Draft: March 28, 2018

Differential Forms

Victor Guillemin & Peter J. Haine

Draft: March 28, 2018
 Draft: March 28, 2018

Contents

Preface v
Introduction v
Organization vi
Notational Conventions x
Acknowledgments xi

Chapter 1. Multilinear Algebra 1

1.1. Background 1
1.2. Quotient spaces & dual spaces 3
1.3. Tensors 8
1.4. Alternating 𝑘-tensors 11
1.5. The space 𝛬𝑘 (𝑉⋆ ) 17
1.6. The wedge product 20
1.7. The interior product 23
1.8. The pullback operation on 𝛬𝑘 (𝑉⋆ ) 25
1.9. Orientations 29

Chapter 2. Differential Forms 33

2.1. Vector fields and one-forms 33
2.2. Integral Curves for Vector Fields 37
2.3. Differential 𝑘-forms 44
2.4. Exterior differentiation 46
2.5. The interior product operation 51
2.6. The pullback operation on forms 54
2.7. Divergence, curl, and gradient 59
2.8. Symplectic geometry & classical mechanics 63

Chapter 3. Integration of Forms 71

3.1. Introduction 71
3.2. The Poincaré lemma for compactly supported forms on rectangles 71
3.3. The Poincaré lemma for compactly supported forms on open subsets of 𝐑𝑛 76
3.4. The degree of a differentiable mapping 77
3.5. The change of variables formula 80
3.6. Techniques for computing the degree of a mapping 85
3.7. Appendix: Sard’s theorem 92

Chapter 4. Manifolds & Forms on Manifolds 97

4.1. Manifolds 97
4.2. Tangent spaces 104
4.3. Vector fields & differential forms on manifolds 109
iii
 Draft: March 28, 2018

iv Contents

4.4. Orientations 116

4.5. Integration of forms on manifolds 124
4.6. Stokes’ theorem & the divergence theorem 128
4.7. Degree theory on manifolds 133
4.8. Applications of degree theory 137
4.9. The index of a vector field 143

Chapter 5. Cohomology via forms 149

5.1. The de Rham cohomology groups of a manifold 149
5.2. The Mayer–Vietoris Sequence 158
5.3. Cohomology of Good Covers 165
5.4. Poincaré duality 171
5.5. Thom classes & intersection theory 176
5.6. The Lefschetz Theorem 183
5.7. The Künneth theorem 191
5.8. Čech Cohomology 194
Appendix a. Bump Functions & Partitions of Unity 201
Appendix b. The Implicit Function Theorem 205

Appendix c. Good Covers & Convexity Theorems 211

Bibliography 215
Index of Notation 217

Glossary of Terminology 219

 Draft: March 28, 2018

Preface

Introduction
For most math undergraduates one’s first encounter with differential forms is the change
of variables formula in multivariable calculus, i.e. the formula

(1) ∫ 𝑓⋆ 𝜙| det 𝐽𝑓 |𝑑𝑥 = ∫ 𝜙 𝑑𝑦

𝑈 𝑉

In this formula, 𝑈 and 𝑉 are bounded open subsets of 𝐑𝑛 , 𝜙 ∶ 𝑉 → 𝐑 is a bounded contin-

uous function, 𝑓 ∶ 𝑈 → 𝑉 is a bijective differentiable map, 𝑓⋆ 𝜙 ∶ 𝑈 → 𝐑 is the function
𝜙 ∘ 𝑓, and det 𝐽𝑓 (𝑥) is the determinant of the Jacobian matrix.

𝜕𝑓𝑖
𝐽𝑓 (𝑥) ≔ [ (𝑥)] ,
𝜕𝑥𝑗
As for the “𝑑𝑥” and “𝑑𝑦”, their presence in (1) can be accounted for by the fact that in single-
variable calculus, with 𝑈 = (𝑎, 𝑏), 𝑉 = (𝑐, 𝑑), 𝑓 ∶ (𝑎, 𝑏) → (𝑐, 𝑑), and 𝑦 = 𝑓(𝑥) a 𝐶1 function
with positive first derivative, the tautological equation 𝑑𝑦
𝑑𝑥
= 𝑑𝑓
𝑑𝑥
can be rewritten in the form
𝑑(𝑓⋆ 𝑦) = 𝑓⋆ 𝑑𝑦
and (1) can be written more suggestively as

(2) ∫ 𝑓⋆ (𝜙 𝑑𝑦) = ∫ 𝜙 𝑑𝑦
𝑈 𝑉

One of the goals of this text on differential forms is to legitimize this interpretation of equa-
tion (1) in 𝑛 dimensions and in fact, more generally, show that an analogue of this formula
is true when 𝑈 and 𝑉 are 𝑛-dimensional manifolds.
Another related goal is to prove an important topological generalization of the change
of variables formula (1). This formula asserts that if we drop the assumption that 𝑓 be a
bijection and just require 𝑓 to be proper (i.e., that pre-images of compact subsets of 𝑉 to be
compact subsets of 𝑈) then the formula (1) can be replaced by

(3) ∫ 𝑓⋆ (𝜙 𝑑𝑦) = deg(𝑓) ∫ 𝜙 𝑑𝑦

𝑈 𝑉

where deg(𝑓) is a topological invariant of 𝑓 that roughly speaking counts, with plus and
minus signs, the number of pre-image points of a generically chosen point of 𝑉.1
This degree formula is just one of a host of results which connect the theory of differ-
ential forms with topology, and one of the main goals of this book will explore some of the
other examples. For instance, for 𝑈 an open subset of 𝐑2 , we define 𝛺0 (𝑈) to be the vector

1It is our feeling that this formula should, like formula (1), be part of the standard calculus curriculum,
particularly in view of the fact that there now exists a beautiful elementary proof of it by Peter Lax (see [6, 8, 9]).

v
 Draft: March 28, 2018

vi Preface

space of 𝐶∞ functions on 𝑈. We define the vector space 𝛺1 (𝑈) to be the space of formal
sums
(4) 𝑓1 𝑑𝑥1 + 𝑓2 𝑑𝑥2 ,
where 𝑓1 , 𝑓2 ∈ 𝐶 (𝑈). We define the vector space 𝛺2 (𝑈) to be the space of expressions of
∞

the form
(5) 𝑓 𝑑𝑥1 ∧ 𝑑𝑥2 ,
where 𝑓 ∈ 𝐶 (𝑈), and for 𝑘 > 2 define 𝛺𝑘 (𝑈) to the zero vector space.
∞

On these vector spaces one can define operators

(6) 𝑑 ∶ 𝛺𝑖 (𝑈) → 𝛺𝑖+1 (𝑈)
by the recipes
𝜕𝑓 𝜕𝑓
(7) 𝑑𝑓 ≔ 𝑑𝑥1 + 𝑑𝑥
𝜕𝑥1 𝜕𝑥2 2
for 𝑖 = 0,
𝜕𝑓2 𝜕𝑓1
(8) 𝑑(𝑓1 𝑑𝑥1 + 𝑓2 𝑑𝑥2 ) = ( − ) 𝑑𝑥1 ∧ 𝑑𝑥
𝜕𝑥1 𝜕𝑥2
for 𝑖 = 1, and 𝑑 = 0 for 𝑖 > 1. It is easy to see that the operator
(9) 𝑑2 ∶ 𝛺𝑖 (𝑈) → 𝛺𝑖+2 (𝑈)
is zero. Hence,
im(𝑑 ∶ 𝛺𝑖−1 (𝑈) → 𝛺𝑖 (𝑈)) ⊂ ker(𝑑 ∶ 𝛺𝑖 (𝑈) → 𝛺𝑖+1 (𝑈)) ,
and this enables one to define the de Rham cohomology groups of 𝑈 as the quotient vector
space
ker(𝑑 ∶ 𝛺𝑖 (𝑈) → 𝛺𝑖+1 (𝑈))
(10) 𝐻𝑖 (𝑈) ≔ .
im(𝑑 ∶ 𝛺𝑖−1 (𝑈) → 𝛺𝑖 (𝑈))
It turns out that these cohomology groups are topological invariants of 𝑈 and are, in
fact, isomorphic to the cohomology groups of 𝑈 defined by the algebraic topologists. More-
over, by slightly generalizing the definitions in equations (4), (5) and (7) to (10) one can
define these groups for open subsets of 𝐑𝑛 and, with a bit more effort, for arbitrary 𝐶∞
manifolds (as we will do in Chapter 5); and their existence will enable us to describe inter-
esting connections between problems in multivariable calculus and differential geometry
on the one hand and problems in topology on the other.
To make the context of this book easier for our readers to access we will devote the
rest of this introduction to the following annotated table of contents, chapter by chapter
descriptions of the topics that we will be covering.

Organization
Chapter 1: Multilinear algebra
As we mentioned above one of our objectives is to legitimatize the presence of the 𝑑𝑥
and 𝑑𝑦 in formula (1), and translate this formula into a theorem about differential forms.
However a rigorous exposition of the theory of differential forms requires a lot of algebraic
preliminaries, and these will be the focus of Chapter 1. We’ll begin, in Sections 1.1 and 1.2, by
reviewing material that we hope most of our readers are already familiar with: the definition
of vector space, the notions of basis, of dimension, of linear mapping, of bilinear form, and
 Draft: March 28, 2018

Organization vii

of dual space and quotient space. Then in Section 1.3 we will turn to the main topics of
this chapter, the concept of 𝑘-tensor and (the future key ingredient in our exposition of
the theory of differential forms in Chapter 2) the concept of alternating 𝑘-tensor. Those 𝑘
tensors come up in fact in two contexts: as alternating 𝑘-tensors, and as exterior forms, i.e.,
in the first context as a subspace of the space of 𝑘-tensors and in the second as a quotient
space of the space of 𝑘-tensors. Both descriptions of 𝑘-tensors will be needed in our later
applications. For this reason the second half of Chapter 1 is mostly concerned with exploring
the relationships between these two descriptions and making use of these relationships to
define a number of basic operations on exterior forms such as the wedge product operation
(see §1.6), the interior product operation (see §1.7) and the pullback operation (see §1.8).
We will also make use of these results in Section 1.9 to define the notion of an orientation for
an 𝑛-dimensional vector space, a notion that will, among other things, enable us to simplify
the change of variables formula (1) by getting rid of the absolute value sign in the term
| det 𝐽𝑓 |.

Chapter 2: Differential Forms

The expressions in equations (4), (5), (7) and (8) are typical examples of differential
forms, and if this were intended to be a text for undergraduate physics majors we would
define differential forms by simply commenting that they’re expressions of this type. We’ll
begin this chapter, however, with the following more precise definition: Let 𝑈 be an open
subset of 𝐑𝑛 . Then a 𝑘-form 𝜔 on 𝑈 is a “function” which to each 𝑝 ∈ 𝑈 assigns an element
of 𝛬𝑘 (𝑇𝑝⋆ 𝑈), 𝑇𝑝 𝑈 being the tangent space to 𝑈 at 𝑝, 𝑇𝑝⋆ 𝑈 its vector space dual, and 𝛬𝑘 (𝑇𝑝⋆ )
the 𝑘th order exterior power of 𝑇𝑝⋆ 𝑈. (It turns out, fortunately, not to be too hard to recon-
cile this definition with the physics definition above.) Differential 1-forms are perhaps best
understood as the dual objects to vector fields, and in Sections 2.1 and 2.2 we elaborate on
this observation, and recall for future use some standard facts about vector fields and their
integral curves. Then in Section 2.3 we will turn to the topic of 𝑘-forms and in the exercises
at the end of Section 2.3 discuss a lot of explicit examples (that we strongly urge readers of
this text to stare at). Then in Sections 2.4 to 2.7 we will discuss in detail three fundamen-
tal operations on differential forms of which we’ve already gotten preliminary glimpses in
the equation (3) and equations (5) to (9), namely the wedge product operation, the exterior
differential operation, and the pullback operation. Also, to add to this list in Section 2.5 we
will discuss the interior product operation of vector fields on differential forms, a general-
ization of the duality pairing of vector fields with one-forms that we alluded to earlier. (In
order to get a better feeling for this material we strongly recommend that one take a look
at Section 2.7 where these operations are related to the div, curl, and grad operations in
freshman calculus.) In addition this section contains some interesting applications of the
material above to physics, in Section 2.8 to electrodynamics and Maxwell’s equation, as well
as to classical mechanisms and the Hamilton–Jacobi equations.

Chapter 3: Integration of Forms

As we mentioned above, the change of variables formula in integral calculus is a special
case of a more general result: the degree formula; and we also cited a paper of Peter Lax
which contains an elementary proof of this formula which will hopefully induce future au-
thors of elementary text books in multivariate calculus to include it in their treatment of the
Riemann integral. In this chapter we will also give a proof of this result but not, regrettably,
Lax’s proof. The reason why not is that we want to relate this result to another result which
 Draft: March 28, 2018

viii Preface

will have some important de Rham theoretic applications when we get to Chapter 5. To de-
scribe this result let 𝑈 be a connected open set in 𝐑𝑛 and 𝜔 = 𝑓𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 a compactly
supported 𝑛-form on 𝑈. We will prove:
Theorem 11. The integral ∫𝑈 𝜔 = ∫𝑈 𝑓𝑑𝑥1 ⋯ 𝑑𝑥𝑛 of 𝜔 over 𝑈 is zero if and only if 𝜔 = 𝑑𝜈
where 𝜈 is a compactly supported (𝑛 − 1)-form on 𝑈.
An easy corollary of this result is the following weak version of the degree theorem:
Corollary 12. Let 𝑈 and 𝑉 be connected open subsets of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝑉 a proper mapping.
Then there exists a constant 𝛿(𝑓) with the property that for every compactly supported 𝑛-form
𝜔 on 𝑉
∫ 𝑓⋆ 𝜔 = 𝛿(𝑓) ∫ 𝜔 .
𝑈 𝑉
Thus to prove the degree theorem it suffices to show that this 𝛿(𝑓) is the deg(𝑓) in formula
(3) and this turns out not to be terribly hard.
The degree formula has a lot of interesting applications and at the end of Chapter 3 we
will describe two of them. The first, the Brouwer fixed point theorem, asserts that if 𝐵𝑛 is
the closed unit ball in 𝐑𝑛 and 𝑓 ∶ 𝐵𝑛 → 𝐵𝑛 is a continuous map, then 𝑓 has a fixed point,
i.e. there is a point 𝑥0 ∈ 𝐵𝑛 that gets mapped onto itself by 𝑓. (We will show that if this were
not true the degree theorem would lead to an egregious contradiction.)
The second is the fundamental theorem of algebra. If
𝑝(𝑧) = 𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑛−1 𝑧𝑛−1 + 𝑧𝑛
is a polynomial function of the complex variable 𝑧, then it has to have a complex root 𝑧0
satisfying 𝑝(𝑧0 ) = 0. (We’ll show that both these results are more-or-less one line corolaries
of the degree theorem.)
Chapter 4: Forms on Manifolds
In the previous two chapters the differential forms that we have been considering have
been forms defined on open subsets of 𝐑𝑛 . In this chapter we’ll make the transition to forms
defined on manifolds; and to prepare our readers for this transition, include at the beginning
of this chapter a brief introduction to the theory of manifolds. (It is a tradition in under-
graduate courses to define 𝑛-dimensional manifolds as 𝑛-dimensional submanifolds of 𝐑𝑁 ,
i.e., as 𝑛-dimensional generalizations of curves and surfaces in 𝐑3 , whereas the tradition in
graduate level courses is to define them as abstract entities. Since this is intended to be a text
book for undergraduates, we’ll adopt the first of these approaches, our reluctance to doing
so being somewhat tempered by the fact that, thanks to the Whitney embedding theorem,
these two approaches are the same.) In section Section 4.1 we will review a few general facts
about manifolds, in section Section 4.2, define the crucial notion of the tangent space 𝑇𝑝 𝑋
to a manifold 𝑋 at a point 𝑝 and in sections Sections 4.3 and 4.4 show how to define dif-
ferential forms on 𝑋 by defining a 𝑘-form to be, as in § 2.4, a function 𝜔 which assigns to
each 𝑝 ∈ 𝑋 an element 𝜔𝑝 of 𝛬𝑘 (𝑇𝑝⋆ 𝑋). Then in Section 4.4 we will define what it means
for a manifold to be oriented and in Section 4.5 show that if 𝑋 is an oriented 𝑛-dimensional
manifold and 𝜔 a compactly supported 𝑛-form, the integral of 𝜔 is well-defined. Moreover,
we’ll prove in this section a manifold version of the change of variables formula and in Sec-
tion 4.6 prove manifold versions of two standard results in integral calculus: the divergence
theorem and Stokes theorem, setting the stage for the main results of Chapter 4: the mani-
fold version of the degree theorem (see §4.7) and a number of applications of this theorem
(among them the Jordan–Brouwer separation theorem, the Gauss–Bonnet theorem and the
Index theorem for vector fields (see §§4.8 and 4.9).
 Draft: March 28, 2018

Organization ix

Chapter 5: Cohomology via Forms

Given an 𝑛-dimensional manifold let 𝛺𝑘 (𝑋) be the space of differential forms on 𝑋 of
degree 𝑘, let 𝑑 ∶ 𝛺𝑘 (𝑋) → 𝛺𝑘+1 (𝑋) be exterior differentiation and let 𝐻𝑘 (𝑋) be the 𝑘th de
Rham cohomology group of 𝑋: the quotient of the vector space
𝑍𝑘 (𝑋) ≔ ker(𝑑 ∶ 𝛺𝑘 (𝑋) → 𝛺𝑘+1 (𝑋))
by the vector space
𝐵𝑘 (𝑋) ≔ im(𝑑 ∶ 𝛺𝑘−1 (𝑋) → 𝛺𝑘 (𝑋)) .
In Chapter 5 we will make a systematic study of these groups and also show that some of
the results about differential forms that we proved in earlier chapters can be interpreted as
cohomological properties of these groups. For instance, in Section 5.1 we will show that
the wedge product and pullback operations on forms that we discussed in Chapter 2 gives
rise to analogous operations in cohomology and that for a compact oriented 𝑛-dimensional
manifold the integration operator on forms that we discussed in Chapter 4 gives rise to a
pairing
𝐻𝑘 (𝑋) × 𝐻𝑛−𝑘 (𝑋) → 𝐑 ,
and in fact more generally if 𝑋 is non-compact, a pairing
(13) 𝐻𝑐𝑘 (𝑋) × 𝐻𝑛−𝑘 (𝑋) → 𝐑 ,
where the groups 𝐻𝑐𝑖 (𝑋) are the DeRham cohomology groups that one gets by replacing
the spaces of differential forms 𝛺𝑖 (𝑋) by the corresponding spaces 𝛺𝑖𝑐 (𝑋) of compactly sup-
ported differential forms.
One problem one runs into in the definition of these cohomology groups is that since
the spaces 𝛺𝑘 (𝑋) and 𝛺𝑘𝑐 (𝑋) are infinite dimensional in general, there’s no guarantee that the
same won’t be the case for the spaces 𝐻𝑘 (𝑋) and 𝐻𝑐𝑘 (𝑋), and we’ll address this problem in
Section 5.3. More explicitly, we will say that 𝑋 has finite topology if it admits a finite covering
by open sets 𝑈1 , … 𝑈𝑁 such that for every multi-index 𝐼 = (𝑖1 , … , 𝑖𝑘 ), where 1 ≤ 𝑖𝑟 ≤ 𝑁,
the intersection
(14) 𝑈𝐼 ≔ 𝑈𝑖1 ∩ ⋯ ∩ 𝑈𝑖𝑘
is either empty or is diffeomorphic to a convex open subset of 𝐑𝑛 . We will show that if 𝑋 has
finite topology, then its cohomology groups are finite dimensional and, secondly, we will
show that many manifolds have finite topology. (For instance, if 𝑋 is compact then 𝑋 has
finite topology; for details see §§5.2 and 5.3.)
We mentioned above that if 𝑋 is not compact it has two types of cohomology groups:
the groups 𝐻𝑘 (𝑋) and the groups 𝐻𝑐𝑛−𝑘 (𝑋). In Section 5.5 we will show that if 𝑋 is oriented
then for 𝜈 ∈ 𝛺𝑘 (𝑋) and 𝜔 ∈ 𝛺𝑛−𝑘 𝑐 (𝑋) the integration operation

(𝜈, 𝜔) ↦ ∫ 𝜈 ∧ 𝜔
𝑋
gives rise to a pairing
𝐻𝑘 (𝑋) × 𝐻𝑐𝑛−𝑘 (𝑋) → 𝐑 ,
and that if 𝑋 is connected this pairing is non-degenerate, i.e., defines a bijective linear trans-
formation
𝐻𝑐𝑛−𝑘 (𝑋) ⥲ 𝐻𝑘 (𝑋)⋆ .
This results in the Poincaré duality theorem and it has some interesting implications which
we will explore in Sections 5.5 to 5.7. For instance in Section 5.5 we will show that if 𝑌 and
𝑍 are closed oriented submanifolds of 𝑋 and 𝑍 is compact and of codimension equal to
 Draft: March 28, 2018

x Preface

the dimension of 𝑌, then the intersection number 𝐼(𝑌, 𝑍) of 𝑌 and 𝑍 in 𝑋 is well-defined

(no matter how badly 𝑌 and 𝑍 intersect). In Section 5.6 we will show that if 𝑋 is a compact
oriented manifold and 𝑓 ∶ 𝑋 → 𝑋 a 𝐶∞ map one can define the Lefschetz number of 𝑓 as
the intersection number 𝐼(𝛥𝑋 , graph(𝑓)) where 𝛥𝑋 is the diagonal in 𝑋 × 𝑋, and view this
as a “topological count” of the number of fixed points of the map 𝑓.
Finally in Section 5.8 we will show that if 𝑋 has finite topology (in the sense we de-
scribed above), then one can define an alternative set of cohomology groups for 𝑋, the Čech
cohomology groups of a cover U, and we will sketch a proof of the association that this Čech
cohomology and de Rham cohomology coincide, leaving a lot of details as exercises (how-
ever, with some hints supplied).

Appendices
In Appendix a we review some techniques in multi-variable calculus that enable one to
reduce global problems to local problems and illustrate these techniques by describing some
typical applications. In particular we show how these techniques can be used to make sense
of “improper integrals” and to extend differentiable maps 𝑓 ∶ 𝑋 → 𝐑 on arbitrary subsets
𝑋 ⊂ 𝐑𝑚 to open neighborhoods of these sets.
In Appendix b we discuss another calculus issue: how to solve systems of equations

{ 𝑓 (𝑥) = 𝑦1
{ 1
{ ⋮
{
𝑓
{ 𝑁 (𝑥) = 𝑦𝑁

for 𝑥 in terms of 𝑦 (here 𝑓1 , …, 𝑓𝑁 are differentiable functions on an open subset of 𝐑𝑛 ). To

answer this question we prove a somewhat refined version of what in elementary calculus
texts is referred to as the implicit function theorem and derive as corollaries of this theorem
two results that we make extensive use of in Chapter 4: the canonical submersion theorem
and the canonical immersion theorem.
Finally, in Appendix c we discuss a concept which plays an important role in the for-
mulation of the “finite topology” results that we discussed in Chapter 5, namely the concept
of a “good cover”. In more detail let 𝑋 be an 𝑛-dimensional manifold an let U = {𝑈𝛼 }𝛼∈𝐼 be
a collection of open subsets of 𝑋 with the property that ⋃𝛼∈𝐼 𝑈𝛼 = 𝑋.
Then U is a good cover of 𝑋 if for every finite subset
{𝛼1 , …, 𝛼𝑟 } ⊂ 𝐼
the intersection 𝑈𝛼1 ∩ ⋯ ∩ 𝑈𝛼𝑟 is either empty or is diffeomorphic to a convex open subset
of 𝐑𝑛 . In Appendix c we will prove that good covers always exist.

Notational Conventions
Below we provide a list of a few of our common notational conventions.
≔ used to define a term; the term being defined appears to the left of the colon
⊂ subset
→ denotes an map (of sets, vector spaces, etc.)
↪ inclusion map
↠ surjective map
⥲ a map that is an isomorphism
∖ difference of sets: if 𝐴 ⊂ 𝑋, then 𝑋 ∖ 𝐴 is the complement of 𝐴 in 𝑋
 Draft: March 28, 2018

Acknowledgments xi

Acknowledgments
To conclude this introduction we would like to thank Cole Graham and Farzan Vafa
for helping us correct earlier versions of this text, a task which involved compiling long list
of typos and errata. (In addition a lot of pertinent comments of theirs helped us improve
the readability of the text itself.)
Draft: March 28, 2018
 Draft: March 28, 2018

CHAPTER 1

Multilinear Algebra

1.1. Background
We will list below some definitions and theorems that are part of the curriculum of a
standard theory-based course in linear algebra.1
Definition 1.1.1. A (real) vector space is a set 𝑉 the elements of which we refer to as vectors.
The set 𝑉 is equipped with two vector space operations:
(1) Vector space addition: Given vectors 𝑣1 , 𝑣2 ∈ 𝑉 one can add them to get a third vector,
𝑣1 + 𝑣2 .
(2) Scalar multiplication: Given a vector 𝑣 ∈ 𝑉 and a real number 𝜆, one can multiply 𝑣 by
𝜆 to get a vector 𝜆𝑣.
These operations satisfy a number of standard rules: associativity, commutativity, dis-
tributive laws, etc. which we assume you’re familiar with. (See Exercise 1.1.i.) In addition
we assume that the reader is familiar with the following definitions and theorems.
(1) The zero vector in 𝑉 is the unique vector 0 ∈ 𝑉 with the property that for every vector
𝑣 ∈ 𝑉, we have 𝑣 + 0 = 0 + 𝑣 = 𝑣 and 𝜆𝑣 = 0 if 𝜆 ∈ 𝐑 is nonzero.
(2) Linear independence: A (finite) set of vectors, 𝑣1 , …, 𝑣𝑘 ∈ 𝑉 is linearly independent if
the map
(1.1.2) 𝐑𝑘 → 𝑉 , (𝑐1 , …, 𝑐𝑘 ) ↦ 𝑐1 𝑣1 + ⋯ + 𝑐𝑘 𝑣𝑘
is injective.
(3) A (finite) set of vectors 𝑣1 , …, 𝑣𝑘 ∈ 𝑉 spans 𝑉 if the map (1.1.2) is surjective.
(4) Vectors 𝑣1 , …, 𝑣𝑘 ∈ 𝑉 are a basis of 𝑉 if they span 𝑉 and are linearly independent; in
other words, if the map (1.1.2) is bijective. This means that every vector 𝑣 can be written
uniquely as a sum
𝑛
(1.1.3) 𝑣 = ∑ 𝑐𝑖 𝑣𝑖 .
𝑖=1
(5) If 𝑉 is a vector space with a basis 𝑣1 , …, 𝑣𝑘 , then 𝑉 is said to be finite dimensional, and 𝑘
is the dimension of 𝑉. (It is a theorem that this definition is legitimate: every basis has to
have the same number of vectors.) In this chapter all the vector spaces we’ll encounter
will be finite dimensional. We write dim(𝑉) for the dimension of 𝑉.
(6) A subset 𝑈 ⊂ 𝑉 is a subspace if it is vector space in its own right, i.e., for all 𝑣, 𝑣1 , 𝑣2 ∈ 𝑈
and 𝜆 ∈ 𝐑, both 𝜆𝑣 and 𝑣1 + 𝑣2 are in 𝑈 (where the addition and scalar multiplication
is given as vectors of 𝑉).
(7) Let 𝑉 and 𝑊 be vector spaces. A map 𝐴 ∶ 𝑉 → 𝑊 is linear if for 𝑣, 𝑣1 , 𝑣2 ∈ 𝑈 and 𝜆 ∈ 𝐑
we have
(1.1.4) 𝐴(𝜆𝑣) = 𝜆𝐴𝑣
1Such a course is a prerequisite for reading this text.

1
 Draft: March 28, 2018

2 Chapter 1: Multilinear Algebra

and
𝐴(𝑣1 + 𝑣2 ) = 𝐴𝑣1 + 𝐴𝑣2 .
(8) Let 𝐴 ∶ 𝑉 → 𝑊 be a linear map of vector spaces. The kernel of 𝐴 is the subset of 𝑉
defined by
ker(𝐴) ≔ { 𝑣 ∈ 𝑉 | 𝐴(𝑣) = 0 } ,
i.e, is the set of vectors in 𝑉 which 𝐴 sends to the zero vector in 𝑊. By equation (1.1.4)
and item 7, the subset ker(𝐴) is a subspace of 𝑉.
(9) By (1.1.4) and (7) the set-theoretic image im(𝐴)of 𝐴 is a subspace of 𝑊. We call im(𝐴)
the image 𝐴 of 𝐴. The following is an important rule for keeping track of the dimensions
of ker 𝐴 and im 𝐴:
(1.1.5) dim(𝑉) = dim(ker(𝐴)) + dim(im(𝐴)) .
(10) Linear mappings and matrices: Let 𝑣1 , …, 𝑣𝑛 be a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 a basis of 𝑊.
Then by (1.1.3) 𝐴𝑣𝑗 can be written uniquely as a sum,
𝑚
(1.1.6) 𝐴𝑣𝑗 = ∑ 𝑐𝑖,𝑗 𝑤𝑖 , 𝑐𝑖,𝑗 ∈ 𝐑 .
𝑖=1
The 𝑚 × 𝑛 matrix of real numbers, [𝑐𝑖,𝑗 ], is the matrix associated with 𝐴. Conversely,
given such an 𝑚 × 𝑛 matrix, there is a unique linear map 𝐴 with the property (1.1.6).
(11) An inner product on a vector space is a map
𝐵∶ 𝑉 × 𝑉 → 𝐑
with the following three properties:
(a) Bilinearity: For vectors 𝑣, 𝑣1 , 𝑣2 , 𝑤 ∈ 𝑉 and 𝜆 ∈ 𝐑 we have
𝐵(𝑣1 + 𝑣2 , 𝑤) = 𝐵(𝑣1 , 𝑤) + 𝐵(𝑣2 , 𝑤)
and
𝐵(𝜆𝑣, 𝑤) = 𝜆𝐵(𝑣, 𝑤) .
(b) Symmetry: For vectors 𝑣, 𝑤 ∈ 𝑉 we have 𝐵(𝑣, 𝑤) = 𝐵(𝑤, 𝑣).
(c) Positivity: For every vector 𝑣 ∈ 𝑉 we have 𝐵(𝑣, 𝑣) ≥ 0. Moreover, if 𝑣 ≠ 0 then
𝐵(𝑣, 𝑣) > 0.
Remark 1.1.7. Notice that by property (11b), property (11a) is equivalent to
𝐵(𝑤, 𝜆𝑣) = 𝜆𝐵(𝑤, 𝑣)
and
𝐵(𝑤, 𝑣1 + 𝑣2 ) = 𝐵(𝑤, 𝑣1 ) + 𝐵(𝑤, 𝑣2 ) .
Example 1.1.8. The map (1.1.2) is a linear map. The vectors 𝑣1 , …, 𝑣𝑘 span 𝑉 if its image of
this map is 𝑉, and the 𝑣1 , …, 𝑣𝑘 are linearly independent if the kernel of this map is the zero
vector in 𝐑𝑘 .
The items on the list above are just a few of the topics in linear algebra that we’re assum-
ing our readers are familiar with. We have highlighted them because they’re easy to state.
However, understanding them requires a heavy dollop of that indefinable quality “mathe-
matical sophistication”, a quality which will be in heavy demand in the next few sections of
this chapter. We will also assume that our readers are familiar with a number of more low-
brow linear algebra notions: matrix multiplication, row and column operations on matrices,
transposes of matrices, determinants of 𝑛 × 𝑛 matrices, inverses of matrices, Cramer’s rule,
recipes for solving systems of linear equations, etc. (See [10, §§1.1 & 1.2] for a quick review
of this material.)
 Draft: March 28, 2018

§1.2 Quotient spaces & dual spaces 3

Exercises for §1.1

Exercise 1.1.i. Our basic example of a vector space in this course is 𝐑𝑛 equipped with the
vector addition operation
(𝑎1 , …, 𝑎𝑛 ) + (𝑏1 , …, 𝑏𝑛 ) = (𝑎1 + 𝑏1 , …, 𝑎𝑛 + 𝑏𝑛 )
and the scalar multiplication operation
𝜆(𝑎1 , …, 𝑎𝑛 ) = (𝜆𝑎1 , …, 𝜆𝑎𝑛 ) .
Check that these operations satisfy the axioms below.
(1) Commutativity: 𝑣 + 𝑤 = 𝑤 + 𝑣.
(2) Associativity: 𝑢 + (𝑣 + 𝑤) = (𝑢 + 𝑣) + 𝑤.
(3) For the zero vector 0 ≔ (0, …, 0) we have 𝑣 + 0 = 0 + 𝑣.
(4) 𝑣 + (−1)𝑣 = 0.
(5) 1𝑣 = 𝑣.
(6) Associative law for scalar multiplication: (𝑎𝑏)𝑣 = 𝑎(𝑏𝑣).
(7) Distributive law for scalar addition: (𝑎 + 𝑏)𝑣 = 𝑎𝑣 + 𝑏𝑣.
(8) Distributive law for vector addition: 𝑎(𝑣 + 𝑤) = 𝑎𝑣 + 𝑎𝑤.
Exercise 1.1.ii. Check that the standard basis vectors of 𝐑𝑛 : 𝑒1 = (1, 0, …, 0), 𝑒2 = (0, 1, 0, …, 0),
etc. are a basis.
Exercise 1.1.iii. Check that the standard inner product on 𝐑𝑛
𝑛
𝐵((𝑎1 , …, 𝑎𝑛 ), (𝑏1 , …, 𝑏𝑛 )) = ∑ 𝑎𝑖 𝑏𝑖
𝑖=1
is an inner product.

1.2. Quotient spaces & dual spaces

In this section we will discuss a couple of items which are frequently, but not always,
covered in linear algebra courses, but which we’ll need for our treatment of multilinear al-
gebra in §§1.3 to 1.8.
The quotient spaces of a vector space
Definition 1.2.1. Let 𝑉 be a vector space and 𝑊 a vector subspace of 𝑉. A 𝑊-coset is a set
of the form
𝑣 + 𝑊 ≔ {𝑣 + 𝑤|𝑤 ∈ 𝑊} .
It is easy to check that if 𝑣1 − 𝑣2 ∈ 𝑊, the cosets 𝑣1 + 𝑊 and 𝑣2 + 𝑊, coincide while if
𝑣1 −𝑣2 ∉ 𝑊, the cosets 𝑣1 +𝑊 and 𝑣2 +𝑊 are disjoint. Thus the distinct 𝑊-cosets decompose
𝑉 into a disjoint collection of subsets of 𝑉.
Notation 1.2.2. Let 𝑉 be a vector space and 𝑊 ⊂ 𝑉 a subspace. We write 𝑉/𝑊 for the set
of distinct 𝑊-cosets in 𝑉.
Definition 1.2.3. Let 𝑉 be a vector space and 𝑊 ⊂ 𝑉 a subspace. Define a vector addition
operation on 𝑉/𝑊 by setting
(1.2.4) (𝑣1 + 𝑊) + (𝑣2 + 𝑊) ≔ (𝑣1 + 𝑣2 ) + 𝑊
and define a scalar multiplication operation on 𝑉/𝑊 by setting
(1.2.5) 𝜆(𝑣 + 𝑊) ≔ (𝜆𝑣) + 𝑊 .
These operations make 𝑉/𝑊 into a vector space, called quotient space of 𝑉 by 𝑊.
 Draft: March 28, 2018

4 Chapter 1: Multilinear Algebra

It is easy to see that the operations on 𝑉/𝑊 from Definition 1.2.3 are well-defined. For
instance, suppose 𝑣1 + 𝑊 = 𝑣′1 + 𝑊 and 𝑣2 + 𝑊 = 𝑣′2 + 𝑊. Then 𝑣1 − 𝑣′1 and 𝑣2 − 𝑣′2 are in
𝑊, so
(𝑣1 + 𝑣2 ) − (𝑣′1 + 𝑣′2 ) ∈ 𝑊 ,
and hence (𝑣1 + 𝑣2 ) + 𝑊 = (𝑣′1 + 𝑣′2 ) + 𝑊.
Definition 1.2.6. Let 𝑉 be a vector space and 𝑊 ⊂ 𝑉 a subspace. Define a map quotient
map
(1.2.7) 𝜋 ∶ 𝑉 → 𝑉/𝑊
by setting 𝜋(𝑣) ≔ 𝑣 + 𝑊. It is clear from equations (1.2.4) and (1.2.5) that 𝜋 is linear, and
that it maps 𝑉 surjectively onto 𝑉/𝑊.
Observation 1.2.8. Note that the zero vector in the vector space 𝑉/𝑊 is the zero coset
0 + 𝑊 = 𝑊. Hence 𝑣 ∈ ker(𝜋) if and only if 𝑣 + 𝑊 = 𝑊, i.e., 𝑣 ∈ 𝑊. In other words,
ker(𝜋) = 𝑊.
In the Definitions 1.2.3 and 1.2.6, 𝑉 and 𝑊 do not have to be finite dimensional, but if
they are then by equation (1.1.5) we have
dim(𝑉/𝑊) = dim(𝑉) − dim(𝑊) .
We leave the following easy proposition as an exercise.
Proposition 1.2.9. Let 𝐴 ∶ 𝑉 → 𝑈 a linear map of vector spaces. If 𝑊 ⊂ ker(𝐴) there exists
a unique linear map 𝐴♯ ∶ 𝑉/𝑊 → 𝑈 with the property that 𝐴 = 𝐴♯ ∘ 𝜋, where 𝜋 ∶ 𝑉 → 𝑉/𝑊
is the quotient map.
The dual space of a vector space
Definition 1.2.10. Let 𝑉 be a vector space. Write 𝑉⋆ the set of all linear functions ℓ ∶ 𝑉 → 𝐑.
Define a vector space structure on 𝑉⋆ as follows: if ℓ1 , ℓ2 ∈ 𝑉⋆ , then define the sum ℓ1 + ℓ2
to be the map 𝑉 → 𝐑 given by
(ℓ1 + ℓ2 )(𝑣) ≔ ℓ1 (𝑣) + ℓ2 (𝑣) ,
which is clearly is linear. If ℓ ∈ 𝑉⋆ is a linear function and 𝜆 ∈ 𝐑, define 𝜆ℓ by
(𝜆ℓ)(𝑣) ≔ 𝜆 ⋅ ℓ(𝑣) ;
then 𝜆ℓ is clearly linear.
The vector space 𝑉⋆ is called the dual space of 𝑉.
Suppose 𝑉 is 𝑛-dimensional, and let 𝑒1 , …, 𝑒𝑛 be a basis of 𝑉. Then every vector 𝑣 ∈ 𝑉
can be written uniquely as a sum
𝑣 = 𝑐1 𝑒1 + ⋯ + 𝑐𝑛 𝑒𝑛 , 𝑐𝑖 ∈ 𝐑 .
Let
(1.2.11) 𝑒𝑖⋆ (𝑣) = 𝑐𝑖 .
If 𝑣 = 𝑐1 𝑒1 + ⋯ + 𝑐𝑛 𝑒𝑛 and 𝑣′ = 𝑐1′ 𝑒1 + ⋯ + 𝑐𝑛′ 𝑒𝑛 then 𝑣 + 𝑣′ = (𝑐1 + 𝑐1′ )𝑒1 + ⋯ + (𝑐𝑛 + 𝑐𝑛′ )𝑒𝑛 , so
𝑒𝑖⋆ (𝑣 + 𝑣′ ) = 𝑐𝑖 + 𝑐𝑖′ = 𝑒𝑖⋆ (𝑣) + 𝑒𝑖⋆ (𝑣′ ) .
This shows that 𝑒𝑖⋆ (𝑣) is a linear function of 𝑣 and hence 𝑒𝑖⋆ ∈ 𝑉⋆ .
Claim 1.2.12. If 𝑉 is an 𝑛-dimensional vector space with basis 𝑒1 , …, 𝑒𝑛 , then 𝑒1⋆ , …, 𝑒𝑛⋆ is a
basis of 𝑉⋆ .
 Draft: March 28, 2018

§1.2 Quotient spaces & dual spaces 5

Proof. First of all note that by (1.2.11)

1, 𝑖 = 𝑗
(1.2.13) 𝑒𝑖⋆ (𝑒𝑗 ) = {
0, 𝑖 ≠ 𝑗 .
If ℓ ∈ 𝑉⋆ let 𝜆𝑖 = ℓ(𝑒𝑖 ) and let ℓ′ = ∑ 𝜆𝑖 𝑒𝑖⋆ . Then by (1.2.13)
𝑛
ℓ′ (𝑒𝑗 ) = ∑ 𝜆𝑖 𝑒𝑖⋆ (𝑒𝑗 ) = 𝜆𝑗 = ℓ(𝑒𝑗 ) ,
𝑖=1

i.e., ℓ and ℓ take identical values on the basis vectors, 𝑒𝑗 . Hence ℓ = ℓ′ .

′
𝑛 𝑛
Suppose next that ∑𝑖=1 𝜆𝑖 𝑒𝑖⋆ = 0. Then by (1.2.13), 𝜆𝑗 = (∑𝑖=1 𝜆𝑖 𝑒𝑖⋆ )(𝑒𝑗 ) = 0 for 𝑗 =
⋆ ⋆
1, …, 𝑛. Hence the vectors 𝑒1 , …, 𝑒𝑛 are linearly independent. □
Let 𝑉 and 𝑊 be vector spaces and 𝐴 ∶ 𝑉 → 𝑊, a linear map. Given ℓ ∈ 𝑊⋆ the com-
position, ℓ ∘ 𝐴, of 𝐴 with the linear map, ℓ ∶ 𝑊 → 𝐑, is linear, and hence is an element of
𝑉⋆ . We will denote this element by 𝐴⋆ ℓ, and we will denote by
𝐴⋆ ∶ 𝑊⋆ → 𝑉⋆
the map, ℓ ↦ 𝐴⋆ ℓ. It is clear from the definition that
𝐴⋆ (ℓ1 + ℓ2 ) = 𝐴⋆ ℓ1 + 𝐴⋆ ℓ2
and that
𝐴⋆ (𝜆ℓ) = 𝜆𝐴⋆ ℓ ,
⋆
i.e., that 𝐴 is linear.
Definition 1.2.14. Let 𝑉 and 𝑊 be vector spaces and 𝐴 ∶ 𝑉 → 𝑊, a linear map. We call the
map 𝐴⋆ ∶ 𝑊⋆ → 𝑉⋆ defined above the transpose of the map 𝐴.
We conclude this section by giving a matrix description of 𝐴⋆ . Let 𝑒1 , …, 𝑒𝑛 be a basis
of 𝑉 and 𝑓1 , …, 𝑓𝑚 a basis of 𝑊; let 𝑒1⋆ , …, 𝑒𝑛⋆ and 𝑓1⋆ , …, 𝑓𝑚⋆ be the dual bases of 𝑉⋆ and
𝑊⋆ . Suppose 𝐴 is defined in terms of 𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑚 by the 𝑚 × 𝑛 matrix, [𝑎𝑖,𝑗 ], i.e.,
suppose
𝑛
𝐴𝑒𝑗 = ∑ 𝑎𝑖,𝑗 𝑓𝑖 .
𝑖=1

Claim 1.2.15. The linear map 𝐴 is defined, in terms of 𝑓1⋆ , …, 𝑓𝑚⋆ and 𝑒1⋆ , …, 𝑒𝑛⋆ by the trans-
⋆

pose matrix (𝑎𝑗,𝑖 ).

Proof. Let
𝑚
𝐴⋆ 𝑓𝑖⋆ = ∑ 𝑐𝑗,𝑖 𝑒𝑗⋆ .
𝑗=1
Then
𝑛
𝐴⋆ 𝑓𝑖⋆ (𝑒𝑗 ) = ∑ 𝑐𝑘,𝑖 𝑒𝑘⋆ (𝑒𝑗 ) = 𝑐𝑗,𝑖
𝑘=1
by (1.2.13). On the other hand
𝑚 𝑚
𝐴⋆ 𝑓𝑖⋆ (𝑒𝑗 ) = 𝑓𝑖⋆ (𝐴𝑒𝑗 ) = 𝑓𝑖⋆ ( ∑ 𝑎𝑘,𝑗 𝑓𝑘 ) = ∑ 𝑎𝑘,𝑗 𝑓𝑖⋆ (𝑓𝑘 ) = 𝑎𝑖,𝑗
𝑘=1 𝑘=1
so 𝑎𝑖,𝑗 = 𝑐𝑗,𝑖 . □
 Draft: March 28, 2018

6 Chapter 1: Multilinear Algebra

Exercises for §1.2

Exercise 1.2.i. Let 𝑉 be an 𝑛-dimensional vector space and 𝑊 a 𝑘-dimensional subspace.
Show that there exists a basis, 𝑒1 , …, 𝑒𝑛 of 𝑉 with the property that 𝑒1 , …, 𝑒𝑘 is a basis of 𝑊.
Hint: Induction on 𝑛 − 𝑘. To start the induction suppose that 𝑛 − 𝑘 = 1. Let 𝑒1 , …, 𝑒𝑛−1
be a basis of 𝑊 and 𝑒𝑛 any vector in 𝑉 ∖ 𝑊.
Exercise 1.2.ii. In Exercise 1.2.i show that the vectors 𝑓𝑖 ≔ 𝜋(𝑒𝑘+𝑖 ), 𝑖 = 1, …, 𝑛 − 𝑘 are a
basis of 𝑉/𝑊, where 𝜋 ∶ 𝑉 → 𝑉/𝑊 is the quotient map.
Exercise 1.2.iii. In Exercise 1.2.i let 𝑈 be the linear span of the vectors 𝑒𝑘+𝑖 for 𝑖 = 1, …, 𝑛−𝑘.
Show that the map
𝑈 → 𝑉/𝑊 , 𝑢 ↦ 𝜋(𝑢) ,
is a vector space isomorphism, i.e., show that it maps 𝑈 bijectively onto 𝑉/𝑊.
Exercise 1.2.iv. Let 𝑈, 𝑉 and 𝑊 be vector spaces and let 𝐴 ∶ 𝑉 → 𝑊 and 𝐵 ∶ 𝑈 → 𝑉 be
linear mappings. Show that (𝐴𝐵)⋆ = 𝐵⋆ 𝐴⋆ .
Exercise 1.2.v. Let 𝑉 = 𝐑2 and let 𝑊 be the 𝑥1 -axis, i.e., the one-dimensional subspace
{ (𝑥1 , 0) | 𝑥1 ∈ 𝐑 }
2
of 𝐑 .
(1) Show that the 𝑊-cosets are the lines, 𝑥2 = 𝑎, parallel to the 𝑥1 -axis.
(2) Show that the sum of the cosets, “𝑥2 = 𝑎” and “𝑥2 = 𝑏” is the coset “𝑥2 = 𝑎 + 𝑏”.
(3) Show that the scalar multiple of the coset, “𝑥2 = 𝑐” by the number, 𝜆, is the coset, “𝑥2 =
𝜆𝑐”.
Exercise 1.2.vi.
(1) Let (𝑉⋆ )⋆ be the dual of the vector space, 𝑉⋆ . For every 𝑣 ∈ 𝑉, let ev𝑣 ∶ 𝑉⋆ → 𝐑 be the
evaluation function ev𝑣 (ℓ) = ℓ(𝑣). Show that the ev𝑣 is a linear function on 𝑉⋆ , i.e., an
element of (𝑉⋆ )⋆ , and show that the map
(1.2.16) ev = ev(−) ∶ 𝑉 → (𝑉⋆ )⋆ , 𝑣 ↦ ev𝑣
is a linear map of 𝑉 into (𝑉⋆ )⋆ .
(2) If 𝑉 is finite dimensional, show that the map (1.2.16) is bijective. Conclude that there is
a natural identification of 𝑉 with (𝑉⋆ )⋆ , i.e., that 𝑉 and (𝑉⋆ )⋆ are two descriptions of
the same object.
Hint: dim(𝑉⋆ )⋆ = dim 𝑉⋆ = dim 𝑉, so by equation (1.1.5) it suffices to show that
(1.2.16) is injective.
Exercise 1.2.vii. Let 𝑊 be a vector subspace of a finite dimensional vector space 𝑉 and let
𝑊⟂ = { ℓ ∈ 𝑉⋆ | ℓ(𝑤) = 0 for all 𝑤 ∈ 𝑊 } .
𝑊⟂ is called the annihilator of 𝑊 in 𝑉⋆ . Show that 𝑊⟂ is a subspace of 𝑉⋆ of dimension
dim 𝑉 − dim 𝑊.
Hint: By Exercise 1.2.i we can choose a basis, 𝑒1 , …, 𝑒𝑛 of 𝑉 such that 𝑒1 , …𝑒𝑘 is a basis
⋆
of 𝑊. Show that 𝑒𝑘+1 , …, 𝑒𝑛⋆ is a basis of 𝑊⟂ .
Exercise 1.2.viii. Let 𝑉 and 𝑉′ be vector spaces and 𝐴 ∶ 𝑉 → 𝑉′ a linear map. Show that if
𝑊 ⊂ ker(𝐴), then there exists a linear map 𝐵 ∶ 𝑉/𝑊 → 𝑉′ with the property that 𝐴 = 𝐵 ∘ 𝜋
(where 𝜋 is the quotient map (1.2.7)). In addition show that this linear map is injective if
and only if ker(𝐴) = 𝑊.
 Draft: March 28, 2018

§1.2 Quotient spaces & dual spaces 7

Exercise 1.2.ix. Let 𝑊 be a subspace of a finite dimensional vector space, 𝑉. From the
inclusion map, 𝜄 ∶ 𝑊⟂ → 𝑉⋆ , one gets a transpose map,
𝜄⋆ ∶ (𝑉⋆ )⋆ → (𝑊⟂ )⋆
and, by composing this with (1.2.16), a map
𝜄⋆ ∘ ev ∶ 𝑉 → (𝑊⟂ )⋆ .
Show that this map is onto and that its kernel is 𝑊. Conclude from Exercise 1.2.viii that
there is a natural bijective linear map
𝜈 ∶ 𝑉/𝑊 → (𝑊⟂ )⋆
with the property 𝜈 ∘ 𝜋 = 𝜄⋆ ∘ ev. In other words, 𝑉/𝑊 and (𝑊⟂ )⋆ are two descriptions of the
same object. (This shows that the “quotient space” operation and the “dual space” operation
are closely related.)
Exercise 1.2.x. Let 𝑉1 and 𝑉2 be vector spaces and 𝐴 ∶ 𝑉1 → 𝑉2 a linear map. Verify that for
the transpose map 𝐴⋆ ∶ 𝑉2⋆ → 𝑉1⋆ we have:
ker(𝐴⋆ ) = im(𝐴)⟂
and
im(𝐴⋆ ) = ker(𝐴)⟂ .
Exercise 1.2.xi. Let 𝑉 be a vector space.
(1) Let 𝐵 ∶ 𝑉 × 𝑉 → 𝐑 be an inner product on 𝑉. For 𝑣 ∈ 𝑉 let
ℓ𝑣 ∶ 𝑉 → 𝐑
be the function: ℓ𝑣 (𝑤) = 𝐵(𝑣, 𝑤). Show that ℓ𝑣 is linear and show that the map
(1.2.17) 𝐿 ∶ 𝑉 → 𝑉⋆ , 𝑣 ↦ ℓ𝑣
is a linear mapping.
(2) If 𝑉 is finite dimensional, prove that 𝐿 bijective. Conclude that if 𝑉 has an inner product
one gets from it a natural identification of 𝑉 with 𝑉⋆ .
Hint: Since dim 𝑉 = dim 𝑉⋆ it suffices by equation (1.1.5) to show that ker(𝐿) = 0.
Now note that if 𝑣 ≠ 0 ℓ𝑣 (𝑣) = 𝐵(𝑣, 𝑣) is a positive number.
Exercise 1.2.xii. Let 𝑉 be an 𝑛-dimensional vector space and 𝐵 ∶ 𝑉 × 𝑉 → 𝐑 an inner
product on 𝑉. A basis, 𝑒1 , …, 𝑒𝑛 of 𝑉 is orthonormal if

1, 𝑖=𝑗
(1.2.18) 𝐵(𝑒𝑖 , 𝑒𝑗 ) = {
0, 𝑖≠𝑗.
(1) Show that an orthonormal basis exists.
Hint: By induction let 𝑒1 , …, 𝑒𝑘 be vectors with the property (1.2.18) and let 𝑣 be a
vector which is not a linear combination of these vectors. Show that the vector
𝑘
𝑤 = 𝑣 − ∑ 𝐵(𝑒𝑖 , 𝑣)𝑒𝑖
𝑖=1
1
is nonzero and is orthogonal to the 𝑒𝑖 ’s. Now let 𝑒𝑘+1 = 𝜆𝑤, where 𝜆 = 𝐵(𝑤, 𝑤)− 2 .
 Draft: March 28, 2018

8 Chapter 1: Multilinear Algebra

(2) Let 𝑒1 , …𝑒𝑛 and 𝑒1′ , …𝑒𝑛′ be two orthonormal bases of 𝑉 and let
𝑛
𝑒𝑗′ = ∑ 𝑎𝑖,𝑗 𝑒𝑖 .
𝑖=1
Show that
𝑛
1, 𝑗 = 𝑘
(1.2.19) ∑ 𝑎𝑖,𝑗 𝑎𝑖,𝑘 = {
𝑖=1 0, 𝑗 ≠ 𝑘 .
(3) Let 𝐴 be the matrix [𝑎𝑖,𝑗 ]. Show that equation (1.2.19) can be written more compactly
as the matrix identity
(1.2.20) 𝐴𝐴⊺ = id𝑛
where id𝑛 is the 𝑛 × 𝑛 identity matrix.
(4) Let 𝑒1 , …, 𝑒𝑛 be an orthonormal basis of 𝑉 and 𝑒1⋆ , …, 𝑒𝑛⋆ the dual basis of 𝑉⋆ . Show that
the mapping (1.2.17) is the mapping, 𝐿𝑒𝑖 = 𝑒𝑖⋆ , 𝑖 = 1, …𝑛.

1.3. Tensors
Definition 1.3.1. Let 𝑉 be an 𝑛-dimensional vector space and let 𝑉𝑘 be the set of all 𝑘-tuples
(𝑣1 , …, 𝑣𝑘 ), where 𝑣1 , …, 𝑣𝑘 ∈ 𝑉, that is, the 𝑘-fold direct sum of 𝑉 with itself. A function
𝑇 ∶ 𝑉𝑘 → 𝐑
is said to be linear in its 𝑖th variable if, when we fix vectors, 𝑣1 , …, 𝑣𝑖−1 , 𝑣𝑖+1 , …, 𝑣𝑘 , the map
𝑉 → 𝐑 defined by
(1.3.2) 𝑣 ↦ 𝑇(𝑣1 , …, 𝑣𝑖−1 , 𝑣, 𝑣𝑖+1 , …, 𝑣𝑘 )
is linear.
If 𝑇 is linear in its 𝑖th variable for 𝑖 = 1, …, 𝑘 it is said to be 𝑘-linear, or alternatively
is said to be a 𝑘-tensor. We write L𝑘 (𝑉) for the set of all 𝑘-tensors in 𝑉. We will agree that
0-tensors are just the real numbers, that is L0 (𝑉) ≔ 𝐑.
Let 𝑇1 , 𝑇2 ∶ 𝑉𝑘 → 𝐑 be linear functions. It is clear from (1.3.2) that if 𝑇1 and 𝑇2 are
𝑘-linear, so is 𝑇1 + 𝑇2 . Similarly if 𝑇 is 𝑘-linear and 𝜆 is a real number, 𝜆𝑇 is 𝑘-linear. Hence
L𝑘 (𝑉) is a vector space. Note that for 𝑘 = 1, “𝑘-linear” just means ‘linear”, so L1 (𝑉) = 𝑉⋆ .
Definition 1.3.3. Let 𝑛 and 𝑘 be positive integers. A multi-index of 𝑛 of length 𝑘 is a 𝑘-tuple
𝐼 = (𝑖1 , …, 𝑖𝑘 ) of integers with 1 ≤ 𝑖𝑟 ≤ 𝑛 for 𝑟 = 1, …, 𝑘.
Example 1.3.4. Let 𝑛 be a positive integer. The multi-indices of 𝑛 of length 2 are in bijection
the square of pairs of integers integers
(𝑖, 𝑗) , 1 ≤ 𝑖, 𝑗 ≤ 𝑛 ,
2
and there are exactly 𝑛 of them.
We leave the following generalization as an easy exercise.
Lemma 1.3.5. Let 𝑛 and 𝑘 be positive integers. There are exactly 𝑛𝑘 multi-indices of 𝑛 of length
𝑘.
Now fix a basis 𝑒1 , …, 𝑒𝑛 of 𝑉. For 𝑇 ∈ L𝑘 (𝑉) write
(1.3.6) 𝑇𝐼 ≔ 𝑇(𝑒𝑖1 , …, 𝑒𝑖𝑘 )
for every multi-index 𝐼 of length 𝑘.
 Draft: March 28, 2018

§1.3 Tensors 9

Proposition 1.3.7. The real numbers 𝑇𝐼 determine 𝑇, i.e., if 𝑇 and 𝑇′ are 𝑘-tensors and 𝑇𝐼 =
𝑇𝐼′ for all 𝐼, then 𝑇 = 𝑇′ .
Proof. By induction on 𝑛. For 𝑛 = 1 we proved this result in § 1.1. Let’s prove that if this
assertion is true for 𝑛 − 1, it is true for 𝑛. For each 𝑒𝑖 let 𝑇𝑖 be the (𝑘 − 1)-tensor
(𝑣1 , …, 𝑣𝑛−1 ) ↦ 𝑇(𝑣1 , …, 𝑣𝑛−1 , 𝑒𝑖 ) .
Then for 𝑣 = 𝑐1 𝑒1 + ⋯ 𝑐𝑛 𝑒𝑛
𝑛
𝑇(𝑣1 , …, 𝑣𝑛−1 , 𝑣) = ∑ 𝑐𝑖 𝑇𝑖 (𝑣1 , …, 𝑣𝑛−1 ) ,
𝑖=1

so the 𝑇𝑖 ’s determine 𝑇. Now apply the inductive hypothesis. □

The tensor product operation

Definition 1.3.8. If 𝑇1 is a 𝑘-tensor and 𝑇2 is an ℓ-tensor, one can define a 𝑘 + ℓ-tensor,
𝑇1 ⊗ 𝑇2 , by setting
(𝑇1 ⊗ 𝑇2 )(𝑣1 , …, 𝑣𝑘+ℓ ) = 𝑇1 (𝑣1 , …, 𝑣𝑘 )𝑇2 (𝑣𝑘+1 , …, 𝑣𝑘+ℓ ) .
This tensor is called the tensor product of 𝑇1 and 𝑇2 .
We note that if 𝑇1 is a 0-tensor, i.e., scalar, then tensor product with 𝑇1 is just scalar
multiplication by 𝑇1 , and similarly if 𝑇2 is a 0-tensor. That is 𝑎 ⊗ 𝑇 = 𝑇 ⊗ 𝑎 = 𝑎𝑇 for 𝑎 ∈ 𝐑
and 𝑇 ∈ L𝑘 (𝑉).
Properties 1.3.9. Suppose that we are given a 𝑘1 -tensor 𝑇1 , a 𝑘2 -tensor 𝑇2 , and a 𝑘3 -tensor
𝑇3 on a vector space 𝑉.
(1) Associativity: One can define a (𝑘1 + 𝑘2 + 𝑘3 )-tensor 𝑇1 ⊗ 𝑇2 ⊗ 𝑇3 by setting
(𝑇1 ⊗ 𝑇2 ⊗ 𝑇3 )(𝑣1 , …, 𝑣𝑘+ℓ+𝑚 ) ≔ 𝑇1 (𝑣1 , …, 𝑣𝑘 )𝑇2 (𝑣𝑘+1 , …, 𝑣𝑘+ℓ )𝑇3 (𝑣𝑘+ℓ+1 , …, 𝑣𝑘+ℓ+𝑚 ) .
Alternatively, one can define 𝑇1 ⊗ 𝑇2 ⊗ 𝑇3 by defining it to be the tensor product of
(𝑇1 ⊗ 𝑇2 ) ⊗ 𝑇3 or the tensor product of 𝑇1 ⊗ (𝑇2 ⊗ 𝑇3 ). It is easy to see that both these
tensor products are identical with 𝑇1 ⊗ 𝑇2 ⊗ 𝑇3 :
(𝑇1 ⊗ 𝑇2 ) ⊗ 𝑇3 = 𝑇1 ⊗ 𝑇2 ⊗ 𝑇3 = 𝑇1 ⊗ (𝑇2 ⊗ 𝑇3 ) .
(2) Distributivity of scalar multiplication: We leave it as an easy exercise to check that if 𝜆 is
a real number then
𝜆(𝑇1 ⊗ 𝑇2 ) = (𝜆𝑇1 ) ⊗ 𝑇2 = 𝑇1 ⊗ (𝜆𝑇2 )
(3) Left and right distributive laws: If 𝑘1 = 𝑘2 , then
(𝑇1 + 𝑇2 ) ⊗ 𝑇3 = 𝑇1 ⊗ 𝑇3 + 𝑇2 ⊗ 𝑇3
and if 𝑘2 = 𝑘3 , then
𝑇1 ⊗ (𝑇2 + 𝑇3 ) = 𝑇1 ⊗ 𝑇2 + 𝑇1 ⊗ 𝑇3 .
A particularly interesting tensor product is the following. For 𝑖 = 1, …, 𝑘 let ℓ𝑖 ∈ 𝑉⋆
and let
(1.3.10) 𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 .
Thus, by definition,
(1.3.11) 𝑇(𝑣1 , …, 𝑣𝑘 ) = ℓ1 (𝑣1 ) ⋯ ℓ𝑘 (𝑣𝑘 ) .
 Draft: March 28, 2018

10 Chapter 1: Multilinear Algebra

A tensor of the form (1.3.11) is called a decomposable 𝑘-tensor. These tensors, as we will
see, play an important role in what follows. In particular, let 𝑒1 , …, 𝑒𝑛 be a basis of 𝑉 and
𝑒1⋆ , …, 𝑒𝑛⋆ the dual basis of 𝑉⋆ . For every multi-index 𝐼 of length 𝑘 let
𝑒𝐼⋆ = 𝑒𝑖⋆1 ⊗ ⋯ ⊗ 𝑒𝑖⋆𝑘 .
Then if 𝐽 is another multi-index of length 𝑘,

1, 𝐼 = 𝐽
(1.3.12) 𝑒𝐼⋆ (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = {
0, 𝐼 ≠ 𝐽
by (1.2.13), (1.3.10) and (1.3.11). From (1.3.12) it is easy to conclude:
Theorem 1.3.13. Let 𝑉 be a vector space with basis 𝑒1 , …, 𝑒𝑛 and 0 ≤ 𝑘 ≤ 𝑛 be an integer.
The 𝑘-tensors 𝑒𝐼⋆ of (1.3.12) are a basis of L𝑘 (𝑉).

Proof. Given 𝑇 ∈ L𝑘 (𝑉), let

𝑇′ = ∑ 𝑇𝐼 𝑒𝐼⋆
𝐼
where the 𝑇𝐼 ’s are defined by (1.3.6). Then
(1.3.14) 𝑇′ (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = ∑ 𝑇𝐼 𝑒𝐼⋆ (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = 𝑇𝐽
𝐼

by (1.3.12); however, by Proposition 1.3.7 the 𝑇𝐽 ’s determine 𝑇, so 𝑇′ = 𝑇. This proves that

the 𝑒𝐼⋆ ’s are a spanning set of vectors for L𝑘 (𝑉). To prove they’re a basis, suppose
∑ 𝐶𝐼 𝑒𝐼⋆ = 0
𝐼

for constants, 𝐶𝐼 ∈ 𝐑. Then by (1.3.14) with 𝑇′ = 0, 𝐶𝐽 = 0, so the 𝑒𝐼⋆ ’s are linearly indepen-
dent. □

As we noted in Lemma 1.3.5, there are exactly 𝑛𝑘 multi-indices of length 𝑘 and hence
𝑛 basis vectors in the set {𝑒𝐼⋆ }𝐼 , so we have proved
𝑘

Corollary 1.3.15. Let 𝑉 be an 𝑛-dimensional vector space. Then dim(L𝑘 (𝑉)) = 𝑛𝑘 .

The pullback operation
Definition 1.3.16. Let 𝑉 and 𝑊 be finite dimensional vector spaces and let 𝐴 ∶ 𝑉 → 𝑊 be
a linear mapping. If 𝑇 ∈ L𝑘 (𝑊), we define
𝐴⋆ 𝑇 ∶ 𝑉𝑘 → 𝐑
to be the function
(1.3.17) (𝐴⋆ 𝑇)(𝑣1 , …, 𝑣𝑘 ) ≔ 𝑇(𝐴𝑣1 , …, 𝐴𝑣𝑘 ) .
It is clear from the linearity of 𝐴 that this function is linear in its 𝑖th variable for all 𝑖, and
hence is a 𝑘-tensor. We call 𝐴⋆ 𝑇 the pullback of 𝑇 by the map 𝐴.
Proposition 1.3.18. The map
𝐴⋆ ∶ L𝑘 (𝑊) → L𝑘 (𝑉) , 𝑇 ↦ 𝐴⋆ 𝑇 ,
is a linear mapping.
 Draft: March 28, 2018

§1.4 Alternating 𝑘-tensors 11

We leave this as an exercise. We also leave as an exercise the identity

(1.3.19) 𝐴⋆ (𝑇1 ⊗ 𝑇2 ) = 𝐴⋆ (𝑇1 ) ⊗ 𝐴⋆ (𝑇2 )
for 𝑇1 ∈ L𝑘 (𝑊) and 𝑇2 ∈ L𝑚 (𝑊). Also, if 𝑈 is a vector space and 𝐵 ∶ 𝑈 → 𝑉 a linear
mapping, we leave for you to check that
(1.3.20) (𝐴𝐵)⋆ 𝑇 = 𝐵⋆ (𝐴⋆ 𝑇)
for all 𝑇 ∈ L𝑘 (𝑊).

Exercises for §1.3

Exercise 1.3.i. Verify that there are exactly 𝑛𝑘 multi-indices of length 𝑘.
Exercise 1.3.ii. Prove Proposition 1.3.18.
Exercise 1.3.iii. Verify equation (1.3.19).
Exercise 1.3.iv. Verify equation (1.3.20).
Exercise 1.3.v. Let 𝐴 ∶ 𝑉 → 𝑊 be a linear map. Show that if ℓ𝑖 , 𝑖 = 1, …, 𝑘 are elements of
𝑊⋆
𝐴⋆ (ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 ) = 𝐴⋆ (ℓ1 ) ⊗ ⋯ ⊗ 𝐴⋆ (ℓ𝑘 ) .
Conclude that 𝐴⋆ maps decomposable 𝑘-tensors to decomposable 𝑘-tensors.
Exercise 1.3.vi. Let 𝑉 be an 𝑛-dimensional vector space and ℓ𝑖 , 𝑖 = 1, 2, elements of 𝑉⋆ .
Show that ℓ1 ⊗ ℓ2 = ℓ2 ⊗ ℓ1 if and only if ℓ1 and ℓ2 are linearly dependent.
Hint: Show that if ℓ1 and ℓ2 are linearly independent there exist vectors, 𝑣𝑖 , 𝑖 =, 1, 2 in
𝑉 with property
1, 𝑖 = 𝑗
ℓ𝑖 (𝑣𝑗 ) = {
0, 𝑖 ≠ 𝑗 .
Now compare (ℓ1 ⊗ ℓ2 )(𝑣1 , 𝑣2 ) and (ℓ2 ⊗ ℓ1 )(𝑣1 , 𝑣2 ). Conclude that if dim 𝑉 ≥ 2 the tensor
product operation is not commutative, i.e., it is usually not true that ℓ1 ⊗ ℓ2 = ℓ2 ⊗ ℓ1 .
Exercise 1.3.vii. Let 𝑇 be a 𝑘-tensor and 𝑣 a vector. Define 𝑇𝑣 ∶ 𝑉𝑘−1 → 𝐑 to be the map
(1.3.21) 𝑇𝑣 (𝑣1 , …, 𝑣𝑘−1 ) ≔ 𝑇(𝑣, 𝑣1 , …, 𝑣𝑘−1 ) .
Show that 𝑇𝑣 is a (𝑘 − 1)-tensor.
Exercise 1.3.viii. Show that if 𝑇1 is an 𝑟-tensor and 𝑇2 is an 𝑠-tensor, then if 𝑟 > 0,
(𝑇1 ⊗ 𝑇2 )𝑣 = (𝑇1 )𝑣 ⊗ 𝑇2 .
Exercise 1.3.ix. Let 𝐴 ∶ 𝑉 → 𝑊 be a linear map mapping, and 𝑣 ∈ 𝑉. Write 𝑤 ≔ 𝐴𝑣. Show
that for 𝑇 ∈ L𝑘 (𝑊), 𝐴⋆ (𝑇𝑤 ) = (𝐴⋆ 𝑇)𝑣 .

1.4. Alternating 𝑘-tensors

We will discuss in this section a class of 𝑘-tensors which play an important role in multi-
variable calculus. In this discussion we will need some standard facts about the “permutation
group”. For those of you who are already familiar with this object (and I suspect most of you
are) you can regard the paragraph below as a chance to re-familiarize yourselves with these
facts.
 Draft: March 28, 2018

12 Chapter 1: Multilinear Algebra

Permutations
Definition 1.4.1. Let 𝛴𝑘 be the 𝑘-element set 𝛴𝑘 ≔ {1, 2, …, 𝑘}. A permutation of order
𝑘 is a bijecton 𝜎 ∶ 𝛴𝑘 ⥲ 𝛴𝑘 . Given two permutations 𝜎1 and 𝜎2 , their product 𝜎1 𝜎2 is the
composition of 𝜎1 ∘ 𝜎2 , i.e., the map,
𝑖 ↦ 𝜎1 (𝜎2 (𝑖)) .
For every permutation 𝜎, one denotes by 𝜎−1 the inverse permutation given by the inverse
bijection of 𝜎, i.e., defined by
𝜎(𝑖) ≔ 𝑗 ⟺ 𝜎−1 (𝑗) = 𝑖 .
Let 𝑆𝑘 be the set of all permutations of order 𝑘. One calls 𝑆𝑘 the permutation group of 𝛴𝑘 or,
alternatively, the symmetric group on 𝑘 letters.
It is easy to check the following.
Lemma 1.4.2. The group 𝑆𝑘 has 𝑘! elements.
Definition 1.4.3. Let 𝑘 be a positive integer. For every 1 ≤ 𝑖 < 𝑗 ≤ 𝑘, let 𝜏𝑖,𝑗 be the permu-
tation
{𝑗, ℓ = 𝑖
{
𝜏𝑖,𝑗 (ℓ) ≔ {𝑖, ℓ = 𝑗
{
{ℓ, ℓ ≠ 𝑖, 𝑗 .
The permuation 𝜏𝑖,𝑗 is called a transposition, and if 𝑗 = 𝑖 + 1, then 𝜏𝑖,𝑗 is called an elementary
transposition.
Theorem 1.4.4. Every permutation in 𝑆𝑘 can be written as a product of (a finite number of)
transpositions.
Proof. We prove this by induction on 𝑘. The base case when 𝑘 = 2 is obvious.
For the induction step, suppose that we know the claim for 𝑆𝑘−1 . Given 𝜎 ∈ 𝑆𝑘 , we
have 𝜎(𝑘) = 𝑖 if and only if 𝜏𝑖,𝑘 𝜎(𝑘) = 𝑘. Thus 𝜏𝑖,𝑘 𝜎 is, in effect, a permutation of 𝛴𝑘−1 . By
induction, 𝜏𝑖,𝑘 𝜎 can be written as a product of transpositions, so
𝜎 = 𝜏𝑖,𝑘 (𝜏𝑖,𝑘 𝜎)
can be written as a product of transpositions. □
Theorem 1.4.5. Every transposition can be written as a product of elementary transpositions.
Proof. Let 𝜏 = 𝜏𝑖𝑗 , 𝑖 < 𝑗. With 𝑖 fixed, argue by induction on 𝑗. Note that for 𝑗 > 𝑖 + 1
𝜏𝑖𝑗 = 𝜏𝑗−1,𝑗 𝜏𝑖,𝑗−1 𝜏𝑗−1,𝑗 .
Now apply the inductive hypothesis to 𝜏𝑖,𝑗−1 . □
Corollary 1.4.6. Every permutation can be written as a product of elementary transpositions.
The sign of a permutation
Definition 1.4.7. Let 𝑥1 , …, 𝑥𝑘 be the coordinate functions on 𝐑𝑘 . For 𝜎 ∈ 𝑆𝑘 we define
𝑥𝜎(𝑖) − 𝑥𝜎(𝑗)
(1.4.8) (−1)𝜎 ≔ ∏ .
𝑖<𝑗 𝑥𝑖 − 𝑥𝑗
Notice that the numerator and denominator in (1.4.8) are identical up to sign. Indeed,
if 𝑝 = 𝜎(𝑖) < 𝜎(𝑗) = 𝑞, the term, 𝑥𝑝 − 𝑥𝑞 occurs once and just once in the numerator and
once and just once in the denominator; and if 𝑞 = 𝜎(𝑖) > 𝜎(𝑗) = 𝑝, the term, 𝑥𝑝 − 𝑥𝑞 , occurs
 Draft: March 28, 2018

§1.4 Alternating 𝑘-tensors 13

once and just once in the numerator and its negative, 𝑥𝑞 − 𝑥𝑝 , once and just once in the
numerator. Thus
(−1)𝜎 = ±1 .
In light of this, we call (−1)𝜎 the sign of 𝜎.
Claim 1.4.9. For 𝜎, 𝜏 ∈ 𝑆𝑘 we have
(−1)𝜎𝜏 = (−1)𝜎 (−1)𝜏 .
That is, the sign defines a group homomorphism 𝑆𝑘 → {±1}.
Proof. By definition,
𝑥𝜎𝜏(𝑖) − 𝑥𝜎𝜏(𝑗)
(−1)𝜎𝜏 = ∏ .
𝑖<𝑗 𝑥𝑖 − 𝑥𝑗
We write the right hand side as a product of
𝑥𝜏(𝑖) − 𝑥𝜏(𝑗)
∏ = (−1)𝜏
𝑖<𝑗 𝑥𝑖 − 𝑥𝑗

and
𝑥𝜎𝜏(𝑖) − 𝑥𝜎𝜏(𝑗)
(1.4.10) ∏
𝑖<𝑗 𝑥𝜏(𝑖) − 𝑥𝜏(𝑗)

For 𝑖 < 𝑗, let 𝑝 = 𝜏(𝑖) and 𝑞 = 𝜏(𝑗) when 𝜏(𝑖) < 𝜏(𝑗) and let 𝑝 = 𝜏(𝑗) and 𝑞 = 𝜏(𝑖) when
𝜏(𝑗) < 𝜏(𝑖). Then
𝑥𝜎𝜏(𝑖) − 𝑥𝜎𝜏(𝑗) 𝑥𝜎(𝑝) − 𝑥𝜎(𝑞)
=
𝑥𝜏(𝑖) − 𝑥𝜏(𝑗) 𝑥𝑝 − 𝑥𝑞
(i.e., if 𝜏(𝑖) < 𝜏(𝑗), the numerator and denominator on the right equal the numerator and
denominator on the left and, if 𝜏(𝑗) < 𝜏(𝑖) are negatives of the numerator and denominator
on the left). Thus equation (1.4.10) becomes
𝑥𝜎(𝑝) − 𝑥𝜎(𝑞)
∏ = (−1)𝜎 . □
𝑝<𝑞 𝑥 𝑝 − 𝑥𝑞

We’ll leave for you to check that if 𝜏 is a transposition, (−1)𝜏 = −1 and to conclude from
this:
Proposition 1.4.11. If 𝜎 is the product of an odd number of transpositions, (−1)𝜎 = −1 and
if 𝜎 is the product of an even number of transpositions (−1)𝜎 = +1.
Alternation
Definition 1.4.12. Let 𝑉 be an 𝑛-dimensional vector space and 𝑇 ∈ L𝑘 (𝑣) a 𝑘-tensor. For
𝜎 ∈ 𝑆𝑘 , define 𝑇𝜎 ∈ L𝑘 (𝑉) to be the 𝑘-tensor
(1.4.13) 𝑇𝜎 (𝑣1 , …, 𝑣𝑘 ) ≔ 𝑇(𝑣𝜎−1 (1) , …, 𝑣𝜎−1 (𝑘) ) .
Proposition 1.4.14.
(1) If 𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 , ℓ𝑖 ∈ 𝑉⋆ , then 𝑇𝜎 = ℓ𝜎(1) ⊗ ⋯ ⊗ ℓ𝜎(𝑘) .
(2) The assignment 𝑇 ↦ 𝑇𝜎 is a linear map L𝑘 (𝑉) → L𝑘 (𝑉).
(3) If 𝜎, 𝜏 ∈ 𝑆𝑘 , we have 𝑇𝜎𝜏 = (𝑇𝜎 )𝜏 .
 Draft: March 28, 2018

14 Chapter 1: Multilinear Algebra

Proof. To (1), we note that by equation (1.4.13)

(ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 )𝜎 (𝑣1 , …, 𝑣𝑘 ) = ℓ1 (𝑣𝜎−1 (1) ) ⋯ ℓ𝑘 (𝑣𝜎−1 (𝑘) ) .
Setting 𝜎−1 (𝑖) = 𝑞, the 𝑖th term in this product is ℓ𝜎(𝑞) (𝑣𝑞 ); so the product can be rewritten
as
ℓ𝜎(1) (𝑣1 )…ℓ𝜎(𝑘) (𝑣𝑘 )
or
(ℓ𝜎(1) ⊗ ⋯ ⊗ ℓ𝜎(𝑘) )(𝑣1 , …, 𝑣𝑘 ) .
We leave the proof of (2) as an exercise.
Now we prove (3). Let 𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 . Then
𝑇𝜎 = ℓ𝜎(1) ⊗ ⋯ ⊗ ℓ𝜎(𝑘) = ℓ1′ ⊗ ⋯ ⊗ ℓ𝑘′
where ℓ𝑗′ = ℓ𝜎(𝑗) . Thus
(𝑇𝜎 )𝜏 = ℓ𝜏(1)
′ ′
⊗ ⋯ ⊗ ℓ𝜏(𝑘) .
′
But if 𝜏(𝑖) = 𝑗, ℓ𝜏(𝑗) = ℓ𝜎(𝜏(𝑗)) . Hence
(𝑇𝜎 )𝜏 ℓ𝜎𝜏(1) ⊗ ⋯ ⊗ ℓ𝜎𝜏(𝑘) = 𝑇𝜎𝜏 . □

Definition 1.4.15. Let 𝑉 be a vector space and 𝑘 ≥ 0 an integer. A 𝑘-tensor 𝑇 ∈ L𝑘 (𝑉) is

alternating if 𝑇𝜎 = (−1)𝜎 𝑇 for all 𝜎 ∈ 𝑆𝑘 .
We denote by A 𝑘 (𝑉) the set of all alternating 𝑘-tensors in L𝑘 (𝑉). By (2) this set is a
vector subspace of L𝑘 (𝑉).
It is not easy to write down simple examples of alternating 𝑘-tensors. However, there is
a method, called the alternation operation Alt for constructing such tensors.
Definition 1.4.16. Let 𝑉 be a vector space and 𝑘 a nonnegative integer. The alternation
operation on L𝑘 (𝑉) is defined as follows: given 𝑇 ∈ L𝑘 (𝑉) let
Alt(𝑇) ≔ ∑ (−1)𝜏 𝑇𝜏 .
𝜏∈𝑆𝑘

The alternation operation enjoys the following properties.

Proposition 1.4.17. For 𝑇 ∈ L𝑘 (𝑉) and 𝜎 ∈ 𝑆𝑘 ,
(1) Alt(𝑇)𝜎 = (−1)𝜎 Alt 𝑇
(2) If 𝑇 ∈ A 𝑘 (𝑉), then Alt 𝑇 = 𝑘!𝑇.
(3) Alt(𝑇𝜎 ) = Alt(𝑇)𝜎
(4) The map
Alt ∶ L𝑘 (𝑉) → L𝑘 (𝑉) , 𝑇 ↦ Alt(𝑇)
is linear.
Proof. To prove (1) we note that by Proposition 1.4.14:
Alt(𝑇)𝜎 = ∑ (−1)𝜏 𝑇𝜏𝜎 = (−1)𝜎 ∑ (−1)𝜏𝜎 𝑇𝜏𝜎 .
𝜏∈𝑆𝑘 𝜏∈𝑆𝑘

But as 𝜏 runs over 𝑆𝑘 , 𝜏𝜎 runs over 𝑆𝑘 , and hence the right hand side is (−1)𝜎 Alt(𝑇).
To prove (2), note that if 𝑇 ∈ A 𝑘 (𝑉)
Alt 𝑇 = ∑ (−1)𝜏 𝑇𝜏 = ∑ (−1)𝜏 (−1)𝜏 𝑇 = 𝑘!𝑇 .
𝜏∈𝑆𝑘 𝜏∈𝑆𝑘
 Draft: March 28, 2018

§1.4 Alternating 𝑘-tensors 15

To prove (3), we compute:

Alt(𝑇𝜎 ) = ∑ (−1)𝜏 𝑇𝜏𝜎 = (−1)𝜎 ∑ (−1)𝜏𝜎 𝑇𝜏𝜎
𝜏∈𝑆𝑘 𝜏∈𝑆𝑘
𝜎 𝜎
= (−1) Alt(𝑇) = Alt(𝑇) .
Finally, (4) is an easy corollary of (2) of Proposition 1.4.14. □

We will use this alternation operation to construct a basis for A 𝑘 (𝑉). First, however, we
require some notation.
Definition 1.4.18. Let 𝐼 = (𝑖1 , …, 𝑖𝑘 ) be a multi-index of length 𝑘.
(1) 𝐼 is repeating if 𝑖𝑟 = 𝑖𝑠 for some 𝑟 ≠ 𝑠.
(2) 𝐼 is strictly increasing if 𝑖1 < 𝑖2 < ⋯ < 𝑖𝑟 .
(3) For 𝜎 ∈ 𝑆𝑘 , write 𝐼𝜎 ≔ (𝑖𝜎(1) , …, 𝑖𝜎(𝑘) ).
Remark 1.4.19. If 𝐼 is non-repeating there is a unique 𝜎 ∈ 𝑆𝑘 so that 𝐼𝜎 is strictly increasing.
Let 𝑒1 , …, 𝑒𝑛 be a basis of 𝑉 and let
𝑒𝐼⋆ = 𝑒𝑖⋆1 ⊗ ⋯ ⊗ 𝑒𝑖⋆𝑘
and
𝜓𝐼 = Alt(𝑒𝐼⋆ ) .
Proposition 1.4.20.
(1) 𝜓𝐼𝜎 = (−1)𝜎 𝜓𝐼 .
(2) If 𝐼 is repeating, 𝜓𝐼 = 0.
(3) If 𝐼 and 𝐽 are strictly increasing,
1, 𝐼 = 𝐽
𝜓𝐼 (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = {
0, 𝐼 ≠ 𝐽 .
Proof. To prove (1) we note that (𝑒𝐼⋆ )𝜎 = 𝑒𝐼⋆𝜎 ; so
Alt(𝑒𝐼⋆𝜎 ) = Alt(𝑒𝐼⋆ )𝜎 = (−1)𝜎 Alt(𝑒𝐼⋆ ) .
To prove (2), suppose 𝐼 = (𝑖1 , …, 𝑖𝑘 ) with 𝑖𝑟 = 𝑖𝑠 for 𝑟 ≠ 𝑠. Then if 𝜏 = 𝜏𝑖𝑟 ,𝑖𝑠 , 𝑒𝐼⋆ = 𝑒𝐼⋆𝜏 so
𝜓𝐼 = 𝜓𝐼𝜏 = (−1)𝜏 𝜓𝐼 = −𝜓𝐼 .
To prove (3), note that by definition
𝜓𝐼 (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = ∑(−1)𝜏 𝑒𝐼⋆𝜏 (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) .
𝜏

But by (1.3.12)
1, 𝐼𝜏 = 𝐽
(1.4.21) 𝑒𝐼⋆𝜏 (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = {
0, 𝐼𝜏 ≠ 𝐽 .
Thus if 𝐼 and 𝐽 are strictly increasing, 𝐼𝜏 is strictly increasing if and only if 𝐼𝜏 = 𝐼, and
(1.4.21) is nonzero if and only if 𝐼 = 𝐽. □

Now let 𝑇 be in A 𝑘 (𝑉). By Theorem 1.3.13,

𝑇 = ∑ 𝑎𝐽 𝑒𝐽⋆ , 𝑎𝐽 ∈ 𝐑 .
𝐽
 Draft: March 28, 2018

16 Chapter 1: Multilinear Algebra

Since 𝑘!𝑇 = Alt(𝑇),

1
𝑇= ∑ 𝑎𝐽 Alt(𝑒𝐽⋆ ) = ∑ 𝑏𝐽 𝜓𝐽 .
𝑘!
We can discard all repeating terms in this sum since they are zero; and for every non-repeating
term, 𝐽, we can write 𝐽 = 𝐼𝜎 , where 𝐼 is strictly increasing, and hence 𝜓𝐽 = (−1)𝜎 𝜓𝐼 .
Conclusion 1.4.22. We can write 𝑇 as a sum
(1.4.23) 𝑇 = ∑ 𝑐𝐼 𝜓𝐼 .
𝐼
with 𝐼’s strictly increasing.
Claim 1.4.24. The 𝑐𝐼 ’s are unique.
Proof. For 𝐽 strictly increasing
(1.4.25) 𝑇(𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = ∑ 𝑐𝐼 𝜓𝐼 (𝑒𝑗1 , …, 𝑒𝑗𝑘 ) = 𝑐𝐽 .
𝐼

By (1.4.23) the 𝜓𝐼 ’s, 𝐼 strictly increasing, are a spanning set of vectors for A 𝑘 (𝑉), and by
(1.4.25) they are linearly independent, so we have proved:
Proposition 1.4.26. The alternating tensors, 𝜓𝐼 , 𝐼 strictly increasing, are a basis for A 𝑘 (𝑉).
Thus dim A 𝑘 (𝑉) is equal to the number of strictly increasing multi-indices 𝐼 of length 𝑘.
We leave for you as an exercise to show that this number is equal to the binomal coefficient
𝑛 𝑛!
(1.4.27) ( )≔
𝑘 (𝑛 − 𝑘)!𝑘!
if 1 ≤ 𝑘 ≤ 𝑛. □
Hint: Show that every strictly increasing multi-index of length 𝑘 determines a 𝑘 element
subset of {1, …, 𝑛} and vice-versa.
Note also that if 𝑘 > 𝑛 every multi-index
𝐼 = (𝑖1 , …, 𝑖𝑘 )
of length 𝑘 has to be repeating: 𝑖𝑟 = 𝑖𝑠 for some 𝑟 ≠ 𝑠 since the 𝑖𝑝 ’s lie on the interval 1 ≤ 𝑖 ≤ 𝑛.
Thus by Proposition 1.4.17
𝜓𝐼 = 0
for all multi-indices of length 𝑘 > 0 and
A 𝑘 (𝑉) = 0 .
Exercises for §1.4
Exercise 1.4.i. Show that there are exactly 𝑘! permutations of order 𝑘.
Hint: Induction on 𝑘: Let 𝜎 ∈ 𝑆𝑘 , and let 𝜎(𝑘) = 𝑖, 1 ≤ 𝑖 ≤ 𝑘. Show that 𝜏𝑖,𝑘 𝜎 leaves 𝑘
fixed and hence is, in effect, a permutation of 𝛴𝑘−1 .
Exercise 1.4.ii. Prove that if 𝜏 ∈ 𝑆𝑘 is a transposition, (−1)𝜏 = −1 and deduce from this
Proposition 1.4.11.
Exercise 1.4.iii. Prove assertion 2 in Proposition 1.4.14.
Exercise 1.4.iv. Prove that dim A 𝑘 (𝑉) is given by (1.4.27).
Exercise 1.4.v. Verify that for 𝑖 < 𝑗 − 1
𝜏𝑖,𝑗 = 𝜏𝑗−1,𝑗 𝜏𝑖,𝑗−1 , 𝜏𝑗−1,𝑗 .
 Draft: March 28, 2018

§1.5 The space 𝛬𝑘 (𝑉⋆ ) 17

Exercise 1.4.vi. For 𝑘 = 3 show that every one of the six elements of 𝑆3 is either a transpo-
sition or can be written as a product of two transpositions.
Exercise 1.4.vii. Let 𝜎 ∈ 𝑆𝑘 be the “cyclic” permutation
𝜎(𝑖) ≔ 𝑖 + 1 , 𝑖 = 1, …, 𝑘 − 1
and 𝜎(𝑘) ≔ 1. Show explicitly how to write 𝜎 as a product of transpositions and compute
(−1)𝜎 .
Hint: Same hint as in Exercise 1.4.i.
Exercise 1.4.viii. In Exercise 1.3.vii show that if 𝑇 is in A 𝑘 (𝑉), 𝑇𝑣 is in A 𝑘−1 (𝑉). Show in
addition that for 𝑣, 𝑤 ∈ 𝑉 and 𝑇 ∈ A 𝑘 (𝑉) we have (𝑇𝑣 )𝑤 = −(𝑇𝑤 )𝑣 .
Exercise 1.4.ix. Let 𝐴 ∶ 𝑉 → 𝑊 be a linear mapping. Show that if 𝑇 is in A 𝑘 (𝑊), 𝐴⋆ 𝑇 is in
A 𝑘 (𝑉).
Exercise 1.4.x. In Exercise 1.4.ix show that if 𝑇 is in L𝑘 (𝑊) then Alt(𝐴⋆ 𝑇) = 𝐴⋆ (Alt(𝑇)),
i.e., show that the Alt operation commutes with the pullback operation.

1.5. The space 𝛬𝑘 (𝑉⋆ )

In §1.4 we showed that the image of the alternation operation, Alt ∶ L𝑘 (𝑉) → L𝑘 (𝑉) is
A 𝑘 (𝑉). In this section we will compute the kernel of Alt.
Definition 1.5.1. A decomposable 𝑘-tensor ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 , with ℓ1 , …, ℓ𝑘 ∈ 𝑉⋆ , is redundant
if for some index 𝑖 we have ℓ𝑖 = ℓ𝑖+1 .
Let I𝑘 (𝑉) ⊂ L𝑘 (𝑉) be the linear span of the set of redundant 𝑘-tensors.
Note that for 𝑘 = 1 the notion of redundant does not really make sense; a single vector
ℓ ∈ L1 (𝑉⋆ ) cannot be “redundant” so we decree
I1 (𝑉) ≔ 0 .
Proposition 1.5.2. If 𝑇 ∈ I𝑘 (𝑉) then Alt(𝑇) = 0.
Proof. Let 𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 with ℓ𝑖 = ℓ𝑖+1 . Then if 𝜏 = 𝜏𝑖,𝑖+1 , 𝑇𝜏 = 𝑇 and (−1)𝜏 = −1. Hence
Alt(𝑇) = Alt(𝑇𝜏 ) = Alt(𝑇)𝜏 = − Alt(𝑇); so Alt(𝑇) = 0. □
Proposition 1.5.3. If 𝑇 ∈ I𝑟 (𝑉) and 𝑇′ ∈ L𝑠 (𝑉) then 𝑇 ⊗ 𝑇′ and 𝑇′ ⊗ 𝑇 are in I𝑟+𝑠 (𝑉).
Proof. We can assume that 𝑇 and 𝑇′ are decomposable, i.e., 𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑟 and 𝑇′ =
ℓ1′ ⊗ ⋯ ⊗ ℓ𝑠′ and that 𝑇 is redundant: ℓ𝑖 = ℓ𝑖+1 . Then
𝑇 ⊗ 𝑇′ = ℓ1 ⊗ ⋯ ℓ𝑖−1 ⊗ ℓ𝑖 ⊗ ℓ𝑖 ⊗ ⋯ ℓ𝑟 ⊗ ℓ1′ ⊗ ⋯ ⊗ ℓ𝑠′
is redundant and hence in I𝑟+𝑠 . The argument for 𝑇′ ⊗ 𝑇 is similar. □
𝑘
Proposition 1.5.4. If 𝑇 ∈ L (𝑉) and 𝜎 ∈ 𝑆𝑘 , then
(1.5.5) 𝑇𝜎 = (−1)𝜎 𝑇 + 𝑆
where 𝑆 is in I𝑘 (𝑉).
Proof. We can assume 𝑇 is decomposable, i.e., 𝑇 = ℓ1 ⊗⋯⊗ℓ𝑘 . Let’s first look at the simplest
possible case: 𝑘 = 2 and 𝜎 = 𝜏1,2 . Then
𝑇𝜎 − (−1)𝜎 𝑇 = ℓ1 ⊗ ℓ2 + ℓ2 ⊗ ℓ1
1
= ((ℓ1 + ℓ2 ) ⊗ (ℓ1 + ℓ2 ) − ℓ1 ⊗ ℓ1 − ℓ2 ⊗ ℓ2 ) ,
2
 Draft: March 28, 2018

18 Chapter 1: Multilinear Algebra

and the terms on the right are redundant, and hence in I2 (𝑉). Next let 𝑘 be arbitrary and
𝜎 = 𝜏𝑖,𝑖+1 . If 𝑇1 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑖−2 and 𝑇2 = ℓ𝑖+2 ⊗ ⋯ ⊗ ℓ𝑘 . Then
𝑇 − (−1)𝜎 𝑇 = 𝑇1 ⊗ (ℓ𝑖 ⊗ ℓ𝑖+1 + ℓ𝑖+1 ⊗ ℓ𝑖 ) ⊗ 𝑇2
is in I𝑘 (𝑉) by Proposition 1.5.3 and the computation above.
The general case: By Theorem 1.4.5, 𝜎 can be written as a product of 𝑚 elementary transpo-
sitions, and we’ll prove (1.5.5) by induction on 𝑚.
We have just dealt with the case 𝑚 = 1.
The induction step: the“𝑚 − 1” case implies the “𝑚” case. Let 𝜎 = 𝜏𝛽 where 𝛽 is a product of
𝑚 − 1 elementary transpositions and 𝜏 is an elementary transposition. Then
𝑇𝜎 = (𝑇 𝛽 )𝜏 = (−1)𝜏 𝑇 𝛽 + ⋯
= (−1)𝜏 (−1)𝛽 𝑇 + ⋯
= (−1)𝜎 𝑇 + ⋯
where the “dots” are elements of I𝑘 (𝑉), and the induction hypothesis was used in the second
line. □
Corollary 1.5.6. If 𝑇 ∈ L𝑘 (𝑉), then
(1.5.7) Alt(𝑇) = 𝑘!𝑇 + 𝑊 ,
𝑘
where 𝑊 is in I (𝑉).
Proof. By definition Alt(𝑇) = ∑𝜎∈𝑆𝑘 (−1)𝜎 𝑇𝜎 , and by Proposition 1.5.4, 𝑇𝜎 = (−1)𝜎 𝑇 + 𝑊𝜎 ,
with 𝑊𝜎 ∈ I𝑘 (𝑉). Thus
Alt(𝑇) = ∑ (−1)𝜎 (−1)𝜎 𝑇 + ∑ (−1)𝜎 𝑊𝜎 = 𝑘!𝑇 + 𝑊
𝜎∈𝑆𝑘 𝜎∈𝑆𝑘
𝜎
where 𝑊 = ∑𝜎∈𝑆𝑘 (−1) 𝑊𝜎 . □

Corollary 1.5.8. Let 𝑉 be a vector space and 𝑘 ≥ 1. Then

I𝑘 (𝑉) = ker(Alt ∶ L𝑘 (𝑉) → A 𝑘 (𝑉)) .
Proof. We have already proved that if 𝑇 ∈ I𝑘 (𝑉), then Alt(𝑇) = 0. To prove the converse
assertion we note that if Alt(𝑇) = 0, then by (1.5.7)
1
𝑇=− 𝑊.
𝑘!
with 𝑊 ∈ I𝑘 (𝑉) . □
Putting these results together we conclude:
Theorem 1.5.9. Every element 𝑇 ∈ L𝑘 (𝑉) can be written uniquely as a sum 𝑇 = 𝑇1 + 𝑇2 ,
where 𝑇1 ∈ A 𝑘 (𝑉) and 𝑇2 ∈ I𝑘 (𝑉).
Proof. By (1.5.7), 𝑇 = 𝑇1 + 𝑇2 with
1
𝑇1 = Alt(𝑇)
𝑘!
and
1
𝑇2 = − 𝑊.
𝑘!
 Draft: March 28, 2018

§1.5 The space 𝛬𝑘 (𝑉⋆ ) 19

To prove that this decomposition is unique, suppose 𝑇1 + 𝑇2 = 0, with 𝑇1 ∈ A 𝑘 (𝑉) and

𝑇2 ∈ I𝑘 (𝑉). Then
0 = Alt(𝑇1 + 𝑇2 ) = 𝑘!𝑇1
so 𝑇1 = 0, and hence 𝑇2 = 0. □
Definition 1.5.10. Let 𝑉 be a finite dimensional vector space and 𝑘 ≥ 0. Define
(1.5.11) 𝛬𝑘 (𝑉⋆ ) ≔ L𝑘 (𝑉)/I𝑘 (𝑉) ,
i.e., let 𝛬𝑘 (𝑉⋆ ) be the quotient of the vector space L𝑘 (𝑉) by the subspace I𝑘 (𝑉). By (1.2.7)
one has a linear map:
(1.5.12) 𝜋 ∶ L𝑘 (𝑉) → 𝛬𝑘 (𝑉⋆ ) , 𝑇 ↦ 𝑇 + I𝑘 (𝑉)
which is onto and has I𝑘 (𝑉) as kernel.
Theorem 1.5.13. The map 𝜋 ∶ L𝑘 (𝑉) → 𝛬𝑘 (𝑉⋆ ) maps A 𝑘 (𝑉) bijectively onto 𝛬𝑘 (𝑉⋆ ).
Proof. By Theorem 1.5.9 every I𝑘 (𝑉) coset 𝑇+I𝑘 (𝑉) contains a unique element 𝑇1 of A 𝑘 (𝑉).
Hence for every element of 𝛬𝑘 (𝑉⋆ ) there is a unique element of A 𝑘 (𝑉) which gets mapped
onto it by 𝜋. □
Remark 1.5.14. Since 𝛬𝑘 (𝑉⋆ ) and A 𝑘 (𝑉) are isomorphic as vector spaces many treatments
of multilinear algebra avoid mentioning 𝛬𝑘 (𝑉⋆ ), reasoning that A 𝑘 (𝑉) is a perfectly good
substitute for it and that one should, if possible, not make two different definitions for what
is essentially the same object. This is a justifiable point of view (and is the point of view taken
by in [4,10,12]). There are, however, some advantages to distinguishing between A 𝑘 (𝑉) and
𝛬𝑘 (𝑉⋆ ), as we shall see in §1.6.
Exercises for §1.5
𝑘
Exercise 1.5.i. A 𝑘-tensor 𝑇 ∈ L (𝑉) is symmetric if 𝑇𝜎 = 𝑇 for all 𝜎 ∈ 𝑆𝑘 . Show that the
set 𝒮 𝑘 (𝑉) of symmetric 𝑘 tensors is a vector subspace of L𝑘 (𝑉).
Exercise 1.5.ii. Let 𝑒1 , …, 𝑒𝑛 be a basis of 𝑉. Show that every symmetric 2-tensor is of the
form
∑ 𝑎𝑖,𝑗 𝑒𝑖⋆ ⊗ 𝑒𝑗⋆
1≤𝑖,𝑗≤𝑛
where 𝑎𝑖,𝑗 = 𝑎𝑗,𝑖 and 𝑒1⋆ , …, 𝑒𝑛⋆ are the dual basis vectors of 𝑉⋆ .

Exercise 1.5.iii. Show that if 𝑇 is a symmetric 𝑘-tensor, then for 𝑘 ≥ 2, then 𝑇 is in I𝑘 (𝑉).
Hint: Let 𝜎 be a transposition and deduce from the identity, 𝑇𝜎 = 𝑇, that 𝑇 has to be in
the kernel of Alt.
Exercise 1.5.iv (a warning). In general 𝒮 𝑘 (𝑉) ≠ I𝑘 (𝑉). Show, however, that if 𝑘 = 2 these
two spaces are equal.
Exercise 1.5.v. Show that if ℓ ∈ 𝑉⋆ and 𝑇 ∈ L𝑘−2 (𝑉), then ℓ ⊗ 𝑇 ⊗ ℓ is in I𝑘 (𝑉).
Exercise 1.5.vi. Show that if ℓ1 and ℓ2 are in 𝑉⋆ and 𝑇 ∈ L𝑘−2 (𝑉), then ℓ1 ⊗𝑇⊗ℓ2 +ℓ2 ⊗𝑇⊗ℓ1
is in I𝑘 (𝑉).
Exercise 1.5.vii. Given a permutation 𝜎 ∈ 𝑆𝑘 and 𝑇 ∈ I𝑘 (𝑉), show that 𝑇𝜎 ∈ I𝑘 (𝑉).
Exercise 1.5.viii. Let W(𝑉) be a subspace of L𝑘 (𝑉) having the following two properties.
(1) For 𝑆 ∈ 𝒮 2 (𝑉) and 𝑇 ∈ L𝑘−2 (𝑉), 𝑆 ⊗ 𝑇 is in W(𝑉).
 Draft: March 28, 2018

20 Chapter 1: Multilinear Algebra

(2) For 𝑇 in W(𝑉) and 𝜎 ∈ 𝑆𝑘 , 𝑇𝜎 is in W(𝑉).

Show that W(𝑉) has to contain I𝑘 (𝑉) and conclude that I𝑘 (𝑉) is the smallest subspace
𝑘
of L (𝑉) having properties (1) and (2).
Exercise 1.5.ix. Show that there is a bijective linear map
𝛼 ∶ 𝛬𝑘 (𝑉⋆ ) ⥲ A 𝑘 (𝑉)
with the property
1
(1.5.15) 𝛼𝜋(𝑇) = Alt(𝑇)
𝑘!
for all 𝑇 ∈ L𝑘 (𝑉), and show that 𝛼 is the inverse of the map of A 𝑘 (𝑉) onto 𝛬𝑘 (𝑉⋆ ) described
in Theorem 1.5.13.
Hint: Exercise 1.2.viii.
Exercise 1.5.x. Let 𝑉 be an 𝑛-dimensional vector space. Compute the dimension of 𝒮 𝑘 (𝑉).
Hints:
(1) Introduce the following symmetrization operation on tensors 𝑇 ∈ L𝑘 (𝑉):
Sym(𝑇) = ∑ 𝑇𝜏 .
𝜏∈𝑆𝑘

Prove that this operation has properties (2)–(4) of Proposition 1.4.17 and, as a substitute
for (1), has the property: Sym(𝑇)𝜎 = Sym(𝑇).
(2) Let 𝜙𝐼 = Sym(𝑒𝐼⋆ ), 𝑒𝐼⋆ = 𝑒𝑖⋆1 ⊗ ⋯ ⊗ 𝑒𝑖⋆𝑛 . Prove that { 𝜙𝐼 | 𝐼 is non-decreasing } form a basis
of 𝑆𝑘 (𝑉).
(3) Conclude that dim(𝒮 𝑘 (𝑉)) is equal to the number of non-decreasing multi-indices of
length 𝑘: 1 ≤ 𝑖1 ≤ 𝑖2 ≤ ⋯ ≤ ℓ𝑘 ≤ 𝑛.
(4) Compute this number by noticing that the assignment
(𝑖1 , …, 𝑖𝑘 ) ↦ (𝑖1 + 0, 𝑖2 + 1, …, 𝑖𝑘 + 𝑘 − 1)
is a bijection between the set of these non-decreasing multi-indicesand the set of in-
creasing multi-indices 1 ≤ 𝑗1 < ⋯ < 𝑗𝑘 ≤ 𝑛 + 𝑘 − 1.

1.6. The wedge product

The tensor algebra operations on the spaces L𝑘 (𝑉) which we discussed in §§1.2 and 1.3,
i.e., the “tensor product operation” and the “pullback” operation, give rise to similar oper-
ations on the spaces, 𝛬𝑘 (𝑉⋆ ). We will discuss in this section the analogue of the tensor
product operation.
Definition 1.6.1. Given 𝜔𝑖 ∈ 𝛬𝑘𝑖 (𝑉⋆ ), 𝑖 = 1, 2 we can, by equation (1.5.12), find a 𝑇𝑖 ∈
L𝑘𝑖 (𝑉) with 𝜔𝑖 = 𝜋(𝑇𝑖 ). Then 𝑇1 ⊗ 𝑇2 ∈ L𝑘1 +𝑘2 (𝑉). The wedge product 𝜔1 ∧ 𝜔2 is defined by
(1.6.2) 𝜔1 ∧ 𝜔2 ≔ 𝜋(𝑇1 ⊗ 𝑇2 ) ∈ 𝛬𝑘1 +𝑘2 (𝑉⋆ ) .
Claim 1.6.3. This wedge product is well defined, i.e., does not depend on our choices of 𝑇1 and
𝑇2 .
Proof. Let 𝜋(𝑇1 ) = 𝜋(𝑇1′ ) = 𝜔1 . Then 𝑇1′ = 𝑇1 + 𝑊1 for some 𝑊1 ∈ I𝑘1 (𝑉), so
𝑇1′ ⊗ 𝑇2 = 𝑇1 ⊗ 𝑇2 + 𝑊1 ⊗ 𝑇2 .
But 𝑊1 ∈ I𝑘1 (𝑉) implies 𝑊1 ⊗ 𝑇2 ∈ I𝑘1 +𝑘2 (𝑉) and this implies:
𝜋(𝑇1′ ⊗ 𝑇2 ) = 𝜋(𝑇1 ⊗ 𝑇2 ) .
 Draft: March 28, 2018

§1.6 The wedge product 21

A symmetric argument shows that 𝜔1 ∧ 𝜔2 is well-defined, independent of the choice of

𝑇2 . □
More generally let 𝜔𝑖 ∈ 𝛬𝑘𝑖 (𝑉⋆ ), for 𝑖 = 1, 2, 3, and let 𝜔𝑖 = 𝜋(𝑇𝑖 ), 𝑇𝑖 ∈ L𝑘𝑖 (𝑉). Define
𝜔1 ∧ 𝜔2 ∧ 𝜔3 ∈ 𝛬𝑘1 +𝑘2 +𝑘3 (𝑉⋆ )
by setting
𝜔1 ∧ 𝜔2 ∧ 𝜔3 = 𝜋(𝑇1 ⊗ 𝑇2 ⊗ 𝑇3 ) .
As above it is easy to see that this is well-defined independent of the choice of 𝑇1 , 𝑇2 and 𝑇3 .
It is also easy to see that this triple wedge product is just the wedge product of 𝜔1 ∧ 𝜔2 with
𝜔3 or, alternatively, the wedge product of 𝜔1 with 𝜔2 ∧ 𝜔3 , i.e.,
𝜔1 ∧ 𝜔2 ∧ 𝜔3 = (𝜔1 ∧ 𝜔2 ) ∧ 𝜔3 = 𝜔1 ∧ (𝜔2 ∧ 𝜔3 ).
We leave for you to check: for 𝜆 ∈ 𝐑
(1.6.4) 𝜆(𝜔1 ∧ 𝜔2 ) = (𝜆𝜔1 ) ∧ 𝜔2 = 𝜔1 ∧ (𝜆𝜔2 )
and verify the two distributive laws:
(1.6.5) (𝜔1 + 𝜔2 ) ∧ 𝜔3 = 𝜔1 ∧ 𝜔3 + 𝜔2 ∧ 𝜔3
and
(1.6.6) 𝜔1 ∧ (𝜔2 + 𝜔3 ) = 𝜔1 ∧ 𝜔2 + 𝜔1 ∧ 𝜔3 .
𝑘
As we noted in §1.4, I (𝑉) = 0 for 𝑘 = 1, i.e., there are no nonzero “redundant” 𝑘 tensors
in degree 𝑘 = 1. Thus
𝛬1 (𝑉⋆ ) = 𝑉⋆ = L1 (𝑉).
A particularly interesting example of a wedge product is the following. Let ℓ1 , …, ℓ𝑘 ∈
𝑉⋆ = 𝛬1 (𝑉⋆ ). Then if 𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑘
(1.6.7) ℓ1 ∧ ⋯ ∧ ℓ𝑘 = 𝜋(𝑇) ∈ 𝛬𝑘 (𝑉⋆ ) .
We will call (1.6.7) a decomposable element of 𝛬𝑘 (𝑉⋆ ).
We will prove that these elements satisfy the following wedge product identity. For 𝜎 ∈
𝑆𝑘 :
(1.6.8) ℓ𝜎(1) ∧ ⋯ ∧ ℓ𝜎(𝑘) = (−1)𝜎 ℓ1 ∧ ⋯ ∧ ℓ𝑘 .

Proof. For every 𝑇 ∈ L𝑘 (𝑉), 𝑇 = (−1)𝜎 𝑇 + 𝑊 for some 𝑊 ∈ I𝑘 (𝑉) by Proposition 1.5.4.
Therefore since 𝜋(𝑊) = 0
𝜋(𝑇𝜎 ) = (−1)𝜎 𝜋(𝑇) .
𝜎
In particular, if 𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 , 𝑇 = ℓ𝜎(1) ⊗ ⋯ ⊗ ℓ𝜎(𝑘) , so
𝜋(𝑇𝜎 ) = ℓ𝜎(1) ∧ ⋯ ∧ ℓ𝜎(𝑘) = (−1)𝜎 𝜋(𝑇)
= (−1)𝜎 ℓ1 ∧ ⋯ ∧ ℓ𝑘 . □
In particular, for ℓ1 and ℓ2 ∈ 𝑉⋆
(1.6.9) ℓ1 ∧ ℓ2 = −ℓ2 ∧ ℓ1
and for ℓ1 , ℓ2 and ℓ3 ∈ 𝑉⋆
ℓ1 ∧ ℓ2 ∧ ℓ3 = −ℓ2 ∧ ℓ1 ∧ ℓ3 = ℓ2 ∧ ℓ3 ∧ ℓ1 .
More generally, it is easy to deduce from equation (1.6.8) the following result (which we’ll
leave as an exercise).
 Draft: March 28, 2018

22 Chapter 1: Multilinear Algebra

Theorem 1.6.10. If 𝜔1 ∈ 𝛬𝑟 (𝑉⋆ ) and 𝜔2 ∈ 𝛬𝑠 (𝑉⋆ ) then

(1.6.11) 𝜔1 ∧ 𝜔2 = (−1)𝑟𝑠 𝜔2 ∧ 𝜔1 .

Hint: It suffices to prove this for decomposable elements i.e., for 𝜔1 = ℓ1 ∧ ⋯ ∧ ℓ𝑟 and
𝜔2 = ℓ1′ ∧ ⋯ ∧ ℓ𝑠′ . Now make 𝑟𝑠 applications of (1.6.9).
Let 𝑒1 , …, 𝑒𝑛 be a basis of 𝑉 and let 𝑒1⋆ , …, 𝑒𝑛⋆ be the dual basis of 𝑉⋆ . For every multi-
index 𝐼 of length 𝑘,

(1.6.12) 𝑒𝑖⋆1 ∧ ⋯ ∧ 𝑒𝑖⋆𝑘 = 𝜋(𝑒𝐼⋆ ) = 𝜋(𝑒𝑖⋆1 ⊗ ⋯ ⊗ 𝑒𝑖⋆𝑘 ) .

Theorem 1.6.13. The elements (1.6.12), with 𝐼 strictly increasing, are basis vectors of 𝛬𝑘 (𝑉⋆ ).

Proof. The elements 𝜓𝐼 = Alt(𝑒𝐼⋆ ), for 𝐼 strictly increasing, are basis vectors of A 𝑘 (𝑉) by
Proposition 1.4.26; so their images, 𝜋(𝜓𝐼 ), are a basis of 𝛬𝑘 (𝑉⋆ ). But

𝜋(𝜓𝐼 ) = 𝜋 (∑𝜎∈𝑆𝑘 (−1)𝜎 (𝑒𝐼⋆ )𝜎 ) = ∑ (−1)𝜎 𝜋(𝑒𝐼⋆ )𝜎

𝜎∈𝑆𝑘

= ∑ (−1) (−1) 𝜎 𝜎
𝜋(𝑒𝐼⋆ ) = 𝑘!𝜋(𝑒𝐼⋆ ) . □
𝜎∈𝑆𝑘

Exercises for §1.6

Exercise 1.6.i. Prove the assertions (1.6.4), (1.6.5), and (1.6.6).

Exercise 1.6.ii. Verify the multiplication law in equation (1.6.11) for the wedge product.

Exercise 1.6.iii. Given 𝜔 ∈ 𝛬𝑟 (𝑉⋆ ) let 𝜔𝑘 be the 𝑘-fold wedge product of 𝜔 with itself, i.e.,
let 𝜔2 = 𝜔 ∧ 𝜔, 𝜔3 = 𝜔 ∧ 𝜔 ∧ 𝜔, etc.
(1) Show that if 𝑟 is odd then for 𝑘 > 1, 𝜔𝑘 = 0.
(2) Show that if 𝜔 is decomposable, then for 𝑘 > 1, 𝜔𝑘 = 0.

Exercise 1.6.iv. If 𝜔 and 𝜇 are in 𝛬𝑟 (𝑉⋆ ) prove:

𝑘
𝑘
(𝜔 + 𝜇)𝑘 = ∑ ( ) 𝜔ℓ ∧ 𝜇𝑘−ℓ .
ℓ=0 ℓ

Hint: As in freshman calculus, prove this binomial theorem by induction using the iden-
tity: ( 𝑘ℓ ) = ( 𝑘−1
ℓ−1
) + ( 𝑘−1
ℓ
).

Exercise 1.6.v. Let 𝜔 be an element of 𝛬2 (𝑉⋆ ). By definition the rank of 𝜔 is 𝑘 if 𝜔𝑘 ≠ 0 and

𝜔𝑘+1 = 0. Show that if
𝜔 = 𝑒1 ∧ 𝑓1 + ⋯ + 𝑒𝑘 ∧ 𝑓𝑘
with 𝑒𝑖 , 𝑓𝑖 ∈ 𝑉⋆ , then 𝜔 is of rank ≤ 𝑘.
Hint: Show that
𝜔𝑘 = 𝑘!𝑒1 ∧ 𝑓1 ∧ ⋯ ∧ 𝑒𝑘 ∧ 𝑓𝑘 .

Exercise 1.6.vi. Given 𝑒𝑖 ∈ 𝑉⋆ , 𝑖 = 1, …, 𝑘 show that 𝑒1 ∧ ⋯ ∧ 𝑒𝑘 ≠ 0 if and only if the 𝑒𝑖 ’s

are linearly independent.
Hint: Induction on 𝑘.
 Draft: March 28, 2018

§1.7 The interior product 23

1.7. The interior product

We’ll describe in this section another basic product operation on the spaces 𝛬𝑘 (𝑉⋆ ). As
above we’ll begin by defining this operator on the L𝑘 (𝑉)’s.
Definition 1.7.1. Let 𝑉 be a vector space and 𝑘 a nonnegative integer. Given 𝑇 ∈ L𝑘 (𝑉) and
𝑣 ∈ 𝑉 let 𝜄𝑣 𝑇 be the be the (𝑘 − 1)-tensor which takes the value
𝑘
(𝜄𝑣 𝑇)(𝑣1 , …, 𝑣𝑘−1 ) ≔ ∑ (−1)𝑟−1 𝑇(𝑣1 , …, 𝑣𝑟−1 , 𝑣, 𝑣𝑟 , …, 𝑣𝑘−1 )
𝑟=1

on the 𝑘 − 1-tuple of vectors, 𝑣1 , …, 𝑣𝑘−1 , i.e., in the 𝑟th summand on the right, 𝑣 gets in-
serted between 𝑣𝑟−1 and 𝑣𝑟 . (In particular the first summand is 𝑇(𝑣, 𝑣1 , …, 𝑣𝑘−1 ) and the last
summand is (−1)𝑘−1 𝑇(𝑣1 , …, 𝑣𝑘−1 , 𝑣).)
It is clear from the definition that if 𝑣 = 𝑣1 + 𝑣2
(1.7.2) 𝜄𝑣 𝑇 = 𝜄 𝑣 1 𝑇 + 𝜄 𝑣 2 𝑇 ,
and if 𝑇 = 𝑇1 + 𝑇2
(1.7.3) 𝜄𝑣 𝑇 = 𝜄𝑣 𝑇1 + 𝜄𝑣 𝑇2 ,
and we will leave for you to verify by inspection the following two lemmas.
Lemma 1.7.4. If 𝑇 is the decomposable 𝑘-tensor ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 then
𝑘
(1.7.5) 𝜄𝑣 𝑇 = ∑ (−1)𝑟−1 ℓ𝑟 (𝑣)ℓ1 ⊗ ⋯ ⊗ ℓ𝑟̂ ⊗ ⋯ ⊗ ℓ𝑘 ,
𝑟=1

where the “hat” over ℓ𝑟 means that ℓ𝑟 is deleted from the tensor product.
Lemma 1.7.6. If 𝑇1 ∈ L𝑝 (𝑉) and 𝑇2 ∈ L𝑞 (𝑉)
(1.7.7) 𝜄𝑣 (𝑇1 ⊗ 𝑇2 ) = 𝜄𝑣 𝑇1 ⊗ 𝑇2 + (−1)𝑝 𝑇1 ⊗ 𝜄𝑣 𝑇2 .
We will next prove the important identity.
Lemma 1.7.8. Let 𝑉 be a vector space and 𝑇 ∈ L𝑘 (𝑉). Then for all 𝑣 ∈ 𝑉 we have
(1.7.9) 𝜄𝑣 (𝜄𝑣 𝑇) = 0 .
Proof. It suffices by linearity to prove this for decomposable tensors and since (1.7.9) is
trivially true for 𝑇 ∈ L1 (𝑉), we can by induction assume (1.7.9) is true for decomposable
tensors of degree 𝑘 − 1. Let ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 be a decomposable tensor of degree 𝑘. Setting
𝑇 ≔ ℓ1 ⊗ ⋯ ⊗ ℓ𝑘−1 and ℓ = ℓ𝑘 we have
𝜄𝑣 (ℓ1 ⊗ ⋯ ⊗ ℓ𝑘 ) = 𝜄𝑣 (𝑇 ⊗ ℓ) = 𝜄𝑣 𝑇 ⊗ ℓ + (−1)𝑘−1 ℓ(𝑣)𝑇
by (1.7.7). Hence
𝜄𝑣 (𝜄𝑣 (𝑇 ⊗ ℓ)) = 𝜄𝑣 (𝜄𝑣 𝑇) ⊗ ℓ + (−1)𝑘−2 ℓ(𝑣)𝜄𝑣 𝑇 + (−1)𝑘−1 ℓ(𝑣)𝜄𝑣 𝑇 .
But by induction the first summand on the right is zero and the two remaining summands
cancel each other out. □

From (1.7.9) we can deduce a slightly stronger result: For 𝑣1 , 𝑣2 ∈ 𝑉

(1.7.10) 𝜄𝑣1 𝜄𝑣2 = −𝜄𝑣2 𝜄𝑣1 .
 Draft: March 28, 2018

24 Chapter 1: Multilinear Algebra

Proof. Let 𝑣 = 𝑣1 + 𝑣2 . Then 𝜄𝑣 = 𝜄𝑣1 + 𝜄𝑣2 so

0 = 𝜄𝑣 𝜄𝑣 = (𝜄𝑣1 + 𝜄𝑣2 )(𝜄𝑣1 + 𝜄𝑣2 )
= 𝜄𝑣1 𝜄𝑣1 + 𝜄𝑣1 𝜄𝑣2 + 𝜄𝑣2 𝜄𝑣1 + 𝜄𝑣2 𝜄𝑣2
= 𝜄 𝑣 1 𝜄𝑣 2 + 𝜄 𝑣 2 𝜄𝑣 1
since the first and last summands are zero by (1.7.9). □

We’ll now show how to define the operation 𝜄𝑣 on 𝛬𝑘 (𝑉⋆ ). We’ll first prove:
Lemma 1.7.11. If 𝑇 ∈ L𝑘 (𝑉) is redundant then so is 𝜄𝑣 𝑇.
Proof. Let 𝑇 = 𝑇1 ⊗ ℓ ⊗ ℓ ⊗ 𝑇2 where ℓ is in 𝑉⋆ , 𝑇1 is in L𝑝 (𝑉) and 𝑇2 is in L𝑞 (𝑉). Then by
equation (1.7.7)
𝜄𝑣 𝑇 = 𝜄𝑣 𝑇1 ⊗ ℓ ⊗ ℓ ⊗ 𝑇2 + (−1)𝑝 𝑇1 ⊗ 𝜄𝑣 (ℓ ⊗ ℓ) ⊗ 𝑇2 + (−1)𝑝+2 𝑇1 ⊗ ℓ ⊗ ℓ ⊗ 𝜄𝑣 𝑇2 .
However, the first and the third terms on the right are redundant and
𝜄𝑣 (ℓ ⊗ ℓ) = ℓ(𝑣)ℓ − ℓ(𝑣)ℓ
by equation (1.7.5). □

Now let 𝜋 be the projection (1.5.12) of L𝑘 (𝑉) onto 𝛬𝑘 (𝑉⋆ ) and for 𝜔 = 𝜋(𝑇) ∈ 𝛬𝑘 (𝑉⋆ )
define
(1.7.12) 𝜄𝑣 𝜔 = 𝜋(𝜄𝑣 𝑇) .
To show that this definition is legitimate we note that if 𝜔 = 𝜋(𝑇1 ) = 𝜋(𝑇2 ), then 𝑇1 − 𝑇2 ∈
I𝑘 (𝑉), so by Lemma 1.7.11 𝜄𝑣 𝑇1 − 𝜄𝑣 𝑇2 ∈ I𝑘−1 and hence
𝜋(𝜄𝑣 𝑇1 ) = 𝜋(𝜄𝑣 𝑇2 ) .
Therefore, (1.7.12) does not depend on the choice of 𝑇.
By definition, 𝜄𝑣 is a linear map 𝛬𝑘 (𝑉⋆ ) → 𝛬𝑘−1 (𝑉⋆ ). We call this the interior product
operation. From the identities (1.7.2)–(1.7.12) one gets, for 𝑣, 𝑣1 , 𝑣2 ∈ 𝑉, 𝜔 ∈ 𝛬𝑘 (𝑉⋆ ), 𝜔1 ∈
𝛬𝑝 (𝑉⋆ ), and 𝜔2 ∈ 𝛬2 (𝑉⋆ )
(1.7.13) 𝜄(𝑣1 +𝑣2 ) 𝜔 = 𝜄𝑣1 𝜔 + 𝜄𝑣2 𝜔
(1.7.14) 𝜄𝑣 (𝜔1 ∧ 𝜔2 ) = 𝜄𝑣 𝜔1 ∧ 𝜔2 + (−1)𝑝 𝜔1 ∧ 𝜄𝑣 𝜔2
(1.7.15) 𝜄𝑣 (𝜄𝑣 𝜔) = 0
and
𝜄𝑣1 𝜄𝑣2 𝜔 = −𝜄𝑣2 𝜄𝑣1 𝜔 .
Moreover if 𝜔 = ℓ1 ∧ ⋯ ∧ ℓ𝑘 is a decomposable element of 𝛬𝑘 (𝑉⋆ ) one gets from (1.7.5)
𝑘
𝜄𝑣 𝜔 = ∑ (−1)𝑟−1 ℓ𝑟 (𝑣)ℓ1 ∧ ⋯ ∧ ℓ𝑟̂ ∧ ⋯ ∧ ℓ𝑘 .
𝑟=1

In particular if 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉, 𝑒1⋆ , …, 𝑒𝑛⋆ the dual basis of 𝑉⋆ and 𝜔𝐼 = 𝑒𝑖⋆1 ∧ ⋯ ∧ 𝑒𝑖⋆𝑘 ,
1 ≤ 𝑖1 < ⋯ < 𝑖𝑘 ≤ 𝑛, then 𝜄𝑒𝑗 𝜔𝐼 = 0 if 𝑗 ∉ 𝐼 and if 𝑗 = 𝑖𝑟

(1.7.16) 𝜄𝑒𝑗 𝜔𝐼 = (−1)𝑟−1 𝜔𝐼𝑟

where 𝐼𝑟 = (𝑖1 , …, 𝚤𝑟̂ , …, 𝑖𝑘 ) (i.e., 𝐼𝑟 is obtained from the multi-index 𝐼 by deleting 𝑖𝑟 ).
 Draft: March 28, 2018

§1.8 The pullback operation on 𝛬𝑘 (𝑉⋆ ) 25

Exercises for §1.7

Exercise 1.7.i. Prove Lemma 1.7.4.
Exercise 1.7.ii. Prove Lemma 1.7.6.
Exercise 1.7.iii. Show that if 𝑇 ∈ A 𝑘 , 𝜄𝑣 𝑇 = 𝑘𝑇𝑣 where 𝑇𝑣 is the tensor (1.3.21). In particular
conclude that 𝜄𝑣 𝑇 ∈ A 𝑘−1 (𝑉). (See Exercise 1.4.viii.)
Exercise 1.7.iv. Assume the dimension of 𝑉 is 𝑛 and let 𝛺 be a nonzero element of the one
dimensional vector space 𝛬𝑛 (𝑉⋆ ). Show that the map
(1.7.17) 𝜌 ∶ 𝑉 → 𝛬𝑛−1 (𝑉⋆ ) , 𝑣 ↦ 𝜄𝑣 𝛺 ,
is a bijective linear map. Hint: One can assume 𝛺 = 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ where 𝑒1 , …, 𝑒𝑛 is a basis of
𝑉. Now use equation (1.7.16) to compute this map on basis elements.
Exercise 1.7.v (cross-product). Let 𝑉 be a 3-dimensional vector space, 𝐵 an inner product
on 𝑉 and 𝛺 a nonzero element of 𝛬3 (𝑉⋆ ). Define a map
− × −∶ 𝑉 × 𝑉 → 𝑉
by setting
𝑣1 × 𝑣2 ≔ 𝜌−1 (𝐿𝑣1 ∧ 𝐿𝑣2 )
where 𝜌 is the map (1.7.17) and 𝐿 ∶ 𝑉 → 𝑉⋆ the map (1.2.17). Show that this map is linear
in 𝑣1 , with 𝑣2 fixed and linear in 𝑣2 with 𝑣1 fixed, and show that 𝑣1 × 𝑣2 = −𝑣2 × 𝑣1 .
Exercise 1.7.vi. For 𝑉 = 𝐑3 let 𝑒1 , 𝑒2 and 𝑒3 be the standard basis vectors and 𝐵 the standard
inner product. (See § 1.1.) Show that if 𝛺 = 𝑒1⋆ ∧ 𝑒2⋆ ∧ 𝑒3⋆ the cross-product above is the
standard cross-product:
𝑒1 × 𝑒2 = 𝑒3
𝑒2 × 𝑒3 = 𝑒1
𝑒3 × 𝑒1 = 𝑒2 .
Hint: If 𝐵 is the standard inner product, then 𝐿𝑒𝑖 = 𝑒𝑖⋆ .
Remark 1.7.18. One can make this standard cross-product look even more standard by
using the calculus notation: 𝑒1 = 𝚤,̂ 𝑒2 = 𝚥,̂ and 𝑒3 = ̂𝑘.

1.8. The pullback operation on 𝛬𝑘 (𝑉⋆ )

Let 𝑉 and 𝑊 be vector spaces and let 𝐴 be a linear map of 𝑉 into 𝑊. Given a 𝑘-tensor
𝑇 ∈ L𝑘 (𝑊), recall that the pullback 𝐴⋆ 𝑇 is the 𝑘-tensor
(𝐴⋆ 𝑇)(𝑣1 , …, 𝑣𝑘 ) = 𝑇(𝐴𝑣1 , …, 𝐴𝑣𝑘 )
in L𝑘 (𝑉). (See §1.3 and equation (1.3.17).) In this section we’ll show how to define a similar
pullback operation on 𝛬𝑘 (𝑉⋆ ).
Lemma 1.8.1. If 𝑇 ∈ I𝑘 (𝑊), then 𝐴⋆ 𝑇 ∈ I𝑘 (𝑉).
Proof. It suffices to verify this when 𝑇 is a redundant 𝑘-tensor, i.e., a tensor of the form
𝑇 = ℓ1 ⊗ ⋯ ⊗ ℓ𝑘
⋆
where ℓ𝑟 ∈ 𝑊 and ℓ𝑖 = ℓ𝑖+1 for some index, 𝑖. But by equation (1.3.19),
𝐴⋆ 𝑇 = 𝐴⋆ ℓ1 ⊗ ⋯ ⊗ 𝐴⋆ ℓ𝑘
and the tensor on the right is redundant since 𝐴⋆ ℓ𝑖 = 𝐴⋆ ℓ𝑖+1 . □
 Draft: March 28, 2018

26 Chapter 1: Multilinear Algebra

Now let 𝜔 be an element of 𝛬𝑘 (𝑊⋆ ) and let 𝜔 = 𝜋(𝑇) where 𝑇 is in L𝑘 (𝑊). We define
(1.8.2) 𝐴⋆ 𝜔 = 𝜋(𝐴⋆ 𝑇) .
Claim 1.8.3. The left hand side of equation (1.8.2) is well-defined.

Proof. If 𝜔 = 𝜋(𝑇) = 𝜋(𝑇′ ), then 𝑇 = 𝑇′ + 𝑆 for some 𝑆 ∈ I𝑘 (𝑊), and 𝐴⋆ 𝑇 = 𝐴⋆ 𝑇′ + 𝐴⋆ 𝑆.

But 𝐴⋆ 𝑆 ∈ I𝑘 (𝑉), so
𝜋(𝐴⋆ 𝑇′ ) = 𝜋(𝐴⋆ 𝑇) . □

Proposition 1.8.4. The map 𝐴⋆ ∶ 𝛬𝑘 (𝑊⋆ ) → 𝛬𝑘 (𝑉⋆ ) sending 𝜔 ↦ 𝐴⋆ 𝜔 is linear. Moreover,

(1) If 𝜔𝑖 ∈ 𝛬𝑘𝑖 (𝑊⋆ ), 𝑖 = 1, 2, then
𝐴⋆ (𝜔1 ∧ 𝜔2 ) = 𝐴⋆ (𝜔1 ) ∧ 𝐴⋆ (𝜔2 ) .
(2) If 𝑈 is a vector space and 𝐵 ∶ 𝑈 → 𝑉 a linear map, then for 𝜔 ∈ 𝛬𝑘 (𝑊⋆ ),
𝐵⋆ 𝐴⋆ 𝜔 = (𝐴𝐵)⋆ 𝜔 .
We’ll leave the proof of these three assertions as exercises. As a hint, they follow im-
mediately from the analogous assertions for the pullback operation on tensors. (See equa-
tions (1.3.19) and (1.3.20).)
As an application of the pullback operation we’ll show how to use it to define the notion
of determinant for a linear mapping.

Definition 1.8.5. Let 𝑉 be a 𝑛-dimensional vector space. Then dim 𝛬𝑛 (𝑉⋆ ) = ( 𝑛𝑛 ) = 1;

i.e., 𝛬𝑛 (𝑉⋆ ) is a one-dimensional vector space. Thus if 𝐴 ∶ 𝑉 → 𝑉 is a linear mapping, the
induced pullback mapping:
𝐴⋆ ∶ 𝛬𝑛 (𝑉⋆ ) → 𝛬𝑛 (𝑉⋆ ) ,
is just “multiplication by a constant”. We denote this constant by det(𝐴) and call it the de-
terminant of 𝐴, Hence, by definition,
(1.8.6) 𝐴⋆ 𝜔 = det(𝐴)𝜔
for all 𝜔 ∈ 𝛬𝑛 (𝑉⋆ ).
From equation (1.8.6) it is easy to derive a number of basic facts about determinants.
Proposition 1.8.7. If 𝐴 and 𝐵 are linear mappings of 𝑉 into 𝑉, then
det(𝐴𝐵) = det(𝐴) det(𝐵) .
Proof. By (2) and
(𝐴𝐵)⋆ 𝜔 = det(𝐴𝐵)𝜔 = 𝐵⋆ (𝐴⋆ 𝜔)
= det(𝐵)𝐴⋆ 𝜔 = det(𝐵) det(𝐴)𝜔 ,
so det(𝐴𝐵) = det(𝐴) det(𝐵). □

Proposition 1.8.8. Write id𝑉 ∶ 𝑉 → 𝑉 for the identity map. Then det(id𝑉 ) = 1.
We’ll leave the proof as an exercise. As a hint, note that id⋆𝑉 is the identity map on
𝛬 (𝑉⋆ ).
𝑛

Proposition 1.8.9. If 𝐴 ∶ 𝑉 → 𝑉 is not surjective, then det(𝐴) = 0.

 Draft: March 28, 2018

§1.8 The pullback operation on 𝛬𝑘 (𝑉⋆ ) 27

Proof. Let 𝑊 be the image of 𝐴. Then if 𝐴 is not onto, the dimension of 𝑊 is less than 𝑛,
so 𝛬𝑛 (𝑊⋆ ) = 0. Now let 𝐴 = 𝑖𝑊 𝐵 where 𝑖𝑊 is the inclusion map of 𝑊 into 𝑉 and 𝐵 is the
mapping, 𝐴, regarded as a mapping from 𝑉 to 𝑊. Thus if 𝜔 is in 𝛬𝑛 (𝑉⋆ ), then by (2)
𝐴⋆ 𝜔 = 𝐵⋆ 𝑖⋆𝑊 𝜔
and since 𝑖⋆𝑊 𝜔 is in 𝛬𝑛 (𝑊⋆ ) it is zero. □

We will derive by wedge product arguments the familiar “matrix formula” for the deter-
minant. Let 𝑉 and 𝑊 be 𝑛-dimensional vector spaces and let 𝑒1 , …, 𝑒𝑛 be a basis for 𝑉 and
𝑓1 , …, 𝑓𝑛 a basis for 𝑊. From these bases we get dual bases, 𝑒1⋆ , …, 𝑒𝑛⋆ and 𝑓1⋆ , …, 𝑓𝑛⋆ , for 𝑉⋆
and 𝑊⋆ . Moreover, if 𝐴 is a linear map of 𝑉 into 𝑊 and [𝑎𝑖,𝑗 ] the 𝑛 × 𝑛 matrix describing
𝐴 in terms of these bases, then the transpose map, 𝐴⋆ ∶ 𝑊⋆ → 𝑉⋆ , is described in terms of
these dual bases by the 𝑛 × 𝑛 transpose matrix, i.e., if
𝑛
𝐴𝑒𝑗 = ∑ 𝑎𝑖,𝑗 𝑓𝑖 ,
𝑖=1

then
𝑛
𝐴⋆ 𝑓𝑗⋆ = ∑ 𝑎𝑗,𝑖 𝑒𝑖⋆ .
𝑖=1
(See §1.2.) Consider now 𝐴⋆ (𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ ). By (1),
𝐴⋆ (𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ ) = 𝐴⋆ 𝑓1⋆ ∧ ⋯ ∧ 𝐴⋆ 𝑓𝑛⋆
= ∑ (𝑎1,𝑘1 𝑒𝑘⋆1 ) ∧ ⋯ ∧ (𝑎𝑛,𝑘𝑛 𝑒𝑘⋆𝑛 ) .
1≤𝑘1 ,…,𝑘𝑛 ≤𝑛

Thus,
𝐴⋆ (𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ ) = ∑ 𝑎1,𝑘1 …𝑎𝑛,𝑘𝑛 𝑒𝑘⋆1 ∧ ⋯ ∧ 𝑒𝑘⋆𝑛 .
If the multi-index, 𝑘1 , …, 𝑘𝑛 , is repeating, then 𝑒𝑘⋆1 ∧ ⋯ ∧ 𝑒𝑘⋆𝑛 is zero, and if it is not repeating
then we can write
𝑘𝑖 = 𝜎(𝑖) 𝑖 = 1, …, 𝑛
for some permutation, 𝜎, and hence we can rewrite 𝐴⋆ (𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ ) as
𝐴⋆ (𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ ) = ∑ 𝑎1,𝜎(1) ⋯ 𝑎𝑛,𝜎(𝑛) (𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ )𝜎 .
𝜎∈𝑆𝑛

But
(𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ )𝜎 = (−1)𝜎 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆
so we get finally the formula
(1.8.10) 𝐴⋆ (𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ ) = det([𝑎𝑖,𝑗 ])𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆
where
(1.8.11) det([𝑎𝑖,𝑗 ]) = ∑ (−1)𝜎 𝑎1,𝜎(1) ⋯ 𝑎𝑛,𝜎(𝑛)
𝜎∈𝑆𝑛

summed over 𝜎 ∈ 𝑆𝑛 . The sum on the right is (as most of you know) the determinant of the
matric [𝑎𝑖,𝑗 ].
Notice that if 𝑉 = 𝑊 and 𝑒𝑖 = 𝑓𝑖 , 𝑖 = 1, …, 𝑛, then 𝜔 = 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ = 𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ ,
hence by (1.8.6) and (1.8.10),
det(𝐴) = det[𝑎𝑖,𝑗 ] .
 Draft: March 28, 2018

28 Chapter 1: Multilinear Algebra

Exercises for §1.8

Exercise 1.8.i. Verify the three assertions of Proposition 1.8.4.
Exercise 1.8.ii. Deduce from Proposition 1.8.9 a well-known fact about determinants of
𝑛 × 𝑛 matrices: If two columns are equal, the determinant is zero.
Exercise 1.8.iii. Deduce from Proposition 1.8.7 another well-known fact about determi-
nants of 𝑛 × 𝑛 matrices: If one interchanges two columns, then one changes the sign of the
determinant.
Hint: Let 𝑒1 , …, 𝑒𝑛 be a basis of 𝑉 and let 𝐵 ∶ 𝑉 → 𝑉 be the linear mapping: 𝐵𝑒𝑖 = 𝑒𝑗 ,
𝐵𝑒𝑗 = 𝑒𝑖 and 𝐵𝑒ℓ = 𝑒ℓ , ℓ ≠ 𝑖, 𝑗. What is 𝐵⋆ (𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ )?
Exercise 1.8.iv. Deduce from Propositions 1.8.3 and 1.8.4 another well-known fact about
determinants of 𝑛 × 𝑛 matrix: If (𝑏𝑖,𝑗 ) is the inverse of [𝑎𝑖,𝑗 ], its determinant is the inverse of
the determinant of [𝑎𝑖,𝑗 ].
Exercise 1.8.v. Extract from (1.8.11) a well-known formula for determinants of 2 × 2 ma-
trices:
𝑎 𝑎12
det ( 11 ) = 𝑎11 𝑎22 − 𝑎12 𝑎21 .
𝑎21 , 𝑎22
Exercise 1.8.vi. Show that if 𝐴 = [𝑎𝑖,𝑗 ] is an 𝑛 × 𝑛 matrix and 𝐴⊺ = [𝑎𝑗,𝑖 ] is its transpose
det 𝐴 = det 𝐴⊺ .
Hint: You are required to show that the sums
∑ (−1)𝜎 𝑎1,𝜎(1) …𝑎𝑛,𝜎(𝑛)
𝜎∈𝑆𝑛

and
∑ (−1)𝜎 𝑎𝜎(1),1 …𝑎𝜎(𝑛),𝑛
𝜎∈𝑆𝑛
are the same. Show that the second sum is identical with
−1
∑ (−1)𝜎 𝑎𝜎−1 (1),1 …𝑎𝜎−1 (𝑛),𝑛 .
𝜎∈𝑆𝑛

Exercise 1.8.vii. Let 𝐴 be an 𝑛 × 𝑛 matrix of the form

𝐵 ∗
𝐴=( ),
0 𝐶
where 𝐵 is a 𝑘 × 𝑘 matrix and 𝐶 the ℓ × ℓ matrix and the bottom ℓ × 𝑘 block is zero. Show
that
det(𝐴) = det(𝐵) det(𝐶) .
Hint: Show that in equation (1.8.11) every nonzero term is of the form
(−1)𝜎𝜏 𝑏1,𝜎(1) …𝑏𝑘,𝜎(𝑘) 𝑐1,𝜏(1) …𝑐ℓ,𝜏(ℓ)
where 𝜎 ∈ 𝑆𝑘 and 𝜏 ∈ 𝑆ℓ .
Exercise 1.8.viii. Let 𝑉 and 𝑊 be vector spaces and let 𝐴 ∶ 𝑉 → 𝑊 be a linear map. Show
that if 𝐴𝑣 = 𝑤 then for 𝜔 ∈ 𝛬𝑝 (𝑊⋆ ),
𝐴⋆ 𝜄𝑤 𝜔 = 𝜄𝑣 𝐴⋆ 𝜔 .
Hint: By equation (1.7.14) and Proposition 1.8.4 it suffices to prove this for 𝜔 ∈ 𝛬1 (𝑊⋆ ),
i.e., for 𝜔 ∈ 𝑊⋆ .
 Draft: March 28, 2018

§1.9 Orientations 29

1.9. Orientations
Definition 1.9.1. We recall from freshman calculus that if ℓ ⊂ 𝐑2 is a line through the
origin, then ℓ ∖ {0} has two connected components and an orientation of ℓ is a choice of
one of these components (as in the Figure 1.9.1 below).

Figure 1.9.1. A line in 𝐑2

More generally, if 𝐿 is a one-dimensional vector space then 𝐿 ∖ {0} consists of two com-
ponents: namely if 𝑣 is an element of 𝐿 ∖ {0}, then these two components are
𝐿1 = { 𝜆𝑣 | 𝜆 > 0 }
and
𝐿2 = { 𝜆𝑣 | 𝜆 < 0 } .
Definition 1.9.2. Let 𝐿 be a one-dimensional vector space. An orientation of 𝐿 is a choice of
one of a connected component of 𝐿 ∖ {0}. Usually the component chosen is denoted 𝐿+ , and
called the positive component of 𝐿∖{0} and the other component 𝐿− the negative component
of 𝐿 ∖ {0}.
Definition 1.9.3. Let (𝐿, 𝐿+ ) be an oriented one-dimensional vector space. A vector 𝑣 ∈ 𝐿
is positively oriented if 𝑣 ∈ 𝐿+ .
Definition 1.9.4. Let 𝑉 be an 𝑛-dimensional vector space. An orientation of 𝑉 is an orien-
tation of the one-dimensional vector space 𝛬𝑛 (𝑉⋆ ).
One important way of assigning an orientation to 𝑉 is to choose a basis, 𝑒1 , …, 𝑒𝑛 of 𝑉.
Then, if 𝑒1⋆ , …, 𝑒𝑛⋆ is the dual basis, we can orient 𝛬𝑛 (𝑉⋆ ) by requiring that 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ be
in the positive component of 𝛬𝑛 (𝑉⋆ ).
Definition 1.9.5. Let 𝑉 be an oriented 𝑛-dimensional vector space. We say that an ordered
basis (𝑒1 , …, 𝑒𝑛 ) of 𝑉 is positively oriented if 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ is in the positive component of
𝛬𝑛 (𝑉⋆ ).
Suppose that 𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑛 are bases of 𝑉 and that
𝑛
(1.9.6) 𝑒𝑗 = ∑ 𝑎𝑖,𝑗, 𝑓𝑖 .
𝑖=1
Then by (1.7.10)
𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ = det[𝑎𝑖,𝑗 ]𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆
so we conclude:
 Draft: March 28, 2018

30 Chapter 1: Multilinear Algebra

Proposition 1.9.7. If 𝑒1 , …, 𝑒𝑛 is positively oriented, then 𝑓1 , …, 𝑓𝑛 is positively oriented if and

only if det[𝑎𝑖,𝑗 ] is positive.
Corollary 1.9.8. If 𝑒1 , …, 𝑒𝑛 is a positively oriented basis of 𝑉, then the basis
𝑒1 , …, 𝑒𝑖−1 , −𝑒𝑖 , 𝑒𝑖+1 , …, 𝑒𝑛
is negatively oriented.
Now let 𝑉 be a vector space of dimension 𝑛 > 1 and 𝑊 a subspace of dimension 𝑘 < 𝑛.
We will use the result above to prove the following important theorem.
Theorem 1.9.9. Given orientations on 𝑉 and 𝑉/𝑊, one gets from these orientations a natural
orientation on 𝑊.
Remark 1.9.10. What we mean by “natural” will be explained in the course of the proof.
Proof. Let 𝑟 = 𝑛 − 𝑘 and let 𝜋 be the projection of 𝑉 onto 𝑉/𝑊 . By Exercises 1.2.i and 1.2.ii
we can choose a basis 𝑒1 , …, 𝑒𝑛 of 𝑉 such that 𝑒𝑟+1 , …, 𝑒𝑛 is a basis of 𝑊 and 𝜋(𝑒1 ), …, 𝜋(𝑒𝑟 )
a basis of 𝑉/𝑊. Moreover, replacing 𝑒1 by −𝑒1 if necessary we can assume by Corollary 1.9.8
that 𝜋(𝑒1 ), …, 𝜋(𝑒𝑟 ) is a positively oriented basis of 𝑉/𝑊 and replacing 𝑒𝑛 by −𝑒𝑛 if neces-
sary we can assume that 𝑒1 , …, 𝑒𝑛 is a positively oriented basis of 𝑉. Now assign to 𝑊 the
orientation associated with the basis 𝑒𝑟+1 , …, 𝑒𝑛 .
Let’s show that this assignment is “natural”(i.e., does not depend on our choice of basis
𝑒1 , …, 𝑒𝑛 ). To see this let 𝑓1 , …, 𝑓𝑛 be another basis of 𝑉 with the properties above and let
𝐴 = [𝑎𝑖,𝑗 ] be the matrix (1.9.6) expressing the vectors 𝑒1 , …, 𝑒𝑛 as linear combinations of the
vectors 𝑓1 , …𝑓𝑛 . This matrix has to have the form
𝐵 𝐶
(1.9.11) 𝐴=( )
0 𝐷
where 𝐵 is the 𝑟 × 𝑟 matrix expressing the basis vectors 𝜋(𝑒1 ), …, 𝜋(𝑒𝑟 ) of 𝑉/𝑊 as linear com-
binations of 𝜋(𝑓1 ), …, 𝜋(𝑓𝑟 ) and 𝐷 the 𝑘 × 𝑘 matrix expressing the basis vectors 𝑒𝑟+1 , …, 𝑒𝑛
of 𝑊 as linear combinations of 𝑓𝑟+1 , …, 𝑓𝑛 . Thus
det(𝐴) = det(𝐵) det(𝐷) .
However, by Proposition 1.9.7, det 𝐴 and det 𝐵 are positive, so det 𝐷 is positive, and hence
if 𝑒𝑟+1 , …, 𝑒𝑛 is a positively oriented basis of 𝑊 so is 𝑓𝑟+1 , …, 𝑓𝑛 . □

As a special case of this theorem suppose dim 𝑊 = 𝑛 − 1. Then the choice of a vector
𝑣 ∈ 𝑉 ∖ 𝑊 gives one a basis vector 𝜋(𝑣) for the one-dimensional space 𝑉/𝑊 and hence if 𝑉
is oriented, the choice of 𝑣 gives one a natural orientation on 𝑊.
Definition 1.9.12. Let 𝐴 ∶ 𝑉1 → 𝑉2 a bijective linear map of oriented 𝑛-dimensional vector
spaces. We say that 𝐴 is orientation-preserving if, for 𝜔 ∈ 𝛬𝑛 (𝑉2⋆ )+ , we have that 𝐴⋆ 𝜔 is in
𝛬𝑛 (𝑉1⋆ )+ .
Example 1.9.13. If 𝑉1 = 𝑉2 then 𝐴⋆ 𝜔 = det(𝐴)𝜔 so 𝐴 is orientation preserving if and only
if det(𝐴) > 0.
The following proposition we’ll leave as an exercise.
Proposition 1.9.14. Let 𝑉1 , 𝑉2 , and 𝑉3 be oriented 𝑛-dimensional vector spaces and 𝐴 𝑖 ∶ 𝑉𝑖 ⥲
𝑉𝑖+1 for 𝑖 = 1, 2 be bijective linear maps. Then if 𝐴 1 and 𝐴 2 are orientation preserving, so is
𝐴2 ∘ 𝐴1 .
 Draft: March 28, 2018

§1.9 Orientations 31

Exercises for §1.9

Exercise 1.9.i. Prove Corollary 1.9.8.
Exercise 1.9.ii. Show that the argument in the proof of Theorem 1.9.9 can be modified to
prove that if 𝑉 and 𝑊 are oriented then these orientations induce a natural orientation on
𝑉/𝑊.
Exercise 1.9.iii. Similarly show that if 𝑊 and 𝑉/𝑊 are oriented these orientations induce
a natural orientation on 𝑉.
Exercise 1.9.iv. Let 𝑉 be an 𝑛-dimensional vector space and 𝑊 ⊂ 𝑉 a 𝑘-dimensional sub-
space. Let 𝑈 = 𝑉/𝑊 and let 𝜄 ∶ 𝑊 → 𝑉 and 𝜋 ∶ 𝑉 → 𝑈 be the inclusion and projection
maps. Suppose 𝑉 and 𝑈 are oriented. Let 𝜇 be in 𝛬𝑛−𝑘 (𝑈⋆ )+ and let 𝜔 be in 𝛬𝑛 (𝑉⋆ )+ . Show
that there exists a 𝜈 in 𝛬𝑘 (𝑉⋆ ) such that 𝜋⋆ 𝜇 ∧ 𝜈 = 𝜔. Moreover show that 𝜄⋆ 𝜈 is intrinsically
defined (i.e., does not depend on how we choose 𝜈) and sits in the positive part 𝛬𝑘 (𝑊⋆ )+ of
𝛬𝑘 (𝑊⋆ ).
Exercise 1.9.v. Let 𝑒1 , …, 𝑒𝑛 be the standard basis vectors of 𝐑𝑛 . The standard orientation of
𝐑𝑛 is, by definition, the orientation associated with this basis. Show that if 𝑊 is the subspace
of 𝐑𝑛 defined by the equation 𝑥1 = 0, and 𝑣 = 𝑒1 ∉ 𝑊 then the natural orientation of 𝑊
associated with 𝑣 and the standard orientation of 𝐑𝑛 coincide with the orientation given by
the basis vectors 𝑒2 , …, 𝑒𝑛 of 𝑊.
Exercise 1.9.vi. Let 𝑉 be an oriented 𝑛-dimensional vector space and 𝑊 an 𝑛−1-dimensional
subspace. Show that if 𝑣 and 𝑣′ are in 𝑉∖𝑊 then 𝑣′ = 𝜆𝑣+𝑤, where 𝑤 is in 𝑊 and 𝜆 ∈ 𝐑∖{0}.
Show that 𝑣 and 𝑣′ give rise to the same orientation of 𝑊 if and only if 𝜆 is positive.
Exercise 1.9.vii. Prove Proposition 1.9.14.
Exercise 1.9.viii. A key step in the proof of Theorem 1.9.9 was the assertion that the matrix
𝐴 expressing the vectors, 𝑒𝑖 as linear combinations of the vectors 𝑓𝐽 , had to have the form
(1.9.11). Why is this the case?
Exercise 1.9.ix.
(1) Let 𝑉 be a vector space, 𝑊 a subspace of 𝑉 and 𝐴 ∶ 𝑉 → 𝑉 a bijective linear map which
maps 𝑊 onto 𝑊. Show that one gets from 𝐴 a bijective linear map
𝐵 ∶ 𝑉/𝑊 → 𝑉/𝑊
with the property
𝜋𝐴 = 𝐵𝜋 ,
where 𝜋 ∶ 𝑉 → 𝑉/𝑊 is the quotient map.
(2) Assume that 𝑉, 𝑊 and 𝑉/𝑊 are compatibly oriented. Show that if 𝐴 is orientation-
preserving and its restriction to 𝑊 is orientation preserving then 𝐵 is orientation pre-
serving.
Exercise 1.9.x. Let 𝑉 be a oriented 𝑛-dimensional vector space, 𝑊 an (𝑛 − 1)-dimensional
subspace of 𝑉 and 𝑖 ∶ 𝑊 → 𝑉 the inclusion map. Given 𝜔 ∈ 𝛬𝑛 (𝑉⋆ )+ and 𝑣 ∈ 𝑉 ∖ 𝑊 show
that for the orientation of 𝑊 described in Exercise 1.9.v, 𝑖⋆ (𝜄𝑣 𝜔) ∈ 𝛬𝑛−1 (𝑊⋆ )+ .
Exercise 1.9.xi. Let 𝑉 be an 𝑛-dimensional vector space, 𝐵 ∶ 𝑉 × 𝑉 → 𝐑 an inner prod-
uct and 𝑒1 , …, 𝑒𝑛 a basis of 𝑉 which is positively oriented and orthonormal. Show that the
volume element
vol ≔ 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ ∈ 𝛬𝑛 (𝑉⋆ )
 Draft: March 28, 2018

32 Chapter 1: Multilinear Algebra

is intrinsically defined, independent of the choice of this basis.

Hint: Equations (1.2.20) and (1.8.10).
Exercise 1.9.xii.
(1) Let 𝑉 be an oriented 𝑛-dimensional vector space and 𝐵 an inner product on 𝑉. Fix an
oriented orthonormal basis, 𝑒1 , …, 𝑒𝑛 , of 𝑉 and let 𝐴 ∶ 𝑉 → 𝑉 be a linear map. Show
that if
𝑛
𝐴𝑒𝑖 = 𝑣𝑖 = ∑ 𝑎𝑗,𝑖 𝑒𝑗
𝑗=1
and 𝑏𝑖,𝑗 = 𝐵(𝑣𝑖 , 𝑣𝑗 ), the matrices 𝑀𝐴 = [𝑎𝑖,𝑗 ] and 𝑀𝐵 = (𝑏𝑖,𝑗 ) are related by: 𝑀𝐵 =
𝑀𝐴⊺ 𝑀𝐴 .
(2) Show that if vol is the volume form 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ , and 𝐴 is orientation preserving
1
𝐴⋆ vol = det(𝑀𝐵 ) 2 vol .
(3) ByTheorem 1.5.13 one has a bijective map
𝛬𝑛 (𝑉⋆ ) ≅ A 𝑛 (𝑉) .
Show that the element, 𝛺, of 𝐴𝑛 (𝑉) corresponding to the form vol has the property
|𝛺(𝑣1 , …, 𝑣𝑛 )|2 = det((𝑏𝑖,𝑗 ))
where 𝑣1 , …, 𝑣𝑛 are any 𝑛-tuple of vectors in 𝑉 and 𝑏𝑖,𝑗 = 𝐵(𝑣𝑖 , 𝑣𝑗 ).
 Draft: March 28, 2018

CHAPTER 2

Differential Forms

The goal of this chapter is to generalize to 𝑛 dimensions the basic operations of three
dimensional vector calculus: divergence, curl, and gradient. The divergence and gradient
operations have fairly straightforward generalizations, but the curl operation is more subtle.
For vector fields it does not have any obvious generalization, however, if one replaces vector
fields by a closely related class of objects, differential forms, then not only does it have a
natural generalization but it turns out that divergence, curl, and gradient are all special cases
of a general operation on differential forms called exterior differentiation.

2.1. Vector fields and one-forms

In this section we will review some basic facts about vector fields in 𝑛 variables and
introduce their dual objects: one-forms. We will then take up in §2.3 the theory of 𝑘-forms
for 𝑘 > 1. We begin by fixing some notation.
Definition 2.1.1. Let 𝑝 ∈ 𝐑𝑛 . The tangent space to 𝐑𝑛 at 𝑝 is the set of pairs
𝑇𝑝 𝐑𝑛 ≔ { (𝑝, 𝑣) | 𝑣 ∈ 𝐑𝑛 } .
The identification
(2.1.2) 𝑇𝑝 𝐑𝑛 ⥲ 𝐑𝑛 , (𝑝, 𝑣) ↦ 𝑣
makes 𝑇𝑝 𝐑𝑛 into a vector space. More explicitly, for 𝑣, 𝑣1 , 𝑣2 ∈ 𝐑𝑛 and 𝜆 ∈ 𝐑 we define the
addition and scalar multiplication operations on 𝑇𝑝 𝐑𝑛 by setting
(𝑝, 𝑣1 ) + (𝑝, 𝑣2 ) ≔ (𝑝, 𝑣1 + 𝑣2 )
and
𝜆(𝑝, 𝑣) ≔ (𝑝, 𝜆𝑣) .
Let 𝑈 be an open subset of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝐑𝑚 a 𝐶1 map. We recall that the derivative
𝐷𝑓(𝑝) ∶ 𝐑𝑛 → 𝐑𝑚
of 𝑓 at 𝑝 is the linear map associated with the 𝑚 × 𝑛 matrix
𝜕𝑓𝑖
[ (𝑝)] .
𝜕𝑥𝑗
It will be useful to have a “base-pointed” version of this definition as well. Namely, if 𝑞 = 𝑓(𝑝)
we will define
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝐑𝑛 → 𝑇𝑞 𝐑𝑚
to be the map
𝑑𝑓𝑝 (𝑝, 𝑣) ≔ (𝑞, 𝐷𝑓(𝑝)𝑣) .
It is clear from the way we have defined vector space structures on 𝑇𝑝 𝐑𝑛 and 𝑇𝑞 𝐑𝑚 that this
map is linear.
33
 Draft: March 28, 2018

34 Chapter 2: Differential Forms

Suppose that the image of 𝑓 is contained in an open set 𝑉, and suppose 𝑔 ∶ 𝑉 → 𝐑𝑘 is

1
a 𝐶 map. Then the “base-pointed” version of the chain rule asserts that
𝑑𝑔𝑞 ∘ 𝑑𝑓𝑝 = 𝑑(𝑓 ∘ 𝑔)𝑝 .
(This is just an alternative way of writing 𝐷𝑔(𝑞)𝐷𝑓(𝑝) = 𝐷(𝑔 ∘ 𝑓)(𝑝).)
In 3-dimensional vector calculus a vector field is a function which attaches to each point
𝑝 of 𝐑3 a base-pointed arrow (𝑝, 𝑣) ∈ 𝑇𝑝 𝐑3 . The 𝑛-dimensional version of this definition is
essentially the same.
Definition 2.1.3. Let 𝑈 be an open subset of 𝐑𝑛 . A vector field 𝒗 on 𝑈 is a function which
assigns to each point 𝑝 ∈ 𝑈 a vector 𝒗(𝑝) ∈ 𝑇𝑝 𝐑𝑛 .
Thus a vector field is a vector-valued function, but its value at 𝑝 is an element of a vector
space 𝑇𝑝 𝐑𝑛 that itself depends on 𝑝.
Examples 2.1.4.
(1) Given a fixed vector 𝑣 ∈ 𝐑𝑛 , the function
(2.1.5) 𝑝 ↦ (𝑝, 𝑣)
𝑛
is a vector field on 𝐑 . Vector fields of this type are called constant vector fields.
(2) In particular let 𝑒1 , …, 𝑒𝑛 be the standard basis vectors of 𝐑𝑛 . If 𝑣 = 𝑒𝑖 we will denote the
vector field (2.1.5) by 𝜕/𝜕𝑥𝑖 . (The reason for this “derivation notation” will be explained
below.)
(3) Given a vector field 𝒗 on 𝑈 and a function, 𝑓 ∶ 𝑈 → 𝐑 we denote by 𝑓𝒗 the vector field
on 𝑈 defined by
𝑝 ↦ 𝑓(𝑝)𝒗(𝑝) .
(4) Given vector fields 𝒗1 and 𝒗2 on 𝑈, we denote by 𝒗1 + 𝒗2 the vector field on 𝑈 defined
by
𝑝 ↦ 𝒗1 (𝑝) + 𝒗2 (𝑝) .
(5) The vectors, (𝑝, 𝑒𝑖 ), 𝑖 = 1, …, 𝑛, are a basis of 𝑇𝑝 𝐑𝑛 , so if 𝒗 is a vector field on 𝑈, 𝒗(𝑝)
can be written uniquely as a linear combination of these vectors with real numbers
𝑔1 (𝑝), …, 𝑔𝑛 (𝑝) as coefficients. In other words, using the notation in example (2) above,
𝒗 can be written uniquely as a sum
𝑛
𝜕
(2.1.6) 𝒗 = ∑ 𝑔𝑖
𝑖=1 𝜕𝑥𝑖
where 𝑔𝑖 ∶ 𝑈 → 𝐑 is the function 𝑝 ↦ 𝑔𝑖 (𝑝).
Definition 2.1.7. We say that 𝒗 is a 𝐶∞ vector field if the 𝑔𝑖 ’s are in 𝐶∞ (𝑈).
Definition 2.1.8. A basic vector field operation is Lie differentiation. If 𝑓 ∈ 𝐶1 (𝑈) we define
𝐿𝒗 𝑓 to be the function on 𝑈 whose value at 𝑝 is given by
(2.1.9) 𝐿𝒗 𝑓(𝑝) ≔ 𝐷𝑓(𝑝)𝑣 ,
where 𝒗(𝑝) = (𝑝, 𝑣).
Example 2.1.10. If 𝒗 is the vector field (2.1.6) then
𝑛
𝜕𝑓
𝐿𝒗 𝑓 = ∑ 𝑔𝑖 ,
𝑖=1 𝜕𝑥𝑖
(motivating our “derivation notation” for 𝒗).
 Draft: March 28, 2018

§2.1 Vector fields and one-forms 35

We leave the following generalization of the product rule for differentiation as an exer-
cise.
Lemma 2.1.11. Let 𝑈 be an open subset of 𝐑𝑛 , 𝒗 a vector field on 𝑈, and 𝑓1 , 𝑓2 ∈ 𝐶1 (𝑈). Then
𝐿𝒗 (𝑓1 ⋅ 𝑓2 ) = 𝐿𝒗 (𝑓1 ) ⋅ 𝑓2 + 𝑓1 ⋅ 𝐿𝒗 (𝑓2 ) .
We now discuss a class of objects which are in some sense “dual objects” to vector fields.
Definition 2.1.12. Let 𝑝 ∈ 𝐑𝑛 . The cotangent space to 𝐑𝑛 at 𝑝 is the dual vector space
𝑇𝑝⋆ 𝐑𝑛 ≔ (𝑇𝑝 𝐑)⋆ .
An element of 𝑇𝑝⋆ 𝐑𝑛 is called a cotangent vector to 𝐑𝑛 at 𝑝.
Definition 2.1.13. Let 𝑈 be an open subset of 𝐑𝑛 . A differential one-form, or simply one-
form on 𝑈 is a function 𝜔 which assigns to each point 𝑝 ∈ 𝑈 a vector 𝜔𝑝 in 𝑇𝑝⋆ 𝐑𝑛 .
Examples 2.1.14.
(1) Let 𝑓 ∶ 𝑈 → 𝐑 be a 𝐶1 function. Then for 𝑝 ∈ 𝑈 and 𝑞 = 𝑓(𝑝) one has a linear map
(2.1.15) 𝑑𝑓𝑝 ∶ 𝑇𝑝 𝐑𝑛 → 𝑇𝑞 𝐑
and by making the identification 𝑇𝑞 𝐑 ⥲ 𝐑 by sending (𝑞, 𝑣) ↦ 𝑣, the differential 𝑑𝑓𝑝
can be regarded as a linear map from 𝑇𝑝 𝐑𝑛 to 𝐑, i.e., as an element of 𝑇𝑝⋆ 𝐑𝑛 . Hence the
assignment
𝑝 ↦ 𝑑𝑓𝑝
defines a one-form on 𝑈 which we denote by 𝑑𝑓.
(2) Given a one-form 𝜔 and a function 𝜙 ∶ 𝑈 → 𝐑 the pointwise product of 𝜙 with 𝜔 defines
one-form 𝜙𝜔 on 𝑈 by (𝜙𝜔)𝑝 ≔ 𝜙(𝑝)𝜔𝑝 .
(3) Given two one-forms 𝜔1 and 𝜔2 their pointwise sum defines a one-form 𝜔1 + 𝜔2 by
(𝜔1 + 𝜔2 )𝑝 ≔ (𝜔1 )𝑝 + (𝜔2 )𝑝 .
(4) The one-forms 𝑑𝑥1 , …, 𝑑𝑥𝑛 play a particularly important role. By equation (2.1.15)
𝜕
(2.1.16) (𝑑𝑥𝑖 ) ( ) = 𝛿𝑖,𝑗
𝜕𝑥𝑗 𝑝
i.e., is equal to 1 if 𝑖 = 𝑗 and zero if 𝑖 ≠ 𝑗. Thus (𝑑𝑥1 )𝑝 , …, (𝑑𝑥𝑛 )𝑝 are the basis of
𝑇𝑝⋆ 𝐑𝑛 dual to the basis (𝜕/𝜕𝑥𝑖 )𝑝 . Therefore, if 𝜔 is any one-form on 𝑈, 𝜔𝑝 can be written
uniquely as a sum
𝑛
𝜔𝑝 = ∑ 𝑓𝑖 (𝑝)(𝑑𝑥𝑖 )𝑝 , 𝑓𝑖 (𝑝) ∈ 𝐑 ,
𝑖=1
and 𝜔 can be written uniquely as a sum
𝑛
(2.1.17) 𝜔 = ∑ 𝑓𝑖 𝑑𝑥𝑖
𝑖=1
where 𝑓𝑖 ∶ 𝑈 → 𝐑 is the function 𝑝 ↦ 𝑓𝑖 (𝑝). We say that 𝜔 is a 𝐶∞ one-form or smooth
one-form if the functions 𝑓1 , …, 𝑓𝑛 are 𝐶∞ .
We leave the following as an exercise.
Lemma 2.1.18. Let 𝑈 be an open subset of 𝐑𝑛 . If 𝑓 ∶ 𝑈 → 𝐑 is a 𝐶∞ function, then
𝑛
𝜕𝑓
𝑑𝑓 = ∑ 𝑑𝑥𝑖 .
𝑖=1 𝜕𝑥𝑖
 Draft: March 28, 2018

36 Chapter 2: Differential Forms

Suppose that 𝒗 is a vector field and 𝜔 a one-form on 𝑈 ⊂ 𝐑𝑛 . Then for every 𝑝 ∈ 𝑈 the
vectors 𝒗(𝑝) ∈ 𝑇𝑝 𝐑𝑛 and 𝜔𝑝 ∈ (𝑇𝑝 𝐑𝑛 )⋆ can be paired to give a number 𝜄𝒗(𝑝) 𝜔𝑝 ∈ 𝐑, and
hence, as 𝑝 varies, an 𝐑-valued function 𝜄𝒗 𝜔 which we will call the interior product of 𝒗 with
𝜔.
Example 2.1.19. For instance if 𝒗 is the vector field (2.1.6) and 𝜔 the one-form (2.1.17),
then
𝑛
𝜄𝒗 𝜔 = ∑ 𝑓𝑖 𝑔𝑖 .
𝑖=1
Thus if 𝒗 and 𝜔 are 𝐶∞ , so is the function 𝜄𝒗 𝜔. Also notice that if 𝜙 ∈ 𝐶∞ (𝑈), then as we
observed above
𝑛
𝜕𝜙 𝜕
𝑑𝜙 = ∑
𝑖=1 𝜕𝑥𝑖 𝜕𝑥𝑖
so if 𝒗 is the vector field (2.1.6) then we have
𝑛
𝜕𝜙
𝜄𝒗 𝑑𝜙 = ∑ 𝑔𝑖 = 𝐿𝒗 𝜙 .
𝑖=1 𝜕𝑥𝑖
Exercises for §2.1
Exercise 2.1.i. Prove Lemma 2.1.18
Exercise 2.1.ii. Prove Lemma 2.1.11
Exercise 2.1.iii. Let 𝑈 be an open subset of 𝐑𝑛 and 𝒗1 and 𝒗2 vector fields on 𝑈. Show that
there is a unique vector field 𝒘, on 𝑈 with the property
𝐿𝒘 𝜙 = 𝐿𝒗1 (𝐿𝒗2 𝜙) − 𝐿𝒗2 (𝐿𝒗1 𝜙)
∞
for all 𝜙 ∈ 𝐶 (𝑈).
Exercise 2.1.iv. The vector field 𝒘 in Exercise 2.1.iii is called the Lie bracket of the vector
fields 𝒗1 and 𝒗2 and is denoted by [𝒗1 , 𝒗2 ]. Verify that the Lie bracket is skew-symmetric, i.e.,
[𝒗1 , 𝒗2 ] = −[𝒗2 , 𝒗1 ] ,
and satisfies the Jacobi identity
[𝒗1 , [𝒗2 , 𝒗3 ]] + [𝒗2 , [𝒗3 , 𝒗1 ]] + [𝒗3 , [𝒗1 , 𝒗2 ]] = 0 .
(And thus defines the structure of a Lie algebra.)
Hint: Prove analogous identities for 𝐿𝒗1 , 𝐿𝒗2 and 𝐿𝒗3 .
𝑛
Exercise 2.1.v. Let 𝒗1 = 𝜕/𝜕𝑥𝑖 and 𝒗2 = ∑𝑗=1 𝑔𝑗 𝜕/𝜕𝑥𝑗 . Show that
𝑛
𝜕𝑔𝑗 𝜕
[𝒗1 , 𝒗2 ] = ∑ .
𝑗=1 𝜕𝑥𝑖 𝜕𝑥𝑗

Exercise 2.1.vi. Let 𝒗1 and 𝒗2 be vector fields and 𝑓 a 𝐶∞ function. Show that
[𝒗1 , 𝑓𝒗2 ] = 𝐿𝒗1 𝑓𝒗2 + 𝑓[𝒗1 , 𝒗2 ] .
Exercise 2.1.vii. Let 𝑈 be an open subset of 𝐑𝑛 and let 𝛾 ∶ [𝑎, 𝑏] → 𝑈, 𝑡 ↦ (𝛾1 (𝑡), …, 𝛾𝑛 (𝑡))
𝑛
be a 𝐶1 curve. Given a 𝐶∞ one-form 𝜔 = ∑𝑖=1 𝑓𝑖 𝑑𝑥𝑖 on 𝑈, define the line integral of 𝜔 over
𝛾 to be the integral
𝑛 𝑏
𝑑𝛾
∫ 𝜔 ≔ ∑ ∫ 𝑓𝑖 (𝛾(𝑡)) 𝑖 𝑑𝑡 .
𝛾 𝑖=1 𝑎 𝑑𝑡
 Draft: March 28, 2018

§2.2 Integral Curves for Vector Fields 37

Show that if 𝜔 = 𝑑𝑓 for some 𝑓 ∈ 𝐶∞ (𝑈)

∫ 𝜔 = 𝑓(𝛾(𝑏)) − 𝑓(𝛾(𝑎)) .
𝛾
In particular conclude that if 𝛾 is a closed curve, i.e., 𝛾(𝑎) = 𝛾(𝑏), this integral is zero.
Exercise 2.1.viii. Let 𝜔 be the 𝐶∞ one-form on 𝐑2 ∖ {0} defined by
𝑥 𝑑𝑥 − 𝑥2 𝑑𝑥1
𝜔 = 1 22 ,
𝑥1 + 𝑥22
and let 𝛾 ∶ [0, 2𝜋] → 𝐑2 ∖{0} be the closed curve 𝑡 ↦ (cos 𝑡, sin 𝑡). Compute the line integral
∫𝛾 𝜔, and note that ∫𝛾 𝜔 ≠ 0. Conclude that 𝜔 is not of the form 𝑑𝑓 for 𝑓 ∈ 𝐶∞ (𝐑2 ∖ {0}).

Exercise 2.1.ix. Let 𝑓 be the function

{arctan 𝑥𝑥2 𝑥1 > 0

{𝜋 1
𝑓(𝑥1 , 𝑥2 ) = { 2 , 𝑥2 > 0 or 𝑥1 = 0
{ 𝑥2
{arctan 𝑥1 + 𝜋 𝑥1 < 0 .
Recall that − 𝜋2 < arctan(𝑡) < 𝜋2 . Show that 𝑓 is 𝐶∞ and that 𝑑𝑓 is the 1-form 𝜔 in Exer-
cise 2.1.viii. Why does not this contradict what you proved in Exercise 2.1.viii?

2.2. Integral Curves for Vector Fields

In this section we’ll discuss some properties of vector fields which we’ll need for the
manifold segment of this text. We’ll begin by generalizing to 𝑛-variables the calculus notion
of an “integral curve” of a vector field.
Definition 2.2.1. Let 𝑈 ⊂ 𝐑𝑛 be open and 𝒗 a vector field on 𝑈. A 𝐶1 curve 𝛾 ∶ (𝑎, 𝑏) → 𝑈
is an integral curve of 𝒗 if for all 𝑡 ∈ (𝑎, 𝑏) we have
𝒗(𝛾(𝑡)) = (𝛾(𝑡), 𝑑𝛾
𝑑𝑡
(𝑡)) .
Remark 2.2.2. If 𝒗 is the vector field (2.1.6) and 𝑔 ∶ 𝑈 → 𝐑𝑛 is the function (𝑔1 , …, 𝑔𝑛 ),
the condition for 𝛾(𝑡) to be an integral curve of 𝒗 is that it satisfy the system of differential
equations
𝑑𝛾
(2.2.3) (𝑡) = 𝑔(𝛾(𝑡)) .
𝑑𝑡
We will quote without proof a number of basic facts about systems of ordinary differ-
ential equations of the type (2.2.3).1
Theorem 2.2.4 (existence of integral curves). Let 𝑈 be an open subset of 𝐑𝑛 and 𝒗 a vector
field on 𝑈. Given a point 𝑝0 ∈ 𝑈 and 𝑎 ∈ 𝐑, there exists an interval 𝐼 = (𝑎 − 𝑇, 𝑎 + 𝑇), a
neighborhood 𝑈0 of 𝑝0 in 𝑈, and for every 𝑝 ∈ 𝑈0 an integral curve 𝛾𝑝 ∶ 𝐼 → 𝑈 for 𝒗 such that
𝛾𝑝 (𝑎) = 𝑝.
Theorem 2.2.5 (uniqueness of integral curves). Let 𝑈 be an open subset of 𝐑𝑛 and 𝒗 a vector
field on 𝑈. For 𝑖 = 1, 2, let 𝛾𝑖 ∶ 𝐼𝑖 → 𝑈, 𝑖 = 1, 2 be integral curves for 𝒗. If 𝑎 ∈ 𝐼1 ∩ 𝐼2 and
𝛾1 (𝑎) = 𝛾2 (𝑎) then 𝛾1 |𝐼1 ∩𝐼2 = 𝛾2 |𝐼1 ∩𝐼2 and the curve 𝛾 ∶ 𝐼1 ∪ 𝐼2 → 𝑈 defined by
𝛾1 (𝑡), 𝑡 ∈ 𝐼1
𝛾(𝑡) ≔ {
𝛾2 (𝑡), 𝑡 ∈ 𝐼2
1A source for these results that we highly recommend is [2, Ch. 6].
 Draft: March 28, 2018

38 Chapter 2: Differential Forms

is an integral curve for 𝒗.

Theorem 2.2.6 (smooth dependence on initial data). Let 𝑉 ⊂ 𝑈 ⊂ 𝐑𝑛 be open subsets. Let
𝒗 be a 𝐶∞ -vector field on 𝑉, 𝐼 ⊂ 𝐑 an open interval, 𝑎 ∈ 𝐼, and ℎ ∶ 𝑉 × 𝐼 → 𝑈 a map with
the following properties.
(1) ℎ(𝑝, 𝑎) = 𝑝.
(2) For all 𝑝 ∈ 𝑉 the curve
𝛾𝑝 ∶ 𝐼 → 𝑈 , 𝛾𝑝 (𝑡) ≔ ℎ(𝑝, 𝑡)
is an integral curve of 𝒗.
Then the map ℎ is 𝐶∞ .
One important feature of the system (2.2.3) is that it is an autonomous system of differ-
ential equations: the function 𝑔(𝑥) is a function of 𝑥 alone (it does not depend on 𝑡). One
consequence of this is the following.
Theorem 2.2.7. Let 𝐼 = (𝑎, 𝑏) and for 𝑐 ∈ 𝐑 let 𝐼𝑐 = (𝑎 − 𝑐, 𝑏 − 𝑐). Then if 𝛾 ∶ 𝐼 → 𝑈 is an
integral curve, the reparametrized curve
(2.2.8) 𝛾𝑐 ∶ 𝐼𝑐 → 𝑈 , 𝛾𝑐 (𝑡) ≔ 𝛾(𝑡 + 𝑐)
is an integral curve.
We recall that 𝐶1 -function 𝜙 ∶ 𝑈 → 𝐑 is an integral of the system (2.2.3) if for every
integral curve 𝛾(𝑡), the function 𝑡 ↦ 𝜙(𝛾(𝑡)) is constant. This is true if and only if for all 𝑡
𝑑 𝑑𝛾
0= 𝜙(𝛾(𝑡)) = (𝐷𝜙)𝛾(𝑡) ( ) = (𝐷𝜙)𝛾(𝑡) (𝑣)
𝑑𝑡 𝑑𝑡
where 𝒗(𝑝) = (𝑝, 𝑣). By equation (2.1.6) the term on the right is 𝐿𝒗 𝜙(𝑝). Hence we conclude:
Theorem 2.2.9. Let 𝑈 be an open subset of 𝐑𝑛 and 𝜙 ∈ 𝐶1 (𝑈). Then 𝜙 is an integral of the
system (2.2.3) if and only if 𝐿𝒗 𝜙 = 0.
Definition 2.2.10. Let 𝑈 be an open subset of 𝐑𝑛 and 𝒗 a vector field on 𝑈. We say that 𝒗 is
complete if for every 𝑝 ∈ 𝑈 there exists an integral curve 𝛾 ∶ 𝐑 → 𝑈 with 𝛾(0) = 𝑝, i.e., for
every 𝑝 there exists an integral curve that starts at 𝑝 and exists for all time.
To see what completeness involves, we recall that an integral curve
𝛾 ∶ [0, 𝑏) → 𝑈 ,
with 𝛾(0) = 𝑝 is called maximal if it cannot be extended to an interval [0, 𝑏′ ), where 𝑏′ > 𝑏.
(See for instance [2, §6.11].) For such curves it is known that either
(1) 𝑏 = +∞,
(2) |𝛾(𝑡)| → +∞ as 𝑡 → 𝑏,
(3) or the limit set of
{ 𝛾(𝑡) | 0 ≤ 𝑡 < 𝑏 }
contains points on the boundary of 𝑈.
Hence if we can exclude (2) and (3), we have shown that an integral curve with 𝛾(0) = 𝑝
exists for all positive time. A simple criterion for excluding (2) and (3) is the following.
Lemma 2.2.11. The scenarios (2) and (3) cannot happen if there exists a proper 𝐶1 -function
𝜙 ∶ 𝑈 → 𝐑 with 𝐿𝒗 𝜙 = 0.
 Draft: March 28, 2018

§2.2 Integral Curves for Vector Fields 39

Proof. The identity 𝐿𝒗 𝜙 = 0 implies that 𝜙 is constant on 𝛾(𝑡), but if 𝜙(𝑝) = 𝑐 this implies
that the curve 𝛾(𝑡) lies on the compact subset 𝜙−1 (𝑐) ⊂ 𝑈, hence it cannot “run off to infinity”
as in scenario (2) or “run off the boundary” as in scenario (3). □

Applying a similar argument to the interval (−𝑏, 0] we conclude:

Theorem 2.2.12. Suppose there exists a proper 𝐶1 -function 𝜙 ∶ 𝑈 → 𝐑 with the property
𝐿𝒗 𝜙 = 0. Then the vector field 𝒗 is complete.
Example 2.2.13. Let 𝑈 = 𝐑2 and let 𝑣 be the vector field
𝜕 𝜕
𝑣 = 𝑥3 −𝑦 .
𝜕𝑦 𝜕𝑥
Then 𝜙(𝑥, 𝑦) = 2𝑦2 + 𝑥4 is a proper function with the property above.
Another hypothesis on 𝒗 which excludes (2) and (3) is the following.
Definition 2.2.14. The support of 𝒗 is set

supp(𝒗) = { 𝑞 ∈ 𝑈 | 𝒗(𝑞) ≠ 0 } .
We say that 𝒗 is compactly supported if supp(𝒗) is compact.
We now show that compactly supported vector fields are complete.
Theorem 2.2.15. If a vector field 𝒗 is compactly supported, then 𝒗 is complete.
Proof. Notice first that if 𝒗(𝑝) = 0, the constant curve, 𝛾0 (𝑡) = 𝑝, −∞ < 𝑡 < ∞, satisfies the
equation
𝑑
𝛾 (𝑡) = 0 = 𝒗(𝑝) ,
𝑑𝑡 0
so it is an integral curve of 𝒗. Hence if 𝛾(𝑡), −𝑎 < 𝑡 < 𝑏, is any integral curve of 𝒗 with the
property 𝛾(𝑡0 ) = 𝑝 for some 𝑡0 , it has to coincide with 𝛾0 on the interval (−𝑎, 𝑎), and hence
has to be the constant curve 𝛾(𝑡) = 𝑝 on (−𝑎, 𝑎).
Now suppose the support supp(𝒗) of 𝒗 is compact. Then either 𝛾(𝑡) is in supp(𝒗) for all
𝑡 or is in 𝑈 ∖ supp(𝒗) for some 𝑡0 . But if this happens, and 𝑝 = 𝛾(𝑡0 ) then 𝒗(𝑝) = 0, so 𝛾(𝑡)
has to coincide with the constant curve 𝛾0 (𝑡) = 𝑝, for all 𝑡. In neither case can it go off to ∞
or off to the boundary of 𝑈 as 𝑡 → 𝑏. □

One useful application of this result is the following. Suppose 𝒗 is a vector field on
𝑈 ⊂ 𝐑𝑛 , and one wants to see what its integral curves look like on some compact set 𝐴 ⊂ 𝑈.
Let 𝜌 ∈ 𝐶0∞ (𝑈) be a bump function which is equal to one on a neighborhood of 𝐴. Then
the vector field 𝒘 = 𝜌𝒗 is compactly supported and hence complete, but it is identical with
𝒗 on 𝐴, so its integral curves on 𝐴 coincide with the integral curves of 𝒗.
If 𝒗 is complete then for every 𝑝, one has an integral curve 𝛾𝑝 ∶ 𝐑 → 𝑈 with 𝛾𝑝 (0) = 𝑝,
so one can define a map
𝑓𝑡 ∶ 𝑈 → 𝑈
by setting 𝑓𝑡 (𝑝) ≔ 𝛾𝑝 (𝑡). If 𝒗 is 𝐶 , this mapping is 𝐶∞ by the smooth dependence on initial
∞

data theorem, and, by definition, 𝑓0 = id𝑈 . We claim that the 𝑓𝑡 ’s also have the property
(2.2.16) 𝑓𝑡 ∘ 𝑓𝑎 = 𝑓𝑡+𝑎 .
 Draft: March 28, 2018

40 Chapter 2: Differential Forms

Indeed if 𝑓𝑎 (𝑝) = 𝑞, then by the reparametrization theorem, 𝛾𝑞 (𝑡) and 𝛾𝑝 (𝑡 + 𝑎) are both
integral curves of 𝒗, and since 𝑞 = 𝛾𝑞 (0) = 𝛾𝑝 (𝑎) = 𝑓𝑎 (𝑝), they have the same initial point,
so
𝛾𝑞 (𝑡) = 𝑓𝑡 (𝑞) = (𝑓𝑡 ∘ 𝑓𝑎 )(𝑝)
= 𝛾𝑝 (𝑡 + 𝑎) = 𝑓𝑡+𝑎 (𝑝)
for all 𝑡. Since 𝑓0 is the identity it follows from (2.2.16) that 𝑓𝑡 ∘ 𝑓−𝑡 is the identity, i.e.,
𝑓−𝑡 = 𝑓𝑡−1 ,
so 𝑓𝑡 is a 𝐶∞ diffeomorphism. Hence if 𝒗 is complete it generates a one-parameter group 𝑓𝑡 ,
−∞ < 𝑡 < ∞, of 𝐶∞ -diffeomorphisms of 𝑈.
If 𝒗 is not complete there is an analogous result, but it is trickier to formulate precisely.
Roughly speaking 𝒗 generates a one-parameter group of diffeomorphisms 𝑓𝑡 but these dif-
feomorphisms are not defined on all of 𝑈 nor for all values of 𝑡. Moreover, the identity
(2.2.16) only holds on the open subset of 𝑈 where both sides are well-defined.
We devote the remainder of this section to discussing some “functorial” properties of
vector fields and one-forms.
Definition 2.2.17. Let 𝑈 ⊂ 𝐑𝑛 and 𝑊 ⊂ 𝐑𝑚 be open, and let 𝑓 ∶ 𝑈 → 𝑊 be a 𝐶∞ map. If 𝒗
is a 𝐶∞ -vector field on 𝑈 and 𝒘 a 𝐶∞ -vector field on 𝑊 we say that 𝒗 and 𝒘 are 𝑓-related
if, for all 𝑝 ∈ 𝑈 we have
𝑑𝑓𝑝 (𝒗(𝑝)) = 𝒘(𝑓(𝑝)) .
Writing
𝑛
𝜕
𝒗 = ∑ 𝑣𝑖 , 𝑣𝑖 ∈ 𝐶𝑘 (𝑈)
𝑖=1 𝜕𝑥𝑖
and
𝑚
𝜕
𝒘 = ∑ 𝑤𝑗 , 𝑤𝑗 ∈ 𝐶𝑘 (𝑉)
𝑗=1 𝜕𝑦𝑗
this equation reduces, in coordinates, to the equation
𝑛
𝜕𝑓𝑖
𝑤𝑖 (𝑞) = ∑ (𝑝)𝑣𝑗 (𝑝) .
𝑗=1 𝜕𝑥𝑗

In particular, if 𝑚 = 𝑛 and 𝑓 is a 𝐶∞ diffeomorphism, the formula (3.2) defines a 𝐶∞ -vector

field on 𝑊, i.e.,
𝑛
𝜕
𝒘 = ∑ 𝑤𝑖
𝑗=1 𝜕𝑦𝑗
is the vector field defined by the equation
𝑛
𝜕𝑓𝑖
𝑤𝑖 = ∑ ( 𝑣 ) ∘ 𝑓−1 .
𝑗=1 𝜕𝑥𝑗 𝑗
Hence we have proved:
Theorem 2.2.18. If 𝑓 ∶ 𝑈 ⥲ 𝑊 is a 𝐶∞ diffeomorphism and 𝒗 a 𝐶∞ -vector field on 𝑈, there
exists a unique 𝐶∞ vector field 𝒘 on 𝑊 having the property that 𝒗 and 𝒘 are 𝑓-related.
Definition 2.2.19. In the setting of Theorem 2.2.18, we denote the vector field 𝒘 by 𝑓⋆ 𝒗,
and call 𝑓⋆ 𝒗 the pushforward of 𝒗 by 𝑓 .
 Draft: March 28, 2018

§2.2 Integral Curves for Vector Fields 41

We leave the following assertions as easy exercises (but provide some hints).
Theorem 2.2.20. For 𝑖 = 1, 2, let 𝑈𝑖 be an open subset of 𝐑𝑛𝑖 , 𝒗𝑖 a vector field on 𝑈𝑖 and
𝑓 ∶ 𝑈1 → 𝑈2 a 𝐶∞ -map. If 𝒗1 and 𝒗2 are 𝑓-related, every integral curve
𝛾 ∶ 𝐼 → 𝑈1
of 𝒗1 gets mapped by 𝑓 onto an integral curve 𝑓 ∘ 𝛾 ∶ 𝐼 → 𝑈2 of 𝒗2 .
Proof sketch. Theorem 2.2.20 follows from the chain rule: if 𝑝 = 𝛾(𝑡) and 𝑞 = 𝑓(𝑝)
𝑑 𝑑
𝑑𝑓𝑝 ( 𝛾(𝑡)) = 𝑓(𝛾(𝑡)) . □
𝑑𝑡 𝑑𝑡
Corollary 2.2.21. In the setting of Theorem 2.2.20, suppose 𝒗1 and 𝒗2 are complete. Let
(𝑓𝑖,𝑡 )𝑡∈𝐑 ∶ 𝑈𝑖 ⥲ 𝑈𝑖 be the one-parameter group of diffeomorphisms generated by 𝒗𝑖 . Then
𝑓 ∘ 𝑓1,𝑡 = 𝑓2,𝑡 ∘ 𝑓.
Proof sketch. To deduce Corollary 2.2.21 from Theorem 2.2.20 note that for 𝑝 ∈ 𝑈, 𝑓1,𝑡 (𝑝)
is just the integral curve 𝛾𝑝 (𝑡) of 𝒗1 with initial point 𝛾𝑝 (0) = 𝑝.
The notion of 𝑓-relatedness can be very succinctly expressed in terms of the Lie differ-
entiation operation. For 𝜙 ∈ 𝐶∞ (𝑈2 ) let 𝑓⋆ 𝜙 be the composition, 𝜙 ∘ 𝑓, viewed as a 𝐶∞
function on 𝑈1 , i.e., for 𝑝 ∈ 𝑈1 let 𝑓⋆ 𝜙(𝑝) = 𝜙(𝑓(𝑝)). Then
𝑓⋆ 𝐿𝒗2 𝜙 = 𝐿𝒗1 𝑓⋆ 𝜙 .
To see this, note that if 𝑓(𝑝) = 𝑞 then at the point 𝑝 the right hand side is
(𝑑𝜙)𝑞 ∘ 𝑑𝑓𝑝 (𝒗1 (𝑝))
by the chain rule and by definition the left hand side is
𝑑𝜙𝑞 (𝒗2 (𝑞)) .
Moreover, by definition
𝒗2 (𝑞) = 𝑑𝑓𝑝 (𝒗1 (𝑝))
so the two sides are the same. □
Another easy consequence of the chain rule is:
Theorem 2.2.22. For 𝑖 = 1, 2, 3, let 𝑈𝑖 be an open subset of 𝐑𝑛𝑖 , 𝒗𝑖 a vector field on 𝑈𝑖 , and
for 𝑖 = 1, 2 let 𝑓𝑖 ∶ 𝑈𝑖 → 𝑈𝑖+1 be a 𝐶∞ -map. Suppose that 𝒗1 and 𝒗2 are 𝑓1 -related and that 𝒗2
and 𝒗3 are 𝑓2 -related. Then 𝒗1 and 𝒗3 are (𝑓2 ∘ 𝑓1 )-related.
In particular, if 𝑓1 and 𝑓2 are diffeomorphisms writing 𝒗 ≔ 𝒗1 we have
(𝑓2 )⋆ (𝑓1 )⋆ 𝒗 = (𝑓2 ∘ 𝑓1 )⋆ 𝒗 .
The results we described above have “dual” analogues for one-forms.
Construction 2.2.23. Let 𝑈 ⊂ 𝐑𝑛 and 𝑉 ⊂ 𝐑𝑚 be open, and let 𝑓 ∶ 𝑈 → 𝑉 be a 𝐶∞ -map.
Given a one-form 𝜇 on 𝑉 one can define a pullback one-form 𝑓⋆ 𝜇 on 𝑈 by the following
method. For 𝑝 ∈ 𝑈, by definition 𝜇𝑓(𝑝) is a linear map
𝜇𝑓(𝑝) ∶ 𝑇𝑓(𝑝) 𝐑𝑚 → 𝐑
and by composing this map with the linear map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝐑𝑛 → 𝑇𝑓(𝑝) 𝐑𝑛
we get a linear map
𝜇𝑓(𝑝) ∘ 𝑑𝑓𝑝 ∶ 𝑇𝑝 𝐑𝑛 → 𝐑 ,
 Draft: March 28, 2018

42 Chapter 2: Differential Forms

i.e., an element 𝜇𝑓(𝑝) ∘ 𝑑𝑓𝑝 of 𝑇𝑝⋆ 𝐑𝑛 .

The pullback one-form 𝑓⋆ 𝜇 is defined by the assignment 𝑝 ↦ (𝜇𝑓(𝑝) ∘ 𝑑𝑓𝑝 ).
In particular, if 𝜙 ∶ 𝑉 → 𝐑 is a 𝐶∞ -function and 𝜇 = 𝑑𝜙 then
𝜇𝑞 ∘ 𝑑𝑓𝑝 = 𝑑𝜙𝑞 ∘ 𝑑𝑓𝑝 = 𝑑(𝜙 ∘ 𝑓)𝑝
i.e.,
𝑓⋆ 𝜇 = 𝑑𝜙 ∘ 𝑓 .
In particular if 𝜇 is a one-form of the form 𝜇 = 𝑑𝜙, with 𝜙 ∈ 𝐶∞ (𝑉), 𝑓⋆ 𝜇 is 𝐶∞ . From this
it is easy to deduce:
Theorem 2.2.24. If 𝜇 is any 𝐶∞ one-form on 𝑉, its pullback 𝑓⋆ 𝜔 is 𝐶∞ . (See Exercise 2.2.ii.)
Notice also that the pullback operation on one-forms and the pushforward operation
on vector fields are somewhat different in character. The former is defined for all 𝐶∞ maps,
but the latter is only defined for diffeomorphisms.

Exercises for §2.2

Exercise 2.2.i. Prove the reparameterization result Theorem 2.2.20.
Exercise 2.2.ii. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝑉 a 𝐶𝑘
map.
(1) Show that for 𝜙 ∈ 𝐶∞ (𝑉) (2.2) can be rewritten
𝑓⋆ 𝑑𝜙 = 𝑑𝑓⋆ 𝜙 .
(2) Let 𝜇 be the one-form
𝑚
𝜇 = ∑ 𝜙𝑖 𝑑𝑥𝑖 , 𝜙𝑖 ∈ 𝐶∞ (𝑉)
𝑖=1
on 𝑉. Show that if 𝑓 = (𝑓1 , …, 𝑓𝑚 ) then
𝑚
𝑓⋆ 𝜇 = ∑ 𝑓⋆ 𝜙𝑖 𝑑𝑓𝑖 .
𝑖=1
∞
(3) Show that if 𝜇 is 𝐶 and 𝑓 is 𝐶 , 𝑓 𝜇 is 𝐶∞ .
∞ ⋆

Exercise 2.2.iii. Let 𝒗 be a complete vector field on 𝑈 and 𝑓𝑡 ∶ 𝑈 → 𝑈, the one parameter
group of diffeomorphisms generated by 𝒗. Show that if 𝜙 ∈ 𝐶1 (𝑈)
𝑑 ⋆
𝐿𝒗 𝜙 = 𝑓 𝜙| .
𝑑𝑡 𝑡 𝑡=0
Exercise 2.2.iv.
(1) Let 𝑈 = 𝐑2 and let 𝒗 be the vector field, 𝑥1 𝜕/𝜕𝑥2 − 𝑥2 𝜕/𝜕𝑥1 . Show that the curve
𝑡 ↦ (𝑟 cos(𝑡 + 𝜃), 𝑟 sin(𝑡 + 𝜃)) ,
for 𝑡 ∈ 𝐑, is the unique integral curve of 𝒗 passing through the point, (𝑟 cos 𝜃, 𝑟 sin 𝜃),
at 𝑡 = 0.
𝑛
(2) Let 𝑈 = 𝐑𝑛 and let 𝒗 be the constant vector field: ∑𝑖=1 𝑐𝑖 𝜕/𝜕𝑥𝑖 . Show that the curve
𝑡 ↦ 𝑎 + 𝑡(𝑐1 , …, 𝑐𝑛 ) ,
for 𝑡 ∈ 𝐑, is the unique integral curve of 𝒗 passing through 𝑎 ∈ 𝐑𝑛 at 𝑡 = 0.
 Draft: March 28, 2018

§2.2 Integral Curves for Vector Fields 43

𝑛
(3) Let 𝑈 = 𝐑𝑛 and let 𝒗 be the vector field, ∑𝑖=1 𝑥𝑖 𝜕/𝜕𝑥𝑖 . Show that the curve
𝑡 ↦ 𝑒𝑡 (𝑎1 , …, 𝑎𝑛 ) ,
for 𝑡 ∈ 𝐑, is the unique integral curve of 𝒗 passing through 𝑎 at 𝑡 = 0.
Exercise 2.2.v. Show that the following are one-parameter groups of diffeomorphisms:
(1) 𝑓𝑡 ∶ 𝐑 → 𝐑 , 𝑓𝑡 (𝑥) = 𝑥 + 𝑡
(2) 𝑓𝑡 ∶ 𝐑 → 𝐑 , 𝑓𝑡 (𝑥) = 𝑒𝑡 𝑥
(3) 𝑓𝑡 ∶ 𝐑2 → 𝐑2 , 𝑓𝑡 (𝑥, 𝑦) = (𝑥 cos(𝑡) − 𝑦 sin(𝑡), 𝑥 sin(𝑡) + 𝑦 cos(𝑡)).
Exercise 2.2.vi. Let 𝐴 ∶ 𝐑𝑛 → 𝐑𝑛 be a linear mapping. Show that the series
∞
(𝑡𝐴)𝑛 𝑡2 𝑡3
exp(𝑡𝐴) ≔ ∑ = id𝑛 +𝑡𝐴 + 𝐴2 + 𝐴3 + ⋯
𝑛=0 𝑛! 2! 3!
converges and defines a one-parameter group of diffeomorphisms of 𝐑𝑛 .
Exercise 2.2.vii.
(1) What are the infinitesimal generators of the one-parameter groups in Exercise 2.2.v?
(2) Show that the infinitesimal generator of the one-parameter group in Exercise 2.2.vi is
the vector field
𝜕
∑ 𝑎𝑖,𝑗 𝑥𝑗
1≤𝑖,𝑗≤𝑛 𝜕𝑥 𝑖

where (𝑎𝑖,𝑗 ) is the defining matrix of 𝐴.

𝑑
Exercise 2.2.viii. Let 𝒗 be the vector field on 𝐑 given by 𝑥2 𝑑𝑥 . Show that the curve
𝑎
𝑥(𝑡) =
𝑎 − 𝑎𝑡
is an integral curve of 𝒗 with initial point 𝑥(0) = 𝑎. Conclude that for 𝑎 > 0 the curve
𝑎 1
𝑥(𝑡) = , 0<𝑡<
1 − 𝑎𝑡 𝑎
is a maximal integral curve. (In particular, conclude that 𝒗 is not complete.)
Exercise 2.2.ix. Let 𝑈 and 𝑉 be open subsets of 𝐑𝑛 and 𝑓 ∶ 𝑈 ⥲ 𝑉 a diffeomorphism. If 𝒘
is a vector field on 𝑉, define the pullback of 𝒘 to 𝑈 to be the vector field
𝑓⋆ 𝒘 ≔ (𝑓⋆−1 𝒘) .
Show that if 𝜙 is a 𝐶∞ function on 𝑉
𝑓⋆ 𝐿 𝒘 𝜙 = 𝐿 𝑓 ⋆ 𝒘 𝑓⋆ 𝜙 .
Hint: Equation (2.2.25).
Exercise 2.2.x. Let 𝑈 be an open subset of 𝐑𝑛 and 𝒗 and 𝒘 vector fields on 𝑈. Suppose 𝒗 is
the infinitesimal generator of a one-parameter group of diffeomorphisms
𝑓𝑡 ∶ 𝑈 ⥲ 𝑈 , − ∞ < 𝑡 < ∞ .
Let 𝒘𝑡 = 𝑓𝑡⋆ 𝒘. Show that for 𝜙 ∈ 𝐶∞ (𝑈) we have
𝐿[𝒗,𝒘] 𝜙 = 𝐿𝒘̇ 𝜙 ,
where
𝑑 ⋆
𝒘̇ = 𝑓 𝒘| .
𝑑𝑡 𝑡 𝑡=0
 Draft: March 28, 2018

44 Chapter 2: Differential Forms

Hint: Differentiate the identity

𝑓𝑡⋆ 𝐿𝒘 𝜙 = 𝐿𝒘𝑡 𝑓𝑡⋆ 𝜙
with respect to 𝑡 and show that at 𝑡 = 0 the derivative of the left hand side is 𝐿𝒗 𝐿𝒘 𝜙 by
Exercise 2.2.iii, and the derivative of the right hand side is
𝐿𝒘̇ + 𝐿𝒘 (𝐿𝒗 𝜙) .
Exercise 2.2.xi. Conclude from Exercise 2.2.x that
𝑑 ⋆
(2.2.25) [𝒗, 𝒘] = 𝑓 𝒘| .
𝑑𝑡 𝑡 𝑡=0

2.3. Differential 𝑘-forms

One-forms are the bottom tier in a pyramid of objects whose 𝑘th tier is the space of 𝑘-
forms. More explicitly, given 𝑝 ∈ 𝐑𝑛 we can, as in Section 1.5, form the 𝑘th exterior powers
𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ) , 𝑘 = 1, 2, 3, …, 𝑛
of the vector space 𝑇𝑝⋆ 𝐑𝑛 , and since
(2.3.1) 𝛬1 (𝑇𝑝⋆ 𝐑𝑛 ) = 𝑇𝑝⋆ 𝐑𝑛
one can think of a one-form as a function which takes its value at 𝑝 in the space (2.3.1). This
leads to an obvious generalization.
Definition 2.3.2. Let 𝑈 be an open subset of 𝐑𝑛 . A 𝑘-form 𝜔 on 𝑈 is a function which
assigns to each point 𝑝 ∈ 𝑈 an element 𝜔𝑝 ∈ 𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ).
The wedge product operation gives us a way to construct lots of examples of such ob-
jects.
Example 2.3.3. Let 𝜔1 , …, 𝜔𝑘 be one-forms. Then 𝜔1 ∧ ⋯ ∧ 𝜔𝑘 is the 𝑘-form whose value
at 𝑝 is the wedge product
(2.3.4) (𝜔1 ∧ ⋯ ∧ 𝜔𝑘 )𝑝 ≔ (𝜔1 )𝑝 ∧ ⋯ ∧ (𝜔𝑘 )𝑝 .
Notice that since (𝜔𝑖 )𝑝 is in 𝛬1 (𝑇𝑝⋆ 𝐑𝑛 ) the wedge product (2.3.4) makes sense and is an
element of 𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ).
Example 2.3.5. Let 𝑓1 , …, 𝑓𝑘 be a real-valued 𝐶∞ functions on 𝑈. Letting 𝜔𝑖 = 𝑑𝑓𝑖 we get
from (2.3.4) a 𝑘-form
𝑑𝑓1 ∧ ⋯ ∧ 𝑑𝑓𝑘
whose value at 𝑝 is the wedge product
(2.3.6) (𝑑𝑓1 )𝑝 ∧ ⋯ ∧ (𝑑𝑓𝑘 )𝑝 .
Since (𝑑𝑥1 )𝑝 , …, (𝑑𝑥𝑛 )𝑝 are a basis of 𝑇𝑝⋆ 𝐑𝑛 , the wedge products
(2.3.7) (𝑑𝑥𝑖1 )𝑝 ∧ ⋯ ∧ (𝑑𝑥1𝑘 )𝑝 , 1 ≤ 𝑖1 < ⋯ < 𝑖𝑘 ≤ 𝑛
are a basis of 𝛬𝑘 (𝑇𝑝⋆ ). To keep our multi-index notation from getting out of hand, we denote
these basis vectors by (𝑑𝑥𝐼 )𝑝 , where 𝐼 = (𝑖1 , …, 𝑖𝑘 ) and the multi-indices 𝐼 range over multi-
indices of length 𝑘 which are strictly increasing. Since these wedge products are a basis of
𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ), every element of 𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ) can be written uniquely as a sum
∑ 𝑐𝐼 (𝑑𝑥𝐼 )𝑝 , 𝑐𝐼 ∈ 𝐑
𝐼
 Draft: March 28, 2018

§2.3 Differential 𝑘-forms 45

and every 𝑘-form 𝜔 on 𝑈 can be written uniquely as a sum

(2.3.8) 𝜔 = ∑ 𝑓𝐼 𝑑𝑥𝐼
𝐼
where 𝑑𝑥𝐼 is the 𝑘-form 𝑑𝑥𝑖1 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑘 , and 𝑓𝐼 is a real-valued function 𝑓𝐼 ∶ 𝑈 → 𝐑.
Definition 2.3.9. The 𝑘-form (2.3.8) is of class 𝐶𝑟 if each function 𝑓𝐼 is in 𝐶𝑟 (𝑈).
Henceforth we assume, unless otherwise stated, that all the 𝑘-forms we consider are of
class 𝐶∞ . We denote the set of 𝑘-forms of class 𝐶∞ on 𝑈 by 𝛺𝑘 (𝑈).
We will conclude this section by discussing a few simple operations on 𝑘-forms. Let 𝑈
be an open subset of 𝐑𝑛 .
(1) Given a function 𝑓 ∈ 𝐶∞ (𝑈) and a 𝑘-form 𝜔 ∈ 𝛺𝑘 (𝑈) we define 𝑓𝜔 ∈ 𝛺𝑘 (𝑈) to be the
𝑘-form defined by
𝑝 ↦ 𝑓(𝑝)𝜔𝑝 .
(2) Given 𝜔1 , 𝜔2 ∈ 𝛺𝑘 (𝑈), we define 𝜔1 + 𝜔2 ∈ 𝛺𝑘 (𝑈) to be the 𝑘-form
𝑝 ↦ (𝜔1 )𝑝 + (𝜔2 )𝑝 .
(Notice that this sum makes sense since each summand is in 𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ).)
(3) Given 𝜔1 ∈ 𝛺𝑘1 (𝑈) and 𝜔2 ∈ 𝛺𝑘2 (𝑈) we define their wedge product, 𝜔1 ∧𝜔2 ∈ 𝛺𝑘1 +𝑘2 (𝑈)
to be the (𝑘1 + 𝑘2 )-form
𝑝 ↦ (𝜔1 )𝑝 ∧ (𝜔2 )𝑝 .
We recall that 𝛬0 (𝑇𝑝⋆ 𝐑𝑛 ) = 𝐑, so a zero-form is an 𝐑-valued function and a zero form
of class 𝐶∞ is a 𝐶∞ function, i.e., 𝛺0 (𝑈) = 𝐶∞ (𝑈).
The addition and multiplication by functions operations naturally give the sets 𝛺𝑘 (𝑈)
the structures of 𝐑-vector spaces — we always regard 𝛺𝑘 (𝑈) as a vector space in this manner.
A fundamental operation on forms is the exterior differention operation which asso-
ciates to a function 𝑓 ∈ 𝐶∞ (𝑈) the 1-form 𝑑𝑓. It is clear from the identity (2.2.3) that 𝑑𝑓
is a 1-form of class 𝐶∞ , so the 𝑑-operation can be viewed as a map
𝑑 ∶ 𝛺0 (𝑈) → 𝛺1 (𝑈) .
In the next section we show that an analogue of this map exists for every 𝛺𝑘 (𝑈).
Exercises for §2.3
2 4
Exercise 2.3.i. Let 𝜔 ∈ 𝛺 (𝐑 ) be the 2-form 𝑑𝑥1 ∧ 𝑑𝑥2 + 𝑑𝑥3 ∧ 𝑑𝑥4 . Compute 𝜔 ∧ 𝜔.
Exercise 2.3.ii. Let 𝜔1 , 𝜔2 , 𝜔3 ∈ 𝛺1 (𝐑3 ) be the 1-forms
𝜔1 = 𝑥2 𝑑𝑥3 − 𝑥3 𝑑𝑥2 ,
𝜔2 = 𝑥3 𝑑𝑥1 − 𝑥1 𝑑𝑥3 ,
𝜔3 = 𝑥1 𝑑𝑥2 − 𝑥2 𝑑𝑥1 .
Compute the following.
(1) 𝜔1 ∧ 𝜔2 .
(2) 𝜔2 ∧ 𝜔3 .
(3) 𝜔3 ∧ 𝜔1 .
(4) 𝜔1 ∧ 𝜔2 ∧ 𝜔3 .
Exercise 2.3.iii. Let 𝑈 be an open subset of 𝐑𝑛 and 𝑓1 , …, 𝑓𝑛 ∈ 𝐶∞ (𝑈). Show that
𝜕𝑓𝑖
𝑑𝑓1 ∧ ⋯ ∧ 𝑑𝑓𝑛 = det [ ] 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
𝜕𝑥𝑗
 Draft: March 28, 2018

46 Chapter 2: Differential Forms

Exercise 2.3.iv. Let 𝑈 be an open subset of 𝐑𝑛 . Show that every (𝑛 − 1)-form 𝜔 ∈ 𝛺𝑛−1 (𝑈)
can be written uniquely as a sum
𝑛
̂ 𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛 ,
∑ 𝑓𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑖=1

̂ 𝑖 indicates that 𝑑𝑥𝑖 is to be omitted from the wedge product

where 𝑓𝑖 ∈ 𝐶∞ (𝑈) and 𝑑𝑥
𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
𝑛
Exercise 2.3.v. Let 𝜇 = ∑𝑖=1 𝑥𝑖 𝑑𝑥𝑖 . Show that there exists an (𝑛−1)-form, 𝜔 ∈ 𝛺𝑛−1 (𝐑𝑛 ∖{0})
with the property
𝜇 ∧ 𝜔 = 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
Exercise 2.3.vi. Let 𝐽 be the multi-index (𝑗1 , …, 𝑗𝑘 ) and let 𝑑𝑥𝐽 = 𝑑𝑥𝑗1 ∧ ⋯ ∧ 𝑑𝑥𝑗𝑘 . Show
that 𝑑𝑥𝐽 = 0 if 𝑗𝑟 = 𝑗𝑠 for some 𝑟 ≠ 𝑠 and show that if the numbers 𝑗1 , …, 𝑗𝑘 are all distinct,
then
𝑑𝑥𝐽 = (−1)𝜎 𝑑𝑥𝐼 ,
where 𝐼 = (𝑖1 , …, 𝑖𝑘 ) is the strictly increasing rearrangement of (𝑗1 , …, 𝑗𝑘 ) and 𝜎 is the per-
mutation
(𝑗1 , …, 𝑗𝑘 ) ↦ (𝑖1 , …, 𝑖𝑘 ) .
Exercise 2.3.vii. Let 𝐼 be a strictly increasing multi-index of length 𝑘 and 𝐽 a strictly increas-
ing multi-index of length ℓ. What can one say about the wedge product 𝑑𝑥𝐼 ∧ 𝑑𝑥𝐽 ?

2.4. Exterior differentiation

Let 𝑈 be an open subset of 𝐑𝑛 . In this section we define an exterior differentiation oper-
ation
(2.4.1) 𝑑 ∶ 𝛺𝑘 (𝑈) → 𝛺𝑘+1 (𝑈) .
This operation is the fundamental operation in 𝑛-dimensional vector calculus.
For 𝑘 = 0 we already defined the operation (2.4.1) in §2.1. Before defining it for 𝑘 > 0,
we list some properties that we require to this operation to satisfy.
Properties 2.4.2 (desired properties of exterior differentiation). Let 𝑈 be an open subset of
𝐑𝑛 .
(1) For 𝜔1 and 𝜔2 in 𝛺𝑘 (𝑈), we have
𝑑(𝜔1 + 𝜔2 ) = 𝑑𝜔1 + 𝑑𝜔2 .
(2) For 𝜔1 ∈ 𝛺𝑘 (𝑈) and 𝜔2 ∈ 𝛺ℓ (𝑈) we have
(2.4.3) 𝑑(𝜔1 ∧ 𝜔2 ) = 𝑑𝜔1 ∧ 𝜔2 + (−1)𝑘 𝜔1 ∧ 𝑑𝜔2 .
(3) For all 𝜔 ∈ 𝛺𝑘 (𝑈) we have
(2.4.4) 𝑑(𝑑𝜔) = 0
Let is point out a few consequences of these properties.
Lemma 2.4.5. Let 𝑈 be an open subset of 𝐑𝑛 . For any functions 𝑓1 , …, 𝑓𝑘 ∈ 𝐶∞ (𝑈) we have
𝑑(𝑑𝑓1 ∧ ⋯ ∧ 𝑑𝑓𝑘 ) = 0 .
 Draft: March 28, 2018

§2.4 Exterior differentiation 47

Proof. We prove this by induction on 𝑘. For the base case note that by (3)
(2.4.6) 𝑑(𝑑𝑓1 ) = 0
∞
for every function 𝑓1 ∈ 𝐶 (𝑈).
For the induction step, suppose that we know the result for 𝑘 − 1 functions and that
we are given functions 𝑓1 , …, 𝑓𝑘 ∈ 𝐶∞ (𝑈). Let 𝜇 = 𝑑𝑓2 ∧ ⋯ ∧ 𝑑𝑓𝑘 . Then by the induction
hypothesis 𝑑𝜇 = 0 Combining (2.4.3) with (2.4.6) we see that
𝑑(𝑑𝑓1 ∧ 𝑑𝑓2 ∧ ⋯ ∧ 𝑑𝑓𝑘 ) = 𝑑(𝑑𝑓1 ∧ 𝜇)
= 𝑑(𝑑𝑓1 ) ∧ 𝜇 + (−1) 𝑑𝑓1 ∧ 𝑑𝜇
=0. □
Example 2.4.7. As a special case of Lemma 2.4.5, given a multi-index 𝐼 = (𝑖1 , …, 𝑖𝑘 ) with
1 ≤ 𝑖𝑟 ≤ 𝑛 we have
(2.4.8) 𝑑(𝑑𝑥𝐼 ) = 𝑑(𝑑𝑥𝑖1 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑘 ) = 0 .
Recall that every 𝑘-form 𝜔 ∈ 𝛺𝑘 (𝑈) can be written uniquely as a sum
𝜔 = ∑ 𝑓𝐼 𝑑𝑥𝐼 , 𝑓𝐼 ∈ 𝐶∞ (𝑈)
𝐼
where the multi-indices 𝐼 are strictly increasing. Thus by (2.4.3) and (2.4.8)
(2.4.9) 𝑑𝜔 = ∑ 𝑑𝑓𝐼 ∧ 𝑑𝑥𝐼 .
𝐼

This shows that if there exists an operator 𝑑 ∶ 𝛺∗ (𝑈) → 𝛺∗+1 (𝑈) with Properties 2.4.2, then
𝑑 is necessarily given by the formula (2.4.9). Hence all we have to show is that the operator
defined by this equation (2.4.9) has these properties.
Proposition 2.4.10. Let 𝑈 be an open subset of 𝐑𝑛 . There is a unique operator 𝑑 ∶ 𝛺∗ (𝑈) →
𝛺∗+1 (𝑈) satisfying Properties 2.4.2.
Proof. The property (1) is obvious.
To verify (2) we first note that for 𝐼 strictly increasing equation (2.4.8) is a special case
of equation (2.4.9) (take 𝑓𝐼 = 1 and 𝑓𝐽 = 0 for 𝐽 ≠ 𝐼). Moreover, if 𝐼 is not strictly increasing
it is either repeating, in which case 𝑑𝑥𝐼 = 0, or non-repeating in which case there exists a
permuation 𝜎 ∈ 𝑆𝑘 such that 𝐼𝜎 is strictly increasing. Moreover,
(2.4.11) 𝑑𝑥𝐼 = (−1)𝜎 𝑑𝑥𝐼𝜎 .
Hence (2.4.9) implies (2.4.8) for all multi-indices 𝐼. The same argument shows that for any
sum ∑𝐼 𝑓𝐼 𝑑𝑥𝐼 over multi-indices 𝐼 of length 𝑘, we have
(2.4.12) 𝑑(∑𝐼 𝑓𝐼 𝑑𝑥𝐼 ) = ∑ 𝑑𝑓𝐼 ∧ 𝑑𝑥𝐼 .
𝐼
(As above we can ignore the repeating multi-indices 𝑑𝑥𝐼 = 0 if 𝐼 is repeating, and by (2.4.11)
we can replace the non-repeating multi-indices by strictly increasing multi-indices.)
Suppose now that 𝜔1 ∈ 𝛺𝑘 (𝑈) and 𝜔2 ∈ 𝛺ℓ (𝑈). Write 𝜔1 ≔ ∑𝐼 𝑓𝐼 𝑑𝑥𝐼 and 𝜔2 ≔
∑𝐽 𝑔𝐽 𝑑𝑥𝐽 , where 𝑓𝐼 , 𝑔𝐽 ∈ 𝐶∞ (𝑈). We see that the wedge product 𝜔1 ∧ 𝜔2 is given by
(2.4.13) 𝜔1 ∧ 𝜔2 = ∑ 𝑓𝐼 𝑔𝐽 𝑑𝑥𝐼 ∧ 𝑑𝑥𝐽 ,
𝐼,𝐽

and by equation (2.4.12)

(2.4.14) 𝑑(𝜔1 ∧ 𝜔2 ) = ∑ 𝑑(𝑓𝐼 𝑔𝐽 ) ∧ 𝑑𝑥𝐼 ∧ 𝑑𝑥𝐽 .
𝐼,𝐽
 Draft: March 28, 2018

48 Chapter 2: Differential Forms

(Notice that if 𝐼 = (𝑖1 , …, 𝑖𝑘 ) and 𝐽 = (𝑗𝑖 , …, 𝑖ℓ ), we have 𝑑𝑥𝐼 ∧ 𝑑𝑥𝐽 = 𝑑𝑥𝐾 , where 𝐾 ≔
(𝑖1 , …, 𝑖𝑘 , 𝑗1 , …, 𝑗ℓ ). Even if 𝐼 and 𝐽 are strictly increasing, 𝐾 is not necessarily strictly in-
creasing. However, in deducing (2.4.14) from (2.4.13) we have observed that this does not
matter.)
Now note that by equation (2.2.8)
𝑑(𝑓𝐼 𝑔𝐽 ) = 𝑔𝐽 𝑑𝑓𝐼 + 𝑓𝐼 𝑑𝑔𝐽 ,
and by the wedge product identities of §1.6,

𝑑𝑔𝐽 ∧ 𝑑𝑥𝐼 = 𝑑𝑔𝐽 ∧ 𝑑𝑥𝑖1 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑘 = (−1)𝑘 𝑑𝑥𝐼 ∧ 𝑑𝑔𝐽 ,

so the sum (2.4.14) can be rewritten:
∑ 𝑑𝑓𝐼 ∧ 𝑑𝑥𝐼 ∧ 𝑔𝐽 𝑑𝑥𝐽 + (−1)𝑘 ∑ 𝑓𝐼 𝑑𝑥𝐼 ∧ 𝑑𝑔𝐽 ∧ 𝑑𝑥𝐽 ,
𝐼,𝐽 𝐼,𝐽

or

(∑ 𝑑𝑓𝐼 ∧ 𝑑𝑥𝐼 ) ∧ (∑ 𝑔𝐽 𝑑𝑥𝐽 ) + (−1)𝑘 (∑ 𝑑𝑔𝐽 ∧ 𝑑𝑥𝐽 ) ∧ (∑ 𝑑𝑓𝐼 ∧ 𝑑𝑥𝐼 ) ,

𝐼 𝐽 𝐽 𝐼

or, finally,
𝑑𝜔1 ∧ 𝜔2 + (−1)𝑘 𝜔1 ∧ 𝑑𝜔2 .
Thus the the operator 𝑑 defined by equation (2.4.9) has (2).
Let is now check that 𝑑 satisfies (3). If 𝜔 = ∑𝐼 𝑓𝐼 𝑑𝑥𝐼 , where 𝑓𝐼 ∈ 𝐶∞ (𝑈), then by
definition, 𝑑𝜔 = ∑𝐼 𝑑𝑓𝐼 ∧ 𝑑𝑥𝐼 and by (2.4.8) and (2.4.3)

𝑑(𝑑𝜔) = ∑ 𝑑(𝑑𝑓𝐼 ) ∧ 𝑑𝑥𝐼 ,

𝐼

so it suffices to check that 𝑑(𝑑𝑓𝐼 ) = 0, i.e., it suffices to check (2.4.6) for zero forms 𝑓 ∈
𝐶∞ (𝑈). However, by (2.2.3) we have
𝑛
𝜕𝑓
𝑑𝑓 = ∑ 𝑑𝑥
𝑗=1 𝜕𝑥𝑗 𝑗

so by equation (2.4.9)
𝑛 𝑛 𝑛
𝜕𝑓 𝜕2 𝑓
𝑑(𝑑𝑓) = ∑ 𝑑 ( ) 𝑑𝑥𝑗 = ∑ (∑ 𝑑𝑥𝑖 ) ∧ 𝑑𝑥𝑗
𝑗=1 𝜕𝑥𝑗 𝑗=1 𝑖=1 𝜕𝑥𝑖 𝜕𝑥𝑗

𝜕2 𝑓
= ∑ 𝑑𝑥𝑖 ∧ 𝑑𝑥𝑗 .
1≤𝑖,𝑗≤𝑛 𝜕𝑥𝑖 𝜕𝑥𝑗

Notice, however, that in this sum, 𝑑𝑥𝑖 ∧ 𝑑𝑥𝑗 = −𝑑𝑥𝑗 ∧ 𝑑𝑥𝑖 and

𝜕2 𝑓 𝜕2 𝑓
=
𝜕𝑥𝑖 𝜕𝑥𝑗 𝜕𝑥𝑗 𝜕𝑥𝑖
so the (𝑖, 𝑗) term cancels the (𝑗, 𝑖) term, and the total sum is zero. □

Definition 2.4.15. Let 𝑈 be an open subset of 𝐑𝑛 . A 𝑘-form 𝜔 ∈ 𝛺𝑘 (𝑈) is closed if 𝑑𝜔 = 0

and is exact if 𝜔 = 𝑑𝜇 for some 𝜇 ∈ 𝛺𝑘−1 (𝑈).
 Draft: March 28, 2018

§2.4 Exterior differentiation 49

By (3) every exact form is closed, but the converse is not true even for 1-forms (see
Exercise 2.1.iii). In fact it is a very interesting (and hard) question to determine if an open
set 𝑈 has the following property: For 𝑘 > 0 every closed 𝑘-form is exact.2
Some examples of spaces with this property are described in the exercises at the end of
§2.6. We also sketch below a proof of the following result (and ask you to fill in the details).
Lemma 2.4.16 (Poincaré lemma). If 𝜔 is a closed form on 𝑈 of degree 𝑘 > 0, then for every
point 𝑝 ∈ 𝑈, there exists a neighborhood of 𝑝 on which 𝜔 is exact.
Proof. See Exercises 2.4.v and 2.4.vi. □
Exercises for §2.4
Exercise 2.4.i. Compute the exterior derivatives of the following differential forms.
(1) 𝑥1 𝑑𝑥2 ∧ 𝑑𝑥3
(2) 𝑥1 𝑑𝑥2 − 𝑥2 𝑑𝑥1
𝑛
(3) 𝑒−𝑓 𝑑𝑓 where 𝑓 = ∑𝑖=1 𝑥𝑖2
𝑛
(4) ∑𝑖=1 𝑥𝑖 𝑑𝑥𝑖
𝑛 ̂𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛
(5) ∑𝑖=1 (−1)𝑖 𝑥𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
Exercise 2.4.ii. Solve the equation 𝑑𝜇 = 𝜔 for 𝜇 ∈ 𝛺1 (𝐑3 ), where 𝜔 is the 2-form:
(1) 𝑑𝑥2 ∧ 𝑑𝑥3
(2) 𝑥2 𝑑𝑥2 ∧ 𝑑𝑥3
(3) (𝑥12 + 𝑥22 )𝑑𝑥1 ∧ 𝑑𝑥2
(4) cos(𝑥1 )𝑑𝑥1 ∧ 𝑑𝑥3
Exercise 2.4.iii. Let 𝑈 be an open subset of 𝐑𝑛 .
(1) Show that if 𝜇 ∈ 𝛺𝑘 (𝑈) is exact and 𝜔 ∈ 𝛺ℓ (𝑈) is closed then 𝜇 ∧ 𝜔 is exact.
Hint: Equation (2.4.3).
(2) In particular, 𝑑𝑥1 is exact, so if 𝜔 ∈ 𝛺ℓ (𝑈) is closed 𝑑𝑥1 ∧ 𝜔 = 𝑑𝜇. What is 𝜇?
Exercise 2.4.iv. Let 𝑄 be the rectangle (𝑎1 , 𝑏1 ) × ⋯ × (𝑎𝑛 , 𝑏𝑛 ). Show that if 𝜔 is in 𝛺𝑛 (𝑄),
then 𝜔 is exact.
Hint: Let 𝜔 = 𝑓𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 with 𝑓 ∈ 𝐶∞ (𝑄) and let 𝑔 be the function
𝑥1
𝑔(𝑥1 , …, 𝑥𝑛 ) = ∫ 𝑓(𝑡, 𝑥2 , …, 𝑥𝑛 )𝑑𝑡 .
𝑎1
Show that 𝜔 = 𝑑(𝑔𝑑𝑥2 ∧ ⋯ ∧ 𝑑𝑥𝑛 ).
Exercise 2.4.v. Let 𝑈 be an open subset of 𝐑𝑛−1 , 𝐴 ⊂ 𝐑 an open interval and (𝑥, 𝑡) product
coordinates on 𝑈 × 𝐴. We say that a form 𝜇 ∈ 𝛺ℓ (𝑈 × 𝐴) is reduced if 𝜇 can be written as a
sum
(2.4.17) 𝜇 = ∑ 𝑓𝐼 (𝑥, 𝑡)𝑑𝑥𝐼 ,
𝐼
(i.e., with no terms involving 𝑑𝑡).
(1) Show that every form, 𝜔 ∈ 𝛺𝑘 (𝑈 × 𝐴) can be written uniquely as a sum:
(2.4.18) 𝜔 = 𝑑𝑡 ∧ 𝛼 + 𝛽
where 𝛼 and 𝛽 are reduced.
2For 𝑘 = 0, 𝑑𝑓 = 0 does not imply that 𝑓 is exact. In fact “exactness” does not make much sense for zero
forms since there are not any “(−1)-forms”. However, if 𝑓 ∈ 𝐶∞ (𝑈) and 𝑑𝑓 = 0 then 𝑓 is constant on connected
components of 𝑈 (see Exercise 2.2.iii).
 Draft: March 28, 2018

50 Chapter 2: Differential Forms

(2) Let 𝜇 be the reduced form (2.4.17) and let

𝑑𝜇 𝑑
= ∑ 𝑓𝐼 (𝑥, 𝑡)𝑑𝑥𝐼
𝑑𝑡 𝑑𝑡
and
𝑛
𝜕
𝑑 𝑈 𝜇 = ∑ (∑ 𝑓𝐼 (𝑥, 𝑡)𝑑𝑥𝑖 ) ∧ 𝑑𝑥𝐼 .
𝐼 𝑖=1 𝜕𝑥𝑖
Show that
𝑑𝜇
𝑑𝜇 = 𝑑𝑡 ∧ + 𝑑𝑈 𝜇 .
𝑑𝑡
(3) Let 𝜔 be the form (2.4.18). Show that
𝑑𝛽
𝑑𝜔 = −𝑑𝑡 ∧ 𝑑𝑈 𝛼 + 𝑑𝑡 ∧ + 𝑑𝑈 𝛽
𝑑𝑡
and conclude that 𝜔 is closed if and only if
𝑑𝛽
= 𝑑𝑈 𝛼
(2.4.19) { 𝑑𝑡
𝑑𝑈 𝛽 = 0 .
(4) Let 𝛼 be a reduced (𝑘 − 1)-form. Show that there exists a reduced (𝑘 − 1)-form 𝜈 such
that
𝑑𝜈
(2.4.20) =𝛼.
𝑑𝑡
Hint: Let 𝛼 = ∑𝐼 𝑓𝐼 (𝑥, 𝑡)𝑑𝑥𝐼 and 𝜈 = ∑ 𝑔𝐼 (𝑥, 𝑡)𝑑𝑥𝐼 . The equation (2.4.20) reduces
to the system of equations
𝑑
(2.4.21) 𝑔 (𝑥, 𝑡) = 𝑓𝐼 (𝑥, 𝑡) .
𝑑𝑡 𝐼
Let 𝑐 be a point on the interval, 𝐴, and using calculus show that equation (2.4.21) has a
unique solution, 𝑔𝐼 (𝑥, 𝑡), with 𝑔𝐼 (𝑥, 𝑐) = 0.
(5) Show that if 𝜔 is the form (2.4.18) and 𝜈 a solution of (2.4.20) then the form
(2.4.22) 𝜔 − 𝑑𝜈
is reduced.
(6) Let
𝛾 = ∑ ℎ𝐼 (𝑥, 𝑡)𝑑𝑥𝐼
𝑑𝛾
be a reduced 𝑘-form. Deduce from (2.4.19) that if 𝛾 is closed then 𝑑𝑡
= 0 and 𝑑𝑈 𝛾 = 0.
Conclude that ℎ𝐼 (𝑥, 𝑡) = ℎ𝐼 (𝑥) and that
𝛾 = ∑ ℎ𝐼 (𝑥)𝑑𝑥𝐼
𝐼

is effectively a closed 𝑘-form on 𝑈. Prove that if every closed 𝑘-form on 𝑈 is exact, then
every closed 𝑘-form on 𝑈 × 𝐴 is exact.
Hint: Let 𝜔 be a closed 𝑘-form on 𝑈 × 𝐴 and let 𝛾 be the form (2.4.22).
Exercise 2.4.vi. Let 𝑄 ⊂ 𝐑𝑛 be an open rectangle. Show that every closed form on 𝑄 of
degree 𝑘 > 0 is exact.
Hint: Let 𝑄 = (𝑎1 , 𝑏1 ) × ⋯ × (𝑎𝑛 , 𝑏𝑛 ). Prove this assertion by induction, at the 𝑛th stage
of the induction letting 𝑈 = (𝑎1 , 𝑏1 ) × ⋯ × (𝑎𝑛−1 , 𝑏𝑛−1 ) and 𝐴 = (𝑎𝑛 , 𝑏𝑛 ).
 Draft: March 28, 2018

§2.5 The interior product operation 51

2.5. The interior product operation

In §2.1 we explained how to pair a one-form 𝜔 and a vector field 𝒗 to get a function 𝜄𝒗 𝜔.
This pairing operation generalizes.
Definition 2.5.1. Let 𝑈 be an open subset of 𝐑𝑛 , 𝒗 a vector field on 𝑈, and 𝜔 ∈ 𝛺𝑘 (𝑈) The
interior product of 𝒗 with 𝜔 is (𝑘 − 1)-form 𝜄𝒗 𝜔 on 𝑈 defined by declaring the value of 𝜄𝒗 𝜔
value at 𝑝 ∈ 𝑈 to be the interior product
(2.5.2) 𝜄𝒗(𝑝) 𝜔𝑝 .
Note that 𝒗(𝑝) is in 𝑇𝑝 𝐑𝑛 and 𝜔𝑝 in 𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ), so by definition of interior product (see §1.7),
the expression (2.5.2) is an element of 𝛬𝑘−1 (𝑇𝑝⋆ 𝐑𝑛 ).
From the properties of interior product on vector spaces which we discussed in § 1.7,
one gets analogous properties for this interior product on forms. We list these properties,
leaving their verification as an easy exercise.
Properties 2.5.3. Let 𝑈 be an open subset of 𝐑𝑛 , 𝒗 and 𝒘 vector fields on 𝑈, 𝜔1 and 𝜔2 both
𝑘-forms on 𝑈, and 𝜔 a 𝑘-form and 𝜇 an ℓ-form on 𝑈.
(1) Linearity in the form: we have
𝜄𝒗 (𝜔1 + 𝜔2 ) = 𝜄𝒗 𝜔1 + 𝜄𝒗 𝜔2 .
(2) Linearity in the vector field: we have
𝜄𝒗+𝒘 𝜔 = 𝜄𝒗 𝜔 + 𝜄𝒘 𝜔 .
(3) Derivation property: we have
𝜄𝒗 (𝜔 ∧ 𝜇) = 𝜄𝒗 𝜔 ∧ 𝜇 + (−1)𝑘 𝜔 ∧ 𝜄𝒗 𝜇
(4) The interior product satisfies the identity
𝜄𝒗 (𝜄𝒘 𝜔) = −𝜄𝒘 (𝜄𝒗 𝜔) .
(5) As a special case of (4), the interior product satisfies the identity
𝜄𝒗 (𝜄𝒗 𝜔) = 0 .
(6) Moreover, if 𝜔 is decomposable, i.e., is a wedge product of one-forms
𝜔 = 𝜇1 ∧ ⋯ ∧ 𝜇𝑘 ,
then
𝑘
𝜄𝒗 𝜔 = ∑ (−1)𝑟−1 𝜄𝒗 (𝜇𝑟 )𝜇1 ∧ ⋯ ∧ 𝜇̂𝑟 ∧ ⋯ ∧ 𝜇𝑘 .
𝑟=1

We also leave for you to prove the following two assertions, both of which are special
cases of (6).
Example 2.5.4. If 𝒗 = 𝜕/𝜕𝑥𝑟 and 𝜔 = 𝑑𝑥𝐼 = 𝑑𝑥𝑖1 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑘 then
𝑘
(2.5.5) 𝜄𝒗 𝜔 = ∑(−1)𝑖−1 𝛿𝑖,𝑖𝑟 𝑑𝑥𝐼𝑟
𝑖=1
where
1, 𝑖 = 𝑖𝑟
𝛿𝑖,𝑖𝑟 ≔ {
0, 𝑖 ≠ 𝑖𝑟 .
and 𝐼𝑟 = (𝑖1 , …, 𝚤𝑟̂ , …, 𝑖𝑘 ).
 Draft: March 28, 2018

52 Chapter 2: Differential Forms

𝑛
Example 2.5.6. If 𝒗 = ∑𝑖=1 𝑓𝑖 𝜕/𝜕𝑥𝑖 and 𝜔 = 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 then
𝑛
(2.5.7) ̂ 𝑟 ⋯ ∧ 𝑑𝑥𝑛 .
𝜄𝒗 𝜔 = ∑ (−1)𝑟−1 𝑓𝑟 𝑑𝑥1 ∧ ⋯ 𝑑𝑥
𝑟=1
By combining exterior differentiation with the interior product operation one gets an-
other basic operation of vector fields on forms: the Lie differentiation operation. For zero-
forms, i.e., for 𝐶∞ functions, we defined this operation by the formula (2.1.16). For 𝑘-forms
we define it by a slightly more complicated formula.
Definition 2.5.8. Let 𝑈 be an open subset of 𝐑𝑛 , 𝒗 a vector field on 𝑈, and 𝜔 ∈ 𝛺𝑘 (𝑈). The
Lie derivative of 𝜔 with respect to 𝒗 is the 𝑘-form
(2.5.9) 𝐿𝒗 𝜔 ≔ 𝜄𝒗 (𝑑𝜔) + 𝑑(𝜄𝒗 𝜔) .
Notice that for zero-forms the second summand is zero, so (2.5.9) and (2.1.16) agree.
Properties 2.5.10. Let 𝑈 be an open subset of 𝐑𝑛 , 𝒗 a vector field on 𝑈, 𝜔 ∈ 𝛺𝑘 (𝑈), and
𝜇 ∈ 𝛺ℓ (𝑈). The Lie derivative enjoys the following properties:
(1) Commutativity with exterior differentiation: we have
(2.5.11) 𝑑(𝐿𝒗 𝜔) = 𝐿𝒗 (𝑑𝜔) .
(2) Interaction with wedge products: we have
(2.5.12) 𝐿𝒗 (𝜔 ∧ 𝜇) = 𝐿𝒗 𝜔 ∧ 𝜇 + 𝜔 ∧ 𝐿𝒗 𝜇
From Properties 2.5.10 it is fairly easy to get an explicit formula for 𝐿𝒗 𝜔. Namely let 𝜔
be the 𝑘-form
𝜔 = ∑ 𝑓𝐼 𝑑𝑥𝐼 , 𝑓𝐼 ∈ 𝐶∞ (𝑈)
𝐼
and 𝒗 the vector field
𝑛
𝜕
𝒗 = ∑ 𝑔𝑖 , 𝑔𝑖 ∈ 𝐶∞ (𝑈) .
𝑖=1 𝜕𝑥𝑖
By equation (2.5.12)
𝐿𝒗 (𝑓𝐼 𝑑𝑥𝐼 ) = (𝐿𝒗 𝑓𝐼 )𝑑𝑥𝐼 + 𝑓𝐼 (𝐿𝒗 𝑑𝑥𝐼 )
and
𝑘
𝐿𝒗 𝑑𝑥𝐼 = ∑ 𝑑𝑥𝑖1 ∧ ⋯ ∧ 𝐿𝒗 𝑑𝑥𝑖𝑟 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑘 ,
𝑟=1
and by equation (2.5.11)
𝐿𝒗 𝑑𝑥𝑖𝑟 = 𝑑𝐿𝒗 𝑥𝑖𝑟
so to compute 𝐿𝒗 𝜔 one is reduced to computing 𝐿𝒗 𝑥𝑖𝑟 and 𝐿𝒗 𝑓𝐼 .
However by equation (2.5.12)
𝐿𝒗 𝑥𝑖𝑟 = 𝑔𝑖𝑟
and
𝑛
𝜕𝑓
𝐿𝒗 𝑓𝐼 = ∑ 𝑔𝑖 𝐼 .
𝑖=1 𝜕𝑥𝑖
We leave the verification of Properties 2.5.10 as exercises, and also ask you to prove (by the
method of computation that we have just sketched) the following divergence formula.
𝑛
Lemma 2.5.13. Let 𝑈 ⊂ 𝐑𝑛 be open, let 𝑔1 , …, 𝑔𝑛 ∈ 𝐶∞ (𝑈), and set 𝒗 ≔ ∑𝑖=1 𝑔𝑖 𝜕/𝜕𝑥𝑖 . Then
𝑛
𝜕𝑔𝑖
(2.5.14) 𝐿𝒗 (𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ) = ∑ ( ) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
𝑖=1 𝜕𝑥𝑖
 Draft: March 28, 2018

§2.5 The interior product operation 53

Exercises for §2.5

Exercise 2.5.i. Verify the Properties 2.5.3 (1)–(6).
Exercise 2.5.ii. Show that if 𝜔 is the 𝑘-form 𝑑𝑥𝐼 and 𝒗 the vector field 𝜕/𝜕𝑥𝑟 , then 𝜄𝒗 𝜔 is
given by equation (2.5.5).
𝑛
Exercise 2.5.iii. Show that if 𝜔 is the 𝑛-form 𝑑𝑥1 ∧⋯∧𝑑𝑥𝑛 and 𝒗 the vector field ∑𝑖=1 𝑓𝑖 𝜕/𝜕𝑥𝑖 ,
then 𝜄𝒗 𝜔 is given by equation (2.5.7).
Exercise 2.5.iv. Let 𝑈 be an open subset of 𝐑𝑛 and 𝒗 a 𝐶∞ vector field on 𝑈. Show that for
𝜔 ∈ 𝛺𝑘 (𝑈)
𝑑𝐿𝒗 𝜔 = 𝐿𝒗 𝑑𝜔
and
𝜄𝒗 (𝐿𝒗 𝜔) = 𝐿𝒗 (𝜄𝒗 𝜔) .
Hint: Deduce the first of these identities from the identity 𝑑(𝑑𝜔) = 0 and the second
from the identity 𝜄𝒗 (𝜄𝒗 𝜔) = 0.
Exercise 2.5.v. Given 𝜔𝑖 ∈ 𝛺𝑘𝑖 (𝑈), for 𝑖 = 1, 2 show that
𝐿𝒗 (𝜔1 ∧ 𝜔2 ) = 𝐿𝒗 𝜔1 ∧ 𝜔2 + 𝜔1 ∧ 𝐿𝒗 𝜔2 .
Hint: Plug 𝜔 = 𝜔1 ∧ 𝜔2 into equation (2.5.9) and use equation (2.4.3) and the derivation
property of the interior product to evaluate the resulting expression.
Exercise 2.5.vi. Let 𝒗1 and 𝒗2 be vector fields on 𝑈 and let 𝒘 be their Lie bracket. Show that
for 𝜔 ∈ 𝛺𝑘 (𝑈)
𝐿𝒘 𝜔 = 𝐿𝒗1 (𝐿𝒗2 𝜔) − 𝐿𝒗2 (𝐿𝒗1 𝜔) .
Hint: By definition this is true for zero-forms and by equation (2.5.11) for exact one-
forms. Now use the fact that every form is a sum of wedge products of zero-forms and
one-forms and the fact that 𝐿𝒗 satisfies the product identity (2.5.12).
Exercise 2.5.vii. Prove the divergence formula (2.5.14).
Exercise 2.5.viii.
(1) Let 𝜔 = 𝛺𝑘 (𝐑𝑛 ) be the form
𝜔 = ∑ 𝑓𝐼 (𝑥1 , …, 𝑥𝑛 )𝑑𝑥𝐼
𝐼
and 𝒗 the vector field 𝜕/𝜕𝑥𝑛 . Show that
𝜕
𝐿𝒗 𝜔 = ∑ 𝑓 (𝑥 , …, 𝑥𝑛 )𝑑𝑥𝐼 .
𝜕𝑥𝑛 𝐼 1
(2) Suppose 𝜄𝒗 𝜔 = 𝐿𝒗 𝜔 = 0. Show that 𝜔 only depends on 𝑥1 , …, 𝑥𝑛−1 and 𝑑𝑥1 , …, 𝑑𝑥𝑛−1 ,
i.e., is effectively a 𝑘-form on 𝐑𝑛−1 .
(3) Suppose 𝜄𝒗 𝜔 = 𝑑𝜔 = 0. Show that 𝜔 is effectively a closed 𝑘-form on 𝐑𝑛−1 .
(4) Use these results to give another proof of the Poincaré lemma for 𝐑𝑛 . Prove by induction
on 𝑛 that every closed form on 𝐑𝑛 is exact.
Hints:
• Let 𝜔 be the form in part (1) and let
𝑥𝑛
𝑔𝐼 (𝑥1 , …, 𝑥𝑛 ) = ∫ 𝑓𝐼 (𝑥1 , …, 𝑥𝑛−1 , 𝑡)𝑑𝑡 .
0
Show that if 𝜈 = ∑𝐼 𝑔𝐼 𝑑𝑥𝐼 , then 𝐿𝒗 𝜈 = 𝜔.
 Draft: March 28, 2018

54 Chapter 2: Differential Forms

• Conclude that
(2.5.15) 𝜔 − 𝑑𝜄𝒗 𝜈 = 𝜄𝒗 𝑑𝜈 .
• Suppose 𝑑𝜔 = 0. Conclude from equation (2.5.15) and (5) of Properties 2.5.3 that
the form 𝛽 = 𝜄𝒗 𝑑𝜈 satisfies 𝑑𝛽 = 𝜄(𝒗)𝛽 = 0.
• By part (3), 𝛽 is effectively a closed form on 𝐑𝑛−1 , and by induction, 𝛽 = 𝑑𝛼. Thus
by equation (2.5.15)
𝜔 = 𝑑𝜄𝒗 𝜈 + 𝑑𝛼 .

2.6. The pullback operation on forms

Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑚 and 𝑓 ∶ 𝑈 → 𝑉 a 𝐶∞ map. Then
for 𝑝 ∈ 𝑈 and the derivative of 𝑓 at 𝑝
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝐑𝑛 → 𝑇𝑓(𝑝) 𝐑𝑚
is a linear map, so (as explained in §1.7) we get from a pullback map
(2.6.1) 𝑑𝑓𝑝⋆ ≔ (𝑑𝑓𝑝 )⋆ ∶ 𝛬𝑘 (𝑇𝑓(𝑝)
⋆
𝐑𝑚 ) → 𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ) .
In particular, let 𝜔 be a 𝑘-form on 𝑉. Then at 𝑓(𝑝) ∈ 𝑉, 𝜔 takes the value
𝜔𝑓(𝑝) ∈ 𝛬𝑘 (𝑇𝑞⋆ 𝐑𝑚 ) ,
so we can apply to it the operation (2.6.1), and this gives us an element:
(2.6.2) 𝑑𝑓𝑝⋆ 𝜔𝑓(𝑝) ∈ 𝛬𝑘 (𝑇𝑝⋆ 𝐑𝑛 ) .
In fact we can do this for every point 𝑝 ∈ 𝑈, and the assignment
(2.6.3) 𝑝 ↦ (𝑑𝑓𝑝 )⋆ 𝜔𝑓(𝑝)
defines a 𝑘-form on 𝑈, which we denote by 𝑓⋆ 𝜔. We call 𝑓⋆ 𝜔 the pullback of 𝜔 along the
map 𝑓. A few of its basic properties are described below.
Properties 2.6.4. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑚 and 𝑓 ∶ 𝑈 → 𝑉 a
𝐶∞ map.
(1) Let 𝜙 be a zero-form, i.e., a function 𝜙 ∈ 𝐶∞ (𝑉). Since
𝛬0 (𝑇𝑝⋆ ) = 𝛬0 (𝑇𝑓(𝑝)
⋆
)=𝐑
the map (2.6.1) is just the identity map of 𝐑 onto 𝐑 when 𝑘 = 0. Hence for zero-forms
(2.6.5) (𝑓⋆ 𝜙)(𝑝) = 𝜙(𝑓(𝑝)) ,
that is, 𝑓⋆ 𝜙 is just the composite function, 𝜙 ∘ 𝑓 ∈ 𝐶∞ (𝑈).
(2) Let 𝜙 ∈ 𝛺0 (𝑈) and let 𝜇 ∈ 𝛺1 (𝑉) be the 1-form 𝜇 = 𝑑𝜙. By the chain rule (2.6.2)
unwinds to:
(2.6.6) (𝑑𝑓𝑝 )⋆ 𝑑𝜙𝑞 = (𝑑𝜙)𝑞 ∘ 𝑑𝑓𝑝 = 𝑑(𝜙 ∘ 𝑓)𝑝
and hence by (2.6.5) we have
(2.6.7) 𝑓⋆ 𝑑𝜙 = 𝑑𝑓⋆ 𝜙 .
(3) If 𝜔1 , 𝜔2 ∈ 𝛺𝑘 (𝑉) from equation (2.6.2) we see that
(𝑑𝑓𝑝 )⋆ (𝜔1 + 𝜔2 )𝑞 = (𝑑𝑓𝑝 )⋆ (𝜔1 )𝑞 + (𝑑𝑓𝑝 )⋆ (𝜔2 )𝑞 ,
and hence by eq:2.5.3 we have
𝑓⋆ (𝜔1 + 𝜔2 ) = 𝑓⋆ 𝜔1 + 𝑓⋆ 𝜔2 .
 Draft: March 28, 2018

§2.6 The pullback operation on forms 55

(4) We observed in §1.7 that the operation (2.6.1) commutes with wedge-product, hence if
𝜔1 ∈ 𝛺𝑘 (𝑉) and 𝜔2 ∈ 𝛺ℓ (𝑉)
𝑑𝑓𝑝⋆ (𝜔1 )𝑞 ∧ (𝜔2 )𝑞 = 𝑑𝑓𝑝⋆ (𝜔1 )𝑞 ∧ 𝑑𝑓𝑝⋆ (𝜔2 )𝑞 .
In other words
(2.6.8) 𝑓⋆ 𝜔1 ∧ 𝜔2 = 𝑓⋆ 𝜔1 ∧ 𝑓⋆ 𝜔2 .
(5) Let 𝑊 be an open subset of 𝐑𝑘 and 𝑔 ∶ 𝑉 → 𝑊 a 𝐶∞ map. Given a point 𝑝 ∈ 𝑈, let
𝑞 = 𝑓(𝑝) and 𝑤 = 𝑔(𝑞). Then the composition of the map
(𝑑𝑓𝑝 )⋆ ∶ 𝛬𝑘 (𝑇𝑞⋆ ) → 𝛬𝑘 (𝑇𝑝⋆ )
and the map
(𝑑𝑔𝑞 )⋆ ∶ 𝛬𝑘 (𝑇𝑤⋆ ) → 𝛬𝑘 (𝑇𝑞⋆ )
is the map
(𝑑𝑔𝑞 ∘ 𝑑𝑓𝑝 )⋆ ∶ 𝛬𝑘 (𝑇𝑤⋆ ) → 𝛬𝑘 (𝑇𝑝⋆ )
by formula (1.7.5). However, by the chain rule
(𝑑𝑔𝑞 ) ∘ (𝑑𝑓)𝑝 = 𝑑(𝑔 ∘ 𝑓)𝑝
so this composition is the map
𝑑(𝑔 ∘ 𝑓)𝑝⋆ ∶ 𝛬𝑘 (𝑇𝑤⋆ ) → 𝛬𝑘 (𝑇𝑝⋆ ) .
Thus if 𝜔 ∈ 𝛺𝑘 (𝑊)
𝑓⋆ (𝑔⋆ 𝜔) = (𝑔 ∘ 𝑓)⋆ 𝜔 .
Let us see what the pullback operation looks like in coordinates. Using multi-index
notation we can express every 𝑘-form 𝜔 ∈ 𝛺𝑘 (𝑉) as a sum over multi-indices of length 𝑘
(2.6.9) 𝜔 = ∑ 𝜙𝐼 𝑑𝑥𝐼 ,
𝐼
the coefficient 𝜙𝐼 of 𝑑𝑥𝐼 being in 𝐶∞ (𝑉). Hence by (2.6.5)
𝑓⋆ 𝜔 = ∑ 𝑓⋆ 𝜙𝐼 𝑓⋆ (𝑑𝑥𝐼 )
where 𝑓⋆ 𝜙𝐼 is the function of 𝜙 ∘ 𝑓.
What about 𝑓⋆ 𝑑𝑥𝐼 ? If 𝐼 is the multi-index, (𝑖1 , …, 𝑖𝑘 ), then by definition
𝑑𝑥𝐼 = 𝑑𝑥𝑖1 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑘
so
𝑓⋆ (𝑑𝑥𝐼 ) = 𝑓⋆ (𝑑𝑥𝑖1 ) ∧ ⋯ ∧ 𝑓⋆ (𝑑𝑥𝑖𝑘 )
by (2.6.8), and by (2.6.7), we have
𝑓⋆ 𝑑𝑥𝑖 = 𝑑𝑓⋆ 𝑥𝑖 = 𝑑𝑓𝑖
where 𝑓𝑖 is the 𝑖th coordinate function of the map 𝑓. Thus, setting
𝑑𝑓𝐼 ≔ 𝑑𝑓𝑖1 ∧ ⋯ ∧ 𝑑𝑓𝑖𝑘 ,
we see that for each multi-index 𝐼 we have
(2.6.10) 𝑓⋆ (𝑑𝑥𝐼 ) = 𝑑𝑓𝐼 ,
and for the pullback of the form (2.6.9)
(2.6.11) 𝑓⋆ 𝜔 = ∑ 𝑓⋆ 𝜙𝐼 𝑑𝑓𝐼 .
𝐼
 Draft: March 28, 2018

56 Chapter 2: Differential Forms

We will use this formula to prove that pullback commutes with exterior differentiation:
(2.6.12) 𝑑(𝑓⋆ 𝜔) = 𝑓⋆ (𝑑𝜔) .
To prove this we recall that by (2.3.6) we have 𝑑(𝑑𝑓𝐼 ) = 0, hence by (2.3.1) and (2.6.10) we
have
𝑑(𝑓⋆ 𝜔) = ∑ 𝑑(𝑓⋆ 𝜙𝐼 ) ∧ 𝑑𝑓𝐼 = ∑ 𝑓⋆ (𝑑𝜙𝐼 ) ∧ 𝑓⋆ (𝑑𝑥𝐼 )
𝐼 𝐼
⋆ ⋆
= 𝑓 ∑ 𝑑𝜙𝐼 ∧ 𝑑𝑥𝐼 = 𝑓 (𝑑𝜔) .
𝐼
A special case of formula (2.6.10) will be needed in Chapter 4: Let 𝑈 and 𝑉 be open
subsets of 𝐑𝑛 and let 𝜔 = 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 . Then by (2.6.10) we have
𝑓⋆ 𝜔𝑝 = (𝑑𝑓1 )𝑝 ∧ ⋯ ∧ (𝑑𝑓𝑛 )𝑝
for all 𝑝 ∈ 𝑈. However,
𝜕𝑓𝑖
(𝑑𝑓𝑖 )𝑝 = ∑ (𝑝)(𝑑𝑥𝑗 )𝑝
𝜕𝑥𝑗
and hence by formula (1.7.10)
𝜕𝑓𝑖
𝑓⋆ 𝜔𝑝 = det [ (𝑝)] (𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 )𝑝 .
𝜕𝑥𝑗
In other words
𝜕𝑓𝑖
(2.6.13) 𝑓⋆ (𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ) = det [ ] 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
𝜕𝑥𝑗
In Exercise 2.6.iv and equation (2.6.6) we outline the proof of an important topological
property of the pullback operation.
Definition 2.6.14. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑚 , 𝐴 ⊂ 𝐑 an open
interval containing 0 and 1 and 𝑓0 , 𝑓1 ∶ 𝑈 → 𝑉 two 𝐶∞ maps. A 𝐶∞ map 𝐹 ∶ 𝑈 × 𝐴 → 𝑉 is
a 𝐶∞ homotopy between 𝑓0 and 𝑓1 if 𝐹(𝑥, 0) = 𝑓0 (𝑥) and 𝐹(𝑥, 1) = 𝑓1 (𝑥).
If there exists a homotopy between 𝑓0 and 𝑓1 , we say that 𝑓0 and 𝑓1 are homotopic and
write 𝑓0 ≃ 𝑓1 .
Intuitively, 𝑓0 and 𝑓1 are homotopic if there exists a family of 𝐶∞ maps 𝑓𝑡 ∶ 𝑈 → 𝑉,
where 𝑓𝑡 (𝑥) = 𝐹(𝑥, 𝑡), which “smoothly deform 𝑓0 into 𝑓1 ”. In Exercise 2.6.iv and equa-
tion (2.6.6) you will be asked to verify that for 𝑓0 and 𝑓1 to be homotopic they have to
satisfy the following criteria.
Theorem 2.6.15. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑚 , and 𝑓0 , 𝑓1 ∶ 𝑈 → 𝑉
two 𝐶∞ maps. If 𝑓0 and 𝑓1 are homotopic then for every closed form 𝜔 ∈ 𝛺𝑘 (𝑉) the form
𝑓1⋆ 𝜔 − 𝑓0⋆ 𝜔 is exact.
This theorem is closely related to the Poincaré lemma, and, in fact, one gets from it
a slightly stronger version of the Poincaré lemma than that described in Exercise 2.3.iv
and equation (2.3.7).
Definition 2.6.16. An open subset 𝑈 of 𝐑𝑛 is contractible if, for some point 𝑝0 ∈ 𝑈, the
identity map id𝑈 ∶ 𝑈 → 𝑈 is homotopic to the constant map
𝑓0 ∶ 𝑈 → 𝑈 , 𝑓0 (𝑝) = 𝑝0
at 𝑝0 .
 Draft: March 28, 2018

§2.6 The pullback operation on forms 57

From Theorem 2.6.15 it is easy to see that the Poincaré lemma holds for contractable
open subsets of 𝐑𝑛 . If 𝑈 is contractable every closed 𝑘-form on 𝑈 of degree 𝑘 > 0 is exact. To
see this, note that if 𝜔 ∈ 𝛺𝑘 (𝑈), then for the identity map id⋆𝑈 𝜔 = 𝜔 and if 𝑓 is the constant
map at a point of 𝑈, then 𝑓⋆ 𝜔 = 0.
Exercises for §2.6
3 3
Exercise 2.6.i. Let 𝑓 ∶ 𝐑 → 𝐑 be the map
𝑓(𝑥1 , 𝑥2 , 𝑥3 ) = (𝑥1 𝑥2 , 𝑥2 𝑥32 , 𝑥33 ) .
Compute the pullback, 𝑓⋆ 𝜔 for the following forms.
(1) 𝜔 = 𝑥2 𝑑𝑥3
(2) 𝜔 = 𝑥1 𝑑𝑥1 ∧ 𝑑𝑥3
(3) 𝜔 = 𝑥1 𝑑𝑥1 ∧ 𝑑𝑥2 ∧ 𝑑𝑥3
Exercise 2.6.ii. Let 𝑓 ∶ 𝐑2 → 𝐑3 be the map
𝑓(𝑥1 , 𝑥2 ) = (𝑥12 , 𝑥22 , 𝑥1 𝑥2 ) .
Complete the pullback, 𝑓⋆ 𝜔, for the following forms.
(1) 𝜔 = 𝑥2 𝑑𝑥2 + 𝑥3 𝑑𝑥3
(2) 𝜔 = 𝑥1 𝑑𝑥2 ∧ 𝑑𝑥3
(3) 𝜔 = 𝑑𝑥1 ∧ 𝑑𝑥2 ∧ 𝑑𝑥3
Exercise 2.6.iii. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑚 , 𝑓 ∶ 𝑈 → 𝑉 a 𝐶∞
map and 𝛾 ∶ [𝑎, 𝑏] → 𝑈 a 𝐶∞ curve. Show that for 𝜔 ∈ 𝛺1 (𝑉)

∫ 𝑓⋆ 𝜔 = ∫ 𝜔
𝛾 𝛾1

where 𝛾1 ∶ [𝑎, 𝑏] → 𝑉 is the curve, 𝛾1 (𝑡) = 𝑓(𝛾(𝑡)). (See Exercise 2.2.viii.)

Exercise 2.6.iv. Let 𝑈 be an open subset of 𝐑𝑛 , 𝐴 ⊂ 𝐑 an open interval containing the
points, 0 and 1, and (𝑥, 𝑡) product coordinates on 𝑈 × 𝐴. Recall from Exercise 2.3.v that a
form 𝜇 ∈ 𝛺ℓ (𝑈 × 𝐴) is reduced if it can be written as a sum
𝜇 = ∑ 𝑓𝐼 (𝑥, 𝑡)𝑑𝑥𝐼
𝐼

(i.e., none of the summands involve 𝑑𝑡). For a reduced form 𝜇, let 𝑄𝜇 ∈ 𝛺ℓ (𝑈) be the form
1
(2.6.17) 𝑄𝜇 ≔ (∑ ∫ 𝑓𝐼 (𝑥, 𝑡)𝑑𝑡) 𝑑𝑥𝐼
𝐼 0
ℓ
and let 𝜇1 , 𝜇2 ∈ 𝛺 (𝑈), be the forms
𝜇0 = ∑ 𝑓𝐼 (𝑥, 0)𝑑𝑥𝐼
𝐼
and
𝜇1 = ∑ 𝑓𝐼 (𝑥, 1)𝑑𝑥𝐼 .
Now recall from Exercise 2.4.v that every form 𝜔 ∈ 𝛺𝑘 (𝑈 × 𝐴) can be written uniquely as a
sum
(2.6.18) 𝜔 = 𝑑𝑡 ∧ 𝛼 + 𝛽
where 𝛼 and 𝛽 are reduced.
(1) Prove:
 Draft: March 28, 2018

58 Chapter 2: Differential Forms

Theorem 2.6.19. If the form (2.6.18) is closed then

(2.6.20) 𝛽1 − 𝛽0 = 𝑑𝑄𝛼 .
Hint: Equation (2.4.19).Let 𝜄0 and 𝜄1 be the maps of 𝑈 into 𝑈×𝐴 defined by 𝜄0 (𝑥) = (𝑥, 0)
and 𝜄1 (𝑥) = (𝑥, 1). Show that (2.6.20) can be rewritten
(2.6.21) 𝜄⋆1 𝜔 − 𝜄⋆0 𝜔 = 𝑑𝑄𝛼 .
Exercise 2.6.v. Let 𝑉 be an open subset of 𝐑𝑚 and 𝑓0 , 𝑓1 ∶ 𝑈 → 𝑉 be 𝐶∞ maps. Suppose 𝑓0
and 𝑓1 are homotopic. Show that for every closed form, 𝜇 ∈ 𝛺𝑘 (𝑉), 𝑓1⋆ 𝜇 − 𝑓0⋆ 𝜇 is exact.
Hint: Let 𝐹 ∶ 𝑈 × 𝐴 → 𝑉 be a homotopy between 𝑓0 and 𝑓1 and let 𝜔 = 𝐹⋆ 𝜇. Show that
𝜔 is closed and that 𝑓0⋆ 𝜇 = 𝜄⋆0 𝜔 and 𝑓1⋆ 𝜇 = 𝜄⋆1 𝜔. Conclude from equation (2.6.21)that
𝑓1⋆ 𝜇 − 𝑓0⋆ 𝜇 = 𝑑𝑄𝛼
where 𝜔 = 𝑑𝑡 ∧ 𝛼 + 𝛽 and 𝛼 and 𝛽 are reduced.
Exercise 2.6.vi. Show that if 𝑈 ⊂ 𝐑𝑛 is a contractable open set, then the Poincaré lemma
holds: every closed form of degree 𝑘 > 0 is exact.
Exercise 2.6.vii. An open subset, 𝑈, of 𝐑𝑛 is said to be star-shaped if there exists a point
𝑝0 ∈ 𝑈, with the property that for every point 𝑝 ∈ 𝑈, the line segment,
{ 𝑡𝑝 + (1 − 𝑡)𝑝0 | 0 ≤ 𝑡 ≤ 1 } ,
joining 𝑝 to 𝑝0 is contained in 𝑈. Show that if 𝑈 is star-shaped it is contractable.
Exercise 2.6.viii. Show that the following open sets are star-shaped.
(1) The open unit ball
{ 𝑥 ∈ 𝐑𝑛 | |𝑥| < 1 } .
(2) The open rectangle, 𝐼1 × ⋯ × 𝐼𝑛 , where each 𝐼𝑘 is an open subinterval of 𝐑.
(3) 𝐑𝑛 itself.
(4) Product sets
𝑈1 × 𝑈2 ⊂ 𝐑𝑛 ≅ 𝐑𝑛1 × 𝐑𝑛2
where 𝑈𝑖 is a star-shaped open set in 𝐑𝑛𝑖 .
Exercise 2.6.ix. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑓𝑡 ∶ 𝑈 → 𝑈, 𝑡 ∈ 𝐑, a one-parameter group
of diffeomorphisms and 𝑣 its infinitesimal generator. Given 𝜔 ∈ 𝛺𝑘 (𝑈) show that at 𝑡 = 0
𝑑 ⋆
(2.6.22) 𝑓 𝜔 = 𝐿𝒗 𝜔 .
𝑑𝑡 𝑡
Here is a sketch of a proof:
(1) Let 𝛾(𝑡) be the curve, 𝛾(𝑡) = 𝑓𝑡 (𝑝), and let 𝜙 be a zero-form, i.e., an element of 𝐶∞ (𝑈).
Show that
𝑓𝑡⋆ 𝜙(𝑝) = 𝜙(𝛾(𝑡))
and by differentiating this identity at 𝑡 = 0 conclude that (2.6.22) holds for zero-forms.
(2) Show that if (2.6.22) holds for 𝜔 it holds for 𝑑𝜔.
Hint: Differentiate the identity
𝑓𝑡⋆ 𝑑𝜔 = 𝑑𝑓𝑡⋆ 𝜔
at 𝑡 = 0.
 Draft: March 28, 2018

§2.7 Divergence, curl, and gradient 59

(3) Show that if (2.6.22) holds for 𝜔1 and 𝜔2 it holds for 𝜔1 ∧ 𝜔2 .

Hint: Differentiate the identity
𝑓𝑡⋆ (𝜔1 ∧ 𝜔2 ) = 𝑓𝑡⋆ 𝜔1 ∧ 𝑓𝑡⋆ 𝜔2
at 𝑡 = 0.
(4) Deduce (2.6.22) from (1)–(3).
Hint: Every 𝑘-form is a sum of wedge products of zero-forms and exact one-forms.
Exercise 2.6.x. In Exercise 2.6.ix show that for all 𝑡
𝑑 ⋆
(2.6.23) 𝑓 𝜔 = 𝑓𝑡⋆ 𝐿𝒗 𝜔 = 𝐿𝒗 𝑓𝑡⋆ 𝜔 .
𝑑𝑡 𝑡
Hint: By the definition of a one-parameter group, 𝑓𝑠+𝑡 = 𝑓𝑠 ∘ 𝑓𝑡 = 𝑓𝑟 ∘ 𝑓𝑠 , hence:
⋆
𝑓𝑠+𝑡 𝜔 = 𝑓𝑡⋆ (𝑓𝑠⋆ 𝜔) = 𝑓𝑠⋆ (𝑓𝑡⋆ 𝜔) .
Prove the first assertion by differentiating the first of these identities with respect to 𝑠 and
then setting 𝑠 = 0, and prove the second assertion by doing the same for the second of these
identities.
In particular conclude that
𝑓𝑡⋆ 𝐿𝒗 𝜔 = 𝐿𝒗 𝑓𝑡⋆ 𝜔 .
Exercise 2.6.xi.
(1) By massaging the result above show that
𝑑 ⋆
𝑓 𝜔 = 𝑑𝑄𝑡 𝜔 + 𝑄𝑡 𝑑𝜔
𝑑𝑡 𝑡
where
𝑄𝑡 𝜔 = 𝑓𝑡⋆ 𝜄𝒗 𝜔 .
Hint: Equation (2.5.9).
(2) Let
1
𝑄𝜔 = ∫ 𝑓𝑡⋆ 𝜄𝒗 𝜔𝑑𝑡 .
0
Prove the homotopy identity
𝑓1⋆ 𝜔 − 𝑓0⋆ 𝜔 = 𝑑𝑄𝜔 + 𝑄𝑑𝜔 .
Exercise 2.6.xii. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑚 , 𝒗 a vector field on
𝑈, 𝒘 a vector field on 𝑉 and 𝑓 ∶ 𝑈 → 𝑉 a 𝐶∞ map. Show that if 𝒗 and 𝒘 are 𝑓-related
𝜄𝒗 𝑓⋆ 𝜔 = 𝑓⋆ 𝜄𝒘 𝜔 .

2.7. Divergence, curl, and gradient

The basic operations in 3-dimensional vector calculus: gradient, curl and divergence
are, by definition, operations on vector fields. As we see below these operations are closely
related to the operations
(2.7.1) 𝑑 ∶ 𝛺𝑘 (𝐑3 ) → 𝛺𝑘+1 (𝐑3 )
in degrees 𝑘 = 0, 1, 2. However, only the gradient and divergence generalize to 𝑛 dimensions.
(They are essentially the 𝑑-operations in degrees zero and 𝑛 − 1.) And, unfortunately, there
is no simple description in terms of vector fields for the other 𝑑-operations. This is one of
the main reasons why an adequate theory of vector calculus in 𝑛-dimensions forces on one
the differential form approach that we have developed in this chapter.
 Draft: March 28, 2018

60 Chapter 2: Differential Forms

Even in three dimensions, however, there is a good reason for replacing the gradient,
divergence, and curl by the three operations (2.7.1). A problem that physicists spend a lot of
time worrying about is the problem of general covariance: formulating the laws of physics in
such a way that they admit as large a set of symmetries as possible, and frequently these for-
mulations involve differential forms. An example is Maxwell’s equations, the fundamental
laws of electromagnetism. These are usually expressed as identities involving div and curl.
However, as we explain below, there is an alternative formulation of Maxwell’s equations
based on the operations (2.7.1), and from the point of view of general covariance, this for-
mulation is much more satisfactory: the only symmetries of 𝐑3 which preserve div and curl
are translations and otations, whereas the operations (2.7.1) admit all diffeomorphisms of
𝐑3 as symmetries.
To describe how the gradient, divergence, and curl are related to the operations (2.7.1)
we first note that there are two ways of converting vector fields into forms.
The first makes use of the natural inner product 𝐵(𝑣, 𝑤) = ∑ 𝑣𝑖 𝑤𝑖 on 𝐑𝑛 . From this
inner product one gets by Exercise 1.2.ix a bijective linear map
𝐿 ∶ 𝐑𝑛 ⥲ (𝐑𝑛 )⋆
with the defining property: 𝐿(𝑣) = ℓ if and only if ℓ(𝑤) = 𝐵(𝑣, 𝑤). Via the identification
(2.1.2) 𝐵 and 𝐿 can be transferred to 𝑇𝑝 𝐑𝑛 , giving one an inner product 𝐵𝑝 on 𝑇𝑝 𝐑𝑛 and a
bijective linear map
𝐿𝑝 ∶ 𝑇𝑝 𝐑𝑛 ⥲ 𝑇𝑝⋆ 𝐑𝑛 .
Hence if we are given a vector field 𝒗 on 𝑈 we can convert it into a 1-form 𝒗♯ by setting
(2.7.2) 𝒗♯ (𝑝) ≔ 𝐿𝑝 𝒗(𝑝)
and this sets up a bijective correspondence between vector fields and 1-forms.
For instance
𝜕
𝒗= ⟺ 𝒗♯ = 𝑑𝑥𝑖 ,
𝜕𝑥𝑖
(see Exercise 2.7.iii) and, more generally,
𝑛 𝑛
𝜕
(2.7.3) 𝒗 = ∑ 𝑓𝑖 ⟺ 𝒗♯ = ∑ 𝑓𝑖 𝑑𝑥𝑖 .
𝑖=1 𝜕𝑥𝑖 𝑖=1

Example 2.7.4. In particular if 𝑓 is a 𝐶∞ function on 𝑈 ⊂ 𝐑𝑛 the gradient of 𝑓 is the vector

field
𝑛
𝜕𝑓
grad(𝑓) ≔ ∑ ,
𝑖=1 𝜕𝑥𝑖
and this gets converted by (2.7.5) into the 1-form 𝑑𝑓. Thus the gradient operation in vector
calculus is basically just the exeterior derivative operation 𝑑 ∶ 𝛺0 (𝑈) → 𝛺1 (𝑈).
The second way of converting vector fields into forms is via the interior product oper-
ation. Namely let 𝛺 be the 𝑛-form 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 . Given an open subset 𝑈 of 𝐑𝑛 and a 𝐶∞
vector field
𝑛
𝜕
(2.7.5) 𝒗 = ∑ 𝑓𝑖
𝑖=1 𝜕𝑥𝑖
on 𝑈 the interior product of 𝑣 with 𝛺 is the (𝑛 − 1)-form
𝑛
(2.7.6) ̂ 𝑟 ⋯ ∧ 𝑑𝑥𝑛 .
𝜄𝒗 𝛺 = ∑ (−1)𝑟−1 𝑓𝑟 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑟=1
 Draft: March 28, 2018

§2.7 Divergence, curl, and gradient 61

Moreover, every (𝑛 − 1)-form can be written uniquely as such a sum, so (2.7.5) and (2.7.6)
set up a bijective correspondence between vector fields and (𝑛 − 1)-forms. Under this corre-
spondence the 𝑑-operation gets converted into an operation on vector fields
𝒗 ↦ 𝑑𝜄𝒗 𝛺 .
Moreover, by (2.5.9)
𝑑𝜄𝒗 𝛺 = 𝐿𝒗 𝛺
and by (2.5.14)
𝐿𝒗 𝛺 = div(𝒗)𝛺
where
𝑛
𝜕𝑓𝑖
(2.7.7) div(𝒗) = ∑ .
𝑖=1 𝜕𝑥𝑖

In other words, this correspondence between (𝑛 − 1)-forms and vector fields converts the
𝑑-operation into the divergence operation (2.7.7) on vector fields.
Notice that divergence and gradient are well-defined as vector calculus operations in 𝑛
dimensions, even though one usually thinks of them as operations in 3-dimensional vector
calculus. The curl operation, however, is intrinsically a 3-dimensional vector calculus oper-
ation. To define it we note that by (2.7.6) every 2-form 𝜇 on an open subset 𝑈 ⊂ 𝐑3 can be
written uniquely as an interior product,
(2.7.8) 𝜇 = 𝜄𝒘 𝑑𝑥1 ∧ 𝑑𝑥2 ∧ 𝑑𝑥3 ,
for some vector field 𝒘, and the left-hand side of this formula determines 𝒘 uniquely.
Definition 2.7.9. Let 𝑈 be an open subset of 𝐑3 and 𝒗 a vector field on 𝑈. From 𝒗 we get by
(2.7.3) a 1-form 𝒗♯ , and hence by (2.7.8) a vector field 𝒘 satisfying
𝑑𝒗♯ = 𝜄𝒘 𝑑𝑥1 ∧ 𝑑𝑥2 ∧ 𝑑𝑥3 .
The curl of 𝒗 is the vector field
curl(𝒗) ≔ 𝒘 .
We leave for you to check that this definition coincides with the definition one finds in
calculus books. More explicitly we leave for you to check that if 𝒗 is the vector field
𝜕 𝜕 𝜕
𝒗 = 𝑓1 + 𝑓2 + 𝑓3
𝜕𝑥1 𝜕𝑥2 𝜕𝑥3
then
𝜕 𝜕 𝜕
curl(𝒗) = 𝑔1 + 𝑔2 + 𝑔3
𝜕𝑥1 𝜕𝑥2 𝜕𝑥3
where
𝜕𝑓2 𝜕𝑓3
𝑔1 = −
𝜕𝑥3 𝜕𝑥2
𝜕𝑓 𝜕𝑓
(2.7.10) 𝑔2 = 3 − 1
𝜕𝑥1 𝜕𝑥3
𝜕𝑓 𝜕𝑓
𝑔3 = 1 − 2 .
𝜕𝑥2 𝜕𝑥1
To summarize: the gradient, curl, and divergence operations in 3-dimensions are basi-
cally just the three operations (2.7.1). The gradient operation is the operation (2.7.1) in de-
gree zero, curl is the operation (2.7.1) in degree one, and divergence is the operation (2.7.1)
 Draft: March 28, 2018

62 Chapter 2: Differential Forms

in degree two. However, to define gradient we had to assign an inner product 𝐵𝑝 to the tan-
gent space 𝑇𝑝 𝐑𝑛 for each 𝑝 ∈ 𝑈; to define divergence we had to equip 𝑈 with the 3-form
𝛺 and to define curl, the most complicated of these three operations, we needed the inner
products 𝐵𝑝 and the form 𝛺. This is why diffeomorphisms preserve the three operations
(2.7.1) but do not preserve gradient, curl, and divergence. The additional structures which
one needs to define grad, curl and div are only preserved by translations and rotations.
Maxwell’s equations and differential forms
We conclude this section by showing how Maxwell’s equations, which are usually for-
mulated in terms of divergence and curl, can be reset into “form” language. The discussion
below is an abbreviated version of [5, §1.20].
Maxwell’s equations assert:
(2.7.11) div(𝒗𝐸 ) = 𝑞
𝜕
(2.7.12) curl(𝒗𝐸 ) = − 𝒗
𝜕𝑡 𝑀
(2.7.13) div(𝒗𝑀 ) = 0
𝜕
(2.7.14) 𝑐2 curl(𝒗𝑀 ) = 𝒘 + 𝒗
𝜕𝑡 𝐸
where 𝒗𝐸 and 𝒗𝑀 are the electric and magnetic fields, 𝑞 is the scalar charge density, 𝒘 is the
current density and 𝑐 is the velocity of light. (To simplify (2.7.15) slightly we assume that
our units of space–time are chosen so that 𝑐 = 1.) As above let 𝛺 = 𝑑𝑥1 ∧ 𝑑𝑥2 ∧ 𝑑𝑥3 and let
𝜇𝐸 = 𝜄(𝒗𝐸 )𝛺
and
𝜇𝑀 = 𝜄(𝒗𝑀 )𝛺 .
We can then rewrite equations (2.7.11) and (2.7.13) in the form
(2.7.11′ ) 𝑑𝜇𝐸 = 𝑞𝛺
and
(2.7.13′ ) 𝑑𝜇𝑀 = 0 .
What about (2.7.12) and (2.7.14)? We leave the following “form” versions of these equa-
tions as an exercise:
♯ 𝜕
(2.7.12′ ) 𝑑𝒗𝐸 = − 𝜇𝑀
𝜕𝑡
and
♯ 𝜕
(2.7.14′ ) 𝑑𝒗𝑀 = 𝜄𝒘 𝛺 + 𝜇𝐸
𝜕𝑡
♯ ♯
where the 1-forms, 𝒗𝐸 and 𝒗𝑀 , are obtained from 𝒗𝐸 and 𝒗𝑀 by the operation (2.7.2).
These equations can be written more compactly as differential form identities in 3 + 1
dimensions. Let 𝜔𝑀 and 𝜔𝐸 be the 2-forms
♯
𝜔𝑀 = 𝜇𝑀 − 𝒗𝐸 ∧ 𝑑𝑡
and
♯
(2.7.15) 𝜔𝐸 = 𝜇𝐸 − 𝒗𝑀 ∧ 𝑑𝑡
and let 𝛬 be the 3-form
𝛬 ≔ 𝑞𝛺 + 𝜄𝒘 𝛺 ∧ 𝑑𝑡 .
 Draft: March 28, 2018

§2.8 Symplectic geometry & classical mechanics 63

We will leave for you to show that the four equations (2.7.11)–(2.7.14) are equivalent
to two elegant and compact (3 + 1)-dimensional identities
𝑑𝜔𝑀 = 0
and
𝑑𝜔𝐸 = 𝛬 .
Exercises for §2.7
Exercise 2.7.i. Verify that the curl operation is given in coordinates by equation (2.7.10).
Exercise 2.7.ii. Verify that Maxwell’s equations (2.7.11) and (2.7.12) become the equations (2.7.13)
and (2.7.14) when rewritten in differential form notation.
Exercise 2.7.iii. Show that in (3 + 1)-dimensions Maxwell’s equations take the form of equa-
tions (2.7.10) and (2.7.11).
Exercise 2.7.iv. Let 𝑈 be an open subset of 𝐑3 and 𝒗 a vector field on 𝑈. Show that if 𝒗 is
the gradient of a function, its curl has to be zero.
Exercise 2.7.v. If 𝑈 is simply connected prove the converse: If the curl of 𝒗 vanishes, 𝒗 is
the gradient of a function.
Exercise 2.7.vi. Let 𝒘 = curl(𝒗). Show that the divergence of 𝒘 is zero.
Exercise 2.7.vii. Is the converse to Exercise 2.7.vi true? Suppose the divergence of 𝒘 is zero.
Is 𝒘 = curl(𝒗) for some vector field 𝒗?

2.8. Symplectic geometry & classical mechanics

In this section we describe some other applications of the theory of differential forms
to physics. Before describing these applications, however, we say a few words about the geo-
metric ideas that are involved.
Definition 2.8.1. Let 𝑥1 , …, 𝑥2𝑛 be the standard coordinate functions on 𝐑2𝑛 and for 𝑖 =
1, …, 𝑛 let 𝑦𝑖 ≔ 𝑥𝑛+𝑖 . The two-form
𝑛 𝑛
𝜔 ≔ ∑ 𝑑𝑥𝑖 ∧ 𝑑𝑦𝑖 = ∑ 𝑑𝑥𝑖 ∧ 𝑑𝑥𝑛+𝑖
𝑖=1 𝑖=1

is called the Darboux form on 𝐑2𝑛 .

From the identity
𝑛
(2.8.2) 𝜔 = −𝑑 (∑𝑖=1 𝑦𝑖 𝑑𝑥𝑖 ) .
it follows that 𝜔 is exact. Moreover computing the 𝑛-fold wedge product of 𝜔 with itself we
get
𝑛 𝑛
𝜔𝑛 = ( ∑ 𝑑𝑥𝑖1 ∧ 𝑑𝑦𝑖1 ) ∧ ⋯ ∧ ( ∑ 𝑑𝑥𝑖𝑛 ∧ 𝑑𝑦𝑖𝑛 )
𝑖𝑖 =1 𝑖𝑛 =1

= ∑ 𝑑𝑥𝑖1 ∧ 𝑑𝑦𝑖1 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑛 ∧ 𝑑𝑦𝑖𝑛 .

𝑖1 ,…,𝑖𝑛

We can simplify this sum by noting that if the multi-index 𝐼 = (𝑖1 , …, 𝑖𝑛 ) is repeating, the
wedge product
(2.8.3) 𝑑𝑥𝑖1 ∧ 𝑑𝑦𝑖1 ∧ ⋯ ∧ 𝑑𝑥𝑖𝑛 ∧ 𝑑𝑥𝑖𝑛
 Draft: March 28, 2018

64 Chapter 2: Differential Forms

involves two repeating 𝑑𝑥𝑖𝑗 and hence is zero, and if 𝐼 is non-repeating we can permute the
factors and rewrite (2.8.3) in the form
𝑑𝑥1 ∧ 𝑑𝑦1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ∧ 𝑑𝑦𝑛 .
(See Exercise 1.6.v.)
Hence since these are exactly 𝑛! non-repeating multi-indices
𝜔𝑛 = 𝑛!𝑑𝑥1 ∧ 𝑑𝑦1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ∧ 𝑑𝑦𝑛 ,
i.e.,
1 𝑛
(2.8.4) 𝜔 =𝛺
𝑛!
where
(2.8.5) 𝛺 = 𝑑𝑥1 ∧ 𝑑𝑦1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ∧ 𝑑𝑦𝑛
is the symplectic volume form on 𝐑2𝑛 .
Definition 2.8.6. Let 𝑈 and 𝑉 be open subsets of 𝐑2𝑛 . A diffeomorphism 𝑓 ∶ 𝑈 ⥲ 𝑉 is a
symplectic diffeomorphism or symplectomorphism if 𝑓⋆ 𝜔 = 𝜔, where 𝜔 denotes the Dar-
boux form on 𝐑2𝑛 .
Definition 2.8.7. Let 𝑈 be an open subset of 𝐑2𝑛 , let
(2.8.8) 𝑓𝑡 ∶ 𝑈 ⥲ 𝑈 , − ∞ < 𝑡 < ∞
be a one-parameter group of diffeomorphisms of 𝑈, and 𝒗 be the vector field generating
(2.8.8) We say that 𝒗 is a symplectic vector field if the diffeomorphisms (2.8.8) are symplec-
tomorphisms, i.e., for all 𝑡 we have 𝑓𝑡⋆ 𝜔 = 𝜔.
Let us see what such vector fields have to look like. Note that by (2.6.23)
𝑑 ⋆
(2.8.9) 𝑓 𝜔 = 𝑓𝑡⋆ 𝐿𝒗 𝜔 ,
𝑑𝑡 𝑡
hence if 𝑓𝑡⋆ 𝜔 = 𝜔 for all 𝑡, the left hand side of (2.8.9) is zero, so
𝑓𝑡⋆ 𝐿𝒗 𝜔 = 0 .
In particular, for 𝑡 = 0, 𝑓𝑡 is the identity map so 𝑓𝑡⋆ 𝐿𝒗 𝜔 = 𝐿𝒗 𝜔 = 0. Conversely, if 𝐿𝒗 𝜔 = 0,
then 𝑓𝑡⋆ 𝐿𝒗 𝜔 = 0 so by (2.8.9) 𝑓𝑡⋆ 𝜔 does not depend on 𝑡. However, since 𝑓𝑡⋆ 𝜔 = 𝜔 for 𝑡 = 0
we conclude that 𝑓𝑡⋆ 𝜔 = 𝜔 for all 𝑡. To summarize, we have proved:
Theorem 2.8.10. Let 𝑓𝑡 ∶ 𝑈 → 𝑈 be a one-parameter group of diffeomorphisms and 𝒗 the
infinitesmal generator of this group. Then 𝒗 is symplectic of and only if 𝐿𝒗 𝜔 = 0.
There is an equivalent formulation of Theorem 2.8.10 in terms of the interior product
𝜄𝒗 𝜔. By (2.5.9)
𝐿𝒗 𝜔 = 𝑑𝜄𝒗 𝜔 + 𝜄𝒗 𝑑𝜔 .
But by (2.8.2) we have 𝑑𝜔 = 0, so
𝐿𝒗 𝜔 = 𝑑𝜄𝒗 𝜔 .
Thus we have shown:
Theorem 2.8.11. A vector field 𝒗 on an open subset of 𝐑2𝑛 is symplectic if and only if the form
𝜄𝒗 𝜔 is closed.
 Draft: March 28, 2018

§2.8 Symplectic geometry & classical mechanics 65

Let 𝑈 be an open subset of 𝐑2𝑛 and 𝒗 a vector field on 𝑈. If 𝜄𝒗 𝜔 is not only closed but is
exact we say that 𝒗 is a Hamiltonian vector field. In other words, 𝒗 is Hamiltonian if
(2.8.12) 𝜄𝒗 𝜔 = 𝑑𝐻
∞ ∞
for some 𝐶 function 𝐻 ∈ 𝐶 (𝑈).
Let is see what this condition looks like in coordinates. Let
𝑛
𝜕 𝜕
(2.8.13) 𝒗 = ∑ (𝑓𝑖 + 𝑔𝑖 ).
𝑖=1 𝜕𝑥𝑖 𝜕𝑦𝑖
Then
𝜄𝒗 𝜔 = ∑ 𝑓𝑖 𝜄𝜕/𝜕𝑥𝑖 (𝑑𝑥𝑗 ∧ 𝑑𝑦𝑗 ) + ∑ 𝑔𝑖 𝜄𝜕/𝜕𝑦𝑖 (𝑑𝑥𝑗 ∧ 𝑑𝑦𝑖 ) .
1≤𝑖,𝑗≤𝑛 1≤𝑖,𝑗≤𝑛
But
1, 𝑖 = 𝑗
𝜄𝜕/𝜕𝑥𝑖 𝑑𝑥𝑗 = {
0, 𝑖 ≠ 𝑗
and
𝜄𝜕/𝜕𝑥𝑖 𝑑𝑦𝑗 = 0
𝑛
so the first summand above is ∑𝑖=1 𝑓𝑖
𝑑𝑦𝑖 , and a similar argument shows that the second
𝑛
summand is − ∑𝑖=1 𝑔𝑖 𝑑𝑥𝑖 .
Hence if 𝒗 is the vector field (2.8.13), then
𝑛
(2.8.14) 𝜄𝒗 𝜔 = ∑(𝑓𝑖 𝑑𝑦𝑖 − 𝑔𝑖 𝑑𝑥𝑖 ) .
𝑖=1
Thus since
𝑛
𝜕𝐻 𝜕𝐻
𝑑𝐻 = ∑ ( 𝑑𝑥 + 𝑑𝑦 )
𝑖=1 𝜕𝑥𝑖 𝑖 𝜕𝑦𝑖 𝑖
we get from (2.8.12)–(2.8.14)
𝜕𝐻 𝜕𝐻
𝑓𝑖 = and 𝑔𝑖 = −
𝜕𝑦𝑖 𝜕𝑥𝑖
so 𝒗 has the form:
𝑛
𝜕𝐻 𝜕 𝜕𝐻 𝜕
(2.8.15) 𝒗 = ∑( − ).
𝑖=1 𝜕𝑦𝑖 𝜕𝑥𝑖 𝜕𝑥𝑖 𝜕𝑦𝑖
In particular if 𝛾(𝑡) = (𝑥(𝑡), 𝑦(𝑡)) is an integral curve of 𝒗 it has to satisfy the system of
differential equations
𝑑𝑥𝑖 𝜕𝐻
{
{ = (𝑥(𝑡), 𝑦(𝑡))
{ 𝑑𝑡 𝜕𝑦𝑖
(2.8.16) {
{ 𝑑𝑦𝑖 = − 𝜕𝐻 (𝑥(𝑡), 𝑦(𝑡)) .
{
{ 𝑑𝑡 𝜕𝑥𝑖
The formulas (2.8.13) and (2.8.14) exhibit an important property of the Darboux form 𝜔.
Every one-form on 𝑈 can be written uniquely as a sum
𝑛
∑ (𝑓𝑖 𝑑𝑦𝑖 − 𝑔𝑖 𝑑𝑥𝑖 )
𝑖=1
with 𝑓𝑖 and 𝑔𝑖 in 𝐶∞ (𝑈) and hence (2.8.13) and (2.8.14) imply:
Theorem 2.8.17. The map 𝒗 ↦ 𝜄𝒗 𝜔 defines a bijective between vector field and one-forms.
 Draft: March 28, 2018

66 Chapter 2: Differential Forms

In particular, for every 𝐶∞ function 𝐻, we get by correspondence a unique vector field

𝒗 = 𝒗𝐻 with the property (2.8.12). We next note that by (1.7.9)
𝐿𝒗 𝐻 = 𝜄𝒗 𝑑𝐻 = 𝜄𝒗 (𝜄𝒗 𝜔) = 0 .
Thus
(2.8.18) 𝐿𝒗 𝐻 = 0
i.e., 𝐻 is an integral of motion of the vector field 𝒗. In particular, if the function 𝐻 ∶ 𝑈 → 𝐑 is
proper, then by Theorem 2.2.12 the vector field, 𝒗, is complete and hence by Theorem 2.8.10
generates a one-parameter group of symplectomorphisms.
Remark 2.8.19. If the one-parameter group (2.8.8) is a group of symplectomorphisms, then
𝑓𝑡⋆ 𝜔𝑛 = 𝑓𝑡⋆ 𝜔 ∧ ⋯ ∧ 𝑓𝑡⋆ 𝜔 = 𝜔𝑛 , so by (2.8.4)
(2.8.20) 𝑓𝑡⋆ 𝛺 = 𝛺 ,
where 𝛺 is the symplectic volume form (2.8.5).
Application 2.8.21. The application we want to make of these ideas concerns the descrip-
tion, in Newtonian mechanics, of a physical system consisting of 𝑁 interacting point-masses.
The configuration space of such a system is
𝐑𝑛 = 𝐑3 × ⋯ × 𝐑3 ,
where there are 𝑁 copies of 𝐑3 in the product, with position coordinates 𝑥1 , …, 𝑥𝑛 and
the phase space is 𝐑2𝑛 with position coordinates 𝑥1 , …, 𝑥𝑛 and momentum coordinates,
𝑦1 , …, 𝑦𝑛 . The kinetic energy of this system is a quadratic function of the momentum co-
ordinates
𝑛
1 1
∑ 𝑦𝑖2 ,
2 𝑖=1 𝑚𝑖
and for simplicity we assume that the potential energy is a function 𝑉(𝑥1 , …, 𝑥𝑛 ) of the
position coordinates alone, i.e., it does not depend on the momenta and is time-independent
as well. Let
𝑛
1 1
(2.8.22) 𝐻= ∑ 𝑦2 + 𝑉(𝑥1 , …, 𝑥𝑛 )
2 𝑖=1 𝑚𝑖 𝑖
be the total energy of the system.
We now show that Newton’s second law of motion in classical mechanics reduces to the
assertion.
Proposition 2.8.23. The trajectories in phase space of the system above are just the integral
curves of the Hamiltonian vector field 𝒗𝐻 .
Proof. For the function (2.8.22) the equations (2.8.16) become
𝑑𝑥𝑖 1
{
{ = 𝑦𝑖
{
{ 𝑑𝑡 𝑚 𝑖
{
{
{ 𝑑𝑦𝑖 = − 𝜕𝑉 .
{
{ 𝑑𝑡 𝜕𝑥𝑖
The first set of equation are essentially just the definitions of momentum, however, if we
plug them into the second set of equations we get
𝑑2 𝑥𝑖 𝜕𝑉
(2.8.24) 𝑚𝑖 =−
𝑑𝑡2 𝜕𝑥𝑖
 Draft: March 28, 2018

§2.8 Symplectic geometry & classical mechanics 67

and interpreting the term on the right as the force exerted on the 𝑖th point-mass and the
term on the left as mass times acceleration this equation becomes Newton’s second law. □
In classical mechanics the equations (2.8.16) are known as the Hamilton–Jacobi equa-
tions. For a more detailed account of their role in classical mechanics we highly recommend
[1]. Historically these equations came up for the first time, not in Newtonian mechanics,
but in geometric optics and a brief description of their origins there and of their relation to
Maxwell’s equations can be found in [5].
We conclude this chapter by mentioning a few implications of the Hamiltonian descrip-
tion (2.8.16) of Newton’s equations (2.8.24).
(1) Conservation of energy: By (2.8.18) the energy function (2.8.22) is constant along the
integral curves of 𝒗, hence the energy of the system (2.8.16) does notchange in time.
(2) Noether’s principle: Let 𝛾𝑡 ∶ 𝐑2𝑛 → 𝐑2𝑛 be a one-parameter group of diffeomorphisms of
phase space and 𝒘 its infinitesimal generator. Then (𝛾𝑡 )𝑡∈𝐑 is a symmetry of the system
above if each 𝛾𝑡 preserves the function (2.8.22) and the vector field 𝒘 is Hamiltonian.
The Hamiltonian condition means that
𝜄𝒘 𝜔 = 𝑑𝐺
∞
for some 𝐶 function 𝐺, and what Noether’s principle asserts is that this function is an
integral of motion of the system (2.8.16), i.e., satisfies 𝐿𝒗 𝐺 = 0. In other words stated
more succinctly: symmetries of the system (2.8.16) give rise to integrals of motion.
(3) Poincaré recurrence: An important theorem of Poincaré asserts that if the function
𝐻 ∶ 𝐑2𝑛 → 𝐑
defined by (2.8.22) is proper then every trajectory of the system (2.8.16) returns arbi-
trarily close to its initial position at some positive time 𝑡0 , and, in fact, does this not just
once but does so infinitely often. We sketch a proof of this theorem, using (2.8.20), in
the next chapter.

Exercises for §2.8

Exercise 2.8.i. Let 𝒗𝐻 be the vector field (2.8.15). Prove that div(𝒗𝐻 ) = 0.
Exercise 2.8.ii. Let 𝑈 be an open subset of 𝐑𝑚 , 𝑓𝑡 ∶ 𝑈 → 𝑈 a one-parameter group of
diffeomorphisms of 𝑈 and 𝒗 the infinitesmal generator of this group. Show that if 𝛼 is a 𝑘-
form on 𝑈 then 𝑓𝑡⋆ 𝛼 = 𝛼 for all 𝑡 if and only if 𝐿𝒗 𝛼 = 0 (i.e., generalize to arbitrary 𝑘-forms
the result we proved above for the Darboux form).
𝑛
Exercise 2.8.iii (the harmonic oscillator). Let 𝐻 be the function ∑𝑖=1 𝑚𝑖 (𝑥𝑖2 + 𝑦𝑖2 ) where the
𝑚𝑖 ’s are positive constants.
(1) Compute the integral curves of 𝒗𝐻 .
(2) Poincaré recurrence: Show that if (𝑥(𝑡), 𝑦(𝑡)) is an integral curve with initial point (𝑥0 , 𝑦0 ) ≔
(𝑥(0), 𝑦(0)) and 𝑈 an arbitrarily small neighborhood of (𝑥0 , 𝑦0 ), then for every 𝑐 > 0
there exists a 𝑡 > 𝑐 such that (𝑥(𝑡), 𝑦(𝑡)) ∈ 𝑈.
Exercise 2.8.iv. Let 𝑈 be an open subset of 𝐑2𝑛 and let 𝐻1 , 𝐻2 ∈ 𝐶∞ (𝑈). Show that
[𝒗𝐻1 , 𝒗𝐻2 ] = 𝒗𝐻
where
𝑛
𝜕𝐻1 𝜕𝐻2 𝜕𝐻2 𝜕𝐻1
(2.8.25) 𝐻=∑ − .
𝑖=1 𝜕𝑥𝑖 𝜕𝑦𝑖 𝜕𝑥𝑖 𝜕𝑦𝑖
 Draft: March 28, 2018

68 Chapter 2: Differential Forms

Exercise 2.8.v. The expression (2.8.25) is known as the Poisson bracket of 𝐻1 and 𝐻2 , de-
noted by {𝐻1 , 𝐻2 }. Show that it is anti-symmetric
{𝐻1 , 𝐻2 } = −{𝐻2 , 𝐻1 }
and satisfies the Jacobi identity
0 = {𝐻1 , {𝐻2 , 𝐻3 }} + {𝐻2 , {𝐻3 , 𝐻1 }} + {𝐻3 , {𝐻1 , 𝐻2 }} .
Exercise 2.8.vi. Show that
{𝐻1 , 𝐻2 } = 𝐿𝒗𝐻 𝐻2 = −𝐿𝒗𝐻 𝐻1 .
1 2

Exercise 2.8.vii. Prove that the following three properties are equivalent.
(1) {𝐻1 , 𝐻2 } = 0.
(2) 𝐻1 is an integral of motion of 𝒗2 .
(3) 𝐻2 is an integral of motion of 𝒗1 .
Exercise 2.8.viii. Verify Noether’s principle.
Exercise 2.8.ix (conservation of linear momentum). Suppose the potential 𝑉 in equation (2.8.22)
is invariant under the one-parameter group of translations
𝑇𝑡 (𝑥1 , …, 𝑥𝑛 ) = (𝑥1 + 𝑡, …, 𝑥𝑛 + 𝑡) .
(1) Show that the function (2.8.22) is invariant under the group of diffeomorphisms
𝛾𝑡 (𝑥, 𝑦) = (𝑇𝑡 𝑥, 𝑦) .
(2) Show that the infinitesmal generator of this group is the Hamiltonian vector field 𝒗𝐺
𝑛
where 𝐺 = ∑𝑖=1 𝑦𝑖 .
(3) Conclude from Noether’s principle that this function is an integral of the vector field
𝒗𝐻 , i.e., that “total linear moment” is conserved.
(4) Show that “total linear momentum” is conserved if 𝑉 is the Coulomb potential
𝑚𝑖
∑ .
𝑖≠𝑗 |𝑥𝑖 − 𝑥𝑗 |

Exercise 2.8.x. Let 𝑅𝑖𝑡 ∶ 𝐑2𝑛 → 𝐑2𝑛 be the rotation which fixes the variables, (𝑥𝑘 , 𝑦𝑘 ), 𝑘 ≠ 𝑖
and rotates (𝑥𝑖 , 𝑦𝑖 ) by the angle, 𝑡:
𝑅𝑖𝑡 (𝑥𝑖 , 𝑦𝑖 ) = (cos 𝑡 𝑥𝑖 + sin 𝑡 𝑦𝑖 ,− sin 𝑡 𝑥𝑖 + cos 𝑡 𝑦𝑖 ) .
(1) Show that 𝑅𝑖𝑡 , −∞ < 𝑡 < ∞, is a one-parameter group of symplectomorphisms.
(2) Show that its generator is the Hamiltonian vector field, 𝒗𝐻𝑖 , where 𝐻𝑖 = (𝑥𝑖2 + 𝑦𝑖2 )/2.
𝑗
(3) Let 𝐻 be the harmonic oscillator Hamiltonian from Exercise 2.8.iii. Show that the 𝑅𝑡 ’s
preserve 𝐻.
(4) What does Noether’s principle tell one about the classical mechanical system with en-
ergy function 𝐻?
Exercise 2.8.xi. Show that if 𝑈 is an open subset of 𝐑2𝑛 and 𝒗 is a symplectic vector field on 𝑈
then for every point 𝑝0 ∈ 𝑈 there exists a neighborhood 𝑈0 of 𝑝0 on which 𝒗 is Hamiltonian.
Exercise 2.8.xii. Deduce from Exercises 2.8.iv and 2.8.xi that if 𝒗1 and 𝒗2 are symplectic
vector fields on an open subset 𝑈 of 𝐑2𝑛 their Lie bracket [𝒗1 , 𝒗2 ] is a Hamiltonian vector
field.
𝑛
Exercise 2.8.xiii. Let 𝛼 = ∑𝑖=1 𝑦𝑖 𝑑𝑥𝑖 .
(1) Show that 𝜔 = −𝑑𝛼.
 Draft: March 28, 2018

§2.8 Symplectic geometry & classical mechanics 69

(2) Show that if 𝛼1 is any one-form on 𝐑2𝑛 with the property 𝜔 = −𝑑𝛼1 , then
𝛼 = 𝛼1 + 𝑑𝐹
∞
for some 𝐶 function 𝐹.
(3) Show that 𝛼 = 𝜄𝒘 𝜔 where 𝒘 is the vector field
𝑛
𝜕
𝒘 ≔ − ∑ 𝑦𝑖 .
𝑖=1 𝜕𝑦𝑖
Exercise 2.8.xiv. Let 𝑈 be an open subset of 𝐑2𝑛 and 𝒗 a vector field on 𝑈. Show that 𝒗 has
the property, 𝐿𝒗 𝛼 = 0, if and only if
𝜄𝒗 𝜔 = 𝑑𝜄𝒗 𝛼 .
In particular conclude that if 𝐿𝒗 𝛼 = 0 then 𝑣 is Hamiltonian.
Hint: Equation (2.8.2).
Exercise 2.8.xv. Let 𝐻 be the function
𝑛
(2.8.26) 𝐻(𝑥, 𝑦) = ∑ 𝑓𝑖 (𝑥)𝑦𝑖 ,
𝑖=1
where the 𝑓𝑖 ’s are 𝐶∞ functions on 𝐑𝑛 . Show that
(2.8.27) 𝐿𝒗𝐻 𝛼 = 0 .
Exercise 2.8.xvi. Conversely show that if 𝐻 is any 𝐶∞ function on 𝐑2𝑛 satisfying equa-
tion (2.8.27) it has to be a function of the form (2.8.26).
Hints:
(1) Let 𝒗 be a vector field on 𝐑2𝑛 satisfying 𝐿𝒗 𝛼 = 0. By equation (2.8.16) we have 𝒗 = 𝒗𝐻 ,
where 𝐻 = 𝜄𝒗 𝛼.
(2) Show that 𝐻 has to satisfy the equation
𝑛
𝜕𝐻
∑ 𝑦𝑖 =𝐻.
𝑖=1 𝜕𝑦𝑖
𝜕𝐻
(3) Conclude that if 𝐻𝑟 = 𝜕𝑦𝑟
then 𝐻𝑟 has to satisfy the equation
𝑛
𝜕
∑ 𝑦𝑖 𝐻 =0.
𝑖=1 𝜕𝑦𝑖 𝑟
(4) Conclude that 𝐻𝑟 has to be constant along the rays (𝑥, 𝑡𝑦), for 0 ≤ 𝑡 < ∞.
(5) Conclude finally that 𝐻𝑟 has to be a function of 𝑥 alone, i.e., does not depend on 𝑦.
Exercise 2.8.xvii. Show that if 𝒗𝐑𝑛 is a vector field
𝑛
𝜕
∑ 𝑓𝑖 (𝑥)
𝑖=1 𝜕𝑥𝑖
on configuration space there is a unique lift of 𝒗𝐑𝑛 to phase space
𝑛
𝜕 𝜕
𝒗 = ∑ 𝑓𝑖 (𝑥) + 𝑔𝑖 (𝑥, 𝑦)
𝑖=1 𝜕𝑥𝑖 𝜕𝑦𝑖
satisfying 𝐿𝒗 𝛼 = 0.
Draft: March 28, 2018
 Draft: March 28, 2018

CHAPTER 3

Integration of Forms

3.1. Introduction
The change of variables formula asserts that if 𝑈 and 𝑉 are open subsets of 𝐑𝑛 and
𝑓 ∶ 𝑈 → 𝑉 a 𝐶1 diffeomorphism then, for every continuous function 𝜙 ∶ 𝑉 → 𝐑 the integral

∫ 𝜙(𝑦)𝑑𝑦
𝑉
exists if and only if the integral

∫ (𝜙 ∘ 𝑓)(𝑥) |det 𝐷𝑓(𝑥)|𝑑𝑥

𝑈
exists, and if these integrals exist they are equal. Proofs of this can be found in [10] or [12].
This chapter contains an alternative proof of this result. This proof is due to Peter Lax. Our
version of his proof in § 3.5 below makes use of the theory of differential forms; but, as
Lax shows in the article [8] (which we strongly recommend as collateral reading for this
course), references to differential forms can be avoided, and the proof described in §3.5 can
be couched entirely in the language of elementary multivariable calculus.
The virtue of Lax’s proof is that is allows one to prove a version of the change of variables
theorem for other mappings besides diffeomorphisms, and involves a topological invariant,
the degree of a map, which is itself quite interesting. Some properties of this invariant, and
some topological applications of the change of variables formula will be discussed in §3.6
of these notes.
Remark 3.1.1. The proof we are about to describe is somewhat simpler and more transpar-
ent if we assume that 𝑓 is a 𝐶∞ diffeomorphism. We’ll henceforth make this assumption.

3.2. The Poincaré lemma for compactly supported forms on rectangles

Definition 3.2.1. Let 𝜈 be a 𝑘-form on 𝐑𝑛 . We define the support of 𝜈 by
supp(𝜈) ≔ { 𝑥 ∈ 𝐑𝑛 | 𝜈𝑥 ≠ 0 } ,
and we say that 𝜈 is compactly supported if supp(𝜈) is compact.
We will denote by 𝛺𝑘𝑐 (𝐑𝑛 ) the set of all 𝐶∞ 𝑘-forms which are compactly supported,
and if 𝑈 is an open subset of 𝐑𝑛 , we will denote by 𝛺𝑘𝑐 (𝑈) the set of all compactly supported
𝑘-forms whose support is contained in 𝑈.
Let 𝜔 = 𝑓 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 be a compactly supported 𝑛-form with 𝑓 ∈ 𝐶0∞ (𝐑𝑛 ). We will
define the integral of 𝜔 over 𝐑𝑛 :
∫ 𝜔
𝐑𝑛
to be the usual integral of 𝑓 over 𝐑𝑛
∫ 𝑓 𝑑𝑥 .
𝐑𝑛
71
 Draft: March 28, 2018

72 Chapter 3: Integration of Forms

(Since 𝑓 is 𝐶∞ and compactly supported this integral is well-defined.)

Now let 𝑄 be the rectangle
[𝑎1 , 𝑏1 ] × ⋯ × [𝑎𝑛 , 𝑏𝑛 ] .
The Poincaré lemma for rectangles asserts:
Theorem 3.2.2 (Poincaré lemma for rectangles). Let 𝜔 be a compactly supported 𝑛-form with
supp(𝜔) ⊂ int(𝑄). Then the following assertions are equivalent.
(1) ∫ 𝜔 = 0.
(2) There exists a compactly supported (𝑛 − 1)-form 𝜇 with supp(𝜇) ⊂ int(𝑄) satisfying 𝑑𝜇 =
𝜔.
We will first prove that (2)⇒(1).
Proof that (2)⇒(1). Let
𝑛
̂ 𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛 ,
𝜇 = ∑ 𝑓𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑖=1

(the “hat” over the 𝑑𝑥𝑖 meaning that 𝑑𝑥𝑖 has to be omitted from the wedge product). Then
𝑛
𝜕𝑓𝑖
𝑑𝜇 = ∑(−1)𝑖−1 𝑑𝑥 ∧ ⋯ ∧ 𝑑𝑥𝑛 ,
𝑖=1 𝜕𝑥𝑖 1
and to show that the integral of 𝑑𝜇 is zero it suffices to show that each of the integrals
𝜕𝑓𝑖
(3.2.2)𝑖 ∫ 𝑑𝑥
𝐑𝑛 𝜕𝑥𝑖
is zero. By Fubini we can compute (3.2.2)𝑖 by first integrating with respect to the variable,
𝑥𝑖 , and then with respect to the remaining variables. But
𝜕𝑓𝑖 𝑥𝑖 =𝑏𝑖
∫ 𝑑𝑥𝑖 = 𝑓(𝑥)| =0
𝜕𝑥𝑖 𝑥𝑖 =𝑎𝑖

since 𝑓𝑖 is supported on 𝑈. □
We will prove that (1)⇒(2) by proving a somewhat stronger result. Let 𝑈 be an open
subset of 𝐑𝑚 . We’ll say that 𝑈 has property 𝑃 if every form 𝜔 ∈ 𝛺𝑚
𝑐 (𝑈) such that ∫𝑈 𝜔 = 0
we have 𝜔 ∈ 𝑑 𝛺𝑚−1
𝑐 (𝑈). We will prove:
Theorem 3.2.3. Let 𝑈 be an open subset of 𝐑𝑛−1 and 𝐴 ⊂ 𝐑 an open interval. Then if 𝑈 has
property 𝑃, 𝑈 × 𝐴 does as well.
Remark 3.2.4. It is very easy to see that the open interval 𝐴 itself has property 𝑃. (See
Exercise 3.2.i below.) Hence it follows by induction from Theorem 3.2.3 that
int 𝑄 = 𝐴 1 × ⋯ × 𝐴 𝑛 , 𝐴 𝑖 = (𝑎𝑖 , 𝑏𝑖 )
has property 𝑃, and this proves “(1)⇒(2)”.
Proof of Theorem 3.2.3. Let (𝑥, 𝑡) = (𝑥1 , …, 𝑥𝑛−1 , 𝑡) be product coordinates on 𝑈 × 𝐴. Given
𝜔 ∈ 𝛺𝑛𝑐 (𝑈 × 𝐴) we can express 𝜔 as a wedge product, 𝑑𝑡 ∧ 𝛼 with 𝛼 = 𝑓(𝑥, 𝑡) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛−1
and 𝑓 ∈ 𝐶0∞ (𝑈 × 𝐴). Let 𝜃 ∈ 𝛺𝑛−1
𝑐 (𝑈) be the form

(3.2.5) 𝜃 = (∫ 𝑓(𝑥, 𝑡)𝑑𝑡) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛−1 .

𝐴
 Draft: March 28, 2018

§3.2 The Poincaré lemma for compactly supported forms on rectangles 73

Then
∫ 𝜃 = ∫ 𝑓(𝑥, 𝑡)𝑑𝑥𝑑𝑡 = ∫ 𝜔
𝐑𝑛−1 𝐑𝑛 𝐑𝑛
so if the integral of 𝜔 is zero, the integral of 𝜃 is zero. Hence since 𝑈 has property 𝑃, 𝜃 = 𝑑𝜈
for some 𝜈 ∈ 𝛺𝑛−2 ∞
𝑐 (𝑈). Let 𝜌 ∈ 𝐶 (𝐑) be a bump function which is supported on 𝐴 and
whose integral over 𝐴 is 1. Setting
𝜅 = −𝜌(𝑡)𝑑𝑡 ∧ 𝜈
we have
𝑑𝜅 = 𝜌(𝑡)𝑑𝑡 ∧ 𝑑𝜈 = 𝜌(𝑡)𝑑𝑡 ∧ 𝜃 ,
and hence
𝜔 − 𝑑𝜅 = 𝑑𝑡 ∧ (𝛼 − 𝜌(𝑡)𝜃)
= 𝑑𝑡 ∧ 𝑢(𝑥, 𝑡)𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛−1
where
𝑢(𝑥, 𝑡) = 𝑓(𝑥, 𝑡) − 𝜌(𝑡) ∫ 𝑓(𝑥, 𝑡)𝑑𝑡
𝐴
by (3.2.5). Thus

(3.2.6) ∫ 𝑢(𝑥, 𝑡) 𝑑𝑡 = 0 .

Let 𝑎 and 𝑏 be the end points of 𝐴 and let

𝑡
(3.2.7) 𝑣(𝑥, 𝑡) = ∫ 𝑢(𝑥, 𝑠) 𝑑𝑠 .
𝑎
By equation (3.2.6) 𝑣(𝑎, 𝑥) = 𝑣(𝑏, 𝑥) = 0, so 𝑣 is in 𝐶0∞ (𝑈 × 𝐴) and by (3.2.7), 𝜕𝑣/𝜕𝑡 = 𝑢.
Hence if we let 𝛾 be the form, 𝑣(𝑥, 𝑡) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛−1 , we have:
𝑑𝛾 = 𝑑𝑡 ∧ 𝑢(𝑥, 𝑡) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛−1 = 𝜔 − 𝑑𝜅
and
𝜔 = 𝑑(𝛾 + 𝜅) .
Since 𝛾 and 𝜅 are both in 𝛺𝑛−1 𝑛−1
𝑐 (𝑈 × 𝐴) this proves that 𝜔 is in 𝑑 𝛺𝑐 (𝑈 × 𝐴) and hence
that 𝑈 × 𝐴 has property 𝑃. □
Exercises for §3.2
Exercise 3.2.i. Let 𝑓 ∶ 𝐑 → 𝐑 be a compactly supported function of class 𝐶𝑟 with support
on the interval (𝑎, 𝑏). Show that the following are equivalent.
𝑏
(1) ∫𝑎 𝑓(𝑥)𝑑𝑥 = 0.
𝑑𝑔
(2) There exists a function 𝑔 ∶ 𝐑 → 𝐑 of class 𝐶𝑟+1 with support on (𝑎, 𝑏) with 𝑑𝑥
= 𝑓.
Hint: Show that the function
𝑥
𝑔(𝑥) = ∫ 𝑓(𝑠)𝑑𝑠
𝑎
is compactly supported.
Exercise 3.2.ii. Let 𝑓 = 𝑓(𝑥, 𝑦) be a compactly supported function on 𝐑𝑘 × 𝐑ℓ with the
property that the partial derivatives
𝜕𝑓
(𝑥, 𝑦) , for 𝑖 = 1, …, 𝑘 ,
𝜕𝑥𝑖
 Draft: March 28, 2018

74 Chapter 3: Integration of Forms

exist and are continuous as functions of 𝑥 and 𝑦. Prove the following “differentiation under
the integral sign” theorem (which we implicitly used in our proof of Theorem 3.2.3).
Theorem 3.2.8. The function 𝑔(𝑥) ≔ ∫𝐑ℓ 𝑓(𝑥, 𝑦)𝑑𝑦 is of class 𝐶1 and
𝜕𝑔 𝜕𝑓
(𝑥) = ∫ (𝑥, 𝑦)𝑑𝑦 .
𝜕𝑥𝑖 𝜕𝑥𝑖
Hints: For 𝑦 fixed and ℎ ∈ 𝐑𝑘 ,
𝑓(𝑥 + ℎ, 𝑦) − 𝑓(𝑥, 𝑦) = 𝐷𝑥 𝑓(𝑐, 𝑦)ℎ
for some point 𝑐 on the line segment joining 𝑥 to 𝑥+ℎ. Using the fact that 𝐷𝑥 𝑓 is continuous
as a function of 𝑥 and 𝑦 and compactly supported, conclude the following.
Lemma 3.2.9. Given 𝜀 > 0 there exists a 𝛿 > 0 such that for |ℎ| ≤ 𝛿
|𝑓(𝑥 + ℎ, 𝑦) − 𝑓(𝑥, 𝑦) − 𝐷𝑥 𝑓(𝑥, 𝑦)ℎ| ≤ 𝜀|ℎ| .
Now let 𝑄 ⊂ 𝐑ℓ be a rectangle with supp(𝑓) ⊂ 𝐑𝑘 × 𝑄 and show that

|𝑔(𝑥 + ℎ) − 𝑔(𝑥) − (∫ 𝐷𝑥 𝑓(𝑥, 𝑦) 𝑑𝑦) ℎ| ≤ 𝜀 vol(𝑄)|ℎ| .

Conclude that 𝑔 is differentiable at 𝑥 and that its derivative is

∫ 𝐷𝑥 𝑓(𝑥, 𝑦)𝑑𝑦 .

Exercise 3.2.iii. Let 𝑓 ∶ 𝐑𝑘 × 𝐑ℓ → 𝐑 be a compactly supported continuous function. Prove

the following theorem.
Theorem 3.2.10. If all the partial derivatives of 𝑓(𝑥, 𝑦) with respect to 𝑥 of order ≤ 𝑟 exist
and are continuous as functions of 𝑥 and 𝑦 the function

𝑔(𝑥) = ∫ 𝑓(𝑥, 𝑦)𝑑𝑦

is of class 𝐶𝑟 .
Exercise 3.2.iv. Let 𝑈 be an open subset of 𝐑𝑛−1 , 𝐴 ⊂ 𝐑 an open interval and (𝑥, 𝑡) product
coordinates on 𝑈 × 𝐴. Recall from Exercise 2.3.v that every form 𝜔 ∈ 𝛺𝑘 (𝑈 × 𝐴) can be
written uniquely as a sum 𝜔 = 𝑑𝑡 ∧ 𝛼 + 𝛽 where 𝛼 and 𝛽 are reduced, i.e., do not contain a
factor of 𝑑𝑡.
(1) Show that if 𝜔 is compactly supported on 𝑈 × 𝐴 then so are 𝛼 and 𝛽.
(2) Let 𝛼 = ∑𝐼 𝑓𝐼 (𝑥, 𝑡)𝑑𝑥𝐼 . Show that the form

(3.2.11) 𝜃 = ∑ (∫ 𝑓𝐼 (𝑥, 𝑡) 𝑑𝑡) 𝑑𝑥𝐼

𝐼 𝐴

is in 𝛺𝑘−1
𝑐 (𝑈).
(3) Show that if 𝑑𝜔 = 0, then 𝑑𝜃 = 0.
Hint: By equation (3.2.11),
𝜕𝑓𝐼
𝑑𝜃 = ∑ (∫ (𝑥, 𝑡) 𝑑𝑡) 𝑑𝑥𝑖 ∧ 𝑑𝑥𝐼 = ∫ (𝑑𝑈 𝛼) 𝑑𝑡
𝐼,𝑖 𝐴 𝜕𝑥𝑖 𝐴
𝑑𝛽
and by equation (2.4.19) we have 𝑑𝑈 𝛼 = 𝑑𝑡
.
Exercise 3.2.v. In Exercise 3.2.iv, show that if 𝜃 is in 𝑑𝛺𝑘−2 𝑘−1
𝑐 (𝑈) then 𝜔 is in 𝑑𝛺𝑐 (𝑈 × 𝐴)
as follows.
 Draft: March 28, 2018

§3.2 The Poincaré lemma for compactly supported forms on rectangles 75

(1) Let 𝜃 = 𝑑𝜈, with 𝜈 ∈ 𝛺𝑘−2 ∞

𝑐 (𝑈) and let 𝜌 ∈ 𝐶 (𝐑) be a bump function which is supported
on 𝐴 and whose integral over 𝐴 is one. Setting 𝑘 = −𝜌(𝑡)𝑑𝑡 ∧ 𝜈 show that

𝜔 − 𝑑𝜅 = 𝑑𝑡 ∧ (𝛼 − 𝜌(𝑡)𝜃) + 𝛽
= 𝑑𝑡 ∧ (∑𝐼 𝑢𝐼 (𝑥, 𝑡)𝑑𝑥𝐼 ) + 𝛽 ,

where
𝑢𝐼 (𝑥, 𝑡) = 𝑓𝐼 (𝑥, 𝑡) − 𝜌(𝑡) ∫ 𝑓𝐼 (𝑥, 𝑡)𝑑𝑡 .
𝐴
(2) Let 𝑎 and 𝑏 be the end points of 𝐴 and let
𝑡
𝑣𝐼 (𝑥, 𝑡) = ∫ 𝑢𝐼 (𝑥, 𝑡) 𝑑𝑡 .
𝑎

Show that the form ∑𝐼 𝑣𝐼 (𝑥, 𝑡) 𝑑𝑥𝐼 is in 𝛺𝑘−1

𝑐 (𝑈 × 𝐴) and that

𝑑𝛾 = 𝜔 − 𝑑𝜅 − 𝛽 + 𝑑𝑈 𝛾 .

(3) Conclude that the form 𝜔 − 𝑑(𝜅 + 𝛾) is reduced.

(4) Prove that if 𝜆 ∈ 𝛺𝑘𝑐 (𝑈 × 𝐴) is reduced and 𝑑𝜆 = 0 then 𝜆 = 0.
Hint: Let 𝜆 = ∑𝐼 𝑔𝐼 (𝑥, 𝑡) 𝑑𝑥𝐼 . Show that 𝑑𝜆 = 0 ⇒ 𝜕𝑡𝜕 𝑔𝐼 (𝑥, 𝑡) = 0 and exploit the
fact that for fixed 𝑥, 𝑔𝐼 (𝑥, 𝑡) is compactly supported in 𝑡.

Exercise 3.2.vi. Let 𝑈 be an open subset of 𝐑𝑚 . We say that 𝑈 has property 𝑃𝑘 , for 1 ≤ 𝑘 < 𝑚,
if every closed 𝑘-form 𝜔 ∈ 𝛺𝑘𝑐 (𝑈) is in 𝑑 𝛺𝑘−1
𝑐 (𝑈). Prove that if the open set 𝑈 ⊂ 𝐑
𝑛−1
in
Exercise 3.2.iv has property 𝑃𝑘−1 then 𝑈 × 𝐴 has property 𝑃𝑘 .

Exercise 3.2.vii. Show that if 𝑄 is the rectangle [𝑎1 , 𝑏1 ] × ⋯ × [𝑎𝑛 , 𝑏𝑛 ] and 𝑈 = int 𝑄 then
𝑈 has property 𝑃𝑘 .

Exercise 3.2.viii. Let 𝐇𝑛 be the half-space

(3.2.12) 𝐇𝑛 ≔ { (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐑𝑛 | 𝑥1 ≤ 0 }

and let 𝜔 ∈ 𝛺𝑛𝑐 (𝐑𝑛 ) be the 𝑛-form 𝜔 ≔ 𝑓𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 with 𝑓 ∈ 𝐶0∞ (𝐑𝑛 ). Define

(3.2.13) ∫ 𝜔 ≔ ∫ 𝑓(𝑥1 , …, 𝑥𝑛 )𝑑𝑥1 ⋯ 𝑑𝑥𝑛

𝐇𝑛 𝐇𝑛

where the right hand side is the usual Riemann integral of 𝑓 over 𝐇𝑛 . (This integral makes
sense since 𝑓 is compactly supported.) Show that if 𝜔 = 𝑑𝜇 for some 𝜇 ∈ 𝛺𝑛−1 𝑛
𝑐 (𝐑 ) then

(3.2.14) ∫ 𝜔=∫ 𝜄⋆ 𝜇
𝐇𝑛 𝐑𝑛−1

where 𝜄 ∶ 𝐑𝑛−1 → 𝐑𝑛 is the inclusion map

(𝑥2 , …, 𝑥𝑛 ) ↦ (0, 𝑥2 , …, 𝑥𝑛 ) .
̂ 𝑖 ⋯ ∧ 𝑑𝑥𝑛 . Mimicking the “(2)⇒(1)” part of the proof
Hint: Let 𝜇 = ∑𝑖 𝑓𝑖 𝑑𝑥1 ∧ ⋯ 𝑑𝑥
of Theorem 3.2.2 show that the integral (3.2.13) is the integral over 𝐑𝑛−1 of the function
0
𝜕𝑓1
∫ (𝑥 , 𝑥 , …, 𝑥𝑛 )𝑑𝑥1 .
−∞ 𝜕𝑥1 1 2
 Draft: March 28, 2018

76 Chapter 3: Integration of Forms

3.3. The Poincaré lemma for compactly supported forms on open subsets of 𝐑𝑛
In this section we will generalize Theorem 3.2.2 to arbitrary connected open subsets of
𝐑𝑛 .
Theorem 3.3.1 (Poincaré lemma for compactly supported forms). Let 𝑈 be a connected open
subset of 𝐑𝑛 and let 𝜔 be a compactly supported 𝑛-form with supp(𝜔) ⊂ 𝑈. The the following
assertions are equivalent:
(1) ∫𝐑𝑛 𝜔 = 0.
(2) There exists a compactly supported (𝑛 − 1)-form 𝜇 with supp 𝜇 ⊂ 𝑈 and 𝜔 = 𝑑𝜇.
Proof that (2)⇒(1). The support of 𝜇 is contained in a large rectangle, so the integral of 𝑑𝜇
is zero by Theorem 3.2.2. □
Proof that (1)⇒(2). Let 𝜔1 and 𝜔2 be compactly supported 𝑛-forms with support in 𝑈. We
will write
𝜔1 ∼ 𝜔2
as shorthand notation for the statement: There exists a compactly supported (𝑛 − 1)-form,
𝜇, with support in 𝑈 and with 𝜔1 − 𝜔2 = 𝑑𝜇. We will prove that (1)⇒(2) by proving an
equivalent statement: Fix a rectangle, 𝑄0 ⊂ 𝑈 and an 𝑛-form, 𝜔0 , with supp 𝜔0 ⊂ 𝑄0 and
integral equal to one.
Theorem 3.3.2. If 𝜔 is a compactly supported 𝑛-form with supp(𝜔) ⊂ 𝑈 and 𝑐 = ∫ 𝜔 then
𝜔 ∼ 𝑐𝜔0 .
Thus in particular if 𝑐 = 0, Theorem 3.3.2 says that 𝜔 ∼ 0 proving that (1)⇒(2). □
To prove Theorem 3.3.2 let 𝑄𝑖 ⊂ 𝑈, 𝑖 = 1, 2, 3, …, be a collection of rectangles with
∞
𝑈 = ⋃𝑖=1 int(𝑄𝑖 ) and let 𝜙𝑖 be a partition of unity with supp(𝜙𝑖 ) ⊂ int(𝑄𝑖 ). Replacing 𝜔
𝑚
by the finite sum ∑𝑖=1 𝜙𝑖 𝜔 for 𝑚 large, it suffices to prove Theorem 3.3.2 for each of the
summands 𝜙𝑖 𝜔. In other words we can assume that supp(𝜔) is contained in one of the open
rectangles int(𝑄𝑖 ). Denote this rectangle by 𝑄. We claim that one can join 𝑄0 to 𝑄 by a
sequence of rectangles as in the figure below.

𝑄0

Figure 3.3.1. A sequence of rectangles joining the rectangles 𝑄0 and 𝑄

Lemma 3.3.3. There exists a sequence of rectangles 𝑅0 , …, 𝑅𝑁+1 such that 𝑅0 = 𝑄0 , 𝑅𝑁+1 = 𝑄
and int(𝑅𝑖 ) ∩ int(𝑅𝑖+1 ) is nonempty.
Proof. Denote by 𝐴 the set of points, 𝑥 ∈ 𝑈, for which there exists a sequence of rectangles,
𝑅𝑖 , 𝑖 = 0, …, 𝑁 + 1 with 𝑅0 = 𝑄0 , with 𝑥 ∈ int 𝑅𝑁+1 and with int 𝑅𝑖 ∩ int 𝑅𝑖+1 nonempty.
It is clear that this set is open and that its complement is open; so, by the connectivity of 𝑈,
𝑈 = 𝐴. □
 Draft: March 28, 2018

§3.4 The degree of a differentiable mapping 77

To prove Theorem 3.3.2 with supp 𝜔 ⊂ 𝑄, select, for each 𝑖, a compactly supported
𝑛-form 𝜈𝑖 with supp(𝜈𝑖 ) ⊂ int(𝑅𝑖 ) ∩ int(𝑅𝑖+1 ) and with ∫ 𝜈𝑖 = 1. The difference, 𝜈𝑖 − 𝜈𝑖+1
is supported in int 𝑅𝑖+1 , and its integral is zero. So by Theorem 3.2.2, 𝜈𝑖 ∼ 𝜈𝑖+1 . Similarly,
𝜔0 ∼ 𝜈0 and, setting 𝑐 ≔ ∫ 𝜔, we have 𝜔 ∼ 𝑐𝜈𝑁 . Thus
𝑐𝜔0 ∼ 𝑐𝜈0 ∼ ⋯ ∼ 𝑐𝜈𝑁 = 𝜔
proving the theorem.

3.4. The degree of a differentiable mapping

Definition 3.4.1. Let 𝑈 and 𝑉 be open subsets of 𝐑𝑛 and 𝐑𝑘 . A continuous map 𝑓 ∶ 𝑈 → 𝑉,
is proper if for every compact subset 𝐾 ⊂ 𝑉, the preimage 𝑓−1 (𝐾) is compact in 𝑈.
Proper mappings have a number of nice properties which will be investigated in the
exercises below. One obvious property is that if 𝑓 is a 𝐶∞ mapping and 𝜔 is a compactly
supported 𝑘-form with support on 𝑉, 𝑓⋆ 𝜔 is a compactly supported 𝑘-form with support
on 𝑈. Our goal in this section is to show that if 𝑈 and 𝑉 are connected open subsets of 𝐑𝑛
and 𝑓 ∶ 𝑈 → 𝑉 is a proper 𝐶∞ mapping then there exists a topological invariant of 𝑓, which
we will call its degree (and denote by deg(𝑓)), such that the “change of variables” formula:

(3.4.2) ∫ 𝑓⋆ 𝜔 = deg(𝑓) ∫ 𝜔
𝑈 𝑉
holds for all 𝜔 ∈ 𝛺𝑛𝑐 (𝑉).
Before we prove this assertion let’s see what this formula says in coordinates. If
𝜔 = 𝜙(𝑦)𝑑𝑦1 ∧ ⋯ ∧ 𝑑𝑦𝑛
then at 𝑥 ∈ 𝑈
𝑓⋆ 𝜔 = (𝜙 ∘ 𝑓)(𝑥) det(𝐷𝑓(𝑥))𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
Hence, in coordinates, equation (3.4.2) takes the form

(3.4.3) ∫ 𝜙(𝑦)𝑑𝑦 = deg(𝑓) ∫ 𝜙 ∘ 𝑓(𝑥) det(𝐷𝑓(𝑥))𝑑𝑥 .

𝑉 𝑈

Proof of equation (3.4.2). Let 𝜔0 be a compactly-supported 𝑛-form with supp(𝜔0 ) ⊂ 𝑉 and

with ∫ 𝜔0 = 1. If we set deg 𝑓 ≔ ∫𝑈 𝑓⋆ 𝜔0 then (3.4.2) clearly holds for 𝜔0 . We will prove
that (3.4.2) holds for every compactly supported 𝑛-form 𝜔 with supp(𝜔) ⊂ 𝑉. Let 𝑐 ≔ ∫𝑉 𝜔.
Then by Theorem 3.3.1 𝜔 − 𝑐𝜔0 = 𝑑𝜇, where 𝜇 is a compactly supported (𝑛 − 1)-form with
supp 𝜇 ⊂ 𝑉. Hence
𝑓⋆ 𝜔 − 𝑐𝑓⋆ 𝜔0 = 𝑓⋆ 𝑑𝜇 = 𝑑𝑓⋆ 𝜇 ,
and by part (1) of Theorem 3.3.1

∫ 𝑓⋆ 𝜔 = 𝑐 ∫ 𝑓⋆ 𝜔0 = deg(𝑓) ∫ 𝜔 . □
𝑈 𝑉

We will show in § 3.6 that the degree of 𝑓 is always an integer and explain why it is a
“topological” invariant of 𝑓.
Proposition 3.4.4. For the moment, however, we’ll content ourselves with pointing out a sim-
ple but useful property of this invariant. Let 𝑈, 𝑉 and 𝑊 be connected open subsets of 𝐑𝑛 and
𝑓 ∶ 𝑈 → 𝑉 and 𝑔 ∶ 𝑉 → 𝑊 proper 𝐶∞ maps. Then
(3.4.5) deg(𝑔 ∘ 𝑓) = deg(𝑔) deg(𝑓) .
 Draft: March 28, 2018

78 Chapter 3: Integration of Forms

Proof. Let 𝜔 be a compactly supported 𝑛-form with support on 𝑊. Then

(𝑔 ∘ 𝑓)⋆ 𝜔 = 𝑔⋆ 𝑓⋆ 𝜔 ;
so

∫ (𝑔 ∘ 𝑓)⋆ 𝜔 = ∫ 𝑔⋆ (𝑓⋆ 𝜔) = deg(𝑔) ∫ 𝑓⋆ 𝜔

𝑈 𝑈 𝑉

= deg(𝑔) deg(𝑓) ∫ 𝜔 . □
𝑊

From this multiplicative property it is easy to deduce the following result (which we
will need in the next section).
Theorem 3.4.6. Let 𝐴 be a non-singular 𝑛 × 𝑛 matrix and 𝑓𝐴 ∶ 𝐑𝑛 → 𝐑𝑛 the linear mapping
associated with 𝐴. Then deg(𝑓𝐴 ) = +1 if det 𝐴 is positive and −1 if det 𝐴 is negative.
A proof of this result is outlined in Exercises 3.4.v to 3.4.ix below.

Exercises for §3.4

Exercise 3.4.i. Let 𝑈 be an open subset of 𝐑𝑛 and (𝜙𝑖 )𝑖≥1 be a partition of unity on 𝑈. Show
that the mapping 𝑓 ∶ 𝑈 → 𝐑 defined by
∞
𝑓 ≔ ∑ 𝑘𝜙𝑘
𝑘=1
∞
is a proper 𝐶 mapping.
Exercise 3.4.ii. Let 𝑈 and 𝑉 be open subsets of 𝐑𝑛 and 𝐑𝑘 and let 𝑓 ∶ 𝑈 → 𝑉 be a proper
continuous mapping. Prove:
Theorem 3.4.7. If 𝐵 is a compact subset of 𝑉 and 𝐴 = 𝑓−1 (𝐵) then for every open subset 𝑈0
with 𝐴 ⊂ 𝑈0 ⊂ 𝑈, there exists an open subset 𝑉0 with 𝐵 ⊂ 𝑉0 ⊂ 𝑉 and 𝑓−1 (𝑉0 ) ⊂ 𝑈0 .
Hint: Let 𝐶 be a compact subset of 𝑉 with 𝐵 ⊂ int 𝐶. Then the set 𝑊 = 𝑓−1 (𝐶) ∖ 𝑈0 is
compact; so its image 𝑓(𝑊) is compact. Show that 𝑓(𝑊) and 𝐵 are disjoint and let
𝑉0 = int 𝐶 ∖ 𝑓(𝑊) .
Exercise 3.4.iii. Show that if 𝑓 ∶ 𝑈 → 𝑉 is a proper continuous mapping and 𝑋 is a closed
subset of 𝑈, then 𝑓(𝑋) is closed.
Hint: Let 𝑈0 = 𝑈 − 𝑋. Show that if 𝑝 is in 𝑉 ∖ 𝑓(𝑋), then 𝑓−1 (𝑝) is contained in 𝑈0 and
conclude from Exercise 3.4.ii that there exists a neighborhood 𝑉0 of 𝑝 such that 𝑓−1 (𝑉0 ) is
contained in 𝑈0 . Conclude that 𝑉0 and 𝑓(𝑋) are disjoint.
Exercise 3.4.iv. Let 𝑓 ∶ 𝐑𝑛 → 𝐑𝑛 be the translation 𝑓(𝑥) = 𝑥 + 𝑎. Show that deg(𝑓) = 1.
Hint: Let 𝜓 ∶ 𝐑 → 𝐑 be a compactly supported 𝐶∞ function. For 𝑎 ∈ 𝐑, the identity

(3.4.8) ∫ 𝜓(𝑡) 𝑑𝑡 = ∫ 𝜓(𝑡 − 𝑎) 𝑑𝑡

𝐑 𝐑
is easy to prove by elementary calculus, and this identity proves the assertion above in di-
mension one. Now let
(3.4.9) 𝜙(𝑥) = 𝜓(𝑥1 ) ⋯ 𝜓(𝑥𝑛 )
and compute the right and left sides of equation (3.4.3) by Fubini’s theorem.
 Draft: March 28, 2018

§3.4 The degree of a differentiable mapping 79

Exercise 3.4.v. Let 𝜎 be a permutation of the numbers, 1, …, 𝑛 and let 𝑓𝜎 ∶ 𝐑𝑛 → 𝐑𝑛 be the

diffeomorphism, 𝑓𝜎 (𝑥1 , …, 𝑥𝑛 ) = (𝑥𝜎(1) , …, 𝑥𝜎(𝑛) ). Prove that deg 𝑓𝜎 = (−1)𝜎 .
Hint: Let 𝜙 be the function (3.4.9). Show that if 𝜔 = 𝜙(𝑥) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 , then we have
𝑓⋆ 𝜔 = (−1)𝜎 𝜔.
Exercise 3.4.vi. Let 𝑓 ∶ 𝐑𝑛 → 𝐑𝑛 be the mapping
𝑓(𝑥1 , …, 𝑥𝑛 ) = (𝑥1 + 𝜆𝑥2 , 𝑥2 , …, 𝑥𝑛 ).
Prove that deg(𝑓) = 1.
Hint: Let 𝜔 = 𝜙(𝑥1 , …, 𝑥𝑛 ) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 where 𝜙 ∶ 𝐑𝑛 → 𝐑 is compactly supported
and of class 𝐶∞ . Show that
∫ 𝑓⋆ 𝜔 = ∫ 𝜙(𝑥1 + 𝜆𝑥2 , 𝑥2 , …, 𝑥𝑛 ) 𝑑𝑥1 ⋯ 𝑑𝑥𝑛

and evaluate the integral on the right by Fubini’s theorem; i.e., by first integrating with re-
spect to the 𝑥1 variable and then with respect to the remaining variables. Note that by equa-
tion (3.4.8)
∫ 𝑓(𝑥1 + 𝜆𝑥2 , 𝑥2 , …, 𝑥𝑛 ) 𝑑𝑥1 = ∫ 𝑓(𝑥1 , 𝑥2 , …, 𝑥𝑛 ) 𝑑𝑥1 .

Exercise 3.4.vii. Let 𝑓 ∶ 𝐑𝑛 → 𝐑𝑛 be the mapping

𝑓(𝑥1 , …, 𝑥𝑛 ) = (𝜆𝑥1 , 𝑥2 , …, 𝑥𝑛 )
with 𝜆 ≠ 0. Show that deg 𝑓 = +1 if 𝜆 is positive and −1 if 𝜆 is negative.
Hint: In dimension 1 this is easy to prove by elementary calculus techniques. Prove it
in 𝑑-dimensions by the same trick as in the previous exercise.
Exercise 3.4.viii.
(1) Let 𝑒1 , …, 𝑒𝑛 be the standard basis vectors of 𝐑𝑛 and 𝐴, 𝐵 and 𝐶 the linear mappings
defined by
𝑒1 , 𝑖=1
𝐴𝑒𝑖 = { 𝑛
∑𝑗=1 𝑎𝑗,𝑖 𝑒𝑗 , 𝑖 ≠ 1 ,
𝑛
∑𝑗=1 𝑏𝑗 𝑒𝑗 , 𝑖 = 1
(3.4.10) 𝐵𝑒𝑖 = {
𝑒𝑖 , 𝑖≠1,
𝑒1 , 𝑖=1
𝐶𝑒𝑖 = {
𝑒𝑖 + 𝑐𝑖 𝑒1 , 𝑖 ≠ 1 .
Show that
𝑚
∑ 𝑏𝑗 𝑒𝑗 , 𝑖=1
𝐵𝐴𝐶𝑒𝑖 = { 𝑛𝑗=1
∑𝑗=1 (𝑎𝑗,𝑖 + 𝑐𝑖 𝑏𝑗 )𝑒𝑗 + 𝑐𝑖 𝑏1 𝑒1 , 𝑖 ≠ 1 .
for 𝑖 > 1.
(2) Let 𝐿 ∶ 𝐑𝑛 → 𝐑𝑛 be the linear mapping
𝑛
(3.4.11) 𝐿𝑒𝑖 = ∑ ℓ𝑗,𝑖 𝑒𝑗 , 𝑖 = 1, …, 𝑛 .
𝑗=1

Show that if ℓ1,1 ≠ 0 one can write 𝐿 as a product, 𝐿 = 𝐵𝐴𝐶, where 𝐴, 𝐵 and 𝐶 are linear
mappings of the form (3.4.10).
Hint: First solve the equations
ℓ𝑗,1 = 𝑏𝑗
 Draft: March 28, 2018

80 Chapter 3: Integration of Forms

for 𝑗 = 1, …, 𝑛. Next solve the equations

ℓ1,𝑖 = 𝑏1 𝑐𝑖
for 𝑖 > 1. Finally, solve the equations
ℓ𝑗,𝑖 = 𝑎𝑗,𝑖 + 𝑐𝑖 𝑏𝑗
for 𝑖, 𝑗 > 1.
(3) Suppose 𝐿 is invertible. Conclude that 𝐴, 𝐵 and 𝐶 are invertible and verify that Theo-
rem 3.4.6 holds for 𝐵 and 𝐶 using the previous exercises in this section.
(4) Show by an inductive argument that Theorem 3.4.6 holds for 𝐴 and conclude from
(3.4.5) that it holds for 𝐿.
Exercise 3.4.ix. To show that Theorem 3.4.6 holds for an arbitrary linear mapping 𝐿 of the
form (3.4.11) we’ll need to eliminate the assumption: ℓ1,1 ≠ 0. Show that for some 𝑗, ℓ𝑗,1 is
nonzero, and show how to eliminate this assumption by considering 𝑓𝜏1,𝑗 ∘ 𝐿 where 𝜏1,𝑗 is
the transposition 1 ↔ 𝑗.
Exercise 3.4.x. Here is an alternative proof of Theorem 3.4.6 which is shorter than the proof
outlined in Exercise 3.4.ix but uses some slightly more sophisticated linear algebra.
(1) Prove Theorem 3.4.6 for linear mappings which are orthogonal, i.e., satisfy 𝐿⊺ 𝐿 = id𝑛 .
Hints:
• Show that 𝐿⋆ (𝑥12 + ⋯ + 𝑥𝑛2 ) = 𝑥12 + ⋯ + 𝑥𝑛2 .
• Show that 𝐿⋆ (𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ) is equal to 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 or −𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛
depending on whether 𝐿 is orientation-preserving or orinetation reversing. (See
Exercise 1.2.x.)
• Let 𝜓 be as in Exercise 3.4.iv and let 𝜔 be the form
𝜔 = 𝜓(𝑥12 + ⋯ + 𝑥𝑛2 ) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
Show that 𝐿⋆ 𝜔 = 𝜔 if 𝐿 is orientation-preserving and 𝐿⋆ 𝜔 = −𝜔 if 𝐿 is orientation
reversing.
(2) Prove Theorem 3.4.6 for linear mappings which are self-adjoint (satisfy 𝐿⊺ = 𝐿).
Hint: A self-adjoint linear mapping is diagonizable: there exists an intervertible lin-
ear mapping, 𝑀 ∶ 𝐑𝑛 → 𝐑𝑛 such that
(3.4.12) 𝑀−1 𝐿𝑀𝑒𝑖 = 𝜆𝑖 𝑒𝑖 , 𝑖 = 1, …, 𝑛 .
(3) Prove that every invertible linear mapping, 𝐿, can be written as a product, 𝐿 = 𝐵𝐶 where
𝐵 is orthogonal and 𝐶 is self-adjoint.
Hints:
• Show that the mapping, 𝐴 = 𝐿⊺ 𝐿, is self-adjoint and that it is eigenvalues (the 𝜆𝑖 ’s
in equation (3.4.12)) are positive.
• Show that there exists an invertible self-adjoint linear mapping, 𝐶, such that 𝐴 =
𝐶2 and 𝐴𝐶 = 𝐶𝐴.
• Show that the mapping 𝐵 = 𝐿𝐶−1 is orthogonal.

3.5. The change of variables formula

Let 𝑈 and 𝑉 be connected open subsets of 𝐑𝑛 . If 𝑓𝑈 ⥲ 𝑉 is a diffeomorphism, the
determinant of 𝐷𝑓(𝑥) at 𝑥 ∈ 𝑈 is nonzero, and hence, since 𝐷𝑓(𝑥) is a continuous function
of 𝑥, its sign is the same at every point. We will say that 𝑓 is orientation-preserving if this
sign is positive and orientation reversing if it is negative. We will prove below:
 Draft: March 28, 2018

§3.5 The change of variables formula 81

Theorem 3.5.1. The degree of 𝑓 is +1 if 𝑓 is orientation-preserving and −1 if 𝑓 is orientation

reversing.
We will then use this result to prove the following change of variables formula for dif-
feomorphisms.
Theorem 3.5.2. Let 𝜙 ∶ 𝑉 → 𝐑 be a compactly supported continuous function. Then

(3.5.3) ∫ (𝜙 ∘ 𝑓)(𝑥)|det(𝐷𝑓(𝑥))| = ∫ 𝜙(𝑦) 𝑑𝑦 .

𝑈 𝑉

Proof of Theorem 3.5.1. Given a point 𝑎1 ∈ 𝑈, let 𝑎2 = −𝑓(𝑎1 ) and for 𝑖 = 1, 2 let 𝑔𝑖 ∶ 𝐑𝑛 →
𝐑𝑛 be the translation, 𝑔𝑖 (𝑥) = 𝑥 + 𝑎𝑖 . By equation (3.4.2) and Exercise 3.4.iv the composite
diffeomorphism
(3.5.4) 𝑔2 ∘ 𝑓 ∘ 𝑔1
has the same degree as 𝑓, so it suffices to prove the theorem for this mapping. Notice however
that this mapping maps the origin onto the origin. Hence, replacing 𝑓 by this mapping, we
can, without loss of generality, assume that 0 is in the domain of 𝑓 and that 𝑓(0) = 0.
Next notice that if 𝐴 ∶ 𝐑𝑛 ⥲ 𝐑𝑛 is a bijective linear mapping the theorem is true for 𝐴
(by Exercise 3.4.ix), and hence if we can prove the theorem for 𝐴−1 ∘ 𝑓, equation (3.4.2) will
tell us that the theorem is true for 𝑓. In particular, letting 𝐴 = 𝐷𝑓(0), we have
𝐷(𝐴−1 ∘ 𝑓)(0) = 𝐴−1 𝐷𝑓(0) = id𝑛
where id𝑛 is the identity mapping. Therefore, replacing 𝑓 by 𝐴−1 ∘ 𝑓, we can assume that
the mapping 𝑓 (for which we are attempting to prove Theorem 3.5.1) has the properties:
𝑓(0) = 0 and 𝐷𝑓(0) = id𝑛 . Let 𝑔(𝑥) = 𝑓(𝑥) − 𝑥. Then these properties imply that 𝑔(0) = 0
and 𝐷𝑔(0) = 0. □
Lemma 3.5.5. There exists a 𝛿 > 0 such that |𝑔(𝑥)| ≤ 12 |𝑥| for |𝑥| ≤ 𝛿.
Proof. Let 𝑔(𝑥) = (𝑔1 (𝑥), …, 𝑔𝑛 (𝑥)). Then
𝜕𝑔𝑖
(0) = 0 ;
𝜕𝑥𝑗
so there exists a 𝛿 > 0 such that
𝜕𝑔𝑖 1
|
(𝑥)| ≤
𝜕𝑥𝑗 2
for |𝑥| ≤ 𝛿. However, by the mean value theorem,
𝑛
𝜕𝑔𝑖
𝑔𝑖 (𝑥) = ∑ (𝑐)𝑥𝑗
𝑗=1 𝜕𝑥𝑗

for 𝑐 = 𝑡0 𝑥, 0 < 𝑡0 < 1. Thus, for |𝑥| < 𝛿,

1 1
|𝑔𝑖 (𝑥)| ≤ sup |𝑥𝑖 | = |𝑥| ,
2 2
so
1
|𝑔(𝑥)| = sup |𝑔𝑖 (𝑥)| ≤
|𝑥| . □
2
Let 𝜌 be a compactly supported 𝐶∞ function with 0 ≤ 𝜌 ≤ 1 and with 𝜌(𝑥) = 0 for
|𝑥| ≥ 𝛿 and 𝜌(𝑥) = 1 for |𝑥| ≤ 𝛿2 and let 𝑓 ̃∶ 𝐑𝑛 → 𝐑𝑛 be the mapping
̃ ≔ 𝑥 + 𝜌(𝑥)𝑔(𝑥) .
𝑓(𝑥)
 Draft: March 28, 2018

82 Chapter 3: Integration of Forms

It is clear that
(3.5.6) ̃ =𝑥
𝑓(𝑥) for |𝑥| ≥ 𝛿
and, since 𝑓(𝑥) = 𝑥 + 𝑔(𝑥),
̃ = 𝑓(𝑥) 𝛿
(3.5.7) 𝑓(𝑥) for |𝑥| ≤ .
2
In addition, for all 𝑥 ∈ 𝐑𝑛 :
̃ 1
(3.5.8) |𝑓(𝑥)| ≥ |𝑥| .
2
̃
Indeed, by (3.5.6), |𝑓(𝑥)| ≥ |𝑥| for |𝑥| ≥ 𝛿, and for |𝑥| ≤ 𝛿
̃
|𝑓(𝑥)| ≥ |𝑥| − 𝜌(𝑥)|𝑔(𝑥)|
1
≥ |𝑥| − |𝑔(𝑥)| ≥ |𝑥| − |𝑥|
2
1
= |𝑥|
2
by Lemma 3.5.5.
Now let 𝑄𝑟 be the cube 𝑄𝑟 ≔ { 𝑥 ∈ 𝐑𝑛 | |𝑥| ≤ 𝑟 }, and let 𝑄𝑐𝑟 ≔ 𝐑𝑛 ∖ 𝑄𝑟 . From (3.5.8)
we easily deduce that
(3.5.9) ̃ (𝑄𝑟 ) ⊂ 𝑄2𝑟
𝑓−1
for all 𝑟, and hence that 𝑓 ̃ is proper. Also notice that for 𝑥 ∈ 𝑄𝛿 ,
̃ 3
|𝑓(𝑥)| ≤ |𝑥| + |𝑔(𝑥)| ≤ |𝑥|
2
by Lemma 3.5.5 and hence
(3.5.10) ̃ (𝑄𝑐3 ) ⊂ 𝑄𝑐 .
𝑓−1 𝛿 𝛿
2

We will now prove Theorem 3.5.1.

Proof of Theorem 3.5.1. Since 𝑓 is a diffeomorphism mapping 0 to 0, it maps a neighbor-
hood 𝑈0 of 0 in 𝑈 diffeomorphically onto a neighborhood 𝑉0 of 0 in 𝑉, and, by shrinking
𝑈0 if necessary, we can assume that 𝑈0 is contained in 𝑄𝛿/2 and 𝑉0 contained in 𝑄𝛿/4 . Let 𝜔
be an 𝑛-form with support in 𝑉0 whose integral over 𝐑𝑛 is equal to one. Then 𝑓⋆ 𝜔 is sup-
ported in 𝑈0 and hence in 𝑄𝛿/2 . Also by (3.5.9) 𝑓⋆̃ 𝜔 is supported in 𝑄𝛿/2 . Thus both of these
forms are zero outside 𝑄𝛿/2 . However, on 𝑄𝛿/2 , 𝑓 ̃ = 𝑓 by (3.5.7), so these forms are equal
everywhere, and hence

deg(𝑓) = ∫ 𝑓⋆ 𝜔 = ∫ 𝑓⋆̃ 𝜔 = deg(𝑓)̃ .

Next let 𝜔 be a compactly supported 𝑛-form with support in 𝑄𝑐3𝛿/2 and with integral
equal to one. Then 𝑓⋆̃ 𝜔 is supported in 𝑄𝛿𝑐 by (3.5.10), and hence since 𝑓(𝑥) = 𝑥 on 𝑄𝛿𝑐 , we
have 𝑓⋆̃ 𝜔 = 𝜔. Thus
deg(𝑓)̃ = ∫ 𝑓⋆̃ 𝜔 = ∫ 𝜔 = 1 .
Putting these two identities together we conclude that deg(𝑓) = 1. □
If the function, 𝜙, in equation (3.5.4) is a 𝐶∞ function, the identity (3.5.3) is an im-
mediate consequence of the result above and the identity (3.4.3). If 𝜙 is not 𝐶∞ , but is just
continuous, we will deduce equation (3.5.4) from the following result.
 Draft: March 28, 2018

§3.5 The change of variables formula 83

Theorem 3.5.11. Let 𝑉 be an open subset of 𝐑𝑛 . If 𝜙 ∶ 𝐑𝑛 → 𝐑 is a continuous function of

compact support with supp 𝜙 ⊂ 𝑉; then for every 𝜀 > 0 there exists a 𝐶∞ function of compact
support, 𝜓 ∶ 𝐑𝑛 → 𝐑 with supp 𝜓 ⊂ 𝑉 and
sup |𝜓(𝑥) − 𝜙(𝑥)| < 𝜀 .
Proof. Let 𝐴 be the support of 𝜙 and let 𝑑 be the distance in the sup norm from 𝐴 to the
complement of 𝑉. Since 𝜙 is continuous and compactly supported it is uniformly continu-
ous; so for every 𝜀 > 0 there exists a 𝛿 > 0 with 𝛿 < 𝑑2 such that |𝜙(𝑥) − 𝜙(𝑦)| < 𝜀 when
|𝑥 − 𝑦| ≤ 𝛿. Now let 𝑄 be the cube: |𝑥| < 𝛿 and let 𝜌 ∶ 𝐑𝑛 → 𝐑 be a non-negative 𝐶∞
function with supp 𝜌 ⊂ 𝑄 and

(3.5.12) ∫ 𝜌(𝑦)𝑑𝑦 = 1 .
Set
𝜓(𝑥) = ∫ 𝜌(𝑦 − 𝑥)𝜙(𝑦)𝑑𝑦 .

By Theorem 3.2.10 𝜓 is a 𝐶∞ function. Moreover, if 𝐴𝛿 is the set of points in 𝐑𝑑 whose

distance in the sup norm from 𝐴 is ≤ 𝛿 then for 𝑥 ∉ 𝐴𝛿 and 𝑦 ∈ 𝐴 , |𝑥 − 𝑦| > 𝛿 and hence
𝜌(𝑦 − 𝑥) = 0. Thus for 𝑥 ∉ 𝐴𝛿

∫ 𝜌(𝑦 − 𝑥)𝜙(𝑦)𝑑𝑦 = ∫ 𝜌(𝑦 − 𝑥)𝜙(𝑦)𝑑𝑦 = 0 ,

𝐴

so 𝜓 is supported on the compact set 𝐴𝛿 . Moreover, since 𝛿 < 𝑑2 , supp 𝜓 is contained in 𝑉.

Finally note that by (3.5.12) and Exercise 3.4.iv

(3.5.13) ∫ 𝜌(𝑦 − 𝑥) 𝑑𝑦 = ∫ 𝜌(𝑦) 𝑑𝑦 = 1

and hence
𝜙(𝑥) = ∫ 𝜙(𝑥)𝜌(𝑦 − 𝑥)𝑑𝑦
so
𝜙(𝑥) − 𝜓(𝑥) = ∫(𝜙(𝑥) − 𝜙(𝑦))𝜌(𝑦 − 𝑥)𝑑𝑦
and
|𝜙(𝑥) − 𝜓(𝑥)| ≤ ∫ |𝜙(𝑥) − 𝜙(𝑦)|𝜌(𝑦 − 𝑥)𝑑𝑦 .
But 𝜌(𝑦 − 𝑥) = 0 for |𝑥 − 𝑦| ≥ 𝛿; and |𝜙(𝑥) − 𝜙(𝑦)| < 𝜀 for |𝑥 − 𝑦| ≤ 𝛿, so the integrand on
the right is less than
𝜀 ∫ 𝜌(𝑦 − 𝑥) 𝑑𝑦 ,
and hence by equation (3.5.13) we have
|𝜙(𝑥) − 𝜓(𝑥)| ≤ 𝜀 . □
To prove the identity (3.5.3), let 𝛾 ∶ 𝐑𝑛 → 𝐑 be a 𝐶∞ cut-off function which is one on
a neighborhood 𝑉1 of the support of 𝜙 is non-negative, and is compactly supported with
supp 𝛾 ⊂ 𝑉, and let
𝑐 = ∫ 𝛾(𝑦) 𝑑𝑦 .
By Theorem 3.5.11 there exists, for every 𝜀 > 0, a 𝐶∞ function 𝜓, with support on 𝑉1 satis-
fying
𝜀
(3.5.14) |𝜙 − 𝜓| ≤ .
2𝑐
 Draft: March 28, 2018

84 Chapter 3: Integration of Forms

Thus
|∫ (𝜙 − 𝜓)(𝑦)𝑑𝑦| ≤ ∫ |𝜙 − 𝜓|(𝑦)𝑑𝑦
𝑉 𝑉

≤ ∫ 𝛾|𝜙 − 𝜓|(𝑥𝑦)𝑑𝑦
𝑉
𝜀 𝜀
≤ ∫ 𝛾(𝑦)𝑑𝑦 ≤
2𝑐 2
so
𝜀
(3.5.15) |∫ 𝜙(𝑦) 𝑑𝑦 − ∫ 𝜓(𝑦) 𝑑𝑦| ≤ .
𝑉 𝑉 2
Similarly, the expression

|∫ (𝜙 − 𝜓) ∘ 𝑓(𝑥)|det 𝐷𝑓(𝑥)| 𝑑𝑥|

𝑈
is less than or equal to the integral

∫ (𝛾 ∘ 𝑓)(𝑥)|(𝜙 − 𝜓) ∘ 𝑓(𝑥)| |det 𝐷𝑓(𝑥)| 𝑑𝑥

𝑈
𝜀
and by (3.5.14), |(𝜙 − 𝜓) ∘ 𝑓(𝑥)| ≤ 2𝑐
, so this integral is less than or equal to
𝜀
∫(𝛾 ∘ 𝑓)(𝑥)|det 𝐷𝑓(𝑥)| 𝑑𝑥
2𝑐
and hence by (3.5.3) is less than or equal to 2𝜀 . Thus
𝜀
(3.5.16) |∫ (𝜙 ∘ 𝑓)(𝑥) | det 𝐷𝑓(𝑥)|𝑑𝑥 − ∫ 𝜓 ∘ 𝑓(𝑥)| det 𝐷𝑓(𝑥)| 𝑑𝑥| ≤ .
𝑈 𝑈 2
Combining (3.5.15), (3.5.16) and the identity

∫ 𝜓(𝑦) 𝑑𝑦 = ∫ 𝜓 ∘ 𝑓(𝑥)|det 𝐷𝑓(𝑥)| 𝑑𝑥

𝑉
we get, for all 𝜀 > 0,

|∫ 𝜙(𝑦) 𝑑𝑦 − ∫ (𝜙 ∘ 𝑓)(𝑥)|det 𝐷𝑓(𝑥)| 𝑑𝑥| ≤ 𝜀

𝑉 𝑈
and hence
∫ 𝜙(𝑦) 𝑑𝑦 = ∫(𝜙 ∘ 𝑓)(𝑥)|det 𝐷𝑓(𝑥)| 𝑑𝑥 .

Exercises for §3.5

Exercise 3.5.i. Let ℎ ∶ 𝑉 → 𝐑 be a non-negative continuous function. Show that if the
improper integral
∫ ℎ(𝑦)𝑑𝑦
𝑉
is well-defined, then the improper integral

∫ (ℎ ∘ 𝑓)(𝑥)|det 𝐷𝑓(𝑥)|𝑑𝑥
𝑈
is well-defined and these two integrals are equal.
Hint: If (𝜙𝑖 )𝑖≥1 is a partition of unity on 𝑉 then 𝜓𝑖 = 𝜙𝑖 ∘ 𝑓 is a partition of unity on 𝑈
and
∫ 𝜙𝑖 ℎ𝑑𝑦 = ∫ 𝜓𝑖 (ℎ ∘ 𝑓(𝑥))|det 𝐷𝑓(𝑥)|𝑑𝑥 .
 Draft: March 28, 2018

§3.6 Techniques for computing the degree of a mapping 85

Now sum both sides of this identity over 𝑖.

Exercise 3.5.ii. Show that the result above is true without the assumption that ℎ is non-
negative.
Hint: ℎ = ℎ+ − ℎ− , where ℎ+ = max(ℎ, 0) and ℎ− = max(−ℎ, 0).
Exercise 3.5.iii. Show that in equation (3.4.3) one can allow the function 𝜙 to be a continu-
ous compactly supported function rather than a 𝐶∞ compactly supported function.
Exercise 3.5.iv. Let 𝐇𝑛 be the half-space (3.2.12) and 𝑈 and 𝑉 open subsets of 𝐑𝑛 . Suppose
𝑓 ∶ 𝑈 → 𝑉 is an orientation-preserving diffeomorphism mapping 𝑈 ∩ 𝐇𝑛 onto 𝑉 ∩ 𝐇𝑛 .
Show that for 𝜔 ∈ 𝛺𝑛𝑐 (𝑉)

(3.5.17) ∫ 𝑓⋆ 𝜔 = ∫ 𝜔.
𝑈∩𝐇𝑛 𝑉∩𝐇𝑛
Hint: Interpret the left and right hand sides of this formula as improper integrals over
𝑈 ∩ int(𝐇𝑛 ) and 𝑉 ∩ int(𝐇𝑛 ).
Exercise 3.5.v. The boundary of 𝐇𝑛 is the set
𝜕𝐇𝑛 ≔ { (0, 𝑥2 , …, 𝑥𝑛 ) | (𝑥2 , …, 𝑥𝑛 ) ∈ 𝐑𝑛−1 }
so the map
𝜄 ∶ 𝐑𝑛−1 → 𝐇𝑛 , (𝑥2 , …, 𝑥𝑛 ) ↦ (0, 𝑥2 , …, 𝑥𝑛 )
in Exercise 3.2.viii maps 𝐑𝑛−1 bijectively onto 𝜕𝐇𝑛 .
(1) Show that the map 𝑓 ∶ 𝑈 → 𝑉 in Exercise 3.5.iv maps 𝑈 ∩ 𝜕𝐇𝑛 onto 𝑉 ∩ 𝜕𝐇𝑛 .
(2) Let 𝑈′ = 𝜄−1 (𝑈) and 𝑉′ = 𝜄−1 (𝑉). Conclude from (1) that the restriction of 𝑓 to 𝑈∩𝜕𝐇𝑛
gives one a diffeomorphism
𝑔 ∶ 𝑈′ → 𝑉 ′
satisfying:
𝜄∘𝑔=𝑓∘𝜄.
(3) Let 𝜇 be in 𝛺𝑛−1
𝑐 (𝑉). Conclude from equations (3.2.14) and (3.5.17):

∫ 𝑔 ⋆ 𝜄⋆ 𝜇 = ∫ 𝜄⋆ 𝜇
𝑈′ 𝑉′
and in particular show that the diffeomorphism 𝑔 ∶ 𝑈′ → 𝑉′ is orientation-preserving.

3.6. Techniques for computing the degree of a mapping

Let 𝑈 and 𝑉 be open subsets of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝑉 a proper 𝐶∞ mapping. In this
section we will show how to compute the degree of 𝑓 and, in particular, show that it is always
an integer. From this fact we will be able to conclude that the degree of 𝑓 is a topological
invariant of 𝑓: if we deform 𝑓 smoothly, its degree doesn’t change.
Definition 3.6.1. A point 𝑥 ∈ 𝑈 is a critical point of 𝑓 if the derivative
𝐷𝑓(𝑥) ∶ 𝐑𝑛 → 𝐑𝑛
fails to be bijective, i.e., if det(𝐷𝑓(𝑥)) = 0.
We will denote the set of critical points of 𝑓 by 𝐶𝑓 . It is clear from the definition that
this set is a closed subset of 𝑈 and hence, by Exercise 3.4.iii, 𝑓(𝐶𝑓 ) is a closed subset of 𝑉. We
will call this image the set of critical values of 𝑓 and the complement of this image the set of
regular values of 𝑓. Notice that 𝑉 ∖ 𝑓(𝑈) is contained in 𝑓(𝑈) ∖ 𝑓(𝐶𝑓 ), so if a point 𝑞 ∈ 𝑉
is not in the image of 𝑓, it is a regular value of 𝑓 “by default”, i.e., it contains no points of
 Draft: March 28, 2018

86 Chapter 3: Integration of Forms

𝑈 in the pre-image and hence, a fortiori, contains no critical points in its pre-image. Notice
also that 𝐶𝑓 can be quite large. For instance, if 𝑐 ∈ 𝑉 and 𝑓 ∶ 𝑈 → 𝑉 is the constant map
which maps all of 𝑈 onto 𝑐, then 𝐶𝑓 = 𝑈. However, in this example, 𝑓(𝐶𝑓 ) = {𝑐}, so the set
of regular values of 𝑓 is 𝑉 ∖ {𝑐}, and hence (in this example) is an open dense subset of 𝑉.
We will show that this is true in general.
Theorem 3.6.2 (Sard). If 𝑈 and 𝑉 are open subsets of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝑉 a proper 𝐶∞ map,
the set of regular values of 𝑓 is an open dense subset of 𝑉.
We will defer the proof of this to Section 3.7 and in this section explore some of its
implications. Picking a regular value 𝑞 of 𝑓 we will prove:
Theorem 3.6.3. The set 𝑓−1 (𝑞) is a finite set. Moreover, if 𝑓−1 (𝑞) = {𝑝1 , …, 𝑝𝑛 } there exist
connected open neighborhoods 𝑈𝑖 of 𝑝𝑖 in 𝑌 and an open neighborhood 𝑊 of 𝑞 in 𝑉 such that:
(1) for 𝑖 ≠ 𝑗 the sets 𝑈𝑖 and 𝑈𝑗 are disjoint;
(2) 𝑓−1 (𝑊) = 𝑈1 ∪ ⋯ ∪ 𝑈𝑛 ,
(3) 𝑓 maps 𝑈𝑖 diffeomorphically onto 𝑊.
Proof. If 𝑝 ∈ 𝑓−1 (𝑞) then, since 𝑞 is a regular value, 𝑝 ∉ 𝐶𝑓 ; so
𝐷𝑓(𝑝) ∶ 𝐑𝑛 → 𝐑𝑛
is bijective. Hence by the inverse function theorem, 𝑓 maps a neighborhood, 𝑈𝑝 of 𝑝 diffeo-
morphically onto a neighborhood of 𝑞. The open sets
{ 𝑈𝑝 | 𝑝 ∈ 𝑓−1 (𝑞) }
are a covering of 𝑓−1 (𝑞); and, since 𝑓 is proper, 𝑓−1 (𝑞) is compact. Thus we can extract a
finite subcovering
{𝑈𝑝1 , …, 𝑈𝑝𝑁 }
and since 𝑝𝑖 is the only point in 𝑈𝑝𝑖 which maps onto 𝑞, we have that 𝑓−1 (𝑞) = {𝑝1 , …, 𝑝𝑁 }.
Without loss of generality we can assume that the 𝑈𝑝𝑖 ’s are disjoint from each other; for,
if not, we can replace them by smaller neighborhoods of the 𝑝𝑖 ’s which have this property.
By Theorem 3.4.7 there exists a connected open neighborhood 𝑊 of 𝑞 in 𝑉 for which
𝑓−1 (𝑊) ⊂ 𝑈𝑝1 ∪ ⋯ ∪ 𝑈𝑝𝑁 .
To conclude the proof let 𝑈𝑖 ≔ 𝑓−1 (𝑊) ∩ 𝑈𝑝𝑖 . □
The main result of this section is a recipe for computing the degree of 𝑓 by counting
the number of 𝑝𝑖 ’s above, keeping track of orientation.
Theorem 3.6.4. For each 𝑝𝑖 ∈ 𝑓−1 (𝑞) let 𝜎𝑝𝑖 = +1 if 𝑓 ∶ 𝑈𝑖 → 𝑊 is orientation-preserving
and −1 if 𝑓 ∶ 𝑈𝑖 → 𝑊 is orientation reversing. Then
𝑁
(3.6.5) deg(𝑓) = ∑ 𝜎𝑝𝑖 .
𝑖=1
Proof. Let 𝜔 be a compactly supported 𝑛-form on 𝑊 whose integral is one. Then
𝑁
deg(𝑓) = ∫ 𝑓⋆ 𝜔 = ∑ ∫ 𝑓⋆ 𝜔 .
𝑈 𝑖=1 𝑈𝑖
Since 𝑓 ∶ 𝑈𝑖 → 𝑊 is a diffeomorphism
1, 𝑓 is orientation-preserving
∫ 𝑓⋆ 𝜔 = ± ∫ 𝜔 = {
𝑈𝑖 𝑊 −1, 𝑓 is not orientation-preserving .
 Draft: March 28, 2018

§3.6 Techniques for computing the degree of a mapping 87

Thus deg(𝑓) is equal to the sum (3.6.5). □

As we pointed out above, a point 𝑞 ∈ 𝑉 can qualify as a regular value of 𝑓 “by default”,
i.e., by not being in the image of 𝑓. In this case the recipe (3.6.5) for computing the degree
gives “by default” the answer zero. Let’s corroborate this directly.
Theorem 3.6.6. If 𝑓 ∶ 𝑈 → 𝑉 is not surjective, then deg(𝑓) = 0.
Proof. By Exercise 3.4.iii, 𝑉 ∖ 𝑓(𝑈) is open; so if it is nonempty, there exists a compactly
supported 𝑛-form 𝜔 with support in 𝑉 ∖ 𝑓(𝑈) and with integral equal to one. Since 𝜔 = 0
on the image of 𝑓, 𝑓⋆ 𝜔 = 0; so

0 = ∫ 𝑓⋆ 𝜔 = deg(𝑓) ∫ 𝜔 = deg(𝑓) . □
𝑈 𝑉
Remark 3.6.7. In applications the contrapositive of Theorem 3.6.6 is much more useful
than the theorem itself.
Theorem 3.6.8. If deg(𝑓) ≠ 0, then 𝑓 maps 𝑈 surjectively onto 𝑉.
In other words if deg(𝑓) ≠ 0 the equation
𝑓(𝑥) = 𝑦
has a solution, 𝑥 ∈ 𝑈 for every 𝑦 ∈ 𝑉.
We will now show that the degree of 𝑓 is a topological invariant of 𝑓: if we deform 𝑓 by
a “homotopy” we do not change its degree. To make this assertion precise, let’s recall what
we mean by a homotopy between a pair of 𝐶∞ maps. Let 𝑈 be an open subset of 𝐑𝑚 , 𝑉
an open subset of 𝐑𝑛 , 𝐴 an open subinterval of 𝐑 containing 0 and 1, and 𝑓1 , 𝑓2 ∶ 𝑈 → 𝑉
a pair of 𝐶∞ maps. Then a 𝐶∞ map 𝐹 ∶ 𝑈 × 𝐴 → 𝑉 is a homotopy between 𝑓0 and 𝑓1 if
𝐹(𝑥, 0) = 𝑓0 (𝑥) and 𝐹(𝑥, 1) = 𝑓1 (𝑥). (See Definition 2.6.14.) Suppose now that 𝑓0 and 𝑓1 are
proper.
Definition 3.6.9. A homotopy 𝐹 between 𝑓0 and 𝑓1 is a proper homotopy if the map
𝐹♯ ∶ 𝑈 × 𝐴 → 𝑉 × 𝐴
defined by (𝑥, 𝑡) ↦ (𝐹(𝑥, 𝑡), 𝑡) is proper.
Note that if 𝐹 is a proper homotopy between 𝑓0 and 𝑓1 , then for every 𝑡 between 0 and
1, the map
𝑓𝑡 ∶ 𝑈 → 𝑉 , 𝑓𝑡 (𝑥) ≔ 𝐹(𝑥, 𝑡)
is proper.
Now let 𝑈 and 𝑉 be open subsets of 𝐑𝑛 .
Theorem 3.6.10. If 𝑓0 and 𝑓1 are properly homotopic, then deg(𝑓0 ) = deg(𝑓1 ).
Proof. Let
𝜔 = 𝜙(𝑦)𝑑𝑦1 ∧ ⋯ ∧ 𝑑𝑦𝑛
be a compactly supported 𝑛-form on 𝑉 whose integral over 𝑉 is 1. The the degree of 𝑓𝑡 is
equal to

(3.6.11) ∫ 𝜙(𝐹1 (𝑥, 𝑡), …, 𝐹𝑛 (𝑥, 𝑡)) det 𝐷𝑥 𝐹(𝑥, 𝑡)𝑑𝑥 .

𝑈
The integrand in (3.6.11) is continuous and for 0 ≤ 𝑡 ≤ 1 is supported on a compact subset
of 𝑈 × [0, 1], hence (3.6.11) is continuous as a function of 𝑡. However, as we’ve just proved,
deg(𝑓𝑡 ) is integer valued so this function is a constant. □
(For an alternative proof of this result see Exercise 3.6.ix below.)
 Draft: March 28, 2018

88 Chapter 3: Integration of Forms

Applications
We’ll conclude this account of degree theory by describing a couple applications.
Application 3.6.12 (The Brouwer fixed point theorem). Let 𝐵𝑛 be the closed unit ball in 𝐑𝑛 :
𝐵𝑛 ≔ { 𝑥 ∈ 𝐑𝑛 | ‖𝑥‖ ≤ 1 } .
Theorem 3.6.13. If 𝑓 ∶ 𝐵𝑛 → 𝐵𝑛 is a continuous mapping then 𝑓 has a fixed point, i.e., maps
some point, 𝑥0 ∈ 𝐵𝑛 onto itself.
The idea of the proof will be to assume that there isn’t a fixed point and show that this
leads to a contradiction. Suppose that for every point 𝑥 ∈ 𝐵𝑛 we have 𝑓(𝑥) ≠ 𝑥. Consider
the ray through 𝑓(𝑥) in the direction of 𝑥:
𝑓(𝑥) + 𝑠(𝑥 − 𝑓(𝑥)) , 𝑠 ∈ [0, ∞) .
This ray intersects the boundary 𝑆𝑛−1 ≔ 𝜕𝐵𝑛 in a unique point 𝛾(𝑥) (see Figure 3.6.1 below);
and one of the exercises at the end of this section will be to show that the mapping 𝛾 ∶ 𝐵𝑛 →
𝑆𝑛−1 given by 𝑥 ↦ 𝛾(𝑥), is a continuous mapping. Also it is clear from Figure 3.6.1 that
𝛾(𝑥) = 𝑥 if 𝑥 ∈ 𝑆𝑛−1 , so we can extend 𝛾 to a continuous mapping of 𝐑𝑛 into 𝐑𝑛 by letting 𝛾
be the identity for ‖𝑥‖ ≥ 1. Note that this extended mapping has the property
‖𝛾(𝑥)‖ ≥ 1
𝑛
for all 𝑥 ∈ 𝐑 and
(3.6.14) 𝛾(𝑥) = 𝑥
for all ‖𝑥‖ ≥ 1. To get a contradiction we’ll show that 𝛾 can be approximated by a 𝐶∞ map
which has similar properties. For this we will need the following corollary of Theorem 3.5.11.
Lemma 3.6.15. Let 𝑈 be an open subset of 𝐑𝑛 , 𝐶 a compact subset of 𝑈 and 𝜙 ∶ 𝑈 → 𝐑 a
continuous function which is 𝐶∞ on the complement of 𝐶. Then for every 𝜀 > 0, there exists a
𝐶∞ function 𝜓 ∶ 𝑈 → 𝐑, such that 𝜙 − 𝜓 has compact support and |𝜙 − 𝜓| < 𝜀.
Proof. Let 𝜌 be a bump function which is in 𝐶0∞ (𝑈) and is equal to 1 on a neighborhood
of 𝐶. By Theorem 3.5.11 there exists a function 𝜓0 ∈ 𝐶0∞ (𝑈) such that |𝜌𝜙 − 𝜓0 | < 𝜀. To
complete the proof, let 𝜓(1 − 𝜌)𝜙 + 𝜓0 , and note that
𝜙 − 𝜓 = (1 − 𝜌)𝜙 + 𝜌𝜙 − (1 − 𝜌)𝜙 − 𝜓0
= 𝜌𝜙 − 𝜓0 . □
∞
By applying Lemma 3.6.15 to each of the coordinates of the map 𝛾, one obtains a 𝐶
map 𝑔 ∶ 𝐑𝑛 → 𝐑𝑛 such that
‖𝑔 − 𝛾‖ < 𝜀 < 1
and such that 𝑔 = 𝛾 on the complement of a compact set. However, by (3.6.14), this means
that 𝑔 is equal to the identity on the complement of a compact set and hence (see Exer-
cise 3.6.ix) that 𝑔 is proper and deg(𝑔) = 1. On the other hand by (3.6.19) and (3.6.14) we
have ‖𝑔(𝑥)‖ > 1 − 𝜀 for all 𝑥 ∈ 𝐑𝑛 , so 0 ∉ im(𝑔) and hence by Theorem 3.6.4 we have
deg(𝑔) = 0, which provides the desired contradiction.

Application 3.6.16 (The fundamental theorem of algebra). Let

𝑝(𝑧) = 𝑧𝑛 + 𝑎𝑛−1 𝑧𝑛−1 + ⋯ + 𝑎1 𝑧 + 𝑎0
be a polynomial of degree 𝑛 with complex coefficients. If we identify the complex plane
𝐂 ≔ { 𝑧 = 𝑥 + 𝑖𝑦 | 𝑥, 𝑦 ∈ 𝐑 }
 Draft: March 28, 2018

§3.6 Techniques for computing the degree of a mapping 89

𝑓(𝑥)

𝛾(𝑥)

Figure 3.6.1. Brouwer fixed point theorem

with 𝐑2 via the map 𝐑2 → 𝐂 given by (𝑥, 𝑦) ↦ 𝑧 = 𝑥 + 𝑖𝑦, we can think of 𝑝 as defining a
mapping
𝑝 ∶ 𝐑2 → 𝐑2 , 𝑧 ↦ 𝑝(𝑧) .
We will prove:
Theorem 3.6.17. The mapping 𝑝 ∶ 𝐑2 → 𝐑2 is proper and deg(𝑝) = 𝑛.
Proof. For 𝑡 ∈ 𝐑 let
𝑛−1
𝑝𝑡 (𝑧) ≔ (1 − 𝑡)𝑧𝑛 + 𝑡𝑝(𝑧) = 𝑧𝑛 + 𝑡 ∑ 𝑎𝑖 𝑧𝑖 .
𝑖=0
We will show that the mapping
𝑔 ∶ 𝐑2 × 𝐑 → 𝐑2 , (𝑧, 𝑡) ↦ 𝑝𝑡 (𝑧)
is a proper homotopy. Let
𝐶 = sup |𝑎𝑖 | .
0≤𝑖≤𝑛−1
Then for |𝑧| ≥ 1 we have
|𝑎0 + ⋯ + 𝑎𝑛−1 𝑧𝑛−1 | ≤ |𝑎0 | + |𝑎1 ||𝑧| + ⋯ + |𝑎𝑛−1 | |𝑧|𝑛−1
≤ 𝐶𝑛|𝑧|𝑛−1 ,
and hence, for |𝑡| ≤ 𝑎 and |𝑧| ≥ 2𝑎𝐶𝑛,
|𝑝𝑡 (𝑧)| ≥ |𝑧|𝑛 − 𝑎𝐶𝑛|𝑧|𝑛−1
≥ 𝑎𝐶𝑛|𝑧|𝑛−1 .
If 𝐴 ⊂ 𝐂 is compact, then for some 𝑅 > 0, 𝐴 is contained in the disk defined by |𝑤| ≤ 𝑅,
and hence the set
{ 𝑧 ∈ 𝐂 | (𝑡, 𝑝𝑡 (𝑧)) ∈ [−𝑎, 𝑎] × 𝐴 }
is contained in the compact set
{ 𝑧 ∈ 𝐂 | 𝑎𝐶|𝑧|𝑛−1 ≤ 𝑅 } .
This shows that 𝑔 is a proper homotopy.
Thus for each 𝑡 ∈ 𝐑, the map 𝑝𝑡 ∶ 𝐂 → 𝐂 is proper and
deg(𝑝𝑡 ) = deg(𝑝1 ) = deg(𝑝) = deg(𝑝0 )
However, 𝑝0 ∶ 𝐂 → 𝐂 is just the mapping 𝑧 ↦ 𝑧𝑛 and an elementary computation (see
Exercises 3.6.v and 3.6.vi below) shows that the degree of this mapping is 𝑛. □
 Draft: March 28, 2018

90 Chapter 3: Integration of Forms

In particular for 𝑛 > 0 the degree of 𝑝 is nonzero; so by Theorem 3.6.4 we conclude that
𝑝 ∶ 𝐂 → 𝐂 is surjective and hence has zero in its image.
Theorem 3.6.18 (Fundamental theorem of algebra). Every positive-degree polynomial
𝑝(𝑧) = 𝑧𝑛 + 𝑎𝑛−1 𝑧𝑛−1 + ⋯ + 𝑎0
with complex coefficients has a complex root: 𝑝(𝑧0 ) = 0 for some 𝑧0 ∈ 𝐂.
Exercises for §3.6
Exercise 3.6.i. Let 𝑊 be a subset of 𝐑𝑛 and let 𝑎(𝑥), 𝑏(𝑥) and 𝑐(𝑥) be real-valued functions
on 𝑊 of class 𝐶𝑟 . Suppose that for every 𝑥 ∈ 𝑊 the quadratic polynomial
𝑎(𝑥)𝑠2 + 𝑏(𝑥)𝑠 + 𝑐(𝑥)
has two distinct real roots, 𝑠+ (𝑥) and 𝑠− (𝑥), with 𝑠+ (𝑥) > 𝑠− (𝑥). Prove that 𝑠+ and 𝑠− are
functions of class 𝐶𝑟 .
Hint: What are the roots of the quadratic polynomial: 𝑎𝑠2 + 𝑏𝑠 + 𝑐?
Exercise 3.6.ii. Show that the function 𝛾(𝑥) defined in Figure 3.6.1 is a continuous surjec-
tion 𝐵𝑛 → 𝑆2𝑛−1 .
Hint: 𝛾(𝑥) lies on the ray,
𝑓(𝑥) + 𝑠(𝑥 − 𝑓(𝑥)) , 𝑠 ∈ [0, ∞)
and satisfies ‖𝛾(𝑥)‖ = 1. Thus
𝛾(𝑥) = 𝑓(𝑥) + 𝑠0 (𝑥 − 𝑓(𝑥)) ,
where 𝑠0 is a non-negative root of the quadratic polynomial
‖𝑓(𝑥) + 𝑠(𝑥 − 𝑓(𝑥))‖2 − 1 .
Argue from Figure 3.6.1 that this polynomial has to have two distinct real roots.
Exercise 3.6.iii. Show that the Brouwer fixed point theorem isn’t true if one replaces the
closed unit ball by the open unit ball.
Hint: Let 𝑈 be the open unit ball (i.e., the interior of 𝐵𝑛 ). Show that the map
𝑥
ℎ ∶ 𝑈 → 𝐑𝑛 , ℎ(𝑥) ≔
1 − ‖𝑥‖2
is a diffeomorphism of 𝑈 onto 𝐑𝑛 , and show that there are lots of mappings of 𝐑𝑛 onto 𝐑𝑛
which do not have fixed points.
Exercise 3.6.iv. Show that the fixed point in the Brouwer theorem doesn’t have to be an
interior point of 𝐵𝑛 , i.e., show that it can lie on the boundary.
Exercise 3.6.v. If we identify 𝐂 with 𝐑2 via the mapping (𝑥, 𝑦) ↦ 𝑥 + 𝑖𝑦, we can think of a
𝐂-linear mapping of 𝐂 into itself, i.e., a mapping of the form
𝑧 ↦ 𝑐𝑧
for a fixed 𝑐 ∈ 𝐂, as an 𝐑-linear mapping of 𝐑2 into itself. Show that the determinant of this
mapping is |𝑐|2 .
Exercise 3.6.vi.
 Draft: March 28, 2018

§3.6 Techniques for computing the degree of a mapping 91

(1) Let 𝑓 ∶ 𝐂 → 𝐂 be the mapping 𝑓(𝑧) ≔ 𝑧𝑛 . Show that 𝐷𝑓(𝑧) is the linear map
𝐷𝑓(𝑧) = 𝑛𝑧𝑛−1
given by multiplication by 𝑛𝑧𝑛−1 .
Hint: Argue from first principles. Show that for ℎ ∈ 𝐂 = 𝐑2
(𝑧 + ℎ)𝑛 − 𝑧𝑛 − 𝑛𝑧𝑛−1 ℎ
|ℎ|
tends to zero as |ℎ| → 0.
(2) Conclude from Exercise 3.6.v that
det(𝐷𝑓(𝑧)) = 𝑛2 |𝑧|2𝑛−2 .
(3) Show that at every point 𝑧 ∈ 𝐂 ∖ {0}, 𝑓 is orientation preserving.
(4) Show that every point, 𝑤 ∈ 𝐂 ∖ {0} is a regular value of 𝑓 and that
𝑓−1 (𝑤) = {𝑧1 , …, 𝑧𝑛 }
with 𝜎𝑧𝑖 = +1.
(5) Conclude that the degree of 𝑓 is 𝑛.
Exercise 3.6.vii. Prove that the map 𝑓 from Exercise 3.6.vi has degree 𝑛 by deducing this
directly from the definition of degree.
Hints:
• Show that in polar coordinates, 𝑓 is the map (𝑟, 𝜃) ↦ (𝑟𝑛 , 𝑛𝜃).
• Let 𝜔 be the 2-form 𝜔 ≔ 𝑔(𝑥2 + 𝑦2 )𝑑𝑥 ∧ 𝑑𝑦, where 𝑔(𝑡) is a compactly supported
𝐶∞ function of 𝑡. Show that in polar coordinates 𝜔 = 𝑔(𝑟2 )𝑟𝑑𝑟 ∧ 𝑑𝜃, and compute
the degree of 𝑓 by computing the integrals of 𝜔 and 𝑓⋆ 𝜔 in polar coordinates and
comparing them.
Exercise 3.6.viii. Let 𝑈 be an open subset of 𝐑𝑛 , 𝑉 an open subset of 𝐑𝑚 , 𝐴 an open subin-
terval of 𝐑 containing 0 and 1, 𝑓0 , 𝑓1 ∶ 𝑈 → 𝑉 a pair of 𝐶∞ mappings, and 𝐹 ∶ 𝑈 × 𝐴 → 𝑉
a homotopy between 𝑓0 and 𝑓1 .
(1) In Exercise 2.4.iv you proved that if 𝜇 ∈ 𝛺𝑘 (𝑉) and 𝑑𝜇 = 0, then
(3.6.19) 𝑓0⋆ 𝜇 − 𝑓1⋆ 𝜇 = 𝑑𝜈
where 𝜈 is the (𝑘 − 1)-form 𝑄𝛼 in equation (2.6.17). Show (by careful inspection of the
definition of 𝑄𝛼) that if 𝐹 is a proper homotopy and 𝜇 ∈ 𝛺𝑘𝑐 (𝑉) then 𝜈 ∈ 𝛺𝑘−1
𝑐 (𝑈).
(2) Suppose in particular that 𝑈 and 𝑉 are open subsets of 𝐑𝑛 and 𝜇 is in 𝛺𝑛𝑐 (𝑉). Deduce
from equation (3.6.19) that
∫ 𝑓0⋆ 𝜇 = ∫ 𝑓1⋆ 𝜇
and deduce directly from the definition of degree that degree is a proper homotopy
invariant.
Exercise 3.6.ix. Let 𝑈 be an open connected subset of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝑈 a proper 𝐶∞
map. Prove that if 𝑓 is equal to the identity on the complement of a compact set 𝐶, then 𝑓
is proper and deg(𝑓) = 1.
Hints:
• Show that for every subset 𝐴 ⊂ 𝑈, we have 𝑓−1 (𝐴) ⊂ 𝐴 ∪ 𝐶, and conclude from
this that 𝑓 is proper.
• Use the recipe (1.6.2) to compute deg(𝑓) with 𝑞 ∈ 𝑈 ∖ 𝑓(𝐶).
 Draft: March 28, 2018

92 Chapter 3: Integration of Forms

Exercise 3.6.x. Let (𝑎𝑖,𝑗 ) be an 𝑛 × 𝑛 matrix and 𝐴 ∶ 𝐑𝑛 → 𝐑𝑛 the linear mapping associated
with this matrix. Frobenius’ Theorem asserts: If the 𝑎𝑖,𝑗 are non-negative then 𝐴 has a non-
negative eigenvalue. In other words there exists a 𝑣 ∈ 𝐑𝑛 and a 𝜆 ∈ 𝐑, 𝜆 ≥ 0, such that
𝐴𝑣 = 𝜆𝑣. Deduce this linear algebra result from the Brouwer fixed point theorem.
Hints:
• We can assume that 𝐴 is bijective, otherwise 0 is an eigenvalue. Let 𝑆𝑛−1 be the
(𝑛 − 1)-sphere, defined by |𝑥| = 1, and 𝑓 ∶ 𝑆𝑛−1 → 𝑆𝑛−1 the map,
𝐴𝑥
𝑓(𝑥) = .
‖𝐴𝑥‖
Show that 𝑓 maps the set
𝑄 ≔ { (𝑥1 , …, 𝑥𝑛 ) ∈ 𝑆𝑛−1 | 𝑥1 , …, 𝑥𝑛 ≥ 0 }
into itself.
• It is easy to prove that 𝑄 is homeomorphic to the unit ball 𝐵𝑛−1 , i.e., that there
exists a continuous map 𝑔 ∶ 𝑄 → 𝐵𝑛−1 which is invertible and has a continuous
inverse. Without bothering to prove this fact deduce from it Frobenius’ theorem.

3.7. Appendix: Sard’s theorem

The version of Sard’s theorem stated in §3.5 is a corollary of the following more general
result.
Theorem 3.7.1. Let 𝑈 be an open subset of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝐑𝑛 a 𝐶∞ map. Then 𝐑𝑛 ∖ 𝑓(𝐶𝑓 )
is dense in 𝐑𝑛 .
Before undertaking to prove this we will make a few general comments about this result.
Remark 3.7.2. If (𝑈𝑚 )𝑚≥1 are open dense subsets of 𝐑𝑛 , the intersection ⋂𝑚≥1 𝑈𝑚 is dense
in 𝐑𝑛 . (This follows from the Baire category theorem; see, for instance, [7, Ch. 6 Thm. 34;
11, Thm. 48.2] or Exercise 3.7.iv.)
Remark 3.7.3. If (𝐴 𝑛 )𝑛≥1 is a covering of 𝑈 by compact sets, 𝑂𝑛 ≔ 𝐑𝑛 ∖ 𝑓(𝐶𝑓 ∩ 𝐴 𝑛 ) is open,
so if we can prove that it is dense then by Remark 3.7.2 we will have proved Sard’s theorem.
Hence since we can always cover 𝑈 by a countable collection of closed cubes, it suffices to
prove: for every closed cube 𝐴 ⊂ 𝑈, the subspace 𝐑𝑛 ∖ 𝑓(𝐶𝑓 ∩ 𝐴) is dense in 𝐑𝑛 .
Remark 3.7.4. Let 𝑔 ∶ 𝑊 → 𝑈 be a diffeomorphism and let ℎ = 𝑓 ∘ 𝑔. Then
(3.7.5) 𝑓(𝐶𝑓 ) = ℎ(𝐶ℎ )
so Sard’s theorem for ℎ implies Sard’s theorem for 𝑓.
We will first prove Sard’s theorem for the set of super-critical points of 𝑓, the set:
♯
𝐶𝑓 ≔ { 𝑝 ∈ 𝑈 | 𝐷𝑓(𝑝) = 0 } .
♯
Proposition 3.7.6. Let 𝐴 ⊂ 𝑈 be a closed cube. Then the open set 𝐑𝑛 ∖ 𝑓(𝐴 ∩ 𝐶𝑓 ) is a dense
subset of 𝐑𝑛 .
We’ll deduce this from the lemma below.
♯
Lemma 3.7.7. Given 𝜀 > 0 one can cover 𝑓(𝐴 ∩ 𝐶𝑓 ) by a finite number of cubes of total
volume less than 𝜀.
 Draft: March 28, 2018

§3.7 Appendix: Sard’s theorem 93

Proof. Let the length of each of the sides of 𝐴 be ℓ. Given 𝛿 > 0 one can subdivide 𝐴 into 𝑁𝑛
cubes, each of volume (ℓ/𝑁)𝑛 , such that if 𝑥 and 𝑦 are points of any one of these subcubes
𝜕𝑓𝑖 𝜕𝑓
(3.7.8) | (𝑥) − 𝑖 (𝑦)| < 𝛿 .
𝜕𝑥𝑗 𝜕𝑥𝑗
♯ ♯
Let 𝐴 1 , …, 𝐴 𝑚 be the cubes in this collection which intersect 𝐶𝑓 . Then for 𝑧0 ∈ 𝐴 𝑖 ∩ 𝐶𝑓 ,
𝜕𝑓𝑖
(𝑧 )
𝜕𝑥𝑗 0
= 0, so for 𝑧 ∈ 𝐴 𝑖 we have

𝜕𝑓𝑖
(3.7.9) | (𝑧)| < 𝛿
𝜕𝑥𝑗
by equation (3.7.8). If 𝑥 and 𝑦 are points of 𝐴 𝑖 then by the mean value theorem there exists
a point 𝑧 on the line segment joining 𝑥 to 𝑦 such that
𝑛
𝜕𝑓𝑖
𝑓𝑖 (𝑥) − 𝑓𝑖 (𝑦) = ∑ (𝑧)(𝑥𝑗 − 𝑦𝑗 )
𝑗=1 𝜕𝑥𝑗

and hence by (3.7.9)

𝑛
ℓ
|𝑓𝑖 (𝑥) − 𝑓𝑖 (𝑦)| ≤ 𝛿 ∑ |𝑥𝑗 − 𝑦𝑗 | ≤ 𝑛𝛿 .
𝑗=1 𝑁
𝑛
Thus 𝑓(𝐶𝑓 ∩ 𝐴 𝑖 ) is contained in a cube 𝐵𝑖 of volume (𝑛 𝛿ℓ
𝑁
) , and 𝑓(𝐶𝑓 ∩ 𝐴) is contained in
a union of cubes 𝐵𝑖 of total volume less than
𝛿𝑛 ℓ𝑛
𝑁𝑛 𝑛𝑛 = 𝑛 𝑛 𝛿𝑛 ℓ𝑛
𝑁𝑛
so if we choose 𝛿 such that 𝑛𝑛 𝛿𝑛 ℓ𝑛 < 𝜀, we’re done. □
Proof of Proposition 3.7.6. To prove Proposition 3.7.6 we have to show that for every point
♯
𝑝 ∈ 𝐑𝑛 and neighborhood, 𝑊 of 𝑝, the set 𝑊 ∖ 𝑓(𝐶𝑓 ∩ 𝐴) is nonempty. Suppose
♯
(3.7.10) 𝑊 ⊂ 𝑓(𝐶𝑓 ∩ 𝐴) .
Without loss of generality we can assume 𝑊 is a cube of volume 𝜀, but the lemma tells us
♯
that 𝑓(𝐶𝑓 ∩ 𝐴) can be covered by a finite number of cubes whose total volume is less than 𝜀,
and hence by (3.7.10) 𝑊 can be covered by a finite number of cubes of total volume less than
𝜀, so its volume is less than 𝜀. This contradiction proves that the inclusion (3.7.10) cannot
hold. □
Now we prove Theorem 3.7.1.
𝜕𝑓𝑖
Proof of Theorem 3.7.1. Let 𝑈𝑖,𝑗 be the subset of 𝑈 where 𝜕𝑥𝑗
≠ 0. Then
♯
𝑈 = 𝐶𝑓 ∪ ⋃ 𝑈𝑖,𝑗 ,
1≤𝑖,𝑗≤𝑛

so to prove the theorem it suffices to show that 𝐑𝑛 ∖ 𝑓(𝑈𝑖,𝑗 ∩ 𝐶𝑓 ) is dense in 𝐑𝑛 , i.e., it

suffices to prove the theorem with 𝑈 replaced by 𝑈𝑖,𝑗 . Let 𝜎𝑖 ∶ 𝐑𝑛 ⥲ 𝐑𝑛 be the involution
which interchanges 𝑥1 and 𝑥𝑖 and leaves the remaining 𝑥𝑘 ’s fixed. Letting 𝑓new = 𝜎𝑖 𝑓old 𝜎𝑗
and 𝑈new = 𝜎𝑗 𝑈old , we have, for 𝑓 = 𝑓new and 𝑈 = 𝑈new
𝜕𝑓1
(3.7.11) (𝑝) ≠ 0
𝜕𝑥1
 Draft: March 28, 2018

94 Chapter 3: Integration of Forms

for all 𝑝 ∈ 𝑈 so we’re reduced to proving Theorem 3.7.1 for maps 𝑓 ∶ 𝑈 → 𝐑𝑛 having the
property (3.7.11). Let 𝑔 ∶ 𝑈 → 𝐑𝑛 be defined by
(3.7.12) 𝑔(𝑥1 , …, 𝑥𝑛 ) = (𝑓1 (𝑥), 𝑥2 , …, 𝑥𝑛 ) .
Then
𝑔⋆ 𝑥1 = 𝑓⋆ 𝑥1 = 𝑓1 (𝑥1 , …, 𝑥𝑛 )
and
𝜕𝑓1
det(𝐷𝑔) = ≠0.
𝜕𝑥1
Thus, by the inverse function theorem, 𝑔 is locally a diffeomorphism at every point 𝑝 ∈ 𝑈.
This means that if 𝐴 is a compact subset of 𝑈 we can cover 𝐴 by a finite number of open
subsets 𝑈𝑖 ⊂ 𝑈 such that 𝑔 maps 𝑈𝑖 diffeomorphically onto an open subset 𝑊𝑖 in 𝐑𝑛 . To
conclude the proof of the theorem we’ll show that 𝐑𝑛 ∖ 𝑓(𝐶𝑓 ∩ 𝑈𝑖 ∩ 𝐴) is a dense subset of
𝐑𝑛 . Let ℎ ∶ 𝑊𝑖 → 𝐑𝑛 be the map ℎ = 𝑓∘𝑔−1 . To prove this assertion it suffices by Remark 3.7.4
to prove that the set 𝐑𝑛 ∖ ℎ(𝐶ℎ ) is dense in 𝐑𝑛 . This we will do by induction on 𝑛. First note
♯
that for 𝑛 = 1, we have 𝐶𝑓 = 𝐶𝑓 , so we’ve already proved Theorem 3.7.1 in dimension one.
Now note that by (3.7.12) we have ℎ⋆ 𝑥1 = 𝑥1 , i.e., ℎ is a mapping of the form
ℎ(𝑥1 , …, 𝑥𝑛 ) = (𝑥1 , ℎ2 (𝑥), …, ℎ𝑛 (𝑥)) .
Thus if we let
𝑊𝑐 ≔ { (𝑥2 , …, 𝑥𝑛 ) ∈ 𝐑𝑛−1 | (𝑐, 𝑥2 , …, 𝑥𝑛 ) ∈ 𝑊𝑖 }
and let ℎ𝑐 ∶ 𝑊𝑐 → 𝐑𝑛−1 be the map
ℎ𝑐 (𝑥2 , …, 𝑥𝑛 ) = (ℎ2 (𝑐, 𝑥2 , …, 𝑥𝑛 ), …, ℎ𝑛 (𝑐, 𝑥2 , …, 𝑥𝑛 )) .
Then
det(𝐷ℎ𝑐 )(𝑥2 , …, 𝑥𝑛 ) = det(𝐷ℎ)(𝑐, 𝑥2 , …, 𝑥𝑛 ) ,
and hence
(3.7.13) (𝑐, 𝑥) ∈ 𝑊𝑖 ∩ 𝐶ℎ ⟺ 𝑥 ∈ 𝐶ℎ𝑐 .
Now let 𝑝0 = (𝑐, 𝑥0 ) be a point in 𝐑𝑛 . We have to show that every neighborhood 𝑉 of 𝑝0
contains a point 𝑝 ∈ 𝐑𝑛 ∖ ℎ(𝐶ℎ ). Let 𝑉𝑐 ⊂ 𝐑𝑛−1 be the set of points 𝑥 for which (𝑐, 𝑥) ∈ 𝑉. By
induction 𝑉𝑐 contains a point 𝑥 ∈ 𝐑𝑛−1 ∖ ℎ𝑐 (𝐶ℎ𝑐 ) and hence 𝑝 = (𝑐, 𝑥) is in 𝑉 by definition
and in 𝐑𝑛 ∖ ℎ(𝐶ℎ ) by (3.7.13). □
Exercises for §3.7
Exercise 3.7.i. What are the set of critical points and the image of the set of critical points
for the following maps 𝐑 → 𝐑?
(1) The map 𝑓1 (𝑥) = (𝑥2 − 1)2 .
(2) The map 𝑓2 (𝑥) = sin(𝑥) + 𝑥.
(3) The map
0, 𝑥≤0
𝑓3 (𝑥) = { − 1
𝑒 𝑥, 𝑥 > 0 .
Exercise 3.7.ii (Sard’s theorem for affine maps). Let 𝑓 ∶ 𝐑𝑛 → 𝐑𝑛 be an affine map, i.e., a
map of the form
𝑓(𝑥) = 𝐴(𝑥) + 𝑥0
where 𝐴 ∶ 𝐑𝑛 → 𝐑𝑛 is a linear map and 𝑥0 ∈ 𝐑𝑛 . Prove Sard’s theorem for 𝑓.
 Draft: March 28, 2018

§3.7 Appendix: Sard’s theorem 95

Exercise 3.7.iii. Let 𝜌 ∶ 𝐑 → 𝐑 be a 𝐶∞ function which is supported in the interval (−1/2, 1/2)
and has a maximum at the origin. Let (𝑟𝑖 )𝑖≥1 be an enumeration of the rational numbers, and
let 𝑓 ∶ 𝐑 → 𝐑 be the map
∞
𝑓(𝑥) = ∑ 𝑟𝑖 𝜌(𝑥 − 𝑖) .
𝑖=1
Show that 𝑓 is a 𝐶∞ map and show that the image of 𝐶𝑓 is dense in 𝐑.
The moral of this example: Sard’s theorem says that the complement of 𝐶𝑓 is dense in 𝐑,
but 𝐶𝑓 can be dense as well.
Exercise 3.7.iv. Prove the assertion made in Remark 3.7.4.
Hint: You need to show that for every point 𝑝 ∈ 𝐑𝑛 and every neighborhood 𝑉 of 𝑝,
𝑉 ∩ ⋂𝑛≥1 𝑈𝑛 is nonempty. Construct, by induction, a family of closed balls (𝐵𝑘 )𝑘≥1 such that
• 𝐵𝑘 ⊂ 𝑉,
• 𝐵𝑘+1 ⊂ 𝐵𝑘 ,
• 𝐵𝑘 ⊂ ⋂𝑛≤𝑘 𝑈𝑛 ,
• The radius of 𝐵𝑘 is less than 𝑘1 ,
and show that ⋂𝑘≥1 𝐵𝑘 ≠ ∅.
Exercise 3.7.v. Verify equation (3.7.5).
Draft: March 28, 2018
 Draft: March 28, 2018

CHAPTER 4

Manifolds & Forms on Manifolds

4.1. Manifolds
Our agenda in this chapter is to extend to manifolds the results of Chapters 2 and 3 and
to formulate and prove manifold versions of two of the fundamental theorems of integral
calculus: Stokes’ theorem and the divergence theorem. In this section we’ll define what we
mean by the term manifold, however, before we do so, a word of encouragement. Having
had a course in multivariable calculus, you are already familiar with manifolds, at least in
their one and two-dimensional emanations, as curves and surfaces in 𝐑3 , i.e., a manifold is
basically just an 𝑛-dimensional surface in some high dimensional Euclidean space. To make
this definition precise let 𝑋 be a subset of 𝐑𝑁 , 𝑌 a subset of 𝐑𝑛 and 𝑓 ∶ 𝑋 → 𝑌 a continuous
map. We recall:
Definition 4.1.1. The map 𝑓 is a 𝐶∞ map if for every 𝑝 ∈ 𝑋, there exists a neighborhood
𝑈𝑝 of 𝑝 in 𝐑𝑁 and a 𝐶∞ map 𝑔𝑝 ∶ 𝑈𝑝 → 𝐑𝑛 which coincides with 𝑓 on 𝑈𝑝 ∩ 𝑋.
We also recall:
Theorem 4.1.2. If 𝑓 ∶ 𝑋 → 𝑌 is a 𝐶∞ map, there exists a neighborhood 𝑈, of 𝑋 in 𝐑𝑁 and a
𝐶∞ map 𝑔 ∶ 𝑈 → 𝐑𝑛 such that 𝑔 coincides with 𝑓 on 𝑋.
(A proof of this can be found in Appendix a.)
We will say that 𝑓 is a diffeomorphism if 𝑓 is a bijection and 𝑓 and 𝑓−1 are both 𝐶∞
maps. In particular if 𝑌 is an open subset of 𝐑𝑛 , 𝑋 is an example of an object which we will
call a manifold. More generally:
Definition 4.1.3. Let 𝑁 and 𝑛 be nonnegative integers with 𝑛 ≤ 𝑁. A subset 𝑋 ⊂ 𝐑𝑁 is an
𝑛-manifold if for every 𝑝 ∈ 𝑋 there exists a neighborhood 𝑉 of 𝑝 in 𝐑𝑁 , an open subset 𝑈
in 𝐑𝑛 , and a diffeomorphism 𝜙 ∶ 𝑈 ⥲ 𝑋 ∩ 𝑉.
Thus 𝑋 is an 𝑛-manifold if, locally near every point 𝑝, 𝑋 “looks like” an open subset of
𝐑𝑛 .
Examples 4.1.4.
(1) Graphs of functions: Let 𝑈 be an open subset of 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝐑 a 𝐶∞ function. Its
graph
𝛤𝑓 = { (𝑥, 𝑓(𝑥)) ∈ 𝐑𝑛+1 | 𝑥 ∈ 𝑈 }
is an 𝑛-manifold in 𝐑𝑛+1 . In fact the map
𝜙 ∶ 𝑈 → 𝐑𝑛+1 , 𝑥 ↦ (𝑥, 𝑓(𝑥))
is a diffeomorphism of 𝑈 onto 𝛤𝑓 . (It’s clear that 𝜙 is a 𝐶∞ map, and it is a diffeomor-
phism since its inverse is the map 𝜋 ∶ 𝛤𝑓 → 𝑈 given by 𝜋(𝑥, 𝑡) ≔ 𝑥, which is also clearly
𝐶∞ .)
97
 Draft: March 28, 2018

98 Chapter 4: Manifolds & Forms on Manifolds

(2) Graphs of mappings: More generally if 𝑓 ∶ 𝑈 → 𝐑𝑘 is a 𝐶∞ map, its graph 𝛤𝑓 is an

𝑛-manifold in 𝐑𝑛+𝑘 .
(3) Vector spaces: Let 𝑉 be an 𝑛- dimensional vector subspace of 𝐑𝑁 , and (𝑒1 , …, 𝑒𝑛 ) a basis
of 𝑉. Then the linear map
𝑛
(4.1.5) 𝜙 ∶ 𝐑𝑛 → 𝑉 , (𝑥1 , …, 𝑥𝑛 ) ↦ ∑ 𝑥𝑖 𝑒𝑖
𝑖=1

is a diffeomorphism of 𝐑 onto 𝑉. Hence every 𝑛-dimensional vector subspace of 𝐑𝑁

𝑛

is automatically an 𝑛-dimensional submanifold of 𝐑𝑁 . Note, by the way, that if 𝑉 is any

𝑛-dimensional vector space, not necessarily a subspace of 𝐑𝑁 , the map (4.1.5) gives
us an identification of 𝑉 with 𝐑𝑛 . This means that we can speak of subsets of 𝑉 as
being 𝑘-dimensional submanifolds if, via this identification, they get mapped onto 𝑘-
dimensional submanifolds of 𝐑𝑛 . (This is a trivial, but useful, observation since a lot of
interesting manifolds occur “in nature” as subsets of some abstract vector space rather
than explicitly as subsets of some 𝐑𝑛 . An example is the manifold, 𝑂(𝑛), of orthogonal
𝑛 × 𝑛 matrices. This manifold occurs in nature as a submanifold of the vector space of
𝑛 by 𝑛 matrices.)
(4) Affine subspaces of 𝐑𝑛 : These are manifolds of the form 𝑝 + 𝑉, where 𝑉 is a vector sub-
space of 𝐑𝑁 and 𝑝 is a specified point in 𝐑𝑁 . In other words, they are diffeomorphic
copies of vector subspaces with respect to the diffeomorphism
𝜏𝑝 ∶ 𝐑𝑁 ⥲ 𝐑𝑁 , 𝑥 ↦ 𝑝 + 𝑥 .
If 𝑋 is an arbitrary submanifold of 𝐑𝑁 its tangent space a point 𝑝 ∈ 𝑋, is an example of
a manifold of this type. (We’ll have more to say about tangent spaces in Section 4.2.)
(5) Product manifolds: For 𝑖 = 1, 2, let 𝑋𝑖 be an 𝑛𝑖 -dimensional submanifold of 𝐑𝑁𝑖 . Then
the Cartesian product of 𝑋1 and 𝑋2
𝑋1 × 𝑋2 = { (𝑥1 , 𝑥2 ) | 𝑥𝑖 ∈ 𝑋𝑖 }
is an (𝑛1 + 𝑛2 )-dimensional submanifold of 𝐑𝑁1 +𝑁2 = 𝐑𝑁1 × 𝐑𝑁2 .
We will leave for you to verify this fact as an exercise.
Hint: For 𝑝𝑖 ∈ 𝑋𝑖 , 𝑖 = 1, 2, there exists a neighborhood, 𝑉𝑖 , of 𝑝𝑖 in 𝐑𝑁𝑖 , an open set,
𝑈𝑖 in 𝐑𝑛𝑖 , and a diffeomorphism 𝜙 ∶ 𝑈𝑖 → 𝑋𝑖 ∩ 𝑉𝑖 . Let 𝑈 = 𝑈1 × 𝑈2 , 𝑉 = 𝑉1 × 𝑉2 and
𝑋 = 𝑋1 × 𝑋2 , and let 𝜙 ∶ 𝑈 → 𝑋 ∩ 𝑉 be the product diffeomorphism, (𝜙(𝑞1 ), 𝜙2 (𝑞2 )).
(6) The unit 𝑛-sphere: This is the set of unit vectors in 𝐑𝑛+1 :
𝑆𝑛 = { 𝑥 ∈ 𝐑𝑛+1 | 𝑥12 + ⋯ + 𝑥𝑛+1
2
= 1} .
To show that 𝑆𝑛 is an 𝑛-manifold, let 𝑉 be the open subset of 𝐑𝑛+1 on which 𝑥𝑛+1 is
positive. If 𝑈 is the open unit ball in 𝐑𝑛 and 𝑓 ∶ 𝑈 → 𝐑 is the function, 𝑓(𝑥) = (1 − (𝑥12 +
⋯ + 𝑥𝑛2 ))1/2 , then 𝑆𝑛 ∩ 𝑉 is just the graph 𝛤𝑓 of 𝑓. Hence we have a diffeomorphism
𝜙 ∶ 𝑈 → 𝑆𝑛 ∩ 𝑉 .
More generally, if 𝑝 = (𝑥1 , …, 𝑥𝑛+1 ) is any point on the unit sphere, then 𝑥𝑖 is nonzero
for some 𝑖. If 𝑥𝑖 is positive, then letting 𝜎 be the transposition, 𝑖 ↔ 𝑛+1 and 𝑓𝜎 ∶ 𝐑𝑛+1 →
𝐑𝑛+1 , the map
𝑓𝜎 (𝑥1 , …, 𝑥𝑛 ) = (𝑥𝜎(1) , …, 𝑥𝜎(𝑛) )
one gets a diffeomorphism 𝑓𝜎 ∘𝜙 of 𝑈 onto a neighborhood of 𝑝 in 𝑆𝑛 and if 𝑥𝑖 is negative
one gets such a diffeomorphism by replacing 𝑓𝜎 by −𝑓𝜎 . In either case we have shown
that for every point 𝑝 in 𝑆𝑛 , there is a neighborhood of 𝑝 in 𝑆𝑛 which is diffeomorphic
to 𝑈.
 Draft: March 28, 2018

§4.1 Manifolds 99

(7) The 2-torus: In calculus books this is usually described as the surface of rotation in 𝐑3
obtained by taking the unit circle centered at the point, (2, 0), in the (𝑥1 , 𝑥3 ) plane and
rotating it about the 𝑥3 -axis. However, a slightly nicer description of it is as the product
manifold 𝑆1 × 𝑆1 in 𝐑4 . (As an exercise, reconcile these two descriptions.)
We’ll now turn to an alternative way of looking at manifolds: as solutions of systems of
equations. Let 𝑈 be an open subset of 𝐑𝑁 and 𝑓 ∶ 𝑈 → 𝐑𝑘 a 𝐶∞ map.

Definition 4.1.6. A point 𝑎 ∈ 𝐑𝑘 is a regular value of 𝑓 if for every point 𝑝 ∈ 𝑓−1 (𝑎), the
map 𝑓 is a submersion at 𝑝.

Note that for 𝑓 to be a submersion at 𝑝, the differential 𝐷𝑓(𝑝) ∶ 𝐑𝑁 → 𝐑𝑘 has to be

surjective, and hence 𝑘 has to be less than or equal to 𝑁. Therefore this notion of regular
value is interesting only if 𝑁 ≥ 𝑘.
Theorem 4.1.7. Let 𝑛 ≔ 𝑁−𝑘. If 𝑎 is a regular value of 𝑓, the set 𝑋 ≔ 𝑓−1 (𝑎) is an 𝑛-manifold.
Proof. Replacing 𝑓 by 𝜏−𝑎 ∘ 𝑓 we can assume without loss of generality that 𝑎 = 0. Let 𝑝 ∈
𝑓−1 (0). Since 𝑓 is a submersion at 𝑝, the canonical submersion theorem (see Theorem b.16)
tells us that there exists a neighborhood 𝑂 of 0 in 𝐑𝑁 , a neighborhood 𝑈0 , of 𝑝 in 𝑈 and a
diffeomorphism 𝑔 → 𝑂 ⥲ 𝑈0 such that
(4.1.8) 𝑓∘𝑔=𝜋
where 𝜋 is the projection map
𝐑𝑁 = 𝐑𝑘 × 𝐑𝑛 → 𝐑𝑘 , (𝑥, 𝑦) ↦ 𝑥 .
Hence 𝜋−1 (0) = {0}×𝐑𝑛 = 𝐑𝑛 and by equation (4.1.5)), 𝑔 maps 𝑂∩𝜋−1 (0) diffeomorphically
onto 𝑈0 ∩ 𝑓−1 (0). However, 𝑂 ∩ 𝜋−1 (0) is a neighborhood of 0 in 𝐑𝑛 and 𝑈0 ∩ 𝑓−1 (0) is a
neighborhood of 𝑝 in 𝑋, and, as remarked, these two neighborhoods are diffeomorphic. □

Some examples:
Examples 4.1.9.
(1) The 𝑛-sphere: Let
𝑓 ∶ 𝐑𝑛+1 → 𝐑
be the map,
(𝑥1 , …, 𝑥𝑛+1 ) ↦ 𝑥12 + ⋯ + 𝑥𝑛+1
2
−1.
Then
𝐷𝑓(𝑥) = 2(𝑥1 , …, 𝑥𝑛+1 )
so, if 𝑥 ≠ 0, then 𝑓 is a submersion at 𝑥. In particular, 𝑓 is a submersion at all points 𝑥
on the 𝑛-sphere
𝑆𝑛 = 𝑓−1 (0)
so the 𝑛-sphere is an 𝑛-dimensional submanifold of 𝐑𝑛+1 .
(2) Graphs: Let 𝑔 ∶ 𝐑𝑛 → 𝐑𝑘 be a 𝐶∞ map and, as in (2), let
𝛤𝑔 ≔ { (𝑥, 𝑦) ∈ 𝐑𝑛 × 𝐑𝑘 | 𝑦 = 𝑔(𝑥) } .

We claim that 𝛤𝑔 is an 𝑛-dimensional submanifold of 𝐑𝑛+𝑘 = 𝐑𝑛 × 𝐑𝑘 .

 Draft: March 28, 2018

100 Chapter 4: Manifolds & Forms on Manifolds

Proof. Let
𝑓 ∶ 𝐑 𝑛 × 𝐑𝑘 → 𝐑 𝑘
be the map 𝑓(𝑥, 𝑦) ≔ 𝑦 − 𝑔(𝑥). Then
𝐷𝑓(𝑥, 𝑦) = (−𝐷𝑔(𝑥) id𝑘 )
where id𝑘 is the identity map of 𝐑𝑘 onto itself. This map is always of rank 𝑘. Hence
𝛤𝑓 = 𝑓−1 (0) is an 𝑛-dimensional submanifold of 𝐑𝑛+𝑘 . □
(3) Let M 𝑛 be the set of all 𝑛 × 𝑛 matrices and let 𝒮𝑛 be the set of all symmetric 𝑛 × 𝑛
matrices, i.e., the set
𝒮𝑛 = { 𝐴 ∈ M 𝑛 | 𝐴 = 𝐴⊺ } .
The map
𝐴 = (𝑎𝑖,𝑗 ) ↦ (𝑎1,1 , 𝑎1,2 , …, 𝑎1,𝑛 , 𝑎2,1 , …, 𝑎2,𝑛 , …)
gives us an identification
2
M 𝑛 ⥲ 𝐑𝑛
and the map
𝐴 = (𝑎𝑖,𝑗 ) ↦ (𝑎1,1 , …, 𝑎1,𝑛 , 𝑎2,2 , …, 𝑎2,𝑛 , 𝑎3,3 , …, 𝑎3,𝑛 , …)
gives us an identification
𝑛(𝑛+1)
𝒮𝑛 ⥲ 𝐑 2 .
(Note that if 𝐴 is a symmetric matrix, then 𝑎𝑖,𝑗 = 𝑎𝑗,𝑖 , so this map avoids redundancies.)
Let
𝑂(𝑛) = { 𝐴 ∈ M 𝑛 | 𝐴⊺ 𝐴 = id𝑛 } .
This is the set of orthogonal 𝑛 × 𝑛 matrices, and we will leave for you as an exercise to
show that it’s an 𝑛(𝑛 − 1)/2-manifold.
Hint: Let 𝑓 ∶ M 𝑛 → 𝒮𝑛 be the map 𝑓(𝐴) = 𝐴⊺ 𝐴 − id𝑛 . Then
𝑂(𝑛) = 𝑓−1 (0) .
These examples show that lots of interesting manifolds arise as zero sets of submersions
𝑓 ∶ 𝑈 → 𝐑𝑘 . This is, in fact, not just an accident. We will show that locally every manifold
arises this way. More explicitly let 𝑋 ⊂ 𝐑𝑁 be an 𝑛-manifold, 𝑝 a point of 𝑋, 𝑈 a neighbor-
hood of 0 in 𝐑𝑛 , 𝑉 a neighborhood of 𝑝 in 𝐑𝑁 and 𝜙 ∶ (𝑈, 0) ⥲ (𝑉 ∩ 𝑋, 𝑝) a diffeomor-
phism. We will for the moment think of 𝜙 as a 𝐶∞ map 𝜙 ∶ 𝑈 → 𝐑𝑁 whose image happens
to lie in 𝑋.
Lemma 4.1.10. The linear map 𝐷𝜙(0) ∶ 𝐑𝑛 → 𝐑𝑁 is injective.
Proof. Since 𝜙−1 ∶ 𝑉 ∩ 𝑋 ⥲ 𝑈 is a diffeomorphism, shrinking 𝑉 if necessary, we can assume
that there exists a 𝐶∞ map 𝜓 ∶ 𝑉 → 𝑈 which coincides with 𝜙−1 on 𝑉 ∩ 𝑋. Since 𝜙 maps 𝑈
onto 𝑉 ∩ 𝑋, 𝜓 ∘ 𝜙 = 𝜙−1 ∘ 𝜙 is the identity map on 𝑈. Therefore,
𝐷(𝜓 ∘ 𝜙)(0) = (𝐷𝜓)(𝑝)𝐷𝜙(0) = id𝑛
by the chain rule, and hence if 𝐷𝜙(0)𝑣 = 0, it follows from this identity that 𝑣 = 0. □
Lemma 4.1.10 says that 𝜙 is an immersion at 0, so by the canonical immersion theorem
(see Theorem b.18) there exists a neighborhood 𝑈0 of 0 in 𝑈, a neighborhood 𝑉𝑝 of 𝑝 in 𝑉,
and a diffeomorphism
(4.1.11) 𝑔 ∶ (𝑉𝑝 , 𝑝) ⥲ (𝑈0 × 𝐑𝑁−𝑛 , 0)
 Draft: March 28, 2018

§4.1 Manifolds 101

such that 𝑔 ∘ 𝜙 = 𝜄, where 𝜄 is, as in Appendix b, the canonical immersion

(4.1.12) 𝜄 ∶ 𝑈0 → 𝑈0 × 𝐑𝑁−𝑛 , 𝑥 ↦ (𝑥, 0) .
By (4.1.11) 𝑔 maps 𝜙(𝑈0 ) diffeomorphically onto 𝜄(𝑈0 ). However, by (4.1.8) and (4.1.11)
𝜄(𝑈0 ) is defined by the equations, 𝑥𝑖 = 0, 𝑖 = 𝑛 + 1, …, 𝑁. Hence if 𝑔 = (𝑔1 , …, 𝑔𝑁 ) the set,
𝜙(𝑈0 ) = 𝑉𝑝 ∩ 𝑋 is defined by the equations
(4.1.13) 𝑔𝑖 = 0 , 𝑖 = 𝑛 + 1, …, 𝑁 .
Let ℓ = 𝑁 − 𝑛, let
𝜋 ∶ 𝐑 𝑁 = 𝐑 𝑛 × 𝐑ℓ → 𝐑 ℓ
be the canonical submersion,
𝜋(𝑥1 , …, 𝑥𝑁 ) = (𝑥𝑛+1 , …, 𝑥𝑁 )
and let 𝑓 = 𝜋 ∘ 𝑔. Since 𝑔 is a diffeomorphism, 𝑓 is a submersion and (4.1.12) can be
interpreted as saying that
(4.1.14) 𝑉𝑝 ∩ 𝑋 = 𝑓−1 (0) .
Thus to summarize we have proved:
Theorem 4.1.15. Let 𝑋 be an 𝑛-dimensional submanifold of 𝐑𝑁 and let ℓ ≔ 𝑁 − 𝑛. Then for
every 𝑝 ∈ 𝑋 there exists a neighborhood 𝑉𝑝 of 𝑝 in 𝐑𝑁 and a submersion
𝑓 ∶ (𝑉𝑝 , 𝑝) → (𝐑ℓ , 0)
such that 𝑋 ∩ 𝑉𝑝 is defined by equation (4.1.13).
A nice way of thinking about Theorem 4.1.2 is in terms of the coordinates of the map-
ping, 𝑓. More specifically if 𝑓 = (𝑓1 , …, 𝑓𝑘 ) we can think of 𝑓−1 (𝑎) as being the set of
solutions of the system of equations
𝑓𝑖 (𝑥) = 𝑎𝑖 , 𝑖 = 1, …, 𝑘
and the condition that 𝑎 be a regular value of 𝑓 can be interpreted as saying that for every
solution 𝑝 of this system of equations the vectors
𝑛
𝜕𝑓𝑖
(𝑑𝑓𝑖 )𝑝 = ∑ (0)𝑑𝑥𝑗
𝑗=1 𝜕𝑥𝑗

in 𝑇𝑝⋆ 𝐑𝑛 are linearly independent, i.e., the system (4.1.14) is an “independent system of
defining equations” for 𝑋.

Exercises for §4.1

Exercise 4.1.i. Show that the set of solutions of the system of equations
𝑥12 + ⋯ + 𝑥𝑛2 = 1
{
𝑥1 + ⋯ + 𝑥𝑛 = 0
is an (𝑛 − 2)-dimensional submanifold of 𝐑𝑛 .
Exercise 4.1.ii. Let 𝑆𝑛−1 ⊂ 𝐑𝑛 be the (𝑛 − 1)-sphere and let
𝑋𝑎 = { 𝑥 ∈ 𝑆𝑛−1 | 𝑥1 + ⋯ + 𝑥𝑛 = 𝑎 } .
For what values of 𝑎 is 𝑋𝑎 an (𝑛 − 2)-dimensional submanifold of 𝑆𝑛−1 ?
 Draft: March 28, 2018

102 Chapter 4: Manifolds & Forms on Manifolds

Exercise 4.1.iii. Show that if 𝑋𝑖 is an 𝑛𝑖 -dimensional submanifold of 𝐑𝑁𝑖 , for 𝑖 = 1, 2, then

𝑋1 × 𝑋2 ⊂ 𝐑𝑁1 × 𝐑𝑁2
is an (𝑛1 + 𝑛2 )-dimensional submanifold of 𝐑𝑁1 × 𝐑𝑁2 .
Exercise 4.1.iv. Show that the set
𝑋 = { (𝑥, 𝑣) ∈ 𝑆𝑛−1 × 𝐑𝑛 | 𝑥 ⋅ 𝑣 = 0 }
𝑛
is a (2𝑛 − 2)-dimensional submanifold of 𝐑𝑛 × 𝐑𝑛 . (Here 𝑥 ⋅ 𝑣 ≔ ∑𝑖=1 𝑥𝑖 𝑣𝑖 is the dot product.)
Exercise 4.1.v. Let 𝑔 ∶ 𝐑𝑛 → 𝐑𝑘 be a 𝐶∞ map and let 𝑋 = 𝛤𝑔 be the graph of 𝑔. Prove
directly that 𝑋 is an 𝑛-manifold by proving that the map
𝛾𝑔 ∶ 𝐑𝑛 → 𝑋 , 𝑥 ↦ (𝑥, 𝑔(𝑥))
is a diffeomorphism.
Exercise 4.1.vi. Prove that the orthogonal group 𝑂(𝑛) is an 𝑛(𝑛 − 1)/2-manifold.
Hints:
• Let 𝑓 ∶ M 𝑛 → 𝒮𝑛 be the map
𝑓(𝐴) = 𝐴⊺ 𝐴 − id𝑛 .
Show that 𝑂(𝑛) = 𝑓−1 (0).
• Show that
𝑓(𝐴 + 𝜀𝐵) = 𝐴⊺ 𝐴 + 𝜀(𝐴⊺ 𝐵 + 𝐵⊺ 𝐴) + 𝜀2 𝐵⊺ 𝐵 − id𝑛 .
• Conclude that the derivative of 𝑓 at 𝐴 is the map given by
(4.1.16) 𝐵 ↦ 𝐴⊺ 𝐵 + 𝐵⊺ 𝐴 .
• Let 𝐴 be in 𝑂(𝑛). Show that if 𝐶 is in 𝒮𝑛 and 𝐵 = 𝐴𝐶/2 then the map (4.1.16)
maps 𝐵 onto 𝐶.
• Conclude that the derivative of 𝑓 is surjective at 𝐴.
• Conclude that 0 is a regular value of the mapping 𝑓.
The next five exercises, which are somewhat more demanding than the exercises above,
are an introduction to Grassmannian geometry.
Exercise 4.1.vii.
(1) Let 𝑒1 , …, 𝑒𝑛 be the standard basis of 𝐑𝑛 and let 𝑊 = span(𝑒𝑘+1 , …, 𝑒𝑛 ). Prove that if 𝑉
is a 𝑘-dimensional subspace of 𝐑𝑛 and
(4.1.17) 𝑉∩𝑊=0,
then one can find a unique basis of 𝑉 of the form
ℓ
(4.1.18) 𝑣𝑖 = 𝑒𝑖 + ∑ 𝑏𝑖,𝑗 𝑒𝑘+𝑗 , 𝑖 = 1, …, 𝑘 ,
𝑗=1

where ℓ = 𝑛 − 𝑘.
(2) Let 𝐺𝑘 be the set of 𝑘-dimensional subspaces of 𝐑𝑛 having the property (4.1.17) and
let M 𝑘,ℓ be the vector space of 𝑘 × ℓ matrices. Show that one gets from the identities
(4.1.18) a bijective map:
(4.1.19) 𝛾 ∶ M 𝑘,ℓ → 𝐺𝑘 .
Exercise 4.1.viii. Let 𝑆𝑛 be the vector space of linear mappings of 𝐑𝑛 into itself which are
self-adjoint, i.e., have the property 𝐴 = 𝐴⊺ .
 Draft: March 28, 2018

§4.1 Manifolds 103

(1) Given a 𝑘-dimensional subspace, 𝑉 of 𝐑𝑛 let 𝜋𝑉 ∶ 𝐑𝑛 → 𝐑𝑛 be the orthogonal projection

of 𝐑𝑛 onto 𝑉. Show that 𝜋𝑉 is in 𝑆𝑛 and is of rank 𝑘, and show that (𝜋𝑉 )2 = 𝜋𝑉 .
(2) Conversely suppose 𝐴 is an element of 𝑆𝑛 which is of rank 𝑘 and has the property, 𝐴2 =
𝐴. Show that if 𝑉 is the image of 𝐴 in 𝐑𝑛 , then 𝐴 = 𝜋𝑉 .
Definition 4.1.20. We call an 𝐴 ∈ 𝑆𝑛 of the form 𝐴 = 𝜋𝑉 above a rank 𝑘 projection operator.
Exercise 4.1.ix. Composing the map
𝜌 ∶ 𝐺𝑘 → 𝑆𝑛 , 𝑉 ↦ 𝜋𝑉
with the map (4.1.19) we get a map
𝜙 ∶ M 𝑘,ℓ → 𝑆𝑛 , 𝜙 = 𝜌 ⋅ 𝛾 .
∞
Prove that 𝜙 is 𝐶 .
Hints:
• By Gram–Schmidt one can convert (4.1.18) into an orthonormal basis
𝑒1,𝐵 , …, 𝑒𝑛,𝐵
of 𝑉. Show that the 𝑒𝑖,𝐵 ’s are 𝐶∞ functions of the matrix 𝐵 = (𝑏𝑖,𝑗 ).
• Show that 𝜋𝑉 is the linear mapping
𝑘
𝑣 ↦ ∑(𝑣 ⋅ 𝑒𝑖,𝐵 )𝑒𝑖,𝐵 .
𝑖=1

Exercise 4.1.x. Let 𝑉0 = span(𝑒1 , …, 𝑒𝑘 ) and let 𝐺̃𝑘 = 𝜌(𝐺𝑘 ). Show that 𝜙 maps a neighbor-
hood of 𝑉0 in M 𝑘,ℓ diffeomorphically onto a neighborhood of 𝜋𝑉0 in 𝐺 ̃𝑘 .
̃𝑘 if and only if 𝑉 satisfies (4.1.17). For 1 ≤ 𝑖 ≤ 𝑘 let
Hints: 𝜋𝑉 is in 𝐺
𝑘 ℓ
𝑤𝑖 = 𝜋𝑉 (𝑒𝑖 ) = ∑ 𝑎𝑖,𝑗 𝑒𝑗 + ∑ 𝑐𝑖,𝑟 𝑒𝑘+𝑟 .
𝑗=1 𝑟=1

̃𝑘 .
• Show that if the matrix 𝐴 = (𝑎𝑖,𝑗 ) is invertible, 𝜋𝑉 is in 𝐺
̃
• Let 𝑂 ⊂ 𝐺𝑘 be the set of all 𝜋𝑉 ’s for which 𝐴 is invertible. Show that 𝜙−1 ∶ 𝑂 →
M 𝑘,ℓ is the map
𝜙−1 (𝜋𝑉 ) = 𝐵 = 𝐴−1 𝐶
where 𝐶 = [𝑐𝑖,𝑗 ].
Exercise 4.1.xi. Let 𝐺(𝑘, 𝑛) ⊂ 𝑆𝑛 be the set of rank 𝑘 projection operators. Prove that 𝐺(𝑘, 𝑛)
𝑛(𝑛+1)
is a 𝑘ℓ-dimensional submanifold of the Euclidean space 𝑆𝑛 ≅ 𝐑 2 .
Hints:
• Show that if 𝑉 is any 𝑘-dimensional subspace of 𝐑𝑛 there exists a linear mapping,
𝐴 ∈ 𝑂(𝑛) mapping 𝑉0 to 𝑉.
• Show that 𝜋𝑉 = 𝐴𝜋𝑉0 𝐴−1 .
• Let 𝐾𝐴 ∶ 𝑆𝑛 → 𝑆𝑛 be the linear mapping,
𝐾𝐴 (𝐵) = 𝐴𝐵𝐴−1 .
Show that
𝐾𝐴 ∘ 𝜙 ∶ M 𝑘,ℓ → 𝑆𝑛
maps a neighborhood of 𝑉0 in M 𝑘,ℓ diffeomorphically onto a neighborhood of
𝜋𝑉 in 𝐺(𝑘, 𝑛).
 Draft: March 28, 2018

104 Chapter 4: Manifolds & Forms on Manifolds

Remark 4.1.21. Let Gr(𝑘, 𝑛) be the set of all 𝑘-dimensional subspaces of 𝐑𝑛 . The identifi-
cation of Gr(𝑘, 𝑛) with 𝐺(𝑘, 𝑛) given by 𝑉 ↦ 𝜋𝑉 allows us to restate the result above in the
form.
Theorem 4.1.22. The Grassmannian Gr(𝑘, 𝑛) of 𝑘-dimensional subspaces of 𝐑𝑛 is a 𝑘ℓ-dimens-
𝑛(𝑛+1)
ional submanifold of 𝑆𝑛 ≅ 𝐑 2 .
Exercise 4.1.xii. Show that Gr(𝑘, 𝑛) is a compact submanifold of 𝑆𝑛 .
Hint: Show that it’s closed and bounded.

4.2. Tangent spaces

We recall that a subset 𝑋 of 𝐑𝑁 is an 𝑛-dimensional manifold if for every 𝑝 ∈ 𝑋 there
exists an open set 𝑈 ⊂ 𝐑𝑛 , a neighborhood 𝑉 of 𝑝 in 𝐑𝑁 , and a 𝐶∞ -diffeomorphism 𝜙 ∶ 𝑈 ⥲
𝑋 ∩ 𝑉.
Definition 4.2.1. We will call 𝜙 a parametrization of 𝑋 at 𝑝.
Our goal in this section is to define the notion of the tangent space 𝑇𝑝 𝑋 to 𝑋 at 𝑝 and
describe some of its properties. Before giving our official definition we’ll discuss some simple
examples.
Example 4.2.2. Let 𝑓 ∶ 𝐑 → 𝐑 be a 𝐶∞ function and let 𝑋 = 𝛤𝑓 be the graph of 𝑓.

𝑦
ℓ

𝛤𝑓

𝑝0
𝑥

Figure 4.2.1. The tangent line ℓ to the graph of 𝑓 at 𝑝0

Then in this figure above the tangent line ℓ to 𝑋 at 𝑝0 = (𝑥0 , 𝑦0 ) is defined by the
equation
𝑦 − 𝑦0 = 𝑎(𝑥 − 𝑥0 )
where 𝑎 = 𝑓′ (𝑥0 ) In other words if 𝑝 is a point on ℓ then 𝑝 = 𝑝0 + 𝜆𝑣0 where 𝑣0 = (1, 𝑎)
and 𝜆 ∈ 𝐑. We would, however, like the tangent space to 𝑋 at 𝑝0 to be a subspace of the
tangent space to 𝐑2 at 𝑝0 , i.e., to be the subspace of the space: 𝑇𝑝0 𝐑2 = {𝑝0 } × 𝐑2 , and this
we’ll achieve by defining
𝑇𝑝0 𝑋 ≔ { (𝑝0 , 𝜆𝑣0 ) | 𝜆 ∈ 𝐑 } .
 Draft: March 28, 2018

§4.2 Tangent spaces 105

Example 4.2.3. Let 𝑆2 be the unit 2-sphere in 𝐑3 . The tangent plane to 𝑆2 at 𝑝0 is usually
defined to be the plane
{ 𝑝0 + 𝑣 | 𝑣 ∈ 𝐑3 and 𝑣 ⟂ 𝑝0 } .
However, this tangent plane is easily converted into a subspace of 𝑇𝑝 𝐑3 via the map 𝑝0 +𝑣 ↦
(𝑝0 , 𝑣) and the image of this map
{ (𝑝0 , 𝑣) | 𝑣 ∈ 𝐑3 and 𝑣 ⟂ 𝑝0 } .
will be our definition of 𝑇𝑝0 𝑆2 .
Let’s now turn to the general definition. As above let 𝑋 be an 𝑛-dimensional submani-
fold of 𝐑𝑁 , 𝑝 a point of 𝑋, 𝑉 a neighborhood of 𝑝 in 𝐑𝑁 , 𝑈 an open set in 𝐑𝑛 and
𝜙 ∶ (𝑈, 𝑞) ⥲ (𝑋 ∩ 𝑉, 𝑝)
a parameterization of 𝑋. We can think of 𝜙 as a 𝐶∞ map
𝜙 ∶ (𝑈, 𝑞) → (𝑉, 𝑝)
whose image happens to lie in 𝑋 ∩ 𝑉 and we proved in §4.1 that its derivative at 𝑞
(𝑑𝜙)𝑞 ∶ 𝑇𝑞 𝐑𝑛 → 𝑇𝑝 𝐑𝑁
is injective.
Definition 4.2.4. The tangent space 𝑇𝑝 𝑋 to 𝑋 at 𝑝 is the image of the linear map
(𝑑𝜙)𝑞 ∶ 𝑇𝑞 𝐑𝑛 → 𝑇𝑝 𝐑𝑁 .
In other words, 𝑤 ∈ 𝑇𝑝 𝐑𝑁 is in 𝑇𝑝 𝑋 if and only if 𝑤 = 𝑑𝜙𝑞 (𝑣) for some 𝑣 ∈ 𝑇𝑞 𝐑𝑛 . More
succinctly,
(4.2.5) 𝑇𝑝 𝑋 ≔ (𝑑𝜙𝑞 )(𝑇𝑞 𝐑𝑛 ) .
(Since 𝑑𝜙𝑞 is injective, 𝑇𝑝 𝑋 is an 𝑛-dimensional vector subspace of 𝑇𝑝 𝐑𝑁 .)
One problem with this definition is that it appears to depend on the choice of 𝜙. To
get around this problem, we’ll give an alternative definition of 𝑇𝑝 𝑋. In §4.1 we showed that
there exists a neighborhood 𝑉 of 𝑝 in 𝐑𝑁 (which we can without loss of generality take to
be the same as 𝑉 above) and a 𝐶∞ map
(4.2.6) 𝑓 ∶ (𝑉, 𝑝) → (𝐑𝑘 , 0) , where 𝑘 = 𝑁 − 𝑛 ,
such that 𝑋 ∩ 𝑉 = 𝑓−1 (0) and such that 𝑓 is a submersion at all points of 𝑋 ∩ 𝑉, and in
particular at 𝑝. Thus
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝐑𝑁 → 𝑇0 𝐑𝑘
is surjective, and hence the kernel of 𝑑𝑓𝑝 has dimension 𝑛. Our alternative definition of 𝑇𝑝 𝑋
is
(4.2.7) 𝑇𝑝 𝑋 ≔ ker(𝑑𝑓𝑝 ) .

Lemma 4.2.8. The spaces (4.2.5) and (4.2.7) are both 𝑛-dimensional subspaces of 𝑇𝑝 𝐑𝑁 , and
we claim that these spaces are the same.
(Notice that the definition (4.2.7) of 𝑇𝑝 𝑋 does not depend on 𝜙, so if we can show that
these spaces are the same, the definitions (4.2.5) and (4.2.7)) will depend neither on 𝜙 nor
on 𝑓.)
 Draft: March 28, 2018

106 Chapter 4: Manifolds & Forms on Manifolds

Proof. Since 𝜙(𝑈) is contained in 𝑋 ∩ 𝑉 and 𝑋 ∩ 𝑉 is contained in 𝑓−1 (0), 𝑓 ∘ 𝜙 = 0, so by

the chain rule
(4.2.9) 𝑑𝑓𝑝 ∘ 𝑑𝜙𝑞 = 𝑑(𝑓 ∘ 𝜙)𝑞 = 0 .
𝑛
Hence if 𝑣 ∈ 𝑇𝑝 𝐑 and 𝑤 = 𝑑𝜙𝑞 (𝑣), 𝑑𝑓𝑝 (𝑤) = 0. This shows that the space (4.2.5) is
contained in the space (4.2.7). However, these two spaces are 𝑛-dimensional so they co-
incide. □
From the proof above one can extract a slightly stronger result.
Theorem 4.2.10. Let 𝑊 be an open subset of 𝐑ℓ and ℎ ∶ (𝑊, 𝑞) → (𝐑𝑁 , 𝑝) a 𝐶∞ map. Sup-
pose ℎ(𝑊) is contained in 𝑋. Then the image of the map
𝑑ℎ𝑞 ∶ 𝑇𝑞 𝐑ℓ → 𝑇𝑝 𝐑𝑁
is contained in 𝑇𝑝 𝑋.
Proof. Let 𝑓 be the map (4.2.6). We can assume without loss of generality that ℎ(𝑊) is con-
tained in 𝑉, and so, by assumption, ℎ(𝑊) ⊂ 𝑋 ∩ 𝑉. Therefore, as above, 𝑓 ∘ ℎ = 0, and hence
𝑑ℎ𝑞 (𝑇𝑞 𝐑ℓ ) is contained in the kernel of 𝑑𝑓𝑝 . □
This result will enable us to define the derivative of a mapping between manifolds. Ex-
plicitly: Let 𝑋 be a submanifold of 𝐑𝑁 , 𝑌 a submanifold of 𝐑𝑚 , and 𝑔 ∶ (𝑋, 𝑝) → (𝑌, 𝑦0 )
a 𝐶∞ map. By Definition 4.1.1 there exists a neighborhood 𝑂 of 𝑋 in 𝐑𝑁 and a 𝐶∞ map
𝑔̃ ∶ 𝑂 → 𝐑𝑚 extending to 𝑔. We will define
(4.2.11) (𝑑𝑔)𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑦0 𝑌
to be the restriction of the map
̃ 𝑝 ∶ 𝑇𝑝 𝐑𝑁 → 𝑇𝑦0 𝐑𝑚
(𝑑𝑔)
to 𝑇𝑝 𝑋. There are two obvious problems with this definition:
(1) Is the space (𝑑𝑔)
̃ 𝑝 (𝑇𝑝 𝑋) contained in 𝑇𝑦0 𝑌?
(2) Does the definition depend on 𝑔? ̃
To show that the answer to (1) is yes and the answer to (2) is no, let
𝜙 ∶ (𝑈, 𝑥0 ) ⥲ (𝑋 ∩ 𝑉, 𝑝)
be a parametrization of 𝑋, and let ℎ = 𝑔̃ ∘ 𝜙. Since 𝜙(𝑈) ⊂ 𝑋, ℎ(𝑈) ⊂ 𝑌 and hence by
Theorem 4.2.10
𝑑ℎ𝑥0 (𝑇𝑥0 𝐑𝑛 ) ⊂ 𝑇𝑦0 𝑌 .
But by the chain rule
𝑑ℎ𝑥0 = 𝑑𝑔̃𝑝 ∘ 𝑑𝜙𝑥0 ,
so by (4.2.5)
(𝑑𝑔̃𝑝 )(𝑇𝑝 𝑋) ⊂ 𝑇𝑝 𝑌
and
(𝑑𝑔̃𝑝 )(𝑇𝑝 𝑋) = (𝑑ℎ)𝑥0 (𝑇𝑥0 𝐑𝑛 )
Thus the answer to (1) is yes, and since ℎ = 𝑔̃ ∘ 𝜙 = 𝑔 ∘ 𝜙, the answer to (2) is no.
From equations (4.2.9) and (4.2.11) one easily deduces:
Theorem 4.2.12 (Chain rule for mappings between manifolds). Let 𝑍 be a submanifold of
𝐑ℓ and 𝜓 ∶ (𝑌, 𝑦0 ) → (𝑍, 𝑧0 ) a 𝐶∞ map. Then 𝑑𝜓𝑦0 ∘ 𝑑𝑔𝑝 = 𝑑(𝜓 ∘ 𝑔)𝑝 .
We will next prove manifold versions of the inverse function theorem and the canonical
immersion and submersion theorems.
 Draft: March 28, 2018

§4.2 Tangent spaces 107

Theorem 4.2.13 (Inverse function theorem for manifolds). Let 𝑋 and 𝑌 be 𝑛-manifolds and
𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map. Suppose that at 𝑝 ∈ 𝑋 the map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑞 𝑌 , 𝑞 ≔ 𝑓(𝑝) ,
is bijective. Then 𝑓 maps a neighborhood 𝑈 of 𝑝 in 𝑋 diffeomorphically onto a neighborhood
𝑉 of 𝑞 in 𝑌.
Proof. Let 𝑈 and 𝑉 be open neighborhoods of 𝑝 in 𝑋 and 𝑞 in 𝑌 and let 𝜙0 ∶ (𝑈0 , 0) → (𝑈, 𝑝)
and 𝜓0 ∶ (𝑉0 , 0) → (𝑉, 𝑞) be parametrizations of these neighborhoods. Shrinking 𝑈0 and 𝑈
we can assume that 𝑓(𝑈) ⊂ 𝑉. Let
𝑔 ∶ (𝑈0 , 𝑝0 ) → (𝑉0 , 𝑞0 )
be the map 𝜓0−1 ∘ 𝑓 ∘ 𝜙0 . Then 𝜓0 ∘ 𝑔 = 𝑓 ∘ 𝜙, so by the chain rule
(𝑑𝜓0 )𝑞0 ∘ (𝑑𝑔)𝑝0 = (𝑑𝑓)𝑝 ∘ (𝑑𝜙0 )𝑝0 .
Since (𝑑𝜓0 )𝑞0 and (𝑑𝜙0 )𝑝0 are bijective it’s clear from this identity that if 𝑑𝑓𝑝 is bijective
the same is true for (𝑑𝑔)𝑝0 . Hence by the inverse function theorem for open subsets of 𝐑𝑛 ,
𝑔 maps a neighborhood of 𝑝0 in 𝑈0 diffeomorphically onto a neighborhood of 𝑞0 in 𝑉0 .
Shrinking 𝑈0 and 𝑉0 we assume that these neighborhoods are 𝑈0 and 𝑉0 and hence that 𝑔 is
a diffeomorphism. Thus since 𝑓 ∶ 𝑈 → 𝑉 is the map 𝜓0 ∘ 𝑔 ∘ 𝜙−1 0 , it is a diffeomorphism as
well. □

Theorem 4.2.14 (The canonical submersion theorem for manifolds). Let 𝑋 and 𝑌 be man-
ifolds of dimension 𝑛 and 𝑚, where 𝑚 < 𝑛, and let 𝑓 ∶ 𝑋 → 𝑌 be a 𝐶∞ map. Suppose that at
𝑝 ∈ 𝑋 the map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑞 𝑌 , 𝑞 ≔ 𝑓(𝑝) ,
is surjective. Then there exists an open neighborhood 𝑈 of 𝑝 in 𝑋, and open neighborhood, 𝑉
of 𝑓(𝑈) in 𝑌 and parametrizations 𝜙0 ∶ (𝑈0 , 0) → (𝑈, 𝑝) and 𝜓0 ∶ (𝑉0 , 0) → (𝑉, 𝑞) such that
in the square
𝜓0−1 𝑓𝜙0
𝑈0 𝑉0

𝜙0 𝜓0

𝑈 𝑉
𝑓

the map 𝜓0−1 ∘ 𝑓 ∘ 𝜙0 is the canonical submersion 𝜋.

Proof. Let 𝑈 and 𝑉 be open neighborhoods of 𝑝 and 𝑞 and 𝜙0 ∶ (𝑈0 , 0) → (𝑈, 𝑝) and
𝜓0 ∶ (𝑉0 , 0) → (𝑉, 𝑞) be parametrizations of these neighborhoods. Composing 𝜙0 and 𝜓0
with the translations we can assume that 𝑝0 is the origin in 𝐑𝑛 and 𝑞0 the origin in 𝐑𝑚 ,
and shrinking 𝑈 we can assume 𝑓(𝑈) ⊂ 𝑉. As above let 𝑔 ∶ (𝑈0 , 0) → (𝑉0 , 0) be the map
𝜓0−1 ∘ 𝑓 ∘ 𝜙0 . By the chain rule
(𝑑𝜓0 )0 ∘ (𝑑𝑔)0 = 𝑑𝑓𝑝 ∘ (𝑑𝜙0 )0 ,
therefore, since (𝑑𝜓0 )0 and (𝑑𝜙0 )0 are bijective it follows that (𝑑𝑔)0 is surjective. Hence, by
Theorem 4.1.2, we can find an open neighborhood 𝑈 of the origin in 𝐑𝑛 and a diffeomor-
phism 𝜙1 ∶ (𝑈1 , 0) ⥲ (𝑈0 , 0) such that 𝑔 ∘ 𝜙1 is the canonical submersion. Now replace 𝑈0
by 𝑈1 and 𝜙0 by 𝜙0 ∘ 𝜙1 . □
 Draft: March 28, 2018

108 Chapter 4: Manifolds & Forms on Manifolds

Theorem 4.2.15 (The canonical immersion theorem for manifolds). Let 𝑋 and 𝑌 be mani-
folds of dimension 𝑛 and 𝑚, where 𝑛 < 𝑚, and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map. Suppose that at 𝑝 ∈ 𝑋
the map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑞 𝑌 , 𝑞 = 𝑓(𝑝)
is injective. Then there exists an open neighborhood 𝑈 of 𝑝 in 𝑋, an open neighborhood 𝑉 of
𝑓(𝑈) in 𝑌 and parametrizations 𝜙0 ∶ (𝑈0 , 0) → (𝑈, 𝑝) and 𝜓0 ∶ (𝑉0 , 0) → (𝑉, 𝑞) such that in
the square
𝜓0−1 𝑓𝜙0
𝑈0 𝑉0

𝜙0 𝜓0

𝑈 𝑉
𝑓

the map 𝜓0−1 ∘ 𝑓 ∘ 𝜙0 is the canonical immersion 𝜄.

Proof. The proof is identical with the proof of Theorem 4.2.14 except for the last step. In the
last step one converts 𝑔 into the canonical immersion via a map 𝜓1 ∶ (𝑉1 , 0) → (𝑉0 , 0) with
the property 𝑔 ∘ 𝜓1 = 𝜄 and then replaces 𝜓0 by 𝜓0 ∘ 𝜓1 . □

Exercises for §4.2

Exercise 4.2.i. What is the tangent space to the quadric

𝑄 = { (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐑𝑛 | 𝑥𝑛 = 𝑥12 + ⋯ + 𝑥𝑛−1

2
}

at the point (1, 0, …, 0, 1)?

Exercise 4.2.ii. Show that the tangent space to the (𝑛 − 1)-sphere 𝑆𝑛−1 at 𝑝 is the space of
vectors (𝑝, 𝑣) ∈ 𝑇𝑝 𝐑𝑛 satisfying 𝑝 ⋅ 𝑣 = 0.

Exercise 4.2.iii. Let 𝑓 ∶ 𝐑𝑛 → 𝐑𝑘 be a 𝐶∞ map and let 𝑋 = 𝛤𝑓 . What is the tangent space
to 𝑋 at (𝑎, 𝑓(𝑎))?

Exercise 4.2.iv. Let 𝜎 ∶ 𝑆𝑛−1 → 𝑆𝑛−1 be the antipodal map 𝜎(𝑥) = −𝑥. What is the derivative
of 𝜎 at 𝑝 ∈ 𝑆𝑛−1 ?

Exercise 4.2.v. Let 𝑋𝑖 ⊂ 𝐑𝑁𝑖 , 𝑖 = 1, 2, be an 𝑛𝑖 -manifold and let 𝑝𝑖 ∈ 𝑋𝑖 . Define 𝑋 to be the

Cartesian product
𝑋1 × 𝑋2 ⊂ 𝐑𝑁1 × 𝐑𝑁2
and let 𝑝 = (𝑝1 , 𝑝2 ). Show that 𝑇𝑝 𝑋 ≅ 𝑇𝑝1 𝑋1 ⊕ 𝑇𝑝2 𝑋2 .

Exercise 4.2.vi. Let 𝑋 ⊂ 𝐑𝑁 be an 𝑛-manifold and for 𝑖 = 1, 2, let 𝜙𝑖 ∶ 𝑈𝑖 → 𝑋 ∩ 𝑉𝑖 be two

parametrizations. From these parametrizations one gets an overlap diagram

𝜙−1
2 ∘𝜙1
𝜙−1
1 (𝑋 ∩ 𝑉1 ∩ 𝑉2 ) 𝜙−1
2 (𝑋 ∩ 𝑉1 ∩ 𝑉2 )
(4.2.16)
𝜙1 𝜙2
𝑋 ∩ 𝑉1 ∩ 𝑉2 .
 Draft: March 28, 2018

§4.3 Vector fields & differential forms on manifolds 109

(1) Let 𝑝 ∈ 𝑋 ∩ 𝑉 and let 𝑞𝑖 = 𝜙−1

𝑖 (𝑝). Derive from the overlap diagram (4.2.16) an overlap
diagram of linear maps
𝑑(𝜙−1
2 ∘𝜙1 )𝑞1
𝑇𝑞1 𝐑𝑛 𝑇𝑞2 𝐑𝑛

𝑑(𝜙1 )𝑞1 𝑑(𝜙2 )𝑞2

𝑁
𝑇𝑝 𝐑 .

(2) Use overlap diagrams to give another proof that 𝑇𝑝 𝑋 is intrinsically defined.

4.3. Vector fields & differential forms on manifolds

A vector field on an open subset 𝑈 ⊂ 𝐑𝑛 is a function 𝒗 which assigns to each 𝑝 ∈ 𝑈 an
element 𝒗(𝑝) of 𝑇𝑝 𝑈, and a 𝑘-form is a function 𝜔 which assigns to each 𝑝 ∈ 𝑈 an element
𝜔𝑝 of 𝛬𝑘 (𝑇𝑝⋆ 𝑈). These definitions have obvious generalizations to manifolds:
Definition 4.3.1. Let 𝑋 be a manifold. A vector field on 𝑋 is a function 𝒗 which assigns to
each 𝑝 ∈ 𝑋 an element 𝒗(𝑝) of 𝑇𝑝 𝑋. A 𝑘-form is a function 𝜔 which assigns to each 𝑝 ∈ 𝑋
an element 𝜔𝑝 of 𝛬𝑘 (𝑇𝑝⋆ 𝑋).
We’ll begin our study of vector fields and 𝑘-forms on manifolds by showing that, like
their counterparts on open subsets of 𝐑𝑛 , they have nice pullback and pushforward proper-
ties with respect to mappings. Let 𝑋 and 𝑌 be manifolds and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ mapping.
Definition 4.3.2. Given a vector field 𝒗 on 𝑋 and a vector field 𝒘 on 𝑌, we’ll say that 𝒗 and
𝒘 are 𝑓-related if for all 𝑝 ∈ 𝑋 and 𝑞 = 𝑓(𝑝) we have
(4.3.3) (𝑑𝑓)𝑝 𝒗(𝑝) = 𝒘(𝑞) .
In particular, if 𝑓 is a diffeomorphism, and we’re given a vector field 𝒗 on 𝑋 we can
define a vector field 𝒘 on 𝑌 by requiring that for every point 𝑞 ∈ 𝑌, the identity (4.3.3)
holds at the point 𝑝 = 𝑓−1 (𝑞). In this case we’ll call 𝒘 the pushforward of by 𝑓 and denote
it by 𝑓⋆ 𝒗. Similarly, given a vector field 𝒘 on 𝑌 we can define a vector field 𝒗 on 𝑋 by applying
the same construction to the inverse diffeomorphism 𝑓−1 ∶ 𝑌 → 𝑋. We will call the vector
field (𝑓−1 )⋆ 𝒘 the pullback of 𝒘 by 𝑓 (and also denote it by 𝑓⋆ 𝒘).
For differential forms the situation is even nicer. Just as in §2.6 we can define the pull-
back operation on forms for any 𝐶∞ map 𝑓 ∶ 𝑋 → 𝑌. Specifically: Let 𝜔 be a 𝑘-form on 𝑌.
For every 𝑝 ∈ 𝑋, and 𝑞 = 𝑓(𝑝) the linear map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑞 𝑌
induces by (1.8.2)a pullback map
(𝑑𝑓𝑝 )⋆ ∶ 𝛬𝑘 (𝑇𝑞⋆ 𝑌) → 𝛬𝑘 (𝑇𝑝⋆ 𝑋)
and, as in §2.6, we’ll define the pullback 𝑓⋆ 𝜔 of 𝜔 to 𝑋 by defining it at 𝑝 by the identity
(𝑓⋆ 𝜔)(𝑝) = (𝑑𝑓𝑝 )⋆ 𝜔(𝑞) .
The following results about these operations are proved in exactly the same way as in
§2.6.
Proposition 4.3.4. Let 𝑋, 𝑌, and 𝑍 be manifolds and 𝑓 ∶ 𝑋 → 𝑌 and 𝑔 ∶ 𝑌 → 𝑍 𝐶∞ maps.
Then if 𝜔 is a 𝑘-form on 𝑍
𝑓⋆ (𝑔⋆ 𝜔) = (𝑔 ∘ 𝑓)⋆ 𝜔 .
 Draft: March 28, 2018

110 Chapter 4: Manifolds & Forms on Manifolds

If 𝒗 is a vector field on 𝑋 and 𝑓 and 𝑔 are diffeomorphisms, then

(4.3.5) (𝑔 ∘ 𝑓)⋆ 𝑣 = 𝑔⋆ (𝑓⋆ 𝑣) .
Our first application of these identities will be to define what one means by a “𝐶∞ vector
field” and a “𝐶∞ 𝑘-form”.
Definition 4.3.6. Let 𝑋 be an 𝑛-manifold and 𝑈 an open subset of 𝑋. The set 𝑈 is a parametriz-
able open set if there exists an open set 𝑈0 in 𝐑𝑛 and a diffeomorphism 𝜙0 ∶ 𝑈0 → 𝑈.
In other words, 𝑈 is parametrizable if there exists a parametrization having 𝑈 as its im-
age. (Note that 𝑋 being a manifold means that every point is contained in a parametrizable
open set.)
Now let 𝑈 ⊂ 𝑋 be a parametrizable open set and 𝜙0 ∶ 𝑈0 → 𝑈 a parametrization of 𝑈.
Definition 4.3.7. A 𝑘-form 𝜔 on 𝑈 is 𝐶∞ or smooth if 𝜙⋆0 𝜔 is a 𝐶∞ .
This definition appears to depend on the choice of the parametrization 𝜙, but we claim
it does not. To see this let 𝜙1 ∶ 𝑈1 → 𝑈 be another parametrization of 𝑈 and let
𝜓 ∶ 𝑈0 → 𝑈1
be the composite map 𝜙−1
0 ∘ 𝜙0 . Then 𝜙0 = 𝜙1 ∘ 𝜓 and hence by Proposition 4.3.4

𝜙⋆0 𝜔 = 𝜓⋆ 𝜙⋆1 𝜔 ,
so by (2.5.11) 𝜙⋆0 𝜔 is 𝐶∞ if 𝜙⋆1 𝜔 is 𝐶∞ . The same argument applied to 𝜓−1 shows that 𝜙⋆1 𝜔
is 𝐶∞ if 𝜙⋆0 𝜔 is 𝐶∞ .
The notion of “𝐶∞ ” for vector fields is defined similarly:
Definition 4.3.8. A vector field 𝒗 on 𝑈 is smooth, or simply 𝐶∞ , if 𝜙⋆0 𝒗 is 𝐶∞ .
By Proposition 4.3.4 𝜙⋆0 𝒗 = 𝜓⋆ 𝜙⋆1 𝒗, so, as above, this definition is independent of the
choice of parametrization.
We now globalize these definitions.
Definition 4.3.9. A 𝑘-form 𝜔 on 𝑋 is 𝐶∞ if, for every point 𝑝 ∈ 𝑋, 𝜔 is 𝐶∞ on a neigh-
borhood of 𝑝. Similarly, a vector field 𝒗 on 𝑋 is 𝐶∞ if, for every point 𝑝 ∈ 𝑋, 𝒗 is 𝐶∞ on a
neighborhood of𝑝.
We will also use the identities (4.3.5) and (4.3.18) to prove the following two results.
Proposition 4.3.10. Let 𝑋 and 𝑌 be manifolds and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map. Then if 𝜔 is a 𝐶∞
𝑘-form on 𝑌, the pullback 𝑓⋆ 𝜔 is a 𝐶∞ 𝑘-form on 𝑋.
Proof. For 𝑝 ∈ 𝑋 and 𝑞 = 𝑓(𝑝) let 𝜙0 ∶ 𝑈0 → 𝑈 and 𝜓0 ∶ 𝑉0 → 𝑉 be parametrizations with
𝑝 ∈ 𝑈 and 𝑞 ∈ 𝑉. Shrinking 𝑈 if necessary we can assume that 𝑓(𝑈) ⊂ 𝑉. Let 𝑔 ∶ 𝑈0 → 𝑉0
be the map 𝑔 = 𝜓0−1 ∘ 𝑓 ∘ 𝜙0 . Then 𝜓0 ∘ 𝑔 = 𝑓 ∘ 𝜙0 , so 𝑔⋆ 𝜓0⋆ 𝜔 = 𝜙⋆0 𝑓⋆ 𝜔. Since 𝜔 is 𝐶∞ , we
see that 𝜓0⋆ 𝜔 is 𝐶∞ , so by equation (2.6.11) we have 𝑔⋆ 𝜓0⋆ 𝜔 is 𝐶∞ , and hence 𝜙⋆0 𝑓⋆ 𝜔 is 𝐶∞ .
Thus by definition 𝑓⋆ 𝜔 is 𝐶∞ on 𝑈. □
By exactly the same argument one proves.
Proposition 4.3.11. If 𝒘 is a 𝐶∞ vector field on 𝑌 and 𝑓 is a diffeomorphism, then 𝑓⋆ 𝒘 is a
𝐶∞ vector field on 𝑋.
Notation 4.3.12.
(1) We’ll denote the space of 𝐶∞ 𝑘-forms on 𝑋 by 𝛺𝑘 (𝑋) .
 Draft: March 28, 2018

§4.3 Vector fields & differential forms on manifolds 111

(2) For 𝜔 ∈ 𝛺𝑘 (𝑋) we’ll define the support of 𝜔 by

supp(𝜔) ≔ { 𝑝 ∈ 𝑋 | 𝜔(𝑝) ≠ 0 }
and we’ll denote by 𝛺𝑘𝑐 (𝑋) ⊂ 𝛺𝑘 (𝑋) the space of compactly supported 𝑘-forms.
(3) For a vector field 𝒗 on 𝑋 we’ll define the support of 𝒗 to be the set
supp(𝒗) ≔ { 𝑝 ∈ 𝑋 | 𝒗(𝑝) ≠ 0 } .
We will now review some of the results about vector fields and the differential forms
that we proved in Chapter 2 and show that they have analogues for manifolds.
Integral curves
Let 𝐼 ⊂ 𝐑 be an open interval and 𝛾 ∶ 𝐼 → 𝑋 a 𝐶∞ curve. For 𝑡0 ∈ 𝐼 we will call
𝑢⃗ = (𝑡0 , 1) ∈ 𝑇𝑡0 𝐑 the unit vector in 𝑇𝑡0 𝐑 and if 𝑝 = 𝛾(𝑡0 ) we will call the vector
𝑑𝛾𝑡0 (𝑢)⃗ ∈ 𝑇𝑝 𝑋
the tangent vector to 𝛾 at 𝑝. If 𝒗 is a vector field on 𝑋 we will say that 𝛾 is an integral curve
of 𝒗 if for all 𝑡0 ∈ 𝐼
𝒗(𝛾(𝑡0 )) = 𝑑𝛾𝑡0 (𝑢)⃗ .
Proposition 4.3.13. Let 𝑋 and 𝑌 be manifolds and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map. If 𝒗 and 𝒘 are
vector fields on 𝑋 and 𝑌 which are 𝑓-related, then integral curves of 𝒗 get mapped by 𝑓 onto
integral curves of 𝒘.
Proof. If the curve 𝛾 ∶ 𝐼 → 𝑋 is an integral curve of 𝒗 we have to show that 𝑓 ∘ 𝛾 ∶ 𝐼 → 𝑌 is
an integral curve of 𝒘. If 𝛾(𝑡) = 𝑝 and 𝑞 = 𝑓(𝑝) then by the chain rule
𝒘(𝑞) = 𝑑𝑓𝑝 (𝒗(𝑝)) = 𝑑𝑓𝑝 (𝑑𝛾𝑡 (𝑢))
⃗
= 𝑑(𝑓 ∘ 𝛾)𝑡 (𝑢)⃗ . □
From this result it follows that the local existence, uniqueness and “smooth dependence
on initial data” results about vector fields that we described in §2.1 are true for vector fields
on manifolds. More explicitly, let 𝑈 be a parametrizable open subset of 𝑋 and 𝜙 ∶ 𝑈0 → 𝑈
a parametrization. Since 𝑈0 is an open subset of 𝐑𝑛 these results are true for the vector field
𝒘 = 𝜙⋆0 𝒗 and hence since 𝒘 and 𝒗 are 𝜙0 -related they are true for 𝒗. In particular:
Proposition 4.3.14 (local existence). For every 𝑝 ∈ 𝑈 there exists an integral curve 𝛾(𝑡)
defined for −𝜀 < 𝑡 < 𝜀, of 𝒗 with 𝛾(0) = 𝑝.
Proposition 4.3.15 (local uniqueness). For 𝑖 = 1, 2, let 𝛾𝑖 ∶ 𝐼𝑖 → 𝑈 be integral curves of 𝒗
and let 𝐼 = 𝐼1 ∩ 𝐼2 . Suppose 𝛾2 (𝑡) = 𝛾1 (𝑡) for some 𝑡 ∈ 𝐼. Then there exists a unique integral
curve, 𝛾 ∶ 𝐼1 ∪ 𝐼2 → 𝑈 with 𝛾 = 𝛾1 on 𝐼1 and 𝛾 = 𝛾2 on 𝐼2 .
Proposition 4.3.16 (smooth dependence on initial data). For every 𝑝 ∈ 𝑈, there exists a
neighborhood 𝑂 of 𝑝 in 𝑈, an interval (−𝜀, 𝜀) and a 𝐶∞ map ℎ ∶ 𝑂 × (−𝜀, 𝜀) → 𝑈 such that
for every 𝑝 ∈ 𝑂 the curve
𝛾𝑝 (𝑡) = ℎ(𝑝, 𝑡) , − 𝜀 < 𝑡 < 𝜀 ,
is an integral curve of 𝒗 with 𝛾𝑝 (0) = 𝑝.
As in Chapter 2 we will say that 𝒗 is complete if, for every 𝑝 ∈ 𝑋 there exists an integral
curve 𝛾(𝑡) defined for −∞ < 𝑡 < ∞, with 𝛾(0) = 𝑝. In Chapter 2 we showed that one simple
criterium for a vector field to be complete is that it be compactly supported. We will prove
that the same is true for manifolds.
Theorem 4.3.17. If 𝑋 is compact or, more generally, if 𝒗 is compactly supported, 𝒗 is complete.
 Draft: March 28, 2018

112 Chapter 4: Manifolds & Forms on Manifolds

Proof. It’s not hard to prove this by the same argument that we used to prove this theorem
for vector fields on 𝐑𝑛 , but we’ll give a simpler proof that derives this directly from the 𝐑𝑛
result. Suppose 𝑋 is a submanifold of 𝐑𝑁 . Then for 𝑝 ∈ 𝑋,
𝑇𝑝 𝑋 ⊂ 𝑇𝑝 𝐑𝑁 = { (𝑝, 𝑣) | 𝑣 ∈ 𝐑𝑁 } ,

so 𝒗(𝑝) can be regarded as a pair (𝑝, 𝒗(𝑝)), where 𝒗(𝑝) is in 𝐑𝑁 . Let

(4.3.18) 𝑓𝒗 ∶ 𝑋 → 𝐑𝑁
be the map 𝑓𝒗 (𝑝) = 𝑣(𝑝). It is easy to check that 𝒗 is 𝐶∞ if and only if 𝑓𝒗 is 𝐶∞ . (See
Exercise 4.3.xi.) Hence (see Appendix b) there exists a neighborhood, 𝑂 of 𝑋 and a map
𝑔 ∶ → 𝐑𝑁 extending 𝑓𝒗 . Thus the vector field 𝒘 on 𝑂 defined by 𝒘(𝑞) = (𝑞, 𝑔(𝑞)) extends
the vector field 𝒗 to 𝑂. In other words if 𝜄 ∶ 𝑋 ↪ 𝑂 is the inclusion map 𝒗 and 𝒘 are 𝜄-related.
Thus by Proposition 4.3.13 the integral curves of 𝒗 are just integral curves of 𝒘 that are
contained in 𝑋.
Suppose now that 𝒗 is compactly supported. Then there exists a function 𝜌 ∈ 𝐶0∞ (𝑂)
which is 1 on the support of 𝒗, so, replacing 𝒘 by 𝜌𝒘, we can assume that 𝒘 is compactly
supported. Thus 𝒘 is complete. Let 𝛾(𝑡), defined for −∞ < 𝑡 < ∞, be an integral curve of
𝒘. We will prove that if 𝛾(0) ∈ 𝑋, then this curve is an integral curve of 𝒗. We first observe:
Lemma 4.3.19. The set of points 𝑡 ∈ 𝐑 for which 𝛾(𝑡) ∈ 𝑋 is both open and closed.
Proof of Lemma 4.3.19. If 𝑝 ∉ supp(𝒗) then 𝒘(𝑝) = 0 so if 𝛾(𝑡) = 𝑝, 𝛾(𝑡) is the constant
curve 𝛾 = 𝑝, and there’s nothing to prove. Thus we are reduced to showing that the set
(4.3.20) { 𝑡 ∈ 𝐑 | 𝛾(𝑡) ∈ supp(𝒗) }
is both open and closed. Since supp(𝒗) is compact this set is clearly closed. To show that it’s
open suppose 𝛾(𝑡0 ) ∈ supp(𝒗). By local existence there exist an interval (−𝜀 + 𝑡0 , 𝜀 + 𝑡0 ) and
an integral curve 𝛾1 (𝑡) of 𝒗 defined on this interval and taking the value 𝛾1 (𝑡0 ) = 𝛾(𝑡0 ) at 𝑝.
However since 𝒗 and 𝒘 are 𝜄-related 𝛾1 is also an integral curve of 𝒘 and so it has to coincide
with 𝛾 on the interval (−𝜀 + 𝑡0 , 𝜀 + 𝑡0 ). In particular, for 𝑡 on this interval 𝛾(𝑡) ∈ supp(𝒗) so
the set (4.3.20) is open. □

To conclude the proof of Theorem 4.3.17 we note that since 𝐑 is connected it follows
that if 𝛾(𝑡0 ) ∈ 𝑋 for some 𝑡0 ∈ 𝐑 then 𝛾(𝑡) ∈ 𝑋 for all 𝑡 ∈ 𝐑, and hence 𝛾 is an integral curve
of 𝒗. Thus in particular every integral curve of 𝒗 exists for all time, so 𝒗 is complete. □

Since 𝒘 is complete it generates a one-parameter group of diffeomorphisms 𝑔𝑡 ∶ 𝑂 ⥲ 𝑂,

defined for −∞ < 𝑡 < ∞, having the property that the curve
𝑔𝑡 (𝑝) = 𝛾𝑝 (𝑡) , − ∞ < 𝑡 < ∞
is the unique integral curve of 𝒘 with initial point 𝛾𝑝 (0) = 𝑝. But if 𝑝 ∈ 𝑋 this curve is an
integral curve of 𝒗, so the restriction
𝑓𝑡 = 𝑔𝑡 |𝑋
is a one-parameter group of diffeomorphisms of 𝑋 with theproperty that for 𝑝 ∈ 𝑋 the
curve
𝑓𝑡 (𝑝) = 𝛾𝑝 (𝑡) , − ∞ < 𝑡 < ∞
is the unique integral curve of 𝒗 with initial point 𝛾𝑝 (0) = 𝑝.
 Draft: March 28, 2018

§4.3 Vector fields & differential forms on manifolds 113

The exterior differentiation operation

∞
Let 𝜔 be a 𝐶 𝑘-form on 𝑋 and 𝑈 ⊂ 𝑋 a parametrizable open set. Given a parametriza-
tion 𝜙0 ∶ 𝑈0 → 𝑈 we define the exterior derivative 𝑑𝜔 of 𝜔 on 𝑋 by the formula
(4.3.21) 𝑑𝜔 = (𝜙−1 ⋆ ⋆
0 ) 𝑑𝜙0 𝜔 .
(Notice that since 𝑈0 is an open subset of 𝐑𝑛 and 𝜙⋆0 𝜔 a 𝑘-form on 𝑈0 , the “𝑑” on the right is
well-defined.) We claim that this definition does not depend on the choice of parametriza-
tion. To see this let 𝜙1 ∶ 𝑈1 → 𝑈 be another parametrization of 𝑈 and let 𝜓 ∶ 𝑈0 → 𝑈1 be
the diffeomorphism 𝜙−1 1 ∘ 𝜙0 . Then 𝜙0 = 𝜙1 ∘ 𝜓 and hence
𝑑𝜙⋆0 𝜔 = 𝑑𝜓⋆ 𝜙⋆1 𝜔 = 𝜓⋆ 𝑑𝜙⋆1 𝜔
= 𝜙⋆0 (𝜙−1 ⋆ ⋆
1 ) 𝑑𝜙1 𝜔
so
(𝜙−1 ⋆ ⋆ −1 ⋆
0 ) 𝑑𝜙0 𝜔 = (𝜙1 ) 𝑑𝜙1 𝜔
⋆

as claimed. We can, therefore, define the exterior derivative 𝑑𝜔 globally by defining it to be

equal to (4.3.21) on every parametrizable open set.
It’s easy to see from the definition (4.3.21) that this exterior differentiation operation
inherits from the exterior differentiation operation on open subsets of 𝐑𝑛 the properties
(2.4.3) and (2.4.4) and that for zero forms, i.e., 𝐶∞ functions 𝑓 ∶ 𝑋 → 𝐑, 𝑑𝑓 is the “intrinsic”
𝑑𝑓 defined in Section 2.1, i.e., for 𝑝 ∈ 𝑋, 𝑑𝑓𝑝 is the derivative of 𝑓
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝐑
viewed as an element of 𝛬1 (𝑇𝑝⋆ 𝑋). Let’s check that it also has the property (2.6.12).

Theorem 4.3.22. Let 𝑋 and 𝑌 be manifolds and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map. Then for 𝜔 ∈ 𝛺𝑘 (𝑌)
we have
𝑓⋆ (𝑑𝜔) = 𝑑(𝑓⋆ 𝜔) .
Proof. For every 𝑝 ∈ 𝑋 we’ll check that this equality holds in a neighborhood of 𝑝. Let
𝑞 = 𝑓(𝑝) and let 𝑈 and 𝑉 be parametrizable neighborhoods of 𝑝 and 𝑞. Shrinking 𝑈 if
necessary we can assume 𝑓(𝑈) ⊂ 𝑉. Given parametrizations
𝜙 ∶ 𝑈0 → 𝑈
and
𝜓 ∶ 𝑉0 → 𝑉
we get by composition a map
𝑔 ∶ 𝑈0 → 𝑉0 , 𝑔 = 𝜓−1 ∘ 𝑓 ∘ 𝜙
with the property 𝜓 ∘ 𝑔 = 𝑓 ∘ 𝜙. Thus
𝜙⋆ 𝑑(𝑓⋆ 𝜔) = 𝑑𝜙⋆ 𝑓⋆ 𝜔 (by definition of 𝑑)
⋆
= 𝑑(𝑓 ∘ 𝜙) 𝜔
= 𝑑(𝜓 ∘ 𝑔)⋆ 𝜔
= 𝑑𝑔⋆ (𝜓⋆ 𝜔)
= 𝑔⋆ 𝑑𝜙⋆ 𝜔 (by (2.6.12))
⋆ ⋆
= 𝑔 𝜓 𝑑𝜔 (by definition of 𝑑)
⋆ ⋆
= 𝜙 𝑓 𝑑𝜔 .
⋆
Hence 𝑑𝑓 𝜔 = 𝑓 𝑑𝜔. ⋆
□
 Draft: March 28, 2018

114 Chapter 4: Manifolds & Forms on Manifolds

The interior product and Lie derivative operation

Given a 𝑘-form 𝜔 ∈ 𝛺𝑘 (𝑋) and a 𝐶∞ vector field 𝒗 we will define the interior product
(4.3.23) 𝜄𝒗 𝜔 ∈ 𝛺𝑘−1 (𝑋) ,
as in §2.5, by setting
(𝜄𝒗 𝜔)𝑝 = 𝜄𝒗(𝑝) 𝜔𝑝
𝑘
and the Lie derivative 𝐿𝒗 𝜔 ∈ 𝛺 (𝑋) by setting
(4.3.24) 𝐿𝒗 𝜔 ≔ 𝜄𝒗 𝑑𝜔 + 𝑑𝜄𝒗 𝜔 .
It’s easily checked that these operations satisfy Properties 2.5.3 and Properties 2.5.10 (since,
just as in §2.5, these identities are deduced from the definitions (4.3.23) and (4.3.3) by purely
formal manipulations). Moreover, if 𝒗 is complete and
𝑓𝑡 ∶ 𝑋 → 𝑋 , − ∞ < 𝑡 < ∞
is the one-parameter group of diffeomorphisms of 𝑋 generated by 𝒗 the Lie derivative op-
eration can be defined by the alternative recipe
𝑑 ⋆
(4.3.25) 𝐿𝒗 𝜔 = 𝑓 𝜔|
𝑑𝑡 𝑡 𝑡=0
as in (2.6.22). (Just as in § 2.6 one proves this by showing that the operation (4.3.25) has
satisfies Properties 2.5.10 and hence that it agrees with the operation (4.3.24) provided the
two operations agree on zero-forms.)
Exercises for §4.3
Exercise 4.3.i. Let 𝑋 ⊂ 𝐑 be the paraboloid defined by 𝑥3 = 𝑥12 + 𝑥22 and let 𝒘 be the vector
3

field
𝜕 𝜕 𝜕
𝒘 ≔ 𝑥1 + 𝑥2 + 2𝑥3 .
𝜕𝑥1 𝜕𝑥2 𝜕𝑥3
(1) Show that 𝒘 is tangent to 𝑋 and hence defines by restriction a vector field 𝒗 on 𝑋.
(2) What are the integral curves of 𝒗?
Exercise 4.3.ii. Let 𝑆2 be the unit 2-sphere, 𝑥12 + 𝑥22 + 𝑥32 = 1, in 𝐑3 and let 𝒘 be the vector
field
𝜕 𝜕
𝒘 ≔ 𝑥1 − 𝑥2 .
𝜕𝑥2 𝜕𝑥1
(1) Show that 𝒘 is tangent to 𝑆2 , and hence by restriction defines a vector field 𝒗 on 𝑆2 .
(2) What are the integral curves of 𝒗?
Exercise 4.3.iii. As in Exercise 4.3.ii let 𝑆2 be the unit 2-sphere in 𝐑3 and let 𝒘 be the vector
field
𝜕 𝜕 𝜕 𝜕
𝒘≔ − 𝑥3 (𝑥1 + 𝑥2 + 𝑥3 )
𝜕𝑥3 𝜕𝑥1 𝜕𝑥2 𝜕𝑥3
(1) Show that 𝒘 is tangent to 𝑆2 and hence by restriction defines a vector field 𝒗 on 𝑆2 .
(2) What do its integral curves look like?
Exercise 4.3.iv. Let 𝑆1 be the unit circle, 𝑥12 + 𝑥22 = 1, in 𝐑2 and let 𝑋 = 𝑆1 × 𝑆1 in 𝐑4 with
defining equations
𝑓1 = 𝑥12 + 𝑥22 − 1 = 0
𝑓2 = 𝑥32 + 𝑥42 − 1 = 0 .
 Draft: March 28, 2018

§4.3 Vector fields & differential forms on manifolds 115

(1) Show that the vector field

𝜕 𝜕 𝜕 𝜕
𝒘 ≔ 𝑥1 − 𝑥2 + 𝜆 (𝑥4 − 𝑥3 ),
𝜕𝑥2 𝜕𝑥1 𝜕𝑥3 𝜕𝑥4
𝜆 ∈ 𝐑, is tangent to 𝑋 and hence defines by restriction a vector field 𝒗 on 𝑋.
(2) What are the integral curves of 𝒗?
(3) Show that 𝐿𝒘 𝑓𝑖 = 0.
Exercise 4.3.v. For the vector field 𝒗 in Exercise 4.3.iv, describe the one-parameter group
of diffeomorphisms it generates.
Exercise 4.3.vi. Let 𝑋 and 𝒗 be as in Exercise 4.3.i and let 𝑓 ∶ 𝐑2 → 𝑋 be the map 𝑓(𝑥1 , 𝑥2 ) =
(𝑥1 , 𝑥2 , 𝑥12 + 𝑥22 ). Show that if 𝒖 is the vector field,
𝜕 𝜕
𝒖 ≔ 𝑥1 + 𝑥2 ,
𝜕𝑥1 𝜕𝑥2
then 𝑓⋆ 𝒖 = 𝒗.
Exercise 4.3.vii. Let 𝑋 be a submanifold of 𝑋 in 𝐑𝑁 and let 𝒗 and 𝒘 be the vector fields on
𝑋 and 𝑈. Denoting by 𝜄 the inclusion map of 𝑋 into 𝑈, show that 𝒗 and 𝒘 are 𝜄-related if
and only if 𝒘 is tangent to 𝑋 and its restriction to 𝑋 is 𝒗.
Exercise 4.3.viii. Let 𝑋 be a submanifold of 𝐑𝑁 and 𝑈 an open subset of 𝐑𝑁 containing 𝑋,
and let 𝒗 and 𝒘 be the vector fields on 𝑋 and 𝑈. Denoting by 𝜄 the inclusion map of 𝑋 into
𝑈, show that 𝒗 and 𝒘 are 𝜄-related if and only if 𝒘 is tangent to 𝑋 and its restriction to 𝑋 is
𝒗.
Exercise 4.3.ix. An elementary result in number theory asserts:
Theorem 4.3.26. A number 𝜆 ∈ 𝐑 is irrational if and only if the set
{ 𝑚 + 𝜆𝑛 | 𝑚, 𝑛 ∈ 𝐙 }
is a dense subset of 𝐑.
Let 𝒗 be the vector field in Exercise 4.3.iv. Using Theorem 4.3.26 prove that if 𝜆 is irra-
tional then for every integral curve 𝛾(𝑡), defined for −∞ < 𝑡 < ∞, of 𝒗 the set of points on
this curve is a dense subset of 𝑋.
Exercise 4.3.x. Let 𝑋 be an 𝑛-dimensional submanifold of 𝐑𝑁 . Prove that a vector field 𝒗
on 𝑋 is 𝐶∞ if and only if the map (4.3.18) is 𝐶∞ .
Hint: Let 𝑈 be a parametrizable open subset of 𝑋 and 𝜙 ∶ 𝑈0 → 𝑈 a parametrization of
𝑈. Composing 𝜙 with the inclusion map 𝜄 ∶ 𝑋 → 𝐑𝑁 one gets a map 𝜄 ∘ 𝜙 ∶ 𝑈 → 𝐑𝑁 . Show
that if
𝜕
𝜙⋆ 𝒗 = ∑ 𝑣𝑖
𝜕𝑥𝑗
then
𝜕𝜙
𝜙⋆ 𝑓𝑖 = ∑ 𝑖 𝑣𝑗
𝜕𝑥𝑗
where 𝑓1 , …, 𝑓𝑁 are the coordinates of the map 𝑓𝒗 and 𝜙1 , …, 𝜙𝑁 the coordinates of 𝜄 ∘ 𝜙.
Exercise 4.3.xi. Let 𝒗 be a vector field on 𝑋 and let 𝜙 ∶ 𝑋 → 𝐑 be a 𝐶∞ function. Show that
if the function
𝐿𝒗 𝜙 = 𝜄𝒗 𝑑𝜙
is zero 𝜙 is constant along integral curves of 𝒗.
 Draft: March 28, 2018

116 Chapter 4: Manifolds & Forms on Manifolds

Exercise 4.3.xii. Suppose that 𝜙 ∶ 𝑋 → 𝐑 is proper. Show that if 𝐿𝒗 𝜙 = 0, 𝒗 is complete.

Hint: For 𝑝 ∈ 𝑋 let 𝑎 = 𝜙(𝑝). By assumption, 𝜙−1 (𝑎) is compact. Let 𝜌 ∈ 𝐶0∞ (𝑋) be a
“bump” function which is one on 𝜙−1 (𝑎) and let 𝒘 be the vector field 𝜌𝒗. By Theorem 4.3.17,
𝒘 is complete and since
𝐿𝒘 𝜙 = 𝜄𝜌𝒗 𝑑𝜙 = 𝜌𝜄𝒗 𝑑𝜙 = 0
𝜙 is constant along integral curves of 𝒘. Let 𝛾(𝑡), −∞ < 𝑡 < ∞, be the integral curve of 𝒘
with initial point 𝛾(0) = 𝑝. Show that 𝛾 is an integral curve of 𝒗.

4.4. Orientations
The last part of Chapter 4 will be devoted to the “integral calculus” of forms on mani-
folds. In particular we will prove manifold versions of two basic theorems of integral calculus
on 𝐑𝑛 , Stokes’ theorem and the divergence theorem, and also develop a manifold version
of degree theory. However, to extend the integral calculus to manifolds without getting in-
volved in horrendously technical “orientation” issues we will confine ourselves to a special
class of manifolds: orientable manifolds. The goal of this section will be to explain what this
term means.
Definition 4.4.1. Let 𝑋 be an 𝑛-manifold. An orientation of 𝑋 is a rule for assigning to each
𝑝 ∈ 𝑋 an orientation of 𝑇𝑝 𝑋.
Thus by Definition 1.9.3 one can think of an orientation as a “labeling” rule which, for
every 𝑝 ∈ 𝑋, labels one of the two components of the set 𝛬𝑛 (𝑇𝑝⋆ 𝑋)∖{0} by 𝛬𝑛 (𝑇𝑝⋆ 𝑋)+ , which
we’ll henceforth call the “plus” part of 𝛬𝑛 (𝑇𝑝⋆ 𝑋), and the other component by 𝛬𝑛 (𝑇𝑝⋆ 𝑋)− ,
which we’ll henceforth call the “minus” part of 𝛬𝑛 (𝑇𝑝⋆ 𝑋).
Definition 4.4.2. An orientation of a manifold 𝑋 is smooth, or simply 𝐶∞ , if for every
𝑝 ∈ 𝑋, there exists a neighborhood 𝑈 of 𝑝 and a non-vanishing 𝑛-form 𝜔 ∈ 𝛺𝑛 (𝑈) with the
property
𝜔𝑞 ∈ 𝛬𝑛 (𝑇𝑞⋆ 𝑋)+
for every 𝑞 ∈ 𝑈.
Remark 4.4.3. If we’re given an orientation of 𝑋 we can define another orientation by as-
signing to each 𝑝 ∈ 𝑋 the opposite orientation to the orientation we already assigned, i.e.,
by switching the labels on 𝛬𝑛 (𝑇𝑝⋆ )+ and 𝛬𝑛 (𝑇𝑝⋆ )− . We will call this the reversed orientation
of 𝑋. We will leave for you to check as an exercise that if 𝑋 is connected and equipped with
a smooth orientation, the only smooth orientations of 𝑋 are the given orientation and its
reversed orientation.
Hint: Given any smooth orientation of 𝑋 the set of points where it agrees with the given
orientation is open, and the set of points where it does not is also open. Therefore one of
these two sets has to be empty.
Note that if 𝜔 ∈ 𝛺𝑛 (𝑋) is a non-vanishing 𝑛-form one gets from 𝜔 a smooth orientation
of 𝑋 by requiring that the “labeling rule” above satisfy
𝜔𝑝 ∈ 𝛬𝑛 (𝑇𝑝⋆ 𝑋)+
for every 𝑝 ∈ 𝑋. If 𝜔 has this property we will call 𝜔 a volume form. It’s clear from this defi-
nition that if 𝜔1 and 𝜔2 are volume forms on 𝑋 then 𝜔2 = 𝑓2,1 𝜔1 where 𝑓2,1 is an everywhere
positive 𝐶∞ function.
Example 4.4.4. Let 𝑈 be an open subset 𝐑𝑛 . We will usually assign to 𝑈 its standard orien-
tation, by which we will mean the orientation defined by the 𝑛-form 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
 Draft: March 28, 2018

§4.4 Orientations 117

Example 4.4.5. Let 𝑓 ∶ 𝐑𝑁 → 𝐑𝑘 be a 𝐶∞ map. If zero is a regular value of 𝑓, the set

𝑋 ≔ 𝑓−1 (0) is a submanifold of 𝐑𝑁 of dimension 𝑛 ≔ 𝑁 − 𝑘 (by Theorem 4.2.13). Moreover,
for 𝑝 ∈ 𝑋, 𝑇𝑝 𝑋 is the kernel of the surjective map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝐑𝑁 → 𝑇0 𝐑𝑘
so we get from 𝑑𝑓𝑝 a bijective linear map
(4.4.6) 𝑇𝑝 𝐑𝑁 /𝑇𝑝 𝑋 → 𝑇0 𝐑𝑘 .

As explained in Example 4.4.4, 𝑇𝑝 𝐑𝑁 and 𝑇0 𝐑𝑘 have “standard” orientations, hence if we

require that the map (4.4.6) be orientation preserving, this gives 𝑇𝑝 𝐑𝑁 /𝑇𝑝 𝑋 an orientation
and, by Theorem 1.9.9, gives 𝑇𝑝 𝑋 an orientation. It’s intuitively clear that since 𝑑𝑓𝑝 varies
smoothly with respect to 𝑝 this orientation does as well; however, this fact requires a proof,
and we’ll supply a sketch of such a proof in the exercises.
Example 4.4.7. A special case of Example 4.4.5 is the 𝑛-sphere
𝑆𝑛 = { (𝑥1 , …, 𝑥𝑛+1 ) ∈ 𝐑𝑛+1 | 𝑥12 + ⋯ + 𝑥𝑛+1
2
= 1} ,
which acquires an orientation from its defining map 𝑓 ∶ 𝐑𝑛+1 → 𝐑 given by
𝑓(𝑥) ≔ 𝑥12 + ⋯ + 𝑥𝑛+1
2
−1.
Example 4.4.8. Let 𝑋 be an oriented submanifold of 𝐑𝑁 . For every 𝑝 ∈ 𝑋, 𝑇𝑝 𝑋 sits inside
𝑇𝑝 𝐑𝑁 as a vector subspace, hence, via the identification 𝑇𝑝 𝐑𝑁 ≅ 𝐑𝑁 , one can think of 𝑇𝑝 𝑋
as a vector subspace of 𝐑𝑁 . In particular from the standard Euclidean inner product on 𝐑𝑁
one gets, by restricting this inner product to vectors in 𝑇𝑝 𝑋, an inner product
𝐵𝑝 ∶ 𝑇𝑝 𝑋 × 𝑇𝑝 𝑋 → 𝐑
on 𝑇𝑝 𝑋. Let 𝜎𝑝 be the volume element in 𝛬𝑛 (𝑇𝑝⋆ 𝑋) associated with 𝐵𝑝 (see Exercise 1.9.x)
and let 𝜎 = 𝜎𝑋 be the non-vanishing 𝑛-form on 𝑋 defined by the assignment
𝑝 ↦ 𝜎𝑝 .
In the exercises at the end of this section we’ll sketch a proof of the following.
Theorem 4.4.9. The form 𝜎𝑋 is 𝐶∞ and hence, in particular, is a volume form. (We will call
this form the Riemannian volume form.)
Example 4.4.10 (Möbius strip). The Möbius strip is a surface in 𝐑3 which is not orientable.
It is obtained from the rectangle
𝑅 ≔ [0, 1] × (−1, 1) = { (𝑥, 𝑦) ∈ 𝐑2 | 0 ≤ 𝑥 ≤ 1 and − 1 < 𝑦 < 1 }
by gluing the ends together in the wrong way, i.e., by gluing (1, 𝑦) to (0, −𝑦). It is easy to see
that the Möbius strip cannot be oriented by taking the standard orientation at 𝑝 = (1, 0) and
moving it along the line (𝑡, 0), 0 ≤ 𝑡 ≤ 1 to the point (0, 0) (which is also the point 𝑝 after
we have glued the ends of the rectangle together).
We’ll next investigate the “compatibility” question for diffeomorphisms between ori-
ented manifolds. Let 𝑋 and 𝑌 be 𝑛-manifolds and 𝑓 ∶ 𝑋 ⥲ 𝑌 a diffeomorphism. Suppose
both of these manifolds are equipped with orientations. We will say that 𝑓 is orientation
preserving if for all 𝑝 ∈ 𝑋 and 𝑞 = 𝑓(𝑝), the linear map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑞 𝑌
 Draft: March 28, 2018

118 Chapter 4: Manifolds & Forms on Manifolds

is orientation preserving. It’s clear that if 𝜔 is a volume form on 𝑌 then 𝑓 is orientation

preserving if and only if 𝑓⋆ 𝜔 is a volume form on 𝑋, and from (1.9.5) and the chain rule
one easily deduces the following theorem.
Theorem 4.4.11. If 𝑍 is an oriented 𝑛-manifold and 𝑔 ∶ 𝑌 → 𝑍 a diffeomorphism, then if
both 𝑓 and 𝑔 are orientation preserving, so is 𝑔 ∘ 𝑓.
If 𝑓 ∶ 𝑋 ⥲ 𝑌 is a diffeomorphism then the set of points 𝑝 ∈ 𝑋 at which the linear map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑞 𝑌 , 𝑞 = 𝑓(𝑝) ,
is orientation preserving is open, and the set of points at which is orientation reversing is
open as well. Hence if 𝑋 is connected, 𝑑𝑓𝑝 has to be orientation preserving at all points or
orientation reversing at all points. In the latter case we’ll say that 𝑓 is orientation reversing.
If 𝑈 is a parametrizable open subset of 𝑋 and 𝜙 ∶ 𝑈0 → 𝑈 a parametrization of 𝑈 we’ll
say that this parametrization is an oriented parametrization if 𝜙 is orientation preserving
with respect to the standard orientation of 𝑈0 and the given orientation on 𝑈. Notice that
if this parametrization isn’t oriented we can convert it into one that is by replacing every
connected component 𝑉0 of 𝑈0 on which 𝜙 isn’t orientation preserving by the open set
♯
(4.4.12) 𝑉0 = { (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐑𝑛 | (𝑥1 , …, 𝑥𝑛−1 , −𝑥𝑛 ) ∈ 𝑉0 }
and replacing 𝜙 by the map
(4.4.13) 𝜓(𝑥1 , …, 𝑥𝑛 ) = 𝜙(𝑥1 , …, 𝑥𝑛−1 , −𝑥𝑛 ) .
If 𝜙𝑖 ∶ 𝑈𝑖 → 𝑈, 𝑖 = 0, 1, are oriented parametrizations of 𝑈 and 𝜓 ∶ 𝑈0 → 𝑈1 is the
diffeomorphism 𝜙−11 ∘ 𝜙0 , then by the theorem above 𝜓 is orientation preserving or in other
words
𝜕𝜓𝑖
(4.4.14) det ( )>0
𝜕𝑥𝑗
at every point on 𝑈0 .
We’ll conclude this section by discussing some orientation issues which will come up
when we discuss Stokes’ theorem and the divergence theorem in §4.6. First a definition.
Definition 4.4.15. An open subset 𝐷 of 𝑋 is a smooth domain if
(1) The boundary 𝜕𝐷 is an (𝑛 − 1)-dimensional submanifold of 𝑋, and
(2) The boundary of 𝐷 coincides with the boundary of the closure of 𝐷.
Examples 4.4.16.
(1) The 𝑛-ball, 𝑥12 + ⋯ + 𝑥𝑛2 < 1, whose boundary is the sphere, 𝑥12 + ⋯ + 𝑥𝑛2 = 1.
(2) The 𝑛-dimensional annulus,
1 < 𝑥12 + ⋯ + 𝑥𝑛2 < 2
whose boundary consists of the union of the two spheres

𝑥12 + ⋯ + 𝑥𝑛2 = 1
{
𝑥12 + ⋯ + 𝑥𝑛2 = 2 .

(3) Let 𝑆𝑛−1 be the unit sphere, 𝑥12 + ⋯ + 𝑥22 = 1 and let 𝐷 = 𝐑𝑛 ∖ 𝑆𝑛−1 . Then the boundary
of 𝐷 is 𝑆𝑛−1 but 𝐷 is not a smooth domain since the boundary of its closure is empty.
 Draft: March 28, 2018

§4.4 Orientations 119

(4) The simplest example of a smooth domain is the half-space

𝐇𝑛 = { (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐑𝑛 | 𝑥1 < 0 }
whose boundary
𝜕𝐇𝑛 = { (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐑𝑛 | 𝑥1 = 0 }
we can identify with 𝐑𝑛−1 via the map 𝐑𝑛−1 → 𝜕𝐇𝑛 given by
(𝑥2 , …, 𝑥𝑛 ) ↦ (0, 𝑥2 , …, 𝑥𝑛 ) .
We will show that every bounded domain looks locally like this example.
Theorem 4.4.17. Let 𝐷 be a smooth domain and 𝑝 a boundary point of 𝐷. Then there exists
a neighborhood 𝑈 of 𝑝 in 𝑋, an open set 𝑈0 in 𝐑𝑛 , and a diffeomorphism 𝜓 ∶ 𝑈0 ⥲ 𝑈 such
that 𝜓 maps 𝑈0 ∩ 𝐇𝑛 onto 𝑈 ∩ 𝐷.
Proof. Let 𝑍 ≔ 𝜕𝐷. First we prove the following lemma.
Lemma 4.4.18. For every 𝑝 ∈ 𝑍 there exists an open set 𝑈 in 𝑋 containing 𝑝 and a parametriza-
tion
𝜓 ∶ 𝑈0 → 𝑈
of 𝑈 with the property
(4.4.19) 𝜓(𝑈0 ∩ 𝜕𝐇𝑛 ) = 𝑈 ∩ 𝑍 .
Proof. Since 𝑋 is locally diffeomorphic at 𝑝 to an open subset of 𝐑𝑛 so it suffices to prove
this assertion for 𝑋 = 𝐑𝑛 . However, if 𝑍 is an (𝑛 − 1)-dimensional submanifold of 𝐑𝑛 then
by Theorem 4.2.15 there exists, for every 𝑝 ∈ 𝑍, a neighborhood 𝑈 of 𝑝 in 𝐑𝑛 and a function
𝜙 ∈ 𝐶∞ (𝑈) with the properties
(4.4.20) 𝑥 ∈ 𝑈 ∩ 𝑍 ⟺ 𝜙(𝑥) = 0
and
(4.4.21) 𝑑𝜙𝑝 ≠ 0 .
Without loss of generality we can assume by equation (4.4.21) that
𝜕𝜙
(𝑝) ≠ 0 .
𝜕𝑥1
Hence if 𝜌 ∶ 𝑈 → 𝐑𝑛 is the map
(4.4.22) 𝜌(𝑥1 , …, 𝑥𝑛 ) ≔ (𝜙(𝑥), 𝑥2 , …, 𝑥𝑛 )
(𝑑𝜌)𝑝 is bijective, and hence 𝜌 is locally a diffeomorphism at 𝑝. Shrinking 𝑈 we can assume
that 𝜌 is a diffeomorphism of 𝑈 onto an open set 𝑈0 . By (4.4.20) and (4.4.22) 𝜌 maps 𝑈 ∩ 𝑍
onto 𝑈0 ∩ 𝜕𝐇𝑛 hence if we take 𝜓 to be 𝜌−1 , it will have the property (4.4.19). □
We will now prove Theorem 4.4.9. Without loss of generality we can assume that the
open set 𝑈0 in Lemma 4.4.18 is an open ball with center at 𝑞 ∈ 𝜕𝐇𝑛 and that the diffeomor-
phism 𝜓 maps 𝑞 to 𝑝. Thus for 𝜓−1 (𝑈 ∩ 𝐷) there are three possibilities:
(i) 𝜓−1 (𝑈 ∩ 𝐷) = (𝐑𝑛 − 𝜕𝐇𝑛 ) ∩ 𝑈0 .
𝑛
(ii) 𝜓−1 (𝑈 ∩ 𝐷) = (𝐑𝑛 − 𝐇 ) ∩ 𝑈0 .
−1 𝑛
(iii) 𝜓 (𝑈 ∩ 𝐷) = 𝐇 ∩ 𝑈0 .
However, (i) is excluded by the second hypothesis in Definition 4.4.15 and if (ii) occurs we
can rectify the situation by composing 𝜙 with the map (𝑥1 , …, 𝑥𝑛 ) ↦ (−𝑥1 , 𝑥2 , …, 𝑥𝑛 ). □
 Draft: March 28, 2018

120 Chapter 4: Manifolds & Forms on Manifolds

Definition 4.4.23. We will call an open set 𝑈 with the properties above a 𝐷-adapted parame-
trizable open set.
We will now show that if 𝑋 is oriented and 𝐷 ⊂ 𝑋 is a smooth domain then the bound-
ary 𝑍 ≔ 𝜕𝐷 of 𝐷 acquires from 𝑋 a natural orientation. To see this we first observe:
Lemma 4.4.24. The diffeomorphism 𝜓 ∶ 𝑈0 ⥲ 𝑈 in Theorem 4.4.17 can be chosen to be
orientation preserving.
Proof. If it is not, then by replacing 𝜓 with the diffeomorphism
𝜓♯ (𝑥1 , …, 𝑥𝑛 ) ≔ 𝜓(𝑥1 , …, 𝑥𝑛−1 , −𝑥𝑛 ) ,
we get a 𝐷-adapted parametrization of 𝑈 which is orientation preserving. (See (4.4.12)–
(4.4.13).) □
Let 𝑉0 = 𝑈0 ∩ 𝐑𝑛−1 be the boundary of 𝑈0 ∩ 𝐇𝑛 . The restriction of 𝜓 to 𝑉0 is a diffeomor-
phism of 𝑉0 onto 𝑈 ∩ 𝑍, and we will orient 𝑈 ∩ 𝑍 by requiring that this map be an oriented
parametrization. To show that this is an “intrinsic” definition, i.e., does not depend on the
choice of 𝜓. We’ll prove:
Theorem 4.4.25. If 𝜓𝑖 ∶ 𝑈𝑖 → 𝑈, 𝑖 = 0, 1, are oriented parametrizations of 𝑈 with the property
𝜓𝑖 ∶ 𝑈𝑖 ∩ 𝐇𝑛 → 𝑈 ∩ 𝐷
the restrictions of 𝜓𝑖 to 𝑈𝑖 ∩ 𝐑𝑛−1 induce compatible orientations on 𝑈 ∩ 𝑋.
Proof. To prove this we have to prove that the map 𝜙−1 𝑛
1 ∘ 𝜙0 , restricted to 𝑈 ∩ 𝜕𝐇 is an
𝑛−1 𝑛−1
orientation preserving diffeomorphism of 𝑈0 ∩ 𝐑 onto 𝑈1 ∩ 𝐑 . Thus we have to prove
the following:
Proposition 4.4.26. Let 𝑈0 and 𝑈1 be open subsets of 𝐑𝑛 and 𝑓 ∶ 𝑈0 → 𝑈1 an orientation
preserving diffeomorphism which maps 𝑈0 ∩ 𝐇𝑛 onto 𝑈1 ∩ 𝐇𝑛 . Then the restriction 𝑔 of 𝑓 to
the boundary 𝑈0 ∩ 𝐑𝑛−1 of 𝑈0 ∩ 𝐇𝑛 is an orientation preserving diffeomorphism
𝑔 ∶ 𝑈0 ∩ 𝐑𝑛−1 ⥲ 𝑈1 ∩ 𝐑𝑛−1 .
Let 𝑓(𝑥) = (𝑓1 (𝑥), …, 𝑓𝑛 (𝑥)). By assumption 𝑓1 (𝑥1 , …, 𝑥𝑛 ) is less than zero if 𝑥1 is less
than zero and equal to zero if 𝑥1 is equal to zero, hence
𝜕𝑓1
(4.4.27) (0, 𝑥2 , …, 𝑥𝑛 ) ≥ 0
𝜕𝑥1
and
𝜕𝑓1
(0, 𝑥2 , …, 𝑥𝑛 ) = 0
𝜕𝑥𝑖
for 𝑖 > 1 Moreover, since 𝑔 is the restriction of 𝑓 to the set 𝑥1 = 0
𝜕𝑓𝑖 𝜕𝑔
(0, 𝑥2 , …, 𝑥𝑛 ) = 𝑖 (𝑥2 , …, 𝑥1 )
𝜕𝑥𝑗 𝜕𝑥𝑗
for 𝑖, 𝑗 ≥ 2. Thus on the set defined by 𝑥1 = 0 we have
𝜕𝑓𝑖 𝜕𝑓 𝜕𝑔
(4.4.28) det ( ) = 1 det ( 𝑖 ) .
𝜕𝑥𝑗 𝜕𝑥1 𝜕𝑥𝑗
Since 𝑓 is orientation preserving the left hand side of equation (4.4.28) is positive at all
𝜕𝑓1 𝜕𝑔𝑖
points (0, 𝑥2 , …, 𝑥𝑛 ) ∈ 𝑈0 ∩ 𝐑𝑛−1 hence by (4.4.27) the same is true for 𝜕𝑥 and det ( 𝜕𝑥 ).
1 𝑗
Thus 𝑔 is orientation preserving. □
 Draft: March 28, 2018

§4.4 Orientations 121

Remark 4.4.29. For an alternative proof of this result see Exercise 3.2.viii and Exercise 3.6.iv.
We will now orient the boundary of 𝐷 by requiring that for every 𝐷-adapted parametriz-
able open set 𝑈 the orientation of 𝑍 coincides with the orientation of 𝑈∩𝑍 that we described
above. We will conclude this discussion of orientations by proving a global version of Propo-
sition 4.4.26.
Proposition 4.4.30. For 𝑖 = 1, 2, let 𝑋𝑖 be an oriented manifold, 𝐷𝑖 ⊂ 𝑋𝑖 a smooth domain
and 𝑍𝑖 ≔ 𝜕𝐷𝑖 its boundary. Then if 𝑓 is an orientation preserving diffeomorphism of (𝑋1 , 𝐷1 )
onto (𝑋2 , 𝐷2 ) the restriction 𝑔 of 𝑓 to 𝑍1 is an orientation preserving diffeomorphism of 𝑍1
onto 𝑍2 .
Proof. Let 𝑈 be an open subset of 𝑋1 and 𝜙 ∶ 𝑈0 → 𝑈 an oriented 𝐷1 -compatible parametriza-
tion of 𝑈. Then if 𝑉 = 𝑓(𝑈) the map 𝑓𝜙 ∶ 𝑈 → 𝑉 is an oriented 𝐷2 -compatible parametriza-
tion of 𝑉 and hence 𝑔 ∶ 𝑈 ∩ 𝑍1 → 𝑉 ∩ 𝑍2 is orientation preserving. □
Exercises for §4.4
Exercise 4.4.i. Let 𝑉 be an oriented 𝑛-dimensional vector space, 𝐵 an inner product on 𝑉
and 𝑒1 , …, 𝑒𝑛 ∈ 𝑉 an oriented orthonormal basis. Given vectors 𝑣1 , …, 𝑣𝑛 ∈ 𝑉, show that if
(4.4.31) 𝑏𝑖,𝑗 = 𝐵(𝑣𝑖 , 𝑣𝑗 )
and
𝑛
𝑣𝑖 = ∑ 𝑎𝑗,𝑖 𝑒𝑗 ,
𝑗=1

the matrices 𝐀 = (𝑎𝑖,𝑗 ) and 𝐁 = (𝑏𝑖,𝑗 ) satisfy the identity:

𝐁 = 𝐀⊺ 𝐀
and conclude that det(𝐁) = det(𝐀)2 . (In particular conclude that det(𝐁) > 0.)
Exercise 4.4.ii. Let 𝑉 and 𝑊 be oriented 𝑛-dimensional vector spaces. Suppose that each
of these spaces is equipped with an inner product, and let 𝑒1 , …, 𝑒𝑛 ∈ 𝑉 and 𝑓1 , …, 𝑓𝑛 ∈ 𝑊
be oriented orthonormal bases. Show that if 𝐴 ∶ 𝑊 → 𝑉 is an orientation preserving linear
mapping and 𝐴𝑓𝑖 = 𝑣𝑖 then
1

𝐴⋆ vol𝑉 = (det(𝑏𝑖,𝑗 )) 2 vol𝑊

where vol𝑉 = 𝑒1⋆ ∧ ⋯ ∧ 𝑒𝑛⋆ , vol𝑊 = 𝑓1⋆ ∧ ⋯ ∧ 𝑓𝑛⋆ and (𝑏𝑖,𝑗 ) is the matrix (4.4.31).
Exercise 4.4.iii. Let 𝑋 be an oriented 𝑛-dimensional submanifold of 𝐑𝑛 , 𝑈 an open subset
of 𝑋, 𝑈0 an open subset of 𝐑𝑛 , and 𝜙 ∶ 𝑈0 → 𝑈 an oriented parametrization. Let 𝜙1 , …, 𝜙𝑁
be the coordinates of the map
𝑈0 → 𝑈 ↪ 𝐑𝑁 .
the second map being the inclusion map. Show that if 𝜎 is the Riemannian volume form on
𝑋 then
1

𝜙⋆ 𝜎 = (det(𝜙𝑖,𝑗 )) 2 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛

where
𝑁
𝜕𝜙𝑘 𝜕𝜙𝑘
(4.4.32) 𝜙𝑖,𝑗 = ∑ for 1 ≤ 𝑖, 𝑗 ≤ 𝑛 .
𝑘=1 𝜕𝑥𝑖 𝜕𝑥𝑗

Conclude that 𝜎 is a smooth 𝑛-form and hence that it is a volume form.

 Draft: March 28, 2018

122 Chapter 4: Manifolds & Forms on Manifolds

Hint: For 𝑝 ∈ 𝑈0 and 𝑞 = 𝜙(𝑝) apply Exercise 4.4.ii with 𝑉 = 𝑇𝑞 𝑋, 𝑊 = 𝑇𝑝 𝐑𝑛 , 𝐴 =

(𝑑𝜙)𝑝 and 𝑣𝑖 = (𝑑𝜙)𝑝 ( 𝜕𝑥𝜕 ) .
𝑖 𝑝
∞
Exercise 4.4.iv. Given a 𝐶 function 𝑓 ∶ 𝐑 → 𝐑, its graph
𝑋 ≔ 𝛤𝑓 = { (𝑥, 𝑓(𝑥)) ∈ 𝐑 × 𝐑 | 𝑥 ∈ 𝐑 }
is a submanifold of 𝐑2 and
𝜙 ∶ 𝐑 → 𝑋 , 𝑥 ↦ (𝑥, 𝑓(𝑥))
is a diffeomorphism. Orient 𝑋 by requiring that 𝜙 be orientation preserving and show that
if 𝜎 is the Riemannian volume form on 𝑋 then
1
2 2
⋆ 𝑑𝑓
𝜙 𝜎 = (1 + ( ) ) 𝑑𝑥 .
𝑑𝑥
Hint: Exercise 4.4.iii
Exercise 4.4.v. Given a 𝐶∞ function 𝑓 ∶ 𝐑𝑛 → 𝐑 the graph 𝛤𝑓 of 𝑓 is a submanifold of 𝐑𝑛+1
and
(4.4.33) 𝜙 ∶ 𝐑𝑛 → 𝑋 , 𝑥 ↦ (𝑥, 𝑓(𝑥))
is a diffeomorphism. Orient 𝑋 by requiring that 𝜙 is orientation preserving and show that
if 𝜎 is the Riemannian volume form on 𝑋 then
1
𝑛 2 2
𝜕𝑓
𝜙⋆ 𝜎 = (1 + ∑ ( ) ) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
𝑖=1 𝜕𝑥𝑖
Hints:
• Let 𝑣 = (𝑐1 , …, 𝑐𝑛 ) ∈ 𝐑𝑛 . Show that if 𝐶 ∶ 𝐑𝑛 → 𝐑 is the linear mapping defined
𝑛
by the matrix (𝑐𝑖 𝑐𝑗 ) then 𝐶𝑣 = (∑𝑖=1 𝑐𝑖2 )𝑣 and 𝐶𝑤 = 0 if 𝑤 ⋅ 𝑣 = 0 .
• Conclude that the eigenvalues of 𝐶 are 𝜆1 = ∑ 𝑐𝑖2 and 𝜆2 = ⋯ = 𝜆𝑛 = 0.
𝑛
• Show that the determinant of 𝐼 + 𝐶 is 1 + ∑𝑖=1 𝑐𝑖2 .
• Compute the determinant of the matrix (4.4.32) where 𝜙 is the mapping (4.4.33).
Exercise 4.4.vi. Let 𝑉 be an oriented 𝑁-dimensional vector space and ℓ𝑖 ∈ 𝑉⋆ , 𝑖 = 1, …, 𝑘,
𝑘 linearly independent vectors in 𝑉⋆ . Define
𝐿 ∶ 𝑉 → 𝐑𝑘
to be the map 𝑣 ↦ (ℓ1 (𝑣), …, ℓ𝑘 (𝑣)).
(1) Show that 𝐿 is surjective and that the kernel 𝑊 of 𝐿 is of dimension 𝑛 = 𝑁 − 𝑘.
(2) Show that one gets from this mapping a bijective linear mapping
(4.4.34) 𝑉/𝑊 → 𝐑𝑘
and hence from the standard orientation on 𝐑𝑘 an induced orientation on 𝑉/𝑊 and on
𝑊.
Hint: Exercise 1.2.viii and Theorem 1.9.9.
(3) Let 𝜔 be an element of 𝛬𝑁 (𝑉⋆ ). Show that there exists a 𝜇 ∈ 𝛬𝑛 (𝑉⋆ ) with the property
(4.4.35) ℓ1 ∧ ⋯ ∧ ℓ𝑘 ∧ 𝜇 = 𝜔 .
Hint: Choose an oriented basis, 𝑒1 , …, 𝑒𝑁 of 𝑉 such that 𝜔 = 𝑒1⋆ ∧⋯∧𝑒𝑁
⋆
and ℓ𝑖 = 𝑒𝑖⋆
⋆ ⋆
for 𝑖 = 1, …, 𝑘, and let 𝜇 = 𝑒𝑖+1 ∧ ⋯ ∧ 𝑒𝑁 .
 Draft: March 28, 2018

§4.4 Orientations 123

(4) Show that if 𝜈 is an element of 𝛬𝑛 (𝑉⋆ ) with the property

ℓ1 ∧ ⋯ ∧ ℓ𝑘 ∧ 𝜈 = 0
then there exist elements, 𝜈𝑖 , of 𝛬𝑛−1 (𝑉⋆ ) such that
𝜈 = ∑ ℓ𝑖 ∧ 𝜈𝑖 .
Hint: Same hint as in part (3).
(5) Show that if 𝜇1 and 𝜇2 are elements of 𝛬𝑛 (𝑉⋆ ) with the property (4.4.35) and 𝜄 ∶ 𝑊 → 𝑉
is the inclusion map then 𝜄⋆ 𝜇1 = 𝜄⋆ 𝜇2 .
Hint: Let 𝜈 = 𝜇1 − 𝜇2 . Conclude from part (4) that 𝜄⋆ 𝜈 = 0.
(6) Conclude that if 𝜇 is an element of 𝛬𝑛 (𝑉⋆ ) satisfying (4.4.35) the element, 𝜎 = 𝜄⋆ 𝜇, of
𝛬𝑛 (𝑊⋆ ) is intrinsically defined independent of the choice of 𝜇.
(7) Show that 𝜎 lies in 𝛬𝑛 (𝑊⋆ )+ .
Exercise 4.4.vii. Let 𝑈 be an open subset of 𝐑𝑁 and 𝑓 ∶ 𝑈 → 𝐑𝑘 a 𝐶∞ map. If zero is a
regular value of 𝑓, the set, 𝑋 = 𝑓−1 (0) is a manifold of dimension 𝑛 = 𝑁 − 𝑘. Show that this
manifold has a natural smooth orientation.
Some suggestions:
• Let 𝑓 = (𝑓1 , …, 𝑓𝑘 ) and let
𝑑𝑓1 ∧ ⋯ ∧ 𝑑𝑓𝑘 = ∑ 𝑓𝐼 𝑑𝑥𝐼
summed over multi-indices which are strictly increasing. Show that for every 𝑝 ∈
𝑋 𝑓𝐼 (𝑝) ≠ 0 for some multi-index, 𝐼 = (𝑖1 , …, 𝑖𝑘 ), 1 ≤ 𝑖1 < ⋯ < 𝑖𝑘 ≤ 𝑁.
• Let 𝐽 = (𝑗1 , …, 𝑗𝑛 ), 1 ≤ 𝑗1 < ⋯ < 𝑗𝑛 ≤ 𝑁 be the complementary multi-index to 𝐼,
i.e., 𝑗𝑟 ≠ 𝑖𝑠 for all 𝑟 and 𝑠. Show that
𝑑𝑓1 ∧ ⋯ ∧ 𝑑𝑓𝑘 ∧ 𝑑𝑥𝐽 = ±𝑓𝐼 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑁
and conclude that the 𝑛-form
1
𝜇=± 𝑑𝑥
𝑓𝐼 𝐽
is a 𝐶∞ 𝑛-form on a neighborhood of 𝑝 in 𝑈 and has the property:
𝑑𝑓1 ∧ ⋯ ∧ 𝑑𝑓𝑘 ∧ 𝜇 = 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑁 .
• Let 𝜄 ∶ 𝑋 → 𝑈 be the inclusion map. Show that the assignment
𝑝 ↦ (𝜄⋆ 𝜇)𝑝
defines an intrinsic nowhere vanishing 𝑛-form 𝜎 ∈ 𝛺𝑛 (𝑋) on 𝑋.
• Show that the orientation of 𝑋 defined by 𝜎 coincides with the orientation that
we described earlier in this section.
Exercise 4.4.viii. Let 𝑆𝑛 be the 𝑛-sphere and 𝜄 ∶ 𝑆𝑛 → 𝐑𝑛+1 the inclusion map. Show that
𝑛+1 ̂ 𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛+1 , the 𝑛-form
if 𝜔 ∈ 𝛺𝑛 (𝐑𝑛+1 ) is the 𝑛-form 𝜔 = ∑𝑖=1 (−1)𝑖−1 𝑥𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
⋆ 𝑛 𝑛
𝜄 𝜔 ∈ 𝛺 (𝑆 ) is the Riemannian volume form.
Exercise 4.4.ix. Let 𝑆𝑛+1 be the (𝑛 + 1)-sphere and let
𝑆𝑛+1
+ = { (𝑥1 , …, 𝑥𝑛+2 ) ∈ 𝑆
𝑛+1
| 𝑥1 < 0 }
be the lower hemisphere in 𝑆𝑛+1 .
(1) Prove that 𝑆𝑛+1
+ is a smooth domain.
(2) Show that the boundary of 𝑆𝑛+1 𝑛
+ is 𝑆 .
 Draft: March 28, 2018

124 Chapter 4: Manifolds & Forms on Manifolds

(3) Show that the boundary orientation of 𝑆𝑛 agrees with the orientation of 𝑆𝑛 in Exer-
cise 4.4.viii.

4.5. Integration of forms on manifolds

In this section we will show how to integrate differential forms over manifolds. In what
follows 𝑋 will be an oriented 𝑛-manifold and 𝑊 an open subset of 𝑋, and our goal will be
to make sense of the integral

(4.5.1) ∫ 𝜔
𝑊

where 𝜔 is a compactly supported 𝑛-form. We’ll begin by showing how to define this integral
when the support of 𝜔 is contained in a parametrizable open set 𝑈. Let 𝑈0 be an open subset
of 𝐑𝑛 and 𝜙0 ∶ 𝑈0 → 𝑈 a parametrization. As we noted in §4.4 we can assume without loss
of generality that this parametrization is oriented. Making this assumption, we’ll define

(4.5.2) ∫ 𝜔 = ∫ 𝜙⋆0 𝜔
𝑊 𝑊0

where 𝑊0 = 𝜙−1 ⋆
0 (𝑈 ∩ 𝑊). Notice that if 𝜙 𝜔 = 𝑓𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 , then, by assumption, 𝑓 is
∞
in 𝐶0 (𝑈0 ). Hence since
∫ 𝜙⋆0 𝜔 = ∫ 𝑓𝑑𝑥1 …𝑑𝑥𝑛
𝑊0 𝑊0
and since 𝑓 is a bounded continuous function and is compactly supported the Riemann
integral on the right is well-defined. (See Appendix b.) Moreover, if 𝜙1 ∶ 𝑈1 → 𝑈 is another
oriented parametrization of 𝑈 and 𝜓 ∶ 𝑈0 → 𝑈1 is the map 𝜓 = 𝜙−1 1 ∘ 𝜙0 then 𝜙0 = 𝜙1 ∘ 𝜓,
so by Proposition 4.3.4
𝜙⋆0 𝜔 = 𝜓⋆ 𝜙⋆1 𝜔 .
Moreover, by (4.3.18) 𝜓 is orientation preserving. Therefore since
𝑊1 = 𝜓(𝑊0 ) = 𝜙−1
1 (𝑈 ∩ 𝑊)

Theorem 3.5.2 tells us that

(4.5.3) ∫ 𝜙⋆1 𝜔 = ∫ 𝜙0⋆ 𝜔 .

𝑊1 𝑊0

Thus the definition (4.5.2) is a legitimate definition. It does not depend on the parametriza-
tion that we use to define the integral on the right. From the usual additivity properties
of the Riemann integral one gets analogous properties for the integral (4.5.2). Namely for
𝜔1 , 𝜔2 ∈ 𝛺𝑛𝑐 (𝑈),

(4.5.4) ∫ (𝜔1 + 𝜔2 ) = ∫ 𝜔1 + ∫ 𝜔2
𝑊 𝑊 𝑊

and for 𝜔 ∈ 𝛺𝑛𝑐 (𝑈) and 𝑐 ∈ 𝐑

∫ 𝑐𝜔 = 𝑐 ∫ 𝜔 .
𝑊 𝑊
We will next show how to define the integral (4.5.1) for any compactly supported 𝑛-
form. This we will do in more or less the same way that we defined improper Riemann inte-
grals in Appendix b: by using partitions of unity. We’ll begin by deriving from the partition
of unity theorem in Appendix b a manifold version of this theorem.
 Draft: March 28, 2018

§4.5 Integration of forms on manifolds 125

Theorem 4.5.5. Let

(4.5.6) U = {𝑈𝛼 }𝛼∈𝐼
be a covering of 𝑋 be open subsets. Then there exists a family of functions 𝜌𝑖 ∈ 𝐶0∞ (𝑋), for
𝑖 ≥ 1, with the properties
(1) 𝜌𝑖 ≥ 0.
(2) For every compact set 𝐶 ⊂ 𝑋 there exists a positive integer 𝑁 such that if 𝑖 > 𝑁 we have
supp(𝜌𝑖 ) ∩ 𝐶 = ∅.
∞
(3) ∑𝑖=1 𝜌𝑖 = 1.
(4) For every 𝑖 ≥ 1 there exists an index 𝛼 ∈ 𝐼 such that supp(𝜌𝑖 ) ⊂ 𝑈𝛼 .
Remark 4.5.7. Conditions (1)–(3) say that the 𝜌𝑖 ’s are a partition of unity and (4) says that
this partition of unity is subordinate to the covering (4.5.5).
Proof. For each 𝑝 ∈ 𝑋 and for some 𝑈𝛼 containing a 𝑝 choose an open set 𝑂𝑝 in 𝐑𝑁 with
𝑝 ∈ 𝑂𝑝 and with
(4.5.8) 𝑂𝑝 ∩ 𝑋 ⊂ 𝑈𝛼 .
Let 𝑂 ≔ ⋃𝑝∈𝑋 𝑂𝑝 and let 𝜌̃𝑖 ∈ 𝐶0∞ (𝑂),for 𝑖 ≥ 1, be a partition of unity subordinate to
the covering of 𝑂 by the 𝑂𝑝 ’s. By (4.5.8) the restriction 𝜌𝑖 of 𝜌̃𝑖 to 𝑋 has compact support
and it is clear that the 𝜌𝑖 ’s inherit from the 𝜌̃𝑖 ’s the properties (1)–(4). □
Now let the covering (4.5.6) be any covering of 𝑋 by parametrizable open sets and let
𝜌𝑖 ∈ 𝐶0∞ (𝑋), for 𝑖 ≥ 1 be a partition of unity subordinate to this covering. Given 𝜔 ∈ 𝛺𝑛𝑐 (𝑋)
we will define the integral of 𝜔 over 𝑊 by the sum
∞
(4.5.9) ∑ ∫ 𝜌𝑖 𝜔 .
𝑖=1 𝑊
Note that since each 𝜌𝑖 is supported in some 𝑈𝛼 the individual summands in this sum are
well-defined and since the support of 𝜔 is compact all but finitely many of these summands
are zero by part (2) of Theorem 4.5.5. Hence the sum itself is well-defined. Let’s show that
this sum does not depend on the choice of U and the 𝜌𝑖 ’s. Let U ′ be another covering of 𝑋
by parametrizable open sets and (𝜌𝑗′ )𝑗≥1 a partition of unity subordinate to U ′ . Then
∞ ∞ ∞ ∞ ∞
(4.5.10) ∑ ∫ 𝜌𝑗′ 𝜔 = ∑ ∫ ∑ 𝜌𝑗′ 𝜌𝑖 𝜔 = ∑ (∑ ∫ 𝜌𝑗′ 𝜌𝑖 𝜔)
𝑗=1 𝑊 𝑗=1 𝑊 𝑖=1 𝑗=1 𝑖=1 𝑊

by equation (4.5.4). Interchanging the orders of summation and resuming with respect to
the 𝑗’s this sum becomes
∞ ∞
∑ ∫ ∑ 𝜌𝑗′ 𝜌𝑖 𝜔
𝑖=1 𝑊 𝑗=1
or
∞
∑ ∫ 𝜌𝑖 𝜔 .
𝑖=1 𝑊
Hence
∞ ∞
∑ ∫ 𝜌𝑗′ 𝜔 = ∑ ∫ 𝜌𝑖 𝜔 ,
𝑖=1 𝑊 𝑖=1 𝑊
so the two sums are the same.
From equations (4.5.4) and (4.5.9) one easily deduces:
 Draft: March 28, 2018

126 Chapter 4: Manifolds & Forms on Manifolds

Proposition 4.5.11. For 𝜔1 , 𝜔2 ∈ 𝛺𝑛𝑐 (𝑋),

∫ 𝜔1 + 𝜔2 = ∫ 𝜔1 + ∫ 𝜔2
𝑊 𝑊 𝑊
and for 𝜔 ∈ 𝛺𝑛𝑐 (𝑋) and 𝑐 ∈ 𝐑
∫ 𝑐𝜔 = 𝑐 ∫ 𝜔 .
𝑊 𝑊

The definition of the integral (4.5.1) depends on the choice of an orientation of 𝑋, but
it’s easy to see how it depends on this choice. We pointed out in Section 4.4 that if 𝑋 is
connected, there is just one way to orient it smoothly other than by its given orientation,
namely by reversing the orientation of 𝑇𝑝 𝑋 at each point 𝑝 and it’s clear from the definitions
(4.5.2) and (4.5.9) that the effect of doing this is to change the sign of the integral, i.e., to
change ∫𝑋 𝜔 to − ∫𝑋 𝜔.
In the definition of the integral (4.5.1) we have allowed 𝑊 to be an arbitrary open subset
of 𝑋 but required 𝜔 to be compactly supported. This integral is also well-defined if we allow
𝜔 to be an arbitrary element of 𝛺𝑛 (𝑋) but require the closure of 𝑊 in 𝑋 to be compact. To
see this, note that under this assumption the sum (4.5.8) is still a finite sum, so the definition
of the integral still makes sense, and the double sum on the right side of (4.5.10) is still a
finite sum so it’s still true that the definition of the integral does not depend on the choice
of partitions of unity. In particular if the closure of 𝑊 in 𝑋 is compact we will define the
volume of 𝑊 to be the integral,
vol(𝑊) = ∫ 𝜎vol ,
𝑊
where 𝜎vol is the Riemannian volume form and if 𝑋 itself is compact we’ll define its volume
to be the integral
vol(𝑋) = ∫ 𝜎vol .
𝑋
We’ll next prove a manifold version of the change of variables formula (3.5.3).
Theorem 4.5.12 (change of variables formula). Let 𝑋′ and 𝑋 be oriented 𝑛-manifolds and
𝑓 ∶ 𝑋′ → 𝑋 an orientation preserving diffeomorphism. If 𝑊 is an open subset of 𝑋 and
𝑊′ ≔ 𝑓−1 (𝑊), then

(4.5.13) ∫ 𝑓⋆ 𝜔 = ∫ 𝜔
𝑊′ 𝑊
for all 𝜔 ∈ 𝛺𝑛𝑐 (𝑋).
Proof. By (4.5.9) the integrand of the integral above is a finite sum of 𝐶∞ forms, each of
which is supported on a parametrizable open subset, so we can assume that 𝜔 itself as this
property. Let 𝑉 be a parametrizable open set containing the support of 𝜔 and let 𝜙0 ∶ 𝑈 ⥲ 𝑉
be an oriented parameterization of 𝑉. Since 𝑓 is a diffeomorphism its inverse exists and is
a diffeomorphism of 𝑋 onto 𝑋1 . Let 𝑉′ ≔ 𝑓−1 (𝑉) and 𝜙′0 ≔ 𝑓−1 ∘ 𝜙0 . Then 𝜙′0 ∶ 𝑈 ⥲ 𝑉′ is
an oriented parameterization of 𝑉′ . Moreover, 𝑓 ∘ 𝜙′0 = 𝜙0 so if 𝑊0 = 𝜙−1
0 (𝑊) we have
𝑊0 = (𝜙′0 )−1 (𝑓−1 (𝑊)) = (𝜙′0 )−1 (𝑊′ )
and by the chain rule we have
𝜙⋆0 𝜔 = (𝑓 ∘ 𝜙′0 )⋆ 𝜔 = (𝜙′0 )⋆ 𝑓⋆ 𝜔
hence
∫ 𝜔 = ∫ 𝜙⋆0 𝜔 = ∫ (𝜙′0 )⋆ (𝑓⋆ 𝜔) = ∫ 𝑓⋆ 𝜔 . □
𝑊 𝑊0 𝑊0 𝑊′
 Draft: March 28, 2018

§4.5 Integration of forms on manifolds 127

As an exercise, show that if 𝑓 ∶ 𝑋′ → 𝑋 is orientation reversing

(4.5.14) ∫ 𝑓⋆ 𝜔 = − ∫ 𝜔 .
𝑊′ 𝑊
We’ll conclude this discussion of “integral calculus on manifolds” by proving a prelim-
inary version of Stokes’ theorem.
Theorem 4.5.15. If 𝜇 is in 𝛺𝑛−1
𝑐 (𝑋) then

∫ 𝑑𝜇 = 0 .
𝑋
Proof. Let (𝜌𝑖 )𝑖≥1 be a partition of unity with the property that each 𝜌𝑖 is supported in a
parametrizable open set 𝑈𝑖 = 𝑈. Replacing 𝜇 by 𝜌𝑖 𝜇 it suffices to prove the theorem for
𝜇 ∈ 𝛺𝑛−1
𝑐 (𝑈). Let 𝜙 ∶ 𝑈0 → 𝑈 be an oriented parametrization of 𝑈. Then

∫ 𝑑𝜇 = ∫ 𝜙⋆ 𝑑𝜇 = ∫ 𝑑(𝜙⋆ 𝜇) = 0
𝑈 𝑈0 𝑈0

by Theorem 3.3.1. □
Exercises for §4.5
Exercise 4.5.i. Let 𝑓 ∶ 𝐑 → 𝐑 be a 𝐶∞ function. Orient the graph 𝑋 ≔ 𝛤𝑓 of 𝑓 by requiring
𝑛

that the diffeomorphism

𝜙 ∶ 𝐑𝑛 → 𝑋 , 𝑥 ↦ (𝑥, 𝑓(𝑥))
be orientation preserving. Given a bounded open set 𝑈 in 𝐑𝑛 compute the Riemannian
volume of the image
𝑋𝑈 = 𝜙(𝑈)
of 𝑈 in 𝑋 as an integral over 𝑈.
Hint: Exercise 4.4.v.
Exercise 4.5.ii. Evaluate this integral for the open subset 𝑋𝑈 of the paraboloid defined by
𝑥3 = 𝑥12 + 𝑥22 , where 𝑈 is the disk 𝑥12 + 𝑥22 < 2.
Exercise 4.5.iii. In Exercise 4.5.i let 𝜄 ∶ 𝑋 ↪ 𝐑𝑛+1 be the inclusion map of 𝑋 onto 𝐑𝑛+1 .
(1) If 𝜔 ∈ 𝛺𝑛 (𝐑𝑛+1 ) is the 𝑛-form 𝑥𝑛+1 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 , what is the integral of 𝜄⋆ 𝜔 over the
set 𝑋𝑈 ? Express this integral as an integral over 𝑈.
2
(2) Same question for 𝜔 = 𝑥𝑛+1 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
(3) Same question for 𝜔 = 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
Exercise 4.5.iv. Let 𝑓 ∶ 𝐑𝑛 → (0, +∞) be a positive 𝐶∞ function, 𝑈 a bounded open subset
of 𝐑𝑛 , and 𝑊 the open set of 𝐑𝑛+1 defined by the inequalities
0 < 𝑥𝑛+1 < 𝑓(𝑥1 , …, 𝑥𝑛 )
and the condition (𝑥1 , …, 𝑥𝑛 ) ∈ 𝑈.
(1) Express the integral of the (𝑛 + 1)-form 𝜔 = 𝑥𝑛+1 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛+1 over 𝑊 as an integral
over 𝑈.
2
(2) Same question for 𝜔 = 𝑥𝑛+1 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛+1 .
(3) Same question for 𝜔 = 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
Exercise 4.5.v. Integrate the “Riemannian area” form
𝑥1 𝑑𝑥2 ∧ 𝑑𝑥3 + 𝑥2 𝑑𝑥3 ∧ 𝑑𝑥1 + 𝑥3 𝑑𝑥1 ∧ 𝑑𝑥2
2
over the unit 2-sphere 𝑆 . (See Exercise 4.4.viii.)
 Draft: March 28, 2018

128 Chapter 4: Manifolds & Forms on Manifolds

Hint: An easier problem: Using polar coordinates integrate 𝜔 = 𝑥3 𝑑𝑥1 ∧ 𝑑𝑥2 over the
hemisphere defined by 𝑥3 = √1 − 𝑥12 − 𝑥22 , where 𝑥12 + 𝑥22 < 1.
𝑛
Exercise 4.5.vi. Let 𝛼 be the one-form ∑𝑖=1 𝑦𝑖 𝑑𝑥𝑖 in equation (2.8.2) and let 𝛾(𝑡), 0 ≤ 𝑡 ≤ 1,
be a trajectory of the Hamiltonian vector field (2.8.2). What is the integral of 𝛼 over 𝛾(𝑡)?

4.6. Stokes’ theorem & the divergence theorem

Let 𝑋 be an oriented 𝑛-manifold and 𝐷 ⊂ 𝑋 a smooth domain. We showed in §4.4 that
𝜕𝐷 acquires from 𝐷 a natural orientation. Hence if 𝜄 ∶ 𝜕𝐷 → 𝑋 is the inclusion map and 𝜇
is in 𝛺𝑛−1
𝑐 (𝑋), the integral

∫ 𝜄⋆ 𝜇
𝜕𝐷
is well-defined. We will prove:
Theorem 4.6.1 (Stokes’ theorem). For 𝜇 ∈ 𝛺𝑘−1
𝑐 (𝑋) we have

(4.6.2) ∫ 𝜄⋆ 𝜇 = ∫ 𝑑𝜇 .
𝜕𝐷 𝐷

Proof. Let (𝜌𝑖 )𝑖≥1 be a partition of unity such that for each 𝑖, the support of 𝜌𝑖 is contained
in a parametrizable open set 𝑈𝑖 = 𝑈 of one of the following three types:
(a) 𝑈 ⊂ int(𝐷).
(b) 𝑈 ⊂ ext(𝐷).
(c) There exists an open subset 𝑈0 of 𝐑𝑛 and an oriented 𝐷-adapted parametrization
𝜙 ∶ 𝑈0 ⥲ 𝑈 .
∞
Replacing 𝜇 by the finite sum ∑𝑖=1 𝜌𝑖 𝜇 it suffices to prove (4.6.2) for each 𝜌𝑖 𝜇 separately. In
other words we can assume that the support of 𝜇 itself is contained in a parametrizable open
set 𝑈 of type (a), (b), or (c). But if 𝑈 is of type (a)

∫ 𝑑𝜇 = ∫ 𝑑𝜇 = ∫ 𝑑𝜇
𝐷 𝑈 𝑋

and 𝜄⋆ 𝜇 = 0. Hence the left hand side of equation (4.6.2) is zero and, by Theorem 4.5.15,
the right hand side is as well. If 𝑈 is of type (b) the situation is even simpler: 𝜄⋆ 𝜇 is zero
and the restriction of 𝜇 to 𝐷 is zero, so both sides of equation (4.6.2) are automatically zero.
Thus one is reduced to proving (4.6.2) when 𝑈 is an open subset of type (c). In this case the
restriction of the map (4.6.2) to 𝑈0 ∩ 𝜕𝐇𝑛 is an orientation preserving diffeomorphism
(4.6.3) 𝜓 ∶ 𝑈0 ∩ 𝜕𝐇𝑛 → 𝑈 ∩ 𝑍
and
(4.6.4) 𝜄𝑍 ∘ 𝜓 = 𝜙 ∘ 𝜄𝐑𝑛−1
where the maps 𝜄 = 𝜄𝑍 and
𝜄𝐑𝑛−1 ∶ 𝐑𝑛−1 ↪ 𝐑𝑛
are the inclusion maps of 𝑍 into 𝑋 and 𝜕𝐇𝑛 into 𝐑𝑛 . (Here we’re identifying 𝜕𝐇𝑛 with 𝐑𝑛−1 .)
Thus
∫ 𝑑𝜇 = ∫ 𝜙⋆ 𝑑𝜇 = ∫ 𝑑𝜙⋆ 𝜇
𝐷 𝐇𝑛 𝐇𝑛
 Draft: March 28, 2018

§4.6 Stokes’ theorem & the divergence theorem 129

and by (4.6.4)

∫ 𝜄𝑍⋆ 𝜇 = ∫ 𝜓⋆ 𝜄𝑍⋆ 𝜇
𝑍 𝐑𝑛−1

=∫ 𝜄⋆𝐑𝑛−1 𝜙⋆ 𝜇
𝐑𝑛−1

=∫ 𝜄⋆𝐑𝑛−1 𝜙⋆ 𝜇 .
𝜕𝐇𝑛
Thus it suffices to prove Stokes’ theorem with 𝜇 replaced by 𝜙⋆ 𝜇, or, in other words, to prove
Stokes’ theorem for 𝐇𝑛 ; and this we will now do.
Theorem 4.6.5 (Stokes’ theorem for 𝐇𝑛 ). Let
𝑛
̂ 𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
𝜇 = ∑(−1)𝑖−1 𝑓𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑖=1

Then
𝑛
𝜕𝑓𝑖
𝑑𝜇 = ∑ 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛
𝑖=1 𝜕𝑥𝑖
and
𝑛
𝜕𝑓𝑖
∫ 𝑑𝜇 = ∑ ∫ 𝑑𝑥 ⋯ 𝑑𝑥𝑛 .
𝐇𝑛 𝑖=1 𝐇𝑛 𝜕𝑥𝑖 1
We will compute each of these summands as an iterated integral doing the integration
with respect to 𝑑𝑥𝑖 first. For 𝑖 > 1 the 𝑑𝑥𝑖 integration ranges over the interval −∞ < 𝑥𝑖 < ∞
and hence since 𝑓𝑖 is compactly supported
∞ 𝑥𝑖 =+∞
𝜕𝑓𝑖
∫ 𝑑𝑥𝑖 = 𝑓𝑖 (𝑥1 , …, 𝑥𝑖 , …, 𝑥𝑛 )| =0.
−∞ 𝜕𝑥𝑖 𝑥𝑖 =−∞

On the other hand the 𝑑𝑥1 integration ranges over the interval −∞ < 𝑥1 < 0 and
0
𝜕𝑓1
∫ 𝑑𝑥 = 𝑓(0, 𝑥2 , …, 𝑥𝑛 ) .
−∞ 𝜕𝑥1 1
Thus integrating with respect to the remaining variables we get

(4.6.6) ∫ 𝑑𝜇 = ∫ 𝑓(0, 𝑥2 , …, 𝑥𝑛 ) 𝑑𝑥2 ∧ ⋯ ∧ 𝑑𝑥𝑛 .

𝐇𝑛 𝐑𝑛−1
On the other hand, since 𝜄⋆𝐑𝑛−1 𝑥1 = 0 and 𝜄⋆𝐑𝑛−1 𝑥𝑖 = 𝑥𝑖 for 𝑖 > 1,
𝜄⋆𝐑𝑛−1 𝜇 = 𝑓1 (0, 𝑥2 , …, 𝑥𝑛 ) 𝑑𝑥2 ∧ ⋯ ∧ 𝑑𝑥𝑛
so
(4.6.7) ∫ 𝜄⋆𝐑𝑛−1 𝜇 = ∫ 𝑓(0, 𝑥2 , …, 𝑥𝑛 ) 𝑑𝑥2 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
𝐑𝑛−1 𝐑𝑛−1
Hence the two sides, (4.6.6) and (4.6.7), of Stokes’ theorem are equal. □
One important variant of Stokes’ theorem is the divergence theorem: Let 𝜔 be in 𝛺𝑛𝑐 (𝑋)
and let 𝒗 be a vector field on 𝑋. Then
𝐿𝒗 𝜔 = 𝜄𝒗 𝑑𝜔 + 𝑑𝜄𝒗 𝜔 = 𝑑𝜄𝒗 𝜔 ,
hence, denoting by 𝜄𝑍 the inclusion map of 𝑍 into 𝑋 we get from Stokes’ theorem, with
𝜇 = 𝜄𝒗 𝜔:
 Draft: March 28, 2018

130 Chapter 4: Manifolds & Forms on Manifolds

Theorem 4.6.8 (divergence theorem for manifolds).

∫ 𝐿𝒗 𝜔 = ∫ 𝜄𝑍⋆ (𝜄𝒗 𝜔) .
𝐷 𝑍

If 𝐷 is an open domain in 𝐑𝑛 this reduces to the usual divergence theorem of multi-

𝑛
variable calculus. Namely if 𝜔 = 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 and 𝒗 = ∑𝑖=1 𝑣𝑖 𝜕𝑥𝜕 then by (2.5.14)
𝑖

𝐿𝒗 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 = div(𝒗) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛

where
𝑛
𝜕𝑣𝑖
div(𝒗) = ∑ .
𝑖=1 𝜕𝑥𝑖
Thus if 𝑍 = 𝜕𝐷 and 𝜄𝑍 is the inclusion 𝑍 ↪ 𝐑𝑛 ,

(4.6.9) ∫ div(𝒗)𝑑𝑥 = ∫ 𝜄𝑍⋆ (𝜄𝒗 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ) .

𝐷 𝑍
The right hand side of this identity can be interpreted as the “flux” of the vector field 𝒗
through the boundary of 𝐷. To see this let 𝑓 ∶ 𝐑𝑛 → 𝐑 be a 𝐶∞ defining function for 𝐷, i.e.,
a function with the properties
𝑝 ∈ 𝐷 ⟺ 𝑓(𝑝) < 0
and 𝑑𝑓𝑝 ≠ 0 if 𝑝 ∈ 𝜕𝐷. This second condition says that zero is a regular value of 𝑓 and
hence that 𝑍 ≔ 𝜕𝐷 is defined by the non-degenerate equation:
(4.6.10) 𝑝 ∈ 𝑍 ⟺ 𝑓(𝑝) = 0 .
Let 𝒘 be the vector field
𝑛 2 −1 𝑛
𝜕𝑓 𝜕𝑓𝑖 𝜕
𝒘 ≔ (∑ ( ) ) ∑ .
𝑖=1 𝜕𝑥𝑖 𝑖=1 𝜕𝑥𝑖 𝜕𝑥𝑖

In view of (4.6.10), this vector field is well-defined on a neighborhood 𝑈 of 𝑍 and satisfies

𝜄𝒘 𝑑𝑓 = 1 .
Now note that since 𝑑𝑓 ∧ 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 = 0
0 = 𝜄𝒘 (𝑑𝑓 ∧ 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 )
= (𝜄𝒘 𝑑𝑓)𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 − 𝑑𝑓 ∧ 𝜄𝒘 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛
= 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 − 𝑑𝑓 ∧ 𝜄𝒘 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ,
hence letting 𝜈 be the (𝑛 − 1)-form 𝜄𝒘 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 we get the identity
(4.6.11) 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 = 𝑑𝑓 ∧ 𝜈
and by applying the operation 𝜄𝒗 to both sides of (4.6.11) the identity
(4.6.12) 𝜄𝒗 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 = (𝐿𝒗 𝑓)𝜈 − 𝑑𝑓 ∧ 𝜄𝒗 𝜈 .
Let 𝜈𝑍 = 𝜄𝑍⋆ 𝜈 be the restriction of 𝜈 to 𝑍. Since 𝜄𝑍⋆ = 0, 𝜄𝑍⋆ 𝑑𝑓 = 0 and hence by (4.6.12)
𝜄𝑍⋆ (𝜄𝒗 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ) = 𝜄𝑍⋆ (𝐿𝒗 𝑓)𝜈𝑍 ,
and the formula (4.6.9) now takes the form

(4.6.13) ∫ div(𝒗)𝑑𝑥 = ∫ 𝐿𝒗 𝑓𝜈𝑍

𝐷 𝑍
 Draft: March 28, 2018

§4.6 Stokes’ theorem & the divergence theorem 131

where the term on the right is by definition the flux of 𝒗 through 𝑍. In calculus books this
is written in a slightly different form. Letting
1
𝑛 2 2
𝜕𝑓
𝜎𝑍 = (∑ ( ) ) 𝜈𝑍
𝑖=1 𝜕𝑥𝑖

and letting
𝑛 2 − 21
𝜕𝑓 𝜕𝑓 𝜕𝑓
𝑛⃗ ≔ (∑ ( ) ) ( , …, )
𝑖=1 𝜕𝑥𝑖 𝜕𝑥1 𝜕𝑥𝑛
and 𝑣 ⃗ ≔ (𝑣1 , …, 𝑣𝑛 ) we have
𝐿𝒗 𝜈𝑍 = (𝑛⃗ ⋅ 𝑣)𝜎
⃗ 𝑍
and hence

(4.6.14) ∫ div(𝒗)𝑑𝑥 = ∫ (𝑛⃗ ⋅ 𝑣)𝜎

⃗ 𝑍.
𝐷 𝑍

In three dimensions 𝜎𝑍 is just the standard “infinitesimal element of area” on the surface 𝑍
and 𝑛𝑝 the unit outward normal to 𝑍 at 𝑝, so this version of the divergence theorem is the
version one finds in most calculus books.
As an application of Stokes’ theorem, we’ll give a very short alternative proof of the
Brouwer fixed point theorem. As we explained in § 3.6 the proof of this theorem basically
comes down to proving the following theorem.
Theorem 4.6.15. Let 𝐵𝑛 be the closed unit ball in 𝐑𝑛 and 𝑆𝑛−1 its boundary. Then the identity
map id𝑆𝑛−1 on 𝑆𝑛−1 cannot be extended to a 𝐶∞ map 𝑓 ∶ 𝐵𝑛 → 𝑆𝑛−1 .
Proof. Suppose that 𝑓 is such a map. Then for every (𝑛 − 1)-form 𝜇 ∈ 𝛺𝑛−1 (𝑆𝑛−1 ),

(4.6.16) ∫ 𝑑(𝑓⋆ 𝜇) = ∫ (𝜄𝑆𝑛−1 )⋆ 𝑓⋆ 𝜇 .

𝐵𝑛 𝑆𝑛−1

But 𝑑(𝑓⋆ 𝜇) = 𝑓⋆ (𝑑𝜇) = 0 since 𝜇 is an (𝑛−1)-form and 𝑆𝑛−1 is an (𝑛−1)-manifold, and since
𝑓 is the identity map on 𝑆𝑛−1 , (𝜄𝑆𝑛−1 )⋆ 𝑓⋆ 𝜇 = (𝑓 ∘ 𝜄𝑆𝑛−1 )⋆ 𝜇 = 𝜇. Thus for every 𝜇 ∈ 𝛺𝑛−1 (𝑆𝑛−1 ),
equation (4.6.16) says that the integral of 𝜇 over 𝑆𝑛−1 is zero. Since there are lots of (𝑛 − 1)-
forms for which this is not true, this shows that a map 𝑓 with the property above cannot
exist. □

Exercises for §4.6

Exercise 4.6.i. Let 𝐵 be the open unit ball in 𝐑𝑛 and 𝑆𝑛−1 the unit (𝑛 − 1)-sphere. Show
𝑛

that vol(𝑆𝑛−1 ) = 𝑛 vol(𝐵𝑛 ).

Hint: Apply Stokes’ theorem to the (𝑛 − 1)-form
𝑛
̂ 𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛
𝜇 ≔ ∑(−1)𝑖−1 𝑥𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑖=1

and note by Exercise 4.4.ix that 𝜇 is the Riemannian volume form of 𝑆𝑛−1 .
Exercise 4.6.ii. Let 𝐷 ⊂ 𝐑𝑛 be a smooth domain. Show that there exists a neighborhood 𝑈
of 𝜕𝐷 in 𝐑𝑛 and a 𝐶∞ defining function 𝑔 ∶ 𝑈 → 𝐑 for 𝐷 with the properties:
(1) 𝑝 ∈ 𝑈 ∩ 𝐷 if and only if 𝑔(𝑝) < 0,
(2) and 𝑑𝑔𝑝 ≠ 0 if 𝑝 ∈ 𝑍
 Draft: March 28, 2018

132 Chapter 4: Manifolds & Forms on Manifolds

Hint: Deduce from Theorem 4.4.9 that a local version of this result is true. Show that
you can cover 𝑍 by a family U = {𝑈𝛼 }𝛼∈𝐼 of open subsets of 𝐑𝑛 such that for each there
exists a function 𝑔𝛼 ∶ 𝑈𝛼 → 𝐑 with the properties (1) and (2). Now let (𝜌𝑖 )𝑖≥1 be a partition
∞
of unity subordinate to U and let 𝑔 = ∑𝑖=1 𝜌𝑖 𝑔𝛼𝑖 where supp(𝜌𝑖 ) ⊂ 𝑈𝛼𝑖 .
Exercise 4.6.iii. In Exercise 4.6.ii suppose 𝑍 is compact. Show that there exists a global
defining function 𝑓 ∶ 𝐑𝑛 → 𝐑 for 𝐷 with properties (1) and (2).
Hint: Let 𝜌 ∈ 𝐶0∞ (𝑈) be a function such that 0 ≤ 𝜌(𝑥) ≤ 1 for all 𝑥 ∈ 𝑈 and 𝜌 is
identically 1 on a neighborhood of 𝑍, and replace 𝑔 by the function
𝑛
{𝜌(𝑥)𝑔(𝑥) + (1 − 𝜌(𝑥)), 𝑥 ∈ 𝐑 ∖ 𝐷
{
𝑓(𝑥) = {𝑔(𝑥), 𝑥 ∈ 𝜕𝐷
{
{𝜌(𝑥) − (1 − 𝜌(𝑥))𝑔(𝑥), 𝑥 ∈ int(𝐷) .
Exercise 4.6.iv. Show that the form 𝐿𝒗 𝑓𝜈𝑍 in equation (4.6.13) does not depend on what
choice we make of a defining function 𝑓 for 𝐷.
Hints:
• Show that if 𝑔 is another defining function then, at 𝑝 ∈ 𝑍, 𝑑𝑓𝑝 = 𝜆𝑑𝑔𝑝 , where 𝜆
is a positive constant.
• Show that if one replaces 𝑑𝑓𝑝 by (𝑑𝑔)𝑝 the first term in the product (𝐿𝒗 𝑓)(𝑝)(𝜈𝑍 )𝑝
changes by a factor 𝜆 and the second term by a factor 1/𝜆.
Exercise 4.6.v. Show that the form 𝜈𝑍 is intrinsically defined in the sense that if 𝜈 is any
(𝑛 − 1)-form satisfying equation (4.6.11), then 𝜈𝑍 = 𝜄𝑍⋆ 𝜈.
Hint: Exercise 4.5.vi.
Exercise 4.6.vi. Show that the form 𝜎𝑍 in equation (4.6.14) is the Riemannian volume form
on 𝑍.
Exercise 4.6.vii. Show that the (𝑛 − 1)-form
𝑛
̂ 𝑟 ∧ ⋯ 𝑑𝑥𝑛
𝜇 = (𝑥12 + ⋯ + 𝑥𝑛2 )−𝑛 ∑ (−1)𝑟−1 𝑥𝑟 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑟=1

is closed and prove directly that Stokes’ theorem holds for the annulus 𝑎 < 𝑥12 + ⋯ + 𝑥𝑛2 < 𝑏
by showing that the integral of 𝜇 over the sphere, 𝑥12 + ⋯ + 𝑥𝑛2 = 𝑎, is equal to the integral
over the sphere, 𝑥12 + ⋯ + 𝑥𝑛2 = 𝑏.
Exercise 4.6.viii. Let 𝑓 ∶ 𝐑𝑛−1 → 𝐑 be an everywhere positive 𝐶∞ function and let 𝑈 be a
bounded open subset of 𝐑𝑛−1 . Verify directly that Stokes’ theorem is true if 𝐷 is the domain
defined by
0 < 𝑥𝑛 < 𝑓(𝑥1 , …, 𝑥𝑛−1 ) , (𝑥1 , …, 𝑥𝑛−1 ) ∈ 𝑈
and 𝜇 an (𝑛 − 1)-form of the form
𝜙(𝑥1 , …, 𝑥𝑛 )𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛−1
where 𝜙 is in 𝐶0∞ (𝐑𝑛 ).
Exercise 4.6.ix. Let 𝑋 be an oriented 𝑛-manifold and 𝒗 a vector field on 𝑋 which is complete.
Verify that for 𝜔 ∈ 𝛺𝑛𝑐 (𝑋)
∫ 𝐿𝒗 𝜔 = 0
𝑋
in the following ways:
(1) Directly by using the divergence theorem.
 Draft: March 28, 2018

§4.7 Degree theory on manifolds 133

(2) Indirectly by showing that

∫ 𝑓𝑡⋆ 𝜔 = ∫ 𝜔
𝑋 𝑋
where 𝑓𝑡 ∶ 𝑋 → 𝑋, −∞ < 𝑡 < ∞, is the one-parameter group of diffeomorphisms of 𝑋
generated by 𝒗.
Exercise 4.6.x. Let 𝑋 be an oriented 𝑛-manifold and 𝐷 ⊂ 𝑋 a smooth domain whose closure
is compact. Show that if 𝑍 ≔ 𝜕𝐷 is the boundary of 𝐷 and 𝑔 ∶ 𝑍 → 𝑍 a diffeomorphism 𝑔
cannot be extended to a smooth map 𝑓 ∶ 𝐷 → 𝑍.

4.7. Degree theory on manifolds

In this section we’ll show how to generalize to manifolds the results about the “degree”
of a proper mapping that we discussed in Chapter 3. We’ll begin by proving the manifold
analogue of Theorem 3.3.1.
Theorem 4.7.1. Let 𝑋 be an oriented connected 𝑛-manifold and 𝜔 ∈ 𝛺𝑛𝑐 (𝑋) a compactly
supported 𝑛-form. Then the following are equivalent
(1) ∫𝑋 𝜔 = 0.
(2) 𝜔 = 𝑑𝜇 for some 𝜇 ∈ 𝛺𝑛−1
𝑐 (𝑋).

Proof. We have already verified the assertion (2) ⇒(1) (see Theorem 4.5.15), so what is left
to prove is the converse assertion. The proof of this is more or less identical with the proof
of the “(1) ⇒(2)” part of Theorem 3.2.2:
Step 1. Let 𝑈 be a connected parametrizable open subset of 𝑋. If 𝜔 ∈ 𝛺𝑛𝑐 (𝑈) has property (1),
then 𝜔 = 𝑑𝜇 for some 𝜇 ∈ 𝛺𝑛−1
𝑐 (𝑈).

Proof of Step 1. Let 𝜙 ∶ 𝑈0 → 𝑈 be an oriented parametrization of 𝑈. Then

∫ 𝜙⋆ 𝜔 = ∫ 𝜔 = 0
𝑈0 𝑋

and since 𝑈0 is a connected open subset of 𝐑𝑛 , 𝜙⋆ 𝜔 = 𝑑𝜈 for some 𝜈 ∈ 𝛺𝑛−1

𝑐 (𝑈0 ) by Theo-
rem 3.3.1. Let 𝜇 = (𝜙−1 )⋆ 𝜈. Then 𝑑𝜇 = (𝜙−1 )⋆ 𝑑𝜈 = 𝜔. □

Step 2. Fix a base point 𝑝0 ∈ 𝑋 and let 𝑝 be any point of 𝑋. Then there exists a collection
of connected parametrizable open sets 𝑊1 , …, 𝑊𝑁 with 𝑝0 ∈ 𝑊1 and 𝑝 ∈ 𝑊𝑁 such that for
1 ≤ 𝑖 ≤ 𝑁 − 1, the intersection 𝑊𝑖 ∩ 𝑊𝑖+1 is non-empty.
Proof of Step 2. The set of points, 𝑝 ∈ 𝑋, for which this assertion is true is open and the set
for which it is not true is open. Moreover, this assertion is true for 𝑝 = 𝑝0 . □

Step 3. We deduce Theorem 4.7.1 from a slightly stronger result. Introduce an equivalence
relation on 𝛺𝑛𝑐 (𝑋) by declaring that two 𝑛-forms 𝜔1 , 𝜔2 ∈ 𝛺𝑛𝑐 (𝑋) are equivalent if 𝜔1 − 𝜔2 ∈
𝑑𝛺𝑥𝑛−1 (𝑋). Denote this equivalence relation by a 𝜔1 ∼ 𝜔2 .
We will prove:
Theorem 4.7.2. For 𝜔1 and 𝜔2 ∈ 𝛺𝑛𝑐 (𝑋) the following are equivalent:
(1) ∫𝑋 𝜔1 = ∫𝑋 𝜔2
(2) 𝜔1 ∼ 𝜔2 .
 Draft: March 28, 2018

134 Chapter 4: Manifolds & Forms on Manifolds

Applying this result to a form 𝜔 ∈ 𝛺𝑛𝑐 (𝑋) whose integral is zero, we conclude that
𝜔 ∼ 0, which means that 𝜔 = 𝑑𝜇 for some 𝜇 ∈ 𝛺𝑛−1 𝑐 (𝑋). Hence Theorem 4.7.2 implies
Theorem 4.7.1. Conversely, if ∫𝑋 𝜔1 = ∫𝑋 𝜔2 , then ∫𝑋 (𝜔1 − 𝜔2 ) = 0, so 𝜔1 − 𝜔2 = 𝑑𝜇 for
some 𝜇 ∈ 𝛺𝑛𝑐 (𝑋). Hence Theorem 4.7.2 implies Theorem 4.7.1.
Step 4. By a partition of unity argument it suffices to prove Theorem 4.7.2 for 𝜔1 ∈ 𝛺𝑛𝑐 (𝑈1 )
and 𝜔2 ∈ 𝛺𝑛𝑐 (𝑈2 ) where 𝑈1 and 𝑈2 are connected parametrizable open sets. Moreover, if the
integrals of 𝜔1 and 𝜔2 are zero then 𝜔𝑖 = 𝑑𝜇𝑖 for some 𝜇𝑖 ∈ 𝛺𝑛𝑐 (𝑈𝑖 ) by Step 1, so in this case,
the theorem is true. Suppose on the other hand that

∫ 𝜔1 = ∫ 𝜔2 = 𝑐 ≠ 0 .
𝑋 𝑋
Then dividing by 𝑐, we can assume that the integrals of 𝜔1 and 𝜔2 are both equal to 1.
Step 5. Let 𝑊1 , …, 𝑊𝑁 be, as in Step 2, a sequence of connected parametrizable open sets with
the property that the intersections, 𝑊1 ∩ 𝑈1 , 𝑊𝑁 ∩ 𝑈2 and 𝑊𝑖 ∩ 𝑊𝑖+1 , for 𝑖 = 1, …, 𝑁 − 1, are all
non-empty. Select 𝑛-forms, 𝛼0 ∈ 𝛺𝑛𝑐 (𝑈1 ∩ 𝑊1 ), 𝛼𝑁 ∈ 𝛺𝑛𝑐 (𝑊𝑁 ∩ 𝑈2 ) and 𝛼𝑖 ∈ 𝛺𝑛𝑐 (𝑊𝑖 ∩ 𝑊𝑖+1 ),
for 𝑖 = 1, …, 𝑁 − 1, such that the integral of each 𝛼𝑖 over 𝑋 is equal to 1 By Step 1 we see that
Theorem 4.7.1 is true for 𝑈1 , 𝑈2 and the 𝑊1 , …, 𝑊𝑁 , hence Theorem 4.7.2 is true for 𝑈1 , 𝑈2
and the 𝑊1 , …, 𝑊𝑁 , so
𝜔1 ∼ 𝛼 0 ∼ 𝛼 1 ∼ ⋯ ∼ 𝛼 𝑁 ∼ 𝜔 2
and thus 𝜔1 ∼ 𝜔2 . □
Just as in (3.4.2) we get as a corollary of the theorem above the following “definition–
theorem” of the degree of a differentiable mapping:
Theorem 4.7.3. Let 𝑋 and 𝑌 be compact oriented 𝑛-manifolds and let 𝑌 be connected. Given a
proper 𝐶∞ mapping, 𝑓 ∶ 𝑋 → 𝑌, there exists a topological invariant deg(𝑓) with the defining
property:

(4.7.4) ∫ 𝑓⋆ 𝜔 = deg(𝑓) ∫ 𝜔 .
𝑋 𝑌

Proof. As in the proof of Theorem 3.4.6 pick an 𝑛-form 𝜔0 ∈ 𝛺𝑛𝑐 (𝑌) whose integral over 𝑌
is one and define the degree of 𝑓 to be the integral over 𝑋 of 𝑓⋆ 𝜔0 , i.e., set

(4.7.5) deg(𝑓) ≔ ∫ 𝑓⋆ 𝜔0 .
𝑋
Now let 𝜔 be any 𝑛-form in 𝛺𝑛𝑐 (𝑌) and let

(4.7.6) 𝑐≔∫ 𝜔.
𝑌

Then the integral of 𝜔 − 𝑐𝜔0 over 𝑌 is zero so there exists an (𝑛 − 1)-form 𝜇 ∈ 𝛺𝑛−1
𝑐 (𝑌) for
which 𝜔 − 𝑐𝜔0 = 𝑑𝜇. Hence 𝑓⋆ 𝜔 = 𝑐𝑓⋆ 𝜔0 + 𝑑𝑓⋆ 𝜇, so

∫ 𝑓⋆ 𝜔 = 𝑐 ∫ 𝑓⋆ 𝜔0 = deg(𝑓) ∫ 𝜔
𝑋 𝑋 𝑌
by equations (4.7.5) and (4.7.6). □

It’s clear from the formula (4.7.4) that the degree of 𝑓 is independent of the choice of
𝜔0 . (Just apply this formula to any 𝜔 ∈ 𝛺𝑛𝑐 (𝑌) having integral over 𝑌 equal to one.) It’s also
clear from (4.7.4) that “degree” behaves well with respect to composition of mappings:
 Draft: March 28, 2018

§4.7 Degree theory on manifolds 135

Theorem 4.7.7. Let 𝑍 be an oriented, connected 𝑛-manifold and 𝑔 ∶ 𝑌 → 𝑍 a proper 𝐶∞

map. Then
(4.7.8) deg(𝑔 ∘ 𝑓) = deg(𝑓) deg(𝑔) .
Proof. Let 𝜔 ∈ 𝛺𝑛𝑐 (𝑍) be a compactly supported form with the property that ∫𝑍 𝜔 = 1. Then

deg(𝑔 ∘ 𝑓) = ∫ (𝑔 ∘ 𝑓)⋆ 𝜔 = ∫ 𝑓⋆ ∘ 𝑔⋆ 𝜔 = deg(𝑓) ∫ 𝑔⋆ 𝜔

𝑋 𝑋 𝑌
= deg(𝑓) deg(𝑔) . □
We will next show how to compute the degree of 𝑓 by generalizing to manifolds the
formula for deg(𝑓) that we derived in §3.6.
Definition 4.7.9. A point 𝑝 ∈ 𝑋 is a critical point of 𝑓 if the map
(4.7.10) 𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑓(𝑝) 𝑌
is not bijective.
We’ll denote by 𝐶𝑓 the set of all critical points of 𝑓, and we’ll call a point 𝑞 ∈ 𝑌 a critical
value of 𝑓 if it is in the image 𝑓(𝐶𝑓 ) of 𝐶𝑓 under 𝑓 and a regular value if it’s not. (Thus the
set of regular values is the set 𝑌 ∖ 𝑓(𝐶𝑓 ).) If 𝑞 is a regular value, then as we observed in Sec-
tion 3.6, the map (4.7.10) is bijective for every 𝑝 ∈ 𝑓−1 (𝑞) and hence by Theorem 4.2.13, 𝑓
maps a neighborhood 𝑈𝑝 of 𝑝 diffeomorphically onto a neighborhood 𝑉𝑝 of 𝑞. In particular,
𝑈𝑝 ∩ 𝑓−1 (𝑞) = 𝑝. Since 𝑓 is proper the set 𝑓−1 (𝑞) is compact, and since the sets 𝑈𝑝 are a
covering of 𝑓−1 (𝑞), this covering must be a finite covering. In particular, the set 𝑓−1 (𝑞) itself
has to be a finite set. As in § 2.7 we can shrink the 𝑈𝑝 ’s so as to insure that they have the
following properties:
(i) Each 𝑈𝑝 is a parametrizable open set.
(ii) 𝑈𝑝 ∩ 𝑈𝑝′ is empty for 𝑝 ≠ 𝑝′ .
(iii) 𝑓(𝑈𝑝 ) = 𝑓(𝑈𝑝′ ) = 𝑉 for all 𝑝 and 𝑝′ .
(iv) 𝑉 is a parametrizable open set.
(v) 𝑓−1 (𝑉) = ⋃𝑝∈𝑓−1 (𝑞) 𝑈𝑝 .
To exploit these properties let 𝜔 be an 𝑛-form in 𝛺𝑛𝑐 (𝑉) with integral equal to 1. Then
by (v):
deg(𝑓) = ∫ 𝑓⋆ 𝜔 = ∑ ∫ 𝑓⋆ 𝜔 .
𝑋 𝑝 𝑈𝑝

But 𝑓 ∶ 𝑈𝑝 → 𝑉 is a diffeomorphism, hence by (4.5.13) and (4.5.14)

∫ 𝑓⋆ 𝜔 = ∫ 𝜔
𝑈𝑝 𝑉

if 𝑓 ∶ 𝑈𝑝 → 𝑉 is orientation preserving and

∫ 𝑓⋆ 𝜔 = − ∫ 𝜔
𝑈𝑝 𝑉

if 𝑓 ∶ 𝑈𝑝 → 𝑉 is orientation reversing. Thus we have proved:

Theorem 4.7.11. The degree of 𝑓 is equal to the sum
(4.7.12) deg(𝑓) = ∑ 𝜎𝑝
𝑝∈𝑓−1 (𝑞)
 Draft: March 28, 2018

136 Chapter 4: Manifolds & Forms on Manifolds

where 𝜎𝑝 = +1 if the map (4.7.10) is orientation preserving and 𝜎𝑝 = −1 if the map (4.7.10)
is orientation reversing.
We will next show that Sard’s Theorem is true for maps between manifolds and hence
that there exist lots of regular values. We first observe that if 𝑈 is a parametrizable open
subset of 𝑋 and 𝑉 a parametrizable open neighborhood of 𝑓(𝑈) in 𝑌, then Sard’s Theorem
is true for the map 𝑓 ∶ 𝑈 → 𝑉 since, up to diffeomorphism, 𝑈 and 𝑉 are just open subsets
of 𝐑𝑛 . Now let 𝑞 be any point in 𝑌, let 𝐵 be a compact neighborhood of 𝑞, and let 𝑉 be a
parametrizable open set containing 𝐵. Then if 𝐴 = 𝑓−1 (𝐵) it follows from Theorem 3.4.7
that 𝐴 can be covered by a finite collection of parametrizable open sets, 𝑈1 , …, 𝑈𝑁 such that
𝑓(𝑈𝑖 ) ⊂ 𝑉. Hence since Sard’s Theorem is true for each of the maps 𝑓 ∶ 𝑈𝑖 → 𝑉 and 𝑓−1 (𝐵)
is contained in the union of the 𝑈𝑖 ’s we conclude that the set of regular values of 𝑓 intersects
the interior of 𝐵 in an open dense set. Thus, since 𝑞 is an arbitrary point of 𝑌, we have proved:
Theorem 4.7.13. If 𝑋 and 𝑌 are 𝑛-manifolds and 𝑓 ∶ 𝑋 → 𝑌 is a proper 𝐶∞ map the set of
regular values of 𝑓 is an open dense subset of 𝑌.
Since there exist lots of regular values the formula (4.7.12) gives us an effective way
of computing the degree of 𝑓. We’ll next justify our assertion that deg(𝑓) is a topological
invariant of 𝑓. To do so, let’s generalize Definition 2.6.14 to manifolds.
Definition 4.7.14. Let 𝑋 and 𝑌 be manifolds and 𝑓0 , 𝑓1 ∶ 𝑋 → 𝑌 be 𝐶∞ maps. A 𝐶∞ map
(4.7.15) 𝐹 ∶ 𝑋 × [0, 1] → 𝑌
is a homotopy between 𝑓0 and 𝑓1 if for all 𝑥 ∈ 𝑋 we have 𝐹(𝑥, 0) = 𝑓0 (𝑥) and 𝐹(𝑥, 1) = 𝑓1 (𝑥).
Moreover, if 𝑓0 and 𝑓1 are proper maps, the homotopy, 𝐹, is a proper homotopy if 𝐹 is proper
as a 𝐶∞ map, i.e., for every compact set 𝐶 of 𝑌, 𝐹−1 (𝐶) is compact.
Let’s now prove the manifold analogue of Theorem 3.6.10.
Theorem 4.7.16. Let 𝑋 and 𝑌 be oriented 𝑛-manifolds and assume that 𝑌 is connected. If
𝑓0 , 𝑓2 ∶ 𝑋 → 𝑌 are proper maps and the map (4.7.8) is a proper homotopy, then deg(𝑓1 ) =
deg(𝑓2 ).
Proof. Let 𝜔 be an 𝑛-form in 𝛺𝑛𝑐 (𝑌) whose integral over 𝑌 is equal to 1, and let 𝐶 ≔ supp(𝜔)
be the support of 𝜔. Then if 𝐹 is a proper homotopy between 𝑓0 and 𝑓1 , the set, 𝐹−1 (𝐶), is
compact and its projection on 𝑋
(4.7.17) { 𝑥 ∈ 𝑋 | (𝑥, 𝑡) ∈ 𝐹−1 (𝐶) for some 𝑡 ∈ [0, 1] }
is compact. Let 𝑓𝑡 ∶ 𝑋 → 𝑌 be the map: 𝑓𝑡 (𝑥) = 𝐹(𝑥, 𝑡). By our assumptions on 𝐹, 𝑓𝑡 is a
proper 𝐶∞ map. Moreover, for all 𝑡 the 𝑛-form 𝑓𝑡⋆ 𝜔 is a 𝐶∞ function of 𝑡 and is supported
on the fixed compact set (4.7.17). Hence it’s clear from the definitions of the integral of a
form and 𝑓𝑡 that the integral
∫ 𝑓𝑡⋆ 𝜔
𝑋
is a 𝐶∞ function of 𝑡. On the other hand this integral is by definition the degree of 𝑓𝑡 and
hence by Theorem 4.7.3, deg(𝑓𝑡 ) is an integer, so it does not depend on 𝑡. In particular,
deg(𝑓0 ) = deg(𝑓1 ). □
Exercises for §4.7
Exercise 4.7.i. Let 𝑓 ∶ 𝐑 → 𝐑 be the map 𝑥 ↦ 𝑥𝑛 . Show that deg(𝑓) = 0 if 𝑛 is even and 1
if 𝑛 is odd.
 Draft: March 28, 2018

§4.8 Applications of degree theory 137

Exercise 4.7.ii. Let 𝑓 ∶ 𝐑 → 𝐑 be the polynomial function,

𝑓(𝑥) = 𝑥𝑛 + 𝑎1 𝑥𝑛−1 + ⋯ + 𝑎𝑛−1 𝑥 + 𝑎𝑛 ,
where the 𝑎𝑖 ’s are in 𝐑. Show that if 𝑛 is even, deg(𝑓) = 0 and if 𝑛 is odd, deg(𝑓) = 1.
Exercise 4.7.iii. Let 𝑆1 be the unit circle in the complex plane and let 𝑓 ∶ 𝑆1 → 𝑆1 be the
map 𝑒𝑖𝜃 ↦ 𝑒𝑖𝑁𝜃 , where 𝑁 ias a positive integer. What is the degree of 𝑓?
Exercise 4.7.iv. Let 𝑆𝑛−1 be the unit sphere in 𝐑𝑛 and 𝜎 ∶ 𝑆𝑛−1 → 𝑆𝑛−1 the antipodal map
𝑥 ↦ −𝑥. What is the degree of 𝜎?
Exercise 4.7.v. Let 𝐴 ∈ 𝑂(𝑛) be an orthogonal 𝑛 × 𝑛 matrix and let
𝑓𝐴 ∶ 𝑆𝑛−1 → 𝑆𝑛−1
be the map 𝑥 ↦ 𝐴𝑥. What is the degree of 𝑓𝐴 ?
Exercise 4.7.vi. A manifold 𝑌 is contractable if for some point 𝑝0 ∈ 𝑌, the identity map of
𝑌 onto itself is homotopic to the constant map 𝑓𝑝0 ∶ 𝑌 → 𝑌 given by 𝑓𝑝0 (𝑦) ≔ 𝑝0 . Show
that if 𝑌 is an oriented contractable 𝑛-manifold and 𝑋 an oriented connected 𝑛-manifold
then for every proper mapping 𝑓 ∶ 𝑋 → 𝑌 we have deg(𝑓) = 0. In particular show that if
𝑛 > 0 and 𝑌 is compact then 𝑌 cannot be contractable.
Hint: Let 𝑓 be the identity map of 𝑌 onto itself.
Exercise 4.7.vii. Let 𝑋 and 𝑌 be oriented connected 𝑛-manifolds and 𝑓 ∶ 𝑋 → 𝑌 a proper
𝐶∞ map. Show that if deg(𝑓) ≠ 0, then 𝑓 is surjective.
Exercise 4.7.viii. Using Sard’s Theorem prove that if 𝑋 and 𝑌 are manifolds of dimension
𝑘 and ℓ, with 𝑘 < ℓ and 𝑓 ∶ 𝑋 → 𝑌 is a proper 𝐶∞ map, then the complement of the image
of 𝑋 in 𝑌 is open and dense.
Hint: Let 𝑟 ≔ ℓ − 𝑘 and apply Sard’s Theorem to the map
𝑔 ∶ 𝑋 × 𝑆𝑟 → 𝑌 , 𝑔(𝑥, 𝑎) ≔ 𝑓(𝑥) .
Exercise 4.7.ix. Prove that the 2-sphere 𝑆2 and the torus 𝑆1 × 𝑆1 are not diffeomorphic.

4.8. Applications of degree theory

The purpose of this section will be to describe a few typical applications of degree theory
to problems in analysis, geometry, and topology. The first of these applications will be yet
another variant of the Brouwer fixed point theorem.
Application 4.8.1. Let 𝑋 be an oriented (𝑛 + 1)-dimensional manifold, 𝐷 ⊂ 𝑋 a smooth
domain and 𝑍 ≔ 𝜕𝐷 the boundary of 𝐷. Assume that the closure 𝐷 = 𝑍∪𝐷 of 𝐷 is compact
(and in particular that 𝑋 is compact).
Theorem 4.8.2. Let 𝑌 be an oriented connected 𝑛-manifold and 𝑓 ∶ 𝑍 → 𝑌 a 𝐶∞ map. Sup-
pose there exists a 𝐶∞ map 𝐹 ∶ 𝐷 → 𝑌 whose restriction to 𝑍 is 𝑓. Then deg(𝑓) = 0.
Proof. Let 𝜇 be an element of 𝛺𝑛𝑐 (𝑌). Then 𝑑𝜇 = 0, so 𝑑𝐹⋆ 𝜇 = 𝐹⋆ 𝑑𝜇 = 0. On the other hand
if 𝜄 ∶ 𝑍 → 𝑋 is the inclusion map,

∫ 𝑑𝐹⋆ 𝜇 = ∫ 𝜄⋆ 𝐹⋆ 𝜇 = ∫ 𝑓⋆ 𝜇 = deg(𝑓) ∫ 𝜇
𝐷 𝑍 𝑍 𝑌
by Stokes’ theorem since 𝐹 ∘ 𝜄 = 𝑓. Hence deg(𝑓) has to be zero. □
 Draft: March 28, 2018

138 Chapter 4: Manifolds & Forms on Manifolds

Application 4.8.3 (a non-linear eigenvalue problem). This application is a non-linear gen-

eralization of a standard theorem in linear algebra. Let 𝐴 ∶ 𝐑𝑛 → 𝐑𝑛 be a linear map. If 𝑛 is
even, 𝐴 may not have real eigenvalues. (For instance for the map
𝐴 ∶ 𝐑2 → 𝐑2 , (𝑥, 𝑦) ↦ (−𝑦, 𝑥)
the eigenvalues of 𝐴 are ±√−1.) However, if 𝑛 is odd it is a standard linear algebra fact that
there exists a vector, 𝑣 ∈ 𝐑𝑛 ∖ {0}, and a 𝜆 ∈ 𝐑 such that 𝐴𝑣 = 𝜆𝑣. Moreover replacing 𝑣
𝑣
by |𝑣| one can assume that |𝑣| = 1. This result turns out to be a special case of a much more
general result. Let 𝑆𝑛−1 be the unit (𝑛 − 1)-sphere in 𝐑𝑛 and let 𝑓 ∶ 𝑆𝑛−1 → 𝐑𝑛 be a 𝐶∞ map.
Theorem 4.8.4. There exists a vector 𝑣 ∈ 𝑆𝑛−1 and a number 𝜆 ∈ 𝐑 such that 𝑓(𝑣) = 𝜆𝑣.
Proof. The proof will be by contradiction. If the theorem isn’t true the vectors 𝑣 and 𝑓(𝑣),
are linearly independent and hence the vector
(4.8.5) 𝑔(𝑣) = 𝑓(𝑣) − (𝑓(𝑣) ⋅ 𝑣)𝑣
is nonzero. Let
𝑔(𝑣)
(4.8.6) ℎ(𝑣) = .
|𝑔(𝑣)|
By (4.8.5)–(4.8.6), |𝑣| = |ℎ(𝑣)| = 1 and 𝑣 ⋅ ℎ(𝑣) = 0, i.e., 𝑣 and ℎ(𝑣) are both unit vectors and
are perpendicular to each other. Let
(4.8.7) 𝛾𝑡 ∶ 𝑆𝑛−1 → 𝑆𝑛−1 , 0 ≤ 𝑡 ≤ 1
be the map
(4.8.8) 𝛾𝑡 (𝑣) = cos(𝜋𝑡)𝑣 + sin(𝜋𝑡)ℎ(𝑣) .
For 𝑡 = 0 this map is the identity map and for 𝑡 = 1, it is the antipodal map 𝜎(𝑣) = 𝑣, hence
(4.8.7) asserts that the identity map and the antipodal map are homotopic and therefore
that the degree of the antipodal map is one. On the other hand the antipodal map is the
restriction to 𝑆𝑛−1 of the map (𝑥1 , …, 𝑥𝑛 ) ↦ (−𝑥1 , …, −𝑥𝑛 ) and the volume form 𝜔 on 𝑆𝑛−1
is the restriction to 𝑆𝑛−1 of the (𝑛 − 1)-form
𝑛
(4.8.9) ̂ 𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛 .
∑(−1)𝑖−1 𝑥𝑖 𝑑𝑥𝑖 ∧ ⋯ ∧ 𝑑𝑥
𝑖=1
If we replace 𝑥𝑖 by −𝑥𝑖 in equation (4.8.9) the sign of this form changes by (−1)𝑛 hence
𝜎 𝜔 = (−1)𝑛 𝜔. Thus if 𝑛 is odd, 𝜎 is an orientation reversing diffeomorphism of 𝑆𝑛−1 onto
⋆

𝑆𝑛−1 , so its degree is −1, and this contradicts what we just deduced from the existence of the
homotopy (4.8.8). □
From this argument we can deduce another interesting fact about the sphere 𝑆𝑛−1 when
𝑛 − 1 is even. For 𝑣 ∈ 𝑆𝑛−1 the tangent space to 𝑆𝑛−1 at 𝑣 is just the space,
𝑇𝑣 𝑆𝑛−1 = { (𝑣, 𝑤) | 𝑤 ∈ 𝐑𝑛 and 𝑣 ⋅ 𝑤 = 0 } ,
so a vector field on 𝑆𝑛−1 can be viewed as a function 𝑔 ∶ 𝑆𝑛−1 → 𝐑𝑛 with the property
𝑔(𝑣) ⋅ 𝑣 = 0
𝑛−1
for all 𝑣 ∈ 𝑆 . If this function is nonzero at all points, then, letting ℎ be the function (4.8.6),
and arguing as above, we’re led to a contradiction. Hence we conclude:
Theorem 4.8.10. If 𝑛 − 1 is even and 𝒗 is a vector field on the sphere 𝑆𝑛−1 , then there exists a
point 𝑝 ∈ 𝑆𝑛−1 at which 𝒗(𝑝) = 0.
 Draft: March 28, 2018

§4.8 Applications of degree theory 139

Note that if 𝑛 − 1 is odd this statement is not true. Thevector field

𝜕 𝜕 𝜕 𝜕
𝑥1 − 𝑥2 + ⋯ + 𝑥2𝑛−1 − 𝑥2𝑛
𝜕𝑥2 𝜕𝑥1 𝜕𝑥2𝑛 𝜕𝑥2𝑛−1
is a counterexample. It is nowhere vanishing and at 𝑝 ∈ 𝑆𝑛−1 is tangent to 𝑆𝑛−1 .
Application 4.8.11 (The Jordan–Brouwer separation theorem). Let 𝑋 be a compact ori-
ented (𝑛 − 1)-dimensional submanifold of 𝐑𝑛 . In this subsection of §4.8 we’ll outline a proof
of the following theorem (leaving the details as a string of exercises).
Theorem 4.8.12. If 𝑋 is connected, the complement of 𝐑𝑛 ∖ 𝑋 of 𝑋 has exactly two connected
components.
This theorem is known as the Jordan–Brouwer separation theorem (and in two dimen-
sions as the Jordan curve theorem). For simple, easy to visualize, submanifolds of 𝐑𝑛 like the
(𝑛 − 1)-sphere this result is obvious, and for this reason it’s easy to be misled into thinking
of it as being a trivial (and ot very interesting) result. However, for submanifolds of 𝐑𝑛 like
the curve in 𝐑2 depicted in the Figure 4.8.2 it’s much less obvious. (In ten seconds or less: is
the point 𝑝 in this figure inside this curve or outside?)
To determine whether a point 𝑝 ∈ 𝐑𝑛 ∖ 𝑋 is inside 𝑋 or outside 𝑋, one needs a topolog-
ical invariant to detect the difference; such an invariant is provided by the winding number.
Definition 4.8.13. For 𝑝 ∈ 𝐑𝑛 ∖ 𝑋 let
𝛾𝑝 ∶ 𝑋 → 𝑆𝑛−1
be the map
𝑥−𝑝
𝛾𝑝 (𝑥) = .
|𝑥 − 𝑝|
The winding number 𝑊(𝑋, 𝑝) of 𝑋 about 𝑝 is the degree
𝑊(𝑋, 𝑝) ≔ deg(𝛾𝑝 ) .
We show below that 𝑊(𝑋, 𝑝) = 0 if 𝑝 is outside, and if 𝑝 is inside 𝑋, then 𝑊(𝑋, 𝑝) = ±1
(depending on the orientation of 𝑋). Hence the winding number tells us which of the two
components of 𝐑𝑛 ∖ 𝑋 the point 𝑝 is contained in.
Exercises for §4.8
Exercise 4.8.i. Let 𝑈 be a connected component of 𝐑𝑛 ∖ 𝑋. Show that if 𝑝0 and 𝑝1 are in 𝑈,
𝑊(𝑋, 𝑝0 ) = 𝑊(𝑋, 𝑝1 ).
Hints:
• First suppose that the line segment,
𝑝𝑡 = (1 − 𝑡)𝑝0 + 𝑡𝑝1 , 0 ≤ 𝑡 ≤ 1
lies in 𝑈. Conclude from the homotopy invariance of degree that
𝑊(𝑋, 𝑝0 ) = 𝑊(𝑋, 𝑝𝑡 ) = 𝑊(𝑋, 𝑝1 ) .
• Show that there exists a sequence of points 𝑞1 , …, 𝑞𝑁 ∈ 𝑈, with 𝑞1 = 𝑝0 and
𝑞𝑁 = 𝑝1 , such that the line segment joining 𝑞𝑖 to 𝑞𝑖+1 is in 𝑈.
Exercise 4.8.ii. Let 𝑋 be a connected (𝑛 − 1)-dimensional submanifold of 𝐑𝑛 . Show that
𝐑𝑛 ∖ 𝑋 has at most two connected components.
Hints:
• Show that if 𝑞 is in 𝑋 there exists a small 𝜀-ball 𝐵𝜀 (𝑞) centered at 𝑞 such that
𝐵𝜀 (𝑞) − 𝑋 has two components. (See Theorem 4.2.15.
 Draft: March 28, 2018

140 Chapter 4: Manifolds & Forms on Manifolds

• Show that if 𝑝 is in 𝐑𝑛 ∖ 𝑋, there exists a sequence 𝑞1 , …, 𝑞𝑁 ∈ 𝐑𝑛 ∖ 𝑋 such that

𝑞1 = 𝑝, 𝑞𝑁 ∈ 𝐵𝜀 (𝑞) and the line segments joining 𝑞𝑖 to 𝑞𝑖+1 are in 𝐑𝑛 ∖ 𝑋.
Exercise 4.8.iii. For 𝑣 ∈ 𝑆𝑛−1 , show that 𝑥 ∈ 𝑋 is in 𝛾𝑝−1 (𝑣) if and only if 𝑥 lies on the ray
(4.8.14) 𝑝 + 𝑡𝑣 , 0 < 𝑡 < ∞ .
Exercise 4.8.iv. Let 𝑥 ∈ 𝑋 be a point on this ray. Show that
(4.8.15) (𝑑𝛾𝑝 )𝑥 ∶ 𝑇𝑝 𝑋 → 𝑇𝑣 𝑆𝑛−1
is bijective if and only if 𝑣 ∉ 𝑇𝑝 𝑋, i.e., if and only if the ray (4.8.14) is not tangent to 𝑋 at 𝑥.
Hint: 𝛾𝑝 ∶ 𝑋 → 𝑆𝑛−1 is the composition of the maps
𝜏𝑝 ∶ 𝑋 → 𝐑𝑛 ∖ {0} , 𝑥 ↦ 𝑥 − 𝑝 ,
and
𝑦
𝜋 ∶ 𝐑𝑛 ∖ {0} → 𝑆𝑛−1 , 𝑦 ↦ .
|𝑦|
Show that if 𝜋(𝑦) = 𝑣, then the kernel of (𝑑𝜋)𝑔 is the one-dimensional subspace of 𝐑𝑛
spanned by 𝑣. Conclude that if 𝑦 = 𝑥 − 𝑝 and 𝑣 = 𝑦/|𝑦| the composite map
(𝑑𝛾𝑝 )𝑥 = (𝑑𝜋)𝑦 ∘ (𝑑𝜏𝑝 )𝑥
is bijective if and only if 𝑣 ∉ 𝑇𝑥 𝑋.
Exercise 4.8.v. From Exercises 4.8.iii and 4.8.iv deduce that 𝑣 is a regular value of 𝛾𝑝 if
and only if the ray (4.8.14) intersects 𝑋 in a finite number of points and at each point of
intersection is not tangent to 𝑋 at that point.
Exercise 4.8.vi. In Exercise 4.8.v show that the map (4.8.15) is orientation preserving if
the orientations of 𝑇𝑥 𝑋 and 𝑣 are compatible with the standard orientation of 𝑇𝑝 𝐑𝑛 . (See
Exercise 1.6.v.)
Exercise 4.8.vii. Conclude that deg(𝛾𝑝 ) counts (with orientations) the number of points
where the ray (4.8.14) intersects 𝑋.
Exercise 4.8.viii. Let 𝑝1 ∈ 𝐑𝑛 ∖ 𝑋 be a point on the ray (4.8.14). Show that if 𝑣 ∈ 𝑆𝑛−1 is a
regular value of 𝛾𝑝 , it is a regular value of 𝛾𝑝1 and show that the number
deg(𝛾𝑝 ) − deg(𝛾𝑝1 ) = 𝑊(𝑋, 𝑝) − 𝑊(𝑋, 𝑝1 )
counts (with orientations) the number of points on the ray lying between 𝑝 and 𝑝1 .
Hint: Exercises 4.8.v and 4.8.vii.
Exercise 4.8.ix. Let 𝑥 ∈ 𝑋 be a point on the ray (4.8.14). Suppose 𝑥 = 𝑝 + 𝑡𝑣. Show that if
𝜀 is a small positive number and
𝑝± = 𝑝 + (𝑡 ± 𝜀)𝑣
then
𝑊(𝑋, 𝑝+ ) = 𝑊(𝑋, 𝑝− ) ± 1 ,
and from Exercise 4.8.i conclude that 𝑝+ and 𝑝− lie in different components of 𝐑𝑛 ∖ 𝑋. In
particular conclude that 𝐑𝑛 ∖ 𝑋 has exactly two components.
 Draft: March 28, 2018

§4.8 Applications of degree theory 141

Exercise 4.8.x. Finally show that if 𝑝 is very large the difference

𝑝
𝛾𝑝 (𝑥) − , 𝑥∈𝑋,
|𝑝|
is very small, i.e., 𝛾𝑝 is not surjective and hence the degree of 𝛾𝑝 is zero. Conclude that for
𝑝 ∈ 𝐑𝑛 ∖ 𝑋, 𝑝 is in the unbounded component of 𝐑𝑛 ∖ 𝑋 if 𝑊(𝑋, 𝑝) = 0 and in the bounded
component if 𝑊(𝑋, 𝑝) = ±1 (the “±” depending on the orientation of 𝑋).
Remark 4.8.16. The proof of Jordan–Brouwer sketched above gives us an effective way of
deciding whether the point 𝑝 in Figure 4.8.2, is inside 𝑋 or outside 𝑋. Draw a non-tangential
ray from 𝑝. If it intersects 𝑋 in an even number of points, 𝑝 is outside 𝑋 and if it intersects
𝑋 is an odd number of points 𝑝 inside.

inside
𝑞

outside

Figure 4.8.1. A Jordan curve with 𝑝 inside of the curve and 𝑞 outside

Application 4.8.17 (The Gauss–Bonnet theorem). Let 𝑋 ⊂ 𝐑𝑛 be a compact, connected, ori-

ented (𝑛 − 1)-dimensional submanifold. By the Jordan–Brouwer theorem 𝑋 is the boundary
of a bounded smooth domain, so for each 𝑥 ∈ 𝑋 there exists a unique outward pointing unit
normal vector 𝑛𝑥 . The Gauss map
𝛾 ∶ 𝑋 → 𝑆𝑛−1
is the map 𝑥 ↦ 𝑛𝑥 . Let 𝜎 be the Riemannian volume form of 𝑆𝑛−1 , or, in other words, the
restriction to 𝑆𝑛−1 of the form,
𝑛
̂ 𝑖 ∧ ⋯ ∧ 𝑑𝑥𝑛 ,
∑(−1)𝑖−1 𝑥𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑖=1
and let 𝜎𝑋 be the Riemannian volume form of 𝑋. Then for each 𝑝 ∈ 𝑋
(4.8.18) (𝛾⋆ 𝜎)𝑝 = 𝐾(𝑝)(𝜎𝑋 )𝑞
where 𝐾(𝑝) is the scalar curvature of 𝑋 at 𝑝. This number measures the extent to which
“𝑋 is curved” at 𝑝. For instance, if 𝑋𝑎 is the circle, |𝑥| = 𝑎 in 𝐑2 , the Gauss map is the map
𝑝 → 𝑝/𝑎, so for all 𝑝 we have 𝐾𝑎 (𝑝) = 1/𝑎, reflecting the fact that for 𝑎 < 𝑏, 𝑋𝑎 is more
curved than 𝑋𝑏 .
The scalar curvature can also be negative. For instance for surfaces 𝑋 in 𝐑3 , 𝐾(𝑝) is
positive at 𝑝 if 𝑋 is convex at 𝑝 and negative if 𝑋 is convex–concave at 𝑝. (See Figures 4.8.2
 Draft: March 28, 2018

142 Chapter 4: Manifolds & Forms on Manifolds

and 4.8.3 below. The surface in Figure 4.8.2 is convex at 𝑝, and the surface in Figure 4.8.3 is
convex–concave.)

𝑝 𝛾

𝛾(𝑝)

Figure 4.8.2. Positive scalar curvature at 𝑝

𝛾
𝑝
𝛾(𝑝)

Figure 4.8.3. Negative scalar curvature at 𝑝

Let vol(𝑆𝑛−1 ) be the Riemannian volume of the (𝑛 − 1)-sphere, i.e., let

2𝜋𝑛/2
vol(𝑆𝑛−1 ) =
𝛤(𝑛/2)
(where 𝛤 is the Gamma function). Then by (4.8.18) the quotient
∫ 𝐾𝜎𝑋
(4.8.19)
vol(𝑆𝑛−1 )
is the degree of the Gauss map, and hence is a topological invariant of the surface of 𝑋. For

𝛾 𝑣
𝑝1 𝑝2 … 𝑝𝑔 𝑝0 0 𝑞

surface 𝑋 of genus 𝑔 the sphere 𝑆2

Figure 4.8.4. The Gauss map from a surface of genus 𝑔 to the sphere
 Draft: March 28, 2018

§4.9 The index of a vector field 143

𝑛 = 3 the Gauss–Bonnet theorem asserts that this topological invariant is just 1 − 𝑔 where
𝑔 is the genus of 𝑋 or, in other words, the “number of holes”. Figure 4.8.4 gives a pictorial
proof of this result. (Notice that at the points 𝑝1 , …, 𝑝𝑔 the surface 𝑋 is convex–concave so
the scalar curvature at these points is negative, i.e., the Gauss map is orientation reversing.
On the other hand, at the point 𝑝0 the surface is convex, so the Gauss map at this point is
orientation preserving.)

4.9. The index of a vector field

𝑛
Let 𝐷 be a bounded smooth domain in 𝐑𝑛 and 𝒗 = ∑𝑖=1 𝑣𝑖 𝜕/𝜕𝑥𝑖 a 𝐶∞ vector field de-
fined on the closure 𝐷 of 𝐷. We will define below a topological invariant which, for “generic”
𝒗’s, counts (with appropriate ±-signs) the number of zeros of 𝒗. To avoid complications we’ll
assume 𝒗 has no zeros on the boundary of 𝐷.
Definition 4.9.1. Let 𝜕𝐷 be the boundary of 𝐷 and let
𝑓𝒗 ∶ 𝜕𝐷 → 𝑆𝑛−1
be the mapping
𝒗(𝑝)
𝑝↦ ,
|𝒗(𝑝)|
where 𝒗(𝑝) = (𝑣1 (𝑝), …, 𝑣𝑛 (𝑝)). The index of 𝒗 with respect to 𝐷 is by definition the degree
of 𝑓𝒗 .
We’ll denote this index by ind(𝒗, 𝐷) and as a first step in investigating its properties
we’ll prove:
Theorem 4.9.2. Let 𝐷1 be a smooth domain in 𝐑𝑛 whose closure is contained in 𝐷. Then
ind(𝒗, 𝐷) = ind(𝒗, 𝐷1 )
provided that 𝒗 has no zeros in the set 𝐷 ∖ 𝐷1 .

Proof. Let 𝑊 = 𝐷 ∖ 𝐷1 . Then the map 𝑓𝒗 ∶ 𝜕𝐷 → 𝑆𝑛−1 extends to a 𝐶∞ map

𝒗(𝑝)
𝐹 ∶ 𝑊 → 𝑆𝑛−1 , 𝑝 ↦ .
|𝒗(𝑝)|
Moreover,
𝜕𝑊 = 𝑋 ∪ 𝑋1−
where 𝑋 is the boundary of 𝐷 with its natural boundary orientation and 𝑋1− is the boundary
of 𝐷1 with its boundary orientation reversed. Let 𝜔 be an element of 𝛺𝑛−1 (𝑆𝑛−1 ) whose
integral over 𝑆𝑛−1 is 1. Then if 𝑓 = 𝑓𝒗 and 𝑓1 = (𝑓1 )𝒗 are the restrictions of 𝐹 to 𝑋 and 𝑋1
we get from Stokes’ theorem (and the fact that 𝑑𝜔 = 0) the identity

0 = ∫ 𝐹⋆ 𝑑𝜔 = ∫ 𝑑𝐹⋆ 𝜔
𝑊 𝑊

= ∫ 𝑓 𝜔 − ∫ 𝑓1⋆ 𝜔 = deg(𝑓) − deg(𝑓1 ) .

⋆
𝑋 𝑋1

Hence
ind(𝒗, 𝐷) = deg(𝑓) = deg(𝑓1 ) = ind(𝒗, 𝐷) . □
 Draft: March 28, 2018

144 Chapter 4: Manifolds & Forms on Manifolds

Suppose now that 𝒗 has a finite number of isolated zeros 𝑝1 , …, 𝑝𝑘 in 𝐷. Let 𝐵𝜀 (𝑝𝑖 ) be
the open ball of radius 𝜀 with center at 𝑝𝑖 . By making 𝜀 small enough we can assume that
each of these balls is contained in 𝐷 and that they’re mutually disjoint. We will define the
local index ind(𝒗, 𝑝𝑖 ) of 𝒗 at 𝑝𝑖 to be the index of 𝒗 with respect to 𝐵𝜀 (𝑝𝑖 ). By Theorem 4.9.2
these local indices are unchanged if we replace 𝜀 by a smaller 𝜀, and by applying this theorem
to the domain
𝑘
𝐷1 = ⋃ 𝐵𝑝𝑖 (𝜀)
𝑖=1
we get, as a corollary of the theorem, the formula
𝑘
(4.9.3) ind(𝒗, 𝐷) = ∑ ind(𝒗, 𝑝𝑖 )
𝑖=1

which computes the global index of 𝒗 with respect to 𝐷 in terms of these local indices.
Let’s now see how to compute these local indices. Let’s first suppose that the point 𝑝 = 𝑝𝑖
is at the origin of 𝐑𝑛 . Then near 𝑝 = 0
𝒗 = 𝒗lin + 𝒗′
where
𝜕
𝒗 = 𝒗lin = ∑ 𝑎𝑖,𝑗 𝑥𝑖
1≤𝑖,𝑗≤𝑛 𝜕𝑥𝑗
the 𝑎𝑖,𝑗 ’s being constants and
𝑛
𝜕
𝒗𝑖 = ∑ 𝑓𝑖𝑗
𝑗=1 𝜕𝑥𝑗
where the 𝑓𝑖𝑗 ’s vanish to second order near zero, i.e., satisfy
(4.9.4) |𝑓𝑖,𝑗 (𝑥)| ≤ 𝐶|𝑥|2
for some constant 𝐶 > 0.
Definition 4.9.5. We’ll say that the point 𝑝 = 0 is a non-degenerate zero of 𝒗 is the matrix
𝐴 = (𝑎𝑖,𝑗 ) is non-singular.
This means in particular that
𝑛 𝑛
(4.9.6) ∑ |∑ 𝑎𝑖,𝑗 𝑥𝑖 | ≥ 𝐶1 |𝑥|
𝑗=1 𝑖=1

form some constant 𝐶1 > 0. Thus by (4.9.4) and (4.9.6) the vector field
𝒗𝑡 = 𝒗lin + 𝑡𝒗′ , 0 ≤ 𝑡 ≤ 1
has no zeros in the ball 𝐵𝜀 (0), other than the point 𝑝 = 0 itself, provided that 𝜀 is small
enough. Therefore if we let 𝑋𝜀 be the boundary of 𝐵𝜀 (0) we get a homotopy
𝐹 ∶ 𝑋𝜀 × [0, 1] → 𝑆𝑛−1 , (𝑥, 𝑡) ↦ 𝑓𝒗𝑡 (𝑥)
between the maps 𝑓𝒗lin ∶ 𝑋𝜀 → 𝑆𝑛−1 and 𝑓𝒗 ∶ 𝑋𝜀 → 𝑆𝑛−1 , thus by Theorem 4.7.16 we see that
deg(𝑓𝒗 ) = deg(𝑓𝒗lin ). Therefore,
(4.9.7) ind(𝒗, 𝑝) = ind(𝒗lin , 𝑝) .
 Draft: March 28, 2018

§4.9 The index of a vector field 145

We have assumed in the above discussion that 𝑝 = 0, but by introducing at 𝑝 a translated

coordinate system for which 𝑝 is the origin, these comments are applicable to any zero 𝑝 of
𝒗. More explicitly if 𝑐1 , …, 𝑐𝑛 are the coordinates of 𝑝 then as above
𝒗 = 𝒗lin + 𝒗′
where
𝜕
𝒗lin ≔ ∑ 𝑎𝑖,𝑗 (𝑥𝑖 − 𝑐𝑖 )
1≤𝑖,𝑗≤𝑛 𝜕𝑥𝑗

and 𝒗′ vanishes to second order at 𝑝, and if 𝑝 is a non-degenerate zero of 𝑞, i.e., if the matrix
𝐴 = (𝑎𝑖,𝑗 ) is non-singular, then exactly the same argument as before shows that
ind(𝒗, 𝑝) = ind(𝒗lin , 𝑝) .
We will next prove:
Theorem 4.9.8. If 𝑝 is a non-degenerate zero of 𝒗, the local index ind(𝒗, 𝑝) is +1 or −1 de-
pending on whether the determinant of the matrix, 𝐴 = (𝑎𝑖,𝑗 ) is positive or negative.
Proof. As above we can, without loss of generality, assume that 𝑝 = 0. By (4.9.7) we can
assume 𝒗 = 𝒗lin . Let 𝐷 be the domain defined by
𝑛 𝑛 2
(4.9.9) ∑ (∑ 𝑎𝑖,𝑗 𝑥𝑖 ) < 1 ,
𝑗=1 𝑖=1

and let 𝑋 ≔ 𝜕𝐷 be its boundary. Since the only zero of 𝒗lin inside this domain is 𝑝 = 0 we
get from (4.9.3) and (4.9.4) that
ind(𝒗, 𝑝) = ind(𝒗lin , 𝑝) = ind(𝒗lin , 𝐷) .
Moreover, the map
𝑓𝒗lin ∶ 𝑋 → 𝑆𝑛−1
is, in view of (4.9.9), just the linear map 𝑣 ↦ 𝐴𝑣, restricted to 𝑋. In particular, since 𝐴 is
a diffeomorphism, this mapping is a diffeomorphism as well, so the degree of this map is
+1 or −1 depending on whether this map is orientation preserving or not. To decide which
𝑛 ̂ 𝑖 ∧ ⋯ 𝑑𝑥𝑛 be the Riemannian
of these alternatives is true let 𝜔 = ∑𝑖=1 (−1)𝑖 𝑥𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥
𝑛−1
volume form on 𝑆 then

∫ 𝑓𝒗⋆lin 𝜔 = deg(𝑓𝒗lin ) vol(𝑆𝑛−1 ) .

𝑋

Since 𝒗 is the restriction of 𝐴 to 𝑋. By Stokes’ theorem this is equal to

∫ 𝐴⋆ 𝑑𝜔 = 𝑛 ∫ 𝐴⋆ 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛
𝐷 𝐷

= 𝑛 det(𝐴) ∫ 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛

𝐷
= 𝑛 det(𝐴) vol(𝐷)
which gives us the formula
𝑛 det 𝐴 vol(𝐷)
deg(𝑓𝒗lin , 𝐷) = . □
vol(𝑆𝑛−1 )
 Draft: March 28, 2018

146 Chapter 4: Manifolds & Forms on Manifolds

We’ll briefly describe the various types of non-degenerate zeros that can occur for vector
fields in two dimensions. To simplify this description a bit we’ll assume that the matrix
𝑎11 𝑎12
𝐴=( )
𝑎21 𝑎22
is diagonalizable and that its eigenvalues are not purely imaginary. Then the following five
scenarios can occur:
(1) The eigenvalues of 𝐴 are real and positive. In this case the integral curves of 𝒗lin in a neigh-
borhood of 𝑝 look like the curves in Figure 4.9.1 and hence, in a small neighborhood

Figure 4.9.1. Real and positive eigenvalues

of 𝑝, the integral sums of 𝒗 itself look approximately like the curve in Figure 4.9.1.
(2) The eigenvalues of 𝐴 are real and negative. In this case the integral curves of 𝒗lin look like
the curves in Figure 4.9.1, but the arrows are pointing into 𝑝, rather than out of 𝑝, i.e.,

Figure 4.9.2. Real and negative eigenvalues

(3) The eigenvalues of 𝐴 are real, but one is positive and one is negative. In this case the
integral curves of 𝒗lin look like the curves in Figure 4.9.3.
(4) The eigenvalues of 𝐴 are complex and of the form 𝑎 ± √−𝑏, with 𝑎 positive. In this case
the integral curves of 𝒗lin are spirals going out of 𝑝 as in Figure 4.9.4.
(5) The eigenvalues of 𝐴 are complex and of the form 𝑎 ± √−𝑏, with 𝑎 negative. In this case
the integral curves are as in Figure 4.9.4 but are spiraling into 𝑝 rather than out of 𝑝.

Definition 4.9.10. Zeros of 𝒗 of types (1) and (4) are called sources; those of types (2) and
(5) are called sinks and those of type (3) are called saddles.
 Draft: March 28, 2018

§4.9 The index of a vector field 147

Figure 4.9.3. Real eigenvalues, one positive and one negative

Figure 4.9.4. Complex eigenvalues of the form 𝑎 ± √−𝑏, with 𝑎 positive.

Thus in particular the two-dimensional version of equation (4.9.3) tells us:

Theorem 4.9.11. If the vector field 𝒗 on 𝐷 has only non-degenerate zeros then ind(𝒗, 𝐷) is
equal to the number of sources plus the number of sinks minus the number of saddles.
Draft: March 28, 2018
 Draft: March 28, 2018

CHAPTER 5

Cohomology via forms

5.1. The de Rham cohomology groups of a manifold

In the last four chapters we’ve frequently encountered the question: When is a closed
𝑘-form on an open subset of 𝐑𝑁 (or, more generally on a submanifold of 𝐑𝑁 ) exact? To inves-
tigate this question more systematically than we’ve done heretofore, let 𝑋 be an 𝑛-manifold
and let
𝑍𝑘 (𝑋) ≔ { 𝜔 ∈ 𝛺𝑘 (𝑋) | 𝑑𝜔 = 0 }
and
𝐵𝑘 (𝑋) ≔ 𝑑(𝛺𝑘−1 (𝑋)) ⊂ 𝛺𝑘 (𝑋)
be the vector spaces of closed and exact 𝑘-forms. Since 𝐵𝑘 (𝑋) is a vector subspace of 𝑍𝑘 (𝑋),
we can form the quotient space
𝐻𝑘 (𝑋) ≔ 𝑍𝑘 (𝑋)/𝐵𝑘 (𝑋) ,
and the dimension of this space is a measure of the extent to which closed forms fail to be
exact. We will call this space the 𝑘th de Rham cohomology group of the manifold 𝑋. Since
the vector spaces 𝑍𝑘 (𝑋) and 𝐵𝑘 (𝑋) are both infinite dimensional in general, there is no
guarantee that this quotient space is finite dimensional, however, we’ll show later in this
chapter that it is in a lot of interesting cases.
The spaces 𝐻𝑘 (𝑋) also have compactly supported counterparts. Namely let
𝑍𝑐𝑘 (𝑋) ≔ { 𝜔 ∈ 𝛺𝑘𝑐 (𝑋) | 𝑑𝜔 = 0 }
and
𝐵𝑐𝑘 (𝑋) ≔ 𝑑(𝛺𝑘−1 𝑘
𝑐 (𝑋)) ⊂ 𝛺𝑐 (𝑋) .
Then as above 𝐵𝑐𝑘 (𝑋) is a vector subspace of 𝑍𝑐𝑘 (𝑋) and the vector space quotient
𝐻𝑐𝑘 (𝑋) ≔ 𝑍𝑐𝑘 (𝑋)/𝐵𝑐𝑘 (𝑋)
is the 𝑘th compactly supported de Rham cohomology group of 𝑋.
Given a closed 𝑘-form 𝜔 ∈ 𝑍𝑘 (𝑋) we will denote by [𝜔] the image of 𝜔 in the quotient
space 𝐻𝑘 (𝑋) = 𝑍𝑘 (𝑋)/𝐵𝑘 (𝑋) and call [𝜔] the cohomology class of 𝜔. We will also use the
same notation for compactly supported cohomology. If 𝜔 is in 𝑍𝑐𝑘 (𝑋) we’ll denote by [𝜔]
the cohomology class of 𝜔 in the quotient space 𝐻𝑐𝑘 (𝑋) = 𝑍𝑐𝑘 (𝑋)/𝐵𝑐𝑘 (𝑋).
Some cohomology groups of manifolds we’ve already computed in the previous chap-
ters (although we didn’t explicitly describe these computations as “computing cohomol-
ogy”). We’ll make a list below of some of the properties that we have already discovered
about the de Rham cohomology of a manifold 𝑋:
(1) If 𝑋 is connected, 𝐻0 (𝑋) = 𝐑. To see this, note that a closed zero form is a function
𝑓 ∈ 𝐶∞ (𝑋) having the property 𝑑𝑓 = 0, and if 𝑋 is connected the only such functions
are constants.
149
 Draft: March 28, 2018

150 Chapter 5: Cohomology via forms

(2) If 𝑋 is connected and non-compact 𝐻𝑐0 (𝑋) = 0. To see this, note that if 𝑓 is in 𝐶0∞ (𝑋)
and 𝑋 is non-compact, 𝑓 has to be zero at some point, and hence if 𝑑𝑓 = 0, then 𝑓 has
to be identically zero.
(3) If 𝑋 is 𝑛-dimensional,
𝛺𝑘 (𝑋) = 𝛺𝑘𝑐 (𝑋) = 0
for 𝑘 < 0 or 𝑘 > 𝑛, hence
𝐻𝑘 (𝑋) = 𝐻𝑐𝑘 (𝑋) = 0
for 𝑘 < 0 or 𝑘 > 𝑛.
(4) If 𝑋 is an oriented, connected 𝑛-manifold, the integration operation is a linear map

(5.1.1) ∫ ∶ 𝛺𝑛𝑐 (𝑋) → 𝐑

𝑋
and, by Theorem 4.8.2, the kernel of this map is 𝐵𝑐𝑛 (𝑋). Moreover, in degree 𝑛, 𝑍𝑐𝑛 (𝑋) =
𝛺𝑛𝑐 (𝑋) and hence by the definition of 𝐻𝑐𝑘 (𝑋) as the quotient 𝑍𝑐𝑘 (𝑋)/𝐵𝑐𝑘 (𝑋), we get from
(5.1.1) a bijective map
(5.1.2) 𝐼𝑋 ∶ 𝐻𝑐𝑛 (𝑋) ⥲ 𝐑 .
In other words 𝐻𝑐𝑛 (𝑋) = 𝐑.
(5) Let 𝑈 be a star-shaped open subset of 𝐑𝑛 . In Exercises 2.6.iv to 2.6.vii we sketched a
proof of the assertion: For 𝑘 > 0 every closed form 𝜔 ∈ 𝑍𝑘 (𝑈) is exact. Translating this
assertion into cohomology language, we showed that 𝐻𝑘 (𝑈) = 0 for 𝑘 > 0.
(6) Let 𝑈 ⊂ 𝐑𝑛 be an open rectangle. In Exercises 3.2.iv to 3.2.vii we sketched a proof of the
assertion: If 𝜔 ∈ 𝛺𝑘𝑐 (𝑈) is closed and 𝑘 is less than 𝑛, then 𝜔 = 𝑑𝜇 for some (𝑘 − 1)-form
𝜇 ∈ 𝛺𝑘−1 𝑘
𝑐 (𝑈). Hence we showed 𝐻𝑐 (𝑈) = 0 for 𝑘 < 𝑛.
(7) Poincaré lemma for manifolds: Let 𝑋 be an 𝑛-manifold and 𝜔 ∈ 𝑍𝑘 (𝑋), 𝑘 > 0 a closed
𝑘-form. Then for every point 𝑝 ∈ 𝑋 there exists a neighborhood 𝑈 of 𝑝 and a (𝑘 − 1)-
form 𝜇 ∈ 𝛺𝑘−1 (𝑈) such that 𝜔 = 𝑑𝜇 on 𝑈. To see this note that for open subsets of 𝐑𝑛
we proved this result in Section 2.4 and since 𝑋 is locally diffeomorphic at 𝑝 to an open
subset of 𝐑𝑛 this result is true for manifolds as well.
(8) Let 𝑋 be the unit sphere 𝑆𝑛 in 𝐑𝑛+1 . Since 𝑆𝑛 is compact, connected and oriented 𝐻0 (𝑆𝑛 ) =
𝐻𝑛 (𝑆𝑛 ) = 𝐑.
We will show that for 𝑘 ≠, 0, 𝑛
𝐻𝑘 (𝑆𝑛 ) = 0 .
To see this let 𝜔 ∈ 𝛺𝑘 (𝑆𝑛 ) be a closed 𝑘-form and let 𝑝 = (0, …, 0, 1) ∈ 𝑆𝑛 be the “north
pole” of 𝑆𝑛 . By the Poincaré lemma there exists a neighborhood 𝑈 of 𝑝 in 𝑆𝑛 and a 𝑘 − 1-
form 𝜇 ∈ 𝛺𝑘−1 (𝑈) with 𝜔 = 𝑑𝜇 on 𝑈. Let 𝜌 ∈ 𝐶0∞ (𝑈) be a “bump function” which is
equal to one on a neighborhood 𝑈0 of 𝑈 in 𝑝. Then
(5.1.3) 𝜔1 = 𝜔 − 𝑑𝜌𝜇
is a closed 𝑘-form with compact support in 𝑆𝑛 ∖ {𝑝}. However stereographic projection
gives one a diffeomorphism
𝜙 ∶ 𝐑𝑛 → 𝑆𝑛 ∖ {𝑝}
(see Exercise 5.1.i), and hence 𝜙⋆ 𝜔1 is a closed compactly supported 𝑘-form on 𝐑𝑛 with
support in a large rectangle. Thus by (5.1.3) 𝜙⋆ 𝜔 = 𝑑𝜈, for some 𝜈 ∈ 𝛺𝑘−1 𝑛
𝑐 (𝐑 ), and by
(5.1.3)
(5.1.4) 𝜔 = 𝑑(𝜌𝜇 + (𝜙−1 )⋆ 𝜈)
 Draft: March 28, 2018

§5.1 The de Rham cohomology groups of a manifold 151

with (𝜙−1 )⋆ 𝜈 ∈ 𝛺𝑘−1 𝑛 𝑘 𝑛

𝑐 (𝑆 ∖{𝑝}) ⊂ 𝛺 (𝑆 ), so we’ve proved that for 0 < 𝑘 < 𝑛 every closed
𝑛
𝑘-form on 𝑆 is exact.
We will next discuss some “pullback” operations in de Rham theory. Let 𝑋 and 𝑌 be
manifolds and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map. For 𝜔 ∈ 𝛺𝑘 (𝑌), 𝑑𝑓⋆ 𝜔 = 𝑓⋆ 𝑑𝜔, so if 𝜔 is closed, 𝑓⋆ 𝜔
is as well. Moreover, if 𝜔 = 𝑑𝜇, 𝑓⋆ 𝜔 = 𝑑𝑓⋆ 𝜇, so if 𝜔 is exact, 𝑓⋆ 𝜔 is as well. Thus we have
linear maps
𝑓⋆ ∶ 𝑍𝑘 (𝑌) → 𝑍𝑘 (𝑋)
and
(5.1.5) 𝑓⋆ ∶ 𝐵𝑘 (𝑌) → 𝐵𝑘 (𝑋)
and composing 𝑓⋆ ∶ 𝑍𝑘 (𝑌) → 𝑍𝑘 (𝑋) with the projection
𝜋 ∶ 𝑍𝑘 (𝑋) → 𝑍𝑘 (𝑋)/𝐵𝑘 (𝑋)
we get a linear map
(5.1.6) 𝑍𝑘 (𝑌) → 𝐻𝑘 (𝑋) .
In view of (5.1.5), 𝐵𝑘 (𝑌) is in the kernel of this map, so by Proposition 1.2.9 one gets an
induced linear map
𝑓♯ ∶ 𝐻𝑘 (𝑌) → 𝐻𝑘 (𝑌) ,
such that 𝑓♯ ∘ 𝜋 is the map (5.1.6). In other words, if 𝜔 is a closed 𝑘-form on 𝑌, then 𝑓♯ has
the defining property
(5.1.7) 𝑓♯ [𝜔] = [𝑓⋆ 𝜔] .
This “pullback” operation on cohomology satisfies the following chain rule: Let 𝑍 be a
manifold and 𝑔 ∶ 𝑌 → 𝑍 a 𝐶∞ map. Then if 𝜔 is a closed 𝑘-form on 𝑍
(𝑔 ∘ 𝑓)⋆ 𝜔 = 𝑓⋆ 𝑔⋆ 𝜔
by the chain rule for pullbacks of forms, and hence by (5.1.7)
(5.1.8) (𝑔 ∘ 𝑓)♯ [𝜔] = 𝑓♯ (𝑔♯ [𝜔]) .
The discussion above carries over verbatim to the setting of compactly supported de
Rham cohomology: If 𝑓 ∶ 𝑋 → 𝑌 is a proper 𝐶∞ map 𝑓 induces a pullback map on coho-
mology
𝑓♯ ∶ 𝐻𝑐𝑘 (𝑌) → 𝐻𝑐𝑘 (𝑋)
and if 𝑓 ∶ 𝑋 → 𝑌 and 𝑔 ∶ 𝑌 → 𝑍 are proper 𝐶∞ maps then the chain rule (5.1.8) holds for
compactly supported de Rham cohomology as well as for ordinary de Rham cohomology.
Notice also that if 𝑓 ∶ 𝑋 → 𝑌 is a diffeomorphism, we can take 𝑍 to be 𝑋 itself and 𝑔 to be
𝑓−1 , and in this case the chain rule tells us that the induced maps maps 𝑓♯ ∶ 𝐻𝑘 (𝑌) → 𝐻𝑘 (𝑌)
and 𝑓♯ ∶ 𝐻𝑐𝑘 (𝑌) → 𝐻𝑐𝑘 (𝑋) are bijections, i.e., 𝐻𝑘 (𝑋) and 𝐻𝑘 (𝑌) and 𝐻𝑐𝑘 (𝑋) and 𝐻𝑐𝑘 (𝑌) are
isomorphic as vector spaces.
We will next establish an important fact about the pullback operation 𝑓♯ ; we’ll show
that it’s a homotopy invariant of 𝑓. Recall that two 𝐶∞ maps
(5.1.9) 𝑓0 , 𝑓1 ∶ 𝑋 → 𝑌
∞
are homotopic if there exists a 𝐶 map
𝐹 ∶ 𝑋 × [0, 1] → 𝑌
with the property 𝐹(𝑝, 0) = 𝑓0 (𝑝) and 𝐹(𝑝, 1) = 𝑓1 (𝑝) for all 𝑝 ∈ 𝑋. We will prove:
 Draft: March 28, 2018

152 Chapter 5: Cohomology via forms

Theorem 5.1.10. If the maps (5.1.9) are homotopic then, for the maps they induce on coho-
mology
♯ ♯
(5.1.11) 𝑓0 = 𝑓1 .
Our proof of this will consist of proving this for an important special class of homo-
topies, and then by “pullback” tricks deducing this result for homotopies in general. Let 𝒗
be a complete vector field on 𝑋 and let
𝑓𝑡 ∶ 𝑋 → 𝑋 , − ∞ < 𝑡 < ∞
be the one-parameter group of diffeomorphisms that 𝒗 generates. Then
𝐹 ∶ 𝑋 × [0, 1] → 𝑋 , 𝐹(𝑝, 𝑡) ≔ 𝑓𝑡 (𝑝) ,
is a homotopy between 𝑓0 and 𝑓1 , and we’ll show that for this homotopic pair (5.1.11) is
true.

Proof. Recall that for 𝜔 ∈ 𝛺𝑘 (𝑋)

𝑑 ⋆
𝑓 𝜔 | = 𝐿𝒗 = 𝜄𝒗 𝑑𝜔 + 𝑑𝜄𝒗 𝜔
𝑑𝑡 𝑡 𝑡=0
and more generally for all 𝑡 we have
𝑑 ⋆ 𝑑 ⋆ 𝑑
𝑓𝑡 𝜔 = 𝑓𝑠+𝑡 𝜔 | = (𝑓 ∘ 𝑓 )⋆ 𝜔 |
𝑑𝑡 𝑑𝑠 𝑠=0 𝑑𝑠 𝑠 𝑡 𝑠=0
𝑑 ⋆ ⋆ 𝑑
= 𝑓 𝑓 𝜔 | = 𝑓𝑡⋆ 𝑓⋆ 𝜔 |
𝑑𝑠 𝑡 𝑠 𝑠=0 𝑑𝑠 𝑠 𝑠=0
= 𝑓𝑡⋆ 𝐿𝒗 𝜔
= 𝑓𝑡⋆ 𝜄𝒗 𝑑𝜔 + 𝑑𝑓𝑡⋆ 𝜄𝒗 𝜔 .
Thus if we set
𝑄𝑡 𝜔 ≔ 𝑓𝑡⋆ 𝜄𝒗 𝜔
we get from this computation:
𝑑 ⋆
𝑓 𝜔 = 𝑑𝑄𝑡 + 𝑄𝑡 𝑑𝜔
𝑑𝑡
and integrating over 0 ≤ 𝑡 ≤ 1:
(5.1.12) 𝑓1⋆ 𝜔 − 𝑓0⋆ 𝜔 = 𝑑𝑄𝜔 + 𝑄𝑑𝜔
where 𝑄 ∶ 𝛺𝑘 (𝑌) → 𝛺𝑘−1 (𝑋) is the operator
1
𝑄𝜔 = ∫ 𝑄𝑡 𝜔𝑑𝑡 .
0
The identity (5.1.11) is an easy consequence of this “chain homotopy” identity. If 𝜔 is in
𝑍𝑘 (𝑋), then 𝑑𝜔 = 0, so
𝑓1⋆ 𝜔 − 𝑓0⋆ 𝜔 = 𝑑𝑄𝜔
and
♯ ♯
𝑓1 [𝜔] − 𝑓0 [𝜔] = [𝑓1⋆ 𝜔 − 𝑓0⋆ 𝜔] = 0 . □

We’ll now describe how to extract from this result a proof of Theorem 5.1.10 for any
pair of homotopic maps. We’ll begin with the following useful observation.
 Draft: March 28, 2018

§5.1 The de Rham cohomology groups of a manifold 153

Proposition 5.1.13. If 𝑓0 , 𝑓1 ∶ 𝑋 → 𝑌 are homotopic 𝐶∞ mappings, then there exists a 𝐶∞

map
𝐹∶ 𝑋 × 𝐑 → 𝑌
such that the restriction of 𝐹 to 𝑋 × [0, 1] is a homotopy between 𝑓0 and 𝑓1 .
Proof. Let 𝜌 ∈ 𝐶0∞ (𝐑), 𝜌 ≥ 0, be a bump function which is supported on the interval [ 14 , 34 ]
and is positive at 𝑡 = 21 . Then
𝑡
∫−∞ 𝜌(𝑠)𝑑𝑠
𝜒(𝑡) = ∞
∫−∞ 𝜌(𝑠)𝑑𝑠
is a function which is zero on the interval (−∞, 14 ], is 1 on the interval [ 43 , ∞), and, for all 𝑡,
lies between 0 and 1. Now let
𝐺 ∶ 𝑋 × [0, 1] → 𝑌
be a homotopy between 𝑓0 and 𝑓1 and let 𝐹 ∶ 𝑋 × 𝐑 → 𝑌 be the map
(5.1.14) 𝐹(𝑥, 𝑡) = 𝐺(𝑥, 𝜒(𝑡)) .
∞
This is a 𝐶 map and since
𝐹(𝑥, 1) = 𝐺(𝑥, 𝜒(1)) = 𝐺(𝑥, 1) = 𝑓1 (𝑥) ,
and
𝐹(𝑥, 0) = 𝐺(𝑥, 𝜒(0)) = 𝐺(𝑥, 0) = 𝑓0 (𝑥) ,
it gives one a homotopy between 𝑓0 and 𝑓1 . □
We’re now in position to deduce Theorem 5.1.10 from the version of this result that we
proved above.
Proof of Theorem 5.1.10. Let
𝛾𝑡 ∶ 𝑋 × 𝐑 → 𝑋 × 𝐑 , − ∞ < 𝑡 < ∞
be the one-parameter group of diffeomorphisms
𝛾𝑡 (𝑥, 𝑎) = (𝑥, 𝑎 + 𝑡)
and let 𝒗 = 𝜕/𝜕𝑡 be the vector field generating this group. For a 𝑘-form 𝜇 ∈ 𝛺𝑘 (𝑋 × 𝐑), we
have by (5.1.12) the identity
𝛾1⋆ 𝜇 − 𝛾0⋆ 𝜇 = 𝑑𝛤𝜇 + 𝛤𝑑𝜇
where
1
(5.1.15) 𝛤𝜇 = ∫ 𝛾𝑡⋆ (𝜄𝜕/𝜕𝑡 𝜇)𝑑𝑡 .
0
Now let 𝐹, as in Proposition 5.1.13, be a 𝐶∞ map
𝐹∶ 𝑋 × 𝐑 → 𝑌
whose restriction to 𝑋 × [0, 1] is a homotopy between 𝑓0 and 𝑓1 . Then for 𝜔 ∈ 𝛺𝑘 (𝑌)
𝛾1⋆ 𝐹⋆ 𝜔 − 𝛾0⋆ 𝐹⋆ 𝜔 = 𝑑𝛤𝐹⋆ 𝜇 + 𝛤𝐹⋆ 𝑑𝜇
by the identity (5.1.14). Now let 𝜄 ∶ 𝑋 → 𝑋 × 𝐑 be the inclusion 𝑝 ↦ (𝑝, 0), and note that
(𝐹 ∘ 𝛾1 ∘ 𝜄)(𝑝) = 𝐹(𝑝, 1) = 𝑓1 (𝑝)
and
(𝐹 ∘ 𝛾0 ∘ 𝜄)(𝑝) = 𝐹(𝑝, 0) = 𝑓0 (𝑝)
 Draft: March 28, 2018

154 Chapter 5: Cohomology via forms

i.e.,
𝐹 ∘ 𝛾1 ∘ 𝜄 = 𝑓1
and
𝐹 ∘ 𝛾0 ∘ 𝜄 = 𝑓0 .
Thus
𝜄⋆ (𝛾1⋆ 𝐹⋆ 𝜔 − 𝛾0⋆ 𝐹⋆ 𝜔) = 𝑓1⋆ 𝜔 − 𝑓0⋆ 𝜔
and on the other hand by (5.1.15)
𝜄⋆ (𝛾1⋆ 𝐹⋆ 𝜔 − 𝛾0⋆ 𝐹⋆ 𝜔) = 𝑑𝜄⋆ 𝛤𝐹⋆ 𝜔 + 𝜄⋆ 𝛤𝐹⋆ 𝑑𝜔 .
Letting
𝑄 ∶ 𝛺𝑘 (𝑌) → 𝛺𝑘−1 (𝑋)
be the “chain homotopy” operator
(5.1.16) 𝑄𝜔 ≔ 𝜄⋆ 𝛤𝐹⋆ 𝜔
we can write the identity above more succinctly in the form
(5.1.17) 𝑓1⋆ 𝜔 − 𝑓0⋆ 𝜔 = 𝑑𝑄𝜔 + 𝑄𝑑𝜔
and from this deduce, exactly as we did earlier, the identity (5.1.11).
This proof can easily be adapted to the compactly supported setting. Namely the oper-
ator (5.1.16) is defined by the integral
1
(5.1.18) 𝑄𝜔 = ∫ 𝜄⋆ 𝛾𝑡⋆ (𝜄𝜕/𝜕𝑡 𝐹⋆ 𝜔)𝑑𝑡 .
0

Hence if 𝜔 is supported on a set 𝐴 in 𝑌, the integrand of (5.1.17) at 𝑡 is supported on the set

(5.1.19) { 𝑝 ∈ 𝑋 | 𝐹(𝑝, 𝑡) ∈ 𝐴 } ,
and hence 𝑄𝜔 is supported on the set
𝜋(𝐹−1 (𝐴) ∩ 𝑋 × [0, 1])
where 𝜋 ∶ 𝑋 × [0, 1] → 𝑋 is the projection map 𝜋(𝑝, 𝑡) = 𝑝. □

Suppose now that 𝑓0 and 𝑓1 are proper mappings and

𝐺 ∶ 𝑋 × [0, 1] → 𝑌
a proper homotopy between 𝑓0 and 𝑓1 , i.e., a homotopy between 𝑓0 and 𝑓1 which is proper
as a 𝐶∞ map. Then if 𝐹 is the map (5.1.14) its restriction to 𝑋 × [0, 1] is also a proper map,
so this restriction is also a proper homotopy between 𝑓0 and 𝑓1 . Hence if 𝜔 is in 𝛺𝑘𝑐 (𝑌) and
𝐴 is its support, the set (5.1.19) is compact, so 𝑄𝜔 is in 𝛺𝑘−1
𝑐 (𝑋). Therefore all summands
in the “chain homotopy” formula (5.1.17) are compactly supported. Thus we’ve proved:
Theorem 5.1.20. If 𝑓0 , 𝑓1 ∶ 𝑋 → 𝑌 are proper 𝐶∞ maps which are homotopic via a proper
homotopy, the induced maps on cohomology
♯ ♯
𝑓0 , 𝑓1 ∶ 𝐻𝑐𝑘 (𝑌) → 𝐻𝑐𝑘 (𝑋)
are the same.
 Draft: March 28, 2018

§5.1 The de Rham cohomology groups of a manifold 155

We’ll conclude this section by noting that the cohomology groups 𝐻𝑘 (𝑋) are equipped
with a natural product operation. Namely, suppose 𝜔𝑖 ∈ 𝛺𝑘𝑖 (𝑋), 𝑖 = 1, 2, is a closed form
and that 𝑐𝑖 = [𝜔𝑖 ] is the cohomology class represented by 𝜔𝑖 . We can then define a product
cohomology class 𝑐1 ⋅ 𝑐2 in 𝐻𝑘1 +𝑘2 (𝑋) by the recipe
(5.1.21) 𝑐1 ⋅ 𝑐2 ≔ [𝜔1 ∧ 𝜔2 ] .
To show that this is a legitimate definition we first note that since 𝜔2 is closed
𝑑(𝜔1 ∧ 𝜔2 ) = 𝑑𝜔1 ∧ 𝜔2 + (−1)𝑘1 𝜔1 ∧ 𝑑𝜔2 = 0 ,
so 𝜔1 ∧ 𝜔2 is closed and hence does represent a cohomology class. Moreover if we replace
𝜔1 by another representative 𝜔1 + 𝑑𝜇1 = 𝜔′ of the cohomology class 𝑐1 , then
𝜔′1 ∧ 𝜔2 = 𝜔1 ∧ 𝜔2 + 𝑑𝜇1 ∧ 𝜔2 .
But since 𝜔2 is closed,
𝑑𝜇1 ∧ 𝜔2 = 𝑑(𝜇1 ∧ 𝜔2 ) + (−1)𝑘1 𝜇1 ∧ 𝑑𝜔2
= 𝑑(𝜇1 ∧ 𝜔2 )
so
𝜔′1 ∧ 𝜔2 = 𝜔1 ∧ 𝜔2 + 𝑑(𝜇1 ∧ 𝜔2 )
and [𝜔′1∧ 𝜔2 ] = [𝜔1 ∧ 𝜔2 ]. Similarly (5.1.21) is unchanged if we replace 𝜔2 by 𝜔2 + 𝑑𝜇2 ,
so the definition of (5.1.21) depends neither on the choice of 𝜔1 nor 𝜔2 and hence is an
intrinsic definition as claimed.
There is a variant of this product operation for compactly supported cohomology classes,
𝑘
and we’ll leave for you to check that it’s also well defined. Suppose 𝑐1 is in 𝐻𝑐 1 (𝑋) and 𝑐2 is in
𝐻𝑘2 (𝑋) (i.e., 𝑐1 is a compactly supported class and 𝑐2 is an ordinary cohomology class). Let
𝑘
𝜔1 be a representative of 𝑐1 in 𝛺𝑐 1 (𝑋) and 𝜔2 a representative of 𝑐2 in 𝛺𝑘2 (𝑋). Then 𝜔1 ∧ 𝜔2
𝑘1 +𝑘2
is a closed form in 𝛺𝑐 (𝑋) and hence defines a cohomology class
(5.1.22) 𝑐1 ⋅ 𝑐2 ≔ [𝜔1 ∧ 𝜔2 ]
𝑘 +𝑘
in 𝐻𝑐 1 2 (𝑋). We’ll leave for you to check that this is intrinsically defined. We’ll also leave
for you to check that (5.1.22) is intrinsically defined if the roles of 𝑐1 and 𝑐2 are reversed, i.e.,
𝑘
if 𝑐1 is in 𝐻𝑘1 (𝑋) and 𝑐2 in 𝐻𝑐 2 (𝑋) and that the products (5.1.21) and (5.1.22) both satisfy
𝑐1 ⋅ 𝑐2 = (−1)𝑘1 𝑘2 𝑐2 ⋅ 𝑐1 .
Finally we note that if 𝑌 is another manifold and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map then for 𝜔1 ∈ 𝛺𝑘1 (𝑌)
and 𝜔2 ∈ 𝛺𝑘2 (𝑌)
𝑓⋆ (𝜔1 ∧ 𝜔2 ) = 𝑓⋆ 𝜔1 ∧ 𝑓⋆ 𝜔2
by (2.6.8) and hence if 𝜔1 and 𝜔2 are closed and 𝑐𝑖 = [𝜔𝑖 ], then
(5.1.23) 𝑓♯ (𝑐1 ⋅ 𝑐2 ) = 𝑓♯ 𝑐1 ⋅ 𝑓♯ 𝑐2 .
Exercises for §5.1
Exercise 5.1.i (stereographic projection). Let 𝑝 ∈ 𝑆𝑛 be the point 𝑝 = (0, …, 0, 1). Show
that for every point 𝑥 = (𝑥1 , …, 𝑥𝑛+1 ) of 𝑆𝑛 ∖ {𝑝} the ray
𝑡𝑥 + (1 − 𝑡)𝑝 , 𝑡 > 0
intersects the plane 𝑥𝑛+1 = 0 in the point
1
𝛾(𝑥) = (𝑥 , …, 𝑥𝑛 )
1 − 𝑥𝑛+1 1
 Draft: March 28, 2018

156 Chapter 5: Cohomology via forms

and that the map 𝛾 ∶ 𝑆𝑛 ∖ {𝑝} → 𝐑𝑛 is a diffeomorphism.

Exercise 5.1.ii. Show that the operator
𝑄𝑡 ∶ 𝛺𝑘 (𝑌) → 𝛺𝑘−1 (𝑋)
in the integrand of equation (5.1.18), i.e., the operator
𝑄𝑡 𝜔 = 𝜄⋆ 𝛾𝑡⋆ (𝜄𝜕/𝜕𝑡 )𝐹⋆ 𝜔
has the following description. Let 𝑝 be a point of 𝑋 and let 𝑞 = 𝑓𝑡 (𝑝). The curve 𝑠 ↦ 𝑓𝑠 (𝑝)
passes through 𝑞 at time 𝑠 = 𝑡. Let 𝑣(𝑞) ∈ 𝑇𝑞 𝑌 be the tangent vector to this curve at 𝑡. Show
that
(5.1.24) (𝑄𝑡 𝜔)(𝑝) = (𝑑𝑓𝑡⋆ )𝑝 𝜄𝑣(𝑞) 𝜔𝑞 .
Exercise 5.1.iii. Let 𝑈 be a star-shaped open subset of 𝐑𝑛 , i.e., a subset of 𝐑𝑛 with the prop-
erty that for every 𝑝 ∈ 𝑈 the ray { 𝑡𝑝 | 0 ≤ 𝑡 ≤ 1 } is in 𝑈.
(1) Let 𝒗 be the vector field
𝑛
𝜕
𝒗 = ∑ 𝑥𝑖
𝑖=1 𝜕𝑥𝑖
and 𝛾𝑡 ∶ 𝑈 → 𝑈 the map 𝑝 ↦ 𝑡𝑝. Show that for every 𝑘-form 𝜔 ∈ 𝛺𝑘 (𝑈)
𝜔 = 𝑑𝑄𝜔 + 𝑄𝑑𝜔
where
1
𝑑𝑡
𝑄𝜔 = ∫ 𝛾𝑡⋆ 𝜄𝒗 𝜔 .
0 𝑡
(2) Show that if 𝜔 = ∑𝐼 𝑎𝐼 (𝑥)𝑑𝑥𝐼 then

(5.1.25) 𝑄𝜔 = ∑ (∫ 𝑡𝑘−1 (−1)𝑟−1 𝑥𝑖𝑟 𝑎𝐼 (𝑡𝑥)𝑑𝑡) 𝑑𝑥𝐼𝑟

𝐼,𝑟

̂ 𝑖 ∧ ⋯ 𝑑𝑥𝑖
where 𝑑𝑥𝐼𝑟 ≔ 𝑑𝑥𝑖1 ∧ ⋯ ∧ 𝑑𝑥 𝑟 𝑘

Exercise 5.1.iv. Let 𝑋 and 𝑌 be oriented connected 𝑛-manifolds, and 𝑓 ∶ 𝑋 → 𝑌 a proper

map. Show that the linear map 𝐿 defined by the commutative square
𝑓♯
𝐻𝑐𝑛 (𝑌) 𝐻𝑐𝑛 (𝑋)
𝐼𝑌 ≀ ≀ 𝐼𝑋

𝐑 𝐿
𝐑

is multiplication by deg(𝑓).
Exercise 5.1.v. A homotopy equivalence between 𝑋 and 𝑌 is a pair of maps 𝑓 ∶ 𝑋 → 𝑌 and
𝑔 ∶ 𝑌 → 𝑋 such that 𝑔 ∘ 𝑓 ≃ id𝑋 and 𝑓 ∘ 𝑔 ≃ id𝑌 . Show that if 𝑋 and 𝑌 are homotopy
equivalent their cohomology groups are the same “up to isomorphism”, i.e., 𝑓 and 𝑔 induce
inverse isomorphisms 𝑓♯ ∶ 𝐻𝑘 (𝑌) → 𝐻𝑘 (𝑋) and 𝑔♯ ∶ 𝐻𝑘 (𝑋) → 𝐻𝑘 (𝑌).
Exercise 5.1.vi. Show that 𝐑𝑛 ∖ {0} and 𝑆𝑛−1 are homotopy equivalent.
Exercise 5.1.vii. What are the cohomology groups of the 𝑛-sphere with two points deleted?
Hint: The 𝑛-sphere with one point deleted is homeomorphic to 𝐑𝑛 .
 Draft: March 28, 2018

§5.1 The de Rham cohomology groups of a manifold 157

Exercise 5.1.viii. Let 𝑋 and 𝑌 be manifolds and 𝑓0 , 𝑓1 , 𝑓2 ∶ 𝑋 → 𝑌 three 𝐶∞ maps. Show

that if 𝑓0 and 𝑓1 are homotopic and 𝑓1 and 𝑓2 are homotopic then 𝑓0 and 𝑓2 are homotopic.
Hint: The homotopy (5.1.7) has the property that
𝐹(𝑝, 𝑡) = 𝑓𝑡 (𝑝) = 𝑓0 (𝑝)
1
for 0 ≤ 𝑡 ≤ 4
and
𝐹(𝑝, 𝑡) = 𝑓𝑡 (𝑝) = 𝑓1 (𝑝)
3
for ≤ 𝑡 < 1. Show that two homotopies with these properties: a homotopy between 𝑓0
4
and 𝑓1 and a homotopy between 𝑓1 and 𝑓2 , are easy to “glue together” to get a homotopy
between 𝑓0 and 𝑓2 .
Exercise 5.1.ix.
(1) Let 𝑋 be an 𝑛-manifold. Given points 𝑝0 , 𝑝1 , 𝑝2 ∈ 𝑋, show that if 𝑝0 can be joined to 𝑝1
by a 𝐶∞ curve 𝛾0 ∶ [0, 1] → 𝑋, and 𝑝1 can be joined to 𝑝2 by a 𝐶∞ curve 𝛾1 ∶ [0, 1] → 𝑋,
then 𝑝0 can be joined to 𝑝2 by a 𝐶∞ curve, 𝛾 ∶ [0, 1] → 𝑋.
Hint: A 𝐶∞ curve, 𝛾 ∶ [0, 1] → 𝑋, joining 𝑝0 to 𝑝2 can be thought of as a homotopy
between the maps
𝛾𝑝0 ∶ ∗ → 𝑋 , ∗ ↦ 𝑝0
and
𝛾𝑝1 ∶ ∗ → 𝑋 , ∗ ↦ 𝑝1
where ∗ is the zero-dimensional manifold consisting of a single point.
(2) Show that if a manifold 𝑋 is connected it is arc-wise connected: any two points can by
joined by a 𝐶∞ curve.
Exercise 5.1.x. Let 𝑋 be a connected 𝑛-manifold and 𝜔 ∈ 𝛺1 (𝑋) a closed one-form.
(1) Show that if 𝛾 ∶ [0, 1] → 𝑋 is a 𝐶∞ curve there exists a partition:
0 = 𝑎 0 < 𝑎 1 < ⋯ < 𝑎𝑛 = 1
of the interval [0, 1] and open sets 𝑈1 , …, 𝑈𝑛 in 𝑋 such that 𝛾 ([𝑎𝑖−1 , 𝑎𝑖 ]) ⊂ 𝑈𝑖 and such
that 𝜔|𝑈𝑖 is exact.
(2) In part (1) show that there exist functions 𝑓𝑖 ∈ 𝐶∞ (𝑈𝑖 ) such that 𝜔|𝑈𝑖 = 𝑑𝑓𝑖 and
𝑓𝑖 (𝛾(𝑎𝑖 )) = 𝑓𝑖+1 (𝛾(𝑎𝑖 )).
(3) Show that if 𝑝0 and 𝑝1 are the end points of 𝛾
1
𝑓𝑛 (𝑝1 ) − 𝑓1 (𝑝0 ) = ∫ 𝛾⋆ 𝜔 .
0
(4) Let
(5.1.26) 𝛾𝑠 ∶ [0, 1] → 𝑋 , 0 ≤ 𝑠 ≤ 1
be a homotopic family of curves with 𝛾𝑠 (0) = 𝑝0 and 𝛾𝑠 (1) = 𝑝1 . Prove that the integral
1
∫ 𝛾𝑠⋆ 𝜔
0
is independent of 𝑠0 .
Hint: Let 𝑠0 be a point on the interval, [0, 1]. For 𝛾 = 𝛾𝑠0 choose 𝑎𝑖 ’s and 𝑓𝑖 ’s as in
parts (a)–(b) and show that for 𝑠 close to 𝑠0 , 𝛾𝑠 [𝑎𝑖−1 , 𝑎𝑖 ] ⊂ 𝑈𝑖 .
(5) A connected manifold 𝑋 is simply connected if for any two curves 𝛾0 , 𝛾1 ∶ [0, 1] → 𝑋
with the same end-points 𝑝0 and 𝑝1 , there exists a homotopy (5.1.22) with 𝛾𝑠 (0) = 𝑝0
and 𝛾𝑠 (1) = 𝑝1 , i.e., 𝛾0 can be smoothly deformed into 𝛾1 by a family of curves all having
the same end-points. Prove the following theorem.
 Draft: March 28, 2018

158 Chapter 5: Cohomology via forms

Theorem 5.1.27. If 𝑋 is simply-connected, then 𝐻1 (𝑋) = 0.

Exercise 5.1.xi. Show that the product operation (5.1.21) is associative and satisfies left and
right distributive laws.
Exercise 5.1.xii. Let 𝑋 be a compact oriented 2𝑛-dimensional manifold. Show that the map
𝐵 ∶ 𝐻𝑛 (𝑋) × 𝐻𝑛 (𝑋) → 𝐑
defined by
𝐵(𝑐1 , 𝑐2 ) = 𝐼𝑋 (𝑐1 ⋅ 𝑐2 )
𝑛
is a bilinear form on 𝐻 (𝑋) and that 𝐵 is symmetric if 𝑛 is even and alternating if 𝑛 is odd.

5.2. The Mayer–Vietoris Sequence

In this section we’ll develop some techniques for computing cohomology groups of
manifolds. (These techniques are known collectively as “diagram chasing” and the mastering
of these techniques is more akin to becoming proficient in checkers or chess or the Sunday
acrostics in the New York Times than in the areas of mathematics to which they’re applied.)
Let 𝐶0 , 𝐶1 , 𝐶2 , … be vector spaces and 𝑑 ∶ 𝐶𝑖 → 𝐶𝑖+1 a linear map. The sequence of vector
spaces and maps
𝑑 𝑑 𝑑
(5.2.1) 𝐶0 𝐶1 𝐶2 ⋯.
is called a cochain complex, or simply a complex, if 𝑑2 = 0, i.e., if for 𝑎 ∈ 𝐶𝑘 , 𝑑(𝑑𝑎) = 0. For
instance if 𝑋 is a manifold the de Rham complex
𝑑 𝑑 𝑑
(5.2.2) 𝛺0 (𝑋) 𝛺1 (𝑋) 𝛺2 (𝑋) ⋯.
is an example of a complex, and the complex of compactly supported de Rham forms
𝑑 𝑑 𝑑
(5.2.3) 𝛺0𝑐 (𝑋) 𝛺1𝑐 (𝑋) 𝛺2𝑐 (𝑋) ⋯.
is another example. One defines the cohomology groups of the complex (5.2.1) in exactly
the same way that we defined the cohomology groups of the complexes (5.2.2) and (5.2.3)
in §5.1. Let
𝑍𝑘 ≔ ker(𝑑 ∶ 𝐶𝑘 → 𝐶𝑘+1 ) = { 𝑎 ∈ 𝐶𝑘 | 𝑑𝑎 = 0 }
and
𝐵𝑘 ≔ 𝑑𝐶𝑘−1
i.e., let 𝑎 be in 𝐵𝑘 if and only if 𝑎 = 𝑑𝑏 for some 𝑏 ∈ 𝐶𝑘−1 . Then 𝑑𝑎 = 𝑑2 𝑏 = 0, so 𝐵𝑘 is a
vector subspace of 𝑍𝑘 , and we define 𝐻𝑘 (𝐶) — the 𝑘th cohomology group of the complex
(5.2.1) — to be the quotient space
(5.2.4) 𝐻𝑘 (𝐶) ≔ 𝑍𝑘 /𝐵𝑘 .
Given 𝑐 ∈ 𝑍𝑘 we will, as in §5.1, denote its image in 𝐻𝑘 (𝐶) by [𝑐] and we’ll call 𝑐 a represen-
tative of the cohomology class [𝑐].
We will next assemble a small dictionary of “diagram chasing” terms.
Definition 5.2.5. Let 𝑉0 , 𝑉1 , 𝑉2 , … be vector spaces and 𝛼𝑖 ∶ 𝑉𝑖 → 𝑉𝑖+1 linear maps. The
sequence
𝛼0 𝛼1 𝛼2
𝑉0 𝑉1 𝑉2 ⋯.
is an exact sequence if, for each 𝑖, the kernel of 𝛼𝑖+1 is equal to the image of 𝛼𝑖 .
 Draft: March 28, 2018

§5.2 The Mayer–Vietoris Sequence 159

For example the sequence (5.2.1) is exact if 𝑍𝑖 = 𝐵𝑖 for all 𝑖, or, in other words, if
𝐻𝑖 (𝐶) = 0 for all 𝑖. A simple example of an exact sequence that we’ll encounter a lot below
is a sequence of the form
𝛼1 𝛼2
0 𝑉1 𝑉2 𝑉3 0,

i.e., a five term exact sequence whose first and last terms are the vector space, 𝑉0 = 𝑉4 = 0,
and hence 𝛼0 = 𝛼3 = 0. This sequence is exact if and only if the following conditions hold:
(1) 𝛼1 is injective,
(2) the kernel of 𝛼2 equals the image of 𝛼1 , and
(3) 𝛼2 is surjective.
We will call an exact sequence of this form a short exact sequence. (We’ll also encounter
a lot below an even shorter example of an exact sequence, namely a sequence of the form

𝛼1
0 𝑉1 𝑉2 0.

This is an exact sequence if and only if 𝛼1 is bijective.)

Another basic notion in the theory of diagram chasing is the notion of a commutative
diagram. The square diagram of vector spaces and linear maps

𝑓
𝐴 𝐵
𝑖 𝑗

𝐶 𝑔 𝐷

commutes if 𝑗 ∘ 𝑓 = 𝑔 ∘ 𝑖, and a more complicated diagram of vector spaces and linear maps
like the diagram

𝐴1 𝐴2 𝐴3

𝐵1 𝐵2 𝐵3

𝐶1 𝐶2 𝐶3

commutes if every sub-square in the diagram commutes.

We now have enough “diagram chasing” vocabulary to formulate the Mayer–Vietoris
theorem. For 𝑟 = 1, 2, 3 let

𝑑 𝑑 𝑑
(5.2.6) 𝐶𝑟0 𝐶𝑟1 𝐶𝑟2 ⋯.

be a complex and, for fixed 𝑘, let

𝑖 𝑗
(5.2.7) 0 𝐶1𝑘 𝐶2𝑘 𝐶3𝑘 0
 Draft: March 28, 2018

160 Chapter 5: Cohomology via forms

be a short exact sequence. Assume that the diagram below commutes:

⋮ ⋮ ⋮

𝑖 𝑗
0 𝐶1𝑘−1 𝐶2𝑘−1 𝐶3𝑘−1 0

𝑑 𝑑 𝑑

0 𝐶1𝑘 𝑖 𝐶2𝑘 𝑗 𝐶3𝑘 0

𝑑 𝑑 𝑑

0 𝐶1𝑘+1 𝑖
𝐶2𝑘+1 𝑗
𝐶3𝑘+1 0

⋮ ⋮ ⋮ .

That is, assume that in the left hand squares, 𝑑𝑖 = 𝑖𝑑, and in the right hand squares, 𝑑𝑗 = 𝑗𝑑.
The Mayer–Vietoris theorem addresses the following question: If one has information
about the cohomology groups of two of the three complexes (5.2.6), what information about
the cohomology groups of the third can be extracted from this diagram? Let us first observe
that the maps 𝑖 and 𝑗 give rise to mappings between these cohomology groups. Namely, for
𝑟 = 1, 2, 3 let 𝑍𝑟𝑘 be the kernel of the map 𝑑 ∶ 𝐶𝑟𝑘 → 𝐶𝑟𝑘+1 , and 𝐵𝑟𝑘 the image of the map
𝑑 ∶ 𝐶𝑟𝑘−1 → 𝐶𝑟𝑘 . Since 𝑖𝑑 = 𝑑𝑖, 𝑖 maps 𝐵1𝑘 into 𝐵2𝑘 and 𝑍1𝑘 into 𝑍2𝑘 , therefore by (5.2.4) it gives
rise to a linear mapping
𝑖♯ ∶ 𝐻𝑘 (𝐶1 ) → 𝐻𝑘 (𝐶2 ) .
Similarly since 𝑗𝑑 = 𝑑𝑗, 𝑗 maps 𝐵2𝑘 into 𝐵3𝑘 and 𝑍2𝑘 into 𝑍3𝑘 , and so by (5.2.4) gives rise to a
linear mapping
𝑗♯ ∶ 𝐻𝑘 (𝐶2 ) → 𝐻𝑘 (𝐶3 ) .
Moreover, since 𝑗 ∘ 𝑖 = 0 the image of 𝑖♯ is contained in the kernel of 𝑗♯ . We’ll leave as an
exercise the following sharpened version of this observation:
Proposition 5.2.8. The kernel of 𝑗♯ equals the image of 𝑖♯ , i.e., the three term sequence
𝑖♯ 𝑗♯
(5.2.9) 𝐻𝑘 (𝐶1 ) 𝐻𝑘 (𝐶2 ) 𝐻𝑘 (𝐶3 )
is exact.
Since (5.2.7) is a short exact sequence one is tempted to conjecture that (5.2.9) is also
a short exact sequence (which, if it were true, would tell us that the cohomology groups
of any two of the complexes (5.2.6) completely determine the cohomology groups of the
third). Unfortunately, this is not the case. To see how this conjecture can be violated Let us
try to show that the mapping 𝑗♯ is surjective. Let 𝑐3𝑘 be an element of 𝑍3𝑘 representing the
cohomology class [𝑐3𝑘 ] in 𝐻3 (𝐶3 ). Since (5.2.7) is exact there exists a 𝑐2𝑘 in 𝐶2𝑘 which gets
mapped by 𝑗 onto 𝑐3𝑘 , and if 𝑐3𝑘 were in 𝑍2𝑘 this would imply
𝑗♯ [𝑐2𝑘 ] = [𝑗𝑐2𝑘 ] = [𝑐3𝑘 ] ,
 Draft: March 28, 2018

§5.2 The Mayer–Vietoris Sequence 161

i.e., the cohomology class [𝑐3𝑘 ] would be in the image of 𝑗♯ . However, since there’s no reason
for 𝑐2𝑘 to be in 𝑍2𝑘 , there’s also no reason for [𝑐3𝑘 ] to be in the image of 𝑗♯ . What we can say,
however, is that 𝑗𝑑𝑐2𝑘 = 𝑑𝑗𝑐2𝑘 = 𝑑𝑐3𝑘 = 0 since 𝑐3𝑘 is in 𝑍3𝑘 . Therefore by the exactness of
(5.2.7) in degree 𝑘 + 1 there exists a unique element 𝑐1𝑘+1 in 𝐶1𝑘+1 with property
(5.2.10) 𝑑𝑐2𝑘 = 𝑖𝑐1𝑘+1 .
Moreover, since
0 = 𝑑(𝑑𝑐2𝑘 ) = 𝑑𝑖(𝑐1𝑘+1 ) = 𝑖 𝑑𝑐1𝑘+1
and 𝑖 is injective, 𝑑𝑐1𝑘+1 = 0, i.e.,
(5.2.11) 𝑐1𝑘+1 ∈ 𝑍1𝑘+1 .
Thus via (5.2.10) and (5.2.11) we’ve converted an element 𝑐3𝑘 of 𝑍3𝑘 into an element 𝑐1𝑘+1 of
𝑍1𝑘+1 and hence set up a correspondence
(5.2.12) 𝑐3𝑘 ∈ 𝑍3𝑘 → 𝑐1𝑘+1 ∈ 𝑍1𝑘+1 .
Unfortunately this correspondence isn’t, strictly speaking, a map of 𝑍3𝑘 into 𝑍1𝑘+1 ; the 𝑐1𝑘 in
(5.2.12) isn’t determined by 𝑐3𝑘 alone but also by the choice we made of 𝑐2𝑘 . Suppose, how-
ever, that we make another choice of a 𝑐2𝑘 with the property 𝑗(𝑐2𝑘 ) = 𝑐3𝑘 . Then the difference
between our two choices is in the kernel of 𝑗 and hence, by the exactness of (5.2.6) at level
𝑘, is in the image of 𝑖. In other words, our two choices are related by
(𝑐2𝑘 )new = (𝑐2𝑘 )old + 𝑖(𝑐1𝑘 )
for some 𝑐1𝑘 in 𝐶1𝑘 , and hence by (5.2.10)
(𝑐1𝑘+1 )new = (𝑐1𝑘+1 )old + 𝑑𝑐1𝑘 .
Therefore, even though the correspondence (5.2.12) isn’t strictly speaking a map it does give
rise to a well-defined map
𝑍3𝑘 → 𝐻𝑘+1 (𝐶1 ) , 𝑐3𝑘 ↦ [𝑐3𝑘+1 ] .
Moreover, if 𝑐3𝑘 is in 𝐵3𝑘 , i.e., 𝑐3𝑘 = 𝑑𝑐3𝑘−1 for some 𝑐3𝑘−1 ∈ 𝐶3𝑘−1 , then by the exactness of (5.2.6)
at level 𝑘−1, 𝑐3𝑘−1 = 𝑗(𝑐2𝑘−1 ) for some 𝑐2𝑘−1 ∈ 𝐶2𝑘−1 and hence 𝑐3𝑘 = 𝑗(𝑑𝑐2𝑘−2 ). In other words we
can take the 𝑐2𝑘 above to be 𝑑𝑐2𝑘−1 in which case the 𝑐1𝑘+1 in equation (5.2.10) is just zero. Thus
the map (5.2.12) maps 𝐵3𝑘 to zero and hence by Proposition 1.2.9 gives rise to a well-defined
map
𝛿 ∶ 𝐻𝑘 (𝐶3 ) → 𝐻𝑘+1 (𝐶1 )
mapping [𝑐3𝑘 ] → [𝑐1𝑘+1 ]. We will leave it as an exercise to show that this mapping measures
the failure of the arrow 𝑗♯ in the exact sequence (5.2.9) to be surjective (and hence the failure
of this sequence to be a short exact sequence at its right end).
Proposition 5.2.13. The image of the map 𝑗♯ ∶ 𝐻𝑘 (𝐶2 ) → 𝐻𝑘 (𝐶3 ) is equal to the kernel of
the map 𝛿 ∶ 𝐻𝑘 (𝐶3 ) → 𝐻𝑘+1 (𝐶1 ).
Hint: Suppose that in the correspondence (5.2.12), 𝑐1𝑘+1 is in 𝐵1𝑘+1 . Then 𝑐1𝑘+1 = 𝑑𝑐1𝑘 for
some 𝑐1𝑘 in 𝐶1𝑘 . Show that 𝑗(𝑐2𝑘 − 𝑖(𝑐1𝑘 )) = 𝑐3𝑘 and 𝑑(𝑐2𝑘 − 𝑖(𝑐1𝑘 )) = 0, i.e., 𝑐2𝑘 − 𝑖(𝑐1𝑘 ) is in 𝑍2𝑘 and
hence 𝑗♯ [𝑐2𝑘 − 𝑖(𝑐1𝑘 )] = [𝑐3𝑘 ].
Let us next explore the failure of the map 𝑖♯ ∶ 𝐻𝑘+1 (𝐶1 ) → 𝐻𝑘+1 (𝐶2 ), to be injective. Let
𝑐1𝑘+1 be in 𝑍1𝑘+1 and suppose that its cohomology class, [𝑐1𝑘+1 ], gets mapped by 𝑖♯ into zero.
This translates into the statement
(5.2.14) 𝑖(𝑐1𝑘+1 ) = 𝑑𝑐2𝑘
 Draft: March 28, 2018

162 Chapter 5: Cohomology via forms

for some 𝑐2𝑘 ∈ 𝐶2𝑘 . Moreover since 𝑑𝑐2𝑘 = 𝑖(𝑐1𝑘+1 ), 𝑗(𝑑𝑐2𝑘 ) = 0. But if
(5.2.15) 𝑐3𝑘 ≔ 𝑗(𝑐2𝑘 )
then 𝑑𝑐3𝑘 = 𝑑𝑗(𝑐2𝑘 ) = 𝑗(𝑑𝑐2𝑘 ) = 𝑗(𝑖(𝑐1𝑘+1 )) = 0, so 𝑐3𝑘 is in 𝑍3𝑘 , and by (5.2.14), (5.2.15) and the
definition of 𝛿
[𝑐1𝑘+1 ] = 𝛿[𝑐3𝑘 ] .
In other words the kernel of the map 𝑖♯ ∶ 𝐻𝑘+1 (𝐶1 ) → 𝐻𝑘+1 (𝐶2 ) is contained in the image of
the map 𝛿 ∶ 𝐻𝑘 (𝐶3 ) → 𝐻𝑘+1 (𝐶1 ). We will leave it as an exercise to show that this argument
can be reversed to prove the converse assertion and hence to prove
Proposition 5.2.16. The image of the map 𝛿 ∶ 𝐻𝑘 (𝐶1 ) → 𝐻𝑘+1 (𝐶1 ) is equal to the kernel of
the map 𝑖♯ ∶ 𝐻𝑘+1 (𝐶1 ) → 𝐻𝑘+1 (𝐶2 ).
Putting together the Propositions 5.2.8, 5.2.13 and 5.2.16 we obtain the main result of
this section: the Mayer–Vietoris theorem. The sequence of cohomology groups and linear
maps
𝛿 𝑖♯ 𝑗♯ 𝛿 𝑖♯
(5.2.17) ⋯ 𝐻𝑘 (𝐶1 ) 𝐻𝑘 (𝐶2 ) 𝐻𝑘 (𝐶3 ) 𝐻𝑘+1 (𝐶1 ) ⋯
is exact.
Remark 5.2.18. To see an illuminating real world example of an application of these ideas,
we strongly recommend that our readers stare at Exercise 5.2.v with gives a method for com-
puting the de Rham cohomology of the 𝑛-sphere 𝑆𝑛 in terms of the de Rham cohomologies
of 𝑆𝑛 ∖ {(0, …, 0, 1)} and 𝑆𝑛 ∖ {(0, …, 0, −1)} via an exact sequence of the form (5.2.17).
Remark 5.2.19. In view of the “⋯”’s this sequence can be a very long sequence and is com-
monly referred to as the long exact sequence in cohomology associated to the short exact
sequence of complexes (5.2.7).
Before we discuss the applications of this result, we will introduce some vector space
notation. Given vector spaces 𝑉1 and 𝑉2 , we’ll denote by 𝑉1 ⊕ 𝑉2 the vector space sum of 𝑉1
and 𝑉2 , i.e., the set of all pairs
(𝑢1 , 𝑢2 ) , 𝑢𝑖 ∈ 𝑉𝑖
with the addition operation
(𝑢1 , 𝑢2 ) + (𝑣1 + 𝑣2 ) ≔ (𝑢1 + 𝑣1 , 𝑢2 + 𝑣2 )
and the scalar multiplication operation
𝜆(𝑢1 , 𝑢2 ) ≔ (𝜆𝑢1 , 𝜆𝑢2 ) .
Now let 𝑋 be a manifold and let 𝑈1 and 𝑈2 be open subsets of 𝑋. Then one has a linear map
𝑖 ∶ 𝛺𝑘 (𝑈1 ∪ 𝑈2 ) → 𝛺𝑘 (𝑈1 ) ⊕ 𝛺𝑘 (𝑈2 )
defined by
(5.2.20) 𝜔 ↦ (𝜔|𝑈1 , 𝜔|𝑈2 )
where 𝜔|𝑈𝑖 is the restriction of 𝜔 to 𝑈𝑖 . Similarly one has a linear map
𝑗 ∶ 𝛺𝑘 (𝑈1 ) ⊕ 𝛺𝑘 (𝑈2 ) → 𝛺𝑘 (𝑈1 ∩ 𝑈2 )
defined by
(𝜔1 , 𝜔2 ) ↦ 𝜔1 |𝑈1 ∩𝑈2 − 𝜔2 |𝑈1 ∩𝑈2 .
We claim:
 Draft: March 28, 2018

§5.2 The Mayer–Vietoris Sequence 163

Theorem 5.2.21. The sequence

𝑖 𝑗
0 𝛺𝑘 (𝑈1 ∩ 𝑈2 ) 𝛺𝑘 (𝑈1 ) ⊕ 𝛺𝑘 (𝑈2 ) 𝛺𝑘 (𝑈1 ∩ 𝑈2 ) 0.

is a short exact sequence.

Proof. If the right hand side of (5.2.20) is zero, 𝜔 itself has to be zero so the map (5.2.20) is
injective. Moreover, if the right hand side of (5.2) is zero, 𝜔1 and 𝜔2 are equal on the overlap,
𝑈1 ∩ 𝑈2 , so we can glue them together to get a 𝐶∞ 𝑘-form on 𝑈1 ∪ 𝑈2 by setting 𝜔 = 𝜔1 on
𝑈1 and 𝜔 = 𝜔2 on 𝑈2 . Thus by (5.2.20) 𝑖(𝜔) = (𝜔1 , 𝜔2 ), and this shows that the kernel of 𝑗 is
equal to the image of 𝑖. Hence to complete the proof we only have to show that 𝑗 is surjective,
i.e., that every form 𝜔 on 𝛺𝑘 (𝑈1 ∩ 𝑈2 ) can be written as a difference, 𝜔1 |𝑈1 ∩ 𝑈2 − 𝜔2 |𝑈1 ∩ 𝑈2 ,
where 𝜔1 is in 𝛺𝑘 (𝑈1 ) and 𝜔2 in in 𝛺𝑘 (𝑈2 ). To prove this we’ll need the following variant of
the partition of unity theorem.
Theorem 5.2.22. There exist functions, 𝜙𝛼 ∈ 𝐶∞ (𝑈1 ∪ 𝑈2 ), 𝛼 = 1, 2, such that support 𝜙𝛼 is
contained in 𝑈𝛼 and 𝜙1 + 𝜙2 = 1.
Before proving this Let us use it to complete our proof of Theorem 5.2.21. Given 𝜔 ∈
𝛺𝑘 (𝑈1 ∩ 𝑈2 ) let

𝜙2 𝜔, on 𝑈1 ∩ 𝑈2
(5.2.23) 𝜔1 ≔ {
0, on 𝑈1 ∖ 𝑈1 ∩ 𝑈2
and let
−𝜙1 𝜔, on 𝑈1 ∩ 𝑈2
(5.2.24) 𝜔2 ≔ {
0, on 𝑈2 ∖ 𝑈1 ∩ 𝑈2 .
Since 𝜙2 is supported on 𝑈2 the form defined by (5.2.23) is 𝐶∞ on 𝑈1 and since 𝜙1 is sup-
ported on 𝑈1 the form defined by (5.2.24) is 𝐶∞ on 𝑈2 and since 𝜙1 + 𝜙2 = 1, 𝜔1 − 𝜔2 =
(𝜙1 + 𝜙2 )𝜔 = 𝜔 on 𝑈1 ∩ 𝑈2 . □

Proof of Theorem 5.2.22. Let 𝜌𝑖 ∈ 𝐶0∞ (𝑈1 ∪ 𝑈2 ), 𝑖 = 1, 2, 3, … be a partition of unity subordi-

nate to the cover {𝑈1 , 𝑈2 } of 𝑈1 ∪ 𝑈2 and let 𝜙1 be the sum of the 𝜌𝑖 ’s with support on 𝑈1 and
𝜙2 the sum of the remaining 𝜌𝑖 ’s. It’s easy to check (using part (b) of Theorem 4.6.1) that 𝜙𝛼
is supported in 𝑈𝛼 and (using part (a) of Theorem 4.6.1) that 𝜙1 + 𝜙2 = 1. □

Now let
𝑑 𝑑 𝑑
(5.2.25) 0 𝐶10 𝐶11 𝐶12 ⋯.

be the de Rham complex of 𝑈1 ∪ 𝑈2 , let

𝑑 𝑑 𝑑
(5.2.26) 0 𝐶30 𝐶31 𝐶32 ⋯.

be the de Rham complex of 𝑈1 ∩ 𝑈2 , and let

𝑑 𝑑 𝑑
(5.2.27) 0 𝐶20 𝐶21 𝐶22 ⋯.

be the vector space direct sum of the de Rham complexes of 𝑈1 and 𝑈2 , i.e., the complex
whose 𝑘th term is
𝐶2𝑘 = 𝛺𝑘 (𝑈1 ) ⊕ 𝛺𝑘 (𝑈2 )
 Draft: March 28, 2018

164 Chapter 5: Cohomology via forms

with 𝑑 ∶ 𝐶2𝑘 → 𝐶2𝑘+1 defined to be the map 𝑑(𝜇1 , 𝜇2 ) = (𝑑𝜇1 , 𝑑𝜇2 ). Since 𝐶1𝑘 = 𝛺𝑘 (𝑈1 ∪ 𝑈2 )
and 𝐶3𝑘 = 𝛺𝑘 (𝑈1 ∩ 𝑈2 ) we have, by Theorem 5.2.21, a short exact sequence
𝑖 𝑗
0 𝐶1𝑘 𝐶2𝑘 𝐶3𝑘 0,
and it’s easy to see that 𝑖 and 𝑗 commute with the 𝑑’s:
𝑑𝑖 = 𝑖𝑑 and 𝑑𝑗 = 𝑗𝑑 .
Hence we’re exactly in the situation to which Mayer–Vietoris applies. Since the cohomology
groups of the complexes (5.2.25) and (5.2.26) are the de Rham cohomology groups 𝐻𝑘 (𝑈1 ∪
𝑈2 ) and 𝐻𝑘 (𝑈1 ∩ 𝑈2 ), and the cohomology groups of the complex (5.2.27) are the vector
space direct sums, 𝐻𝑘 (𝑈1 ) ⊕ 𝐻𝑘 (𝑈2 ), we obtain from the abstract Mayer–Vietoris theorem,
the following de Rham theoretic version of Mayer–Vietoris.
Theorem 5.2.28. Letting 𝑈 = 𝑈1 ∪ 𝑈2 and 𝑉 = 𝑈1 ∩ 𝑈2 one has a long exact sequence in de
Rham cohomology:
𝛿 𝑖♯ 𝑗♯ 𝛿 𝑖♯
⋯ 𝐻𝑘 (𝑈) 𝐻𝑐𝑘 (𝑈1 ) ⊕ 𝐻𝑐𝑘 (𝑈2 ) 𝐻𝑐𝑘 (𝑉) 𝐻𝑐𝑘+1 (𝑈) ⋯.

This result also has an analogue for compactly supported de Rham cohomology. Let
(5.2.29) 𝑖 ∶ 𝛺𝑘𝑐 (𝑈1 ∩ 𝑈2 ) → 𝐻𝑐𝑘 (𝑈1 ) ⊕ 𝛺𝑘𝑐 (𝑈2 )
be the map
𝑖(𝜔) ≔ (𝜔1 , 𝜔2 )
where
𝜔, on 𝑈1 ∩ 𝑈2
(5.2.30) 𝜔𝑖 ≔ {
0, on 𝑈𝑖 ∖ 𝑈1 ∩ 𝑈2 .
(Since 𝜔 is compactly supported on 𝑈1 ∩ 𝑈2 the form defined by equation (5.2.29) is a 𝐶∞
form and is compactly supported on 𝑈𝑖 .) Similarly, let
𝑗 ∶ 𝛺𝑘𝑐 (𝑈1 ) ⊕ 𝛺𝑘𝑐 (𝑈2 ) → 𝛺𝑘𝑐 (𝑈1 ∪ 𝑈2 )
be the map
𝑗(𝜔1 , 𝜔2 ) ≔ 𝜔̃1 − 𝜔̃2
where:
𝜔𝑖 , on 𝑈𝑖
𝜔̃𝑖 ≔ {
0, on (𝑈1 ∪ 𝑈2 ) − 𝑈𝑖 .
As above it’s easy to see that 𝑖 is injective and that the kernel of 𝑗 is equal to the image of 𝑖.
Thus if we can prove that 𝑗 is surjective we’ll have proved
Theorem 5.2.31. The sequence
𝑖 𝑗
0 𝛺𝑘𝑐 (𝑈1 ∩ 𝑈2 ) 𝛺𝑘𝑐 (𝑈1 ) ⊕ 𝛺𝑘𝑐 (𝑈2 ) 𝛺𝑘𝑐 (𝑈1 ∩ 𝑈2 ) 0.
is a short exact sequence.
Proof. To prove the surjectivity of 𝑗 we mimic the proof above. Given 𝜔 ∈ 𝛺𝑘𝑐 (𝑈1 ∪ 𝑈2 ), let
𝜔1 ≔ 𝜙1 𝜔|𝑈1
and
𝜔2 ≔ −𝜙2 𝜔|𝑈2 .
Then by (5.2.30) we have that𝜔 = 𝑗(𝜔1 , 𝜔2 ). □
 Draft: March 28, 2018

§5.3 Cohomology of Good Covers 165

Thus, applying Mayer–Vietoris to the compactly supported versions of the complexes

(5.2.6), we obtain the following theorem.
Theorem 5.2.32. Letting 𝑈 = 𝑈1 ∪ 𝑈2 and 𝑉 = 𝑈1 ∩ 𝑈2 there exists a long exact sequence in
compactly supported de Rham cohomology
𝛿 𝑖♯ 𝑗♯ 𝛿 𝑖♯
⋯ 𝐻𝑐𝑘 (𝑈) 𝐻𝑐𝑘 (𝑈1 ) ⊕ 𝐻𝑐𝑘 (𝑈2 ) 𝐻𝑐𝑘 (𝑉) 𝐻𝑐𝑘+1 (𝑈) ⋯.

Exercises for §5.2

Exercise 5.2.i. Prove Proposition 5.2.8.
Exercise 5.2.ii. Prove Proposition 5.2.13.
Exercise 5.2.iii. Prove Proposition 5.2.16.
Exercise 5.2.iv. Show that if 𝑈1 , 𝑈2 and 𝑈1 ∩ 𝑈2 are non-empty and connected the first
segment of the Mayer–Vietoris sequence is a short exact sequence
𝑖♯ 𝑗♯
0 𝐻0 (𝑈1 ∪ 𝑈2 ) 𝐻0 (𝑈1 ) ⊕ 𝐻0 (𝑈2 ) 𝐻0 (𝑈1 ∩ 𝑈2 ) 0.

Exercise 5.2.v. Let 𝑋 = 𝑆𝑛 and let 𝑈1 and 𝑈2 be the open subsets of 𝑆𝑛 obtained by removing
from 𝑆𝑛 the points, 𝑝1 = (0, …, 0, 1) and 𝑝2 = (0, …, 0, −1).
(1) Using stereographic projection show that 𝑈1 and 𝑈2 are diffeomorphic to 𝐑𝑛 .
(2) Show that 𝑈1 ∪ 𝑈2 = 𝑆𝑛 and 𝑈1 ∩ 𝑈2 is homotopy equivalent to 𝑆𝑛−1 . (See Exercise 5.1.v.)
Hint: 𝑈1 ∩ 𝑈2 is diffeomorphic to 𝐑𝑛 ∖ {0}.
(3) Deduce from the Mayer–Vietoris sequence that 𝐻𝑖+1 (𝑆𝑛 ) = 𝐻𝑖 (𝑆𝑛−1 ) for 𝑖 ≥ 1.
(4) Using part (3) give an inductive proof of a result that we proved by other means in
Section 5.1: 𝐻𝑘 (𝑆𝑛 ) = 0 for 1 ≤ 𝑘 < 𝑛.
Exercise 5.2.vi. Using the Mayer–Vietoris sequence of Exercise 5.2.v with cohomology re-
placed by compactly supported cohomology show that
𝐑, 𝑘 = 1, 𝑛
𝐻𝑐𝑘 (𝐑𝑛 ∖ {0}) ≅ {
0, otherwise .
Exercise 5.2.vii. Let 𝑛 a positive integer and let
𝑓1 𝑓𝑛−1
0 𝑉1 𝑉2 ⋯ 𝑉𝑛−1 𝑉𝑛 0
𝑛
be an exact sequence of finite dimensional vector spaces. Prove that ∑𝑖=1 dim(𝑉𝑖 ) = 0.

5.3. Cohomology of Good Covers

In this section we will show that for compact manifolds (and for lots of other manifolds
besides) the de Rham cohomology groups which we defined in §5.1 are finite dimensional
vector spaces and thus, in principle, “computable” objects. A key ingredient in our proof of
this fact is the notion of a good cover of a manifold.
Definition 5.3.1. Let 𝑋 be an 𝑛-manifold, and let U = {𝑈𝛼 }𝛼∈𝐼 be an open cover of 𝑋. Then
U is a good cover if for every finite set of indices 𝛼1 , …, 𝛼𝑘 ∈ 𝐼 the intersection 𝑈𝛼1 ∩⋯∩𝑈𝛼𝑘
is either empty or is diffeomorphic to 𝐑𝑛 .
One of our first goals in this section will be to show that good covers exist. We will
sketch below a proof of the following.
 Draft: March 28, 2018

166 Chapter 5: Cohomology via forms

Theorem 5.3.2. Every manifold admits a good cover.

The proof involves an elementary result about open convex subsets of 𝐑𝑛 .
Proposition 5.3.3. A bounded open convex subset 𝑈 ⊂ 𝐑𝑛 is diffeomorphic to 𝐑𝑛 .
A proof of this will be sketched in Exercises 5.3.i to 5.3.iv. One immediate consequence
of Proposition 5.3.3 is an important special case of Theorem 5.3.2.
Theorem 5.3.4. Every open subset 𝑈 of 𝐑𝑛 admits a good cover.
Proof. For each 𝑝 ∈ 𝑈 let 𝑈𝑝 be an open convex neighborhood of 𝑝 in 𝑈 (for instance an
𝜀-ball centered at 𝑝). Since the intersection of any two convex sets is again convex the cover,
{𝑈𝑝 }𝑝∈𝑈 is a good cover by Proposition 5.3.3. □

For manifolds the proof of Theorem 5.3.2 is somewhat trickier. The proof requires a
manifold analogue of the notion of convexity and there are several serviceable candidates.
The one we will use is the following. Let 𝑋 ⊂ 𝐑𝑁 be an 𝑛-manifold and for 𝑝 ∈ 𝑋 let 𝑇𝑝 𝑋
be the tangent space to 𝑋 at 𝑝. Recalling that 𝑇𝑝 𝑋 sits inside 𝑇𝑝 𝐑𝑁 and that

𝑇𝑝 𝐑𝑁 = { (𝑝, 𝑣) | 𝑣 ∈ 𝐑𝑁 }
we get a map
𝑇𝑝 𝑋 ↪ 𝑇𝑝 𝐑𝑁 → 𝐑𝑁 , (𝑝, 𝑥) ↦ 𝑝 + 𝑥 ,
and this map maps 𝑇𝑝 𝑋 bijectively onto an 𝑛-dimensional “affine” subspace 𝐿𝑝 of 𝐑𝑁 which
is tangent to 𝑋 at 𝑝. Let 𝜋𝑝 ∶ 𝑋 → 𝐿𝑝 be, as in the figure below, the orthogonal projection
of 𝑋 onto 𝐿𝑝 .

𝐿𝑝
𝑝 𝜋𝑝(𝑥)

Figure 5.3.1. The orthogonal projection of 𝑋 onto 𝐿𝑝

Definition 5.3.5. An open subset 𝑉 of 𝑋 is convex if for every 𝑝 ∈ 𝑉 the map 𝜋𝑝 ∶ 𝑋 → 𝐿𝑝

maps 𝑉 diffeomorphically onto a convex open subset of 𝐿𝑝 .
It’s clear from this definition of convexity that the intersection of two open convex sub-
sets of 𝑋 is an open convex subset of 𝑋 and that every open convex subset of 𝑋 is diffeo-
morphic to 𝐑𝑛 . Hence to prove Theorem 5.3.2 it suffices to prove that every point 𝑝 ∈ 𝑋 is
contained in an open convex subset 𝑈𝑝 of 𝑋. Here is a sketch of how to prove this. In the
figure above let 𝐵𝜀 (𝑝) be the ball of radius 𝜀 about 𝑝 in 𝐿𝑝 centered at 𝑝. Since 𝐿𝑝 and 𝑇𝑝
are tangent at 𝑝 the derivative of 𝜋𝑝 at 𝑝 is just the identity map, so for 𝜀 small 𝜋𝑝 maps a
neighborhood, 𝑈𝑝𝜀 of 𝑝 in 𝑋 diffeomorphically onto 𝐵𝜀 (𝑝). We claim:
Proposition 5.3.6. For 𝜀 small, 𝑈𝑝𝜀 is a convex subset of 𝑋.
 Draft: March 28, 2018

§5.3 Cohomology of Good Covers 167

Intuitively this assertion is pretty obvious: if 𝑞 is in 𝑈𝑝𝜀 and 𝜀 is small the map
𝜋𝑝−1 𝜋𝑞
𝐵𝑝𝜀 𝑈𝑝𝜀 𝐿𝑞

is to order 𝜀2 equal to the identity map, so it’s intuitively clear that its image is a slightly
warped, but still convex, copy of 𝐵𝜀 (𝑝). We won’t, however, bother to write out the details
that are required to make this proof rigorous.
A good cover is a particularly good “good cover” if it is a finite cover. We’ll codify this
property in the definition below.
Definition 5.3.7. An 𝑛-manifold 𝑋 is said to have finite topology if 𝑋 admits a finite cov-
ering by open sets 𝑈1 , …, 𝑈𝑁 with the property that for every multi-index, 𝐼 = (𝑖1 , …, 𝑖𝑘 ),
1 ≤ 𝑖1 ≤ 𝑖2 ⋯ < 𝑖𝐾 ≤ 𝑁, the set
(5.3.8) 𝑈𝐼 ≔ 𝑈𝑖1 ∩ ⋯ ∩ 𝑈𝑖𝑘
is either empty or is diffeomorphic to 𝐑𝑛 .
If 𝑋 is a compact manifold and U = {𝑈𝛼 }𝛼∈𝐼 is a good cover of 𝑋 then by the Heine–
Borel theorem we can extract from U a finite subcover 𝑈1 , …, 𝑈𝑁 , where
𝑈𝑖 ≔ 𝑈𝛼𝑖 , for 𝛼1 , …, 𝛼𝑁 ∈ 𝐼 ,
hence we conclude:
Theorem 5.3.9. Every compact manifold has finite topology.
More generally, for any manifold 𝑋, let 𝐶 be a compact subset of 𝑋. Then by Heine–
Borel we can extract from the cover U a finite subcollection 𝑈1 , …, 𝑈𝑁 , where
𝑈𝑖 ≔ 𝑈𝛼𝑖 , for 𝛼1 , …, 𝛼𝑁 ∈ 𝐼 ,
that covers 𝐶, hence letting 𝑈 ≔ 𝑈1 ∪ ⋯ ∪ 𝑈𝑁 , we’ve proved:
Theorem 5.3.10. If 𝑋 is an 𝑛-manifold and 𝐶 a compact subset of 𝑋, then there exists an open
neighborhood 𝑈 of 𝐶 in 𝑋 with finite topology.
We can in fact even strengthen this further. Let 𝑈0 be any open neighborhood of 𝐶 in
𝑋. Then in the theorem above we can replace 𝑋 by 𝑈0 to conclude:
Theorem 5.3.11. Let 𝑋 be a manifold, 𝐶 a compact subset of 𝑋, and 𝑈0 an open neighborhood
of 𝐶 in 𝑋. Then there exists an open neighborhood 𝑈 of 𝐶 in 𝑋, 𝑈 contained in 𝑈0 , so that 𝑈
has finite topology.
We will justify the term “finite topology” by devoting the rest of this section to proving.
Theorem 5.3.12. Let 𝑋 be an 𝑛-manifold. If 𝑋 has finite topology the de Rham cohomology
groups 𝐻𝑘 (𝑋), for 𝑘 = 0, …, 𝑛, and the compactly supported de Rham cohomology groups
𝐻𝑐𝑘 (𝑋), for 𝑘 = 0, …, 𝑛 are finite dimensional vector spaces.
The basic ingredients in the proof of this will be the Mayer–Vietoris techniques that we
developed in §5.2 and the following elementary result about vector spaces.
Lemma 5.3.13. Let 𝑉1 , 𝑉2 , and 𝑉3 be vector spaces and
𝛼 𝛽
(5.3.14) 𝑉1 𝑉2 𝑉3
an exact sequence of linear maps. Then if 𝑉1 and 𝑉3 are finite dimensional, so is 𝑉2 .
 Draft: March 28, 2018

168 Chapter 5: Cohomology via forms

Proof. Since 𝑉3 is finite dimensional, the im(𝛽) is of finite dimension 𝑘. Hence there exist
vectors 𝑣1 , …, 𝑣𝑘 ∈ 𝑉2 such that
im(𝛽) = span(𝛽(𝑣1 ), …, 𝛽(𝑣𝑘 )) .
𝑘
Now let 𝑣 ∈ 𝑉2 . Then 𝛽(𝑣) is a linear combination 𝛽(𝑣) = ∑𝑖=1 𝑐𝑖 𝛽(𝑣𝑖 ), where 𝑐1 , …, 𝑐𝑘 ∈
𝐑. So
𝑘
(5.3.15) 𝑣′ ≔ 𝑣 − ∑ 𝑐𝑖 𝑣𝑖
𝑖=1

is in the kernel of 𝛽 and hence, by the exactness of (5.3.14), in the image of 𝛼. But 𝑉1 is finite
dimensional, so im(𝛼) is finite dimensional. Letting 𝑣𝑘+1 , …, 𝑣𝑚 be a basis of im(𝛼) we can
𝑚
by (5.3.15)) write 𝑣 as a sum 𝑣 = ∑𝑖=1 𝑐𝑖 𝑣𝑖 . In other words 𝑣1 , …, 𝑣𝑚 is a basis of 𝑉2 . □

Proof of Theorem 5.3.4. Our proof will be by induction on the number of open sets in a good
cover of 𝑋. More specifically, let
U = {𝑈1 , …, 𝑈𝑁 }
be a good cover of 𝑋. If 𝑁 = 1, 𝑋 = 𝑈1 and hence 𝑋 is diffeomorphic to 𝐑𝑛 , so 𝐻𝑘 (𝑋) = 0
for 𝑘 > 0, and 𝐻0 (𝑋) ≅ 𝐑, so the theorem is certainly true in this case. Let us now prove
it’s true for arbitrary 𝑁 by induction. Let 𝑈 ≔ 𝑈2 ∪ ⋯ ∪ 𝑈𝑁 . Then 𝑈 is a submanifold of 𝑋,
and, thinking of 𝑈 as a manifold in its own right, {𝑈2 , …, 𝑈𝑁 } is a good cover of 𝑈 involving
only 𝑁 − 1 sets. Hence the cohomology groups of 𝑈 are finite dimensional by the induction
hypothesis. The manifold 𝑈 ∩ 𝑈1 has a good cover given by {𝑈 ∩ 𝑈2 , …, 𝑈 ∩ 𝑈𝑁 }, so by the
induction hypothesis the cohomology groups of 𝑈∩𝑈1 are finite-dimensional. To prove that
the theorem is true for 𝑋 we note that 𝑋 = 𝑈1 ∪ 𝑈, and and the Mayer–Vietoris sequence
gives an exact sequence

𝛿 𝑖♯
𝐻𝑘−1 (𝑈1 ∩ 𝑈) 𝐻𝑘 (𝑋) 𝐻𝑘 (𝑈1 ) ⊕ 𝐻𝑘 (𝑈) .

Since the right-hand and left-hand terms are finite dimensional it follows from Lemma 5.3.13
that the middle term is also finite dimensional. □

The proof works practically verbatim for compactly supported cohomology. For 𝑁 = 1
𝐻𝑐𝑘 (𝑋) = 𝐻𝑐𝑘 (𝑈1 ) = 𝐻𝑐𝑘 (𝐑𝑛 )
so all the cohomology groups of 𝐻𝑘 (𝑋) are finite in this case, and the induction “𝑁 − 1” ⇒
“𝑁” follows from the exact sequence
𝑗♯ 𝛿
𝐻𝑐𝑘 (𝑈1 ) ⊕ 𝐻𝑐𝑘 (𝑈) 𝐻𝑐𝑘 (𝑋) 𝐻𝑐𝑘+1 (𝑈1 ∩ 𝑈) .

Remark 5.3.16. A careful analysis of the proof above shows that the dimensions of the
vector spaces 𝐻𝑘 (𝑋) are determined by the intersection properties of the open sets 𝑈𝑖 , i.e.,
by the list of multi-indices 𝐼 for which th intersections (5.3.8) are non-empty.
This collection of multi-indices is called the nerve of the cover U, and this remark
suggests that there should be a cohomology theory which has as input the nerve of U and
as output cohomology groups which are isomorphic to the de Rham cohomology groups.
Such a theory does exist, and we further address it in §5.8. (A nice account of it can also be
found in [13, Ch. 5]).
 Draft: March 28, 2018

§5.3 Cohomology of Good Covers 169

Exercises for §5.3

Exercise 5.3.i. Let 𝑈 be a bounded open subset of 𝐑𝑛 . A continuous function
𝜓 ∶ 𝑈 → [0, ∞)
is called an exhaustion function if it is proper as a map of 𝑈 into [0, ∞); i.e., if, for every
𝑎 > 0, 𝜓−1 ([0, 𝑎]) is compact. For 𝑥 ∈ 𝑈 let
𝑑(𝑥) = inf{ |𝑥 − 𝑦| | 𝑦 ∈ 𝐑𝑛 − 𝑈 } ,
i.e., let 𝑑(𝑥) be the “distance” from 𝑥 to the boundary of 𝑈. Show that 𝑑(𝑥) > 0 and that 𝑑(𝑥)
is continuous as a function of 𝑥. Conclude that 𝜓0 = 1/𝑑 is an exhaustion function.
Exercise 5.3.ii. Show that there exists a 𝐶∞ exhaustion function, 𝜙0 ∶ 𝑈 → [0, ∞), with the
property 𝜙0 ≥ 𝜓02 where 𝜓0 is the exhaustion function in Exercise 5.3.i.
Hints: For 𝑖 = 2, 3, … let
𝐶𝑖 = { 𝑥 ∈ 𝑈 | 1𝑖 ≤ 𝑑(𝑥) ≤ 1
𝑖−1
}
and
1 1
𝑈𝑖 = { 𝑥 ∈ 𝑈 | 𝑖+1 < 𝑑(𝑥) < 𝑖−2
} .
Let 𝜌𝑖 ∈ 𝐶0∞ (𝑈𝑖 ), 𝜌𝑖 ≥ 0, be a “bump” function which is identically one on 𝐶𝑖 and let 𝜙0 =
∞
∑𝑖=2 𝑖2 𝜌𝑖 + 1.
Exercise 5.3.iii. Let 𝑈 be a bounded open convex subset of 𝐑𝑛 containing the origin. Show
that there exists an exhaustion function
𝜓 ∶ 𝑈 → 𝐑 , 𝜓(0) = 1 ,
having the property that 𝜓 is a monotonically increasing function of 𝑡 along the ray, 𝑡𝑥,
0 ≤ 𝑡 ≤ 1, for all points 𝑥 ∈ 𝑈.
Hints:
• Let 𝜌(𝑥), 0 ≤ 𝜌(𝑥) ≤ 1, be a 𝐶∞ function which is one outside a small neighbor-
hood of the origin in 𝑈 and is zero in a still smaller neighborhood of the origin.
Modify the function, 𝜙0 , in the previous exercise by setting 𝜙(𝑥) = 𝜌(𝑥)𝜙0 (𝑥) and
let
1
𝑑𝑠
𝜓(𝑥) = ∫ 𝜙(𝑠𝑥) + 1 .
0 𝑠
Show that for 0 ≤ 𝑡 ≤ 1
𝑑𝜓
(𝑡𝑥) = 𝜙(𝑡𝑥)/𝑡
𝑑𝑡
and conclude from equation (5.3.15) that 𝜓 is monotonically increasing along the
ray, 𝑡𝑥, 0 ≤ 𝑡 ≤ 1.
• Show that for 0 < 𝜀 < 1,
𝜓(𝑥) ≥ 𝜀𝜙(𝑦)
where 𝑦 is a point on the ray, 𝑡𝑥, 0 ≤ 𝑡 ≤ 1 a distance less than 𝜀|𝑥| from 𝑋.
• Show that there exist constants, 𝐶0 and 𝐶1 , 𝐶1 > 0 such that
𝐶1
𝜓(𝑥) = + 𝐶0 .
𝑑(𝑥)
(Take 𝜀 to be equal to 12 𝑑(𝑥)/|𝑥|.)
 Draft: March 28, 2018

170 Chapter 5: Cohomology via forms

Exercise 5.3.iv. Show that every bounded, open convex subset, 𝑈, of 𝐑𝑛 is diffeomorphic
to 𝐑𝑛 .
Hints:
• Let 𝜓(𝑥) be the exhaustion function constructed in Exercise 5.3.iii and let
𝑓 ∶ 𝑈 → 𝐑𝑛
be the map: 𝑓(𝑥) = 𝜓(𝑥)𝑥. Show that this map is a bijective map of 𝑈 onto 𝐑𝑛 .
• Show that for 𝑥 ∈ 𝑈 and 𝑣 ∈ 𝐑𝑛
(𝑑𝑓)𝑥 𝑣 = 𝜓(𝑥)𝑣 + 𝑑𝜓𝑥 (𝑣)𝑥
and conclude that 𝑑𝑓𝑥 is bijective at 𝑥, i.e., that 𝑓 is locally a diffeomorphism of a
neighborhood of 𝑥 in 𝑈 onto a neighborhood of 𝑓(𝑥) in 𝐑𝑛 .
• Putting these together show that 𝑓 is a diffeomorphism of 𝑈 onto 𝐑𝑛 .
Exercise 5.3.v. Let 𝑈 ⊂ 𝐑 be the union of the open intervals, 𝑘 < 𝑥 < 𝑘 + 1, 𝑘 an integer.
Show that 𝑈 does not have finite topology.
Exercise 5.3.vi. Let 𝑉 ⊂ 𝐑2 be the open set obtained by deleting from 𝐑2 the points, 𝑝𝑛 =
(0, 𝑛), for every integer 𝑛. Show that 𝑉 does not have finite topology.
Hint: Let 𝛾𝑛 be a circle of radius 21 centered about the point 𝑝𝑛 . Using Exercises 2.1.viii
and 2.1.ix show that there exists a closed smooth one-form, 𝜔𝑛 on 𝑉 with the property that
∫𝛾 𝜔𝑛 = 1 and ∫𝛾 𝜔𝑛 = 0 for 𝑚 ≠ 𝑛.
𝑛 𝑚

Exercise 5.3.vii. Let 𝑋 be an 𝑛-manifold and U = {𝑈1 , 𝑈2 } a good cover of 𝑋. What are the
cohomology groups of 𝑋 if the nerve of this cover is:
(1) {1}, {2}.
(2) {1}, {2}, {1, 2}.
Exercise 5.3.viii. Let 𝑋 be an 𝑛-manifold and U = {𝑈1 , 𝑈2 , 𝑈3 } a good cover of 𝑋. What
are the cohomology groups of 𝑋 if the nerve of this cover is:
(1) {1}, {2}, {3}.
(2) {1}, {2}, {3}, {1, 2}.
(3) {1}, {2}, {3}, {1, 2}, {1, 3}
(4) {1}, {2}, {3}, {1, 2}, {1, 3}, {2, 3}.
(5) {1}, {2}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3}.
Exercise 5.3.ix. Let 𝑆1 be the unit circle in 𝐑3 parametrized by arc length: (𝑥, 𝑦) = (cos 𝜃, sin 𝜃).
Let 𝑈1 be the set: 0 < 𝜃 < 2𝜋
3
, 𝑈2 the set: 𝜋2 < 𝜃 < 3𝜋
2
, and 𝑈3 the set: − 2𝜋
3
< 𝜃 < 𝜋3 .
(1) Show that the 𝑈𝑖 ’s are a good cover of 𝑆1 .
(2) Using the previous exercise compute the cohomology groups of 𝑆1 .
Exercise 5.3.x. Let 𝑆2 be the unit 2-sphere in 𝐑3 . Show that the sets
𝑈𝑖 = { (𝑥1 , 𝑥2 , 𝑥3 ) ∈ 𝑆2 | 𝑥𝑖 > 0 }
𝑖 = 1, 2, 3 and
𝑈𝑖 = { (𝑥1 , 𝑥2 , 𝑥3 ) ∈ 𝑆2 | 𝑥𝑖−3 < 0 } ,
𝑖 = 4, 5, 6, are a good cover of 𝑆2 . What is the nerve of this cover?
Exercise 5.3.xi. Let 𝑋 and 𝑌 be manifolds. Show that if they both have finite topology, their
product 𝑋 × 𝑌 does as well.
Exercise 5.3.xii.
 Draft: March 28, 2018

§5.4 Poincaré duality 171

(1) Let 𝑋 be a manifold and let 𝑈1 , …, 𝑈𝑁 , be a good cover of 𝑋. Show that 𝑈1 ×𝐑, …, 𝑈𝑁 ×𝐑
is a good cover of 𝑋 × 𝐑 and that the nerves of these two covers are the same.
(2) By Remark 5.3.16,
𝐻𝑘 (𝑋 × 𝐑) ≅ 𝐻𝑘 (𝑋) .
Verify this directly using homotopy techniques.
(3) More generally, show that for all ℓ ≥ 0
𝐻𝑘 (𝑋 × 𝐑ℓ ) ≅ 𝐻𝑘 (𝑋)
(a) by concluding that this has to be the case in view of the Remark 5.3.16,
(b) and by proving this directly using homotopy techniques.

5.4. Poincaré duality

In this chapter we’ve been studying two kinds of cohomology groups: the ordinary de
Rham cohomology groups 𝐻𝑘 and the compactly supported de Rham cohomology groups
𝐻𝑐𝑘 . It turns out that these groups are closely related. In fact if 𝑋 is a connected oriented
𝑛-manifold and has finite topology, then 𝐻𝑐𝑛−𝑘 (𝑋) is the vector space dual of 𝐻𝑘 (𝑋). We
give a proof of this later in this section, however, before we do we need to review some basic
linear algebra.
Given finite dimensional vector spaces 𝑉 and 𝑊, a bilinear pairing between 𝑉 and 𝑊
is a map
(5.4.1) 𝐵∶ 𝑉 × 𝑊 → 𝐑
which is linear in each of its factors. In other words, for fixed 𝑤 ∈ 𝑊, the map
ℓ𝑤 ∶ 𝑉 → 𝐑 , 𝑣 ↦ 𝐵(𝑣, 𝑤)
is linear, and for 𝑣 ∈ 𝑉, the map
ℓ𝑣 ∶ 𝑊 → 𝐑 , 𝑤 ↦ 𝐵(𝑣, 𝑤)
is linear. Therefore, from the pairing (5.4.1) one gets a map
(5.4.2) 𝐿𝐵 ∶ 𝑊 → 𝑉⋆ , 𝑤 ↦ ℓ𝑤
and since ℓ𝑤1 + ℓ𝑤2 (𝑣) = 𝐵(𝑣, 𝑤1 + 𝑤2 ) = ℓ𝑤1 +𝑤2 (𝑣), this map is linear. We’ll say that (5.4.1)
is a non-singular pairing if (5.4.2) is bijective. Notice, by the way, that the roles of 𝑉 and
𝑊 can be reversed in this definition. Letting 𝐵♯ (𝑤, 𝑣) ≔ 𝐵(𝑣, 𝑤) we get an analogous linear
map
𝐿𝐵 ♯ ∶ 𝑉 → 𝑊 ⋆
and in fact
(5.4.3) (𝐿𝐵♯ (𝑣))(𝑤) = (𝐿𝐵 (𝑤))(𝑣) = 𝐵(𝑣, 𝑤) .
Thus if
𝜇 ∶ 𝑉 → (𝑉⋆ )⋆
is the canonical identification of 𝑉 with (𝑉⋆ )⋆ given by the recipe
𝜇(𝑣)(ℓ) = ℓ(𝑣)
for 𝑣 ∈ 𝑉 and ℓ ∈ 𝑉⋆ , we can rewrite equation (5.4.3) more suggestively in the form
(5.4.4) 𝐿𝐵♯ = 𝐿⋆𝐵 𝜇
i.e., 𝐿𝐵 and 𝐿𝐵♯ are just the transposes of each other. In particular 𝐿𝐵 is bijective if and only
if 𝐿𝐵♯ is bijective.
 Draft: March 28, 2018

172 Chapter 5: Cohomology via forms

Let us now apply these remarks to de Rham theory. Let 𝑋 be a connected, oriented
𝑛-manifold. If 𝑋 has finite topology the vector spaces, 𝐻𝑐𝑛−𝑘 (𝑋) and 𝐻𝑘 (𝑋) are both finite
dimensional. We will show that there is a natural bilinear pairing between these spaces, and
hence by the discussion above, a natural linear mapping of 𝐻𝑘 (𝑋) into the vector space dual
of 𝐻𝑐𝑛−1 (𝑋). To see this let 𝑐1 be a cohomology class in 𝐻𝑐𝑛−𝑘 (𝑋) and 𝑐2 a cohomology class
in 𝐻𝑘 (𝑋). Then by (5.1.22) their product 𝑐1 ⋅ 𝑐2 is an element of 𝐻𝑐𝑛 (𝑋), and so by (5.1.2) we
can define a pairing between 𝑐1 and 𝑐2 by setting
(5.4.5) 𝐵(𝑐1 , 𝑐2 ) ≔ 𝐼𝑋 (𝑐1 ⋅ 𝑐2 ) .
Notice that if 𝜔1 ∈ 𝛺𝑛−𝑘
𝑐 (𝑋) and 𝜔2 ∈ 𝛺𝑘 (𝑋) are closed forms representing the cohomology
classes 𝑐1 and 𝑐2 , respectively, then by (5.1.22) this pairing is given by the integral

𝐵(𝑐1 , 𝑐2 ) ≔ ∫ 𝜔1 ∧ 𝜔2 .
𝑋
We’ll next show that this bilinear pairing is non-singular in one important special case:
Proposition 5.4.6. If 𝑋 is diffeomorphic to 𝐑𝑛 the pairing defined by (5.4.5) is non-singular.
Proof. To verify this there is very little to check. The vector spaces 𝐻𝑘 (𝐑𝑛 ) and 𝐻𝑐𝑛−𝑘 (𝐑𝑛 )
are zero except for 𝑘 = 0, so all we have to check is that the pairing
𝐻𝑐𝑛 (𝑋) × 𝐻0 (𝑋) → 𝐑
is non-singular. To see this recall that every compactly supported 𝑛-form is closed and that
the only closed zero-forms are the constant functions, so at the level of forms, the pairing
(5.4.5) is just the pairing

𝛺𝑛 (𝑋) × 𝐑 → 𝐑 , (𝜔, 𝑐) ↦ 𝑐 ∫ 𝜔 ,
𝑋
and this is zero if and only if 𝑐 is zero or 𝜔 is in 𝑑𝛺𝑛−1
𝑐 (𝑋). Thus at the level of cohomology
this pairing is non-singular. □
We will now show how to prove this result in general.
Theorem 5.4.7 (Poincaré duality). Let 𝑋 be an oriented, connected 𝑛-manifold having finite
topology. Then the pairing (5.4.5) is non-singular.
The proof of this will be very similar in spirit to the proof that we gave in the last sec-
tion to show that if 𝑋 has finite topology its de Rham cohomology groups are finite dimen-
sional. Like that proof, it involves Mayer–Vietoris plus some elementary diagram-chasing.
The “diagram-chasing” part of the proof consists of the following two lemmas.
Lemma 5.4.8. Let 𝑉1 , 𝑉2 and 𝑉3 be finite dimensional vector spaces, and let
𝛼 𝛽
𝑉1 𝑉2 𝑉3
be an exact sequence of linear mappings. Then the sequence of transpose maps
𝛽⋆ 𝛼⋆
𝑉3⋆ 𝑉2⋆ 𝑉1⋆
is exact.
Proof. Given a vector subspace, 𝑊2 , of 𝑉2 , let
𝑊2⟂ = { ℓ ∈ 𝑉2⋆ | ℓ(𝑤) = 0 for all 𝑤 ∈ 𝑊 } .
We’ll leave for you to check that if 𝑊2 is the kernel of 𝛽, then 𝑊2⟂ is the image of 𝛽⋆ and that
if 𝑊2 is the image of 𝛼, 𝑊2⟂ is the kernel of 𝛼⋆ . Hence if 𝛽 = im(𝛼), im(𝛽⋆ ) = ker(𝛼⋆ ). □
 Draft: March 28, 2018

§5.4 Poincaré duality 173

Lemma 5.4.9 (5-lemma). Consider a commutative diagram of vector spaces

𝛽1 𝛽2 𝛽3 𝛽4
𝐵1 𝐵2 𝐵3 𝐵4 𝐵5
(5.4.10) 𝛾1 𝛾2 𝛾3 𝛾4 𝛾5

𝐴1 𝛼1 𝐴2 𝛼2 𝐴3 𝛼3 𝐴4 𝛼4 𝐴5 .

satisfying the following conditions:

(1) All the vector spaces are finite dimensional.
(2) The two rows are exact.
(3) The linear maps, 𝛾1 , 𝛾2 , 𝛾4 , and 𝛾5 are bijections.
Then the map 𝛾3 is a bijection.
Proof. We’ll show that 𝛾3 is surjective. Given 𝑎3 ∈ 𝐴 3 there exists a 𝑏4 ∈ 𝐵4 such that
𝛾4 (𝑏4 ) = 𝛼3 (𝑎3 ) since 𝛾4 is bijective. Moreover, 𝛾5 (𝛽4 (𝑏4 )) = 𝛼4 (𝛼3 (𝑎3 )) = 0, by the exactness
of the top row. Therefore, since 𝛾5 is bijective, 𝛽4 (𝑏4 ) = 0, so by the exactness of the bottom
row 𝑏4 = 𝛽3 (𝑏3 ) for some 𝑏3 ∈ 𝐵3 , and hence
𝛼3 (𝛾3 (𝑏3 )) = 𝛾4 (𝛽3 (𝑏3 )) = 𝛾4 (𝑏4 ) = 𝛼3 (𝑎3 ) .
Thus 𝛼3 (𝑎3 − 𝛾3 (𝑏3 )) = 0, so by the exactness of the top row
𝑎3 − 𝛾3 (𝑏3 ) = 𝛼2 (𝑎2 )
for some 𝑎2 ∈ 𝐴 2 . Hence by the bijectivity of 𝛾2 there exists a 𝑏2 ∈ 𝐵2 with 𝑎2 = 𝛾2 (𝑏2 ), and
hence
𝑎3 − 𝛾3 (𝑏3 ) = 𝛼2 (𝑎2 ) = 𝛼2 (𝛾2 (𝑏2 )) = 𝛾3 (𝛽2 (𝑏2 )) .
Thus finally
𝑎3 = 𝛾3 (𝑏3 + 𝛽2 (𝑏2 )) .
Since 𝑎3 was any element of 𝐴 3 this proves the surjectivity of 𝛾3 .
One can prove the injectivity of 𝛾3 by a similar diagram-chasing argument, but one can
also prove this with less duplication of effort by taking the transposes of all the arrows in
(5.4.10) and noting that the same argument as above proves the surjectivity of 𝛾3⋆ ∶ 𝐴⋆3 →
𝐵3⋆ . □
To prove Theorem 5.4.7 we apply these lemmas to the diagram below. In this diagram
𝑈1 and 𝑈2 are open subsets of 𝑋, 𝑀 = 𝑈1 ∪ 𝑈2 , we write 𝑈1,2 ≔ 𝑈1 ∩ 𝑈2 , and the vertical
arrows are the mappings defined by the pairing (5.4.5). We will leave for you to check that
this is a commutative diagram “up to sign”. (To make it commutative one has to replace
some of the vertical arrows, 𝛾, by their negatives: −𝛾.) This is easy to check except for the
commutative square on the extreme left. To check that this square commutes, some serious
diagram-chasing is required.

⋯ 𝐻𝑐𝑘−1 (𝑀) 𝐻𝑐𝑘 (𝑈1,2 ) 𝐻𝑐𝑘 (𝑈1 ) ⊕ 𝐻𝑐𝑘 (𝑈2 ) 𝐻𝑐𝑘 (𝑀) ⋯

⋯ 𝐻𝑛−(𝑘−1) (𝑀)⋆ 𝐻𝑛−𝑘 (𝑈1,2 )⋆ 𝐻𝑛−𝑘 (𝑈1 )⋆ ⊕ 𝐻𝑛−𝑘 (𝑈2 )⋆ 𝐻𝑛−𝑘 (𝑀)⋆ ⋯

By Mayer–Vietoris the bottom row of the diagram is exact and by Mayer–Vietoris and
Lemma 5.4.8 the top row of the diagram is exact. Hence we can apply the “five lemma” to
the diagram to conclude:
 Draft: March 28, 2018

174 Chapter 5: Cohomology via forms

Lemma 5.4.11. If the maps

𝐻𝑘 (𝑈) → 𝐻𝑐𝑛−𝑘 (𝑈)⋆
defined by the pairing (5.4.5) are bijective for 𝑈1 , 𝑈2 and 𝑈1,2 = 𝑈1 ∩ 𝑈2 , they are also bijective
for 𝑀 = 𝑈1 ∪ 𝑈2 .
Thus to prove Theorem 5.4.7 we can argue by induction as in §5.3. Let 𝑈1 , 𝑈2 , …, 𝑈𝑁
be a good cover of 𝑋. If 𝑁 = 1, then 𝑋 = 𝑈1 and, hence, since 𝑈1 is diffeomorphic to
𝐑𝑛 , the map (5.4.12) is bijective by Proposition 5.4.6. Now Let us assume the theorem is
true for manifolds involving good covers by 𝑘 open sets where 𝑘 is less than 𝑁. Let 𝑈′ =
𝑈1 ∪ ⋯ ∪ 𝑈𝑁−1 and 𝑈″ = 𝑈𝑁 . Since
𝑈′ ∩ 𝑈″ = 𝑈1 ∩ 𝑈𝑁 ∪ ⋯ ∪ 𝑈𝑁−1 ∩ 𝑈𝑁
it can be covered by a good cover by 𝑘 open sets, 𝑘 < 𝑁, and hence the hypotheses of the
lemma are true for 𝑈′ , 𝑈″ and 𝑈′ ∩ 𝑈″ . Thus the lemma says that (5.4.12) is bijective for
the union 𝑋 of 𝑈′ and 𝑈″ . □
Exercises for §5.4
Exercise 5.4.i (proper pushforward for compactly supported de Rham cohomology). Let
𝑋 be an 𝑚-manifold, 𝑌 an 𝑛-manifold and 𝑓 ∶ 𝑋 → 𝑌 a 𝐶∞ map. Suppose that both of
these manifolds are oriented and connected and have finite topology. Show that there exists
a unique linear map
(5.4.12) 𝑓! ∶ 𝐻𝑐𝑚−𝑘 (𝑋) → 𝐻𝑐𝑛−𝑘 (𝑌)
with the property
(5.4.13) 𝐵𝑌 (𝑓! 𝑐1 , 𝑐2 ) = 𝐵𝑋 (𝑐1 , 𝑓♯ 𝑐2 )
for all 𝑐1 ∈ 𝐻𝑐𝑚−𝑘 (𝑋) and 𝑐2 ∈ 𝐻𝑘 (𝑌). (In this formula 𝐵𝑋 is the bilinear pairing (5.4.5) on
𝑋 and 𝐵𝑌 is the bilinear pairing (5.4.5) on 𝑌.)
Exercise 5.4.ii. Suppose that the map 𝑓 in Exercise 5.4.i is proper. Show that there exists a
unique linear map
𝑓! ∶ 𝐻𝑚−𝑘 (𝑋) → 𝐻𝑛−𝑘 (𝑌)
with the property
𝐵𝑌 (𝑐1 , 𝑓! 𝑐2 ) = (−1)𝑘(𝑚−𝑛) 𝐵𝑋 (𝑓♯ 𝑐1 , 𝑐2 )
for all 𝑐1 ∈ 𝐻𝑐𝑘 (𝑌) and 𝑐2 ∈ 𝐻𝑚−𝑘 (𝑋), and show that, if 𝑋 and 𝑌 are compact, this mapping
is the same as the mapping 𝑓! in Exercise 5.4.i.
Exercise 5.4.iii. Let 𝑈 be an open subset of 𝐑𝑛 and let 𝑓 ∶ 𝑈 × 𝐑 → 𝑈 be the projection,
𝑓(𝑥, 𝑡) = 𝑥. Show that there is a unique linear mapping
𝑓⋆ ∶ 𝛺𝑘+1 𝑘
𝑐 (𝑈 × 𝐑) → 𝛺𝑐 (𝑈)
with the property

(5.4.14) ∫ 𝑓⋆ 𝜇 ∧ 𝜈 = ∫ 𝜇 ∧ 𝑓⋆ 𝜈
𝑈 𝑈×𝐑

for all 𝜇 ∈ 𝛺𝑘+1

𝑐 (𝑈 × 𝐑) and 𝜈 ∈ 𝛺
𝑛−𝑘
(𝑈).
Hint: Let 𝑥1 , …, 𝑥𝑛 and 𝑡 be the standard coordinate functions on 𝐑𝑛 × 𝐑. By Exer-
cise 2.3.v every (𝑘 + 1)-form 𝜔 ∈ 𝛺𝑘+1
𝑐 (𝑈 × 𝐑) can be written uniquely in “reduced form” as
a sum
𝜔 = ∑ 𝑓𝐼 𝑑𝑡 ∧ 𝑑𝑥𝐼 + ∑ 𝑔𝐽 𝑑𝑥𝐽
𝐼 𝐽
 Draft: March 28, 2018

§5.4 Poincaré duality 175

over multi-indices, 𝐼 and 𝐽, which are strictly increasing. Let

𝑓⋆ 𝜔 = ∑ (∫ 𝑓𝐼 (𝑥, 𝑡)𝑑𝑡) 𝑑𝑥𝐼 .

𝐼 𝐑

Exercise 5.4.iv. Show that the mapping 𝑓⋆ in Exercise 5.4.iii satisfies 𝑓⋆ 𝑑𝜔 = 𝑑𝑓⋆ 𝜔.
Exercise 5.4.v. Show that if 𝜔 is a closed compactly supported (𝑘 + 1)-form on 𝑈 × 𝐑 then
[𝑓⋆ 𝜔] = 𝑓! [𝜔]
where 𝑓! is the mapping (5.4.13) and 𝑓⋆ the mapping (5.4.14).
Exercise 5.4.vi.
(1) Let 𝑈 be an open subset of 𝐑𝑛 and let 𝑓 ∶ 𝑈 × 𝐑ℓ → 𝑈 be the projection, 𝑓(𝑥, 𝑡) = 𝑥.
Show that there is a unique linear mapping
𝑓⋆ ∶ 𝛺𝑘+ℓ ℓ 𝑘
𝑐 (𝑈 × 𝐑 ) → 𝛺𝑐 (𝑈)
with the property
∫ 𝑓⋆ 𝜇 ∧ 𝜈 = ∫ 𝜇 ∧ 𝑓⋆ 𝜈
𝑈 𝑈×𝐑ℓ
for all 𝜇 ∈ 𝛺𝑘+ℓ ℓ
𝑐 (𝑈 × 𝐑 ) and 𝜈 ∈ 𝛺
𝑛−𝑘
(𝑈).
Hint: Exercise 5.4.iii plus induction on ℓ.
(2) Show that for 𝜔 ∈ 𝛺𝑘+ℓ
𝑐 (𝑈 × 𝐑 )
ℓ

𝑑𝑓⋆ 𝜔 = 𝑓⋆ 𝑑𝜔 .
(3) Show that if 𝜔 is a closed, compactly supported (𝑘 + ℓ)-form on 𝑋 × 𝐑ℓ
[𝑓⋆ 𝜔] = 𝑓! [𝜔]
where 𝑓! ∶ 𝐻𝑐𝑘+ℓ (𝑈 × 𝐑ℓ ) → 𝐻𝑐𝑘 (𝑈) is the map (5.4.13).
Exercise 5.4.vii. Let 𝑋 be an 𝑛-manifold and 𝑌 an 𝑚-manifold. Assume 𝑋 and 𝑌 are com-
pact, oriented and connected, and orient 𝑋 × 𝑌 by giving it its natural product orientation.
Let
𝑓∶ 𝑋 × 𝑌 → 𝑌
be the projection map 𝑓(𝑥, 𝑦) = 𝑦. Given
𝜔 ∈ 𝛺𝑚 (𝑋 × 𝑌)
and 𝑝 ∈ 𝑌, let

(5.4.15) 𝑓⋆ 𝜔(𝑝) ≔ ∫ 𝜄𝑝⋆ 𝜔 ,

𝑋
where 𝜄𝑝 ∶ 𝑋 → 𝑋 × 𝑌 is the inclusion map 𝜄𝑝 (𝑥) = (𝑥, 𝑝).
(1) Show that the function 𝑓⋆ 𝜔 defined by equation (5.4.15) is 𝐶∞ , i.e., is in 𝛺0 (𝑌).
(2) Show that if 𝜔 is closed this function is constant.
(3) Show that if 𝜔 is closed
[𝑓⋆ 𝜔] = 𝑓! [𝜔]
𝑛 0
where 𝑓! ∶ 𝐻 (𝑋 × 𝑌) → 𝐻 (𝑌) is the map (5.4.13).
Exercise 5.4.viii.
(1) Let 𝑋 be an 𝑛-manifold which is compact, connected and oriented. Combining Poincaré
duality with Exercise 5.3.xii show that
𝐻𝑐𝑘+ℓ (𝑋 × 𝐑ℓ ) ≅ 𝐻𝑐𝑘 (𝑋) .
 Draft: March 28, 2018

176 Chapter 5: Cohomology via forms

(2) Show, moreover, that if 𝑓 ∶ 𝑋 × 𝐑ℓ → 𝑋 is the projection 𝑓(𝑥, 𝑎) = 𝑥, then

𝑓! ∶ 𝐻𝑐𝑘+ℓ (𝑋 × 𝐑ℓ ) → 𝐻𝑐𝑘 (𝑋)
is a bijection.
Exercise 5.4.ix. Let 𝑋 and 𝑌 be as in Exercise 5.4.i. Show that the pushforward operation
(5.4.13) satisfies the projection formula
𝑓! (𝑐1 ⋅ 𝑓♯ 𝑐2 ) = 𝑓! (𝑐1 ) ⋅ 𝑐2 ,
for 𝑐1 ∈ 𝐻𝑐𝑘 (𝑋) and 𝑐2 ∈ 𝐻ℓ (𝑌).

5.5. Thom classes & intersection theory

Let 𝑋 be a connected, oriented 𝑛-manifold. If 𝑋 has finite topology its cohomology
groups are finite-dimensional, and since the bilinear pairing 𝐵 defined by (5.4.5) is non-
singular we get from this pairing bijective linear maps
(5.5.1) 𝐿𝐵 ∶ 𝐻𝑐𝑛−𝑘 (𝑋) → 𝐻𝑘 (𝑋)⋆
and
(5.5.2) 𝐿⋆𝐵 ∶ 𝐻𝑛−𝑘 (𝑋) → 𝐻𝑐𝑘 (𝑋)⋆ .
In particular, if ℓ ∶ 𝐻𝑘 (𝑋) → 𝐑 is a linear function (i.e., an element of 𝐻𝑘 (𝑋)⋆ ), then by
(5.5.1) we can convert ℓ into a cohomology class
(5.5.3) 𝐿−1 𝑛−𝑘
𝐵 (ℓ) ∈ 𝐻𝑐 (𝑋) ,
and similarly if ℓ𝑐 ∶ 𝐻𝑐𝑘 (𝑋) → 𝐑 is a linear function, we can convert it by (5.5.2) into a
cohomology class
(𝐿⋆𝐵 )−1 (ℓ) ∈ 𝐻𝑛−𝑘 (𝑋) .
One way that linear functions like this arise in practice is by integrating forms over
submanifolds of 𝑋. Namely let 𝑌 be a closed oriented 𝑘-dimensional submanifold of 𝑋.
Since 𝑌 is oriented, we have by (5.1.2) an integration operation in cohomology
𝐼𝑌 ∶ 𝐻𝑐𝑘 (𝑌) → 𝐑 ,
and since 𝑌 is closed the inclusion map 𝜄𝑌 of 𝑌 into 𝑋 is proper, so we get from it a pullback
operation on cohomology
♯
𝜄𝑌 ∶ 𝐻𝑐𝑘 (𝑋) → 𝐻𝑐𝑘 (𝑌)
♯
and by composing these two maps, we get a linear map ℓ𝑌 = 𝐼𝑌 ∘ 𝜄𝑌 ∶ 𝐻𝑐𝑘 (𝑋) → 𝐑. The
cohomology class
𝑇𝑌 ≔ 𝐿−1 𝑘
𝐵 (ℓ𝑌 ) ∈ 𝐻𝑐 (𝑋)
associated with ℓ𝑌 is called the Thom class of the manifold 𝑌 and has the defining property
♯
(5.5.4) 𝐵(𝑇𝑌 , 𝑐) = 𝐼𝑌 (𝜄𝑌 𝑐)
for 𝑐 ∈ 𝐻𝑐𝑘 (𝑋). Let us see what this defining property looks like at the level of forms. Let
𝜏𝑌 ∈ 𝛺𝑛−𝑘 (𝑋) be a closed 𝑘-form representing 𝑇𝑌 . Then by (5.4.5), the formula (5.5.4) for
𝑐 = [𝜔], becomes the integral formula

(5.5.5) ∫ 𝜏𝑌 ∧ 𝜔 = ∫ 𝜄⋆𝑌 𝜔 .
𝑋 𝑌
𝑛−𝑘
In other words, for every closed form 𝜔 ∈ 𝛺𝑐 (𝑋),
the integral of 𝜔 over 𝑌 is equal to
the integral over 𝑋 of 𝜏𝑌 ∧ 𝜔. A closed form 𝜏𝑌 with this “reproducing” property is called a
Thom form for 𝑌. Note that if we add to 𝜏𝑌 an exact (𝑛 − 𝑘)-form 𝜇 ∈ 𝑑𝛺𝑛−𝑘−1 (𝑋), we get
 Draft: March 28, 2018

§5.5 Thom classes & intersection theory 177

another representative 𝜏𝑌 + 𝜇 of the cohomology class 𝑇𝑌 , and hence another form with this
reproducing property. Also, since equation (5.5.5) is a direct translation into form language
of equation (5.5.4), any closed (𝑛 − 𝑘)-form 𝜏𝑌 with the reproducing property (5.5.5) is a
representative of the cohomology class 𝑇𝑌 .
These remarks make sense as well for compactly supported cohomology. Suppose 𝑌 is
compact. Then from the inclusion map we get a pullback map
♯
𝜄𝑌 ∶ 𝐻𝑘 (𝑋) → 𝐻𝑘 (𝑌)
and since 𝑌 is compact, the integration operation 𝐼𝑌 is a map 𝐻𝑘 (𝑌) → 𝐑, so the composi-
tion of these two operations is a map,
ℓ𝑌 ∶ 𝐻𝑘 (𝑋) → 𝐑
which by (5.5.3) gets converted into a cohomology class
𝑇𝑌 = 𝐿−1 𝑛−𝑘
𝐵 (ℓ𝑌 ) ∈ 𝐻𝑐 (𝑋) .

Moreover, if 𝜏𝑌 ∈ 𝛺𝑛−𝑘
𝑐 (𝑋) is a closed form, it represents this cohomology class if and only
if it has the reproducing property

(5.5.6) ∫ 𝜏𝑌 ∧ 𝜔 = ∫ 𝜄⋆𝑌 𝜔
𝑋 𝑌
𝑛−𝑘
for closed forms 𝜔 ∈ 𝛺 (𝑋). (There’s a subtle difference, however, between formula (5.5.5)
and formula (5.5.6). In (5.5.5) 𝜔 has to be closed and compactly supported and in (5.5.6) it
just has to be closed.)
As above we have a lot of latitude in our choice of 𝜏𝑌 : we can add to it any element of
𝑑𝛺𝑛−𝑘−1
𝑐 (𝑋). One consequence of this is the following.
Theorem 5.5.7. Given a neighborhood 𝑈 of 𝑌 in 𝑋, there exists a closed form 𝜏𝑌 ∈ 𝛺𝑛−𝑘
𝑐 (𝑈)
with the reproducing property

(5.5.8) ∫ 𝜏𝑌 ∧ 𝜔 = ∫ 𝜄⋆𝑌 𝜔
𝑈 𝑌
𝑘
for all closed forms 𝜔 ∈ 𝛺 (𝑈).
Hence, in particular, 𝜏𝑌 has the reproducing property (5.5.6) for all closed forms 𝜔 ∈
𝛺𝑛−𝑘 (𝑋). This result shows that the Thom form 𝜏𝑌 can be chosen to have support in an
arbitrarily small neighborhood of 𝑌.
Proof of Theorem 5.5.7. By Theorem 5.3.9 we can assume that 𝑈 has finite topology and
hence, in our definition of 𝜏𝑌 , we can replace the manifold 𝑋 by the open submanifold 𝑈.
This gives us a Thom form 𝜏𝑌 with support in 𝑈 and with the reproducing property (5.5.8)
for closed forms 𝜔 ∈ 𝛺𝑛−𝑘 (𝑈). □
Let us see what Thom forms actually look like in concrete examples. Suppose 𝑌 is de-
fined globally by a system of ℓ independent equations, i.e., suppose there exists an open
neighborhood 𝑂 of 𝑌 in 𝑋, a 𝐶∞ map 𝑓 ∶ 𝑂 → 𝐑ℓ , and a bounded open convex neighbor-
hood 𝑉 of the origin in 𝐑𝑛 satisfying the following properties.
Properties 5.5.9.
(1) The origin is a regular value of 𝑓.
(2) 𝑓−1 (𝑉) is closed in 𝑋.
(3) 𝑌 = 𝑓−1 (0).
 Draft: March 28, 2018

178 Chapter 5: Cohomology via forms

Then by (1) and (3) 𝑌 is a closed submanifold of 𝑂 and by (2) it’s a closed submanifold
of 𝑋. Moreover, it has a natural orientation: For every 𝑝 ∈ 𝑌 the map
𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇0 𝐑ℓ
is surjective, and its kernel is 𝑇𝑝 𝑌, so from the standard orientation of 𝑇0 𝐑ℓ one gets an
orientation of the quotient space
𝑇𝑝 𝑋/𝑇𝑝 𝑌 ,
and hence since 𝑇𝑝 𝑋 is oriented, one gets by Theorem 1.9.9 an orientation on 𝑇𝑝 𝑌. (See
Example 4.4.5.) Now let 𝜇 be an element of 𝛺ℓ𝑐 (𝑋). Then 𝑓⋆ 𝜇 is supported in 𝑓−1 (𝑉) and
hence by property (2) we can extend it to 𝑋 by setting it equal to zero outside 𝑂. We will
prove:
Theorem 5.5.10. If ∫𝑉 𝜇 = 1, then 𝑓⋆ 𝜇 is a Thom form for 𝑌.
To prove this we’ll first prove that if 𝑓⋆ 𝜇 has property (5.5.5) for some choice of 𝜇 it has
this property for every choice of 𝜇.
Lemma 5.5.11. Let 𝜇1 and 𝜇2 be forms in 𝛺ℓ𝑐 (𝑉) such that ∫𝑉 𝜇1 = ∫𝑉 𝜇2 = 1. Then for every
closed 𝑘-form 𝜈 ∈ 𝛺𝑘𝑐 (𝑋) we have

∫ 𝑓⋆ 𝜇1 ∧ 𝜈 = ∫ 𝑓⋆ 𝜇2 ∧ 𝜈 .
𝑋 𝑋

Proof. By Theorem 3.2.2, 𝜇1 − 𝜇2 = 𝑑𝛽 for some 𝛽 ∈ 𝛺ℓ−1

𝑐 (𝑉), hence, since 𝑑𝜈 = 0
(𝑓⋆ 𝜇1 − 𝑓⋆ 𝜇2 ) ∧ 𝜈 = 𝑑𝑓⋆ 𝛽 ∧ 𝜈 = 𝑑(𝑓⋆ 𝛽 ∧ 𝜈) .
Therefore, by Stokes theorem, the integral over 𝑋 of the expression on the left is zero. □
Now suppose 𝜇 = 𝜌(𝑥1 , …, 𝑥ℓ )𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥ℓ , for 𝜌 in 𝐶0∞ (𝑉). For 𝑡 ≤ 1 let
𝑥 𝑥
𝜇𝑡 = 𝑡ℓ 𝜌 ( 1 , …, ℓ ) 𝑑𝑥1 ∧ ⋯ 𝑑𝑥ℓ .
𝑡 𝑡
This form is supported in the convex set 𝑡𝑉, so by Lemma 5.5.11

∫ 𝑓⋆ 𝜇𝑡 ∧ 𝜈 = ∫ 𝑓⋆ 𝜇 ∧ 𝜈
𝑋 𝑋
for all closed forms 𝜈 ∈ 𝛺𝑘𝑐 (𝑋). Hence to prove that 𝑓⋆ 𝜇 has the property (5.5.5), it suffices
to prove that

(5.5.12) lim ∫ 𝑓⋆ 𝜇𝑡 ∧ 𝜈 = ∫ 𝜄⋆𝑌 𝜈 .

𝑡→0 𝑌
We prove this by proving a stronger result.
Lemma 5.5.13. The assertion (5.5.12) is true for every 𝑘-form 𝜈 ∈ 𝛺𝑘𝑐 (𝑋).
Proof. The canonical sumbersion theorem (see Theorem b.17) says that for every 𝑝 ∈ 𝑌
there exists a neighborhood 𝑈𝑝 of 𝑝 in 𝑌, a neighborhood, 𝑊 of 0 in 𝐑𝑛 , and an orientation
preserving diffeomorphism 𝜓 ∶ (𝑊, 0) → (𝑈𝑝 , 𝑝) such that
(5.5.14) 𝑓∘𝜓=𝜋
𝑛 ℓ
where 𝜋 ∶ 𝐑 → 𝐑 is the canonical submersion, 𝜋(𝑥1 , …, 𝑥𝑛 ) = (𝑥1 , …, 𝑥ℓ ). Let U be the
cover of 𝑂 by the open sets, 𝑂 ∖ 𝑌 and the 𝑈𝑝 ’s. Choosing a partition of unity subordinate
to this cover it suffices to verify (5.5.12) for 𝜈 in 𝛺𝑘𝑐 (𝑂 ∖ 𝑌) and 𝜈 in 𝛺𝑘𝑐 (𝑈𝑝 ). Let us first
suppose 𝜈 is in 𝛺𝑘𝑐 (𝑂 ∖ 𝑌). Then 𝑓(supp 𝜈) is a compact subset of 𝐑ℓ ∖ {0} and hence for 𝑡
 Draft: March 28, 2018

§5.5 Thom classes & intersection theory 179

small 𝑓(supp 𝜈) is disjoint from 𝑡𝑉, and both sides of (5.5.12) are zero. Next suppose that 𝜈
is in 𝛺𝑘𝑐 (𝑈𝑝 ). Then 𝜓⋆ 𝜈 is a compactly supported 𝑘-form on 𝑊 so we can write it as a sum
𝜓⋆ 𝜈 = ∑ ℎ𝐼 (𝑥)𝑑𝑥𝐼 , ℎ𝐼 ∈ 𝐶0∞ (𝑊)
𝐼
the 𝐼’s being strictly increasing multi-indices of length 𝑘. Let 𝐼0 = (ℓ + 1, ℓ2 + 2, …, 𝑛). Then
𝑥 𝑥
(5.5.15) 𝜋⋆ 𝜇𝑡 ∧ 𝜓⋆ 𝜈 = 𝑡ℓ 𝜌 ( 1 , …, ℓ ) ℎ𝐼0 (𝑥1 , …, 𝑥𝑛 )𝑑𝑥𝑟 ∧ ⋯ 𝑑𝑥𝑛
𝑡 𝑡
and by (5.5.14)
𝜓⋆ (𝑓⋆ 𝜇𝑡 ∧ 𝜈) = 𝜋⋆ 𝜇𝑡 ∧ 𝜓⋆ 𝜈
and hence since 𝜓 is orientation preserving
𝑥 𝑥
∫ 𝑓⋆ 𝜇𝑡 ∧ 𝜈 = ∫ 𝑓⋆ 𝜇𝑡 ∧ 𝜈 = 𝑡ℓ ∫ 𝜌 ( 1 , …, ℓ ) ℎ𝐼0 (𝑥1 , …, 𝑥𝑛 )𝑑𝑥
𝑋 𝑈𝑝 𝐑𝑛 𝑡 𝑡

= ∫ 𝜌(𝑥1 , …, 𝑥ℓ )ℎ𝐼0 (𝑡𝑥1 , …, 𝑡𝑥ℓ , 𝑥ℓ+1 , …, 𝑥𝑛 )𝑑𝑥

𝐑𝑛
and the limit of this expression as 𝑡 tends to zero is

∫ 𝜌(𝑥1 , …, 𝑥ℓ )ℎ𝐼0 (0, …, 0, 𝑥ℓ+1 , …, 𝑥𝑛 )𝑑𝑥1 ⋯ 𝑑𝑥𝑛

or
(5.5.16) ∫ ℎ𝐼 (0, …, 0, 𝑥ℓ+1 , …, 𝑥𝑛 ) 𝑑𝑥ℓ+1 ⋯ 𝑑𝑥𝑛 .

This, however, is just the integral of 𝜓⋆ 𝜈 over the set 𝜋−1 (0)∩𝑊. By (5.5.12), 𝜓 maps 𝜋−1 (0)∩
𝑊 diffeomorphically onto 𝑌 ∩ 𝑈𝑝 and by our recipe for orienting 𝑌 this diffeomorphism is
an orientation-preserving diffeomorphism, so the integral (5.5.16) is equal to ∫𝑌 𝜈. □
We’ll now describe some applications of Thom forms to topological intersection theory.
Let 𝑌 and 𝑍 be closed, oriented submanifolds of 𝑋 of dimensions 𝑘 and ℓ where 𝑘 + ℓ = 𝑛,
and Let us assume one of them (say 𝑍) is compact. We will show below how to define an
intersection number 𝐼(𝑌, 𝑍), which on the one hand will be a topological invariant of 𝑌
and 𝑍 and on the other hand will actually count, with appropriate ±-signs, the number of
points of intersection of 𝑌 and 𝑍 when they intersect non-tangentially. (Thus this notion is
similar to the notion of the degree deg(𝑓) for a 𝐶∞ mapping 𝑓. On the one hand deg(𝑓)
is a topological invariant of 𝑓. It’s unchanged if we deform 𝑓 by a homotopy. On the other
hand if 𝑞 is a regular value of𝑓, then deg(𝑓) counts with appropriate ±-signs the number of
points in the set 𝑓−1 (𝑞).)
We’ll first give the topological definition of this intersection number. This is by the for-
mula
(5.5.17) 𝐼(𝑌, 𝑍) = 𝐵(𝑇𝑌 , 𝑇𝑍 )
ℓ
where 𝑇𝑌 ∈ 𝐻 (𝑋) and 𝑇𝑍 ∈ and 𝐵 is the bilinear pairing (5.4.5). If 𝜏𝑌 ∈ 𝛺ℓ (𝑋)
𝐻𝑐𝑘 (𝑋)
and 𝜏𝑍 ∈ 𝛺𝑘𝑐 (𝑋) are Thom forms representing 𝑇𝑌 and 𝑇𝑍 , (5.5.17) can also be defined as the
integral
𝐼(𝑌, 𝑍) = ∫ 𝜏𝑌 ∧ 𝜏𝑍
𝑋
or by (5.5.8), as the integral over 𝑌,

(5.5.18) 𝐼(𝑌, 𝑍) = ∫ 𝜄⋆𝑌 𝜏𝑍

𝑌
 Draft: March 28, 2018

180 Chapter 5: Cohomology via forms

or, since 𝜏𝑌 ∧ 𝜏𝑍 = (−1)𝑘ℓ 𝜏𝑍 ∧ 𝜏𝑌 , as the integral over 𝑍

𝐼(𝑋, 𝑌) = (−1)𝑘ℓ ∫ 𝜄𝑍⋆ 𝜏𝑌 .

𝑍
In particular
𝐼(𝑌, 𝑍) = (−1)𝑘ℓ 𝐼(𝑍, 𝑌).
As a test case for our declaring 𝐼(𝑌, 𝑍) to be the intersection number of 𝑌 and 𝑍 we will first
prove:
Proposition 5.5.19. If 𝑌 and 𝑍 do not intersect, then 𝐼(𝑌, 𝑍) = 0.
Proof. If 𝑌 and 𝑍 don’t intersect then, since 𝑌 is closed, 𝑈 = 𝑋 ∖ 𝑌 is an open neighborhood
of 𝑍 in 𝑋, therefore since 𝑍 is compact there exists by Theorem 5.5.7 a Thom form 𝜏𝑍 ∈
𝛺ℓ𝑐 (𝑈). Thus 𝜄⋆𝑌 𝜏𝑍 = 0, and so by (5.5.18) we have 𝐼(𝑌, 𝑍) = 0. □
We’ll next indicate how one computes 𝐼(𝑌, 𝑍) when 𝑌 and 𝑍 intersect “non-tangentially”,
or, to use terminology more in current usage, when their intersection is transversal. Recall
that at a point of intersection, 𝑝 ∈ 𝑌 ∩ 𝑍, 𝑇𝑝 𝑌 and 𝑇𝑝 𝑍 are vector subspaces of 𝑇𝑝 𝑋.
Definition 5.5.20. 𝑌 and 𝑍 intersect transversally if for every 𝑝 ∈ 𝑌 ∩ 𝑍 we have
𝑇𝑝 𝑌 ∩ 𝑇𝑝 𝑍 = 0 .
Since 𝑛 = 𝑘 + ℓ = dim 𝑇𝑝 𝑌 + dim 𝑇𝑝 𝑍 = dim 𝑇𝑝 𝑋, this condition is equivalent to
(5.5.21) 𝑇𝑝 𝑋 = 𝑇𝑝 𝑌 ⊕ 𝑇𝑝 𝑍 ,
i.e., every vector, 𝑢 ∈ 𝑇𝑝 𝑋, can be written uniquely as a sum, 𝑢 = 𝑣 + 𝑤, with 𝑣 ∈ 𝑇𝑝 𝑌 and
𝑤 ∈ 𝑇𝑝 𝑍. Since 𝑋, 𝑌 and 𝑍 are oriented, their tangent spaces at 𝑝 are oriented, and we’ll
say that these spaces are compatibly oriented if the orientations of the two sides of (5.5.21)
agree. (In other words if 𝑣1 , …, 𝑣𝑘 is an oriented basis of 𝑇𝑝 𝑌 and 𝑤1 , …, 𝑤ℓ is an oriented
basis of 𝑇𝑝 𝑍, the 𝑛 vectors, 𝑣1 , …, 𝑣𝑘 , 𝑤1 , …, 𝑤ℓ , are an oriented basis of 𝑇𝑝 𝑋.) We will define
the local intersection number 𝐼𝑝 (𝑌, 𝑍) of 𝑌 and 𝑍 at 𝑝 to be equal to +1 if 𝑋, 𝑌, and 𝑍 are
compatibly oriented at 𝑝 and to be equal to −1 if they’re not. With this notation we’ll prove:
Theorem 5.5.22. If 𝑌 and 𝑍 intersect transversally then 𝑌 ∩ 𝑍 is a finite set and
𝐼(𝑌, 𝑍) = ∑ 𝐼𝑝 (𝑌, 𝑍) .
𝑝∈𝑌∩𝑍

To prove this we first need to show that transverse intersections look nice locally.
Theorem 5.5.23. If 𝑌 and 𝑍 intersect transversally, then for every 𝑝 ∈ 𝑌 ∩ 𝑍, there exists
an open neighborhood 𝑉𝑝 of 𝑝 in 𝑋, an open neighborhood 𝑈𝑝 of the origin in 𝐑𝑛 , and an
orientation-preserving diffeomorphism 𝜓𝑝 ∶ 𝑉𝑝 ⥲ 𝑈𝑝 which maps 𝑉𝑝 ∩ 𝑌 diffeomorphically
onto the subset of 𝑈𝑝 defined by the equations: 𝑥1 = ⋯ = 𝑥ℓ = 0, and maps 𝑉 ∩ 𝑍 onto the
subset of 𝑈𝑝 defined by the equations: 𝑥ℓ+1 = ⋯ = 𝑥𝑛 = 0.
Proof. Since this result is a local result, we can assume that 𝑋 = 𝐑𝑛 and hence by Theo-
rem 4.2.15 that there exists a neighborhood 𝑉𝑝 of 𝑝 in 𝐑𝑛 and submersions 𝑓 ∶ (𝑉𝑝 , 𝑝) →
(𝐑ℓ , 0) and 𝑔 ∶ (𝑉𝑝 , 𝑝) → (𝐑𝑘 , 0) with the properties
(5.5.24) 𝑉𝑝 ∩ 𝑌 = 𝑓−1 (0)
and
(5.5.25) 𝑉𝑝 ∩ 𝑍 = 𝑔−1 (0) .
 Draft: March 28, 2018

§5.5 Thom classes & intersection theory 181

Moreover, by (4.3.5) 𝑇𝑝 𝑌 = (𝑑𝑓𝑝 )−1 (0) and 𝑇𝑝 𝑍 = (𝑑𝑔𝑝 )−1 (0). Hence by (5.5.21), the equa-
tions
(5.5.26) 𝑑𝑓𝑝 (𝑣) = 𝑑𝑔𝑝 (𝑣) = 0
for 𝑣 ∈ 𝑇𝑝 𝑋 imply that 𝑣 = 0. Now let 𝜓𝑝 ∶ 𝑉𝑝 → 𝐑𝑛 be the map
(𝑓, 𝑔) ∶ 𝑉𝑝 → 𝐑ℓ × 𝐑𝑘 = 𝐑𝑛 .
Then by (5.5.26), 𝑑𝜓𝑝 is bijective, therefore, shrinking 𝑉𝑝 if necessary, we can assume that
𝜓𝑝 maps 𝑉𝑝 diffeomorphically onto a neighborhood 𝑈𝑝 of the origin in 𝐑𝑛 , and hence by
(5.5.24) and (5.5.25)), 𝜓𝑝 maps 𝑉𝑝 ∩ 𝑌 onto the set: 𝑥1 = ⋯ = 𝑥ℓ = 0 and maps 𝑉𝑝 ∩ 𝑍 onto
the set: 𝑥ℓ+1 = ⋯ = 𝑥𝑛 = 0. Finally, if 𝜓 isn’t orientation preserving, we can make it so by
composing it with the involution, (𝑥1 , …, 𝑥𝑛 ) ↦ (𝑥1 , 𝑥2 , …, 𝑥𝑛−1 , −𝑥𝑛 ). □
From this result we deduce:
Theorem 5.5.27. If 𝑌 and 𝑍 intersect transversally, their intersection is a finite set.
Proof. By Theorem 5.5.23 the only point of intersection in 𝑉𝑝 is 𝑝 itself. Moreover, since 𝑌
is closed and 𝑍 is compact, 𝑌 ∩ 𝑍 is compact. Therefore, since the 𝑉𝑝 ’s cover 𝑌 ∩ 𝑍 we can
extract a finite subcover by the Heine–Borel theorem. However, since no two 𝑉𝑝 ’s cover the
same point of 𝑌 ∩ 𝑍, this cover must already be a finite subcover. □
We will now prove Theorem 5.5.22.
Proof of Theorem 5.5.22. Since 𝑌 is closed, the map 𝜄𝑌 ∶ 𝑌 ↪ 𝑋 is proper, so by Theo-
rem 3.4.7 there exists a neighborhood 𝑈 of 𝑍 in 𝑋 such that 𝑈 ∩ 𝑌 is contained in the union
of the open sets 𝑉𝑝 above. Moreover by Theorem 5.5.7 we can choose 𝜏𝑍 to be supported
in 𝑈 and by Theorem 5.3.2 we can assume that 𝑈 has finite topology, so we’re reduced to
proving the theorem with 𝑋 replaced by 𝑈 and 𝑌 replaced by 𝑌 ∩ 𝑈. Let
𝑂 ≔ 𝑈 ∩ ⋃ 𝑉𝑝 ,
𝑝∈𝑋
ℓ
let 𝑓 ∶ 𝑂 → 𝐑 be the map whose restriction to 𝑉𝑝 ∩ 𝑈 is 𝜋 ∘ 𝜓𝑝 where 𝜋 is, as in equa-
tion (5.5.14), the canonical submersion of 𝐑𝑛 onto 𝐑ℓ , and finally let 𝑉 be a bounded con-
vex neighborhood of 𝐑ℓ , whose closure is contained in the intersection of the open sets,
𝜋 ∘ 𝜓𝑝 (𝑉𝑝 ∩ 𝑈). Then 𝑓−1 (𝑉) is a closed subset of 𝑈, so if we replace 𝑋 by 𝑈 and 𝑌 by
𝑌 ∩ 𝑈, the data (𝑓, 𝑂, 𝑉) satisfy Properties 5.5.9. Thus to prove Theorem 5.5.22 it suffices by
Theorem 5.5.10 to prove this theorem with
𝜏𝑌 = 𝜎𝑝 (𝑌)𝑓⋆ 𝜇
on 𝑉𝑝 ∩ 𝑂 where 𝜎𝑝 (𝑌) = +1 or −1 depending on whether the orientation of 𝑌 ∩ 𝑉𝑝 in
Theorem 5.5.10 coincides with the given orientation of 𝑌 or not. Thus
𝐼(𝑌, 𝑍) = (−1)𝑘ℓ 𝐼(𝑍, 𝑌)

= (−1)𝑘ℓ ∑ 𝜎𝑝 (𝑌) ∫ 𝜄𝑍⋆ 𝑓⋆ 𝜇

𝑝∈𝑋 𝑍

= (−1)𝑘ℓ ∑ 𝜎𝑝 (𝑌) ∫ 𝜄𝑍⋆ 𝜓𝑝⋆ 𝜋⋆ 𝜇

𝑝∈𝑋 𝑍

= ∑ (−1)𝑘ℓ 𝜎𝑝 (𝑌) ∫ (𝜋 ∘ 𝜓𝑝 ∘ 𝜄𝑍 )⋆ 𝜇 .
𝑝∈𝑋 𝑍∩𝑉𝑝
 Draft: March 28, 2018

182 Chapter 5: Cohomology via forms

But 𝜋 ∘ 𝜓𝑝 ∘ 𝜄𝑍 maps an open neighborhood of 𝑝 in 𝑈𝑝 ∩ 𝑍 diffeomorphically onto 𝑉, and

𝜇 is compactly supported in 𝑉, so since ∫𝑉 𝜇 = 1,

∫ (𝜋 ∘ 𝜓𝑝 ∘ 𝜄𝑍 )⋆ 𝜇 = 𝜎𝑝 (𝑍) ∫ 𝜇 = 𝜎𝑝 (𝑍) ,
𝑍∩𝑈𝑝 𝑉

where 𝜎𝑝 (𝑍) = +1 or −1 depending on whether 𝜋 ∘ 𝜓𝑝 ∘ 𝜄𝑍 is orientation preserving or not.

Thus finally
𝐼(𝑌, 𝑍) = ∑ (−1)𝑘ℓ 𝜎𝑝 (𝑌)𝜎𝑝 (𝑍) .
𝑝∈𝑋

We will leave as an exercise the task of unraveling these orientations and showing that
(−1)𝑘ℓ 𝜎𝑝 (𝑌)𝜎𝑝 (𝑍) = 𝐼𝑝 (𝑌, 𝑍)
and hence that 𝐼(𝑌, 𝑍) = ∑𝑝∈𝑋 𝐼𝑝 (𝑌, 𝑍). □

Exercises for §5.5

Exercise 5.5.i. Let 𝑋 be a connected oriented 𝑛-manifold, 𝑊 a connected oriented ℓ-dimensional
manifold, 𝑓 ∶ 𝑋 → 𝑊 a 𝐶∞ map, and 𝑌 a closed submanifold of 𝑋 of dimension 𝑘 ≔ 𝑛 − ℓ.
Suppose 𝑌 is a “level set” of the map 𝑓, i.e., suppose that 𝑞 is a regular value of 𝑓 and that
𝑌 = 𝑓−1 (𝑞). Show that if 𝜇 is in 𝛺ℓ𝑐 (𝑍) and its integral over 𝑍 is 1, then one can orient 𝑌 so
that 𝜏𝑌 = 𝑓⋆ 𝜇 is a Thom form for 𝑌.
Hint: Theorem 5.5.10.
Exercise 5.5.ii. In Exercise 5.5.i show that if 𝑍 ⊂ 𝑋 is a compact oriented ℓ-dimensional
submanifold of 𝑋 then
𝐼(𝑌, 𝑍) = (−1)𝑘ℓ deg(𝑓 ∘ 𝜄𝑍 ) .
Exercise 5.5.iii. Let 𝑞1 be another regular value of the map 𝑓 ∶ 𝑋 → 𝑊 and let 𝑌1 = 𝑓−1 (𝑞).
Show that
𝐼(𝑌, 𝑍) = 𝐼(𝑌1 , 𝑍) .
Exercise 5.5.iv.
(1) Show that if 𝑞 is a regular value of the map 𝑓 ∘ 𝜄𝑍 ∶ 𝑍 → 𝑊, then 𝑍 and 𝑌 intersect
transversally.
(2) Show that this is an “if and only if ” proposition: If 𝑌 and 𝑍 intersect transversally, then
𝑞 is a regular value of the map 𝑓 ∘ 𝜄𝑍 .
Exercise 5.5.v. Suppose 𝑞 is a regular value of the map 𝑓 ∘ 𝜄𝑍 . Show that 𝑝 is in 𝑌 ∩ 𝑍 if and
only if 𝑝 is in the preimage (𝑓 ∘ 𝜄𝑍 )−1 (𝑞) of 𝑞 and that
𝐼𝑝 (𝑋, 𝑌) = (−1)𝑘ℓ 𝜎𝑝
where 𝜎𝑝 is the orientation number of the map 𝑓 ∘ 𝜄𝑍 , at 𝑝, i.e., 𝜎𝑝 = 1 if 𝑓 ∘ 𝜄𝑍 is orientation-
preserving at 𝑝 and 𝜎𝑝 = −1 if 𝑓 ∘ 𝜄𝑍 is orientation-reversving at 𝑝.
Exercise 5.5.vi. Suppose the map 𝑓 ∶ 𝑋 → 𝑊 is proper. Show that there exists a neighbor-
hood, 𝑉, of 𝑞 in 𝑊 having the property that all points of 𝑉 are regular values of 𝑓.
Hint: Since 𝑞 is a regular value of 𝑓 there exists, for every 𝑝 ∈ 𝑓−1 (𝑞) a neighborhood
𝑈𝑝 of 𝑝 on which 𝑓 is a submersion. Conclude, by Theorem 3.4.7, that there exists a neigh-
borhood 𝑉 of 𝑞 with 𝑓−1 (𝑉) ⊂ ⋃𝑝∈𝑓−1 (𝑞) 𝑈𝑝 .
 Draft: March 28, 2018

§5.6 The Lefschetz Theorem 183

Exercise 5.5.vii. Show that in every neighborhood 𝑉1 of 𝑞 in 𝑉 there exists a point 𝑞1 whose
preimage
𝑌1 ≔ 𝑓−1 (𝑞1 )
intersects 𝑍 transversally. Conclude that one can “deform 𝑌 an arbitrarily small amount so
that it intersects 𝑍 transversally”.
Hint: Exercise 5.5.iv plus Sard’s theorem.
Exercise 5.5.viii (Intersection theory for mappings). Let 𝑋 be an oriented, connected 𝑛-
manifold, 𝑍 a compact, oriented ℓ-dimensional submanifold, 𝑌 an oriented manifold of
dimension 𝑘 ≔ 𝑛 − ℓ and 𝑓 ∶ 𝑌 → 𝑋 a proper 𝐶∞ map. Define the intersection number of
𝑓 with 𝑍 to be the integral
𝐼(𝑓, 𝑍) ≔ ∫ 𝑓⋆ 𝜏𝑍 .
𝑌
(1) Show that 𝐼(𝑓, 𝑍) is a homotopy invariant of 𝑓, i.e., show that if 𝑓0 , 𝑓1 ∶ 𝑌 → 𝑋 are
proper 𝐶∞ maps and are properly homotopic, then
𝐼(𝑓0 , 𝑍) = 𝐼(𝑓1 , 𝑍) .
(2) Show that if 𝑌 is a closed submanifold of 𝑋 of dimension 𝑘 = 𝑛 − ℓ and 𝜄𝑌 ∶ 𝑌 → 𝑋 is
the inclusion map, then 𝐼(𝜄𝑌 , 𝑍) = 𝐼(𝑌, 𝑍).
Exercise 5.5.ix.
(1) Let 𝑋 be an oriented connected 𝑛-manifold and let 𝑍 be a compact zero-dimensional
submanifold consisting of a single point 𝑧0 ∈ 𝑋. Show that if 𝜇 is in 𝛺𝑛𝑐 (𝑋) then 𝜇 is a
Thom form for 𝑍 if and only if its integral is 1.
(2) Let 𝑌 be an oriented 𝑛-manifold and 𝑓 ∶ 𝑌 → 𝑋 a 𝐶∞ map. Show that for 𝑍 = {𝑧0 } as
in part (1) we have 𝐼(𝑓, 𝑍) = deg(𝑓).

5.6. The Lefschetz Theorem

In this section we’ll apply the intersection techniques that we developed in § 5.5 to a
concrete problem in dynamical systems: counting the number of fixed points of a differen-
tiable mapping. The Brouwer fixed point theorem, which we discussed in §3.6, told us that
a 𝐶∞ map of the unit ball into itself has to have at least one fixed point. The Lefschetz the-
orem is a similar result for manifolds. It will tell us that a 𝐶∞ map of a compact manifold
into itself has to have a fixed point if a certain topological invariant of the map, its global
Lefschetz number, is nonzero.
Before stating this result, we will first show how to translate the problem of counting
fixed points of a mapping into an intersection number problem. Let 𝑋 be an oriented com-
pact 𝑛-manifold and 𝑓 ∶ 𝑋 → 𝑋 a 𝐶∞ map. Define the graph of 𝑓 in 𝑋 × 𝑋 to be the
set
𝛤𝑓 ≔ { (𝑥, 𝑓(𝑥)) | 𝑥 ∈ 𝑋 } ⊂ 𝑋 × 𝑋 .
It’s easy to see that this is an 𝑛-dimensional submanifold of 𝑋 × 𝑋 and that this manifold is
diffeomorphic to 𝑋 itself. In fact, in one direction, there is a 𝐶∞ map
(5.6.1) 𝛾𝑓 ∶ 𝑋 → 𝛤𝑓 , 𝑥 ↦ (𝑥, 𝑓(𝑥)) ,
∞
and, in the other direction, a 𝐶 map
𝜋 ∶ 𝛤𝑓 → 𝑋 , (𝑥, 𝑓(𝑥)) ↦ 𝑥 ,
and it’s obvious that these maps are inverses of each other and hence diffeomorphisms. We
will orient 𝛤𝑓 by requiring that 𝛾𝑓 and 𝜋 be orientation-preserving diffeomorphisms.
 Draft: March 28, 2018

184 Chapter 5: Cohomology via forms

An example of a graph is the graph of the identity map of 𝑋 onto itself. This is the
diagonal in 𝑋 × 𝑋
𝛥 ≔ { (𝑥, 𝑥) | 𝑥 ∈ 𝑋 } ⊂ 𝑋 × 𝑋
and its intersection with 𝛤𝑓 is the set
𝛤𝑓 ∩ 𝛥 = { (𝑥, 𝑥) | 𝑓(𝑥) = 𝑥 } ,
which is just the set of fixed points of 𝑓. Hence a natural way to count the fixed points of 𝑓 is
as the intersection number of 𝛤𝑓 and 𝛥 in 𝑋 × 𝑋. To do so we need these three manifolds to
be oriented, but, as we noted above, 𝛤𝑓 and 𝛥 acquire orientations from the identifications
(5.6.1) and, as for 𝑋×𝑋, we’ll give it its natural orientation as a product of oriented manifolds.
(See §4.5.)
Definition 5.6.2. The global Lefschetz number of 𝑓 is the intersection number
𝐿(𝑓) ≔ 𝐼(𝛤𝑓 , 𝛥) .
In this section we’ll give two recipes for computing this number: one by topological
methods and the other by making transversality assumptions and computing this number
as a sum of local intersection numbers via Theorem 5.5.22. We first show what one gets from
the transversality approach.
Definition 5.6.3. The map 𝑓 ∶ 𝑋 → 𝑋 is a Lefschetz map, or simply Lefschetz, if 𝛤𝑓 and 𝛥
intersect transversally.
Let us see what being Lefschetz entails. Suppose 𝑝 is a fixed point of 𝑓. Then at the point
𝑞 = (𝑝, 𝑝) of 𝛤𝑓
(5.6.4) 𝑇𝑞 (𝛤𝑓 ) = (𝑑𝛾𝑓 )𝑝 𝑇𝑝 𝑋 = { (𝑣, 𝑑𝑓𝑝 (𝑣)) | 𝑣 ∈ 𝑇𝑝 𝑋 }
and, in particular, for the identity map,
𝑇𝑞 (𝛥) = { (𝑣, 𝑣) | 𝑣 ∈ 𝑇𝑝 𝑋 } .
Therefore, if 𝛥 and 𝛤𝑓 are to intersect transversally, the intersection of 𝑇𝑞 (𝛤𝑓 ) ∩ 𝑇𝑞 (𝛥) inside
𝑇𝑞 (𝑋 × 𝑋) has to be the zero space. In other words if
(5.6.5) (𝑣, 𝑑𝑓𝑝 (𝑣)) = (𝑣, 𝑣)
then 𝑣 = 0. But the identity (5.6.5) says that 𝑣 is a fixed point of 𝑑𝑓𝑝 , so transversality at 𝑝
amounts to the assertion
𝑑𝑓𝑝 (𝑣) = 𝑣 ⟺ 𝑣 = 0 ,
or in other words the assertion that the map
(5.6.6) (id𝑇𝑝 𝑋 −𝑑𝑓𝑝 ) ∶ 𝑇𝑝 𝑋 → 𝑇𝑝 𝑋
is bijective.
Proposition 5.6.7. The local intersection number 𝐼𝑝 (𝛤𝑓 , 𝛥) is 1 if (5.6.6) is orientation-preserving
and −1 otherwise.
In other words 𝐼𝑝 (𝛤𝑓 , 𝛥) is the sign of det(id𝑇𝑝 𝑋 −𝑑𝑓𝑝 ).

Proof. To prove this let 𝑒1 , …, 𝑒𝑛 be an oriented basis of 𝑇𝑝 𝑋 and let

𝑛
(5.6.8) 𝑑𝑓𝑝 (𝑒𝑖 ) = ∑ 𝑎𝑗,𝑖 𝑒𝑗 .
𝑗=1
 Draft: March 28, 2018

§5.6 The Lefschetz Theorem 185

Now set 𝑣𝑖 ≔ (𝑒𝑖 , 0) ∈ 𝑇𝑞 (𝑋 × 𝑋) and 𝑤𝑖 ≔ (0, 𝑒𝑖 ) ∈ 𝑇𝑞 (𝑋 × 𝑋). Then by the definition of

the product orientation on 𝑋 × 𝑋, we have that (𝑣1 , …, 𝑣𝑛 , 𝑤1 , …, 𝑤𝑛 ) is an oriented basis
of 𝑇𝑞 (𝑋 × 𝑋), and by (5.6.4)
𝑛 𝑛
(𝑣1 + ∑𝑗=1 𝑎𝑗,𝑖 𝑤𝑗 , …, 𝑣𝑛 + ∑𝑗=1 𝑎𝑗,𝑛 𝑤𝑗 )
is an oriented basis of 𝑇𝑞 𝛤𝑓 and 𝑣1 + 𝑤1 , …, 𝑣𝑛 + 𝑤𝑛 is an oriented basis of 𝑇𝑞 𝛥. Thus
𝐼𝑝 (𝛤𝑓 , 𝛥) = ±1, depending on whether or not the basis
𝑛 𝑛
(𝑣1 + ∑𝑗=1 𝑎𝑗,𝑖 𝑤𝑗 , …, 𝑣𝑛 + ∑𝑗=1 𝑎𝑗,𝑛 𝑤𝑗 , 𝑣1 + 𝑤1 , …, 𝑣𝑛 + 𝑤𝑛 )
of 𝑇𝑞 (𝑋 × 𝑋) is compatibly oriented with the basis (5.6.8). Thus 𝐼𝑝 (𝛤𝑓 , 𝛥) = ±1, the sign
depending on whether the determinant of the 2𝑛 × 2𝑛 matrix relating these two bases:
id𝑛 𝐴
𝑃≔( )
id𝑛 id𝑛
is positive or negative, where 𝐴 = (𝑎𝑖,𝑗 ). However, by the block matrix determinant formula,
we see that to see that
det(𝑃) = det(id𝑛 −𝐴 id−1
𝑛 id𝑛 ) det(id𝑛 ) = det(id𝑛 −𝐴) ,
hence by (5.6.8) we see that det(𝑃) = det(id𝑛 −𝑑𝑓𝑝 ). (If you are not familiar with this, it is
easy to see that by elementary row operations 𝑃 can be converted into the matrix
id𝑛 𝐴
( )
0 id𝑛 −𝐴
apply Exercise 1.8.vii.) □
Let us summarize what we have shown so far.
Theorem 5.6.9. A 𝑓 ∶ 𝑋 → 𝑋, is a Lefschetz map if and only if for every fixed point 𝑝 of 𝑓,
the map
id𝑇𝑝 𝑋 −𝑑𝑓𝑝 ∶ 𝑇𝑝 𝑋 → 𝑇𝑝 𝑋
is bijective. Moreover for Lefschetz maps 𝑓 we have
𝐿(𝑓) = ∑ 𝐿𝑝 (𝑓)
𝑝=𝑓(𝑝)

where 𝐿𝑝 (𝑓) = +1 if id𝑇𝑝 𝑋 −𝑑𝑓𝑝 is orientation-preserving and 𝐿𝑝 (𝑓) = −1 if id𝑇𝑝 𝑋 −𝑑𝑓𝑝 is

orientation-reversing.
We’ll next describe how to compute 𝐿(𝑓) as a topological invariant of 𝑓. Let 𝜄𝛤𝑓 be the
inclusion map of 𝛤𝑓 into 𝑋 × 𝑋 and let 𝑇𝛥 ∈ 𝐻𝑛 (𝑋 × 𝑋) be the Thom class of 𝛥. Then by
(5.5.18)
𝐿(𝑓) = 𝐼𝛤𝑓 (𝜄⋆𝛤𝑓 𝑇𝛥 )
and hence since the mapping 𝛾𝑓 ∶ 𝑋 → 𝑋×𝑋 defined by (5.6.1) is an orientation-preserving
diffeomorphism of 𝑋 ⥲ 𝛤𝑓 we have
(5.6.10) 𝐿(𝑓) = 𝐼𝑋 (𝛾𝑓⋆ 𝑇𝛥 ) .
To evaluate the expression on the right we’ll need to know some facts about the cohomology
groups of product manifolds. The main result on this topic is the Künneth theorem, and we’ll
take up the formulation and proof of this theorem in §5.7. First, however, we’ll describe a
result which follows from the Künneth theorem and which will enable us to complete our
computation of 𝐿(𝑓).
 Draft: March 28, 2018

186 Chapter 5: Cohomology via forms

Let 𝜋1 and 𝜋2 be the projection of 𝑋 × 𝑋 onto its first and second factors, i.e., for 𝑖 = 1, 2
let
𝜋𝑖 ∶ 𝑋 × 𝑋 → 𝑋
be the projection 𝜋𝑖 (𝑥1 , 𝑥2 ) ≔ 𝑥𝑖 . Then by equation (5.6.1) we have
(5.6.11) 𝜋1 ⋅ 𝛾𝑓 = id𝑋
and
(5.6.12) 𝜋2 ⋅ 𝛾𝑓 = 𝑓 .
𝑛
Lemma 5.6.13. If 𝜔1 and 𝜔2 are in 𝛺 (𝑋) then

(5.6.14) ∫ 𝜋⋆1 𝜔1 ∧ 𝜋⋆2 𝜔2 = (∫ 𝜔1 ) (∫ 𝜔2 ) .

𝑋×𝑋 𝑋 𝑋

Proof. By a partition of unity argument we can assume that 𝜔𝑖 has compact support in a
parametrizable open set, 𝑉𝑖 . Let 𝑈𝑖 be an open subset of 𝐑𝑛 and 𝜙𝑖 ∶ 𝑈𝑖 → 𝑉𝑖 an orientation-
preserving diffeomorphism. Then
𝜙⋆𝑖 𝜔 = 𝜌𝑖 𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛
with 𝜌𝑖 ∈ 𝐶0∞ (𝑈𝑖 ), so the right hand side of (5.6.14) is the product of integrals over 𝐑𝑛 :

(5.6.15) (∫ 𝜌1 (𝑥)𝑑𝑥) (∫ 𝜌2 (𝑥)𝑑𝑥) .

𝐑𝑛 𝐑𝑛
Moreover, since 𝑋 × 𝑋 is oriented by its product orientation, the map
𝜓 ∶ 𝑈1 × 𝑈2 → 𝑉1 × 𝑉2
is given by (𝑥, 𝑦) ↦ (𝜙1 (𝑥) 𝜙
, 2 (𝑦)) is an orientation-preserving diffeomorphism and since
𝜋𝑖 ∘ 𝜓 = 𝜙𝑖
𝜓⋆ (𝜋⋆1 𝜔1 ∧ 𝜋⋆2 𝜔2 ) = 𝜙⋆1 𝜔1 ∧ 𝜙⋆2 𝜔2
= 𝜌1 (𝑥)𝜌2 (𝑦)𝑑𝑥1 ∧ ⋯ ∧ 𝑑𝑥𝑛 ∧ 𝑑𝑦1 ∧ ⋯ ∧ 𝑑𝑦𝑛
and hence the left hand side of (5.6.14) is the integral over 𝐑2𝑛 of the function, 𝜌1 (𝑥)𝜌2 (𝑦),
and therefore, by integration by parts, is equal to the product (5.6.15). □
As a corollary of this lemma we get a product formula for cohomology classes.
Lemma 5.6.16. If 𝑐1 and 𝑐2 are in 𝐻𝑛 (𝑋) then
𝐼𝑋×𝑋 (𝜋⋆1 𝑐1 ⋅ 𝜋⋆2 𝑐2 ) = 𝐼𝑋 (𝑐1 )𝐼𝑋 (𝑐2 ) .
Now let 𝑑𝑘 ≔ dim 𝐻𝑘 (𝑋) and note that since 𝑋 is compact, Poincaré duality tells us
that 𝑑𝑘 = 𝑑ℓ when ℓ = 𝑛 − 𝑘. In fact it tells us even more. Let 𝜇𝑘1 , …, 𝜇𝑘𝑑𝑘 be a basis of 𝐻𝑘 (𝑋).
Then, since the pairing (5.4.5) is non-singular, there exists for ℓ = 𝑛 − 𝑘 a “dual” basis
𝜈𝑗ℓ , 𝑗 = 1, …, 𝑑ℓ
of 𝐻ℓ (𝑋) satisfying
(5.6.17) 𝐼𝑋 (𝜇𝑘𝑖 ⋅ 𝜈𝑗ℓ ) = 𝛿𝑖,𝑗 .
Lemma 5.6.18. The cohomology classes
♯ ♯
𝜋1 𝜈𝑟ℓ ⋅ 𝜋2 𝜇𝑘𝑠 , 𝑘 + ℓ = 𝑛
for 𝑘 = 0, …, 𝑛 and 1 ≤ 𝑟, 𝑠 ≤ 𝑑𝑘 , are a basis for 𝐻𝑛 (𝑋 × 𝑋).
 Draft: March 28, 2018

§5.6 The Lefschetz Theorem 187

This is the corollary of the Künneth theorem that we alluded to above (and whose proof
we’ll give in §5.7). Using these results we prove the following.
Theorem 5.6.19. The Thom class 𝑇𝛥 ∈ 𝐻𝑛 (𝑋 × 𝑋) is given explicitly by the formula
𝑑𝑘
♯ ♯
(5.6.20) 𝑇𝛥 = ∑ (−1)ℓ ∑ 𝜋1 𝜇𝑘𝑖 ⋅ 𝜋2 𝜈𝑖ℓ .
𝑘+ℓ=𝑛 𝑖=1

Proof. We have to check that for every cohomology class 𝑐 ∈ 𝐻𝑛 (𝑋×𝑋), the class 𝑇𝛥 defined
by (5.6.20) has the reproducing property
♯
(5.6.21) 𝐼𝑋×𝑋 (𝑇𝛥 ⋅ 𝑐) = 𝐼𝛥 (𝜄𝛥 𝑐)
where 𝜄𝛥 is the inclusion map of 𝛥 into 𝑋 × 𝑋. However the map
𝛾𝛥 ∶ 𝑋 → 𝑋 × 𝑋 , 𝑥 ↦ (𝑥, 𝑥)
is an orientation-preserving diffeomorphism of 𝑋 onto 𝛥, so it suffices to show that
♯
(5.6.22) 𝐼𝑋×𝑋 (𝑇𝛥 ⋅ 𝑐) = 𝐼𝑋 (𝛾𝛥 𝑐)
and by Lemma 5.6.18 it suffices to verify (5.6.22) for 𝑐’s of the form
♯ ♯
𝑐 = 𝜋1 𝜈𝑟ℓ ⋅ 𝜋2 𝜇𝑘𝑠 .
The product of this class with a typical summand of (5.6.20), for instance, the summand
′ ♯ ′ ♯ ′
(5.6.23) (−1)ℓ 𝜋1 𝜇𝑘𝑖 ⋅ 𝜋2 𝜈𝑖ℓ , 𝑘′ + ℓ′ = 𝑛 ,
is equal, up to sign to,
♯ ′ ♯ ′
𝜋1 𝜇𝑘𝑖 ⋅ 𝜈𝑟ℓ ⋅ 𝜋2 𝜇𝑘𝑠 ⋅ 𝜈𝑖ℓ .
Notice, however, that if 𝑘 ≠ 𝑘′ this product is zero: For 𝑘 < 𝑘′ , 𝑘′ + ℓ is greater than 𝑘 + ℓ
and hence greater than 𝑛. Therefore
′ ′
𝜇𝑘𝑖 ⋅ 𝜈𝑟ℓ ∈ 𝐻𝑘 +ℓ (𝑋)
′
is zero since 𝑋 is of dimension 𝑛, and for 𝑘 > 𝑘′ , ℓ′ is greater than ℓ and 𝜇𝑘𝑠 ⋅ 𝜈𝑖ℓ is zero for
the same reason. Thus in taking the product of 𝑇𝛥 with 𝑐 we can ignore all terms in the sum
except for the terms, 𝑘′ = 𝑘 and ℓ′ = ℓ. For these terms, the product of (5.6.23) with 𝑐 is
♯ ♯
(−1)𝑘ℓ 𝜋1 𝜇𝑘𝑖 ⋅ 𝜈𝑟ℓ ⋅ 𝜋2 𝜇𝑘3 ⋅ 𝜈𝑖ℓ .
2
(Exercise: Check this. Hint: (−1)ℓ (−1)ℓ = 1.) Thus
𝑑𝑘
♯ ♯
𝑇𝛥 ⋅ 𝑐 = (−1)𝑘ℓ ∑ 𝜋1 𝜇𝑘𝑖 ⋅ 𝜈𝑟ℓ ⋅ 𝜋2 𝜇𝑘𝑠 ⋅ 𝜈𝑖ℓ
𝑖=1

and hence by Lemma 5.6.13 and (5.6.17)

𝑑𝑘
𝐼𝑋×𝑋 (𝑇𝛥 ⋅ 𝑐) = (−1)𝑘ℓ ∑ 𝐼𝑋 (𝜇𝑘𝑖 ⋅ 𝜈𝑟ℓ )𝐼𝑋 (𝜇𝑘𝑠 ⋅ 𝜈𝑖ℓ )
𝑖=1
𝑑𝑘
= (−1)𝑘ℓ ∑ 𝛿𝑖,𝑟 𝛿𝑖,𝑠
𝑖=1
𝑘ℓ
= (−1) 𝛿𝑟,𝑠 .
 Draft: March 28, 2018

188 Chapter 5: Cohomology via forms

♯ ♯
On the other hand for 𝑐 = 𝜋1 𝜈𝑟ℓ ⋅ 𝜋2 𝜇𝑘𝑠
♯ ♯ ♯ ♯ ♯
𝛾𝛥 𝑐 = 𝛾𝛥 𝜋1 𝜈𝑟ℓ ⋅ 𝛾𝛥 𝜋2 𝜇𝑘𝑠
= (𝜋1 ⋅ 𝛾𝛥 )♯ 𝜈𝑟ℓ (𝜋2 ⋅ 𝛾𝛥 )♯ 𝜇𝑘𝑠
= 𝜈𝑟ℓ ⋅ 𝜇𝑘𝑠
since
𝜋1 ⋅ 𝜈𝛥 = 𝜋2 ⋅ 𝛾𝛥 = id𝑋 .
So
♯
𝐼𝑋 (𝛾𝛥 𝑐) = 𝐼𝑋 (𝜈𝑟ℓ ⋅ 𝜇𝑘𝑠 ) = (−1)𝑘ℓ 𝛿𝑟,𝑠
by (5.6.17). Thus the two sides of (5.6.21) are equal. □
We’re now in position to compute 𝐿(𝑓) , i.e., to compute the expression on 𝐼𝑋 (𝛾𝑓⋆ 𝑇𝛥 )
ℓ ℓ ℓ
the right hand side of (5.6.10). Since 𝜈1 , …, 𝜈𝑑ℓ is a basis of 𝐻 (𝑋) the linear mapping
(5.6.24) 𝑓♯ ∶ 𝐻ℓ (𝑋) → 𝐻ℓ (𝑋)
ℓ
can be described in terms of this basis by a matrix [𝑓𝑖,𝑗 ] with the defining property
𝑑ℓ
𝑓♯ 𝜈𝑖ℓ = ∑ 𝑓𝑗,𝑖
ℓ ℓ
𝜈𝑗 .
𝑗=1

Thus by equations (5.6.11), (5.6.12) and (5.6.20) we have

𝑑𝑘
♯ ♯ ♯ ♯
𝛾𝑓 𝑇𝛥 = 𝛾𝑓 ∑ (−1)ℓ ∑ 𝜋1 𝜇𝑘𝑖 ⋅ 𝜋2 𝜈𝑖ℓ
𝑘+ℓ=𝑛 𝑖=1
𝑑𝑘
= ∑ (−1)ℓ ∑(𝜋1 ⋅ 𝛾𝑓 )♯ 𝜇𝑘𝑖 ⋅ (𝜋2 ⋅ 𝜈𝑓 )♯ 𝜈𝑖ℓ
𝑘+ℓ=𝑛 𝑖=1
𝑑𝑘
= ∑ (−1)ℓ ∑ 𝜇𝑘𝑖 ⋅ 𝑓♯ 𝜈𝑖ℓ
𝑘+ℓ=𝑛 𝑖=1
𝑑𝑘
= ∑ (−1)ℓ ∑ 𝑓𝑗,𝑖
ℓ 𝑘
𝜇𝑖 ⋅ 𝜈𝑗ℓ .
𝑘+ℓ=𝑛 𝑖=1
Thus by (5.6.17)
♯
𝐼𝑋 (𝛾𝑓 𝑇𝛥 ) = ∑ (−1)ℓ ∑ 𝑓𝑗,𝑖
ℓ
𝐼𝑋 (𝜇𝑘𝑖 ⋅ 𝜈𝑗ℓ )
𝑘+ℓ=𝑛 1≤𝑖,𝑗≤𝑑ℓ
ℓ ℓ
= ∑ (−1) ∑ 𝑓𝑗,𝑖 𝛿𝑖,𝑗
𝑘+ℓ=𝑛 1≤𝑖,𝑗≤𝑑ℓ
𝑛 𝑑ℓ
= ∑ (−1)ℓ ∑ 𝑓𝑖,𝑖ℓ .
ℓ=0 𝑖=1
𝑑ℓ
But ∑𝑖=1 𝑓𝑖,𝑖ℓ
is just the trace of the linear mapping (5.6.24) (see Exercise 5.6.xii), so we end
up with the following purely topological prescription of 𝐿(𝑓).
Theorem 5.6.25. The Lefschetz number 𝐿(𝑓) is the alternating sum
𝑛
𝐿(𝑓) = ∑ (−1)ℓ tr(𝑓♯ | 𝐻ℓ (𝑋))
ℓ=0
 Draft: March 28, 2018

§5.6 The Lefschetz Theorem 189

where tr(𝑓♯ | 𝐻ℓ (𝑋)) is the trace of the map 𝑓♯ ∶ 𝐻ℓ (𝑋) → 𝐻ℓ (𝑋).

Exercises for §5.6
Exercise 5.6.i. Show that if 𝑓0 ∶ 𝑋 → 𝑋 and 𝑓1 ∶ 𝑋 → 𝑋 are homotopic 𝐶∞ mappings
𝐿(𝑓0 ) = 𝐿(𝑓1 ).
Exercise 5.6.ii.
(1) The Euler characteristic 𝜒(𝑋) of 𝑋 is defined to be the intersection number of the diag-
onal with itself in 𝑋 × 𝑋, i.e., the “self-intersection” number
𝜒(𝑋) ≔ 𝐼(𝛥, 𝛥) = 𝐼𝑋×𝑋 (𝑇𝛥 , 𝑇𝛥 ) .
∞
Show that if a 𝐶 map 𝑓 ∶ 𝑋 → 𝑋 is homotopic to the identity, then 𝐿(𝑓) = 𝜒(𝑋).
(2) Show that
𝑛
𝜒(𝑋) = ∑ (−1)ℓ dim 𝐻ℓ (𝑋) .
ℓ=0
(3) Show that if 𝑋 is an odd-dimensional (compact) manifold, then 𝜒(𝑋) = 0.
Exercise 5.6.iii.
(1) Let 𝑆𝑛 be the unit 𝑛-sphere in 𝐑𝑛+1 . Show that if 𝑔 ∶ 𝑆𝑛 → 𝑆𝑛 is a 𝐶∞ map, then
𝐿(𝑔) = 1 + (−1)𝑛 deg(𝑔) .
(2) Conclude that if deg(𝑔) ≠ (−1)𝑛+1 , then 𝑔 has to have a fixed point.
Exercise 5.6.iv. Let 𝑓 be a 𝐶∞ mapping of the closed unit ball 𝐵𝑛+1 into itself and let
𝑔 ∶ 𝑆𝑛 → 𝑆𝑛 be the restriction of 𝑓 to he boundary of 𝐵𝑛+1 . Show that if deg(𝑔) ≠ (−1)𝑛+1
then the fixed point of 𝑓 predicted by Brouwer’s theorem can be taken to be a point on the
boundary of 𝐵𝑛+1 .
Exercise 5.6.v.
(1) Show that if 𝑔 ∶ 𝑆𝑛 → 𝑆𝑛 is the antipodal map 𝑔(𝑥) ≔ −𝑥, then deg(𝑔) = (−1)𝑛+1 .
(2) Conclude that the result in Exercise 5.6.iv is sharp. Show that the map
𝑓 ∶ 𝐵𝑛+1 → 𝐵𝑛+1 , 𝑓(𝑥) = −𝑥 ,
has only one fixed point, namely the origin, and in particular has no fixed points on the
boundary.
Exercise 5.6.vi. Let 𝒗 be a vector field on a compact manifold 𝑋. Since 𝑋 is compact, 𝒗
generates a one-parameter group of diffeomorphisms
(5.6.26) 𝑓𝑡 ∶ 𝑋 → 𝑋 , − ∞ < 𝑡 < ∞ .
(1) Let 𝛴𝑡 be the set of fixed points of 𝑓𝑡 . Show that this set contains the set of zeros of 𝒗,
i.e., the points 𝑝 ∈ 𝑋 where 𝒗(𝑝) = 0.
(2) Suppose that for some 𝑡0 , 𝑓𝑡0 is Lefschetz. Show that for all 𝑡, 𝑓𝑡 maps 𝛴𝑡0 into itself.
(3) Show that for |𝑡| < 𝜀, 𝜀 small, the points of 𝛴𝑡0 are fixed points of 𝑓𝑡 .
(4) Conclude that 𝛴𝑡0 is equal to the set of zeros of 𝑣.
(5) In particular, conclude that for all 𝑡 the points of 𝛴𝑡0 are fixed points of 𝑓𝑡 .
Exercise 5.6.vii.
(1) Let 𝑉 be a finite dimensional vector space and
𝐹(𝑡) ∶ 𝑉 → 𝑉 , − ∞ < 𝑡 < ∞
𝑑𝐹
a one-parameter group of linear maps of 𝑉 onto itself. Let 𝐴 = 𝑑𝑡
(0) and show that
𝐹(𝑡) = exp 𝑡𝐴. (See Exercise 2.2.viii.)
 Draft: March 28, 2018

190 Chapter 5: Cohomology via forms

(2) Show that if id𝑉 −𝐹(𝑡0 ) ∶ 𝑉 → 𝑉 is bijective for some 𝑡0 , then 𝐴 ∶ 𝑉 → 𝑉 is bijective.
Hint: Show that if 𝐴𝑣 = 0 for some 𝑣 ∈ 𝑉 − {0}, 𝐹(𝑡)𝑣 = 𝑣.
Exercise 5.6.viii. Let 𝒗 be a vector field on a compact manifold 𝑋 and let (5.6.26) be the
one-parameter group of diffeomorphisms generated by 𝒗. If 𝒗(𝑝) = 0 then by part (1) of
Exercise 5.6.vi, 𝑝 is a fixed point of 𝑓𝑡 for all 𝑡.
(1) Show that
(𝑑𝑓𝑡 ) ∶ 𝑇𝑝 𝑋 → 𝑇𝑝 𝑋
is a one-parameter group of linear mappings of 𝑇𝑝 𝑋 onto itself.
(2) Conclude from Exercise 5.6.vii that there exists a linear map
(5.6.27) 𝐿𝒗 (𝑝) ∶ 𝑇𝑝 𝑋 → 𝑇𝑝 𝑋
with the property
exp 𝑡𝐿𝒗 (𝑝) = (𝑑𝑓𝑡 )𝑝 .
Exercise 5.6.ix. Suppose 𝑓𝑡0 is a Lefschetz map for some 𝑡0 . Let 𝑎 = 𝑡0 /𝑁 where 𝑁 is a
positive integer. Show that 𝑓𝑎 is a Lefschetz map.
Hints:
• Show that
𝑓𝑡0 = 𝑓𝑎 ∘ ⋯ ∘ 𝑓𝑎 = 𝑓𝑎𝑁
(i.e., 𝑓𝑎 composed with itself 𝑁 times).
• Show that if 𝑝 is a fixed point of 𝑓𝑎 , it is a fixed pointof 𝑓𝑡0 .
• Conclude from Exercise 5.6.vi that the fixed points of 𝑓𝑎 are the zeros of 𝑣.
• Show that if 𝑝 is a fixed point of 𝑓𝑎 ,
(𝑑𝑓𝑡0 )𝑝 = (𝑑𝑓𝑎 )𝑝𝑁 .
• Conclude that if (𝑑𝑓𝑎 )𝑝 𝑣 = 𝑣 for some 𝑣 ∈ 𝑇𝑝 𝑋 ∖ {0}, then (𝑑𝑓𝑡0 )𝑝 𝑣 = 𝑣.
Exercise 5.6.x. Show that for all 𝑡, 𝐿(𝑓𝑡 ) = 𝜒(𝑋).
Hint: Exercise 5.6.ii.
Exercise 5.6.xi (Hopf Theorem). A vector field 𝒗 on a compact manifold 𝑋 is a Lefschetz
vector field if for some 𝑡0 ∈ 𝐑 the map 𝑓𝑡0 is a Lefschetz map.
(1) Show that if 𝒗 is a Lefschetz vector field then it has a finite number of zeros and for each
zero 𝑝 the linear map (5.6.27) is bijective.
(2) For a zero 𝑝 of 𝒗 let 𝜎𝑝 (𝒗) = +1 if the map (5.6.27) is orientation-preserving and let
𝜎𝑝 (𝒗) = −1 if it’s orientation-reversing. Show that
𝜒(𝑋) = ∑ 𝜎𝑝 (𝒗) .
𝒗(𝑝)=0

Hint: Apply the Lefschetz theorem to 𝑓𝑎 , 𝑎 = 𝑡0 /𝑁, 𝑁 large.

Exercise 5.6.xii (review of the trace). For 𝐴 = (𝑎𝑖,𝑗 ) an 𝑛 × 𝑛 matrix define
𝑛
tr(𝐴) ≔ ∑ 𝑎𝑖,𝑖 .
𝑖=1
(1) Show that if 𝐴 and 𝐵 are 𝑛 × 𝑛 matrices
tr(𝐴𝐵) = tr(𝐵𝐴) .
(2) Show that if 𝐵 is an invertible 𝑛 × 𝑛 matrix
tr(𝐵𝐴𝐵−1 ) = tr(𝐴) .
 Draft: March 28, 2018

§5.7 The Künneth theorem 191

(3) Let 𝑉 be and 𝑛-dimensional vector space and 𝐿 ∶ 𝑉 → 𝑉 a linear map. Fix a basis
𝑣1 , …, 𝑣𝑛 of 𝑉 and define the trace of 𝐿 to be the trace of 𝐴 where 𝐴 is the defining
matrix for 𝐿 in this basis, i.e.,
𝑛
𝐿𝑣𝑖 = ∑ 𝑎𝑗,𝑖 𝑣𝑗 .
𝑗=1

Show that this is an intrinsic definition not depending on the basis 𝑣1 , …, 𝑣𝑛 .

5.7. The Künneth theorem

Let 𝑋 be an 𝑛-manifold and 𝑌 an 𝑟-dimensional manifold, both of these manifolds
having finite topology. Let
𝜋∶ 𝑋 × 𝑌 → 𝑋
be the projection map 𝜋(𝑥, 𝑦) = 𝑥 and
𝜌∶ 𝑋 × 𝑌 → 𝑌
the projection map (𝑥, 𝑦) → 𝑦. Since 𝑋 and 𝑌 have finite topology their cohomology groups
are finite dimensional vector spaces. For 0 ≤ 𝑘 ≤ 𝑛 let
𝜇𝑘𝑖 , 1 ≤ 𝑖 ≤ dim 𝐻𝑘 (𝑋) ,
be a basis of 𝐻𝑘 (𝑋) and for 0 ≤ ℓ ≤ 𝑟 let
𝜈𝑗ℓ , 1 ≤ 𝑗 ≤ dim 𝐻ℓ (𝑌)

be a basis of 𝐻ℓ (𝑌). Then for 𝑘 + ℓ = 𝑚 the product, 𝜋♯ 𝜇𝑘𝑖 ⋅ 𝜌♯ 𝜈𝑗ℓ , is in 𝐻𝑚 (𝑋 × 𝑌). The
Künneth theorem asserts
Theorem 5.7.1. The product manifold 𝑋 × 𝑌 has finite topology and hence the cohomology
groups 𝐻𝑚 (𝑋 × 𝑌) are finite dimensional. Moreover, the products over 𝑘 + ℓ = 𝑚
(5.7.2) 𝜋♯ 𝜇𝑘𝑖 ⋅ 𝜌♯ 𝜈𝑗ℓ , 0 ≤ 𝑖 ≤ dim 𝐻𝑘 (𝑋) , 0 ≤ 𝑗 ≤ dim 𝐻ℓ (𝑌) ,
are a basis for the vector space 𝐻𝑚 (𝑋 × 𝑌).
The fact that 𝑋 × 𝑌 has finite topology is easy to verify. If 𝑈1 , …, 𝑈𝑀 , is a good cover of
𝑋 and 𝑉1 , …, 𝑉𝑁 , is a good cover of 𝑌 the products of these open sets 𝑈𝑖 × 𝑉𝑗 , for 1 ≤ 𝑖 ≤ 𝑀
and 1 ≤ 𝑗 ≤ 𝑁 is a good cover of 𝑋 × 𝑌: For every multi-index 𝐼, 𝑈𝐼 is either empty or
diffeomorphic to 𝐑𝑛 , and for every multi-index 𝐽, 𝑉𝐽 is either empty or diffeomorphic to 𝐑𝑟 ,
hence for any product multi-index (𝐼, 𝐽), the product 𝑈𝐼 ×𝑉𝐽 is either empty or diffeomorphic
to 𝐑𝑛 × 𝐑𝑟 . The tricky part of the proof is verifying that the products from (5.7.2) are a basis
of 𝐻𝑚 (𝑋 × 𝑌), and to do this it will be helpful to state the theorem above in a form that
avoids our choosing specified bases for 𝐻𝑘 (𝑋) and 𝐻ℓ (𝑌). To do so we’ll need to generalize
slightly the notion of a bilinear pairing between two vector space.
Definition 5.7.3. Let 𝑉1 , 𝑉2 and 𝑊 be finite dimensional vector spaces. A map 𝐵 ∶ 𝑉1 ×𝑉2 →
𝑊 of sets is a bilinear map if it is linear in each of its factors, i.e., for 𝑣2 ∈ 𝑉2 the map
𝑣1 ∈ 𝑉1 ↦ 𝐵(𝑣1 , 𝑣2 )
is a linear map of 𝑉1 into 𝑊 and for 𝑣1 ∈ 𝑉1 so is the map
𝑣2 ∈ 𝑉2 ↦ 𝐵(𝑣1 , 𝑣2 ) .
 Draft: March 28, 2018

192 Chapter 5: Cohomology via forms

It’s clear that if 𝐵1 and 𝐵2 are bilinear maps of 𝑉1 × 𝑉2 into 𝑊 and 𝜆1 and 𝜆2 are real
numbers the function
𝜆1 𝐵1 + 𝜆2 𝐵2 ∶ 𝑉1 × 𝑉2 → 𝑊
is also a bilinear map of 𝑉1 × 𝑉2 into 𝑊, so the set of all bilinear maps of 𝑉1 × 𝑉2 into 𝑊
forms a vector space. In particular the set of all bilinear maps of 𝑉1 × 𝑉2 into 𝐑 is a vector
space, and since this vector space will play an essential role in our intrinsic formulation of
the Künneth theorem, we’ll give it a name. We’ll call it the tensor product of 𝑉1⋆ and 𝑉2⋆ and
denote it by 𝑉1⋆ ⊗ 𝑉2⋆ . To explain where this terminology comes from we note that if ℓ1 and
ℓ2 are vectors in 𝑉1⋆ and 𝑉2⋆ then one can define a bilinear map
ℓ1 ⊗ ℓ2 ∶ 𝑉1 × 𝑉2 → 𝐑
by setting (ℓ1 ⊗ ℓ2 )(𝑣1 , 𝑣2 ) ≔ ℓ1 (𝑣1 )ℓ2 (𝑣2 ). In other words one has a tensor product map:
(5.7.4) 𝑉1⋆ × 𝑉2⋆ → 𝑉1⋆ ⊗ 𝑉2⋆
mapping (ℓ1 , ℓ2 ) to ℓ1 ⊗ ℓ2 . We leave for you to check that this is a bilinear map of 𝑉1⋆ × 𝑉2⋆
into 𝑉1⋆ ⊗ 𝑉2⋆ and to check as well
Proposition 5.7.5. If ℓ11 , …, ℓ𝑚
1
is a basis of 𝑉1⋆ and ℓ12 , …, ℓ𝑛2 is a basis of 𝑉2⋆ then ℓ𝑖1 ⊗ ℓ𝑗2 , for
1 ≤ 𝑖 ≤ 𝑚 and 1 ≤ 𝑗 ≤ 𝑛, is a basis of 𝑉1⋆ ⊗ 𝑉2⋆ .
Hint: If 𝑉1 and 𝑉2 are the same vector space you can find a proof of this in §1.3 and the
proof is basically the same if they’re different vector spaces.
Corollary 5.7.6. Let 𝑉1 and 𝑉2 be finite-dimensional vector spaces. Then
dim(𝑉1⋆ ⊗ 𝑉2⋆ ) = dim(𝑉1⋆ ) dim(𝑉2⋆ ) = dim(𝑉1 ) dim(𝑉2 ) .
We’ll now perform some slightly devious maneuvers with “duality” operations. First
note that for any finite dimensional vector space 𝑉, the evaluation pairing
𝑉 × 𝑉⋆ → 𝐑 , (𝑣, ℓ) ↦ ℓ(𝑣)
is a non-singular bilinear pairing, so, as we explained in §5.4 it gives rise to a bijective linear
mapping
(5.7.7) 𝑉 → (𝑉⋆ )⋆ .
Next note that if
(5.7.8) 𝐿 ∶ 𝑉1 × 𝑉2 → 𝑊
is a bilinear mapping and ℓ ∶ 𝑊 → 𝐑 a linear mapping (i.e., an element of 𝑊⋆ ), then the
composition of ℓ and 𝐿 is a bilinear mapping
ℓ ∘ 𝐿 ∶ 𝑉1 × 𝑉2 → 𝐑
and hence by definition an element of 𝑉1⋆ ⊗ 𝑉2⋆ . Thus from the bilinear mapping (5.7.8) we
get a linear mapping
(5.7.9) 𝐿♯ ∶ 𝑊⋆ → 𝑉1⋆ ⊗ 𝑉2⋆ .
We’ll now define a notion of tensor product for the vector spaces 𝑉1 and 𝑉2 themselves.
Definition 5.7.10. The vector space 𝑉1 ⊗ 𝑉2 is the vector space dual of 𝑉1⋆ ⊗ 𝑉2⋆ , i.e., is the
space
(5.7.11) 𝑉1 ⊗ 𝑉2 ≔ (𝑉1⋆ ⊗ 𝑉2⋆ )⋆ .
 Draft: March 28, 2018

§5.7 The Künneth theorem 193

One implication of (5.7.11) is that there is a natural bilinear map

(5.7.12) 𝑉1 × 𝑉2 → 𝑉1 ⊗ 𝑉2 .
(In (5.7.4) replace 𝑉𝑖 by 𝑉𝑖⋆ and note that by (5.7.7) (𝑉𝑖⋆ )⋆ = 𝑉𝑖 .) Another is the following:
Proposition 5.7.13. Let 𝐿 be a bilinear map of 𝑉1 × 𝑉2 into 𝑊. Then there exists a unique
linear map
𝐿# ∶ 𝑉1 ⊗ 𝑉2 → 𝑊
with the property
(5.7.14) 𝐿# (𝑣1 ⊗ 𝑣2 ) = 𝐿(𝑣1 , 𝑣2 )
where 𝑣1 ⊗ 𝑣2 is the image of (𝑣1 , 𝑣2 ) with respect to (5.7.12).
Proof. Let 𝐿# be the transpose of the map 𝐿♯ in (5.7.9) and note that by (5.7.7) (𝑊⋆ )⋆ =
𝑊. □

Notice that by Proposition 5.7.13 the property (5.7.14) is the defining property of 𝐿# ;
it uniquely determines this map. (This is in fact the whole point of the tensor product con-
struction. Its purpose is to convert bilinear maps, which are not linear, into linear maps.)
After this brief digression (into an area of mathematics which some mathematicians
unkindly refer to as “abstract nonsense”), let us come back to our motive for this digression:
an intrinsic formulation of the Künneth theorem. As above let 𝑋 and 𝑌 be manifolds of
dimension 𝑛 and 𝑟, respectively, both having finite topology. For 𝑘+ℓ = 𝑚 one has a bilinear
map
𝐻𝑘 (𝑋) × 𝐻ℓ (𝑌) → 𝐻𝑚 (𝑋 × 𝑌)
mapping (𝑐1 , 𝑐2 ) to 𝜋⋆ 𝑐1 ⋅ 𝜌⋆ 𝑐2 , and hence by Proposition 5.7.13 a linear map
(5.7.15) 𝐻𝑘 (𝑋) ⊗ 𝐻ℓ (𝑌) → 𝐻𝑚 (𝑋 × 𝑌) .
Let
𝐻1𝑚 (𝑋 × 𝑌) ≔ ⨁ 𝐻𝑘 (𝑋) ⊗ 𝐻ℓ (𝑌) .
𝑘+ℓ=𝑚
The maps (5.7.15) can be combined into a single linear map
(5.7.16) 𝐻1𝑚 (𝑋 × 𝑌) → 𝐻𝑚 (𝑋 × 𝑌)
and our intrinsic version of the Künneth theorem asserts the following.
Theorem 5.7.17. The map (5.7.16) is bijective.
Here is a sketch of how to prove this. (Filling in the details will be left as a series of
exercises.) Let 𝑈 be an open subset of 𝑋 which has finite topology and let
𝑚
H 1 (𝑈) ≔ ⨁ 𝐻𝑘 (𝑈) ⊗ 𝐻ℓ (𝑌)
𝑘+ℓ=𝑚

and
𝑚
H 2 (𝑈) ≔ 𝐻𝑚 (𝑈 × 𝑌) .
As we’ve just seen there’s a Künneth map
𝑚 𝑚
𝜅 ∶ H 1 (𝑈) → H 2 (𝑈) .
 Draft: March 28, 2018

194 Chapter 5: Cohomology via forms

Exercises for §5.7

Exercise 5.7.i. Let 𝑈1 and 𝑈2 be open subsets of 𝑋, both having finite topology, and let
𝑈 = 𝑈1 ∪ 𝑈2 . Show that there is a long exact sequence:
𝛿 𝑚 𝑚 𝑚 𝑚 𝛿 𝑚+1
⋯ H 1 (𝑈) H 1 (𝑈1 ) ⊕ H 1 (𝑈2 ) H 1 (𝑈1 ∩ 𝑈2 ) H1 (𝑈) ⋯,
Hint: Take the usual Mayer–Vietoris sequence
𝛿 𝛿
⋯ 𝐻𝑘 (𝑈) 𝐻𝑘 (𝑈1 ) ⊕ 𝐻𝑘 (𝑈2 ) 𝐻𝑘 (𝑈1 ∩ 𝑈2 ) 𝐻𝑘+1 (𝑈) ⋯,

tensor each term in this sequence with 𝐻ℓ (𝑌), and sum over 𝑘 + ℓ = 𝑚.
Exercise 5.7.ii. Show that for H 2 there is a similar sequence.
Hint: Apply Mayer–Vietoris to the open subsets 𝑈1 × 𝑌 and 𝑈2 × 𝑌 of 𝑀.
Exercise 5.7.iii. Show that the diagram below commutes:
𝛿 𝑚 𝑚 𝑚 𝑚 𝛿 𝑚+1
⋯ H 1 (𝑈) H 1 (𝑈1 ) ⊕ H 1 (𝑈2 ) H 1 (𝑈1 ∩ 𝑈2 ) H1 (𝑈) ⋯

𝜅 𝜅 𝜅 𝜅

𝛿 𝑚 𝑚 𝑚 𝑚 𝛿 𝑚+1
⋯ H 2 (𝑈) H 2 (𝑈1 ) ⊕ H 2 (𝑈2 ) H 2 (𝑈1 ∩ 𝑈2 ) H2 (𝑈) ⋯
(This looks hard but is actually very easy: just write down the definition of each arrow in the
language of forms.)
Exercise 5.7.iv. Conclude from Exercise 5.7.iii that if the Künneth map is bijective for 𝑈1 ,
𝑈2 and 𝑈1 ∩ 𝑈2 it is bijective for 𝑈.
Exercise 5.7.v. Prove the Künneth theorem by induction on the number of open sets in a
good cover of 𝑋. To get the induction started, note that
𝐻𝑘 (𝐑𝑛 × 𝑌) ≅ 𝐻𝑘 (𝑌) .
(See Exercise 5.3.xi.)

5.8. Čech Cohomology

We pointed out at the end of §5.3 that if a manifold 𝑋 admits a finite good cover
U = {𝑈1 , …, 𝑈𝑑 } ,
then in principle its cohomology can be read off from the intersection properties of the 𝑈𝑖 ’s.
In this section we’ll explain in more detail how to do this. More explicitly, we will show how
to construct from the intersection properties of the 𝑈𝑖 ’s a sequence of cohomology groups
𝐻𝑘 (U; 𝐑), and we will prove that these Čech cohomology groups are isomorphic to the de
Rham cohomology groups of 𝑋. (However we will leave for the reader a key ingredient in
the proof: a diagram chasing argument that is similar to the diagram chasing arguments that
were used to prove the Mayer–Vietoris theorem in §5.2 and the five lemma in §5.4.)
We will denote by 𝑁𝑘 (U) the set of all multi-indices
𝐼 = (𝑖0 , …, 𝑖𝑘 ) , 1 ≤ 𝑖0 , …, 𝑖𝑘 ≤ 𝑑
with the property that the intersection
𝑈𝐼 ≔ 𝑈𝑖0 ∩ ⋯ ∩ 𝑈𝑖𝑘
 Draft: March 28, 2018

§5.8 Čech Cohomology 195

is nonempty. (For example, for 1 ≤ 𝑖 ≤ 𝑑, the multi-index 𝐼 = (𝑖0 , …, 𝑖𝑘 ) with 𝑖0 = ⋯ = 𝑖𝑘 = 𝑖

is in 𝑁𝑘 (U) since 𝑈𝐼 = 𝑈𝑖 .) The disjoint union 𝑁(U) ≔ ∐𝑘≥0 𝑁𝑘 (U) is called the nerve of
the cover U and 𝑁𝑘 (U) is the 𝑘-skeleton of the nerve 𝑁(U). Note that if we delete 𝑖𝑟 from
a multi-index 𝐼 = (𝑖0 , …, 𝑖𝑘 ) in 𝑁𝑘 (U), i.e., replace 𝐼 by
𝐼𝑟 ≔ (𝑖0 , …, 𝑖𝑟−1 , 𝑖𝑟+1 , …, 𝑖𝑘 ) ,
𝑘−1
then 𝐼𝑟 ∈ 𝑁 (U).
We now associate a cochain complex to 𝑁(U) by defining 𝐶𝑘 (U; 𝐑) to be the finite
dimensional vector space consisting of all (set-theoretic) maps 𝑐 ∶ 𝑁𝑘 (U) → 𝐑:
𝐶𝑘 (U; 𝐑) ≔ {𝑐 ∶ 𝑁𝑘 (U) → 𝐑} ,
with addition and scalar multiplication of functions. Define a coboundary operator
𝛿 ∶ 𝐶𝑘−1 (U; 𝐑) → 𝐶𝑘 (U; 𝐑)
by setting
𝑘
(5.8.1) 𝛿(𝑐)(𝐼) ≔ ∑ (−1)𝑟 𝑐(𝐼𝑟 ) .
𝑟=0

We claim that 𝛿 actually is a coboundary operator, i.e., 𝛿 ∘ 𝛿 = 0. To see this, we note that
for 𝑐 ∈ 𝐶𝑘−1 (U; 𝐑) and 𝐼 ∈ 𝑁𝑘 (U), by (5.8.1) we have
𝑘+1
𝛿(𝛿(𝑐))(𝐼) = ∑ (−1)𝑟 𝛿(𝑐)(𝐼𝑟 )
𝑟=0
𝑘+1
= ∑ (−1)𝑟 (∑(−1)𝑠 𝑐(𝐼𝑟,𝑠 ) + ∑(−1)𝑠−1 𝑐(𝐼𝑟,𝑠 )) .
𝑟=0 𝑠<𝑟 𝑠>𝑟

Thus the term 𝑐(𝐼𝑟,𝑠 ) occurs twice in the sum, but occurs once with opposite signs, and hence
𝛿(𝛿(𝑐))(𝐼) = 0.
The cochain complex
𝛿 𝛿 𝛿 𝛿
(5.8.2) 0 𝐶0 (U; 𝐑) 𝐶1 (U; 𝐑) ⋯ 𝐶𝑘 (U; 𝐑) ⋯

is called the Čech cochain complex of the cover U. The Čech cohomology groups of the cover
U are the cohomology groups of the Čech cochain complex:
ker(𝛿 ∶ 𝐶𝑘 (U; 𝐑) → 𝐶𝑘+1 (U; 𝐑))
𝐻𝑘 (U; 𝐑) ≔ .
im(𝛿 ∶ 𝐶𝑘−1 (U; 𝐑) → 𝐶𝑘 (U; 𝐑))
The rest of the section is devoted to proving the following theorem:
Theorem 5.8.3. Let 𝑋 be a manifold and U a finite good cover of 𝑋. Then for all 𝑘 ≥ 0 we
have isomorphisms
𝐻𝑘 (U; 𝐑) ≅ 𝐻𝑘 (𝑋) .
Remarks 5.8.4.
(1) The definition of 𝐻𝑘 (U; 𝐑) only involves the nerve 𝑁(U) of this cover so Theorem 5.8.3
is in effect a proof of the claim we made above: the cohomology of 𝑋 is in principle deter-
mined by the intersection properties of the 𝑈𝑖 ’s.
(2) Theorem 5.8.3 gives us another proof of an assertion we proved earlier: if 𝑋 admits a
finite good cover the cohomology groups of 𝑋 are finite-dimensional.
 Draft: March 28, 2018

196 Chapter 5: Cohomology via forms

The proof of Theorem 5.8.3 will involve an interesting de Rham theoretic generalization
of the Čech cochain complex (5.8.2). Namely, for 𝑘 and ℓ nonnegative integers we define a
Čech cochain of degree 𝑘 with values in 𝛺ℓ to be a map 𝑐 which assigns each 𝐼 ∈ 𝑁𝑘 (U) an
ℓ-form 𝑐(𝐼) ∈ 𝛺ℓ (𝑈𝐼 ).
The set of these cochains forms an infinite dimensional vector space 𝐶𝑘 (U; 𝛺ℓ ). We
will show how to define a coboundary operator
𝛿 ∶ 𝐶𝑘−1 (U; 𝛺ℓ ) → 𝐶𝑘 (U; 𝛺ℓ )
similar to the Čech coboundary operator (5.8.1). To define this operator, let 𝐼 ∈ 𝑁𝑘 (U) and
for 0 ≤ 𝑟 ≤ 𝑘 let 𝛾𝑟 ∶ 𝛺ℓ (𝑈𝐼𝑟 ) → 𝛺ℓ (𝑈𝐼 ) be the restriction map defined by
𝜔 ↦ 𝜔|𝑈𝐼 .
(Note that the restiction 𝜔|𝑈𝐼 is well-defined since 𝑈𝐼 ⊂ 𝑈𝐼𝑟 .) Thus given a Čech cochain
𝑐 ∈ 𝐶𝑘−1 (U; 𝛺ℓ ) we can, mimicking (5.8.1), define a Čech cochain 𝛿(𝑐) ∈ 𝐶𝑘 (U; 𝛺ℓ ) by
setting
𝑘 𝑘
(5.8.5) 𝛿(𝑐)(𝐼) ≔ ∑ 𝛾𝑟 (𝑐(𝐼𝑟 )) = ∑ 𝑐(𝐼𝑟 )|𝑈𝐼 .
𝑟=0 𝑟=0
In other words, except for the 𝛾𝑟 ’s the definition of this cochain is formally identical with
the definition (5.8.1). We leave it as an exercise that the operators
𝛿 ∶ 𝐶𝑘−1 (U; 𝛺ℓ ) → 𝐶𝑘 (U; 𝛺ℓ )
satisfy 𝛿 ∘ 𝛿 = 0.1
Thus, to summarize, for every nonnegative integer ℓ we have constructed a Čech cochain
complex with values in 𝛺ℓ
𝛿 𝛿 𝛿 𝛿
(5.8.6) 0 𝐶0 (U; 𝛺ℓ ) 𝐶1 (U; 𝛺ℓ ) ⋯ 𝐶𝑘 (U; 𝛺ℓ ) ⋯.
Our next task in this section will be to compute the cohomology of the complex (5.8.6). We’ll
begin with the 0th cohomology group of (5.8.6), i.e., the kernel of the coboundary operator
𝛿 ∶ 𝐶0 (U; 𝛺ℓ ) → 𝐶1 (U; 𝛺ℓ ) .
An element 𝑐 ∈ 𝐶0 (U; 𝛺ℓ ) is, by definition, a map which assigns to each 𝑖 ∈ {1, …, 𝑑} an
element 𝑐(𝑖) ≕ 𝜔𝑖 of 𝛺ℓ (𝑈𝑖 ).
Now let 𝐼 = (𝑖0 , 𝑖1 ) be an element of 𝑁1 (U). Then, by definition, 𝑈𝑖0 ∩ 𝑈𝑖1 is nonempty
and
𝛿(𝑐)(𝐼) = 𝛾𝑖0 𝜔𝑖1 − 𝛾𝑖1 𝜔𝑖0 .
Thus 𝛿(𝑐) = 0 if and only if
𝜔𝑖0 |𝑈𝑖 ∩𝑈𝑖 = 𝜔𝑖1 |𝑈𝑖 ∩𝑈𝑖
0 1 0 1

for all 𝐼 ∈ 𝑁1 (U). This is true if and only if each ℓ-form 𝜔𝑖 ∈ 𝛺ℓ (𝑈𝑖 ) is the restriction to 𝑈𝑖
of a globally defined ℓ-form 𝜔 on 𝑈1 ∪ ⋯ ∪ 𝑈𝑑 = 𝑋. Thus
ker(𝛿 ∶ 𝐶0 (U; 𝛺ℓ ) → 𝐶1 (U; 𝛺ℓ )) = 𝛺ℓ (𝑋) .
Inserting 𝛺ℓ (𝑋) into the sequence (5.8.6) we get a new sequence
𝛿 𝛿 𝛿
(5.8.7) 0 𝛺ℓ (𝑋) 𝐶0 (U; 𝛺ℓ ) ⋯ 𝐶𝑘 (U; 𝛺ℓ ) ⋯,
and we prove:
1Hint: Except for keeping track of the 𝛾 ’s the proof is identical to the proof of the analogous result for the
𝑟
coboundary operator (5.8.1).
 Draft: March 28, 2018

§5.8 Čech Cohomology 197

Theorem 5.8.8. Let 𝑋 be a manifold and U a finite good cover of 𝑋. Then for each nonnegative
integer ℓ, the sequence (5.8.7) is exact.
We’ve just proved that the sequence (5.8.7) is exact in the first spot. To prove that (5.8.7)
is exact in its 𝑘th spot, we will construct a chain homotopy operator
𝑄 ∶ 𝐶𝑘 (U; 𝛺ℓ ) → 𝐶𝑘+1 (U; 𝛺ℓ )
with the property that
(5.8.9) 𝛿𝑄(𝑐) + 𝑄𝛿(𝑐) = 𝑐
𝑘 ℓ
for all 𝑐 ∈ 𝐶 (U; 𝛺 ). To define this operator we’ll need a slight generalization of Theo-
rem 5.2.22.
Theorem 5.8.10. Let 𝑋 be a manifold with a finite good cover {𝑈1 , …, 𝑈𝑁 }. Then there exist
𝑁
functions 𝜙1 , …, 𝜙𝑁 ∈ 𝐶∞ (𝑋) such that supp(𝜙𝑖 ) ⊂ 𝑈𝑖 for 𝑖 = 1, …, 𝑁 and ∑𝑖=1 𝜙𝑖 = 1.
Proof of Theorem 5.8.10. Let (𝜌𝑖 )𝑖≥1 , where 𝜌𝑖 ∈ 𝐶0∞ (𝑋), be a partition of unity subordinate
to the cover U and let 𝜙1 be the sum of the 𝜌𝑗 ’s with support in 𝑈1 :
𝜙1 ≔ ∑ 𝜌𝑗 .
𝑗≥1
supp(𝜌𝑗 )⊂𝑈1

Deleting those 𝜌𝑗 such that supp(𝜌𝑗 ) ⊂ 𝑈1 from the sequence (𝜌𝑖 )𝑖≥1 , we get a new sequence
(𝜌′𝑖 )𝑖≥1 . Let 𝜙2 be the sum of the 𝜌𝑗′ ’s with support in 𝑈1 :
𝜙2 ≔ ∑ 𝜌𝑗′ .
𝑗≥1
supp(𝜌𝑗′ )⊂𝑈2

Now we delete those 𝜌𝑗′ such that supp(𝜌𝑗′ ) ⊂ 𝑈2 from the sequence (𝜌′𝑖 )𝑖≥1 and construct
𝜙3 by the same method we used to construct 𝜙1 and 𝜙2 , and so on. □
Now we define the operator 𝑄 and prove Theorem 5.8.8.
Proof of Theorem 5.8.8. To define the operator (5.8.9) let 𝑐 ∈ 𝐶𝑘 (U; 𝛺ℓ ). Then we define
𝑄𝑐 ∈ 𝐶𝑘−1 (U; 𝛺ℓ ) by defining its value at 𝐼 ∈ 𝑁𝑘−1 (𝐼) to be
𝑑
𝑄𝑐(𝐼) ≔ ∑ 𝜙𝑖 𝑐(𝑖, 𝑖0 , …, 𝑖𝑘−1 ) ,
𝑖=0
th
where 𝐼 = (𝑖0 , …, 𝑖𝑘−1 ) and the 𝑖 summand on the right is defined by be equal to 0 when 𝑈𝑖 ∩
𝑈𝐼 is empty and defined to be the product of the function 𝜙𝑖 and the ℓ-form 𝑐(𝑖, 𝑖0 , …, 𝑖𝑘−1 )
when 𝑈𝑖 ∩ 𝑈𝐼 is nonempty. (Note that in this case, the multi-index (𝑖, 𝑖0 , …, 𝑖𝑘−1 ) is in 𝑁𝑘 (U)
and so by definition 𝑐(𝑖, 𝑖0 , …, 𝑖𝑘−1 ) is an element of 𝛺ℓ (𝑈𝑖 ∩ 𝑈𝐼 ). However, since 𝜙𝑖 is sup-
ported on 𝑈𝑖 , we can extend the ℓ-form 𝜙𝑖 𝑐(𝑖, 𝑖0 , …, 𝑖𝑘−1 ) ti 𝑈𝐼 by setting it equal to zero
outside of 𝑈𝑖 ∩ 𝑈𝐼 .)
To prove (5.8.9) we note that for 𝑐 ∈ 𝐶𝑘 (U; 𝛺ℓ ) we have
𝑑
𝑄𝛿𝑐(𝑖0 , …, 𝑖𝑘 ) = ∑ 𝜙𝑖 𝛿𝑐(𝑖, 𝑖0 , …, 𝑖𝑘 ) ,
𝑖=1
so by (5.8.5) the right hand side is equal to
𝑑 𝑑 𝑘
(5.8.11) ∑ 𝜙𝑖 𝛾𝑖 𝑐(𝑖0 , …, 𝑖𝑘 ) + ∑ ∑(−1)𝑠+1 𝜙𝑖 𝛾𝑖𝑠 𝑐(𝑖, 𝑖0 , …, 𝚤𝑠̂ , …, 𝑖𝑘 ) ,
𝑖=1 𝑖=1 𝑠=0
 Draft: March 28, 2018

198 Chapter 5: Cohomology via forms

where (𝑖, 𝑖0 , …, 𝚤𝑠̂ , …, 𝑖𝑘 ) is the multi-index (𝑖, 𝑖0 , …, 𝑖𝑘 ) with 𝑖𝑠 deleted. Similarly,

𝑘
𝛿𝑄𝑐(𝑖0 , …, 𝑖𝑘 ) = ∑(−1)𝑠 𝛾𝑖𝑠 𝑄𝑐(𝑖0 , …, 𝚤𝑠̂ , …, 𝑖𝑘 )
𝑠=0
𝑑 𝑘
= ∑ ∑(−1)𝑠 𝜙𝑖 𝛾𝑖𝑠 𝑐(𝑖, 𝑖0 , …, 𝚤𝑠̂ , …, 𝑖𝑘 ) .
𝑖=1 𝑠=0

However, this is the negative of the second summand of (5.8.11); so, by adding these two
summands we get
𝑑
(𝑄𝛿𝑐 + 𝛿𝑄𝑐)(𝑖0 , …, 𝑖𝑘 ) = ∑ 𝜙𝑖 𝛾𝑖 𝑐(𝑖0 , …, 𝑖𝑘 )
𝑖=1
𝑑
and since ∑𝑖=1 𝜙𝑖 = 1, this identity reduced to
(𝑄𝛿𝑐 + 𝛿𝑄𝑐)(𝐼) = 𝑐(𝐼) ,
𝑘
or, since 𝐼 ∈ 𝑁 (U) is any element, 𝑄𝛿𝑐 + 𝛿𝑄𝑐 = 𝑐.
Finally, note the Theorem 5.8.8 is an immediate consequence of the existence of this
chain homotopy operator. Namely, if 𝑐 ∈ 𝐶𝑘 (U; 𝛺ℓ ) is in the kernel of 𝛿, then
𝑐 = 𝑄𝛿(𝑐) + 𝛿𝑄(𝑐) = 𝛿𝑄(𝑐) ,
so 𝑐 is in the image of 𝛿 ∶ 𝐶 𝑘−1
(U; 𝛺ℓ ) → 𝐶𝑘 (U; 𝛺ℓ ). □
To prove Theorem 5.8.3, we note that in addition to the exact sequence
𝛿 𝛿 𝛿
0 𝛺ℓ (𝑋) 𝐶0 (U; 𝛺ℓ ) ⋯ 𝐶𝑘 (U; 𝛺ℓ ) ⋯,
we have another exact sequence
(5.8.12)
𝑑 𝑑 𝑑 𝑑 𝑑
0 𝐶𝑘 (U; 𝐑) 𝐶𝑘 (U; 𝛺0 ) 𝐶𝑘 (U; 𝛺1 ) ⋯ 𝐶𝑘 (U; 𝛺ℓ ) ⋯,

where the 𝑑’s are defined as follows: given 𝑐 ∈ 𝐶𝑘 (U; 𝛺ℓ ) and 𝐼 ∈ 𝑁𝑘 (U), we have that
𝑐(𝐼) ∈ 𝛺ℓ (𝑈𝐼 ) and 𝑑(𝑐(𝐼)) ∈ 𝛺ℓ+1 (𝑈𝐼 ), and we define 𝑑(𝑐) ∈ 𝐶𝑘 (U; 𝛺ℓ+1 ) to be the map
given by
𝐼 ↦ 𝑑(𝑐(𝐼)) .
It is clear from this definition that 𝑑(𝑑(𝑐)) = 0 so the sequence
𝑑 𝑑 𝑑 𝑑
(5.8.13) 𝐶𝑘 (U; 𝛺0 ) 𝐶𝑘 (U; 𝛺1 ) ⋯ 𝐶𝑘 (U; 𝛺ℓ ) ⋯
is a complex and the exactness of this sequence follows from the fact that, since U is a good
cover, the sequence
𝑑 𝑑 𝑑 𝑑
𝛺0 (𝑈𝐼 ) 𝛺1 (𝑈𝐼 ) ⋯ 𝛺ℓ (𝑈𝐼 ) ⋯

is exact. Moreover, if 𝑐 ∈ 𝐶𝑘 (U; 𝛺0 ) and 𝑑𝑐 = 0, then for every 𝐼 ∈ 𝑁𝑘 (U), we have

𝑑𝑐(𝐼) = 0, i.e., the 0-form 𝑐(𝐼) in 𝛺0 (𝑈𝐼 ) is constant. Hence the map given by 𝐼 ↦ 𝑐(𝐼) is
just a Čech cochain of the type we defined earlier, i.e., a map 𝑁𝑘 (U) → 𝐑. Therefore, the
kernel of this map 𝑑 ∶ 𝐶𝑘 (U; 𝛺0 ) → 𝐶𝑘 (U; 𝛺1 ) is 𝐶0 (U; 𝐑); and, adjoining this term to the
sequence (5.8.13) we get the exact sequence (5.8.12).
Lastly we note that the two operations we’ve defined above, the 𝑑 operation
𝑑 ∶ 𝐶𝑘 (U; 𝛺ℓ ) → 𝐶𝑘 (U; 𝛺ℓ+1 )
 Draft: March 28, 2018

§5.8 Čech Cohomology 199

and the 𝛿 operation 𝛿 ∶ 𝐶𝑘 (U; 𝛺ℓ ) → 𝐶𝑘+1 (U; 𝛺ℓ ), commute. i.e., for 𝑐 ∈ 𝐶𝑘 (U; 𝛺ℓ ) we
have
(5.8.14) 𝛿𝑑𝑐 = 𝑑𝛿𝑐 .

Proof of equation (5.8.14). For 𝐼 ∈ 𝑁𝑘+1 (U)

𝑘
𝛿𝑐(𝐼) = ∑ (−1)𝑟 𝛾𝑟 𝑐(𝐼𝑟 ) ,
𝑟=0

so if 𝑐(𝐼𝑟 ) ≕ 𝜔𝐼𝑟 ∈ 𝛺ℓ (𝑈𝐼𝑟 ), then

𝑘
𝛿𝑐(𝐼) = ∑ (−1)𝑟 𝜔𝐼𝑟 |𝑈𝐼
𝑟=0

and
𝑘
𝑑𝛿(𝐼) = ∑ 𝑑𝜔𝐼𝑟 |𝑈𝐼 = 𝛿𝑑𝑐(𝐼) . □
𝑟=0

Putting these results together we get the following commutative diagram:

⋮ ⋮ ⋮ ⋮

𝛿 𝛿
0 𝛺2 (𝑋) 𝐶0 (U; 𝛺2 ) 𝐶1 (U; 𝛺2 ) 𝐶2 (U; 𝛺2 ) ⋯

𝑑 𝑑 𝑑 𝑑

0 𝛺1 (𝑋) 𝐶0 (U; 𝛺1 ) 𝐶1 (U; 𝛺1 ) 𝐶2 (U; 𝛺1 ) ⋯

𝛿 𝛿

𝑑 𝑑 𝑑 𝑑

0 𝛺0 (𝑋) 𝐶0 (U; 𝛺0 ) 𝐶1 (U; 𝛺0 ) 𝐶2 (U; 𝛺0 ) ⋯

𝛿 𝛿

0 𝐶0 (U; 𝐑) 𝐶1 (U; 𝐑) 𝐶2 (U; 𝐑) ⋯

𝛿 𝛿

0 0 0 .
In this diagram all columns are exact except for the extreme left and column, which is the
usual de Rham complex, and all rows are exact except for the bottom row which is the usual
Čech cochain complex.

Exercises for §5.8

Exercise 5.8.i. Deduce from this diagram that if a manifold 𝑋 admits a finite good cover
U, then 𝐻𝑘 (U; 𝐑) ≅ 𝐻𝑘 (𝑋).
Hint: Let 𝑐 ∈ 𝛺𝑘+1 (𝑋) be a form such that 𝑑𝑐 = 0. Show that one can, by a sequence of
“chess moves” (of which the first few stages are illustrated in the (5.8.15) below), convert 𝑐
 Draft: March 28, 2018

200 Chapter 5: Cohomology via forms

into an element 𝑐 ̌ of 𝐶𝑘+1 (U; 𝐑) satisfying 𝛿(𝑐)̌ = 0

𝑐 𝑐𝑘+1,0
𝑑

𝑐𝑘,0 𝑐𝑘,1
𝛿
𝑑

𝑐𝑘−1,1 𝑐𝑘−1,2
(5.8.15) 𝛿
𝑑

𝑐𝑘−2,2 𝑐𝑘−2,3
𝛿
𝑑

𝑐𝑘−3,3 ⋯
𝛿
 Draft: March 28, 2018

APPENDIX a

Bump Functions & Partitions of Unity

We will discuss in this section a number of “global to local” techniques in multi-variable

calculus: techniques which enable one to reduce global problems on manifolds and large
open subsets of 𝐑𝑛 to local problems on small open subsets of these sets. Our starting point
will be the function 𝜌 on 𝐑 defined by

1
𝑒− 𝑥 , 𝑥 > 0
(a.1) 𝜌(𝑥) ≔ {
0, 𝑥≤0,
which is positive for 𝑥 positive, zero for 𝑥 negative and everywhere 𝐶∞ . (We will sketch a
proof of this last assertion at the end of this appendix). From 𝜌 one can construct a number
of other interesting 𝐶∞ functions.
Example a.2. For 𝑎 > 0 the function 𝜌(𝑥) + 𝜌(𝑎 − 𝑥) is positive for all 𝑥 so the quotient
𝜌𝑎 (𝑥) of 𝜌(𝑥) by this function is a well-defined 𝐶∞ function with the properties:

{𝜌 (𝑥) = 0, for 𝑥 ≤ 0
{ 𝑎
{0 ≤ 𝜌𝑎 (𝑥) ≤ 1
{
{𝜌𝑎 (𝑥) = 1, for 𝑥 ≥ 𝑎 .
Example a.3. Let 𝐼 be the open interval (𝑎, 𝑏), and 𝜌𝐼 the product of 𝜌(𝑥 − 𝑎) and 𝜌(𝑏 − 𝑥).
Then 𝜌𝐼 (𝑥) is positive for 𝑥 ∈ 𝐼 and zero on the complement of 𝐼.
Example a.4. More generally let 𝐼1 , …, 𝐼𝑛 be open intervals and let 𝑄 = 𝐼1 × ⋯ × 𝐼𝑛 be the
open rectangle in 𝐑𝑛 having these intervals as sides. Then the function
𝜌𝑄 (𝑥) ≔ 𝜌𝐼1 (𝑥1 ) ⋯ 𝜌𝐼𝑛 (𝑥𝑛 )
is a 𝐶∞ function on 𝐑𝑛 that’s positive on 𝑄 and zero on the complement of 𝑄.
Using these functions we will prove:
Lemma a.5. Let 𝐶 be a compact subset of 𝐑𝑛 and 𝑈 an open set containing 𝐶. Then there
exists function 𝜙 ∈ 𝐶0∞ (𝑈) such that
𝜙(𝑥) ≥ 0, for all 𝑥 ∈ 𝑈
{
𝜙(𝑥) > 0, for 𝑥 ∈ 𝐶 .

Proof. For each 𝑝 ∈ 𝐶 let 𝑄𝑝 be an open rectangle with 𝑝 ∈ 𝑄𝑝 and 𝑄𝑝 ⊂ 𝑈. The 𝑄𝑝 ’s cover
𝑁
𝐶𝑖 so, by Heine–Borel there exists a finite subcover 𝑄𝑝1 , …, 𝑄𝑝𝑁 . Now let 𝜙 ≔ ∑𝑖=1 𝜌𝑄𝑝 □
𝑖

This result can be slightly strengthened. Namely we claim:

Theorem a.6. There exists a function 𝜓 ∈ 𝐶0∞ (𝑈) such that 0 ≤ 𝜓 ≤ 1 and 𝜓 = 1 on 𝐶.
201
 Draft: March 28, 2018

202 Appendix a: Bump Functions & Partitions of Unity

Proof. Let 𝜙 be as in Lemma a.5 and let 𝑎 > 0 be the greatest lower bound of the restriction
of 𝜙 to 𝐶. Then if 𝜌𝑎 is the function in Example a.2 the function 𝜌𝑎 ∘ 𝜙 has the properties
indicated above. □
Remark a.7. The function 𝜓 in this theorem is an example of a bump function. If one wants
to study the behavior of a vector field 𝒗 or a 𝑘-form 𝜔 on the set 𝐶, then by multiplying 𝒗
(or 𝜔) by 𝜓 one can, without loss or generality, assume that 𝒗 (or 𝜔) is compactly supported
on a small neighborhood of 𝐶.
Bump functions are one of the standard tools in calculus for converting global problems
to local problems. Another such tool is partitions of unity.
Let 𝑈 be an open subset of 𝐑𝑛 and U = {𝑈𝛼 }𝛼∈𝐼 a covering of 𝑈 by open subsets (in-
dexed by the elements of the index set 𝐼). Then the partition of unity theorem asserts:
Theorem a.8. There exists a sequence of functions 𝜌1 , 𝜌2 , … ∈ 𝐶0∞ (𝑈) such that
(1) 𝜌𝑖 ≥ 0
(2) For every 𝑖 there is an 𝛼 ∈ 𝐼 with 𝜌𝑖 ∈ 𝐶0∞ (𝑈𝛼 )
(3) For every 𝑝 ∈ 𝑈 there exists a neighborhood 𝑈𝑝 of 𝑝 in 𝑈 and an 𝑁𝑝 > 0 such that
𝜌𝑖 |𝑈𝑝 = 0 for 𝑢 > 𝑁𝑝 .
(4) ∑ 𝜌𝑖 = 1
Remark a.9. Because of item (3) the sum in item (4) is well defined. We will derive this
result from a somewhat simpler set theoretical result.
Theorem a.10. There exists a countable covering of 𝑈 by open rectangles (𝑄𝑖 )𝑖≥1 such that
(1) 𝑄𝑖 ⊂ 𝑈
(2) For each 𝑖 there is an 𝛼 ∈ 𝐼 with 𝑄𝑖 ⊂ 𝑈𝛼
(3) For every 𝑝 ∈ 𝑈 there exists a neighborhood 𝑈𝑝 of 𝑝 in 𝑈 and 𝑁𝑝 > 0 such that 𝑄𝑖 ∩ 𝑈𝑝
is empty for 𝑖 > 𝑁𝑝
We first note that Theorem a.10 implies Theorem a.8. To see this note that the func-
tions 𝜌𝑄𝑖 in Example a.4 above have all the properties indicated in Theorem a.8 except for
∞
property (4). Moreover since the 𝑄𝑖 ’s are a covering of 𝑈 the sum ∑𝑖=1 𝜌𝑄𝑖 is everywhere
positive. Thus we get a sequence of 𝜌𝑖 ’s satisfying (1)–(4) by taking 𝜌𝑖 to be the quotient of
𝜌𝑄𝑖 by this sum.
Proof of Theorem a.10. Let 𝑑(𝑥, 𝑈𝑐 ) be the distance of a point 𝑥 ∈ 𝑈 to the complement
𝑈𝑐 ≔ 𝐑𝑛 ∖ 𝑈 of 𝑈 in 𝐑𝑛 , and let 𝐴 𝑟 be the compact subset of 𝑈 consisting of points, 𝑥 ∈ 𝑈,
satisfying 𝑑(𝑥, 𝑈𝑐 ) ≥ 1/𝑟 and |𝑥| ≤ 𝑟. By Heine–Borel we can find, for each 𝑟, a collection of
open rectangles, 𝑄𝑟,𝑖 , 𝑖 = 1, …, 𝑁𝑟 , such that 𝑄𝑟,𝑖 is contained in int(𝐴 𝑟+1 ∖𝐴 𝑟−2 ) and in some
𝑈𝛼 and such that the 𝑄𝑖 , 𝑟’s are a covering of 𝐴 𝑟 ∖int(𝐴 𝑟−1 ). Thus the 𝑄𝑟,𝑖 ’s have the properties
listed in Theorem a.10, and by relabelling, i.e., setting 𝑄𝑖 = 𝑄1,𝑖 for 1 ≤ 𝑖 ≤ 𝑁𝑖 , 𝑄𝑁1 +𝑖 ≔ 𝑄2,𝑖 ,
for 1 ≤ 𝑖 ≤ 𝑁2 , etc., we get a sequence, 𝑄1 , 𝑄2 , … with the desired properties. □
Applications
We will next describe a couple of applications of Theorem a.10.
Application a.11 (Improper integrals). Let 𝑓 ∶ 𝑈 → 𝐑 be a continuous function. We will
say that 𝑓 is integrable over 𝑈 if the infinite sum
∞
(a.12) ∑ ∫ 𝜌𝑖 (𝑥) |𝑓(𝑥)| 𝑑𝑥
𝑖=1 𝑈
 Draft: March 28, 2018

203

converges and if so we will define the improper integral of 𝑓 over 𝑈 to be the sum
∞
(a.13) ∑ ∫ 𝜌𝑖 (𝑥)𝑓(𝑥)𝑑𝑥
𝑖=1 𝑈
(Notice that each of the summands in this series is the integral of a compactly supported
continuous function over 𝐑𝑛 so the individual summands are well-defined. Also it’s clear
that if 𝑓 itself is a compactly supported function on 𝐑𝑛 , (a.12) is bounded by the integral of
|𝑓| over 𝐑𝑛 , so for every 𝑓 ∈ 𝐶0 (𝐑𝑛 ) the improper integral of 𝑓 over 𝑈 is well-defined.)
Application a.14 (An extension theorem for 𝐶∞ maps). Let 𝑋 be a subset of 𝐑𝑛 and 𝑓 ∶
𝑋 → 𝐑𝑚 a continuous map. We will say that 𝑓 is 𝐶∞ if, for every 𝑝 ∈ 𝑋, there exists an
open neighborhood, 𝑈𝑝 , of 𝑝 in 𝐑𝑛 and a 𝐶∞ map, 𝑔𝑝 ∶ 𝑈𝑝 → 𝐑𝑚 such that 𝑔𝑝 = 𝑓 on
𝑈𝑝 ∩ 𝑋.
Theorem a.15 (Extension Theorem). If 𝑓 ∶ 𝑋 → 𝐑𝑚 is 𝐶∞ there exists an open neighbor-
hood 𝑈 of 𝑋 in 𝐑𝑛 and a 𝐶∞ map 𝑔 ∶ 𝑈 → 𝐑𝑚 , such that 𝑔 = 𝑓 on 𝑋.
Proof. Let 𝑈 = ⋃𝑝∈𝑋 𝑈𝑝 and let (𝜌𝑖 )𝑖≥1 be a partition of unity subordinate to the covering
U = {𝑈𝑝 }𝑝∈𝑋 of 𝑈. Then, for each 𝑖, there exists a 𝑝 such that the support of 𝜌𝑖 is contained
in 𝑈𝑝 . Let
𝜌𝑖 𝑔𝑝 (𝑥), 𝑥 ∈ 𝑈𝑝
𝑔𝑖 (𝑥) = {
0, 𝑥 ∈ 𝑈𝑝𝑐 .
∞
Then 𝑔 = ∑𝑖=1 𝑔𝑖 is well defined by item (3) of Theorem a.8 and the restriction of 𝑔 to 𝑋 is
∞
given by ∑𝑖=1 𝜌𝑖 𝑓, which is equal to 𝑓. □
Exercises for Appendix a
Exercise a.i. Show that the function (a.1) is 𝐶∞ .
Hints:
(1) From the Taylor series expansion
∞
𝑥𝑘
𝑒𝑥 = ∑
𝑘=0 𝑘!
conclude that for 𝑥 > 0
𝑥𝑘+𝑛
𝑒𝑥 ≥ .
(𝑘 + 𝑛)!
(2) Replacing 𝑥 by 1/𝑥 conclude that for 𝑥 > 0,
1 1
𝑒1/𝑥 ≥ .
(𝑛 + 𝑘)! 𝑥𝑛+𝑘
(3) From this inequality conclude that for 𝑥 > 0,
1
𝑒− 𝑥 ≤ (𝑛 + 𝑘)!𝑥𝑛+𝑘 .
(4) Let 𝑓𝑛 (𝑥) be the function
1
𝑒− 𝑥 𝑥−𝑛 , 𝑥 > 0
𝑓𝑛 (𝑥) ≔ {
0, 𝑥≤0.
Conclude from (3) that for 𝑥 > 0,
𝑓𝑛 (𝑥) ≤ (𝑛 + 𝑘)!𝑥𝑘
for all 𝑘
 Draft: March 28, 2018

204 Appendix a: Bump Functions & Partitions of Unity

(5) Conclude that 𝑓𝑛 is 𝐶1 differentiable.

(6) Show that
𝑑
𝑓 = 𝑓𝑛+2 − 𝑛𝑓𝑛+1
𝑑𝑥 𝑛
(7) Deduce by induction that the 𝑓𝑛 ’s are 𝐶𝑟 differentiable for all 𝑟 and 𝑛.
Exercise a.ii. Show that the improper integral (a.13) is well-defined independent of the
choice of partition of unity.
Hint: Let (𝜌𝑗′ )𝑗≥1 be another partition of unity. Show that (a.13) is equal to
∞ ∞
∑ ∑ ∫ 𝜌𝑖 (𝑥)𝜌𝑗′ (𝑥)𝑓(𝑥)𝑑𝑥 .
𝑖=1 𝑗=1 𝑈
 Draft: March 28, 2018

APPENDIX b

The Implicit Function Theorem

Let 𝑈 be an open neighborhood of the origin in 𝐑𝑛 and let 𝑓1 , …, 𝑓𝑘 be 𝐶∞ functions

on 𝑈 with the property 𝑓1 (0) = ⋯ = 𝑓𝑘 (0) = 0. Our goal in this appendix is to prove the
following implicit function theorem.
Theorem b.1. Suppose the matrix
𝜕𝑓
(b.2) [ 𝜕𝑥𝑗𝑖 (0)]
is non-singular. Then there exists a neighborhood 𝑈0 of 0 in 𝐑𝑛 and a diffeomorphism 𝑔 ∶ (𝑈0 , 0) ↪
(𝑈, 0) of 𝑈0 onto an open neighborhood of 0 in 𝑈 such that
𝑔⋆ 𝑓𝑖 = 𝑥𝑖 , 𝑖 = 1, …, 𝑘
and
𝑔⋆ 𝑥𝑗 = 𝑥𝑗 , 𝑗 = 𝑘 + 1, …, 𝑛 .
Remarks b.3.
(1) Let 𝑔(𝑥) = (𝑔1 (𝑥), …, 𝑔𝑛 (𝑥)). Then the second set of equations just say that
𝑔𝑗 (𝑥1 , …, 𝑥𝑛 ) = 𝑥𝑗
for 𝑗 = 𝑘 + 1, …, 𝑛, so the first set of equations can be written more concretely in the
form
(b.4) 𝑓𝑖 (𝑔1 (𝑥1 , …, 𝑥𝑛 ), …, 𝑔𝑘 (𝑥1 , …, 𝑥𝑛 ), 𝑥𝑘+1 , …, 𝑥𝑛 ) = 𝑥𝑖
for 𝑖 = 1, …, 𝑘.
(2) Letting 𝑦𝑖 = 𝑔𝑖 (𝑥1 , …, 𝑥𝑛 ) the equations (b.4) become
(b.5) 𝑓𝑖 (𝑦1 , …, 𝑦𝑘 , 𝑥𝑘+1 , …, 𝑥𝑛 ) = 𝑥𝑖 .
Hence what the implicit function theorem is saying is that, modulo the assumption (b.2)
the system of equations (b.5) can be solved for the 𝑦𝑖 ’s in terms of the 𝑥𝑖 ’s.
(3) By a linear change of coordinates:
𝑘
𝑥𝑖 ↦ ∑ 𝑎𝑖,𝑟 𝑥𝑟 , 𝑖 = 1, …, 𝑘
𝑟=1
we can arrange without loss of generality for the matrix (b.2) to be the identity matrix,
i.e.,
𝜕𝑓𝑖
(b.6) (0) = 𝛿𝑖,𝑗 , 1 ≤ 𝑖, 𝑗 ≤ 𝑘 .
𝜕𝑥𝑗
Our proof of this theorem will be by induction on 𝑘.
First, let’s suppose that 𝑘 = 1 and, for the moment, also suppose that 𝑛 = 1. Then the
theorem above is just the inverse function theorem of freshman calculus. Namely if 𝑓(0) = 0
and 𝑑𝑓/𝑑𝑥(0) = 1, there exists an interval (𝑎, 𝑏) about the origin on which 𝑑𝑓/𝑑𝑥 is greater
205
 Draft: March 28, 2018

206 Appendix b: The Implicit Function Theorem

than 1/2, so 𝑓 is strictly increasing on this interval and its graph looks like the curve in the
figure below with 𝑐 = 𝑓(𝑎) ≤ − 12 𝑎 and 𝑑 = 𝑓(𝑏) ≥ 12 𝑏.

𝑎 𝑏
𝑥

Figure b.1. The implicit function theorem

The graph of the inverse function 𝑔 ∶ [𝑐, 𝑑] → [𝑎, 𝑏] is obtained from this graph by just
rotating it through ninety degrees, i.e., making the 𝑦-axis the horizontal axis and the 𝑥-axis
the vertical axis. (From the picture it’s clear that 𝑦 = 𝑓(𝑥) ⟺ 𝑥 = 𝑔(𝑦) ⟺ 𝑓(𝑔(𝑦)) =
𝑦.)
Most elementary text books regard this intuitive argument as being an adequate proof
of the inverse function theorem; however, a slightly beefed-up version of this proof (which is
completely rigorous) can be found in [12, Ch. 12]. Moreover, as Spivak points out in [12, Ch.
12], if the slope of the curve in the figure above at the point (𝑥, 𝑦) is equal to 𝜆 the slope of
the rotated curve at (𝑦, 𝑥) is 1/𝜆, so from this proof one concludes that if 𝑦 = 𝑓(𝑥)
−1 −1
𝑑𝑔 𝑑𝑓 𝑑𝑓
(b.7) (𝑦) = ( (𝑥)) =( (𝑔(𝑦))) .
𝑑𝑦 𝑑𝑥 𝑑𝑥
Since 𝑓 is a continuous function, its graph is a continuous curve and, therefore, since the
graph of 𝑔 is the same curve rotated by ninety degrees, 𝑔 is also a continuous functions.
Hence by (b.7), 𝑔 is also a 𝐶1 function and hence by (b.7), 𝑔 is a 𝐶2 function, etc. In other
words 𝑔 is in 𝐶∞ ([𝑐, 𝑑]).
Let’s now prove that the implicit function theorem with 𝑘 = 1 and 𝑛 arbitrary. This
amounts to showing that if the function 𝑓 in the discussion above depends on the param-
eters, 𝑥2 , …𝑥𝑛 in a 𝐶∞ fashion, then so does its inverse 𝑔. More explicitly let’s suppose
𝑓(𝑥1 , …, 𝑥𝑛 ) is a 𝐶∞ function on a subset of 𝐑𝑛 of the form [𝑎, 𝑏] × 𝑉 where 𝑉 is a compact,
convex neighborhood of the origin in 𝐑𝑛 and satisfies 𝜕𝑓/𝜕𝑥1 ≥ 21 on this set. Then by the
argument above there exists a function, 𝑔 = 𝑔(𝑦, 𝑥2 , …, 𝑥𝑛 ) defined on the set [𝑎/2, 𝑏/2] × 𝑉
with the property
(b.8) 𝑓(𝑥1 , 𝑥2 , …, 𝑥𝑛 ) = 𝑦 ⟺ 𝑔(𝑦, 𝑥2 , …, 𝑥𝑛 ) = 𝑥1 .
Moreover by (b.7) 𝑔 is a 𝐶∞ function of 𝑦 and
𝜕𝑔 𝜕𝑓
(b.9) (𝑦, 𝑥2 , …, 𝑥𝑛 ) = (𝑥 , 𝑥 , …, 𝑥𝑛 )−1
𝜕𝑦 𝜕𝑥1 1 2
 Draft: March 28, 2018

207

𝜕𝑓 1
at 𝑥1 = 𝑔(𝑦). In particular, since 𝜕𝑥1
≥ 2

𝜕𝑔
0< <2
𝜕𝑦
and hence
(b.10) |𝑔(𝑦′ , 𝑥2 , …, 𝑥𝑛 ) − 𝑔(𝑦, 𝑥2 , …, 𝑥𝑛 )| < 2|𝑦′ − 𝑦|
for points 𝑦 and 𝑦′ in the interval [𝑎/2, 𝑏/2].
The 𝑘 = 1 case of Theorem b.1 is almost implied by (b.8) and (b.9) except that we must
still show that 𝑔 is a 𝐶∞ function, not just of 𝑦, but of all the variables, 𝑦, 𝑥2 , …, 𝑥𝑛 , and this
we’ll do by quoting another theorem from freshman calculus (this time a theorem from the
second semester of freshman calculus).
Theorem b.11 (Mean Value Theorem in 𝑛 variables). Let 𝑈 be a convex open set in 𝐑𝑛 and
𝑓 ∶ 𝑈 → 𝐑 a 𝐶∞ function. Then for 𝑎, 𝑏 ∈ 𝑈 there exists a point 𝑐 on the line interval joining
𝑎 to 𝑏 such that
𝑛
𝜕𝑓
𝑓(𝑏) − 𝑓(𝑎) = ∑ (𝑐)(𝑏𝑖 − 𝑎𝑖 ) .
𝑖=1 𝜕𝑥 𝑖

Proof. Apply the one-dimensional mean value theorem to the function

ℎ(𝑡) ≔ 𝑓((1 − 𝑡)𝑎 + 𝑡𝑏) . □
Let’s now show that the function 𝑔 in (b.8) is a 𝐶∞ function of the variables, 𝑦 and 𝑥2 .
To simplify our notation we’ll suppress the dependence of 𝑓 on 𝑥3 , …, 𝑥𝑛 and write 𝑓 as
𝑓(𝑥1 , 𝑥2 ) and 𝑔 as 𝑔(𝑦, 𝑥2 ). For ℎ ∈ (−𝜀, 𝜀), 𝜀 small, we have
𝑦 = 𝑓(𝑔(𝑦, 𝑥2 + ℎ), 𝑥2 + ℎ) = 𝑓(𝑔(𝑦, 𝑥2 ), 𝑥2 ) ,
and, hence, setting 𝑥1′ = 𝑔(𝑦, 𝑥2 + ℎ), 𝑥1 = 𝑔(𝑦, 𝑥2 ) and 𝑥2′ = 𝑥2 + ℎ, we get from the mean
value theorem
0 = 𝑓(𝑥1′ , 𝑥2′ ) − 𝑓(𝑥1 , 𝑥2 )
𝜕𝑓 𝜕𝑓
= (𝑐)(𝑥1′ − 𝑥1 ) + (𝑐)(𝑥2′ − 𝑥2 )
𝜕𝑥1 𝜕𝑥2
and therefore
−1
𝜕𝑓 𝜕𝑓
(b.12) 𝑔(𝑦, 𝑥2 + ℎ) − 𝑔(𝑦, 𝑥2 ) = (− (𝑐)) (𝑐) ℎ
𝜕𝑥1 𝜕𝑥2
for some 𝑐 on the line segment joining (𝑥1 , 𝑥2 ) to (𝑥1′ , 𝑥2′ ). Letting ℎ tend to zero we can draw
from this identity a number of conclusions:
(1) Since 𝑓 is a 𝐶∞ function on the compact set [𝑎, 𝑏] × 𝑉 its derivatives are bounded on
this set, so the right hand side of (b.12) tends to zero as ℎ tends to zero, i.e., for fixed 𝑦,
𝑔 is continuous as a function of 𝑥2 .
(2) Now divide the right hand side of (b.12) by ℎ and let ℎ tend to zero. Since 𝑔 is continuous
in 𝑥2 , the quotient of (b.12) by ℎ tends to its value at (𝑥1 , 𝑥2 ). Hence for fixed 𝑦 𝑓 is
differentiable as a function of 𝑥2 and
−1
𝜕𝑔 𝜕𝑓 𝜕𝑓
(b.13) (𝑦, 𝑥2 ) = − ( ) (𝑥1 , 𝑥2 ) (𝑥 , 𝑥 )
𝜕𝑥2 𝜕𝑥1 𝜕𝑥2 1 2
where 𝑥1 = 𝑔(𝑦, 𝑥2 ).
 Draft: March 28, 2018

208 Appendix b: The Implicit Function Theorem

(3) Moreover, by the inequality (b.10) and the triangle inequality

|𝑔(𝑦′ , 𝑥2′ ) − 𝑔(𝑦, 𝑥2 )| ≤ |𝑔(𝑦′ , 𝑥2′ ) − 𝑔(𝑦, 𝑥2′ )| + |𝑔(𝑦, 𝑥2′ ) − 𝑔(𝑦, 𝑥2 )|
≤ 2|𝑦′ − 𝑦| + |𝑔(𝑦, 𝑥2′ ) − 𝑔(𝑦, 𝑥2 )| ,
hence 𝑔 is continuous as a function of 𝑦 and 𝑥2 .
(4) Hence by (b.9) and (b.13), 𝑔 is a 𝐶1 function of 𝑦 and 𝑥2 , and hence ….
In other words 𝑔 is a 𝐶∞ function of 𝑦 and 𝑥2 . This argument works, more or less
verbatim, for more than two 𝑥𝑖 ’s and proves that 𝑔 is a 𝐶∞ function of 𝑦, 𝑥2 , …, 𝑥𝑛 . Thus
with 𝑓 = 𝑓1 and 𝑔 = 𝑔1 Theorem b.1 is proved in the special case 𝑘 = 1.
Proof of Theorem b.1 for 𝑘 > 1. We’ll now prove Theorem b.1 for arbitrary 𝑘 by induction
on 𝑘. By induction we can assume that there exists a neighborhood 𝑈0′ , of the origin in 𝐑𝑛
and a 𝐶∞ diffeomorphism 𝜙 ∶ (𝑈0′ , 0) ↪ (𝑈, 0) such that
(b.14) 𝜙⋆ 𝑓𝑖 = 𝑥𝑖
for 2 ≤ 𝑖 ≤ 𝑘 and
(b.15) 𝜙⋆ 𝑥𝑗 = 𝑥𝑗
for 𝑗 = 1 and 𝑘 + 1 ≤ 𝑗 ≤ 𝑛. Moreover, by (b.6)
𝑘
𝜕 ⋆ 𝜕𝑓 𝜕 𝜕𝜙
( 𝜙 𝑓1 ) (0) = ∑ 𝑖 (0) 𝜙𝑖 (0) = 1 = 1
𝜕𝑥1 𝑖=1 𝜕𝑥1 𝜕𝑥 𝑖 𝜕𝑥1
since 𝜙1 = 𝜙⋆ 𝑥1 = 𝑥1 . Therefore we can apply Theorem b.1, with 𝑘 = 1, to the function, 𝜙⋆ 𝑓1 ,
to conclude that there exists a neighborhood 𝑈0 of the origin in 𝐑𝑛 and a diffeomorphism
𝜓 ∶ (𝑈0 , 0) → (𝑈0′ , 0) such that 𝜓⋆ 𝜙⋆ 𝑓1 = 𝑥1 and 𝜓⋆ 𝜙⋆ 𝑥𝑖 = 𝑥𝑖 for 1 ≤ 𝑖 ≤ 𝑛. Thus by (b.14)
𝜓⋆ 𝜙⋆ 𝑓𝑖 = 𝜓⋆ 𝑥𝑖 = 𝑥𝑖 for 2 ≤ 𝑖 ≤ 𝑘, and by (b.15) 𝜓⋆ 𝜙⋆ 𝑥𝑗 = 𝜓⋆ 𝑥𝑗 = 𝑥𝑗 for 𝑘 + 1 ≤ 𝑗 ≤ 𝑛.
Hence if we let 𝑔 = 𝜙 ∘ 𝜓 we see that:
𝑔⋆ 𝑓𝑖 = (𝜙 ∘ 𝜓)⋆ 𝑓𝑖 = 𝜓⋆ 𝜙⋆ 𝑓𝑖 = 𝑥𝑖
for 𝑖 ≤ 𝑖 ≤ 𝑘 and
𝑔⋆ 𝑥𝑗 = (𝜙 ∘ 𝜓)⋆ 𝑥𝑗 = 𝜓⋆ 𝜙⋆ 𝑥𝑗 = 𝑥𝑗
for 𝑘 + 1 ≤ 𝑗 ≤ 𝑛. □
We’ll derive a number of subsidiary results from Theorem b.1. The first of these is the
𝑛-dimensional version of the inverse function theorem:
Theorem b.16. Let 𝑈 and 𝑉 be open subsets of 𝐑𝑛 and 𝜙 ∶ (𝑈, 𝑝) → (𝑉, 𝑞) a 𝐶∞ map. Suppose
that the derivative of 𝜙 at 𝑝
𝐷𝜙(𝑝) ∶ 𝐑𝑛 → 𝐑𝑛
is bijective. Then 𝜙 maps a neighborhood of 𝑝 in 𝑈 diffeomorphically onto a neighborhood of
𝑞 in 𝑉.
Proof. By pre-composing and post-composing 𝜙 by translations we can assume that 𝑝 =
𝑞 = 0. Let 𝜙 = (𝑓1 , …, 𝑓𝑛 ). Then the condition that 𝐷𝜙(0) be bijective is the condition that
the matrix (b.2) be non-singular. Hence, by Theorem b.1, there exists a neighborhood 𝑉0 of
0 in 𝑉, a neighborhood 𝑈0 of 0 in 𝑈, and a diffeomorphism 𝑔 ∶ 𝑉0 → 𝑈0 such that
𝑔⋆ 𝑓𝑖 = 𝑥𝑖
for 𝑖 = 1, …, 𝑛. However, these equations simply say that 𝑔 is the inverse of 𝜙, and hence
that 𝜙 is a diffeomorphism. □
 Draft: March 28, 2018

209

A second result which we’ll extract from Theorem b.1 is the canonical submersion the-
orem.
Theorem b.17. Let 𝑈 be an open subset of 𝐑𝑛 and 𝜙 ∶ (𝑈, 𝑝) → (𝐑𝑘 , 0) a 𝐶∞ map. Suppose
𝜙 is a submersion at 𝑝, i.e., suppose its derivative
𝐷𝜙(𝑝) ∶ 𝐑𝑛 → 𝐑𝑘
is onto. Then there exists a neighborhood 𝑉 of 𝑝 in 𝑈, a neighborhood 𝑈0 of the origin in 𝐑𝑛 ,
and a diffeomorphism 𝑔 ∶ (𝑈0 , 0) → (𝑉, 𝑝) such that the map 𝜙 ∘ 𝑔 ∶ (𝑈0 , 0) → (𝐑𝑛 , 0) is the
restriction to 𝑈0 of the canonical submersion:
𝜋 ∶ 𝐑𝑛 → 𝐑 𝑘 , 𝜋(𝑥1 , …, 𝑥𝑛 ) ≔ (𝑥1 , …, 𝑥𝑘 ) .
Proof. Let 𝜙 = (𝑓1 , …, 𝑓𝑘 ). Composing 𝜙 with a translation we can assume 𝑝 = 0 and by a
permutation of the variables 𝑥1 , …, 𝑥𝑛 we can assume that the matrix (b.2) is non-singular.
By Theorem b.1 we conclude that there exists a diffeomorphism 𝑔 ∶ (𝑈0 , 0) ↪ (𝑈, 𝑝) with
the properties 𝑔⋆ 𝑓𝑖 = 𝑥𝑖 for 𝑖 = 1, …, 𝑘, and hence,
𝜙 ∘ 𝑔(𝑥1 , …, 𝑥𝑛 ) = (𝑥1 , …, 𝑥𝑘 ) . □
As a third application of Theorem b.1 we’ll prove a theorem which is similar in spirit to
Theorem b.17, the canonical immersion theorem.
Theorem b.18. Let 𝑈 be an open neighborhood of the origin in 𝐑𝑘 and 𝜙 ∶ (𝑈, 0) → (𝐑𝑛 , 𝑝)
a 𝐶∞ map. Suppose that the derivative of 𝜙 at 0
𝐷𝜙(0) ∶ 𝐑𝑘 → 𝐑𝑛
is injective. Then there exists a neighborhood 𝑈0 of 0 in 𝑈, a neighborhood 𝑉 of 𝑝 in 𝐑𝑛 , and
a diffeomorphism.
𝜓 ∶ 𝑉 → 𝑈0 × 𝐑𝑛−𝑘
such that the map 𝜓 ∘ 𝜙 ∶ 𝑈0 → 𝑈0 × 𝐑𝑛−𝑘 is the restriction to 𝑈0 of the canonical immersion
𝜄 ∶ 𝐑𝑘 → 𝐑𝑘 × 𝐑𝑛−𝑘 , 𝜄(𝑥1 , …, 𝑥𝑘 ) = (𝑥1 , …, 𝑥𝑘 , 0, …, 0) .
Proof. Let 𝜙 = (𝑓1 , …, 𝑓𝑛 ). By a permutation of the 𝑓𝑖 ’s we can arrange that the matrix
𝜕𝑓
[ 𝜕𝑥𝑗𝑖 (0)] , 1 ≤ 𝑖, 𝑗 ≤ 𝑘
is non-singular and by composing 𝜙 with a translation we can arrange that 𝑝 = 0. Hence by
Theorem b.16 the map
𝜒 ∶ (𝑈, 0) → (𝐑𝑘 , 0) 𝑥 ↦ (𝑓1 (𝑥), …, 𝑓𝑘 (𝑥))
maps a neighborhood 𝑈0 of 0 diffeomorphically onto a neighborhood 𝑉0 of 0 in 𝐑𝑘 . Let
𝜓 ∶ (𝑉0 , 0) → (𝑈0 , 0) be its inverse and let
𝛾 ∶ 𝑉0 × 𝐑𝑛−𝑘 → 𝑈0 × 𝐑𝑛−𝑘
be the map
𝛾(𝑥1 , …, 𝑥𝑛 ) ≔ (𝜓(𝑥1 , …, 𝑥𝑘 ), 𝑥𝑘+1 , …, 𝑥𝑛 ) .
Then
(b.19) (𝛾 ∘ 𝜙)(𝑥1 , …, 𝑥𝑘 ) = 𝛾(𝜒(𝑥1 , …, 𝑥𝑘 ), 𝑓𝑘+1 (𝑥), …, 𝑓𝑛 (𝑥))
= (𝑥1 , …, 𝑥𝑘 , 𝑓𝑘+1 (𝑥), …, 𝑓𝑛 (𝑥)) .
Now note that the map
ℎ ∶ 𝑈0 × 𝐑𝑛−𝑘 → 𝑈0 × 𝐑𝑛−𝑘
 Draft: March 28, 2018

210 Appendix b: The Implicit Function Theorem

defined by
ℎ(𝑥1 , …, 𝑥𝑛 ) ≔ (𝑥1 , …, 𝑥𝑘 , 𝑥𝑘+1 − 𝑓𝑘+1 (𝑥), …, 𝑥𝑛 − 𝑓𝑛 (𝑥))
is a diffeomorphism. (Why? What is its inverse?) and, by (b.19),
(ℎ ∘ 𝛾 ∘ 𝜙)(𝑥1 , …, 𝑥𝑘 ) = (𝑥1 , …, 𝑥𝑘 , 0, …, 0) ,
i.e., (ℎ ∘ 𝛾) ∘ 𝜙 = 𝜄. Thus we can take the 𝑉 in Theorem b.18 to be 𝑉0 × 𝐑𝑛−𝑘 and the 𝜓 to be
ℎ ∘ 𝛾. □
Remark b.20. The canonical submersion and immersion theorems can be succinctly sum-
marized as saying that every submersion “looks locally like the canonical submersion” and
every immersion “looks locally like the canonical immersion”.
 Draft: March 28, 2018

APPENDIX c

Good Covers & Convexity Theorems

Let 𝑋 be an 𝑛-dimensional submanifold of 𝐑𝑁 . Our goal in this appendix is to prove

that 𝑋 admits a good cover. To do so we’ll need some basic facts about convexity.
Definition c.1. Let 𝑈 be an open subset of 𝐑𝑛 and 𝜙 a 𝐶∞ function on 𝑈. We say that 𝜙 is
convex if the matrix
2
𝜕 𝜙
(c.2) [ 𝜕𝑥 𝜕𝑥 (𝑝)]
𝑖 𝑗

is positive-definite for all 𝑝 ∈ 𝑃.

Suppose now that 𝑈 itself is convex and 𝜙 is a convex function on 𝑈 which has as image
the half open interval [0, 𝑎) and is a proper map 𝑈 → [0, 𝑎). (In other words, for 0 ≤ 𝑐 < 𝑎,
the set { 𝑥 ∈ 𝑈 | 𝜙(𝑥) ≤ 𝑐 } is compact).
Theorem c.3. For 0 < 𝑐 < 𝑎, the set
𝑈𝑐 ≔ { 𝑥 ∈ 𝑈 | 𝜙(𝑥) < 𝑐 }
is convex.
Proof. For 𝑝, 𝑞 ∈ 𝜕𝑈𝑐 with 𝑝 ≠ 𝑞, let
𝑓(𝑡) ≔ 𝜙((1 − 𝑡)𝑝 + 𝑡𝑞) ,
for 0 ≤ 𝑡 ≤ 1. We claim that
𝑑2 𝑓
(c.4) (𝑡) > 0
𝑑𝑡2
and
𝑑𝑓 𝑑𝑓
(c.5) (0) < 0 < (1) .
𝑑𝑡 𝑑𝑡
To prove (c.4) we note that
𝑑2 𝑓 𝜕2 𝜙
2
(𝑡) = ∑ ((1 − 𝑡)𝑝 + 𝑡𝑞)(𝑝𝑖 − 𝑞𝑖 )(𝑝𝑗 − 𝑞𝑗 )
𝑑𝑡 1≤𝑖,𝑗≤𝑛 𝜕𝑥𝑖 𝜕𝑥𝑗

and that the right-hand side is positive because of the positive-definiteness of (c.2).
To prove (c.5) we note that 𝑑𝑓 𝑑𝑡
is strictly increasing by (c.4). Hence 𝑑𝑓
𝑑𝑡
(0) > 0 would
𝑑𝑓
imply that 𝑑𝑡 (𝑡) > 0 for all 𝑡 ∈ [0, 1] and hence would imply that 𝑓 is strictly increasing.
But
𝑓(0) = 𝑓(1) = 𝜙(𝑝) = 𝜙(𝑞) = 𝑐 ,
so this would be a contradiction. We conclude that 𝑑𝑓 𝑑𝑡
(0) < 0. A similar argument shows
𝑑𝑓
that 𝑑𝑡 (1) > 0.
Now we use equations (c.4) and (c.5) to show that 𝑈𝑐 is convex. Since 𝑑𝑓 𝑑𝑡
is strictly
𝑑𝑓 𝑑𝑓
increasing it follows from (c.5) that 𝑑𝑡 (𝑡0 ) = 0 for some 𝑡0 ∈ (0, 1) and that 𝑑𝑡 (𝑡) < 0 for
211
 Draft: March 28, 2018

212 Appendix c: Good Covers & Convexity Theorems

𝑡 ∈ [0, 𝑡0 ) and 𝑑𝑓
𝑑𝑡
(𝑡) > 0 for 𝑡 ∈ (𝑡0 , 1]. Therefore, since 𝑓(0) = 𝑓(1) = 𝑐, we have 𝑓(𝑡) < 𝑐
for all 𝑡 ∈ (0, 1); so 𝜙(𝑥) < 𝑐 on the line segment defined by (1 − 𝑡)𝑝 + 𝑡𝑞 for 𝑡 ∈ (0, 1). Hence
𝑈𝑐 is convex. □

Coming back to the manifold 𝑋 ⊂ 𝐑𝑁 , let 𝑝 ∈ 𝑋 and let 𝐿𝑝 ≔ 𝑇𝑝 𝑋. We denote by 𝜋𝑝

the orthogonal projection
𝜋𝑝 ∶ 𝑋 → 𝐿𝑝 + 𝑝 ,
i.e., for 𝑞 ∈ 𝑋 and 𝑥 = 𝜋𝑝 (𝑞), we have that 𝑞 − 𝑥 is orthogonal to 𝐿𝑝 (it’s easy to see that 𝜋𝑝
is defined by this condition). We will prove that this projection has the following convexity
property.
Theorem c.6. There exists a positive constant 𝑐(𝑝) such that if 𝑞 ∈ 𝑋 satisfies |𝑝 − 𝑞|2 < 𝑐(𝑝)
and 𝑐 < 𝑐(𝑝), then the set
(c.7) 𝑈(𝑞, 𝑐) ≔ { 𝑞′ ∈ 𝑋 | |𝑞′ − 𝑞| < 𝑐 }
is mapped diffeomorphically by 𝜋𝑝 onto a convex open set in 𝐿𝑝 + 𝑝.
Proof. We can, without loss of generality, assume that 𝑝 = 0 and that
𝐿 = 𝐑𝑛 = 𝐑𝑛 × {0}
in 𝐑𝑛 × 𝐑𝑘 = 𝐑𝑁 , where 𝑘 ≔ 𝑁 − 𝑛. Thus 𝑋 can be described, locally over a convex neigh-
borhood 𝑈 of the origin in 𝐑𝑛 as the graph of a 𝐶∞ function 𝑓 ∶ (𝑈, 0) → (𝐑𝑘 , 0) and since
𝐑𝑛 × {0} is tangent to 𝑋 at 0, we have
𝜕𝑓
(c.8) =0 for 𝑖 = 1, …, 𝑛 .
𝜕𝑥𝑖
Given 𝑞 = (𝑥𝑞 , 𝑓(𝑥𝑞 )) ∈ 𝑋, let 𝜙𝑞 ∶ 𝑈 → 𝐑 by the function
(c.9) 𝜙𝑞 (𝑥) ≔ |𝑥 − 𝑥𝑞 |2 + |𝑓(𝑥) − 𝑓(𝑥𝑞 )|2 .
Then by (c.7) 𝜋𝑝 (𝑈(𝑞, 𝑐)) is just the set
{ 𝑥 ∈ 𝑈 | 𝜙𝑞 (𝑥) < 𝑐 } ,
so to prove the theorem it suffices to prove that if 𝑐(𝑝) is sufficiently small and 𝑐 < 𝑐(𝑝), this
set is convex. However, at 𝑞 = 0, by equations (c.8) and (c.9) we have
𝜕2 𝜙𝑞
= 𝛿𝑖,𝑗 ,
𝜕𝑥𝑖 𝜕𝑥𝑗
and hence by continuity 𝜙𝑞 is convex on the set { 𝑥 ∈ 𝑈 | |𝑥|2 < 𝛿 } provided that 𝛿 and
|𝑞| are sufficiently small. Hence if 𝑐(𝑝) is sufficiently small, then 𝜋𝑝 (𝑈(𝑞, 𝑐)) is convex for
𝑐 < 𝑐(𝑝). □

We will now use Theorem c.6 to prove that 𝑋 admits a good cover: for every 𝑝 ∈ 𝑋, let
√𝑐(𝑝)
𝜀(𝑝) ≔
3
and let
𝑈𝑝 ≔ { 𝑞 ∈ 𝑋 | |𝑝 − 𝑞| < 𝜀(𝑝) } .
Theorem c.10. The 𝑈𝑝 ’s are a good cover of 𝑋.
 Draft: March 28, 2018

213

Proof. Suppose that the intersection

(c.11) 𝑈𝑝1 ∩ ⋯ ∩ 𝑈𝑝𝑘
is nonempty and that
𝜀(𝑝1 ) ≥ 𝜀(𝑝2 ) ≥ ⋯ ≥ 𝜀(𝑝𝑘 ) .
Then if 𝑝 is a point in the intersection 𝑈𝑝1 ∩ ⋯ ∩ 𝑈𝑝𝑘 , then for 𝑖 = 1, …, 𝑘 we have |𝑝 − 𝑝𝑖 | <
𝜀(𝑝1 ), hence |𝑝𝑖 −𝑝1 | < 2𝜀(𝑝1 ). Moreover, if 𝑞 ∈ 𝑈𝑝𝑖 and |𝑞−𝑝𝑖 | < 𝜀(𝑝1 ), then |𝑞−𝑝1 | < 3𝜀(𝑝1 )
and so |𝑞 − 𝑝1 |2 < 𝑐(𝑝1 ). Therefore the set 𝑈𝑝𝑖 is contained in 𝑈(𝑝𝑖 , 𝑐(𝑝1 )) and consequently
by Theorem c.6 is mapped diffeomorphically onto a convex open subset of 𝐿𝑝1 + 𝑝1 by the
projection 𝜋𝑝1 . Consequently the intersection (c.11) is mapped diffeomorphically onto a
convex open subset of 𝐿𝑝1 + 𝑝1 by 𝜋𝑝1 . □
Draft: March 28, 2018
 Draft: March 28, 2018

Bibliography

1. Vladimir Igorevich Arnol’d, Mathematical methods of classical mechanics, Vol. 60, Springer Science & Business
Media, 2013.
2. Garrett Birkhoff and Gian-Carlo Rota, Ordinary differential equations, 4th ed., John Wiley & Sons, 1989.
3. Raoul Bott and Loring W. Tu, Differential forms in algebraic topology, Vol. 82, Springer Science & Business
Media, 2013.
4. Victor Guillemin and Alan Pollack, Differential topology, Vol. 370, American Mathematical Society, 2010.
5. Victor Guillemin and Shlomo Sternberg, Symplectic techniques in physics, Cambridge University Press, 1990.
6. Nikolai V. Ivanov, A differential forms perspective on the Lax proof of the change of variables formula, The Amer-
ican Mathematical Monthly 112 (2005), no. 9, 799–806.
7. John L. Kelley, General topology, Graduate Texts in Mathematics, vol. 27, Springer, 1975.
8. Peter D. Lax, Change of variables in multiple integrals, The American Mathematical Monthly 106 (1999), no. 6,
497–501.
9. , Change of variables in multiple integrals II, The American Mathematical Monthly 108 (2001), no. 2,
115–119.
10. James R. Munkres, Analysis on manifolds, Westview Press, 1997.
11. , Topology, 2nd ed., Prentice Hall, 2000 (English).
12. Michael Spivak, Calculus on manifolds: a modern approach to classical theorems of advanced calculus, Mathe-
matics Monographs, Westview Press, 1971.
13. Frank W. Warner, Foundations of differentiable manifolds and Lie groups, Graduate Texts in Mathematics,
vol. 94, Springer-Verlag New York, 1983.

215
Draft: March 28, 2018
 Draft: March 28, 2018

Index of Notation

A 𝑘 (𝑉), 14 𝛬𝑘 (𝑉⋆ ), 19
Alt, 14 𝐿𝒗 , 34

𝐵𝑘 (𝑋), 149
𝐵𝑐𝑘 (𝑋), 149 𝑁(U), 195
𝑁𝑘 (U), 194
( 𝑛𝑘 ), 16

𝐶𝑘 (U; 𝛺ℓ ), 196 𝛺𝑘 (𝑋), 110

𝐶𝑘 (U; 𝐑), 195 𝛺𝑘𝑐 (𝑋), 111
𝜒(𝑋), 189
𝑓⋆ 𝒗, 40
deg, 77
𝐷𝑓, 33
𝑑𝑓, 33 𝑆𝑘 , 12
dim, 1 𝛴𝑘 , 12
𝑑𝑥𝑖 , 35 (−1)𝜎 , 12
𝜕/𝜕𝑥𝑖 , 34 supp, 39, 71

𝐻𝑘 (𝑋), 149
𝑇𝑝 𝐑𝑛 , 33
𝐻𝑘 (U; 𝐑), 195
𝑇𝑝 (𝑋), 104, 105
𝐻𝑐𝑘 (𝑋), 149
⊗, 9
≃, 56
𝑇𝑌 , 176
I𝑘 (𝑉), 17 𝑇𝜎 , 13
im, 2 𝜏𝑌 , 176
𝜄𝑣 , 23
𝜄𝒗 , 36 𝑊(𝑋, 𝑝), 139
∧, 20
ker, 2

𝐿(𝑓), 184 𝑍𝑘 (𝑋), 149

L𝑘 (𝑉), 8 𝑍𝑐𝑘 (𝑋), 149

217
Draft: March 28, 2018
 Draft: March 28, 2018

Glossary of Terminology

alternation operation, 14 decomposable element, 21

annihilator of a subspace, 6 degree
of a map, 77
basis, 1 derivative
positively oriented, 29 of a 𝐶∞ map between manifolds, 106
bilinear, 191 determinant, 26
bilinear pairing, 171 diagonal, 184
non-singular, 171 diffeomorphism, 97
orientation preserving, 117
canonical immersion theorem, 108, 209
orientation reversing, 118
canonical submersion theorem, 107, 209
differential form
chain rule
closed, 48
for manifolds, 106
compactly supported, 71, 111
change of variables formula
exact, 48
for manifolds, 126
on an open subset of 𝐑𝑛 , 44
cochain complex, 158
one-form, 35
cohomology class, 149
dimension
configuration space, 66
of a vector space, 1
contractable, 56
divergence theorem
manifold, 137
for manifolds, 130
convex, 141
dual vector space, 4
open set, 166
convex–concave, 141 electric field, 62
coset, 3 energy
cotangent space conservation of, 67
to 𝐑𝑛 , 35 kinetic, 66
critial point, 135 total, 66
critical Euler characteristic, 189
point, 85 exact sequence, 158
value, 85 short, 159
critical value, 135 exterior derivative, 46
current density, 62
Čech 𝑓-related
cochain complex, 195 vector fields, 40
cohomology groups, 195 vector fields on manifolds, 109
finite topology, 167
Darboux form, 63 Frobenius’ Theorem, 92
de Rham cohomology, 149
compactly supported, 149 genus, 143
219
 Draft: March 28, 2018

220 Glossary of Terminology

global Lefschetz number, 184 line integral, 36

good cover, 165 linear independence, 1
gradient, 60 long exact sequence in cohomology, 162

homotopic, 56 magnetic field, 62

homotopy, 56 manifold, 97
equivalence, 156 Mayer–Vietoris Theorem, 162
invariant, 151 multi-index, 8
proper, 87 repeating, 15
strictly increasing, 15
image
of a linear map, 2 nerve, 168
immersion nerve of a cover, 195
canonical, 209 Noether’s principle, 67
implicit function theorem, 205
index one-parameter group
local, 144 associated to a complete vector field,
of a vector field, 143 40
inner product, 2 open set
integral parametrizable, 110
of a system, 38 star-shaped, 58
integral curve, 37, 111 orientation
maximal, 38 compatible, 180
of a line in 𝐑2 , 29
interior product
of a manifold, 116
for vector spaces, 24
of a one-dimensional vector space, 29
of differential forms, 51
of a vector space (in general), 29
of forms, 36
reversed, 116
intersection
smooth, 116
transverse, 180
standard, 116
intersection number
orientation preserving
local, 180
diffeomorphism, 117
of a map with a submanifold, 183
orientation reversing
inverse function theorem
diffeomorphism, 118
for manifolds, 107
orientation-preserving
Jacobi identity map of vector spaces, 30
for vector fields, 36 orthonormal vectors, 7

𝑘-form parametrizable open set

on an open subset of 𝐑𝑛 , 44 𝐷-adapted, 120
𝑘-skeleton, 195 parametrization, 104
kernel oriented, 118
of a linear map, 2 permutation, 12
transposition, 12
Lefschetz map, 184 elementary, 12
Lie algebra, 36 phase space, 66
Lie bracket Poincaré lemma
of vector fields, 36 for compactly supported forms, 76
Lie derivative, 34 for manifolds, 150
of a differential form, 52 for rectangles, 72
 Draft: March 28, 2018

221

Poincaré recurrence, 67 tangent space

Poisson bracket, 68 of a manifold, 98, 104, 105
projection formula, 176 to 𝐑𝑛 , 33
proper tensor, 8
map, 77 alternating, 14
pullback decomposable, 10
of a one-form, 41 redundant, 17
of forms, 54 symmetric, 19
of vector fields, 43, 109 tensor product, 192
pushforward of linear maps, 9
of a vector field, 40 Thom
of vector fields, 109 class, 176
form, 176
quotient
vector space, 3 vector, 1
positively oriented, 29
regular value, 85, 99, 135 vector field, 34
saddle, 146 compactly supported, 39
Sard’s theorem, 86 complete, 38, 111
scalar charge density, 62 Hamiltonian, 65
scalar curvature, 141 Lefschetz, 190
sign on a manifold, 109
of a permutation, 13 smooth, 110
simply connected, 157 symplectic, 64
sink, 146 vector space, 1
smooth domain, 118 volume element
source, 146 of a vector space, 31
submersion, 209 volume form
sumbersion on a manifold, 116
canonical, 209 Riemannian, 117
support symplectic, 64
of a differential form wedge product
on 𝐑𝑛 , 71 of forms, 45
on a manifold, 111 winding number, 139
of a vector field, 39
symmetric group, 12 zero
symplectomorphism, 64 non-degenerate, 144