Topology As Fluid Geometry Two Dimensional Spaces Volume 2 PDF

Two-Dimensional Spaces, Volume 2
TOPOLOGY AS FLUID
GEOMETRY
James W. Cannon
TOPOLOGY AS FLUID
GEOMETRY
TOPOLOGY AS FLUID
GEOMETRY
James W. Cannon
AMERICAN MATHEMATICAL SOCIETY

Providence, Rhode Island
2010 Mathematics Subject Classification. Primary 57-01, 57M20.
For additional information and updates on this book, visit

www.ams.org/bookpages/mbk-109
Library of Congress Cataloging-in-Publication Data

Names: Cannon, James W., author.
Title: Two-dimensional spaces / James W. Cannon.
Description: Providence, Rhode Island : American Mathematical Society, [2017] | Includes bibli-
ographical references.
Identifiers: LCCN 2017024690 | ISBN 9781470437145 (v. 1) | ISBN 9781470437152 (v. 2) | ISBN
9781470437169 (v. 3)
Subjects: LCSH: Geometry. | Geometry, Plane. | Non-Euclidean geometry. | AMS: Geometry –
Instructional exposition (textbooks, tutorial papers, etc.). msc
Classification: LCC QA445 .C27 2017 | DDC 516–dc23
LC record available at https://lccn.loc.gov/2017024690
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 select pages 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. Permissions to reuse
portions of AMS publication content are handled by Copyright Clearance Center’s RightsLink
service. For more information, please visit: http://www.ams.org/rightslink.
Send requests for translation rights and licensed reprints to [email protected].
Excluded from these provisions is material for which the author holds copyright. In such cases,
requests for permission to reuse or reprint material should be addressed directly to the author(s).
Copyright ownership is indicated on the copyright page, or on the lower right-hand corner of the
first page of each article within proceedings volumes.

c 2017 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 22 21 20 19 18 17
Contents
Preface to the Three Volume Set ix
Preface to Volume 2 xiii
Chapter 1. The Fundamental Theorem of Algebra 1

1.1. Complex Arithmetic 2
1.2. First Proof of the Fundamental Theorem 5
1.3. Second Proof 7
1.4. Exercises 10
Chapter 2. The Brouwer Fixed Point Theorem 11

2.1. Statement of the Theorem 11
2.2. Introducing Extra Structure into a Problem 12
2.3. Two Elementary Problems 12
2.4. Three Advanced Problems 18
2.5. Exercises 27
Chapter 3. Tools 29
3.1. Polyhedral complexes 29
3.2. Urysohn’s Lemma and the Tietze Extension Theorem 31
3.3. Set Convergence 34
3.4. Exercises 36
Chapter 4. Lebesgue Covering Dimension 37

4.1. Definition of Covering Dimension 37
4.2. Euclidean n-Dimensional Space Rn Has Covering Dimension n 38
4.3. Construction of Partitions of Unity 40
4.4. Techniques Needed in Higher Dimensions 41
4.5. Exercises 42
Chapter 5. Fat Curves and Peano Curves 45

5.1. The Constructions 45
5.2. The Topological Lemmas 49
5.3. The Analytical Lemmas 51
5.4. Characterization of Peano Curves 52
5.5. Exercises 54
Chapter 6. The Arc, the Simple Closed Curve, and the Cantor Set 57
6.1. Characterizing the Arc and Simple Closed Curve 57
6.2. The Cantor Set and Its Characterization 61
6.3. Interesting Cantor Sets 63
v
vi CONTENTS
6.4. Cantor Sets in the Plane Are Tame 69

6.5. Exercises 73
Chapter 7. Algebraic Topology 75

7.1. Facts Assumed from Algebraic Topology 76
7.2. The Reduced Homology of a Sphere 77
7.3. The Homology of a Ball Complement 77
7.4. The Homology of a Sphere Complement 78
7.5. Proof of the Arc Non-Separation Theorem and the Jordan Curve
Theorem 79
Chapter 8. Characterization of the 2-Sphere 81

8.1. Statement and Proof of the Characterization Theorem 81
8.2. Exercises 89
Chapter 9. 2-Manifolds 91
9.1. Definition and Examples 91
9.2. Exercises 91
Chapter 10. Arcs in S2 Are Tame 95

10.1. Arcs in S2 Are Tame 95
10.2. Disk Isotopies 97
10.3. Exercises 100
Chapter 11. R. L. Moore’s Decomposition Theorem 101

11.1. Examples and Applications 101
11.2. Decomposition Spaces 102
11.3. Proof of the Moore Decomposition Theorem 104
11.4. Exercises 107
Chapter 12. The Open Mapping Theorem 109

12.1. Tools 109
12.2. Two Lemmas 110
12.3. Proof of the Open Mapping Theorem 112
12.4. Exercise 112
Chapter 13. Triangulation of 2-Manifolds 113

13.1. Statement of the Triangulation Theorem 113
13.2. Tools 113
13.3. Proof of the Triangulation Theorem 115
13.4. Exercises 116
Chapter 14. Structure and Classification of 2-Manifolds 117

14.1. Statement of the Structure Theorem 117
14.2. Edge-pairings 118
14.3. Proof of the Structure Theorem 119
14.4. Statement and Proof of the Classification Theorem 123
14.5. Exercises 125
Chapter 15. The Torus 129

15.1. Lines and Arcs in the Plane 129
CONTENTS vii
15.2. The Torus as a Euclidean Surface 131

15.3. Curve Straightening 134
15.4. Construction of the Simple Closed Curve with Slope k/ 136
15.5. Exercises 137
Chapter 16. Orientation and Euler Characteristic 139
16.1. Orientation 139
16.2. Euler Characteristic 141
16.3. Exercises 149
Chapter 17. The Riemann-Hurwitz Theorem 151
17.1. Setting 151
17.2. Elementary Facts from Trigonometry 151
17.3. Branched Maps of S2 154
17.4. Statement of the Riemann-Hurwitz Theorem 155
17.5. Proof of the Riemann-Hurwitz Theorem 155
17.6. Rational Maps 156
17.7. Exercises 157
Bibliography 159
Preface to the Three Volume Set
Geometry measures space (geo = earth, metry = measurement). Einstein’s

theory of relativity measures space-time and might be called geochronometry (geo
= space, chrono= time, metry = measurement). The arc of mathematical history
that has led us from the geometry of the plane of Euclid and the Greeks after 2500
years to the physics of space-time of Einstein is an attractive mathematical story.
Geometrical reasoning has proved instrumental in our understanding of the real and
complex numbers, algebra and number theory, the development of calculus with its
elaboration in analysis and differential equations, our notions of length, area, and
volume, motion, symmetry, topology, and curvature.
These three volumes form a very personal excursion through those parts of the
mathematics of 1- and 2-dimensional geometry that I have found magical. In all
cases, this point of view is the one most meaningful to me. Every section is designed
around results that, as a student, I found interesting in themselves and not just
as preparation for something to come later. Where is the magic? Why are these
things true?
Where is the tension? Every good theorem should have tension
between hypothesis and conclusion. — Dennis Sullivan
Where is the Sullivan tension in the statement and proofs of the theorems?
What are the key ideas? Why is the given proof natural? Are the theorems almost
false? Is there a nice picture? I am not interested in quoting results without proof.
I am not afraid of a little algebra, or calculus, or linear algebra. I do not care
about complete rigor. I want to understand. If every formula in a book cuts the
readership in half, my audience is a small, elite audience. This book is for the
student who likes the magic and wants to understand.
A scientist is someone who is always a child, asking ‘Why?
why? why?’. — Isidor Isaac Rabi, Nobel Prize in Physics
1944
Wir müssen wissen, wir werden wissen. [We must know, we
will know.] — David Hilbert
The three volumes indicate three natural parts into which the material on 2-
dimensional spaces may be divided:
Volume 1: The geometry of the plane, with various historical attempts to
understand lengths and areas: areas by similarity, by cut and paste, by counting,
by slicing. Applications to the understanding of the real numbers, algebra, number
theory, and the development of calculus. Limitations imposed on the measurement
of size given by nonmeasurable sets and the wonderful Hausdorff-Banach-Tarski
paradox.
ix
x PREFACE TO THE THREE VOLUME SET
Volume 2: The topology of the plane, with all of the standard theorems of
1 and 2-dimensional topology, the Fundamental Theorem of Algebra, the Brouwer
Fixed-Point Theorem, space-filling curves, curves of positive area, the Jordan Curve
Theorem, the topological characterization of the plane, the Schoenflies Theorem,
the R. L. Moore Decomposition Theorem, the Open Mapping Theorem, the trian-
gulation of 2-manifolds, the classification of 2-manifolds via orientation and Euler
characteristic, dimension theory.
Volume 3: An introduction to non-Euclidean geometry and curvature. What
is the analogy between the standard trigonometric functions and the hyperbolic
trig functions? Why is non-Euclidean geometry called hyperbolic? What are the
gross intuitive differences between Euclidean and hyperbolic geometry?
The approach to curvature is backwards to that of Gauss, with definitions
that are obviously invariant under bending, with the intent that curvature should
obviously measure the degree to which a surface cannot be flattened into the plane.
Gauss’s Theorema Egregium then comes at the end of the discussion.
Prerequisites: An undergraduate student with a reasonable memory of cal-
culus and linear algebra, but with no fear of proofs, should be able to understand
almost all of the first volume. A student with the rudiments of topology—open and
closed sets, continuous functions, compact sets and uniform continuity—should be
able to understand almost all of the second volume with the exeption of a little
bit of algebraic topology used to prove results that are intuitively reasonable and
can be assumed if necessary. The final volume should be well within the reach of
someone who is comfortable with integration and change of variables. We will make
an attempt in many places to review the tools needed.
Comments on exercises: Most exercises are interlaced with the text in those
places where the development suggests them. They are an essential part of the
text, and the reader should at least make note of their content. Exercise sections
which appear at the end of most chapters refer back to these exercises, sometimes
with hints, occasionally with solutions, and sometimes add additional exercises.
Readers should try as many exercises as attract them, first without looking at hints
or solutions.
Comments on difficulty: Typically, sections and chapters become more diffi-
cult toward the end. Don’t be afraid to quit a chapter when it becomes too difficult.
Digest as much as interests you and move on to the next chapter or section.
Comments on the bibliography: The book was written with very little
direct reference to sources, and many of the proofs may therefore differ from the
standard ones. But there are many wonderful books and wonderful teachers that
we can learn from. I have therefore collected an annotated bibliography that you
may want to explore. I particularly recommend [1, G. H. Hardy, A Mathematician’s
Apology], [2, G. Pólya, How to Solve It], and [3, T. W. Körner, The Pleasure of
Counting], just for fun, light reading. For a bit of hero worship, I also recommend
the biographical references [21, E. T. Bell, Men of Mathematics], [22, C. Henrion,
Women of Mathematics], and [23, W. Dunham, Journey Through Genius]. And
I have to thank my particular heroes: my brother Larry, who taught me about
uncountable sets, space-filling curves, and mathematical induction; Georg Pólya,
who invited me into his home and showed me his mathematical notebooks; my ad-
visor C. E. Burgess, who introduced me to the wonders of Texas-style mathematics;
R. H. Bing, whose Sling, Dogbone Space, Hooked Rug, Baseball Move, epslums and
PREFACE TO THE THREE VOLUME SET xi
deltas, and Crumpled Cubes added color and wonder to the study of topology; and
W. P. Thurston, who often made me feel like Gary Larson’s character of little brain
(“Stop, professor, my brain is full.”) They were all kind and encouraging to me.
And then there are those whom I only know from their writing: especially Euclid,
Archimedes, Gauss, Hilbert, and Poincaré.
Finally, I must thank Bill Floyd and Walter Parry for more than three decades
of mathematical fun. When we would get together, we would work hard every
morning, then talk mathematics for the rest of the day as we hiked the cities, coun-
trysides, mountains, and woods of Utah, Virginia, Michigan, Minnesota, England,
France, and any other place we could manage to get together. And special thanks
to Bill for cleaning up and improving almost all of those figures in these books
which he had not himself originally drawn.
Preface to Volume 2
The first of three volumes in this set was devoted to the measurement of lengths
and areas, and to some of the consequences that study had in number theory,
algebra, and analysis. Euclid was able to solve quadratic equations by geometric
construction. But when mathematicians tried to extend those results to equations
of higher degree and to differential equations, a number of fascinating difficulties
arose, all involving limits and continuity, best modelled by topology.
In this second volume we assume that the reader has had a first course in
topology and is comfortable with open and closed sets, connected sets, compact
sets, limits, and continuity. Two good references are W. S. Massey [25] and J. R.
Munkres [24].
The following discoveries led to the topics of this second volume.
(1) The solution of cubic and quartic equations required serious consideration of
complex numbers, thought at first to be mysterious. But the mystery disappeared
when it was seen that complex numbers simply model the Euclidean plane. Abel
and Galois proved that equations of degrees 5 and higher could not be solved in
the relatively simple manner by formula as had sufficed in equations of degrees 1
through 4. But Gauss, without giving explicit solutions, managed to prove the
Fundamental Theorem of Algebra that ensured that complex numbers sufficed for
their solution. Gauss gave proofs involving the geometry and topology of the plane.
(2) Newton showed that the study of motion could be greatly simplified if, in-
stead of examining standard equations, one examined differential equations. Prov-
ing the existence of solutions to rather general differential equations led to problems
in topology. One of the standard proof techniques involves Brouwer’s Fixed Point
Theorem. This volume proves that theorem in dimension 2 and outlines the proof
in general dimensions.
(3) Descartes demonstrated that mechanical devices other than straight edge
and compass can construct curves of very high degree. Once curves of very general
form are accepted as interesting, further delicate questions of length and area arise:
finite curves of infinite length, finite curves of positive area, space filling curves,
disks whose interiors have smaller areas than their closures, 0-dimensional sets
through which no light rays can penetrate, continuous functions that are nowhere
differentiable, sets of fractional dimension. This volume gives examples of many of
these phenomena.
(4) The study of solutions to equations became more unified when all variables
were considered to be complex variables. Riemann modelled complex curves by
surfaces, which are 2-dimensional manifolds and are called Riemann surfaces. The
analysis of 2-dimensional manifolds led naturally to notions, such as triangulation,
genus, and Euler characteristic. These notions are explained in this volume.
xiii
xiv PREFACE TO VOLUME 2
All of these considerations required the study of limits and continuity, and the
abstract notion that models limits and continuity in their most general settings is
the notion of topology. Henri Poincaré wrote:
As for me, all of the diverse paths which I have successively
followed have led me to topology. I have needed the gifts of this
science to pursue my studies of the curves defined by differential
equations and for the generalization to differential equations of
higher order, and, in particular, to those of the three body
problem. I have needed topology for the study of nonuniform
functions of two variables. I have needed it for the study of
the periods of multiple integrals and for the application of that
study to the expansion of perturbed functions. Finally, I have
glimpsed in topology a means to attack an important problem
in the theory of groups, the search for discrete or finite groups
contained in a given continuous group.
CHAPTER 1
The Fundamental Theorem of Algebra
Our first excursion into the topology of the plane will be in the proof of the
Fundamental Theorem of Algebra:
Theorem 1.1 (Fundamental Theorem of Algebra). If f (x) = xn + an−1 xn−1 +
· · · + a1 x + a0 = 0 is a polynomial equation in the unknown x and if the constant
coefficients an−1 , . . ., a1 , a0 are complex numbers, then there is a complex number
x = α that satisfies the equation: f (α) = 0. (Of course, since the real numbers are
a subset of the complex numbers (the line lies in the plane), these coefficients are
allowed to be real numbers.)
The Greeks solved linear and quadratic equations geometrically. They rejected
solutions that were not real numbers as being of no application or interest. Complex
numbers were first acknowledged as important in the solution of cubic equations.
General solution formulas for the cubic and quartic equations were found only by
great effort and cleverness.
Georg Pólya wrote to me when I was a young Mormon missionary in Austria.
He said that I should solve a hard mathematical problem every week so that I
wouldn’t rust (“Wer rastet, der rostet”). He also gave me a list of books that I
might find in a used bookstore. I learned the following argument from one of them:
Here is a general solution to the cubic equation, f (x) = x3 + ax2 + bx + c = 0:
We simplify the equation by translation, setting x = z + k. By subsititution, we
find
f (x) = z 3 + (3k + a)z 2 + (3k2 + 2ak + b)z + (k3 + ak2 + bk + c) = 0.
If we choose k = −a/3, the equation has the form
z 3 + pz + q = 0.
Setting z = u + v and substituting, we find
(u3 + v 3 + q) + (3uv + p)(u + v) = 0.
If we are able to choose u and v so that
u3 + v 3 + q = 0 and 3uv + p = 0,
the equation will be satisfied. We can solve this pair of equations for u and v in
the standard way by substitution:
u = −p/3v and, consequently,
3
−p
+ v 3 + q = 0.
3v
1
2 1. THE FUNDAMENTAL THEOREM OF ALGEBRA
Multiplying by v 3 ,

3
−p
v 6 + qv 3 + = 0.
3
Though this equation has degree 6 in the unknown v, it is only quadratic in the
unknown v 3 , so that by the quadratic formula,
3
3 2
q 1 p q q p
v =− ±
3
q2 + 4 =− ± + .
2 2 27 2 2 3
Assuming that we know how to take cube roots, we obtain v, from which we find
u3 = −v 3 − q or u = −p/3v, z = u + v, and x = z − (a/3).
For example, if√f (x) = x3 − 15x − 4 (already simplified), we have p = −15,
q = −4, v 3 = 2 + 4 − 125 = 2 + 11i, and u3 = 2 − 11i. But (2 + i)3 = 2 + 11i
and (2 − i)3 = 2 − 11i. The sum of these cube roots is (2 + i) + (2 − i) = 4,
which is a real root of the original equation. The fact that real roots could be
mediated by formulas involving complex numbers was a huge motivating factor for
the acceptance of complex numbers.
We have ignored some of the obvious difficulties: How do we take cube roots?
Further, if we take the three cube roots of v 3 and three of u3 , we would be able
to put together 9 possibilities for z, and there can only be three solutions to the
original equation. To choose those that are in fact roots, we must satisfy the more
restrictive equation, u = −p/3v. In the example,
−p 15 6 − 3i 15 · (6 − 3i)
= · = = 2 − i.
3(2 + i) 6 + 3i 6 − 3i 36 + 9
Mathematicians hoped to find similar solutions for polynomial equations of
higher degree, solutions that only required the arithmetic operations of addition,
subtraction, multiplication, and division together with the extraction of roots. This
hope was dashed by the work of N. H. Abel and E. Galois, who showed that the
four arithmetic operations and extraction of roots were inadequate in general for
expressing the roots of a quintic equation in terms of its coefficients.
I would love to include their proofs here, but Abel’s proof required twenty pages
of work and Galois’s development requires a substantial development of the theory
of fields and finite groups. Instead, we shall only prove the Fundamental Theorem
of Algebra. We first need to review some fundamentals from the arithmetic of
complex numbers.
1.1. Complex Arithmetic

Complex numbers were fully accepted when mathematicians learned to in-
terpret them as the points of the plane, with a + bi represented by the pair
(a, b). Addition is then simply vector addition or parallelogram addition, with
(a + bi) + (c + di) = (a + c) + (b + d)i corresponding to (a, b) + (c, d) = (a + c, b + d).
See Figure 1.
Multiplication also has a beautiful geometric interpretation that is based
on the polar representation of a complex number (see Figure 2) and on the sum
formulas for the sin and cos. We review these trigonometric sum formulas and their
proofs here.
The sum formulas for sin and cos are consequences of a simple projection prin-
ciple, that is essentially the definition of sin and cos. See Figure 3. It is helpful
1.1. COMPLEX ARITHMETIC 3
(a + c) + (b + d)i
c + di
a + bi
Figure 1. Addition of complex numbers
(r cos α, r sin α) = r(cos α, sin α)
(0, 0)
Figure 2. Polar coordinates
to think of the projection principle as determining the effect on lengths when one
length is projected orthogonally onto another.
The projection principle. Consider a right triangle with one of its acute
angles equal to α, and suppose that the length of the hypotenuse is r. Then the
leg adjacent to the angle α has length (cos α) · r, and the leg opposite the angle α
has length (sin α) · r.
To obtain the sum formulas for sin and cos, we apply the projection principle to
the following two diagrams; see Figure 4. Each term of the sum formulas represents
a well-defined geometric segment in the diagrams.
(sin α) · r
(cos α) · r
α
Figure 3. The projection principle
s(α)s(β) s(α)s(β)
s(α)c(β)
s(α)c(β) s(α)
s(α)
1 1
c(α) c(α)
α c(α)s(β) α c(α)s(β)
β β
(0, 0) c(α)c(β) (0, 0) c(α)c(β)
Figure 4. The addition formulas
We use c and s as shorthand for cos(α + β) and sin(α + β). We use c(α) and
s(α) as shorthand for cos(α) and sin(α), and similarly for cos(β) and sin(β). We
have drawn two right triangles in each of the diagrams, one with angle α, the other
with angle β. We have scaled the triangles so that the larger one has hypotenuse
1 so that the vertex emphasized by the large dot has coordinates (c, s). All of the
other entries are consequences of the projection principle. From the figures, it is
clear that
c = c(α)c(β) − s(α)s(α), or
cos(α + β) = cos(α) cos(β) − sin(α) sin(β), and similarly
sin(α + β) = cos(α) sin(β) + sin(α) cos(β).
These are the sum formulas for sin and cos.
Admittedly, the diagrams deal only with positive angles α and β whose sum is
≤ π, but all other angles can readily be reduced to these cases.
1.2. FIRST PROOF OF THE FUNDAMENTAL THEOREM 5
If we add to these two formulas the Pythagorean Theorem,

cos2 (α) + sin2 (α) = 1,
we have the three basic identities of trigonometry.
If we apply these identities to the multiplication of complex numbers, as ex-
pressed in polar coordinates, we obtain the geometric interpretation of complex
multiplication.
We multiply two complex numbers by applying the standard principles of mul-
tiplication and addition, then set i2 = −1:
α) · s(cos β + i sin β)
r(cos α+i sin
= rs(cos α · cos β − sin α · sin β)+ i(cos α · sin β + sin α · cos β)
= rs cos(α + β) + i · sin(α + β) .
The second equality is a consequence of the addition formulas for sin and cos. In
other words, to multiply, we multiply lengths and add angles.
To raise to the nth power,
n
r(cos α + i · sin α) = r n cos(nα) + i sin(nα) ,
we raise length to the nth power and multiply the angle by n.
Division is just as easy: divide lengths and subtract angles.
Taking the nth root is more sophisticated. Of course, we may take the nth
root of length and divide the angle by n. But the angle of a complex number is not
well-defined, for we can add any multiple of 2π to a given angle without changing
the complex number. When we divide the angle by n, we actually find n possible
angles:
α/n, α/n + 2π/n, α/n + 4π/n, ..., α/n + (n − 1)2π/n.
In other words, for every nonzero complex number there are exactly n nth roots,
and they are equally spaced on a circle of radius r 1/n centered at the origin.
1.2. First Proof of the Fundamental Theorem

We recall the statement of the Fundamental Theorem of Algebra:
Theorem 1.2. Suppose that p(z) is a nonconstant polynomial whose coefficients
are complex numbers. Then there is a complex number z0 such that p(z0 ) = 0.
I learned the first proof from [7, H. Dörrie]. The first proof is based on three
simple facts:
Lemma 1.3. If c ∈ C and if n > 0 is an integer, then there is a number d such
that dn = c.
Proof of the Lemma. We use the basic fact about multiplication of complex
numbers a and b: Express a and b in polar coordinates, so that a = (|a|, θ(a))pol
and b = (|b|, θ(b))pol . Then a · b = (|a| · |b|, θ(a) + θ(b))pol . It follows immediately
that an = (|a|n , n · θ(a))pol .
Consequently, the nth roots of c = (|c|, θ(c))pol are the complex numbers
θ(c) i
(|c|1/n , + · 2π)pol for i = 0, 1, . . . , n − 1.
n n

Lemma 1.4. If p(z) is a nonconstant complex polynomial, then limz→∞ p(z) =

∞.
n−1 + · · · + a1 z + a0 , with an = 0 and |z| > 1, then

Proof. If p(z) = an z n + an−1
|p(z)| ≥ (|z|n · |an | − |z|n−1 · (|an−1 | + · · · + |a1 | + |a0 |) > M,
provided that
|z| · |an | − (|an−1 | + · · · + |a1 | + |a0 |) > M/|z|n−1 .

Corollary 1.5. If p(z) is a complex polynomial, then there is a complex num-
ber z0 so that m = |p(z0 )| is the minimum value of |p(z)|.
Proof. Let m = inf{|p(z)| : z ∈ C}. Let z1 , z2 , . . . denote a sequence of
complex numbers such that m = limn→∞ |p(zn )|. Then the sequence must be
bounded because of the previous lemma. Hence there is a convergent subsequence,
so that we may assume zn → z0 ∈ C. By continuity, |p(z0 )| = m.
Lemma 1.6. If p(z) is a complex polynomial and z0 ∈ C, then there is a complex
polynomial of the form q(Z) = q(z − z0 ) such that p(z) = q(z − z0 ).
Proof. The coefficients of q(z−z0 ) = an (z−z0 )n +· · ·+a0 are easily calculated
by successive divisions (most efficiently implemented as synthetic division):
p(z) = p1 (z) · (z − z0 ) + a0 ,
p1 (z) = p2 (z) · (z − z0 ) + a1 ,
p2 (z) = p3 (z) · (z − z0 ) + a2 ,
p3 (z) = p4 (z) · (z − z0 ) + a3 , etc.
Reversing the steps demonstrates the desired equality. For example,
2z 3 + 17z 2 + 50z + 56 = (2z 2 + 11z + 17) · (z + 3) + 5,
2z 2 + 11z + 17 = (2z + 5) · (z + 3) + 2,
2z + 5 = (2) · (z + 3) − 1, and
2 = (0) · (z + 3) + 2,
so that
2z 3 + 17z 2 + 50z + 56 = 2(z + 3)3 − 1(z + 3)2 + 2(z + 3) + 5.

First proof of the Fundamental Theorem of Algebra. Suppose that,
to the contrary, the minimum value (Corollary 1.5) of |p(z)| is m = |p(z0 )| > 0.
By Lemma 1.6, we may translate the domain so that z0 = 0. Multiplying p(z) by
a nonzero complex constant, we may assume that |p(z0 )| = p(0) = 1, so that p(z)
has the form
p(z) = 1 + bk z k + bk+1 z k+1 + · · · + bn z n ,
where bk is the first nonzero coefficient after the constant 1.
The (nonzero) complex number −1/bk has a kth root d (1.3). Let λ denote a
positive number very close to 0. Then bk (λd)k = −λk so that
|p(λd)| ≤ 1 − λk + λk+1 · (|bk+1 dk+1 | + · · · + |bn dn |).
1.3. SECOND PROOF 7
If λ is chosen so small that λ(|bk+1 dk+1 | + · · · + |bn dn |) < 1, we conclude that

|p(λd)| < 1, a contradiction. We conclude that m = 0 so that p(z) has a root.
1.3. Second Proof

Our second proof is based on the geometrically intuitive idea of deforming or
dragging a closed curve in the plane. The technical term for deforming or dragging
is homotopy. The technical parts of the proof are more involved than our first proof
of the theorem, but the concepts are powerful and allow significant generalization.
Definition 1.7. Maps f, g : X → Y are said to be homotopic if there is a
continuous function F : X × [0, 1] → Y such that, for each x ∈ X, F (x, 0) = f (x)
and F (x, 1) = g(x). The mapping F is called a homotopy from f to g.
The curves in question are the images of circles under the polynomial mapping
p : C → C : z
→ p(z). Recall that S1 = S1 denotes the circle of radius 1 in the
complex plane centered at the origin. In complex notation,
S1 = {eiθ | θ ∈ R} = {eiθ | θ ∈ [0, 2π)}.
Using inner products, we obtain for each positive number r a concentric circle
Sr = r · S1 = {r · z : z ∈ S1 } of radius r. We may think of the point S0 = 0 · S1
as a degenerate curve that has been collapsed to a point. The polynomial mapping
p provides a homotopy deforming the image of any one of these circles Sr to the
point p(0) via the definition
Pr : Sr × [0, 1] → C : (x, t)
→ p((1 − t)x).
Note that Pr restricted to Sr × {0} is the same as p restricted to Sr , while Pr
restricted to Sr × {1} is the constant map to p(0).
We also use the fact already noted:
Lemma 1.8. If c ∈ C has polar coordinates c = (|c|, θ(c))pol and if n is a positive
integer, then cn = (|c|n , nθ(c))pol .
Corollary 1.9. The mapping q : C → C : z
→ z n , with n > 0 a positive
integer, wraps the circle Sr n times around the circle Ss , where s = r n .
We are now prepared to give the intuitive outline of our second proof of the
Fundamental Theorem of Algebra. We start with an arbitrary nonconstant poly-
nomial, and, dividing by the leading coefficient, find that we may assume it has
the form p(z) = z n + an−1 z n−1 + · · · + a1 z + a0 . We assume that, contrary to
the theorem, p(z) is never 0. Then the homotopy Pr : Sr × [0, 1] → C actually
has image in C \ {0}. That is, the curve p : Sr → C \ {0} can be dragged to
a point without hitting the origin. But we shall see that, for r sufficiently large,
the curve q : Sr → C \ {0} of the previous corollary is homotopic in C \ {0} to
p : Sr → C \ {0}. That is, q|Sr can be dragged to p|Sr missing the origin, and
from there, using the polynomial p, can be dragged to the constant map missing
the origin. That is, the curve that maps n times around the origin can be dragged
to a point without going through the origin. This is physically absurd since the
curve q|Sr is most thoroughly hooked about the origin. This completes the intuitive
proof of the theorem.
The difficulty that remains is that of turning these statements into technical
mathematics. The necessary details are supplied by the following two theorems.
Theorem 1.10. If q(z) = z n and p(z) = z n + an−1 z n−1 + · · · + a1 z + a0 , then,

for all sufficiently large r > 1, the curves q|Sr and p|Sr are homotopic in C \ {0}.
Proof. Define F : Sr ×[0, 1] → C by the formula F (s, t) = (1−t)·q(s)+t·p(s).
Then F is definitely a homotopy from q|Sr to p|Sr in C. We need to show that
F (s, t) is never 0, provided that r is big enough:
|F (s, t)| = |(1 − t)sn + t(sn + an−1 sn−1 + · · · + a1 s + a0 )|

≥ r n − r n−1 (|an−1 | + · · · + |a1 | + |a0 |) > 0,
provided that r > (|an−1 | + · · · + |a1 | + |a0 |).
Theorem 1.11. The map q|Sr is not homotopic to a constant map in C \ {0}.
Proof. This theorem is the technically complex bit of our second proof of the
Fundamental Theorem of Algebra.
Suppose that, contrary to the theorem, there is a continuous function Q :
Sr × [0, 1] → C \ {0} such that, for each s ∈ Sr , Q(s, 0) = q(s) = sn and Q|Sr × {1}
is a constant map.
We divide the product Sr ×[0.1] into tiny triangles, as, for example, in Figure 5.
Sr × 0
Sr × 1
Figure 5. Dividing a ring into layers of triangles
Since the set Sr × [0, 1] is compact and its image misses 0, there is a positive
distance from the image to the origin. The map Q is uniformly continuous, so that,
if the triangles are small, so are the images of those triangles. We may move the
images of the vertices a tiny bit so that the images of the vertices of each triangle are
vertices of a rectilinear triangle in the plane. We replace each (possibly curvilinear
and singular) image triangle by the corresponding rectilinear triangle spanned by
the new image vertices.
We take care in the adjustment that the following conditions are satisfied. The
vertex images from Sr × {0} are not moved. The image of the entire set Sr × {1}
is still very small. No adjusted triangle hits the origin 0.
1.3. SECOND PROOF 9
We now pick a ray R from 0 to ∞ in the plane in such a way that it misses all
of the new vertex images and misses the (very small) image of Sr × {1}. We lose
no generality in assuming that R is the positive x-axis.
We complete the proof by counting the number of times various triangle edges
intersect and cross the ray R. We orient the edges of Sr ×{0} in the counterclockwise
direction. The image of Sr ×{0} intersects R exactly n times, and each time crosses
R from the lower side to the upper side. We assign the crossing number +1 to each
oriented edge that crosses R from the bottom to the top. When we later come
upon oriented edges that cross R from top to bottom, we assign such crossings the
crossing number −1. See Figure 7. If an edge misses R, we assign it number 0.
The sum of crossing numbers for the image of Sr × {0} is n.
We delete triangles one at a time, layer by layer, as indicated by Figure 6. We
first remove the green triangles, then the orange triangles, etc.
Sr × 0
Sr × 1
Figure 6. Peeling away the layers
Each time a triangle is deleted, we obtain a new boundary curve. Either one
oriented boundary edge is replaced by two new oriented boundary edges or two ori-
ented boundary edges are replaced by one new oriented boundary edge. It is easy
to see that the algebraic sum of assigned crossing numbers remains unchanged with
each triangle deletion since the sum obviously remains unchanged on each triangle.
In order to see this, examine Figure 7.
Eventually, one moves through the diagram from the image of Sr × {0}, which
has algebraic sum n, to the image of Sr × {1}, which has algebraic sum 0, a con-
tradiction since the algebraic sum remains unchanged.
With this contradiction, the proof of the theorem is complete.
This argument completes our second proof of the Fundamental Theorem of

Algebra.
Figure 7. Calculating intersection numbers
1.4. Exercises
Abel and Galois proved that there are no formulas like the quadratic formula
for finding the roots of polynomials with real or complex coefficients of degrees
≥ 5. The Fundamental Theorem of Algebra, for which we have just given two
proofs, shows that there are always roots, in fact, if we count possible multiplicity
of roots, exactly as many roots as the degree of the polynomial. The problem
becomes one of finding the roots, or at least approximating the roots to a desired
degree of accuracy. There is a large literature concerning efficient ways of finding
the roots, some of which have been programmed into most pocket calculators. In
the following exercises, we suggest exploring one of the standard methods, namely,
Newton’s method.
1.1. (The equation of the tangent line) Let p(x) be a polynomial with real
coefficients and let xn be some real number. Find the equation of the tangent line
to the graph of p(x) at the point (xn , p(xn )).
Answer: y = p(xn ) + p (xn ) · (x − xn ).
1.2. Find the point xn+1 at which the tangent line crosses the x-axis.
Answer: xn+1 = xn − p(xn )/p (xn ).
Newton’s method: Newton notes that if xn is a root of p(x), then xn+1 = xn .
That is, a root of p(x) is a fixed point of the operation xn
→ xn+1 , and when xn is
very close to a root r of p(x), then xn+1 is much√closer to r.
1.3. Use Newton’s method to approximate 5. (That is, find an approximate
root to the equation x2 − 5 = 0.)
1.4. Apply Newton’s method in an attempt to find a root of the equation
x2 − 2x + 10 = 0. Why does the method fail?
1.5. Use the quadratic formula to find the two roots of x2 − 2x + 10.
Newton’s method often works even when there are no real roots. Use the same
iterative formula xn+1 = xn − p(xn )/p (xn ), but start the process with x0 equal to
some nonreal complex number. This requires the ability to do complex arithmetic.
1.6. Apply Newton’s method to the polynomial x2 − 2x + 10 but with x0 equal
to the complex number i.
1.7. Construct a computer program to color the plane. Starting with an initial
complex number x0 representing a pixel in the plane, iterate Newton’s process
enough times that the result seems to be getting close to one of the roots of the
polynomial and color the original pixel a color assigned to that root. If the method
doesn’t seem to converge given that initial value, leave that pixel uncolored. Iterate
the process with different initial pixels.
CHAPTER 2
The Brouwer Fixed Point Theorem
As the climax of this chapter, we will prove the Brouwer Fixed Point Theorem.
Our proof will be complete only in dimension 2, but the technique we use is valid in
all dimensions and lacks only a tiny bit of technique to complete, technique that we
will outline in exercises. This theorem is one of the two basic tools used in the proof
of existence of solutions to differential equations, the other being the contraction
mapping principle. A generalized version of the theorem is used in game theory to
prove the existence of optima.
In order to completely understand this chapter, the reader should have a rea-
sonable grasp of limits, continuity, open sets, and closed sets in Euclidean space.
2.1. Statement of the Theorem

First we recall some definitions.
Definition 2.1. The n-dimensional ball Bn in n-dimensional Euclidean space
R is the set
n
Bn = {x = (x1 , x2 , . . . , xn ) ∈ Rn : |x|2 = x21 + x22 + · · · + x2n ≤ 1}.

The boundary of the n-ball Bn is the (n − 1)-sphere Sn−1 , where
Sn−1 = {x = (x1 , x2 , . . . , xn ) ∈ Rn : |x|2 = x21 + x22 + · · · + x2n = 1}.
Thus the 1-ball is the unit interval of “radius 1” and length 2, the 0-sphere is
a pair of points, the 2-ball is the circular disk of radius 1, and the 1-sphere is the
circle of radius 1. The 3-ball is the solid ball of radius 1 in Euclidean 3-dimensional
space. The 2-sphere is the surface of that ball. Up to homeomorphism (topological
equivalence) there are many realizations of the ball and sphere. Every (finite closed)
interval is homeomorphic to the unit interval. The planar triangle is homeomorphic
to the circle, the planar triangle together with its interior is homeomorphic to the
circular disk of radius 1, etc.
Theorem 2.2 (Brouwer Fixed Point Theorem). If f : Bn → Bn is a continuous
function, then there is at least one point x ∈ Bn such that f (x) = x. Such a point
is called a fixed point of the function f .
Exercise 2.3. Prove the 1-dimensional version of the fixed point theorem: If
f : [0, 1] → [0, 1] is a continuous function, then there is at least one real number
x ∈ [0, 1] such that f (x) = x.
A version of the 1-dimensional theorem was exhibited as a puzzle in a magazine
story, related to me, that apparently appeared many, many years ago: A monk set
out at 6:00 AM to take the trail to a retreat at the top of the mountain, which
he reached early in the afternoon and where he prayed and worshiped through the
11
12 2. THE BROUWER FIXED POINT THEOREM
night. The next morning he left the mountain top, again at 6:00 AM and returned
down the trail to his monastery at the foot of the mountain. Explain why there was
a time of day and a place on the trail so that the monk was at that place on the
trail at that same time on each of the two days. [The answer to the puzzle given
in the story was “The monk had to meet himself.”]
The proof of the Brouwer Fixed Point Theorem, as well as the proof of many
theorems in topology, relies on the introduction of extra structure to the problem.
We will illustrate the value of additional structure in a series of arguments. The
first will be an elementary puzzle involving the checkerboard. The second will use a
childhood puzzle to motivate the use of the sign of a permutation; this notion will be
important when we treat the mathematical notion of right and left, or orientation
in classifying 2-dimensional surfaces. The third will show how to justify the (?)
obvious (?) fact that a polygonal closed curve in the plane separates the plane
into two pieces, an inside and an outside; this argument will introduce the powerful
topological tool of general position. The idea of general position, together with a
trick called the one-ended arc trick, will be used to prove the famous No Retraction
Theorem. Finally, we will easily deduce the Brouwer Fixed Point Theorem from
the No Retraction Theorem.
2.2. Introducing Extra Structure into a Problem

Often we understand a problem by imposing extra structure. The simplest
extra structure we can impose is that of counting. Many theorems in 2-dimensional
topology are proved by clever yet sophisticated counting arguments or by intro-
ducing extra structure. After considering these two elementary problems in this
subsection, we shall move on to some rather famous graduate level problems that
can be solved by fairly elementary, yet sophisticated, techniques that involve im-
posing extra structure and observing the geometric consequences of that structure.
Sherlock Holmes: How often have I said to you that when you have eliminated
the impossible, whatever remains, however improbable, must be the truth? - The
Sign of Four, Chapter 6.
2.3. Two Elementary Problems

Problem 2.4. I learned about this problem when I was a high school student.
It appeared in a paperback book that I read and reread. I thank that unknown
author and very interesting book. I lost the book many decades ago. Consider a
square 8x8 board of squares. See Figure 1. Remove two corner squares from the
board, diagonally across the board from each other.
Is it possible to tile the remaining squares with dominoes of the shape 1x2
(Figure 2, using those dominoes both vertically and horizontally?
We will give a solution after we describe the second problem.
Problem 2.5. Before the 3-dimensional Rubik’s Cube, there was the 2-dimen-
sional Fifteen Puzzle.
The puzzle consists of fifteen tabbed and slotted plastic tiles set in a plastic
case, with one blank slot (pictured as black) into which adjacent tiles can be slid.
See Figures 3 and 4.
2.3. TWO ELEMENTARY PROBLEMS 13
Figure 1. The uncolored board with two corners deleted
Figure 2. The uncolored domino
The goal was to transform a scrambled arrangement of tiles into the unscram-
bled version. This puzzle was, for a time, the rage — just as the Rubik’s Cube had
its period — and the puzzle is still manufactured and sold as a children’s diversion.
I played with the puzzle at church when I was a child when I was bored with the
sermon. Puzzle enthusiasts began to ask whether all arrangements of tiles could
be realized. In particular, they tried unsuccessfully to transform the standard ar-
rangement into the reverse arrangement. See Figure 5.
Books were written about the puzzle. A museum at the Indiana University,
Bloomington, collects and displays the wonderful artifacts concerning this puzzle.
Why, or why not, can the reverse position be realized?
2.3.1. Solution to the Tiling Problem. First approximation to a solution

for the domino tiling problem. The board with corners deleted has 62 squares,
which is even. The domino has 2 squares, which is even. Thus, any solution would
require exactly 31 = 62/2 dominoes.
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15
Figure 3. The 15 Puzzle
1 2 3 4
5 6 8
9 10 7 11
13 14 15 12
Figure 4. Tile slides
Second approximation for the tiling problem. Introduce extra structure to both
the board and to the domino by coloring every other square black. See Figures 6
and 7.
Solution to the tiling problem: Every domino, whether horizontally placed or
vertically placed, will cover one white and one black square. Thus, if the tiling
exists, there must be the same number of black and white squares. But there are
32 black squares and only 30 white squares. Therefore the tiling is impossible.
2.3.2. Solution to the 15 Puzzle. Solution to the 15 puzzle: As a matter

of fact, exactly half of the potential arrangements of tiles can be realized, and the
situation is explained by a simple, but useful, mathematical notion, namely the
15 14 13 12
11 10 9 8
7 6 5 4
3 2 1
Figure 5. The 15 puzzle with numbers reversed
Figure 6. The colored board with two corners omitted
Figure 7. The colored domino
sign or parity of a permutation — is a permutation even (plus sign) or odd (minus

sign)?
Definition 2.6. Let Sn = {1, 2, . . . , n} denote a finite set with n elements. A
permutation of Sn is a reordering p of the set. Equivalently, a permutation is a
bijection p : Sn → Sn .
There are standard ways of picturing a permutation p. For example, the 2-row
form
1 2 3 4 5 6 7 8
p=
7 2 4 3 1 8 5 6
exhibits each i above its image p(i), so that, for example, p(1) = 7 and p(7) = 5.
The disjoint cycle form
(1 7 5)(2)(3 4)(6 8)
indicates that p naturally divides the elements of Sn into circles or cycles:
p : 1
→ 7
→ 5
→ 1,
p : 2
→ 2,
p : 3
→ 4
→ 3,
p : 6
→ 8
→ 6.
The cycle (1 7 5) is called a 3-cycle; the cycle (2) is a 1-cycle, and the cycles (3 4)
and (6 8) are 2-cycles. The 1-cycles of a permutation are often omitted in the
disjoint-cycle form. A 2-cycle is also called a transposition.
Each cycle can itself be considered as a permutation (a function), where all
elements of Sn not explicitly mentioned are to be considered as (omitted) 1-cycles.
Since functions are typically composed from right to left, cycles (as functions) can
also be composed, whether they are disjoint or not, to form new permutations. For
example,
(1 3 5 7)(2 5 3) = (2 7 1 3)(5).
Theorem 2.7. Every permutation of Sn can be realized as a product of 2-cycles
(transpositions).
Proof. It suffices to show that a cycle (a1 a2 . . . ak ) is a product of 2-cycles,
but
(a1 a2 . . . ak ) = (a1 ak )(a1 ak−1 ) · · · (a1 a3 )(a1 a2 ).

Definition 2.8. A permutation is even if it can be expressed as the product
of an even number of 2-cycles. A permutation is odd if it can be expressed as the
product of an odd number of 2-cycles.
Definition 2.9. Suppose that p is a permutation of Sn . Consider the un-
ordered pairs (i, j). We say that (i, j) is inverted by p if p changes the order of
i and j; that is, (i, j) is inverted by p if either (i < j and p(j) < p(i)) or (j < i
and p(i) < p(j)). Let inv(p) denote the number of unordered pairs (i, j) that are
inverted by p.
Here is an example, namely the permutation

1 2 3 4 5 6 7
.
5 1 6 3 7 4 2
If we draw arrows from each integer to its image spot, then the number of inversions
is the number of crossings of these arrows. See Figure 8.
Theorem 2.10. The permutation p is even iff inv(p) is even. The permutation
p is odd iff inv(p) is odd. (Consequently, no permutation is both even and odd.)
1 2 3 4 5 6 7
5 1 6 3 7 4 2
Figure 8. The inversions of a permutation
Proof. It suffices to show that each 2-cycle changes the number of inversions
by an odd number. We start with the 2-row representation

1 ... i (i + 1) . . . (k − 1) k ... n
p(1) . . . p(i) p(i + 1) . . . p(k − 1) p(k) . . . p(n)
and compose with the transposition (p(i) p(k)), with i < k. The result, after the
composition is

1 ... i (i + 1) . . . (k − 1) k ... n
.
p(1) . . . p(k) p(i + 1) . . . p(k − 1) p(i) . . . p(n)
If k = i + 1, so that ai and ak are adjacent before the move, then the only
pair whose status of inversion is affected by the transposition is the pair (i, k): if
this pair is inverted before the transposition, then it is in order after; if it is in
order before the transposition, then it is inverted after. Thus the parity of inv(p)
is changed by this adjacency transposition. If there are = k − i − 1 elements
between ai and ak , then adjacency moves can move ai so that it is in position
k − 1 adjacent to ak , with changes in parity. Then ai and ak can be interchanged,
with a change of 1 in parity. Finally, ak can be moved back to position i by another
adjacency switches, with changes in parity. The result is 2 + 1 parity changes,
an odd number of changes. Hence, the original transposition resulted in a change
in the parity of inv(p).
We are now in the position to show that the reverse position cannot be reached
from the initial position.
We think of the blank slot as a tile labelled 16. We can view each position of
the puzzle as a permutation of the numbers 1 through 16. Each move in the puzzle
is a transposition that interchanges 16 with another tile. Hence, each move changes
the parity of the permutation. The tile 16 begins in a position in the checkerboard
pattern that is dark. Each move takes 16 from a dark position to a light position, or
vice versa. Hence, the position of 16 in the checkerboard pattern indicates whether
the permutation is even or odd: dark position = even permutation; light position
= odd permutation.
In the reverse position, tile 16 is in dark position (Figure 9); hence the permu-
tation, if attainable, must be even. We count the number of inversions. (Compare
with the original position, Figure 10. The number 15 is inverted with fourteen tiles,
the number 14 with 13 tiles, the number 13 with 12 tiles, etc. That is, the number
15 14 13 12
11 10 9 8
7 6 5 4
3 2 1
Figure 9. The reversed position again
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15
Figure 10. The initial position again
of inversions is
14 + 13 + 12 + 11 + 10 + 9 + 8 + 7 + 6 + 5 + 4 + 3 + 2 + 1,
which is odd. Thus this position is not attainable.
Exercise. Show that a tile position can be attained iff the parity of the permu-
tation agrees with the parity dictated by the position of tile 16.
2.4. Three Advanced Problems

We give a sequence of applications to geometry, all referring to important prop-
erties of the 2-dimensional plane R2 . We are interested in the ways a circle, which
2.4. THREE ADVANCED PROBLEMS 19
we denote by S1 , can be placed in the plane. For the remainder of this section,
it is necessary to know about continuous functions.
2.4.1. Polygonal Simple Closed Curves in the Plane.
Definition 2.11. A simple closed curve in the plane is the image of a contin-
uous function J : S1 → R2 from the circle S1 into the plane R2 , where the function
J is one to one (that is, has no self-intersections).
For the next few theorems, we shall assume that the simple closed curve is
polygonal, that is, formed by a finite sequence of straight line segments. Figure 11
gives two examples, the first very simple, the second rather elaborate.
u
v
Example 1
Example 2
Figure 11. The simple and the complex simple closed curve
For Example 1, the triangle, there is evidently both an inside and an outside
to the curve. This is likewise true for a round circle. It is intuitively obvious that
every convoluted “circle” (simple closed curve) in the plane should have an inside
and an outside, but the proof that it is so becomes complex as we observe Example
2 and ask ourselves whether the points u and v that we have marked are inside the
curve or outside the curve.
Exercise 2.12. Is the point v inside the curve? The point u?
For a polygonal simple closed curve, there are at most two pieces in the com-
plement R2 \ J of J, for, locally, there are only two sides (Figure 12), and, as one
traverses the curve, the local sides near one point are connected to local sides of
surrounding points.
The question is only whether, when we follow those two local sides around
the full length of J, does the black Side 1 come back to connect again with the
Side 1, or does the black Side 1 come back to connect with grey Side 2 to form a
twisted Möbius band. We will give a proof of the following theorem. Although the
theorem seems obvious, the proof is not trivial and the techniques of its proof will
be used to prove the very important No Retraction Theorem and Brouwer Fixed
Point Theorem.
Side 1
Side 2 J
Figure 12. The two local sides of a polygonal simple closed curve
Theorem 2.13. Each polygonal simple closed curve J in the plane R2 separates
the plane. That is, the complement R2 \ J of J in R2 is not connected. In fact, it
has exactly two pieces (called components), an inside (called the interior) and the
outside (called the exterior).
Remark. The theorem is completely obvious for the triangle, and our proof
will make use of that fact. The proof that we will give is elementary in the sense
that it uses only basic geometric ideas, but the logic involved invokes ideas that,
in fully developed form, are very important in the development of geometric and
algebraic topology. We will try to point out the critical ideas as we go along. End
remark.
Here are a few of the basic ideas:
Idea (1) A triangle T in the plane has both an inside and an outside. Every
time you cross the boundary of the triangle you pass from outside to inside, or from
inside to outside. See Figure 13. If we traverse a polygonal simple closed curve
J that crosses ∂T at every intersection point, then every crossing point where J
crosses into T is followed by a subpath of J in J ∩ T that is terminated by a point
where J crosses the boundary ∂T out of T .
Note the 5 pairs of crossings.
Figure 13. Curve crossing a triangle

The components (pieces) of J ∩ T pair the points of J ∩ ∂T .
Idea (2) If T1 , T2 , . . ., Tn are triangles in the plane R2 , and if J is a polygonal

simple closed curve in the plane, then J can be translated an arbitrarily small
amount so that it misses each vertex of each Ti and crosses the boundary of each
Ti at each point of J ∩ ∂Ti . This is a very simple case of an exceedingly important
property called general position.
Idea (3) Every finite edge path that has at least one end point has, in fact, two
end points. The only other alternative for the path is that it form a closed path
(initial point = terminal point) with no end points.
Proof (theorem). The proof is called the one-ended arc trick. We will show
that, unless the theorem is true, there is a finite edge path that has only one
endpoint, in contradiction to Idea (3) above. We therefore begin the proof by
assuming the ridiculous fact that the curve J does not separate the plane.
The proof requires that we introduce extra structure in our picture. This struc-
ture will take place in two different copies of the plane. We call the plane that
contains our polygonal simple closed curve J the image plane. The other plane
that we introduce we will call the model plane.
If our curve J contains n vertices, then we shall first construct n model triangles
in the model plane and then n corresponding image triangles in the image plane.
See Figures 14 and 15. After we have constructed the image triangles, we will
construct a second polygonal simple closed curve K in the image plane that is
in general position with respect to each of the n image triangles. See Figure 16.
Corresponding to each intersection arc of K with an image triangle, there is a
corresponding arc in the model triangle. It is in this collection of model arcs that
we will discover a one-ended arc, an obvious contradiction. We will conclude that
J must separate the image plane. (I find such an argument completely wild! I love
it.)
We first introduce the triangles in the model plane. If our first polygonal simple
closed curve has n vertices, then we pick n vertices on the unit circle S1 in the model
plane and join them by straight arcs to form a polygonal simple closed curve. We
then add a radius of the circle from each vertex A to the center V of the circle. We
obtain a model array of triangles ABV inside the disk of unit radius. It is in this
model that we will trace certain important paths:
S1
B
V
A
Figure 14. The model disk

We map this model disk of triangles into the plane in the following way. We
map the outer boundary of the model to our original polygonal simple closed curve,
vertex to vertex, edge to edge. We map the center V of our model to some rather
arbitrary point V in the plane, subject only to the condition that it not be in any
of the lines defined by the edges of J. Then, given any edge e of J with vertices A
and B , the points A , B and V are the vertices of a triangle A B V in the plane.
The corresponding points A, B, and V of our model bound a triangle ABV in our
model. We map the image triangle A B V to the model triangle ABV linearly.
B
A
Z
V
Figure 15. The image of the model disk
While the model triangles intersect only along edges, the image triangles do
intersect along the corresponding common edge but may also fold back over each
other along that edge. For example, the triangles A B V and Z A V fold over each
other in the figure.
We have now completed the introduction of triangles into our problem. We
have still one additional structure to add to our picture. This is the addition of a
second polygonal simple closed curve in the image plane.
Since we are assuming that J does not separate the image plane, it does not
separate points just opposite each other across the arc A B . We may thus construct
a polygonal simple closed curve K in the image plane that intersects J only at one
point X ∈ A B where it crosses J.
This construction completes the added structure to our problem: n triangles
in the model plane, n triangles in the image plane, and a polygonal simple closed
curve K in the image plane.
2.4.2. The One-sided Arc Trick. We now examine the way K intersects
each of the image triangles. After a slight translation, we may assume that K misses
all of the image vertices and crosses each triangle edge at each point of intersection
(general position, Idea (2)).
Thus by Idea (1), if K intersects a triangle, it does so in a finite collection
of polygonal arcs, each with its endpoints on the boundary of the triangle. The
corresponding model triangle likewise is crossed by a corresponding finite collection
of polygonal arcs (Figure 17), pushed over from the image triangle by the linear
correspondence between the two triangles.
B
X
A
Z
Figure 16. The curve K
Image Model
B
X
A
V
Figure 17. Pull-back arcs in the model disk
The important property of these crossing arcs is the following. Whether adja-
cent triangles in the image plane fold back on each other or not, any intersection
arc from the one triangle that meets the common edge extends into the neighboring
triangle at that point. If the triangles did not fold on each other, then the arc is
simply an extension into the adjacent triangle. If the triangles did fold on each
other, then the arc enters the common edge in one triangle and folds right back on
top of itself in the adjacent triangle.
As a conseqence, in the model plane the intersection arcs form nonsingular
curves, which can only be finite edge paths or polygonal simple closed curves.
The key observation is this: Look at the intersection path in the model that
begins at the point X. Because X is an endpoint of that path, the path must have
a second endpoint by Idea (3). That endpoint must lie on a triangle edge and that
edge cannot be an interior edge of the model because intersection paths continue at
every interior edge. But no other boundary point is available since X is the only
point at which K hits a boundary edge.
We have discovered a one-ended arc, a contradiction.
2.4.3. The Brouwer Theorem and the No Retraction Theorem. The

theorem we have just proved is both “obvious” and “uninteresting”—obvious be-
cause we have a hard time imagining that it could be false, and uninteresting
because there is a much more general theorem that does not assume the simple
closed curve is polygonal. Much of the content in a beginning algebraic topology
course was developed in order to prove the general case of this theorem and other
related facts. One generalization is the Riemann Mapping Theorem in the theory
of complex variables. Riemann’s proof had some gaps that required about 50 years
of effort by a cadre of mathematicians to fill. Riemann’s theorem is an analytic gen-
eralization of this theorem. We will return to general versions of this theorem using
the techniques of geometric, set theoretic, and algebraic topology in later chap-
ters. At that point we will prove theorems with famous names: the Jordan Curve
Theorem, the Schoenflies Theorem, the Zippin Characterization of the Sphere, the
R. L. Moore Decomposition Theorem, and others.
But at this point we want to enjoy the beauty of the one-ended arc proof for a
while and use it to prove other results. I remember asking R. H. Bing what he could
possibly do with a certain unresolved problem called the Free Surface Conjecture
if he could resolve it. He said, “Oh, I think that I’d enjoy it for a while.”
Here are a couple of exercises for the reader. The first is rather direct from what
we have just done. The second is a generalization to dimension 3 and requires some
imaginative thought. In particular, it requires that we understand what general
position would mean in dimension 3, and that we understand how two triangular
disks in dimension 3 could be expected to intersect.
Exercise 2.14. Prove that there is no polyhedral Möbius strip in the Euclidean
plane.
Exercise 2.15. Show that every polyhedral 2-dimensional surface S in 3-

dimensional Euclidean space R3 separates R3 .
It requires less imagination to prove the 2-dimensional version of the famous

Brouwer Fixed Point Theorem.
Theorem 2.16 (Brouwer Fixed Point Theorem). If D is the unit square disk,
and f : D → D is a continuous function that maps D into itself, then there is at
least one point x ∈ D such that f (x) = x.
We will deduce the 2-dimensional Brouwer Fixed Point Theorem from the 2-
dimensional version of an equally famous theorem called the No Retraction Theo-
rem. We will prove the No Retraction Theorem by the one-ended arc trick.
Theorem 2.17 (No Retraction Theorem). If Bn is the unit n-dimensional ball
(the ball of radius 1), then it is impossible to find a continuous function f : Bn →
(Sn−1 = ∂Bn ) such that, for each x ∈ Sn−1 , f (x) = x. Such a map, if it existed,
would be called a retraction.
Proof of the No Retraction Theorem in dimension 2. In place of the
round disk B2 we use the square disk D to which it is topologically equivalent. As in
our previous proof using the one-ended arc trick, we make the ridiculous assumption
that the theorem is false. Hence there is a continuous function f : D → ∂D such
that, for each x ∈ ∂D, f (x) = x.
Again, we consider two planes: the model plane and the image plane. In the
symbol f : D → ∂D, we view the first D to be in the model plane and the second
D to be in the image plane. As before, we add structure to the model disk D by
subdividing it into very tiny triangles. See Figure 18.
Figure 18. The model triangulated square
We now use the continuous function f to carry the triangles of the square in
the model plane over to tiny triangles in the image plane. If T = ABC is any of the
tiny triangles, with vertices A, B, and C, then the images f (A), f (B), and f (C)
will be three points of ∂D that are very close to one another. We may move them
a tiny bit so that they are still near one another and still near ∂D, yet they are the
vertices of an actual triangle in the image plane. This process does not require that
we move any vertices that were originally in ∂D. In this manner, we modify all
those image vertices that must be moved. We may replace the original map f with
a new continuous function g : D → R2 that takes each triangle of the model space
linearly to the corresponding triangle of the image space. The new map perhaps
does not take D into ∂D, but it certainly takes D into the complement of the center
of the disk and does not move any point of ∂D.
We now have to add to the structure a path R analogous to the polygonal
simple closed curve K in the separation theorem. For R we take an infinite ray
that begins at the center of D in the image plane and misses every image vertex
under the map g, and passes onward to infinity. Then R either misses an image
triangle g(T ) entirely or else intersects g(T ) precisely in an arc that enters g(T )
through one edge and exits g(T ) through another edge. See Figure 19.
g(B)
g(A)
g(C)
Figure 19. The intersection of a ray with a triangle
The intersection arc can be carried back to an intersection arc in the model
triangle T = ABC by the linear correspondence between T and g(T ).
As in the previous proof, we concentrate on the union of the intersection arcs
as viewed in the model disk. As before, this union must be a finite collection of
finite edge paths, each having two endpoints, and a finite collection of polygonal
simple closed curves. As before, we consider that component that contains the
only intersection point of R with ∂D. That component must be an arc with two
endpoints by Idea (3). The second endpoint can only lie on the boundary of D, and
there is no other intersection with ∂D, a contradiction. That is, this component is
a one-ended arc.
As before, we conclude that the theorem is true, that our original assumption
was truly ridiculous.
Proof of the Brouwer Fixed Point Theorem. Since the round disk E
and the square disk D are topologically equivalent, we may replace D by E in both
the No Retraction Theorem and the Brouwer Fixed Point Theorem. We assume
2.5. EXERCISES 27
that, contrary to the latter, there is a continuous function g : E → E such that, for
each x ∈ E, g(x) = x. We then define a function f : E → ∂E as follows.
Consider the ray R(x) that begins at g(x) in E and passes through x on its
way to ∞. Define f (x) to be the last point of E in the ray R(x). Then the function
f : E → ∂E is a continuous function that fixes each point of ∂E, in contradiction
to the No Retraction Theorem.
2.5. Exercises
2.1. Solve Exercise 2.3 on page 11.
2.4. Generalize the No Retraction Theorem to dimension 3. Consider the unit
cube. Show how to divide the unit cube into tiny tetrahedra. View these tetrahedra
as lying in the model cube. Show how to carry those tetrahedra to tiny tetrahedra
in the image cube. Determine what the appropriate general position property is for
the intersection of a ray with a tetrahedron. Pull the intersection segments back
into the model cube. Show that you obtain a one-ended arc.
2.5. Contemplate what would have to be done to prove a No Retraction Theo-
rem in every dimension. Think how you would deduce a generalized Brouwer Fixed
Point Theorem in that dimension.
The Brouwer Fixed Point Theorem does not suggest a method for finding a
fixed point. If the map f : B2 → B2 is a contraction mapping, a particularly nice
property that we will now explain, then it is an easy matter to approximate a fixed
point. (Contraction mappings defined on other suitable spaces are one of the main
tools used to find solutions to differential equations.)
Definition 2.18. A map f : B2 → B2 is a contraction mapping if there is
a number 0 < λ < 1 such that, for each two points x, y ∈ B2 , the distances
d(f (x), f (y)) from f (x) to f (y) and d(x, y) from x to y satisfy the inequality
d(f (x), f (y)) ≤ λ · d(x, y).
In the following exercise you will need to use one of the fundamental properties
of a compact metric space: Every Cauchy sequence converges. (Recall that a
sequence x1 , x2 , . . . is a Cauchy sequence if, for each > 0, there is an integer N
such that n, m ≥ N implies d(xn , xm ) < .)
2.6. Prove that, if f : B2 → B2 is a contraction mapping, then f has a unique
fixed point x0 . If x1 ∈ B2 and if xn+1 = f (xn ) for each n ≥ 1, then the sequence
x1 , x2 , . . . is a Cauchy sequence converging to x0 .
2.7. Show that a tile position can be attained iff the parity of the permutation
agrees with the parity dictated by the position of tile 16.
2.8. Consider puzzles like the 15 puzzle, but with dimensions m × n. Which
configurations can be realized by such a puzzle?
2.9. Suppose that J is a polygonal simple closed curve in the plane bounding
a disk D. Show that D can be divided into triangles using only the vertices of J as
vertices of the triangles used.
2.10. By induction on the number of triangles used in the previous exercise,
show that D is homeomorphic to a single triangle T . (That is, you must prove the
existence of a continuous bijection f : D → T .)
CHAPTER 3
Tools
We take time to introduce three tools that are interesting in and of themselves.
The most useful additional structure added to problems in geometric topology
is the polyhedral complex; complexes are essentially used as finitely describable
approximations to spaces that have continuous structure.
The space being approximated can often be mapped into the finite polyhedral-
complex approximation by means of either Urysohn’s Lemma or the Tietze Exten-
sion Theorem.
Sequences of simple spaces often have interesting limit spaces. The limiting
process is described by Set Convergence.
3.1. Polyhedral complexes

In the last chapter we talked about the power of using additional structure in
the solution to a problem. In that chapter and in the previous one, we introduced
the extra structure of disks formed from triangles. Such things are often the type of
extra structure we employ. We formalize the concepts now for later use. In partic-
ular, we want to define convex sets, simplexes, vertices, faces, simplicial complexes,
polyhedra, triangulable spaces, and triangulations.
One of our major goals in this book is the proof that 2-dimensional surfaces
are triangulable.
Definition 3.1. As we see from Figure 1, if we begin with vectors A and B,
then we obtain the points of the segment [A, B] from A to B by adding to the
vector A a multiple λ · (B − A) of the vector B − A, with λ ∈ [0, 1]. That is, the
interval [A, B] from A to B is given by the formula
[A, B] = {x = μA + λB | μ, λ ≥ 0, μ + λ = 1}.
A set X is said to be convex if, for each A, B ∈ X, the interval [A, B] is also in
X.
Exercise 3.2. If X is any subset of Rn , then there is a smallest convex set
containing X.
If X is any set, then the smallest convex set H(X) containing X is called the
convex hull of X.
Exercise 3.3. The line (infinite in both directions) containing vectors a, b
consists of the sums
λ1 a + λ2 b,
where λ1 + λ2 = 1.
29
30 3. TOOLS
A + λ · (B − A) = (1 − λ) · A + λ · B
Figure 1. Parametrizing the interval
Exercise 3.4. If a, b, c are given vectors, then the convex hull of the set {a, b, c}
consists of the sums
λ1 a + λ2 b + λ3 c,
where each λi ≥ 0 and λ1 + λ2 + λ3 = 1.
Exercise 3.5. The smallest 2-dimensional plane in 3-dimensional space con-
taining (noncollinear) vectors a, b, c consists of the sums
λ1 a + λ2 b + λ3 c,
where λ1 + λ2 + λ3 = 1.
The simplest convex sets are the simplexes. A 1-simplex is a closed interval. A
2-simplex is a triangle (together with its interior). A 3-simplex is a solid tetrahe-
dron.
Definition 3.6. Let O denote the origin in Rn . Let v1 , v2 , . . ., vn denote n
linearly independent vectors. Then the convex hull

σ = H({O, v1 , v2 , . . . , vn }) = {λi vi | li ≥ 0 and λi ≤ 1}
i>0 i>0
is called an n-simplex.
If w0 , w1 , w2 , . . ., wn are any n + 1 vectors, then we may map σ linearly onto
the convex hull H = H({w0 , w1 , w2 , . . . , wn }) by the formula
f (λ1 v1 + λ2 v2 + · · · + λn vn ) = λ0 w0 + λ1 w1 + · · · + λn wn ,
3.2. URYSOHN’S LEMMA AND THE TIETZE EXTENSION THEOREM 31

where λ0 = 1 − i>0 λi . If the map f is 1 to 1, so that it is a homeomorphism,
then f (σ) is also called an n-simplex.
The vectors w0 , w1 , . . ., wn are called the vertices of the simplex. The faces of
the simplex are the convex hulls of the subsets of the vertices. Consequently, the
endpoints of a 1-simplex are faces, as are the edges and vertices of a 2-simplex, and
the 4 triangular faces, the 6 edges, and the 4 vertices of a 3-simplex, and, of course,
the simplex σ is a face of itself.
Definition 3.7. A simplicial complex K is a collection of simplexes such that,
if two of the simplexes intersect, then they do so in a face of each. The union
|K| of the simplexes of K is called the carrier of K. We assign a topology to |K|
by declaring a subset to be closed (open) if its intersection with each face of each
simplex of K is closed (open) in that face.
Definition 3.8. We say that a space X is a polyhedron if it is the carrier of a
simplicial complex. We say X is triangulable if it is homeomorphic to a polyhedron.
A topological triangulation of X is a homeomorphism h : |K| → X from the carrier
of some simplicial complex K.
3.2. Urysohn’s Lemma and the Tietze Extension Theorem

In the next chapter we will see how Urysohn’s Lemma can be used to create a
family of functions called a partition of unity and how partitions of unity allow us
to map spaces into polyhedral complexes. These theorems allow us to approximate
very general spaces by polyhedral spaces. There is a companion theorem, a gener-
alization of Urysohn’s Lemma, called the Tietze Extension Theorem. Our intent is
to prove very general versions of these two powerful tools in this chapter.
For example, the Tietze Extension Theorem shows us how to retract all of
Euclidean 3-dimensional space R3 onto the infinitely knotted arc K indicated by
Figure 2. It is not at all obvious how to construct such a function. Below the figure,
we suggest the construction of such a function as a puzzling exercise.
Figure 2. The infinitely knotted arc
Exercise 3.9. Construct a continuous function f : R3 → K such that, for each

point x ∈ K of the arc K, f (x) = x.
We first state and prove Urysohn’s Lemma for metric spaces.
32 3. TOOLS
Theorem 3.10 (Urysohn’s Lemma for metric spaces). Suppose that X is a

metric space and that A and B are disjoint closed subsets of X. Then there is a
continuous function f : X → [0, 1] such that f (A) = {0} and f (B) = {1}.
Proof. Define
d(x, A)
f (x) = .
d(x, A) + d(x, B)
(We use d(x, y) to denote the distance from x to y.)
Corollary 3.11. If X is a metric space, then X has the topological property
called normality: If A and B are disjoint closed subsets of X, there exist disjoint
open sets U and V such that A ⊂ U and B ⊂ V .
Proof. If f : X → [0, 1] is continuous with f (A) = {0} and f (B) = {1}, then
we may take U = f −1 ([0, 1/2)) and V = f −1 ((1/2, 1]).
In proving Urysohn’s Lemma for metric spaces, we have seen how easy it is
to prove that such spaces have the topological property called normality. It is
therefore not a surprise that normality should become a central hypothesis in the
general forms of Urysohn’s Lemma and the Tietze Extension Theorem.
Definition 3.12. A space X is normal if, given any two disjoint closed subsets
A and B, there exist disjoint open sets U and V such that A ⊂ U and B ⊂ V .
Theorem 3.13 (Urysohn’s Lemma). Suppose that X is a normal space and
that A and B are disjoint closed subsets of X. Then there is a continuous function
f : X → [0, 1] such that f (A) = {0} and f (B) = {1}.
Theorem 3.14 (Tietze’s Extension Theorem). Suppose that X is a normal
space, A ⊂ X is closed, and f : A → [0, 1] is continuous. Then there is a continuous
function F : X → [0, 1] such that, for each a ∈ A, F (a) = f (a).
We first prove Tietze’s Extension Theorem in the special case where X is a
metric space. Then we prove both theorems in the general case.
Proof of Tietze’s Extension Theorem when X is a metric space. The

Tietze theorem is more complicated than Urysohn’s Lemma so that we have to
proceed by approximation. The statement of the proof is simplified slightly if we
assume that the map f : A → [0, 1] is onto. That guarantees that each of the sets
we define is nonempty.
We define two closed sets
X(1, 1) = {x ∈ X | d(x, f −1 ([0, 1/2])) ≤ d(x, f −1 ([1/2, 1]))}
and
X(1, 2) = {x ∈ X | d(x, f −1 ([1/2, 1])) ≤ d(x, f −1 ([0, 1/2]))}.
In the end, it will be natural to define F so that F (X(1, 1)) = [0, 1/2] and F (X(1, 2))
= [1/2, 1]. In particular, we may begin by defining F (X(1, 1) ∩ X(1, 2)) = 1/2.
We assume inductively that the interval [0, 1] has been divided into 2n successive
subintervals I(n, 1), I(n, 2), . . ., I(n, 2n ) of equal length and that we have deter-
mined closed sets X(n, k) that are to be mapped by F to the interval I(n, k). We
assume that F has already been defined not only on X(n, k) ∩ A but also on the in-
tersections X(n, k−1)∩X(n, k) so that F takes the intersection X(n, k−1)∩X(n, k)
to the single point I(n, k − 1) ∩ I(n, k).
3.2. URYSOHN’S LEMMA AND THE TIETZE EXTENSION THEOREM 33
Proceding inductively, we divide each of the intervals I(n, k) into two subin-
tervals I = I(n + 1, 2k − 1) and J = I(n + 1, 2k) of equal length. We set A(I) =
f −1 (I) ∪ (X(n, k) ∩ X(n, k − 1)) and A(J) = f −1 (J) ∪ (X(n, k) ∩ X(n, k + 1)). We
define X(I) = {x ∈ X(n, k) | d(x, A(I)) ≤ d(x, A(J))} and X(J) = {x ∈ X(n, k) |
d(x, A(J)) ≤ d(x, A(I))}. We set X(n+1, 2k −1) = X(I) and X(n+1, 2k) = X(J).
Finally, every point x ∈ X is in a descending sequence of closed sets X(1, i1 ) ⊃
X(2, i2 ) ⊃ X(3, i3 ) ⊃ · · · , and the corresponding intervals I(1, i1 ) ⊃ I(2, i2 ) ⊃
I(3, i3 ) ⊃ · · · intersect at a single point y. We define F (x) = y. It is an easy
matter to see that F is continuous and is an extension of the original map f .
The general proof of both Urysohn’s Lemma and Tietze’s Extension Theorem
involves the construction of sets like the X(i, k)’s of the metric proof of Tietze’s
Extension Theorem. For Urysohn’s Lemma, the construction can be quite loose
and easy. For Tietze’s theorem, the construction requires more care.
Proof of Urysohn’s Lemma in the general case. Since X is normal
and since A and B are disjoint closed sets, there are disjoint open sets U ⊃ A
and V ⊃ B. Let X(1, 1) = X \ V and X(1, 2) = X \ U .
Assume inductively that closed sets X(n, 1), X(n, 2), . . ., X(n, 2n ) have been
chosen whose union is X. Assume further that X(n, j) and X(n, k) are disjoint
unless |j − k| ≤ 1. Then, for each k, the sets X(n, 1) ∪ · · · ∪ X(n, k − 1) and
X(n, k +1)∪· · ·∪X(n, 2n ) are disjoint closed sets. By normality, there exist disjoint
open sets U containing X(n, 1) ∪ · · · ∪ X(n, k − 1) and V containing X(n, k + 1) ∪
· · · ∪ X(n, 2n ). Let X(n, 2k − 1) = X(n, k) \ V and X(n, 2k) = X(n, k) \ U .
As in the proof of the metric Tietze theorem, every point x ∈ X is in a de-
scending sequence of closed sets X(1, i1 ) ⊃ X(2, i2 ) ⊃ X(3, i3 ) ⊃ · · · , and the
corresponding intervals I(1, i1 ) ⊃ I(2, i2 ) ⊃ I(3, i3 ⊃ · · · intersect at a single point
y. We define F (x) = y. It is an easy matter to see that F is continuous and is an
extension of the original map f .
Proof of Tietze’s Extension Theorem in the general case. The proof

is exactly like that of the general case of Urysohn’s Lemma except that, each time
disjoint open sets U and V are chosen, they must contain the appropriate portion
of A. More precisely, suppose X(n, k) has been chosen which corresponds to the
interval I(n, k) ⊂ [0, 1]. The interval I(n, k) = [a, c] is to be divided into two closed
subintervals I = I(n + 1, 2k − 1) = [a, b] and J = I(n + 1, 2k) = [b, c] of equal
length. Then U must contain f −1 ([a, b)), V must contain f −1 ((b, c]), and neither
may contain any point of f −1 (b). The choices we made in the metric case actually
satisfied these requirements. But in the general setting we obtain such sets U and
V by an infinite sequence of applications of the normality axiom.
Define F = X(n, k−1)∩X(n, k). Define G = X(n, k)∩X(n, k+1). In I = [a, b]
we take an increasing sequence x1 , x2 , . . . converging to the singleton {b} = I ∩ J.
In J = [b, c] we take a decreasing sequence y1 , y2 , . . . converging

∞ to the same point
b. We define open sets U (1) ⊂ U
(2) ⊂ U (3) ⊂ · · · ⊂ i=1 U (i) = U and open
∞
sets V (1) ⊂ V (2) ⊂ V (3) ⊂ · · · ⊂ j=1 V (j) = V inductively, using normality, as
follows:
The set U (1) must contain F ∪ f −1 ([a, x1 ]) and (using normality) have closure
missing G ∪ f −1 (J).
The set V (1) must contain G ∪ f −1 [y1 , c] and (using normality) have closure
missing both the closure of U (1) and the set f −1 (I).
34 3. TOOLS
In general, the set U (i) must contain the closure of U (i − 1) and f −1 ([a, xi ])
and (using normality) have closure missing both the closure of V (i − 1) and the set
f −1 (J).
Likewise, the set V (j) must contain the closure of V (j − 1) and f −1 ([yj , c]) and
(using normality) have closure missing both the closure of U (j) and the set f −1 (I).
The sets U and V so defined satisfy the requirements of the theorem. This
construction completes the proof of the Tietze Extension Theorem.
3.3. Set Convergence

Just as Urysohn’s Lemma and partitions of unity will allow us to approximate
quite general spaces by polyhedral spaces, we can take limits of rather general
spaces.
Definition 3.15. Suppose that X1 , X2 , . . . is a sequence of sets in a separable
metric space S. We define lim sup Xn to be the set of points x ∈ S such that
each neighborhood of x in S intersects infinitely many of the sets Xn . We define
lim inf Xn to be the set of points x ∈ S such that each neighborhood of x in S
intersects all but finitely many of the sets Xn . We say that X1 , X2 , . . . converges
to a set X if lim sup Xn = X = lim inf Xn . Note that, in any case, lim inf Xn ⊂
lim sup Xn .
Theorem 3.16. If X1 , X2 , . . . is a sequence of sets in a separable metric space
S, then some subsequence converges.
Remark. We were assigned this theorem as an exercise in a graduate topology
class. The only published place where I have seen it stated and proved is in [39,
Saks]. End remark.
Proof. Let U1 , U2 , . . . denote a countable basis for the topology of S. Proceed
by induction on j. For j = 0, consider the sequence
X(j) = X(0) = {X(0, 1) = X1 , X(0, 2) = X2 , . . . , X(0, n) = Xn , . . .}.
Assume that
X(j) = {X(j, 1), X(j, 2), . . .}
has been defined. If there is a subsequence X(j + 1, 1), X(j + 1, 2), . . . of X(j) that
misses Uj , let X(j + 1) be such a subsequence. Otherwise, set X(j + 1) = X(j).
We claim that the subsequence X(1, 1), X(2, 2), X(3, 3), . . . of X(0) converges.
It suffices to prove that every point x ∈ lim sup X(n, n) also lies in lim inf X(n, n).
Suppose to the contrary that some neighborhood of x misses infinitely many of
the sets X(n, n). We may take that neighborhood to be an element Um of the
countable basis. Then a subsequence of X(1, 1), X(2, 2), X(3, 3), . . . misses Um
so that the subsequence X(m) was also chosen to miss Um . However, all of the
sets X(m + k, m + k) lie in X(m) and infinitely many of them intersect Um since
x ∈ lim sup X(n, n), a contradiction.
Theorem 3.17. If the space S is compact, the sequence {X1 , X2 , . . .} of non-
empty sets converges to a set X, and each of the sets Xn is connected, then the
limit X is nonempty, closed, and connected.
Proof. Since the sets Xi are nonempty and the space S is compact, lim sup Xn
is automatically nonempty. The lim sup is always closed. It remains only to show
that X is connected. Suppose X = A ∪ B, with A and B separated. There are
3.3. SET CONVERGENCE 35
disjoint open sets U and V containing A and B, respectively. Since X = lim inf Xn ,
we may assume that every one of the sets Xn intersects both U and V . Since Xn
is connected, it must contain a point xn of S \ (U ∪ V ). Passing to a subsequence
if necessary, we may assume the sequence xn converges to a point x ∈ / U ∪ V . This
point must be in lim sup Xn = X, a contradiction.
Remark. We need not assume that the sets Xn are actually connected. It
suffices to assume that they are more and more nearly connected as n → ∞. Below
is the appropriate defintion. End remark.
Definition 3.18. A set Y is said to be δ-connected, if each two points x and
y of Y are connected by a chain x = x0 , x1 , . . . , xn = y of points of Y having the
property that the distance between xi and xi−1 is < δ for each i = 1, . . . , n.
Theorem 3.19. If the space S is compact, the sequence {X1 , X2 , . . .} of non-
empty sets converges to a set X, and, for each n, the set Xn is 1/n-connected, then
the limit X is nonempty, closed, and connected.
Proof. The limit is closed, hence compact. If X is not connected, say X =
A ∪ B, with A and B separated, then A and B lie in disjoint open sets U and V
that have disjoint closures. There is a positive distance δ between the sets U and
V . When 1/n < δ, any 1/n chain joining points of U and V must have a point xn in
the complement of U ∪ V . A subsequence converges to a point x ∈ (lim sup Xn ) \ X,
a contradiction.
Since the compact subsets of a separable metric space can be metrized (the
Hausdorff Metric) so as to form a separable metric space, the convergence theorem
has interesting applications to that space.
Definition 3.20 (Definition of the Hausdorff Metric). If X is a separable
metric space and A and B are compact subsets of X, then we define

d(A, B) = max sup{d(a, B) | a ∈ A}, sup{d(b, A) | b ∈ B} .
Exercise 3.21. The compact subsets of a separable metric space X form a

separable metric space with the Hausdorff Metric.
Exercise 3.22. The product of two separable metric spaces is separable metric.
Exercise 3.23. The space C(X, Y ) of continuous functions from a compact
metric space X into a compact metric space Y forms a separable metric space with
the sup metric
d(f, g) = sup{d(f (x), g(x)) | x ∈ X}.
Exercise 3.24. The set of embeddings Emb(X, Y ) of a compact metric space
X in a separable metric space Y is a separable metric space.
Exercise 3.25. Every uncountable subset of a separable metric space contains
a limit point of itself.
Definition 3.26. A triod is the union of three arcs that meet at a single
endpoint of each. See Figure 3.
36 3. TOOLS
Figure 3. The triod
Exercise 3.27. Suppose that F = {fα : T → R2 | α ∈ A} is an uncountable

family of embeddings of a triod T in the plane R2 . Then there is a sequence f1 , f2 ,
. . . from F that converges to an element f ∈ F, in the sense that
d(fn , f ) = sup{d(fn (x), f (x)) | x ∈ X} → 0 as n → ∞.
Remark. This result can be used in conjunction with the Jordan Curve The-
orem and Schoenflies Theorem, which we shall prove later, to prove that it is
impossible to find uncountably many disjoint triods in the plane. Disjoint triods
can simply not get close together, arc by arc. End remark.
3.4. Exercises
[Hint: Consider the function f : X → Y as the compact subset {(x, f (x)) | x ∈
X} of the separable metric space X × Y .]
The set convergence theorem has some rather surprising consequences. Certain
spaces of functions can be realized as separable metric spaces:
3.11. If F = {fα : S1 → R2 : α ∈ A} is an uncountable family of simple closed
curves in the plane, then there is an element fβ of F and sequences fαn and fβn of
elements of F so that the curves fαn converge to fβ metrically from the exterior of
the simple closed curve fβ while the curves fβn converge to fβ metrically from the
interior of the simple closed curve fβ .
3.12. [Baire Category Theorem] Suppose that C is a nonempty compact metric
space and that C is the union of countably many closed subsets C1 , C2 , . . .. Then
at least one of the sets Cn contains a nonempty open subset of C.
3.13. [The Slicing Theorem] Suppose that C is a compact subset of the disk
D = [0, 1] × [0, 1] ⊂ R2 , and suppose that every horizontal slice [0, 1] × {y} in
the disk D contains a nonempty interval [ay , by ] × {y}. Show that C contains a
nonempty open subset of D so that the dimension of D is 2.
3.14. Can the Slicing Theorem be generalized to compact subsets of the 3-
dimensional cube?
CHAPTER 4
Lebesgue Covering Dimension
Prerequisites: We assume the reader knows about metric spacea, a countable

basis for a topology, and compactness. My favorite introduction to dimension
theory is [31, W. Hurewicz and H. Wallman, Dimension Theory].
In the next chapter we shall see by examples which utilize the arc that dimension
cannot be tied to
(1) The number of parameters needed to define a space.
(2) The measure (length, area, or volume) of the space.
The most successful method of determining dimension is Lebesgue’s notion of
covering dimension.
Urysohn’s Lemma from the previous chapter will be a major ingredient in
proofs.
4.1. Definition of Covering Dimension

We restrict our definition to locally compact metric spaces with a countable
basis.
Definition 4.1. Suppose that X is a locally compact metric space with a
countable basis. Then the space X has covering dimension dim(X) ≤ n iff, for
each open covering T for X, there is a locally-finite open (or closed) cover U of X,
every element of which lies in some element of T, such that no point of U is in more
than n + 1 elements of U .
We say that X has covering dimension dim(X) = n, if it has covering dimension
≤ n and does not have covering dimension ≤ n − 1.
Remark. Since the definition is given only in terms of open sets, covering
dimension is a topological property. End remark.
Definition 4.2. A cover is locally finite if every point has a neighborhood
that intersects only finitely many elements of the cover. Covers by closed sets are
essentially useless unless they are locally finite; for if every point is a closed set,
then the cover of X by its points is a closed cover such that each point lies in only
a single element of the cover.
Exercise 4.3. Let U be an open cover of Rn each element of which has diameter
< 1. Then there is a countable subcover V each element of which intersects only
finitely many elements of V.
Solution. For each positive integer n, let Cn = {x ∈ Rn | |x| ∈ [n − 1, n]}.
The sets Cn are compact. Hence there is a finite subset Un of U that covers Cn .

from Un any elements that do not actually intersect Cn . Then we
We may remove
may set V = n Un .
37
38 4. LEBESGUE COVERING DIMENSION
4.2. Euclidean n-Dimensional Space Rn Has Covering Dimension n

Theorem 4.4 (Dimension Theorem). Euclidean n-dimensional space Rn has
covering dimension n.
Corollary 4.5. Euclidean spaces Rm and Rn are homeomorphic only if m =
n.
Since this book is devoted to 2-dimensional spaces, we shall only give the details
of the proof for n = 2. However, the outline and techniques of the proof are valid in
all dimensions. The necessary techniques in high dimensions are outlined in section
4.4.
Proof that dim(R2 ) ≤ 2. Given an open covering T of the plane, we tile R2
with triangles, each of which lies in some element of T. We may require that, if two
triangles intersect, they do so in a vertex or an edge. We call such a tiling of the
plane a triangulation of the plane.
We cover the vertices with small disjoint open sets. See Figure 1. We take care
that each of these open sets intersects only those edges and triangles that contain
that vertex.
Each edge will have its ends already covered. Hence there will be a central
subinterval in its interior which contains all of that edge not already covered. Cover
these central subintervals by disjoint open sets. We take care that each of these
open sets intersects only those triangles containing that edge and intersects only
those open sets already constructed around the endpoints of that edge.
Each triangle will have vertices and edges already covered. Hence there will be
a central disk in its interior which contains all of that triangle not already covered.
Cover these central disks by disjoint open sets. We take care that each of these
open sets intersects only those open sets already constructed corresponding to its
vertices and edges.
The resulting cover of the plane shows that dim(X) ≤ 2.
Proof that dim(R2 ) ≥ 2. We assume that, contrary to the theorem, there is

a locally finite open cover of R2 by sets of diameter ≤ 1 such that no point lies in
more than 2 elements of the cover, rather than 3 as required by the theorem. We
shall show how this open cover allows us to construct a retraction of the unit ball
B2 onto its boundary, in contradiction to the No Retraction Theorem, Theorem
2.17 on page 25.
We map the plane onto the interior of the unit disk by mapping x to x/(1+|x|).
This map is a homeomorphism with inverse y
→ y/(1 − |y|). This homeomorphism
takes our open cover of the plane to a locally finite open cover U of the interior
of the unit disk by sets with diameter approaching 0 near the boundary of the
disk and such that no point lies in more than 2 elements of the cover. We note
that U is necessarily a countable cover (exercise) and enumerate its elements U =
{u1 , u2 , . . .}.
In each of the open sets ui we choose a point xi ; and if ui and uj intersect,
we join xi and xj by the interval [xi , xj ]. Let X denote this collection of intervals.
These intervals may or may not intersect one another. Some of them may even be
a single point if xi = xj . These apparent difficulties are irrelevant to us, though in
this dimension we could, with care, avoid them. If we took care to avoid them, we
would have constructed a locally finite, but infinite, polygonal graph in the interior
4.2. EUCLIDEAN n-DIMENSIONAL SPACE Rn HAS COVERING DIMENSION n 39
Figure 1. Covering a two-dimensional complex
of B2 . We can certainly choose the points xi so that none of the intervals of X

contains the origin O, and we do that for convenience of description only.
Our next aim is to show that all of the interior of B2 can be mapped continuously
into the union of the intervals of X and that this function extends continuously to
the identity map on the circle boundary S1 = ∂B2 . Since
no interval of X contains
the origin O, we may then compose this map f : B2 → {[xi , xj ] ∈ X}∪S1 with the
obvious retraction r : B2 \ {O} → S1 : x
→ x/|x| to obtain a forbidden retraction
from B2 onto its boundary
S . This will complete the proof that dim(R ) ≥ 2.
1 2
The map int(B ) → {[xi , xj ] ∈ X} is defined by what is called a partition of

2
unity, a standard analytic tool.
Definition 4.6. Let U denote a locally finite open covering of a separable

metric space X. A partition of unity subordinate to the covering U is a family
of continuous functions fi : X → [0, 1] such that, for each i, the support of the
function fi lies in the set ui and, for each x ∈ X, i fi (x) = 1.
We will prove the existence of a partition of unity at the end of this section.
But for purposes of this proof, we assume the existence. That is, we apply the
existence for the space X = int(B2 ) and the given cover U.
If x ∈ int(B2 ), we define f (x) = fi (x) · xi . Since the support of fi lies in ui ,
fi (x) can only be nonzero for the one or two elements of U that contain x. If x is
in only one element ui , then f (x) = xi . If x is in the two elements ui and uj , then
f (x) = fi (x) · xi + fj (x) · xj , with fi (x), fj (x) ≥ 0 and fi (x) + fj (x) = 1. That is,
f (x) is a convex combination of xi and xj , hence lies in the interval [xi , xj ] ∈ X.
Since the cover U is locally finite, f (x) is locally a finite sum of continuous
functions, hence continuous.
Since the diameters of the sets ui , and also of unions ui ∪ uj with ui and uj
intersecting, go to 0 near ∂B2 , the map f extends continuously to the boundary of
B2 by the identity on S1 = ∂B2 .
This completes the proof.
4.3. Construction of Partitions of Unity

In order to prove the existence of partitions of unity, we need two lemmas,
namely a shrinking lemma and Urysohn’s Lemma.
Lemma 4.7 (Shrinking Lemma). Given a locally finite open cover U of a sepa-
rable metric space X, there is an open cover V = {vα | α ∈ A} of X such that the
closure of the set vα is a subset of uα . This cover V is called a shrinking of the
cover U.
Proof. Because the cover U is locally finite, the cover U is countable: U =
{u1 , u2 , . . .}. We need to use the fact (normality — a corollary to Urysohn’s Lemma
for metric spaces, Theorem 3.10 on page 32) that disjoint closed subsets A and B
of X are contained in disjoint open sets V and W . We add one empty element u0
to the cover U for convenience and define C0 = ∅,
v0 = ∅ and
w0 = X.
We proceed inductively. Define Ci = ui \ j<i vj ∪ k>i uk . Then there
are disjoint open sets vi and wi containing the disjoint closed sets Ci and X \ ui ,
respectively.
Because wi contains X \ ui , the closure of vi lies in ui .
The sets vi cover each x ∈ X. In order
to see this, consider the last ui to
contain x. Then
x cannot be an element of k>i uk , hence must either already be
covered by j<i vj or be in the set Ci , hence in vi .
The metric version of Urysohn’s Lemma is adequate for our purpose here. We
recall its statement.
Lemma 4.8 (Urysohn’s Lemma, metric case). Suppose that X is a metric space
and that A and B are disjoint closed subsets of X. Then there is a continuous
function f : X → [0, 1] such that f (A) = {0} and f (B) = {1}.
Proof of the existence of a partition of unity. We are given a locally
finite open cover U = {u1 , u2 , . . .} of a separable metric space X, and we seek a
partition of unity subordinate to the cover U.
By the shrinking lemma, there is an open cover V = {v1 , v2 , . . .} of X such
that, for each i, the closure of the set vi lies in ui . By Urysohn’s Lemma, there is
a continuous function gi : X → [0, 1] that is 0 on the complement of ui and is 1 on
4.4. TECHNIQUES NEEDED IN HIGHER DIMENSIONS 41
the closure of vi . We define

gi (x)
fi (x) = .
j (gj (x))
The functions fi form a partition of unity subordinate to U.
4.4. Techniques Needed in Higher Dimensions

We gave the details of the No Retraction Theorem only in dimension 2. That
must first be generalized to higher dimensions.
We described triangulations of R2 in terms of triangles. This notion has to
be generalized to higher dimensions. If we represent the n-dimensional ball Bn
topologically by the n-dimensional cube [0, 1]n , then it is easy to subdivide it into
tiny cubes by subdividing each factor [0, 1]. Each of the tiny cubes can be subdi-
vided into tiny n-simplexes by coning each face from an interior point, dimension
by dimension: divide edges in the middle; then cone each 2-dimensional face from
a point in its interior to subdivide it into triangles; then cone each 3-dimensional
face from an interior point to subdivide it into tetrahedra; etc. Having divided the
cube into tiny n-dimensional simplexes, a putative retraction onto its boundary can
be approximated by maps on the tiny simplexes. From that point on, the proof is
very much like that of the 2-dimensional case.
Exercise 4.9. See to the details of the No Retraction Theorem in dimension
n, if you can.
Exercise 4.10. Given an open cover U of a metric space X such that no
element of X is covered by more than n elements, construct a simplicial complex
as follows:
For each u ∈ U, let v(u) be an abstract vertex. If elements u0 , u1 , . . ., uk
intersect, then construct an abstract k-simplex with vertices v(u0 ), v(u1 ), . . ., v(uk )
by identifying the points of that k-simplex with expressions

{λ0 · v(u0 ) + λ1 · v(u1 ) + · · · + λk · v(uk ) | λi ≥ 0, λi = 1},
as if this were the convex hull of the vertices.
The collection of abstract simplexes thus formed create a simplicial complex,
called the nerve of the cover U.
Show that a partition of unity subordinate to the cover U defines a continuous
function from X into the nerve.
Once those two things have been done, the proof of the Dimension Theorem
proceeds without essential change.
Here are some beginnings of these generalizations:
In large measure, this definition of covering dimension is designed to generalize
the dimension of a simplicial complex.
Definition 4.11. A complex K is a locally finite union of simplexes in some
Euclidean space Rm subject to the condition that, if two simplexes σ and τ of the
complex intersect, then they do so in a face of each. The dimension of K is the
largest k such that K has a k-simplex.
Theorem 4.12. A k-dimensional complex K has covering dimension ≤ k.
Proof outline. We have already essentially proved the theorem for k = 2 in

showing that the dimension of R2 is 2. The reader should review that proof and
use it to prove the theorem at least for complexes of dimensions 0, 1, and 2. Here
is a repeat of that outline:
Given > 0, we need to cover the complex K with open sets of diameter ≤ .
Hence we subdivide the simplexes of K into smaller simplexes, each of diameter
< .
We repeat the appropriate figure in Figure 2. We cover the 0-simplexes with
small disjoint open sets. We take care that each of these open sets intersects only
those simplexes of which the given 0-simplex is a face.
Each 1-simplex will have its ends already covered. Hence there will be a central
subinterval in its interior which contains all of that 1-simplex not already covered.
Cover these central subintervals by disjoint open sets. We take care that each of
these open sets intersects only those simplexes of which the given 1-simplex is a
face and intersects only those open sets already constructed corresponding to its
faces.
Each 2-simplex will have its boundary edges already covered. Hence there will
be a central disk in its interior which contains all of that 2-simplex not already
covered. Cover these central disks by disjoint open sets. We take care that each
of these open sets intersects only those simplexes of which the given 2-simplex is
a face and intersects only those open sets already constructed corresponding to its
faces. The picture is exactly that from dimension 2.
Etc., inductively.
We might think of the complex as the prototypical polyhedral object. Paul
Alexandroff wrote [28],
I would formulate the basic problem of set-theoretic topology
as follows:
To determine which set-theoretic structures have a connec-
tion with the intuitively given material of elementary polyhe-
dral topology and hence deserve to be considered as geometrical
figures—even if very general ones.
4.5. Exercises
Solution. For each positive integer n, let Cn = {x ∈ Rn | |x| ∈ [n − 1, n]}.
The sets Cn are compact. Hence there is a finite subset Un of U that covers Cn .

from Un any elements that do not actually intersect Cn . Then we
We may remove
may set V = n Un .
4.5. EXERCISES 43
Figure 2. Covering a two-dimensional complex

CHAPTER 5
Fat Curves and Peano Curves
In the preceding chapter on Lebesgue covering dimension we promised to give

examples that show that dimension cannot be tied to the number of parameters
needed to define a space nor to the measure (length, area, or volume) of the space.
In this chapter we fulfill that promise.
The reader should understand topological spaces, compactness, connectedness,
and continuity of maps.
Although Euclidean m- and n-dimensional spaces Rm and Rn are most easily
described by a different number of parameters, namely m and n, Giuseppe Peano,
in 1890 ([81] and [82]), showed that, in general, the number of parameters required
in describing a space is not a suitable criterion for determining the dimension of
a space. He defined a space-filling curve, that is, a continuous map from the 1-
dimensional unit interval [0, 1] onto the 2-dimensional unit square D = [0, 1]×[0, 1].
In fact, this can be done in such a manner that the length of a subinterval [a, b]
in [0, 1] is the 2-dimensional area of the image of [a, b] in D. Consequently, the
2-dimensional space D and the areas of its subsets can be described by a single
parameter. The same is true of the n-dimensional Euclidean cube and the n-
dimensional volumes of its subsets in all dimensions n.
We shall describe three particular space-filling curves, namely, those described
by Peano [81] and [82], by David Hilbert [83], and by Georg Pólya [84]. See also
Lax [85]. In the process, we will define 1-dimensional curves in the plane that have
positive 2-dimensional area. We call such curves fat curves. So we see that areas
and volumes are also not suitable as determinants of dimension.
Those spaces that can be realized as the continuous image of the interval [0, 1]
can be characterized topologically. They are called Peano curves. We will give a
topological characterization of such spaces in the last portion of this chapter.
Though parameters and volumes are ineffective in defining dimension, we have
already discussed one of the ways in which dimension can be determined topologi-
cally, namely, by Lebesgue covering dimension.
5.1. The Constructions

The idea in constructing a space-filling curve through D = [0, 1] × [0, 1] is to
map the interval [0, 1] into D so that, by wiggling, the image comes fairly close to
every point of D. Then, one introduces more small, local wiggles, to make the image
come even closer to every point of D. This process, when iterated, creates a very
kinky sequence of maps fn : [0, 1] → D that converges uniformly to a continuous
function f : [0, 1] → D whose image is all of D.
In order to completely justify the constructions, a number of fairly intuitive
topological lemmas and standard analytical lemmas are required. Nevertheless, an
45
46 5. FAT CURVES AND PEANO CURVES
intuitive understanding of the constructions does not require those details. There-
fore, we will describe the constructions without complete justification, then include
the topological and analytical lemmas and their proofs at the end of the section.
Example 5.1 (Pólya’s example). See Figure 1. Begin with a triangle T = abc
having two distinguished vertices a and b. Choose a point c ∈ int(ab). Let C
denote a triangle in T that has c as one vertex and whose opposite side is a tiny
neighborhood N of c in int(ab). Remove int(N ) ∪ int(C) from T so as to form two
new, smaller triangles T0 and T1 , the first having a and c as distinguished vertices,
the second having c and b as distinguished vertices. (See Figure 1.) In the same
manner, cut these two triangles into four triangles T00 , T01 , T10 , and T11 . Iterate,
making sure that, in the limit, triangle size approaches 0. Stage 0 consists of the
single triangle A0 = T . Stage 1 is the union A1 of the two triangles T0 and T1 . Stage
n
2 is the union A2 of four triangles. In general,
∞ An consists of 2 triangles, strung
together linearly. The intersection A = n=0 An , as an intersection of compact,
connected subspaces of T , is also compact and connected. Since the triangles of
stage n are strung together in a linear fashion and since triangle size approaches 0,
the intersection A is obviously linearly ordered with the linear-order topology. It
follows that A is an arc.
T
T0
T1
T00
T01
T11
T10
Figure 1. Pólya’s Curve
Example 5.2 (Hilbert’s example). See Figure 2. We begin with a planar

quadrilateral Q with two adjacent vertices distinguished. We cut Q into four
quadrilaterals Q0 , Q1 , Q2 , and Q3 , each with two adjacent vertices distinguished,
as indicated in the figure. We take A0 = Q as the initial stage. We take A1 =
Q0 ∪ Q1 ∪ Q2 ∪ Q3 as stage 1. As with the previous example, we iterate and obtain
an arc as the intersection A = ∞ n=0 An .
Example 5.3 (Peano’s example). See Figure 3. As in the previous example,

we begin with a quadrilateral Q, but this time we assume that vertices in oppo-
site corners are distinguished. We divide Q into nine such quadrilaterals, strung
together linearly, as in the figure. We iterate and obtain an arc A in the limit.
5.1. THE CONSTRUCTIONS 47
Q1 Q2
Q0 Q3
Figure 2. Hilbert’s Curve
Figure 3. Peano’s Curve
The three arcs that we have described are interesting for many reasons:
First, all three arcs are incredibly twisted and kinked. In particular, they
are fractal arcs in the intuitive sense of Benoit Mandelbrot. See Mandelbrot [86],
Devaney [88], and Falconer [87].
Second, we see even more precisely that these arcs can be constructed to satisfy
the technical definition of fractal as described by Mandelbrot, in that they have 2-
dimensional area > 0 while their topological dimension is 1. We see this by noting
that the amount of area removed at each stage can be made as small as desired.
Hence, we may assume that the total area removed in the process can be made as
small as we like: < for any prescribed > 0. Thus the 1-dimensional arc can be
chosen to have positive 2-dimensional area greater than (original area - ). That
is, 1-dimensional sets can have positive 2-dimensional area.
Third, we can string such arcs of positive area together to form a simple closed
curve S. The curve S will separate the plane into two pieces, an interior U and an
exterior V . The closure of U will be U ∪ S. The area of U will be strictly less than
the area of its closure. That is, the boundary of a set can be fat.
Fourth, the edges of the gaps that we cut in creating the arcs can be pinched
back together to recover the original triangle and quadrilaterals (See Figure 4). The
result maps our kinky arcs continuously onto the original triangle and quadrilater-
als. That is, we obtain space-filling curves — first a triangle-filling curve, then two
quadrilateral-filling curves. Figure 5 shows approximations to the quadrilateral-
filling curves.
Figure 4. Peano’s second subdivision
Figure 5. Hilbert and Peano approximations
Fifth, each of the three space-filling curves we have just described is famous
in its own right. The third curve is the original curve described by Peano, for
which Peano curves were named; Peano described the curve by formula depending
essentially on a ternary representation of the real numbers. The second is Hilbert’s
space-filling curve for which Hilbert drew several approximations and gave the first
visual description of a space-filling curve; Hilbert’s curve is essentially based on
expressing numbers in base 4. The first of the three examples is the most recent. It
is Pólya’s space-filling curve which depends rather obviously on binary expansions
of the real numbers as indicated by the subscripts on the labels of the triangles Tij .
Pólya described the image of each point x ∈ [0, 1] under his space-filling curve
algorithmically. Lax [85] made that algorithm particularly visual as follows: We
assume that x is expressed in binary notation as an unending string of 0’s and 1’s.
x = .10 · · ·
We take as the triangle a right triangle whose legs are not equal in length.
We draw the altitude P0 to the hypotenuse. This path divides the triangle into
two similar right triangles, one of which is smaller and one of which is larger. If the
first entry in the binary expansion of x is 1 (as it is in the indicated example), we
draw the altitude P1 in the larger triangle. If the first entry in the binary expansion
of x is 0, we draw the altitude P1 in the smaller triangle. In either case, the path
5.2. THE TOPOLOGICAL LEMMAS 49
P0
P1
P2
x = .10 · · ·
Figure 6. Lax’s Algorithm
P1 divides a triangle into two similar triangles, one of which is smaller and one of
which is larger. If the second entry in x is 0 (as it is in the indicated example of
Figure 6), we draw the altitude P2 in the smaller triangle. If the second entry in
x is 1, we draw P2 in the larger triangle. We iterate, considering successive entries
in the binary expansion of x. The sequence P0 , P1 , P2 , . . . converges to a unique
point of the triangle. This convergence point is the image of x under Pólya’s map,
as described by Lax. Note that the paths described by .0111111 · · · and .10000 · · · ,
both of which describe the real number 1/2, converge to the same point of the
triangle, namely the vertex between the two original legs of the triangle. Although
some numbers have two binary expansions, the algorithm gives the same image
point for each.
The examples show that there are unexpected difficulties in defining dimen-
sion. One cannot say that a set is 1-dimensional if it can be parametrized by a
single variable. In fact, the unit cube of dimension n can be parametrized by a
single variable in every dimension n. One cannot say that a set is n-dimensional
topologically just because it has positive n-dimensional volume.
The examples show that the size of a set and the size of its closure may be
different. This causes technical difficulties in assigning size (volume) to a set.
All of these difficulties caused headaches for an entire generation of mathemati-
cians.
5.2. The Topological Lemmas

We begin with a characterization of the arc. We learned this characterization
from C. E. Burgess in his very effective class introducing problem solving, proofs,
and topology. Note that the real line R and the interval [0, 1] are linearly-ordered
sets whose topologies can be described as follows:
Definition 5.4. Suppose (A, <) is a linearly ordered set. Add to A a new first
point −∞ and a new last point +∞. Define an open interval in A to be a subset
of A of the form (x, y) = {z ∈ A | x < z < y}, where x and y are any two elements
of A ∪ {±∞}. Then the open intervals in A form a basis for a topology on A that
is called the linear-order topology.
Theorem 5.5 (Characterization of the arc). A space A is an arc if it is linearly
ordered with the linear-order topology, has distinct first and last points, is connected,
and has a countable dense subset.
We give only an outline of the proof. The reader is invited to fill in the details:
Proof outline for the characterization of the arc. Let

S = {x0 , x1 , a1 , a2 , . . .} denote a countable dense subset of A, where x0 is the first
point of A and x1 is the last.
Construct a countable dense subset of [0, 1] of the form T = {0, 1, b1 , b2 , . . .} so
that S and T are order isomorphic.
Then the order-preserving bijection x0
→ 0, x1
→ 1, ai
→ bi , extends uniquely
to a homeomorphism from A to [0, 1]. (End of proof outline.)
Corollary 5.6 (Characterization of the simple closed curve). A space S is a

simple closed curve if it is the union of two arcs A1 and A2 , closed subsets of S,
whose intersection is the set of endpoints of each.
Proof. Map the upper semicircle C1 of the circle C homeomorphically onto
A1 and the lower semicircle C2 onto A2 .
The characterization of the arc has a very interesting consequence.

Corollary 5.7. Suppose f : [0, 1] → X is a continuous function from the unit
interval [0, 1] into a metric space X such that f (0) = f (1) and such that, for each
x ∈ X, f −1 (x) is connected. Then f ([0, 1]) is an arc joining f (0) and f (1).
Here is an alternative way of stating the corollary.
Definition 5.8. Let G denote a family of disjoint points and closed intervals
filling [0, 1], and let p : [0, 1] → G denote the projection map that assigns to each
x ∈ [0, 1] the element p(x) ∈ G that contains x.
Topologize G as follows. A set U ⊂ G is said to be open if and only if p−1 (U )
is open in [0, 1].
The topological space G is usually denoted by the symbol [0, 1]/G and is called
a monotone decomposition space of [0, 1].
Theorem 5.9. A monotone decomposition space [0, 1]/G of [0, 1] is either a
single point (when [0, 1] is the only element of G) or an arc.
Proof. Assume [0, 1]/G has at least two elements. Since p : [0, 1] → [0, 1]/G
is continuous (why?), [0, 1]/G is connected and has a countable dense subset. It
remains only to prove that [0, 1]/G is linearly ordered with the linear-order topology
with distinct first and last points.
If g1 and g2 are distinct elements of G, define g1 < g2 if and only if there
exist x1 ∈ g1 and x2 ∈ g2 with x1 < x2 . Since each g ∈ G is connected, this
order is clearly a linear order on G with first element containing 0 and last element
containing 1.
Let g ∈ U ⊂ [0, 1]/G, with U open in the decomposition space topology. Then
p−1 (U ) is, by definition, open in the [0, 1] topology, hence a disjoint union of open
intervals. Let (a, b) be the one containing the connected set g. Then (p(a), p(b)) is
an open interval in the linear-order topology in U containing g.
Let g ∈ (g1 , g2 ), with (g1 , g2 ) an open interval in the linear-order topology. Let
x be the last point of g1 and y the first point of g2 . Then p−1 (g1 , g2 ) = (x, y), which
is open in the topology of [0, 1].
We conclude that the decomposition space topology and the linear-order topol-
ogy coincide. Hence [0, 1]/G is an arc.
5.3. THE ANALYTICAL LEMMAS 51
5.3. The Analytical Lemmas

All of the following lemmas are probably familiar to someone who has taken
advanced calculus.
Definition 5.10. Metric spaces X, Y , etc., in which each sequence x1 , x2 , . . .
has a convergent subsequence xn1 , xn2 , → x are called compact. (This definition
of compactness is not the most useful definition, but it requires the fewest new
concepts.)
Definition 5.11. We denote the distance between two points x and y by the
symbol d(x, y). A sequence xn is Cauchy if, for each > 0, there is an integer N
such that n, m ≥ N implies d(xn , xm ) ≤ .
Theorem 5.12. In a compact metric space X, every Cauchy sequence xn of
points converges.
Proof. Let xni denote a subsequence that converges to a point x ∈ X. Sup-
pose that > 0 is given. Since the sequence xn is Cauchy, there is an N such that
n, m ≥ N implies d(xn , xm ) < /2. Likewise, we may choose a point xni of the
convergent subsequence with ni > N such that d(xni , x) < /2. Then for n ≥ N ,
d(xn , x) ≤ d(xn , xni ) + d(xni , x) < /2 + /2 = .
Thus xn → x.
Definition 5.13. A sequence fn : X → Y of functions is Cauchy if, for each

> 0, there is an integer N such that, for each x ∈ X and for each n, m ≥ N ,
d(fn (x), fm (x)) ≤ . We use the symbol d(fn , fm ) to denote the distance between
the functions fn and fm , which is the supremum of the distances d(fn (x), fm (x)).
Theorem 5.14. For compact metric spaces X and Y , every Cauchy sequence
fn : X → Y of continuous functions converges to a continuous function f : X → Y .
Proof. Since the sequence fn is Cauchy, for each x ∈ X the sequence fn (x)
of points is also Cauchy. Hence fn (x) converges to a point of Y which we denote
by f (x). Note that if d(fn , fm ) < for n, m ≥ N , then d(f (x), fN (x)) ≤ .
We need to show that f is continuous. To that end, suppose that x ∈ X and
> 0 are given. Since the sequence fn is Cauchy, there is an integer N such that
n, m ≥ N implies d(fn , fm ) < /3. Since fN is continuous at x, there is a positive
number δ > 0 such that, if d(x, y) < δ, then d(fN (x), fN (y)) < /3. Then, if
d(x, y) < δ,
d(f (x), f (y)) ≤ d(f (x), fN (x)) + d(fN (x), fN (y)) + d(fN (y), f (y)) ≤ 3 · /3.
Thus f is continuous at x.
Definition 5.15. We say that a function f : X → Y is -dense in Y , if, for

each point y ∈ Y , there is a point x ∈ X such that d(f (x), y) ≤ .
Theorem 5.16. For compact metric spaces X and Y , suppose that fn : X → Y
is a Cauchy sequence of continuous functions such that each of the functions fn is
1/n-dense in Y . Then the limit function f : X → Y is surjective (= onto). That
is, for each y ∈ Y , there is an x ∈ X such that f (x) = y.
Proof. Suppose that a point y ∈ Y is given. We must find a point x0 ∈ X such

that f (x0 ) = y. We pick points x1 , x2 , . . . such that d(fn (xn ), y) ≤ 1/n. Passing to
a subsequence if necessary, we lose no generality in assuming that the sequence xn
converges to a point x0 ∈ X. We must show that f (x0 ) = y. Passing to a further
subsequence if necessary, we lose no generality in assuming that d(fn , fn+k ) < 1/n
for all positive integers n and k.
We must show that f (x0 ) = y. To that end, we bound successive distances
between points y, fm (xm ), fN (xm ),fN (x0 ) and f (x0 ), so that y and f (x0 ) have to
be close to one another.
We pick N large so that 1/N is small. Using the continuity of fN , we may choose
m > N so that the distance d(xm , x0 ) is small and so that d(fN (xm ), fN (x0 )) is
consequently also small.
The distance between y and fm (xm ) is less than 1/m by choice.
The distance between fm (xm ) and fN (xm ) is less than 1/N provided that
m > N.
The distance between fN (xm ) and fN (x0 ) is small, as indicated above.
The distance between fN (x0 ) and f (x0 ) is less than 1/N since d(fN , fN +k ) <
1/N for each positive integer k.
We conclude that the distance from f (x0 ) to y is as close as we like to 1/m +
1/N + + 1/N , with m and N as large as we like and as small as we like. Hence
f (x0 ) = y.
5.4. Characterization of Peano Curves

Definition 5.17. Properties are said to characterize a space topologically if
those properties are preserved under homeomorphism, are satisfied by, and only by,
spaces homeomorphic to the given space.
Our goal in this section is to give a very nice characterization of those metric
spaces that can be realized as the continuous image of the unit interval.
Definition 5.18. A continuum is a compact, connected metric space.
Definition 5.19. A space X is locally connected iff each component of each
open subset of X is also open.
Definition 5.20. A Peano continuum is a locally connected continuum.
Example 5.21. The topologist’s sine curve (the closure of the curve sin(1/x)
for x ∈ (0, 1]) is an example of a continuum that is not locally connected. See
Figure 7.
Theorem 5.22. A continuum X is locally connected (that is, X is a Peano

continuum) if and only if there is a continuous surjection f : [0, 1] → X.
Remark. Note that every sphere and ball of dimension ≥ 1 is a Peano con-
tinuum, hence, by the theorem, is the continuous image of the interval [0, 1]. End
remark.
Proof. Suppose that there is a continuous surjection f : [0, 1] → X. Suppose
U ⊂ X is open, and let V denote a component of U . We must show that V is
open. Suppose to the contrary that V is not open so that there is a sequence x1 ,
x2 , . . . from U \ V that converges to x ∈ V . Let y1 , y2 , . . . ∈ f −1 (U ) be such
5.4. CHARACTERIZATION OF PEANO CURVES 53
Figure 7. The topologist’s sine curve
that f (yn ) = xn . Since [0, 1] is compact, we may assume that yn → y ∈ [0, 1]. By
continuity, f (y) = x. Therefore, y lies in the open set f −1 (U ). Since [0, 1] is locally
connected, the component V of f −1 (U ) that contains y is open in [0, 1], hence
contains all but finitely many of the points yn . The image of V is a connected
subset of U and contains x. Hence f (V ) ⊂ V , and all but finitely many of the
points xn = f (yn ) lie in V , a contradiction. We conclude that V is open and that
X is locally connected.
Remark. Our proof that X is locally connected used only the facts that [0, 1]
is compact and locally connected. End remark.
Conversely, assume that X is locally connected. We must construct a continu-
ous surjection f : [0, 1] → X. For each integer n > 0, let U(n) be a finite open cover
of X by connected open subsets of diameter < 1/2n . Let V(n) be the collection of
closures of the elements of U(n). Note that the elements of V(n) are also connected
of diameter < 1/2n .
There is a sequence V(1, 1), V(1, 2), V(1, 3), . . . , V(1, n(1)) of elements of V(1)
such that:
(1) Every element of V(1) appears at least once in the sequence, and
(2) For each i > 1, V(1, i) ∩ V(1, i − 1) is nonempty.
Pick points x(1, 1) ∈ V(1, 1), x(1, 2) ∈ V(1, 2) ∩ V(1, 1), . . . , x(1, n(1)) ∈
V(1, n(1)) ∩ V(1, n(1) − 1), and x(1, n(1) + 1) ∈ V(1, n(1)). Let
0 = y(1, 1) < y(1, 2) < · · · < y(1, n(1)) < y(1, n(1) + 1) = 1
denote equally spaced points in the interval [0, 1]. We will eventually map y(1, i)
to x(1, i) for each i.
We proceed by induction. We assume that a sequence V(k, 1), V(k, 2), . . . ,
V(k, n(k)) of elements of V(k) have been defined such that:
(1) Every element of V(k) appears at least once in the sequence.
(2) For each i > 1, V(k, i) ∩ V(k, i − 1) is nonempty.
We assume further that points x(k, 1) ∈ V(k, 1), x(k, 2) ∈ V(k, 2) ∩ V(k, 1),
V(k, 3) ∈ V(k, 3)∩V(k, 2), . . ., x(k, n(k)) ∈ V(k, n(k))∩V(k, n(k)−1)), and x(k, n(k)+
1) have been chosen, along with corresponding points y(k, i) ∈ [0, 1] which are to
be mapped to the points x(k, i).
We concentrate on a single one of the sets V(k, i) with its special points x(k, i)
and x(k, i + 1) and corresponding points y(k, i) and y(k, i + 1) of [0, 1]. There is
a finite subcollection of the cover U(k) that irreducibly covers V(k, i). (That is,
no smaller subcollection covers V(k, i).) The elements of this subcollection can be
placed in a sequence such that:
(1) Each element appears at least once in the sequence.
(2) Each element intersect the next element in the sequence.
(3) x(k, i) lies in the first element in the sequence.
(4) x(k, i + 1) lies in the last element in the sequence.
Choose corresponding x’s in successive intersections and equally spaced y’s in
the interval [y(k, i), y(k, i + 1)] ⊂ [0, 1]. Use the element x(k, i) as the first of the
x’s and x(k, i + 1) as the last of the x’s. Arrange these covers of the V(k, i)’s into
a single sequence with the corresponding sequence of x’s and y’s.
Iterate. Then the map sending y(k, i)’s to the x(k, i)’s is defined on a dense sub-
set of [0, 1] and is uniformly continuous. It follows that there is a unique continuous
extension to all of [0, 1], and this extension is surjective.
Corollary 5.23. If X is a Peano continuum, U is a connected open subset
of X, and x, y ∈ U , then there is a continuous function f : [0, 1] → U such that
f (0) = x and f (1) = y.
Proof. Cover U with connected open sets whose closures lie in U . There is
a sequence of these open sets such that x lies in the first and y in the last, while
each open set intersects the next one in the sequence. Let X1 , X2 , . . ., Xn denote
the corresponding sequence of closures. Let this chain replace the first chain in the
proof of the previous theorem. Proceed thereafter as in the proof of the previous
theorem. The result is a continuous function f : [0, 1] → U with f (0) = x and
f (1) = y.
Theorem 5.24. Suppose that f : [0, 1] → X is a continuous function into a
metric space X, with f (0) = f (1). Then there is an arc g : [0, 1] → f ([0, 1]) that
joins f (0) with f (1).
Proof. Let I(1) = [a(1), b(1)] denote a largest metric interval in [0, 1] such
that f (a(1)) = f (b(1)). Let f1 (x) = f (x) for each x ∈ / I(1), and let f1 (I(1)) =
f (a(1)). Let I(2) = [a(2), b(2)] denote a largest metric interval in [0, 1] \ I(1) such
that f (a(2)) = f (b(2)). Change f1 only on I2 so as to define f2 with f2 (I2 ) =
f1 (a(2)). Iterate. Since there can only be finitely many subintervals of any positive
size, this process creates a limit function g such that fn → g, g : [0, 1] → f ([0, 1]),
and every point preimage is closed and connected. Thus g([0, 1]) is an arc by
Corollary 5.7.
5.5. Exercises
5.1. Construct a physical model of Pólya’s triangle-filling curve out of two
colors of poster board, each cut into triangles, attached together in a line. Show
how that line can be folded into one large triangle, so that the folding approximates
Pólya’s map. See Figures 8 and 9.
5.2. Construct physical models for the Peano and Hilbert space-filling curves.
5.3. Pólya’s curve can be modified by basing it on right triangles of various
shapes. What shape or shapes minimize the number of points sent to the points of
the large triangle? (See Lax [85].
5.5. EXERCISES 55
Figure 8. Polya’s curve unfolded
Figure 9. Polya’s curve folded

CHAPTER 6
The Arc, the Simple Closed Curve, and the

Cantor Set
Prerequisites: We have previously assumed the reader understands at least

intuitively the notions of compactness, connectedness, and continuity. We are going
to use the precise definition of connectedness again and again. Hence we review:
The reader should understand the definition of a connected space and of a separation
of a space. We say that a space X is not connected if it can be written as a union
X = A∪B of nonempty sets A and B such that neither A nor B contains a point or
a limit point of the other. We write “X = A ∪ B, separated.” We say that a point
x ∈ X separates the space X if we can write X \ {x} as a union X \ {x} = A ∪ B,
separated. A pair of points x and y separates X if X \ {x, y} = A ∪ B, separated.
We devote this chapter to the topological characterization of three important
spaces, namely, the arc [0, 1], the simple closed curve S1 , and the Cantor set C. The
first two are undoubtedly the two most important spaces used in understanding the
topological structure of the plane, and their wonderful characterizations supply the
model for our later topological characterization of the plane itself. The Cantor set
has played a role in some of the most important historical examples in calculus and
its generalizations.
6.1. Characterizing the Arc and Simple Closed Curve

An arc is any space homeomorphic with the interval [0, 1]. A simple closed
curve is any space homeomorphic with the circle S1 . Arcs and simple closed curves
can be recognized by the way they are divided into pieces when points are removed.
Interior points of an arc cut the arc into two pieces. Pairs of points in a simple
closed curve cut the simple closed curve into two pieces.
We have already given one characterization of the arc in terms of its linear
order topology. In this chapter we will give a second characterization and this
second characterization will depend on the first, which we recall here.
Theorem 6.1 (First characterization of the arc). A space X is an arc if it is
linearly ordered with the linear-order topology, has distinct first and last points, is
connected, and has a countable dense subset.
Theorem 6.2 (Second characterization of the arc). A compact, connected met-
ric space X is an arc iff there are exactly two points a and b of X that do not separate
X. (The points a and b are called the endpoints of X.)
Theorem 6.3 (Characterization of the simple closed curve). A compact, con-
nected metric space S is a simple closed curve iff S has more than one point and
is separated by each pair of its points.
57
58 6. THE ARC, THE SIMPLE CLOSED CURVE, AND THE CANTOR SET
The characterization of the simple closed curve S will be reduced to the second
characterization of the arc by showing that any two points a and b of S divide S
into two subspaces, each being an arc with a and b as endpoints.
Before we begin the proofs of the characterization theorems, we need two facts.
The first is very easy. The second is the most complicated ingredient in the char-
acterization theorems.
Lemma 6.4. Suppose that X is a compact, connected metric space, that x ∈ X,
and that X \ {x} = A ∪ B, separated. Then each of the subspaces A ∪ {x} and
B ∪ {x} is connected.
Proof. Suppose to the contrary that A ∪ {x} = C ∪ D, separated, with x ∈ C.
Then X = (B ∪ C) ∪ D, separated, so that X is not connected, a contradiction.
Lemma 6.5. Suppose that X is a compact, connected metric space having at

least two points. Then there are at least two points a, b ∈ X neither of which
separates X.
The intuitive idea behind the lemma is that there are at least two (usually
infinitely many) extreme edges in any such space X and that the points on the
edges do not separate X. These edges are found by cutting away interior portions of
X iteratively. The idea of the proof is simple and is illustrated by the simplest case
X = [0, 1]: Cut off a subarc of [0, 1], cut the subarc to get an even smaller subarc
nearer the end, and iterate, getting smaller and smaller until, in the limit, only
the endpoint remains. This endpoint does not separate the space. The technical
difficulty lies in making the successive remnants sufficiently small. Since X has
infinitely many points, we need only consider the case where most points of X
separate X. Thus it suffices to prove the following lemma.
Lemma 6.6. If X is a compact, connected metric space that is separated by a
point x into nonempty, disjoint open sets A and B, then each of A and B contains
a point that does not separate X.
Proof. Our proof will be indirect. We assume the existence of a separation
X \ {x} = A ∪ B, where, contrary to the theorem, every point of B separates
X. Setting x0 = x, A0 = A, and B0 = B, we will use this initial separation
X \{x0 } = A0 ∪B0 to construct an entire sequence of separations X \{xn } = An ∪Bn
with the sizes of the sets Bn getting smaller in an efficient manner. In the limit we
will find a separation X \ {x∞ } = A∞ ∪ B∞ with B∞ ⊂ B, but with no point y of
B∞ separating X, a contradiction.
Here is the inductive construction. See Figure 1:
Let U1 , U2 , . . . denote a countable basis of open sets for X. Suppose that xn−1
and Bn−1 have been defined. Choose xn ∈ Bn−1 , with X \ {xn } = An ∪ Bn ,
separated. Since An−1 ∪ {xn−1 } is a connected subset of X \ {xn }, we may choose
the notation so that An−1 ∪ {xn−1 } lies in An . Note that Bn ∪ {xn } ⊂ Bn−1 . We
have many choices for xn and for the associated Bn . We choose xn and Bn so that
Bn is small in the following sense: if possible, choose xn and Bn so that Bn misses
Un .
Let x∞ be a point of the nested intersection of the nonempty compact sets
B0 ∪ {x0 } ⊃ B1 ∪ {x1 } ⊃ · · · .
6.1. CHARACTERIZING THE ARC AND SIMPLE CLOSED CURVE 59
B0
Bn
B∞
B
X x0 xn x∞ y
A0
An
A∞
A
Figure 1. Finding the points that do not separate
Since every point of B0 separates X by hypothesis, X \{x∞ } = A∞ ∪B∞ , separated,

with the ascending union of connected sets
A0 ∪ {x0 } ⊂ A1 ∪ {x1 } ⊂ · · ·
lying in A∞ .
We obtain a contradiction as follows: If y ∈ B∞ , then X \ {y} = A ∪ B ,
separated, with A∞ ∪ {x∞ } ⊂ A . Let Un be a basic open set that contains x∞ but
misses the compact set {y} ∪ B . Thus at stage n, it was possible to choose xn so
that Bn misses Un . But Bn contains x∞ ∈ Un , a contradiction.
We conclude that some point of B fails to separate X.
We now prove a series of lemmas designed to prove the character-

ization theorem for the arc. Since every compact metric space has a
countable dense subset and since our set X is connected by hypothesis,
it follows from the first characterization that it suffices to show that
there is a linear order on X with distinct first and last points with re-
spect to which X has the linear-order topology. From this point on, we
assume the hypotheses of the second characterization of the arc: X is a
compact, connected metric space with exactly two points a and b that
do not separate X.
Lemma 6.7. If X \ {z0 } = A ∪ B, separated, with a ∈ A, then b ∈ B.
Proof. By Lemma 1.6, each of A and B must contain a point that does not
separate X. Since A contains a, B must contain the other nonseparating point
b.
Definition 6.8. For each point z ∈ X, we define sets A(z) and B(z) as follows.
If z ∈ A \ {a, b}, then, by hypothesis, A \ {z} = A(z) ∪ B(z), separated, with
a ∈ A(z). By the previous lemma, b ∈ B(z). For the nonseparating point a, we
define A(a) = ∅ and B(a) = A \ {a}. Similarly, A(b) = A \ {b} and B(b) = ∅. We
see that these sets are uniquely defined as follows.
Lemma 6.9. The sets A(z) and B(z) are connected, hence uniquely defined.
Proof. If B(z) = C ∪D, separated, with b ∈ C, then X \{z} = (A(z)∪C)∪D,
separated, so that D contains a third nonseparating point of X, a contradiction.
Definition 6.10. We define x < y in X if x ∈ A(y).
Lemma 6.11. The relation x < y is satisfied iff y ∈ B(x). It is impossible to
have both x < y and y < x. The relation < is also transitive. Therefore < is a
linear order on X.
Proof. We have x < y iff x ∈ A(y). In that case, the connected set {y} ∪ B(y)
misses x and contains both y and b. Hence {y} ∪ B(y) ⊂ B(x) and, in particular,
y ∈ B(x). Conversely, if y ∈ B(x), then the connected set {x} ∪ A(x) misses y and
contains both x and a. Hence{x} ∪ A(x) ⊂ A(y) and, in particular, x ∈ A(y) and
x < y.
To have x < y and y < x we would have to have y ∈ B(x) by the previous
paragraph, and y ∈ A(x) by definition. But B(x) and A(x) are disjoint.
The order is transitive because x < y < z implies that {x}∪A(x) ⊂ {y}∪A(y) ⊂
A(z), so that the order is transitive, hence a linear order.
Lemma 6.12. If X is a compact, connected metric space with exactly two non-
separating points a and b, then the linear-order topology on X defined by the linear
order above agrees with the given metric topology on X.
Proof. Intervals of the form [a, y) = A(y) and (x, b] = B(x) are open and
form a subbasis for the linear-order topology on X. Hence the map from the metric
topology to the linear-order topology on X is continuous.
We need to show that, conversely, if z is a point of the open set U in the metric
topology, then z is contained in an open interval in U in the linear-order topology.
Suppose z ∈ U is distinct from the initial endpoint a. Since the set {z} ∪ A(z)
is connected, there is a sequence z1 < z2 < · · · from A(z) converging to z in the
metric topology. We claim that, for some i, the order interval (zi , z) lies in U .
If not, then we may assume the existence of points w1 , w2 , . . . ∈ A(z) \ U with
z1 < w1 < z2 < w2 < · · · such that w1 , w2 , · · · → w ∈ A(z) \ U . Choose w with
w < w < z. Since zi → z, we may assume that all of the zi ’s lie in the open
6.2. THE CANTOR SET AND ITS CHARACTERIZATION 61
set B(w ). It follows that the wi ’s lie in B(w ) as well. But the wi ’s converge to
w ∈ A(w ), a contradiction.
By a similar argument, there is an interval (z, zi ) ⊂ U .
This last lemma completed the final requirement for the application
of first characterization. Hence the proof of second characterization of
the arc is complete.
We now prove three lemmas that, together, establish the character-
ization of the simple closed curve. From here on we assume that S is
a compact, connected metric space having at least two points that is
separated by each pair of its points.
Lemma 6.13. No point of S separates S.
Proof. If S \ {x} = A ∪ B, separated, then A ∪ {x} and B ∪ {x} are connected,
and each has a nonseparating point distinct fom x, say a ∈ A and b ∈ B. Then
S \ {a, b} = (A ∪ {x} \ {a}) ∪ (B ∪ {x} \ {b}) is the union of two connected sets with
a common point, hence is connected, a contradiction.
Lemma 6.14. If S \ {x, y} = A ∪ B, separated, then each of the sets A ∪ {x, y}
and B ∪ {x, y} is connected.
Proof. If A ∪ {x, y} = C ∪ D, separated, we consider two cases.
Case 1. Suppose x and y are in the same one of the sets, say x, y ∈ C. Then
S = (B ∪ C) ∪ D, separated, a contradiction.
Case 2. Suppose x ∈ C and y ∈ D. Then S \ {x} = (C \ {x}) ∪ (B ∪ D),
separated, a contradiction.
Lemma 6.15. If S \ {x, y} = A ∪ B, separated, then each of the sets A ∪ {x, y}
and B ∪ {x, y} is an arc with x and y as endpoints.
Proof. Suppose to the contrary that b ∈ B does not separate B ∪ {x, y}. If
a ∈ A does not separate A ∪ {x, y}, then S \ {a, b} is the union of two connected
sets with x, y as common points, a contradiction. Hence every point of A separates
A∪{x, y} so that A∪{x, y} is an arc with endpoints x and y by the characterization
of the arc. But thus S \{a, b} is the union of the connected set C = (B ∪{x, y}\{b})
and the two components of A ∪ {x, y} \ {a}, each of which shares a point with C,
a contradiction. Hence every point of B separates B ∪ {x, y}, so that B ∪ {x, y} is
also an arc with endpoints x and y.
The characterization of the simple closed curve is an immediate con-
sequence of this last lemma. Hence the characterization is complete.
6.2. The Cantor Set and Its Characterization

Definition 6.16. See Figure 2. Let C0 denote the unit interval [0, 1]. Assume
inductively that Cn has been defined and that it is the disjoint union of finitely
many closed intervals in the real line R. From each of those closed intervals, delete
the open middle third. This process creates two new closed intervals, each of which
Cn+1 denote the union of these smaller
is one third of the original interval. Let
closed intervals. Finally, define C = ∞ n=1 Cn . Then C is called the standard
middle-third Cantor set. Any set homeomorphic with C is also called a Cantor set.
C0
C1
C2
C3
C4
Figure 2. The Cantor set
Theorem 6.17 (Characterization of the Cantor set.). A compact metric space

X is a Cantor set if and only if each component of X is a point and each point
x ∈ X is a limit point of its complement X \ {x}.
The proof relies on two lemmas. The second is a fairly easy consequence of
the first. Both lemmas are obviously true of the middle-thirds Cantor set. We first
prove the characterization theorem on the basis of the two lemmas, then give the
proofs of the lemmas.
Lemma 6.18. Every point x ∈ X has arbitrarily small neighborhoods that are
both open and closed in X.
Lemma 6.19. Given > 0, there is a positive integer K such that, for each
k ≥ K, there is a cover of X by k nonempty disjoint sets that are both open and
closed in X and have diameter < .
Proof of the characterization theorem. By the lemma, there exist
covers X(1), . . ., X(k) of X and C(1), . . ., C(k) of the middle-thirds Cantor set
C such that the elements of the two covers are disjoint, nonempty, both open and
closed, and have diameter < 1. We match each X(i) with C(i) to form a pair
(X(i), C(i)), and treat each of these pairs separately.
We assume inductively that we have pairs (X , C ) of nonempty, open and
closed subsets of X and C, respectively, each of diameter < 1/n. The sets X and
C satisfy the conditions of the characterization theorem. Hence, we may apply the
lemma to find covers X (1), . . ., X (k ) of X and C (1), . . ., C (k ) of C such that
the elements of the two covers are disjoint, nonempty, both open and closed, and
have diameter < 1/(n + 1). We pair X (i) with C (i).
Each point of x ∈ X is then defined by a sequence of sets X (0) (i0 ) ⊃ X (1) (i1 ) ⊃
X (i2 ) ⊃ · · · , where X (n) (in ) has diameter < 1/n. These sets are paired with sets
(2)
C (n) (in ) from C whose intersection is a single point c(x) ∈ C. The map x
→ c(x)
is the desired homeomorphism from X to C.
Proof of the first lemma. Consider an arbitrary point x ∈ X. We claim
that x has arbitrarily small neighborhoods in X that are both open and closed in
X. If this is not the case we shall show the component containing x has more than
one point.
In order to see this, we consider two positive numbers and δ. Let N (δ) denote
the set of points y that can be joined to x by a chain y = y0 , . . . , yn = x of points
such that for i = 1, . . . , n, the distance from yi to yi−1 is less than δ. We call such
a chain a δ chain from y to x. Note that the set N (δ) is both open and closed.
We claim that, for δ sufficiently small, the set N (δ) lies in the neighborhdood
6.3. INTERESTING CANTOR SETS 63
of x in X. Otherwise, let C(1), C(2), . . . denote a sequence of finite chains, each

ending at x, with C(n) being a 1/n chain that begins outside the neighborhood of
x. A subsequence converges (Theorem 3.16) to a compact, connected set C which
contains both x and some point outside the neighborhood of x, a contradiction.
Proof of the second lemma. We conclude from the first lemma that, for
each x ∈ X, x has arbitrarily small neighborhoods in X that are both open and
closed. For fixed > 0, we cover X by finitely many sets X1 , X2 , . . ., XK of
diameter < that are both open and closed. Since intersections and differences
of such sets are also both open and closed, taking intersections and differences as
needed, we find that we may assume these sets are disjoint.
By splitting these K sets again by the same argument, we may express each
Xi as a union of as many such sets as we wish, and, then joining them together in
finite unions, we may assume each is the union of precisely as many such sets as
we wish. That is, we may divide X into any number k of disjoint open and closed
sets X1 , X2 , . . ., Xk provided only that k ≥ K.
6.3. Interesting Cantor Sets

General construction. Given the characterization of the Cantor set, it is
an easy matter to construct Cantor sets having interesting properties. One starts
with a set X0 that is the union of a finite collection of nonempty disjoint compact
sets X0 (1), . . ., X0 (k0 ) in some space such as Euclidean n-dimensional space Rn .
Assuming inductively that Xi has been chosen, one replaces Xi with a subspace
Xi+1 ⊂ X( i) that is also a union of a finite collection of nonempty disjoint compact
sets Xi+1 (1), . . ., Xi+1 (ki+1 ), with at least two of the sets Xi+1 (j) lying in each
of the sets Xi (). Finally, we require that the sets Xi+1 (j) each have diameter
< 1/(i + 1). ∞
It follows immediately that the intersection X = i=0 Xi is a set satisfying the
requirements of the characterization of the Cantor set, hence is a Cantor set.
Five examples of the general construction.
• An opaque Cantor set.
• A fat Cantor set.
• A wild Cantor set.
• A fractal Cantor set.
• A nonconstant continuous function that has slope 0 almost everywhere.
6.3.1. An Opaque Cantor Set. Given a rectangle X0 in the plane, it is an
easy matter to construct a finite collection X1 (1), . . ., X1 (k1 ) of disjoint rectangles
in X0 , each of diameter < 1, and each having sides parallel to those of R, such that
the following property is satisfied. If S is a straight line segment joining the top
and bottom of X0 , then there is at least one of the rectangles X1 (j) such that some
subsegment of S joins the top and bottom of X1 (j). Here is one way to construct
such a collection of rectangles:
Partition the top and bottom of X0 into n segments of equal length < 1.
Consider all of the n2 parallelograms that have one of the top segments as one end
and one of the bottom segments as the opposite end. Then there is a rectangle
in X0 having diameter < 1 that cuts the top end of the parallelogram from the
bottom end. See Figure 3. These rectangles may be chosen to be disjoint.
Figure 3. Blocking rays through a rectangle

k1
We set X1 = j=1 X1 (j) and proceed by induction. Assuming that Xi has
been defined, find a collection of rectangles in each rectangle of Xi that block the
bottom from the top, each of the new rectangles having diameter less than 1/(i+1).
∞ rectangles is Xi+1 .
The union of these smaller
The intersection i=0 Xi is an opaque Cantor set in the sense that every seg-
ment from the bottom of X0 to the top must intersect the Cantor set.
6.3.2. A Fat Cantor Set. The total length of the segments removed from
[0, 1] in forming the middle-thirds Cantor set is

1 1 2 1 2 2
+ · + · · + · · · = 1.
3 3 3 3 3 3
We conclude that the middle-thirds Cantor set has length 1 − 1 = 0.
But the only critical condition required in the construction of the general Cantor
set is that the pieces remaining at each stage have diameters approaching 0 in the
limit. In order to satisfy that requirement, it suffices, for example, to make sure
that each remaining piece have diameter less than one-half the diameter of the
preceding pieces. Thus we may remove a total length as little as we please. As a
result, there are Cantor sets in [0, 1] with length as close to 1 as we choose.
6.3.3. A Wild Cantor Set. A Cantor set X in 3-space R3 is tame if there

is a homeomorphism of R3 to itself that takes X to the middle-thirds Cantor set Y
viewed as a subset of the x-axis. Otherwise X is called wild.
Although each Cantor set is 0-dimensional, there are Cantor sets in 3-space
that are so complicated and tangled that one could form a necklace out of them
that could be placed around the neck and would not fall off. The standard example
X is, in fact, called Antoine’s Necklace. The Cantor set X must be wild because it
does not share the following property with the middle-thirds Cantor set Y .
Exercise 6.20. View the standard middle-thirds Cantor set Y as a subset of

the x-axis in 3-space R3 . Let f : S1 → R3 \ Y . Then the map f can be extended
to a map F : B2 → R3 \ Y .
Here is a description of Antoine’s necklace; see Figure 4:
We take as X0 a solid torus (a solid doughnut). Inside of it, forming X1 , is a
finite collection of smaller disjoint solid tori linked together in a chain that circles
the interior of X0 . The set X2 is formed by a collection of even smaller chains, one
such chain circling the interior of each of the links making up X1 , and so forth.
The necklace X is the set of points that lie in each of the smaller and smaller sets
X0 , X1 , X2 , . . ..
Here is a property that distinguishes X from Y topologically:
Consider the circle C that goes through the hole in the doughnut X0 . Then C
does not bound even a singular disk in R3 \ X. (Proving this fact is nontrivial.)
Figure 4. Antoine’s Necklace
6.3.4. Hausdorff Dimension and a Fractal Cantor Set. Felix Hausdorff

suggested another measure of dimension based on measure rather than topology.
The middle-thirds Cantor set which we introduced in Definition 6.16 has, according
to Hausdorff, dimension equal to log 2/log 3 ≈ .63092975... as opposed to covering
dimension 0. Mandelbrot has defined a fractal as a set whose Hausdorff dimension
is not equal to its Lebesgue covering dimension [86, 87, 88]. Since the covering
dimension of the middle-thirds Cantor set is equal to 0, that set is, according to
Mandelbrot’s definition, a fractal.
Hausdorff observed that the volume of an n-dimensional cube is proportional

to the nth power of its diameter and suggested that, for an arbitrary set X, the nth
power of the diameter be taken as an approximation to the n-dimensional volume of
X, regardless of the fine-scale shape of the object X. To refine the approximation,
take an open (or closed) cover U of X by small sets and define the approximate n-
dimensional volume H n (X, U ) to be the sum of the nth powers |u|n of the diameters
|u| of the elements u ∈ U . Hausdorff noted that, in this manner it is possible to
assign every subset of a separable metric space an n-dimensional volume for every
n. Here are the details:
In order to make the n-dimensional volume of X well-defined, the covers of X
by small open sets should be as efficient as possible. To that end, fix a positive
number δ, consider all possible countable open covers U of X by sets of diameter
< δ, and define the δ-approximate n-dimensional volume Hδn (X) of X to be the
infimum of the approximations H n (X, U ). Finally, define the n-dimensional volume
H n (X) of X by the formula
H n (X) = lim Hδn (X).
δ→0
There is nothing in this definition that requires that n be a nonnegative integer.

Thus Hausdorff noted that one could assign an r-dimensional volume H r (X) to X
for every nonnegative number r by adding rth powers of diameters.
Hausdorff proved as a consequence of the following theorem that, given the set
X, there is a unique nonnegative number d such that H r (X) = ∞ for r < d and
H r (X) = 0 for r > d. He called this number d the dimension of X. We call it the
Hausdorff dimension of X.
Theorem 6.21. Suppose that r and s are nonnegative real numbers such that
r < s and that H r (X) < ∞. Then H s (X) = 0.
Proof. Pick a positive number P such that H r (X) < P . Then, for each
η > 0, there is a countable open cover Uη of X by sets of diameter < η such that
H r (X, Uη ) < P . Thus, for each δ > 0 and for each η < δ,

Hδs (X) ≤ H s (X, Uη ) = {|u|s : u ∈ Uη }

≤ η s−r {|u|r : u ∈ Uη } = η s−r H r (X, Uη ) < η s−r P.
Hence, for each δ > 0, Hδs (X) ≤ η s−r P . It follows that Hδs (X) = 0 and hence that
H s (X) = 0.
Hausdorff gave interesting examples of sets whose (Hausdorff) dimension is not

an integer. Here is one of his examples.
Theorem 6.22. The Hausdorff dimension of the middle-thirds Cantor set C
is r = log 2/log 3. Furthermore, H r (C) = 1.
Proof. We first show why it is natural to conjecture that r = log 2/ log 3 is
the Hausdorff dimension of C:
It is hard to imagine
∞covers of C that are more efficient than the defining
covers Cn , where C = n=1 Cn and Cn is the disjoint union of 2n intervals of
length (1/3n ). The 2n intervals of Cn define a closed cover Un of C that, for the
arbitrary dimension s, gives the following approximation to H s (C):
n
2
s
J (C, Un ) = {|u| : u ∈ Un } =
s
= en(log 2−s log 3) ,
3s
where we have used the identity ab = eb log a .
If log 2 − s log 3 = 0, which implies s = log 2/ log 3, we have H s (C, Un ) =
en·0
= 1, which is neither ∞ nor 0. For s < log 2/ log 3, H s (C, Un ) → ∞. For
x > log 2/ log 3, H s (C, Un ) → 0. We conclude that s = (log 2/ log 3 = r) is the
most probable dimension of C In view of the previous theorem, it remains only to
show that H r (C) = 1.
The estimates that we have already made show that H r (C) ≤ 1. It remains to
prove that H r (C) ≥ 1.
To that end, we suppose to the contrary that H r (C) < 1. We fix a positive
number δ, and consider an open cover U of C by sets of diameter < δ such that
H r (C, U ) < 1. Since C is compact, we may assume that the cover U is finite.
Without increasing the sum, we may assume that each element of U is an open
interval. Expanding each element of U a tiny bit if necessary, we may assume that
the end points of the interval lie in the complement of C. We replace these open
intervals by their corresponding closed intervals, and proceed to modify this cover
by closed intervals as follows.
If I = [a, b] is one of the intervals, we may shorten I until its endpoints a and
b actually lie in C. Let E denote the largest gap in C between the endpoints of I.
This gap divides I into three subintervals P , E, and P , with a ∈ P and b ∈ P .
We need three facts:
(1) 3r = 2. (Take the logarithm of both sides.)
(2) (3/2)(|P | + |P |) ≤ |P | + |E| + |P |. (This follows from the fact that
|P |, |P | ≤ |E|. In order to see that |P |, |P | ≤ |E|, consider the last n ≥ 0 such

that the endpoints a and b of the interval I lie in the same interval J of Cn . Then
E is the middle third of J, P is a subinterval of the first third of J, and P is a
subinterval of the last third of J.)
(3) ((|P | + |P |)/2)r ≥ (|P |r + |P |r )/2. (This follows since r < 1 so that the
function xr is concave. See Figure 5)
From (1), (2), and (3) we deduce

3
|I|r = (|P | + |E| + |P |)r ≥ ( (|P | + |P |))r
2
= 2((|P | + |P |)/2) ≥ 2(|P | + |P | )/2 = |P |r + |P |r .
r r r
Thus we may replace I by P and P .

A finite number of such operations can be used to reduce the covering to one
by intervals of the same Cn . But at that stage, the approximate area is ≥ 1 by
the calculation of the first half of the proof, contradicting our (false) assumption
that the approximate area was < 1, a contradiction, as we wished to show. Hence
H r (C, U ) ≥ 1.
Thus H r (C) = 1; and the proof is complete.
6.3.5. The Cantor Function. There is a natural map from the Cantor set
onto the unit interval. See an approximation to the graph of this function in Figure
7. This mapping is at the basis of the Cantor function: a continuous function that
has slope 0 almost everywhere, yet increases in such a way that it maps the unit
interval [0, 1] onto itself.
f (avg) ≥ avg(f )
f (b)
f ((a + b)/2)
A concave curve
f (a)+f (b)
2
f (a)
a (a + b)/2 b
Figure 5. A concave curve
The idea of the construction is to create a monotone decomposition of the

interval. Recall the definition of monotone decomposition in Section 5.2, Definition
5.8, and both Corollary 5.7 and Theorem 5.9 from that section. We describe the
construction in generality:
We partition [0, 1] into a collection X of points and closed subintervals. See
Figure 6. In the case of the Cantor set, the closed subintervals are the closures of
the open intervals removed in defining the Cantor set. The elements of X are then
equivalence classes of an equivalence relation on [0, 1]. We assume that there are
at least two, hence uncountably many, equivalence classes, most of which contain
only a single point of [0, 1]. At most countably many of the equivalence classes are
closed intervals, since each contains an open set and there can be no more than
countably many disjoint open sets.
Figure 6. Monotone decomposition of an interval
We recall the identification space or decomposition space whose points are the
elements of X. That is, each element of X, even if it consists of more than one
point of [0, 1], is abstractly considered to be one point of X. We assigned a topology
to X in the usual manner as follows. We considered the function p : [0, 1] → X
6.4. CANTOR SETS IN THE PLANE ARE TAME 69
which assigned to each point x ∈ [0, 1] the equivalence class p(x) = [x] ∈ X which
contains x. We declared a subset U ⊂ X to be an open subset of X iff the union
U ∗ = p−1 (U ) of the elements of U formed an open subset of [0, 1]. Note that,
by this definition, the map p : [0, 1] → X is continuous. We proved the following
earlier.
Theorem 6.23. The decomposition space X is an arc.
Definition 6.24. Let I denote the set of closed intervals [xα , yα ] in [0, 1] that
are closures of the open intervals deleted from [0, 1] in defining the middle-third
Cantor set C. Let X denote the partition of [0, 1] into the elements of I and the
singleton sets containing the points not included in any of the intervals of I. We
assign X the identification-space topology. Let p : [0, 1] → X denote the continuous
function that assigns each point x ∈ [0, 1] the element p(x) = [x] ∈ X which contains
x. Note that X is an arc by the previous theorem. Hence we may identify X with
[0, 1] and p with a map from [0, 1] to itself which fixes 0 and 1.
Theorem 6.25. Let I, X = [0, 1], and p : [0, 1] → X = [0, 1] be as in the
previous definition. Then we may assume that p(I) is the set of rational numbers
in the open interval (0, 1).
Proof. The set p(I) is a countable dense subset of (0, 1) ⊂ [0, 1] = X. Every
countable dense subset of (0, 1) is embedded in [0, 1] in exactly the same way.
Corollary 6.26. The set of points of the middle-thirds Cantor set C that are
not endpoints of any of the open intervals removed in forming C is homeomorphic
to the set of irrational points between 0 and 1.
Theorem 6.27. There is a (weakly) monotone continuous function f : [0, 1] →
[0, 1], with f (0) = 0 and f (1) = 1 that has derivative 0 almost everywhere.
Proof. Map the first middle third to 1/2. Map the next larger intervals to
1/4 and 3/4. Map the four next intervals to 1/8, 3/8, 5/8, and 7/8, etc. Extend
continuously to the remaining points of [0, 1]. Since the sum of the lengths of the
intervals defining the Cantor set is

1 2 n 1 1
· 1 + 2/3 + 2/3 + · · · + 2/3 + · · · = · = 1,
3 3 1 − (2/3)
the graph is horizontal on a set of measure 1 yet rises from 0 to 1.

6.4. Cantor Sets in the Plane Are Tame

We saw that there are Cantor sets in 3-dimensional space which are so entangled
that they can act like a necklace and not fall off when placed around the neck. Are
there Cantor sets in the plane with similar entanglement? We show in this section
that there are no such Cantor sets in the plane. Not only are Cantor sets in the
plane homeomorphic by definition to the standard middle-thirds Cantor set, but
the homeomorphisms can be realized as the restrictions of homeomorphisms from
the plane to itself. That is, every embedding of the Cantor set in the plane is
equivalent to the standard embedding. We call all standard embeddings of an
object tame embeddings.
Figure 7. The graph of the Cantor function
Theorem 6.28 (All Cantor sets in the plane are tame). Suppose C is the
middle-thirds Cantor set and f : C → R2 is an embedding of C in the plane. Then
the map f extends to a homeomorphism F : R2 → R2 .
In order to keep the proof of this theorem in reasonable bounds, we will assume
that finite families of disjoint polygonal disks in the plane can be moved around
at will, in the sense that any two such families consisting of the same number of
disks are embedded in the plane in the same way. If each of the disks is a triangle,
this is a fairly easy fact. With some tedium, the general fact can be reduced to the
following two exercises.
Exercise 6.29. If D is a polygonal disk in the plane, then, without introducing
any new vertices, D can be cut into triangles.
Exercise 6.30. If D is a polygonal disk in the plane and N is any open
neighborhood of D, then there is a homeomorphism of the plane fixed outside of N
that takes D to a triangular disk. [Hint: Use the previous exercise to assume that
D is a union of triangles. Inductively, move D into itself so as to make D have one
triangle fewer until there is only a single triangle remaining.]
The first step in proving this theorem is to prove that f (C) can be covered by
a collection of disjoint polygonal disks in the plane, each of small diameter. This
fact follows from the following more general theorem.
Theorem 6.31. Suppose that K is a compact subset of the plane and that > 0.
Then there is a finite collection D1 , . . . , Dk of disks-with-holes in the plane whose
interiors cover K, each having polygonal boundary, such that, for each i, Di lies in
the neighborhood of some component of K.
Proof. Fix a positive number δ. For each x ∈ K, let E(x) denote a polygonal
disk of diameter < δ whose interior contains x. Since K is compact, there is a finite
collection E(x1 ), . . . , E(x ) whose interiors already cover K. Each can be expanded
slightly so that their boundaries are in general position. (General position means
6.4. CANTOR SETS IN THE PLANE ARE TAME 71
in this case that two boundaries intersect, if they intersect at all, in finitely many
points, each in the interior of edges that cross one another at those points.) Then

the union Kδ = i=1 E(xi ) is a finite collection of disks-with-holes whose interiors
cover K, each having polygonal boundary.
We will show that, for some positive integer n, K1/n satisfies the conclusions of
the theorem. Suppose not. Then for each n > 0, there is a component Cn of K1/n
that is not in the neighborhood of any component of K. Passing to a subsequence
of C1 , C2 , . . . if necessary, we may assume that the sequence C1 , C2 , . . . converges
to a compact connected set C that necessarily is a subset of K since every point
of Cn is within 1/n of a point of K. See Theorem 3.16. Let N denote the
neighborhood of C. Then, for n sufficiently large, each of the sets Cn must lie in
N ; otherwise, again passing to a subsequence if necessary, we may assume that
there is a convergent sequence x1 , x2 , . . ., with xn ∈ Cn \ N , which converges to a
point x in the lim sup of the Cn ’s but not in C, a contradiction. This completes
the proof of the theorem.
Corollary 6.32 (Covering Lemma). If K is a Cantor set in the plane and

if > 0, then there is a finite collection of disjoint disks D1 , . . . , Dk in the plane
whose interiors cover K, each having polygonal boundary, such that, for each i, Di
has diameter < .
It will be convenient to collect groups of small disks into single disks. Since a
small disk can be shrunk inward to become as close to a single point as we like, the
following lemma about finite collections of points will suffice for our purposes.
Theorem 6.33 (Clumping Lemma). Suppose that F1 , . . ., Fk are disjoint finite
subsets of the plane, such that each Fi lies in the interior of some xy-square Si of
edge-length 1. Then F1 , . . ., Fk lie in the interiors of disjoint polyhedral disks D1 ,
. . ., Dk in the plane, where Di ⊂ Si .
Proof. See Figure 8. There are squares T1 , . . ., Tk in the interiors of the
squares S1 , . . ., Sk , respectively, all with the same edge length < 1, such that:
(1) For each i, Fi ⊂ int(Ti ), and
(2) No two boundary edges of the Ti ’s lie in the same horizontal line.
After a slight adjustment, we may assume:

(3) No two points of ki=1 Fi lie in the same vertical line.

Let Gi denote the union of the bottom edge Bi of Ti with the vertical segments
that begin on Bi and end at a point of Fi . A vertical segment S of one Gi may
intersect the bottom edges Bj of one or more Tj ’s. At such an intersection point,
we push a small segment of each of those Bj ’s up over the top of S to remove
the intersection. We obtain thereby disjoint, finite, connected, polygonal graphs
Gi ⊂ int(Si ) containing the points of Fi . The disjoint polyhedral disks D1 , . . ., Dk
desired can then be taken as small neighborhoods of the graphs Gi .
Corollary 6.34. Suppose C is the middle-third Cantor set and f : C → R2

is an embedding of C in the plane R2 . Let C1 , . . ., Ck denote disjoint open and
closed subsets of the middle-third Cantor set whose union is C such that each of the
sets f (Ci ) lies in some xy-square in the plane of edge length < . Then there exist
disjoint polygonal disks D1 , . . ., Dk in the plane, each lying in some xy-square of
edge length < such that, for each i, f (Ci ) ⊂ Di .
Figure 8. Enclosing disjoint sets into disjoint disks
Proof. Choose δ > 0 so small that any disk containing a point of some f (Ci )
cannot contain a point of another f (Cj ). Also choose δ > 0 so small that the δ
neighborhood of f (Ci ) still lies in an xy-square of edge length < . Use the Covering
Lemma to cover f (C) by disjoint disks of diameter < δ. Then these disks fall into
finite subcollections, namely, those containing a point of f (C1 ), those containing a
point of f (C2 ), etc. These subcollections can then be consolidated by the Clumping
Lemma 6.33 into single disks Di for each subcollection.
Proof that all Cantor sets in the plane are tame. The middle-third
Cantor set C is naturally covered by a collection I(n) of 2n intervals of length 1/3n
for every n. See Figure 9. Each of those collections of intervals can be thickened
slightly to create coverings D(n) by disks.
Figure 9. Standard partitions of the Cantor set
We wish to copy a subcollection of these standard disk covers by disk covers

for f (C).
We pick a large integer n1 . For each interval I ∈ I(n1 ), we cover f (C ∩ I) by
the interior of a polygonal disk D(I). We may choose these disks to be disjoint;
and, if n1 is sufficiently large, we may assume that the diameter of each D(I) is
< 1. Let the collection of these disjoint disks be D(n1 ).
6.5. EXERCISES 73
Assuming that D(nj ) has been chosen and that nj+1 > nj , for each interval
I ∈ I(nj+1 ), we cover f (C ∩ I) by the interior of a polygonal disk D(I). We may
choose these disks to be disjoint; and, if nj+1 is sufficiently large, we may assume
that the diameter of each D(I) is < j + 1 and that D(I) lies in the interior of a
disk element of D(nj ). Let the collection of these disjoint disks be D(nj+1 ).
Disjoint polygonal disks A1 , . . . , Ak in the interior of a disk A can be moved, in
order, to any other collection B1 , . . . , Bk in the interior of A by a homeomorphism
of A that fixes the boundary of A. Using this fact, we move the disks of D(n1 ) to
the disks of D(n1 ) by a homeomorphism of R2 . Then, fixing all points outside the
union of these disks, we can move the images of the disks of D(n2 ) to the disks of
D(n2 ). In general, fixing all points outside the union of the disks of D(nj ), it is
possible to move the images of the disks of D(nj+1 ) to the disks of D(nj+1 ). The
limit of these homeomorphisms is a homeomorphism g : R2 → R2 that, restricted
to f (C) is the inverse of f . This completes the proof of the theorem.
6.5. Exercises
6.4. The points of the middle-thirds Cantor set can be expressed explicitly as
infinite expansions using ternary numbers (symbols using only 0, 1, and 2). How
is that done? Likewise the points of the interval [0, 1] can be expressed as infinite
binary expansions. How is that done? Use these expansions to define an explicit
mapping from the middle-thirds Cantor set onto the interval [0, 1]. This, of course,
shows that there are as many points in the middle-thirds Cantor set as in the
interval [0.1].
6.5. How can you define different Cantor sets with different Hausdorff dimen-
sions?
CHAPTER 7
Algebraic Topology
Prerequisites: Algebraic topology was originally designed to prove the very

theorems we hope to prove in this section: the fact that every simple closed curve
in the plane separates the plane into two pieces and the fact that no arc in the
plane separates the plane. Our own very effective introduction to the topology
of the plane in graduate school, taught by C. E. Burgess, took these results as
foundational axioms. The reader may choose to do the same. These facts are
fundamental to the topology of 2-dimensional surfaces.
Nevertheless, since most of the readers we expect to enjoy the remainder of
this book will have had an introduction to algebraic topology, we will use four
results from the beginning of an algebraic topology course to prove these facts.
We will use reduced homology with integer coefficients, and will assume as known
the homology of a finite set, homotopy invariance, and the Mayer-Vietoris exact
sequence. My own personally preferred introductions to algebraic topology are
Massey [25], Munkres [27], and Hatcher [26].
The most important topological spaces are Euclidean space, the ball, and the
sphere.
Definition 7.1. The n-dimensional Euclidean space Rn is the set
{x = (x1 , . . . , xn ) | xi ∈ R},
where R denotes the line of real numbers. Euclidean space Rn has a standard
Euclidean metric,

d(x, y) = (x1 − y1 )2 + · · · + (xn − yn )2 .
The n-dimensional ball Bn is the subspace Bn = {x ∈ Rn | d(x, 0) ≤ 1}. The
(n − 1)-dimensional sphere Sn−1 is the subspace Sn−1 = {x ∈ Rn | d(x, 0) = 1}.
In dimension n = 2, the two most important facts are the following.
Theorem 7.2 (Arc Non-separation Theorem). If f : [0, 1] → S2 is an embed-
ding, then f ([0, 1]) does not separate S2 .
Theorem 7.3 (The Jordan Curve Theorem). If f : S1 → S2 is an embedding,
then f (S1 ) separates S2 into two components U and V and is the boundary of each.
Many introductions to algebraic topology are almost completely algebraic.
Therefore, it is probably not amiss to review some of the intuitive underlying ideas.
Homology theory counts the number of holes of various dimensions in a space.
See Figure 1.
A 0-hole arises when two points of the space cannot be connected by a path in
the space. Thus 0-dimensional homology H0 essentially counts the number of path
components of the space.
75
76 7. ALGEBRAIC TOPOLOGY
A 1-dimensional hole is represented by a (possibly singular) family of oriented

simple closed curves in the space. (The closed curves are called 1-cycles.) The hole
is trivial if the curves, taken together, are the oriented boundary of a surface in the
space (called 2-chains). Thus 1-dimensional homology H1 represents the nontriv-
ial 1-dimensional holes as the nonzero elements of an Abelian group, each element
represented by 1-cycles, where different families of curves are said to be equiva-
lent or to define the same hole if, their algebraic difference does bound a surface
(= 2-chain) in the space.
A 2-dimensional hole is represented by a (possibly singular) family of compact,
oriented surfaces without boundary (called 2-cycles). This 2-dimensional hole is
trivial if the oriented surfaces, taken together, are the oriented boundary of a 3-
dimensional “surface” (3-chain) in the space. Thus 2-dimensional homology H2
represents the nontrivial 2-dimensional holes as the nonzero elements of an Abelian
group, again up to equivalence.
Points in different components

represent a 0-dimensional hole.
The curve around the torus

(doughnut) represents a hole.
Figure 1. Intuitive homology
7.1. Facts Assumed from Algebraic Topology

The Jordan Curve Theorem and the Arc Non-separation Theorem are most
easily proved by the techniques of algebraic topology, and, in fact, the beginnings
of algebraic topology were designed to organize the ideas necessary for the proof. We
will take four simple facts from algebraic topology as known. We can easily deduce
from these facts three theorems that we shall prove about the reduced homology of
spheres Sn and of sphere- and ball-complements in Sn . We shall always use reduced
homology with integer coefficients, and all the spaces we consider will be nonempty,
locally connected metric spaces.
The minimal facts we assume from algebraic topology are the following:
Fact 0. If X is a finite collection of k points, then the reduced homology of X
is 0 in all dimensions except 0, and in dimension 0, H0 (X) = Zk−1 .
7.3. THE HOMOLOGY OF A BALL COMPLEMENT 77
Fact 1. If the reduced, 0-dimensional homology of the space X is Zk , then X

has k + 1 path components.
Fact 2. If two spaces are homotopy equivalent, then their reduced homologies
are isomorphic.
Fact 3. If a space X is a union of two open sets U and V with nonempty
intersection U ∩ V , then there is a long exact sequence in reduced homology of the
form
· · · → Hj (U ∩ V ) → Hj (U ) ⊕ Hj (V ) → Hj (U ∪ V ) → Hj−1 (U ∩ V ) → · · · ,
induced by inclusions.
7.2. The Reduced Homology of a Sphere

Theorem 7.4 (The reduced homology of a sphere). The reduced homology
Hk (Sn ) is trivial (= 0) in all dimensions except k = n, and, in that dimension
Hn (Sn ) = Z.
Proof. The proof is by induction on the dimension n of the sphere. The 0-
sphere S1 consists of two points, so that the reduced homology in dimension 0 is Z,
while the homology in all other dimensions is 0 (Fact 0, section 7.1).
Assuming the theorem in dimension n−1, with n > 0, we realize S n as the union
of two contractible open sets U and V , where U is the complement of the south pole
S = (0, . . . , 0, −1) and V is the complement of the north pole N = (0, . . . , 0, 1).
Hence the reduced homology of both U and V is trivial (Facts 0 and 2, section
7.1). The intersection U ∩ V is homotopy equivalent to Sn−1 . Hence, its reduced
homology is Z in dimension n − 1 and trivial in all other dimensions (Fact 2, section
7.1, and inductive hypothesis).
We examine a segment of the Mayer-Vietoris sequence of reduced homology of
the pair (U, V ) (Fact 3, section 7.1):
0 = Hj (U ) ⊕ Hj (V ) → Hj (U ∪ V ) → Hj−1 (U ∩ V ) → Hj−1 (U ) ⊕ Hj−1 (V ) = 0.
From the exactness of the sequence, we deduce that the central homomorphism is
an isomorphism. Hence, since U ∪ V = Sn and U ∩ V is homotopy equivalent to
Sn−1 , we find that
Hj (Sn ) ≈ Hj−1 (Sn−1 ).
The desired result is immediate.
7.3. The Homology of a Ball Complement

Theorem 7.5 (The homology of a ball complement). If f : Bk → Sn is an
embedding then the reduced homology H∗ (Sn \ f (Bk )) is trivial (= 0).
Proof. If k = 0, then the complement of f (Bk ) is contractible so that H∗ (Sn \
k
f (B )) is trivial as claimed (Facts 2 and 0, section 7.1). We consider k > 0 and
assume inductively that H∗ (Sn \ f (Bk−1 )) is trivial.
We realize Bk as a product Bk−1 × [0, 1], and we adopt the temporary notation
B[a, b] for the image f (Bk−1 × [a, b]) ⊂ f (Bk−1 × [0, 1]). If a = b, then we denote
B[a, b] by B[a].
We assume that, contrary to the theorem, there is a cycle z representing a
nonzero element of Hj (Sn \ B[0, 1]). We assume inductively that [a, b] is a subset
of [0, 1] and that z represents a nonzero element of Hj (Sn \ B[a, b]).
78 7. ALGEBRAIC TOPOLOGY
We claim that there is at least one half [a , b ] of the interval [a, b] such that z
represents a nonzero element of Hj (Sn \ B[a , b ]).
To that end, we let c denote the midpoint c = (a + b)/2 of the interval [a, b].
We set U = Sn \ B[a, c] and V = Sn \ B[c, b]. Then U ∩ V = Sn \ B[a, b] and
U ∪ V = Sn \ B[c]. We examine a segment of the reduced Mayer-Vietoris sequence
of the pair (U, V ) (Fact 3, section 7.1):
Hj+1 (U ∪ V ) → Hj (U ∩ V ) → Hj (U ) ⊕ Hj (V ) → Hj (U ∪ V ).
Since U ∪V = Sn \B[c] the inductive hypothesis implies that the first and last terms
of the segment are equal to 0. Hence, the central homomorphism is an isomorphism,
induced by inclusion of cycles. In particular, z represents a nonzero element either
of Hj (U ) or of Hj (V ), as claimed.
We conclude, inductively, that there are intervals [a1 , b1 ] = [0, 1], [a2 , b2 ],
[a3 , b3 ], . . ., each equal to one-half of the preceding, such that z represents a nonzero
element of Hj (Sn \ B[ai , bi ]) for each i = 1, 2, . . . . Let a denote the intersection of
these intervals. But B[a] is an embedding of the (k − 1)-ball in Sn so that, by
induction, z can only represent the trivial element of Hj (Sn \ B[a]). Hence, there is
a (j + 1)-dimensional chain c in Sn \ B[a] with boundary z. Since the chain c misses
B[a], it also misses all but finitely many of the sets B[ai , bi ]. But that contradicts
the fact that z represents a nontrivial element of Hj (Sn \ B[ai , bi ]).
We conclude that no such cycle z can exist, so that H∗ (Sn \B[0, 1]) is trivial.
7.4. The Homology of a Sphere Complement

Theorem 7.6 (The homology of a sphere complement). If f : Sk → Sn , with
k < n, is an embedding, then, the reduced homology Hj (Sn \ f (Sk )) is 0 in all
dimensions except j = n − k − 1, and in that dimension this reduced homology is Z.
Proof. Again we induct on the dimension k. If k = 0 so that Sk is a pair
of points, then Sn \ f (Sk ) is homotopically equivalent to Sn−1 . Hence its reduced
homology is Z in dimension n − 1 = n − k − 1 and 0 otherwise (Fact 2, section 7.1,
and homology of a sphere, theorem 7.4).
We consider k > 0 and assume inductively that the result is true for embeddings
of Sk−1 in Sn . Let f : Sk → Sn be an embedding. We let B1 denote the image
under f of the upper hemisphere of Sk and B2 the image of the lower hemisphere.
Then S = B1 ∩ B2 is the image of the equator Sk−1 of Sk , hence an embedded
image of a k − 1 sphere in Sn .
We let U = Sn \ B1 and V = Sn \ B2 . Then U ∪ V = Sn \ f (Sk−1 ) and
U ∩ V = Sn \ f (Sk ). We examine a segment of the reduced Mayer-Vietoris sequence
for the pair (U, V ):
Hj (U ) ⊕ Hj (V ) → Hj (U ∪ V ) → Hj−1 (U ∩ V ) → Hj−1 (U ) ⊕ Hj−1 (V ) or
0 ⊕ 0 → Hj (Sn \ f (Sk−1 )) → Hj−1 (Sn \ f (Sk ) → 0 ⊕ 0.

The first and last terms of this sequence are equal to 0 since they are the reduced ho-
mologies of ball-complements. Hence the central homomorphism is an isomorphism.
By inductive hypothesis, Hj (Sn \ f (Sk−1 )) is reduced homology of a sphere comple-
ment where the embedded sphere has dimension k−1. Hence this reduced homology
is nonzero only for j = n − (k − 1) − 1 and, in that dimension, is equal to Z (induc-
tive hypothesis). Hence the reduced homology Hj−1 (Sn \ f (Sk )) = Hj−1 (U ∩ V ) is
7.5. PROOF OF THE ARC NON-SEPARATION THEOREM 79
nonzero only for j − 1 = n − (k − 1) − 2 = n − k − 1 and, in that dimension, is equal

to Z.
7.5. Proof of the Arc Non-Separation Theorem

and the Jordan Curve Theorem
Proof that no arc separates S2 . This is a special case of the theorem on
ball complements (Theorem 7.5).
Proof of the Jordan Curve Theorem. By the theorem on sphere com-
plements, Theorem 7.6, the reduced homology of S2 \ f (S1 ) is Z in dimension
2 − 1 − 1 = 0. That is, S2 \ f (S1 ) has two components U and V (Fact 1, section
7.1).
Let x ∈ f (S1 ), and let N be a neighborhood of x in S2 . Let A be an arc in
f (S ) whose complement in f (S1 ) is an open arc B in N containing x. Since A
1
does not separate S2 , there is an arc A(U ) that connects a point u ∈ U irreducibly
to f (S1 ) but misses A. (That is, only the endpoints of A(U ) lie in {u} ∪ f (S1 ).)
The arc A(U ) lies, except for its terminal endpoint u in U , and u ∈ B. That is,
u ∈ ∂U . Thus U ∩ N is not empty, so that x is in the boundary of U . Similarly,
x is in the boundary of V . We conclude that f (S1 ) is the boundary of both U and
V.
CHAPTER 8
Characterization of the 2-Sphere
8.1. Statement and Proof of the Characterization Theorem

Recall that a Peano continuum X is a metric space that is compact, connected,
and locally connected. Recall also that, if U is a component of an open subset of
X, then U is open and arcwise connected. (See Corollary 5.23 and Theorem 5.24.)
Theorem 8.1 (Characterization of the 2-Sphere). Suppose that X is a Peano
continuum having more than one point, that X is not separated by any arc, and
that X satisfies the Jordan Curve Theorem: each simple closed curve J in X sep-
arates X into two components U and V and is the boundary of each. Then X is
homeomorphic to the 2-sphere S2 .
Remark. A small modification of this characterization of S2 is called the Zippin
sphere characterization. I learned a proof from Wilder [90, p. 88], though the proof
here is a modification of his. End remark.
Furthermore, if J is a simple closed curve in X with complementary domains
U and V , then each of J ∪ U and J ∪ V is homeomorphic to the standard 2-ball B2 .
Remark. That the 2-sphere satisfies the hypotheses was proved in the previous
chapter. End remark.
Schoenflies’ Theorem is an obvious special case of the theorem.
Corollary 8.2 (Schoenflies’ Theorem). If H is a simple closed curve in the
2-dimensional plane, then J bounds a disk in that plane.
Proof. By stereographic projection, we may complete the plane by adding
one point at infinity to form a 2-sphere. Then J bounds two disks in that 2-sphere,
one of them in the original plane.
Before we proceed to the proof, we explain some convenient terminology.
Definition 8.3. We say that an arc A irreducibly joins disjoint sets B and C
if A has one endpoint in B, the other endpoint in C, and interior in the complement
of B ∪ C.
Definition 8.4. Suppose that J ⊂ X is a simple closed curve, U is one of the
two complementary domains of J in X, and D = J ∪ U . We say that an arc A is
properly embedded in D if the endpoints of A are in J and the interior of A is in U .
We say that A spans J through U . See Figure 1.
Remark. That X contains a simple closed curve J is an easy exercise, as

we shall explain next. By the Jordan Curve Theorem, J separates X into two
components U and V . That X is a 2-sphere will follow easily once we prove that
81
82 8. CHARACTERIZATION OF THE 2-SPHERE
A
U
Figure 1. The properly embedded arc
J ∪ U and J ∪ V are disks. The proof that each of the sets J ∪ U and J ∪ V is a
disk will require a number of important separation lemmas, which will occupy the
majority of this chapter. End remark.
Proof that X contains a simple closed curve. Since X has more than
one point and, as a Peano continuum, is arcwise connected, we see that X contains
an arc A joining two points x and y. Let B denote a subarc of A in the interior of
A. Since no arc separates, B does not separate x from y. Hence, since components
of open sets in a Peano continuum are arcwise connected, there is an arc C joining
x and y that does not intersect B. The arc C contains a subarc D that irreducibly
joins one of the two components of A \ B with the other component. The arc A
contains a subarc E that joins the endpoints of D. Then J = D ∪ E is a simple
closed curve.
Our first lemma shows how to create spanning arcs with controlled endpoints.
Lemma 8.5 (Spanning-Arc Lemma). Suppose that J ⊂ X is a simple closed
curve separating X into two components U and V . Suppose that α and β are
disjoint subarcs of J. Then there is an arc A properly embedded in D = J ∪ U with
one endpoint in α and the other in β. (We say that A spans J through U from α
to β.) See Figure 2.
Proof. Since every point of J is in the closure of U and since X is locally

arcwise connected, there is a small arc α that begins at a point x ∈ U , ends at a
point x ∈ α, and intersects J only in α. Passing to a subarc if necessary, we may
assume that α irreducibly joins x to J, with the second endpoint x in α. Similarly,
there is an arc β irreducibly joining a point y ∈ U to J, with the second endpoint
y in β. Since U is arcwise connected by (1), there is an arc B in U that joins x to
y. The union α ∪ B ∪ β contains an arc A spanning D from α to β.
The following very intuitive lemma is the most important of the separation
lemmas.
Lemma 8.6 (Arc-crossing Lemma). Suppose that J ⊂ X is a simple closed curve
with complementary domains U and V , that D = J ∪ U and E = J ∪ V . Suppose
further that A and B are arcs that span J through U and that the endpoints a1 and
a2 of A separate the endpoints b1 and b2 of B on J. Then the arcs A and B must
intersect. See Figure 3.
8.1. STATEMENT AND PROOF OF THE CHARACTERIZATION THEOREM 83
A
U
Figure 2. The Spanning-arc Lemma

a1
b1 B b2
J a2
Figure 3. The Arc-crossing Lemma
Proof. Suppose to the contrary that the arcs A and B are disjoint. We may
assume that the endpoints appear in the order a1 , b1 , a2 , b2 on J. By the Spanning-
Arc Lemma 2, there is an arc C that spans J through V with endpoint c1 between
b1 and a2 and with endpoint c2 between b2 and a1 .
We obtain a contradiction by showing that the simple closed curve K = A ∪
a2 c1 ∪ C ∪ c2 a1 does not separate X. In fact, we show that every point x in
the complement of K can be joined by an arc in X \ K to the connected set
L = B ∪ (J \ K). If x is not already in L, then x must either lie in U \ K or
in V \ K. If x ∈ U \ K, then there is an arc α irreducibly joining x to J \ K in
the complement of the arc c2 a1 ∪ A ∪ a2 c1 . The arc α must miss K entirely. If
x ∈ V \ K, then there is an arc β irreducibly joining x to J \ K in the complement
of the arc a2 c1 ∪ C ∪ c2 a1 . The arc β must miss K entirely.
We show the strength of the Arc-crossing Lemma by deducing some important
corollaries.
Corollary 8.7 (Theta Curve Theorem). Consider a space Θ that is the union
of three arcs A, B, and C having disjoint interiors but that share their endpoints x
and y. Then the complement X \ Θ has three components, namely U with boundary
A ∪ B, V with boundary B ∪ C, and W with boundary C ∪ A. See Figure 4.
x B y
W
C
Figure 4. The Theta Curve Theorem
Proof. Since it requires two of the arcs A, B, and C to separate X, given any
point z in the complement of Θ, there are two arcs α(z) and β(z) with initial point
z and terminal points α (z) and β (z), each arc irreducibly connecting z to Θ, such
that α (z) and β (z) lie in the interiors of different arcs among the three A, B, and
C. We suppose for purposes of argument that α (z) ∈ A and β (z) ∈ B.
The union α(z) ∪ β(z) contains an arc spanning J = A ∪ B through one of the
components U of X \ (A ∪ B). Since its terminal endpoints α (z) and β (z) separate
the ends x and y of C in the simple closed curve A ∪ B, and since α(z) and β(z)
do not intersect C, it follows from the Arc-crossing Lemma 8.6 that C lies in the
other component U of X \ (A ∪ B).
Since all points of U can be irreducibly joined to Θ by an arc ending in the
interior of C, and since all the points of U can be irreducibly joined to Θ by arcs
ending in the interiors of both in A and B, we conclude that U is characterized by
that property.
We find similarly that the points joinable to both B and C form an open set V
bounded by B ∪ C and that the points joinable to both C and A form an open set
W bounded by C ∪ A. The three sets U , V , and W exhaust the points of X \ Θ,
and the proof of the corollary is complete.
Corollary 8.8 (Arc-separation Lemma). Assume that J ⊂ X is a simple

closed curve. Assume that, in addition, B and C are disjoint arcs in D = J ∪ U ,
intersecting J only at single endpoints b ∈ B and c ∈ C. Then there is an arc
A properly embedded in D which misses B ∪ C and whose endpoints separate the
points b and c in J. The arc A separates B from C in D. See Figure 5.
B
C
U
J
Figure 5. The Arc-separation Lemma
Proof. By the Spanning-arc Lemma 8.5, there is an arc γ properly embedded

in E = J ∪ V having one endpoint b near b and the other endpoint c near c. The
arc α = B ∪ bb ∪ γ ∪ c c ∪ C does not separate X. Hence, there is an arc A in X \ α
that irreducibly joins one component of J \ (bb ∪ c c) to the other component. Since
its endpoints separate b from c in J, it follows from the Arc-crossing Lemma 8.6
that int(A) cannot be in V ; for otherwise, A would have to intersect γ. Thus int(A)
lies in U , so that A is properly embedded in D, as desired. If A did not separate
B from C in D, then B and C could be joined in D \ A, and we would find a new
arc B properly embedded in D, disjoint from A, but with endpoints separating the
endpoints of A in J, a contradiction. Thus A separates B from C in D.
Corollary 8.9 (Contrary Neighbor Theorem). It is impossible to have six

points A, B, C, x, y, and z in X and arcs Ax, Ay, Az, Bx, By, Bz, Cx, Cy, Cz
in X that intersect only in their endpoints. (This graph is usually denoted K3,3 .)
See Figure 6.
A B C
x y z
Figure 6. The Contrary Neighbor Graph
Remark. The points A, B, and C are interpreted as neighbors, each of which

needs a line to x=electricity, y=water, and z=telephone. They dislike each other
so much that they insist that their individual lines not cross one another. The
theorem says that their desires cannot be honored. End remark.
y x z
Figure 7. The impossibility of embedding the contrary neighbor

graph K3,3
Proof. Assume to the contrary that the indicated graph can be embedded
in X. See Figure 7. Consider the simple closed curve J = Ax ∪ xC ∪ Cy ∪ yA.
The arcs AzC and xBy would have to be properly embedded arcs whose endpoints
separate one another on J. By the Arc-crossing Lemma 8.6, one of these two arcs
would have to have interior lying in one domain U of X \ J, and the other would
have to have interior lying in the other domain V of X \ J. But the arc Bz would
then have to cross the curve J, which separates B from z, a contradiction.
Corollary 8.10. It is impossible to embed the complete graph K5 on five
vertices in X. See Figure 8.
Proof. Assume that K5 can be embedded in X. Let a, b, c, d, e denote the

five vertices. Define J = ab ∪ bc ∪ cd ∪ da. The edges ae, be, ce, de must all lie in
one component U of X \ J. The endpoints of the edge ac separate the endpoints
of be ∪ ed in J. Hence, by the Arc-crossing Lemma 8.6, ac must lie in the other
component V of X \ J. Similarly, the edge bd must also lie in V . But the endpoints
of ac and bd separate one another in J. Hence ac and bd must intersect by the
Arc-crossing Lemma, a contradiction.
Our goal is to show that D can be divided into small pieces by finitely many
arcs in D that intersect one another nicely. To organize the argument, we make
the following simple, yet technical, definition.
a b
d c
Figure 8. The complete graph on five vertices
Definition 8.11. We assume X, J, U , and D = J ∪ U given as above. A

separation triple (A, B, C) consists of an arc A that spans J through U and compact,
connected subsets B, C ⊂ D that intersect J and are separated in D by A. See
Figure 9.
The goal of the next lemma is to show that D can be divided into small pieces
by a finite number of arcs properly embedded in D. At this point we do not require
that the arcs intersect one another nicely.
Lemma 8.12 (Separation-triples Lemma). Given > 0, there is a finite collec-
tion (A1 , B1 , C1 ), . . . , (Ak , Bk , Ck ) of separation triples in D such that, if x and y
are two points of D at distance at least from each other, then there is one of the
separation triples (Ai , Bi , Ci ) such that x is in one of the two sets Bi and Ci , while
y is in the other.
Proof. We first concentrate on separating a single pair of points. Let x and y

be points of D that are at distance at least from each other. There are disjoint arcs
β and γ joining x to J and y to J, respectively. By the Arc-separation Lemma 8.8,
there is an arc A properly embedded in D that separates β from γ in D. Let N (x)
and N (y) denote small connected open sets about x and y, respectively. Define B
to be the union of β with the closure of N (x). Define C to be the union of γ with
the closure of N (y). Then (A, B, C) is a separation triple, and A not only separates
x from y but also separates all points close to x from all points close to y.
We then note that the set S of all pairs (x, y), where x and y are points of
D that are at least apart is a compact set and that A separates an entire open
neighborhood of such pairs from one another. Hence there is a finite collection of
separation triples in D that separate all such pairs.
We now have to define what we mean when we say that finitely many arcs divide
D into small pieces and that the arcs intersect one another nicely. We shall make
B
C
Figure 9. A separation triple
use of the Theta Curve Theorem 8.7 which, in application, says that a spanning
arc in D divides D into two domains bounded by simple closed curves.
Definition 8.13. Let X, J, U , and D = J ∪ U be given, as above. A simple
subdivision of D is a tuple (D, α1 , α2 , . . . , αk ) such that α1 is a spanning arc in D
that divides D into two closed domains D0 and D1 , α2 is a spanning arc either
of D0 or of D1 that divides that domain into two closed domains, etc. Thus,
thinking inductively, (D, α1 , . . . , αk−1 ) is a simple subdivision that divides D into
finitely many closed domains, each bounded by a simple closed curve; and αk is a
spanning arc in one of those domains that divides it into two closed domains. The
closed domains into which α1 , . . . , αk divide D are called the cells of the simple
subdivision. The arcs α1 , . . ., αk are called the arcs of the simple subdivision. The
union of these arcs with J is called the 1-skeleton of the subdivision.
Lemma 8.14 (Subdivision Lemma). Given > 0, there is a simple subdivision
(D, α1 , . . . , αk ) of D each of whose cells has diameter < .
Proof. The Separation-Triples Lemma 8.12 states that there is a finite col-
lection (A1 , B1 , C1 ), . . . , (Ak , Bk , C ) of separation triples in D such that, if x and
y are two points of D at distance at least from each other, then there is one of
the separation triples (Ai , Bi , Ci ) such that x is in one of the two sets Bi and Ci ,
while y is in the other.
There is a positive number δ that is smaller than the distance from each of the
arcs Ai to its companion sets Bi and Ci . We now process each of the arcs Ai in
turn. Each will supply us with a finite number of the curves αj .
First, we keep the arc A1 as α1 .
Assume inductively that arcs A1 , . . ., Ai−1 have been processed, yielding a
simple subdivision (D, α1 , . . . , αk(i−1) ) of D. Consider the components of Ai \ (J ∪
α1 ∪ · · · αk(i−1) ) whose endpoints cannot be joined in J ∪ α1 ∪ · · · αk(i−1) by an arc
of length < δ. Since J ∪ α1 ∪ · · · αk(i−1) is locally connected and the components of
Ai \ (J ∪ α1 ∪ · · · αk(i−1) ) are countable in number with diameters approaching 0,
there are only finitely many components whose endpoints cannot be so joined. We
add the closure of each such component to the collection of αj ’s.
8.2. EXERCISES 89
After we have processed each of the arcs Ai , we have a simple subdivision

(D, α1 , . . ., αk ) of D. We need to show that each cell of the subdivision has
diameter < .
It clearly suffices to show that the arcs αj separate each Bi from the correspond-
ing Ci . In order to see that, we consider all of the countably many components of
Ai \ (J ∪ α1 ∪ · · · αk(i−1) ). Those that we have not already incorporated as arcs
of our subdivision have endpoints that can be joined by arcs of diameter < δ in
the 1-skeleton of the subdivision. We replace them all in that manner, with the
diameters of the replacements going to 0 in diameter. The result is a continuous
function from the arc Ai into the 1-skeleton of the subdivision. The continuous
image of Ai misses Bi ∪ Ci since δ is so small. This image therefore contains an arc
that separates Bi from Ci in D.
We are finally ready to prove the final portion part (ii) of the main theorem.
We restate that theorem for the reader’s convenience.
Theorem 8.15. The set D = J ∪ U is homeomorphic with the standard unit
disk B2 .
Proof. It is an easy matter to show that each simple subdivision of D can be
copied by a simple subdivision of B2 so that arcs, cells and adjacencies of the two
subdivisions are combinatorially isomorphic.
On the other hand, it is an easy matter, using the Spanning-Arc Lemma 8.5,
to show that every simple subdivision of B2 in which no successive subdivision arc
has an endpoint coinciding with an endpoint of any previous subdivision arc can
be copied combinatorially by a subdivision of D.
Using these facts, we proceed inductively as follows.
Let S1 be a subdivision of D so that no cell has diameter larger than 1.
Let T1 be an isomorphic subdivision of B2 .
Let T1 be a subdivision of T1 so that no cell of T1 has diameter larger than 1
and whose new endpoints do not coincide with previous endpoints.
Let S1 be an isomorphic subdivision of S1 .
Let S2 be a subdivision of S1 so that no cell has diameter larger than 1/2.
Let T2 be a subdivision of T1 isomorphic to S2 .
Let T2 be a subdivision of T2 so that no cell of T2 has diameter larger than 1/2
and whose new endpoints do not coincide with previous endpoints.
Let S2 be an isomorphic subdivision of S2 .
Etc.
The process creates, in the limit, a bijection between D and B2 that is contin-
uous in both directions.
8.2. Exercises
8.1. Show that a space containing either of the graphs in Figure 10 cannot
satisfy the Jordan Curve Theorem.
8.2. Show that each of the two graphs in Figure 10 can be embedded in a
sphere with handles. Does a torus suffice? (One handle.)
8.3. If D denotes the square and A is an arc with its endpoints on opposite
sides of D, then A separates the top and bottom of D. (See Figure 11.) However,
there is a connected subset K of D that contains both sides of D but does not
separate the top of D from the bottom of D. Construct such a set K.
Figure 10. Skew curves
Figure 11. Arcs crossing a disk
[Hint: If C is a connected set and C ⊃ C is contained in the closure of C, then

C is also connected.]
8.4. Construct a connected subset of the plane that is the union of countably
many disjoint arcs.
CHAPTER 9
2-Manifolds
9.1. Definition and Examples

A 2-manifold is a 2-dimensional surface with or without boundary. Riemann
associated a surface called a Riemann surface with each complex function (see [14]).
Surfaces are used in complex analysis, in computer aided design, in number theory,
in algebraic geometry, and as foundation for higher-dimensional topology.
The characterization of the sphere and of the disk open the way to the study
of more general surfaces. At this point, we simply give the basic definition and
describe the basic examples so that we can refer to them when we need to. In a
later chapter, we will prove theorems that classify and characterize the compact,
connected 2-manifolds and 2-manifolds-with-boundary. For the sake of simplicity,
we restrict our considerations to metric spaces.
Definition 9.1. A metric space X is a 2-manifold or a surface without bound-
ary if each point has an open neighborhood homeomorphic to the plane.
A metric space Y is a 2-manifold-with-boundary or surface with boundary if
each point has an open neighborhood homeomorphic to an open subset of the
closed upper half-plane
{x = (x1 , x2 ) ∈ R2 | x2 ≥ 0}.
Example 9.2 (The Möbius band). There are two types of compact, connected
2-manifolds: the nonorientable manifolds that contain a Möbius strip and the ori-
entable ones that do not. The Möbius strip is the space obtained by identifying the
ends of a strip of paper after a half twist. See Figure 1.
Example 9.3. The compact, connected, orientable 2-manifolds are the 2-

spheres with some number ≥ 0 of handles. The torus is a sphere with one handle.
See Figure 2.
Example 9.4. The compact, connected, nonorientable 2-manifolds arise from

the orientable ones when one cuts some number of disks from the sphere with
handles and inserts, in their place, Möbius bands.
9.2. Exercises
The Four-color Theorem says that every map on the sphere can be colored
with four colors in such a way that no two countries share a common edge. (A map
consists of a number of polygonal disks with disjoint interiors, but possibly sharing
parts of their boundary.) This theorem is extremely difficult. The same result is
not true for other surfaces.
91
92 9. 2-MANIFOLDS
a b
b a
Figure 1. The Möbius strip
Figure 2. The sphere with handles
9.1. Construct 4 disks in the plane that have disjoint interiors such that each
shares an edge with the other three.
9.2. Show that it is impossible to have 5 disks in the plane that have disjoint
interiors such that each shares an edge with the other four. (This result is a
necessary property in order that the Four-color Theorem be true, but it is not
sufficient. The subtle question is whether some exotic complicated map might force
two countries to share an edge, not because of some local configuration, but because
of some global configuration.)
9.2. EXERCISES 93
9.3. Construct 5 disks in the torus that have disjoint interiors such that each
shares an edge with the other 4.
9.4. On a sphere with n handles, how many disks with disjoint interiors can
share a common edge with each of the others?
CHAPTER 10
Arcs in S2 Are Tame
Our goal is to understand arcs in all 2-manifolds. The fundamental case is the
arc in the 2-sphere S2 .
Definition 10.1. An embedding f : [0, 1] → Sn is called tame if there is a
homeomorphism h : Sn → Sn such that h ◦ f ([0, 1]) lies in a great circle. That is, f
is equivalent to a standard embedding. Otherwise, the embedding f is called wild
(and similarly for embeddings in Rn ).
Before considering arcs in the general 2-manifold, we show that every arc in
the 2-sphere S2 is tame and that tame arcs can be pushed around at will.
Example 10.2. The diagram illustrates a wild embedding of [0, 1] in R3 .
Figure 1. An Infinitely Knotted Arc
10.1. Arcs in S2 Are Tame

Theorem 10.3. If f : [0, 1] → S2 is an embedding of the unit interval [0, 1] ⊂
R ⊂ R2 ∪ {∞} ⊂ S2 in the 2-sphere S2 , then the map f extends to a homeomor-
2
phism F : S2 → S2 . In other words, every arc in S2 is tame.

Proof. It suffices to prove that A = f ([0, 1]) is a subset of a simple closed
curve J in S2 , since we have proved that every simple closed curve is embedded in
S2 in exactly the same way a circle is embedded in S2 . Thus the theorem follows
from the following lemma.
Lemma 10.4. Every arc A in the 2-sphere is a subset of a simple closed curve
J in the 2-sphere.
Proof. Let B ⊂ A denote a subarc of the interior of A that separates A into
two components C1 and C2 . Since no arc separates the plane, there is an arc B
that irreducibly connects C1 to C2 . The endpoints of B bound an arc B in A
95
96 10. ARCS IN S2 ARE TAME
with B ⊂ int(B ). The union J = B ∪ B is a simple closed curve. All simple

closed curves in S2 are embedded in the same way, so that we lose no generality in
assuming that J is a circle. See Figure 2. Since a circle locally divides S2 into two
components, the same is true of B. Furthermore, int(B) can be pushed slightly
into each component of S2 \ J so as to form a simple closed curve J of which B is
a spanning arc.
B
J
Figure 2. Each arc locally separates the plane
Repeat the argument in order to construct a doubly infinite sequence

. . . , J−2 , J−1 , J0 , J1 , J2 , . . .
of simple closed curves enclosing overlapping portions of the arc A as in the following
diagram, Figure 3.
J−1
J2
J−2 J1
J0
Figure 3. A sequence of thickenings
Have the diameters of the curves Jn approach 0 as n → ±∞ and have the curves

±∞
Jn approach the ends of A as n → ±∞. Then the union n Jn \ A, together with
the endpoints of A, will contain two arcs joining the endpoints of A. The union of
A with one of these arcs will form the desired simple closed curve containing A.
10.2. DISK ISOTOPIES 97
Scholium 10.5. The simple closed curve J ⊃ A may be chosen arbitrarily close
to A.
Corollary 10.6. Every arc A = f ([0, 1]) in the plane R2 locally separates R2
into two components.
The proof of the lemma is essentially local so that we can conclude the following.
Lemma 10.7. If A is an arc in a 2-manifold, then A is a subset of a simple
closed curve in the 2-manifold.
10.2. Disk Isotopies

Definition 10.8. An isotopy of the space Y is a map F : Y × [0, 1] → Y such
that, for each t ∈ [0, 1], F |(Y × {t}) is a homeomorphism from Y to itself and
F |Y × {0} is the identity map on Y .
If f : X → Y is an embedding, then an isotopy of f in Y is a map F :
X × [0, 1] → Y such that, for each t ∈ [0, 1], F |X × {t} is an embedding and
F |X × {0} = f .
The isotopy F is said to be a global isotopy or an ambient isotopy if there is an
isotopy G of Y such that, for each (x, t) ∈ X × [0, 1], G((f (x), t)) = F (x, t).
As far as 2-dimensional spaces go, our goal is show that spanning arcs with the
same endpoints in a disk are unique up to isotopy. In the process, we will explain
Alexander’s Trick which shows that homeomorphisms of a ball fixing the points of
the boundary are isotopic to the identity.
Theorem 10.9 (Boundary extension). If f : ∂Bn → ∂Bn is a homeomorphism
from the sphere Sn−1 = ∂Bn to itself, then there is a homeomorphism F : Bn → Bn
that extends f .
Proof. Every point x ∈ Bn \ {0} has a unique representation of the form
x = |x| · (x/|x|).
Define F (0) = 0 and F (x) = |x| · (f (x/|x|)).
Theorem 10.10 (Alexander’s Trick). Suppose that f : Bn → Bn is a homeo-
morphism from the disk to itself that is the identity on the boundary of Bn . Then
f is isotopic to the identity.
Proof. We must construct a map F : Bn × [0, 1] → Bn such that, for each
t ∈ [0, 1], F |Bn × {t} is a homeomorphism onto Bn ; for each x ∈ Bn , F (x, 0) = f (x)
and F (x, 1) = x. The idea is to shrink f toward 0 until it disappears at time t = 1
and to spread the identity from the boundary of Bn until it completely fills B2 at
time t = 1. In formulae:
F (x, t) = x if |x| ≥ 1 − t,
x
F (x, t) = (1 − t)f if 0 < |x| < 1 − t,
1−t
F (0, t) = (1 − t)f (0).

Corollary 10.11 (Uniqueness of spanning arcs). If D is a disk and A and
B are arcs spanning D through the interior of D with ∂A = ∂B, then there is an
isotopy of D fixing the boundary of D that takes A to B.
Proof. Let X and Y denote the arcs into which ∂A divides ∂D. Let D1
and D2 denote the two disks into which A divides D, with ∂D1 = A ∪ X and
∂D2 = A ∪ Y . Let E1 and E2 denote the two disks into which B divides D, with
∂E1 = B ∪ X and ∂E2 = B ∪ Y . Let g : A → B denote a homeomorphism
that is the identity on ∂A = ∂B. By the Boundary Extension Theorem, there are
homeomorphisms G1 : D1 → E1 and G2 : D2 → E2 that are equal to the identity
on ∂D and are equal to g on A. Then G = G1 ∪ G2 : D → D is a homeomorphism
that is the identity on ∂D. By Alexander’s trick, Theorem 10.10, there is an isotopy
from the identity to G. This isotopy takes A to B.
Exercise 10.12. If C and D are disks and C ∩ D is an arc in the boundary of

each, then C ∪ D is a disk.
Exercise 10.13. If A is an arc in a 2-manifold X, then there is a disk D in X
containing A in its interior.
Lemma 10.14 (Arc-sliding Lemma). If A is an arc in a 2-manifold X and
f : A → A is a homeomorphism fixing ∂A, then there is an isotopy of X, fixed
outside an arbitrarily small neighborhood U of int(A) that takes id|A to f |A.
Proof. There is a disk D in U ∪ ∂A of which A is a spanning arc. There is
a homeomorphism h : D → D that is the identity on ∂D and that restricts to f
on A. By Alexander’s Trick, Theorem 10.10, there is an isotopy of X, fixed on ∂D
and on X \ D, that takes id|D to h|D. Thus this isotopy takes id|A to f |A.
Lemma 10.15 (Edge-adjustment Lemma). If A is an arc in a 2-manifold X,

F and F are finite subsets of int(A) having the same number of points, and U is
an arbitrarily small neighborhood of int(A), then there is an isotopy of X that is
the identity at level 0, takes A to itself at every level, and takes F to F at level 1.
Proof. There is a disk D in U ∪ ∂A of which A is a spanning arc. There
is a homeomorphism f : A → A that takes F to F . There is a homeomorphism
F : D → D that is the identity on ∂D and restricts to f on A. By the Alexander
Trick, , Theorem 10.10, there is an isotopy from the identity on D to F . This isotopy
takes F to F and satisfies the requirements of the Edge-adjustment Lemma.
Lemma 10.16 (Straightening Lemma). Let P denote a convex polygon in the

plane R2 , and let A1 , . . . , Ak denote a finite, disjoint family of spanning arcs for
P having the property that no Ai has both of its endpoints in the same side of P .
Then there is an isotopy of P fixed on ∂P such that each of the arcs Ai has image
that is straight.
Proof. By hypothesis, there is a straight arc A1 in P joining the endpoints of
A1 . By the corollary above, there is an isotopy of R2 that fixes every point outside
of P that takes A1 to A1 . The arc A1 subdivides P into two convex polygons in
the plane. The images of the arcs A2 , . . . , Ak lie in these convex polygons as the
arcs A1 , . . . , Ak lay in P . By induction, these new arcs can be straightened.
Lemma 10.17 (Polygonal-curve Lemma). Suppose that J is a simple closed

curve in a triangulated 2-manifold X and that U is a neighborhood of J. Then
there is an isotopy of X that if fixed outside of U and takes J to a polygonal simple
closed curve.
10.2. DISK ISOTOPIES 99
Proof. Suppose that a ∈ J cuts J into the open arc α and that, in some
natural linear order on α, points b, c, d appear in the order b < c < d. See Figure 4.
d
a
Figure 4. The setting
We have seen that there is a simple closed curve K1 bounding a disk D1 con-
taining the arc abcd ⊂ J as a spanning arc and excluding the complementary arc
da ⊂ J. See Figure 5.
a d
K1
a
d
b D1
b c
c
Figure 5. The polygonal replacement
There is a (dotted) arc B = a d irreducibly joining ab and cd in D1 that is a

finite union of straight line segments. We obtain a new spanning arc for the disk
D1 by replacing the subarc of abcd bounded by the endpoints a and d of B by
the arc B itself. Since all spanning arcs of D1 joining points a and d are equivalent
by a homeomorphism of D1 that fixes the boundary of D1 , we find that we may
assume B is a subarc of our new simple closed curve J .
By Alexander’s trick, the old curve can be moved to the new curve by isotopy.
We now consider points a , b , c , d in J , where a and d are the endpoints of
the path B and b and c are points in the interior of B. We construct a simple closed
curve K2 very close to the arc c d a b , with K2 bounding a disk D2 that contains
the arc c d a b as a spanning arc and excludes the arc b c . The arc c d a b can
be replaced by a spanning arc that is a finite union of straight line segments. The
curve J can then be moved to the resulting simple closed curve J that satisfies
the requirements of the lemma.
10.3. Exercises
10.3. Remove a disk from a torus to create a hole in the torus. Show how to
turn this punctured torus inside out through the hole.
10.4. In the punctured torus, insert a rope and tie it around the center of the
torus. When the punctured torus is turned inside out, what happens to the rope.
It ends up on the outside of the torus. Does it fall away from the torus?
10.5. As we have seen, there is only one embedding of an arc or simple closed
curve in the 2-sphere S2 up to equivalence. That is, if f, g : [0, 1] → S2 are
embeddings, then there is a homeomorphism h : S2 → S2 such that, for each
x ∈ [0, 1], h(f (x)) = g(x). Construct a graph Γ such that there are two embed-
dings f, g : Γ → S2 for which no such homeomorphism h exists.
The Straight Arc
Two Infinitely Knotted Arcs
Figure 6. Three arcs
10.6. Figure 6 indicates three embeddings of [0, 1] in R3 . Two of them are

equivalently embedded in R3 . That is, if we denote those two by f, g : [0, 1] →
R3 , then there is a homeomorphism h : R3 → R3 such that, for each x ∈ [0, 1],
h(f (x)) = g(x). Which two are equivalently embedded? Can you construct the
homeomorphism?
CHAPTER 11
R. L. Moore’s Decomposition Theorem
When R. L. Moore completed his own characterization of the 2-dimensional

sphere, he realized that there was a wonderful corollary. (See [89].)
Theorem 11.1 (The Moore Decomposition Theorem [89]). Suppose that f :
S2 → X is a continuous map from the 2-sphere S2 onto a Hausdorff space X such
that, for each x ∈ X, the sets f −1 (x) and S2 \ f −1 (x) are nonempty and connected.
Then X is homeomorphic with S2 .
A good way to view X is as the result of collapsing each set f −1 (x) ⊂ S2 to a
single point. The theorem says that certain collections of sets can be collapsed to
points without changing the topology of S2 .
11.1. Examples and Applications

Example 11.2. Certain set collapses alter topology: Consider a single circle J
in S2 . Collapse J to a point to form a new space X. This space X is homeomorphic
to the union of two copies of S2 that are joined to one another at a point. There is,
of course, a natural continuous function f : S2 → X, but there is one point x ∈ X,
namely the image of J, such that J = f −1 (x) is connected, but S2 \ f −1 (x) is not.
See Figure 1.
S2
Figure 1. Pinching a simple closed curve in a 2-sphere
Example 11.3. Consider the middle-thirds Cantor C set in the real line R.
Multiply by the unit interval to form a Cantor set of intervals, C × [0, 1] ⊂ R2 ⊂
R2 ∪ {∞} = S2 in the 2-sphere S2 . Collapse each of the intervals {x} × [0, 1] with
x ∈ C to a point to form a decomposition space S2 / ∼ . Then this decomposition
space is homeomorphic to S2 . See Figure 2.
101
102 11. R. L. MOORE’S DECOMPOSITION THEOREM
C × [0, 1] f (C × [0, 1])
Figure 2. Collapsing a Cantor set of arcs
Example 11.4. Let D1 , D2 , D3 , . . . denote a sequence

of disjoint disks in the
2-sphere S2 such that limn→∞ diam(Dn ) = 0 and n Dn is dense in S2 . Collapse
each of the disks Dn to a point to form a new space X. Then X is homeomorphic
with S2 and the images of the disks Dn form a countable dense subset of X. It is
interesting to note that S2 \∪n int(Dn ) is a 1-dimensional set Y called the Sierpinski
curve. The Moore Decomposition Theorem shows that the set of points of Y that
do not lie in any ∂(Dn ) is homeomorphic to the complement of a countable dense
subset of S2 . See Figure 3.
Figure 3. Collapsing the disks defining a Sierpinski curve
Mathematicians studying complex dynamics have found R. L. Moore’s theorem

wonderfully suited for analyzing the structure of Julia and Fatou sets, the Man-
delbrot set, and other fractals. The theorem has also spawned a huge literature
exploring generalizations to higher dimensions.
11.2. Decomposition Spaces

2
The space X = f (S ) can be considered to be a decomposition space of the
2-sphere. We will take a moment to review the appropriate concepts.
11.2. DECOMPOSITION SPACES 103
Definition 11.5. Let S denote a topological space and G a collection of disjoint

subsets of S whose union is S. Then we call G a decomposition or partition of S.
There is a natural projection or identification map p : S → G which takes each
point x ∈ S to the set p(x) ∈ G that contains x. The projection map p can be used
to define a topology on the set G by declaring U ⊂ G to be open iff p−1 (U ) is open
in S. This new space is called a decomposition space or identification space of S.
The topology is called either the quotient topology or the identification topology.
The space is often denoted by the symbol S/G in order to indicate both the space
S from which it arose and the collection G of subsets into which S is partitioned.
Partitions of S give rise to, and arise from, equivalence relations ∼ on S: x ∼ y
iff x and y lie in the same element of G. Hence we also denote S/G by the symbol
S/ ∼. One natural way in which a decomposition arises is the following. Suppose
that f : S → Y is a function. Then the nonempty sets of the form f −1 (y), with
y ∈ Y , partition S and form a decomposition of S.
The identification topology satisfies a universal mapping property:
Definition 11.6. Suppose that G is a decomposition of S and that S/G has
the identification topology, with p : S → S/G the identification map. Suppose
further that f : S → Y is a continuous function such that, if p(x1 ) = p(x2 ), then
f (x1 ) = f (x2 ). Then there is a unique continuous function f˜ : S/G → Y such that
f˜ ◦ p = f .
Proof. If p(x) ∈ S/G, define f˜(p(x)) = f (x). Since p(x) = p(x ) =⇒
f (x) = f (x ), the function f˜ is well-defined, and f˜ ◦ p = f .
Suppose U ⊂ Y is open. Then p(x) ∈ f˜−1 (U ) iff f (x) ∈ U . Thus p−1 (f˜−1 (U ))
−1
= f (U ) is open since f is continuous. Therefore, by the definition of the identi-
fication topology, f˜−1 (U ) is open, and, consequently, f˜ is continuous.
Definition 11.7. Assume that S is a compact metric space. A collection G
of disjoint closed sets filling S is said to be upper semicontinuous if any of the
following equivalent conditions is satisfied:
(1) (Upper semicontinuity condition) If U is an open subset of S, then the
union of those elements of G lying in U forms an open (possibly empty) subset of
S.
(2) (Hausdorff condition) The decomposition space S/G is Hausdorff.
(3) (Continuity condition) The projection map p : S → S/G is a closed map. (If
the multi-valued function p−1 : S/G → S were a function, this continuity condition
would say that p−1 is continuous.)
Theorem 11.8. Suppose S is a compact metric space and that G is a decom-
position of S into closed subsets. Then the three conditions of the definition are
equivalent.
Proof. Assume condition (1), the upper semicontinuity condition. Let g1 and
g2 denote two distinct elements of the decomposition G. Since S is a metric space,
there are disjoint open subsets U1 and U2 in S containing g1 and g2 , respectively. Let
V1 denote the union of the elements of G lying in U1 . Let V2 denote the union of the
elements of G lying in U2 . By condition (1), the sets V1 and V2 are open. Since they
are unions of elements of G, V1 = p−1 ◦p(V1 ) and V2 = p−1 ◦p(V2 ). Consequently, by
definition of the identification topology, p(V1 ) and p(V2 ) are open — disjoint open
sets containing p(g1 ) and p(g2 ), respectively. Hence S/G is Hausdorff and condition
(2) is satisfied.
Assume condition (2), the Hausdorff condition. Suppose that F is a closed
subset of S. Then F is compact since S is compact. Hence p(F ) is compact
by continuity. Since S/G is Hausdorff, compact subsets of S/G are also closed.
Therefore p(F ) is closed and condition (3) is satisfied.
Assume condition (3), the continuity condition. Let g be an element of G and
U an open set containing g. Then the complement S \ U is closed, and the image
p(S \ U ) is therefore a closed set C that misses p(g). Thus V = (S/G) \ C is an open
subset of S/G. Hence p−1 (V ) is open and consists of the union of the elements of
G contained in U . Hence condition (1) is satisfied.
11.3. Proof of the Moore Decomposition Theorem

Proof of the Moore Decomposition Theorem. We use our earlier char-
acterization of the 2-sphere, Theorem 8.1. Recall:
A space X is a 2-sphere if it satisfies the following axioms:
(1) The space X is a nondegenerate Peano continuum. That is, X is compact,
connected, locally connected, metric having more than one point.
(2) The space X is not separated by any arc.
(3) The space X satisfies the Jordan Curve Theorem. That is, each simple
closed curve J in X separates X into two componenents U and V and is the
boundary of each.
Remark. The condition that J be the boundary of each complementary do-
main U follows from the other conditions. Indeed, suppose J is a simple closed
curve in X, p is a point of J, and A is an open arc in J containing p, with comple-
mentary arc B. Since the arc B does not separate X, there is an arc C in X \ B
joining a point of U irreducibly to A. The arc C must, except for its endpoint in A,
lie in U . Thus there are points of U arbitrarily close to A. Since A can be chosen
arbitrarily near to p, we conclude that p is in the boundary of U . End remark.
We verify the conditions (1), (2), and (3) in turn for the space X = S2 /G.
The space X has more than one point because, for each x ∈ X, both f −1 (x)
and S2 \ f −1 (x) are nonempty.
Our first goal is to show that X is a compact, connected metric space.
Remark. It is tempting to try to write down an explicit metric on X, but that
is a surprisingly difficult task. End remark.
We use the Urysohn Metrization Theorem: If X is a regular space with a
countable basis, then X is metrizable. (See [[24], Theorem 34.1, p. 215].)
Since S2 is compact and connected, X is also compact and connected. Since X
is a compact Hausdorff space, X is a regular space.
We must show that X has a countable basis. Let U = {U1 , U2 , . . .} denote a
countable basis for S2 . Let V = {V1 , V2 , . . .} denote the collection of sets formed
by taking finite unions of elements of U. For each set Vn , let Wn denote the union
of those elements of G that lie in Vn . Since X is Hausdorff, the decomposition G
satisfies the upper semicontinuity condition. That is, the sets Wn are open and
have open image p(Wn ) in X. We claim that the images {p(W1 ), p(W2 ), . . .} form
a countable basis for X. Indeed, let x = p(g) denote an element of X, with g ∈ G,
11.3. PROOF OF THE MOORE DECOMPOSITION THEOREM 105
and let U denote an open neighborhood of x in X. Then p−1 (U ) is an open

neighborhood of g in S2 . There is a finite collection of elements of U in p−1 (U ) that
suffice to cover the compact set g. The union of these elements is one of the sets
Vn . The image p(Wn ) is therefore a neighborhood of x in U .
The 2-sphere S2 is locally connected, hence a Peano continuum, hence a con-
tinuous image of [0, 1]. Hence X is also a continuous image of [0, 1], hence also
a Peano continuum. (As we noted earlier, our proof that a continuous image of
[0, 1] is locally connected actually used only, as properties of [0, 1], that [0, 1] is
locally connected, compact, metric. Hence, we could, if we wanted, make no direct
mention of [0, 1] in this argument.)
This completes the proof that condition (1) of the 2-sphere characterization is
satisfied.
Exercise 11.9. If A is a connected, closed subset of X, then p−1 (A) is con-
nected.
The proofs that no arc separates X and that each simple closed curve separates
X into two components are modelled exactly on the proofs that the same properties
are satisfied for a 2-sphere. Review that argument from section 7.3. We use the
same elementary algebraic topology. In order to simplify notation, for each subset
A ⊂ S, let A denote the inverse image p−1 (A) of A in S2 .
Suppose that, contrary to condition (2), there is an arc A ⊂ X that separates
X, and let x and y denote two points separated by A. Let p ∈ A divide A into
two subarcs A1 and A2 . We claim that either A1 or A2 separates x from y. We
see that by examining a segment of the reduced Mayer-Vietoris sequence for the
pair U = S2 \ A1 and V = S2 \ A2 . We use the fact that the first homology of the
complement of a connected set in S2 is trivial.
(0 = H1 (S2 \ {p }) → H0 (S2 \ A )−→H0 (S2 \ A1 ) ⊕ H0 (S2 \ A2 ).
α
Thus the map α is one-to-one. Since A separates x from y, A separates x from

y . If x ⊂ x and y ⊂ y , then x − y represents a nontrivial element in the
center group H0 (S2 \ A), hence also a nontrivial element either in H0 (S2 \ A1 ) or in
H0 (S2 \A2 ). It follows that either A1 or A2 separates x from y in S2 . Consequently,
since each inverse image of a point is connected, either A1 or A2 separates x from
y in X.
By induction, one obtains
intervals (I0 = A) ⊃ I1 ⊃ I2 ⊃ · · · which separate x
from y in X such that ∞
n=0 In is a single point q. But S \ q is connected, so that
2
q does not separate x from y. A path from x to y in X \ {q} also misses all but
finitely many of the sets In , a contradiction. We conclude that A does not separate
X.
Suppose that J is a simple closed curve in X. Let p1 , p2 ∈ J cut J into two
arcs A1 and A2 . Then A1 and A2 are compact, connected, and have nonconnected
intersection p1 ∪ p2 . The reduced Mayer-Vietoris homology sequence for the pair
U = S2 \ A1 and V = S2 \ A2 contains the segment
H1 (S2 \ A1 ) ⊕ H1 (S2 \ A2 ) → H1 (S2 \ (p1 ∪ p2 )−→H0 (S2 \ J ).
α
The groups H1 (S2 \ A1 ) and H1 (S2 \ A2 ) are = 0 since A1 and A2 are connected;
hence the homomorphism α is one-to one. The group H1 (U ∪ V ) is not 0 since
A1 ∩ A2 is not connected; hence the group H0 (S2 \ J ) is not 0. That is, J separates
S2 . Since each element of G not in J is connected, hence must lie in a single

component of S2 \ J , J also separates X.
It remains only to prove that X \ J has only two components. If X \ J has
at least three components U, V, W , we shall obtain a contradiction by showing the
existence of a simple closed curve K that does not separate X. See Figure 4. As
argued before, J is the common boundary of each of its complementary domains.
Hence it is possible to construct arcs AU , AV , AW spanning J through U , V , and W ,
respectively, having endpoints situated on J approximately as indicated in Figure 4.
AW
d
AU
AV
Figure 4. Showing that there are only two complementary domains
It is easy to see that the simple closed curve K = AU ∪ d ∪ AV ∪ c does not

separate X. For every point can be joined to one of the two complementary arcs in
J without hitting K since the possible obstacles are always contained in a subarc
of K, and the two complementary arcs in J are connected by the arc AW . This
contradiction shows that X \ J has only two components.
11.4. EXERCISES 107
11.4. Exercises
11.2. Figure 5 indicates three decompositions of the 2-sphere. In the first, the
equator is to be identified to a single point. In the second, the black disk is to be
identified to a single point. In the third, the components of a product of a Cantor
set with an interval are to identified to points. Identify the spaces that result.
Figure 5. Three decompositions

CHAPTER 12
The Open Mapping Theorem
Theorem 12.1 (The Open Mapping Theorem). Suppose that f : Rn → Rn is

a continuous function from Euclidean n-space into itself that is one-to-one. Then
the image of Rn is an open subset of Rn .
Corollary 12.2. If n > m, then Rn and Rm are not homeomorphic.
Proof of corollary. Otherwise we would have one-to-one mappings
Rn → Rm → Rn ,
with the first of these maps a homeomorphism and the latter the natural inclusion
map Rm = Rm × {0} ⊂ Rn whose image is not open. The composite would
contradict the Open Mapping Theorem.
12.1. Tools
Our proof of the Open Mapping Theorem depends on the n-dimensional ver-
sions of four theorems that we have already discussed. We have proved all of these
except for the general version of the No Retraction Theorem, and we outlined the
necessary additions that must be carried out to complete the general version.
Theorem 12.3 (Jordan Separation Theorem). If f : Sn → Sn+1 is continuous
and one-to-one, then the image f (Sn ) separates Sn+1 into two components U and
V and is the boundary of each. See Theorem 7.6.
Theorem 12.4 (No Retraction Theorem). There is no continuous function
f : Bn → (Sn−1 = ∂Bn ) such that f (x) = x for each x ∈ Sn−1 . (See Theorem
2.17.)
Lemma 12.5 (Urysohn’s Lemma, easy metric case). If X is a metric space
and A and B are disjoint closed subsets of X, then there is a continuous function
f : X → [0, 1] such that f (A) = {0} and f (B) = {1}. (See Theorem 3.10.)
Lemma 12.6 (Tietze’s Extension Theorem, easier metric case). If X is a metric
space, A is a closed subspace, and f : A → [0, 1] is a continuous function, then there
is a continuous function F : X → [0, 1] such that, for each x ∈ A, F (x) = f (x).
See Theorem 3.14.
Because we have completed the proof of the No Retraction Theorem only in
dimension 2, our proof will be complete only in dimension 2. (However, we recall,
that the reader with some facility working with higher-dimensional simplicial com-
plexes should have no difficulty in generalizing the 2-dimensional proof of the No
Retraction Theorem to higher dimensions. One uses a triangulation of the ball by
very small n-simplexes, uses the forbidden retraction to map this triangulation near
to the boundary of the ball, intersects the ray with the image simplexes, and finds
a one-ended arc in the preimage.)
109
110 12. THE OPEN MAPPING THEOREM
12.2. Two Lemmas

We first use these tools to prove two lemmas, an easy nullhomotopy lemma,
and a harder lemma called the Borsuk Lemma.
Lemma 12.7 (Nullhomotopy Lemma). Suppose that x and y are points of Sn
and that A is a compact subset of Sn \ {x, y}. If x and y lie in the same component
of Sn \ A, then A can be contracted to a point in Sn \ {x, y}. See Figure 1.
Proof. We may view Sn as a polyhedron, say as the boundary of the n + 1
simplex. If x and y are in the same component of Sn \ A, then there is a polygonal
path P from x to y in the complement of A. A standard ball around x can be
stretched along the path P by a homeomorphism of Sn that fixes A and x so that
the image of the ball contains both x and y. The inverse of this homeomorphism
shows that we can assume that both x and y lie in the interior of a standard ball in
the complement of A. Since the complement of a standard ball in Sn is contractible,
its contraction will also contract A.
y
x
Figure 1. The Nullhomotopy Lemma
Lemma 12.8 (Borsuk’s Lemma). Suppose that x and y are points of Sn and
that A is a compact subset of Sn \ {x, y}. If A can be contracted to a point in
Sn \ {x, y}, then x and y lie in the same component of Sn \ A.
Proof. See Figure 2.

We may view Sn as the one-point compactification of Rn , with x = 0 and
y = ∞. We must show that 0 is in the unbounded component of Rn \ A. We
proceed indirectly by assuming to the contrary that 0 is in a bounded component
C of Rn \ A.
We claim that, because A can be contracted to a point in Rn \ {0}, there is a
map F : A ∪ C → Rn \ {0} that is the identity on A. Assuming this for the moment,
12.2. TWO LEMMAS 111
(C ∪ A) × 1
U A × [0, 1] U
A × [0, 1]
A × [0, 1]
(C ∪ A) × 0
Figure 2. The homotopy extension construction
we complete the proof as follows. Define a continuous function F : Rn → Rn \ {0}

by defining F |A ∪ C = F and F (z) = z for all other points of Rn . Let B denote a
very large round ball centered at x = 0 and containing the image of F in its interior.
Let r : B \ {0} → ∂B denote the radial retraction of B \ {0} onto ∂B. Then the
composite, r ◦ (F |B) is a retraction of B onto its boundary and contradicts the No
Retraction Theorem.
To construct F , we have to assume that A is contractible in Rn \ {0} and
we have to apply both the Tietze Extension Theorem and Urysohn’s Lemma: Let
J : A × [0, 1] → Rn \ {0} be a contraction of A, so that J(a, 0) = a for each
a ∈ A and, for some constant z and every a ∈ A, J(a, 1) = z. We may extend this
map to C × {1} by sending each (c, 1) to z as well. This partial map defined on
(A × [0, 1]) ∪ (C × {1}) may be extended by the Tietze Extension Theorem to a
map J : (A ∪ C) × [0, 1] → Rn . An open neighborhood U of (A × [0, 1]) ∪ (C × {1})
is mapped thereby into Rn \ {0}. The neighborhood U contains a set of the form
W ×[0, 1], where W is a neighborhood of A in A∪C. By Urysohn’s Lemma, there is
a continuous function j : A ∪ C → [0, 1] such that h(C) = 0 and h((A ∪ C) \ W ) = 1.
Define K : A ∪ C → (A ∪ C) × [0, 1] by the formula
K(α) = (α, h(α)), ∀ α ∈ (A ∪ C).
The image of A ∪ C is indicated in the diagram by a bold line. Then we may set
F (α) = J ◦ K. This shows the existence of F and completes the proof of the
Borsuk Lemma.
112 12. THE OPEN MAPPING THEOREM
12.3. Proof of the Open Mapping Theorem

Proof of the Open Mapping Theorem. We are given a continuous map
f : Rn → Rn that is one-to-one. We must show that the image of f is an open set.
To that end, let x ∈ Rn , and let B denote a small ball around x. We will show that
f (B) is a neighborhood of f (x) in Rn .
The image f (∂B) separates Rn into two components U and V and is the bound-
ary of each, by the Jordan Separation Theorem. We may assume that U is the
bounded component of Rn \ f (∂B). We claim that U = f (int(B)). At any rate,
f (int(B)) is connected and lies in the complement of f (∂B), hence must lie either
in U or in V . The image f (B) is contractible, hence cannot separate Rn by the
Borsuk Lemma. Hence f (B) must contain every point of U , for otherwise f (B)
would separate such a point from ∞, a contradiction. Thus f (x) ∈ U ⊂ f (B), so
that the image of f is open.
12.4. Exercise
12.1. The Open Mapping Theorem is often proved in the differentiable case.
Consult any standard graduate text in analysis for the very interesting proof. It
involves the contraction mapping principle which we have already considered and
proved. See, for example, Serge Lang [41].
CHAPTER 13
Triangulation of 2-Manifolds
13.1. Statement of the Triangulation Theorem

We recall:
Definition 13.1. An n-dimensional manifold X is a space such that each point
x ∈ X has a neighborhood N such that N is homeomorphic with n-dimensional
Euclidean space Rn . In this section, we tacitly assume that our n-manifolds are
separable metric spaces.
Example 13.2. The only connected 1-manifolds are the line R1 itself and the
1-sphere S1 . In dimension 2 the standard examples are R2 and the 2-sphere S2 . In
a previous Chapter 9 we claimed as fact that all compact connected 2-manifolds
are spheres with handles and Möbius bands inserted. Triangulation, the subject of
this chapter, is the major first step in proving that fact.
The standard procedure in topology texts is to assume that 2-manifolds are
triangulable. Our aim is to prove that fact. The original proof of this fact is
usually attributed to Radó [91].
Definition 13.3. A triangulation T of a 2-manifold X consists of a family
→ X | α ∈ A} of embeddings of a standard triangle Δ in X such that
{fα : Δ
(1) α fα (Δ) = X
(2) If two images fα (Δ) and fβ (Δ) intersect, then they do so in the image of a
vertex or the image of an edge of each.
13.2. Tools
We will assume one difficult theorem in the proof, namely, the Schoenflies
Theorem 8.2 on page 81, which we proved in an earlier section. We also assume
the key lemma used in the proof of the Schoenflies Theorem, which we called the
Arc-Crossing Lemma 8.6 on page 82.
Theorem 13.4 (The Schoenflies Theorem). If J is a simple closed curve in
R2 , then J separates R2 into two disjoint open sets, namely, the interior int(J) of
J which has compact closure in R2 and the exterior ext(J) which has noncompact
closure in R2 . The closure of int(J) is homeomorphic to the triangular disk Δ.
(We actually only proved the corresponding result for J ⊂ S2 , of which this
theorem is an easy corollary: the one-point compactification of R2 is S2 so that the
earlier theorem applies.)
Lemma 13.5 (Arc-crossing Lemma). Suppose that J is a simple closed curve
in R2 and that A and B are arcs spanning J across int(J). If the endpoints a1 , a2
of A separate the endpoints b1 , b2 of B on J, then the arcs A and B must intersect.
See Figure 1.
113
114 13. TRIANGULATION OF 2-MANIFOLDS
Figure 1. The Arc-crossing Lemma
Lemma 13.6 (Curve-intersection Lemma). Suppose that J is a simple closed

curve, K ⊂ int(J) is a compact set, and Γ is a finite graph, all in R2 . Then, if
> 0, there is a simple closed curve J in the -neighborhood of J which contains
K in its interior such that J ∩ Γ has only finitely many components.
Proof. There is a positive number δ < such that, if p and q are two points
within δ of one another in Γ, then there is an arc in Γ joining p and q of diameter
less than . Divide J into finitely many subarcs Ji , each of diameter < δ. For each
Ji , if Ji ∩ Γ is not connected, then there is both a first point pi and a last point qi
of the intersection. Replace that portion of Ji between these two points by an arc
of diameter < in Γ. The result is a closed curve J , probably not simple, whose
intersection with Γ has only finitely many components. See Figure 2.
It remains to be shown that, if is sufficiently small, then the closed curve J
contains a simple closed curve that satisfies the conclusions of the lemma.
To that end, we prepare for the arc replacement of the first paragraph in the
following way. If J ⊂ Γ, then there is nothing to prove. Otherwise, we may choose
an arc A ⊂ (J \ Γ). Let B denote the complementary arc in J. Pick a third arc C
with the same endpoints that is very close to B so that A ∪ C is a simple closed
curve containing int(B) in its interior. Expand the compact set K to be a disk in
the interior of J and connect K to int(A) by an arc; call the result K . (See Figure
2.)
Now proceed as in the first paragraph of the proof. If δ and are very small,
then an entire neighborhood of A in J will remain unchanged because it lies in the
complement of Γ. Likewise, int(B) will be moved such a small distance that its
replacement remains in int(A ∪ C) \ K . The replacement of B will contain an arc
B that spans the simple closed curve A ∪ C through int(A ∪ C), misses K , and
has only finitely many components of intersection with Γ. We claim that K lies
13.3. PROOF OF THE TRIANGULATION THEOREM 115
K B
K
A
Figure 2. The setting for the replacement of J by J
in the interior of the simple closed curve J = A ∪ B . For otherwise, K could be

joined by an arc L through int(A ∪ C) \ B to C, K ∪ L and B would contradict
the Arc-crossing Lemma 13.5.
13.3. Proof of the Triangulation Theorem

We now reiterate the Triangulation Theorem and give its proof.
Theorem 13.7 (Triangulation Theorem). Every (separable metric) 2-manifold
X can be triangulated.
Proof. We assume for convenience that X is connected. Our intent is to
define a locally finite graph Γ∞ dividing X into disks. Each of these disks will have
boundary in Γ consisting of finitely many segments. We add enough vertices to the
segments of Γ that every complementary disk has at least three edges. We insert
a vertex v(D) in the middle of each complementary disk D, and join v(D) by arcs
to the vertices on the boundary of D. The result is the desired triangulation. See
Figure 3.
There is a countable family of (closed) disks {Di | i = 1, 2, . . .} in X with interi-
ors covering X, with each Di lying in a coordinate neighborhood R2i homeomorphic
to R2 . We may assume that the collection {R2i | i = 1, 2, . . .} is locally finite. We
pick simple closed curves J1 , J2 , . . ., with Ji ⊂ R2i such that Di lies in the interior
of Ji in R2i . We may assume the disks ordered so that each Di with i > 1 intersects
the union D1 ∪ D2 ∪ · · · ∪ Di−1 .
At this point, we want to apply the result of the following exercise.
Exercise 13.8. Suppose that a set K is connected. We say that a point p ∈ K
is a cut point if K \ {p} is not connected.
Suppose that D is a disk and K a connected finite graph in D without cut
points, with ∂D ⊂ K. Then K separates D into subdisks E, ∂E ⊂ K, K ∩int(E) =
∅.
116 13. TRIANGULATION OF 2-MANIFOLDS
v(D)
Figure 3. The cellular subdivision
We adjust the simple closed curves Ji inductively by the Curve-intersection

Lemma so that, for each i, Ji has only finitely many components of intersection
with Γ(i − 1) = J1 ∪ J2 ∪ · · · Ji−1 . It is also easy to modify Ji slightly so that either
Ji ∩ Γ(i − 1) is either empty or contains more than one point. If it is empty, that

i−1
means that Di is already covered by j=1 int(Jj ), and we can discard Ji and define
Γ(i) = Γ(i−1). If Ji ∩Γ(i−1) is nonempty, hence has more than one point, then we

i
add Ji to Γ(i − 1) to form Γ(i). The union j=1 int(Ji ) is cellulated by the graph
Γ(i) and each of its cells is indeed a disk with boundary a simple closed curve in
Γ(i).
∞

∞ we define Γ∞ = i=1 Γ(i). Note that every component of X \ Γ∞ that
Finally,
lies in i=1 int(Ji ) is the interior of a disk with boundary a simple closed curve in
Γ∞ , and since each Di is covered by some int(Ji ), every point of X is in one of
these disks. The graph Γ∞ is locally finite because the coordinate neighborhoods
R2i were locally finite.
13.4. Exercises
[Hint: By our results on disk isotopy, every arc in a 2-manifold looks like every
other. If U is a component of D \ K, trace the boundary of U . It either closes up
as a simple closed curve, or it comes back on itself to form a cut point.]
13.2. Suppose that S is a compact, connected 2-manifold with nonempty
boundary. Show that S can be retracted onto a finite graph.
CHAPTER 14
Structure and Classification of 2-Manifolds
The Structure Theorem for compact, connected, metric 2-manifolds says that
each is composed from copies of four standard pieces by an operation called con-
nected sum.
The Classification Theorem says that, up to homeomorphism, such 2-manifolds
are completely determined by two invariants, namely orientability and Euler char-
acteristic.
14.1. Statement of the Structure Theorem

The Structure Theorem for compact metric 2-manifolds depends on the knowl-
edge of the four simplest examples and a construction that creates more complex
examples by a form of addition. The four examples are the 2-sphere S2 which we
have already met, the torus T2 which is the surface of a doughnut, the projective
plane P2 which is formed by sewing a Möbius band to a disk along their common
circle boundaries, and the Klein bottle K2 which is formed by sewing together two
Möbius bands. The construction, or addition, is called connected sum.
D1 D2
Figure 1. Connected sum
Definition 14.1. If M1 and M2 are 2-manifolds, then the connected sum

M1 #M2 is formed by removing the interiors of disks D1 ⊂ M1 and D2 ⊂ M2 and
sewing the boundaries ∂D1 and ∂D2 together by a homeomorphism. See Figure 1.
117
118 14. STRUCTURE AND CLASSIFICATION OF 2-MANIFOLDS
Theorem 14.2 (Structure Theorem). Every compact, connected 2-manifold is

a connected sum of finitely many copies of S2 , T2 , P2 , and K2 .
There is a very simple and straightforward way of proving the Structure Theo-
rem inductively, which seems now to have been discovered and rediscovered in some
form or another independently by at least five different mathematicians over the
last twenty or thirty years. We give here our own version of this proof.
Every proof of the Structure Theorem begins with the following fact.
Lemma 14.3 (Edge-pairing Lemma). Every surface can be realized up to home-
omorphism as an edge-pairing of a disk. (Definitions and proof of this lemma appear
in the next section.)
14.2. Edge-pairings
Most students of topology are familiar with the description of a 2-manifold M
as a quotient of a disk D under an edge pairing. See, for example, Figure 2.
b
a a
a
b
v b
Figure 2. The standard presentation of the torus as an identifi-
cation space
Edges with the same label are to be identified by a homeomorphism which

matches the arrows on those labelled edges. The pairing, and the manifold M are
determined up to homeomorphism by the boundary word W of D, which in Figure
2 is given by
W = aba−1 b−1 .
We write π : D → D/W = M (W ) for the identification map. If one traverses the
boundary in the clockwise direction, beginning at vertex v, then one encounters first
clockwise-edge a, then clockwise-edge b, then counterclockwise-edge a = clockwise-
edge a−1 , and finally counterclockwise-edge b = clockwise-edge b−1 .
In general, the cyclic boundary word W can be any finite word whatsoever that
is quadratic, in the sense that, if either edge label x or x−1 appears in W , then
exactly one additional label x or x−1 also appears. For example,
W = abc−1 c−1 b−1 dad
is a quadratic word which corresponds to the disk of Figure 3.
We even allow the degenerate case where W is the empty word and the iden-
tification map π : D → D/W identifies the entire boundary of D to a point; the
space D/W is then the 2-sphere S2 .
14.3. PROOF OF THE STRUCTURE THEOREM 119
a
b
v
d
c
a
c
d b
Figure 3. The disk associated with the word W = abc−1 c−1 b−1 dad
Proof of the Edge-pairing Lemma 14.3. Let T denote a triangulation of

the surface S. Cut S apart along the edges of the triangles of T to form a finite,
disjoint union of triangles T1 , T2 , . . ., Tk . If Ti originally shared an edge with Tj ,
then we say that the resulting edges correspond. Since S is connected, the triangles
may be ordered in such a fashion that each Ti has at least one edge that corresponds
to some edge of T1 ∪. . .∪Ti−1 . Inductively, partially sew the triangles back together
in the following fashion. Let D1 = T1 . Sew T2 to D1 by identifying exactly one edge
of T2 with a corresponding edge of T1 in D1 to form D2 . In general, sew Ti to Di−1
by identifying exactly one edge of Ti with a corresponding edge of T1 ∪ . . . ∪ Ti−1
in Di−1 to form Di . According to the following exercise, each of the sets D1 , D2 ,
. . ., Dk is a triangulated disk.
Exercise 14.4. Suppose that two disks D1 and D2 intersect in an arc A which
lies in the boundary of each. Then D1 ∪ D2 is also a disk.
The edges of Dk appear in pairs that need to be identified to reconstitute S.
This edge-pairing of Dk is the edge-pairing required by the Edge-pairing Lemma.

14.3. Proof of the Structure Theorem

Proof. We now begin with an arbitrary surface S. By the Edge-pairing
Lemma 14.3, there is a quadratic word and an edge-pairing π : D → (S = D/W ).
We need to show that S is a finite connected sum of copies of S2 , P2 , T2 , and K2 .
We prove this by induction on the length |W | of the word W .
If W is the empty word so that |W | = 0, then D/W = S2 , as noted before.
If |W | = 2, then W has the form W = aa−1 with D/W = S2 or W = aa with
D/W = P2 .
Assume inductively that |W | > 2 and that the result is true for any quadratic
word of length less than |W |. Let α and β denote two edges of the boundary of D
that are to be identified, so that α and β have the same label a in the word W up
to sign: label(α), label(β) ∈ {a, a−1 }. Attach small disks D1 and D2 to D along
the arcs α and β, with ∂D1 = α ∪ α and ∂D2 = β ∪ β . Collapse the arcs α and
β to points. The quotient D will be a disk with the images of D1 and D2 being
disks in D , and the boundary of D will be labelled by the word W obtained by
deleting the labels of α and β from W . The two possible situations are pictured in
Figures 4 and 5.
D1
α D2
α
ββ α β
D D
Figure 4. Induction collapse
D1
β D2
α β
α
β
α
D D
Figure 5. Induction collapse2
By inductive hypothesis, S = D /W is a 2-manifold that is a connected sum

of copies of our four model 2-manifolds S2 , P2 , T2 , and K2 ; and in that 2-manifold
will be two disks D1 and D2 that are the images of D1 and D2 . The union of the
two disks D1 and D2 will lie in a disk E ⊂ D /W and the disk E will take one of
the four forms pictured in Figure 6.
E E E E
D1 D2 D1 D2 D1 D2 D1 D2
a a a a a a a a
Case 1 Case 2 Case 3 Case 4
Figure 6. Four Cases
In cases 1 and 2, there is nothing in the word W , before the collapse of α with
β to form D , that requires the pinching of D1 with D2 at the vertex in the center
of the picture of E. Hence the diagrams of cases 1 and 2 split at that central vertex
and become pictures of S2 minus a disk and P2 minus a disk, shown in Figures 7
and 8. We conclude that S = D/W is the connected sum of S = D /W with S2
in case 1 and of S = D /W with P2 in case 2:
14.3. PROOF OF THE STRUCTURE THEOREM 121
S 2 \ int(D)
a a
Figure 7. Case 1 after vertex splitting
P2 \ int(D)
a a
Figure 8. Case 2 after vertex splitting
In case 3, one obtains S = D/W from S = D /W by identifying the edges

labelled a in Figure 9. But that implies that E, with identifications, represents T2
with a disk removed. Thus S = D/W is the connected sum of S = D /W with
T2 . Similarly, in case 4, we see that S = D/W is the connected sum of S = D /W
with K2 (Figure 10):
T 2 \ int(D)
a a
Figure 9. Case 3: Torus minus a disk
K 2 \ int(D)
a a
Figure 10. Case 4: Klein bottle minus a disk

That is, the reintroduction of the edges α and β leads to a connected sum of
D /W with one of the four model 2-manifolds. This completes the inductive proof
of the Structure Theorem.
Alternative description. Every surface can be built from the following build-
ing blocks (Figure 11): the 2-dimensional sphere S2 , the annulus A2 , and the Möbius
strip M2 . The 2-sphere is the surface of a ball. The annulus is an open cylinder
with two circles as boundary. The Möbius strip is formed by sewing two ends of a
strip together with a half twist; it has one boundary component.
S2
A2
M2
Figure 11. The building blocks
Figure 12. The sphere with handles and Möbius bands
The construction can be interpreted as follows (Figure 12): Cut 2n + k circular

holes in the 2-sphere S2 . For each of the n pairs, sew in a copy of A2 , one boundary
component to the boundary of one of the holes, the other boundary component to
the other hole boundary. In each of the remaining k holes, sew in a copy of M2 ,
14.4. STATEMENT AND PROOF OF THE CLASSIFICATION THEOREM 123
boundary to boundary. In the figure, the annuli appear as handles on the sphere
and the Möbius bands are indicated by crosses.
Proof of equivalence. Connected sums with a 2-sphere do not change the

surface. Hence we lose no generality in beginning with a 2-sphere, and only one
copy of the 2-sphere is needed. The Möbius band is the projective plane with a
disk removed. Hence each inserted Möbius band is simply a connected sum with a
projective plane. The Klein bottle is the connected sum of two projective planes, so
no Klein bottles are required. Finally, the sewing on of an annulus results either in
the connected sum with a torus or the connected sum with a Klein bottle, depending
on the orientations with which boundary components are inserted. Since no Klein
bottles are required, the annuli may be attached in the fashion indicated by the
figure.
Remark. Over the last 25 years, a number of mathematicians have discovered

very nice inductive proofs of the Structure Theorem which are in some sense much
more straightforward. All of these, despite their differences in description, seem
really to be variants of the same proof. They appear, for example, in [92, (1988)],
[93, (1979)], [94, (1984)], and [95, (1999)]. They seem to deserve to be called the
“natural” proof of the theorem. Our own description is simply a variant of this
natural proof in a form suggested by our studies of face pairings of 3-manifolds.
End remark.
Remark. The standard proof of the Structure Theorem for surfaces appears in
[30], [25], and [24]; this standard proof has the advantage of putting each surface
in a standard “algebraic” form. End remark.
14.4. Statement and Proof of the Classification Theorem

The classification of compact, metric 2-manifolds, up to homeomorphism, de-
pends on two concepts, orientability and Euler characteristic. The goal of this
section is to suggest exercises that can be used to prove the following classification
theorem.
Theorem 14.5 (Classification of 2-manifolds, Classification Theorem). Two
compact, metric 2-manifolds (without boundary) are homeomorphic if and only if
they have the same Euler characteristic, and both are orientable or both are nonori-
entable.
We now define the two important concepts: orientability and Euler character-
istic.
Definition 14.6. A 2-manifold is nonorientable if and only if it is impossible
to determine a well-defined notion of right and left in the manifold, if and only if it
contains a Möbius band. See Figure 13 to see how the existence of a Möbius band
in a surface disrupts the notion of left and right.
(As one walks around a Möbius band and returns to the beginning, the notion
of right and left will have switched.)
Definition 14.7. Euler characteristic is defined for spaces called finite cell
complexes. The prototypical example is the triangulated polyhedron, where the
vertices are 0-cells, the edges are 1-cells, the triangles are 2-cells, and in general the
Left tfeL
Left Left
Left
Figure 13. Reverse of Orientation Along a Möbius Band
n-simplexes are n-cells. The Euler characteristic is the number of even dimensional
cells minus the number of odd dimensional cells.
Some beautiful examples of 2-dimensional cell complexes are the Platonic sur-
faces (the surfaces of the solid tetrahedron, cube, octahedron, dodecahedron, icosa-
hedron) which are 2-spheres tiled by 2-dimensional polygonal disks, with associated
1-dimensional edges and 0-dimensional vertices. See Figure 5 on page 143. In Chap-
ter 16, Euler characteristic will be discussed further.
Remark. By means of algebraic topology, one proves that the Euler charac-
teristic of a cell complex depends only on the topology of the space in question and
not on the particular division into cells. In particular, any triangulation may be
used to calculate the Euler characteristic of a 2-manifold. End remark.
Remark (Motivation for the definition of Euler characteristic). The simplest
equivalence relation on cell-complexes is simple subdivision: cutting one n-cell into
two n-cells by inserting an (n − 1)-cell between them. See Figure 14.
dim n − 1
dim n
dim n
simple subdivision dim n
Figure 14. Simple subdivision
Euler characteristic is the simplest cell-counting formula that remains un-

changed under this operation: the difference between the number of even and the
number of odd dimensional cells is not changed by simple subdivision: one adds
one n-dimensional cell and one n − 1-dimensional cell. End remark.
14.5. EXERCISES 125
14.5. Exercises
14.2. Show that each of the five Platonic surfaces can be obtained by repeated
simple subdivisions from the following cell structure on the 2-sphere which has two
0-cells on the equator, two 1-cells on the equator joining the two 0-cells, and two
2-cells, namely, a northern hemisphere and a southern hemisphere, with resulting
Euler characteristic of 4 − 1 = 2. In particular, all of the Platonic solids have Euler
characteristic equal to 2. This is of course a consequence of the fact that each
is a cellular subdivision of the 2-dimensional sphere and the fact that the Euler
characteristic of a space does not depend on any particular cellulation.
For example, see Figures 15 and 16.
Figure 15. Cell structure on the sphere
7-5=2
8-6=2
4-2=2 6-4=2
5-3=2
The tetrahedron by simple subdivision.

Figure 16. The tetrahedron by simple subdivision
14.3. Show that the model manifold K2 is the connected sum of two copies of
P .
2
14.4. Show that if M is a surface that has P2 as a connected summand, then

the manifolds M #T2 and M #K2 are homeomorphic.
[Hint: Examine cases 3 and 4 in the proof of the Structure Theorem of the last
chapter. Slide D2 once around the Möbius band represented by P2 minus a disk to
change case 4 to case 3.]
14.5. Show that the Euler characteristics of the four model surfaces are these:
χ(S2 ) = 2, χ(P2 ) = 1, χ(T2 ) = 0, χ(K2 ) = 0.
14.6. Show that the Euler characteristic χ(M #N ) of the connected sum M #N
is χ(M ) + χ(N ) − 2.
[Hint: Triangulate the surfaces M and N so that the disks removed in forming
the connected sum are triangles in the triangulations.]
14.7. Show that every surface M is a connected sum of a sphere with k ≥ 0
copies of T2 and = 0, 1, or 2 copies of P2 . Show that χ(M ) = 2 − 2k − .
14.8. Show that two surfaces are homeomorphic iff they have the same Euler
characteristic and both are orientable or both are nonorientable.
The following nontrivial exercises can be used to show that connected sum is
well-defined in dimension 2.
14.9. Suppose that D1 and D2 are disks in the interior of a disk D. Then show
that there is an isotopy of D, fixing the boundary of D, which takes D1 at stage 0
of the isotopy to D2 at stage 1 of the isotopy.
14.10. Suppose that D1 and D2 are two disks in a connected 2-manifold M .
Then show that there is an isotopy of M that takes D1 at stage 0 to D2 at stage 2
of the isotopy.
14.11. Show that if D is a disk in a 2-manifold M , then there is a homeomor-
phism h : M → M such that h(D) = D and such that h|∂D reverses the orientation
of ∂D.
14.12. Show that if h : S1 → S1 is orientation preserving, then h is isotopic to
the identity map from S1 to itself.
[Hint: Without loss of generality, there is a point x such that h(x) = x. Cut
S1 at x to form an interval (a ball of dimension 1). Apply the Alexander trick.]
14.13. Show that the connected sum M1 #M2 of two compact, connected 2-
manifolds M1 and M2 is well-defined.
A standard way to construct surfaces is to identify the edges of a polygonal
disk in pairs.
14.14. Given the edge pairings from Figures 17 and 18, identify the surfaces as
connected sums of spheres, tori, projective planes and Klein bottles.
14.5. EXERCISES 127
−1
e
c−1
d−1
a−1
a
b−1
c
Figure 17. An edge identification

g c
d
g −1
b
e
f −1
a−1
a
j
i
i
e
h−1
c−1
b
h
d f
j
Figure 18. Another edge identification

CHAPTER 15
The Torus
The torus T2 is the surface of the doughnut. It is the next simplest surface
after the 2-sphere S2 . Our goal is to understand the simple closed curves on the
torus. A good reference is Rolfsen [96].
As we shall see, the torus T2 is closely related to the Euclidean plane. We begin
our preparations for the study of curves on the torus by examining some delightful
properties of lines and arcs in the plane.
15.1. Lines and Arcs in the Plane

Exercise 15.1. Suppose that L ⊂ R2 is homeomorphic to the line R and that
both ends of L approach ∞. Show that L separates R2 into two components U and
V and is the boundary of each.
Problem 15.2. Suppose that A = A(0) is an arc joining (0, 0) and (1, 0) in the
plane. Let A(1/2) denote the arc formed by translating A(0) 1/2 unit to the right.
Prove that A(0) and A(1/2) must intersect (Figure 1). See [96].
A = A(0)
A(1)
q
A(1/2)
Figure 1. A(0) and A(1/2) must intersect
129
130 15. THE TORUS
Solution. Consider A = A(0) and its two translates A(1/2) and A(1). As-
sume that, contrary to the statement, A(0) does not intersect A(1/2). Then A(1/2)
intersects neither A(0) nor A(1).
Let p ∈ A(1/2) denote a highest point of A(1/2). Let q denote a lowest point
of A(1/2). Form a homeomorphic copy L of the line R by adding to the arc
pq ⊂ A(1/2) a rising vertical ray from p to ∞ and a falling vertical ray from q to
∞. By Exercise 15.1, L separates R2 , and in fact must separate A(0) from A(1).
But A(0) and A(1) intersect at (1, 0), a contradiction.
Exercise 15.3. Let A(0) and A(1) be as above. Let d be any distance between
0 and 1, and let A(d) denote the image of A(0) translated a distance d to the right.
Then show that A(d) must intersect either A(0) or A(1). See Rolfsen [96].
Definition 15.4. If A(0) and A(d) intersect at a point (a, b), then the points
(a, b) and (a − d, b) are points of A(0) at the same height b and at distance d from
each other. We call the subarc of A(0) from (a − d, b) to (a, b) a d-chord in the arc
A(0).
A simple scaling argument then shows that the Exercise 15.3 implies the fol-
lowing corollary.
Corollary 15.5. If A is an arc in the plane that is a d-chord between its
endpoints (that is, has its endpoints at the same height and at distance d from one
another), and if e is any number between 0 and 1, then there is either an e · d-chord
in A or a (1 − e) · d-chord in A. See Rolfsen [96].
Theorem 15.6 (Chord Theorem). Let n denote a positive integer. Then A(0)
must intersect A(1/n). (That is, A(0) must contain a (1/n)-chord.) See Rolfsen
[96] and Figure 2.
Proof. By the Corollary 15.5, A(0) contains either a (1/n)-chord, in which

case we are done, or an (n−1)/n-chord. In the latter case we may apply the corollary
to the (n − 1)/n-chord to find that A(0) contains either a (1/(n − 1)) · ((n − 1)/n) =
1/n chord or an (n − 2)/(n − 1) · (n − 1)/n = (n − 2)/n-chord. By induction, the
numerator of the fraction (n − 2)/n may be reduced step by step to 1, in which
case we are finally done.
The distances 1/n are the only distances d for which we can guarantee that
A(0) must intersect A(d):
Theorem 15.7. If d ∈ (0, 1) is not of the form d = 1/n, then there is an arc
Q(0) joining (0, 0) to (1, 0) such that Q(0) ∩ Q(d) = ∅.
Remark. A student, whose name I do not recall, showed me the following
argument. End remark.
Proof. By hypothesis, there is a positive integer n such that 1/n < d <
1/(n − 1). Construct a polygonal path Q(0) whose consecutive vertices are
(0, 0), (1/n, −1 + 1/n), A = (1/n, 1/n), B = (2/n, −1 + 2/n), (2/n, 2/n),
(3/n, −1 + 3/n), (3/n, 3/n), . . . , (n/n, −1 + n/n) = (1, 0).
Examples appear in Figures 3 and 4.
We denote the point at which the segment AB crosses the x-axis by C.
15.2. THE TORUS AS A EUCLIDEAN SURFACE 131
A = A(0)
A(1/n)
A(0)
Figure 2. Intersection paths
Exercise 15.8. C = (1/(n − 1), 0).

The point (d, 0) lies between the vertical line through A and the point C. The
translate Q(d) lies below the path Q(0) and does not intersect it. This proves the
theorem.
15.2. The Torus as a Euclidean Surface

Definition 15.9. The torus T2 is the product of two circles:
T2 = S1 × S1 .
The easiest way to understand surfaces is to impose nice geometries on them.
In the case of the torus, the appropriate geometry is Euclidean geometry.
Theorem 15.10. The torus T2 can be realized as the quotient space of the
Euclidean plane R2 via an equivalence relation ∼, with identification map p : R2 →
(T2 = R2 / ∼), where (x, y) ∼ (x + m, y + n) for every pair of integers m and n.
Recall that the integer vectors (a, b) ∈ R2 are called lattice points. If we call
the horizontal and vertical lines through the lattice points of the plane lattice lines,
then the lattice lines divide the plane into squares. The projection map p maps
each of these squares onto the torus T2 and identifies the top edge with the bottom
edge, the left edge with the right edge.
Alternatively, we can think of R2 as R × R, we can view S1 as {e2πti | t ∈ R}
and then define p : R2 → T2 by formula.
Theorem 15.11. The projection map p : R2 → T2 can be realized by formula
as
p((s, t)) = (e2πsi , e2πti ).
132 15. THE TORUS
Figure 3. A zig zag curve
Since translations of the plane preserve distances and lengths, the torus inherits
local distances and lengths, hence a notion of “straight”, from the plane via the
projection p. We call a curve in the torus T2 straight if and only if it lies in the
image of a straight line L under the projection p : R2 → T2 . We shall eventually
show that every simple closed curve J on T2 that does not bound a disk is isotopic
in T2 to a straight simple closed curve J .
Likewise, T2 inherits a notion of angles and slopes from R2 via p since vectors
A and B that are identified have neighborhoods that are identified by a translation
that preserves angles and slopes.
Theorem 15.12 (Straight-curve Theorem). The projection map p : R2 → T2
takes straight line L to a (straight) simple closed curve J ⊂ T2 if and only if the
slope of L is rational. In that case, if we express the slope as a fraction a/b in
lowest terms, take A as a point of L, and set B = A + (b, a), then the segment AB
lies in L, p maps AB onto J, p(A) = p(B), and p is one-to-one on the remainder
of the segment AB.
15.2. THE TORUS AS A EUCLIDEAN SURFACE 133
Figure 4. Another zig zag curve
Proof. Since p preserves distances locally, L has infinite length, and J has
finite length (by compactness), it follows that the map p|L cannot be one-to-one.
If p(A) = p(B) with A, B ∈ L, then A ∼ B, so that there is an integer vector
(b, a) such that B = A + (b, a). But that implies that the slope of L is the rational
number b/a.
Conversely, if the slope of L is rational, then there are points A and B on L
that differ by an integer vector (b, a) so that A and B are identified by p. Fix A ∈ L
and let B denote one of the two points of L \ {A} nearest to A that are identified
with A. Then B − A = (b, a), with b and a relatively prime, so that no two points
of AB other than A and B are identified by p. The slopes of L at A and at B
are, of course, the same; and p preserves those slopes. Hence p(AB) is straight at
p(A) = p(B), so that the image of L entering B extends straight into the image of
AB near A. It follows that p(L) is the simple closed curve p(AB).
134 15. THE TORUS
Theorem 15.13 ((a, b) Theorem). If p : R2 → T2 maps an arc α onto a simple

closed curve J, identifying the endpoints A with B of α and being otherwise one-
to-one, then the vector B − A is an integer vector (a, b), with a and b relatively
prime.
Proof. Since p identifies A and B, the vector B − A is, by definition of ∼,

an integer vector (a, b). If a and b had a common divisor n > 1, then by the
Chord Theorem, there are points A and B on the arc α such that the vector
B − A = (a/n, b/n). But that implies that B ∼ A so that the projection map
p identifies A and B . Since n > 1, at least one of the points A and B is not an
endpoint of α. Consequently, p(α) is not a simple closed curve, a contradiction.
15.3. Curve Straightening

Theorem 15.14. Suppose that J is a simple closed curve on the torus T2 that
is not nullhomotopic. Then there is an isotopy of T2 that takes J to a curve that
is straight.
Proof. Here are the main steps in the proof:

(1) The arguments of Chapter 10 allow us to assume that J is a finite union
of straight line segments and that the lift p−1 (J) of J to the plane R2 misses the
lattice points of R2 and crosses each lattice line at each point of intersection.
We assume J adjusted by isotopy so that the intersection of p−1 (J) with a
lattice square has the fewest possible number of components. Then:
(2) If Sq is a lattice square in R2 , then no component J of p−1 (J) ∩ Sq has
both of its endpoints on the same edge of Sq. Furthermore, we may assume that
each component J is, in fact, a straight line interval.
(3) If J is oriented so that each component J of p−1 (J) inherits a direction,
then J is monotone in the sense that it crosses each horizontal lattice line in the
same direction and crosses each vertical lattice line in the same direction.
(4) The curve J can be straightened by isotopy so that p−1 (J) consists of
straight parallel lines of the same rational slope.
Proof of (1). That J can be made polygonal is a consequence of the Polygonal-
curve Lemma 10.17. Then a small translation of the plane R2 descends to an isotopy
of T2 and moves J into the position required by (1).
Proof of (2). Assume that some component J of p−1 (J) has both of its end-
points on the same edge of Sq. By passing to still smaller components if necessary,
we may assume that no point of p−1 (J) ∩ Sq intersects the open arc between the
endpoints of J . See Figure 5.
The component J extends slightly into the neighboring square at each end of
J . Join the endpoints of the extended J with an arc J very close to the edge

of Sq. The arcs J and J bound a disk. Expand that disk slightly to form a
disk D such that J and J are spanning arcs of the disk D. We may require that
the extended J is the entire intersection of D with p−1 (J). There is an isotopy
fixed outside of D that moves J to J . This process removes the component J of
p−1 (J) ∩ Sq and joins together the two components that contained the translation
into Sq of the two end pieces of the extended J . This reduces the number of
components of p−1 (J) ∩ Sq, which contradicts our insistence that this number is
already minimal.
15.3. CURVE STRAIGHTENING 135
Sq
D

J
J
Figure 5. Removing trivial intersections
We conclude that every component J of Sq ∩ p−1 (J) joins distinct sides of Sq.
Hence, by the Straightening Lemma 10.16, we may assume that each component
J is a straight line interval.
This completes the proof of (2).
Proof of (3). If some component J of p−1 (J) does not cross all horizontal
lattice lines in the same direction, then it crosses some horizontal lattice line twice,
say at points a and b, and intersects no horizontal lattice line between those two
crossings. Because of (2), the interval between a and b in J is monotone in the
horizontal direction. In fact, we claim that the interval from a to b can only consist
of two segments as in the Figure 6.
a b
Figure 6. Two allowable segments
We assume this claim for the moment. The two segments joining a to b are
drawn as solid segments. Since the set p−1 (J) is invariant under translations by
integer vectors, there are further segments that are translates of the two segments
of ab and are drawn as dotted segments. The dotted segments limit the horizontal
distance between a and b to be less than 1 so that ab acts just like the component
136 15. THE TORUS
J of (2) with respect to a translate of the square Sq. Thus the arc from a to b
can be pushed across the horizontal line to reduce the number of components of
p−1 (J) ∩ Sq, a contradiction. Thus we can complete the proof of (3) by proving
the claim.
If there were a third segment in the path from a to b, we would have the
following situation (Figure 7):
g
f
e
Figure 7. Trapped segments
The second edge e has a dotted translate that is forced by its nonintersection
with the first edge to rise higher. The third edge f has a dotted translate that is
forced by its nonintersection with the dotted translate of e to rise even higher, and
so on. The path can therefore never return to the level of a. This proves the claim,
and completes the proof of (3).
Proof of (4). By invariance under translation by integer vectors, p−1 (J)
intersects the top of Sq the same number of times it intersects the bottom, and it
intersects the left side the same number of times it intersects the right side. By
the Edge-adjustment Lemma 10.15, we may assume that the points of p−1 (J) on
any horizontal lattice line are evenly spaced, as are those on any vertical lattice
line. By the Straightening Lemma 10.16, we may assume that all segments are
straight. It follows that the components of p−1 (J) are parallel straight lines, evenly
spaced in R2 . Since the projection is a closed curve, there are points on these lines
that are equivalent under p, hence differ by an integer vector. Hence the slopes are
rational.
15.4. Construction of the Simple Closed Curve with Slope k/

Let L be a line in R2 with rational slope k/, where k and are integers that
are relatively prime. Then p(L) is the simple closed curve in T2 . obtained from
the straight segment joining the point (a, b) ∈ L to the point (a, b) + (, k) ∈ L by
identifying segment endpoints. No points of L that are nearer to one another differ
by an integer vector.
If k and are positive, for example, then the construction may be carried out
explicitly as follows in the unit square S which has sides parallel to the x and y
axes and which has the point (a, b) as the lower-left corner. See Figure 8.
15.5. EXERCISES 137
Slope = 3/5
Figure 8. Constructing a simple closed curve
Let p0 , p1 , . . . , pk denote k + 1 points equally spaced on the bottom edge of

S, with p0 denoting the left bottom corner and pk the right bottom corner. Let
p0 , p1 , . . . , pk denote the corresponding points on the top edge. Let q0 , q1 , . . . , q
denote + 1 points equally spaced on the left edge of S, with q0 = p0 denoting
the left bottom corner and q denoting the top left corner. Let q0 , q1 , . . . , q denote
the corresponding points on the right edge. Consider the set K that is the union
of those segments in S of slope k/ joining points labelled either pg , qh , pi , or qj .
Then p(K) is the simple closed curve p(L).
End of construction.
15.5. Exercises
[Hint: Form S2 by adding a point at infinity to R2 . (Remember stereographic
projection.) Apply the Jordan Curve Theorem.]
15.2. Solve Problem 15.2 on page 129.
CHAPTER 16
Orientation and Euler Characteristic
We used both orientation and Euler characteristic in our classification of com-

pact, connected, metric 2-manifolds. Here we shall consider these concepts in more
depth.
16.1. Orientation
It is surprisingly delicate to define the notions of right and left mathematically.
We considered the notion only by means of the Möbius band in earlier sections.
In general, however, these notions are defined in terms of the determinant. For
2-manifolds, this is done in detail as follows.
Definition 16.1. We defined a 2-manifold as a space X that is locally home-
omorphic with the plane. That is, there are an open cover U = {uα | α ∈ A}
of X and homeomorphisms hα : uα → R2 . These homeomorphisms give rise to
transition functions hβ ◦ h−1
α taking the hα -image of uα ∩ uβ to the hβ -image of the
same set. The open cover U and the accompanying homeomorphisms hα are said
to form an atlas (= a collection of charts or road maps—navigational charts) for
the 2-manifold X.
Definition 16.2. We consider the points of the plane R2 = R×R to be column
vectors (a, b)T . (The superscript T indicates transpose.) We may consider each to
be the tip of an arrow with base at the origin (0, 0)T . We consider a pair (a, b)T and
(c, d)T of such arrows to form a right-handed pair if the determinant with columns
(a, b)T and (c, d)T is positive. That is,

a c

b d = ad − bc > 0.
This is true if, when the right-hand is placed over the origin, then the fingers curl
from the point (a, b)T toward the point (c, d)T .
For example, with standard unit vectors e1 = (1, 0)T and e2 = (0, 1)T (T
denoting transpose),

1 0
|e1 e2 | = = 1 > 0,
0 1
so that the pair is right-handed.
In order to localize this definition, we need to allow vectors to be based at
points (e, f )T not equal to the origin of R2 .
Definition 16.3. We consider points (a, b)T and (c, d)T to form a right-handed
pair with respect to (e, f )T if the determinant with columns (a, b)T − (e, f )T and
(c, d)T − (e, f )T is positive.
139
140 16. ORIENTATION AND EULER CHARACTERISTIC
Definition 16.4. A 2-manifold X is said to be orientable, if it admits a family

of local charts {hα : Uα → R2 } such that the transition functions hβ ◦ h−1 α , for each
point (e, f )T in the image of hα , take the pairs that are right-handed with respect
to (e, f )T to pairs that are right-handed with respect to hβ ◦ h−1 T
α ((e, f ) ).
Example 16.5. The compact, connected, orientable 2-manifolds are the 2-

sphere S2 ; the torus T2 , which is the surface of the doughnut, or the 2-sphere with
one handle attached; and the sphere with more than one handle attached. See
Figure 1.
Figure 1. Sphere with two handles
Definition 16.6. In an orientable 2-manifold, every simple closed curve J has

two sides. If the curve J is also oriented, then it is easy to designate one side as
the right-hand side of J and the other side as the left-hand side of J. See Figure 2.
Take a base point B on the curve J. Pick a center point C on J in the direction
from B of the orientation of the curve J. Then a nearby point R is on the right-hand
side of J if the pair (R, C) is right-handed with respect to B.
Definition 16.7. We say that a path K has intersection number +1 with J if
it crosses J from left-hand side to right-hand side. If K crosses from right to left,
we assign intersection number −1. See Figure 3.
The notions of manifold and orientation can be generalized to all dimensions

in a straightforward way.
Definition 16.8. We define an n-manifold as a space X that is locally home-
omorphic with the Euclidean n-dimensional space Rn . That is, there are an open
cover U = {uα | α ∈ A} of X and homeomorphisms hα : uα → Rn . These
16.2. EULER CHARACTERISTIC 141
Figure 2. Defining the right-hand side
K
+1
Figure 3. Defining a positive crossing
homeomorphisms give rise to transition functions hβ ◦ h−1

α taking the hα -image of
uα ∩ uβ to the hβ -image of the same set. The open cover U and the accompanying
homeomorphisms hα are said to form an atlas for the n-manifold X.
Definition 16.9. We consider the points of Rn to be column vectors (a1 , a2 ,
. . . , an )T . (The superscript T indicates transpose.) We may consider each to be the
tip of an arrow with base at the origin (0, 0, . . . , 0)T . We consider an n-tuple of such
arrows to be right-handed if the determinant with the given vectors as columns is
positive.
Definition 16.10. An n-manifold X is said to be orientable, if it admits a
family of local charts {hα : Uα → R2 } such that the transition functions hβ ◦ h−1
α ,
for each point in the image of hα , take the n-tuples that are right-handed to image
n-tuples that are also right-handed.
16.2. Euler Characteristic

Setting. We assume, as given, a finite collection C of objects Ci which we
call open cells, each with an assigned dimension, |Ci | = dim(Ci ) ≥ 0. With such a
collection, there is an operation called simple subdivision which removes one cell
Ci of dimension n = |Ci | > 0 from the collection C and inserts in its place two new
cells of dimension n and one new cell of dimension n − 1. These three new cells are
said to subdivide the original cell. See Figure 4.
dim n − 1
dim n
dim n
dim n
Figure 4. Simple subdivision
Definition 16.11. The Euler characteristic of the collection C gives a first

approximate answer to the following question: If n(k) denotes the number of open
cells having dimension k, what functions of the numbers n(k) are invariant under
simple subdivision? Subdivision increases the count n(k) of cells having dimension
k by 1 and also increases the count n(k − 1) of cells having dimension k − 1 by 1.
Thus the difference n(k) − n(k − 1) remains unchanged under simple subdivision.
This suggests that we consider the alternating sum

χ(C) = n(0) − n(1) + n(2) − n(3) + · · · = (−1)k n(k).
k
We call the number χ(C) the Euler characteristic of the collection C.

Theorem 16.12. Euler characteristic χ(C) is invariant under simple subdivi-
sion.
When we define collections C and simple subdivision geometrically, then Euler
characteristic likewise takes on geometric significance. Before we give a precise
general geometric definition, we consider as examples finite graphs and Platonic
solids.
Example 16.13 (The finite graphs). As open 0-cells we take the vertices of the
graph. As open 1-cells we take the open edges of the graph. As simple subdivision
we take the insertion of a new vertex into the interior of an edge of the graph.
That new vertex divides the old edge into two new edges and inserts one new 0-cell.
Since finite graphs and simple subdivision satisfy the requirements of the setting,
Euler characteristic is invariant under simple subdivision.
Exercise 16.14. The Euler characteristic of a finite tree T is 1. The Euler
characteristic of a finite connected graph Γ is equal to 1 − n, where n is the number
of edges remaining when the edges of a maximal tree have been omitted from the
count.
Exercise 16.15 (The platonic solids). There are five regular convex polyhedra
whose boundaries consist of faces that are all of the same polygonal shape. See
Figure 5.
Figure 5. The platonic solids
Discussion: The polygons meeting at a vertex must have total angle sum at
that vertex of less than 2π. For triangles, this angle restriction dictates that 3, 4, or
5 triangles meet at every vertex. For squares, the only possibility is 3 squares at a
vertex. For pentagons, the only possibility is also 3 at a vertex. These requirements
immediately dictate the combinatorial structure of the polyhedra and lead to the
tetrahedron, the octahedron, and the icosahedron for triangles; the cube for squares;
and the dodecahedron for pentagons.
In a platonic solid, we count the vertices as open 0-cells, the interiors of edges
as open 1-cells, and the open polygons as open 2-cells. Counting vertices, edges,
and faces of each of the Platonic solids yields the following data:
Polyhedron Vertices Edges Faces
Tetrahedron 4 6 4
Octahedron 6 12 8
Icosahedron 12 30 20
Cube 8 12 6
Dodecahedron 20 30 12
In each case,
χ = vertices − edges + faces = 2.
This number χ is called the Euler characteristic of the surface of the polyhedron.
It is, in fact, completely natural to expect that the Euler characteristic of each of
these surfaces should be the same because, combinatorially, each can be derived by
simple subdivision from the same cell complex, as we shall now explain.
Setting. Let Γ denote a finite graph that divides the 2-sphere S2 into disks in
the sense that each component of S2 \ Γ is the interior of a disk whose boundary
is a simple closed curve carried by a subgraph of Γ. The vertices of the graph are
called the open 0-cells of the complex. The interiors of the edges of the graph are
called the open 1-cells of the complex. The components of S2 \ Γ are called the open
2-cells of the complex. There are two sorts of simple subdivision: (1) insertion of
a new vertex into the interior of an edge of the graph. (2) Insertion of a spanning
arc through one of the disks of the complex whose boundary vertices are already
vertices of the graph. The Theta-curve Theorem 8.7 implies that simple subdivision
satisfies the requirements of the setting leading to Euler characteristic.
Exercise 16.16. Let Γ denote a graph in the 2-sphere S2 that has one vertex
and one edge. Let C denote the resulting complex. The Euler characteristic is
therefore χ(C) = 1 − 1 + 2 = 2. If P is a platonic solid, then up to homeomorphism,
P can be derived from C by simple subdivision. Consequently, since χ(C) = 2,
χ(P ) is also 2. If we actually consider the 3-dimensional solid which includes the
interior of the 2-sphere as an open 3-cell, then the Euler characteristic of this solid
is 2 − 1 = 1.
Definition 16.17. The natural setting for the Euler characteristic is the finite
CW-complex C. This is a space formed as the identification space of a finite number
of finite-dimensional balls built up inductively, by dimension, as follows.
A finite CW-complex C 0 of dimension 0 is a finite collection of points, each
comprising an open set in C 0 . These points are called open 0-cells.
A finite CW-complex C n of dimension n is formed from a finite CW-complex
C n−1
of dimension n − 1 and a disjoint union of finitely many n-dimensional balls
B1 , . . . , Bk by identification as follows. For each Bi , there is a continuous function
fi : ∂Bi → C n−1 called an attaching map. Let X n denote the disjoint union of
the space C n−1 and the balls B1 , . . . , Bk . Let ∼ denote the equivalence relation
generated by requiring that each element x of each set ∂Bi be equivalent to its
image fi (x) under the attaching map fi . Then C n = X n / ∼. This space C n is
called the n-skeleton of C.
Note that the sets int(Bi ) are embedded in C n . They are called the open n-cells
of the complex C n .
Definition 16.18 (Simple subdivision). Suppose that C is a CW-complex and
that C is an open cell in C of dimension n. Then C is the interior of an n-dimensional
ball B that has been attached to the n−1 dimensional skeleton C n−1 by an attaching
map f : ∂B → C n−1 . Let B denote an (n−1)-dimensional ball in B with ∂B ⊂ ∂B
and with int(B ) ⊂ int(B) that divides B into two n-dimensional balls B1 and B2 .
Assume in addition that f |∂B : ∂B → C n−1 actually takes ∂B into C n−2 . Then
we may form a new CW-complex by omitting the open n-cell int(B), attaching one
new (n − 1) cell by the attaching map f |∂B , and attaching two new n-cells by the
attaching maps f ∪ id|∂B1 and f ∪ id|∂B2. We call the resulting CW-complex a
simple subdivision of C.
Corollary 16.19. The Euler characteristic χ(C) of a finite CW-complex is
invariant under simple subdivision.
There is a stronger invariance theorem that we shall assume but not prove.
Theorem 16.20 (Topological invariance of Euler characteristic). If C and D are
two finite CW-complexes whose underlying identification spaces are homeomorphic,
then χ(C) = χ(D). In fact, if the homology of C is isomorphic to the homology
of D, then χ(C) = χ(D), which happens for example when C and D are homotopy
equivalent.
Remark. The theorem on topological invariance of Euler characteristic allows
us to speak of the Euler characteristic of a sphere or ball, of a compact 2-manifold,
of n-space, etc. End remark.
Corollary 16.21. The Euler characteristic of the n-sphere Sn is equal to 2
in even dimensions and is equal to 0 in odd dimensions.
Proof. The n-sphere can be realized as a CW-complex with one 0-cell and
one n-dimensional cell. If n is even, then χ(Sn ) = 10 + 1n = 2. If n is odd, then
χ(Sn ) = 10 − 1n = 0.
Corollary 16.22. The Euler characteristic of the n-ball Bn is equal to 1.

Proof. The n-ball can be realized as a CW-complex with one 0-cell, one (n −
1)-dimensional cell (forming the (n − 1)-sphere), and one n-cell attached by the
identity map on the (n − 1)-sphere. Thus, for n even, χ(Bn ) = 10 − 1n−1 + 1n = 1
and, for n odd, χ(Bn ) = 10 + 1n−1 − 1n = 1. Alternatively, Bn is homotopy
equivalent to a single point, hence has the Euler characteristic of a point.
Corollary 16.23. The Euler characteristic of n-dimensional projective space

Pn is equal to 1 if n is even, and is equal to 0 if n is odd.
Proof. Another CW-complex realizing the n-sphere can be described as fol-
lows. The 0-sphere consists of two points. The 1-sphere can be realized by attaching
two 1-cells to the 0-sphere, each with boundary mapping homeomorphically to the
0-sphere. In general, the n-sphere can be realized by attaching two n-cells to the
(n − 1)-sphere, each with boundary mapping homeomorphically to the (n − 1)-
sphere. Then n-dimensional projective space is a quotient of Sn by declaring an-
tipodal points of Sn to be equivalent. This map identifies the two 0-cells of the
0-sphere, the two 1-cells of the 1-sphere, the two 2-cells of the 2-sphere, etc. Thus
projective n-space can be realized as a CW-complex with one cell in each dimension
0, 1, . . . , n. Thus the Euler characteristic is
χ(Pn ) = 10 − 11 + 12 − 13 + · · · + (−1)n 1n .
That is, χ(Pn ) = 1 if n is even, and 0 if n is odd.
Exercise 16.24 (Euler characteristic of a union). Suppose that C is a CW-

complex that is the union of two subcomplexes C1 and C2 that intersect in a sub-
complex D. Then χ(C) = χ(C1 ) + χ(C2 ) − χ(D).
Corollary 16.25 (Euler characteristic of a connected sum). Suppose that M1
and M2 are two compact, connected 2-manifolds and that N = M1 #M2 is their
connected sum. Then χ(N ) = χ(M1 ) + χ(M2 ) − 2.
Proof. We may assume that M1 and M2 are triangulated, that the interior
of a triangle is deleted from each to form 2-manifolds-with-boundary M1 and M2 ,
and that the boundaries of those two triangles are identified to form N . By the
exercise on the Euler characteristic of a union,
χ(N ) = χ(M1 ) + χ(M2 ) − χ(M1 ∩ M2 ).
But χ(Mj ) = χ(Mj ) − 1 since Mj has one fewer 2-cells than does Mj . The intersec-
tion of M1 and M2 after identification of boundaries is a 1-sphere, which has Euler
characteristic 0. The corollary follows.
Theorem 16.26 (Euler characteristic of a 2-manifold). Suppose that the com-
pact, connected 2-manifold M is the connected sum of S2 with k copies of the torus
and = 0, 1, or 2 copies of the projective plane. Then the Euler characteristic of
M is
χ(M ) = 2 + (k · 0 − 2k) + ( · 1 − 2) = 2 − 2k − .
Proof. The Euler characteristic of S2 is 2. Each of the k-connected summands
that is a torus adds the characteristic 0 of the torus and subtracts 2. Each of the
-connected summands that is a projective plane adds the characteristic 1 of the
projective plane and subtracts 2.
Corollary 16.27. Two compact connected 2-manifolds M1 and M2 are home-
omorphic iff the following two conditions are satisfied:
(1) Both are orientable or both are nonorientable.
(2) χ(M1 ) = χ(M2 ).
Proof. The two conditions are clearly necessary in order that M1 and M2
be homeomorphic. Suppose both are satisfied. We may realize Mj as a connected
sum of the 2-sphere with kj copies of T2 and j = 0, 1, or 2 copies of P2 . If both
are orientable, then 1 = 2 = 0. If both are nonorientable, then 1 = 2 = 2 if
χ(M1 ) = χ(M2 ) is even, and 1 = 2 = 1 if χ(M1 ) = χ(M2 ) is odd. Hence, in any
case, 1 = 2 . It follows that k1 = k2 . Hence M1 and M2 are homeomorphic.
Theorem 16.28 (Euler characteristic of an odd-dimensional manifold). The
Euler characteristic of an odd-dimensional manifold M is 0.
Proof. Though the result is true in general by homological considerations
(Poincaré duality), we shall only prove it for manifolds that are triangulated. We
employ Alexander’s algebraic representation which describes each simplex as the
product (join) of its vertices and the manifold itself as the formal sum M of its
principal simplexes (simplexes of maximal dimension). If σ is a simplex of M , then
M can be factored in the form M = σL + K. The subcomplex L is called the
link of σ in M , and the product σL is called the star of σ in M . In the standard
(combinatorial) triangulations of an n-manifold, the link of a k-simplex is always
a sphere of dimension n − k − 1. We assume this to be true of our triangulations.
The algebraic product of disjoint simplexes is called their join.
We will sum the Euler characteristics of the links of every simplex of M , di-
mension by dimension. We calculate the answers in two different ways. In the
first, we use our knowledge that vertex links are (n − 1)-spheres, edge links are
(n − 2)-spheres, triangle links are (n − 3)-spheres, etc. In the second, we consider
the contribution of each simplex to all link Euler characteristics and sum over all
simplexes.
First calculation: The Euler characteristic of a link is 2 for simplexes of even
dimension, 0 for simplexes of odd dimension. Therefore the sum over all simplexes
is
2 · #(0) + #(2) + #(4) + · · · + #(n − 1) ,
where #(k) denotes the number of simplexes of dimension k.
Second calculation: We fix a simplex α of dimension k and consider any j-

dimensional face β of α, with j < k. Then there is a complementary face β ∗ of α
such that α is the join of β and β ∗ . The dimension of β ∗ is k − j − 1, and β ∗ is in
the link of β in M . We see that α contributes (−1)k−j−1 to the Euler characteristic
k+1
of the link of β. Note that there are exactly j+1 j-simplexes β having link Euler
characteristics to which α contributes. Thus we have:

2 3 4 n+1
χ(link(v)) = #(1) − #(2) + #(3) − · · · + #(n) ,
v a vertex
1 1 1 1

3 4 n+1
χ(link(e)) = +#(2) − #(3) + · · · − #(n) ,
2 2 2
e an edge

4 n+1
χ(link(t)) = +#(3) − · · · + #(n) ,
3 3
t a triangle
... ...

n+1
χ(link(α)) = +#(n) .
n
dim(α)=n−1
The link of an n-simplex is empty and thus does not appear. Summing, we
obtain

2 · #(1) + #(3) + #(5) + · · · + #(n) .
Equating our two calculations, we find

2 · #(0) + #(2) + #(4) + · · · + #(n − 1) = 2 · #(1) + #(3) + #(5) + · · · + #(n) ,
or χ(M ) = 0.

Theorem 16.29. Every polygonal knot in R3 is the boundary of an orientable

surface-with-boundary.
Proof. In the figure we see a standard projection of a knot, with six crossings
and with an orientation as indicated by the arrows. See Figure 6.
At each crossing we attach a twisted strip, blue on one side, green on the other,
in such a way that two opposite sides of the strip follow the knot while the other
two sides are free.
We replace the two segments of the knot that lie on any strip by the opposite
pair of sides of the strip. The result is a family of disjoint simple closed curves in the
plane called Seifert circles, named after the mathematician Seifert who described
this construction.
These Seifert circles bound disjoint disks, called Seifert disks, in R3 in such a
way that the union of these disks with the half-twisted strips forms a 2-manifold M
with boundary equal to the original knot. This surface is called a Seifert surface.
In order to make sure that the 2-manifold M is orientable, we must take care to
insert the twisted strips so that the Seifert circles inherit an orientation compatible
with the original orientation of the knot. There are two choices at each crossing for
the insertion of the strip, and one of the two has the required orientation.
Twisted Strip at a Crossing
Knot
Knot wih Twisted Strips Attached
Seifert Circles
Figure 6. Seifert circles
The relevance of the construction to our discussion of Euler characteristic is

the following.
Theorem 16.30. The Euler characteristic of the Seifert surface is circles −
crossings, where circles is the number of Seifert circles and crossings is the num-
ber crossings of the knot, hence the number of half-twisted strips.
Proof. The Seifert disks can be joined by circles − 1 twisted strips to form a
single disk, which has Euler characteristic equal to 1. Each additional twisted strip
adds 0 vertices, 2 edges, and 1 face, hence decreases the Euler characteristic by 1.
Therefore the final Euler characteristic is
1 (from the disk) − [crossings − (circles − 1)]( the number of additional strips)
= circles − crossings.
16.3. EXERCISES 149

In the example, circles = 3 and crossings = 6. Therefore the Euler characteris-
tic is 3 − 6 = −3. If we were to fill in the boundary of the surface with a disk, we
would increase the Euler characteristic by 1 to form a closed orientable surface of
Euler characteristic = −2. Thus the Seifert surface is a disk with two handles.
16.3. Exercises
16.4.Solve Exercise 16.24 on page 145.
16.5. Construct orientable Seifert surfaces for the two knots in Figure 7. Cal-
culate the Euler characteristic of each surface.
Figure 7. Two knots
16.6. Given the edge pairings in Figures 8 and 9, determine whether the re-
sulting surfaces are orientable and calculate the Euler characteristic.
e−1
c−1
d−1
a−1
a
b−1
c
Figure 8. An edge-pairing
g c
d
g −1
b
e
f −1
a−1
a
j
i
i
e
h−1
c−1
b
h
d f
j
Figure 9. Another edge-pairing

CHAPTER 17
The Riemann-Hurwitz Theorem
We can use Euler characteristic to prove a foundational theorem in the study

of rational mappings of the 2-sphere, called the Riemann-Hurwitz Theorem. The
ways in which a rational map of the 2-sphere twists and turns the sphere are very
complex and very beautiful. There is a huge literature, with remarkable computer
graphics, designed to show just how interesting these maps are. For a start, see
Mandelbrot [86], Devaney [88], and Falconer [87].
17.1. Setting
We first have to prepare the setting.
We consider functions f (z) of one complex variable z which are rational in the
sense that they are the quotient f (z) = p(z)/q(z) of two polynomials p(z) and q(z)
with complex coefficients.
The 2-dimensional sphere S2 can be interpreted as the Riemann sphere, which is
the union of the complex plane C with a single point at infinity. The correspondence
which supports this interpretation is stereographic projection, which we described
in detail in the section on Pythagorean triples, Volume 1, Chapter 2.
A rational function of one complex variable can therefore be interpreted as a
function that maps the two-dimensional sphere S2 onto itself. The poles of the
rational function are the points that are mapped to infinity. The image of infinity
can be calculated by taking the limit of the function as domain points approach
infinity.
Such a function can be iterated to form a discrete dynamical system, and the
properties of such systems are unendingly fascinating: Near which points is the
system infinitely complicated (the Julia set)? Near which points is the system
simple (the Fatou set)? The computer visualizations of such sets are intricate and
beautiful. The study of even the family of quadratic polynomials of the form z 2 + c
is complicated and leads to the famous Mandelbrot set, which is parametrized by
the complex numbers c. A first approximation to the complexity is supplied by the
singularities of the rational map, and the nature of the singularities is controlled to
some extent by Euler characteristic as appearing in the Riemann-Hurwitz Theorem.
Before explaining that theorem, we need to describe the local properties of a rational
map.
17.2. Elementary Facts from Trigonometry

The first thing to learn is that, locally, every rational map f (z) acts like a
power map z
→ z n , with n a positive integer called the local degree of f at the
point in question. In order to understand this power map, we repeat our earlier
review elementary facts from trigonometry.
151
152 17. THE RIEMANN-HURWITZ THEOREM
Theorem 17.1. (d/dx) sin(x) = cos(x) and (d/dx) cos(x) = − sin(x).

Proof. Let p(t) = (c(t), s(t)) denote the position of a particle moving at
speed 1 counterclockwise around the unit circle beginning at (1, 0) at time 0. Since
|p(t)| = 1,
1 = |p(t)|2 = (c(t), s(t)) · (c(t), s(t)) = c2 (t) + s2 (t).
Taking derivatives, we find

0 = 2 c(t) · c (t) + s(t) · s (t) .
That is p(t) · p (t) = 0, and the position vector p(t) is perpendicular to the velocity
vector p (t). By hypothesis, the speed |p (t)| is equal to 1. Thus p (t) is a vector
of length 1 perpendicular to the unit circle at p(t) and pointing counterclockwise.
See Figure 1.
p (t)
(−s(t), c(t))
p (t) (c(t), s(t))

p(t)
(1, 0)
Figure 1. Derivative of the sine and cosine
There is obviously only one vector p (t) satisfying these conditions, namely

p (t) = (−s(t), c(t)). Since c(t) = cos(t), and s(t) = sin(t), the proof is complete.

Theorem 17.2 (Projection principle). If Δ is a right triangle in the plane with

hypotenuse of length r and if α is one of the acute angles of the triangle, then the
length of the leg adjacent to α is r · cos(α) and the length of the leg opposite the
angle α is r · sin(α). See Figure 2.
Theorem 17.3 (Sum formulas).

cos(α + β) = cos(β) · cos(α) − sin(β) · sin(α) and
sin(α + β) = sin(β) · cos(α) + cos(β) · sin(α).

17.2. ELEMENTARY FACTS FROM TRIGONOMETRY 153
(sin α) · r
(cos α) · r
α
Figure 2. The projection principle

sin β · sin α
(C, S)
cos β · sin α
1 β
α sin β · cos α
cos β · cos α
Figure 3. The sum formulas
Proof. Graph the point (c, s) = (cos(α + β), sin(α + β)), and apply the pro-
jection principle to evaluate the four lengths indicated by arrows (Figure 3).
It is obvious from the projection principle as applied to the diagram, that the
products that appear in the sum formulas have clear geometric meaning so that
cos(α + β) = c = cos(β) · cos(α) − sin(β) · sin(α)
and that
sin(α + β) = s = sin(β) · cos(α) + cos(β) · sin(α),
as required by the theorem.
Definition 17.4. Every complex number z can be expressed in polar coordi-

nates as z = [r, θ]pol , where r = |z| is the length of the vector from 0 to z and θ is
the angle as measured from the positive x-axis to the same vector. The angle θ is
only well-defined up to multiples of 2π. In coordinates,
z = x + iy = r(cos θ + i sin θ),
so that x = r cos θ and y = r sin θ.
Theorem 17.5. [r, θ]pol · [s, η]pol = [rs, θ + η]pol .
Proof.
[r, θ]pol · [s, η]pol = r(cos θ + i sin θ) · s(cos η + i sin η)

= rs (cos θ · cos η − sin θ · sin η) + i(sin θ · cos η + cos θ · sin η)
= rs(cos(θ + η) + i(sin(θ + η))
= [rs, θ + η]pol .

Corollary 17.6. The power map z

→ z n takes [r, θ]pol to [r n , n · θ], hence
maps the circle of radius r counterclockwise n times around the circle of radius
rn .
17.3. Branched Maps of S2

Definition 17.7. A continuous function f : S1 → S2 from one closed, oriented
surface S1 to another S2 is called a branched map if, at each point p, there is a
positive integer n(p) such that f is locally equivalent to a power map z
→ z n(p) .
The integer n(p) is called the local degree of f at p. To be more precise, there are
neighborhoods M and N of p and of f (p), respectively, and orientation preserving
homeomorphisms uM and vN from M and N , respectively, onto neighborhoods of
0 in the complex plane C that take p and f (p) to 0 such that
vN ◦ f ◦ u−1
M (z) = z
n(p)
for each z ∈ uM (M ). It follows immediately that f is locally a homeomorphism

except at the points for which n(p) > 1, and there are only finitely many such
points. These points are called the critical points of the branched map f , and their
images are called the critical values of f .
17.5. PROOF OF THE RIEMANN-HURWITZ THEOREM 155
17.4. Statement of the Riemann-Hurwitz Theorem

Theorem 17.8 (The Riemann-Hurwitz Theorem). Suppose that S1 and S2 are
compact, connected, oriented surfaces and that f : S1 → S2 is a branched map.
(1) If V is the set of critical values of the map f , then f |S1 \ f −1 (V ) : S1 \
−1
f (V ) → S2 \ V is a finite sheeted covering map, so that every point of S2 \ V has
the same number of preimages in S1 . This number is called the degree deg(f ) of
f.
(2) The Euler characteristics of S1 and S2 satisfy the equation

χ(S1 ) = deg(f ) · χ(S2 ) − {n(p) − 1 | p ∈ S1 }.
(3) If S1 = S = S2 , then

(deg(f ) − 1) · χ(S) = {n(p) − 1 | p ∈ S} ≥ 0.
Corollary 17.9 (Corollary to the Riemann-Hurwitz Theorem). If S1 = S =

S2 , then there are three possibilities:
(1) If χ(S) < 0, then deg(f ) = 1 and f is a homeomorphism.
(2) If χ(S) = 0 so that S is the torus, then f is a local homeomorphism and f
is a covering map.
(3) If χ(S) > 0 so that S is the 2-sphere (χ(S2 ) = 2), then

2 deg(f ) − 2 = {n(p) − 1 | p ∈ S}.
17.5. Proof of the Riemann-Hurwitz Theorem

Proof. (1) A power map is both open and finite to one. Hence, since S1 is
compact and locally a power map, f is globally finite to one and has both open
and compact image. Hence the map f is onto since the only open, compact, and
nonempty subset of S2 is S2 itself. The map f can have only finitely many critical
points since S1 is compact and each point has a neighborhood with at most one
critical point. Hence the set V ⊂ S2 of critical values is also finite and its inverse
image f −1 (V ) is also finite. Thus we have an induced surjective, finite-to-one map
f |S1 \ f −1 (V ) : S1 \ f −1 (V ) → S2 \ V . We must show that this restricted map is a
covering map.
Let q ∈ S2 \ V . Let p1 , . . . , pk be the points in the preimage of q. Each has
a neighborhood mapped homeomorphically onto a neighborhood of q in S2 \ V .
The intersection of these neighborhoods is a neighborhood mapped evenly by the
restriction of f . Thus the restriction of f is a covering map.
Since f −1 (V ) is finite, S1 \ f −1 (V ) is a connected set. Hence every point of
S2 \ V has the same number of preimages by standard properties of a covering map.
This number is called the degree deg(f ) of f .
(2) Let T2 triangulate S2 in such a way that each element of V is a vertex.
Lift this triangulation to a triangulation T1 of S1 . Assign to each vertex p of T1
its degree n(p). We then calculate Euler characteristics. Note that each point of
S2 \V has exactly deg(f ) preimages so that each edge of T2 has exactly deg(f ) edges
in its preimage and each face of T2 also has exactly deg(f ) faces in its preimage.
The behavior at vertices is more complicated. If q is a vertex, q is a point of
S2 very close to q, and if p1 , . . . , pk are the preimages of q, then q has exactly
deg(f ) preimages, but n(pi ) of these are arrayed in a small circle around pi . That

is deg(f ) = {n(pi ) | i = 1, . . . , k}, and, consequently, q has only k = deg(f ) −

{n(pi ) − 1 | i = 1, . . . , k} preimages. We thus obtain the following equation:
χ(S1 ) = V (T1 ) − E(T1 ) + F (T1 )

= deg(f )(V (T2 ) − E(T2 ) + F (F2 )) − {n(p) − 1 | p ∈ S1 }

deg(f )χ(S2 ) − {n(p) − 1 | p ∈ S1 }
(3) If S1 = S = S2 , then the equation

(deg(f ) − 1) · χ(S) = {n(p) − 1 | p ∈ S} ≥ 0
is an immediate consequence of (2).
Proof of the Corollary 17.9 to the Riemann-Hurwitz Theorem.
(1) If χ(S) < 0, then (deg(f ) − 1) · χ(S) is both ≤ 0 and ≥ 0, hence = 0. Thus
deg(f ) = 1 and f is a homeomorphism.
If χ(S) = 0 so that S is a torus, then the (deg(f ) − 1) · χ(S) = 0 so
(2)
that {n(p) − 1 | p ∈ S} = 0. Consequently, each n(p) is 1 and f is a local
homeomorphism, hence a covering map by (1) of the Riemann-Hurwitz Theorem.
S is the 2-sphere, then χ(S) = 2 so that (deg(f ) − 1) ·
(3) If χ(S) > 0 so that
χ(S) = 2 deg(f ) − 2 = {n(p) − 1 | p ∈ S}.
17.6. Rational Maps

In order to apply the Riemann-Hurwitz Theorem to rational maps, we must
show that each rational map is a branched map of the 2-sphere.
Theorem 17.10. A map f : S1 → S2 is a branched map if it is finite-to-one
and there is a finite subset V ⊂ S2 such that f |S1 \ f −1 (V ) : S1 \ f −1 (V ) → S2 \ V
is an orientation-preserving covering map. See Figure 4.
Proof. Consider a point q ∈ V and a point p ∈ f −1 (q). Let N denote an
open neighborhood of p that contains no other point of f −1 (V ), and let D denote
a disk neighborhood of q in the open neighborhood f (N ) of q. Let J denote the
simple closed curve that is the boundary of D, and let A denote an arc in D that
irreducibly connects J to q.
We may view D cut along A as a singular disk in which two boundary arcs
have been identified to form A. By basic covering theory, this disk has lifts
D1 , D2 , . . . , Dn to N , and the union of these lifts forms a neighborhood of p in
N . (In the figure, there are three lifts D1 , D2 , D3 of D and three lifts of the arc
A.) Let fi : Di
→ D denote the restriction of f to Di . Then these maps twist the
neighborhood Di n times around q just as the map z
→ z n twists a neighborhood
of the origin 01 in one copy of C around the origin 02 of the image copy of C. Hence
the map f is branched at p with local degree n.
Theorem 17.11. The composite of two branched maps is a branched map.
Proof. The composite of two finite sheeted covering maps is a covering map.
Hence this theorem follows from the previous one.
Theorem 17.12. Every rational map f : S2 → S2 is a branched map.
Proof. A rational map is a composite of maps of the form z
→ Cz, z
→ z n ,
z
→ z + D, z
→ z −n , each of which is obviously a branched map. Hence this
theorem follows from the previous one.
17.7. EXERCISES 157
f (N )
N
q
D3
p D2 A
D
D1
Figure 4. The local view of a branched cover
17.7. Exercises
17.1. For what surfaces S1 and S2 is it possible to construct a branched map
f : S1 → S2 ?
17.2. The study of branched maps from the 2-sphere to itself has spawned a
huge literature. Look up the Mandelbrot set on the internet. Find other references
there. Find the relationship between branched maps and fractals.
Bibliography
Plain Fun
(top recommendations for easy, but rewarding, pleasure).
[1] Hardy, G. H., A Mathematician’s Apology, Cambridge University Press, 2004 (eighth print-
ing).
[2] Pólya, G., How to Solve It, Princeton Univerity Press, 2004.
[3] Körner, T. W., The Pleasures of Counting, Cambridge University Press, 1996.
- More Fun
[4] Davis, P. J. and Hersh, R., The Mathematical Experience, Houghton Mifflin Company, 1981.
[5] Rademacher, H., Higher Mathematics from an Elementary Point of View, Birkhäuser, 1983.
[6] Hilbert, D., and Cohn-Vossen, S., Geometry and the Imagination, (translated by P. Nemeyi),
Chelsea Publishing Company, New York, 1952. [College level exposition of rich ideas from
low-dimensional geometry, with many figures.]
[7] Dörrie, H., 100 Great Problems of Elementary Mathematics: Their History and Solution,
Dover Publications, Inc., 1965, pp. 108-112. [We learned our first proof of the fundamental
theorem of algebra here.]
[8] Courant, R. and Robbins, H., What is Mathematics?, Oxford University Press, 1941.
Classics
(a chance to see the thinking of the very best, in chronological order).
[9] Euclid, The Thirteen Books of Euclid’s Elements, Vol. 1-3, 2nd Ed., (edited by T. L. Heath)
Cambridge University Press, Cambridge, 1926. [Reprinted by Dover, New York, 1956.]
[10] Archimedes, The Works of Archimedes, edited by T. L. Heath, Dover Publications, In.,
Mineola, New York, 2002. See also the exposition in Pólya, G., Mathematics and Plausible
Reasoning, Vol. 1. Induction and Analogy in Mathematics, Chapter IX. Physical Mathemat-
ics, pp. 155-158, Princeton University Press, 1954. [How Archimedes discovered the integral
calculus.]
[11] Wallis, J., in A Source Book in Mathematics, 1200-1800, edited by D. J. Struik., Harvard
University Press, 1969, pp. 244-253. [Wallis’s product formula for π.]
[12] Gauss, K. F., General Investigations of Curved Surfaces of 1827 and 1825, Princeton Uni-
versity Library, 1902. [Available online. Difficult reading.]
[13] Fourier, J., The Analytical Theory of Heat, translated by Alexander Freeman, Cambridge
University Press, 1878. [Available online, 508 pages. The introduction explains Fourier’s
thoughts in approaching the problem of the mathematical treatment of heat. Chapter 3 ex-
plains his discovery of Fourier series.]
[14] Riemann, B., Collected Papers, edited by Roger Baker, Kendrick Press, Heber City, Utah,
2004. [English translation of Riemann’s wonderful papers.]
159
160 BIBLIOGRAPHY
[15] Poincaré, H., Science and Method, Dover Publications, Inc., 2003. [Discusses the role of the
subconscious in mathematical discovery.] Also, The Value of Science, translated by G. B.
Halstead, Dover Publications, Inc., 1958.
[16] Klein, F., Vorlesungen über Nicht-Euklidische Geometrie, Verlag von Julius Springer, Berlin,
1928. [In German. An algebraic development of non-Euclidean geometry with respect to the
Klein and projective models. Beautiful figures. Elegant exposition.]
[17] Hilbert, D., Gesammelte Abhandlungen (Collected Works), 3 volumes, Springer-Verlag,
1970. [In German. The transcendence of e and π appears in Volume 1, pp. 1-4. Hilbert’s
space-filling curve appears in Volume 3, pp. 1-2.]
[18] Einstein, A., The Meaning of Relativity, Princeton University Press, 1956.
[19] Thurston, W. P., Three-Dimensional Geometry and Topology, edited by Silvio Levy, Prince-
ton University Press, 1997. [An intuitive introduction to dimension 3 by the foremost ge-
ometer of our generation.]
[20] W. P. Thurston’s theorems on surface diffeomorphisms as exposited in Fathi, A., and Lau-
denbach, F., and Poénaru, V., Travaux de Thurston sur les Surfaces, Séminaire Orsay,
Société Mathématique de France, 1991/1979. [In French.]
History
(concentrating on famous mathematicians).
[21] Bell, E. T., Men of Mathematics, Simon and Schuster, Inc., 1937. [The book that convinced
me that mathematics is exciting and romantic.]
[22] Henrion, C., Women of Mathematics, Indiana University Press, 1997.
[23] Dunham, W., Journey Through Genius, Penguin Books, 1991.
Supporting Textbooks
- Topology
[24] Munkres, J. R., Topology, a First Course, Prentice-Hall, Inc., 1975. [The early chapters
explain the basics of topology that form the prerequisites for the latter half of this book. The
later chapters contain rather different views of some of the later theorems in our book.]
[25] Massey, W. S., Algebraic Topology: An Introduction. Springer-Verlag, New York -
Heidelberg-Berlin, 1967 (Sixth printing: 1984), Chapter I, pp. 1-54. [A particularly nice
introduction to covering spaces.]
[26] Hatcher, A., Algebraic Topology , Cambridge University Press, 2001. [A very nice introduc-
tion to algebraic topology, a bit of which we need in Volume 2.]
[27] Munkres, J. R., Elements of Algebraic Topology, Addison-Wesley, 1984. [Another nice in-
troduction.]
[28] Alexandroff, P., Elementary Concepts of Topology, translated by Alan E. Farley, Dover
Publications, Inc., 1932. [A wonderful small book.]
[29] Alexandrov, P. S., Combinatorial Topology, 3 volumes, translated by Horace Komm, Gray-
lock Press, Rochester, NY, 1956.
[30] Seifert, H., and Threlfall, W., A Textbook of Topology, translated by Michael A. Goldman;
and Seifert, H., Topology of 3-Dimensional Fibered Spaces, translated by Wolfgang Heil,
Academic Press, 1980. [Available online.]
[31] Hurewicz, W., and Wallman, H., Dimension Theory, Princeton University Press, 1941. [See
Chapters 4, 5, and 6 of Volume 2.]
BIBLIOGRAPHY 161
- Algebra
[32] Herstein, I. N., Abstract Algebra, third edition, John Wiley & Sons, Inc., 1999. [See our
Chapter 6 of Volume 1.]
[33] Dummit, D. S., and Foote, R. M., Abstract Algebra, third edition, John Wiley & Sons, Inc.,
2004. [See our Chapter 6 of Volume 1.]
[34] Hardy, G. H., and Wright, E. M., An Introduction to the Theory of Numbers, fourth
edition, Oxford University Press, 1960. [See our Chapter 5 of Volume 1.]
- Analysis
[35] Apostol, T. M., Mathematical Analysis , Addison-Wesley, 1957. [Good background for Rie-
mannian metrics in Chapter 1 of Volume 1, and also the chapters of Volume 3.]
[36] Lang, S., Real and Functional Analysis, third edition, Springer, 1993. [Chapter XIV gives the
differentiable version of the open mapping theorem. The proof uses the contraction mapping
principle. See our Chapter 12 of Volume 2 for the topological version of the open mapping
theorem.]
[37] Spivak, M., Calculus on Manifolds, W. A. Benjamin, Inc., New York, N. Y., 1965. [Good
background for Riemannian metrics in Chapter 1 of Volume 1 and for Volume 3.]
[38] Jänich, K., Vector Analysis, translated by Leslie Kay, Springer, 2001. [Good background for
Riemannian metrics in Chapter 1 of Volume 1 and for Volume 3.]
[39] Saks, S., Theory of the Integral, second revised edition, translated by L. C. Young, Dover
Publications, Inc., New York, 1964. [Wonderfully readable.]
[40] H. L. Royden, H. L., Real Analysis, third edition, Macmillan, 1988. [The place where we
first learned about nonmeasurable sets.]
References from our Paper on Hyperbolic Geometry in Flavors of

Geometry (reprinted here as our Volume 3, Chapter 2)
[41] Flavors of Geometry, edited by Silvio Levy, Cambridge University Press, 1997.
[42] Alonso, J. M., Brady, T., Cooper, D., Ferlini, V., Lustig, M., Mihalik, M., Shapiro, M.,
Short, H., Notes on word hyperbolic groups, Group Theory from a Geometrical Viewpoint:
21 March — 6 April 1990, ICTP, Trieste, Italy, (E. Ghys, A. Haefliger, and A. Verjovsky,
eds.), World Scientific, Singapore, 1991, pp. 3–63.
[43] Benedetti, R., and Petronio, C., Lectures on Hyperbolic Geometry, Universitext, Springer-
Verlag, Berlin, 1992. [Expounds many of the facts about hyperbolic geometry outlined in
Thurston’s influential notes.]
[44] Bolyai, W., and Bolyai, J., Geometrische Untersuchungen, B. G. Teubner, Leipzig and
Berlin, 1913. (reprinted by Johnson Reprint Corp., New York and London, 1972) [Historical
and biographical materials.]
[45] Cannon, J. W., The combinatorial structure of cocompact discrete hyperbolic groups, Geom.
Dedicata 16 (1984), 123–148.
[46] Cannon, J. W., The theory of negatively curved spaces and groups, Ergodic Theory, Symbolic
Dynamics and Hyperbolic Spaces, (T. Bedford, M. Keane, and C. Series, eds.) Oxford
University Press, Oxford and New York, 1991, pp. 315–369.
[47] Cannon, J. W., The combinatorial Riemann mapping theorem, Acta Mathematica 173
(1994), 155–234.
[48] Cannon, J. W., Floyd, W. J., Parry, W. R., Squaring rectangles: the finite Riemann mapping
theorem, The Mathematical Heritage of Wilhelm Magnus — Groups, Geometry & Special
162 BIBLIOGRAPHY
Functions, Contemporary Mathematics 169, American Mathematics Society, Providence,

1994, pp. 133–212.
[49] Cannon, J. W., Floyd, W. J., Parry, W. R., Sufficiently rich families of planar rings,
preprint.
[50] Cannon, J. W., Swenson, E. L., Recognizing constant curvature groups in dimension 3,
preprint.
[51] Coornaert, M., Delzant, T., Papadopoulos, A., Geometrie et theorie des groupes: les groupes
hyperboliques de Gromov, Lecture Notes 1441, Springer-Verlag, Berlin-Heidelberg-NewYork,
1990.
[52] Euclid, The Thirteen Books of Euclid’s Elements, Vol. 1-3, 2nd Ed., (T. L. Heath, ed.)
Cambridge University Press, Cambridge, 1926 (reprinted by Dover, New York, 1956).
[53] Gabai, D., Homotopy hyperbolic 3-manifolds are virtually hyperbolic, J. Amer. Math. Soc.
7 (1994), 193–198.
[54] Gabai, D., On the geometric and topological rigidity of hyperbolic 3-manifolds, Bull. Amer.
Math. Soc. 31 (1994), 228–232.
[55] Ghys, E., de la Harpe, P., Sur les groupes hyperboliques d’après Mikhael Gromov, Progress
in Mathematics 83, Birkhäuser, Boston, 1990.
[56] Gromov, M., Hyperbolic groups, Essays in Group Theory, (S. Gersten, ed.), MSRI Publi-
cation 8, Springer-Verlag, New York, 1987. [Perhaps the most influential recent paper in
geometric group theory.]
[57] Hilbert, D., Cohn-Vossen, S., Geometry and the Imagination, Chelsea Publishing Company,
New York, 1952. [College level exposition of rich ideas from low-dimensional geometry with
many figures.]
[58] Iversen, B., Hyperbolic Geometry, London Mathematical Society Student Texts 25, Cam-
bridge University Press, Cambridge, 1993. [Very clean algebraic approach to hyperbolic ge-
ometry.]
[59] Klein, F., Vorlesungen über Nicht-Euklidische Geometrie, Verlag von Julius Springer, Berlin,
1928. [Mostly algebraic development of non-Euclidean geometry with respect to Klein and
projective models. Beautiful figures. Elegant exposition.]
[60] Kline, M. Mathematical Thought from Ancient to Modern Times, Oxford University Press,
New York, 1972. [A 3-volume history of mathematics. Full of interesting material.]
[61] Lobatschefskij, N. I., Zwei Geometrische Abhandlungen, B. G. Teubner, Leipzig and Berlin,
1898. (reprinted by Johnson Reprint Corp., New York and London, 1972) [Original papers.]
[62] Mosher, L., Geometry of cubulated 3-manifolds, Topology 34 (1995), 789–814.
[63] Mosher, L., Oertel, U., Spaces which are not negatively curved, preprint.
[64] Mostow, G. D., Strong Rigidity of Locally Symmetric Spaces, Annals of Mathematics Studies
78, Princeton University Press, Princeton, 1973.
[65] Poincaré, H., Science and Method, Dover Publications, New York, 1952. [One of Poincaré’s
several popular expositions of science. Still worth reading after almost 100 years.]
[66] Ratcliffe, J. G., Foundations of Hyperbolic Manifolds, Graduate Texts in Mathematics 149,
Springer-Verlag, New York, 1994. [Fantastic bibliography, careful and unified exposition.]
[67] Riemann, B., Collected Papers, Kendrick Press, Heber City, Utah, 2004. [English translation
of Riemann’s wonderful papers]
[68] Swenson, E. L., Negatively curved groups and related topics, Ph.D. dissertation, Brigham
Young University, 1993.
[69] Thurston, W. P., The Geometry and Topology of 3-Manifolds, lecture notes, Princeton Uni-
versity, Princeton, 1979. [Reintroduced hyperbolic geometry to the topologist. Very exciting
and difficult.]
[70] Weyl, H., Space—Time—Matter, Dover, New York, 1922. [Weyl’s exposition and develop-
ment of relativity and gauge theory which begins at the beginning with motivation, philoso-
phy, and elementary developments as well as advanced theory.]
BIBLIOGRAPHY 163
Further Technical References (arranged by chapter)
For the entirety of Volume 2
[71] Newman, M. H. A., Elements of the Topology of Plane Sets of Points, Cambridge University
Press, 1939.[A good alternative introduction to the topology of the plane.]
- Volume 1, Chapter 1
[72] Feynman, R., The Character of Physical Law, The M.I.T. Press, 1989, p. 47.[All of
Feynman’s writing is fun and thought provoking.]
[73] Gilbert, W. J., and Vanstone, S. A., An Introduction to Mathematical Thinking, Pearson
Prentice Hall, 2005. [The place where I learned the algorithmic calculations about the
Euclidean algorithm. See our Chapter 2.]
[74] Reid, C., Hilbert, Springer Verlag, 1970. [A wonderful biography of Hilbert, with an extended
discussion of the Hilbert address in which he stated the Hilbert problems. See our Chapter
4.]
[75] Apostol , T. M., Calculus , Volume 1, Blaisdell Publishing Company, New York, 1961. [The
place where I first learned areas by counting. See our Chapter 5.]
[76] Hilton, P., and Pedersen, J., Approximating any regular polygon by folding paper, Math.
Mag. 56 (1983), 141-155. [Method for approximating many angles algorithmically by paper-
folding.]
[77] Hilton, P., and Pedersen, J., Folding regular star polygons and number theory Math. Intel-
ligencer 7 (1985), 15-26. [More paper-folding.]
[78] Burkard Polster, Variations on a Theme in Paper Folding, Amer. Math. Monthly 111 (2004),
39-47. [More paper-folding approximations to angles. See Chapter 6 and the impossibility of
trisecting an angle.]
164 BIBLIOGRAPHY
[79] Wagon, S., The Banach-Tarski Paradox, Cambridge University Press, 1994.[A wonderful
exposition of the Hausdorff-Banach-Tarski paradox, without the emphasis on the graph of
the free group. See our Chapter 7.]
- Volume 1, Chapter 7; Volume 3, Chapter 1.
[80] Coxeter, H. S. M., and Moser, W. O., Generators and Relations for Discrete Groups, second
edition, Springer-Verlag, 1964. [The place where I learned that groups can be viewed as graphs
(the Cayley graph or the Dehn Gruppenbild). See our Chapter 7 where we use the graph of
the free group on two generators and Chapter 25 where we use graphs as approximations
to non Euclidean geometry.]
[81] Peano, G. , Sur une courbe, qui remplit toute une aire plane, Mathematische Annalen 36
(1), 1890, pp. 157-160. [The first space-filling curve, described algebraically. See our Chapter
12.]
[82] Peano, G., Selected works of Giuseppe Peano, edited by Kennedy, Hubert C., and translated.
With a biographical sketch and bibliography, Allen & Unwin, London, 1973.
[83] Hilbert, D., Über die stetige Abbildung einer Line auf ein Flächenstück, Mathematische
Annalen 38 (3), 1891, pp. 459-460. [Hilbert gave the first pictures of a space-filling curve.
See our Chapter 12.]
[84] G. Pólya, Über eine Peanosche Kurve, Bull. Acad. Sci. Cracovie, A, 1913, pp. 305-313.
[Pólya’s triangle-filling curve. See our Chapter 12.]
[85] Lax, P. D., The differentiability of Pólya’s function, Adv. Math., 10, 1973, pp. 456-464.
[Lax recommends the non-isosceles triangle in Pólya’s construction since it simplifies the
description of the path followed to the point represented by a binary expansion. See our
Chapter 12.]
[86] Mandelbrot, B., The Fractal Geometry of Nature, W. H . Freeman & Co, 1982. [Mandelbrot
suggests the use of Hausdorff dimension as a means of recognizing sets that are locally
complicated or chaotic. He defines these to be fractals. See our Chapter 13.]
[87] Falconer, K. J., The Geometry of Fractal Sets, Cambridge University Press, 1985. [See
reference [86] and our Chapter 13.]
[88] Devaney, R. L., Differential Equations, Dynamical Systems, and an Introduction to Chaos
with Morris Hirsch and Stephen Smale, 2nd edition, Academic Press, 2004; 3rd edition,
Academic Press, 2013. [See reference [84] and our Chapter 13.]
- Volume 2, Chapter 8 and 11
[89] Moore, R. L., Concerning upper semi-continuous collections of continua, Trans. Amer.
Math. Sc. 27 (1925), pp. 416-428. [Moore shows that his topological characterization of the
plane or 2-sphere allows him to prove his theorem about decompositions of the 2-sphere. See
our Volume 2, Chapters 8 and 11.]
BIBLIOGRAPHY 165
[90] Wilder, R. L., Topology of Manifolds, American Mathematical Society, 1949 . [Our proof of
the topological characterization of the sphere is primarily modelled on Wilder’s proof, with
what we consider to be conceptual simplifications. See our Chapter 8.]
[91] Radó, T., Über den Begriff der Riemannschen Fläche, Acts. Litt. Sci. Szeged 2 (1925), pp.
101-121. [The first proof that 2-manifolds can be triangulated. See our Chapter 20.]
[92] Andrews, Peter, The classification of surfaces, Amer. Math. Monthly 95 (1988), 861-867l
[93] Armstrong, M. A., Basic Topology, McGraw-Hill, London, 1979.
[94] Burgess, C. E., Classification of surfaces, Amer. Math. Monthly 92 (1985), 349-354.
[95] Francis, George K., Weeks, Jeffrey R., Conway’s ZIP proof, Amer. Math. Monthly 106
(1999), 393-399.
[96] Rolfsen, D., Knots and Links, AMS Chelsea, vol 346, 2003. [See our Chapter 22.]
For the entirety of Volume 3, see the references above taken from our article in
Flavors of Geometry, beginning with reference [41].
[97] Misner, C. W., and Thorne, K. S., and Wheeler, J. A., Gravitation, W. H. Freeman and
Company, 1973.
- Volume 3, Chapters 4 and 5
[98] Abelson, H., and diSessa, A., Turtle Geometry, MIT Press, 1986. [The authors use the paths
of a computer turtle to model straight paths on a curved surface.]
This is the second of a three volume collection devoted to the geometry,
topology, and curvature of 2-dimensional spaces. The collection provides
a guided tour through a wide range of topics by one of the twentieth
century’s masters of geometric topology. The books are accessible to
college and graduate students and provide perspective and insight to
mathematicians at all levels who are interested in geometry and topology.
The second volume deals with the topology of 2-dimensional spaces. The
attempts encountered in Volume 1 to understand length and area in the
plane lead to examples most easily described by the methods of topology
(fluid geometry): finite curves of infinite length, 1-dimensional curves of positive area, space-
filling curves (Peano curves), 0-dimensional subsets of the plane through which no straight
path can pass (Cantor sets), etc. Volume 2 describes such sets. All of the standard topological
results about 2-dimensional spaces are then proved, such as the Fundamental Theorem of
Algebra (two proofs), the No Retraction Theorem, the Brouwer Fixed Point Theorem, the
Jordan Curve Theorem, the Open Mapping Theorem, the Riemann-Hurwitz Theorem, and
the Classification Theorem for Compact 2-manifolds. Volume 2 also includes a number of
theorems usually assumed without proof since their proofs are not readily available, for example,
the Zippin Characterization Theorem for 2-dimensional spaces that are locally Euclidean, the
Schoenflies Theorem characterizing the disk, the Triangulation Theorem for 2-manifolds, and
the R. L. Moore’s Decomposition Theorem so useful in understanding fractal sets.
For additional information

and updates on this book, visit
www.ams.org/bookpages/mbk-109
AMS on the Web

www.ams.org
MBK/109

Topology As Fluid Geometry Two Dimensional Spaces Volume 2 PDF

Uploaded by

Copyright:

Available Formats

Topology As Fluid Geometry Two Dimensional Spaces Volume 2 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Topology As Fluid Geometry Two Dimensional Spaces Volume 2 PDF

Uploaded by

Copyright:

Available Formats

Two-Dimensional Spaces, Volume 2

AMERICAN MATHEMATICAL SOCIETY

For additional information and updates on this book, visit

Library of Congress Cataloging-in-Publication Data

Preface to the Three Volume Set ix

Preface to Volume 2 xiii

Chapter 1. The Fundamental Theorem of Algebra 1

Chapter 2. The Brouwer Fixed Point Theorem 11

Chapter 4. Lebesgue Covering Dimension 37

Chapter 5. Fat Curves and Peano Curves 45

6.4. Cantor Sets in the Plane Are Tame 69

Chapter 7. Algebraic Topology 75

Chapter 8. Characterization of the 2-Sphere 81

Chapter 10. Arcs in S2 Are Tame 95

Chapter 11. R. L. Moore’s Decomposition Theorem 101

Chapter 12. The Open Mapping Theorem 109

Chapter 13. Triangulation of 2-Manifolds 113

Chapter 14. Structure and Classiﬁcation of 2-Manifolds 117

Chapter 15. The Torus 129

15.2. The Torus as a Euclidean Surface 131

Geometry measures space (geo = earth, metry = measurement). Einstein’s

The Fundamental Theorem of Algebra

1.1. Complex Arithmetic

Figure 1. Addition of complex numbers

(r cos α, r sin α) = r(cos α, sin α)

Figure 2. Polar coordinates

Figure 3. The projection principle

Figure 4. The addition formulas

If we add to these two formulas the Pythagorean Theorem,

1.2. First Proof of the Fundamental Theorem

Lemma 1.4. If p(z) is a nonconstant complex polynomial, then limz→∞ p(z) =

n−1 + · · · + a1 z + a0 , with an = 0 and |z| > 1, then

If λ is chosen so small that λ(|bk+1 dk+1 | + · · · + |bn dn |) < 1, we conclude that

1.3. Second Proof

Theorem 1.10. If q(z) = z n and p(z) = z n + an−1 z n−1 + · · · + a1 z + a0 , then,

|F (s, t)| = |(1 − t)sn + t(sn + an−1 sn−1 + · · · + a1 s + a0 )|

Figure 5. Dividing a ring into layers of triangles

Figure 6. Peeling away the layers

This argument completes our second proof of the Fundamental Theorem of

Figure 7. Calculating intersection numbers

The Brouwer Fixed Point Theorem

2.1. Statement of the Theorem

Bn = {x = (x1 , x2 , . . . , xn ) ∈ Rn : |x|2 = x21 + x22 + · · · + x2n ≤ 1}.

2.2. Introducing Extra Structure into a Problem

2.3. Two Elementary Problems

Figure 1. The uncolored board with two corners deleted

Figure 2. The uncolored domino

2.3.1. Solution to the Tiling Problem. First approximation to a solution

Figure 3. The 15 Puzzle

Figure 4. Tile slides

2.3.2. Solution to the 15 Puzzle. Solution to the 15 puzzle: As a matter

Figure 5. The 15 puzzle with numbers reversed

Figure 6. The colored board with two corners omitted

Figure 7. The colored domino

sign or parity of a permutation — is a permutation even (plus sign) or odd (minus

Figure 8. The inversions of a permutation

Figure 9. The reversed position again

Figure 10. The initial position again

2.4. Three Advanced Problems

n−1 + · · · + a1 z + a0 , with an = 0 and |z| > 1, then