D IFFERENTIALTOPOLOGY

L ECTURE NOTES

G EREON Q UICK
L AST UPDATE: 02 FEBRUARY 2024
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.1 Topology in ℝ𝑛 14

2.2 Smooth maps 24

2.3 Smooth manifolds 28

2.4 Tangent spaces and the derivative 38

2.5 Tangent Bundle 51

2.6 Exercises and more examples 55

3 The Inverse Function Theorem, immersions and embeddings . . . . . . 58

3.1 The Inverse Function Theorem and local diffeomorphisms 58

3.2 Immersions and embeddings 63

3.3 Embeddings 69

3.4 Exercises and more examples 73

4 Submersions and regular values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.1 Submersions 75

4.2 Regular values 78

4.3 The Stack of Records Theorem 84

4.4 Milnor’s proof of the Fundamental Theorem of Algebra 87

3
4.5 Exercises and more examples 93

5 A brief excursion to Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5.1 Lie groups - the definition 96

5.2 Examples of Lie groups 100

5.3 Topology of Lie groups 102

5.4 Lie subgroups 104

6 Transversality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

6.1 Transversality and preimages 109

6.2 Transverse intersections 113

6.3 Exercises and more examples 122

7 Sard’s theorem and Morse functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

7.1 The Theorem of Brown and Sard 124

7.2 Morse Functions 129

8 Smooth Homotopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

8.1 Homotopies 138

8.2 Smooth homotopies and bump functions 140

8.3 Stable properties 143

8.4 Exercises and more examples 149

9 Abstract Smooth Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

9.1 Abstract manifolds - the definition 151

9.2 Real projective space 155

9.3 Further examples 158

9.4 Grassmannian 162

9.5 Embedding Manifolds in Euclidean Space 169

9.6 Exercises and more examples 183

10 Manifolds with Boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

10.1 Motivation: A first glimpse at intersection theory 186

10.2 Manifolds with Boundary 187

10.3 One-Manifolds 197

10.4 Brouwer Fixed Point Theorem 201

10.5 Exercises and more examples 208

11 The Brouwer Degree modulo 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

11.1 The Brouwer Degree of maps modulo 2 211

11.2 Winding Numbers and the Borsuk–Ulam Theorem 219

11.3 Linking numbers and the Hopf invariant modulo 2 232

11.4 Exercises and more examples 240

12 Tubular Neighborhoods and Transversality . . . . . . . . . . . . . . . . . . . . . . 242

12.1 Normal bundles and tubular neighborhoods 242

12.2 Whitney Approximation Theorem 252

12.3 Ehresmann’s Fibration Theorem 256

12.4 Thom’s Transversality Theorem 257

12.5 Exercises and more examples 266

13 Intersection Theory modulo 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

13.1 Intersection Numbers modulo 2 268

13.2 Intersection of manifolds and self-intersections 272

13.3 Obstruction to extensions to boundaries 273

13.4 Intersecting circles on 𝕊𝑛 versus ℝP𝑛 274

13.5 ℝ𝑛 as a commutative division algebra 279

13.6 Exercises and more examples 281

14 Orientation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284

14.1 Towards integer-valued invariants 284

14.2 Orientation 286

14.3 Induced orientation on the boundary 291

14.4 Oriented Homotopy 293

14.5 Orientation of transverse preimage 294

14.6 Example: Fibers of the Hopf fibration 296

14.7 Orientation on boundary of preimage 298

14.8 Example: Simply-connected manifolds are orientable 300

14.9 Summary 302

14.10 Exercises and more examples 304

15 The Brouwer Degree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306

15.1 The integral-valued Brouwer Degree 306

15.2 Hopf Degree Theorem in dimension one 313

15.3 Exercises and more examples 317

15.4 Linking number and the Hopf invariant via exercises 320
16 Hopf Degree Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323

16.1 Strategy for proving Hopf’s theorem 323

16.2 The Special Case 328

16.3 The Extension Theorem for degree zero maps 331

16.4 The proof of Hopf’s theorem 332

17 Vector Fields and the Poincaré–Hopf Index Theorem . . . . . . . . . . . . 333

17.1 Vector Fields 333

17.2 Index of a vector field at a zero 334

17.3 The Euler characteristic - algebraic topology in a nutshell 340

17.4 The Poincaré–Hopf Index Theorem 344

17.5 Poincaré–Hopf Theorem - Independence 346

17.6 Poincaré–Hopf Theorem - Euler characterstic 352

17.7 Existence of vector fields with no zeros 355

A Solutions to exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357

A.2 Smooth manifolds 357

A.3 The Inverse Function Theorem, immersions and embeddings 369

A.4 Submersions and regular values 373

A.5 Lie groups 380

A.6 Transversality 386

A.7 Smooth Homotopy 391

A.8 Abstract Smooth Manifolds 395

A.9 Manifolds with Boundary 401

A.10 Brouwer Degree mod 2 and Borsuk–Ulam Theorem 406

 8
A.11 Transversality 411

A.12 Intersection Theory modulo 2 416

A.13 Orientation 423

A.14 The Brouwer Degree 429

A.15 Linking Number and the Hopf Invariant 435

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
Preface

These are the lecture notes accompanying a one-semester course. The aim of the course is to
pick up the readers with a background in multi-variable calculus and linear algebra and to take
them on a direct path to the fascinating world of differential topology.

In order to make the ideas and techniques as accessible as possible the arguments are ex-
plained in great detail. Thus instead of aiming for the most elegant and shortest argument we
often take a longer walk and pick up every flower along the way by hand. We hope that the
amount of detail will make it easier for a relatively unexperienced mathematician to witness
and understand what is happening and to appreciate how some relatively straight-forward ideas
lead to exciting and deep results.

This has the consequence that these notes are not brief and do not just summarise the main
ideas and theorems in differential topology. For a brief and comprehensive account we recom-
mend the excellent books by Milnor [13] and Guillemin–Pollack [5] on which these notes are
based on.

The chapters are accompanied by a list of exercises. We highly recommend to work through
all the exercises. At the end of the book there are suggestions for solutions to all exercises. For
some it may be tempting to glimpse at the solutions before trying to solve the problems, other
will rather try on their own first. We think that the readers may decide for themselves how they
prefer to learn new mathematics.

We are grateful to many colleagues and students who helped to improve these notes. In
particular, we would like to thank Torgeir Aambø, Johanna Aigner, Anders Alexander Ander-
sen, Torgeir Aambø, Daria Barjaktarevic, Denis Bergmann, Erlend Bergtun, Dagmar Coelle,
Eivind Xu Djurhuus, Robin Fissum, Sigurd Gaukstad, Knut Bjarte Haus, Håvard Skjetne Lille-
heie, Abigail Linton, Martin Löcsei, Chileshe Mwamba, Trygve Poppe Oldervoll, Leona Ro-
denkirchen, Sindre Strøm, Marius Thaule, Inga Maria Tillmann, Melvin Vaupel, and Glen
Matthew Wilson.

9
1. Introduction

∙ Geometry vs topology

Classical geometers were interested in measuring angles and distances. For example, two
things are the ‘same’ – more mathematically speaking they are congruent – in classical geom-
etry1 if you can transform one into the other by moving or rotating them. A first variation to
allow flexibility, is projective geometry: Two things are considered the same if they are both
views of the same object. For example, an ellipse and a circle can be projectively equivalent:
for one can look like the other when you look at them from the right perspective. In topology,
we take this idea a big step further and consider two objects the same if we can continuously
transform one into the other. For example, a triangle is equivalent to a circle and both are
equivalent to a square. In differential topology, the part we study in this class, we only allow
smooth transformations. Then square and circle are different, because a square has vertices
which are not smooth points while the circle does not. As a slogan we may say that differential
topology is the study of properties that do not change under smooth transformations. Smooth-
ness of an object is something we check locally, i.e., by looking at every point and a small
neighborhood around it. The smoothest space we know is Euclidean space. This leads to the
following first working definition of what kind of objects we are going to study:

Definition 1.1 (Working definition: What is a manifold?) A manifold is a geometric

object which locally, i.e., in a small neighborhood of every point, looks like ℝ𝑛 .

In order to get a first idea, let us look at a fundamental example:

Example 1.2 (Spheres) Let us look at the unit circle

𝕊1 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 + 𝑦2 = 1} ⊂ ℝ2 .

The circle is something one-dimensional. But how do we describe that precisely when
we need two coordinates to describe its points. One way is that if we zoom in at any
point it just looks like a bended line segment. Looking very closely it even looks almost
like a straight line segment. See Figure 1.1. So, locally, and we will make precise what
that means very soon, the circle looks like a segment of ℝ1 . The unit circle 𝕊1 , more
generally, the 𝑛-dimensional sphere

𝕊𝑛 = {(𝑥1 , … , 𝑥𝑛+1 ) ∈ ℝ𝑛+1 ∶ 𝑥21 + ⋯ + 𝑥2𝑛+1 = 1}

is an example of a smooth manifold.

1
A word on colors: We make extensive use of colors in the text. Sometimes NTNU orange is used for defi-
nitions, while NTNU blue is used for key words. The use of the latter is quite frequent. The reason is that these
notes are also used as actual printed notes during lectures where it is desirable to spot the key words of an argument
easily. This might be a distraction for the reader for which we apologize in advance.

10
Chapter 1. Introduction 11

Figure 1.1: Each segment of the circle looks like an open interval in ℝ.

While not being precise at all, our working definition may sound at the same time quite strict.
In a small neighborhood every point looks the same. How can this lead to interesting objects?
We will see that there is in fact a universe of examples of smooth manifolds of very different
kind. The point is that even though all points look pretty much the same locally, manifolds may
look very different globally. As a first simple example consider the two-dimensional sphere
and the two-dimensional torus. They look the same locally, but they are quite different globally:
the torus has a whole in the middle while the sphere does not. In fact, one of the main goals in
topology is to classify all types of manifolds.

∙ Important idea: introduce invariants

A key method to analyse spaces and maps, or more generally any kind of complicated object
in mathematics, is to attach to them interesting invariants. These are usually numbers, groups,
vector spaces, …, any sort of algebraic objects which are much easier to understand and to
distinguish than the spaces and maps we started with. The name invariant comes from the
fact that we require that the algebraic objects, for example numbers 𝑖(𝑋) and 𝑗(𝑓 ) we attach to
each object 𝑋 and each map 𝑓 , does not change under the geometric transformations we allow.
For smooth manifolds, the transformations we allow are diffeomorphisms, or, later on, smooth
homotopy equivalences. For example, assume we would like to understand two manifolds 𝑋
and 𝑌 . Let us assume they are defined in complicated ways so that it is not easy at all to check
if they are maybe the same objects after all. If we can calculate the invariants 𝑖(𝑋) and 𝑖(𝑌 ) and
get 𝑖(𝑋) ≠ 𝑖(𝑌 ), then 𝑋 and 𝑌 could not have been the same to begin with. If 𝑋 and 𝑌 were
vector spaces, then we are very familiar with the idea of an invariant, namely the dimension.
That is, if dim 𝑋 ≠ dim 𝑌 , then we know that there is no linear isomorphism between 𝑋 and
𝑌 . If dim 𝑋 = dim 𝑌 , then we can construct an isomorphism 𝑋 ≅ 𝑌 by choosing bases for
both 𝑋 and 𝑌 . In fact, a smooth manifold has a tangent space at every point. The tangent
space is a real vector space, and it will be one of the main tools for our study. However, tangent
spaces and the dimension are not sufficient to study smooth manifolds and we will need more
sophisticated tools and invariants. For example, we will develop intersection theory and will
show that the smooth manifolds 𝕊𝑛 , the 𝑛-sphere, and ℝP𝑛 , the 𝑛-dimensional real projective
space, are not diffeomorphic for 𝑛 ≥ 2. In fact, we will show that they are not even homotopy
equivalent for 𝑛 ≥ 2.
12

∙ Here are some examples of invariants in differential topology:

∙ dimension

∙ Brouwer degree

∙ intersection number

∙ Euler characteristic

∙ index of a vector field

∙ Some important theorems

After developing the theory of smooth manifolds with and without boundaries we will prove
the following famous theorem: Let 𝔻𝑛 be the 𝑛-dimensional unit disc
𝔻𝑛 = {(𝑥1 , … , 𝑥𝑛 ) ∈ ℝ𝑛 ∶ 𝑥21 + ⋯ + 𝑥2𝑛 ≤ 1}.

Theorem 1.3 (Brouwer Fixed Point Theorem) Every continuous map 𝑓 ∶ 𝔻𝑛 → 𝔻𝑛

has a fixed point, i.e., there is an 𝑥 ∈ 𝔻𝑛 such that 𝑓 (𝑥) = 𝑥.

This theorem has a lot of very important applications. In particular, one way to apply it is
to first transform a problem into finding a solution 𝑥0 of an equation of the form 𝑓 (𝑥) = 𝑥 and
to use Brouwer’s theorem to show that such a solution exists. It will require a good amount of
exciting work to prove the theorem.

Then we will introduce the most important invariant in differential topology: the Brouwer
degree of a smooth map. It will turn out to be an extremely powerful tool. We will first introduce
a mod 2-version of the degree and use it for example to show the following important and deep
result of Hopf:

Theorem 1.4 (A Hopf invariant one primer) There is a smooth map 𝕊3 → 𝕊2 which
is not homotopic to a constant map.

The map 𝕊3 → 𝕊2 of the theorem is an example of Hopf fibration which only exists in
certain dimensions. We will use this important map as a running example throughout these
notes. The way we talked about invariants so far may suggest that they only allow to show that
something does not exist. However, there are also some very important situations where an
equality of invariants implies that spaces are equivalent. A very famous example of such a case
is the following theorem which uses the integer-valued Brouwer degree which we will study in
this course after introducing orientations:

Theorem 1.5 (Hopf Degree Theorem) Let 𝑋 be a compact, connected, oriented

smooth 𝑘-manifold without boundary. Then two continuous maps 𝑋 → 𝕊𝑘 are homo-
topic if and only if they have the same degree.
Chapter 1. Introduction 13
As a generalization of the degree we will develop mod 2-intersection theory on smooth
manifolds. Surprisingly, we can use it to prove the following purely algebraic result: Let ℝ𝑛 ×
ℝ𝑛 → ℝ𝑛 , (𝑎, 𝑏) → 𝑎 ⋅ 𝑏, be an ℝ-bilinear map with no zero-divisors, i.e., 𝑎 ⋅ 𝑏 implies 𝑎 = 0
or 𝑏 = 0. Assume that we have 𝑎 ⋅ 𝑏 = 𝑏 ⋅ 𝑎 for all 𝑎, 𝑏 ∈ ℝ𝑛 . Such a map is called a
commutative division algebra structure on ℝ𝑛 . We are very familiar with commutative division
algebra structures on ℝ𝑛 for 𝑛 = 1 and 𝑛 = 2: They are given by the field of real numbers ℝ
for 𝑛 = 1 and the field of complex numbers ℂ ≅ ℝ2 for 𝑛 = 2. We will show that there are no
other possible cases:

Theorem 1.6 There is no commutative division algebra structure on ℝ𝑛 for 𝑛 > 2.

A vector field on a smooth manifold 𝑋 is a map which assigns to each point 𝑥 a tangent
vector to 𝑋 at 𝑥. A zero of a vector field is a point to which the field assigns the zero vector.
The Brouwer degree allows us to define the index of a zero which is an integer that characterises
the geometry of the vector field around the zero. The sum of the indices of the zeros of a vector
field a priori depends on the smooth structure of the manifold. On the other hand, the manifold
𝑋 is also equipped with an integer, the Euler characteristic 𝜒(𝑋). This is a purely topological
invariant which means it does not change if we transform 𝑋 continuously and it does not depend
on the fact that 𝑋 is not just a space but a smooth manifold. One of the highlights of this course
is the proof of the following famous result which relates the geometry of vector fields on a
smooth manifold to its Euler characteristic. This is a first example of an index theorem which
is part of a fascinating and very influential area in mathematics:

Theorem 1.7 (Poincaré–Hopf Index Theorem) Let 𝑋 be a compact, oriented smooth

manifold and let 𝐯 be a vector field on 𝑋 with only finitely many zeros. Then the sum
of the indices of 𝐯 equals the Euler characteristic of 𝑋.

Before we are able to prove these exciting results we set out to develop the basic theory of
smooth manifolds in the following chapters.
2. Smooth manifolds

2.1 Topology in ℝ𝑛

2.1.1 Open and closed subsets in Euclidean space

Recall from Calculus that the norm of a vector 𝑥 = (𝑥1 , … , 𝑥𝑛 ) ∈ ℝ𝑛 is defined as the non-
negative real number √
|𝑥| = 𝑥21 + 𝑥22 + ⋯ + 𝑥2𝑛 ∈ ℝ≥0 .
The norm defines a distance between two points 𝑥, 𝑦 in ℝ𝑛 by taking the norm |𝑥 − 𝑦| of the
difference of 𝑥 and 𝑦. For any 𝑛, the space ℝ𝑛 together with this norm is called 𝑛-dimensional
Euclidean space. It is a topological space in the following way:

∙ (Open sets in ℝ𝑛 ) Here are important examples of open subsets in ℝ𝑛 :

∙ Let 𝑥 be a point in ℝ𝑛 and 𝑟 > 0 a real number. We define the 𝑛-dimensional open
ball with radius 𝑟 around 𝑥

𝔹𝑟 (𝑥) = 𝔹𝑛𝑟 (𝑥) = {𝑦 ∈ ℝ𝑛 ∶ |𝑥 − 𝑦| < 𝑟}.

Note that we drop the superscript 𝑛 whenever possible.

∙ The open balls 𝔹𝑟 (𝑥) are the prototypes of open sets in ℝ𝑛 .

∙ A non-empty subset 𝑈 ⊆ ℝ𝑛 is called open if for every point 𝑥 ∈ 𝑈 there exists

a real number 𝜀 > 0 such that 𝔹𝜀 (𝑥) is contained in 𝑈 .

∙ The empty set ∅ are defined to be open.

∙ The whole space ℝ𝑛 is open.

∙ A subset 𝑍 ⊆ ℝ𝑛 is called closed if its complement ℝ𝑛 ⧵ 𝑍 is open.

∙ Arbitrary unions of open sets are open and finite intersections of open sets are
open.

Examples and remarks:

∙ Familiar examples of open sets in ℝ are open intervals, e.g., the open interval (−2, 1).
∙ The cartesian product of 𝑛 open intervals (an open ‘rectangle’) is open in ℝ𝑛 .
∙ Similarly, closed intervals are examples of closed sets in ℝ, .e.g., the closed interval
[−2, 1].

14
Chapter 2. Smooth manifolds 15

Figure 2.1: An open ball in ℝ3 .

∙ An important example of a closed set is the 𝑛-dimensional sphere 𝕊𝑛 defined as

{ }
𝕊𝑛 = 𝑥 ∈ ℝ𝑛+1 ∶ |𝑥| = 1 .

∙ The cartesian product of 𝑛 closed intervals (a closed ’rectangle’) is closed in ℝ𝑛 .

∙ The empty set ∅ and ℝ𝑛 itself are by definition both open and closed sets.

∙ Not every subset of ℝ𝑛 is open or closed. There are a lot of subsets which are neither
open nor closed. For example, the interval (0, 1] in ℝ or the product of an open and a
closed interval in ℝ2 .

Figure 2.2: Examples of subsets in ℝ2 .

Definition 2.1 (Relative open sets) Let 𝑋 be a subset in ℝ𝑛 . Then we say that 𝑉 ⊆ 𝑋
is open in 𝑋 (or relatively open) if there is an open subset 𝑈 ⊆ ℝ𝑛 which is open in ℝ𝑛
with 𝑉 = 𝑈 ∩ 𝑋. More concretely: 𝑉 ⊆ 𝑋 is open in 𝑋 if and only if for every point
𝑥 ∈ 𝑉 there exists a real number 𝜀 > 0 such that 𝔹𝑛𝜀 (𝑥) ∩ 𝑋 ⊆ 𝑉 . See Figure 2.3.

In order to decide whether a subspace is open or closed it is very important to take the
ambient space into account:

Remark 2.2 (Warning) It is important to note that the property of being an open
subset very much depends on the bigger space we are looking at. Hence open always
refers to being open in some given space. For example, a set can be open in a space
𝑋 ⊂ ℝ2 , but not be open in ℝ2 , see Figure 2.4.
16 2.1. Topology in ℝ𝑛

Figure 2.3: A relative open subset of 𝑋: 𝑋 ∩ 𝑈 is open in 𝑋.

Figure 2.4: The relative open subset of 𝑋 ∩ 𝑈 is open in 𝑋, but is not open in ℝ2 .
Chapter 2. Smooth manifolds 17
Examples:

∙ Let 𝑋 = 𝕊2 be the two-dimensional sphere. We consider it as a subset in ℝ3 with the

subspace topology. Let 𝑥0 ∈ 𝕊2 be a point on 𝕊2 . An example of a subset in 𝕊2 which is
open in 𝕊2 and contains 𝑥0 is the set

{ }
1
𝕊2 ∩ 𝔹31 (𝑥0 ) with 𝔹31 (𝑥0 ) = 𝑦 ∈ ℝ3 ∶ |𝑦| < .
2 2 2

In fact, every subset which is open in 𝕊2 and contains 𝑥0 contains a subset if the form
𝕊2 ∩ 𝔹3𝜀 (𝑥0 ) with 𝜀 > 0 sufficiently small.

∙ However, the set 𝕊2 ∩ 𝔹31 (𝑥0 ) is not open in ℝ3 . For there is no three-dimensional open
2
ball 𝔹3𝜀 (𝑥0 ) which is completely contained in 𝕊2 ∩ 𝔹31 (𝑥0 ).
2

Open sets are nice for a lot of reasons. First of all, they provide us with a way to talk about
things that happen close to a point.

Definition 2.3 (Open neighborhoods) We say that a subset 𝑉 ⊆ 𝑋 containing a point

𝑥 ∈ 𝑋 is a neighborhood of 𝑥 if there is an open subset 𝑈 ⊆ 𝑉 with 𝑥 ∈ 𝑈 . If 𝑉 itself
is open, we call 𝑉 an open neighborhood.

Second, a collection of open subsets in a set 𝑋, define a topology on 𝑋:

Definition 2.4 (Spaces) We establish the following convention for the use of the word
space:

∙ A set together with a topology, i.e., a collection of open sets, is called a topological
space.

∙ From now on, when we talk about a space, we mean a topological space, i.e., a
set together with a specified topology.

Here we observe that the word topology is used in several ways. On the one hand, it is
the name of a whole area in mathematics. On the other hand, it is the name for an additional
structure on a set. We are familiar with this phenomenon: for example, the word algebra
denotes both an area in mathematics and a structure on a set.

2.1.2 Continuous maps

The type of maps that preserve open sets, i.e., respect the topology on a set, are called contin-
uous maps:
18 2.1. Topology in ℝ𝑛

Definition 2.5 (Continuous maps: abstract definition) Continuous maps are charac-
terized as follows:

∙ Let 𝑋 and 𝑌 be topological spaces. A map 𝑓 ∶ 𝑋 → 𝑌 is called continuous if,

for every open subset 𝑈 ⊆ 𝑌 , the subset 𝑓 −1 (𝑈 ) is open in 𝑋.

∙ In the subspace topology in ℝ𝑛 : If 𝐴 is a subset of ℝ𝑛 with the subset topology,

then a map 𝑓 ∶ 𝐴 → ℝ𝑚 is continuous if and only if, for every open subset
𝑈 ⊆ ℝ𝑚 , there is some open subset 𝑉 ⊆ ℝ𝑛 with 𝑓 −1 (𝑈 ) = 𝑉 ∩ 𝐴 (in other
words 𝑓 −1 (𝑈 ) is open in 𝐴).

Remark 2.6 Just in case you have heard of categories before: topological spaces form
a category with morphisms given by continuous maps.

We are familiar with continuous maps from Calculus. The 𝜀-𝛿-characterization of continuity
looks as follows:

Lemma 2.7 (Continuous maps: familiar description in ℝ𝑛 ) Let 𝐴 be a subset in ℝ𝑛

and 𝑎 ∈ 𝐴 be a point. A map 𝑓 ∶ 𝐴 → ℝ𝑚 is continuous at 𝑎 if it satisfies the following
condition: for every 𝜀 > 0, there is a 𝛿 > 0 such that

0 < |𝑥 − 𝑎| < 𝛿 and 𝑥 ∈ 𝐴 ⇒ |𝑓 (𝑥) − 𝑓 (𝑎)| < 𝜀.

In our new fancy notation, we can reformulate the last condition as follows: for every
𝜀 > 0, there is a 𝛿 > 0 such that

𝑥 ∈ 𝔹𝑛𝛿 (𝑎) ∩ 𝐴 ⇒ 𝑓 (𝑥) ∈ 𝔹𝑚

𝜀 (𝑓 (𝑎)).

Proof of Lemma 2.7: First, assume 𝑓 satisfies 𝜀-𝛿-continuity. Let 𝑈 ⊆ ℝ𝑚 be an open set
in ℝ𝑚 . If 𝑓 −1 (𝑈 ) is empty, it is open by definition. So let 𝑎 ∈ 𝑓 −1 (𝑈 ) be a point in 𝑓 −1 (𝑈 ).
The fact that 𝑈 is open means that there is an 𝜀 > 0 such that 𝔹𝑚 𝜖 (𝑓 (𝑎)) ⊆ 𝑈 . Given this 𝜀, the
fact that 𝑓 is continuous means that

there is a 𝛿 > 0 such that 𝑥 ∈ 𝔹𝑛𝛿 (𝑎) ∩ 𝐴 ⇒ 𝑓 (𝑥) ∈ 𝔹𝑚

𝜀 (𝑓 (𝑎)).

But

𝑓 (𝑥) ∈ 𝔹𝑚
𝜀 (𝑓 (𝑎)) ⊆ 𝑈

which implies 𝑓 (𝑥) ∈ 𝑈 and hence 𝑥 ∈ 𝑓 −1 (𝑈 ). Thus, for every 𝑥 ∈ 𝔹𝑛𝛿 (𝑎) ∩ 𝐴, we have
𝑥 ∈ 𝑓 −1 (𝑈 ). In other words,
𝔹𝑛𝛿 (𝑎) ∩ 𝐴 ⊆ 𝑓 −1 (𝑈 ).
Since 𝑎 was an arbitrary point in 𝑓 −1 (𝑈 ), this shows that 𝑓 −1 (𝑈 ) is open in 𝐴.

Second, assume 𝑓 −1 (𝑈 ) is open in 𝐴 for every open subset 𝑈 ⊆ ℝ𝑚 . Given 𝑎 ∈ 𝐴 and

𝜀 > 0, let 𝔹𝑚𝜀 (𝑓 (𝑎)) ⊂ ℝ be the open ball around 𝑓 (𝑎) with radius 𝜀. Since 𝔹𝜀 (𝑓 (𝑎)) is open
𝑚 𝑚

in ℝ𝑚 , our assumption tells us that 𝑓 −1 (𝔹𝑚

𝜀 (𝑓 (𝑎))) is open in 𝐴. Since 𝑎 ∈ 𝑓 (𝔹𝜀 (𝑓 (𝑎))) this
−1 𝑚
Chapter 2. Smooth manifolds 19
means that

there is a 𝛿 > 0 such that 𝔹𝑛𝛿 (𝑎) ∩ 𝐴 ⊆ 𝑓 −1 (𝔹𝑚

𝜀 (𝑓 (𝑎))).

But that means

𝑥 ∈ 𝔹𝑛𝛿 (𝑎) ∩ 𝐴 ⇒ 𝑓 (𝑥) ∈ 𝔹𝑚

𝜀 (𝑓 (𝑎)).

Hence 𝑓 is continuous at 𝑎. Since 𝑎 was arbitrary, 𝑓 is continuous.

Next we specify the maps which have inverses in the category of topological spaces:

Definition 2.8 (Homeomorphisms) Let 𝑋 and 𝑌 be topological spaces. A continuous

map 𝑓 ∶ 𝑋 → 𝑌 is a homeomorphism if it is one-to-one and onto and its inverse 𝑓 −1
is continuous as well. Homeomorphisms preserve the topology in the sense that: if
𝑓 ∶ 𝑋 → 𝑌 is a homeomorphism then 𝑈 ⊂ 𝑋 is open in 𝑋 if and only if 𝑓 (𝑈 ) ⊂ 𝑌 is
open in 𝑌 .

Examples:

∙ The map tan ∶ (−𝜋∕2, 𝜋∕2) → ℝ is a homeomorphism.

∙ The map 𝑓 ∶ ℝ → ℝ, 𝑥 → 𝑥3 is a homeomorphism.

∙ However, the map 𝑔 ∶ ℝ → ℝ, 𝑥 → 𝑥2 is not a homeomorphism, since it is neither

one-to-one nor onto.

There are also more interesting examples:

∙ Example: A bijection which is not a homeomorphism

Example 2.9 (A bijection which is not a homeomorphism) Let

𝕊1 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 + 𝑦2 = 1} ⊂ ℝ2

be the unit circle considered as a subspace of ℝ2 . Define a map

𝑓 ∶ [0, 1) → 𝕊1 , 𝑡 → (cos(2𝜋𝑡), sin(2𝜋𝑡)).

We know that 𝑓 is bijective and continuous from calculus and trigonometry. However, [ the )
function 𝑓 is not continuous. For example, the image of the open subset 𝑈 = 0, 41
−1
[ )
under 𝑓 , i.e., the subset 𝑓 (𝑈 ), is not open in 𝕊1 . Let us first remark that 𝑈 = 0, 14 is
( )
indeed open in [0, 1), since for examples 𝑈 = [0, 1) ∩ − 14 , 14 . Now we look at the point
𝑓 (0) = (1, 0) ∈ 𝕊1 . Since 0 ∈ 𝑈 , 𝑓 (0) is a point in 𝑓 (𝑈 ). However, for every open ball
𝔹2𝜀 ((1, 0)) ⊂ ℝ2 , the intersection 𝔹2𝜀 ((1, 0)) ∩ 𝕊1 contains points with strictly negative
𝑦-coordinate. In particular, 𝔹2𝜀 ((1, 0)) ∩ 𝕊1 contains points which are not in 𝑓 (𝑈 ), i.e.,

𝔹2𝜀 ((1, 0)) ∩ 𝕊1 ) ⊄ 𝑓 (𝑈 ).

20 2.1. Topology in ℝ𝑛

Alternatively, we could observe that 𝑓 −1 (𝔹2𝜀 ((1, 0)) ∩ 𝕊1 ) ⊂ [0, 1) contains points which
are close to 1 in [0, 1) and therefore do not lie in 𝑈 , i.e.,

𝑓 −1 (𝔹2𝜀 ((1, 0)) ∩ 𝕊1 ) ⊄ 𝑈 .

This shows that we cannot find an open subset 𝑉 of ℝ2 such that

𝑉 ∩ 𝕊1 = 𝑓 (𝑈 ).

Hence 𝑓 (𝑈 ) is not open in 𝕊1 .

Figure 2.5: Wrapping the interval around the circle via 𝑓 is a bijection, but not a homeomor-
phism, since 𝑓 (𝑈 ) may not be open in 𝕊1 .

2.1.3 Topological properties

One could characterize Topology as the study of properties which are preserved under homeo-
morphisms. Hence we may call a property that is preserved under homeomorphisms a topolog-
ical property. We often refer such a topological property as a global property, since we cannot
check that a space has it just by looking at small neighborhoods of all points. Many interesting
phenomena in differential topology require an interplay of local and global properties. We will
now recall some such topological properties that will play an important role for our study of
smooth manifolds. We will recall here what it means for a space to be

∙ compact,

∙ connected,

∙ path-connected.

∙ Compactness

Definition 2.10 (Compact space) A topological space 𝑋 is called compact if every

⋃
open cover {𝑈𝑖 }𝑖 of 𝑋, i.e., a collection of open subsets 𝑈𝑖 ⊂ 𝑋 such that 𝑋 = 𝑖 𝑈𝑖 is
the union of them, has a finite subcover. That is, among the {𝑈𝑖 }𝑖 it is always possible
Chapter 2. Smooth manifolds 21
to pick finitely many 𝑈𝑖1 , … , 𝑈𝑖𝑛 with

𝑍 = 𝑈𝑖1 ∪ … ∪ 𝑈𝑖𝑛 .

Recall that a subset 𝑍 ⊂ ℝ𝑛 is called bounded if there is some, possibly big, 𝑟 > 0 such
that 𝑍 ⊂ 𝐵𝑟 (0). For subspaces in Euclidean space we then have the following important char-
acterization of compact subsets.

Theorem 2.11 (Theorem of Heine–Borel) A subset 𝑍 ⊂ ℝ𝑛 is compact if and only

if it is closed and bounded.

Examples and remarks:

∙ Closed balls in ℝ𝑛 are compact.

∙ The 𝑛-dimensional sphere 𝕊𝑛 is an important example of a compact space. According to

the previous theorem we can show this by remembering that 𝕊𝑛 is a closed subset of ℝ𝑛+1
and to observe that it is bounded as it is contained in, for example, 𝔹𝑛+1
2
(0), the open ball
of radius 2 around the origin.

∙ Theorem 2.11 tells us that open subsets in ℝ𝑛 cannot be compact. For example, open
balls in ℝ𝑛 are never compact.

∙ Compactness makes a lot of things easier. On the one hand, it makes it easier to keep
track of things, as we can cover the space with finitely many open sets. On the other
hand, the previous theorem tells us that points cannot lie too far off, at least for subspaces
in ℝ𝑛 , since compact spaces in ℝ𝑛 are bounded. Hence we can think of compactness as
a general condition which helps to avoid trouble.

Lemma 2.12 (Compact and discrete implies finite) Every compact and discrete
subset 𝑆 of ℝ𝑛 is finite.

Proof: Assume 𝑆 was not finite. Compact subsets of ℝ𝑛 are bounded. Hence there is an
𝜀 > 0 such that 𝑆 is contained in the 𝑛-dimensional box with edges of length 𝜀 and center 0.
Divide this box into 2𝑛 many 𝑛-dimensional boxes of equal size. The length of their edges is
𝜀∕2. If 𝑆 was infinite there must be at least one small box which still contains infinitely many
points of 𝑆. We take this box and divide it again into 2𝑛 many 𝑛-dimensional boxes of equal
size. The length of their edges is now 𝜀∕4. Again, if 𝑆 was infinite there must be at least one of
the smaller boxes which still contains infinitely many points of 𝑆. By repeating the argument,
we see that we can find an infinite sequence of points in 𝑆 which converges. Since 𝑆 is also
closed, any convergent infinite sequence of points in 𝑆 must have a limit in 𝑆. Call this limit
𝑠. But then the subset {𝑠} would not be open in 𝑆, since every open subset of ℝ𝑛 containing 𝑠
would also contain other points of 𝑆. Hence 𝑆 would not be discrete.

∙ Connectedness
22 2.1. Topology in ℝ𝑛

Definition 2.13 (Connected spaces) Recall that a topological space 𝑋 is called con-
nected if 𝑋 cannot be written as the union of two nonempty disjoint open subsets; or
equivalently, if 𝑋 and ∅ are the only subsets which are both open and closed in 𝑋.

∙ Familiar examples of connected spaces are intervals in ℝ. For example, the closed inter-
val [0, 1] is connected.

∙ If 𝑋 is not connected, it has subsets 𝑍𝛼 ⊂ 𝑋 which are both open and closed. Each such
𝑍𝛼 is called a connected component of 𝑋. Hence 𝑋 can be considered as the possibly
infinite union of its connected components.

∙ Here is a type of argument we will meet frequently: Let 𝑋 ⊂ ℝ𝑘 and 𝑌 ⊂ ℝ𝑛 be two

topological spaces and we would like to show that there cannot be a homeomorphism
between them. First, if the numbers of connected components of 𝑋 and 𝑌 are different,
𝑓
then there cannot be a homeomorphism 𝑋 ←←←→
← 𝑌.

∙ The previous argument is often used indirectly in the following way: if 𝑋 and 𝑌 have
the same number connected components, we remove a suitable point 𝑥 ∈ 𝑋 and count
the number of connected components of 𝑋 ⧵ {𝑥} and 𝑌 ⧵ {𝑓 (𝑥)}. If these numbers are
𝑓
different, there cannot be a homeomorphism 𝑋 ←←←→
← 𝑌 . For if such an 𝑓 exists, then the
𝑓
restriction 𝑋 ⧵ {𝑥} ←←←→
← 𝑌 ⧵ {𝑓 (𝑥)} is still a homeomorphism. Continuing this or a similar
process often leads to the desired conclusion.

∙ Another frequent application of connectedness is the following: Given a map 𝑓 ∶ 𝑋 → 𝑆

from a topological space 𝑋 to any set 𝑆. Recall that 𝑓 is called locally constant if for
every 𝑥 ∈ 𝑋 there is an open neighborhood 𝑈𝑥 ⊂ 𝑋 such that 𝑓|𝑈𝑥 is constant.

Lemma 2.14 Let 𝑋 be a connected space and 𝑓 ∶ 𝑋 → 𝑆 be locally constant. Then

𝑓 is constant.

Proof: Let 𝑠 ∈ 𝑆 be a value of 𝑓 , i.e., 𝑠 = 𝑓 (𝑥) for some 𝑥 ∈ 𝑋. We can write 𝑋 as the
disjoint union of the sets

𝐴 = {𝑥 ∈ 𝑋 ∶ 𝑓 (𝑥) = 𝑠} and 𝐵 = {𝑥 ∈ 𝑋 ∶ 𝑓 (𝑥) ≠ 𝑠}.

Since 𝑓 is locally constant, both 𝐴 and 𝐵 are open. For, if 𝑎 ∈ 𝐴, then there is an open
neighborhood 𝑈𝑎 ⊂ 𝐴 with 𝑓 (𝑈𝑎 ) = {𝑠}, i.e. 𝑈𝑎 ⊂ 𝐴. Similarly, if 𝑏 ∈ 𝐵, then there is an
open neighborhood 𝑈𝑏 ⊂ 𝑋 with 𝑓 (𝑈𝑏 ) = {𝑓 (𝑏)}, i.e. 𝑈𝑏 ⊂ 𝐵. But since 𝑋 is connected and
𝐴 ≠ ∅, we must have 𝐴 = 𝑋, and 𝑓 is constant.

∙ Path-connectedness
Chapter 2. Smooth manifolds 23
The criterion for connectedness is elegant to state, but also rather abstract. For example, it
does not tell us if we can walk, i.e., draw a line without interruptions, from one point to another,
as one would intuitively expect for a connected space. This leads to a related and more concrete
property:

Definition 2.15 (Path-connected spaces) A topological space 𝑋 is called path-

connected if for any two points 𝑥, 𝑦 ∈ 𝑋 there is a continuous map 𝛾 ∶ [0, 1] → 𝑋
with 𝛾(0) = 𝑥 and 𝛾(1) = 𝑦.

Path-connectedness is the stronger property:

Lemma 2.16 (Path-connected implies connected) If a space is path-connected, then

it is also connected.

Proof: Suppose 𝑋 is path-connected. If 𝑋 was not connected, then there would be two
disjoint nonempty open subsets 𝐴 and 𝐵 with 𝑋 = 𝐴 ∪ 𝐵. Since 𝐴 and 𝐵 are nonempty, we
can choose two points 𝑎 ∈ 𝐴 and 𝑏 ∈ 𝐵. Since 𝑋 is path-connected, there is a continuous map
𝛾 ∶ [0, 1] → 𝑋 with 𝛾(0) = 𝑎 and 𝛾(1) = 𝑏. Hence 0 ∈ 𝛾 −1 (𝐴) ⊂ [0, 1] and 1 ∈ 𝛾 −1 (𝐵) ⊂ [0, 1].
Since 𝐴 and 𝐵 are disjoint and open, the subsets 𝛾 −1 (𝐴) and 𝛾 −1 (𝐵) are disjoint and open in
[0, 1]. Since 𝑋 = 𝐴 ∪ 𝐵, we would have [0, 1] = 𝛾 −1 (𝐴) ∪ 𝛾 −1 (𝐵) which contradicts the fact
that [0, 1] is connected. Hence 𝑋 must be connected.

Figure 2.6: Not all points may be connected by a path.

∙ We will show later that smooth manifolds, however, are connected if and only if they are
path-connected.

∙ But be aware that there are connected spaces which are not path-connected. A standard
example, illustrated in Figure 2.7, is the subspace

𝑋 = {(𝑥, sin(log 𝑥)) ∈ ℝ2 ∶ 𝑥 > 0} ∪ (0 × [−1, 1]) ⊂ ℝ2 .

24 2.2. Smooth maps

Figure 2.7: A connected but not path-connected space.

2.2 Smooth maps

2.2.1 Maps on open domains

Let 𝑈 ⊆ ℝ𝑛 and 𝑉 ⊆ ℝ𝑚 be open sets. Recall that a map 𝑓 ∶ 𝑈 → 𝑉 is called totally

differentiable at 𝑎 ∈ 𝑈 if there exists a linear map 𝐿𝑎 ∶ ℝ𝑛 → ℝ𝑚 such that

|𝑓 (𝑎 + ℎ) − 𝑓 (𝑎) − 𝐿𝑎 (ℎ)|
lim = 0.
ℎ→0 |ℎ|

Note that if such an 𝐿𝑎 exists, it is unique and is the best possible linear approximation of 𝑓 at
𝑎. Moreover, if 𝐿𝑎 exists it can be represented in the standard bases of ℝ𝑛 and ℝ𝑚 , respectively,
𝜕𝑓
by the Jacobian matrix at 𝑎, the 𝑚 × 𝑛-matrix with (𝑖, 𝑗)-entry the partial derivative 𝜕𝑥 𝑖 (𝑎).
𝑗
Recall that a differentiable map is in particular also continuous.

Conversely, if we know that all partial derivatives at 𝑎 exist and are continuous, then 𝑓 is
differentiable at 𝑎. We say that 𝑓 is differentiable if it is differentiable at every point 𝑎 in 𝑈 .

In differential topology we usually require that maps are not just once but infinitely many
times differentiable. In this case, we call them smooth. More precisely, we define:

Definition 2.17 (Smooth maps on open subsets) Let 𝑈 ⊆ ℝ𝑛 and 𝑉 ⊆ ℝ𝑚 be open

sets. A map 𝑓 ∶ 𝑈 → 𝑉 is called smooth if, at every point 𝑥 ∈ 𝑈 , the partial derivatives
𝜕 𝑘 𝑓𝑖
of 𝑓 of all orders exist and are continuous, i.e., all the partial derivatives 𝜕𝑥 …𝜕𝑥 (𝑎)
𝑗1 𝑗𝑘
exist and are continuous for all 𝑘 ≥ 1.

Examples:

∙ The familiar maps exp, cos, sin and all polynomials are smooth maps from ℝ to ℝ.

∙ Let 𝑝(𝑥1 , 𝑥2 , … , 𝑥𝑛 ) be a polynomial with coefficients in ℝ in the variables 𝑥1 , 𝑥2 , … , 𝑥𝑛 .

Then 𝑝 induces a smooth map 𝑝 ∶ ℝ𝑛 → ℝ by evaluating 𝑝 on the coordinates of ℝ𝑛 , i.e.,
Chapter 2. Smooth manifolds 25
𝜕𝑘 𝑝
by sending 𝑎 = (𝑎1 , … , 𝑎𝑛 ) to 𝑝(𝑎1 , … , 𝑎𝑛 ). The partial derivative 𝜕𝑥𝑗1 …𝜕𝑥𝑗𝑘
(𝑎) is just the
partial derivative of the polynomial 𝑝. The latter always exists and is continuous.

∙ Let 𝑝1 , 𝑝2 , … , 𝑝𝑚 be polynomials with coefficients in ℝ in the variables 𝑥1 , 𝑥2 , … , 𝑥𝑛 .

They induce a smooth map 𝑃 ∶ ℝ𝑛 → ℝ𝑚 by evaluating 𝑝1 , … , 𝑝𝑚 on the coordinates of
ℝ𝑛 , i.e., by sending 𝑎 = (𝑎1 , … , 𝑎𝑛 ) to the 𝑚-tuple
( )
𝑝1 (𝑎1 , … , 𝑎𝑛 ), 𝑝2 (𝑎1 , … , 𝑎𝑛 ), … , 𝑝𝑚 (𝑎1 , … , 𝑎𝑛 ) .

𝜕 𝑘 𝑃𝑖
The partial derivative 𝜕𝑥𝑗1 …𝜕𝑥𝑗𝑘
(𝑎) is just the partial derivative of the polynomial 𝑝𝑖 . The
latter always exists and is continuous.

∙ For example, the map ℝ2 → ℝ2 defined by (𝑥, 𝑦) → (𝑥2 − 𝑦2 , 2𝑥𝑦) is smooth.

∙ Another example of this type is the smooth map

⎛𝑥1 ⎞ ⎛ 2𝑥 𝑥 + 2𝑥 𝑥 ⎞
⎜𝑥 ⎟ 1 3 2 4
𝐹 ∶ ℝ4 → ℝ3 , ⎜ 2 ⎟ → ⎜ 2𝑥2 𝑥3 − 2𝑥1 𝑥4 ⎟ . (2.1)
⎜𝑥3 ⎟ ⎜⎝𝑥2 + 𝑥2 − 𝑥2 − 𝑥2 ⎟⎠
⎝𝑥4 ⎠ 1 2 3 4

To convince ourselves let us calculate some partial derivatives. For example, at a point
𝑎 = (𝑎1 , 𝑎2 , 𝑎3 , 𝑎4 ) ∈ ℝ4 we get

𝜕𝐹1 𝜕 2 𝐹1
(𝑎) = 2𝑎1 , (𝑎) = 0,
𝜕𝑥3 𝜕𝑥3 𝜕𝑥4
𝜕𝐹2 𝜕 2 𝐹2
(𝑎) = −2𝑎4 , (𝑎) = −2,
𝜕𝑥1 𝜕𝑥1 𝜕𝑥4
𝜕𝐹3 𝜕 2 𝐹3
(𝑎) = 2𝑎2 , (𝑎) = 2, …
𝜕𝑥2 𝜕𝑥2 𝜕𝑥2

2.2.2 Extension to arbitrary subsets

Now we would like to extend smoothness to maps between arbitrary sets subsets of ℝ𝑛 . But
there is an issue we need to discuss:

In Calculus, we learned what it means for a function 𝑓 ∶ (𝑎 − 𝜀, 𝑎 + 𝜀) → ℝ defined on

an open interval to be differentiable at the point 𝑎. However, the definition only makes sense
if there is some space on the left and right hand side of 𝑎 in the interval, i.e., if 𝜀 > 0. For
example, we cannot talk about differentiability of a function 𝑓 ∶ [𝑎, 𝑎 + 𝜀) → ℝ at 𝑎. The
definition requires that we can approach 𝑎 from both the left and the right when we take the
limit. That is why we required all maps to be at least defined on an open neighborhood of the
point 𝑎.
26 2.2. Smooth maps
However, there is a way out of this: Given a map 𝑓 ∶ 𝑋 → ℝ𝑚 where 𝑋 ⊂ ℝ𝑛 is an
arbitrary subset. For 𝑓 to be smooth at 𝑎 ∈ 𝑋, we require that 𝑓 is actually just a shadow
of a map which is indeed defined on an open ball 𝔹𝑛𝜀 (𝑎) in ℝ𝑛 . We will see that this simple
trick makes the whole machinery work very nicely. To make things more precise, here is the
definition:

Definition 2.18 (Smooth maps on arbitrary subsets) Let 𝑋 ⊆ ℝ𝑛 and 𝑌 ⊆ ℝ𝑚 be

arbitrary subsets. A map 𝑓 ∶ 𝑋 → 𝑌 is called smooth if for each 𝑥 ∈ 𝑋 there exist an
open subset 𝑈 ⊆ ℝ𝑛 containing 𝑥 and a smooth map 𝐹 ∶ 𝑈 → ℝ𝑚 that coincides with
𝑓 on all of 𝑋 ∩ 𝑈 , i.e.,

𝐹 is smooth and 𝐹𝑋∩𝑈 = 𝑓𝑋∩𝑈 .

Figure 2.8: Smoothness of a map 𝑓 with an arbitrary domain is defined by finding at each point
a smooth map 𝐹 that restrict to 𝑓 on relatively open subset.

Note that smoothness at a point 𝑥 is a local property, i.e., we need to check it only in a
small neighborhood of 𝑥.

Examples and remarks:

∙ The identity map of any set 𝑋 is smooth.

∙ If 𝑓 ∶ 𝑋 → 𝑌 and 𝑔 ∶ 𝑌 → 𝑍 are smooth, then the composition 𝑔◦𝑓 is also smooth.
∙ The projection map
𝜋 ∶ 𝕊1 → ℝ, (𝑥, 𝑦) → 𝑥
is smooth, since it can be extended to the projection ℝ2 → ℝ onto the first coordinate
which is smooth.
∙ Let 𝑓 ∶ 𝕊3 → 𝕊2 be the map defined by

⎛𝑥1 ⎞ ⎛ 2𝑥 𝑥 + 2𝑥 𝑥 ⎞
⎜𝑥 ⎟ 1 3 2 4
𝑓 ∶ 𝕊3 → 𝕊2 , ⎜ 2 ⎟ → ⎜ 2𝑥2 𝑥3 − 2𝑥1 𝑥4 ⎟ .
⎜𝑥3 ⎟ ⎜⎝𝑥2 + 𝑥2 − 𝑥2 − 𝑥2 ⎟⎠
⎝𝑥4 ⎠ 1 2 3 4
Chapter 2. Smooth manifolds 27
This map is the restriction of the map 𝐹 defined in (2.1) and hence smooth. Since 𝐹 is
smooth, it remains to check that if 𝑥 = (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) is in 𝕊3 ⊂ ℝ4 , then 𝑓 (𝑥) ∈ 𝕊2 ⊂
ℝ2 .1

2.2.3 Diffeomorphisms

Definition 2.19 (Diffeomorphism) A smooth map 𝑓 ∶ 𝑋 → 𝑌 is called a diffeomor-

phism if 𝑓 is one-to-one and onto, and its inverse 𝑓 −1 is smooth as well. We say that
𝑋 and 𝑌 are diffeomorphic if there exists a diffeomorphism 𝑓 ∶ 𝑋 → 𝑌 .

Note that every diffeomorphism is a homeomorphism, but not the other way around. Here
are some examples for which it is an exercise to verify the assertions.

Examples:

∙ The map 𝑔 ∶ ℝ → ℝ, 𝑥 → 𝑥3 + 𝑥 is a diffeomorphism.

∙ However, 𝑓 ∶ ℝ → ℝ, 𝑥 → 𝑥3 is a homeomorphism but not a diffeomorphism, since

the inverse map is not differentiable and therefore not smooth at the origin.

∙ The map ℝ2 ⧵ {(0, 0)} → ℝ2 ⧵ {(0, 0)}, (𝑥, 𝑦) → (𝑥2 − 𝑦2 , 2𝑥𝑦), is not a diffeomorphism
– even though its derivative is invertible everywhere – because it is not one-to-one.

∙ Let 𝕊2 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 + 𝑧2 = 1} be the two-dimensional sphere. The map

1
𝑓 ∶ 𝕊2 ⧵ {(0, 0, 1)} → ℝ2 , (𝑥, 𝑦, 𝑧) → (𝑥, 𝑦)
1−𝑧
is a diffeomorphism. We will meet it again soon and see that it is quite useful.

∙ The map 𝑓 ∶ 𝕊3 → 𝕊2 defined by

⎛𝑥1 ⎞ ⎛ 2𝑥 𝑥 + 2𝑥 𝑥 ⎞
⎜𝑥 ⎟ 1 3 2 4
𝑓 ∶ 𝕊3 → 𝕊2 , ⎜ 2 ⎟ → ⎜ 2𝑥2 𝑥3 − 2𝑥1 𝑥4 ⎟
⎜𝑥3 ⎟ ⎜⎝𝑥2 + 𝑥2 − 𝑥2 − 𝑥2 ⎟⎠
⎝𝑥4 ⎠ 1 2 3 4

is not a diffeomorphism, since it is not one-to-one. For example, the whole unit circle
in the 𝑥1 -𝑥2 -plane on 𝕊3 , i.e., points on 𝕊3 with 𝑥3 = 𝑥4 = 0, is mapped to the north
pole (0, 0, 1) on 𝕊2 . The whole unit circle in the 𝑥3 -𝑥4 -plane on 𝕊3 , i.e., points on 𝕊3
with 𝑥1 = 𝑥2 = 0, is mapped to the south pole (0, 0, −1) on 𝕊2 . In fact, we will see
in Exercise 2.8 that each fiber of the Hopf fibration 𝑓 is diffeomorphic to a circle on
𝕊3 . However, as we will show later, none of these circles intersect even though they are
all linked into each other. This is a fascinating and very rare phenomenon. More on
1
We will study this map further though in different disguise in the exercises and will meet it many times in these
notes. But let us remark anyway that if 𝑥21 + 𝑥22 + 𝑥23 + 𝑥24 = 1, then

(2𝑥1 𝑥3 + 2𝑥2 𝑥4 )2 + (2𝑥2 𝑥3 − 2𝑥1 𝑥4 )2 + (𝑥21 + 𝑥22 − 𝑥23 − 𝑥24 )2 = (𝑥21 + 𝑥22 + 𝑥23 + 𝑥24 )2 = 1.
28 2.3. Smooth manifolds
this later. For the moment, we conclude this example with the remark that after defining
tangent spaces for manifolds we will see that, in fact, there cannot exist a diffeomorphism
between 𝕊3 and 𝕊2 .

Remark 2.20 (Diffeomorphic spaces are equivalent) From the point of view of dif-
ferential topology, diffeomorphic spaces are equivalent, and we may (and will) consider
them as copies of the same abstract space, which may happen to be differently situated
in their surrounding Euclidean spaces.

2.3 Smooth manifolds

2.3.1 How to describe a space of solutions?

Many interesting spaces are given as the set of solutions of an equation of the form

𝑓 (𝑥) = 𝑏

where 𝑓 ∶ 𝑋 → 𝑌 is a map and 𝑏 ∈ 𝑌 is some specified point. It is a natural and important

question: how we can best describe the space

𝐒 = {𝑥 ∈ 𝑋 ∶ 𝑓 (𝑥) = 𝑏} ⊂ 𝑋?

∙ (Goal) We would like to describe the space 𝐒 in a simple and efficient way while
still expressing all its interesting properties.

Let us look at a familiar situation and consider the linear map

⎛𝑥1 ⎞ ⎛ 2𝑥2 + 2𝑥4 ⎞

4 3 ⎜𝑥2 ⎟ ⎜ ⎟
𝐴 ∶ ℝ → ℝ , ⎜ ⎟ → 2𝑥3 − 2𝑥1
⎜ ⎟ ⎝𝑥1 + 𝑥2 − 𝑥3 − 𝑥4 ⎟⎠
𝑥3 ⎜
⎝𝑥4 ⎠

ans solve the equation

𝐴(𝐱) = 𝟎.

We can use, for example, Gauss elimination to get the set of solutions

𝐋 = {𝐱 ∈ ℝ4 ∶ 𝑥2 = 0, 𝑥3 = 𝑥1 , 𝑥4 = −𝑥2 } ⊂ ℝ4 .

This is a line in ℝ4 . In particular, it is something one-dimensional, i.e., we can describe all the
points in 𝐋 by using just one variable, say 𝑡:

⎧ ⎛1⎞⎫
⎪ ⎜0⎟⎪
𝐋 = ⎨𝐱 ∈ ℝ4 ∶ 𝐱 = 𝑡 ⋅ ⎜ ⎟⎬ .
⎪ ⎜1⎟⎪
⎩ ⎝0⎠⎭
Chapter 2. Smooth manifolds 29
We think of the variable 𝑡 as a parameter and would like to say that 𝑡 parametrizes the set of
solutions 𝐋. In a more formal way we have the map

⎛1⎞
⎜0⎟
𝜓 ∶ ℝ → 𝐋, 𝑡 → 𝑡 ⋅ ⎜ ⎟ .
⎜1⎟
⎝0⎠

In fact, 𝜓 is a linear isomorphism. Hence we may think of 𝜓 as a mean to express

∙ the one-dimensionality of 𝐋, and

∙ that 𝐋 has a linear structure.

We call the map 𝜓 a parametrization of 𝐋.

Now let us look at the map2

⎛𝑥1 ⎞ ⎛ 2𝑥 𝑥 + 2𝑥 𝑥 ⎞
⎜𝑥 ⎟ 1 3 2 4
𝑓 ∶ ℝ4 → ℝ3 , ⎜ 2 ⎟ → ⎜ 2𝑥2 𝑥3 − 2𝑥1 𝑥4 ⎟ .
⎜𝑥3 ⎟ ⎜⎝𝑥2 + 𝑥2 − 𝑥2 − 𝑥2 ⎟⎠
⎝𝑥4 ⎠ 1 2 3 4

This map is not linear, since we multiply variables. W have seen such maps in multivariable
calculus and know how to calculate their derivatives. In fact, 𝑓 is a smooth map, since each
of its coordinates is a polynomial.The equation 𝑓 (𝐱) = 𝟎 has only a single solution – the zero
vector. So let us rather determine the solutions the equation

⎛0⎞
𝑓 (𝐱) = 𝐛 =∶ ⎜1⎟ .
⎜ ⎟
⎝0⎠

After some calculation we arrive at the set of solutions

𝐒 = {𝐱 ∈ ℝ4 ∶ 𝑥3 = 𝑥2 , 𝑥4 = −𝑥1 , 𝑥21 + 𝑥22 = 1∕2} ⊂ ℝ4 .

This is not a straight line in ℝ4 . However, it looks like something one-dimensional, since
knowing one of the variables, say 𝑥1 , determines 𝑥2 , 𝑥3 and 𝑥4 , and thereby 𝐱. Well, hold on:
Let us fix a value 𝑥1 = 𝑡. Then we get
√
1∕2 = 𝑥21 + 𝑥22 = 𝑡2 + 𝑥22 ⇒ 𝑥2 = ± 1∕2 − 𝑡2 .

Hence 𝑥2 is only determined up to a choice. To remedy this defect, let us restrict our attention
to points 𝐱 ∈ 𝐒 with 𝑥2 ≥ 0, then 𝑥1 = 𝑡 determines 𝐱 completely. We write 𝐒𝑥2 ≥0 for the set
of such points.
2
The map 𝑓 is modelled on the famous Hopf fibration 𝕊3 → 𝕊2 which is one of the very few smooth maps
between spheres whose fibers are all spheres themselves. We will meet the Hopf map many times during this class.
30 2.3. Smooth manifolds
In addition, we need
[ to make sure that
] the square root is defined, i.e., we need that 𝑡 only
√ √
varies in the range 𝑡 ∈ − 1∕2, 1∕2 . Hence we can use the map

⎛ 𝑡 ⎞
[ √ ] ⎜√ ⎟
√ 1∕2 − 𝑡2 ⎟
𝜙̃ + ∶ − 1∕2, 1∕2 → 𝐒𝑥2 ≥0 ⊂ ℝ4 , 𝑡 → ⎜√
⎜ 1∕2 − 𝑡2 ⎟
⎜ −𝑡 ⎟
⎝ ⎠

to describe one part of the set of solutions 𝐒. And we check that this map is a bijection.

This is very similar to the parametrization we used to describe 𝐋. However, this map is
far from linear. There is no way to fix this, since 𝑓 was not linear in the first place. But we can
check that 𝑓 is differentiable at many points. In particular, it is continuously differentiable at
all points in 𝐒. Hence we would like our map 𝜙̃ + to be differentiable as well. In fact, we would
like it to be smooth, since 𝑓 is smooth at all points in 𝐒.

To achieve this, we need to make sure that the domain of 𝜙̃ + is open and the partial deriva-
tives are defined. Hence we replace 𝜙̃ + with the map

⎛ 𝑡 ⎞
( √ √ ) ⎜√ 2 ⎟
1∕2 − 𝑡 ⎟
𝜙+ ∶ − 1∕2, 1∕2 → 𝐒𝑥2 >0 ⊂ ℝ4 , 𝑡 → ⎜√
⎜ 1∕2 − 𝑡2 ⎟
⎜ −𝑡 ⎟
⎝ ⎠

defined on an open interval.

The map 𝜙+ is now a diffeomorphism, the best we can hope for, and does a similar job as
the parametrization 𝜓 above: it expresses

∙ the one-dimensionality of 𝐒 for points in 𝐒𝑥2 >0 , and

∙ that 𝐒 has a smooth structure3 for points in 𝐒𝑥2 >0 .

Since 𝜙+ describes only some part of 𝐒, we call the map 𝜙+ a local parametrization of 𝐒.

Finally, we observe that we are missing out on some points of 𝐒, in particular where 𝑥2 ≤ 0.
Hence we need further local parametrizations similar to 𝜙+ to cover all of 𝐒. The collection
of such maps will then express 𝐒 as a one-dimensional and smooth subspace of ℝ4 . In fact,
these maps give 𝐒 the structure of a one-dimensional smooth manifold, a notion we will now
define rigorously based on what we learned from this example.

2.3.2 Smooth manifolds - the definition

Let 𝑋 ⊆ ℝ𝑛 be an arbitrary subset. We learned what it means for subsets in 𝑋 ⊆ ℝ𝑛 to be open.

One reason why open sets are useful is that they give us a way to talk about things that happen
3
This is not yet a well-defined term. So for the moment we may think of it as saying that 𝐒 does not have any
corners or any other nasty points.
Chapter 2. Smooth manifolds 31
close to a point. In order to facilitate this way of thinking we are going to use the following
terminology:

Definition 2.21 (Neighborhoods) We say that a subset 𝑉 ⊆ 𝑋 containing a point

𝑥 ∈ 𝑋 is a neighborhood of 𝑥 if there is an open subset 𝑈 ⊆ 𝑉 with 𝑥 ∈ 𝑈 . If 𝑉 itself
is open, we call 𝑉 an open neighborhood.

We will also use the following abbreviation:

Remark 2.22 (A way of speaking: Local properties) If we refer to something that

happens in the neighborhood of a point 𝑥 ∈ 𝑋, then we are often going to say that it
happens locally at 𝑥. Moreover, a property of a space or a function that we only need to
test for a neighborhood of each point is a local property. For example, smoothness
of a map is a local property, since we test it in a neighborhood of each point. In contrast,
there are global properties which are properties that describe the whole space.

Manifolds are now spaces that locally look like Euclidean spaces in the following sense:

Definition 2.23 (Smooth manifolds) Let ℝ𝑁 be some big Euclidean space.

∙ A subset 𝑋 ⊆ ℝ𝑁 is a 𝑘-dimensional smooth manifold if it is locally diffeo-

morphic to ℝ𝑘 . The latter means that for every point 𝑥 ∈ 𝑋 there is an open
subset 𝑉 ⊂ 𝑋 containing 𝑥 and an open subset 𝑈 ⊆ ℝ𝑘 such that 𝑈 and 𝑉 are
diffeomorphic. Note that the number 𝑘 is the same for all points in 𝑋.

∙ Any such diffeomorphism 𝜙 ∶ 𝑈 → 𝑉 is called a local parametrization.

∙ The inverse diffeomorphism 𝜙−1 ∶ 𝑉 → 𝑈 is called a local coordinate system

on 𝑉 .

The natural number 𝑁 in the previous definition is not specified. We just assume that there
is some ℝ𝑁 big enough to fit 𝑋 into it. We are going to discuss what we can say about the
minimal 𝑁 later. It is actually a very interesting question.

Figure 2.9: Points on 𝕊2 and on 𝕋 2 have both open nighborhoods diffeomorphic to open subsets
in ℝ2 . However, 𝕊2 and 𝕋 2 have different global properties.
32 2.3. Smooth manifolds

Figure 2.10: A hyperboloid is an example of a smooth 2-manifold. The cone, however, is not
a manifold, since it has a point without an open neighborhood diffeomorphic to an open subset
of ℝ2 . More about these two spaces in Exercise 2.5.

Remark 2.24 (Local coordinates) The set 𝑈 in the definition of a local parametrization
is a subset of ℝ𝑘 , and it may therefore seem plausible to express a point 𝑢 ∈ 𝑈 by its
coordinates 𝑢 = (𝑢1 , 𝑢2 , … , 𝑢𝑘 ). More precisely, given a coordinate system

𝜙−1 ∶ 𝑉 → 𝑈

on 𝑉 , we can talk about the coordinates (𝜙−11

(𝑥), 𝜙−1
2
(𝑥), … , 𝜙−1
𝑘
(𝑥)) of a point 𝑥 ∈ 𝑉 ⊂
𝑋. Writing 𝑢𝑖 (𝑥) = 𝜙𝑖 (𝑥) for 𝑖 = 1, … , 𝑘, we usually drop mentioning 𝜙−1 and just
−1

talk about the coordinates (𝑢1 (𝑥), 𝑢2 (𝑥), … , 𝑢𝑘 (𝑥)) of 𝑥. Hence we need to remember
that the 𝑢1 , … , 𝑢𝑘 are really coordinate functions.

Remark 2.25 (Simplified notation for parametrizations) Let 𝑋 ⊆ ℝ𝑁 be 𝑘-

dimensional manifold and 𝑥 ∈ 𝑋 a point. Let 𝜙 ∶ 𝑈 → 𝑉 be a local parametrization
around 𝑥, i.e., 𝑈 ⊆ ℝ𝑘 and 𝑉 ⊆ 𝑋 are open subsets with 𝑥 ∈ 𝑉 and 𝜙 ∶ 𝑈 → 𝑉 is a
𝜙
diffeomorphism. Then we also write 𝜙 ∶ 𝑈 → 𝑋 for the composite 𝑈 ←←←→
← 𝑉 → 𝑋. We
usually assume that 𝜙 is adjusted such that 𝜙(0) = 𝑥.

2.3.3 First examples

We are very well familiar with some simple examples:

∙ An open subset 𝑈 ⊆ ℝ𝑘 is a 𝑘-dimensional manifold. The identity map 𝑈 → 𝑈 is a

parametrization for all of 𝑈 . For example, any 𝑘-dimensional open ball 𝔹𝑘𝑟 (𝑥) around
some point 𝑥 ∈ ℝ𝑘 is a manifold of dimension 𝑘.

∙ In particular, Euclidean space ℝ𝑘 is a 𝑘-dimensional manifold.

∙ A 0-dimensional manifold 𝑀 just consists of a collection of discrete points. Given 𝑥 ∈

Chapter 2. Smooth manifolds 33
𝑀, the set {𝑥} ⊂ 𝑀 consisting of 𝑥 alone is open in 𝑀 and is diffeomorphic to the
one-point set ℝ0 = {0}.

A fundamental example that will play an important role during the whole semester is the
𝑛-dimensional sphere. We start with the one-dimensional case: the unit circle.

Example 2.26 (The unit circle) We start with 𝑛 = 1: Let

𝕊1 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 + 𝑦2 = 1} ⊂ ℝ2

be the unit circle. We are going to show that 𝕊1 is a one-dimensional manifold. First,
suppose that (𝑥, 𝑦) lies in the upper semicircle where 𝑦 > 0. Then
√
𝜙1 (𝑥) = (𝑥, 1 − 𝑥2 )

maps the open interval 𝑊 = (−1, 1) ⊂ ℝ bijectively onto the upper semicircle. It is a
smooth map (−1, 1) → ℝ2 , since its partial derivatives exist and are continuous. Here
it is important that we do not include the endpoints of the interval (−1, 1). Its inverse is
the projection map
𝜙−1
1
(𝑥, 𝑦) = 𝑥
which is defined on the upper semicircle. This 𝜙−1
1
is smooth, since it extends to a smooth
map of all of ℝ to ℝ . Therefore, 𝜙1 is a parametrization.
2 1

A parametrization of the lower semicircle where 𝑦 < 0 is similarly defined by

√
𝜙2 (𝑥) = (𝑥, − 1 − 𝑥2 ) with inverse 𝜙−1
2
(𝑥, 𝑦) = 𝑥.

These two maps give local parametrizations of 𝕊1 around any point except the two points
(1, 0) and (−1, 0). To cover these points, we can use the maps
√ √
𝜙3 (𝑦) = ( 1 − 𝑦2 , 𝑦) and 𝜙4 (𝑦) = (− 1 − 𝑦2 , 𝑦)

which map 𝑊 to the right and left semicircles, respectively.

This shows that 𝕊1 is a 1-dimensional manifold.

More generally, we will show in the exercises:

Example 2.27 (𝑛-sphere) The 𝑛-sphere

𝕊𝑛 = {𝑥 ∈ ℝ𝑛+1 ∶ |𝑥| = 1} ⊂ ℝ𝑛+1

is an 𝑛-dimensional smooth manifold.

∙ Another example is the set of solutions 𝐒 ⊂ ℝ4 of the equation 𝑓 (𝐱) = 𝐛 that we have
seen at the beginning of the chapter. We can check that the map 𝜙+ we defined is a local
parametrization of 𝐒. Using what we learned from the local parametrizations of 𝕊1 it
should not be too difficult to write down the missing local parametrizations for 𝐒. We
just need to adjust for how 𝐒 sits inside ℝ4 . It is a good exercise to work this out on your
own. Note that we will meet the map 𝑓 and the set 𝐒 again during this course.
34 2.3. Smooth manifolds

Figure 2.11: A simple parametrization of the 1-manifold 𝕊1 .

∙ In fact, we will see later that many smooth manifolds arise as the set of solutions of a
suitable equation involving a smooth function.4

The definition of a manifold requires parametrizations that cover the whole space. It is
a natural question, what the minimal number of such maps is. The answer depends on the
manifold we look at. Here is a first thought about this number for the sphere:

Remark 2.28 (Need at least two parametrizations on the sphere) Note that we
have used four parametrization maps in the above example. It is an exercise to show
that it is possible to cover 𝕊1 with only two parametrizations. But note that just one
parametrization cannot be enough, because 𝕊1 is compact. For, if there was a dif-
feomorphism 𝜙 ∶ 𝕊1 → 𝑈 ⊂ ℝ1 to an open subset, it would mean that 𝑈 is compact
contradicting the Theorem of Heine-Borel which says that the compact subsets of ℝ1 are
closed and bounded. This argument actually holds for the 𝑛-sphere in every dimension
𝑛 ≥ 1.

There are many different ways to choose parametrizations for a sphere. There is a very
economical one which shows that two parametrizations suffice to cover the 𝑛-sphere:

Remark 2.29 (Stereographic projection) The method of stereographic projection

yields a cover of the 𝑛-sphere with only two parametrizations. In Exercise 2.7 we find
the formulae for the corresponding diffeomorphisms.

This is an illustration of the stereographic projection for the 2-sphere 𝕊2 . We study the
formulae of the maps involved in the exercises and will show that this actually defines a sufficient
parametrization.

4
If you cannot wait, you may want to fast-forward to Section 4.2.1 on regular values to have a first glimpse.
Chapter 2. Smooth manifolds 35

Figure 2.12: A diffeomorphism between 𝕊𝑛 ⧵ {𝑁} and ℝ𝑛 .

Definition 2.30 (Morphisms between smooth manifolds: smooth maps) Let 𝑋 ⊂

ℝ𝑁 and 𝑌 ⊂ ℝ𝑀 be two smooth manifolds. Then a smooth map between the manifolds
𝑋 and 𝑌 is just a smooth map in the sense we defined previously. In fact, those are the
maps which respect the smooth manifold structure on 𝑋 and 𝑌 . Hence smooth maps are
the morphisms in the category of smooth manifolds.

Manifolds have subsets. We are interested in those subsets which are manifolds on their
own, possibly of lower dimension:

Definition 2.31 (Submanifolds) Let 𝑋 ⊂ ℝ𝑁 and 𝑍 ⊂ 𝑋 be a subset considered

as a topological space with relative topology induced from 𝑋 and hence from ℝ𝑁 . If
both 𝑍 and 𝑋 are manifolds — possibly of different dimensions — then 𝑍 is called a
submanifold of 𝑋. In particular, 𝑋 itself is a submanifold of ℝ𝑁 . Any open subset of
𝑋 is a submanifold of 𝑋.

Examples and remarks:

∙ We could consider the equator in 𝕊2 as a copy of 𝕊1 and hence as a submanifold.

∙ Similarly, we have basically two ways of considering a copy of the circle on the two-
dimensional torus: once as a horizontal circle, once as a vertical circle.

∙ We can generalize these examples to consider copies of 𝕊1 , 𝕊2 , … , 𝕊𝑛−1 as submanifolds

of 𝕊𝑛 . As in the previous cases there are different ways of how these submanifolds sit
inside the bigger manifold. Understanding all possible ways of how submanifolds can
sit inside a bigger manifold is actually a very interesting and useful problem to study.
36 2.3. Smooth manifolds
We will get back to this question when we discuss embeddings and intersection theory.

2.3.4 Product manifolds

To find submanifolds in already existing manifolds is an important way to define and study new
manifolds. But there are also other ways to produce new manifolds:

Lemma 2.32 (Creating new manifolds out of old ones) Let 𝑋 ⊆ ℝ𝑁 and 𝑌 ⊆ ℝ𝑀
be manifolds of dimensions 𝑘 and 𝑙, respectively. Then 𝑋 × 𝑌 ⊆ ℝ𝑁+𝑀 is a manifold
of dimension 𝑘 + 𝑙.

For, let 𝑊 ⊂ ℝ𝑘 an open set with 𝜙 ∶ 𝑊 → 𝑋 a local parametrization around 𝑥 ∈ 𝑋, and

𝑈 ⊂ ℝ𝑘 an open set with 𝜓 ∶ 𝑈 → 𝑌 a local parametrization around 𝑦 ∈ 𝑌 . Then we can
define the map
𝜙 × 𝜓 ∶ 𝑊 × 𝑈 → 𝑋 × 𝑌 , (𝜙 × 𝜓)(𝑤, 𝑢) = (𝜙(𝑤), 𝜓(𝑢)).
from the open set 𝑊 × 𝑈 ⊆ ℝ𝑘 × ℝ𝑙 = ℝ𝑘+𝑙 to 𝑋 × 𝑌 . This map defines a local parametrization
around (𝑥, 𝑦). We recommend it to check this assertion as an exercise.

Example 2.33 (Torus) One way to define the two-dimensional torus is to think of
it as the product 𝕋 2 = 𝕊1 × 𝕊1 . Then the general statement above implies that 𝕋 2 is a
two-dimensional manifold. This is convenient. However, this way we consider 𝕋 2 as a
subset of ℝ4 , since 𝕊1 being a subspace of ℝ2 forces us to take the product of ℝ2 with
itself to embed 𝕋 2 . Since we are more used to visualise the torus as a subspace in three
dimensions, we will discuss a way to describe 𝕋 2 as a subspace in ℝ3 in the exercises.

2.3.5 A non-example

Finally, we are now going to discuss the case of a space which is not a manifold:

Let 𝑋 denote the union of the 𝑥- and the 𝑦-axis in ℝ2 , in other words,
𝑋 = {(𝑥, 𝑦) ∈ ℝ2 such that 𝑥𝑦 = 0}.
The critical point is the origin 𝑂 = (0, 0), as every other point on 𝑋 has an open neighborhood
which is diffeomorphic to an open interval in ℝ. But no point in ℝ𝑑 with 𝑑 ≥ 2 has an
open neighborhood in ℝ𝑑 diffeomorphic to an open interval in ℝ1 .5 Hence 𝑋 could only be
1-dimensional.

Now let us check the point 𝑂 = (0, 0). If 𝑋 was a manifold of dimension one, there would
be an open subset 𝑉 ⊆ 𝑋 around 𝑂 diffeomorphic to an open interval in ℝ1 . By definition of
open sets in a subset of ℝ2 , there must be an open ball 𝔹2𝜀 (𝑂) such that 𝔹2𝜀 (𝑂) ∩ 𝑋 is contained
in 𝑉 . Let 𝐼 be the open interval in ℝ homeomorphic to 𝔹2𝜀 (𝑂) ∩ 𝑋.
5
We can prove this fact when we have introduced tangent spaces.
Chapter 2. Smooth manifolds 37
The subset 𝔹𝜀 (𝑂) ∩ 𝑋 looks like a cross and contains, in particular, the points

𝑃1 = (−𝜀∕2, 0), 𝑃2 = (0, 𝜀∕2), and 𝑃3 = (𝜀∕2, 0).

In 𝔹2𝜀 (𝑂) ∩ 𝑋, there are paths, i.e., continuous maps 𝛾 ∶ [0, 1] → 𝔹2𝜀 (𝑂) ∩ 𝑋,

∙ 𝛾1 from 𝑃2 to 𝑃3 not passing through 𝑃1 .

∙ 𝛾2 from 𝑃1 to 𝑃3 not passing through 𝑃2 ,

∙ 𝛾3 from 𝑃1 to 𝑃2 not passing through 𝑃3 .

But there is no triple of distinct points with this property in the open intervall 𝐼 ⊂ ℝ. In
more detail, we can argue as follows:

Since 𝑃1 , 𝑃2 , 𝑃3 are pairwise distinct points, their images under 𝜙 must be pairwise distinct
as well. Since ℝ is a totally ordered set, we can order these three points. Assume first 𝜙(𝑃1 ) <
𝜙(𝑃2 ) < 𝜙(𝑃3 ). Then the Intermediate Value Theorem of Calculus implies that for the path

𝜙◦𝛾2 ∶ [0, 1] → 𝔹2𝜀 (𝑂) ∩ 𝑋 with 𝛾2 (0) = 𝑃1 and 𝛾2 (1) = 𝑃3

there is an 𝑠 ∈ (0, 1) such that (𝜙◦𝛾2 )(𝑠) = 𝜙(𝑃2 ). This would imply 𝛾2 (𝑠) = 𝑃2 contradicting
the choice of 𝛾2 . Thus the diffeomorphism 𝜙 with the assumed ordering 𝜙(𝑃1 ) < 𝜙(𝑃2 ) < 𝜙(𝑃3 )
cannot exist.

Now we can adjust and repeat this argument for any ordering of the three points 𝜙(𝑃1 ),
𝜙(𝑃2 ) and 𝜙(𝑃3 ) and get contradictions to the choices of paths 𝛾1 , 𝛾2 and 𝛾3 .

Hence the homeomorphism 𝜙 ∶ 𝔹2𝜀 (𝑂) ∩ 𝑋 → 𝐼 cannot exist. We conclude that 𝑂 does
not have a neighborhood homeomorphic to an open interval in ℝ, and 𝑋 is not a manifold.

Paths and path-connectedness - an alternative argument

In the discussion above, we used implicitly that we were looking at path-connected spaces.
Recall that a topological space 𝑋 is called path-connected if for any two points 𝑥, 𝑦 ∈ 𝑋 there
is a continuous map 𝛾 ∶ [0, 1] → 𝑋 from the unit interval to 𝑋 with 𝛾(0) = 𝑥 and 𝛾(1) = 𝑦.
Path-connectedness is a topological property, i.e. it is preserved under homeomorphisms.

The union of the coordinate axes in ℝ2 is an example of a path-connected space and every
interval in ℝ is path-connected. Now assume 𝜙 ∶ 𝔹2𝜀 (𝑂) ∩ 𝑋 → 𝐼 was a homeomorphism to
an interval 𝐼 ⊂ ℝ. Let 𝜙(𝑂) ∈ 𝐼 be the image of the origin. If we remove 𝜙(𝑂) from 𝐼, we
get two components of the interval. Points in the same component can be connected by a path,
whereas points from different components cannot be connected to each other via a path without
crossing 𝜙(𝑂).

If we remove 𝑂 from 𝔹2𝜀 (𝑂) ∩ 𝑋 we get a space with four components. Again, points in the
same component can be connected by a path, whereas points from different components cannot
be connected to each other via a path.

We call these subsets the path-components of the spaces 𝐼 ⧵{𝜙(𝑂)} and (𝔹2𝜀 (𝑂)∩𝑋)⧵{𝑂}.
The key observation is that if 𝜙 was a homeomorphism, 𝜙|(𝔹2 (𝑂)∩𝑋)⧵{𝑂} would still be a home-
𝜀
omorphism. But homeomorphic spaces need to have the same number of path-components,
assuming that number is finite. This is the background for the argument we used above.
38 2.4. Tangent spaces and the derivative

Figure 2.13: Three points on the coordinate axes, two of which can be connected without pass-
ing through the other one.

2.4 Tangent spaces and the derivative

We are now going to introduce one of the key tools to study smooth manifolds.

2.4.1 The tangent space - motivation

Let 𝑥 be a point on the smooth manifold 𝑋. By definition, we can choose a local parametrization
𝜙 ∶ 𝑈 → 𝑋 around 𝑥 which tells us that, at least locally at 𝑥, 𝑋 is the image of 𝑈 ⊂ ℝ𝑘 under
the diffeomorphism 𝜙. Images under diffeomorphism are nice, but images under linear maps
are even better since the latter are vector spaces.

So how could we describe 𝑋, at least locally at 𝑥, via a vector space? Well, 𝑋 is itself not a
vector space, but what we can look for is a linear approximation of 𝑋 at 𝑥. This is the purpose
of the tangent space at 𝑥.

In order to motivate our construction we begin with a familiar situation. Let 𝑓 ∶ ℝ → ℝ

be a smooth function and let 𝑓 ′ (𝑥) be the derivative of 𝑓 at 𝑥. In Calculus, we think of the
derivative often as the slope of the tangent line at the graph of 𝑓 at the point (𝑥, 𝑓 (𝑥)). The
graph of 𝑓 , i.e., the subset Γ(𝑓 ) = {(𝑥, 𝑓 (𝑥)) ∈ ℝ2 ∶ 𝑥 ∈ ℝ} ⊂ ℝ2 , is an example of a smooth
manifold. We have the natural map

𝜙 ∶ ℝ → Γ(𝑓 ), 𝑥 → (𝑥, 𝑓 (𝑥)).

This map is smooth since 𝑓 is smooth, and 𝜙 is injective because of the first coordinate and
surjective by definition of Γ(𝑓 ). Moreover, the projection (𝑥, 𝑓 (𝑥)) → 𝑥 defines a smooth
inverse. Thus, 𝜙 is a diffeomorphism and yields us a parametrization for all points of Γ(𝑓 ).
Chapter 2. Smooth manifolds 39
The tangent line at the point (𝑥, 𝑓 (𝑥)) is the prototype of an example of a tangent space of
smooth manifold. More precisely, we prefer to consider the parallel translate of the tangent
line to the origin, since we want the tangent space to be a vector space. See Figure 2.14. How
can we describe the tangent line 𝐿𝑥 passing through the origin? It is determined by its slope,
i.e., { ( ) }
1 2
𝐿𝑥 = 𝑡 ⋅ ∈ ℝ ∶ 𝑡 ∈ ℝ .
𝑓 ′ (𝑥)
( )
1
Now we observe that is exactly the derivative at 𝑥 of the map 𝜙 we defined above,
𝑓 ′ (𝑥)
( )
1
i.e., 𝑑𝜙𝑥 = ′ . Thus, the tangent line 𝐿𝑥 is the image in ℝ2 of the linear map
𝑓 (𝑥)

𝑑𝜙𝑥 ∶ ℝ → ℝ2 .

Figure 2.14: The tangent line at the graph Γ(𝑓 ) of 𝑓 is the parallel translate of the tangent space
of Γ(𝑓 ).

Our goal is to generalize this observation to an arbitrary smooth manifold by following

the same recipe: the tangent space at a point should be the image of the derivative of a local
parametrization. To make this precise we recall some facts about the derivative of a smooth
map. Let
𝜙∶ 𝑈 → 𝑉
where 𝑈 ⊂ ℝ𝑛 and 𝑉 ⊂ ℝ𝑁 are open subsets. Let 𝑢 ∈ 𝑈 be a point in the domain of 𝑓 and
ℎ ∈ ℝ𝑛 be a vector in ℝ𝑛 . Then the derivative of 𝜙 in the direction ℎ can be defined as the
limit
𝜙(𝑢 + 𝑡ℎ) − 𝜙(𝑢)
𝑑𝜙𝑢 (ℎ) = lim .
𝑡→0 𝑡
For a fixed 𝑢, the derivative is a map

𝑑𝜙𝑢 ∶ ℝ𝑛 → ℝ𝑁

sending a vector ℎ ∈ ℝ𝑛 to the vector 𝑑𝜙𝑢 (ℎ) ∈ ℝ𝑁 . In Calculus we learned that this map is
ℝ-linear, i.e., 𝑑𝜙𝑢 (ℎ + 𝑔) = 𝑑𝜙𝑢 (ℎ) + 𝑑𝜙𝑢 (𝑔) and 𝑑𝜙𝑢 (𝜆ℎ) = 𝜆𝑑𝜙𝑢 (ℎ) for all ℎ, 𝑔 ∈ ℝ𝑛 and
40 2.4. Tangent spaces and the derivative
𝜆 ∈ ℝ. In particular, the derivative of 𝜙 is a map on its own which is defined on all of ℝ𝑛 even
when 𝜙 may not be. Recall that we can calculate 𝑑𝜙𝑢 in the standard bases of Euclidean spaces
𝜕𝜙
as the Jacobian matrix. Its entry in row 𝑖 and column 𝑗 is the partial derivative 𝜕𝑥 𝑖 (𝑢).
𝑗

Remark 2.34 (The derivative is a linear approximation) One way to appreciate the
significance of the derivative is to think of it as a simple and useful approximation to 𝜙
at 𝑢, i.e., knowing 𝜙(𝑢) and 𝑑𝜙𝑢 gives us a good guess for what 𝜙(𝑢 + ℎ) might, namely
something close to 𝜙(𝑢) + 𝑑𝜙𝑢 (ℎ).

2.4.2 The tangent space - definition

Now we are ready to define tangent spaces in general:

Definition 2.35 (Tangent space) Let 𝑋 ⊆ ℝ𝑁 be 𝑘-dimensional manifold and 𝑥 ∈ 𝑋

a point. Let 𝜙 ∶ 𝑈 → 𝑉 be a local parametrization around 𝑥 with 𝜙(𝑢) = 𝑥. We define
the tangent space of 𝑋 at 𝑥 to be the image of the linear map 𝑑𝜙𝑢 ∶ ℝ𝑘 → ℝ𝑁 . We
denote it by 𝑇𝑥 (𝑋). This is a vector subspace of ℝ𝑁 .

By this definition, a tangent vector to 𝑋 ⊆ ℝ𝑁 at 𝑥 is a point 𝑣 ∈ ℝ𝑁 that lies in the vector

subspace 𝑇𝑥 (𝑋) ⊂ ℝ𝑁 . However, we usually picture 𝑣 geometrically as the arrow pointing
from 𝑥 to 𝑥 + 𝑣 in the translate 𝑥 + 𝑇𝑥 (𝑋). See Figure 2.15.

Figure 2.15: The tangent space is the isomorphic image of a ℝ𝑘 in ℝ𝑁 . We visualize it as the
parallel translate of this plane attached to the point of the manifold.

∙ Tangent spaces are useful: While tangent spaces may look quite boring, since they are
just vector spaces, we will see very soon that they are extremely useful for understanding
manifolds. Many important geometric conditions can be stated in terms of tangent spaces.
The most important example for us might be transversality, a key condition for making
intersection theory work.

Lemma 2.36 (Dimension of 𝑇𝑥 (𝑋)) If 𝑋 is a 𝑘-dimensional manifold, then 𝑇𝑥 (𝑋) is

a 𝑘-dimensional vector space over ℝ.
Chapter 2. Smooth manifolds 41
Proof: Since a local parametrization 𝜙 is a diffeomorphism onto its image, its derivative
𝑑𝜙𝑢 is injective. Hence by definition of the vector space 𝑇𝑥 (𝑋) of the image of ℝ𝑘 under 𝑑𝜙𝑢 ,
the dimension of 𝑇𝑥 (𝑋) is 𝑘.

2.4.3 Independence of choices: 𝑇𝑥 (𝑋) is well-defined

In order to define 𝑇𝑥 (𝑋) we made a choice of a parametrization 𝜙. We have to check what

happens if we choose a different parametrization.

∙ Question: Do we get the same tangent space?

We can find an answer to this question by taking another local parametrization and check
whether 𝑇𝑥 (𝑋) changes. So let 𝜓 ∶ 𝑉 → 𝑋 be another local parametrization around 𝑥 with
𝜓(0) = 𝑥. If necessary, we shrink 𝑈 and 𝑉 , i.e., we replace 𝑈 with 𝜙−1 (𝜙(𝑈 ) ∩ 𝜓(𝑉 )) ⊂ 𝑈
and 𝑉 with 𝜓 −1 (𝜙(𝑈 ) ∩ 𝜓(𝑉 )) ⊂ 𝑉 . After doing this we can assume

𝜙(𝑈 ) = 𝜓(𝑉 ).

Then the map

𝜃 ∶= 𝜓 −1 ◦𝜙 ∶ 𝑈 → 𝑉
is a diffeomorphism, since it is the composite of two diffeomorphisms and the chain rule
implies that this yields a diffeomorphism as well. By definition of 𝜃, we have 𝜙 = 𝜓◦𝜃. Taking
derivatives yields by the chain rule

𝑑𝜙0 = 𝑑𝜓0 ◦𝑑𝜃0 .

This implies that the image of 𝑑𝜙0 is contained in the image of 𝑑𝜓0 :

𝐝𝜙𝟎 (ℝ𝐤 ) ⊆ 𝐝𝜓𝟎 (ℝ𝐤 ) in ℝ𝐍 .

By switching the roles of 𝜙 and 𝜓 in the argument, we also get:

𝐝𝜓𝟎 (ℝ𝐤 ) ⊆ 𝐝𝜙𝟎 (ℝ𝐤 ) in ℝ𝐍 .

Hence 𝑑𝜙0 (ℝ𝑘 ) = 𝑑𝜓0 (ℝ𝑘 ) in ℝ𝑁 . This shows that whatever local parametrization around 𝑥
we start with, the vector subspace 𝑇𝑥 (𝑋) ⊆ ℝ𝑁 is always the same. In mathematical terms we
say that 𝑇𝑥 (𝑋) is well-defined.

2.4.4 Some examples

Example 2.37 (Tangent spaces of the unit circle) Let 𝑝 = (𝑎, 𝑏) ∈ 𝕊1 be a point with
𝑏 > 0. A local parametrization around 𝑝 with 𝜙(0) = 𝑝 is given by
√
𝜙 ∶ (−𝜀, 𝜀) → 𝕊1 , 𝑡 → (𝑡 + 𝑎, 1 − (𝑡 + 𝑎)2 )
42 2.4. Tangent spaces and the derivative
for some small enough real number 𝜀 > 0. The derivative at 𝑡 is the linear map
( )
2
1
𝑑𝜙𝑡 ∶ ℝ → ℝ , 𝑑𝜙𝑡 = − √ 𝑡+𝑎 .
2 1−(𝑡+𝑎)

Hence
√ the image of ℝ under 𝑑𝜙0 in ℝ is the line spanned by (−𝑏, 𝑎) where we use
2

𝑏 = 1 − 𝑎2 .

We can extend this to dimension two:

Example 2.38 (Tangent spaces of the two-sphere 𝕊2 ) Let 𝑝 = (𝑎, 𝑏, 𝑐) be point on 𝕊2

which is not the north pole. Then we can use the stereographic projection 𝜙𝑁 ∶ ℝ2 → 𝕊2
as a local parametrization.a
Recall that
1 ( 2 2
)
𝜙𝑁 (𝑥, 𝑦) = 2𝑥, 2𝑦, 𝑥 + 𝑦 − 1 .
1 + 𝑥2 + 𝑦2
The derivative at (𝑥, 𝑦) is the linear map 𝑑𝜙𝑁 ∶ ℝ2 → ℝ3 which, with respect to the
standard bases of ℝ2 and ℝ3 , is given by the matrix

⎛1 − 𝑥2 + 𝑦2 −2𝑥𝑦 ⎞
2 ⎜ −2𝑥𝑦
𝑑(𝜙𝑁 )(𝑥,𝑦) = 1 + 𝑥2 − 𝑦2 ⎟ .
(1 + 𝑥2 + 𝑦2 )2 ⎜ ⎟
⎝ 2𝑥 2𝑦 ⎠

The image of ℝ2 under the linear map 𝑑(𝜙𝑁 )(𝑥,𝑦) is the tangent space 𝑇𝜙𝑁 (𝑥,𝑦) 𝕊2 . This
image is spanned by the two column vectors of the matrix 𝑑(𝜙𝑁 )(𝑥,𝑦) . Let us check that
we get the space we would have expected, i.e., the plane which is orthogonal to the
vector 𝜙𝑁 (𝑥, 𝑦):

) ⎛1 − 𝑥 + 𝑦 ⎞
2 2
(
2𝑥, 2𝑦, 𝑥 + 𝑦 − 1 ⋅ ⎜ −2𝑥𝑦 ⎟
2 2
⎜ ⎟
⎝ 2𝑥 ⎠
= 2𝑥(1 − 𝑥2 + 𝑦2 ) − 4𝑥𝑦2 + 2𝑥(𝑥2 + 𝑦2 − 1)
= 2𝑥 − 2𝑥3 + 2𝑥𝑦2 − 4𝑥𝑦2 + 2𝑥3 + 2𝑥𝑦2 − 2𝑥
= 0.

Similarly,

( ) ⎛ −2𝑥𝑦 ⎞
2𝑥, 2𝑦, 𝑥2 + 𝑦2 − 1 ⋅ ⎜1 + 𝑥2 − 𝑦2 ⎟
⎜ ⎟
⎝ 2𝑦 ⎠
= −4𝑥2 𝑦 + 2𝑦(1 + 𝑥2 − 𝑦2 ) + 2𝑦(𝑥2 + 𝑦2 − 1)
= −4𝑥2 𝑦 + 2𝑦 + 2𝑥2 𝑦 − 2𝑦3 + 2𝑥2 𝑦 + 2𝑦3 − 2𝑦
= 0.

Hence the plane spanned by the column vectors is orthogonal to 𝜙𝑁 (𝑥, 𝑦).
a
We do not have to translate first to get 𝜙𝑁 (0) = 𝑝. That is up to us.
Chapter 2. Smooth manifolds 43
∙ Open subsets and tangent spaces

Let 𝑋 ⊂ ℝ𝑁 be a 𝑘-dimensional manifold and 𝑊 be an open subset. Then 𝑊 is also a

𝑘-dimensional manifold, since we can restrict all local parametrizations to the intersection with
𝑊 (which is again open in 𝑋). In fact, for 𝑥 ∈ 𝑊 , let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization
around 𝑥 of 𝑋. We can assume 𝜙(0) = 𝑥. Then 𝜙|𝑊 ∩𝑈 ∶ 𝜙−1 (𝑊 ∩ 𝑈 ) → 𝑊 is a local
parametrization around 𝑥 of 𝑊 . Since the derivative only depends on an open neighborhood
around a point, we get 𝑑𝜙0 = 𝑑(𝜙|𝑊 ∩𝑈 )0 . In particular, for the tangent spaces at 𝑥, we get

𝑇𝑥 (𝑋) = 𝑑𝜙0 (ℝ𝑘 ) = 𝑑(𝜙|𝑊 ∩𝑈 )0 (ℝ𝑘 ) = 𝑇𝑥 (𝑊 )

as vector subspaces of ℝ𝑁 .

We summarize this discussion as a lemma:

Lemma 2.39 (Tangent spaces and open subsets) Let 𝑋 be a 𝑘-dimensional manifold
and 𝑊 be an open subset. At any point 𝑥 ∈ 𝑊 , we have

𝑇𝑥 (𝑋) = 𝑇𝑥 (𝑊 ).

A simple way to produce new manifolds is by taking products. We have already met an
important example of this construction: the two-dimensional torus 𝕋 2 = 𝕊1 × 𝕊1 . The tangent
space of such a product behaves as nicely as we can imagine:

Lemma 2.40 (Tangent space of a product) Given two smooth manifolds 𝑋 ⊆ ℝ𝑁

and 𝑌 ⊆ ℝ𝑀 and points 𝑥 ∈ 𝑋, 𝑦 ∈ 𝑌 , then the tangent space of the product 𝑋 and 𝑌
is the product of the tangent spaces, i.e.

𝑇(𝑥,𝑦) (𝑋 × 𝑌 ) = 𝑇𝑥 (𝑋) × 𝑇𝑦 (𝑌 ).

Proof: This follows from the fact that we can choose neighborhoods in 𝑋 × 𝑌 by taking
the product of neighborhoods in 𝑋 and 𝑌 , respectively. Moreover, it is easy to check that
𝑓 ∶ 𝑋 → 𝑋 ′ and 𝑔 ∶ 𝑌 → 𝑌 ′ are smooth maps, then the derivative of the product map is the
product of the derivatives, i.e.,

𝑑(𝑓 × 𝑔)(𝑥,𝑦) = 𝑑𝑓𝑥 × 𝑑𝑔𝑦

for all (𝑥, 𝑦) ∈ 𝑋 × 𝑌 .

2.4.5 The induced derivative

Now we will turn to the effect of smooth maps on tangent spaces. In fact, every smooth map
between manifolds induces a linear map between tangent spaces. These linear maps are very
useful.

Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map from a 𝑘-dimensional smooth manifold 𝑋 ⊆ ℝ𝑁 to an

𝑙-dimensional smooth manifold 𝑌 ⊆ ℝ𝑀 . We would like to define a map best linear approx-
imation of 𝑓 at a point (𝑥, 𝑓 (𝑥)). For 𝑦 = 𝑓 (𝑥), this should result in a linear map of vector
44 2.4. Tangent spaces and the derivative
spaces
𝑇𝑥 (𝑋) → 𝑇𝑦 (𝑌 ).

Suppose that 𝜙 ∶ 𝑈 → 𝑋 is a local parametrization around 𝑥 with 𝑈 ⊆ ℝ𝑘 , and 𝜓 ∶ 𝑉 → 𝑌

a local parametrization around 𝑦 with 𝑉 ⊆ ℝ𝑙 . We can assume 𝜙(0) = 𝑥 and 𝜓(0) = 𝑦. Then
we define a map 𝜃 ∶ 𝑈 → 𝑉 by the following commutative diagram:6

𝑓
𝑋O /𝑌
O
𝜙 𝜓 𝜓 −1

𝑈 /𝑉.
𝜃=𝜓 −1 ◦𝑓 ◦𝜙

Definition 2.41 (The derivative 𝑑𝑓𝑥 ) Taking derivatives yields a diagram of linear
maps and we define 𝑑𝑓𝑥 to be the linear map which makes the diagram commutative:

𝑑𝑓𝑥
𝑇𝑥 (𝑋) / 𝑇 (𝑌 )
O 𝑦
O
𝑑𝜙0 𝑑(𝜙−1 )𝑥 𝑑𝜓0

ℝ𝑘 / ℝ𝑙 .
𝑑𝜃0

Since 𝑑𝜙0 is an isomorphism, we can define 𝑑𝑓𝑥 as

𝐝𝐟𝐱 ∶= 𝐝𝜓𝟎 ◦𝐝𝜃𝟎 ◦𝐝(𝜙−𝟏 )𝐱 .

We call 𝑑𝑓𝑥 the derivative of 𝑓 at 𝑥.

Remark 2.42 (Why so complicated?) You may wonder why we need to take this
detour to define 𝑑𝑓𝑥 when we could also consider 𝑓 as a map of subsets of Euclidean
spaces and take the derivative of that map, since we assume 𝑋 ⊂ 𝑅𝑁 and 𝑌 ⊂ ℝ𝑀
for some 𝑁 and 𝑀 anyway. This works nicely if 𝑋 is an open subset in ℝ𝑁 . For if
𝑋 ⊂ ℝ𝑁 is open, we can choose 𝜙 as the identity map and have 𝑇𝑥 (𝑋) = 𝑇𝑥 (ℝ𝑁 ).
Then the derivative 𝑑𝐹𝑥 ∶ 𝑇𝑥 (𝑋) = 𝑇𝑥 (ℝ𝑁 ) = ℝ𝑁 → 𝑇𝑦 (𝑌 ) is actually the derivative
as a smooth map between Euclidean space.
However, if 𝑋 is not open in ℝ𝑁 we need to work a bit harder. By definition of
smoothness, there is an open subset 𝑊 ⊂ ℝ𝑁 and a smooth map 𝐹 ∶ 𝑊 → ℝ𝑀 such
that 𝐹|𝑊 ∩𝑋 = 𝑓|𝑊 ∩𝑋 . The derivative of 𝐹 at 𝑥 is a linear map 𝑑𝐹𝑥 ∶ ℝ𝑁 → ℝ𝑀 . But
what we want is a linear map defined on 𝑇𝑥 (𝑋) ⊂ ℝ𝑁 and with image in 𝑇𝑦 (𝑌 ) ⊂ ℝ𝑀 .
When we look at the gymnastics we do to define 𝑑𝑓𝑥 , we see that this is exactly what
we do: we restrict and adjust 𝑑𝐹𝑥 to the vector subspace 𝑇𝑥 (𝑋) ⊂ ℝ𝑁 such that it has
image in 𝑇𝑦 (𝑌 ). Thus, in the end, the seemingly complicated definition is just the linear
algebra necessary to assure that 𝑑𝑓𝑥 has the correct domain and codomain. We will see
later when we learn about regular values and transversal intersections that there is often
a short cut to make our first intuition work.

6
which means that it does not matter which way we walk around from 𝑈 to 𝑌 .
Chapter 2. Smooth manifolds 45
2.4.6 The derivative is well-defined:

We should check that the derivative is well-defined, i.e., that 𝑑𝑓𝑥 does not depend on the choice
of local parametrizations around 𝑥 and 𝑦 = 𝑓 (𝑥). So let 𝜙′ ∶ 𝑈 → 𝑋 and 𝜓 ′ ∶ 𝑉 ′ → 𝑌
be another choice of local parametrizations around 𝑥 and 𝑦, respectively. Again after possibly
shrinking both 𝑈 , 𝑈 ′ , 𝑉 and 𝑉 ′ we can assume that 𝜙(𝑈 ) = 𝜙′ (𝑈 ′ ) ⊆ 𝑋 and 𝜓(𝑉 ) = 𝜓 ′ (𝑉 ′ ) ⊆
𝑌.

Then 𝑑𝜙0 and 𝑑𝜙′0 differ by a linear isomorphism of ℝ𝑘 , say 𝛼: 𝑑𝜙0 = 𝑑𝜙′0 ◦𝛼. Similarly,
there is a linear isomorphism 𝛽 of ℝ𝑙 such that 𝑑𝜓0 = 𝑑𝜓0′ ◦𝛽. Let 𝜃 ′ ∶ 𝑈 → 𝑉 be defined
similarly to 𝜃, i.e., we set 𝜃 ′ = 𝜓 ′ −1 ◦𝑓 ◦𝜙′ . This gives us the following diagram in which each
square commutes

𝑑𝑓𝑥
𝑇𝑥 (𝑋) / 𝑇 (𝑌 )
@ O 𝑦
O ]
𝑑𝜙′0 𝑑𝜓0′

𝑑𝜙0 ℝO 𝑘 / ℝ𝑙 𝑑𝜓0
𝑑𝜃0′ O
𝛼 𝛽

ℝ𝑘 / ℝ𝑙 .
𝑑𝜃0

The relation between 𝑑𝜃0 and 𝑑𝜃0′ is given by

𝑑𝜃0′ = 𝛽◦𝑑𝜃0 ◦𝛼 −1 .

Putting all relations together we get

−1 −1
𝑑𝜓0′ ◦𝑑𝜃0′ ◦𝑑(𝜙′ )𝑥 = 𝑑𝜓0′ ◦(𝛽◦𝑑𝜃0 ◦𝛼 −1 )◦𝑑(𝜙′ )𝑥
−1
= (𝑑𝜓0′ ◦𝛽)◦𝑑𝜃0 ◦(𝛼 −1 )◦𝑑(𝜙′ )𝑥 )
= 𝑑𝜓0 ◦𝑑𝜃0 ◦𝑑(𝜙−1 )𝑥 .

This implies the desired identity for the a priori different constructions of 𝑑𝑓𝑥 .

Before we look at an example, we would like to know that the new derivative satisfies a
chain rule, since this is a very useful rule.

∙ The chain rule:

Let 𝑔 ∶ 𝑌 → 𝑍 be another smooth map. Let 𝜂 ∶ 𝑊 → 𝑍 be a local parametrization around

𝑧 = 𝑔(𝑦) with an open subset 𝑊 ⊆ ℝ𝑚 and 𝜂(0) = 𝑧. Then we have a commutative diagram

𝑓 𝑔
𝑋O /𝑌 /𝑍
O O
𝜙 𝜓 𝜂

𝑈 /𝑉 /𝑊
𝜃=𝜓 −1 ◦𝑓 ◦𝜙 𝜄=𝜂 −1 ◦𝑔◦𝜓
46 2.4. Tangent spaces and the derivative
which gives us the commutative square
𝑔◦𝑓
𝑋O /𝑍
O
𝜙 𝜂

𝑈 /𝑊.
𝜄◦𝜃

Thus, by definition,
𝑑(𝑔◦𝑓 )𝑥 = 𝑑𝜂0 ◦𝑑(𝜄◦𝜃)0 ◦𝑑(𝜙−1 )𝑥 .
The chain rule from Calculus for maps of open sets of Euclidean spaces, then gives
𝑑(𝜄◦𝜃)0 = (𝑑𝜄0 )◦(𝑑𝜃0 ).
Thus
𝑑(𝑔◦𝑓 )𝑥 = (𝑑𝜂0 ◦𝑑𝜄0 ◦𝑑(𝜓 −1 )𝑦 )◦(𝑑𝜓0 ◦𝑑𝜃0 ◦𝑑(𝜙−1 )𝑥 ) = 𝑑𝑔𝑦 ◦𝑑𝑓𝑥 .
Hence we have in fact the desired rule.

𝑓 𝑔
Theorem 2.43 (Chain Rule) If 𝑋 ←←←→ ← 𝑍 are smooth maps of manifolds, then
← 𝑌 ←←→

𝑑(𝑔◦𝑓 )𝑥 = 𝑑𝑔𝑓 (𝑥) ◦𝑑𝑓𝑥 .

2.4.7 Example: The Hopf fibration

We conclude this section with another important example and some concrete calculations. Re-
call the Hopf map 𝑓 ∶ 𝕊3 → 𝕊2 defined by

⎛𝑥1 ⎞ ⎛ 2𝑥 𝑥 + 2𝑥 𝑥 ⎞
⎜𝑥 ⎟ 1 3 2 4
𝑓 ∶ 𝕊3 → 𝕊2 , ⎜ 2 ⎟ → ⎜ 2𝑥2 𝑥3 − 2𝑥1 𝑥4 ⎟ .
⎜𝑥3 ⎟ ⎜⎝𝑥2 + 𝑥2 − 𝑥2 − 𝑥2 ⎟⎠
⎝𝑥4 ⎠ 1 2 3 4

We will now compute the derivative of 𝑓 at two points concrete points in 𝕊3 . Later we will be
able to appreciate the relevance of these computations and the choice of points much better. For
the moment, we consider this just as training and illustration.

First, let us pick a point which is mapped to the south pole 𝐬2 = (0, 0, −1) ∈ 𝕊2 . The
formula for 𝑓 shows that all points (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) ∈ 𝕊3 with 𝑥1 = 𝑥2 = 0 are mapped to 𝐬2 .
Hence, in particular, the north pole 𝐧3 = (0, 0, 0, 1) ∈ 𝕊3 is mapped to 𝐬2 = (0, 0, −1) ∈ 𝕊2 . So
let us look at the point 𝐧3 which is sent to 𝐬2 .

Since the formula for the stereographic projections can be quite involved when the number
of variables increases, we use the local parametrizations given as the inverse of the projection
onto the first coordinates. We used them in Example
√ 2.26 for the 𝕊 . We choose the open ball
1

𝔹3 (𝟎3 ) around the origin 𝟎3 in ℝ3 of radius7 1∕ 2 and use the local parametrization
( √ )
𝜙 ∶ 𝑈 = 𝔹3 √ (𝟎3 ) → 𝑊3 , (𝑥, 𝑦, 𝑧) → 𝑥, 𝑦, 𝑧, 1 − (𝑥2 + 𝑦2 + 𝑧2 )
1∕ 2
7
The choice of the radius will become apparent in a minute when we compute 𝑓 ◦𝜙 which needs to have image
in the subset of points in 𝕊2 with 𝑥3 < 0.
Chapter 2. Smooth manifolds 47
where 𝑊3 ⊂ 𝕊3𝑥 ⊂ 𝕊3 denotes the open subset of 𝕊3 consisting of points with coordinate
4 >0
𝑥4 > 0 and 2𝑥21 + 2𝑥22 < 1. Note that 𝜙 maps 𝟎3 to 𝐧3 . Similarly, we choose the open ball
𝔹21 (𝟎2 ) around the origin 𝟎2 in ℝ2 of radius 1 and use the local parametrization
( √ )
𝜓 ∶ 𝑉 = 𝔹21 (𝟎2 ) → 𝑊2 , (𝑥, 𝑦) → 𝑥, 𝑦, − 1 − (𝑥2 + 𝑦2 )

where 𝑊2 = 𝕊2𝑥 <0 ⊂ 𝕊2 denotes the open subset of 𝕊2 consisting of points with coordinate
3
𝑥3 < 0. Note that 𝜓 maps 𝟎2 to 𝐬2 .

Now we need to calculate the induced map 𝜃 ∶ ℝ3 → ℝ2 such that the diagram commutes
𝑓
𝑊3 /𝑊
O O2
𝜙 𝜓 𝜓 −1

𝔹3 √ (𝟎3 ) / 𝔹2 (𝟎 ).
1 2
1∕ 2 𝜃=𝜓 −1 ◦𝑓 ◦𝜙

We calculate the effect of the maps step by step. Recall that we write |𝐱| for the norm of
points 𝐱 ∈ ℝ𝑘 . First we apply 𝜙 to a point in 𝔹3 (𝟎3 ):

⎛ 𝑥 ⎞
⎛𝑥⎞ ⎜ 𝑦 ⎟
𝜙 ∶ 𝐱 = ⎜𝑦⎟ → ⎜ ⎟.
⎜ ⎟ ⎜√ 𝑧 ⎟
⎝𝑧⎠ ⎜ ⎟
⎝ 1 − |𝐱| ⎠
2

Next we apply the composite 𝑓 ◦𝜙 to a point in 𝑈 :

√
⎛𝑥⎞ ⎛ 2𝑥𝑧 + 2𝑦√1 − |𝐱|2 ⎞
𝑓 ◦𝜙 ∶ 𝐱 = ⎜𝑦⎟ → ⎜ 2𝑦𝑧 − 2𝑥 1 − |𝐱|2 ⎟ .
⎜ ⎟ ⎜ 2 ⎟
⎝𝑧⎠ ⎝𝑥 + 𝑦2 − 𝑧2 − 1 + |𝐱|2 ⎠

Recall that the inverse 𝜓 −1 is just the projection onto the first two coordinates. Hence applying
the composite 𝜓 −1 ◦𝜙 to a point in ℝ3 yields:
( √ )
⎛𝑥⎞
2𝑥𝑧 + 2𝑦 1 − |𝐱| 2
𝜓 −1 ◦𝑓 ◦𝜙 ∶ 𝐱 = ⎜𝑦⎟ → √ .
⎜ ⎟ 2𝑦𝑧 − 2𝑥 1 − |𝐱|2
⎝ ⎠
𝑧

This is the map 𝜓 −1 ◦𝑓 ◦𝜙 ∶ 𝑈 → 𝑉 .8

Now we need to apply this result to compute the horizontal map 𝑑𝑓𝑛3 such that following
diagram for tangent spaces commutes:
𝑑𝑓𝐧3
𝑇𝐧3 (𝕊3 ) / 𝑇 (𝕊2 )
𝐬2
O O
𝑑𝜙𝟎3 𝑑(𝜙−1 )𝐧3 𝑑𝜓𝟎2

ℝ3 / ℝ2 .
𝑑𝜃𝟎3

8
A good reality check is that 𝜃 does map the origin 𝟎3 to the origin 𝟎2 as we claimed.
48 2.4. Tangent spaces and the derivative
In particular, we need to calculate the derivative 𝑑𝜃𝟎3 ∶ ℝ3 → ℝ2 at the origin 𝟎3 . We do
this by first computing the Jacobian matrix 𝐽𝜃 of 𝜃:
√ 2𝑦2
⎛ 2𝑧 − √ 2𝑥𝑦 2 2 1 − |𝐱|2 − √ 2𝑥 − √
2𝑦𝑧 ⎞

𝐽𝜃 = ⎜ √ 1−|𝐱|
2
1−|𝐱|2 1−|𝐱|2 ⎟
.
⎜−2 1 − |𝐱|2 + √ 2𝑥 2𝑧 + √
2𝑥𝑦
2𝑦 + √
2𝑥𝑧 ⎟
⎝ 1−|𝐱|2 1−|𝐱|2 1−|𝐱|2 ⎠

This looks annoyingly complicated. However, there is good news. We want to calculate the
matrix representing 𝑑𝜃𝟎3 at the origin. Hence we set 𝑥 = 𝑦 = 𝑧 = 0 and see that 𝑑𝜃𝟎3 ∶ ℝ3 → ℝ2
is given by the matrix
( )
0 2 0
𝑑𝜃𝟎3 =
−2 0 0

with respect to the standard bases of ℝ3 and ℝ2 .

By our definition of 𝑇𝐧3 (𝕊3 ) as the image of 𝑑𝜙𝟎3 ∶ ℝ3 → ℝ4 . As a basis of 𝑇𝐧3 (𝕊3 ) we
can hence choose the images of the standard basis 𝐞31 , 𝐞32 , 𝐞33 of ℝ3 under 𝑑𝜙𝟎3 . With respect this
basis 𝑑𝜙𝟎3 (𝐞31 ), 𝑑𝜙𝟎3 (𝐞32 ), 𝑑𝜙𝟎3 (𝐞33 ) for 𝑇𝐧3 (𝕊3 ) and the standard basis for ℝ3 , the derivative

𝑑(𝜙−1 )𝐧3 ∶ 𝑇𝐧3 (𝕊3 ) → ℝ3

is represented by the 3 × 3-identity matrix.

Similarly, for 𝑇𝐬2 (𝕊2 ) we can choose the image of the standard basis 𝐞21 , 𝐞22 of ℝ2 under
𝑑𝜓𝟎2 . With respect to the standard basis for ℝ2 and the basis 𝑑𝜓𝟎2 (𝐞21 ), 𝑑𝜓𝟎2 (𝐞22 ) for 𝑇𝐬2 (𝕊2 ) the
derivative
𝑑𝜓𝟎2 ∶ ℝ2 → 𝑇𝐬2 (𝕊2 )

is represented by the 2 × 2-identity matrix.

Hence — with respect to these bases for 𝑇𝐧3 (𝕊3 ) and 𝑇𝐬2 (𝕊2 ) — we see that the composition

𝑑𝑓𝐧3 = 𝑑𝜓𝟎2 ◦𝑑𝜃𝟎3 ◦𝑑(𝜙−1 )𝐧3

is given by the matrix

( )
0 2 0
𝑑𝑓𝐧3 = .
−2 0 0

Note that our choices of bases make it very easy to compute the matrix for 𝑑𝑓𝐧3 . So in the
end, there is not as much to compute as one might fear. To make things even more explicit we
observe that 𝑑𝑓𝐧3 has the effect on the basis vectors:

𝑑𝜙𝟎3 (𝐞31 ) → 0 ⋅ 𝑑𝜓𝟎2 (𝐞21 ) − 2 ⋅ 𝑑𝜓𝟎2 (𝐞22 ),

𝑑𝜙𝟎3 (𝐞32 ) → 2 ⋅ 𝑑𝜓𝟎2 (𝐞21 ) + 0 ⋅ 𝑑𝜓𝟎2 (𝐞22 ), and
𝑑𝜙𝟎3 (𝐞33 ) → 0 ⋅ 𝑑𝜓𝟎2 (𝐞21 ) + 0 ⋅ 𝑑𝜓𝟎2 (𝐞22 ).
Chapter 2. Smooth manifolds 49

Remark 2.44 (Something we learn from this example) Among other things we see
in this example that — once we have computed 𝜃 — there is a straight-forward way to
compute a matrix which describes 𝑑𝑓𝑥 . For, in this setting, there is a canonical choice
for the bases of 𝑇𝑥 (𝑋) and 𝑇𝑦 (𝑌 ): the images of the standard basis vectors of ℝ𝑛 and ℝ𝑚
under the isomorphisms 𝑑𝜙0 and 𝑑𝜓0 , respectively.
Then we can compute the matrix which represents the linear map 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) →
𝑇𝑦 (𝑌 ) with respect to these bases just as the matrix which represents 𝑑𝜃0 with respect to
the standard bases of ℝ𝑛 and ℝ𝑚 . And we get this matrix as the Jacobian matrix of 𝜃 at
the origin.

Second, let us apply what we just learned and practice a bit more. So let us look at a point
on 𝕊3 which is mapped to the north pole 𝐧2 = (0, 0, 1) ∈ 𝕊2 . The formula for 𝑓 shows that all
points (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) ∈ 𝑆 3 with 𝑥3 = 𝑥4 = 0 are mapped to 𝐧2 . Hence, in particular, the point
𝐪3 = (1, 0, 0, 0) ∈ 𝑆 3 is mapped to 𝐧2 = (0, 0, 1) ∈ 𝕊2 . So let us pick that 𝐪3 mapping to 𝐧2 .

Much of the calculation is the same as in the previous case. However, there are some in-
teresting changes, in particular, of some signs. Again, we will be able to appreciate this more
later.
√
We choose the open ball 𝔹3 (𝟎3 ) around the origin 𝟎3 in ℝ3 of radius 1∕ 2 and use the local
parametrization
(√ )
𝜙 ∶ 𝑈 = 𝔹3 √ (𝟎3 ) → 𝑊3 , (𝑥, 𝑦, 𝑧) → 1 − (𝑥2 + 𝑦2 + 𝑧2 ), 𝑥, 𝑦, 𝑧
1∕ 2

where 𝑊3 ⊂ 𝕊3𝑥 ⊂ 𝕊3 denotes the open subset of 𝕊3 consisting of points with coordinate
1 >0
𝑥1 > 0 and 2𝑥23 + 2𝑥24 < 1. We choose the open ball 𝔹21 (𝟎2 ) around the origin 𝟎2 in ℝ2 of radius
1 and use the local parametrization
( √ )
𝜓 ∶ 𝑉 = 𝔹21 (𝟎2 ) → 𝑊2 , (𝑥, 𝑦) → 𝑥, 𝑦, 1 − (𝑥2 + 𝑦2 )

where 𝑊2 ⊂ 𝕊2𝑥 denotes the open subset of 𝕊2 consisting of points with coordinate 𝑥3 > 0.
3 >0

We need to calculate the induced map 𝜃 ∶ ℝ3 → ℝ2 such that the diagram commutes
𝑓
𝑊3 /𝑊
O O2
𝜙 𝜓 𝜓 −1

𝔹3 √ (𝟎3 ) / 𝔹2 (𝟎 ).
1∕ 2 𝜃=𝜓 −1 ◦𝑓 ◦𝜙 1 2

We calculate the effect of the maps step by step. Recall that we write |𝐱| for the norm of
points 𝐱 ∈ ℝ𝑘 . First we apply 𝜙 to a point in 𝔹3 (𝟎3 ):
√
⎛ 1 − |𝐱|2 ⎞
⎛𝑥⎞ ⎜ ⎟
𝑥
𝜙 ∶ 𝐱 = ⎜𝑦⎟ → ⎜ ⎟.
⎜ ⎟ ⎜ 𝑦 ⎟
⎝𝑧⎠ ⎜ ⎟
⎝ 𝑧 ⎠
50 2.4. Tangent spaces and the derivative
Next we apply the composite 𝑓 ◦𝜙:
√
⎛𝑥⎞ ⎛ 2𝑦 1 −√ |𝐱|2 + 2𝑥𝑧 ⎞
𝑓 ◦𝜙 ∶ 𝐱 = ⎜𝑦⎟ → ⎜ 2𝑥𝑦 − 2 1 − |𝐱|2 𝑧 ⎟
⎜ ⎟ ⎜ ⎟
⎝𝑧⎠ ⎝1 − |𝐱|2 + 𝑥2 − 𝑦2 − 𝑧2 ⎠
√
⎛2𝑦 1 −√|𝐱|2 + 2𝑥𝑧⎞
= ⎜2𝑥𝑦 − 2 1 − |𝐱|2 𝑧⎟ .
⎜ ⎟
⎝ 1 − 2𝑦2 − 2𝑧2 ⎠

Applying the composite 𝜓 −1 ◦𝜙 to a point in ℝ3 yields:

( √ )
⎛𝑥⎞
2𝑦 1 − |𝐱|2 + 2𝑥𝑧
𝜓 −1
◦𝑓 ◦𝜙 ∶ 𝐱 = ⎜𝑦⎟ → √ .
⎜ ⎟ 2𝑥𝑦 − 2 1 − |𝐱|2 𝑧
⎝𝑧⎠

This is the map 𝜓 −1 ◦𝑓 ◦𝜙 ∶ 𝑈 → 𝑉 .

Now we need to apply this result to compute the horizontal map 𝑑𝑓𝑞3 such that following
diagram for tangent spaces commutes:
𝑑𝑓𝐪3
𝑇𝐪3 (𝕊3 ) / 𝑇 (𝕊2 )
𝐧2
O O
𝑑𝜙𝟎3 𝑑(𝜙−1 )𝐪3 𝑑𝜓𝟎2

ℝ3 / ℝ2 .
𝑑𝜃𝟎3

In particular, we need to calculate the derivative 𝑑𝜃𝟎3 ∶ ℝ3 → ℝ2 at the origin 𝟎3 . We do

this again by computing the Jacobian matrix 𝐽𝜃 of 𝜃:
√ 2
⎛2𝑧 − √ 2𝑥𝑦 2 1 − |𝐱| 2 − √ 2𝑦 2𝑥 − √ 2𝑦𝑧 2 ⎞
𝐽𝜃 = ⎜ ⎟.
2 2
1−|𝐱| 1−|𝐱|
2𝑧2
√
1−|𝐱|
⎜2𝑦 + √ 2𝑥𝑧 2𝑥 + √
2𝑦𝑧
√ − 2 1 − 𝐱| 2⎟
⎝ 2
1−|𝐱| 2 2
1−|𝐱| ⎠
1−|𝐱|

At the origin we set 𝑥 = 𝑦 = 𝑧 = 0 and see that 𝑑𝜃𝟎3 ∶ ℝ3 → ℝ2 is given by the matrix
( )
0 2 0
𝑑𝜃𝟎3 =
0 0 −2

with respect to the standard bases of ℝ3 and ℝ2 , respectively.

Now we make our standard choice of bases of 𝑇𝐪3 (𝕊3 ), i.e., 𝑑𝜙𝟎3 (𝐞31 ), 𝑑𝜙𝟎3 (𝐞32 ), 𝑑𝜙𝟎3 (𝐞33 ),
and of 𝑇𝐧2 (𝕊2 ), i.e., 𝑑𝜓𝟎2 (𝐞21 ), 𝑑𝜓𝟎2 (𝐞22 ). Then — with respect to these bases for 𝑇𝐪3 (𝕊3 ) and
𝑇𝐧2 (𝕊2 ) — the composition

𝑑𝑓𝐪3 = 𝑑𝜓𝟎2 ◦𝑑𝜃𝟎3 ◦𝑑(𝜙−1 )𝐪3

is given by the matrix

( )
0 2 0
𝑑𝑓𝐪3 = .
0 0 −2
Chapter 2. Smooth manifolds 51
In this case, we observe that 𝑑𝑓𝐪3 has the effect on the basis vectors that it sends 𝑑𝜙𝟎3 (𝐞31 ) to
the zero vector, 𝑑𝜙𝟎3 (𝐞32 ) to the vector 2 ⋅ 𝑑𝜓𝟎2 (𝐞21 ) + 0 ⋅ 𝑑𝜓𝟎2 (𝐞22 ), and 𝑑𝜙𝟎3 (𝐞33 ) to 0 ⋅ 𝑑𝜓𝟎2 (𝐞21 ) −
2 ⋅ 𝑑𝜓𝟎2 (𝐞22 ).

Remark 2.45 (Outlook to orientations) We will appreciate this example even more
when we have learned about orientations. For the above computation shows that 𝑑𝑓𝐪3
sends a positively oriented basis to a negatively oriented basis. In other words, 𝑑𝑓𝐪3
reverses orientations. More on this later.

2.5 Tangent Bundle

In this section we look at yet another example of an interesting space, the tangent bundle, which
is formed by the collection of tangent spaces of a given smooth manifold 𝑋. We will see that
it is itself a smooth manifold. Later on we will learn that it tells us quite a lot about the
geometry of 𝑋. Moreover, the tangent bundle will turn out to be an extremely useful tool for
many constructions. See for example Section 9.5.

Advice: The reader who may not feel comfortable yet with tangent spaces and what they
are may want to skip this section first and get back to it later when we use the tangent bundle.

2.5.1 Tangent Bundle - the definition

Let 𝑋 ⊂ ℝ𝑁 be a smooth manifold. For every 𝑥 ∈ 𝑋, the tangent space 𝑇𝑥 (𝑋) to 𝑋 at 𝑥 is a

vector subspace of ℝ𝑁 . If we let 𝑥 vary, these tangent space will in general overlap in ℝ𝑁 . For
example, if 𝑋 is a vector space itself, they will all be equal.

Hence in order to be able to keep track of the information contained in all the different
tangent spaces we need a smart device that keeps those spaces apart:

Definition 2.46 (Tangent bundle) The tangent bundle of 𝑋, denoted 𝑇 (𝑋), is the
subset of 𝑋 × ℝ𝑁 defined by

𝑇 (𝑋) ∶= {(𝑥, 𝑣) ∈ 𝑋 × ℝ𝑁 ∶ 𝑣 ∈ 𝑇𝑥 (𝑋)}.

In particular, 𝑇 (𝑋) contains a natural copy of 𝑋 consisting of the points (𝑥, 0).
In the direction perpendicular to 𝑋0 , it contains copies of each tangent space 𝑇𝑥 (𝑋)
embedded as the sets

{(𝑥, 𝑣) ∈ 𝑇 (𝑋) ∶ for a fixed 𝑥}.

There is a natural projection map

𝜋 ∶ 𝑇 (𝑋) → 𝑋, (𝑥, 𝑣) → 𝑥.
52 2.5. Tangent Bundle
Any smooth map 𝑓 ∶ 𝑋 → 𝑌 induces a global derivative map

𝑑𝑓 ∶ 𝑇 (𝑋) → 𝑇 (𝑌 ), (𝑥, 𝑣) → (𝑓 (𝑥), 𝑑𝑓𝑥 (𝑣)).

Note that, since 𝑋 ⊂ ℝ𝑁 and 𝑇𝑥 (𝑋) ⊂ ℝ𝑁 for every 𝑥, 𝑇 (𝑋) is also a subset of Euclidean
space:
𝑇 (𝑋) ⊂ ℝ𝑁 × ℝ𝑁 .
Therefore, if 𝑌 ⊂ ℝ𝑀 , then 𝑑𝑓 maps a subset of ℝ2𝑁 to ℝ2𝑀 .

Claim: 𝑑𝑓 is smooth.

Proof: Since 𝑓 ∶ 𝑋 → ℝ𝑀 is smooth, it extends by definition around any point 𝑥 ∈ 𝑋 to

a smooth map 𝐹 ∶ 𝑈 → ℝ𝑀 , where 𝑈 is an open set of ℝ𝑁 . Then 𝑑𝐹 ∶ 𝑇 (𝑈 ) → ℝ2𝑀 locally
extends 𝑑𝑓 . However, since 𝑈 ⊂ ℝ𝑁 is open and hence 𝑇𝑢 (𝑈 ) = ℝ𝑁 for every 𝑢 ∈ 𝑈 , 𝑇 (𝑈 )
is all of 𝑈 × ℝ𝑁 . Since 𝑈 × ℝ𝑁 is an open set in ℝ2𝑁 , 𝑑𝐹 is a linear and hence smooth
map defined on an open subset of ℝ2𝑁 . This shows that 𝑑𝑓 ∶ 𝑇 (𝑋) → ℝ2𝑀 may be locally
extended to a smooth map on an open subset of ℝ2𝑁 , meaning that 𝑑𝑓 is smooth.

We can also say something about the derivative of the composition of smooth maps: Given
smooth maps 𝑓 ∶ 𝑋 → 𝑌 and 𝑔 ∶ 𝑌 → 𝑍, the global derivative of the composite is equal to
the composite of global derivatives:

𝑑(𝑔◦𝑓 ) = 𝑑𝑔◦𝑑𝑓 ∶ 𝑇 (𝑋) → 𝑇 (𝑍).

For, the chain rule implies that, for any (𝑥, 𝑣) ∈ 𝑇 (𝑋),

𝑑(𝑔◦𝑓 )(𝑥, 𝑣) = ((𝑔◦𝑓 )(𝑥), 𝑑(𝑔◦𝑓 )𝑥 (𝑣)

= ((𝑔(𝑓 (𝑥)), (𝑑𝑔𝑓 (𝑥) ◦𝑑𝑓𝑥 )(𝑣))
= 𝑑𝑔(𝑑𝑓 (𝑥, 𝑣))
= 𝑑𝑔◦𝑑𝑓 (𝑥, 𝑣).

Now, if 𝑓 ∶ 𝑋 → 𝑌 is a diffeomorphism, then 𝑑𝑓 ∶ 𝑇 (𝑋) → 𝑇 (𝑌 ) is a diffeomorphism:

Theorem 2.47 (Tangent bundles are intrinsic) Diffeomorphic manifolds have dif-
feomorphic tangent bundles. As a result, 𝑇 (𝑋) is an object intrinsically associated to
𝑋, i.e., it does not depend on the ambient Euclidean space.

2.5.2 Tangent bundles are manifolds

Finally, we are going to show that 𝑇 (𝑋) is in fact itself a smooth manifold. We will use this
fact in the proof of Whitney’s theorem below.

Let 𝑊 be an open set of 𝑋. In particular, 𝑊 is also a manifold, and we can consider its
tangent bundle 𝑇 (𝑊 ). Since 𝑇𝑥 (𝑊 ) = 𝑇𝑥 (𝑋) for every 𝑥 ∈ 𝑊 , 𝑇 (𝑊 ) is by definition

𝑇 (𝑊 ) = {(𝑥, 𝑣) ∈ 𝑇 (𝑋) ∶ 𝑥 ∈ 𝑊 } = 𝑇 (𝑋) ∩ (𝑊 × ℝ𝑁 ) ⊂ 𝑇 (𝑋).

Chapter 2. Smooth manifolds 53
Since 𝑊 × ℝ𝑁 is open in 𝑋 × ℝ𝑁 , 𝑇 (𝑊 ) is open in 𝑇 (𝑋).

Now suppose that 𝑊 is the image of a local parametrization 𝜙 ∶ 𝑈 → 𝑊 , where 𝑈 is

an open set in ℝ𝑘 . Then the global derivative 𝑑𝜙 ∶ 𝑇 (𝑈 ) → 𝑇 (𝑊 ) is a diffeomorphism. But
𝑇 (𝑈 ) = 𝑈 × ℝ𝑘 is an open subset of ℝ2𝑘 , so 𝑑𝜙 is a parametrization of the open set 𝑇 (𝑊 )
in 𝑇 (𝑋). Since every point of 𝑇 (𝑋) sits in such a neighborhood, we have proved the following
useful result:

Theorem 2.48 (Tangent bundles are manifolds) The tangent bundle of a manifold
𝑋 is a smooth manifold of dimension dim 𝑇 (𝑋) = 2 dim 𝑋.

In the next section, we are going to use the tangent bundle as a tool to construct new maps.
The key will be that the tangent bundle gives us extra space for manoeuvring. But before we
do this we take a brief detour.

2.5.3 Tangent bundles are vector bundles

Tangent bundles are examples of a more general class of spaces, called vector bundles. They
can be defined on any topological space. But let us assume we have a manifold 𝑋. Roughly
speaking, an 𝑛-dimensional vector bundle 𝐸 consists of two data:

∙ an assignment of an 𝑛-dimensional vector space 𝑉 to each point 𝑥 ∈ 𝑋;

∙ a rule for how to glue all these vector spaces together in a nice way.

More precisely, an 𝑛-dimensional vector bundle 𝐸 over 𝑋 consists of a topological space

𝐸 together with a continuous map 𝜋 ∶ 𝐸 → 𝑋 which satisfy the following condition:

∙ for every point 𝑥 ∈ 𝑋 there is an open subset 𝑈 ⊂ 𝑋 around 𝑥 and a homeomorphism

ℎ ∶ 𝑈 × ℝ𝑛 → 𝜋 −1 (𝑈 ) such that for every 𝑦 ∈ 𝑋 the map 𝑣 → (𝑦, 𝑣) defines a linear
isomorphism between the vector space ℝ𝑛 and 𝜋 −1 (𝑦).

If it is possible to choose the open subset around 𝑥 in the above condition to be all of 𝑋,
then we call 𝐸 → 𝑋 a trivial bundle.

We can refine this definition and say that 𝐸 → 𝑋 is a smooth vector bundle if we require
in addition that

∙ 𝐸 is a smooth manifold

∙ 𝜋 ∶ 𝐸 → 𝑋 is a smooth map

∙ each of the ℎ above is a diffeomorphism.

The tangent bundle is an important example a smooth vector bundle.

54 2.5. Tangent Bundle
Vector bundles have a rich and very interesting theory and many problems can bee formu-
lated in terms of vector bundles. For example, we will see in a later chapter that there is a nice
classifying space for vector bundles, called the Grassmannian. See Section 9.4 for more about
this important object. The idea of a classifying space is extremely powerful and we will not be
able to appreciate it in this course. However, we encourage everybody to continue reading in
this.

We conclude the detour with a famous example of a problem which can be phrased in terms
of vector bundles:

Remark 2.49 (Parallelizable spheres) A manifold for which the tangent bundle is
trivial is called parallelizable. Examples of manifolds which are parallelizable are 𝕊1 ,
𝕊3 and 𝕊7 , whereas 𝕊2 is not parallelizable. In fact, it is a famous and deep result that 𝕊𝑛
is parallelizable if and only if 𝑛 = 0, 1, 3 or 7. This is a consequence of the famous and
fundamental result on the possible multiplicative structures on ℝ𝑛 . For the above
statement follows from: Let ℝ𝑛 × ℝ𝑛 → ℝ𝑛 be a map with two-sided identity element
and no zero-divisors. Then 𝑛 must be either 1, 2, 4 or 8.
Chapter 2. Smooth manifolds 55
2.6 Exercises and more examples

2.6.1 Smooth maps and manifolds

Exercise 2.1 Consider the map 𝑓 ∶ ℝ2 → ℝ2 , (𝑥, 𝑦) → (𝑥2 − 𝑦2 , 2𝑥𝑦).

(a) Check that 𝑓 is smooth by calculating the partial derivatives.

(b) Show that the Jacobian matrix of the restriction

𝑓|𝑈 ∶ 𝑈 = ℝ2 ⧵ {(0, 0)} → ℝ2 ⧵ {(0, 0)}, (𝑥, 𝑦) → (𝑥2 − 𝑦2 , 2𝑥𝑦), is invertible for
every point in 𝑈 .

(c) Is 𝑓|𝑈 a diffeomorphism?

Exercise 2.2 Let 𝑋 ⊆ ℝ𝑁 , 𝑌 ⊆ ℝ𝑀 and 𝑍 ⊆ ℝ𝐿 be arbitrary subsets, and let

𝑓 ∶ 𝑋 → 𝑌 , 𝑔 ∶ 𝑌 → 𝑍 be smooth maps with 𝑓 (𝑋) ⊆ 𝑌 .

(a) Show the composite 𝑔◦𝑓 ∶ 𝑋 → 𝑍 is smooth if 𝑔 and 𝑓 are smooth.

Hint: If all subsets are open, this is just the Chain Rule from Calculus.

(b) Show that if 𝑔 and 𝑓 are diffeomorphisms, so is 𝑔◦𝑓 .

Exercise 2.3 For a real number 𝑟 > 0, let 𝔹𝑟 = 𝔹𝑘𝑟 (0) = {𝑥 ∈ ℝ𝑘 ∶ |𝑥| < 𝑟} be the
open ball around the origin with radius 𝑟 in ℝ𝑘 .

(a) Show that the map

𝑟𝑥
𝑓 ∶ 𝔹𝑟 → ℝ𝑘 , 𝑥 → √
𝑟 − |𝑥|2
2

is a diffeomorphism from 𝔹𝑟 to ℝ𝑘 .
Hint: Compute the inverse directly, and use the previous exercise to show smooth-
ness.

(b) Suppose that 𝑋 is a 𝑘-dimensional manifold. Show that every point in 𝑋 has a
neighborhood diffeomorphic to an open ball in ℝ𝑘 around the origin.

(c) Suppose that 𝑋 is a 𝑘-dimensional manifold. Show that every point in 𝑋 has a
neighborhood diffeomorphic to all of ℝ𝑘 .

Exercise 2.4 Show that every 𝑘-dimensional vector subspace 𝑉 of ℝ𝑁 is a manifold

diffeomorphic to ℝ𝑘 and that any linear map 𝑉 → ℝ𝑚 is smooth.

Comment: Recall that choosing a basis for 𝑉 corresponds to choosing a linear iso-
morphism 𝜙 ∶ ℝ𝑘 → 𝑉 . Expressing a vector in 𝑉 in terms of this basis means to attach
coordinates to this vector. Since 𝜙 is linear, we refer to the corresponding coordinates as
linear coordinates.
56 2.6. Exercises and more examples

Exercise 2.5 Recall the hyperboloid and the cone drawn in Figure 2.10:

(a) Prove that the subspace of ℝ3 , defined by 𝑥2 + 𝑦2 − 𝑧2 = 𝑎, is a manifold if 𝑎 > 0.

(b) Explain why 𝑥2 + 𝑦2 − 𝑧2 = 0 does not define a manifold.

Exercise 2.6 The torus 𝕋 (𝑎, 𝑏) is the set of points in ℝ3 at distance 𝑏 from the circle
of radius 𝑎 in the 𝑥𝑦-plane, where 0 < 𝑏 < 𝑎. Prove that each 𝕋 (𝑎, 𝑏) is diffeomorphic to
𝕊1 × 𝕊1 ⊂ ℝ4 . What happens when 𝑏 = 𝑎?

Exercise 2.7 Let 𝑁 = (0, … , 0, 1) ∈ 𝕊𝑘 be the ‘north pole’ on the 𝑘-dimensional

sphere. The stereographic projection 𝜙−1 𝑁
from 𝕊𝑘 ⧵ {𝑁} onto ℝ𝑘 is the map which
sends a point 𝑝 to the point at which the line through 𝑁 and 𝑝 intersects the subspace in
ℝ𝑘+1 defined by 𝑥𝑘+1 = 0. See Figure 2.12 for the case 𝑘 = 2.

(a) Show that 𝜙−1

𝑁
is given by the formula

1
(𝑥1 , … , 𝑥𝑘+1 ) → (𝑥 , … , 𝑥𝑘 ).
1 − 𝑥𝑘+1 1

(b) Find a formula for the inverse 𝜙𝑁 of 𝜙−1

𝑁
, and check that both maps are smooth.

(c) Let 𝑆 = (0, … , 0, −1) ∈ 𝕊𝑘 be the ‘south pole’. Describe the parametrization
using the stereographic projection starting in 𝑆 instead of 𝑁, and conclude that 𝕊𝑘
is a 𝑘-dimensional manifold.

Exercise 2.8 We consider 𝕊3 as a subset of ℂ2 , i.e., 𝕊3 = {(𝑧0 , 𝑧1 ) ∈ ℂ2 ∶ |𝑧0 |2 +

|𝑧1 |2 = 1}, and 𝕊2 as a subset of ℂ × ℝ, i.e., 𝕊2 = {(𝑧, 𝑥) ∈ ℂ × ℝ ∶ |𝑧|2 + 𝑥2 = 1}.
Then the Hopf map 𝜋 is the map 𝕊3 → 𝕊2 given by
( )
𝜋(𝑧0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2 .

(a) Check that this actually defines a map 𝕊3 → 𝕊2 .

(b) Show that 𝜋(𝑧0 , 𝑧1 ) = 𝜋(𝑤0 , 𝑤1 ) if and only if there is a complex number 𝛼 with
|𝛼|2 = 𝛼 𝛼̄ = 1 such that (𝑤0 , 𝑤1 ) = (𝛼𝑧0 , 𝛼𝑧1 ).

(c) Show that, for every point 𝑝 ∈ 𝕊2 , the fiber 𝜋 −1 (𝑝) is diffeomorphic to 𝕊1 .

We conclude that the Hopf map 𝜋 realizes 𝕊3 as a disjoint union of fibers which each
look like 𝕊1 .
Chapter 2. Smooth manifolds 57
2.6.2 Tangent spaces

Exercise 2.9 Let 𝑉 be a vector subspace of ℝ𝑁 . Show that 𝑇𝑥 (𝑉 ) = 𝑉 for 𝑥 ∈ 𝑉 .

Exercise 2.10 Determine the tangent space to the torus 𝕊1 × 𝕊1 ⊂ ℝ4 at an arbitrary

point 𝑝. Recall the description of the torus 𝕋 (𝑎, 𝑏) ⊂ ℝ3 from the previous exercise set.
Can you describe the tangent space at a point in 𝕋 (𝑎, 𝑏)?

√ Determine the tangent space to the subspace of ℝ defined by 𝑥 + 𝑦 −

Exercise 2.11 3 2 2

𝑧2 = 𝑎 at ( 𝑎, 0, 0) for 𝑎 > 0.

Exercise 2.12 The graph of a map 𝑓 ∶ 𝑋 → 𝑌 is the subset of 𝑋 × 𝑌 defined by

Γ(𝑓 ) = {(𝑥, 𝑓 (𝑥)) ∈ 𝑋 × 𝑌 ∶ 𝑥 ∈ 𝑋}.

Define 𝐹 ∶ 𝑋 → Γ(𝑓 ) by 𝐹 (𝑥) = (𝑥, 𝑓 (𝑥)). We assume that 𝑋 and 𝑌 are smooth
manifolds and 𝑓 is a smooth map.

(a) Show 𝐹 is a diffeomorphism, and conclude that Γ(𝑓 ) is a smooth manifold.

(b) We also write 𝐹 for the composite map 𝐹 ∶ 𝑋 → 𝑋 × 𝑌 , 𝑥 → (𝑥, 𝑓 (𝑥)). Show
that 𝑑𝐹𝑥 (𝑣) = (𝑣, 𝑑𝑓𝑥 (𝑣)). (You can use 𝑇(𝑥,𝑦) (𝑋 × 𝑌 ) = 𝑇𝑥 (𝑋) × 𝑇𝑦 (𝑌 ).)

(c) Show that the tangent space to Γ(𝑓 ) at the point (𝑥, 𝑓 (𝑥)) is the graph of
𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 ).

Exercise 2.13 A curve in a manifold 𝑋 is a smooth map 𝑡 → 𝑐(𝑡) of an open interval

of ℝ into 𝑋. The velocity vector of the curve 𝑐 at time 𝑡0 in 𝑥0 = 𝑐(𝑡0 ), denoted simply
𝑑𝑐
(𝑡 ), is defined to be the vector 𝑑𝑐𝑡0 (1) ∈ 𝑇𝑥0 (𝑋), where 𝑑𝑐𝑡0 ∶ ℝ1 → 𝑇𝑥0 (𝑋).
𝑑𝑡 0

(a) For 𝑋 = ℝ𝑘 and 𝑐(𝑡) = (𝑐1 (𝑡), … , 𝑐𝑘 (𝑡)), show that

𝑑𝑐
(𝑡 ) = 𝑑𝑐𝑡0 (1) = (𝑐1′ (𝑡0 ), … , 𝑐𝑘′ (𝑡0 )) ∈ 𝑇𝑥0 ℝ𝑘 .
𝑑𝑡 0
(b) For an arbitrary 𝑘-dimensional smooth manifold, use the above observation and
local parametrizations to prove that every vector in 𝑇𝑥0 (𝑋) is the velocity vector
of some curve in 𝑋.
Aside: This shows that there is a correspondence between tangent vectors at 𝑥0 ∈ 𝑋
and velocity vectors at 𝑡0 of curves 𝑐 ∶ 𝐼 → 𝑋 with 𝑐(𝑡0 ) = 𝑥0 . Note that two curves
𝑐1 ∶ 𝐼 → 𝑋 and 𝑐2 ∶ 𝐽 → 𝑋, with 𝐼 and 𝐽 open in ℝ, have the same velocity vector
in 𝑐1 (𝑡1 ) = 𝑥0 = 𝑐2 (𝑡2 ) if 𝑑(𝑐1 )𝑡1 (1) = 𝑑(𝑐2 )𝑡2 (1) ∈ 𝑇𝑥0 (𝑋). One can show that having
the same velocity vector in a point of 𝑋 is an equivalence relation on the set of curves
through 𝑥0 in 𝑋. Using this relation, we have shown that there is a unique correspondence
between tangent vectors at 𝑋 in 𝑥 and equivalence classes of smooth curves through 𝑥0
in 𝑋.
3. The Inverse Function Theorem, immersions and
embeddings

3.1 The Inverse Function Theorem and local diffeomorphisms

For understanding smooth manifolds, it can be smart to study maps between manifolds even
though it sounds like making things even more difficult. But assume we know something about
𝑋 and about a map 𝑓 ∶ 𝑋 → 𝑌 , then we might be able to say something interesting about 𝑌 .
In addition, there are a lot of interesting problems which can be stated in terms of properties of
maps.

So let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map between smooth manifolds. Remember that the deriva-
tive at 𝑥 ∈ 𝑋, 𝑑𝑓𝑥 ∶ 𝑇𝑥 𝑋 → 𝑇𝑓 (𝑥) 𝑌 , is a linear map between vector spaces. We have learned
that we may think of the derivative as the best linear approximation at a point. Since it is eas-
ier to understand linear maps, it would be nice if we could classify maps like 𝑓 by the behavior
of 𝑑𝑓𝑥 (with 𝑥 varying in 𝑋).

Question How much does 𝑑𝑓𝑥 tell us about the map 𝑓 ?

For the behaviour 𝑑𝑓𝑥 , there are three cases which we are going to study:

∙ dim 𝑋 = dim 𝑌 in which case the nicest possible behaviour of 𝑓 at 𝑥 is that 𝑑𝑓𝑥 an
isomorphism.

∙ dim 𝑋 < dim 𝑌 in which case the nicest possible behaviour of 𝑓 at 𝑥 is that 𝑑𝑓𝑥 one-to-
one.

∙ dim 𝑋 > dim 𝑌 in which case the nicest possible behaviour of 𝑓 at 𝑥 is that 𝑑𝑓𝑥 onto.

We are going to consider these cases separately.

∙ First case: 𝑑𝑓𝑥 is an isomorphism. We begin with the nicest case when 𝑑𝑓𝑥 is an
isomorphism. This implies in particular: dim 𝑋 = dim 𝑌 .

Manifolds are characterized by the way they look in a neighborhood around any point. So
let us think locally. In the nicest case, 𝑓 sends a neighborhood of a point 𝑥 diffeomorphically
to a neighborhood of 𝑦 = 𝑓 (𝑥). In this case, 𝑓 is called a local diffeomorphism at 𝑥. More
precisely, we define:

Definition 3.1 (Local diffeomorphism) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map between

smooth manifolds. Then 𝑓 is called a local diffeomorphism at 𝑥 if there is an open

58
Chapter 3. The Inverse Function Theorem, immersions and embeddings 59
subset 𝑈 ⊂ 𝑋 containing 𝑥 such that 𝑓 (𝑈 ) ⊂ 𝑌 is open in 𝑌 and

𝑓|𝑈 ∶ 𝑈 → 𝑓 (𝑈 )

is a diffeomorphism. We say that 𝑓 is a local diffeomorphism if it is a local diffeomor-

phism at every 𝑥 ∈ 𝑋.

If 𝑓 is a diffeomorphism 𝑈 → 𝑉 between neighborhoods 𝑈 around 𝑥 ∈ 𝑋 and 𝑦 = 𝑓 (𝑥) ∈

𝑌 , respectively, let 𝑓 −1 be its smooth inverse. Then we have 𝑓 −1 ◦𝑓 = Id𝑈 and 𝑓 ◦𝑓 −1 = 𝐼𝑑𝑉 .
The chain rule implies

𝑑(𝐼𝑑𝑈 )𝑥 = 𝑑(𝑓 −1 )𝑦 ◦𝑑𝑓𝑥 , and 𝑑(𝐼𝑑𝑉 )𝑦 = 𝑑𝑓𝑥 ◦𝑑(𝑓 −1 )𝑦 .

But we obviously have 𝑑(Id𝑋 ) = Id𝑇𝑥 (𝑋) for any manifold 𝑋 and any point 𝑥 ∈ 𝑋. Hence 𝑑𝑓𝑥
is an isomorphism with inverse 𝑑(𝑓 −1 )𝑓 (𝑥) .

Thus a necessary condition for 𝑓 to be a local diffeomorphism at 𝑥 is that its derivative

𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑦 (𝑌 ) is an isomorphism. It is an important result that this is actually a suffi-
cient condition. In order to prove this, we recall the corresponding result for Euclidean space
from Calculus:

Theorem 3.2 (The Inverse Function Theorem in Calculus) Suppose that 𝑓 ∶ ℝ𝑛 →

ℝ𝑛 is continuously differentiable in an open set containing a point 𝑎 ∈ ℝ𝑛 , and
det 𝑑𝑓𝑎 ≠ 0, i.e., 𝑑𝑓𝑎 is an invertible linear map ℝ𝑛 → ℝ𝑛 . Then there is an open set
𝑉 ⊆ ℝ𝑛 containing 𝑎 and an open set 𝑊 ⊆ ℝ𝑛 containing 𝑓 (𝑎) such that 𝑓 ∶ 𝑉 → 𝑊
has a continuous inverse 𝑓 −1 ∶ 𝑊 → 𝑉 which is differentiable and for all 𝑦 ∈ 𝑊
satisfies
( )−1
𝑑(𝑓 −1 )𝑦 = 𝑑𝑓𝑓 −1 (𝑦) .

Remark 3.3 (It’s a map not a fraction) Note that this is exactly the formula you are
used to from Calculus 1 where we learned

(𝑓 −1 )′ (𝑦) = (𝑓 ′ (𝑓 −1 (𝑦)))−1 .
1
You may be used to this formula as (𝑓 −1 )′ (𝑦) = 𝑓 ′ (𝑓 −1 (𝑦))
from Calculus. But the fraction
here is misleading, since (𝑓 −1 )′ (𝑦) is a linear map. The superscript “to the −1" really
means take the inverse map! In dimension 1, the inverse map happens to be given by
multiplication by the inverse number. But for linear maps or matrices in dimensions > 1,
we cannot write the inverse as a fraction.

Theorem 3.4 (Inverse Function Theorem) Let 𝑋 and 𝑌 be smooth manifolds. Sup-
pose that 𝑓 ∶ 𝑋 → 𝑌 is a smooth map whose derivative

𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 )

at a point 𝑥 ∈ 𝑋 is an isomorphism. Then 𝑓 is a local diffeomorphism at 𝑥.

60 3.1. The Inverse Function Theorem and local diffeomorphisms
The great thing about the Inverse Function Theorem (IFT) is that it tells us that in order to
check that 𝑓 is a diffeomorphism in a neighborhood of a point 𝑥, we just need to check that a
single number, the determinant of 𝑑𝑓𝑥 , is nonzero.

Idea of Proof: We can assume that 𝑋 and 𝑌 are subsets in ℝ𝑁 for some large 𝑁. Let
𝜙 ∶ 𝑈 → 𝑋 be a local parametrization around 𝑥 ∈ 𝑋, and 𝜓 ∶ 𝑊 → 𝑌 a local parametrization
around 𝑦 = 𝑓 (𝑥) ∈ 𝑌 with 𝑈 ⊂ ℝ𝑛 and 𝑊 ⊂ ℝ𝑛 open and 𝜙(0) = 𝑥 and 𝜓(0) = 𝑦.1

We define the map 𝜃 ∶ 𝑈 → 𝑊 as in the following diagram:

𝑓
𝑋O /𝑌
O
𝜙 𝜓 𝜓 −1

𝑈 /𝑊.
𝜃∶=𝜓 −1 ◦𝑓 ◦𝜙

Then recall that 𝑑𝑓𝑥 is defined such that the following diagram commutes
𝑑𝑓𝑥
𝑇𝑥 (𝑋) / 𝑇 (𝑌 )
O 𝑦
O
𝑑𝜙0 𝑑(𝜙−1 )𝑥 𝑑𝜓0

ℝ𝑘 / ℝ𝑙 .
𝑑𝜃0

Our assumption is that 𝑑𝑓𝑥 is an isomorphism which implies that 𝑑𝜃0 is an isomorphism.
By the IFT in Calculus, this implies that

∙ there is an open neighborhood 𝑉 ⊆ 𝑈 around 0 and

∙ there is an open neighborhood 𝑉 ′ ⊆ 𝑊 around 0 such that

∙ 𝜃|𝑉 ∶ 𝑉 → 𝑉 ′ is a diffeomorphism.

Since 𝜙 and 𝜓 are diffeomorphisms, 𝜙(𝑉 ) ⊆ 𝑋 and 𝜓(𝑉 ′ ) ⊆ 𝑌 are open neighborhoods
of 𝑥 and 𝑦, respectively. Moreover, 𝜙|𝑉 and 𝜓|𝑉 ′ are local parametrizations around 𝑥 and 𝑦,
respectively, and
𝑓|𝜙(𝑉 ) ∶ 𝜙(𝑉 ) → 𝜓(𝑉 ′ )
is a diffeomorphism.

Note that this is a local statement, i.e., if 𝑑𝑓𝑥 is invertible, it only tells us that 𝑓 is invertible
in a neighborhood of 𝑥. Even if 𝑑𝑓𝑥 is invertible for every 𝑥 ∈ 𝑋, one cannot conclude that
𝑓 ∶ 𝑋 → 𝑌 is globally a diffeomorphism. But such an 𝑓 is a local diffeomorphism for every
point 𝑥 ∈ 𝑋. We call such a map a local diffeomorphism (without having to refer to a point).

Example 3.5 (A global diffeomorphism) The map

(−𝜋∕2, 𝜋∕2) → ℝ, 𝑡 → tan 𝑡

is a global diffeomorphism.
1
Note that the dimension has to be the same when the tangent spaces are isomorphic.
Chapter 3. The Inverse Function Theorem, immersions and embeddings 61

Example 3.6 (A local but not global diffeomorphism) The map

𝑓 ∶ ℝ1 → 𝕊1 ⊂ ℝ2 , 𝑡 → (cos 𝑡, sin 𝑡)
is a local diffeomorphism but not a global diffeomorphism. Let us check how this ex-
ample works:
First, 𝑓 is not a global diffeomorphism because it is not injective. And we have seen
that 𝑓 is not a homeomorphism even when we restrict it to [0, 2𝜋) → 𝕊1 . We could also
argue that 𝕊1 is compact and ℝ is not, so there is no chance of finding a diffeomorphism
between them.
However, the Inverse Function Theorem 3.4 tells us that 𝑓 is indeed a local dif-
feomorphism since 𝑑𝑓𝑡 is an isomorphism for every 𝑡 ∈ ℝ. For, let 𝑡0 ∈ ℝ such that
cos(𝑡0 ) > 0 (for other points the argument is similar, we just want to be able to choose a
parametrization), and consider the local parametrization
√
𝜓 ∶ 𝑊 = (−1, 1) → 𝑉 ⊂ ℝ2 , 𝑦 → ( 1 − 𝑦2 , 𝑦)
of 𝕊1 around 𝑓 (𝑡0 ) with 𝑉 = {(𝑥, 𝑦) ∈ 𝕊1 ∶ 𝑥 < 0}. The inverse is given by projecting
onto the second coordinate: 𝜓 −1 (𝑥, 𝑦) = 𝑦.
Let 𝜖 > 0 be such that both cos(𝑡0 − 𝜖) > 0 and cos(𝑡0 + 𝜖) > 0. We let 𝜙 ∶ 𝑈 =
(𝑡0 − 𝜖, 𝑡0 + 𝜖) → ℝ be the local parametrization around 𝑡0 to be the inclusion (we don’t
shift 𝑈 to be centred around 0). Then the map 𝜃 ∶ 𝑈 → 𝑊 (see proof of the IFT) is
defined as
𝜃 = 𝜓 −1 ◦𝑓 ◦𝜙, 𝑡 → sin 𝑡.
Then we get
𝑑𝜃𝑡 ∶ ℝ → ℝ, 𝑧 → (cos 𝑡) ⋅ 𝑧.
at any point 𝑡 in 𝑈 . Since 𝜙 is the identity at each point it is defined, we know 𝑑𝜙𝑡 = id.
To calculate the derivative of 𝜓, we consider it first as a map 𝑊 → ℝ2 . Then we get
( )
𝑦
𝑑𝜓𝑦 ∶ ℝ → ℝ2 , 𝑧 → − √ , 1 ⋅ 𝑧.
1 − 𝑦2
Remember that the tangent space 𝑇𝑓 (𝑡0 ) 𝕊1 is by definition the image of 𝑑𝜓𝑦0 ∶ ℝ → ℝ2
where 𝑦0 is the point in 𝑊 which maps to 𝑓 (𝑡0 ). Thus we have
𝑇𝑓 (𝑡0 ) 𝕊1 = 𝑑𝜓𝑡0 (ℝ) = span{(− sin(𝑡0 ), cos(𝑡0 ))} ⊂ ℝ2 .
Now we collect all this information
𝑑𝑓𝑡0 = 𝑑𝜓sin 𝑡0 ◦𝑑𝜃𝑡0
and hence
( )
sin 𝑡0
𝑑𝑓𝑡0 (𝑧) = − , 1 (cos 𝑡0 ) ⋅ 𝑧
cos 𝑡0
=(− sin 𝑡0 , cos 𝑡0 ) ⋅ 𝑧.
Summarizing we have
𝑑𝑓𝑡0 ∶ 𝑇𝑡0 ℝ → 𝑇𝑓 (𝑡0 ) 𝕊1
𝑧 → (− sin(𝑡0 ), cos(𝑡0 )) ⋅ 𝑧
which is an isomorphism. For any other point in ℝ, there is a similar argument.
62 3.1. The Inverse Function Theorem and local diffeomorphisms

Figure 3.1: Locally, the map 𝑓 is a diffeomorphism at any point. But it cannot be a global
diffeomorphism, since it is not injective.

Lemma 3.7 A bijective local diffeomorphism is a global diffeomorphism.

Proof: See Exercise 3.2.

Remark 3.8 (𝑑𝑓𝑥 looks like the identity) In some situations it would be nice if we
could assume that the linear isomorphism 𝑑𝑓𝑥 was the identity. This is usually not the
case of course. But our freedom of choosing local parametrizations allows us to do the
following. Assume that 𝑑𝑓𝑥 is an isomorphism as in the IFT. Then we can choose local
parametrizationsa 𝜙 ∶ 𝑈 → 𝑋 and 𝜓 ∶ 𝑈 → 𝑌 around 𝑥 and 𝑓 (𝑥), respectively, with
the same open domain 𝑈 ⊂ ℝ𝑛 , such that the diagram commutes:

𝑓
𝑋O /𝑌
O
𝜙 𝜓

𝑈 / 𝑈.
Id𝑈

For example, in Example 3.6 above, we would replace

∙ 𝑊 = (−1, 1) with 𝑈 = (𝑡0 − 𝜀, 𝑡0 + 𝜀) and

∙ 𝜓 with
√
𝜓 ∶ 𝑡 → (− 1 − sin2 𝑡, sin 𝑡) = (cos 𝑡, sin 𝑡).
a
We are going to explain how to choose suitable parametrizations in the next sections.
Chapter 3. The Inverse Function Theorem, immersions and embeddings 63
3.2 Immersions and embeddings

3.2.1 Immersions

We continue our study of smooth maps between manifolds using the behaviour of their deriva-
tive. Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. We would like to understand how much do we know
about 𝑓 if the derivative is injective. Note that this is only possible if dim 𝑋 ≤ dim 𝑌 , so this
is a silent assumption in this chapter.

Let us introduce some terminology for this case.

Definition 3.9 (Immersion) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. If 𝑑𝑓𝑥 is injective, we

say that 𝑓 is an immersion at 𝑥. If 𝑓 is an immersion at every point, we say that 𝑓 is
an immersion.

Let us look at some first examples:

∙ Every linear injective map 𝐿 ∶ ℝ𝑘 → ℝ𝑛 is an immersion. This follows from the fact
that 𝐿 equals its own derivative 𝑑𝐿𝑥 at every point 𝑥 ∈ ℝ𝑘 .

∙ Every local diffeomorphism 𝑓 ∶ 𝑋 → 𝑌 is also an immersion. Considering our exam-

ples for local diffeomorphisms, this also shows that an immersion 𝑓 does not have to be
injective itself, only its derivative 𝑑𝑓𝑥 is injective at every 𝑥.

∙ Let 𝑓 be the map defined by

( )
2 𝑒𝑡 + 𝑒−𝑡 𝑒𝑡 − 𝑒−𝑡
𝑓 ∶ ℝ → ℝ , 𝑡 → , .
2 2

We check in the exercises that this map is an immersion.

∙ Let 𝑓 ∶ ℝ2 → ℝ3 be the map

𝑓 ∶ (𝑠, 𝑡) → ((1 + 2 cos 𝑠) cos 𝑡, (1 + 2 cos 𝑠) sin 𝑡, 2 sin 𝑠)

is smooth, since all its components are smooth functions. The derivative of 𝑓 at a point
(𝑠, 𝑡) can be described in the standard bases of ℝ2 and ℝ3 , respectively, by the matrix

⎛−2 sin 𝑠 cos 𝑡 −(2 cos 𝑠 + 1) sin 𝑡⎞

⎜ −2 sin 𝑠 sin 𝑡 (2 cos 𝑠 + 1) cos 𝑡 ⎟ .
⎜ ⎟
⎝ 2 cos 𝑠 0 ⎠
We claim that this matrix has rank 2 for all (𝑠, 𝑡). This follows from the following two
observations: From the third row
{ we get that
} the two columns can only be linearly depen-
𝜋
dent if cos 𝑠 = 0, i.e., if 𝑠 ∈ 𝑛 2 |𝑛 ∈ ℤ . However, the determinant of the submatrix
of the two top rows is

− 2 sin 𝑠(2 cos 𝑠 + 1) cos2 𝑡 − 2 sin 𝑠(2 cos 𝑠 + 1) sin2 𝑡

= − 2 sin 𝑠(2 cos 𝑠 + 1)
64 3.2. Immersions and embeddings
which is ≠ 0 for all 𝑠 ∉ {𝑛𝜋 ∶ 𝑛 ∈ ℤ}. Thus the two columns are always linearly
independent which proves the claim. Hence 𝑓 is an immersion. The image of 𝑓 is the
torus 𝑇 (1, 2) consisting of all points in ℝ3 at distance 𝑏 from the circle of radius 𝑎 in the
𝑥𝑦-plane. We learn from this example, that 𝑓 itself does not have to be one-to-one to be
an immersion. It is about the derivative. We also observe that the image of 𝑓 is a smooth
manifold in ℝ3 .

Before we study more examples and interesting phenomena, we pause for a moment and
show a technical and useful result about immersions.

Towards the local immersion theorem

We have all seen many injective maps before. In fact, among all the injective maps ℝ𝑘 →
ℝ𝑛 with 𝑘 ≤ 𝑛 there is a simplest one, namely the map

𝜄𝑘𝑛 ∶ ℝ𝑘 → ℝ𝑛 , (𝑥1 , … , 𝑥𝑘 ) → (𝑥1 , … , 𝑥𝑘 , 0, … , 0)

which sends the coordinates of a point 𝐱 ∈ ℝ𝑘 to the first 𝑘 coordinates in ℝ𝑛 and adds 0 at the
remaining 𝑛 − 𝑘 positions. This corresponds to the inclusion ℝ𝑘 × {𝟎} ⊂ ℝ𝑛 where 𝟎 denotes
the (𝑛 − 𝑘)-tuple of zeros. The map 𝜄𝑘𝑛 is called the canonical immersion from ℝ𝑘 into ℝ𝑛 .
The matrix which represents 𝜄𝑘𝑛 with respect to the standard bases of ℝ𝑘 and ℝ𝑛 , respectively,
is the 𝑛 × 𝑘-matrix 𝐽𝑛𝑘 which has the 𝑘 × 𝑘-identity matrix in the first 𝑘 rows and only zeros in
the remaining 𝑛 − 𝑘 rows at the bottom. Not all injective maps ℝ𝑘 → ℝ𝑛 are as simple as 𝜄𝑘𝑛 .
However, we will see very soon that it does not get that worse either, at least locally.

To motivate the next theorem let us start with a linear injective map 𝐿 ∶ ℝ𝑘 → ℝ𝑛 . Say
we do not like transformations between spaces of different dimensions. The only one we think
is OK is the canonical one 𝜄𝑘,𝑛 . And say we do like isomorphisms of ℝ𝑛 to itself. Then we could
write 𝐿 as the composition
𝜄𝑘𝑛 𝛼
ℝ𝑘 ←←←→
← ℝ𝑛 ←←→
← ℝ𝑛

where 𝛼 is a linear isomorphism determined by sending the first 𝑘 standard basis vectors 𝐞𝑛1 , … , 𝐞𝑛𝑘
of ℝ𝑛 to the images 𝐯1 = 𝐿(𝐞𝑘1 ), … , 𝐯𝑘 = 𝐿(𝐞𝑘𝑘 ) in ℝ𝑛 of the standard basis vectors of ℝ𝑘 and
by sending the remaining basis vectors 𝐞𝑛𝑘+1 , … , 𝐞𝑛𝑛 of ℝ𝑛 to basis vectors 𝐰𝑘+1 , … , 𝐰𝑛 of the
orthogonal complement of the image of 𝐿 in ℝ𝑛 . Since 𝐿 is injective, this complement has
dimension 𝑛 − 𝑘.

So what happened is that we placed a copy of ℝ𝑘 into ℝ𝑛 as the 𝑘-dimensional plane con-
taining the origin via 𝜄𝑘𝑛 and then we move this plane inside ℝ𝑛 to the position of 𝐿(ℝ𝑘 ).

Actually, we are used to such a manoeuvre. For what 𝛼 does is changing the basis of ℝ𝑛
from the standard basis to the new basis which consists of the vectors 𝐯1 , … , 𝐯𝑘 , 𝐰𝑘+1 , … , 𝐰𝑛 .
With respect to this new basis for ℝ𝑛 and the standard basis for ℝ𝑘 , the matrix which repre-
sents 𝐿 is the 𝑛 × 𝑘-matrix 𝕊𝑘𝑛 described above.

We can translate this idea to immersions between smooth manifolds. In other words — up
to diffeomorphisms — the canonical immersion is locally the only immersion:
Chapter 3. The Inverse Function Theorem, immersions and embeddings 65

Theorem 3.10 (Local Immersion Theorem) Let 𝑋 and 𝑌 be smooth manifolds of

dimensions 𝑘 and 𝑛, respectively, with 𝑘 ≤ 𝑛. Suppose that 𝑓 ∶ 𝑋 → 𝑌 is an immersion
at 𝑥, and write 𝑦 = 𝑓 (𝑥). Then there exist local coordinates around 𝑥 and 𝑦 such that

𝑓 (𝑥1 , … , 𝑥𝑘 ) = (𝑥1 , … , 𝑥𝑘 , 0, … , 0).

More precisely, we can choose local parametrizations 𝜙 ∶ 𝑈 → 𝑋 around 𝑥 and

𝜓 ∶ 𝑊 → 𝑌 around 𝑦 such that in the commutative diagram

𝑓
𝑋O /𝑌
O
𝜙 𝜓
canonical immersion /𝑊
𝑈
𝜃=𝜓 −1 ◦𝑓 ◦𝜙

the map 𝜃 ∶ 𝑈 → 𝑊 is the canonical immersion restricted to 𝑈 .

Proof of the Local Immersion Theorem 3.10: We start by choosing any local parametriza-
tion 𝜙 ∶ 𝑈 → 𝑋 with 𝜙(0) = 𝑥 and 𝜓 ∶ 𝑊 → 𝑌 with 𝜓(0) = 𝑦:

𝑓
𝑋O /𝑌
O
𝜙 𝜓

𝑈 /𝑊.
𝜃=𝜓 −1 ◦𝑓 ◦𝜙

The plan is to manipulate 𝜙 and 𝜓 such that 𝜃 becomes the canonical immersion.

By the assumption, 𝑑𝜃0 ∶ ℝ𝑘 → ℝ𝑛 is injective. Hence, after choosing a suitable basis for
ℝ𝑛 , we can assume that 𝑑𝜃0 is given by the 𝑛 × 𝑘-matrix 𝐽𝑛𝑘 which has the 𝑘 × 𝑘-identity matrix
in the first 𝑘 rows and only zeros in the remaining 𝑛 − 𝑘 rows. Now we define a new map

Θ ∶ 𝑈 × ℝ𝑛−𝑘 → ℝ𝑛 , by Θ(𝑥, 𝑧) = 𝜃(𝑥) + (0, 𝑧).

It is related to 𝜃 by the commutative diagram

canonical / 𝑈 × ℝ𝑛−𝑘
𝑈
immersion

𝜃 Θ
y
ℝ𝑛 .

Since 𝜃 is a local diffeomorphism at 0, we can choose 𝑈 and 𝑉 small enough such that 𝜃 sends
open sets to open sets. By the assumption on 𝑑𝜃0 and the construction of Θ, 𝑑Θ0 is represented
by the 𝑛 × 𝑛-identity matrix. By the Inverse Function Theorem, this implies that Θ is a local
diffeomorphism of ℝ𝑛 to itself at 0. Hence we can find an open subset 𝑉 ⊂ 𝑈 × ℝ𝑛−𝑘 and
𝑊 ′ ⊂ ℝ𝑛 such that Θ|𝑉 is a diffeomorphism. After possibly shrinking 𝑈 to an open subset 𝑈 ′
66 3.2. Immersions and embeddings
we get the commutative diagram
𝑓
𝑋O /𝑌
O
𝜙 𝜓

𝑈′
𝜃 /𝑊
O
canonical Θ
immersion ,𝑊′

Since 𝜓 and Θ are local diffeomorphisms at 0, so is the composition 𝜓◦Θ. Hence we can
use 𝜓◦Θ as a local parametrization around 𝑦.

Thus, we have constructed the desired commutative diagram

𝑓
𝑋O /𝑌
O
𝜙 𝜓◦𝚯
canonical /𝑊′
𝑈′
immersion

which finishes the proof.

Remark 3.11 We observe from the proof that to be an immersion is a local condition,
i.e., if 𝑓 ∶ 𝑋 → 𝑌 is an immersion at 𝑥, then it is also an immersion for all points
in a neighborhood of 𝑥. For, the local parametrization 𝜙 ∶ 𝑈 → 𝑋 of the proof also
parametrizes any point in the image of 𝜙. This is an open subset around 𝑥 because 𝜙 is
a diffeomorphism onto its image. Hence in order to say more about 𝑓 we need to add
some global topological properties to the local differential data. Recall that for a local
diffeomorphism to be a global one, it suffices to require to be one-to-one and onto. We
will now study what kind of additional property we wish to impose on immersions to
behave nicely.

3.2.2 The image of an immersion

Even though an immersion does not have to be one-to-one itself, the injectivity of the derivative
does not leave much room for different phenomena in the fibers. So let us have a closer look at
the image of an immersion 𝑓 ∶ 𝑋 → 𝑌 , i.e., the subset 𝑓 (𝑋) ⊂ 𝑌 .

Question We can ask at least these two questions:

∙ Is 𝑓 (𝑋) always a submanifold?a

∙ Is 𝑓 a diffeomorphism onto its image?

a
Recall that a subset 𝑍 ⊂ 𝑋 of a smooth manifold 𝑋 is called a submanifold if and only if 𝑍 is itself
a smooth manifold, possibly of lower dimension.

First observations:
Chapter 3. The Inverse Function Theorem, immersions and embeddings 67
∙ Recall the map 𝑓 ∶ ℝ2 → ℝ3 in the above examples for which the image is the two-
dimensional torus. This provides an example where 𝑓 (𝑋) is a submanifold. However, 𝑓
is not a diffeomorphism onto its image, since 𝑓 is not one-to-one.
∙ Note also that we can, in principle, answer the second question independently of the
first. For we can check if 𝑓 has an inverse which is defined on the subset 𝑓 (𝑋) ⊂ 𝑌 and
check if this map is smooth. However, we will see below that if 𝑓 is a diffeomorphism
onto 𝑓 (𝑋), then this will provide the subspace 𝑓 (𝑋) ⊂ with the structure of a smooth
manifold and hence this will be a submanifold in 𝑌 .

We have seen that 𝑓 (𝑋) often is a manifold. But there are also examples where this is not
the case. We will have a look at them now:

Example 3.12 (Figure eight as twist) We just look at Figure 3.2 without making the
formula for 𝑓 explicit. We see here that 𝑓 (𝑋) is not a manifold as a subset of ℝ2 , since
the intersection point of the two branches does not have a local parametrization. We also
see that 𝑓 is not a diffeomorphism onto its image, since 𝑓 is not one-to-one.

Figure 3.2: The map 𝑓 twists the circle once. It is an immersion, but it is not injective. The
image of 𝑓 , considered as a subspace of ℝ2 , is not a submanifold.

Example 3.13 (Figure eight immersion as wrap) So let us modify the map to make it
one-to-one. Consider the map
𝑓 ∶ (−𝜋, 𝜋) → ℝ2 , 𝑡 → (sin 2𝑡, sin 𝑡).
The image of 𝑓 is called a lemniscate, the locus of points (𝑥, 𝑦) satisfying 𝑥2 = 4𝑦2 (1 −
𝑦2 ). See Figure 3.3.
We can check that 𝑓 is smooth, one-to-one and an immersion. For the latter note
that 𝑑𝑓𝑡 can be represented by the 2 × 1-matrix
( )
2 cos 2𝑡
𝐽𝑓 (𝑡) =
cos 𝑡
which is never zero for 𝑡 ∈ (−𝜋, 𝜋) and hence, as a linear map between one-dimensional
vector spaces, 𝑑𝑓𝑡 is an isomorphism for all 𝑡.
Nevertheless, 𝑓 is still not a diffeomorphism onto its image. In fact, we see that 𝑓 is
not a homeomorphism onto its image 𝑓 (𝑋). Moreover, the image 𝑓 (𝑋) is the same as
above and not a manifold.
68 3.2. Immersions and embeddings

Figure 3.3: The map 𝑓 wraps ℝ along the figure eight. It is an immersion and one-to-one.
However, the image of 𝑓 , considered as a subspace of ℝ2 , is still not a submanifold. Note that,
even though 𝑓 is a local diffeomorphism and bijective onto its image, this is not a contradiction
to Lemma 3.7, since the image of 𝑓 , i.e., the figure eight, is not a manifold in ℝ2 .

3.2.3 A curve on the torus

Let us look at a classical example of another case of a map which is a one-to-one immersion,
but not a homeomorphism — and hence not a diffeomorphism — onto its image. We will not,
however, show that 𝑓 (𝑋) is not a submanifold.

Let 𝑔 ∶ ℝ → 𝕊1 ⊂ ℂ be the local diffeomorphism 𝑡 → 𝑒2𝜋𝑖𝑡 . We define

𝐺 ∶ ℝ2 → 𝕊1 × 𝕊1 =∶ 𝕋 2 ⊂ ℂ2 , 𝐺(𝑠, 𝑡) = (𝑔(𝑠), 𝑔(𝑡))

The map 𝐺 is a local diffeomorphism from the plane onto the two-dimensional torus 𝕋 2 .

We define the map 𝛾 by

𝛾 ∶ ℝ → 𝕋 2 , 𝛾(𝑡) = (𝑔(𝑡), 𝑔(𝛼𝑡)).

where 𝛼 is an irrational number.

Figure 3.4: The map 𝛾 wraps a line around the torus. If the slope is irrational, then the image
will never meet itself and is dense on 𝕋 2 .
Chapter 3. The Inverse Function Theorem, immersions and embeddings 69

Example 3.14 (Image of a line with irrational slope) The map 𝛾 is an immersion
because 𝑑𝛾𝑡 is nonzero for every 𝑡, and, as above, a nonzero linear map from a one-
dimensional vector space to another is automatically injective. Moreover, 𝛾 itself is in-
jective, since 𝛾(𝑡1 ) = 𝛾(𝑡2 ) implies

𝑔(𝑡1 ) = 𝑔(𝑡2 ) and 𝑔(𝛼𝑡1 ) = 𝑔(𝛼𝑡2 )

⇒ 𝑒2𝜋𝑖𝑡1 = 𝑒2𝜋𝑖𝑡2 and 𝑒2𝜋𝑖𝛼𝑡1 = 𝑒2𝜋𝑖𝛼𝑡2
⇒ 𝑡1 − 𝑡2 ∈ ℤ and 𝛼(𝑡1 − 𝑡2 ) ∈ ℤ.

Since 𝛼 is irrational, this implies 𝑡1 = 𝑡2 . One can show that the image of 𝛾 is a dense
subset in 𝕋 2 .
However, 𝛾 is not a diffeomorphism onto its image, since it is not even a homeo-
morphism:
For, look at the set 𝛾(ℤ) = {𝛾(𝑛) ∶ 𝑛 ∈ ℤ}. By Dirichlet’s Approximation Theorem,
for every 𝜀 > 0, there are integers 𝑛 and 𝑚 such that

|𝛼𝑛 − 𝑚| < 𝜖.

Since the line segment between two points 𝑒2𝜋𝑖𝑡1 and 𝑒2𝜋𝑖𝑡2 on the unit circle is shorter
than the circular arc of length |𝑡1 − 𝑡2 |, we have

|𝑒2𝜋𝑖𝛼𝑛 − 𝑒2𝜋𝑖𝑚 | ≤ 2𝜋|𝛼𝑛 − 𝑚| < 2𝜋𝜀.

Therefore, with coordinates in ℂ2 , we get

|𝛾(𝑛) − 𝛾(0)|
( )
= |(𝑒2𝜋𝑖𝑛 , 𝑒2𝜋𝑖𝛼𝑛 ) − (𝑒0 , 𝑒0 )| = | 1, 𝑒2𝜋𝑖𝛼𝑛 − (1, 1) |
= |𝑒2𝜋𝑖𝛼𝑛 − 𝑒2𝜋𝑖𝑚 | ≤ 2𝜋|𝛼𝑛 − 𝑚| < 2𝜋𝜀.

Thus, there is a sequence of integers such that 𝛾(𝑛) converges to 𝛾(0), i.e., 𝛾(0) is a limit
point in 𝛾(ℤ). The image of a convergent sequence under a continuous map is again
a convergent sequence.a Hence if 𝛾 −1 was continuous, then 0 = 𝛾 −1 (𝛾(0)) had to be a
limit point as well. However, ℤ does not have any limit points in ℝ. Hence 𝛾 is not a
homeomorphism onto its image.
We can also show that 𝛾(ℝ) with the subspace topology in 𝕋 2 is not a manifold. We
leave this as an exercise for the moment.
a
This is actually an alternative way to define what continuity means.

3.3 Embeddings

3.3.1 Embeddings

We have seen that both our questions we asked earlier may have negative answers. Let us now
focus on the situation when they do have a positive answer. We give this case a name:
70 3.3. Embeddings

Definition 3.15 (Embedding) An immersion 𝑓 ∶ 𝑋 → 𝑌 is a called an embedding

if 𝑓 (𝑋) ⊂ 𝑌 considered with the subspace topology is a manifold and 𝑓 is a diffeo-
morphsim onto its image.

Note that an embedding must be one-to-one, since it is a diffeomorphism onto its image.

Now let 𝑓 ∶ 𝑋 → 𝑌 be an immersion which is also one-to-one. Let us try to show what 𝑓
is an embedding to see which additional assumption we have to make:

∙ By the Local Immersion Theorem 3.10, we can choose local parametrizations 𝜙 ∶ 𝑈 →

𝑊 ⊂ 𝑋 around 𝑥 and 𝜓 ∶ 𝑉 → 𝑊 ′ ⊂ 𝑌 around 𝑦 = 𝑓 (𝑥) such that the induced map
𝜃 ∶ 𝑈 → 𝑉 is the canonical immersion. This provides the commutative diagram

𝑓
𝑋O / 𝑓 (𝑋)
O
𝐨𝐩𝐞𝐧 ✓ 𝐨𝐩𝐞𝐧 ?
? 𝑓|𝑊 ≅ ?
𝑊O / 𝑓 (𝑊 )
4 O
𝑓|𝑊 ◦𝜙 𝜓|𝑉 ∩𝜓 −1 (𝑓 (𝑊 ))
𝜙 ≅
≅
canonical / 𝑉 ∩ 𝜓 −1 (𝑓 (𝑊 )).
𝑈
immersion

Since the bottom horizontal map is the canonical immersion, the restriction 𝑓|𝑊 is a
diffeomorphism. As 𝜙 is one as well, we see that the composite 𝑓|𝑊 ◦𝜙 ∶ 𝑈 → 𝑓 (𝑊 ) is
a diffeomorphism. Thus 𝑓 (𝑊 ) is diffeomorphic to an open subset in ℝ𝑛 .

∙ Since we can do this for every point 𝑦 ∈ 𝑓 (𝑋), we would like to say that the collection
of diffeomorphisms 𝑓|𝑊 ◦𝜙 for varying 𝜙 provide local parametrizations for 𝑓 (𝑋).

∙ However, we do not know that 𝑓 (𝑊 ) is open in 𝑓 (𝑋). This does not follow from the
given assumptions on 𝑓 .

∙ Let us assume for a moment that we were lucky and 𝑓 (𝑊 ) was open for each open subset
𝑊 ⊂ 𝑋. Then we could conclude that 𝑓 is a diffeomorphism onto its image. For then 𝑓
would be a bijective and a local diffeomorphism, since for every point 𝑥 ∈ 𝑋 there would
be an open subset 𝑥 ∈ 𝑊 ⊂ 𝑋 and an open subset 𝑓 (𝑊 ) ⊂ 𝑓 (𝑋) such that 𝑓|𝑊 is a
diffeomorphism 𝑊 → 𝑓 (𝑊 ). A bijective local diffeomorphism is a diffeomorphism
by Lemma 3.7 and Exercise 3.2.

∙ Thus we can conclude: If 𝑓 (𝑊 ) is open for each open subset 𝑊 ⊂ 𝑋, then 𝑓 is a

diffeomorphism onto its image.

3.3.2 The embedding theorem

We learn from this discussion that we need to find conditions which ensure that 𝑓 (𝑊 ) is open
in 𝑓 (𝑋). This is the case if 𝑓 is an open map, i.e., if 𝑓 sends every open subset in 𝑋 to an
Chapter 3. The Inverse Function Theorem, immersions and embeddings 71

Figure 3.5: One might think that 𝑓 (𝑊 ) provides a local parametrization, but it is not neces-
sarily open in 𝑓 (𝑋). Hence it is notguaranteed that 𝑓 (𝑋), considered as a subspace of 𝑌 , is a
manifold.

open subset in 𝑌 . Equivalently, we could require that 𝑓 is a closed map, i.e., 𝑓 sends every
closed subset in 𝑋 to a closed subset in 𝑌 . Note that this fits well into what we have seen in
the last two examples. For, there 𝑓 was neither closed nor open and 𝑓 ∶ 𝑋 → 𝑓 (𝑋) failed to
be a homeomorphism. A condition that is often easier to test is the following: Recall that for a
general continuous map, the image of any compact set is compact. However, the preimage of
a compact subset is, in general, not compact.

Definition 3.16 (Proper maps) A map 𝑓 ∶ 𝑋 → 𝑌 between topological spaces is said

to be proper if the preimage of every compact subset is a compact subset.

Being proper turns out to be a sufficient global topological constraint for our purposes:

Theorem 3.17 (Embedding Theorem) Let 𝑓 ∶ 𝑋 → 𝑌 be a one-to-one immersion

which satisfies one of the following conditions:

∙ 𝑓 is an open map, or

∙ 𝑓 is a closed map, or

∙ 𝑓 is a proper map.

Then 𝑓 is an embedding.

Before we prove the theorem, let us assume for a moment that 𝑋 is compact. Then every
continuous map 𝑓 ∶ 𝑋 → 𝑌 is proper. This follows from the fact that closed subsets of compact
sets are compact. Hence we can deduce from the theorem the following important special case:

Corollary 3.18 (Compact domain) For a compact smooth manifold 𝑋, every one-to-
one immersion 𝑓 ∶ 𝑋 → 𝑌 is an embedding.
72 3.3. Embeddings
Proof of the Embedding Theorem 3.17:

We know from our previous arguments that 𝑓 ∶ 𝑋 → 𝑓 (𝑋) is a diffeomorphism if 𝑓 is

open. Since a bijective continuous map is open if and only if it is closed, 𝑓 ∶ 𝑋 → 𝑓 (𝑋) is
a diffeomorphism if 𝑓 is closed. Now assume that 𝑓 is proper. We are going to prove that
this implies that 𝑓 is a closed map and thereby prove the theorem. Actually, we show the
following general statement, formulated in Lemma 3.19. Proving the lemma will finish the
proof of Theorem 3.17.

Lemma 3.19 A continuous bijective proper map 𝑓 ∶ 𝑋 → 𝑌 is a homeomorphism.

Proof of Lemma 3.19:2 We have 𝑌 = 𝑓 (𝑋). Let 𝑍 be a closed subset in 𝑋. Let 𝑦 ∈

𝑌 ⧵ 𝑓 (𝑍). We are going to show that this is an open subset of 𝑓 (𝑋) which then implies that
𝑓 (𝑍) is closed in 𝑌 . Since 𝑌 is a subspace in Euclidean space, 𝑌 is locally compact which
means that we can find a compact neighborhood 𝐾 of 𝑦 in 𝑌 , i.e., there is an open subset
𝑈 and a compact subset 𝐾 such that 𝑦 ∈ 𝑈 ⊂ 𝐾. Then 𝑓 −1 (𝐾) is compact, since 𝑓 is
proper. Hence 𝑓 −1 (𝐾) ∩ 𝑍 is compact, since it is a closed subset in 𝑓 −1 (𝐾) ⊂ ℝ𝑁 . Hence
𝑓 (𝑓 −1 (𝐾) ∩ 𝑍) = 𝐾 ∩ 𝑓 (𝑍) is compact, since the continuous image of a compact set is always
compact. Since 𝐾 is a subspace in Euclidean space and thereby Hausdorff, this implies that
𝐾 ∩ 𝑓 (𝑍) is closed in 𝐾 and hence also closed in 𝑌 . Thus, since 𝑈 is open in 𝑌 , the subset
𝑈 ⧵ (𝐾 ∩ 𝑓 (𝑍)) is open in 𝑌 as well. Now, by choice of 𝑦, we know 𝑦 ∈ 𝑈 ⧵ (𝐾 ∩ 𝑓 (𝑍)). This
shows that 𝑦 has an open neighborhood in 𝑌 ⧵ 𝑓 (𝑍), and the latter is an open subset of 𝑌 .

2
We will use a little bit more general topology in this proof than we recalled so far. We hope that is ok.
Chapter 3. The Inverse Function Theorem, immersions and embeddings 73
3.4 Exercises and more examples

3.4.1 Diffeomorphisms, immersions and embeddings

Exercise 3.1 Let 𝐴 ∶ ℝ𝑛 → ℝ𝑛 be a linear map, and 𝑏 ∈ ℝ𝑛 . Show that the mapping

𝑓 ∶ ℝ𝑛 → ℝ𝑛 , 𝑥 → 𝐴𝑥 + 𝑏

is a diffeomorphism of ℝ𝑛 if and only if 𝐴 is invertible.

Exercise 3.2 Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. Assume that 𝑓 is a local diffeomor-

phism, and assume that 𝑓 is bijective as a map of sets. Show that 𝑓 is a diffeomorphism.

Exercise 3.3 Show that the map

( )
2 𝑒𝑡 + 𝑒−𝑡 𝑒𝑡 − 𝑒−𝑡
𝑓 ∶ ℝ → ℝ , 𝑡 → ,
2 2

is an embedding.

Exercise 3.4 Show that the map 𝑓 ∶ ℝ2 → ℝ3 defined by

(𝑠, 𝑡) → ((2 + cos(2𝜋𝑠)) cos(2𝜋𝑡), (2 + cos(2𝜋𝑠)) sin(2𝜋𝑡), sin(2𝜋𝑠))

is an immersion. Is it an embedding?

Exercise 3.5 Let 𝛾𝑎,𝑏 be the curve on the torus defined by

𝛾𝑎,𝑏 ∶ ℝ → 𝕊1 × 𝕊1 , 𝑡 → (𝑒2𝜋𝑖𝑎𝑡 , 𝑒2𝜋𝑖𝑏𝑡 )

where we consider 𝕊1 as a subset of ℂ ≅ ℝ2 . Let 𝑎 and 𝑏 are integers with 𝑎 ≠ 0. Show

that 𝛾𝑎,𝑏 factors through an embedding 𝕊1 → 𝕊1 ×𝕊1 , i.e., find a map 𝑔𝑎,𝑏 ∶ 𝕊1 → 𝕊1 ×𝕊1
which is an embedding such that 𝛾𝑎,𝑏 is the composite of a map ℝ → 𝕊1 , 𝑡 → 𝑒2𝜋𝑖𝑡 ,
followed by 𝑔𝑎,𝑏 .

Exercise 3.6 Consider the map 𝑓 ∶ (0, 3𝜋∕4) → ℝ2 , 𝑡 → sin(2𝑡)(cos 𝑡, sin 𝑡).

(a) Show that 𝑓 is an immersion.

(b) Let Im (𝑓 ) = 𝑓 ((0, 3𝜋∕4)) ⊂ ℝ2 be the image of 𝑓 considered as a subspace in

ℝ2 . Show that 𝑓 ∶ (0, 3𝜋∕4) → Im (𝑓 ) is not a homeomorphism. (Draw a picture
of the image of 𝑓 .)

(c) To test your understanding answer the following questions (and give reasons for
your answer):
74 3.4. Exercises and more examples
∙ What is the difference between Im (𝑓 ) and the graph Γ(𝑓 )?
∙ Is the map 𝐹 ∶ (0, 3𝜋∕4) → (0, 3𝜋∕4) × ℝ2 , 𝑡 → (𝑡, 𝑓 (𝑡)), an embedding?
∙ Would 𝑓 be an embedding if it was defined on the closed interval [0, 3𝜋∕4]?
∙ Is the map 𝑔 ∶ (0, 3𝜋∕4) → ℝ3 , 𝑡 → sin(2𝑡)(cos 𝑡, sin 𝑡, 𝑡) an embedding?
∙ Is the map ℎ ∶ [0, 3𝜋∕4] → ℝ3 , 𝑡 → (sin(2𝑡) cos 𝑡, sin(2𝑡) sin 𝑡, 2𝑡) an embed-
ding?

Exercise 3.7 Let 𝑋 be an 𝑛-dimensional smooth manifold, 𝑍 be a 𝑘-dimensional

smooth submanifold of 𝑋, and let 𝑧 ∈ 𝑍. Show that there exists a local coordinate
system (𝑥1 , … , 𝑥𝑛 ) defined in a neighborhood 𝑈 of 𝑧 in 𝑋 such that 𝑍 ∩ 𝑈 is defined by
the equations 𝑥𝑘+1 = 0, … , 𝑥𝑛 = 0, i.e., 𝑍 ∩ 𝑈 is the subset of points in 𝑈 for which the
functions 𝑥𝑘+1 , … , 𝑥𝑛 all vanish.
4. Submersions and regular values

4.1 Submersions

4.1.1 Submersions

Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. We will now turn to the question how much do we know
about 𝑓 if the derivative is surjective. Note that this will require that dim 𝑋 ≤ dim 𝑌 , and
this is a silent assumption in this chapter. We will see that this case will lead to a very useful
observation about the fibers of smooth maps.

We begin with some terminology:

Definition 4.1 (Submersion) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. If 𝑑𝑓𝑥 is surjective,

we say that 𝑓 is a submersion at 𝑥. If 𝑓 is a submersion at every point, we say that 𝑓
is a submersion.

Let us look at some first examples:

∙ Every linear surjective map 𝐿 ∶ ℝ𝑘 → ℝ𝑛 is a submersion. This follows from the fact
that 𝐿 equals its own derivative 𝑑𝐿𝑥 at every point 𝑥 ∈ ℝ𝑘 .

∙ Every local diffeomorphism 𝑓 ∶ 𝑋 → 𝑌 is also a submersion. Considering our exam-

ples for local diffeomorphisms, this also shows that a submersion 𝑓 does not have to be
surjective itself, only its derivative 𝑑𝑓𝑥 is surjective at every 𝑥.

∙ Let 𝑓 be the map defined by

𝑔 ∶ ℝ2 → ℝ, (𝑥, 𝑦) → 𝑥2 − 𝑦2 .

We check in Exercise 4.2 that 𝑔 is not a submersion. However, it is a submersion at every

point (𝑥, 𝑦) ≠ (0, 0).

∙ The Hopf map 𝑓 ∶ 𝕊3 → 𝕊2 is a submersion. We show this in the exercises. So far,

our previous calculations show that the derivative of 𝑓 is surjective at the north pole
𝐧3 = (0, 0, 0, 1) ∈ 𝕊3 and at 𝐪3 = (1, 0, 0, 0) ∈ 𝕊3 . For we showed that 𝑑𝑓𝐧3 can be
( )
0 2 0
represented by the matrix , while 𝑑𝑓𝐪3 can be represented by the matrix
−2 0 0
( )
0 2 0
. Both these matrices have rank two and the corresponding linear maps are
0 −2
surjective. Hence 𝑓 is a submersion at 𝐧3 and 𝐪3 . In the exercises we develop a more
elegant way and generalize these computations and show that 𝑓 is a submersion.

75
76 4.1. Submersions
∙ We define the map

𝑓 ∶ ℝ4 → ℝ,
(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) → 𝑥1 + 𝑥22 + 𝑥33 + 𝑥44 .

The derivative 𝑑𝑓𝑧 at a point 𝑧 = (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) is a linear map ℝ4 → ℝ given in the

standard basis by the 1 × 4-matrix
( )
𝑑𝑓𝑧 = 1 2𝑥2 3𝑥23 4𝑥34 .

Since 𝑑𝑓𝑧 is a linear map with values in ℝ, it suffices to observe that 𝑑𝑓𝑧 is not the zero
map to conclude that 𝑑𝑓𝑧 is surjective for all 𝑧 ∈ ℝ4 . Hence 𝑓 is a submersion.
2
∙ Let 𝑀(𝑛) denote the space of real 𝑛×𝑛-matrices. It is isomorphic as a vector space to ℝ𝑛 ,
since we can write every 𝑛 × 𝑛-matrix as a column vector of length 𝑛2 . Hence 𝑀(𝑛) is
smooth 𝑛2 -dimensional manifold. Let 𝐺𝐿(𝑛) denote the group of invertible 𝑛×𝑛-matrices
with group operation given by matrix multiplication. This is an open subset and hence a
submanifold of 𝑀(𝑛), also of dimension 𝑛2 . The determinant map

det ∶ 𝑀(𝑛) → ℝ

is a submersion at every matrix 𝐴 ∈ 𝐺𝐿(𝑛). We will discuss this claim later in this
section and in Exercise 4.9.

As for immersions, we can — at least locally — make submersions look as simple as pos-
sible. We will discuss this useful observation next.

4.1.2 The Local Submersion Theorem

The possibly simplest surjective map ℝ𝑛 → ℝ𝑚 with 𝑛 ≥ 𝑚 is the map

𝜎𝑚𝑛 ∶ ℝ𝑛 → ℝ𝑚 , (𝑥1 , … , 𝑥𝑚 , 𝑥𝑚+1 , … , 𝑥𝑛 ) → (𝑥1 , … , 𝑥𝑚 )

which sends the coordinates of a point 𝐱 ∈ ℝ𝑛 to the point in ℝ𝑚 with the first 𝑚 coordinates
and forgets the remaining 𝑛 − 𝑚 ones. The map 𝜎𝑚𝑛 is called the canonical submersion from
ℝ𝑛 onto ℝ𝑚 . The matrix which represents 𝜎𝑚𝑛 with respect to the standard bases of ℝ𝑛 and ℝ𝑚 ,
respectively, is the 𝑚 × 𝑛-matrix 𝑆𝑚𝑛 which has the 𝑚 × 𝑚-identity matrix in the first 𝑚 columns
and only zeros in the remaining 𝑛 − 𝑚 columns on the right hand side.

Starting with a linear surjective map 𝐿 ∶ ℝ𝑛 → ℝ𝑚 we can split 𝐿 into a composition of

the form 𝑛
𝛼 𝜎𝑚
ℝ𝑛 ←←→
← ℝ𝑛 ←←←←→
← ℝ𝑚
where 𝛼 is a linear isomorphism of ℝ𝑛 as follows: Since 𝐿 is surjective, the kernel of 𝐿 has
dimension 𝑛−𝑚, and the quotient vector space space ℝ𝑛 ∕Ker 𝐿 has dimension 𝑚. Let 𝐯1 , … , 𝐯𝑚
be vectors in ℝ𝑛 which map to a basis of ℝ𝑛 ∕Ker 𝐿 and such that 𝐿(𝐯1 ) = 𝐞𝑛1 , … , 𝐿(𝐯𝑚 ) = 𝐞𝑛𝑚
in ℝ𝑛 . Let 𝐰1 , … , 𝐰𝑛−𝑚 be a basis of Ker 𝐿. We define 𝛼 by sending the first 𝑚 standard basis
vectors 𝐞𝑛1 , … , 𝐞𝑛𝑚 of ℝ𝑚 to 𝐯1 , … , 𝐯𝑚 and the remaining 𝑛 − 𝑚 basis vectors 𝐞𝑛𝑚+1 , … , 𝐞𝑛𝑛 of ℝ𝑛
to 𝐰1 , … , 𝐰𝑛−𝑚 .
Chapter 4. Submersions and regular values 77
In this way we move all the action into while the second step, where the dimension drops,
ℝ𝑛
always consists of one fixed map which just forgets coordinates.

Again, this is a familiar manoeuvre. For 𝛼 changes the basis of ℝ𝑛 from the standard basis
to a new basis which consists of the vectors 𝐯1 , … , 𝐯𝑚 , 𝐰1 , … , 𝐰𝑛−𝑚 . With respect to this new
basis for ℝ𝑛 and the standard basis for ℝ𝑚 , the matrix which represents 𝐿 is the 𝑚 × 𝑛-matrix
𝑆𝑚𝑛 described above.

We can translate this idea to submersions between smooth manifolds. In other words — up
to diffeomorphisms — the canonical submersion is locally the only submersion:

Theorem 4.2 (Local Submersion Theorem) Let 𝑋 and 𝑌 be smooth manifolds of

dimensions 𝑛 and 𝑚, respectively, with 𝑛 ≥ 𝑚. Suppose that 𝑓 ∶ 𝑋 → 𝑌 is an submer-
sion at 𝑥, and write 𝑦 = 𝑓 (𝑥). Then there exist local coordinates around 𝑥 and 𝑦 such
that

𝑓 (𝑥1 , … , 𝑥𝑚 , 𝑥𝑚+1 , … , 𝑥𝑛 ) = (𝑥1 , … , 𝑥𝑚 ).

More precisely, we can choose local parametrizations 𝜙 ∶ 𝑈 → 𝑋 around 𝑥 and

𝜓 ∶ 𝑊 → 𝑌 around 𝑦 such that in the commutative diagram

𝑓
𝑋O /𝑌
O
𝜙 𝜓
canonical submersion /𝑊
𝑈
𝜃=𝜓 −1 ◦𝑓 ◦𝜙

the map 𝜃 ∶ 𝑈 → 𝑊 is the canonical submersion from ℝ𝑛 to ℝ𝑚 restricted to the open

subset 𝑈 .

Proof of the Local Submersion Theorem: We start by choosing any local parametrizations
𝜙 ∶ 𝑈 → 𝑋 with 𝜙(0) = 𝑥 and 𝜓 ∶ 𝑉 → 𝑌 with 𝜓(0) = 𝑦:

𝑓
𝑋O /𝑌
O
𝜙 𝜓

𝑈 /𝑉
𝜃=𝜓 −1 ◦𝑓 ◦𝜙

Now we are going to manipulate 𝜙 and 𝜓 such that 𝜃 becomes the canonical submersion.

By the assumption, we know 𝑑𝜃0 ∶ ℝ𝑛 → ℝ𝑚 is surjective. Hence, after choosing a suitable

basis for ℝ𝑛 , we can assume that 𝑑𝜃0 is given by the 𝑚 × 𝑛-matrix 𝑆𝑚𝑛 which has the 𝑚 × 𝑚-
identity matrix in the first 𝑚 columns and only zeros in the remaining 𝑛 − 𝑚 columns. We define
a new map

Θ ∶ 𝑈 → ℝ𝑛 , by Θ(𝐱) = (𝜃(𝐱), 𝑥𝑚+1 , … , 𝑥𝑛 )

78 4.2. Regular values
for a point 𝐱 = (𝑥1 , … , 𝑥𝑛 ). It is related to 𝜃 by the commutative diagram

𝑈
𝜃 / ℝ𝑚
<

Θ ! canonical submersion
ℝ𝑛 .
By the construction, the derivative 𝑑Θ0 at 0 is given by the 𝑛 × 𝑛-identity matrix. Hence Θ
is a local diffeomorphism at 0. Thus we can find a small enough neighborhood 𝑈 ′ around 0
in ℝ𝑛 such that Θ−1 exists as a diffeomorphism from 𝑈 ′ ⊂ ℝ𝑛 onto some small neighborhood
around 0 in 𝑈 . By construction, 𝜃 equals the composition of the canonical submersion with Θ,
i.e., we have 𝜃◦Θ−1 = 𝜎𝑚𝑛 on 𝑈 ′ . This gives us the commutative diagram

𝑓
/𝑌
A 𝑋O O
𝜙 𝜓
𝜙◦Θ−1 𝜃 /6 𝑉 .
>𝑈
canonical
Θ−1
submersion
𝑈′

Hence it suffices to replace 𝑈 with 𝑈 ′ and 𝜙 with 𝜙◦Θ−1 to get the desired commutative diagram
𝑓
𝑋O /𝑌
O
𝜙◦𝚯−𝟏 𝜓
canonical /𝑉
𝑈′
submersion

which proves the theorem.

Remark 4.3 (Local property) We observe from this proof that if 𝑓 ∶ 𝑋 → 𝑌 is a

submersion at 𝑥, then it is also a submersion for all points in a neighborhood of 𝑥. For
if 𝑑𝜃0 is surjective, then 𝑑𝜃𝑢 is surjective for all points 𝑢 in a small open neighborhood
of 0 in 𝑈 .

4.2 Regular values

4.2.1 Manifolds as preimages

Suppose we have a smooth map between manifolds 𝑓 ∶ 𝑋 → 𝑌 and a point 𝑦 ∈ 𝑌 . We would

like to understand the geometry or topology of the fiber of 𝑓 over 𝑦, i.e., the set

𝑓 −1 (𝑦) = {𝑥 ∈ 𝑋 ∶ 𝑓 (𝑥) = 𝑦} ⊆ 𝑋.

For example, we would very much like 𝑓 −1 (𝑦) to be a smooth manifold itself. Remember
that this was a situation we started with when we defined manifolds. The fiber 𝑓 −1 (𝑦) is the set
of solutions of the equation 𝑓 (𝑥) = 𝑦.
Chapter 4. Submersions and regular values 79

Remark 4.4 (Be aware) Unfortunately, in general, there is no reason for the set 𝑓 −1 (𝑦)
to be a manifold.

∙ However, life is much nicer in the world of submersions:

We now assume that 𝑓 ∶ 𝑋 → 𝑌 is a submersion at a point 𝑥0 ∈ 𝑋. Assume 𝑛 = dim 𝑋 ≥

dim 𝑌 = 𝑚. Set 𝑦 = 𝑓 (𝑥0 ) or in other words 𝑥0 ∈ 𝑓 −1 (𝑦). We choose local parametrizations
𝜙 ∶ 𝑈 → 𝑉 ⊂ 𝑋 and 𝜓 ∶ 𝑊 → 𝑊 ′ ⊂ 𝑌 around 𝑥 and 𝑦, respectively, with 𝜙(0) = 𝑥0 and
𝜓(0) = 𝑦 and open subsets 𝑈 ⊂ ℝ𝑛 and 𝑊 ⊂ ℝ𝑚 . By the Local Submersion Theorem 4.2, we
can choose these parametrizations such that 𝜃 = 𝜓 −1 ◦𝑓 ◦𝜙 becomes the canonical submersion
ℝ𝑛 → ℝ𝑚 restricted to 𝑈 :
𝑓
𝑋O /𝑌
O
open open
? ?
 /𝑉
𝑓|𝑉
/𝑊′
𝑉 ∩ 𝑓 −1 (𝑦) O O
𝜙 ≅ ≅ 𝜓
𝑢1 =⋯=𝑢𝑚 =0
$ 𝜃 = canonical /𝑊.
𝑈
submersion

Since 𝜓 −1 (𝑦) = 0 and since 𝜙 and 𝜓 are diffeomorphisms, we have

𝑥 ∈ 𝑉 ∩ 𝑓 −1 (𝑦) ⇐⇒ 𝜃(𝜙−1 (𝑥)) = 0.

We now write ( )
𝜙−1 (𝑥) =∶ 𝑢(𝑥) = 𝑢1 (𝑥), 𝑢2 (𝑥), … , 𝑢𝑛 (𝑥)
to make the local coordinates in 𝑈 of points in 𝑥 ∈ 𝑉 explicit. With this notation, we can
rewrite the above equivalence as
( )
𝑥 ∈ 𝑉 ∩ 𝑓 −1 (𝑦) ⇐⇒ 𝜃 𝑢1 (𝑥), 𝑢2 (𝑥), … , 𝑢𝑛 (𝑥) = 0.

Now, since 𝜃 is the canonical submersion, we can rewrite this yet again as

𝑥 ∈ 𝑉 ∩ 𝑓 −1 (𝑦) ⇐⇒ 𝑢1 (𝑥) = 𝑢2 (𝑥) = ⋯ = 𝑢𝑚 (𝑥) = 0.

Hence in terms of sets we get

𝑓 −1 (𝑦) ∩ 𝑉 = {𝑥 ∈ 𝑉 ∶ 𝑢1 (𝑥) = ⋯ = 𝑢𝑚 (𝑥) = 0}.

However, on 𝑓 −1 (𝑦) ∩ 𝑉 , the remaining local coordinate functions 𝑢𝑚+1 (𝑥), … , 𝑢𝑛 (𝑥) do not
have to satisfy any additional condition. Hence we can use them to define a local coordinate
system in ℝ𝑛−𝑚 for points 𝑥 ∈ 𝑓 −1 (𝑦) ∩ 𝑉 . In fact, the subset 𝑓 −1 (𝑦) ∩ 𝑉 ⊂ 𝑓 −1 (𝑦) is open
in 𝑓 −1 (𝑦), since 𝑉 is open in 𝑋 and 𝑓 −1 (𝑦) has the subspace topology. Similarly, the subset
({0} × ℝ𝑛−𝑚 ) ∩ 𝑈 is open in {0} × ℝ𝑛−𝑚 , since 𝑈 is open in ℝ𝑛 . We now identify {0} × ℝ𝑛−𝑚
with ℝ𝑛−𝑚 and write 𝑈 ′ ⊂ ℝ𝑛−𝑚 for the open subset corresponding to ({0} × ℝ𝑛−𝑚 ) ∩ 𝑈 .

Summarizing, we have shown that the restriction of 𝜙−1 to 𝑓 −1 (𝑦) ∩ 𝑉

( )
𝑓 −1 (𝑦) ∩ 𝑉 → ({0} × ℝ𝑛−𝑚 ) ∩ 𝑈 , 𝑥 → 𝑢𝑚+1 (𝑥), … , 𝑢𝑛 (𝑥)
80 4.2. Regular values
defines a local coordinate system for 𝑓 −1 (𝑦) around 𝑥0 . In particular, this is a diffeomorphism
between open subsets. We write
𝜙′ ∶ ({0} × ℝ𝑛−𝑚 ) ∩ 𝑈 =∶ 𝑈 ′ → 𝑓 −1 (𝑦) ∩ 𝑉
for the inverse diffeomorphism where 𝑈 ′ ⊂ ℝ𝑛−𝑚 is an open subset. That is 𝜙′ is a local
parametrization around 𝑥0 in 𝑓 −1 (𝑦).

We would like this to be possible for every point in the fiber 𝑓 −1 (𝑦). This is not always the
case. So let us give the desired case a name:

Definition 4.5 (Regular value) For a smooth map of manifolds 𝑓 ∶ 𝑋 → 𝑌 , a point

𝑦 ∈ 𝑌 is called a regular value for 𝑓 if 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑦 (𝑌 ) is surjective at every
point 𝑥 ∈ 𝑋 such that 𝑓 (𝑥) = 𝑦.

The above argument shows the following very important result:

Theorem 4.6 (Preimage Theorem) If 𝑦 is a regular value for 𝑓 ∶ 𝑋 → 𝑌 , then the

fiber 𝑓 −1 (𝑦) over 𝑦 is a submanifold of 𝑋, with dim 𝑓 −1 (𝑦) = dim 𝑋 − dim 𝑌 .

Remark 4.7 The above Theorem 4.6 is actually one of the main tools to show that a
space is a smooth manifold. It is an important mile stone on our journey and we should
always have it in mind whenever we are asked to show that a space has the structure of
a smooth manifold.

Before we look at some examples, we determine the tangent spaces of the submanifold given
as the preimage of a regular value:

Lemma 4.8 (Tangent space of a regular fiber) Let 𝑍 = 𝑓 −1 (𝑦) be the fiber over
a regular value 𝑦 ∈ 𝑌 under the smooth map 𝑓 ∶ 𝑋 → 𝑌 . Then the tangent space to
𝑇𝑥 (𝑍) at a point 𝑥 ∈ 𝑍 is the kernel of the derivative

𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑦 (𝑌 ),

i.e., 𝑇𝑥 (𝑍) = Ker (𝑑𝑓𝑥 ).

Proof: Since 𝑓 (𝑍) = 𝑦, 𝑓 is constant on 𝑍. Therefore, 𝑑𝑓𝑥 vanishes on the subspace

𝑇𝑥 (𝑍) ⊂ 𝑇𝑥 (𝑋). Hence 𝑑𝑓𝑥 sends all of 𝑇𝑥 (𝑍) to zero. More concretely, using the notation of
the argument to prove Theorem 4.6, we have the following commutative diagram:

𝑇𝑥 (𝑍) 
𝑑𝑓𝑥
/ 𝑇 (𝑋) / 𝑇 (𝑌 )
O 𝑥 O 𝑦
O
𝑑𝜙′0 𝑑𝜙0 𝑑𝜓0
𝑛−𝑚  
𝑑𝜃0 =𝜎𝑚𝑛
ℝ / ℝ𝑛 / ℝ𝑚 .

Since 𝜎𝑚𝑛 is the canonical submersion, the composition of the lower horizontal map is zero.
Since 𝑇𝑥 (𝑍) is defined as the image of 𝑑𝜙′0 and since the diagram commutes, this shows that
𝑑𝑓𝑥 sends all of 𝑇𝑥 (𝑍) to 0. This proves 𝑇𝑥 (𝑍) ⊆ Ker 𝑑𝑓𝑥 .
Chapter 4. Submersions and regular values 81
On the other hand, since 𝑦 is a regular value, 𝑑𝑓𝑥 is surjective. Hence the dimension of the
kernel of 𝑑𝑓𝑥 is dim 𝑇𝑥 (𝑋) − dim 𝑇𝑦 (𝑌 ) = dim 𝑋 − dim 𝑌 = dim 𝑍. This shows that 𝑇𝑥 (𝑍) is
a subspace of the kernel of 𝑑𝑓𝑥 of the same dimension as Ker 𝑑𝑓𝑥 . Thus 𝑇𝑥 (𝑍) = Ker 𝑑𝑓𝑥 .

4.2.2 First examples

∙ Let 𝑓 ∶ ℝ𝑛+1 → ℝ be the map

𝑥 = (𝑥1 , … , 𝑥𝑛+1 ) → |𝑥|2 = 𝑥21 + ⋯ + 𝑥2𝑛+1 .

The derivative 𝑑𝑓𝑥 at the point 𝑥 = (𝑥1 , … , 𝑥𝑛+1 ) is the linear map given by the matrix
(2𝑥1 … 2𝑥𝑛+1 ) expressed in the standard basis. Thus 𝑑𝑓𝑥 ∶ ℝ𝑛+1 → ℝ is surjective un-
less 𝑓 (𝑥) = 0, so every nonzero real number is a regular value of 𝑓 . In particular, we get
again that the sphere 𝕊𝑛 = 𝑓 −1 (1) is an 𝑛-dimensional manifold.

∙ We will show in the exercises that 1 ∈ ℝ is a regular value of the determinant function
det ∶ 𝐺𝐿(𝑛) → ℝ. This will show that the subgroup 𝑆𝐿(𝑛) of matrices with determinant
one is a smooth manifold. It is is called the special linear group. More on such groups
later.

∙ Recall the map

𝑓 ∶ ℝ4 → ℝ,
(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) → 𝑥1 + 𝑥22 + 𝑥33 + 𝑥44 .

We showed above that 𝑓 is a submersion. In particular, this implies that 0 is a regular

value for 𝑓 . Hence the preimage

𝑍 = 𝑓 −1 (0) = {(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) ∈ ℝ4 ∶ 𝑥1 + 𝑥22 + 𝑥33 + 𝑥44 = 0}

is a submanifold in ℝ4 . Its dimension is dim ℝ4 − dim ℝ = 3.

∙ Let 𝑝1 , … , 𝑝𝑛 be polynomials with real coefficients in variables 𝑡1 , … , 𝑡𝑚 . We can con-

sider the collection (𝑝1 , … , 𝑝𝑛 ) as a smooth map 𝐩 ∶ ℝ𝑚 → ℝ𝑛 . Then the set of common
zeros of 𝑝1 , … , 𝑝𝑛 is a smooth submanifold of ℝ𝑚 if 𝟎 = (0, … , 0) ∈ ℝ𝑛 is a regu-
lar value of 𝑝. This is the case if the Jacobian matrix 𝐽𝐬 at a point 𝐬 = (𝑠1 , … , 𝑠𝑚 ),
𝜕𝑝
i.e., the matrix with (𝑖, 𝑗)-entry the partial derivative 𝜕𝑡 𝑖 (𝐬), has maximal rank when-
𝑗
ever 𝐩(𝐬) = 𝟎. For example, if 𝑛 = 𝑚, this requires that det 𝐽𝐬 ≠ 0. While, if 𝑛 = 1
and 𝑚 > 1, then this only requires that the partial derivatives do not vanish simultane-
ously. Such a set of zeros of polynomials is often referred to as an algebraic variety.

Remark 4.9 (Algebraic Geometry in a nutshell) The study of the zeroes of poly-
nomials is the central theme in Algebraic Geometry. This is a classical and fascinating
82 4.2. Regular values
part of pure mathematics. In the past three decades, deep and fascinating connections
between algebraic geometry and homotopy theory have been developed. This is the field
of Motivic Homotopy Theory.

Since 𝑓 −1 (𝑦) can be very complicated if 𝑦 is not regular, the values which are not regular
get the following name:

Definition 4.10 (Critical values) For a smooth map of manifolds 𝑓 ∶ 𝑋 → 𝑌 , a point

𝑦 ∈ 𝑌 which is not a regular value, is called a critical value for 𝑓 .

∙ For example, we will show in Exercise 4.9 that the only critical value of the determinant
map det ∶ 𝑀(𝑛) → ℝ is 0.

∙ Note: All values 𝑦 which are not in the image of 𝑓 also are regular values for 𝑓 .
For, if 𝑓 −1 (𝑦) is the empty set, then there is no condition to be satisfied.

Depending on the relative dimensions of 𝑋 and 𝑌 we can now say how regular values can
arise.

Remark 4.11 (Summary for regular values) Suppose 𝑓 ∶ 𝑋 → 𝑌 is a smooth map

of manifolds. Then 𝑦 being a regular value for 𝑓 has the following meaning:

∙ if dim 𝑋 > dim 𝑌 : 𝑦 is a regular value if and only if 𝑓 is a submersion at each

point 𝑥 ∈ 𝑓 −1 (𝑦);

∙ if dim 𝑋 = dim 𝑌 : 𝑦 is a regular value if and only if 𝑓 is a local diffeomorphism

at each point 𝑥 ∈ 𝑓 −1 (𝑦);

∙ if dim 𝑋 < dim 𝑌 : 𝑦 is a regular value if and only if 𝑦 is not in the image of 𝑓 ; for,
all values in the image are critical (𝑑𝑓𝑥 cannot be surjective when dim 𝑇𝑥 (𝑋) <
dim 𝑇𝑓 (𝑥) (𝑌 )).

4.2.3 More examples: Matrix subgroups are manifolds

A very important application of the Preimage Theorem is that we can use it to show that various
matrix groups are smooth manifolds. Let 𝑀(𝑛) again denote the space of real 𝑛 × 𝑛-matrices
and 𝐺𝐿(𝑛) denote the group of invertible matrices in 𝑀(𝑛).

Let 𝑂(𝑛) be the subgroup of matrices 𝐴 of 𝐺𝐿(𝑛) which satisfy 𝐴𝐴𝑡 = 𝐼 where 𝐴𝑡 denotes
the transpose of 𝐴 and 𝐼 is the 𝑛×𝑛-identity matrix. Note that 𝑂(𝑛) is the subgroup of invertible
matrices which preserve the scalar product of vectors. In particular, matrices in 𝑂(𝑛) preserve
distances and angles in ℝ𝑛 .

Our goal is to show:

Chapter 4. Submersions and regular values 83

Figure 4.1: Orthogonal transformations preserve the geometry of the plane.

Theorem 4.12 (Orthogonal matrices form a smooth manifold) The group 𝑂(𝑛) of
orthogonal matrices is a smooth manifold of dimension 𝑛(𝑛 − 1)∕2.

Proof: First, we note that 𝐴𝐴𝑡 is a symmetric matrix:

(𝐴𝐴𝑡 )𝑡 = (𝐴𝑡 )𝑡 𝐴𝑡 = 𝐴𝐴𝑡 .

The subspace 𝑆(𝑛) of symmetric matrices is a smooth submanifold of 𝑀(𝑛) diffeomorphic

to ℝ𝑘 with 𝑘 = 𝑛(𝑛 + 1)∕2. The dimension is determined by the fact that the entries below the
diagonal are determined by the entries above the diagonal such that there are 𝑛(𝑛 + 1)∕2 free
choices. We consider the map

𝑓 ∶ 𝑀(𝑛) → 𝑆(𝑛), 𝐴 → 𝐴𝐴𝑡 .

This map is smooth, since multiplication of matrices is smooth and taking transposes is smooth
as well as we just reshuffle the entries in the matrix. Now we observe 𝑂(𝑛) = 𝑓 −1 (𝐼). Hence,
in order to show that 𝑂(𝑛) is a smooth manifold, we just need to show that 𝐼 is a regular value
for 𝑓 . So let us compute the derivative of 𝑓 at a matrix 𝐴:
𝑓 (𝐴 + 𝑠𝐵) − 𝑓 (𝐴) (𝐴 + 𝑠𝐵)(𝐴 + 𝑠𝐵)𝑡 − 𝐴𝐴𝑡
𝑑𝑓𝐴 (𝐵) = lim = lim
𝑠→0 𝑠 𝑠→0 𝑠
(𝐴 + 𝑠𝐵)(𝐴𝑡 + 𝑠𝐵 𝑡 ) − 𝐴𝐴𝑡 𝐴𝐴𝑡 + 𝑠𝐵𝐴𝑡 + 𝑠𝐴𝐵 𝑡 + 𝑠2 𝐵𝐵 𝑡 − 𝐴𝐴𝑡
= lim = lim
𝑠→0 𝑠 𝑠→0 𝑠
𝑠𝐵𝐴𝑡 + 𝑠𝐴𝐵 𝑡 + 𝑠2 𝐵𝐵 𝑡
= lim = lim 𝐵𝐴𝑡 + 𝐴𝐵 𝑡 + 𝑠𝐵𝐵 𝑡
𝑠→0 𝑠 𝑠→0
= 𝐴𝐵 𝑡 + 𝐵𝐴𝑡 .

In order to check that 𝐼 is a regular value, we need to show that

𝑑𝑓𝐴 ∶ 𝑇𝐴 (𝑀(𝑛)) → 𝑇𝑓 (𝐴) (𝑆(𝑛))

2
is surjective for all 𝐴 ∈ 𝑂(𝑛). Since 𝑀(𝑛) ≅ ℝ𝑛 and 𝑆(𝑛) ≅ ℝ𝑛(𝑛+1)∕2 are vector spaces, we
have
𝑇𝐴 (𝑀(𝑛)) = 𝑀(𝑛) and 𝑇𝑓 (𝐴) (𝑆(𝑛)) = 𝑆(𝑛).
84 4.3. The Stack of Records Theorem
Hence, given a matrix 𝐶 ∈ 𝑆(𝑛), we need to show that there is a matrix 𝐵 ∈ 𝑀(𝑛) with
𝑑𝑓𝐴 (𝐵) = 𝐵𝐴𝑡 + 𝐴𝐵 𝑡 = 𝐶. Since 𝐶 is symmetric, we have
1 1 1( )
𝐶 = (2𝐶) = (𝐶 + 𝐶) = 𝐶 + 𝐶𝑡 .
2 2 2
Now we set 𝐵 = 21 𝐶𝐴 and compute, using 𝐴𝐴𝑡 = 𝐼:
( ) ( )𝑡
1 1
𝑑𝑓𝐴 (𝐵) = 𝐶𝐴 𝐴𝑡 + 𝐴 𝐶𝐴
2 2
1 𝑡 1 𝑡 𝑡
= 𝐶𝐴𝐴 + 𝐴𝐴 𝐶
2 2
1 1 𝑡
= 𝐶+ 𝐶
2 2
= 𝐶.

Thus 𝐼 is a regular value, and 𝑂(𝑛) is a submanifold of 𝑀(𝑛). We can also calculate the
dimension of 𝑂(𝑛):
𝑛(𝑛 + 1) 𝑛(𝑛 − 1)
dim 𝑂(𝑛) = dim 𝑀(𝑛) − dim 𝑆(𝑛) = 𝑛2 − = .
2 2

We end this section with an outlook to an exciting field which has ramifications to almost
all areas of mathematics:

Remark 4.13 (Lie groups) The manifold 𝑂(𝑛) is an example of a very important
class of smooth manifolds. For, 𝑂(𝑛) is both a smooth manifold and a group such that
the group operations are smooth. For both the multiplication map

𝑂(𝑛) × 𝑂(𝑛) → 𝑂(𝑛), (𝐴, 𝐵) → 𝐴𝐵

and the map of forming the inverse

𝑂(𝑛) → 𝑂(𝑛), 𝐴 → 𝐴−1

are smooth. For the latter note 𝐴−1 = 𝐴𝑡 for 𝐴 ∈ 𝑂(𝑛), though taking inverses is also
smooth for other matrix groups.
In general, a group which is also a manifold such that the group operations are
smooth is called a Lie group. Lie groups are extremely interesting and important
and have a rich and exciting theory. For example, the tangent space at a Lie group at the
identity element is a Lie algebra, a vector space with a certain additional operation. Such
Lie algebras can be classified completely. Lie groups and Lie algebras play an important
role in many different areas of mathematics and physics. We will take a closer look,
though by now way sufficient for their very rich theory, at Lie groups in chapter 5.

4.3 The Stack of Records Theorem

We will now take a closer look at a specific and important situation for regular values. This will
be useful later at several occasions.
Chapter 4. Submersions and regular values 85
Suppose 𝑓 ∶ 𝑋 → 𝑌 is a smooth map with dim 𝑋 = dim 𝑌 and 𝑋 compact. Our goal is to
understand the fibers 𝑓 −1 (𝑦) of regular values. We do this in a series of observations:

∙ Claim: The set 𝑅 of regular values for 𝑓 is open in 𝑌 .

Proof of the claim: Let 𝑥0 ∈ 𝑋 be a regular point for 𝑓 , i.e., 𝑑𝑓𝑥0 is surjective. Since
we assume dim 𝑋 = dim 𝑌 , this actually implies that 𝑑𝑓𝑥0 is an isomorphism. By the Inverse
Function Theorem 3.4, there is an open neighborhood 𝑊 around 𝑥0 such that 𝑑𝑓𝑥 is an iso-
morphism for all 𝑥 ∈ 𝑊 . Hence the set of regular points is open in 𝑋. Consequently, its
complement in 𝑋, the set of critical points 𝐶 is a closed subset in 𝑋. Since 𝑋 is compact,
this implies that 𝐶 is a compact subset. Since 𝑓 is continuous, 𝑓 (𝐶) is a compact subset in
𝑌 . In particular, 𝑓 (𝐶) is closed in 𝑌 .1 The set 𝑓 (𝐶) is the set of critical values of 𝑓 and its
complement is the set 𝑅 of regular values. Hence 𝑅 is open in 𝑌 .

∙ Now let 𝑦 ∈ 𝑌 be a regular value for 𝑓 such that 𝑓 −1 (𝑦) is not empty. Let 𝑥 be a point in
𝑓 −1 (𝑦). Since 𝑦 is a regular value, 𝑥 is a regular point, i.e., 𝑑𝑓𝑥 is surjective and hence an
isomorphism as dim 𝑋 = dim 𝑌 . Hence 𝑓 is a local diffeomorphism at 𝑥. Let 𝑉 ⊂ 𝑋
and 𝑈 ⊂ 𝑌 be open neighborhoods around 𝑥 and 𝑦, respectively, such that 𝐟|𝐕 ∶ 𝑉 → 𝑈
is a diffeomorphism.

∙ If 𝑥′ is another point in 𝑓 −1 (𝑦) with 𝑥 ≠ 𝑥′ . Then 𝑑𝑓𝑥′ is an isomorphism as well, and

we can choose an open neighborhood 𝑉 ′ ⊂ 𝑋 around 𝑥′ such that 𝑓|𝑉 ′ ∶ 𝑉 ′ → 𝑈 ′ is a
diffeomorphism onto an open subset 𝑈 ′ ⊂ 𝑌 containing 𝑦.

∙ Claim: 𝑉 and 𝑉 ′ are disjoint.

Proof of the claim: If 𝑉 ∩ 𝑉 ′ ≠ ∅, then 𝑓 restricts to a diffeomorphism from 𝑉 ∩ 𝑉 ′ onto

𝑈 ∩ 𝑈 ′ . Since 𝑦 ∈ 𝑈 ∩ 𝑈 ′ and 𝑓 (𝑥) = 𝑦 = 𝑓 (𝑥′ ), this would imply 𝑥 = 𝑥′ ∈ 𝑉 ∩ 𝑉 ′ , since 𝑥
and 𝑥′ are the unique points in 𝑉 and 𝑉 ′ , respectively, which map to 𝑦. So if 𝑥 ≠ 𝑥′ , we must
have 𝑉 ∩ 𝑉 ′ = ∅.

∙ Hence all the points in 𝑓 −1 (𝑦) lie in pairwise disjoint open subsets of 𝑋. We conclude
that 𝑓 −1 (𝑦) is discrete, i.e., it is a space in which every subset is open. Since the subset
{𝑦} is closed in 𝑌 , the fiber 𝑓 −1 (𝑦) is a closed subset of 𝑋. Since 𝑋 is compact and
since closed subsets in compact spaces are compact, this implies that 𝑓 −1 (𝑦) is compact
as well. Hence it is a compact and discrete subset of Euclidean space. We have shown
previously that this implies that 𝑓 −1 (𝑦) is a finite set.

∙ Let 𝑓 −1 (𝑦) = {𝑥1 , … , 𝑥𝑛 }. We can thus pick pairwise disjoint open subsets 𝑊1 , … , 𝑊𝑛
in 𝑋 with 𝑥𝑖 ∈ 𝑊𝑖 and open subsets 𝑈1 , … , 𝑈𝑛 in 𝑌 , each containing 𝑦, such that 𝑓 maps
𝑊𝑖 diffeomorphically onto open 𝑈𝑖 . We define the set

𝑈 ∶= (𝑈1 ∩ ⋯ ∩ 𝑈𝑛 ) − 𝑓 (𝑋 − (𝑊1 ∪ … ∪ 𝑊𝑛 )).

1
This follows from the fact that 𝑌 is Hausdorff or from our assumption that 𝑌 is a subspace in ℝ𝑁 and there
compact subsets are in particular closed.
86 4.3. The Stack of Records Theorem

Figure 4.2: Every point 𝑦 in 𝑌 has an open neighborhood 𝑈 such that 𝑓 −1 (𝑈 ) consists of disjoint
copies of open subsets 𝑉𝑖 in 𝑋 which are diffeomorphic to 𝑈 . In three-dimensional space this
may be drawn as several old school records stacked on top of each other. One key consequence
is that the fiber over 𝑦 is a finite and discrete set.

∙ Claim: 𝑈 is an open subset in 𝑌 .

Proof: First, the finite intersection 𝑈 ∶= 𝑈1 ∩ ⋯ ∩ 𝑈𝑛 is open in 𝑌 . Second, the set
𝑊 ∶= 𝑊1 ∪ … ∪ 𝑊𝑛 is open in 𝑋 and its complement 𝑋 − 𝑊 in 𝑋 is closed. Since 𝑋
is compact, 𝑋 − 𝑊 is compact and hence 𝑓 (𝑋 − 𝑊 ) is compact. This implies again that
it also is a closed subset in 𝑌 . Hence 𝑓 (𝑋 − 𝑊 ) is closed in 𝑈1 ∩ ⋯ ∩ 𝑈𝑛 . This shows
that 𝑈 is open.
∙ We also know 𝑦 ∈ 𝑈 , since 𝑦 ∈ 𝑈1 ∩ ⋯ ∩ 𝑈𝑛 and 𝑦 ∉ 𝑓 (𝑋 − 𝑊 ) by our choices of the
𝑈𝑖 ’s and 𝑊𝑖 ’s.
∙ We set 𝑉𝑖 ∶= 𝑊𝑖 ∩ 𝑓 −1 (𝑈 ). Then 𝑓|𝑉𝑖 ∶ 𝑉𝑖 → 𝑈 is a diffeomorphism and maps 𝑥𝑖 to 𝑦.

∙ Thus the inverse image 𝑓 −1 (𝑈 ) is a disjoint union of open subsets 𝑉1 , … , 𝑉𝑛 with 𝑥𝑖 ∈ 𝑉𝑖

and the restriction of 𝑓 to any of the 𝑉𝑖 is a diffeomorphism onto 𝑈 .

Hence we have shown the following very useful result:

Theorem 4.14 (Stack of Records Theorem) Suppose dim 𝑋 = dim 𝑌 , 𝑓 ∶ 𝑋 → 𝑌

is a smooth map and 𝑋 is compact. Let 𝑦 ∈ 𝑌 be a regular value for 𝑓 such that
𝑓 −1 (𝑦) ≠ ∅. Then:

∙ the set 𝑓 −1 (𝑦) is a discrete finite subset {𝑥1 , … , 𝑥𝑛 } of 𝑋, and

∙ we can choose an open neighborhood 𝑈 ⊂ 𝑌 around 𝑦 such that 𝑓 −1 (𝑈 ) ⊂ 𝑋 is

Chapter 4. Submersions and regular values 87
the disjoint union 𝑉1 ∪⋯∪𝑉𝑛 of open subsets of 𝑋 with 𝑥𝑖 ∈ 𝑉𝑖 and the restriction
of 𝑓 to any of the 𝑉𝑖 is a diffeomorphism onto 𝑈 .

This is a particular instance of an important phenomenon:

∙ If in addition to the assumptions of the theorem all values in 𝑌 are regular, then 𝑋 → 𝑌 is
an example of a covering. In Topology, a continuous map 𝑓 ∶ 𝑋 → 𝑌 is an unramified
covering if every point in 𝑌 has an open neighborhood 𝑈 such that 𝑓 −1 (𝑈 ) is the disjoint
union of open sets 𝑉𝑖 such that 𝑓 maps each 𝑉𝑖 homeomorphically onto 𝑈 . Coverings
play an important role in topology and homotopy theory.

Since 𝑓 −1 (𝑦) is either empty or finite, it makes sense to talk about the number of elements
in the set 𝑓 −1 (𝑦). We denote this number by #𝑓 −1 (𝑦).

Lemma 4.15 (Locally constant fiber) The function 𝑦 → #𝑓 −1 (𝑦) defined on the set
of regular points for 𝑓 is locally constant, i.e., for every regular value 𝑦0 there is an
open neighborhood 𝑈 ⊂ 𝑌 of 𝑦0 such that #𝑓 −1 (𝑦) = #𝑓 −1 (𝑦0 ) for all 𝑦 ∈ 𝑈 .

Proof: We split the proof into considering the cases 𝑓 −1 (𝑦0 ) ≠ ∅ and 𝑓 −1 (𝑦0 ) = ∅ sepa-
rately.

∙ First we assume 𝑓 −1 (𝑦0 ) ≠ ∅. Then we can choose the open neighborhood 𝑈 of 𝑦0 in 𝑌

such that 𝑓 −1 (𝑈 ) = 𝑉1 ∪ ⋯ ∪ 𝑉𝑛 is the pairwise disjoint union of open neighborhoods
𝑉𝑖 of 𝑥𝑖 which are all mapped diffeomorphically onto the open subset 𝑈 . Thus, for every
point 𝑦 ∈ 𝑈 , there is exactly one point in 𝑉𝑖 which maps to 𝑦. And these are the only
points in 𝑋 which map onto 𝑦 by the choice of 𝑈 . Thus

#𝑓 −1 (𝑦) = #𝑓 −1 (𝑦0 ) for all 𝑦 ∈ 𝑈

and the function 𝑦 → #𝑓 −1 (𝑦) is constant on the open subset 𝑈 .

∙ Now assume 𝑓 −1 (𝑦0 ) = ∅. Since 𝑋 is compact, 𝑓 (𝑋) is compact and hence closed in 𝑌 .
Thus 𝑌 − 𝑓 (𝑋) is open in 𝑌 . Since by our assumption 𝑦0 ∈ 𝑌 − 𝑓 (𝑋), there is an open
neighborhood 𝑈 ∶ 𝑌 − 𝑓 (𝑋) around 𝑦0 such that

#𝑓 −1 (𝑦) = #𝑓 −1 (𝑦0 ) for all 𝑦 ∈ 𝑈 .

4.4 Milnor’s proof of the Fundamental Theorem of Algebra

As an application of regular values we will now study Milnor’s proof of the Fundamental The-
orem of Algebra:
88 4.4. Milnor’s proof of the Fundamental Theorem of Algebra

Theorem 4.16 (Fundamental Theorem of Algebra) Every non-constant complex

polynomial
𝑃 (𝑧) = 𝑧𝑛 + 𝑎𝑛−1 𝑧𝑛−1 + ⋯ + 𝑎1 𝑧 + 𝑎0
must have a zero in ℂ.

As a consequence, 𝑃 (𝑧) must have exactly 𝑛 zeroes when we count them with multiplici-
ties.

∙ (Outline of the proof) Let us first summarize the idea of the proof:

∙ We will show that 𝑃 ∶ ℂ → ℂ is onto. So 0 is a value of 𝑃 .

∙ We do this by replacing 𝑃 with a smooth map 𝑓 ∶ 𝕊2 → 𝕊2 such that: 𝑓 is onto

if and only if 𝑃 is onto. The key feature is that 𝕊2 is compact while ℂ is not!

∙ We show that the assumption that 𝑓 is not onto leads to a contradiction as fol-
lows:

∙ The Stack of Records Theorem implies that the function 𝑦 → #𝑓 −1 (𝑦)

is constant on the set of regular values on 𝕊2 . Here we use that 𝑃 is a
polynomial such that 𝑑𝑃𝑧 has only finitely many critical points.
∙ If 𝑓 was not onto, #𝑓 −1 (𝑦) = 0 for all regular values.
∙ Then 𝑓 would have to be constant, since 𝕊2 ⧵{critical values} is connected.
∙ But 𝑓 is not constant, since 𝑃 is not constant.

If you wonder why this does not work for real polynomials, then note:

Remark 4.17 (ℝ is too one-dimensional) If we tried the above strategy for a real
polynomial 𝑃 ∶ ℝ → ℝ, we would replace 𝑃 with a smooth map 𝑓 ∶ 𝕊1 → 𝕊1 . Then 𝑓
is smooth defined on a compact domain and 𝑓 has only finitely many critical values. But
the conclusion that the function 𝑦 → #𝑓 −1 (𝑦) is constant on the set of regular values on
𝕊1 would not work anymore. For, after removing at least two different points from 𝕊1
we get a space which is not connected.

Now we are going to see the proof.

The proof:

We are going to identify the complex numbers ℂ with the points in real plane ℝ2 , but we
keep in mind how that we can multiply and form inverses for points in ℂ. To prove the theorem
we need to extend the map 𝑃 ∶ ℂ → ℂ to a map on a compact space. Recall that 𝕊2 is a compact
subspace of ℝ3 and that we can relate 𝕊2 and the real plane ℝ2 via stereographic projection,
recall Figure 2.12.
Chapter 4. Submersions and regular values 89
The formulae for the projection from the north pole 𝑁 = (0, 0, 1) ∈ 𝕊2 are

1
𝜙−1 2 2
𝑁 ∶ 𝕊 ⧵ {𝑁} → ℝ , (𝑥1 , 𝑥2 , 𝑥3 ) → (𝑥 , 𝑥 ) and
1 − 𝑥3 1 2
1 ( )
𝜙𝑁 ∶ ℝ2 → 𝕊2 ⧵ {𝑁}, (𝑥1 , 𝑥2 ) → 2𝑥1 , 2𝑥2 , |𝑥|2 − 1 .
1 + |𝑥| 2

The formulae for the projection from the south pole 𝑆 = (0, 0, −1) ∈ 𝕊2 :

1
𝜙−1 2 2
𝑆 ∶ 𝕊 ⧵ {𝑆} → ℝ , (𝑥1 , 𝑥2 , 𝑥3 ) → (𝑥 , 𝑥 ) and
1 + 𝑥3 1 2
1 ( )
𝜙𝑆 ∶ ℝ2 → 𝕊2 ⧵ {𝑆}, (𝑥1 , 𝑥2 ) → 2𝑥1 , 2𝑥2 , 1 − |𝑥|2 .
1 + |𝑥| 2

Considering our polynomial 𝑃 as a map from ℂ to ℂ we define a new map

{
𝑓 (𝑥) ≔ 𝜙𝑁 ◦𝑃 ◦𝜙−1 (𝑥) for all 𝑥 ∈ 𝕊2 ⧵ {𝑁}
𝑓 ∶ 𝕊2 → 𝕊2 , 𝑁 (4.1)
𝑓 (𝑁) ≔ 𝑁 for 𝑥 = 𝑁.

Lemma 4.18 (First claim: Smoothness) The map 𝑓 ∶ 𝕊2 → 𝕊2 is smooth.

Proof: Since 𝜙𝑁 and 𝜙−1 𝑁

are smooth and polynomials are smooth as well, it is clear that 𝑓
is smooth at all points which are not the north pole.

∙ It remains to show that it is also smooth in a neighborhood of 𝑁.

In order to do this we use the projection from the south pole and define a map

𝑄 ∶ ℂ → ℂ by 𝑄 ∶= 𝜙−1
𝑆 ◦𝑓 ◦𝜙𝑆 .

∙ Note that 𝑄(0) = 0, since 𝜙𝑆 (0) = 𝑁, 𝑓 (𝑁) = 𝑁 and 𝜙−1

𝑆
(𝑁) = 0.

To compute 𝑄 for other values we need to calculate the composite 𝜙−1 𝑁

◦𝜙𝑆 . For 𝑥 =
(𝑥1 , 𝑥2 ) ≠ (0, 0) we get
( )
−1 −1 1 2
𝜙𝑁 ◦𝜙𝑆 (𝑥1 , 𝑥2 ) = 𝜙𝑁 (2𝑥1 , 2𝑥2 , 1 − |𝑥| )
1 + |𝑥|2
( )
1 2𝑥1 2𝑥2
= ,
1−|𝑥|2
1 − 1+|𝑥| 1 + |𝑥|2 1 + |𝑥|2
2
( )
1 + |𝑥| 2 2𝑥1 2𝑥2
= ,
2|𝑥|2 1 + |𝑥|2 1 + |𝑥|2
1
= (𝑥1 , 𝑥2 ).
|𝑥|2
90 4.4. Milnor’s proof of the Fundamental Theorem of Algebra
Remembering complex conjugation 𝑧 → 𝑧̄ on ℂ, we can rewrite this as:
𝑧
𝜙−1
𝑁 ◦𝜙𝑆 (𝑧) = = 1∕𝑧̄ for all 𝑧 ∈ ℂ ⧵ {0}.
|𝑧|2

Similarly, we get
𝑧
𝜙−1
𝑆 ◦𝜙𝑁 (𝑧) = = 1∕𝑧̄ for all 𝑧 ∈ ℂ ⧵ {0}.
|𝑧|2

Thus, for 𝑧 ≠ 0, we get

𝑄(𝑧) = 𝜙−1 −1
𝑆 ◦𝜙𝑁 ◦𝑃 ◦𝜙𝑁 ◦𝜙𝑆 (𝑧)
= 𝜙−1
𝑆 ◦𝜙𝑁 (𝑃 (1∕𝑧))
̄
= 𝜙−1
𝑆 ◦𝜙𝑁 (𝑎𝑛 𝑧̄
−𝑛
+ 𝑎𝑛−1 𝑧̄ −(𝑛−1) + ⋯ + 𝑎1 𝑧̄ −1 + 𝑎0 )
( −𝑛 )
= 1∕ 𝑎̄𝑛 𝑧 + 𝑎̄𝑛−1 𝑧−(𝑛−1) + ⋯ + 𝑎̄1 𝑧−1 + 𝑎̄0
( )
= 𝑧𝑛 ∕ 𝑎̄𝑛 + 𝑎̄𝑛−1 𝑧 + ⋯ + 𝑎̄1 𝑧𝑛−1 + 𝑎̄0 𝑧𝑛 .

Note that 𝑧 = 0 is not a zero of 𝐷(𝑧) ∶= 𝑎̄𝑛 + 𝑎̄𝑛−1 𝑧 + ⋯ + 𝑎̄1 𝑧𝑛−1 + 𝑎̄0 𝑧𝑛 , since 𝑎̄𝑛 ≠ 0. Thus
in small a neighborhood of 𝑧 = 0 in ℂ, the denominator in the expression for 𝑄(𝑧) is bounded
below, i.e., there is a small 𝜀 > 0 and a 𝑐 > 0 such that |𝐷(𝑧)| ≥ 𝑐 for all 𝑧 ∈ 𝐵𝜀2 (0) ⊂ ℂ. Since
𝑄(𝑧) = 𝑧𝑛 ∕𝐷(𝑧), this shows that 𝑄 is smooth at 𝑧 = 0.

We can also calculate the derivative of 𝑄 at 0:

𝑄(0) = 𝜙−1 −1 −1
𝑆 (𝑓 (𝜙𝑆 (0)) = 𝜙𝑆 (𝑓 (𝑁)) = 𝜙𝑆 (𝑁) = 0

and hence
𝑄(ℎ) − 𝑄(0)
lim
(ℎ
ℎ→0
)
ℎ𝑛 ∕ 𝑎̄𝑛 + 𝑎̄𝑛−1 ℎ + ⋯ + 𝑎̄0 ℎ𝑛 − 0
= lim
ℎ→0 ℎ
ℎ 𝑛−1
= lim
ℎ→0 𝑎̄ 𝑛 + ℎ(𝑎̄ 𝑛−1 + ⋯ + 𝑎̄ 0 ℎ𝑛−1 )
{
1∕𝑎̄𝑛 if 𝑛 = 1
=
0 if 𝑛 ≥ 2.

Since smoothness is a local property, 𝑄 is smooth in a small open neighborhood of 0. Since

𝜙𝑆 and 𝜙−1
𝑆
are diffeomorphisms and since 𝜙𝑆 sends an open neighborhood of 𝑁 in 𝕊2 to an
open neighborhood of 0 in ℂ, this implies that

𝑓 = 𝜙−1
𝑆 ◦𝑄◦𝜙𝑆

is smooth in an open neighborhood of 𝑁.

∙ Now we use the fact that 𝑃 is a complex polynomial. This will give us a simple way to
calculate its derivative. For the derivative of the real function 𝑃 ∶ ℝ2 → ℝ2 is the same
Chapter 4. Submersions and regular values 91
as the underlying derivative of 𝑃 ∶ ℂ → ℂ as a complex function.2 More importantly,
the derivative of a polynomial has at most finitely many zeros.

Lemma 4.19 (Second claim: Finitely many critical points) The map 𝑓 ∶ 𝕊2 → 𝕊2
has only finitely many critical points, i.e., points 𝑥 where 𝑑𝑓𝑥 fails to be surjective.

Proof: Since 𝜙𝑁 and 𝜙−1 𝑁

are diffeomorphisms, the only points that might be critical for
𝑓 are the points where 𝑃 fails to be a local diffeomorphism, and possibly 𝑁. But the derivative
of 𝑃 , as a linear map
𝑑𝑃𝑧 ∶ ℂ → ℂ, 𝑤 → 𝑑𝑃𝑧 ⋅ 𝑤,
is given by
∑
𝑛
′
𝑑𝑃𝑧 = 𝑃 (𝑧) = 𝑗𝑎𝑗 𝑧𝑗−1
𝑗=1

Hence 𝑑𝑃𝑧 fails to be an isomorphism, only if it is the zero map, i.e., 𝑑𝑃𝑧 = 0. However, as a
function of 𝑧, 𝑑𝑃𝑧 is a polynomial of degree 𝑛 − 1 and therefore has at most 𝑛 − 1 zeros. Hence
there are only finitely many 𝑧 where 𝑑𝑃𝑧 is not an isomorphism.

∙ Thus the set 𝑅 of regular values for 𝑓 is 𝕊2 with finitely many points removed and is
therefore connected. Since 𝕊2 is compact, we can apply the Stack of Records Theo-
rem 4.14 and see that the function

𝑅 = 𝕊2 − {critical values} → ℤ, 𝑦 → #𝑓 −1 (𝑦),

is locally constant. Since it is defined on a connected space, it must be constant.

∙ (Third claim: Infinitely many values) The map 𝑓 has infinitely many different
values.

Proof: Assume 𝑓 had only finitely many different values 𝑦1 , … , 𝑦𝑘 ∈ 𝕊2 . Then 𝕊2 =

1 ) ∪ ⋯ ∪ 𝑓 (𝑦𝑘 ) with 𝑓 (𝑦𝑖 ) ∩ 𝑓 (𝑦𝑗 ) = ∅ and each 𝑓 (𝑦𝑖 ) would be nonempty and
𝑓 −1 (𝑦 −1 −1 −1 −1

open and closed, since 𝑓 is continuous. This is impossible, since 𝕊2 is connected, unless
𝑘 = 1, This would imply that 𝑓 was constant. However, 𝑃 is not constant, and 𝜙𝑁 and 𝜙−1
𝑁
are
diffeomorphisms. Thus 𝑓 is not constant, and our initial assumption had to be wrong.

This enables us to show:

Lemma 4.20 (Fourth claim: Surjectivity) The smooth map 𝑓 is onto.

Proof: Assume there is a 𝑦0 ∈ 𝕊2 with 𝑓 −1 (𝑦0 ) = ∅, i.e., #𝑓 −1 (𝑦0 ) = 0. Then 𝑦0 is a

regular value for 𝑓 by definition. Since the function 𝑦 →
 #𝑓 −1 (𝑦) is constant on the set of
2
We also have the additional information that the derivative is even ℂ-linear, not just ℝ-linear.
92 4.4. Milnor’s proof of the Fundamental Theorem of Algebra
regular values, it would have to be zero for every regular value. Hence #𝑓 −1 (𝑦) ≠ 0 only for
critical values 𝑦. But that would mean that 𝑓 had only finitely many values, as 𝑓 has only
finitely many critical points. This contradicts the previous claim. Thus 𝑓 must be onto.

∙ Conclusion: In particular, 𝑓 −1 (𝑆) ≠ ∅ for the south pole 𝑆 on 𝕊2 . Hence there must be
at least one point 𝑥 ∈ 𝕊2 with 𝑓 (𝑥) = 𝑆. Since 𝜙𝑁 is a diffeomorphism and 𝜙𝑁 (0) = 𝑆,
𝑥 must satisfy 𝑃 (𝜙−1
𝑁
(𝑥)) = 0. Hence 𝑧 ∶= 𝜙−1𝑁
(𝑥) ∈ ℂ is a zero of 𝑃 .
Chapter 4. Submersions and regular values 93
4.5 Exercises and more examples

4.5.1 Submersions and regular values

Exercise 4.1 Let 𝑓 ∶ 𝑋 → 𝑌 be a submersion and 𝑈 an open subset of 𝑋. Show that

𝑓 (𝑈 ) is open in 𝑌 . In other words, submersions are open maps.

Exercise 4.2 We define the map 𝑔 by

𝑔 ∶ ℝ2 → ℝ, (𝑥, 𝑦) → 𝑥2 − 𝑦2 .

Determine the set of regular values of 𝑔, and determine the set of critical values of 𝑔. Is
𝑔 a submersion?

Exercise 4.3 Recall that a space 𝑌 is called connected if 𝑌 cannot be written as the
union of two nonempty disjoint open subsets; or equivalently, if 𝑌 and ∅ are the only
subsets which are both open and closed in 𝑌 .

(a) Show that if 𝑋 is compact and 𝑌 is connected, then every nontrivial submersion
𝑓 ∶ 𝑋 → 𝑌 is surjective.

(b) Show that there exist no submersions from compact manifolds to ℝ𝑛 for any 𝑛.

Exercise 4.4 Show that the orthogonal group 𝑂(𝑛) is compact.

∑
Hint: Show that if 𝐴 = (𝑎𝑖𝑗 ) lies in 𝑂(𝑛), then for each 𝑖, 𝑗 𝑎2𝑖𝑗 = 1.

Exercise 4.5 Show that the tangent space to 𝑂(𝑛) at the identity matrix 𝐼 is the vector
space of skew symmetric 𝑛 × 𝑛-matrices, i.e., matrices 𝐵 satisfying 𝐵 𝑇 = −𝐵.

Exercise 4.6 Let 𝑀(2) denote the set of all real 2 × 2-matrices.

(a) Show that the determinant function is a submersion on the open submanifold of
nonzero 2 × 2-matrices 𝑀(2) ⧵ {0}.

(b) Conclude that the set 𝑅1 of all 2 × 2-matrices of rank 1 is a three-dimensional

submanifold of ℝ4 = 𝑀(2).

Exercise 4.7 In this exercise we prove Euler’s identity for homogeneous polynomials:
Let 𝑃 (𝑥1 , … , 𝑥𝑘 ) be a homogeneous polynomial of degree 𝑚 in 𝑘 variables, i.e.,

𝑃 (𝑡𝑥1 , … , 𝑡𝑥𝑘 ) = 𝑡𝑘 𝑃 (𝑥1 , … , 𝑥𝑘 ) for all 𝑡.

94 4.5. Exercises and more examples
Show Euler’s identity
∑
𝑥𝑖 𝜕𝑃 ∕𝜕𝑥𝑖 = 𝑚𝑃 . (4.2)
𝑖

Hint: Define a new function 𝑄 by

𝑄(𝑥1 , … , 𝑥𝑘 , 𝑡) ∶= 𝑃 (𝑡𝑥1 , … , 𝑡𝑥𝑘 ) − 𝑡𝑚 𝑃 (𝑥1 , … , 𝑥𝑘 )

and compute its derivative.

Exercise 4.8 In this exercise we show that the fibers of homogeneous polynomials form
manifolds.
Let 𝑃 (𝑥1 , … , 𝑥𝑘 ) be a homogeneous polynomial of degree 𝑚 in 𝑘 variables We con-
sider 𝑃 as a map
ℝ𝑘 → ℝ, (𝑥1 , … , 𝑥𝑘 ) → 𝑃 (𝑥1 , … , 𝑥𝑘 ).

(a) Show that 0 is the only critical value of 𝑃 . Conclude that 𝑃 −1 (𝑎) is a 𝑘 − 1-
dimensional submanifold of ℝ𝑘 for all 𝑎 ≠ 0.
Hint: You may want to use Euler’s identity for homogeneous polynomials.

(b) For two positive real numbers 𝑎, 𝑏 > 0, show that 𝑃 −1 (𝑎) is diffeomorphic to
𝑃 −1 (𝑏). Similarly, For two negative real numbers 𝑎, 𝑏 < 0, show that 𝑃 −1 (𝑎) is
diffeomorphic to 𝑃 −1 (𝑏).

Exercise 4.9 The set 𝑆𝐿(𝑛) of 𝑛 × 𝑛- matrices with determinant +1 form a subgroup
of 𝐺𝐿(𝑛). In this exercise we show that 𝑆𝐿(𝑛) is a smooth manifold of dimension 𝑛2 − 1
and compute its tangent space.

(a) Show that 0 is the only critical value of det ∶ 𝑀(𝑛) → ℝ.

Hint: You may either want to think of det as a homogeneous polynomial given by
Leibniz’ formula
∑ ∏
𝑛
det(𝐴) = (sgn(𝜎) 𝑎𝑖𝜎(𝑖) ) (4.3)
𝜎 𝑖=1

where the sum runs over all permutations of the set {1, … , 𝑛} and sgn(𝜎) denotes
the sign of the permutation 𝜎. Then use a previous exercise.
∑
Or you use the formula det 𝐴 = 𝑛𝑖=1 (−1)𝑖+𝑗 𝑎𝑖𝑗 det 𝐴𝑖𝑗 where 𝐴𝑖𝑗 denotes the
(𝑛 − 1) × (𝑛 − 1)-matrix defined by removing the 𝑖th row and 𝑗th column from 𝐴.

(b) Conclude that 𝑆𝐿(𝑛) is a submanifold of 𝑀(𝑛) of dimension 𝑛2 − 1.

(c) Show that the tangent space to 𝑆𝐿(𝑛) at the identity matrix consists of all matrices
with trace equal to zero.
Hint: Recall that we proved: If 𝑍 = 𝑓 −1 (𝑦) ⊆ 𝑋 is a submanifold defined by a
regular value 𝑦 of a smooth map 𝑓 ∶ 𝑋 → 𝑌 , then 𝑇𝑥 (𝑍) = Ker (𝑑𝑓𝑥 ) ⊆ 𝑇𝑥 (𝑋).
You may also want to use Leibniz’ formula for det.
Chapter 4. Submersions and regular values 95

Exercise 4.10 Recall the Hopf fibration 𝜋 that we have seen previously: We consider
𝕊3 as a subset of ℂ2 , i.e., 𝕊3 = {(𝑧0 , 𝑧1 ) ∈ ℂ2 ∶ |𝑧0 |2 + |𝑧1 |2 = 1}, and 𝕊2 as a subset
of ℂ × ℝ, i.e., 𝕊2 = {(𝑧, 𝑥) ∈ ℂ × ℝ ∶ |𝑧|2 + 𝑥2 = 1}. Then the Hopf fibration 𝜋 is the
map 𝕊3 → 𝕊2 given by
( )
𝜋(𝑧0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2 .

(a) First we consider 𝜋 as a map 𝜋̃ ∶ ℝ4 ≅ ℂ2 → ℂ × ℝ ≅ ℝ3 using the same formula

as for 𝜋, i.e., 𝜋 = 𝜋̃|𝕊3 . Compute the derivative of 𝜋̃ at a point 𝑞 = (𝑧0 , 𝑧1 ).

(b) Let 𝑞 ∈ 𝕊3 . Explain why the restriction of 𝑑 𝜋̃𝑞 to 𝑇𝑞 𝕊3 has image contained in
𝑇𝜋(𝑞) 𝕊2 .

(c) Consider the points 𝑎 = (0, 0, 1) and 𝑏 = (0, 1, 0) on 𝕊2 ⊂ ℝ3 ≅ ℂ × ℝ. Determine

the fibers 𝜋 −1 (𝑎) and 𝜋 −1 (𝑏) and show that 𝑎 and 𝑏 are regular values for 𝜋.

(d) Now show that actually each point in 𝕊2 is a regular value for 𝜋.
5. A brief excursion to Lie groups

Very important examples of smooth manifolds are given by Lie groups. We have seen first
examples in a previous section and will now discuss them further taking the group structure
into account. The study of Lie groups and their associated Lie algebras is a fascinating subject
in mathematics and we highly recommend to read more about them in other books. Here we
will only give a brief introduction considering Lie groups as examples of smooth manifolds.

5.1 Lie groups - the definition

Definition 5.1 (Lie groups) A Lie group is a group 𝐺 which is also a smooth manifold
such that the two maps

𝜇 ∶ 𝐺 × 𝐺 → 𝐺, (𝑔, ℎ) → 𝑔 ⋅ ℎ =∶ 𝑔ℎ

and

𝜄 ∶ 𝐺 → 𝐺, 𝑔 → 𝑔 −1

corresponding to the two group operations of multiplication and taking inverses, respec-
tively, are both smooth.
In fact, we can summarize the condition that 𝜇 and 𝜄 are smooth by requiring that

𝐺 × 𝐺 → 𝐺, (𝑔, ℎ) → 𝑔ℎ−1

is smooth.

Translations and tangent spaces

∙ If 𝐺 is a Lie group, then any element 𝑔 ∈ 𝐺 defines maps

𝐿𝑔 and 𝑅𝑔 ∶ 𝐺 → 𝐺,

called left translation and right translation, respectively, by

𝐿𝑔 (ℎ) = 𝑔ℎ and 𝑅𝑔 (ℎ) = ℎ𝑔.

Since 𝐿𝑔 can be expressed as the composition of smooth maps

𝑖𝑔 𝜇
← 𝐺 × 𝐺 ←←→
𝐺 ←←←→ ← 𝐺,

96
Chapter 5. A brief excursion to Lie groups 97
with 𝑖𝑔 (ℎ) = (𝑔, ℎ), it follows that 𝐿𝑔 is smooth. It is actually a diffeomorphism of 𝐺,
because 𝐿𝑔−1 is a smooth inverse for it. Similarly, 𝑅𝑔 ∶ 𝐺 → 𝐺 is a diffeomorphism.

∙ In fact, many of the important properties of Lie groups follow from the fact that we can
systematically map any point to any other via a canonical global diffeomorphism given by
translation by a suitable element in 𝐺. This translation makes the study of Lie groups
much more accessible compared to arbitrary smooth manifolds. In particular, we can
move an open neighborhood around any point in 𝐺 to make it an open neighborhood of
the identity element. Hence, in a Lie group, we basically only need to study neighbor-
hoods of the identity element.

∙ This observation has important consequence for the tangent spaces of Lie groups. In
fact, the translation property of Lie groups implies that the tangent space to a Lie group
𝐺 at any matrix in 𝐺 is isomorphic to tangent space to 𝐺 at the identity element. It is
a vector space with an additional structure, a Lie bracket, and is an example of a Lie
algebra. The classification of Lie algebras and thereby Lie groups is a highlight in the
history of mathematics.

Here are some simple first examples of Lie groups:

∙ The real numbers ℝ and Euclidean space ℝ𝑛 are Lie groups under addition, because the
coordinates of 𝑥 − 𝑦 are linear and therefore smooth functions of (𝑥, 𝑦).

∙ Similarly, ℂ and ℂ𝑛 are Lie groups under addition.

∙ Any finite group with the discrete topology is a (compact) Lie group.

∙ Suppose 𝐺 is a Lie group and 𝐻 ⊆ 𝐺 is an open subgroup, i.e., a subgroup which is also
an open subspace. Then 𝐻 is a Lie group as well.

∙ The set ℝ∗ = ℝ ⧵ {0} of nonzero real numbers is a 1-dimensional Lie group under multi-
plication. The subset ℝ>0 of positive real numbers is an open subgroup, and is thus itself
a 1-dimensional Lie group — still with multiplication as the group operation.

∙ The set ℂ∗ of nonzero complex numbers is a 2-dimensional Lie group under complex
multiplication.

∙ The unit circle 𝕊1 ⊂ ℂ∗ is a Lie group under the operations induced by multiplication of
complex numbers.
98 5.1. Lie groups - the definition
∙ A finite product of 𝑘 copies of 𝕊1 is a Lie group. We denote it by 𝕋 𝑘 . In particular, the
2-dimensional torus 𝕋 2 = 𝕊1 × 𝕊1 is a Lie group.

∙ More generally, the product of Lie groups is again a Lie group.

We will see more examples below. But before, we introduce the notion of maps between
Lie groups which respect the Lie group structure.

Definition 5.2 (Lie group homomorphisms) If 𝐺 and 𝐻 are Lie groups, a Lie group
homomorphism from 𝐺 to 𝐻 is a smooth map 𝐹 ∶ 𝐺 → 𝐻 that is also a group homo-
morphism. It is called a Lie group isomorphism if it is also a diffeomorphism, which
implies that it has an inverse that is also a Lie group homomorphism. In this case, we
say that 𝐺 and 𝐻 are isomorphic Lie groups.

Here are some examples of Lie group homomorphisms:

∙ The inclusion map 𝕊1 → ℂ is a Lie group homomorphism.

∙ Considering ℝ as a Lie group under addition, and ℝ∗ as a Lie group under multiplication,
the map
exp ∶ ℝ → ℝ∗ , 𝑡 → 𝑒𝑡

is smooth, and is a Lie group homomorphism, since 𝑒𝑠+𝑡 = 𝑒𝑠 𝑒𝑡 . The image of exp is the
open subgroup ℝ>0 consisting of positive real numbers. In fact, exp ∶ ℝ → ℝ>0 is a Lie
group isomorphism with inverse log ∶ ℝ>0 → ℝ.

∙ Similarly, exp ∶ ℂ → ℂ∗ given by exp(𝑧) = 𝑒𝑧 is a Lie group homomorphism. It is sur-

jective but not injective, because its kernel consists of the complex numbers of the form
2𝜋𝑖𝑘, where 𝑘 is an integer.

∙ The map
𝜖 ∶ ℝ → 𝕊1 , 𝑡 → 𝑒2𝜋𝑖𝑡

is a Lie group homomorphism whose kernel is the set ℤ of integers.

∙ Similarly, the map

𝜖 𝑛 ∶ ℝ𝑛 → 𝕋 𝑛 , (𝑡1 , … , 𝑡𝑛 ) → (𝑒2𝜋𝑖𝑡1 , … , 𝑒2𝜋𝑖𝑡𝑛 )

is a Lie group homomorphism whose kernel is ℤ𝑛 .

Chapter 5. A brief excursion to Lie groups 99
∙ If 𝐺 is a Lie group and 𝑔 ∈ 𝐺, conjugation by 𝑔 is the map 𝐶𝑔 ∶ 𝐺 → 𝐺 given by
𝐶𝑔 (ℎ) = 𝑔ℎ𝑔 −1 . Because group multiplication and inversion are smooth, 𝐶𝑔 is smooth
and it is a group homomorphism:

𝐶𝑔 (ℎℎ′ ) = 𝑔ℎℎ′ 𝑔 −1 = (𝑔ℎ𝑔 −1 )(𝑔ℎ′ 𝑔 −1 ) = 𝐶𝑔 (ℎ)𝐶𝑔 (ℎ′ ).

In fact, it is a Lie group isomorphism, because it has 𝐶𝑔−1 as an inverse. A subgroup

𝐻 ⊆ 𝐺 is said to be normal if 𝐶𝑔 (𝐻) = 𝐻 for every 𝑔 ∈ 𝐺.

Lie group homomorphisms behave much nicer in many respects than arbitrary smooth maps
between manifolds. For example, the rank of the derivative is constant:

Theorem 5.3 (Constant Rank Theorem) Let 𝑓 ∶ 𝐺 → 𝐻 be a Lie group homomor-

phism. Then, as a linear map, the derivative 𝑑𝑓𝑔 has the same rank for all 𝑔 ∈ 𝐺.

Proof: Let 𝑒𝐺 and 𝑒𝐻 denote the identity elements in 𝐺 and 𝐻, respectively. Suppose 𝑔0
is an arbitrary element of 𝐺. We will show that 𝑑𝑓𝑔0 has the same rank as 𝑑𝑓𝑒 . The fact that 𝑓
is a homomorphism means that for all 𝑔 ∈ 𝐺,

𝑓 (𝐿𝑔0 (𝑔)) = 𝑓 (𝑔0 𝑔) = 𝑓 (𝑔0 )𝑓 (𝑔) = 𝐿𝑓 (𝑔0 ) (𝑓 (𝑔));

or in other words, 𝑓 ◦𝐿𝑔0 = 𝐿𝑓 (𝑔0 ) ◦𝑓 . Taking differentials of both sides at the identity and
using the chain rule yields

𝑑𝑓𝑔0 ◦𝑑(𝐿𝑔0 )𝑒𝐺 = 𝑑(𝐿𝑓 (𝑔0 ) )𝑒𝐻 ◦𝑑𝑓𝑒𝐺 .

Recall that left multiplication by any element of a Lie group is a diffeomorphism, so both
𝑑(𝐿𝑔0 )𝑒𝐺 and 𝑑(𝐿𝑓 (𝑔0 ) )𝑒𝐻 are isomorphisms. Because composing with an isomorphism does
not change the rank of a linear map, it follows that 𝑑𝑓𝑔0 and 𝑑𝑓𝑒𝐺 have the same rank.

Remark 5.4 (Lie group isomorphisms revisited) Using the constant rank theorem,
one can now show that every bijective Lie group homomorphism 𝑓 ∶ 𝐺 → 𝐻 is auto-
matically a Lie group isomorphism. This is yet another point which makes Lie groups
special among all smooth manifolds. Here is a hint why this could be true: We will learn
soon about Sard’s theorem which will tell us that the subspace of regular values of a
smooth map is dense in the codomain. In particular, there is at least one regular value,
say ℎ0 ∈H, for our Lie group homomorphism 𝑓 . Since 𝑓 is bijective, ℎ0 must be in the
image of 𝑓 . Hence, at the unique point 𝑔0 ∈ 𝐺 with 𝑓 (𝑔0 ) = ℎ0 , we know that 𝑑𝑓𝑔0 is
surjective. But then 𝑑𝑓𝑔 must be an isomorphism, since otherwise the Local Submer-
sion Theorem would imply that 𝑓 looked like the canonical submersion and would have
nontrivial kernel. Hence 𝑓 would not be bijective. According to the previous theorem,
this implies that 𝑑𝑓𝑔 is an isomorphism for all 𝑔 ∈ 𝐺. Hence 𝑓 is a bijective local dif-
feomorphism everywhere. Bijective local diffeomorphisms are global diffeomorphisms.
Since the map is a Lie group homomorphism, it is a Lie group isomorphism.
100 5.2. Examples of Lie groups
5.2 Examples of Lie groups

Now let us study some more interesting examples:

∙ The General Linear Group

The general linear group

𝐺𝐿(𝑛) = {𝐴 ∈ 𝑀(𝑛) ∶ det 𝐴 ≠ 0}

of all invertible 𝑛 × 𝑛-matrices with entries in ℝ, is a smooth manifold of dimension 𝑛2 , since it

2
is an open subset of 𝑀(𝑛) ≅ ℝ𝑛 . To check that it is open, look at its complement

𝑀(𝑛) ⧵ 𝐺𝐿(𝑛) = {𝐴 ∈ 𝑀(𝑛) ∶ det 𝐴 = 0} = det −1 (0).

Since det ∶ 𝑀(𝑛) → ℝ is continuous, being a polynomial in the entries of the matrix, and since
{0} is a closed subset of ℝ, det −1 (0) is closed in 𝑀(𝑛).

We claim that 𝐺𝐿(𝑛) is a Lie group: To show this we need to check that multiplication
and taking inverses are smooth operations. Given two matrices 𝐴 and 𝐵 in 𝐺𝐿(𝑛), the entry in
position (𝑖, 𝑗) in 𝐴𝐵 is given by

∑
𝑛
(𝐴𝐵)𝑖𝑗 = 𝑎𝑖𝑘 𝑏𝑘𝑗 .
𝑘=1

Hence (𝐴𝐵)𝑖𝑗 is a polynomial in the coordinates of 𝐴 and 𝐵. Thus matrix multiplication

𝜇 ∶ 𝐺𝐿(𝑛) × 𝐺𝐿(𝑛) → 𝐺𝐿(𝑛)

is a smooth map.

Recall that the (𝑖, 𝑗)-minor of a matrix 𝐴 is the determinant of the submatrix 𝐴𝑖𝑗 of 𝐴 ob-
tained by deleting the 𝑖th row and the 𝑗th column of 𝐴. By Cramer’s rule from linear algebra,
the (𝑖, 𝑗)-entry of 𝐴−1 is
1
(𝐴−1 )𝑖𝑗 = ⋅ (−1)𝑖+𝑗 ((𝑗, 𝑖)-minor of 𝐴),
det 𝐴
which is a smooth function of the 𝑎𝑖𝑗 ’s provided det 𝐴 ≠ 0, i.e., the map

𝑀(𝑛) → ℝ, 𝐴 → (𝐴−1 )𝑖𝑗

is smooth because it depends smoothly on the entries of 𝐴. Therefore, the map of taking inverses

𝜄 ∶ 𝐺𝐿(𝑛) → 𝐺𝐿(𝑛)

is also smooth.

Remark 5.5 (𝐺𝐿(𝑛) exists over many bases) In fact, we can matrices with entries
in any ring 𝐾. We denote the corresponding matrix groups by 𝑀(𝑛, 𝐾), 𝐺𝐿(𝑛, 𝐾), ….
Chapter 5. A brief excursion to Lie groups 101
Since 𝐾 = ℝ is the most important case for us, we omit mentioning the base when it is
clear that we work over ℝ.
Another very important case is 𝐾 = ℂ. The complex general linear group 𝐺𝐿(𝑛, ℂ)
is also a Lie group. It is a group under matrix multiplication, and it is an open subman-
ifold of 𝑀(𝑛, ℂ) and thus a 2𝑛2 -dimensional smooth manifold. It is a Lie group, since
matrix products and inverses are smooth functions of the real and imaginary parts of the
matrix entries.
Note that the determinant is a Lie group homomorphism for both ℝ and ℂ:

det ∶ 𝐺𝐿(𝑛, ℝ) → ℝ∗ and det ∶ 𝐺𝐿(𝑛, ℂ) → ℂ∗ .

∙ The Special Linear Group: 𝑆𝐿(𝑛) = {𝐴 ∈ 𝑀(𝑛) ∶ det 𝐴 = 1}

In terms of geometry, note that 𝑆𝐿(𝑛) consists of all transformations of ℝ𝑛 into itself which
preserve volumes and orientations.We have shown in the exercises that 𝑆𝐿(𝑛) = det −1 (1) is
a smooth manifold of dimension 𝑛2 − 1. Since it is a subset of the Lie group 𝐺𝐿(𝑛) with the
operation inherited from the one of 𝐺𝐿(𝑛), 𝑆𝐿(𝑛) is also a Lie group. In the exercises we cal-
culate the tangent space of 𝑆𝐿(𝑛) at the identity to be the subspace in 𝑀(𝑛) of all matrices with
trace zero.

∙ The Special Orthogonal Group

Recall that the orthogonal group 𝑂(𝑛) is defined as the subset of matrices 𝐴 in 𝑀(𝑛) such
𝐴𝐴𝑇 = 𝐼. This equation implies, in particular, that every 𝐴 ∈ 𝑂(𝑛) is invertible with 𝐴−1 =
𝐴𝑇 . Hence the determinant of an 𝐴 ∈ 𝑂(𝑛) must satisfy (det 𝐴)2 = 1, i.e., det 𝐴 = ±1. Thus,
𝑂(𝑛) splits into two disjoint parts, the subset of matrices with determinant +1 and the subset of
matrices with determinant −1.

If 𝐴 and 𝐵 have determinant −1, then their product 𝐴𝐵 has determinant +1. Hence the
subset of matrices with determinant −1 is not closed under multiplication and therefore not
a subgroup of 𝑂(𝑛). But the other part is a Lie subgroup of 𝑂(𝑛) and is called the Special
Orthogonal Group denoted by 𝑆𝑂(𝑛)
𝑆𝑂(𝑛) = {𝐴 ∈ 𝑂(𝑛) ∶ det 𝐴 = 1} ⊂ 𝑂(𝑛).
The subgroup 𝑆𝑂(𝑛) is a Lie group

∙ Unitary and Special Unitary Groups

The unitary group 𝑈 (𝑛) is defined to be

𝑈 (𝑛) ∶= {𝐴 ∈ 𝐺𝐿(𝑛, ℂ) ∶ 𝐴̄ 𝑇 𝐴 = 𝐼},
where 𝐴̄ denotes the complex conjugate of 𝐴, the matrix obtained from 𝐴 by conjugating every
entry of 𝐴. A similar argument as for 𝑂(𝑛) shows that 𝑈 (𝑛) is a submanifold of 𝐺𝐿(𝑛, ℂ) and
that dim 𝑈 (𝑛) = 𝑛2 .
102 5.3. Topology of Lie groups
The special unitary group 𝑆𝑈 (𝑛) is defined to be the subgroup of 𝑈 (𝑛) of matrices of
determinant 1.

Remark 5.6 (Spin groups) There are other important examples of Lie groups which, in
general, do not arise as closed subgroups of 𝐺𝐿(𝑛, ℝ) or 𝐺𝐿(𝑛, ℂ). For example, the 𝑛th
Spin group Spin(𝑛) is the 𝑛-dimensional Lie group which is a double cover of 𝑆𝑂(𝑛).
The latter means that Spin(𝑛) is equipped with a smooth surjective map 𝜋 ∶ Spin(𝑛) →
𝑆𝑂(𝑛) such that each point in 𝑆𝑂(𝑛) has an open neighborhood 𝑈 such that 𝜋 −1 (𝑈 ) is
a disjoint union of open subsets in Spin(𝑛) each of which is mapped diffeomorphically
onto 𝑈 by 𝜋. The map 𝜋 is part of a short exact sequence of groups

1 → ℤ∕2 → Spin(𝑛) → 𝑆𝑂(𝑛) → 1.

Spin groups can be constructed for example via Clifford algebras. However, there are
some exceptional isomorphisms in low dimensions which we can write down:

Spin(1) ≅ 𝑂(1),
Spin(2) ≅ 𝑆𝑂(2),
Spin(3) ≅ 𝑆𝑈 (2),
Spin(4) ≅ 𝑆𝑈 (2) × 𝑆𝑈 (2),
Spin(6) ≅ 𝑆𝑈 (4).

5.3 Topology of Lie groups

For some important Lie groups, we will now study the topological properties we singled out in
the beginning: compactness, connectedness and path-connectedness.

We showed in the exercises that 𝑂(𝑛) is compact. As a closed subset , 𝑆𝑂(𝑛) is compact
we well. Similarly, 𝑈 (𝑛) and 𝑆𝑈 (𝑛) are compact. The general linear group 𝐺𝐿(𝑛), however, is
not compact as an open subset of 𝑀(𝑛).

Moreover, note that both 𝑆𝑂(𝑛) and its complement are both open and closed in 𝑂(𝑛). They
are the two connected components of 𝑂(𝑛). In particular, there is no continuous path in 𝑂(𝑛)
from a matrix with determinant +1 to one with determinant −1. In fact, there is no such path
in 𝐺𝐿(𝑛):

Theorem 5.7 The real general linear group is not connected Let 𝛾 be a path in
𝐺𝐿(𝑛), i.e. a continuous map

𝛾 ∶ [0, 1] → 𝐺𝐿(𝑛).

Since 𝛾 and det are continuous, so is their composite

𝛾 det
det ◦𝛾 ∶ [0, 1] ←←→
← 𝐺𝐿(𝑛) ←←←←←→
← ℝ.
Chapter 5. A brief excursion to Lie groups 103
Hence if det(𝛾(0)) > 0 and det(𝛾(1)) < 0, then the Intermediate Value Theorem from
Calculus implies that there must be a real number 𝑡0 ∈ (0, 1) such that det(𝛾(𝑡0 )) = 0 ∉
𝐺𝐿(𝑛). Hence 𝛾 would have to leave 𝐺𝐿(𝑛).

Thus also 𝐺𝐿(𝑛) has two connected components, one of which is an open subgroup con-
sisting to all matrices 𝐴 with det 𝐴 > 0. The other one is just an open subset consisting to all
matrices 𝐴 with det 𝐴 < 0.

Theorem 5.8 The complex general linear group is connected However, 𝐺𝐿(𝑛, ℂ)
is path-connected. We see the difference between 𝐺𝐿(𝑛, ℝ) and 𝐺𝐿(𝑛, ℂ) most clearly
for the case 𝑛 = 1: 𝐺𝐿(1, ℝ) = ℝ∗ is not path-connected, since we cannot cross 0;
whereas 𝐺𝐿(1, ℂ) = ℂ∗ is path-connected, since we can just walk around 0 in the plane.
More generally, to show that 𝐺𝐿(𝑛, ℂ) is path-connected, it suffices to show that
there is a path from any matrix 𝐴 ∈ 𝐺𝐿(𝑛, ℂ) to the identity matrix 𝐼 ∈ 𝐺𝐿(𝑛, ℂ).
Therefore, we define first the function

𝑃 ∶ ℂ → ℂ, 𝑧 → det(𝐴 + 𝑧(𝐼 − 𝐴)).

Then we have 𝑃 (0) = det 𝐴 ≠ 0 and 𝑃 (1) = det 𝐼 = 1 ≠ 0. Since 𝑃 is a polynomial

of degree 𝑛, it has only finitely many zeroes. We also note that neither 0 ∈ ℂ nor 1 ∈ ℂ
are zeroes of 𝑃 . Since ℂ ⧵ {finitely many points} is path-connected, we can find a path
𝛾 ∶ [0, 1] → ℂ with 𝛾(0) = 0, 𝛾(1) = 1 and which avoids the zeroes of 𝑃 , i.e.,

𝑃 (𝛾(𝑡)) ≠ 0 for all 𝑡.

Then the continuous map

Γ ∶ [0, 1] → 𝐺𝐿(𝑛, ℂ), 𝑡 → 𝐴 + 𝛾(𝑡)(𝐼 − 𝐴)

is the desired path from 𝐴 to 𝐼.

Figure 5.1: Two points in 𝐺𝐿(𝑛, ℝ) may not be connected. In 𝐺𝐿(𝑛, ℂ), however, we can
always find a path between two matrices. This reduces via the determinant function to the fact
that the plane remains path-connected after removing finitely many points.
104 5.4. Lie subgroups

Remark 5.9 The fact that GL(𝑛, ℂ) is connected while 𝐺𝐿(𝑛, ℝ) is not plays a crucial
role for orientations of vector spaces, vector bundles, manifolds etc. For, every complex
vector space, complex vector bundle, complex manifold, etc has a natural orientation.
We will get back to this later.

Open neighborhoods of the identity:

Recall that if 𝐺 is a group and 𝑆 ⊂ 𝐺 is a subset, the subgroup generated by 𝑆 is the

smallest subgroup containing 𝑆, i.e., the intersection of all subgroups containing 𝑆. One can
check that the subgroup generated by 𝑆 is equal to the set of all elements of 𝐺 that can be
expressed as finite products of elements of 𝑆 and their inverses.

Lemma 5.10 (Neighborhoods of the identity) Suppose 𝐺 is a Lie group, and 𝑊 ⊂ 𝐺

is any neighborhood of the identity. Then

∙ 𝑊 generates an open subgroup of 𝐺.

∙ If 𝐺 is connected, then 𝑊 generates 𝐺. In particular, an open subgroup in a

connected Lie group must be equal to the whole group.

Proof:

∙ Let 𝑊 ⊂ 𝐺 be any neighborhood of the identity, and let 𝐻 be the subgroup generated by
𝑊 . To simplify notation, if 𝐴 and 𝐵 are subsets of 𝐺, we write

𝐴𝐵 ∶= {𝑎𝑏 ∶ 𝑎 ∈ 𝐴, 𝑏 ∈ 𝐵}, and 𝐴−1 ∶= {𝑎−1 ∶ 𝑎 ∈ 𝐴}.

For each positive integer 𝑘, let 𝑊𝑘 denote the set of all elements of 𝐺 that can be expressed
as products of 𝑘 or fewer elements of 𝑊 ∪ 𝑊 −1 . As mentioned above, 𝐻 is the union of
all the sets 𝑊𝑘 as 𝑘 ranges over the positive integers.
Now, 𝑊 −1 is open because it is the image of 𝑊 under the inversion map, which is a
diffeomorphism. Thus, 𝑊1 = 𝑊 ∪ 𝑊 −1 is open, and, for each 𝑘 > 1, we have

𝑊𝑘 = 𝑊1 𝑊𝑘−1 = ∪𝑔∈𝑊1 𝐿𝑔 (𝑊𝑘−1 ).

Because each 𝐿𝑔 is a diffeomorphism, it follows by induction that each 𝑊𝑘 is open, and

thus 𝐻 is open as a union of open subsets.

∙ Assume G is connected. We just showed that 𝐻 is an open subgroup of 𝐺. It is an

exercise to show that an open subgroup in a connected Lie group is equal to the whole
group.

5.4 Lie subgroups

In the previous paragraph we talked about subgroups of a Lie group. But we did not discuss how
the subgroup structure relates to the structure as a smooth manifold. Actually, this is a subtle
Chapter 5. A brief excursion to Lie groups 105
and interesting point that illustrates the importance of the distinction between immersions and
embeddings once again. So here is the definition of a Lie subgroup:

Definition 5.11 (Lie subgroups) A Lie subgroup of a Lie group 𝐺 is an abstract

subgroup 𝐻 ⊆ 𝐺 such that

∙ there exists a smooth manifold 𝑋 and an immersion 𝑓 ∶ 𝑋 → 𝐺 such that 𝐻 =

Im (𝑓 ) ⊆ 𝐺 is the image of 𝑓 , and

∙ the group operations on 𝐻 are smooth, in the sense that the compositions
𝑓 ×𝑓 𝜇
𝑋 × 𝑋 ←←←←←←←←→ ← 𝐺, and
← 𝐺 × 𝐺 ←←→
𝑓 𝜄
𝑋 ←←←→
← 𝐺 ←→
← 𝐺

are smooth.

∙ Note: It is important to note that, in the above definition, we do not require 𝐻 to have
the subspace topology induced by being a subset in 𝐺. Instead we can think of the map 𝑓
to define a topology on 𝐻 in the sense that a subset 𝑈 ⊂ 𝐻 is open if and only if 𝑓 −1 (𝑈 )
is open in 𝑋.

Let us have a closer look at this rather complicated definition:

∙ An abstract subgroup simply means a subgroup in the algebraic sense. The group oper-
ations on the subgroup 𝐻 are the restrictions of the multiplication map 𝜇 and the inverse
map 𝜄 from 𝐺 to 𝐻.

∙ If 𝐻 were defined to be a submanifold of 𝐺, then the multiplication map 𝐻 × 𝐻 → 𝐻

and similarly the inverse map 𝐻 → 𝐻 would automatically be smooth, and the defini-
tion would be much shorter. However, since a Lie subgroup is defined to be an immersed
submanifold, it is necessary to impose the last condition.

∙ If 𝐻 is in fact also a submanifold, then 𝐻 is a Lie subgroup as the following result shows.

Lemma 5.12 (Embedded Lie subgroups) If 𝐻 is an abstract subgroup and a subman-

ifold of a Lie group 𝐺, then it is a Lie subgroup of 𝐺. In this case, the inclusion map
𝐻 → 𝐺 is an embedding, and we call 𝐻 an embedded subgroup.

Proof: Since 𝐻 is a subgroup, multiplication and taking inverses in 𝐻 are just the re-
strictions of multiplication and taking inverses in 𝐺 and both have image in 𝐻. Since 𝐻 is a
submanifold we can take 𝑋 = 𝐻 in the above definition, the restrictions of smooth maps to 𝐻
are again smooth.
106 5.4. Lie subgroups
Here are some examples:

∙ The subgroups 𝑆𝐿(𝑛) and 𝑂(𝑛) of 𝐺𝐿(𝑛) are both submanifolds, and therefore embedded
Lie subgroups.

∙ One easily verifies that

( )
∗ 𝑥 𝑦
ℂ = GL(1, ℂ) → 𝐺𝐿(2, ℝ), 𝑧 = 𝑥 + 𝑖𝑦 →
−𝑦 𝑥
is an embedding.

∙ More generally, this map induces an embedding

𝐺𝐿(𝑛, ℂ) → 𝐺𝐿(2𝑛, ℝ)
) (
𝑥 𝑦
by replacing each entry 𝑧 = 𝑥 + 𝑖𝑦 in 𝐴 ∈ 𝐺𝐿(𝑛, ℂ) by the block :
−𝑦 𝑥

⎛𝑥11 −𝑦11 𝑥 −𝑦1𝑛 ⎞

… 1𝑛
⎛𝑥11 + 𝑖𝑦11 … 𝑥1𝑛 + 𝑖𝑦1𝑛 ⎞ ⎜𝑦11 𝑥11 𝑦1𝑛 𝑥1𝑛 ⎟
⎜ ⋮ ⋱ ⋮ ⎟ → ⎜⎜ ⋮ ⋱ ⋮
⎟
⎟.
⎜ ⎟ ⎜ ⎟
⎝ 𝑛1
𝑥 + 𝑖𝑦 𝑛1 … 𝑥𝑛𝑛 + 𝑖𝑦𝑛𝑛 ⎠ 𝑥𝑛1 −𝑦 𝑛1 𝑥 −𝑦
⎜𝑦 … 𝑛𝑛 𝑛𝑛
⎝ 𝑛1 𝑥𝑛1 𝑦𝑛𝑛 𝑥𝑛𝑛 ⎟⎠

This way, 𝐺𝐿(𝑛, ℂ) is an embedded Lie subgroup of 𝐺𝐿(2𝑛, ℝ).

Now let us get back to understanding the definition of a Lie subgroup. The subtle differ-
ences of immersed and embedded subgroups can be illustrated by a familiar example:

Example 5.13 (An immersed but not embedded Lie subgroup) Recall the maps
𝑔∶ ℝ → 𝕊1 ⊂ ℂ, 𝑡 → 𝑒𝑖𝑡 , and

𝐺 ∶ ℝ2 → 𝕊1 × 𝕊1 = 𝕋 2 , 𝐺(𝑠, 𝑡) = (𝑔(𝑠), 𝑔(𝑡))

The map 𝐺 is a local diffeomorphism from the plane onto the torus 𝕋 2 .
Given a real number 𝛼, we defined the map 𝛾𝛼 by

𝛾𝛼 ∶ ℝ → 𝕋 2 , 𝛾(𝑡) = (𝑔(𝑡), 𝑔(𝛼 ⋅ 𝑡)).

We learned that 𝛾𝛼 is always an immersion, but its image is not a submanifold of 𝕋 2

if𝛼 is an irrational number. However, when 𝛼 is rational, then 𝛾𝛼 (ℝ) is a submanifold of
𝑇 2 . Recall Figure 3.4.
After checking that 𝛾𝛼 (ℝ) is an abstract subgroup, we see that 𝛾𝛼 (ℝ) is in fact a Lie
subgroup of 𝕋 2 for every real number 𝛼.

For an explanation of why a Lie subgroup is defined in such a complicated way, we refer to
a fact we will only be able to appreciate when we learn more about Lie theory:
Chapter 5. A brief excursion to Lie groups 107

Remark 5.14 (Why so complicated?) A fundamental theorem in Lie group theory

asserts the existence of a one-to-one correspondence between the connected Lie sub-
groups of a Lie group 𝐺 and the Lie subalgebras of its Lie algebra 𝔤, i.e., tangent space
at the identity with its Lie bracket:
1−1
{connected Lie subgroups in 𝐺} ←→ {Lie subalgebras in 𝔤}.

In the previous example, the Lie algebra of 𝕋 2 has ℝ2 as the underlying vector space,
and the one-dimensional Lie subalgebras are all the lines through the origin with addition
as group operation. Such a line is determined by its slope 𝛼. Hence every 𝛼 should
correspond to a Lie subgroup 𝛾𝛼 (ℝ) in 𝕋 2 .
However, if a Lie subgroup had been defined as a subgroup that is also a submani-
fold, then one would have to exclude all the lines with irrational slopes as Lie subgroups
of the torus. In this case it would not be possible to have a one-to-one correspondence
between the connected subgroups of a Lie group and the Lie subalgebras of its Lie al-
gebra. But this correspondence is extremely useful in Lie theory.

The following theorem is a very useful fact which we state here without proof. See for
example [11, Theorem 7.21].

Theorem 5.15 (Closed Subgroup Theorem) Suppose 𝐺 is a Lie group and 𝐻 ⊆ 𝐺

is a Lie subgroup. Then 𝐻 is closed in 𝐺 if and only if it is an embedded Lie subgroup.

5.4.1 Lie groups

Exercise 5.1 (a) Show that 𝑆𝑂(2) is diffeomorphic to 𝕊1 .

(b) Show that 𝑆𝑈 (2) is diffeomorphic to 𝕊3 .

Exercise 5.2 (a) Show that a local diffeomorphism 𝑓 ∶ 𝑋 → 𝑌 which is bijective is

a diffeomorphism.

(b) Show that a local diffeomorphism 𝑓 ∶ 𝑋 → 𝑌 which is one-to-one is a diffeomor-

phism of 𝑋 onto an open subset of 𝑌 .

(c) Show that a bijective smooth map 𝑓 ∶ 𝑋 → 𝑌 of constant rank is a diffeomor-

phism.
Comment: You can assume that 𝑓 is a submersion to simplify things. If you want
to challenge yourself, you could only assume that 𝑋 is compact. Showing that 𝑓
also is a submersion in general requires the use of Baire’s category theorem.

(d) Show that a bijective Lie group homomorphism is a Lie group isomorphism.
108 5.4. Lie subgroups

Exercise 5.3 Show that an open subgroup 𝐻, i.e., a subgroup which is also an open
subset, of a connected Lie group 𝐺 is equal to 𝐺.

Exercise 5.4 Let 𝐺 be a Lie group and let 𝑒 ∈ 𝐺 be the identity element.

(a) Let 𝜇 ∶ 𝐺 × 𝐺 → 𝐺 denote the multiplication map, and let 𝑔, ℎ ∈ 𝐺. Recall that
we denote by 𝐿𝑔 the left translation in 𝐺 by 𝑔, and by 𝑅ℎ the right translation by ℎ.
Using the identification 𝑇(𝑔,ℎ) (𝐺 × 𝐺) = 𝑇𝑔 (𝐺) × 𝑇ℎ (𝐺), show that the differential
of 𝜇 at (𝑔, ℎ)
𝑑𝜇(𝑔,ℎ) ∶ 𝑇𝑔 (𝐺) × 𝑇ℎ (𝐺) → 𝑇𝑔ℎ (𝐺)
is given by

𝑑𝜇(𝑔,ℎ) (𝑋, 𝑌 ) = 𝑑𝜇(𝑔,ℎ) (𝑋, 0) + 𝑑𝜇(𝑔,ℎ) (0, 𝑌 ) = 𝑑(𝑅ℎ )𝑔 (𝑋) + 𝑑(𝐿𝑔 )ℎ (𝑌 ).

Hint: Calculate 𝑑𝜇(𝑔,ℎ) (𝑋, 0) and 𝑑𝜇(𝑔,ℎ) (0, 𝑌 ) separately.

(b) Let 𝜄 ∶ 𝐺 → 𝐺 denote the inversion map. Show that

𝑑𝜄𝑒 ∶ 𝑇𝑒 (𝐺) → 𝑇𝑒 (𝐺)

is given by 𝑑𝜄𝑒 (𝑋) = −𝑋.

(c) Use the previous point to show that, for any 𝑔 ∈ 𝐺, the derivative of 𝜄 at 𝑔 is given
by

𝑑𝜄𝑔 ∶ 𝑇𝑔 (𝐺) → 𝑇𝑔−1 (𝐺), 𝑌 → −𝑑(𝑅𝑔−1 )𝑒 (𝑑(𝐿𝑔−1 )𝑔 (𝑌 )) for all 𝑌 ∈ 𝑇𝑔 (𝐺).

Exercise 5.5 Show that for any Lie group 𝐺, the multiplication map 𝜇 ∶ 𝐺 × 𝐺 → 𝐺
is a submersion.

Exercise 5.6 Show that the differential of the determinant map det ∶ 𝐺𝐿(𝑛, ℝ) → ℝ at
𝐴 ∈ 𝐺𝐿(𝑛, ℝ) is given by

𝑑(det)𝐴 (𝐵) = (det 𝐴) ⋅ (tr 𝐴−1 𝐵) for all 𝐵 ∈ 𝑀(𝑛).

In particular, 𝑑(det)𝐴 (𝐴𝐵) = (det 𝐴) ⋅ (tr 𝐴𝐵) for all 𝐵 ∈ 𝑀(𝑛).

6. Transversality

6.1 Transversality and preimages

Now we would like to understand what happens when we do not take the preimage of just a
single point, but the preimage of a whole submanifold.

Let 𝑋 and 𝑌 be smooth manifolds and let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. Assume that
𝑍 ⊂ 𝑌 is a submanifold of 𝑌 . We would like to understand:

Question Under which conditions is the subset 𝑓 −1 (𝑍) ⊆ 𝑋 a smooth manifold?

6.1.1 Transverse preimages of submanifolds

∙ Let us denote 𝑛 = dim 𝑋, 𝑚 = dim 𝑌 and 𝑘 = dim 𝑍. Let 𝑥0 ∈ 𝑋 with 𝑓 (𝑥0 ) = 𝑧0 ∈ 𝑍.

The inclusion of subsets 𝑍 → 𝑌 is an immersion. Hence we can apply the Local Im-
mersion Theorem 3.10.

∙ We can choose local parametrizations as follows:

∙ 𝜓 ∶ 𝑈 → 𝑉 with 𝜓(0) = 𝑧0 and 𝑉 ⊂ 𝑌 and 𝑈 ⊂ ℝ𝑚 open subsets,

∙ 𝜙 ∶ 𝑊 → 𝑍 ∩ 𝑉 with 𝜙(0) = 𝑧0 and 𝑊 ⊂ ℝ𝑘 and 𝑍 ∩ 𝑉 ⊂ 𝑍 open subsets, such

that

𝑍 ∩O 𝑉 
inclusion /𝑉
O (6.1)
𝜙 𝜓
canonical /𝑈
𝑊
immersion

commutes.

∙ The inverse map 𝜓 −1 ∶ 𝑉 → 𝑈 is a local coordinate system on the open neighborhood

𝑉 around 𝑧0 ∈ 𝑌 . We write 𝑢𝑖 ∶ 𝑉 → ℝ for the 𝑖th component of 𝜓 −1 , i.e., a point
𝑦 ∈ 𝜓(𝑊 ) has the local coordinates
𝜓 −1 (𝑦) = (𝑢1 (𝑦), … , 𝑢𝑚 (𝑦)) ∈ ℝ𝑚 .

∙ Since the lower horizontal map in (6.1) is the canonical immersion, the points in 𝑍 ∩ 𝑉
are exactly those on which the coordinate functions 𝑢𝑘+1 , … , 𝑢𝑚 vanish. Hence we have
𝑍 ∩ 𝑉 = {𝑦 ∈ 𝑉 ∶ 𝑢𝑘+1 (𝑦) = ⋯ = 𝑢𝑚 (𝑦) = 0}.

109
110 6.1. Transversality and preimages
( )
∙ Let us write 𝑔 ∶ 𝑉 → ℝ𝑚−𝑘 for the map given by 𝑦 → 𝑢𝑘+1 (𝑦), … , 𝑢𝑚 (𝑦) . The above
relation then reads
𝑍 ∩ 𝑉 = 𝑔 −1 (0).

∙ Now we take the preimage of 𝑍 ∩ 𝑉 along 𝑓 and get

𝑓 −1 (𝑍 ∩ 𝑉 ) = 𝑓 −1 (𝑔 −1 (0)) = (𝑔◦𝑓 )−1 (0). (6.2)

Since 𝑓 is continuous, 𝑉 is open in 𝑌 and 𝑓 −1 (𝑍) ⊂ 𝑋 has the subspace topology,

𝑓 −1 (𝑍 ∩ 𝑉 ) = 𝑓 −1 (𝑍) ∩ 𝑓 −1 (𝑉 )

is an open subset containing 𝑥0 .

∙ Now if 𝑓 −1 (𝑍 ∩ 𝑉 ) is a manifold, then 𝑥0 has an open neighborhood in 𝑓 −1 (𝑍) which

is diffeomorphic to an open subset in ℝ𝑚−𝑘 . Hence, if every point 𝑥 ∈ 𝑓 −1 (𝑍) has such
an open neighborhood, then 𝑓 −1 (𝑍) is a smooth manifold.

∙ And we do have a criterion that guarantees that this is the case. For, according to (6.2),
𝑓 −1 (𝑍 ∩ 𝑉 ) is a manifold if 0 is a regular value for the smooth map 𝑔◦𝑓 .

∙ Thus we would like to have that 𝑔◦𝑓 is a submersion at every point

𝑥 ∈ 𝑓 −1 (𝑍 ∩ 𝑉 ) = (𝑔◦𝑓 )−1 (0).

∙ One way to check this is to show that 0 is a regular value of the composite 𝑔◦𝑓 . The
chain rule tells us
𝑑(𝑔◦𝑓 )𝑥 = 𝑑𝑔𝑓 (𝑥) ◦𝑑𝑓𝑥 .

∙ Thus, the map

𝑑(𝑔◦𝑓 )𝑥 ∶ 𝑇𝑥 (𝑋) → ℝ𝑚−𝑘 is surjective

⇐⇒ 𝑑𝑔𝑓 (𝑥) maps the image of 𝑑𝑓𝑥 onto ℝ𝑚−𝑘 .

∙ Since 0 is a regular value for 𝑔 ∶ 𝑉 → ℝ𝑚−𝑘 , we know that

𝑑𝑔𝑓 (𝑥) ∶ 𝑇𝑓 (𝑥) (𝑉 ) = 𝑇𝑓 (𝑥) (𝑌 ) → ℝ𝑚−𝑘 ,

on the whole tangent space to 𝑌 at 𝑓 (𝑥), is a surjective linear map.

∙ By Lemma 4.8, which was a consequence of the Preimage Theorem 4.6, the kernel of
𝑑𝑔𝑓 (𝑥) is the subspace 𝑇𝑧 (𝑍). Thus 𝑑𝑔𝑓 (𝑥) induces an isomorphism
≅
← ℝ𝑚−𝑘 .
𝑑 𝑔̄𝑓 (𝑥) ∶ 𝑇𝑓 (𝑥) (𝑌 )∕𝑇𝑓 (𝑥) (𝑍) ←←←→

∙ This means that (𝑑𝑔𝑓 (𝑥) )|Im (𝑑𝑓𝑥 ) can only be surjective if Im (𝑑𝑓𝑥 ) generates the quo-
tient space 𝑇𝑓 (𝑥) (𝑌 )∕𝑇𝑓 (𝑥) (𝑍). In other words, (𝑑𝑔𝑓 (𝑥) )|Im (𝑑𝑓𝑥 ) can only be surjective if
Im (𝑑𝑓𝑥 ) and 𝑇𝑓 (𝑥) (𝑍) together span all of 𝑇𝑓 (𝑥) (𝑌 ).
Chapter 6. Transversality 111
∙ We conclude that 𝑔◦𝑓 is a submersion at 𝑥 ∈ 𝑓 −1 (𝑍 ∩ 𝑉 ) if and only if

Im (𝑑𝑓𝑥 ) + 𝑇𝑓 (𝑥) (𝑍) = 𝑇𝑓 (𝑥) (𝑌 ).

We give this condition a name:

Definition 6.1 (Transversality) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map and 𝑍 ⊆ 𝑌 a

submanifold. Then 𝑓 is said to be transverse to the submanifold 𝑍, denoted 𝑓 −
⋔ 𝑍,
if
Im (𝐝𝐟𝐱 ) + 𝐓𝐟(𝐱) (𝐙) = 𝐓𝐟(𝐱) (𝐘)
at every point 𝑥 ∈ 𝑓 −1 (𝑍) in the preimage of 𝑍.

The above discussion then provides the proof for the following fundamental result:

Theorem 6.2 (Transverse intersections yield submanifolds) If the smooth map

𝑓 ∶ 𝑋 → 𝑌 is transverse to a submanifold 𝑍 ⊆ 𝑌 , then 𝑓 −1 (𝑍) is a submanifold
of 𝑋. Moreover, the codimension of 𝑓 −1 (𝑍) in 𝑋 equals the codimension of 𝑍 in 𝑌 ,
i.e., we have
dim 𝑓 −1 (𝑍) = dim 𝑋 − (dim 𝑌 − dim 𝑍).

∙ Note that regularity is a special case of transversality: If 𝑍 = {𝑦} consists of a single

point, then 𝑇𝑦 (𝑍) is a trivial vector space consisting just of the zero vector. Hence, in
this case, transversality means Im 𝑑𝑓𝑥 = 𝑇𝑦 (𝑌 ) at all points with 𝑓 (𝑥) = 𝑦, i.e., 𝑦 is a
regular value of 𝑓 .

∙ Note that transversality tells us something about how the images of 𝑓 and 𝑍 meet in
𝑌 . We will give further geometric intuition for transversality soon.

For now, let us look at some simple examples of transversality and non-transversality. We
consider 𝑌 = ℝ2 with the submanifold 𝑍 being the 𝑥-axis. Then

∙ The map 𝑓 ∶ ℝ1 → ℝ2 defined by 𝑓 (𝑡) = (0, 𝑡) is transverse to 𝑍, with 𝑓 −1 (𝑍) =

{(0, 0)}.

∙ The map 𝑓 ∶ ℝ1 → ℝ2 defined by 𝑓 (𝑡) = (𝑡, cos 𝑡 − 1) is not transverse to 𝑍, with

𝑓 −1 (𝑍) = {(0, 0)}.

∙ The map 𝑓 ∶ ℝ1 → ℝ2 defined by 𝑓 (𝑡) = (𝑡, 𝑡2 ), however, is not transverse to 𝑍, with

𝑓 −1 (𝑍) = {(0, 0)}.

∙ The map 𝑓 ∶ ℝ1 → ℝ2 defined by 𝑓 (𝑡) = (𝑡, 𝑡2 − 1) is transverse to 𝑍, with 𝑓 −1 (𝑍) =

{(−1, 0), (1, 0)}.

∙ More generally, let 𝑝(𝑡) = 𝑡𝑛 + 𝑎𝑛−1 𝑡𝑛−1 + ⋯ + 𝑎0 be a polynomial with real coefficients.
We can consider 𝑝(𝑡) as a smooth map ℝ → ℝ. The map 𝑓 ∶ ℝ1 → ℝ2 defined by
112 6.1. Transversality and preimages

Figure 6.1: On the left, 𝑓 is transverse to the 𝑥-axis, since Im (𝑑𝑓0 ) and 𝑇0 𝑍 span all of 𝑇0 𝑌 =
ℝ2 . On the right, however, 𝑓 is not transverse to 𝑍, since both Im (𝑑𝑓0 ) and 𝑇0 𝑍 span the same
one-dimensional subspace.

𝑓 (𝑡) = (𝑡, 𝑝(𝑡)) is transverse to the 𝑥-axis 𝑍 if and only if all zeros of 𝑝 are simple, i.e.,
if and only if we can write 𝑝(𝑡) as a product

𝑝(𝑡) = (𝑡 − 𝑟1 ) ⋅ (𝑡 − 𝑟2 ) ⋅ … ⋅ (𝑡 − 𝑟𝑛 )

with 𝑟𝑖 ≠ 𝑟𝑗 if 𝑖 ≠ 𝑗.
The reason is that the derivative of 𝑝 ∶ ℝ → ℝ at 𝑡0 is given by multiplication by 𝑝′ (𝑡0 ),
i.e.,
𝑑𝑝𝑡0 ∶ ℝ1 → ℝ1 , 𝑠 → 𝑝′ (𝑡0 ) ⋅ 𝑠.
This map is nontrivial and hence surjective if and only if 𝑝′ (𝑡0 ) ≠ 0, i.e., if and only if 𝑡0
is not a zero of 𝑝′ . In other words, 𝑑𝑝′ (𝑡0 ) is trivial only if 𝑡0 is a multiple zero of 𝑝.
Now the tangent space of 𝑍 at a point 𝑓 (𝑡0 ) ∈ ℝ2 = 𝑌 is the 𝑥-axis 𝑇𝑓 (𝑡0 ) 𝑍 = ℝ × {0} ⊂
ℝ2 , and the derivative 𝑑𝑓𝑡0 of 𝑓 at 𝑡0 is the map
( )
2 1
𝑑𝑓𝑡0 ∶ ℝ → ℝ , 𝑠 → 𝑠 ⋅ ′ .
𝑝 (𝑡0 )
( ) ( )
1 1
Now we observe that the vectors and ′ are linearly independent in ℝ2 if and
0 𝑝 (𝑡0 )
only if 𝑝′ (𝑡0 ) ≠ 0. Hence 𝑇𝑓 (𝑡0 ) 𝑍 and Im 𝑑𝑓 (𝑡0 ) span 𝑇𝑓 (𝑡0 ) 𝑌 if and only if 𝑝′ (𝑡0 ) ≠ 0.
For transversality of 𝑓 and 𝑍, we need to check that 𝑇𝑓 (𝑡0 ) 𝑍 + Im 𝑑𝑓 (𝑡0 ) = 𝑇𝑓 (𝑡0 ) 𝑌 for all
𝑡 ∈ 𝑓 −1 (𝑍), i.e., for all zeros of 𝑝.

6.1.2 Tangent space of a transverse preimages

We return to the general situation: 𝑋 and 𝑌 are smooth manifolds, 𝑍 ⊂ 𝑌 is a submanifold,

and 𝑓 ∶ 𝑋 → 𝑌 is a smooth map. We assume now that 𝑓 meets 𝑍 transversally. Hence 𝑓 −1 (𝑍)
is a manifold and we would like to have a formula for the tangent space of 𝑓 −1 (𝑍) at a given
point.
Chapter 6. Transversality 113

Figure 6.2: We begin to see a pattern in the plane: 𝑓 is transverse to the 𝑥-axis unless the 𝑥-
axis is tangent to the graph of 𝑓 . In the latter case, the tangent spaces do not provide enough
information.

So let 𝑥 ∈ 𝑓 −1 (𝑍) and 𝑧 ∶= 𝑓 (𝑥). As above we can choose suitable local parametrizations
and find an open subset 𝑉 ⊂ 𝑌 and a smooth map 𝑔 ∶ 𝑉 → ℝ𝑚−𝑘 such that 𝑍 ∩ 𝑉 = 𝑔 −1 (0).
Since tangent spaces are determined locally, the tangent spaces 𝑇𝑧 (𝑍 ∩ 𝑉 ) and 𝑇𝑧 (𝑍) are equal
as subspaces of 𝑇𝑧 (𝑌 ). Since 0 is a regular value for 𝑔, we know that the tangent space of 𝑍 at
𝑧 is given by

𝑇𝑧 (𝑍) = Ker 𝑑𝑔𝑧 ⊂ 𝑇𝑧 (𝑌 ) (6.3)

where 𝑑𝑔𝑧 ∶ 𝑇𝑧 (𝑌 ) → ℝ𝑚−𝑘 .

Now we use that 0 is also a regular value for the composite 𝑔◦𝑓 . Again, since tangent spaces
are determined locally, we get

𝑇𝑥 (𝑓 −1 (𝑍)) = Ker (𝑑(𝑔◦𝑓 )𝑥 )

= Ker (𝑑𝑔𝑓 (𝑥) ◦𝑑𝑓𝑥 ) (by the Chain Rule)
= (𝑑𝑓𝑥 )−1 (Ker (𝑑𝑔𝑓 (𝑥) ))
= (𝑑𝑓𝑥 )−1 (𝑇𝑓 (𝑥) (𝑍)) (by (6.3) above).

Thus we have proved the following very useful result:

Theorem 6.3 (Tangent space of preimage) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map and

𝑍 be a submanifold in 𝑌 . Assume that 𝑓 is transverse to 𝑍. Then 𝑇𝑥 (𝑓 −1 (𝑍)) is the
preimage of 𝑇𝑓 (𝑥) (𝑍) under the linear map

𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 )

As a formula:

𝑇𝑥 (𝑓 −1 (𝑍)) = (𝑑𝑓𝑥 )−1 (𝑇𝑓 (𝑥) (𝑍)).

6.2 Transverse intersections

114 6.2. Transverse intersections
We will now study the most important case, at least for us, of transversality: the intersection
of two submanifolds 𝑋 and 𝑍 in the same given bigger manifold 𝑌 . This becomes, in fact, a
special case of our previous studies by letting 𝑓 to be the inclusion map 𝑖 ∶ 𝑋 → 𝑌 . For to
say a point 𝑥 ∈ 𝑋 belongs to the intersection 𝑋 ∩ 𝑍 is equivalent to say that 𝑥 belongs to the
preimage 𝑖−1 (𝑍).

We would like the intersection 𝑋 ∩ 𝑍 to be a submanifold in 𝑌 . Transversality is the crucial

condition to ensure this is the case:

The derivative 𝑑𝑖𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑥 (𝑌 ) is the inclusion map of 𝑇𝑥 (𝑋) into 𝑇𝑥 (𝑌 ) as a subspace.
Hence we get 𝑖 −
⋔ 𝑍 if and only if, for every 𝑥 ∈ 𝑋 ∩ 𝑍,

𝑇𝑥 (𝑋) + 𝑇𝑥 (𝑍) = 𝑇𝑥 (𝑌 ).

∙ Notice that this equation is symmetric in 𝑋 and 𝑍. When it holds, we shall say that the
two submanifolds 𝑋 and 𝑍 are transverse, and write 𝑋 −⋔ 𝑍.

∙ Warning: For equation 𝑇𝑥 (𝑋) + 𝑇𝑥 (𝑍) = 𝑇𝑥 (𝑌 ) to be true, it is not sufficient that

dim 𝑇𝑥 (𝑋)+dim 𝑇𝑥 (𝑍) = dim 𝑇𝑥 (𝑌 ). The two subspaces together must span all of 𝑇𝑥 (𝑌 ).
A simple example if the intersection of 𝑍 with itself, i.e., 𝑋 = 𝑍, with dim 𝑍 = 12 dim 𝑌 .
Then we have dim 𝑇𝑧 (𝑋) + dim 𝑇𝑧 (𝑍) = dim 𝑇𝑧 (𝑌 ), but the tangent spaces 𝑇𝑧 (𝑍) and
𝑇𝑧 (𝑋) are equal and do not span 𝑇𝑧 (𝑌 ).

∙ However, we will see later a way to make self-intersections interesting. This will be
made possible by Thom’s transversality theorem and invariance under homotopy. More
on this later in the chapter on intersection theory.

6.2.1 Intersection of submanifolds

The theorem on transversality for this special case says:

Theorem 6.4 (Intersection of transverse submanifolds) The intersection of two

transverse submanifolds 𝑋 and 𝑍 of 𝑌 is again a submanifold. Moreover, the codi-
mensions in 𝑌 satisfy

codim (𝑋 ∩ 𝑍) = codim 𝑋 + codim 𝑍.

∙ The additivity of codimensions follows from the codimension formula of the previous
theorem:

codim 𝑖−1 (𝑍) in 𝑋 = codim 𝑍 in 𝑌

⇒ dim 𝑋 − dim 𝑋 ∩ 𝑍 = dim 𝑌 − dim 𝑍
⇒ dim 𝑌 − dim 𝑋 ∩ 𝑍 = (dim 𝑌 − dim 𝑍) + (dim 𝑌 − dim 𝑋)
⇒ codim 𝑋 ∩ 𝑍 = codim 𝑍 + codim 𝑋.
Chapter 6. Transversality 115

Figure 6.3: The intersection of a sphere and a circle is transverse unless the there is only one
intersection point.

Remark 6.5 (Intersect as little as possible) We have just learned that two manifolds
intersect transversally if their tangent spaces together span the whole ambient space. A
different way to think of transversality is: Two manifolds intersect transversally if they
intersect as little as possible at every point. And we measure the degree of intersection
in terms of tangent spaces: If two submanifolds intersect, then they intersect transver-
sally if the intersection of their tangent spaces in the ambient tangent space is minimal.

∙ Note that the converse of the theorem is not true. We have seen a simple example last
time: the submanifolds 𝑋 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑦 = 𝑥2 } and 𝑍 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑦 = 0}
do not intersect transversally at 0 in 𝑌 = ℝ2 , but their intersection 𝑋 ∩ 𝑍 = {0} is a
zero-dimensional manifold. However, there do, of course, exist intersections which are
not transversal and where the intersection is not a manifold. We will look at an example
below!

∙ It is useful to note that any smooth map 𝑓 ∶ 𝑋 → 𝑌 whose image does not meet a
submanifold 𝑍 of 𝑌 , i.e., 𝑓 −1 (𝑍) = ∅, is transverse to 𝑍 for trivial reasons. For in this
case there is no condition to be satisfied. In particular, two submanifolds which do not
intersect at all are transverse.

∙ On the other hand, if 𝑓 ∶ 𝑋 → 𝑌 is a submersion, then 𝑓 is transverse to every sub-

manifold 𝑍 of 𝑌 , since then Im (𝑑𝑓𝑥 ) = 𝑇𝑓 (𝑥) (𝑌 ) for every 𝑥.

Remark 6.6 (The ambient space matters!) Transversality of 𝑋 and 𝑍 also depends
on the ambient space 𝑌 . For example, the two coordinate axes intersect transversally
in ℝ2 , but not when considered to be submanifolds of ℝ3 . In general, if the dimensions
of 𝑋 and 𝑍 do not add up to at least the dimension of 𝑌 , then they can only intersect
transversally by not intersecting at all. For example, if 𝑋 and 𝑍 are curves in ℝ3 ,
then 𝑋 − ⋔ 𝑌 if and only if 𝑋 ∩ 𝑌 = ∅.

By applying the formula we got for the tangent spaces of 𝑓 −1 (𝑍) to 𝑓 being the inclusion
map 𝑋 ⊂ 𝑌 we get the following useful result:
116 6.2. Transverse intersections

Figure 6.4: The intersection of two circles in the plane is transverse unless the there is exactly
one intersection point.

Theorem 6.7 (Tangent space of intersections) If 𝑋 and 𝑍 are submanifolds which

meet transversally in 𝑌 , then the tangent space to 𝑋 ∩ 𝑍 is the intersection of the
tangent spaces, i.e.,

𝑇𝑥 (𝑋 ∩ 𝑍) = 𝑇𝑥 (𝑋) ∩ 𝑇𝑥 (𝑍) for all 𝑥 ∈ 𝑋 ∩ 𝑍.

6.2.2 Examples

Let us have a look at some examples:

Example 6.8 (A primer to Brieskorn manifolds) Recall the map

𝑓 ∶ ℝ4 → ℝ,
(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) → 𝑥1 + 𝑥22 + 𝑥33 + 𝑥44 .

We showed before that 𝑓 is a submersion and hence that the preimage

𝑍 = 𝑓 −1 (0) = {(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) ∈ ℝ4 ∶ 𝑥1 + 𝑥22 + 𝑥33 + 𝑥44 = 0}

is a manifold in ℝ4 .
Now we show 𝑍 and 𝕊3 intersect transversally in ℝ4 where 𝕊3 ⊂ ℝ4 denotes the
three-dimensional sphere. This will imply that 𝑍 ∩ 𝕊3 is a smooth manifold.
To do this we need to check that 𝑇𝑧 (𝑍) + 𝑇𝑧 (𝕊3 ) = 𝑇𝑧 (ℝ4 ) = ℝ4 for all 𝑧 ∈ 𝑍 ∩ 𝕊3 .
Since 𝑇𝑧 (𝑍) and 𝑇𝑧 (𝕊3 ) are both three-dimensional subspaces of ℝ4 , it suffices to show
that, for every 𝑧 ∈ 𝑍 ∩ 𝕊3 , there is at least one vector 𝐯 in 𝑇𝑧 (𝑍) which is not contained
in 𝑇𝑧 (𝕊3 ).
The tangent space to 𝑍 in a point 𝑧 ∈ 𝑍 is the subspace in ℝ4 given by the kernel of
the derivative 𝑑𝑓𝑧 which in the standard bases is given by the 1 × 4-matrix
( )
𝑑𝑓𝑧 = 1 2𝑥2 3𝑥23 4𝑥34 .
Chapter 6. Transversality 117

Let 𝑧 = (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) be any point in 𝑍 ∩ 𝕊3 . Then the vector

⎛12𝑥1 ⎞
⎜ 6𝑥 ⎟
𝐯 ∶= ⎜ 2 ⎟
⎜ 4𝑥3 ⎟
⎝ 2𝑥4 ⎠

lies in 𝑇𝑧 (𝑍), since

⎛12𝑥1 ⎞
( ) ⎜ 6𝑥 ⎟
𝑑𝑓𝑧 (𝐯) = 1 2𝑥2 3𝑥23 4𝑥34 ⋅ ⎜ 2 ⎟
⎜ 4𝑥3 ⎟
⎝ 3𝑥4 ⎠
= 12𝑥1 + 12𝑥22 + 12𝑥33 + 12𝑥34
= 12𝑓 (𝑧)
= 0.

But 𝐯 is not an element in 𝑇𝑧 (𝕊3 ). For, recall that 𝑇𝑧 (𝕊3 ) is the subspace in ℝ4 which
is orthogonal to the vector 𝑧, i.e.,

𝑇𝑧 (𝕊3 ) = {𝑤 ∈ ℝ4 ∶ 𝑧 ⟂ 𝑤 = 0}.

We can check orthogonality via the scalar product in ℝ4 :

𝑧 ⟂ 𝑤 ⇐⇒ 𝑧 ⋅ 𝑤 = 0.

For 𝐯 we calculate

⎛12𝑥1 ⎞
( ) ⎜ 6𝑥2 ⎟ 2 2 2 2
𝑧 ⋅ 𝐯 = 𝑥1 𝑥2 𝑥3 𝑥4 ⋅ ⎜ ⎟ = 12𝑥1 + 6𝑥2 + 4𝑥3 + 3𝑥4 > 0.
4𝑥
⎜ 3⎟
⎝ 3𝑥4 ⎠

Thus 𝐯 is not an element in 𝑇𝑧 (𝕊3 ). Hence 𝑍 and 𝕊3 meet transversally in ℝ4 .

By the theorem, the codimension of 𝑍 ∩ 𝕊3 in 𝕊3 equals the codimension of 𝑍 in ℝ4 .
Thus dim 𝑍 ∩ 𝕊3 = 2.

Example 6.9 (Hyperboloid meets a sphere) In 𝑌 = ℝ3 , we consider the two subman-

ifolds

𝑋 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 − 𝑧2 = 1}

and the sphere

𝑍𝑎 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 + 𝑧2 = 𝑎}

with 𝑎 > 0. See Figure 6.5. We would like to understand for which 𝑎 do these two
submanifolds intersect transversally in 𝑌 .
Therefore, we need to determine the tangent spaces of 𝑋 and 𝑍𝑎 at points where
118 6.2. Transverse intersections

they intersect. We observe that 𝑋 = 𝑓 −1 (1) for the map

𝑓 ∶ ℝ3 → ℝ, (𝑥, 𝑦, 𝑧) → 𝑥2 + 𝑦2 − 𝑧2 .

Expressed as a matrix in the standard bases the derivative 𝑑𝑓𝑝 at a point 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 𝑧𝑝 )
has the form

𝑑𝑓𝑝 = (2𝑥𝑝 2𝑦𝑝 − 2𝑧𝑝 ) ∶ ℝ3 → ℝ.

This map is surjective for all 𝑝 ∈ ℝ3 ⧵ {0}. Hence 1 is a regular value of 𝑓 and the
tangent space to 𝑋 at 𝑝 is the kernel of 𝑑𝑓𝑝 . For 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 𝑧𝑝 ) ∈ 𝑋 with 𝑧𝑝 ≠ 0, this
is

𝑇𝑝 (𝑋) = Ker (𝑑𝑓𝑝 ) = span{(𝑧𝑝 , 0, 𝑥𝑝 ), (0, 𝑧𝑝 , 𝑦𝑝 )} ⊂ ℝ3 .

For all 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 0) ∈ 𝑋, this is

𝑇𝑝 (𝑋) = Ker (𝑑𝑓𝑝 ) = span{(−𝑦𝑝 , 𝑥𝑝 , 0), (0, 0, 1)} ⊂ ℝ3 .

Similarly, we observe that 𝑍𝑎 = 𝑔 −1 (𝑎) for the map

𝑔 ∶ ℝ3 → ℝ, (𝑥, 𝑦, 𝑧) → 𝑥2 + 𝑦2 + 𝑧2 .

Expressed as a matrix in the standard bases the derivative 𝑑𝑔𝑝 at a point 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 𝑧𝑝 )
has the form

𝑑𝑔𝑝 = (2𝑥𝑝 2𝑦𝑝 2𝑧𝑝 ) ∶ ℝ3 → ℝ.

This map is surjective for all 𝑝 ∈ ℝ3 ⧵ {0}. Hence 𝑎 > 0 is a regular value of 𝑔 and the
tangent space to 𝑍 at 𝑝 is the kernel of 𝑑𝑔𝑝 . For 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 𝑧𝑝 ) ∈ 𝑍𝑎 with 𝑧𝑝 ≠ 0, this
is

𝑇𝑝 (𝑍𝑎 ) = Ker (𝑑𝑔𝑝 ) = span{(−𝑧𝑝 , 0, 𝑥𝑝 ), (0, −𝑧𝑝 , 𝑦𝑝 )} ⊂ ℝ3 .

For all 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 0) ∈ 𝑍𝑎 , this is

𝑇𝑝 (𝑍𝑎 ) = Ker (𝑑𝑔𝑝 ) = span{(−𝑦𝑝 , 𝑥𝑝 , 0), (0, 0, 1)} ⊂ ℝ3 .

Now 𝑋 and 𝑍 intersect in the points 𝑝 = (𝑥, 𝑦, 𝑧) which satisfy

𝑥2 + 𝑦2 − 𝑧2 − 1 = 0 = 𝑥2 + 𝑦2 + 𝑧2 − 𝑎.

Subtracting both equations yields the condition

2𝑧2 = 𝑎 − 1. (6.4)

This gives us three cases for the intersection 𝑋 ∩ 𝑍𝑎 :

∙ If 𝑎 < 1, then 𝑋 and 𝑍𝑎 do not intersect, i.e., 𝑋 ∩ 𝑍𝑎 = ∅, since there is no 𝑧 ∈ ℝ

which satisfies condition (6.4).

∙ If 𝑎 = 1, then we have 𝑧 = 0 and 𝑋 and 𝑍1 intersect in the circle with radius 1

in the 𝑥𝑦-plane in ℝ3 with the origin as center, i.e.,

𝑋 ∩ 𝑍1 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 = 1 and 𝑧 = 0}.

Chapter 6. Transversality 119
∙ If 𝑎 > 1, then 𝑋 and 𝑍𝑎 intersect in two disjoint circles
√ which lie in the planes
parallel to the 𝑥𝑦-plane in ℝ3 with 𝑧-coordinate 𝑧 = ± (𝑎 − 1)∕2:
{ √ }
𝑎+1
𝑋 ∩ 𝑍𝑎 = (𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 = and 𝑧 = ± (𝑎 − 1)∕2 .
2

Now we check transversality:a

∙ if 𝑎 < 1, then the intersection is empty and therefore transversal.

∙ if 𝑎 = 1, then we see 𝑇𝑝 (𝑋) = 𝑇𝑝 (𝑍1 ) from the above descriptions of the tangent
spaces, since 𝑧𝑝 = 0 for all points in 𝑋 ∩ 𝑍1 . In particular, the tangent spaces
span the same plane in ℝ3 , and not all of ℝ3 , at every 𝑝 ∈ 𝑋 ∩ 𝑍1 . Thus the
intersection is not transversal.

∙ if 𝑎 > 1, let 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 𝑧𝑝 ) ∈ 𝑋 ∩ 𝑍𝑎 . In this case, we have 𝑧𝑝 ≠ 0. Then 𝑇𝑝 (𝑋)

and 𝑇𝑝 (𝑍𝑎 ) together span all of ℝ3 : For any point 𝑝 = (𝑥𝑝 , 𝑦𝑝 , 𝑧𝑝 ) ∈ 𝑋 ∩ 𝑍𝑎 , 𝑥𝑝
and 𝑦𝑝 cannot both be zero. If 𝑥𝑝 ≠ 0, then the vector (−𝑧𝑝 , 0, 𝑥𝑝 ) ∈ 𝑇𝑝 (𝑍𝑎 ) is not
a linear combination of the vectors (𝑧𝑝 , 0, 𝑥𝑝 ) and (0, 𝑧𝑝 , 𝑦𝑝 ) which span 𝑇𝑝 (𝑋).
And if 𝑦𝑝 ≠ 0, then the vector (0, −𝑧𝑝 , 𝑦𝑝 ) ∈ 𝑇𝑝 (𝑍𝑎 ) is not a linear combination
of the vectors (𝑧𝑝 , 0, 𝑥𝑝 ) and (0, 𝑧𝑝 , 𝑦𝑝 ). Since 𝑇𝑝 (𝑋) is 2-dimensional, this shows

𝑇𝑝 (𝑋) + 𝑇𝑝 (𝑍𝑎 ) = ℝ3 at every 𝑝 ∈ 𝑋 ∩ 𝑍𝑎 .

Thus the intersection is transverse.

a
Recall 𝑇𝑝 (ℝ3 ) = ℝ3 at every 𝑝

Figure 6.5: The intersection is transverse if 𝑎 > 1, and not transverse if 𝑎 = 1. In both cases,
however, the intersection is a smooth manifold.

Here is an example of an intersection which is not transverse and where the intersection
is not a manifold:
120 6.2. Transverse intersections

Example 6.10 (A non-transverse intersection) Let 𝑌 = ℝ3 and let 𝑍 be the hyper-

plane defined by

𝑍 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥 = 1}

and let 𝑋 be the hyperboloid defined by

𝑋 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 − 𝑧2 = 1}.

The intersection of 𝑋 and 𝑍 is given by the points satisfying 𝑥 = 1 and 𝑥2 + 𝑦2 − 𝑧2 = 1,

i.e., all points such that 𝑥 = 1 and 𝑦2 = 𝑧2 . This means

𝑋 ∩ 𝑍 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥 = 1, 𝑦 = ±𝑧}.

We have seen in Section 2.3.5 that a space consisting of two lines crossing each other
is not a manifold. The intersection point, i.e., the point 𝑝 = (1, 0, 0), does not have a
neighborhood in 𝑋 ∩ 𝑍 which is diffeomorphic to an open subset in Euclidean space.
Thus 𝑋 ∩ 𝑍 is not a manifold. See Figure 6.6.
As a reality check, let us look at the tangent spaces to 𝑋 and 𝑍 at 𝑝: Since 𝑍 is
a parallel translate of a vector subspace of ℝ3 , we see that 𝑇𝑝 (𝑍) is the 𝑦𝑧-plane in ℝ3
(all points with 𝑥 = 0). The tangent space to 𝑋 was calculated in the previous example
(and in an exercise). At 𝑝 = (1, 0, 0), 𝑇𝑝 (𝑋) is the vector subspace in ℝ3 spanned by the
vectors (0, 1, 0) and (0, 0, 1). In other words, 𝑇𝑝 (𝑋) is the 𝑥𝑦-plane in ℝ3 . Thus 𝑇𝑝 (𝑍)
and 𝑇𝑝 (𝑋) do not span 𝑇𝑝 (𝑌 ) = ℝ3 . The problem here is that 𝑍 is’the tangent plane to
𝑋 at 𝑝.

Figure 6.6: The intersection is not transverse and 𝑋 ∩ 𝑍 is not a manifold, since we get two
axes that cross each other.

Famous examples of transverse intersections are provided by Brieskorn manifolds:

Remark 6.11 (Exotic Spheres) Consider the following intersections in ℂ5 ⧵ {0}:

𝕊7𝑘 ={𝑧21 + 𝑧22 + 𝑧23 + 𝑧34 + 𝑧6𝑘−1

5
= 0}
∩ {|𝑧1 |2 + |𝑧2 |2 + |𝑧3 |2 + |𝑧4 |2 + |𝑧5 |2 = 1}.
Chapter 6. Transversality 121
In the exercises, we show that this is a transverse intersection. One can show that, for
each value 𝑘 = 1, … , 28, the space 𝑆𝑘7 is a smooth manifold which is homeomorphic
to the seven-sphere 𝕊7 . But none of these manifolds are diffeomorphic.
These are so called exotic 7-spheres. They were constructed by Brieskorn and repre-
sent each of the 28 diffeomorphism classes on the space 𝕊7 . That such exotic 7-spheres
is a famous and groundbreaking result of Milnor. Milnor’s work started an amazing
story about the diffeomorphic structures on spheres which culminated in the solution of
the Kervaire Invariant One Problem by Hill, Hopkins and Ravenel in 2009.
122 6.3. Exercises and more examples
6.3 Exercises and more examples

6.3.1 Transversality

Exercise 6.1 As a first test, answer the following questions:

(a) Let 𝑧 = (𝑎, 𝑏) ∈ 𝕊1 ⊆ ℝ2 and let 𝑁𝑧 = {(𝑎, 𝑦) ∶ 𝑦 ∈ ℝ} be the vertical line

intersecting the circle at 𝑧. When is 𝕊1 ⊆ ℝ2 transverse to 𝑁𝑧 ⊆ ℝ2 ?

(b) Which of the following linear spaces intersect transversally?

∙ The 𝑥𝑦-plane and the 𝑧-axis.

∙ The 𝑥𝑦-plane and the plane spanned by {(3, 2, 0), (0, 4, −1)}.
∙ The plane spanned by {(1, 0, 0), (2, 1, 0)} and the 𝑦-axis in ℝ3 .
∙ ℝ𝑘 × {0} and {0} × ℝ𝑙 in ℝ𝑛 . (The answer depends on 𝑘, 𝑙, and 𝑛.)
∙ 𝑉 × {0} and the diagonal in 𝑉 × 𝑉 , for a real vector space 𝑉 .
∙ The spaces of symmetric (𝐴𝑡 = 𝐴) and skew symmetric (𝐴𝑡 = −𝐴) matrices
in 𝑀(𝑛).

(c) Do 𝑆𝐿(𝑛) and 𝑂(𝑛) meet transversally in 𝑀(𝑛)?

Exercise 6.2 Recall the maps

( )
𝑒𝑡 + 𝑒−𝑡 𝑒𝑡 − 𝑒−𝑡
2
𝑓 ∶ ℝ → ℝ , 𝑡 → , and 𝑔 ∶ ℝ2 → ℝ, (𝑥, 𝑦) → 𝑥2 − 𝑦2
2 2

from previous exercise sets. Is the set Im (𝑓 ), the image of 𝑓 in ℝ2 , a manifold? Is the
set (𝑔◦𝑓 )−1 (1) a manifold?

𝑓 𝑔
Exercise 6.3 Let 𝑋 ←←←→ ← 𝑍 be a sequence of smooth maps between manifolds, and
← 𝑌 ←←→
let 𝑊 ⊂ 𝑍 be a submanifold. Assume that 𝑔 is transversal to 𝑊 . Show:

𝑓−
⋔ 𝑔 −1 (𝑊 ) if and only if (𝑔◦𝑓 ) −
⋔𝑊.

Exercise 6.4 Let 𝑉 be a vector space, and let Δ be the diagonal of 𝑉 × 𝑉 . For a linear
map 𝐴 ∶ 𝑉 → 𝑉 , consider the graph Γ(𝐴) = {(𝑣, 𝐴𝑣) ∶ 𝑣 ∈ 𝑉 }. Show that Γ(𝐴) − ⋔ Δ if
and only if +1 is not an eigenvalue of 𝐴.

Exercise 6.5 Let 𝑓 ∶ 𝑋 → 𝑋 be a map, and let 𝑥 be a fixed point of 𝑓 , i.e., 𝑓 (𝑥) = 𝑥. If
+1 is not an eigenvalue of 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑥 (𝑋), then 𝑥 is called a Lefschetz fixed point
of 𝑓 . The map 𝑓 is called a Lefschetz map if all its fixed points are Lefschetz. Prove that
if 𝑋 is compact and 𝑓 is Lefschetz, then 𝑓 has only finitely many fixed points.
Hint: Show that the intersection of the graph of 𝑓 and the diagonal of 𝑋 is a 0-
Chapter 6. Transversality 123
dimensional submanifold of 𝑋 × 𝑋.

Exercise 6.6 Consider the following intersections in ℂ5 ⧵ {0}:

𝑆𝑘7 = {𝑧21 + 𝑧22 + 𝑧23 + 𝑧34 + 𝑧6𝑘−1

5
= 0} ∩ {|𝑧1 |2 + |𝑧2 |2 + |𝑧3 |2 + |𝑧4 |2 + |𝑧5 |2 = 1}.

Prove that 𝑆𝑘7 is a 7-dimensional manifold by showing that the intersection is transver-
sal in ℂ5 ⧵ {0}.
Hint:
( At some point you may want ) to show that, at a point 𝑧 = (𝑧1 , … , 𝑧5 ), the vector
𝑤 ≔ 𝑚2 𝑧1 , 𝑚2 𝑧2 , 𝑚2 𝑧3 , 𝑚3 𝑧4 , 6𝑘−1
𝑚
𝑧5 , with 𝑚 ≔ 2 ⋅ 3 ⋅ (6𝑘 − 1), lies in one of the tangent
spaces but not in the other.
7. Sard’s theorem and Morse functions

7.1 The Theorem of Brown and Sard

In the previous sections we have seen how useful regular values are. This motivates to ask:

Question Given a smooth map 𝑓 ∶ 𝑋 → 𝑌 . How many regular values are there in
𝑌?

Recall that a 𝑦 ∈ 𝑌 which is not a regular value for 𝑓 is called a critical value. Hence we
may ask the equivalent question: How many critical values are there?

We have seen an answer to this question in one situation:

∙ In Milnor’s proof of the Fundamental Theorem of Algebra, we showed that the smooth
map in question had only finitely many critical values. Actually, we showed that the set
of critical points of the map 𝕊2 → 𝕊2 we defined using the given polynomial was finite.
∙ We might hope that set of critical values is always finite.
∙ However, finiteness is too good to be true when we allow 𝑋 and 𝑌 to be arbitrary smooth
manifolds.

So what is the correct analog of finiteness for our situation? Here is our first answer:

Theorem 7.1 (Sard’s Theorem) Let 𝑋 and 𝑌 be smooth manifolds and let 𝑓 ∶ 𝑋 → 𝑌
be a smooth map. Then the set of regular values of 𝑓 is a dense subset of 𝑌 , i.e., every
open subset of 𝑌 contains a regular value.

∙ This theorem is a key result in differential topology. We will see later that it will be
crucial for Thom’s Transversality Theorem which will make intersection theory work.

We will now reformulate and simplify the theorem. To do this and for the proof we are
going to use the following terminology:

Definition 7.2 (The interior of a set) Let 𝑋 be a topological space, and 𝑆 a subset
of 𝑋. Then the interior of 𝑆, denoted int(𝑆), is the union of all open subsets of 𝑋
contained in 𝑆. By definition, the interior of any 𝑆 is an open subset of 𝑋. In fact, it
is the largest open subset of 𝑋 which is contained in 𝑆. See Figure 7.1.
If 𝑆 ⊂ ℝ𝑁 then int(𝑆) is the set of all points 𝑠 ∈ 𝑆 such that there is a small open

124
Chapter 7. Sard’s theorem and Morse functions 125
ball centred at 𝑥 which is contained in 𝑆.
If 𝑈 is an open subset of 𝑋 then int(𝑈 ) = 𝑈 . In particular, if 𝑋 ⊂ ℝ𝑁 is open then
int(𝑋) = 𝑋. In general, however, int(𝑆) is a proper subset of 𝑋.

Figure 7.1: The interior of a subset 𝑆 in the topological space 𝑋 is the union of all open subsets
of 𝑋 which are contained in 𝑆.

We recall the following key facts from general topology:

∙ A subset 𝐴 ⊂ 𝑋 has empty interior if and only if its complement 𝑋 − 𝐴 is dense in 𝑋.

∙ The countable intersection of open dense subsets is a dense subset in ℝ𝑝 .

∙ Equivalently, the countable union of compact subsets with empty interior in ℝ𝑝 is a subset
with empty interior.

∙ The analogous statements hold in every smooth 𝑝-dimensional manifold.1

We will show that Sard’s theorem is a consequence of the following result:

Theorem 7.3 (Euclidean case) Let 𝑓 ∶ ℝ𝑛 → ℝ𝑝 be a smooth map. Then the set of
regular values of 𝑓 is a dense subset of ℝ𝑝 . Equivalently, the set of critical values of 𝑓
is a set with empty interior.

Before we start proving things, we introduce some more notation:

∙ For a smooth map 𝑓 ∶ 𝑋 → 𝑌 we denote by 𝐶𝑓 the set of critical points of 𝑓 , by 𝐷𝑓 the

set of critical values, and by 𝑅𝑓 the set of regular values.

∙ Note that it follows from the Local Submersion Theorem 4.2 that 𝐶𝑓 is a closed subset
of 𝑋.

1
This means that ℝ𝑝 and smooth manifolds in general are Baire spaces.
126 7.1. The Theorem of Brown and Sard
Proof that Theorem 7.3 implies Sard’s Theorem 7.1: First assume 𝑌 = ℝ𝑝 . Let 𝑥 ∈ 𝑋
be a point and let 𝜙 ∶ ℝ𝑛 → 𝑋 be a local parametrization.2 We set 𝑊 ∶= 𝜙(ℝ𝑛 ) and 𝐾 =
𝜙(𝔹̄ 𝑛1 (0)), where 𝔹̄ 𝑛1 (0) is the closed unit ball in ℝ𝑛 . In particular, 𝐾 is a compact subspace in
𝑋. By Theorem 7.3, the set of critical values of 𝑓 ◦𝜙

𝐷𝑓 ◦𝜙 = 𝑓 (𝐶𝑓 ∩ 𝑊 )

has empty interior. Thus, since 𝐶𝑓 is closed and hence 𝐶𝑓 ∩ 𝐾 is compact, 𝑓 (𝐶𝑓 ∩ 𝐾) is
a compact subset of ℝ𝑝 with empty interior. Since 𝑋 can be covered by countably many
neighborhoods of the form 𝐾, 𝐷𝑓 = 𝑓 (𝐶𝑓 ) is a countable union union of compact subsets
with empty interior.

Second, for an arbitrary smooth 𝑝-dimensional manifold 𝑌 , we can find countably many
local parametrizations 𝜓𝑖 of 𝑌 such that the union of the compact sets 𝐾𝑖 ∶= 𝜓𝑖 (𝔹̄ 𝑝1 (0)) covers
𝑌 . By the first case, the intersection of the 𝐷𝑓 with any of these sets 𝐾𝑖 , is a compact subset
with empty interior. Hence 𝐷𝑓 is a countable union of compact subsets with empty interior
and has itself empty interior.

Proof of Theorem 7.3: The proof will proceed in two steps: First we prove the result for
function ℝ𝑛 → ℝ. Second we use this to reduce the case ℝ𝑝 to ℝ𝑝−1 and will conclude by
induction.

∙ The case 𝑝 = 1:

If 𝑛 = 0, there is nothing to prove. Now assume that the assertion is true for 𝑛 − 1. Let
𝐶𝑖 ⊂ ℝ𝑛 be the closed set of points where all partial derivatives of 𝑓 of order ≤ 𝑖 vanish.
Then 𝐶𝑓 = 𝐶1 and 𝐶1 ⊃ 𝐶2 ⊃ 𝐶3 ⊃ ⋯. Hence

𝐶𝑓 = (𝐶1 − 𝐶2 ) ∪ ⋯ ∪ (𝐶𝑛−1 − 𝐶𝑛 ) ∪ 𝐶𝑛 ,

and thus
𝐷𝑓 = 𝑓 (𝐶1 − 𝐶2 ) ∪ ⋯ ∪ 𝑓 (𝐶𝑛−1 − 𝐶𝑛 ) ∪ 𝑓 (𝐶𝑛 ).

Hence the theorem follows for 𝑝 if we can show that each of the sets 𝑓 (𝐶𝑖 − 𝐶𝑖+1 ) for all
𝑖 ≥ 1 and 𝑓 (𝐶𝑖 ) for 𝑖 ≥ 𝑛 are countable union of closed subsets with empty interior. This will
be shown in the following two lemmas:

Lemma 7.4 The set 𝑓 (𝐶𝑖 − 𝐶𝑖+1 ) is a countable union of closed subsets with empty
interior for all 𝑖 ≥ 1.

Proof: We claim the following: For each 𝑢 ∈ 𝐶𝑖 − 𝐶𝑖+1 there is a compact neighborhood 𝐾
disjoint with 𝐶𝑖+1 and an (𝑛 − 1)-dimensional submanifold 𝑍 ⊂ ℝ𝑛 such that 𝐶𝑖 ∩ 𝐾 ⊂ 𝑍. Then
every point of 𝐶𝑖 ∩ 𝐾 is critical for 𝑓|𝑍 , since it is critical for 𝑓 . By our induction hypothesis,
𝑓 (𝐶𝑖 ∩ 𝐾) is then a closed subset in ℝ without interior points. Since 𝐶𝑖 − 𝐶𝑖+1 can be covered
by countably many arbitrarily small compact neighborhoods 𝐾, this will prove the lemma.
2
Note that we can choose the domain of 𝜙 to be all of ℝ𝑛 , by stretching an open ball 𝔹𝑛𝑟 (0) to all of ℝ𝑛 via a
diffeomorphism.
Chapter 7. Sard’s theorem and Morse functions 127
Now we prove the claim. Since 𝑢 ∈ 𝐶𝑖 − 𝐶𝑖+1 , there is some 𝑖-th order partial derivative of
𝑓 whose first order partial derivatives do not all vanish at 𝑢. We denote this partial derivative
by 𝑔. Then 𝑢 is a regular point for 𝑔. By definition of 𝐶𝑖 , we know 𝑔(𝑢) = 0. By the Local
Submersion Theorem, we can then find an open neighborhood 𝑈 of 𝑢 such that 0 is a regular
value for 𝑔|𝑈 . Hence, by the Preimage Theorem, 𝑔 −1 (0)∩𝑈 is an (𝑛−1)-dimensional subman-
ifold in 𝑈 . Moreover, since 𝐶𝑖 ⊆ 𝑔 −1 (0) by definition of 𝑔 and 𝐶𝑖 , we can set 𝑍 ∶= 𝑔 −1 (0) ∩ 𝑈
and choose a sufficiently small compact neighborhood 𝐾 of 𝑢 in 𝑈 . By definition of 𝑔, we also
know that 𝐾 is disjoint with 𝐶𝑖+1 .

Lemma 7.5 The set 𝑓 (𝐶𝑖 ) is a countable union of closed subsets with empty interior
for 𝑖 ≥ 𝑛.

Proof: By Taylor’s Theorem we have

1 𝑖 1
𝑓 (𝑥 + 𝑢) = 𝑓 (𝑥) + 𝜕𝑢 𝑓 (𝑥) + ⋯ + 𝜕𝑢 𝑓 (𝑥) + 𝜕 𝑖+1 𝑓 (𝑥 + 𝜆𝑢)
𝑖! (𝑖 + 1)! 𝑢
for any two points 𝑥 and 𝑢 in ℝ𝑛 , where 0 ≤ 𝜆 ≤ 1 is a real number and 𝜕𝑢 is the differential
operator
𝜕𝑢 = 𝑢1 𝜕∕𝜕𝑥1 + ⋯ + 𝑢𝑛 𝜕∕𝜕𝑥𝑛 .
Thus, if 𝑥 ∈ 𝐶𝑖 , then by definition of 𝐶𝑖 :

1
𝑓 (𝑥 + 𝑢) − 𝑓 (𝑥) = 𝜕 𝑖+1 𝑓 (𝑥 + 𝜆𝑢).
(𝑖 + 1)! 𝑢
If in addition 𝑥 and 𝑦 ∶= 𝑥 + 𝑢 lie in a convex set 𝐾, then 𝑥 + 𝜆𝑢 is also in 𝐾, and we get the
inequality
|𝑓 (𝑦) − 𝑓 (𝑥)|max ≤ 𝑐|𝑦 − 𝑥|𝑖+1
max

where |𝑤|max = max{|𝑤1 |, … , |𝑤𝑛 |} and 𝑐 is a constant depending on 𝐾 and 𝑓 only.

Now we let 𝐾 be the unit cube in ℝ𝑛 and consider the subdivision of 𝐾 into 𝑘𝑛 subcubes
with sides of length 1∕𝑘. Let 𝐾 ′ be one of these subcubes and suppose 𝑥 ∈ 𝐶𝑖 ∩ 𝐾 and 𝑦 ∈ 𝐾 ′ .
Then
|𝑦 − 𝑥|max ≤ 1∕𝑘.
This implies that 𝑓 (𝐶𝑖 ∩ 𝐾 ′ ) is contained in an interval of length 𝑐∕𝑘𝑖+1 .

Thus 𝑓 (𝐶𝑖 ∩ 𝐾) is contained in a union of 𝑘𝑛 intervals of joint length

𝑘𝑛 𝑐∕𝑘𝑖+1 ≤ 𝑐∕𝑘

where we use that 𝑖 ≥ 𝑛 by our assumption. Since 𝑘 is any positive integer, this length can be
arbitrarily small. Hence the set 𝑓 (𝐶𝑖 ∩ 𝐾) must have empty interior. Finally, ℝ𝑛 and therefore
𝐶𝑖 is contained in a countable union of unit cubes 𝐾. This proves the lemma.

∙ The case 𝑝 ≥ 2:

We assume the assertion holds for every smooth map 𝑓 ∶ ℝ𝑛 → ℝ𝑝−1 .

128 7.1. The Theorem of Brown and Sard
Let 𝑂 ⊂ be an nonempty open subset of
ℝ𝑝 ℝ𝑝 .
We will show that 𝑓 has regular values in
𝑂. This implies that the set of regular values for 𝑓 is dense in ℝ𝑝 .

If 𝑂 contains points which are not in 𝑓 (ℝ𝑛 ), then these points are regular values for 𝑓 , and
we are done.

Hence assume 𝑂 ⊆ 𝑓 (ℝ𝑛 ). Let 𝜋 ∶ ℝ𝑝 → ℝ𝑝−1 be the projection onto the first 𝑝 − 1
coordinates. Then 𝜋(𝑂) is open in ℝ𝑝−1 , since 𝜋 is a submersion and hence an open map.3 By
our induction hypothesis, the map 𝜋◦𝑓 ∶ ℝ𝑛 → ℝ𝑝−1 has a regular value 𝑦′ ∈ 𝜋(𝑂). In other
words, 𝑓 is transverse to the line 𝑌 ′ = 𝜋 −1 (𝑦′ ) in ℝ𝑝 .

Now let 𝑓 ′ ∶ 𝑋 ′ → 𝑌 ′ be the induced map. Since 𝑂 ⊆ 𝑓 (ℝ𝑛 ), the open set 𝑂 meets 𝑌 ′ ,
i.e., 𝑌 ′ ∩ 𝑂 ≠ ∅. Since 𝑌 ′ is diffeomorphic to ℝ, 𝑓 ′ has a regular value 𝑦′′ ∈ 𝑌 ′ ∩ 𝑂 by the
case 𝑝 = 1.

Thus, we have shown

∙ 𝑓 is transverse to 𝑌 ′ in ℝ𝑝 , and

∙ 𝑓 ′ ∶ 𝑓 −1 (𝑌 ′ ) → 𝑌 ′ has a regular value 𝑦′′ in 𝑌 ′ ∩ 𝑂.

By the chain rule and what we learned about transversality, this implies that 𝑦′′ ∈ 𝑌 ′ ∩ 𝑂
is a regular value for 𝑓 . This finishes the proof of Theorem 7.3.

∙ At the end we used a fact we proved in the exercises: If 𝑓 ∶ 𝑋 → ℝ𝑝 is transverse to a

submanifold 𝑌 ′ ⊂ ℝ𝑝 and the induced map 𝑓 ′ ∶ 𝑓 −1 (𝑌 ′ ) → 𝑌 ′ is transverse to 𝑌 ′′ ⊂ 𝑌 ′ ,
then 𝑓 is transverse to 𝑌 ′′ .
∙ Holm therefore thinks of the argument for the general case as splitting off a transverse
component of lower dimension to use induction.

To conclude this section, we remark that Sard’s Theorem is often formulated as follows:

Theorem 7.6 (Sard’s Theorem revisited) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map of mani-

folds with dim 𝑌 = 𝑝. Define the set 𝐶 to be

𝐶 = {𝑥 ∈ 𝑋 ∶ rank(𝑑𝑓𝑥 ) < 𝑝}.

Then the subset 𝑓 (𝐶) ⊂ 𝑌 of critical values has measure zero in 𝑌 .

Before we recall some basic measure theory, we observe how the different versions of the
theorem are related:

∙ This result is stronger than the previous version, and we invite the reader to investigate
the relationship between measure zero sets and sets with empty interior. However, the
version we proved suffices for our purposes. For, we will apply the theorem for knowing
that any small open neighborhood of a point contains a regular value.
3
We proved this fact in the exercises.
Chapter 7. Sard’s theorem and Morse functions 129

∙ (Measure zero in a not-measure zero box) A rectangular solid in ℝ𝑛 is just a

cartesian product of 𝑛 intervals in ℝ𝑛 , and its volume is the product of the lengths of the
𝑛 intervals. An arbitrary set 𝐴 in ℝ𝑛 is said to have (Lebesgue) measure zero if, for
every 𝜀 > 0, there exists a countable collection {𝑆1 , 𝑆2 , …} of rectangular solids in ℝ𝑛 ,
such that 𝐴 is contained in the union of the 𝑆𝑖 , and

∑
∞
vol (𝑆𝑖 ) < 𝜀.
𝑖=1

Then in a manifold 𝑋, an arbitrary subset 𝐶 ⊂ 𝑋 has measure zero if, for every local
parametrization 𝜙 of 𝑋, the preimage 𝜙−1 (𝐶) has measure zero in Euclidean space.
Note that measures and volumes depend on the ambient space!

∙ An example of a measure zero subset is given by the set of rational numbers in ℝ.

∙ Hence for measure theorists, almost every real number is irrational. This example il-
lustrates that something that happens almost never, can still happen often enough to be
noticed.

∙ By definition, no nonempty rectangular solid in ℝ𝑛 has measure zero. Hence it cannot be

contained in a set of measure zero.

∙ Now, every nonempty open subset of ℝ𝑛 contains some nonempty rectangular solid.
Thus, no nonempty open subset of ℝ𝑛 has measure zero.

∙ Hence, no nonempty open subset of a manifold 𝑌 has measure zero. In other words, no
set of measure zero in a manifold 𝑌 can contain a nonempty open subset of 𝑌 .

7.2 Morse Functions

We will now study a very interesting application of Sard’s Theorem. We understand the
local behaviour of smooth maps at regular points by the Local Submersion Theorem 4.2. But
what about the local behaviour at critical points? In fact, it is often at critical points that
the interesting stuff happens. For example, it is often at critical points that the topology of a
manifold can change.

For example, let 𝑓 ∶ 𝑋 → ℝ be a smooth function. If 𝑋 is compact, then we know that

𝑓 must have a maximum and a minimum. At a point 𝑥 ∈ 𝑋 where 𝑓 (𝑥) is either a maximal
or a minimal value, 𝑓 does not change in any direction in 𝑋. In other words, the derivative
𝑑𝑓𝑥 must vanish (recall 𝑑𝑓𝑥 (ℎ) is a measure for the change of 𝑓 in direction ℎ). Hence 𝑥 is a
critical point in our terminology.

A standard example is given by the height function on a torus, see Figure 7.2:

Change of homotopy types: We observe in this example that the homotopy type of the
fiber can change at critical points. You may have noticed that have not defined what the term
‘homotopy type’ means. Roughly speaking it is equivalence class of a space under the relation
130 7.2. Morse Functions

Figure 7.2: The critical points of the height function yield a decomposition of the torus.

of homotopy equivalence (which we have not defined either). For example, all contractible
spaces have the same homotopy type (with some assumptions in place). Anyway, we are not
going to fix this lack of precise definitions now. Instead we look at what happens with the
fibers on the torus in this concrete example: For 𝑠 < 0, the preimage ℎ−1 ([0, 𝑠)) under the
height functions is empty. The first critical value is 𝑠 = 0. For sufficiently small 𝑠 > 0, the
preimage ℎ−1 ([0, 𝑠)) has changed and looks like a two-dimensional disk, just a bit punched
in at the center. In particular, the preimage is contractible in this range. But when we pass the
next critical value, the light green dot on the vertical number line, another significant change
happens. For after it, the preimage ℎ−1 ([0, 𝑠)) looks like a bent cylinder. On this cylinder, there
are loops which are not homotopic to a constant map, e.g., the dark green circle. Thus, above
the critical value, the preimage ℎ−1 ([0, 𝑠)) is not contractible anymore. Hence the homotopy
type of the preimage has changed at a critical value. The next change of the homotopy type
happens when we pass the next critical value. The preimage ℎ−1 ([0, 𝑠]) (of the closed interval
[0, 𝑠]) becomes homeomorphic to a compact surface of genus one, i.e., it has one hole, with a
circle as boundary. Another list of things we need to and are going to define towards the end
of the semester. Finally, after passing the last critical value, the preimage is the whole torus.
The torus has the homotopy type of a compact surface with genus one, i.e., still one hole,
but without boundary. In total, we see that a lot of interesting stuff happens at the critical
values.

Back to the general study:

So let us stick to smooth functions, i.e., smooth maps to ℝ. We want to understand how
critical points look like locally.

Let us look a smooth function 𝑓 ∶ ℝ𝑘 → ℝ. Locally around a point 𝑐 ∈ ℝ𝑘 , we can describe

𝑓 by

∑𝑘
𝜕𝑓 1 ∑ 𝜕2𝑓
𝑘
𝑓 (𝑥) = 𝑓 (𝑐) + (𝑐) ⋅ (𝑥𝑖 − 𝑐𝑖 ) + (𝑐) ⋅ (𝑥𝑖 − 𝑐𝑖 )(𝑥𝑗 − 𝑐𝑗 ) + 𝑜(|𝑥|3 ).
𝑖=1
𝜕𝑥 𝑖 2 𝑖,𝑗=1
𝜕𝑥𝑖 𝜕𝑥 𝑗
Chapter 7. Sard’s theorem and Morse functions 131
If 𝑐 is a critical point, then by definition
( )
𝜕𝑓 𝜕𝑓
𝑑𝑓𝑐 = (𝑐), … , (𝑐) =0
𝜕𝑥1 𝜕𝑥𝑘

(otherwise 𝑑𝑓𝑐 was surjective as a linear map ℝ𝑘 → ℝ). Hence the best possible approximation
for the local behavior of 𝑓 at 𝑐 is the Hessian matrix of the second partial derivatives. Critical
points where the Hessian matrix is invertible is the best we can hope for.

Definition 7.7 (Non-degenerate critical points and Morse functions) For a smooth
function 𝑓(∶ ℝ𝑘 →)ℝ, a point 𝑐 ∈ ℝ𝑘 where 𝑑𝑓𝑐 vanishes, but the Hessian matrix
2𝑓
𝐻(𝑓 )𝑐 = 𝜕𝑥𝜕 𝜕𝑥 (𝑐) is invertible at 𝑐, is called a non-degenerate critical point.
𝑖 𝑗
A smooth function 𝑓 ∶ ℝ𝑘 → ℝ for which all critical points are non-degenerate is
called a Morse function.

Non-degenerate critical points are much easier to study than arbitrary critical points, since
they are isolated from the other critical points, i.e., there is an open neighborhood which does
not contain any other critical points. Hence Morse functions are easier to understand than
arbitrary smooth functions.

We check that non-degenerate critical points are isolated: We define a map

( )
𝜕𝑓 𝜕𝑓
𝑔 ∶ ℝ → ℝ by the formula 𝑔 =
𝑘 𝑘
,…, . (7.1)
𝜕𝑥1 𝜕𝑥𝑘

Then: 𝑑𝑓𝑥 = 0 ⇐⇒ 𝑔(𝑥) = 0.

Moreover, the matrix representing the derivative 𝑑𝑔𝑥 is the Hessian of 𝑓 at 𝑥. So if 𝑥 is non-
degenerate, then not only is 𝑔(𝑥) = 0, but 𝑔 maps a neighborhood of 𝑥 diffeomorphically
onto a neighborhood of 0 as well. In particular, 𝑔 is injective in that neighborhood of 𝑥. Thus
𝑔 can be zero at no other points in this neighborhood, and 𝑓 has no other critical point in this
neighborhood.

Another reason to be interested in Morse functions is the fact that there are a lot of them.

Theorem 7.8 (Morse functions on ℝ𝑘 are generic) Let 𝑓 ∶ 𝑈 → ℝ be a smooth

function defined on some open 𝑈 ⊆ ℝ𝑘 and 𝑎 ∈ ℝ𝑘 , define

𝑓𝑎 (𝑥) = 𝑓 (𝑥) + 𝑎 ⋅ 𝑥.

Then, for almost all 𝑎 ∈ ℝ𝑘 , 𝑓𝑎 is a Morse function.

Proof: We us again the function 𝑔 from (7.1).The derivative of 𝑓𝑎 at a point 𝑝 ∈ 𝑈 then

satisfies
( )
𝜕𝑓𝑎 𝜕𝑓𝑎
(𝑑𝑓𝑎 )𝑝 = (𝑝), … , (𝑝) = 𝑔(𝑝) + 𝑎.
𝜕𝑥1 𝜕𝑥𝑘
132 7.2. Morse Functions
Hence the critical points of 𝑓𝑎 are the points 𝑝 ∈ 𝑈 with 𝑔(𝑝) + 𝑎 = 0. Moreover, the Hessian
of 𝑓𝑎 at 𝑝 is the matrix 𝑑𝑔𝑝 , i.e.

𝐻(𝑓𝑎 )𝑝 = 𝐻(𝑓 )𝑝 = 𝑑𝑔𝑝 .

Hence

𝑓𝑎 is Morse ⇐⇒ det(𝐻(𝑓𝑎 )𝑝 ) ≠ 0 at all critical points 𝑝

⇐⇒ det(𝑑𝑔𝑝 ) ≠ 0 at all 𝑝 with 𝑔(𝑝) + 𝑎 = 0
⇐⇒ − 𝑎 is a regular value of 𝑔.

By Sard’s Theorem, −𝑎 is a regular value of 𝑔 for almost all 𝑎 ∈ ℝ𝑘 . Therefore almost every
𝑓𝑎 is a Morse function.

Now we would like to transport the concept of non-degenerate critical points to manifolds.

Lemma 7.9 (Independence of choice) So let 𝑋 be a smooth manifold. Suppose that

𝑓 ∶ 𝑋 → ℝ has a critical point at 𝑥 and that 𝜙 ∶ 𝑈 → 𝑋 is a local parametrization
with 𝜙(0) = 𝑥. Then

𝑑(𝑓 ◦𝜙)0 = 𝑑𝑓𝑥 ◦𝑑𝜙0

and hence 0 is a critical point for the function 𝑓 ◦𝜙. We call 𝑥 a non-degenerate critical
point for 𝑓 if 0 is a non-degenerate critical point for 𝑓 ◦𝜙.

Independence of choice:

Since we made a choice of a local parametrization for this definition, we need to make sure
that the criterion is independent of the choice.

So let 𝜓 ∶ 𝑉 → 𝑋 be another local parametrization with 𝜓(0) = 𝑥. We define 𝜃 ∶=

𝜓 −1 ◦𝜙 ∶ 𝑈 → 𝑉 . Since 𝜃 is a diffeomorphism, the critical points of 𝑓 ◦𝜙 and 𝑓 ◦𝜓◦𝜃 are the
same.

Assuming that 𝑥 is a critical point of 𝑓 , i.e., 𝑑𝑓𝑥 = 0, the chain rule implies for the two
Hessian matrices at 0:
𝐻(𝑓 ◦𝜙)0 = (𝑑𝜃0 )𝑡 𝐻(𝑓 ◦𝜓)0 𝑑𝜃0 .
Since 𝑑𝜃0 is invertible, we see

𝐻(𝑓 ◦𝜙)0 is invertible ⇐⇒ 𝐻(𝑓 ◦𝜓)0 is invertible.

An important result on Morse functions is that they can be described in some sort of canon-
ical form. It extends our understanding of the local behavior of smooth maps and is a key
result in this section:

Lemma 7.10 (Morse Lemma) Let 𝑋 be a smooth manifold and 𝑓 ∶ 𝑋 → ℝ be a

smooth function. Suppose that 𝑎 ∈ 𝑋 is a non-degenerate critical point of 𝑓 . Then
there is a local parametrization 𝜙 ∶ 𝑈 → 𝑋 with 𝜙(0) = 𝑎 and a local coordinate
Chapter 7. Sard’s theorem and Morse functions 133

system 𝜙−1 = (𝑥1 , … , 𝑥𝑘 ) around 𝑎 such that

𝑓 (𝑥) = 𝑓 (𝑎) − 𝑥21 − … − 𝑥2𝑠 + 𝑥2𝑠+1 + … + 𝑥2𝑘

for all 𝑥 ∈ 𝜙(𝑈 ) where 𝑠 is the number of negative eigenvalues of the Hessian of 𝑓 at
𝑎.

For the proof of Lemma 7.10 we follow [12, Part I, §2.1]. We are going to use the following
general observation:

Lemma 7.11 Let 𝑓 ∶ 𝔹𝑘 → ℝ be a smooth function defined on an open ball around

the origin in ℝ𝑘 with 𝑓 (0) = 0. Then we have

∑
𝑘
𝑓 (𝑥1 , … , 𝑥𝑘 ) = 𝑥𝑖 𝑔𝑖 (𝑥1 , … 𝑥𝑘 )
𝑖=1

𝜕𝑓
for some suitable smooth functions 𝑔𝑖 ∶ 𝔹𝑘 → ℝ with 𝑔𝑖 (0) = 𝜕𝑥𝑖
(0).

Proof: We have
1
𝑑𝑓 (𝑡𝑥1 , … , 𝑡𝑥𝑘 )
𝑓 (𝑥1 , … , 𝑥𝑘 ) = 𝑑𝑡
∫0 𝑑𝑡
1∑
𝑘
𝜕𝑓
= (𝑡𝑥 , … , 𝑡𝑥𝑘 ) ⋅ 𝑥𝑖 𝑑𝑡.
∫0
𝑖=1
𝜕𝑥𝑖 1
1 𝜕𝑓
Hence we may set 𝑔𝑖 (𝑥1 , … 𝑥𝑘 ) ∶= ∫0 𝜕𝑥𝑖
(𝑡𝑥1 , … , 𝑡𝑥𝑘 ) ⋅ 𝑥𝑖 𝑑𝑡.

Proof of Lemma 7.10: Without loss of generality, we can assume that 𝑎 is the origin and
that 𝑓 (𝑎) = 𝑓 (0) = 0. By Lemma 7.11 we can write

∑
𝑘
𝑓 (𝑥1 , … , 𝑥𝑘 ) = 𝑥𝑗 𝑔𝑗 (𝑥1 , … 𝑥𝑘 )
𝑗=1

in a small neighborhood of 0. Since 0 is a critical point, we must have

𝜕𝑓
𝑔𝑗 (0) = (0) = 0.
𝜕𝑥𝑗
Hence we can apply Lemma 7.11 to each 𝑔𝑗 and get

∑
𝑘
𝑔𝑗 (𝑥1 , … , 𝑥𝑘 ) = 𝑥𝑗 ℎ𝑖𝑗 (𝑥1 , … 𝑥𝑘 )
𝑗=1

for suitable smooth functions ℎ𝑖𝑗 defined on the small neighborhood around the origin. Com-
bining these expressions we get

∑
𝑘
𝑓 (𝑥1 , … , 𝑥𝑘 ) = 𝑥𝑖 𝑥𝑗 ℎ𝑖𝑗 (𝑥1 , … 𝑥𝑘 )
𝑖,𝑗=1
134 7.2. Morse Functions
We can assume that ℎ𝑖𝑗 = ℎ𝑗𝑖 since we can replace ℎ𝑖𝑗 with ℎ̄ 𝑖𝑗 = 21 (ℎ𝑖𝑗 + ℎ𝑗𝑖 ) and then get
∑
ℎ̄ 𝑖𝑗 = ℎ̄ 𝑗𝑖 and 𝑓 = 𝑥𝑖 𝑥𝑗 ℎ̄ 𝑖𝑗 . By construction we then get

𝜕2𝑓
ℎ𝑖𝑗 (0) = (0).
𝜕𝑥𝑖 𝜕𝑥𝑗

Thus the matrix 𝐻 with entry ℎ𝑖𝑗 (0) in position (𝑖, 𝑗) is the Hessian matrix of 𝑓 , computed in
the local coordinate system defined by 𝜙. Since 0 is a nondegenerate critical point, the Hessian
is invertible.

Now we have to show that we can choose a coordinate system such that 𝑓 has the simple
form of the lemma. To do so, we write 𝐻 ∶= 𝐻(𝑓 ◦𝜙))0 for the Hessian matrix and 𝐱 =
(𝑥1 , … , 𝑥𝑘 ). Then we have locally
𝑓 = 𝐱𝑡 𝐻𝐱.
Since the Hessian matrix 𝐻 is real symmetric, there is an orthonormal matrix 𝑃 such that
𝑃 𝑡 𝐻𝑃 is a diagonal matrix. Note that multiplying with 𝑃 𝑡 , or 𝑃 , corresponds to a change of
basis ℝ𝑘 → ℝ𝑘 . We write 𝐲 = 𝑃 𝑡 𝐱 for the coordinates with respect to the new basis. Let
𝜆1 , … , 𝜆𝑘 be the eigenvalues of 𝐻, where we order them, and hence the columns in 𝑃 , such
that the first 𝑠 𝜆𝑖 are negative and the remaining 𝑘 − 𝑠 are positive. Note that 𝐻 does not have
eigenvalue 0, since it is invertible by the assumption on 𝑎 being a nondegenerate critical point.
Then we get
∑𝑘
𝑡 𝑡 𝑡 𝑡 𝑡
𝑓 = 𝐱 𝑃 𝑃 𝐻𝑃 𝑃 𝐱 = 𝐲 𝑃 𝐻𝑃 𝐲 = 𝜆𝑖 𝑦2𝑖 .
𝑖=1

Finally, we make the change of basis ℝ𝑘

→ ℝ𝑘given by rescaling
√
𝑧𝑖 = |𝜆𝑖 |𝑦𝑖 for each 𝑖 = 1, … , 𝑘.

In the new local coordinate system 𝐳 we have

𝑓 = −𝑧21 − … − 𝑧2𝑠 + 𝑧2𝑠+1 + … + 𝑧2𝑘 .

Example 7.12 Let us look at the example of the height function on the torus. To make
the notation compatible with the above theorem, we denote the hight function by 𝑓 .
This is just a projection onto the vertical coordinate of the points on the torus. At the
point 𝑝 on the torus with 𝑓 (𝑝) = 0, we can choose local coordinates 𝑥, 𝑦 and write

𝑓 (𝑥, 𝑦) = 𝑓 (𝑝) + 𝑥2 + 𝑦2 = 𝑥2 + 𝑦2 .

At the next two critical values, we can choose local coordinates 𝑥, 𝑦 write 𝑓 as

𝑓 (𝑥, 𝑦) = constant + 𝑥2 − 𝑦2 .

At the final critical value, we can choose local coordinates 𝑥, 𝑦 such that

𝑓 (𝑥, 𝑦) = constant − 𝑥2 − 𝑦2 .

We are now going to discuss some situations where the Morse Lemma 7.10 is useful. We
will see another important application in Section 17.6 when we discuss the Poincaré–Hopf Index
Chapter 7. Sard’s theorem and Morse functions 135
Theorem 17.16. First, we can generalize the fact that almost all functions are Morse to the level
of manifolds: Suppose 𝑋 ⊂ ℝ𝑁 , and let 𝑥1 , … , 𝑥𝑁 ∈ ℝ𝑁 be the usual coordinate functions on
ℝ𝑁 . If 𝑓 ∶ 𝑋 → ℝ is a smooth function on 𝑋 and 𝑎 = (𝑎1 , … , 𝑎𝑁 ) is an 𝑁-tuple of numbers,
we define again a new function 𝑓𝑎 ∶ 𝑋 → ℝ by

𝑓𝑎 ∶= 𝑓 + 𝑎1 𝑥1 + ⋯ + 𝑎𝑁 𝑥𝑁 .

Theorem 7.13 (Morse functions on manifolds are generic) For every smooth function
𝑓 ∶ 𝑋 → ℝ and for almost every 𝑎 ∈ ℝ𝑁 , 𝑓𝑎 is a Morse function on 𝑋, i.e., all its
critical points are nondegenerate.

Proof: We would like to use the above result for 𝑈 ⊂ ℝ𝑘 open. Since 𝑋 ⊂ ℝ𝑁 is in general
not open (in fact, it is never open if dim 𝑋 < 𝑁), the strategy is to cover 𝑋 by open subsets
and then try to lift the 𝑘-dimensional result to open sets in ℝ𝑁 .

So let 𝑥 be any point in 𝑋. First we are going to choose a suitable local coordinate system
around 𝑥. Let 𝑣1 , … , 𝑣𝑘 ∈ ℝ𝑁 be a basis of 𝑇𝑥 (𝑋) (for 𝑘 = dim 𝑋). Then the matrix [𝑣1 ⋯ 𝑣𝑘 ],
having the 𝑣𝑖 ’s as columns, has rank 𝑘. Hence it has 𝑘 linearly independent rows, say 𝑖1 , … , 𝑖𝑘 .
Let 𝜋 ∶ ℝ𝑁 → ℝ𝑘 be projection defined by (𝑥1 , … , 𝑥𝑁 ) → (𝑥𝑖1 , … , 𝑥𝑖𝑘 ) where the 𝑥1 , … , 𝑥𝑁
denote the standard coordinates on ℝ𝑁 . Then

(𝑑𝜋𝑥 )|𝑇𝑥 (𝑋) ∶ 𝑇𝑥 (𝑋) → ℝ𝑘 is an isomorphism

by construction. Hence, by the Inverse Function Theorem,

𝜋|𝑋 ∶ 𝑋 → ℝ𝑘 is a local diffeomorphism.

Hence we can take the 𝑘-tuple of functions (𝑥𝑖1 , … , 𝑥𝑖𝑘 ) ∶ 𝑋 → ℝ𝑘 to define a local coordinate
system around 𝑥.

Therefore we can cover 𝑋 with open subsets 𝑈𝛼 ⊆ ℝ𝑁 such that on each 𝑈𝛼 some 𝑘-tuple of
the functions 𝑥1 , … , 𝑥𝑁 on ℝ𝑁 form a coordinate system. Moreover, by some general nonsense
on the topology of Euclidean space, it is always possible to choose a countable subfamily of
the 𝑈𝛼 ’s. Hence we may assume there are only countably many 𝑈𝛼 .

Let 𝑆 ⊂ ℝ𝑁 be the subset of 𝑎 such that 𝑓𝑎 is not Morse. Since the countable union of
sets with measure zero has measure zero, it suffices to show that for each 𝑈𝛼 the set 𝑆𝛼 of 𝑎’s
such that 𝑓𝑎 ∶ 𝑈𝛼 → ℝ is n̊ot Morse, has measure zero.

So let us look at one of the 𝑈𝛼 ’s. We want to show that 𝑆𝛼 has measure zero in ℝ𝑁 .

To simplify the notation, assume 𝑥1 , … , 𝑥𝑘 form a coordinate system around 𝑥 on 𝑈𝛼 . We

can write any 𝑎 ∈ ℝ𝑁 as 𝑎 = (𝑏, 𝑐), where 𝑏 denotes the first 𝑘 coordinates and 𝑐 denotes the
last 𝑁 − 𝑘 coordinates. Around a given point 𝑥, we can thus write

𝑓𝑎 (𝑥) = 𝑓 (𝑥) + 𝑐 ⋅ (𝑥𝑘+1 , … , 𝑥𝑁 ) + 𝑏 ⋅ (𝑥1 , … , 𝑥𝑘 ).

The function 𝑥 → 𝑓 (𝑥) + 𝑐 ⋅ (𝑥𝑘+1 , … , 𝑥𝑁 ) is smooth. Hence we can apply our previous result
on genericity of Morse functions on open subsets in ℝ𝑘 to this function and get that 𝑓𝑎 is a
Morse function for almost every 𝑏 ∈ ℝ𝑘 .
136 7.2. Morse Functions
Thus, for a fixed 𝑐, the subset of all 𝑏 ∈ ℝ𝑘 where 𝑓𝑎 is not Morse, has measure zero in
ℝ𝑘 . Hence 𝑆𝛼 ∩ (ℝ𝑘 × {0}) has measure zero in ℝ𝑁 . It is a classical result in Measure Theory,
called Fubini’s Theorem, which then implies that the set 𝑆𝛼 of all 𝑎 = (𝑏, 𝑐) where 𝑎 does not
yield a Morse function has measure zero in ℝ𝑁 . Hence 𝑓𝑎 is a Morse function for almost every
𝑎.

Finally, we can also show that being a Morse function is a stable property. In order to prove
stability, we start with a little lemma:

Lemma 7.14 Let 𝑓 be a smooth function on an open set 𝑈 ⊂ ℝ𝑘 . For each 𝑥 ∈ 𝑈 , let
𝐻(𝑓 )𝑥 be the Hessian matrix of 𝑓 at 𝑥. Then 𝑓 is a Morse function if and only if
( )2
∑
𝑘
𝜕𝑓
2
(det(𝐻(𝑓 )𝑥 )) + (𝑥) > 0 for all 𝑥 ∈ 𝑈 . (7.2)
𝑖=1
𝜕𝑥𝑖

𝜕𝑓 𝜕𝑓
Proof: A point 𝑥 is regular if 𝑑𝑓𝑥 = ( 𝜕𝑥 (𝑥), … , 𝜕𝑥 (𝑥)) ≠ 0, and 𝑥 is a nondegenerate
1 𝑘
critical point if 𝑑𝑓𝑥 = 0 and det(𝐻(𝑓 )𝑥 ) ≠ 0. Hence 𝑓 is Morse if and only if (7.2) is satisfied.

Lemma 7.15 Suppose that 𝑓𝑡 is a homotopic family of functions on ℝ𝑘 . If 𝑓0 is a

Morse function on some open subset 𝑈 ⊂ ℝ𝑘 containing a compact set 𝐾 ⊂ ℝ𝑘 , then
every 𝑓𝑡 for 𝑡 sufficiently small is a Morse function.

Proof: We define the map

( )2
∑
𝑘
𝜕𝑓
2
𝐹 ∶ 𝑈 × [0, 1] → ℝ, (𝑥, 𝑡) → (det(𝐻(𝑓𝑡 )𝑥 )) + (𝑥) .
𝑖=1
𝜕𝑥𝑖
Since 𝑓 is smooth, 𝐹 depends smoothly on both variables. By Lemma 7.14 and the assumption,
we know 𝐹 (𝑥, 0) > 0 for all 𝑥 ∈ 𝑈 × {0}. Since 𝐾 ⊂ 𝑈 is compact, 𝐹 has a minimum on
𝐾 × {0}, i.e. there is a 𝛿 > 0 such that 𝐹 (𝑥, 0) ≥ 2𝛿 for all 𝑥 ∈ 𝐾. Since 𝐹 is continuous,
there an open neighborhood 𝑊 ⊂ 𝑈 × [0, 1] containing 𝐾 × {0} such that 𝐹 (𝑥, 𝑡) > 𝛿 for all
(𝑥, 𝑡) ∈ 𝑊 . In fact, we can cover 𝐾 × {0} by open subsets 𝑊𝑖 ⊂ 𝑈 × [0, 1] such that 𝐹 (𝑥, 𝑡) > 𝛿
for all (𝑥, 𝑡) ∈ 𝑊𝑖 . Each such open subset 𝑊𝑖 has the form 𝑉𝑖 × [0, 𝜖𝑖 ) for some open 𝑉𝑖 ⊂ 𝑈
and 𝜖𝑖 > 0. Since 𝐾 is compact, finitely many such open 𝑊𝑖 suffice to cover 𝐾 × {0}. Let 𝜖
be the minimum of the finitely many 𝜖𝑖 . Then we have 𝐹 (𝑥, 𝑡) > 𝛿 for all (𝑥, 𝑡) ∈ 𝐾 × [0, 𝜖).
Since 𝐹 is continuous, for any fixed 𝑡 ∈ [0, 𝜖), there is again an open subset 𝑉 ℝ𝑘 containing 𝐾
such that 𝐹 (𝑥, 𝑡) > 0 for all (𝑥, 𝑡) ∈ 𝑉 × {𝑡}. Thus 𝑓𝑡 is Morse in a neighborhood of 𝐾 for all
sufficiently small 𝑡.

Finally, we can prove stability of Morse functions.

Theorem 7.16 (Stability of Morse functions) Let 𝑋 be a compact smooth manifold,

let 𝑓0 ∶ 𝑋 → ℝ be a smooth function and 𝑓𝑡 be a homotopy of 𝑓0 . If 𝑓0 is Morse, then
there is an 𝜀 > 0 such that 𝑓𝑡 is a Morse function for all 𝑡 ∈ [0, 𝜀).

Proof: For 𝑥 ∈ 𝑋, let 𝜙𝑥 ∶ 𝑈𝑥 → 𝑋 be a local parametrization around 𝑥. Then 𝑓0 ◦𝜙𝑥 is

Chapter 7. Sard’s theorem and Morse functions 137
a Morse function on 𝑈 . Since {0} is a compact subset of 𝑈 , Lemma 7.15 implies that there
is an open subset 𝑉𝑥 ⊂ 𝑈𝑥 containing {0} and an 𝜀(𝑥) > 0 such that 𝑓𝑡 is Morse on 𝑉𝑥 for all
𝑡 ∈ [0, 𝜀(𝑥)). The images 𝜙𝑥 (𝑉𝑥 ) are open in 𝑋 and cover 𝑋. Since 𝑋 is compact, finitely many
suffice to cover 𝑋, say

𝑋 = 𝜙𝑥1 (𝑉𝑥1 ) ∪ ⋯ ∪ 𝜙𝑥𝑛 (𝑉𝑥𝑛 )

Then we can set 𝜀 ∶= minimum of 𝜀(𝑥1 ), … , 𝜀(𝑥𝑛 ). Then 𝑓𝑡 ∶ 𝑋 → ℝ is a Morse function for
all 𝑡 ∈ [0, 𝜀).
8. Smooth Homotopy

In this chapter we are going to introduce one of the most important concepts in topology: homo-
topies between maps. The idea of studying objects up to homotopy has turned out be extremely
successful in many areas in mathematics.

8.1 Homotopies

We begin with a fundamental definition:

Definition 8.1 (Homotopy) Let 𝑋 and 𝑌 be two topological spaces and let 𝐼 = [0, 1]
denote the unit interval in ℝ. We say that two continuous maps 𝑓0 and 𝑓1 from 𝑋 to 𝑌
are homotopic, denoted 𝑓0 ∼ 𝑓1 , if there exists a continuous map

𝐹 ∶ 𝑋 × [0, 1] → 𝑌

such that 𝐹 (𝑥, 0) = 𝑓0 (𝑥) and 𝐹 (𝑥, 1) = 𝑓1 (𝑥). 𝐹 is called a homotopy between 𝑓0 and
𝑓1 . We also write 𝑓𝑡 (𝑥) for 𝐹 (𝑥, 𝑡). In other words, a homotopy is a family of continuous
functions 𝑓𝑡 which continuously interpolates between 𝑓0 and 𝑓1 .

Here are some first examples:

Example 8.2 Consider 𝑓0 ∶ ℝ → ℝ2 , 𝑥 → (𝑥, 0) and 𝑓1 ∶ ℝ → ℝ2 , 𝑥 → (𝑥, sin 𝑥)

with homotopy 𝐹 ∶ ℝ × [0, 1] → ℝ2 , (𝑥, 𝑡) → (𝑥, 𝑡 sin 𝑥). See Figure 8.1.

Figure 8.1: A family of maps which interpolates between a constant map and the sine function.

Example 8.3 Let 𝛾 ∶ [0, 1] → ℝ2 be a loop, i.e., a continuous map where start and end
points agree: 𝛾(0) = 𝛾(1). Then 𝛾 is homotopic to the constant map [0, 1] → {0} ⊂
ℝ2 . See Figure 8.2. In fact, this is true when we replace ℝ2 with any ℝ𝑘 , since ℝ𝑘 is

138
Chapter 8. Smooth Homotopy 139
contractible as we will show in the exercises.

Figure 8.2: Every loop in ℝ2 can be shrunk to a point, since ℝ2 is contractible.

Example 8.4 In Exercise 8.4 we will show that the antipodal map on the 𝑘-sphere
𝕊𝑘→ 𝕊𝑘 ,
𝑥 → −𝑥 (which sends a point to the point on the other side of the sphere) is
homotopic to the identity on 𝕊𝑘 if 𝑘 is odd. See Figure 8.3.

Figure 8.3: The antipodal map on the sphere may or may not be homotopic to the identity map.
This is a subtle phenomenon which we will study in more detail later.

Example 8.5 (Two maps which are not homotopic) The constant map 𝑓 ∶ 𝕊1 →
ℝ2 ⧵ {0}, 𝑝 → (1, 0) is not homotopic to the map 𝑔 ∶ 𝕊1 → ℝ2 ⧵ {0}, 𝑝 → 𝑝. We are
not yet able to prove this claim. It often turns out that showing that a homotopy cannot
exist is much harder to find a homotopy. We will later develop techniques and invariants
that will allow us to handle such problems. In particular, degrees and winding numbers
will help us.a
There are also other ways to prove the claim, for example using Brouwer’s fixed point theorem which
a

we will prove soon.

Homotopy is an equivalence relation on the set of continuous maps 𝑋 → 𝑌 . We will prove

this fact soon for a smooth refinement of homotopies.
140 8.2. Smooth homotopies and bump functions

Remark 8.6 (The homotopy category) Homotopy is one of the most crucial concepts
in topology. In fact, a lot of properties in topology are invariant under homotopy.
Therefore, they can be studied by considering maps only ‘up to homotopy’. This leads
to the construction of the homotopy category of spaces in which morphisms are ho-
motopy classes of maps, i.e., continuous maps 𝑓 and 𝑔 represent the same morphism if
and only if 𝑓 and 𝑔 are homotopic. Passing to the homotopy category is a very power-
ful idea which has had tremendous influence in many areas of mathematics. We will,
however, not be able to fully appreciate the homotopy category this semester. You will
meet homotopies also in algebraic topology, homological algebra, model categories, ∞-
categories, the theory of motives in algebraic geometry, new foundations of logic and
set theory and in many other areas of mathematics.

8.2 Smooth homotopies and bump functions

For the study of smooth manifolds, it is desirable to strengthen our assumptions on what a
homotopy is and to require it to be a smooth family of maps. We first present the necessary
definition and will then see that it actually is not such a strong restriction after all.

Definition 8.7 (Smooth homotopy) Let 𝑋 and 𝑌 be smooth manifolds. We say that
two smooth maps 𝑓0 and 𝑓1 from 𝑋 to 𝑌 are smoothly homotopic, denoted 𝑓0 ∼ 𝑓1 ,
if there exists a smooth map 𝐹 ∶ 𝑋 × [0, 1] → 𝑌 such that

𝐹 (𝑥, 0) = 𝑓0 (𝑥) and 𝐹 (𝑥, 1) = 𝑓1 (𝑥).

𝐹 is called a smooth homotopy between 𝑓0 and 𝑓1 . We also write 𝑓𝑡 (𝑥) for 𝐹 (𝑥, 𝑡). In
other words, a homotopy is a family of smooth functions 𝑓𝑡 which smoothly interpo-
lates between 𝑓0 and 𝑓1 .

∙ Recall that smoothness of 𝐹 in the second variable means that we can extend 𝐹 to a
smooth map 𝑋 × [−𝜀, 1 + 𝜀] → 𝑌 for some small 𝜀 > 0.

To allow only smooth homotopies might seem like a strong restriction. However, it turns
out that we can always approximate a continuous map by a smooth map within its homotopy
class:

Theorem 8.8 (Whitney Approximation Theorem) Let 𝑋 and 𝑌 be smooth manifolds.

∙ Let 𝑔 ∶ 𝑋 → 𝑌 be a continuous map. Then 𝑔 is homotopic to a smooth map

𝑓 ∶ 𝑋 → 𝑌 . Moreover, if 𝑔 was already smooth on a closed subset 𝑍 ⊂ 𝑋, then
one can choose 𝑓 such that 𝑓|𝐴 = 𝑔|𝐴 .

∙ Let 𝑔0 , 𝑔1 ∶ 𝑋 → 𝑌 be two smooth maps which are homotopic. Then they are
smoothly homotopic.
Chapter 8. Smooth Homotopy 141
We will prove this important result in Section 12.2. See Theorem 12.20 in particular. The
key tool we will use for the proof are tubular neighborhoods that we will introduce and study in
Section 12.1. For the moment, however, we rather move on and show that smooth homotopy is
an equivalence relation and will then see an application of homotopies.

In order to show that being smoothly homotopic defines an equivalence relation we need to
introduce a new tool: smooth bump functions. It will turn out to be extremely useful in many
applications later. It is a special feature of differential topology that we have the following type
of functions at our disposal:

∙ (Smooth bump functions) We define a smooth bump function in the following

steps:

∙ We start with the function

{ 2
𝑒−1∕𝑥 𝑥>0
𝑓 ∶ ℝ → ℝ, 𝑓 (𝑥) =
0 𝑥 ≤ 0.

∙ We observe that 𝑓 is smooth: We only need to think about 𝑥 ≥ 0. The 𝑖th

2
derivative has the form 𝑒−1∕𝑥 times a rational polynomial. Such a product is
differentiable and

lim 𝑓 (𝑖) (𝑥) = 0,

𝑥→0

2
since 𝑒−1∕𝑥 goes faster to 0 than any rational polynomial can go to ±∞.

∙ For given real numbers 𝑎 < 𝑏, we define a function

𝑔(𝑥) ∶= 𝑓 (𝑥 − 𝑎)𝑓 (𝑏 − 𝑥).

∙ As a product of two smooth functions, 𝑔 is smooth, and

⎧𝑔(𝑥) = 0 𝑥 ≤ 𝑎 since 𝑓 (𝑥 − 𝑎) = 0
⎪
⎨𝑔(𝑥) > 0 𝑎 < 𝑥 < 𝑏
⎪𝑔(𝑥) = 0 𝑥 ≥ 𝑏 since 𝑓 (𝑏 − 𝑥) = 0.
⎩

∙ Next we define yet another function

𝑥
∫−∞ 𝑔(𝑡)𝑑𝑡
ℎ ∶ ℝ → ℝ, ℎ(𝑥) ∶= ∞ .
∫−∞ 𝑔(𝑡)𝑑𝑡

By the Fundamental Theorem of Calculus, ℎ is smooth, nondecreasing, and

⎧ℎ(𝑥) = 0 𝑥≤𝑎
⎪
⎨0 < ℎ(𝑥) < 1 𝑎 < 𝑥 < 𝑏
⎪ℎ(𝑥) = 1 𝑥 ≥ 𝑏.
⎩

∙ Then ℎ is what we call a smooth bump function. See Figure 8.4.

142 8.2. Smooth homotopies and bump functions
∙ Finally, we can also define higher dimensional smooth bump functions by set-
ting
𝐻 ∶ ℝ𝑘 → ℝ, 𝐻(𝑥) ∶= 1 − ℎ(|𝑥|).
Then 𝐻(𝑥) is equal 1 on the closed ball with radius 𝑎 around the origin, is 0
outside the open ball with radius 𝑏, and between 0 and 1 on the intermediate
points:

⎧𝐻(𝑥) = 1 𝑥 ∈ 𝔹̄ 𝑎 (0)
⎪
⎨0 < 𝐻(𝑥) < 1 𝑎 < |𝑥| < 𝑏
⎪𝐻(𝑥) = 0 𝑥 ∈ ℝ𝑘 ⧵ 𝔹𝑏 (0).
⎩

Figure 8.4: A smooth bump function ℝ → ℝ. Its higher dimensional analogs are based on
this one. The function ℎ has value zero up to 𝑎 and value one from 𝑏 onwards. in between ℎ
increases smoothly.

Now we are ready to improve the following important fact.

Lemma 8.9 (Smooth homotopy is an equivalence relation) Let 𝑋 and 𝑌 be smooth

manifolds. Smooth homotopy defines an equivalence relation on the set of smooth
maps from 𝑋 to 𝑌 . The equivalence class to which a mapping belongs to is called its
homotopy class.

Proof: We need to check that ∼ is reflexive, symmetric, and transitive:

∙ Reflexivity: This is clear as every map is homotopic to itself via the homotopy 𝑓𝑡 = 𝑓
for all 𝑡.

∙ Symmetry: Suppose 𝑓 ∼ 𝑔 and let 𝐹 be a homotopy. Then the map defined by (𝑥, 𝑡) →
𝐹 (𝑥, 1 − 𝑡) is a homotopy from 𝑔 to 𝑓 . Hence 𝑔 ∼ 𝑓 as well.
Chapter 8. Smooth Homotopy 143
∙ Transitivity: Suppose 𝑓 ∼ 𝑔 and 𝑔 ∼ ℎ, and let 𝐹 be a homotopy from 𝑓 to 𝑔 and 𝐺
be a homotopy from 𝑔 to ℎ. We would like to compose 𝐹 and 𝐺 to get a homotopy from
𝑓 to ℎ. Since we require our homotopies to be smooth, we need to make sure that the
transition from 𝐹 to 𝐺 is smooth.
In order to this, we need to manipulate 𝐹 and 𝐺 a bit. And here we are lucky that we have
our smooth bump functions at our disposal. So let 𝜑 ∶ ℝ → ℝ be a smooth function
such that
{
0 𝑥 ≤ 1∕4
𝜑(𝑡) =
1 𝑥 ≥ 3∕4

and define new homotopies 𝐹̃ from 𝑓 to 𝑔 and 𝐻̃ from 𝑔 to ℎ by

𝐹̃ (𝑥, 𝑡) ∶= 𝐹 (𝑥, 𝜑(𝑡)) and 𝐺(𝑥,
̃ 𝑡) ∶= 𝐺(𝑥, 𝜑(𝑡)).

Now we can define the map

{
𝐹̃ (𝑥, 2𝑡) 𝑡 ∈ [0, 1∕2]
𝐻 ∶ 𝑋 × [0, 1] → 𝑌 , 𝐻(𝑥, 𝑡) =
̃ 2𝑡 − 1)
𝐺(𝑥, 𝑡 ∈ [1∕2, 1].

This map is well-defined and smooth, since 𝐹̃ (𝑥, 2𝑡) = 𝐺(𝑥,

̃ 2𝑡 − 1) for 𝑡 ∈ [3∕8, 5∕8].
Thus 𝐻 is a smooth homotopy from 𝑓 to ℎ. Hence ∼ is also transitive and an equivalence
relation.

8.3 Stable properties

We will now study the following question: Assume that 𝑓0 has a given property, for example let
us say that 𝑓0 is a submersion. If there is a smooth homotopy between 𝑓0 and 𝑓1 , do we know
that 𝑓1 also is a submersion?

In other words, we would like to know which properties of maps are or are not invariant
under homotopy.

In fact, many of the properties we have studied so far are not invariant, i.e., if 𝑓0 has a
property 𝑃 and 𝑓𝑡 is a homotopy from 𝑓0 to 𝑓1 , then it is often not true that 𝑓1 has property 𝑃 .
For example, we could start with an embedding 𝑓0 and end up with a constant map 𝑓1 .

So let us ask a more modest question: given 𝑓0 has property 𝑃 , is there always a small
𝜀 > 0 such that 𝑓𝑡 has property 𝑃 for all 𝑡 ∈ [0, 𝜖)? For example, if 𝑓0 is an embedding there
is always a small 𝜀 > 0 such that 𝑓𝑡 remains an embedding for 0 ≤ 𝑡 < 𝜀. We will say that
being an embedding is stable:

Definition 8.10 (Stable properties) A property 𝑃 is stable provided that whenever

𝑓0 ∶ 𝑋 → 𝑌 possesses the property and 𝑓𝑡 ∶ 𝑋 → 𝑌 is a homotopy of 𝑓0 then, for some
𝜀 > 0, each 𝑓𝑡 with 𝑡 < 𝜀 also possesses property 𝑃 .

∙ We also call the maps which have a stable property, a stable class. Examples are the
classes of embeddings, local diffeomorphisms, submersions,... as we will learn soon.
144 8.3. Stable properties
∙ Note that stability is a very natural condition to ask for: For real-world measurements,
only stable properties are interesting, since any tiny perturbation of the data would make
an unstable property appear or disappear.

In order to get a better idea of stability, let us look at the difference between requiring that
things merely intersect or that they intersect transversally:

∙ That a smooth map 𝑓0 ∶ ℝ → ℝ2 passes through a fixed point in ℝ2 is not a stable

property. It disappears immediately. See Figure 8.5.

Figure 8.5: Passing through a point is not stable.

∙ That a smooth map 𝑓0 ∶ ℝ → ℝ2 merely intersects the 𝑥-axis is not a stable property.
It disappears immediately. See Figure 8.6.

Figure 8.6: Mere intersection is not stable.

∙ However, that a smooth map 𝑓0 ∶ ℝ → ℝ2 intersects the 𝑥-axis transversely is a stable

property. It persists after a small perturbation. See Figure 8.7. This reveals yet another
very important feature of transversality.

Figure 8.7: Transverse intersection, however, is indeed stable.

∙ That two smooth curves (connected 1-dimensional manifolds) meet in ℝ3 is not a stable
property. It disappears immediately. See Figure 8.8.
Chapter 8. Smooth Homotopy 145

Figure 8.8: Intersection of curves in three-dimensional space is not stable.

∙ That a smooth curve and a smooth surface (2-dimensional manifold) intersect transver-
sally in ℝ3 is a stable property. It persists after a small perturbation. See Figure 8.9.

Figure 8.9: Intersection of a surface and a curve in three-dimensional space is not stable. The
dimensions have to add up.

The following theorem tells us that the properties which turned out to be useful for us so far
are all stable.

Theorem 8.11 (Stability Theorem) Let 𝑋 and 𝑌 be smooth manifolds. We assume

that 𝑋 is compact. The following classes of smooth maps from 𝑋 to 𝑌 are stable
classes:

(a) local diffeomorphisms,

(b) immersions,

(c) submersions,

(d) maps which are transversal to any fixed closed submanifold 𝑍 ⊂ 𝑌 ,

(e) embeddings,

(f) diffeomorphisms.

∙ Note that the assumption that 𝑋 is compact is crucial and not just made for convenience.
146 8.3. Stable properties
The next example will show that we cannot drop this assumption for any of the properties
in theorem:

Example 8.12 (Compactness matters) The Stability Theorem fails when 𝑋 is not
compact. For a simple example, let 𝜌 ∶ ℝ → ℝ be a smooth function with 𝜌(𝑠) = 1 for
|𝑠| < 1 and 𝜌(𝑠) = 0 for |𝑠| > 2. Then we define

𝑓𝑡 ∶ ℝ → ℝ, 𝑓𝑡 (𝑥) = 𝑥𝜌(𝑡𝑥).

For 𝑡 = 0, 𝑓0 (𝑥) = 𝑥 for all 𝑥, i.e., 𝑓0 = Id. Hence 𝑓0 is a local diffeomorphism, an

immersion, a submersion, an embedding, a diffeomorphism and transversal to every
submanifold of ℝ.
However, for any fixed 𝑡 > 0, we have |𝑡𝑥| > 2 when 𝑥 > 2∕|𝑡|. Hence, for this fixed
𝑡, 𝑓𝑡 (𝑥) = 0 for all 𝑥 > 2∕|𝑡|.
Thus 𝑓𝑡 is neither a local diffeomorphism, an immersion, a submersion, an embed-
ding, nor a diffeomorphism, and is not transversal to {0} ⊂ ℝ.

To emphasise what is going wrong, let us replace the domain with a closed interval,
say 𝑋 = [𝑎, 𝑏] with 𝑏 > 0, which is a compact subspace of ℝ. Then we can choose
𝜀 > 0 which is small enough such that 1∕𝜀 > max(|𝑎|, |𝑏|). This implies that 𝑥 is not
bigger than 1∕|𝑡|. Then we have 𝑓𝑡 (𝑥) = 𝑥 for all 𝑥 and all 𝑡 < 𝜀.

Proof of Theorem 8.11:

(a) First we note that local diffeomorphisms are just immersions in the special case when
dim 𝑋 = dim 𝑌 , so (a) follows from (𝑏).

(b) Assume 𝑓0 ∶ 𝑋 → 𝑌 is an immersion and dim 𝑋 = 𝑚. Let𝑓𝑡 be a homotopy of 𝑓0 .

That 𝑓0 is an immersion means that 𝑑(𝑓0 )𝑥 is injective for all 𝑥 ∈ 𝑋. We need to show that
there is an 𝜖 > 0 such that 𝑑(𝑓𝑡 )𝑥 is injective for all points (𝑥, 𝑡) in 𝑋 × [0, 𝜀) ⊂ 𝑋 × 𝐼:

Given a point 𝑥0 ∈ 𝑋, that 𝑑(𝑓0 )𝑥0 is injective implies that the matrix representing 𝑑(𝑓0 )𝑥0
(in local coordinates) has an 𝑚×𝑚-submatrix 𝐴(𝑥0 , 0) with nonvanishing determinant. Since
the determinant is continuous, this submatrix will have nonvanishing determinant in an
open neighborhood of (𝑥0 , 0) in 𝑋 × [0, 1]. Since 𝑋 is compact, finitely many such neighbor-
hoods suffice to cover all of 𝑋 × {0}. Hence there is a small 𝜀 > 0 (it is the minimum for the
open intervals [0, 𝜀𝑖 ) covering {0}) such that the intersection of these finitely many neighbor-
hoods contains 𝑋 × [0, 𝜀). Thus 𝑑(𝑓𝑡 )𝑥 is injective for all (𝑥, 𝑡) ∈ 𝑋 × [0, 𝜀). This is what we
needed.

(c) If 𝑓0 is a submersion, almost the same argument works. We just need to choose an
𝑛 × 𝑛-submatrix of the surjective map 𝑑(𝑓0 )𝑥 with 𝑛 = dim 𝑌 .

(d) Let 𝑍 ⊂ 𝑌 be a closed submanifold, and assume that 𝑓0 is a map which is transversal
to 𝑍. Then we have shown that, for every point 𝑥 ∈ 𝑋, there is a smooth function 𝑔 which
sends a neighborhood of 𝑓 (𝑥) to 0 ∈ ℝcodim 𝑍 and such that 𝑔◦𝑓0 is a submersion. Since 𝑍 is
closed in 𝑌 , 𝑓 −1 (𝑍) is closed in 𝑋 and therefore also compact. Therefore, by (c), there is an
𝜀 > 0 such that 𝑔◦𝑓𝑡 is still a submersion for all 𝑡 < 𝜀. This is means that 𝑓𝑡 is still transversal
to 𝑍 for all 𝑡 < 𝜀.
Chapter 8. Smooth Homotopy 147
(e) Assume that 𝑓0 is an embedding, and let 𝑓𝑡 be a homotopy of 𝑓0 . Since 𝑋 is compact,
𝑓0 and each 𝑓𝑡 are automatically proper maps. Hence we need to show that when 𝑓0 is a
one-to-one immersion, then so is 𝑓𝑡 in a small neighborhood. We just checked that being an
immersion is stable. Hence it remains to show that 𝑓𝑡 is still one-to-one if 𝑡 is small enough.

Therefor we define a smooth map

𝐺 ∶ 𝑋 × 𝐼 → 𝑌 × 𝐼, 𝐺(𝑥, 𝑡) ∶= (𝑓𝑡 (𝑥), 𝑡).

Then if (e) was false, i.e., if 𝑓𝑡 was not one-to-one in some small neighborhood of 0, then, for
every 𝜀 > 0, we can find a 𝑡 with 0 < 𝑡 < 𝜀 and 𝑥, 𝑦 ∈ 𝑋 such that 𝑓𝑡 (𝑥) = 𝑓𝑡 (𝑦). For example,
for every 𝜀𝑖 = 1∕𝑖, we could find such a 𝑡𝑖 , 𝑥𝑖 and 𝑦𝑖 . Thus there is an infinite sequence 𝑡𝑖 → 0,
and an infinite sequence of points 𝑥𝑖 ≠ 𝑦𝑖 ∈ 𝑋 where 𝑓𝑡𝑖 fails to be injective, i.e., such that

𝑓𝑡𝑖 (𝑥𝑖 ) = 𝐺(𝑥𝑖 , 𝑡𝑖 ) = 𝐺(𝑦𝑖 , 𝑡𝑖 ) = 𝑓𝑡𝑖 (𝑦𝑖 ).

Since 𝑋 is compact, we may pass to subsequences which converge 𝑥𝑖 → 𝑥0 and 𝑦𝑖 → 𝑦0 .

Since 𝐺 is continuous, this implies

𝐺(𝑥0 , 0) = lim 𝐺(𝑥𝑖 , 𝑡𝑖 ) = lim 𝐺(𝑦𝑖 , 𝑡𝑖 ) = 𝐺(𝑦0 , 0).

𝑖 𝑖

But 𝐺(𝑥0 , 0) = 𝑓0 (𝑥0 ) and 𝐺(𝑦0 , 0) = 𝑓0 (𝑥0 ). By assumption, 𝑓0 is injective, and hence
𝑥0 = 𝑦0 .

Now, after choosing local coordinates, we can express the derivative of 𝐺 at (𝑥0 , 0) by
the matrix

⎛ ∗⎞
⎜𝑑(𝑓 ) ⋮⎟
= ⎜ 0 𝑥0
∗⎟⎟
𝑑𝐺(𝑥0 ,0)
⎜
⎝ 0⋯0 1⎠

where the 0s in the lowest row arise from the fact that the first coordinates do not depend on
𝑡, and the 1 in the bottom right hand corner is the derivative of the function 𝑡 → 𝑡.

Since 𝑓0 is an immersion, 𝑑(𝑓0 )𝑥0 has 𝑘 = dim 𝑋 many independent rows. Thus the
matrix of 𝑑𝐺(𝑥0 ,0) has 𝑘 + 1 independent rows, and hence 𝑑𝐺(𝑥0 ,0) is an injective linear map.
In other words, 𝐺 is an immersion around (𝑥0 , 0) and hence 𝐺 must be one-to-one on some
neighborhood of (𝑥0 , 0).

But we have shown above that the sequences (𝑥𝑖 , 𝑡𝑖 ) and (𝑦𝑖 , 𝑡𝑖 ) both converge to (𝑥0 , 0).
This means that for large 𝑖 both (𝑥𝑖 , 𝑡𝑖 ) and (𝑦𝑖 , 𝑡𝑖 ) belong to this neighborhood. This contra-
dicts the injectivity of 𝐺.

(f) Assume that 𝑓0 ∶ 𝑋 → 𝑌 is a diffeomorphism. Since 𝑋 is compact, this implies that

𝑌 is compact as well. Let 𝑓𝑡 be a homotopy of 𝑓0 . We need to show that there is an 𝜀 > 0 such
that 𝑓𝑡 is diffeomorphism for all 0 ≤ 𝑡 < 𝜀.

Since 𝑋 is compact, 𝑋 has only finitely many connected components, and so does 𝑌 .
Hence we can check the statement for each of these connected components separately. For, this
gives us an 𝜀𝑖 for each component. Since there are finitely many components, we can just take
the minimum of the 𝜖𝑖 ’s as the 𝜖 for all of 𝑋 and 𝑌 .
148 8.3. Stable properties
Thus we may assume that 𝑋 and 𝑌 are connected. By (a) and (e), we know that being a
local diffeomorphism and being an embedding is a stable property. Thus there is a 𝜀 > 0 such
that 𝑓𝑡 is a local diffeomorphism and an embedding. For 𝑓𝑡 being a diffeomorphism, it remains
to show that 𝑓𝑡 is surjective.

We fix a 0 < 𝑡 < 𝜀. Since 𝑓𝑡 is a local diffeomorphism, 𝑓𝑡 is an open map and hence
𝑓𝑡 (𝑋) is open in 𝑌 . But 𝑓𝑡 (𝑋) is also closed, since it is compact being the image of a compact
space. Since 𝑌 is connected, this implies 𝑓𝑡 (𝑋) = 𝑌 .
Chapter 8. Smooth Homotopy 149
8.4 Exercises and more examples

Exercise 8.1 A manifold 𝑋 is called contractible if its identity map is homotopic to

some constant map 𝑋 → {𝑥} where 𝑥 is any point of 𝑋.

(a) Show that if 𝑋 is contractible, then all smooth maps 𝑌 → 𝑋 from an arbitrary
manifold 𝑌 into 𝑋 are homotopic.

(b) Conversely, show that if all maps of an arbitrary manifold 𝑌 into 𝑋 are homotopic,
then 𝑋 is contractible.

(c) Show that ℝ𝑘 is contractible.

Exercise 8.2 Let 𝑋 be a smooth manifold of dimension 𝑘. Show that every smooth
map 𝑓 ∶ 𝑋 → 𝕊𝑛 is homotopic to a constant map if 𝑘 < 𝑛.
Hint: Use Sard’s Theorem.

Exercise 8.3 Show that the antipodal map 𝕊𝑘 → 𝕊𝑘 , 𝑥 → −𝑥, is homotopic to the
identity if 𝑘 is odd. (We will see later that this is not true if 𝑛 is even.)
Hint: Start off with 𝑘 = 1 by using the linear maps defined by
( )
cos(𝜋𝑡) − sin(𝜋𝑡)
[0, 1] → 𝑀(2), 𝑡 → .
sin(𝜋𝑡) cos(𝜋𝑡)

Exercise 8.4 A manifold 𝑋 is called simply-connected if it is connected and if every

smooth map from the circle 𝕊1 into 𝑋 is homotopic to a constant map. Show that all
contractible manifolds are simply connected.
Note that the converse is false. As an example consider 𝑋 = 𝕊2 . The 2-sphere
is simply connected, but it is not contractible. For example, the antipodal map is not
homotopic to the identity. We will have to develop further techniques to be able to check
this.

Exercise 8.5 Show that every connected smooth manifold 𝑋 is path-connected, i.e.,
given any two points 𝑥0 , 𝑥1 ∈ 𝑋, there exists a smooth map 𝑓 ∶ [0, 1] → 𝑋 with 𝑓 (0) =
𝑥0 and 𝑓 (1) = 𝑥1 .
Hint: Use the fact that homotopy is an equivalence relation to show that the relation
𝑥0 and 𝑥1 can be joined by a smooth curve is an equivalence relation on 𝑋. Then show
that the equivalence classes are both open and closed subsets of 𝑋.

Exercise 8.6 Prove that the sphere 𝕊𝑘 is simply connected if 𝑘 > 1.

Hint: If 𝑓 ∶ 𝕊1 → 𝕊𝑘 and 𝑘 > 1, Sard’s Theorem gives us a point 𝑝 ∉ 𝑓 (𝕊1 ). Now
use stereographic projection.
150 8.4. Exercises and more examples

Exercise 8.7 Let 𝑋 be a smooth manifold. Let 𝑓 , 𝑔 ∶ 𝑋 → 𝕊𝑘 be two continuous maps

such that |𝑓 (𝑥) − 𝑔(𝑥)| < 2 for all 𝑥 ∈ 𝑋, where the norm is taken as elements in ℝ𝑘+1 .

(a) Show that 𝑓 and 𝑔 are homotopic.

Hint: What does the assumption say geometrically about 𝑓 (𝑥) and 𝑔(𝑥)?

(b) Show that 𝑓 and 𝑔 are smooth, then they are smoothly homotopic.

Exercise 8.8 Let 𝑋 ⊂ ℝ𝑁 be a smooth manifold. A vector field on 𝑋 is a smooth

section of 𝜋 ∶ 𝑇 (𝑋) → 𝑋, i.e., a smooth map 𝜎 ∶ 𝑋 → 𝑇 (𝑋) such that 𝜋◦𝜎 = Id𝑋 .
An equivalent way to describe such a section is to give a map 𝑠 ∶ 𝑋 → ℝ𝑁 such that
𝑠(𝑥) ∈ 𝑇𝑥 (𝑋) ⊂ ℝ𝑁 for all 𝑥 (with corresponding 𝜎(𝑥) = (𝑥, 𝑠(𝑥))). A point 𝑥 ∈ 𝑋 is a
zero of the vector field 𝜎 if 𝜎(𝑥) = (𝑥, 0) or equivalently 𝑠(𝑥) = 0.

(a) Show that if 𝑘 is odd, there exists a vector field on 𝕊𝑘 having no zeros.
Hint: For 𝑘 = 1, use (𝑥1 , 𝑥2 ) → (−𝑥2 , 𝑥1 ).

(b) Prove that if 𝕊𝑘 has a vector field which has no zeros, then its antipodal map 𝑥 →
−𝑥 is homotopic to the identity.
Hint: Show that you may assume |𝑠(𝑥)| = 1 everywhere. Now contemplate about
(cos(𝜋𝑡))𝑥 + (sin(𝜋𝑡))𝑠(𝑥) when 𝑡 varies from 0 to 1.

(c) Show that if 𝑘 is even, then the antipodal map on 𝕊𝑘 is homotopic to the reflection
map
𝑟 ∶ 𝕊𝑘 → 𝕊𝑘 , (𝑥1 , … , 𝑥𝑘+1 ) → (−𝑥1 , 𝑥2 , … , 𝑥𝑘+1 ).
Hint: Consider also the reflections

𝑟𝑖 (𝑥1 , … , 𝑥𝑘+1 ) = (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑘+1 ).

Show that 𝑟𝑖 ◦𝑟𝑖+1 is homotopic to the identity on 𝕊𝑘 .

9. Abstract Smooth Manifolds

9.1 Abstract manifolds - the definition

We would like to define manifolds without referring to a given embedding into some ℝ𝑁 . The
key idea that should be preserved in any new definition is that a manifold is a space which
locally looks like Euclidean space. First, we recall an important concept from topology:

Definition 9.1 (Hausdorff spaces) A topological space 𝑋 is called Hausdorff if, for
any two distinct points 𝑥, 𝑦 ∈ 𝑋, there are two disjoint open subsets 𝑈 , 𝑉 ⊂ 𝑋 such
that 𝑥 ∈ 𝑈 and 𝑦 ∈ 𝑉 . In other words, in a Hausdorff space we can separate points by
open neighborhoods.

∙ Every subspace of ℝ𝑁 with the relative topology is a Hausdorff space. However, there
are spaces which are not Hausdorff.

∙ For a typical example of a space which is not Hausdorff, consider two copies of the real
line 𝑌1 ∶= ℝ × {1} and 𝑌2 ∶= ℝ × {2} as subspaces of ℝ2 . On 𝑌1 ∪ 𝑌2 , we define the
equivalence relation (𝑥, 1) ∼ (𝑥, 2) for all 𝑥 ≠ 0. Let 𝑋 be the set of equivalence classes.
In fact, 𝑋 looks like the real line except that the origin is replaced with two different
copies of the origin:

Figure 9.1: A space that looks like a line with a double-point. The topology is such that we
cannot separate the two points by open subsets.

The topology on 𝑋 is the quotient topology: a subset 𝑊 ⊂ 𝑋 is open in 𝑋 if and only

if both its preimages in ℝ × {1} and ℝ × {2} are open. Away from the double origin,
𝑋 looks perfectly nice like a one-dimensional manifold. But every neighborhood of one
of the origins contains the other. Hence we cannot separate the two origins by open
subsets, and 𝑋 is not Hausdorff.

151
152 9.1. Abstract manifolds - the definition
In the definition of an abstract manifold, we want to avoid such pathological spaces. There
are several reasons for this choice. First of all, manifolds are characterised by how open neigh-
borhoods of points look like, and we would like to be able to find separate neighborhoods for
distinct points. That is exactly the Hausdorff property. But there are also slightly deeper rea-
sons. For example, we would like to use the fact that a compact subset 𝑍 of a closed subset
𝑌 ⊂ 𝑋 is itself closed in 𝑋. This general conclusion requires that 𝑋 is a Hausdorff space.

Definition 9.2 (Charts) Let 𝑋 be a topological space. A chart on 𝑋 is a pair (𝑉 , 𝜓)

where 𝑉 ⊂ 𝑋 is an open subset and 𝜓 ∶ 𝑉 → 𝑈 is a homeomorphism from 𝑉 to an
open subset 𝑈 ⊂ ℝ𝑘 .

Now we can define manifolds:

Definition 9.3 (Abstract manifolds) An abstract smooth 𝑘-manifold is a Hausdorff

and second-countable space 𝑋, i.e., there exists a countable basis of the topology,
together with a collection of charts (𝑉𝛼 , 𝜓𝛼 ) on 𝑋 such that

∙ every point in 𝑋 is in the domain of some chart, and

∙ for every pair of overlapping charts 𝜓𝛼 and 𝜓𝛽 , i.e.,

𝑉𝛼𝛽 ∶= 𝑉𝛼 ∩ 𝑉𝛽 ≠ ∅,

the change-of-coordinates map

𝜓𝛽 ◦𝜓𝛼−1 ∶ 𝜓𝛼 (𝑉𝛼𝛽 ) → 𝜓𝛽 (𝑉𝛼𝛽 )

is smooth as a map between open subsets of ℝ𝑘 . In fact, this means that the
change-of-coordinates maps are diffeomorphisms, since they are mutual smooth
inverses to each other.

Definition 9.4 (Smooth maps between abstract manifolds) Let 𝑋 be an abstract

smooth 𝑘-manifold.

∙ A continuous map 𝑓 ∶ 𝑋 → ℝ𝑛 is called smooth if for every chart 𝜓𝛼 ∶ 𝑉𝛼 → 𝑈𝛼 ,

the composition
𝑓 ◦𝜓𝛼−1 ∶ 𝑈𝛼 → ℝ𝑛
is smooth as a map from an open subset of ℝ𝑘 to ℝ𝑛 .

∙ More generally, let 𝑌 an abstract smooth 𝑚-manifold and 𝑓 ∶ 𝑋 → 𝑌 a contin-

uous map. Then 𝑓 is smooth at 𝑥 ∈ 𝑋 if, for every chart 𝜓 𝑋 ∶ 𝑉 → 𝑈 on 𝑋
around 𝑥 and every chart 𝜓 𝑌 ∶ 𝑉 ′ → 𝑈 ′ on 𝑌 around 𝑓 (𝑥), the map

𝜓 𝑌 ◦𝑓|𝑉 ∩𝑓 −1 (𝑉 ′ ) ◦(𝜓 𝑋 )−1

|𝑈 ∩𝜓 𝑋 (𝑉 ∩𝑓 −1 (𝑉 ′ ))
∶ 𝑈 ∩ 𝜓 𝑋 (𝑉 ∩ 𝑓 −1 (𝑉 ′ )) → 𝑈 ′
Chapter 9. Abstract Smooth Manifolds 153

is a smooth map as a map from an open subset of ℝ𝑘 to an open subset of ℝ𝑚 . We

call 𝑓 smooth if it is smooth at every 𝑥 ∈ 𝑋.

Figure 9.2: On a manifold every point is contained in the domain of a chart. If two charts
overlap we can look at the composition of the chart maps and get a map between open subsets
in ℝ𝑘 . We call this the change-of-coordinate map and require this map to be smooth in the
usual sense.

Note that the smooth 𝑘-dimensional manifolds 𝑋 ⊂ ℝ𝑁 we have been studying so far are
examples of abstract smooth 𝑘-manifolds:

∙ The Hausdorff property is satisfied in ℝ𝑁 and therefore also for every subspace of ℝ𝑁
with the relative subspace topology.

∙ Moreover, every open cover {𝑈𝛼 } of ℝ𝑁 has a countable refinement. For, we can take
the collection of all open balls which are contained in some 𝑈𝛼 , which have rational
radii, and which are centred at points having only rational coordinates.

∙ For an open cover {𝑉𝛼 } of a subset 𝑋 ⊂ ℝ𝑁 , we can write 𝑉𝛼 = 𝑈𝛼 ∩ 𝑋 for some open
subsets 𝑈𝛼 of ℝ𝑁 . Then let {𝑈̃ 𝑖 } be a countable refinement of {𝑈𝛼 } in 𝑅𝑁 , and define
𝑉̃𝑖 = 𝑈̃ 𝑖 ∩ 𝑋.

∙ The charts are just what we called local coordinates and the inverses of charts are what
we called local parametrizations. One difference is that we required local parametriza-
tions to be diffeomorphisms. For an abstract manifold 𝑋, we use the charts to define
what smoothness means for a map on 𝑋. Hence a priori it makes only sense to talk
about the smoothness of the change of coordinate maps. A posteriori we can then check
that charts are in fact diffeomorphisms.

∙ Similarly for smooth maps between manifolds. We only know what smoothness of maps
between Euclidean spaces means. Hence we need to use the charts to first translate the
maps into maps between Euclidean spaces.
154 9.1. Abstract manifolds - the definition
∙ In the abstract definition, we take care of the fact that the images of the charts/local
parametrizations overlap. In fact, we use the overlap to define the smooth structure.
∙ Finally, a chosen collection of charts is called an atlas on the manifold. One can show that
every manifold has a maximal atlas, i.e., the images of the charts are as big as possible.

Luckily, our initial definition fits nicely into this picture:

Lemma 9.5 (Our smooth manifolds are also abstract manifolds) Let 𝑋 be a smooth
manifold according to our initial definition. Then 𝑋 is also an abstract smooth mani-
fold.

Proof: Subspaces in ℝ𝑁 are Hausdorff. By our definition of a smooth manifold, we can

cover 𝑋 by the open subsets associated to local parametrizations. Since 𝑋 ⊂ ℝ𝑁 , it suffices to
use countable many such open sets to cover 𝑋. It remains to show that 𝑋 has charts.

Let 𝑥 ∈ 𝑋 be a point in 𝑋 and let 𝜙1 ∶ 𝑈1 → 𝑉1 and 𝜙2 ∶ 𝑈2 → 𝑉2 be local parametrizations

around 𝑥 with open subsets 𝑈1 , 𝑈2 ⊂ ℝ𝑘 and 𝑉1 , 𝑉2 ⊂ 𝑋. The overlap of 𝑉1 and 𝑉2 is not
empty, since it contains 𝑥. Moreover, 𝑉1 ∩ 𝑉2 is an open subset of 𝑋. Since both 𝜙1 and 𝜙2 are
homeomorphisms, 𝜙−1 1
(𝑉1 ∩ 𝑉2 ) ⊂ 𝑈1 and 𝜙−1
2
(𝑉1 ∩ 𝑉2 ) ⊂ 𝑈2 are both open subsets in ℝ𝑘 .
Since both 𝜙1 and 𝜙2 are diffeomorphisms, the transition map
𝜙−1
2
◦𝜙1 ∶ 𝜙−1
1
(𝑉1 ∩ 𝑉2 ) → 𝜙−1
2
(𝑉1 ∩ 𝑉2 )
is a diffeomorphism between open subsets in ℝ𝑘 . Thus our local parametrizations (or rather
our local coordinate systems) do equip 𝑋 with the structure of an abstract smooth manifold.

Remark 9.6 (It all fits together) It is nice and important to have such an intrinsic
definition of a manifold. However, the definition is quite abstract indeed. And, in
fact, we are going to show that every abstract smooth manifold can be embedded
into Euclidean space and is therefore a manifold for our previous definition. Hence all
the machinery we have developed can be applied to abstract manifolds.

Abstract smooth manifolds also have tangent spaces:

Tangent space of an abstract manifold

There are several different, though equivalent, ways to define the tangent space of an abstract
manifold. We will look at just one way which is closest to our intuitive and concrete approach.
In addition, we have seen it in earlier exercises.

Let 𝑋 be an abstract smooth manifold of dimension 𝑘 and let 𝑥 ∈ 𝑋 be a point. We consider

the set of all smooth curves through 𝑥 on 𝑋, i.e., the set of all smooth maps 𝛾 ∶ ℝ → 𝑋 with
𝛾(0) = 𝑥. On this set we define the following equivalence relation:

∙ Let 𝛾1 ∶ ℝ → 𝑋 and 𝛾2 ∶ ℝ → 𝑋 be two smooth curves with 𝛾1 (0) = 𝑥 = 𝛾2 (0). We

consider 𝛾1 and 𝛾2 to be equivalent, written 𝛾1 ∼ 𝛾2 , if for all charts 𝜓 ∶ 𝑉 → 𝑈 with
𝑥 ∈ 𝑉 and 𝑈 ⊂ ℝ𝑘 we have an equality of derivatives at 0
(𝜓◦𝛾1 )′ (0) = (𝜓◦𝛾2 )′ (0)
Chapter 9. Abstract Smooth Manifolds 155
where the derivative is taken as maps ℝ → ℝ𝑘 . Hence (𝜓◦𝛾1 )′ (0) = (𝜓◦𝛾2 )′ (0) is just
an equality of vectors in ℝ𝑘 .

Definition 9.7 (Abstract tangent spaces) The tangent space of 𝑋 at 𝑥, denoted

𝑇𝑥 (𝑋), is defined as the set of all equivalence classes of curves through 𝑥.

∙ Actually, it is not necessary that the curves 𝛾 are defined on all of ℝ. It suffices that
there is an open subset 𝑊 ⊂ ℝ containing 0 and 𝛾 ∶ 𝑊 → 𝑋 is a smooth map de-
fined on 𝑊 with 𝛾(0) = 𝑥. In a more sophisticated terminology, it suffices to consider
the set of germs of curves through 𝑥. Then we consider the set of all such curves and
the equivalence relation only requires that the derivatives of the composite with any chart
are equal. Anyway, we see that 𝑇𝑥 (𝑋) only depends on the local structure of 𝑋 at 𝑥.

9.2 Real projective space

This is an important example which we can easily be described with the new definition of an
abstract manifold, but for which it is not obvious at all how we can embed it into ℝ𝑁 .

Actually, it is a difficult question how to embed these guys into ℝ𝑁 with 𝑁 as small as
𝑟 𝑟
possible. In fact, if 𝑛 = 2𝑟 for some 𝑚 and if there is an immersion ℝP2 → ℝ2 +𝑘 , then 𝑘
must be at least 2𝑟 − 1. You will learn about the techniques to show this in the more advanced
algebraic topology course and can read about it in [14].

Definition 9.8 (Real Projective Space) The real projective 𝑛-space ℝP𝑛 is the set of
all straight lines through the origin in ℝ𝑛+1 . As a topological space, ℝP𝑛 is the quotient
space
ℝP𝑛 = (ℝ𝑛+1 ⧵ {0})∕ ∼
where the equivalence relation is given by 𝑥 ∼ 𝑦 if there is a nonzero real number 𝜆
such that 𝑥 = 𝜆𝑦. This means that a subset 𝑉 is open in ℝP𝑛 if and only if its preimage
𝑈 = {𝑥 ∈ ℝ𝑛+1 ⧵ {0} ∶ [𝑥] ∈ 𝑉 } is open in ℝ𝑛+1 ⧵ {0}.

∙ Note that since each line through the origin intersects the unit sphere in two (antipodal)
points, ℝP𝑛 can alternatively be described as

𝕊𝑛 ∕ ∼

where the equivalence relation is 𝑥 ∼ −𝑥. As a quotient of 𝕊𝑛 , we see that ℝP𝑛 is actually
compact. As we have seen, this is always very good to know about a space.
156 9.2. Real projective space

Theorem 9.9 ℝP𝑛 is a smooth manifold] Real projective 𝑛-space ℝP𝑛 is an abstract
𝑛-dimensional smooth manifold.

Proof: If 𝑥 = (𝑥0 , … , 𝑥𝑛 ) ∈ ℝ𝑛+1 ⧵ {0}, we write [𝑥] for its equivalence class considered
as a point in ℝP𝑛 . One also often writes [𝑥] = [𝑥0 ∶ … ∶ 𝑥𝑛 ].

For 0 ≤ 𝑖 ≤ 𝑛, let
𝑉𝑖 ∶= {[𝑥] ∈ ℝP𝑛 ∶ 𝑥𝑖 ≠ 0}.

The preimage of 𝑉𝑖 in ℝ𝑛+1 is the open subset {𝑥 ∈ ℝ𝑛+1 ∶ 𝑥𝑖 ≠ 0}. Hence each 𝑉𝑖 is open in
ℝP𝑛 . By varying 𝑖, this gives an open cover of ℝP𝑛 because every representative (𝑥0 , … , 𝑥𝑛 )
of a point [𝑥] ∈ ℝP𝑛 must have at least one coordinate ≠ 0 (otherwise it would be the origin
which is excluded).

For each 𝑖, we have the maps 𝜙𝑖 ∶ ℝ𝑛 → 𝑉𝑖

(𝑥0 , … , 𝑥̂𝑖 , … , 𝑥𝑛 ) → [𝑥0 ∶ … ∶ 𝑥𝑖−1 ∶ 1 ∶ 𝑥𝑖+1 ∶ … ∶ 𝑥𝑛 ].

and 𝜙−1
𝑖 ∶ 𝑉𝑖 → ℝ
𝑛

1
[𝑥0 ∶ … ∶ 𝑥𝑖 ∶ … ∶ 𝑥𝑛 ] → (𝑥 , … , 𝑥̂𝑖 , … , 𝑥𝑛 )
𝑥𝑖 0

1
where 𝑥̂𝑖 means that 𝑥𝑖 is omitted. Note that the quotient 𝑥𝑖
is well-defined on all of 𝑉𝑖 .

𝑖 , we need to
Since we use a representative of an equivalence class for the definition of 𝜙−1
check that the definition is independent of the chosen representative: If [𝑥0 ∶ … ∶ 𝑥𝑖 ∶ … ∶
𝑥𝑛 ] = [𝜆𝑥0 ∶ … ∶ 𝜆𝑥𝑖 ∶ … ∶ 𝜆𝑥𝑛 ] for some 𝜆 ≠ 0, then

1
𝜙−1
𝑖 ([𝜆𝑥]) = (𝜆𝑥0 , … , 𝜆𝑥𝑖−1 , 𝜆𝑥𝑖+1 , … , 𝜆𝑥𝑛 )
𝜆𝑥𝑖
1
= (𝑥0 , … , 𝑥𝑖−1 , 𝑥𝑖+1 , … , 𝑥𝑛 ) = 𝜙−1
𝑘 ([𝑥]).
𝑥𝑖

It is easy to see that 𝜙𝑖 and 𝜙−1

𝑖 are mutual inverses which are both continuous.

Finally, the change-of-coordinate maps are smooth: For the composite

𝜙𝑖 𝜙−1
𝑗
𝜙−1
𝑖 (𝑉𝑖 ∩ 𝑉𝑗 ) ←←←←→ ← 𝜙−1
← 𝑉𝑖 ∩ 𝑉𝑗 ←←←←←←→ 𝑗 (𝑉𝑖 ∩ 𝑉𝑗 )

is just

1
(𝑥0 , … , 𝑥̂𝑖 , … , 𝑥𝑛 ) → (𝑥 , … , 𝑥𝑖−1 , 1, 𝑥𝑖+1 , … , 𝑥̂𝑗 , … , 𝑥𝑛 )
𝑥𝑗 0

which is smooth whenever 𝑥𝑗 ≠ 0.1

1
Our notation seems to indicate that 𝑗 must be bigger than 𝑖. But this is of course not the case. Both cases 𝑖 < 𝑗
and 𝑖 > 𝑗 work out in the same way.
Chapter 9. Abstract Smooth Manifolds 157

Remark 9.10 (Tangent space of ℝP𝑛 ) Intuitively, we can think of the tangent space of
ℝP𝑛 at a point [𝑥] ∈ ℝP𝑛 as follows: We can consider the point [𝑥] as a pair of antipodal
points (𝑥, −𝑥) in 𝕊𝑛 . Then a tangent vector at [𝑥] may be viewed as a pair of ’antipodal
vectors’ (𝑣, −𝑣) where 𝑣 ∈ 𝑇𝑥 𝕊𝑛 and −𝑣 ∈ 𝑇−𝑥 𝕊𝑛 . To write −𝑣 may be justified by
identifying both 𝑇𝑥 𝕊𝑛 and 𝑇−𝑥 𝕊𝑛 with the subspace in ℝ𝑛+1 which is orthogonal to the
line 𝐿 through 𝑥 and −𝑥:

𝑇[𝑥] ℝP𝑛 = {(𝑣, −𝑣) ∶ 𝑣 ∈ 𝐿⟂ = 𝑇𝑥 𝕊𝑛 ⊂ ℝ𝑛+1 }. (9.1)

In Remark 9.20, and in the arguments leading up to it, we will see that there is a more
precise and canonical description of the tangent space of ℝP𝑛 at a point [𝑥] ∈ ℝP𝑛 as

𝑇[𝑥] ℝP𝑛 = Homℝ (𝐿, 𝐿⟂ ) (9.2)

where 𝐿 ⊂ ℝ𝑛+1 is the line through the origin determined by [𝑥] and 𝐿⟂ denotes the
orthogonal complement of 𝐿 in ℝ𝑛+1 .

There are many reasons why real projective space is important. One is that it comes equipped
with a very useful additional structure:

Remark 9.11 (Canonical line bundle) We discussed vector bundles briefly in Sec-
tion 2.5.3. One class of vector bundles are line bundles, i.e., each vector space at a point
is one-dimensional. Real projective 𝑛-space has a canonical line bundle, often also
called the tautological line bundle. A point in ℝP𝑛 consists of a line 𝐿 in ℝ𝑛+1 through
the origin. Considering 𝐿 both as a point in ℝP𝑛 and as a one-dimensional vector space
defines a line bundle on ℝP𝑛 .
It turns out that for every sufficiently nice topological space 𝑌 with a line bundle
 → 𝑌 , there is a continuous map 𝑌 → ℝP𝑛 (for 𝑛 large enough) such that  is the
pullback of the canonical line bundle along this map. Up to homotopy, this map is in
fact unique. We refer to this phenomenon as that ℝP𝑛 is a classifying space for line
bundles. The idea is classifying spaces is very powerful and important.

Another, more geometric reason for considering projective spaces is the following:

∙ Intersection at ∞:

A plane 𝑉 in ℝ3 can be described as the orthogonal complement of given vector 𝑣 ≠ 0 in

ℝ3 :

𝑉 = {(𝑥0 , 𝑥1 , 𝑥2 ) ∈ ℝ3 ∶ 𝑥0 𝑣0 + 𝑥1 𝑣1 + 𝑥2 𝑣2 = 0}.

Since multiplying the equation 𝑥0 𝑣0 + 𝑥1 𝑣1 + 𝑥2 𝑣2 = 0 with a nonzero real number does

not change the set of solutions, we can consider the equivalence classes in ℝP2 of the points
of 𝑉 . This gives us a line  in ℝP2 :

 = {[𝑥0 ∶ 𝑥1 ∶ 𝑥2 ] ∈ ℝP2 ∶ 𝑥0 𝑣0 + 𝑥1 𝑣1 + 𝑥2 𝑣2 = 0}.

158 9.3. Further examples
In fact, every line in ℝP2 is represented by a plane through the origin in ℝ3 and is hence
determined by a nonzero vector 𝑣 in ℝ3 . Moreover, 𝑣 and 𝜆𝑣 with 𝜆 ≠ 0 determine the same
line.

Now assume we are given two distinct lines 1 and 2 in ℝP2 determined by two distinct
vectors 𝑣, 𝑤 ≠ 0 in ℝ3 , i.e.,

1 = {[𝑥0 ∶ 𝑥1 ∶ 𝑥2 ] ∈ ℝP2 ∶ 𝑥0 𝑣0 + 𝑥1 𝑣1 + 𝑥2 𝑣2 = 0}
2 = {[𝑥0 ∶ 𝑥1 ∶ 𝑥2 ] ∈ ℝP2 ∶ 𝑥0 𝑤0 + 𝑥1 𝑤1 + 𝑥2 𝑤2 = 0}.

The orthogonal complements of 𝑣 and 𝑤, respectively, are two planes through the origin.
Hence they meet in a line through the origin in ℝ3 which is the set of solutions of the two
linear equations defining 1 and 2 above. This is a one-dimensional vector subspace of ℝ3
(the kernel of a 2 × 3-matrix). By definition of ℝP2 , this subspace corresponds to a point in
ℝP2 . This is the intersection point of 1 and 2 in ℝP2 .

If this intersection line happens to be the 𝑧-axis, i.e., when 1 and 2 are represented by the
planes given by the 𝑥𝑧-plane and the 𝑦𝑧-plane, then the intersection point is [0 ∶ 0 ∶ 1] ∈ ℝP2 .
We can think of it as the point at infinity in ℝP2 .

However, in the Euclidean plane ℝ2 it may very well happen that two lines are paral-
lel and hence do not intersect. The idea for ℝP2 is to add a point at infinity which is the
intersection point for all parallel lines.

9.3 Further examples

9.3.1 The torus as an abstract manifold

We already know the torus as an important example of a two-dimensional compact manifold.

So far we have constructed it via coordinates in ℝ3 or as a product 𝕊1 ×𝕊1 in ℝ4 . Here is another
way to construct a two-dimensional torus 𝕋 2 without referring to an ambient space:

On the product ℝ × ℝ we consider the relation

(𝑥1 , 𝑦1 ) ∼ (𝑥2 , 𝑦2 ) ⇐⇒ 𝑥1 − 𝑥2 ∈ ℤ and 𝑦1 − 𝑦2 ∈ ℤ.

We readily verify that this is an equivalence relation. W equip the quotient space with the
quotient topology and can check that it is just the torus
ℝ×ℝ
≅ 𝕊1 × 𝕊1 = 𝕋 2 .
∼

However, let us have a closer look because we would like to use this picture to equip 𝕋 with
the structure of a manifold:

Claim: The quotient map

ℝ×ℝ
𝑞∶ ℝ × ℝ →
∼
is an open map, i.e., 𝑞 sends open subsets to open subsets.
Chapter 9. Abstract Smooth Manifolds 159

Figure 9.3: The torus 𝕋 2 can be constructed by gluing together the opposite edges of a square.

Let 𝑈 ⊂ ℝ × ℝ be open. By definition of the quotient topology, the image 𝑉 = 𝑞(𝑈 ) is

open in 𝕋 if and only if 𝑞 −1 (𝑉 ) = 𝑞 −1 (𝑞(𝑈 )) is open in ℝ × ℝ. The latter subset is
⋃
𝑞 −1 (𝑉 ) = 𝑞 −1 (𝑞(𝑈 )) = 𝑈(𝑛,𝑚) with 𝑈(𝑛,𝑚) = {(𝑥, 𝑦) ∈ ℝ × ℝ ∶ (𝑥 − 𝑛, 𝑦 − 𝑚) ∈ 𝑈 }.
𝑛,𝑚∈ℤ

This is a union of open subsets, each homeomorphic to 𝑈 , and hence open in ℝ × ℝ.

Now we can check that 𝕋 = ℝ×ℝ ∼

is a Hausdorff space. Moreover, it has a countable basis
for its topology. Let us just accept these facts. For we are much more interested in the charts
that turn 𝕋 2 into a manifold. In fact, defining these charts is almost trivial:

Let (𝑥, 𝑦) ∈ ℝ × ℝ and let [𝑥, 𝑦] be its equivalence class in 𝕋 . Let 𝑈𝑟 = 𝐵𝑟2 (𝑥, 𝑦) ⊂ ℝ × ℝ
be the open ball of radius 𝑟 > 0 around (𝑥, 𝑦). Then the map

ℝ×ℝ
𝑞|𝑈𝑟 ∶ 𝑈𝑟 →
∼

is a homeomorphism onto its image for all 𝑟 < 12 . For it is surjective and it is injective, since
for any two points in 𝑈𝑟 the difference of the 𝑥- or 𝑦-coordinates, respectively, is strictly less
than 2𝑟 < 1 by the choice of 𝑟. Moreover, the inverse is continuous, since we checked that 𝑞 is
an open map. We use 𝑞|𝑈𝑟 as a chart around the point [𝑥, 𝑦]. The change-of-coordinate maps
are of the form (or almost, we need to restrict the maps accordingly, but the notation would
become too annoying, so we simplify):

𝐵𝑟1 (𝑥1 , 𝑦1 ) → 𝑇 → 𝐵𝑟2 (𝑥2 , 𝑦2 ), (𝑥, 𝑦) → (𝑥 + (𝑥2 − 𝑥1 ), 𝑦 + (𝑦2 − 𝑦1 )).

This is a linear map and hence smooth.

In summary: We have shown that the quotient space 𝕋 = ℝ×ℝ ∼

is an abstract smooth mani-
fold. And if we ignore or accept the topological general nonsense in the beginning, this was a
very simple way to show that the torus is a manifold.

9.3.2 The Klein bottle

Another famous and slightly more complicated space is the Klein bottle defined as follows:
This time we consider the relation on ℝ × ℝ defined by

(𝑥1 , 𝑦1 ) ∼ (𝑥2 , 𝑦2 ) ⇐⇒ 𝑥1 + 𝑥2 ∈ ℤ and 𝑦1 − 𝑦2 ∈ ℤ.

160 9.3. Further examples

Figure 9.4: The Klein bottle can be constructed by gluing together the opposite edges of a
square, but we twist one pair of the two edges.

This is again an equivalence relation and the quotient map is open.

The quotient space 𝐾 = ℝ×ℝ ∼

is a Hausdorff space and is the famous so-called Klein
bottle. It is a two-dimensional space which we cannot embed into ℝ3 . However, one can draw
its shadows in three-dimensional space. We see that it may not be the best strategy to first embed
𝐾 into an ℝ𝑁 . However, to show that 𝐾 is an abstract smooth manifold is not so difficult.
The charts for 𝐾 are as simple as for the torus: Again, let (𝑥, 𝑦) ∈ ℝ × ℝ and let [𝑥, 𝑦] be its
equivalence class in 𝐾. Let 𝑈𝑟 = 𝐵𝑟2 (𝑥, 𝑦) ⊂ ℝ × ℝ be the open ball of radius 𝑟 > 0 around
(𝑥, 𝑦). Then the restriction of the quotient map

𝑞|𝑈𝑟 ∶ 𝑈𝑟 → 𝐾

is a homeomorphism onto its image for all 𝑟 < 21 for the same reason as for the torus. We use
𝑞|𝑈𝑟 as a chart around the point [𝑥, 𝑦]. The change-of-coordinate maps are of the form (again
we cheat a bit here):

𝐵𝑟1 (𝑥1 , 𝑦1 ) → 𝐾 → 𝐵𝑟2 (𝑥2 , 𝑦2 ), (𝑥, 𝑦) → (−𝑥 + (𝑥2 + 𝑥1 ), 𝑦 + (𝑦2 − 𝑦1 )).

This is a linear map and hence smooth.

Remark 9.12 In the exercises we study Hopf manifolds as another important and
interesting class of examples. They play a crucial role in complex geometry, since they
provide the simplest examples of compact complex manifolds which do not admit a
Kähler metric and therefore cannot be embedded into complex projective space.

9.3.3 Stiefel manifolds

Definition 9.13 (Stiefel manifold) A 𝑘-frame in ℝ𝑛+𝑘 is a 𝑘-tuple [𝑣1 , … , 𝑣𝑘 ] of

orthonormal vectors in ℝ𝑛+𝑘 . Define the Stiefel manifold 𝑉𝑘 (ℝ𝑛+𝑘 ) as the subset

𝑉𝑘 (ℝ𝑛+𝑘 ) = {𝑘-frames in ℝ𝑛+𝑘 } ⊂ ℝ(𝑛+𝑘)𝑘 .

Chapter 9. Abstract Smooth Manifolds 161
Note that we have already met a Stiefel manifold before, since 𝑉1 (ℝ𝑛+1 ) may be identified
with 𝕊𝑛 .

The topology on 𝑉𝑘 (ℝ𝑛+𝑘 ) is given as follows: We consider 𝑉𝑘 (ℝ𝑛+𝑘 ) as the subspace of

𝕊𝑛+𝑘−1 × … × 𝕊𝑛+𝑘−1 of 𝑘 copies of spheres 𝕊𝑛+𝑘−1 given by all orthonormal 𝑘-tuples and
equip 𝑉𝑘 (ℝ𝑛+𝑘 ) with the subspace topology. It is a closed subspace since orthogonality of two
vectors can be expressed by an algebraic equation. In particular, 𝑉𝑘 (ℝ𝑛+𝑘 ) is compact, since
the product of spheres is compact (and closed subspaces of compact spaces are compact). In
fact, we can show that 𝑉𝑘 (ℝ𝑛+𝑘 ) actually deserves the name manifold:

Theorem 9.14 (Stiefel manifolds are manifolds) The space 𝑉𝑘 (ℝ𝑛+𝑘 ) of 𝑘-frames is
𝑘(𝑘−1)
a compact smooth manifold of dimension 𝑛𝑘 + 2
.

Remark 9.15 We observe that 𝑉𝑘 (ℝ𝑛+𝑘 ) can be identified with the quotient 𝑂(𝑛 +
𝑘)∕𝑂(𝑛). For, if [𝑣1 , … , 𝑣𝑘 ] and 𝑇 ∈ 𝑂(𝑛 + 𝑘), then multiplying each 𝑣𝑖 with 𝑇 yields
another 𝑘-frame [𝑇 𝑣1 , … , 𝑇 𝑣𝑘 ]. In fact, any two 𝑘-frames are connected in this way,
i.e., if [𝑣1 , … , 𝑣𝑘 ] and [𝑣′1 , … , 𝑣′𝑘 ] are two 𝑘-frames then there is an 𝑇 ∈ 𝑂(𝑛 + 𝑘) such
that
[𝑣′1 , … , 𝑣′𝑘 ] = [𝑇 𝑣1 , … , 𝑇 𝑣𝑘 ].
In other words, 𝑂(𝑛+𝑘) acts transitively on the set of 𝑘-frames. Moreover, the stabilizer
subgroup of a given frame is the subgroup isomorphic to 𝑂(𝑛) which acts nontrivially
on the orthogonal complement of the space spanned by that frame.

Proof of Theorem 9.14: We have already proven that 𝑉𝑘 (ℝ𝑛+𝑘 ) is compact. It remains
to show the manifold part. Note that a 𝑘-frame [𝑣1 , … , 𝑣𝑘 ] corresponds to an (𝑛 + 𝑘) × 𝑘-
matrix 𝐴 with 𝑣𝑖 as 𝑖th column. That the column vectors are orthonormal is equivalent to
that the product of the transpose of the 𝑖th column with the 𝑗th column is 1 if 𝑖 = 𝑗 and 0
otherwise. In a formula: {
1 if 𝑖 = 𝑗
𝑣𝑡𝑖 ⋅ 𝑣𝑗 =
0 if 𝑖 ≠ 𝑗.
In other words,
𝐴 represents a 𝑘-frame in ℝ𝑛+𝑘 ⇐⇒ 𝐴𝑡 𝐴 = 𝐼𝑘 ,
where 𝐼𝑘 denotes the 𝑘 × 𝑘-identity matrix.

Let 𝑀(𝑛 + 𝑘, 𝑘) be the space of (𝑛 + 𝑘) × 𝑘-matrices and 𝑆(𝑘) denote the space of symmetric
𝑘 × 𝑘-matrices. We define the map
𝑓 ∶ 𝑀(𝑛 + 𝑘, 𝑘) → 𝑆(𝑘), 𝐴 → 𝐴𝑡 𝐴
and we observe 𝑉𝑘 (ℝ𝑛+𝑘 ) = 𝑓 −1 (𝐼𝑘 ).

Hence, in order to show that 𝑉𝑘 (ℝ𝑛+𝑘 ) is a smooth manifold, we need to show that 𝐼𝑘 is a
regular value for 𝑓 . We have computed the derivative of 𝑓 at a matrix 𝐴 ∈ 𝑉𝑘 (ℝ𝑛+𝑘 ) in the
proof of Theorem 4.12. There we showed
𝑑𝑓𝐴 (𝐵) = 𝐴𝑡 𝐵 + 𝐵 𝑡 𝐴.
162 9.4. Grassmannian
In order to check that 𝐼𝑘 is a regular value, we need to show that

𝑑𝑓𝐴 ∶ 𝑇𝐴 (𝑀(𝑛 + 𝑘, 𝑘)) = 𝑀(𝑛 + 𝑘, 𝑘) → 𝑇𝑓 (𝐴) (𝑆(𝑘)) = 𝑆(𝑘)

is surjective for all 𝐴 ∈ 𝑉𝑘 (ℝ𝑛+𝑘 ). Recall for the above computation that 𝑀(𝑛 + 𝑘, 𝑘) ≅ 𝑅(𝑛+𝑘)𝑘
and 𝑆(𝑘) ≅ ℝ𝑘(𝑘+1)∕2 are vector spaces and that the tangent space of a vector space equals the
vector space.

For 𝐶 ∈ 𝑆(𝑘), we set 𝐵 = 12 𝐴𝐶 ∈ 𝑀(𝑛 + 𝑘, 𝑘) and get

( ) ( )𝑡
1 1 1 1 1 1
𝑑𝑓𝐴 (𝐵) = 𝐴𝑡 𝐴𝐶 + 𝐴𝐶 𝐴 = 𝐴𝑡 𝐴𝐶 + 𝐶 𝑡 𝐴𝑡 𝐴 = 𝐶 + 𝐶 𝑡 = 𝐶.
2 2 2 2 2 2

By the Preimage Theorem 4.6, this shows that 𝑉𝑘 (ℝ𝑛+𝑘 ) is a smooth manifold of dimen-
sion

dim 𝑉𝑘 (ℝ𝑛+𝑘 ) = dim 𝑀(𝑛 + 𝑘, 𝑘) − dim 𝑆(𝑘)

𝑘(𝑘 + 1) 𝑘(𝑘 − 1)
= (𝑛 + 𝑘)𝑘 − = 𝑛𝑘 + .
2 2

9.4 Grassmannian

There is another very important space that arises from Stiefel manifolds. Any 𝑘-frame in ℝ𝑛+𝑘
spans a 𝑘-dimensional linear subspace in ℝ𝑛+𝑘 .

Definition 9.16 (Grassmannian) The set of all 𝑘-dimensional linear subspaces in ℝ𝑛+𝑘
is called the Grassmann manifold, or short the Grassmannian,

Gr 𝑘 (ℝ𝑛+𝑘 ) = {𝑘-dimensional linear subspaces in ℝ𝑛+𝑘 }.

The Grassmannian Gr 𝑘 (ℝ𝑛+𝑘 ) can be identified with the quotient of the Stiefel manifold
𝑉𝑘 (ℝ𝑛+𝑘 ) of orthonormal sequences
[𝑣1 , … , 𝑣𝑘 ]

of vectors 𝑣𝑖 ∈ ℝ𝑛+𝑘 , modulo the equivalence relation

[𝑣′1 , … , 𝑣′𝑘 ] ∼ [𝑣1 , … , 𝑣𝑘 ] ⋅ 𝑇 ,

with 𝑇 any orthogonal 𝑘 × 𝑘-matrix.

In other words, Gr 𝑘 (ℝ𝑛+𝑘 ) is the quotient of 𝑉𝑘 (ℝ𝑛+𝑘 ) that we get by identifying 𝑘-frames
which span the same subspace in ℝ𝑛+𝑘 :

𝑞 ∶ 𝑉𝑘 (ℝ𝑛+𝑘 ) → Gr 𝑘 (ℝ𝑛+𝑘 ), [𝑣1 , … , 𝑣𝑘 ] → span(𝑣1 , … , 𝑣𝑘 ) in ℝ𝑛+𝑘 .

We have already seen an example of a Grassmannian:

Chapter 9. Abstract Smooth Manifolds 163

Example 9.17 (Projective spaces are Grassmannians) For 𝑘 = 1, Gr 1 (ℝ𝑛+1 ) is the

space of lines, i.e., one-dimensional subspaces, in ℝ𝑛+1 . Thus we have

Gr 1 (ℝ𝑛+1 ) = ℝP𝑛

the real projective space of dimension 𝑛.

It is actually not so easy to embed projective spaces and Grassmannians into Euclidean
spaces. For the moment, we content ourselves with noting that Gr 𝑘 (ℝ𝑛+𝑘 ) can be embedded
at least in ℝ2𝑛𝑘+1 . However, it is an interesting and difficult question what the minimal
dimension 𝑁 such that we can embed them into ℝ𝑁 . We will get back to this point in a general
context in Section 9.5.

There are many reasons why Grassmannians are important. We will now sketch one
of them. If it does not make complete sense to you yet, consider it as a advert for a future
adventure.

9.4.1 The canonical bundle

We mentioned vector bundles briefly in Section 2.5. The tangent bundle on a smooth manifold is
an important example. The Grassmannian is equipped with a canonical or tautological vector
bundle, denoted 𝛾𝑘𝑛+𝑘 . It is defined as follows: A point in Gr 𝑘 (ℝ𝑛+𝑘 ) consists of a 𝑘-plane 𝑉 in
ℝ𝑛+1 . We can consider 𝑉 both as a point in Gr 𝑘 (ℝ𝑛+𝑘 ) and as a 𝑘-dimensional vector space,
and we can pick a vector 𝑣 ∈ 𝑉 . The space 𝛾𝑘𝑛+𝑘 consists of the collection of such pairs, i.e.,

𝛾𝑘𝑛+𝑘 = {(𝑉 , 𝑣) ∶ 𝑉 ∈ Gr 𝑘 (ℝ𝑛+𝑘 ), 𝑣 ∈ 𝑉 }.

Together with the projection 𝜋 ∶ 𝛾𝑘𝑛+𝑘 → Gr 𝑘 (ℝ𝑛+𝑘 ) defined by sending (𝑉 , 𝑣) to 𝑉 turns 𝛾𝑘𝑛+𝑘
into a a vector bundle on Gr 𝑘 (ℝ𝑛+𝑘 ).

This vector bundle is extremely useful. Let 𝑋 ⊂ ℝ𝑁 be a smooth manifold of dimension

𝑘. Let 𝑛 be a natural number such that 𝑁 = 𝑛 + 𝑘. Let 𝑥 ∈ 𝑋 be a point and 𝑇𝑥 𝑋 be the
tangent space at 𝑥. This is a 𝑘-dimensional vector subspace of ℝ𝑛+𝑘 . Hence we can consider
𝑇𝑥 𝑋 as an element in Gr 𝑘 (ℝ𝑛+𝑘 ). This defines a map 𝑓 ∶ 𝑋 → Gr 𝑘 (ℝ𝑛+𝑘 ). Now recall that the
tangent bundle 𝑇 𝑋 consists of pairs (𝑥, 𝑣) with 𝑥 ∈ 𝑋 and 𝑣 ∈ 𝑇𝑥 𝑋. Sending the paris (𝑥, 𝑣)
to the pair (𝑇𝑥 𝑋, 𝑣) defines a map 𝑓̄ ∶ 𝑇 𝑋 → 𝛾𝑘𝑛+𝑘 . We can show, assuming the results we will
prove in this section below, that both these maps 𝑓 and 𝑓̄ are smooth. Moreover, they fit into a
commutative diagram
𝑓̄
𝑇𝑋 / 𝛾 𝑛+𝑘
𝑘

 
𝑋 / Gr (ℝ𝑛+𝑘 ).
𝑓 𝑘

We can show that this actually defines a morphism of vector bundles. This diagram is often
called a generalized Gauss map. The maps 𝑓 and 𝑓̄ are very useful for studying manifolds.
More importantly, this picture actually has a fascinating generalization:
164 9.4. Grassmannian

Theorem 9.18 (Grassmannians classify vector bundles) Let 𝑌 be a sufficiently nice

topological space 𝑌 with a 𝑘-dimensional vector bundle 𝐸 → 𝑌 . Then, for 𝑛 large
enough, there is a continuous map 𝑓 ∶ 𝑌 → Gr 𝑘 (ℝ𝑛+𝑘 ) such that 𝐸 is the pullback of
the canonical vector bundle along this map:

𝑓̄
𝐸 ≅ 𝑓 ∗𝛾 /𝛾

 
𝑌 / Gr (ℝ𝑛+𝑘 ).
𝑓 𝑘

Up to homotopy, the map 𝑓 is in fact unique for the isomorphism class of 𝐸 → 𝑌 .

We refer to this phenomenon by saying that Gr 𝑘 (ℝ𝑛+𝑘 ) is a classifying space for 𝑘-
dimensional vector bundles. The concept of classifying spaces is very powerful and we
strongly encourage to learn more about it in the future.

9.4.2 Grassmannians are manifolds - geometric proof

Now we show the following key fact:

Theorem 9.19 (Grassmannians are manifolds) The Grassmannian Gr 𝑘 (ℝ𝑛+𝑘 ) is a

compact smooth manifold of dimension 𝑘 ⋅ 𝑛.

∙ Note that we will be cheating a bit for the moment. For, we define a map

𝑓 ∶ Gr 𝑘 (ℝ𝑛+𝑘 ) → ℝ𝑁

to be smooth if and only if the composite 𝑓 ◦𝑞 ∶ 𝑉𝑘 (ℝ𝑛+𝑘 ) → ℝ𝑁 is smooth. We will

get back to this point later. Recall that in order to shorten our sentences we will often
refer to a 𝑘-dimensional linear subspace as a 𝑘-plane.

Proof: Since the quotient map 𝑞 ∶ 𝑉𝑘 (ℝ𝑛+𝑘 ) → Gr 𝑘 (ℝ𝑛+𝑘 ) is continuous and surjective
and since 𝑉𝑘 (ℝ𝑛+𝑘 ) is compact, we see that Gr 𝑘 (ℝ𝑛+𝑘 ) is compact. It remains to show that it is
a smooth manifold.

Let 𝑉 ⊂ ℝ𝑛+𝑘 be a 𝑘-dimensional linear subspace in ℝ𝑛+𝑘 , and let 𝑉 ⟂ be the orthogonal
complement of 𝑉 in ℝ𝑛+𝑘 (with respect to the standard inner product). Define the subspace
𝑉 ⊂ Gr 𝑘 (ℝ𝑛+𝑘 ) consisting of 𝑘-dimensional linear subspaces 𝑉 ′ of ℝ𝑛+𝑘 with the property
that 𝑉 ′ ∩ 𝑉 ⟂ = {0}:

𝑉 = {𝑉 ′ ∈ Gr 𝑘 (ℝ𝑛+𝑘 ) ∶ 𝑉 ′ ∩ 𝑉 ⟂ = {0}}.

Equivalently, 𝑉 is the set of all 𝑘-dimensional subspaces 𝑉 ′ ⊂ ℝ𝑛+𝑘 which are mapped sur-
jectively onto 𝑉 by the projection 𝑝 ∶ ℝ𝑛+𝑘 = 𝑉 ⊕ 𝑉 ⟂ → 𝑉 .
Chapter 9. Abstract Smooth Manifolds 165
We will use subsets of the form 𝑉 to construct local parametrizations. To do this we need
to check several things:

First claim: 𝑈 is an open neighborhood of 𝑉 .

To see that 𝑉 is open Gr 𝑘 (ℝ𝑛+𝑘 ) it suffices to show that its preimage ̃ 𝑉 under 𝑞 is open in
𝑉𝑘 (ℝ𝑛+𝑘 ), since we defined the topology on the Grassmannian as the quotient topology induced
by
𝑞 ∶ 𝑉𝑘 (ℝ𝑛+𝑘 ) → Gr 𝑘 (ℝ𝑛+𝑘 ), [𝑣1 , … , 𝑣𝑘 ] → span(𝑣1 , … , 𝑣𝑘 ) in ℝ𝑛+𝑘 .
By definition of 𝑞 and 𝑉 , the set ̃ 𝑉 consists of all orthonormal 𝑘-frames [𝑣′1 , … , 𝑣′𝑘 ] such
that

span(𝑣′1 , … , 𝑣′𝑘 ) ∩ 𝑉 ⟂ = {0}. (9.3)

Let {𝑤1 , … , 𝑤𝑛 } be a basis of 𝑊 = 𝑉 ⟂ . Then condition (9.3) is equivalent to saying

that the set of vectors {𝑣′1 , … , 𝑣′𝑘 , 𝑤1 , … , 𝑤𝑛 } forms a basis of ℝ𝑛+𝑘 , since they are linearly
independent and span an (𝑛+𝑘)-dimensional subspace of ℝ𝑛+𝑘 . Equivalently, the (𝑛+𝑘)×(𝑛+
𝑘)-matrix 𝑀(𝑣′1 , … , 𝑣′𝑘 , 𝑤1 , … , 𝑤𝑛 ) with column vectors 𝑣′1 , … , 𝑣′𝑘 , 𝑤1 , … , 𝑤𝑛 has nonzero
determinant. Keeping the basis of 𝑊 fixed, we see that ̃ 𝑉 consists of the preimage of the
open subset ℝ ⧵ {0} under the map

det ∶ 𝑉𝑘 (ℝ𝑛+𝑘 ) → ℝ, det(𝑀(𝑣′1 , … , 𝑣′𝑘 , 𝑤1 , … , 𝑤𝑛 )).

Hence ̃ 𝑉 is an open subset. This proves the first claim.

Now we are going to define a map from 𝑉 to Euclidean space.

We can think of each 𝑉 ′ ∈ 𝑉 as the graph of a linear map 𝑉 → 𝑊 = 𝑉 ⟂ :

′ ′
Given 𝑉 ′ ∈ 𝑉 , let pr 𝑉𝑉 be the orthogonal projection of 𝑉 ′ onto 𝑉 and pr 𝑉𝑊 be the
orthogonal projection of 𝑉 ′ onto 𝑊 . Note that since 𝑉 ′ ∩ 𝑉 ⟂ = {0} and dim 𝑉 ′ = dim 𝑉 ,
′
we know that pr 𝑉𝑉 is an isomorphism. Hence we can define a map

′
𝜓𝑉 ∶ 𝑉 → Homℝ (𝑉 , 𝑊 ), 𝑉 ′ → 𝜓𝑉 (𝑉 ′ ) = pr 𝑊
𝑉′
◦(pr 𝑉𝑉 )−1
from 𝑉 to Homℝ (𝑉 , 𝑊 ). See Figure 9.5.

The space Homℝ (𝑉 , 𝑊 ) is a real vector space isomorphic to ℝ𝑛𝑘 .

We can define an inverse to 𝜓𝑉 as well:

Given a linear map 𝑓 ∶ 𝑉 → 𝑊 , the graph of 𝑓 ,

Γ(𝑓 ) = {𝑣 + 𝑓 (𝑣) ∈ 𝑉 ⊕ 𝑊 } ⊂ 𝑉 ⊕ 𝑊 ≅ ℝ𝑛+𝑘 ,

is a linear subspace of dimension 𝑘 which satisfies

Γ(𝑓 ) ∩ 𝑊 = 𝑉 ∩ 𝑊 = {0}.

Hence Γ(𝑓 ) ∈ 𝑉 and we define the inverse by

𝜓𝑉−1 ∶ Homℝ (𝑉 , 𝑊 ) → 𝑉 , 𝑓 → Γ(𝑓 ).

166 9.4. Grassmannian

Figure 9.5: A vector 𝑣′ in 𝑉 ′ defines a linear transformation from 𝑉 to 𝑊 by taking orthogonal

projections.

Second claim: 𝜓𝑉 and 𝜓𝑉−1 are mutual inverses.

First, note that Γ(𝜓𝑉 (𝑉 ′ )) = 𝑉 ′ by definition of 𝜓𝑉 (𝑉 ′ ) by considering 𝑉 ′ as the graph of

a map 𝑉 → 𝑊 . Now let 𝑓 ∶ 𝑉 → 𝑊 be a linear map. Then
𝜓𝑉 (Γ(𝑓 ))(𝑣) = pr Γ(𝑓
𝑊
)
◦(pr Γ(𝑓
𝑉
) −1
) (𝑣) = pr Γ(𝑓
𝑊
)
(𝑣 + 𝑓 (𝑣)) = 𝑓 (𝑣).
Hence 𝜓𝑉 (Γ(𝑓 )) = 𝑓 . This proves the second claim.

Third claim: 𝜓𝑉 is a diffeomorphism.

To prove the claim it remains to show that 𝜓𝑉 and 𝜓𝑉−1 are smooth. Note that we consider
Homℝ (𝑉 , 𝑊 ) as a topological space by identifying it with ℝ𝑛𝑘 . That is we fix an orthonormal
basis 𝑣1 , … , 𝑣𝑘 for 𝑉 and an orthonormal basis 𝑤1 , … , 𝑤𝑛 for 𝑊 = 𝑉 ⟂ . Then we identify
linear maps 𝑓 ∶ 𝑉 → 𝑊 with their associated matrix 𝐴𝑓 with respect to these bases. The
entries in 𝐴𝑓 are real numbers and we can collect them to 𝑛 ⋅ 𝑘-tuples, i.e., elements in ℝ𝑛⋅𝑘 .

We show first that 𝑇 is smooth: By definition of the topology on the Grassmannian and our
definition of smoothness on Gr 𝑘 (ℝ𝑛+𝑘 ), 𝑇 is smooth if and only if the composite
𝑞
̃ 𝑉 = 𝑞 −1 (𝑉 ) /
𝑉

𝜓𝑉
𝜓𝑉 ◦𝑞
( 
Homℝ (𝑉 , 𝑊 )

is smooth. Let [𝑣′1 , … , 𝑣′𝑛 ] be an orthonormal 𝑘-frame which spans 𝑉 ′ . The entries in the matri-
′
ces of the maps pr 𝑊𝑉′
and (pr 𝑉𝑉 )−1 , respectively, expressed in the bases of 𝑉 , 𝑊 and 𝑉 ′ , depend
′
smoothly on the coordinates of the 𝑣′𝑖 . Hence matrix of the map 𝜓𝑉 (𝑉 ′ ) = pr 𝑊 𝑉′
◦(pr 𝑉𝑉 )−1
depends smoothly on the 𝑘-frame [𝑣′1 , … , 𝑣′𝑛 ] and 𝜓𝑉 ◦𝑞 is smooth.
Chapter 9. Abstract Smooth Manifolds 167
Now let 𝑓 ∶ 𝑉 → 𝑊 be a linear map and let 𝐴𝑓 be its representing matrix (in the bases we
chose). Then there is a unique orthonormal basis 𝑣′1 , … , 𝑣′𝑛 of Γ(𝑓 ) with
𝑣′𝑖 = 𝑣𝑖 + 𝑓 (𝑣𝑖 ) = 𝑣𝑖 + 𝐴𝑓 (𝑣𝑖 )
for all 𝑖. Hence the coordinates of each 𝑣′𝑖 depend smoothly on the entries in 𝐴𝑓 . Hence the
𝑘-frame [𝑣′1 , … , 𝑣′𝑛 ] in 𝑉𝑘 (ℝ𝑛+𝑘 ) depends smoothly on 𝑓 . Since the 𝑘-frame [𝑣′1 , … , 𝑣′𝑛 ] spans
Γ(𝑓 ) = 𝜓𝑉−1 (𝑓 ), we can interpret this as Γ(𝑓 ) depends smoothly on 𝑓 and 𝜓𝑉−1 is also smooth.
This proves the third claim.

In total we have shown that the point 𝑉 ∈ Gr 𝑘 (ℝ𝑛+𝑘 ) has an open neighborhood 𝑉 which
is diffeomorphic to ℝ𝑛𝑘 . In other words, Gr 𝑘 (ℝ𝑛+𝑘 ) is indeed a smooth manifold of dimension
𝑛 ⋅ 𝑘.

Remark 9.20 (Tangent space of the Grassmannian) The local coordinate chart 𝜓𝑉
gives us also a way to understand the tangent space of Gr 𝑘 (ℝ𝑛+𝑘 ) at 𝑉 . The 𝜓𝑉 behaves
like a linear map (if we keep vectors small enough) and the image Homℝ (𝑉 , 𝑉 ⟂ ) of 𝜓𝑉
is a vector space. Hence the tangent space of Homℝ (𝑉 , 𝑉 ⟂ ) at 0 is Homℝ (𝑉 , 𝑉 ⟂ ) itself.
Therefore, we can use 𝑑𝜓𝑉 to get a linear isomorphism

𝑇𝑉 (Gr 𝑘 (ℝ𝑛+𝑘 )) ≅ 𝑇0 (Homℝ (𝑉 , 𝑉 ⟂ )) = Homℝ (𝑉 , 𝑉 ⟂ ).

9.4.3 Grassmannians are manifolds - linear algebraic proof

It is useful to also introduce a slightly different argument for proving the theorem:

An alternative argument using row vectors and more linear algebra

Any 𝑘-dimensional linear subspace in ℝ𝑛+𝑘 is spanned by 𝑘 row vectors in ℝ𝑛+𝑘 . Hence
any such 𝑘-dimensional linear subspace 𝑉 can be represented by a 𝑘 × (𝑛 + 𝑘)-matrix
⎛𝑣11 ⋯ 𝑣1𝑛+𝑘 ⎞
𝐴=⎜ ⋮ ⋱ ⋮ ⎟
⎜ ⎟
⎝𝑣𝑘1 ⋯ 𝑣𝑘𝑛+𝑘 ⎠
where any two such matrices 𝐴 and 𝐴′ represent the same element in Gr 𝑘 (ℝ𝑛+𝑘 ) if and only if
there is an invertible matrix 𝑘 × 𝑘-matrix 𝑇 such that
𝐴′ = 𝑇 ⋅ 𝐴.

Now let 𝐼 = {𝑖1 , … , 𝑖𝑘 } ⊂ {1, … , 𝑛 + 𝑘} be a 𝑘-tuple with 𝑖1 < 𝑖2 < ⋯ < 𝑖𝑘 . We define
the subset 𝐼 of Gr 𝑘 (ℝ𝑛+𝑘 ) to be the set of all 𝑘-planes which are represented by a 𝑘 × (𝑛 + 𝑘)-
matrix 𝐴 such that the 𝑘 × 𝑘-submatrix consisting of the columns 𝑖1 , … , 𝑖𝑘 is invertible. We
call this submatrix the I-th 𝑘 × 𝑘-minor of 𝐴. Recall that elementary row operations that
we learned about in our very first encounter with linear algebra do not change the row space,
i.e., the space spanned by the row vectors. Hence, after performing suitable row operations,
we see that every 𝑉 ∈ 𝐼 is represented by a unique 𝑘 × (𝑛 + 𝑘)-matrix in which this I-th
𝑘 × 𝑘-minor is the 𝑘 × 𝑘-identity matrix. In other words, every 𝑉 ∈ 𝐼 is represented by a
unique 𝑘 × (𝑛 + 𝑘)-matrix 𝐴𝐼 (𝑉 ) in which the 𝑗th column is the standard basis vector 𝑒𝑗 of
length 𝑘 if 𝑗 ∈ 𝐼 and an arbitrary vector of length 𝑘 if 𝑗 ∉ 𝐼.
168 9.4. Grassmannian

Example 9.21 (Representing planes by minors) Let us look at how this works:

∙ If 𝐼 = {1, 2, … , 𝑘}, then any 𝑘-plane in {1,…,𝑘} has a unique representative of

the form
⎛ 1 ⋯ 0 ∗ ⋯ ∗⎞
𝐴𝐼 = ⎜⋮ ⋱ ⋮ ⋮ ⋱ ⋮⎟ .
⎜ ⎟
⎝ 0 ⋯ 1 ∗ ⋯ ∗⎠

∙ More concretely, let 𝑘 = 3 and 𝑛 + 𝑘 = 5 and 𝐼 = {1, 3, 4}. Let 𝑉 be the 3-plane,
i.e., 3-dimensional linear subspace, spanned by the rows of the matrix

⎛3 6 0 1 −18⎞
𝐴 = ⎜3 1 1 0 −7 ⎟ .
⎜ ⎟
⎝5 6 1 1 −21⎠

The I-th minor is the matrix

⎛3 0 1⎞
𝑀 = ⎜3 1 0⎟
⎜ ⎟
⎝5 1 1⎠

consisting of the first, third and fourth column of 𝐴. This minor is an invertible
matrix with inverse
⎛1 1 −1⎞
𝑀 −1 = ⎜−3 −2 3 ⎟ .
⎜ ⎟
⎝−2 −3 3 ⎠
Then 𝑉 ∈ 𝐼 is uniquely represented by the matrix

⎛1 1 0 0 −4⎞
𝐴𝐼 = 𝑀 −1
⋅ 𝐴 = ⎜0 −2 1 0 5 ⎟ .
⎜ ⎟
⎝0 3 0 1 −6⎠

Back to the general discussion:

Conversely, any 𝑘×(𝑛+𝑘)-matrix with invertible I-th 𝑘×𝑘-minor defines a unique 𝑘-plane
in 𝐼 .

Thus, for each 𝐼, the 𝑘 ⋅ 𝑛 entries in the remaining columns of any 𝑘 × (𝑛 + 𝑘)-matrix of this
form define a bijection of sets
≅
← ℝ𝑘⋅𝑛 .
𝜓𝐼 ∶ 𝐼 ←←←→

Example 9.22 (Real projective space) For 𝑘 = 1, we have seen that Gr 1 (ℝ𝑛+1 ) = ℝP𝑛 .
A point in ℝP𝑛 is represented by an (𝑛 + 1)-tuple 𝑥1 , … , 𝑥𝑛+1 . Recall that we denote the
equivalence class by [𝑥] = [𝑥1 ∶ … ∶ 𝑥𝑛+1 ]. The sets 𝐼 then consist each of just one
number 𝑖 ∈ {1, … , 𝑛 + 1}. The set 𝑖 is then equal to the subset 𝑉𝑖 we have defined
previously:
𝑖 = 𝑉𝑖 = {[𝑥] ∈ ℝP𝑛 ∶ 𝑥𝑖 ≠ 0}.
The representing matrix 𝐴𝑖 of [𝑥] is the matrix with just one row
Chapter 9. Abstract Smooth Manifolds 169
(𝑥1 ∕𝑥𝑖 , … , 1, … , 𝑥𝑛+1 ∕𝑥𝑖 ) and a 𝑒1 = 1 in column 𝑖.

Claim: For each 𝐼, the subset 𝐼 is open in Gr 𝑘 (ℝ𝑛+𝑘 ).

By definition of 𝐼 , each 𝑘-plane in 𝐼 is represented by a 𝑘 × (𝑛 + 𝑘)-matrix 𝐴 such that

I-th 𝑘 × 𝑘-minor consisting of the columns 𝑖1 , … , 𝑖𝑘 is invertible. We define det 𝐼 to be the map
which sends a 𝑘 × (𝑛 + 𝑘)-matrix to the determinant of its I-th 𝑘 × 𝑘-minor. Then det 𝐼 is a map
on the space of all 𝑘 × (𝑛 + 𝑘)-matrices with values in ℝ. This map is continuous, since the
determinant of the I-th minor is a polynomial in the entries of this minor. Thus det 𝐼 depends
continuously on these entries. Finally, 𝐼 is a preimage of the open subset ℝ ⧵ {0} under det 𝐼 .

We also note that the subset 𝜓𝐼 (𝐼 ∩ 𝐼 ′ ) is open in ℝ𝑘⋅𝑛 for every pair 𝐼, 𝐼 ′ . Now we
can give a second proof of Theorem 9.19:

Proof: We already know that Grassmannians are compact. To show it once again anyway
we could argue as follows: Let 𝑉𝑘 be the 𝑘-plane in ℝ𝑛+𝑘 spanned by the first 𝑘 standard basis
𝑒1 , … , 𝑒𝑘 ∈ ℝ𝑛+𝑘 of length 𝑛 + 𝑘. We deduce from what we learned above that we have a
surjective map
𝑞 ∶ 𝑂(𝑛 + 𝑘) → Gr 𝑘 (ℝ𝑛+𝑘 ), 𝑇 → 𝑇 ⋅ 𝑉𝑘 .
Since 𝑂(𝑛 + 𝑘) is compact and the continuous image of compact sets is compact, this shows
that Gr 𝑘 (ℝ𝑛+𝑘 ) is compact.

We need to show that, for every pair 𝐼, 𝐼 ′ , the map

𝜓𝐼 ◦𝜓𝐼 ′
𝜓𝐼 ′ (𝐼 ∩ 𝐼 ′ ) ←←←←←←←←←←←←→
← 𝜓𝐼 (𝐼 ∩ 𝐼 ′ )
is a smooth map between open subsets of ℝ𝑘⋅𝑛 , where we omit to denote the appropriate re-
strictions of 𝜓𝐼 and 𝜓𝐼 ′ to simplify the notation. To prove this let 𝑉 be a 𝑘-plane in 𝐼 ∩ 𝐼 ′ .
Let 𝐴𝐼 be a matrix representing 𝑉 with its I-th 𝑘 × 𝑘-minor being the identity matrix and let
𝐴𝐼 ′ be a matrix representing 𝑉 with its 𝐼 ′ th 𝑘 × 𝑘-minor being the identity matrix. Let 𝑇𝐼𝐼′ be
the 𝐼 ′ th 𝑘 × 𝑘-minor of 𝐴𝐼 . Then we have
𝑇𝐼𝐼′ ⋅ 𝐴𝐼 ′ = 𝐴𝐼 .
Since the entries of 𝑇𝐼𝐼′ vary smoothly with the entries in 𝐴𝐼 , it follows that the entries in 𝐴𝐼 ′
also vary smoothly with the entries in 𝐴𝐼 . Thus 𝜓𝐼 ◦𝜓𝐼 ′ is smooth.

9.5 Embedding Manifolds in Euclidean Space

We will now study the following fundamental question:

Question Given an abstract smooth manifold 𝑋 in the sense of Definition 9.3: Is it

possible to embed 𝑋 into Euclidean space? That is, is there a natural number 𝑁 and
an embedding 𝑋 → ℝ𝑁 ?

The answer to this question is yes which may justify our initial approach to smooth man-
ifolds as subsets of Euclidean space. We will discuss the proof only for compact manifolds
170 9.5. Embedding Manifolds in Euclidean Space
in Section 9.5.2. The non-compact requires much more familiarity with arguments in general
topology and we leave this as an extra homework.

Once we know that every smooth abstract manifold can be considered as subspace of some
Euclidean space we will try to answer the following question:

Question Given a 𝑘-dimensional manifold 𝑋 ⊂ ℝ𝑁 : What is the minimal 𝑛 such

that we can be sure that there is an embedding 𝑋 ⊂ ℝ𝑛 ?

We will approach an answer to this question in two steps: In Section 9.5.3 we show that
there always is an one-to-one immersion of 𝑋 into ℝ2𝑘+1 . If 𝑋 is compact, this immersion
will automatically be an embedding as we learned in Section 3.3. In Section 9.5.2 we will then
show that also for non-compact smooth 𝑘-manifolds there is always an embedding into ℝ2𝑘+1 .
In fact, Whitney showed that 𝑁 = 2𝑘 always works. The proof is much harder, and we will not
discuss it in these notes.

In several proofs in this section we will employ an important and very useful tool. It will
also turn out handy in several occasions later on. This will be the content of the next subsection.

9.5.1 Partition of unity

We recall some terminology from general topology:

Definition 9.23 (The closure of a subset) Let 𝑋 be a topological space and 𝐴 be

an arbitrary subset. The closure of 𝐴 in 𝑋, denoted 𝐴, is the intersection of all closed
subsets in 𝑋 which contain 𝐴.

We are familiar with the closure of subsets in many cases. For example, the closure of an
open ball 𝔹𝜀 (0) in ℝ𝑁 is just the closed ball

𝔹𝜀 (0) = {𝑥 ∈ ℝ𝑁 ∶ |𝑥| ≤ 𝜀}.

We need the closure of a subset for example when we want to talk about the support of a
function:

Definition 9.24 (Support of a function) Let 𝑋 be a smooth manifold and 𝑓 ∶ 𝑋 → ℝ

be a smooth function. The closed subset

supp(𝑓 ) ∶= {𝑥 ∈ 𝑋 ∶ 𝑓 (𝑥) ≠ 0}

is called the support of 𝑓 .

We are now going to introduce a fundamental tool for studying manifolds:

Definition 9.25 (Partition of unity) Let 𝑋 be a smooth manifold and let {𝑈𝛼 } be an
⋃
open cover, i.e., a collection of open subsets in 𝑋 such that 𝛼 𝑈𝛼 = 𝑋. A sequence
Chapter 9. Abstract Smooth Manifolds 171
of smooth functions {𝜌𝑖 ∶ 𝑋 → ℝ} is called a partition of unity subordinate to the
open cover {𝑈𝛼 } if it has the following properties:

(a) 0 ≤ 𝜌𝑖 (𝑥) ≤ 1 for all 𝑥 ∈ 𝑋 and all 𝑖.

(b) Each 𝑥 ∈ 𝑋 has a neighborhood on which all but finitely many functions 𝜌𝑖 are
identically zero.

(c) For each 𝑖, supp(𝜌𝑖 ) ⊂ 𝑈𝛼 for some 𝛼.

∑
(d) For each 𝑥 ∈ 𝑋, 𝑖 𝜌𝑖 (𝑥) = 1. Note that according to (b), this sum is always
finite.

The most general existence result for partitions of unity (only requiring that each 𝜌𝑖 is merely
continuous and not smooth) is that they exist on every paracompact space, i.e., spaces on
which every open cover has a locally finite refinement. The latter means that every point has
a neighborhood that intersects only finitely many sets in the cover.

We will postpone the proof of the existence of partitions of unity on smooth manifolds to
the end of this section. See Section 9.5.6. It is an interesting proof and we encourage everybody
to go through it at some point. However, it is also rather technical and diverts our attention too
far from our main goal. We will therefore first look at some applications of partitions of unity
and will then show that they actually exist.

9.5.2 Embedding abstract manifolds in Euclidean space

We use partitions of unity to prove the following important result:

Theorem 9.26 (Embedding abstract manifolds in Euclidean space) Let 𝑋 be a

compact abstract smooth 𝑘-manifold. Then there is an embedding, i.e., an injective
proper map, 𝑋 → ℝ𝑁 for some large 𝑁.

Proof:

The collection of all 𝑉𝛼 for all charts (𝑉𝛼 , 𝜙𝛼 ) is an open cover of 𝑋. Since 𝑋 is compact,
we can cover 𝑋 by the images of a finite number of charts 𝑉1 , … , 𝑉𝑛 .

Let {𝜌𝑖 } be a partition of unity subordinate to the open cover defined by the 𝑉𝑖 ’s.

For a chart 𝜙𝑖 ∶ 𝑉𝑖 → 𝑈𝑖 ⊂ ℝ𝑘 , we define a new map

{
𝑘 𝜌𝑖 (𝑥) ⋅ 𝜙𝑖 (𝑥) for 𝑥 ∈ 𝑉𝑖
𝑔𝑖 ∶ 𝑋 → ℝ , 𝑔𝑖 (𝑥) =
0 for 𝑥 ∈ 𝑋 ⧵ supp(𝜌𝑖 ).

The map 𝑔𝑖 is well-defined, since if 𝑥 ∈ 𝑉𝑖 ⧵ supp(𝜌𝑖 ), then both definitions agree to be 0.

Moreover, 𝑔𝑖 is smooth, since its restrictions to the two open subsets 𝑉𝑖 and 𝑋 ⧵ supp(𝜌𝑖 ) are
smooth.
172 9.5. Embedding Manifolds in Euclidean Space

Remark 9.27 At this point we see why we do not use 𝑋 ⧵ 𝑉𝑖 in the definition of 𝑔𝑖
because that would be a closed subset. We also observe that it would not work to drop
the 𝜌𝑖 and just use the 𝜙𝑖 , since there is no reason why 𝜙𝑖 would get closer to zero the
closer we get to the boundary of 𝑉𝑖 . Hence there are several reasons why continuity and
smoothness would not be guaranteed without the 𝜌𝑖 .

Back to the proof: Now we define a map

𝑔 ∶ 𝑋 → ℝ𝑛 × ℝ𝑛𝑘 , 𝑥 → (𝜌1 (𝑥), … , 𝜌𝑛 (𝑥), 𝑔1 (𝑥), … , 𝑔𝑛 (𝑥)).

We observe that 𝑔 is continuous, since the 𝑔𝑖 ’s and the 𝜌𝑖 ’s are continuous.

Claim: 𝑔 is an injective proper map.

Since 𝑋 is compact, 𝑔 is a proper map. Now we show that 𝑔 is injective. So assume

𝑔(𝑥) = 𝑔(𝑦). Then 𝜌𝑖 (𝑥) = 𝜌𝑖 (𝑦) for all 𝑖 by the definition of 𝑔. However, by the definition of a
partition of unity, for at least one 𝑖, we must have 𝜌𝑖 (𝑥) = 𝜌𝑖 (𝑦) ≠ 𝟎.

Thus 𝑥 and 𝑦 must lie in the same 𝑉𝑖 , since 𝜌𝑖 is supported on 𝑉𝑖 , i.e., 𝜌𝑖 (𝑥) ≠ 0 implies
𝑥 ∈ 𝑉𝑖 and similarly for 𝑦. Hence, since we have 𝑔𝑖 (𝑥) = 𝑔𝑖 (𝑦) and 𝜌𝑖 (𝑥) = 𝜌𝑖 (𝑦) ≠ 𝟎, we must
have
𝜙𝑖 (𝑥) = 𝜙𝑖 (𝑦).

Since 𝜙𝑖 is a bijection, this shows 𝑥 = 𝑦. Thus 𝑔 is injective.

Actually, 𝑔 is also an immersion, but we have not defined what that means for an abstract
manifold. However, this is just an exercise in translating the definitions, and we omit this point
and rather move on.

Remark 9.28 (All manifolds can be embedded in Euclidean space) In fact, every ab-
stract 𝑘-manifold 𝑋 can be embedded in Euclidean space. One can just keep on going
with the above argument in the non-compact case and use local charts to map pieces
of 𝑋 into ℝ𝑘 . Though when using only finitely many copies of ℝ𝑘 to accommodate
infinitely many neighborhoods of 𝑋, we loose injectivity. The key tool that restores
injectivity is a partition of unity which evens out the troubles occurring because of
overlapping neighborhoods.
For this to work, it is crucial that the topology on 𝑋 has a countable basis. This is
a technical point which we do not discuss any further because it would divert us too far
from the main story.
We just remark that it is possible to construct topological spaces without a count-
able basis which are locally homeomorphic to Euclidean space, but which cannot be
embedded into Euclidean space.

9.5.3 Whitney’s Theorem for smooth manifolds

We are now ready to prove the following fundamental result:

Chapter 9. Abstract Smooth Manifolds 173

Theorem 9.29 (Whitney’s Theorem) Every 𝑘-dimensional manifold admits a one-

to-one immersion in ℝ2𝑘+1 .

Proof: Let 𝑋 ⊂ ℝ𝑁 be 𝑘-dimensional manifold which is a subset in ℝ𝑁 for some 𝑁 >

2𝑘 + 1. In particular, we are given an injective immersion 𝑋 → ℝ𝑁 . Our goal is to show that
we can choose 𝑁 to be 2𝑘 + 1 and still have an injective immersion. Therefor we are going to
construct a linear projection ℝ𝑁 → ℝ2𝑘+1 that restricts to a one-to-one immersion 𝑋 → ℝ2𝑘+1
on 𝑋.

The construction will proceed by induction:

Whenever we are given an injective immersion 𝑓 ∶ 𝑋 → ℝ𝑀 with 𝑀 > 2𝑘 + 1, then we

will show that there exists a vector 𝑎 ∈ ℝ𝑀 such that the composition of 𝑓 with the projection
map carrying ℝ𝑀 onto the orthogonal complement of 𝑎 is still an injective immersion. The
complement 𝐻 ∶= {𝑏 ∈ ℝ𝑀 ∶ 𝑏 ⟂ 𝑎} is an 𝑀 −1-dimensional vector subspace of ℝ𝑀 , hence
isomorphic to ℝ𝑀−1 . Thus, after choosing a basis for 𝐻, we obtain an injective immersion
into ℝ𝑀−1 .

Continuing this procedure yields a chain of linear maps

ℝ𝑁 → ℝ𝑁−1 → ⋯ → ℝ2𝑘+1

such that the composition 𝑋 → ℝ𝑁 → ℝ2𝑘+1 is still an injective immersion.

So let us assume we have an injective immersion

𝑓 ∶ 𝑋 → ℝ𝑀 with 𝑀 > 2𝑘 + 1.

We define two smooth maps

𝑋×𝑋×ℝ
ℎ

𝑇 (𝑋) / ℝ𝑀
𝑔

by

ℎ ∶ 𝑋 × 𝑋 × ℝ → ℝ𝑀 , (𝑥, 𝑦, 𝑡) → 𝑡(𝑓 (𝑥) − 𝑓 (𝑦))

and, using 𝑇𝑦 (ℝ𝑀 ) = ℝ𝑀 at any 𝑦 ∈ ℝ𝑀 and hence 𝑑𝑓𝑥 is a map 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → ℝ𝑀 ,

𝑔 ∶ 𝑇 (𝑋) → ℝ𝑀 , (𝑥, 𝑣) → 𝑑𝑓𝑥 (𝑣).

By Sard’s Theorem 7.1 the sets of regular values 𝑅𝑔 and 𝑅ℎ for 𝑔 and ℎ, respectively, are
dense subsets in ℝ𝑀 . Thus their intersection is non-empty and there exists a point in ℝ𝑀 which
is a regular value for both 𝑔 and ℎ simultaneously.

Since dim 𝑇 (𝑋) = 2𝑘, dim 𝑋 × 𝑋 × ℝ = 2𝑘 + 1, but 𝑀 > 2𝑘 + 1, the only regular values
of 𝑔 and ℎ are the points in ℝ𝑀 which are not in the image of 𝑔 or ℎ. Hence there exists a
point 𝑎 ∈ ℝ𝑀 which is neither in the image of 𝑔 nor in the image of ℎ. Note that, since 0
belongs to both images, we must have 𝑎 ≠ 0.
174 9.5. Embedding Manifolds in Euclidean Space

Remark 9.30 (Cannot reduce beyond 2𝑘 + 1) Before we move on, note that we could
only make this argument for 𝑀 > 2𝑘 + 1. Hence this induction argument of shrinking
𝑀 further and further cannot be extended beyond 2𝑘 + 1.

Let 𝜋 be the projection of ℝ𝑀 onto the orthogonal complement 𝐻 of 𝑎.

First claim: 𝜋◦𝑓 ∶ 𝑋 → 𝐻 is injective.

For suppose that 𝜋◦𝑓 (𝑥) = 𝜋◦𝑓 (𝑦). Then, since 𝜋 is linear, we have
𝜋(𝑓 (𝑥) − 𝑓 (𝑦)) = 0,
i.e.,
𝑓 (𝑥) − 𝑓 (𝑦) ∈ Ker (𝜋) = span(𝑎) in ℝ𝑀
= {𝑤 ∈ ℝ𝑀 ∶ 𝑤 = 𝑡 ⋅ 𝑎 for some 𝑡 ∈ ℝ}.
Thus there is a 𝑡 ∈ ℝ with 𝑓 (𝑥) − 𝑓 (𝑦) = 𝑡𝑎. If 𝑥 ≠ 𝑦 then 𝑡 ≠ 0, since 𝑓 is injective. But then
𝑎 = 1∕𝑡(𝑓 (𝑥) − 𝑓 (𝑦)) = ℎ(𝑥, 𝑦, 1∕𝑡)
which contradicts the choice of 𝑎 not being in the image of ℎ.

Second claim: 𝜋◦𝑓 ∶ 𝑋 → 𝐻 is an immersion.

For suppose there was a nonzero vector 𝑣 in 𝑇𝑥 (𝑋) for which 𝑑(𝜋◦𝑓 )𝑥 = 0. Because 𝜋 is
linear, we have 𝑑𝜋𝑓 (𝑥) = 𝜋 and the chain rule yields
𝑑(𝜋◦𝑓 )𝑥 = 𝜋◦𝑑𝑓𝑥 .
Thus 𝜋(𝑑𝑓𝑥 (𝑣)) = 0, so 𝑑𝑓𝑥 (𝑣) = 𝑡𝑎 for some 𝑡 ∈ ℝ. Because 𝑓 is an immersion, we must
have 𝑡𝑎 ≠ 0. Since we know 𝑎 ≠ 0, this implies 𝑡 ≠ 0. Thus, since 𝑑𝑓𝑥 is linear,
( ) ( )
1 1 1
𝑎 = 𝑑𝑓𝑥 (𝑣) = 𝑑𝑓𝑥 𝑣 = 𝑔 𝑥, 𝑣
𝑡 𝑡 𝑡
which again contradicts the choice of 𝑎 not being in the image of 𝑔.

9.5.4 Whitney’s Theorem - the compact case

For compact manifolds, one-to-one immersions are the same as embeddings. So we have just
proved the embedding theorem in the compact case.

Theorem 9.31 (Whitney’s Embedding for compact manifolds) Every compact 𝑘-

dimensional manifold admits an embedding in ℝ2𝑘+1 .

Note that Whitney’s result does not give us the minimal 𝑁 for an individual manifold. For
example, we know that 𝕊𝑛 is embedded in ℝ𝑛+1 for every 𝑛. The result tells us that, in general,
𝑁 = 2𝑘 + 1 will always work.

In order to extend Whitney’s theorem (for 𝑁 = 2𝑘 + 1) to non-compact manifolds, we have

to modify the immersion to make it proper. This is a topological, not a differential problem.
The key tool to solve this problem will again be partitions of unity.
Chapter 9. Abstract Smooth Manifolds 175
9.5.5 Whitney’s Embedding Theorem

Partitions of unity allow us to prove the following key lemma:

Lemma 9.32 (Existence of proper functions on manifolds) On every smooth manifold

𝑋, there is a proper smooth function 𝑝 ∶ 𝑋 → ℝ.

Proof: Let {𝑈𝛼 } be the collection of open subsets of 𝑋 that have compact closure, and let
{𝜌𝑖 } be a subordinate partition of unity. Then
∑
∞
𝑝(𝑥) = 𝑖𝜌𝑖 (𝑥)
𝑖=1
is a well-defined smooth function, since, in a neighborhood of every point, it is a finite sum of
smooth functions.

In order to show that 𝑝 is proper, we need to show that the preimage of any compact subset
of ℝ is again compact. Every compact subset 𝐾 ⊂ ℝ is contained in a closed interval of the
form [−𝑗, 𝑗] for some large enough natural number 𝑗. Hence if we can show that 𝑝−1 ([−𝑗, 𝑗])
is compact, then 𝑝−1 (𝐾) is a closed subset of a compact set and therefore also compact.

For a given natural number 𝑗, assume we have 𝜌1 (𝑥) = ⋯ = 𝜌𝑗 (𝑥) = 0 for any 𝑥 ∈ 𝑋.
Then, by definition of a partition of unity, we have
∑
∞
𝜌𝑖 (𝑥) = 1
𝑖=𝑗+1

and therefore
∑
∞
𝑝(𝑥) ≥ (𝑗 + 1) 𝜌𝑖 (𝑥) = 𝑗 + 1 >𝑗.
𝑖=𝑗+1
This shows
⋃
𝑗
−1
𝑝 ([−𝑗, 𝑗]) ⊂ {𝑥 ∈ 𝑋 ∶ 𝜌𝑖 (𝑥) ≠ 0}.
𝑖=1
Since {𝜌𝑖 } is a partition of unity subordinate to {𝑈𝛼 }, we have supp(𝜌𝑖 ) ⊂ 𝑈𝑖 . Since 𝑈𝑖 has
compact closure and the finite union of compact sets is compact, this shows that 𝑝−1 ([−𝑗, 𝑗])
is a closed subset in a compact set and therefore it is also compact.

Theorem 9.33 (Whitney’s Embedding Theorem) Every smooth 𝑘-dimensional man-

ifold 𝑋 ⊂ ℝ𝑁 admits an embedding into ℝ2𝑘+1 .

Recall that the strongest result is that 𝑁 = 2𝑘 suffices. But this is much harder to prove.
Note also that, for many manifolds, an even lower dimension suffices, e.g., 𝕊𝑛 ⊂ ℝ𝑛+1 the
𝑛-sphere is embedded in ℝ𝑛+1 .

Proof of Whitney’s Theorem 9.33:

The idea is to replace the injective immersion 𝑓 ∶ 𝑋 → ℝ𝑁 with the map (𝑓 , 𝑝) ∶ 𝑋 →

ℝ𝑁+1 with a proper 𝑝 ∶ 𝑋 → ℝ. Then (𝑓 , 𝑝) is still an injective immersion, and it is proper,
since 𝑝 is proper. It remains to reduce the dimension 𝑁 +1. The details are a bit more involved:
176 9.5. Embedding Manifolds in Euclidean Space
Starting with 𝑋 ⊂ ℝ𝑁 we have seen that we can find an injective immersion 𝑓 ∶ 𝑋 →
ℝ2𝑘+1 . By composing 𝑓 with the injective immersion
𝑥
ℝ2𝑘+1 → 𝐵1 (0), 𝑥 → ,
1 + |𝑥|2
we can assume that |𝑓 (𝑥)| < 1 for all 𝑥 ∈ 𝑋.

Let 𝑝 ∶ 𝑋 → ℝ be a proper function which we know to exist by the previous lemma. We

define a new injective immersion

𝐹 ∶ 𝑋 → ℝ2𝑘+2 , 𝑥 → (𝑓 (𝑥), 𝑝(𝑥)).

Since 2𝑘 + 2 > 2𝑘 + 1, we can apply the argument from the proof of the previous theorem
by Whitney and find a nonzero vector 𝑎 ∈ ℝ2𝑘+2 such that

𝜋◦𝐹 ∶ 𝑋 → 𝐻

is still an injective immersion, where 𝜋 is the projection onto the orthogonal complement
𝐻 = {𝑏 ∈ ℝ2𝑘+2 ∶ 𝑏 ⟂ 𝑎} of 𝑎 in ℝ2𝑘+2 . By rescaling we can assume |𝑎| = 1. In other words,
we can assume 𝑎 ∈ 𝕊2𝑘+1 .

Since 𝜋◦𝐹 is an injective immersion for almost every 𝑎 ∈ 𝕊2𝑘+1 , we can assume that 𝑎 is
neither the north nor the south pole on 𝕊2𝑘+1 . For if 𝜋◦𝐹 failed to be an injective immersion
on these two points, it would suffice to rotate 𝕊2𝑘+1 a bit (which is a diffeomorphism of 𝕊2𝑘+1 )
in order to avoid north and south pole.

This will allow us to show that 𝜋◦𝐹 is proper as follows:

Claim: Given any bound 𝑐, there exists another number 𝑑 such that

{𝑥 ∈ 𝑋 ∶ |(𝜋◦𝐹 )(𝑥)| ≤ 𝑐} ⊂ {𝑥 ∈ 𝑋 ∶ |𝑝(𝑥)| ≤ 𝑑}.

The claim implies properness: Since 𝑝 is proper,

{𝑥 ∈ 𝑋 ∶ |𝑝(𝑥)| ≤ 𝑑} = 𝑝−1 ([−𝑑, 𝑑])

is a compact subset of 𝑋. Thus the preimage under 𝜋◦𝐹 of every closed ball in 𝐻 is a compact
subset of 𝑋. Since every compact subset 𝐾 of 𝐻 is a closed subset of some closed ball in 𝑋,
this shows that (𝜋◦𝐹 )−1 (𝐾) is a closed subset of a compact subset in 𝑋 and therefore also
compact.

Proof of the claim: If the claim was false, then there exists a 𝑐 and a sequence of points
{𝑥𝑖 } in 𝑋 for which

|(𝜋◦𝐹 )(𝑥𝑖 )| ≤ 𝑐, but |𝑝(𝑥𝑖 )| → ∞

as there would be no 𝑑 bounding |𝑝(𝑥𝑖 )|.

By definition of the projection onto an orthogonal complement, for every 𝑧 ∈ ℝ2𝑘+2 ,

𝜋(𝑧) is the one point in 𝐻 for which 𝑧 − 𝜋(𝑧) is a multiple of 𝑎. In particular,

𝐹 (𝑥𝑖 ) − 𝜋(𝐹 (𝑥𝑖 )) is a multiple of a for each 𝑖,

Chapter 9. Abstract Smooth Manifolds 177
and hence so is the vector
1
𝑤𝑖 ∶= (𝐹 (𝑋𝑖 ) − 𝜋(𝐹 (𝑥𝑖 ))).
𝑝(𝑥𝑖 )

Let us look at what happens when 𝑖 → ∞:

( )
𝐹 (𝑥𝑖 ) 𝑓 (𝑥𝑖 )
= , 1 → (0, … , 0, 1)
𝑝(𝑥𝑖 ) 𝑝(𝑥𝑖 )
because |𝑓 (𝑥𝑖 )| < 1 for all 𝑖 and 𝑝(𝑥𝑖 ) → ∞. We have
| 𝜋(𝐹 (𝑥𝑖 )) | 𝑐
| |
| 𝑝(𝑥 ) | ≤ |𝑝(𝑥 )| .
| 𝑖 | 𝑖
Thus
𝜋(𝐹 (𝑥𝑖 ))
→ 0 ⇒ 𝑤𝑖 → (0, … , 0, 1).
𝑝(𝑥𝑖 )

But each 𝑤𝑖 is a multiple of 𝑎. Hence the limit of the 𝑤𝑖 must be a multiple of 𝑎 as well.
We conclude that 𝑎 must be either the north or south pole of 𝕊𝑘+1 which contradicts our
assumption on 𝑎. This proves the claim and finishes the proof of the theorem.

9.5.6 Existence of partitions of unity on smooth manifolds

Now we return to partitions of unity and prove that they exist on smooth manifolds. Before we
can start the proof, we need some further preparation:

Lemma 9.34 (Separating closed subsets) Let 𝐴 and 𝐶 be disjoint closed subsets in
ℝ𝑁 . Then there are disjoint open subsets 𝑈 and 𝑉 such that 𝐴 ⊂ 𝑈 and 𝐶 ⊂ 𝑉 .

Proof: For each 𝑎 ∈ 𝐴, choose an 𝜀𝑎 > 0 such that 𝐵2𝜀𝑎 (𝑎) ∩ 𝐶 = ∅. This is possible since
𝐶 is closed. Similarly, for each 𝑐 ∈ 𝐶, choose an 𝜀𝑐 > 0 such that 𝐵2𝜀𝑐 (𝑐) ∩ 𝐴 = ∅. We define

𝑈 ∶= ∪𝑎∈𝐴 𝐵𝜖𝑎 (𝑎) and 𝑉 ∶= ∪𝑐∈𝐶 𝐵𝜀𝑐 (𝑐).

Then 𝑈 and 𝑉 are open subsets with 𝐴 ⊂ 𝑈 and 𝐶 ⊂ 𝑉 . We claim that 𝑈 and 𝑉 are disjoint.

For, if 𝑥 ∈ 𝑈 ∩ 𝑉 , then
𝑥 ∈ 𝐵𝜀𝑎 (𝑎) ∩ 𝐵𝜀𝑐 (𝑐)
for some 𝑎 ∈ 𝐴 and 𝑐 ∈ 𝐶. By the triangle inequality, this implies
|𝑎 − 𝑐| < 𝜀𝑎 + 𝜀𝑐 .

But, if 𝜀𝑎 ≤ 𝜀𝑐 , then |𝑎 − 𝑐| < 2𝜀𝑐 and 𝑎 ∈ 𝐵2𝜀𝑐 (𝑐). And, if 𝜀𝑐 ≤ 𝜀𝑎 , then |𝑎 − 𝑐| < 2𝜀𝑎 and
𝑐 ∈ 𝐵2𝜀𝑎 (𝑎). Both cases are impossible.

Another important tool that we will need are smooth bump functions which we introduced
in Section 8.2. Now we will need them in a slightly more interesting form:
178 9.5. Embedding Manifolds in Euclidean Space

Lemma 9.35 (Smooth bump functions revisited) Let 𝑈 ⊂ ℝ𝑁 be open and 𝐾 ⊂ 𝑈

be compact. Then there is a smooth function 𝜑 ∶ ℝ𝑁 → ℝ with 𝜑(𝑥) = 1 for all 𝑥 ∈ 𝐾
and 𝜑(𝑥) = 0 for all 𝑥 ∈ ℝ𝑁 ⧵ 𝐶 for some closed subset 𝐶 with 𝐾 ⊂ 𝐶 ⊂ 𝑈 .

Proof: Recall the smooth function

{ 2
𝑒−1∕𝑥 𝑥>0
𝑓 ∶ ℝ → ℝ, 𝑓 (𝑥) =
0 𝑥 ≤ 0.

For any given 𝜖 > 0, we define a function

𝑓𝜀 (𝑥) ∶= 𝑓 (𝑥)𝑓 (𝑥 − 𝜀).

As a product of two smooth functions, 𝑓𝜀 is smooth.

Next we define yet another function

𝑥
∫0 𝑓𝜀 (𝑡)𝑑𝑡
𝑔𝜀 ∶ ℝ → ℝ, 𝑔𝜀 (𝑥) ∶= 𝜀 .
∫0 𝑓𝜀 (𝑡)𝑑𝑡
By the Fundamental Theorem of Calculus, 𝑔𝜀 is smooth, nondecreasing, and
⎧𝑔𝜀 (𝑥) = 0 𝑥≤0
⎪
⎨0 < 𝑔𝜀 (𝑥) < 1 0<𝑥<𝜀
⎪𝑔 (𝑥) = 1 𝑥≥𝜀
⎩ 𝜀

Finally, for any fixed point 𝑎 ∈ ℝ𝑁 and for any given 𝑟 > 0, we define
ℎ𝑟𝜀 ∶ ℝ𝑁 → ℝ, ℎ𝑟𝜀 (𝑥) = 1 − 𝑔𝜀 (|𝑥 − 𝑎| − 𝑟).

Then ℎ𝑟𝜀 is smooth, nonincreasing, and

⎧ℎ𝑟 (𝑥) = 1 |𝑥 − 𝑎| ≤ 𝑟
⎪ 𝜀 𝑟
⎨0 < ℎ𝜀 (𝑥) < 1 𝑟 < |𝑥 − 𝑎| < 𝑟 + 𝜀
⎪ℎ𝑟 (𝑥) = 0 |𝑥 − 𝑎| ≥ 𝑟 + 𝜀
⎩ 𝜀
Chapter 9. Abstract Smooth Manifolds 179
This gives us a smooth function 𝜙 ∶ ℝ𝑁 → ℝ which has value 1 on the compact subset
𝔹𝑟 (𝑎) and has value 0 outside the closed subset 𝔹𝑟+𝜀 (𝑎).

Now let 𝑈 ⊂ ℝ𝑁 be open and 𝐾 ⊂ 𝑈 be compact. For this general situation we need to
work a bit harder and rearrange the argument as follows:

Let 𝜓 be the function

{ 2
𝑒−1∕|𝑥| |𝑥| < 1
𝜓𝜀 ∶ ℝ𝑁 → ℝ, 𝜓𝜀 (𝑥) =
0 |𝑥| ≥ 1.

This is a smooth map with ∫ℝ𝑁 𝜓𝑑𝑥 = 1 where 𝑑𝑥 denotes the standard Lebesgue measure on
ℝ𝑁 .

For a given 𝜀 > 0, define 𝜓𝜀 ∶ 𝑅𝑁 → ℝ by 𝜓𝜀 (𝑥) ∶= 𝜀−𝑁 𝜓(𝑥∕𝜀). This is still a smooth
function with ∫ℝ𝑁 𝜓𝜀 𝑑𝑥 = 1.

Since ℝ𝑁 ⧵ 𝑈 is closed and 𝐾 is compact, we can choose a small 𝜀 > 0 such that, for each
point 𝑥 ∈ 𝐾, we have 𝐵2𝜀 (𝑥) ∩ 𝑈 = ∅. Then the 𝑉 ∶= ∪𝑥∈𝐾 𝐵𝜀 (𝑥) is an open set containing 𝐾
with compact closure 𝑉̄ ⊂ 𝑈 contained in 𝑈 .

Let 𝜒𝑉 be the characteristic function on 𝑉 , i.e. the function

{
𝜒𝑉 (𝑥) = 1 for 𝑥 ∈ 𝑉
𝜒𝑉 ∶ ℝ𝑁 → ℝ,
𝜒𝑉 (𝑥) = 0 for 𝑥 ∉ 𝑉 .

The function 𝜒𝑉 is identically 1 on 𝐾 and has compact support contained in 𝑈 . But it is

of course not smooth on ℝ𝑁 , not even continuous. Hence we need to modify it, to make it
smooth. The function 𝜓𝜀 , for the fixed 𝜀, will serve as a tool to ‘smoothen’ 𝜒𝑉 .

Then the desired smooth function 𝜑 is the convolution 𝜓𝜀 ∗ 𝜒𝑉 of 𝜒𝑉 and 𝜓𝜀 :

𝜑 ∶ ℝ𝑁 → ℝ, 𝑥 → 𝜓𝜀 (𝑥 − 𝑦)𝜒𝑉 (𝑦)𝑑𝑦.
∫ℝ 𝑁

Note that the integral is well-defined, since the closure of 𝑉 , which is the support of 𝜒𝑉 , is
compact.

Existence of partitions of unity

We are going to show that partitions of unity exist on manifolds step by step with increasing
difficulty. We start with the case of compact subspaces in ℝ𝑁 . Then we are going to transport
this result to compact smooth manifolds. Finally, we discuss arbitrary compact smooth 𝑘-
manifolds. There is no need to restrict to compact manifolds. In fact, partitions of unity exist
on every paracompact topological space (every open cover has a locally finite refinement), a
class of spaces much larger than abstract manifolds.

First case: 𝑋 ⊂ ℝ𝑁 compact.

Let {𝑈𝛼 } be an open cover of 𝑋. Since 𝑋 is compact, {𝑈𝛼 } has a finite subcover {𝑈1 , … , 𝑈𝑛 }.
A partition of unity subordinate to the finite subcover is also a partition of unity subordinate
to the original cover.
180 9.5. Embedding Manifolds in Euclidean Space
Step 1: We are going to show that we can shrink the covering to an open covering {𝑉1 … , 𝑉𝑛 }
such that 𝑉̄𝑖 ⊂ 𝑈𝑖 for each 𝑖.

Consider the closed subset

𝐴 ∶= 𝑋 ⧵ (𝑈2 ∪ ⋯ ∪ 𝑈𝑛 )

of 𝑋. Since {𝑈1 , … , 𝑈𝑛 } cover 𝑋, we know 𝐴 ⊂ 𝑈1 . Since 𝐴 and 𝑋 ⧵ 𝑈1 are closed disjoint,

we can choose an open subset 𝑉1 containing 𝐴 such that 𝑉1 is disjoint to an open subset 𝑊
which contains 𝑋 ⧵𝑈1 . Thus 𝑉1 is contained in the complement 𝑋 ⧵𝑊 . Since 𝑋 ⧵𝑊 is a closed
subset which contains 𝑉1 , we know 𝑉̄1 ⊂ 𝑋 ⧵ 𝑊 , since the closure of 𝑉1 is the intersection
of all closed subsets which contain 𝑉1 . Since 𝑋 ⧵ 𝑈1 ⊂ 𝑊 by the choice of 𝑊 , we have
𝑋 ⧵ 𝑊 ⊂ 𝑋 ⧵ (𝑋 ⧵ 𝑈1 ) = 𝑈1 . Thus we have 𝑉̄1 ⊂ 𝑈1 . Since 𝑉1 contains the complement of
𝑈2 ∪ ⋯ ∪ 𝑈𝑛 in 𝑋, the collection {𝑉1 , 𝑈2 , … , 𝑈𝑛 } covers 𝑋.

Now we proceed by induction as follows: Given open subsets 𝑉1 , … , 𝑉𝑘−1 such that

𝑋 = {𝑉1 , … , 𝑉𝑘−1 , 𝑈𝑘 , 𝑈𝑘+1 , … , 𝑈𝑛 },

let 𝐴𝑘 be the subset

𝐴𝑘 = 𝑋 ⧵ (𝑉1 ∪ ⋯ ∪ 𝑉𝑘−1 ) ∪ (𝑈𝑘+1 ∪ ⋯ ∪ 𝑈𝑛 ).

Then 𝐴𝑘 is a closed subset of 𝑋 which is contained in the open set 𝑈𝑘 . Choose an open subset
𝑉𝑘 containing 𝐴𝑘 such that 𝑉̄𝑘 ⊂ 𝑈𝑘 . Then {𝑉1 , … , 𝑉𝑘−1 , 𝑉𝑘 , 𝑈𝑘+1 , … , 𝑈𝑛 } covers 𝑋. At the
𝑛th step of the induction we are done.

Step 2: Given the open covering {𝑈1 , … , 𝑈𝑛 } of 𝑋, we use Step 1 to choose an open cover
{𝑉1 , … , 𝑉𝑛 } of 𝑋 such that 𝑉̄𝑖 ⊂ 𝑈𝑖 for each 𝑖. Then we repeat this process and choose an
open cover {𝑊1 , … , 𝑊𝑛 } of 𝑋 such that 𝑊̄ 𝑖 ⊂ 𝑉𝑖 for each 𝑖.

For each 𝑖, we choose a smooth bump function

𝜑𝑖 ∶ 𝑋 → [0, 1] such that 𝜑𝑖 (𝑊̄ 𝑖 ) = {1} and 𝜑𝑖 (𝑋 − 𝑉𝑖 ) = {0}.

𝑖 (ℝ ⧵ {0}) ⊂ 𝑉𝑖 , we have
Since 𝜑−1

supp(𝜑𝑖 ) ⊂ 𝑉̄𝑖 ⊂ 𝑈𝑖 .

Note: Here is the point where we see why we need to apply Step 1 twice: If we were working
with the 𝑉𝑖 ’s instead of 𝑊𝑖 ’s, then we would have supp(𝜑) ⊂ 𝑈̄ 𝑖 instead of supp(𝜑) ⊂ 𝑈𝑖 as
required for a partition subordinate to the cover {𝑈𝑖 }.

Since {𝑊1 , … , 𝑊𝑛 } covers 𝑋, we have

∑
𝑛
𝜑(𝑥) ∶= 𝜑𝑖 (𝑥) > 0 for all 𝑥 ∈ 𝑋.
𝑖=1

Finally, for each 𝑖, we define

𝜑𝑖 (𝑥)
𝜌𝑖 (𝑥) ∶= .
𝜑(𝑥)
Chapter 9. Abstract Smooth Manifolds 181
Second case: 𝑋 ⊂ ℝ𝑁 and 𝑋 = 𝑋1 ∪ 𝑋2 ∪ 𝑋3 ∪ ⋯ where each 𝑋𝑖 is compact and
𝑋𝑖 ⊂ int(𝑋𝑖+1 ).

Let {𝑈𝛼 } be an open cover of 𝑋. For each 𝑖, we define

𝑈𝛼𝑖 ∶= 𝑈𝛼 ∩ (𝑋𝑖+1 ⧵ int(𝑋𝑖−2 )).

Then {𝑈𝛼𝑖 } is an open cover of 𝑌𝑖 ∶= 𝑋𝑖 ⧵ int(𝑋𝑖−1 ). Since int(𝑋𝑖−1 ) is an open subset, 𝑌𝑖

is a closed subset of 𝑋𝑖 and therefore 𝑌𝑖 is also compact. Then, for each 𝑖, the first case implies
that there is a partition of unity 𝜑𝑖𝛼 on 𝑌𝑖 subordinate to the cover {𝑈𝛼𝑖 }.

For each 𝑥 ∈ 𝑋, there is an 𝑖 such that 𝑥 ∈ 𝑋𝑖 and hence 𝜑𝑗𝛼 (𝑥) = 0 for all 𝑗 ≥ 𝑖 + 2. Hence,
for each 𝑥 ∈ 𝑋, the sum
∑
𝜑(𝑥) ∶= 𝜑𝑖𝛼 (𝑥)
𝛼,𝑖

is a finite sum in some open set containing 𝑥.

Now for each 𝛼 we define

𝜑𝑖𝛼 (𝑥)
𝜌𝑖𝛼 (𝑥) ∶=
𝜑(𝑥)
This is a partition of unity subordinate to the open cover {𝑈𝛼 }.

Third case: 𝑋 ⊂ ℝ𝑁 is open.

Define subsets

𝑋𝑖 ∶= {𝑥 ∈ 𝑋 ∶ |𝑥| ≤ 𝑖 and the distance to ℝ𝑁 ⧵ 𝑋 ≥ 1∕𝑗}.

Then these subsets satisfy:

∙ each 𝑋𝑖 is compact, since it is the intersection 𝑋 ∩ 𝐵𝑖 (0) ∩ (𝑋 ⧵ (∪𝑝∈ℝ𝑁 ⧵𝑋 𝐵1∕𝑖 (𝑝)) and
therefore closed and bounded in ℝ𝑁 ;

∙ for each 𝑖: 𝑋𝑖 ⊂ int(𝑋𝑖+1 );

∙ 𝑋 = 𝑋1 ∪ 𝑋2 ∪ ⋯.

Hence we can apply the second case.

Fourth case: 𝑋 ⊂ ℝ𝑁 arbitrary.

Let {𝑈𝛼 } be an open cover of 𝑋. By the definition of the subspace topology on 𝑋, for each
𝛼, there is a subset 𝑉𝛼 ⊂ ℝ𝑁 open in ℝ𝑁 such that 𝑈𝛼 = 𝑋 ∩ 𝑉𝛼 . Let 𝑌 be the union of all the
𝑉𝛼 in ℝ𝑁 . By the third case, there is a partition of unity on 𝑌 subordinate to the open cover
{𝑉𝛼 }. This is also a partition of unity on 𝑋 subordinate to the open cover {𝑈𝛼 }.

Last case: 𝑋 is a compact abstract smooth 𝑘-manifold.

Let {𝑉𝛼 } be an open over of 𝑋. By intersecting with the domains of charts on 𝑋, we get a
refinement of the cover. Hence we can assume that 𝑉𝛼 are the domains of charts on 𝑋. Since
182 9.5. Embedding Manifolds in Euclidean Space
𝑋 is compact, the domains of finitely many charts on 𝑋 suffice to cover 𝑋. Let us label them
(𝑉1 , 𝜙1 ), … , (𝑉𝑛 , 𝜙𝑛 ). Then each 𝑈𝑖 = 𝜙𝑖 (𝑉𝑖 ) is an open subset in ℝ𝑘 .

Now we can proceed exactly as in the case of a compact subspace in ℝ𝑁 for the finite
cover {𝑈1 , … , 𝑈𝑛 } of the space 𝑌 ∶= 𝑈1 ∪ ⋯ ∪ 𝑈𝑛 ⊂ ℝ𝑘 . This yields a partition of unity
{𝜌𝑖 } subordinate to the cover {𝑈1 , … , 𝑈𝑛 }. Composition of each 𝜌𝑖 with 𝜙𝑖 yields a partition
of unity {𝜌𝑖 ◦𝜙𝑖 } on 𝑋 subordinate to the cover {𝑉1 , … , 𝑉𝑛 }.
Chapter 9. Abstract Smooth Manifolds 183
9.6 Exercises and more examples

Exercise 9.1 Let 𝑋 be the set of all straight lines in ℝ2 (not just lines through the
origin).

(a) Show that 𝑋 is an abstract smooth 2-manifold by showing that we can identify 𝑋
with an open subset of the real projective plane ℝP2 .a

(b) Show that there is a bijection between 𝑋 and the set of equivalence classes

(𝕊1 × ℝ)∕ ∼

where ∼ is the equivalence relation defined by

(𝑠, 𝑥) ∼ (𝑦, 𝑡) ⇐⇒ 𝑡 = ±𝑠 and 𝑦 = 𝑥.

a
Here we use that open subsets of abstract smooth 𝑘-manifolds are again abstract smooth 𝑘-manifolds.

Exercise 9.2 Recall the Hopf map 𝜋 that we have seen previously: We consider 𝕊3 as
a subset of ℂ2 , i.e., 𝕊3 = {(𝑧, 𝑤) ∈ ℂ2 ∶ |𝑧|2 + |𝑤|2 = 1}, and 𝕊2 as a subset of ℂ × ℝ,
i.e., 𝕊2 = {(𝑧, 𝑥) ∈ ℂ × ℝ ∶ |𝑧|2 + 𝑥2 = 1}. Then the Hopf map 𝜋 is the map 𝕊3 → 𝕊2
given by ( )
𝜋(𝑧, 𝑤) = 2𝑧𝑤,̄ |𝑧|2 − |𝑤|2 .
In this exercise we study another way to define the Hopf map:
Let ℂP1 denote the complex projective space consisting of all complex-one-
dimensional linear subspaces in ℂ2 . We can consider ℂP1 as the set of pairs (𝑧, 𝑤) ≠
(0, 0) of complex numbers modulo the equivalence relation
(𝑧0 , 𝑤0 ) ∼ (𝑧1 , 𝑤1 ) ⇐⇒ 𝑧1 = 𝜆𝑧0 and 𝑤1 = 𝜆𝑤0 for some 𝜆 ∈ ℂ ⧵ {0}.
We consider ℂP1 as a topological space as the quotient (ℂ2 ⧵ {(0, 0)})∕ ∼ and write
[𝑧 ∶ 𝑤] for the equivalence class of (𝑧, 𝑤).
(a) Show that ℂP1 is a two-dimensional abstract smooth manifold.
Hint: Follow the outline for ℝP𝑛 .
(b) We consider 𝕊3 as a subset in ℂ2 . Show that ℂP1 also can be defined as a quotient
ℂP1 = 𝕊3 ∕ ∼𝑠 by an appropriate equivalence relation ∼𝑠 .
(c) We consider 𝕊3 as a subset in ℂ2 and define the map
𝜑 ∶ 𝕊3 → ℂP1 , (𝑧, 𝑤) → [𝑧 ∶ 𝑤].
Find a map ℎ ∶ ℂP1 → 𝕊2 such that the composition
𝜑 ℎ
𝕊3 ←←←→
← ℂP1 ←←→
← 𝕊2
equals 𝜋.
Hint: Think of the stereographic projection.
(c) Let [𝑧0 , 𝑤0 ] be a point in ℂP1 . Describe the fiber 𝜑−1 ([𝑧0 ∶ 𝑤0 ]).
184 9.6. Exercises and more examples
In the next exercise we let a group act on a space. If you are not familiar with group actions
yet, you may want to skip this exercise or first read about group actions in a textbook of your
choice.

Exercise 9.3 In this exercise we study examples of Hopf manifolds:

Let 0 < 𝜆 < 1 be a fixed real number. We let 𝑘 ∈ ℤ act on ℂ𝑛 ⧵ {0} by

(𝑧1 , … , 𝑧𝑛 ) → (𝜆𝑘 𝑧1 , … , 𝜆𝑘 𝑧𝑛 ).

The Hopf manifold 𝐻𝜆2𝑛 is defined as the quotient

𝐻𝜆2𝑛 ∶= (ℂ𝑛 ⧵ {0})∕ℤ,

i.e., we identify points 𝑧 and 𝜆𝑘 𝑧 for every 𝑘 ∈ ℤ.

We can show that 𝐻𝜆2𝑛 is a 2𝑛-dimensional abstract smooth manifold.a To focus on
the important features, we will restrict our study to the cases 𝑛 = 1 and 𝑛 = 2. Though
the general argument follows the same outline. We are also going to fix a 𝜆 and write
𝐻 2𝑛 for 𝐻𝜆2𝑛 .

(a) Show that 𝐻 2 is a 2-dimensional abstract smooth manifold.b

(b) Let 𝐴 = {𝑧 ∈ ℂ ∶ 1 ≤ |𝑧| ≤ 2} ⊂ ℂ. Show that 𝐻 2 homeomorphic to the

quotient 𝐴∕ℤ and deduce that 𝐻 2 is a compact space.

(c) Show that 𝐻 4 is a 4-dimensional abstract smooth manifold.c

Hint: It is the same argument as for 𝑛 = 1. We are just practising a bit more.

(d) Show that there is a homeomorphism between 𝐻 4 and 𝕊3 × 𝕊1 . Then show that
this actually a diffeomorphism by showing that the compositions with coordinate
charts are smooth.d

Note: The Hopf surface 𝐻𝜆4 is the simplest example of a compact complex manifold
which cannot be equipped with a Kähler metric and cannot be holomorphically embedded
into complex projective space. This follows from that fact that 𝐻𝜆4 ≅ 𝕊3 × 𝕊1 implies
that the first cohomology of a Hopf surface is one-dimensional, whereas every Kähler
manifold must have an even-dimensional first cohomology. This is a first glimpse at the
exciting theory of complex geometry.
a
Once we have learned more general theory, we can say that this follows from the fact that the action of
ℤ on ℂ𝑛 ⧵ {0} is free and discrete.
b
Note that 𝐻 2 is called the Hopf curve, since it is usually considered as a complex manifold of complex
dimension one.
c
Note that 𝐻 4 is called the Hopf surface, since it is usually considered as a complex manifold of complex
dimension two.
d
In general, 𝐻 2𝑛 is diffeomorphic to 𝕊2𝑛−1 × 𝕊1 . The previous point for 𝑛 = 1 is a special case of this
fact, since 𝐴∕ℤ ≅ 𝕊1 × 𝕊1 .
Chapter 9. Abstract Smooth Manifolds 185

Exercise 9.4 For 0 < 𝑚 ≤ 𝑛, the real Milnor hypersurface 𝐻(𝑚, 𝑛) is defined by
{ }
∑
𝑚
𝐻(𝑚, 𝑛) = ([𝑥0 ∶ … ∶ 𝑥𝑚 ], [𝑦0 ∶ … ∶ 𝑦𝑛 ]) ∈ ℝP𝑚 × ℝP𝑛 ∶ 𝑥𝑖 𝑦𝑖 = 0 .
𝑖=0

Show that 𝐻(𝑚, 𝑛) ⊂ ℝP𝑚 × ℝP𝑛 is an abstract smooth manifold of dimension 𝑚 + 𝑛 − 1.

Hint: Show that the subsets 𝑉𝑖𝑗 = {([𝑥0 ∶ … ∶ 𝑥𝑚 ], [𝑦0 ∶ … ∶ 𝑦𝑛 ]) ∈ 𝐻(𝑚, 𝑛) ∶

𝑥𝑖 ≠ 0, 𝑦𝑗 ≠ 0} are open and their union for all 𝑖 ≠ 𝑗 covers 𝐻(𝑚, 𝑛).
10. Manifolds with Boundary

10.1 Motivation: A first glimpse at intersection theory

Before we introduce the main new definition of this chapter, we look at an interesting and prob-
lem we like to solve.

Consider the two smooth manifolds 𝕊2 and ℝP2 , real projective 2-space. We have seen that
both are smooth manifolds of dimension 2, both are connected and compact. So one might
wonder if there is any topological feature that distinguishes these two spaces. In other words,
one might even wonder: are they homeomorphic or even diffeomorphic?

∙ (A new challenge) The answer is no: 𝕊2 and ℝP2 are not homeomorphic and
therefore not diffeomorphic. But how can we prove this?

To attack this problem, we need a new idea. For example, we could study the homotopy
classes of loops on 𝕊2 and ℝP2 , i.e., of continuous maps 𝕊1 → 𝑋 for 𝑋 being either 𝕊2 or ℝP2 .
If these two spaces were homeomorphic, the sets of of such homotopy classes of loops would
be the same. But we can show that the sets we get for 𝕊2 and ℝP2 , respectively, are different
and hence 𝕊2 ≉ ℝP2 :

∙ For 𝕊2 , every continuous map 𝕊1 → 𝕊2 is homotopic to a constant map. In fact, 𝕊2 is

a simply-connected space. One way to show this is to use the stereographic projection
and to use that ℝ2 is contractible. This requires, however, to show that every loop is
homotopic to a loop which does not pass through a given point on 𝕊2 . This is not difficult,
but takes some time to write down. We skip this for the moment, but we could do it!
∙ For ℝP2 , however, there is a map 𝑓̄ ∶ 𝕊1 → ℝP2 that is not homotopic to a constant map.

Let us see how such a map 𝑓̄ ∶ 𝕊1 → ℝP2 could look like: We define 𝑓̄ via the commutative
diagram
𝑓
[0, 2𝜋] / 𝕊2

𝜋
 
𝕊1 / ℝP2 = 𝕊2 ∕ ∼
𝑓̄

where 𝜋 is the quotient map, the left vertical map sends 𝑡 to (cos 𝑡, sin 𝑡) ∈ 𝕊1 and 𝑓 is defined
by
𝑓 ∶ [0, 2𝜋] → 𝕊2 , 𝑡 → (cos(𝑡∕2), sin(𝑡∕2), 0).
Since (cos 0, sin 0, 0) = (−1) ⋅ (cos 𝜋, sin 𝜋, 0), the map 𝜋◦𝑓 induces a continuous map
𝑓̄ ∶ [0, 2𝜋]∕(0 ∼ 2𝜋) = 𝕊1 → ℝP2 .

186
Chapter 10. Manifolds with Boundary 187
We can even show that 𝑓̄ is smooth: We show this locally for the induced map on local
coordinate charts. On ℝP2 we use the charts we defined previously. Then we can check:

∙ For the charts (0, 𝜋) → 𝕊1 and (𝜋, 2𝜋) → 𝕊1 defined by 𝑡 → (cos 𝑡, sin 𝑡), 𝑓̄ induces the
map ( )
cos(𝑡∕2)
𝑡 → ,0 .
sin(𝑡∕2)

∙ For the charts (𝜋∕2, 3𝜋∕2) → 𝕊1 and (3𝜋∕2, 5𝜋∕2) → 𝕊1 defined by 𝑡 → (cos 𝑡, sin 𝑡), 𝑓̄
induces the map ( )
sin(𝑡∕2)
𝑡 → ,0 .
cos(𝑡∕2)

All of these maps are smooth and hence 𝑓̄ is smooth. We admit that we cheat here a bit and
use a bit of the abstract manifold language. We hope this is OK for this motivating story.

But now we need to answer the question: How do we show that 𝑓̄ is not homotopic to a
constant map?

∙ (Intersection Theory will do it!) We will develop a theory of intersections up to

homotopy that will help us solve the problem. For we will be able to show: If 𝑔 ∶ 𝕊1 →
ℝP2 is any curve homotopic to 𝑓̄ so that 𝑔 −
⋔ Im (𝑓̄), then #𝑔 −1 (Im (𝑓̄) is an odd number.
In particular, we will have 𝑔 −1 (Im (𝑓̄)) ≠ ∅. However, the constant map 𝑐 ∶ 𝕊1 →
ℝP2 with image [1 ∶ 1 ∶ 1] has no intersection with Im (𝑓̄). Hence 𝑐 − ⋔ Im (𝑓̄), but
𝑐 (Im (𝑓̄)) = ∅. Thus 𝑓̄ cannot be homotopic to 𝑐.
−1

In order to obtain such a theory of intersections, we already have many tools available, for
example, homotopy and transversality. But we still lack one important piece of the puzzle:
manifolds with boundary. For allowing manifolds to have a boundary will make it possible to
include, in particular, the closed interval [0, 1] that we use in the definition of homotopies in
the category of manifolds.

10.2 Manifolds with Boundary

10.2.1 Manifolds with boundary - the definition

In order to be able to analyse a wider class of phenomena we would like to enlarge the class of
manifolds. A typical example which we would like to include is the domain of a homotopy
𝑋 × [0, 1] for a smooth 𝑘-dimensional manifold 𝑋. The points on 𝑋 × {0} and 𝑋 × {1} do not
have an open neighborhood which is diffeomorphic to ℝ𝑘 . Another example is the closed unit
ball in ℝ𝑘 .

So far such guys do not qualify as a manifold. From now on, we would like to allow such
subsets. This sounds like an laborious endeavour, but we will see that most of the theorems we
have proved so far are also valid for manifolds with boundary.
188 10.2. Manifolds with Boundary
The idea for what a manifold with boundary should be is the same as before: it is a space
which locally looks like some model space with boundary which we understand well. Hence
we need to choose a suitable new model space:

Definition 10.1 (New Euclidean model) The standard model of a Euclidean space
with boundary is the half-plane

ℍ𝑘 = {(𝑥1 , … , 𝑥𝑘 ) ∈ ℝ𝑘 ∶ 𝑥𝑘 ≥ 0}

in ℝ𝑘 . The boundary of ℍ𝑘 , denoted 𝜕ℍ𝑘 , is given by

𝜕ℍ𝑘 = {(𝑥1 , … , 𝑥𝑘 ) ∈ ℝ𝑘 ∶ 𝑥𝑘 = 0} = ℝ𝑘−1 × {0} ⊂ ℝ𝑘 .

Now a manifold with boundary is a space which locally looks like ℍ𝑘 :

Definition 10.2 (Manifolds with boundary) A subset 𝑋 of ℝ𝑁 is a smooth 𝑘-

dimensional manifold with boundary if every point 𝑥 of 𝑋 there is an open neighbor-
hood 𝑉 ⊂ 𝑋 containing 𝑥 and an open subset 𝑈 ⊂ ℍ𝑘 together with a diffeomorphism
𝜙 ∶ 𝑈 → 𝑉 . As before, any such a diffeomorphism is called a local parametrization
of 𝑋.
The boundary of 𝑋, denoted 𝜕𝑋, consists of those points that belong to the image
of the boundary of ℍ𝑘 under some local parametrization. Its complement is called the
interior of 𝑋, denoted Int(𝑋) = 𝑋 ⧵ 𝜕𝑋.
A manifold 𝑋 with 𝜕𝑋 = ∅ is just a smooth manifold in our previous terminology.
In order to make the distinction clear, if necessary, we call them also boundaryless
manifolds or manifolds without a boundary.

Example 10.3 (The closed unit interval) The unit interval [0, 1] is a one-dimensional
smooth manifold with boundary and the boundary consists of the two endpoints {0, 1}.
We can choose local parametrizations

𝜙0 ∶ [0, 1) → [0, 1], 𝑥 → 𝑥, and 𝜙1 ∶ [0, 1) → [0, 1], 𝑥 → 1 − 𝑥

defined on [0, 1) which is an open subset on ℍ1 = [0, ∞) ⊂ ℝ1 . The interior is the open
interval (0, 1).

Warning: The interior of 𝑋 ⊂ ℝ𝑁 as a manifold is in general different from the interior

of 𝑋 as a subspace of ℝ𝑁 . The interior of 𝑋 as a manifold is the complement of the boundary,
whereas the interior of the topological space 𝑋 is the union of all its open subsets. But also
every point in 𝜕𝑋 lies in some open neighborhood of 𝑋.

The interior is well-defined:

Let 𝑋 be a manifold with boundary. We need to check that our definition of points in the
interior and on the boundary is independent of the choice of a local parametrization.

So let 𝑥 ∈ 𝑋 be a point which is in the image of a local parametrization 𝜙 ∶ 𝑈 → 𝑉 ⊂ 𝑋

such that 𝑈 ⊂ ℍ𝑘 is an open set of ℍ𝑘 which is contained in the interior of ℍ𝑘 . Then ℝ𝑘 is
Chapter 10. Manifolds with Boundary 189

Figure 10.1: There are two types of open balls in ℍ𝑘 . The ones where all points satisfy 𝑥𝑘 > 0
are well-known. The ones which allow 𝑥𝑘 = 0 are new.

an open subset of ℝ𝑘 . Now assume 𝑥 is also in the image of another local parametrization
𝜙′ ∶ 𝑈 ′ → 𝑉 ′ ⊂ 𝑋. Then 𝑥 ∈ 𝑊 ∶= 𝑉 ∩ 𝑉 ′ ⊂ 𝑋, and the composition 𝜙′ ◦𝜙−1 ∶ 𝜙−1 (𝑊 ) →
(𝜙′ )−1 (𝑊 ) is a diffeomorphism. Hence, after possibly shrinking 𝑈 ′ , we see that 𝑈 ′ is also an
open subset in ℝ𝑘 . Thus 𝑥 is being an interior point is well-defined.

This shows in particular: if 𝑋 is a manifold with boundary, then the interior of 𝑋, Int(𝑋),
is a boundaryless manifold of the same dimension as 𝑋.

It remains to show that being a boundary point is also well-defined. We show this by
proving the following interest result:

Theorem 10.4 (Boundaries are manifolds) If 𝑋 is a 𝑘-dimensional manifold with

boundary, then 𝜕𝑋 is a (𝑘 − 1)-dimensional manifold without boundary.

Proof: Let 𝑥 ∈ 𝑋 and let 𝜙 and 𝜓 be two local parametrizations around 𝑥. After possibly
shrinking the domains and codomains, we can assume that 𝜙 ∶ 𝑈 → 𝑉 and 𝜓 ∶ 𝑊 → 𝑉 are
both diffeomorphisms from open sets 𝑈 ⊂ ℍ𝑘 , 𝑊 ⊂ ℍ𝑘 to the same open subset 𝑉 ⊂ 𝑋.

We would like to show 𝜙(𝜕𝑈 ) = 𝜓(𝜕𝑊 ). For then 𝜕𝑉 = 𝜙(𝜕𝑈 ) is independent of our
choice of local parametrization and therefore well-defined. Moreover, since 𝜕𝑈 = 𝑈 ∩𝜕ℍ𝑘 is an
open subset of ℝ𝑘−1 , we would get that every point 𝑦 ∈ 𝜕𝑋 is contained in a local parametriza-
tion 𝜙|𝜕𝑈 ∶ 𝑈 ∩ 𝜕ℍ𝑘 → 𝜕𝑋. This will show that 𝜕𝑋 is a manifold of dimension 𝑘 − 1.
190 10.2. Manifolds with Boundary
By our assumption on 𝜙 and 𝜓, it suffices to show 𝜓(𝜕𝑊 ) ⊂ 𝜙(𝜕𝑈 ). The other inclusion
will follow by symmetry. Hence we would like to show:

Claim: 𝜙−1 (𝜓(𝜕𝑊 )) ⊂ 𝜕𝑈 .

To simplify notation, we define the map 𝑔 ∶= 𝜙−1 ◦𝜓 ∶ 𝑊 → 𝑈 .

Suppose that the claim is false and there is a point 𝑤 ∈ 𝜕𝑊 which is mapped to an interior
point 𝑢 = 𝑔(𝑤) of 𝑈 by 𝑔. Since both 𝜙 and 𝜓 are diffeomorphisms, 𝑔 is a diffeomorphism
of 𝑊 onto an open subset 𝑔(𝑊 ) of 𝑈 . The chain rule implies that the derivative 𝑑(𝑔 −1 )𝑢 of its
inverse is bijective. But, since 𝑢 ∈ Int(𝑈 ), 𝑔(𝑊 ) contains a neighborhood of 𝑢 that is open in
ℝ𝑘 . Thus the Inverse Function Theorem, applied to the map 𝑔 −1 defined on this open subset
of ℝ𝑘 , implies that the image of 𝑔 −1 contains a neighborhood of 𝑤 that is open in ℝ𝑘 . This
contradicts the assumption 𝑤 ∈ 𝜕𝑊 .

10.2.2 Derivatives and tangent spaces vs boundaries

Tangent spaces and derivatives are still defined in the setting of manifolds with boundary.
Derivatives of smooth maps can be defined as before: Since smoothness at a point requires
a functions to be defined on open neighborhhod around that point, we need to be a bit more
careful at boundary points:

∙ Derivatives on ℍ𝑘

Suppose that 𝑔 is a smooth map of an open set 𝑈 of ℍ𝑘 to 𝑅𝑙 . If 𝑢 is an interior point of 𝑈 ,

then the derivative 𝑑𝑔𝑢 is defined as before.

If 𝑢 ∈ 𝜕𝑈 is a boundary point, the smoothness of 𝑔 means that it may be extended to a

smooth map 𝐺 defined in an open neighborhood of 𝑢 in ℝ𝑘 .

We define 𝑑𝑔𝑢 to be the derivative 𝑑𝐺𝑢 ∶ ℝ𝑘 → ℝ𝑙 .

Claim: This definition is independent of the choice of 𝐺.

So let 𝐺′ be another local extension of 𝑔. We need to show 𝑑𝐺𝑢′ = 𝑑𝐺𝑢 . The equality of
the two derivatives is no problem at points in the interior int(𝑈 ) of 𝑈 , because then we have
a small open neighborhood which is still in int(𝑈 ). We are going to use this and approximate
𝑢 be a sequence {𝑢𝑖 } of interior points 𝑢𝑖 ∈ int(𝑈 ) which converge to 𝑢.

Since 𝐺 and 𝐺′ agree with 𝑔 on int(𝑈 ), we have

𝑑𝐺𝑢𝑖 = 𝑑𝐺𝑢′ for all 𝑖.
𝑖

Since the derivative of a smooth map at a point depends continuously on changing the point,
this implies that 𝑑𝐺𝑢𝑖 → 𝑑𝐺𝑢 and 𝑑𝐺𝑢′ → 𝑑𝐺𝑢′ when 𝑢𝑖 → 𝑢 and both limits agree. Hence
𝑖
𝑑𝑔𝑢 is well-defined at boundary points too.

Note: It is important to observe that, at all points, 𝑑𝑔𝑢 is still a linear map of all of ℝ𝑘 to
ℝ𝑙 . For we have defined 𝑑𝑔𝑢 as the derivative 𝑑𝐺𝑢 of an extension 𝐺 to an open subset of ℝ𝑘 .

∙ Tangent spaces
Chapter 10. Manifolds with Boundary 191
Let 𝑋 ⊂ ℝ𝑁 be a smooth manifold with boundary, and 𝑥 ∈ 𝑋. Let 𝜙 ∶ 𝑈 → 𝑋 be a local
parametrization with 𝑈 ⊂ ℍ𝑘 open. Let 𝑢 ∈ 𝑈 be the point with 𝜙(𝑢) = 𝑥. Note that we cannot
assume 𝑢 = 0 when 𝑥 is an interior point. Then we have just learned that we can form the
derivative
𝑑𝜙𝑢 ∶ ℝ𝑘 → ℝ𝑁
no matter what kind of point 𝑥 is.

Definition 10.5 (Tangent spaces revisited) Let 𝑋 be a smooth 𝑘-manifold with

boundary. As before, we define the tangent space to 𝑋 at 𝑥, denoted 𝑇𝑥 (𝑋), to be
the image of ℝ𝑘 in ℝ𝑁 under the linear map 𝑑𝜙𝑢 .

It follows again from the Chain Rule that 𝑇𝑥 (𝑋) does not depend as a subspace of ℝ𝑁 on
the choice of 𝜙.

∙ Derivatives on tangent spaces

Now let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map between manifolds with boundaries with 𝑋 ⊂ ℝ𝑁

and 𝑌 ⊂ ℝ𝑀 . Given a point 𝑥 ∈ 𝑋. Then after choosing local parametrizations 𝜙 ∶ 𝑈 → 𝑋
with 𝜙(𝑢) = 𝑥 and 𝜓 ∶ 𝑉 → 𝑌 with 𝜓(𝑣) = 𝑓 (𝑥), then we define

𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 )

as the unique linear map which makes the following diagram commutative

𝑑𝑓𝑥
𝑇𝑥 (𝑋) / 𝑇 (𝑌 )
O 𝑦
O
𝑑𝜙𝑢 𝑑𝜓𝑣

ℝ𝑘 / ℝ𝑙 .
𝑑𝜃𝑢

where 𝜃 is the map 𝜓 −1 ◦𝑓 ◦𝜙 (note 𝑣 = 𝜃(𝑢)).

We see that things work out nicely so far for manifolds with boundaries. But:

Warning: Sometimes we do have to be careful when we apply our developed concepts

to manifolds with boundaries. For example, the product of two manifolds with boundary may
not be a manifold anymore.

Example 10.6 (Square is not a manifold) The square 𝑆 = [0, 1] × [0, 1] is not a
smooth manifold with boundary.
For, suppose 𝑆 was a smooth manifold with boundary. Then the corner 𝑠 = (0, 0)
has an open neighborhood 𝑉 ⊂ 𝑆 and there is a diffeomorphism 𝑓 ∶ 𝑈 → 𝑉 to an open
𝑈 ⊂ ℍ2 such that 𝑓 (𝜕𝑉 ) ⊂ 𝜕ℍ2 .
After shrinking 𝑉 if necessary, let 𝐹 ∶ 𝑉̃ → ℝ2 be a smooth extension of 𝑓 on a
subset 𝑉 ⊂ 𝑉̃ ⊂ ℝ2 open in ℝ2 . Then the derivative 𝑑𝐹𝑠 is an isomorphism, since 𝑓 is
a diffeomorphism. Writing 𝐹 = (𝐹1 , 𝐹2 ) we have 𝐹2 (𝑥, 0) = 0 = 𝐹2 (0, 𝑦) for any (𝑥, 0)
and (0, 𝑦) in 𝑈 , since 𝜕𝑉 is mapped to 𝜕ℍ2 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑦 = 0}. Thus taking partial
192 10.2. Manifolds with Boundary
derivatives yields
𝜕𝐹2 𝜕𝐹
(𝑠) = 0, and 2 (𝑠) = 0.
𝜕𝑥 𝜕𝑦
But this implies that 𝑑𝐹𝑠 (𝑒1 ) and 𝑑𝐹𝑠 (𝑒2 ) both lie in ℝ × {0} ⊂ ℝ2 , where 𝑒1 and 𝑒2
denote the first and second standard basis vector in ℝ2 , respectively. In particular, 𝑑𝐹𝑠 (𝑒1 )
and 𝑑𝐹𝑠 (𝑒2 ) are linearly dependent. This contradicts that 𝑑𝐹2 is an isomorphism and
therefore the existence of the diffeomorphism 𝑓 .

But if only one manifold has a boundary we are ok:

Lemma 10.7 (Products and Boundaries) The product of a manifold without bound-
ary 𝑋 and a manifold with boundary 𝑌 is a manifold with boundary. Furthermore,

𝜕(𝑋 × 𝑌 ) = 𝑋 × 𝜕𝑌 ,

and

dim(𝑋 × 𝑌 ) = dim 𝑋 + dim 𝑌 .

Proof: If 𝑈 ⊂ ℝ𝑘 and 𝑉 ⊂ ℍ𝑙 are open, then

𝑈 × 𝑉 ⊂ ℝ𝑘 × ℍ𝑙 = ℍ𝑘+𝑙

is open. Moreover, if 𝜙 ∶ 𝑈 → 𝑋 and 𝜓 ∶ 𝑉 → 𝑌 are local parametrizations, so is 𝜙 × 𝜓 ∶ 𝑈 ×

𝑉 → 𝑋 ×𝑌.

Example 10.8 (Closed cylinder) The closed cylinder 𝕊1 × [0, 1] is a manifold with
boundary, see Figure 10.2, and the boundary consists of

𝕊1 × {0} ∪ 𝕊1 × {1}.

Example 10.9 (Domain of a homotopy) More generally, if 𝑋 is a manifold without

boundary, then the domain of a homotopy 𝑋 × [0, 1] defined on 𝑋 is a manifold with
boundary.

10.2.3 Regular values and transversality

One of the most important concepts we have studied is transversality of smooth maps to
submanifolds. We would like to extend this to manifolds with boundary. This is possible, but
requires some care.

We start with the special case of regular values for functions on manifolds without boundary.
This is a well-known case, but it turns out that it actually produces manifolds with boundary as
follows:
Chapter 10. Manifolds with Boundary 193

Figure 10.2: A cylinder has the two outer circles as its boundary.

Lemma 10.10 (Regular values for real-valued functions) Suppose that 𝑆 is a mani-
fold without boundary and that 𝜋 ∶ 𝑆 → ℝ is a smooth function with regular value
0. Then the subset {𝑠 ∈ 𝑆 ∶ 𝜋(𝑠) ≥ 0} is a manifold with boundary, and the boundary
is 𝜋 −1 (0).

Proof: The set {𝑥 ∈ 𝑆 ∶ 𝜋(𝑥) > 0} is open in 𝑆, since it is the preimage of the open
subset (0, ∞) ⊂ ℝ under the continuous map 𝜋. It is therefore a submanifold of the same
dimension as 𝑆. Hence every point in {𝑥 ∈ 𝑆 ∶ 𝜋(𝑥) > 0} has an open neighborhood which
is diffeomorphic to an open subset of ℝ𝑘 , 𝑘 = dim 𝑆. It remains to study the preimage of 0.

So suppose that 𝑠 is a point with 𝜋(𝑠) = 0. By assumption, 0 is a regular value which

means that 𝑠 is a regular point of 𝜋. Hence, locally near 𝑠, we can show as before that 𝜋 is
equivalent to the canonical submersion. But for the canonical submersion
𝜋 ∶ ℍ𝑘 → ℝ, (𝑥1 , … , 𝑥𝑘 ) → 𝑥𝑘
the lemma just states the definition of the boundary of ℍ𝑘 :
𝜕ℍ𝑘 = 𝜋 −1 (0) = {(𝑥1 , … , 𝑥𝑘 ) ∈ ℝ𝑘 ∶ 𝑥𝑘 = 0}.

An immediate consequence of this fact is:

Example 10.11 (Spheres are boundaries) Let 𝜋 be the smooth function defined by
∑
𝜋 ∶ ℝ𝑘 → ℝ, (𝑥1 , … , 𝑥𝑘 ) → 1 − 𝑥2𝑖 .
𝑖

Then 0 is a regular value of 𝜋, and the closed unit ball 𝔻𝑘 in ℝ𝑘 can be described as
𝔻𝑘 = {𝑥 ∈ ℝ𝑘 ∶ 𝜋(𝑥) ≥ 0}.
The boundary of 𝔻𝑘 is the (𝑘 − 1)-sphere 𝕊𝑘−1 = 𝜋 −1 (0).
194 10.2. Manifolds with Boundary
10.2.4 Towards transversality

Now we try to enter the next level: Recall that transversality is formulated as a criterion on tan-
gent spaces and derivatives. We would like to formulate a similar criterion for maps between
manifolds with boundary.

As we learned above, the boundary 𝜕𝑋 of a 𝑘-manifold with boundary 𝑋 is a manifold

of dimension 𝑘 − 1 without boundary. Let 𝑥 ∈ 𝜕𝑋 be a point on the boundary. We have
dim 𝑇𝑥 (𝜕𝑋) = 𝑘 − 1 and dim 𝑇𝑥 (𝑋) = 𝑘. Moreover, since 𝜕𝑋 is a submanifold of 𝑋, we know
that
𝑇𝑥 (𝜕𝑋) ⊂ 𝑇𝑥 (𝑋)
is a vector subspace of codimension 1 in 𝑇𝑥 (𝑋) (recall that the latter means the difference of
dimensions equals 1).

Definition 10.12 (Notation: Restriction to boundary) For any smooth map 𝑓 ∶ 𝑋 →

𝑌 , we introduce the notation
𝜕𝑓 = 𝑓|𝜕𝑋
for the restriction of 𝑓 to 𝜕𝑋. The derivative of 𝜕𝑓 at 𝑥 is just the restriction of 𝑑𝑓𝑥 to
the subspace 𝑇𝑥 (𝜕𝑋):

𝑑(𝜕𝑓 )𝑥 = (𝑑𝑓𝑥 )|𝑇𝑥 (𝜕𝑋) ∶ 𝑇𝑥 (𝜕𝑋) → 𝑇𝑓 (𝑥) (𝑌 ).

Now let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map from a smooth manifold with boundary 𝑋 to a

boundaryless manifold 𝑌 , and let 𝑍 ⊂ 𝑌 be a submanifold. We would like to know under
which circumstances is 𝑓 −1 (𝑍) a submanifold with boundary of 𝑋, i.e., a subset of 𝑋 which
is itself a smooth manifold with boundary, such that we have control over the boundary in the
sense that we know
𝜕𝑓 −1 (𝑍) = 𝑓 −1 (𝑍) ∩ 𝜕𝑋. (10.1)

It turns out that it is not enough to ask that 𝑓 is transversal to 𝑍 in the previous sense, i.e.,
Im (𝑑𝑓𝑥 ) + 𝑇𝑓 (𝑥) (𝑍) = 𝑇𝑓 (𝑥) (𝑌 ).

Example 10.13 Even for the restriction of the canonical submersion

𝜋 ∶ ℍ2 → ℝ, (𝑥1 , 𝑥2 ) → 𝑥2
our previous condition for transversality is not sufficient to guarantee that equation (10.1)
is satisfied:
For, the derivative 𝑑𝜋(𝑥1 ,𝑥2 ) ∶ ℝ2 → ℝ is just the projection onto the second factor.
Hence it is surjective at every point (𝑥1 , 𝑥2 ). In particular, 0 is a regular value fo 𝜋.
Let 𝑍 ∶= {0}. Then
𝜋 −1 (𝑍) = {(𝑥1 , 𝑥2 ) ∈ ℝ2 ∶ 𝑥2 = 0} = 𝜕ℍ2 .
Since 0 is regular value, we know that 𝜋 −1 (𝑍) is a submanifold. The problem is that the
boundary does not satisfy condition (10.1):
𝜕𝜋 −1 (𝑍) = ∅, whereas 𝜋 −1 (𝑍) ∩ 𝜕𝑋 = 𝜕ℍ2 ≠ ∅.
Chapter 10. Manifolds with Boundary 195
In order to make sure that the boundary behaves well, we need to impose an additional
transversality condition on 𝜕𝑓 .

We start again with regular values:

Theorem 10.14 (With boundary: fibers of regular values) Let 𝑔 be a smooth map of
a 𝑘-manifold 𝑋 with boundary onto a boundaryless 𝑛-manifold 𝑌 , and suppose that
𝑦 ∈ 𝑌 is a regular value for both 𝑔 ∶ 𝑋 → 𝑌 and 𝜕𝑔 ∶ 𝜕𝑋 → 𝑌 . Then the preimage
𝑔 −1 (𝑦) is a (𝑘 − 𝑛)-dimensional manifold with boundary and

𝜕(𝑔 −1 (𝑦)) = 𝑔 −1 (𝑦) ∩ 𝜕𝑋.

Proof: To show that 𝑔 −1 (𝑦) is a manifold with boundary is a local question, i.e., it suffices
that each point in 𝑔 −1 (𝑦) has an open neighborhood which is a manifold with boundary. So
let 𝑥 ∈ 𝑋 be a point with 𝑔(𝑥) = 𝑦. After choosing local coordinates, we can assume that 𝑔 is
a map of the form
𝑔 ∶ ℍ𝑘 → ℝ𝑛 .

If 𝑥 is an interior point in 𝑋, then 𝑔 −1 (𝑦) is a manifold without boundary in an open

neighborhood around 𝑥 by the Preimage Theorem 4.6 for boundaryless manifolds.

So let us look at what happens if 𝑥 ∈ 𝜕𝑋. That 𝑔 is smooth at 𝑥 means by definition that
there is an open subset 𝑈 ⊂ ℝ𝑘 and a smooth map

𝐺 ∶ 𝑈 → ℝ𝑛 such that 𝐺𝑈 ∩ℍ𝑘 = 𝑔𝑈 ∩ℍ𝑘 .

After possibly replacing 𝑈 with a smaller subset, we can assume that 𝐺 has no critical
points in 𝑈 . Then 𝐺−1 (𝑦) is a smooth manifold by the Preimage Theorem 4.6 for boundaryless
manifolds. We need to show that

𝑔 −1 (𝑦) = 𝐺−1 (𝑦) ∩ ℍ𝐾 is a manifold with boundary.

In order to show this, we define a new smooth function 𝜋 on the manifold 𝑆 ∶= 𝐺−1 (𝑦)

𝜋 ∶ 𝑆 → ℝ, (𝑥1 , … , 𝑥𝑘 ) → 𝑥𝑘

as the projection to the last coordinate. Then

𝑆 ∩ ℍ𝑘 = {𝑠 ∈ 𝑆 ∶ 𝜋(𝑠) ≥ 0}.

Claim: 0 is a regular value of 𝜋.

If we can show the claim, then our previous lemma shows that 𝑆 ∩ ℍ𝑘 is a manifold with
boundary and the boundary is 𝜋 −1 (0).

To show the claim, assume there was an 𝑠 ∈ 𝑆 with both 𝜋(𝑠) = 0, i.e. 𝑠 ∈ 𝑆 ∩ 𝜕ℍ𝑘 , and
𝑑𝜋𝑠 = 0. We want to show that the assumption 𝑑𝜋𝑠 = 0 leads to a contradiction.

To do so, first note that 𝜋 is a linear map, and therefore 𝑑𝜋𝑠 = 𝜋. Thus,

𝑑𝜋𝑠 = 𝜋 ∶ 𝑇𝑠 (𝑆) → ℝ
196 10.2. Manifolds with Boundary
being trivial, just means that the last coordinate of every vector in 𝑇𝑠 (𝑋) is 0, i.e.

𝑑𝜋𝑠 = 0 ⇒ 𝑇𝑠 (𝑆) ⊂ 𝑇𝑠 (𝜕ℍ𝑘 ) = ℝ𝑘−1 .

Hence we want to show 𝑇𝑠 (𝑆) ⊄ ℝ𝑘−1 .

The tangent space to 𝑆 = 𝐺−1 (𝑦) at 𝑠 is the kernel of 𝑑𝐺𝑠 :

𝑇𝑠 (𝑆) = 𝑇𝑠 (𝐺−1 (𝑦)) = Ker (𝑑𝐺𝑠 = 𝑑𝑔𝑠 ∶ ℝ𝑘 → ℝ𝑛 )

where 𝑑𝑔𝑠 = 𝑑𝐺𝑠 by definition of 𝑑𝑔𝑠 .

We know that 𝑑(𝜕𝑔)𝑠 is the restriction of 𝑑𝑔𝑠 ∶ ℝ𝑘 → ℝ to ℝ𝑘−1 :

𝑑(𝜕𝑔)𝑠 = (𝑑𝑔𝑠 )|ℝ𝑘−1 .

Thus, if 𝑇𝑠 (𝑆) = Ker (𝑑𝑔𝑠 ) ⊆ ℝ𝑘−1 , then

Ker (𝑑𝑔𝑠 ) = Ker (𝑑(𝜕𝑔)𝑠 ). (10.2)

Now, finally, we apply the assumption of regularity of 𝑦. Since 𝑦 is a regular value of both
𝑔 and 𝜕𝑔, we know that both 𝑑𝑔𝑠 and 𝑑(𝜕𝑔)𝑠 are surjective. This implies

dim Ker (𝑑𝑔𝑠 ) = 𝑘 − 𝑛 and dim Ker (𝑑(𝜕𝑔)𝑠 ) = 𝑘 − 1 − 𝑛.

This contradicts assertion (10.2) about the kernels when Ker (𝑑𝑔𝑠 ) ⊂ ℝ𝑘−1 . Thus this
assumption must be false, i.e.,

𝑇𝑠 (𝑆) = Ker (𝑑𝑔𝑠 ) ⊈ ℝ𝑘−1

and hence 𝑑𝜋𝑠 ≠ 0 and therefore 𝑑𝜋𝑠 is surjective. In other words, 0 is a regular value.

Theorem 10.15 (Preimages of manifolds with boundary) Let 𝑓 be a smooth map

of a manifold 𝑋 with boundary onto a boundaryless manifold 𝑌 , and suppose that both
𝑓 ∶ 𝑋 → 𝑌 and 𝜕𝑓 ∶ 𝜕𝑋 → 𝑌 are transversal with respect to a boundaryless subman-
ifold 𝑍 in 𝑌 . Then the preimage 𝑓 −1 (𝑍) is a manifold with boundary

𝜕(𝑓 −1 (𝑍)) = 𝑓 −1 (𝑍) ∩ 𝜕𝑋,

and the codimension of 𝑓 −1 (𝑍) in 𝑋 equals the codimension of 𝑍 in 𝑌 .

Proof: The restriction of 𝑓 to the boundaryless manifold Int(𝑋) is transversal to 𝑍. Hence,

by the Preimage Theorem for boundaryless manifolds, 𝑓 −1 (𝑍) ∩ Int(𝑋) is a boundaryless
manifold of correct codimension. Thus it remains to examine 𝑓 −1 (𝑍) in a neighborhood of a
point 𝑥 ∈ 𝑓 −1 (𝑍) ∩ 𝜕𝑋.

Let 𝑙 ∶= codim 𝑍 in 𝑌 . As in the boundaryless case, we can choose a submersion ℎ ∶ 𝑊 →

ℝ𝑙 defined on an open neighborhood 𝑊 of 𝑓 (𝑥) in 𝑌 to ℝ𝑙 such that 𝑍 ∩ 𝑊 = ℎ−1 (0). Then
ℎ◦𝑓 is defined in a neighborhood 𝑉 of 𝑥 in 𝑋, and 𝑓 −1 (𝑍) ∩ 𝑉 = (ℎ◦𝑓 )−1 (0).
Chapter 10. Manifolds with Boundary 197
Now let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization around 𝑥, where 𝑈 is an open subset of ℍ𝑘 .
Then define
𝑔 ∶= ℎ◦𝑓 ◦𝜙 ∶ 𝑈 → ℝ𝑙 .

Since 𝜙 ∶ 𝑉 → 𝜙(𝑉 ) is a diffeomorphism, the set

𝑓 −1 (𝑍) is a manifold with boundary in a neighhorhood of 𝑥

⇐⇒ (𝑓 ◦𝜙)−1 (𝑍) = 𝑔 −1 (0) is a manifold with boundary near 𝑢 = 𝜙−1 (𝑥) ∈ 𝜕𝑈 .

But the transversality assumptions of 𝑓 and 𝜕𝑓 with respect to 𝑍 imply the 0 is a regular
value of 𝑔. Hence we can apply the previous theorem and we are done.

10.2.5 Sard’s Theorem with boundary

Finally, also Sard’s Theorem has a version with boundary.

Theorem 10.16 (Sard’s Theorem with boundary) For any smooth map 𝑓 ∶ 𝑋 → 𝑌
of a manifold 𝑋 with boundary to a manifold 𝑌 without a boundary, almost every
point of 𝑌 is a regular value of both 𝑓 and 𝜕𝑓 .

Proof: For any point 𝑥 ∈ 𝜕𝑋 on the boundary of 𝑋,

𝑑(𝜕𝑓 )𝑥 = (𝑑𝑓𝑥 )|𝑇𝑥 (𝜕𝑋) ∶ 𝑇𝑥 (𝜕𝑋) → 𝑇𝑓 (𝑥) (𝑌 ).

Hence if 𝑑(𝜕𝑓 )𝑥 is surjective, then 𝑑𝑓𝑥 is surjective. Hence if 𝜕𝑓 is regular at 𝑥, so is 𝑓 .

Thus a point 𝑦 ∈ 𝑌 fails to be a regular value of both 𝑓 and 𝜕𝑓 only if it is critical value
for 𝑓 ∶ Int(𝑋) → 𝑌 or 𝜕𝑓 ∶ 𝜕𝑋 → 𝑌 , i.e., if 𝑑𝑓𝑥 fails to be surjective for at least one 𝑥 ∈
𝑓 −1 (𝑦) ∩ Int(𝑋) or 𝑑(𝜕𝑓 )𝑥 fails to be surjective for at least one 𝑥 ∈ 𝑓 −1 (𝑦) ∩ 𝜕𝑋. But since
Int(𝑋) and 𝜕𝑋 are manifolds without boundary, both sets of critical values have measure
zero by Sard’s Theorem we have shown before. Thus the complement of the set of common
regular values for 𝑓 and 𝜕𝑓 is the union of two sets of measure zero. It is therefore itself a set
of measure zero.

10.3 One-Manifolds

The following theorem gives us a complete list of smooth one-dimensional manifolds. Note that
in general, since every manifold is the disjoint union of its connected components, it suffices to
classify connected manifold.

Theorem 10.17 (Classification of one-Manifolds) Let 𝑋 be a connected smooth one-

dimensional manifold.

(a) If 𝑋 is compact and without boundary, then 𝑋 is diffeomorphic to 𝕊1 .

(b) If 𝑋 is compact with boundary, then 𝑋 is diffeomorphic to [0, 1].

198 10.3. One-Manifolds
(c) If 𝑋 is noncompact without boundary, then 𝑋 is diffeomorphic (0, 1).

(d) If 𝑋 is noncompact with boundary, then 𝑋 is diffeomorphic to either [0, 1) or

(0, 1].

10.3.1 Some heuristics for the classification

Before we prove the theorem, we begin with the rough idea why it should be true.

(a): Assume 𝑋 is a nonempty, compact, connected 1-manifold. Each point has a neighbor-
hood diffeomorphic to (−1, 1). By compactness, finitely many such neighborhoods 𝑈1 , … , 𝑈𝑛
cover 𝑋. If 𝑛 was equal 1, then 𝑋 ≅ (−1, 1). But an open interval is not compact. Thus, there
must be at least two neighborhoods. Since 𝑋 is connected, these two charts must overlap.
The union of these two intervals has to be either an open interval (if they overlap on one side of
each) or a circle (if they overlap on both sides). But if their union is an open interval, there has
to be another chart, by the compactness of 𝑋. Since there are only finitely many 𝑈𝑖 ’s, we must
eventually arrive at the situation where the neighborhoods intersect on both sides and form a
circle. Then one has to use this to construct a diffeomorphism to 𝕊1 .

(b): Let 𝑋 be a compact, connected, one-dimensional smooth manifold with boundary.

Since 𝑋 has at least one boundary point, there must be a neighborhood in 𝑋 containing that
boundary point. This neighborhood must be diffeomorphic to [𝑎, 𝑏) for some 𝑎, 𝑏. Since this
interval is not compact, there must be another neighborhood in 𝑋. This neighborhood either
intersects another boundary point which would yield us 𝑋 ≅ [𝑎, 𝑐] for some 𝑐, or it does not
contain a boundary point. In the latter case, the union of the neighborhoods is diffeomorphic to
a half-open interval [𝑎, 𝑑) which is not compact. Hence there has to be another neighborhood.
Since 𝑋 is compact, this process will end after finitely many steps when we eventually get that
𝑋 is the union of neighborhoods which is diffeomorphic to a closed interval.

(c) and (d): When 𝑋 is not compact, we repeat the above processes. The difference is that
the process may not terminate and we end up with open or half-open intervals.

Figure 10.3: We can cover a compact one-manifold by finitely many open subsets looking like
an open interval. Either we get back to the beginning and get a circle or we stop with another
boundary point. This is the heuristic idea for the classification. There are some details to be
straightened out though.
Chapter 10. Manifolds with Boundary 199
10.3.2 Proof of the classification theorem

We will follow Milnor’s proof using arc-lengths given in the appendix of [13]. In this subsection
𝑋 will aways denote a smooth one-manifold.

Definition 10.18 (Parametrization by arc-length) Let 𝐼 be an interval, i.e., a con-

nected subset of ℝ. A map 𝑓 ∶ 𝐼 → 𝑋 is called a parametrization by arc-length
if 𝑓 maps 𝐼 diffeomorphically onto an open subset of 𝑋, and if the velocitiy vector
𝑑𝑓𝑡 (1) ∈ 𝑇𝑓 (𝑡) 𝑋 has length one for each 𝑡 ∈ 𝐼, where 𝑑𝑓𝑡 ∶ ℝ = 𝑇𝑡 𝐼 → 𝑇𝑓 (𝑡) 𝑋.

Note that the definition forces that 𝐼 can have a boundary only if 𝑋 has a boundary. The
proof of the classification theorem relies on the folllowing lemma:

Lemma 10.19 (Number of components) Let 𝑓 ∶ 𝐼 → 𝑋 and 𝑓 𝑔 ∶ 𝐽 → 𝑋 be two

parametrizations by arc-length. Then 𝑓 (𝐼) ∩ 𝑔(𝐽 ) has at most two components. If it has
only one component, then 𝑓 can be extended to a parametrization by arc-length of the
union 𝑓 (𝐼) ∪ 𝑔(𝐽 ). If it has two components, then 𝑋 must be diffeomorphic to 𝕊1 .

Proof: We write 𝑈 ∶= 𝑓 (𝐼) ∩ 𝑔(𝐽 ) ⊂ 𝑋. Note that 𝑈 is open in 𝑋, since both 𝑓 (𝐼) and
𝑔(𝐽 ) are open. In order to determine the number of components of 𝑈 we define the map
𝑓| (𝑔| )−1
𝜅 ∶ 𝑓 −1 (𝑈 ) ←←←→ ← 𝑔 −1 (𝑈 )
← 𝑈 ←←←←←←←←←→
where 𝑓| and 𝑔| denote the restrictions of 𝑓 and 𝑔, respectively, such that the maps are defined.
By definition of parametrizations by arc-length, 𝜅 is a local diffeomorphism with derivative
equal to ±1 and maps open subset of 𝐼 diffeomorphically onto open subsets of 𝐽 . In particular,
it is an open map. We consider the graph of 𝜅, denoted by Γ ⊂ 𝐼 × 𝐽 , given by
Γ = {(𝑠, 𝑡) ∈ 𝐼 × 𝐽 ∶ 𝑓 (𝑠) = 𝑔(𝑡)}.
Since Γ is the preimage of the diagonal Δ in 𝑋 × 𝑋 under the map (𝑓 , 𝑔) ∶ 𝐼 × 𝐽 → 𝑋 × 𝑋
and since Δ is closed in 𝑋 × 𝑋, Γ is a closed subset of 𝐼 × 𝐽 . Since 𝑑𝜅(𝑠,𝑡) = ±1, 𝑑𝜅(𝑠,𝑡) takes
values in the discrete set {±1}. Hence 𝑑𝜅(𝑠,𝑡) is locally constant and is therefore constant on
each connected component of Γ. Thus, thinking of 𝐼 × 𝐽 as a rectangle with sides 𝐼 and 𝐽 , Γ
consists of line segments of slope either +1 or −1. We need to understand how many such line
segmens there are. Note that the image of 𝜅 equals the projection of the components of Γ to 𝐽 .
Now let 𝐾 be a connected component of Γ.

∙ Claim: The endpoints of 𝐾 lie on the boundary of 𝐼 × 𝐽 .

Proof: Since Γ is closed, 𝐾 is closed in 𝐼 × 𝐽 as well. Assume that 𝐾 ends in the interior
of 𝐼 × 𝐽 with endpoint 𝑎. Then the projection of 𝐾 onto 𝐽 has endpoint 𝜅(𝑎). But since 𝜅 is a
local diffeomorphism, the image of 𝜅 has to contain an open neighborhood around 𝜅(𝑎). Hence
𝜅(𝑎) cannot be an endpoint of a subinterval of 𝐽 .

Moreover, since 𝜅 is one-to-one, each of the four edges of the rectangle 𝐼 × 𝐽 can contain
at most one endpoint of a component of Γ. Since each component has two endpoints, Γ can
have at most two components. This proves the first assertion in the lemma.
200 10.3. One-Manifolds
Now we show the second by considering the two cases:

∙ Assume Γ has one component: Then 𝜅 has constant slope and Γ is just a linesegment in
the rectangle 𝐼 × 𝐽 . Hence we can extend 𝜅 to a linear map 𝐿 ∶ ℝ𝑡𝑜ℝ. Since both 𝑓 , 𝑔
and 𝐿 are smooth and have equal derivatives, we can glue 𝑓 and 𝑔◦𝐿𝐿−1 (𝐽 ) together to
get an extension
𝐹 ∶ 𝐼 ∪ 𝐿−1 (𝐽 ) → 𝑓 (𝐼) ∪ 𝑔(𝐽 )
which is a parametrization by arc-length of the union 𝑓 (𝐼) ∪ 𝑔(𝐽 ).

∙ Assume Γ has two components: Then all four edges are hit by the components. This
implies that the slope of both components must be the same. Let us assume the slope is
+1. The case with slope −1 is analogous. Let 𝑎 < 𝑑 be the endpoints of 𝐼 and 𝛾 < 𝛽
be the endpoints of 𝐽 . Let 𝛼 ∈ 𝐽 be the starting point and 𝑏 ∈ 𝐼 be the endpoint of the
first component of Γ, and let 𝑐 ∈ 𝐼 be the starting point and 𝛿 ∈ 𝐽 be the endpoint of
the second component. After possibly translating the interval 𝐽 in ℝ we can assume that
𝛾 = 𝑐 and 𝛿 = 𝑑. By the previous arguments, we then have

𝑎 < 𝑏 ≤ 𝑐 = 𝛾 < 𝑑 = 𝛿 ≤ 𝛼 < 𝛽.

Figure 10.4: The case when Γ has two components. The slope is the same for both components,
here +1. All four edges of the rectangle 𝐼 × 𝐽 are hit by Γ.

Now we can define a map ℎ ∶ 𝕊1 → 𝑋 as follows: We define the parameter 𝜃 ∶= 2𝜋𝑡∕(𝛼−

𝑎) and define {
𝑓 (𝑡) for 𝑎 < 𝑡 < 𝑑,
ℎ(cos 𝜃, sin 𝜃) =
𝑔(𝑡) for 𝑐 < 𝑡 < 𝛽.
We need to check that ℎ is actually well-defined. First, if 𝑡 varies between 𝑎 and 𝛽, 𝜃 takes
all values in [0, 2𝜋], since 𝛼 < 𝛽. Thus ℎ is defined on all points of 𝕊1 . Second, we need
to check what happens on the overlaps. We have two intervals on which ℎ is defined by
Chapter 10. Manifolds with Boundary 201
two apriori different terms: for 𝑡 between 𝑐 = 𝛾 and 𝑑 = 𝛿, and for 𝑡 between 𝛼 and 𝛽.
However, by definition of Γ, we have 𝑓 (𝑐 + 𝑟) = 𝑔(𝛾 + 𝑟) for 0 ≤ 𝑟 ≤ 𝑑 − 𝑐 = 𝛿 − 𝛾.
Similarly, we have 𝑓 (𝑎 + 𝑟) = 𝑔(𝛼 + 𝑟) for 0 ≤ 𝑟 ≤ 𝑏 − 𝑎 = 𝛽 − 𝛼.
Hence ℎ is well-defined and since 𝑓 and 𝑔 are smooth, ℎ is smooth. Moreover, ℎ is an
open map, since 𝑓 and 𝑔 are. Thus ℎ(𝕊1 ) is an open subset of 𝑋. Since ℎ is continuous
and 𝕊1 is compact, ℎ(𝕊1 ) compact and hence a closed subset of 𝑋, as 𝑋 is Hausdorff. As
𝑋 is connected, this implies that ℎ(𝕊1 ) = 𝑋. Since 𝑓 and 𝑔 are local diffeomorphisms,
so is ℎ. Thus, ℎ is a diffeomorphism.

Proof of the Classification Theorem 10.17: Let 𝑓 ∶ 𝐼 → 𝑋 be any parametrization by

arc-length. If there exists a parametrization by arc-length 𝑔 ∶ 𝐽 → 𝑋 with 𝐼 ∩ 𝐽 ≠ ∅, then we
can glue these parametrizations together to get a parametrization by arc-length 𝐼 ∪ 𝐽 → 𝑋 as
described in the previous lemma. Hence after possibly extending 𝑓 as far as possible to the
left and then as far as possible to the right, we can assume that 𝑓 is maximal in the sense that
it cannot be extended over any larger interval as a parametrization by arc-length. If 𝑋 is not
diffeomorphic to 𝕊1 , we will prove that 𝑓 is onto. This will imply that 𝑓 is a diffeomorphism,
since we already know it is injective and a local diffeomorphism.

If the open set 𝑓 (𝐼) ⊂ 𝑋 was not all of 𝑋, then there wouldbe a limit point 𝑥 of 𝑓 (𝐼) in
𝑋 ⧵ 𝑓 (𝐼). Then we can parametrize a neighborhood of 𝑥 by arc-length. The previous lemma
then implies that 𝑓 can be extended over a larger interval. This contradicts the assumption
that 𝑓 is maximal. The assertion of the theorem now follows from the comparison with the
different types of intervals in ℝ.

At least as interesting as the theorem are its consequences which are surprisingly rich. We
will begin to study them next.

10.4 Brouwer Fixed Point Theorem

Now we study the surprising consequences of the classification of one-manifolds.

10.4.1 Boundaries and retractions

Lemma 10.20 (Boundary of One-Manifold) The boundary of any compact one-

dimensional manifold with boundary consists of an even number of points.

Proof: Every compact one-manifold with boundary 𝑋 is the disjoint union of finitely many
connected components. Each component is diffeomorphic to a copy of [0, 1]. Hence the
boundary of each component consists of two points. The boundary of 𝑋 consists of these
finitely many pairs of points.

Definition 10.21 (Retraction) Let 𝑋 be a smooth manifold and 𝑍 ⊂ 𝑋 be a subman-

202 10.4. Brouwer Fixed Point Theorem
ifold. Then a retraction is a smooth map 𝑓 ∶ 𝑋 → 𝑍 such that 𝑓|𝑍 is the identity.

𝑍 _
id /𝑍
>

 𝑓
𝑋

There is an important restriction for the existence of such retractions for manifolds with
boundary:

Lemma 10.22 (No retraction onto boundary) If 𝑋 is any compact manifold with
boundary, then there is no retraction of 𝑋 onto its boundary.

Proof: Suppose there is such a smooth map 𝑔 ∶ 𝑋 → 𝜕𝑋 such that 𝜕𝑔 ∶ 𝜕𝑋 → 𝜕𝑋 is

the identity. By Sard’s Theorem, we can choose a regular value 𝑧 ∈ 𝜕𝑋 of 𝑔. Since 𝜕𝑔 is the
identity, all values in 𝜕𝑋 a regular for 𝜕𝑔. Hence 𝑧 is regular for both 𝑔 and 𝜕𝑔. By the Preimage
Theorem 10.15 for manifolds with boundary, we know that 𝑔 −1 (𝑧) is a submanifold of 𝑋 with
boundary given by
𝜕(𝑔 −1 (𝑧)) = 𝑔 −1 (𝑧) ∩ 𝜕𝑋.

Moreover, the codimension of 𝑔 −1 (𝑧) in 𝑋 equals the codimension of {𝑧} in 𝜕𝑋, namely
dim 𝑋 −1 as {𝑧} has dimension 0. Hence 𝑔 −1 (𝑧) is one-dimensional. Since it is a closed subset
in the compact manifold 𝑋, it is also compact.

By definition of 𝜕𝑔 as the restriction of 𝑔 to 𝜕𝑋, we have for just set-theoretic reasons

(𝜕𝑔)−1 (𝑧) = (𝑔|𝜕𝑋 )−1 (𝑧) = (𝑔◦𝑖)−1 (𝑧) = 𝑖−1 (𝑔(𝑧)) = 𝑔 −1 (𝑧) ∩ 𝜕𝑋

where 𝑖 ∶ 𝜕𝑋 → 𝑋 denotes the inclusion of 𝜕𝑋 into 𝑋.

Hence, using our previous observation, we have

(𝜕𝑔)−1 (𝑧) = 𝜕(𝑔 −1 (𝑧))

Since 𝜕𝑔 = Id𝜕𝑋 ,

{𝑧} = (𝜕𝑔)−1 (𝑧) = 𝜕(𝑔 −1 (𝑧)).

This contradicts Lemma 10.20 that the boundary 𝜕(𝑔 −1 (𝑧)) of the compact one-dimensional
manifold 𝑔 −1 (𝑧) consists of an even number of points.

10.4.2 Brouwer Fixed Point Theorem - smooth case

This theorem has a famous consequence:

Chapter 10. Manifolds with Boundary 203

Theorem 10.23 (Brouwer Fixed Point Theorem for smooth maps) Let 𝑓 ∶ 𝔻𝑛 → 𝔻𝑛
be a smooth map of the closed unit ball 𝔻𝑛 = {𝑥 ∈ ℝ𝑛 ∶ |𝑥| ≤ 1} ⊂ ℝ𝑛 into itself.
Then 𝑓 has a fixed point, i.e., there is an 𝑥 ∈ 𝐷𝑛 with 𝑓 (𝑥) = 𝑥.

Before we prove the theorem, let us have a look at dimension one, where the result is very
familiar:

Remark 10.24 (Familiar in dimension one) Note that we have seen this theorem for
𝑛 = 1 in Calculus 1: Every continuous map 𝑓 ∶ [0, 1] → [0, 1] has a fixed point. For
define the function 𝑔(𝑥) = 𝑓 (𝑥) − 𝑥 which is a continuous map from [0, 1] to itself. We
have 𝑔(0) ≥ 0 and 𝑔(1) ≤ 0, since 𝑓 (0) ≥ 0 and 𝑓 (1) ≤ 1. If 𝑔(0) = 0 or 𝑔(1) = 1, we
are done. And if 𝑔(0) > 0 and 𝑔(1) < 1, then the Intermediate Value Theorem implies
that there is an 𝑥0 ∈ (0, 1) with 𝑔(𝑥0 ) = 0, i.e., 𝑓 (𝑥0 ) = 𝑥0 .

Figure 10.5: In dimension one this is the Intermediate Value Theorem. The graph has to cross
the diagonal which consists of fixed points of 𝑓 .

Proof of Brouwer Fixed Point Theorem 10.23:

Suppose that there exists an 𝑓 without fixed points. We will show that such an 𝑓 would
allow us to construct a retraction 𝑔 ∶ 𝔻𝑛 → 𝜕𝔻𝑛 . But, since 𝔻𝑛 is compact, we have just
proved in Lemma 10.22 that such a retraction cannot exist.

So suppose 𝑓 (𝑥) ≠ 𝑥 for all 𝑥 ∈ 𝔻𝑛 . Then, for every 𝑥 ∈ 𝔻𝑛 , the two different points 𝑥
and 𝑓 (𝑥) determine a line. Let 𝑔(𝑥) be the point where the line segment starting at 𝑓 (𝑥) and
passing through 𝑥 hits the boundary 𝜕𝔻𝑛 . This defines a map 𝑔 ∶ 𝔻𝑛 → 𝜕𝔻𝑛 . See Figure 10.6.

If 𝑥 ∈ 𝜕𝔻𝑛 , then 𝑔(𝑥) = 𝑥 by construction of 𝑔. Hence 𝑔 ∶ 𝔻𝑛 → 𝜕𝔻𝑛 is the identity on

𝜕𝔻𝑛 . Thus, in order to show that 𝑔 is a retraction, it remains to show that 𝑔 is smooth.

So let us describe 𝑔(𝑥) explicitly. As a point on the line from 𝑓 (𝑥) to 𝑥, 𝑔(𝑥) can be written
in the form
𝑥 − 𝑓 (𝑥)
𝑔(𝑥) = 𝑥 + 𝑡𝑣, where 𝑣 ∶=
|𝑥 − 𝑓 (𝑥)|
for some real number 𝑡. Note that, since we assume 𝑥 ≠ 𝑓 (𝑥), the vector 𝑣 is always defined.
In fact, it is the unit vector pointing from 𝑓 (𝑥) to 𝑥. Moreover, since 𝑓 is smooth, 𝑣 depends
smoothly on 𝑥.
204 10.4. Brouwer Fixed Point Theorem

Figure 10.6: We construct 𝑔 by extending the line from 𝑓 (𝑥) to 𝑥 until we hit the boundary.
If this map existed for all 𝑥, then we would have found a retraction. That is a contradiction to
Lemma 10.22.

We need to calculate 𝑡 and show that 𝑡 depends smoothly on 𝑥. Since 𝑔(𝑥) is a point on
boundary of 𝔻𝑛 , we know |𝑔(𝑥)| = 1, and 𝑡 is determined by the equation

1 = |𝑔(𝑥)|2 = (𝑥 + 𝑡𝑣) ⋅ (𝑥 + 𝑡𝑣) = 𝑥 ⋅ 𝑥 + 2𝑡𝑥 ⋅ 𝑣 + 𝑡2 𝑣 ⋅ 𝑣

or, equivalently,

0 = (𝑣 ⋅ 𝑣)𝑡2 + (2𝑥 ⋅ 𝑣)𝑡 + 𝑥 ⋅ 𝑥 − 1. (10.3)

By definition of 𝑣, we know 𝑣 ⋅ 𝑣 = |𝑣|2 = 1. Since 𝑣 points from 𝑓 (𝑥) to 𝑥, we know that 𝑡

must be positive. Now we just need to find the positive solution of the quadratic equation (10.3)
for 𝑡 and get
√
−2𝑥 ⋅ 𝑣 +4(𝑥 ⋅ 𝑣)2 − 4(𝑥 ⋅ 𝑥 − 1)
𝑡=
√ 2
= −𝑥 ⋅ 𝑣 + (𝑥 ⋅ 𝑣)2 − 𝑥 ⋅ 𝑥 + 1

where (𝑥 ⋅ 𝑣)2 − 𝑥 ⋅ 𝑥 + 1 is positive, since 𝑥 ⋅ 𝑥 = |𝑥|2 ≤ 1 and (𝑥 ⋅ 𝑣)2 > 0. Since the scalar
products and square roots involved depend smoothly on 𝑥, we see that 𝑡 depends smoothly on
𝑥. Hence 𝑔 is smooth.

∙ Note that, for 𝑛 = 1, in the above proof we would have construct a map 𝑔 ∶ [0, 1] →
{0, 1} which would send 0 to 0 and 1 to 1. Such a map cannot be continuous by the
Intermediate Value Theorem.

10.4.3 Brouwer Fixed Point Theorem - continuous case

Actually, just as in the one-dimensional case, the theorem also holds for continuous maps:
Chapter 10. Manifolds with Boundary 205

Theorem 10.25 (Brouwer Fixed-Point Theorem for continuous maps) Every con-
tinuous map 𝐹 ∶ 𝔻𝑛 → 𝔻𝑛 has a fixed point.

Proof: The idea is to reduce this theorem to the statement on smooth maps by approxi-
mating 𝐹 by a smooth mapping. This is possible by Weierstrass’ Approximation Theorem,
an important result from Calculus. It applies as 𝔻𝑛 is compact and says:

Given 𝜖 > 0, there is a polynomial function 𝑄 ∶ 𝔻𝑛 → ℝ𝑛 with

|𝑄(𝑥) − 𝐹 (𝑥)| < 𝜀 for all 𝑥 ∈ 𝔻𝑛 .

However, it is possible that 𝑄 sends points in 𝔻𝑛 to points outside of 𝔻𝑛 . In order to

remedy this defect, we replace 𝑄 with
𝑄(𝑥)
𝑃 (𝑥) ∶= .
1+𝜀

Since |𝐹 (𝑥)| ≤ 1, this new polynomial 𝑃 satisfies:

(1 + 𝜀)|𝑃 (𝑥)| = |𝑄(𝑥)| ≤ |𝑄(𝑥) − 𝐹 (𝑥)| + |𝐹 (𝑥)| < 𝜀 + 1

where we apply the triangle inequality. Hence |𝑃 (𝑥)| ≤ 1 and 𝑃 is a map 𝔻𝑛 → 𝔻𝑛 . More-
over,

(1 + 𝜀)|𝑃 (𝑥) − 𝐹 (𝑥)| = |𝑄(𝑥) − (1 + 𝜀)𝐹 (𝑥)| = |𝑄(𝑥) − 𝐹 (𝑥) + 𝜀𝐹 (𝑥)|

≤ |𝑄(𝑥) − 𝐹 (𝑥)| + 𝜀|𝐹 (𝑥)| < 2𝜀

where we use that |𝐹 (𝑥)| ≤ 1. Since 1 + 𝜀 > 1, this shows

|𝑃 (𝑥) − 𝐹 (𝑥)| < 2𝜀. (10.4)

Now suppose that 𝐹 (𝑥) ≠ 𝑥 for all 𝑥 ∈ 𝔻𝑛 . Then the continuous function

𝔻𝑛 → 𝔻𝑛 , 𝑥 → |𝐹 (𝑥) − 𝑥|

must have a minimum 𝜇, since 𝔻𝑛 is compact. Since 𝐹 (𝑥) ≠ 𝑥 for all 𝑥, we must have 𝜇 > 0.

Now, for 𝜀 = 𝜇∕2, we choose polynomials 𝑄 and then 𝑃 as above. Since |𝐹 (𝑥) − 𝑥| ≥ 𝜇
for all 𝑥 ∈ 𝔻𝑛 , the triangle inequality yields

𝜇 ≤ |𝐹 (𝑥) − 𝑥| = |𝐹 (𝑥) − 𝑃 (𝑥) + 𝑃 (𝑥) − 𝑥|

≤ |𝐹 (𝑥) − 𝑃 (𝑥)| + |𝑃 (𝑥) − 𝑥|.

But by (10.4), we know

|𝐹 (𝑥) − 𝑃 (𝑥)| < 𝜇 for all 𝑥 ∈ 𝔻𝑛 .

Thus |𝑃 (𝑥) − 𝑥| > 0, and therefore 𝑃 (𝑥) ≠ 𝑥 for all 𝑥 ∈ 𝔻𝑛 .

Hence 𝑃 ∶ 𝔻𝑛 → 𝔻𝑛 is a smooth map from 𝔻𝑛 to itself without a fixed point. This

contradicts the statement on smooth maps and completes the proof.
206 10.4. Brouwer Fixed Point Theorem
10.4.4 Counter-example on an open ball

The assertion of Theorem 10.23 is not true for the open ball:

Let 𝔹𝑘1 (0) = {𝑥 ∈ ℝ𝑘 ∶ |𝑥| < 1} be the open ball in ℝ𝑘 . We define the map
𝑥
𝜑 ∶ 𝔹𝑘1 (0) → ℝ𝑘 , 𝑥 → √ .
1 − |𝑥|2
This is a smooth map with smooth inverse
𝑦
𝜑−1 ∶ ℝ𝑘 → 𝔹𝑘1 (0), 𝑦 → √
1 + |𝑦|2

Thus 𝜑 is a diffeomorphism 𝔹𝑘1 (0) → ℝ𝑘 .

The translation 𝑇 ∶ ℝ𝑘 → ℝ𝑘 , 𝑥 → 𝑥 + 1 does not have a fixed point. Hence the

composite map
𝜑−1 ◦𝑇 ◦𝜑 ∶ 𝔹𝑘1 (0) → 𝔹𝑘1 (0)
is a smooth map which does not have a fixed point. For, if it had a fixed point 𝑥, then

𝜑−1 (𝑇 (𝜑(𝑥))) = 𝑥 ⇒ 𝑇 (𝜑(𝑥)) = 𝜑(𝑥)

and 𝑇 had a fixed point, which is not the case.

10.4.5 Some interesting consequences

Brouwer’s Fixed-Point Theorem has many important applications. Here is one of them:

Theorem 10.26 (Brouwer Invariance of Domain) Let 𝑈 be an open subset of ℝ𝑛 ,

and let 𝑓 ∶ 𝑈 → ℝ𝑛 be a continuous injective map. Then 𝑓 (𝑈 ) is also open.

Instead of studying the proof of this theorem, let us note a consequence of this result. You
can read more about this story and the proof on Terence Tao’s blog.

Theorem 10.27 (Topological Invariance of Dimension) If 𝑛 > 𝑚, and 𝑈 is a nonempty

open subset of ℝ𝑛 , then there is no continuous injective map from 𝑈 to ℝ𝑚 . In partic-
ular, since a homeomorphism from ℝ𝑛 to ℝ𝑚 would be such a continuous injective map,
ℝ𝑛 and ℝ𝑚 are not homeomorphic whenever 𝑛 ≠ 𝑚.

Even though it sounds like an obvious fact, this is a rather deep theorem. Note that there
exist weird things like a continuous surjection from ℝ𝑚 to ℝ𝑛 for 𝑚 < 𝑛 due to variants of the
Peano curve construction. Hence often we have to be careful with our topological intuition.

Proof of Theorem 10.27: If there was such a continuous injective map from 𝑈 to ℝ𝑚 , then
≅
we could compose it with the embedding ℝ𝑚 ←←←→
← (ℝ𝑚 × {0}) ⊂ ℝ𝑛 . Hence the composite would
yield a continuous injective map from 𝑈 into ℝ𝑛 . By the theorem, the image would be both
open in ℝ𝑛 and lie in the subspace ℝ𝑚 × {0}. But no open subset of ℝ𝑛 can be contained in
Chapter 10. Manifolds with Boundary 207
ℝ𝑚 × {0}, since we must be able to fit at least a tiny open ball of
ℝ𝑛 into that subset and there
is no room for such a ball in the direction of the remaining 𝑛 − 𝑚 coordinates.

Note that invariance of domain and dimension for smooth injective maps is just a con-
sequence of the Inverse Function Theorem. But for maps which are just continuous and
injective, it is much harder to achieve.
208 10.5. Exercises and more examples
10.5 Exercises and more examples

10.5.1 Manifolds with boundary

Exercise 10.1 If 𝑈 ⊂ ℝ𝑘 and 𝑉 ⊂ ℍ𝑘 are open neighborhoods of 0, prove that there

exists no diffeomorphism of 𝑉 with 𝑈 .
Hint: Use the Inverse Function Theorem 3.4.

Exercise 10.2 Prove that if 𝑓 ∶ 𝑋 → 𝑌 is a diffeomorphism of manifolds with bound-

ary, then 𝜕𝑓 maps 𝜕𝑋 diffeomorphically onto 𝜕𝑌 .
Hint: Use the Inverse Function Theorem 3.4.

Exercise 10.3 We define the smooth maps

𝐹 ∶ ℝ × [−1∕2, 1∕2] → ℝ3 , (𝑡, 𝑠) → (cos 𝑡, sin 𝑡, 𝑠)

and

𝐺 ∶ ℝ × [−1∕2, 1∕2] → ℝ3 ,
(𝑡, 𝑠) → ((1 + 𝑠 cos(𝑡∕2)) cos 𝑡, (1 + 𝑠 cos(𝑡∕2)) sin 𝑡, 𝑠 sin(𝑡∕2)).

We define 𝑋 to be the image of 𝐹 in ℝ3 , and 𝑌 = 𝐺 to be the image of 𝐺 in ℝ3 .

(a) Show that 𝑋 is a 2-dimensional manifold with boundary whose boundary is diffeo-
morphic to the disjoint union of two copies of the unit circle. (Convince yourself
that 𝑋 is a cylinder obtained by starting with a rectangular surface and then glueing
two opposite edges together.)

(b) Show that 𝑌 is a 2-dimensional manifold with boundary whose boundary is diffeo-
morphic to just one copy of the unit circle. (Convince yourself that 𝑌 is a Möbius
band obtained by starting with a rectangular surface and then glueing two opposite
edges after twisting one edge once. If you do not get through all the formulae,
make sure you understand the answer visually at least.)

Exercise 10.4 Suppose that 𝑋 is a manifold with boundary and 𝑥 ∈ 𝜕𝑋. Let 𝜙 ∶ 𝑈 →
𝑋 be a local parametrization with 𝜙(0) = 𝑥, where 𝑈 is an open subset of ℍ𝑘 . Then
𝑑𝜙0 ∶ ℝ𝑘 → 𝑇𝑥 (𝑋) is an isomorphism. Define the upper halfspace 𝐻𝑥 (𝑋) in 𝑇𝑥 (𝑋) to
be the image of ℍ𝑘 under 𝑑𝜙0 , 𝐻𝑥 (𝑋) ∶= 𝑑𝜙0 (ℍ𝑘 ).
(a) Prove that 𝐻𝑥 (𝑋) does not depend on the choice of local parametrization.
(b) Show that there are precisely two unit vectors in 𝑇𝑥 (𝑋) that are perpendicular to
𝑇𝑥 (𝜕𝑋) and that one lies inside 𝐻𝑥 (𝑋), the other outside. The one in 𝐻𝑥 (𝑋) is
called the inward unit normal vector to the boundary, and the other is the outward
unit normal vector to the boundary. Denote the outward unit normal vector by 𝑛(𝑥).
(c) If 𝑋 ⊂ ℝ𝑁 , we consider 𝑛(𝑥) as an element in ℝ𝑁 and get a map 𝑛 ∶ 𝜕𝑋 → ℝ𝑁 .
Chapter 10. Manifolds with Boundary 209
Show that 𝑛 is smooth.

Exercise 10.5 Let 𝑋 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥 ≥ −1}, 𝑌 = ℝ and

𝑓 ∶ 𝑋 → 𝑌 , (𝑥, 𝑦) → 𝑥2 + 𝑦2 .

(a) What is the boundary of 𝑋? Show that 1 is a regular value of 𝑓 . Is 1 a regular

value of 𝜕𝑓 ?

(b) Determine 𝑓 −1 (1), 𝜕(𝑓 −1 (1)) and 𝑓 −1 (1)∩𝜕𝑋. Why does the answer not contradict
the assertion of the Preimage Theorem for manifolds with boundary?

10.5.2 Brouwer Fixed Point Theorem

Exercise 10.6 Prove the Theorem of Perron–Frobenius: Let 𝐴 be an 𝑛 × 𝑛-matrix 𝐴

with only nonnegative entries. Then 𝐴 has a real nonnegative eigenvalue.

Hint: We can assume that 𝐴 is invertible, otherwise 0 is an eigenvalue. Let 𝐴 also

denote the associated linear map of ℝ𝑛 , and consider the map 𝑣 → 𝐴𝑣∕|𝐴𝑣| restricted to
𝕊𝑛−1 → 𝕊𝑛−1 . Show that this maps the first quadrant

𝑄 = {(𝑥𝑙 , … , 𝑥𝑛 ) ∈ 𝕊𝑛−1 ∶ all 𝑥𝑖 ≥ 0}

into itself. Now use the fact that there is a homeomorphism 𝔹𝑛−1 → 𝑄 to get a continuous
map 𝔹𝑛−1 → 𝔹𝑛−1 .

Exercise 10.7 Here is an alternative and equivalent definition to our previous one: A
space 𝑋 is simply-connected if it is path-connected and every continuous map 𝑓 ∶ 𝕊1 →
𝑋 can be extended to a continuous map 𝐹 ∶ 𝔹2 → 𝑋 such that 𝐹|𝕊1 = 𝑓 . In this exercise
we are going to show that 𝑋 = ℝ2 ⧵ {(0, 0)} is not simply-connected using Brouwer’s
fixed point theorem.

(a) Consider the antipodal map 𝑎 ∶ 𝕊1 → 𝕊1 which sends 𝑝 to −𝑝. Composing with
the inclusion 𝑆 1 ⊂ 𝑋, we consider 𝑎 as a map 𝑓 ∶ 𝕊1 → 𝑋. Show that, if 𝑋 =
ℝ2 ⧵ {(0, 0)} was simply-connected, then this would induce a map

𝑔 ∶ 𝔹2 → 𝑋 → 𝕊1 → 𝔹2 .

(b) Still assuming 𝑋 = ℝ2 ⧵ {(0, 0)} was simply-connected, show that 𝑔 does not have
a fixed point.

(c) Conclude that 𝑋 = ℝ2 ⧵ {(0, 0)} is not simply-connected.

(d) Deduce from the previous point that the composed map

id
𝜄 ∶ 𝕊1 ←←←→
← 𝕊1 ⊂ ℝ2 ⧵ {(0, 0)}, 𝑝 → 𝑝,
210 10.5. Exercises and more examples

is not homotopic to the constant map 𝑐 ∶ 𝕊1 → ℝ2 ⧵ {(0, 0)} sending all points to
(1, 0).
Hint: Use that ℝ2 ⧵ {(0, 0)} is not contractible, since it is not-simply connected.
11. The Brouwer Degree modulo 2

11.1 The Brouwer Degree of maps modulo 2

We are now ready to define a very important and powerful invariant for smooth maps, the degree
modulo 2. We will see that despite its rather simple definition, the degree determines a lot of
the interesting properties of a map.

11.1.1 A well-defined invariant

The definition of the degree relies on the following fact:

Lemma 11.1 (Homotopy Lemma mod 2) Let 𝑋 and 𝑌 be smooth manifolds of the
same dimension, where 𝑋 is compact and without boundary. Let 𝑓 , 𝑔 ∶ 𝑋 → 𝑌 be two
smooth maps. Assume 𝑓 and 𝑔 are smoothly homotopic. If 𝑦 ∈ 𝑌 is a regular value
for both 𝑓 and 𝑔, then

𝑓 −1 (𝑦) = 𝑔 −1 (𝑦) mod 𝟐.

Proof: Let 𝐹 ∶ 𝑋 × [0, 1] → 𝑌 be a smooth homotopy between 𝐹0 = 𝑓 and 𝐹1 = 𝑔.

First we suppose that 𝑦 is a regular value for 𝐹 . By the Preimage Theorem 10.14 with
boundary, this implies 𝐹 −1 (𝑦) is a submanifold with boundary of 𝑋 × [0, 1] of dimension
dim 𝑋 × [0, 1] − dim 𝑌 = 1, since we assume dim 𝑋 = dim 𝑌 . Moreover, the boundary of
𝐹 −1 (𝑦) is

𝜕𝐹 −1 (𝑦) = 𝐹 −1 (𝑦) ∩ 𝜕(𝑋 × [0, 1]) = 𝐹0−1 (𝑦) × {0} ∪ 𝐹1−1 (𝑦) × {1}
= 𝑓 −1 (𝑦) × {0} ∪ 𝑔 −1 (𝑦) × {1}.

Since 𝑋 is compact, 𝑋 × [0, 1] is compact. Since {𝑦} is a closed subset of 𝑌 , 𝐹 −1 (𝑦) is closed
in 𝑋 × [0, 1]. This implies that 𝐹 −1 (𝑦) is compact. Hence Lemma 10.20, which follows from
the classification of compact one-manifolds, implies that 𝜕𝐹 −1 (𝑦) must have an even number
of points. Thus, computing mod 2 we get

#𝑓 −1 (𝑦) + #𝑔 −1 (𝑦) ≡ 0,

and hence

#𝑓 −1 (𝑦) ≡ #𝑔 −1 (𝑦) mod 𝟐.

Now suppose that 𝑦 is not a regular value of 𝐹 . By Lemma 4.15, the functions 𝑦′ →  #𝑓 −1 (𝑦′ )
and 𝑦′ → #𝑔 −1 (𝑦′ ) are both locally constant on the set of regular values 𝑦′ ∈ 𝑌 of 𝑓 and 𝑔,

211
212 11.1. The Brouwer Degree of maps modulo 2
respectively. Thus there is an open neighborhood 𝑈1 ⊂ 𝑌 of 𝑦 consisting only of regular values
of 𝑓 so that

#𝑓 −1 (𝑦′ ) = #𝑓 −1 (𝑦) for all 𝑦′ ∈ 𝑈1 .

Similarly, there is an open neighborhood 𝑈2 ⊂ 𝑌 of 𝑦 consisting only of regular values of 𝑔 so

that

#𝑔 −1 (𝑦′ ) = #𝑔 −1 (𝑦) for all 𝑦′ ∈ 𝑈2 .

The intersection 𝑈1 ∩ 𝑈2 is an open subset of 𝑌 . Hence, by Sard’s Theorem 7.1, there is a

regular value 𝑧 of 𝐹 in 𝑈1 ∩ 𝑈2 . By the first case we have #𝑓 −1 (𝑧) = #𝑔 −1 (𝑧) modulo 2 and we
get

#𝑓 −1 (𝑦) = #𝑓 −1 (𝑧) ≡ #𝑔 −1 (𝑧) = 𝑔 −1 (𝑦) mod 𝟐.

We will need an important special type of homotopy which preserves more information than
homotopies in general:

Definition 11.2 (Isotopy) An isotopy is a homotopy ℎ𝑡 in which each map ℎ𝑡 is a

diffeomorphism. We say that two diffeomorphisms are isotopic if they can be joined
by an isotopy. An isotopy is called compactly supported if the maps ℎ𝑡 are all equal to
the identity map when restricted to the complement of some fixed compact set.

The following result will allow us to move points on connected manifolds via a family of
diffeomorphisms. The fact that every map in the homotopy family is a diffeomorphism makes
it much easier to keep track of the orientation numbers at preimages.

Lemma 11.3 (Isotopy Lemma) Given any two points 𝑦 and 𝑧 in the connected man-
ifold 𝑌 , there exists a diffeomorphism ℎ ∶ 𝑌 → 𝑌 such that

∙ ℎ(𝑦) = 𝑧 and

∙ ℎ is isotopic to the identity.

Moreover, the isotopy may be taken to be compactly supported.

We proof Lemma 11.3 and an extension to finitely many points in Section 11.1.3.

Using the Homotopy Lemma 11.1 and the Isotopy Lemma 11.3 we can now prove the fol-
lowing key result:

Theorem 11.4 (Degree modulo 2 is well-defined) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map

from a compact manifold 𝑋 to a connected manifold 𝑌 and dim 𝑋 = dim 𝑌 . Assume
𝑦 and 𝑧 are regular values of 𝑓 . Then

#𝑓 −1 (𝑦) ≡ #𝑓 −1 (𝑧) mod 𝟐.

This common residue class #𝑓 −1 (𝑦) modulo 2 only depends on the homotopy class of
𝑓.
Chapter 11. The Brouwer Degree modulo 2 213
Proof: Let 𝑦 and 𝑧 be two regular values of 𝑓 . Since 𝑌 is connected, we can apply the
Isotopy Lemma 11.3 and find a diffeomorphism ℎ ∶ 𝑌 → 𝑌 which is isotopic to the identity and
with ℎ(𝑦) = 𝑧. Since ℎ is a diffeomorphism, 𝑧 is a regular value of the composite ℎ◦𝑓 . Since
ℎ is homotopic to the identity, the composite ℎ◦𝑓 is homotopic to 𝑓 . Hence the Homotopy
Lemma 11.1 implies that

#(ℎ◦𝑓 )−1 (𝑧) ≡ #𝑓 −1 (𝑧) mod 𝟐.

The left-hand fiber is given by

(ℎ◦𝑓 )−1 (𝑧) = 𝑓 −1 (ℎ−1 (𝑧)) = 𝑓 −1 (𝑦).

Hence we have

#(ℎ◦𝑓 )−1 (𝑧) = #𝑓 −1 (𝑦).

Thus we can conclude

#𝑓 −1 (𝑦) ≡ #𝑓 −1 (𝑧) mod 𝟐.

To prove the second assertion, assume that 𝑔 ∶ 𝑋 → 𝑌 is a smooth map which is smoothly
homotopic to 𝑓 . By Sard’s Theorem 7.1, there exists an element 𝑦 ∈ 𝑌 which is a regular
value for both 𝑓 and 𝑔. By the Homotopy Lemma 11.1, this implies

#𝑓 −1 (𝑦) ≡ #𝑔 −1 (𝑦) mod 𝟐.

This completes the proof of the theorem.

Thus the number #𝑓 −1 (𝑦) is independent of 𝑦 and only depends on the smooth homotopy
class of 𝑓 . Hence we make the following definition:

Definition 11.5 (Degree modulo 2) Let 𝑋 be a compact smooth manifold 𝑋 without

boundary and 𝑌 a connected smooth manifold with dim 𝑋 = dim 𝑌 . Then the mod
2 degree of 𝑓 , denoted deg2 (𝑓 ), is the mod 2 residue class of #𝑓 −1 (𝑦) for any regular
value 𝑦 of 𝑓 .

Note: The degree mod 2 is defined only when the range manifold 𝑌 is connected, the
domain 𝑋 is compact, and dim 𝑋 = dim 𝑌 . Whenever we write deg2 , we assume that these
assumptions are satisfied.

Now that we have the invariant deg2 , let us contemplate on what it is good for. First of all,
we note that there are upsides and downsides equipped to deg2 :

∙ The first good news is that deg2 (𝑓 ) is a powerful invariant as we will see soon.

∙ The second good news is that deg2 (𝑓 ) is easy to calculate: just pick any regular value 𝑦
for 𝑓 and count preimage points

deg2 (𝑓 ) = #𝑓 −1 (𝑦) mod 2.

214 11.1. The Brouwer Degree of maps modulo 2
∙ The bad news is that its power is limited. For example, the map

ℂ → ℂ, 𝑧 → 𝑧𝑛 ,

which wraps the circle 𝕊1 smoothly around 𝕊1 𝑛 times, has mod 2 degree zero if 𝑛 is
even, and one if 𝑛 is odd. For example, deg2 of the constant map 𝕊1 → 𝕊1 is equal to
deg2 of the squaring map 𝕊1 → 𝕊1 sending 𝑧 → 𝑧2 . Hence deg2 cannot distinguish
between many different maps. This is rather unfortunate.

∙ One may wonder why Theorem 11.4 does not contradict our assertion in Remark 4.17
when we discussed Milnor’s proof of the FTA that it is not sufficient for 𝑌 to be connected
that the number #𝑓 −1 (𝑦) is constant for all regular values. In the proof of Theorem 11.4
we compose 𝑓 with the diffeomorphism ℎ and then use that deg2 is homotopy-invariant.
However, the Homotopy Lemma 11.1 only shows that #𝑓 −1 (𝑦) modulo 2 is constant
within the homotopy class of 𝑓 . Hence we the fiber 𝑓 −1 (𝑦) can be empty for some regular
value, while it is non-empty for others. All we know is that the parity of #𝑓 −1 (𝑦) does
not change.

Lemma 11.6 (deg2 is additive) Let 𝑋 and 𝑍 be compact smooth manifolds without
boundary and let 𝑌 a connected smooth manifold with dim 𝑋 = dim 𝑍 = dim 𝑌 . Let
𝑓 ∶ 𝑋 ⊔ 𝑍 → 𝑌 be a smooth map. Then we have

deg2 (𝑓 ) = deg2 (𝑓|𝑋 ) + deg2 (𝑓|𝑍 ).

Proof: Let 𝑦 be a regular value for 𝑓 . Then deg2 (𝑓 ) = #𝑓 −1 (𝑦), and the number of elements
in #𝑓 −1 (𝑦) is the sum of the number of elements lying in 𝑋 and 𝑍, respectively. These two
numbers are deg2 (𝑓|𝑋 ) and deg2 (𝑓|𝑍 ), respectively.

11.1.2 Some applications of the mod 2 degree

There are many important and powerful applications of deg2 . Here we just have a glimpse at a
few of them.

∙ Application: Compact manifolds without boundary are not contractible

Let 𝑋 be a compact smooth manifold without boundary of dimension at least one. The
identity map on 𝑋 has degree one, since #(id−1 (𝑦)) = 1 for every 𝑦 ∈ 𝑋. Thus deg2 (id) = 1.
Now let 𝑐 ∶ 𝑋 → 𝑋 be a constant map with value 𝑥0 ∈ 𝑋. Then every 𝑥 ≠ 𝑥0 is a regular
value for 𝑐, since every point not in the image is regular. Thus deg2 (𝑐) = 0. Hence we can
conclude deg2 (id) ≠ deg2 (𝑐). Since deg2 is invariant under smooth homotopy, we have proven
the following interesting fact:

Theorem 11.7 A compact manifold without boundary of dimension at least one is

not smoothly contractible.
Chapter 11. The Brouwer Degree modulo 2 215
An immediate consequence of this result is a useful result we have proven previously. See
Lemma 10.22.

∙ (No retraction to the boundary) Then there is no smooth map 𝑓 ∶ 𝔻𝑛+1 → 𝕊𝑛

which leaves 𝕊𝑛 pointwise fix, i.e., we cannot have 𝑓|𝕊𝑛 = id𝕊𝑛 .

Proof: If such a map 𝑓 ∶ 𝔻𝑛+1 → 𝕊𝑛 existed, then we could define a map

𝐹 ∶ 𝕊𝑛 × [0, 1] → 𝕊𝑛 , 𝐹 (𝑥, 𝑡) = 𝑓 (𝑡𝑥).

This map would be a smooth homotopy between a constant map and the identity map which
contradicts Theorem 11.7.

∙ Application: Obstruction to extending maps to boundary

Theorem 11.8 (Boundary Theorem for deg2 ) Let 𝑌 be a connected smooth manifold.
Assume 𝑋 = 𝜕𝑊 is the boundary for some compact manifold 𝑊 and that 𝑓 ∶ 𝑋 →
𝑌 is a smooth map which can be extended to all of 𝑊 , i.e., there is a smooth map
𝐹 ∶ 𝑊 → 𝑌 with 𝐹|𝜕𝑊 = 𝑓 . Then deg2 (𝑓 ) = 0.

∙ Note that when 𝑊 is compact, then the closed subset 𝑋 = 𝜕𝑊 is also compact. Hence
deg2 (𝑓 ) is defined.

Proof: Let 𝐹 ∶ 𝑊 → 𝑌 be an extension of 𝑓 . By Sard’s Theorem 7.1, there is an element

𝑦 ∈ 𝑌 which is a regular value for both 𝑓 and 𝐹 . By the Preimage Theorem 10.14 with
boundary, this implies 𝐹 −1 (𝑦) is a submanifold of 𝑊 of dimension dim 𝑊 − dim 𝑌 = 1, since
we assume dim 𝑋 = dim 𝑌 and dim 𝑊 = dim 𝑋 + 1. Since 𝑊 is compact and since {𝑦}
is a closed subset of 𝑌 , 𝐹 −1 (𝑦) is closed in 𝑊 . This implies that 𝐹 −1 (𝑦) is compact. Hence
Lemma 10.20, which follows from the classification of compact one-manifolds, implies that
#𝜕(𝐹 −1 (𝑦)) must be even. Moreover, the boundary of 𝐹 −1 (𝑦) is

𝜕𝐹 −1 (𝑦) = 𝐹 −1 (𝑦) ∩ 𝜕𝑊 = 𝐹 −1 (𝑦) ∩ 𝑋 = (𝐹|𝑋 )−1 (𝑦)

= 𝑓 −1 (𝑦).

Thus, computing mod 2 we get

deg2 (𝑓 ) ≡ #𝑓 −1 (𝑦) = #𝜕(𝐹 −1 (𝑦)) ≡ 0.

This has a useful consequence:

∙ (Obstruction for extending maps) Let 𝑌 be a connected smooth manifold. Let 𝑊

be a compact manifold with dim 𝑊 = dim 𝑌 + 1 and 𝑓 ∶ 𝜕𝑊 → 𝑌 a smooth map. If
deg2 (𝑓 ) ≠ 0, then 𝑓 cannot be extended to a smooth map 𝑊 → 𝑌 on all of 𝑊 .
216 11.1. The Brouwer Degree of maps modulo 2
∙ Application: Existence of zeros for complex valued functions

Suppose that 𝑝 ∶ ℂ → ℂ is a smooth, as a map ℝ2 → ℝ2 , complex function and 𝑊 ⊂ ℂ

is a smooth compact region in the plane, i.e., a two-dimensional compact manifold with
boundary.

Question: Is there a 𝑧 ∈ 𝑊 with 𝑝(𝑧) = 0?

Assume that 𝑝 has no zeros on the boundary 𝜕𝑊 . Then

𝑝
∶ 𝜕𝑊 → 𝕊1
|𝑝|
is defined and smooth as a map of compact one-manifolds. Now if 𝑝 has no zeros inside 𝑊 ,
𝑝 𝑝
then |𝑝| is defined on all of 𝑊 , i.e., the map |𝑝| ∶ 𝜕𝑊 → 𝕊1 can be extended to a smooth map
𝑝
𝑊 → 𝕊1 on all of 𝑊 . If this is the case, we just learned that we must have deg2 ( |𝑝| ) = 0. See
Figure 11.1.

𝑝
Theorem 11.9 (Existence of zeros via deg2 ) If the mod 2 degree of ∶ 𝜕𝑊 → 𝕊1
|𝑝|
is nonzero, then the function 𝑝 has a zero inside 𝑊 .

( )
𝑝
∙ Note that calculating deg2 |𝑝| simply consists of picking a point 𝑧 ∈ 𝕊1 , we could
think of it as a direction, and just counting the number of times we find a 𝑤 ∈ 𝜕𝑊 with
𝑝(𝑤) = 𝑧, i.e., how often 𝑝(𝑤) points in the chosen direction.

∙ Theorem 11.9 says that this simple procedure tells us whether 𝑝 has a zero inside 𝑊 .

∙ Finally, if you have learned about Complex Analysis, then this should remind you of the
Residue Theorem and Cauchy’s formula.

∙ We look at some applications of Theorem 11.9 in the exercises.

11.1.3 Proof of the Isotopy Lemma

Proof of Lemma 11.3: For the proof, we call two points 𝑦 and 𝑧 isotopic if the statement of the
theorem holds. This defines an equivalence relation on the set of points in 𝑌 . The proof of the
lemma will consist in showing that each equivalence class is open. Since equivalence classes
are disjoint, this will show that 𝑌 is the disjoint union of open subsets. Since 𝑌 is connected,
this implies there can only be one equivalence class, i.e., the lemma holds for all points in 𝑌 .

To prove that the equivalence classes are open, we will first construct an isotopy ℎ𝑡 on ℝ𝑘
such that

∙ ℎ0 is the identity,

∙ each ℎ𝑡 is the identity outside some specified small ball around 0, and

∙ ℎ1 (0) is any desired point sufficiently close to 0.

Chapter 11. The Brouwer Degree modulo 2 217

Figure 11.1: The degree is able to detect a zero of the polynomial in the region 𝑊 .

Then we transport the isotopy to 𝑌 : Given 𝑦 ∈ 𝑌 , choose a local parametrization 𝜙 ∶ ℝ𝑘 →

𝑉 ⊂ 𝑌 with 𝜙(0) = 𝑦. Let 𝔹 be a small open ball around 0 such that ℎ𝑡 on ℝ𝑘 is the identity
outside 𝔹. Since 𝜙 is a diffeomorphism onto 𝑉 , 𝜙(𝔹) is an open subset in 𝑌 contained in 𝑉 .
Then the map 𝐻𝑡 ∶= 𝜙◦ℎ𝑡 ◦𝜙−1 extends to an isotopy on all of 𝑌 by defining 𝐻𝑡 to be the
identity outside 𝜙(𝔹). Then 𝐻𝑡 shows that 𝑦 is isotopic to all points in the open subset 𝜙(𝔹) of
𝑌 . Hence the isotopy-equivalence class of 𝑦 is open.

Now we construct the isotopy ℎ𝑡 on ℝ𝑘 :

∙ We begin with the case 𝑘 = 1. Given 𝜀 > 0, let 𝜌 be smooth function that vanishes
outside (−𝜀, 𝜀) and equals 1 at 0. If 𝑧 ∈ ℝ1 , we define

ℎ𝑡 (𝑥) = 𝑥 + 𝑡 ⋅ 𝜌(𝑥) ⋅ 𝑧.

Then ℎ𝑡 (𝑥) = 𝑥 if 𝑥 ∈ (−𝜀, 𝜀) or if 𝑡 = 0, and ℎ1 (0) = 𝑧. To show that ℎ𝑡 is an isotopy

we compute its derivative with respect to 𝑥:

ℎ′𝑡 (𝑥) = 1 + 𝑡 ⋅ 𝜌′ (𝑥) ⋅ 𝑧.

Since |𝜌′ (𝑥)| vanishes outside a compact set, it must be bounded. Hence as long as |𝑧|
is small enough, |𝑡 ⋅ 𝜌′ (𝑥) ⋅ 𝑧| < 1 for all 𝑡 ∈ [0, 1] and 𝑥 ∈ ℝ. Thus ℎ′𝑡 (𝑥) > 0 for all 𝑡
and 𝑥 and ℎ𝑡 is strictly increasing. By the Inverse Function Theorem 3.2 this implies
that its inverse function is also smooth. Hence ℎ𝑡 is a diffeomorphism of ℝ1 for all 𝑧
with |𝑧| small enough.

∙ Now let 𝑘 be arbitrary. We are going to use what we just learned from the case 𝑘 = 1.
Given any point 𝑝 in ℝ𝑘 near the origin, we may rotate the coordinate axes so that the
point lies on the first axis. Hence, writing ℝ𝑘 = ℝ1 × ℝ𝑘−1 , we can assume that 𝑝 has
coordinates of the form 𝑝 = (𝑧, 0) with 𝑧 ∈ ℝ1 .
First we choose a function 𝜌 as for 𝑘 = 1. Given 𝜀 > 0, let 𝜌 be smooth function that
vanishes outside (−𝜀, 𝜀) and equals 1 at 0. Now we choose a function 𝜎 ∶ ℝ𝑘−1 → ℝ
218 11.1. The Brouwer Degree of maps modulo 2
such that 𝜎(0) = 1 and 𝜎 is zero outside some small ball of radius 𝛿. We then define
ℎ𝑡 ∶ ℝ𝑘−1 → ℝ𝑘−1 as follows: for (𝑥, 𝑦) ∈ ℝ1 × ℝ𝑘−1 we set

ℎ𝑡 (𝑥, 𝑦) = (𝑥 + 𝑡 ⋅ 𝜎(𝑦) ⋅ 𝜌(𝑥) ⋅ 𝑧, 𝑦)

We need to check that ℎ𝑡 has the desired properties:

∙ We have ℎ𝑡 (𝑥, 𝑦) = (𝑥, 𝑦) unless |𝑥| < 𝜀, |𝑦| < 𝛿, and 𝑡 > 0.

∙ ℎ1 (0, 0) = (𝑧, 0) by the choice of 𝜎 and 𝜌.

∙ It remains to show that ℎ𝑡 is a diffeomorphism of ℝ𝑘 when |𝑧| is small. First, we

observe that ℎ𝑡 is one-to-one and onto: We can show as above that |𝑡 ⋅ 𝜎(𝑦) ⋅ 𝜌′ (𝑥) ⋅
𝑧| < 1 for all 𝑡 and 𝑥 if |𝑧| is small enough. Hence, for each fixed 𝑦 ∈ ℝ𝑘−1 , ℎ𝑡
restricts to a diffeomorphism of ℝ1 × {𝑦}. Since ℎ𝑡 is the identity on ℝ𝑘−1 , ℎ𝑡 must
be one-to-one and onto on all of ℝ𝑘 . Since ℎ𝑡 is smooth, it remains to check this its
inverse is also smooth: The derivative of ℎ𝑡 at (𝑥, 𝑦) ∈ ℝ1 × ℝ𝑘−1 is represented by
the following matrix in the standard basis:
′
⎛1 + 𝑡 ⋅ 𝜎(𝑦) ⋅ 𝜌 (𝑥) ⋅ 𝑧 ∗ ⋯ ∗⎞
⎜ 0 ⎟
⎜ ⋮ 𝐼𝑘−1 ⎟⎟
⎜
⎝ 0 ⎠

where 𝐼𝑘−1 is the (𝑘 − 1) × (𝑘 − 1)-identity matrix. When |𝑧| is small enough, we

can show as above that the left-hand corner entry 1 + 𝑡 ⋅ 𝜎(𝑦) ⋅ 𝜌′ (𝑥) ⋅ 𝑧 is always
positive. Hence the matrix always has strictly positive determinant. By the Inverse
Function Theorem 3.4 this implies that the inverse of ℎ𝑡 is smooth.

We can actually extend the assertion to any finite number of points. We will use this result
later.

Theorem 11.10 (Isotopy Theorem) Suppose that 𝑌 is a connected manifold of di-

mension bigger than 1, and let {𝑦1 , … , 𝑦𝑛 } and {𝑧1 , … , 𝑧𝑛 } be two sets of distinct
points in 𝑌 . Then there exists a diffeomorphism ℎ ∶ 𝑌 → 𝑌 which is isotopic to the
identity with

ℎ(𝑦1 ) = 𝑧1 , … , ℎ(𝑦𝑛 ) = 𝑧𝑛 .

Moreover, the isotopy may be taken to be compactly supported.

Proof of Theorem 11.10: The proof is by induction. The Isotopy Lemma 11.3 is the case
𝑛 = 1. Now we assume the assertion being true for 𝑛−1. Then we have a compactly supported
isotopy ℎ′𝑡 ∶ 𝑌 ⧵ {𝑦𝑛 , 𝑧𝑛 } → 𝑌 ⧵ {𝑦𝑛 , 𝑧𝑛 } such that ℎ′1 (𝑦𝑖 ) = 𝑧𝑖 for all 𝑖 < 𝑛 and ℎ′0 = Id.

Since dim 𝑌 > 1, the punctured manifold 𝑌 ⧵ {𝑦𝑛 , 𝑧𝑛 } is connected. Since the isotopy ℎ′𝑡
has compact support, there are open neighborhoods around 𝑦𝑛 and 𝑧𝑛 in 𝑌 on which the ℎ′𝑡 are
all equal to the identity. Hence we can extend the family ℎ′𝑡 to a family of diffeomorphisms of
𝑌 that fix those two points.
Chapter 11. The Brouwer Degree modulo 2 219
Now we apply the induction hypothesis again to the punctured manifold
𝑌 ⧵ {𝑦1 , … , 𝑦𝑛−1 , 𝑧1 , … , 𝑧𝑛−1 } and the points 𝑦𝑛 , 𝑧𝑛 .
Then we get a compactly supported isotopy ℎ′′𝑡 with ℎ′′1 (𝑦𝑛 ) = 𝑧𝑛 and ℎ′′0 = Id. By the same
argument as for ℎ′𝑡 , we can extend ℎ′′𝑡 to an isotopy on all of 𝑌 such that all ℎ′′𝑡 satisfy ℎ′′𝑡 (𝑦𝑖 ) = 𝑧𝑖
for all 𝑖 < 𝑛. Then
ℎ𝑡 ∶= ℎ′′𝑡 ◦ℎ′𝑡
is the desired isotopy.

11.2 Winding Numbers and the Borsuk–Ulam Theorem

Now we study an important application of the degree modulo 2 and prove a famous theorem.
First we introduce a useful new invariant.

11.2.1 Winding numbers modulo 2

Let 𝑋 be a compact, connected smooth manifold, and let

𝑓 ∶ 𝑋 → ℝ𝑛
be a smooth map. We assume dim 𝑋 = 𝑛 − 1.

Let 𝑧 be a point of ℝ𝑛 not lying in the image 𝑓 (𝑋). We would like to understand how
𝑓 (𝑥) winds around 𝑧. To do this, we look at the unit vector
𝑓 (𝑥) − 𝑧
𝑢(𝑥) = .
|𝑓 (𝑥) − 𝑧|
It points in the direction from 𝑧 to 𝑓 (𝑥) and has length one.

With 𝑧 fixed and 𝑥 varying, we can consider 𝑢 as a map

𝑢 ∶ 𝑋 → 𝕊𝑛−1 .

We would like to know how often this vector points in a given direction, i.e., how often
𝑢(𝑥) has a given value. We learned in Section 11.1.1 that the degree of 𝑢 is an invariant that
encodes this information:

The number #𝑢−1 (𝑦) modulo 2 is by definition deg2 (𝑢). This number is constant for regular
values 𝑦 of the map 𝑢. To be a regular value means that 𝑦 − 𝑧 hits 𝑓 (𝑋) transversally, or in
other words, the line through 𝑧 and 𝑦 must hit 𝑓 (𝑋) transversally. See Figure 11.2.

Definition11.11 (Winding number) We give this number a name and call it the
winding number of 𝑓 around 𝑧. We denote it by

𝑊2 (𝑓 , 𝑧) ∶= deg2 (𝑢).

Our goal is to prove the following famous result:

220 11.2. Winding Numbers and the Borsuk–Ulam Theorem

Figure 11.2: The winding number of 𝑓 around 𝑧: we count the number of times the vector
𝑓 (𝑥) − 𝑧 points in a given direction where we neglect tangential points. Note also that some
points on the graph may contribute multiple times. In the end we take our count modulo 2.

Theorem 11.12 (Borsuk–Ulam Theorem) Let 𝑓 ∶ 𝕊𝑘 → ℝ𝑘+1 ⧵ {0} be a smooth

map, and suppose that 𝑓 is odd, i.e., 𝑓 satisfies the symmetry condition

𝑓 (−𝑥) = −𝑓 (𝑥) for all 𝑥 ∈ 𝕊𝑘 . (11.1)

Then the winding number of 𝑓 equals one, i.e., 𝑊2 (𝑓 , 0) = 1.

In other words, any map that is anti-symmetric around the origin must wind around the
origin an odd number of times.

Aside: A we will see below, there is a nice interpretation of this result for the meteorologists
among us: At any given time, there are two antipodal points on the Earth that have the same
temperature and pressure. Assuming temperature and pressure vary smoothly on the surface
of the Earth.

Before we approach the proof, we observe that the Borsuk–Ulam theorem is equivalent to
the following assertion:

Theorem 11.13 (Equivalent formulation of the Borsuk–Ulam Theorem) If 𝑓 ∶ 𝕊𝑘 →

𝕊𝑘 is a map which sends antipodal points to antipodal points, i.e., 𝑓 (−𝑥) = −𝑓 (𝑥),
then deg2 (𝑓 ) = 1.

Proof: First assume Borsuk–Ulam Theorem 11.12 is true: Given a smooth map 𝑓 ∶ 𝕊𝑘 →
𝕊𝑘 with 𝑓 (−𝑥) = −𝑓 (𝑥), we can consider it as a map 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 ⊂ ℝ𝑘+1 . Then we have
𝑓 = 𝑓 ∕|𝑓 | and therefore

1 = 𝑊2 (𝑓 , 0) = deg2 (𝑓 ∕|𝑓 |) = deg2 (𝑓 ).

Chapter 11. The Brouwer Degree modulo 2 221
Now assume the assertion is true: Given a smooth map 𝑓 ∶ 𝕊𝑘 → ℝ𝑘+1 ⧵ {0} with 𝑓 (−𝑥) =
−𝑓 (𝑥), then 𝑓 ∕|𝑓 | is a well-defined smooth map 𝑓 ∕|𝑓 | ∶ 𝕊𝑘 → 𝕊𝑘 . Hence we have

1 = deg2 (𝑓 ∕|𝑓 |) = 𝑊2 (𝑓 , 0)

by definition of winding number of 𝑓 around 0.

Recall that we called real functions 𝑓 with 𝑓 (−𝑥) = −𝑓 (𝑥) odd. Hence, as a slogan, we
can remember the Borsuk–Ulam Theorem for a smooth map 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 as follows:

∙ (Borsuk–Ulam Theorem in a nutshell) If 𝑓 is odd, its degree is odd.

11.2.2 Winding numbers and boundaries

In order to prove the theorem, we first need to investigate the relationship of winding numbers
and boundaries:

Theorem 11.14 (Winding numbers and boundaries) Suppose that 𝑋 = 𝜕𝐷 is the

boundary of a compact manifold 𝐷 of dimension 𝑛 with boundary, and let 𝐹 ∶ 𝐷 → ℝ𝑛
be a smooth map extending 𝑓 ∶ 𝑋 → ℝ𝑛 , i.e., 𝜕𝐹 = 𝑓 . Suppose that 𝑧 is a regular
value of 𝐹 that does not belong to the image of 𝑓 .
Then 𝐹 −1 (𝑧) is a finite set, and

𝑊2 (𝑓 , 𝑧) = #𝐹 −1 (𝑧) mod 2.

In other words, 𝑓 winds 𝑋 around 𝑧 as often as 𝐹 hits 𝑧, at least modulo 2.

Proof:

∙ First case: 𝐹 −1 (𝑧) = ∅, i.e., #𝐹 −1 (𝑧) = 0.

In this case, the map

𝑓 (𝑥) − 𝑧
𝑢 ∶ 𝑋 = 𝜕𝐷 → 𝕊𝑛−1 , 𝑥 →
|𝑓 (𝑥) − 𝑧|

can be extended to a map

𝐹 (𝑥) − 𝑧
𝐷 → 𝕊𝑛−1 , 𝑥 →
|𝐹 (𝑥) − 𝑧|
since 𝐹 (𝑥) − 𝑧 is never 0. Hence by the Boundary Theorem 11.8,

𝑊2 (𝑓 , 𝑧) = deg2 (𝑢) = 0 mod 2.

∙ Second case: 𝐹 −1 (𝑧) ≠ ∅.

222 11.2. Winding Numbers and the Borsuk–Ulam Theorem
Since 𝐷 is compact and of dimension 𝑛, 𝐹 −1 (𝑧) is a zero-dimensional closed submanifold
of 𝐷, and hence compact and hence a finite set. Suppose

𝐹 −1 (𝑧) = {𝑦1 , … , 𝑦𝑚 }.

Then we can choose local parametrizations around each 𝑦𝑖 in 𝐷 and let 𝐵𝑖 be the image of a
closed ball in ℝ𝑛 around 𝑦𝑖 . See Figure 11.3. Since 𝑧 is a regular value, the Stack of Records
Theorem 4.14 shows that 𝐹 −1 (𝑧) is discrete and disjoint to 𝑋 = 𝜕𝐷. Thus we can choose the
radii small enough such that these balls satisfy

𝔹𝑖 ∩ 𝔹𝑗 = ∅ and 𝔹𝑖 ∩ 𝑋 = ∅ for all 𝑖 ≠ 𝑗, and 𝑖 = 1, … , 𝑚.

We define
𝑓𝑖 ∶= 𝐹|𝜕𝔹𝑖 ∶ 𝜕𝔹𝑖 → ℝ𝑛 .
to be the restriction of 𝐹 to 𝜕𝔹𝑖 .

Figure 11.3: We split the contributions of the 𝑓𝑖 and 𝑓 by considering them as restrictions of
𝐹.

Now we observe that the subset

𝐷̃ ∶= 𝐷 ⧵ (∪𝑖 Int(𝔹𝑖 ))

is a closed submanifold of 𝐷 with boundary

𝜕 𝐷̃ = 𝜕𝐷 ∪̇ 𝜕𝔹1 ∪̇ ⋯ ∪̇ 𝜕𝔹𝑚

the disjoint union of the boundaries of 𝐷 and the 𝔹𝑖 ’s.

By the choice of the 𝔹𝑖 ’s, we have 𝐹 −1 (𝑧) ∩ 𝐷̃ = ∅. Hence

𝐹 −1 (𝑧) ∩ 𝐷̃ = (𝐹|𝐷̃ )−1 (𝑧) = ∅.

Hence the winding number of 𝜕𝐹|𝐷̃ at 𝑧 is zero.

Since degrees and hence winding numbers are additive with respect to connected com-
ponents this yields

0 = 𝑊2 (𝜕𝐹|𝐷̃ , 𝑧) = 𝑊2 (𝑓 , 𝑧) + 𝑊2 (𝑓1 , 𝑧) + ⋯ + 𝑊2 (𝑓𝑚 , 𝑧) mod 2.

Chapter 11. The Brouwer Degree modulo 2 223
Since we are working modulo 2, this implies

𝑊2 (𝑓 , 𝑧) = 𝑊2 (𝑓1 , 𝑧) + ⋯ + 𝑊2 (𝑓𝑚 , 𝑧) mod 2.

Now it remains to show 𝑊2 (𝑓𝑖 , 𝑧) = 1 for each 𝑖 = 1, … , 𝑚. For then

∑
#𝐹 −1 (𝑧) = 𝑚 = 𝑊2 (𝑓𝑖 , 𝑧) = 𝑊 (𝑓 , 𝑧) mod 2.
𝑖

Since 𝑧 is a regular value, 𝑑𝐹𝑦𝑖 is an isomorphism (remember dim 𝐷 = 𝑛). Thus, by

the Inverse Function Theorem, we can choose the radius of 𝐵𝑖 small enough such that 𝐹|𝐵𝑖
is a diffeomorphism onto its image (which contains 𝑧). By continuity, this implies also that
𝑓𝑖 = 𝜕𝐹|𝔹𝑖 is one-to-one onto the boundary of 𝐹 (𝔹𝑖 ).

By possibly rescaling and translating, we are reduced to showing:

Let 𝔹 be the closed unit ball in ℝ𝑛 and 𝐹 ∶ 𝔹 → 𝔹 be a diffeomorphism. Let 𝑓 =

𝜕𝐹 ∶ 𝕊𝑛−1 → 𝕊𝑛−1 . Then

#𝐹 −1 (0) = 𝑊2 (𝑓 , 0) = 1 mod 2.

And this follows from 𝑊2 (𝑓 , 0) = deg2 (𝑓 ) = #𝑓 −1 (𝑣) = 1 for any 𝑣 ∈ 𝕊𝑛−1 .

11.2.3 Key lemma: Lifting self-maps of the circle

Now we are almost ready to attack the proof of the Borsuk–Ulam Theorem. The proof will
proceed by induction on the dimension. To treat the case 𝑘 = 1, we need one more ingredient
which we will discuss next. It will turn out, however, that the effort we put in proving the
following lemma will pay off later on.

Lemma 11.15 (Lifting self-maps of the circle) Let 𝑓 ∶ 𝕊1 → 𝕊1 be a smooth map,

and let 𝑝 ∶ ℝ → 𝕊1 ⊂ ℂ be the map defined by 𝑡 → exp(2𝜋𝑖𝑡) = 𝑒2𝜋𝑖𝑡 . Then we have:

∙ There exists a smooth map 𝑔 ∶ ℝ → ℝ such that the following diagram commutes
𝑔
ℝ /ℝ
𝑝 𝑝
 
𝕊1 / 𝕊1 .
𝑓

∙ The map 𝑔 satisfies

𝑔(𝑡 + 1) = 𝑔(𝑡) + 𝑞 for all 𝑡 ∈ ℝ for some fixed 𝑞 ∈ ℤ.

∙ deg2 (𝑓 ) = 𝑞 modulo 2, i.e., deg2 (𝑓 ) = 1 if and only if 𝑞 is odd.

Proof of Lemma 11.15:

224 11.2. Winding Numbers and the Borsuk–Ulam Theorem

Figure 11.4: We think of ℝ lying as a spiral above 𝕊1 . That makes it easier to imagine how we
lift 𝑓 to a map 𝑔, first by local lifts which then are patched together.
Chapter 11. The Brouwer Degree modulo 2 225
The idea is that, given any smooth map 𝑓 ∶ 𝕊1 → 𝕊1 , we can lift 𝑓 locally and then patch
the pieces together to get a smooth map, see Figure 11.4,

𝑔 ∶ ℝ → ℝ such that 𝑝(𝑔(𝑡)) = 𝑓 (𝑝(𝑡))

where 𝑝 is the covering map

𝑝 ∶ ℝ → 𝕊1 , 𝑡 → exp(2𝜋𝑖𝑡) = 𝑒2𝜋𝑖𝑡 .

If such a map 𝑔 exists, we must have

𝑝(𝑔(𝑡 + 1)) = 𝑓 (𝑝(𝑡 + 1)) = 𝑓 (𝑝(𝑡)) = 𝑝(𝑔(𝑡)) for all 𝑡.

Since 𝑝(𝑡1 ) = 𝑝(𝑡2 ) if and only if 𝑡1 − 𝑡2 ∈ ℤ, we get

𝑔(𝑡 + 1) − 𝑔(𝑡) ∈ ℤ.

Since the function 𝑡 → 𝑔(𝑡 + 1) − 𝑔(𝑡) is continuous and takes only values in the discrete space
ℤ, it is locally constant. Since ℝ is connected, the function must be constant. In other words,
for all 𝑡 ∈ ℝ, we have

𝑔(𝑡 + 1) = 𝑔(𝑡) + 𝑞 for some fixed 𝑞 ∈ ℤ.

This 𝑞 is a fixed integer depending only on 𝑓 . We can actually think of 𝑞 as measuring how
fast 𝑓 wraps 𝕊1 around itself.

∙ Now we should explain why such a lift 𝑔 exists:

We first consider the composite ℎ = 𝑓 ◦𝑝:

[0, 1] ℝ
𝑝 ℎ 𝑝
 
𝕊1 /' 𝕊1
𝑓

where we restrict 𝑝 to [0, 1] ⊂ ℝ. We make a choice of a preimage point 𝑠0 ∈ ℝ with 𝑝(𝑠0 ) =

ℎ(0). The restriction 𝑝𝑠0 of the covering map

𝑝𝑠0 ∶ (𝑠0 − 1∕2, 𝑠0 + 1∕2) → 𝕊1 ⧵ {𝑝(𝑠0 + 1∕2)}

is a diffeomorphism. Hence we can use the inverse 𝑝−1

𝑠0 to construct a local lift 𝑝𝑠0 ◦ℎ in a
−1

neighborhood 𝑈0 of 0 in ℝ small enough such that ℎ(𝑈0 ) ⊆ 𝕊1 ⧵ {𝑝(𝑠0 + 1∕2)} =∶ 𝕊1− :

𝑝−1
𝑠 ◦ℎ
𝑈0
0
/ℝ
]
ℎ 𝑝
𝑝 𝑝−1
𝑠 0
 & 
𝕊1 / 𝕊1 .
𝑓 −
226 11.2. Winding Numbers and the Borsuk–Ulam Theorem
Now we continue this procedure to construct enough local lifts to cover all of [0, 1]. That
this works can be shown as follows:1 We have just constructed a lift of ℎ on the interval [0, 𝑠]
for some 𝑠 with 0 < 𝑠 ≤ 1. Thus the set

𝐷 = {𝑡 ∈ [0, 1]|ℎ ∶ [0, 𝑡] → 𝕊1 can be lifted to a map

𝑔̃ ∶ [0, 𝑡] → ℝ beginning at 𝑠0 }

is not empty. Since 𝐷 is bounded by 1, it has a least upper bound 𝑑 ∈ [0, 1].

First claim: 𝑑 ∈ 𝐷.

To prove the claim, let 𝑈 be an open subset of 𝕊1 containing ℎ(𝑑) such that 𝑝 maps each
component of 𝑝−1 (𝑈 ) diffeomorphically onto 𝑈 . Since ℎ is continuous, there is an open
subset 𝑊 ⊂ [0, 1] containing 𝑑 with ℎ(𝑊 ) ⊂ 𝑈 . By definition of 𝐷 and 𝑑, there is a point
𝑠′ ∈ 𝑊 such that 0 < 𝑠′ < 𝑑 and 𝑠′ ∈ 𝐷. Since we can lift ℎ to 𝑔̃ ∶ [0, 𝑠′ ] → ℝ, we have a
̃ ′ ) ∈ ℝ. We also have 𝑔(𝑠
unique point 𝑔(𝑠 ̃ ′ ) ∈ 𝑝−1 (ℎ(𝑠′ )).

Let 𝑉 ⊂ ℝ denote the component of 𝑝−1 (𝑈 ) with 𝑔(𝑠 ̃ ′ ) ∈ 𝑉 . Since ℎ(𝑠′ ) and ℎ(𝑑) are both
in 𝑈 , we can use (𝑝|𝑉 )−1 to define a local lift (𝑝|𝑉 )−1 ◦ℎ ∶ [0, 𝑠′ ] → ℝ of ℎ on [𝑠′ , 𝑑]. Now the
previous lift of ℎ on [0, 𝑠′ ] and the new one on [𝑠′ , 𝑑] agree at 𝑠′ by construction. Hence we
can glue them together to get a lift 𝑔̃ ∶ [0, 𝑑] → ℝ of ℎ on the entire interval [0, 𝑑].2 Thus
𝑑 ∈ 𝐷 as claimed.

Second claim: 𝑑 = 1.

Suppose 𝑑 < 1. For the open set 𝑊 ⊂ [0, 1] as above, we have 𝑑 ∈ 𝑊 . Since 𝑊 is open,
it still contains the open interval (𝑑 − 2𝜀, 𝑑 + 2𝜀) for some small 𝜖 > 0. Then we could use the
above argument to extend the lift of ℎ to [0, 𝑑 + 𝜀]. Thus we would have 𝑑 + 𝜀 ∈ 𝐷 which
contradicts that 𝑑 is the least upper bound of 𝐷. Hence we must have 𝑑 = 1 as claimed.

Hence we have shown that there is a smooth map 𝑔̃ ∶ [0, 1] → ℝ such that
𝑔̃
[0, 1] /ℝ
𝑝 𝑝
 
𝕊1 / 𝕊1
𝑓

commutes. We then have

𝑝(𝑔(1))
̃ = 𝑓 (𝑝(1)) = 𝑓 (𝑝(0)) = 𝑝(𝑔(0))
̃

and hence
𝑔(1) = 𝑔(0) + 𝑞 for some fixed 𝑞 ∈ ℤ.
1
You find a proof of this fact in almost every textbook on Algebraic Topology in the section on covering spaces
and lifting of paths. Here we follow Vick’s argument in his book Homology Theory, page 88. You may want to have
a look at other books if you do not like this one.
2
Actually, this gives us only a continuous lift a priori. But we can compose with appropriate smooth bump
functions to smoothen out any possible singular points. Since the notation would get extremely annoying, we skip
this step.
Chapter 11. The Brouwer Degree modulo 2 227
Finally, we define 𝑔 ∶ ℝ → ℝ by setting
{
̃ for all 𝑡 ∈ [0, 1], and
𝑔(𝑡) = 𝑔(𝑡)
𝑔(𝑡 + 1) = 𝑔(𝑡) + 𝑞 for all 𝑡 ∈ ℝ.

∙ It remains to show deg2 (𝑓 ) = 𝑞 modulo 2:

First, note that if 𝑓 is not surjective, then we can pick a point 𝑦 ∉ 𝑓 (𝕊1 ). This 𝑦 is auto-
matically a regular value. Since #𝑓 −1 (𝑦) = 0, we must have deg(𝑓 ) = 0. In this case, we need
to have 𝑞 = 0, i.e., 𝑔(𝑡 + 1) = 𝑔(𝑡). For otherwise 𝑝◦𝑔 was surjective and hence 𝑓 would be
surjective.

Note that, since the stereographic projection map 𝕊1 ⧵ {𝑦} → ℝ is a diffeomorphism and
ℝ is contractible, this shows that 𝕊1 ⧵ {𝑦} is contractible. Hence 𝑓 is a map to a contractible
space and therefore homotopic to a constant map and has degree 0.

Now we assume that 𝑓 is surjective. Let 𝑦 be a regular value of 𝑓 . We have deg2 (𝑓 ) =

#𝑓 −1 (𝑦). Hence we must count the preimages of 𝑦. Let 𝑥 ∈ 𝑓 −1 (𝑦) and 𝑠 ∈ ℝ be such that
𝑝(𝑠) = 𝑥. We then have

𝑦 = 𝑓 (𝑥) = 𝑓 (𝑝(𝑠)) = 𝑝(𝑔(𝑠)) = 𝑝(𝑔(𝑠) + 𝑘) for every integer 𝑘.

And since 𝑝(𝑢) = 𝑝(𝑢′ ) if and only if 𝑢 − 𝑢′ ∈ ℤ, the points of the form 𝑔(𝑠) + 𝑘 are the
only points with 𝑝(𝑢) = 𝑦. Hence we need to count how often we have 𝑔(𝑡) = 𝑔(𝑠) + 𝑘 for
𝑡 ∈ [𝑠, 𝑠 + 1]. See Figure 11.5.

We assume that 𝑞 ≥ 0. If 𝑞 < 0, then we use a similar argument with −𝑞. By the Interme-
diate Value Theorem in Calculus, we know that the smooth function 𝑔|[𝑠,𝑠+1] ∶ [𝑠, 𝑠 + 1] → ℝ
takes the value 𝑔(𝑠) + 𝑘 for each integer 𝑘 ∈ {0, 1, … , 𝑞 − 1} an odd number of times and the
values 𝑔(𝑠) + 𝑘 for any other integer 𝑘 an even number of times.3 Hence, modulo 2, there are
𝑞 many points 𝑡 in the interval4 [𝑠, 𝑠 + 1) such that

𝑝(𝑔(𝑡)) = 𝑝(𝑔(𝑠)) = 𝑓 (𝑦).

Thus, modulo 2, there are 𝑞 many points in 𝑓 −1 (𝑦), and we have shown deg2 (𝑓 ) = 𝑞.

Remark 11.16 (ℝ is a covering space of 𝕊1 ) There is a deeper reason why this

works. For ℝ is a (universal) covering space of 𝕊1 , and continuous paths can always
be lifted to a covering space. You will learn more about this phenomenon later.

11.2.4 Proof of the Borsuk–Ulam Theorem

Now we are ready to prove Theorem 11.12. The proof will proceed by induction.

The case 𝑘 = 1:
3
Here we use that 𝑦 being regular implies det(𝑑𝑓𝑥 ) ≠ 0 and hence 𝑑𝑔(𝑡𝑘 ) ≠ 0 for all 𝑡𝑘 with 𝑔(𝑡𝑘 ) = 𝑔(𝑠) + 𝑘.
4
Remember that 𝑝(𝑔(𝑠 + 1)) = 𝑝(𝑔(𝑠)). Hence we only need to take one of 𝑠 and 𝑠 + 1 into account.
228 11.2. Winding Numbers and the Borsuk–Ulam Theorem

Figure 11.5: An application of important theorems of Calculus. We count how many times the
graph of 𝑔 hits integer values, marked by the horizontal lines. We know that if the graph drops
below a given value it has to pass it again to reach the value 𝑔(𝑠) + 𝑞 eventually. Counting
modulo 2 we get 𝑞 crossings of the horizontal lines.

By the previous observation, to show the theorem is equivalent to showing that a map
𝑓 ∶ 𝕊1 → 𝕊1 with 𝑓 (−𝑥) = −𝑓 (𝑥) has deg2 (𝑓 ) = 1. By the previous lemma, we can find a
smooth map 𝑔 ∶ ℝ → ℝ with 𝑓 ◦𝑝 = 𝑝◦𝑔 and 𝑔(𝑡 + 1) = 𝑔(𝑡) for some 𝑞 ∈ ℤ.

If 𝑓 is odd, then
𝑓 (𝑝(𝑡 + 1∕2)) = 𝑓 (−𝑝(𝑡)) = −𝑓 (𝑝(𝑡))
where we use 𝑝(𝑡 + 1∕2) = exp(2𝜋𝑖(𝑡 + 1∕2)) = − exp(2𝜋𝑖𝑡) = −𝑝(𝑡). Hence
𝑝(𝑔(𝑡 + 1∕2)) = 𝑓 (𝑝(𝑡 + 1∕2)) = −𝑓 (𝑝(𝑡)) = −𝑝(𝑔(𝑡)).
We also have
𝑝(𝑠1 ) = −𝑝(𝑠2 )
⇐⇒ exp(2𝜋𝑖𝑠1 ) = − exp(2𝜋𝑖𝑠2 )
⇐⇒ exp(2𝜋𝑖𝑠1 ) = exp(2𝜋𝑖𝑠2 + 𝜋𝑖)
⇐⇒ 𝑠1 = 𝑠2 + 𝑚∕2 for some odd 𝑚 ∈ ℤ.
This implies
𝑔(𝑡 + 1∕2) = 𝑔(𝑡) + 𝑚∕2 for some odd 𝑚 ∈ ℤ.
Applied to 𝑡 = 1∕2, this yields
𝑔(1) = 𝑔(1∕2 + 1∕2) = 𝑔(1∕2) + 𝑚∕2 = 𝑔(0) + 𝑚∕2 + 𝑚∕2 = 𝑔(0) + 𝑚.
Thus 𝑞 = 𝑚 is odd. Hence, by the previous lemma, we have deg2 (𝑓 ) = 𝑞 = 1 mod 2. This
finishes the case 𝑘 = 1.
Chapter 11. The Brouwer Degree modulo 2 229
Induction step: Assume the theorem is true for 𝑘 − 1 and 𝑘 ≥ 2. Let 𝑓 ∶ 𝕊𝑘 → ℝ𝑘+1 ⧵ {0}
satisfy the symmetry condition (11.1). We consider 𝕊𝑘−1 to be the equator of 𝕊𝑘 , embedded
by

(𝑥1 , … , 𝑥𝑘 ) → (𝑥1 , … , 𝑥𝑘 , 0).

The idea is to compute 𝑊2 (𝑓 , 0) by counting how often 𝑓 intersects a line 𝐿 in ℝ𝑘+1 . By

choosing 𝐿 disjoint from the image of the equator, we can use the induction hypothesis to
show that the equator winds around 𝐿 an odd number of times. Finally, it is easy to calculate
the intersection of 𝑓 with 𝐿 once we know the behavior of 𝑓 on the equator.

Let 𝑔 ∶ 𝕊𝑘−1 → ℝ𝑘+1 ⧵ {0} be the restriction of 𝑓 to the equator. By Sard’s Theorem, we
can choose a value 𝑦 ∈ 𝕊𝑘 which is regular for both smooth maps

𝑔 𝑓
∶ 𝕊𝑘−1 → 𝕊𝑘 , and ∶ 𝕊𝑘 → 𝕊𝑘 .
|𝑔| |𝑓 |

The symmetry condition implies that 𝑦 is regular for both these maps if and only if −𝑦 is
regular for both maps, since the derivatives at preimages of 𝑦 and −𝑦 just differ by multiplying
with (−1).
𝑔
Since dim 𝕊𝑘−1 < dim 𝕊𝑘 , the only way 𝑦 can be a regular value of |𝑔|
is when 𝑦 is not in
𝑔
the image. Hence neither 𝑦 nor −𝑦 are in the image of |𝑔|
.

Thus, for the line 𝐿 ∶= ℝ ⋅ 𝑦 = span(𝑦), we have

𝑦 is a regular value of 𝑔 ⇐⇒ Im (𝑔) ∩ 𝐿 = ∅.

𝑓
That 𝑦 is regular for |𝑓 |
means by definition

( ( ) )
𝑓
Im 𝑑 = 𝑇𝑦 (𝕊𝑘 ).
|𝑓 | 𝑥

The tangent space to 𝕊𝑘 at 𝑦 is the orthogonal complement of the line pointing in direction
of 𝑦. The map 𝑥 → |𝑓𝑓 (𝑥)
(𝑥)|
is the composite of 𝑓 and the map 𝑥 → 𝑥∕|𝑥| which is smooth in
dimensions 𝑘 ≥ 2.

The derivative of the latter map satisfies

Im (𝑑(𝑥∕|𝑥|)𝑥 ) = (span(𝑥))⟂ ⊂ ℝ𝑘+1 , i.e., Ker (𝑑(𝑥∕|𝑥|)𝑥 ) = span(𝑥).

For 𝑓 ∕|𝑓 |, this means

( ( ) )
𝑓
Ker 𝑑 = span(𝑓 (𝑥)) ∩ Im (𝑑𝑓𝑥 ).
|𝑓 | 𝑥
230 11.2. Winding Numbers and the Borsuk–Ulam Theorem
Thus
( ( ) ) ( ( ) )
𝑓 𝑘 𝑓
Im 𝑑 = 𝑇𝑦 (𝕊 ) ⇐⇒ Ker 𝑑 = {0}
|𝑓 | 𝑥 |𝑓 | 𝑥
⇐⇒ span(𝑓 (𝑥)) ∩ Im (𝑑𝑓𝑥 ) = {0}
⇐⇒ span(𝑓 (𝑥)) ⊄ Im (𝑑𝑓𝑥 )
⇐⇒ 𝐿 + Im (𝑑𝑓𝑥 ) = ℝ𝑘+1
⇐⇒ 𝑓 −
⋔ 𝐿.

Summarizing the argument, we have obtained

𝑓
𝑦 is a regular value of ⇐⇒ 𝑓 −
⋔ 𝐿 = span(𝑦). (11.2)
|𝑓 |

Now we are going to exploit these two observations for calculating 𝑊2 (𝑓 , 0). By defini-
tion, we have
( ) ( ) ( )−1
𝑓 −0 𝑓 𝑓
𝑊2 (𝑓 , 0) = deg2 = deg2 =# (𝑦) mod 2.
|𝑓 − 0| |𝑓 | |𝑓 |

By symmetry, we have
( )−1 ( )−1
𝑓 𝑓
# (𝑦) = # (−𝑦).
|𝑓 | |𝑓 |

From (11.2) we know

𝑓 −1 (𝐿) = {𝑥 ∈ 𝕊𝑘 ∶ 𝑓 (𝑥) ∈ 𝐿}
𝑓 (𝑥)
= {𝑥 ∈ 𝕊𝑘 ∶ = ±𝑦}
|𝑓 (𝑥)|
( )−1 ( )−1
𝑓 𝑓
= (𝑦) ∪ (−𝑦).
|𝑓 | |𝑓 |

Thus
( )−1
𝑓 1
# (𝑦) = #𝑓 −1 (𝐿).
|𝑓 | 2

Hence we need to calculate the number 12 #𝑓 −1 (𝐿), at least modulo 2.

By symmetry, we can do this on the upper hemisphere 𝕊𝑘+ of 𝕊𝑘 , i.e., the points on 𝕊𝑘
with 𝑥𝑘+1 ≥ 0. Let 𝑓+ be the restriction of 𝑓 to 𝕊𝑘+ . By the choice of 𝑦, 𝐿 does not meet the
equator, and hence no point on the equator is in 𝑓 −1 (𝐿). This implies

1 −1
#𝑓 (𝐿) = #𝑓+−1 (𝐿).
2
Chapter 11. The Brouwer Degree modulo 2 231
The upper hemisphere is a manifold with boundary
∑
𝜕𝕊𝑘+ = {𝑥 = (𝑥1 , … , 𝑥𝑘+1 ) ∶ 𝑥2𝑖 = 1 and 𝑥𝑘+1 = 0} = 𝕊𝑘−1
𝑖

being the equator.

Now we would like to apply the previous theorem to the 𝑓+ and 𝑔 = 𝜕𝑓+ and use the
induction hypothesis. However, the target of 𝑓+ has dimension 𝑘 + 1, whereas for both the
theorem and the induction hypothesis we need as target a Euclidean space of dimension 𝑘. So
we need to fix this.

The key is that the orthogonal complement of 𝐿 in ℝ𝑘+1 , denoted by 𝑉 , is a vector space
of dimension 𝑘. By choosing a basis of 𝑉 , we can identify it with ℝ𝑘 .

To complete the argument, let 𝜋 ∶ ℝ𝑘+1 → 𝑉 be the orthogonal projection onto 𝑉 . Since
𝑔 is symmetric and 𝜋 is linear,

𝜋◦𝑔 ∶ 𝕊𝑘−1 → 𝑉 is symmetric ∶ 𝜋(𝑔(−𝑥)) = 𝜋(−𝑔(𝑥)) = −𝜋(𝑔(𝑥)).

Moreover, we have

𝜋(𝑔(𝑥)) = 0 ⇐⇒ 𝑔(𝑥) ∈ 𝐿, hence 𝜋(𝐠(𝐱)) ≠ 𝟎 for all 𝑥 ∈ 𝕊𝑘−1

by the definition of 𝜋 and the choice of 𝐿.

Thus, after choosing a basis for 𝑉 , we can consider 𝜋◦𝑔 as a map

𝜋◦𝑔 ∶ 𝕊𝑘−1 → ℝ𝑘 ⧵ {0}.

Now we apply the induction hypothesis to 𝜋◦𝑔 and get 𝑊2 (𝜋◦𝑔, 0) = 1.

To finish, recall 𝑓+ −
⋔ 𝐿 and hence for

𝜋◦𝑓+ ∶ 𝕊𝑘 → 𝑉 , (𝜋◦𝑓+ ) −
⋔ {0}.

In other words, 0 is a regular value of 𝜋◦𝑓+ .

Hence we can apply the previous theorem to 𝜋◦𝑓+ and its boundary map 𝜕(𝜋◦𝑓+ ) = 𝜋◦𝑔
to get

𝑊2 (𝜋◦𝑔, 0) = #(𝜋◦𝑓+ )−1 (0).

But, by the choice of 𝐿, we have

𝜋(𝑓+ (𝑥)) = 0 ⇐⇒ 𝑓+ (𝑥) ∈ 𝐿, and hence (𝜋◦𝑓+ )−1 (0) = 𝑓+−1 (𝐿).

Thus

𝑊2 (𝑓 , 0) = #𝑓+−1 (𝐿) = 𝑊2 (𝜋◦𝑔, 0) = 1.

232 11.3. Linking numbers and the Hopf invariant modulo 2

Remark 11.17 Going back to the definition of 𝑊2 (𝑓 , 𝑧) and the picture at the be-
ginning, we learn from the proof, in particular, that lines tangential to 𝑓 (𝑋) are not
allowed for calculating 𝑊2 (𝑓 , 𝑧).

Theorem 11.18 (Corollary of the Borsuk–Ulam Theorem) If 𝑓 ∶ 𝕊𝑘 → ℝ𝑘+1 ⧵ {0} is

symmetric about the origin, i.e., 𝑓 (−𝑥) = −𝑓 (𝑥), then 𝑓 intersects every line through
0 at least once.

Proof: Let 𝐿 be a line in ℝ𝑘+1 through the origin. If 𝑓 never hits 𝐿, then #𝑓 −1 (𝐿) = 0 and
−
𝑓 ⋔ 𝐿. By repeating the above proof using this 𝑓 and 𝐿 for calculating 𝑊2 (𝑓 , 0), we would get
the contradiction to the Borsuk–Ulam Theorem

𝑊2 (𝑓 , 0) = #𝑓 −1 (𝐿) = 0.

In the exercises we study further consequences of the Borsuk–Ulam Theorem. So have a

look!

11.3 Linking numbers and the Hopf invariant modulo 2

In this section we will have a first glimpse at the Hopf invariant in a modulo 2 version. We will
discuss the actual ℤ-valued Hopf invariant in the exercises later.

There are many different ways to define this tremendously influential invariant. Here we
follow Milnor’s outline [13, Problems 13-15] of Hopf’s original approach using linking num-
bers.

11.3.1 Mod 2 linking number

We begin with the following definition:

Definition 11.19 (Linking number) For 𝑘 ≥ 1, let 𝑋, 𝑌 ⊂ ℝ𝑘+1 be two disjoint

smooth manifolds. We assume that 𝑋 and 𝑌 are compact and without boundary, of
dimensions dim 𝑋 = 𝑚 and dim 𝑌 = 𝑛 such that 𝑚 + 𝑛 = 𝑘. The mod 2 linking number
𝐿2 (𝑋, 𝑌 ) ∶= deg2 (𝜆) of 𝑋 and 𝑌 is defined to be the mod 2 degree of the map 𝜆 given
by
𝑥−𝑦
𝜆 ∶ 𝑋 × 𝑌 → 𝑆 𝑘 , (𝑥, 𝑦) → .
|𝑥 − 𝑦|

∙ Note that, since 𝑋 and 𝑌 are disjoint, the map 𝜆 is well-defined and smooth.

∙ The order of 𝑋 and 𝑌 does not matter modulo 2, i.e., 𝐿2 (𝑋, 𝑌 ) = 𝐿2 (𝑌 , 𝑋). For the
oriented linking number that we will study later this will be different. Then we have to
Chapter 11. The Brouwer Degree modulo 2 233
consider a sign and have 𝐿(𝑌 , 𝑋) = (−1)(𝑚+1)(𝑛+1) 𝐿(𝑋, 𝑌 ). This will imply that the Hopf
invariant vanishes when𝑛 is odd.

∙ Say something about linking... See Figure 11.6.

Figure 11.6: The red circle is the boundary of a compact manifold 𝐷. In both cases the green
curve intersects the manifold 𝐷. The linking number detects the intersection. However, multiple
intersections may occur. The circles on the left-hand side are linked and cannot be moved apart.
The curves on the right-hand side are not linked and can be moved.

Lemma 11.20 (Linking and boundary) Assume that 𝑋 is the boundary of a smooth
(𝑚 + 1)-manifold 𝑊 which is disjoint from 𝑌 . Then we have 𝐿2 (𝑋, 𝑌 ) = 0.

Proof: Since 𝑌 does not have a boundary, the product 𝑊 × 𝑌 is a smooth manifold with
boundary 𝜕(𝑊 × 𝑌 ) = 𝜕𝑊 × 𝑌 = 𝑋 × 𝑌 . Since 𝑊 and 𝑌 are disjoint, 𝜆 extends to a smooth
map on 𝑊 × 𝑌 . Then the Boundary Theorem 11.8 for degrees implies that deg2 (𝜆) = 0.

Now we extend the definition of the linking number to submanifolds 𝑋, 𝑌 in 𝕊𝑘+1 as follows:

Definition 11.21 (Linking number on spheres) Let 𝑋 and 𝑌 be compact submanifolds

of 𝕊𝑘+1 without boundary with dim 𝑋 + dim 𝑌 = 𝑘. Since the sphere is connected and
since 𝑋 and 𝑌 are closed subsets, there must be a point 𝑝 which is not contained in
either 𝑋 or 𝑌 . We identity 𝕊𝑘+1 ⧵ {𝑝} with ℝ𝑘+1 via the diffeomorphism defined by
stereographic projection 𝜙𝑝 from 𝑝. Then we consider the images 𝜙𝑝 (𝑋) and 𝜙𝑝 (𝑌 )
of 𝑋 and 𝑌 , respectively, under this stereographic projection as submanifolds of ℝ𝑘+1 .
Now we define the linking number 𝐿2 (𝑋, 𝑌 ) of 𝑋 and 𝑌 to be 𝐿2 (𝜙𝑝 (𝑋), 𝜙𝑝 (𝑌 )) of
𝜙𝑝 (𝑋) and 𝜙𝑝 (𝑌 ) as in Definition 11.19.

11.3.2 Mod 2 Hopf invariant

The main situation we are interested in is the following:

Definition 11.22 (Hopf invariant modulo 2) For 𝑛 ≥ 1, consider a smooth map

𝑓∶ 𝕊2𝑛−1 → 𝕊𝑛 . Let 𝑤 ≠ 𝑧 ∈ 𝕊𝑛 be two regular values for 𝑓 . Then 𝑓 −1 (𝑤) and
𝑓 (𝑧) are compact submanifolds of 𝕊𝑛 without boundary and their linking number
−1

𝐿2 (𝑓 −1 (𝑤), 𝑓 −1 (𝑧)) is defined. We define the mod 2 Hopf invariant of 𝑓 , denoted

234 11.3. Linking numbers and the Hopf invariant modulo 2

𝐻2 (𝑓 ), to be the mod 2 linking number of 𝑓 −1(𝑤) and 𝑓 −1 (𝑧), i.e.,

( )
𝐻2 (𝑓 ) ∶= 𝐿2 𝑓 −1 (𝑤), 𝑓 −1 (𝑧) .

Our next goal is to show that this number does not depend on the choice of 𝑤 and 𝑧 and
only depends on the homotopy class of 𝑓 .

Lemma 11.23 (Linking number is homotopy invariant) Let 𝐹 ∶ 𝕊2𝑛−1 × [0, 1] → 𝕊𝑛

be a smooth homotopy between 𝑓0 , 𝑓1 ∶ 𝕊2𝑛−1 → 𝕊𝑛 . Let 𝑤, 𝑧 ∈ 𝕊𝑛 be regular values
for 𝐹 , and both 𝑓0 and 𝑓1 . Then we have the equalities
( ) ( ) ( )
𝐿2 𝑓0−1 (𝑤), 𝑓0−1 (𝑧) = 𝐿2 𝑓1−1 (𝑤), 𝑓0−1 (𝑧) = 𝐿2 𝑓1−1 (𝑤), 𝑓1−1 (𝑧) . (11.3)

Proof: Since 𝑤 is a regular value for 𝐹 , the Boundary Theorem 10.14 implies that the
subset 𝐹 −1 (𝑤) ⊂ 𝕊2𝑛−1 × [0, 1] ⊂ ℝ2𝑛+1 is a compact submanifold with boundary 𝜕𝐹 −1 (𝑤)
given by

𝜕𝐹 −1 (𝑤) = 𝐹 −1 (𝑤) ∩ (𝕊2𝑛−1 × {1} ⊔ 𝕊2𝑛−1 × {0})

= 𝑓1−1 (𝑤) × {1} ⊔ 𝑓0−1 (𝑤) × {0}.

Since 𝑓0−1 (𝑧) is a compact manifold without boundary, the product 𝐹 −1 (𝑤)×𝑓0−1 (𝑧) is a compact
manifold of with boundary given by

𝜕(𝐹 −1 (𝑤) × 𝑓0−1 (𝑧)) = (𝑓1−1 (𝑤) × {1} × 𝑓0−1 (𝑧)) ⊔ (𝑓0−1 (𝑤) × {0} × 𝑓0−1 (𝑧)).

Thus the maps 𝜆0 ∶ 𝑓0−1 (𝑤) × 𝑓0−1 (𝑧) → 𝕊2𝑛−2 and 𝜆1 ∶ 𝑓1−1 (𝑤) × 𝑓0−1 (𝑧) → 𝕊2𝑛−2 are the
restrictions of the map
𝑥−𝑦
𝜆𝐹 ∶ 𝐹 −1 (𝑤) × 𝑓0−1 (𝑧) → 𝕊2𝑛−2 , (𝑥, 𝑦) →
|𝑥 − 𝑦|
to the two boundary components. By the Boundary Theorem 11.8 for deg2 , this implies

deg2 (𝜆0 ⊔ 𝜆1 ) = 0

where 𝜆0 ⊔ 𝜆1 denotes the induced map (𝑓0−1 (𝑤) × 𝑓0−1 (𝑧)) ⊔ (𝑓1−1 (𝑤) × 𝑓0−1 (𝑧)) → 𝕊2𝑛−2 . Since
deg2 is by construction additive on connected components, this implies that
( ) ( )
𝐿2 𝑓0−1 (𝑤), 𝑓0−1 (𝑧) + 𝐿2 𝑓1−1 (𝑤), 𝑓0−1 (𝑧) ≡ 0 mod 2.

Thus we have
( ) ( )
𝐿2 𝑓0−1 (𝑤), 𝑓0−1 (𝑧) ≡ 𝐿2 𝑓1−1 (𝑤), 𝑓0−1 (𝑧) mod 2.

This proves the first equality in (11.3). Applying the same argument for the second factor proves
the second equality.

( )
Lemma 11.24 The linking number 𝐿2 𝑓 −1 (𝑤), 𝑓 −1 (𝑧) is locally constant as a func-
tion of 𝑤 and of 𝑧.
Chapter 11. The Brouwer Degree modulo 2 235
Proof: Because of the symmetry we only prove the assertion for 𝑤. The Local Submersion
Theorem 4.2 implies that the set of regular values of is an open subset of 𝕊𝑛 . Hence we can
choose an 𝜀 so that the open 𝜀-neighborhood of 𝑤 contains only regular values. Let 𝑣 be a
point in 𝕊𝑛 with |𝑣 − 𝑤| < 𝜀. We can choose a family of smooth rotations 𝑟𝑡 ∶ 𝕊𝑛 → 𝕊𝑛 so that
𝑟1 (𝑤) = 𝑣, and so that

∙ 𝑟𝑡 is the identity for 0 ≤ 𝑡 < 𝛿, for some 0 < 𝛿 < 1,

∙ 𝑟𝑡 equals 𝑟1 for 1 − 𝛿 < 𝑡 ≤ 1, and
∙ each 𝑟−1𝑡 (𝑣) lies on the great circle from 𝑤 to 𝑣, which implies that it is a regular value
for 𝑓 .

Now we can define a homotopy

𝐹 ∶ 𝕊2𝑛−1 × [0, 1] → 𝕊𝑛 , 𝐹 (𝑥, 𝑡) = 𝑟𝑡 (𝑓 (𝑥)).

By our choice of 𝑟𝑡 and 𝑣, note that, for each fixed 𝑡, 𝑣 is a regular value for the composition
𝑟𝑡 ◦𝑓 ∶ 𝕊2𝑛−1 → 𝕊𝑛 . This implies that 𝑣 also is a regular value for the map 𝐹 . Now we can
apply Lemma 11.23 to deduce the assertion.

) is well-defined) For a smooth map 𝑓 ∶ 𝕊

Lemma 11.25( (Hopf invariant 2𝑛−1 → 𝕊𝑛 ,

the number 𝐿2 𝑓 −1 (𝑤), 𝑓 −1 (𝑧) does not depend on the choice of regular values 𝑤 and
𝑧 and only depends on the homotopy class of 𝑓 .

( )
Proof: By Lemma 11.24, the assignment (𝑤, 𝑧) → 𝐿2 𝑓 −1 (𝑤), 𝑓 −1 (𝑧) is a locally con-
stant function. Since 𝑌 is connected, the assignment is constant. Now if 𝑔 ∶ 𝕊2𝑛−1 → 𝕊𝑛
is a smooth map homotopic to 𝑓 , then there is a homotopy 𝐹 between 𝑓 and 𝑔. By Sard’s
Theorem 7.1, the set of regular values for 𝑓 , 𝑔 and 𝐹 are dense in 𝑌 . Hence we can find ele-
ments 𝑤 and 𝑧 in 𝑌 which are regular values for 𝑓 , 𝑔 and 𝐹 simultaneously. Now we can apply
Lemma 11.23 to deduce that 𝐻2 (𝑓 ) = 𝐻2 (𝑔).

Remark 11.26 (𝐻2 is a map on homotopy groups) Denoting by 𝜋2𝑛−1 (𝕊𝑛 ) the (2𝑛−1)-
homotopy group of 𝑆 𝑛 , we can view the mod 2 Hopf invariant as a map

𝐻2 ∶ 𝜋2𝑛−1 (𝕊𝑛 ) → ℤ∕2.

11.3.3 The mod 2 Hopf invariant of the Hopf fibration

Now we are going to compute the Hopf invariant for the Hopf fibration 𝜋 ∶ 𝕊3 → 𝕊2 , an
extremely important example of a smooth map. We recall the definition of 𝜋: We consider 𝕊3
as a subset of ℂ2 , i.e., 𝕊3 = {(𝑧0 , 𝑧1 ) ∈ ℂ2 ∶ |𝑧0 |2 + |𝑧1 |2 = 1}, and 𝕊2 as a subset of ℂ × ℝ,
i.e., 𝕊2 = {(𝑧, 𝑥) ∈ ℂ × ℝ ∶ |𝑧|2 + 𝑥2 = 1}. Then the Hopf fibration 𝜋 is the map 𝕊3 → 𝕊2
given by ( )
𝜋(𝑧0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2 .

We checked in Exercise 2.8 that this actually defines a map 𝕊3 → 𝕊2 .

236 11.3. Linking numbers and the Hopf invariant modulo 2
We will show that 𝐻2 (𝜋) = 1 in ℤ∕2. This result has a lot of important consequences. Here
are just a few:

∙ We can conclude that 𝜋 is not homotopic to a constant map. This is a famous result
of Heinz Hopf providing examples of non-contractible maps between spheres where the
dimension of the domain is bigger than the codomain.
∙ We have shown in Exercise 2.8 that the Hopf map realizes 𝕊3 as a disjoint union of
fibers which each look like 𝕊1 . Since the Hopf invariant is the linking number of any two
distinct circles on 𝑆 3 , this shows that all these disjoint circles are linked in each other
and cannot be pulled apart.
∙ Hence 𝜋 exhibits a very special behavior of maps between spheres that only exists in a
handful of dimensions. The latter is a famous result of Frank Adams, known as the Hopf
invariant one problem, and is actually concerned with the integral version of the Hopf
invariant that we will study later. Adams’ theorem had enormous consequences on the
development of mathematics.

Theorem 11.27 (Hopf invariant of the Hopf fibration is non-trivial) We have

𝐻2 (𝜋) = 1 in ℤ∕2.

Proof: We have shown previously that 𝜋 is a submersion. Thus every point in 𝕊2 is a

regular value. Here we consider 𝑎 = (0, 0, 1) and 𝑏 = (0, 1, 0) on 𝕊2 ⊂ ℝ3 ≅ ℂ × ℝ. The fiber
of 𝜋 over 𝑎 is { }
𝜋 −1 (𝑎) = (𝑧0 , 0) ∈ 𝕊3 ⊂ ℂ2 ∶ |𝑧0 |2 = 1 .
To determine the fiber over 𝑏, we write 𝑧0 = 𝑥0 + 𝑖𝑦0 and 𝑧1 = 𝑥1 + 𝑖𝑦1 . Then we get
1
𝜋(𝑧0 , 𝑧1 ) = (0, 1, 0) ⇒ 2𝑧0 𝑧̄ 1 = 𝑖 and |𝑧0 |2 = |𝑧1 |2 =
2
1
⇒ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 and 𝑥20 + 𝑥21 = .
2
Thus the fiber over 𝑏 has the form
{ }
−1 3 𝑖
𝜋 (𝑏) = (𝑧0 , 𝑧1 ) ∈ 𝕊 ∶ 𝑧̄ 1 =
2𝑧0
{ 3
}
= (𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) ∈ 𝕊 ∶ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 .

By definition of 𝐻2 (𝜋), we need to choose two distinct regular values 𝑤 and 𝑧 of 𝜋 and
calculate the mod 2 linking number of 𝜋 −1 (𝑤) and 𝜋 −1 (𝑧). Since we showed that each value is
regular, we can for example choose 𝑤 = 𝑎 = (0, 0, 1) and 𝑧 = 𝑏 = (0, 1, 0) on 𝕊2 ⊂ ℝ3 ≅ ℂ×ℝ.

To calculate the linking number of 𝜋 −1 (𝑎) and 𝜋 −1 (𝑏) we need to choose a point on 𝕊3
disjoint from these two subsets and stereographically project 𝕊3 from this point onto ℝ3 . By
our choice of 𝑎 and 𝑏, we get that the north pole 𝑁 = (0, 0, 0, 1) is neither on 𝜋 −1 (𝑎) nor on
𝜋 −1 (𝑏). Recall that the formula for the stereographic projection 𝜙−1
𝑁
∶ 𝕊3 ⧵ {𝑁} → ℝ3 is, with
the notation we use here, given by
1
(𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) → (𝑥 , 𝑦 , 𝑥 ).
1 − 𝑦1 0 0 1
Chapter 11. The Brouwer Degree modulo 2 237
Hence we get

𝑁 (𝜋 (𝑎)) = {𝐱 = (𝑥0 , 𝑥1 , 𝑥2 ) ∈ ℝ ∶ 𝑥0 + 𝑥1 = 1 and 𝑥2 = 0}

𝕊𝑎 ∶= 𝜙−1 −1 3 2 2

and
{ }
(1 − 𝑦0 )2
𝕊𝑏 ∶= 𝜙−1 −1
𝑁 (𝜋 (𝑏)) = 𝐲 = (𝑦0 , 𝑦1 , 𝑦2 ) ∈ ℝ ∶ 𝑦1 = 𝑦2 and
3
𝑦20 + 𝑦21 = .
2

Now we can calculate 𝐻2 (𝜋) as deg2 (𝜆) with

𝐱−𝐲
𝜆 ∶ 𝕊𝑎 × 𝕊𝑏 → 𝕊2 , (𝐱, 𝐲) → .
|𝐱 − 𝐲|

To compute the degree of 𝜆 we pick a convenient point of 𝕊2 and determine the fiber over
this point. Then we check that we actually picked a regular value.

So let us look at 𝑝 = (1, 0, 0). The equation 𝜆(𝐱, 𝐲) = 𝑝 then implies

𝑥1 = 𝑦1 = 0 and 𝑥0 − 𝑦0 = |𝑥0 − 𝑦0 |.

The latter condition implies that 𝑣0 − 𝑤0 is positive. This does not look very helpful at first
glance, but we also know
1 = 𝑥20 + 𝑥21 = 𝑥20 , i.e., 𝑥0 = ±1,
and
(1 − 𝑦0 )2 √
𝑦20 = ⇐⇒ 𝑦0 = ± 2 − 1.
2
Hence we can check for the four possible permutations
√ of the signs whether they yield
√ 𝑥0 −𝑦0 ≥
0 and get three points: one√with 𝑥0 = 1, 𝑦0 = 2 − 1, one with 𝑥0 = 1, 𝑦0 = − 2 − 1, and
one with 𝑥0 = −1, 𝑦0 = − 2 − 1. Hence we get three points (𝐱, 𝐲) in 𝕊𝑎 × 𝕊𝑏 with 𝜆(𝐱, 𝐲) = 𝑝.
Hence, once we have shown that 𝑝 is a regular value, we will have proved 𝐻2 (𝜋) = deg2 (𝜆) ≡ 1
mod 2.

It remains to check the derivatives of 𝜆 at these points and to show that 𝑝 is a regular value.
Hence we have to show that the determinants at each point are nonzero.

Since 𝕊𝑎 is the unit circle in the 𝑥𝑦-plane, the tangent space of 𝕊𝑎 at a point 𝐱 is given by

𝑇𝐱 𝕊𝑎 = {𝐮 = (𝑢0 , 𝑢1 , 𝑢2 ) ∈ ℝ3 ∶ 𝑢2 = 0 and 𝑢0 𝑥0 + 𝑢1 𝑥1 = 0}.

Similarly, 𝕊𝑏 lies in the plane 𝑃 in ℝ3 of points 𝐲 = (𝑦0 , 𝑦1 , 𝑦2 ) with 𝑦1 = 𝑦2 . Then 𝕊𝑏 is the

fiber of the map
(1 − 𝑦0 )2
𝑔𝑏 ∶ 𝑃 → ℝ, 𝐲 = (𝑦0 , 𝑦1 , 𝑦1 ) → 𝑦20 + 𝑦21 = .
2
After a simple computation we get

𝕊𝑏 = 𝑔𝑏−1 (1) = {𝐲 = (𝑦0 , 𝑦1 , 𝑦1 ) ∈ 𝑃 ∶ 2𝑦21 + 𝑦20 + 2𝑦0 = 1}.

The derivative of 𝑔𝑏 as a map from 𝑃 → ℝ is given by the (2×1)-matrix (we could also consider
it as a map ℝ3 → ℝ2 )
𝑑(𝑔𝑏 )𝐲 = (2𝑦0 + 2, 2𝑦1 ).
238 11.3. Linking numbers and the Hopf invariant modulo 2
Hence we get

𝑇𝐰 𝕊𝑏 = {𝐮 = (𝑢0 , 𝑢1 , 𝑢2 ) ∈ ℝ3 ∶ 𝑢2 = 𝑢1 and 𝑢0 (𝑦0 + 1) + 𝑢1 𝑦1 = 0}.

Now we calculate the derivative of 𝜆. First we do this as a map

𝐱−𝐲
𝜆̃ ∶ ℝ3 × ℝ3 → ℝ3 , (𝐱, 𝐲) → .
|𝐱 − 𝐲|

This will help us, since 𝑔3 ◦𝜆̃ is constant on 𝕊𝑎 × 𝕊𝑏 . Hence 𝑑 𝜆̃(𝐱,𝐲) sends the subspace 𝑇𝐱 𝕊𝑎 ×
𝑇𝐲 𝕊𝑏 ∈ ℝ3 × ℝ3 to 𝑇𝑝 𝑆 2 ⊂ ℝ3 .
𝜕 𝜆̃ 𝜕 𝜆̃
We determine 𝑑 𝜆̃(𝐱,𝐲) by computing its partial derivatives 𝜕𝑥 𝑖 (𝐱, 𝐲) and 𝜕𝑦 𝑖 (𝐱, 𝐲) with respect
𝑗 𝑗
to the variables 𝑥0 , 𝑥1 , 𝑥2 and 𝑦0 , 𝑦1 , 𝑦2 :

For 𝑖 ≠ 𝑗, we have

𝜕 𝜆̃𝑖 (𝑥𝑖 − 𝑦𝑖 )(𝑥𝑗 − 𝑦𝑗 ) 𝜕 𝜆̃𝑖

(𝐱, 𝐲) = = (𝐱, 𝐲).
𝜕𝑥𝑗 |𝐱 − 𝐲|3 𝜕𝑦𝑗

For 𝑖 = 𝑗, we get

⎧(𝑥1 − 𝑦1 )2 + (𝑥2 − 𝑦2 )2 if 𝑖 = 0
𝜕 𝜆̃𝑖 1 ⎪
(𝐱, 𝐲) = ⋅ ⎨(𝑥0 − 𝑦0 )2 + (𝑥2 − 𝑦2 )2 if 𝑖 = 1
𝜕𝑥𝑖 |𝐱 − 𝐲| ⎪
3
2
⎩(𝑥0 − 𝑦0 ) + (𝑥1 − 𝑦1 )
2 if 𝑖 = 2

and

𝜕 𝜆̃𝑖 𝜕 𝜆̃
(𝐱, 𝐲) = − 𝑖 (𝐱, 𝐲).
𝜕𝑦𝑖 𝜕𝑥𝑖

Now we evaluate these formulae at the points (𝐱, 𝐲) with 𝜆(𝐱, 𝐲) = 𝑝. For each such point
we found 𝑥1 = 𝑥2 = 𝑦1 = 𝑦2 = 0. Hence we get

𝜕 𝜆̃𝑖 𝜕 𝜆̃
(𝐱, 𝐲) = 0 = 𝑖 (𝐱, 𝐲)
𝜕𝑥𝑗 𝜕𝑦𝑗

for 𝑖 ≠ 𝑗,
{
𝜕 𝜆̃𝑖 1 0 if 𝑖 = 0
(𝐱, 𝐲) = ⋅
𝜕𝑥𝑖 |𝑥0 − 𝑦0 | 1 if 𝑖 = 1, 2

and
{
𝜕 𝜆̃𝑖 1 0 if 𝑖 = 0
(𝐱, 𝐲) = ⋅
𝜕𝑦𝑖 |𝑥0 − 𝑦0 | −1 if 𝑖 = 1, 2.

Now we are equipped to study the linear map 𝑑𝜆(𝐱,𝐲) ∶ 𝑇𝐱 𝕊𝑎 × 𝑇𝐲 𝕊𝑏 → 𝑇𝑝 𝕊2 :

Chapter 11. The Brouwer Degree modulo 2 239
⎛0⎞ ⎛0⎞
A basis of 𝑇𝑝 𝕊2 is given by the vectors 𝐞2 = ⎜1⎟ and 𝐞3 = ⎜0⎟.
⎜ ⎟ ⎜ ⎟
⎝0⎠ ⎝1⎠

⎛0⎞
At each of the points (𝐱, 𝐲) we found with 𝜆(𝐱, 𝐲) = 𝐩, the vector 𝐚 = ⎜1⎟ is a basis of 𝑇𝐱 𝕊𝑎
⎜ ⎟
⎝0⎠
⎛0⎞ ⎛0⎞
and the vector 𝐛 = ⎜1⎟ provides a basis of 𝑇𝐲 𝕊𝑏 . The map 𝑑𝜆(𝐱,𝐲) sends 𝐚 to |𝑥 −𝑦
1
⋅ ⎜1⎟ and 𝐛
⎜ ⎟ 0 0| ⎜ ⎟
⎝1⎠ ⎝0⎠
⎛0⎞
1
to |𝑥 −𝑦 ⋅ ⎜−1⎟. Hence we have
0 0| ⎜ ⎟
⎝−1⎠

1 1 1
𝑑𝜆(𝐱,𝐲) (𝐚) = ⋅ 𝐞 and 𝑑𝜆(𝐱,𝐲) (𝐛) = − ⋅𝐞 − ⋅𝐞 .
|𝑥0 − 𝑦0 | 2 |𝑥0 − 𝑦0 | 2 |𝑥0 − 𝑦0 | 3

These two vectors form a basis of 𝑇𝑝 𝕊2 and we see that (𝐱, 𝐲) is a regular point. Since this is
true for all points in the fiber of 𝑝 ∈ 𝕊2 , we conclude that 𝑝 actually is a regular value.
240 11.4. Exercises and more examples
11.4 Exercises and more examples

11.4.1 Degree modulo 2

𝑓 𝑔
Exercise 11.1 Let 𝑋 ←←←→ ← 𝑍 be a sequence of smooth maps between manifolds with
← 𝑌 ←←→
𝑋 and 𝑌 compact, 𝑌 and 𝑍 connected. Assume that all three manifolds are boundaryless
and dim 𝑋 = dim 𝑌 = dim 𝑍. Show that

deg2 (𝑔◦𝑓 ) = deg2 (𝑔) ⋅ deg2 (𝑓 ).

Exercise 11.2 Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map between smooth manifolds with 𝑋

compact, 𝑌 connected and dim 𝑋 = dim 𝑌 .

(a) Show that, if deg2 (𝑓 ) ≠ 0, then 𝑓 is surjective.

(b) Show that if 𝑌 is not compact, then deg2 (𝑓 ) = 0.

(c) Let 𝑋 = 𝑌 = 𝕊1 ⊂ ℝ2 be the unit circle and assume that 𝑓 ∶ 𝕊1 → 𝕊1 is a smooth

map without fixed points. Show that 𝑓 is surjective.
Hint: Show that 𝑓 is homotopic to the antipodal map 𝛼 ∶ 𝕊1 → 𝕊1 , 𝑥 → −𝑥.
What is deg2 (𝛼)? Use the previous points.

Exercise 11.3 Show that there exists a complex number 𝑧 such that

𝑧7 + cos(|𝑧|2 )(1 + 93𝑧4 ) = 0.

Exercise 11.4 Let 𝑚 be an odd number and let

𝑝(𝑧) = 𝑧𝑚 + 𝑎1 𝑧𝑚−1 + ⋯ + 𝑎𝑚

be a monic complex polynomial. Show that there exists a 𝑤 ∈ ℂ such that 𝑝(𝑤) = 0.

11.4.2 Borsuk–Ulam Theorem

Exercise 11.5 Let 𝑓1 , … , 𝑓𝑘 be 𝑘 smooth real-valued odd functions, i.e., 𝑓𝑖 ∶ 𝕊𝑘 → ℝ

with 𝑓 (−𝑥) = −𝑓 (𝑥). Show that 𝑓1 , … , 𝑓𝑘 must have a common zero, i.e., there is an
𝑥 ∈ 𝕊𝑘 such that 𝑓1 (𝑥) = ⋯ = 𝑓𝑘 (𝑥).
Hint: Use the Borsuk–Ulam Theorem and its consequences.

Exercise 11.6 Let 𝑔1 , … , 𝑔𝑘 on 𝕊𝑘 be smooth real-valued functions. Show that there

Chapter 11. The Brouwer Degree modulo 2 241

exists a point 𝑝 ∈ 𝕊𝑘 such that

𝑔1 (𝑝) = 𝑔1 (−𝑝), … , 𝑔𝑘 (𝑝) = 𝑔𝑘 (−𝑝).

Hint: Use the previous exercise.

Exercise 11.7 Let 𝑝1 , … , 𝑝𝑛 be real polynomials in 𝑛 + 1 variables. Assume each 𝑝𝑖 is

homogeneous of odd order, i.e., there is an odd number 𝑚𝑖 such that 𝑝𝑖 (𝜆𝑥) = 𝜆𝑚𝑖 𝑝𝑖 (𝑥)
for all 𝜆 ∈ ℝ. We consider each 𝑝𝑖 as a smooth function ℝ𝑛+1 → ℝ by sending 𝑥 to 𝑝𝑖 (𝑥).
Show that there is a line through the origin in ℝ𝑛+1 on which all the 𝑝𝑖 ’s simultane-
ously vanish.
Hint: Use the Borsuk–Ulam Theorem and the previous exercises.

Exercise 11.8 Let 𝕊1 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 + 𝑦2 = 1} be the unit circle and 𝕊2 =

{(𝑥, 𝑦, 𝑧) ∈ℝ3 ∶ 𝑥2 + 𝑦2 + 𝑧2 = 1} be the two-dimensional sphere.
Show that there is no continuous map 𝑓 ∶ 𝕊2 → 𝕊1 with 𝑓 (−𝑝) = −𝑓 (𝑝) for all
𝑝 ∈ 𝕊2 .
Hint: Assume such a map 𝑓 existed. Then we could define the continuous map
√
𝑔 ∶ 𝔹2 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 + 𝑦2 ≤ 1} → 𝕊1 , 𝑔(𝑥, 𝑦) ∶= 𝑓 (𝑥, 𝑦, 1 − 𝑥2 − 𝑦2 ).

Show that 𝑔 satisfies 𝑔(−𝑞) = −𝑔(𝑞) for all 𝑞 ∈ 𝕊1 = 𝜕𝔹2 . What is the degree modulo 2
of 𝑔? Conclude that 𝑓 cannot exist.

Exercise 11.9 Use the previous exercise to prove the following special case of our
previous observations: For every smooth map 𝕊2 → ℝ2 , there is a point 𝑝 on 𝕊2 such
that 𝑓 (𝑝) = 𝑓 (−𝑝).

Exercise 11.10 Assume the following continuous version of the assertion of Exer-
cise 11.9: For every continuous map 𝕊2 → ℝ2 , there is a point 𝑝 on 𝕊2 such that
𝑓 (𝑝) = 𝑓 (−𝑝).
Deduce from this fact the following version of Invariance of Dimension:
An open subset in ℝ2 cannot be homeomorphic to an open subset in ℝ𝑛 for 𝑛 ≥ 3.
12. Tubular Neighborhoods and Transversality

12.1 Normal bundles and tubular neighborhoods

In this section we study two key tools in differential topology. We will see several important
application in the next sections. We are going to use a generalization of the Inverse Function
Theorem that we will prove first.

12.1.1 A generalization of the Inverse Function Theorem

We begin with the compact case:

Theorem 12.1 (Generalization of the IFT - compact case) Let 𝑓 ∶ 𝑋 → 𝑌 be a

smooth map that is one-to-one on a compact submanifold 𝑍 of 𝑋. Suppose that for all
𝑥 ∈ 𝑍,

𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 )

is an isomorphism. Then 𝑓 maps an open neighborhood of 𝑍 in 𝑋 diffeomorphically

onto an open neighborhood of 𝑓 (𝑍) in 𝑌 . If 𝑍 is a single point, this is just the usual
IFT.

Proof: We know that 𝑓 maps 𝑍 diffeomorphically onto its image 𝑓 (𝑍), since 𝑓 ∶ 𝑍 →
𝑓 (𝑍) is a bijective local diffeomorphism and therefore a diffeomorphism. We would like to
show that we can extend this to an open neighborhood around 𝑍.

Since 𝑑𝑓𝑥 is an isomorphism for all 𝑥 ∈ 𝑍, for each 𝑥 ∈ 𝑍, there exists an open neighbor-
hood 𝑈𝑥 in 𝑋 around 𝑥 on which 𝑓|𝑈𝑥 is a diffeomorphism. The collection {𝑈𝑥 } is an open cover
of 𝑍. Since 𝑍 is compact, we can choose a finite subcover {𝑈1 , … , 𝑈𝑛 }. We set 𝑈 ∶= ∪𝑖 𝑈𝑖 .
Restricted to the open subset 𝑈 , 𝑓|𝑈 is a local diffeomorphism. It remains to shrink 𝑈 if nec-
essary to ensure that 𝑓|𝑈 is also injective, i.e., we need to show that there is some open subset
𝑉 in 𝑋 which contains 𝑍 such that 𝑓|𝑉 is injective. Then 𝑓|𝑈 ∩𝑉 is an injective local diffeo-
morphism. We have shown previously that this implies that 𝑓|𝑈 ∩𝑉 is a diffeomorphism onto
an open subset of 𝑌 . Since 𝑍 ⊂ 𝑈 and 𝑍 ⊂ 𝑉 , we also have 𝑍 ⊂ 𝑈 ∩ 𝑉 and the assertion is
proven.

We are going to show that 𝑉 exists by assuming the contrary, i.e., we assume that there
exists at least one point 𝑧 ∈ 𝑍 such that in any small open neighborhood 𝑊 of 𝑧 there are points
𝑎 ≠ 𝑏 with 𝑓 (𝑎) = 𝑓 (𝑏). By choosing open neighborhoods 𝑊𝑖 smaller and smaller around 𝑧
and by choosing subsequences 𝑎𝑖 ≠ 𝑏𝑖 with 𝑓 (𝑎𝑖 ) = 𝑓 (𝑏𝑖 ), we can assume that both the 𝑎𝑖 and
𝑏𝑖 converge to 𝑧. Since 𝑓 (𝑎𝑖 ) = 𝑓 (𝑏𝑖 ) for all 𝑖 and 𝑓 is continuous, we have 𝑓 (𝑎𝑖 ) → 𝑓 (𝑧) and

242
Chapter 12. Tubular Neighborhoods and Transversality 243
𝑓 (𝑏𝑖 ) → 𝑓 (𝑧). But since 𝑑𝑓𝑧 is an isomorphism, the usual Inverse Function Theorem implies
that there is a small open neighborhood 𝑊𝑧 in 𝑋 around 𝑧 such that 𝑓|𝑊𝑧 is a diffeomorphism.
Since 𝑎𝑖 → 𝑧 and 𝑏𝑖 → 𝑧, for 𝑁 large enough, we have 𝑎𝑖 , 𝑏𝑖 ∈ 𝑊𝑧 and hence 𝑓 (𝑎𝑖 ) = 𝑓 (𝑏𝑖 ) ∈
𝑓 (𝑊𝑧 ) for all 𝑖 ≥ 𝑁. But since 𝑓|𝑊𝑧 is injective, this implies 𝑎𝑖 = 𝑏𝑖 for all 𝑖 ≥ 𝑁. This
contradicts the choice of the 𝑎𝑖 and 𝑏𝑖 .

The existence of partitions of unity allows us to move from the compact to the general
case. First we recall the following notion from general topology:

Definition 12.2 (Locally finite subcovers) An open cover {𝑉𝛼 } of a manifold 𝑋 is

called locally finite if each point of 𝑋 possesses a neighborhood that intersects only
finitely many of the sets 𝑉𝛼 .

Partitions of unity can be used to show the following lemma:

Lemma 12.3 (Local finiteness lemma) Every open cover {𝑈𝛼 } on a manifold admits
a locally finite refinement {𝑉𝛼 }.

Since the above lemma is rather a statement in general topology, we omit its proof here. So
we move on and use it to generalize the Inverse Function Theorem:

Theorem 12.4 (Generalization of the IFT - general case) Let 𝑓 ∶ 𝑋 → 𝑌 be a

smooth map that is one-to-one on a submanifold 𝑍 of 𝑋. Suppose that for all 𝑥 ∈ 𝑍,
𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 ) is an isomorphism. Assume that 𝑓 maps 𝑍 diffeomorphi-
cally onto 𝑓 (𝑍).
Then 𝑓 maps an open neighborhood of 𝑍 in 𝑋 diffeomorphically onto an open
neighborhood of 𝑓 (𝑍) in 𝑌 .

Proof: For each 𝑥 ∈ 𝑍, there exists an open neighborhood 𝑉𝑥 in 𝑋 around 𝑥 on which 𝑓|𝑉𝑥
is a diffeomorphism, since 𝑑𝑓𝑥 is an isomorphism for all 𝑥 ∈ 𝑍. Let 𝑈𝑥 = 𝑓 (𝑉𝑥 ) be the open
image in 𝑌 . The collection of all 𝑈𝑥 is an open cover of 𝑓 (𝑍), since each 𝑓 (𝑥) ∈ 𝑓 (𝑍) lies
in some 𝑈𝑥 = 𝑓 (𝑉𝑥 ). By the previous lemma, we can choose a locally finite subcover {𝑈𝑖 } of
𝑓 (𝑍) in 𝑌 . For each 𝑈𝑖 , there is a local inverse 𝑔𝑖 ∶ 𝑈𝑖 → 𝑉𝑖 ⊂ 𝑋 of 𝑓|𝑉𝑖 . We define

𝑊 ∶= {𝑦 ∈ 𝑈𝑖 ∶ 𝑔𝑖 (𝑦) = 𝑔𝑗 (𝑦) whenever 𝑦 ∈ 𝑈𝑖 ∩ 𝑈𝑗 }.

On the subset 𝑊 , we can define an inverse

𝑔 ∶ 𝑊 → 𝑋, 𝑔(𝑦) = 𝑔𝑖 (𝑦) for any 𝑖.

This is well-defined by the construction of 𝑊 , as 𝑔(𝑦) = 𝑔𝑖 (𝑦) = 𝑔𝑗 (𝑦) whenever 𝑦 ∈ 𝑈𝑖 ∩ 𝑈𝑗 .

Since the 𝑔𝑖 ’s are local inverses of 𝑓 , we have 𝑓 (𝑍) ⊂ 𝑊 .

It remains to show that 𝑊 contains an open subset which still contains 𝑓 (𝑍). Let 𝑥 ∈ 𝑍.
Since 𝑓 (𝑥) ∈ 𝑓 (𝑍), we can find a 𝑘 such that 𝑓 (𝑥) ∈ 𝑈𝑘 . If 𝑈𝑘 ⊂ 𝑊 , we are done, since then
every point in 𝑓 (𝑍) has an open neighborhood which is contained in 𝑊 . So assume 𝑈𝑘 is not
contained in 𝑊 . After shrinking 𝑈𝑗 if necessary, we can assume by the local finiteness of the
cover {𝑈𝑖 }, that there are only finitely many of the 𝑈𝑖 ’s which intersect 𝑈𝑘 , say 𝑈1 , … , 𝑈𝑛 .
244 12.1. Normal bundles and tubular neighborhoods
Then, for 𝑖 = 1, … , 𝑛, we set 𝐶𝑖𝑘 to be the closure of the set {𝑦 ∈ 𝑈𝑖 ∩ 𝑈𝑘 ∶ 𝑔𝑖 (𝑦) ≠ 𝑔𝑘 (𝑦)},
i.e.,
𝐶𝑖𝑘 = {𝑦 ∈ 𝑈𝑖 ∩ 𝑈𝑘 ∶ 𝑔𝑖 (𝑦) ≠ 𝑔𝑘 (𝑦)}.
Since the union of a finite collection of closed subsets is closed, 𝐶𝑘 ∶= 𝐶1𝑘 ∪ ⋯ ∪ 𝐶𝑛𝑘 is closed
in 𝑌 . Hence

𝑈 ∶= 𝑈𝑘 ⧵ 𝐶𝑘

is open in 𝑌 . By definition of 𝑊 and the 𝐶𝑘 , we know 𝑈 ⊂ 𝑊 . It remains to make sure that

𝑓 (𝑥) is still in 𝑈 , i.e., that it does not belong to one of the closures 𝐶𝑖𝑘 .

Note that 𝑓 (𝑥) satisfies 𝑔𝑖 (𝑓 (𝑥)) = 𝑥 = 𝑔𝑘 (𝑓 (𝑥)) for all 𝑖 = 1, … , 𝑛. Since 𝑑𝑓𝑥 is an
isomorphism, the usual Inverse Function Theorem implies that there is a small open neigh-
borhood 𝑉𝜀 ⊂ 𝑈 around 𝑥 such that 𝑓|𝑉𝜀 is a diffeomorphism. Hence, for each 𝑖 = 1, … , 𝑛, we
have

𝑔𝑖 (𝑓 (𝑥′ )) = 𝑥′ = 𝑔𝑘 (𝑓 (𝑥′ )) for all 𝑥′ ∈ 𝑉𝜀 ∩ 𝑔𝑖 (𝑈𝑖 ) ∩ 𝑔𝑘 (𝑈𝑘 ).

Hence the finite intersection 𝑓 (𝑉𝜀 ) ∩ 𝑈𝑘 ∩ 𝑈1 ∩ ⋯ ∩ 𝑈𝑛 is an open subset which is not contained
in any of the sets {𝑦 ∈ 𝑈𝑖 ∩ 𝑈𝑘 ∶ 𝑔𝑖 (𝑦) ≠ 𝑔𝑘 (𝑦)}. Thus 𝑓 (𝑥) is not contained in 𝐶𝑘 . Hence we
have shown that 𝑈 ⊂ 𝑊 is an open subset containing 𝑓 (𝑥).

12.1.2 Normal bundle

Let 𝑋 ⊂ ℝ𝑚 be a smooth manifold without boundary. We would like to understand a little bit
of the geometry of 𝑋 with respect to its environment. In order to prove the 𝜀-Neighborhood
Theorem we introduce an important geometric tool similar to the tangent bundle.

Definition 12.5 (The Normal Bundle) For each 𝑥 ∈ 𝑋, define 𝑁𝑥 (𝑋), the normal
space of 𝑋 at 𝑥, to be the orthogonal complement of 𝑇𝑥 (𝑋) in ℝ𝑚 . The normal bundle
𝑁(𝑌 ) is then defined to be the set

𝑁(𝑋) = {(𝑥, 𝑣) ∈ 𝑋 × ℝ𝑚 ∶ 𝑣 ∈ 𝑁𝑥 (𝑋)}.

There is a natural projection map 𝜎 ∶ 𝑁(𝑋) → 𝑌 defined by 𝜎(𝑥, 𝑣) = 𝑥.

∙ Warning: Note that, unlike 𝑇 (𝑋), 𝑁(𝑋) is not intrinsic to the manifold 𝑌 but depends
on the specific relationship between 𝑌 and the surrounding ℝ𝑚 .

The normal bundle 𝑁(𝑋) is actually a smooth manifold itself. In order to show this, we
must recall a fact from linear algebra: Suppose that 𝐴 ∶ ℝ𝑚 → ℝ𝑘 is a linear map. Its trans-
pose is a linear map 𝐴𝑇 ∶ ℝ𝑘 → ℝ𝑚 characterised by the dot product equation

𝐴𝑣 ⋅ 𝑤 = 𝑣 ⋅ 𝐴𝑇 𝑤 for all 𝑣 ∈ ℝ𝑚 , 𝑤 ∈ ℝ𝑘 .
Chapter 12. Tubular Neighborhoods and Transversality 245

Lemma 12.6 (Help from Linear Algebra) If 𝐴 is surjective, then the transpose 𝐴𝑇
maps ℝ𝑘 isomorphically onto the orthogonal complement of the kernel of 𝐴.

Proof: First we note that 𝐴𝑇 is injective. For if 𝐴𝑇 𝑤 = 0, then 𝐴𝑣 ⋅ 𝑤 = 𝑣 ⋅ 𝐴𝑇 𝑤 = 0,

so that 𝑤 ⟂ 𝐴(ℝ𝑚 ). Since 𝐴 is surjective, 𝑤 must be zero. Now, if 𝑣 ∈ Ker (𝐴), i.e., 𝐴𝑣 = 0,
then 0 = 𝐴𝑣 ⋅ 𝑤 = 𝑣 ⋅ 𝐴𝑇 𝑤. Thus 𝐴𝑇 (ℝ𝑘 ) ⟂ Ker (𝐴). Hence 𝐴𝑇 maps ℝ𝑘 one-to-one into
the orthogonal complement of Ker (𝐴). As Ker (𝐴) has dimension 𝑚 − 𝑘, its complement has
dimension 𝑘, so 𝐴𝑇 is surjective, too.

Now we can prove:

Theorem 12.7 (Normal bundles are manifolds) Let 𝑋 ⊂ ℝ𝑚 be a smooth 𝑛-

dimensional manifold without boundary. Then 𝑁(𝑋) is a smooth manifold of dimen-
sion 𝑚 and the projection 𝜎 ∶ 𝑁(𝑋) → 𝑋 is a submersion.

Proof: We need to find local parametrizations for 𝑁(𝑋). Let 𝑥 be any point in 𝑋. We
learned in the discussion of the Local Immersion Theorem that we can find an open subset
𝑉 ⊂ ℝ𝑚 containing 𝑥 and local coordinate functions (𝑢1 , … , 𝑢𝑚 ) ∶ 𝑉 → ℝ𝑚 such that

𝑋 ∩ 𝑉 = {𝑣 ∈ 𝑉 ∶ 𝑢𝑛+1 (𝑣) = ⋯ = 𝑢𝑚 (𝑣) = 0}.

We define 𝜑 to be the smooth map

𝜑 ∶ 𝑉 → ℝ𝑚−𝑛 , 𝑣 → (𝑢𝑛+1 (𝑣), … , 𝑢𝑚 (𝑣)).

We set 𝑈 = 𝜑−1 (0) = 𝑋 ∩ 𝑉 . Note that, since 𝑉 is open in ℝ𝑚 , we deduce that 𝑈 is open in
𝑋. Let 𝑁(𝑈 ) be the normal bundle of 𝑈 considered as a smooth manifold in ℝ𝑚 . We observe
that 𝑁(𝑈 ) equals 𝑁(𝑋) ∩ (𝑈 × ℝ𝑚 ), thus it is open in 𝑁(𝑋).

For each 𝑥 ∈ 𝑈 , 𝑑𝜑𝑥 ∶ ℝ𝑚 → ℝ𝑚−𝑛 is surjective and has kernel 𝑇𝑥 (𝑈 ) = 𝑇𝑥 (𝑋) by the
Preimage Theorem 6.3. Therefore the transpose of 𝑑𝜑𝑥 maps ℝ𝑚−𝑛 isomorphically onto the
orthogonal complement of Ker (𝑑𝜑𝑥 ) = 𝑇𝑥 (𝑋). But this is 𝑁𝑥 (𝑋) by definition. Hence we
get an isomorphism
≅
← (𝑇𝑥 (𝑋))⟂ = 𝑁𝑥 (𝑋).
(𝑑𝜑𝑥 )𝑇 ∶ ℝ𝑘 ←←←→

This shows that the map

𝜓 ∶ 𝑈 × ℝ𝑚−𝑛 → 𝑁(𝑈 ), (𝑥, 𝑣) → (𝑥, 𝑑𝜑𝑇𝑥 (𝑣))

is bijective. It is also an embedding of 𝑈 ×ℝ𝑚−𝑛 into 𝑈 ×ℝ𝑚 , since it is the identity on the first
factor and an injective linear map on the second factor. Hence 𝜓 is a diffeomorphism. Thus
we have shown that 𝑁(𝑈 ) is a smooth manifold with local parametrization 𝜓. The dimension
of 𝑁(𝑈 ) is
dim 𝑁(𝑈 ) = dim 𝑈 + 𝑚 − 𝑛 = 𝑛 + 𝑚 − 𝑛 = 𝑚.
Since every point of 𝑁(𝑋) has such a neighborhood 𝑁(𝑈 ), 𝑁(𝑋) is a smooth manifold.
Note that 𝜎◦𝜓 ∶ 𝑈 ×ℝ𝑘 → 𝑈 is just the projection onto the first factor, which is a submersion.
Thus 𝑑(𝜎◦𝜓)(𝑢,𝑤) , is surjective at every point (𝑢, 𝑤). Hence 𝑑𝜎𝑢 is surjective at every 𝑢, and 𝜎
is a submersion.
246 12.1. Normal bundles and tubular neighborhoods
12.1.3 The Tubular Neighborhood Theorem

Now we are going to study an important application of normal bundles. Let 𝑋 ⊂ ℝ𝑛 be a

smooth manifold without boundary. The next theorem will provide us with the desired infor-
mation about the geometry of how 𝑌 sits inside its ambient space. More precisely, our goal is to
show that every smooth manifold has a special type of neighborhood. We begin with a lemma
that will give us the existence of interesting neighborhoods of 𝑋 in ℝ𝑛 :

Lemma 12.8 (𝜀-Neighborhood Lemma) Let 𝑋 ⊂ ℝ𝑛 be a smooth manifold without

boundary. There is a smooth function 𝜀 ∶ 𝑋 → ℝ>0 such that every open subset 𝑈 of
ℝ𝑛 with 𝑋 ⊂ 𝑈 contains the open subspace in ℝ𝑛

𝑋 𝜀 = {𝑦 ∈ ℝ𝑛 ∶ |𝑦 − 𝑥| < 𝜀(𝑥) for some 𝑥 ∈ 𝑋}.

If 𝑋 is compact, 𝜀 can be chosen to be constant.

Proof: For each point 𝑥 ∈ 𝑋, we can find a small radius 𝜀𝑥 such that the open ball 𝐵2𝜀𝑥 (𝑥) ⊂
𝑈 is contained in 𝑈 . We set
𝑈𝑥 ∶= 𝑋 ∩ 𝐵𝜀𝑥 (𝑥).

𝜀
∙ Claim: 𝑈𝑥 𝑥 = {𝑦 ∈ ℝ𝑛 ∶ |𝑦 − 𝑥′ | < 𝜀𝑥 for some 𝑥′ ∈ 𝑈𝑥 } ⊂ 𝑈 .

𝜀
For, 𝑦 ∈ 𝑈𝑥 𝑥 means there is an 𝑥′ ∈ 𝑈𝑥 with |𝑦−𝑥′ | < 𝜀𝑥 . But 𝑥′ ∈ 𝑈𝑥 means |𝑥′ −𝑥| < 𝜀𝑥 .
Thus the triangle inequality implies

|𝑦 − 𝛼| ≤ |𝑦 − 𝑥′ | + |𝑥′ − 𝑥| < 2𝜀𝑥 .

Thus 𝑣 ∈ 𝐵2𝜀𝑥 (𝑥) ⊂ 𝑈 by the choice of 𝜀𝑥 . The collection of all 𝑈𝑥 forms an open cover {𝑈𝑥 }
of 𝑋 ⊂ ℝ𝑛 . By the existence of partitions of unity for subsets in ℝ𝑛 , we can choose a partition
of unity {𝜌𝑖 } subordinate to the cover {𝑈𝑥 }. Now we define the function
∑
𝜀 ∶ 𝑋 → ℝ>0 , 𝑥 → 𝜌𝑖 (𝑥)𝜀𝑥
𝑖

Note that 𝜀 is a smooth function, since all the 𝜌𝑖 ’s are smooth.

∙ Claim: 𝑋 𝜀 ⊂ 𝑈 .

Let 𝑦 ∈ 𝑋 𝜀 . Then there is a 𝑥 ∈ 𝑋 such that |𝑦 − 𝑥| < 𝜀(𝑥). For this 𝑥, only finitely many
of the numbers 𝜌𝑖 (𝑥) are nonzero, say 𝜌𝑖1 (𝑥), … , 𝜌𝑖𝑛 (𝑥). This implies 𝑦 ∈ 𝑈𝑖1 ∩ ⋯ ∩ 𝑈𝑖𝑛 . Let
∑
𝜀𝑖𝑚 be the maximum of the finitely many numbers 𝜀𝑖1 , … , 𝜀𝑖𝑛 . Then, since 𝑖 𝜌𝑖 (𝑥) = 1, we
have 𝜀(𝑥) ≤ 𝜀𝑖𝑚 . Hence
𝜀𝑖𝑚
|𝑦 − 𝑥| < 𝜀(𝑥) ≤ 𝜀𝑖𝑚 implies 𝑣 ∈ 𝑈𝑖 ⊂ 𝑈.
𝑚

Thus 𝑋 𝜀 ⊂ 𝑈 .

If 𝑋 is compact, we can reduce {𝑈𝑥 } to a finite cover 𝑈𝑥1 , … , 𝑈𝑥𝑛 and let 𝜀 be the maxi-
mum of the 𝜀𝑥𝑗 .
Chapter 12. Tubular Neighborhoods and Transversality 247
Now we would like to realize such neighborhoods 𝑋𝜀 of 𝑋 using the normal bundle 𝑁(𝑋)
of 𝑋 in ℝ𝑛 . We define the map

ℎ ∶ 𝑁(𝑋) → ℝ𝑛 , (𝑥, 𝑣) → 𝑥 + 𝑣.

Since 𝑁(𝑋) is a smooth manifold in ℝ𝑛 × ℝ𝑛 and ℎ is just the restriction of the addition map
on ℝ𝑛 , ℎ is smooth. Geometrically, ℎ maps each normal space 𝑁𝑥 (𝑋) to the affine subspace
containing 𝑥 which is orthogonal to 𝑇𝑥 𝑋. The map ℎ provides us with the connection of the
neighborhood 𝑋 𝜀 with the normal bundle in the following way:

Definition 12.9 (Tubular neighborhoods) Let 𝑋 ⊂ ℝ𝑛 be a smooth manifold without

boundary. A tubular nighborhood of 𝑋 is an open subset 𝑋 𝜀 of ℝ𝑛 containing 𝑋 such
that ℎ maps an open subspace 𝑁 𝜀 (𝑋) ⊂ 𝑁(𝑋) diffeomorphically onto 𝑋 𝜀 where the
space 𝑁 𝜀 (𝑋) is given by

𝑁 𝜀 (𝑋) = {(𝑥, 𝑣) ∈ ℝ𝑛 ∶ |𝑣| < 𝜀(𝑥)}

for a smooth function 𝜀 ∶ 𝑋 → ℝ>0 .

A key feature of smooth manifolds embedded in some Euclidean space is that they always
have a tubular neighborhood. We will first prove this important fact and then study some of its
consequences.

Theorem 12.10 (Tubular Neighborhood Theorem) Let 𝑋 ⊂ ℝ𝑛 be a smooth man-

ifold without boundary. Then 𝑋 has a tubular neighborhood 𝑋 𝜀 such that there is a
submersion 𝜋 ∶ 𝑋 𝜀 → 𝑋 for which the restriction of 𝜋 to 𝑋 is the identity.

Proof: Let again 𝑋0 = 𝑁 0 (𝑋) ⊂ 𝑁(𝑋) denote the zero-section of 𝑁(𝑋), i.e., the subspace
𝑋0 = {(𝑥, 0) ∶ 𝑥 ∈ 𝑋}. The map ℎ maps 𝑋0 diffeomorphically onto 𝑋. Our goal is to show
that this extends to an open neighborhood around 𝑋0 . To do this we show that 𝑑ℎ(𝑥,0) is an
isomorphism for every point (𝑥, 0) in 𝑋0 as follows:

∙ First, we observe that, since ℎ|𝑋0 ∶ 𝑋0 → 𝑋 is a diffeomorphism, 𝑑ℎ(𝑥,0) maps 𝑇(𝑥,0) (𝑋0 ) ⊂
𝑇(𝑥,0) 𝑁(𝑋) isomorphically onto 𝑇𝑥 (𝑋) ⊂ ℝ𝑛 .

∙ Second, since the restriction of ℎ to the normal space 𝑁𝑥 (𝑋) ⊂ 𝑁(𝑋) is given by 𝑣 →
𝑥 + 𝑣, we see that 𝑑ℎ(𝑥,0) maps 𝑇(𝑥,0) ({𝑥} × 𝑁𝑥 (𝑋)) ⊂ 𝑇(𝑥,0) (𝑁(𝑋)) isomorphically onto
𝑁𝑥 (𝑋) ⊂ ℝ𝑛 .

∙ Since we have 𝑇𝑥 ℝ𝑛 = 𝑇𝑥 𝑋 ⊕ 𝑁𝑥 𝑋, this shows that 𝑑ℎ(𝑥,0) is surjective for every 𝑥.

∙ Since dim 𝑇(𝑥,0) (𝑁(𝑋)) = 𝑛, this implies that 𝑑ℎ(𝑥,0) is an isomorphism for every 𝑥.

Hence the assumptions of the generalized Inverse Function Theorem 12.4 are satisfied
and we can conclude that ℎ maps an open neighborhood of 𝑋0 in 𝑁(𝑋) diffeomorphically
onto an open neighborhood of 𝑋 in ℝ𝑛 . By the 𝜀-Neighborhood Lemma 12.8, every open
neighborhood of 𝑋 contains an 𝑋 𝜀 for a smooth function 𝜀 ∶ 𝑋 → ℝ>0 . It is clear from the
definition of ℎ that 𝑋 𝜀 is the image the open neighborhood 𝑁 𝜀 (𝑋) of 𝑋0 in 𝑁(𝑋) for the same
function 𝜀.
248 12.1. Normal bundles and tubular neighborhoods
Finally, we have shown that there is a smooth map ℎ−1 ∶ 𝑋 𝜀 → 𝑁 𝜀 (𝑋) ⊂ 𝑁(𝑋). We define
the map 𝜋 by

𝜋 = 𝜎|𝑁 𝜀 (𝑋) ◦ℎ−1 ∶ 𝑋 𝜀 → 𝑋.

Since ℎ−1 is a diffeomorphism and 𝜎 is a submersion, we conclude that 𝜋 is submersion. It

follows directly from the definition of ℎ and 𝜎 that restriction of 𝜋 to 𝑋 ⊂ 𝑋 𝜀 is the identity.

∙ We emphasize that a key point for the existence of 𝜋 is that we do not only have 𝑋 𝜀 but
also that we know that it is the diffeomorphic image of 𝑁 𝜀 (𝑋) under ℎ. Hence we really
need to use the normal bundle.

∙ For some applications, the important point of the theorem is not so much the existence
of the 𝑋 𝜀 , but rather that they come equipped with the submersion 𝜋.

∙ It is crucial that we can find an open neighborhood of 𝑋0 in 𝑁(𝑋) which is diffeomor-

phic to an open neighborhood of 𝑍 in 𝑌 . For it is clear that 𝑋 is diffeomorphic to 𝑋0
which is closed in 𝑁(𝑋). The difference may become apparent when we look at the
dimensions: dim 𝑋 𝜀 = 𝑛, whereas dim 𝑋 < 𝑛 in general.

∙ Recall from Lemma 12.8 that, if 𝑋 is compact, then 𝜀 > 0 can be chosen constant, and
𝑋 𝜀 is the open set of points in ℝ𝑛 with distance less than 𝜀 from 𝑋. If 𝜀 is sufficiently
small, then each point 𝑤 ∈ 𝑋 𝜀 possesses a unique closest point in 𝑋. Writing 𝜋(𝑤) for
this unique point, defines the map 𝜋 ∶ 𝑋 𝜀 → 𝑋 in this case. After studying the proof of
the theorem, it is a good exercise to check that this actually yields the desired map 𝜋.

∙ Tubular neighborhoods are very useful, for example for the Pontryagin–Thom construc-
tion which is key for proving more advanced results in differential topology. We will
see some applications in the next section.

12.1.4 Tubular neighborhood - the relative case

We can actually consider normal bundles more generally whenever we have a submanifold
𝑍 ⊂ 𝑋 in order to understand the geometry of 𝑍 in 𝑋:

Definition 12.11 (Normal Bundle to a submanifold) Let 𝑋 ⊂ ℝ𝑚 be a boundaryless

manifold, and let 𝑍 be a submanifold of 𝑋. We define the normal bundle to 𝑍 in 𝑋 to
be the set

𝑁(𝑍, 𝑋) ∶= {(𝑧, 𝑣) ∶ 𝑧 ∈ 𝑍, 𝑣 ∈ 𝑇𝑧 (𝑋) and 𝑣 ⟂ 𝑇𝑧 (𝑍)}.

∙ We think of this as the relative normal bundle, since we take the normal space within the
tangent space of 𝑋, not in the ambient Euclidean space.

Also relative normal bundles are manifolds on their own:

Chapter 12. Tubular Neighborhoods and Transversality 249

Theorem 12.12 (Normal bundles are manifolds revisited) The normal bundle
𝑁(𝑍, 𝑋) is a smooth manifold of dimension equal to dim 𝑋. The canonical map

𝜎 ∶ 𝑁(𝑍, 𝑋) → 𝑍, 𝜎(𝑧, 𝑣) = 𝑧,

is a submersion.

Proof: Let 𝑧 be a point in 𝑍. We can find an open subset 𝑉 ⊂ ℝ𝑚 containing 𝑧 and local
coordinate functions (𝑢1 , … , 𝑢𝑚 ) ∶ 𝑉 → ℝ𝑚 such that

𝑍 ∩ 𝑉 = {𝑣 ∈ ℝ𝑚 ∶ 𝑢1 (𝑣) = ⋯ = 𝑢𝑛 (𝑣) = 0}
and 𝑋 ∩ 𝑉 = {𝑢 ∈ ℝ𝑚 ∶ 𝑢𝑘+1 (𝑣) = ⋯ = 𝑢𝑛 (𝑣) = 0}

where 𝑛 is the codimension of 𝑍 in ℝ𝑚 and 𝑘 is the codimension of 𝑍 in 𝑋. Let 𝜑 be the

smooth map given by
𝜑 = (𝑢1 , … , 𝑢𝑛 ) ∶ 𝑉 → ℝ𝑛 .
We set 𝑈 ∶= 𝑍 ∩ 𝑉 . We observed above that the map

𝜓 ∶ 𝑈 × ℝ𝑛 → 𝑁𝑈 (𝑍, ℝ𝑚 ) ∶= (𝑈 × ℝ𝑚 ) ∩ 𝑁(𝑍, ℝ𝑚 ),
(𝑢, 𝑣) → (𝑢, 𝑑𝜑𝑇𝑢 (𝑣))

is a local parametrization of 𝑁(𝑍, ℝ𝑚 ) = 𝑁(𝑍).

By restricting 𝜓 to elements in 𝑈 × ℝ𝑘 ⊂ 𝑈 × ℝ𝑛 , we get a smooth map 𝜙 defined as the

composite
𝜓
𝑈 × ℝ𝑘 / 𝑁 (𝑍, ℝ𝑚 )
𝑈

id×𝑝
𝜙
& 
𝑁𝑈 (𝑍, 𝑋)

where 𝑁𝑈 (𝑍, 𝑋) ∶= (𝑈 × ℝ𝑚 ) ∩ 𝑁(𝑍, 𝑋) and 𝑝 is the map induced by the orthogonal pro-
jection 𝑝𝑧 ∶ ℝ𝑚 → 𝑇𝑧 (𝑋) at each 𝑧. Note that, for a vector 𝑤 ∈ ℝ𝑚 which satisfies 𝑤 ⟂ 𝑇𝑧 (𝑍),
we have 𝑝(𝑤) ∈ 𝑇𝑧 (𝑋) and 𝑝(𝑤) ⟂ 𝑇𝑧 (𝑍). Let 𝜑̃ = (𝑢𝑘+1 , … , 𝑢𝑛 ) ∶ 𝑉 → ℝ𝑛−𝑘 . We observe
that, by our choice of 𝜑 and 𝜑,
̃ we know

𝑇𝑧 (𝑍) = (Ker 𝑑𝜑𝑧 ) ⊂ Ker (𝑑 𝜑̃ 𝑧 ) = 𝑇𝑧 (𝑋)

and the orthogonal projection 𝑝𝑧 varies smoothly with 𝑧. At each 𝑧 ∈ 𝑈 , the dimension of the
kernel of the composite
𝑑𝜑𝑇𝑧 𝑝𝑧
ℝ𝑛 ←←←←←←←→
← 𝑁𝑧 (𝑍, ℝ𝑚 ) ←←←←→
← 𝑁𝑧 (𝑍, 𝑋)

is
dim Ker (𝑝𝑧 ) = dim 𝑁𝑧 (𝑍, ℝ𝑚 ) − dim 𝑁𝑧 (𝑍, 𝑋),
since 𝑑𝜑𝑇𝑧 is an isomorphism. We can calculate this dimension by

dim 𝑁𝑧 (𝑍, ℝ𝑚 ) − dim 𝑁𝑧 (𝑍, 𝑋) = 𝑚 − dim 𝑍 − (dim 𝑋 − dim 𝑍) = 𝑛 − 𝑘.

250 12.1. Normal bundles and tubular neighborhoods
Thus, 𝜙 is a diffeomorphism being the identity on the factor and a linear isomorphism on
the second factor. Hence 𝜙 ∶ 𝑈 × ℝ𝑘 → 𝑁𝑈 (𝑍, 𝑌 ) is a local parametrization of 𝑁(𝑍, 𝑋).
Since 𝑁𝑈 (𝑍, 𝑋) is open in 𝑁(𝑍, 𝑋) and every point in 𝑁(𝑍, 𝑋) lies in such an 𝑁𝑈 (𝑍, 𝑋),
we conclude that 𝑁(𝑍, 𝑋) is a smooth manifold. Its dimension is

dim 𝑁(𝑍, 𝑋) = dim 𝑈 + dim ℝ𝑘 = dim 𝑍 + dim 𝑋 − dim 𝑍 = dim 𝑋.

We note again that 𝜎◦𝜙 ∶ 𝑈 × ℝ𝑘 → 𝑈 is just the projection onto the first factor, which is a
submersion. Thus 𝑑(𝜎◦𝜙)(𝑢,𝑣) , is surjective at every point (𝑢, 𝑣). Hence 𝑑𝜎𝑢 is surjective at
every 𝑢, and 𝜎 is a submersion.

Let us look at an example of a normal bundle of an embedded submanifold:

Example 12.13 (Normal bundle to sphere) Consider 𝕊𝑘−1 as a submanifold of 𝕊𝑘

via the embedding
(𝑥1 , … , 𝑥𝑘 ) → (𝑥1 , … , 𝑥𝑘 , 0).
The tangent space 𝑇𝑝 (𝕊𝑘−1 ) is embedded in 𝑇𝑝 (𝕊𝑘 ) as the subspace consisting of vectors
with last coordinate being 0. Hence the orthogonal complement of 𝑇𝑝 (𝕊𝑘−1 ) in 𝑇𝑝 (𝕊𝑘 )
is spanned by the vector with coordinates 𝑣𝑘 ∶= (0, … , 0, 1) (in 𝑇𝑝 (𝕊𝑘 )). Hence we can
define a map

𝕊𝑘−1 × ℝ → 𝑁(𝕊𝑘−1 , 𝕊𝑘 ), (𝑝, 𝜆) → (𝑝, 𝜆𝑣𝑘 ).

This map is a diffeomorphism with inverse (𝑝, 𝜆𝑣𝑘 ) → (𝑝, 𝜆).

∙ Note that an 𝑛-dimensional vector bundle which is diffeomorphic to the product of the
base space with ℝ𝑛 is called trivial. Hence we just showed that 𝑁(𝕊𝑘−1 , 𝕊𝑘 ) is a trivial
one-dimensional bundle.

∙ We get a similar result when we consider 𝕊𝑘−1 ⊂ ℝ𝑘 for 𝑘 ≥ 2. Then, at any 𝑝 ∈ 𝕊𝑘−1 ,
the unit vector 𝑝∕|𝑝| spans the normal complement to 𝑇𝑝 (𝕊𝑘−1 ) in ℝ𝑘 . Hence there is a
diffeomorphism

𝕊𝑘−1 × ℝ → 𝑁(𝕊𝑘−1 , ℝ𝑘 ), (𝑝, 𝜆) → (𝑝, 𝜆𝑝∕|𝑝|).

Hence 𝑁(𝕊𝑘−1 , ℝ𝑘 ) is a trivial one-dimensional bundle over 𝕊𝑘−1 .

∙ However, there are a lot of nontrivial vector bundles as well. Important examples
are the tangent bundle 𝑇 (𝕊2 ) over 𝕊2 and the universal bundle over the Grassmannian
Gr 𝑘 (ℝ𝑛+𝑘 ).

∙ Note that for any 𝑧 ∈ 𝑍, the preimage 𝜎 −1 (𝑧) =∶ 𝑁𝑧 (𝑍, 𝑋) is the space of normal
vectors to 𝑍 at 𝑧 in 𝑇𝑧 (𝑋) that we have met before.

∙ Warning: As for 𝑁(𝑋), 𝑁(𝑍, 𝑋) is not intrinsic to the manifold 𝑍 but depends on the
specific relationship between 𝑍 and the surrounding space 𝑋.

Now we extend the tubular neighborhood theorem to submanifolds:

Chapter 12. Tubular Neighborhoods and Transversality 251

Theorem 12.14 (Tubular Neighborhoods - relative version) Let 𝑋 ⊂ ℝ𝑛 be a smooth

manifold without boundary, and let 𝑍 be a submanifold of 𝑋. Then there is a diffeo-
morphism of an open neighborhood 𝑍 𝜀 of 𝑍 in 𝑋 to an open neighborhood 𝑁 𝜀 (𝑍, 𝑋)
of 𝑍0 ≔ 𝑍 × {0} in 𝑁(𝑍, 𝑋). There is a submersion 𝑍 𝜀 → 𝑍 which restricts to the
identity on 𝑍.

Proof: By the Tubular Neighborhood Theorem 12.10 applied to 𝑋 ⊂ ℝ𝑛 , we have the

open neighborhood 𝑋 𝜀𝑋 ⊂ ℝ𝑛 for some positive smooth function 𝜀𝑋 and a submersion

𝜋𝑋 ∶ 𝑋 𝜀𝑋 → 𝑋.

We define again a smooth map

ℎ ∶ 𝑁(𝑍, 𝑋) → ℝ𝑛 , (𝑧, 𝑣) → 𝑧 + 𝑣.

The inverse image

𝑊 ≔ ℎ−1 (𝑋 𝜀𝑋 ) ⊂ 𝑁(𝑍, 𝑋)

is an open neighborhood of 𝑍0 in 𝑁(𝑍, 𝑋). Since ℎ(𝑧, 0) = 𝑧 for all 𝑧 ∈ 𝑍, the composition

ℎ 𝜋
← 𝑋 𝜀𝑋 ←←→
𝑓 ∶ 𝑊 ←←→ ← 𝑋

is the identity when we restrict it to 𝑍0 . By the same argument as above, we can show that
𝑑ℎ(𝑧,0) is an isomorphism at every point of 𝑍0 in 𝑁(𝑍, 𝑋). Since 𝑑𝜋𝑧 is the identity for all 𝑧 ∈
𝑍 ⊂ 𝑋 𝜀𝑋 , 𝑑𝑓(𝑧,0) is an isomorphism for every (𝑧, 0) ∈ 𝑍0 ⊂ 𝑁(𝑍, 𝑋). Hence the assumptions
of the generalized Inverse Function Theorem 12.4, are satisfied, and we can conclude that
there is an open neighborhood 𝑉 of 𝑍0 in 𝑁(𝑍, 𝑋) which is mapped diffeomorphically onto
an open neighborhood 𝑈 of 𝑍 in 𝑋 by 𝑓 = 𝜋◦ℎ. Then we can find a positive smooth function
𝜀 ∶ 𝑍 → ℝ>0 such that 𝑍 ⊂ 𝑍 𝜀 ⊂ 𝑈 and 𝑍 𝜀 is diffeomorphic to an open neighborhood
𝑁 𝜀 (𝑍, 𝑋) of 𝑍0 in 𝑁(𝑍, 𝑋). The composition

≅ 𝜎𝑁 𝜀 (𝑍,𝑋)
𝜋 ∶ 𝑍 𝜀 ←←←→
← 𝑁 𝜀 (𝑍, 𝑋) ←←←←←←←←←←←←←←←→
← 𝑍

is the desired submersion.

Finally, we make the following observation that we will use later:

Theorem 12.15 (Trivial relative normal bundle) Let 𝑋 ⊂ ℝ𝑛 be a smooth manifold

without boundary, and let 𝑍 be a submanifold of 𝑋. Then 𝑁(𝑍, 𝑋) is a trivial bundle
over 𝑍 if and only if there is a submersion 𝑔 ∶ 𝑈 → ℝ𝑘 defined on an open subset
𝑈 ⊂ 𝑋 such that 𝑍 = 𝑔 −1 (0).

Proof: First, we assume that 𝑁(𝑍, 𝑋) is trivial, i.e., there is a diffeomorphism 𝜑 ∶ 𝑍 ×

ℝ𝑘 → 𝑁(𝑍, 𝑋). By Theorem 12.14 there are open neighborhoods 𝑍 ⊂ 𝑍 𝜀 ⊂ 𝑋 and
𝑁 𝜀 (𝑍, 𝑋) ⊂ 𝑁(𝑍, 𝑋) with a diffeomorphism 𝑍 𝜀 → 𝑁 𝜀 (𝑍, 𝑋). Composition and restric-
tion yield the desired submersion 𝑔 ∶ 𝑍 𝜀 → 𝑁 𝜀 (𝑍, 𝑋) → ℝ𝑘 such that 𝑍 = 𝑔 −1 (0).
252 12.2. Whitney Approximation Theorem
Second, we assume that there is a submersion 𝑔 ∶ 𝑈 → ℝ𝑘 with 𝑍 = 𝑔 −1 (0). Hence,
for each 𝑧 ∈ 𝑍 ⊂ 𝑈 ⊂ 𝑋, 𝑑𝑔𝑧 ∶ 𝑇𝑧 𝑈 = 𝑇𝑧 𝑋 → ℝ𝑘 is a surjective linear map. This im-
plies by Lemma 12.6 that the transpose (𝑑𝑔𝑧 )𝑇 maps ℝ𝑘 isomorphically onto the orthogonal
complement of 𝑇𝑧 𝑍 in 𝑇𝑧 𝑋 for every 𝑧 ∈ 𝑍. Thus, by definition of 𝑁(𝑍, 𝑋), the map

𝜑 ∶ 𝑍 × ℝ𝑘 → 𝑁(𝑍, 𝑋), (𝑧, 𝑣) → (𝑧, (𝑑𝑔𝑧 )𝑇 (𝑣))

is a diffeomorphism.

12.2 Whitney Approximation Theorem

In this section we show that every continuous map between smooth manifolds can be approx-
imated by a smooth map. In particular, we will show that every continuous map is homotopic
to a smooth one. The key tool that makes this work are tubular neighborhoods.

12.2.1 Whitney Approximation Theorem for Functions

We begin with the approximation of maps to Euclidean space. We will show that such maps
can be approximated in the following sense:

Definition 12.16 (𝜀-close functions) Let 𝑋 be a smooth manifold and 𝑓 , 𝑔 ∶ 𝑋 → ℝ𝑛

two maps. Let 𝜀 ∶ 𝑋 → ℝ>0 be a positive continuous function. Then we say that 𝑓 and
𝑔 are said to be 𝜀-close if |𝑓 (𝑥) − 𝑔(𝑥)| < 𝜀(𝑥) for all 𝑥 ∈ 𝑋.

Theorem 12.17 (Approximating maps to Euclidean space) Let 𝑋 be a smooth

manifold with or without boundary, and let 𝑓 ∶ 𝑋 → ℝ𝑛 be a continuous map. Let
𝜀 ∶ 𝑋 → ℝ>0 be a positive continuous function. Then there exists a smooth map
𝑔 ∶ 𝑋 → ℝ𝑛 which is 𝜀-close to 𝑓 . Moreover, if 𝑓 is already smooth on a closed subset
𝐴 ⊂ 𝑋, then 𝑔 can be chosen equal to 𝑓 on 𝐴.

Proof: Let 𝐴 ⊂ 𝑋 be a closed subset and assume that 𝑓|𝐴 is smooth. By definition, this
means that there is an open subset 𝑈 ⊂ 𝑋 with 𝐴 ⊂ 𝑈 and a smooth map 𝐹 ∶ 𝑈 → ℝ𝑛 such
that 𝑓|𝐴 = 𝐹|𝐴 . We define the set 𝑈 as

𝑈0 ∶= {𝑥 ∈ 𝑋 ∶ |𝐹 (𝑥) − 𝑓 (𝑥)| < 𝜀(𝑥)}.

This is an open subset in 𝑋, since the function 𝑥 → 𝜀(𝑥) − |𝐹 (𝑥) − 𝑓 (𝑥)| is continuous and 𝑈
is the inverse image of the open subset ℝ>0 under this function. Note that 𝑋 ⧵ 𝐴 ⊂ 𝑈0 , since
𝐹 (𝑥) − 𝑓 (𝑥) = 0 for all 𝑥 ∈ 𝐴. If 𝐴 = ∅, we set 𝑈0 ∶= ∅.

Now let 𝑥 ∈ 𝑋 ⧵ 𝐴 and 𝑈𝑥 be an open neighborhood of 𝑥 in contained in 𝑋 ⧵ 𝐴 which is

small enough such that

1
|𝐹 (𝑦) − 𝑓 (𝑥)| < 𝜀(𝑥) < 𝜀(𝑦) for all 𝑦 ∈ 𝑈𝑥 .
2
Chapter 12. Tubular Neighborhoods and Transversality 253
We can find
{ such a neighborhood,
} since 𝑥 is fixed and 𝜀 and 𝐹 are continuous maps. The
collection 𝑈𝑥 ∶ 𝑥 ∈ 𝑋 ⧵ 𝐴 of all such neighborhoods for all 𝑥 is an open cover of 𝑋 ⧵ 𝐴.
{ }∞
Now we can choose a countable subcover 𝑈𝑥𝑖 and we set 𝑈𝑖 ∶= 𝑈𝑥𝑖 to simplify the
𝑖=1
notation.

Let {𝜌0 , 𝜌𝑖 } be a smooth partition of unity subordinate to the open cover {𝑈0 , 𝑈𝑖 } of 𝑋. We
define the map 𝑔 ∶ 𝑋 → ℝ𝑛 by
∑
∞
𝑔(𝑥) ∶= 𝜌0 (𝑥)𝐹 (𝑥) + 𝜌𝑖 (𝑥)𝑓 (𝑥𝑖 ).
𝑖=1

Since 𝐹 , 𝜌 and the 𝜌𝑖 are smooth and the 𝑓 (𝑥𝑖 ) are fixed values, 𝑔 is smooth. Moreover, we
∑
have 𝑔|𝐴 = 𝐹|𝐴 = 𝑓|𝐴 . It remains to show that 𝑔 and 𝑓 are 𝜀-close: Since 𝑖≥0 𝜌𝑖 (𝑥) = 1 and
|𝐹 (𝑥) − 𝑓 (𝑥𝑖 )| < 21 𝜀(𝑥) for all 𝑥 ∈ 𝑋, we see

| ( ) |
| ∑∞
∑∞
|
|𝑔(𝑥) − 𝑓 (𝑥)| = ||𝜌0 (𝑥)𝐹 (𝑥) + 𝜌𝑖 (𝑥)𝑓 (𝑥𝑖 ) − 𝜌0 (𝑥) + 𝜌𝑖 (𝑥) 𝑓 (𝑥)||
| |
| 𝑖=1 𝑖=1 |
∑
∞
≤ 𝜌0 (𝑥)|𝐹 (𝑥) − 𝑓 (𝑥)| + 𝜌𝑖 (𝑥)|𝑓 (𝑥𝑖 ) − 𝑓 (𝑥)|
𝑖=1
∑
∞
< 𝜌0 (𝑥)𝜀(𝑥) + 𝜌𝑖 (𝑥)𝜀(𝑥) = 𝜀(𝑥).
𝑖=1

12.2.2 Whitney Approximation Theorem between manifolds

Now we can extend this result to smooth manifolds embedded in ℝ𝑛 . Recall that two maps
𝑓 , 𝑔 ∶ 𝑋 → 𝑌 are said to be homotopic relative to a subset 𝐴 ⊂ 𝑋 if there is a homotopy
𝐻 ∶ 𝑋 × [0, 1] → 𝑌 such that 𝑓 (𝑥) = 𝐻(𝑥, 𝑡) = 𝑔(𝑥) for all 𝑥 ∈ 𝐴 and all 𝑡 ∈ [0, 1].

Theorem 12.18 (Whitney Approximation Theorem) Let 𝑋 be a smooth manifold

with or without boundary and let 𝑌 ⊂ ℝ𝑛 be a smooth manifold without boundary. Let
𝑓 ∶ 𝑋 → 𝑌 be a continuous map. Then 𝑓 is homotopic to a smooth map 𝑋 → 𝑌 .
Moreover, if 𝑓 is already smooth on a closed subset 𝐴 ⊂ 𝑋, then the homotopy can be
chosen relative to 𝐴.

Proof: Let 𝑌 𝜀 ⊂ ℝ𝑛 be a tubular neighborhood of 𝑌 in ℝ𝑛 and let 𝜀 ∶ 𝑌 → ℝ>0 be

the associated positive smooth function. Let 𝜋 ∶ 𝑌 𝜀 → 𝑌 be the submersion of the Tubular
Neighborhood Theorem 12.10 for which the restriction of 𝜋 to 𝑋 is the identity. We define 𝜀̃
to be the positive continuous function given as the composition

𝜀̃ = 𝜀◦𝑓 ∶ 𝑋 → ℝ>0 .

By the previous theorem, there is a smooth map 𝑔 ∶ 𝑋 → ℝ𝑛 with 𝑔|𝐴 = 𝑓|𝐴 and which is
𝜀-close
̃ to 𝑓 . Now we define the map 𝐻 ∶ 𝑋 × [0, 1] → 𝑌 as the composition of 𝜋 with a
straight-line homotopy between 𝑓 and 𝑔:

𝐻(𝑥, 𝑡) = 𝜋 ((1 − 𝑡)𝑓 (𝑥) + 𝑡𝑔(𝑥)) .

254 12.2. Whitney Approximation Theorem
∙ Claim: 𝐻 is well-defined.

We know |𝑔(𝑥) − 𝑓 (𝑥)| < 𝜀(𝑥)

̃ = 𝜀(𝑓 (𝑥)) for all 𝑥 ∈ 𝑋. Hence 𝑔(𝑥) is contained in the
open ball around 𝑓 (𝑥) with radius 𝜀(𝑓 (𝑥)). Thus, 𝑔(𝑥) and the whole line segment between
𝑓 (𝑥) and 𝑔(𝑥) are contained in 𝑌 𝜀 for all 𝑥 by definition of 𝑌 𝜀 . This proves the claim.

Hence we can conclude that 𝐻 is a homotopy between 𝐻(𝑥, 0) = 𝜋(𝑓 (𝑥)) = 𝑓 (𝑥) and
𝐻(𝑥, 1) = 𝜋(𝑔(𝑥)). Moreover, we have 𝐻(𝑥, 𝑡) = 𝑓 (𝑥) for all 𝑥 ∈ 𝐴. The map 𝜋◦𝑔 is smooth
and the desired map homotopic to 𝑓 .

Remark 12.19 (No boundary is important) The assumption on 𝑌 being without

boundary is actually crucial. For a simple example, let 𝑋 = ℝ and 𝑌 = [0, ∞). Con-
sider the continuous map 𝑓 ∶ 𝑋 → 𝑌 given by 𝑓 (𝑥) = |𝑥| and let 𝐴 = [0, ∞). Then
there does not exist a smooth map 𝑔 ∶ 𝑋 → 𝑌 with 𝑔|𝐴 = 𝑓|𝐴 . For 𝑓 is smooth on
𝑑
the open interval (0, ∞) with 𝑑𝑥 𝑓 (𝑥) = 1 for all 𝑥 ∈ (0, ∞). However, 𝑔 would have
to satisfy 𝑔(𝑥) ∈ 𝑌 , i.e., 𝑔(𝑥) ≥ 0 for all 𝑥 ∈ 𝑋. From what we learned in Calculus
we can deduce that such a map 𝑔 cannot be smooth. As a consequence we get that there
is no smooth homotopy ℎ ∶ 𝑋 × [0, 1] → 𝑌 with ℎ(𝑥, 𝑡) = 𝑓 (𝑥) for all 𝑥 ∈ 𝐴 and all
𝑡 ∈ [0, 1].

Theorem 12.20 (Homotopic maps are smoothly homotopic) Let 𝑋 be a smooth

manifold with or without boundary and let 𝑌 be a smooth manifold without boundary.
Let 𝑓 , 𝑔 ∶ 𝑋 → 𝑌 be smooth maps. Then, if 𝑓 and 𝑔 are homotopic, they are smoothly
homotopic. Moreover, if 𝑓 and 𝑔 are homotopic are relative to some closed subset
𝐴 ⊂ 𝑋, then they are smoothly homotopic relative to 𝐴.

Proof: Let ℎ ∶ 𝑋×[0, 1] → 𝑌 be a continuous homotopy between 𝑓 and 𝑔 relative to 𝐴. We

would like to apply Whitney’s Theorem 12.18 to ℎ. However, 𝑋 × [0, 1] may not be a manifold
with boundary in our sense. Hence, we first extend ℎ to a continuous map ℎ̃ ∶ 𝑋 × ℝ → 𝑌 by
setting

⎧ℎ(𝑥, 𝑡), 𝑡 ∈ [0, 1]

̃ 𝑡) = ⎪
ℎ(𝑥, ⎨ℎ(𝑥, 0), 𝑡≤0
⎪ℎ(𝑥, 1), 𝑡 ≥ 1.
⎩

The restriction of ℎ̃ to 𝑋 × {0} ∪ 𝑋 × {1} is smooth, since it is equal to 𝑓 ◦𝜋1 on 𝑋 × {1} and
𝑔◦𝜋1 where 𝜋1 ∶ 𝑋 × ℝ → 𝑋 denotes the projection. If ℎ is a homotopy relative to 𝐴, then ℎ̃
is also smooth on 𝐴 × [0, 1].

Now, since 𝑋 × ℝ is a smooth manifold, with or without boundary, Whitney’s Theo-

rem 12.18 implies that there is a smooth map 𝐻 ∶ 𝑋 × ℝ → 𝑌 whose restriction to 𝑋 ×
{0} ∪ 𝑋 × {1} ∪ 𝐴 × [0, 1] agrees with ℎ.
̃ By restricting 𝐻 to 𝑋 × [0, 1] we get the desired
smooth homotopy between 𝑓 and 𝑔 relative to 𝐴.
Chapter 12. Tubular Neighborhoods and Transversality 255
12.2.3 Compact manifolds are not contractible

The ideas we used to prove Brouwer’s fixed point Theorem 10.23 together with the approxima-
tion theorems of this section lead to the following stronger form of Theorem 11.7:

Theorem 12.21 (Compact manifolds are not contractible) Let 𝑋 be a connected

compact smooth manifold of dimension at least one without boundary. Then the identity
map from 𝑋 to 𝑋 is not homotopic to a constant map.

∙ The theorem says that no compact manifold of dimension 𝑛 ≥ 1 has the homotopy type
of a point.

∙ In Algebraic Topology, we learn about this fact when we show that the 𝑛th singular ho-
mology group of a connected compact manifold is nontrivial, more specifically we have
𝐻𝑛 (𝑋; ℤ∕2) ≅ ℤ∕2. Since homology is invariant under homotopy and 𝐻𝑛 ({𝑥0 }; ℤ∕2) =
0 for 𝑛 ≥ 1, there cannot be a homotopy equivalence between 𝑋 and a one-point space.

∙ Here, however, we give a more geometric explanation. For the theorem follows again
from the classification of one-manifolds in Theorem 10.17 together with Whitney’s Ap-
proximation Theorem 12.18.

Proof: Assume there was a smooth homotopy 𝐻 ∶ 𝑋 × [0, 1] → 𝑋 between the identity on
𝑋 and a constant map with image 𝑥0 ∈ 𝑋, i.e., 𝐻(𝑥, 0) = 𝑥 and 𝐻(𝑥, 1) = 𝑥0 for all 𝑥 ∈ 𝑋. By
Sard’s Theorem 10.16, the map 𝐻 has at least one regular value 𝑦 ∈ 𝑋. We note that, since
dim 𝑋 ≥ 1, we must have 𝑦 ≠ 𝑥0 . This follows from the fact that 𝑑𝐻(𝑥,1) fails to be surjective,
since 𝐻(−, 1) ∶ 𝑋 → 𝑋 is constant.

The fiber 𝐻 −1 (𝑦) is a smooth manifold by the Preimage Theorem 10.14. Since 𝑋 is com-
pact, so is 𝑋 × [0, 1]. Since 𝐻 −1 (𝑦) is closed in 𝑋 × [0, 1], 𝐻 −1 (𝑦) is compact as well. Since
𝐻 −1 (𝑦) has codimension dim 𝑋 in 𝑋 × [0, 1], we have dim 𝐻 −1 (𝑦) = 1.

Moreover, 𝐻 −1 (𝑦) may have a boundary, since 𝑋 has no boundary and hence 𝑋 × [0, 1] is
a manifold with boundary. The Preimage Theorem 10.14 for manifolds with boundary implies
that this boundary has the form

𝜕𝐻 −1 (𝑦) = 𝜕(𝑋 × [0, 1]) ∩ 𝐻 −1 (𝑦) = (𝑋 × {0} ∪ 𝑋 × {1}) ∩ 𝐻 −1 (𝑦).

Since 𝐻 satisfies 𝐻(𝑥, 0) = 𝑥 for all 𝑥 ∈ 𝑋, we have

(𝑋 × {0}) ∩ 𝐻 −1 (𝑦) = (𝑦, 0).

For 𝑡 = 1, however, 𝐻(−, 1) is constant with 𝐻(𝑥, 1) = 𝑥0 for all 𝑥 ∈ 𝑋. Since we know
𝑥0 ≠ 𝑦, we have

(𝑋 × {1}) ∩ 𝐻 −1 (𝑦) = ∅.

Thus we have 𝜕𝐻 −1 (𝑦) = (𝑦, 0). Hence the boundary of 𝐻 −1 (𝑦) consists of a single point. This
contradicts Lemma 10.20 which says that a compact one-manifold with boundary always has
an even number of boundary points.
256 12.3. Ehresmann’s Fibration Theorem
By Whitney’s Approximation Theorem 12.18, we know that if there was a smooth homo-
topy between the identity and a constant map, then there was a continuous homotopy between
the identity and a constant map. Since such a smooth homotopy cannot exist, we get the theo-
rem.

12.3 Ehresmann’s Fibration Theorem

In this section we study another consequence of the existence of tubular neighborhoods.

We begin with an important type of maps:

Definition 12.22 (Locally trivial fibrations) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map between

manifolds. Then 𝑓 is called a locally trivial fibration if for each 𝑦 ∈ 𝑌 there is an open
neighborhood 𝑈 ⊂ 𝑌 of 𝑦 and a diffeomorphism 𝜑 ∶ 𝑓 −1 (𝑈 ) → 𝑈 × 𝑓 −1 (𝑦) such that
the diagram
𝜑
𝑓 −1 (𝑈 ) / 𝑈 × 𝑓 −1 (𝑦)

𝑓|𝑓 −1 (𝑈 ) 𝑝𝑈
# z
𝑈

commutes where 𝑝𝑈 denotes the projection onto the first factor.

Here are some examples and remarks:

∙ Every projection 𝑝𝑋 ∶ 𝑋 × 𝑍 → 𝑋 is a locally trivial fibration. In fact, it is globally

trivial.
∙ Every vector bundle 𝜋 ∶ 𝐸 → 𝑌 is a locally trivial fibration.
∙ Assume we have a vector bundle 𝜋 ∶ 𝐸 → 𝑌 such that each fiber 𝜋 −1 (𝑦) has a metric
| ⋅ |𝑦 . Then we can form the sphere bundle 𝑓 ∶ 𝑆(𝐸) → 𝑌 as follows: we set 𝑆(𝐸) =
{𝑣 ∈ 𝐸 ∶ |𝑣|𝑦 = 1 if 𝜋(𝑣) = 𝑦} ⊂ 𝐸, and 𝑓 is just the restriction to 𝑆(𝐸). Then 𝑓 is a
locally trivial fibration.
∙ If 𝑌 is connected, then the fibers of a locally trivial fibration 𝑓 ∶ 𝑋 → 𝑌 are all diffeo-
morphic.
∙ The Hopf fibration 𝑓 ∶ 𝕊3 → 𝕊2 is a locally trivial fibration. This will become clear
after we have proven the main theorem of this section.

The following famous theorem gives us a sufficient criterion for a map to be locally trivial
based on the notions we have studied before:

Theorem 12.23 (Ehresmann Fibration Theorem) Let 𝑓 ∶ 𝑋 → 𝑌 be a proper sub-

mersion. Then 𝑓 is a locally trivial fibration. In particular, if 𝑋 is compact, every
submersion is a locally trivial fibration.
Chapter 12. Tubular Neighborhoods and Transversality 257
∙ The second assertion follows from the first, since every continuous map with a compact
domain is proper.

∙ Ehresmann’s fibration theorem is highly influential in many areas of geometry and topol-
ogy. For example, in complex and algebraic geometry it implies that that the higher direct
images along 𝑓 of a constant sheaf form a local system on 𝑌 (see [18]).

∙ Let 𝑦0 ∈ 𝑌 be a given basepoint in 𝑌 . The theorem also tells us that we can think of a
proper submersion 𝑓 ∶ 𝑋 → 𝑌 as a family of diffeomorphic copies of the fiber 𝑓 −1 (𝑦0 )
varying over the points in 𝑌 .

Proof of Theorem 12.23: The assertion is clear if 𝑋 is the empty set. So assume 𝑋 and
𝑓 −1 (𝑦) are nonempty for 𝑦 ∈ 𝑌 . Since 𝑓 is a submersion, 𝑦 is a regular value for 𝑓 . Hence
𝑍 ∶= 𝑓 −1 (𝑦) is a submanifold of 𝑋. Then the relative version of the Tubular Neighborhood
Theorem 12.14 says that there is a neighborhood 𝑍 𝜀 of 𝑍 which is open in 𝑋 and a submersion
𝜋 ∶ 𝑍 𝜀 → 𝑍 for which the restriction to 𝑍 is the identity. Then we can form the map

𝑞 ∶= (𝜋, 𝑓|𝑍 𝜀 ) ∶ 𝑍 𝜀 → 𝑍 × 𝑌 .

Since 𝑓 is proper, 𝑍 = 𝑓 −1 (𝑦) is compact. Moreover, the Preimage Theorem 4.6 tells us
that 𝑍 is of dimension dim 𝑋 − dim 𝑌 and 𝑇𝑧 (𝑍) = Ker (𝑑𝑓𝑧 ). This implies that the derivative
𝑑𝑞𝑧 of 𝑞 at any point 𝑧 ∈ 𝑍 ⊂ 𝑍 𝜀 is an isomorphism, as 𝑑𝑓𝑧 is surjective onto 𝑇𝑦 𝑌 and 𝑑𝜋𝑧
is surjective onto 𝑇𝑧 𝑍. Hence 𝑑𝑞𝑧 is a surjective linear map and since the dimension of 𝑇𝑧 𝑍 𝜀
is equal to dim 𝑋 = dim 𝑍 + dim 𝑌 , 𝑑𝑞𝑧 must be an isomorphism. Thus we can apply the
generalized Inverse Function Theorem 12.4 to 𝑍 ⊂ 𝑍 𝜀 and 𝑞. Hence we get that there is a
neighborhood 𝑍 ⊂ 𝑊 ⊂ 𝑍 𝜀 of 𝑍 which is open in 𝑍 𝜀 such that 𝑞|𝑊 is a diffeomorphism
onto an open neighborhood of 𝑞(𝑍) in 𝑍 × 𝑌 .

Finally, since 𝑓 is proper, it is a closed map. Hence, Lemma 12.24 below implies that
there is an open neighborhood 𝑈 ⊂ 𝑌 around 𝑦 such that 𝑍 ⊂ 𝑓 −1 (𝑈 ) ⊂ 𝑊 . Then we
have 𝑞(𝑓 −1 (𝑈 )) = 𝑈 × 𝑍 and have shown that 𝜑 ∶= 𝑞𝑓 −1 (𝑈 ) ∶ 𝑓 −1 (𝑈 ) → 𝑈 × 𝑓 −1 (𝑦) is a
diffeomorphism such that 𝑓|𝑓 −1 (𝑈 ) = 𝑝𝑈 ◦𝜑.

Lemma 12.24 (Closed maps) Let 𝑓 ∶ 𝑋 → 𝑌 be a continuous map between topo-

logical spaces. Then 𝑓 is closed if and only if for all 𝑦 ∈ 𝑌 and open subset 𝑊 ⊆ 𝑋
satisfying 𝑓 −1 (𝑦) ⊆ 𝑊 , there is an open neighborhood 𝑈 of 𝑦 satisfying 𝑓 −1 (𝑈 ) ⊆ 𝑊 .

Proof: We only show the implication we need. The other direction is left as an exercise.
Since 𝑊 is open, its complement 𝑋 ⧵ 𝑊 is closed in 𝑋. Since 𝑓 is closed, the image 𝑓 (𝑋 ⧵ 𝑊 )
is closed in 𝑌 . Hence the complement 𝑈 ∶= 𝑌 ⧵ 𝑓 (𝑋 ⧵ 𝑊 ) is open in 𝑌 . As 𝑓 −1 (𝑦) ⊂ 𝑊 , we
know 𝑦 ∉ 𝑓 (𝑋 ⧵ 𝑊 ) and hence 𝑦 ∈ 𝑈 and 𝑓 −1 (𝑦) ⊆ 𝑓 −1 (𝑈 ) ⊆ 𝑊 .

12.4 Thom’s Transversality Theorem

We are going to review what we have learned about transversality and show that transversality
is actually a generic property. This is a quite long and technical chapter. You may want to skip
258 12.4. Thom’s Transversality Theorem
some details for the first reading and get back to them later. The good news is that the results we
prove here will lay the ground for the intersection theory we will develop in the next chapter.
So hang in there, it is worth it!

∙ Main application:

Our main application will be to reach our goal to define an interesting intersection theory
for smooth manifolds which helps us deciding difficult questions. For example, we would like
to use it to show that 𝕊2 and ℝP2 are not homeomorphic. So assume we have a compact smooth
𝑘-dimensional manifold (without boundary), 𝑌 an 𝑛-dimensional manifold, and 𝑍 a closed 𝑚-
dimensional submanifold of 𝑌 . Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map which is transverse to 𝑍.
Then the Preimage Theorem 6.2 tells us that 𝑓 −1 (𝑍) is a 𝑘 + 𝑚 − 𝑛-dimensional submanifold
of 𝑋. Since it is a closed subset in the compact space 𝑋, 𝑓 −1 (𝑍) is also compact.

In the special case 𝑚 = 𝑛 − 𝑘, 𝑓 −1 (𝑍) is a compact manifold of dimension 0. Thus it is

a finite set of points. This is the starting point for intersection theory. For we can ask: How
many points are in the preimage 𝑓 −1 (𝑍)? Actually, we will define the mod 2-intersection
number 𝐼2 (𝑓 , 𝑍) as the number #𝑓 −1 (𝑍) modulo 2. The reason for modulo 2 will become
apparent later. It will be necessary to make 𝐼2 (𝑓 , 𝑍) homotopy-invariant. We will remedy
this defect later when we introduce orientations. Now we would like to do this with a general
smooth map. However, the assumption that 𝑓 meets 𝑍 transverse is crucial for the above to
work. The goal of this section is to show that every smooth map 𝑓 ∶ 𝑋 → 𝑌 can be replaced
up to homotopy with one a map that is transverse to 𝑍, and the resulting intersection number
will not depend on the chosen homotopic map. This will make it possible to define 𝐼2 (𝑓 , 𝑍)
in the next chapter for an arbitrary smooth map.

12.4.1 The Transversality Theorem

We start with the following extension of Sard’s Theorem:

Theorem 12.25 (Transversality Theorem) Let 𝐹 ∶ 𝑋 × 𝑆 → 𝑌 be a smooth map

of manifolds, where only 𝑋 has a boundary. Let 𝑍 be a submanifold of 𝑌 without
boundary. If both 𝐹 and 𝜕𝐹 are transverse to 𝑍, then for almost every 𝑠 ∈ 𝑆, both 𝑓𝑠
and 𝜕𝑓𝑠 are transverse to 𝑍 where 𝑓𝑠 denotes the map 𝑥 → 𝑓𝑠 (𝑥) = 𝐹 (𝑥, 𝑠).

∙ Roughly speaking, the difference between requiring that 𝐹 is transverse to 𝑍 versus 𝑓𝑠

is transversal to 𝑍 is that, for 𝐹 , the image of 𝑇(𝑥,𝑠) (𝑋 × 𝑆) under 𝑑𝐹(𝑥,𝑠) has to be big
enough, whereas for 𝑓𝑠 we look at the potentially smaller image of 𝑇(𝑥,𝑠) (𝑋 × 𝑆) under
𝑑(𝑓𝑠 )𝑥 . Similarly for 𝜕𝐹 and 𝜕𝑓𝑠 .

Proof: Since both 𝐹 and 𝜕𝐹 are transversal to 𝑍, the Preimage Theorem 10.15 implies
that 𝑊 ∶= 𝐹 −1 (𝑍) is a submanifold of 𝑋 × 𝑆 with boundary

𝜕𝑊 = 𝑊 ∩ 𝜕(𝑋 × 𝑆) = 𝑊 ∩ (𝜕𝑋 × 𝑆).

Chapter 12. Tubular Neighborhoods and Transversality 259
Let 𝜋 ∶ 𝑋 × 𝑆 → 𝑆 be the natural projection map. We will show that whenever 𝑠 ∈ 𝑆 is a
regular value for the restriction 𝜋 ∶ 𝑊 → 𝑆, then 𝑓𝑠 −
⋔ 𝑍, and whenever 𝑠 is a regular value
−
for 𝜕𝜋 ∶ 𝜕𝑊 → 𝑆, then 𝜕𝑓𝑠 ⋔ 𝑍. By Sard’s Theorem 10.16 for manifolds with boundary
almost every 𝑠 ∈ 𝑆 is a regular value for both maps, so the theorem follows.

∙ Claim: 𝑓𝑠 −
⋔ 𝑍.

Suppose 𝑓𝑠 (𝑥) = 𝑧 ∈ 𝑍. Because 𝐹 (𝑥, 𝑠) = 𝑧 and 𝐹 −

⋔ 𝑍 by the assumption, we know
that
𝑑𝐹(𝑥,𝑠) (𝑇(𝑥,𝑠) (𝑋 × 𝑆)) + 𝑇𝑧 (𝑍) = 𝑇𝑧 (𝑌 ).
Hence, given any vector 𝑎 ∈ 𝑇𝑧 (𝑌 ), there exists a vector 𝑏 ∈ 𝑇(𝑥,𝑠) (𝑋 × 𝑆) such that

𝑑𝐹(𝑥,𝑠) (𝑏) − 𝑎 ∈ 𝑇𝑧 (𝑍).

We need to find a vector 𝑣 ∈ 𝑇𝑥 (𝑋) such that

𝑑(𝑓𝑠 )𝑥 (𝑣) − 𝑎 ∈ 𝑇𝑧 (𝑍),
as that would show that 𝑑(𝑓𝑠 )𝑥 (𝑇𝑥 (𝑋)) + 𝑇𝑧 (𝑍) = 𝑇𝑧 (𝑌 ). Since
𝑇(𝑥,𝑠) (𝑋 × 𝑆) = 𝑇𝑥 (𝑋) × 𝑇𝑠 (𝑆),
we can write 𝑏 as a pair (𝑤, 𝑒) for vectors 𝑤 ∈ 𝑇𝑥 (𝑋) and 𝑒 ∈ 𝑇𝑠 (𝑆).

If 𝑒 is zero, we are done: Since 𝑓𝑠 is the restriction of 𝐹 to 𝑋 × {𝑠}, it follows that 𝑑(𝑓𝑠 )𝑥
is the restriction of 𝑑𝐹(𝑥,𝑠) to 𝑇𝑥 (𝑋) × {0} ⊂ 𝑇𝑥 (𝑋) × 𝑇𝑠 (𝑆), i.e., the diagram

𝑇𝑥 (𝑋) × {0}  / 𝑇 (𝑋) × 𝑇 (𝑆)
𝑥 𝑠

𝑑𝐹(𝑥,𝑠)
𝑑(𝑓𝑠 )𝑥 ( 
𝑇𝑧 (𝑌 )

commutes and hence

𝑑𝐹(𝑥,𝑠) (𝑏) = 𝑑𝐹(𝑥,𝑠) (𝑤, 0) = 𝑑(𝑓𝑠 )𝑥 (𝑤).

However, 𝑒 need not be zero. But we may use the projection 𝜋 to modify 𝑤 and 𝑒 as
follows: It is an exercise to check that
𝑑𝜋(𝑥,𝑠) ∶ 𝑇𝑥 (𝑋) × 𝑇𝑠 (𝑆) → 𝑇𝑠 (𝑆)
is just projection onto the second factor. In fact, this holds for every projection map from a
product of manifolds.

Now we use the assumption that 𝑠 is a regular value of 𝜋. For this implies that the restric-
tion of 𝑑𝜋(𝑥,𝑠) to 𝑇(𝑥,𝑠) (𝑊 )

𝑇(𝑥,𝑠) (𝑊 )  / 𝑇 (𝑋) × 𝑇 (𝑆)
𝑥 𝑠

𝑑𝜋(𝑥,𝑠)
𝑑𝜋(𝑥,𝑠) )|𝑇(𝑥,𝑠) (𝑊 ) ' 
𝑇𝑠 (𝑆)
260 12.4. Thom’s Transversality Theorem
is surjective. In particular, the fiber over 𝑒 ∈ 𝑇𝑠 (𝑆) is nonempty, and there is some vector of
the form (𝑢, 𝑒) in 𝑇(𝑥,𝑠) (𝑊 ).

But 𝐹|𝑊 ∶ 𝑊 = 𝐹 −1 (𝑍) → 𝑍 is a map to 𝑍, so the vector 𝑑𝐹(𝑥,𝑠) (𝑢, 𝑒) is an element in

𝑇𝑧 (𝑍). Consequently, the vector 𝑣 ∶= 𝑤 − 𝑢 ∈ 𝑇𝑥 (𝑋) is the vector in 𝑇𝑥 (𝑋) we were looking
for:

𝑑(𝑓𝑠 )𝑥 (𝑣) − 𝑎 = 𝑑𝐹(𝑥,𝑠) (𝑤 − 𝑢, 0) − 𝑎

= 𝑑𝐹(𝑥,𝑠) ((𝑤, 𝑒) − (𝑢, 𝑒)) − 𝑎
= (𝑑𝐹(𝑥,𝑠) (𝑤, 𝑒) − 𝑎) − 𝑑𝐹(𝑥,𝑠) (𝑢, 𝑒),

and we remember that both 𝑑𝐹(𝑥,𝑠) (𝑤, 𝑒)−𝑎 and 𝑑𝐹(𝑥,𝑠) (𝑢, 𝑒) belong to 𝑇𝑧 (𝑍). Hence 𝑑(𝑓𝑠 )𝑥 (𝑣)−
𝑎 is in 𝑇𝑧 (𝑍) as we wished to prove.

∙ Claim: 𝜕𝑓𝑠 −
⋔ 𝑍.

This is a special instance of what we just proved, for the case of the boundaryless manifold
𝜕𝑋 and the map 𝜕𝐹 ∶ (𝜕𝑋) × 𝑆 → 𝑌 .

12.4.2 Transversality is generic

This shows that transversality for smooth maps 𝑋 → ℝ𝑁 is generic in the following sense:

∙ Let 𝑓 ∶ 𝑋 → ℝ𝑁 be a smooth map. Let 𝑆 be an open ball in ℝ𝑀 , and define

𝐹 ∶ 𝑋 × 𝑆 → ℝ𝑁 , 𝐹 (𝑥, 𝑠) = 𝑓 (𝑥) + 𝑠.

The derivative of 𝐹 at (𝑥, 𝑠) is

𝑑𝐹(𝑥,𝑠) = (𝑑𝑓𝑥 , Idℝ𝑁 ) ∶ 𝑇𝑥 (𝑋) × ℝ𝑁 → ℝ𝑁 .

Thus 𝑑𝐹(𝑥,𝑠) is surjective at any (𝑥, 𝑠). Hence 𝐹 is a submersion. This implies that 𝐹
is transverse to every submanifold 𝑍 ⊂ ℝ𝑁 . Now we can apply the Transversality
Theorem 12.25 we have just proven:
Let 𝑍 ⊂ ℝ𝑁 be a manifold. Since 𝐹 and 𝜕𝐹 are transverse to 𝑍, for almost every 𝑠 ∈ 𝑆,
the map 𝑓𝑠 (𝑥) = 𝑓 (𝑥) + 𝑠 is transversal to 𝑍. Moreover, the map

𝑋 × [0, 1] → ℝ𝑁 , (𝑥, 𝑡) → 𝑓𝑡𝑠 (𝑥) = 𝑓 (𝑥) + 𝑡𝑠

provides a smooth homotopy between 𝑓 and 𝑓𝑠 .

∙ This shows us that transversality is generic for maps 𝑋 → ℝ𝑁 . We would like to gen-
eralize this result to an arbitrary boundaryless smooth manifold 𝑌 ⊂ ℝ𝑁 and smooth
map 𝑓 ∶ 𝑋 → 𝑌 .
Chapter 12. Tubular Neighborhoods and Transversality 261
∙ Given a submanifold 𝑍 ⊂ 𝑌 , we have just learned how to vary a smooth map 𝑓 ∶ 𝑋 →
𝑌 ⊂ ℝ𝑁 as a family of maps 𝑋 → ℝ𝑁 such that 𝑓𝑠 − ⋔ 𝑍 for arbitrarily small 𝑠, where we
consider 𝑍 as a submanifold in ℝ𝑁 . It remains to project these maps down to 𝑌 such
that a small perturbation 𝑓𝑠 of 𝑓 remains transversal to the given submanifold 𝑍 ⊂ 𝑌 .
We can do this using normal bundles and tubular neighboorhoods we constructred in
the previous section.

As a first consequence we get:

Theorem 12.26 (Creating families of submersions) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map

where 𝑌 is a manifold without boundary. Let 𝑆 be an open ball in ℝ𝑁 . Then there is
a smooth map 𝐹 ∶ 𝑋 × 𝑆 → 𝑌 such that 𝐹 (𝑥, 0) = 𝑓 (𝑥), and for any fixed 𝑥 ∈ 𝑋, the
map

𝑆 → 𝑌 , 𝑠 → 𝐹 (𝑥, 𝑠) is a submersion.

In particular, both 𝐹 and 𝜕𝐹 are submersions.

Proof: Let 𝑌 ⊂ ℝ𝑁 and 𝑆 be the open unit ball in ℝ𝑁 . We define

𝐹 ∶ 𝑋 × 𝑆 → 𝑌 , 𝐹 (𝑥, 𝑠) = 𝜋(𝑓 (𝑥) + 𝜀(𝑓 (𝑥))𝑠). (12.1)

Recall the map

𝜋∶ 𝑌𝜀 → 𝑌

from the Tubular Neighborhood Theorem 12.10. Since 𝜋 restricts to the identity on 𝑌 , we
have
𝐹 (𝑥, 0) = 𝜋(𝑓 (𝑥) + 0) = 𝑓 (𝑥).

For fixed 𝑥, the map

𝜑 ∶ 𝑆 → 𝑌 𝜀 , 𝑠 → 𝑓 (𝑥) + 𝜀(𝑓 (𝑥))𝑠

is the translation of a linear map. Thus 𝑑𝜑𝑠 is just given by multiplying a vector in 𝑇𝑠 (𝑆) =
ℝ𝑀 by the real number 𝜀(𝑓 (𝑥)) > 0 to get a vector in 𝑇𝜑(𝑠) (𝑌 𝜀 ) ⊂ ℝ𝑁 . This derivative is just
𝜀(𝑓 (𝑥)) times the identity of ℝ𝑁 , and therefore surjective. Thus 𝜑 is a submersion.

As the composition of two submersions is a submersion, we get that

𝜑 𝜋
← 𝑌 , 𝑠 → 𝐹 (𝑥, 𝑠) is a submersion.
← 𝑌 𝜀 ←←→
𝐹 (𝑥, −) = 𝜋◦𝜑 ∶ 𝑆 ←←←→

Hence the restriction 𝐹{𝑥}×𝑆 ∶ {𝑥} × 𝑆 → 𝑌 of 𝐹 to the submanifold {𝑥} × 𝑆 is a submersion

for every 𝑥 ∈ 𝑋. Since every point of 𝑋 × 𝑆 lies in one of these submanifolds, 𝐹 must be a
submersion as well. The same argument applied to 𝜕𝐹 and 𝜕𝑋, shows that 𝜕𝐹 is a submersion.

Now we can prove that transversality is generic:

262 12.4. Thom’s Transversality Theorem

Theorem 12.27 (Transversality Homotopy Theorem) For every smooth map

𝑓 ∶ 𝑋 → 𝑌 and any boundaryless submanifold 𝑍 of the boundaryless manifold 𝑌 ,
there exists a smooth map 𝑔 ∶ 𝑋 → 𝑌 such that 𝑔 is homotopic to 𝑓 and transversal to
𝑍, i.e., such that
𝑔 ≃ 𝑓 and 𝑔 −
⋔ 𝑍 and 𝜕𝑔 −⋔ 𝑍.

Proof: For the family of mappings 𝐹 of the previous corollary of the 𝜀-Neighborhood
Theorem, the Transversality Theorem implies that 𝑓𝑠 −
⋔ 𝑍 and 𝜕𝑓𝑠 −
⋔ 𝑍 for almost all 𝑠 ∈ 𝑆.
But each 𝑓𝑠 is homotopic to 𝑓 , the homotopy being
𝑋 × 𝐼 → 𝑌 , (𝑥, 𝑡) → 𝐹 (𝑥, 𝑡𝑠).

12.4.3 The Extension Theorem

In fact, Tubular Neighborhood Theorem 12.10 allows us to prove a stronger form of the
Transversality Homotopy Theorem 12.27. In order to be able to formulate it, we need some
terminology.

Definition 12.28 (Transversality on subsets) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map,

𝑍 ⊂ 𝑌 a submanifold, and 𝐶 ⊂ 𝑋 be a subset. We will say 𝑓 is transversal to 𝑍 on
𝐶, if the transversality condition

Im (𝑑𝑓𝑥 ) + 𝑇𝑓 (𝑥) (𝑍) = 𝑇𝑓 (𝑥) (𝑌 ) (12.2)

is satisfied for all 𝑥 ∈ 𝐶 ∩ 𝑓 −1 (𝑍).

Note that, even if 𝐶 is a submanifold, this is different than requiring 𝑓|𝐶 − ⋔ 𝑍,
since (12.2) involves Im (𝑑𝑓𝑥 ) = 𝑑𝑓𝑥 (𝑇𝑥 (𝑋)), not Im (𝑑(𝑓|𝐶 )𝑥 ) = 𝑑𝑓𝑥 (𝑇𝑥 (𝐶)), which is
smaller in general.

Now we can formulate an important technical result.

Theorem 12.29 (Extension Theorem) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map, 𝑌 without

boundary, and 𝑍 a closed submanifold of 𝑌 without boundary. Let 𝐶 be a closed subset
of 𝑋. Assume that 𝑓 −⋔ 𝑍 on 𝐶 and 𝜕𝑓 − ⋔ 𝑍 on 𝐶 ∩ 𝜕𝑋. Then there exists a smooth
map 𝑔 ∶ 𝑋 → 𝑌 homotopic to 𝑓 , such that 𝑔 − ⋔ 𝑍 and 𝜕𝑔 − ⋔ 𝑍, and on an open
neighborhood of 𝐶 we have 𝑔 = 𝑓 .

Since the boundary 𝜕𝑋 is closed in 𝑋, we obtain the important special case:

Theorem 12.30 (Extension of maps on boundaries) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth

map such that 𝜕𝑓 − ⋔ 𝑍. Then there exists a map 𝑔 ∶ 𝑋 → 𝑌 homotopic to 𝑓 such that
𝜕𝑔 = 𝜕𝑓 and 𝑔 − ⋔ 𝑍.
In particular, suppose there is a smooth map ℎ ∶ 𝜕𝑋 → 𝑌 transverse to 𝑍. Then, if
ℎ extends to any smooth map 𝑋 → 𝑌 , it extends to a smooth map 𝑓 ∶ 𝑋 → 𝑌 that is
transverse to 𝑍.
Chapter 12. Tubular Neighborhoods and Transversality 263
For the proof of the Extension Theorem we need a lemma:

Lemma 12.31 (Extending functions) If 𝑈 is an open subset which contains the

closed set 𝐶 in 𝑋, then there exists a smooth function 𝛾 ∶ 𝑋 → [0, 1] that is equal to
one outside 𝑈 but that is zero on a neighborhood of 𝐶.

Proof: Let 𝐶 ′ be any closed set contained in 𝑈 that contains 𝐶 in its interior, and let {𝜌𝑖 }
be a partition of unity subordinate to the open cover {𝑈 , 𝑋 ⧵ 𝐶 ′ }. Here it comes handy that
we proved the existence of partition of unity for arbitrary subsets of ℝ𝑁 . Then just take 𝛾 to
be the sum of those 𝜌𝑖 that vanish outside of 𝑋 ⧵ 𝐶 ′ .

Proof of the Extension Theorem 12.29:

First we show that 𝑓 − ⋔ 𝑍 on a neighborhood of 𝐶 i.e. an open subset containing 𝐶. If

𝑥 ∈ 𝐶 but 𝑥 ∉ 𝑓 −1 (𝑍), then since 𝑍 is closed, 𝑋 ⧵ 𝑓 −1 (𝑍) is a neighborhood of 𝑥 on which
𝑓−⋔ 𝑍 automatically.

If 𝑥 ∈ 𝑓 −1 (𝑍), then there is a neighborhood 𝑊 of 𝑓 (𝑥) in 𝑌 and a submersion 𝜑 ∶ 𝑊 → ℝ𝑘

such that 𝑓 −⋔ 𝑍 at a point 𝑥′ ∈ 𝑓 −1 (𝑍 ∩ 𝑊 ) precisely when 𝜑◦𝑓 is regular at 𝑥′ . But if 𝜑◦𝑓
is regular at 𝑥, so it is regular in a neighborhood of 𝑥. Thus 𝑓 −⋔ 𝑍 on a neighborhood of every
point 𝑥 ∈ 𝐶, and so
𝑓− ⋔ 𝑍 on a neighborhood 𝑈 of 𝐶 in 𝑋.

Second, let 𝛾 be the function in the above lemma for the closed subset 𝐶 and the open
neighborhood 𝑈 of 𝐶 in 𝑋. We set 𝜏 ∶= 𝛾 2 . Since

𝑑𝜏𝑥 = 2𝛾(𝑥)𝑑𝛾𝑥 , hence 𝛾(𝑥) = 0 = 𝜏(𝑥) ⇒ 𝑑𝜏𝑥 = 0.

Now we modify the map 𝐹 ∶ 𝑋 × 𝑆 → 𝑌 which we defined in (12.1) in proving the Homo-
topy Theorem 12.27, where 𝑆 is the unit ball in ℝ𝑀 . and set

𝐺 ∶ 𝑋 × 𝑆 → 𝑌 , 𝐺(𝑥, 𝑠) ∶= 𝐹 (𝑥, 𝜏(𝑥)𝑠).

∙ Claim: 𝐺 −
⋔ 𝑍.

For suppose that (𝑥, 𝑠) ∈ 𝐺−1 (𝑍), and let us assume first 𝜏(𝑥) ≠ 0. Then the map

𝑆 → 𝑌 , 𝑟 → 𝐺(𝑥, 𝑟),

is a submersion, since it is the composition of the

diffeomorphism 𝑟 → 𝜏(𝑥)𝑟 with the submersion 𝑟 → 𝐹 (𝑥, 𝑟).

Hence 𝐺 is regular at (𝑥, 𝑠), so certainly 𝐺 −

⋔ 𝑍 at (𝑥, 𝑠).

To show the claim when 𝜏(𝑥) = 0, we need to check that the image of the derivative 𝑑𝐺(𝑥,𝑠)
is big enough. To do this, we introduce the map

𝑚 ∶ 𝑋 × 𝑆 → 𝑋 × 𝑆, (𝑥, 𝑠) → (𝑥, 𝜏(𝑥)𝑠).

264 12.4. Thom’s Transversality Theorem
We would like to calculate the derivative of 𝑚. Therefor, we apply the product rule to the
second coordinate and remember that 𝜏 ∶ 𝑋 → [0, 1], i.e. 𝜏(𝑥) and 𝑑𝜏𝑥 (𝑣) are both in ℝ for any
𝑣 ∈ 𝑇𝑥 (𝑋). Then we get

𝑑𝑚(𝑥,𝑠) (𝑣, 𝑤) = (𝑣, 𝜏(𝑥) ⋅ 𝑤 + 𝑑𝜏𝑥 (𝑣) ⋅ 𝑠)

where 𝑤 and 𝑠 are vectors in ℝ𝑁 .

We observe that 𝐺 = 𝐹 ◦𝑚. Hence in order to calculate the derivative of 𝐺, we can apply
the chain rule. Since we are interested in the case where 𝜏(𝑥) = 0 and 𝑑𝜏𝑥 = 0 we get

𝑑𝐺(𝑥,𝑠) (𝑣, 𝑤) = 𝑑𝐹(𝑥,𝑠) (𝑣, 0).

Moreover, since 𝐹 (𝑥, 0) = 𝑓 (𝑥) for all 𝑥 by construction of 𝐹 , we know 𝐹|𝑋×{0} = 𝑓 . This
implies
𝑑𝐹(𝑥,𝑠) (𝑣, 0) = 𝑑𝐹(𝑥,0) (𝑣, 0) = 𝑑𝑓𝑥 (𝑣).

Hence we get

𝑑𝐺(𝑥,𝑠) (𝑣, 𝑤) = 𝑑𝑓𝑥 (𝑣)

and therefore

Im (𝑑𝐺(𝑥,𝑠) ) = Im (𝑑𝑓𝑥 (𝑣)) ⊂ 𝑇𝑓 (𝑥) (𝑌 ). (12.3)

Now 𝜏(𝑥) = 0, implies 𝑥 ∈ 𝑈 by definition of 𝛾 and 𝜏. But by the choice of 𝑈 above, this
implies 𝑓 −
⋔ 𝑍 at 𝑥. Hence (12.3) implies 𝐺 −
⋔ 𝑍 at (𝑥, 𝑠).

The same argument shows 𝜕𝐺 −

⋔ 𝑍.

Now we can apply the Transversality Theorem 12.25 to 𝐺 ∶ 𝑋 × 𝑆 → 𝑌 and conclude

that we can pick and fix an 𝑠 for which the map

𝑔(𝑥) ∶= 𝐺(𝑥, 𝑠) satisfies 𝑔 −

⋔ 𝑍 and 𝜕𝑔 −
⋔ 𝑍.

The map 𝐺 is then a homotopy

𝑓 = 𝐹|𝑋×{0} = 𝐺|𝑋×{0} ∼ 𝐺|𝑋×{𝑠} = 𝑔.

Finally, if 𝑥 belongs to the neighborhood of 𝐶 on which 𝜏 = 0, then we even have

𝑔(𝑥) = 𝐺(𝑥, 𝑠) = 𝐹 (𝑥, 0) = 𝑓 (𝑥).

Let us summarize what we have achieved:

∙ (Summary) We have proven three key results about transversality:

(a) The Transversality Theorem 12.25 says that when a homotopy 𝐹 is transversal
to 𝑍, then, in this homotopy family, almost every 𝑓𝑠 = 𝐹 (−, 𝑠) is transversal to
𝑍.
Chapter 12. Tubular Neighborhoods and Transversality 265
(b) The Transversality Homotopy Theorem 12.27 says that given a map 𝑓 and a sub-
manifold 𝑍, then there exists a map 𝑔 transversal to 𝑍 and 𝑔 is homotopic to
𝑓.

(c) The Extension Theorem 12.29 says that, given a map 𝑓 which is transversal to 𝑍
on a subset 𝐶, then we can always replace 𝑓 with a homotopic map 𝑔 which is
transversal to 𝑍 everywhere (not only on 𝐶) and 𝑓 = 𝑔 on an open set containing
𝐶.

(a) is a generalization of Sard’s Theorem 10.16. For (b) and (c), the key for the proof
is the Tubular Neighborhood Theorem 12.10.
266 12.5. Exercises and more examples
12.5 Exercises and more examples

12.5.1 Transversality

Exercise 12.1 As a first test, answer the following questions:

(a) Let 𝑧 = (𝑎, 𝑏) ∈ 𝕊1 ⊆ ℝ2 and let 𝑁𝑧 = {(𝑎, 𝑦) ∶ 𝑦 ∈ ℝ} be the vertical line

intersecting the circle at 𝑧. When is 𝕊1 ⊆ ℝ2 transverse to 𝑁𝑧 ⊆ ℝ2 ?

(b) Which of the following linear spaces intersect transversally?

∙ The 𝑥𝑦-plane and the 𝑧-axis.

∙ The 𝑥𝑦-plane and the plane spanned by {(3, 2, 0), (0, 4, −1)}.
∙ The plane spanned by {(1, 0, 0), (2, 1, 0)} and the 𝑦-axis in ℝ3 .
∙ ℝ𝑘 × {0} and {0} × ℝ𝑙 in ℝ𝑛 . (The answer depends on 𝑘, 𝑙, and 𝑛.)
∙ 𝑉 × {0} and the diagonal in 𝑉 × 𝑉 , for a real vector space 𝑉 .
∙ The spaces of symmetric (𝐴𝑡 = 𝐴) and skew symmetric (𝐴𝑡 = −𝐴) matrices
in 𝑀(𝑛).

(c) Do 𝑆𝐿(𝑛) and 𝑂(𝑛) meet transversally in 𝑀(𝑛)?

Exercise 12.2 Recall the maps

( )
𝑒𝑡 + 𝑒−𝑡 𝑒𝑡 − 𝑒−𝑡
2
𝑓 ∶ ℝ → ℝ , 𝑡 → , and 𝑔 ∶ ℝ2 → ℝ, (𝑥, 𝑦) → 𝑥2 − 𝑦2
2 2

from previous exercise sets. Is the set Im (𝑓 ), the image of 𝑓 in ℝ2 , a manifold? Is the
set (𝑔◦𝑓 )−1 (1) a manifold?

𝑓 𝑔
Exercise 12.3 Let 𝑋 ←←←→ ← 𝑍 be a sequence of smooth maps between manifolds,
← 𝑌 ←←→
and let 𝑊 ⊂ 𝑍 be a submanifold. Assume that 𝑔 is transversal to 𝑊 . Show:

𝑓−
⋔ 𝑔 −1 (𝑊 ) if and only if (𝑔◦𝑓 ) −
⋔𝑊.

Exercise 12.4 Let 𝑉 be a vector space, and let Δ be the diagonal of 𝑉 × 𝑉 . For a linear
map 𝐴 ∶ 𝑉 → 𝑉 , consider the graph Γ(𝐴) = {(𝑣, 𝐴𝑣) ∶ 𝑣 ∈ 𝑉 }. Show that Γ(𝐴) − ⋔ Δ if
and only if +1 is not an eigenvalue of 𝐴.

Exercise 12.5 Let 𝑓 ∶ 𝑋 → 𝑋 be a map, and let 𝑥 be a fixed point of 𝑓 , i.e., 𝑓 (𝑥) = 𝑥.
If +1 is not an eigenvalue of 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑥 (𝑋), then 𝑥 is called a Lefschetz fixed
point of 𝑓 . The map 𝑓 is called a Lefschetz map if all its fixed points are Lefschetz. Prove
that if 𝑋 is compact and 𝑓 is Lefschetz, then 𝑓 has only finitely many fixed points.
Hint: Show that the intersection of the graph of 𝑓 and the diagonal of 𝑋 is a 0-
Chapter 12. Tubular Neighborhoods and Transversality 267
dimensional submanifold of 𝑋 × 𝑋.

Exercise 12.6 Consider the following intersections in ℂ5 ⧵ {0}:

𝑆𝑘7 = {𝑧21 + 𝑧22 + 𝑧23 + 𝑧34 + 𝑧6𝑘−1

5
= 0} ∩ {|𝑧1 |2 + |𝑧2 |2 + |𝑧3 |2 + |𝑧4 |2 + |𝑧5 |2 = 1}.

Prove that 𝑆𝑘7 is a 7-dimensional manifold by showing that the intersection is transver-
sal in ℂ5 ⧵ {0}.
Hint:
( At some point you may want ) to show that, at a point 𝑧 = (𝑧1 , … , 𝑧5 ), the vector
𝑤 ≔ 𝑚2 𝑧1 , 𝑚2 𝑧2 , 𝑚2 𝑧3 , 𝑚3 𝑧4 , 6𝑘−1
𝑚
𝑧5 , with 𝑚 ≔ 2 ⋅ 3 ⋅ (6𝑘 − 1), lies in one of the tangent
spaces but not in the other.
13. Intersection Theory modulo 2

13.1 Intersection Numbers modulo 2

A classical geometric approach to classifying maps is to study their fibres. This is directly
related to other fundamental problems in mathematics. For example, if 𝑓 ∶ 𝑋 → 𝑌 is a map
defined by an equation and given a value 𝑦 ∈ 𝑌 , the set {𝑥 ∈ 𝑋 ∶ 𝑓 (𝑥) = 𝑦} is the set of
solutions of the equation 𝑓 (𝑥) = 𝑦. In geometric terms, we could rephrase the question which
𝑥 solve equation 𝑓 (𝑥) = 𝑦 by asking how 𝑓 meets or intersects the subspace {𝑦} in 𝑌 .

Building on the methods we have developed so far, we are going to exploit this geometric
idea to derive interesting and powerful invariants. We will start with intersection numbers
modulo 2, i.e., intersection numbers with values in ℤ∕2. In order to define a ℤ-valued invariant
we will have to introduce orientations later.

13.1.1 Intersecting manifolds

Let us start with a natural situation.

Definition 13.1 (Intersecting manifolds) Two submanifolds 𝑋 and 𝑍 inside 𝑌 have

complementary dimension if

dim 𝐗 + dim 𝐙 = dim 𝐘.

We assume all manifolds are without boundary for the moment.

If 𝑋 −
⋔ 𝑍, the Preimage Theorem 4.6 tells us that their intersection 𝑋 ∩𝑍 is manifold
with codim (𝑋 ∩ 𝑍) in 𝑋 being equal to codim 𝑍 in 𝑌 . Since codim 𝑍 = dim 𝑋, 𝑋 ∩ 𝑍
is a zero-dimensional manifold.
If we further assume that both 𝑋 and 𝑍 are closed and that at least one of them, say
𝑋, is compact, then 𝑋 ∩ 𝑍 must be a finite set of points. We are going to think of this
finite number of points in 𝑋 ∩ 𝑍 as the intersection number of 𝑋 and 𝑍, denoted by
#(𝑋 ∩ 𝑍).

We would like to generalize the notion of intersection numbers. A first obstacle is that if 𝑋
and 𝑍 do not intersect transversally, then it makes in general no sense to count the points in
𝑋 ∩ 𝑍. Hence, once again, transversality is key.

Luckily, we have learned how to move or deform manifolds to make intersections transver-
sal: we can alter them in homotopic families. And since embeddings form a stable class of
maps, i.e., for any homotopy 𝑖𝑡 of an embedding 𝑖0 , there is an 𝜀 > 0 such that 𝑖𝑡 is still an
embedding for all 𝑡 < 𝜀, any small homotopy of 𝑖 gives us another embedding 𝑋 → 𝑌 and thus

268
Chapter 13. Intersection Theory modulo 2 269

Figure 13.1: A tangential or non-transverse intersection is not stable. Any slight perturbation
will change the number of intersection points. We need to avoid such situations.

produces an image manifold that is a diffeomorphic copy of 𝑋 adjacent to the original.

But we still have to be careful. For the intersection number may depend on how we move
or deform the manifold.

Figure 13.2: We can move the two circles in the plane and either move them apart or make them
intersect transversely. In both cases we get a stable number of intersection points. However, 0
is not equal 2, unless we work modulo 2. That is what we are going to do.

For example, take two circles in ℝ2 . Assume that they intersect non-transversally, i.e.,
they touch each other in a point such that both tangent spaces agree and together just span a line.
Then we can move the circles by a simple translations 𝑥 → 𝑥 + 𝑡𝑎 in direction 𝑎 such that they
intersect either in two points or in no points. In both cases, the intersection is transversal, but
the intersection numbers do not agree. We observe, however, that the parity of the intersection
numbers is preserved, i.e., up to a multiple of 2 the intersection numbers after moving into a
transversal intersection agree.

13.1.2 Intersection Number modulo 2

This observation is the starting point for the following generalization:

270 13.1. Intersection Numbers modulo 2

Definition 13.2 (Mod 2 Intersection numbers) Let 𝑋 be a compact manifold, and

let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map transverse to the closed manifold 𝑍 in 𝑌 . Assume
dim 𝑋 + dim 𝑍 = dim 𝑌 . Then 𝑓 −1 (𝑍) is a closed submanifold of 𝑋 of codimension
equal to dim 𝑋. Hence 𝑓 −1 (𝑍) is of dimension zero, and therefore a finite set. We
define the mod 2 intersection number of the map 𝑓 with 𝑍, denoted 𝐼2 (𝑓 , 𝑍), to be
the number of points in 𝑓 −1 (𝑍) modulo 2:

𝐼2 (𝑓 , 𝑍) ∶= #𝑓 −1 (𝑍) mod 2.

For an arbitrary smooth map 𝑔 ∶ 𝑋 → 𝑌 , we can choose a map 𝑓 ∶ 𝑋 → 𝑌

that is homotopic to 𝑔 and transversal to 𝑍 by the Transversality Homotopy Theo-
rem 12.27. Then we define 𝐼2 (𝑔, 𝑍) ∶= 𝐼2 (𝑓 , 𝑍).

Of course, we need to check that the intersection number does not depend on the choice of
homotopic map. The key technical result that allows us to show independence is the Extension
Theorem 12.29 which says the following: Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map, 𝑌 boundaryless,
and 𝑍 a closed submanifold of 𝑌 without boundary. Let 𝐶 be a closed subset of 𝑋. Assume
that 𝑓 −
⋔ 𝑍 on 𝐶 and 𝜕𝑓 − ⋔ 𝑍 on 𝐶 ∩ 𝜕𝑋. Then there exists a smooth map 𝑔 ∶ 𝑋 → 𝑌
homotopic to 𝑓 such that 𝑔 −⋔ 𝑍 and 𝜕𝑔 − ⋔ 𝑍, and on a neighborhood of 𝐶 we have 𝑔 = 𝑓 .

∙ Independence of the chosen homotopy:

Let 𝑋, 𝑌 and 𝑍 ⊂ 𝑌 are boundaryless manifolds. The product 𝑋 × [0, 1] is then a manifold
with boundary. We let 𝐶 be the boundary of 𝑋 × [0, 1], i.e., 𝐶 is the closed subset

𝐶 ∶= 𝜕(𝑋 × [0, 1]) = 𝑋 × {0} ∪ 𝑋 × {1}.

Now we apply the Extension Theorem 12.29 to the case of a smooth homotopy

𝐹 ∶ 𝑋 × [0, 1] → 𝑌 .

Then 𝜕𝐹 , i.e., 𝐹 restricted to the boundary of 𝑋 × [0, 1], is given by the two maps

𝑓0 = 𝐹 (−, 0) ∶ 𝑋 → 𝑌 and 𝑓1 = 𝐹 (−, 1) ∶ 𝑋 → 𝑌 .

The two conditions 𝐹 −⋔ 𝑍 on 𝐶 and 𝜕𝐹 −

⋔ 𝐶 on 𝐶 ∩𝜕𝑋 are thus equivalent, and mean 𝑓0 −
⋔𝑍
− − −
and 𝑓1 ⋔ 𝑍. Hence, assuming 𝑓0 ⋔ 𝑍 and 𝑓1 ⋔ 𝑍, the Extension Theorem 12.29 says that
there is a smooth map

𝐺 ∶ 𝑋 × [0, 1] → 𝑌 with 𝐆 −
⋔ 𝐙 and 𝜕𝐺 −
⋔ 𝑍,

and 𝐺 = 𝐹 on a neighborhood of 𝐶 = 𝜕(𝑋 × [0, 1]). The latter means that

𝐺 is still a homotopy between 𝑓0 = 𝐺(−, 0) and 𝑓1 = 𝐺(−, 1).

Now we are ready to prove the following crucial observation:

Chapter 13. Intersection Theory modulo 2 271

Lemma 13.3 (Mod 2 Intersection Numbers are well-defined) If 𝑓0 ∶ 𝑋 → 𝑌 and

𝑓1 ∶ 𝑋 → 𝑌 are homotopic and both transverse to 𝑍, then 𝐼2 (𝑓0 , 𝑍) = 𝐼2 (𝑓1 , 𝑍).

Proof: Let 𝐹 ∶ 𝑋 × 𝐼 → 𝑌 be a homotopy of 𝑓0 and 𝑓1 . By the above discussion, we may

assume that 𝐹 −⋔ 𝑍. By the Preimage Theorem 10.14 with boundary, this implies 𝐹 −1 (𝑍) is
a submanifold of 𝑋 × [0, 1] such that

codim 𝐹 −1 (𝑍) in 𝑋 × [0, 1] = codim 𝑍 in 𝑌 .

Hence

dim 𝐹 −1 (𝑍) = dim(𝑋 × [0, 1]) + dim 𝑍 − dim 𝑌

= dim 𝑋 + 1 + dim 𝑍 − dim 𝑌
=1

since we assume that dim 𝑋 + dim 𝑍 = dim 𝑌 . Moreover, the boundary of 𝐹 −1 (𝑍) is

𝜕𝐹 −1 (𝑍) = 𝐹 −1 (𝑍) ∩ 𝜕(𝑋 × [0, 1]) = 𝑓0−1 (𝑍) × {0} ∪ 𝑓1−1 (𝑍) × {1}.

Since 𝑋 is compact, 𝐹 −1 (𝑍) is compact. Hence Lemma 10.20, which follows from the classi-
fication of compact one-manifolds, implies that 𝜕𝐹 −1 (𝑍) must have an even number of points.
Thus, computing mod 2, we get

𝐼2 (𝑓0 , 𝑍) = #𝑓0−1 (𝑍) = #𝑓1−1 (𝑍) = 𝐼2 (𝑓1 , 𝑍).

We can generalise this a bit further.

Lemma 13.4 (All homotopic maps have equal intersection numbers) If 𝑔0 ∶ 𝑋 → 𝑌

and 𝑔1 ∶ 𝑋 → 𝑌 are arbitrary homotopic maps, then 𝐼2 (𝑔0 , 𝑍) = 𝐼2 (𝑔1 , 𝑍).

Proof: As before, we can choose maps 𝑓0 − ⋔ 𝑍 and 𝑓1 −⋔ 𝑍 such that 𝑔0 ∼ 𝑓0 , 𝐼2 (𝑔0 , 𝑍) =

𝐼2 (𝑓0 , 𝑍), and 𝑔1 ∼ 𝑓1 , 𝐼2 (𝑔1 , 𝑍) = 𝐼2 (𝑓1 , 𝑍). Since homotopy is a transitive relation1 , we
have

𝑓0 ∼ 𝑔0 ∼ 𝑔1 ∼ 𝑓1 , and hence 𝑓0 ∼ 𝑓1 .

By the previous Lemma 13.3, this implies

𝐼2 (𝑔0 , 𝑍) = 𝐼2 (𝑓0 , 𝑍) = 𝐼2 (𝑓1 , 𝑍) = 𝐼2 (𝑔1 , 𝑍).

Remark 13.5 (Intersection number and Brouwer degree mode 2) Assume 𝑋 is

compact, 𝑌 is connected and dim 𝑋 = dim 𝑌 . Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. Then
we recover the mod 2 Brouwer degree as a mod 2 intersection number as follows. We
let 𝑍 = {𝑦} for any 𝑦 ∈ 𝑌 and get from the definition of both sides that
𝐼2 (𝑓 ; 𝑦) = deg2 (𝑓 ).
1
Recall that we proved that it is an equivalence relation
272 13.2. Intersection of manifolds and self-intersections
The difference in the two approaches to the degree is that in Section 11.1.1 we used the
Isotopy Lemma 11.3 to show that we can use any regular value for 𝑦 to calculate the
degree. For 𝐼2 (𝑓 ; 𝑦), we use the machinery of Thom Transversality to replace 𝑓 with a
homotopic map if necessary. While the approach in Section 11.1.1 is more direct and
simpler, Thom Transversality provides us with a more general theory.

13.2 Intersection of manifolds and self-intersections

Now that we have a solid notion of intersection numbers modulo 2 for maps and submanifolds,
let us return to the situation we started with:

Definition 13.6 (Mod 2 intersection numbers of submanifolds) Assume 𝑋 is a

compact submanifold of 𝑌 and 𝑍 a closed submanifold of 𝑌 . Assume the dimensions
are complementary, i.e., dim 𝑋 + dim 𝑍 = dim 𝑌 . Then we can define the mod 2
intersection number of 𝑋 with 𝑍, denoted by 𝐼2 (𝑋, 𝑍), by

𝐼2 (𝑋, 𝑍) ∶= 𝐼2 (𝑖, 𝑍)

where 𝑖 ∶ 𝑋 → 𝑌 is the inclusion. Note that if 𝑋 −

⋔ 𝑍, then

𝐼2 (𝑋, 𝑍) = #(𝑋 ∩ 𝑍).

In general, we have to move or deform 𝑋 into a transverse position.

Here some important situations:

Example 13.7 (Obstruction to pull apart circles) If 𝐼2 (𝑋, 𝑍) ≠ 0, then no matter

how 𝑋 is moved or deformed, it cannot be pulled entirely away from 𝑍.

∙ For example, on the torus 𝑌 = 𝕊1 × 𝕊1 , the two circles 𝕊1 × {1} and {1} × 𝕊1
have complimentary dimensions and nonzero mod 2 intersection number.

Example 13.8 (Self-intersection number) If dim 𝑌 = 2 dim 𝑋, then we may consider

𝐼2 (𝑋, 𝑋) as the mod 2 self-intersection number of 𝑋.

∙ An important and interesting example is the central circle 𝑋 on the open Möbius
band 𝑌 . See Figure 13.3. In Exercise 13.6 we show that 𝐼2 (𝑋, 𝑋) = 1. With-
out actually calculating the intersection number, we can already deduce from the
Intermediate Value Theorem in Calculus that we cannot continuously deform 𝑋
such that it does not intersect itself anymore.
Chapter 13. Intersection Theory modulo 2 273

Figure 13.3: We can construct the Möbius band by taking a rectangle and glue to opposite sides
after a twist. The central line then becomes a central circle on the band. If we move it a bit from
its initial position, the new curve has to intersect the original central line. This is a consequence
of the Intermediate Value Theorem, since the two ends of the curve must be on opposite sides
of the central line because of the twist.

13.3 Obstruction to extensions to boundaries

If 𝑋 happens to be the boundary of some 𝑊 in 𝑌 , then 𝐼2 (𝑋, 𝑍) = 0. For if 𝑍 −

⋔ 𝑋, then,
roughly speaking, 𝑍 leaves 𝑊 as often as it enters. Hence #(𝑋 ∩ 𝑍) should be even. This
heuristic can be made rigorous as follows:

Theorem 13.9 (Boundary Theorem for intersection numbers) Suppose that 𝑋 is the
boundary of some compact manifold 𝑊 and 𝑔 ∶ 𝑋 → 𝑌 is a smooth map. If 𝑔 can be
extended to all of 𝑊 , then 𝐼2 (𝑔, 𝑍) = 0 for every closed submanifold 𝑍 in 𝑌 which
satisfies dim 𝑋 + dim 𝑍 = dim 𝑌 .

Proof: Let 𝐺 ∶ 𝑊 → 𝑌 be an extension of 𝑔, i.e., 𝐺 is smooth and 𝜕𝐺 = 𝑔. From the

Transversality Homotopy Theorem 12.27, we obtain a map 𝐹 ∶ 𝑊 → 𝑌 homotopic to 𝐺
with 𝐹 −
⋔ 𝑍 and 𝜕𝐹 −⋔ 𝑍. We write 𝑓 ∶= 𝜕𝐹 . Then 𝑓 ∼ 𝑔 and hence

𝐼2 (𝑔, 𝑍) = 𝐼2 (𝑓 , 𝑍) = #𝑓 −1 (𝑍) mod 2.

Now 𝐹 −1 (𝑍) is a compact submanifold whose codimension in 𝑊 is the same as the

codimension of 𝑍 in 𝑌 . Here we use again that 𝑋 is the boundary of 𝑊 , for this implies
dim 𝑊 = dim 𝜕𝑊 + 1 = dim 𝑋 + 1, and hence

dim 𝐹 −1 (𝑍) = dim 𝑋 + 1 − dim 𝑌 + dim 𝑍 = 1.

Hence 𝐹 −1 (𝑍) is a compact one-dimensional manifold with boundary, so #𝜕(𝐹 −1 (𝑍))

274 13.4. Intersecting circles on 𝕊𝑛 versus ℝP𝑛
is even and hence

#𝜕(𝐹 −1 (𝑍)) = #(𝜕𝐹 )−1 (𝑍) = #𝑓 −1 (𝑍) is even.

This has an interesting consequence:

Theorem 13.10 (Obstruction for extending maps) Let 𝑊 be a compact manifold

and 𝑔 ∶ 𝜕𝑊 → 𝑌 be a smooth map. If there is at least one closed submanifold 𝑍 ⊂ 𝑌
such that 𝐼2 (𝑔, 𝑍) ≠ 0, then 𝑔 cannot be extended to a smooth map 𝑊 → 𝑌 on all of
𝑊.

We will see some applications of the previous two theorems in the exercises.

13.4 Intersecting circles on 𝕊𝑛 versus ℝP𝑛

In this section we discuss an example for how intersection theory can be used to show the exis-
tence of non-trivial elements in the fundamental group. Recall from Section 9.2 real projective
𝑛-space ℝP𝑛 which consists of the set of equivalence classes [𝑥0 ∶ 𝑥1 ∶ … ∶ 𝑥𝑛 ] of (𝑛+1)-tuples
of real numbers with the equivalence relation

(𝑥0 , 𝑥1 , … , 𝑥𝑛 ) ∼ (𝜆𝑥0 , 𝜆𝑥1 , … , 𝜆𝑥𝑛 ) for 𝜆 ∈ ℝ ⧵ {0}.

Recall that we showed that ℝP𝑛 is an n-dimensional smooth manifold. Alternatively, we can
describe ℝP𝑛 as a the quotient of 𝕊𝑛 under the equivalence relation 𝑥 ∼ −𝑥, i.e., we identify
antipodal points on 𝕊𝑛 to obtain ℝP𝑛 :

ℝP𝑛 = 𝕊𝑛 ∕(𝑥 ∼ −𝑥),

and we have a smooth map

𝕊𝑛 → ℝP𝑛 , 𝑥 → [𝑥]

which sends a point 𝑥 in 𝕊𝑛 to its equivalence class [𝑥] in ℝP𝑛 . One can show that ℝP1 and 𝕊1
are diffeomorphic. However, this is the exception special to dimension one as we will now find
out.

We begin with a recollection of the following definition which we previously studied in the
exercises. See Exercise 8.3 and Exercise 10.7.

Definition 13.11 (Simply-connected) Let 𝑋 be a connected topological space. Then

𝑋 is called simply-connected if every continuous map 𝕊1 → 𝑋 is homotopic to a con-
stant map.

For every 𝑛 ≥ 2, the 𝑛-dimensional sphere 𝕊𝑛 is an example of a simply-connected space:2

2
A more thorough treatment of this fact can be found for example in [15, Theorems 59.1 and 59.3].
Chapter 13. Intersection Theory modulo 2 275

Theorem 13.12 (𝕊𝑛 is simply-connected for 𝑛 ≥ 2) For 𝑛 ≥ 2, the 𝑛-dimensional

sphere 𝕊𝑛 is simply-connected.

Remark 13.13 This implies that the fundamental group of 𝕊𝑛 , for all 𝑛 ≥ 2, is trivial,
i.e., 𝜋1 (𝕊𝑛 , 𝑥) = 0 for every 𝑥 ∈ 𝕊𝑛 .

In order to prove Theorem 13.12 we are going to use the following lemma:

Lemma 13.14 (Paths avoiding a point) Let 𝑛 ≥ 2 and 𝑝 ∈ 𝕊𝑛 . Every path 𝛾 in 𝕊𝑛

with 𝛾(0) ≠ 𝑝 ≠ 𝛾(1) is homotopic to a path in 𝕊𝑛 ⧵ {𝑝}.

Proof of Lemma 13.14: Let 𝑈 = 𝕊𝑛 ⧵{𝑝} and 𝑉 = 𝕊𝑛 ⧵{−𝑝}. Then 𝕊𝑛 is the union the two
open subsets 𝑈 and 𝑉 . Both 𝑈 and 𝑉 are diffeomorphic to ℝ𝑛 via stereographic projection. By
Lebesgue’s number lemma3 , we can find a subdivision 𝑏0 < ⋯ < 𝑏𝑚 of the unit interval [0, 1]
such that, for each 𝑖, the set 𝛾([𝑏𝑖−1 , 𝑏𝑖 ]) is contained either in 𝑈 or in 𝑉 . If there is an index
𝑖 such that 𝛾(𝑏𝑖 ) ∉ 𝑈 ∩ 𝑉 , i.e., if 𝛾(𝑏𝑖 ) ∈ {𝑝, −𝑝}, then we modify the subdivision as follows:
Each of the sets 𝛾([𝑏𝑖−1 , 𝑏𝑖 ]) and 𝛾([𝑏𝑖 , 𝑏𝑖+1 ]) is either contained in 𝑈 or in 𝑉 . If 𝛾([𝑏𝑖 ]) ∈ 𝑈 ,
then both of these sets must lie in 𝑈 . Then we may delete 𝑏𝑖 from our subdivision to obtain a
new subdivision 𝑐0 < … < 𝑐𝑚−1 that still satisfies the condition that 𝛾([𝑐𝑖−1 , 𝑐𝑖 ]) is contained
either in 𝑈 or in 𝑉 for each 𝑖. If 𝛾([𝑏𝑖 ]) ∈ 𝑉 we proceed similarly. After a finite number of
steps, this leads to a subdivision 𝑎0 < ⋯ < 𝑎𝑛 of [0, 1] such that 𝛾([𝑎𝑖−1 , 𝑎𝑖 ]) is contained either
in 𝑈 or in 𝑉 and 𝛾(𝑎𝑖 ) ∈ 𝑈 ∩ 𝑉 for each 𝑖.

For each 𝑖, we choose a path 𝛾𝑖 ∶ [𝑎𝑖−1 , 𝑎𝑖 ] → 𝕊𝑛 as follows: If 𝛾([𝑎𝑖−1 , 𝑎𝑖 ]) ⊂ 𝑈 , we set

𝛾𝑖 = 𝛾|[𝑎𝑖−1 ,𝑎𝑖 ] . If 𝛾([𝑎𝑖−1 , 𝑎𝑖 ]) ⊂ 𝑉 , then we let 𝛾𝑖 be a path which is homotopic to 𝛾|[𝑎𝑖−1 ,𝑎𝑖 ]
which avoids 𝑝 with 𝛾𝑖 (𝑎𝑖−1 ) = 𝛾(𝑎𝑖−1 ) and 𝛾𝑖 (𝑎𝑖 ) = 𝛾(𝑎𝑖 ). We can find such a path, since
𝑉 is homeomorphic to ℝ𝑛 . Composing the 𝛾𝑖 ’s as paths yields a path [0, 1] → 𝕊𝑛 which is
homotopic to 𝛾 with the same start- and endpoint and which avoids 𝑝.

Proof of Theorem 13.12: Let 𝑝 ∈ 𝕊𝑛 be any point. Stereographic projection defines a

diffeomorphism 𝕊𝑛 ⧵ {𝑝} ≅ ℝ𝑛 . In ℝ𝑛 every continuous map is homotopic to a constant map.
Hence the theorem follows from Lemma 13.14.

As a consequence, we can show that the intersection theory for circles on 𝕊𝑛 is rather simple.

Example 13.15 (Intersecting circles on 𝕊2 ) Let us look at the concrete case for 𝕊2 .
We define a smooth map
𝑗 ∶ 𝕊1 → 𝕊2 , (cos(2𝜋𝑡), sin(2𝜋𝑡)) → (cos(2𝜋𝑡), sin(2𝜋𝑡), 0)
which maps the circle on the intersection of the sphere with the 𝑥𝑦-plane in ℝ3 . This map
is homotopic to the constant map 𝕊1 → 𝕊2 , 𝑡 → (0, 0, 1). We can show this directly us-
ing the homotopy ((cos(2𝜋𝑡), sin(2𝜋𝑡)), 𝑠) → (𝑠 cos(2𝜋𝑡), 𝑠 sin(2𝜋𝑡), 1−𝑠). Hence we can
smoothly
( 1 move 𝑗(𝕊1 ) away from itself and get that, on 𝕊2 , the self-intersection number
)
𝐼2 𝑗(𝕊 ), 𝑗(𝕊1 ) is zero.

3
A fact from general topology: Let (𝑋, 𝑑) be a compact metric space, and let  be an open cover of 𝑋. Then
there is a number 𝜆 > 0 such that for every 𝑥 ∈ 𝑋, there is a 𝑈 ∈  with 𝐵𝜆 (𝑥) ⊂ 𝑈 .
276 13.4. Intersecting circles on 𝕊𝑛 versus ℝP𝑛

Note that if we define, for example, a map 𝑗2 ∶ 𝕊1 → 𝕊2 , 𝑡 → (0, sin(2𝜋𝑡), cos(2𝜋𝑡)),

then 𝑗(𝕊1 ) ∩ 𝑗2 (𝕊1 ) consists of two points: the points (0, 1, 0) and (0, −1, 0). Hence
modulo 2, we still get 𝐼2 (𝑗(𝕊1 ), 𝑗2 (𝕊1 )) = 0.

Our next goal is to use the fact that 𝕊𝑛 is simply-connected for 𝑛 ≥ 2 to show that 𝕊𝑛 and
ℝP𝑛 are not homeomorphic for 𝑛 ≥ 2. In fact, we will prove a stronger statement. We therefore
recall the following definition:

Definition 13.16 (Homotopy equivalence) Let 𝑋 and 𝑌 be topological spaces. A

continuous map 𝑓 ∶ 𝑋 → 𝑌 is called a homotopy equivalence if there exists a contin-
uous map 𝑔 ∶ 𝑌 → 𝑋 such that 𝑓 ◦𝑔 ≃ id𝑌 and 𝑔◦𝑓 ≃ id𝑌 . We call 𝑔 a homotopy
inverse of 𝑓 . We say that 𝑋 and 𝑌 are homotopy equivalent if there exists a homotopy
equivalence 𝑋 → 𝑌 .

∙ Every homeomorphism is also a homotopy equivalence. However, the converse is not

true. A contractible space 𝑋 homotopy equivalent to a one-point space, but it may not
be homeomorphic to it. For example, ℝ𝑛 is homotopy equivalent to the one-point space
{0} consisting of the origin in ℝ𝑛 , but ℝ𝑛 is not homeomorphic to {0} for 𝑛 ≥ 1.

Here we are going to exploit the following observation:

Lemma 13.17 Let 𝑋 and 𝑌 be topological spaces. Assume that 𝑋 and 𝑌 are homotopy
equivalent. Then 𝑋 is simply-connected if and only if 𝑌 is simply-connected.

Proof: Let 𝑓 ∶ 𝑋 → 𝑌 be a homotopy equivalence with homotopy inverse 𝑔 ∶ 𝑌 → 𝑋. Let

𝐻 ∶ 𝑌 × [0, 1] → 𝑌 be a homotopy between 𝐻(𝑦, 1) = 𝑓 ◦𝑔(𝑦) and 𝐻(𝑦, 0) = 𝑦 for all 𝑦 ∈ 𝑌 .
Assume that 𝑋 is simply-connected, and let 𝛾 ∶ 𝕊1 → 𝑌 be a continuous map. The composition
𝛾 𝑔
𝑔◦𝛾 ∶ 𝕊1 ←←→ ← 𝑋 is homotopic to a constant map 𝑐. Let ℎ ∶ 𝕊1 × [0, 1] → 𝑋 be a homotopy
← 𝑌 ←←→
with ℎ(𝑠, 1) = 𝑔(𝛾(𝑠)) and ℎ(𝑠, 0) = 𝑐. Then we can define a homotopy Γ ∶ 𝕊1 × [0, 1] → 𝑌 by
{
𝐻(𝛾(𝑠), 2𝑡) if 0 ≤ 𝑡 ≤ 1∕2
Γ(𝑠, 𝑡) =
𝑓 (ℎ(𝑠, 2(1 − 𝑡))) if 1∕2 ≤ 𝑡 ≤ 1.

We need to check that this map is well-defined. For 𝑡 = 1∕2 we compute

𝐻(𝛾(𝑠), 1) = 𝑓 ◦𝑔(𝛾(𝑠)) = 𝑓 (𝑔(𝛾(𝑠)) = 𝑓 (ℎ(𝑠, 1)).

Moreover, we have

Γ(𝑠, 0) = 𝛾(𝑠) and Γ(𝑠, 1) = 𝑓 (𝑐) for all 𝑠 ∈ 𝕊1 .

Thus Γ is a homotopy between 𝛾 and the constant map with value 𝑓 (𝑐) ∈ 𝑌 . This shows that
𝑌 is simply-connected. By applying the argument with the roles of 𝑋 and 𝑌 reversed, we get
the assertion.
Chapter 13. Intersection Theory modulo 2 277

Theorem 13.18 (ℝP𝑛 is not simply-connected) For every 𝑛 ≥ 1, there is a loop

𝕊1 → ℝP𝑛 which is not homotopic to a constant map.

Proof: The case 𝑛 = 1 follows from ℝP1 ≅ 𝕊1 and the fact that 𝕊1 is not simply-connected.
This follows from deg2 (𝑧 → 𝑧) = 1. So now we assume 𝑛 ≥ 2. Our strategy is to find a smooth
map 𝑓 ∶ 𝕊1 → ℝP𝑛 and a suitable closed submanifold 𝑍 ⊂ ℝP𝑛 such that 𝐼2 (𝑓 , 𝑍) is defined
and non-zero. Then we know that 𝑓 is not smoothly homotopic to a constant map. For, if
it was, then 𝑓 would be homotopic to a constant map whose image was disjoint to 𝑍. Then
we would have 𝑓 − ⋔ 𝑍, and hence 𝐼2 (𝑓 , 𝑍) = 0. This is not possible if our strategy works
out. Then we use Whitney’s Approximation Theorem 12.20 which implies that if there was a
continuous homotopy between 𝑓 and a constant map, then there also was a smooth homotopy.
Since the latter does not exist, the former cannot exist either. So let us find suitable 𝑓 and 𝑍.

Recall that, given a smooth manifold 𝑋, a smooth map 𝑓 ∶ 𝕊1 → 𝑋 is equivalent to a

smooth map on the open interval 𝑔 ∶ (−𝜀, 2𝜋 + 𝜀) → 𝑋 such that 𝑔(0) = 𝑔(2𝜋) for some 𝜀 > 0.
Hence we may define a map 𝑓 by
𝑓 ∶ 𝕊1 → ℝP𝑛 , 𝑡 → [cos(𝑡∕2) ∶ sin(𝑡∕2) ∶ 0 ∶ … ∶ 0].
Now we embed 𝕊𝑛−1 as a submanifold of 𝕊𝑛 by requiring the coordinate 𝑥0 to be zero and let
𝑍 ⊂ ℝP𝑛 be its image under the quotient map 𝕊𝑛 → ℝP𝑛 :
𝑍 = {[𝑥] = [𝑥0 ∶ 𝑥1 ∶ … ∶ 𝑥𝑛 ] ∈ ℝP𝑛 |𝑥0 = 0}.
We can check that 𝑍 is a submanifold of dimension 𝑛 − 1 of ℝP𝑛 by restricting the standard
charts (𝑉𝑖 , 𝜙𝑖 ) for ℝP𝑛 : Recall that we defined in Section 9.2, for 0 ≤ 𝑖 ≤ 𝑛, the subsets
𝑉𝑖 = {[𝑥] ∈ 𝑍 ⊂ ℝP𝑛 ∶ 𝑥𝑖 ≠ 0}
and homeomorphisms
1
𝜙−1 𝑛
𝑖 ∶ 𝑉𝑖 → ℝ , [𝑥0 ∶ … ∶ 𝑥𝑖 ∶ … ∶ 𝑥𝑛 ] → (𝑥 , … , 𝑥̂𝑖 , … , 𝑥𝑛 ).
𝑥𝑖 0
Then the subsets 𝑉𝑖𝑍 ∶ 𝑍 ∩ 𝑉𝑖 and maps (𝜙−1 𝑍
𝑖 )|𝑉𝑖𝑍 ∶ 𝑉𝑖 → ℝ
𝑛−1 provide charts for 𝑍. Since

𝕊𝑛−1 is a closed subset in 𝕊𝑛 , 𝑍 is closed in ℝP𝑛 by definition of the quotient topology on ℝP𝑛 .
Hence the assumptions for applying intersection theory are satisfied.

The only point of 𝑍 which is hit by the image of 𝑓 is the point [0 ∶ 1 ∶ 0 ∶ … ∶ 0] ∈ 𝑍.

Hence the preimage 𝑓 −1 (𝑍) consists of exactly one point, namely the only point in 𝕊1 that is
mapped to [0 ∶ 1 ∶ 0 ∶ … ∶ 0]. Since we describe 𝑓 as a map [0, 2𝜋] → ℝP𝑛 which extends
to a smooth map over the boundary points of [0, 2𝜋], we get 𝑓 −1 (𝑍) = {𝜋} ⊂ [0, 2𝜋]. Now
we need to check that 𝑓 and 𝑍 are transverse in ℝP𝑛 . We can do this by studying the impact
of the derivative 𝑑𝑓𝜋 of 𝑓 at 𝜋 and the tangent space of 𝑍 locally, i.e., in a chart around the
point [0 ∶ 1 ∶ 0 ∶ … ∶ 0]. Under the map 𝜙1 ∶ ℝ𝑛 → 𝑉1 the point [0 ∶ 1 ∶ 0 ∶ … ∶ 0] ∈ 𝑉1
corresponds to the origin. For 𝕊1 , we choose the local parametrization 𝜓 ∶ (𝜋 − 𝜀, 𝜋 + 𝜀) → 𝕊1 ,
𝑡 → (cos(𝑡∕2), sin(𝑡∕2)). Then we consider the diagram
𝑓
𝕊O 1 / ℝP𝑛
𝜓 𝜙−1
1

(𝜋 − 𝜀, 𝜋 + 𝜀) / ℝ𝑛
𝑓̃
278 13.4. Intersecting circles on 𝕊𝑛 versus ℝP𝑛
where 𝑓̃ is defined as the composition of the other maps such that the diagram commutes. We
can compute 𝑓̃ as
1
𝑓̃(𝑡) = ⋅ (cos(𝑡∕2), 0, … , 0).
sin(𝑡∕2)
The derivative of 𝑓̃ at 𝑡 = 𝜋 is then the linear map

𝑑 𝑓̃𝜋 ∶ ℝ → ℝ𝑛 , 𝑥 → (−1∕2) ⋅ (𝑥, 0, … , 0)

given by multiplying 𝑥 with the number 𝑓̃′ (𝑡) = − 21 1

and embedding it as the first co-
sin2 (𝑡∕2)
ordinate. The key consequence of this computation is that the image of 𝑑 𝑓̃𝜋 in ℝ𝑛 is the line
ℝ × {0} ⊂ ℝ𝑛 . On the other hand, the tangent space at the origin of the image of 𝑍 under
𝜙−1
1
is the subspace {0} × ℝ𝑛−1 ⊂ ℝ𝑛 . Thus this tangent space and the image of 𝑑 𝑓̃𝜋 together
span all of ℝ𝑛 . By definition of the smooth structure on the (abstract) manifold ℝP𝑛 , the map
𝜙1 ∶ ℝ𝑛 → 𝑉1 is a diffeomorphism. Thus the induced map on tangent spaces is an isomor-
phism. Since tangent spaces are determined locally, this shows that 𝑓 and 𝑍 are transverse in
ℝP𝑛 . Thus we have proved

𝐼2 (𝑓 , 𝑍) = #𝑓 −1 (𝑍) = 1

as we set out to show.

Theorem 13.18 implies that the fundamental group of ℝP𝑛 is non-trivial. Using techniques
from algebraic topology we can say even more:

Remark 13.19 (Fundamental group of ℝP𝑛 ) In Exercise 13.7 we will show that
2[𝑓 ] = 0 in 𝜋1 (ℝP2 ), and hence [𝑓 ] generates a subgroup ℤ∕2 inside 𝜋1 (ℝP2 ). One
can then check that 𝜋1 (ℝP2 ) ≅ ℤ∕2 and extend this computation, for example using
induction and the Seifert–van Kampen Theorem, to show

𝜋1 (ℝP𝑛 , 𝑥0 ) ≅ ℤ∕2 for 𝑛 ≥ 2

where 𝑥0 is the point [1 ∶ 0 ∶ … ∶ 0].

Example 13.20 (Intersecting circles on ℝP2 ) Let us look at the concrete case 𝑛 = 2.
The map 𝑓 of the above proof is given by
𝑓 ∶ 𝕊1 → ℝP2 , 𝑡 → [cos(𝑡∕2) ∶ sin(𝑡∕2) ∶ 0].
In Exercise 13.7, we study the map 𝑓 in more detail and check that 𝑓 is proper and
injective. The submanifold 𝑍
𝑍 = {[𝑥] = [𝑥0 ∶ 𝑥1 ∶ 𝑥2 ] ∈ ℝP2 |𝑥0 = 0}
corresponds to the image of 𝕊1 under the map
𝕊1 → ℝP2 , 𝑡 → [0 ∶ sin(𝑡∕2) ∶ cos(𝑡∕2)].
The intersection of the image of 𝑓 and 𝑍 therefore is the intersection of the images of
the two circles in ℝP2 . This intersection consists of the single point
[0 ∶ 1 ∶ 0] ∈ 𝑓 (𝕊1 ) ∩ 𝑍.
Chapter 13. Intersection Theory modulo 2 279

Hence, in ℝP2 , there is exactly one intersection point. If we had considered the map
and submanifold corresponding to 𝑓 and 𝑍 for 𝕊2 instead of ℝP2 , we would have gotten
two intersection points: (0, 1, 0) and (0, −1, 0). Hence on 𝕊2 , the corresponding mod 2
intersection number is zero. This fits well with our the fact that 𝕊2 is simply-connected.
On ℝP2 , however, we have 𝐼2 (𝑓 , 𝑍) which implies that it is impossible to move 𝑓 (𝕊1 )
and 𝑍 within ℝP2 such that they do not meet.

By Lemma 13.17, Theorem 13.12 and Theorem 13.18 imply that there cannot be a homo-
topy equivalence between 𝕊𝑛 and ℝP𝑛 for 𝑛 ≥ 2. Since there cannot be a homeomorphism
between spaces which are not homotopy equivalent, we have proven the following theorem:

Theorem 13.21 (𝑛 ≥ 2: 𝕊𝑛 and ℝP𝑛 are not homeomorphic) For 𝑛 ≥ 2, there is no

homotopy equivlance between 𝕊𝑛 and ℝP𝑛 . In particular, there is no homeomorphism
between 𝕊𝑛 and ℝP𝑛 for 𝑛 ≥ 2.

13.5 ℝ𝑛 as a commutative division algebra

We will now discuss an important and at first glance maybe surprising application of Theo-
rem 13.21. An algebra structure on ℝ𝑛 is a ℝ-bilinear multiplication map

ℝ𝑛 × ℝ𝑛 → ℝ𝑛 , (𝑎, 𝑏) → 𝑎𝑏.

Note that it is not required that this multiplication has an identity element. Such a map is called a
division algebra structure if there are no zero-divisors, i.e., 𝑎𝑏 = 0 implies 𝑎 = 0 or 𝑏 = 0. It is
called commutative if 𝑎𝑏 = 𝑏𝑎 for all 𝑎, 𝑏 ∈ ℝ𝑛 . Examples of finite-dimensional commutative
division algebras over ℝ with an identity element are given by ℝ and ℂ = ℝ2 . However, there
is not much more:

Theorem 13.22 (Commutative algebra structures on ℝ𝑛 ) There is no commutative

division algebra structure on ℝ𝑛 for 𝑛 ≥ 3.

Proof:4 Suppose there is a commutative division algebra structure on ℝ𝑛 . Then we can

define a map by
𝑓 ∶ 𝕊𝑛−1 → 𝕊𝑛−1 , 𝑓 (𝑥) = 𝑥2 ∕|𝑥2 |.

This is well-defined, since 𝑥 ≠ 0 implies 𝑥2 ≠ 0 in a division algebra. The map 𝑓 is smooth,

since the multiplication map ℝ𝑛 × ℝ𝑛 → ℝ𝑛 is bilinear and hence smooth. Since 𝑓 satisfies
𝑓 (−𝑥) = 𝑓 (𝑥) for all 𝑥, 𝑓 induces a smooth map on the quotient

𝑓̄ ∶ ℝP𝑛−1 → 𝕊𝑛−1 .

∙ Claim: 𝑓̄ is injective.
4
The following proof follows Hatcher’s book [6] on Algebraic Topology.
280 13.5. ℝ𝑛 as a commutative division algebra
To show the claim we assume 𝑓 (𝑥) = 𝑓 (𝑦). This implies
√
|𝑥2 |
𝑥2 = 𝛼 2 𝑦2 for 𝛼 = > 0.
|𝑦2 |

Thus we have
𝑥2 − 𝛼 2 𝑦2 = 0.
Using commutativity and the fact that 𝛼 is a real number and multiplication is ℝ-bilinear we
see that this equation factors into

(𝑥 + 𝛼𝑦)(𝑥 − 𝛼𝑦) = 0.

Since we assume that the multiplication has no zero-divisors, this implies

𝑥 = ±𝛼𝑦.

Since |𝑥| = |𝑦| = 1 and 𝛼 is a real number, this implies |𝛼| = 1 and hence

𝑥 = ±𝑦.

Thus 𝑥 and 𝑦 determine the same point in ℝP𝑛−1 . This proves the claim that 𝑓̄ is injective.

Now we use the nice topological properties of ℝP𝑛−1 and 𝕊𝑛−1 . First we observe that 𝑓̄ is a
map between compact Hausdorff spaces. Since 𝑓̄ is injective, it is a homeomorphism onto its
image. Moreover, since ℝP𝑛−1 is compact, its image in 𝕊𝑛−1 is also compact and hence closed.
Since ℝP𝑛−1 and 𝕊𝑛−1 are smooth manifolds of the same dimension, invariance of domain5 ,
which we stated in Theorem 10.26, implies that the image is also open in 𝕊𝑛−1 . Since 𝕊𝑛−1 is
a connected, 𝑓̄ must be a homeomorphism. By Theorem 13.21, we know that for 𝑛 ≥ 3 there
cannot be a homeomorphism between ℝP𝑛−1 and 𝕊𝑛−1 . Hence our initial assumption must
have been wrong.

There is a fascinating much deeper story attached to this problem:

Remark 13.23 (Adams’ Theorem: Hopf invariant one) There is a vast generalisation
of the above result: Let ℝ𝑛 × ℝ𝑛 → ℝ𝑛 be a division algebra structure on ℝ𝑛 , without
assuming it is commutative. For 𝑛 = 4, there are the Hamiltonians, or Quaternions,
ℍ ≅ ℝ4 with a multiplication which is associative and almost as good as the one in ℂ
and ℝ, but it is not commutative. For 𝑛 = 8, there are the Octonions 𝕆 ≅ ℝ8 . The
multiplication is neither commutative nor associative. And that’s it! This is a deep
result. And surprisingly it is equivalent to a topological problem on the behavior of
tangent spaces on spheres. It was solved first by Adams. The prove goes way beyond
the methods of this class, unfortunately. However, it turned out to be not that hard using
complex 𝐾-theory. A key tool for the study is the Hopf invariant. We have seen its
mod 2 version in Section 11.3 and we will define the ℤ-valued version later via a series
of exercises in Section 15.4 using intersection theory and linking numbers.

5
Let 𝑈 ⊂ ℝ𝑛 be an open subset and ℎ ∶ 𝑈 → ℝ𝑛 an injective continuous map. Then ℎ(𝑈 ) is open in ℝ𝑛 .
Chapter 13. Intersection Theory modulo 2 281
13.6 Exercises and more examples

𝑓 𝑔
Exercise 13.1 Let 𝑋 ←←←→ ← 𝑍 be a sequence of smooth maps between manifolds
← 𝑌 ←←→
with 𝑋 compact, and let 𝑊 ⊂ 𝑍 be a submanifold. Assume that 𝑔 is transversal to 𝑊 ,
so that 𝑔 −1 (𝑊 ) is a submanifold in 𝑌 . Show that

𝐼2 (𝑓 , 𝑔 −1 (𝑊 )) = 𝐼2 (𝑔◦𝑓 , 𝑊 ).

In particular, verify that if one of these two intersection numbers is defined, then the other
one is defined as well.

Exercise 13.2 Let 𝑋 be a compact manifold without boundary.

(a) Assume dim 𝑋 ≥ 1: Show that if 𝑓 ∶ 𝑋 → 𝑌 is homotopic to a constant map,

then 𝐼2 (𝑓 , 𝑍) = 0 for all complementary dimensional closed submanifolds 𝑍 in
𝑌.
Hint: Show that if dim 𝑍 < dim 𝑌 , then 𝑓 is homotopic to a constant map 𝑋 →
{𝑦} for some 𝑦 ∉ 𝑍. You may assume that 𝑌 is connected.

(b) For dim 𝑋 = 0, show that this assertion is wrong.

Hint: If 𝑋 consists of just one point, for which 𝑍 will 𝐼2 (𝑓 , 𝑍) ≠ 0?

(c) Show that 𝕊1 is not simply-connected. Recall that we call a manifold 𝑋 simply-
connected if it is connected and if every map of the circle 𝕊1 into 𝑋 is homotopic
to a constant map.
Hint: Consider the identity map.

Exercise 13.3 (a) Show that intersection theory is trivial in contractible boundaryless
manifolds: if 𝑌 is boundaryless and contractible, i.e., its identity map is homotopic
to a constant map, and dim 𝑌 > 0, then 𝐼2 (𝑓 , 𝑍) = 0 for every smooth map
𝑓 ∶ 𝑋 → 𝑌 such that 𝑋 compact and 𝑍 closed, dim 𝑋 ≥ 1 and dim 𝑋 + dim 𝑍 =
dim 𝑌 . In particular, intersection theory is trivial in Euclidean space.

(b) Prove that no compact boundaryless manifold - other than the one-point space - is
contractible.
Hint: Consider the identity map.

Exercise 13.4(a) Let 𝑓 ∶ 𝑋 → 𝕊𝑘 be a smooth map with 𝑋 compact and

0 < dim 𝑋 < 𝑘. Show that, for all closed submanifolds 𝑍 ⊂ 𝕊𝑘 of dimension
complementary to 𝑋, 𝐼2 (𝑓 , 𝑍) = 0.
Hint: Use Sard’s Theorem 7.1 to show that there exists a 𝑝 ∉ 𝑓 (𝑋) ∩ 𝑍. Now use
a stereographic projection and the previous exercises.

(b) Show that 𝕊2 and the torus 𝕋 = 𝕊1 × 𝕊1 are not diffeomorphic.

282 13.6. Exercises and more examples

Exercise 13.5 Two compact manifolds 𝑋 and 𝑍 of the same dimension in 𝑌 are called
cobordant in 𝑌 if there exists a compact manifold with boundary 𝑊 ⊂ 𝑌 × [0, 1] such
that

𝜕𝑊 = 𝑋 × {0} ∪ 𝑍 × {1}.

The manifold 𝑊 is called a cobordism between 𝑋 and 𝑍. See Figure 13.4.

(a) Show that if we can deform 𝑋 into 𝑍, i.e., if there is a smooth homotopy from the
embedding 𝑖0 ∶ 𝑋 → 𝑌 of 𝑋 in 𝑌 to an embedding 𝑖1 ∶ 𝑋 → 𝑌 with 𝑖1 (𝑋) = 𝑍
such that each 𝑖𝑡 is an embedding, then 𝑋 and 𝑍 are cobordant.
Note that the standard image of a cobordism, a pair of pants as in Figure 13.4,
illustrates that the converse is false: 𝑋 and 𝑍 are cobordant, but we cannot deform
𝑋 into 𝑍, since 𝑋 has one connected component whereas 𝑍 has two.

(b) Show that if 𝑋 and 𝑍 are cobordant in 𝑌 , then for every compact submanifold 𝐶
in 𝑌 with dimension complementary to 𝑋 and 𝑍, i.e., dim 𝑋 + dim 𝐶 = dim 𝑍 +
dim 𝐶 = dim 𝑌 (where dim 𝑋 = dim 𝑍 because they are cobordant), we have

𝐼2 (𝐶, 𝑋) = 𝐼2 (𝐶, 𝑍).

Hint: Let 𝑓 be the restriction to 𝑊 of the projection map 𝑌 × [0, 1] → 𝑌 , and use
the Boundary Theorem 13.9.

Figure 13.4: A pair of pants is a cobordism between the upper boundary 𝑋 and the lower
boundary 𝑍. While 𝑋 is connected, 𝑍 has two connected components.

Exercise 13.6 Let 𝑍 ⊂ 𝑋 be a compact submanifold of 𝑋 with dim 𝑍 = 12 dim 𝑋.

Show that if there is a submersion 𝑔 ∶ 𝑈 → ℝ𝑘 defined on an open subset 𝑈 ⊂ 𝑋 such
Chapter 13. Intersection Theory modulo 2 283

that 𝑍 = 𝑔 −1 (0), then 𝐼2 (𝑍, 𝑍) = 0.

Hint: Use Theorem 12.14 and Theorem 12.15.

The following exercises require some computations for abstract manifolds. They shed light
on very interesting phenomenons and provide the proof for important results we mentioned in
the main text.

Exercise 13.7 Let 𝑋 be the open Möbius band given as the quotient of [−1, 1] × (−1, 1)
modulo the equivalence relation (−1, 𝑥) ∼ (1, −𝑥). Let 𝑍 denote the central circle given
by 𝑍 = {(𝑥, 0) ∶ 𝑥 ∈ [−1, 1]} ⊂ 𝑋.

(a) Show that 𝐼2 (𝑍, 𝑍) = 1. See Figure 13.3.

(b) Conclude using Exercise 13.6 that there cannot be a submersion 𝑔 ∶ 𝑋 → ℝ such
that 𝑍 = 𝑔 −1 (0).

Exercise 13.8 Let 𝑓 ∶ 𝕊1 → ℝP2 be the map defined by sending 𝑡 ∈ [0, 2𝜋] to
[cos(𝑡∕2) ∶ sin(𝑡∕2) ∶ 0].

(a) Show that 𝑓 is a smooth embedding, i.e., explain why it is smooth, injective, and
proper.

(b) Write 𝑍 for the image of 𝑓 in ℝP2 and consider it as a submanifold of ℝP2 . Prove
that 𝐼2 (𝑍, 𝑍) = 1.

(c) Conclude that 𝑓 is not homotopic to a constant map.

Aside: Recall that, since every continuous map on 𝕊2 is homotopic to a constant
map, this implies that there cannot be a homeomorphism between 𝕊2 and ℝP2 .
Aside: Note that the map [0, 2𝜋] → 𝕊2 defined by sending 𝑡 to the point
(cos(𝑡∕2), sin(𝑡∕2), 0) is not a loop on 𝕊2 . Hence the image has a boundary and
our intersection theory fails to work. And if we remove the boundary points and
define the map on (0, 2𝜋), then we loose compactness and even closedness, so in-
tersection theory will fail as well. However, if we look at the loop [0, 2𝜋] → 𝕊2
defined by sending 𝑡 to (cos(𝑡), sin(𝑡), 0) on 𝕊2 , then we see that the vertical loop
intersects the original loop twice. And we can convince us that any other loop
would intersect the original loop in an even number of points. Hence the mod
2-self intersection number of the loop is indeed zero.

(d) We write 2𝑓 for the map 2𝑓 ∶ 𝕊1 → ℝP2 defined by sending 𝑡 ∈ [0, 2𝜋] to [cos(𝑡) ∶
sin(𝑡) ∶ 0]. Show that 2𝑓 is homotopic to a constant map by constructing a concrete
homotopy. Do you see why the homotopy works for 2𝑓 but not for 𝑓 ?
Aside: The observations for 𝑓 and 2𝑓 are an indicator for the fact that the fun-
damental group of ℝP2 is cyclic of order two, i.e., 𝜋1 (ℝP2 ) ≅ ℤ∕2, where 𝑓
represents the non-trivial element and [2𝑓 ] = 0. Apart from the formula for 2𝑓
this is another explanation of our choice for the name 2𝑓 .
14. Orientation

Our next goal is to improve our definition of degree and intersection number and remedy the
defect that they only have even and odd values.

14.1 Towards integer-valued invariants

One of the reasons for this limitation was that a homotopy can move a non-transversal inter-
section into, for example, either an empty intersection or an intersection in two points. The
idea to handle this phenomenon is to take into account in which direction the intersection hap-
pens and to count intersection points with signs:

Figure 14.1: We can again move the two circles in the plane and either move them apart or
make them intersect transversely. In both cases we get a stable number of intersection points.
However, 0 is still not equal 2. But this time we keep track of how the points are oriented by
looking at how the relation of the directions of tangent vectors change.

The technical solution to implement this idea is to introduce orientations. We will see that,
unfortunately, not all manifolds are orientable. For those manifolds that orientable, however,
we will introduce and study integer-valued invariants in the next chapters. In order to get a first
idea why the claimed solution might be reasonable we look back at an important example first.

∙ A look back: Mod 2 Brouwer Degree of maps 𝕊1 → 𝕊1

In Section 11.2.3 we studied self-maps of the circle. Given a smooth map 𝑓 ∶ 𝕊1 → 𝕊1 , we

showed that it has a lift 𝑔 ∶ ℝ → ℝ such that 𝑔(𝑡 + 1) = 𝑔(𝑡) + 𝑞 for all 𝑡 for some fixed 𝑞 ∈ ℤ.
The integer 𝑞 is a number that is attached to 𝑓 . We then showed that deg2 (𝑓 ) = 𝑞 modulo 2.
How nice would it be if we had a notion of a degree for 𝑓 with integer values so that we could
write an equation like
deg(𝑓 ) = 𝑞.

284
Chapter 14. Orientation 285
If we knew in addition that the number deg(𝑓 ) is homotopy-invariant, i.e., it only depends on
the class of 𝑓 under the equivalence relation given by homotopy between maps, then we could
use the degree to distinguish between possibly infinitely many different homotopy classes of
maps 𝕊1 → 𝕊1 .

The way we came up with the formula deg2 (𝑓 ) = 𝑞 modulo 2 was to look at the graph of
𝑔 in Figure 11.5. Then we used the Intermediate Value Theorem and observed that the graph
crosses the horizontal lines 𝑔(𝑠) + 𝑘 for an integer 𝑘 ∈ {0, 1, … , 𝑞 − 1} an odd number of times
and lines 𝑔(𝑠) + 𝑘 for any other integer 𝑘 an even number of times. We would like to improve
this count by taking all the values with 𝑔(𝑡) = 𝑔(𝑠) + 𝑘 into account. But there might be many
more such values than just 𝑞. How can we fix this?

The idea is to remember how the graph of 𝑔 crosses the lines 𝑔(𝑠) + 𝑘: when 𝑡 increases
𝑔(𝑡) either increases while crossing 𝑔(𝑠) + 𝑘 or decreases. In other words, the derivative of 𝑔
is either positive or negative at 𝑡𝑘 with 𝑔(𝑡𝑘 ) = 𝑔(𝑠) + 𝑘. Note that we cannot have 𝑔 ′ (𝑡𝑘 ) = 0,
since 𝑦 is regular and hence det(𝑑𝑓𝑥 ) ≠ 0 and therefore also 𝑔 ′ (𝑡) ≠ 0 for all 𝑡 with 𝑝(𝑔(𝑡)) = 𝑦.

Now if 𝑔 ′ (𝑡𝑘 ) > 0, we count 𝑡𝑘 with value +1, if 𝑔 ′ (𝑡𝑘 ) < 0, we count 𝑡𝑘 with value −1.
See Figure 14.2. Counting all 𝑡 ∈ [𝑠, 𝑠 + 1) with 𝑔(𝑡) = 𝑔(𝑠) + 𝑘 for 𝑘 ∈ ℤ, this gives us the
desired value 𝑞. Hence if we can improve our definition of the degree such that we can use this
way of counting points on the graph of 𝑔, we would get deg(𝑓 ) = 𝑞.

Figure 14.2: The idea is to keep track of how the graph passes the horizontal lines: if 𝑔 increases
while passing the horizontal line we count the intersection as +1 and if 𝑔 decreases we count it
as −1.

To generalise this procedure, we need to describe 𝑔 ′ (𝑡𝑘 ) > 0 or 𝑔 ′ (𝑡𝑘 ) < 0 in more general
terms. In higher dimensions, we cannot just look at the sign of 𝑔 ′ (𝑡) or 𝑓 ′ (𝑡). A more fancy
way of describing the sign of the derivative of 𝑔 ′ (𝑡) is to say that the tangent space of the graph
of 𝑔 has a positive or a negative orientation. A line has exactly two orientations, depending
on if we walk in one direction or the other. It turns out that this can be generalised to higher
dimensional vector spaces and thereby to higher dimensional tangent spaces too. This is what
we are going to do next.
286 14.2. Orientation
14.2 Orientation

We begin with a look back at Linear Algebra.

14.2.1 Orientation on vector spaces

An orientation for a finite dimensional real vector space 𝑉 is an equivalence class of ordered
bases where the relation is defined as follows: the ordered basis (𝑣1 , … , 𝑣𝑛 ) has the same ori-
entation as the basis (𝑣′1 , … , 𝑣′𝑛 ) if the matrix 𝐴 with

𝑣′𝑖 = 𝐴𝑣𝑖 for all 𝑖 has determinant det(𝐴) > 0.

It has the opposite orientation if det(𝐴) < 0. The fact that this an equivalence relation
follows from the multiplicativity of the determinant function. Thus each finite dimensional
vector space has precisely two orientations, corresponding to the two equivalence classes of
ordered bases.

So an orientation of 𝑉 is a choice of an equivalence class of ordered bases. To make it

easier to talk about the choice of orientation, we attach to the chosen orientation a positive sign
and a negative sign to the other orientation. We say then that an ordered basis is positively
oriented (respectively negatively oriented) if its equivalence class belongs to the orientation
+1 (respectively −1). We often confuse an orientations with their corresponding signs +1 or
−1.

Remark 14.1 (Warning) The ordering of the basis elements is essential. Interchang-
ing the positions of two basis vectors changes the sign of the orientation since the the
determinant of the corresponding permutation matrix is negative.

In the case of the zero dimensional vector space it is convenient to define an orientation
as the symbol +1 or −1.

Example 14.2 (Standard orientation on ℝ𝑛 ) The vector space ℝ𝑛 has a standard

orientation given by the ordered basis (𝑒1 , … , 𝑒𝑛 ). We always assign +1 to the standard
orientation of ℝ𝑛 .

If 𝜑 ∶ 𝑉 → 𝑊 is an isomorphism of vector spaces, then 𝜑 either preserves or reverses

the orientation. For, given two ordered bases 𝛽 and 𝛽 ′ of 𝑉 belonging to the the same equiva-
lence class, the ordered bases 𝜑(𝛽) and 𝜑(𝛽 ′ ) either still belong to the same equivalence class of
ordered bases of 𝑊 or not. Whether 𝜑 preserves or reverses the orientation is determined by
its determinant. If det(𝜑) is positive, then 𝜑 preserves orientations, and if det(𝜑) is negative,
then 𝜑 reserves orientations.
Chapter 14. Orientation 287
14.2.2 Orientation on manifolds

Now we translate orientations from vector spaces to orientations for manifolds. The idea is to
orient each tangent space. However, this needs to be done in a compatible way.

Definition 14.3 (Orientation on manifolds) Let 𝑋 be a smooth manifold. An orien-

tation of 𝑋 is a smooth choice of orientations for all the tangent spaces 𝑇𝑥 𝑋 at each
point 𝑥 ∈ 𝑋. That means: around each point 𝑥 ∈ 𝑋 there exists a local parametrization
𝜙 ∶ 𝑈 → 𝑋 such that the isomorphism 𝑑𝜙𝑢 ∶ ℝ𝑘 → 𝑇𝜙(𝑢) 𝑋 preserves orientations
at each point 𝑢 of 𝑈 ⊆ ℍ𝑘 where the orientation on ℝ𝑘 is always assumed to be the
standard one.
We use the symbol 𝔬𝑋 to denote the data of an orientation on 𝑋. We will sometimes
write (𝑋, 𝔬𝑋 ) to express that 𝑋 is oriented by the orientation 𝔬𝑋 .

We now introduce an important distinction to our terminology and distinguish between the
possibility to make a choice or that such a choice actually has been made:

Definition 14.4 (Orientable and oriented manifolds) A manifold 𝑋 is called ori-

entable if a smooth choice of orientations of its tangent spaces exists. A manifold is
called oriented if it is orientable and a choice of orientation has been made.

Hence an oriented manifold is a pair consisting of a manifold together with a chosen ori-
entation.

Example 14.5 For zero-dimensional manifolds, orientations are very simple: to each
point 𝑥 ∈ 𝑋 we assign an orientation number +1 or −1.

Definition 14.6 (Maps and orientations) A smooth map 𝑓 ∶ 𝑋 → 𝑌 between oriented

manifolds is called orientation preserving if its derivative 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 )
preserves orientations at every point 𝑥 ∈ 𝑋.

Lemma 14.7 (Diffeomorphisms and orientation) Let 𝑋 and 𝑌 be oriented smooth

manifolds and let 𝑓 ∶ 𝑋 → 𝑌 be a diffeomorphism. Then 𝑓 is either preserves or
reverses orientation.

Proof: We show in Exercise 14.4 that if 𝑑𝑓𝑥 preserves orientation at one point 𝑥, then 𝑓
preserves orientation at every point.

Example 14.8 (Reflections and orientation) Let 𝑟𝑖 ∶ ℝ𝑛 → ℝ𝑛 be the reflection of

the 𝑖th coordinate, i.e.,

𝑟𝑖 (𝑥1 , … , 𝑥𝑛 ) = (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑛 ).

Then 𝑟𝑖 is an isomorphism which reverses the orientation. The composition 𝑟𝑖 ◦𝑟𝑗 of two
reflections is an isomorphism which preserves the orientation. This shows that −Idℝ𝑛 ,
288 14.2. Orientation
the negative of the identity of ℝ𝑛 , preserves the orientation if 𝑛 is even and −Idℝ𝑛 reverses
the orientation if 𝑛 is odd.
Now we can consider the reflection 𝑟𝑖 ∶ 𝕊𝑛 → 𝕊𝑛 on the 𝑛-sphere given by

𝑟𝑖 (𝑥1 , … , 𝑥𝑛+1 ) = (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑛+1 ).

Since the derivative of 𝑟𝑖 is a reflection on a vector space, we see that 𝑟𝑖 is a diffeomor-

phism which reverses the orientation.

The above observation on the composition of reflections and the negative of the identity
yield the following result:

Lemma 14.9 (Orientation and the antipodal map) The antipodal map 𝑎 ∶ 𝕊𝑛 → 𝕊𝑛 ,
𝑎(𝑥) = −𝑥, is a diffeomorphism which preserves the orientation if 𝑛 is odd and reverses
the orientation if 𝑛 is even.

We just learned that a manifold may or may not be orientable. To assign +1 or −1 to

the orientation of 𝑇𝑥 (𝑋) for every point is a locally constant function. If 𝑋 is orientable this
assignment is continuous. If 𝑋 is in addition connected, then this assignment must be constant.
Hence on every connected component of an orientable manifold, the orientation is constant +1
or −1.

Here is a rigorous proof of this fact:

Lemma 14.10 (Orientable manifolds have exactly two orientations) A connected,

orientable manifold with boundary admits exactly two orientations.

Proof: Assume we are given two orientations on 𝑋. In fact, we know that there are at least
two, since given one, we can reverse signs everywhere and get another orientation. We need to
show that there are not more than two.

We do this by showing that the set of points at which two orientations agree and the set
where they disagree are both open. Consequently, two orientations of a connected manifold
are either identical or opposite.

Since 𝑋 is orientable, we can choose local parametrizations 𝜙 ∶ 𝑈 → 𝑋 and 𝜙′ ∶ 𝑈 ′ → 𝑋

around 𝑥 ∈ 𝑋 with 𝜙(0) = 𝑥 = 𝜙′ (0) such that 𝑑𝜙𝑢 preserves the first orientation and 𝑑𝜙′𝑢′
preserves the second, for all 𝑢 ∈ 𝑈 and 𝑢′ ∈ 𝑈 ′ . After possibly shrinking we can assume
𝜙(𝑈 ) = 𝜙′ (𝑈 ′ ) (replace 𝑈 and 𝑈 ′ with 𝜙−1 (𝜙(𝑈 ) ∩ 𝜙′ (𝑈 ′ )) and 𝜙′ −1 (𝜙(𝑈 ) ∩ 𝜙′ (𝑈 ′ )), respec-
tively). If the two orientations of 𝑇𝑥 (𝑋) agree, then the map

𝑑(𝜙−1 ◦𝜙′ )0 ∶ ℝ𝑘 → ℝ𝑘

is an orientation preserving isomorphism. Thus the determinant of 𝑑(𝜙−1 ◦𝜙′ )0 is positive.

Hence the function

𝜑 ∶ 𝑈 ′ → ℝ, 𝑢′ → det(𝑑(𝜙−1 ◦𝜙′ )𝑢′ )

satisfies 𝜑(0) > 0.

Chapter 14. Orientation 289
Since the derivative depends continuously on and the determinant function is continu-
𝑢′
ous, 𝜑 is continuous. Hence, since 𝜑(0) > 0, there is an open neighborhood 𝑉 ′ around 0 in
𝑈 ′ on which 𝜑|𝑉 ′ > 0. But this implies that the orientations of 𝑇𝑥 (𝑋) induced by 𝜙 and 𝜙′ ,
respectively, agree for all 𝑥 in the open subset 𝜙′ (𝑉 ′ ). Since every point on 𝑋 has such an
open neighborhood, the set of points where the orientations agree is open.

If the orientations on 𝑇𝑥 (𝑋) induced by 𝜙 and 𝜙′ , respectively, do not agree, the same
argument shows that the set of points where the orientations do not agree is open.

Definition 14.11 (Reversed orientation) Hence if (𝑋, 𝔬𝑋 ) is an oriented manifold

𝑋, then we can talk about the manifold 𝑋 with the opposite orientation. We write
−𝔬𝑋 for the opposite orientation. We will often just write −𝑋 for the oriented manifold
(𝑋, −𝔬𝑋 ).

The product of oriented manifolds inherits an orientation as follows:

Definition 14.12 (Product orientation) Assume 𝑋 and 𝑌 are oriented and one of
them is without boundary. Then 𝑋 × 𝑌 is a manifold with boundary and inherits an
orientation in the following way: At a point (𝑥, 𝑦) ∈ 𝑋 × 𝑌 , let 𝛼 = (𝑣1 , … , 𝑣𝑘 ) and
𝛽 = (𝑤1 , … , 𝑤𝑚 ) be ordered bases of 𝑇𝑥 (𝑋) and 𝑇𝑦 (𝑌 ), respectively. We denote the
ordered basis ((𝑣1 , 0), … , (𝑣𝑘 , 0), (0, 𝑤1 ), … , (0, 𝑤𝑚 )) of 𝑇𝑥 (𝑋) × 𝑇𝑦 (𝑌 ) = 𝑇(𝑥,𝑦) (𝑋 × 𝑌 )
by (𝛼 × 0, 0 × 𝛽) . Now it comes handy that we related orientations of ordered bases to
signs. For we can define the orientation of 𝑇𝑥 (𝑋) × 𝑇𝑦 (𝑌 ) simply by determining a sign
by setting

sign (𝛼 × 0, 0 × 𝛽) ∶= sign (𝛼) ⋅ sign (𝛽).

Remark 14.13 (Classification) Not all manifolds are orientable. The most famous
example being the Möbius band. We will give a rigorous proof of this fact in Theo-
rem 14.16. Other examples of non-orientable manifolds include the Klein bottle and
the real projective plane ℝP2 . We will discuss the latter case in Theorem 14.15.
As a consequence, we see that the question whether a smooth manifold is orientable
or not helps classifying manifolds up to diffeomorphism: There is the class of orientable
manifolds, and the class of non-orientable manifolds.

Now we study how orientations behave with respect to smooth maps:

Theorem 14.14 (Pullback of an orientation) Let 𝑋 and 𝑌 be smooth manifolds with

or without boundary. Assume that 𝑌 = (𝑌 , 𝔬𝑌 ) is oriented and that there is a local
diffeomorphism 𝑓 ∶ 𝑋 → 𝑌 . Then there is a unique orientation on 𝑋 such that 𝑓
preserves orientations. We call the induced orientation on 𝑋 the pullback orientation
induced by 𝑓 and denote it by 𝑓 ∗ 𝔬𝑌 . Moreover, if 𝑔 ∶ 𝑌 → 𝑍 is another smooth map
and (𝑍, 𝔬𝑍 ) is an oriented manifold, then the pullback orientation on 𝑋 induced by 𝑔◦𝑓
is the same as the orientation on 𝑋 obtained by pulling back the orientation on 𝑍 via 𝑔
290 14.2. Orientation
to 𝑌 and then further to 𝑋, i.e.,

(𝑔◦𝑓 )∗ 𝔬𝑍 = 𝑓 ∗ (𝑔 ∗ 𝔬𝑍 ) as orientations on 𝑋.

In other words, the pullback of orientations behaves functorially.

Proof: Since 𝑓 is a local diffeomorphism, at every point 𝑥 ∈ 𝑋, there are open subsets
𝑈 ⊂ 𝑋 and 𝑉 ⊂ 𝑌 with 𝑥 ∈ 𝑈 and 𝑓 (𝑥) ∈ 𝑉 such that 𝑓|𝑈 ∶ 𝑈 → 𝑉 is a diffeomorphism. By
choosing 𝑈 and 𝑉 small enough, we can assume that there are local parametrizations 𝜙 ∶ 𝑊 →
𝑈 ⊂ 𝑋 and 𝜓 ∶ 𝑊 → 𝑉 ⊂ 𝑌 such that 𝑓 ◦𝜙 = 𝜓 with 𝑊 ⊂ ℝ𝑛 open. Since 𝑌 is oriented,
we can choose 𝜓 such that 𝑑𝜓𝑤 determines the orientation on 𝑇𝑦 𝑌 for all 𝑦 = 𝜓(𝑤) in 𝑉 .
Since 𝑑𝑓𝑥 is an isomorphism for all 𝑥 by assumption, we can define an orientation on 𝑇𝑥 𝑋 by
requiring det(𝑑𝑓𝑥 ) > 0. This defines an orientation on 𝑇𝑥 𝑋 for all 𝑥 ∈ 𝑈 . It remains to check
that this orientation on the tangent spaces for points in 𝑈 ⊂ 𝑋 does not depend on the choice
of 𝜙. So let 𝜙′ ∶ 𝑊 ′ → 𝑈 ′ ⊂ 𝑋 be another local parametrization of 𝑋 with 𝑈 ∩ 𝑈 ′ ≠ ∅. Then
we can find a local parametrization 𝜓 ′ ∶ 𝑊 ′ → 𝑉 ′ ⊂ 𝑌 of 𝑌 such that 𝑓|𝑈 ′ ∶ 𝑈 ′ → 𝑉 ′ is a
diffeomorphism. Since 𝑌 is oriented, the orientation of 𝑇𝑦 𝑌 determined by 𝑑𝜓𝑤 and by 𝑑𝜓𝑤′ ,
with 𝜓(𝑤) = 𝑦 = 𝜓(𝑤′ ) are the same for all 𝑦 in 𝑉 ∩ 𝑉 ′ . In other words, we have

det(𝑑(𝜓 −1 ◦𝜓 ′ )𝑤′ > 0 for all 𝑤′ ∈ 𝑊 ∩ 𝑊 ′ ⊂ ℝ𝑛 .

Since det(𝑑𝑓𝑥 ) > 0, this implies

det(𝑑(𝜙−1 ◦𝜙′ )𝑤′ > 0 for all 𝑤′ ∈ 𝑊 ∩ 𝑊 ′ ⊂ ℝ𝑛

as well. Hence the orientation of 𝑇𝑥 𝑋 via 𝑑𝜙𝑤 and 𝑑𝜙′𝑤′ is the same for all 𝑤, 𝑤′ ∈ 𝑊 ∩ 𝑊
with 𝜙(𝑤) = 𝑥 = 𝜙′ (𝑤′ ). This shows that we have defined an orientation on 𝑋. The facts
that 𝑓 preserves orientations, uniqueness and functoriality follow from the construction of the
induced orientation.

Here is an interesting application of Theorem 14.14:

Theorem 14.15 (Orientability of real projective space) Real projective space ℝP𝑛 is
orientable if and only if 𝑛 is odd.

Proof: Let 𝑞 ∶ 𝕊𝑛 → ℝP𝑛 , 𝑥 → [𝑥], be the canonical projection, and let 𝑎 ∶ 𝕊𝑛 → 𝕊𝑛

be the antipodal map. We have 𝑞(𝑥) = 𝑞(−𝑥) for all 𝑥 ∈ 𝕊𝑛 , i.e., we have 𝑞 = 𝑞◦𝑎. Both 𝑞
and 𝑎 are local diffeomorphisms. Hence, if ℝP𝑛 is oriented by an orientation 𝔬ℝP𝑛 , then the
orientation on 𝕊𝑛 induced by 𝑞 and by 𝑞◦𝑎 must be the same. Using the functoriality of the
pullback orientation, we need to have i.e., we must have

𝑞 ∗ 𝔬ℝP𝑛 = (𝑞◦𝑎)∗ 𝔬ℝP𝑛 = 𝑎∗ (𝑞 ∗ 𝔬ℝP𝑛 ).

This shows that ℝP𝑛 can be oriented only if 𝑎 preserves orientation. By Lemma 14.9 this is
the case if only if 𝑛 is odd. This proves that ℝP𝑛 is not orientable if 𝑛 is even.

Now we assume that 𝑛 is odd. We will now define an orientation of ℝP𝑛 . At every point
[𝑥] ∈ ℝP𝑛 , the fiber under 𝑞 consists of two antipodal points 𝑞 −1 ([𝑥]) = {𝑥, −𝑥} ⊂ 𝕊𝑛 . Given
a point 𝑥 ∈ 𝕊𝑛 , we can use stereographic projection from a point ≠ ±𝑥 on 𝕊𝑛 as one local
Chapter 14. Orientation 291
parametrization 𝜙 for both 𝑥 and −𝑥. The orientation of 𝑇𝑥 and 𝑇−𝑥 is then determined
𝕊𝑛 𝕊𝑛
such that the determinants of 𝑑𝜙𝜙−1 (𝑥) and 𝑑𝜙𝜙−1 (−𝑥) both have positive sign. Now we observe
that, for any choice of basis of 𝑇[𝑥] ℝP𝑛 , det(𝑑𝑞𝑥 ) and det(𝑑𝑞−𝑥 ) must have the same sign since
the products det(𝑑𝜙𝜙−1 (𝑥) ) ⋅ det(𝑑𝑞𝑥 ) and det(𝑑𝜙𝜙−1 (−𝑥) ) ⋅ det(𝑑𝑞−𝑥 ) must be either both positive
or both negative. This shows that we can define an orientation on 𝑇[𝑥] ℝP𝑛 by requiring that
det(𝑑𝑞𝑥 ) has positive sign and if we had used 𝑑𝑞−𝑥 instead of 𝑑𝑞𝑥 we would get the same
orientation. Since 𝕊𝑛 is oriented, this defines an orientation on ℝP𝑛 for 𝑛 odd.

We now look at an important example of a non-orientable manifold.

Theorem 14.16 The Möbius band is not orientable.

Proof: Let 𝑋 denote the Möbius band. We assume that 𝑋 is orientable and assume that we
have chosen an orientation. We will now show that this leads to a contradiction to previously
obtained results. Let 𝑍 ⊂ 𝑋 be the central circle. Then 𝑍 is orientable as a circle, and it
is a submanifold of codimension one. By Exercise 14.6, this implies that the normal bundle
𝑁(𝑍, 𝑋) is trivial. By Theorem 12.15, this implies that there is a submersion 𝑔 ∶ 𝑈 → ℝ1
defined on an open subset 𝑈 ⊂ 𝑋 such that 𝑍 = 𝑔 −1 (0). By Exercise 13.7, such a 𝑔 does not
exist. This shows that 𝑋 cannot be equipped with an orientation.

It is now a long and technical endeavour to check how orientations behave under the main
constructions and relate to the concepts we have developed so far. We will go through them one
by one.

14.3 Induced orientation on the boundary

Let 𝑋 be an oriented smooth manifold with boundary. Then the boundary submanifold 𝜕𝑋
inherits an orientation as follows:

At every point 𝑥 ∈ 𝜕𝑋, 𝑇𝑥 (𝜕𝑋) is a subspace of codimension one in 𝑇𝑥 (𝑋). Its orthogo-
nal complement in 𝑇𝑥 (𝑋), is a line which contains exactly two unit vectors: one is pointing
inward into 𝑇𝑥 (𝑋), the other one is pointing outward away from 𝑇𝑥 (𝑋).

Figure 14.3: We define an outward normal vector 𝑛𝑥 as the image 𝑑𝜙0 (−𝑒𝑘 ) of the outward
pointing standard basis vector −𝑒𝑘 .

This can be made precise by choosing a local parametrization 𝜙 ∶ 𝑈 → 𝑋 around 𝑥 with

292 14.3. Induced orientation on the boundary
𝑈 ⊂ ℍ𝑘 open and 𝜙(0) = 𝑥. The derivative 𝑑𝜙0 ∶ ℝ𝑘 → 𝑇𝑥 (𝑋) is by definition of 𝑇𝑥 (𝑋)
an isomorphism. In ℝ𝑘 , there are two unit vectors: 𝑒𝑘 = (0, … , 0, 1) pointing into ℍ𝑘 , and
−𝑒𝑘 = (0, … , 0, −1) pointing out of ℍ𝑘 . See Figure 14.3. Using the Gram–Schmidt process
we can orthonormalize the image of 𝑒𝑘 under 𝑑𝜙0 with respect to 𝑇𝑥 (𝜕𝑋) and get the inward
pointing unit normal vector. The orthonormalization with respect to 𝑇𝑥 (𝜕𝑋) of 𝑑𝜙0 (−𝑒𝑘 ) is the
outward pointing unit normal vector.1

We denote the outward pointing unit normal vector by 𝑛𝑥 . We checked in Exercise 10.4
that the construction of 𝑛𝑥 does not depend on the choice of 𝜙 and that the assignment 𝑥 → 𝑛𝑥
is a smooth map on 𝜕𝑋. See Figure 14.4.

Now we are ready to orient 𝑇𝑥 (𝜕𝑋):

Definition 14.17 (Boundary orientation) We define the sign of an ordered basis

(𝑣1 , … , 𝑣𝑘−1 ) in 𝑇𝑥 (𝜕𝑋) to be the sign of the ordered basis (𝑛𝑥 , 𝑣1 , … , 𝑣𝑘−1 ) in 𝑇𝑥 (𝑋),
i.e.,

sign (𝑣1 , … , 𝑣𝑘−1 ) in 𝑇𝑥 (𝜕𝑋) = sign (𝑛𝑥 , 𝑣1 , … , 𝑣𝑘−1 ) in 𝑇𝑥 (𝑋).

Since both the assignment 𝑥 → 𝑛𝑥 and the choice of sign for ordered bases on 𝑇𝑥 (𝑋) are
smooth, this defines an orientation on 𝜕𝑋 which is called the boundary orientation.

Figure 14.4: The outward pointing unit vector 𝑛𝑥 is the vector orthogonal to the tangent line at
𝑥 to 𝕊1 and pointing away from the origin.

Lemma 14.18 (Orientation of one-manifolds) Let us apply what we just learned to

the case of a one-manifold with boundary. The boundary 𝜕𝑋 is zero dimensional. The
orientation of the zero-dimensional vector space 𝑇𝑥 (𝜕𝑋) equals the sign of the basis of
𝑇𝑥 (𝑋) consisting of the outward-pointing unit vector 𝑛𝑥 .

1
Note that the inner product on 𝑇𝑥 (𝑋) is induced by the standard inner product on ℝ𝑁 , where 𝑋 ⊂ ℝ𝑁 and
hence 𝑇𝑥 (𝑋) ⊂ ℝ𝑁 .
Chapter 14. Orientation 293

Example 14.19 (Unit interval) Let us look at the compact interval 𝑋 = [0, 1] with its
standard orientation inherited from being a subset in ℝ. Note that local parametrizations
of [0, 1] are given by

𝜙 ∶ [0, 1) → [0, 1], 𝑥 → 𝑥

around 0 ∈ [0, 1] and

𝜓 ∶ [0, 1) → [0, 1], 𝑥 → 1 − 𝑥

around 1 ∈ [0, 1].

Hence, at 𝑥 = 1, the outward-pointing normal vector is 1 ∈ ℝ = 𝑇𝑥 (𝑋). The
basis consisting of this vector is positively oriented. At 𝑥 = 0 the outward-pointing
normal vector is the negatively oriented −1 ∈ ℝ = 𝑇0 (𝑋). Thus the orientation of
𝑇1 (𝜕𝑋) is +1, and the orientation of 𝑇0 (𝜕𝑋) is −1. Reversing the orientation on [0, 1]
simply reverses the orientations at each boundary point. Thus the sum of both orientation
numbers at the boundary points of [0, 1] is always zero.

Since, by Theorem 10.17, every compact one-manifold with boundary is diffeomorphic

is the disjoint union of copies of [0, 1], we conclude:

Lemma 14.20 (Boundary orientation of one-manifolds) The sum of the orientation

numbers at the boundary points of any compact oriented one-dimensional manifold
with boundary is zero. In particular, the boundary points of a smooth path 𝛾 on an
oriented manifold 𝑋, i.e., a smooth map 𝛾 ∶ [0, 1] → 𝑋, must have opposite orientation
signs.

Remark: This is an important observation which makes it possible to define homotopy

invariant degree of a smooth map with values in ℤ.

14.4 Oriented Homotopy

As an application of product and boundary orientations, we would like to orient the product
[0, 1] × 𝑋 for a boundaryless smooth oriented manifold 𝑋 which is the domain of all homo-
topies on 𝑋. This will be crucial for the homotopy invariance of intersection numbers in the
next section.

We just learned that products and boundaries inherit orientations. For each 𝑡 ∈ [0, 1], the
slice 𝑋𝑡 ∶= {𝑡} × 𝑋 is diffeomorphic to 𝑋, and the orientation on 𝑋𝑡 should be such that the
diffeomorphism

𝑋 → 𝑋𝑡 , 𝑥 → (𝑡, 𝑥) preserves orientations.

For future applications, we are particularly interested in the orientation of the boundary

𝜕([0, 1] × 𝑋) = {0} × 𝑋 ∪ {1} × 𝑋.

294 14.5. Orientation of transverse preimage
So let us try to understand the induced orientation on the boundary:

We start with 𝑋1 : We see from the local parametrization 𝜓 above that along 𝑋1 the outward-
pointing normal vector is

𝑛(1,𝑥) = (1, 0) = (1, 0, … , 0) ∈ 𝑇1 ([0, 1]) × 𝑇𝑥 (𝑋),

If 𝛽 = (𝑣1 , … , 𝑣𝑘 ) is an ordered basis of 𝑇𝑥 (𝑋), then 0 × 𝛽 = ((0, 𝑣1 ), … , (0, 𝑣𝑘 )) is an

ordered basis of 𝑇𝑥 (𝑋1 ). By definition of the boundary orientation, (𝑛(1,0) , (0×𝛽)) is positively
oriented if and only if 𝛽 is positively oriented. In terms of signs:

sign (𝑛(1,0) , (0 × 𝛽)) = sign (𝛽).

If we calculate the orientation induced from the product structure, then we get

sign ((1, 0), (0 × 𝛽)) = (+1) ⋅ sign (𝛽) = sign (𝛽).

We learn from these two equations, that the boundary orientation of 𝑋1 is just the orien-
tation of 𝑋 as a copy in the product [0, 1] × 𝑋.

This sounds obvious, but pay attention:

We see from the local parametrization 𝜙 that along 𝑋0 the outward-pointing normal vector
is
𝑛(0,𝑥) = (−1, 0) = (−1, 0, … , 0) ∈ 𝑇0 ([0, 1]) × 𝑇𝑥 (𝑋).
Hence the orientation on 𝑇0 ([0, 1]) is opposite to the standard orientation of ℝ. Hence the
formula for product orientations yields

sign ((−1, 0), 0 × 𝛽)) = sign (−1) ⋅ sign (𝛽) = −sign (𝛽).

Thus the boundary orientation on 𝑋0 is the reverse of its orientation as a copy of 𝑋 in the
product [0, 1] × 𝑋.

Thus the orientation on the boundary is

𝜕([0, 1] × 𝑋) = 𝑋1 ∪ (−𝑋0 ).

We will also express this fact by using the notation

𝜕([0, 1] × 𝑋) = 𝑋1 − 𝑋0 .

14.5 Orientation of transverse preimage

Our next goal is to orient preimages. In order to do so, we will have to look at direct sums (not
just products) of vector spaces, and we need to orient those guys.
Chapter 14. Orientation 295

Figure 14.5: The vectors 𝑛𝑥0 and 𝑛𝑥1 point in opposite directions. This provides the two bound-
ary circles with opposite orientations. This fact will turn out to be crucial for the definition of
intersection numbers later.

Definition 14.21 (Orientation on a direct sum of vector spaces) Suppose that 𝑉 =

𝑉1 ⊕ 𝑉2 is a direct sum of vector spaces. Then orientations on any two of these vector
spaces automatically induces a direct sum orientation on the third, as follows:a
Choose ordered bases 𝛽1 of 𝑉1 and 𝛽2 of 𝑉2 . Let 𝛽 = (𝛽1 , 𝛽2 ) be the combined ordered
basis of 𝑉 (in this order!). For orientations or signs to be compatible with the structure
as a direct sum, we require the formula

sign (𝛽) = sign (𝛽1 ) ⋅ sign (𝛽2 ).

It follows immediately from the way matrices on direct sums are put together that
this formula determines an orientation on the third space if two orientations are given.
Note again that the order of the summands 𝑉1 and 𝑉2 is crucial.
a
Note that this not only means, orientations on 𝑉1 and 𝑉2 determine an orientation on 𝑉 , but also ori-
entations on 𝑉 and, say, 𝑉2 determine an orientation on 𝑉1 .

Now let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map and 𝑍 ⊂ 𝑌 be a submanifold with 𝑓 − ⋔ 𝑍 and

−
𝜕𝑓 ⋔ 𝑍. We assume that 𝑋, 𝑌 , and 𝑍 are all oriented and 𝑌 and 𝑍 are without boundary. We
would like to define a preimage orientation on the manifold with boundary 𝑆 = 𝑓 −1 (𝑍):

If 𝑓 (𝑥) = 𝑧 ∈ 𝑍, then

𝑇𝑥 (𝑆) = (𝑑𝑓𝑥 )−1 (𝑇𝑧 (𝑍)) ⊂ 𝑇𝑥 (𝑋).

Let 𝑁𝑥 (𝑆; 𝑋) be the orthogonal complement to 𝑇𝑥 (𝑆) in 𝑇𝑥 (𝑋). By definition, we have a

direct sum decomposition

𝑁𝑥 (𝑆; 𝑋) ⊕ 𝑇𝑥 (𝑆) = 𝑇𝑥 (𝑋).

Hence, by Definition 14.21 on the orientation of a direct sum, we only need to choose an ori-
296 14.6. Example: Fibers of the Hopf fibration
entation on 𝑁𝑥 (𝑆; 𝑋) to obtain a direct sum orientation on 𝑇𝑥 (𝑆). Since 𝑓 −
⋔ 𝑍, we have

𝑇𝑧 (𝑌 ) = 𝑑𝑓𝑥 (𝑇𝑥 (𝑋)) + 𝑇𝑧 (𝑍)

= 𝑑𝑓𝑥 (𝑁𝑥 (𝑆; 𝑋) ⊕ 𝑇𝑥 (𝑆)) + 𝑇𝑧 (𝑍)
= 𝑑𝑓𝑥 (𝑁𝑥 (𝑆; 𝑋)) ⊕ 𝑇𝑧 (𝑍)

where we use 𝑑𝑓𝑥 (𝑇𝑥 (𝑆)) = 𝑇𝑧 (𝑍) for the last step. Thus the orientations on 𝑍 and 𝑌 induce
a direct sum orientation on 𝑑𝑓𝑥 (𝑁𝑥 (𝑆; 𝑋)). It remains to show that this also induces an
orientation on 𝑁𝑥 (𝑆; 𝑋): We have

{0} ⊂ 𝑇𝑧 (𝑍) ⇒ Ker (𝑑𝑓𝑥 ) ⊂ (𝑑𝑓𝑥 )−1 (𝑇𝑧 (𝑍)) = 𝑇𝑥 (𝑆),

and hence the restriction of 𝑑𝑓𝑥 to 𝑁𝑥 (𝑆; 𝑋) is in fact an isomorphism onto its image. There-
fore the induced orientation on 𝑑𝑓𝑥 (𝑁𝑥 (𝑆; 𝑋)) defines an orientation on 𝑁𝑥 (𝑆; 𝑋) via the iso-
morphism 𝑑𝑓𝑥 . Since the orientations on 𝑋, 𝑌 and 𝑍 vary smoothly and 𝑑𝑓𝑥 also depends
smoothly on 𝑥, the induced orientation on 𝑇𝑥 (𝑆) varies smoothly with 𝑥.

We put this into a formula in the following definition:

Definition 14.22 (Preimage orientation) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map and 𝑍 ⊂ 𝑌

be a submanifold with 𝑓 − ⋔ 𝑍 and 𝜕𝑓 − ⋔ 𝑍. Let 𝑋, 𝑌 , and 𝑍 be oriented and 𝑌 and 𝑍
be without boundary. We choose and denote by 𝛽𝑉 an ordered basis for the vector space
𝑉 which runs through the spaces 𝑇𝑥 𝑓 −1 (𝑍), 𝑇𝑥 𝑋, 𝑇𝑓 (𝑥) 𝑍, 𝑇𝑓 (𝑥) 𝑌 , and 𝑁𝑥 (𝑓 −1 (𝑍); 𝑋).
We define the sign of 𝛽𝑁𝑥 (𝑓 −1 (𝑍);𝑋) , and thereby the orientation of 𝑁𝑥 (𝑓 −1 (𝑍); 𝑋), such
that
( ( )) ( ) ( )
sign 𝑑𝑓𝑥 𝛽𝑁𝑥 (𝑓 −1 (𝑍);𝑋) ⋅ sign 𝛽𝑇𝑓 (𝑥) 𝑍) = sign 𝛽𝑇𝑓 (𝑥) 𝑌 .

Then we define the sign of 𝛽𝑇𝑥 𝑓 −1 (𝑍) and hence the orientation of 𝑓 −1 (𝑍) such that
( ) ( ) ( )
sign 𝛽𝑁𝑥 (𝑓 −1 (𝑍);𝑋) ⋅ sign 𝛽𝑇𝑥 𝑓 −1 (𝑍) = sign 𝛽𝑇𝑥 𝑋 .

Remark 14.23 (Orthogonality is not important) Note that we did not really use
that 𝑁𝑥 (𝑆; 𝑋) is orthogonal to 𝑇𝑥 (𝑆). All we needed was a direct sum decomposition
𝐻 ⊕ 𝑇𝑥 (𝑆) = 𝑇𝑥 (𝑋) with a space 𝐻 with an orientation induced by the orientation of
𝑋. We will exploit this fact in the proof below.

14.6 Example: Fibers of the Hopf fibration

Recall the Hopf fibration 𝜋 that we have seen previously: We consider 𝕊3 as a subset of ℂ2 ,
i.e., 𝕊3 = {(𝑧0 , 𝑧1 ) ∈ ℂ2 ∶ |𝑧0 |2 + |𝑧1 |2 = 1}, and 𝕊2 as a subset of ℂ × ℝ, i.e., 𝕊2 = {(𝑧, 𝑥) ∈
ℂ × ℝ ∶ |𝑧|2 + 𝑥2 = 1}. Then the Hopf fibration 𝜋 is the map 𝕊3 → 𝕊2 given by
( )
𝜋(𝑧0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2 .
Chapter 14. Orientation 297
Consider the point 𝑎 = (0, 0, 1) on 𝕊2 ⊂ ℝ3 ≅ ℂ × ℝ. In a previous exercise we have
determined the fiber 𝜋 −1 (𝑎). Now we would like to compute its orientation as a preimage
under 𝜋.

Recall that the fiber over 𝑎 is

𝜋 −1 (𝑎) = {(𝑧0 , 0) ∈ 𝕊3 ⊂ ℂ2 ∶ |𝑧0 |2 = 1}.

Let 𝑞 = (𝑥0 , 𝑦0 , 0, 0) ∈ 𝜋 −1 (𝑎) be a point in the fiber over 𝑎. The tangent space 𝑇𝑞 𝕊3 is the
vector space

𝑇𝑞 𝕊3 = {𝐮 ∈ ℝ4 ∶ 𝐮 ⟂ 𝑞}
⎧ ⎛−𝑦0 ⎞ ⎛0⎞ ⎛0⎞⎫
⎪ ⟂ ⎜ 𝑥0 ⎟ 4 ⎜0⎟ 4 ⎜0⎟⎪
= span ⎨𝑞 = ⎜ ⎟ , 𝐞3 = ⎜1⎟ , 𝐞4 = ⎜0⎟⎬ .
⎪ ⎜ 0 ⎟ ⎜ ⎟ ⎜ ⎟⎪
⎩ ⎝ 0 ⎠ ⎝0⎠ ⎝1⎠⎭

The orientation of 𝑇𝑞 𝕊3 as a boundary of the unit ball is such that the outward pointing
vector 𝑞 together with the basis vectors of 𝑇𝑞 𝕊3 form a positively oriented basis of ℝ4 . The
determinant of the matrix expressing the basis (𝑞, 𝑞 ⟂ , 𝐞43 , 𝐞44 ) in the standard basis of ℝ4 equals
𝑥20 + 𝑦20 = 1 > 0. In particular, it is positive and the basis (𝑞 ⟂ , 𝐞43 , 𝐞44 ) is a positively oriented
basis of 𝑇𝑞 𝕊3 .

The tangent space 𝑇𝑞 𝜋 −1 (𝑎) equals the kernel of 𝑑 𝜋̃𝑞 , where we consider the map 𝜋̃ ∶ ℝ4 ≅
ℂ2 → ℂ × ℝ ≅ ℝ3 using the same formula as for 𝜋, i.e., 𝜋 = 𝜋̃|𝕊3 . In an exercise we computed
this map as represented by the matrix

⎛ 0 0 𝑥0 𝑦0 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ ⎜ 0 0 𝑦0 −𝑥0 ⎟ .
⎜ ⎟
⎝𝑥0 𝑦0 0 0 ⎠

The kernel of this map is the span of the vector 𝑞 ⟂ that we have just seen above. The normal
space 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ) ⊂ 𝑇𝑞 𝕊3 of vectors which are orthogonal to 𝑇𝑞 𝜋 −1 (𝑎) is the span of (𝐞43 , 𝐞44 ).
The map 𝑑 𝜋̃𝑞 sends 𝐞3 and 𝐞4 to, respectively,

⎛𝑥0 ⎞ ⎛ 𝑦0 ⎞
⎜ 𝑦 ⎟ ⎜−𝑥 ⎟
𝑑 𝜋̃𝑞 (𝐞43 ) = 2 ⎜ 0 ⎟ , 𝑑 𝜋̃𝑞 (𝐞44 ) = 2 ⎜ 0 ⎟ .
⎜0⎟ ⎜ 0 ⎟
⎝0⎠ ⎝ 0 ⎠

These two vectors form a basis (𝑑 𝜋̃𝑞 (𝐞33 ), 𝑑 𝜋̃𝑞 (𝐞44 )) of 𝑇𝑎 𝕊2 .

We need to check the orientation of this basis: The tangent space 𝑇𝑎 𝕊2 has a basis (𝐞31 , 𝐞32 )
as a subspace in ℝ3 . This basis is positively oriented since, together with the outward pointing
vector 𝑎 = 𝐞33 , the basis (𝐞33 , 𝐞31 , 𝐞32 ) is a positively oriented basis of ℝ3 .2 The matrix 𝐴 which
expresses (𝑑 𝜋̃𝑞 (𝐞33 ), 𝑑 𝜋̃𝑞 (𝐞44 )) in terms of the basis (𝐞31 , 𝐞32 ) is given by
( )
𝑥0 𝑦0
𝐴=2⋅ .
𝑦0 −𝑥0
2
We make two permutations which lead to multiplying with (−1)2 = +1.
298 14.7. Orientation on boundary of preimage
We see that det 𝐴 = 4(−𝑥20 − 𝑦20 ) = −4 < 0 is negative. Hence the basis (𝑑 𝜋̃𝑞 (𝐞33 ), 𝑑 𝜋̃𝑞 (𝐞44 )) is a
negatively oriented basis of 𝑇𝑎 𝕊2 . This defines an orientation on the normal space 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 )
by declaring the orientation of the basis (𝐞43 , 𝐞44 ) to be negative.

Finally, the orientation of 𝑇𝑞 𝜋 −1 (𝑎) is such that the direct sum

𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑎) = 𝑇𝑞 𝕊3

induces the given orientation on 𝑇𝑞 𝕊3 .

We check this by looking at the basis (𝐞43 , 𝐞44 , 𝑞 ⟂ ) of 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑎). As a basis
of 𝑇𝑞 𝕊3 this basis is positively oriented, since it arises by two permutations from the positively
oriented basis (𝑞 ⟂ , 𝐞43 , 𝐞44 ). Since the sign of (𝐞43 , 𝐞44 ) is negative as a basis of 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ), we
need that 𝑞 ⟂ also has negative sign. Hence the vector 𝑞 ⟂ provides a negatively oriented basis
of 𝑇𝑞 𝜋 −1 (𝑎).

Comparing this orientation with the standard orientation of 𝕊1 ⊂ ℂ ⊂ ℂ2 , we see that

𝜋 −1 (𝑎)has the opposite orientation.

We will compute the orientation of the fiber of 𝜋 at another point in the exercises. You
should try to solve this exercise even though it is a bit involved.

14.7 Orientation on boundary of preimage

Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map with 𝑓 −⋔ 𝑍 and 𝜕𝑓 −

⋔ 𝑍, where 𝑋, 𝑌 , and 𝑍 are all oriented,
𝑌 and 𝑍 are boundaryless, and 𝑋 has a boundary.

Then the manifold 𝜕𝑓 −1 (𝑍) acquires two orientations:

∙ an orientation as the boundary of the manifold 𝑓 −1 (𝑍), and

∙ an orientation as the preimage of 𝑍 under the map 𝜕𝑓 ∶ 𝜕𝑋 → 𝑌 .

It turns out that there is a formula that relates these two orientations:

Theorem 14.24 (Orientation on boundary of preimage) Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth

map with 𝑓 − ⋔ 𝑍 and 𝜕𝑓 −
⋔ 𝑍, where 𝑋, 𝑌 , and 𝑍 are all oriented, 𝑌 and 𝑍 are without
boundary, while 𝑋 has a boundary. Then the orientation of the boundary of 𝑓 −1 (𝑍)
satisfies the formula:

𝜕(𝑓 −1 (𝑍)) = (−1)codim 𝑍 (𝜕𝑓 )−1 (𝑍).

This means the orientations of 𝜕𝑓 −1 (𝑍), induced by being a boundary or by being a

preimage, are the same if codim 𝑍 is even, and opposite if codim 𝑍 is odd.

Proof: Denote 𝑓 −1 (𝑍) again by 𝑆. Let 𝐻 be a subspace of 𝑇𝑥 (𝜕𝑋) complementary to

Chapter 14. Orientation 299
𝑇𝑥 (𝜕𝑆), i.e.,

𝐻 ⊕ 𝑇𝑥 (𝜕𝑆) = 𝑇𝑥 (𝜕𝑋).

Note that 𝐻 is also complementary to 𝑇𝑥 (𝑆) in 𝑇𝑥 (𝑋), i.e.,

𝐻 ⊕ 𝑇𝑥 (𝑆) = 𝑇𝑥 (𝑋).

For we have

𝐻 ∩ 𝑇𝑥 (𝑆) = {0} and 𝑇𝑥 (𝑆) ∩ 𝑇𝑥 (𝜕𝑋) = 𝑇𝑥 (𝜕𝑆),

and

dim 𝐻 = dim 𝑇𝑥 (𝜕𝑋) − dim 𝑇𝑥 (𝜕𝑆) = dim 𝑇𝑥 (𝑋) − dim 𝑇𝑥 (𝑆).

Hence we may use 𝐻 to define the direct sum orientation of both 𝑆 and 𝜕𝑆 at 𝑥.

Since 𝐻 ⊂ 𝑇𝑥 (𝜕𝑋) ⊂ 𝑇𝑥 (𝑋), the maps 𝑑𝑓𝑥 and 𝑑(𝜕𝑓 )𝑥 agree on 𝐻, i.e.,

𝑑𝑓𝑥 (𝐻) = 𝑑(𝜕𝑓 )𝑥 (𝐻).

As in the case of 𝑁𝑥 (𝑆; 𝑋), since

{0} ⊂ 𝑇𝑧 (𝑍) ⇒ Ker (𝑑𝑓𝑥 ) ⊂ 𝑓 −1 (𝑇𝑧 (𝑍)) = 𝑇𝑥 (𝑆),

the intersection Ker (𝑑𝑓𝑥 ) ∩ 𝐻 is {0}. Hence the restrictions of 𝑑𝑓𝑥 and 𝑑(𝜕𝑓 )𝑥 to 𝐻 are
isomorphisms onto their common image.

Thus 𝑓 − ⋔ 𝑍 and 𝜕𝑓 − ⋔ 𝑍 imply that we have two direct sum decompositions 𝑑𝑓𝑥 (𝐻) ⊕
𝑇𝑧 (𝑍) = 𝑇𝑧 (𝑌 ) = 𝑑(𝜕𝑓 )𝑥 (𝐻) ⊕ 𝑇𝑧 (𝑍), and the two orients of 𝐻 via these direct sums agree.

To conclude, we obtained that 𝐻 has a well-defined orientation. Hence we can use this
unique orientation on 𝐻 to orient

𝑆 via 𝐻 ⊕ 𝑇𝑥 (𝑆) = 𝑇𝑥 (𝑋) and 𝜕𝑆 via 𝐻 ⊕ 𝑇𝑥 (𝜕𝑆) = 𝑇𝑥 (𝜕𝑋).

It remains to check how this orientation of 𝑇𝑥 (𝜕𝑆) relates to the orientation of the boundary
induced from the orientation of 𝑇𝑥 (𝑆). Let 𝑛𝑥 be the outward unit vector to 𝜕𝑆 in 𝑇𝑥 (𝑆), and
let ℝ ⋅ 𝑛𝑥 represent the one-dimensional subspace spanned by 𝑛𝑥 . We orient this space by
assigning the sign +1 to the basis (𝑛𝑥 ). Even though 𝑛𝑥 need not be perpendicular to all of
𝑇𝑥 (𝜕𝑋), it suffices to know that 𝑛𝑥 lies in the half-space pointing away from 𝑇𝑥 (𝑋) to know
that the orientations of ℝ ⋅ 𝑛𝑥 , 𝑇𝑥 (𝜕𝑋) and 𝑇𝑥 (𝑋) are related by the direct sum

𝑇𝑥 (𝑋) = ℝ ⋅ 𝑛𝑥 ⊕ 𝑇𝑥 (𝜕𝑋).

Now we use that 𝐻 is complementary to both 𝑇𝑥 (𝑆) in 𝑇𝑥 (𝑋) and 𝑇𝑥 (𝜕𝑆) in 𝑇𝑥 (𝜕𝑋) and
plugg this into the above direct sum to get

𝐻 ⊕ 𝑇𝑥 (𝑆) = ℝ ⋅ 𝑛𝑥 ⊕ 𝐻 ⊕ 𝑇𝑥 (𝜕𝑆).

This equation is already almost what we need, since we would like to compare the orientations
𝑇𝑥 (𝑆) and ℝ ⋅ 𝑛𝑥 ⊕ 𝑇𝑥 (𝜕𝑆). For doing so, we need to move ℝ ⋅ 𝑛𝑥 passed 𝐻. If dim 𝐻 = 𝑚, 𝐻
300 14.8. Example: Simply-connected manifolds are orientable
has 𝑚 basis vectors (𝑤1 , … , 𝑤𝑚 ). Remembering the rule for orienting direct sums, this means
we have to apply exactly 𝑚 transpositions to the ordered set

(𝑛𝑥 , 𝑤1 , … , 𝑤𝑚 ) to get to (𝑤1 , … , 𝑤𝑚 , 𝑛𝑥 ).

This results in 𝑚 shifts of signs. Hence we get

𝐻 ⊕ 𝑇𝑥 (𝑆) = (−1)codim 𝑍 𝐻 ⊕ ℝ ⋅ 𝑛𝑥 ⊕ 𝑇𝑥 (𝜕𝑆).

Since 𝐻 appears on both sides as the first summand, we can disregard it for the computation
and get that if 𝜕𝑆 is oriented as a preimage under 𝜕𝑓 , then its orientation relates to the one of
𝑇𝑥 (𝑆) by

𝑇𝑥 (𝑆) = (−1)codim 𝑍 ℝ ⋅ 𝑛𝑥 ⊕ 𝑇𝑥 (𝜕𝑆).

Now, if 𝜕𝑆 is oriented as a boundary, then we have

𝑇𝑥 (𝑆) = ℝ ⋅ 𝑛𝑥 ⊕ 𝑇𝑥 (𝜕𝑆).

Thus

sign (𝜕𝑆) as a boundary = (−1)codim 𝑍 ⋅ sign (𝜕𝑆) as a preimage.

14.8 Example: Simply-connected manifolds are orientable

The following theorem shows that a lot of manifolds are orientable. Recall that a manifold 𝑋
is called simply-connected if it is connected and every smooth map 𝕊1 → 𝑋 is homotopic to a
constant map.

Theorem 14.25 (Simply-connected implies orientable) Every simply-connected man-

ifold is orientable.

Proof: We start by picking any point 𝑥 ∈ 𝑋, and choose an orientation for the tangent
space 𝑇𝑥 (𝑋). Since 𝑇𝑥 (𝑋) is a vector space, this is always possible. Now let 𝑦 ∈ 𝑋 be any
other point in 𝑋. Since 𝑋 is simply-connected, 𝑋 is connected. By a previous exercise, since
𝑋 is a smooth manifold, 𝑋 is therefore even path-connected. Hence there is a smooth map
𝛾 ∶ [0, 1] → 𝑋 with 𝛾(0) = 𝑥 and 𝛾(1) = 𝑦. For every point in 𝑧 ∈ 𝛾([0, 1]) we choose a local
parametrization 𝜙𝑧 ∶ 𝑉𝑧 → 𝑈𝑧 around 𝑧. By shrinking 𝑉𝑧 if necessary, we can assume that
each 𝑉𝑧 is an open ball in ℝ𝑘 .

The sets 𝑈𝑧 ∩ 𝛾([0, 1]) is open in 𝛾([0, 1]), and the collection of {𝑈𝑧 ∩ 𝛾([0, 1])} for all
𝑧 ∈ 𝛾([0, 1]) is an open covering of 𝛾([0, 1]). Since [0, 1] is compact and 𝛾 continuous, the
image 𝛾([0, 1]) is compact. Hence finitely many of the 𝑈𝑧 suffice to cover 𝛾([0, 1]). We label
these open sets 𝑈1 , … , 𝑈𝑚 and order them such that 𝑈𝑖 ∩ 𝑈𝑖+1 ≠ ∅ and 𝑥 ∈ 𝑈1 , 𝑦 ∈ 𝑈𝑚 .

For 𝑈1 , we choose the orientation which is compatible with the chosen orientation of
𝑇𝑥 (𝑋). This means: Let 𝜙1 ∶ 𝑈1 → 𝑋 be the associated local parametrization with 𝜙1 (0) = 𝑥.
Chapter 14. Orientation 301
∙ If 𝑑(𝜙1 )0 ∶ ℝ𝑘 → 𝑇𝑥 (𝑋) is orientation preserving, we orient the vector space 𝑇𝑎 (𝑈1 )
such that 𝑑(𝜙1 )𝜙−1 (𝑧) ∶ ℝ𝑘 → 𝑇𝑎 (𝑋) is orientation preserving for all 𝑎 ∈ 𝑈1 .
1

∙ If 𝑑(𝜙1 )0 → ℝ𝑘 → 𝑇𝑥 (𝑋) reverses orientation, we first replace 𝜙1 with 𝜙̃ 1 ∶ 𝑉1 →

𝑋, 𝑣 → 𝜙1 (−𝑣). This new map 𝜙̃ 1 is also a local parametrization of 𝑋 with domain 𝑉1 ,
since 𝑉1 is an open ball in ℝ𝑘 and 𝜙1 is therefore symmetric with respect to the origin.

Hence after possibly replacing 𝜙1 with 𝜙̃ 1 , we can assume that 𝑑(𝜙1 )0 is orientation pre-
serving, and we orient all 𝑇𝑎 (𝑈1 ) as above. For 𝑈2 , we choose the orientation which is compat-
ible with the orientation of the 𝑇𝑎 (𝑋) for all points 𝑎 ∈ 𝑈1 ∩ 𝑈2 : If 𝑑(𝜙2 )𝜙−1 (𝑎) is orientation
2
preserving on 𝑇𝑎 (𝑋) for 𝑎 ∈ 𝑈1 ∩ 𝑈2 , we orient 𝑇𝑎 (𝑋) such that 𝑑(𝜙2 )𝜙−1 (𝑎) ∶ ℝ𝑘 → 𝑇𝑎 (𝑋)
2
is orientation preserving for all 𝑎 ∈ 𝑈2 . If it is not orientation preserving, then we replace
𝜙2 (𝑣) with 𝜙2 (−𝑣). Continuing this way, we obtain an orientation for 𝑈𝑚 and therefore an
orientation for 𝑇𝑦 (𝑋) after finitely many steps. See Figure 14.6.

It remains to show that the induced orientation on 𝑇𝑦 (𝑋) does not depend on the choice
of 𝛾 and the 𝑈𝑖 ’s. So let 𝜔 ∶ [0, 1] → 𝑋 be another smooth path with 𝜔(0) = 𝑥 and 𝜔(1) = 𝑦.
As for 𝛾, we choose open sets 𝑊1 , … , 𝑊𝑙 covering all points in 𝜔([0, 1]) with 𝑥 ∈ 𝑊1 and
𝑦 ∈ 𝑊𝑙 and 𝑊𝑖 ∩ 𝑊𝑖+1 ≠ ∅. Then we orient 𝑇𝑦 (𝑋) following the same procedure using the
𝑊𝑖 ’s.

Arriving at 𝑦, we do not know a priori whether the orientation of 𝑇𝑦 (𝑋) induced by 𝛾 and
the orientation of 𝑇𝑦 (𝑋) induced by 𝜔 agree. But now we use that 𝑋 is simply-connected.

Figure 14.6: We choose a path from 𝑥 to 𝑦 and open sets which overlap the path. Then we orient
the tangent spaces such that everything is compatible on overlaps. Simply-connectedness will
make sure that the choice of path and open sets did not matter, since we can shrink any path to
a point.

For, walking first along 𝛾 and then back on 𝜔 defines, after readjusting the speed and
smoothing things out, a loop 𝛼 ∶ [0, 1] → 𝑋 with 𝛼(0) = 𝑥 = 𝛼(1), i.e., a smooth map 𝛼 ∶ 𝕊1 →
302 14.9. Summary
𝑋. Walking along 𝛼, we obtain an isomorphism
≅
𝑇 (𝛼) ∶ 𝑇𝑥 (𝑋) = 𝑇𝛼(0) (𝑋) ←←←→
← 𝑇𝛼(1) (𝑋) = 𝑇𝑥 (𝑋)

by composing
𝑑(𝜙1 )−1
∙ 𝑑(𝜙2 )∙ 𝑑(𝜙2 )−1
∙ 𝑑(𝜙2 )∙ 𝑑(𝜓𝑚−1 )−1
∙ 𝑑(𝜓𝑚 )∙
← ℝ𝑘 ←←←←←←←←←←→
𝑇𝑥 (𝑋) ←←←←←←←←←←←←←→ ← ℝ𝑘 ←←←←←←←←←←→
← 𝑇𝑧 (𝑋) ←←←←←←←←←←←←←→ ← ℝ𝑘 ←←←←←←←←←←←→
← ⋯ ←←←←←←←←←←←←←←←←←→ ← 𝑇𝑥 (𝑋)

where the subscript ∙ stands for the varying points at which we take derivatives.

The isomorphism 𝑇 (𝛼) is either orientation preserving or reversing. If it preserves the

orientation, then its determinant is positive. If it reverses the orientation, then its determinant
is negative. And 𝑇 (𝛼) is orientation preserving if and only if the two orientations on 𝑇𝑦 (𝑋)
induced by 𝛾 and 𝜔, respectively, agree.

Since 𝑋 is simply-connected 𝛼 is homotopic to the constant map 𝑐𝑥 ∶ 𝕊1 → {𝑥}.

Let 𝐹 ∶ 𝕊1 ×[0, 1] → 𝑋 be a homotopy from 𝛼 to 𝑐𝑥 . Since 𝕊1 ×[0, 1] is compact, its image

in 𝑋 is compact and we can add finitely many open subsets to the collection 𝑈1 , … , 𝑈𝑚 , 𝑊1 , … , 𝑊𝑙
to cover 𝐹 (𝕊1 × [0, 1]) with the codomains of local parametrizations.

For each 𝑡 ∈ [0, 1], 𝐹 (−, 𝑡) defines a smooth loop from 𝑥 to 𝑥. Using the above procedure
for orienting tangent spaces along a path, we obtain an isomorphism
≅
← 𝑇𝐹 (1,𝑡) (𝑋) = 𝑇𝑥 (𝑋) for each 𝑡 ∈ [0, 1].
𝑇 (𝐹 (−, 𝑡)) ∶ 𝑇𝑥 (𝑋) = 𝑇𝐹 (0,𝑡) (𝑋) ←←←→

Taking the determinant of 𝑇 (𝐹 (−, 𝑡)) defines a map

[0, 1] → ℝ, 𝑡 → det(𝑇 (𝐹 (−, 𝑡)))

which is continuous, since each point of 𝑋 is contained an open neighborhood on which the
orientation is determined by the derivatives of local parametrizations, and these derivatives
vary smoothly with the base-points.

Since each 𝑇 (𝐹 (−, 𝑡)) is an isomorphism, its determinant is either strictly positive > 0 or
strictly negative < 0. Since [0, 1] is connected and 𝑡 → det(𝑇 (𝐹 (−, 𝑡))) is continuous, we have

either det(𝑇 (𝐹 (−, 𝑡))) > 0 or det(𝑇 (𝐹 (−, 𝑡))) < 0 for all 𝑡 ∈ [0, 1].

But we know that, for 𝑡 = 1, 𝐹 (−, 1) = 𝑐𝑥 is the constant loop at 𝑥. Thus

det(𝑇 (𝐹 (−, 1))) = det(Id𝑇𝑥 (𝑋) ) > 0.

Hence we must have det(𝑇 (𝐹 (−, 𝑡))) > 0 for all 𝑡 ∈ [0, 1]. In other words, 𝑇 (𝐹 (−, 𝑡)) must
be orientation preserving for all 𝑡, and in particular, 𝑇 (𝛼) is orientation preserving.

This shows that the orientation of 𝑇𝑦 (𝑋) does not depend on the choice of 𝛾.

14.9 Summary

Let us summarise the key points we should remember from this chapter:
Chapter 14. Orientation 303
∙ An orientation of a vector space is a choice of a sign, +1 or −1, for an equivalence of
orderings of a bases. We can think of it as choosing a positive and negative direction.

∙ An orientation on a manifold is a smooth choice of orientations of the tangent spaces

for each point. Such a smooth choice may or may not exist. Hence manifolds can be
orientable or not.

∙ Orientations help us classifying manifolds: there is a box with orientable and a box with
non-orientable manifolds.

∙ For any compact oriented one-dimensional manifold with boundary, the sum of the
orientation numbers at the boundary points is zero.

∙ The boundary of a cylinder has opposite orientations:

𝜕([0, 1] × 𝑋) = 𝑋1 − 𝑋0 .

This is the key point for defining homotopy invariant intersection numbers soon.

∙ There is a formula for the boundary of preimages:

sign (𝜕𝑓 −1 (𝑍)) as a boundary

= (−1)codim 𝑍 ⋅ sign (𝜕𝑓 −1 (𝑍)) as a preimage.

∙ An important class of orientable manifolds consists of simply-connected manifolds.

304 14.10. Exercises and more examples
14.10 Exercises and more examples

Exercise 14.1 Let 𝛽 = (𝑣1 , … , 𝑣𝑘 ) be an ordered basis of a vector space 𝑉 .

(a) Show that replacing one 𝑣𝑖 by a multiple 𝑐𝑣𝑖 yields an equivalently oriented ordered
basis if 𝑐 > 0, and an oppositely oriented one if 𝑐 < 0.

(b) Show that transposing two elements, i.e., interchanging the places of 𝑣𝑖 and 𝑣𝑗 for
𝑖 ≠ 𝑗, yields an oppositely oriented ordered basis.

(c) Show that subtracting from one 𝑣𝑖 a linear combination of the others yields an
equivalently oriented ordered basis.

(d) Suppose that 𝑉 is the direct sum of 𝑉1 and 𝑉2 . Show that the direct sum orientation
of 𝑉 from 𝑉1 ⊕ 𝑉2 equals (−1)(dim 𝑉1 )(dim 𝑉2 ) times the orientation from 𝑉2 ⊕ 𝑉1 .

Exercise 14.2 The upper half space ℍ𝑘 is oriented by the standard orientation of ℝ𝑘 .
Thus 𝜕ℍ𝑘 acquires a boundary orientation. But 𝜕ℍ𝑘 may be identified with ℝ𝑘−1 . Show
that the boundary orientation agrees with the standard orientation of ℝ𝑘−1 if and only if
𝑘 is even.

Exercise 14.3 In this exercise we study the orientation on spheres:

(a) Write down the orientation of 𝕊2 as the boundary of the closed unit ball 𝔹3 in
ℝ3 , by specifying a positively oriented ordered basis for the tangent space at each
(𝑎, 𝑏, 𝑐) ∈ 𝕊2 .

(b) Show that the boundary orientation of 𝕊𝑘 equals the orientation of 𝕊𝑘 = 𝑔 −1 (1) as
the preimage under the map

𝑔 ∶ ℝ𝑘+1 → ℝ, 𝑥 → |𝑥|2 .

Exercise 14.4 Suppose that 𝑓 ∶ 𝑋 → 𝑌 is a diffeomorphism of connected oriented

manifolds with boundary. Show that if 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝑋) → 𝑇𝑓 (𝑥) (𝑌 ) preserves orientation at
one point 𝑥, then 𝑓 preserves orientation at every point.

Exercise 14.5 Let 𝑋 and 𝑍 be transversal submanifolds in 𝑌 and assume 𝑋, 𝑍 and 𝑌

are oriented. Let 𝑖 ∶ 𝑋 → 𝑌 be the inclusion of 𝑋 into 𝑌 , 𝑗 ∶ 𝑍 → 𝑌 be the inclusion of
𝑍 into 𝑌 . We orient the intersection 𝑋 ∩ 𝑍 as the preimage 𝑖−1 (𝑍), and the intersection
𝑍 ∩ 𝑋 as the preimage 𝑗 −1 (𝑋). Show that the orientations of 𝑋 ∩ 𝑍 and 𝑍 ∩ 𝑋 are
related by
𝑋 ∩ 𝑍 = (−1)(codim 𝑋)(codim 𝑍) 𝑍 ∩ 𝑋.
Hint: Show that the orientation of 𝑆 = 𝑋 ∩ 𝑍 at any 𝑦 is induced by the direct sum
(𝑁𝑦 (𝑆, 𝑋) ⊕ 𝑁𝑦 (𝑆, 𝑍)) ⊕ 𝑇𝑦 (𝑆) = 𝑇𝑦 (𝑌 ).
Chapter 14. Orientation 305
What happens when you consider 𝑍 ∩ 𝑋 instead?

Exercise 14.6 Let 𝑋 be an oriented manifold and let 𝑍 ⊂ 𝑋 be a submanifold of

codimension one. Show that 𝑍 is orientable if and only if the relative normal bundle
𝑁(𝑍, 𝑋) is trivial.

Exercise 14.7 (a) Let 𝑉 be a vector space. Show that both orientations on 𝑉 define
the same product orientation on 𝑉 × 𝑉 .

(b) Let 𝑋 be an orientable manifold. Show that the product orientation on 𝑋 × 𝑋 is

the same for all choices of orientation on 𝑋.

(c) Suppose that 𝑋 is not orientable. Show that 𝑋 × 𝑌 is never orientable, no matter
what manifold 𝑌 may be. In particular, 𝑋 × 𝑋 is not orientable.
Hint: First show that 𝑋 × ℝ𝑚 is not orientable, and then use that every 𝑌 has an
open subset diffeomorphic to ℝ𝑚 .

(d) Prove that there exists a natural orientation on some neighborhood of the diagonal
Δ in 𝑋 × 𝑋, whether or not 𝑋 can be oriented.
But note that Δ itself is orientable if and only if 𝑋 × 𝑋 is orientable. Why?
Hint: Cover a neighborhood of Δ by local parametrizations 𝜙 × 𝜙 ∶ 𝑈 × 𝑈 →
𝑋 × 𝑋, where 𝜙 ∶ 𝑈 → 𝑋 is a local parametrization of 𝑋, then apply the previous
observations.

Exercise 14.8 Recall the Hopf fibration 𝜋 that we have seen previously: We consider
𝕊3 as a subset of ℂ2 , i.e., 𝕊3 = {(𝑧0 , 𝑧1 ) ∈ ℂ2 ∶ |𝑧0 |2 + |𝑧1 |2 = 1}, and 𝕊2 as a subset
of ℂ × ℝ, i.e., 𝕊2 = {(𝑧, 𝑥) ∈ ℂ × ℝ ∶ |𝑧|2 + 𝑥2 = 1}. Then the Hopf fibration 𝜋 is the
map 𝕊3 → 𝕊2 given by
( )
𝜋(𝑧0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2 .

Consider the point 𝑏 = (0, 1, 0) on 𝕊2 ⊂ ℝ3 ≅ ℂ × ℝ. Recall that that computed

the fiber 𝜋 −1 (𝑏) in a previous exercise. Determine the orientation of the fiber 𝜋 −1 (𝑏) as a
preimage under 𝜋.
15. The Brouwer Degree

15.1 The integral-valued Brouwer Degree

We have seen in Section 11.1 that the mod 2 degree is a powerful invariant of smooth maps
between manifolds. Now we remedy the defect that it was merely an element in ℤ∕2 and define
an integer-valued degree. We will then study several applications.

15.1.1 A well-defined invariant

Let 𝑋 and 𝑌 be oriented 𝑛-dimensional smooth manifolds without boundary. Let 𝑓 ∶ 𝑋 → 𝑌

be a smooth map. We assume that 𝑋 is compact and that 𝑌 is connected.

Then the degree of 𝑓 is defined as follows: Let 𝑥 ∈ 𝑋 be a regular point of 𝑓 . With our
assumptions this means that 𝑑𝑓𝑥 ∶ 𝑇𝑥 𝑋 → 𝑇𝑓 (𝑥) 𝑌 is a linear isomorphism between oriented
vector spaces. The sign of 𝑑𝑓𝑥 is defined to be +1 if 𝑑𝑓𝑥 preserves orientations and it is defined
to be −1 if 𝑑𝑓𝑥 reverses orientations. Now, for a regular value 𝑦 ∈ 𝑌 of 𝑓 , we define the integer
∑
deg(𝑓 ; 𝑦) = sign 𝑑𝑓𝑥 .
𝑥∈𝑓 −1 (𝑦)

Figure 15.1: We calculate the degree by counting the number of points in the fiber of a regular
value. The number is well-defined if we take signs determined by orientation into account. In
dimension one, this boils down to looking at the sign of the derivative or by checking whether
the graph hits the horizontal line corresponding to a value from below or above.

In the following we keep the above assumptions.

306
Chapter 15. The Brouwer Degree 307

Lemma 15.1 (deg is locally constant) The function 𝑌 → ℤ, 𝑦 → deg(𝑓 ; 𝑦) is locally

constant on the set of regular values of 𝑓 .

Proof: By the Stack of Records Theorem 4.14, we can find a neighborhood 𝑈 of 𝑦 such
that the preimage 𝑓 −1 (𝑈 ) is a disjoint union 𝑉1 ∪ ⋯ ∪ 𝑉𝑛 , where each 𝑉𝑖 is an open set in 𝑋
mapped by 𝑓 diffeomorphically onto 𝑈 . Hence, for all points 𝑧 ∈ 𝑈 , we have #𝑓 −1 ({𝑧}) = 𝑛.
It remains to take orientations into account. Since 𝑓|𝑉𝑖 ∶ 𝑉𝑖 → 𝑈 is a diffeomorphism, we
know that
𝑑𝑓𝑥𝑖 ∶ 𝑇𝑥𝑖 (𝑋) → 𝑇𝑦 (𝑌 )
is an isomorphism. Now both 𝑇𝑥𝑖 (𝑋) and 𝑇𝑦 (𝑌 ) are oriented, and hence 𝑑𝑓𝑥𝑖 is either orientation
preserving or reversing. But by our definition of orientations on manifolds, we have

∙ either det(𝑑𝑓𝑥𝑖 ) > 0 and hence, for all 𝑧 ∈ 𝑈 , det(𝑑𝑓𝑤𝑖 ) > 0, where 𝑤𝑖 is the unique
point in 𝑉𝑖 with 𝑓 (𝑤𝑖 ) = 𝑧; in other words, 𝑑𝑓𝑤𝑖 preserves orientations for all points
𝑤𝑖 ∈ 𝑉𝑖 ;

∙ or det(𝑑𝑓𝑥𝑖 ) < 0 and hence, for all 𝑧 ∈ 𝑈 , det(𝑑𝑓𝑤𝑖 ) < 0, where 𝑤𝑖 is the unique point
in 𝑉𝑖 with 𝑓 (𝑤𝑖 ) = 𝑧; in other words, 𝑑𝑓𝑤𝑖 reverses orientations for all points 𝑤𝑖 ∈ 𝑉𝑖 .

Thus the orientation number is the same for all points in 𝑉𝑖 . Hence the sum of orientation
numbers of the points in 𝑓 −1 (𝑧) is the same for all points 𝑧 ∈ 𝑈 . Consequently, the function

𝑌 → ℤ, 𝑦 → deg(𝑓 ; 𝑦)

is locally constant.

Now we show a very useful theorem about the degree which generalizes Theorem 11.8.

Theorem 15.2 (Boundary Theorem for deg) Assume that 𝑋 = 𝜕𝑊 is the boundary
of a compact oriented manifold 𝑊 and that 𝑋 is oriented as the boundary of 𝑊 . If
𝑓 ∶ 𝑋 → 𝑌 can be extended to a smooth map 𝐹 ∶ 𝑊 → 𝑌 , then deg(𝑓 ; 𝑦) = 0 for
every regular value 𝑦 of 𝑓 .

Proof: First we suppose that 𝑦 is a regular value for both 𝐹 and 𝑓 . By Theorem 10.14,
𝐹 −1 (𝑦) is a compact submanifold of 𝑊 with boundary of dimension dim 𝑊 − dim 𝑌 = 1.
Thus, by Theorem 10.17, 𝐹 −1 (𝑦) is the disjoint union of finitely many connected components
which are diffeomorphic to either 𝑆 1 or [0, 1]. Since 𝑆 1 does not have boundary points, only
the boundary points of the components diffeomorphic to [0, 1] lie on 𝑋 = 𝜕𝑊 . Let 𝐴 be one
such component diffeomorphic to [0, 1] and let 𝑎 and 𝑏 denote the boundary points of 𝐴, i.e.,
𝜕𝐴 = {𝑎} ∪ {𝑏}. We will show that

sign 𝑑𝑓𝑎 + sign 𝑑𝑓𝑏 = 0. (15.1)

Since 𝑓 −1 (𝑦) = (𝐹|𝜕𝑊 )−1 (𝑦) consists of the finitely many boundary points of the components
diffeomorphic to [0, 1], this implies that the sum of the signs of all 𝑑𝑓𝑥 for 𝑥 ∈ 𝑓 −1 (𝑦) is zero
as claimed.
308 15.1. The integral-valued Brouwer Degree
To show (15.1) we look at how the tangent spaces at points in 𝐴 are oriented. So let 𝑥 ∈ 𝐴
and let (𝑣1 , 𝑣2 , … , 𝑣𝑛 , 𝑤) be a positively oriented bases for 𝑇𝑥 𝑊 such that 𝑤 is tangent to the
one-dimensional submanifold 𝐴, i.e., 𝑤 forms a basis of the one-dimensional vector subspace
𝑇𝑥 𝐴 ⊂ 𝑇𝑥 𝑊 , while (𝑣1 , … , 𝑣𝑛 ) is an ordered basis of the normal space 𝑁𝑥 (𝐴; 𝑊 ) ⊂ 𝑇𝑥 𝑊 .
Since we have a direct sum decomposition 𝑁𝑥 (𝐴; 𝑊 ) ⊕ 𝑇𝑥 𝐴 = 𝑇𝑥 𝑊 , we get the equation

sign (𝑣1 , … , 𝑣𝑛 ) ⋅ sign (𝑤) = sign (𝑣1 , 𝑣2 , … , 𝑣𝑛 , 𝑤) = +1

Thus we have

sign (𝑤) = +1 in 𝑇𝑥 𝐴 ⇐⇒ sign (𝑣1 , … , 𝑣𝑛 ) = +1 in 𝑁𝑥 (𝐴; 𝑊 ).

Since 𝐴 is a component of 𝐹 −1 (𝑦), it inherits an orientation as a preimage as explained in

Section 14.5. This means that sign (𝑣1 , … , 𝑣𝑛 ) in 𝑁𝑥 (𝐴; 𝑊 ) is determined by the effect of
𝑑𝐹𝑥 on (𝑣1 , … , 𝑣𝑛 ). As in Definition 14.22 we have
( ( ))
sign (𝑣1 , … , 𝑣𝑛 ) = sign 𝑑𝐹𝑥 𝑣1 , … , 𝑣𝑛

where the right-hand side is determined by the given orientation of 𝑇𝑦 𝑌 . Thus, we conclude
that 𝑤 is a positively oriented basis of 𝑇𝑥 𝐴 if and only if 𝑑𝐹𝑥 sends the basis (𝑣1 , … , 𝑣𝑛 ) to a
positively oriented basis in 𝑇𝑦 𝑌 .

Let 𝑤(𝑥) be the positively oriented unit vector 𝑇𝑥 𝐴. Then 𝑤(𝑥) depends smoothly of 𝑥 and
points outward at one boundary point, say 𝑏, and inward at the other boundary point 𝑎. At the
boundary points we have the commutative diagrams

𝑇𝑎 𝑋 ⊕ 𝑇𝑎 𝐴 𝑇𝑎 𝑊 and 𝑇𝑏 𝑋 ⊕ 𝑇𝑏 𝐴 𝑇𝑏 𝑊
𝑑𝑓𝑎 𝑑𝑓𝑏
𝑑𝐹𝑎 𝑑𝐹𝑏
 y  y
𝑇𝑦 𝑌 𝑇𝑦 𝑌 .

The kernel of 𝑑𝐹𝑥 𝑇𝑥 𝐴 and, at the boundary points, 𝑑𝐹𝑥 restricts to 𝑑𝑓𝑥 . We just learned that
the sign of 𝑤(𝑥) is positive if and only if sign 𝑑𝑓𝑥 = +1. Thus, since 𝑤(𝑏) is positively oriented
as the outward pointing vector and 𝑤(𝑎) negatively, we conclude

sign 𝑑𝑓𝑎 = −1 and sign 𝑑𝑓𝑏 = +1.

Summing over all arcs in 𝐹 −1 (𝑦), we have proved the assertion when 𝑦 is a regular value for 𝐹 .

Now suppose that 𝑦0 is a regular value for 𝑓 , but not for 𝐹 . The function 𝑦 → deg(𝑓 ; 𝑦)
is constant within some open neighborhood 𝑈 of 𝑦0 by Lemma 15.1. By Sard’s Theorem 7.1
the open subset 𝑈 contains a regular value for 𝐹 . Hence we can choose a regular value 𝑦 for 𝐹
within 𝑈 and get

deg(𝑓 ; 𝑦0 ) = deg(𝑓 ; 𝑦).

The previous case then implies the assertion of the theorem.

Lemma 15.3 (Homotopy Lemma for deg) Let 𝑓 , 𝑔 ∶ 𝑋 → 𝑌 be two smooth maps.
Assume 𝑓 and 𝑔 are smoothly homotopic. If 𝑦 ∈ 𝑌 is a regular value for both 𝑓 and
Chapter 15. The Brouwer Degree 309
𝑔, then

𝑓 −1 (𝑦) = 𝑔 −1 (𝑦).

Proof: Let 𝐹 ∶ [0, 1] × 𝑋 → 𝑌 be a smooth homotopy between 𝐹0 = 𝑓 and 𝐹1 = 𝑔.

Recall that the boundary of [0, 1] × 𝑋 is given by {0} × 𝑋 ∪ {1} × 𝑋 where the orientation
of {1} × 𝑋 is the one of 𝑋 and the orientation of {0} × 𝑋 is the opposite one. Since the
degree is additive on connected components, we deduce that the degree at 𝑦 of the restriction
𝜕𝐹 ∶= 𝐹|𝜕([0,1]×𝑋 ∶ {0} × 𝑋 ∪ {1} × 𝑋 → 𝑌 of 𝐹 to the boundary of [0, 1] × 𝑋 is equal to the
difference

deg(𝑔; 𝑦) − deg(𝑓 ; 𝑦).

Since 𝜕𝐹 can be extended to the compact oriented manifold [0, 1] × 𝑋, this difference must be
zero by Theorem 15.2.

Now we can prove the following key result:

Theorem 15.4 (The Brouwer Degree is well-defined) Let 𝑋 and 𝑌 be oriented

𝑛-dimensional smooth manifolds without boundary, with 𝑋 compact and 𝑌 connected.
Let 𝑓 ∶ 𝑋 → 𝑌 be a smooth map. The number deg(𝑓 ; 𝑦) does not depend on the choice
of the regular value 𝑦, and we denote it by deg 𝑓 . Moreover, deg 𝑓 only depends on the
homotopy class of 𝑓 , i.e., if 𝑓 and 𝑔 are smoothly homotopic then deg 𝑓 = deg 𝑔.

Proof: Let 𝑦 and 𝑧 be two regular values of 𝑓 . By the Isotopy Lemma 11.3 we can choose a
diffeomorphism ℎ ∶ 𝑋 → 𝑌 which isotopic to the identity and with ℎ(𝑦) = 𝑧. Every diffeomor-
phism either preserves or reverses the orientation at every point. Hence we know sign 𝑑𝑥 ℎ = ±1
for all 𝑥. Since ℎ is isotopic to the identity, we must have sign 𝑑ℎ𝑥 = +1 for all 𝑥. Thus ℎ pre-
serves orientation, and we get

deg(𝑓 ; 𝑦) = deg(ℎ◦𝑓 ; ℎ(𝑦)).

Since ℎ is homotopic to the identity, 𝑓 is homotopic to ℎ◦𝑓 . Hence

deg(ℎ◦𝑓 ; 𝑧) = deg(𝑓 ; 𝑧)

by Lemma 15.3. Thus we get deg(𝑓 ; 𝑦) = deg(𝑓 ; 𝑧).

Now let 𝑔 ∶ 𝑋 → 𝑌 be a smooth map which is smoothly homotopic to 𝑓 . By Sard’s Theo-

rem 7.1 we can choose a 𝑦 ∈ 𝑌 which is regular value for both 𝑓 and 𝑔. By the first assertion,
we can use 𝑦 to calculate deg(𝑓 ) = deg(𝑓 ; 𝑦) and deg(𝑔) = deg(𝑔; 𝑦). Then Lemma 15.3 applies
and shows deg(𝑓 ) = deg(𝑔).

In the exercises we are going to show that the degree is multiplicative in the following sense:

Lemma 15.5 (The degree is multiplicative) Let 𝑓 ∶ 𝑋 → 𝑌 and 𝑔 ∶ 𝑌 → 𝑍 be

smooth maps between manifolds with 𝑋 and 𝑌 compact, 𝑌 and 𝑍 connected. Assume
that all three manifolds are oriented, without boundary and that dim 𝑋 = dim 𝑌 =
310 15.1. The integral-valued Brouwer Degree
dim 𝑍. Then we have
deg(𝑔◦𝑓 ) = deg(𝑔) ⋅ deg(𝑓 ).

Proof: This is Exercise 15.6.

15.1.2 Some examples

In all examples we keep the assumption that 𝑋 and 𝑌 are smooth manifolds without boundary,
𝑋 is compact, 𝑌 is connected and dim 𝑋 = dim 𝑌 .

Example 15.6 (Constant maps have degree zero) Let 𝑓 ∶ 𝑋 → 𝑌 be a constant map
with value 𝑦0 . Then deg(𝑓 ) = 0. For, any value 𝑦 ≠ 𝑦0 is a regular value and 𝑓 −1 (𝑦) = ∅.

Example 15.7 (Degree of a diffeomorphism) Let 𝑓 ∶ 𝑋 → 𝑌 be a diffeomorphism.

Then we have #𝑓 −1 (𝑦) = 1 for every 𝑦 ∈ 𝑌 , and deg(𝑓 ) = +1 if 𝑓 preserves the
orientation and deg(𝑓 ) = −1 if 𝑓 reverses the orientation.

In particular, we get:

Lemma 15.8 (Obstruction to a homotopy to the identity) An orientation reversing

diffeomorphism of a compact boundaryless manifold is not smoothly homotopic to the
identity.

Example 15.9 (Reflection on 𝕊𝑛 ) An example of an orientation reversing diffeomor-

phism is provided by the reflection 𝑟𝑖 ∶ 𝕊𝑛 → 𝕊𝑛 which we have seen in the exercises
before:

𝑟𝑖 (𝑥1 , … , 𝑥𝑛+1 ) = (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑛+1 ).

By Example 15.7 we have deg(𝑟𝑖 ) = −1.

Example 15.10 (Degree of self-maps of 𝕊1 ) Recall that the restriction of complex

multiplication 𝑧 → 𝑧𝑚 defines a smooth map 𝑓𝑚 ∶ 𝕊1 → 𝕊1 for every 𝑚 ∈ ℤ. For 𝑚 ≠ 0,
let us calculate the derivative 𝑑(𝑓𝑚 )𝑧 ∶ 𝑇𝑧 (𝕊1 ) → 𝑇𝑓𝑚 (𝑧) (𝕊1 ). We use the parametrization
𝜙 ∶ 𝑡 → (cos 𝑡, sin 𝑡). We have the commutative diagram

𝑓𝑚
𝕊O 1 / 𝕊1
O
𝜙 𝜙

ℝ / ℝ.
𝑡→𝑚𝑡

Taking derivatives yields, where we note that 𝑡 → 𝑚𝑡 is a linear map and therefore equal
Chapter 15. The Brouwer Degree 311
to its derivative:
𝑑(𝑓𝑚 )𝑧
𝑇𝑧 (𝕊1 ) / 𝑇 𝑚 (𝕊1 )
O 𝑧 O

𝑑𝜙𝑡 𝑑𝜙𝑚𝑡

ℝ / ℝ.
𝑡→𝑚𝑡

In order to determine 𝑑(𝑓𝑚 )𝑧 , recall that 𝑑𝜙𝑡 has the form

𝑑𝜙𝑡 ∶ ℝ → ℝ2 , 𝑠 → (− sin 𝑡, cos 𝑡) ⋅ 𝑠

and hence at 𝑧 = 𝜙(𝑡):

𝑇𝑧 (𝕊1 ) = (− sin 𝑡, cos 𝑡) ⋅ ℝ.

Putting these information together we obtain

𝑑(𝑓𝑚 )𝑧 ∶ 𝑇𝑧 (𝕊1 ) → 𝑇𝑧𝑚 (𝕊1 ),

(− sin 𝑡, cos 𝑡) ⋅ 𝑠 → 𝑚 ⋅ (− sin(𝑚𝑡), cos(𝑚𝑡)) ⋅ 𝑠

which means that 𝑑(𝑓𝑚 )𝑧 is the linear map given by multiplication by 𝑚. Hence, when
𝑚 > 0, 𝑓𝑚 wraps the circle uniformly around itself 𝑚 times preserving orientation. The
map is everywhere regular and orientation preserving, so its degree is the number
of preimages of any point. And that number is 𝑚. Similarly, when 𝑚 < 0 the map
is everywhere regular but orientation reversing. As each point has |𝑚| preimages, the
degree is −|𝑚| = 𝑚. Finally, when 𝑚 = 0 the map is constant, so its degree is zero.

Remark 15.11 (One homotopy class 𝕊1 → 𝕊1 for each integer) One immediate
consequence of this calculation (which could not have been proven with mod 2 theory)
is the interesting fact that the circle admits an infinite number of homotopically distinct
maps since deg(𝑧𝑚 ) = 𝑚 implies that 𝑓𝑛 and 𝑓𝑚 are not homotopic if 𝑛 ≠ 𝑚. We provide
a more complete picture of the self-maps of 𝕊1 in Theorem 15.18.

Example 15.12 (Self-maps of 𝕊2 and complex polynomials) Let 𝑝(𝑧) = 𝑧𝑚 +𝑎1 𝑧𝑚−1 +
⋯ + 𝑎𝑚 be a complex monic polynomial. We may consider 𝑝 as a smooth map ℂ → ℂ,
and as in Equation 4.1, it induces a smooth map 𝑓 ∶ 𝕊2 → 𝕊2 which sends ∞ to ∞, i.e.,
𝑓 sends the north pole to the north pole. We claim that the map 𝑓 has degree 𝑚.
To justify our claim we first observe that the homotopy

𝐻(𝑧, 𝑡) = 𝑡𝑧𝑚 + (1 − 𝑡)(𝑎1 𝑧𝑚−1 + ⋯ + 𝑎𝑚 )

induces a smooth homotopy 𝐹 ∶ 𝕊2 × [0, 1] → 𝕊2 between the maps induced by 𝑝 and

the monomial 𝑧𝑚 . The invariance of the degree under homotopy then shows that we can
assume that 𝑝(𝑧) = 𝑧𝑚 is a monomial. Then the construction of 𝑓 and the computation of
Example 15.10 yield the claim. We recommend to look at Exercise 15.3, Exercise 15.4
and Exercise 15.5 which study this situation further. Hopf’s Degree Theorem 16.1 yields
a more general statement.
312 15.1. The integral-valued Brouwer Degree

Theorem 15.13 (Self-maps of 𝕊𝑘 ) Let 𝑘 ≥ 1 and 𝑚 ∈ ℤ be an integer. Then there is

a smooth map 𝑓𝑚 ∶ 𝕊𝑘 → 𝕊𝑘 with deg(𝑓𝑚 ) = 𝑚.

Proof: For 𝑚 = 0 it suffices to take any constant map. For 𝑘 = 1 and 𝑚 ∈ ℤ, we proved
the assertion in Example 15.10. We can generalize this case as follows. First we assume that
𝑚 is positive,√i.e., we assume√𝑚 ≥ 1. We can describe the coordinates of any point in 𝕊𝑘 as
the tuple (𝑥, 1 − |𝑥|2 cos 𝑡, 1 − |𝑥|2 sin 𝑡) where 𝑥 = (𝑥1 , … , 𝑥𝑘−1 ) denotes the first 𝑘 − 1
coordinates and |𝑥|2 = 𝑥21 + ⋯ 𝑥2𝑘−1 . We define the smooth map 𝑓𝑚 ∶ 𝕊𝑘 → 𝕊𝑘 defined by
√ √ √ √
(𝑥, 1 − |𝑥|2 cos 𝑡, 1 − |𝑥|2 sin 𝑡) → (𝑥, 1 − |𝑥|2 cos(𝑚𝑡), 1 − |𝑥|2 sin(𝑚𝑡)).

The map 𝑓𝑚 is surjective and by Sard’s Theorem 7.1 we can find a regular value 𝑝 ∈ 𝕊𝑘 . The
preimage of 𝑝 consists of exactly 𝑚 points. It remains to check the sign of the orientation at each
preimage point 𝑞 ∈ 𝑓 −1 (𝑝). With respect to the bases of 𝑇𝑞 𝕊𝑘 and 𝑇𝑓𝑚 (𝑞) 𝕊𝑘 that are induced by
the standard basis in ℝ𝑘 , the derivative 𝑑(𝑓𝑚 )𝑞 at 𝑞 is given by the (𝑘 × 𝑘)-matrix which consists
of the (𝑘 − 1 × 𝑘 − 1)-identity matrix and the integer 𝑚 in the bottom right-hand corner. This
matrix has determinant 𝑚 > 0. Thus, the orientation is preserved at each point in 𝑞 ∈ 𝑓 −1 (𝑝),
and we get deg(𝑓𝑚 ) = 𝑚.

If 𝑚 < 0 is negative, we consider the composition 𝑓|𝑚| ◦𝑟1 where 𝑟1 denotes the reflection
in the first coordinate. Since the degree is multiplicative by Lemma 15.5, this map has degree
𝑚 by the first case and Example 15.9.

Remark 15.14 (Homotopy groups of spheres - surjectivity) Let [𝕊𝑘 , 𝕊𝑘 ] denote the
set of smooth maps 𝕊𝑘 → 𝕊𝑘 modulo homotopy. By Theorem 15.4 we can think of the
degree as a map

deg ∶ [𝕊𝑘 , 𝕊𝑘 ] ←→ ℤ.

By Theorem 15.13, this map is surjective. We will later prove Hopf’s Theorem 16.1
which tells us that this map is injective as well.

In the exercises we are going ro prove the following two applications of the degree:

Theorem 15.15 (Fixed points of self-maps of the sphere) Let 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 be a

smooth map. Assume that deg(𝑓 ) ≠ (−1)𝑘+1 . Then 𝑓 must have a fixed point.

Proof: We prove the assertion in Exercise 15.7.

Theorem 15.16 (Self-map of the sphere of odd degree) Let 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 be a

smooth map. Assume deg(𝑓 ) is odd. Then 𝑓 must send some pair of antipodal points
to antipodal points, i.e., there is at least one pair of antipodal points 𝑥0 , −𝑥0 such that
𝑓 (−𝑥0 ) = −𝑓 (𝑥0 ).

Proof: We prove the assertion in Exercise 15.9.

Chapter 15. The Brouwer Degree 313
∙ Note that Theorem 15.15 does not follow from the Borsuk–Ulam Theorem 11.12, but
rather is a partial converse.

Example 15.17 (Self-maps of the torus) Let 𝕋 2 = 𝕊1 × 𝕊1 be the two-dimensional

torus. We have seen that we also can view it as a quotient of a square where we identify
opposite sides. Yet another presentation of 𝕋 2 is as the quotient ℝ2 ∕ℤ2 where two points
𝑝, 𝑞 ∈ ℝ2 are identified if 𝑝 − 𝑞 ∈ ℤ2 . We equip this space with the quotient topology
induced by the quotient map ℝ2 → ℝ2 ∕ℤ2 . Now let 𝑛 be a fixed positive integer. We
consider the two linear maps ℝ2 → ℝ2 represented by the matrices
( ) ( )
𝑛 0 1 0
𝐴= and 𝐵 = .
0 1 0 𝑛

Since 𝐴 and 𝐵 send ℤ2 to ℤ2 , they induce well-defined maps

𝕋 2 = ℝ2 ∕ℤ2 ←→ ℝ2 ∕ℤ2 = 𝕋 2 .

We denote the induced maps on 𝕋 2 by the same symbols 𝐴 and 𝐵 respectively. The
effect of 𝐴 on 𝕋 2 is to revolve 𝑛 times around the circle 𝕊1 × {𝑦} in 𝕋 2 , for some 𝑦 ∈ 𝕊1 .
While the effect of 𝐵 on 𝕋 2 is to revolve 𝑛 times around the circle {𝑥} × 𝕊1 in 𝕋 2 , for
some 𝑥 ∈ 𝕊1 . Hence, for any point 𝑞 ∈ 𝕋 2 , the preimages 𝐴−1 (𝑞) and 𝐵 −1 (𝑞) consist of
exactly 𝑛 points.
The derivative 𝑑𝐴𝑝 of 𝐴 at any point 𝑝 ∈ 𝕋 2 equals the linear map ℝ2 → ℝ2 repre-
sented by 𝐴, i.e., 𝑑𝐴𝑝 = 𝐴, and similarly 𝑑𝐵𝑝 = 𝐵. In particular, we get

det(𝑑𝐴𝑝 ) = det(𝐴) = 𝑛 = det(𝐵) = det(𝑑𝐵𝑝 ).

Since 𝑛 > 0, the determinant of 𝑑𝐴𝑝 and 𝑑𝐵𝑝 is positive at every point 𝑝 which is sent
to 𝑞. Thus we can conclude
deg(𝐴) = 𝑛 = deg(𝐵).

15.2 Hopf Degree Theorem in dimension one

We return to self-maps of 𝕊1 . We learned that there is a homotopy class of maps 𝕊1 → 𝕊1

for every integer 𝑚. Actually, the following theorem, the one-dimensional case of a famous
theorem of Hopf, shows that the degree is a bijective map
deg ∶ [𝕊1 , 𝕊1 ] → ℤ, 𝑓 → deg(𝑓 ),
where [𝕊1 , 𝕊1 ] = Hom(𝕊1 , 𝕊1 )∕∼ denotes the set of equivalence classes of smooth maps from
𝕊1 to 𝕊1 modulo the homotopy relation.

Theorem 15.18 (Hopf Degree Theorem in dimension one) Two smooth maps
𝑓0 , 𝑓1 ∶ 𝕊1 → 𝕊1 are homotopic if and only if they have the same degree.

∙ We generalize the assertion to all 𝑛 in Theorem 16.1.

314 15.2. Hopf Degree Theorem in dimension one
∙ Note that the statement also holds for continuous maps. Since we know that every con-
tinuous map between smooth manifolds is homotopic to a smooth map by Whitney’s
Approximation Theorem 12.18, all we need to extend the assertion is a definition of the
degree for merely continuous maps. This can be done using the techniques of algebraic
topology, e.g., singular homology. See for example the notes [16].

Proof of Theorem 15.18: We already know that if 𝑓0 and 𝑓1 are homotopic, then we have
deg(𝑓0 ) = deg(𝑓1 ). So assume deg(𝑓0 ) = deg(𝑓1 ), and we need to show 𝑓0 ∼ 𝑓1 . Recall from
Figure 11.4 the map 𝑝 defined by

𝑝 ∶ ℝ → 𝕊1 , 𝑡 → 𝑒2𝜋𝑖𝑡 ,

and showed that every smooth map 𝑓 ∶ 𝕊1 → 𝕊1 can be lifted1 to a map 𝑔 ∶ ℝ → ℝ with

𝑔(𝑡 + 1) = 𝑔(𝑡) + 𝑞 for some 𝑞 ∈ ℤ such that 𝑓 (𝑝(𝑡)) = 𝑝(𝑔(𝑡)).

If we can show 𝑞 = deg(𝑓 ), then we get a homotopy 𝑓0 ∼ 𝑓1 as follows:

Let 𝑔0 and 𝑔1 be smooth maps ℝ → ℝ which both satisfy

𝑔0 (𝑡 + 1) = 𝑔0 (𝑡) + 𝑞 and 𝑔1 (𝑡 + 1) = 𝑔1 (𝑡) + 𝑞

such that 𝑓0 (𝑝(𝑡)) = 𝑝(𝑔0 (𝑡)), 𝑓1 (𝑝(𝑡)) = 𝑝(𝑔1 (𝑡)). Then the map

𝑔𝑠 (𝑡) ∶= 𝑠𝑔1 + (1 − 𝑠)𝑔0 also satisfies 𝑔𝑠 (𝑡 + 1) = 𝑔𝑠 (𝑡) + 𝑞.

Note 𝑔𝑠 (𝑡) defines a homotopy 𝐺 from 𝑔0 to 𝑔1 by 𝐺(𝑡, 𝑠) = 𝑔𝑠 (𝑡).

But any homotopy

𝐺 ∶ ℝ × [0, 1] → ℝ with 𝐺(𝑡 + 1, 𝑠) = 𝐺(𝑡, 𝑠) + 𝑞 for all 𝑡, 𝑠

induces a well-defined homotopy

𝐹 ∶ 𝕊1 × [0, 1] → 𝕊1 , (𝑧, 𝑠) → 𝑝(𝐺(𝑡, 𝑠)) for any 𝑡 ∈ 𝑝−1 (𝑧).

Hence the above 𝑔𝑠 (𝑡) induces a homotopy from

𝑓0 = 𝑝◦𝑔0 to 𝑝◦𝑔1 = 𝑓1 .

It remains to prove:

Claim: 𝑞 = deg(𝑓 ).

∙ First, note that if 𝑓 is not surjective, then we can pick a point 𝑦 ∉ 𝑓 (𝕊1 ). Thus 𝑦 is
automatically a regular value. Since #𝑓 −1 (𝑦) = 0, we must have deg(𝑓 ) = 0. In this
case, we need to have 𝑞 = 0, i.e., 𝑔(𝑡 + 1) = 𝑔(𝑡). For otherwise 𝑝◦𝑔 was surjective and
hence 𝑓 would be surjective.
Note that, since the stereographic projection map 𝕊1 ⧵ {𝑦} → ℝ is a diffeomorphism
and ℝ is contractible, this shows that 𝕊1 ⧵ {𝑦} is contractible. Hence 𝑓 is a map to a
contractible space and therefore homotopic to a constant map and has degree 0.
1
The idea was to lift piecewise locally and to patch the pieces together.
Chapter 15. The Brouwer Degree 315
∙ Now we assume that 𝑓 is surjective. Let 𝑦 ∈ 𝕊1 be a regular value of 𝑓 , and let 𝑧 ∈
𝑓 −1 (𝑦). Since 𝑝 is surjective, there is a 𝑡 ∈ ℝ with 𝑝(𝑡) = 𝑧. Since 𝑦 is a regular value, 𝑓
is a local diffeomorphism around 𝑧. Its derivative is related to the one of 𝑔 by the chain
rule

𝑑𝑓𝑧 ◦𝑑𝑝𝑡 = 𝑑𝑝𝑔(𝑡) ◦𝑑𝑔𝑡 .

The derivative of 𝑝 ∶ ℝ → 𝕊1 at any 𝑡 is

𝑑𝑝𝑡 ∶ ℝ → 𝑇𝑝(𝑡) (𝕊1 ), 𝑤 → 2𝜋 ⋅ (− sin(2𝜋𝑡), cos(2𝜋𝑡)) ⋅ 𝑤.

Hence the determinant of 𝑑𝑝𝑡 at any 𝑡 is positive (in fact equal +2𝜋). Thus the sign of the
determinant of 𝑑𝑓𝑧 equals the sign of 𝑑𝑔𝑡 ∈ ℝ.
As above, let 𝑦 ∈ 𝕊1 be a regular value of 𝑓 and 𝑧 ∈ 𝑓 −1 (𝑦). Let us fix a 𝑡0 ∈ ℝ with
𝑝(𝑡0 ) = 𝑧. When we walk from 𝑡0 to 𝑡0 + 1 we need to count how many preimages of 𝑦
we collect along the way, with their orientation (!).

∙ We start with the case 𝑞 = 0, i.e., 𝑔(𝑡 + 1) = 𝑔(𝑡). It will actually teach us the key ideas
we need to remember from this proof.
We need to count how often 𝑔(𝑠) = 𝑔(𝑡0 ) with 𝑑𝑔𝑠 = 𝑔 ′ (𝑠) > 0 and how often 𝑔(𝑠) = 𝑔(𝑡0 )
with 𝑑𝑔𝑠 = 𝑔 ′ (𝑠) < 0. Note that since 𝑦 is regular, 𝑑𝑔𝑠 is always ≠ 0 for such an 𝑠.
Since 𝑔 is a smooth function ℝ → ℝ, this is now just an exercise in Calculus. Using the
periodicity of 𝑔, i.e., that 𝑔 ′ (𝑡0 ) must have the same sign as 𝑔 ′ (𝑡0 + 1), we see that there
are exactly as many points 𝑠 with 𝑔(𝑠) = 𝑔(𝑡0 ) and 𝑑𝑔𝑠 = 𝑔 ′ (𝑠) > 0 as there are points
with 𝑔(𝑠) = 𝑔(𝑡0 ) and 𝑑𝑔𝑠 = 𝑔 ′ (𝑠) < 0. Thus deg(𝑓 ) = 0. See Figure 15.2.

Figure 15.2: We count the intersection points with the sign of the derivative. If 𝑞 = 0, there are
as many crossing points from below with positive derivative as there are crossing points from
above with negative derivative.

∙ Now assume 𝑞 > 0, and 𝑔(𝑡 + 1) = 𝑔(𝑡) + 𝑞.

Again, we walk from 𝑡0 to 𝑡0 + 1 and sum up the orientation numbers of all the preimages
of 𝑦 that we collect along the way. This corresponds to counting how often we have
𝑔(𝑠) = 𝑔(𝑡0 ) + 𝑖 for some 𝑖 = 0, 1, … , 𝑞 − 1 and 𝑠 ∈ [𝑡0 , 𝑡0 + 1].
316 15.2. Hopf Degree Theorem in dimension one
Let us look at one interval [𝑔(𝑡0 ) + 𝑖, 𝑔(𝑡0 ) + 𝑖 + 1] at a time. We would like to know how
many 𝑠 ∈ [𝑡0 , 𝑡0 + 1] are sent to either 𝑔(𝑡0 ) + 𝑖 or 𝑔(𝑡0 ) + 𝑖 + 1 together with the sign of
the derivative.
Therefore we look at the preimage 𝑔 −1 ([𝑔(𝑡0 ) + 𝑖, 𝑔(𝑡0 ) + 𝑖 + 1]). This set is a disjoint
union of closed intervals. For each of these intervals the start and endpoints are sent to
either 𝑔(𝑡0 ) + 𝑖 or 𝑔(𝑡0 ) + 𝑖 + 1.
Let us think of the graph of 𝑔 passing 𝑔(𝑡0 ) + 𝑖with a positive sign of the derivative as
going in with +1 and passing 𝑔(𝑡0 ) + 𝑖 + 1 with a positive sign of the derivative as going
out +1, and the other two alternatives as the ones with −1. Then we see that the graph
has to go in with +1 for a first time, and has to go out with +1 for a last time (since the
graph starts at 𝑔(𝑡0 ) ≤ 𝑔(𝑡0 ) + 𝑖 and ends at 𝑔(𝑡0 ) + 𝑞 ≥ 𝑔(𝑡0 ) + 𝑖 + 1). In between those
two points, the graph is going out with −1 as often as it goes in +1 and goes in with −1
as often as it goes out with +1. See Figure 15.3.

Figure 15.3: We count the intersection points with the sign of the derivative. If 𝑞 > 0, there
are exactly 𝑞 more crossing points from below with positive derivative than there are crossing
points from above with negative derivative.

Thus in total the orientation numbers for 𝑔 −1 ([𝑔(𝑡0 )+𝑖, 𝑔(𝑡0 )+𝑖+1]) add up to +2. Repeating
this for all 𝑖 = 0, 1, … , 𝑞 − 1 gives a sum of orientation numbers equal to 𝑞, since we have to
account for that we counted the inner points twice.

Since the sum of orientation numbers of 𝑓 equals the one of 𝑔, this shows deg(𝑓 ) = 𝑞.

∙ Finally, if 𝑞 < 0, the same argument works with signs and directions reversed.
Chapter 15. The Brouwer Degree 317
15.3 Exercises and more examples

15.3.1 The Degree

Exercise 15.1 Let 𝑎 ∶ 𝕊𝑘 → 𝕊𝑘 , 𝑎(𝑥) = −𝑥, be the antipodal map on 𝕊𝑘 .

(a) Show that 𝑎 has degree (−1)𝑘+1 .

Hint: Consider the reflections

𝑟𝑖 (𝑥1 , … , 𝑥𝑘+1 ) = (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑘+1 ).

Recall that reflections are diffeomorphisms which reverse orientations.

(b) Show that 𝑎 is homotopic to the identity if and only if 𝑘 is odd.

Hint: Recall and use that we showed in a previous exercise that 𝑎 is homotopic to
the identity if 𝑘 is odd.

Aside: Note that we would not have been able to make the only if -conclusion with
the mod 2-degree. For in ℤ∕2, we cannot distinguish between 1 and −1.

Exercise 15.2 Let 𝑋 ⊂ ℝ𝑁 be a smooth manifold. Recall that a vector field on 𝑋

is a smooth section of 𝜋 ∶ 𝑇 (𝑋) → 𝑋, i.e., a smooth map 𝜎 ∶ 𝑋 → 𝑇 (𝑋) such that
𝜋◦𝜎 = Id𝑋 . An equivalent way to describe such a section is to give a map 𝑠 ∶ 𝑋 → ℝ𝑁
such that 𝑠(𝑥) ∈ 𝑇𝑥 (𝑋) ⊂ ℝ𝑁 for all 𝑥 (with corresponding 𝜎(𝑥) = (𝑥, 𝑠(𝑥))). A point
𝑥 ∈ 𝑋 is a zero of the vector field 𝜎 if 𝜎(𝑥) = (𝑥, 0) or equivalently 𝑠(𝑥) = 0.
Prove the following important theorem:
The 𝑛-dimensional sphere 𝕊𝑛 admits a vector field 𝑣 without zeros, i.e., 𝑣(𝑥) ≠ 0 for
all 𝑥 ∈ 𝕊𝑛 , if and only if 𝑛 is odd.
Hint: Note that we have proven most of the statement on a previous exercise set. You
can assume those results. What is new is the only if -part for which we need the improved
degree. To show this, write the antipodal map as a composition of reflections and use the
results from a previous exercise set on vector fields.

Exercise 15.3 Use the degree to prove the Fundament Theorem of Algebra: Given a
monic complex polynomial

𝑝(𝑧) = 𝑧𝑚 + 𝑎1 𝑧𝑚−1 + ⋯ + 𝑎𝑚 .

Show that there exists a 𝑤 ∈ ℂ such that 𝑝(𝑤) = 0.

Exercise 15.4 Explain why the degree cannot be used to prove that there is a zero in ℝ
for every monic real polynomial.
318 15.3. Exercises and more examples

Exercise 15.5 Let

𝑝(𝑧) = 𝑧𝑚 + 𝑎1 𝑧𝑚−1 + ⋯ + 𝑎𝑚
be a monic complex polynomial. At every point 𝑧0 ∈ ℂ, there is an integer 𝑚 ≥ 0 such
that we can factor
𝑝(𝑧) = (𝑧 − 𝑧0 )𝑚 𝑞(𝑧) with 𝑞(𝑧0 ) ≠ 0.
Note that 𝑧0 is a zero of 𝑝 if and only if 𝑚 > 0. In this case we call 𝑚 the multiplicity of
the zero 𝑧0 . In this exercise we relate multiplicities and degrees as follows:
Let 𝔻0 ⊂ ℂ be a disk with center 𝑧0 and radius 𝑟 small enough such that there is
no other zero of 𝑝 in 𝔻0 except 𝑧0 . Recall that we showed that this is possible when we
discussed Milnor’s proof of the FTA in Section 4.4.

(a) Check that the formula 𝑔(𝑧) = 𝑧0 + 𝑟𝑧 defines an orientation-preserving diffeo-

morphism 𝕊1 → 𝜕𝔻0 where 𝔻0 inherits an orientation from ℂ = ℝ2 .
𝑝 𝑝◦𝑔
(b) Explain why |𝑝|
∶ 𝜕𝔻0 → 𝕊1 and |𝑝◦𝑔|
∶ 𝕊1 → 𝕊1 have the same degree.
𝑝◦𝑔
(c) Find a smooth homotopy ℎ𝑡 ∶ 𝕊1 → 𝕊1 between ℎ1 = |𝑝◦𝑔|
and ℎ0 (𝑧) = 𝑐 ⋅ 𝑧𝑚 with
𝑞(𝑧𝑖 )
𝑐 being the constant 𝑐 = |𝑞(𝑧𝑖 )|
.
( )
𝑝
(d) Conclude that deg |𝑝|
= 𝑚.

𝑓 𝑔
Exercise 15.6 Let 𝑋 ←←←→ ← 𝑍 be a sequence of smooth maps between manifolds
← 𝑌 ←←→
with 𝑋 and 𝑌 compact, 𝑌 and 𝑍 connected. Assume that all three manifolds are oriented
and boundaryless and dim 𝑋 = dim 𝑌 = dim 𝑍. Show that

deg(𝑔◦𝑓 ) = deg(𝑔) ⋅ deg(𝑓 ).

𝑓
Exercise 15.7 Let 𝕊𝑘 ←←←→
← 𝕊𝑘 be a smooth map. Show that if deg(𝑓 ) ≠ (−1)𝑘+1 , then 𝑓
must have a fixed point.
Hint: Assume that 𝑓 had no fixed point. Show that then the composition of 𝑓 with
the antipodal map would be homotopic to the identity map.

Exercise 15.8 Let 𝑘 ≥ 1 be an odd integer. Let 𝑞 ∶ 𝕊𝑘 → ℝP𝑘 , 𝑥 → [𝑥], be the

canonical projection. Recall from Theorem 14.15 that ℝP𝑘 is compact and orientable so
that deg(𝑞) is defined. Show that the degree of 𝑞 is ±2, and that it is +2 if the orientation
on ℝP𝑘 is chosen such that 𝑞 preserves orientation.
Hint: Use that 𝑞 does not change after composition with the antipodal map.

𝑓
← 𝕊𝑘 be a smooth map. In this exercise we prove the following
Exercise 15.9 Let 𝕊𝑘 ←←←→
statement: if deg(𝑓 ) is odd, then 𝑓 must send some pair of antipodal points to antipodal
points.
Chapter 15. The Brouwer Degree 319

To prove the claim we assume that 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 sends no pair of antipodal points to

antipodal points. We will show that this implies that deg(𝑓 ) must be even.

(a) Assume that 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 sends no pair of antipodal points to antipodal points.

Show that then 𝑓 and 𝑓 ◦𝑎 are both smoothly homotopic to the map 𝑔 ∶ 𝕊𝑘 → 𝕊𝑘
defined by 𝑔(𝑥) ∶= 21 (𝑓 (𝑥)+𝑓 (−𝑥)), and deduce that deg(𝑓 ) = deg(𝑓 ◦) = deg(𝑔).

(b) Now further assume that 𝑘 is even. Show that this implies deg(𝑔) = 0.

(c) Now assume that 𝑘 is odd. Use Exercise 15.6 and Exercise 15.8 to deduce that this
implies deg(𝑔) is even.

(d) Deduce the claim and thereby prove Theorem 15.16.

320 15.4. Linking number and the Hopf invariant via exercises
15.4 Linking number and the Hopf invariant via exercises

In the following exercises, we will study some basic properties of the Hopf invariant. Recall
that we have already defined the linking number and the Hopf invariant in a mod 2-version in
Section 11.3. Now we improve the constructions to get ℤ-valued invariants. In particular, we
show that the Hopf invariant of the Hopf fibration equals one. This is a much stronger result
than what we proved previously in Theorem 11.27.

As in Definition 11.19 we define the linking number of submanifolds in Euclidean space as

follows: For 𝑘 ≥ 1, let 𝑋, 𝑌 ⊂ ℝ𝑘+1 be two disjoint smooth manifolds. The linking map
𝜆 ∶ 𝑋 × 𝑌 → 𝕊𝑘
is defined by
𝑥−𝑦
𝜆(𝑥, 𝑦) = .
|𝑥 − 𝑦|
Note that, since 𝑋 and 𝑌 are disjoint, this map is well-defined and smooth. Now we assume
that 𝑋 and 𝑌 are compact, oriented, and without boundary, of dimensions dim 𝑋 = 𝑚 and
dim 𝑌 = 𝑛 such that 𝑚 + 𝑛 = 𝑘. Then the linking number 𝐿(𝑋, 𝑌 ) of 𝑋 and 𝑌 is defined to be
the degree of 𝜆, i.e.,
𝐿(𝑋, 𝑌 ) ∶= deg(𝜆).

Figure 15.4: The red circle is the boundary of a compact manifold 𝐷. In both cases the green
curve intersects the manifold 𝐷. The linking number detects the intersection. The circles on
the left-hand side are linked and cannot be moved apart. However, multiple intersections may
occur. The curves on the right-hand side are not linked and can be moved. The linking number
detects this by counting the linking points with signs.

Exercise 15.10 (a) Show that

𝐿(𝑌 , 𝑋) = (−1)(𝑚+1)(𝑛+1) 𝐿(𝑋, 𝑌 ).

Hint: Check what happens with the orientation numbers at points when we switch
the order of 𝑋 and 𝑌 , and think of our computation of the degree of the antipodal
map on 𝕊𝑘 .
(b) Assume that 𝑋 is the boundary of an oriented manifold 𝑊 which is disjoint from
𝑌 . Show that this implies 𝐿(𝑋, 𝑌 ) = 0.
Hint: Use the Boundary Theorem 15.2 for degrees.
Chapter 15. The Brouwer Degree 321
As in Definition 11.21, we extend the definition of the linking number to submanifolds 𝑋, 𝑌
in 𝕊𝑘+1 : Assume 𝑋 and 𝑌 are compact, oriented and boundaryless, and dim 𝑋 + dim 𝑌 = 𝑘.
Since the sphere is connected and 𝑋 and 𝑌 are closed subsets, there must be a point 𝑝 which
is not contained in either 𝑋 or 𝑌 . We identity 𝕊𝑘+1 ⧵ {𝑝} with ℝ𝑘+1 via the diffeomorphism
defined by stereographic projection from 𝑝. Then we consider 𝑋 and 𝑌 as submanifolds of
ℝ𝑘+1 and define the linking number 𝐿(𝑋, 𝑌 ) as above.

For 𝑛 ≥ 1, consider a smooth map 𝑓 ∶ 𝕊2𝑛−1 → 𝕊𝑛 . Let 𝑤 ≠ 𝑧 ∈ 𝕊𝑛 be two regular values

for 𝑓 . Then 𝑓 −1 (𝑤) and
( 𝑓−1 (𝑧) are
−1 compact,
) oriented, boundaryless submanifolds of 𝕊𝑛 and
their linking number 𝐿 𝑓 (𝑤), 𝑓 (𝑧) is defined.
−1

The number ( )
𝐻(𝑓 ) ∶= 𝐿 𝑓 −1 (𝑤), 𝑓 −1 (𝑧)
is called the Hopf invariant of 𝑓 . This is a famous invariant that played a crucial role in the
development of mathematics. As in Section 11.3.2 one can show that 𝐻(𝑓 ) does not depend on
the choice of 𝑤 and 𝑧 and only depends on the homotopy class of 𝑓 . We skip this verification
here. Denoting by 𝜋2𝑛−1 (𝕊𝑛 ) the (2𝑛−1)-homotopy group of 𝕊𝑛 , we can view the Hopf invariant
as a map
𝐻 ∶ 𝜋2𝑛−1 (𝕊𝑛 ) → ℤ.

Exercise 15.11 We are going to study some of the basic properties of the Hopf invariant
in this exercise:

(a) Show that if 𝑛 is odd, then 𝐻(𝑓 ) = 0.

(b) Let 𝑔 ∶ 𝕊𝑛 → 𝕊𝑛 be a smooth map. Consider the composition

𝑓 𝑔
𝕊2𝑛−1 ←←←→
← 𝕊𝑛 ←←→
← 𝕊𝑛 .

Show that
𝐻(𝑔◦𝑓 ) = 𝐻(𝑓 ) ⋅ deg(𝑔)2 .

Now we are going to compute the Hopf invariant for the Hopf fibration 𝜋 ∶ 𝕊3 → 𝕊2 . We
have met 𝜋 several times before and have solved several of the following problems in the main
text and previous exercises. We collect and reprove all the results for computing 𝐻(𝜋). We
recommend that you solve each of the tasks again to get more practice. The calculation of
𝐻(𝜋) = 1 is a key new step which goes beyond the computation modulo 2.

Exercise 15.12 We consider 𝕊3 as a subset of ℂ2 , i.e., 𝕊3 = {(𝑧0 , 𝑧1 ) ∈ ℂ2 ∶ |𝑧0 |2 +

|𝑧1 |2 = 1}, and 𝕊2 as a subset of ℂ × ℝ, i.e., 𝕊2 = {(𝑧, 𝑥) ∈ ℂ × ℝ ∶ |𝑧|2 + 𝑥2 = 1}.
Then the Hopf fibration 𝜋 is the map 𝕊3 → 𝕊2 given by
( )
𝜋(𝑧0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2 .

(a) Check that this actually defines a map 𝕊3 → 𝕊2 .

(b) Show that 𝜋(𝑧0 , 𝑧1 ) = 𝜋(𝑤0 , 𝑤1 ) if and only if there is a complex number 𝛼 with
|𝛼|2 = 𝛼 𝛼̄ = 1 such that (𝑤0 , 𝑤1 ) = (𝛼𝑧0 , 𝛼𝑧1 ).
322 15.4. Linking number and the Hopf invariant via exercises

(c) Show that, for every point 𝑝 ∈ 𝕊2 , the fiber 𝜋 −1 (𝑝) is diffeomorphic to 𝕊1 .

(d) Show that each point in 𝕊2 is a regular value for 𝜋.

(e) Consider 𝑎 = (0, 0, 1) and 𝑏 = (0, 1, 0) on 𝕊2 ⊂ ℝ3 ≅ ℂ × ℝ. Determine the fibers

𝜋 −1 (𝑎) and 𝜋 −1 (𝑏) together with their orientations as preimages under 𝜋.
Hint: We have worked out the orientation of 𝜋 −1 (𝑎) in Section 14.6 and Exer-
cise 14.8. Follow the same outline to figure it out for 𝜋 −1 (𝑏).

(f) Show that 𝐻(𝜋) = 1.

Hint: You may want to choose 𝑎 and 𝑏 as regular values for 𝜋 and compute 𝐻(𝜋)
as the degree of 𝜆 ∶ 𝑆𝑎 × 𝑆𝑏 → 𝕊2 with 𝑆𝑎 = 𝜙−1 𝑁
(𝜋 −1 (𝑎)) and 𝑆𝑏 = 𝜙−1
𝑁
(𝜋 −1 (𝑏))
where 𝜙−1𝑁
denotes the stereographic projection from 𝑁 = (0, 0, 0, 1) ∈ 𝕊3 . To
calculate deg(𝜆), pick a regular value for 𝜆, say 𝑝 = (1, 0, 0). Then, for a point
(𝐯, 𝐰) with 𝜆(𝐯, 𝐰) = 𝑝, compute 𝑑𝜆(𝐯,𝐰) ∶ 𝑇𝐯 𝑆𝑎 × 𝑇𝐰 𝑆𝑏 → 𝑇𝑝 𝕊2 . The final step is
to check if 𝑑𝜆(𝐯,𝐰) preserves or reverses orientations. To do this, use the previous
point to determine the orientations of 𝑆𝑎 and 𝑆𝑏 .
16. Hopf Degree Theorem

We proved previously that there is exactly one homotopy class of maps 𝕊1 → 𝕊1 for every
integer 𝑛 ∈ ℤ. By our classification of one-manifolds, we can read this also as follows:

For every compact, connected, boundaryless one-manifold 𝑋, there is exactly one homo-
topy class of maps 𝑋 → 𝕊1 for every integer 𝑛 ∈ ℤ.

Now we are going to generalize this result to higher dimensions. Recall that the degree of
a map 𝑓 ∶ 𝑋 → 𝕊𝑘 as in the theorem is defined as
∑
deg(𝑓 ) = sign (𝑑𝑓𝑥 )
𝑥∈𝑓 −1 (𝑦)

where 𝑦 is a regular value of 𝑓 and sign (𝑑𝑓𝑥 ) is +1 if 𝑑𝑓𝑥 preserves orientations and −1 if 𝑑𝑓𝑥
reverses orientations. We will refer to this sign rule as our usual orientation convention.

Theorem 16.1 (Hopf Degree Theorem) Let 𝑋 be a compact, connected, oriented

smooth 𝑘-manifold without boundary. Then two continuous maps 𝑋 → 𝕊𝑘 are homo-
topic if and only if they have the same degree. In other words, the degree defines an
injective map [𝑋, 𝕊𝑘 ] → ℤ.

Together with Theorem 15.13 and since every continuous map between spheres is homo-
topic to a smooth map we then get:

Corollary 16.2 (Homotopy groups of spheres - injectivity) Let 𝜋𝑘 (𝕊𝑘 ) denote the 𝑘th
homotopy group of 𝕊𝑘 . The degree defines a bijection

deg ∶ 𝜋𝑘 (𝕊𝑘 ) → ℤ.

Remark 16.3 Note that the situation is similar, though different for non-orientable
manifolds: Two maps of a compact, connected, non-orientable, boundaryless 𝑘-
manifold 𝑋 to 𝕊𝑘 are homotopic if and only if they have the same degree modulo
2.

16.1 Strategy for proving Hopf’s theorem

Now we start our march towards a proof Hopf’s theorem. We will follow the guideline of
Guillemin-Pollack [5]. But it is worth noting that there are many different ways to prove this
theorem and we will present Pontryagin’s proof as presented in Milnor’s book [13] in ??.

323
324 16.1. Strategy for proving Hopf’s theorem

∙ (Strategy for the proof) Assume given two maps 𝑓0 and 𝑓1 from 𝑋 to 𝕊𝑘 .

∙ Set 𝑊 ∶= 𝑋 × [0, 1], define 𝑓 ∶ 𝜕𝑊 → 𝕊𝑘 by 𝑓 ∶= 𝑓0 on 𝑋 × {0} and 𝑓 ∶= 𝑓1

on 𝑋 × {1}. Then deg(𝑓 ) = deg(𝑓1 ) − deg(𝑓0 ) = 0, and a homotopy between 𝑓0
and 𝑓1 is a global extension of 𝑓 to 𝑊 .

∙ Show the Extension Theorem 16.13: 𝑓 ∶ 𝜕𝑊 → 𝕊𝑘 has a global extension

𝑊 → 𝕊𝑘 if and only if deg(𝑓 ) = 0, for any compact, connected, oriented 𝑘 + 1-
manifold 𝑊 . We knew already: existence of global extensions ⇒ deg(𝑓 ) = 0.

∙ For the proof of the Extension Theorem 16.13, use the Isotopy Lemma 11.3 to
move 𝑊 inside some ball 𝔹 ⊂ ℝ𝑘+1 with Int(𝑊 ) ⊂ 𝔹. This reduces to checking
an extension statement on balls and spheres.

∙ Use winding numbers to show that a map which is homotopic to a constant

map on the boundary of a ball 𝔹 extends to all of 𝔹.

∙ Show the Special Case 𝑋 = 𝕊𝑘 : For 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 ,

deg(𝑓 ) = 0 ⇒ 𝑓 ∼ constant map.

This follows by induction on the dimension 𝑘 of 𝕊𝑘 . We have shown previously

that 𝑓 , 𝑔 ∶ 𝕊1 → 𝕊1 are homotopic if and only if deg(𝑓 ) = deg(𝑔).
The induction step is actually a zigzag argument using winding numbers. The
Isotopy Lemma is frequently used to move points into appropriate open neigh-
borhoods and balls.

In order to make this strategy work, we need to prove a series of technical results. This
will occupy the rest of the chapter. Two main technical ingredients are isotopies which allow
to move points, and winding numbers which help us calculating degrees. We will also need
the following version of the Isotopy Lemma:

Lemma 16.4 (Linear Isotopy Lemma) Suppose that 𝐸 is a linear isomorphism of

ℝ𝑘 that preserves orientation. Then there exists a homotopy 𝐸𝑡 consisting of linear
isomorphisms such that 𝐸0 = 𝐸 and 𝐸1 is the identity. If 𝐸 reverses orientation, then
there exists a homotopy 𝐸𝑡 consisting of linear isomorphisms such that 𝐸0 = 𝐸 and 𝐸1
is the reflection map

𝐸(𝑥1 , 𝑥2 , … , 𝑥𝑘 ) = (−𝑥1 , 𝑥2 , … , 𝑥𝑘 ).

Proof: First we remark that it suffices to deal with the case that 𝐸 preserves orientations.
For if 𝐸 is orientation reversing, then 𝑟1 ◦𝐸 preserves orientations. Then if there is a homotopy
𝐹 between 𝑟1 ◦𝐸 and Id, then, after composing all maps with 𝑟1 , 𝑟1 ◦𝐹 is a homotopy between
Chapter 16. Hopf Degree Theorem 325
𝐸 = 𝑟1 ◦𝑟1 ◦𝐸 and 𝑟1 .

So let 𝐸 be a linear isomorphism of ℝ𝑘 that preserves orientations. The proof is by in-

duction on the dimension 𝑘. We need to check two initial cases:

First, let 𝑘 = 1. Then 𝐸 ∶ ℝ → ℝ is given by multiplication by a real number 𝜆 > 0. Then

𝐸𝑡 = 𝑡 ⋅ 1 + (1 − 𝑡) ⋅ 𝜆 is a homotopy between 𝐸 = 𝜆 and Id = 1. Note that each 𝐸𝑡 is nonzero
and therefore a linear isomorphism.

Now let 𝑘 = 2 and assume that 𝐸 has only complex eigenvalues. Then 𝐸𝑡 = 𝑡𝐸 + (1 − 𝑡)Id
is a linear homotopy between Id and 𝐸. Moreover, each ( 𝐸𝑡 is)a linear isomorphism. To show
𝑎 𝑏
this we show that det(𝐸𝑡 ) ≠ 0 for all 𝑡 ∈ [0, 1]. If 𝐸 = , then we get
𝑐 𝑑

det(𝐸𝑡 ) = (𝑡(𝑎 − 1) + 1)(𝑡(𝑑 − 1) + 1) − 𝑡2 𝑏𝑐

= 𝑡2 (𝑎 − 1)(𝑑 − 1) + 𝑡(𝑎 + 𝑑 − 2) + 1 − 𝑡2 𝑏𝑐
= 𝑡2 (𝑎𝑑 − 𝑏𝑐 − 𝑎 − 𝑑 + 1) + 𝑡(𝑎 + 𝑑 − 2) + 1.

The discriminant of this quadratic equation in 𝑡 is

(𝑎 + 𝑑 − 2)2 − 4(𝑎𝑑 − 𝑏𝑐 − 𝑎 − 𝑑 + 1)
=(𝑎 + 𝑑)2 − 4(𝑎 + 𝑑) + 4 − 4(𝑎𝑑 − 𝑏𝑐) + 4(𝑎 + 𝑑) − 4
=(𝑎 + 𝑑)2 − 4(𝑎𝑑 − 𝑏𝑐).

But this is exactly the discriminant of the equation

𝑡2 + 𝑡(𝑎 + 𝑑) − (𝑎𝑑 − 𝑏𝑐) = 0

which is the characteristic polynomial in 𝑡 of 𝐸. By assumption, this polynomial has only

complex roots, i.e., its discriminant is negative. Hence there is no real 𝑡 such that det(𝐸𝑡 ) = 0.

Now we show the induction step: So assume 𝑘 ≥ 2 and the assertion to be true in all
dimensions < 𝑘. Then 𝐸 has either at least one real eigenvalue or at least one complex eigen-
value. Let 𝑉 ⊂ ℝ𝑘 be the corresponding eigenspace, which is either one- or two-dimensional.
Then 𝐸 maps 𝑉 into itself. Hence ℝ𝑘 splits into a direct sum ℝ𝑘 = 𝑉 ⊕ 𝑊 . By choosing a
basis of ℝ𝑘 consisting of a basis of 𝑉 and one for 𝑊 , we can represent 𝐸 as a matrix of the
form
( )
𝐴 𝐵
𝐸= .
0 𝐶

Here 𝐴 is either a 1 × 1- or a 2 × 2-matrix determined by the type of the eigenvalues. Then we

can define a linear homotopy 𝐸𝑡 by
( )
𝐴 𝑡𝐵
𝐸𝑡 = .
0 𝐶

Since 𝐸 is a linear isomorphism and the determinant is multiplicative, we have

0 ≠ det(𝐸) = det(𝐴) det(𝐶) = det(𝐸𝑡 ).

326 16.1. Strategy for proving Hopf’s theorem
Thus 𝐸𝑡 is also a linear isomorphism for every 𝑡. For 𝑡 = 0, we see that 𝐸0 maps 𝑉 to 𝑉 by 𝐴
and 𝑊 to 𝑊 by 𝐶. Since dim 𝑊 is strictly less than 𝑘, we can apply the induction hypothesis
to 𝐶 and 𝑊 and the initial cases to 𝐴 and 𝑉 , respectively. Hence we have a homotopy𝐶𝑡
consisting of linear isomorphisms between 𝐶 and the identity and a homotopy 𝐴𝑡 between 𝐴
and the identity. Then
( )
𝐴𝑡 𝑡𝐵
0 𝐶𝑡

is a homotopy consisting of linear homotopies between 𝐸 and the identity map of ℝ𝑘 .

16.1.1 Winding numbers revisited

As for many results on maps between spheres, the winding number is useful concept. We used
it before with values modulo 2. Now we need an integral version:

Definition 16.5 (Integer-valued winding numbers) Let 𝑋 be a compact, oriented

𝑘-dimensional smooth manifold, and let 𝑓 ∶ 𝑋 → ℝ𝑘+1 be a smooth map. The winding
number of 𝑓 , denoted 𝑊 (𝑓 , 𝑧), around any point 𝑧 ∈ ℝ𝑘+1 ⧵ 𝑓 (𝑋) is defined as the
degree of the map

𝑓 (𝑥) − 𝑧
𝑢 ∶ 𝑋 → 𝕊𝑘 , 𝑥 → .
|𝑓 (𝑥) − 𝑧|
As a formula:

𝑊 (𝑓 , 𝑧) = deg(𝑢).

The winding number will be the main tool in the proof of Hopf’s theorem. In order to
exploit it effectively, we investigate some of its properties:

Lemma 16.6 (Step 1) Let 𝑓 ∶ 𝑈 → ℝ𝑘 be a smooth map defined on an open subset

𝑈 of ℝ𝑘 , and let 𝑥 be a regular point, with 𝑓 (𝑥) = 𝑧. Let 𝔹 be a sufficiently small closed
ball centred at 𝑥, and define 𝜕𝑓 ∶ 𝜕𝔹 → ℝ𝑘 to be the restriction of 𝑓 to the boundary
of 𝔹. Then we have
{
+1 if 𝑓 preserves orientation at 𝑥,
𝑊 (𝜕𝑓 , 𝑧) =
−1 if 𝑓 reverses orientation at 𝑥.

Proof: After possibly translating things, we can assume 𝑥 = 0 = 𝑧, which keeps the
notation simpler. We set 𝐴 = 𝑑𝑓0 . We are going to show that 𝑊 (𝐴, 0) can be used to calculate
𝑊 (𝜕𝑓 , 0). This will follow if we show that we can choose 𝔹 small enough such that there is a
homotopy 𝐹𝑡 ∶ 𝜕𝔹 × [0, 1] → 𝕊𝑘−1 between 𝐴𝑥∕|𝐴𝑥| and 𝜕𝑓 (𝑥)∕|𝜕𝑓 (𝑥)|. For then
( ) ( )
𝜕𝑓 (𝑥) 𝐴𝑥
𝑊 (𝜕𝑓 , 0) = deg = deg = 𝑊 (𝐴, 0).
|𝜕𝑓 (𝑥)| |𝐴𝑥|
Chapter 16. Hopf Degree Theorem 327
Now we are going to construct the homotopy 𝐹𝑡 : By Taylor theory, we can write

𝑓 (𝑥) = 𝐴𝑥 + 𝜀(𝑥), where 𝜀(𝑥)∕|𝑥| → 0 when 𝑥 → 0. (16.1)

We define

𝑓𝑡 (𝑥) = 𝐴𝑥 + 𝑡𝜀(𝑥) for 𝑡 ∈ [0, 1].

Then, 𝑓𝑡 is a homotopy between 𝑓0 (𝑥) = 𝐴𝑥 and 𝑓1 (𝑥) = 𝑓 (𝑥).

Since 𝑥 = 0 is a regular point, we know that 𝐴 is an isomorphism. Hence the image of the
unit ball in ℝ𝑘 under 𝐴 strictly contains a closed ball of some radius 𝑟 > 0. Since every linear
isomorphism is a diffeomorphism, we also know that 𝐴 maps boundaries to boundaries, i.e.,
𝕊𝑘−1 to the boundary of the closed ball of radius 𝑟. Hence

|𝐴𝑥| > 𝑟 for all 𝑥 ∈ 𝕊𝑘−1 .

As a consequence,

|𝐴(𝑥∕|𝑥|)| > 𝑐 and thus |𝐴𝑥| > |𝑟𝑥| for all 𝑥 ∈ ℝ𝑘 ⧵ {0}.

Now we use (16.1). Since 𝜀(𝑥)∕|𝑥| → 0 as 𝑥 → 0, we can choose a ball 𝔹 small enough
such that
𝑟
𝜀(𝑥)∕|𝑥| < for all 𝑥 ∈ 𝜕𝔹.
2
Then we have
𝑟 𝑟
|𝑓𝑡 (𝑥)| = |𝐴𝑥| − 𝑡|𝜀(𝑥)| > 𝑟|𝑥| − |𝑥| = |𝑥|,
2 2
𝑖.𝑒., |𝑓𝑡 (𝑥)| > 0 for all 𝑥 ∈ 𝜕𝔹.

Hence we can define the desired homotopy 𝐹𝑡 by

𝑓𝑡 (𝑥)
𝐹𝑡 ∶ 𝜕𝔹 × [0, 1] → 𝕊𝑘 , 𝑥 → .
|𝑓𝑡 (𝑥)|

Now we compute 𝑊 (𝐴, 0). Therefor we apply the Linear Isotopy Lemma 16.4 and get
that 𝐴 is homotopic to the identity if it preserves orientations, and homotopic to the reflection
map (𝑥1 , 𝑥2 , … , 𝑥𝑘 ) → (−𝑥1 , 𝑥2 , … , 𝑥𝑘 ) if it reverses orientations. In the former case, we have
𝑊 (𝐴, 0) = +1, and in the latter case 𝑊 (𝐴, 0) = −1.

This result determines how local diffeomorphisms can wind. Now we are going to use this
information to count preimages:

Lemma 16.7 (Step 2) Let 𝑓 ∶ 𝔹 → ℝ𝑘 be a smooth map defined on some closed ball
𝔹 in ℝ𝑘 . Suppose that 𝑧 is a regular value of 𝑓 that has no preimages on the boundary
sphere 𝜕𝔹, and let 𝜕𝑓 ∶ 𝜕𝔹 → ℝ𝑘 be its restriction to the boundary. Then the number
of preimages of 𝑧, counted with our usual orientation convention, equals the winding
number 𝑊 (𝜕𝑓 , 𝑧).
328 16.2. The Special Case
Proof: By the Stack of Records Theorem 4.14, we know that 𝑓 −1 (𝑧) is a finite set {𝑥1 , … , 𝑥𝑛 },
and we can choose disjoint balls 𝔹𝑖 around each 𝑥𝑖 . Since 𝑓 −1 (𝑧) is disjoint from 𝜕𝔹 by as-
sumption, we can shrink these balls such that 𝔹𝑖 ∩ 𝜕𝔹 = ∅ and so that each 𝔹𝑖 is sufficiently
small so that Step 1 can be applied. Let 𝜕𝑓𝑖 = 𝑓|𝜕𝔹𝑖 . Then Step 1, i.e., Lemma 16.6, implies
that the number of preimage points, counted with our usual orientation convention, equals
∑𝑛
𝑖=1 𝑊 (𝜕𝑓𝑖 , 𝑧). Let 𝔹 ∶= 𝔹 ⧵ ∪𝑖 𝔹𝑖 and consider the map
′

𝑓 (𝑥) − 𝑧
𝑢 ∶ 𝜕𝔹 → 𝕊𝑘−1 , 𝑥 → .
|𝑓 (𝑥) − 𝑧|
Since 𝑓 (𝑥) ≠ 𝑧 on 𝔹′ , this map extends to all of 𝔹′ . This implies

𝑊 (𝑓|𝜕𝔹′ , 𝑧) = deg(𝑢) = 0.

The orientations of the boundaries are related by

𝜕𝔹′ = 𝜕𝔹 ∪𝑛𝑖=1 (−𝜕𝔹𝑖 ).

This implies
∑
𝑛
𝑊 (𝑓|𝜕𝔹′ , 𝑧) = 𝑊 (𝜕𝑓 , 𝑧) − 𝑊 (𝜕𝑓𝑖 , 𝑧).
𝑖=1
∑𝑛
Hence in total we get 𝑊 (𝜕𝑓 , 𝑧) = 𝑖=1 𝑊 (𝜕𝑓𝑖 , 𝑧).

Lemma 16.8 (Step 3) Let 𝔹 be a closed ball in ℝ𝑘 , and let 𝑓 ∶ ℝ𝑘 ⧵ Int(𝔹) → 𝑌 be a

smooth map defined outside the open ball Int(𝔹). Let 𝜕𝑓 ∶ 𝜕𝔹 → 𝑌 be the restriction
to the boundary. Assume that 𝜕𝑓 is homotopic to a constant map. Then 𝑓 extends
to a smooth map defined on all of ℝ𝑘 into 𝑌 .

Proof: For simplicity, we assume that 𝔹 is centred at 0. Then we can write every non-zero
point 𝑥 ∈ 𝔹 uniquely as 𝑥 = 𝑡𝑦 for some 𝑦 ∈ 𝜕𝔹 and some 𝑡 ∈ [0, 1]. By assumption, there is
a homotopy 𝑔𝑡 ∶ 𝜕𝔹 → 𝑌 with 𝑔1 = 𝜕𝑓 and 𝑔0 being a constant map.

Now we define the map 𝐹 ∶ ℝ𝑘 → 𝑌 by setting

{
𝑓 (𝑥) if 𝑥 ∈ ℝ𝑘 ⧵ Int(𝔹)
𝐹 (𝑥) =
𝑔𝑡 (𝑥) if 𝑥 ∈ 𝔹 and 𝑥 = 𝑡𝑦 for some 𝑦 ∈ 𝜕𝔹 and 𝑡 ∈ [0, 1].

Note that 𝐹 is well-defined on ℝ𝑘 ⧵ Int(𝔹), since 𝑓 and 𝑔𝑡 agree on 𝜕𝔹 = 𝔹 ∩ (ℝ𝑘 ⧵ Int(𝔹))

where we have 𝑓 = 𝜕𝑓 = 𝑔1 . Note also that 𝐹 (0) is well-defined as the constant value of 𝑔0 .

Now it remains to use smooth bump function to turn 𝐹 into a smooth homotopy.1

16.2 The Special Case

1
It is already smooth except, possibly, on 𝜕𝔹.
Chapter 16. Hopf Degree Theorem 329

Lemma 16.9 (The Special Case) Let 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 be a smooth map of degree zero.
Then 𝑓 is homotopic to a constant map.

Before we prove this case, we look at a consequence:

Theorem 16.10 (Winding number zero) Any smooth map 𝑓 ∶ 𝕊𝑘 → ℝ𝑘+1 ⧵ {0}
having winding number zero with respect to the origin is homotopic to a constant
map.

Proof of Theorem 16.10: By assumption, the degree of the map 𝑓 ∕|𝑓 | is zero. By the
special case, this implies that 𝑓 ∕|𝑓 | is homotopic to a constant map. But 𝑓 ∕|𝑓 | and 𝑓 are
homotopic via the homotopy

𝐹 ∶ 𝕊𝑘 × [0, 1] → ℝ𝑘+1 ⧵ {0}, (𝑥, 𝑡) → 𝑡𝑓 (𝑥) + (1 − 𝑡) ⋅ 𝑓 ∕|𝑓 |

Since homotopy is a transitive relation, 𝑓 is also homotopic to a constant map.

Proof of the special case Lemma 16.9:

The proof is by induction on the dimension 𝑘. We have established the case 𝑘 = 1 in

Theorem 15.18. So we assume the special case being true for 𝑘 − 1 and want to deduce it for
𝑘. We need to prove a lemma first:

Lemma 16.11 Let 𝑓 ∶ ℝ𝑘 → ℝ𝑘 be a smooth map with 0 as a regular value. Suppose

that 𝑓 −1 (0) is finite and that the sum of preimage points in 𝑓 −1 (0) is zero when counted
with the usual orientation convention. Assuming the special case in dimension 𝑘 − 1.
Then there exists a map 𝑔 ∶ ℝ𝑘 → ℝ𝑘 ⧵ {0} such that 𝑔 = 𝑓 outside a compact set. In
particular, the homotopy 𝑡𝑓 + (1 − 𝑡)𝑔 from 𝑔 to 𝑓 is constant outside this compact set.

Proof: Since 𝑓 −1 (0) is a finite, we can choose a ball 𝔹 centred at the origin with 𝑓 −1 (0) ⊂
Int(𝔹). By assumption, the sum of preimage points is zero when counted with the usual ori-
entation convention. By Step 2, i.e., Lemma 16.7, the map 𝜕𝑓 ∶ 𝜕𝔹 → ℝ𝑘 ⧵ {0} has winding
number zero. Since 𝜕𝔹 is diffeomorphic to 𝕊𝑘−1 , we may consider 𝜕𝑓 as a map from 𝕊𝑘−1 to
ℝ𝑘 ⧵ {0}.

Since we are assuming the special case being true in dimension 𝑘 − 1, we can apply its
consequence, i.e., Theorem 16.10, in that dimension. Thus, 𝜕𝑓 is homotopic to a constant map.
Hence
𝑓|ℝ𝑘 ⧵Int(𝔹) ∶ ℝ𝑘 ⧵ Int(𝔹) → ℝ𝑘 ⧵ {0}
is a map to which we can apply Step 3, i.e., Lemma 16.8. This implies that 𝑓 extends to a
smooth map 𝑔 ∶ ℝ𝑘 → ℝ𝑘 ⧵ {0} with 𝑓 = 𝑔 outside the compact space 𝔹.

Now we get back to the proof of the special case. So we are given a smooth map 𝑓 ∶ 𝕊𝑘 →
𝕊𝑘 with deg(𝑓 ) = 0.

The idea of the proof is to show that 𝑓 is homotopic to a map ℎ ∶ 𝕊𝑘 → 𝕊𝑘 ⧵ {𝑏}, where
𝑏 is some point in 𝕊𝑘 . But 𝕊𝑘 ⧵ {𝑏} is diffeomorphic to ℝ𝑘 via stereographic projection from
330 16.2. The Special Case
𝑏. Since ℝ𝑘 is contractible, this implies ℎ is homotopic to a constant map. Then 𝑓 is also
homotopic to a constant map.

So we need to show:

∙ Claim: 𝑓 is homotopic to a smooth map 𝑔 ∶ 𝕊𝑘 → 𝕊𝑘 ⧵ {𝑏}.

By Sard’s Theorem 7.1, we can choose distinct regular values 𝑎 and 𝑏 of 𝑓 . By the Stack of
Records Theorem 4.14, the preimage sets are finite, say 𝑓 −1 (𝑎) = {𝑎1 , … , 𝑎𝑛 } and 𝑓 −1 (𝑏) =
{𝑏1 , … , 𝑏𝑚 }.

Moreover, we can find an open neighborhood 𝑈 of 𝑎1 such that 𝑈 is diffeomorphic to ℝ𝑘

via a diffeomorphism 𝛼 ∶ ℝ𝑘 → 𝑈 and such that 𝑏𝑖 ∉ 𝑈 for all 𝑖 = 1, … , 𝑚.

Since 𝑘 > 1, we can apply the Isotopy Theorem 11.10, which is an extension of the Isotopy
Lemma 11.3, to the points {𝑎2 , … , 𝑎𝑛 } in 𝑌 ∶= 𝕊𝑘 ⧵ {𝑏} to get a diffeomorphism which is
isotopic to the identity, compactly supported, and moves the points 𝑎𝑖 into 𝑈 .

Since homotopy is a transitive relation, we can therefore assume that 𝑈 is an open neigh-
borhood of 𝑓 −1 (𝑎) with 𝑏 ∉ 𝑓 (𝑈 ).

Now let 𝛽 ∶ 𝕊𝑘 ⧵ {𝑏} → ℝ𝑘 be a diffeomorphism with 𝛽(𝑎) = 0. Then

𝛼 𝑓 𝛽
𝛽◦𝑓 ◦𝛼 ∶ ℝ𝑘 ←←→ ← 𝕊𝑘 ⧵ {𝑏} ←←→
← 𝑈 ←←←→ ← ℝ𝑘

is a smooth map from ℝ𝑘 to ℝ𝑘 . Since 𝑎 is a regular value of 𝑓 , 0 is a regular value of 𝛽◦𝑓 ◦𝛼.
Moreover, since 𝑓 −1 (𝑎) is finite, (𝛽◦𝑓 ◦𝛼)−1 (0) is finite as well.

Now we use the assumption deg(𝑓 ) = 0. For this means that the number of preimages of
𝑎 under 𝑓 is zero when counted with our usual orientation convention. Hence the number
of preimages of 0 under 𝛽◦𝑓 ◦𝛼 is zero when counted with the usual orientation convention.

Thus, we can apply Lemma 16.11 to 𝛽◦𝑓 ◦𝛼 ∶ ℝ𝑘 → ℝ𝑘 and get a map 𝑔 ∶ ℝ𝑘 → ℝ𝑘 ⧵ {0}
such that 𝑔 = 𝛽◦𝑓 ◦𝛼 outside a compact set 𝐵 and 𝑔 is homotopic to 𝛽◦𝑓 ◦𝛼 on ℝ𝑘 .

Since 𝛼 and 𝛽 are diffeomorphisms, this implies that 𝑓 is homotopic to 𝛽 −1 ◦𝑔◦𝛼 −1 as a

map from 𝑈 to 𝕊𝑘 ⧵ {𝑏}. As 𝑔 = 𝛽◦𝑓 ◦𝛼 outside 𝐵, we have

𝛽 −1 ◦𝑔◦𝛼 −1 = 𝑓 on 𝑈 ⧵ 𝛼 −1 (𝐵).

Thus, the map

ℎ ∶ 𝕊𝑘 → 𝕊𝑘 ⧵ {𝑏}

defined by setting
{
𝑓 on 𝕊𝑘 ⧵ 𝛼 −1 (𝐵)
ℎ=
𝛽 −1 ◦𝑔◦𝛼 −1 on 𝛼 −1 (𝐵)

is smooth, and ℎ is the desired map homotopic to 𝑓 . This proves the special case.
Chapter 16. Hopf Degree Theorem 331
16.3 The Extension Theorem for degree zero maps

Next we need to learn how to extend maps defined on the boundary to the whole manifold. This
is possible for maps of degree zero. This is an important result. We start with the simpler case
of maps to Euclidean space:

Lemma 16.12 (Extending maps to Euclidean spaces) Let 𝑊 be a compact smooth

manifold with boundary, and let 𝑓 ∶ 𝜕𝑊 → ℝ𝑘 be a smooth map. Then 𝑓 can be
extended to a globally defined map 𝐹 ∶ 𝑊 → ℝ𝑘 .

Proof: As always we assume that 𝑊 is a subset of some ℝ𝑁 . Since 𝑊 is compact, it is

a closed subset of ℝ𝑁 , and so is 𝜕𝑊 . Since 𝑓 is a smooth map defined on a closed subset
of ℝ𝑁 , it may be locally extended to a smooth map on open sets. Since 𝜕𝑊 is compact and
boundaryless, we can apply the Tubular Neighborhood Theorem 12.10 to extend 𝑓 to a map
𝐹 defined on a neighborhood 𝑈 of 𝜕𝑊 in ℝ𝑁 .

Now we choose a smooth bump function 𝜌 that is constant 1 on 𝜕𝑊 and 0 outside some
compact subset of 𝑈 . Then we can extend 𝑓 to all of 𝑊 by letting it be

𝜌 ⋅ 𝐹 on 𝑈 , and 0 outside of 𝑈 .

This is a smooth function defined on all of ℝ𝑁 with values in ℝ𝑘 and being 𝑓 = 1 ⋅ 𝐹 on 𝜕𝑊 .

Now we apply this lemma to maps with values in spheres:

Theorem 16.13 (Extension Theorem for degree zero maps) Let 𝑊 be a com-
pact, connected, oriented 𝑘 + 1-dimensional smooth manifold with boundary, and
let 𝑓 ∶ 𝜕𝑊 → 𝕊𝑘 be a smooth map. Then 𝑓 extends to a globally defined map
𝐹 ∶ 𝑊 → 𝕊𝑘 with 𝜕𝐹 = 𝑓 if and only if deg(𝑓 ) = 0.

Proof: We already know that if 𝑓 can be extended to all of 𝑊 , then deg(𝑓 ) = 0. It remains
to show the opposite direction.

So let 𝑓 be as in the theorem, and assume deg(𝑓 ) = 0. By Lemma 16.12 above, we can
extend 𝑓 to a smooth map 𝐹 ∶ 𝑊 → ℝ𝑘+1 . By the Transversality Extension Theorem 12.29,
we can assume that 0 is a regular value of 𝐹 . Since 𝑊 is compact of dimension 𝑘 + 1, we know
that 𝐹 −1 (0) is a finite set. Hence we can apply the corollary to the Isotopy Lemma 11.3 to this
finite set, and move 𝐹 −1 (0) inside Int(𝔹) where 𝔹 is a closed ball contained Int(𝑊 ).
𝐹
In particular, since 𝐹 −1 (0) ⊂ Int(𝔹), the map |𝐹 |
extends to 𝑊 ′ ∶= 𝑊 ⧵ Int(𝔹). Hence

𝑊 (𝐹 ∕|𝐹 |, 0) = deg (𝐹 ∕|𝐹 |) = 0.

On the other hand, we know by our assumption that

𝑊 (𝐹|𝜕𝑊 , 0) = 𝑊 (𝑓 , 0) = deg(𝑓 ) = 0,

where we use 𝑓 = 𝑓 ∕|𝑓 |, since 𝑓 has values in 𝕊𝑘 .

332 16.4. The proof of Hopf’s theorem
Now let

𝜕𝐹 = 𝐹𝜕𝔹 ∶ 𝜕𝔹 → ℝ𝑘+1 ⧵ {0}

be the restriction to the boundary. By the definition of 𝑊 ′ and of boundary orientations, we

have
𝜕𝑊 ′ = (𝜕𝑊 ) ∪ (−𝜕𝔹).

Hence we get

𝑊 (𝐹|𝜕𝑊 ′ , 0) = 𝑊 (𝐹|𝜕𝑊 , 0) − 𝑊 (𝐹|𝜕𝔹 , 0)

and therefore 𝑊 (𝐹|𝜕𝔹 , 0) by our previous observations.

Now Theorem 16.10, i.e., the consequence of the special case, implies that 𝜕𝐹 is homotopic
to a constant map. By Step 3, i.e., Lemma 16.8, this implies that 𝜕𝐹 extends to a map 𝐺 ∶ 𝑊 →
ℝ𝑘+1 ⧵ {0}. Then the map 𝐺∕|𝐺| ∶ 𝑊 → 𝕊𝑘 is the global extension of 𝑓 .

16.4 The proof of Hopf’s theorem

Let 𝑓0 and 𝑓1 be two maps 𝑋 → 𝕊𝑘 and let 𝑊 ∶= 𝑋 × [0, 1]. We define a map 𝑓 ∶ 𝜕𝑊 → 𝕊𝑘
by setting
{
𝑓0 on 𝑋 × {0}
𝑓=
𝑓1 on 𝑋 × {1}.

By the Extension Theorem 16.13, 𝑓 extends to a map on all of 𝑊 if and only if deg(𝑓 ) = 0.
By definition, such an extension would be a homotopy between 𝑓0 and 𝑓1 . Thus we have

𝑓0 ∼ 𝑓1 ⇐⇒ deg(𝑓 ) = 0.

It remains to relate deg(𝑓 ) to deg(𝑓0 ) and deg(𝑓1 ). But, since 𝜕𝑊 = (𝑋 × {1}) ∪ (𝑋 × {0})
with the opposite orientation on 𝑋 × {0}, it follows that

deg(𝑓 ) = deg(𝑓1 ) − deg(𝑓0 ).

Thus

𝑓0 ∼ 𝑓1 ⇐⇒ deg(𝑓1 ) = deg(𝑓0 ).
17. Vector Fields and the Poincaré–Hopf Index Theorem

17.1 Vector Fields

Let 𝑋 ⊂ ℝ𝑁 be a smooth manifold. Recall from the exercises that a vector field on 𝑋 is a
smooth assignment of a vector tangent to 𝑋 at each point 𝑥, i.e., a smooth map 𝐯 ∶ 𝑋 → ℝ𝑁
such that 𝐯(𝑥) ∈ 𝑇𝑥 (𝑋) for every 𝑥.

Vector fields play a crucial role in many applications, for example in mathematical physics.
To provide efficient tools to understand them is an important motivation for differential topol-
ogy. A priori one might expect that vector fields are free to move in every way they want. It
turns out, however, that the topology of 𝑋 provides pretty strong restrictions for what is al-
lowed and what is not. This is the content of the Poincaré–Hopf Index Theorem 17.16 that we
are going to prove. Since 𝐯(𝑥) varies smoothly with 𝑥, the most interesting points 𝑥 ∈ 𝑋 are
where 𝐯(𝑥) = 0. For, in a small neighborhood of such a point, 𝐯 can change directions radically.
Hence the main points we need to investigate are the zeros of the vector field.

We have already seen some examples of vector fields in the exercises. Here are some more
examples:

Example 17.1 (Vector field on an open subset) Let 𝑈 ⊂ ℝ𝑘 be an open subset. Recall
that the tangent space 𝑇𝑢 𝑈 is just ℝ𝑘 for every 𝑢 ∈ 𝑈 . Hence a vector field 𝐯 on 𝑈 is just
a smooth map 𝑈 → ℝ𝑘 . Let 𝑒1 , … , 𝑒𝑘 be the standard basis of ℝ𝑘 . Then 𝐯 is of the form
∑
𝐯 = 𝑘𝑖=1 𝑣𝑖 𝑒𝑖 for smooth functions 𝑣𝑖 ∶ 𝑈 → ℝ.

Example 17.2 (A vector field on a sphere) On the 𝑛-dimensional sphere 𝕊𝑛 , let

𝑁 = (0, … , 0, 1) ∈ 𝕊𝑛 denote the north pole. We can construct a vector field 𝐯 on 𝕊𝑛 by
setting
𝐯(𝑥) = 𝑁 − (𝑁 ⋅ 𝑥)𝑥
where ⋅ denotes the inner product defined by considering 𝑝𝑁 and 𝑥 as vectors in ℝ𝑛+1 . To
check that this actually defines a vector field, it suffices to check that 𝐯(𝑥) is perpendicular
to every 𝑥:
𝐯(𝑥) ⋅ 𝑥 = 𝑁 ⋅ 𝑥 − (𝑁 ⋅ 𝑥)(𝑥 ⋅ 𝑥) = 0
where we use 𝑥 ⋅ 𝑥 = 1 for every 𝑥 ∈ 𝕊𝑛 .

Example 17.3 (Gradient vector field of a function) Let 𝑋 ⊂ ℝ𝑁 be a smooth 𝑘-

dimensional manifold, and let 𝑓 ∶ 𝑋 → ℝ be a smooth real-valued function on 𝑋. For
each 𝑥 ∈ 𝑋, the derivative 𝑑𝑓𝑥 ∶ 𝑇𝑥 𝑋 → ℝ defines a linear functional on 𝑇𝑥 𝑋. Recall

333
334 17.2. Index of a vector field at a zero
from linear algebra that for any such functional there is a vector 𝐯(𝑥) ∈ 𝑇𝑥 𝑋 such that

𝑑𝑓𝑥 (𝑤) = 𝐯(𝑥)𝑇 ⋅ 𝑤

where 𝐯(𝑥)𝑇 denotes the transpose of 𝐯(𝑥) and ⋅ is matrix multiplication. We may think
of this assignment 𝑥 → 𝐯(𝑥) ∈ 𝑇𝑥 𝑋 as a vector field. This is called the gradient field
of 𝑓 and we denote it by 𝐠𝐫𝐚𝐝(𝑓 ). We note that if 𝑧 ∈ 𝑋 is a zero of 𝐠𝐫𝐚𝐝(𝑓 ), then
𝑑𝑓𝑧 (𝑤) = 𝐠𝐫𝐚𝐝(𝑓 )(𝑧)𝑇 ⋅ 𝑤 = 0 for all 𝑤. Hence 𝑑𝑓𝑧 = 0 and 𝑧 is a critical point of
𝑓 . Conversely, if 𝑧 is a critical point of 𝑓 , then 𝑑𝑓𝑧 = 0 and 𝐠𝐫𝐚𝐝(𝑓 )(𝑧) must be zero.
Thus, 𝑧 is a zero the gradient field of 𝑓 if and only if 𝑧 is a critical point.
In the special case 𝑋 = ℝ𝑘 , the gradient field of 𝑓 can be represented by the Jacobian
(1 × 𝑘)-matrix at 𝑥, i.e.,
( )
𝜕𝑓 𝜕𝑓
𝐠𝐫𝐚𝐝(𝑓 )(𝑥) = (𝑥), … , (𝑥) ,
𝜕𝑥1 𝜕𝑥𝑘

i.e., for the standard basis vectors 𝑒1 , … , 𝑒𝑘 of ℝ𝑘 we have

∑𝑘
𝜕𝑓
𝐠𝐫𝐚𝐝(𝑓 )(𝑥) = (𝑥)𝑒𝑖 .
𝑖=1
𝜕𝑥𝑖

17.2 Index of a vector field at a zero

We begin our investigation in Euclidean space, i.e., we assume first 𝑋 = ℝ𝑛 .

Definition 17.4 (Index at the origin) We assume that the vector field 𝐯 on ℝ𝑛 has an
isolated zero at the origin, i.e., there is a small radius 𝜀 > 0 such that 𝐯 has no other
zeros in the closed ball 𝔹̄ 𝑛𝜀 around the origin with radius 𝜀. Then we can define the map

𝐯(𝑦)
𝐯̄ ∶ 𝕊𝑛−1 → 𝕊𝑛−1 , 𝑦 → .
𝜀 |𝐯(𝑦)|
We equip both spheres with the standard orientation, i.e., they are oriented as the
boundary of 𝔹̄ 𝑛𝜀 and 𝔹̄ 𝑛1 , respectively. We define the index of 𝐯 at 0, denoted ind0 (𝐯), the
be the degree of 𝐯:̄
̄
ind0 (𝐯) ∶= deg(𝐯).

Note:

∙ We can think the map 𝑦 → 𝐯(𝑦)∕|𝐯(𝑦)| as a measure of the variation of the direction of
𝐯 around the origin.

∙ The choice of the radius does not matter for the definition of ind0 , since if we choose
another radius, say 𝜀′ < 𝜀, then 𝐯̄ can be extended to a compact manifold 𝑊 = 𝔹̄ 𝑛𝜀 ⧵ 𝔹𝑛𝜀′
where 𝔹̄ denotes the closed ball and 𝔹 the open ball. The boundary of 𝑊 is 𝜕𝑊 =
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 335

𝜀 − 𝕊𝜀′ . Hence the Boundary Theorem 15.2 implies that the two degrees that arise
𝑛−1
𝕊𝑛−1
from using 𝕊𝑛−1
𝜀 and 𝕊𝑛−1
𝜀′
, respectively, agree.

∙ We can easily construct vector fields with a zero with arbitrary index as follows: In the
plane of complex numbers the polynomial 𝑧 → 𝑧𝑘 defines a smooth vector field with a
zero of index 𝑘 at the origin, and the function 𝑧 → 𝑧̄ 𝑘 defines a vector field with a zero
of index −𝑘. See Figure 17.1.

∙ More examples of vector fields, or for some the curves that the fields are tangent to, in
Figure 17.2 and Figure 17.3.

Figure 17.1: Vector fields arising from the maps 𝑧 → 𝑧1 and 𝑧 → 𝑧̄ with indices 1 and −1,
respectively.

Figure 17.2: Source and sink have index +1, while saddle points have index −1.

Now let 𝑋 be a smooth manifold and let 𝐯 be a vector field on 𝑋. We would like to transfer
the definition of the index of 𝐯 at a point 𝑥 ∈ 𝑋 from Euclidean space to 𝑋. The idea is of how
to do this is the familiar one: we use a local parametrization to transfer information between 𝑋
and Euclidean space.

First we introduce a useful construction:

Definition 17.5 (Pullback of vector fields) Let 𝑋 and 𝑌 be smooth manifolds and
let 𝐯 be a vector field on 𝑌 . Assume that there is a diffeomorphism 𝑓 ∶ 𝑋 → 𝑌 . Then
we define the pullback vector field, denoted 𝑓 ∗ 𝐯, by assigning to 𝑥 the vector which
336 17.2. Index of a vector field at a zero

Figure 17.3: Two types of spiral movements. One has index +1, the other one has index +2,
since we have a positive contribution from above and below.

corresponds to the value of 𝐯 at 𝑓 (𝑥):

𝑓 ∗ 𝐯(𝑥) ∶= 𝑑𝑓𝑥−1 (𝐯(𝑓 (𝑥))).

Example 17.6 (Gradient vector field - pullback) Let 𝑋 ⊂ ℝ𝑁 be a smooth 𝑘-

dimensional manifold, and let 𝑓 ∶ 𝑋 → ℝ be a real-valued function on 𝑋. Recall from
Example 17.3 the gradient field of 𝑓 . Let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization of 𝑋.
We will now have a closed look at the pullback vector field 𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 ) ∶ 𝑈 → ℝ𝑘 on 𝑈 .
By definition we have

𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 )(𝑢) = 𝑑𝜙−1

𝑢 (𝐠𝐫𝐚𝐝(𝑓 )(𝜙(𝑢)).

To compute the right-hand side, let (𝑒1 , … , 𝑒𝑘 ) denote the standard basis of ℝ𝑘 . For
𝑢 ∈ 𝑈 , the vectors 𝑑𝜙𝑢 (𝑒1 ), … , 𝑑𝜙𝑢 (𝑒𝑘 ) form a basis of 𝑇𝜙(𝑢) 𝑋. We need to determine
how 𝑓𝜙(𝑢) acts on each 𝑑𝜙𝑢 (𝑒𝑖 ). By definition of the induced derivative we have

𝑑𝑓𝜙(𝑢) (𝑑𝜙𝑢 (𝑒𝑖 )) = 𝑑(𝑓 ◦𝜙)𝑢 (𝑑𝜙−1

𝑢 ◦𝑑𝜙𝑢 (𝑒𝑖 )) = 𝑑(𝑓 ◦𝜙)𝑢 (𝑒𝑖 )

𝜙 𝑓
where 𝑑(𝑓 ◦𝜙)𝑢 is the derivative of the smooth map 𝑈 ←←←→ ← ℝ. Hence, with respect
← 𝑋 ←←←→
to the basis (𝑑𝜙𝑢 (𝑒1 ), … , 𝑑𝜙𝑢 (𝑒𝑘 )), we can represent 𝑑𝑓𝜙(𝑢) by the Jacobian (1×𝑘)-matrix
( )
𝜕(𝑓 ◦𝜙) 𝜕(𝑓 ◦𝜙)
𝑑𝑓𝜙(𝑢) = (𝑢), … , (𝑢) .
𝜕𝑢1 𝜕𝑢𝑘

The vector 𝐠𝐫𝐚𝐝(𝑓 )(𝜙(𝑢)) ∈ 𝑇𝜙(𝑢) 𝑋 such that 𝑑𝑓𝜙(𝑢) (𝑤) = (𝐠𝐫𝐚𝐝(𝑓 )(𝜙(𝑢)))𝑇 ⋅ 𝑤 for all
𝑤 ∈ 𝑇𝜙(𝑢) 𝑋, is then given by

∑
𝑘
𝜕(𝑓 ◦𝜙)
𝐠𝐫𝐚𝐝(𝑓 )(𝜙(𝑢)) = (𝑢)𝑑𝜙𝑢 (𝑒𝑖 ) ∈ 𝑇𝜙(𝑢) 𝑋.
𝑖=1
𝜕𝑢𝑖
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 337
For a basis vector 𝑑𝜙𝑢 (𝑒𝑗 ) ∈ 𝑇𝜙(𝑢) 𝑋, we then get

𝑑𝑓𝜙(𝑢) (𝑑𝜙𝑢 (𝑒𝑗 )) = (𝐠𝐫𝐚𝐝(𝑓 )(𝜙(𝑢)))𝑇 ⋅ 𝑑𝜙𝑢 (𝑒𝑗 )

( 𝑘 )𝑇
∑ 𝜕(𝑓 ◦𝜙)
= (𝑢)𝑑𝜙𝑢 (𝑒𝑖 ) ⋅ 𝑑𝜙𝑢 (𝑒𝑗 ) ∈ ℝ. (17.1)
𝑖=1
𝜕𝑢𝑖

In order to obtain 𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 ) it remains to apply 𝑑𝜙−1 𝑢 and to express the effect of
𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 )(𝑢) in terms of the standard basis 𝑒1 , … , 𝑒𝑘 of ℝ𝑘 . For the latter we observe
that we have
( ∗ )𝑇
𝜙 𝐠𝐫𝐚𝐝(𝑓 )(𝑢) ⋅ 𝑒𝑗 = (𝐠𝐫𝐚𝐝(𝑓 )(𝜙(𝑢)))𝑇 ⋅ 𝑑𝜙𝑢 (𝑒𝑗 ) ∈ ℝ.

Moreover, Equation 17.1 tells us which real number that is. Thus, in terms of the stan-
dard basis of ℝ𝑘 , we can express 𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 )(𝑢) as follows: We define smooth functions
𝑔𝑖𝑗 ∶ 𝑈 → ℝ by the formula

𝑔𝑖𝑗 (𝑢) = (𝑑𝜙𝑢 (𝑒𝑖 ))𝑇 ⋅ 𝑑𝜙𝑢 (𝑒𝑗 ) ∈ ℝ.

Then we get, for all 𝑢 ∈ 𝑈 ,

( 𝑘 )
∑𝑘
∑ 𝜕(𝑓 ◦𝜙)
(𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 ))(𝑢) = (𝑢)𝑔𝑖𝑗 (𝑢) ⋅ 𝑒𝑗 ∈ ℝ𝑘 . (17.2)
𝑗=1 𝑖=1
𝜕𝑢 𝑖

Finally, recall from Example 17.1 that every vector field 𝐯 on 𝑈 ⊂ ℝ𝑘 is of the form
∑
𝐯 = 𝑗 𝑣𝑗 𝑒𝑗 for smooth functions 𝑣𝑗 ∶ 𝑈 → ℝ. Then we can write 𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 ) as
∑
𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 ) = 𝑣𝑗 𝑒𝑗
𝑗

for the smooth functions

∑
𝑘
𝜕(𝑓 ◦𝜙)
𝑣𝑗 = 𝑔𝑖𝑗 .
𝑖=1
𝜕𝑢𝑖

Now we extend the definition of the index to arbitrary smooth manifolds:

Definition 17.7 (Index of a zero) Let 𝑋 be a smooth manifold and let 𝐯 be a vector
field on 𝑋 with an isolated zero at 𝑧 ∈ 𝑋. Let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization
with 𝜙(0) = 𝑧. We define the index of 𝐯 at 𝑧 to be

ind𝑧 (𝐯) ∶= ind0 (𝜙∗ 𝐯).

We need to show that this definition does not depend on the choice of the local parametriza-
tion 𝜙. We do this in two steps via the following two lemmata:

Lemma 17.8 Every orientation preserving diffeomorphism 𝑓 ∶ ℝ𝑛 → ℝ𝑛 is

smoothly isotopic to the identity.
338 17.2. Index of a vector field at a zero
Proof: We may assume that 𝑓 (0) = 0. Since the evaluation at a point 𝑥 in 𝑇𝑥 ℝ𝑛 = ℝ𝑛 of
the derivative of 𝑓 at 0 can be defined by
𝑑𝑓0 (𝑥) = lim 𝑓 (𝑡𝑥)∕𝑡,
𝑡→0

we define an isotopy 𝐹 ∶ ℝ𝑛 × [0, 1] → by the formula

ℝ𝑛
{
𝐹 (𝑥, 𝑡) ∶= 𝑓 (𝑡𝑥)∕𝑡 for 0 < 𝑡 ≤ 1,
𝐹 (𝑥, 0) ∶= 𝑑𝑓0 for 𝑡 = 0.

We need to check that 𝐹 is smooth even when 𝑡 → 0. To do so we write 𝑓 (𝑥) = 𝑥1 𝑔1 (𝑥) + ⋯ +

𝑥𝑛 𝑔𝑛 (𝑥) for smooth functions 𝑔𝑖 ∶ ℝ𝑛 → ℝ defined by
1
𝜕𝑓
𝑔𝑖 (𝑥1 , … , 𝑥𝑛 ) ∶= (𝑡𝑥 , … , 𝑡𝑥𝑛 )𝑑𝑡.
∫0 𝜕𝑥𝑖 1
We then have
𝐹 (𝑥, 𝑡) = 𝑥1 𝑔1 (𝑡𝑥) + ⋯ + 𝑥𝑛 𝑔𝑛 (𝑡𝑥) for all 𝑡 ∈ [0, 1].
Thus 𝑓 is smoothly isotopic to the linear map 𝑑𝑓0 . Since 𝑓 is an orientation preserving
diffeomorphism, we know det 𝑑𝑓0 > 0. Since GL𝑛 (ℝ)+ is path-connected, this implies that
𝑑𝑓0 is smoothly isotopic to the identity.

∙ Note that the assertion of Lemma 17.8 is quite different from the situation on the sphere
𝕊𝑛 where there exists orientation preserving diffeomorphism which is not smoothly ho-
motopic to the identity.

That the index of a zero of a vector field on a smooth manifold is well-defined will now
follow from the next lemma:

Lemma 17.9 Let 𝑈 and 𝑈 ′ be open neighborhoods of the origin in ℝ𝑛 . Let 𝐯 be a

vector field on 𝑈 with an isolated zero at the origin. Assume there is a diffeomorphism
𝑓 ∶ 𝑈 ′ → 𝑈 with 𝑓 (0) = 0. Then the index of 𝐯 at 0 is equal to the index of 𝑓 ∗ 𝐯 at 0:

ind0 (𝐯) = ind0 (𝑓 ∗ 𝐯).

Proof: Since we only need to study the zeros in a small open neighborhood of the origin,
we may assume that 𝑈 and 𝑈 ′ are open balls around the origin with sufficiently small radii 𝜀
and 𝜀′ , respectively. First we assume that 𝑓 preserves orientation. Then by Lemma 17.8 we
can find a smooth isotopy
𝑓𝑡 ∶ 𝑈 ′ × [0, 1] → ℝ𝑛 , with 𝑓0 = id, 𝑓1 = 𝑓 , and 𝑓𝑡 (0) = 0 for all 𝑡.
For each 𝑡, we set 𝐯𝑡 ∶= 𝑓𝑡∗ 𝐯 on 𝑓𝑡−1 (𝑈 ) ⊂ ℝ𝑛 . Since 𝑓𝑡 is a smooth isotopy, each 𝑓𝑡 is a
diffeomorphism and the vector field 𝐯𝑡 is well-defined with an isolated zero at the origin for
every 𝑡 ∈ [0, 1]. For each 𝑡 ∈ [0, 1], let 𝜀𝑡 > 0 be the maximal radius such that 𝐯𝑡 has no zeros
other than the origin in 𝔹𝑛𝜀 . The function [0, 1] → ℝ+ , 𝑡 → 𝜀𝑡 is continuous and hence takes
𝑡
a minimum when 𝑡 varies in [0, 1]. Let 𝜀 > 0 be that minimum. Then we have a well-defined
smooth map
𝐯 (𝑦)
𝐯̄ 𝑡 ∶ 𝕊𝑛−1 → 𝕊𝑛−1 , 𝑦 → 𝑡 .
𝜀 |𝐯𝑡 (𝑦)|
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 339
As 𝐯0 = 𝐯 and 𝐯1 = 𝑓 ∗ 𝐯, we deduce

ind0 (𝐯) = deg(𝐯̄ 0 ) = deg(𝐯̄ 1 ) = ind0 (𝑓 ∗ 𝐯)

from the invariance of deg under homotopy. This proves the assertion for an orientation pre-
serving diffeomorphism.

Now we assume that 𝑓 reverses orientation. Then we can write 𝑓 as the composition of
an orientation preserving diffeomorphism and a reflection 𝑟 ∶ ℝ𝑛 → ℝ𝑛 which switches the sign
of exactly one coordinate, i.e., 𝑟 sends (𝑥1 , … , 𝑥𝑖 , … , 𝑥𝑛 ) to (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑛 ). By the first
case, we may therefore assume that 𝑓 = 𝑟. We write 𝐯′ = 𝑓 ∗ 𝐯. Then the associated map 𝐯̄ ′ on
the 𝜀-sphere 𝕊𝑛−1
𝜀 → 𝕊𝑛−1 with 𝐯̄ ′ (𝑦) = 𝐯′ (𝑦)∕|𝐯′ (𝑦)| satisfies

𝐯̄ ′ = 𝑟−1 ◦𝐯◦𝑟
̄

since 𝑟 is a linear isomorphism and therefore 𝑑𝑟0 = 𝑟. Using the definition of the degree this
shows deg(𝐯) ̄ = deg(𝐯̄ ′ ) and completes the proof of the lemma.

Now we can show that the definition of the index is well-defined:

Lemma 17.10 (Index of a zero is well-defined) Let 𝑋 be a smooth manifold and let
𝐯 be a vector field on 𝑋 with an isolated zero at 𝑧 ∈ 𝑋. Then the index of 𝐯 at 𝑧 does
not depend on the choice of a local parametrization around 𝑧.

Proof: Let 𝜙 ∶ 𝑈 → 𝑋 and 𝜓 ∶ 𝑈 ′ → 𝑋 be local parametrizations with 𝜙(0) = 𝑧 = 𝜓(0).

By choosing 𝑈 and 𝑈 ′ small enough, the composition 𝑓 ∶= 𝜓 −1 ◦𝜙 defines a diffeomorphism
𝑓 ∶ 𝑈 ′ → 𝑈 . We check using the definition of the pullback that we have

𝜙∗ 𝐯 = (𝜓 −1 ◦𝜙)∗ (𝜓 ∗ 𝐯).

By Lemma 17.9, we then get

( )
ind0 (𝜙∗ 𝐯) = ind0 (𝜓 −1 ◦𝜙)∗ (𝜓 ∗ 𝐯) = ind0 (𝜓 ∗ 𝐯).

This proves the assertion.

Example 17.11 (Index sum on spheres) Recall the the vector field 𝐯 on the 𝑛-
dimensional sphere 𝕊𝑛 defined in Example 17.2 given by

𝐯(𝑥) = 𝑁 − (𝑁 ⋅ 𝑥)𝑥

where 𝑁 ∈ 𝕊𝑛 denotes the north pole. and ⋅ denotes the inner product defined by consid-
ering 𝑁 and 𝑥 as vectors in ℝ𝑛+1 . We compute the effect of 𝐯 on 𝑥 = (𝑥1 , … , 𝑥𝑛+1 ) ∈ 𝕊𝑛
as

𝐯(𝑥) = (−𝑥1 𝑥𝑛+1 , −𝑥2 𝑥𝑛+1 , … , −𝑥𝑛 𝑥𝑛+1 , 1 − 𝑥2𝑛+1 ). (17.3)

This shows that 𝐯 has exactly two zeros: the north pole 𝑁 = (0, … , 0, 1) and the south
pole 𝑆 = (0, … , 0, −1). Now we compute the index of 𝐯 at both zeros. By definition,
we compute the index at a zero by choosing a local parametrization and pulling back the
vector field to ℝ𝑛 . For this example, it is convenient to use the stereographic projections
from the poles of 𝕊𝑛 . First, we look at the zero at 𝑆. Let 𝜙𝑁 ∶ ℝ𝑛 → 𝕊𝑛 ⧵ 𝑝𝑁 be the
340 17.3. The Euler characteristic - algebraic topology in a nutshell
stereographic projection form the north pole. We then restrict 𝜙𝑁 to the open lower
hemisphere on §𝑛 as an open neighborhood 𝑉𝑆 of 𝑆 which does not contain any other
zeros of 𝐯. This has the advantage that the restriction of 𝜙𝑁 to 𝑉𝑆 is a diffeomorphism of
𝑉𝑆 to the open ball 𝑈𝑆 ∶= 𝔹𝑛1 ⊂ ℝ𝑛 . Thus the boundary of the closure 𝑈 𝑆 of 𝑈𝑆 is just
the unit sphere 𝕊𝑛−1 ∈ ℝ𝑛 . Hence 𝜙𝑁 send a point 𝑢 ∈ 𝜕𝑈 𝑆 to 𝜙𝑁 (𝑢) = (𝑢1 , … , 𝑢𝑛 , 0) ∈
𝕊𝑛 ⊂ ℝ𝑛+1 . Thus, by Equation 17.3, we get

𝐯(𝜙𝑁 (𝑢) = (0, … , 0, 1) for all 𝑢 ∈ 𝜕𝑈 𝑆 .

This does not mean that 𝜙∗𝑁 𝐯 is constant, since we still have to apply the derivative
(𝑑(𝜙𝑁 )𝑢 )−1 . We computed this derivative in Exercise 2.7. The formula there yields

⎛1 0 … 0 𝑢1 ⎞ ⎛0⎞ ⎛𝑢1 ⎞
⎜0 1 … 0 𝑢2 ⎟ ⎜⋮⎟ ⎜𝑢2 ⎟
𝜙∗𝑁 𝐯(𝑢) = ⎜ ⎟ ⋅ ⎜ ⎟ = ⎜ ⎟ for all 𝑢 ∈ 𝜕𝑈 𝑆
⎜⋮ ⋮ … ⋮ ⋮ ⎟ ⎜0⎟ ⎜ ⋮ ⎟
⎝ 0 … 0 1 𝑢𝑛 ⎠ ⎝ 1 ⎠ ⎝ 𝑢𝑛 ⎠

Thus, on the boundary of the closure of the neighborhood around the zero 𝑆, 𝜙∗𝑁 𝐯 acts
as the identity. This shows that the index of 𝐯 at 𝑆 is +1:

ind𝑆 (𝐯) = ind0 (𝜙∗𝑁 𝐯) = deg(id𝕊𝑛−1 ) = +1.

To compute the index at the north pole 𝑁 we use the stereographic projection 𝜙𝑆 and
use the upper hemisphere as a neighborhood 𝑉𝑁 of 𝑁. We then get again 𝐯(𝜙𝑁 (𝑢) =
(0, … , 0, 1) for all 𝑢 ∈ 𝜕𝑈 𝑁 . However, the vector field 𝜙∗𝑆 𝐯 now acts as

⎛1 0 … 0 −𝑢1 ⎞ ⎛0⎞ ⎛−𝑢 ⎞

⎜0 1 … 0 −𝑢2 ⎟ ⎜⋮⎟ ⎜ 1 ⎟
𝜙∗𝑆 𝐯(𝑢) = ⎜ ⎟ ⋅ ⎜ ⎟ = ⋮ for all 𝑢 ∈ 𝜕𝑈 𝑁
⎜⋮ ⋮ … ⋮ ⋮ ⎟ ⎜0⎟ ⎜⎝−𝑢𝑛 ⎟⎠
⎝0 … 0 1 −𝑢𝑛 ⎠ ⎝1⎠

This shows that 𝜙∗𝑆 𝐯 acts as the antipodal map 𝑎𝕊𝑛−1 on 𝕊𝑛−1 . Hence the index of 𝐯 at
𝑁 is equals the degree of the antipodal map on 𝕊𝑛−1 which is (−1)𝑛 by Exercise 15.1:

ind𝑁 (𝐯) = ind0 (𝜙∗𝑆 𝐯) = deg(𝑎𝕊𝑛−1 ) = (−1)𝑛 .

We observe that, in this example, the sum of the indices of the zeros of the vector field
𝐯 is 0 if 𝑛 is odd and is 2 if 𝑛 is even. We will see in the next section that this sum is
actually independent of the particular choice of vector field.

17.3 The Euler characteristic - algebraic topology in a nutshell

In order to state the main result of this chapter we have to take a brief detour to algebraic topol-
ogy and introduce the Euler characteristic. There are of course many different ways to define
the Euler characteristic of a smooth manifold 𝑋. However, the definitions that are accessible
with the techniques developed in this book, for example as the self-intersection number of the
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 341
diagonal embedded into 𝑋 × 𝑋, do not convey the property that we are interested here: the
Euler characteristic is a purely topology invariant, i.e., it only depends on the topology of 𝑋
and not it structure as a smooth manifold. Other definition, for example via triangulation, are
more intuitive and seem more elementary. To make them precise and independent of choices,
however, requires a lot of work. Since we assume the reader to be interested or familiar with
algebraic topology anyway we will define the Euler characteristic via singular homology.

We will now briefly introduce singular homology groups. We refer the reader to any general
introduction to algebraic topology for any details and further explanation and motivation, for
example [6] or [16]. One may view the underlying idea as to map certain model spaces with
precisely understood properties into the topological spaces we would like to understand. There
are different choices for these model spaces, for example one can use spheres and define homo-
topy groups. For singular homology one uses spaces with a very nice combinatorial behavior.
A rigorous way to do this is In algebraic topology is the following.

For 𝑛 ≥ 0, the standard 𝑛-simplex Δ𝑛 is the set Δ𝑛 ⊂ ℝ𝑛+1 defined by

∑
𝑛
𝑛
Δ = {(𝑡0 , … , 𝑡𝑛 ) ∈ ℝ 𝑛+1
∶ 𝑡𝑖 = 1, 𝑡𝑖 ≥ 0 for all 𝑖}.
𝑖=0

The standard simplices are related by face maps for 0 ≤ 𝑖 ≤ 𝑛 which can be described as

𝜙𝑛𝑖 (𝑡0 , … , 𝑡𝑛−1 ) = (𝑡0 , … , 𝑡𝑖−1 , 0, 𝑡𝑖 , … , 𝑡𝑛−1 )

with the 0 inserted at the 𝑖th coordinate where 𝑡0 is the 0th coordinate. Let 𝑋 be a topological
space. A singular 𝑛-simplex in 𝑋 is a continuous map 𝜎 ∶ Δ𝑛 → 𝑋. We denote by Sing𝑛 (𝑋)
the set of all 𝑛-simplices in 𝑋. For example, Sing0 (𝑋) is just the set of points of 𝑋. But, in
general, Sing𝑛 (𝑋) carries more interesting information for 𝑛 ≥ 1. For 0 ≤ 𝑖 ≤ 𝑛, we can use
the face maps 𝜙𝑛𝑖 to define maps

𝑑𝑖𝑛 ∶ Sing𝑛 (𝑋) → Sing𝑛−1 (𝑋), 𝜎 → 𝜎◦𝜙𝑛𝑖

by sending an 𝑛-simplex 𝜎 to the 𝑛 − 1-simplex defined by precomposition with the 𝑖th face
inclusion. The image 𝑑𝑖𝑛 (𝜎) = 𝜎◦𝜙𝑛𝑖 is called the 𝑖th face of 𝜎.

Figure 17.4: The face maps from the one-dimensional simplex into the two-dimensional one.
There are three maps corresponding to the three edges of Δ2 . Then 𝜎 maps the simplices to 𝑋.
342 17.3. The Euler characteristic - algebraic topology in a nutshell
To make this construction accessible to algebraic tools, we define the group 𝑆𝑛 (𝑋) of sin-
gular 𝑛-chains in 𝑋 as the free abelian group generated by 𝑛-simplices, i.e.,

𝑆𝑛 (𝑋) ∶= ℤSing𝑛 (𝑋).

Thus an 𝑛-chain is a finite ℤ-linear combination of simplices. We define the boundary operator
as the homomorphism of abelian groups determined by

∑
𝑛
𝜕𝑛 ∶ 𝑆𝑛 (𝑋) → 𝑆𝑛−1 (𝑋), 𝜕𝑛 (𝜎) = (−1)𝑖 𝑑𝑖𝑛 𝜎.
𝑖=0

A key property of the operators 𝜕𝑛 is that their composition vanishes, i.e., 𝜕𝑛+1 ◦𝜕𝑛 = 0. This
implies that the sequence of pairs {𝑆𝑛 (𝑋), 𝜕𝑛 } forms a chain complex:

𝜕𝑛+1 𝜕𝑛 𝜕𝑛−1 𝜕2 𝜕1 𝜕0
⋯ ←←←←←←←→
← 𝑆𝑛 (𝑋) ←←←←→ ← ⋯ ←←←←→
← 𝑆𝑛−1 (𝑋) ←←←←←←←→ ← 𝑆1 (𝑋) ←←←←→
← 𝑆0 (𝑋) ←←←←→
← 0.

This allows to make the following definition:

Definition 17.12 (Singular homology) The 𝑛th singular homology group of 𝑋 with
coefficients in ℤ is defined to be the quotient group

Ker (𝜕 ∶ 𝑆𝑛 (𝑋) → 𝑆𝑛−1 (𝑋))

𝐻𝑛 (𝑋; ℤ) = .
Im (𝜕 ∶ 𝑆𝑛+1 (𝑋) → 𝑆𝑛 (𝑋))

∙ The construction of singular homology is functorial, i.e., a continuous map 𝑓 ∶ 𝑌 → 𝑌

induces a homomorphism of abelian groups 𝑓∗ ∶ 𝐻𝑛 (𝑋; ℤ) → 𝐻𝑛 (𝑌 ; ℤ).

∙ For each 𝑛, the group 𝐻𝑛 (𝑋; ℤ) only depends on the homotopy type of 𝑋, i.e. 𝑓 ∶ 𝑌 →
𝑌 is a homotopy equivalence, then 𝑓∗ ∶ 𝐻𝑛 (𝑋; ℤ) → 𝐻𝑛 (𝑌 ; ℤ) is an isomorphism.

∙ This follows from the fact that the assignment 𝑓 → 𝑓∗ is invariant under homotopy, i.e.,
𝑓 ≃ 𝑔 implies 𝑓∗ = 𝑔∗ .

∙ Intuitively, the 𝑛th singular homology measures the following property of 𝑋: A 𝑛-cycle
𝛼, i.e., an element in the kernel of 𝜕𝑛 , is a closed 𝑛-dimensional loop on 𝑋. The cycle 𝛼
vanishes in homology if it is a boundary, i.e., if there is an element 𝛽 in 𝑆𝑛+1 (𝑋) such that
𝜕𝑛+1 (𝛽) = 𝛼. In other words, 𝛼 vanishes if it can be filled by a chain of one dimension
higher. Hence we may think of a nonzero element in 𝐻𝑛 (𝑋; ℤ) as a tool to detect an
𝑛-dimensional hole in 𝑋.

∙ In particular, the group 𝐻0 (𝑋; ℤ) is determined by the number 𝑚 of connected compo-

nents of 𝑋, i.e., 𝐻0 (𝑋; ℤ) is a sum of 𝑚 copies of ℤ. The 𝐻1 (𝑋; ℤ) is an abelianization
of the fundamental group of 𝑋, i.e., it is determined by the closed loops on 𝑋 up to
continuous deformations.

∙ The homology of the 𝑘-dimensional sphere are given by 𝐻0 (𝕊𝑘 ; ℤ) = ℤ, 𝐻𝑘 (𝕊𝑘 ; ℤ) = ℤ,

and 𝐻𝑖 (𝕊𝑘 ; ℤ = 0 for all 𝑖 ≠ 0, 𝑘.
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 343

Remark 17.13 (Relative homology and attaching cells) Given a subspace 𝑌 ⊂ 𝑋,

there are relative homology groups denoted 𝐻𝑛 (𝑋, 𝑌 ; ℤ) which, roughly speaking, mea-
sure only the features of the complement of 𝑌 in 𝑋. They are defined as the homology
of the induced chain complex {𝑆𝑛 (𝑋)∕𝑆𝑛 (𝑌 ), 𝜕̄𝑛 }𝑛 . A special case of this situation is
of particular interest for us: Suppose we are given a map 𝑓 ∶ 𝕊𝑛−1 → 𝑌 . Then we can
attach 𝔹̄ 𝑛 to 𝑌 via 𝑓 by forming the pushout diagram

𝑓
𝕊𝑛−1
_ /𝑌

 
𝔹̄ 𝑛 /𝑌 ∪ 𝔹 ̄ 𝑛.
𝑓

We think of this process as obtaining 𝑋 by gluing an 𝑛-dimensional cell to 𝑌 . In this

case, the relative homology groups are given as follows:
{
ℤ if 𝑗 = 𝑛
𝐻𝑗 (𝑋, 𝑌 ; ℤ) =
0 otherwise

The Euler characteristic may now be viewed as a very rough summary of the information the
homology groups contain. First we reduce the information of homology groups to their rank,
i.e., we let 𝑏𝑖 denote the rank of 𝐻𝑖 (𝑋; ℤ) as an abelian group. The number 𝑏𝑖 is called the 𝑖th
Betti number of 𝑋. If 𝑋 is a 𝑘-dimensional manifold, then we have 𝑏𝑖 = 0 for all 𝑖 ≥ 𝑘 + 1.
Hence we may take the alternating sum of the Betti numbers:

Definition 17.14 (Euler characteristic) Let 𝑋 be a smooth 𝑘-dimensional manifold.

The Euler characteristic of 𝑋, denoted by 𝜒(𝑋), is defined as the alternating sum of
the Betti numbers of 𝑋, i.e.,

∑
𝑘
∑
𝑘
𝜒(𝑋) ∶= (−1)𝑖 𝑏𝑖 = (−1)𝑖 rank𝐻𝑖 (𝑋; ℤ).
𝑖=1 𝑖=1

∙ Since, for all 𝑖, the groups 𝐻𝑖 (𝑋; ℤ) only depend on the topology of 𝑋, it follows that
Euler characteristic of 𝑋 only depends on the topology of 𝑋. In fact, 𝜒(𝑋) only depends
on the homotopy type of 𝑋.

The Euler characteristic can often be computed by other methods. The following situation
is of particular interest in Morse theory which will play again a role later in this chapter.

Remark 17.15 (Euler characteristic of a cell complex) Assume that the manifold 𝑋 is
homotopy equivalent to a cell complex 𝑌 , i.e., a space which is obtained by successively
gluing 𝑖-dimensional cells together for 𝑖 = 1, … , 𝑘. Let 𝑐𝑖 denote the number of 𝑖-
dimensional cells we attach. In this case, the Euler characteristic of 𝑋 can be computed
344 17.4. The Poincaré–Hopf Index Theorem
as

∑
𝑘
𝜒(𝑋) = (−1)𝑖 𝑐𝑖 . (17.4)
𝑖=1

Equation 17.4 comes quite close to the familiar formula of Euler which says that the
alternating sum 𝑣 − 𝑒 + 𝑓 is an invariant of a surface 𝑆, where 𝑣, 𝑒, 𝑓 denote the number
of vertices, edges and faces, respectively, of a polygon which covers 𝑆. For example,
the Euler characteristic of the 2-sphere is two and the Euler characteristic of the torus is
zero.

17.4 The Poincaré–Hopf Index Theorem

We are now ready to state the following famous result:

Theorem 17.16 (Poincaré–Hopf Index Theorem) Let 𝑋 be a compact smooth man-

ifold and let 𝐯 be a vector field on 𝑋 with only finitely many isolated zeros. Then the
sum of the indices of 𝐯 at its zeros is equal to the Euler characteristic of 𝑋.

∙ Theorem 17.16 consists of two remarkable statements: First, the sum of the indices is
independent of the vector field, i.e., the sum is entirely determined by the manifold 𝑋.
Secondly, it is just the topology of 𝑋 that matters since the latter determines the Euler
characteristic of 𝑋.

∙ The Euler characteristic is a topological invariant while the nature of the index is an-
alytic. Hence we may think of the theorem as a version of an index theorem. The
generalizations of such results are extremely influential and important in many fields of
mathematics, mostly for the reason that they actually connect different branches.

∙ We have seen in Example 17.11 that the index sum for some vector field on 𝕊2𝑛 is 2.
Hence the index sum for any vector field on 𝕊2𝑛 must be 2. In particular, every vector
field on 𝕊2𝑛 must have at least one zero.

∙ Similarly, every vector field on 𝕊2𝑛+1 has index sum 0.

∙ The theorem can be extended to manifolds with boundary by requiring that the vector field
has to point outwards at every boundary point. We will omit the proof of this extended
result here.

∙ The Eisenbud–Levine–Khimshiashvili signature formula provides an algebraic method

to compute the index at isolated zeros by assigning a quadratic form to the germ of a
smooth 𝑓 ∶ ℝ𝑛 → ℝ𝑛 with 𝑓 (0) = 0 and an isolated zero at the origin, see [2] and [3].

As an immediate consequence of Theorem 17.16, we observe:

Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 345

Corollary 17.17 (Vanishing of Euler characteristic) If a compact manifold 𝑋 admits

a vector field without zeros, then the Euler characteristic of 𝑋 must be zero.

∙ For example, we know from the exercises that odd-dimensional spheres admit non-
vanishing vector fields. This again shows 𝜒(𝕊2𝑘+1 ) = 0.

Corollary 17.18 (Non-vanishing of Euler characteristic implies existence of zeros)

Let 𝑋 be a compact smooth manifold. If 𝜒(𝑋) ≠ 0, then every vector field on 𝑋 must
have at least one zero.

In fact, the dimension of the manifold has a strong influence on the Euler characteristic.

Corollary 17.19 (Odd dimensional manifolds have vanishing Euler characteristic)

Let 𝑋 be a compact smooth manifold. If the dimension 𝑛 of 𝑋 is odd, then the index
sum of any vector field on 𝑋 and the Euler characteristic of 𝑋 are zero.

Proof: Let 𝐯 be a vector field on 𝑋. Then we can multiply each 𝐯(𝑥) by −1 to get a new field
we denote by −𝐯. If 𝑧 is a zero of 𝐯, 𝑧 is a zero of −𝐯 as well. The induced maps 𝕊𝑛−1
𝜀 → 𝕊𝑛−1
we use to compute the indices at 𝑧 of 𝐯 and −𝐯, respectively, differ by a composition with the
antipodal map 𝑎𝕊𝑛−1 on 𝕊𝑛−1 . This changes the degree by multiplication by (−1)𝑛 which is −1,
since 𝑛 is odd. Thus, by Theorem 17.16, the index sum of 𝐯 and −𝐯 satisfy
∑ ∑ ∑
ind𝑧 (𝐯) = ind𝑧 (−𝐯) = − ind𝑧 (𝐯).
𝑧∈𝐯−1 (0) 𝑧∈(−𝐯)−1 (0) 𝑧∈𝐯−1 (0)

Hence the index sum must be zero. By Theorem 17.16 this implies that the Euler characteristic
is zero as well.

Example 17.20 (Nowhere vanishing vector field on the torus) Let 𝕋 = 𝕊1 × 𝕊1 ⊂ ℝ4

be the two-dimensional torus. We define a vector field 𝐯 on 𝕋 by

𝐯 ∶ 𝕋 → ℝ4 , 𝑥 = (𝑥1 , 𝑦1 , 𝑥2 , 𝑦2 ) → (−𝑦1 , 𝑥1 , −𝑦2 , 𝑥2 ).

We can verify that 𝐯(𝑥) ∈ 𝑇𝑥 𝕋 = 𝑇(𝑥1 ,𝑦1 ) 𝕊1 × 𝑇(𝑥2 ,𝑦2 ) 𝕊1 ⊂ ℝ4 . Moreover, we see 𝐯(𝑥)
is never zero. Hence there are nowhere vanishing vector fields on the torus. We deduce
from this example and Theorem 17.16 that the index sum of any vector field on the torus
and that the Euler characteristic of the torus are both zero.

Remark 17.21 (Hopf’s theorem on non-vanishing vector fields) In fact, it is a theorem

of Hopf in [8] that the vanishing of the index sum is not just a necessary condition for the
existence of a non-vanishing vector field but also sufficient. That is, if the index sum of
sum of some vector field on a connected compact oriented smooth manifold 𝑋 is zero,
then there exists a vector field on 𝑋 with no zeros at all. We will discuss a proof of this
result in Section 17.7.
346 17.5. Poincaré–Hopf Theorem - Independence
17.5 Poincaré–Hopf Theorem - Independence

We split the proof of Theorem 17.16 into two steps: the assertion on the independence of the
index sum of the vector field, and the identification with the Euler characteristic. We begin with
independence and will prove the following theorem:

Theorem 17.22 (Poincaré–Hopf Index Theorem - Independence) Let 𝑋 be a com-

pact smooth manifold and let 𝐯 be a vector field on 𝑋 with only finitely many isolated
zeros. If 𝑋 has a boundary we require that 𝐯 points outward at all boundary points of
𝑋. Then the index sum does not depend on the choice of the vector field.

We prove Theorem 17.22 in several steps. The key idea is show that the index sum equals
the degree of a particular map that only depends on 𝑋.

Definition 17.23 (Gauss map) Let 𝑋 ⊂ ℝ𝑛 be a compact smooth 𝑛-dimensional

manifold with boundary. The Gauss map

𝑔 ∶ 𝜕𝑋 → 𝕊𝑛−1

is defined by sending 𝑥 ∈ 𝜕𝑋 to the outward pointing unit normal vector at 𝑥.

In the following we are going to use without mentioning the following observation:

Remark 17.24 The assumption that 𝑋 ⊂ ℝ𝑛 is an 𝑛-dimensional manifold implies

that its tangent space, which is an 𝑛-dimensional subspace of ℝ𝑛 , is equal to ℝ𝑛 at every
point. Hence a vector field on 𝑋 is a smooth map 𝐯 ∶ 𝑋 → ℝ𝑛 .

Now we prove the first case of Theorem 17.22.

Lemma 17.25 Let 𝑋 ⊂ ℝ𝑛 be a compact smooth 𝑛-dimensional manifold with bound-

ary. Let 𝐯 ∶ 𝑋 → ℝ𝑛 be a vector field on 𝑋 with finitely many isolated zeros and such
that 𝐯 points outward at all boundary points of 𝑋. Then the index sum is equal to the
degree of the Gauss map. In particular, the index sum does not depend on the choice
of 𝐯.

Proof: Let 𝑧1 , … , 𝑧𝑘 be the zeros of 𝐯 which are not on the boundary of 𝑋. After removing
an 𝜀-ball around each zero for a sufficiently small 𝜀 we obtain again a smooth manifold with
boundary which we denote by 𝑌 . The map 𝐯(𝑦) ̄ = 𝐯(𝑦)∕|𝐯(𝑦)| defines a smooth map 𝑌 → 𝕊𝑛−1 .
Since 𝐯̄ is defined on all of 𝑌 , the degree of the restriction of 𝐯̄ to 𝜕𝑌 is zero by the Boundary
Theorem 15.2. However, the degree of the restriction of 𝐯̄ to 𝜕𝑌 equals the sum of the degrees
of the restrictions of 𝐯̄ to the components of 𝜕𝑌 . One of the components of 𝜕𝑌 is 𝜕𝑋. By
assumption, 𝐯 points outward at all points on 𝜕𝑋. This implies 𝐯̄ |𝜕𝑋 is homotopic to the Gauss
map 𝑔 which sends 𝑥 ∈ 𝜕𝑋 to the outward unit normal vector at 𝑥. Since the degree is invariant
under homotopy, this implies deg(𝐯̄ |𝜕𝑋 ) = deg(𝑔).
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 347
The other boundary components of 𝜕𝑌 are 𝑛 − 1-dimensional 𝜀-spheres with the boundary
orientation induced by the orientation of 𝑋. By construction of 𝑌 the boundary orientation is
such that the normal vector pointing into the 𝜀-sphere completes to a positive oriented basis.
That is, the orientation of each of the 𝜀-spheres, which are the components of 𝜕𝑌 , is the opposite
of the standard orientation. Thus, by definition of the index of a zero of 𝐯, the degree of the
∑
restriction of 𝐯̄ to the other boundary components of 𝜕𝑌 equals − 𝑘𝑖=0 ind𝑧𝑖 (𝐯), the negative
of the index sum. Hence, in total, we have shown

∑
𝑘
deg(𝑔) − ind𝑧𝑖 (𝐯) = deg(𝐯̄ |𝜕𝑌 ) = 0.
𝑖=0

This proves the lemma.

In order to extend Lemma 17.25 to arbitrary manifolds we will now study the derivative of
the vector field.

Definition 17.26 (Nondegenerate zeros - Euclidean case) Let 𝑈 ⊂ ℝ𝑛 be an open

subset. Let 𝐯 ∶ 𝑈 → ℝ𝑛 be a vector field on 𝑈 and let 𝑧 ∈ 𝑈 be a zero of 𝐯. Then the
vector field 𝐯 is said to be nondegenerate at 𝑧 if the linear transformation 𝑑𝐯𝑧 ∶ ℝ𝑛 → ℝ𝑛
is an isomorphism.

∙ Note that it follows from the Inverse Function Theorem 12.4 that, if 𝐯 is nondegenerate
at 𝑧, then 𝑧 is an isolated zero.

Example 17.27 (Gradient vector field - nondegenerate zeros - Euclidean case)

Let 𝑈 ⊂ ℝ𝑛 be an open subset. Let 𝑓 ∶ 𝑈 → ℝ be a smooth function and let 𝐯 ∶=
𝐠𝐫𝐚𝐝(𝑓 ) ∶ 𝑈 → ℝ𝑛 denote its gradient vector field. Recall from Example 17.3 that 𝐯(𝑢)
is represented by the (1 × 𝑘)-Jacobian matrix with respect to the standard basis of ℝ𝑛 .
Hence the derivative 𝑑𝐯𝑢 is represented by the Hessian matrix of 𝑓 . This shows that a
zero of 𝐠𝐫𝐚𝐝(𝑓 ) is nondegenerate if and only if 𝑧 is a nondegenerate critical point of
𝑓 , i.e., the Hessian matrix of 𝑓 at 𝑧 is invertible.

Lemma 17.28 Let 𝑈 ⊂ ℝ𝑛 be an open subset. Let 𝐯 ∶ 𝑈 → ℝ𝑛 be a vector field on

𝑋 and let 𝑢 ∈ 𝑈 be a zero of 𝐯. Assume that 𝐯 is nondegenerate at 𝑢. Then we have
ind𝑧 (𝐯) = det 𝑑𝐯𝑧 ∕| det 𝑑𝐯𝑧 |. That is, the index of 𝐯 at 𝑢 is either +1 or −1 according to
whether the determinant of 𝑑𝐯𝑧 is positive or negative.

Proof: We may assume 𝑧 = 0. We can find a sufficiently small open ball 𝔹 around the
origin inside 𝑈 We may then consider 𝐯|𝔹 as a diffeomorphism into ℝ𝑛 . Now we can apply
Lemma 17.8 and the method used in the proof of Lemma 17.9. If 𝐯 preserves orientation, then
we can smoothly deform it into the identity without introducing any new zeros. In this case
the index equals +1. If 𝐯 reverses orientation, then we can smoothly deform it into a reflection
without introducing any new zeros. In this case the index equals −1. This proves the lemma.
348 17.5. Poincaré–Hopf Theorem - Independence

Lemma 17.29 (Derivative of a vector field) Let 𝑋 ⊂ ℝ𝑁 be a smooth 𝑛-dimensional

manifold, 𝐰 a vector field on 𝑋, and let 𝑧 be a zero of 𝐰. Then the derivative
𝑑𝐰𝑧 ∶ 𝑇𝑧 𝑋 → ℝ𝑛 has image 𝑇𝑧 𝑋 ⊂ ℝ𝑁 , and we may consider 𝑑𝐰𝑧 as a linear transfor-
mation from 𝑇𝑧 𝑋 into itself.
If this linear transformation has determinant ≠ 0, then 𝑧 is an isolated zero of 𝐰 with
index equal det 𝑑𝐰𝑧 ∕| det 𝑑𝐰𝑧 |, i.e, with index +1 or −1 according to whether the de-
terminant of 𝑑𝐰𝑧 is positive or negative.

Proof: We choose a local parametrization 𝜙 ∶ 𝑈 → 𝑋 around 𝑧. Let 𝑒𝑖 denote the 𝑖th basis
vector of ℝ𝑛 . For a point 𝑢 = (𝑢1 , … , 𝑢𝑛 ) ∈ 𝑈 and writing 𝜙 = (𝜙1 , … , 𝜙𝑁 ), we set
1
⎛ 𝜕𝜙 (𝑢) ⎞
⎜ 𝜕𝑢𝑖 ⎟ 𝜕𝜙
𝑡𝑖 (𝑢) ∶= 𝑑𝜙𝑢 (𝑒𝑖 ) = ⎜ ⋮ ⎟ =∶ (𝑢). (17.5)
⎜ 𝜕𝜙 (𝑢)⎟
𝑁 𝜕𝑢𝑖
⎝ 𝜕𝑢
𝑖
⎠

By definition of the tangent space and the derivative, the 𝑡1 (𝑢), … , 𝑡𝑛 (𝑢) then form a basis of
the tangent space 𝑇𝜙(𝑢) 𝑋. We need to determine the image of the 𝑡𝑖 (𝑢) under the linear transfor-
mation 𝑑𝐰𝜙(𝑢) . We fix an 𝑖 and then have

𝜕(𝐰◦𝜙)
𝑑𝐰𝜙(𝑢) (𝑡𝑖 (𝑢)) = 𝑑(𝐰◦𝜙)𝑢 (𝑒𝑖 ) = (𝑢) (17.6)
𝜕𝑢𝑖

using the shortened notation of (17.5) on the right-hand side. Now we let 𝐯 = 𝜙∗ 𝐰 denote
pullback vector field on 𝑈 along 𝜙. Since 𝑈 is an open subset of ℝ𝑛 , the tangent space 𝑇𝑢 𝑈 at
∑
any 𝑢 ∈ 𝑈 is just ℝ𝑛 . We may then write 𝐯 as 𝐯 = 𝑛𝑖=1 𝑣𝑖 𝑒𝑖 with suitable smooth functions
𝑣1 , … , 𝑣𝑛 ∶ 𝑈 → ℝ. By definition we have 𝐯(𝑢) = 𝑑𝜙−1
𝑢 ◦𝐰◦𝜙(𝑢) and 𝑡 (𝑢) = 𝑑𝜙𝑢 (𝑒 ) so that
𝑖 𝑖

∑
𝐰(𝜙(𝑢)) = 𝑑𝜙𝑢 (𝐯(𝑢)) = 𝑣𝑖 (𝑢)𝑡𝑖 (𝑢).
𝑖

Taking partial derivatives with respect to 𝑢𝑖 and using the product rule then yields

𝜕(𝐰◦𝜙) ∑ 𝜕𝑣𝑗 ∑ 𝜕𝑡𝑗

(𝑢) = (𝑢)𝑡𝑗 (𝑢) + 𝑣𝑗 (𝑢) (𝑢). (17.7)
𝜕𝑢𝑖 𝑗
𝜕𝑢𝑖 𝑗
𝜕𝑢𝑖
∑
Now we apply (17.7) at the zero 𝑢̄ = 𝜙−1 (𝑧) of 𝐯, i.e., 𝐯(𝑢) ̄ 𝑖 = 0. Since the 𝑒1 , … , 𝑒𝑛
̄ = 𝑖 𝑣𝑖 (𝑢)𝑒
form a basis of ℝ , we must have 𝑣1 (𝑢)
𝑛 ̄ = … = 𝑣𝑛 (𝑢) ̄ = 0. Thus, at the zero 𝑢̄ we just get

𝜕(𝐰◦𝜙) ∑ 𝜕𝑣𝑗
𝑑𝐰𝑧 (𝑡𝑖 (𝑢))
̄ = (𝑢)
̄ = ̄ 𝑗 (𝑢).
(𝑢)𝑡 ̄ (17.8)
𝜕𝑢𝑖 𝑗
𝜕𝑢𝑖

This shows that, at the zero 𝑧, the image of the basis vector 𝑡𝑖 under 𝑑𝐰𝑧 is a linear combination
of the basis vectors of 𝑇𝑧 𝑋. Thus, the image of 𝑑𝐰𝑧 in ℝ𝑁 is contained in the subspace 𝑇𝑧 𝑋.
This proves the first assertion of the lemma.

Moreover, (17.8) also shows that the determinant of the linear transformation

𝑑𝐰𝑧 ∶ 𝑇𝑧 𝑋 → 𝑇𝑧 𝑋
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 349
𝜕𝑣
equals the determinant of the matrix with (𝑖, 𝑗)th entry 𝜕𝑢𝑗 (𝑢).
̄ Thus we get det 𝑑𝐰𝑧 = det 𝑑𝐯𝑢̄ .
𝑖
Since 𝜙 is a diffeomorphism between open neighborhoods of the zeros 𝑢̄ of 𝐯 and 𝑧 of 𝐰, the
indices ind𝑧 (𝐰) and ind𝑢̄ (𝐯) are equal. Moreover, by the Inverse Function Theorem 12.4 the
zero 𝑧 is isolated if det 𝑑𝐰𝑧 ≠ 0, since then there is a small open neighborhood of 𝑧 such that
𝑧 is the only point where 𝐰(𝑥) vanishes.
By Lemma 17.28 we know ind𝑢̄ (𝐯) is +1 or −1 according to whether det 𝑑𝐯𝑢̄ is positive or
negative. Thus, ind𝑧 (𝐰) is +1 or −1 according to whether det 𝑑𝐰𝑧 is positive or negative. This
proves the lemma.

Lemma 17.29 allows us, in particular, to extend the definition of nondegenerate zeros to
arbitrary vector fields.

Definition 17.30 (Nondegenerate zeros - general case) Let 𝑋 ⊂ ℝ𝑛 be a smooth

manifold and 𝐯 be a vector field on 𝑋. Let 𝑧 ∈ 𝑋 be a zero of 𝐯. Then the zero 𝑧 is said
to be nondegenerate if the linear transformation 𝑑𝐯𝑧 ∶ 𝑇𝑧 𝑋 → 𝑇𝑧 𝑋 is an isomorphism.

Nondegenerate zeros of the gradient field of a function can be detected by studying the
Hessian matrix:

Lemma 17.31 (Gradient vector field - nondegenerate zeros) Let 𝑋 ⊂ ℝ𝑛 be a

smooth 𝑘-dimensional manifold. Let 𝑓 ∶ 𝑋 → ℝ be a real-valued function on 𝑋 and
𝐯 ∶= 𝐠𝐫𝐚𝐝(𝑓 ) be its gradient field defined in Example 17.3. A zero 𝑧 of 𝐠𝐫𝐚𝐝(𝑓 ) is
nondegenerate if and only if 𝑧 is a nondegenerate critical point of 𝑓 , i.e., the Hessian
matrix of 𝑓 at 𝑧, computed in any local coordinate system, is invertible.

Proof: We already know that 𝑧 ∈ 𝑋 is a zero of 𝐠𝐫𝐚𝐝(𝑓 ) if and only if 𝑧 is a critical point
of 𝑓 . Let 𝑧 ∈ 𝑋 be a zero of 𝐠𝐫𝐚𝐝(𝑓 ). Let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization of 𝑋 with
𝜙(𝑢) = 𝑧. Then it follows from Equation 17.2 that the derivative 𝑑(𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 ))𝑢 of 𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 )
at 𝑢 is represented, with respect to the standard basis of ℝ𝑘 , by a matrix which is the product
of the Hessian matrix of 𝑓 ◦𝜙 at 𝑢 and the real (𝑘 × 𝑘)-matrix 𝐺(𝑢) with (𝑖, 𝑗)-entry 𝑔 𝑖 𝑗(𝑢). By
definition, 𝑧 is a nondegenerate critical point of 𝑓 if and only if the Hessian matrix of 𝑓 ◦𝜙 at 𝑢
is invertible. Thus, to prove the assertion it suffices to show that det 𝐺(𝑢) ≠ 0.
Let 𝐴𝑢 denote the matrix which represents 𝑑𝜙𝑢 with respect to the bases 𝑒1 , … , 𝑒𝑘 of ℝ𝑘 and
𝑑𝜙𝑢 (𝑒1 ), … , 𝑑𝜙𝑢 (𝑒𝑘 ) of 𝑇𝜙(𝑢) 𝑋. Then 𝑑𝜙𝑢 (𝑒𝑖 )𝑇 is the 𝑖th row and 𝑑𝜙𝑢 (𝑒𝑗 ) is the 𝑗th column of
𝑀𝑢 . Thus, by definition of 𝐺(𝑢) and the 𝑔𝑖𝑗 (𝑢), we have 𝐺(𝑢) = 𝑀𝑢𝑇 ⋅𝑀𝑢 . Since 𝑀𝑢 is invertible,
this shows that 𝐺(𝑢) is an invertible real symmetric matrix. Hence it is positive definite and all
its eigenvalues are strictly positive. In particular, we see that det 𝐺(𝑢) > 0.

Remark 17.32 (Existence of vector fields with finitely many zeros on compact
manifolds) Since functions with only nondegenerate critical points are generic by The-
orem 7.13, it follows from Lemma 17.31 that there are vector fields on every smooth
manifold with only nondegenerate zeros. In particular, every smooth manifold admits a
vector field with only isolated zeros. Moreover, if 𝑋 is compact, then the set of isolated
zeros is closed and hence finite. Thus, on every compact manifold there exists a vector
field with only finite many zeros which are all isolated.
350 17.5. Poincaré–Hopf Theorem - Independence
Now we let 𝑋 ⊂ ℝ𝑛 be a compact 𝑘-dimensional smooth manifold without boundary. For
𝜀 > 0, we let
𝑋 𝜀 = {𝑦 ∈ ℝ𝑛 ∶ |𝑦 − 𝑥| < 𝜀(𝑥) for some 𝑥 ∈ 𝑋}
be the open subspace in ℝ𝑛 containing 𝑋 of the 𝜀-Neighborhood Lemma 12.8. We then write
𝑋̄ 𝜀 = {𝑦 ∈ ℝ𝑛 ∶ |𝑦 − 𝑥| ≤ 𝜀(𝑥) for some 𝑥 ∈ 𝑋}
for the closure of 𝑋 𝜀 . Note that if we choose 𝜀 > 0 small enough, then 𝑋̄ 𝜀 is a smooth 𝑛-
dimensional manifold with boundary.

Theorem 17.33 (Index sum and Gauss map on 𝜀-neighborhood) Let 𝑋 and 𝑋̄ 𝜀 be
as above. Let 𝐰 be a vector field on 𝑋 with only finitely many zeros all of which are
nondegenerate. Then the index sum of 𝐰 is equal to the degree of the Gauss map

𝑔 ∶ 𝜕 𝑋̄ 𝜀 → 𝕊𝑘−1 .

In particular, the index sum does not depend on the choice of vector field.

Proof: Given a point 𝑦 ∈ 𝑋̄ 𝜀 . We know from ?? that, if we choose 𝜀 small enough, we can
find the point 𝜋(𝑦) ∈ 𝑋 which is closest to 𝑦 and such that the vector 𝑦 − 𝜋(𝑦) is perpendicular
to the tangent space of 𝑋 at 𝜋(𝑦). Moreover, we can consider the assignment 𝑦 → 𝜋(𝑦) as a
smooth map 𝜋 ∶ 𝑋̄ 𝜀 → 𝑋. Since 𝑦 − 𝜋(𝑦) is perpendicular to 𝑇𝜋(𝑦) 𝑋, the tangent space 𝑇𝑦 𝑋̄ 𝜀
at 𝑦 equals the tangent space at 𝜋(𝑦). For a point 𝑦 on the boundary 𝜕 𝑋̄ 𝜀 , we have |𝑦 − 𝜋(𝑦)| = 𝜀
and the vector 𝑦 − 𝜋(𝑦) points outward from 𝜋(𝑦) to 𝑦. Thus the outward unit normal vector
at 𝑦 ∈ 𝜕 𝑋̄ 𝜀 is
𝑔(𝑦) = (𝑦 − 𝜋(𝑦))∕𝜀.
Thus the Gauss map on 𝜕 𝑋̄ 𝜀 is given by
𝑔 ∶ 𝜕 𝑋̄ 𝜀 → 𝕊𝑛−1 , 𝑦 → (𝑦 − 𝜋(𝑦))∕𝜀.
Now we extend the vector field 𝐯 to a vector field 𝐰 on 𝑋̄ 𝜀 by setting
𝐰(𝑦) = (𝑦 − 𝜋(𝑦)) + 𝐯(𝜋(𝑦)).
We will now show that 𝑋̄ 𝜀 and 𝐰 satisfy the hypotheses of Lemma 17.25. Since 𝑋 is compact,
𝑋̄ 𝜀 is a bounded and closed subset of ℝ𝑛 and therefore a compact 𝑛-dimensional submanifold
of ℝ𝑛 with boundary.

∙ Claim: 𝐰 points outward along the boundary.

Since 𝑔(𝑦) points outward and the inner product of 𝐰(𝑦) and 𝑔(𝑦), considered as vectors in ℝ𝑛 ,
is given by
𝐰(𝑦) ⋅ 𝑔(𝑦) = [(𝑦 − 𝜋(𝑦)) + 𝐯(𝜋(𝑦))] ⋅ [(𝑦 − 𝜋(𝑦))∕𝜀]
= [(𝑦 − 𝜋(𝑦)) ⋅ (𝑦 − 𝜋(𝑦))]∕𝜀 + [𝐯(𝜋(𝑦) ⋅ (𝑦 − 𝜋(𝑦))]∕𝜀
= 𝜀2 ∕𝜀 + 0
=𝜀>0
where we use that 𝑦 is on the boundary of 𝑋̄ 𝜀 and the 𝑦 − 𝜋(𝑦) is perpendicular to the tangent
vector 𝐯(𝜋(𝑦) ∈ 𝑇𝜋(𝑦) 𝑋.
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 351
∙ Claim: 𝐯 and 𝐰 have the same zeros.

Since the vectors 𝑦−𝜋(𝑦) and 𝐯(𝜋(𝑦)) are perpendicular, they cannot cancel each others when-
ever they are nonzero. Thus, 𝐰(𝑦) can only be zero if both 𝑦 − 𝜋(𝑦) and 𝐯(𝜋(𝑦)) are the zero
vector. Since 𝑦 − 𝜋(𝑦) = 0 only if 𝑦 ∈ 𝑋, we see that 𝐰 vanishes exactly when 𝐯 does.

∙ Claim: ind𝑧 (𝐰) = ind𝑧 (𝐯) at every zero 𝑧 ∈ 𝑋.

Let 𝑧 ∈ 𝑋 be a zero of 𝐰 and therefore of 𝐯 by the previous claim. By Lemma 17.29 we can
compute both ind𝑧 (𝐰) and ind𝑧 (𝐯) as the the derivatives 𝑑𝐰𝑧 and 𝑑𝐯𝑧 , respectively, and we
know that 𝑑𝐯𝑧 is a linear transformation of 𝑇𝑧 𝑋 into itself, and 𝑑𝐰𝑧 is a linear transformation
of 𝑇𝑧 𝑋̄ 𝜀 into itself. Since 𝑇𝑧 𝑋 ⊂ 𝑇𝑧 𝑋̄ 𝜀 and hence 𝑇𝑧 𝑋̄ 𝜀 = 𝑇𝑧 𝑋 ⊕ 𝑇𝑧 𝑋 ⟂ , we can determine the
effect of 𝑑𝐰𝑧 by computing its effect on vectors in 𝑇𝑧 𝑋 and the orthogonal complement 𝑇𝑧 𝑋 ⟂
separately. Since the restriction of 𝜋 to 𝑋 is the identity, the derivative of 𝑦 → 𝑦 − 𝜋(𝑦) in the
direction of 𝑇𝑧 𝑋 acts trivially. Thus we get

𝑑𝐰𝑧 (ℎ) = 𝑑𝐯𝑧 (ℎ) for all ℎ ∈ 𝑇𝑧 𝑋.

In the direction orthogonal to 𝑇𝑧 𝑋, however, 𝑑𝐯𝑧 acts trivially and 𝜋 is constant along a fixed
line perpendicular to 𝑋. Thus we get

𝑑𝐰𝑧 (ℎ) = ℎ for all ℎ ∈ 𝑇𝑧 𝑋 ⟂ .

Since 𝑇𝑧 𝑋 and 𝑇𝑧 𝑋 ⟂ are orthogonal to each other and together span all of 𝑇𝑧 𝑋̄ 𝜀 , we can com-
pute the determinant of 𝑑𝐰𝑧 as the product

det 𝑑𝐰𝑧 = det(𝑑𝐰𝑧 )|𝑇𝑧 𝑋 ⋅ det(𝑑𝐰𝑧 )|𝑇𝑧 𝑋 ⟂ = det 𝑑𝐯𝑧 ⋅ 1 = det 𝑑𝐯𝑧 .

This proves the claim by Lemma 17.29. Hence, in particular, the index sums of 𝐰 and of 𝐯 are
equal. Moreover, since det 𝑑𝐰𝑧 = det 𝑑𝐯𝑧 ≠ 0, all zeros of 𝐰 are isolated by Lemma 17.29.
Thus we can apply Lemma 17.25 which shows that the index sum of 𝐰 is equal to the degree
of the Gauss map 𝑔. Thus the index sum of 𝐯 is equal to the degree of the Gauss map 𝑔 which
proves the theorem.

It remains to remove the assumption that the zeros of the vector field have to be nondegen-
erate.

Theorem 17.34 (Index sums of fields with degenerate zeros) Let 𝑋 ⊂ ℝ𝑛 be a

compact 𝑘-dimensional smooth manifold without boundary. Let 𝐯 be a vector field on
𝑋 with finitely many isolated zeros. Then there is a vector field 𝐰 on 𝑋 with finitely
many zeros all of which are nondegenerate such that the index sums of 𝐯 and 𝐰 are
equal.

Proof: First we assume that 𝐯 is a vector field defined on an open subset 𝑈 ⊂ ℝ𝑘 , and
assume that 𝑧 is the only zero of 𝐯. We choose a smooth bump function 𝜆 ∶ 𝑈 → [0, 1] which
has the value 1 on a small ball 𝔹𝜀′ around 𝑧 and has the value 0 outside a ball 𝔹𝜀 around 𝑧 with
slightly larger radius 𝜀. Since 𝐯 is a smooth map and has no other zeros than 𝑧, we can choose
a regular value 𝑦 of 𝐯. We then define a new vector field 𝐰 on 𝑈 by

𝐰(𝑢) = 𝐯(𝑢) − 𝜆(𝑢)𝑦.

352 17.6. Poincaré–Hopf Theorem - Euler characterstic
Then 𝐰(𝑢) can only be zero if 𝐯(𝑢) = 𝜆(𝑢)𝑦. Since 𝑧 is the only zero of 𝐯, our choice of 𝜆 implies
that 𝐰 does not have any zeros outside of 𝔹𝜀 . After possibly rescaling by multiplying with a
constant, we can choose 𝑦 small enough such that |𝐯(𝑢) − 𝜆(𝑢)𝑦| > 0 on 𝔹𝜀 ⧵ 𝔹𝜀′ . Moreover,
within 𝔹𝜀′ , we have 𝜆(𝑢) = 1, i.e., 𝐰(𝑢) = 𝐯(𝑢) − 𝑦 for all 𝑢 ∈ 𝔹𝜀′ . Thus, inside 𝔹𝜀′ we can only
have 𝐰(𝑢) if 𝐯(𝑢) = 𝑦. Since 𝑦 is a regular value, this implies that 𝑑𝐰𝑢 = 𝑑𝐯𝑢 ∶ 𝑇𝑢 𝑈 = ℝ𝑘 → ℝ𝑘
is an isomorphism at every point 𝑢 in the set 𝐯−1 (𝑦). This proves that 𝐰 only has nondegenerate
zeros. Since the closed ball 𝔹̄ 𝜀′ is compact, 𝐰 can only be finitely many zeros on 𝑈 .

Again, since 𝔹̄ 𝜀′ is compact, we can apply Lemma 17.25 to see that the index sum of 𝐯 and
the index sum of 𝐰 on 𝔹̄ 𝜀′ equals the degree of the Gauss map 𝑔 ∶ 𝜕 𝔹̄ 𝜀′ = 𝕊𝑘−1
𝜀′
→ 𝕊𝑘−1 , i.e.,

∑ ∑
ind𝑧 (𝐯) = deg(𝑔) = ind𝑧 (𝐰).
𝑧∈𝐯−1 (0) 𝑢∈𝐰−1 (0)

Thus, the index sums of 𝐯 and 𝐰 are equal.

Now we let 𝑋 ⊂ ℝ𝑛 and 𝐯 be as assumed in the theorem. Let 𝑧𝑖 be one of the finitely many
isolated zeros 𝑧1 , … , 𝑧𝑚 of 𝐯. Then we can choose a local parametrization 𝜙𝑖 ∶ 𝑈𝑖 → 𝑋 around
𝑧. Let 𝜆𝑖 ∶ 𝑈𝑖 → [0, 1] smooth bump function and 𝑦𝑖 ∈ ℝ𝑘 be a regular value of 𝜙∗𝑖 𝐯 as in the
first case above. Then we can form the new vector field 𝐰𝑖 on 𝑋 by setting

𝐰𝑖 (𝑥) ∶= 𝐯(𝑥) − 𝜆𝑖 (𝜙−1

𝑖 (𝑥))𝑑𝜙𝜙−1 (𝑥) (𝑦𝑖 ).
𝑖

Now we can check as above that 𝐰𝑖 has finitely many zeros and the zeros of 𝐰𝑖 which lie in the
image of 𝜙𝑖 are all nondegenerate. Moreover, the index sums of 𝐯 and 𝐰𝑖 are equal. Choosing
the local parametrizations around all the 𝑧𝑖 small enough so that they do not overlap, we can
perform this process for each zero and define a vector field 𝐰 on 𝑋 by

∑
𝑚
𝐰(𝑥) ∶= 𝐯(𝑥) − 𝜆𝑖 (𝜙−1
𝑖 (𝑥))𝑑𝜙𝜙−1 (𝑥) (𝑦𝑖 ).
𝑖
𝑖=1

The vector field 𝐰 satisfies the properties claimed in the theorem.

Now we can conclude that Theorem 17.33 and Theorem 17.34 together imply Theorem 17.22.

17.6 Poincaré–Hopf Theorem - Euler characterstic

Theorem 17.22 reduces the proof of Theorem 17.16 to finding at least one vector field on a
given manifold 𝑋 with index sum equal to 𝜒(𝑋). The idea is to find a suitable Morse function
𝑓 ∶ 𝑋 → ℝ such that the index sum of its gradient field equals the Euler characteristic of 𝑋.
We will now sketch how this can be achieved.

Given a smooth real-valued function 𝑓 ∶ 𝑋 → ℝ. By Lemma 17.31 we know that the

nondegenerate zeros of 𝐠𝐫𝐚𝐝(𝑓 ) are exactly the nondegenerate critical points of 𝑓 . We now
need to express the index of a zero of 𝐠𝐫𝐚𝐝(𝑓 ) in terms of the behavior of the critical points of
𝑓.
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 353

Definition 17.35 (Gradient vector field - index of nondegenerate critical points)

Let 𝑋 ⊂ ℝ𝑛 be a smooth manifold. Let 𝑓 ∶ 𝑋 → ℝ be a real-valued function on 𝑋 with
a nondegenerate critical point 𝑧 ∈ 𝑋. The index of the critical point 𝑧 is defined to
be the number of negative eigenvalues of the Hessian matrix of 𝑓 at 𝑧 computed in a
local coordinate system of 𝑋 around 𝑧. We denote the index by ind𝑧 (𝑓 ).

∙ One can show that the index is well-defined, i.e., the number of negative eigenvalues the
Hessian of 𝑓 at 𝑧 is the same for every local coordinate system.

Remark 17.36 (Index of a critical point and bilinear forms) Recall that the index of
a bilinear form 𝐵 on a vector space 𝑉 is defined to be maximal dimension of a vector
subspace of 𝑉 on which 𝐵 is negative definite. The Hessian matrix 𝐻𝑧 (𝑓 ) of a smooth
function 𝑓 ∶ 𝑋 → ℝ at 𝑧, computed in a local coordinate system around 𝑧 and with
respect to a suitable basis, is a symmetric matrix and hence defines a symmetric bilinear
form 𝐻𝑧 (𝑓 ) on 𝑇𝑧 𝑋 by setting 𝑤 → 𝑤𝑇 ⋅ 𝐻𝑧 (𝑓 ) ⋅ 𝑤. It then follows that the index of 𝑧
as a critical point equals the index of the Hessian at 𝑧 as a symmetric bilinear form.

Now we can relate the index of a critical point to the index of a zero as follows:

Lemma 17.37 (Gradient vector field - index of a nondegenerate zero) Let 𝑋 ⊂ ℝ𝑛

be a smooth 𝑘-dimensional manifold. Let 𝑓 ∶ 𝑋 → ℝ be a real-valued function on 𝑋
with a nondegenerate critical point 𝑧 ∈ 𝑋. Let 𝑠 = ind𝑧 (𝑓 ) be the index of 𝑓 at 𝑧. Then
the index of 𝐠𝐫𝐚𝐝(𝑓 ) at the zero 𝑧 is (−1)𝑠 , i.e., ind𝑧 (𝐠𝐫𝐚𝐝(𝑓 )) = (−1)𝑠 .

Proof: Let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization of 𝑋 around 𝑧 with 𝜙(0) = 𝑧. we

may choose 𝑈 small enough such that 𝑈 does not contain any other critical points of 𝑓 . By
definition, the index of 𝐠𝐫𝐚𝐝(𝑓 ) at 𝑧 is the index of 𝜙∗ 𝐠𝐫𝐚𝐝(𝑓 ) at 0. Thus, we may assume that
𝑋 = 𝑈 is an open subset of ℝ𝑘 and 0 is the only critical point of 𝑓 ∶ 𝑈 → ℝ. Moreover, by
Lemma 7.10 we can choose 𝜙 such that we may assume that 𝑓 is of the form

𝑓 (𝑥) = 𝑓 (0) − 𝑥21 − ⋯ − 𝑥2𝑠 + 𝑥2𝑠+1 + ⋯ + 𝑥2𝑘 (17.9)

for all 𝑥 ∈ 𝑋 where 𝑠 is the number of negative eigenvalues of the Hessian 𝐻(𝑓 )0 of 𝑓 at 0.
From Equation 17.9 we deduce that

𝐠𝐫𝐚𝐝(𝑓 )(𝑥) = (−2𝑥1 , … , −2𝑥𝑠 , 2𝑥𝑠+1 , … , 2𝑥𝑘 ) ∈ ℝ𝑘 .

Writing 𝐯 = 𝐠𝐫𝐚𝐝(𝑓 ) this shows that the induced map 𝐯̄ ∶ 𝕊𝑘−1 → 𝕊𝑘−1 is the composition of
the reflection of each of the first 𝑠 coordinate. Since the reflection of an individual coordinate
has degree −1, this implies that 𝐯̄ has degree (−1)𝑠 .

Theorem 17.38 (Gradient vector field - index sum) Let 𝑋 ⊂ ℝ𝑛 be a compact

smooth 𝑘-dimensional manifold. Let 𝑓 ∶ 𝑋 → ℝ be a Morse function on 𝑋, i.e, a
smooth real-valued function with only nondegenerate critical points. For 𝜆 ∈ ℤ, let 𝑐𝜆
denote the number of critical points with index 𝜆. Then the index sum of 𝐠𝐫𝐚𝐝(𝑓 ) is
354 17.6. Poincaré–Hopf Theorem - Euler characterstic
∑
equal to 𝜆 (−1)
𝜆𝑐 .
𝜆

Proof: Let 𝑧 be a zero of 𝐠𝐫𝐚𝐝(𝑓 ) and let 𝜆 be its index as a critical point. By Lemma 17.37,
𝑧 contributes with (−1)𝜆 to the index sum of 𝐠𝐫𝐚𝐝(𝑓 ). Hence, if there are 𝑐𝜆 many such zeros,
they contribute with (−1)𝜆 𝑐𝜆 to the index sum. Since the zeros of 𝐠𝐫𝐚𝐝(𝑓 ) are the critical points
∑
of 𝑓 , this shows that the sum 𝜆 (−1)𝜆 𝑐𝜆 is exactly the index sum of 𝑓 .

By Theorem 7.13 every smooth manifold admits a Morse function. Hence we can always
∑
find a Morse function and we know that the corresponding sum 𝜆 (−1)𝜆 𝑐𝜆 is independent of
our choice of function. It remains to relate this sum to the Euler characteristic, i.e., it remains
∑
to show 𝜆 (−1)𝜆 𝑐𝜆 = 𝜒(𝑋). The details of this identification are beyond the scope of these
notes. The idea, however, is based on the discussion of homotopy types at the beginning of
Section 7.2 where we sketched how a Morse function on 𝑋 helps us finding a cell complex
which is homotopy equivalent to 𝑋.

We now hint briefly at the general procedure and refer to [12, Part I] for the details. Let 𝑋
be a compact smooth 𝑘-dimensional manifold and let 𝑓 ∶ 𝑋 → ℝ be a Morse function. For
𝑎 ∈ ℝ, set
𝑋 𝑎 ∶= 𝑓 −1 ((−∞, 𝑎]) = {𝑥 ∈ 𝑋 ∶ 𝑓 (𝑥) ≤ 𝑎}.

Now let 𝑎1 < ⋯ < 𝑎𝑛 be real numbers such that 𝑋 𝑎𝑖 contains exactly 𝑖 critical points, and
𝑋 𝑎𝑛 = 𝑋. We then have ⊂ 𝑋 𝑎1 ⊂ … ⊂ 𝑋 𝑎𝑛 = 𝑋. Let 𝔹̄ 𝜆 denote the unit ball in ℝ𝜆 . We think
of 𝔹̄ 𝜆 as a 𝜆-dimensional cell. In 𝑋 𝑎𝑖 there is exactly one critical point which is not contained
in 𝑋 𝑎𝑖−1 . Let 𝜆𝑖 denote the index of this critical point. We then get

𝐻𝑗 (𝑋 𝑎𝑖 , 𝑋 𝑎𝑖−1 ; ℤ) ≅ 𝐻𝑗 (𝑋 𝑎𝑖−1 ∪ 𝔹̄ 𝜆𝑖 , 𝑋 𝑎𝑖−1 ; ℤ)

≅ 𝐻𝑗 (𝔹̄ 𝜆𝑖 , 𝔹𝜆𝑖 ; ℤ)
{
ℤ if 𝑗 = 𝜆𝑖
≅
0 otherwise

where the first isomorphism can be deduced from the idea we sketched at beginning of Sec-
tion 7.2 and the second follows from excision.
Now let 𝑏𝜆 (𝑋 𝑎𝑖 , 𝑋 𝑎𝑖−1 ) denote the relative Betti number, i.e., the rank of the abelian group
𝐻𝜆 (𝑋 𝑎𝑖 , 𝑋 𝑎𝑖−1 ; ℤ). We then get

∑
𝑘
𝜒(𝑋 𝑎𝑖 , 𝑋 𝑎𝑖−1 ) ∶= 𝑏𝜆 (𝑋 𝑎𝑖 , 𝑋 𝑎𝑖−1 ) = 𝑐𝜆
𝑖=1

where 𝑐𝜆 denotes the number of critical points of 𝑓 of index 𝜆. Since the Euler characteristic is
additive (see [12, Part I, §5]), this shows

∑
𝑘
∑
𝑘
𝜒(𝑋) = 𝜒(𝑋 𝑎𝑖 , 𝑋 𝑎𝑖−1 ) = 𝑐0 − 𝑐1 + 𝑐2 − + ⋯ ± 𝑐𝑘 = (−1)𝜆 𝑐𝜆 .
𝑖=1 𝑖=1

Together with Theorem 17.38 this shows that the index sum of the gradient field equals the
Euler characteristic and finishes the proof of Theorem 17.16.
Chapter 17. Vector Fields and the Poincaré–Hopf Index Theorem 355
17.7 Existence of vector fields with no zeros

Given a compact oriented smooth manifold 𝑋, it is a natural question whether we can find a
nowhere vanishing vector field 𝐯 on 𝑋. Theorem 17.16 provides a necessary condition: if 𝐯
has no zeros, then the index sum is zero and hence the Euler characteristic of 𝑋 must be zero
too. Hopf showed in [8] that this is also a sufficient condition:

Theorem 17.39 (Hopf - Existence of nowhere vanishing vector fields) Let 𝑋 be a

compact, connected smooth manifold. Then 𝑋 possesses a nowhere vanishing vector
field if and only if the Euler characteristic of 𝑋 is zero.

∙ The proof is based on the ideas we developed for the proof of the Hopf Degree Theo-
rem 16.1. It shows once more the power of Brouwer degree as an invariant.

Since smooth manifolds of odd dimension have Euler characteristic zero by Corollary 17.19,
we get the following consequence.

Corollary 17.40 (Odd dimensional manifolds have nowhere vanishing vector fields)
Let 𝑋 be a compact smooth manifold without boundary. If the dimension 𝑛 of 𝑋 is odd,
then 𝑋 possesses a nowhere vanishing vector field.

Proof of Theorem 17.39: First we assume that 𝑋 = ℝ𝑘 and that we have a vector field 𝐯
on 𝑋 with only finitely many zeros. Then the vector field is just a smooth map 𝐯 ∶ ℝ𝑘 → ℝ𝑘 .
Since 𝐯 has only finitely many zeros, there is a closed ball 𝔹̄ = 𝔹̄ 𝑘𝑟 containing all the zeros of 𝐯 in
its interior. Using the technique of the proof of Theorem 17.34 we can assume that the zeros of
𝐯 are nondegenerate. Moreover, we can assume that the origin in ℝ𝑘 is not a zero of 𝐯. Since
there are no zeros of 𝐯 on the boundary of 𝔹,̄ 𝐯 points either inward or outward at all boundary
points of 𝔹.̄ Since the index sum of 𝐯 is zero, we can replace 𝐯 with −𝐯 if 𝐯 points inwards.
Hence we can assume that 𝐯 points outward at every boundary point of 𝔹. ̄ Since 𝔹̄ is compact,
we can apply Lemma 17.25 and obtain that the index sum of 𝐯 equals the degree of the Gauss
map 𝑔 ∶ 𝜕 𝔹̄ = 𝕊𝑘−1
𝑟 → 𝕊𝑘−1 . By assumption, the index sum of 𝐯 is zero, and hence the degree
of 𝑔 is zero. Since 𝑔 and 𝐯̄ = 𝐯∕|𝐯| are homotopic, this show that 𝐯̄ has degree zero. Hence
the winding number of 𝐯|𝔹̄ around the origin is zero. Thus, 𝜕𝐯|𝔹̄ = 𝐯|𝜕𝔹̄ ∶ 𝜕 𝔹̄ → ℝ𝑘 ⧵ {0} is
homotopic to a constant map by Theorem 16.10. Hence

𝐯|ℝ𝑘 ⧵𝔹 ∶ ℝ𝑘 ⧵ 𝔹 → ℝ𝑘 ⧵ {0}

is a map to which we can apply Lemma 16.8. This implies that 𝐯|ℝ𝑘 ⧵𝔹 extends to a smooth map

𝐰 ∶ ℝ𝑘 → ℝ𝑘 ⧵ {0}

with 𝐰 = 𝐯 outside the compact space 𝔹. ̄ Hence we have constructed a vector field 𝐰 on ℝ𝑘
without zeros which equals 𝐯 outside a compact subset.

Now we let 𝑋 ⊂ ℝ𝑁 be a compact, oriented, connected 𝑘-dimensional smooth manifold

with 𝜒(𝑋) = 0. For 𝑘 = 1, the assertion follows from the classification of compact one-
manifolds in Theorem 10.17. Hence we may assume 𝑘 ≥ 2. As pointed out in Remark 17.32
356 17.7. Existence of vector fields with no zeros
there exists a vector field 𝐯 on 𝑋 with only finitely many nondegenerate zeros. Since 𝜒(𝑋) = 0,
Theorem 17.16 implies the index sum of 𝐯 is zero. Let 𝜙 ∶ 𝑈 → 𝑉 ⊂ 𝑋 be a local parametriza-
tion of 𝑋. By the Isotopy Theorem 11.10, we can find a diffeomorphism of 𝑋 to itself such that
𝑓 ∗ 𝐯 has finitely many nondegenerate zeros all of which lie in 𝑉 . Since 𝑉 is diffeomorphic to 𝑈
via 𝜙 and 𝑈 is diffeomorphic to ℝ𝑘 , we may assume that the zeros of 𝐯 all lie in an open subset
which is diffeomorphic to ℝ𝑘 . Thus, we can apply the case Euclidean space we considered
above to replace 𝐯 with a vector field that has no zeros. This proves the theorem.
A. Solutions to exercises

A.2 Smooth manifolds

A.2.1 Smooth maps and manifolds

Solution (Exercise 2.1) (a) We have remarked in the main text that 𝑓 is smooth,
since each component of 𝑓 is a polynomial. To get some more exercise, we could
calculate all partial derivatives (in all degrees) and check that they exist and are
continuous. So let us calculate the partial derivatives. We denote the two compo-
nents of 𝑓 by 𝑓1 (𝑥, 𝑦) = 𝑥2 − 𝑦2 and 𝑓2 (𝑥, 𝑦) = 2𝑥𝑦. The first partial derivatives
are
𝜕𝑓1 𝜕𝑓1 𝜕𝑓2 𝜕𝑓2
= 2𝑥, = −2𝑦, = 2𝑦, = 2𝑥.
𝜕𝑥 𝜕𝑦 𝜕𝑥 𝜕𝑦
All these functions are differentiable. Hence we can calculate the second deriva-
tives:
𝜕 2 𝑓1 𝜕 2 𝑓1 𝜕 2 𝑓1 𝜕 2 𝑓1
= 2, = 0, = 0, = −2,
𝜕𝑥𝜕𝑥 𝜕𝑥𝜕𝑦 𝜕𝑦𝜕𝑥 𝜕𝑦𝜕𝑦
𝜕 2 𝑓2 𝜕 2 𝑓2 𝜕 2 𝑓2 𝜕 2 𝑓2
= 0, = 2, = 2, = 0.
𝜕𝑥𝜕𝑥 𝜕𝑥𝜕𝑦 𝜕𝑦𝜕𝑥 𝜕𝑦𝜕𝑦
The second derivatives are again all differentiable and we see that the next deriva-
tives will all vanish. This implies that all further partial derivatives vanish and
therefor are differentiable. This shows that 𝑓 is smooth.

(b) We denote the two components of 𝑓 by 𝑓1 (𝑥, 𝑦) = 𝑥2 − 𝑦2 and 𝑓2 (𝑥, 𝑦) = 2𝑥𝑦. We

calculate the Jacobian 𝐽 (𝑓 )(𝑥, 𝑦) of 𝑓|𝑈 at a point 𝑝 = (𝑥, 𝑦) ∈ 𝑈 :
( 𝜕𝑓 𝜕𝑓1 ) ( )
1
𝜕𝑥
(𝑝) 𝜕𝑦
(𝑝) 2𝑥 −2𝑦
𝐽 (𝑓 )(𝑥, 𝑦) = 𝜕𝑓2 𝜕𝑓2 = .
(𝑝) (𝑝) 2𝑦 2𝑥
𝜕𝑥 𝜕𝑦

The determinant of the Jacobian at 𝑝 = (𝑥, 𝑦) is 4𝑥2 + 4𝑦2 . This is a positive real
number for every 𝑝 in 𝑈 = ℝ2 ⧵{(0, 0)}. This implies that the Jacobian is invertible
at every point in 𝑈 .

(c) Even though 𝑓|𝑈 has an invertible derivative at every point, it is not invertible itself.
For 𝑓|𝑈 is not injective. For example, 𝑓 (1, 1) = (0, 2) = 𝑓 (−1, −1).

357
358 A.2. Smooth manifolds

Solution (Exercise 2.2) (a) The assertion is true if 𝑋, 𝑌 and 𝑍 are open subsets
of ℝ𝑁 , ℝ𝑀 , ℝ𝐿 , respectively. For in this case this is just the Chain Rule from
Calculus. Now let 𝑥 ∈ 𝑋 and 𝑦 = 𝑓 (𝑥). Since 𝑔 is smooth at 𝑦, there is an
open subset 𝑉 ⊂ ℝ𝑀 with 𝑦 ∈ 𝑉 and a smooth map 𝐺 ∶ 𝑉 → ℝ𝐿 such that
𝐺|𝑉 ∩𝑌 = 𝑔|𝑉 ∩𝑌 . Since 𝑓 is smooth at 𝑥, there is an open subset 𝑈 ⊂ ℝ𝑁 with
𝑥 ∈ 𝑈 and a smooth map 𝐹 ∶ 𝑈 → ℝ𝑀 such that 𝐹|𝑈 ∩𝑋 = 𝑓|𝑈 ∩𝑋 . After replacing
𝑈 with 𝑈 ∩ 𝑓 −1 (𝑉 ) if necessary (to make sure that the image of 𝐹 lies in 𝑉 ), we
have
𝐺◦𝐹 = 𝑔◦𝑓 on 𝑋 ∩ 𝑈 , since 𝑓 (𝑋 ∩ 𝑈 ) ⊂ 𝑌 ∩ 𝑉 .
By the Chain Rule, we know that 𝐺◦𝐹 ∶ 𝑈 → ℝ𝐿 is smooth, since 𝑈 is open in
ℝ𝑁 . This shows that 𝑔◦𝑓 is smooth.

(b) By assumption 𝑓 and 𝑔 are smooth and have smooth inverses 𝑓 −1 and 𝑔 −1 respec-
tively. Since 𝑓 and 𝑔 are bijective, so is 𝑔◦𝑓 . By the previous point, both 𝑔◦𝑓 and
(𝑔◦𝑓 )−1 = 𝑓 −1 ◦𝑔 −1 are smooth.

Solution (Exercise 2.3) (a) The inverse is

𝑟𝑦
𝑔 ∶ ℝ𝑘 → 𝐵𝑟 , 𝑦 → √ .
𝑟2 + |𝑦|2
For we can calculate
𝑟 √ 2𝑟𝑦 2 √
𝑟𝑔(𝑦) 𝑟 +|𝑦| 𝑟2 𝑦 𝑟2 + |𝑦|2
𝑓 (𝑔(𝑦)) = √ =√ =√ ⋅√
𝑟2 − |𝑔(𝑦)|2 𝑟2 − 𝑟2|𝑟𝑦| 𝑟2 + |𝑦|2 𝑟2 (𝑟2 + |𝑦|2 − |𝑦|2 )
2

+|𝑦|2
=𝑦
and
𝑟 √ 2𝑟𝑥 2 √
𝑟𝑓 (𝑥) 𝑟 −|𝑥| 𝑟2 𝑥 𝑟2 − |𝑥|2
𝑔(𝑓 (𝑥)) = √ =√ =√ ⋅√
𝑟2 + 𝑟2|𝑟𝑥| 𝑟2 − |𝑥|2 𝑟2 (𝑟2 − |𝑥|2 + |𝑥|2 )
2
𝑟2 + 𝑓 (𝑥)|2
−|𝑥|2
= 𝑥.
Both 𝑓 and 𝑔 are smooth, since they are the composite of several smooth maps.
Hence the previous exercise shows that they are both smooth.
(b) Let 𝑥 ∈ 𝑋. By definition of a smooth manifold, there is an open subset 𝑉 ⊂ 𝑋
with 𝑥 ∈ 𝑉 , an open subset 𝑈 ⊂ ℝ𝑘 and a diffeomorphism 𝜙 ∶ 𝑈 → 𝑉 . Let
𝑢 ∈ 𝑈 be the inverse 𝜙−1 (𝑥). If 𝑢 is not the origin in ℝ𝑘 , we compose 𝜙 with the
translation 𝑇𝑢 ∶ ℝ𝑘 → ℝ𝑘 defined by 𝑇𝑢 (𝑦) = 𝑦 + 𝑢 which satisfies 𝑇𝑢 (0) = 𝑢. Note
that 𝑇𝑢 is a diffeomorphism, since it is invertible and both 𝑇𝑢 and its inverse 𝑇−𝑢
have the identity matrix as their Jacobian matrix at any point. Hence all higher
partial derivatives vanish and exist.
Thus after composing 𝜙 with 𝑇𝑢 , we can assume 𝜙(0) = 𝑥. Now it suffices to
choose a small enough radius 𝑟 such that 𝜙(𝐵𝑟𝑘 (0)) ⊂ 𝑉 . Then 𝜙|𝐵𝑘 (0) ∶ 𝐵𝑟𝑘 (0) →
𝑟
𝜙(𝐵𝑟𝑘 (0)) ⊂ 𝑋 is the desired local parametrization.
Appendix A. Solutions to exercises 359

(c) By the previous point, for every 𝑥 ∈ 𝑋 there is a diffeomorphism 𝜙 ∶ 𝐵𝑟𝑘 (0) → 𝑉
for some open subset 𝑉 ⊂ 𝑋 with 𝑥 ∈ 𝑉 . Now it suffices to precompose 𝜙 with
the diffeomorphism 𝑔 ∶ ℝ𝑘 → 𝐵𝑟𝑘 (0) of the first point in this exercise.

Solution (Exercise 2.4) Every linear map ℝ𝑛 → ℝ𝑚 is smooth, and the derivative
is equal to the map itself. Hence, given a 𝑘-dimensional vector subspace 𝑉 of ℝ𝑁 , it
suffices to choose a basis in 𝑉 to get a linear isomorphism 𝜙 ∶ 𝑉 → ℝ𝑘 . This map serves
as a parametrization, since it is a diffeomorphism.
Now given a linear map 𝑓 ∶ 𝑉 → ℝ𝑚 , the composite 𝜙◦𝑓 ∶ ℝ𝑘 → ℝ𝑚 is linear and
therefore smooth. Since 𝜙 is a diffeomorphism, this implies that 𝑓 must be smooth too.

Solution (Exercise 2.5) (a) Let 𝑎 > 0 be a real number. We want to show that the
subset
𝐻𝑎 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 − 𝑧2 = 𝑎} ⊂ ℝ3
is a 2-dimensional manifold. Hence we need to find local parametrizations. First,
there is a diffeomorphism
( √ )
𝜙 ∶ ℝ2 ⧵ 𝐵𝑎 ((0, 0)) → 𝐻 ∩ {𝑧 > 0}, (𝑥, 𝑦) → 𝑥, 𝑦, 𝑥2 + 𝑦2 − 𝑎

where 𝐵𝑎 ((0, 0)) denotes the closed ball of radius 𝑎 around the origin, i.e.,

𝐵√𝑎 ((0, 0)) = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 + 𝑦2 ≤ 𝑎}.

Note that 𝐻 ∩{𝑧 > 0} is an open subset of 𝐻, because it is equal to the intersection
of 𝐻 with the open subset {𝑧 > 0} ⊂ ℝ3 . The inverse to 𝜙 is the projection map
𝜙−1 ∶ 𝐻 ∩ {𝑧 > 0} → ℝ2 , (𝑥, 𝑦, 𝑧) → (𝑥, 𝑦).
This map is smooth, since it can be extended to a smooth map on the whole of
ℝ3 to ℝ2 . We know that 𝜙 is smooth, since its component functions are infinitely
often differentiable functions in each variable on the open subset ℝ2 ⧵ 𝐵𝑎 ((0, 0)).
We can for example calculate the Jacobian matrix in the standard basis at a point
( )
3 2 𝜕𝜙1 ∕𝜕𝑥 𝜕𝜙2 ∕𝜕𝑥 𝜕𝜙3 ∕𝜕𝑥
𝑑𝜙𝑥 ∶ ℝ → ℝ ,
𝜕𝜙1 ∕𝜕𝑦 𝜕𝜙2 ∕𝜕𝑦 𝜕𝜙3 ∕𝜕𝑦
( )
1 0 𝑥(𝑥2 + 𝑦2 − 𝑎)−1∕2
= .
0 1 𝑦(𝑥2 + 𝑦2 − 𝑎)−1∕2

On the open set ℝ2 ⧵ 𝐵𝑎 ((0, 0)), the entries of this matrix are continuously differ-
entiable functions.
The local parametrization for 𝐻 ∩ {𝑧 < 0} is similarly given by
( √ )
𝜙 ∶ ℝ2 ⧵ 𝐵𝑎 ((0, 0)) → 𝐻 ∩ {𝑧 < 0}, (𝑥, 𝑦) → 𝑥, 𝑦, − 𝑥2 + 𝑦2 − 𝑎 .

It remains to cover the points in 𝐻 ∩ {𝑧 = 0}. We are going to cover those points
by the following four open sets together with local parametrizations:
( √ )
𝐵√𝑎 ((0, 0)) → 𝐻 ∩ {𝑥2 + 𝑧2 < 𝑎}, (𝑥, 𝑧) → 𝑥, 𝑧2 − 𝑥2 + 𝑎, 𝑧
360 A.2. Smooth manifolds
√ ( )
𝐵√𝑎 ((0, 0)) → 𝐻 ∩ {𝑥2 + 𝑧2 < 𝑎}, (𝑥, 𝑧) → 𝑥, − 𝑧2 − 𝑥2 + 𝑎, 𝑧
(√ )
𝐵√𝑎 ((0, 0)) → 𝐻 ∩ {𝑦2 + 𝑧2 < 𝑎}, (𝑦, 𝑧) → 𝑧2 − 𝑦2 + 𝑎, 𝑦, 𝑧
( √ )
𝐵√𝑎 ((0, 0)) → 𝐻 ∩ {𝑦2 + 𝑧2 < 𝑎}, (𝑦, 𝑧) → − 𝑧2 − 𝑦2 + 𝑎, 𝑦, 𝑧 .

(b) If 𝑎 = 0, then the set

𝐻0 = {(𝑥, 𝑦, 𝑧) ∈ ℝ3 ∶ 𝑥2 + 𝑦2 − 𝑧2 = 0} ⊂ ℝ3

is not a manifold, since there is no local parametrization around the point

(0, 0, 0) ∈ 𝐻. For, assume there was such a local parametrization 𝜙 ∶ 𝑈 → 𝑉 with
both 𝑈 ⊆ ℝ2 and (0, 0, 0) ∈ 𝑉 ⊆ 𝐻 open. We can assume 𝜙((0, 0)) = (0, 0, 0).
(Remember that 𝑉 being open in 𝐻 means that there is an open subset 𝑉̃ ⊆ ℝ3 with
𝑉 = 𝐻 ∩ 𝑉̃ . In particular, 𝑉̃ must contain a small open ball 𝐵𝑟 ((0, 0, 0)) around
the origin.) Then removing the point (0, 0, 0) from 𝐻 splits 𝐻 and therefore 𝑉
into two disjoint connected components. But 𝑈 ⧵ {(0, 0)} ⊂ ℝ2 is connected.
Hence 𝜙 cannot be a diffeomorphism (because if it was, 𝜙|𝑈 ⧵{(0,0)} would also be a
diffeomorphism and that would imply that 𝑉 ⧵ {(0, 0, 0)} were connected as well;
a contradiction).

Solution (Exercise 2.6) For 0 < 𝑏 < 𝑎, let 𝑇 (𝑎, 𝑏) denote the set of points in ℝ3 at
distance 𝑏 from the circle of radius 𝑎 in the 𝑥𝑦-plane. We can parametrize these points
as follows: First the points in the 𝑥𝑦-plane which lie on the circle of radius 𝑎 satisfy

(𝑎 cos 𝑡, 𝑎 sin 𝑡, 0) for 𝑡 ∈ [0, 2𝜋).

A point in the plane in the direction of a fixed point (𝑎 cos 𝑡, 𝑎 sin 𝑡, 0) which lies on the
circle of radius 𝑏 around the point (𝑎 cos 𝑡, 𝑎 sin 𝑡, 0) has coordinates (𝑎+𝑏 cos 𝑠) cos 𝑡, (𝑎+
𝑏 cos 𝑠) sin 𝑡, 𝑏 sin 𝑠) with 𝑠 ∈ [0, 2𝜋). For we have

| ((𝑎 + 𝑏 cos 𝑠) cos 𝑡, (𝑎 + 𝑏 cos 𝑠) sin 𝑡, 𝑏 sin 𝑠) − (𝑎 cos 𝑡, 𝑎 sin 𝑡, 0) |2

=| ((𝑏 cos 𝑠) cos 𝑡, (𝑏 cos 𝑠) sin 𝑡, 𝑏 sin 𝑠) |2
=𝑏2 cos2 𝑠(cos2 𝑡 + sin2 𝑡) + 𝑏2 sin2 𝑠 = 𝑏2 .

Hence the points on 𝑇 (𝑎, 𝑏) ⊂ ℝ3 are given by

𝑇 (𝑎, 𝑏) = {((𝑎 + 𝑏 cos 𝑠) cos 𝑡, (𝑎 + 𝑏 cos 𝑠) sin 𝑡, 𝑏 sin 𝑠) ∶ 𝑠, 𝑡 ∈ [0, 2𝜋)}.

The points on 𝑆 1 × 𝑆 1 ⊂ ℝ4 are given by

( )
𝑆 1 × 𝑆 1 = { 𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ∈ ℝ4 ∶ 𝑥21 + 𝑥22 = 𝑥23 + 𝑥24 = 1}
= {(cos 𝑡, sin 𝑡, cos 𝑠, sin 𝑠) ∈ ℝ4 ∶ 𝑡 ∈ [0, 2𝜋), 𝑠 ∈ [0, 2𝜋)}.

Now it is clear how we can define a continuous map

𝜙 ∶ 𝑆 1 × 𝑆 1 → 𝑇 (𝑎, 𝑏),
(cos 𝑡, sin 𝑡, cos 𝑠, sin 𝑠) → ((𝑎 + 𝑏 cos 𝑠) cos 𝑡, (𝑎 + 𝑏 cos 𝑠) sin 𝑡, 𝑏 sin 𝑠) .
Appendix A. Solutions to exercises 361

In order to check that 𝜙 is smooth, we use the coordinates of ℝ4 again. Then 𝜙 is

given by
( ) ( )
𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 → (𝑎 + 𝑏𝑥3 )𝑥1 , (𝑎 + 𝑏𝑥3 )𝑥2 , 𝑏𝑥4 .

Its derivative in the standard basis at a point 𝑝 = (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) is then given by

⎛𝑎 + 𝑏𝑥3 0 𝑏𝑥1 0⎞
4 3
𝑑𝜙𝑝 ∶ ℝ → ℝ , 𝑑𝜙𝑝 = ⎜ 0 𝑎 + 𝑏𝑥3 𝑏𝑥2 0⎟ .
⎜ ⎟
⎝ 0 0 0 𝑏⎠

Since all partial derivatives are smooth maps, 𝜙 is a smooth map.

Its inverse is the map 𝜓 ∶ ℝ3 → ℝ4 defined by
√
⎛ ⎞
( ) ⎜ 𝑦 𝑦 𝑦21 + 𝑦22 − 𝑎 ⎟
𝑦1 , 𝑦2 , 𝑦3 , → ⎜ √ 1 ,√ 2 , , 𝑦3 ∕𝑏⎟ .
⎜ 𝑦2 + 𝑦2 2 + 𝑦2 𝑏 ⎟
𝑦
⎝ 1 2 1 2 ⎠

The image of 𝜓 lies in 𝑆 1 × 𝑆 1 since

2 2
⎛ ⎞ ⎛ ⎞
⎜ 𝑦1 ⎟ ⎜ 𝑦2 ⎟ 𝑦21 + 𝑦22
⎜√ ⎟ + ⎜√ ⎟ = 2 = 1.
⎜ 𝑦2 + 𝑦2 ⎟ ⎜ 𝑦2 + 𝑦2 ⎟ 𝑦1 + 𝑦22
⎝ 1 2⎠ ⎝ 1 2⎠

To check the other equation, we recall that 𝑦1 , 𝑦2 and 𝑦3 are connected by the condition
of being on 𝑇 (𝑎, 𝑏) which means
2 2
⎛ ⎞ ⎛ ⎞
⎜ 𝑎𝑦1 ⎟ ⎜ 𝑎𝑦2 ⎟ 2 2
⎜𝑦1 − √ ⎟ + ⎜𝑦2 − √ ⎟ + 𝑦3 = 𝑏
⎜ 𝑦21 + 𝑦22 ⎟ ⎜ 𝑦21 + 𝑦22 ⎟
⎝ ⎠ ⎝ ⎠
(√ )2
⇐⇒ 𝑦21 + 𝑦22 − 𝑎 + 𝑦23 = 𝑏2
√
⇐⇒ 𝑦21 + 𝑦22 + 𝑦23 + 𝑎2 − 2𝑎 𝑦21 + 𝑦22 = 𝑏2 .

Now we can calculate

√ 2 √
⎛ 2 2 ⎞
⎜ 𝑦1 + 𝑦2 − 𝑎 ⎟ 𝑦21 + 𝑦22 + 𝑦23 + 𝑎2 − 2𝑎 𝑦21 + 𝑦22
2
⎜ ⎟ + (𝑦3 ∕𝑏) = = 1.
⎜ 𝑏 ⎟ 𝑏2
⎝ ⎠

Hence the image of (𝑦1 , 𝑦2 , 𝑦3 ) does lie on 𝑆 1 × 𝑆 1 .

We can easily check that 𝜓◦𝜙 is the identity. For example, using 𝑦21 + 𝑦22 = (𝑎 + 𝑏𝑥3 )2 ,
we get
( )
(𝑎 + 𝑏𝑥3 )𝑥1 (𝑎 + 𝑏𝑥3 )𝑥2 (𝑎 + 𝑏𝑥3 ) − 𝑎 𝑏𝑥4
𝜓(𝜙(𝑥)) = √ ,√ , , = (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ),
(𝑎 + 𝑏𝑥 )2 (𝑎 + 𝑏𝑥 )2 𝑏 𝑏
3 3
362 A.2. Smooth manifolds
√ √
and, setting 𝑦12 ∶= 𝑦21 + 𝑦22 ,
(( √ ) ( √ ) )
𝑦12 − 𝑎 𝑦1 𝑦12 − 𝑎 𝑦2 𝑏𝑦3
𝜙(𝜓(𝑦)) = 𝑎+𝑏 √ , 𝑎+𝑏 √ ,
𝑏 𝑦12 𝑏 𝑦12 𝑏
=(𝑦1 , 𝑦2 , 𝑦3 ),

It remains to check that 𝜓 is smooth. The derivative of 𝜓 in the standard basis at a

point 𝑞 = (𝑦1 , 𝑦2 , 𝑦3 ) is then given by

⎛ 𝑦22 𝑦 𝑦 ⎞
⎜ (𝑦21 +𝑦22 )3∕2 − (𝑦2 +𝑦1 22)3∕2 0 ⎟
1 2
⎜ 𝑦1 𝑦2 𝑦21 ⎟
𝑑𝜓𝑞 ∶ ℝ3 → ℝ4 , 𝑑𝜓𝑞 = ⎜− (𝑦21 +𝑦22 )3∕2 (𝑦1 +𝑦22 )3∕2
2 0 ⎟.
⎜ 𝑦1 𝑦2 ⎟
⎜ 𝑏(𝑦2 +𝑦2 )1∕2 𝑏(𝑦21 +𝑦22 )1∕2
0 ⎟
⎜ 1
0
2
0 1∕𝑏⎟⎠
⎝

Since 𝑏 < 𝑎 we know 𝑦21 + 𝑦22 ≠ 0 and all partial derivatives are continuous smooth
functions. Hence 𝜓 is smooth. This proves that 𝜙 is a global diffeomorphism

𝜙 ∶ 𝑆 1 × 𝑆 1 → 𝑇 (𝑎, 𝑏) for all 0 < 𝑏 < 𝑎.

Solution (Exercise 2.7) (a) Let 𝑁 = (0, … , 0, 1) ∈ 𝑆 𝑘 be the north pole on the
𝑘-dimensional sphere. The stereographic projection 𝜙−1𝑁
from 𝑆 𝑘 ⧵ {𝑁} onto ℝ𝑘
is the map which sends a point 𝑝 to the point at which the line through 𝑁 and 𝑝
intersects the subspace in ℝ𝑘+1 defined by 𝑥𝑘+1 = 0. In order to get from 𝑁 to 𝑝,
we walk in the direction of the vector

𝑣 ∶= 𝑝 − 𝑁 = (𝑥1 , … , 𝑥𝑘+1 − 1).

To find 𝜙−1
𝑁
(𝑝) we need to find the real number 𝜆 such that the (𝑘 + 1)-st coordinate
of 𝑁 + 𝜆 ⋅ 𝑣 is 0. Hence we need to solve
1
1 + 𝜆(𝑥𝑘+1 − 1) = 0 ⇐⇒ 𝜆 = .
1 − 𝑥𝑘+1

Thus
1
𝜙−1
𝑁 (𝑥1 , … , 𝑥𝑘+1 ) = (𝑥 , … , 𝑥𝑘 ).
1 − 𝑥𝑘+1 1

(b) We calculate the inverse 𝜙𝑁 : Given the point 𝑥 = (𝑥1 , … , 𝑥𝑘 ) ∈ ℝ𝑘 . To find its
image under 𝜙𝑁 , we walk from 𝑁 in the direction of the vector 𝑤 = 𝑥 − 𝑁 until
we reach the sphere, i.e. we need to find 𝜆 ∈ ℝ such that

𝑁 + 𝜆 ⋅ 𝑤 = (𝜆𝑥1 , … , 𝜆𝑥𝑘 , 1 − 𝜆)
Appendix A. Solutions to exercises 363

lies on 𝕊𝑘 . Hence we need to solve

𝜆2 (𝑥21 + ⋯ + 𝑥2𝑘 ) + (1 − 𝜆)2 = 1
⇐⇒ 𝜆2 |𝑥|2 + 1 − 2𝜆 + 𝜆2 = 1
⇐⇒ 𝜆2 (1 + |𝑥|2 ) − 2𝜆 = 0
2
⇒𝜆 = (since 0 is not a valid solution).
1 + |𝑥|2
Thus
1 ( )
𝜙𝑁 (𝑥) = 2𝑥1 , … , 2𝑥𝑘 , |𝑥|2 − 1 .
1 + |𝑥| 2

We need to check that both 𝜙𝑆 and 𝜙−1 𝑆

are smooth. All entries in 𝜙𝑁 are smooth
functions (they are polynomials on the domain and these are infinitely often dif-
ferentiable functions in each variable). This implies that 𝜙𝑁 is smooth. But let us
calculate the derivative anyway since we are going to use it later:
⎛1 + |𝑥|2 − 𝑥21 −2𝑥1 𝑥2 −2𝑥1 𝑥3 … −2𝑥1 𝑥𝑘 ⎞
⎜ −2𝑥 𝑥 1 + |𝑥| 2 − 𝑥2 −2𝑥 𝑥 … −2𝑥2 𝑥𝑘 ⎟
2 ⎜ 2 1 2 2 3 ⎟
𝑑(𝜙𝑁 )𝑥 = ⎜ ⋮ ⋮ ⋮ ⋮ ⋮ ⎟.
(1 + |𝑥|2 )2 ⎜
−2𝑥𝑘 𝑥1 … … … 1 + |𝑥|2 − 𝑥2𝑘 ⎟
⎜ ⎟
⎝ 2𝑥1 2𝑥2 … … 2𝑥𝑘 ⎠
All entries are smooth functions on the domain and 𝜙𝑁 is smooth. Similarly for
𝜙−1
𝑁
, all entries are smooth functions, since 𝑥𝑘+1 ≠ 1. Its derivative is given by

⎛ 1 0 … 0
𝑥1 ⎞
⎜ 1−𝑥𝑘+1 1
(1−𝑥𝑘+1 )2 ⎟
𝑥2
⎜ 0 … 0 ⎟
𝑑(𝜙−1
𝑁 )𝑥 = ⎜
1−𝑥𝑘+1 (1−𝑥𝑘+1 )2 .
⎟
⎜ ⋮ ⋮ … ⋮ ⋮ ⎟
⎜ 0 0 … 1 𝑥𝑘 ⎟
⎝ 1−𝑥𝑘+1 (1−𝑥𝑘+1 )2 ⎠

(c) The formulae for the projection from the south pole 𝑆 = (0, … , 0, −1) ∈ 𝕊𝑘 are
similar:
1
𝜙−1
𝑆 (𝑥1 , … , 𝑥𝑘+1 ) = 1 + 𝑥 (𝑥1 , … , 𝑥𝑘 ).
𝑘+1
and, while we get the same 𝜆,
1 ( 2
)
𝜙𝑆 (𝑥) = 𝑆 + 𝜆(𝑥 − 𝑆) = 2𝑥1 , … , 2𝑥𝑘 , 1 − |𝑥| .
1 + |𝑥|2
To check that both 𝜙𝑆 and 𝜙−1 𝑆
are both smooth is completely analogous. Since
all points are covered by these two parametrizations, 𝕊𝑘 is a smooth manifold. We
note that the formula for the derivative of 𝜙−1
𝑆
is given by

⎛ 1 0 … 0
−𝑥1 ⎞
⎜ 1+𝑥𝑘+1 1
(1+𝑥𝑘+1 )2 ⎟
−𝑥2
⎜ 0 … 0 ⎟
𝑑(𝜙−1
𝑆 )𝑥 = ⎜
1+𝑥𝑘+1 (1+𝑥𝑘+1 )2 .
⎟
⎜ ⋮ ⋮ … ⋮ ⋮ ⎟
⎜ 0 0 … 1 −𝑥𝑘 ⎟
⎝ 1+𝑥𝑘+1 (1+𝑥𝑘+1 )2 ⎠
364 A.2. Smooth manifolds

Solution (Exercise 2.8) (a) For (𝑧0 , 𝑧1 ) ∈ 𝕊3 , we need to check 𝜋(𝑧0 , 𝑧1 ) ∈ 𝕊2 .

To do this, we recall that |𝑧| = 𝑧𝑧̄ and 𝑧̄ = 𝑧 for any complex number 𝑧. Now we
compute:

( )2
(2𝑧0 𝑧̄ 1 ) ⋅ (2𝑧̄ 0 𝑧1 ) + |𝑧0 |2 − |𝑧1 |2
= 4|𝑧0 |2 |𝑧1 |2 + |𝑧0 |4 − 2|𝑧0 |2 |𝑧1 |2 + |𝑧1 |4
( )2
= |𝑧0 |2 + |𝑧1 |2
=1

where the final step uses that (𝑧0 , 𝑧1 ) ∈ 𝑆 3 .

(b) First we assume 𝜋(𝑧0 , 𝑧1 ) = 𝜋(𝑤0 , 𝑤1 ): Then we get

( ) ( )
2𝑤0 𝑤̄ 1 , |𝑤0 |2 − |𝑤1 |2 = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2
⇐⇒ 𝑤0 𝑤̄ 1 = 𝑧0 𝑧̄ 1 and |𝑤0 |2 − |𝑤1 |2 = |𝑧0 |2 − |𝑧1 |2 .

We have |𝑤0 |2 − |𝑤1 |2 = |𝑧0 |2 − |𝑧1 |2 and |𝑤0 |2 + |𝑤1 |2 = 1 = |𝑧0 |2 + |𝑧1 |2 .
Putting these together implies

𝑧0 𝑧̄ 0 = |𝑧0 |2 = |𝑤0 |2 = 𝑤0 𝑤̄ 0 and 𝑧1 𝑧̄ 1 = 𝑧21 = 𝑤21 = 𝑤1 𝑤̄ 1 .

Remembering that neither of the numbers can be zero, by rewriting these equations
we get
𝑤0 𝑧̄ 𝑤 𝑧̄
= 0 and 1 = 1 .
𝑧0 𝑤̄ 0 𝑧1 𝑤̄ 1
Hence, looking at the left-hand side, there is a complex number 𝛼 ≠ 0 such that
𝑤0 𝑧̄ 1
𝛼= = 0 = , and thus 𝑤0 = 𝛼𝑧0 with 𝛼 𝛼̄ = 1.
𝑧0 𝑤̄ 0 𝛼̄
𝑤0 𝑧̄ 1
On the other hand, we also know 𝑧0
= 𝑤̄ 1
. Combining these equations yields

𝑤0 𝑧̄ 𝑤
𝛼= = 1 = 1 , and thus 𝑤1 = 𝛼𝑧1 .
𝑧0 𝑤̄ 1 𝑧1

This shows that the desired 𝛼 with 𝛼 𝛼̄ = 1 exists.

Now we assume that (𝑤0 , 𝑤1 ) = (𝛼𝑧0 , 𝛼𝑧1 ) with |𝛼|2 = 𝛼 𝛼̄ = 1: Then we compute
( )
𝜋(𝑤0 , 𝑤1 ) = 2𝑤0 𝑤̄ 1 , |𝑤0 |2 − |𝑤1 |2
( )
= 2𝛼𝑧0 𝛼̄ 𝑧̄ 1 , |𝛼|2 |𝑧0 |2 − |𝛼|2 |𝑧1 |2
( )
= 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2
= 𝜋(𝑧0 , 𝑧1 ).

(c) Let 𝑝 ∈ 𝕊2 be a fixed point and fix a point (𝑧0 , 𝑧1 ) in 𝕊3 with 𝜋(𝑧0 , 𝑧1 ) = 𝑝. By the
previous point, we have that the points in 𝜋 −1 (𝑝) are parametrized by the complex
Appendix A. Solutions to exercises 365

number 𝛼 with |𝛼|2 = 1. The latter condition means 𝛼 ∈ 𝕊1 ⊂ ℂ. Hence we get a

bijective map
𝑓 ∶ 𝕊1 → 𝜋 −1 (𝑝), 𝛼 → (𝛼𝑧0 , 𝛼𝑧1 ).
Since this map just consists of multiplication with nonzero complex numbers, we
can conclude that it is a smooth map where we consider 𝕊1 ⊂ ℂ = ℝ2 and 𝜋 −1 (𝑝) ⊂
ℝ4 as subsets in real Euclidean space. The previous point implies that 𝑓 is onto.
Since not both of 𝑧0 and 𝑧1 can be zero at the same time, we see that 𝑓 is one-
to-one. For, say 𝑧0 ≠ 0, then 𝛼𝑧0 = 𝛽𝑧0 implies 𝛼 = 𝛽. The inverse is given
by sending (𝑤0 , 𝑤1 ) ∈ 𝜋 −1 (𝑝) to 𝛼 ∈ 𝕊1 such that 𝑤0 = 𝛼𝑧0 and 𝑤1 = 𝛼𝑧1 . In
every open subset of 𝜋 −1 (𝑝) ⊂ 𝕊3 where 𝑧0 ≠ 0, the map 𝑤0 → 𝑤0 𝑧0 = 𝛼 is
smooth. Similarly, in every open subset of 𝜋 −1 (𝑝) ⊂ 𝕊3 where 𝑧1 ≠ 0, the map
𝑤1 → 𝑤1 𝑧1 = 𝛼 is smooth. Hence 𝑓 −1 is a smooth map as well, and 𝑓 is a
diffeomorphism.

A.2.2 Tangent spaces

Solution (Exercise 2.9) We can choose any basis of 𝑉 to define a linear isomorphism
𝜙 ∶ ℝ𝑘 → 𝑉 which is a diffeomorphism. Given a point 𝑥 ∈ 𝑉 , we modify 𝜙 by adding
𝑥 and get a new diffeomorphism (not linear anymore!)

𝜓 ∶ ℝ𝑘 → 𝑉 , 𝑤 → 𝜙(𝑤) + 𝑥.

The derivative 𝑑𝜓0 of 𝜓 at 0 is just 𝜙 ∶ ℝ𝑘 → ℝ𝑁 (independent of 𝑥). The tangent space

of 𝑉 at 𝑥 is by our definition the image of 𝑑𝜓0 in ℝ𝑁 , which is equal to the image of 𝜙
in ℝ𝑁 which is by definition of 𝜙 equal to 𝑉 .

Solution (Exercise 2.10) We only answer the question about the tangent space of
𝕋 (𝑎, 𝑏) in ℝ3 . For all points apart from (𝑎 + 𝑏, 0, 0) we can parametrize 𝕋 (𝑎, 𝑏) ⊂ ℝ3 by

𝜙 ∶ (0, 2𝜋) × (0, 2𝜋) → 𝑇 (𝑎, 𝑏)

(𝑠, 𝑡) → ((𝑎 + 𝑏 cos 𝑠) cos 𝑡, (𝑎 + 𝑏 cos 𝑠) sin 𝑡, 𝑏 sin 𝑠).

The derivative of 𝜙 at (𝑠, 𝑡) is

⎛−𝑏 sin 𝑠 cos 𝑡 −(𝑎 + 𝑏 cos 𝑠) sin 𝑡⎞

2 3
𝑑𝜙(𝑠,𝑡) ∶ ℝ → ℝ , 𝑑𝜙(𝑠,𝑡) = ⎜ −𝑏 sin 𝑠 sin 𝑡 (𝑎 + 𝑏 cos 𝑠) cos 𝑡 ⎟ .
⎜ ⎟
⎝ 𝑏 cos 𝑠 0 ⎠

The tangent space to 𝑇 (𝑎, 𝑏) at the point 𝜙(𝑠, 𝑡) is 𝑑𝜙(𝑠,𝑡) (ℝ2 ) ⊂ ℝ3 .

Let us check that the column vectors of the matrix 𝑑𝜙(𝑠,𝑡) , and hence the whole tangent
space, is orthogonal to the vector pointing from the center of the circle with radius 𝑏 to
𝜙(𝑠, 𝑡). The center point is (𝑎 cos 𝑡, 𝑎 sin 𝑡, 0, and hence the vector we need to look at is
366 A.2. Smooth manifolds
(𝑏 cos 𝑠 cos 𝑡, 𝑏 cos 𝑠 sin 𝑡, 𝑏 sin 𝑠). We calculate the two scalar products:

⎛−𝑏 sin 𝑠 cos 𝑡⎞

(𝑏 cos 𝑠 cos 𝑡, 𝑏 cos 𝑠 sin 𝑡, 𝑏 sin 𝑠) ⋅ ⎜ −𝑏 sin 𝑠 sin 𝑡 ⎟
⎜ ⎟
⎝ 𝑏 cos 𝑠 ⎠
= − 𝑏2 cos 𝑠 sin 𝑠 cos2 𝑡 − 𝑏2 cos 𝑠 sin 𝑠 sin2 𝑡 + 𝑏2 sin 𝑠 cos 𝑠
=𝑏2 cos 𝑠 sin 𝑠(− cos2 𝑡 − sin2 𝑡 + 1)
=𝑏2 cos 𝑠 sin 𝑠(−1 + 1) = 0

and

⎛−(𝑎 + 𝑏 cos 𝑠) sin 𝑡⎞

(𝑏 cos 𝑠 cos 𝑡, 𝑏 cos 𝑠 sin 𝑡, 𝑏 sin 𝑠) ⋅ ⎜ (𝑎 + 𝑏 cos 𝑠) cos 𝑡 ⎟
⎜ ⎟
⎝ 0 ⎠
= − 𝑏(𝑎 + 𝑏 cos 𝑠) cos 𝑠 sin 𝑡 cos 𝑡𝑏(𝑎 + 𝑏 cos 𝑠) cos 𝑠 sin 𝑡 cos 𝑡 + 0
=0.

In order to cover also the point (𝑎 + 𝑏, 0, 0) it suffices to rotate our parametrization by the
angle 𝜋 in the 𝑥𝑦−plane and use the diffeomorphism

𝜙 ∶ (0, 2𝜋) × (0, 2𝜋) → 𝕋 (𝑎, 𝑏)

(𝑠, 𝑡) → ((−𝑎 + 𝑏 cos 𝑠) cos 𝑡, (−𝑎 + 𝑏 cos 𝑠) sin 𝑡, 𝑏 sin 𝑠)

which covers (𝑎 + 𝑏, 0, 0) for (𝑠, 𝑡) = (𝜋, 𝜋).

√
Solution (Exercise 2.11) Around the point ( 𝑎, 0, 0) on 𝐻𝑎 we can choose the local
parametrization
(√ )
𝜙 ∶ 𝐵√𝑎 ((0, 0)) → 𝐻 ∩ {𝑦2 − 𝑧2 < 𝑎}, (𝑦, 𝑧) → 𝑧2 − 𝑦2 + 𝑎, 𝑦, 𝑧 .
The derivative in the standard basis at a point (𝑦, 𝑧) is the linear map
⎛− √ 𝑦 √
𝑧 ⎞
⎜ 𝑧2 −𝑦2 +𝑎 𝑧2 −𝑦2 +𝑎 ⎟
𝑑𝜙(𝑦,𝑧) ∶ ℝ2 → ℝ3 , 𝑑𝜙(𝑦,𝑧) =⎜ 1 0 ⎟.
⎜ 0 1 ⎟
⎝ ⎠
Hence at a point (𝑢, 𝑣, 𝑤) ∈ 𝐻𝑎 the image of the standard basis of ℝ2 is
⎛−𝑣∕𝑢⎞ ⎛𝑤∕𝑢⎞
⎜ 1 ⎟ and ⎜ 0 ⎟
⎜ ⎟ ⎜ ⎟
⎝ 0 ⎠ ⎝ 1 ⎠
√
where we write 𝑢 = 𝑤2 − 𝑣2 + 𝑎). Hence the tangent space at 𝑇(𝑢,𝑣,𝑤) (𝐻𝑎 ) is spanned
by these two vectors.
√
For (𝑢, 𝑣, 𝑤) = ( 𝑎, 0, 0) we get that 𝑇(√𝑎,0,0) (𝐻𝑎 ) is simply spanned by

⎛0⎞ ⎛0⎞
⎜1⎟ and ⎜0⎟ .
⎜ ⎟ ⎜ ⎟
⎝0⎠ ⎝1⎠
Appendix A. Solutions to exercises 367

Solution (Exercise 2.12) (a) The inverse map to 𝐹 is the projection 𝜋 onto the
first factor, for we obviously have 𝜋◦𝐹 = Id𝑋 and 𝐹 ◦𝜋 = IdΓ(𝑓 ) . Let 𝑋 ⊂ ℝ𝑁 ,
𝑌 ⊂ ℝ𝑀 , and let 𝑓 be smooth. Then for any point 𝑥 ∈ 𝑋, there is an open
subset 𝑈 ⊂ ℝ𝑁 and a smooth map 𝑓̃ ∶ 𝑈 → ℝ𝑀 with 𝑓̃𝑋∩𝑈 = 𝑓𝑋∩𝑈 . Then
(𝑓̃ × Id) ∶ 𝑈 → ℝ𝑁 × ℝ𝑀 is a smooth extension of 𝐹𝑋∩𝑈 . Hence 𝐹 is smooth.
The inverse map 𝜋 is a smooth map, since it extends to the smooth projection
on all of ℝ𝑁 × ℝ𝑀 . Hence 𝐹 is a diffeomorphism when 𝑓 is smooth. Hence
any local parametrization 𝜙 ∶ 𝑉 → 𝑋 can be extended to a local parametrization
𝐹 ◦𝜙 ∶ 𝑉 → Γ(𝑓 ). Thus the graph Γ(𝑓 ) is a manifold if 𝑋 is.

(b) For ∈ 𝑋, let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization around 𝑥 with 𝜙(0) = 𝑥

and open 𝑈 ⊆ ℝ𝑘 . Let 𝜓 ∶ 𝑊 → 𝑌 be a local parametrization around 𝑓 (𝑥)
with 𝜓(0) = 𝑓 (𝑥) and open 𝑊 ⊆ ℝ𝑙 . Then 𝜙 × 𝜓 ∶ 𝑈 × 𝑊 → 𝑋 × 𝑌 is a local
parametrization around (𝑥, 𝑓 (𝑥)) with 𝑈 ×𝑊 ⊂ ℝ𝑘+𝑙 open. Then we can construct
a commutative diagram

𝑋O
𝐹 /𝑋×𝑌
O
𝜙 𝜙×𝜓

𝑈 /𝑈 ×𝑊
𝐺=𝜓 −1 ◦𝐹 ◦𝜙

where 𝐺 is the map defined by 𝑣 → (𝑣, 𝜓 −1 (𝑓 (𝜙(𝑣))). Thus 𝐺 is the map Id𝑉 ×
(𝜓 −1 ◦𝑓 𝜙). Hence 𝑑𝐺0 ∶ ℝ𝑘 → ℝ𝑘 × ℝ𝑙 is the linear map

𝑑𝐺0 = Idℝ𝑘 × (𝑑𝜓𝑓−1(𝑥) ◦𝑑𝑓𝑥 ◦𝑑𝜙0 ).

Thus in the commutative diagram below, 𝑑𝐹𝑥 has to be defined as Id𝑇𝑥 (𝑋) × 𝑑𝑓𝑥 :

𝑑𝐹𝑥 =Id𝑇𝑥 (𝑋) ×𝑑𝑓𝑥

𝑇𝑥 (𝑋) / 𝑇 (𝑋) × 𝑇 (𝑌 )
O 𝑥
O 𝑦
𝑑𝜙0 𝑑𝜙0 ×𝑑𝜓0

ℝ𝑘 / ℝ𝑘 × ℝ𝑙 .
𝑑𝐺0 =Idℝ𝑘 ×(𝑑𝜓𝑓−1(𝑥) ◦𝑑𝑓𝑥 ◦𝑑𝜙0 )

Hence 𝑑𝐹𝑥 (𝑣) = (𝑣, 𝑑𝑓𝑥 (𝑣)).

(c) For ∈ 𝑋, let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization around 𝑥 with 𝜙(0) = 𝑥 and

open 𝑈 ⊆ ℝ𝑘 . Since 𝐹 is a diffeomorphism, 𝜓 = 𝐹 ◦𝜙 ∶ 𝑈 → Γ(𝑓 ) is then a local
parametrization around (𝑥, 𝑓 (𝑥)) with 𝜓(0) = 𝐹 (𝜙(0)) = (𝑥, 𝑓 (𝑥)). The tangent
space of Γ(𝑓 ) at (𝑥, 𝑓 (𝑥)) is by definition the image of 𝑑𝜓0 ∶ ℝ𝑘 → ℝ𝑁+𝑀 . By
the chain rule we have
𝑑𝜙0 𝑑𝐹𝑥
𝑑𝜓0 = 𝑑𝐹𝑥 ◦𝑑𝜙0 ∶ ℝ𝑘 ←←←←←←←→
← ℝ𝑁 ←←←←←←←→
← ℝ𝑁+𝑀 .

Hence by our definition of tangent spaces:

𝑇(𝑥,𝑓 (𝑥)) (Γ(𝑓 )) = 𝑑𝜓0 (ℝ𝑘 ) = 𝑑𝐹𝑥 (𝑑𝜙0 (ℝ𝑘 )) = 𝑑𝐹𝑥 (𝑇𝑥 (𝑋)).
368 A.2. Smooth manifolds
Finally, by the previous point, we know 𝑑𝐹𝑥 = Id𝑇𝑥 (𝑋) × 𝑑𝑓𝑥 and get

𝑇(𝑥,𝑓 (𝑥)) (Γ(𝑓 )) = (Id𝑇𝑥 (𝑋) × 𝑑𝑓𝑥 )(𝑇𝑥 (𝑋)) = Γ(𝑑𝑓𝑥 ) ⊂ 𝑇𝑥 (𝑋) × 𝑇𝑓 (𝑥) (𝑌 )

which is the graph of 𝑑𝑓𝑥 in 𝑇𝑥 (𝑋) × 𝑇𝑓 (𝑥) (𝑌 ).

Solution (Exercise 2.13) (a) Given a smooth map 𝑐 ∶ 𝐼 → ℝ𝑘 with 𝑐 = (𝑐1 , … , 𝑐𝑘 )

and 𝑐𝑖 ∶ 𝐼 → ℝ all smooth. The derivative of 𝑐 at 𝑡0 ∈ 𝐼 is a linear map
𝑑𝑐𝑡0 ∶ 𝑇𝑡0 𝐼 = ℝ → ℝ𝑘 = 𝑇𝑥0 (ℝ𝑘 ):

𝑑𝑐𝑡0 (𝑣) = (𝑐1′ (𝑡0 ), … , 𝑐𝑘′ (𝑡0 )) ⋅ 𝑣.

Since 𝑣 ∈ ℝ is just a real number, we get

𝑑𝑐𝑡0 (1) = (𝑐1′ (𝑡0 ), … , 𝑐𝑘′ (𝑡0 )) ∈ ℝ𝑘 .

(b) First, assume 𝑋 = ℝ𝑘 and let 𝑤 = (𝑤1 , … , 𝑤𝑘 ) be a vector in 𝑇𝑥 𝑋 = ℝ𝑘 . Then

define the curve 𝑐𝑤 ∶ ℝ → ℝ𝑘 by 𝑡 → 𝑡 ⋅ 𝑤. The derivative of 𝑐𝑤 at any 𝑡0 is

𝑑(𝑐𝑤 )𝑡0 ∶ ℝ → ℝ𝑘 , 𝑡 → (𝑤1 , … , 𝑤𝑘 ) ⋅ 𝑡.

Thus we have 𝑑(𝑐𝑤 )𝑡0 (1) = 𝑤.

Now let 𝑋 be an arbitrary 𝑘-dimensional smooth manifold, 𝑥 ∈ 𝑋, and let 𝑣 be
a vector in 𝑇𝑥 (𝑋). Let 𝜙 ∶ 𝑈 → 𝑋 be a local parametrization around 𝑥 with
𝜙(0) = 𝑥. By definition, 𝑇𝑥 (𝑋) = 𝑑𝜙0 (ℝ𝑘 ) and there is a unique vector 𝑤 ∈ ℝ𝑘
with 𝑑𝜙0 (𝑤) = 𝑣. Since any open ball around the origin in ℝ𝑘 is diffeomorphic to
ℝ𝑘 , we can assume 𝑈 = ℝ𝑘 and have 𝑤 ∈ 𝑈 . Let 𝑐𝑤 ∶ ℝ → ℝ𝑘 be the linear curve
in ℝ𝑘 defined in the previous point. Then we define 𝑐 ∶ ℝ → 𝑋 by 𝑐 = 𝜙◦𝑐𝑤 , i.e.,
𝑐(𝑡) = 𝜙(𝑡 ⋅ 𝑤). The derivative of 𝑐 at 𝑡0 = 0 ∈ ℝ is

𝑑(𝑐)𝑡0 = 𝑑𝜙0 ◦𝑑(𝑐𝑤 )𝑡0 .

Thus

𝑑(𝑐)𝑡0 (1) = 𝑑𝜙0 (𝑑(𝑐𝑤 )𝑡0 (1)) = 𝑑𝜙0 (𝑤) = 𝑣.

Appendix A. Solutions to exercises 369
A.3 The Inverse Function Theorem, immersions and embeddings

A.3.1 Diffeomorphisms, immersions and embeddings

Solution (Exercise 3.1) The derivative of 𝑓 at 𝑥 ∈ ℝ𝑛 is just given by the linear map
𝑑𝑓𝑥 ∶ ℝ𝑛 → ℝ𝑛 , 𝑣 → 𝐴 ⋅ 𝑣. Hence 𝑑𝑓𝑥 is an isomorphism if and only if 𝐴 is invertible.
So 𝑓 is a local diffeomorphism if and only if 𝐴 is invertible. Now 𝑓 is a diffeomorphism
if and only if the map 𝑥 → 𝑓 (𝑥) − 𝑏 = 𝐴𝑥 is a diffeomorphism which is the case if and
only if 𝐴 is invertible. For, if this is the case then 𝑦 → 𝐴−1 𝑦 + 𝐴−1 𝑏 is the inverse of 𝑓 .
Hence 𝑓 is a diffeomorphism if and only if 𝐴 is invertible.

Solution (Exercise 3.2) Since 𝑓 is bijective there is an inverse map 𝑓 −1 ∶ 𝑌 → 𝑋.

Since 𝑓 is a local diffeomorphism we can find, for every 𝑥 ∈ 𝑋, an open subset 𝑈 ⊂
such that 𝑓|𝑈 ∶ 𝑈 → 𝑓 (𝑈 ) is a diffeomorphism and 𝑓 (𝑈 ) is open in 𝑌 . Hence there
is a smooth inverse (𝑓|𝑈 )−1 of 𝑓|𝑈 . Since the inverse of a map is unique, we must have
(𝑓|𝑈 )−1 = (𝑓 −1 )|𝑓 (𝑈 ) . This shows that 𝑓 −1 is smooth on the open subset 𝑓 (𝑈 ) ⊂ 𝑌 .
Since 𝑓 is bijective, the open subsets 𝑓 (𝑈 ), for all 𝑥 ∈ 𝑋, cover 𝑌 . Hence, at every
point 𝑦 ∈ 𝑌 , we can find an open subset on which the restriction of 𝑓 −1 is smooth. This
shows that 𝑓 −1 is a smooth map. Thus, 𝑓 is a diffeomorphism.

Solution (Exercise 3.3) By definition of embeddings, we need to show that 𝑓 is an

injective, proper immersion.
𝑡 −𝑡 𝑠 −𝑠 𝑡 −𝑡 𝑠 −𝑠
• 𝑓 is injective: If 𝑓 (𝑡) = 𝑓 (𝑠), then 𝑒 +𝑒
2
= 𝑒 +𝑒
2
and 𝑒 −𝑒
2
= 𝑒 −𝑒2
. Adding
these two equations, implies 𝑒 = 𝑒 . Since the exponential function is injective,
𝑡 𝑠

this shows 𝑡 = 𝑠.

• 𝑓 is proper: Let 𝐾 be a compact subset of ℝ2 . That means that 𝐾 is both closed

and bounded in ℝ2 . Since 𝑓 is continuous, 𝑓 −1 (𝐾) is closed in ℝ. Since both coor-
dinates of 𝑓 (𝑡) are unbounded when 𝑡 varies in all of ℝ, 𝑓 −1 (𝐾) must be bounded
as well. Thus 𝑓 −1 (𝐾) is both closed and bounded in ℝ and therefore compact.

• 𝑓 is an immersion: The derivative of 𝑓 at any 𝑡 ∈ ℝ is given in the standard basis

by the 2 × 1-matrix
( 𝑡 −𝑡 )
𝑒 −𝑒
𝑑𝑓𝑡 = 2 .
𝑒𝑡 +𝑒−𝑡
2

For each 𝑡, 𝑑𝑓𝑡 is a linear map ℝ → ℝ2 . Since Ker (𝑑𝑓𝑡 ) is a vector subspace of
ℝ, it is either {0} or ℝ itself. Since 𝑑𝑓𝑡 is not the zero matrix for any 𝑡, 𝑑𝑓𝑡 must
be injective for all 𝑡 ∈ ℝ.

Solution (Exercise 3.4) The map 𝑓 is not an embedding, since it is not injective. But
we can check it is an immersion by showing that the derivative is injective everywhere.
370 A.3. The Inverse Function Theorem, immersions and embeddings
The derivative of 𝑓 at (𝑠, 𝑡) is

⎛− sin 𝑠 cos 𝑡 −(2 + cos 𝑠) sin 𝑡⎞

𝑑𝑓(𝑠,𝑡) ∶ ℝ2 → ℝ3 , 𝑑𝑓(𝑠,𝑡) = ⎜ − sin 𝑠 sin 𝑡 (2 + cos 𝑠) cos 𝑡 ⎟ .
⎜ ⎟
⎝ cos 𝑠 0 ⎠

In order to show that 𝑑𝑓(𝑠,𝑡) is injective, we need to check that it has full or maximal
rank, i.e. rank 2. Hence we need to check that the two column vectors are always linearly
independent. To simplify notation, we set 𝑥 = sin 𝑠, 𝑦 = cos 𝑠, 𝑢 = sin 𝑡, and 𝑣 = cos 𝑡.
Now assume there are two real numbers 𝜆 and 𝜇 such that

𝜆(−𝑥𝑣) + 𝜇(−(2 + 𝑦)𝑢) = 0

𝜆(−𝑥𝑢) + 𝜇((2 + 𝑦)𝑢) = 0
𝜆𝑦 = 0.

We distinguish two cases: 𝑦 = cos 𝑠 = 0 and 𝑦 = cos 𝑠 ≠ 0. If 𝑦 ≠ 0, then we must have

𝜆 = 0 and

𝜇(2 + 𝑦)𝑢 = 0
𝜇(2 + 𝑦)𝑣 = 0.

Since |𝑦| ≤ 1, we know 2 + 𝑦 ≠ 0. Hence we can divide by 2 + 𝑦. Moreover, we know

that not both 𝑢 and 𝑣 can be 0 at the same time. This implies 𝜇 = 0.
If 𝑦 = 0, then 𝑥 = sin 𝑠 = ±1 and still 2 + 𝑦 ≠ 0. Hence we get the system

±𝜆𝑣 − 𝜇(2 + 𝑦)𝑢 = 0

±𝜆𝑢 + 𝜇(2 + 𝑦)𝑣 = 0.

But since 𝑢 and 𝑣 are never both 0, we know that the vectors (𝑣, 𝑢) and (𝑢, 𝑣) are linearly
independent. Hence we must have 𝜆 = 0 and 𝜇(2 + 𝑦) = 0. The latter implies 𝜇 = 0,
since 2 + 𝑦 ≠ 0.

Solution (Exercise 3.5) Let 𝑎 and 𝑏 be two relatively prime integers with 𝑎 ≠ 0.
Recall that, for 𝑡1 , 𝑡2 ∈ 𝑅, we have
𝑒2𝜋𝑖𝑡1 = 𝑒2𝜋𝑖𝑡2 ⇐⇒ 𝑡1 − 𝑡2 ∈ ℤ.
Since 𝑎 and 𝑏 are integers, 𝑡1 − 𝑡2 ∈ ℤ implies 𝑎𝑡1 − 𝑎𝑡2 ∈ ℤ and 𝑏𝑡1 − 𝑏𝑡2 ∈ ℤ. Hence
we have
𝑒2𝜋𝑖𝑡1 = 𝑒2𝜋𝑖𝑡2 ⇒ 𝑒2𝜋𝑖𝑎𝑡1 = 𝑒2𝜋𝑖𝑎𝑡2 and 𝑒2𝜋𝑖𝑏𝑡1 = 𝑒2𝜋𝑖𝑏𝑡2 .
This shows that we have a well-defined map
𝑔𝑎,𝑏 ∶ 𝕊1 → 𝕊1 × 𝕊1 , 𝑒2𝜋𝑖𝑡 → (𝑒2𝜋𝑖𝑎𝑡 , 𝑒2𝜋𝑖𝑏𝑡 ).
Hence, for 𝑓 ∶ ℝ → 𝕊1 , 𝑡 → 𝑒2𝜋𝑖𝑡 , we get a commutative diagram
𝛾𝑎,𝑏
ℝ / 𝕊1 × 𝕊1
;

𝑓 𝑔𝑎,𝑏

𝕊1 .
Appendix A. Solutions to exercises 371

Since 𝕊1 is compact and 𝑔𝑎,𝑏 a one-to-one immersion, 𝑔𝑎,𝑏 is an embedding.

Solution (Exercise 3.6) (a) The derivative of 𝑓 at 𝑡 is the map 𝑑𝑓𝑡 ∶ ℝ → ℝ2 given
by the Jacobian matrix

(2 cos(2𝑡) cos 𝑡 + sin(2𝑡)(− sin 𝑡), 2 cos(2𝑡) sin 𝑡 + sin(2𝑡) cos 𝑡)

= (2(cos2 𝑡 − sin2 𝑡) cos 𝑡 − 2 sin 𝑡 cos 𝑡 sin 𝑡, 2(cos2 𝑡 − sin2 𝑡) sin 𝑡 + 2 sin 𝑡 cos 𝑡 cos 𝑡)
= 2(cos3 𝑡 − 2 sin2 𝑡 cos 𝑡, 2 cos2 𝑡 sin 𝑡 − sin3 𝑡)
= 2(cos 𝑡(cos2 𝑡 − 2 sin2 𝑡), sin 𝑡(2 cos2 𝑡 − sin2 𝑡))
= 2(cos 𝑡(1 − 3 sin2 𝑡), sin 𝑡(3 cos2 𝑡 − 1))

where we have used several trigonometric identities. Since the derivative 𝑑𝑓𝑡 is
nontrivial for all 𝑡, it is always injective as a linear map ℝ → ℝ2 . Hence 𝑓 is an
immersion.

(b) But 𝑓 is not a homeomorphism onto Im (𝑓 ). See Figure A.1. For, consider the open
subset (𝜋∕4, 3𝜋∕4) in (0, 3𝜋∕4). If 𝑓 was a homeomorphism, then 𝑓 ((𝜋∕4, 3𝜋∕4))
had to be open in Im (𝑓 ) as well. That means that around any point, for exam-
ple the point 𝑓 (𝜋∕2) = (0, 0), there had to an open neighborhood contained in
𝑓 ((𝜋∕4, 3𝜋∕4)). By the definition of the open sets in Im (𝑓 ) as a subspace of ℝ2 ,
there had to be an open ball 𝐵𝜀 (0, 0) ∈ ℝ2 with

𝐵𝜀 (0, 0) ∩ Im (𝑓 ) ⊂ 𝑓 ((𝜋∕4, 3𝜋∕4)).

But for every 𝜖 > 0, we have

𝐵𝜀 (0, 0) ∩ 𝑓 ((0, 𝜋∕4)) ≠ ∅,

since | sin(2𝑡)(cos 𝑡, sin 𝑡)| < 𝜀 for all 𝑡 < 𝜖∕2 (where we use sin 𝑥 ≤ 𝑥 and
|(cos 𝑡, sin 𝑡)| = 1). Hence 𝑓 cannot be an open map and therefore not a home-
omorphism.

(c) • What is the difference between Im (𝑓 ) and the graph Γ(𝑓 )?

Answer: The graph of a map 𝑋 → 𝑌 is a subspace of 𝑋 × 𝑌 . In this case,
Γ(𝑓 ) is a subspace of (0, 3𝜋∕4) × ℝ2 , whereas Im (𝑓 ) is a subspace of ℝ2 .
• Is the map 𝐹 ∶ (0, 3𝜋∕4) → (0, 3𝜋∕4) × ℝ2 an embedding?
Answer: Yes, because 𝐹 is a diffeomorphism (0, 3𝜋∕4) → Γ(𝑓 ) (since 𝑓
is smooth, see previous exercise set). Hence it is in particular, a one-to-one
immersion and proper.
• Would 𝑓 be an embedding if it was defined on the closed interval [0, 3𝜋∕4]?
Answer: No, because 𝑓 would not be injective anymore: 𝑓 (0) = (0, 0) =
𝑓 (𝜋∕2).
• Is the map 𝑔 ∶ (0, 3𝜋∕4) → ℝ3 , 𝑡 → sin(2𝑡)(cos 𝑡, sin 𝑡, 𝑡) an embedding?
Answer: No, this map is still just an immersion and it is even one-to-one,
but it is not a homeomorphism onto its image in ℝ3 . The same argument as
in the previous point.
372 A.3. The Inverse Function Theorem, immersions and embeddings

• Is the map ℎ ∶ [0, 3𝜋∕4] → ℝ3 , 𝑡 → (sin(2𝑡) cos 𝑡, sin(2𝑡) sin 𝑡, 2𝑡) an embed-
ding?
Answer: Yes, this time we have a map which is an immersion, it is one-to-
one this time 𝑓 (0) = (0, 0, 0) ≠ (0, 0, 𝜋) = 𝑓 (𝜋∕2), and it is defined on a
compact space and is therefore a proper map.

Figure A.1: The origin is the critical point. We cannot separate the two branches of the graph
with open subsets.

Solution (Exercise 3.7) By the Local Immersion Theorem, we can choose lo-
cal parametrizations 𝜙 ∶ 𝑉 → 𝑍 and 𝜓 ∶ 𝑊 → 𝑋 around 𝑧 with 𝑉 ⊂ ℝ𝑘 and
𝑊 = 𝑉 ⊕ 𝑉 ′ ⊂ ℝ𝑛 such that
inclusion /𝑋
𝑍O O
𝜙 𝜓

canonical / 𝑊 ⊆ ℝ𝑛
ℝ𝑘 ⊇ 𝑉
immersion

commutes. The map 𝜓 is a diffeomorphism onto its image 𝜓(𝑊 ) ⊂ 𝑋. The inverse
map 𝜓 −1 ∶ 𝜓(𝑊 ) → 𝑊 is a local coordinate system on the open neighborhood 𝜓(𝑊 )
around 𝑧 ∈ 𝑋. We write 𝑥𝑖 ∶ 𝜓(𝑊 ) → ℝ for the 𝑖th component of 𝜓 −1 , i.e. a point
𝑝 ∈ 𝜓(𝑊 ) has the local coordinates (𝑥1 (𝑝), … , 𝑥𝑛 (𝑝)) = (𝜓1−1 (𝑝), … , 𝜓𝑛−1 (𝑝)). Since
the above diagram commutes and 𝜙 is a diffeomorphism onto its image, we have
𝜙(𝑉 ) = 𝑍 ∩ 𝜓(𝑊 ).
Hence, since the lower horizontal map is the canonical immersion, the points in 𝑍∩𝜓(𝑊 )
are exactly those on which the coordinate functions 𝑥𝑘+1 , … , 𝑥𝑛 vanish. Relabelling the
open subset 𝜓(𝑊 ) as 𝑈 we have
𝑍 ∩ 𝑈 = {𝑝 ∈ 𝑈 such that 𝑥𝑘+1 (𝑝) = ⋯ = 𝑥𝑛 (𝑝) = 0}.
Appendix A. Solutions to exercises 373
A.4 Submersions and regular values

A.4.1 Submersions and regular values

Solution (Exercise 4.1) We can assume that 𝑈 = 𝑋, since we just need to consider
𝑓
the map 𝑈 ⊆ 𝑋 ←←←→← 𝑌 . Let 𝑦 be any point in 𝑓 (𝑋). We need to show that there is an
open neighborhood 𝑊 around 𝑦 which is contained in 𝑓 (𝑋). Let 𝑥 be a point in 𝑋 with
𝑓 (𝑥) = 𝑦 which exists since 𝑦 ∈ 𝑓 (𝑋). By the Local Submersion Theorem, we can
choose local parametrizations 𝜙 ∶ 𝑉 → 𝑋 around 𝑥 with 𝑉 ⊂ ℝ𝑛 open and 𝜓 ∶ 𝑉 ′ → 𝑌
around 𝑦 with 𝑉 ′ ⊂ ℝ𝑚 open such that the induced map 𝑉 → 𝑉 ′ is the canonical
submersion:
𝑓
𝑋O /𝑌
O
𝜙 𝜓

canonical /𝑉′
𝑉
submersion

By possibly shrinking 𝑉 and 𝑉 ′ , we can assume that 𝑉 = 𝐵𝜖 (0) ⊂ ℝ𝑛 and 𝑉 ′ = 𝐵𝜖 (0) ⊂

ℝ𝑚 . Then the canonical submersion maps 𝑉 onto 𝑉 ′ . Since the above diagram com-
mutes, we see that 𝑊 ∶= 𝜓(𝑉 ′ ) is contained in 𝑓 (𝑋). Since 𝜓 is a diffeomorphism onto
its image in 𝑌 , 𝑊 is open in 𝑌 .

Solution (Exercise 4.2) The derivative of 𝑔 at a point (𝑥, 𝑦) is given by the 1×2-matrix
( )
𝑑𝑔(𝑥,𝑦) = 2𝑥 −2𝑦 .

As a linear map from ℝ2 to ℝ, 𝑑𝑔(𝑥,𝑦) is surjective whenever it is not the zero map. Hence
𝑑𝑔(𝑥,𝑦) is surjective for all (𝑥, 𝑦) ≠ (0, 0).
Thus the set of regular values of 𝑔 is the subset ℝ ⧵ {0}. Since 𝑔(0, 0) = 0, the only
critical value is 0.
Since the derivative of 𝑔 is not surjective at all points, 𝑔 is not a submersion.

Solution (Exercise 4.3) (a) Let 𝑓 ∶ 𝑋 → 𝑌 be a submersion. We have 𝑌 =

𝑓 (𝑋) ∪ (𝑌 ⧵ 𝑓 (𝑋)). Since 𝑋 is compact and 𝑓 is continuous, we know 𝑓 (𝑋) is
compact and therefore closed in 𝑌 . Since 𝑋 is open in 𝑋 and 𝑓 is a submersion,
the previous exercise shows that 𝑓 (𝑋) is open in 𝑌 . Hence 𝑓 (𝑋) is both open and
closed in 𝑌 . Hence 𝑓 (𝑋) must be either 𝑌 or ∅. Assuming 𝑓 is nontrivial, 𝑓 (𝑋)
must be all of 𝑌 .

(b) Given a compact smooth manifold 𝑋. Assume we had a submersion 𝑓 ∶ 𝑋 → ℝ𝑛

with for some 𝑛. By the previous point, we would have 𝑓 (𝑋) = ℝ𝑛 . But since 𝑋 is
compact, 𝑓 (𝑋) is compact too. But ℝ𝑛 is not compact. Hence such a submersion
cannot exist.
374 A.4. Submersions and regular values

Solution (Exercise 4.4) Given 𝐴 = (𝑎𝑖𝑗 ) ∈ 𝑂(𝑛). Unfolding the matrix-multiplication,

∑ 2 ∑ 2
we see that 𝑗 𝑎𝑖𝑗 is the 𝑖th diagonal entry in 𝐴𝐴 . Moreover, 𝑗 𝑎𝑖𝑗 is also the square
𝑇

of the norm of the 𝑖th row vector of 𝐴. Since 𝐴 ∈ 𝑂(𝑛), we have 𝐴𝐴𝑇 = 𝐼 and the 𝑖th
diagonal entry in 𝐴𝐴𝑇 is equal 1. This shows that 𝑂(𝑛) is contained in the product of
∏ ∏ 2
𝑛 spheres 𝕊𝑛−1 in ℝ𝑛 = ℝ𝑛 = 𝑀(𝑛). Hence 𝑂(𝑛) is bounded. But 𝑂(𝑛) is also
2
closed in ℝ𝑛 , since we can define it as the inverse image of the closed point 𝐼 ∈ 𝑆(𝑛)
under the map 𝑀(𝑛) → 𝑆(𝑛) sending 𝐴 to 𝐴𝐴𝑇 . Thus 𝑂(𝑛) is closed and bounded in
2
ℝ𝑛 and therefore compact.

2
Solution (Exercise 4.5) We consider 𝑂(𝑛) as a subspace in 𝑀(𝑛) = ℝ𝑛 . We
defined 𝑂(𝑛) as 𝑓 −1 (𝐼) under the map 𝑓 ∶ 𝑀(𝑛) → 𝑆(𝑛), 𝑓 (𝐴) = 𝐴𝐴𝑇 . We have
checked in the proof of Theorem 4.12 that 𝐼 is a regular value for 𝑓 . As a consequence
of the Preimage Theorem we saw that this implies that 𝑇𝐼 (𝑂(𝑛)) equals the kernel of
𝑑𝑓𝐼 ∶ 𝑀(𝑛) = 𝑇𝐼 (𝑀(𝑛)) → 𝑇𝐼 (𝑆(𝑛)) = 𝑆(𝑛). We calculated the derivative 𝑑𝑓𝐴 for any
𝐴 ∈ 𝑂(𝑛) in the proof of Theorem 4.12: it is given by 𝑑𝑓𝐴 (𝐵) = 𝐵𝐴𝑇 + 𝐴𝐵 𝑇 . For
𝐴 = 𝐼, this gives 𝑑𝑓𝐼 (𝐵) = 𝐵 + 𝐵 𝑇 . Hence the kernel of 𝑑𝑓𝐼 is the space of matrices
satisfying 𝐵 + 𝐵 𝑇 = 0, i.e., 𝐵 𝑇 = −𝐵.

Solution (Exercise 4.6) (a) Since 𝑑(det)𝐴 is a linear map ℝ4 → ℝ, it suffices to

show 𝑑(det)𝐴 is nonzero. Therefore, it suffices to show that 𝑑(det)𝐴 (𝐵) ≠ 0 for
some matrix 𝐵.
Since 𝐴 ≠ 0, there is at least one entry in 𝐴 which is nonzero. Assume that
( and for
𝑎11 ≠ 0, ) the other cases the argument is similar. Then we take the matrix
0 0
𝐸22 = :
0 1

det(𝐴 + 𝑠𝐸22 ) − det 𝐴

𝑑(det)𝐴 (𝐸22 ) = lim
𝑠→0 𝑠
det(𝐴 + 𝑠𝐸22 ) − 0
= lim
𝑠→0 𝑠
(𝑎11 )(𝑎22 + 𝑠) − 𝑎12 𝑎21
= lim
𝑠→0 𝑠
𝑎11 𝑎22 + 𝑎11 𝑠 − 𝑎12 𝑎21
= lim
𝑠→0 𝑠
𝑎11 𝑠 + det 𝐴
= lim
𝑠→0 𝑠
𝑎11 𝑠
= lim = 𝑎11 ≠ 0.
𝑠→0 𝑠

(b) A 2×2-matrix 𝐴 has rank 0 if and only if it is the zero matrix. Thus 𝐴 ∈ 𝑀(2)⧵{0}
has rank 1 if and only if it does not have rank 2, i.e., if and only if it is not invertible.
Hence 𝐴 ∈ 𝑀(2) ⧵ {0} has rank 1 if and only if det 𝐴 = 0. By the previous point,
the determinant function is a submersion 𝑀(2) ⧵ {0} → ℝ. Hence 𝑅1 = det −1 (0)
is a submanifold of dimension 4 − 1 = 3 by the Preimage Theorem.
Appendix A. Solutions to exercises 375

Solution (Exercise 4.7) We consider the function 𝑄 defined in the hint. Since 𝑃 is
homogeneous, we know 𝑄 is always 0. Hence its derivative with respect to 𝑡 is zero as
well. Hence we get
∑
0 = 𝜕𝑄∕𝜕𝑡 = 𝑥𝑖 𝜕𝑃 ∕𝜕𝑥𝑖 (𝑡𝑥1 , … , 𝑡𝑥𝑘 ) − 𝑚𝑡𝑚−1 𝑃 (𝑡𝑥1 , … , 𝑡𝑥𝑘 ) (A.1)
𝑖

where we apply the chain rule to the first summand of 𝑄 which is the composite 𝑡 →
𝑡𝑥 → 𝑃 (𝑡𝑥). Setting 𝑡 = 1 in (A.1) yields Euler’s identity (4.2).

Solution (Exercise 4.8) (a) The derivative of 𝑃 at a point (𝑥1 , … , 𝑥𝑘 ) is

⎛ 𝑧1 ⎞
𝑑𝑃𝑥 ∶ ℝ𝑘 → ℝ, (𝑧1 , … , 𝑧𝑘 ) →(𝜕𝑃 ∕𝜕𝑥1 (𝑥) … 𝜕𝑃 ∕𝜕𝑥𝑘 (𝑥)) ⋅ ⎜ ⋮ ⎟
⎜ ⎟
⎝𝑧𝑘 ⎠
∑
= 𝑧𝑖 𝜕𝑃 ∕𝜕𝑥𝑖 (𝑥).
𝑖

To show that 𝑑𝑃𝑥 is nonsingular, i.e. surjective, it suffices to show that 𝑑𝑃𝑥 is
nontrivial. But applying 𝑑𝑃𝑥 to 𝑥 and using Euler’s identity yields
∑
𝑑𝑃𝑥 (𝑥) = 𝑥𝑖 𝜕𝑃 ∕𝜕𝑥𝑖 (𝑥1 , … , 𝑥𝑘 ) = 𝑚𝑃 (𝑥1 , … , 𝑥𝑘 ).
𝑖

Hence if 𝑥 = (𝑥1 , … , 𝑥𝑘 ) is not a zero of 𝑃 , then 𝑑𝑃𝑥 (𝑥) is nonzero. Hence all
nonzero real numbers are regular values of 𝑃 . The Preimage Theorem now implies
that 𝑃 −1 (𝑎) is a 𝑘 − 1-dimensional submanifold of ℝ𝑘 for all 𝑎 ≠ 0.

(b) Given two real numbers 𝑎, 𝑏 > 0, then (𝑏∕𝑎)1∕𝑚 exists and we if 𝑃 (𝑥) = 𝑎, we have

𝑃 ((𝑏∕𝑎)1∕𝑚 𝑥1 , … , (𝑏∕𝑎)1∕𝑚 𝑥𝑘 ) = 𝑏∕𝑎𝑃 (𝑥1 , … , 𝑥𝑘 ) = 𝑏.

Multiplying each coordinate with (𝑏∕𝑎)1∕𝑚 corresponds to multiplication with the

diagonal matrix with (𝑏∕𝑎)1∕𝑚 on the diagonal. This map is a linear isomorphism
of ℝ𝑘 to itself. Hence we have the diffeomorphism

𝑃 −1 (𝑎) → 𝑃 −1 (𝑏), (𝑥1 , … , 𝑥𝑘 ) → ((𝑏∕𝑎)1∕𝑚 𝑥1 , … , (𝑏∕𝑎)1∕𝑚 𝑥𝑘 ).

Similarly, if both 𝑎, 𝑏 < 0 are negative, then (𝑏∕𝑎)1∕𝑚 exists and the same argument
shows that 𝑃 −1 (𝑎) and 𝑃 −1 (𝑏) are diffeomorphic.

Solution (Exercise 4.9) (a) If we think of the entries in an 𝑛 × 𝑛-matrix 𝐴 as

variables, then det 𝐴 is a homogeneous polynomial of degree 𝑛 given by Leibniz’
formula. Hence we can apply a previous exercise to
2
𝑃 = det ∶ 𝑀(𝑛) = ℝ𝑛 → ℝ

and conclude that 0 is the only critical value of det.

376 A.4. Submersions and regular values
∑𝑛
Alternatively: Using the formula det 𝐴 = 𝑖=1 (−1)𝑖+𝑗 𝑎𝑖𝑗 det 𝐴𝑖𝑗 we can compute
the partial derivatives of det with respect to each variable 𝑥𝑖𝑗 . Since we remove the
𝑗th column from 𝐴 for each of the 𝐴𝑖𝑗 , the entry 𝑎𝑖𝑗 occurs in this formula exactly
once. Hence the partial derivative with respect to 𝑥𝑖𝑗 is just the factor of 𝑎𝑖𝑗 :
𝜕 det
(𝐴) = (−1)𝑖+𝑗 det 𝐴𝑖𝑗 .
𝜕𝑥𝑖𝑗
The total derivative of det as a function of its entries can be represented by the
𝑛 × 1-matrix with these partial derivatives as entries. This linear map is then not
surjective if and only if all entries are zero, i.e., if and only if det 𝐴𝑖𝑗 = 0 for all 𝑖, 𝑗.
The latter happens if and only if the rank of 𝐴 is < 𝑛, i.e., if and only if det 𝐴 = 0.
Hence 0 the only critical value for det.
(b) By the previous point, 1 is a regular value for det. Hence, by the Preimage Theo-
rem, 𝑆𝐿(𝑛) = det −1 (1) is a smooth manifold of dimension dim 𝑀(𝑛) − dim ℝ =
𝑛2 − 1.
(c) According to the Preimage Theorem and the previous point, we can determine the
tangent space as the kernel of the derivative of det at the identity matrix. Hence
we need to calculate the derivative of det at the identity. To do this we are going
to use Leibniz’ formula (4.3).
Given a matrix 𝐴, in Leibniz’ formula for the determinant of 𝐵 ∶= 𝐼 + 𝑠𝐴, every
summand contains at least a factor 𝑠2 unless it is the product of at least 𝑛 − 1
diagonal entries 𝑏𝑖𝑖 = 1 + 𝑠𝑎𝑖𝑖 . For we need 𝑛 − 1 factors not containing 𝑠 which
is only possible when we multiply 𝑛 − 1 times 1. But if a permutation {1, … , 𝑛}
leaves 𝑛 − 1 numbers fixed, it also has to leave the remaining one fixed. Hence the
only summand in (4.3) which does not contain a factor 𝑠2 is the summand
∏
𝑛
(1 + 𝑠𝑎𝑖𝑖 ) = (1 + 𝑠𝑎11 ) ⋯ (1 + 𝑠𝑎𝑛𝑛 ) = 1 + 𝑠 ⋅ tr (𝐴) + 𝑂(𝑠2 ).
𝑖=1

The derivative of the determinant at the identity

𝑑(det)𝐼 ∶ 𝑇𝐼 (𝑀(𝑛)) = 𝑀(𝑛) → 𝑇1 (ℝ) = ℝ
is then given by
det(𝐼 + 𝑠𝐴) − det 𝐼
𝑑(det)𝐼 (𝐴) = lim
𝑠→0 𝑠
1 + 𝑠 ⋅ tr (𝐴) + 𝑂(𝑠2 ) − 1
= lim
𝑠→0 𝑠
𝑠 ⋅ tr (𝐴) + 𝑂(𝑠2 )
= lim
𝑠→0 𝑠
= lim tr (𝐴) + 𝑂(𝑠)
𝑠→0
= tr (𝐴).

Hence we get
𝑇𝐼 (𝑆𝐿(𝑛)) = Ker (𝑑(det)𝐼 ) = {𝐴 ∈ 𝑀(𝑛) ∶ tr (𝐴) = 0}.

In other words, the tangent space to 𝑆𝐿(𝑛) at the identity is the space of matrices
whose trace vanishes.
Appendix A. Solutions to exercises 377

Solution (Exercise 4.10) (a) We write (𝑧0 , 𝑧1 ) = (𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) for real coordinates
𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 . First we get
( )
̃ 0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2
𝜋(𝑧
( )
= 2(𝑥0 𝑥1 + 𝑦0 𝑦1 ) + 𝑖2(−𝑥0 𝑦1 + 𝑦0 𝑥1 ), (𝑥20 + 𝑦20 ) − (𝑥21 + 𝑦21 ) .

Then we can compute

⎛ 𝑥1 𝑦1 𝑥0 𝑦0 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ −𝑦1 𝑥1 𝑦0 −𝑥0 ⎟ .
⎜
⎜ ⎟
⎝ 𝑥0 𝑦0 −𝑥1 −𝑦1 ⎠

(b) Let 𝑔4 ∶ ℝ4 → ℝ and 𝑔3 ∶ ℝ3 → ℝ be the usual smooth maps such that

𝕊3 = 𝑔4−1 (1) and 𝕊2 = 𝑔3−1 (1) respectively. Then we have 𝑇𝑞 𝕊3 = Ker (𝑑(𝑔4 )𝑞 ) ⊂
𝑇𝑞 ℝ4 = ℝ4 and 𝑇𝑝 𝕊2 = Ker (𝑑(𝑔3 )𝑝 ) ⊂ 𝑇𝑝 ℝ3 = ℝ3 . We know that 𝜋(𝑞) ̃ ∈ 𝕊2 if
𝑞 ∈ 𝕊 . Actually, our calculation above shows that 𝜋(𝑞)
3 ̃ ∈ 𝕊 if and only if 𝑞 ∈ 𝕊3 .
2

This implies 𝕊 = 𝜋̃ (𝑔3 (1)) = (𝑔3 ◦𝜋)

3 −1 −1 ̃ (1). In particular, 𝑔3 (𝜋(𝑞))
−1 ̃ = 1 is con-
stant on 𝕊 . Hence, for every 𝑞 ∈ 𝕊 , the image of the restriction (𝑑 𝜋̃𝑞 )|𝑇𝑞 𝕊3 is
3 3

contained in the kernel of 𝑑(𝑔3 )𝜋(𝑞)

̃ which is 𝑇𝜋(𝑞)
̃ 𝕊 .
2

(c) ∙ The fiber over 𝑎 is

𝜋 −1 (𝑎) = {(𝑧0 , 0) ∈ 𝕊3 ⊂ ℂ2 ∶ |𝑧0 |2 = 1}.
Let 𝑞 = (𝑥0 , 𝑦0 , 0, 0) ∈ 𝜋 −1 (𝑎) be a point in the fiber over 𝑎. The tangent
space 𝑇𝑞 𝕊3 is the vector space
𝑇𝑞 𝕊3 = {𝐮 ∈ ℝ4 ∶ 𝐮 ⟂ 𝑞}
⎧ ⎛−𝑦0 ⎞ ⎛0⎞ ⎛0⎞⎫
⎪ ⟂ ⎜ 𝑥0 ⎟ 4 ⎜0⎟ 4 ⎜0⎟⎪
= span ⎨𝑞 = ⎜ ⎟ , 𝐞3 = ⎜1⎟ , 𝐞4 = ⎜0⎟⎬ .
⎪ ⎜ 0 ⎟ ⎜ ⎟ ⎜ ⎟⎪
⎩ ⎝ 0 ⎠ ⎝0⎠ ⎝1⎠⎭
The vectors 𝑞 ⟂ , 𝐞43 and 𝐞44 are linearly independent and hence form a basis of
𝑇𝑞 𝕊3 .
Now we consider the map 𝑑 𝜋̃𝑞 . We computed this map as represented by the
matrix
⎛ 0 0 𝑥0 𝑦0 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ ⎜ 0 0 𝑦0 −𝑥0 ⎟ .
⎜ ⎟
⎝𝑥0 𝑦0 0 0 ⎠
The kernel of this map is the span of the vector 𝑞 ⟂ that we have just seen.
This implies that 𝑑𝜋𝑞 is surjective onto 𝑇𝑎 𝕊2 .
More concretely, the tangent space 𝑇𝑎 𝕊2 consists of the vectors which are
orthogonal to 𝑎 in ℝ3 . Hence it has a basis (𝐞31 , 𝐞32 ) as a subspace in ℝ3 . The
map 𝑑 𝜋̃𝑞 ∶ 𝑇𝑞 ℝ4 → 𝑇𝑎 ℝ3 sends
⎛2𝑥0 ⎞ ⎛ 2𝑦0 ⎞
⟂
𝑞 → 𝟎, 𝐞43 → 2𝑦0 , 𝐞4 → ⎜−2𝑥0 ⎟ .
⎜ ⎟ 4
⎜ ⎟ ⎜ ⎟
⎝ 0 ⎠ ⎝ 0 ⎠
378 A.4. Submersions and regular values

Hence the map 𝑑𝜋𝑞 ∶ 𝑇𝑞 𝕊3 → 𝑇𝑎 𝕊2 can be represented in the chosen bases

by the matrix ( )
0 𝑥0 𝑦0
𝑑𝜋𝑞 = 2 ⋅ .
0 𝑦0 −𝑥0
This map is surjective, since −𝑥20 − 𝑦20 = −1 ≠ 0. Hence 𝑞 is a regular point
for 𝜋. Since 𝑞 was any point in the fiber over 𝑎, we have shown that 𝑎 is a
regular value for 𝜋.
∙ To determine the fiber over 𝑏, we write 𝑧0 = 𝑥0 + 𝑖𝑦0 and 𝑧1 = 𝑥1 + 𝑖𝑦1 .
Then we get
1
𝜋(𝑧0 , 𝑧1 ) = (0, 1, 0) ⇒ 2𝑧0 𝑧̄ 1 = 𝑖 and |𝑧0 |2 = |𝑧1 |2 =
2
1
⇒ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 and 𝑥20 + 𝑥21 = .
2
Thus the fiber over 𝑏 has the form
𝑖
𝜋 −1 (𝑏) = {(𝑧0 , 𝑧1 ) ∈ 𝕊3 ∶ 𝑧̄ 1 = }
2𝑧0
= {(𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) ∈ 𝕊3 ∶ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 }.

Let 𝑞 = (𝑥0 , 𝑥1 , 𝑥1 , −𝑥0 ) ∈ 𝜋 −1 (𝑏) be a point in the fiber over 𝑏. Since not
both 𝑥0 and 𝑥1 can be zero, we assume that 𝑥0 ≠ 0. The tangent space 𝑇𝑞 𝕊3
is the vector space

𝑇𝑞 𝕊3 = {𝐮 ∈ ℝ4 ∶ 𝐮 ⟂ 𝑞}
⎧ ⎛−𝑥1 ⎞ ⎛−𝑥1 ⎞ ⎛𝑥0 ⎞⎫
⎪ ⟂ ⎜ 𝑥0 ⎟ ⟂ ⎜ 0 ⎟ ⟂ ⎜ 0 ⎟⎪
= span ⎨𝑞1 = ⎜ ⎟ , 𝑞2 = ⎜ 𝑥 ⎟ , 𝑞3 = ⎜ 0 ⎟⎬ .
⎪ ⎜ 0 ⎟ ⎜ 0⎟ ⎜ ⎟⎪
⎩ ⎝ 0 ⎠ ⎝ 0 ⎠ ⎝𝑥0 ⎠⎭

Now we consider the map 𝑑 𝜋̃𝑞 . We computed this map as represented by the
matrix
⎛𝑥1 −𝑥0 𝑥0 𝑥1 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ ⎜𝑥0 𝑥1 𝑥1 −𝑥0 ⎟ .
⎜ ⎟
⎝𝑥0 𝑥1 −𝑥1 𝑥0 ⎠

⎛−𝑥1 ⎞
⎜𝑥 ⎟
The kernel of this map is the span of the vector 𝑞0⟂ = ⎜ 0 ⎟. This implies
⎜ 𝑥0 ⎟
⎝ 𝑥1 ⎠
that 𝑑𝜋𝑞 is surjective onto 𝑇𝑎 𝕊2 .
More concretely, the tangent space 𝑇𝑏 𝕊2 consists of the vectors which are
orthogonal to 𝑏 in ℝ3 . Hence it has a basis (𝐞31 , 𝐞33 ) as a subspace in ℝ3 . The
map 𝑑 𝜋̃𝑞 ∶ 𝑇𝑞 ℝ4 → 𝑇𝑏 ℝ3 sends

⎛−2(𝑥20 + 𝑥21 )⎞ ⎛2(𝑥20 − 𝑥21 )⎞ ⎛2𝑥0 𝑥1 ⎞

⟂
𝑞1 → ⎜ 0 ⎟ ⟂
,𝑞 → ⎜ 0 ⎟ , 𝑞 → ⎜ 0 ⎟ .
⟂
⎜ ⎟ 2 ⎜ ⎟ 3 ⎜ 2 ⎟
⎝ 0 ⎠ ⎝ −2𝑥0 𝑥1 ⎠ ⎝ 2𝑥0 ⎠
Appendix A. Solutions to exercises 379

Hence the map 𝑑𝜋𝑞 ∶ 𝑇𝑞 𝕊3 → 𝑇𝑏 𝕊2 can be represented in the chosen bases

by the matrix
( 2 )
−𝑥0 − 𝑥21 𝑥20 − 𝑥21 𝑥0 𝑥1
𝑑𝜋𝑞 = 2 ⋅ .
0 −𝑥0 𝑥1 𝑥20

This map is surjective, as one can check by using the conditions we have on
𝑥0 and 𝑥1 . Hence 𝑞 is a regular point for 𝜋. Since 𝑞 was any point in the fiber
over 𝑎, we have shown that 𝑏 is a regular value for 𝜋.

(d) By a previous point, we need to show that 𝑑 𝜋̃𝑞 restricted to 𝑇𝑞 𝕊3 is surjective onto
𝑇𝜋(𝑞) 𝕊2 at every 𝑞 ∈ 𝕊3 . Since the tangent space 𝑇𝜋(𝑞) 𝕊2 of 𝕊2 is two-dimensional,
we need to check that the image of 𝑑 𝜋̃𝑞 restricted to Ker (𝑑(𝑔4 )𝑞 ) spans a two-
dimensional subspace. Since Ker (𝑑(𝑔4 )𝑞 ) is of dimension 3, it suffices to show
that 𝑑 𝜋̃𝑞 has rank 3, which implies that the kernel of 𝑑 𝜋̃𝑞 has dimension 1. Hence
we need to show that 𝑑 𝜋̃𝑞 always has 3 linear independent columns.
We can show this for example by calculating the determinants of appropriate 3 × 3-
minors. Ignoring the factor 2 in our formula for 𝑑 𝜋̃𝑞 we look at the minors 𝐴𝑗 of
the remaining matrix where we omit the 𝑗th column:

∙ The determinant of 𝐴4 is −𝑥1 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑥1 .

∙ The determinant of 𝐴3 is −𝑦0 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑦0 .
∙ The determinant of 𝐴2 is −𝑥0 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑥0 .
∙ The determinant of 𝐴1 is −𝑦1 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑦1 .

For every point 𝑞 ∈ 𝕊3 , at least one of the coordinates 𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 is nonzero.

Hence the matrix always has three linear independent columns and 𝑑 𝜋̃𝑞 has rank
3. This shows that each point in 𝕊3 is a regular point for 𝜋, and hence every point
in 𝕊2 is a regular value for 𝜋.
380 A.5. Lie groups
A.5 Lie groups

A.5.1 Lie groups

( )
𝑎 𝑏
Solution (Exercise 5.1) (a) Every 𝐴 = ∈ 𝑆𝑂(2) satisfies
𝑐 𝑑
( )( ) ( 2 ) ( )
𝑇 𝑎 𝑐 𝑎 𝑏 𝑎 + 𝑐 2 𝑎𝑏 + 𝑐𝑑 1 0
𝐴 𝐴= = = .
𝑏 𝑑 𝑐 𝑑 𝑎𝑏 + 𝑐𝑑 𝑏2 + 𝑑 2 0 1

Hence 𝐴 corresponds to two points (𝑎, 𝑐) and (𝑏, 𝑑) on 𝕊1 ⊂ ℝ2 whose correspond-

ing vectors are orthogonal to each other. Since we also know det 𝐴 = 𝑎𝑑 − 𝑏𝑐 = 1,
one of these(points uniquely) determines the other. See Figure A.2. Hence we can
cos 𝑡 − sin 𝑡
write 𝐴 as for some real number 𝑡. Now one can check that the
sin 𝑡 cos 𝑡
map ( )
1 cos 𝑡 − sin 𝑡
𝕊 → 𝑆𝑂(2), (cos 𝑡, sin 𝑡) →
sin 𝑡 cos 𝑡
is a diffeomorphism and Lie group isomorphism.
( )
𝑎 𝑏
(b) Every 𝐴 = ∈ 𝑆𝑈 (2) satisfies
𝑐 𝑑
( )( ) ( ) ( )
̄𝑇 𝑎̄ 𝑐̄ 𝑎 𝑏 ̄ + 𝑐𝑐
𝑎𝑎 ̄ + 𝑐𝑑
̄ 𝑎𝑏 ̄ 1 0
𝐴 𝐴= ̄ ̄ = ̄ ̄ 𝑏𝑏
̄ + 𝑑𝑑
̄ = .
𝑏 𝑑 𝑐 𝑑 𝑏𝑎 + 𝑑𝑐 0 1

Together with det 𝐴 = 𝑎𝑑 − 𝑏𝑐 = 1 we get four linear equations for the complex
numbers 𝑎, 𝑏, 𝑐, 𝑑, and their complex conjugates. Unraveling these equations
shows that we can write 𝐴 as
( )
𝑎 𝑏
𝐴= with 𝑎𝑎̄ + 𝑏𝑏̄ = 1.
−𝑏̄ 𝑎̄

Hence 𝐴 corresponds uniquely to a pair of complex numbers (𝑎, 𝑏) which satisfies

𝑎𝑎̄ + 𝑏𝑏̄ = 1. Since this is exactly the defining condition for elements of 𝕊3 ⊂ ℂ2 ,
we see that
( )
3 𝑎 𝑏
𝕊 → 𝑆𝑈 (2), (𝑎, 𝑏) →
−𝑏̄ 𝑎̄

is a diffeomorphism.

Solution (Exercise 5.2) (a) Since 𝑓 is bijective, it has an inverse 𝑓 −1 ∶ 𝑌 →

𝑋. We need to show that 𝑓 −1 is smooth. Let 𝑦 be a point in 𝑌 . Since 𝑓 is a
local diffeomorphism, there is an open neighborhood 𝑈 ⊂ 𝑋 around the point
𝑓 −1 (𝑦) and an open neighborhood 𝑉 ⊂ 𝑌 around 𝑦 such that 𝑓|𝑈 ∶ 𝑈 → 𝑉 is
a diffeomorphism. Hence there is a smooth inverse (𝑓|𝑈 )−1 ∶ 𝑉 → 𝑈 . Since
Appendix A. Solutions to exercises 381

Figure A.2: The column vectors are orthogonal to each other and they determine each other.

inverses are unique (as maps of sets), (𝑓 −1 )|𝑉 must agree with (𝑓|𝑈 )−1 . Hence 𝑓 −1
is a smooth map on an open neighborhood of 𝑦. Since 𝑦 was arbitrary, we see that
𝑓 −1 is smooth at every point and therefore smooth.

(b) Since 𝑓 is one-to-one, it is a bijection from 𝑋 onto its image Im (𝑓 ) ⊆ 𝑌 . Since

it is a local diffeomorphism, 𝑓 ∶ 𝑋 → Im (𝑓 ) is a bijective local diffeomorphism.
By the previous point, it is a diffeomorphism.

(c) We would like to show dim 𝑋 = rank(𝑓 ) = dim 𝑌 . Because then the Inverse
Function Theorem implies that 𝑓 is a local diffeomorphism, and, since 𝑓 is also
bijective, 𝑓 would be a diffeomorphism by the first point and we were done.
Assume 𝑋 ⊆ ℝ𝑀 and 𝑌 ⊆ 𝑅𝑁 , dim 𝑋 = 𝑚, dim 𝑌 = 𝑛, and set 𝑟 ∶= rank(𝑓 ).
By definition of the rank, we have 𝑚 ≥ 𝑟 and 𝑛 ≥ 𝑟. We want to show 𝑚 = 𝑟 = 𝑛.
For any point 𝑥 ∈ 𝑋, the linear map 𝑑𝑓𝑥 has rank 𝑟. Recall that for a linear
map 𝐿 ∶ ℝ𝑚 → ℝ𝑛 of rank 𝑟, we can choose a basis of ℝ𝑛 such that the first 𝑟
basis vectors 𝑏1 , … , 𝑏𝑟 span the image of 𝐿 and the remaining 𝑛 − 𝑟 basis vectors
𝑏𝑟+1 , … , 𝑏𝑛 span the orthogonal complement of 𝐿 in ℝ𝑛 . Then we choose a basis
of ℝ𝑚 such that the 𝑖th basis vector is sent to 𝑏𝑖 . The matrix representing 𝐿 in
these bases has the 𝑟 × 𝑟-identity matrix sitting in the upper left corner and zeros
elsewhere. Then, as in the proof of the Local Immersion (or Submersion) Theorem,
we can choose local parametrizations 𝜙 ∶ 𝑈 → 𝑋 around 𝑥 and 𝜓 ∶ 𝑉 → 𝑌
around 𝑦 such that the map 𝜃 ∶ 𝑈 → 𝑉 in the commutative diagram

𝑓
𝑋O /𝑌
O
𝜙 𝜓

𝑈 /𝑉
𝜃=𝜓 −1 ◦𝑓 ◦𝜙
382 A.5. Lie groups
has the form 𝜃(𝑥1 , … , 𝑥𝑚 ) = (𝑥1 , … , 𝑥𝑟 , 0) ∈ ℝ𝑛 . (Note that the 0 at the end of
𝜃(𝑥) only occur if 𝑟 < 𝑛.)
If 𝑚 > 𝑟, then for a sufficiently small 𝜖 > 0, 𝜃(𝑥1 , … , 𝑥𝑟 , 𝜖, 0) = (𝑥1 , … , 𝑥𝑟 , 0) and
𝜃 is not injective. Since 𝜙 and 𝜓 are diffeomorphisms, this would imply that 𝑓 is
not injective which contradicts that 𝑓 is bijective. Hence we can assume 𝑚 = 𝑟
and 𝑓 is an immersion.
Assume we had 𝑟 < 𝑛. Then, after possibly shrinking 𝑈 , we can assume that
𝑈 is a small open 𝐵𝜖 (0) around 0 in ℝ𝑚 and that 𝜃(𝐵̄ 𝜖 (0)) ⊆ 𝑉 (where 𝐵̄ 𝜖 (0)
denotes the closed ball of radius 𝜖: 𝐵̄ 𝜖 (0) = {𝑥 ∈ ℝ𝑚 ∶ |𝑥| ≤ 𝜖}). Since 𝐵̄ 𝜖 (0)
is compact, so is 𝜃(𝐵̄ 𝜖 (0)). Hence 𝜃(𝐵̄ 𝜖 (0)) is closed in 𝑉 and is contained in
𝑉 ∩ (ℝ𝑟 × {0}). Hence 𝜃(𝐵̄ 𝜖 (0)) does not contain any open subsets of 𝑉 . Since
𝜙 and 𝜓 are diffeomorphisms, this implies that 𝑓 (𝜙(𝐵̄ 𝜖 (0))) is closed and does
not contain any nonempty open subset of 𝑌 . Since we can cover 𝑋 by such local
parametrizations, we see that 𝑓 (𝑋) is the union of subsets which do not contain
any nonempty open subset of 𝑌 .
Now if 𝑋 could be assumed to be compact, then 𝑓 (𝑋) is compact, and 𝑓 (𝑋) can
be covered by finitely many closed subsets which do not contain any nonempty
open subset of 𝑌 . That would imply that 𝑓 (𝑋) is itself a closed subset which does
not contain any nonempty open subset of 𝑌 . Hence 𝑓 (𝑋) cannot be all of 𝑌 , and
𝑓 would not be surjective.
In general, for any open cover of a subspace in ℝ𝑀 , we can always choose a count-
able subcover. This implies that 𝑓 (𝑋) is the countable union of subsets which do
not contain any nonempty open subset of 𝑌 . By Baire’s Category Theorem, this
implies that 𝑓 (𝑋) does not contain any nonempty open subset of 𝑌 . Hence 𝑓 (𝑋)
cannot be equal 𝑌 and 𝑓 would not be surjective.

(d) We learned in Theorem 5.3 that a Lie group homomorphism has constant rank.
Hence we just need to apply the previous point.

Solution (Exercise 5.3) For any 𝑔 ∈ 𝐺, left multiplication 𝐿𝑔 ∶ 𝐺 → 𝐺 by 𝑔 maps

the subgroup 𝐻 to the left coset 𝑔𝐻 = {𝑔ℎ ∶ ℎ ∈ 𝐻}. Since 𝐻 is open and 𝐿𝑔 is a
diffeomorphism, the coset 𝑔𝐻 is open. Thus, 𝐺 can be written as the union of the open
subsets 𝑔𝐻 where 𝑔 ranges over all elements in 𝐺. But since cosets are pairwise disjoint,
this would give us a way to write 𝐺 as the union of nonempty disjoint open subsets. Since
G is connected, there can be only one coset. Therefore, 𝐻 = 𝐺.

Solution (Exercise 5.4) (a) Let 𝑔, ℎ ∈ 𝐺 be any fixed elements. Let 𝑗 ∶ 𝐺 → 𝐺 ×𝐺

be the map 𝑗(𝑔) = (𝑔, ℎ). Note that the composite 𝜇◦𝑗 = 𝑅ℎ is right translation by
ℎ.
For 𝑥 ∈ 𝐺, let 𝜙𝑥 ∶ 𝑈𝑥 → 𝐺 be a local parametrization around 𝑥 with 𝜙(0) = 𝑥.
Appendix A. Solutions to exercises 383
Then we get the diagram

𝑗 𝜇
𝐺O /𝐺×𝐺 /𝐺
O O
𝜙𝑔 𝜙𝑔 ×𝜙ℎ 𝜙𝑔ℎ

𝑈𝑔 /𝑈 ×𝑈 /𝑈 .
𝛾 𝑔 ℎ 𝜃 𝑔ℎ

where we define the maps 𝛾 and 𝜃 such that the diagram commutes. Since 𝜙𝑔 (0) =
𝑔 and 𝜙ℎ (0) = ℎ, we must have 𝛾(𝑢) = (𝑢, 0) ∈ 𝑈𝑔 × 𝑈ℎ to make the left hand
diagram commute. Moreover, we must have 𝜃(0, 0) = 0 ∈ 𝑈𝑔ℎ .
Taking derivatives at 𝑔 and using 𝑇(𝑔,ℎ) (𝐺 × 𝐺) = 𝑇𝑔 (𝐺) × 𝑇ℎ (𝐺) gives

𝑑𝑗𝑔 𝑑𝜇(𝑔,ℎ)
𝑇𝑔 (𝐺) / 𝑇 (𝐺) × 𝑇 (𝐺) / 𝑇 (𝐺)
𝑔
O O ℎ 𝑔ℎ
O
𝑑(𝜙𝑔 )0 𝑑(𝜙𝑔 )0 ×𝑑(𝜙ℎ )0 𝑑(𝜙𝑔ℎ )0

ℝ𝑛 / ℝ𝑛 × ℝ𝑛 / ℝ𝑛 .
𝑑𝛾0 𝑑𝜃0

Since 𝛾(𝑢) = (𝑢, 0), we have 𝑑𝛾0 (𝑣) = (𝑣, 0) and hence 𝑑𝑗𝑔 (𝑋) = (𝑋, 0). Since
𝜇◦𝑗 = 𝑅ℎ , we have 𝑑𝜇(𝑔,ℎ) ◦𝑑𝑗𝑔 = 𝑑(𝑅ℎ )𝑔 . Thus

𝑑𝜇(𝑔,ℎ) (𝑋, 0) = 𝑑𝜇(𝑔,ℎ) (𝑑𝑗𝑔 (𝑋)) = 𝑑(𝑅ℎ )𝑔 (𝑋).

Repeating this argument with 𝑗 replaced with 𝑗 ∶ ℎ → (𝑔, ℎ) yields

𝑑𝜇(𝑔,ℎ) (0, 𝑌 ) = 𝑑(𝐿𝑔 )ℎ (𝑌 ).

Since 𝑑𝜇(𝑔,ℎ) is linear, it satisfies

𝑑𝜇(𝑔,ℎ) (𝑋, 𝑌 ) = 𝑑𝜇(𝑔,ℎ) (𝑋, 0) + 𝑑𝜇(𝑔,ℎ) (0, 𝑌 ) = 𝑑(𝑅ℎ )𝑔 (𝑋) + 𝑑(𝐿𝑔 )ℎ (𝑌 ).

(b) Let 𝜄 ∶ 𝐺 → 𝐺 denote the inversion map. Show that

𝑑𝜄𝑒 ∶ 𝑇𝑒 (𝐺) → 𝑇𝑒 (𝐺)

is given by 𝑑𝜄𝑒 (𝑋) = −𝑋.

Solution:
Consider the map
(Id,𝜄) 𝜇
← 𝐺, 𝑔 → (𝑔, 𝑔 −1 ) → 𝑔𝑔 −1 = 𝑒.
← 𝐺 × 𝐺 ←←→
𝐺 ←←←←←←←←→

Since this map is constant, its derivative at 𝑒 vanishes. Hence we get

(𝑑Id𝑒 ,𝑑𝜄𝑒 ) 𝑑𝜇(𝑒,𝑒)
𝑇𝑒 (𝐺) ←←←←←←←←←←←←←←←←→
← 𝑇𝑒 (𝐺) × 𝑇𝑒 (𝐺) ←←←←←←←←←←→
← 𝑇𝑒 (𝐺), 𝑋 → (𝑋, 𝑑𝜄𝑒 (𝑋)) → 0.

As we have just learned 𝑑𝜇(𝑒,𝑒) (𝑋, 𝑑𝜄𝑒 (𝑋)) = 𝑋 + 𝑑𝜄𝑒 (𝑋) = 0, and hence 𝑑𝜄𝑒 (𝑋) =
−𝑋.
384 A.5. Lie groups

(c) Given 𝑔 ∈ 𝐺, we consider the diagram

𝐺
𝜄 /𝐺
O
𝐿𝑔−1 𝑅𝑔−1

𝐺 / 𝐺.
𝜄

One easily checks that it commutes. Taking the derivative at 𝑔 of the top map
yields a commutative diagram of derivatives

𝑑𝜄𝑔
𝑇𝑔 (𝐺) / 𝑇 −1 (𝐺)
𝑔
O
𝑑(𝐿𝑔−1 )𝑔 𝑑(𝑅𝑔−1 )𝑒

𝑇𝑒 (𝐺) / 𝑇 (𝐺).
𝑑𝜄𝑒 𝑒

We just calculated the effect of the map 𝑑𝜄𝑒 ∶ 𝑇𝑒 (𝐺) → 𝑇𝑒 (𝐺) as 𝑋 → −𝑋. Hence,
since all maps in the above diagram are linear, we get

𝑑𝜄𝑔 ∶ 𝑇𝑔 (𝐺) → 𝑇𝑔−1 , 𝑌 → −𝑑(𝑅𝑔−1 )𝑒 (𝑑(𝐿𝑔−1 )𝑔 (𝑌 )).

Solution (Exercise 5.5) Given elements 𝑔, ℎ ∈ 𝐺. Let 𝑅ℎ−1 denote the right translation
with ℎ−1 . We define the smooth map 𝑗ℎ by

𝑗ℎ ∶ 𝐺 → 𝐺 × 𝐺, 𝑥 → (𝑅ℎ−1 (𝑥), ℎ).

Note that 𝑗ℎ (𝑔ℎ) = (𝑔ℎℎ−1 , ℎ) = (𝑔, ℎ) ∈ 𝐺 × 𝐺. For the tangent spaces we get

𝑇𝑗ℎ (𝑔ℎ) (𝐺 × 𝐺) = 𝑇(𝑔,ℎ) (𝐺 × 𝐺) ≅ 𝑇𝑔 (𝐺) × 𝑇ℎ (𝐺).

The composite of the map

𝑗ℎ 𝜇
← 𝐺, 𝑥 → (𝑅ℎ−1 (𝑥), ℎ) → 𝜇(𝑥ℎ−1 , ℎ) = 𝑥
← 𝐺 × 𝐺 ←←→
𝐺 ←←←→

is the identity of 𝐺. Taking derivatives at 𝑔ℎ yields

𝑑(𝑗ℎ )𝑔ℎ 𝑑𝜇(𝑔,ℎ)
𝑇𝑔ℎ (𝐺) ←←←←←←←←←←←←→
← 𝑇𝑔 (𝐺) × 𝑇ℎ (𝐺) ←←←←←←←←←←←→
← 𝑇𝑔ℎ (𝐺).

Since 𝜇◦𝑗ℎ = Id𝐺 , we also have 𝑑𝜇(𝑔,ℎ) ◦𝑑(𝑗ℎ )𝑔ℎ = Id𝑇𝑔ℎ (𝐺) . In particular,

𝑑𝜇(𝑔,ℎ) ∶ 𝑇(𝑔,ℎ) (𝐺 × 𝐺) → 𝑇𝑔ℎ (𝐺)

is surjective. Since we started with arbitrary elements 𝑔 and ℎ, this shows that 𝜇 is a
submersion.

Solution (Exercise 5.6) For matrices 𝐴 ∈ 𝐺𝐿(𝑛) and 𝐵 ∈ 𝑀(𝑛), we have

det(𝐵) = (det 𝐴) ⋅ det(𝐴−1 𝐵) = (det 𝐴) ⋅ det(𝐿𝐴−1 𝐵).

Appendix A. Solutions to exercises 385
Taking the derivative at 𝐴 and remembering the chain rule yields

𝑑(det)𝐴 (𝐵) = 𝑑((det 𝐴) ⋅ det ◦𝐿𝐴−1 )(𝐵)

= (det 𝐴) ⋅ 𝑑(det ◦𝐿𝐴−1 )𝐵
= (det 𝐴) ⋅ 𝑑(det)(𝐿 ◦𝑑(𝐿𝐴−1 )𝐴 (𝐵)
𝐴−1 𝐴)

= (det 𝐴) ⋅ 𝑑(det)𝐼 (𝐴−1 𝐵)

= (det 𝐴) ⋅ tr (𝐴−1 𝐵),

where we have used 𝑑(𝐿𝐴−1 )𝐴 (𝐵) = 𝐴−1 𝐵 which can be easily checked, since matrix
multiplication is linear. Hence, after replacing 𝐵 with 𝐴𝐵, we get

𝑑(det)𝐴 (𝐴𝐵) = (det 𝐴) ⋅ (tr 𝐵) for all 𝐵 ∈ 𝑀(𝑛).

386 A.6. Transversality
A.6 Transversality

A.6.1 Transversality

Solution (Exercise 12.1) (a) The tangent space to 𝕊1 at 𝑧 is 𝑇𝑧 (𝕊1 ) = {𝑣 ∈ ℝ2 ∶

𝑣 ⋅ 𝑧 = 0}. The tangent space to 𝑁𝑧 at 𝑧 is 𝑇𝑧 (𝑁𝑧 ) = {(0, 𝑦) ∶ 𝑦 ∈ ℝ} ⊂ ℝ2 . Since
𝑇𝑧 (ℝ2 ) = ℝ2 , 𝑇𝑧 (𝕊1 ) and 𝑇𝑧 (𝑁𝑧 ) span all of 𝑇𝑧 (ℝ2 ) for all 𝑧 ≠ (±1, 0). Hence
𝕊1 − ⋔ 𝑁𝑧 if and only if 𝑧 ≠ (±1, 0). (Drawing a picture explains everything.)

(b) Which of the following linear spaces intersect transversally?

∙ The 𝑥𝑦-plane and the 𝑧-axis.

Answer: Transverse, since the two spaces span all of ℝ3 .
∙ The 𝑥𝑦-plane and the plane spanned by {(3, 2, 0), (0, 4, −1)}.
Answer: Transverse, since the two planes span all of ℝ3 .
∙ The plane spanned by {(1, 0, 0), (2, 1, 0)} and the 𝑦-axis in ℝ3 .
Answer: Not transverse, since the 𝑦-axis and lies in the span of
{(1, 0, 0), (2, 1, 0)}.
∙ ℝ𝑘 × {0} and {0} × ℝ𝑙 in ℝ𝑛 . (The answer depends on 𝑘, 𝑙, and 𝑛.)
Answer: Transverse, if 𝑘 + 𝑙 ≥ 𝑛.
∙ 𝑉 × {0} and the diagonal in 𝑉 × 𝑉 , for a real vector space 𝑉 .
Answer: Transverse, since they span all of 𝑉 × 𝑉 : any (𝑣, 𝑤) ∈ 𝑉 × 𝑉 is
equal to the sum of (𝑤, 𝑤) ∈ Δ𝑉 and (𝑣 − 𝑤, 0) ∈ 𝑉 × {0}.
∙ The spaces of symmetric (𝐴𝑡 = 𝐴) and skew symmetric (𝐴𝑡 = −𝐴) matrices
in 𝑀(𝑛).
Answer: Transverse, since every matrix in 𝑀(𝑛) can be written as a sum of
a symmetric and an antisymmetric matrix:

1 1
𝐶 = (𝐶 + 𝐶 𝑡 ) + (𝐶 − 𝐶 𝑡 ) for any 𝐶 ∈ 𝑀(𝑛).
2 2

(c) Yes, 𝑆𝐿(𝑛) and 𝑂(𝑛) do not meet transversally in 𝑀(𝑛), since 𝑆𝑂(𝑛) is contained
in 𝑆𝐿(𝑛). Hence we also have 𝑇𝐴 (𝑂(𝑛)) = 𝑇𝐴 (𝑆𝑂(𝑛)) ⊆ 𝑇𝐴 (𝑆𝐿(𝑛)), and these
tangent spaces do not span all of 𝑇𝐴 (𝑀(𝑛)) = 𝑀(𝑛).

Solution (Exercise 12.2) The image of 𝑓 is a submanifold of ℝ2 . This follows, for

example, from the fact that 𝑓 is an embedding. We could also observe that Im (𝑓 ) =
𝑔 −1 (1) and remark that 1 is a regular value of 𝑔. The composition 𝑔◦𝑓 is the constant
map ℝ → ℝ with value 1. Hence (𝑔◦𝑓 )−1 (1) = ℝ is a manifold.

Solution (Exercise 12.3) By assumption, 𝑔 is transverse to 𝑊 , i.e.,

Im (𝑑𝑔𝑦 ) + 𝑇𝑔(𝑦) (𝑊 ) = 𝑇𝑔(𝑦) (𝑍) for all 𝑦 ∈ 𝑌 with 𝑔(𝑦) ∈ 𝑊 . (A.2)

Appendix A. Solutions to exercises 387

Now we assume 𝑓 −
⋔ 𝑔 −1 (𝑊 ), i.e.,

Im (𝑑𝑓𝑥 ) + 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) = 𝑇𝑓 (𝑥) (𝑌 ) for all 𝑥 ∈ 𝑋 with 𝑓 (𝑥) ∈ 𝑔 −1 (𝑊 ). (A.3)

We need to show (𝑔◦𝑓 ) − ⋔ 𝑊 . So let 𝑥 ∈ 𝑋 be a point such that 𝑔(𝑓 (𝑥)) ∈ 𝑊 and let
𝑐 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑍). By (A.6), there are vectors 𝑏1 ∈ 𝑇𝑓 (𝑥) (𝑌 ) and 𝑏2 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) such
that
𝑑𝑔𝑓 (𝑥) (𝑏1 ) + 𝑏2 = 𝑐.
By (A.7), there are vectors 𝑎 ∈ 𝑇𝑥 (𝑋) and 𝑏3 ∈ 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) such that

𝑑𝑓𝑥 (𝑎) + 𝑏3 = 𝑏1 .

Putting these two equations together we get

𝑐 = 𝑑𝑔𝑓 (𝑥) (𝑑𝑓𝑥 (𝑎) + 𝑏3 ) + 𝑏2 = 𝑑𝑔𝑓 (𝑥) (𝑑𝑓𝑥 (𝑎1 )) + 𝑑𝑔𝑓 (𝑥) (𝑏3 ) + 𝑏2 .

By the chain rule, we have 𝑑(𝑔◦𝑓 )𝑥 = 𝑑𝑔𝑓 (𝑥) ◦𝑑𝑓𝑥 . By a previous exercise, we
know 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) = (𝑑𝑔𝑓 (𝑥) )−1 (𝑇𝑔(𝑓 (𝑥)) (𝑊 ) in 𝑇𝑓 (𝑥) (𝑌 ). In particular, 𝑑𝑔𝑓 (𝑥) (𝑏3 ) ∈
𝑇𝑔(𝑓 (𝑥)) (𝑊 ) and thus
𝑑𝑔𝑓 (𝑥) (𝑏3 ) + 𝑏2 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ).
Hence we have

𝑐 = 𝑑(𝑔◦𝑓 )𝑥 (𝑎) + 𝑑𝑔𝑓 (𝑥) (𝑏3 ) + 𝑏2 ∈ [Im (𝑑(𝑔◦𝑓 )𝑥 ) + 𝑇𝑔(𝑓 (𝑥)) (𝑊 )] ⊂ 𝑇𝑔(𝑓 (𝑥)) (𝑍).

Since 𝑐 was an arbitrary element in 𝑇𝑔(𝑓 (𝑥)) (𝑌 ), we have proven

Im (𝑑(𝑔◦𝑓 )𝑥 ) + 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) = 𝑇𝑔(𝑓 (𝑥)) (𝑍) for all 𝑥 ∈ 𝑋 with 𝑔(𝑓 (𝑥)) ∈ 𝑊 .

In other words, (𝑔◦𝑓 ) −

⋔𝑊.
Now we assume (𝑔◦𝑓 ) −
⋔ 𝑊 , i.e.,

Im (𝑑(𝑔◦𝑓 )𝑥 ) + 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) = 𝑇𝑔(𝑓 (𝑥)) (𝑍) for all 𝑥 ∈ 𝑋 with 𝑔(𝑓 (𝑥)) ∈ 𝑊 . (A.4)

Let 𝑥 ∈ 𝑋 be a point such that 𝑓 (𝑥) ∈ 𝑔 −1 (𝑊 ) and let 𝑏 ∈ 𝑇𝑓 (𝑥) (𝑌 ). Since 𝑔(𝑓 (𝑥)) ∈ 𝑊 ,
we can use (A.8) to find vectors 𝑎 ∈ 𝑇𝑥 (𝑋) and 𝑐 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) such that

𝑑(𝑔◦𝑓 )𝑥 (𝑎) + 𝑐 = 𝑑𝑔𝑓 (𝑥) (𝑏).

By the chain rule, we have 𝑑(𝑔◦𝑓 )𝑥 (𝑎) = 𝑑𝑔𝑓 (𝑥) (𝑑𝑓𝑥 (𝑎)). Thus

𝑑𝑔𝑓 (𝑥) (𝑏 − 𝑑𝑓𝑥 (𝑎)) = 𝑐 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ),

In other words,

𝑏 − 𝑑𝑓𝑥 (𝑎) ∈ (𝑑𝑔𝑓 (𝑥) )−1 (𝑇𝑔(𝑓 (𝑥)) (𝑊 )) = 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )).

Since 𝑏 was an arbitrary element, we have proven Im (𝑑𝑓𝑥 ) + 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) = 𝑇𝑓 (𝑥) (𝑌 )

for all 𝑥 ∈ 𝑋 with 𝑓 (𝑥) ∈ 𝑔 −1 (𝑊 ). In other words, 𝑓 −
⋔ 𝑔 −1 (𝑊 ).
388 A.6. Transversality

Solution (Exercise 12.4) By definition, Γ(𝐴) −

⋔ Δ if and only if Γ(𝐴)+Δ(𝑉 ) = 𝑉 ×𝑉 .
Let (𝑣1 , 𝑣2 ) be an arbitrary element of 𝑉 × 𝑉 . We need to check under which conditions
we can find 𝑣, 𝑤 ∈ 𝑉 such that

(𝑣1 , 𝑣2 ) = (𝑣, 𝑣) + (𝑤, 𝐴𝑤) = (𝑣 + 𝑤, 𝑣 + 𝐴𝑤), i.e. 𝑣1 = 𝑣 + 𝑤 and 𝑣2 = 𝑣 + 𝐴𝑤.

If we can find a suitable 𝑤, then we just set 𝑣 ∶= 𝑣1 − 𝑤. Hence, by taking the difference
of the two equations, we see that the question is reduced to checking whether we can find
a 𝑤 such that 𝑣2 − 𝑣1 = 𝐴𝑤 − 𝑤 = (𝐴 − 𝐼)𝑤 where 𝐼 is the identity map of 𝑉 . But such
a 𝑤 exists for any choice of 𝑣1 and 𝑣2 if and only if 𝐴 − 𝐼 is invertible, i.e. if and only if
det(𝐴 − 𝐼) ≠ 0 which happens if and only if +1 is not an eigenvalue of 𝐴 (because the
eigenvalues are the 𝜆 such that det(𝐴 − 𝜆𝐼) = 0).

Solution (Exercise 12.5) Let Δ𝑋 = {(𝑥, 𝑥) ∶ 𝑥 ∈ 𝑋} ⊆ 𝑋 × 𝑋 be the diagonal of

𝑋 and Γ(𝑓 ) = {(𝑥, 𝑓 (𝑥)) ∶ 𝑥 ∈ 𝑋} ⊆ 𝑋 × 𝑋 be the graph of 𝑓 . Then the set of fixed
points of 𝑓 in 𝑋 is the intersection Δ𝑋 ∩ Γ(𝑓 ). For
𝑥 = 𝑓 (𝑥) ⇐⇒ (𝑥, 𝑥) = (𝑥, 𝑓 (𝑥)) ⇐⇒ (𝑥, 𝑥) ∈ Γ(𝑓 ).
Recall that a 0-dimensional manifold is just a discrete set of points. Since 𝑋 is compact,
𝑋 × 𝑋 is also compact. Hence a 0-dimensional submanifold is a discrete subset of the
compact space 𝑋 × 𝑋 and is therefore finite. Note that we have used this before: dis-
crete closed subspaces of compact spaces are finite. Thus, in order to prove that 𝑓 has
only finitely many fixed points, it suffices to show that Δ𝑋 ∩ Γ(𝑓 ) is a 0-dimensional
submanifold of 𝑋 × 𝑋.
Hence we would like to show that Δ𝑋 and Γ(𝑓 ) meet transversally in 𝑋 × 𝑋. Since
both have codimension equal to dim 𝑋 =∶ 𝑛 in 𝑋 × 𝑋, the Transversality Theorem
then implies that Δ𝑋 ∩ Γ(𝑓 ) is a 0-dimensional submanifold of 𝑋 × 𝑋. By definition,
Δ𝑋 −⋔ Γ(𝑓 ) means
𝑇(𝑥,𝑥) (Γ(𝑓 )) + 𝑇(𝑥,𝑥) (Δ𝑋 ) = 𝑇(𝑥,𝑥) (𝑋 × 𝑋)
for every point (𝑥, 𝑥) ∈ Δ𝑋 ∩ Γ(𝑓 ). We know 𝑇(𝑥,𝑥) (Γ(𝑓 )) = Γ(𝑑𝑓𝑥 ) and 𝑇(𝑥,𝑥) (Δ𝑋 ) =
Δ𝑇𝑥 (𝑋) by a previous exercise. Moreover, we know 𝑇(𝑥,𝑥) (𝑋 × 𝑋) = 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋).
Hence we need to show
Γ(𝑑𝑓𝑥 ) + Δ𝑇𝑥 (𝑋) = 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋) (A.5)
for every point (𝑥, 𝑥) ∈ Δ𝑋 ∩ Γ(𝑓 ). This means exactly Γ(𝑑𝑓𝑥 ) −⋔ Δ𝑇𝑥 (𝑋) which we have
shown to be true if +1 is not an eigenvalue of 𝑑𝑓𝑥 in the previous exercise. Thus we can
stop here, since 𝑓 is Lefschetz by assumption.
But we could also just continue and give another proof as follows: Since we know
dim Δ𝑇𝑥 (𝑋) = dim Γ(𝑑𝑓𝑥 ) = dim 𝑇𝑥 (𝑋),
equality (A.9) will follow once we show Γ(𝑑𝑓𝑥 ) ∩ Δ𝑇𝑥 (𝑋) = {0}. For then we have
shown that Γ(𝑑𝑓𝑥 ) + Δ𝑇𝑥 (𝑋) is a subspace of 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋) of the same dimension as
𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋). (Recall that in a finite dimensional vector space 𝑉 with subspaces 𝑈 and
𝑊 , the following dimension formula holds:
dim 𝑈 + dim 𝑊 = dim(𝑈 + 𝑊 ) − dim(𝑈 ∩ 𝑊 )
Appendix A. Solutions to exercises 389
where 𝑈 + 𝑊 ⊆ 𝑉 is the subspace of 𝑉 generated by 𝑈 and 𝑊 and 𝑈 ∩ 𝑊 is their
intersection.)
Since 𝑓 is a Lefschetz map, we know that for every fixed point 𝑥 of 𝑓 , +1 is not an
eigenvalue of 𝑑𝑓𝑥 . This means that 𝑑𝑓𝑥 does not have any fixed points, for there is no
𝑣 ∈ 𝑇𝑥 (𝑋) ⧵ {0} with 𝑑𝑓𝑥 (𝑣) = 1 ⋅ 𝑣. As we observed for 𝑓 above, this is equivalent to

Γ(𝑑𝑓𝑥 ) ∩ Δ𝑇𝑥 (𝑋) = {0}.

Solution (Exercise 12.6) We define the map

𝑓 ∶ ℂ5 ⧵ {0} → ℂ,
(𝑧1 , … , 𝑧5 ) → 𝑧21 + 𝑧22 + 𝑧23 + 𝑧34 + 𝑧6𝑘−1
5
.

Setting 𝑍 = 𝑓 −1 (0), we need to show that 𝑍 and 𝕊9 meet transversally. The tangent
space to 𝑍 in a point 𝑧 ∈ 𝑍 is the kernel of the derivative 𝑑𝑓𝑧 . Since 𝑓 is a polynomial
in the variables 𝑧1 , … , 𝑧5 , we can use our usual rules for partial differentiation to get the
following matrix for 𝑑𝑓𝑧 (in the standard basis):
( )
𝑑𝑓𝑧 ∶ ℂ5 → ℂ, 𝑑𝑓𝑧 = 2𝑧1 , 2𝑧2 , 2𝑧3 , 3𝑧24 , (6𝑘 − 1)𝑧6𝑘−25
.
( )
𝑥 −𝑦
Recall that we can represent every element 𝑥+𝑖𝑦 ∈ ℂ by the real 2×2-matrix .
𝑦 𝑥
Then we see that 𝑑𝑓𝑧 is a real 2 × 10-matrix. Its maximal rank (as a matrix with entries in
ℝ) is therefore 2. And, in fact, for every 𝑧 ≠ 0, 𝑑𝑓𝑧 has rank 2, since it maps surjectively
onto ℂ ≅ ℝ2 . Thus 0 is a regular value for 𝑓 and the tangent space 𝑇𝑧 (𝑍) is the kernel
of 𝑑𝑓𝑧 .
Writing a complex number 𝑧 = 𝑥 + 𝑖𝑦, we can express 𝕊9 as the fiber of the smooth
map

𝑔 ∶ ℂ5 ≅ ℝ10 → ℝ,
(𝑧1 , … , 𝑧5 ) → 𝑥21 + 𝑦21 + 𝑥22 + 𝑦22 + ⋯ + 𝑥25 + 𝑦25 − 1

at the regular value 0, i.e., 𝕊9 = 𝑔 −1 (0) ⊂ ℂ5 ≅ ℝ10 . The tangent space to 𝕊9 at 𝑧 is then
given by the kernel of the derivative (in standard bases)
( )
𝑑𝑔𝑧 ∶ ℂ5 = ℝ10 → ℝ, 𝑑𝑔𝑧 = 2𝑥1 , 2𝑦1 , 2𝑥2 , 2𝑦2 , … , 2𝑥5 , 2𝑦5 .

Thus, as expected, the tangent space 𝑇𝑧 (𝕊9 ) consists of all vectors 𝑤 in ℝ10 which are
orthogonal to 𝑧, i.e., which satisfy 𝑤 ⋅ 𝑧 = 0.
The tangent space of 𝕊9 is of dimension 9 and the tangent space of ℝ10 ⧵ {0} is of
dimension 10. Hence in order to show that 𝑍 and 𝕊9 meet transversally in ℝ10 ⧵ {0} we
need to show: For every 𝑧 ∈ 𝑍 ∩ 𝕊9 , there is at least one vector 𝑤 in 𝑇𝑧 (𝑍) which does
not belong to 𝑇𝑧 𝕊9 . Then we have 𝑇𝑧 (𝑍) + 𝑇𝑧 (𝕊9 ) ⊆ 𝑇𝑧 (ℝ10 ⧵ {0}) is a vector subspace
of the same dimension as 𝑇𝑧 (ℝ10 ⧵ {0}) and therefore equal 𝑇𝑧 (ℝ10 ⧵ {0}).
So let 𝑧 = (𝑧1 , … , 𝑧5 ) be a fixed point in 𝑍 ∩ 𝕊9 . The tangent space 𝑇𝑧 (𝑍) is the
kernel of 𝑑𝑓𝑧 . Hence we need to find at least one vector 𝑤 ∈ ℂ5 = ℝ10 with 𝑑𝑓𝑧 (𝑤) = 0
and 𝑤 ⋅ 𝑧 ≠ 0.
390 A.6. Transversality

Set 𝑚 ∶= 2 ⋅ 3 ⋅ (6𝑘 − 1) and 𝑤 ∶= ( 𝑚2 𝑧1 , 𝑚2 𝑧2 , 𝑚2 𝑧3 , 𝑚3 𝑧4 , 6𝑘−1

𝑚
𝑧5 ). Then we have
𝑚
⎛ 2 𝑧1 ⎞
⎜ 𝑚 𝑧2 ⎟
( ) ⎜ 𝑚2 ⎟
𝑑𝑓𝑧 (𝑤) = 2𝑧1 , 2𝑧2 , 2𝑧3 , 3𝑧24 , (6𝑘 − 1)𝑧6𝑘−2
5
⋅ ⎜ 2 𝑧3 ⎟
⎜ 𝑚 𝑧4 ⎟
⎜ 𝑚3 ⎟
⎝ 6𝑘−1 𝑧5 ⎠
( )
= 𝑚 𝑧21 + 𝑧22 + 𝑧23 + 𝑧34 + 𝑧6𝑘−1
5
=0

since 𝑧 is by assumption a point on 𝑍. Hence 𝑤 ∈ 𝑇𝑧 (𝑍).

On the other hand, we can calculate the inner product of 𝑤 and 𝑧 as vectors in ℝ10 ,
using (𝑥𝑖 , 𝑦𝑖 ) for the coordinates of 𝑧𝑖 in ℝ2 ≅ ℂ, and get

⎛𝑥1 ⎞
⎜ 𝑦1 ⎟
( ) ⎜𝑥 ⎟
𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚
𝑤⋅𝑧= 𝑥 , 𝑦 , 𝑥 , 𝑦 , 𝑥 , 𝑦 , 𝑥 , 𝑦 , 𝑥 , 𝑦 ⋅ ⎜ 2⎟
2 1 2 1 2 2 2 2 2 3 2 3 3 4 3 4 6𝑘 − 1 5 6𝑘 − 1 5 ⎜ ⋮ ⎟
⎜𝑥 ⎟
⎜ 5⎟
⎝ 𝑦5 ⎠
𝑚 2 𝑚 2 𝑚 2 𝑚 2 𝑚 2
= |𝑧1 | + |𝑧2 | + |𝑧3 | + |𝑧4 | + |𝑧 |
2 2 2 3 6𝑘 − 1 5
>0

which is bigger than zero, since 𝑧 is a point on 𝕊9 . Thus 𝑤 is a vector in 𝑇𝑧 (ℝ10 ) which
is in 𝑇𝑧 (𝑍), but not in 𝑇𝑧 (𝕊9 ), and we have shown

𝑇𝑧 (𝑍) + 𝑇𝑧 (𝕊9 ) = 𝑇𝑧 (ℝ10 ⧵ {0}).

Hence 𝑍 and 𝕊9 meet transversally in ℝ10 ⧵{0}. The codimension of 𝑍 ∩𝕊9 in ℝ10 ⧵{0}
is 2 + 1 by the codimension formula. Thus dim(𝑍 ∩ 𝕊9 ) = 10 − 3 = 7.
Appendix A. Solutions to exercises 391
A.7 Smooth Homotopy

Solution (Exercise 8.1) (a) Let 𝑓 and 𝑔 be two smooth maps 𝑌 → 𝑋. Let 𝐹 be
a homotopy from the identity map of 𝑋 to the constant map 𝑋 → {𝑥0 } for some
𝑥0 ∈ 𝑋. Then we can use 𝐹 to define a homotopy from 𝑓 to 𝑌 → {𝑥0 } and a
homotopy from 𝑌 → {𝑥0 } to 𝑔. Setting these two homotopies together yields a
homotopy 𝐻 from 𝑓 to 𝑔. It only remains to make sure that 𝐻 is smooth. To
achieve this we apply the technique used in the main text. After composing with
a smooth bump function, we can assume 𝐹 (𝑥, 𝑡) = 𝑥 for all (𝑥, 𝑡) ∈ 𝑋 × [0, 1∕4]
and 𝐹 (𝑥, 𝑡) = 𝑥0 for all (𝑥, 𝑡) ∈ 𝑋 × [3∕4, 1]. Then we can define 𝐻 by
{
𝐹 (𝑓 (𝑦), 2𝑡) 𝑡 ∈ [0, 1∕2]
𝐻 ∶ 𝑌 × [0, 1] → 𝑋, (𝑦, 𝑡) →
𝐹 (𝑔(𝑦), 2(1 − 𝑡)) 𝑡 ∈ [1∕2, 1].

(b) Let 𝑌 = 𝑋 and let 𝑓 ∶ 𝑋 → 𝑋 be the identity and 𝑔 ∶ 𝑋 → {𝑥0 } ⊂ 𝑋 be the

constant map for some point 𝑥0 ∈ 𝑋. By the assumption, 𝑓 and 𝑔 are homotopic.
Hence 𝑋 is contractible.

(c) The map

𝐹 ∶ ℝ𝑘 × [0, 1] → ℝ𝑘 , (𝑥, 𝑡) → (1 − 𝑡)𝑥
is a smooth homotopy from the identity map to the constant map ℝ𝑘 → {0}.

Solution (Exercise 8.2) By Sard’s Theorem 7.1, there is a regular value 𝑦 ∈ 𝕊𝑛 for
𝑓 . Assume there is a point 𝑥 ∈ 𝑓 −1 (𝑦). Since dim 𝑇𝑥 𝑋 = 𝑘 and dim 𝑇𝑦 𝕊𝑛 = 𝑛, 𝑑𝑓𝑥
cannot be surjective if 𝑘 < 𝑛. Thus, if 𝑘 < 𝑛, then 𝑓 −1 (𝑦) must be empty. Hence we can
assume that the image of 𝑓 is contained in 𝑈 ∶= 𝕊𝑛 ⧵{𝑦}. Now we can use stereographic
projection from 𝑦 to define a diffeomorphism 𝜓 ∶ 𝑈 → ℝ𝑛 . Thus, 𝜓◦𝑓 is homotopic to a
constant map. Composing the homotopy with the inverse of 𝜓 defines a homotopy from
𝑓 to a constant map.

Solution (Exercise 8.3) For 𝑘 = 1, the antipodal map is (𝑥, 𝑦) → (−𝑥, −𝑦). The map
( )( )
1 1 cos(𝜋𝑡) − sin(𝜋𝑡) 𝑥
𝐹1 ∶ 𝕊 × [0, 1] → 𝕊 , ((𝑥, 𝑦), 𝑡) → .
sin(𝜋𝑡) cos(𝜋𝑡) 𝑦

is a smooth homotopy from the identity of 𝕊1 to the antipodal map. To convince ourselves
that 𝐹1 (𝑥, 𝑦,(𝑡) is an element in 𝕊
) , we can either just calculate its norm or observe that
1

cos(𝜋𝑡) − sin(𝜋𝑡)
the matrix is an element in 𝑂(2) for every 𝑡. Elements in 𝑂(2)
sin(𝜋𝑡) cos(𝜋𝑡)
preserve the scalar product and hence the norm of vectors in ℝ2 .
For an arbitrary odd 𝑘, we have 𝕊𝑘 ⊂ ℝ𝑘+1 and 𝑘 + 1 is even. Then we define a
smooth homotopy from the identity in 𝕊𝑘 to the antipodal map by

𝐹𝑘 ∶ 𝕊𝑘 × [0, 1] → 𝕊𝑘 ,
((𝑥1 , 𝑦1 ), … , (𝑥(𝑘+1)∕2 , 𝑦(𝑘+1)∕2 ), 𝑡) → (𝐹1 (𝑥1 , 𝑦1 , 𝑡), … , 𝐹1 ((𝑥(𝑘+1)∕2 , 𝑦(𝑘+1)∕2 , 𝑡)
392 A.7. Smooth Homotopy

Again, for every 𝑡, 𝐹𝑘 (−, −, 𝑡) ∶ ℝ𝑘+1 → ℝ𝑘+1 is an element in 𝑂(𝑘 + 1) and preserves
the norm on ℝ𝑘+1 .

Solution (Exercise 8.4) Let 𝑓 ∶ 𝕊1 → 𝑋 be a smooth map. Since 𝑋 is contractible,

there is a smooth homotopy 𝐹 from the identity on 𝑋 and a constant map {𝑥0 }. The
composition
𝕊1 × [0, 1] → 𝑋, (𝑥, 𝑡) → 𝐹 (𝑓 (𝑥), 𝑡)
defines a smooth homotopy from 𝑓 to the constant map 𝕊1 → {𝑥0 }.

Solution (Exercise 8.5) Given two points 𝑥, 𝑦 ∈ 𝑋, we define the relation 𝑥 ∼ 𝑦,

and say 𝑥 and 𝑦 are path-connected, by: 𝑥 ∼ 𝑦 if and only if there is a smooth path
𝛾 ∶ [0, 1] → 𝑋 with 𝛾(0) = 𝑥 and 𝛾(1) = 𝑦. We would like to show that ∼ is an
equivalence relation. Therefor, we are going to link it to the homotopy relation.
Let 𝑓 ∶ {𝑥} → 𝑋, 𝑓 (𝑥) = 𝑥, and 𝑔 ∶ {𝑥} → 𝑋, 𝑔(𝑥) = 𝑦. If 𝐹 ∶ {𝑥} × [0, 1] → 𝑋
is a smooth homotopy from 𝑓 to 𝑔, then 𝛾(𝑡) ∶= 𝐹 (𝑥, 𝑡) is a smooth path from 𝑥 to 𝑦.
Conversely, if 𝛾 is a smooth path from 𝑥 to 𝑦, then 𝐹 (𝑥, 𝑡) ∶= 𝛾(𝑡) is a smooth homotopy
from 𝑓 to 𝑔. Thus 𝑥 ∼ 𝑦 if and onlf if 𝑓 ∼ 𝑔.
Since homotopy is an equivalence relation, we see that path-connectedness is also an
equivalence relation. Recall that the equivalence class [𝑥] of a point 𝑥 ∈ 𝑋 is the set

[𝑥] = {𝑦 ∈ 𝑋 ∶ 𝑥 ∼ 𝑦}.

A crucial feature of equivalence relations is that equivalence classes are either equal or
disjoint, i.e. for any 𝑥 and 𝑦 in 𝑋 we have either [𝑥] = [𝑦] or [𝑥] ∩ [𝑦] = ∅. We are going
to use this fact in the following way: If we can show that every equivalence class [𝑥] is
an open subset of 𝑋, then we know that every [𝑥] is also a closed subset. For, 𝑋 ⧵ [𝑥]
is the union of all the other open classes and is therefore open itself (arbitrary unions of
open sets are open).
So, given an arbitrary point 𝑥0 ∈ 𝑋, we would like to show that [𝑥0 ] is open. Let
𝑥 ∈ 𝑋 be a point in [𝑥]. Since 𝑋 is a smooth manifold, there is a local parametrization
𝜙 ∶ 𝐵𝜖 (0) → 𝑈 with 𝜙(0) = 𝑥, where 𝑈 is open in 𝑋 and 𝐵𝜖 (0) is the open ball of radius
𝜖 in ℝdim 𝑋 . Given any 𝑦 ∈ 𝑈 , let 𝜙−1 (𝑦) be its preimage in 𝐵𝜖 (0). In 𝐵𝜖 (0), all points
are path-connected to 0. Hence there is a smooth path

𝛾 ∶ [0, 1] → 𝐵𝜖 (0), 𝑡 → 𝑡 ⋅ 𝜙−1 (𝑦)

with 𝛾(0) = 0 and 𝛾(1) = 𝜙−1 (𝑦). Since 𝜙 is a diffeomorphism, the composite 𝜙◦𝛾 is a
smooth path from 𝑥 to 𝑦 in 𝑋, i.e. 𝑥 ∼ 𝑦.
This shows that 𝑈 is contained in [𝑥0 ]. Thus [𝑥0 ] is an open subset in 𝑋, since every
point 𝑥 ∈ [𝑥0 ] has an open neighborhood in 𝑋 which is completely contained in [𝑥0 ].
Thus [𝑥0 ] is a nonempty, open and closed subset of 𝑋. Since 𝑋 is connected, this
implies [𝑥0 ] = 𝑋. Thus 𝑋 is path-connected.

Solution (Exercise 8.6) Let 𝑘 > 1 and 𝑓 ∶ 𝕊1 → 𝕊𝑘 be a smooth map. At any point
𝑥 ∈ 𝕊1 , the derivative 𝑑𝑓𝑥 ∶ 𝑇𝑥 (𝕊1 ) → 𝑇𝑓 (𝑥) (𝕊𝑘 ) is a linear map from a one-dimensional
Appendix A. Solutions to exercises 393
space to a 𝑘-dimensional space. Hence 𝑑𝑓𝑥 is not surjective for any 𝑥. Thus the only way
𝑝 ∈ 𝕊𝑘 can be a regular value is when it is not in the image of 𝑓 . But, by Sard’s Theorem,
almost every 𝑝 ∈ 𝕊𝑘 is a regular value. Hence there must be a point 𝑝 ∈ 𝕊𝑘 ⧵ 𝑓 (𝕊1 ).
Fix such a point 𝑝. Then, after possibly rotating our coordinate system, we can use 𝑝 as
a center for stereographic projection. This gives us a diffeomorphism 𝕊𝑘 ⧵ {𝑝} → ℝ𝑘 .
Since ℝ𝑘 is contractible, every smooth map into ℝ𝑘 is homotopic to a constant map.
Since the image of 𝑓 is contained in 𝕊𝑘 ⧵ {𝑝} ≅ ℝ𝑘 , 𝑓 is homootopic to a constant map.
Since 𝕊𝑘 is connected, this shows that 𝕊𝑘 is simply-connected.

Solution (Exercise 8.7) (a) The assumption that |𝑓 (𝑥) − 𝑔(𝑥)| < 2 implies that
𝑓 (𝑥) and 𝑔(𝑥) are never antipodal points. In particular, the straight line segment
from 𝑓 (𝑥) to 𝑔(𝑥) in ℝ𝑘+1 does not go through the origin for all 𝑥 ∈ 𝑋. Hence the
vector (1 − 𝑡)𝑓 (𝑥) + 𝑡𝑔(𝑥) is nonzero for all 𝑥 ∈ 𝑋 and all 𝑡 ∈ [0, 1]. Thus, we can
form the well-defined map

(1 − 𝑡)𝑓 (𝑥) + 𝑡𝑔(𝑥)

𝐻 ∶ 𝑋 × [0, 1] → 𝕊𝑘 , 𝑥 → .
|(1 − 𝑡)𝑓 (𝑥) + 𝑡𝑔(𝑥)|
Since 𝑓 and 𝑔 are continuous, 𝐻 is continuous and defines a homotopy from 𝑓 to
𝑔.

(b) By the previous point, there is a continuous homotopy 𝐻 from 𝑓 to 𝑔. Now it

suffices to compose 𝐻 with a smooth bump function to turn it into a smooth ho-
motopy.

Solution (Exercise 8.8) (a) If 𝑘 is odd, then 𝑘 + 1 is even and we can define the
map

𝑠 ∶ 𝕊𝑘 → ℝ𝑘+1 , (𝑥1 , … , 𝑥𝑘+1 ) → (−𝑥2 , 𝑥1 , −𝑥3 , 𝑥4 , … , −𝑥𝑘+1 , 𝑥𝑘 ).

This map can be extended to a linear map ℝ𝑘+1 → ℝ𝑘+1 and therefore 𝑠 is smooth.
For each 𝑥 ∈ 𝕊𝑘 , 𝑠(𝑥) is nonzero and satisfies 𝑥 ⟂ 𝑠(𝑥). Thus 𝑠(𝑥) is a tangent
vector at 𝑥, i.e. 𝑠(𝑥) ∈ 𝑇𝑥 (𝕊𝑘 ) ⧵ {0}. Hence

𝜎 ∶ 𝕊𝑘 → 𝑇 (𝕊𝑘 ), 𝜎(𝑥) ∶= (𝑥, 𝑠(𝑥))

is the desired non-vanishing vector field on 𝕊𝑘 .

(b) Given a vector field 𝜎 ∶ 𝕊𝑘 → 𝑇 (𝑆 𝑘 ) which has no zeros. Let 𝜎(𝑥) = (𝑥, 𝑠(𝑥)).
Since 𝑠(𝑥) ≠ 0 for every 𝑥 ∈ 𝕊𝑘 , we can define a new vector field by
𝑠(𝑥)
𝑥 → .
|𝑠(𝑥)|
By replacing 𝑠 with this new non-vanishing vector field, we can assume |𝑠(𝑥)| = 1.
Hence we can assume 𝑠(𝑥) ∈ 𝕊𝑘 and 𝑠(𝑥) ⋅ 𝑥 = 0 for every 𝑥 ∈ 𝕊𝑘 .
Now we define the map

𝐹 ∶ 𝕊𝑘 × [0, 1] → 𝕊𝑘 , (𝑥, 𝑡) → cos(𝜋𝑡)𝑥 + sin(𝜋𝑡)𝑠(𝑥).

394 A.7. Smooth Homotopy

We need to check that 𝐹 (𝑥, 𝑡) is in fact an element in 𝕊𝑘 for every 𝑥 ∈ 𝕊𝑘 :

𝐹 (𝑥, 𝑡) ⋅ 𝐹 (𝑥, 𝑡) = (cos(𝜋𝑡)𝑥 + sin(𝜋𝑡)𝑠(𝑥)) ⋅ (cos(𝜋𝑡)𝑥 + sin(𝜋𝑡)𝑠(𝑥))

= cos2 (𝜋𝑡)(𝑥 ⋅ 𝑥) + 2 cos(𝜋𝑡) sin(𝜋𝑡)(𝑥 ⋅ 𝑠(𝑥)) + sin2 (𝜋𝑡)(𝑠(𝑥) ⋅ 𝑠(𝑥))
= cos2 (𝜋𝑡) + sin2 (𝜋𝑡)
=1

where we use 𝑥 ⋅ 𝑥 = 1 = 𝑠(𝑥) ⋅ 𝑠(𝑥) and 𝑥 ⋅ 𝑠(𝑥) = 0. Thus 𝐹 (𝑥, 𝑡) is a vector of

norm 1 for every 𝑥 and every 𝑡. Moreover, 𝐹 is a smooth map with 𝐹 (𝑥, 0) = 𝑥
and 𝐹 (𝑥, 1) = −𝑥, i.e. 𝐹 is a smooth homotopy from the identity to the antipodal
map on 𝕊𝑘 .

(c) For 1 ≤ 𝑖 ≤ 𝑘 + 1, let 𝑟𝑖 be the reflection map on the 𝑖th coordinate:

𝑟𝑖 ∶ 𝕊𝑘 → 𝕊𝑘 , (𝑥1 , … , 𝑥𝑘+1 ) → (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑘+1 ).

Then the map 𝕊𝑘 × [0, 1] → 𝕊𝑘 defined by sending (𝑥1 , … , 𝑥𝑘+1 , 𝑡) to

(𝑥1 , … , 𝑥𝑖−1 , cos(𝜋𝑡)𝑥𝑖 − sin(𝜋𝑡)𝑥𝑖+1 , sin(𝜋𝑡)𝑥𝑖 + cos(𝜋𝑡)𝑥𝑖+1 , 𝑥𝑖+2 , … , 𝑥𝑘+1 )

is a homotopy from the identity on 𝕊𝑘 to the map 𝑟𝑖 ◦𝑟𝑖+1 ∶ 𝕊𝑘 → 𝕊𝑘 .

The antipodal map is equal to the composition of reflections 𝑟1 ◦𝑟2 ◦ ⋯ ◦𝑟𝑘+1 . Since
𝑘 is even 𝑟2 ◦ ⋯ ◦𝑟𝑘+1 is homotopic to the identity. Thus the antipodal map is
homotopic to the reflection 𝑟1 .
Appendix A. Solutions to exercises 395
A.8 Abstract Smooth Manifolds

Solution (Exercise 9.1) (a) A line in ℝ2 is determined by an equation of the form

𝑎𝑥+𝑏𝑦+𝑐 = 0 with fixed (𝑎, 𝑏, 𝑐) ∈ ℝ3 . Since an equation of the form 𝑎𝑥+𝑏𝑦+𝑐 =
0 with 𝑎 = 𝑏 = 0 does not define a line, we have to exclude triples of the form
(0, 0, 𝑐). Moreover, the equations 𝑎𝑥 + 𝑏𝑦 + 𝑐 = 0 and (𝜆𝑎)𝑥( 𝜆𝑏)𝑦 + (𝜆𝑐) = 0
with 𝜆 ≠ 0 determine the same line. Hence 𝑋 can be identified with the set of
equivalence classes

𝑋 = (ℝ3 ⧵ {(0, 0, 0), (0, 0, 1)})∕ ∼

where ∼ is the relation defined by

(𝑎, 𝑏, 𝑐) ∼ (𝜆𝑎, 𝜆𝑏, 𝜆𝑐) if there is a 𝜆 ≠ 0.

But this is the subspace of ℝP2 given by removing the point [0 ∶ 0 ∶ 1]. Since any
subspace consisting of just one point is closed in ℝP2 , we have shown that 𝑋 can
be identified with an open subset of ℝP2 :

𝑋 = ℝP2 ⧵ {[0 ∶ 0 ∶ 1]}.

(b) Every line in ℝ2 is determined by the point where it crosses the 𝑥-axis and a direc-
tion which can be expressed by an angle ∈ [0, 2𝜋]. Since we have not specified a
direction for the line, two angles which differ by adding 𝜋 determine the same line.
Any angle between 0 and 2𝜋 can be described by a point on the unit circle, where
the points 𝑠 and −𝑠 on 𝕊1 correspond to angles which differ by adding 𝜋. Hence
any line in ℝ2 is determined by a (𝑠, 𝑥) ∈ 𝕊1 × ℝ where 𝑠 is uniquely determined
up to multiplying with ±1.

Solution (Exercise 9.2) (a) We define

𝑉1 ∶= {[𝑧 ∶ 𝑤] ∈ ℂP1 ∶ 𝑧 ≠ 0} and 𝑉2 ∶= {[𝑧 ∶ 𝑤] ∈ ℂP1 ∶ 𝑤 ≠ 0}.
The preimage of 𝑉1 in ℂ2 is the open subset {(𝑧, 𝑤) ∈ ℂ2 ∶ 𝑧 ≠ 0} and preimage
of 𝑉2 in ℂ2 is the open subset {(𝑧, 𝑤) ∈ ℂ2 ∶ 𝑤 ≠ 0}. Hence 𝑉1 and 𝑉2 are open in
ℂP1 . Since either 𝑧 or 𝑤 must be nonzero for every point [𝑧 ∶ 𝑤] ∈ ℂP1 , {𝑉1 , 𝑉2 }
provides an open cover of ℂP1 .
We define maps 𝜙1 ∶ ℝ2 → 𝑉1 and 𝜙2 ∶ ℝ2 → 𝑉2
𝜙1 ∶ (𝑥, 𝑦) → [1 ∶ (𝑥 + 𝑖𝑦)] and 𝜙2 ∶ (𝑥, 𝑦) → [(𝑥 + 𝑖𝑦) ∶ 1].
The inverses are 𝜙−1
1
∶ 𝑉1 → ℝ2 and 𝜙−1
2
∶ 𝑉2 → ℝ2 defined by
1
𝜙−1
1
∶ [(𝑥1 + 𝑖𝑦1 ) ∶ (𝑥2 + 𝑖𝑦2 )] → (𝑥1 𝑥2 + 𝑦1 𝑦2 , 𝑥1 𝑦2 − 𝑥2 𝑦1 )
𝑥21 + 𝑦21
and
1
𝜙−1
2
∶ [(𝑥1 + 𝑖𝑦1 ) ∶ (𝑥2 + 𝑖𝑦2 )] → (𝑥1 𝑥2 + 𝑦1 𝑦2 , 𝑥2 𝑦1 − 𝑥1 𝑦2 )
𝑥22 + 𝑦22
396 A.8. Abstract Smooth Manifolds
where the right hand sides arise from calculating the quotients of complex num-
bers.
As in the proof of Theorem 9.9 we can check that these maps do not depend on
the chosen representatives. It is also easy to see that 𝜙 and 𝜙−1
𝑖 are mutual in-
verses which are both continuous. We check that the change-of-coordinate maps
are smooth: Both composites

𝜙1 𝜙−1
2
𝜙−1
1
(𝑉1 ∩ 𝑉2 ) ←←←←→ ← 𝜙−1
← 𝑉1 ∩ 𝑉2 ←←←←←←→ 2
(𝑉1 ∩ 𝑉2 )

and
𝜙2 𝜙−1
1
𝜙−1
2
(𝑉1 ∩ 𝑉2 ) ←←←←→ ← 𝜙−1
← 𝑉1 ∩ 𝑉2 ←←←←←←→ 2
(𝑉1 ∩ 𝑉2 )

are given by
1
(𝑥, 𝑦) → (𝑥, −𝑦)
𝑥2 + 𝑦2
and are therefore smooth maps.

(b) We copy the relation we used before and just add a condition to make sure that the
norms are respected. We define ∼𝑠 by

(𝑧0 , 𝑤0 ) ∼𝑠 (𝑧1 , 𝑤1 )
⇐⇒ 𝑧1 = 𝜆𝑧0 and 𝑤1 = 𝜆𝑤0 for some 𝜆 ∈ ℂ ⧵ {0} with |𝜆| = 1.

Then it is easy to check that 𝕊3 ∼𝑠 ≅ (ℂ2 ⧵ {(0, 0)})∕ ∼.

(c) For 𝑤 ≠ 0, we consider the points [𝑧 ∶ 𝑤] = [𝑧∕𝑤 ∶ 1] in ℂP1 as points in ℂ by

identifying them with 𝑧∕𝑤. This misses only one point in ℂP1 , the point [1 ∶ 0]
’at infinity’. Then we mimic the stereographic projection and define ℎ ∶ ℂP1 → 𝕊2
by { 1
(2𝑧∕𝑤, |𝑧∕𝑤|2 − 1) if 𝑤 ≠ 0
[𝑧 ∶ 𝑤] → |𝑧∕𝑤| +1
2

(0, 1) if 𝑤 = 0
where we consider 𝕊2 as a subset of ℂ × ℝ as before. We need to check that this
map has the desired properties:

∙ We need to check that ℎ([𝑧 ∶ 𝑤]) actually is a point on 𝕊2 . This is true for
ℎ([𝑧 ∶ 0]) = (0, 1) and for the other points we check

2𝑧∕𝑤 ̄ 𝑤̄
2𝑧∕ (|𝑧∕𝑤|2 − 1)2
|ℎ([𝑧 ∶ 𝑤])|2 = ⋅ +
|𝑧∕𝑤|2 + 1 |𝑧∕𝑤|2 + 1 (|𝑧∕𝑤|2 + 1)2
4|𝑧∕𝑤|2 + |𝑧∕𝑤|4 − 2|𝑧∕𝑤|2 + 1
=
(|𝑧∕𝑤|2 + 1)2
|𝑧∕𝑤|4 + 2|𝑧∕𝑤|2 + 1
=
(|𝑧∕𝑤|2 + 1)2
= 1.
Appendix A. Solutions to exercises 397
∙ We need to check that this map is well-defined, i.e., if 𝜆 ≠ 0 ∈ ℂ, we need
to check that ℎ sends [𝑧 ∶ 𝑤] and [𝜆𝑧 ∶ 𝜆𝑤] to the same point. This is true,
since we have |𝑧∕𝑤| = |(𝜆𝑧)∕(𝜆𝑤)|.
∙ For (𝑧, 0) ∈ 𝕊3 , we compute

ℎ(𝜑([𝑧 ∶ 0])) = (0, 1) = 𝜋(𝑧, 0)

where we use that |𝑧|2 = 1 as 𝑧 is a point on 𝕊3 . And for (𝑧, 𝑤) ∈ 𝕊3 with

𝑤 ≠ 0, we get

1
ℎ(𝜑([𝑧 ∶ 𝑤])) = (2𝑧∕𝑤, |𝑧∕𝑤|2 − 1)
|𝑧∕𝑤|2 + 1
|𝑤|2
= (2𝑧∕𝑤, |𝑧∕𝑤|2 − 1)
|𝑧| + |𝑤|
2 2

̄ |𝑧|2 − |𝑤|2 )
= (2𝑧∕𝑤 ⋅ (𝑤𝑤),
̄ |𝑧|2 − |𝑤|2 )
= (2𝑧 ⋅ 𝑤,
= 𝜋(𝑧, 𝑤)

where we used |𝑧|2 + |𝑤|2 = 1. Hence we have ℎ◦𝜑 = 𝜋.

(d) The fiber consists of all points (𝑧, 𝑤) in 𝕊3 with [𝑧 ∶ 𝑤] = [𝑧0 ∶ 𝑤0 ]. By the
definition of ℂP1 , that means (𝑧, 𝑤) is in the fiber if and only if there is a 𝜆 ∈ ℂ⧵{0}
such that (𝑧, 𝑤) = (𝜆𝑧0 , 𝜆𝑤0 ). The additional feature we need to remember is that
(𝑧, 𝑤) is in 𝕊3 just as we did for the relation ∼𝑠 . Hence 𝜆 needs to satisfy |𝜆| = 1.
Summarising, we have

(𝑧, 𝑤) ∈ 𝜑−1 ([𝑧0 ∶ 𝑤0 ])

⇐⇒ (𝑧, 𝑤) = (𝜆𝑧0 , 𝜆𝑤0 ) for some 𝜆 ∈ ℂ ⧵ {0} with |𝜆| = 1.

In other words, the points in the fiber 𝜑−1 ([𝑧0 ∶ 𝑤0 ]) are in one-to-one correspon-
dence to points 𝜆 on the circle 𝕊1 ⊂ ℂ. This shows again that the fiber of the Hopf
map at any point is diffeomorphic to 𝕊1 .

Solution (Exercise 9.3) (a) We write down local coordinate charts. For 𝑧 ∈ ℂ⧵{0},
let [𝑧] be its equivalence class in 𝐻 2 . We pick a point [𝑧0 ] ∈ 𝐻 2 . Choosing
𝜀 > 0 small enough, i.e., in our case 0 < 𝜀 < 𝜆∕2 is enough, the open ball
𝐵𝜀2 (𝑧0 ) ⊂ ℂ⧵{0} does not contain any point 𝑧 with [𝑧] = [𝑧0 ]. Hence the restriction
𝐵𝜀2 (𝑧0 ) → 𝐻 2 of the quotient map to 𝐵𝜀2 (𝑧0 ) is a homeomorphism onto its image
𝐵̄ 𝜀2 ([𝑧0 ]) ⊂ 𝐻 2 . The inverse is given by sending a point [𝑧] ∈ 𝐵̄ 𝜀2 ([𝑧0 ]) to the point
𝑧 ∈ 𝐵𝜀2 (𝑧0 ). Since 𝜀 is small enough, there is a unique such lift in 𝐵𝜀2 (𝑧0 ). This
defines a homeomorphism

𝜙 ∶ 𝐵𝜀2 (𝑧0 ) → 𝐵̄ 𝜀2 ([𝑧0 ])

whose inverse 𝜓 a local chart around [𝑧0 ] ∈ 𝐻 2 .

If 𝜓1 and 𝜓1 are charts around [𝑧1 ] and [𝑧2 ], respectively, then, since we chose 𝜀
398 A.8. Abstract Smooth Manifolds
small enough, the change of coordinate map

𝜓2 ◦𝜓1−1 ∶ 𝐵𝜀2 (𝑧1 ) ∩ 𝐵𝜀2 (𝑧2 ) → 𝐵𝜀2 (𝑧1 ) ∩ 𝐵𝜀2 (𝑧2 )

is just the identity. Since this is a smooth map, we have shown that 𝐻 2 is an abstract
smooth 2-manifold.

(b) We write [𝑧] for the image of 𝑧 ∈ 𝐴 in 𝐴∕ℤ. We claim that the map

ℎ ∶ 𝐴∕ℤ → 𝐻 2 , [𝑧]𝐴 → [𝑧]𝐻

where [𝑧]𝐴 and [𝑧]𝐻 denote the equivalence classes of 𝑧 ∈ ℂ in 𝐴∕ℤ and 𝐻 2 ,
respectively. We claim that ℎ is a homeomorphism:

∙ First we check that ℎ is injective. Assume 𝑤, 𝑧 ∈ 𝐴 with ℎ([𝑤]𝐴 ) = ℎ([𝑧)𝐴 ].

After possibly switching 𝑤 and 𝑧 we can assume |𝑤| ≤ |𝑧|. Then there is a
𝑘 ∈ ℤ, 𝑘 ≤ 0, such that 𝑤 = 𝜆𝑘 𝑧. If 𝑘 = 0, then 𝑤 = 𝑧, and if 𝑘 < 0, then
not both 𝑧 and 𝑤 can be in 𝐴, unless one of 𝑤 and 𝑧 has absolute value 1 and
the other one has absolute value 1∕𝜆. So let us say |𝑤| = 1 and |𝑧| = 1∕𝜆.
But in this case their images in the quotient 𝐴∕ℤ are the same. Hence ℎ is
injective.
∙ Next we check that 𝑓 is surjective. Let 𝑧 ∈ ℂ ⧵ {0}. Then there is a 𝑘 ∈ ℤ
such that 𝜆𝑘 𝑧 ∈ 𝐴. Hence ℎ([𝜆𝑘 𝑧]𝐴 ) = [𝑧]𝐻 ∈ 𝐻 2 , and ℎ is surjective.
∙ Since 𝐴 is a closed subset of ℂ, the map ℎ is continuous. Since 𝐴 is compact
and the quotient map 𝐴 → 𝐴∕ℤ is continuous, 𝐴∕ℤ is compact. Hence ℎ is
a continuous bijection with compact domain and Hausdorff codomain. This
implies, by general topology arguments, that ℎ is a homeomorphism.

This shows that 𝐻 2 is compact, since it is the image of a compact space under a
continuous map.

(c) For (𝑧, 𝑤) ∈ ℂ2 ⧵ {0}, let [𝑧, 𝑤] be its equivalence class in 𝐻 4 . We pick a point
[𝑧0 , 𝑤0 ] ∈ 𝐻 4 . Choosing 𝜀 > 0 small enough, the open ball 𝐵𝜀4 (𝑧0 ) ⊂ ℂ2 ⧵ {0}
does not contain any point (𝑧, 𝑤) with [𝑧, 𝑤] = [𝑧0 , 𝑤0 ]. Hence the restriction
𝐵𝜀4 (𝑧0 , 𝑤0 ) → 𝐻 4 to 𝐵𝜀2 (𝑧0 ) is a homeomorphism onto its image 𝐵̄ 𝜀4 ([𝑧0 , 𝑤0 ]) ⊂
𝐻 4 . The inverse is given by sending a point [𝑧, 𝑤] ∈ 𝐵̄ 𝜀4 ([𝑧0 , 𝑤0 ]) to the point
(𝑧, 𝑤) ∈ 𝐵𝜀4 (𝑧0 , 𝑤0 ). Since 𝜀 is small enough, there is a unique such lift in
𝐵𝜀4 (𝑧0 , 𝑤0 ). This defines a homeomorphism

𝜙 ∶ 𝐵𝜀4 (𝑧0 , 𝑤0 ) → 𝐵̄ 𝜀4 ([𝑧0 , 𝑤0 ])

whose inverse 𝜓 a local chart around [𝑧0 , 𝑤0 ] ∈ 𝐻 4 .

If 𝜓1 and 𝜓1 are charts around [𝑧1 , 𝑤1 ] and [𝑧2 , 𝑤2 ], respectively, then, by chosing
𝜀 small enough, the change of coordinate map

𝜓2 ◦𝜓1−1 ∶ 𝐵𝜀4 (𝑧1 , 𝑤1 ) ∩ 𝐵𝜀4 (𝑧2 , 𝑤2 ) → 𝐵𝜀4 (𝑧1 , 𝑤1 ) ∩ 𝐵𝜀4 (𝑧2 , 𝑤2 )

is just the identity. Since this is a smooth map, we have shown that 𝐻 4 is an abstract
smooth 4-manifold.
Appendix A. Solutions to exercises 399

(d) We consider 𝕊3 as a subset in ℂ2 ⧵ {0} and 𝕊1 as the quotient [1, 1∕𝜆]∕(1 ∼ 1∕𝜆),
i.e., the closed interval [1, 1∕𝜆] ⊂ ℝ where we identify the endpoints. We define
a map

𝑓 ∶ 𝕊3 × 𝕊1 = 𝕊3 × [1, 1∕𝜆]∕(1 ∼ 1∕𝜆) → 𝐻 4 , (𝑧, 𝑤, 𝑠) → [𝑠𝑧, 𝑠𝑤].

We need to check that this is well-defined and compute 𝑓 (𝑧, 𝑤, 1) = [𝑧, 𝑤] =

[1∕𝜆𝑧, 1∕𝜆𝑤] = 𝑓 (𝑧, 𝑤, 1∕𝜆).

∙ To show that 𝑓 is a homeomorphism, we recall that every point in ℂ2 ⧵ {0}

is determined by a direction, i.e., a point in 𝕊3 , and the distance from
the origin. Since we identify points 𝑝1 and 𝑝2 in 𝐻 4 if 𝑝2 = 𝜆𝑘 𝑝1 for
some 𝑘 ∈ ℤ, it suffices to specify the distance from the origin up to a
multiple of 𝜆𝑘 . This shows that 𝑓 is injective. But 𝑓 is also surjective,
since every point 𝑝 in ℂ2 ⧵ {0} determines a point in 𝕊3 , that is the point
(𝑧, 𝑤) where the line from the origin to 𝑝 meets 𝕊3 . Then there is a unique
𝑘 ∈ ℤ such that 1 ≤ 𝜆𝑘 |𝑝| < 1∕𝜆. This real number 𝜆𝑘 |𝑝| is the coordinate 𝑠.

By general topology, that facts that 𝑓 is continuous and has a compact

domain and a Hausdorff codomain, imply that 𝑓 is a homeomorphism.

∙ Our copy of 𝕊1 equals the quotient ℝ∕ℤ where we identify two real numbers
𝑠1 and 𝑠2 if 𝑠1 − 𝑠2 ∈ ℤ. This makes it easy to provide local charts. Around
any 𝑥 ∈ ℝ we can look at the open neighborhood (−𝜀 + 𝑥, 𝑥 + 𝜀) which
maps homeomorphically onto its image in 𝕊1 = ℝ∕ℤ under the quotient map
ℝ → ℝ∕ℤ. On points in ℂ2 ⧵ {0} and 𝕊3 , 𝑓 and its inverse are just scaling
by a real number, some integer power of 𝜆. So composition with local charts
results in a smooth map, since the local charts on 𝕊1 = ℝ∕ℤ are just identity
maps and the local charts on 𝕊3 are diffeomorphisms as we have seen many
times before. Hence 𝑓 is a diffeomorphism as it is a smooth homeomorphism
with a smooth inverse.

Solution (??) Recall the subsets 𝑉𝑖 ∶= {[𝑥0 ∶ … ∶ 𝑥𝑛 ] ∈ ℝP𝑛 ∶ 𝑥𝑖 ≠ 0} which

are open in ℝP𝑛 . The product spaces 𝑉𝑖 × 𝑉𝑗 are open in ℝP𝑚 × ℝP𝑛 and hence the
subsets 𝑉𝑖𝑗 = 𝐻(𝑚, 𝑛) ∩ (𝑉𝑖 × 𝑉𝑗 ) are open in 𝐻(𝑚, 𝑛). The union of all 𝑉𝑖 × 𝑉𝑗 for all
𝑖, 𝑗 covers ℝP𝑚 × ℝP𝑛 . Hence the union of all 𝑉𝑖𝑗 = 𝐻(𝑚, 𝑛) ∩ (𝑉𝑖 × 𝑉𝑗 ) for all 𝑖, 𝑗 covers
𝐻(𝑚, 𝑛). In fact, it suffices to consider all pairs (𝑖, 𝑗) with 𝑖 ≠ 𝑗, since if 𝑥𝑖 is the only
coordinate in [𝑥] with 𝑥𝑖 ≠ 0, then we must have 𝑦𝑖 = 0 in order to satisfy the condition
∑𝑚
𝑖=0 𝑥𝑖 𝑦𝑖 = 0. Thus we cannot have that 𝑥𝑖 and 𝑦𝑖 are the only non-zero coordinates in
[𝑥[ and [𝑦], respectively.
Now we define maps 𝜙𝑖𝑗 ∶ ℝ𝑚+𝑛−1 → 𝑉𝑖𝑗 by sending the (𝑚 + 𝑛 − 1)-tuple
((𝑥1 , … , 𝑥𝑚 ), (𝑦1 , … , 𝑦𝑛−1 )) to
( )
∑𝑚
[𝑥1 ∶ … ∶ 1 ∶ … ∶ 𝑥𝑚 ), [𝑦1 ∶ … ∶ − 𝑥𝑠 𝑦𝑠 ∶ … ∶ 1 ∶ … ∶ 𝑦𝑛−1 ) .
𝑠=1,𝑠≠𝑖

where the first 1 and the sum are at position 𝑖 + 1, and the second 1 is at position 𝑗 + 1.
Note that this actually yields an element in 𝐻(𝑚, 𝑛), and not just ℝP𝑚 × ℝP𝑛 , since the
400 A.8. Abstract Smooth Manifolds

defining condition 𝐻(𝑚, 𝑛) is satisfied. Their inverses 𝜙−1 𝑖𝑗 ∶ 𝑉𝑖𝑗 → ℝ

𝑚+𝑛−1 are given by

sending ([𝑥0 ∶ … ∶ 𝑥𝑚 ], [𝑦0 ∶ … ∶ 𝑦𝑛 ]) to

( )
1 1
(𝑥0 , … , 𝑥̂𝑖 , … , 𝑥𝑚 ), (𝑦0 , … , 𝑦̂𝑗 , … , 𝑦𝑛 ) .
𝑥𝑖 𝑦𝑗

We can check that the change of coordinate maps are smooth just as we did for real
∑
projective space, since the sum 𝑚𝑖=0 𝑥𝑖 𝑦𝑖 is a polynomial and hence smooth.
Appendix A. Solutions to exercises 401
A.9 Manifolds with Boundary

A.9.1 Manifolds with boundary

Solution (Exercise 10.1) Let 𝑈 ⊂ ℝ𝑘 and 𝑉 ⊂ ℍ𝑘 be open neighborhoods of 0.

Suppose there was a diffeomorphism 𝜃 ∶ 𝑈 → 𝑉 . We can assume that 0 is sent to a
boundary point of 𝑉 . In fact, we can assume that 𝜃(0) = 0. Otherwise we just another
pick point 𝑢 ∈ 𝑈 with 𝜃(𝑢) ∈ 𝜕𝑉 . Then 𝑑𝜃0 ∶ ℝ𝑘 → ℝ𝑘 is an isomorphism. By the
Inverse Function Theorem, there are subsets 𝑊1 and 𝑊2 in ℝ𝑘 containing 0 which are
both open in ℝ𝑘 such that 𝜃 maps 𝑊1 diffeomorphically onto 𝑊2 . Since 𝑊2 is open in
ℝ𝑘 and contained in the image of 𝜃, we get that 𝑉 must be open in ℝ𝑘 . But since 𝑉
contains 0, it satisfies 𝑉 ∩ 𝜕ℍ𝑘 ≠ ∅ and cannot be open in ℝ𝑘 .

Solution (Exercise 10.2) Let 𝑓 ∶ 𝑋 → 𝑌 be a diffeomorphism of manifolds with

boundary. Let 𝜙 ∶ 𝑈 → 𝑋 and 𝜓 ∶ 𝑉 → 𝑌 be local parametrizations, where 𝑈 and 𝑉
are open subsets of ℍ𝑘 (check that you know why the dimensions of 𝑋 and 𝑌 must be
equal). Let 𝜃 ∶ 𝑈 → 𝑉 be the induced map. By shrinking 𝑈 and 𝑉 if necessary, we can
assume that 𝜃 is a diffeomorphism with

𝑓 ◦𝜙 = 𝜓◦𝜃.

Boundary points of 𝑋 are those which are in the image 𝜙(𝜕𝑈 ) = 𝜙(𝑈 ∩ 𝜕ℍ𝑘 ). Sim-
ilarly, boundary points of 𝑋 are those which are in the image 𝜓(𝜕𝑉 ) = 𝜓(𝑉 ∩ 𝜕ℍ𝑘 ).
Hence we need to show 𝜃(𝜕𝑈 ) ⊂ 𝜕𝑉 , for then

𝑓 (𝜙(𝜕𝑈 )) = 𝜓(𝜃(𝜕𝑈 )) ⊂ 𝜓(𝜕𝑉 ) ⊂ 𝜕𝑌 .

The argument is again based on the Inverse Function Theorem. Suppose there is
a point 𝑢 ∈ 𝜕𝑈 which is mapped to an interior point 𝑣 = 𝜃(𝑢) in 𝑉 . Since 𝜃 is a
diffeomorphism, the derivative 𝑑(𝜃 −1 )𝑣 ∶ ℝ𝑘 → ℝ𝑘 of its inverse is an isomorphism.
But, since 𝑣 ∈ Int(𝑉 ), 𝑉 contains a neighborhood 𝑊 of 𝑣 that is open in ℝ𝑘 . Thus the
Inverse Function Theorem implies that 𝜃 −1 (𝑊 ) contains a neighborhood of 𝑢 that is open
in ℝ𝑘 . Hence 𝑢 is also an interior point in 𝑈 which contradicts the assumption 𝑢 ∈ 𝜕𝑈 .

Solution (Exercise 10.3) (a) The image of 𝐹 is the product 𝕊1 × [−1∕2, 1∕2]. This
is a product of a manifold without a boundary 𝕊1 and the manifold [−1∕2, 1∕2]
with boundary. The boundary of [−1∕2, 1∕2] constists of the disjoint union of
{−1∕2} and {1∕2}. By Lemma 10.7, we get

𝜕𝑋 = 𝜕(𝕊1 × [−1∕2, 1∕2]) = 𝕊1 × {−1∕2} ∪ 𝕊1 × {1∕2}.

402 A.9. Manifolds with Boundary

(b) We define the maps

𝜙+ ∶ (−𝜋, 𝜋) × [0, 3∕4) → 𝑌 ,

(𝑡, 𝑠) → ((1 + (−1∕2 + 𝑠) cos(𝑡∕2)) cos 𝑡, (1 + (−1∕2 + 𝑠) cos(𝑡∕2)) sin 𝑡, (−1∕2 + 𝑠) sin(𝑡∕2))
𝜙− ∶ (−𝜋, 𝜋) × [0, 3∕4) → 𝑌 ,
(𝑡, 𝑠) →((1 + (1∕2 − 𝑠) cos(𝑡∕2)) cos 𝑡, (1 + (1∕2 − 𝑠) cos(𝑡∕2)) sin 𝑡, (1∕2 − 𝑠) sin(𝑡∕2))
𝜓+ ∶ (0, 2𝜋) × [0, 3∕4) → 𝑌 ,
(𝑡, 𝑠) →((1 + (−1∕2 + 𝑠) cos(𝑡∕2)) cos 𝑡, (1 + (−1∕2 + 𝑠) cos(𝑡∕2)) sin 𝑡, (−1∕2 + 𝑠) sin(𝑡∕2))
𝜓− ∶ (0, 2𝜋) × [0, 3∕4) → 𝑌 ,
(𝑡, 𝑠) →((1 + (1∕2 − 𝑠) cos(𝑡∕2)) cos 𝑡, (1 + (1∕2 − 𝑠) cos(𝑡∕2)) sin 𝑡, (1∕2 − 𝑠) sin(𝑡∕2)).

As one can check by calculating the partial derivatives, each of these maps are
diffeomorphisms, and the union of their images covers 𝑌 . Hence we can use these
four maps as local parametrizations of 𝑌 .
The boundary of 𝑌 is then given by the union of the points

𝜕𝑌 = 𝜙+ ((−𝜋, 𝜋) × {0}) ∪ 𝜙− ((−𝜋, 𝜋) × {0}) ∪ 𝜓+ ((0, 2𝜋) × {0}) ∪ 𝜓− ((0, 2𝜋) × {0}).

Setting 𝑠 = 0 in the fomulae for those maps gives

1 1 1
𝜕𝑌 ={((1 − cos(𝑡∕2)) cos 𝑡, (1 − cos(𝑡∕2)) sin 𝑡, − sin(𝑡∕2)) ∈ ℝ3 ∶ 𝑡 ∈ ℝ}
2 2 2
1 1 1
∪ {((1 + cos(𝑡∕2)) cos 𝑡, (1 + cos(𝑡∕2)) sin 𝑡, sin(𝑡∕2)) ∈ ℝ3 ∶ 𝑡 ∈ ℝ}.
2 2 2
But, in fact, the two sets describing 𝜕𝑌 are the same which we see when we replace
𝑡 with 𝑡 + 2𝜋 and use some simple trigonometric identities:

⎧(1 − 1 cos( 𝑡+2𝜋 )) cos(𝑡 + 2𝜋) = (1 + 1 cos(𝑡∕2)) cos 𝑡,

⎪ 2
1
2
𝑡+2𝜋
2
1
⎨(1 − 2 cos( 2 )) sin(𝑡 + 2𝜋) = ((1 + 2 cos(𝑡∕2)) sin 𝑡
⎪− 1 sin( 𝑡+2𝜋 )) = 12 sin(𝑡∕2).
⎩ 2 2

Hence
1 1 1
𝜕𝑌 = {((1 + cos(𝑡∕2)) cos 𝑡, (1 + cos(𝑡∕2)) sin 𝑡, sin(𝑡∕2)) ∈ ℝ3 ∶ 𝑡 ∈ ℝ}.
2 2 2

Now we would like to show that 𝜕𝑌 is diffeomorphic to 𝕊1 . Remembering the

trigonometric identities

sin 𝑡 = 2 sin(𝑡∕2) cos(𝑡∕2) and cos 𝑡 = cos2 (𝑡∕2) − sin2 (𝑡∕2)

we see that the map

1 1 1
𝜑 ∶ ℝ2 → ℝ3 , (𝑥, 𝑦) → ((1 + 𝑥)(𝑥2 − 𝑦2 ), (1 + 𝑥)2𝑥𝑦, 𝑦)
2 2 2
restricts to a bijection from 𝕊1 onto 𝜕𝑌 (for injectivity, note that the last coordinate
determines 𝑦 uniquely, then the circle equation determines 𝑥 up to sign, and the
first and/or second coordinate determine the sign of 𝑥).
Appendix A. Solutions to exercises 403
It remains to check that 𝜑|𝕊1 is a local diffeomorphism. Since 𝜑|𝕊1 is a bijection
onto its image, this will show that it is a diffeomorphism.
First we observe that 𝜑 is smooth, since the three functions in each coordinate are
just polynomials and hence smooth. To see that 𝜑|𝕊1 is a local diffeomorphism,
we use the maps

𝜙 ∶ 𝑈 → 𝕊1 , 𝑡 → (cos(𝑡∕2), sin(𝑡∕2))

and
1 1 1
𝜓 ∶ 𝑈 → 𝜕𝑌 , 𝑡 → ((1 + cos(𝑡∕2)) cos 𝑡, (1 + cos(𝑡∕2)) sin 𝑡, sin(𝑡∕2))
2 2 2
where 𝑈 ⊂ ℝ is some sufficiently small open subset. Then 𝜙 and 𝜓 serve as local
parametrizations of 𝕊1 and 𝜕𝑌 , respectively, for suitable choices of 𝑈 . But the
induced map 𝜃 ∶ 𝑈 → 𝑈 which arises as the composite 𝜓 −1 ◦𝜑|𝕊1 ◦𝜙 is just the
identity 𝑡 → 𝑡. Hence 𝜑|𝕊1 is a local diffeomorphism.

Solution (Exercise 10.4) (a) We proceed as before when we showed that tangent
spaces are well-defined.
Let 𝜓 ∶ 𝑉 → 𝑋 be another local parametrization around 𝑥 with 𝜓(0) = 𝑥, where 𝑉
is an open subset of ℍ𝑘 . By shrinking both 𝑈 and 𝑉 , we can assume 𝜙(𝑈 ) = 𝜓(𝑉 )
(replace 𝑈 by 𝜙−1 (𝜙(𝑈 ) ∩ 𝜓(𝑉 )) ⊂ 𝑈 and 𝑉 by 𝜓 −1 (𝜙(𝑈 ) ∩ 𝜓(𝑉 )) ⊂ 𝑉 ).
Then the map
𝜃 ∶= 𝜓 −1 ◦𝜙 ∶ 𝑈 → 𝑉
is a diffeomorphism (its the composite of two diffeomorphisms). By definition of
𝜃, we have 𝜙 = 𝜓◦𝜃. Differentiating yields

𝑑𝜙0 = 𝑑𝜓0 ◦𝑑𝜃0

(where we have used the chain rule). This implies that the image of 𝑑𝜙0 is con-
tained in the image of 𝑑𝜓0 :

𝑑𝜙0 (ℝ𝑘 ) ⊆ 𝑑𝜓0 (ℝ𝑘 ) in ℝ𝑁 .

By switching the roles of 𝜙 and 𝜓 in the argument, we also get:

𝑑𝜓0 (ℝ𝑘 ) ⊆ 𝑑𝜙0 (ℝ𝑘 ) in ℝ𝑁 .

Hence 𝑇𝑥 (𝑋) = 𝑑𝜙0 (ℝ𝑘 ) = 𝑑𝜓0 (ℝ𝑘 ) is well-defined in ℝ𝑁 .

In particular, the image of the upper halfplane ℍ𝑘 ⊂ ℝ𝑘 is well-defined:

𝐻𝑥 (𝑋) = 𝑑𝜙0 (ℍ𝑘 ) = 𝑑𝜓0 (ℍ𝑘 ) in ℝ𝑁 .

(b) The codimension of 𝑇𝑥 (𝜕𝑋) in 𝑇𝑥 (𝑋) is one. Thus the orthogonal complement
of 𝑇𝑥 (𝜕𝑋) is one-dimensional and is spanned by one vector. By definition of 𝜕𝑋
as the image of the points in 𝜕𝐻𝐻 𝑘 under local parametrizations, we know that
𝑑𝜙0 (𝑒𝑘 ) spans the complement of 𝑇𝑥 (𝜕𝑋) in 𝑇𝑥 (𝑋), since 𝑑𝜙0 is an isomorphism
404 A.9. Manifolds with Boundary

and 𝑒𝑘 = (0, … , 0, 1) is nonzero and not contained in 𝑑𝜙0 (ℍ𝑘 ). We also know by
the definition of 𝐻𝑥 (𝑋) that 𝑑𝜙0 (𝑒𝑘 ) ∈ 𝐻𝑥 (𝑋), and therefore 𝑑𝜙0 (−𝑒𝑘 ) ∉ 𝐻𝑥 (𝑋).
But we do not know whether 𝑑𝜙0 (−𝑒𝑘 ) is orthogonal to 𝑇𝑥 (𝜕𝑋) in 𝑇𝑥 (𝑋). To make
𝑑𝜙0 (−𝑒𝑘 ) into a vector which is orthogonal to 𝑇𝑥 (𝜕𝑋), we apply the Gram-Schmidt
process. It produces a unit vector which is orthogonal to 𝑇𝑥 (𝜕𝑋). We denote this
vector 𝑛(𝑥), this is the outward unit normal vector to 𝜕𝑋. Note that −𝑛(𝑥) is a
unit vector contained in 𝐻𝑥 (𝑋) and orthogonal to 𝑇𝑥 (𝜕𝑋), this is the inward unit
normal vector to 𝜕𝑋.

(c) From what we have learned in the previous point, we can construct 𝑛(𝑥) by ap-
plying the Gram-Schmidt orthonormalization process to 𝑑𝜙0 (−𝑒𝑘 ). This process
depends smoothly on the coefficients in the matrix representing 𝑑𝜙0 . Since the
derivative 𝑑𝜙𝑢 depends smoothly on 𝑢, 𝑑𝜙𝑢 (−𝑒𝑘 ) depends smoothly on 𝑢. By the
independence of the choice of local parametrization, we see that 𝑛(𝑦) = 𝑑𝜙𝑢 (−𝑒𝑘 )
for all 𝑦 ∈ 𝜙(𝜕𝑈 ) which is an open neighborhood of 𝑥 in 𝜕𝑋, where 𝜙(𝑢) = 𝑦.
Thus, in total we see that 𝑛(𝑥) depends smoothly on 𝑥 in 𝜕𝑋.

Solution (Exercise 10.5) (a) The boundary of 𝑋 is 𝜕𝑋 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥 = −1}.

( )
The derivative of 𝑓 is given by the 1 × 2-matrix 𝑑𝑓(𝑥,𝑦) = 2𝑥 2𝑦 . Hence 𝑑𝑓(𝑥,𝑦)
is a surjective linear map for all (𝑥, 𝑦) ≠ (0, 0). Since 𝑓 (0, 0) = 0 ≠ 1, 𝑑𝑓(𝑥,𝑦) is
surjective for all (𝑥, 𝑦) ∈ 𝑓 −1 (1) and 1 is a regular value of 𝑓 .
The restriction of 𝑓 to the boundary of 𝑋 is

𝜕𝑓 ∶ 𝜕𝑋 → 𝑌 , (−1, 𝑦) → 1 + 𝑦2 .

Hence the derivative of 𝜕𝑓 is given by the 1 × 1-matrix (𝜕𝑓 )(−1,𝑦) = 2𝑦. This is
a linear map which is surjective if and only if 𝑦 ≠ 0. Since (−1, 0) ∈ 𝜕𝑋 and
𝜕𝑓 (−1, 0) = 1, we see that 1 is not a regular value of 𝜕𝑓 .

(b) The preimage 𝑓 −1 (1) is just the unit sphere 𝕊1 . Hence the boundary 𝜕(𝑓 −1 (1)) is
empty. However,

𝑓 −1 (1)∩𝜕𝑋 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 +𝑦2 = 1}∩{(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥 = −1} = {(−1, 0)} ≠ ∅.

In particular, 𝜕(𝑓 −1 (1)) ≠ 𝑓 −1 (1) ∩ 𝜕𝑋.

This is not a contradiction to the Preimage Theorem for manifolds with boundary,
since the conclusion of the theorem required that 1 was a regular of both 𝑓 and 𝜕𝑓 .
But we showed in the first part that 1 is not a regular value of 𝜕𝑓 .

A.9.2 Brouwer Fixed Point Theorem

Solution (Exercise 10.6) If 𝐴 is not invertible, then 0 is an eigenvalue, and we are

done. So assume 𝐴 is nonsingular. For any vector 𝑣 ∈ 𝕊𝑛−1 ⊂ ℝ𝑛 , the vector 𝐴𝑣∕|𝐴𝑣|
has norm one and lies on 𝕊𝑛−1 . (Note that this map is not defined for 𝑣 = 0 and we cannot
continuously extend it on 0. Hence we cannot just consider this as a map 𝔹𝑛 → 𝔹𝑛 !)
Appendix A. Solutions to exercises 405

Let 𝑔 ∶ 𝕊𝑛−1 → 𝕊𝑛−1 be the map 𝑣 → 𝐴𝑣∕|𝐴𝑣|. Now we use the assumption on 𝐴: if

𝑣 ∈ 𝑄 = {(𝑥𝑙 , … , 𝑥𝑛 ) ∈ 𝕊𝑛−1 ∶ all 𝑥𝑖 ≥ 0}

then 𝐴𝑣 has only nonnegative entries, since the entries in 𝑣 are all nonnegative and all
the entries in 𝐴 are by assumption nonnegative. Since |𝐴𝑣| > 0 is nonnegative as well,
we know that 𝑔(𝑣) is an element in 𝑄. Thus we can restrict 𝑔 to a map 𝑔 ∶ 𝑄 → 𝑄.
≅
Now we can compose with a homeomorphism 𝜑 ∶ 𝑄 ←←←→
← 𝔹𝑛−1 to get a continuous
map
𝜑 𝑔 𝜑−1
𝑓 ∶ 𝔹𝑛−1 ←←←→
← 𝑄 ←←→ ← 𝔹𝑛−1 .
← 𝑄 ←←←←←←→
By the Brouwer Fixed Point Theorem for continuous maps, 𝑓 must have a fixed point
𝑦 ∈ 𝔹𝑛−1 with 𝑓 (𝑦) = 𝑦. Hence the image of 𝑤 ∶= 𝜑−1 (𝑦) is a vector in 𝕊𝑛−1 ⊂ ℝ𝑛 with

𝐴𝑤∕|𝐴𝑤| = 𝑤, i.e., 𝐴𝑤 = |𝐴𝑤| ⋅ 𝑤.

Since 𝑤 is nonzero being a point on 𝕊𝑛−1 , 𝑤 is an eigenvector with real nonnegative

eigenvalue |𝐴𝑤|.

Solution (Exercise 10.6) (a) Assuming that 𝑋 is simply-connected, 𝑓 can be ex-

𝑝
tended to a map 𝐹 ∶ → 𝑋. Let 𝑟 ∶ 𝑋 → 𝕊1 be the map defined by 𝑝 → |𝑝|
𝔹2 . We
can compose these maps to get
𝐹 𝑟
𝑔 ∶ 𝔹2 ←←←→ ←← 𝕊1 → 𝔹2 .
← 𝑋 ←→

(b) The image of 𝑔 is contained in 𝕊1 by definition of 𝑔 as a composition with the

inclusion 𝕊1 → 𝔹2 . Hence the only possible fixed points of 𝑔 are contained in 𝕊1 .
But 𝑔 sends points in 𝕊1 ⊂ 𝔹2 , to their antipodal points. Hence 𝑔 does not have
any fixed points.

(c) If 𝑋 = ℝ2 ⧵ {(0, 0)} was simply-connected we could construct a continuous map

𝑔 ∶ 𝔹2 → 𝔹2 without a fixed point. This contradicts Brouwer’s fixed point theo-
rem. Hence 𝑋 = ℝ2 ⧵ {(0, 0)} cannot be simply-connected.

(d) Assume there was such a homotopy 𝐹 ∶ 𝕊1 × [0, 1] → 𝑋 = ℝ2 ⧵ {(0, 0)}. Then we
get a homotopy 𝐻 between the identity map of ℝ2 ⧵ {(0, 0)} and a constant map
as follows: Let 𝐺 ∶ 𝑋 × [0, 1] → 𝑋 be a homotopy between the identity of 𝑋, i.e.,
𝐺(𝑝, 0) = 𝑝, and the map 𝑟, i.e., 𝐺(𝑝, 1) = 𝑝∕|𝑝|. Then we define a homotopy by

𝐻 ∶ ℝ2 ⧵ {(0, 0)} × [0, 1] → ℝ2 ⧵ {(0, 0)}

{
𝐺(𝑝, 2𝑡) for 0 ≤ 𝑡 ≤ 1∕2
(𝑝, 𝑡) →
𝐹 (𝑟(𝑝), 1 − 2𝑡) for 1∕2 ≤ 𝑡 ≤ 1.

This would imply that ℝ2 ⧵ {(0, 0)} is contractible. This contradicts our result that
ℝ2 ⧵ {(0, 0)} is not simply-connected.
406 A.10. Brouwer Degree mod 2 and Borsuk–Ulam Theorem
A.10 Brouwer Degree mod 2 and Borsuk–Ulam Theorem

A.10.1 Degree modulo 2

Solution (Exercise 11.1) Let 𝑧 be a regular value for 𝑔. If 𝑧 is not in the image of 𝑔,
then deg2 (𝑔) and deg2 (𝑔◦𝑓 ) both vanish and the claim is true. So let us assume that 𝑧 is
in the image of 𝑔. We showed in a previous exercise that 𝑧 then is a regular value for 𝑔◦𝑓
if and only if every 𝑦 ∈ 𝑔 −1 (𝑧) is a regular value for 𝑓 . Hence we can use 𝑧 to compute
deg2 (𝑔◦𝑓 ) if and only if we can use 𝑦 ∈ 𝑔 −1 (𝑧) to compute deg2 (𝑓 ). If 𝑧 is not in the
image of 𝑔◦𝑓 , then 𝑦 is not in the image of 𝑓 . In this case both deg2 (𝑔◦𝑓 ) and deg2 (𝑓 )
vanish and the claim is true. So assume there are points 𝑥 ∈ 𝑋 with 𝑔(𝑓 (𝑥)) = 𝑧. Then
it remains to apply the definition of deg2 . By our assumptions, we have

deg2 (𝑔◦𝑓 ) = #(𝑔◦𝑓 )−1 (𝑧)

= #𝑓 −1 (𝑔 −1 (𝑧))
= (#𝑔 −1 (𝑧)) ⋅ (#𝑓 −1 (𝑦))
= deg2 (𝑔) ⋅ deg2 (𝑓 ) mod 2

where we use that the number of points in the fiber 𝑓 −1 (𝑦) is the same for every 𝑦 ∈ 𝑔 −1 (𝑧)
which is a regular value for 𝑓 .

Solution (Exercise 11.2) (a) Assume 𝑓 is not surjective. Then there is a 𝑦 ∈ 𝑌

which is not in the image of 𝑓 . Hence 𝑦 is a regular value for 𝑓 and we can use it
to compute the degree of 𝑓 as

deg2 (𝑓 ) = #𝑓 −1 (𝑦) = 0.

This contradicts the assumption that deg2 (𝑓 ) ≠ 0. Hence 𝑓 must be surjective.

Alternative argument: Since deg2 (𝑓 ) ≠ 0, we must have #𝑓 −1 (𝑦) ≠ 0 mod 2 for

some 𝑦 ∈ 𝑌 . But, since 𝑌 is connected, the function

#𝑓 −1 (−) ∶ 𝑌 → ℤ∕2, 𝑦 → #𝑓 −1 (𝑦) mod 2

is constant. Thus we must have #𝑓 −1 (𝑦) ≠ 0 mod 2 for all 𝑦 ∈ 𝑌 . Hence

𝑓 −1 (𝑦) ≠ ∅ for all 𝑦 ∈ 𝑌 and 𝑓 is surjective.

(b) Let us assume deg2 (𝑓 ) ≠ 0 and derive a contradiction. By the previous point, if
deg2 (𝑓 ) ≠ 0, then 𝑓 is surjective. But that means 𝑌 = 𝑓 (𝑋). Since 𝑓 is continuous
and 𝑋 is compact, the image of 𝑋 under 𝑓 is compact. Hence 𝑌 would be compact
as the continuous image of a compact space. This contradicts the assumption.
Hence we must have deg2 (𝑓 ) = 0.

(c) Let 𝑓 ∶ 𝕊1 → 𝕊1 be a smooth map without fixed points. We define the map

𝐺(𝑥, 𝑡) ∶ 𝕊1 × [0, 1] → ℝ2 , (𝑥, 𝑡) → 𝑓 (𝑥)(1 − 𝑡) − 𝑡𝑥.

Appendix A. Solutions to exercises 407
We would like to turn 𝐺 into a homotopy between 𝑓 and 𝛼. Hence we need to
manipulate 𝐺 such that its image is contained in 𝕊1 ⊂ ℝ2 . We can arrange this if
𝐺(𝑥,𝑡)
𝐺(𝑥, 𝑡) ≠ 0. For then |𝐺(𝑥,𝑡)| is in 𝕊1 . Hence we need to check 𝐺(𝑥, 𝑡) ≠ 0 for all
(𝑥, 𝑡) ∈ 𝕊1 × [0, 1].
For a fixed 𝑥 and varying 𝑡, 𝑓 (𝑥)(1 − 𝑡) − 𝑡𝑥 describes the line segment in ℝ2
between the two points 𝑓 (𝑥) and −𝑥 on 𝕊1 . The only way, this line segment can
pass 0 ∈ ℝ2 , is when 𝑓 (𝑥) = 𝑥 is the antipodal point to −𝑥. But, by the assumption
on 𝑓 , 𝑓 (𝑥) ≠ 𝑥 for all 𝑥 ∈ 𝕊1 .
Thus the smooth map

𝑓 (𝑥)(1 − 𝑡) − 𝑡𝑥
𝐹 (𝑥, 𝑡) ∶ 𝕊1 × [0, 1] → 𝕊1 , (𝑥, 𝑡) →
|𝑓 (𝑥)(1 − 𝑡) − 𝑡𝑥|
is a homotopy between 𝑓 and 𝛼.
Since 𝛼 −1 (𝑥) = −𝑥 for all 𝑥 ∈ 𝕊1 , there is exactly one preimage point for each 𝑥.
Hence deg2 (𝛼) = 1. Since 𝑓 and 𝛼 are homotopic, the invariance of deg2 under
homotopy implies deg2 (𝑓 ) = 1. By the first point, deg2 (𝑓 ) = 1 implies that 𝑓 is
surjective.

Solution (Exercise 11.3) We define

𝑓 ∶ ℂ → ℂ, 𝑧 → 𝑧7 + cos(|𝑧|2 )(1 + 93𝑧4 ).

We can define a homotopy from 𝑓0 (𝑧) = 𝑧7 to 𝑓1 (𝑧) = 𝑓 (𝑧) by

𝑓𝑡 (𝑧) = 𝑡𝑓 (𝑧) + (1 − 𝑡)𝑧7 = 𝑧7 + 𝑡 cos(|𝑧|2 )(1 + 93𝑧4 ).

Since |𝑧|7 dominates the absolute value of 𝑓𝑡 (𝑧) for all 𝑡 ∈ [0, 1], or in other words, since
the second summand in the parantheses goes to 0 when 𝑧 → ∞ in

𝑓𝑡 (𝑧) cos(|𝑧|2 )
= 1 + 𝑡 (1 + 93𝑧4 ),
𝑧7 𝑧7
𝑓𝑡 (𝑧) has no zero on the boundary of a closed ball 𝑊 ⊂ ℂ of large enough radius. Thus
the homotopy

𝑓𝑡 (𝑧)
∶ 𝜕𝑊 → 𝕊1
|𝑓𝑡 (𝑧)|

is defined for all 𝑡. By the homotopy invariance of deg2 , we have

( ) ( )
𝑓 (𝑧) 𝑓0 (𝑧)
deg2 = deg2 = deg2 (𝑧7 ) = 1 mod 2.
|𝑓 (𝑧)| |𝑓0 (𝑧)|

Thus, by the Boundary Theorem 11.8 for deg2 , 𝑓 (𝑧) must have a zero inside 𝑊 .
408 A.10. Brouwer Degree mod 2 and Borsuk–Ulam Theorem

Solution (Exercise 11.4) We define a homotopy from 𝑝0 (𝑧) = 𝑧𝑚 to 𝑝1 (𝑧) = 𝑝(𝑧) by

𝑝𝑡 (𝑧) = 𝑡𝑝(𝑧) + (1 − 𝑡)𝑧𝑚 = 𝑧𝑚 + 𝑡(𝑎1 𝑧𝑚−1 + ⋯ + 𝑎𝑚 ).

For 𝑧 ≠ 0, we consider
𝑝𝑡 (𝑧) (𝑎 𝑎𝑚 )
1
= 1 + 𝑡 ⋅ + ⋯ + .
𝑧𝑚 𝑧 𝑧𝑚
𝑎 𝑎
As 𝑧 → ∞ moves towards infinity, the term 𝑧1 + ⋯ + 𝑧𝑚𝑚 → 0 moves towards zero.
Hence, if 𝑊 is a closed ball around the origin in ℂ with sufficiently large radius, none
of the 𝑝𝑡 has a zero on 𝜕𝑊 .
Thus the homotopy
𝑝𝑡
∶ 𝜕𝑊 → 𝕊1
|𝑝𝑡 |

is defined for all 𝑡 ∈ [0, 1]. This implies

( ) ( )
𝑝 𝑝0
deg2 = deg2 .
|𝑝| |𝑝0 |
𝑧𝑚
Since 𝑝0 (𝑧) = 𝑧𝑚 and #{𝑧 ∈ 𝜕𝑊 ∶ |𝑧|𝑚
= 1} = 𝑚, we have
( )
𝑝0
deg2 = 𝑚 mod 2.
|𝑝0 |
( )
𝑝
Hence, if 𝑚 is odd, then deg2 |𝑝| ≠ 0, and there must be 𝑤 ∈ 𝑊 with 𝑝(𝑤) = 0
by Theorem 11.9.

A.10.2 Borsuk–Ulam Theorem

Solution (Exercise 11.5) Assume that 𝑓1 , … , 𝑓𝑘 did not have a common zero. Then
we can form the smooth odd map

𝑓 ∶= (𝑓1 , … , 𝑓𝑘 , 0) ∶ 𝕊𝑘 → ℝ𝑘+1 ⧵ {0}.

Now we can apply Theorem 11.18 to 𝑓 and 𝐿 being the 𝑥𝑘+1 -axis. Hence 𝑓 intersects
𝐿 at least once. But 𝑥 with 𝑓 (𝑥) ∈ 𝐿 is a common zero of the 𝑓1 , … , 𝑓𝑘 . Hence the
𝑓1 , … , 𝑓𝑘 must have had a common zero after all.

Solution (Exercise 11.6) We define functions 𝑓1 , … , 𝑓𝑘 on 𝕊𝑘 by

𝑓𝑖 (𝑥) ∶= 𝑔𝑖 (𝑥) − 𝑔𝑖 (−𝑥).

Then each 𝑓𝑖 is smooth and odd. The functions 𝑓1 , … , 𝑓𝑘 satisfy the assumption of the
previous exercise. Hence there is a common zero of the 𝑓1 , … , 𝑓𝑘 which is the desired
Appendix A. Solutions to exercises 409

point 𝑝 ∈ 𝕊𝑘 .

Solution (Exercise 11.7) Since the 𝑝𝑖 ’s are all homogeneous of odd order, they satisfy

𝑝𝑖 (−𝑥) = (−1)𝑚𝑖 𝑝𝑖 (𝑥) = −𝑝(𝑥𝑖 ).

Moreover, for any 𝑥 ∈ ℝ𝑛+1 ⧵ {0}, we can consider the associated map
( )
𝑛 𝑥
𝑞𝑖 ∶ 𝕊 → ℝ, 𝑥 → 𝑝𝑖 .
|𝑥|

Since 𝑛 + 1 ≥ 2, this is a smooth map. Hence we have 𝑛 smooth real-valued functions

𝑞1 , … , 𝑞𝑛 on 𝕊𝑛 , all satisfying the symmetry condition

𝑞𝑖 (−𝑥) = 𝑝𝑖 (−𝑥∕|𝑥|) = −𝑝𝑖 (𝑥∕|𝑥|) = −𝑞𝑖 (𝑥).

By the first exercise, we know that these 𝑛 maps 𝑞1 , … , 𝑞𝑛 must have a common zero
on 𝕊𝑛 , say 𝑥0 ∈ 𝕊𝑛 .
Since the 𝑝𝑖 are homogeneous, 𝑞𝑖 (𝑥0 ) = 𝑝𝑖 (𝑥0 ) = 0 implies

𝑝1 (𝑥) = ⋯ = 𝑝𝑛 (𝑥) = 0 for all 𝑥 = 𝜆𝑥0 for some 𝜆 ∈ ℝ.

Hence the line spanned by 𝑥0 in ℝ𝑛+1 is the desired line on which all 𝑝𝑖 vanish simulta-
neously.

Solution (Exercise 11.8) Assume such a map 𝑓 existed. Then we could define the
continuous map
√
𝑔 ∶ 𝔹2 = {(𝑥, 𝑦) ∈ ℝ2 ∶ 𝑥2 + 𝑦2 ≤ 1} → 𝕊1 , 𝑔(𝑥, 𝑦) ∶= 𝑓 (𝑥, 𝑦, 1 − 𝑥2 − 𝑦2 ).

Note that 𝑔 is continuous, since 𝑔 is the composite of 𝑓 with the inverse of the projection
the projection from the upper hemisphere of the sphere to 𝔹2 . Then, by the assumption
on 𝑓 , we have

𝑔(−𝑥, −𝑦) = 𝑓 (−𝑥, −𝑦, 0) = −𝑓 (𝑥, 𝑦, 0) = −𝑔(𝑥, 𝑦)

for any (𝑥, 𝑦) ∈ 𝕊1 = 𝜕𝔹2 , i.e., for (𝑥, 𝑦) such that 1 − 𝑥2 − 𝑦2 = 0. In particular,
𝑔|𝜕𝔹2 ∶ 𝕊1 → 𝕊1 satisfies the assumptions of the Borsuk–Ulam Theorem. Hence, by the
Borsuk–Ulam Theorem, deg2 (𝑔) = 1. But 𝑔|𝜕𝔹2 can be extended to a continuous map
on all of 𝔹2 . Thus, by the Boundary Theorem for degrees, deg2 (𝑔) = 0. Hence the
assumption that 𝑓 exists, leads to a contradiction. Thus 𝑓 cannot exist.

Solution (Exercise 11.9) Assume that there is no such point 𝑝. Then we can consider
the map 𝑔 ∶ 𝕊2 → 𝕊1 given by
𝑓 (𝑝) − 𝑓 (−𝑝)
𝑔(𝑝) = .
|𝑓 (𝑝) − 𝑓 (−𝑝)|
By the assumption, 𝑔 is smooth and satisfies 𝑔(−𝑝) = −𝑔(𝑝) for all 𝑝 ∈ 𝕊2 . This contra-
410 A.10. Brouwer Degree mod 2 and Borsuk–Ulam Theorem
dicts the result of the previous exercise.

Solution (Exercise 11.10) Let 𝑈 ⊂ ℝ2 and 𝑉 ⊂ ℝ𝑛 be open subsets. After possibly

translating we can assume that 𝑉 contains the origin. Moreover, since 𝑉 is open, there
is an 𝜀 > 0 such that 𝔹𝑛𝜀 (0) ⊂ 𝑉 , where 𝔹𝑛𝜀 (0) denotes the closed ball of radius 𝜀 centred
at the origin. We can then consider the subset 𝕊2𝜀 ⊂ 𝔹3𝜀 (0) ⊂ 𝔹𝑛𝜀 (0) ⊂ 𝑉 . Assume
now there was a homeomorphism 𝜑 ∶ 𝑉 → 𝑈 . Then 𝜑|𝕊2 is still a homeomorphism.
𝜀
Since scaling defines a homeomorphism 𝕊2 → 𝕊2𝜀 , composition with 𝜑|𝕊2 would yield
𝜀
a continuous map 𝕊2 → ℝ2 with 𝑓 (𝑝) ≠ 𝑓 (−𝑝) for all 𝑝 ∈ 𝕊2 . This contradicts the
assumed continuous version of the previous exercise.
Appendix A. Solutions to exercises 411
A.11 Transversality

A.11.1 Transversality

Solution (Exercise 12.1) (a) The tangent space to 𝕊1 at 𝑧 is 𝑇𝑧 (𝕊1 ) = {𝑣 ∈ ℝ2 ∶

𝑣 ⋅ 𝑧 = 0}. The tangent space to 𝑁𝑧 at 𝑧 is 𝑇𝑧 (𝑁𝑧 ) = {(0, 𝑦) ∶ 𝑦 ∈ ℝ} ⊂ ℝ2 . Since
𝑇𝑧 (ℝ2 ) = ℝ2 , 𝑇𝑧 (𝕊1 ) and 𝑇𝑧 (𝑁𝑧 ) span all of 𝑇𝑧 (ℝ2 ) for all 𝑧 ≠ (±1, 0). Hence
𝕊1 − ⋔ 𝑁𝑧 if and only if 𝑧 ≠ (±1, 0). (Drawing a picture explains everything.)

(b) Which of the following linear spaces intersect transversally?

∙ The 𝑥𝑦-plane and the 𝑧-axis.

Answer: Transverse, since the two spaces span all of ℝ3 .
∙ The 𝑥𝑦-plane and the plane spanned by {(3, 2, 0), (0, 4, −1)}.
Answer: Transverse, since the two planes span all of ℝ3 .
∙ The plane spanned by {(1, 0, 0), (2, 1, 0)} and the 𝑦-axis in ℝ3 .
Answer: Not transverse, since the 𝑦-axis and lies in the span of
{(1, 0, 0), (2, 1, 0)}.
∙ ℝ𝑘 × {0} and {0} × ℝ𝑙 in ℝ𝑛 . (The answer depends on 𝑘, 𝑙, and 𝑛.)
Answer: Transverse, if 𝑘 + 𝑙 ≥ 𝑛.
∙ 𝑉 × {0} and the diagonal in 𝑉 × 𝑉 , for a real vector space 𝑉 .
Answer: Transverse, since they span all of 𝑉 × 𝑉 : any (𝑣, 𝑤) ∈ 𝑉 × 𝑉 is
equal to the sum of (𝑤, 𝑤) ∈ Δ𝑉 and (𝑣 − 𝑤, 0) ∈ 𝑉 × {0}.
∙ The spaces of symmetric (𝐴𝑡 = 𝐴) and skew symmetric (𝐴𝑡 = −𝐴) matrices
in 𝑀(𝑛).
Answer: Transverse, since every matrix in 𝑀(𝑛) can be written as a sum of
a symmetric and an antisymmetric matrix:

1 1
𝐶 = (𝐶 + 𝐶 𝑡 ) + (𝐶 − 𝐶 𝑡 ) for any 𝐶 ∈ 𝑀(𝑛).
2 2

(c) Yes, 𝑆𝐿(𝑛) and 𝑂(𝑛) do not meet transversally in 𝑀(𝑛), since 𝑆𝑂(𝑛) is contained
in 𝑆𝐿(𝑛). Hence we also have 𝑇𝐴 (𝑂(𝑛)) = 𝑇𝐴 (𝑆𝑂(𝑛)) ⊆ 𝑇𝐴 (𝑆𝐿(𝑛)), and these
tangent spaces do not span all of 𝑇𝐴 (𝑀(𝑛)) = 𝑀(𝑛).

Solution (Exercise 12.2) The image of 𝑓 is a submanifold of ℝ2 . This follows, for

example, from the fact that 𝑓 is an embedding. We could also observe that Im (𝑓 ) =
𝑔 −1 (1) and remark that 1 is a regular value of 𝑔. The composition 𝑔◦𝑓 is the constant
map ℝ → ℝ with value 1. Hence (𝑔◦𝑓 )−1 (1) = ℝ is a manifold.

Solution (Exercise 12.3) By assumption, 𝑔 is transverse to 𝑊 , i.e.,

Im (𝑑𝑔𝑦 ) + 𝑇𝑔(𝑦) (𝑊 ) = 𝑇𝑔(𝑦) (𝑍) for all 𝑦 ∈ 𝑌 with 𝑔(𝑦) ∈ 𝑊 . (A.6)

412 A.11. Transversality

Now we assume 𝑓 −
⋔ 𝑔 −1 (𝑊 ), i.e.,

Im (𝑑𝑓𝑥 ) + 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) = 𝑇𝑓 (𝑥) (𝑌 ) for all 𝑥 ∈ 𝑋 with 𝑓 (𝑥) ∈ 𝑔 −1 (𝑊 ). (A.7)

We need to show (𝑔◦𝑓 ) − ⋔ 𝑊 . So let 𝑥 ∈ 𝑋 be a point such that 𝑔(𝑓 (𝑥)) ∈ 𝑊 and let
𝑐 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑍). By (A.6), there are vectors 𝑏1 ∈ 𝑇𝑓 (𝑥) (𝑌 ) and 𝑏2 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) such
that
𝑑𝑔𝑓 (𝑥) (𝑏1 ) + 𝑏2 = 𝑐.
By (A.7), there are vectors 𝑎 ∈ 𝑇𝑥 (𝑋) and 𝑏3 ∈ 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) such that

𝑑𝑓𝑥 (𝑎) + 𝑏3 = 𝑏1 .

Putting these two equations together we get

𝑐 = 𝑑𝑔𝑓 (𝑥) (𝑑𝑓𝑥 (𝑎) + 𝑏3 ) + 𝑏2 = 𝑑𝑔𝑓 (𝑥) (𝑑𝑓𝑥 (𝑎1 )) + 𝑑𝑔𝑓 (𝑥) (𝑏3 ) + 𝑏2 .

By the chain rule, we have 𝑑(𝑔◦𝑓 )𝑥 = 𝑑𝑔𝑓 (𝑥) ◦𝑑𝑓𝑥 . By a previous exercise, we
know 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) = (𝑑𝑔𝑓 (𝑥) )−1 (𝑇𝑔(𝑓 (𝑥)) (𝑊 ) in 𝑇𝑓 (𝑥) (𝑌 ). In particular, 𝑑𝑔𝑓 (𝑥) (𝑏3 ) ∈
𝑇𝑔(𝑓 (𝑥)) (𝑊 ) and thus
𝑑𝑔𝑓 (𝑥) (𝑏3 ) + 𝑏2 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ).
Hence we have

𝑐 = 𝑑(𝑔◦𝑓 )𝑥 (𝑎) + 𝑑𝑔𝑓 (𝑥) (𝑏3 ) + 𝑏2 ∈ [Im (𝑑(𝑔◦𝑓 )𝑥 ) + 𝑇𝑔(𝑓 (𝑥)) (𝑊 )] ⊂ 𝑇𝑔(𝑓 (𝑥)) (𝑍).

Since 𝑐 was an arbitrary element in 𝑇𝑔(𝑓 (𝑥)) (𝑌 ), we have proven

Im (𝑑(𝑔◦𝑓 )𝑥 ) + 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) = 𝑇𝑔(𝑓 (𝑥)) (𝑍) for all 𝑥 ∈ 𝑋 with 𝑔(𝑓 (𝑥)) ∈ 𝑊 .

In other words, (𝑔◦𝑓 ) −

⋔𝑊.
Now we assume (𝑔◦𝑓 ) −
⋔ 𝑊 , i.e.,

Im (𝑑(𝑔◦𝑓 )𝑥 ) + 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) = 𝑇𝑔(𝑓 (𝑥)) (𝑍) for all 𝑥 ∈ 𝑋 with 𝑔(𝑓 (𝑥)) ∈ 𝑊 . (A.8)

Let 𝑥 ∈ 𝑋 be a point such that 𝑓 (𝑥) ∈ 𝑔 −1 (𝑊 ) and let 𝑏 ∈ 𝑇𝑓 (𝑥) (𝑌 ). Since 𝑔(𝑓 (𝑥)) ∈ 𝑊 ,
we can use (A.8) to find vectors 𝑎 ∈ 𝑇𝑥 (𝑋) and 𝑐 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ) such that

𝑑(𝑔◦𝑓 )𝑥 (𝑎) + 𝑐 = 𝑑𝑔𝑓 (𝑥) (𝑏).

By the chain rule, we have 𝑑(𝑔◦𝑓 )𝑥 (𝑎) = 𝑑𝑔𝑓 (𝑥) (𝑑𝑓𝑥 (𝑎)). Thus

𝑑𝑔𝑓 (𝑥) (𝑏 − 𝑑𝑓𝑥 (𝑎)) = 𝑐 ∈ 𝑇𝑔(𝑓 (𝑥)) (𝑊 ),

In other words,

𝑏 − 𝑑𝑓𝑥 (𝑎) ∈ (𝑑𝑔𝑓 (𝑥) )−1 (𝑇𝑔(𝑓 (𝑥)) (𝑊 )) = 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )).

Since 𝑏 was an arbitrary element, we have proven Im (𝑑𝑓𝑥 ) + 𝑇𝑓 (𝑥) (𝑔 −1 (𝑊 )) = 𝑇𝑓 (𝑥) (𝑌 )

for all 𝑥 ∈ 𝑋 with 𝑓 (𝑥) ∈ 𝑔 −1 (𝑊 ). In other words, 𝑓 −
⋔ 𝑔 −1 (𝑊 ).
Appendix A. Solutions to exercises 413

Solution (Exercise 12.4) By definition, Γ(𝐴) −

⋔ Δ if and only if Γ(𝐴)+Δ(𝑉 ) = 𝑉 ×𝑉 .
Let (𝑣1 , 𝑣2 ) be an arbitrary element of 𝑉 × 𝑉 . We need to check under which conditions
we can find 𝑣, 𝑤 ∈ 𝑉 such that

(𝑣1 , 𝑣2 ) = (𝑣, 𝑣) + (𝑤, 𝐴𝑤) = (𝑣 + 𝑤, 𝑣 + 𝐴𝑤), i.e. 𝑣1 = 𝑣 + 𝑤 and 𝑣2 = 𝑣 + 𝐴𝑤.

If we can find a suitable 𝑤, then we just set 𝑣 ∶= 𝑣1 − 𝑤. Hence, by taking the difference
of the two equations, we see that the question is reduced to checking whether we can find
a 𝑤 such that 𝑣2 − 𝑣1 = 𝐴𝑤 − 𝑤 = (𝐴 − 𝐼)𝑤 where 𝐼 is the identity map of 𝑉 . But such
a 𝑤 exists for any choice of 𝑣1 and 𝑣2 if and only if 𝐴 − 𝐼 is invertible, i.e. if and only if
det(𝐴 − 𝐼) ≠ 0 which happens if and only if +1 is not an eigenvalue of 𝐴 (because the
eigenvalues are the 𝜆 such that det(𝐴 − 𝜆𝐼) = 0).

Solution (Exercise 12.5) Let Δ𝑋 = {(𝑥, 𝑥) ∶ 𝑥 ∈ 𝑋} ⊆ 𝑋 × 𝑋 be the diagonal of

𝑋 and Γ(𝑓 ) = {(𝑥, 𝑓 (𝑥)) ∶ 𝑥 ∈ 𝑋} ⊆ 𝑋 × 𝑋 be the graph of 𝑓 . Then the set of fixed
points of 𝑓 in 𝑋 is the intersection Δ𝑋 ∩ Γ(𝑓 ). For
𝑥 = 𝑓 (𝑥) ⇐⇒ (𝑥, 𝑥) = (𝑥, 𝑓 (𝑥)) ⇐⇒ (𝑥, 𝑥) ∈ Γ(𝑓 ).
Recall that a 0-dimensional manifold is just a discrete set of points. Since 𝑋 is compact,
𝑋 × 𝑋 is also compact. Hence a 0-dimensional submanifold is a discrete subset of the
compact space 𝑋 × 𝑋 and is therefore finite. Note that we have used this before: dis-
crete closed subspaces of compact spaces are finite. Thus, in order to prove that 𝑓 has
only finitely many fixed points, it suffices to show that Δ𝑋 ∩ Γ(𝑓 ) is a 0-dimensional
submanifold of 𝑋 × 𝑋.
Hence we would like to show that Δ𝑋 and Γ(𝑓 ) meet transversally in 𝑋 × 𝑋. Since
both have codimension equal to dim 𝑋 =∶ 𝑛 in 𝑋 × 𝑋, the Transversality Theorem
then implies that Δ𝑋 ∩ Γ(𝑓 ) is a 0-dimensional submanifold of 𝑋 × 𝑋. By definition,
Δ𝑋 −⋔ Γ(𝑓 ) means
𝑇(𝑥,𝑥) (Γ(𝑓 )) + 𝑇(𝑥,𝑥) (Δ𝑋 ) = 𝑇(𝑥,𝑥) (𝑋 × 𝑋)
for every point (𝑥, 𝑥) ∈ Δ𝑋 ∩ Γ(𝑓 ). We know 𝑇(𝑥,𝑥) (Γ(𝑓 )) = Γ(𝑑𝑓𝑥 ) and 𝑇(𝑥,𝑥) (Δ𝑋 ) =
Δ𝑇𝑥 (𝑋) by a previous exercise. Moreover, we know 𝑇(𝑥,𝑥) (𝑋 × 𝑋) = 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋).
Hence we need to show
Γ(𝑑𝑓𝑥 ) + Δ𝑇𝑥 (𝑋) = 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋) (A.9)
for every point (𝑥, 𝑥) ∈ Δ𝑋 ∩ Γ(𝑓 ). This means exactly Γ(𝑑𝑓𝑥 ) −⋔ Δ𝑇𝑥 (𝑋) which we have
shown to be true if +1 is not an eigenvalue of 𝑑𝑓𝑥 in the previous exercise. Thus we can
stop here, since 𝑓 is Lefschetz by assumption.
But we could also just continue and give another proof as follows: Since we know
dim Δ𝑇𝑥 (𝑋) = dim Γ(𝑑𝑓𝑥 ) = dim 𝑇𝑥 (𝑋),
equality (A.9) will follow once we show Γ(𝑑𝑓𝑥 ) ∩ Δ𝑇𝑥 (𝑋) = {0}. For then we have
shown that Γ(𝑑𝑓𝑥 ) + Δ𝑇𝑥 (𝑋) is a subspace of 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋) of the same dimension as
𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋). (Recall that in a finite dimensional vector space 𝑉 with subspaces 𝑈 and
𝑊 , the following dimension formula holds:
dim 𝑈 + dim 𝑊 = dim(𝑈 + 𝑊 ) − dim(𝑈 ∩ 𝑊 )
414 A.11. Transversality
where 𝑈 + 𝑊 ⊆ 𝑉 is the subspace of 𝑉 generated by 𝑈 and 𝑊 and 𝑈 ∩ 𝑊 is their
intersection.)
Since 𝑓 is a Lefschetz map, we know that for every fixed point 𝑥 of 𝑓 , +1 is not an
eigenvalue of 𝑑𝑓𝑥 . This means that 𝑑𝑓𝑥 does not have any fixed points, for there is no
𝑣 ∈ 𝑇𝑥 (𝑋) ⧵ {0} with 𝑑𝑓𝑥 (𝑣) = 1 ⋅ 𝑣. As we observed for 𝑓 above, this is equivalent to

Γ(𝑑𝑓𝑥 ) ∩ Δ𝑇𝑥 (𝑋) = {0}.

Solution (Exercise 12.6) We define the map

𝑓 ∶ ℂ5 ⧵ {0} → ℂ,
(𝑧1 , … , 𝑧5 ) → 𝑧21 + 𝑧22 + 𝑧23 + 𝑧34 + 𝑧6𝑘−1
5
.

Setting 𝑍 = 𝑓 −1 (0), we need to show that 𝑍 and 𝕊9 meet transversally. The tangent
space to 𝑍 in a point 𝑧 ∈ 𝑍 is the kernel of the derivative 𝑑𝑓𝑧 . Since 𝑓 is a polynomial
in the variables 𝑧1 , … , 𝑧5 , we can use our usual rules for partial differentiation to get the
following matrix for 𝑑𝑓𝑧 (in the standard basis):
( )
𝑑𝑓𝑧 ∶ ℂ5 → ℂ, 𝑑𝑓𝑧 = 2𝑧1 , 2𝑧2 , 2𝑧3 , 3𝑧24 , (6𝑘 − 1)𝑧6𝑘−25
.
( )
𝑥 −𝑦
Recall that we can represent every element 𝑥+𝑖𝑦 ∈ ℂ by the real 2×2-matrix .
𝑦 𝑥
Then we see that 𝑑𝑓𝑧 is a real 2 × 10-matrix. Its maximal rank (as a matrix with entries in
ℝ) is therefore 2. And, in fact, for every 𝑧 ≠ 0, 𝑑𝑓𝑧 has rank 2, since it maps surjectively
onto ℂ ≅ ℝ2 . Thus 0 is a regular value for 𝑓 and the tangent space 𝑇𝑧 (𝑍) is the kernel
of 𝑑𝑓𝑧 .
Writing a complex number 𝑧 = 𝑥 + 𝑖𝑦, we can express 𝕊9 as the fiber of the smooth
map

𝑔 ∶ ℂ5 ≅ ℝ10 → ℝ,
(𝑧1 , … , 𝑧5 ) → 𝑥21 + 𝑦21 + 𝑥22 + 𝑦22 + ⋯ + 𝑥25 + 𝑦25 − 1

at the regular value 0, i.e., 𝕊9 = 𝑔 −1 (0) ⊂ ℂ5 ≅ ℝ10 . The tangent space to 𝕊9 at 𝑧 is then
given by the kernel of the derivative (in standard bases)
( )
𝑑𝑔𝑧 ∶ ℂ5 = ℝ10 → ℝ, 𝑑𝑔𝑧 = 2𝑥1 , 2𝑦1 , 2𝑥2 , 2𝑦2 , … , 2𝑥5 , 2𝑦5 .

Thus, as expected, the tangent space 𝑇𝑧 (𝕊9 ) consists of all vectors 𝑤 in ℝ10 which are
orthogonal to 𝑧, i.e., which satisfy 𝑤 ⋅ 𝑧 = 0.
The tangent space of 𝕊9 is of dimension 9 and the tangent space of ℝ10 ⧵ {0} is of
dimension 10. Hence in order to show that 𝑍 and 𝕊9 meet transversally in ℝ10 ⧵ {0} we
need to show: For every 𝑧 ∈ 𝑍 ∩ 𝕊9 , there is at least one vector 𝑤 in 𝑇𝑧 (𝑍) which does
not belong to 𝑇𝑧 𝕊9 . Then we have 𝑇𝑧 (𝑍) + 𝑇𝑧 (𝕊9 ) ⊆ 𝑇𝑧 (ℝ10 ⧵ {0}) is a vector subspace
of the same dimension as 𝑇𝑧 (ℝ10 ⧵ {0}) and therefore equal 𝑇𝑧 (ℝ10 ⧵ {0}).
So let 𝑧 = (𝑧1 , … , 𝑧5 ) be a fixed point in 𝑍 ∩ 𝕊9 . The tangent space 𝑇𝑧 (𝑍) is the
kernel of 𝑑𝑓𝑧 . Hence we need to find at least one vector 𝑤 ∈ ℂ5 = ℝ10 with 𝑑𝑓𝑧 (𝑤) = 0
and 𝑤 ⋅ 𝑧 ≠ 0.
Appendix A. Solutions to exercises 415

Set 𝑚 ∶= 2 ⋅ 3 ⋅ (6𝑘 − 1) and 𝑤 ∶= ( 𝑚2 𝑧1 , 𝑚2 𝑧2 , 𝑚2 𝑧3 , 𝑚3 𝑧4 , 6𝑘−1

𝑚
𝑧5 ). Then we have
𝑚
⎛ 2 𝑧1 ⎞
⎜ 𝑚 𝑧2 ⎟
( ) ⎜ 𝑚2 ⎟
𝑑𝑓𝑧 (𝑤) = 2𝑧1 , 2𝑧2 , 2𝑧3 , 3𝑧24 , (6𝑘 − 1)𝑧6𝑘−2
5
⋅ ⎜ 2 𝑧3 ⎟
⎜ 𝑚 𝑧4 ⎟
⎜ 𝑚3 ⎟
⎝ 6𝑘−1 𝑧5 ⎠
( )
= 𝑚 𝑧21 + 𝑧22 + 𝑧23 + 𝑧34 + 𝑧6𝑘−1
5
=0

since 𝑧 is by assumption a point on 𝑍. Hence 𝑤 ∈ 𝑇𝑧 (𝑍).

On the other hand, we can calculate the inner product of 𝑤 and 𝑧 as vectors in ℝ10 ,
using (𝑥𝑖 , 𝑦𝑖 ) for the coordinates of 𝑧𝑖 in ℝ2 ≅ ℂ, and get

⎛𝑥1 ⎞
⎜ 𝑦1 ⎟
( ) ⎜𝑥 ⎟
𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚 𝑚
𝑤⋅𝑧= 𝑥 , 𝑦 , 𝑥 , 𝑦 , 𝑥 , 𝑦 , 𝑥 , 𝑦 , 𝑥 , 𝑦 ⋅ ⎜ 2⎟
2 1 2 1 2 2 2 2 2 3 2 3 3 4 3 4 6𝑘 − 1 5 6𝑘 − 1 5 ⎜ ⋮ ⎟
⎜𝑥 ⎟
⎜ 5⎟
⎝ 𝑦5 ⎠
𝑚 2 𝑚 2 𝑚 2 𝑚 2 𝑚 2
= |𝑧1 | + |𝑧2 | + |𝑧3 | + |𝑧4 | + |𝑧 |
2 2 2 3 6𝑘 − 1 5
>0

which is bigger than zero, since 𝑧 is a point on 𝕊9 . Thus 𝑤 is a vector in 𝑇𝑧 (ℝ10 ) which
is in 𝑇𝑧 (𝑍), but not in 𝑇𝑧 (𝕊9 ), and we have shown

𝑇𝑧 (𝑍) + 𝑇𝑧 (𝕊9 ) = 𝑇𝑧 (ℝ10 ⧵ {0}).

Hence 𝑍 and 𝕊9 meet transversally in ℝ10 ⧵{0}. The codimension of 𝑍 ∩𝕊9 in ℝ10 ⧵{0}
is 2 + 1 by the codimension formula. Thus dim(𝑍 ∩ 𝕊9 ) = 10 − 3 = 7.
416 A.12. Intersection Theory modulo 2
A.12 Intersection Theory modulo 2

Solution (Exercise 13.1) We showed in a previous exercise that

𝑓−
⋔ 𝑔 −1 (𝑊 ) if and only if (𝑔◦𝑓 ) −
⋔𝑊.

Moreover, since codimensions are preserved under taking preimages, we have

dim 𝑋 + dim 𝑊 = dim 𝑍 if and only if dim 𝑋 + dim 𝑔 −1 (𝑊 ) = dim 𝑌 .

Thus,
𝐼2 (𝑓 , 𝑔 −1 (𝑊 )) is defined if and only if 𝐼2 (𝑔◦𝑓 , 𝑊 ) is defined.
Now it remains to observe that the finite numbers satisfy

𝐼2 (𝑓 , 𝑔 −1 (𝑊 )) = #𝑓 −1 (𝑔 −1 (𝑊 )) = #(𝑔◦𝑓 )−1 (𝑊 ) = 𝐼2 (𝑔◦𝑓 , 𝑊 ).

Solution (Exercise 13.2) (a) Let 𝑍 ⊂ 𝑌 be any closed submanifold with dim 𝑋 +
dim 𝑍 = dim 𝑌 . For dim 𝑋 ≥ 1, this implies dim 𝑌 > 0 and dim 𝑍 = dim 𝑌 −
dim 𝑋 < dim 𝑌 . In particular, 𝑍 is not all of 𝑌 and there exist points in 𝑌 which
are not in 𝑍.
We can assume that 𝑌 is connected, since 𝑓 being homotopic to a constant map
implies that the image of 𝑓 lies in only one connected component of 𝑌 . Hence any
part of 𝑍 in a different connected component satisfies 𝑓 −1 (𝑍) = ∅.
Assume that 𝑓 is homotopic to the constant map 𝑔0 ∶ 𝑋 → {𝑦0 } ⊂ 𝑌 . If 𝑦0 ∈ 𝑍,
then 𝑔 and 𝑍 do not meet transversally, since dim 𝑍 < dim 𝑌 and dim{𝑦0 } = 0.
But since 𝑌 is connected and a smooth manifold, it is path-connected. Hence there
is a path 𝛾 from 𝑦0 to a point 𝑦1 with 𝑦1 ∉ 𝑍. Then 𝑔0 is homotopic to the constant
map 𝑔1 ∶ 𝑋 → {𝑦1 } ⊂ 𝑌 by composing 𝑔0 with the path 𝛾. Since being smoothly
homotopic is a transitive relation, we have 𝑓 ≃ 𝑔1 .
Since 𝑦 ∉ 𝑍, we have 𝑔 −
1 1⋔ 𝑍. This means 𝑔 −1 (𝑍) = ∅, i.e.,
1

𝐼2 (𝑔1 , 𝑍) = #𝑔1−1 (𝑍) = 0.

By the homotopy invariance of 𝐼2 (−, 𝑍) we have 𝐼2 (𝑓 , 𝑍) = 𝐼2 (𝑔1 , 𝑍), and hence
𝐼2 (𝑓 , 𝑍) = 0.
(b) Assume that 𝑋 = {𝑥} is a one-point space. In particular, we have dim 𝑋 = 0.
Thus any closed submanifold 𝑍 ⊂ 𝑌 with dim 𝑋 +dim 𝑍 = dim 𝑌 is of dimension
dim 𝑍 = dim 𝑌 . This implies that every such 𝑍 intersects every other submanifold
of 𝑌 transversally. In particular, {𝑥} −
⋔ 𝑍, i.e., 𝑓 −
⋔ 𝑍 for every map 𝑓 ∶ 𝑋 =
{𝑥} → 𝑌 . Hence, by definition of intersection numbers, we can calculate 𝐼2 (𝑓 , 𝑍)
by assuming 𝑓 (𝑥) = 𝑦 ∈ 𝑍 and get
𝐼2 (𝑓 , 𝑍) = #𝑓 −1 (𝑍) = #{𝑦} = 1 ≠ 0.

(c) Since 𝕊1 is compact, the intersection number 𝐼2 (id𝕊1 , {𝑥}) is defined for every
point 𝑥 ∈ 𝕊1 where we consider intersections in 𝕊1 , i.e., 𝑌 = 𝑋 = 𝕊1 . Since the
Appendix A. Solutions to exercises 417

id−1
𝕊1
(𝑥) = {𝑥}, we have 𝐼2 (id𝕊1 , {𝑥}) = 1 ≠ 0. But we just learned that if id𝕊1 was
homotopic to a constant map, then 𝐼2 (Id𝕊1 , {𝑥}) had to be zero, since dim 𝕊1 = 1.
Thus there is at least one map 𝕊1 → 𝕊1 which is not homotopic to a constant map.
In fact, up to homotopy there is one homotopy class of maps 𝕊1 → 𝕊1 for every
integer 𝑛 ∈ ℤ represented by taking 𝑛th powers in ℂ: 𝑧 → 𝑧𝑛 . We will learn more
about that later.

Solution (Exercise 13.3) (a) Let 𝑌 be contractible and dim 𝑌 > 0. Let 𝑓 ∶ 𝑋 → 𝑌
be a smooth map with 𝑋 compact and 𝑍 ⊂ 𝑌 closed, and dim 𝑋 + dim 𝑍 =
dim 𝑌 . Since 𝑌 is contractible, the identity map id𝑌 is homotopic to a constant
map 𝑌 → {𝑦} ⊂ 𝑌 . Thus 𝑓 is homotopic to the constant map 𝑋 → {𝑦} ⊂ 𝑌 .
Since dim 𝑌 > 0, we have {𝑦} ≠ 𝑌 and there are points in 𝑌 different from 𝑦.
Hence, since dim 𝑋 ≥ 1, the previous exercise implies 𝐼2 (𝑓 , 𝑍) = 0.
Finally, we recall that Euclidean space ℝ𝑘 is contractible for all 𝑘.

(b) If 𝑋 is compact, then the intersection number 𝐼2 (id𝑋 , {𝑥}) is defined for every
point 𝑥 ∈ 𝑋 where we consider intersections in 𝑋 (with 𝑌 = 𝑋 and 𝑍 = {𝑥}). For
the dimensions are complementary and id𝑋 meets every submanifold transversally.
This intersection number satisfies 𝐼2 (id𝑋 , {𝑥}) = #{𝑥} = 1, since 𝑥 has exactly
one preimage under id𝑋 .
If 𝑋 is contractible and dim 𝑋 ≥ 1, we can apply the previous point. Or we note
that we have id𝑋 ∼ 𝑔 where 𝑔 ∶ 𝑋 → {𝑦} is some constant map. Since dim 𝑋 ≥ 1,
we have 𝑔 −⋔ {𝑥} for 𝑥 ∈ 𝑋 if and only if 𝑥 ≠ 𝑦. Hence 𝐼2 (𝑔, {𝑥}) = #𝑔 −1 (𝑥) = 0.
However, since id𝑋 ≃ 𝑔, we have 𝐼2 (id𝑋 , {𝑥}) = 𝐼2 (𝑔, {𝑥}). This is impossible.
Hence, if dim 𝑋 ≥ 1, 𝑋 cannot be both compact and contractible.
If dim 𝑋 = 0, then the only way 𝑋 can be contractible is that it consists of just one
point. This is a compact space.

Solution (Exercise 13.4) (a) Let 𝑓 ∶ 𝑋 → 𝕊𝑘 be a smooth map with 𝑋 compact

and 0 < dim 𝑋 < 𝑘. Let 𝑍 be a closed submanifold 𝑍 ⊂ 𝕊𝑘 of dimension
complementary to 𝑋, i.e., dim 𝑋 + dim 𝑍 = dim 𝕊𝑘 . The image of the derivative
𝑑𝑓𝑥 at any point has at most dimension dim 𝑋 = dim 𝑇𝑥 (𝑋) in 𝑇𝑓 (𝑥) (𝕊𝑘 ). Hence,
since dim 𝑋 < 𝑘, 𝑑𝑓𝑥 cannot be surjective. Hence the only points in 𝕊𝑘 which
are regular values of 𝑓 are the points which are not in the image of 𝑓 . By Sard’s
Theorem, however, we know that regular values exist. Hence there must be a point
𝑝 ∈ 𝕊𝑘 with 𝑝 ∉ 𝑓 (𝑋) ∩ 𝑍.
Now stereographic projection provides a diffeomorphism 𝜙 ∶ 𝕊𝑘 ⧵ {𝑝} → ℝ𝑘 .
Since 𝑝 is not in the image of 𝑓 , the composition 𝜙◦𝑓 ∶ 𝑋 → 𝕊𝑘 ⧵ {𝑝} → ℝ𝑘 is
defined. Similarly, since 𝑝 is not in 𝑍, 𝜙|𝑍 is a diffeomorphism. Hence we can
consider 𝜙(𝑍) as a submanifold of ℝ𝑘 of dimension dim 𝑍 which is diffeomorphic
to 𝑍. This implies
𝐼2 (𝑓 , 𝑍) = 𝐼2 (𝜙◦𝑓 , 𝜙(𝑍))
where the latter is the intersection number mod 2 in ℝ𝑘 . But since ℝ𝑘 is con-
tractible, the previous exercise implies 𝐼2 (𝜙◦𝑓 , 𝜙(𝑍)) = 0. Thus 𝐼2 (𝑓 , 𝑍) = 0.
418 A.12. Intersection Theory modulo 2

(b) Considering 𝕊1 ⊂ ℂ, we can define tow submanifolds 𝑋 and 𝑍 as 𝑋 = 𝕊1 × {1}

and 𝑍 = {1} × 𝕊1 in 𝕊1 × 𝕊1 . They intersect in the single point 𝑞 given by
𝑞 ∶= {1} × {1} ∈ 𝑆 1 × 𝑆 1 . The tangent space to 𝕊1 × 𝕊1 at 𝑞 is

𝑇𝑞 (𝕊1 × 𝕊1 ) = 𝑇1 (𝕊1 ) × 𝑇1 (𝕊1 ) = 𝑇1 (𝑋) × 𝑇1 (𝑍).

Hence 𝑋 and 𝑍 meet transversally in 𝕊1 × 𝕊1 . Thus their intersection number in

𝕊1 × 𝕊1 is 𝐼2 (𝑋, 𝑍) = 1.
Now assume there was a diffeomorphism 𝜑 ∶ 𝕊1 × 𝕊1 → 𝕊2 . The map 𝜄 ∶ 𝕊1 →
𝕊1 × 𝕊1 , 𝑧 → (𝑧, 1) maps 𝕊1 diffeomorphically into 𝑋. Composition with 𝜑 gives
us a map 𝜑◦𝜄 ∶ 𝕊1 → 𝕊2 . Since 𝜑 is a diffeomorphism, 𝜑(𝑍) is the diffeomorphic
image of 𝑍, and we would have

𝐼2 (𝑋, 𝑍) = 𝐼2 (𝜄, 𝑍) = 𝐼2 (𝜑◦𝜄, 𝜑(𝑍))

where the latter is the intersection number in 𝕊2 .

But by the previous point, 𝐼2 (𝜑◦𝜄, 𝜑(𝑍)) = 0 since dim 𝕊1 = 1 < 2.

Solution (Exercise 13.5) (a) Let 𝐼 ∶ 𝑋 × [0, 1] → 𝑌 be a smooth homotopy with

𝐼(𝑥, 0) = 𝑖0 (𝑥) being the embedding of 𝑋 in 𝑌 and 𝑖1 (𝑋) = 𝑍 ⊂ 𝑌 . We define a
new map
𝐹 ∶ 𝑋 × [0, 1] → 𝑌 × [0, 1], (𝑥, 𝑡) → (𝐼(𝑥, 𝑡), 𝑡).
Since 𝐼 is smooth, 𝐹 is smooth.
Now we define 𝑊 ∶= 𝐹 (𝑋 ×[0, 1]) ⊂ 𝑌 ×[0, 1] to be the image of 𝑋 ×[0, 1] under
𝐹 . We claim that 𝑊 is a compact smooth manifold with boundary. First, since 𝑋
is compact, 𝑋 × [0, 1] is a compact, and hence its continuous image 𝐹 (𝑋 × [0, 1])
in 𝑌 × [0, 1] is compact.
Since 𝐼(−, 𝑡) is an embedding for every 𝑡 ∈ [0, 1], and the identity map on the
second component is obviously an embedding as well, 𝐹 is also an embedding.
Thus the image of 𝐹 is a smooth manifold.
The boundary of 𝑊 is then given by
𝜕𝑊 = 𝑊 ∩ 𝜕(𝑌 × [0, 1])
= 𝐹 (𝑋 × {0}) ∩ 𝑌 × {0} ∪ 𝐹 (𝑋 × {1}) ∩ 𝑌 × {1}
= 𝑋 × {0} ∪ 𝑍 × {1}.
Hence 𝑊 is a cobordism between 𝑋 and 𝑍.
(b) Let 𝑋 and 𝑍 be cobordant in 𝑌 , and let 𝐶 be a compact submanifold 𝐶 in 𝑌 with
dimension complementary to 𝑋 and 𝑍. Let 𝑓 denote the restriction to 𝑊 of the
projection map 𝑌 × [0, 1] → 𝑌 . Then 𝜕𝑓 = 𝑓|𝜕𝑊 is a smooth map which can be
extended to a map 𝑓 ∶ 𝑊 → 𝑌 . Hence, by the Boundary Theorem, 𝐼2 (𝜕𝑓 , 𝑉 ) = 0
for any closed submanifold 𝑉 of 𝑌 of dimension dim 𝑉 = dim 𝑌 − dim 𝜕𝑊 . Since
dim 𝜕𝑊 = dim 𝑋 = dim 𝑍, we have, in particular,
𝐼2 (𝜕𝑓 , 𝐶) = 0.
Appendix A. Solutions to exercises 419
By the Transversality Homotopy Theorem 12.27, we can assume 𝑓 − ⋔ 𝐶 and 𝜕𝑓 −
⋔
− −
𝐶. In particular, we can assume 𝑋 ⋔ 𝐶 and 𝑍 ⋔ 𝐶. Hence 𝐼2 (𝑋, 𝐶) = #(𝑋 ∩ 𝐶)
and 𝐼2 (𝑍, 𝐶) = #(𝑍 ∩ 𝐶). By definition of 𝑓 , (𝜕𝑓 )−1 (𝐶) is given by

(𝜕𝑓 )−1 (𝐶) = 𝜕𝑊 ∩ (𝐶 × [0, 1])

= (𝑋 × {0}) ∩ (𝐶 × [0, 1]) ∪ (𝑍 × {1}) ∩ (𝐶 × [0, 1])
= ((𝑋 ∩ 𝐶) × {0}) ∪ ((𝑍 ∩ 𝐶) × {1}).

Thus

0 = 𝐼2 (𝜕𝑓 , 𝐶)
= #(𝜕𝑓 )−1 (𝐶)
= #((𝑋 ∩ 𝐶) × {0}) + #((𝑍 ∩ 𝐶) × {1})
= #(𝑋 ∩ 𝐶) + #(𝑍 ∩ 𝐶)
= 𝐼2 (𝑋, 𝐶) + 𝐼2 (𝑍, 𝐶).

Since we are working modulo 2, this implies

𝐼2 (𝐶, 𝑋) = 𝐼2 (𝐶, 𝑍).

Solution (Exercise 13.6) By Theorem 12.14, there is an open neighborhood 𝑍 𝜀 ⊂ 𝑋

of 𝑍 and a diffeomorphism 𝑍 𝜀 ≅ 𝑁 𝜀 (𝑍, 𝑋) to an open subset 𝑁 𝜀 (𝑍, 𝑋) ⊂ 𝑁(𝑍, 𝑋).
Now we use the assumption on 𝑍. By Theorem 12.15, there is a diffeomorphism 𝑍 ×
ℝ𝑘 → 𝑁(𝑍, 𝑋). Hence we may identify 𝑁(𝑍, 𝑋) = 𝑍 × ℝ𝑘 , and 𝑁 𝜀 (𝑍, 𝑋) = 𝑍 ×
(−𝜀, 𝜀). However, for any 𝑣 ≠ 0 in ℝ𝑘 , the subspaces 𝑍 = 𝑍 × {0} and the deformation
𝑍 × {𝑣} do not intersect in 𝑍 × ℝ𝑘 . Thus, the diffeomorphism 𝑍 𝜀 ≅ 𝑍 × (−𝜀, 𝜀)
implies that there is a deformation 𝑍 ′ of 𝑍 in 𝑋 such that 𝑍 ∩ 𝑍 ′ = ∅. Hence we have
𝐼2 (𝑍, 𝑍) = 0.

Solution (Exercise 13.7) (a) We need to find a deformation 𝑍 ′ of 𝑍 which intersects

𝑍 transversally and then count the intersection points. For example, let us look at
the map
( ( ))
1 𝜋
𝑓 ∶ [−1, 1] → [−1, 1] × (−1, 1), 𝑡 → 𝑡, sin 𝑡 ⋅ .
2 2
Let 𝑍 ′ be the image of 𝑓 in 𝑋. There is a smooth homotopy
( ( ))
1 𝜋
𝑓𝑠 ∶ [−1, 1] × [0, 1] → 𝑋, (𝑡, 𝑠) → 𝑡, 𝑠 ⋅ sin 𝑡 ⋅
2 2
with 𝑓0 being the embedding of 𝑍 into 𝑋, i.e., the map

𝑓0 ∶ [−1, 1] → 𝑋, 𝑡 → (𝑡, 0),

and 𝑓1 being the embedding of 𝑍 ′ into 𝑋. (Note that, by abuse of notation, we have
not distinguished between a point in [−1, 1] × (−1, 1) and its equivalence class in
𝑋.) Moreover, there is only one point where 𝑍 and 𝑍 ′ meet, namely the image of
420 A.12. Intersection Theory modulo 2
𝑓 (0) in 𝑋, i.e., the point with coordinates 𝑝 = (0, 0). Finally, the intersection of 𝑍
and 𝑍 ′ is transverse. We check this locally for the induced maps on the codomains
of local charts: the tangent space to 𝑋 at 𝑝 (is the )
𝑥-axis in ℝ2 and the tangent space
1
to 𝑍 ′ at 𝑝 is the line spanned by the vector ∈ ℝ2 . Hence 𝑇𝑝 (𝑍) and 𝑇𝑝 (𝑍 ′ )
1∕2
span a two-dimensional subspace in the two-dimensional vector space 𝑇𝑝 (𝑋), i.e.,
they span all of 𝑇𝑝 (𝑋).
Thus we can conclude:

𝐼2 (𝑍, 𝑍) = 𝐼2 (𝑍, 𝑍 ′ ) = #(𝑋 ∩ 𝑋 ′ ) = 1.

(b) By Exercise 13.6, the existence of a submersion 𝑔 ∶ 𝑋 → ℝ such that 𝑍 = 𝑔 −1 (0)

would imply 𝐼2 (𝑍, 𝑍) = 0. This contradicts the previous point. Hence such a 𝑔
cannot exist.

Solution (Exercise 13.8) (a) We have seen this map in Section 10.1 when we
introduced the idea of intersection theory. There we showed that 𝑓 is smooth by
checking it for the induced map on local charts. It is proper, since its domain 𝕊1 is
compact. It is injective, since cos(𝑡1 ∕2) = ± cos(𝑡2 ∕2) implies 𝑡2 = 𝑡1 + 2𝜋, and
similarly for sin.
(b) We consider 𝑍 = Im (𝑓 ) as a submanifold of ℝP2 . We now need to show we
can deform 𝑍 such that it intersects itself in an odd number of points. Thinking
of 𝑍 as being the image of the horizontal equator on 𝕊2 under the quotient map
𝕊2 → ℝP2 we could try the image of a vertical equator. So let 𝑔 ∶ 𝕊1 → ℝP2 be
the map defined by sending 𝑡 ∈ [0, 2𝜋] to [0 ∶ sin(𝑡∕2) ∶ cos(𝑡∕2)]. Just as for 𝑓 ,
we can check that 𝑔 is a smooth embedding. Properness and injectivity follow as
for 𝑓 . For smoothness we observe look at the effect of 𝑔 on the codomains of local
charts: We look at the diagram
𝑔
𝕊O 1 / ℝP2

𝜙𝑖 𝜓𝑗

𝑊𝑖 / ℝ2
𝑔

with maps 𝜓𝑗 defined on the subsets 𝑉𝑗 = {[𝑥] ∶ 𝑥𝑗 ≠ 0} of ℝP2 given by

( ) ( )
𝑥 𝑥 𝑥 𝑥
𝜓2 ([𝑥1 ∶ 𝑥2 ∶ 𝑥3 ]) = 𝑥1 , 𝑥3 and 𝜓3 ([𝑥1 ∶ 𝑥2 ∶ 𝑥3 ]) = 𝑥1 , 𝑥2 . Then we have:
2 2 3 3

∙ for the parametrizations 𝜙1 ∶ 𝑊1 = (0, 𝜋) → and 𝜙2 ∶ 𝑊2 = (𝜋, 2𝜋) →

𝕊1
𝕊1 of 𝕊1 and 𝜓2 ∶ 𝑉2 = {[𝑥] ∶ 𝑥2 ≠ 0} → ℝ2 of ℝP2 , 𝑔 induces the map
( )
cos(𝑡∕2)
𝑡 → 0, .
sin(𝑡∕2)
∙ for the local parametrizations 𝜙3 ∶ 𝑊3 = (𝜋∕2, 3𝜋∕2) → 𝕊1 and 𝜙4 ∶ 𝑊4 =
(3𝜋∕2, 5𝜋∕2) → 𝕊1 of 𝕊1 and 𝜓3 ∶ 𝑉3 = {[𝑥] ∶ 𝑥3 ≠ 0} → ℝ2 of ℝP2 , 𝑔
induces the map ( )
sin(𝑡∕2)
𝑡 → 0, .
cos(𝑡∕2)
Appendix A. Solutions to exercises 421
All of these maps are smooth and hence 𝑔 is smooth.
We write 𝑍 ′ for the image of 𝑔 in ℝP2 . We have

𝑓 (𝑡) = 𝑔(𝑡) ⇐⇒ cos(𝑡∕2) = 0, i.e., if and only if 𝑡 = 𝜋.

Hence 𝑍 and 𝑍 ′ have exactly one intersection point 𝑝 = [0 ∶ 1 ∶ 0] in ℝP2 . We

need to check that this intersection is transverse in 𝑝. We can do this locally using
charts on ℝP2 and check that 𝑑𝑓𝜋 and 𝑑𝑔𝜋 together span a two-dimensional space.
We can also check transversality by looking at the actual tangent space of 𝑇𝑝 (ℝP2 ).
So let 𝐿 denote the line in ℝ3 through the origin corresponding to 𝑝, i.e., 𝐿 =
⎛0⎞
span(𝑣𝑝 ) where 𝑣𝑝 denotes the vector 𝑣𝑝 = ⎜1⎟. In Remark 9.20 we pointed out
⎜ ⎟
⎝0⎠
that the tangent space of ℝP = Gr 1 (ℝ ) at 𝑝 = 𝐿 is given by
2 3

𝑇𝑝 (ℝP2 ) = Homℝ (𝐿, 𝐿⟂ ).

To add some visual understanding note that 𝐿 is the 𝑦-axis and 𝐿⟂ is the 𝑥𝑧-plane
in ℝ3 . In particular, 𝑇𝑝 (ℝP2 ) is a two-dimensional vector space. Hence, in order to
show transversality, it suffices to show that the one-dimensional subspaces 𝑇𝑝 (𝑍)
and 𝑇𝑝 (𝑍 ′ ) are spanned by two linearly independent vectors.
⎛0⎞
A linear map 𝐿 → 𝐿⟂ is determined by where it sends the vector 𝑣𝑝 = ⎜1⎟ which
⎜ ⎟
⎝0⎠
spans 𝐿. By abusing notation a bit, we can think of

𝑓|(𝜋∕2,3𝜋∕2) ∶ (𝜋∕2, 3𝜋∕2) → ℝP2 and 𝑔|(𝜋∕2,3𝜋∕2) ∶ (𝜋∕2, 3𝜋∕2) → ℝP2

as local parametrizations of ℝP2 around 𝑝. Then the image of 𝑑𝑓𝜋 is the one-
dimensional subspace in Homℝ (𝐿, 𝐿⟂ ) spanned by linear maps which send 𝐿 to
the 𝑥-axis. And the image of 𝑑𝑔𝜋 is the one-dimensional subspace in Homℝ (𝐿, 𝐿⟂ )
spanned by linear maps which send 𝐿 to the 𝑧-axis. These two subspaces together
span all of 𝑇𝑝 (ℝP2 ). Hence 𝑍 −
⋔ 𝑍 ′.
Finally, we need to check that 𝑓 and 𝑔 are homotopic. Let 𝜑 ∶ ℝ → ℝ be a smooth
bump function such that
{
0 𝑥 ≤ 1∕4
𝜑(𝑡) =
1 𝑥 ≥ 3∕4.

We define a smooth homotopy by

𝐹 ∶ [0, 2𝜋] × [0, 1] → ℝP2 ,

√ √
(𝑡, 𝑠) → [ 1 − 𝜑(𝑠) cos(𝑡∕2) ∶ sin(𝑡∕2) ∶ 𝜑(𝑠) cos(𝑡∕2)]

such that 𝐹 (𝑡, 0) = 𝑓 (𝑡) and 𝐹 (𝑡, 1) = 𝑔(𝑡) for all 𝑡. Summarising we have proven

𝐼2 (𝑍, 𝑍) = 1.
422 A.12. Intersection Theory modulo 2

(c) Recall that any constant map 𝑐 with value not in 𝑍 is transverse to 𝑍 and has mod
2-intersection number 𝐼2 (𝑐, 𝑍) = 0. Since ℝP2 is path-connected, all constant
maps with value in ℝP2 are homotopic to each other. Hence 𝑓 is not homotopic to
a constant map.

(d) We pick a point on ℝP2 , say 𝑄 = [0 ∶ 0 ∶ 1], and let 𝑞 ∶ 𝕊1 → ℝP2 be the constant
map with value 𝑄. Let 𝜑 ∶ ℝ → ℝ again be a smooth bump function such that
{
0 𝑥 ≤ 1∕4
𝜑(𝑡) =
1 𝑥 ≥ 3∕4.

We define a smooth homotopy by

𝐻 ∶ [0, 2𝜋] × [0, 1] → ℝP2 ,

√ √ √
(𝑡, 𝑠) → [ 1 − 𝜑(𝑠) cos(𝑡) ∶ 1 − 𝜑(𝑠) sin(𝑡) ∶ 𝜑(𝑠)]

such that 𝐻(𝑡, 0) = 2𝑓 (𝑡) and 𝐻(𝑡, 1) = 𝑞(𝑡) for all 𝑡.

Finally, note that the map

𝐻̃ ∶ [0, 2𝜋] × [0, 1] → ℝP2 ,

√ √ √
(𝑡, 𝑠) → [ 1 − 𝜑(𝑠) cos(𝑡∕2) ∶ 1 − 𝜑(𝑠) sin(𝑡∕2) ∶ 𝜑(𝑠)]

does not work for 𝑓 , i.e., it does not provide a homotopy between 𝑓 and the constant
map 𝑐. One problem is that it does not induce a map on 𝕊1 × [0, 1], since
√ √
̃ 𝑠) = [ 1 − 𝜑(𝑠) ∶ 0 ∶ 𝜑(𝑠)]
𝐻(0,

whereas √ √
̃
𝐻(2𝜋, 𝑠) = [− 1 − 𝜑(𝑠) ∶ 0 ∶ 𝜑(𝑠)].
There is a minus sign occurring in the first coordinate because

cos(2𝜋∕2) = cos(𝜋) = − cos(0).

Hence
̃ 𝑠) ≠ 𝐻(2𝜋,
𝐻(0, ̃ 𝑠) in ℝP2 when 0 < 𝜑(𝑠) < 1.
This indicates that we cannot continuously deform the loop corresponding to 𝑓 to
a constant map without cutting it open at some point. And the above calculation
of the intersection number for 𝑓 shows that it is impossible to remedy this defect.
Appendix A. Solutions to exercises 423
A.13 Orientation

Solution (Exercise 14.1) (a) Replacing 𝑣𝑖 by a multiple 𝑐𝑣𝑖 corresponds to multi-

plying 𝛽 with the matrix which equals the identity matrix except at the 𝑖th position
on the diagonal where 1 is replaced with 𝑐. The determinant of this matrix is equal
to 𝑐. Hence (𝑣1 , … , 𝑣𝑘 ) and (𝑣1 , … , 𝑐𝑣𝑖 , … , 𝑣𝑘 ) are in the same equivalence class
if and only if 𝑐 > 0. If 𝑐 < 0, they have opposite orientations.

(b) Interchanging the places of 𝑣𝑖 and 𝑣𝑗 for 𝑖 ≠ 𝑗 corresponds to multiplying 𝛽 with

the matrix which equals the identity matrix with the 𝑖th and 𝑗th rows switched. We
know from Linear Algebra that the determinant of this matrix is −1.

(c) Subtracting from one 𝑣𝑖 a linear combination of the others corresponds to multi-
plying 𝛽 with a matrix that we obtain from the identity matrix by subtracting the
corresponding linear combination of rows from the 𝑖th row. We know from Linear
Algebra that this operation does not change the determinant of the matrix. Hence
the determinant of the change-of-basis-matrix is still +1.

(d) Suppose that 𝑉 is the direct sum of 𝑉1 and 𝑉2 . Let (𝑣1 , … , 𝑣𝑘 ) be an ordered
positively oriented basis of 𝑉1 and (𝑤1 , … , 𝑤𝑚 ) an ordered positively oriented
basis of 𝑉2 . Then (𝑣1 , … , 𝑣𝑘 , 𝑤1 , … , 𝑤𝑚 ) is an ordered positively oriented basis
of 𝑉1 ⊕ 𝑉2 , and (𝑤1 , … , 𝑤𝑚 , 𝑣1 , … , 𝑣𝑘 ) is an ordered positively oriented basis of
𝑉2 ⊕ 𝑉1 . Switching from the given positive basis of 𝑉1 ⊕ 𝑉2 to the positive basis of
𝑉2 ⊕ 𝑉1 corresponds to transposing exactly (dim 𝑉1 )(dim 𝑉2 ) many elements in the
basis. Hence the determinant of the change-of-basis-matrix is (−1)(dim 𝑉1 )(dim 𝑉2 ) .

Solution (Exercise 14.2) Let (𝑒1 , … , 𝑒𝑘 ) be the ordered basis of ℝ𝑘 which defines the
standard orientation of ℝ𝑘 . The orientation of ℍ𝑘 is given by the standard orientation
of ℝ𝑘 restricted to the subspace ℍ𝑘 ⊂ ℝ𝑘 . The boundary orientation of 𝜕ℍ𝑘 is given by
requiring that, at any point 𝑥 ∈ 𝜕ℍ𝑘 , the outward pointing unit normal vector 𝑛𝑥 = −𝑒𝑘
fits into a positively oriented basis for ℝ𝑘

(𝑛𝑥 , 𝑒1 , … , 𝑒𝑘−1 ) = (−𝑒𝑘 , 𝑒1 , … , 𝑒𝑘−1 ).

But the matrix which sends (𝑒1 , … , 𝑒𝑘 ) to (−𝑒𝑘 , 𝑒1 , … , 𝑒𝑘−1 ) is given by

⎛0 1 … 0⎞
⎜0 0 1 … 0⎟
⎜ ⎟
𝐴=⎜⋮ ⋮ ⋱ ⋮⎟ .
⎜0 0 … 1⎟
⎜−1 0 … 0⎟⎠
⎝
The matrix 𝐴 can be transormed into the diagonal matrix

⎛1 0 … 0 ⎞
⎜0 1 … 0 ⎟
𝐷=⎜
⎜⋮ ⋱ ⋮ ⎟⎟
⎝0 0 … −1⎠
424 A.13. Orientation

by interchanging two columns exactly 𝑘 − 1 times. Hence det(𝐷) = (−1)𝑘−1 det(𝐴). But
det(𝐷) = −1. Thus det(𝐴) = 1 > 0 if and only if (−1)𝑘 = 1, i.e., if 𝑘 is even.

Solution (Exercise 14.3) (a) At 𝑥 = (𝑎, 𝑏, 𝑐) ∈ 𝕊2 , the tangent space 𝑇𝑥 (𝕊2 )

is the two-dimensional vector subspace of ℝ3 which is orthogonal to 𝑥. Since
(𝑎, 𝑏, 𝑐) ≠ (0, 0, 0), let us assume that, say, 𝑏 ≠ 0. A basis for 𝑇𝑥 (𝕊2 ) is given by, for
example, 𝑣 = (−𝑏, 𝑎, 0) and 𝑣 = (0, 𝑐, −𝑏). The outward pointing normal vector is
given by 𝑛𝑥 = (𝑎, 𝑏, 𝑐). The boundary orientation of 𝕊2 , is the orientation of 𝑇𝑥 (𝕊2 )
determined by the basis (𝑛𝑥 , 𝑣, 𝑤). This basis is positively oriented in 𝑇𝑥 (ℝ3 ) = ℝ3
⎛𝑎 −𝑏 0 ⎞
if and only if the matrix 𝐴 = ⎜𝑏 𝑎 𝑐 ⎟ has positive determinant, since this is
⎜ ⎟
⎝𝑐 0 −𝑏⎠
the matrix that transforms the standard basis of ℝ3 into the basis (𝑛𝑥 , 𝑣, 𝑤). The
determinant of 𝐴 is

det(𝐴) = −𝑎2 𝑏 − 𝑏3 − 𝑏𝑐 2 = −𝑏(𝑎2 + 𝑏2 + 𝑐 2 ) = −𝑏.

Thus, if 𝑏 < 0, (𝑛𝑥 , 𝑣, 𝑤) is a positively oriented basis of 𝑇𝑥 (𝕊2 ). If 𝑏 > 0, we take

the basis (𝑛𝑥 , 𝑤, 𝑣). And if 𝑏 = 0, we start over with either 𝑎 or 𝑐 replacing 𝑏.

(b) The boundary orientation of 𝕊𝑘 is, at any point 𝑥 ∈ 𝕊𝑘 , given on 𝑇𝑥 (𝕊𝑘 ) by chosing
the ordered basis (𝑛𝑥 , 𝑣1 , … , 𝑣𝑘 ) to be positively oriented where 𝑛𝑥 is the outward
pointing unit normal vector in 𝑇𝑥 (ℝ𝑘+1 ) = ℝ𝑘+1 and (𝑣1 , … , 𝑣𝑘 ) is an ordered basis
of 𝑇𝑥 (𝕊𝑘 ). But since 𝕊𝑘 ⊂ ℝ𝑘+1 is of codimension one, 𝑛𝑥 spans the orthogonal
complement 𝑁𝑥 (𝕊𝑘 , ℝ𝑘+1 ) of 𝑇𝑥 (𝕊𝑘 ) in ℝ𝑘+1 . Hence the orientation of 𝑇𝑥 (𝕊𝑘 )
induced by the direct sum

𝑁𝑥 (𝕊𝑘 , ℝ𝑘+1 ) ⊕ 𝑇𝑥 (𝕊𝑘 ) = 𝑇𝑥 (ℝ𝑘+1 ) = ℝ𝑘+1

equals the orientation of 𝕊𝑘 as the preimage under 𝑔.

Solution (Exercise 14.4) Assume that 𝑑𝑓𝑥0 ∶ 𝑇𝑥0 (𝑋) → 𝑇𝑓 (𝑥0 ) (𝑌 ) preserves orien-
tation at some point 𝑥0 ∈ 𝑋. Since 𝑓 is a diffeomorphism, 𝑑𝑓𝑥 is an isomorphism for
all 𝑥 ∈ 𝑋. Hence det(𝑑𝑓𝑥 ) ≠ 0 for all 𝑥 ∈ 𝑋. In particular, the two disjoint open
subsets 𝑈 ∶= {𝑥 ∈ 𝑋 ∶ det(𝑑𝑓𝑥 ) > 0} and 𝑉 ∶= {𝑥 ∈ 𝑋 ∶ det(𝑑𝑓𝑥 ) < 0} cover 𝑋.
By assumption 𝑥0 ∈ 𝑈 , and hence 𝑈 is nonempty. Since 𝑋 is connected, this implies
𝑈 = 𝑋.

Solution (Exercise 14.5) Let 𝑋 and 𝑍 be transversal submanifolds in 𝑌 and assume

𝑋, 𝑍 and 𝑌 are oriented. Let 𝑖 ∶ 𝑋 → 𝑌 be the inclusion of 𝑋 into 𝑌 . The intersection
𝑋 ∩ 𝑍 equals the preimage 𝑖−1 (𝑍). By Section 14.5, the preimage orientation on 𝑆 ∶=
𝑖−1 (𝑍) is induced, at any 𝑦 ∈ 𝑋 ∩ 𝑍, by the direct sum

𝑁𝑦 (𝑆, 𝑋) ⊕ 𝑇𝑦 (𝑆) = 𝑇𝑦 (𝑋),

where 𝑁𝑦 (𝑆, 𝑋) is the orthogonal complement of 𝑇𝑦 (𝑆) in 𝑇𝑦 (𝑋). The orientation on

Appendix A. Solutions to exercises 425
𝑁𝑦 (𝑆, 𝑋) is induced by the direct sum

𝑑𝑖𝑦 (𝑁𝑦 (𝑆, 𝑋)) ⊕ 𝑇𝑦 (𝑍) = 𝑇𝑦 (𝑌 ),

and the fact that 𝑑(𝑖𝑦 )|𝑁𝑦 (𝑆,𝑋) is an isomorphism onto its image. Since all these vector
spaces are subspaces in 𝑇𝑦 (𝑌 ), and are oriented as subspaces of 𝑇𝑦 (𝑌 ), we can identify
𝑁𝑦 (𝑆, 𝑋) with its image under 𝑑𝑖𝑦 in 𝑇𝑦 (𝑌 ) and can rewrite this equation as

𝑁𝑦 (𝑆, 𝑋) ⊕ 𝑇𝑦 (𝑍) = 𝑇𝑦 (𝑌 ).

Now let 𝑁𝑦 (𝑆, 𝑍) be the orthogonal complement of 𝑇𝑦 (𝑆) in 𝑇𝑦 (𝑍). Then the orien-
tation of 𝑇𝑦 (𝑆) is determined by the direct sum

𝑁𝑦 (𝑆, 𝑋) ⊕ 𝑁𝑦 (𝑆, 𝑍) ⊕ 𝑇𝑦 (𝑆) = 𝑇𝑦 (𝑌 ).

Now if we start with the inclusion 𝑗 ∶ 𝑍 → 𝑌 of 𝑍 in 𝑌 instead, we get that the

orientation of 𝑆 considered as the preimage 𝑗 −1 (𝑋) in 𝑍, is determined by the direct
sum

𝑁𝑦 (𝑆, 𝑍) ⊕ 𝑁𝑦 (𝑆, 𝑋) ⊕ 𝑇𝑦 (𝑆) = 𝑇𝑦 (𝑌 ).

We learned in the first exercise that the signs of the orientations of 𝑁𝑦 (𝑆, 𝑋) ⊕
𝑁𝑦 (𝑆, 𝑍) and 𝑁𝑦 (𝑆, 𝑍) ⊕ 𝑁𝑦 (𝑆, 𝑋) differ by (−1)(dim 𝑁𝑦 (𝑆,𝑋))(dim 𝑁𝑦 (𝑆,𝑍)) . Now it re-
mains to remark that, by definition of the normal spaces as orthogonal complements, we
have

dim 𝑁𝑦 (𝑆, 𝑋) = codim 𝑋 ∩ 𝑍 in 𝑋 = codim 𝑍 in 𝑌 , and

dim 𝑁𝑦 (𝑆, 𝑍) = codim 𝑋 ∩ 𝑍 in 𝑍 = codim 𝑋 in 𝑌 .

Solution (Exercise 14.6) First, we assume that 𝑍 is orientable. At any point 𝑧 ∈ 𝑍, let
(𝑣1 , … , 𝑣𝑛 ) be an oriented basis of 𝑇𝑧 𝑍. Since 𝑇𝑧 𝑍 is a vector subspace of codimension
one in 𝑇𝑧 𝑋 we can choose a single vector 𝑛𝑧 ∈ (𝑇𝑧 𝑍)⟂ = 𝑁𝑧 (𝑍, 𝑋) ⊂ 𝑇𝑧 𝑋 such that
(𝑛𝑧 , 𝑣1 , … , 𝑣𝑛 ) is an oriented basis of 𝑇𝑧 𝑋. Since both 𝑋 and 𝑍 are oriented we can
make this choice of 𝑛𝑧 smoothly for every 𝑧 ∈ 𝑍. Since 𝑛𝑧 has to be non-trivial in order
to be part of a basis we can define a diffeomorphism

𝜑 ∶ 𝑍 × ℝ → 𝑁(𝑍, 𝑋), (𝑧, 𝑡) → (𝑧, 𝑡 ⋅ 𝑛𝑧 ).

Hence 𝑁(𝑍, 𝑋) is a trivial line bundle over 𝑍.

Now we assume that 𝑁(𝑍, 𝑋) is trivial and let 𝜑 ∶ 𝑍×ℝ → 𝑁(𝑍, 𝑋) be a trivializing
diffeomorphism. Let 𝑛𝑧 denote the vector in (𝑇𝑧 𝑍)⟂ = 𝑁𝑧 (𝑍, 𝑋) ⊂ 𝑇𝑧 𝑋 such that
𝜑(𝑧, 1) = (𝑧, 𝑛𝑧 ). We then define a basis (𝑣1 , … , 𝑣𝑛 ) of 𝑇𝑧 𝑍 to be positively oriented if
(𝑛𝑧 , 𝑣1 , … , 𝑣𝑛 ) is a positively oriented basis of 𝑇𝑧 𝑋. Since 𝜑 is a diffeomorphism this
defines a smooth choice of an orientation for all 𝑧 ∈ 𝑍. Hence 𝑍 is orientable.

Solution (Exercise 14.7) (a) Any basis of 𝑉 ×𝑉 consists of the product (𝛼 ×0, 0×𝛽)
426 A.13. Orientation
where 𝛼 and 𝛽 are ordered bases of 𝑉 . The sign of this basis satisfies

sign (𝛼 × 0, 0 × 𝛽) = sign (𝛼) ⋅ sign (𝛽).

Switching the orientation of 𝑉 changes both signs, sign (𝛼) and sign (𝛽). Changing
both signs simultaneously results in multiplying with (−1)2 = 1. Hence the sign
of the basis of 𝑉 × 𝑉 is independent of the choice of orientation for 𝑉 .

(b) Let 𝑋 be an orientable manifold. The orientation of 𝑋 × 𝑋 is given by a smooth

choice of orientation of each tangent space

𝑇(𝑥,𝑦) (𝑋 × 𝑋) = 𝑇𝑥 (𝑋) × 𝑇𝑦 (𝑋).

Changing the orientation of 𝑋 means changing the orientation of both 𝑇𝑥 (𝑋) and
𝑇𝑦 (𝑋). As in the previous point, this means multiplying the sign of any ordered
basis of 𝑇(𝑥,𝑦) (𝑋 × 𝑋) by +1. Hence the product orientation on 𝑋 × 𝑋 is the same
for all choices of orientation on 𝑋.

(c) Let 𝑋 be a smooth manifold which is not orientable. Any Euclidean space ℝ𝑚
is oriented as a manifold by the choice of the standard orientation of the tangent
space 𝑇𝑧 (ℝ𝑚 ) = ℝ𝑚 for any 𝑧 ∈ ℝ𝑚 . For any points 𝑥 ∈ 𝑋 and 𝑧 ∈ ℝ𝑚 , the
tangent space 𝑇(𝑥,𝑧) (𝑋 × ℝ𝑚 ) is just 𝑇𝑥 (𝑋) × ℝ𝑚 . If there was a smooth choice
for an orientation of 𝑋 × ℝ𝑚 , then each tangent space 𝑇𝑥 (𝑋) of 𝑋 would inherit a
smooth choice of orientation from the product 𝑇𝑥 (𝑋) × ℝ𝑚 . This contradicts the
non-orientability of 𝑋.
Now let 𝑌 by any smooth manifold. If 𝑋 × 𝑌 was orientable, then also 𝑋 × 𝑈 for
an open subspace 𝑈 ⊂ 𝑌 which is diffeomorphic to some ℝ𝑚 . But then 𝑋 × ℝ𝑚
would also inherit an orientation which is not possible. Applied to 𝑌 = 𝑋, we see
that 𝑋 × 𝑋 is not orientable.

(d) We can cover 𝑋 by local parametrizations 𝜙 ∶ 𝑈 → 𝑋. The union of the images

of the maps 𝜙 × 𝜙 ∶ 𝑈 × 𝑈 → 𝑋 × 𝑋 is then an open subspace 𝑉 of 𝑋 × 𝑋
which includes Δ. We orient each individual 𝜙(𝑈 ) by requiring the diffeomor-
phism 𝜙 ∶ 𝑈 → 𝜙(𝑈 ) to be orientation preserving. This induces an orientation on
(𝜙 × 𝜙)(𝑈 × 𝑈 ) = 𝜙(𝑈 ) × 𝜙(𝑈 ). As we argued before, changing the orientation on
𝜙(𝑈 ) does not change the orientation on the product 𝜙(𝑈 ) × 𝜙(𝑈 ), since we multi-
ply the signs of all tangent spaces by +1. Hence there is a well-defined orientation
on 𝜙(𝑈 ) × 𝜙(𝑈 ) which is independent on the local parametrizations chosen. Thus
𝑉 which is an open neighborhood of Δ in 𝑋 × 𝑋 is orientable.
However, this does not mean that Δ is always orientable. For, the tangent space to
Δ at any point (𝑥, 𝑥) is the diagonal of 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋). This diagonal is isomorphic
to 𝑇𝑥 (𝑋). Hence changing the orientation of 𝑇𝑥 (𝑋) does change the orientation of
the diagonal in 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋). Thus if we had a smooth choice of orientations for
all diagonals in 𝑇𝑥 (𝑋) × 𝑇𝑥 (𝑋), then we had a smooth choice of orientations for
all 𝑇𝑥 (𝑋). In other words, Δ is orientable if and only if 𝑋 is orientable.
Appendix A. Solutions to exercises 427

Solution (Exercise 14.8) To determine the fiber over 𝑏, we write 𝑧0 = 𝑥0 + 𝑖𝑦0 and
𝑧1 = 𝑥1 + 𝑖𝑦1 . Then we get
1
𝜋(𝑧0 , 𝑧1 ) = (0, 1, 0) ⇒ 2𝑧0 𝑧̄ 1 = 𝑖 and |𝑧0 |2 = |𝑧1 |2 =
2
1
⇒ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 and 𝑥20 + 𝑥21 = .
2
Thus the fiber over 𝑏 has the form
𝑖
𝜋 −1 (𝑏) = {(𝑧0 , 𝑧1 ) ∈ 𝕊3 ∶ 𝑧̄ 1 = }
2𝑧0
= {(𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) ∈ 𝕊3 ∶ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 }.
Let 𝑞 = (𝑥0 , 𝑥1 , 𝑥1 , −𝑥0 ) ∈ 𝜋 −1 (𝑏) be a point in the fiber over 𝑏. Since not both 𝑥0
and 𝑥1 can be zero, we assume that 𝑥0 ≠ 0. The tangent space 𝑇𝑞 𝕊3 is the vector space
𝑇𝑞 𝕊3 = {𝐮 ∈ ℝ4 ∶ 𝐮 ⟂ 𝑞}
⎧ ⎛−𝑥1 ⎞ ⎛−𝑥1 ⎞ ⎛𝑥0 ⎞⎫
⎪ ⟂ ⎜ 𝑥0 ⎟ ⟂ ⎜ 0 ⎟ ⟂ ⎜ 0 ⎟⎪
= span ⎨𝑞1 = ⎜ ⎟ , 𝑞2 = ⎜ 𝑥 ⎟ , 𝑞3 = ⎜ 0 ⎟⎬ .
⎪ ⎜ 0 ⎟ ⎜ 0⎟ ⎜ ⎟⎪
⎩ ⎝ 0 ⎠ ⎝ 0 ⎠ ⎝𝑥0 ⎠⎭
The orientation of 𝑇𝑞 𝕊3 as a boundary of the unit ball is such that the outward point-
ing vector 𝑞 together with the basis vectors of 𝑇𝑞 𝕊3 form a positively oriented basis of
ℝ4 . The matrix expressing the basis (𝑞, 𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) in the standard basis of ℝ4 is

⎛ 𝑥0 −𝑥1 −𝑥1 𝑥0 ⎞
⎜ 𝑥1 𝑥0 0 0⎟
⎜𝑥 0 𝑥 0 ⎟.
⎜ 1 0 ⎟
⎝−𝑥0 0 0 𝑥0 ⎠
The determinant of this matrix is
2𝑥40 + 2𝑥20 𝑥21 = 2𝑥20 (𝑥20 + 𝑥21 ) = 𝑥20 > 0.
In particular, it is positive and the basis (𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) is a positively oriented basis of 𝑇𝑞 𝕊3 .
The tangent space 𝑇𝑞 𝜋 −1 (𝑏) equals the kernel of 𝑑 𝜋̃𝑞 . In a previous exercise, we
computed this map as represented by the matrix
⎛𝑥1 −𝑥0 𝑥0 𝑥1 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ ⎜𝑥0 𝑥1 𝑥1 −𝑥0 ⎟ .
⎜ ⎟
⎝𝑥0 𝑥1 −𝑥1 𝑥0 ⎠

⎛−𝑥1 ⎞
⎜𝑥 ⎟
The kernel of this map is the span of the vector 𝑞0⟂ = ⎜ 0 ⎟. The normal space
⎜ 𝑥0 ⎟
⎝ 𝑥1 ⎠
𝑁𝑞 (𝜋 −1 (𝑏); 𝕊3 ) ⊂ 𝑇𝑞 𝕊3 of vectors which are orthogonal to 𝑇𝑞 𝜋 −1 (𝑏) is the span of
(𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ ). The map 𝑑 𝜋̃𝑞 sends 𝑞1⟂ − 𝑞2⟂ and 𝑞3⟂ to, respectively,

⎛ −2𝑥20 ⎞ ⎛−2𝑥0 𝑥1 ⎞
⟂ ⟂
𝑑 𝜋̃𝑞 (𝑞1 − 𝑞2 ) = 2 ⎜ 0 ⎟ , 𝑑 𝜋̃𝑞 (𝑞3 ) = 2 ⎜ 0 ⎟ .
⟂
⎜ ⎟ ⎜ 2 ⎟
⎝−2𝑥0 𝑥1 ⎠ ⎝ 2𝑥0 ⎠
428 A.13. Orientation

These two vectors form a basis (𝑑 𝜋̃𝑞 (𝑞1⟂ − 𝑞2⟂ ), 𝑑 𝜋̃𝑞 (𝑞3⟂ )) of 𝑇𝑏 𝕊2 . We need to check the
orientation of this basis.
The tangent space 𝑇𝑏 𝕊2 has a basis (𝐞31 , 𝐞33 ) as a subspace in ℝ3 . This basis is negatively
oriented, since, together with the outward pointing vector 𝑏 = 𝐞32 , the basis (𝐞32 , 𝐞31 , 𝐞33 )
is a negatively oriented basis of ℝ3 . For we need to make one permutation to get the
standard basis which leads to multiplying the sign with −1. The matrix 𝐵 which expresses
(𝑑 𝜋̃𝑞 (𝑞1⟂ − 𝑞2⟂ ), 𝑑 𝜋̃𝑞 (𝑞3⟂ )) in terms of the basis (𝐞31 , 𝐞33 ) is given by
( )
−𝑥20 −𝑥0 𝑥1
𝐵 =4⋅ .
−𝑥0 𝑥1 𝑥20

We see that det 𝐵 = 16(−𝑥40 − 𝑥20 𝑥21 ) = −16𝑥20 (𝑥20 + 𝑥21 ) = −8𝑥20 < 0 is negative. Hence
the basis (𝑑 𝜋̃𝑞 (𝑞1⟂ − 𝑞2⟂ ), 𝑑 𝜋̃𝑞 (𝑞3⟂ )) is a negatively oriented basis of 𝑇𝑏 𝕊2 . This defines an
orientation on the normal space 𝑁𝑞 (𝜋 −1 (𝑏); 𝕊3 ) by declaring the orientation of the basis
(𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ ) to be negative.

Finally, the orientation of 𝑇𝑞 𝜋 −1 (𝑏) is such that the direct sum

𝑁𝑞 (𝜋 −1 (𝑏); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑏) = 𝑇𝑞 𝕊3

induces the given orientation on 𝑇𝑞 𝕊3 . We check this by looking at the basis

(𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ , 𝑞0⟂ ) of 𝑁𝑞 (𝜋 −1 (𝑏); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑏). The transition matrix from the basis
(𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) to the basis (𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ , 𝑞0⟂ ) is given by

⎛1 0 1 ⎞
⎜−1 0 1 ⎟.
⎜ ⎟
⎝ 0 1 𝑥1 ∕𝑥0 ⎠

The determinant of this matrix is −2. In particular, it is negative. Since we checked that
the basis (𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) is a negatively oriented basis, we see that the orientation of the basis
(𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ , 𝑞0⟂ ) of 𝑇𝑞 𝕊3 is positive. Since the sign of (𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ ) is positive as a basis
of 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ), we need that the basis 𝑞0⟂ also has positive sign. Hence the vector 𝑞0⟂
provides a positively oriented basis of 𝑇𝑞 𝜋 −1 (𝑏).
Appendix A. Solutions to exercises 429
A.14 The Brouwer Degree

A.14.1 Brouwer degree

Solution (Exercise 15.1) (a) For 1 ≤ 𝑖 ≤ 𝑘 + 1, let 𝑟𝑖 be the reflection map on the
𝑖th coordinate:

𝑟𝑖 ∶ 𝕊𝑘 → 𝕊𝑘 , (𝑥1 , … , 𝑥𝑘+1 ) → (𝑥1 , … , −𝑥𝑖 , … , 𝑥𝑘+1 ).

The antipodal map is equal to the composition of reflections 𝑟1 ◦𝑟2 ◦ ⋯ ◦𝑟𝑘+1 . Each
reflection 𝑟𝑖 is a diffeomorphism which reverses the orientation on 𝕊𝑘 . The com-
position of two such reflections 𝑟𝑖 ◦𝑟𝑖+1 , however, is then a diffeomorphism which
preserves the orientation on 𝕊𝑘 . Hence, if 𝑘 = 2𝑛 is even, 𝑎 = 𝑟1 ◦𝑟2 ◦ ⋯ ◦𝑟2𝑛+1
has degree −1. Whereas if 𝑘 = 2𝑛 − 1 is odd, 𝑎 = 𝑟1 ◦𝑟2 ◦ ⋯ ◦𝑟2𝑛 has degree +1.
In other words, deg(𝑎) = (−1)𝑘+1 .

(b) As pointed out, we know that 𝑎 is homotopic to the identity if 𝑘 is odd. By the
previous point, deg(𝑎) = −1 if 𝑘 is even. Since deg is homotopy invariant, deg(𝑎) =
−1 ≠ 1 = deg(1id) implies that the antipodal map is not homotopic to the identity
if 𝑘 is even.

Solution (Exercise 15.2) Recall that we have proven the existence part for 𝑘 odd
before:
If 𝑘 is odd, then 𝑘 + 1 is even and we can define the map
𝑠 ∶ 𝕊𝑘 → ℝ𝑘+1 , (𝑥1 , … , 𝑥𝑘+1 ) → (−𝑥2 , 𝑥1 , −𝑥3 , 𝑥4 , … , −𝑥𝑘+1 , 𝑥𝑘 ).
This map can be extended to a linear map ℝ𝑘+1 → ℝ𝑘+1 and therefore 𝑠 is smooth. For
each 𝑥 ∈ 𝕊𝑘 , 𝑠(𝑥) is nonzero and satisfies 𝑥 ⟂ 𝑠(𝑥). Thus 𝑠(𝑥) is a tangent vector at 𝑥,
i.e. 𝑠(𝑥) ∈ 𝑇𝑥 (𝕊𝑘 ) ⧵ {0}. Hence
𝜎 ∶ 𝕊𝑘 → 𝑇 (𝕊𝑘 ), 𝜎(𝑥) ∶= (𝑥, 𝑠(𝑥))
is the desired non-vanishing vector field on 𝕊𝑘 .

Recall that we also have shown that if 𝕊𝑘 has a vector field which has no zeros, then
its antipodal map 𝑥 → −𝑥 is homotopic to the identity:
Given a vector field 𝜎 ∶ 𝕊𝑘 → 𝑇 (𝕊𝑘 ) which has no zeros. Let 𝜎(𝑥) = (𝑥, 𝑣(𝑥)). Since
𝑣(𝑥) ≠ 0 for every 𝑥 ∈ 𝕊𝑘 , we can define a new vector field by
𝑣(𝑥)
𝑥 → 𝑤(𝑥) = .
|𝑣(𝑥)|
430 A.14. The Brouwer Degree
By replacing 𝑠 with this new non-vanishing vector field, we can assume |𝑣(𝑥)| = 1.
Hence we can assume 𝑣(𝑥) ∈ 𝕊𝑘 and 𝑣(𝑥) ⋅ 𝑥 = 0 for every 𝑥 ∈ 𝕊𝑘 .
Now we define the map

𝐹 ∶ 𝕊𝑘 × [0, 1] → 𝕊𝑘 , (𝑥, 𝑡) → cos(𝜋𝑡)𝑥 + sin(𝜋𝑡)𝑣(𝑥).

We check that 𝐹 (𝑥, 𝑡) is in fact an element in 𝕊𝑘 for every 𝑥 ∈ 𝕊𝑘 :

𝐹 (𝑥, 𝑡) ⋅ 𝐹 (𝑥, 𝑡) = (cos(𝜋𝑡)𝑥 + sin(𝜋𝑡)𝑠(𝑥)) ⋅ (cos(𝜋𝑡)𝑥 + sin(𝜋𝑡)𝑣(𝑥))

= cos2 (𝜋𝑡)(𝑥 ⋅ 𝑥) + 2 cos(𝜋𝑡) sin(𝜋𝑡)(𝑥 ⋅ 𝑣(𝑥)) + sin2 (𝜋𝑡)(𝑠(𝑥) ⋅ 𝑣(𝑥))
= cos2 (𝜋𝑡) + sin2 (𝜋𝑡)
=1

where we use 𝑥⋅𝑥 = 1 = 𝑣(𝑥)⋅𝑣(𝑥) and 𝑥⋅𝑣(𝑥) = 0. Thus 𝐹 (𝑥, 𝑡) is a vector of norm 1 for
every 𝑥 and every 𝑡. Moreover, 𝐹 is a smooth map with 𝐹 (𝑥, 0) = 𝑥 and 𝐹 (𝑥, 1) = −𝑥,
i.e. 𝐹 is a smooth homotopy from the identity to the antipodal map on 𝕊𝑘 .
Hence, by homotopy invariance of deg, if 𝕊𝑘 has a vector field which has no zeros,
then deg(𝑎) = deg(id) = 1. By the previous exercise, since deg(𝑎) = (−1)𝑘+1 and hence
𝑘 must be even.
Note that we would not have been able to make this conclusion with the mod 2-degree.
For in ℤ∕2, we cannot distinguish between 1 and −1.

Solution (Exercise 15.3) We begin as in the solution to Exercise 11.4. We use the
homotopy from 𝑝0 (𝑧) = 𝑧𝑚 to 𝑝1 (𝑧) = 𝑝(𝑧) defined by

𝑝𝑡 (𝑧) = 𝑡𝑝(𝑧) + (1 − 𝑡)𝑧𝑚 = 𝑧𝑚 + 𝑡(𝑎1 𝑧𝑚−1 + ⋯ + 𝑎𝑚 ).

We then observe that, if 𝑊 is a closed ball around the origin in ℂ with sufficiently
large radius, none of the 𝑝𝑡 has a zero on 𝜕𝑊 . Hence the homotopy
𝑝𝑡
∶ 𝜕𝑊 → 𝕊1
|𝑝𝑡 |

is defined for all 𝑡 ∈ [0, 1]. Thus

( ) ( )
𝑝 𝑝0
deg = deg .
|𝑝| |𝑝0 |

Since 𝑝0 (𝑧) = 𝑧𝑚 , the degree of 𝑝0 ∕|𝑝0 | is the same as deg(𝑧𝑚 ) = 𝑚, that is

( )
𝑝
deg = 𝑚.
|𝑝|

Thus, if 𝑚 > 0, 𝑝∕|𝑝| does not extend to all of 𝑊 , since otherwise its degree had to
be zero. Hence 𝑝 must have a zero inside 𝑊 .a
a
Note that the final argument was not available for deg2 if 𝑚 is even.
Appendix A. Solutions to exercises 431

Solution (Exercise 15.4) Let us try to set up the argument we used in Exercise 15.3:
Let
𝑝(𝑥) = 𝑥𝑚 + 𝑎1 𝑥𝑚−1 + ⋯ + 𝑎𝑚
be a monic real polynomial. Define a homotopy from 𝑝0 (𝑥) = 𝑥𝑚 to 𝑝1 (𝑥) = 𝑝(𝑥) by

𝑝𝑡 (𝑥) = 𝑡𝑝(𝑥) + (1 − 𝑡)𝑥𝑚 = 𝑥𝑚 + 𝑡(𝑎1 𝑥𝑚−1 + ⋯ + 𝑎𝑚 ).

If 𝑊 = [−𝑎, 𝑎] is a large enough closed interval in ℝ containing the origin, none of

the 𝑝𝑡 has a zero on 𝜕𝑊 = {−𝑎, 𝑎}. Hence the homotopy
𝑝𝑡
∶ 𝜕𝑊 = {−𝑎, 𝑎} → {−1, +1}
|𝑝𝑡 |

is defined for all 𝑡 ∈ [0, 1]. Thus

( ) ( )
𝑝 𝑝0
deg = deg .
|𝑝| |𝑝0 |
𝑝
However, for each 𝑡 ∈ [0, 1], the map |𝑝𝑡 | (𝑥) is constant with value either −1 or +1.
𝑡
A constant map has degree 0, since all points not(in the ) image of the map are regular
𝑝
values and have empty fibers. Hence we get deg |𝑝| = 0. But this does not imply
𝑝
that we cannot extend |𝑝| to all of 𝑊 . In particular, we do not get a contradiction to the
assumption that 𝑝 does not have a zero in 𝑊 .

Solution (Exercise 15.5) (a) First, 𝑔 is in fact a map 𝕊1 → 𝜕𝔻0 , since any point 𝑧
with |𝑧| = 1 is sent to point with
|𝑔(𝑧) − 𝑧0 | = 𝑧0 + 𝑟𝑧 − 𝑧0 = |𝑟𝑧| = 𝑟 since |𝑧| = 1.
Moreover, 𝑔 is smooth and has an inverse given by
𝑧 − 𝑧0
𝑔 −1 ∶ 𝜕𝔻0 → 𝕊1 , 𝑧 → .
|𝑧 − 𝑧0 |
Note that 𝑔 −1 is also smooth, since both taking norms and dividing by |𝑧 − 𝑧0 | are
smooth operations in ℂ ⧵ {𝑧0 }.
Hence we know that 𝑔 either preserves or reverses orientations. To check that 𝑔
preserves orientations, it suffices to note that the derivative of 𝑔 at any point, for
example 𝑑𝑔1 , is given by multiplication with 𝑟 > 0.
(b) Let 𝑦 ∈ 𝕊1 . Since 𝑔 is a diffeomorphism, 𝑦 is a regular value for 𝑝 if and only if it
is a regular value for 𝑝◦𝑔. Moreover, 𝑔 defines a bijection between the finite sets
{𝑧 ∈ 𝜕𝐷0 |𝑝(𝑧) = 𝑦} and {𝑧 ∈ 𝕊1 |𝑝(𝑔(𝑧)) = 𝑦}. Since 𝑔 preserves orientations,
the orientation numbers at each point in these sets agree. Hence the degrees of the
above maps are the same.
(c) Since 𝑝(𝑧) ≠ 0 for all 𝑧 ∈ 𝔻0 ⧵ {𝑧0 } by our choice of 𝔻0 , we have |𝑞(𝑧0 + 𝑡𝑟𝑧)| ≠ 0
for all |𝑧| = 1. Hence we have a well-defined smooth map
𝑧𝑚 𝑞(𝑧0 + 𝑡𝑟𝑧)
𝐻 ∶ 𝕊1 × [0, 1] → 𝕊1 , (𝑧, 𝑡) → .
|𝑞(𝑧0 + 𝑡𝑟𝑧)|
432 A.14. The Brouwer Degree

Write ℎ𝑡 (𝑧) = 𝐻(𝑧, 𝑡). For all 𝑧 ∈ 𝕊1 , we have

𝑧𝑚 𝑞(𝑧0 ) 𝑞(𝑧0 ) 𝑚
ℎ0 (𝑧) = = ⋅𝑧
𝑞(𝑧0 )| 𝑞(𝑧0 )|

and
𝑧𝑚 𝑞(𝑧0 + 𝑟𝑧) (𝑟𝑧)𝑚 𝑞(𝑧0 + 𝑟𝑧)
ℎ1 (𝑧) = = 𝑚
|𝑞(𝑧0 + 𝑟𝑧)| 𝑟 |𝑞(𝑧0 + 𝑟𝑧)|
𝑚
(𝑧0 + 𝑟𝑧 − 𝑧0 ) 𝑞(𝑧0 + 𝑟𝑧)
=
|𝑧0 + 𝑟𝑧 − 𝑧0 |𝑚 |𝑞(𝑧0 + 𝑟𝑧)|
𝑝(𝑔(𝑧))
= .
|𝑝(𝑔(𝑧))|
Hence ℎ𝑡 is the desired homotopy.

(d) We know from the calculation in Example 15.10 that the map 𝕊1 → 𝕊1 , 𝑧 → 𝑧𝑚 ,
has degree 𝑚. Since multiplying with a constant 𝑐 does not change the degree,
we also have deg(𝑧 → 𝑐 ⋅ 𝑧𝑚 ) = deg(ℎ0 ) = 𝑚. Since degrees are homotopy
invariant, the previous point shows deg(ℎ1 ) = deg(ℎ0 ). As we observed before,
deg(𝑝∕|𝑝|) = deg(ℎ1 ). Hence we can conclude deg(𝑝∕|𝑝|) = 𝑚.

Solution (Exercise 15.6) Let 𝑧 be a regular value for 𝑔. If 𝑧 is not in the image of 𝑔,
then deg(𝑔) and deg(𝑔◦𝑓 ) both vanish and the claim is true. So let us assume that 𝑧 is in
the image of 𝑔. We showed in a previous exercise that 𝑧 then is a regular value for 𝑔◦𝑓
if and only if every 𝑦 ∈ 𝑔 −1 (𝑧) is a regular value for 𝑓 . Hence we can use 𝑧 to compute
deg(𝑔◦𝑓 ) if and only if we can use 𝑦 ∈ 𝑔 −1 (𝑧) to compute deg(𝑓 ). If 𝑧 is not in the image
of 𝑔◦𝑓 , then 𝑦 is not in the image of 𝑓 . In this case both deg(𝑔◦𝑓 ) and deg(𝑓 ) vanish
and the claim is true.
So assume there are points 𝑥 ∈ 𝑋 with 𝑔(𝑓 (𝑥)) = 𝑧. Given such an 𝑥, the chain rule
gives us
𝑑(𝑔◦𝑓 )𝑥 = 𝑑𝑔𝑓 (𝑥) ◦𝑑𝑓𝑥
Since 𝑥 and 𝑓 (𝑥) are regular points by assumption, all three derivatives in this equa-
tion are isomorphisms. Recall that we equip 𝑥 with orientation number +1 if 𝑑𝑓𝑥 pre-
serves orientation and with orientation number −1 if 𝑑𝑓𝑥 reverses orientation. Moreover,
whether 𝑑𝑓𝑥 preserves or reverses orientation is determined by the sign of its determi-
nant. Similarly, for 𝑓 (𝑥) with respect to 𝑑𝑔𝑓 (𝑥) and 𝑥 with respect to 𝑑(𝑔◦𝑓 )𝑥 . By the
chain rule and since the determinant is a multiplicative function, we get the formula for
signs:
sign (det(𝑑(𝑔◦𝑓 )𝑥 )) = sign (det(𝑑𝑔𝑓 (𝑥) )) ⋅ sign (det(𝑑𝑓𝑥 )).
Appendix A. Solutions to exercises 433
Since the degree is the sum of the orientation numbers at all points in the fiber, we get
∑
deg(𝑔◦𝑓 ) = sign (det(𝑑(𝑔◦𝑓 )𝑥 ))
𝑥∈(𝑔◦𝑓 )−1 (𝑧)
∑
= [sign (det(𝑑𝑔𝑓 (𝑥) )) ⋅ sign (det(𝑑𝑓𝑥 ))]
𝑥∈(𝑔◦𝑓 )−1 (𝑧)
( ) ( )
∑ ∑
= sign (det(𝑑𝑔𝑦 )) ⋅ sign (det(𝑑𝑓𝑥 ))
𝑦∈𝑔 −1 (𝑧) 𝑥∈𝑓 −1 (𝑦)

= deg(𝑔) ⋅ deg(𝑓 ).

Note that we used here that we can compute deg 𝑓 using any regular value 𝑦 ∈ 𝑌 for 𝑓
with 𝑔(𝑦) = 𝑧.

Solution (Exercise 15.7) Let 𝑎 ∶ 𝕊𝑘 → 𝕊𝑘 denote the antipodal map. We assume

that 𝑓 had no fixed point. Then 𝑥 and 𝑎(𝑓 (𝑥)) are never antipodal points and we get
|𝑥 − (𝑎◦𝑓 )(𝑥)| < 2 for all 𝑥 ∈ 𝕊𝑘 . By Exercise 8.7, this implies that id𝕊𝑘 and 𝑎◦𝑓 are
homotopic. By Exercise 15.6, we have deg(𝑎◦𝑓 ) = deg(𝑎) ⋅ deg(𝑓 ) and by Exercise 15.1
we know deg(𝑎) = (−1)𝑘+1 . Thus we obtain

1 = deg(id) = (−1)𝑘+1 deg(𝑓 ),

which implies deg(𝑓 ) = (−1)𝑘+1 . Thus, if 𝑓 does not have fixed point, then deg(𝑓 ) =
(−1)𝑘+1 . This proves the claim.

Solution (Exercise 15.8) At every point [𝑥] ∈ ℝP𝑘 , the fiber under 𝑞 consists of two
antipodal points 𝑞 −1 ([𝑥]) = {𝑥, −𝑥} ⊂ 𝕊𝑘 . Let 𝑎 ∶ 𝕊𝑘 → 𝕊𝑘 be the antipodal map. We
have 𝑞(𝑥) = 𝑞(−𝑥) for all 𝑥 ∈ 𝕊𝑘 , i.e., we have 𝑞 = 𝑞◦𝑎. This implies 𝑑𝑞−𝑥 ◦𝑑𝑎𝑥 = 𝑑𝑞𝑥
for all 𝑥 ∈ 𝕊𝑘 . In particular, this implies det(𝑑𝑞−𝑥 ) ⋅ det(𝑑𝑎𝑥 ) = det(𝑑𝑞𝑥 ). If 𝑘 is odd,
then 𝑎 preserves orientation, i.e, det(𝑑𝑎𝑥 ) > 0. Thus, det(𝑑𝑞−𝑥 ) and det(𝑑𝑞𝑥 ) have the
same sign. This implies that 𝑥 and −𝑥 contribute to the degree with the same orientation
number. Hence deg(𝑞) is either +2 or −2. This also shows that, if det(𝑑𝑞𝑥 ) > 0, i.e., if 𝑞
preserves orientations, then deg(𝑞) = +2.

Solution (Exercise 15.9) (a) If 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 sends no pair of antipodal points to

antipodal points, then the straight line in ℝ𝑘+1 between 𝑓 (𝑥) and 𝑓 (−𝑥) does not
go through the origin. Thus the map

(1 − 𝑡)𝑓 (𝑥) + 𝑡𝑓 (−𝑥)

𝐻 ∶ 𝕊𝑘 × [0, 1] → 𝕊𝑘 , 𝑥 →
|(1 − 𝑡)𝑓 (𝑥) + 𝑡𝑓 (−𝑥)|
is well-defined and smooth for every 𝑥 and 𝑡. Hence 𝐻 defines a homotopy between
𝑓 and 𝑓 ◦𝑎. For 𝑡 = 1∕2, we have 𝐻(𝑥, 1∕2) = 12 (𝑓 (𝑥) + 𝑓 (−𝑥)) = 𝑔(𝑥) for all
𝑥 ∈ 𝕊𝑘 . Thus 𝑓 and 𝑓 ◦𝑎 are both smoothly homotopic to the map 𝑔 as well. The
invariance of the degree under homotopy implies deg(𝑓 ) = deg(𝑓 ◦𝑎) = deg(𝑔).
434 A.14. The Brouwer Degree

(b) The previous point gives us, using Exercise 15.6,

deg(𝑔) = deg(𝑓 ◦𝑎) = deg(𝑓 ) ⋅ deg(𝑎) = deg(𝑓 ) = deg(𝑔).

Since we assume that 𝑘 is even, we have deg(𝑎) = −1 by Exercise 15.1. Thus we

get deg(𝑔) = − deg(𝑔) which implies deg(𝑔) = 0.

(c) Since 𝑘 is odd, the degree of the quotient map 𝑞 ∶ 𝕊𝑘 → ℝP𝑘 is defined. Since 𝑔
satisfies 𝑔(𝑥) = 𝑔(−𝑥) and 𝑞 is a quotient map, 𝑔 can be written as a composition
𝑞 [𝑔]
𝕊𝑘 ←←→
← ℝP𝑘 ←←←←←→
← 𝕊𝑘 .

By Exercise 15.8, we have deg(𝑞) = 2. Thus, by Exercise 15.6, we get

deg(𝑔) = deg(𝑞) ⋅ deg([𝑔]) = 2 ⋅ deg([𝑔]).

This shows that deg(𝑔) is even.

(d) The previous points show that the assumption that 𝑓 ∶ 𝕊𝑘 → 𝕊𝑘 sends no pair of
antipodal points to antipodal points implies that deg(𝑓 ) = deg(𝑔) must be even.
Since deg(𝑓 ) is odd, there must be at least one pair of antipodal points 𝑥0 , −𝑥0
such that 𝑓 (−𝑥0 ) = −𝑓 (𝑥0 ). This proves the claim.
Appendix A. Solutions to exercises 435
A.15 Linking Number and the Hopf Invariant

A.15.1 Linking number and the Hopf invariant

Solution (Exercise 15.10) (a) Let 𝑧 ∈ 𝕊𝑘 be a regular value for 𝜆 and let (𝑥, 𝑦) ∈
𝑋 × 𝑌 with 𝜆(𝑥, 𝑦) = 𝑧. Consider the derivative

𝑑𝜆(𝑥,𝑦) ∶ 𝑇𝑥 (𝑋) × 𝑇𝑦 (𝑌 ) = 𝑇(𝑥,𝑦) (𝑋 × 𝑌 ) → 𝑇𝑧 (𝕊𝑘 ).

Then the degree of 𝜆 is defined as the sum

∑
deg(𝜆) = sign (𝑑𝜆(𝑥,𝑦) ).
(𝑥,𝑦)∈𝜆−1 (𝑧)

As we learned in Section 14.2.1, switching the order of the factors in the domain
corresponds to making dim 𝑋 ⋅dim 𝑌 many flips of basis vectors. Each flip requires
to multiply orientation numbers with a factor (−1). Hence the map 𝑠 ∶ 𝑋 × 𝑌 →
𝑌 × 𝑋 induces multiplication by (−1)𝑚𝑛 on the degree. In addition, we change the
value of 𝜆 by sending 𝑥 − 𝑦 to 𝑦 − 𝑥 = −(𝑥 − 𝑦). Hence we compose with the
antipodal map 𝑎 on 𝕊𝑘 . This map has degree (−1)𝑘+1 as we learned in a previous
exercise. Still using that we showed in an exercise that deg sends composition of
maps to products, this implies in total

𝐿(𝑌 , 𝑋) = deg(𝑎◦𝜆◦𝑠) = (−1)𝑘+1 ⋅ deg(𝜆) ⋅ (−1)𝑚𝑛

= (−1)(𝑚+1)(𝑛+1) deg(𝜆) = (−1)(𝑚+1)(𝑛+1) 𝐿(𝑋, 𝑌 )

where we use 𝑘 = 𝑚 + 𝑛.

(b) Since 𝑌 does not have a boundary, the product 𝑊 × 𝑌 is a smooth manifold with
boundary 𝜕(𝑊 × 𝑌 ) = 𝜕𝑊 × 𝑌 = 𝑋 × 𝑌 . Since 𝑊 and 𝑌 are disjoint, 𝜆 extends
to a smooth map on 𝑊 × 𝑌 . Then the Boundary Theorem for degrees implies that
deg(𝜆) = 0.

Solution (Exercise 15.11) (a) Since taking preimages preserves codimensions, the
dimension of 𝑓 −1 (𝑤) and 𝑓 −1 (𝑧) for any regular values 𝑤 and 𝑧 for 𝑓 is given by

dim 𝑓 −1 (𝑤) = dim 𝑓 −1 (𝑧) = dim 𝕊2𝑛−1 − dim 𝕊𝑛 = 𝑛 − 1.

Hence, if 𝑛 is odd, then dim 𝑓 −1 (𝑤) + 1 = dim 𝑓 −1 (𝑧) = 𝑛 are odd and hence
𝐻(𝑓 ) = 𝐿(𝑓 −1 (𝑤), 𝑓 −1 (𝑧)) = −𝐿(𝑓 −1 (𝑧), 𝑓 −1 (𝑤)) = −𝐻(𝑓 ) by a previous exer-
cise. Thus we must have 𝐻(𝑓 ) = 0.

(b) By Sard’s Theorem we can find a point 𝑎 ∈ 𝕊𝑛 which is a regular value for both 𝑔
and 𝑔◦𝑓 . The fiber 𝑔 −1 (𝑎) consists of a finite number of points, say 𝑎1 , … , 𝑎𝑟 in
𝕊𝑛 . We can assume that these points are ordered such that the orientation numbers
of 𝑔 at 𝑎1 , … , 𝑎𝑝 are positive, while the orientation numbers of 𝑔 at 𝑎𝑝+1 , … , 𝑎𝑟
are negative. Then we have deg(𝑔) = 2𝑝 − 𝑟. The fiber 𝑓 −1 (𝑔 −1 (𝑎)) then consists
436 A.15. Linking Number and the Hopf Invariant

of a disjoint union of the oriented submanifolds 𝑓 −1 (𝑎𝑖 ) ⊂ 𝕊2𝑛−1 :

𝑓 −1 (𝑔 −1 (𝑎)) = 𝑓 −1 (𝑎1 ) ⊔ … ⊔ 𝑓 −1 (𝑎𝑝 ) ⊔ (−𝑓 −1 (𝑎𝑝+1 )) ⊔ … ⊔ (−𝑓 −1 (𝑎𝑟 )).

Similarly, we can choose 𝑏 such that the fiber 𝑓 −1 (𝑔 −1 (𝑏)) consists of a disjoint
union of the submanifolds 𝑓 −1 (𝑏𝑗 ) ⊂ 𝕊2𝑛−1 for 𝑗 = 1, … , 𝑠 and such that the
𝑓 −1 (𝑎𝑖 ) and 𝑓 −1 (𝑏𝑗 are mutually disjoint. Again we order these points are ordered
such that the orientation numbers of 𝑔 at 𝑏1 , … , 𝑏𝑠 are positive, while the orienta-
tion numbers of 𝑔 at 𝑏𝑞+1 , … , 𝑏𝑠 are negative. Then we have deg(𝑔) = 2𝑞 − 𝑠. The
fiber 𝑓 −1 (𝑔 −1 (𝑏)) then consists of a disjoint union of the oriented submanifolds
𝑓 −1 (𝑎𝑖 ) ⊂ 𝕊2𝑛−1 :

𝑓 −1 (𝑔 −1 (𝑏)) = 𝑓 −1 (𝑏1 ) ⊔ … ⊔ 𝑓 −1 (𝑏𝑞 ) ⊔ (−𝑓 −1 (𝑏𝑞+1 ) ⊔ … ⊔ (−𝑓 −1 (𝑏𝑠 ).

Since orientations behave multiplicatively on products of manifolds, the product

fiber 𝑓 −1 (𝑔 −1 (𝑎)) × 𝑓 −1 (𝑔 −1 (𝑏)) is the disjoint union of the submanifolds in 𝕊2𝑛−1

𝑓 −1 (𝑔 −1 (𝑎)) × 𝑓 −1 (𝑔 −1 (𝑏))
= ⊔𝑖,𝑗 (±𝑓 −1 (𝑎𝑖 )) × (±𝑓 −1 (𝑏𝑗 ))
= ⊔𝑝,𝑞
𝑖=1,𝑗=1
(𝑓 −1 (𝑎𝑖 ) × 𝑓 −1 (𝑏𝑗 )) ⊔ ⊔𝑟,𝑞
𝑖=𝑝+1,𝑗=1
(−(𝑓 −1 (𝑎𝑖 ) × 𝑓 −1 (𝑏𝑗 )))
⊔ ⊔𝑝,𝑠
𝑖=1,𝑗=𝑞+1
(−(𝑓 −1 (𝑎𝑖 ) × 𝑓 −1 (𝑏𝑗 ))) ⊔ ⊔𝑟,𝑠
𝑖=𝑝+1,𝑗=𝑞+1
(𝑓 −1 (𝑎𝑖 ) × 𝑓 −1 (𝑏𝑗 )).

Now we write 𝜆 for the map

𝑥−𝑦
(𝑔◦𝑓 )−1 (𝑎) × (𝑔◦𝑓 )−1 (𝑏) → 𝕊2𝑛−2 , (𝑥, 𝑦) →
|𝑥 − 𝑦|
and we write 𝜆𝑖𝑗 for the map
𝑥−𝑦
𝑓 −1 (𝑎𝑖 ) × 𝑓 −1 (𝑏𝑗 ) → 𝕊2𝑛−2 , (𝑥, 𝑦) → .
|𝑥 − 𝑦|
Note that 𝐻(𝑓 ) = deg(𝜆𝑖𝑗 ) for every pair 𝑖, 𝑗, since 𝐻(𝑓 ) does not depend on
the choice of regular values. Since the degree is additive on oriented connected
components, the above decomposition implies

𝐻(𝑔◦𝑓 ) = deg(𝜆)
∑
𝑝,𝑞
∑
𝑟,𝑞
∑
𝑝,𝑠
∑
𝑟,𝑠
= deg(𝜆𝑖𝑗 ) − deg(𝜆𝑖𝑗 ) − deg(𝜆𝑖𝑗 ) + deg(𝜆𝑖𝑗 )
𝑖=1,𝑗=1 𝑖=𝑝+1,𝑗=1 𝑖=1,𝑗=𝑞+1 𝑖=𝑝+1,𝑗=𝑞+1

= 𝐻(𝑓 ) ⋅ [𝑝𝑞 + (𝑝 − deg(𝑔))(𝑞 − deg(𝑔)) − (𝑞(𝑝 − deg(𝑔))) − (𝑝(𝑞 − deg(𝑔)))]

= 𝐻(𝑓 ) ⋅ (deg(𝑔))2 .

Solution (Exercise 15.12) (a) For (𝑧0 , 𝑧1 ) ∈ 𝕊3 , we need to check 𝜋(𝑧0 , 𝑧1 ) ∈ 𝕊2 .

To do this, we recall that |𝑧| = 𝑧𝑧̄ and 𝑧̄ = 𝑧 for any complex number 𝑧. Now we
compute:
Appendix A. Solutions to exercises 437

( )2
(2𝑧0 𝑧̄ 1 ) ⋅ (2𝑧̄ 0 𝑧1 ) + |𝑧0 |2 − |𝑧1 |2
= 4|𝑧0 |2 |𝑧1 |2 + |𝑧0 |4 − 2|𝑧0 |2 |𝑧1 |2 + |𝑧1 |4
( )2
= |𝑧0 |2 + |𝑧1 |2
=1

where the final step uses that (𝑧0 , 𝑧1 ) ∈ 𝕊3 .

(b) First we assume 𝜋(𝑧0 , 𝑧1 ) = 𝜋(𝑤0 , 𝑤1 ): Then we get

( ) ( )
2𝑤0 𝑤̄ 1 , |𝑤0 |2 − |𝑤1 |2 = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2
⇐⇒ 𝑤0 𝑤̄ 1 = 𝑧0 𝑧̄ 1 and |𝑤0 |2 − |𝑤1 |2 = |𝑧0 |2 − |𝑧1 |2 .

Remembering that neither of the numbers can be zero, the left hand equation gives
𝑤 𝑧̄
us 𝑧 0 = 𝑤̄1 . Moreover, we have |𝑤0 |2 − |𝑤1 |2 = |𝑧0 |2 − |𝑧1 |2 and |𝑤0 |2 + |𝑤1 |2 =
0 1
1 = |𝑧0 |2 +|𝑧1 |2 . Putting these together implies 𝑧20 = 𝑤20 and 𝑧21 = 𝑤21 . This shows
that the desired 𝛼 with 𝛼 𝛼̄ = 1, i.e., 𝛼̄ = 𝛼1 , exists.
Now we assume that (𝑤0 , 𝑤1 ) = (𝛼𝑧0 , 𝛼𝑧1 ) with |𝛼|2 = 𝛼 𝛼̄ = 1: Then we compute
( )
𝜋(𝑤0 , 𝑤1 ) = 2𝑤0 𝑤̄ 1 , |𝑤0 |2 − |𝑤1 |2
( )
= 2𝛼𝑧0 𝛼̄ 𝑧̄ 1 , |𝛼|2 |𝑧0 |2 − |𝛼|2 |𝑧1 |2
( )
= 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2
= 𝜋(𝑧0 , 𝑧1 ).

(c) Let 𝑝 ∈ 𝕊2 be a fixed point and fix a point (𝑧0 , 𝑧1 ) in 𝕊3 with 𝜋(𝑧0 , 𝑧1 ) = 𝑝. By the
previous point, we have that the points in 𝜋 −1 (𝑝) is parametrized by the complex
number 𝛼 with |𝛼|2 = 1. The latter condition means 𝛼 ∈ 𝕊1 ⊂ ℂ. Hence we get a
bijective map
𝕊1 → 𝜋 −1 (𝑝), 𝛼 → (𝛼𝑧0 , 𝛼𝑧1 ).
Since this map just consists of multiplication with nonzero complex numbers, we
can conclude that it is a diffeomorphism (where we consider𝕊1 and 𝜋 −1 (𝑝) as sub-
sets in real Euclidean space).

(d) First we look at the map 𝜋̃ ∶ ℝ4 ≅ ℂ2 → ℂ × ℝ ≅ ℝ3 using the same formula as

for 𝜋, i.e., 𝜋 = 𝜋̃|𝕊3 , and compute its derivative at a point 𝑞 = (𝑧0 , 𝑧1 ). To do this
we write (𝑧0 , 𝑧1 ) = (𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) for real coordinates 𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 . First we get
( )
̃ 0 , 𝑧1 ) = 2𝑧0 𝑧̄ 1 , |𝑧0 |2 − |𝑧1 |2
𝜋(𝑧
( )
= 2(𝑥0 𝑥1 + 𝑦0 𝑦1 ) + 𝑖2(−𝑥0 𝑦1 + 𝑦0 𝑥1 ), (𝑥20 + 𝑦20 ) − (𝑥21 + 𝑦21 ) .

Then we can compute

⎛ 𝑥1 𝑦1 𝑥0 𝑦0 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ −𝑦1 𝑥1 𝑦0 −𝑥0 ⎟ .
⎜
⎜ ⎟
⎝ 𝑥0 𝑦0 −𝑥1 −𝑦1 ⎠
438 A.15. Linking Number and the Hopf Invariant

Note that if 𝑞 ∈ 𝕊3 , the restriction of 𝑑 𝜋̃𝑞 to 𝑇𝑞 𝕊3 has image contained in 𝑇𝜋(𝑞) 𝕊2 .

Recall: Since it has been a while that we looked at tangent spaces, let us see
why this claim is true. So let 𝑔4 ∶ ℝ4 → ℝ and 𝑔3 ∶ ℝ3 → ℝ be the usual
smooth maps such that 𝕊3 = 𝑔4−1 (1) and 𝕊2 = 𝑔3−1 (1) respectively. Then we have
𝑇𝑞 𝕊3 = Ker (𝑑(𝑔4 )𝑞 ) ⊂ 𝑇𝑞 ℝ4 = ℝ4 and 𝑇𝑝 𝕊2 = Ker (𝑑(𝑔3 )𝑝 ) ⊂ 𝑇𝑝 ℝ3 = ℝ3 .
We know that 𝜋(𝑞) ̃ ∈ 𝕊2 if 𝑞 ∈ 𝕊3 . Actually, our calculation above shows that
𝜋(𝑞)
̃ ∈ 𝕊 if and only if 𝑞 ∈ 𝕊3 . This implies 𝕊3 = 𝜋̃ −1 (𝑔3−1 (1)) = (𝑔3 ◦𝜋)
2 ̃ −1 (1). In
particular, 𝑔3 (𝜋(𝑞))
̃ = 1 is constant on 𝕊3 . Hence, for every 𝑞 ∈ 𝕊3 , the image of
the restriction (𝑑 𝜋̃𝑞 )|𝑇𝑞 𝕊3 is contained in the kernel of 𝑑(𝑔3 )𝜋(𝑞)
̃ which is 𝑇𝜋(𝑞)
̃ 𝕊 .
2

Now back to our task:

We want to show that every 𝑞 ∈ 𝕊3 is a regular point. Hence we need to show that
𝑑 𝜋̃𝑞 restricted to 𝑇𝑞 𝕊3 is surjective onto 𝑇𝜋(𝑞) 𝕊2 . Since the tangent space 𝑇𝜋(𝑞) 𝕊2
of 𝕊2 is two-dimensional, we need to check that the image of 𝑑 𝜋̃𝑞 restricted to
Ker (𝑑(𝑔4 )𝑞 ) spans a two-dimensional subspace. Since Ker (𝑑(𝑔4 )𝑞 ) is of dimen-
sion 3, it suffices to show that 𝑑 𝜋̃𝑞 has rank 3, which implies that the kernel of 𝑑 𝜋̃𝑞
has dimension 1. Hence we need to show that 𝑑 𝜋̃𝑞 always has 3 linear independent
columns.
We can show this for example by calculating the determinants of appropriate 3 × 3-
minors. Ignoring the factor 2 in our formula for 𝑑 𝜋̃𝑞 we look at the minors 𝐴𝑗 of
the remaining matrix where we omit the 𝑗th column:

∙ The determinant of 𝐴4 is −𝑥1 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑥1 .

∙ The determinant of 𝐴3 is −𝑦0 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑦0 .
∙ The determinant of 𝐴2 is −𝑥0 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑥0 .
∙ The determinant of 𝐴1 is −𝑦1 (𝑥20 + 𝑦20 + 𝑥21 + 𝑦21 ) = −𝑦1 .

For every point 𝑞 ∈ 𝕊3 , at least one of the coordinates 𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 is nonzero.

Hence the matrix always has three linear independent columns and 𝑑 𝜋̃𝑞 has rank
3. This shows that each point in 𝕊3 is a regular point for 𝜋, and hence every point
in 𝕊2 is a regular value for 𝜋.

(e) ∙ The fiber over 𝑎 is

𝜋 −1 (𝑎) = {(𝑧0 , 0) ∈ 𝕊3 ⊂ ℂ2 ∶ |𝑧0 |2 = 1}.

Let 𝑞 = (𝑥0 , 𝑦0 , 0, 0) ∈ 𝜋 −1 (𝑎) be a point in the fiber over 𝑎. The tangent

space 𝑇𝑞 𝕊3 is the vector space

𝑇𝑞 𝕊3 = {𝐮 ∈ ℝ4 ∶ 𝐮 ⟂ 𝑞}
⎧ ⎛−𝑦0 ⎞ ⎛0⎞ ⎛0⎞⎫
⎪ ⟂ ⎜ 𝑥0 ⎟ 4 ⎜0⎟ 4 ⎜0⎟⎪
= span ⎨𝑞 = ⎜ ⎟ , 𝐞3 = ⎜1⎟ , 𝐞4 = ⎜0⎟⎬ .
⎪ ⎜ 0 ⎟ ⎜ ⎟ ⎜ ⎟⎪
⎩ ⎝ 0 ⎠ ⎝0⎠ ⎝1⎠⎭
Appendix A. Solutions to exercises 439

The orientation of 𝑇𝑞 𝕊3 as a boundary of the unit ball is such that the outward
pointing vector 𝑞 together with the basis vectors of 𝑇𝑞 𝕊3 form a positively
oriented basis of ℝ4 . The determinant of the matrix expressing the basis
(𝑞, 𝑞 ⟂ , 𝐞43 , 𝐞44 ) in the standard basis of ℝ4 equals 𝑥20 +𝑦20 = 1 > 0. In particular,
it is positive and the basis (𝑞 ⟂ , 𝐞43 , 𝐞44 ) is a positively oriented basis of 𝑇𝑞 𝕊3 .
The tangent space 𝑇𝑞 𝜋 −1 (𝑎) equals the kernel of 𝑑 𝜋̃𝑞 . We computed this map
as represented by the matrix

⎛ 0 0 𝑥0 𝑦0 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ ⎜ 0 0 𝑦0 −𝑥0 ⎟ .
⎜ ⎟
⎝𝑥0 𝑦0 0 0 ⎠

The kernel of this map is the span of the vector 𝑞 ⟂ that we have just seen
above. The normal space 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ) ⊂ 𝑇𝑞 𝕊3 of vectors which are or-
thogonal to 𝑇𝑞 𝜋 −1 (𝑎) is the span of (𝐞43 , 𝐞44 ). The map 𝑑 𝜋̃𝑞 sends 𝐞3 and 𝐞4 to,
respectively,
⎛𝑥0 ⎞ ⎛ 𝑦0 ⎞
⎜ 𝑦 ⎟ ⎜−𝑥 ⎟
𝑑 𝜋̃𝑞 (𝐞43 ) = 2 ⎜ 0 ⎟ , 𝑑 𝜋̃𝑞 (𝐞44 ) = 2 ⎜ 0 ⎟ .
⎜0⎟ ⎜ 0 ⎟
⎝0⎠ ⎝ 0 ⎠
These two vectors form a basis (𝑑 𝜋̃𝑞 (𝐞33 ), 𝑑 𝜋̃𝑞 (𝐞44 )) of 𝑇𝑎 𝕊2 . We need to check
the orientation of this basis. The tangent space 𝑇𝑎 𝕊2 has a basis (𝐞31 , 𝐞32 ) as
a subspace in ℝ3 . This basis is positively oriented, since, together with the
outward pointing vector 𝑎 = 𝐞33 , the basis (𝐞33 , 𝐞31 , 𝐞32 ) is a positively oriented
basis of ℝ3 .a The matrix 𝐴 which expresses (𝑑 𝜋̃𝑞 (𝐞33 ), 𝑑 𝜋̃𝑞 (𝐞44 )) in terms of
the basis (𝐞31 , 𝐞32 ) is given by
( )
𝑥0 𝑦0
𝐴=2⋅ .
𝑦0 −𝑥0

We see that det 𝐴 = 4(−𝑥20 − 𝑦20 ) = −4 < 0 is negative. Hence the basis
(𝑑 𝜋̃𝑞 (𝐞33 ), 𝑑 𝜋̃𝑞 (𝐞44 )) is a negatively oriented basis of 𝑇𝑎 𝕊2 . This defines an
orientation on the normal space 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ) by declaring the orientation
of the basis (𝐞43 , 𝐞44 ) to be negative.

Finally, the orientation of 𝑇𝑞 𝜋 −1 (𝑎) is such that the direct sum

𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑎) = 𝑇𝑞 𝕊3

induces the given orientation on 𝑇𝑞 𝕊3 . We check this by looking at the basis

(𝐞43 , 𝐞44 , 𝑞 ⟂ ) of 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑎). As a basis of 𝑇𝑞 𝕊3 this basis is
positively oriented, since it arises by two permutations from the positively
oriented basis (𝑞 ⟂ , 𝐞43 , 𝐞44 ). Since the sign of (𝐞43 , 𝐞44 ) is negative as a basis of
𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ), we need that 𝑞 ⟂ also has negative sign. Hence the vector 𝑞 ⟂
provides a negatively oriented basis of 𝑇𝑞 𝜋 −1 (𝑎).
Comparing this orientation with the standard orientation of 𝕊1 ⊂ ℂ ⊂ ℂ2 ,
we see that 𝜋 −1 (𝑎) has the opposite orientation.
440 A.15. Linking Number and the Hopf Invariant
∙ To determine the fiber over 𝑏, we write 𝑧0 = 𝑥0 + 𝑖𝑦0 and 𝑧1 = 𝑥1 + 𝑖𝑦1 .
Then we get
1
𝜋(𝑧0 , 𝑧1 ) = (0, 1, 0) ⇒ 2𝑧0 𝑧̄ 1 = 𝑖 and |𝑧0 |2 = |𝑧1 |2 =
2
1
⇒ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 and 𝑥20 + 𝑥21 = .
2
Thus the fiber over 𝑏 has the form
𝑖
𝜋 −1 (𝑏) = {(𝑧0 , 𝑧1 ) ∈ 𝕊3 ∶ 𝑧̄ 1 = }
2𝑧0
= {(𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) ∈ 𝕊3 ∶ 𝑦0 = 𝑥1 , 𝑦1 = −𝑥0 }.

Let 𝑞 = (𝑥0 , 𝑥1 , 𝑥1 , −𝑥0 ) ∈ 𝜋 −1 (𝑏) be a point in the fiber over 𝑏. Since not
both 𝑥0 and 𝑥1 can be zero, we assume that 𝑥0 ≠ 0. The tangent space 𝑇𝑞 𝕊3
is the vector space

𝑇𝑞 𝕊3 = {𝐮 ∈ ℝ4 ∶ 𝐮 ⟂ 𝑞}
⎧ ⎛−𝑥1 ⎞ ⎛−𝑥1 ⎞ ⎛𝑥0 ⎞⎫
⎪ ⟂ ⎜ 𝑥0 ⎟ ⟂ ⎜ 0 ⎟ ⟂ ⎜ 0 ⎟⎪
= span ⎨𝑞1 = ⎜ ⎟ , 𝑞2 = ⎜ 𝑥 ⎟ , 𝑞3 = ⎜ 0 ⎟⎬ .
⎪ ⎜ 0 ⎟ ⎜ 0⎟ ⎜ ⎟⎪
⎩ ⎝ 0 ⎠ ⎝ 0 ⎠ ⎝𝑥0 ⎠⎭

The orientation of 𝑇𝑞 𝕊3 as a boundary of the unit ball is such that the outward
pointing vector 𝑞 together with the basis vectors of 𝑇𝑞 𝕊3 form a positively
oriented basis of ℝ4 . The matrix expressing the basis (𝑞, 𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) in the
standard basis of ℝ4 is
⎛ 𝑥0 −𝑥1 −𝑥1 𝑥0 ⎞
⎜ 𝑥1 𝑥0 0 0⎟
⎜𝑥 0 𝑥 0 ⎟.
⎜ 1 0 ⎟
⎝−𝑥0 0 0 𝑥0 ⎠
The determinant of this matrix is

2𝑥40 + 2𝑥20 𝑥21 = 2𝑥20 (𝑥20 + 𝑥21 ) = 𝑥20 > 0.

In particular, it is positive and the basis (𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) is a positively oriented
basis of 𝑇𝑞 𝕊3 .
The tangent space 𝑇𝑞 𝜋 −1 (𝑏) equals the kernel of 𝑑 𝜋̃𝑞 . We computed this map
as represented by the matrix

⎛𝑥1 −𝑥0 𝑥0 𝑥1 ⎞
𝑑 𝜋̃𝑞 = 2 ⋅ ⎜𝑥0 𝑥1 𝑥1 −𝑥0 ⎟ .
⎜ ⎟
⎝𝑥0 𝑥1 −𝑥1 𝑥0 ⎠

⎛−𝑥1 ⎞
⎜𝑥 ⎟
The kernel of this map is the span of the vector 𝑞0⟂ = ⎜ 0 ⎟. The normal
⎜ 𝑥0 ⎟
⎝ 𝑥1 ⎠
space 𝑁𝑞 (𝜋 (𝑏); 𝕊 ) ⊂ 𝑇𝑞 𝕊 of vectors which are orthogonal to 𝑇𝑞 𝜋 −1 (𝑏) is
−1 3 3
Appendix A. Solutions to exercises 441

the span of (𝑞1⟂ −𝑞2⟂ , 𝑞3⟂ ). The map 𝑑 𝜋̃𝑞 sends 𝑞1⟂ −𝑞2⟂ and 𝑞3⟂ to, respectively,

⎛ −2𝑥20 ⎞ ⎛−2𝑥0 𝑥1 ⎞
𝑑 𝜋̃𝑞 (𝑞1⟂ − 𝑞2⟂ ) = 2 ⎜ 0 ⎟ , 𝑑 𝜋̃𝑞 (𝑞3⟂ ) = 2 ⎜ 0 ⎟ .
⎜ ⎟ ⎜ 2 ⎟
⎝−2𝑥0 𝑥1 ⎠ ⎝ 2𝑥0 ⎠

These two vectors form a basis (𝑑 𝜋̃𝑞 (𝑞1⟂ − 𝑞2⟂ ), 𝑑 𝜋̃𝑞 (𝑞3⟂ )) of 𝑇𝑏 𝕊2 . We need to
check the orientation of this basis. The tangent space 𝑇𝑏 𝕊2 has a basis (𝐞31 , 𝐞33 )
as a subspace in ℝ3 . This basis is negatively oriented, since, together with the
outward pointing vector 𝑏 = 𝐞32 , the basis (𝐞32 , 𝐞31 , 𝐞33 ) is a negatively oriented
basis of ℝ3 . For we need to make one permutation to get the standard basis
which leads to multiplying the sign with −1. The matrix 𝐵 which expresses
(𝑑 𝜋̃𝑞 (𝑞1⟂ − 𝑞2⟂ ), 𝑑 𝜋̃𝑞 (𝑞3⟂ )) in terms of the basis (𝐞31 , 𝐞33 ) is given by
( )
−𝑥20 −𝑥0 𝑥1
𝐵 =4⋅ .
−𝑥0 𝑥1 𝑥20

We see that det 𝐵 = 16(−𝑥40 − 𝑥20 𝑥21 ) = −16𝑥20 (𝑥20 + 𝑥21 ) = −8𝑥20 < 0 is neg-
ative. Hence the basis (𝑑 𝜋̃𝑞 (𝑞1⟂ − 𝑞2⟂ ), 𝑑 𝜋̃𝑞 (𝑞3⟂ )) is a negatively oriented basis
of 𝑇𝑏 𝕊2 . This defines an orientation on the normal space 𝑁𝑞 (𝜋 −1 (𝑏); 𝕊3 ) by
declaring the orientation of the basis (𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ ) to be negative.

Finally, the orientation of 𝑇𝑞 𝜋 −1 (𝑏) is such that the direct sum

𝑁𝑞 (𝜋 −1 (𝑏); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑏) = 𝑇𝑞 𝕊3

induces the given orientation on 𝑇𝑞 𝕊3 . We check this by looking at the basis

(𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ , 𝑞0⟂ ) of 𝑁𝑞 (𝜋 −1 (𝑏); 𝕊3 ) ⊕ 𝑇𝑞 𝜋 −1 (𝑏). The transition matrix from
the basis (𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) to the basis (𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ , 𝑞0⟂ ) is given by

⎛1 0 1 ⎞
⎜−1 0 1 ⎟.
⎜ ⎟
⎝ 0 1 𝑥1 ∕𝑥0 ⎠
The determinant of this matrix is −2. In particular, it is negative. Since we
checked that the basis (𝑞1⟂ , 𝑞2⟂ , 𝑞3⟂ ) is a negatively oriented basis, we see that
the orientation of the basis (𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ , 𝑞0⟂ ) of 𝑇𝑞 𝕊3 is positive. Since the
sign of (𝑞1⟂ − 𝑞2⟂ , 𝑞3⟂ ) is positive as a basis of 𝑁𝑞 (𝜋 −1 (𝑎); 𝕊3 ), we need that
the basis 𝑞0⟂ also has positive sign. Hence the vector 𝑞0⟂ provides a positively
oriented basis of 𝑇𝑞 𝜋 −1 (𝑏).

(f) By definition of 𝐻(𝜋), we need to choose two distinct regular values 𝑎 and 𝑏 of
𝜋 and calculate the linking number of 𝜋 −1 (𝑎) and 𝜋 −1 (𝑏). Since we showed that
each value is regular, we can for example choose 𝑎 = (0, 0, 1) and 𝑏 = (0, 1, 0) on
𝕊2 ⊂ ℝ3 ≅ ℂ × ℝ.
To calculate the linking number of 𝜋 −1 (𝑎) and 𝜋 −1 (𝑏) we need to choose a point
on 𝕊3 disjoint from these two subsets and stereographically project 𝕊3 from this
point onto ℝ3 . By our choice of 𝑎 and 𝑏, we get that the north pole 𝑁 = (0, 0, 0, 1)
442 A.15. Linking Number and the Hopf Invariant

is neither on 𝜋 −1 (𝑎) nor on 𝜋 −1 (𝑏). Recall that the formula for the stereographic
projection 𝜙−1
𝑁
∶ 𝕊3 ⧵ {𝑁} → ℝ3 is, with the notation we use here, given by

1
(𝑥0 , 𝑦0 , 𝑥1 , 𝑦1 ) → (𝑥 , 𝑦 , 𝑥 ).
1 − 𝑦1 0 0 1

Hence we get

𝑆𝑎 ≔ 𝜙−1 −1
𝑁 (𝜋 (𝑎))
= {𝐯 = (𝑣0 , 𝑣1 , 𝑣2 ) ∈ ℝ3 ∶ 𝑣20 + 𝑣21 = 1 and 𝑣2 = 0}

and

𝑆𝑏 ≔ 𝜙−1 −1
𝑁 (𝜋 (𝑏))
{ }
(1 − 𝑤0 )2
= 𝐰 = (𝑤0 , 𝑤1 , 𝑤2 ) ∈ ℝ ∶ 𝑤1 = 𝑤2 and 𝑤0 + 𝑤1 =
3 2 2
.
2

Now we can calculate 𝐻(𝜋) as deg(𝜆) with

𝐯−𝐰
𝜆 ∶ 𝑆𝑎 × 𝑆𝑏 → 𝕊2 , (𝐯, 𝐰) → .
|𝐯 − 𝐰|

To compute the degree of 𝜆 we pick a convenient point of 𝕊2 and determine the

fiber over this point. Then we check that we actually picked a regular value.
So let us look at 𝑝 = (1, 0, 0). The equation 𝜆(𝐯, 𝐰) = 𝑝 then implies

𝑣1 = 𝑤1 = 0 and 𝑣0 − 𝑤0 = |𝑣0 − 𝑤0 |.

The latter condition, implies that 𝑣0 − 𝑤0 is positive. This does not look very
helpful at first glance, but we also know

1 = 𝑣20 + 𝑣21 = 𝑣20 , i.e., 𝑣0 = ±1,

and
(1 − 𝑤0 )2 √
𝑤20 = ⇐⇒ 𝑤0 = ± 2 − 1.
2
Now we can check for the four possible permutations of the signs
√ whether they
yield 𝑣0 − 𝑤0 ≥ √ 0 and get three points: one with 𝑣0 = 1, 𝑤
√0 = 2 − 1, one with
𝑣0 = 1, 𝑤0 = − 2 − 1, and one with 𝑣0 = −1, 𝑤0 = − 2 − 1. Hence we get
three points (𝐯, 𝐰) in 𝑆𝑎 × 𝑆𝑏 with 𝜆(𝐯, 𝐰) = 𝑝.
It remains to check the derivatives of 𝜆 at these points. We have to show that the
determinants at each point are nonzero and that the sum of the signs is +1. For
then we get deg(𝜆) = +1 as claimed.
Since 𝑆𝑎 is the unit circle in the 𝑥𝑦-plane, the tangent space of 𝑆𝑎 at a point 𝐯 is
given by

𝑇𝐯 𝑆𝑎 = {𝐮 = (𝑢0 , 𝑢1 , 𝑢2 ) ∈ ℝ3 ∶ 𝑢2 = 0 and 𝑢0 𝑣0 + 𝑢1 𝑣1 = 0}.

Appendix A. Solutions to exercises 443

Similarly, 𝑆𝑏 lies in the plane 𝑃 in ℝ3 of points 𝐰 = (𝑤0 , 𝑤1 , 𝑤2 ) with 𝑤1 = 𝑤2 .

Then 𝑆𝑏 is the fiber of the map

(1 − 𝑤0 )2
𝑔𝑏 ∶ 𝑃 → ℝ, 𝐰 = (𝑤0 , 𝑤1 , 𝑤1 ) → 𝑤20 + 𝑤21 = .
2
After a simple computation we get

𝑆𝑏 = 𝑔𝑏−1 (1) = {𝐰 = (𝑤0 , 𝑤1 , 𝑤1 ) ∈ 𝑃 ∶ 2𝑤21 + 𝑤20 + 2𝑤0 = 1}.

The derivative of 𝑔𝑏 as a map from 𝑃 → ℝ is given by the matrix (we could also
consider it as a map ℝ3 → ℝ2 )

𝑑(𝑔𝑏 )𝐰 = (2𝑤0 + 2, 2𝑤1 ).

Hence we get

𝑇𝐰 𝑆𝑏 = {𝐮 = (𝑢0 , 𝑢1 , 𝑢2 ) ∈ ℝ3 ∶ 𝑢2 = 𝑢1 and 𝑢0 (𝑤0 + 1) + 𝑢1 𝑤1 = 0}.

Now we calculate the derivative of 𝜆. First we do this as a map

𝐯−𝐰
𝜆̃ ∶ ℝ3 × ℝ3 → ℝ3 , (𝐯, 𝐰) → .
|𝐯 − 𝐰|

This will help us, since 𝑔3 ◦𝜆̃ is constant on 𝑆𝑎 × 𝑆𝑏 . Hence 𝑑 𝜆̃(𝐯,𝐰) sends the
subspace 𝑇𝐯 × 𝑇𝐰 𝑆𝑏 ∈ ℝ3 × ℝ3 to 𝑇𝑝 𝕊2 ⊂ ℝ3 .
𝜕 𝜆̃𝑖 𝜕 𝜆̃𝑖
We determine 𝑑 𝜆̃(𝐯,𝐰) by computing its partial derivatives 𝜕𝑣𝑗
(𝐯, 𝐰) and 𝜕𝑤𝑗
(𝐯, 𝐰)
with respect to the variables 𝑣0 , 𝑣1 , 𝑣2 and 𝑤0 , 𝑤1 , 𝑤2 : For 𝑖 ≠ 𝑗, we have

𝜕 𝜆̃𝑖 (𝑣𝑖 − 𝑤𝑖 )(𝑣𝑗 − 𝑤𝑗 ) 𝜕 𝜆̃𝑖

(𝐯, 𝐰) = = (𝐯, 𝐰).
𝜕𝑣𝑗 |𝐯 − 𝐰|3 𝜕𝑤𝑗

For 𝑖 = 𝑗, we get

⎧(𝑣1 − 𝑤1 )2 + (𝑣2 − 𝑤2 )2 if 𝑖 = 0
𝜕 𝜆̃𝑖 1 ⎪
(𝐯, 𝐰) = ⋅ ⎨(𝑣0 − 𝑤0 )2 + (𝑣2 − 𝑤2 )2 if 𝑖 = 1
𝜕𝑣𝑖 |𝐯 − 𝐰|3 ⎪(𝑣 − 𝑤 )2 + (𝑣 − 𝑤 )2
⎩ 0 0 1 1 if 𝑖 = 2

and
𝜕 𝜆̃𝑖 𝜕 𝜆̃
(𝐯, 𝐰) = − 𝑖 (𝐯, 𝐰).
𝜕𝑤𝑖 𝜕𝑣𝑖

Now we evaluate these formulae at the points (𝐯, 𝐰) with 𝜆(𝐯, 𝐰) = 𝑝. For each
such point we found 𝑣1 = 𝑣2 = 𝑤1 = 𝑤2 = 0. Hence we get

𝜕 𝜆̃𝑖 𝜕 𝜆̃𝑖
(𝐯, 𝐰) = 0 = (𝐯, 𝐰)
𝜕𝑣𝑗 𝜕𝑤𝑗
444 A.15. Linking Number and the Hopf Invariant
for 𝑖 ≠ 𝑗,
{
𝜕 𝜆̃𝑖 1 0 if 𝑖 = 0
(𝐯, 𝐰) = ⋅
𝜕𝑣𝑖 |𝑣0 − 𝑤0 | 1 if 𝑖 = 1, 2
and
{
𝜕 𝜆̃𝑖 1 0 if 𝑖 = 0
(𝐯, 𝐰) = ⋅
𝜕𝑤𝑖 |𝑣0 − 𝑤0 | −1 if 𝑖 = 1, 2.

Now we are equipped to study the linear map 𝑑𝜆(𝐯,𝐰) ∶ 𝑇𝐯 𝑆𝑎 × 𝑇𝐰 𝑆𝑏 → 𝑇𝑝 𝕊2 :

⎛0⎞ ⎛0⎞
A basis of 𝑇𝑝 𝕊2 is given by the vectors 𝐞2 = ⎜1⎟ and 𝐞3 = ⎜0⎟. Since the vector
⎜ ⎟ ⎜ ⎟
⎝0⎠ ⎝1⎠
⎛1⎞
𝐩 = 𝐞1 = ⎜0⎟ is an outward pointing normal vector, this is a positively oriented
⎜ ⎟
⎝0⎠
basis of 𝑇𝑝 𝕊2 .
⎛0⎞
At each of the points (𝐯, 𝐰) we found with 𝜆(𝐯, 𝐰) = 𝐩, the vector 𝐚 = ⎜1⎟ is a
⎜ ⎟
⎝0⎠
⎛0⎞
basis of 𝑇𝐯 𝑆𝑎 and the vector 𝐛 = ⎜1⎟ provides a basis of 𝑇𝐰 𝑆𝑏 . The map 𝑑𝜆(𝐯,𝐰)
⎜ ⎟
⎝1⎠
⎛0⎞ ⎛0⎞
1
sends 𝐚 to |𝑣 −𝑤 ⋅ ⎜1⎟ and 𝐛 to 1 ⋅ ⎜−1⎟. Hence we have
0 0| ⎜ ⎟ |𝑣0 −𝑤0 | ⎜ ⎟
⎝0⎠ ⎝−1⎠
1
𝑑𝜆(𝐯,𝐰) (𝐚) = ⋅𝐞
|𝑣0 − 𝑤0 | 2
and
1 1
𝑑𝜆(𝐯,𝐰) (𝐛) = − ⋅ 𝐞2 − ⋅𝐞 .
|𝑣0 − 𝑤0 | |𝑣0 − 𝑤0 | 3
These two vectors form a basis of 𝑇𝑝 𝕊2 and we see that (𝐯, 𝐰) is a regular point.
Since this is true for all points in the fiber of 𝑝 ∈ 𝕊2 , we conclude that 𝑝 actually
is a regular value.

To check the orientation of the basis, we calculate the determinant of

( )
1 1 −1
𝑑𝜆(𝐯,𝐰) =
|𝑣0 − 𝑤0 | 0 −1
which is also the transition matrix from (𝐞2 , 𝐞3 ) to the new basis. The determinant
of this matrix is negative. Hence the basis (𝑑𝜆(𝐯,𝐰) (𝐚), 𝑑𝜆(𝐯,𝐰) (𝐛)) of 𝑇𝑝 𝕊2 is
negatively oriented.

In order to understand the effect of 𝑑𝜆(𝐯,𝐰) on the orientations, we must determine

whether (𝐚, 𝐛) is a positively or negatively oriented basis of 𝑇𝐯 𝑆𝑎 × 𝑇𝐰 𝑆𝑏 :
Appendix A. Solutions to exercises 445
√
∙ (𝐯, 𝐰) with 𝑣0 = 1, 𝑤0 = 2 − 1 and 𝑣1 = 𝑣2 = 𝑤1 = 𝑤2 = 0: As
we checked in a previous point, the orientation of 𝑆𝑎 is opposite to the
standard orientation of the circle. Hence the basis (𝐚) is a negatively oriented
basis of 𝑇𝐯 𝑆𝑎 . The basis (𝐛), however, is a positively oriented basis of
√
⎛ 1∕ 2 ⎞
⎜ ⎟
0 ⎟
𝑇𝐰 𝑆𝑏 .b For, at this 𝐰, the corresponding vectors are 𝑞 = ⎜ and
⎜ 0√ ⎟
⎜ ⎟
⎝−1∕ 2⎠
⎛ 0 ⎞ ⎛ 0 ⎞
⎜ √ ⎟ √
1∕ 2
𝑞0⟂ = ⎜ √ ⎟, and 𝐛 is a positive multiple of 𝑑(𝜙−1 ⟂ ) = ⎜1∕ 2⎟. Since
⎜1∕ 2⎟ 𝑁
) 𝑞 (𝑞0 ⎜ √ ⎟
⎜ ⎟ ⎜ ⎟
⎝ 0 ⎠ ⎝1∕ 2⎠
𝑞0⟂ has sign +1 and 𝜙𝑁 preserves orientations, 𝐛 also has sign +1. Hence
(𝐚 × 0, 0 × 𝐛) is a negatively oriented basis of 𝑇𝐯 𝑆𝑎 × 𝑇𝐰 𝑆𝑏 . Thus 𝑑𝜆(𝐯,𝐰)
sends a negatively oriented basis to a negatively oriented basis and therefore
preserves orientations. Hence the orientation number at (𝐯, 𝐰) is +1.
√
∙ (𝐯, 𝐰) with 𝑣0 = 1, 𝑤0 = − 2 − 1 and 𝑣1 = 𝑣2 = 𝑤1 = 𝑤2 = 0: The
vector 𝐚 is a negatively oriented basis of 𝑇𝐯 𝑆𝑎 . The basis 𝐛, however, is now
also a negatively oriented basis of 𝑇𝐰 𝑆𝑏 . For, at this 𝐰, the corresponding
√
⎛−1∕ 2⎞ ⎛ 0 ⎞
⎜ ⎟ ⎜ √ ⎟
⎜ 0 ⎟ ⎜−1∕ 2⎟
vector are 𝑞 = and 𝑞0 =
⟂ √ , and 𝐛 is a negative multiple
⎜ 0√ ⎟ ⎜−1∕ 2⎟
⎜ ⎟ ⎜ ⎟
⎝ 1∕ 2 ⎠ ⎝ 0 ⎠
⎛ 0 ⎞
√
of 𝑑(𝜙−1 ⟂ ) = ⎜−1∕ 2⎟. Since 𝑞 ⟂ has sign +1 and 𝜙
)
𝑁 𝑞 0
(𝑞 ⎜ √ ⎟ 0 𝑁 preserves
⎜−1∕ 2⎟
⎝ ⎠
orientations, 𝐛 has sign −1. Hence (𝐚 × 0, 0 × 𝐛) is a positively oriented
basis of 𝑇𝐯 𝑆𝑎 × 𝑇𝐰 𝑆𝑏 . Thus 𝑑𝜆(𝐯,𝐰) sends a positively oriented basis to a
negatively oriented basis and therefore reverses orientations. Hence the
orientation number at (𝐯, 𝐰) is −1.
√
∙ (𝐯, 𝐰) with 𝑣0 = −1, 𝑤0 = − 2 − 1 and 𝑣1 = 𝑣2 = 𝑤1 = 𝑤2 = 0:
Now the vector 𝐚 is a positively oriented basis of 𝑇𝐯 𝑆𝑎 and 𝐛 is a negatively
oriented basis of 𝑇𝐰 𝑆𝑏 . Hence (𝐚 × 0, 0 × 𝐛) is a negatively oriented basis of
𝑇𝐯 𝑆𝑎 × 𝑇𝐰 𝑆𝑏 . Thus 𝑑𝜆(𝐯,𝐰) sends a negatively oriented basis to a negatively
oriented basis and therefore preserves orientations. Hence the orientation
number at (𝐯, 𝐰) is +1.

Thus in total we get that the sum of the orientation numbers is +1 − 1 + 1 = +1.
Hence we have shown 𝐻(𝜋) = deg(𝜆) = 1.
a
We make two permutations which lead to multiplying with (−1)2 = +1.
b
Imagining 𝑆𝑏 as a deformation of the unit circle lying in the plane 𝑦 = 𝑧 in ℝ3 and the vector 𝐛 points
upwards at a point with positive 𝑥-coordinate. We checked that the orientation of 𝑆𝑏 is opposite to the
standard orientation of the circle.
Bibliography

[1] Bjørn Dundas, A short course in differential topology, Cambridge Mathematical Textbooks,
Cambridge University Press, Cambridge, 2018.

[2] David Eisenbud, Harold I. Levine, An algebraic formula for the degree of a 𝐶 ∞ map germ,
Ann. of Math. (2) 106 (1977), no.1, 19–44.

[3] David Eisenbud, An algebraic approach to the topological degree of a smooth map, Bull.
Amer. Math. Soc. 84 (1978), no.5, 751–764.

[4] William Fulton, Algebraic topology, Graduate Texts in Mathematics 153, Springer-Verlag,
New York, 1995.

[5] Victor Guillemin, Alan Pollack, Differential topology, Reprint of the 1974 original, AMS
Chelsea Publishing, Providence, RI, 2010.

[6] Allen Hatcher, Algebraic topology, Cambridge University Press, Cambridge, 2002.

[7] Heinz Hopf, Über die Abbildungen von Sphären auf Sphären niedrigerer Dimension, Fun-
damenta Mathematicae 25 (1935), 427–440.

[8] Heinz Hopf, Vektorfelder in 𝑛-dimensionalen Mannigfaltigkeiten, Math. Annalen 96

(1926), 225–250.

[9] Daniel Huybrechts, Complex Geometry. An introduction. Universitext. Springer-Verlag,

Berlin, 2005.

[10] Klaus Jänich, Topology, Undergraduate Texts in Mathematics. Springer-Verlag, New

York, 1984.

[11] John M. Lee, Introduction to smooth manifolds, Graduate Texts in Mathematics 218,
Springer-Verlag, New York, 2013.

[12] John W. Milnor, Morse Theory, Annals of Mathematics Studies, No. 51, Princeton Uni-
versity Press, Princeton, NJ, 1963.

[13] John W. Milnor, Topology from the differentiable viewpoint, Princeton Landmarks in
Mathematics, Revised reprint of the 1965 original, Princeton University Press, Princeton,
NJ, 1997.

[14] John W. Milnor, James D. Stasheff, Characteristic classes, Annals of Mathematics Stud-
ies, No. 76, Princeton University Press, Princeton, NJ, 1974.

[15] James R. Munkres, Topology, Prentice Hall, Inc., Upper Saddle River, NJ, Second edition,
2000.

[16] Gereon Quick, Algebraic Topology - Lecture Notes, online notes from a course in 2018.

446
Bibliography 447
[17] Loring W. Tu, An introduction to manifolds, Universitext, Springer-Verlag, New York,
2011.

[18] Claire Voisin, Hodge theory and complex algebraic geometry. I. Translated from the
French original by Leila Schneps. Cambridge Studies in Advanced Mathematics, 76. Cam-
bridge University Press, Cambridge, 2002.
Index

𝜀-Neighborhood Lemma, 246, 350 graph of a map, 57

Grassmannian, 54, 162
abstract manifold, 152
antipodal map, 288 Hausdorff, 72, 398
Hausdorff space, 151
boundary of one-manifold, 201 Hessian matrix, 131
Brieskorn manifold, 116, 121 homeomorphism, 19
bump function, 141, 178, 351 homotopy, 138, 157, 211, 212, 216, 220,
235, 270, 293, 307, 308, 311–313
canonical line bundle, 157 equivalence, 276
cell complex, 343 group, 312
chain rule, 46 Hopf invariant, 233, 236, 280, 320, 322,
classifying space, 164 441
cobordism, 282, 418 Hopf manifold, 160, 184
codimension, 111, 194, 196, 202, 249, Hopf map, 26, 27, 29, 46, 56, 75, 183, 235,
270, 298 236, 256, 296, 305
compact, 20 Hopf number, 321
connected, 22, 91 hyperboloid, 32, 56, 117
continuous map, 18
contractible, 214, 281 immersion, 63
critical value, 82 intersection number, 270, 271, 273
self-intersection, 272, 275, 283
degree, 213, 214, 223, 271, 306 isotopy, 212, 216, 338
derivative, 44 Isotopy Lemma, 212, 216, 218, 309, 324,
Hopf map, 46 331
diffeomorphism, 27, 35, 62, 66, 70, 107,
287 Kervaire invariant, 121
local, 58, 59, 62, 63, 65, 66, 68, 70, Klein bottle, 159, 289
75, 85, 106, 107, 135, 145, 199,
242, 290, 327 Lefschetz map, 389, 414
local but not global, 61 Lie group, 84, 96, 380
general linear, 100, 384
embedding, 70, 71 orthogonal, 83, 374, 386, 411
Euler characteristic, 343 special linear, 101, 376, 386, 411
exotic sphere, 120 special orthogonal, 101, 386, 411
special unitary, 102
fiber, 80 spin, 102
fundamental group, 278 unitary, 101
Fundamental Theorem of Algebra, 88 Linear Isotopy Lemma, 324, 327
linking number, 232, 233, 320, 435
Gauss map, 350, 355
degree of, 346, 350 Möbius band, 208, 273, 283, 289, 291

448
Index 449
Milnor hypersurface, 185 Boundary (degree), 233, 234, 307,
Morse function, 131, 354 335, 407, 435
Morse Lemma, 132 Boundary (intersection number), 273,
282
neighborhood, 17 Brouwer Fixed Point, 203, 205, 209
normal bundle, 244, 249, 251, 305 Classification of one-manifolds, 197,
201, 211, 271
open ball, 14 Constant Rank, 99, 382
orientable, 287, 288, 290, 291, 298, 305 Dirichlet’s Approximation, 69
orientation Ehresmann Fibration, 256
boundary, 292, 298, 332 Embedding, 71
opposite, 289 Extension (transversality), 262, 331
preimage, 296, 298 Extension for degree zero maps, 331
pullback, 289 Heine–Borel, 21
real vector space, 286 Hopf Degree, 313, 323
smooth manifold, 287 Hopf nowhere vanishing vector fields,
345, 355
path-connected, 23, 103, 149
Invariance of domain, 206
proper map, 71
Inverse Function, 59, 61, 65, 85, 208,
real projective space, 151, 155, 163, 274, 242, 243, 251, 257
280 Isotopy, 218, 330
regular value, 80, 81, 83, 85, 86, 91, 195, Local Immersion, 65, 70, 109
235, 307, 309 Local Submersion, 77, 79, 125
relative open, 15 Perron–Frobenius, 209
retraction, 202 Poincaré–Hopf Index, 344
Preimage, 80, 162, 195, 245, 258,
simply-connceted, 276 268, 374, 376
simply-connected, 149, 209, 274, 277, Preimage with boundary, 196, 211,
281, 300 271
singular chain, 342 Sard, 124, 128, 149, 173, 213, 235,
singular homology, 342 281, 312
smooth map, 24, 26, 35, 152 Sard with boundary, 197
stable property, 143 Stack of Records, 86, 91, 307, 328,
stereographic projection, 34, 35, 56, 149, 330
186, 233, 236, 275, 281 Topological invariance of domain, 206
Stiefel manifold, 160 Transversality, 258
submanifold, 35, 66, 80, 111, 114 Transversality Homotopy, 262, 273
submersion, 75 Tubular Neighborhood, 247, 253, 257,
331
tangent bundle, 51 Whitney Approximation, 140, 253,
tangent space, 40, 155 277
2-sphere, 42 Whitney Embedding, 174, 175
Grassmannian, 167 Whitney Immersion, 173
real projective space, 157 torus, 36, 43, 57, 64, 68, 130, 134, 158,
unit circle, 41 272, 313, 344, 345
Theorem transverse, 111, 113, 114, 116, 128, 268,
Baire Category, 382 270, 272, 386, 411
Borsuk–Ulam, 220 tubular neighborhood, 247
450 Index
vector bundle, 53, 164 isolated zero, 334
vector field, 150, 333 nondegenerate zero, 347
derivative, 348 pullback, 335
gradient, 333, 336, 353 torus, 345
index of a zero, 334, 337 velocity vector, 57
index sum, 339, 344, 351
index sum of gradient field, 353 winding number, 219, 324, 326, 327, 329