of coordinate charts
in which the overlap functions are local isometries. This atlas is not
maximal, but then we can complete it to a maximal atlas using Zorn’s
lemma. (See any book on set theory, such as [DEV], for a discussion
of Zorn’s lemma.) In this way, we see that Σ is a hyperbolic surface
according to Definition 12.2.

Exercise 1. Prove that a local hyperbolic isometry is a smooth

map. This amounts to showing that a linear fractional is infinitely
differentiable.

Suppose that Σ is a hyperbolic surface according to Definition 12.2.

According to Exercise 1, the system of coordinate charts on Σ has
smooth overlap functions. Therefore, Σ is a smooth surface. We can
define a Riemannian metric on Σ as follows. Let p ∈ Σ be a point.
Let (U, f ) be a coordinate chart about p. This means that U is an
open neighborhood of p and f : U → H 2 is a homeomorphism onto
a disk-like set. Let V, W ∈ Tp (Σ) be two tangent vectors. This is to
say V = [α] and W = [β] where α, β : (−, ) → Σ are smooth curves
12.2. Gluing Recipes 145

with α(0) = β(0) = p. We define

Hp (V, W ) = Gf (p) ((f ◦ α) (0), (f ◦ β) (0)).

Here G is the Riemannian metric on the hyperbolic plane. In other

words, we have just used the coordinate chart to transfer the metric
on H 2 to the tangent space Tp Σ of Σ at p. The fact that the overlap
functions are all hyperbolic isometries implies that the above defini-
tion of the metric is independent of which coordinate chart is used.
This puts a Riemannian metric on Σ with the desired properties.
Equipped with this metric, Σ satisfies Definition 12.1.
Now we know that the two definitions pick out the same objects
as hyperbolic surfaces.

12.2. Gluing Recipes

We would like a way to systematically build hyperbolic surfaces. Re-
call from §10.8 that a convex geodesic polygon is a convex subset of
H 2 whose boundary consists of a simple closed path of geodesic seg-
ments. The idea is to glue together a bunch of geodesic polygons,
taking care to get the angle sums correct.
Let P be a geodesic polygon. Let e ∈ P be an edge. Say a
decoration of e is a labelling of e by both a number and an arrow. Say
that a decoration of P is a decoration of every edge of P . Whenever
we have built surfaces by gluing the sides of a polygon together, we
have always based the construction on a decoration of the polygon.
We say that a gluing recipe for a hyperbolic surface is a finite list
P1 , . . . , Pn of decorated polygons. There are some conditions we want
to force:

• If some number appears as a label, then it appears as the

label for exactly two edges. This condition guarantees that
we will glue the edges together in pairs.
• If two edges have the same numerical label, then they have
the same hyperbolic length. This allows us to make our
gluing using (the restriction of) a hyperbolic isometry.
146 12. Hyperbolic Surfaces

• Any complete circuit of angles adds up to 2π. This condi-

tion guarantees that a neighborhood of each vertex is locally
isometric to H 2 .

4
1
1

3 2

4 4 3

Figure 12.1. A complete circuit

The third condition requires some explanation. A complete circuit

is a collection of edges
e1 , e1 , e2 , e2 , e3 , e3 , . . . , ek , e1 .
with the property, for all j, that ej and ej have the same numerical
label and ej and ej+1 are consecutive edges of the same polygon.
(Here we are taking the indices cyclically, so that k + 1 is set equal
to 1.) Figure 12.1 shows what we have in mind.
There is one subtle condition that we need also to require. Let
vj be the vertex incident to ej and ej+1 . Then the arrow along ej+1
points to vj iff the arrow along ej+1 points to vj+1
. Figure 12.1
depicts a situation where this holds. The point here is that we want
the edges in our chain to emanate from a single vertex in the quotient
space. The edges ej and ej+1 subtend an angle αj and we want
α1 + · · · + αk = 2π.
12.3. Gluing Recipes Lead to Surfaces 147

12.3. Gluing Recipes Lead to Surfaces

Theorem 12.1. Any gluing recipe gives rise to a hyperbolic surface.

Proof. Given a gluing recipe, we can form a surface Σ as follows.

First of all, we start out with the metric space X which is the disjoint
union of P1 , . . . , Pn . We can do this by declaring d(p, q) = 1 if p ∈ Pi
and q ∈ Pj with j = i. For p, q ∈ Pj (the same polygon) we just
use the hyperbolic metric. So, you should picture X approximately
as a stack of polygons hovering in the air, as on the left-hand side of
Figure 12.1.
Now we define an equivalence relation on X using the rule that
p ∼ p iff p and p are corresponding points on like-numbered edges.
Here corresponding should be pretty obvious. Suppose e and e are
two like-numbered edges, both having length λ. Then there is some t
such that p is t units along e measured in the direction of the arrow.
Likewise there is some t such that p is t units along e . Then p and
p are corresponding points iff t = t .
The nontrivial equivalence classes typically have 2 members, with
1 member coming from each edge. However, for the vertices of the
polygons, each of which belongs to two edges, the corresponding
equivalence class might be larger. In Figure 12.1, the equivalence
class of the relevant vertex has 4 elements.
The surface is defined as Σ = X/ ∼. We would like to show that
Σ is indeed a surface, so we have to construct an atlas of coordinate
charts. Suppose that x is an interior point of some polygon P . Then
some open neighborhood Ux of x remains in the interior of P . No
point in Ux is equivalent to any other point of Σ. The inclusion map
Ux → P ⊂ H 2 gives a coordinate chart from Ux to H 2 . We take Ux
to be a metric disk.
Suppose now that p ∈ Σ is an equivalence class consisting of two
points, in the interiors of a pair of edges, that are glued together
when the edges are paired. That is, p = {q, q  }, with q ∈ e and
q  ∈ e , where e and e are open edges. Let P and P  be the polygons
containing e and e , respectively. Let U and U  be small half-disk
neighborhoods of q and q  in P and P  , respectively, as shown in
Figure 12.2.
148 12. Hyperbolic Surfaces

Figure 12.2. Half-disk neighborhoods

We define h : U ∪ U  → H 2 so that the following hold.

• The map h, when restricted to either U or U  , is the inclusion
map composed with a hyperbolic isometry.
• h(e ∩ U ) = h (e ∩ U  ) and the arrows go the right way.
• h(U ) and h(U  ) lie on opposite sides of h(e) = h(e ).
This is pretty obvious. We first define h as the inclusion map on both
halves, and then we compose one half of the map with a suitable
isometry to adjust things. The main point here is that U ∩ e and
U  ∩ e are open geodesic segments of the same length.

Exercise 2. Prove that Δ = (U ∪ U  )/ ∼ is homeomorphic to an

open disk. More precisely, prove that h defines a homeomorphism Δ
to a disk in H 2 . Finally, prove that Δ is an open neighborhood of p
in Σ.

Finally, suppose that p is the equivalence class coming from some

vertices of our polygons. Then we have one of the circuits mentioned
above. Let {q1 , . . . , qk } be the equivalence class of p. In the example
shown in Figure 12.1, we have k = 4. Let Pj be the polygon that has
qj as a vertex. In each Pj we choose a little wedge-shaped neighbor-
hood consisting of all points of Pj within  of qj .
12.4. Some Examples 149

Exercise 3. Prove that the union (U1 ∪ · · · ∪ Uk )/ ∼ is homeo-

morphic to a disk.

We define a map h : U1 ∪ · · · ∪ Uk → H 2 in such a way that the

following holds.
• The map h, when restricted to any Uj , is the inclusion map
composed with a hyperbolic isometry.
• h respects the gluing of edges.
Expressing the last condition is a bit clumsy, but I hope that you
can see what it means. If two edges are glued together, then h sends
them (or at least the portions inside our little pizza slices) to the same
segment in H 2 .

Exercise 4. Prove that (U1 ∪ . . . ∪ Uk )/ ∼ is an open neighbor-

hood of p in Σ and that h gives a homeomorphism from this set onto
an open disk in H 2 . (Hint: The circuit condition guarantees that the
images of h fit together to make a single hyperbolic disk.)

From the way we have defined things, the overlap functions are
all local hyperbolic isometries, so we have found an atlas on Σ whose
overlap functions are local hyperbolic isometries. We can complete
this to a maximal atlas, if we like, using Zorn’s lemma. 

12.4. Some Examples

Here are some additional examples for you to work out. The first ex-
ercise asks you to work out the discussion in §1.5. The next example
points to more flexible and systematic approach.

Exercise 5. Prove that there is a regular convex 4n-gon, with angles

π/2n, provided that n ≥ 2. Call this polygon P4n . Decorate P4n by
giving the opposite sides and making the arrows point in the same
direction. See Figure 1.7. Prove that P4n , as decorated, is a gluing
diagram for a hyperbolic surface.
150 12. Hyperbolic Surfaces

Exercise 6. Prove that there exists a right angled regular hexagon.

Construct a decoration of 4n such hexagons in such a way that it is
the gluing diagram for a hyperbolic surface.

Exercise 7 (Challenge). If you take n = 2 in Exercises 5 and

6 you get homeomorphic surfaces. Prove that they are not isometric.

Exercise 8 (Challenge). Prove that there are uncountably many

surfaces, all homeomorphic to the octagon surface from Exercise 5,
no two of which are isometric to each other.

12.5. Geodesic Triangulations

So far, we have shown how to build some hyperbolic surfaces from
gluing diagrams. In this section we will show that every compact
hyperbolic surface arises from this construction. We begin with a
well-known construction in H 2 .
Let X ⊂ H 2 be a finite collection of points. For each p ∈ X, we
let Np be the set of points that are closer to p than to any point of
X.

Lemma 12.2. Np is convex. If Np is bounded, then Np is the interior

of a convex geodesic hyperbolic polygon.

Proof. Say that a geodesic half-plane is a set of points in H 2 lying

to one side of a hyperbolic geodesic. Geodesic half-planes are convex.
Given any two points p, q ∈ H 2 , the set of points closer to p is a
geodesic half-plane. For this reason, Np is the intersection of finitely
many geodesic half-planes, and the boundary of Np is contained in
a finite union of geodesics. Since the intersection of convex sets is
convex, Np is convex. In case Np is bounded, the boundary evidently
is a convex geodesic polygon. 

Say that a geodesic triangulation of a hyperbolic surface is a de-

composition of the surface as the finite union of geodesic triangles.
Every pair of triangles should either be disjoint or share an edge or
share a vertex. If a hyperbolic surface has a geodesic triangulation,
12.5. Geodesic Triangulations 151

then we can cut the surface open along the triangles and thereby
obtain a description of the surface in terms of a gluing diagram.

Lemma 12.3. Every compact hyperbolic surface has a geodesic tri-

angulation.

Proof. Let S be the surface. By compactness, there is some d ∈ (0, 1)

such that every disk of radius d on the surface is isometric to a disk
of radius d in H 2 . Place a finite number of points on S in such a
way that every disk of radius D/K contains at least one point. The
constant K is yet to be determined. Let X denote this finite set of
points.
Given p ∈ X, let Bd (p) denote the disk of radius d about p. Let
Np ⊂ S be the set of points in S that are closer to p than to any other
point in X. We claim that each Np is isometric to the interior of a
convex geodesic hyperbolic polygon provided that K is large enough.
(This is not an immediate consequence of the previous result because
we are working on a surface and not directly in H 2 .) The boundary
of Np consists of points q such that q is equidistant between p and
some other point p of X. Let Xp denote the set of points p ∈ X such
that some point of Np is equidistant from p and p . We can choose K
large enough so that Np ⊂ Bd (p) and Xp consists entirely of points
in the Bd (p). Now we apply the previous result. This shows that Np
is the interior of a convex geodesic polygon.
We have partitioned S into convex geodesic polygons. To finish
the triangulation, we just add in extra geodesic segments, as needed,
to divide each of the convex polygons into triangles. 

Lemma 12.3 allows us to prove the Gauss–Bonnet Theorem for

hyperbolic surfaces.

Theorem 12.4 (Gauss–Bonnet). The hyperbolic area of a compact

hyperbolic surface S is −2πχ(S), where χ(S) is the Euler character-
istic of S. In particular, the area only depends on the Euler charac-
teristic.
152 12. Hyperbolic Surfaces

Proof. We give S a geodesic triangulation. From §3.4, we have the

formula

(12.1) χ(S) = F − E + V,

where F is the number of faces in the triangulation, E is the number

of edges, and V is the number of vertices.
Each triangle in the triangulation has 3 edges, and each edge
belongs to two triangles. For this reason, E = 3F/2. At the same
time, the total sum of all the interior angles of all the triangles is
2πV , because the sum of these angles around any one vertex is 2π.
Putting these equations together, we get
F F 1 
(12.2) χ(S) = − + V = − + θi .
2 2 2π
angles

For each triangle τ , let θi (τ ), for i = 1, 2, 3, be the three interior

angles of τ . Hence
  
−2πχ(S) = π F − θi
angles
  
= π − θ1 (τ ) − θ2 (τ ) − θ3 (τ )
(12.3) triangles

∗
= area(τ )
triangles

= area(S).

The starred equality comes from Theorem 10.9. 

Theorem 12.4 is a special case of the Gauss-Bonnet Theorem from

differential geometry. See [BAL] for a discussion of the proof of this
general result.

12.6. Riemannian Covers

We say that a Riemannian cover of a Riemannian manifold X is a
Riemannian manifold X such that the covering map E : X → X is
a local isometry. We mean that the differential dE is an isometry
12.6. Riemannian Covers 153

on each tangent plane, measured with respect to the two Riemannian

metrics.
 is
Lemma 12.5. Suppose that X is a Riemannian manifold and X

a covering space of X. Then one can make X into a Riemannian
 → X is a
manifold in such a way that the covering map E : X
Riemannian cover.

 inherits the structure of a manifold. We have

Proof. First of all, X
the covering map E : X  → X. Each point x  ∈ X lies in a small
 
open neighborhood U such that U = U is an evenly covered neigh-
borhood of x = E( x) and also (U, φ) is a coordinate chart for x. The
composition φ ◦ E : U  → Rn gives a coordinate chart for an open
neighborhood of x. The overlap functions for these coordinate charts
on X are the same as for the coordinate charts on X. Hence X is a
smooth manifold and E is a smooth map.
There exists a unique Riemannian metric on X  so that E : X →
X is an isometry. We define the metric g such that
gx(X, Y ) = gx (dE(X), dE(Y )).
Here dE is the differential of E. Here X and Y are tangent vectors to
X at x
. When measured in the local coordinates we have described,
the differential dE is just the identity map. So, the metric g is actually
an inner product.
There is a second way to think about the Riemannian metric
on X  which perhaps is more clear. The Riemannian metric on X
is just a collection of Riemannian metrics on various open sets of
Rn that are compatible in the sense that all overlap functions are
isometries. We may, first of all, restrict our attention to open sets
in X that are evenly covered by the covering map. We can then use
the preimages of these open sets as coordinate charts in X. Since

the overlap functions for the charts on X are the same as on X, the
same collection of compatible metrics defines a Riemannian metric on

X. 

Exercise 9. Show that a Riemannian covering map E : X  → X

is distance nonincreasing. Also, give an example of a Riemannian
covering from a connected space X to a connected space X that is
154 12. Hyperbolic Surfaces

not a global isometry. That is, give an example where there are
points x, y ∈ x
 that are farther apart than their corresponding im-
ages x, y ∈ X.

Recall that a metric space is complete if every Cauchy sequence in

the space converges. For a Riemannian manifold, there is a different
notion, called geodesic completness, which people often mean when
they say that a Riemannian manifold is complete. However, the two
definitions are the same, thanks to the Hopf–Rinow Theorem. See
[DOC] for a proof. We mention this just to keep consistent with
other texts. We only care about the metric completeness.

Lemma 12.6. Let E : X  → X be a Riemannian covering space. If


X is complete, then so is X.

Proof. Let { xn } be a Cauchy sequence in X.  We have constructed


things in such a way that the map E : X → X is distance nonin-
creasing. Setting xn = E( xn ), we now know that {xn } is a Cauchy
sequence in X. Since X is complete, there is some limit point x∗ .
There is an evenly covered neighborhood U of x∗ which contains xn
for n large. But then all the points x n lie in the same component of
E −1 (U ) for n large. But E : U → U is a homeomorphism. In par-
ticular, E maps convergent sequences to convergent sequences and so
does E −1 . Since {xn } is a convergent sequence in U , the sequence
{xn } is a convergent sequence in U. 

12.7. Hadamard’s Theorem

In this section we prove Hadamard’s Theorem, in two dimensions. See
[DOC] for a proof in general. The version of Hadamard’s Theorem
we prove is a technical step in our proof that any complete hyperbolic
surface is covered by H 2 . Just for this section, let H = H 2 stand
for the hyperbolic plane.

Theorem 12.7 (Hadamard). Let H be a complete and simply con-

nected surface that is locally isometric to H 2 . Then H is globally
isometric to H 2 .
12.7. Hadamard’s Theorem 155

A surface is oriented if we can make a continuous choice of basis

for each tangent plane. Any simply connected surface is oriented.
Let h ∈ H be a point and let h ∈ H be a point. Both points have
neighborhoods which are isometric to disks in the hyperbolic plane.
Thus we can find an isometry I between a neighborhood U ⊂ H of
h and a neigborhood U ⊂ H of h. Let x ∈ H be any point. We can
take I to be orientation preserving.
Let γ be a continuous path connecting h to x.

Lemma 12.8. I can be extended to a neighborhood of γ in such a

way that I is a local isometry at every point along γ.

Proof. We think of γ as a map from [0, 1] to H, with γ(0) = h and

γ(1) = x. Say that a point t ∈ [0, 1] is good if this lemma holds for
the restriction of γ to the interval [0, t]. Note that 0 is good. Note
also that if t is good, then so is s ∈ [0, t]. Hence the set J of good
points is an interval that contains 0. Moreover, since local isometries
are defined on open sets, J is an open interval.
We claim that J is a closed interval. Suppose that all points
t ∈ [0, s) are good. We take a sequence of points {sn } ∈ [0, s) such
that sn → t. Then {γ(sn )} is a Cauchy sequence. Since I is not
distance increasing, {I(γ(sn )} is also a Cauchy sequence. Since H is
complete, this Cauchy sequence converges. We define

I(t) = lim I(γ(sn )).

We would like to see that in fact I is defined and a local isometry in

a neighborhood of γ(t).
There is a local isometry I  carrying a neighborhood U of γ(t) to
a disk in H 2 . Since every two points have isometric neighborhoods,
we can assume that I  and I agree on γ(t). Once n is large, we have
γ(sn ) ∈ U . The points I(γ(sn )) and I  (γ(sn )) are the same distance
from I(γ(t)). So, we may adjust I  by a rotation so that I and I 
agree on some γ(sn ). But then I and I  agree on all of γ(sn , t]. The
point is that two orientation-preserving isometries agree everywhere
provided that they agree on two points. This shows that the union
map I ∪ I  is a local isometry at all points of γ[0, t].
156 12. Hyperbolic Surfaces

Our argument shows that t is good, and therefore that J is a

closed interval. Since J is open, closed, and connected, we must have
that J = [0, 1]. 

Now we have a candidate map I : H → H. However, we need

to see that this map is well defined. That is, we need to see that the
point I(x) is independent of the choice of path γ joining h to x. This
is where we use the simple connectivity assumption.
Let γ0 and γ1 be two paths joining h to x. We think of γ0 and γ1
both as maps from [0, 1] into H, with γ0 (0) = γ1 (0) = h and γ0 (1) =
γ1 (1) = x. Since H is simply connected, there is a path homotopy
γt from γ0 to γ1 . The point xt = I(γt (1)) varies continuously with t.
On the other hand, note that the same extension in the above lemma
works for both γs and γt as long as s and t are close together. Hence
xs = xt for s and t close. But this shows that xt does not move at
all.
Our extension gives a local isometry I : H → H. But the exis-
tence of our extension just used the following.
• Completeness of H.
• Local homogeneity of H, in connection with the map I 
above.
• Path connectivity and simple connectivity of H.
All these properties hold with the two spaces reversed. Reversing
the roles of H and H, we construct the inverse map I −1 using the
same method. Hence both I and I −1 are homeomorphisms and local
isometries. Bring local isometries, both maps I and I −1 are globally
distance nonincreasing. This is only possible if both these maps are
global isometries.

12.8. The Hyperbolic Cover

We are almost done with the proof that every complete hyperbolic
surface is covered by the hyperbolic plane. We just need one more
technical result.
Lemma 12.9. A complete hyperbolic surface is good in the sense of
Chapter 7.
12.8. The Hyperbolic Cover 157

Proof. Let X be a complete hyperbolic surface. A sufficiently small

ball about any x ∈ X is isometric to a hyperbolic disk. Such sets
are obviously both conical and simply connected. Indeed, we can join
each point y ∈ B (x) to x by a geodesic. We just need to see that
any path in X is good.
Consider a continuous path f0 : [0, 1] → X. Every point x ∈
f0 [0, 1] has a neighborhood Ux that is isometric to a hyperbolic disk.
By compactness, there is a single positive constant, say 2 that works
for all points of f0 [0, 1]. Let f1 : [0, 1] → X be a path such that
D(f0 , f1 ) < . This means that distance between f0 (t) and f1 (t)
is less than . For each t ∈ [0, 1] there is a geodesic gt [0, 1] → X
connecting f0 (t) to f1 (t) that remains within the -ball about f0 (t).
For s sufficiently near t, the two paths γs and γt lie in the 2 ball
about γ0 (t). Therefore, the path γt varies continuously with t. But
then the map F (s, t) = γs (t) gives a homotopy from f0 = F (0, ∗) to
f1 = F (1, ∗). 

Theorem 12.10. A complete hyperbolic surface is universally cov-

ered by H 2 .

Proof. Let X be a complete hyperbolic surface. We know that X

is a good metric space in the sense of Chapter 7. By Theorem 7.1,
there exists a simply connected covering space X and a covering map
 
E : X → X. The space X is complete by Lemma 12.6. But then, by
Hadamard’s Theorem, X  is isometric to H 2 . 

What I (and many people) find really great about this result is
that it opens the door to beautiful tilings of the hyperbolic plane.
These are the kinds of tilings drawn by M. C. Escher in his Circle
Woodcut series. Here we will sketch the idea behind these tilings. We
begin with a general exercise that justifies the construction we give
below.

Exercise 10. Let X  → X be a Riemannian covering of a complete

Riemannian manifold X. Let U be a simply connected open subset of
X. Let U  and that
 = E −1 (U ). Prove that U is evenly covered by U
the restriction of E to any component of U  is an isometry between
158 12. Hyperbolic Surfaces

that component and U . (Hint: Imitate the proof of Hadamard’s The-

orem to construct an inverse map that is also a local isometry.)

Now consider a description of a hyperbolic surface as one obtained

by gluing together the sides of a hyperbolic polygon. For instance, if
we glue together 4 regular right angled hexagons in a suitable pattern,
we get a hyperbolic surface of Euler characteristic −2; see §12.4. Let
X be a hyperbolic surface obtained by this construction. The interiors
of the right angled hexagons are embedded and simply connected in
X. We can consider the preimages of these open hexagons in H 2 by
pulling them back by the map E. By Exercise 10, the result is an
infinite collection of open right angled hexagons H 2 .
At the same time, X contains a graph whose edges are embedded
geodesic arcs. These arcs are the images of the edges of the hexagons
under the gluing maps. The preimages of these arcs in H 2 are the
interfaces between the open hexagons. The whole picture fits together
to give a tiling of H 2 by right angled hexagons. Being right angled,
these hexagons necessarily meet 4 per vertex. This is a hyperbolic
geometry analogue of the picture we developed in §6.3.
In §6.3, we actually went the other way around. We started with
the tiling and then produced the covering map. The situation here
is so concrete that we can actually do the same thing. We take an
infinite supply of regular right angled hyperbolic hexagons and glue
them together so that they meet 4 per vertex. The same argument as
the one given in Chapter 12 shows that the result is locally isometric
to the hyperbolic plane. With a bit of effort, one can see that the
resulting space is both simply connected and complete, and hence
globally isometric to the hyperbolic plane. Once we have built this
tiling of H 2 by hexagons, we can imitate the construction in §6.3,
directly producing the covering map from H 2 to the surface.
Given that we can construct the universal cover E : H 2 → X
directly in this case, without resorting to Theorem 12.10, you might
wonder why we need this result at all. I suppose that the best an-
swer to this question is that Theorem 12.10 is completely general.
12.8. The Hyperbolic Cover 159

We do not have to fool around with the combinatorics of gluing to-

gether infinite families of polygons every time we want to construct
the universal cover of a hyperbolic surface.
Part 3

Surfaces and Complex

Analysis
Chapter 13

A Primer on Complex
Analysis

The purpose of this chapter is to present some of the foundational

results in complex analysis. I have tried to write this chapter in such
a way that someone who knows no complex analysis could follow
along. However, the development here is rather rapid and terse. The
ideal reader is a person who has already taken a semester of complex
analysis, but who perhaps does not remember the proofs of the main
results. This chapter collects the basic results in one place. All the
material here can be found in any book on complex analysis; see, e.g.,
[AHL].

13.1. Basic Definitions

Throughout the chapter U will denote an open subset of C, the com-
plex plane. Let f : U → C be a continuous map. We say that f has
a complex derivative at z ∈ U if the quotient
f (z + h) − f (z)
f  (z) = lim
h→0 h
exists and is finite. Note that h is allowed to be a complex number.
f is said to be complex analytic in U if f  (z) exists for all z ∈ U and
the function z → f  (z) varies continuously in U . Complex analytic

163
164 13. A Primer on Complex Analysis

functions are sometimes called holomorphic functions. The two terms

are synonyms.
Complex analysis is mainly the study of complex analytic func-
tions. In this chapter we will discuss complex analytic functions from
3 points of view:

• A complex analytic function is a function that has a complex

derivative at each point, as we have just discussed.

• A complex analytic function is a function which satisfies the

Cauchy Integral Formula.

• A complex analytic function is a function which agrees with

its Taylor series in a neighborhood of each point.

Each of these concepts brings out a different characteristic of a com-

plex analytic function. A major part of an undergraduate complex
analysis course is explaining why these three definitions are the same.
Among other things, we will establish the equivalence of the 3 defini-
tions in this chapter.
Here is an overview of this chapter. The next several sections lead
up to the Cauchy Integral Formula. Once we establish the Cauchy
Integral Formula, we will prove a number of results about complex
analytic functions. We will consider the connection to power series
at the end.

Exercise 1. Suppose that f and g are complex analytic in U and

g is never 0 in U . Prove that the functions f + g and f − g and f g
and f /g are all complex analytic in U . Conclude that any function
P (z)/Q(z), where P and Q are polynomials, is complex analytic away
from the roots of Q.

Exercise 2. Suppose that f is complex analytic on U and g is com-

plex analytic on V and f (U ) ⊂ V . Prove g ◦ f is complex analytic
and the complex derivative satisfies (g ◦ f ) (z) = g  (f (z))f  (z). This
is the chain rule.
13.2. Cauchy’s Theorem 165

Being a complex analytic map is rather special. For instance, the

function f (z) = z 2 + 3z is not complex analytic in C. So, not all
smooth maps are complex analytic.
We can think of a complex analytic function f as a map from R2
to R2 by writing
f (x + iy) = u(x + iy) + iv(x + iy).
Recall that f is differentiable at the point (x, y) if the matrix of partial
derivatives  
u uy
df = x
vx vy
exists at p = (x, y) and
f (p + tv) − f (v)
lim = df |p (v).
t→0 t
Here t ∈ R. To say that f has a complex derivative at z = x + iy is
the same as saying that f is differentiable and df |p is the composition
of a rotation and a dilation. That is
   
u x uy r cos(θ) r sin(θ)
= , r ∈ R, θ ∈ [0, 2π).
vx vy −r sin(θ) r cos(θ)
Equating terms, we get
ux = vy , uv = −vx .
These are called the Cauchy–Riemann equations. Thus, if f is com-
plex analytic, then its first partials vary continuously and satisfy the
Cauchy–Riemann equations.
The converse is also true: f is complex analytic provided that df
exists, is continuous, and satisfies the Cauchy–Riemann equations.

13.2. Cauchy’s Theorem

Suppose γ is a smooth oriented arc in C and f is a complex val-
ued function defined in a neighborhood of γ. We define a complex
line integral along γ as follows. Letting g : [a, b] → γ be a smooth
parametrization of γ that respects the orientation of γ, we define
  b
dg
f dz = f (g(t)) dt.
γ a dt
166 13. A Primer on Complex Analysis

The same argument as in §8.6 shows that the answer only depends
on γ and not the parametrization. Also, were we to switch the orien-
tation, the value of the line integral would switch signs.

Exercise 3. Let λ be a counterclockwise

 oriented circle centered
at 0, and let f (z) = 1/z. Prove that λ f dz = 2πi.

If we have a finite union γ = {γj } of smooth oriented arcs, we

define  
f dz = f.
γ j γj

In particular, we want to consider the case when γ is a circular poly-

gon. A circular polygon is an embedded loop made by concatenating
line segments and arcs of circles; see Figure 13.1 for an example.

Theorem 13.1 (Cauchy). Let γ be a circular polygon. Suppose that

f is complex
 analytic in a neighborhood of the domain bounded by γ.
Then γ f dz = 0.

Proof. Let f = u + iv. Letting dx and dy be the usual line elements,

we can write
   
f dz = (u+iv)(dx+idy) = (udx−vdy)+i (vdx+udy).
∂D ∂D ∂D D

By Green’s theorem, the integral on the right-hand side equals

 
(uy + vx )dxdy + i (ux − vy )dxdy.
D D

Both pieces vanish, due to the Cauchy–Riemann equations. 

Remark. In §8.7 we proved Green’s Theorem for polygons. The case

of circular polygons follows from the polygon case and a straightfor-
ward limiting argument. Alternatively, most books on multivariable
calculus have a proof of Green’s Theorem in great generality; see, e.g.,
[SPI]. Cauchy’s Theorem holds in the same generality that Green’s
Theorem holds, but the version we state is sufficient for all the appli-
cations we give.
13.3. The Cauchy Integral Formula 167

13.3. The Cauchy Integral Formula

Here is the beautiful Cauchy Integral Formula.

Theorem 13.2 (Cauchy Integral Formula). Let γ be a circular poly-

gon, oriented counterclockwise around the domain D that it bounds.
Let a ∈ D − γ. Suppose that f is complex analytic in a neighborhood
U of D. Then

1 f (z)
(13.1) f (a) = dz.
2πi γ z − a

Proof. We translate the whole picture and consider without loss of

generality the case when a = 0. The function g(z) = f (z)/z is com-
plex analytic in U − {0}. Let β be the circular polygon shown in
Figure 13.1.

0
λ

Figure 13.1.

We have

(13.2) g dz = 0
β

by Cauchy’s Theorem. We allow the two oppositely oriented vertical

segments in β to approach each other. In the limit, the contributions
168 13. A Primer on Complex Analysis

from the two vertical segments cancel out, and equation (13.2) yields
 
(13.3) g(z) = g(z).
γ λ

Here λ is a counterclockwise circle centered at 0. Define

 
 
(13.4) I =  g(z)dz − 2πif (0).
γ

We want to show that I = 0. Combining Exercise 4 and equation

(13.3), we have
     
 dz   f (z) f (0) 
(13.5) 
I =  g(z)dz − f (0) dz −
= dz .
γ λ z λ z λ z

Now we have a bound on I that is expressed entirely in terms of λ.

Rearranging the terms of the last integral, we have
 
 f (z) − f (0) 
(13.6) 
I= dz  ≤ length(λ) × 2|f  (0)|.
λ z

The last inequality holds once λ is sufficiently small. Letting λ shrink

to a point, we see that I = 0, as desied. 

13.4. Differentiability
Here we use the Cauchy Integral Formula to prove some results about
the differentiability of complex analytic functions. Our first result is
not so important in itself, but it illustrates how one uses the Cauchy
Integral Formula to get a formula for the derivative of a complex
analytic function.

Theorem 13.3. Suppose that f is a complex valued and continuously

differentiable function defined in an open set U . If f satisfies the
Cauchy Integral Formula with respect to every circle in U , then f is
complex analytic in U .
13.4. Differentiability 169

Proof. Let a ∈ U and let γ ⊂ U be a circle surrounding a. Using

the Cauchy Integral Formula, we compute
f (a + h) − f (a)
lim
h→0 h
  
1 f (z) f (z)
= lim dz − dz =
h→0 2πih γ z−a−h γ z−a
(13.7) 
1 f (z)
= lim dz =
h→0 2πi γ (z − a)(z − a − h)

1 f (z)
= dz.
2πi γ (z − a)2
This tells us that f has a complex derivative at a and also gives a
formula for it. 

Theorem 13.4. Suppose that f is a complex analytic function defined

in an open set U . Then f  is also complex analytic in U .

Proof. Note that f  exists just by virtue of the fact that f is complex
analytic. Since f is complex analytic in U , Theorem 13.3 holds for f .
Equation (13.7) gives us a formula for f  . We compute
f  (a + h) − f  (a)
lim
h→0 h
  
1 f (z) f (z)
(13.8) = lim dz − dz =
γ (z − a − h) γ (z − a)
h→0 2πih 2 2

2 f (z)
= dz.
2πi γ (z − a)3
Here γ is some circle that surrounds a. Hence f  has a complex
derivative throughout U and equation (13.8) gives a formula for it.
In light of equation (13.8), the function f  is continuous. Hence f  is
complex analytic in U . 

An immediate corollary is that complex analytic functions are

infinitely differentiable. The calculation in equation (13.8), when done
inductively, yields the following formula for the nth derivative of a
complex analytic function f .

(n) n! f (z)
(13.9) f (a) = dz.
2πi γ (z − a)n+1
170 13. A Primer on Complex Analysis

13.5. The Maximum Principle

Let f be a complex analytic function in a connected open set U . Here
we will show that f cannot take on its maximum value at a point in
U unless f is constant. We will assume that f takes on a maximum
at some point a ∈ U , and we will derive a contradiction. If f has an
interior maximum, we can compose f with translations and dilations
and arrange the following properties.
• |f (0)| = 1.
• U contains the unit disk.
• |F (z)| ≤ 1 for all |z| = 1.
• |F (z)| < 1 for some z such that |z| = 1.
Let γ be the unit circle. By the Cauchy Integral Formula we have
  
1  f (z)  ∗ 1
1 = |f (0)| = ≤ |f (z)|dz < 1.
2π  γ z  2π γ

This is a contradiction. The starred inequality is essentially the tri-

angle inequality. For later purposes we work out some consequences
of the Maximum Principle.

Lemma 13.5. Suppose that f (z)/z n is well defined at 0 and complex

analytic in a neighborhood of the unit disk. Then f (z) ≤ M |z|n ,
where M is the maximum value of |f (z)| on the unit circle.

Proof. From the Maximum Principle we get the result that

|f (z)|/|z n | ≤ M.
Hence |f (z)| ≤ M |z|n . 

Lemma 13.6. Suppose, for all n, that the function f (z)/z n is well
defined at 0 and complex analytic in a neighborhood of the unit disk.
Then f is identically 0 on the unit disk.

Proof. From the preceding result, we have |f (z)| ≤ M |z|n . If |z| < 1,
then
lim M |z|n = 0.
n→∞
Hence |f (z)| = 0 if |z| < 1. By continuity, |f (z)| = 0 if |z| ≤ 1. 
13.6. Removable Singularities 171

13.6. Removable Singularities

Here we will prove the following result:

Theorem 13.7. Let U be an open set that contains a point b. Suppose

that f is complex analytic and bounded on U − {b}. Then f (b) can be
(uniquely) defined so that f is complex analytic in U .

Proof. Let γ and β and λ be the loops used to prove the Cauchy
Integral Formula. So, λ is a small loop surrounding b and γ is a big
loop surrounding b. Let |λ| denote the radius of λ. Let D be the open
domain bounded by γ. We define g : D → C by the integral

1 f (z)
g(a) = dz.
2πi γ z − a
The same calculation as in the proof of Theorem 13.4 shows that g is
complex analytic on all of D. We will show that f (a) = g(a) for all
a ∈ D − {b}. Once we know this, we set f (b) = g(b) and we are done.
Now suppose that a = b. Since f (z) is bounded in a neighborhood
of b we have 
f (z)
lim dz = 0.
|λ|→0 λ z − a

But, by the Cauchy Integral Formula,


1 f (z)
f (a) = dz
2πi β z − a
no matter which choice of λ we make. Therefore
 
1 f (z) 1 f (z)
f (a) = lim dz = dz = g(a).
|λ|→0 2πi β z − a 2πi γ z − a
So f (a) = g(a) for all a ∈ D − {b}. 

Lemma 13.8. Let D denote the unit disk. Suppose that f is complex
analytic in a neighborhood of D and |f (z)|/z n is bounded on D − {0}.
Then f is identically 0 on D.

Proof. The function f (z)/z n is complex analytic in the unit disk by

the above result. Lemma 13.6 now says that f is identically 0. 
172 13. A Primer on Complex Analysis

13.7. Power Series

We say that a sequence {an } of complex numbers satisfies the unit
convergence condition (or UCC) if
(13.10) lim an ρn = 0, ∀ρ ∈ [0, 1).
n→∞

The UCC implies that the terms in the sequence {|an |ρn } decay ex-
ponentially fast for any ρ < 1. To see this, we choose any ρ∗ ∈ (ρ, 1)
and note that
 n  n
ρ ρ
|an |ρn = |an |(ρ∗ )n × <
ρ∗ ρ∗
for n sufficiently large.

Exercise 5. Suppose that {an } satisfies the UCC. Let k > 0 be any
integer and let C be any constant. Prove that the sequence {Cnk an }
also satisfies the UCC.

Now we will discuss the convergence of power series to complex

analytic functions, as well as the term-by-term differentiation of these
series. Let {an } be a sequence satisfying the UCC. First, we define a
“finite series”, which is just a polynomial.

n
(13.11) fn (z) = ak z k .
k=0

Lemma 13.9. The sequence {fn (z)} is a Cauchy sequence of complex

numbers for all |z| < 1.

Proof. If a, b > N and N is sufficiently large, then


b 
b ∞
 δn
|fa (z) − fb (z)| = | an z n | ≤ |an ||z|n ≤ δn = .
n=a n=a
1−δ
N

Here we have chosen some ρ > |z| and taken δ = |z|/ρ∗ . This
∗

calculation establishes what we want. 

Lemma 13.9 says that the limit


(13.12) f (z) = an z n = lim fn (z)
n→∞
13.7. Power Series 173

exists provided that |z| < 1. Here is our main result about this infinite
series.
Theorem 13.10. f (z) is complex analytic in the open unit disk and
f  (z) is obtained by differentiating the series term-by-term.

Proof. Let gN = f − fN . Then

f (z + h) − f (z) fN (z + h) − fN (z) gN (z + h) − gN (z)
= + .
h h h
From Exercise 1 above fN (z) is complex analytic. Also, the sequence

{nan } satisfies the UCC by Exercise 4. Hence, limN →∞ fN (z) exists
at every point in the unit disk. Moreover, this limit is just obtained
by differentiating the series for f (z) term by term. To prove our result
we just have to show that
f (z + h) − f (z) 
lim = lim fN (z).
h→0 h N →∞
This is the same as showing that
gN (z + h) − gN (z)
lim lim = 0.
N →∞ h→0 h
On the individual terms we have the bound
an (z + h)n − an z n (z + h)n − z n
| | = |an || | ≤∗ n|an ||z + h|n−1 .
h h
The starred inequality comes from the fact that the map φ(z) = z n
expands distances in C by at most nδ n−1 as long as |z| ≤ δ.
As long as h is fairly small, we can choose some δ < 1 and restrict
our attention to the case |z + h| < δ < 1. Given the above estimate,
we get
∞ ∞

gN (z + h) − gN (z)
| |≤ n|an |δ n−1
= nδ|an |δ n = RN .
h
n=N n=N

(We are just calling the last expression RN for convenience.) But the
sequence {nδ|an |} satisfies the UCC by Exercise 5. Hence, the terms
comprising RN decay exponentially. Hence, limN →∞ RN = 0. But
the inequality above holds for any h with |z + h| < δ. Hence
gN (z + h) − gN (z)
lim lim | | ≤ lim RN = 0.
N →∞ h→0 h N →∞
This is what we wanted to prove. 
174 13. A Primer on Complex Analysis

The above result, applied iteratively, shows that the kth complex
derivative f (k) (z) is complex analytic in the open unit disk and is
obtained by differentiating the series for f (z) term-by-term k times.
Our discussion, which focused on the unit disk, generalizes in
a straightforward way. Say that the sequence {bn } satisfies the R-
convergence criterion if the sequence {an Rn } satisfies the UCC. In

this case the series bn (z − z0 )n is complex analytic in the open disk
of radius R about z0 and the same result as above applies.

13.8. Taylor Series

The basic result we want to prove is that a complex analytic function
equals its Taylor series. We begin with a technical lemma.

Lemma 13.11. Suppose that f is complex analytic in a neighborhood

of the unit disk. Then the sequence
{f (n) (0)/n!}
is bounded and hence satisfies the UCC.

Proof. It follows immediately from equation (13.9) that |f (n) | ≤

M n!, where M is the maximum value attained by f on the closed
unit disk. 

Lemma 13.11 says that the Taylor series for f about 0 defines a
power series which converges and is complex analytic in a neighbor-
hood of the unit disk. The next result says that f coincides with its
Taylor series in the unit disk.

Theorem 13.12. Suppose that f is complex analytic in a neighbor-

hood of the unit disk. Then f equals its Taylor series on the unit
disk.

Proof. Since the Taylor series f of f is defined and complex analytic

on the unit disk, we can consider the difference function f − f. This
complex analytic function has zero Taylor series. Thus, it suffices to
prove the following special case. If the Taylor series of f vanishes
identically at 0, then f is zero on the whole unit disk.
13.8. Taylor Series 175

If g is any function with g(0) = 0, we have

 1
(13.13) |g(z)| ≤ |g  (tz)|dt.
0

Here g (tz) is the complex derivative of the function z → g(tz). Equa-
tion (13.13) is best seen geometrically. The idea is that |g  (tz)| mea-
sures the speed of the curve t → g(tz) which connects 0 to g(z).
Let Δ be the closed unit disk. Fix n for the moment. Since
f n (0) = 0 we can choose δ > 0 so that |f (n) (z)| < 1 for all |z| < δ.
Applying equation (13.13) to g = f (n−1) , we get
(13.14) |f (n−1) (z)| ≤ |z|, ∀|z| ≤ δ.
Applying equation (13.13) to g = f (n−2) and using the bound in
equation (13.14), we get
(13.15) f (n−2) ≤ |z|2 /2, ∀|z| ≤ δ.
Continuing in this way, we get
(13.16) |f (z)| ≤ |z|n /n!, ∀|z| ≤ δ.
In particular, |f (z)|/|z|n is bounded on Dn −{0}, where Dn is the disk
of radius δ. Note that δ depends on n, but this does not bother us.
By compactness, |f (z)|/|z|n is bounded on Δ − Dn . Hence |f (z)|/|z n |
is bounded on Δ − {0}. Since this holds for all n, Lemma 13.8 says
that f is identically 0 on the unit disk. 

Exercise 6. Define the exponential function

∞
zn
E(z) = .
n=0
n!

Prove that the series defining E(z) converges on all of C. Prove also
that E  (z) = E(z) and that E(z1 + z2 ) = E(z1 )E(z2 ). For this last
part, you can do it by manipulating the series directly and applying
the binomial theorem. The restriction of E to R coincides with the
familiar exponential function.

Exercise 7. Define the two functions

z2 z4 z6 zz z5 z7
C(z) = 1 − + − ··· , S(z) = z − + − ··· .
2! 4! 6! z! 5! 7!
176 13. A Primer on Complex Analysis

Show that these series converge for all z ∈ C and that C(x) = cos(x)
and S(x) = sin(x) for all x ∈ R. Verify that E(z) = C(z) + iS(z).

Exercise 8. Let us define cos(x) and sin(x) such that the map
γ0 (x) = (cos(x), sin(x))
is the unit speed counterclockwise parametrization of the unit circle
such that γ(0) = (1, 0). Prove that C(x) = cos(x) and S(x) = sin(x)
for all x ∈ R. (Hint: Consider the map γ1 (x) = (C(x), S(x)). Check
that  
d 2 2
C (x) + S (x) = 0
dx
using term-by-term differentiation. From here it is not too hard to
show that γ0 and γ1 are the same parametrization of the unit circle.)

Exercise 9. Our main result in this section is definitely false for

smooth functions that are not complex analytic. Consider the func-
tion
f (t) = exp(−1/t2 ), t > 0.
When t ≤ 0 we define f (t) = 0. Prove that f is smooth and has a
trivial Taylor series about 0. This shows that smooth functions need
not equal their Taylor series.
Chapter 14

Disk and Plane Rigidity

In this chapter, we apply some of the complex analysis developed in

the previous chapter, notably the Maximum Principle and Theorem
13.7, to certain holomorphic maps of the disk and plane. The types
of results we prove show that certain weak-seeming conditions placed
on a complex analytic function actually place very strong restrictions
on the function. These kinds of rigidity results provide a link between
complex analysis and geoemtry.
As an application of the results, we will prove that stereographic
projection maps circles on S 2 to circles in C ∪ ∞. While not the most
elementary possible proof, our proof does give an application of the
complex analysis we have been developing. For a geometric proof of
the main result, see [HCV].

14.1. Disk Rigidity

We first prove Theorem 1.1, mentioned in Chapter 1.

Theorem 14.1. Let f be biholomorphism from the unit disk to itself.

Then f is a Möbius transformation.

Proof. If f (0) = 0, then we can find a linear fractional automorphism

H of Δ such that f ◦ H(0) = 0. Thus, it suffices to consider the case
when f (0) = 0. Since f  (0) exists, the function g(z) = f (z)/z is

177
178 14. Disk and Plane Rigidity

bounded in Δ. Hence, this function is complex analytic. Below we

will show that |f (z)| ≤ |z| for all z ∈ Δ, and the same argument,
applied to f −1 , shows that |f −1 (z)| ≤ |z| for all z ∈ Δ. These two
inequalities show that |f (z)| = |z| on Δ. But then g(Δ) is contained
in the unit circle, a 1-dimensional curve. This is impossible unless g
is a constant map. Hence there is a constant C such that f (z) = Cz.
Hence f is a linear fractional transformation.
It remains to show that |f (z)| ≤ |z| for z ∈ Δ. This is the same
as showing that |g(z)| ≤ 1 for all z ∈ Δ. Let Cr be the circle of radius
r < 1 about 0. Then |g(z)| ≤ 1/r on Cr . Hence |g(z)| ≤ 1/r if |z| < r
by the Maximum Principle. Letting r → 1, we see that |g(z)| ≤ 1 on
Δ. This is what we wanted to prove. 

Exercise 1. Prove the same result for a biholomorphism from the

upper half-plane to itself.

The next result shows the distinguished role played by the hyper-
bolic metric on the open unit disk, from the point of view of complex
analysis.

Lemma 14.2. Let Δ be the unit disk, equipped with the hyperbolic
metric from §10.7. Let f : Δ → Δ be a complex analytic map, not
necessarily a biholomorphism. Then f does not expand distances in
the hyperbolic metric.

Proof. We would like to see, at each point p ∈ Δ, that the differ-

ential df maps vectors having hyperbolic length 1 to vectors having
hyperbolic length at most 1. Call this the no-stretch property. It
suffices to prove that the no-stretch property holds for each p ∈ Δ.
We can find Möbius transformations T1 and T2 such that T1 (0) = p
and T2 (f (p)) = 0, respectively. The map g = T2 ◦ f ◦ T1 satisfies
g(0) = 0. Since T1 and T2 are hyperbolic isometries, g has the no-
stretch property at 0 if and only if f has the no-stretch property at p.
Since g(0) = 0, we just need to show that |g  (0)| ≤ 1 to establish the
no-stretch property for g at 0. The same argument as in the previous
proof establishes this fact. 
14.2. Liouville’s Theorem 179

Lemma 14.2 is more flexible than it first appears. Any disk W0 in

the plane has its own hyperbolic metric, so that a similarity carrying
Δ to W0 is a hyperbolic isometry. This principle should help you with
the next exercise.

Exercise 2. Suppose that U is some open set in the plane and

w ∈ U is some point. Suppose also that G : U → Δ is a holomorphic
map. Prove that there is some disk W ⊂ U , centered at w, whose size
does not depend on G, such that the hyperbolic distance from G(w)
to G(w ) is less than 1 for all w ∈ W .

14.2. Liouville’s Theorem

Here is Liouville’s Theorem.

Theorem 14.3 (Liouville). Suppose that f is a bounded holomorphic

function on C. Then f is constant.

Proof. Equation (13.7) says that


 f (z)
(14.1) f (a) = dz.
γ (z − a)2
Taking γ to be a large circle of radius r about 0, we see that the right
hand side of the above equation is at most C/r for some constant
C. Letting r → ∞, we see that f  (a) = 0. Since a is artitrary, f is
constant. 

Exercise 3 (Challenge). A function f : C → R is called harmonic

if it has the following property. For any disk D, the value of f at
the center of D equals the average value of f on D. Prove that a
bounded harmonic function is constant. This result is equivalent to
Liouville’s Theorem. (Sketch: You want to show that f (a) = f (b) for
all a, b ∈ C. Consider the difference Cr = Ar − Br , where Ar is the
average of f on the disk of radius r about a and Br is the average
of f on the disk of radius r about b. Show that limr→∞ Cr = 0, by
analyzing the intersection Ar − Br and observing that there is a lot
of cancellation in the computation of Ar − Br when r is large.)
180 14. Disk and Plane Rigidity

Exercise 4. Give an alternate proof of Liouville’s Theorem by show-

ing that the function g(z) = f (z)/z is holomorphic in the whole plane
and then applying the Maximum Principle.

Exercise 5. Use Liouville’s Theorem to give another proof of the

Fundamental Theorem of Algebra. (Hint: Let P (z) be a complex
polynomial supposedly with no roots. Consider f (z) = 1/P (z).)

Exercise 6. Suppose that g : C → C is a continuous map with the

following properties:
• g(0) = 0.
• g is holomorphic on C − {0}.
• |g(z)| < C|z| for |z| sufficiently large.
Prove that g(z) = Az for some constant A. Hint: First show that g is
holomorphic on all of C, then show the same thing for h(z) = g(z)/z.

Exercise 7 (Challenge). Suppose that f : C → C is a holo-

morphic function such that |f (z)| < |z|n for some n and all z with |z|
sufficiently large. Prove that f is a polynomial.

Lemma 14.4. Suppose f is a homeomorphism of C that is complex

analytic except at finitely many points. Then f (z) = Az + B for some
constants A and B.

Proof. Combining Lemma 2.2 and Theorem 13.7, we see that f is

complex analytic on all of C. The function f  (z) cannot identically
vanish. So, we can compose f with translations and then assume
that f (0) = 0 and f  (0) > 0. But then there is some C > 0 such that
|f (z)| > C|z| provided that |z| is sufficiently small. Now consider the
function
1
(14.2) g(z) = .
f (1/z)
Note that g satisfies the conditions of Exercise 4. Hence g(z) = Az.
But then f (z) = z/A. Remembering that this new version of f is a
14.3. Stereographic Projection Revisited 181

translation of the original, we see that the original version of f has

the form Az + B. 

14.3. Stereographic Projection Revisited

Let φ : S 2 → C ∪ ∞ be stereographic projection.
Lemma 14.5. The differential dφ is a similarity on the tangent plane
Tx at x ∈ S 2 − {(0, 0, 1)}.

Proof. One can prove this result by a direct calculation, but we will
give a geometric proof. Our proof refers to Figure 14.1. We think
of C as the xy-plane. Let T = Tx and let T  be the plane through
x parallel to C. Let L be the line joining (0, 0, 1) to x. Figure 14.1
shows the intersection of all these objects with the plane Π containing
(0, 0, 0) and (0, 0, 1) and x. The X Theorem from §8.3 implies that
the lines T ∩ Π and T  ∩ Π make the same angle with L = L ∩ Π.
Hence, reflection in the plane P = L⊥ carries T isometrically to T  .
The differential dφ has the following description: First reflect T
to T  through P , then radially project T  to C through p. Thus dφ is
the composition of an isometry and a similiarity, which is just another
similarity. 

P T
L
T’
x
C

Figure 14.1. The Differential of Stereographic Projection

Exercise 8. Prove Lemma 14.5 by a direct calculation, using equa-

tion (9.9).
Lemma 14.6. Suppose I is an isometry of S 2 . Then I  = φ ◦ I ◦ φ−1
is a linear fractional transformation.
182 14. Disk and Plane Rigidity

Proof. Here is where complex analysis comes in. We can find a linear
fractional transformation T such that J = T ◦ I  fixes ∞. It suffices
to show that J is a linear fractional transformation. The map J is
smooth except at perhaps a finite list of points. (The points we are
not certain about are various images and preimages of ∞.) Moreover,
by Lemma 14.5, the differential dJ is a similarity at all but finitely
many points. Hence J is a homeomorphism of C that is holomorphic
except at finitely many points. By Lemma 14.4, the map J is linear,
and hence a linear fractional transformation. But then I  is a linear
fractional transformation. 

Now we come to the main application of the results in this section.

Again, this result has a direct geometric proof, but we want to show
how one can get the result from complex analysis.
Lemma 14.7. Stereographic projection maps circles on S 2 to gener-
alized circles in C ∪ ∞.

Proof. Let C be a circle on S 2 . Let I be an isometry of S 2 such that

I(C) contains (0, 0, 1). As we remarked in §9.5, the curve L = φ(I(C))
is a straight line (union ∞). Thus, φ(I(C)) is a generalized circle. But
φ(L) = I  (φ(C)), I  = φ ◦ I ◦ φ−1 .
By Lemma 14.4, the map I  is a linear fractional transformation.
Therefore, so is J = (I  )−1 . But φ(C) = J(L), where J is a linear
fractional transformation and L is a generalized circle. Since linear
fractional transformations map generalized circles to generalized cir-
cles, as we saw in Chapter 10, we see that J(L) = φ(C) is also a
generalized circle. 

Exercise 9. Generalize the definition of stereographic projection so

that it works in all dimensions and prove that generalized stereo-
graphic projection maps spheres to spheres. You should be able to
deduce this from the 2-dimensional case and symmetry.
Chapter 15

The Schwarz–Christoffel
Transformation

In this chapter we will study some examples of Schwarz–Christoffel

transformations. These maps turn out to give biholomorphisms be-
tween the upper half-plane and the interiors of polygons. For ease of
exposition, we will restrict our attention to the case when the sides of
the polygon are parallel to the coordinate axes. We call such polygons
rectilinear polygons; see Figure 15.1.

Figure 15.1. A rectilinear polygon

One remarkable thing about the Schwarz–Christoffel transforma-

tions is that there is, in a sense, an explicit formula for them. The
book [DRT] has a great deal of information about these maps, in-
cluding a discussion of their history.

183
184 15. The Schwarz–Christoffel Transformation

15.1. The Basic Construction

Suppose that x1 < x2 < · · · < xn ∈ R and e1 , · · · , en are numbers
such that ej = ±1/2 for all j and e1 + · · · + en = −2.
Let U ⊂ C denote the upper half-plane. Let U ∗ ⊂ C denote the
region obtained by deleting the closed downward pointing rays which
start at x1 , . . . , xn . We are mainly interested in U , but the larger
region U ∗ is convenient for technical purposes.

Figure 15.2. The region U ∗

Consider the function

(15.1) f (z) = (z − x1 )e1 · · · (z − xn )en .

If we try to define this function in all of C we run into trouble because

we cannot consistently define f all the way around a loop which circles
around xj . Since U ∗ has no loops like this, f is defined and complex
analytic in all of U ∗ .
We define a function F : U ∗ → C as follows. First we set
F (i) = 0. Next, for any z ∈ U ∗ , we let γ be a piecewise smooth
path connecting 0 to z, and we set

(15.2) F (z) = f (z)dz.
γ

Equation (15.2) is well defined by Theorem 13.1. It follows almost

immediately from the Fundamental Theorem of Calculus that F is
holomorphic in U ∗ and F  (z) = f (z). In particular, F  (z) never
vanishes in U ∗ . Here is our main result about F .

Theorem 15.1. F is well defined and continuous on R ∪ ∞. The

image F (R ∪ ∞) is a closed polygonal loop whose sides are alternately
parallel to the real and imaginary axes. If F (R ∪ ∞) is an embedded
polygon, then F is a biholomorphism from U to the polygonal domain
bounded by F (R ∪ ∞).
15.2. The Inverse Function Theorem 185

15.2. The Inverse Function Theorem

As a prelude to proving Theorem 15.1. We prove a special case of the
Inverse Function Theorem. For the general case, see, e.g., [SPI].

Theorem 15.2. Let f be a holomorphic map defined in a neighbor-

hood of z ∈ C. Suppose that f  (z) = 0. Then the restriction of f to
a neighborhood of z has an inverse, and f −1 is also holomorphic.

Proof. We can translate and scale so that z = 0 and f (0) = 0 and

f  (0) = 1. Let Dr be the disk of radius r about 0. For r small, we
have |f  (z) − 1| < 1/100 for all z ∈ Dr . Let z1 = z2 be two points
in Dr . Let L be the straight line joining these points. Given our
bounds on f  (z) along L, we see that the curve f (L) nearly has the
same length as L and points almost in the same direction as L at all
points. Hence f (z1 ) = f (z2 ). Hence f is injective on Dr for r small.
The same argument shows that f (∂Dr ) is a closed loop that is at
least (say) r/2 from 0 and winds once around 0. Let Δr denote the
set of points w such that f (∂Dr ) winds once around w. Note that Δr
is an open neighborhood of 0. Suppose there is some w ∈ Δr −f (Dr ).
Consider the 1-paramater family of loops γt = f (t∂Dr ). For t close
to 0, the loop γ1 winds 0 times around w. On the other hand, γ1
winds once around w. In order for the winding number to change in
this way, γt must contain w for some t. But then w ∈ f (Dr ). Hence
f : Dr → Δr is a surjection.
Now we know that f : Dr → Δr is a bijection. So, f −1 : Δr → Dr
exists. Our injectivity proof also shows that f −1 is continuous: f
cannot map far away points close together. One way to see that f is
differentiable at 0 is that the dilated maps gn (z) = nf (z/n) converge
to a similarity as n → ∞. But the dilated inverse of f is the inverse
of the dilation of f . Hence, the dilations of f −1 also converge to a
similarity. This shows that f −1 is differentiable at 0. The chain rule
now shows that (f −1 ) (0) = 1/f  (0). The same argument works at
any other point z in the interior of Dn . This shows that f −1 has a
continuously varying complex derivative in the interior of Δn . Hence,
f −1 is holomorphic in Δn . 
186 15. The Schwarz–Christoffel Transformation

15.3. Proof of Theorem 15.1

We already know that F is defined on R − {x1 , . . . , xn }, and F is
pretty obviously continuous where defined.

Exercise 1. Prove that F is defined and continuous on R ∪ ∞.

(Hint: Use the same definition at these points as for the other points.
The finiteness of integrals such as
 1  ∞
1
1/2
dx; x−2 dx
0 x 1

is what makes the definition work.)

Now we want to analyze the image F (R ∪ ∞). The points

x1 , . . . , xn divide R into the n + 1 intervals I0 , . . . , In . Actually,
I0 = (−∞, x1 ) and In = (xn , ∞) are rays. Let Jk = F (Ik ).
When we square f , we get

f 2 (z) = (z − x1 )±1 · · · (Z − xn )±1 .

From this we see that f 2 is positive on I0 , negative on I1 , positive on

I2 , and so on. So, f is real on I0 , pure imaginary on I1 , real on I2 ,
pure imaginary on I3 , and so on. But F  (z) = f (z), and the argument
of F  (z) tells us how F rotates points in a neighborhood of z. Hence
J0 is a horizontal segment, J1 is a vertical segment, J2 is a horizontal
segment, and so on. Since F is continuous on R ∪ ∞, we see that
these segments all piece together to give the kind of path described
in Theorem 15.1.
We orient R from −∞ to +∞. If you walk along R, then U
lies to your left. Being complex analytic, the map F is orientation
preserving. This means that, as you walk around F (R), the image
F (U ) (at least locally) lies to your left.

Exercise 2. Show that F (R) turns left at xj if ej = −1/2 and

right if ej = −1/2. Geometrically, f (U ) looks like one quadrant in
a neighborhood of f (xj ) if ej = −1/2 and three quadrants if ej = 1/2.
15.3. Proof of Theorem 15.1 187

Given Exercise 2, and the fact that e1 + · · · + en = −2, the polyg-

onal path F (R) turns once around counterclockwise (the equivalent
of 4 left turns.) Hence F (I0 ) and F (In ) travel in the same direction
and fit together seamlessly.
Now we suppose that F (R ∪ ∞) is an embedded polygon. Let R
be the region bounded by F (R ∪ ∞).
Lemma 15.3. F (U ) ⊂ R.

Proof. Let U = U ∪ R ∪ ∞. The set U is a compact subset of the

Riemann sphere S 2 = C ∪ ∞. Lemma 2.2 tells us that F (U ) is a
bounded subset of C. Since U is a compact subset of S 2 and F is
continuous, F (U ) is compact.
If F (U ) is not a subset of R, we can find a point p ∈ U such that
F (p) lies in the boundary of F (U ) but not in ∂R. Note that p must
lie in U because F (U − U ) = ∂R. By the Inverse Function Theorem,
F maps a neighborhood of p onto a neighborhood of F (p). But then
F (p) could not lie in the boundary of F (U ). This contradiction shows
that F (U ) ⊂ R. 

Exercise 3. Use essentially the same argument that we gave in §5.3,

in connection with the Fundamental Theorem of Algebra, to show
that F (U ) = R.
Lemma 15.4. F is one-to-one on U .

Proof. Let B ⊂ U denote the set of points z such that F (z) = F (z  )

for some other z  ∈ U . Consider the extreme case when B = U .
Choose some z ∈ U ∗ ∩ R, and let {zn } be a sequence of points in
U converging to z. Let {zn } be a sequence of points in U such that
F (zn ) = F (zn ).
By Theorem 15.2, the map F is one-to-one in a neighborhood
of z, so there is some minimum distance between z and zn . Passing
to a subsequence, we can assume that zn converges to some point
z  ∈ R ∪ ∞. From the minimum distance property, z = z  . By
continuity F (z) = F (z  ). But F is one-to-one on R ∪ ∞.
Now we know that B = U . We will show that B is both open
and closed. Since U is connected, the only possibility is that B = ∅.
188 15. The Schwarz–Christoffel Transformation

Essentially, that same argument we just gave to show that B = U

shows that B is closed in U . We just have to show that B is open.
Suppose that z ∈ B and F (z  ) = F (z). By Theorem 15.2, F
maps neighborhoods of z and z  onto neighborhoods of F (z) = F (z  ).
Hence B contains a neighborhood of z. Hence B is open. 

Lemma 15.4 and Exercise 3 combine to show that F : U → R is

a complex analytic bijection. Theorem 15.2 now shows that F −1 is
complex analytic. Hence F is a biholomorphism. This completes the
proof of Theorem 15.1.

15.4. The Range of Possibilities

Theorem 15.1 explains how we can get some rectilinear polygons as
images of the upper half-plane under a Schwarz–Christoffel transfor-
mation. It turns out that, up to scaling, we can get all of them this
way. The idea is to show that we can vary the inputs of the construc-
tion so as to produce every possibility. Here is the main result.

Theorem 15.5. Up to scaling, every rectilinear polygon is the image

of the upper half-plane under a Schwarz–Christoffel transformation.

The proof of Theorem 15.5 is a bit hard going, but I included

it because I like the result and also because I will use Theorem 15.5
in the next chapter to prove the Riemann Mapping Theorem. Once
we know the Riemann Mapping Theorem, we can say right away
that every open solid polygon is the image of the upper half-plane
under a biholomorphsm. However, without knowing Theorem 15.5,
it seems difficult to prove, just from the Riemann Mapping Theorem,
that every biholomorphism from the upper half-plane to a rectilinear
polygon is given (up to composition with Möbius transformations) by
a Schwarz–Christoffel transformation.
One unfortunate thing about our proof of Theorem 15.5 is that it
is not completely self-contained. It relies on a basic result in topology
known as Invariance of Domain. The Invariance of Domain result has
always struck me as obviously true, but the proof is fairly difficult.
15.5. Invariance of Domain 189

15.5. Invariance of Domain

The following result is Theorem 2B.3 in [HAT].

Theorem 15.6 (Invariance of Domain). Suppose that U ⊂ Rn is an

open set, and Φ : U → Rn is a continous and one-to-one map. Then
Φ(U ) is open in Rn .

We are mainly interested in a certain corollary of the Invariance

of Domain result. Suppose that X and Y are spaces, both homeo-
morphic to open subsets of Rn . A map Φ : X → Y is proper if it has
the following property. If K ⊂ Y is compact, then Φ−1 (K) ⊂ X is
compact.

Lemma 15.7. Let X and Y be spaces, both homeomorphic to open

subsets of Rn . Suppose also that X is nonempty and Y is connected.
If Φ : X → Y is a one-to-one, continuous, and proper map, then
Φ(X) = Y .

Proof. We suppose that this result is false and derive a contradiction.

By Invariance of Domain, Φ(X) is an open subset of Y . Moreover,
Φ(X) is nonempty. Since Y is connected, Y itself is the only subset
of Y that is simultaneously open, closed, and nonempty. We conclude
that Φ(X) is not closed. Hence, we can find a point

q ∈ Φ(X) − Φ(X).

Given the location of q, we can find a sequence {pk } ∈ Φ(X) such

that pk → q.
We can choose {pk } so that it lies in a compact subset of Y .
Since Φ is proper, there is a sequence {pk }, contained in a compact
subset of X such that Φ(pk ) = pk . Since {pk } lies in a compact
subset of X, this sequence has a convergent subsequence. Passing
to this subsequence, we let q  = lim pk ∈ X. Since Φ is continuous,
Φ(q  ) = q. This contradicts the fact that q ∈ Φ(X). 

Given Lemma 15.7, the rest of our proof of Theorem 15.5 is self-
contained.
190 15. The Schwarz–Christoffel Transformation

15.6. The Existence Proof

Say that a marked loop is a counterclockwise oriented rectilinear loop
with a preferred edge. We fix some length n sequence Σ of “lefts” and
“rights”, with a total of 4 more “lefts” than ”rights”, and we let YΣ
denote the space of all marked polygons that have this sequence of
turns as we trace around it counterclockwise, starting with the pre-
ferred edge. Let YΣ ⊂ YΣ denote the subset of embedded ones. Using
the side lengths of the polygons, we consider YΣ and YΣ as subsets of
Rn . This makes these sets into metric spaces.

Exercise 4. Let Σ be a sequence of length n, as above. Prove

that YΣ and YΣ are both homeomorphic to open subset of Rn−2 .

Exercise 5 (Challenge). Prove that YΣ is connected. (Hint: The

result is certainly true for the sequence Σ = LLLL. Here YΣ is just
the space of rectangles. In general, do induction on the length of Σ.
Show that a rectilinear polygon always has a “spot” where you can
continuously shrink one of the edges to a point without destroying
the embedding property; see Figure 15.3.)

Figure 15.3. Shrinking an edge

Let Σ1 and Σ2 be two sequences. We write Σ1 → Σ2 if Σ2 is

obtained from Σ1 by the insertion of LR or RL somewhere in Σ1 .
For any sequence Σ2 except LLLL, there is some sequence Σ1 such
that Σ1 → Σ2 . Say that the sequence Σ1 is good if some polygon in
YΣ is the image F (R ∪ ∞) for a Schwarz–Christoffel transform F .

Lemma 15.8. All sequences are good.

15.6. The Existence Proof 191

Proof. The sequence LLLL is certainly good. We will prove the

following statement. If Σ1 → Σ2 and Σ1 is good, then so is Σ2 . This
lemma then follows from induction.
Let P be a polygon in Y1 that is realized as the image F (R ∪ ∞)
for some Schwarz–Christoffel transformation F . Let x1 , ..., xn be the
special points corresponding to F . The exponents e1 , ..., en are chosen
so as to match the sequence of lefts and rights in Σ1 .

Figure 15.4. A zig-zag

Let’s say that Σ2 is obtained from Σ1 by inserting LR after the

kth slot. Then, between xk and xk+1 , we insert two new points x1 and
x2 . We place these points extremely close together, and right near the
middle of the interval bounded by xk and xk+1 . We chose additional
exponents e1 = −1/2 and e2 = 1/2. Let F  be the new Schwarz–
Christoffel transform based on the points x1 , . . . , xk , x1 , x2 , xk+1 ,
. . . , xn and the corresponding exponents. When x1 and x2 are very
close, the images F (R ∪ ∞) and F  (R ∪ ∞) are almost identical, ex-
cept that the single edge F (Ik ) is replaced by a zig-zag, as shown in
Figure 15.4. If this perturbation is small, the polygon in question is
embedded. 

We fix a sequence of exponents e1 , . . . , en , as above. These expo-

nents determine the corresponding sequence Σ of lefts and rights. The
input to our construction is a positive constant c and points x1 = −1
and 0 = x2 < x3 < · · · < xn = 1. Let X  be the set of possible inputs.
X  is homeomorphic to Rn−2 .
Given some particular input p ∈ X  , the output is a polygonal
loop Φ(p) = F (R ∪ ∞). Here F is the Schwarz–Christoffel transfor-
mation from equation (15.2), rescaled by c. We scale by c at the end
for technical purposes. By construction Φ(p) is a point in the space
of Y  = YΣ . Thus, we have a map Φ : X  → Y  . The map Φ is pretty
192 15. The Schwarz–Christoffel Transformation

obviously continuous, given the formula for the Schwarz–Christoffel

transformation.

Lemma 15.9. Φ : X  → Y  is one-to-one.

Proof. Suppose that F1 and F2 are two Schwarz–Christoffel trans-

formations such that F1 (R ∪ ∞) and F2 (R ∪ ∞) trace out the same
polygon. We mean also that F1 (xi ) = F2 (xi ) for all i. Define
G = F1−1 F2 : U → U . This is a biholomorphism that fixes −1,
0, and 1. By Exercise 6 below, G is the identity. 

Exercise 6. Let G : U → U be a biholomorphism from the upper

half-plane to itself. Suppose that G fixes the points −1, 0, and 1.
Prove that G is the identity map.

Recall that Y ⊂ Y  is the subset of embedded polygons. We let

X = Φ−1 (Y ). Since all sequences are good, X is nonempty. Since Φ
is continuous, X is open. To finish the proof of Theorem 15.5, we just
have to show that Φ : X → Y is proper. Then we can apply Lemma
15.7 and conclude that Φ(X) = Y , as desired.
Let K be a compact subset of Y . We want to show that Φ−1 (K)
is a compact subset of X. This is the same as showing that Φ−1 (K)
is a compact subset of X  . We can put this another way. Suppose
{pk } is a sequence of inputs that exits every compact subset of X  .
We want to prove that Φ(pk ) exits every compact subset of Y .
Say that a special interval relative to the index k is an interval
bounded by consecutive points xk,j and xk,j+1 . If {pk } exits every
compact subset of X  , then at least one of 3 things happens on a
subsequence. Either ck → ∞ or ck → 0 or {ck } is bounded. In the
last case, the length of the shortest special interval tends to 0 with k.
Suppose that ck → ∞. Now matter how we choose the input,
all the points in the interval [−2/3, −1/3] ⊂ I0 are at least 1/3 units
away from all the special points. Looking at the formula for F in
equation (15.2), we see that F (I0 ) has length at least (1/3)n+1 . But
then one of the sides of the kth output has length at least ck (1/3)n+1 ,
a number that tends to ∞ with k. Hence Φ(pk ) exits every compact
subset of Y .
15.6. The Existence Proof 193

Suppose that ck → 0. No matter what the input, we can find 3

points y1 , y2 , y3 ∈ R, all in distinct special intervals, that are all at
least 1/2n from any of the endpoints of the special intervals. Looking
at the formula for F in equation (15.2), we see that there is some
constant C, independent of inputs, such that |F (yi )| < C for i =
1, 2, 3. But then the polygon corresponding to Φ(pk ) has three sides
which come within Cck of the origin. But Cck tends to 0. This shows
that points Φ(pk ) exit every compact set of Y .
The following result finishes our proof of Theorem 15.5.

Lemma 15.10. If {ck } is bounded and the length of some special

interval tends to 0, then Φ(pk ) exits every compact subset of Y .

Proof. We will suppose that Φ(pk ) lies in a compact subset of Y and

derive a contradiction. After a bounded amount of scaling, we can
assume that ck = 1 for all k. Let Fk be the Schwarz–Christoffel trans-
formation associated to pk . Let Pk = Fk (R ∪ ∞). By compactness,
there is some D > 0 such that the sides of Pk have length at most
1/D and the distance between any two distinct vertices of Pk is at
least D. Here D is independent of k.
Passing to a subsequence, we can assume that xk1 , . . . , xkn con-
verges to points x∞,1 , . . . , x∞,m . Here m < n because some points
have coalesced. We have associated exponents e∞,1 , . . . , e∞,m , where
e∞,k is the sum of the exponents of the points that coalesce to x∞,k .

Exercise 7. Prove that e∞,k ≥ −1/2. (Hint: Use the fact that
the sides of Pk have length at most 1/D, independent of k.)

Because the integrands for Fk converge at each point of U ∗ , the

sequence {Fk } of maps converges to a map F∞ : U ∗ → C, defined

exactly as in equation (15.2). Because e∞,k ≥ −1/2 and e∞,k =
−2, the map F∞ extends to be continuous on R ∪ ∞.
Choose an index m such that 2 or more points xk,j converge to
x∞,m . Consider the special intervals A∞ and B∞ on either side of
x∞,m . There are special intervals Ak and Bk such that Ak → A∞
and Bk → B∞ . By our choice of index m, the intervals Ak and Bk
are not consecutive.
194 15. The Schwarz–Christoffel Transformation

Exercise 8. Prove the following result. There is some K such that

k > K implies that Fk (Ak ) contains all but D/3 of F∞ (A∞ ). (Hint:
use the finiteness of of all the integrals involved and the convergence
of the integrands. The same result holds for B in place of A.)

By Exercise 8, some endpoint of Fk (Ak ) is within 2D/3 of some

endpoint of Fk (Bk ). This contradicts the existence of D. 
Chapter 16

Riemann Surfaces and

Uniformization

The purpose of this chapter is to define the notion of a Riemann

surface. A Riemann surface is essentially a surface that is built out
of pieces of C glued together with complex analytic maps. Once we
know about Riemann surfaces, we can speak about complex analytic
maps between them. We will prove some basic results about such
maps, relying on the material from the previous 3 chapters.
Following the discussion of Riemann surfaces, we will prove the
Riemann Mapping Theorem. For another proof of this result, one
that does not rely on Theorem 15.5, see [AHL].
The Riemann Mapping Theorem is a special case of the Poincaré
Uniformization Theorem, a result we will state without proof. A proof
can be found in [BE2]. After stating the Uniformization Theorem,
we will deduce some consequences from it.

16.1. Riemann Surfaces

Let S be a surface. Recall that a smooth structure on S is a maximal
collection of coordinate charts which have the property that the over-
lap functions are all smooth. A Riemann surface is defined in a similar
way, with the word complex analytic replacing the word smooth. That

195
196 16. Riemann Surfaces and Uniformization

is, a Riemann surface structure on a surface is a maximal collection

of coordinate charts such that the overlap functions are all smooth.
Here are some examples:

Open Subsets of C. Any open subset of C is a Riemann sur-

face. We can take the coordinate chart maps to be the identity.

The Riemann Sphere. We can think of S 2 as C ∪ ∞. Then

U1 = C is a neighborhood of {0} and U2 = C ∪ ∞ − {0} is a neigh-
borhood of ∞. The identity map is a homeomorphism from U1 to C
and the map f (z) = 1/z is a homeomorphism from U2 to C. The
overlap U1 ∩U2 is C −{0} and the overlap function is just f (z) = 1/z,
a complex analytic function. We already have a collection of (two)
coordinate charts which cover S 2 , and we can complete this collection
to a maximal collection. This makes S 2 into a Riemann surface. This
surface is known as the Riemann sphere.

Flat Tori. Let P be a parallelogram. If we glue the opposite sides

of P together by translations, then we produce a closed surface. We
can find a covering of S by coordinate charts whose overlap functions
are translations, i.e., maps of the form z → z + C for various choices
of the constant C. Such maps are complex analytic, and so we can
make these flat tori into Riemann surfaces in a natural way.

Hyperbolic Surfaces. Recall that a hyperbolic structure on a sur-

face is a maximal collection of coordinate charts into H 2 such that
the overlap functions are all restrictions of hyperbolic isometries. If
we only use orientation preserving hyperbolic isometries, then these
maps are all linear fractional transformations. Linear fractional trans-
formations are complex analytic, and so a hyperbolic structure on a
surface is always a Riemann surface structure.

Exercise 1. In §12.6 we discussed the notion of a Riemannian cover-

ing space. We can similarly define a Riemann surface covering. This
would be a covering map between Riemann surfaces that is complex
analytic. Given a covering map E : S → S, prove that S can be made
into a Riemann surface such that E is a Riemann surface covering.
16.2. Maps Between Riemann Surfaces 197

Exercise 2. Let E(z) be the exponential function, as defined in Ex-

ercise 7 of §13.8. Prove that E is a covering map from C to C − {0}.
(Hint: Use the identities in Exercises 7–9 of §13.8 to get a handle on
the geometry of E.)

Exercise 3 (Challenge). Let X be the space obtained by glu-

ing together two copies of the solid unit square, along all sides (see
Figure 16.1). Give X the structure of a Riemann surface (by finding
local charts) so that there is a biholomorphic map between X and the
Riemann sphere. (Hint: For the coordinate charts, the only tricky
part is thinking about what to do at the vertices and edges. Think
about the Christoffel transform between the square and the upper
half plane.)

1 1

4 2 4 2

3 3
Figure 16.1. Gluing 2 squares

16.2. Maps Between Riemann Surfaces

Suppose S1 and S2 are two Riemann surfaces. A map f : S1 → S2 is
complex analytic in a neighborhood of p1 ∈ S1 if there are neighbor-
hoods U1 of p1 and U2 of p2 = f (p1 ), together with coordinate charts
fj : Uj → C such that the map f2 ◦ f ◦ f1−1 is complex analytic. f is
complex analytic on S1 if f is complex analytic in a sufficiently small
neighborhood of every point. We can use some of the machinery from
Chapter 13 to prove nontrivial results about maps between Riemann
surfaces. This chapter contains a sampler of these results.
Theorem 16.1. There is no nontrivial complex analytic map from a
compact Riemann surface into C.
198 16. Riemann Surfaces and Uniformization

Proof. Suppose f : S → C is complex analytic. Since S is compact

f achieves its maximum at some point p ∈ S. Let U be a coordinate
chart about p and let g : U → C be a coordinate chart. Then h =
f ◦ g −1 is a complex analytic map from the open set g(U ) into C.
Moreover, h takes its maximum value at an interior point of g(U ).
But a nonconstant complex analytic map cannot have an interior
maximum, according to the Maximum Principle from §13.5. 

On the other hand, there are plenty of complex analytic maps

from the Riemann sphere to itself. For instance, any rational function
P (z)
R(z) = Q(z) is a complex analytic map from the Riemann sphere to
itself. Here P and Q are polynomials. The set R−1 (∞) is contained
in the set of zeros of Q.
Theorem 16.2. There is no nonconstant complex analytic map from
C into a hyperbolic surface.

Proof. Let f : C → S be a complex analytic map from C to S. Let

E : H 2 → S be the universal covering map. Using the lifting property
for maps we can find a lifting f : C → H 2 such that E ◦ f = f . (We
produce f by partitioning C into an infinite grid of squares, and
applying the lifting theorem one square at a time.) By construction
f is complex analytic. The point is that on small neighborhoods E −1
is defined and complex analytic; and f = f ◦ E −1 on these small
neighborhoods. However, we can take H 2 as the open unit disk. So,
f is a bounded complex analytic function on C. However, all such
maps are constant. Since f is constant, so is f . 

Any complex analytic homeomorphism from C to C is a linear

map, by Corollary 14.4. Our proof of the next result uses this fact.
Theorem 16.3. Suppose that S is a Riemann surface which has a
non-Abelian fundamental group. Then there is no complex analytic
covering map of the form E : C → S.

Proof. Suppose that E : C → S exists. Let G be the fundamental

group of S. Then G acts on C as the deck group. Each element
g ∈ G acts as a complex analytic homeomorphism of C. Hence g is
a complex linear map. Being an element of the deck group, g acts
16.3. The Riemann Mapping Theorem 199

without any fixed points. Therefore, g must be a translation. In

short G is a group of translations. But any two translations commute
and hence G is Abelian. This contradiction shows that E does not
exist. 

16.3. The Riemann Mapping Theorem

Let Δ denote the open unit disk. Say that a Jordan domain is any
set of the form h(Δ), where h : C → C is a homeomorphism.
Theorem 16.4 (Riemann Mapping Theorem). Let D be any Jordan
domain. There exists a biholomorphism from Δ to D.

Riemann gave an intuitive description of the Riemann map. Imag-

ine that the domain D is a uniformly conducting material, and that
an electric potential of 1 is maintained at some interior point x ∈ D
and a potential of 0 is maintained on the boundary of D. The equipo-
tential lines form loops around x, and the electricity flowing from x
out to ∂D flows along lines perpendicular to the equipotential loops.
The equipotential loops and the flow lines form a kind of wavy coor-
dinate system on D. The Riemann map, if it is normalized to send
0 to x, sends the ordinary polar coordinate system on Δ to the wavy
one.
We will give a proof of the Riemann Mapping Theorem that is
based on Theorem 15.5.

Exercise 4. Prove the following statement. For any  > 0, there

is an embedded rectilinear polygon P such that every point of ∂D is
within  of P , and vice versa.

We scale the picture so that Δ ⊂ D. For each positive integer

n, choose a rectilinear polygon that is within 1/n of ∂D in the sense
of Exercise 4. Let Dn be the region bounded by this polygon. The
polygon itself is ∂Dn .
Since Δ and the upper half plane are biholomorphically equiva-
lent, Theorem 15.5 says that there is a biholomorphism Fn : Δ → Dn .
Composing Fn with a Möbius transformation of Δ, we arrange that
Fn (0) = 0 for all n. The rest of the proof amounts to showing that
200 16. Riemann Surfaces and Uniformization

the sequence {Fn } converges to the desired map.

Exercise 5. Let r < 1 and let Δ(r) denote the disk of radius r
centered at the origin. Prove that there is some constant R, depend-
ing on r but not on n, such that |Fn (z)| < R for all z ∈ Δ(r). (Hint:
Apply equation (14.1), using a circular loop γ ⊂ Δ that bounds Δ(r )
for some r  > 1.)

Since D is bounded, we can pass to a subsequence so that {Fn (z)}

converges on a countable dense subset of points z ∈ Δ. But then,
Exercise 5 guarantees that {Fn (z)} converges uniformly on each disk
Δ(r). That is, for any  > 0, there is some N such that n > N implies
that |Fm (z) − Fn (z)| <  for all m, n > N .
Let F = lim Fn . We have a converging sequence of maps, all of
which satisfy the Cauchy Integral Formula for all loops in Δ. Hence,
F satisfies the Cauchy integral as well. Hence F is holomorphic. The
main thing we want to rule out is that F is the constant map. The
next lemma does this for us.

Lemma 16.5. |F  (0)| ≥ 1.

Proof. Let Gn = Fn−1 . Recall that Fn (0) = 0 and Δ ⊂ D. Hence

Gn (Δ) ⊂ Δ and Gn (0) = 0. By Lemma 14.2, we have the inequality
|Gn (0)| ≤ 1. 

Exercise 6. Imitate the proof of Lemma 16.5 to show that F  (z) > 0
for all z ∈ Δ.

Lemma 16.6. F is one-to-one.

Proof. Suppose that F (z1 ) = F (z2 ). Then, by Theorem 15.2, there

are disjoint open sets U1 and U2 such that F (U1 ) = F (U2 ). But then
Fn (U1 ) and Fn (U2 ) overlap for large n. This contradicts that Fn is
one-to-one. 

Since F  (z) never vanishes, Theorem 15.2 shows that F −1 is holo-

morphic. Now we know that F is a biholomorphism from Δ onto
16.4. The Uniformization Theorem 201

F (Δ). Certainly F (Δ) ⊂ D. To finish the proof, we just have to

show that F (Δ) = D.
Choose w ∈ D. Let zn = Fn−1 (w). Call w good if the sequence
{zn } remains within a compact subset of Δ. Otherwise call w bad .
If w is good, there is at least one accumulation point z ∈ Δ of {zn }.
Since Fn (zn ) = w and we have a uniform bound on |Fn | in a neigh-
borhood of z, we have F (z) = w. We just have to show that every
point in D is good.
Lemma 16.7. w is contained in the interior of a disk W ⊂ D with the
following property. For all w ∈ W , the hyperbolic distance between
Fn (w) and Fn (w ) is less than 1, independent of n.

Proof. Apply Exercise 2 from Chapter 14 to the map G = Fn−1 and

some open set U ⊂ D such that w ∈ U and U ⊂ Fn (Δ) for all n. 

Note that w is good if and only if there is some K such that

{zn } stays within K hyperbolic units of 0. It therefore follows from
Lemma 16.7 and the triangle inequality that the set of good points
is open. If {zn } stays with K hyperbolic units of 0, then {zn } stays
within K + 1 units of 0. Here we have set zn = Fn−1 (w ). Similarly, it
follows from Lemma 16.7 and the triangle inequality that the set of
bad points is open. Finally, 0 is good. So, the set of good points is
open, closed, and nonempty. Hence every point in D is good. Hence
f (Δ) = D.

16.4. The Uniformization Theorem

Here is the Poincaré Uniformization Theorem.
Theorem 16.8 (Poincaré Uniformization). Suppose that A is a sim-
ply connected Riemann surface. Then one of three things is true:
• A is compact, and there is a biholomorphism between A and
the Riemann sphere.
• A is noncompact and there is a biholomorphism between A
and C.
• A is noncompact and there is a biholomorphism between A
and the open unit disk.
202 16. Riemann Surfaces and Uniformization

Note that the Poincaré Uniformization Theorem contains the Rie-

mann Mapping Theorem as a special case, when A is a Jordan do-
main. The main difference between the two results is that, in the
Uniformization Theorem, A is not assumed to be a subset of C.

16.5. The Small Picard Theorem

For the rest of the chapter, we deduce some nice consequences of the
Uniformization Theorem.
Lemma 16.9. There is a complex analytic covering map from the
open unit disk to C − {0, 1}, the twice punctured plane.

Proof. The universal cover X of C − {0, 1} is a simply connected

Riemann surface. Let E : X → C − {0, 1} be the covering map. If X
is compact, then E(X) is also compact, since the image of a compact
set under a continuous map is compact. But E(X) = C−{0, 1}, which
is noncompact. So, X is noncompact. If there is a biholomorphism
between X and C, then we have a complex analytic cover C → C −
{0, 1}. However, the fundamental group of C − {0, 1} is non-Abelian.
This is a contradiction. We have only one alternative left in the
Uniformization Theorem, and so there is a biholomorphism h between
X and the open unit disk. But then E ◦ h−1 is the desired complex
analytic covering map between the open unit disk and C − {0, 1}. 

Remark. In the concrete setting just discussed, it is possible to

prove Lemma 16.9 directly, without appealing to the Uniformization
Theorem. This is done in [AHL].

Lemma 16.9 is the main ingredient in the proof of the the follow-
ing result, which is known as the Small Picard Theorem:
Theorem 16.10. Let f : C → C be a nonconstant analytic map.
Then either f is onto or f omits exactly one value.

Proof. We will suppose that f omits at least two values and show
that f is constant. We can scale f so that two of the omitted values
are 0 and 1. Then f : C → C − {0, 1}. We have our holomorphic
covering from the open unit disk Δ to C − {0, 1}. But then we can
16.6. Implications for Compact Surfaces 203

find a lift f : C → Δ. This map is a bounded complex analytic

function, and hence constant. Hence f is constant as well. 

16.6. Implications for Compact Surfaces

The Uniformization Theorem is stated above in terms of simply con-
nected Riemann surfaces, but it has nice implications for general sur-
faces. Here is a the main consequence for compact surfaces.

Theorem 16.11. Let S be a compact and oriented Riemann surface.

• If S is homeomorphic to a sphere, then there is a biholomor-
phism between S and the Riemann sphere.
• If S is homeomorphic to the torus, then there is a biholo-
morphism between S and a flat torus.
• If S is a Riemann surface of negative Euler characteristic,
then there is a biholomorphism between S and some hyper-
bolic surface.

Proof. The sphere case is immediate from the Uniformization The-

orem.
Suppose that S is not homeomorphic to a torus. Let S be the
universal cover of S. Note that S is a simply connected Riemann
surface. According to the Uniformization Theorem, there is either
a biholomorphism between S and C, or a biholomorphism between
S and Δ, the open unit disk. In the former case, we would have a
complex analytic covering map C → S. But S has non-Abelian fun-
damental group, so Theorem 16.3 rules out this possibility. Therefore,
we have a complex analytic covering Δ → S where Δ is the unit disk.
Let G be the fundamental group of S. Then G acts on Δ as the deck
group. Each element g ∈ G is a biholomorphism of Δ. In Chapter 13
we proved that such maps are hyperbolic isometries. Hence G acts
on Δ as a group of hyperbolic isometries. S is precisely the quotient
of the hyperbolic plane by the orbit equivalence relation: Two points
are equivalent iff there is some element of G which maps one to the
other. Small neighborhoods of points in Δ contain unique members
of equivalence classes, and so these little disks map injectively into S.
The inverse maps give local coordinate charts into Δ, such that the
204 16. Riemann Surfaces and Uniformization

overlap functions are restrictions of hyperbolic isometries. In short,

S inherits its hyperbolic structure from Δ.
Suppose that S is homeomorphic to a torus. If there is a holo-
morphic covering Δ → S then the same argument as just given shows
that S is a hyperbolic surface and the fundamental group Z 2 acts on
Δ by hyperbolic isometries. This is only possible if all the elements
of Z 2 fix a common point on the unit circle. Such maps have the
following property: For any  > 0 there is some point x ∈ Δ which
is moved less than  (as measured in the hyperbolic metric). But
then S would have closed and homotopically nontrivial loop of length
less than . This contradicts the fact that all sufficiently short loops
on S are homotopically trivial. The contradiction shows that there
is no holomorphic cover from Δ to S. Only one alternative for the
Uniformization Theorem holds and so there is a holomorphic cover
C → S. But now the deck transformations are all Euclidean trans-
lations and S inherits a Euclidean structure from C just as in the
previous case. 

The above theorem is true in much more generality. For instance,

suppose that C ⊂ C is a finite set of N > 2 points. Then there is a
biholomorphism between C − C and a hyperbolic surface. The same
result holds if C is a countably infinite set of points, or the middle-
third Cantor set. It is hard to picture the universal cover of the
complement of the middle-third Cantor set, but the Uniformization
Theorem says that it is just the hyperbolic plane in disguise!
Part 4

Flat Cone Surfaces

Chapter 17

Flat Cone Surfaces

In this chapter we revisit the idea of gluing polygons together to form

a surface. In a sense, we return to the question taking the most
naive point of view possible. We keep the Euclidean geometry of the
component pieces and see what we get when they are glued together.
This point of view leads to the definition of a flat cone surface.
After we define flat cone surfaces, we will prove a fundamental
result about them, the combinatorial Gauss–Bonnet Theorem. The
combinatorial Gauss–Bonnet Theorem is am analogue the Gauss–
Bonnet Theorem from differential geometry; compare Theorem 12.4.
Following the proof of the combinatorial Gauss–Bonnet Theorem,
we give an application of flat cone surfaces to the study of polygonal
billiards. This is a theme that will take up both this chapter and the
next. All the material about billiards can be found, in much greater
detail in [MAT].

17.1. Sectors and Euclidean Cones

A sector in R2 is the closure of one of the 2 components of R2 −ρ1 −ρ2 ,
where ρ1 and ρ2 are two distinct rays emanating from the origin. For
example, the nonnegative quadrant is a sector. The angle of the sector
is defined as the angle between ρ1 and ρ2 as measured from inside the
sector. For instance, the angle of the nonnegative quadrant is π/2.

207
208 17. Flat Cone Surfaces

Two sectors in R2 can be glued together isometrically along one of

their edges. A Euclidean cone is a space obtained by gluing together,
in a cyclic pattern, a finite number of sectors. The angle of the
Euclidean cone is the sum of the angles of the sectors. The cone point
is the equivalence class of the origin(s) under the gluing. The cone
point is the only point which potentially does not have a neighborhood
locally isometric to R2 .
Note that two isometric Euclidean cones might have different de-
scriptions. For instance, R2 can be broken into 4 quadrants or 8
sectors of angle π/4.

Exercise 1. Prove that two Euclidean cones are isometric if and

only if they have the same angle.

Exercise 2. Define the unit circle in a Euclidean cone to be the

set of points which are 1 unit away from the cone point. On the cone
of angle 4π find the shortest path between every pair of points on
the unit circle. This problem breaks down into finitely many cases,
depending on where the points are located.

Exercise 3. Let C be a Euclidean cone, with cone point x. Say

that a vector field on C − x is locally constant if an isometry carrying
any open set of C − x into R2 carries the vector field to a constant
vector field. Prove that C − x has a parallel vector field in a neigh-
borhood of x if and only if the cone angle of C is a multiple of 2π.
(Hint: Unroll C into the plane and watch the vector field as you go
once around the cone point.)

17.2. Euclidean Cone Surfaces

We defined in §3.2 what it means for a surface to be oriented—it
does not contain any Möbius bands. For ease of exposition, we only
consider oriented surfaces.
Say that a compact oriented surface Σ is a Euclidean cone surface
if it has the following two properties:
17.3. The Gauss–Bonnet Theorem 209

• Every point p ∈ Σ has a neighborhood which is isometric

to a neighborhood of the cone point in a Euclidean cone of
angle θ(p).
• We have θ(p) = 2π for all but finitely many points.
The points p, where θ(p) = 2π, are called the cone points. The
quantity
δ(p) = 2π − θ(p)
is called the angle deficit. So, there are only finitely many points with
nonzero angle deficit, and these deficits could be positive or negative.
Here are two examples:
• Let P be a convex polyhedron in R3 . Then ∂P is a Eu-
clidean cone surface. The metric on ∂P is the intrinsic one:
the distance between two points is the length of the shortest
curve which remains on ∂P and joins the points.
• Let P1 , . . . , Pn be a finite union of polygons. Suppose that
these polygons can be glued together, isometrically along
their edges, so that the result is a surface. Then the sur-
face in question is a Euclidean cone surface if it is given its
intrinsic metric, i.e., the shortest path metric.
Amazingly, every example of type 2 is also an example of type 1
provided that the underlying surface is a sphere and all the angle
deficits are positive. This result is known as the Alexandrov Theorem.
(To make this strictly true we have to allow for the possibility that
P is contained in a plane in R3 .) One interesting open problem is to
determine the combinatorics of the convex polyhedron you get, based
on the intrinsic geometry of the cone surface.

17.3. The Gauss–Bonnet Theorem

Here is combinatorial version of the Gauss–Bonnet Theorem:
Theorem 17.1. If S is a compact cone surface, then

δ(p) = 4πχ(S).
p

Here the sum is taken over all angle deficits.

210 17. Flat Cone Surfaces

Proof. A Euclidean triangle on a Euclidean cone surface S is a re-

gion isometric to (you guessed it) a Euclidean triangle. For instance,
on the boundary of a tetrahedron, there are 4 obvious maximal Eu-
clidean triangles. Two triangles on a cone surface intersect normally
if they are either disjoint or share a vertex or share an edge. A tri-
angulation of S is a decomposition of S into finitely many triangles,
such that each pair of triangles intersects normally.

Exercise 4. Prove that every Euclidean cone surface has a trian-

gulation.

Choose a triangulation of S. Let T1 , . . . , TF be the list of triangles

in the triangulation. Each Ti has associated to it three angles ai , bi , ci ,
with ai +bi +ci = π. The cone points are all at vertices of the triangles,
and so
 
F 
F 
F
δ(p) = 2πV − ( ai + bi + ci ).
p i=1 i=1 i=1

In other words, we add up all the angles and see how the total sum
differs from the expected 2πV . Given that ai + bi + ci = π, we have

δ(p) = 2πV − πF = 2π(V − F/2) =∗ 2π(V + F − E) = 2πχ(S).
p

The starred equality has the following explanation. Each triangle

contributes 3/2 edges to the total number of edges. That is, E =
3F/2 = F + F/2. Hence −F/2 = F − E. 

For comparison, we mention that the differential geometric ver-

sion of the Gauss–Bonnet Theorem says that the total curvature of
a surface S is 2πχ(S), where χ is the Euler characteristic of S; see
§3.4. One can view the combinatorial Gauss–Bonnet Theorem as the
limit of the differential geometric version, in which all the curvature
is concentrated at finitely many points. At the same time, one can
view the differential geometric version as a limit of the combinatorial
version, in which the curvature gradually diffuses out, over larger and
larger finite sets of points so that it becomes continuously distributed.
17.4. Translation Surfaces 211

17.4. Translation Surfaces

A Euclidean cone surface is a translation surface if all the cone an-
gles are integer multiples of 2π. For instance, the octagon surface
discussed extensively in Chapter 1 is a translation surface when the
octagon is interpreted as a regular Euclidean octagon.

Theorem 17.2. Let S be a flat cone surface, and let C be a finite

list of points in S. Then S − C admits a parallel vector field if and
only if S is a translation surface.

Proof. Suppose first that such a vector field exists. Let x1 , . . . , xn be

the points of C. Let U1 , . . . , Un be disk neighborhoods of x1 , . . . , xn ,
respectively. Since Uk − xk admits a parallel vector field, the cone
angle at xk is an integer multiple of 2π. This is Exercise 3 above.
Now we prove the converse. Choose some basepoint x ∈ Σ − C.
Let v(x) be some unit vector tangent to x. Our goal is to define a
unit vector v(y) for each point y ∈ Σ − C. Here is the construction.
Let γ be any smooth curve which joins x to y and stays in Σ − C.
Say that a vector field along γ is parallel if, in the local coordinates,
the vectors are all translates of each other. Since every point of γ has
a neighborhood which is isometric to a disk in R2 , there is a unique
parallel vector field along γ which agrees with v(x) at x. We define
v(y) to be the vector of this parallel vector field at y. If this is really
well defined, then in small neighborhoods, our vector field consists
entirely of parallel vectors.
To finish our proof, we need to see that this definition is inde-
pendent of the path γ. If γ1 and γ2 are paths connecting x to y, and
are homotopic relative their endpoints, then we can produce a finite
sequence of paths γ1 = β1 , . . . , βn = γ2 such that βi and βi+1 agree
except in a region which is contained in a single Euclidean disk. (You
get the β curves just by doing the homotopy a little bit at a time.)
Within the Euclidean disk, you can see that the vector field along βi
must be parallel to the vector field along βi+1 , because both vector
fields just consist of a bunch of parallel vectors, and the two vector
fields agree at some point in the disk. Since this is true for all i, the
two methods for defining v(y) agree.
212 17. Flat Cone Surfaces

The fundamental group π1 (Σ − C) is generated by loops which

travel from x into a small neighborhood of one of the cone points,
wind around the cone point, and then come back. If γ1 and γ2 are
arbitrary paths joining x to y, then γ1 is homotopic relative to the
endpoints to δ1 ∗ · · · ∗ δk ∗ γ2 , where each δi is one of the special loops
just mentioned. Each loop δi starts and ends at x. We just have to
see that the parallel vector field along δi agrees with v(x) at both
ends. Everything boils down to what happens in a neighborhood of
the cone point.
By Exercise 3, we can define a parallel vector field in the neigh-
borhood of each point of C. Call these vector fields the “background
vector fields”. The parallel vector fields along our looks have constant
length and make constant angles with the relevant background vector
field. So, the parallel vector field along one of our loops comes exactly
back to itself when the loop is done. 

Recall that a gluing diagram for a surface is a list of finitely many

polygons, together with a recipe for gluing together the sides of the
polygon in pairs.

Lemma 17.3. Suppose that S is a flat cone surface obtained from a

gluing diagram in which the two sides in each glued pair are parallel.
Then S is a translation surface.

Proof. Once we show that S is orientable, we will know that S is

a cone surface. On each polygon, we consider the standard pair of
vector fields V1 and V2 . Here Vj consists of vectors parallel to the
basis vector ej . Given the nature of the gluing maps, the vector fields
piece together across the edges to give parallel vector fields V1 and V2
defined on the complement of finitely many points.
We first show that S is orientable. If S is not orientable, then
S contains a Möbius band M . By shrinking M if necessary, we can
arrange that M lies entirely in the region where both V1 and V2
are defined. But then we can define a continuous pair of linearly
independent vector fields on a Möbius band. This is easily seen to be
impossible. Hence S is oriented.
17.5. Billiards and Translation Surfaces 213

It now follows from Lemma 17.2 that S is a translation surface.



In light of Lemma 17.3, the surface obtained by gluing (with

translations) the opposite sides of a regular 2n-gon is a translation
surface.

Translation Principle. Whenever we consider gluing diagrams for

translation surfaces, in which more than one polygon is involved, we
always think of the polygons in the plane as being pairwise disjoint.
How the polygons sit in the plane is really not so important, in the
following sense. Suppose that P1 , . . . , Pn are the polygons involved in
a gluing diagram for some surface. Suppose that Q1 , . . . , Qn are new
polygons, such that Qk is a translation of Pk for all k, and the pat-
tern of gluing for the Q’s is the same as the pattern of gluing for the
P ’s. Then the two resulting surfaces are canonically isometric. The
canonical isometry is obtained by piecing together the translations
that carry each Pk to Qk . We mention this rather obvious principle
because it guarantees that certain constructions, which seem based
on arbitrary choices, are actually well defined independent of these
choices.

17.5. Billiards and Translation Surfaces

Let P be a Euclidean polygon. A billiard path in P is the motion
taken by an infinitesimal frictionless ball as it rolls around inside P ,
bouncing off the walls according to the laws of inelastic collisions:
the angle of incidence equals the angle of reflection; see Figure 17.1
below. We make a convention that a path stops if it lands precisely
at a vertex. (The infinitesimal ball falls into the infinitesimal pocket.)
The billiard path is periodic if it eventually repeats itself. Geo-
metrically, a periodic billiard path corresponds to a polygonal path
Q with the following properties:
• Q ⊂ P (that is, the solid planar region).
• The vertices of Q are contained in the interiors of the edges
of P .
214 17. Flat Cone Surfaces

• Q obeys the angle of incidence rule discussed above.

Figure 17.1. Polygonal billiards

Exercise 5. Find (with proof) all the examples of periodic billiard

paths in a square which do not have self-intersections. So, the path
Q has to be embedded.

The polygon P is called rational if all its angles are rational mul-
tiples of π. For instance, the equilateral triangle is a rational polygon.
In this section I will explain how to associate a translation surface
to a rational polygon. This is a classical construction, attributed by
some people to A. Katok and A.N. Zemylakov. The geometry of the
translation surface encodes many of the features of billiards in the
polygon.
For each edge e of P there is a reflection Re in the line through
the origin parallel to e. Like all reflections, Re has order 2. That is,
Re ◦ Re is the identity map. Let G be the group generated by the
elements R1 , . . . , Rn . Here Rj stands for Rej and e1 , . . . , en is the
complete list of edges. If ei and ej are parallel, then Ri = Rj . If P
is a rational polygon then there is some N such that ej is parallel to
some N th root of unity. But then G is a group of order at most 2N .
In particular, G is a finite group.
For each g ∈ G, we define a polygon

(17.1) Pg = g(Pg ) + Vg .
17.5. Billiards and Translation Surfaces 215

Here Vg is a vector included so that all the polygons {Pg | g ∈ G}

are disjoint. Thanks to the Translation Principle, the surface we will
produce is independent of the choices of the translation vectors.
To form a gluing diagram, we declare that every two edges of the
form

(17.2) e1 = g(e) + Vg , e2 = gr(e) + Vgr , r = Re .

are glued together by a translation. Here e is an arbitrary edge of P .

Since gr(e) = g(e), the edges e1 and e2 are parallel. Hence, it makes
sense to glue them by a translation. Note also that (gr)r = g. So,
our instructions tell us to glue e1 to e2 if and only if they tell us to
glue e2 to e1 . Let S be the space obtained from the gluing diagram.
Since the edges are glued in pairs, S is a surface. By Lemma 17.3, S
is a translation surface.
Here we work out the example where P is an isosceles triangle
with small angles 2π/8. In this case, the group G has order 16 and
our surface will be made from 16 isometric copies of P .

5 4
6 3

7 2

8 1

1 8

2 7

3 6
4 5

Figure 17.2. Gluing diagram for a translation surface

Figure 17.2 shows the resulting gluing diagram. We have chosen

the translations so that all the long sides have already been glued
together. Also, we have colored the triangles alternately light and
216 17. Flat Cone Surfaces

dark so as to better show the pattern. The numbers around the

outside of the figure indicate the gluing pattern for the short edges.
The gluing pattern in Figure 17.2 has an alternate description.
Take two regular Euclidean octagons and glue each side of one to the
opposite side of the other. The smaller inset picture in Figure 17.2
shows one of the two octagons. The other octagon is splayed open,
and made by gluing together the pieces that are outside the octagon
shown.
A path γ ∈ P is called straight if every point p ∈ γ has a neigh-
borhood U with the following property: any isometry between U and
a subset of R2 maps γ ∩ U to a straight line segment. (For concrete-
ness we can always take U to be a little Euclidean ball centered at
p.) There is an obvious map π : X → P . We just forget the group
element involved. This forgetting respects the way we have done the
gluing and so π is a well-defined continuous map from P to P . The
map π is somewhat like a covering map, except that it is not locally
a homeomorphism around points on the edges or vertices.

Lemma 17.4. Suppose γ  is a straight path on P which does not go

through any vertices of P. Then γ = π( γ ) is a billiard path on P .

Proof. By construction γ is a polygonal path whose only vertices

are contained in the interiors of edges of P . We just have to check
the perfect K condition at each vertex. You can see why this works
by building a physical model: Take a piece of paper and make a
crease in it by folding it in half (and then unfolding it.) Now draw
a straight line on the paper which crosses the crease. This straight
 which crosses an edge. When you fold
line corresponds to a piece of γ
the paper in half you see the straight line turn back at the crease and
form a perfect K. This folded path corresponds to γ. 

The converse is also true:

Lemma 17.5. Suppose that γ is a billiard path on P . Then there is

 on P such that π(
a straight path γ γ ) = γ.

Proof. We use the fact that the map π is almost a covering map.
Think of γ as a parametrized path γ : R → P , with γ(0) contained
17.6. Special Maps on a Translation Surface 217

in the interior of P . We define γ (0) to be the corresponding interior

point of (P, g), where g ∈ G is any initial element of G we like. We
can define γ (t) until the first value t1 > 0 such that γ(t1 ) lies on an
edge, say e1 , of P . But then we can define γ  in a neighborhood of t1
in such a way that γ (t1 − s) ∈ (P, g) and γ
(t + s) ∈ (P, e1 g) for s > 0
small. If you think about the folding construction described in the
previous lemma, you will see that the straight path γ(t1 − , t1 + )
(t1 − , t1 + ). Here  is some small value which depends
projects to γ
on the location of γ(t1 ). We can define γ  for t > t1 until we reach
the next time t2 such that γ(t2 ) lies in an edge of P . Then we repeat
the above construction for parameter values in a neighborhood of t2 .
And so on. This process continues indefinitely, and defines γ  for all
t ≥ 0. Now we go in the other direction and define γ  for all t < 0. 

Note that γ is a closed loop in P if and only if γ is a periodic

billiard path. Thus, the closed straight loops in P correspond, via π,
to periodic billiard paths in P .

Exercise 6. Suppose that P is the regular 7-gon. What is the Euler

characteristic of P? As a much harder problem, can you find a for-
mula for the Euler characteristic of P as a function of the angles of P ?

Exercise 7. The same construction can be made when P has some

irrational angles. What do you get if P is a right triangle with the
two small angles irrational multiples of π?

17.6. Special Maps on a Translation Surface

We would like to understand how straight lines move around on the
polygon P . Recall that P is the translation surface made by suitably
gluing together finitely many copies of P . We are going to prove a
dynamical result about the action of certain maps on P. The result
we prove holds in much greater generality, but to keep the discussion
self-contained, we are going to consider only one special map on P.
Choose some direction on P. Given x ∈ P, let f (x) be the point
you get by starting at x and moving for one unit in the given direction.
The map f is defined except at those points x whose corresponding
218 17. Flat Cone Surfaces

path hits a cone point. This means that f is defined except on a

set contained in a finite union of line segments. When f is defined
at x, a sufficiently small disk about x just moves forward in exactly
the same way that x does. This means that f is an isometry when
restricted to sufficiently small disks. On a large disk, which cuts
across the set where f is not defined, f is a piecewise isometry; it has
the effect of splitting the disk into several pieces and mapping each
piece isometrically to some place on P.
Given a set S ⊂ P, we define

n
(17.3) area(S) = area(S ∩ Pi ).
i=1

Here P1 , . . . , Pn are the polygons out of which P is made. This defini-

tion assumes that you know how to compute area inside the Euclidean
plane. For the kinds of complicated sets which arise in dynamical
systems, one actually needs some measure theory to give a rigorous
definition. In our setting here, we are just going to be computing the
areas of sets which are obtained by cutting a disk into finitely many
pieces along straight line segments.

Exercise 8. Let Δ be a disk. Let Δn ⊂ Δ denote the set of points

p ∈ Δ such that f k is well-defined on p for all k = 1, . . . , n. Note
that Δ − Δn is contained in a finite union of line segments. Define
f n (Δ) = f (Δn ). Prove that f n (Δ) and Δ have the same area. (Hint:
f n (Δ) is is obtained by translating small pieces of Δn isometrically
to various parts of P. These pieces do not overlap because f −1 exists
and has the same properties as f .)

Theorem 17.6. Let p ∈ P be any point on which f and all its iterates
are defined, and let  > 0 be arbitrary. Then there is some q ∈ P and
some n such that d(p, q) <  and d(p, f n (q)) < .

Proof. Let Δ be the disk of radius  about p. Let D0 = Δ and let

Dn = F n (Δ). By Exercise 8, the sets D0 , D1 , D2 , . . . all have the
same area. Since P has finite area, these sets cannot all be disjoint
from each other. Hence there are two sets Da and Db , which intersect
at some point xa . We take a < b. But then Da−1 and Db−1 intersect
17.7. Existence of Periodic Billiard Paths 219

at xa−1 = f −1 (xa ). Continuing in this way, we see that D0 and Db−a

intersect at some point x0 . By construction x0 lies within  of p and
f b−a (x0 ) also lies within  of p. 

Theorem 17.6 works more generally when all we know is that f

is an area preserving map of P that is defined except on a set of zero
area (or, more technically zero measure). The proof is essentially the
same, but one has to deal more carefully with the concept of area.
Even for area preserving maps, Theorem 17.6 is a toy version
of a much stronger and more general result known as the Poincaré
Recurrence Theorem.

17.7. Existence of Periodic Billiard Paths

It is a theorem of Howie Masur that every rational polygon has a
periodic billiard path. In fact, Masur gives bounds on the number of
such billiard paths of length at most L. He proves that there are at
least L2 /C − C of them, and at most CL + C of them, for some con-
stant C which depends on the polygon. In some cases, it is possible
to get sharper results. For instance:

Exercise 9. Prove that there is a constant C such that

lim N (L)/L2 = C,
L→∞

where N (L) is the number of periodic billiard paths of length less

than L on the unit square. What is C?

In this section, I will sketch an elementary proof, due to Bosher-

nitsyn, that every rational polygon has at least one periodic billiard
path. You will see that the proof actually gives the existence of many
periodic billiard paths, but no bounds like the ones mentioned above.
We choose a direction perpendicular to one of the sides of P and
let f : P → P be the function considered in the previous section. Let
p ∈ P be some point. We think of p as the lift of γ(0), where γ : R →
P is a billiard path which, at time 0, is travelling perpendicular to a
side of P . That is, γ is travelling parallel to V at time 0. By Theorem
17.6, there is some q very close to p and some n so that q = β(0) and
220 17. Flat Cone Surfaces


f n (q) = β(n) 
are very close together, and β(0) (0).
is very close to γ
 
Here β(0) is a straight path in P which goes through q at time 0.

If β(0) (0) are sufficiently close, then these two points are
and γ
on the same polygon of P. Hence β and γ are travelling in the same
direction at time 0. Likewise β is travelling in the same direction at
times 0 and n. In short, β travels perpendicular to a side of P at
time 0 and also at some much later time n. This means that β hits
the same side of P twice, and both times at right angles. But then β
is periodic. Each time it hits P perpendicularly, β just reverses itself
and retraces its path. Figure 17.3 shows an example of such a path.

Figure 17.3. A periodic billiard path

Chapter 18

Translation Surfaces and

the Veech Group

In the previous chapter we explained a construction which starts with

a rational polygon and produces a translation surface. The straight
line flow on the polygon controls the nature of billiards in the polygon.
In this chapter we will study the group of affine automorphisms of a
translation surface. This group is known as the Veech group.
It turns out that the Veech group can be interpreted as a group
of symmetries of the hyperbolic plane. So, starting with polygonal
billiards, we get back to hyperbolic geometry. We will work out a non-
trivial example of a Veech group at the end of the chapter. I learned
this particular example from Pat Hooper, and the presentation I give
is pretty close to the way he explained it to me.
A lot of the material in this chapter can be found in various
surveys of rational billiards; see, e.g., [MAT].

18.1. Affine Automorphisms

Recall that an affine map of R2 is a map of the form x → Ax + B,
where A is a 2 × 2 invertible and orientation-preserving matrix and
B is another vector. If B = 0, then the map is linear. Note that the
set of affine maps of R2 forms a group under composition.

221
222 18. Translation Surfaces and the Veech Group

Suppose that Σ is a translation surface. An affine automorphism

of Σ is a homeomorphism φ : Σ → Σ such that the following hold:
• φ permutes the nontrivial cone points of Σ.
• Every ordinary point of Σ has a neighborhood in which φ is
an affine map.
The second condition needs a bit more explanation. Let p ∈ Σ be an
ordinary point. This is to say that there is a small disk Δp about p
and an isometry Ip from Δp to a small disk in R2 . The same goes
for the point q = φ(p). The map Iq ◦ φ ◦ Ip−1 is defined on the open
set U = Ip (Δp ) ⊂ R2 and maps it to another open set Iq (Δq ) ⊂ R2 .
The second condition says that this map is the restriction of an affine
map to U .
We denote the set of all affine automorphisms of Σ as A(Σ). It
is easy to see that the composition of two affine automorphisms of
Σ is again an affine automorphism. Likewise, the inverse of an affine
automorphism of Σ is an affine automorphism of Σ. In short, A(Σ)
is a group.

Exercise 1. Let A be a 2 × 2 matrix with integer entries and deter-

minant 1. Let B any vector, and let Σ be the square torus. You can
think of Σ as (R/Z)2 . Let φ be the map φ([x]) = [Ax + B]. Prove
that φ is an affine automorphism of Σ. Thus, the square torus has a
huge affine automorphism group.

Exercise 2. Give an example of a translation surface which has

no nontrivial affine automorphisms.

Exercise 3 (Challenge). The affine automorphisms group of the

square torus is uncountable since it contains any translation. How-
ever, prove that the affine automorphism group of a surface with at
least one cone point is countable. (Hint: It suffices to consider the
subgroup G that preserves all the cone points. Try to show that this
subgroup is discrete, in the sense that any element of G sufficiently
close to the identity must actually be the identity. Draw many seg-
ments connecting all the cone points, and consider the action of an
element near the identity on these many line segments.)
18.2. The Diffential Representation 223

18.2. The Diffential Representation

Let SL2 (R) denote the group of 2×2 matrices having real entries and
determinant 1. Given a group A, a representation of A into SL2 (R)
is a homomorphism ρ : A → SL2 (R). Here is one explanation for
this terminology: The elements of A might be somehow abstract,
but a representation is a way of, well, representing these elements
concretely as matrices. A representation doesn’t have to be one-
to-one or onto, but of course representations with these additional
properties are especially nice.
Here we explain a canonical representation ρ : A(Σ) → SL2 (R).
The basic property of Σ we use is that there are canonical identifi-
cations between any pair of tangent planes Tp (Σ) and Tq (Σ), defined
as follows: By Theorem 17.2, there exists a parallel vector field on
Σ − C, where C is the set of cone points. Given p, q ∈ S − C, we can
find an isometry I from a neighborhood of p to a neighborhood of q
such that I(p) = q. If we insist that I preserves both the orientation
and the parallel field, then I is unique. Moreover, I is independent
of the choice of parallel field. The differential dI isometrically maps
Tp (Σ) to Tq (Σ). We set φpq = dI. So, in short

(18.1) φpq : Tp (Σ) → Tq (Σ)

is a canonical isometry. One immediate consequence of our definition

is that

(18.2) φpr = φqr ◦ φpq , φqp = φ−1

pq .

Now, given an element f ∈ A(Σ) we choose an ordinary point

p ∈ Σ, and let q = f (p). Let dfp be the differential of f at p. This
means that dfp is a linear map from Tp (Σ) to Tq (Σ). Note that the
composition
M (f, p) = φqp ◦ dfp

is a linear isomorphism from Tp (Σ) to itself. Using the isometry Ip ,

we can identify Tp (Σ) with, say, the tangent plane to R2 at the origin.
We let ρ(f ) be the linear transformation of R2 which corresponds to
M (f, p) under the identification.
224 18. Translation Surfaces and the Veech Group

We claim that ρ(f ) is independent of the choice of point p. To

see this, we note that the map ρ(f ) has the following alternate de-
scription. Using the coordinate charts Ip and Iq discussed above, the
map ρ(f ) is just the linear part of

dIq ◦ dfp ◦ dIp−1 .

The linear part of an affine map does not depend on the point. Hence
ρ(f ) has the same definition independent of which point we use inside
our local coordinate chart. But the surface is connected, so ρ(f ) does
not depend on the choice of point at all.
The determinant of ρ(f ) measures the factor by which f increases
area in a neighborhood of any point. Since the whole surface has
finite area and ρ(f ) is an automorphism, ρ(f ) must have determinant
1. Hence we can interpret ρ(f ) as an element of SL2 (R). The map
f → ρ(f ) is a homomorphism because of the chain rule: The linear
differential of a composition of maps is just the composition of the
linear differential of the invididual maps. And composition of linear
maps is the same thing as matrix multiplication in SL2 (R).
We have now constructed the representation ρ : A(Σ) → SL2 (R).
We let V (Σ) = ρ(A(Σ)). The matrix group V (Σ) is sometimes called
the Veech group. Below we will work out the Veech group associated
to the “double octagon” example discussed toward the end of §17.5.
Before we get to examples, however, we need to develop a bit more
of the theory.

18.3. Hyperbolic Group Actions

Recall that H 2 is the hyperbolic plane. We work in the upper half
plane model. Every element of SL2 (R) acts on H 2 isometrically, as
a linear fractional transformation; see §10.3. In particular, the Veech
group V acts on H 2 . The orbit of a point x ∈ H 2 is defined to be
the set
{g(x)| g ∈ V }.
We define an equivalence relation on points in H 2 by saying that two
points are equivalent iff they lie in the same orbit.
18.3. Hyperbolic Group Actions 225

V is said to act properly discontinuously on H 2 if, for every metric

ball B ⊂ H 2 , the set
{g ∈ V | g(B) ∩ B = ∅}
is a finite set. In other words, all but finitely elements of V have such
a drastic action on H 2 that they move the ball B completely off itself.

Exercise 4. Let SL2 (Z) be the group of 2 × 2 integer matrices

having determinant 1. Prove that SL2 (Z) acts properly discontinu-
ously on H 2 .

Before we establish the main result in this section, we give one

more definition. Two groups G1 , G2 ∈ SL2 (R) are conjugate if there
is some g ∈ SL2 (R) such that G2 = gG1 g −1 .

Exercise 5. Suppose that G1 and G2 are conjugate. Prove that

G1 acts properly discontinuously on H 2 if and only if G2 does.

Theorem 18.1. If V is the Veech group of a surface, then V acts

properly discontinuously on H 2 .

We will sketch the proof of Theorem 18.1 in the next section.

Whether or not V acts properly discontinuously, we can form the
quotient H 2 /V as follows. We define two points x, y ∈ H 2 to be
equivalent if there is some g ∈ V such that g(x) = y. Then H 2 /V
is defined to be the set of equivalence classes of points. In the case
where V acts properly discontinuously, the quotient is particularly
nice:

Theorem 18.2. If V acts properly discontinuously on H 2 , then we

can remove a countable discrete set of points T from H 2 such that
the quotient (H 2 − T )/V is a hyperbolic surface.

Proof. Before we start we note that all the elements of V act in an

orientation-preserving way, so that there are no reflections in V . (For
the orientation-reversing case, the statement of the result is slightly
different.)
226 18. Translation Surfaces and the Veech Group

Let T be the set of points x ∈ H 2 such that g(x) = x for some

nontrivial g ∈ V . The set T must be discrete in the sense that there
is some  > 0 such that any ball of radius  contains at most one point
of T . Otherwise we could find some ball B which contained infinitely
many points of T , and we would contradict the proper discontinuity.
Note that T is invariant under V : If x ∈ T is fixed by g, then y = h(x)
is fixed by hgh−1 . Thus, the quotient (H 2 −T )/V makes sense. Every
x ∈ H 2 − T has a neighborhood Δx such that g(Δx ) ∩ Δx = ∅ for
any nontrivial g. To see this, let dg denote the hyperbolic distance
between g(x) and x. Since x ∈ T , the number dg is positive. The
proper discontinuity prevents there being a sequence {gi } with {dgi }
converging to 0. Hence there is some positive lower bound to dg ,
which is what we need.
Now we know that each x ∈ H 2 − T has a little neighborhood
which is moved completely off itself by all of G (except the identity).
This little neighborhood therefore maps injectively into the quotient
(H 2 − T ) and serves as a coordinate chart about x. 

Note that the quotient H 2 /V still makes sense, and actually it

is obtained from (H 2 − T )/V just by adding finitely many points.
We define the covolume of V to be the volume of (H 2 − T )/V . The
group V is said to be a lattice if V has finite covolume. Σ is said to
be a Veech surface if V is a lattice. For instance, SL2 (Z) is a lattice.

18.4. Proof of Theorem 18.1

We first take care of a trivial case of Theorem 18.1.

Exercise 6. Suppose that Σ is a translation surface with no cone

points. Prove that Σ is isometric to a flat torus.

Exercise 7. Prove Theorem 18.1 in the case when the surface has
no cone points.

From now on, we consider the case when Σ has at least one cone
point. In this case, Σ is homeomorphic to a surface having negative
18.4. Proof of Theorem 18.1 227

Euler characteristic. Let C be the set of cone points of Σ. We call a

map γ : [0, 1] → Σ a saddle connection if the follwing hold.
• γ(t) ∈ C if and only if t = 0, 1.
• The restriction of γ to (0, 1) is locally a straight line.
Exercise 8. Prove that Σ has a pair of non-parallel saddle connec-
tions that intersect at a point of Σ − C.

Lemma 18.3. Let f be an affine automorphism of Σ. Let γ1 and γ2

be a pair of saddle connections, as in Exercise 8. Suppose f preserves
the endpoints of γ1 and γ2 , and f (γj ) = γj for j = 1, 2. Then f is in
the kernel of the differential representation ρ.

Proof. The restriction of an affine map to a straight line is just a

dilation. Hence, the restriction of f to γj is just a dilation. Since
f (γj ) = γj , the dilation factor must be one: the total length is pre-
served. So f is the identity on γj .
Let p be an intersection point of γ1 and γ2 . We know that
f (p) = p. Since γ1 and γ2 are nonparallel, we see that dfp fixes
two independent directions at p. Hence dfp is the identity. But then
ρ(f ) is the identity. 

We suppose that there is some ball B and an infinite collection

{gi } ∈ V such that gi (B) ∩ B = ∅. It is a general principle of com-
pactness that there must be elements of our set which are arbitrarily
close to each other. Hence, we can find an infinite list of distinct ele-
ments of V whose action on H 2 converges to the action of the identity
element.
What this means in terms of Σ is that we can find an infinite se-
quence {fj } of affine automorphisms such that ρ(fi ) is not the identity
but ρ(fi ) converges to the identity as i → ∞. All these elements per-
mute the set of cone points somehow. So, by taking suitable powers
of our elements, we can assume that each fi fixes each cone point of
Σ.
Let γ1 and γ2 be the saddle connections from Exercise 8. The
segment fk (γ1 ) is another saddle connection that connects the same
228 18. Translation Surfaces and the Veech Group

two cone points as does γ1 . For k large, fk (γ1 ) and γ1 nearly point in
the same direction and nearly have the same length. If they do not
point in exactly the same direction, they cannot connect the same
two endpoints. The two paths start out at the same cone point but
then slowly diverge, so that one of them misses the cone point at the
other end. Figure 18.1 shows what we mean.

Figure 18.1. Nearly parallel paths

This means that fk (γ1 ) and γ1 point in exactly the same direction
for k large. But then fk (γ1 ) = γ1 . The same argument shows that
fk (γ2 ) = γ2 for k large. But then, by the previous result, ρ(fk ) is the
identity for large k. This contradiction finishes the proof.

18.5. Triangle Groups

Figure 18.2. The hyperbolic triangle of interest

18.6. Linear and Hyperbolic Reflections 229

Recall that a geodesic hyperbolic triangle is a triangle in H 2

whose sides are either geodesic segments, geodesic rays, or geodesics.
The case of interest to us is the geodesic triangle with 2 ideal vertices
and one other vertex having interior angle 2π/8. Figure 18.2 shows
a picture of the triangle we mean, drawn in the disk model. This
triangle is known as the (8, ∞, ∞) triangle.

Lemma 18.4. Let γ be any geodesic in H 2 . Then there is an order

2 hyperbolic isometry which fixes γ.

Proof. Thinking of H 2 as the upper half-plane, the map z → −z

fixes the imaginary axis, which is a geodesic. We have already seen
that any two geodesics are isometric to each other. If g is an isometry
taking the geodesic γ1 to the geodesic γ2 and I is an order 2 isometry
fixing γ1 , then gIg −1 is the desired order 2 isometry fixing γ2 . Thus,
we can start with the one reflection desribed above and construct all
the others by conjugation. 

The order 2 hyperbolic isometry fixing γ is called a hyperbolic

reflection in γ. Given any geodesic triangle Δ, we can form the group
G(Δ) ⊂ SL2 (R) as follows. We let I1 , I2 , I3 be hyperbolic reflections
fixing the 3 sides of Δ and then we let G(Δ) be the group generated by
words of even length in I1 , I2 , I3 . For instance, I1 I2 and I1 I2 I1 I3 all
belong to G but I1 I2 I3 does not. All the elements in G are orientation
preserving and it turns out that we can find matrices in SL2 (R) for
the elements I1 I2 , I2 I3 , and I3 I1 . This is enough to show that G
actually comes from a subgroup of SL2 (R).

18.6. Linear and Hyperbolic Reflections

As preparation for the Veech group example we will work out, we
discuss how to convert between certain linear maps as they act on R2
and the corresponding linear fractional actions on H 2 .
Say that a linear reflection is a linear transformation T : R2 →
R such that T (v) = v and T (w) = −w for some basis {v, w} of R2 .
2

The corresponding linear fractional transformation acting on H 2 is

a hyperbolic reflection. This can be seen by considering the special
230 18. Translation Surfaces and the Veech Group

case when v = (1, 0) and w = (0, 1): all other cases are conjugate to
this one.
The map T is determined by the pair (v, w), but more than one
basis determines T . The basis (C1 v, C2 w) also determines T , where
C1 and C2 are any 2 nonzero constants. For this reason, it is really
the pair (L1 , L2 ) that determines T , where L1 is the line through v
and L2 is the line through w. The map T fixes L1 pointwise and
reverse L2 .
The map −T fixes L2 pointwise and reverses L1 . For this reason,
the unordered pair {L1 , L2 } determines the pair of maps {T, −T }.
The map ±T corresponds to a hyperbolic reflection, and each hy-
perbolic reflection corresponds to a pair ±T of maps. In short, each
hyperbolic reflection is determined by an unordered pair {L1 , L2 } of
lines through the origin. We call such a pair of lines a cross.
Let us first consider the case when L1 and L2 are perpendicular.
In this case, we call {L1 , L2 } a plus, because the two lines look like
a + symbol, up to rotation. If we work in the disk model Δ of the
hyperbolic plane, we can normalize so that the hyperbolic reflections
corresponding to pluses all fix some geodesic through the origin in C.
Figure 18.3 shows two examples. On the left-hand side of Figure 18.3
we show two pluses, one drawn thickly and one drawn thinly. On the
right hand side of Figure 18.3, we show the geodesics in Δ fixed by
the corresponding hyperbolic reflections.

Figure 18.3. Euclidean and hyperbolic reflections

18.6. Linear and Hyperbolic Reflections 231

Exercise 9. Let θ be the smallest angle between the lines of one plus
and the lines of another. Prove that the corresponding geodesics in
Δ meet at an angle of 2θ.

In light of Exercise 9, we can draw 3 crosses whose corresponding

geodesics in Δ are three sides of the (8, ∞, ∞) triangle shown in Figure
18.2. Two of the crosses are pluses and one is not. The crosses are
drawn thickly, and the thin lines are present for reference. The thin
lines are evenly spaced in the radial sense.

Figure 18.4. Three special crosses

Here is why this works. Let ±T1 , ±T2 , and ±T3 be the (pairs of)
hyperbolic reflections corresponding to each of the three crosses. Let
Rj be the hyperbolic reflection corresponding to ±Tj . By construc-
tion R1 and R2 each fix one of the “Euclideanly straight” sides of the
triangle in Figure 18.2.
We claim that R3 fixes the third side of the triangle in Figure 18.2.
The central point is that the third cross shares a line with each of the
first two crosses. If the signs are appropriately chosen, the element
T1 T3 is a parabolic element that fixes the vertical line through the
origin. From this we see that R1 R3 is a parabolic element fixing the
top vertex of our triangle. This is only possible if R3 fixes this same
point. A similar argument shows that R3 fixes the other ideal vertex
of our triangle. Hence, as claimed, R3 fixes the edge joining these
vertices.
We have gone through all this trouble because we want to rec-
ognize the (8, ∞, ∞) triangle group as a subgroup of the group of all
affine automorphisms of a certain translation surface. We will work
this out in the next section.
232 18. Translation Surfaces and the Veech Group

18.7. Behold, The Double Octagon!

We will compute the Veech group of the translation surface associated
to the Euclidean isosceles triangle having small angle 2π/8. As we
saw in §17.5, this surface is obtained from a gluing diagram involving
two regular Euclidean octagons. Each side of one octagon is glued to
the opposite side of the other. Let Σ be this surface.

Theorem 18.5. V (Σ) is the even subgroup of the (8, ∞, ∞) reflection

triangle group.

The (8, ∞, ∞) triangle group is the group generated by the three

hyperbolic reflections R1 , R2 , R3 considered in the previous section.
The even subgroup consists of elements made from composing an even
number of these elements. The even subgroup has index 2 in the whole
group. The point is that every element of the reflection triangle group
is either odd or even.
We will sketch a proof of Theorem 18.5. To make things work
well, we define an anti-affine automorphism to be a homeomorphism
of Σ which is locally anti-affine, meaning that the map locally has the
form x → L(x) + C, where L is an orientation-reversing linear map
and C is some constant vector. The linear reflections considered in
the previous section are of this form.

Let A(Σ) be the group of these maps, and let V = ρ(A),
 where
ρ is the differential representation as above. We will show that V
coincides with the group G  generated by the reflections in the sides
of the (8, ∞, ∞) triangle. The odd elements of A  are orientation
reversing and the even elements are orientation preserving. So, the
Veech group corresponds to the images of the even elements.

Figure 18.5. The first cross

18.7. Behold, The Double Octagon! 233

Figure 18.5 shows the octagons involved in the gluing diagram

for Σ. Again, each side of the left octagon is glued to the opposite
side of the right octagon by a translation. Simultaneous reflection in
the vertical sides of Σ induces an element T1 of A. The differential
of this map, evaluated at the center of the first octagon, fixes the
vertical line through the center and reverses the horizontal line. The
element ±dT1 therefore corresponds to the first plus in Figure 18.4.
Hence ρ(±T1 ) = R1 . Figure 18.6 does for R2 with Figure 18.5 does
for R1 . Here we take T2 to be simultaneous reflection in the diagonals
of positive slope.

Figure 18.6. The second cross

So far we have used fairly trivial symmetries of our surface. Now

we have to do something nontrivial to see the anti-affine automor-
phism that corresponds to the third cross. Figure 18.7 shows the
cross {L1 , L2 } we are aiming for, drawn on one of the octagons. The
auxilliary line L3 will be explained momentarily.

L3

L2 L1

Figure 18.7. The third cross

234 18. Translation Surfaces and the Veech Group

We will produce an automorphism g : Σ → Σ such that g fixes L2

pointwise and g(L1 ) = L3 in a length-preserving and height-reversing
way. That is, g maps the top vertex of L1 to the bottom vertex of
L3 and vice versa. At the same time, the map T2 fixes L2 pointwise
and maps L3 to L1 in a length-preserving way and height-preserving
way. But then the composition T3 = T2 ◦ g fixes L2 pointwise and
reverses L1 . By construction, the maps ±T3 correspond to our third
cross. We set R3 = ρ(±T3 ), and we have the desired map.

C D
A B
B A
D C

Figure 18.8. Cylinder decomposition

Now we turn our attention to the construction of the map g. Fig-

ure 18.8 shows a decomposition of Σ into 4 cylinders, labelled A, B,
C, D. Remember, each side of the left cylinder is glued to the oppo-
site side of the right cylinder. Thus, for instance, the two A pieces on
the left and right glue together to make the A cylinder. The A and B
cylinders are isometric to each other and the C and D cylinders are
isometric to each other. Here is the miracle that makes everything
work.

Exercise 10. Prove that the A and C cylinders are similar to each
other. Hence, all 4 cylinders are similar to each other.

For starters, we have g do the same thing on each octagon. Figure

18.9 shows how g acts on one of the octagons. g maps the points
labelled x to the points labelled y, in the manner suggested by the
arrows. These points are at the midpoints of the relevant edges.
18.7. Behold, The Double Octagon! 235

x
y

y
x
Figure 18.9. Action of the automorphism

Assume for the moment that there really is a locally affine au-
tomorphism of Σ that has this action. That is, assume that g really
exists. By construction g fixes L1 pointwise and g maps L1 to L3 in
a length-preserving and height-reversing way. The point is that L1
connects the two x points and L3 connects the two y points as shown
in Figure 18.9.
It only remains to show that g actually exists. First of all, we
define g in a neighborhood of the “centerline” L2 . We start extending
g outward until it is defined on the A cylinder. The lines connecting
the x points to the y points glue together to form the central loops of
the A and B cylinders. By construction g shifts these central loops
half way around. Hence g extends to be the identity on the boundary
of the A cylinder. Even though g is the identity on the boundary of
the A cylinder, g is not the identity on A: it is what is called a Dehn
twist. The same discussion works for the B cylinder.
Now we consider the C cylinder. So far, g is defined on one
boundary component of the C cylinder, and g is the identity on this
boundary component. Because the A and C cylinders are similar, g
extends to all of the C cylinder in such a way as to be the identity
on both boundary components. The action of g on the C cylinder
is the same, up to scaling, as the action of g on the A cylinder. A
similar thing works for the D cylinder. So, all in all, g is a Dehn
twist of each of the 4 cylinders, and the 4 separate maps fit together
236 18. Translation Surfaces and the Veech Group

seamlessly because g is the identity on every boundary component of

every cylinder. This establishes the existence of g.

Now we know that V (Σ) contains the (8, ∞, ∞) reflection trian-

gle group. Hence, the Veech group V (Σ) contains the even subgroup
of the (8, ∞, ∞) reflection triangle group. To finish our proof, we
will show that V (Σ) is precisely the reflection triangle group. Let
Y denote the (8, ∞, ∞) triangle. Let G  be the group generated by
hyperbolic reflections in the sides of Y .

Exercise 11 (Challenge). Suppose that Γ is a group acting prop-

erly discontinuously on H 2 and G  ⊂ Γ. Prove that either Γ = G 
or else Γ is the group generated by the reflections in the sides of the
geodesic triangle obtained by bisecting the Y in half.

If V does not equal G,

 then Σ has an extra isometric symme-
try which fixes the centers of the octagons. (This corresponds to the
extra element, reflection in the bisector of Y .) But the octagons do
not have any line of symmetry between the two drawn in our figures
above. Hence, this extra symmetry does not exist. Hence V (Σ) = G.

This is what we wanted to prove.

Exercise 12 (Challenge). Do all the same things as above for the

translation surface associated to the isosceles triangle having small
angles π/n for n = 4, 6, 8, . . . .
Part 5

The Totality of Surfaces

Chapter 19

Continued Fractions

The purpose of this chapter is to describe continued fractions and

their connection to hyperbolic geometry. One motivating factor for in-
cluding a chapter on continued fractions (besides their obvious beauty)
is that it gives us a nice way to introduce the modular group. The
modular group makes its appearance several times in subsequent
chapters. See the book [DAV] for an excellent treatment.

19.1. The Gauss Map

Given any x ∈ (0, 1) we define

(19.1) γ(x) = (1/x) − floor(1/x).

Here, floor(y) is the greatest integer less than or equal to y. The

Gauss map has a nice geometric interpretation, as shown in Figure
19.1. We start with a 1 × x rectangle, and remove as many x × x
squares as we can. Then we take the left over (shaded) rectangle
and turn it 90 degrees. The resulting rectangle is proportional to a
1 × γ(x) rectangle. Starting with x0 = x, we can form the sequence
x0 , x1 , x2 , . . . where xk+1 = γ(xk ). This sequence is defined until we
reach an index k for which xk = 0. Once xk = 0, the point xk+1 is
not defined.

239
240 19. Continued Fractions

1
Figure 19.1. Cutting down a rectangle

Exercise 1. Prove that the sequence {xk } terminates at a finite in-

dex if and only if x0 is rational.

Consider the rational case. We have a sequence x0 , . . . , xn , where

xn = 0. We define

(19.2) ak+1 = floor(1/xk ); k = 0, . . . , n − 1.

The numbers ak also have a geometric interpretation. Referring to

Figure 19.1, where x = xk , the number ak+1 tells us the number of
squares we can remove before we are left with the shaded rectangle. In
Figure 19.1, ak+1 = 2. Figure 19.2 shows a more extended example.
Starting with x0 = 7/24, we have the following.

• a1 = floor(24/7) = 3.
• x1 = 24/7 − 3 = 3/7.
• a2 = floor(7/3) = 2.
• x2 = (7/3) − 2 = 1/3.
• a3 = floor(3) = 3.
• x3 = 0.

In Figure 19.2 we can read off the sequence (a1 , a2 , a3 ) = (3, 2, 3) by

looking at the number of squares of each size in the picture. The
overall rectangle is 1 × x0 .
19.2. Continued Fractions 241

Figure 19.2. A 7/24 by 1 rectangle cut into squares

19.2. Continued Fractions

Again, sticking to the rational case, we can get an expression for x0
in terms of a1 , . . . , an . In general, we have
1
xk+1 = − ak+1 ,
xk
which leads to
1
(19.3) xk = .
ak+1 + xk+1
But then we can say that
1 1 1
(19.4) x0 = = = ··· .
a 1 + x1 1 1
a1 + a1 +
a 2 + x2 1
a2 +
a 3 + x3
We introduce the notation
(19.5)
1 1 1
α1 = , α2 = , α3 = ··· .
a1 1 1
a1 + a1 +
a2 1
a2 +
a3
In making these definitions, we are chopping off the xk in each expres-
sion in equation (19.4). The value of αk depends on k, but x0 = αn
because xn = 0.
Cosidering the example from the previous section, we have
1 1 2 7
α1 = , α2 = = , α3 = x0 = .
3 1 7 24
3+
2
242 19. Continued Fractions

We say that two rational numbers p1 /q1 and p2 /q2 are Farey
related if
 
p1 p2
(19.6) det = p1 q2 − p2 q1 = ±1.
q1 q2
In this case, we write p1 /q2 ∼ p2 /q2 . For instance 1/3 ∼ 2/7 and
2/7 ∼ 7/24. This is no accident.

Exercise 2. Starting with any rational x0 ∈ (0, 1) we get a sequence

{αk } as above. Prove that αk ∼ αk+1 for all k.

Exerxise 3. Consider the sequence of differences βk = αk+1 − αk .

Prove that the signs of βk alternate. Thus, the sequence α1 , α2 , α3 , . . .
alternately over-approximates and under-approximates x0 = αn .

Exercise 4. Prove that the denominator of αk+1 is greater than

the denominator of αk for all k. In particular, the α-sequence does
not repeat. With a little bit of extra effort, you can show that the
sequence of denominators grows at least exponentially.

19.3. The Farey Graph

Now we will switch gears and discuss an object in hyperbolic geom-
etry. Let H 2 denote the upper half-plane model of the hyperbolic
plane. We form a geodesic graph G in H 2 as follows. The vertices
of the graph are the rational points in R ∪ ∞, the ideal boundary of
H 2 . The point ∞ counts as rational, and is considered to be the frac-
tion 1/0. The edges of the graph are geodesics joining Farey related
rationals. For instance, the vertices
0 1 0
0= , 1= , ∞=
1 1 1
are the vertices of an ideal triangle T0 whose boundary lies in G.
Let Γ = SL2 (Z) denote the group of integer 2 × 2 matrices acting
on H 2 by linear fractional transformations. As usual, Γ also acts on
R ∪ ∞. The group Γ is known as the modular group.

Technical Remark. Before we launch into a discussion about Γ,

19.3. The Farey Graph 243

there is one technical point we need to clear up. The matrices A and
−A give rise to the same linear fractional transformation, so some-
times people introduce the notation P SL2 (Z) to denote the quotient
group SL2 (Z)/±, in which each element is an equivalence class con-
sisting of {A, −A}. This irritating distinction really plays no role in
our discussions, but you should keep in mind that a matrix is really
not quite the same thing as a linear fractional transformation, due
to the redundancy just mentioned. Nonetheless matrices represent
linear transformations.

Exercise 4. Let g ∈ Γ be some element. Suppose r1 ∼ r2 . Prove

that g(r1 ) ∼ g(r2 ). In particular, g is a symmetry of G.

Now we know that Γ acts as a group of symmetries of G. We can

say more. Suppose e is an edge of G, connecting p1 /q1 to p2 /q2 . The
matrix
 −1
p1 p2
q1 q2
carries e to the edge connecting 0 = 0/1 to ∞ = 1/0. We call this
latter edge our favorite. In other words, we can find a symmetry of G
that carries any edge we like to our favorite edge. Since Γ is a group,
we can find an element of Γ carrying any one edge e1 of G to any
other edge e2 . We just compose the element that carries e1 to our
favorite edge with the inverse of the element that carries e2 to our
favorite edge. In short Γ acts transitively on the edges of G.

Exercise 5. Prove that no two edges of G cross each other.

We have exhibited an ideal triangle T0 whose boundary lies in G.

Our favorite edge is an edge of this triangle. It is also an edge of the
ideal triangle T1 with vertices
0 1 −1
, , .
1 0 1
The boundary of this triangle lies in G as well. Thus, our favorite
edge is flanked by two ideal triangles whose boundaries lie in G. But
then, by symmetry, this holds for every edge of G. Starting out from
T0 and moving outward in a tree-like manner, we recognize that G is
244 19. Continued Fractions

the set of edges of a triangulation of H 2 by ideal triangles. Figure

19.3 shows a finite portion of G. The vertical line on the left is our
favorite line. The vertical line on the right connects 1 to ∞.

Figure 19.3. A portion of the Farey graph

19.4. Structure of the Modular Group

The Farey graph gives a good way to understand the structure of the
modular group. Since it is built out of ideal triangles, the Farey graph
has 3-fold symmetry built into it. The matrix
 
0 1
−1 1

represents the order 3 linear fractional transformation A ∈ Γ that

permutes the vertices of the ideal triangle T0 discussed above. More
precisely, A has the action

0 → 1 → ∞ → 0.

The Farey graph also has 2-fold symmetry. The matrix

 
0 1
−1 0
19.5. Continued Fractions and the Farey Graph 245

represents the order 2 linear fractional transformation B ∈ Γ that has

the action
0 → ∞, 1 → −1.
The element B is a rotation about the “midpoint” of the edge of the
Farey graph that joins 0 to ∞. Put another way, B swaps the triangle
T0 = (0, 1, ∞) with the adjacent triangle T  = (0, −1, ∞).

Exercise 6. Prove that the element BAB rotates the triangle T 

with vertices (0, −1, ∞).

Exercise 7. Some finite string of letters, just using A and B, is

called reduced if the strings AAA and BB do not occur in it. We
make this definition because the hyperbolic isometries A3 and B 2 are
the identity. Prove that any element of the modular group has the
form w(A, B) where w is a reduced word. (Hint: Show, just using
A’s and B’s, that you can move your favorite edge to any other edge
in two ways. This is exactly what the modular group can do.)

Exercise 8. Let w(A, B) be some nontrivial reduced work. Prove

that w(A, B) is a nontrivial element of the modular group.

19.5. Continued Fractions and the Farey Graph

Let’s go back to continued fractions and see how they fit in with the
Farey graph. Let x0 ∈ (0, 1) be a rational number. We have the
sequence of approximations α1 , . . . , αn = x0 as in equation (19.5). It
is convenient to also define
(19.7) α−1 = ∞, α−0 = 0.
If we consider the larger sequence α−1 , . . . , αn , the statements of Ex-
ercises 2 and 3 remain true. In particular, we have a path P (x0 ) in
the Farey graph that connects ∞ to x0 , obtained by connecting ∞ to
0 to α1 , etc. The example given above does not produce such a nice
picture, so we will give some other examples.
Let x0 = 5/8. This gives us
a1 = · · · = a5 = 1
246 19. Continued Fractions

and
1 2 3 5
α1 = 1, α2 = , α3 = , α4 = , α5 = x0 = .
2 3 5 8

Figure 19.4: The Farey path associated to 5/8

Taking x0 = 5/7 gives

a1 = 1, a2 = 2, a3 = 2.
and

2 5
α1 = 1, α2 = , α3 = .
3 7
There are three things we would like to point out about these
pictures. First, they make a zig-zag pattern. This always happens,
thanks to Exercises 3 and 4 above. Exercise 3 says that the path
cannot backtrack on itself, and then Exercise 4 forces the back-and-
forth behavior.
Second, we can read off the numbers a1 , . . . , an by looking at the
“amount of turning” the path makes at each vertex. In Figure 19.4,
our path turns “one click” at α0 , then “two clicks” at α1 , then “two
clicks” at α2 . This corresponds to the sequence (1, 2, 2). Similarly,
the path in Figure 19.3 turns “one click” at each vertex, and this
corresponds to the sequence (1, 1, 1, 1, 1).
19.6. The Irrational Case 247

Figure 19.4. The Farey path associated to 5/7

Exercise 9. Prove that the observation about the turns holds for
any rational x0 ∈ (0, 1).

Third, the diameter of the kth arc in our path is less than 1/k(k − 1).
This is a terrible estimate, but it will serve our purposes below. To
understand this estimate, note that the kth arc connects αk−1 =
pk−1 /qk−1 to αk = pk /qk , and αk−1 ∼ αk . The diameter of the kth
arc is
 
 pk−1 pk  ∗ 1 1

|αk−1 − αk | =  − = ≤ .
qk−1 qk qk−1 qk k(k − 1)
The starred equation comes from the fact that αk−1 and αk are Farey
related. The last inequality comes from Exercise 4. As we mentioned
in Exercise 4, the denominators of the α-sequence grow at least ex-
ponentially. So, actually, the arcs in our path shrink exponentially
fast.

19.6. The Irrational Case

So far, we have concentrated on the case when x0 is rational. If x0
is irrational, then we produce an infinite sequence {αk } of rational
248 19. Continued Fractions

numbers that approximate x. From what we have said above, we have

(19.8) x ∈ [αk , αk+1 ] or x ∈ [αk+1 , αk ]
for each index k, with the choice depending on the parity of k, and
also
(19.9) lim |αk − αk+1 | = 0.
k→∞

Therefore,
(19.10) x0 = lim αk .
k→∞

The corresponding infinite path in the Farey graph starts at ∞ and

zig-zags downward forever, limiting on x.
The nicest possible example is probably
√
5−1
x0 = = 1/φ,
2
where φ is the golden ratio. In this case, ak = 1 for all k and αk is
always ratio of two consecutive Fibonacci numbers. The path in this
case starts out as in Figure 19.3 and continues the pattern forever.
Taking some liberties with the notation, we can write
1 1
=
φ 1
1+
1
1+
1 + ··· .
Since φ = 1 + (1/φ) we can equally well write
1
(19.11) φ=1+
1
1+
1
1+
1 + ··· .
The {ak } sequence is known as the continued fraction expansion
of x0 . In case x0 > 1, we pad the sequence with floor(x0 ). So, 1/φ has
continued fraction expansion 1, 1, 1, . . . and φ has continued fraction
expansion 1, 1, 1, 1 . . . .
√
Exercise 10. Find the continued fraction expansion of k for k =
2, 3, 5, 7.
19.6. The Irrational Case 249

The subject of continued fractions is a vast one. Here are a few

basic facts:
• An irrational number x0 ∈ (0, 1) is the root of an integer
quadratic equation ax2 + bx + c = 0 if and only if it has
a continued fraction expansion that is eventually periodic.
[DAV] has a proof.
• The famous number e has continued fraction expansion
2; 1, 2, 1, 1, 4, 1, 1, 6, 1, 1, 8, 1, 1, 10 · · · .
• The continued fraction expansion of π is not known.
In spite of having a huge literature, the subject of continued fractions
abounds with unsolved problems. For instance, it is unknown whether
the {ak } sequence for the cube root of 2 is unbounded. In fact, this
is unknown for any root of an integer polynomial equation that is
neither quadratic irrational nor rational.
Chapter 20

Teichmüller Space and

Moduli Space

The purpose of this chapter is to introduce the notions of the Te-

ichmüller space and the moduli space of a closed surface. I will also
discuss the mapping class group, which is the group of symmetries
of Teichmüller space. The theory of these objects is vast and deep.
My purpose is just to introduce the basic objects in an intuitive way.
I learned most of the material here (at least in the negative Euler
characteristic case) from [THU]. The book [RAT] has a careful
treatment from a similar point of view. There are a number of more
advanced works devoted entirely to the topics introduced here. For
instance, see [GAR] and [FMA].
First we will deal with the case of tori, and then we will deal with
negative Euler characteristic case.

20.1. Parallelograms
Say that a marked parallelogram is a parallelogram P with a distin-
guished vertex v, a distinguished first side e1 , and a distinguished
second side e2 . The sides e1 and e2 should meet at v, as in Figure
20.1. We say that two marked parallelograms P1 and P2 are equiva-
lent if there is an orientation-preserving similarity, i.e., a translation

251
252 20. Teichmüller Space and Moduli Space

followed by a dilation followed by a rotation, that maps P1 to P2 and

preserves all the markings.

Figure 20.1. A marked parallelogram

We think of P as a subset of C. If we have a marked parallel-

ogram, we can translate it so that v = 0 and e1 points from 0 to 1.
Then e2 points from 0 to some z ∈ C − R. We only consider “half”
of the possibilities, the case when z ∈ H 2 , considered as the upper
half plane of C.

Exercise 1. Prove that z(P1 ) = z(P2 ) if and only if P1 and P2

are equivalent.

We can also reverse the process. Given z ∈ H 2 , we can form a

marked parallelogram P such that z(P ) = z. We simply choose the
parallelogram with vertices (0, 1, z, 1 + z) and mark it in the obvious
way. In short, we can say that there is a natural bijection between
the set T of marked parallelograms and H 2 , the upper half plane.
We can even take this one step further. H 2 has its usual hyperbolic
metric, and we can transfer this metric onto T . That is, the distance
between P1 and P2 is defined to be the hyperbolic distance between
z(P1 ) and z(P2 ).

20.2. Flat Tori

In §6.3 we discussed the surface one gets by gluing the opposite sides
of a parallelogram. Such a surface is known as a flat torus. One
thing we did not discuss in §6.3 is the effect of changing the shape of
the underlying parallelogram, i.e., the topic of the previous section.
20.2. Flat Tori 253

When we consider the idea of doing the construction in §6.3 for all
possible parallelograms, we are led to the notions of Teichmüller space
and moduli space. So, the idea is to unite the discussion in §6.3 with
the discussion in the previous section.
Say that a flat torus is a surface T that is locally Euclidean and
also homeomorphic to a torus. Recall that the universal cover of T is
R2 and the fundamental group of T is Z 2 .

Definition 20.1. Say that a marked flat torus is a flat torus, to-
gether with a distinguished pair of elements γ1 , γ2 ∈ π1 (T ) which
generate π1 (T ). We say that two marked tori T1 and T2 are equiva-
lent if there is an orientation-preserving similarity that carries T1 to
T2 and induces a map on the fundamental groups that carries the one
distinguished generating set to the other.

Given a marked flat torus T , we can produce a marked paral-

lelogram, as follows: We think of π1 (T, v) as the deck group, acting
on R2 by translations. We then consider the parallelogram P with
vertices

(20.1) 0, γ1 (0), γ2 (0), γ1 (0) + γ2 (0).

The distinguished vertex is 0, and the kth distinguished edge points

from 0 to γk (0). We insist that the marking of P is positively oriented
So, again, we weed out redundancy by only considering half the pos-
sibilities. We can also reverse the process. If we start with a marked
parallelogram P , as in Figure 20.1, we can glue the opposite sides of
P . The glued-up sides are loops which represent γ1 and γ2 .

Exercise 2. Prove that two marked tori are equivalent if and only if
the corresponding marked parallelograms are equivalent.

Now we will redo the same construction from another point of

view. Let Σ denote our favorite flat torus, say, the one obtained by
identifying the opposite sides of a square.

Definition 20.2. A marked flat torus is a triple (Σ, T, φ), where

254 20. Teichmüller Space and Moduli Space

T is a flat torus and φ : Σ → T is an orientation-preserving home-

omorphism. We say that two triples (Σ, T1 , φ1 ) and (Σ, T2 , φ1 ) are
equivalent if there is an orientation-preserving similarity f : T1 → T2
such that f ◦ φ1 and φ2 are homotopic maps.

We can convert between the one notion of marked torus and the
other. To go from Definition 20.2 to Definition 20.1, we first choose
a distinguished pair of elements of π1 (Σ): We let γ1 denote the loop
that is the glued-up horizontal edge, and we let γ2 be the loop that
is the glued-up vertical edge. Then φ∗ (γ1 ) and φ∗ (γ2 ) are the distin-
guished elements of π1 (T ). To go from Definition 20.1 to Definition
20.2, we recall that our original notion of a marked torus gives us
a description of the torus as a glued-up parallelogram P . We just
map the unit square to P by an affine map in such a way that the
gluings are respected, and then we interpret this map as a map from
Σ to our torus. This produces a triple (Σ, T, φ), where φ is not just a
homeomorphism but actually locally affine.

Exercise 3. Prove that our conversion between the two notions

of marked tori respects the equivalence relations. That is, the two
notions are really the same notion.

Since all the notions we have discussed are the same, we think
of T as the space of marked tori in the second sense. That is, we
work with equivalence classes of triples (Σ, T, φ). We have a canon-
ical identification of T with H 2 , the hyperbolic plane. With this
interpretation, T is known as the Teichmüller space of (marked) flat
structures on the torus.

20.3. The Modular Group Again

Now we are going to bring the modular group into the picture. We
discussed this group in §19.3 and §19.4. First of all, we can interpret
our favorite flat torus Σ as the quotient R2 /Z 2 . This observation is
important in what we do next.
Let G = SL2 (Z) denote the group of integer 2 × 2 matrices of
determinant 1, the modular group. Any g ∈ G acts on R2 as an
20.3. The Modular Group Again 255

orientation-preserving linear transformation, and g(Z 2 ) = Z 2 . This

means that g induces an orientation-preserving homeomorphism g :
Σ → Σ. We give this homeomorphism the same name as the linear
transformation which induces it.
Given a triple (Σ, T, φ), we define the new triple (Σ, T, φ ◦ g −1 ).
That is, we keep the same surface T , but we change φ : Σ → T to
the map given by the composition Σ → Σ → T , with the first arrow
given by g −1 . We use g −1 in place of the more obvious choice of g for
technical reasons that we will explain momentarily.

Exercise 4. Prove that (Σ, T1 , φ1 ) and (Σ, T2 , φ2 ) are equivalent

if and only if (Σ, T1 , φ1 ◦ g) and (Σ, T2 , φ2 ◦ g) are equivalent.

The group G acts on the space T in the sense that

(20.2) g1 (g2 (x)) = (g1 ◦ g2 )(x),

for all g1 , g2 ∈ G and all x ∈ T . Here g1 ◦ g2 means “first do g2 and

then do g1 ”. To see this, let x be a point represented by the triple
(Σ, T, φ). We compute

g1 (g2 (x)) = g1 (Σ, T, φ ◦ g2−1 )(Σ, T, φ ◦ g2−1 ◦ g1−1 )

= (Σ, T, φ ◦ (g1 ◦ g2 )−1 ) = (g1 ◦ g2 )(x).

From this calculation, you can see why we used the inverse: it makes
the compositions come out the right way.
We have an explicit identification of T with H 2 , and we can see
how a particular matrix
 
a b
(20.3) g=
c d

acts on T in these coordinates. Let x = (Σ, T, φ) as above. We

put T in the best possible position, so that T is obtained by gluing
the opposite sides of the parallelogram (0, 1, z, 1 + z), and so that φ
is induced by the linear transformation carrying (1, 0) to (1, 0) and
(0, 1) to (x, y). Here z = x+iy. When we lift φ to the universal covers
of Σ and T , respectively, we get the same linear transformation. In
256 20. Teichmüller Space and Moduli Space

other words, the linear transformation

 
1 x
(20.4) φ =
0 y
induces the homeomorphism φ. The linear transformation
 
 −1 d − cx −b + ax
(20.5) φ◦g =
−cy ay

induces the homeomorphism φ ◦ g −1 .

To figure out the point g(x) we just have to compute the shape
of the marked parallelogram g(T0 ), where T0 is the unit square whose
distinguished point is the origin, whose first distinguished edge is
1 ≡ (1, 0), and whose second distinguished edge is i ≡ (0, 1). Here we
are listing both the coordinates in C and in R2 . We compute that φ
maps the first and second edges, respectively, to cz − d and az − b.
So, if x ∈ T corresponds to z ∈ H 2 , then g(z) corresponds to
az − b
(20.6) .
cz − d
Except for the minus signs, this is the usual linear fractional action
of g on H 2 .

20.4. Moduli Space

The quotient M = T /G is known as moduli space. To interpret this
quotient in terms of tori, we (temporarily) let M denote the set of
equivalence classes of flat tori. Here we say that two flat tori are
equivalent if there is an orientation-preserving similarity carrying one
to the other. We are going to construct a natural bijection between
M and M . Once we have this bijection, we can forget about M
and simply realize that M is the space of equivalence classes of flat
tori.
There is an obvious map from T to M . Given a triple (Σ, T, φ),
we simply consider the torus T alone. It is a tautology that this map
respects the notions of equivalence on both T and M . The action
of G on T has no effect on the underlying torus—only the map is
changed—so actually we get a map from M to M .
20.4. Moduli Space 257

At the same time, there is a map from M to M. Given a flat

torus T , we arbitrarily choose a homeomorphism φ : Σ → T , and
we consider the image of the triple (Σ, T, φ) ∈ M. To see that this
map is independent of choices, consider the two triples (Σ, T, φ1 ) and
(Σ, T, φ2 ). The map φ−1
2 ◦ φ1 is an orientation-preserving homeomor-
phism of Σ. We claim that this map is homotopic to some linear
homeomorphism of Σ, given by g ∈ G. Assuming this claim for the
moment, we see that (Σ, T, φ2 ) = g(Σ, T, φ1 ) for some g ∈ G. That
is, both choices lead to the same point in M.
Assuming our claim, we see that there is a natural bijection be-
tween M and M , so we may identify M as the space of equivalence
classes of flat tori. The space M is known as moduli space.
It is worth pointing out that we just considered M as a set, but
we might have put a metric on M in some way. Any reasonable choice
would make our bijection between M and M a homeomorphism. We
will refrain from doing this here, because below we will actually do it
for surfaces of negative Euler characteristic.
Our only piece of unfinished business is to show that γ = φ−1
2 ◦ φ1
is homotopic to the action of some g ∈ SL2 (Z). Note that γ acts on
the fundamental group π1 (Σ). Since γ is an orientation-preserving
homeomorphism, γ has the same action on π1 (Σ) as does some g ∈
SL2 (Z). So, replacing γ by γ ◦ g −1 , we can assume that γ acts as
the identity on π1 (Σ). Our task now is to show that γ is homotopic
to the identity map.
A formal proof of this fact is a bit tedious, but we will sketch
the idea. Let e1 and e2 be the usual horizontal and vertical loops on
Σ. Since γ(e1 ) is homotopic to e1 , we first adjust γ so that it is the
identity on e1 . Next, we adjust γ so that it is the identity on e1 ∪ e2 .
But now we can cut Σ open and interpret γ as a continuous map from
the unit square to itself which is the identity on the boundary. The
following exercise finishes the proof.

Exercise 5. Prove that a continuous map from the unit square to

itself, which is the identity on the boundary, is homotopic to the iden-
tity map.
258 20. Teichmüller Space and Moduli Space

There is an important lesson to take away from this section. The

space M has a more direct and simple definition than the space T .
However, it was useful to define T first and then realize M as a
quotient of T . We hope that this lesson motivates the definition
of the Teichmüller space of surfaces having a fixed negative Euler
characteristic.

20.5. Teichmüller Space

We would like to make all the same constructions that we made above
in the negative Euler characteristic case, but there is one fine point
we want to iron out. Above, we considered flat tori and similari-
ties between them, and below we will consider hyperbolic surfaces
and isometries between. Given any flat torus, we can always rescale
the metric so that it has unit area. If we only work with unit area
tori, then the natural maps between them are (orientation-preserving)
isometries. The point here is that an area-preserving and orientation-
preserving similarity is an isometry. So, we might have redone the
whole theory above using unit area tori and isometries. This point of
view is more natural in the negative Euler characteristic case, because
two hyperbolic surfaces with the same topology always have the same
area; see Theorem 12.4.
Now we are ready to go. We will fix a number g ≥ 2, the genus
of the surfaces we consider. Recall that the genus g of a surface S
satisfies the equation
(20.7) χ(S) = 2 − 2g.
Here χ is the Euler characteristic, as discussed in §3.4. Thus, a torus
has genus 1, and the octagon surface has genus 2. In general, a genus
g surface is a “g-holed torus” that is locally isometric to H 2 , the
hyperbolic plane. We are going to build T g , the Teichmüller space of
genus g hyperbolic surfaces.
We first fix our favorite surface of genus g, and call it Σ. Unlike
in the torus case, a “favorite” does not immediately jump out. My
personal favorite is the one obtained by gluing together the opposite
sides of a regular hyperbolic 4g-gon. In any case, we look at triples of
the form (Σ, M, φ), where M is a hyperbolic surface of genus g and
20.5. Teichmüller Space 259

φ : Σ → M is a homeomorphism. We say that two triples (Σ, M1 , φ1 )

and (Σ, M2 , φ2 ) are equivalent if there is an isometry f : M1 → M2
such that f ◦ φ1 and φ2 are homotopic maps.
When we worked out the case of the torus, we had a natural
way of putting coordinates on the space T . The point is that T is
really just H 2 in disguise. This time, it is not so obvious what to
do. So, we will first make T g into a metric space. The idea behind
the next definition is to make sense of surfaces being nearby to each
other without quite being the same. We say that a homeomorphism
f : M1 → M2 is a (1 + )-isometry if the inequality

d2 (x2 , y2 )
(20.8) 1−≤ ≤1+
d1 (x1 , y2 )

holds for all quadruples x1 , y1 , x2 , y2 with x1 , y1 ∈ M1 and x2 = f (x1 )

and y2 = f (y1 ). The functions d1 and d2 are the metrics on M1 and
M2 , respectively.
Define the distance between 2 triples (Σ, M1 , φ1 ) and (Σ, M2 , φ2 )
to be the infimal  with the property that there is a (1 + ) isometry
f : M1 → M2 such that f ◦ φ1 and φ2 are homotopic maps.

Exercise 6. Prove that the equivalence relation we have defined

respects the distance we have defined. Hence, the distance between
two equivalence classes makes sense. This is how we make T g into a
metric space.

In the case of the torus, there was a perfectly canonical metric on

T , namely the hyperbolic metric. In the higher genus case, the metric
we have defined is pretty good but not perfectly canonical. There
are a number of canonical metrics on T g . The two most commonly
used are the Teichmüller metric and the Weil–Petersson metric. One
vexing thing is that these two common metrics are pretty different
from each other. So, while there are several nice ways to view T g ,
there does not seem to be one best way. What is best depends upon
the context.
260 20. Teichmüller Space and Moduli Space

20.6. The Mapping Class Group

When we worked with our favorite flat torus Σ, the one based on
the unit square, we saw that SL2 (Z) arose naturally as the group of
locally linear and orientation-preserving homeomorphisms of Σ. For
a hyperbolic surface, it is not immediately obvious that there is a
similarly natural group of homeomorphisms. However, it turns out
that there is such a group.
Above, we sketched a proof that any orientation-preserving home-
omorphism of the flat square torus Σ is homotopic to the action of
an element of SL2 (Z). In fact, we can equally well say that SL2 (Z)
is the quotient of the group of orientation-preserving similarities of
Σ, modulo homotopy. That is, two such homeomorphisms are equiv-
alent, and considered the same, if they are homotopic. This is a
definition that carries over immediately to the higher genus case.
We fix some initial hyperbolic surface Σ of genus g. The mapping
class group is defined as the group of equivalence classes of homeo-
morphisms of Σ, where two homeomorphisms are equivalent if they
are homotopic. The group is often denoted MCG g . It is a kind of
generalization of SL2 (Z). The definition of the mapping class group
depends only mildly on the choice of Σ. Any other choice would lead
to an isomorphic group.

Exercise 7. The mapping class group is certainly well defined as

a set. Prove that it is well defined as a group, that is, the group law
respects the equivalence classes.

People have focused quite a bit of attention on the mapping class

group in recent years. There are many open problems about this
group. One of the most well-known open problems is as follows.
For each genus g, does there exist some n = ng and a one-to-one
homomorphism φ : MCG g → GLn (C)? Here GLn (C) is the group of
complex valued n×n matrices having nonzero determinant. Or, more
briefly, is the mapping class group linear? . Subgroups of GLn (C) are
called linear.
20.6. The Mapping Class Group 261

The group MCG g acts on T g . The homeomorphism g : Σ → Σ

acts on the triple (M, Σ, φ) by sending it to the triple (M, Σ, φ ◦ g −1 ).
From the way we have defined the mapping class group, this action
respects the equivalence relation used to define T g .

Exercise 8. Recall that we defined a metric on T g above. Prove

that each element of MCG g acts as an isometry on T g .

Now that we have defined MCG g and T g , we define Mg to be

the quotient space.
(20.9) Mg = T g /MCG g .
The space Mg is known as the moduli space of genus g hyperbolic
surfaces.
As in the torus case, we could take the alternate route and define
Mg as the set of hyperbolic surfaces of genus g equipped with a metric
like the one defined above for T g . That is, the distance between two
hyperbolic surfaces is the infimal  such that there is a (1+) isometry
between them.
Chapter 21

Topology of Teichmüller
Space

We defined the Teichmüller space T g in the previous chapter. In

the case of the torus, the Teichmüller space is just a copy of H 2 ; in
particular, it is homeomorphic to R2 . Here will sketch a well-known
proof, mainly through a series of exercises, that T g is homeomorphic
to R6g−6 .
This beautiful result sets up a picture that is in some ways very
similar to the one for the torus. We have the mapping class group
MCG g acting isometrically on a space that is homeomorphic to R6g−g
(but having a funny metric on it). All the complexity in the topology
of the moduli space Mg comes from the operation of taking the quo-
tient of a topologically trivial space by the action of a complicated
group. This is exactly what happens for the torus, except that the
space M and the group SL2 (Z) are not so complicated.

21.1. Pairs of Pants

A pair of pants is a hyperbolic surface-with-boundary that is obtained
by taking two identical copies of a right-angled hexagon and gluing 3
of the sides. Figures 21.1 and 21.2 show this. These kinds of gluings
were considered in detail in Chapter 12.

263
264 21. Topology of Teichmüller Space

Figure 21.1. Hexagon gluing

Exercise 1. Let l1 , l2 , l3 be three positive numbers. Prove that there

is a right-angled hexagon whose “odd” sides have lengths l1 , l2 , l3 .
Prove that this hexagon is unique up to an isometry.

Figure 21.2. A pair of pants

A pair of pants is homeomorphic to a sphere with 3 holes. The

boundaries are totally geodesic in the sense that every point on the
boundary has a neighborhood that is isometric to a half-disk in H 2 .
By half-disk , we mean the portion of a disk that lies to one side of a
diameter.

Exercise 2. Suppose that M is a surface with boundary whose

interior is locally isometric to H 2 and whose boundary is totally ge-
odesic. Suppose also that M is homeomorphic to a sphere with 3
holes. Prove that M is a pair of pants, in the sense that M can be
21.2. Pants Decompositions 265

built by gluing together 2 hexagons in the manner described above.

(Hint: Consider the 3 geodesics arcs obtained by connecting the 3
boundary components, in pairwise fashion, by shortest curves. Cut
M open along these curves and watch M fall apart into two right-
angled hexagons. Use Exercise 1 to show that these hexagons are
copies of each other.)

Exercise 3. Let l1 , l2 , l3 be three positive real numbers. Prove that

there exists a unique pair of pants, up to isometry, whose boundaries
have lengths l1 , l2 , l3 , respectively.

To avoid confusion, we will shorten pair of pants to pant.

21.2. Pants Decompositions

We say that a pants decomposition of a hyperbolic surface is a real-
ization of that surface as a finite union of pants, glued together along
their boundaries. In this section we will prove that every hyperbolic
surface has a pants decomposition. Actually, we will prove that any
hyperbolic surface has many such.
Suppose that M is a hyperbolic surface and γ is a closed loop on
M . Recall that H 2 is the universal cover of M and that M = H 2 /Γ,
where Γ is the deck group. Let γ  denote a lift of γ to H 2 . Since γ
is a closed loop, there must be some nontrivial element g ∈ Γ such
that g(γ) = γ. According to the classification of isometries of the
hyperbolic plane, g is either elliptic, hyperbolic, or parabolic. Since
M is a compact surface, there is some  > 0 such that every  ball on
M is embedded. This means that g moves every point of H 2 by at
least . But this means that g is hyperbolic; see §10.9. In particular,
g has two fixed points on ∂H 2 .
From this picture, we see that γ  has two accumulation points
on ∂H 2 , namely the fixed points of g. There is a unique geodesic
β connecting the two fixed points of g. This geodesic is the axis of
g. The quotient β = β/g  is called the geodesic representative of γ.
Intuitively, if we think of γ as a rubber band that has been perhaps
stretched out of its natural position, then β represents the curve as-
sumed when the rubber band snaps back into position. The following
266 21. Topology of Teichmüller Space

exercise justifies this point of view.

Exercise 4. Prove that γ and β are homotopic.

The curve γ is called simple if γ has no self-intersections. If γ is

simple, then no two lifts of γ to the universal cover intersect. If β the
geodesic representative is not simple, then two lifts β1 and β2 in H 2
cross each other. But then the endpoints of β1 separate the endpoints
of β2 on ∂H 2 . But then we can find corresponding lifts γ 1 and γ2 of
γ whose (ideal) endpoints have the same property. This forces these
lifts to cross, which means that γ does have a self-intersection. Figure
21.3 shows the situation.

Figure 21.3. Intersecting curves

So, if γ is simple, then so is β. A similar argument proves the

following generalization: if {γi } is a finite list of pairwise disjoint
simple loops, then so is {βi }, the list of geodesic representatives.
The process of replacing a simple loop by its geodesic represen-
tative is a magical one. You might imagine that the geodesic replace-
ments could somehow crash into themselves or to each other, even
though the original curves do not. But, as we just explained, this is
21.3. Special Maps and Triples 267

not the case. So, if we want to find a pants decomposition of a hy-

perbolic surface M , all we have to do is find simple loops that divide
M into 3-holed spheres. Then we replace all these loops with their
geodesic representatives, and we have our pants decomposition.
Now we can give the intuitive idea behind the main result of this
chapter. We start by choosing our favorite pants decomposition of
our favorite surface Σ. Let S ⊂ Σ denote the set of curves of this
decomposition. By definition Σ − S is a union of 3-holed spheres. A
point of T g is an equivalence class of triple (Σ, M, φ). We define a
pants decomposition on M by taking the geodesic representatives of
the curves in the set φ(S) ⊂ M .

Exercise 5. Prove that there will be exactly 2g − 2 pants in the

decomposition, with 3g − 3 boundary curves.

We get 3g − 3 real numbers by considering the lengths of the

boundary curves in the pants decomposition. Each curve is contained
in 2 pants. We get an additional 3g − 3 numbers by considering how
two pants are glued together along their common boundary. These
other 3g − 3 numbers are usually called the twist parameters. Our
construction respects the equivalence relation on triples and gives a
well-defined map. We from T g to R6g−6 . will see that this map is a
homeomorphism.

21.3. Special Maps and Triples

In this section we prepare for our main construction. First, we choose
our favorite kind of pant. This is the pant obtained by gluing together
two identical regular hexagons. Each pant, including our favorite, has
two special points on each boundary component. These are the points
which are images of hexagon vertices; see Figure 21.2 above.
We (re)choose our favorite surface Σ so that it made from gluing
together 3g − 3 of our favorite kind of pants. We insist that the pants
are glued in such a way that the marked points are matched up. This
still doesn’t determine the surface exactly, but we just pick one from
amongst the various possibilities.
268 21. Topology of Teichmüller Space

Let P (l1 , l2 , l3 ) denote the pant having boundary lengths l1 , l2 , l3 .

We choose some reasonable homeomorphism from our favorite pant P0
to P (l1 , l2 , l3 ). The best way to to this is to choose a homeomorphism
from the regular right-angled hexagon to a hexagon with side lengths
lk /2 that sends vertices to vertices, and then to double this map, so
to speak.
More generally, given a 6-tuple (l1 , l2 , l3 , θ1 , θ2 , θ3 ), we choose a
map from P0 to P (l1 , l2 , l3 ) which agrees with our original map above,
but which makes a twist of θk /2 (say) clockwise in a small neighbor-
hood of the kth boundary component. This new map is obtained
from the original one by giving a kind of twist. The new map is very
similar to the Dehn twists we discussed in §18.7. We call such a map
special and denote it by μ(l1 , l2 , l3 , θ1 , θ2 , θ3 ).
Given a triple (Σ, M, φ), we get a pants decomposition of M , as
described above. We call the triple special if the following hold:
• The restriction of φ to each pant of Σ is one of our special
maps.
• If the restriction of φ to some pant P is the map
μ(l1 , l2 , l3 , θ1 , θ2 , θ3 ),
then the restriction of φ to the pant P  that meets P along
the kth boundary of P is μ(. . . lk . . . , . . . , θk , . . . ). The other
4 coordinates can be different.
Here we sketch the proof that any triple is equivalent to a special
triple. We warn the reader that a formal construction is filled with
many details that we are not including. We hope that this sketch is
sufficient for the interested reader to give a careful proof. [RAT] has
all the details.
Lemma 21.1. Any triple is equivalent to a special triple.

Proof (sketch). Starting with the triple (Σ, M, φ1 ), we make a

homotopy between φ1 and a homeomorphism φ2 : Σ → M that maps
the set S of geodesics to the set of geodesics on the natural pants
decomposition of M . Next, we make a homotopy between φ2 and
a map φ3 : Σ → M that agrees with a special map outside a small
neighborhood of S, the only place where we do not have control over
21.4. The End of the Proof 269

φ2 . This map agrees with the special map on the outside of 3g − 3

small annuli.
Each annulus on Σ is divided in half. The two halves of an annu-
lus are subsets of the two different pants that glue together along a
common boundary component. The center curve of the annulus is the
common boundary component. We have a foliation of each annulus
by circles that are, in a sense, parallel to the center curve. We have a
similar picture for the corresponding annulus on M . We just adjust
φ3 so that it twists these circles “at a constant rate”, evenly dividing
the total twist between the two halves, so to speak. The final map φ4
is homotopic to the original one, and is special on each pant. Thus
(Σ, M, φ4 ) is equivalent to the original triple (Σ, M, φ1 ).

21.4. The End of the Proof

Now we construct our map from T g to R6g−6 . We start with a triple
(Σ, M, φ) representing a point in T g . By Lemma 21.1, it suffices to
consider a special triple. However, for special triples, we can assign
a pair (l, θ) to each geodesic in the set S ⊂ Σ of pants boundaries.
This gives us the map from T g to R6g−6 . Call this map Φ.
The map Φ is surjective, essentially thanks to Exercise 3. We
can build pants with any boundary lengths we like, and then we can
glue them together with as much twisting as we like. The map Φ is
injective because the coordinates on S give us complete instructions
for how to assemble the surface M and the map φ : Σ → M . Hence
Φ is a bijection between T g and R6g−6 .
The map Φ−1 is continuous. If we have special maps correspond-
ing to nearly identical parameters, the corresponding pants are nearly
isometric to each other, and the twisting is nearly the same. This al-
lows us to build a map between the two surfaces that is nearly an
isometry and in the correct homotopy class.
Showing that Φ is continuous is the most tedious part of the
argument. Here we explain the proof. Let (M, Σj , φj ) for j = 1, 2
be two very nearby special triples. Suppose that f : M1 → M2 is a
(1 + )-isometry.
270 21. Topology of Teichmüller Space

If  is small, the map f carries each pant on M1 to a 3-holed sphere

on M2 whose boundaries are very nearly geodesics. The process of
replacing each near geodesic by an actual geodesic shortens the curve.
Hence, the pants boundaries on M2 are at most (1 + ) times as long
as their counterparts on M1 and vice versa. Hence, the length param-
eters {lk } labelling each curve in S are about the same for each triple.

Exercise 6. Let P1 be a pant on M1 , and let P2 be the corresponding

pant on M2 . Prove that f (∂P1 ) is contained in an  -neighborhood
of P2 , where  → 0 with . (Hint: Lift the picture to the univer-
sal cover, and show that a curve that nearly minimizes length in H 2
must be close to an actual geodesic.)

Exercise 7. Each pant on M1 decomposes into two identical right-

angled hexagons. Let H be such a hexagon. Prove that f (H) is
within  of one of the corresponding hexagons on M2 . Here  → 0 as
 → 0. (Hint: By Exercise 5, f maps the pant boundaries to curves
that are very close to their geodesic representatives. This takes care
of half the sides of H. Next, f maps each other side s of H to a curve
that nearly realizes the distance between two components of a pant
on M2 . Prove that this forces f (s) to be near the true minimizer.)

Each pant boundary β on M1 has 4 special points. Two of these

points come from one pant incident to β and the other two come from
the other pant incident to β. We call these collections of points special
quads.

Exercise 8. Let Q denote one of the special quads. Prove f (Q)

is within  of the corresponding special quad on M2 . Here  → 0 as
 → 0.

For M1 , the parameter θ/(2π) on each curve of S can be deduced

from the distances between the points of the relevant special quads.
We now conclude from Exercise 8 that the θ parameters for M1 are
close mod 2π to the corresponding θ parameters for M2 .
21.4. The End of the Proof 271

Finally, if the θ parameter for M1 differs by nearly an integer

from the corresponding θ parameter for M2 , then we can find an γ on
M1 , crossing over the image of the offending boundary component,
such that f (γ) twists more times around and is considerably longer,
as shown in Figure 21.4. Figure 21.4 shows the case when θ = 0 for
M1 but θ = 1 for M2 .

Figure 21.4. Integer twist

All in all, we have shown (modulo some details) that the map Φ
is a homeomorphism from T g to R6g−6 .
Part 6

Dessert
Chapter 22

The Banach–Tarski
Theorem

The purpose of this chapter is to prove the Banach–Tarski Theorem.

My account is somewhat similar to the one given in [WAG]. At first,
the Banach–Tarski Theorem does not look too much like a result
about surfaces, but in some sense it is a result about rotations of
the sphere. The proof I give also brings in the modular group in an
essential way.

22.1. The Result

We say that A, B ⊂ R3 are equivalent if there are finite partitions
into disjoint pieces,

A = A1 ∪ · · · ∪ An , B = B1 ∪ · · · ∪ Bn ,

and isometries I1 , . . . , In such that Ij (Aj ) = Bj for all j. In this case,

we write A ∼ B. When A ∼ B it means, informally, that one can
cut A into n pieces, like a puzzle, and reassemble those pieces into B.
The implied map A → B is, by definition, a piecewise isometric map.

Exercise 1. Prove that ∼ is an equivalence relation.

275
276 22. The Banach–Tarski Theorem

The Banach–Tarski Theorem requires the Axiom of Choice. See

[DEV] for a discussion of this axiom. Here is the precise version that
is needed.

Real Axiom of Choice (RAC). Let {Xα } be a disjoint union of

subsets of R3 . Then there exists a set S ⊂ Xα such that S contains
exactly one element of Xα for each α.

The RAC might seem obvious, or at least harmless. What the

Banach–Tarski Theorem shows is that a highly counterintuitive result
comes as a consequence of assuming that the RAC is true.
Say that A is a good set if A is bounded and A contains a ball.

Theorem 22.1 (Banach–Tarski). Assume the RAC. If A and B are

arbitrary good sets, then A ∼ B.

In light of the fact that ∼ is an equivalence relation, it suffices

to prove the Banach–Tarski Theorem in the case that A is a ball of
radius 1.
What makes this theorem amazing is that A could be a tiny ball
and B could be an enormous ball. At first you might think that this
result contradicts such physical properties as “conservation of mass”.
The usual reply to this objection is that the pieces needed to make
the puzzle are so complicated that they do not have mass. Another
reply is that we are not talking about physical objects that are made
out of atoms.

22.2. The Schroeder–Bernstein Theorem

The Schroeder–Bernstein Theorem says the following. If there are
injective maps from A into B and from B into A, then there is a
bijection between A and B. This result works for any sets and any
maps. (Any book on set theory has this result, but you can extract
the general proof from the proof of the next result.)
In case A and B are subsets of R3 and the injections are piecewise
isometric maps, then the bijection manufactured by the proof is also
piecewise isometric. To formalize this situation, we write A ≺ B if
22.2. The Schroeder–Bernstein Theorem 277

A ∼ B  for some subset B  ⊂ B. This is another way of saying that

there is a piecewise isometric injection from A into B.

Lemma 22.2. If A ≺ B and B ≺ A, then A ∼ B.

Proof. We have injective and piecewise isometric maps f : A → B

and g : B → A. Say that an n-chain is a sequence of the form
xn → · · · → x0 ∈ A, where

• xk ∈ A if k > 0 is even. In this case f (xk ) = xk−1

• xk ∈ B if k is odd. In this case g(xk ) = xk−1 .

For each a ∈ A, let n(a) denote the length of the longest n-chain that
ends in a = x0 . It might be that n(a) = ∞. Let An = {a ∈ A| n(a) =
n}. Swapping the roles of A and B, define Bn similarly.
Now observe the following:

• f (Ak ) = Bk+1 for k = 0, 2, 4, . . . .

• g −1 (Ak ) = Bk−1 for k = 1, 3, 5.
• f (A∞ ) = B∞ .

The restriction of f to

A = A0 ∪ A2 ∪ · · · ∪ A∞

is an injective piecewise isometry and the restriction of g −1 to

A = A − A = A1 ∪ A3 ∪ A5 · · ·

is also an injective piecewise isometry. (Note that A does not include
A∞ .) Define h(a) = f (a) if a ∈ A and h(a) = g −1 (a) if a ∈ A . By
construction
f (A ) ∩ g −1 (A ) = ∅.
Hence h is an injection. Also,

B = f (A ) ∪ g −1 (A ).

Hence h is a surjection. Hence h is a bijection. By construction h is

a piecewise isometric map. 
278 22. The Banach–Tarski Theorem

22.3. The Doubling Theorem

We write B  A if there is a partition B = B1 ∪· · ·∪Bn and isometries
I1 , . . . , In such that A ⊂ Ij (Bj ). In other words, we can break B
into finitely many pieces and use these pieces to cover A. The sets
I1 (B1 ), . . . , In (Bn ) need not be disjoint from each other.
Here is a result that sounds simpler (but not really much less
surprising) than the Banach–Tarski Theorem.

Theorem 22.3 (Doubling). Assume the RAC. Then there exist 3

disjoint unit balls A, B1 , B2 such that A  B, where B = B1 ∪ B2 .

Now we will reduce the Banach–Tarski Theorem to the Doubling

Theorem.

Lemma 22.4. If B  A, then A ≺ B.

Proof. Assume B  A. Define

• A1 = A ∩ I1 (B1 ).
• A2 = A ∩ I2 (B2 ) − A1 .
• A3 = A ∩ I3 (B3 ) − A1 − A2 , etc.
Then A = A1 ∪ · · · ∪ An is a partition. Let Bj = Ij−1 Aj and let
B  = Bj . Then B1 ∪ · · · ∪ Bn is a partition of B  . By construction
A ∼ B  ⊂ B. Hence A ≺ B. 

Let Br denote the unit ball of radius r centered at the origin.

We will assume the RAC and the Doubling Theorem. We claim
that Br ∼ Bs for any r, s > 0. By scaling, we can assume that
1 = r < s. Clearly, B1 ≺ Bs . In light of the lemmas in the previous
section, it suffices to prove that B1  Bs . There is some n such that
Bs can be covered by 2n translates of B1 . Iterating the Doubling
Theorem n times, we see that B1 is equivalent to 2n disjoint translates
of B1 . But then B1  Bs . This proves what we want. Now we know
that Br ∼ Bs for all r, s > 0.
We have already mentioned that it suffices to prove the Banach–
Tarski Theorem when B = B1 , the unit ball. But Br ⊂ A ⊂ Bs for
some pair of balls Br and Bs . Since Br ∼ Bs and A ⊂ Bs , we have
22.4. Depleted Balls 279

Br  A. This implies that A ≺ Br . But Br ≺ A. Hence A ∼ Br .

But Br ∼ B1 . Hence A ∼ B1 . This finishes the reduction of the
Banach–Tarski Theorem to the Doubling Theorem.

22.4. Depleted Balls

We are left to prove the Doubling Theorem. The Doubling Theorem
is about as simple as can be, but unfortunately some technical com-
plications arise when we try to prove the Doubling Theorem directly.
The way around these complications is to prove a related result in-
stead.
Say that a depleted ball is a set of the form B − X, where B is
a unit ball and X is a countable union of lines through the center of B.

Exercise 2. Prove that any unit ball can be covered by 3 isometric

images of any depleted ball.

Theorem 22.5 (Depleted Ball). Assume the RAC. Then there exists
a depleted ball Σ and a partition Σ = Σ1 ∪ Σ2 ∪ Σ3 such that the
following hold:

• Σi and Σj are isometric for all pairs i, j.

• Σ 3  Σ1 ∪ Σ2 .

Lemma 22.6. Assume the RAC. Then there are 9 disjoint depleted
balls A, B1 , . . . , B8 such that A ∼ B where B = B1 ∪ · · · ∪ B8 .

Proof. Iterating the conclusion of the Depleted Ball Theorem, we

see that Σ1  Y , where Y is any finite union of isometric copies of
Σ1 . Our lemma follows almost immediate from this. 

Exercise 3. Deduce the Doubling Theorem from the last lemma.

To finish the proof of the Banach–Tarski Theorem, we just have

to prove the Depleted Ball Theorem.
280 22. The Banach–Tarski Theorem

22.5. The Depleted Ball Theorem

Proving the Depleted Ball Theorem is the most interesting part of the
proof of the Banach–Tarski Theorem. The rest is really just “window
dressing”. This is the part of the proof that brings in the modular
group.
Consider the countable group
Γ = A, B|A3 = B 2 = identity.
In other words, Γ is the group of all words in A and B subject to the
relations that A3 and B 2 are the identity word.

Exercise 4. Prove that Γ is isomorphic, as a group, to the modular

group discussed in §19.3 and §19.4. (Hint: Put together Exercises 7
and 8 in §19.4.)

We have a partition Γ = Γ1 ∪ Γ2 ∪ Γ3 , where

• Γ1 consists of those words starting with A.
• Γ2 consists of those words starting with A2 .
• Γ3 consists of the empty word and also those words starting
with B.
We have the following structure:
AΓk = Γk+1 , Γ1 ∪ Γ2 ⊂ BΓ3 .
Indices are taken mod 3 for the first equation. These two algebraic
facts are quite close to the conclusion of the Depleted Ball Theorem.
The trick is to convert the algebra into geometry. Let B denote the
unit ball in R3 , and let SO(3) denote the group of rotations of B.
Below we will prove the following technical lemma.
Lemma 22.7. There exists an injective homomorphism ρ : Γ →
SO(3).

We choose our injective homomorphism, and we identify A and

B with their images under ρ. So, A is an order 3 rotation of B and
B is an order 2 rotation of B. In general, we identify elements of Γ
with their images under ρ.
22.5. The Depleted Ball Theorem 281

A nontrivial element of SO(3) is a rotation about some line

through the origin. We say that a line in R3 is bad if it is the line
fixed by some element of Γ. Since Γ is a countable group, there are
only countably many bad lines. Let X denote the union of these bad
lines, and let Σ = B − X. Then Σ is a depleted ball. Moreover, the
group Γ acts freely on Σ in the following sense. If γ(p) = p for some
γ ∈ Γ and some p ∈ Σ, then γ is the identity element.
We have an equivalence relation on Σ. We write p1 ∼ p2 if and
only if p1 = γ(p2 ) for some γ ∈ Γ. The fact that Γ is a group implies
easily that ∼ is an equivalence relation. This gives us an uncountable
partition

Σ= Σα
into the equivalence classes. By the RAC, we can find a new set
S ⊂ Σ such that S has one member in common with each Sα .

Lemma 22.8. Let γ1 , γ2 ∈ Γ be distinct elements. Then γ1 (S) ∩

γ2 (S) = ∅.

Proof. We argue by contradiction. Suppose that p ∈ γ1 (S) ∩ γ2 (S).

We have γj−1 (p) ∈ S for j = 1, 2. But γ1−1 (p) and γ2−1 (p) are in the
same Γ orbit. Since S intersects each Γ orbit exactly once, we have
γ1−1 (p) = γ2−1 (p). But then γ2 γ1−1 (p) = p. Since Γ acts freely on Σ, we
have γ2 γ1−1 = identity. Hence γ1 = γ2 . This is a contradiction. 

Lemma 22.9.

Σ= γ(S).
γ∈Γ

Proof. Choose p ∈ Σ. By construction, there is some q ∈ S such that

p ∼ q. This means that p = γ(q) for some γ ∈ Γ. Hence p ∈ γ(S). 

Now define

Σk = Γk (S) := (S).
γ∈Γk

The previous two results show that Σ = Σ1 ∪ Σ2 ∪ Σ3 is a partition

of Σ. At the same time
A(Σk ) = Σk+1 , B(Σ3 ) = BΓ3 (S) ⊃ (Γ1 ∪ Γ2 )(S) = Σ1 ∪ Σ2 .
282 22. The Banach–Tarski Theorem

The first part of this equation shows that Σi and Σj are isometric for
all i, j. The second part shows that Σ1 ∪ Σ2 is isometric to a subset of
Σ3 . Hence Σ3  (Σ1 ∪ Σ2 ). This proves the Depleted Ball Theorem.

22.6. The Injective Homomorphism

The last piece of unfinished business is to produce an injective ho-
momorphism ρ : Γ → SO(3). Let φ : S 2 → C ∪ ∞ be stereographic
projection, as in §9.5 and §14.3. We say that two points z, w ∈ C ∪ ∞
are partner points if
(22.1) w = −1/z.
In particular, 0 and ∞ are partner points.

Exercise 5. Prove that φ maps antipodal points on S 2 to part-

ner points.

Exercise 6. Let T1 and T2 be two linear fractional transformations,

both of which fix two distinct points z, w ∈ C. Suppose also that the
differentials dT1 and dT2 are the same map at z. Prove that T1 = T2 .
Lemma 22.10. Suppose that α is an order 3 linear fractional trans-
formation that fixes two partner points z and −1/z in C. Then the
map φ−1 ◦ α ◦ φ is an isometric rotation of S 2 .

Proof. Let α be as in Lemma 22.10. We know by Exercise 5 that

the map α = φ−1 ◦ α ◦ φ fixes two antipodal points and has order
3. We can find an isometry I  of S 2 that has order 3 and fixes these
same two points. Let I = φ ◦ I  ◦ φ−1 . By Lemma 14.6, the map I is a
linear fractional transformation. Note that I and α fix the same two
points and dI and dα have the same action at either point. Hence
I = α, by Exercise 6. Hence I  = α , as desired. This completes the
proof. 

Let SL2 (C) denote the group of 2 × 2 matrices, with complex

entries, having determinant 1. As in Chapter 10, these matrices rep-
resent linear fractional transformations.

Exercise 7. Let z, w ∈ C − {0} be distinct points. Prove that

22.6. The Injective Homomorphism 283

there exists an element Tz,w of SL2 (C) that represents a linear frac-
tional transformation that carries 0 to z and ∞ to w.

The matrix  
i 0
β=
0 −i
represents the linear fractional transformation that has order 2 rota-
tion about 0 and ∞. The matrix
 
ω 0 −1
αz,w = Tz,w ◦ ◦ Tz,w , ω = exp(πi/3),
0 ω5
represents an order 3 linear fractional transformation that fixes z and
w. The entries of αz,w are polynomials in z and w.
Once we pick z ∈ C − {0}, we define a homomorphism ρz : Γ →
SL2 (C) by the rule

ρz (A) = αz,w , w = 1/z, ρz (B) = β.

Note that ρz (β) doesn’t depend on z. Now we define ρ : Γ → SO(3)

by the rule
ρ = φ−1 ◦ ρz ◦ φ,
where φ is stereographic projection. Lemma 22.10 guarantees that
ρ(A) is an isometric rotation, and this is obvious for ρ(B). Note
that ρ is injective if and only if ρz is injective. So, at this point,
we can forget about ρ entirely and just worry about choosing z so
that ρz is injective. This is a problem entirely about linear fractional
transformations.
Given any γ ∈ Γ, let S(γ) ⊂ C denote those z ∈ C − {0} such
that ρz (γ) is not the identity matrix. Setting z = x + iy, we see
that the coordinates of w = 1/z are rational functions of x and y.
Therefore, the entries of ρz (γ) are rational functions of x and y. Any
rational function of x and y either vanishes identically or vanishes on
a nowhere dense set. In particular, S(γ) is either empty or else an
open dense set.

Exercise 8. The Baire Category Theorem (for the plane) says that
the intersection of a countable collection of open dense subsets of C
284 22. The Banach–Tarski Theorem

is nonempty. Prove this result.

Suppose for the moment that S(γ) is nonempty for all nontrivial
γ ∈ Γ. Then, by the Baire Category Theorem, the intersection

S(γ)
γ

is nonempty. Choosing any z in this intersection leads to an injective

ρz . So, to finish our proof, we just have to show that S(γ) is always
nonempty.
We will fix γ and show that S(γ) is nonempty. Recall that the
element αz,w is defined for all pair of distinct z, w ∈ C. Accordingly,
we can define the homomorphism ρz,w by sending A to αz,w and B
to β. This lets us speak about the matrix entries of ρz,w (γ). These
are polynomials on the two complex variables z and w. Let Fij be
one of these polynomials.

Lemma 22.11. Fij is nontrivial for some i, j.

Proof. Here is the crucial observation. We can choose (z, w) so that

the image ρ(Γ) is conjugate to the hyperbolic modular group discussed
in §19.3 and §20.3. For this choice of (z, w), the matrix ρz,w (γ) is
not the identity. The point is that the corresponding element in the
modular group does something nontrivial to the hyperbolic plane.
Hence, the matrix coefficients of this matrix, as functions of z and w,
cannot be constant. 

We let F = Fij for the indices guaranteed by the previous result.

Let
RΔ = {(z, −1/z)| z ∈ C − {0}}.
We really only care about the restriction of F to RΔ , because the
other points in C 2 do not correspond to homomorphisms from Γ into
SO(3).
Intuitively, what makes the next lemma work is that RΔ is a
“big” subset of C 2 . An algebraic geometer would say that RΔ is
Zariski dense, and that would be the end of the proof, but we will
work out what we need from scratch. For the interested reader, most
22.6. The Injective Homomorphism 285

books on algebraic geometry will have a discussion of Zariski Density.

See, for instance, [KEN].
Lemma 22.12. F is nonconstant on RΔ .

Proof. This is a general result about polynomials in C 2 and does

not depend on the specific structure of F . We will suppose that F
is constant in RΔ and derive a contradiction. Consider the following
rational map on C 2 :
 
θ(z1 , z2 ) = z1 + 1/z2 , i(z1 − 1/z2 ) .

By construction θ(RΔ ) is open in R2 . The function θ ◦ F ◦ θ −1 is

a rational function on C 2 that is constant on an open subset of R2 .
(A rational function is the ratio of two polynomials.) This forces
θ ◦ F ◦ θ −1 to be globally constant. But then F is globally constant
as well. This contradiction completes the proof. 

Since F is not constant on RΔ , the matrix ρz (γ) cannot be con-

stant on all of RΔ . Hence S(γ) is nonempty. This completes the last
piece of unfinished business. There is an injective homomorphism
ρ : Γ → SO(3).
Chapter 23

Dehn’s Dissection
Theorem

We saw in §8.5 that any two polygons of the same area are dissection
equivalent to each other. The purpose of this chapter is to prove
Dehn’s Dissection Theorem, which shows that the analogous result in
3 dimensions is false.

23.1. The Result

A polyhedron is a solid body whose boundary is a finite union of poly-
gons, called faces. We require that any two faces are either disjoint or
share a common edge or share a common vertex. Finally, we require
that any edge common to two faces is not common to any other face.
A dissection of a polyhedron P is a description of P as a finite
union of smaller polyhedra,

(23.1) P = P1 ∪ · · · ∪ Pn ,

such that the smaller polyhedra have pairwise disjoint interiors. Note
that there is not an additional assumption, say, that the smaller poly-
hedra meet face to face.
Two polyhedra P and Q are scissors congruent if there are dis-
sections P = P1 ∪ · · · ∪ Pn and Q = Q1 ∪ · · · ∪ Qn such that each

287
288 23. Dehn’s Dissection Theorem

Pk is isometric to Qk . Sometimes, one requires that all the isome-

tries are orientation-preserving, but in fact and two shapes that are
scissors congruent via general isometries are also scissors congruent
via orientation-preserving isometries. (This little fact isn’t something
that is important for our purposes.)
In 1900, David Hilbert posed 23 problems, now known as the
Hilbert Problems. Hilbert’s Third Problem asks if every two polyhe-
dra of the same volume are scissors congruent to each other. (Hilbert
conjectured that the answer was no.) The Hilbert Problems have in-
spired a huge amount of mathematics since 1900, but the third one
was solved in 1901, by Max Dehn. In 1901, Dehn proved the following
result.
Theorem 23.1. The cube and the regular tetrahedron (of the same
volume) are not scissors congruent.

Exercise 1. Say that a prism is a polyhedron with 5 faces, two of

which are parallel. So, a prism has a triangular cross-section. Prove
that any two prisms of the same volume are scissors congruent. (Hint:
After some effort you can boil this down to the the polygon dissection
theorem.)

23.2. Dihedral Angles

The dihedral angle is an angle we attach to an edge of a polyhedron.
To define this angle, we rotate the polyhedron so that the edge in
question is vertical, and then we look directly down on the polyhe-
dron. The two faces containing our edge appear as line segments, and
the dihedral angle is the angle between these line segments. We will
measure dihedral angles in such a way that a right angle has measure
1/4. All the dihedral angles of a cube are 1/4.
All edges of a regular tetrahedron have the same dihedral angle.
We are going to prove that this common angle is irrational. Geomet-
rically, this is the same as saying that one cannot fit finitely many
tetrahedra precisely around an edge, even if these tetrahedra are per-
mitted to wrap around more than once before closing back up.
We will place our tetrahedron in space so that one edge is vertical.
Rather than work in R3 , it is useful to work in C × R, where C is the
23.3. Irrationality Proof 289

complex plane. This is a nice way to distinguish the vertical direction.

Consider the complex number
√
1 2 2
(23.2) ω= + i.
3 3
Note that |ω| = 1. Let T0 be the tetrahedron with vertices
   
1 −1
(1, 0), (ω, 0), 0, √ , 0, √ .
3 3
Exercise 2. Check that T0 is a regular tetrahedron.

Consider the new tetrahedron Tn , with vertices

   
1 −1
n
(ω , 0), (ω n+1
, 0), 0, √ , 0, √ .
3 3
The tetrahedra T0 , T1 , T2 , . . . are just rotated copies of T0 . We are
rotating about the vertical axis. Notice that Tn+1 and Tn share a face
for every n. To say that the dihedral angle is irrational is the same
as saying that the list T0 , T1 , T2 , . . . is infinite. This is the same as
saying that there is no n such that ω n = 1.
In the next section, we will rule out the possibility that ω n = 1
for any positive integer n. This means that T0 , T1 , T2 . . . really is
an infinite list. Hence, the common dihedral angle associated to the
edges of a regular tetrahedron is irrational.

23.3. Irrationality Proof

The point of this section is to prove the following result: The complex
number
√
1 2 2
(23.3) ω= + i
3 3
does not satisfy the equation ω n = 1 for any positive integer n.

Exercise 3. Check that ω n = 1 for n = 1, 2, 3, 4, 5, 6. Also check

that ω 2 = (2/3)ω − 1.

In light of Exercise 2, we just have to check the case n ≥ 7. Let

G(ω) be the set of numbers of the form a + bω, where a and b are
290 23. Dehn’s Dissection Theorem

integers. This set is discrete: every disk intersects only finitely many
elements of G(ω). The point here is that ω is not real. So, considered
as vectors in the plane, 1 and ω are linearly independent over the
reals.
Let n ≥ 7 be the smallest value such that (supposedly) ω n = 1.
Let Z[ω] denote the set of numbers of the form

(23.4) a1 ω + a2 ω 2 + · · · + an ω n

where a1 , . . . , an are integers. Z[ω] has the nice property that

(23.5) (ω a − ω b )c ∈ Z[ω]

for any positive integers a, b, c. This comes from the fact that ω n = 1.
There are at least 7 powers of ω crowded on the unit circle, so at
least 2 of them must be closer than 1 unit apart. But that means
we can find integers a and b such that |z| < 1, where z = ω a − ω b .
The numbers z, z 2 , z 3 . . . all belong to Z[ω], and these numbers are
distinct because |z n+1 | = |z||z n | < |z n |. So, Z[ω] is not discrete.
Using Exercise 3, we get
!
ω 3 = ω × ω 2 = ω × (2/3)ω − 1 = (2/3)ω 2 − ω = (5/9)ω − (2/3),

and similarly for higher powers of ω. In general,

(23.6) 3n (a1 ω + · · · + an ω n ) = integer + integer × ω.

for any choice of integers a1 , . . . , an . But then Z[ω] is contained in

a scaled-down copy of G(ω) and hence is discrete. But Z[ω] is not
discrete, and we have a contradiction.

23.4. Rational Vector Spaces

Let R = {r1 , . . . , rn } be a finite list of real numbers. Let V be the
set of all numbers of the form

a 0 + a 1 r1 + · · · + a N rN , a0 , a1 , . . . , aN ∈ Q.

V is a finite-dimensional Q-vector space.

We declare two elements v1 , v2 ∈ V to be equivalent if v1 −v2 ∈ Q.
In this case we write v1 ∼ v2 . Let [v] denote the set of all elements
23.5. Dehn’s Invariant 291

of V that are equivalent to v. Let W denote the set of equivalence

classes of V . The two operations are given by
[v] + [w] = [v + w], r[v] = [rv].
The 0-element is given by [0].

Exercise 4. Prove that these definitions make sense, and turn W

into another finite-dimensional Q-vector space.

Let v1 , . . . , vm be a basis for V , and let w1 , . . . , wn be a basis for

W . The tensor product V ⊗ W is the Q-vector space of formal linear
combinations

(23.7) aij (vi ⊗ wj ), aij ∈ Q
i,j

Here vi ⊗ wj is just a formal symbol, but in a compatible way the

symbol ⊗ defines a bilinear map from V × W into V ⊗ W :
    
(23.8) ai vi ⊗ bj wj = ai bj (vi ⊗ wj ).

The m × n elements {1(vi ⊗ wj )} serve as as a basis for V ⊗ W .

Here is a basic property of V ⊗W . If v ∈ V is nonzero and w ∈ W
is nonzero, then v ⊗ w is nonzero. One sees this simply by writing v
and w out in a basis and considering equation (23.8). At least one
product ai bj will be nonzero. In particular
(23.9) 6 ⊗ δ = 0,
where δ is the dijedral angle of the regular tetrahedron, and R is
chosen so as to contain δ.

23.5. Dehn’s Invariant

Let R = {r1 , . . . , rN } be a finite list of real numbers, and let V and
W be the two examples of vector spaces given in Examples 1 and 2
above. Once again, V is the set of all numbers of the form
a 0 + a 1 r1 + · · · + a n rN , a0 , . . . , aN ∈ Q,
and W is the set of equivalence classes in V .
292 23. Dehn’s Dissection Theorem

Suppose that X is a polyhedron. Let λ1 , . . . , λk denote the side

lengths of all the edges of X. Let θ1 , . . . , θk be the dihedral angles,
listed in the same order. We say that X is adapted to R if

(23.10) λ1 , . . . , λk , θ1 , . . . , θk ∈ R.

If X is adapted to R, we define the Dehn invariant as


k
(23.11) X = (λi ⊗ [θi ]) ∈ V ⊗ W.
i=1

The operation ⊗ is as in equation (23.7), and the addition makes

sense because V ⊗ W is a vector space.
Suppose now that P and Q are a cube and a regular tetrahedron
having the same volume. Assume R is chosen large enough so that
P and Q are both adapted to R. Let λP and λQ denote the side
lengths of P and Q, respectively. Let δP and δQ denote the respective
dihedral angles. We have [δP ] = [1/4] = [0], because 1/4 is rational.
On the other hand, we have already seen that δQ is irrational. Hence
[δQ ] = [0]. This gives us

(23.12) P  = 12λP ⊗ [δP ] = [0], Q = 6λQ ⊗ [δQ ] = [0].

In particular,

(23.13) P  = Q.

To prove Dehn’s Theorem, our strategy is to show that the Dehn

invariant is the same for two polyhedra that are scissors congruent.
The result in the next section is the key step in this argument.

23.6. Clean Dissections

Say that a clean dissection of a polyhedron X is a dissection X =
X1 ∪ · · · ∪ XN , where each pair of polyhedra are either disjoint or
share precisely a lower-dimensional face. Let R be as above.

Lemma 23.2. Suppose that X = X1 ∪ · · · ∪ XN is a clean dissection

and all polyhedra are adapted to R. Then X = X1  + · · · + XN .
23.6. Clean Dissections 293

Proof. We will let Y stand for a typical polyhedron on our list. Say
that a flag is a pair (e, Y ), where e is an edge of Y . Then

X1  + · · · + XN  = S = λ(f ) ⊗ θ  (f ).
f ∈F

Here F is the set of all flags and λ(f ) and θ  (f ) are the length and
dihedral angle associated to the flag f .
We classify the flag (e, Y ) as one of three types:
• Type-1. e does not lie on the boundary of P .
• Type-2. e lies in the boundary of P , but not in an edge.
• Type-3. e lies in an edge of P .
We can write S = S1 + S2 + S3 , where Sj is the sum over flags of
Type j.
Call two flags (e, Y ) and (e , Y  ) strongly equivalent iff e = e .
Given a Type-1 edge e, let θ1 , . . . , θm denote the dijedral angles as-
sociated to the flags involving e. From the clean dissection property,
these polyhedra fit exactly around e, so that (with our special units)
θ1 + · · · + θm = 1. Hence
 
λ(e) ⊗ [θj ] = λ(e) ⊗ [θj ] = λ(e) ⊗ [1] = 0.
Summing over all Type-1 equivalence classes, we find that S1 = 0. A
similar argument shows that S2 = 0. In this case θ1 + · · · + θk = 1/2.
Now we show that S3 = X. Define a weak equivalence class as
follows. (e, P ) and (e , P  ) are weakly equivalent iff e and e lie in
the same edge of X. The set of weak equivalence classes is bijective
with the set of edges of X. Let e be some edge of X, with length
and dihedral angle λ and θ. Let e1 , . . . , em be the different edges
that appear in weak equivalence class named by e. With the obvious
notation λ = λ1 + · · · + λk . Let θj1 , . . . , θjmj denote the dihedral
angles associated to the strong equivalence class involving ej . We
have θj1 + · · · + θjmj = θ. Summing over the weak equivalence class,
we get  
λj ⊗ [θjk ] = λj ⊗ [θ] = λ(e) ⊗ [θ(e)].
jk j
Summing over all weak equivalence classes, we get S3 = X, as
desired. 
294 23. Dehn’s Dissection Theorem

23.7. The Proof

Let P be a cube, and let Q be a tetrahedron. We will suppose that
we have a scissors congruence between P = P1 ∪ · · · ∪ Pn and Q =
Q1 ∪ · · · ∪ Qn .
We first produce new dissections of P and Q that are clean. Here
is the construction. Let Π1 , . . . , Πk denote the union of all the planes
obtained by extending the faces of any polyhedron in the above dis-
section of P . Say that a chunk is the closure of a component of
R − Πj . Then we have clean dissections
(23.14) Pi = Pi1 ∪ · · · ∪ Pini
of each Pi into chunks, and also the clean dissection

(23.15) P = Pij
of P into chunks. We make all the same definitions for Q. The
dissections in equation (23.15) for P and Q might not define a scissors
congruence, but we don’t care.
Let R denote the finite list of lengths and dihedral angles that
arise in any of the polyhedra appearing in our constructions involving
P and Q. Let V ⊗ W be the vector space defined as in the previous
sections, relative to R. Computing the Dehn invariants in V ⊗ W , we
have
   
(23.16) P  = Pij  = Pi  = Qi  = Qij  = Q.
The first equality is obtained by applying Lemma 23.2 to the dissec-
tion in equation (23.15). The second equality is obtained by applying
Lemma 23.2 to each dissection in equation (23.14) and adding the
results. The middle equality comes from the obvious isometric invari-
ance of the Dehn invariant. The last two equalities have the same
explanations as the first two. In short, P  = Q. This contradicts
our computation that P  = Q. The only way out of the contradic-
tion is that the cube and the tetrahedron are not scissors congruent.

Exercise 5. Consider all the unit area platonic solids. Which are
scissors congruent to which?
Chapter 24

The Cauchy Rigidity

Theorem

The purpose of this chapter is to prove the Cauchy Rigidity Theorem

for strictly convex polyhedra. One can find another proof in the book
[AIZ]. As the authors point out therein, Cauchy’s original proof
was flawed, and a correct proof from comes from a letter from I.J.
Schoenberg to K. Zaremba.
The proof is half geometrical and half combinatorial. The geo-
metrical half of the proof I give is very similar to what Aigner and
Ziegler do, except that I spell out some of the intermediate steps
in more detail. The combinatorial half can be done in many ways,
and I give an argument based on the combinatorial Gauss–Bonnet
Theorem; see §17.3.

24.1. The Main Result

A polyhedron is a solid body whose boundary is a finite union of
polygons, called faces. A polyhedron P is called strictly convex if,
for each face f of P , there is a half-space Πf such P ⊂ Πf and
P ∩ ∂Πf = f . The boundary ∂Πf is the plane extending f . The cube
is a classic example of a strictly convex polyhedron.

295
296 24. The Cauchy Rigidity Theorem

Say that two polyhedra P and P  are flexes of each other if there
is a homeomorphism from ∂P to ∂P  which is an isometry when
restricted to each face. In other words, there is a combinatorics-
respecting bijection between the faces of P and the faces of P  such
that corresponding faces are isometric to each other. Making the
same definition for polygons, we observe that any pair of rhombuses,
having unit side length, are flexes of each other. The Cauchy Rigidity
Theorem rules out this behavior in 3 dimensions, at least for strictly
convex polyhedra.
Theorem 24.1 (Cauchy). Let P and P  be two strictly convex polyhe-
dra. If P and P  are flexes of each other then P and P  are isometric.
Exercise 1. Show by example that the Cauchy Rigidity Theorem is
false when the convexity assumption is dropped.
Amazingly, Robert Connelly discovered examples of continuous
families of polyhedra, in which every two are flexes of each other. In
other words, Connelly’s examples actually flex in a literal sense.

24.2. The Dual Graph

There is a nice graph that lies on the surface of P , called the dual
graph. We place one new vertex per face of P , and join two vertices
by an edge if and only if the corresponding faces share an edge.
There is a nice geometric way to picture the dual graph. Let S
denote a set of points, one per interior face of P . Just to be definite,
we choose the center of mass of each face of P . Then, let P ∗ denote
the convex polygon whose vertex set is S. Formally, we can say that
P ∗ is the convex hull of S, namely the intersection of all closed and
convex subsets of R3 that contain S. The dual graph is exactly the
union of edges and vertices of P ∗ .

Exercise 2. When P is a platonic solid, P ∗ is also a platonic

solid. The construction pairs up the cube with the octahedron, the
dodecahedron with the icosahedron, and the tetrahedron with itself,
or, rather, a slightly smaller tetrahedron. Try to draw pictures of
these cases.
24.3. Outline of the Proof 297

To get perhaps the nicest picture of the dual graph, we surround

P ∗ by a large sphere and then project the dual graph onto the surface
of the sphere by a radial projection from some point in the interior
of P ∗ . Finally, we identify this large sphere with S 2 , the unit sphere.
This gives us a graph Γ on S 2 , all of whose edges are arcs of circles.
Each component of S 2 − Γ is a polygon whose boundary is made from
circular arcs. The important thing for us is just that each component
is homeomorphic to a disk.
There is a natural correspondence between the edges of the poly-
gon and the edges of the dual graph. When we draw the dual graph
directly on P , each edge of the dual graph crosses one edge of P , and
vice versa. So, if we have some kind of labelling of the edges of P ,
we can transfer it in the obvious way to a labelling of the edges of
the dual graph Γ, which we think about in its final incarnation, as a
graph on S 2 .

24.3. Outline of the Proof

Each edge e of P has a partner edge e of P  . Let θ(e) be the dihedral
angle of P at e, and let θ(e ) be the corresponding dihedral angle of
P  at e . (Recall that the dihedral angle is the angle made by the
planes incident to the edge.) We label the edge e by (+), (−), or (0)
according to whether the sign of θ(e) − θ(e ) is positive, negative, or
zero.
We transfer our labelling to the dual graph, Γ ⊂ S 2 . Each com-
ponent C of S 2 − Γ is bounded by a circuit in Γ. We get a cyclically
ordered list L(C) of members of {+, −, 0} by reading the labels of
this circle, say, in clockwise order.
We call L = L(C) bad if, after we delete all the 0’s from L, we have
a nonempty list that changes from + to − at most once as we cycle
through it. Otherwise, we call L good . For instance (+0 + − − −00+)
is a bad list, and (+ + − − − + −) is a good list.
Below we will prove two results. The first is geometrical and the
second is combinatorial.
298 24. The Cauchy Rigidity Theorem

Lemma 24.2. For any component C of S 2 − Γ, the list L(C) is good.

Lemma 24.3. Let Γ be a graph on S 2 such that each component of

S 2 − Γ is an embedded topological disk. Suppose that the edges of Γ
are labelled nontrivially by elements of {+, −, 0}. Then there is at
least one component C of S 2 − Γ such that L(C) is a bad list.

Our two lemmas contradict each other unless the labelling of Γ

is completely trivial. But then θ(e) = θ(e ) for all edges e of P . But
this easily implies that P and P  are isometric.

Exercise 3. Build half an octahedron by taping together 4 card-

board equilateral triangles about a vertex. The portion of Γ corre-
sponding to these faces is a quadrilateral. Physically flex the object
and observe that the only possible nontrivial labelling is (+ − +−)
or, of course, (−+−+). Compare this with Exercise 5 from Chapter 9.

Exercise 4. Without looking at the long-winded proof below, prove

Lemma 24.3 for the cube.

Exercises 3 and 4 combine to prove Cauchy’s Theorem for P and

P  , when P is a regular octahedron.

Exercise 5. Imitating Exercises 3 and 4, Give a proof of Cauchy’s

Theorem for the regular icosahedron.

24.4. Proof of Lemma 24.3

Let P be a polygon whose edges are labelled (+) and (−). We say that
P has a good labelling if the list of labels around its edges is good; that
is, there at least 2 sign changes from (+) to (−). A quadrilateral with
a good labelling must be labelled (+, −, +, −), up to cyclic ordering.
24.4. Proof of Lemma 24.3 299

Figure 24.1. Adding edges and vertices

Figure 24.1 shows examples of how one can divide a polygon with
a good labelling into quadrilaterals with good labellings. In the sub-
division process, you are allowed to add both edges and vertices.

Exercise 6. Prove Lemma 24.4 below.

Lemma 24.4. Suppose that P is a polygon with a good labelling. Let
Pv denote the vertex set of P . We can partition P into alternately
labelled quadrilaterals, extending the labelling on P , such that the fol-
lowing is true:
• Let w be a vertex of a quadrtilateral that lies in the interior
of P . Then w is a vertex of at least 4 quadrilaterals.
• Let w be a vertex of a quadrilateral that lies in ∂P − Pv .
Then w is a vertex of 2 quadrilaterals.

First we prove Lemma 24.3 in the special case that all labels of
Γ are nonzero. By Lemma 24.4, we can partition each component of
S 2 − Γ into alternately labelled quadrilaterals. These partitions fit
together to partition S 2 itself into alternately labelled quadrilaterals.
Now we make 3 observations.
• Each quadrilateral vertex in the interior of a component of
S 2 − Γ is a vertex of 4 quadrilaterals.
• Each vertex in the interior of an edge is a vertex of exactly
4 quadrilaterals, two coming from each side.
• Each vertex of Γ is a vertex of at least 3 quadrilaterals, by
the valence condition. However, the edges emanating from a
300 24. The Cauchy Rigidity Theorem

vertex must alternate in sign, given that our quadrilaterals

are all alternately labelled. Hence, each vertex of Γ is a
vertex of at least 4 quadrilaterals.
In short, every quadrilateral vertex is the vertex of at least 4 quadri-
laterals.
Now we build a Euclidean cone surface based on our partition.
We glue together unit squares using the combinatorial pattern given
by our quadrilaterals. Call the resulting surface Σ. By construction,
the cone angle of Σ is at least 2π at each quadrilateral vertex. The
remaining points of Σ are locally Euclidean. Hence, the total com-
binatorial curvature of Σ is nonpositive. But Σ is homeomorphic to
S 2 . This contradicts the combinatorial Gauss–Bonnet Theorem.

Now consider the general case, where there are possibly edges
labelled with a zero. Our proof goes by induction on the number Z
of edges that have the zero label. We already treated the case when
Z = 0. In general, suppose that e is an edge labelled 0. There are
two cases.
In the first case, suppose that the closed edge E is embedded.
We can form a new graph Γe in S 2 by collapsing e to a point and
dragging all the edges of Γ incident to e to this new point; see Figure
24.3.

c
c a
a
e

b b
d d
Figure 24.2. Collapsing an edge

Our operation only changes the two components of S 2 − Γ that

share e. These components remain topological disks: we are just
shrinking one of their edges to a point. Moreover, since e is labelled
0, the lists associated to each of these two components remain good.
In short Γe satisfies the same hypotheses as Γ but has one fewer edge
labelled 0.
24.5. Proof of Lemma 24.2 301

Figure 24.3. Replacing the inner part of the graph with a disk

In the other case, e is a loop in S 2 . Note that e divides S 2 into

two disks. At least one of these disks—say the outer one, as in Figure
24.3—contains some edges of Γ. Pick such a disk, and then erase the
portion of Γ contained in the other disk—the inner disk in Figure
24.3. Finally, erase e. Figure 24.4 shows this operation. The result is
a smaller graph that satisfies the hypotheses of Lemma 24.3 but the
Z value has decreased by one.

24.5. Proof of Lemma 24.2

Say that a spherical arm is a connected polygonal arc contained in
the boundary of a convex spherical polygon. Thus, a spherical arm is
made from a finite union of arcs of great circles, meeting end to end.
We insist that the two endpoints of the spherical arm are distinct, so
that the spherical arm does not make a complete circuit around the
spherical polygon. Given the notion of convexity discussed in Chapter
9, a spherical arm is necessarily contained in a hemisphere.
Suppose that A(0) and A(1) are spherical arms, each consisting of
n geodesic segments. Let A1 (k), . . . , An (k) be the geodesic segments
comprising A(k), taken in order. Let a1 (k), . . . , an (k) be the vertices
of A(k). Finally, let θj (k) be the interior angle of A(k) at aj (k). The
choice of interior angle makes sense, thanks to convexity.
302 24. The Cauchy Rigidity Theorem

A4

θ a4
4

θ3
A3
θ2
A1 a3
A2
a2
Figure 24.4. A spherical arm of length 4

Below we will prove Cauchy’s Arm Lemma:

Lemma 24.5 (Cauchy’s Arm Lemma). Assume that θj (0) ≤ θj (1)

for all j, with strict inequality for at least one index. Then

d(a0 (0), an (0)) < d(a0 (1), an (1)).

Here d denotes spherical distance.

Before proving Cauchy’s Arm Lemma, let’s use it to prove Lemma

24.2. Let v be a vertex of a strictly convex polyhedron. Let Σ be a
small sphere centered at v. The intersection ∂P ∩ Σ is a convex
spherical polygon. Dilating the picture, we think of this polygon as
existing on S 2 , the unit sphere.
We can make this construction for partner vertices v and v  on
P and P  , respectively. This produces two convex spherical polygons
C and C  . The lengths of the edges of C are the same as the lengths
of the corresponding edges of C  . We label the vertices of C as in
Cauchy’s Arm Lemma from the previous chapter, depending on the
comparison between the two internal angles at the vertices.
If our list of labels is not good, we can find a chord of C so that
all the (+) labels occur on one side and all the (−) labels occur on the
other. This is shown in Figure 24.5. Let p and q be the endpoints of
this chord. Let C1 denote one of the arcs of C connecting p to q, and
let C2 denote the other. Let C1 and C2 be the corresponding chords
on C  .
24.6. Euclidean Intuition Does Not Work 303

0 −
+ C1
C2
−

+
p
Figure 24.5. Dividing the polygon in half

Applying Cauchy’s Arm Lemma to C1 and C1 , we conclude that

p − q > p − q  .

Applying the Arm Lemma to C2 and C2 , we get the opposite inequal-
ity. This is a contradiction.

24.6. Euclidean Intuition Does Not Work

The proof of Cauchy’s Arm Lemma is actually rather difficult, though
the result seems obvious based on Euclidean intuition.
Cauchy’s mistake was that he assumed a result from Euclidean
geometry that is false in the spherical case. In this section, I’ll high-
light the difference between the Euclidean and spherical cases. My
reason for doing this is to justify the difficulty it takes to actually
prove Cauchy’s Arm Lemma.
Consider a Euclidean version of the main construction. We say
that a Euclidean arm is a connected arc of a convex Euclidean poly-
gon. Suppose that A(0) is a Euclidean arm. We can make a polygonal
arc A(t), for t > 0 by increasing the last angle of A(0). Call this an-
gle θ(t). We keep everything else fixed. One should picture a person
flexing his finger; Figure 24.5 shows the situation.
304 24. The Cauchy Rigidity Theorem

Figure 24.6. Flexing a Euclidean arm.

Exercise 7. Prove that A(t) is a Euclidean arm provided that

θ(t) < π. That is, the process of opening up one of the angles cannot
destroy convexity.

Exercise 8. Show, by example, that the conclusion of Exercise 7

is no longer true in the spherical case.

The natural approach to proving Cauchy’s Arm Lemma is to

simply open up one of the arms a bit at a time, showing that the
distance between the endpoints keeps increasing. Unfortunately, in
the spherical case, the object can cease to be an arm at some point
because one can lose the convexity.

24.7. Proof of Cauchy’s Arm Lemma

Let’s analyze the problem of flexing a spherical arm. Let B(t) be a
spherical arm for all 0 ≤ t < s. Let b0 , . . . , bn−1 , bn (t) be the points
of B(t). Only the last point changes. Let B1 , . . . , Bn−1 , Bn (t) be the
segments of B(t). We suppose that the angle θ(t) at bn−1 increases
as t → s, but that θ(s) < π.

Lemma 24.6. Whether or not B(s) is a spherical arm, B(s) lies in

some open hemisphere.

Proof. Suppose not. Let B n (t) be the great circle extending Bn (t),
and let H(t) denote the open hemisphere containing B(t) − Bn (t) for
t small. By Exercise 4 of Chapter 9, and continuity, H(t) contains
B(t) − Bn (t) for all t < s. But H(s) cannot contain B(s) − Bn (s),
24.7. Proof of Cauchy’s Arm Lemma 305

because then we could move B(s) by a tiny amount so that it lies in

H(s).

bj H
γ

γ
Bn
b n−1
Figure 24.7. An arm and a great circle

The only possibility is that there is some vertex bj (s) contained

n (s). Here j ≤ n − 2. Let γ be the geodesic joining bn−1 to bj .
in B
The angle between γ and Bn (t) is bounded away from both 0 and
π. The endpoints of γ are not antipodal because they are vertices of
a spherical arm. Therefore, γ is the unique geodesic connecting its
endpoints. But, the condition bj ∈ Bn (s) forces γ ⊂ B
n (s). This is a
contradiction. 

We keep going with the same set-up as in the previous lemma.

Lemma 24.7. If B(s) is not a spherical arm, then b0 , b1 , and bn (s)
lie on the same arc of a great semicircle, with b1 between b0 and bn (s).

Proof. From the previous result, we know that B(t) lies in some
open hemisphere for all t ≤ s. Since B(s) is not a spherical arm,
there are 3 points β0 , β1 , β2 ∈ B(s), not all on the same edge of B(s)
but all lying on the same geodesic segment β. These 3 points cannot
lie in any spherical arm, so one of the points, say β0 , must lie in
Bn (s) − bn−1 .
For the same reason as in the previous lemma, β does not lie in
n (s). At the same time, β cannot be transverse to
the great circle B
306 24. The Cauchy Rigidity Theorem

B(s) at any βj . Otherwise, by stability, we would have a similar triple

of points for all t sufficiently close to s.
Suppose β0 is an interior point of Bn (s). Since β ∈ B n (s), the
segment β is transverse to B(s) at β0 . This is a contradiction. Hence
β0 = Bn (s).
The points β1 and β2 both lie on the spherical arm A = B(s) −
Bn (s). If β1 and β2 do not lie on the same edge of A, then β is
transverse to A at both β1 and β2 . This is a contradiction. Hence β1
and β2 lie on the same edge of A. Since A ⊂ B(s), we see that β1
and β2 lie on the same side, say, the jth side, of B(s).
If j > 1 then we have the topological picture shown (for j = 2)
in Figure 24.8. This picture is implied by the fact that B(s) − Bn (s)
is a spherical arm. But then there is a geodesic nearby β, through
bn (s), which intersects B(s) transversely at the two other intersection
points. The same stability argument as above gives us a contradiction.

B2

bn

Figure 24.8. The limiting shape

Now we know that b0 and b1 and bn (s) lie on the same great
circle. Since all these points lie in an open hemisphere, all these
points lie in the interior of some great semicircle. Finally, observe
that the geodesic connecting b0 to b2 crosses the geodesic connecting
b1 to bn (t) for all t < s. Taking a limit, as t → s, establishes that b1
lies between b0 and bn (s). 

Finally, we prove Cauchy’s Arm Lemma. The proof goes by in-

duction on n. In the case where n = 2, the result follows from Ex-
ercise 5 of Chapter 9. Consider the special case when θj (0) = θj (1)
for some j. Then we can produce a new spherical arm B by re-
placing Aj ∪ Aj+1 by the single geodesic segment connecting aj to
24.7. Proof of Cauchy’s Arm Lemma 307

aj+1 . Here we are just cutting off a corner. The spherical arms

B(0) and B(1) satisfy the same hypotheses as do A(0) and A(1), and
our basic move has not changed the endpoints. Hence, by induction
d(a0 (0), an (0)) < d(a0 (1), an (1)).
For t ∈ [0, 1], let θn−1 (t) be the angle that linearly interpolates
between θn−1 (0) and θn−1 (1). Let B(t) denote the polygonal curve
that is the same as A(0), except that we move An so that the angle
between An and An−1 is θ(t). Let Bj (t) be the jth segment of B(t),
and let bj (t) be the jth vertex. We have set things up so that Bj (t) =
Aj (0) for j = 1, . . . , n − 1 and bj (t) = aj (0) for j = 1, . . . , n − 1. Only
the last segment moves.
Suppose that B(1) is a spherical arm. Only one angle of B(1)
differs from A(1). At the same time, the last angle of B(1) is the
same as the last angle of A(1). Thus, we may apply the special
case we have already considered, twice, to get the following chain of
inequalities:

d(a0 (0), an (0)) < d(b0 (0), bn (0)) < d(a0 (1), an (1)).

Now we come to the hard part of the proof. Suppose that B(1)
is not a spherical arm. By Exercise 8 above, this case really can
happen. If B(1) is not a spherical arm, then there is some s such that
B(t) is a spherical arm for all t < s, but B(s) is not a spherical arm.
By Lemma 24.7, the points b0 (s), b1 (s), bn (s) lie on the same great
half-circle, with b0 (s) between b1 (s) and bn (s). Therefore

(24.1) d(b1 (s), bn (s)) = d(b0 (s), b1 (s)) + d(d0 (s), bn (s)).

We have
d(a0 (1), an (1))
≥1 d(a1 (1), an (1)) − d(a0 (1), a1 (1))
≥2 lim d(a1 (t), an (t)) − d(a0 (1), a1 (1))
t→s
(24.2)
=3 lim d(a1 (t), an (t)) − d(a0 (t), a1 (t))
t→s
=4 d(a1 (s), an (s)) − d(a0 (s), a1 (s)))
=5 d(a0 (s), an (s)).
308 24. The Cauchy Rigidity Theorem

The first inequality is the triangle inequality. The second in-

equality is the induction step applied to the spherical arm obtained
from B(t) by chopping off the first segment B1 (t). The third equality
comes from the fact that b0 (t) and b1 (t) are independent of t. The
fourth equality is continuity. The fifth equality is equation (24.1).
On the other hand, choosing any u ∈ (0, s), we have
d(a0 (s), an (s))
= lim d(a0 (t), an (t))
t→s
(24.3)
≥ d(a0 (u), an (u))
1

> d(a0 (0), an (0).

The first inequality comes from the special case (some angles equal)
applied to B(u) and B(u). The last inequality comes from the special
case appied to B(u) and B(0). Our last two equations combine to give
the statement in Cauchy’s Arm Lemma. This completes the proof.
Bibliography

[AHL] L. Ahlfors, Complex Analysis, McGraw-Hill, New York, 1952.

[AIZ] M. Aigner and G. Ziegler, Proofs from The Book, Springer-Verlag,
1998.
[BAL] T. Banchoff and S. Lovett, Differential Geometry of Curves and
Surfaces, A. K. Peters, Ltd., Natick, MA, 2010.
[BE1] A. Beardon, The Geometry of Discrete Groups, Graduate Texts in
Mathematics 91, Springer-Verlag, New York, 1983.
[BE2] A. Beardon, A Primer on Riemann Surfaces, L.M.S. Lecture Note
Series 78, Cambridge University Press, Cambridge, 1984.
[BRO] M. Beck and S. Robins, Computing the Continuous Discretely, Un-
dergraduate Texts in Mathematics, Springer-Verlag, New York,
2007.
[CHE] C. Chevalley, Theory of Lie Groups, Princeton Mathematical Series
8, Princeton University Press, Princeton, NJ, 1999.
[DAV] H. Davenport, The Higher Arithmetic (8th ed.), Cambridge Uni-
versity Press, Cambridge, 2008.
[DEV] K. Devlin, The Joy of Sets: Fundamentals of Contemporary Set
Theory (2nd Ed.), Undergraduate Texts in Mathematics, Springer-
Verlag, New York, 1993.
[DOC] M. DoCarmo, Riemannian Geometry, Mathematics: Theory & Ap-
plications, Birkhauser, Boston, 1992.
[DRT] A. Driscoll and L. N. Trefethen, Schwarz-Christoffel Mapping, Cam-
bridge Monographs on Applied and Computational Mathematics,
Cambridge University Press, Cambridge, 2002.

309
310 Bibliography

[FMA] B. Farb and D. Margalit, A Primer on Mapping Class Groups,

Princeton University Press, Princeton (to appear).
[GAR] F. Gardiner, Teichmüller Theory and Quadratic Differentials, Pure
and Applied Mathematics, John Wiley and Sons, New York, 1987.
[GPO] V. Guillemin and A. Pollack Differential Topology, Prentice Hall,
Englewood Cliffs, NJ, 1974.
[HAT] A. Hatcher, Algebraic Topology, Cambridge University Press, Cam-
bridge, 2002.
[HCV] D. Hilbert and A. Cohn-Vossen, Geometry and the Imagination,
Chelsea Publishing Company, New York, 1952.
[HER] I. M. Herstein, Topics in Algebra, 2nd Ed., John Wiley and Sons,
Xerox College Publishing, Lexington, MA, 1975.
[KAT] S. Katok, Fuchsian Groups, Chicago Lectures in Mathematics, Uni-
versity of Chicago Press, Chicago, 1992.
[KEN] K. Kendig, Elementary Algebraic Geometry, Graduate Texts in
Mathematics 44, Springer-Verlag, New York–Berlin, 1977.
[KIN] L. C. Kinsey, The Topology of Surfaces, Undergraduate Texts in
Mathematics, Springer-Verlag, New York, 1993.
[MAT] H. Masur and S. Tabachnikov, Rational Billiards and Flat Struc-
tures, Handbook of Dynamical Systems, Vol. 1A, 1015-1089, North-
Holland, Amsterdam, 2002.
[MUN] J. R. Munkries, Topology, Prentice Hall, Englewood Cliffs, NJ,
1975.
[RAT] J. Ratcliff, Foundations of Hyperbolic Manifolds, Graduate Texts in
Mathematics, 149, Springer-Verlag, New York, 1994.
[SPI] M. Spivak, Calculus on Manifolds. A Modern Approach to Classi-
cal Theorems of Advanced Calculus, W. A. Benjamin, New York–
Amsterdam, 1965.
[TAP] K. Tapp, Matrix Groups for Undergraduates, Student Mathematical
Library, 29, American Math Society, Providence, RI, 2005.
[THU] W. P. Thurston, The Geometry and Topology of 3-Manifolds, Lec-
ture Notes, Princeton University Press, Princeton, NJ, 1978.
[WAG] S. Wagon, The Banach-Tarski Paradox, Encyclopedia of Mathe-
matics and Its Applications 24, Cambridge University Press, Cam-
bridge, 1985.
[WAL] W. Wallace, Question 269, in Thomas Leyborne, Math. Repository
III, London, 1814.
Index

1-forms, 98 chain, 32
2-adic metric, 23 chain rule, 164
circle, 56, 116
Abelian, 43 circular polygon, 166
affine automorphisms, 221 classification of isometries, 130
angle, 87
classification of surfaces, 36
angle deficit, 209
Clean Dissections, 292
antipodal points, 69
closed set, 23
area preserving maps, 219
closure, 23
atlas, 136
combinatorial Gauss–Bonnet
Banach–Tarski Theorem, 17, 275 Theorem, 209
basepoint, 46, 48 compact, 26, 36
basic reflection, 105 compact Riemann surfaces, 203
basic rotation, 105 complex analysis, 11
bijection, 21 complex analytic function, 163
billiards, 213 composing loops, 46
binary icosahedral group, 63 composition, 24
Bolyai-Gerwein Theorem, 96 cone surface, 13
Bolzano-Weierstrass Theorem, 72 connect sum, 36
boundary, 23 continued fractions, 241, 245
continuity, 24
Cauchy Integral Formula, 163, 165,
convex geodesic polygon, 145
167
convexity, 110
Cauchy Rigidity Theorem, 17, 295
Cauchy sequence, 154 coordinate chart, 136
Cauchy’s Arm Lemma, 302 covering group, 66
Cauchy’s Theorem, 165 covering space, 8, 65
Cauchy-Riemann equations, 165 curve, 133
Cauchy-Schwarz inequality, 87 cylinder, 34
Cayley’s Theorem, 44 cylinder decomposition, 234

311
312 Index

deck group, 66, 71 Giraud’s Theorem, 107

deck transformation, 66 gluing, 2, 3, 31, 145
Dehn twist, 235 gluing recipe, 147
Dehn’s Dissection Theorem, 17, 98, good metric space, 80
287 Green’s Theorem, 100, 166
determinant, 30 group, 43
diffeomorphism, 135
differential representation, 223 Hadamard’s Theorem, 154
disk model, 125 harmonic function, 179
disk rigidity, 177 holomorphic function, 163
disk-like sets, 143, 144 homeomorphism, 25
dodecahedron, 63 homomorphism, 44
dot product, 22, 87 homotopy, 45
double octagon, 232 homotopy equivalence, 45
Hopf–Rinow Theorem, 154
edges, 5 hyperbolic geometry, 9
elliptic isometry, 130 hyperbolic group actions, 224
equivalence class, 32 hyperbolic isometry, 121, 130, 143
equivalence relation, 32 hyperbolic plane, 9, 224
essential curves, 5 hyperbolic reflections, 229
Euclidean cone, 207 hyperbolic surface, 143, 147, 195
Euclidean cone surface, 208
Euclidean plane, 1 ideal triangle, 127, 242
Euclidean space, 1, 87 ideal vertices, 127
Euler characteristic, 37 Implicit Function Theorem, 28
exponential function, 175 infimum, 31
infinite differentiability, 169
faces, 5 inner product, 121, 134
Farey graph, 242, 245 interior, 23
Farey related, 242 Invariance of Domain, 189
figure 8, 69 Inverse Function Theorem, 135, 185
flat cone surface, 207 irrational numbers, 247
flat torus, 16, 67, 195, 252 isometry, 135
functoriality, 49 isomorphism, 44
fundamental group, 43, 46, 71 isomorphism theorem, 71, 74, 79
Fundamental Theorem of Algebra,
57, 180 Jordan curve, 199
Jordan domain, 199
Gauss map, 239
Gauss–Bonnet Theorem, 107, 127 Klein bottle, 35
Gauss-Bonnet Theorem, 151
generalized circle, 116 latitude, 105
geodesic, 145 lens space, 59
geodesic half-plane, 150 Lie group, 30
geodesic hyperbolic triangle, 228 lifting property, 73
geodesic polygon, 127, 145 line integrals, 98
geodesic triangle, 107 linear fractional transformation, 11,
geodesic triangulation, 150 115, 130
geodesics, 9, 105, 120, 123 linear functionals, 98
Index 313

linear reflections, 229 Poincaré Uniformization Theorem,

Liouville’s Theorem, 179 201
longitude, 105 Polygon Dissection Theorem, 96
loop, 46 polygonal billiards, 213
power series, 172
manifold, 28, 59 product space, 51
map, 24 projective plane, 34, 59, 69
mapping class group, 260, 263 projective space, 59
marked loop, 190 properly discontinuous action, 224
marked parallelograms, 251 Pythagorean Theorem, 90
matrix group, 30
maximal atlas, 136 quadratic irrationals, 248
Maximum Principle, 170 quaternions, 62
metric, 22
real linear transformation, 121
metric completeness, 154
removable singularity, 171
metric space, 22
Riemann Mapping Theorem, 188,
Möbius band, 34
199
Möbius geometry, 11
Riemann surface, 195
Möbius transformation, 11, 115
Riemann surface covering, 196
modular group, 14, 242, 244, 254
Riemannian cover, 152
moduli space, 16, 256, 263
Riemannian geometries, 9
Riemannian isometry, 135
negative genus, 258
Riemannian length, 139
nonorientable, 35
Riemannian manifold, 152
nonorientable surfaces, 41
Riemannian metric, 134, 139
norm, 87
Riemannian surface, 139
right-angled hexagon, 150
octagon, 4
octagon surface, 9, 34 saddle connection, 227
one-to-one, 21 Schwarz–Christoffel
onto, 21 Transformation, 188
open set, 23 scissors congruent, 287
orientable, 35 sectors, 207
overlap function, 136 set, 21
simple connectivity, 70
pairs of pants, 263 Sinai Robins, 92
pants decomposition, 265 Small Picard Theorem, 202
parabolic isometry, 130 smooth curve, 133, 137
parallelogram tiling, 68 smooth surface, 136
path connectivity, 48 sphere, 1, 58
path homotopic, 70 spherical isometries, 103
pathification, 32 spherical metrics, 103
Peano curve, 51 spherical triangle, 107
periodic billiard paths, 219 square torus, 1
permutation group, 43 stereographic projection, 111, 181
Pick’s Theorem, 92 surface, 21, 27
Poincaré homology sphere, 62
Poincaré Recurrence Theorem, 219 tangent plane, 103, 137
314 Index

Taylor series, 163, 174

Teichmüller space, 16, 254, 258, 263
The Riemann sphere, 196
Thomas Harriot, 107
topological space, 25
torus, 1, 58
translation surface, 13, 211, 221
tree, 69
triangle group, 228
triangle inequality, 22, 32, 87

unit convergence condition (UCC),

172
universal cover, 79, 156
universal covering map, 57
universal covering theorem, 80
upper half-plane, 118

Veech group, 14, 224, 228

vertices, 5
volume, 87

William Wallace, 96
winding number, 53

X Theorem, 91
This book presents a number of topics related to surfaces, such as
Euclidean, spherical and hyperbolic geometry, the fundamental
group, universal covering surfaces, Riemannian manifolds, the
Gauss-Bonnet Theorem, and the Riemann mapping theorem. The
main idea is to get to some interesting mathematics without too
much formality. The book also includes some material only tangen-
tially related to surfaces, such as the Cauchy Rigidity Theorem, the
Dehn Dissection Theorem, and the Banach–Tarski Theorem.
The goal of the book is to present a tapestry of ideas from various
areas of mathematics in a clear and rigorous yet informal and
friendly way. Prerequisites include undergraduate courses in real
analysis and in linear algebra, and some knowledge of complex
analysis.

For additional information

and updates on this book, visit
www.ams.org/bookpages/stml-60

AMS on the Web

STML/60 www.ams.org