Lectures on the Geometry of Manifolds

Liviu I. Nicolaescu

May 14, 2025

Introduction

Shape is a fascinating and intriguing subject which has stimulated the imagination of
many people. It suffices to look around to become curious. Euclid did just that and came
up with the first pure creation. Relying on the common experience, he created an abstract
world that had a life of its own. As the human knowledge progressed so did the ability of
formulating and answering penetrating questions. In particular, mathematicians started
wondering whether Euclid’s “obvious” absolute postulates were indeed obvious and/or
absolute. Scientists realized that Shape and Space are two closely related concepts and
asked whether they really look the way our senses tell us. As Felix Klein pointed out
in his Erlangen Program, there are many ways of looking at Shape and Space so that
various points of view may produce different images. In particular, the most basic issue
of “measuring the Shape” cannot have a clear cut answer. This is a book about Shape,
Space and some particular ways of studying them.
Since its inception, the differential and integral calculus proved to be a very versatile
tool in dealing with previously untouchable problems. It did not take long until it found
uses in geometry in the hands of the Great Masters. This is the path we want to follow
in the present book.
In the early days of geometry nobody worried about the natural context in which the
methods of calculus “feel at home”. There was no need to address this aspect since for the
particular problems studied this was a non-issue. As mathematics progressed as a whole
the “natural context” mentioned above crystallized in the minds of mathematicians and
it was a notion so important that it had to be given a name. The geometric objects which
can be studied using the methods of calculus were called smooth manifolds. Special cases
of manifolds are the curves and the surfaces and these were quite well understood. B.
Riemann was the first to note that the low dimensional ideas of his time were particular
aspects of a higher dimensional world.
The first chapter of this book introduces the reader to the concept of smooth manifold
through abstract definitions and, more importantly, through many we believe relevant
examples. In particular, we introduce at this early stage the notion of Lie group. The
main geometric and algebraic properties of these objects will be gradually described as we
progress with our study of the geometry of manifolds. Besides their obvious usefulness in
geometry, the Lie groups are academically very friendly. They provide a marvelous testing
ground for abstract results. We have consistently taken advantage of this feature through-
out this book. As a bonus, by the end of these lectures the reader will feel comfortable
manipulating basic Lie theoretic concepts.
To apply the techniques of calculus we need things to derivate and integrate. These

i
ii

“things” are introduced in Chapter 2. The reason why smooth manifolds have many
differentiable objects attached to them is that they can be locally very well approximated
by linear spaces called tangent spaces . Locally, everything looks like traditional calculus.
Each point has a tangent space attached to it so that we obtain a “bunch of tangent spaces”
called the tangent bundle. We found it appropriate to introduce at this early point the
notion of vector bundle. It helps in structuring both the language and the thinking.
Once we have “things to derivate and integrate” we need to know how to explicitly
perform these operations. We devote the Chapter 3 to this purpose. This is perhaps
one of the most unattractive aspects of differential geometry but is crucial for all further
developments. To spice up the presentation, we have included many examples which
will found applications in later chapters. In particular, we have included a whole section
devoted to the representation theory of compact Lie groups essentially describing the
equivalence between representations and their characters.
The study of Shape begins in earnest in Chapter 4 which deals with Riemann manifolds.
We approach these objects gradually. The first section introduces the reader to the notion
of geodesics which are defined using the Levi-Civita connection. Locally, the geodesics
play the same role as the straight lines in an Euclidian space but globally new phenomena
arise. We illustrate these aspects with many concrete examples. In the final part of this
section we show how the Euclidian vector calculus generalizes to Riemann manifolds.
The second section of this chapter initiates the local study of Riemann manifolds.
Up to first order these manifolds look like Euclidian spaces. The novelty arises when we
study “second order approximations ” of these spaces. The Riemann tensor provides the
complete measure of how far is a Riemann manifold from being flat. This is a very involved
object and, to enhance its understanding, we compute it in several instances: on surfaces
(which can be easily visualized) and on Lie groups (which can be easily formalized). We
have also included Cartan’s moving frame technique which is extremely useful in concrete
computations. As an application of this technique we prove the celebrated Theorema
Egregium of Gauss. This section concludes with the first global result of the book, namely
the Gauss-Bonnet theorem. We present a proof inspired from [38] relying on the fact
that all Riemann surfaces are Einstein manifolds. The Gauss-Bonnet theorem will be a
recurring theme in this book and we will provide several other proofs and generalizations.
One of the most fascinating aspects of Riemann geometry is the intimate correlation
“local-global”. The Riemann tensor is a local object with global effects. There are cur-
rently many techniques of capturing this correlation. We have already described one in
the proof of Gauss-Bonnet theorem. In Chapter 5 we describe another such technique
which relies on the study of the global behavior of geodesics. We felt we had the moral
obligation to present the natural setting of this technique and we briefly introduce the
reader to the wonderful world of the calculus of variations. The ideas of the calculus of
variations produce remarkable results when applied to Riemann manifolds. For example,
we explain in rigorous terms why “very curved manifolds” cannot be “too long” .
In Chapter 6 we leave for a while the “differentiable realm” and we briefly discuss the
fundamental group and covering spaces. These notions shed a new light on the results
of Chapter 5. As a simple application we prove Weyl’s theorem that the semisimple Lie
groups with definite Killing form are compact and have finite fundamental group.
Chapter 7 is the topological core of the book. We discuss in detail the cohomology
 iii

of smooth manifolds relying entirely on the methods of calculus. In writing this chapter
we could not, and would not escape the influence of the beautiful monograph [22], and
this explains the frequent overlaps. In the first section we introduce the DeRham coho-
mology and the Mayer-Vietoris technique. Section 2 is devoted to the Poincaré duality, a
feature which sets the manifolds apart from many other types of topological spaces. The
third section offers a glimpse at homology theory. We introduce the notion of (smooth)
cycle and then present some applications: intersection theory, degree theory, Thom iso-
morphism and we prove a higher dimensional version of the Gauss-Bonnet theorem at the
cohomological level. The fourth section analyzes the role of symmetry in restricting the
topological type of a manifold. We prove Élie Cartan’s old result that the cohomology
of a symmetric space is given by the linear space of its bi-invariant forms. We use this
technique to compute the lower degree cohomology of compact semisimple Lie groups. We
conclude this section by computing the cohomology of complex Grassmannians relying on
Weyl’s integration formula and Schur polynomials. The chapter ends with a fifth section
containing a concentrated description of Čech cohomology.
Chapter 8 is a natural extension of the previous one. We describe the Chern-Weil
construction for arbitrary principal bundles and then we concretely describe the most im-
portant examples: Chern classes, Pontryagin classes and the Euler class. In the process,
we compute the ring of invariant polynomials of many classical groups. Usually, the con-
nections in principal bundles are defined in a global manner, as horizontal distributions.
This approach is geometrically very intuitive but, at a first contact, it may look a bit
unfriendly in concrete computations. We chose a local approach build on the reader’s ex-
perience with connections on vector bundles which we hope will attenuate the formalism
shock. In proving the various identities involving characteristic classes we adopt an invari-
ant theoretic point of view. The chapter concludes with the general Gauss-Bonnet-Chern
theorem. Our proof is a variation of Chern’s proof.
Chapter 9 is the analytical core of the book.1 Many objects in differential geometry
are defined by differential equations and, among these, the elliptic ones play an important
role. This chapter represents a minimal introduction to this subject. After presenting
some basic notions concerning arbitrary partial differential operators we introduce the
Sobolev spaces and describe their main functional analytic features. We then go straight
to the core of elliptic theory. We provide an almost complete proof of the elliptic a
priori estimates (we left out only the proof of the Calderon-Zygmund inequality). The
regularity results are then deduced from the a priori estimates via a simple approximation
technique. As a first application of these results we consider a Kazhdan-Warner type
equation which recently found applications in solving the Seiberg-Witten equations on
a Kähler manifold. We adopt a variational approach. The uniformization theorem for
compact Riemann surfaces is then a nice bonus. This may not be the most direct proof but
it has an academic advantage. It builds a circle of ideas with a wide range of applications.
The last section of this chapter is devoted to Fredholm theory. We prove that the elliptic
operators on compact manifolds are Fredholm and establish the homotopy invariance of the
index. These are very general Hodge type theorems. The classical one follows immediately
from these results. We conclude with a few facts about the spectral properties of elliptic

1
In the current edition, this is Chapter 10
iv

operators.
The last chapter is entirely devoted to a very important class of elliptic operators
namely the Dirac operators. The important role played by these operators was singled
out in the works of Atiyah and Singer and, since then, they continue to be involved in the
most dramatic advances of modern geometry. We begin by first describing a general notion
of Dirac operators and their natural geometric environment, much like in [14]. We then
isolate a special subclass we called geometric Dirac operators. Associated to each such
operator is a very concrete Weitzenböck formula which can be viewed as a bridge between
geometry and analysis, and which is often the source of many interesting applications. The
abstract considerations are backed by a full section describing many important concrete
examples.
In writing this book we had in mind the beginning graduate student who wants to
specialize in global geometric analysis in general and gauge theory in particular. The
second half of the book is an extended version of a graduate course in differential geometry
we taught at the University of Michigan during the winter semester of 1996.
The minimal background needed to successfully go through this book is a good knowl-
edge of vector calculus and real analysis, some basic elements of point set topology and
linear algebra. A familiarity with some basic facts about the differential geometry of curves
of surfaces would ease the understanding of the general theory, but this is not a must.
Some parts of the chapter on elliptic equations may require a more advanced background
in functional analysis.
The theory is complemented by a large list of exercises. Quite a few of them contain
technical results we did not prove so we would not obscure the main arguments. There
are however many non-technical results which contain additional information about the
subjects discussed in a particular section. We left hints whenever we believed the solution
is not straightforward.

Personal note It has been a great personal experience writing this book, and I sincerely
hope I could convey some of the magic of the subject. Having access to the remarkable
science library of the University of Michigan and its computer facilities certainly made my
job a lot easier and improved the quality of the final product.
I learned differential equations from Professor Viorel Barbu, a very generous and en-
thusiastic person who guided my first steps in this field of research. He stimulated my
curiosity by his remarkable ability of unveiling the hidden beauty of this highly technical
subject. My thesis advisor, Professor Tom Parker, introduced me to more than the funda-
mentals of modern geometry. He played a key role in shaping the manner in which I regard
mathematics. In particular, he convinced me that behind each formalism there must be
a picture, and uncovering it, is a very important part of the creation process. Although
I did not directly acknowledge it, their influence is present throughout this book. I only
hope the filter of my mind captured the full richness of the ideas they so generously shared
with me.
My friends Louis Funar and Gheorghe Ionesei 2 read parts of the manuscript. I am
grateful to them for their effort, their suggestions and for their friendship. I want to thank
Arthur Greenspoon for his advice, enthusiasm and relentless curiosity which boosted my
2
He passed away while I was preparing the second edition. He was the ultimate poet of mathematics.
 v

spirits when I most needed it. Also, I appreciate very much the input I received from the
graduate students of my “Special topics in differential geometry” course at the University
of Michigan which had a beneficial impact on the style and content of this book.
At last, but not the least, I want to thank my family who supported me from the
beginning to the completion of this project.
Ann Arbor, 1996.

Preface to the second edition

Rarely in life is a man given the chance to revisit his “youthful indiscretions”. With
this second edition I have been given this opportunity, and I have tried to make the best
of it.
The first edition was generously sprinkled with many typos, which I can only attribute
to the impatience of youth. In spite of this problem, I have received very good feedback
from a very indulgent and helpful audience, from all over the world.
In preparing the new edition, I have been engaged on a massive typo hunting, supported
by the wisdom of time, and the useful comments that I have received over the years from
many readers. I can only say that the number of typos is substantially reduced. However,
experience tells me that Murphy’s Law is still at work, and there are still typos out there
which will become obvious only in the printed version.
The passage of time has only strengthened my conviction that, in the words of Isaac
Newton, “in learning the sciences examples are of more use than precepts”. The new
edition continues to be guided by this principle. I have not changed the old examples, but
I have polished many of my old arguments, and I have added quite a large number of new
examples and exercises.
The only major addition to the contents is a new chapter (Chapter 9) on classical
integral geometry. This is a subject that captured my imagination over the last few years,
and since the first edition developed all the tools needed to understand some of the juiciest
results in this area of geometry, I could not pass the chance to share with a curious reader
my excitement about this line of thought.
One novel feature in our presentation of the classical results of integral geometry is the
use of tame geometry. This is a recent extension of the better know area of real algebraic
geometry which allowed us to avoid many heavy analytical arguments, and present the
geometric ideas in as clear a light as possible.
Notre Dame, 2007.

Preface to the third edition

I started writing the first edition 25 years ago as a fresh PhD and I was warned
by many “adults” that this was a wrong career move. I embarked in this project with
all the energy, enthusiasm, inexperience and confidence of young age. It was meant to
be an honest presentation of the basic elements of differential geometry used in global
analysis. By “honest” presentation I understood that I should include clear and detailed
explanations for many of the folklore results, examples and points of view that are harder
to trace in the literature and had helped in my research.
 i

It was an exciting experience writing the first edition and I was rewarded for my effort
it in many ways. There was an immediate reward for the time spent immersed in the
minutiæ of many examples and proofs. The detailed understanding I achieved allowed me
to push my own research to new directions and in such a depth that I did not believe I
was capable of at that age.
There was a long term reward since it turns out that the useful facts and examples
and computations that were harder to trace are still useful and harder to trace, but now
I can always turn to this book for details. Personal bias aside, I keep a copy of my book
on my desk since I frequently need to look up something in it.
There is a personal reward, as an author. I have been receiving input and acknowl-
edgments from many readers from all the corners of the world. I have implemented all
the corrections and suggestions I have received. A book lives through its readers and
apparently the present one is still alive.
So what is knew in the third edition? There are some obvious additions reflecting
my current research interests. There is the new Chapter 11 on spectral geometry leading
to the original results in Subsection 11.4.5. This is the first place where they appear in
published form.
Describing new results in a monograph rather than in a research journal has allowed
me to go to a level of detail and provide perspective that is not possible in a journal. In
particular, I have added two new subsections 4.2.5 and 5.2.3 on Riemannian geometry
containing facts that, surprisingly, are not familiar to many geometers. These play an
important role in Subsection 11.4.5.
There are less conspicuous changes. I have been using various parts of the book for
graduate courses I taught over the years. The new edition contains the useful feedback
I have received from my students. I have enhanced and cleaned many proofs and I have
added new examples. Of course I have corrected the typos pointed out to me by readers.
It is likely I have introduced new ones.
Notre Dame, 2019.
ii
Contents

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i

1 Manifolds 3
1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Space and Coordinatization . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 The implicit function theorem . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Partitions of unity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.4 How many manifolds are there? . . . . . . . . . . . . . . . . . . . . . 22

2 Natural Constructions on Manifolds 25

2.1 The tangent bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.1 Tangent spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.2 The tangent bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.3 Transversality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.4 Vector bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.5 Some examples of vector bundles . . . . . . . . . . . . . . . . . . . . 39
2.2 A linear algebra interlude . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.1 Tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.2 Symmetric and skew-symmetric tensors . . . . . . . . . . . . . . . . 48
2.2.3 The “super” slang . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.2.4 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.2.5 Some complex linear algebra . . . . . . . . . . . . . . . . . . . . . . 67
2.3 Tensor fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.3.1 Operations with vector bundles . . . . . . . . . . . . . . . . . . . . . 70
2.3.2 Tensor fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.3.3 Fiber bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

3 Calculus on Manifolds 81
3.1 The Lie derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.1.1 Flows on manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.1.2 The Lie derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.1.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.2 Derivations of Ω• (M ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

iii
iv CONTENTS

3.2.1 The exterior derivative . . . . . . . . . . . . . . . . . . . . . . . . . . 91

3.2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.3 Connections on vector bundles . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3.1 Covariant derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3.2 Parallel transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.3.3 The curvature of a connection . . . . . . . . . . . . . . . . . . . . . . 104
3.3.4 Holonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.3.5 The Bianchi identities . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.3.6 Connections on tangent bundles . . . . . . . . . . . . . . . . . . . . 112
3.4 Integration on manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.4.1 Integration of 1-densities . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.4.2 Orientability and integration of differential forms . . . . . . . . . . . 118
3.4.3 Stokes’ formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.4.4 Representations and characters of compact Lie groups . . . . . . . . 129
3.4.5 Fibered calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

4 Riemannian Geometry 141

4.1 Metric properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.1.1 Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . 141
4.1.2 The Levi–Civita connection . . . . . . . . . . . . . . . . . . . . . . . 144
4.1.3 The exponential map and normal coordinates . . . . . . . . . . . . . 150
4.1.4 The length minimizing property of geodesics . . . . . . . . . . . . . 152
4.1.5 Calculus on Riemann manifolds . . . . . . . . . . . . . . . . . . . . . 157
4.2 The Riemann curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.2.1 Definitions and properties . . . . . . . . . . . . . . . . . . . . . . . . 167
4.2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.2.3 Cartan’s moving frame method . . . . . . . . . . . . . . . . . . . . . 173
4.2.4 The geometry of submanifolds . . . . . . . . . . . . . . . . . . . . . 177
4.2.5 Correlators and their geometry . . . . . . . . . . . . . . . . . . . . . 184
4.2.6 The Gauss–Bonnet theorem for oriented surfaces . . . . . . . . . . . 194

5 Elements of the Calculus of Variations 203

5.1 The least action principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
5.1.1 The 1-dimensional Euler-Lagrange equations . . . . . . . . . . . . . 203
5.1.2 Noether’s conservation principle . . . . . . . . . . . . . . . . . . . . 209
5.2 The variational theory of geodesics . . . . . . . . . . . . . . . . . . . . . . . 212
5.2.1 Variational formulæ . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
5.2.2 Jacobi fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
5.2.3 The Hamilton-Jacobi equations . . . . . . . . . . . . . . . . . . . . . 223

6 The Fundamental Group and Covering Spaces 229

6.1 The fundamental group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
6.1.1 Basic notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
6.1.2 Of categories and functors . . . . . . . . . . . . . . . . . . . . . . . . 234
6.2 Covering Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
CONTENTS v

6.2.1 Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . 235

6.2.2 Unique lifting property . . . . . . . . . . . . . . . . . . . . . . . . . 237
6.2.3 Homotopy lifting property . . . . . . . . . . . . . . . . . . . . . . . . 238
6.2.4 On the existence of lifts . . . . . . . . . . . . . . . . . . . . . . . . . 239
6.2.5 The universal cover and the fundamental group . . . . . . . . . . . . 241

7 Cohomology 243
7.1 DeRham cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
7.1.1 Speculations around the Poincaré lemma . . . . . . . . . . . . . . . 243
7.1.2 Čech vs. DeRham . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
7.1.3 Very little homological algebra . . . . . . . . . . . . . . . . . . . . . 249
7.1.4 Functorial properties of the DeRham cohomology . . . . . . . . . . . 256
7.1.5 Some simple examples . . . . . . . . . . . . . . . . . . . . . . . . . . 259
7.1.6 The Mayer–Vietoris principle . . . . . . . . . . . . . . . . . . . . . . 261
7.1.7 The Künneth formula . . . . . . . . . . . . . . . . . . . . . . . . . . 265
7.2 The Poincaré duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
7.2.1 Cohomology with compact supports . . . . . . . . . . . . . . . . . . 267
7.2.2 The Poincaré duality . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
7.3 Intersection theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
7.3.1 Cycles and their duals . . . . . . . . . . . . . . . . . . . . . . . . . . 275
7.3.2 Intersection theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
7.3.3 The topological degree . . . . . . . . . . . . . . . . . . . . . . . . . . 285
7.3.4 The Thom isomorphism theorem . . . . . . . . . . . . . . . . . . . . 287
7.3.5 Gauss–Bonnet revisited . . . . . . . . . . . . . . . . . . . . . . . . . 290
7.4 Symmetry and topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
7.4.1 Symmetric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
7.4.2 Symmetry and cohomology . . . . . . . . . . . . . . . . . . . . . . . 297
7.4.3 The cohomology of compact Lie groups . . . . . . . . . . . . . . . . 300
7.4.4 Invariant forms on Grassmannians and Weyl’s integral formula . . . 301
7.4.5 The Poincaré polynomial of a complex Grassmannian . . . . . . . . 309
7.5 Čech cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
7.5.1 Sheaves and presheaves . . . . . . . . . . . . . . . . . . . . . . . . . 315
7.5.2 Čech cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320

8 Characteristic Classes 331

8.1 Chern–Weil theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
8.1.1 Connections in principal G-bundles . . . . . . . . . . . . . . . . . . . 331
8.1.2 G-vector bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
8.1.3 Invariant polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . 338
8.1.4 The Chern–Weil Theory . . . . . . . . . . . . . . . . . . . . . . . . . 341
8.2 Important examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
8.2.1 The invariants of the torus T n . . . . . . . . . . . . . . . . . . . . . 346
8.2.2 Chern classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
8.2.3 Pontryagin classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
8.2.4 The Euler class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
vi CONTENTS

8.2.5 Universal classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354

8.3 Computing characteristic classes . . . . . . . . . . . . . . . . . . . . . . . . 360
8.3.1 Reductions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
8.3.2 The Gauss–Bonnet–Chern theorem . . . . . . . . . . . . . . . . . . . 366

9 Classical Integral Geometry 375

9.1 The integral geometry of real Grassmannians . . . . . . . . . . . . . . . . . 375
9.1.1 Co-area formulæ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
9.1.2 Invariant measures on linear Grassmannians . . . . . . . . . . . . . . 388
9.1.3 Affine Grassmannians . . . . . . . . . . . . . . . . . . . . . . . . . . 397
9.2 Gauss–Bonnet again?!? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
9.2.1 The shape operator and the second fundamental form . . . . . . . . 400
9.2.2 The Gauss–Bonnet theorem for hypersurfacesof an Euclidean space. 403
9.2.3 Gauss–Bonnet theorem for domains in an Euclidean space . . . . . . 408
9.3 Curvature measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
9.3.1 Tame geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
9.3.2 Invariants of the orthogonal group . . . . . . . . . . . . . . . . . . . 417
9.3.3 The tube formula and curvature measures . . . . . . . . . . . . . . . 421
9.3.4 Tube formula =⇒ Gauss–Bonnet formula . . . . . . . . . . . . . . . 432
9.3.5 Curvature measures of domains in an Euclidean space . . . . . . . . 434
9.3.6 Crofton formulæ for domains of an Euclidean space . . . . . . . . . 436
9.3.7 Crofton formulæ for submanifolds of an Euclidean space . . . . . . . 446

10 Elliptic Equations on Manifolds 455

10.1 Partial differential operators: algebraic aspects . . . . . . . . . . . . . . . . 455
10.1.1 Basic notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
10.1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
10.1.3 Formal adjoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
10.2 Functional framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
10.2.1 Sobolev spaces in RN . . . . . . . . . . . . . . . . . . . . . . . . . . 469
10.2.2 Embedding theorems: integrability properties . . . . . . . . . . . . . 476
10.2.3 Embedding theorems: differentiability properties . . . . . . . . . . . 481
10.2.4 Functional spaces on manifolds . . . . . . . . . . . . . . . . . . . . . 485
10.3 Elliptic partial differential operators: analytic aspects . . . . . . . . . . . . 489
10.3.1 Elliptic estimates in RN . . . . . . . . . . . . . . . . . . . . . . . . . 490
10.3.2 Elliptic regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
10.3.3 An application: prescribing the curvature of surfaces . . . . . . . . . 499
10.4 Elliptic operators on compact manifolds . . . . . . . . . . . . . . . . . . . . 509
10.4.1 Fredholm theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
10.4.2 Spectral theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
10.4.3 Hodge theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
CONTENTS 1

11 Spectral Geometry 529

11.1 Generalized functions and currents . . . . . . . . . . . . . . . . . . . . . . . 529
11.1.1 Generalized functions and operations withy them . . . . . . . . . . . 529
11.1.2 Currents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
11.1.3 Temperate distributions and the Fourier transform . . . . . . . . . . 535
11.1.4 Linear differential equations with distributional data . . . . . . . . . 537
11.2 Important families of generalized functions . . . . . . . . . . . . . . . . . . . 543
11.2.1 Some classical generalized functions on the real axis . . . . . . . . . 543
11.2.2 Homogeneous generalized functions . . . . . . . . . . . . . . . . . . . 549
11.3 The wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555
11.3.1 Fundamental solutions of the wave operator . . . . . . . . . . . . . . 555
11.3.2 The wave family . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
11.3.3 Local parametrices for the wave equation with variable coefficients . 575
11.4 Spectral geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
11.4.1 The spectral function of the Laplacian on a compact manifold . . . . 582
11.4.2 Short time asymptotics for the wave kernel . . . . . . . . . . . . . . 588
11.4.3 Spectral function asymptotics . . . . . . . . . . . . . . . . . . . . . . 597
11.4.4 Spectral estimates of smoothing operators . . . . . . . . . . . . . . . 602
11.4.5 Spectral perestroika . . . . . . . . . . . . . . . . . . . . . . . . . . . 606

12 Dirac Operators 619

12.1 The structure of Dirac operators . . . . . . . . . . . . . . . . . . . . . . . . 619
12.1.1 Basic definitions and examples . . . . . . . . . . . . . . . . . . . . . 619
12.1.2 Clifford algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 622
12.1.3 Clifford modules: the even case . . . . . . . . . . . . . . . . . . . . . 626
12.1.4 Clifford modules: the odd case . . . . . . . . . . . . . . . . . . . . . 630
12.1.5 A look ahead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631
12.1.6 The spin group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
12.1.7 The complex spin group . . . . . . . . . . . . . . . . . . . . . . . . . 642
12.1.8 Low dimensional examples . . . . . . . . . . . . . . . . . . . . . . . . 644
12.1.9 Dirac bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
12.2 Fundamental examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653
12.2.1 The Hodge–DeRham operator . . . . . . . . . . . . . . . . . . . . . . 653
12.2.2 The Hodge–Dolbeault operator . . . . . . . . . . . . . . . . . . . . . 658
12.2.3 The spin Dirac operator . . . . . . . . . . . . . . . . . . . . . . . . . 664
12.2.4 The spinc Dirac operator . . . . . . . . . . . . . . . . . . . . . . . . 669
2 CONTENTS
Chapter 1

Manifolds

1.1 Preliminaries
1.1.1 Space and Coordinatization
Mathematics is a natural science with a special modus operandi. It replaces concrete
natural objects with mental abstractions which serve as intermediaries. One studies the
properties of these abstractions in the hope they reflect facts of life. So far, this approach
proved to be very productive.
The most visible natural object is the Space, the place where all things happen. The
first and most important mathematical abstraction is the notion of number. Loosely
speaking, the aim of this book is to illustrate how these two concepts, Space and Number,
fit together.
It is safe to say that geometry as a rigorous science is a creation of ancient Greeks.
Euclid proposed a method of research that was later adopted by the entire mathematics.
We refer of course to the axiomatic method. He viewed the Space as a collection of points,
and he distinguished some basic objects in the space such as lines, planes etc. He then
postulated certain (natural) relations between them. All the other properties were derived
from these simple axioms.
Euclid’s work is a masterpiece of mathematics, and it has produced many interesting
results, but it has its own limitations. For example, the most complicated shapes one
could reasonably study using this method are the conics and/or quadrics, and the Greeks
certainly did this. A major breakthrough in geometry was the discovery of coordinates
by René Descartes in the 17th century. Numbers were put to work in the study of Space.
Descartes’ idea of producing what is now commonly referred to as Cartesian coor-
dinates is familiar to any undergraduate. These coordinates are obtained using a very
special method (in this case using three concurrent, pairwise perpendicular lines, each one
endowed with an orientation and a unit length standard. What is important here is that
they produced a one-to-one mapping

Euclidian Space → R3 , P 7−→ (x(P ), y(P ), z(P )).

We call such a process coordinatization. The corresponding map is called (in this case)

3
4 CHAPTER 1. MANIFOLDS

r
θ

Figure 1.1: Polar coordinates

Cartesian system of coordinates. A line or a plane becomes via coordinatization an alge-

braic object, more precisely, an equation.
In general, any coordinatization replaces geometry by algebra and we get a two-way
correspondence
Study of Space ←→ Study of Equations.
The shift from geometry to numbers is beneficial to geometry as long as one has efficient
tools do deal with numbers and equations. Fortunately, about the same time with the
introduction of coordinates, Isaac Newton created the differential and integral calculus
and opened new horizons in the study of equations.
The Cartesian system of coordinates is by no means the unique, or the most use-
ful coordinatization. Concrete problems dictate other choices. For example, the polar
coordinates represent another coordinatization of (a piece of the plane) (see Figure 1.1).

P 7→ (r(P ), θ(P )) ∈ (0, ∞) × (−π, π).

This choice is related to the Cartesian choice by the well known formulae
x = r cos θ y = r sin θ. (1.1.1)
A remarkable feature of (1.1.1) is that x(P ) and y(P ) depend smoothly upon r(P ) and
θ(P ).
As science progressed, so did the notion of Space. One can think of Space as a configu-
ration set, i.e., the collection of all possible states of a certain phenomenon. For example,
we know from the principles of Newtonian mechanics that the motion of a particle in the
ambient space can be completely described if we know the position and the velocity of the
particle at a given moment. The space associated with this problem consists of all pairs
(position, velocity) a particle can possibly have. We can coordinatize this space using six
functions: three of them will describe the position, and the other three of them will de-
scribe the velocity. We say the configuration space is 6-dimensional. We cannot visualize
this space, but it helps to think of it as an Euclidian space, only “roomier”.
1.1. PRELIMINARIES 5

There are many ways to coordinatize the configuration space of a motion of a particle,
and for each choice of coordinates we get a different description of the motion. Clearly, all
these descriptions must “agree” in some sense, since they all reflect the same phenomenon.
In other words, these descriptions should be independent of coordinates. Differential ge-
ometry studies the objects which are independent of coordinates.
The coordinatization process had been used by people centuries before mathematicians
accepted it as a method. For example, sailors used it to travel from one point to another
on Earth. Each point has a latitude and a longitude that completely determines its
position on Earth. This coordinatization is not a global one. There exist four domains
delimited by the Equator and the Greenwich meridian, and each of them is then naturally
coordinatized. Note that the points on the Equator or the Greenwich meridian admit two
different coordinatizations which are smoothly related.
The manifolds are precisely those spaces which can be piecewise coordinatized, with
smooth correspondence on overlaps, and the intention of this book is to introduce the
reader to the problems and the methods which arise in the study of manifolds. The next
section is a technical interlude. We will review the implicit function theorem which will
be one of the basic tools for detecting manifolds.

1.1.2 The implicit function theorem

We gather here, with only sketchy proofs, a collection of classical analytical facts. For
more details one can consult [34].
Let X and Y be two Banach spaces and denote by L(X, Y ) the space of bounded
linear operators X → Y . For example, if X = Rn , Y = Rm , then L(X, Y ) can be identified
with the space of m × n matrices with real entries. For any set S we will denote by 1S
the identity map S → S.

Definition 1.1.1. Let F : U ⊂ X → Y be a continuous function (U is an open subset of

X). The map F is said to be (Fréchet) differentiable at u ∈ U if there exists T ∈ L(X, Y )
such that
kF (u0 + h) − F (u0 ) − T hkY = o(khkX ) as h → 0,
i.e.,
1
lim kF (u0 + h) − F (u0 ) − T hkY = 0. ⊔
⊓
h→0 khkX

Loosely speaking, a continuous function is differentiable at a point if, near that point,
it admits a “best approximation” by a linear map.
When F is differentiable at u0 ∈ U , the operator T in the above definition is uniquely
determined by
d 1
Th = |t=0 F (u0 + th) = lim (F (u0 + th) − F (u0 )) .
dt t→0 t

We will use the notation T = Du0 F and we will call T the Fréchet derivative of F at u0 .
Assume that the map F : U → Y is differentiable at each point u ∈ U . Then F is said
to be of class C 1 , if the map u 7→ Du F ∈ L(X, Y ) is continuous. F is said to be of class
C 2 if u 7→ Du F is of class C 1 . One can define inductively C k and C ∞ (or smooth) maps.
6 CHAPTER 1. MANIFOLDS

Example 1.1.2. Consider F : U ⊂ Rn → Rm . Using Cartesian coordinates

x = (x1 , . . . , xn )
in Rn and
u = (u1 , . . . , um )
in Rm , we can think of F as a collection of m functions on U
u1 = u1 (x1 , . . . , xn ), . . . , um = um (x1 , . . . , xn ).
The map F is differentiable at a point p = (p1 , . . . , pn ) ∈ U if and only if the functions ui
are differentiable at p in the usual sense of calculus. The Fréchet derivative of F at p is
the linear operator Dp F : Rn → Rm given by the Jacobian matrix
 ∂u1 ∂u1 ∂u1 
∂x1 (p) ∂x2 (p) · · · ∂xn (p)
 
∂(u1 , . . . , um )   ∂u2 (p) ∂u2 (p) · · · ∂u2 (p) 

Dp F = =  ∂x 1 ∂x 2 ∂x n .
∂(x1 , . . . , xn )  .. .. .. .. 
 . . . . 
∂um ∂um ∂um
∂x1
(p) ∂x2 (p) · · · ∂xn (p)
The map F is smooth if and only if the functions ui (x) are smooth. ⊔
⊓

Exercise 1.1.3. (a) Let U ⊂ L(Rn , Rn ) denote the set of invertible n × n matrices. Show
that U is an open subset of L(Rn , Rn ).
(b) Let F : U → U be defined as A → A−1 . Show that DA F (H) = −A−1 HA−1 for any
n × n matrix H.
(c) Show that the Fréchet derivative of the map det : L(Rn , Rn ) → R, A 7→ det A, at
A = 1Rn ∈ L(Rn , Rn ) is given by tr H, i.e.,
d
|t=0 det(1Rn + tH) = tr H, ∀H ∈ L(Rn , Rn ). ⊔
⊓
dt

Theorem 1.1.4 (Inverse function theorem). Let X, Y be two Banach spaces, U ⊂ X open
and F : U ⊂ X → Y a smooth map. If at a point u0 ∈ U the derivative Du0 F ∈ L(X, Y )
is invertible, then there exits an open neighborhood U1 of u0 in U such that F (U1 ) is an
open neighborhood of v0 = F (u0 ) in Y and F : U1 → F (U1 ) is bijective, with smooth
inverse. ⊔
⊓

The spirit of the theorem is very clear: the invertibility of the derivative Du0 F “prop-
agates” locally to F because Du0 F is a very good local approximation for F .
More formally, if we set T = Du0 F , then
F (u0 + h) = F (u0 ) + T h + r(h),
where r(h) = o(khk) as h → 0. The theorem states that, for every v sufficiently close to
v0 , the equation F (u) = v has a unique solution u = u0 + h, with h very small. To prove
the theorem one has to show that, for kv − v0 kY sufficiently small, the equation below
v0 + T h + r(h) = v
1.2. SMOOTH MANIFOLDS 7

has a unique solution. We can rewrite the above equation as T h = v − v0 − r(h) or,
equivalently, as h = T −1 (v − v0 − r(h)). This last equation is a fixed point problem that
can be approached successfully via the Banach fixed point theorem.
Theorem 1.1.5 (Implicit function theorem). Let X, Y , Z be Banach spaces, U ⊂ X,
V ⊂ Y open sets and
F :U ×V →Z
a smooth map. Let (x0 , y0 ) ∈ U × V , and z0 := F (x0 , y0 ). Set
F2 : V → Z, F2 (y) = F (x0 , y).
Assume that Dy0 F2 ∈ L(Y, Z) is invertible. Then there exist open neighborhoods U ⊂ U
of x0 in X, V ⊂ V of y0 in Y , and a smooth map G : U → V such that the set S of
solution (x, y) of the equation F (x, y) = z0 which lie inside U × V can be identified with
the graph of G, i.e.,
 
(x, y) ∈ U × V ; F (x, y) = z0 = (x, G(x)) ∈ U × V ; x ∈ U .
In pre-Bourbaki times, the classics regarded the coordinate y as a function of x defined
implicitly by the equality F (x, y) = z0 .
Proof. Consider the map
H : X × Y → X × Z, ξ = (x, y) 7→ (x, F (x, y)).
The map H is a smooth map, and at ξ0 = (x0 , y0 ) its derivative Dξ0 H : X × Y → X × Z
has the block decomposition
 
1X 0
Dξ0 H = .
Dξ0 F1 Dξ0 F2
Above, DF1 (respectively DF2 ) denotes the derivative of x 7→ F (x, y0 ) (respectively the
derivative of y 7→ F (x0 , y)). The linear operator Dξ0 H is invertible, and its inverse has
the block decomposition
 
1X 0
(Dξ0 H)−1 =  .
−1 −1
− (Dξ0 F2 ) ◦ (Dξ0 F1 ) (Dξ0 F2 )
Thus, by the inverse function theorem, the equation (x, F (x, y)) = (x, z0 ) has a unique
solution (x̃, ỹ) = H −1 (x, z0 ) in a neighborhood of (x0 , y0 ). It obviously satisfies x̃ = x and
F (x̃, ỹ) = z0 . Hence, the set {(x, y); F (x, y) = z0 } is locally the graph of x 7→ H −1 (x, z0 ).
⊔
⊓

1.2 Smooth manifolds

1.2.1 Basic definitions
We now introduce the object which will be the main focus of this book, namely, the
concept of (smooth) manifold. It formalizes the general principles outlined in Subsection
1.1.1.
8 CHAPTER 1. MANIFOLDS

Uj
Ui

ψi ψ
j

ψ ψ −1
j i
m
m R
R

Figure 1.2: Transition maps

Definition 1.2.1. A smooth manifold of dimension m is a locally compact, second count-

able1 Hausdorff space M together with the following collection of data (henceforth called
atlas or smooth structure) consisting of the following.

(a) An open cover {Ui }i∈I of M ;

(b) A collection of homeomorphisms


Ψi : Ui → Ψi (Ui ) ⊂ Rm ; i ∈ I

(called charts or local coordinates) such that, Ψi (Ui ) is open in Rm and, if Ui ∩Uj 6=
∅, then the transition map

Ψj ◦ Ψ−1 m
i : Ψi (Ui ∩ Uj ) ⊂ R → Ψj (Ui ∩ Uj ) ⊂ R
m

⊔
is smooth. (We say that the various charts are smoothly compatible; see Figure 1.2).⊓

Remark 1.2.2. (a) Each chart Ψi : Ui → Rm can be viewed as a collection of m functions

(x1 , . . . , xm ) on Ui ,
 1 
x (p)
 x2 (p) 
 
Ψi (p) =  .. .
 . 
xm (p)
1
A second countable space is a topological space that admits a countable basis of open sets.
1.2. SMOOTH MANIFOLDS 9

Similarly, we can view another chart Ψj as another collection of functions (y 1 , . . . , y m ).

The transition map Ψj ◦ Ψ−1 i can then be interpreted as a collection of maps


(x1 , . . . , xm ) 7→ y 1 (x1 , . . . , xm ), . . . , y m (x1 , . . . , xm ) .

(b) Since a manifold is a second countable space we can always work with atlases that are
at most countable. ⊔
⊓

The first and the most important example of manifold is Rn itself. The natural smooth
structure consists of an atlas with a single chart, 1Rn : Rn → Rn . To construct more
examples we will use the implicit function theorem .
Definition 1.2.3. (a) Let M , N be two smooth manifolds of dimensions m and respec-
tively n. A continuous map f : M → N is said to be smooth if, for any local charts φ
on M and ψ on N , the composition ψ ◦ f ◦ φ−1 (whenever this makes sense) is a smooth
map Rm → Rn .
(b) A smooth map f : M → N is called a diffeomorphism if it is invertible and its inverse
is also a smooth map. ⊔
⊓

Example 1.2.4. The map t 7→ et is a diffeomorphism (−∞, ∞) → (0, ∞). The map
t 7→ t3 is a homeomorphism R → R, but it is not a diffeomorphism! ⊔
⊓

If M is a smooth m-dimensional manifold, we will denote by C ∞ (M ) the linear space

of all smooth functions f : M → R. Let us point out a simple procedure that we will use
frequently in the sequel. Suppose that f : M → R is a smooth function. If (U, Ψ) is a
local chart on M so Ψ(U ) is an open subset of Rm , then, by definition, the composition
f ◦ Ψ−1 : Ψ(U ) → R is a smooth function on the open set Ψ(U ) ⊂ Rn . If we denote
by x1 , . . . , xm the canonical Euclidean coordinates on Rm , then ψ ◦ Ψ−1 is a function

depending on the m variables x1 , . . . , xm and we will use the notation f x1 , . . . , xm when
referring to this function.
Remark 1.2.5. Let U be an open subset of the smooth manifold M (dim M = m) and

Ψ : U → Rm

a smooth, one-to-one map with open image and smooth inverse. Then Ψ defines local
coordinates over U compatible with the existing atlas of M . Thus (U, Ψ) can be added
to the original atlas and the new smooth structure is diffeomorphic with the initial one.
Using Zermelo’s Axiom we can produce a maximal atlas (no more compatible local chart
can be added to it). ⊔
⊓

Our next result is a general recipe for producing manifolds. Historically, this is how
manifolds entered mathematics.
Proposition 1.2.6. Let M be a smooth manifold of dimension m and f1 , . . . , fk ∈
C ∞ (M ). Define

Z = Z(f1 , . . . , fk ) = p ∈ M ; f1 (p) = · · · = fk (p) = 0 .
10 CHAPTER 1. MANIFOLDS

Assume that the functions f1 , . . . , fk are functionally independent along Z, i.e., for each
p ∈ Z, there exist local coordinates (x1 , . . . , xm ) defined in a neighborhood of p in M such
that xi (p) = 0, i = 1, . . . , m, and the matrix
 ∂f1 ∂f1 ∂f1 
1 2 · · · ∂xm
∂ f~ ∂x
 ..
∂x
. . .. 
|p :=  . .
. .
. . 
∂~x ∂fk ∂fk ∂fk
∂x1 ∂x2 · · · ∂xm x1 =···=xm =0

has rank k. Then Z has a natural structure of smooth manifold of dimension m − k.

Proof. Step 1: Constructing the charts. Let p0 ∈ Z, and denote by (x1 , . . . , xm ) local
coordinates near p0 such that xi (p0 ) = 0. One of the k × k minors of the matrix
 ∂f1 ∂f1 ∂f1 
∂x1 ∂x2
· · · ∂x m
∂f~  ..
|p :=  . .
.. .
.. .. 
.

∂~x ∂fk ∂fk ∂fk
∂x1 ∂x2 · · · ∂xm x1 =···=xm =0

is nonzero. Assume this minor is determined by the last k columns (and all the k lines).
We can think of the functions f1 , . . . , fk as defined on an open subset U of Rm . Split
Rm as Rm−k × Rk , and set

x′ := (x1 , . . . , xm−k ), x′′ := (xm−k+1 , . . . , xm ).

We are now in the setting of the implicit function theorem with

X = Rm−k , Y = Rk , Z = Rk ,

and F : X × Y → Z given by
 
f1 (x)
 .. 
x 7→  .  ∈ Rk .
fk (x))
∂F


In this case, DF2 = ∂x ′′ : Rk → Rk is invertible since its determinant corresponds to
our nonzero minor. Thus, in a product neighborhood Up0 = Up′ 0 × Up′′0 of p0 , the set Z is
the graph of some function

g : Up′ 0 ⊂ Rm−k −→ Up′′0 ⊂ Rk ,

i.e., 
Z ∩ Up 0 = (x′ , g(x′ ) ) ∈ Rm−k × Rk ; x′ ∈ Up′ 0 , |x′ | small .
We now define ψp0 : Z ∩ Up0 → Rm−k by
ψp
( x′ , g(x′ ) ) 7−→
0
x′ ∈ Rm−k .

The map ψp0 is a local chart of Z near p0 .

Step 2. The transition maps for the charts constructed above are smooth. The details
are left to the reader. ⊔
⊓
1.2. SMOOTH MANIFOLDS 11

Exercise 1.2.7. Complete Step 2 in the proof of Proposition 1.2.6. ⊔

⊓

Definition 1.2.8. Let M be a m-dimensional manifold. A codimension k submanifold

of M is a subset S ⊂ M locally defined as the common zero locus of k functionally
independent functions f1 , . . . , fk ∈ C ∞ (M ). More precisely, this means that, for any
p0 ∈ S, there exists an open neighborhood U of p0 ∈ M and k functionally independent
smooth functions
f1 , . . . , fk : U → R
such that p ∈ U ∩ S if and only if f1 (p) = · · · = fk (p) = 0. ⊔
⊓

Proposition 1.2.6 shows that any submanifold N ⊂ M has a natural smooth structure
so it becomes a manifold per se. Moreover, the inclusion map i : N ֒→ M is smooth.
Exercise 1.2.9. Suppose that M is a smooth m-dimensional manifold. Prive that S ⊂ M
is a codimension k-submanifold of M if and only if, for any p0 ∈ M , there exists a
coordinate chart (U, Ψ) with local coordinates (x1 , . . . , xm ) such that p0 ∈ U and

U ∩S = p ∈ U ; x1 (p) = · · · = xk (p) = 0 . ⊔
⊓

1.2.2 Partitions of unity

This is a very brief technical subsection describing a trick we will extensively use in this
book. Recall that manifolds are locally compact, second countable topological spaces. As
such, they are paracompact so they admit continuous partitions of unity; see [30, §3.7]. A
much more precise result is in fact true.
Definition 1.2.10. Let M be a smooth manifold and (Uα )α∈A an open cover of M . A
(smooth) partition of unity subordinated to this cover is a family (fβ )β∈B ⊂ C ∞ (M )
satisfying the following conditions.

(i) 0 ≤ fβ ≤ 1.

(ii) ∃φ : B → A such that supp fβ ⊂ Uφ(β) .

(iii) The family (supp fβ ) is locally finite, i.e., any point x ∈ M admits an open neigh-
borhood intersecting only finitely many of the supports supp fβ .
P
(iv) β fβ (x) = 1 for all x ∈ M . ⊔
⊓

We include here for the reader’s convenience the basic existence result concerning
partitions of unity. For a proof we refer to [128].
Proposition 1.2.11. (a) For any open cover U = (Uα )α∈A of a smooth manifold M there
exists at least one smooth partition of unity (fβ )β∈B subordinated to U such that supp fβ
is compact for any β.
(b) If we do not require compact supports, then we can find a partition of unity in which
B = A and φ = 1A . ⊔
⊓
12 CHAPTER 1. MANIFOLDS

Exercise 1.2.12. Let M be a smooth manifold and S ⊂ M a closed submanifold, i.e.,

S is a closed subset of M . Prove that the restriction map
r : C ∞ (M ) → C ∞ (S) f 7→ f |S
is surjective. Deduce that for any finite set X ⊂ M and any function g : F → R there
exists a smooth, compactly supported gunction f : M → R such that f (x) = g(x), ∀x ∈ X.
⊔
⊓

1.2.3 Examples
Manifolds are everywhere, and in fact, to many physical phenomena which can be modelled
mathematically one can naturally associate a manifold. On the other hand, many problems
in mathematics find their most natural presentation using the language of manifolds. To
give the reader an idea of the scope and extent of modern geometry, we present here a
short list of examples of manifolds. This list will be enlarged as we enter deeper into the
study of manifolds.
Example 1.2.13. (The n-dimensional sphere) This is the codimension 1 submanifold of
Rn+1 given by the equation
n
X
2
|x| = (xi )2 = r 2 , x = (x0 , . . . , xn ) ∈ Rn+1 .
i=0

One checks that, along the sphere, the differential of |x|2 is nowhere zero, so by Proposition
1.2.6, S n is indeed a smooth manifold. In this case one can explicitly construct an atlas
(consisting of two charts) which is useful in many applications. The construction relies on
stereographic projections.
Let N and S denote the North and resp. South pole of S n (N = (0, . . . , 0, 1) ∈ Rn+1 ,
S = (0, . . . , 0, −1) ∈ Rn+1 ). Consider the open sets UN = S n \ {N } and US = S n \ {S}.
They form an open cover of S n . The stereographic projection from the North pole is the
map σN : UN → Rn such that, for any P ∈ UN , the point σN (P ) is the intersection of the
line N P with the hyperplane {xn = 0} ∼ = Rn .
The stereographic projection from the South pole is defined similarly.


• For P ∈ UN we denote by y 1 (P ), . . . , y n (P ) the coordinates of σN (P ).
• For Q ∈ US we denote by z 1 (Q), . . . , z n (Q) ) the coordinates of σS (Q).

A simple argument shows the map

y 1 (P ), . . . , y n (P ) 7→ z 1 (P ), . . . , z n (P ) , P ∈ UN ∩ US ,

is smooth (see the exercise below). Hence (UN , σN ), (US , σS ) defines a smooth structure
on S n . ⊔
⊓

Exercise 1.2.14. Show that the functions y i , z j constructed in the above example satisfy
yi
z i = P  , ∀i = 1, . . . , n. ⊔
⊓
n j 2
j=1 (y )
1.2. SMOOTH MANIFOLDS 13

Example 1.2.15. (The n-dimensional torus) This is the codimension n submanifold of the
vector space R2n with Cartesian coordinates (x1 , y1 ; ... ; xn , yn ), defined by the equalities

x21 + y12 = · · · = x2n + yn2 = 1.

Note that T 1 is diffeomorphic with the 1-dimensional sphere S 1 (unit circle). As a set T n
is a direct product of n circles (see Figure 1.3)
 
T n = x21 + y12 = 1 × x2n + yn2 = 1 = S 1 × · · · × S 1 . ⊔
⊓

Figure 1.3: The 2-dimensional torus

The above example suggests the following general construction.

Example 1.2.16. Let M and N be smooth manifolds of dimension m and respectively
n. Then their topological direct product has a natural structure of smooth manifold of
dimension m + n. ⊔
⊓

Example 1.2.17. (The connected sum of two manifolds) Let M1 and M2 be two manifolds
of the same dimension m. Pick pi ∈ Mi (i = 1, 2), choose small open neighborhoods Ui of
pi in Mi and then local charts ψi identifying each of these neighborhoods with B2 (0), the
ball of radius 2 in Rm .
Let Vi ⊂ Ui correspond (via ψi ) to the annulus {1/2 < |x| < 2} ⊂ Rm . Consider
  x
φ: 1/2 < |x| < 2 → 1/2 < |x| < 2 , φ(x) = .
|x|2

The action of φ is clear: it switches the two boundary components of {1/2 < |x| < 2},
and reverses the orientation of the radial directions.
Now “glue” V1 to V2 using the “prescription” given by ψ2−1 ◦ φ ◦ ψ1 : V1 → V2 . In this
way we obtain a new topological space with a natural smooth structure induced by the
smooth structures on Mi . Up to a diffeomeorphism, the new manifold thus obtained is
independent of the choices of local coordinates ([24]), and it is called the connected sum
of M1 and M2 and is denoted by M1 #M2 (see Figure 1.4). ⊔
⊓

Example 1.2.18. (The real projective space RPn ) As a topological space RPn is the
quotient of Rn+1 modulo the equivalence relation
def
x ∼ y ⇐⇒ ∃λ ∈ R∗ : x = λy.
14 CHAPTER 1. MANIFOLDS

V V
1 2

Figure 1.4: Connected sum of tori

The equivalence class of x = (x0 , . . . , xn ) ∈ Rn+1 \ {0} is usually denoted by [x0 , . . . , xn ].

Alternatively, RPn is the set of all lines (directions) in Rn+1 . Traditionally, one attaches
a point to each direction in Rn+1 , the so called “point at infinity” along that direction, so
that RPn can be thought as the collection of all points at infinity along all the directions
in Rn+1 .
The space RPn has a natural structure of smooth manifold. To describe it consider
the sets

Uk = [x0 , . . . , xn ] ∈ RPn ; xk 6= 0 , k = 0, . . . , n.
Now define

ψk : Uk → Rn [x0 , . . . , xn ] 7→ (x0 /xk , . . . , xk−1 /xk , xk+1 /xk , . . . xn ).

The maps ψk define local coordinates on the projective space. The transition map on the
overlap region Uk ∩ Um = {[x0 , . . . , xn ] ; xk xm 6= 0} can be easily described. Set

ψk ([x0 , . . . , xn ]) = (ξ1 , . . . , ξn ), ψm ([x0 , . . . , xn ]) = (η1 , . . . , ηn ).

The equality

[x0 , . . . , xn ] = [ξ1 , . . . , ξk−1 , 1, ξk , . . . , ξn ] = [η1 , . . . , ηm−1 , 1, ηm , . . . , ηn ]

immediately implies (assume k < m)


 ξ1 = η1 /ηk , . . . , ξk−1 = ηk−1 /ηk , ξk+1 = ηk
ξk = ηk+1 /ηk , . . . , ξm−2 = ηm−1 /ηk , ξm−1 = 1/ηk (1.2.1)

ξm = ηm ηk , . . . , ξn = ηn /ηk
1.2. SMOOTH MANIFOLDS 15

This shows the map ψk ◦ ψm −1 is smooth and proves that RPn is a smooth manifold. Note
1
that when n = 1, RP is diffeomorphic with S 1 . One way to see this is to observe that
the projective space can be alternatively described as the quotient space of S n modulo the
equivalence relation which identifies antipodal points . ⊔
⊓

Example 1.2.19. (The complex projective space CPn ) The definition is formally identical
to that of RPn . CPn is the quotient space of Cn+1 \ {0} modulo the equivalence relation
def
x ∼ y ⇐⇒ ∃λ ∈ C∗ : x = λy.

The open sets Uk are defined similarly and so are the local charts ψk : Uk → Cn . They
satisfy transition rules similar to (1.2.1) so that CPn is a smooth manifold of dimension
2n. ⊔
⊓

Exercise 1.2.20. Prove that CP1 is diffeomorphic to S 2 . ⊔

⊓

In the above example we encountered a special (and very pleasant) situation: the
−1 : U → V
gluing maps not only are smooth, they are also holomorphic as maps ψk ◦ ψm
n
where U and V are open sets in C . This type of gluing induces a “rigidity” in the
underlying manifold and it is worth distinguishing this situation.
Definition 1.2.21 (Complex manifolds). A complex manifold is a smooth, 2n-dimensional
manifold M which admits an atlas {(Ui , ψi ) : Ui → Cn } such that all transition maps are
holomorphic. ⊔
⊓

The complex projective space is a complex manifold. Our next example naturally
generalizes the projective spaces described above.
Example 1.2.22. (The real and complex Grassmannians Grk (Rn ), Grk (Cn ))
Suppose V is a real vector space of dimension n. For every 0 ≤ k ≤ n we denote by
Grk (V ) the set of k-dimensional vector subspaces of V . We will say that Grk (V ) is the
linear Grassmannian of k-planes in E. When V = Rn we will write Grk,n (R) instead of
Grk (Rn ).
We would like to give several equivalent descriptions of the natural structure of smooth
manifold on Grk (V ). To do this it is very convenient to fix an Euclidean metric on V .
Any k-dimensional subspace L ⊂ V is uniquely determined by the orthogonal projec-
tion onto L which we will denote by PL . Thus we can identify Grk (V ) with the set of
rank k projectors

Projk (V ) := P : V → V ; P ∗ = P = P 2 , rank P = k .

Let use observe that the rank of an orthogonal projector is determined by the equality

rank P = tr P.

We have a natural map

P : Grk (V ) → Projk (V ), L 7→ PL
16 CHAPTER 1. MANIFOLDS

with inverse P 7→ Range (P ).

The set Projk (V ) is a subset of the vector space of symmetric endomorphisms

End+ (V ) := A ∈ End(V ), A∗ = A .

The space End+ (V ) is equipped with a natural inner product

1
(A, B) := tr(AB), ∀A, B ∈ End+ (V ). (1.2.2)
2
We denote by k − k the norm on End+ (V ) induced by this inner product,
1
kAk2 = tr(A2 ).
2
Note that the subset Projk (V ) ⊂ End+ (V ) can alternatively be described by the equalities

P 2 = P, tr P = k.

This proves that Projk (V ) is a closed subset of End+ (V ). From the equality

1 1 k
kP k2 = tr P 2 = tr P = , ∀P ∈ Projk (V )
2 2 2
we deduce that Projk (V ) is also bounded subset of End+ (V ). The bijection

P : Grk (V ) → Projk (V ), L 7→ PL

induces a topology on Grk (V ), and with this topology Grk (V ) is a compact metric space.
We want to show that Grk (V ) has a natural structure of smooth manifold compatible
with this topology. To see this, we define for every L ⊂ Grk (V ) the set

Grk (V, L) := U ∈ Grk (V ); U ∩ L⊥ = 0 .

Lemma 1.2.23. (a) Let L ∈ Grk (V ). Then

U ∩ L⊥ = 0 ⇐⇒ 1 − PL + PU : V → V is an isomorphism. (1.2.3)

(b) The set Grk (V, L) is an open subset of Grk (V ).

Proof. (a) Note first that a dimension count implies that

U ∩ L⊥ = 0 ⇐⇒ U + L⊥ = V ⇐⇒ U ⊥ ∩ L = 0.

Let us show that U ∩ L⊥ = 0 implies that 1 − PL + PL is an isomorphism. It suffices to

show that
ker(1 − PL + PU ) = 0.
Suppose v ∈ ker(1 − PL + PU ). Then

0 = PL (1 − PL + PU )v = PL PU v = 0 =⇒ PU v ∈ U ∩ ker PL = U ∩ L⊥ = 0.
1.2. SMOOTH MANIFOLDS 17

Hence PU v = 0, so that v ∈ U ⊥ . From the equality (1 − PL − PU )v = 0 we also deduce

(1 − PL )v = 0 so that v ∈ L. Hence

v ∈ U ⊥ ∩ L = 0.

Conversely, we will show that if 1 − PL + PU = PL⊥ + PU is onto, then U + L⊥ = V .

Indeed, let v ∈ V . Then there exists x ∈ V such that

v = PL⊥ x + PU x ∈ L⊥ + U.

(b) We have to show that, for every K ∈ Grk (V, L), there exists ε > 0 such that any U
satisfying
kPU − PK k < ε
intersects L⊥ trivially. Since K ∈ Grk (V, L) we deduce from (a) that the map 1 −PL −PK :
V → V is an isomorphism. Note that

k(1 − PL − PK ) − (1 − PL − PU )k = kPK − PU k.

The space of isomorphisms of V is an open subset of End(V ). Hence there exists ε > 0
such that, for any subspace U satisfying kPU − PK k < ε, the endomorphism (1 − PL − PU )
is an isomorphism. We now conclude using part (a). ⊔
⊓

Since L ∈ Grk (V, L), ∀L ∈ Grk (V ), the collection

n o
Grk (V, L); L ∈ Grk (V )

is an open cover of Grk (V ). Note that for every L ∈ Grk (V ) we have a natural map

Γ : Hom(L, L⊥ ) → Grk (V, L), (1.2.4)

that associates to each linear map S : L → L⊥ its graph (see Figure 1.5)

ΓS = {x + Sx ∈ L + L⊥ = V ; x ∈ L}.

The map (1.2.4) is obviously injective. We claim that it is in fact a bijection.

Indeed, if U ∈ Grk (V, L), then the restriction of PL to U is injective since

U ∩ ker PL = U ∩ L⊥ = 0.

Thus, the linear map PL U : U → L, U ∋ U u → 7 PL u is a linear isomorphism because

dim U = dim L = k. Denote by HU its inverse, HU : L → U . It is easy to see that U is
the graph of the linear map

S : L → L⊥ , Sx = PL⊥ HU x.

We will show that the bijection (1.2.4) is a homeomorphism. We first prove that it is
continuous by providing an explicit description of the orthogonal projection PΓS
18 CHAPTER 1. MANIFOLDS

v
v Γ
L S

Sx

v x L
L

Figure 1.5: Subspaces as graphs of linear operators.

Observe first that the orthogonal complement of ΓS is the graph of −S ∗ : L⊥ → L.

More precisely,

Γ⊥ ∗ ⊥
S = Γ−S ∗ = y − S y ∈ L + L = V ; y ∈ L
⊥
.
Let v = PL v + PL⊥ v = vL + vL+ ∈ V (see Figure 1.5). Then

PΓS v = x + Sx, x ∈ L ⇐⇒ v − (x + Sx) ∈ Γ⊥

S

⊥ x + S ∗ y = vL
⇐⇒ ∃x ∈ L, y ∈ L such that .
Sx − y = vL⊥

Consider the operator S : L ⊕ L⊥ → L ⊕ L⊥ which has the block decomposition

 
1L S∗
S= .
S −1⊥ L

Then the above linear system can be rewritten as

   
x vL
S· = .
y vL ⊥

Now observe that  

1L + S ∗ S 0
S2 = .
0 1L⊥ + SS ∗
Hence S is invertible, and
 
−1 (1L + S ∗ S)−1 0
S = ·S
0 (1L⊥ + SS ∗ )−1
 
(1L + S ∗ S)−1 (1L + S ∗ S)−1 S ∗
= .
(1L⊥ + SS ∗ )−1 S −(1L⊥ + SS ∗ )−1
We deduce
x = (1L + S ∗ S)−1 vL + (1L + S ∗ S)−1 S ∗ vL⊥
1.2. SMOOTH MANIFOLDS 19

and,  
x
PΓS v = .
Sx
Hence PΓS has the block decomposition
 
1L  
PΓS = · (1L + S ∗ S)−1 (1L + S ∗ S)−1 S ∗
S
 
(1L + S ∗ S)−1 (1L + S ∗ S)−1 S ∗
= . (1.2.5)
S(1L + S ∗ S)−1 S(1L + S ∗ S)−1 S ∗
This proves that the map

Hom(L, L⊥ ) ∋ S 7→ PΓS ∈ Grk (V, L)

is continuous. Note that if U ∈ Grk (V, L), then with respect to the decomposition V =
L + L⊥ the projector PU has the block form
 
  PL PU IL PL PU IL⊥
A B
PU = = ,
C D
PL⊥ PU IL PL⊥ PU IL⊥

where for every subspace K ֒→ V we denoted by IK : K → V the canonical inclusion,

then U = ΓS . If U = ΓS , then (1.2.5) implies

PL PU IL = A(U ) = (1L + S ∗ S)−1 , PL PU IL⊥ = C(U ) = S(1L + S ∗ S)−1 ,

so S = CA−1 . Since A and C depend continuously on PU we deduce shows that the

inverse of the graph map

Hom(L, L⊥ ) ∋ S 7→ ΓS ∈ Grk (V )

is also continuous. Moreover, the above formulæ show that if U ∈ Grk (V, L0 )∩Grk (V, L1 ),
then we can represent U in two ways,

U = ΓS0 = ΓS1 , Si ∈ Hom(Li , L⊥

i ), i = 0, 1,

and the correspondence S0 → S1 is smooth. This shows that Grk (V ) has a natural
structure of smooth manifold of dimension

dim Grk (V ) = dim Hom(L, L⊥ ) = k(n − k).

Grk (Cn ) is defined as the space of complex k-dimensional subspaces of Cn . It can be

structured as above as a smooth manifold of dimension 2k(n − k). Note that Gr1 (Rn ) ∼
=
RPn−1 , and Gr1 (Cn ) ∼
= CPn−1 . The Grassmannians have important applications in many
classification problems. ⊔
⊓
20 CHAPTER 1. MANIFOLDS

Exercise 1.2.24. Show that Grk (Cn ) is a complex manifold of complex dimension k(n −
k)). ⊔
⊓

Example 1.2.25. (Lie groups). A Lie group is a smooth manifold G together with a
group structure on it such that the map

G×G→G (g, h) 7→ g · h−1

is smooth. These structures provide an excellent way to formalize the notion of symmetry.

(a) (Rn , +) is a commutative Lie group.

(b) The unit circle S 1 can be alternatively described as the set of complex numbers of
norm one and the complex multiplication defines a Lie group structure on it. This is a
commutative group. More generally, the torus T n is a Lie group as a direct product of n
circles2 .
(c) The general linear group GL(n, K) defined as the group of invertible n × n matrices
with entries in the field K = R, C is a Lie group. Indeed, GL(n, K) is an open subset (see
Exercise 1.1.3) in the linear space of n × n matrices with entries in K. It has dimension
dK n2 , where dK is the dimension of K as a linear space over R. We will often use the
alternate notation GL(Kn ) when referring to GL(n, K).
(d) The orthogonal group O(n) is the group of real n × n matrices satisfying

T · T t = 1.

To describe its smooth structure we will use the Cayley transform trick as in [113] (see
also the classical [133]). Set

Mn (R)# := T ∈ Mn (R) ; det(1 + T ) 6= 0 .

The matrices in Mn (R)# are called non exceptional. Clearly 1 ∈ O(n)# = O(n)∩Mn (R)#
so that O(n)# is a nonempty open subset of O(n). The Cayley transform is the map
# : Mn (R)# → Mn (R) defined by

A 7→ A# = (1 − A)(1 + A)−1 .

The Cayley transform has some very nice properties.

(i) A# ∈ Mn (R)# for every A ∈ Mn (R)# .
(ii) # is involutive, i.e., (A# )# = A for any A ∈ Mn (R)# .
(iii) For every T ∈ O(n)# the matrix T # is skew-symmetric, and conversely, if A ∈
Mn (R)# is skew-symmetric then A# ∈ O(n).
Thus the Cayley transform is a homeomorphism from O(n)# to the space of non-
exceptional, skew-symmetric, matrices. The latter space is an open subset in the linear
space of real n × n skew-symmetric matrices, o(n).
Any T ∈ O(n) defines a self-homeomorphism of O(n) by left translation in the group

LT : O(n) → O(n) S 7→ LT (S) = T · S.

2
One can show that any connected commutative Lie group has the from T n × Rm .
1.2. SMOOTH MANIFOLDS 21

We obtain an open cover of O(n):

[
O(n) = T · O(n)# .
T ∈O(n)

Define ΨT : T · O(n)# → o(n) by S 7→ (T −1 · S)# . One can show that the collection
 
T · O(n)# , ΨT
T ∈O(n)

defines a smooth structure on O(n). In particular, we deduce

dim O(n) = n(n − 1)/2.

Inside O(n) lies a normal subgroup (the special orthogonal group)

SO(n) = {T ∈ O(n) ; det T = 1}.

The group SO(n) is a Lie group as well and dim SO(n) = dim O(n).
(e) The unitary group U (n) is defined as

U (n) = {T ∈ GL(n, C) ; T · T ∗ = 1},

where T ∗ denotes the conjugate transpose (adjoint) of T . To prove that U (n) is a manifold
one uses again the Cayley transform trick. This time, we coordinatize the group using the
space u(n) of skew-adjoint (skew-Hermitian) n × n complex matrices (A = −A∗ ). Thus
U (n) is a smooth manifold of dimension

dim U (n) = dim u(n) = n2 .

Inside U (n) sits the normal subgroup SU (n), the kernel of the group homomorphism
det : U (n) → S 1 . SU (n) is also called the special unitary group. This a smooth manifold
of dimension n2 − 1. In fact the Cayley transform trick allows one to coordinatize SU (n)
using the space
su(n) = {A ∈ u(n) ; tr A = 0}. ⊔
⊓
Exercise 1.2.26. (a)
Prove the properties (i)-(iii) of the Cayley transform, and then show
#
that T · O(n) , ΨT T ∈O(n) defines a smooth structure on O(n).
(b) Prove that U (n) and SU (n) are manifolds.
(c) Show that O(n), SO(n), U (n), SU (n) are compact spaces.
(d) Prove that SU (2) is diffeomorphic with S 3 (Hint: think of S 3 as the group of unit
quaternions.) ⊔
⊓

Exercise 1.2.27. Let SL(n; K) denote the group of n × n matrices of determinant 1 with
entries in the field K = R, C. Use the implicit function theorem to show that SL(n; K) is
a smooth manifold of dimension dK (n2 − 1), where dK = dimR K. ⊔
⊓
22 CHAPTER 1. MANIFOLDS

Exercise 1.2.28. (Quillen). Suppose V0 , V1 are two real, finite dimensional Euclidean
space, and T : V0 → V1 is a linear map. We denote by T ∗ is adjoint, T ∗ : V1 → V0 , and by
ΓT the graph of T , 
ΓT = (v0 , v1 ) ∈ V0 ⊕ V1 ; v1 = T v0 l .
We form the skew-symmetric operator
     
v0 0 T∗ v0
X : V0 ⊕ V1 → V0 ⊕ V1 , X · = · .
v1 −T 0 v1

We denote by CT the Cayley transform of X,

CT = (1 − X)(1 + X)−1 ,

and by R0 : V0 ⊕ V1 → V0 ⊕ V1 the reflection

 
1V0 0
R0 = .
0 −1V1

Show that RT = CT R0 is an orthogonal involution, i.e.,

RT2 = 1, RT∗ = RT ,

and ker(1 − RT ) = ΓT . In other words, RT is the orthogonal reflection in the subspace

ΓT ,
RT = 2PΓT − 1,
where PΓT denotes the orthogonal projection onto ΓT . ⊔
⊓

Exercise 1.2.29. Suppose G is a Lie group, and H is an abstract subgroup of G. Prove

that the closure of H is also a subgroup of G.
Exercise 1.2.30. (a) Let G be a connected Lie group and denote by U a neighborhood
of 1 ∈ G. If H is the subgroup algebraically generated by U show that H is dense in G.
(b) Let G be a compact Lie group and g ∈ G. Show that 1 ∈ G lies in the closure of
{gn ; n ∈ Z \ {0}}. ⊔
⊓

Remark 1.2.31. If G is a Lie group, and H is a closed subgroup of G, then H is in fact

a smooth submanifold of G, and with respect to this smooth structure H is a Lie group.
For a proof we refer to [61, 128]. In view of Exercise 1.2.29, this fact allows us to produce
many examples of Lie groups. ⊔
⊓

1.2.4 How many manifolds are there?

The list of examples in the previous subsection can go on for ever, so one may ask whether
there is any coherent way to organize the collection of all possible manifolds. This is too
general a question to expect a clear cut answer. We have to be more specific. For example,
we can ask
1.2. SMOOTH MANIFOLDS 23

Question 1: Which are the compact, connected manifolds of a given dimension d?

For d = 1 the answer is very simple: the only compact connected 1-dimensional man-
ifold is the circle S 1 . (Can you prove this?)
We can raise the stakes and try the same problem for d = 2. Already the situation is
more elaborate. We know at least two surfaces: the sphere S 2 and the torus T 2 . They
clearly look different but we have not yet proved rigorously that they are indeed not
diffeomorphic. This is not the end of the story. We can connect sum two tori, three tori
or any number g of tori. We obtain doughnut-shaped surface as in Figure 1.6

Figure 1.6: Connected sum of 3 tori

Again we face the same question: do we get non-diffeomorphic surfaces for different
choices of g? Figure 1.6 suggests that this may be the case but this is no rigorous argument.
We know another example of compact surface, the projective plane RP2 , and we nat-
urally ask whether it looks like one of the surfaces constructed above. Unfortunately,
we cannot visualize the real projective plane (one can prove rigorously it does not have
enough room to exist inside our 3-dimensional Universe). We have to decide this question
using a little more than the raw geometric intuition provided by a picture. To kill the
suspense, we mention that RP2 does not belong to the family of donuts. One reason is
that, for example, a torus has two faces: an inside face and an outside face (think of a car
rubber tube). RP2 has a weird behavior: it has “no inside” and “no outside”. It has only
one side! One says the torus is orientable while the projective plane is not.
We can now connect sum any numbers of RP2 ’s to any donut an thus obtain more
and more surfaces, which we cannot visualize and we have yet no idea if they are pairwise
distinct. A classical result in topology says that all compact surfaces can be obtained in
this way (see [90]), but in the above list some manifolds are diffeomorphic, and we have
to describe which. In dimension 3 things are not yet settled3 and, to make things look
hopeless, in dimension ≥ 4 Question 1 is algorithmically undecidable .
We can reconsider our goals, and look for all the manifolds with a given property X.
In many instances one can give fairly accurate answers. Property X may refer to more
than the (differential) topology of a manifold. Real life situations suggest the study of
manifolds with additional structure. The following problem may give the reader a taste
of the types of problems we will be concerned with in this book.
Question 2 Can we wrap a planar piece of canvas around a metal sphere in a one-
to-one fashion? (The canvas is flexible but not elastic).
3
Things are still not settled in 2007, but there has been considerable progress due to G. Perelman’s
proof of the Poincaré’s and Thursron’s conjectures.
24 CHAPTER 1. MANIFOLDS

A simple do-it-yourself experiment is enough to convince anyone that this is not pos-
sible. Naturally, one asks for a rigorous explanation of what goes wrong. The best ex-
planation of this phenomenon is contained in the celebrated Theorema Egregium (Golden
Theorem) of Gauss. Canvas surfaces have additional structure (they are made of a special
material), and for such objects there is a rigorous way to measure “how curved” are they.
One then realizes that the problem in Question 2 is impossible, since a (canvas) sphere is
curved in a different way than a plane canvas.
There are many other structures Nature forced us into studying them, but they may
not be so easily described in elementary terms.

A word to the reader. The next two chapters are probably the most arid in geometry
but, keep in mind that, behind each construction lies a natural motivation and, even if we
do not always have the time to show it to the reader, it is there, and it may take a while
to reveal itself. Most of the constructions the reader will have to “endure” in the next two
chapters constitute not just some difficult to “swallow” formalism, but the basic language
of geometry. It might comfort the reader during this less than glamorous journey to carry
in the back of his mind Hermann Weyl’s elegantly phrased advise.

“It is certainly regrettable that we have to enter into the purely formal aspect
in such detail and to give it so much space but, nevertheless, it cannot be
avoided. Just as anyone who wishes to give expressions to his thoughts with
ease must spend laborious hours learning language and writing, so here too the
only way we can lessen the burden of formulæ is to master the technique of
tensor analysis to such a degree that we can turn to real problems that concern
us without feeling any encumbrance, our object being to get an insight into
the nature of space [...]. Whoever sets out in quest of these goals must possess
a perfect mathematical equipment from the outset.”
H. Weyl: Space, Time, Matter.
Chapter 2

Natural Constructions on
Manifolds

The goal of this chapter is to introduce the basic terminology used in differential geometry.
The key concept is that of tangent space at a point which is a first order approximation of
the manifold near that point. We will be able to transport many notions in linear analysis
to manifolds via the tangent space.

2.1 The tangent bundle

2.1.1 Tangent spaces
We begin with a simple example which will serve as a motivation for the abstract defini-
tions.
Example 2.1.1. Consider the sphere

(S 2 ) : x2 + y 2 + z 2 = 1 in R3 .

We want to find the plane passing through the North pole N (0, 0, 1) that is “closest” to
the sphere. The classics would refer to such a plane as an osculator plane.
The natural candidate for this osculator plane would be a plane given by a linear
equation that best approximates the defining equation x2 + y 2 + z 2 = 1 in a neighborhood
of the North pole. The linear approximation of x2 + y 2 + z 2 near N seems like the best
candidate. We have
x2 + y 2 + z 2 − 1 = 2(z − 1) + O(2),
where O(2) denotes a quadratic error. Hence, the osculator plane is z = 1, Geometrically,
it is the horizontal affine plane through the North pole. The linear subspace {z = 0} ⊂ R3
is called the tangent space to S 2 at N .
The above construction has one deficiency: it is not intrinsic, i.e., it relies on objects
“outside” the manifold S 2 . There is one natural way to fix this problem. Look at a smooth
path γ(t) on S 2 passing through N at t = 0. Hence, t 7→ γ(t) ∈ R3 , and

|γ(t)|2 = 1. (2.1.1)

25
26 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

If we differentiate (2.1.1) at t = 0 we get (γ̇(0), γ(0)) = 0, i.e., γ̇(0) ⊥ γ(0), so that γ̇(0) lies
in the linear subspace z = 0. We deduce that the tangent space consists of the tangents
to the curves on S 2 passing through N .
This is apparently no major conceptual gain since we still regard the tangent space
as a subspace of R3 , and this is still an extrinsic description. However, if we use the
stereographic projection from the South pole we get local coordinates (u, v) near N , and
any curve γ(t) as above can be viewed as a curve t 7→ (u(t), v(t)) in the (u, v) plane. If
φ(t) is another curve through N given in local coordinates by t 7→ (u(t), v(t)), then



γ̇(0) = φ̇(0) ⇐⇒ u̇(0), v̇(0) = u̇(0), v̇(0) .

The right hand side of the above equality defines an equivalence relation ∼ on the set of
smooth curves passing trough (0, 0). Thus, there is a bijective correspondence between the
tangents to the curves through N , and the equivalence classes of “∼”. This equivalence
relation is now intrinsic modulo one problem: “∼” may depend on the choice of the local
coordinates. Fortunately, as we are going to see, this is a non-issue. ⊔
⊓

Definition 2.1.2. Let M m be a smooth manifold and p0 a point in M . Two smooth

paths α, β : (−ε, ε) → M such that α(0) = β(0) = p0 are said to have a first order contact
at p0 if there exist local coordinates (x) = (x1 , . . . , xm ) near p0 such that

ẋα (0) = ẋβ (0),

where

α(t) = (xα (t)) := x1α (t), . . . , xm
α (t) ,

and

β(t) = (xβ (t)) = x1β (t), . . . , xm
β (t) .

We write this α ∼1 β. ⊔
⊓

Lemma 2.1.3. ∼1 is an equivalence relation.

Sketch of proof. The binary relation ∼1 is obviously reflexive and symmetric, so we
only have to check the transitivity. Let α ∼1 β and β ∼1 γ. Thus there exist local
coordinates (x) = (x1 , . . . , xm ) and (y) = (y 1 , . . . , y m ) near p0 such that




ẋα (0) = (ẋβ (0)) and ẏβ (0) = ẏγ (0) .

The transitivity follows from the equality

X ∂y i j X ∂y i
ẏγi (0) = ẏβi (0) = ẋ (0) = ẋj (0) = ẏαj (0). ⊔
⊓
∂xj β ∂xj α
j j

Definition 2.1.4. A tangent vector to M at p is a first-order-contact equivalence class

of curves through p. The equivalence class of a curve α(t) such that α(0) = p will be
temporarily denoted by [α̇(0)]. The set of these equivalence classes is denoted by Tp M ,
and is called the tangent space to M at p. ⊔
⊓
2.1. THE TANGENT BUNDLE 27

Lemma 2.1.5. Tp M has a natural structure of vector space.

Proof. Choose local coordinates (x1 , . . . , xm ) near p such that xi (p) = 0, ∀i, and let α and
β be two
smooth
curves through p. In the above local coordinates the curves α, β become
xiα (t) , xiβ (t) . Construct a new curve γ through p whose x coordinates are given by



xiγ (t) = xiα (t) + xiβ (t) .

We want to emphasize that the above curve depends on α, β and the choice of local
coordinates (x1 , . . . , xm ). We will indicate this using the notation α +x β when referring
to γ. Set
[α̇(0)] + [β̇(0)] := [γ̇(0)].
The zero vector in Tp M is described by a curve α such that ẋiα (0) = 0, ∀i. For this
operation to be well defined one has to check several things.
(a) The equivalence class [γ̇(0)] is independent of coordinates. In other words, if
(x1 , . . . , xm ) and (y 1 , . . . , y m ) are local coordinates near p such that

xi (p) = y j (p) = 0, ∀i, j,

then
α +x β ∼1 α +y β.

(b) If a curve represents the zero vector in some local coordinates (xi ) at p, then it
represents the zero vector in any other choice of local coordinates at p.
(c) If [α˙1 (0)] = [α˙2 (0)] and [β˙1 (0)] = [β˙2 (0)], then

[α˙1 (0)] + [β˙1 (0)] = [α˙2 (0)] + [β˙2 (0)].

We let the reader supply the routine details. ⊔

⊓

Exercise 2.1.6. Finish the proof of the Lemma 2.1.5. ⊔

⊓

From this point on we will omit the brackets [ – ] in the notation of a tangent vector.
Thus, [α̇(0)] will be written simply as α̇(0).
As one expects, all the above notions admit a nice description using local coordinates.
Let (x1 , . . . , xm ) be coordinates near p ∈ M such that xi (p) = 0, ∀i. Consider the curves

ek (t) = (tδk1 , . . . , tδkm ), k = 1, . . . , m,

where δji denotes Kronecker’s delta symbol. For example

e1 (t) = (t, 0, 0, . . . , 0 < e2 (t) = (0, t, 0, . . . , 0), . . . .

We set

∂xk (p) := ėk (0) = δk1 , . . . , δkn . (2.1.2)
Often, when the point p is clear from the context, we will omit it in the above notation.
☞ Note that the vectors ∂xk (p) depend on the choice of local coordinates (x1 , . . . , xm ).
28 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Lemma 2.1.7. The collection ∂xk (p) 1≤k≤m
is a basis of Tp M .

Proof. Given a smooth path

α(t) = x1α (t), . . . , xm
α (t) , α(0) = p,

we have




α̇(0) = ẋ1α (0), . . . , ẋm 1 m
α (0) = ẋα (0) 1, 0, . . . , 0 + · · · + ẋα (0) 0, . . . , 0, 1
Xm
(2.1.3)
= ẋ1α (0)∂x1 + · · · + ẋm
α (0)∂ x m = ẋiα (0)∂xi .
i=1

⊔
⊓

Exercise 2.1.8. Suppose that M ⊂ Rn is an m-dimensional smooth submanifold of

Rn . Let p0 ∈ M and suppose that α, β : (−ε, ε) → M are two smooth paths such that
α(0) = β(0) = p0 . Prove that the following statements are equivalent.

1. The paths α, β have first order contact at p0 .

dα(0) dβ(0)
2. dt = dt as vectors in Rn .

Deduce from the above that Tp0 M can be identified with the vector subspace of Rn
spanned by the vectors ~v ∈ Rn with the property that there exists a smooth path α :
(ε, ε) → M ⊂ Rn such that
dα(0)
α(0) = p0 , = ~v . ⊔
⊓
dt
Exercise 2.1.9. Let F : RN → Rk be a smooth map. Assume that
(a) M = F −1 (0) 6= ∅;
(b) rank Dx F = k, for all x ∈ M .
Then M is a smooth manifold of dimension N − k and

Tx M ∼
= ker Dx F, ∀x ∈ M. ⊔
⊓

Example 2.1.10. We want to describe T1 G, where G is one of the Lie groups discussed
in Section 1.2.2.
(a) G = O(n). Let (−ε, ε) ∋ s 7→ T (s) be a smooth path of orthogonal matrices such that
T (0) = 1. Then T t (s) · T (s) = 1. Differentiating this equality at s = 0 we get

Ṫ t (0) + Ṫ (0) = 0.

The matrix Ṫ (0) defines a vector in T1 O(n), so the above equality states that this tangent
space lies inside the space of skew-symmetric matrices, i.e., T1 O(n) ⊂ o(n). On the other
hand, we proved in Section 1.2.2 that dim G = dim o(n) so that

T1 O(n) = o(n).
2.1. THE TANGENT BUNDLE 29

(b) G = SL(n; R). Let (−ε, ε) ∋ s 7→ T (s) be a smooth path in SL(n; R) such that
T (0) = 1. Then det T (s) = 1 and differentiating this equality at s = 0 we get (see
Exercise 1.1.3)
tr Ṫ (0) = 0.
Thus, the tangent space at 1 lies inside the space of traceless matrices, i.e. T1 SL(n; R) ⊂
sl(n; R). Since (according to Exercise 1.2.27) dim SL(n; R) = dim sl(n; R) we deduce

T1 SL(n; R) = sl(n; R). ⊔

⊓

Exercise 2.1.11. Show that T1 U (n) = u(n) and T1 SU (n) = su(n). ⊔

⊓

2.1.2 The tangent bundle

In the previous subsection we have naturally associated to an arbitrary point p on a
manifold M a vector space Tp M . It is the goal of the present subsection to coherently
organize the family of tangent spaces (Tp M )p∈M . In particular, we want to give a rigorous
meaning to the intuitive fact that Tp M depends smoothly upon p.
We will organize the disjoint union of all tangent spaces as a smooth manifold T M .
There is a natural surjection
G
π : TM = Tp M → M, π(v) = p ⇐⇒ v ∈ Tp M.
p∈M

Any local coordinate

system x = (xi ) defined over an open set U ⊂
F M produces a natural
∂
basis ∂xi (p) of Tp M , for any p ∈ U . Thus, an element v ∈ T U = p∈U Tp M is completely
determined if we know

• which tangent space does it belong to, i.e., we know the point p = π(v) ∈ M ,


• the coordinates of v in the basis ∂xi (p) ,
X
v= X i (v)∂xi (p).
i

We thus have a bijection

Ψx : T U → U x × Rm ⊂ Rm × Rm ,

where U x is the image of U in Rm via the coordinates (xi ). We can now use the map Ψx
to transfer the topology on Rm × Rm to T U . Again, we have to make sure this topology
is independent of local coordinates.
To see this, pick a different coordinate system y = (y i ) on U . The coordinate inde-
pendence referred to above is equivalent to the statement that the transition map

Ψy ◦ Ψ−1 x m y
x : U × R −→ T U −→ U × R
m

is a homeomorphism.
30 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Let A := (x, X) ∈ U x × Rm . Then Ψ−1

x (A) = (p, α̇(0)), where x(p) = x, and α(t) ⊂ U
is a path through p given in the x coordinates as

α(t) = x + tX.

Denote by F : U x → U y the transition map x 7→ y. Then

Ψy ◦ Ψ−1 1
x (A) = y(x); Y , . . . , Y
m
,
P j

where α̇(0) = (ẏαj (0)) = Y ∂yj (p), and yα (t) is the description of the path α(t) in
the coordinates y j . Applying the chain rule we deduce
X ∂y j X ∂y j
Y j = ẏαj (0) = i
ẋi (0) = X i. (2.1.4)
∂x ∂xi
i i

This proves that Ψy ◦ Ψ−1 x is actually smooth.

The natural topology of T M is obtained by patching together the topologies of T Uγ ,
where (Uγ , φγ )γ is a countable atlas of M . A set D ⊂ T M is open if its intersection with
any T Uγ is open in T Uγ . The above argument shows that T M is a smooth manifold with
(T Uγ , Ψγ ) a defining atlas. Moreover, the natural projection π : T M → M is a smooth
map.
Definition 2.1.12. The smooth manifold T M described above is called the tangent bundle
of M . ⊔
⊓

Proposition 2.1.13. A smooth map f : M → N induces a smooth map Df : T M → T N

such that
(a) Df (Tp M ) ⊂ Tf (p) N , ∀p ∈ M
(b) The restriction to each tangent space Dp F : Tp M → Tf (p) N is linear. The map Df is
called the differential of f , and one often uses the alternate notation f∗ = Df .

Proof. Recall that Tp M is the space of tangent vectors to curves through p. If α(t) is such
a curve (α(0) = p), then β(t) = f (α(t)) is a smooth curve in N through q = f (p), and we
define
Df (α̇(0)) := β̇(0).
One checks easily that if α1 ∼1 α2 , then f (α1 ) ∼1 f (α2 ), so that Df is well defined. To
prove that the map Df : Tp M → Tq N is linear it suffices to verify this in any particular
local coordinates (x1 , . . . , xm ) near p, and (y 1 , . . . , y n ) near q, such that xi (p) = 0, y j (q) =
0, ∀i, j, since any two choices differ (infinitesimally) by a linear substitution. Hence, we
can regard f as a collection of maps

(x1 , . . . , xm ) 7→ (y 1 (x1 , . . . , xm ), . . . , y n (x1 , . . . , xm )).

 
A basis in Tp M is given by ∂xi , while a basis of Tq N is given by ∂yj .
If α, β : (−ε, ε) → M are two smooth paths such that α(0) = β(0) = p, then in local
coordinates they have the description

α(t) = (x1α (t), . . . , xm 1 m

α (t)), β(t) = (xβ (t), . . . , xβ (t)),
2.1. THE TANGENT BUNDLE 31

α̇(0) = (ẋ1α (0), . . . , ẋm 1 m

α (0)), β̇(0) = (ẋβ (0), . . . , ẋβ (0) )

Then


F (α(t)) = yα1 (t), . . . , yαn (t) = y 1 ( xiα (t) ), . . . , y n ( xiα (t) ) ,



F (β(t)) = yβ1 (t), . . . , yβn (t) = y 1 ( xiβ (t) ), . . . , y n ( xiβ (t) ) ,


F (α(t) + β(t) ) = y 1 ( xiα (t) + xiβ (t) ), . . . , y n ( xiα (t) + xiβ (t) ) .
n n m
!

d (2.1.3) X X X ∂y j
DF α̇(0) = t=0
F (α(t)) = ẏαj (0)∂yj = ẋiα (0) ∂yj ,
dt ∂xi
j=1 j=1 i=1
n n m
!

d (2.1.3 ) X X X ∂y j i
DF β̇(0) = F (β(t)) = ẏβj (0)∂yj = ẋ (0) ∂yj ,
dt t=0 ∂xi β
j=1 j=1 i=1
n X

d
X m
∂y j i i


DF α̇(0) + β̇(0) = t=0
F α(t) + β(t) = ẋ α (0) + ẋ β (0) ∂yj
dt ∂xi
j=1 i=1



= DF α̇(0) + DF β̇(0) .
This shows
 jthat Df : Tp M → Tq N is the linear operator given in these bases by the
∂y
matrix ∂xi . In particular, this implies that Df is also smooth. ⊔
⊓
1≤j≤n, 1≤i≤m

2.1.3 Transversality
In this subsection we want to explain rigorously a phenomenon with which the reader may
already be intuitively acquainted. We describe it first in a special case.
Suppose M is a submanifold of dimension 2 in R3 . Then, a simple thought experiment
suggests that most horizontal planes will not be tangent to M . Equivalently, if we denote
by f the restriction to M of the function (x, y, z) 7→ z , then for most real numbers h the
level set f −1 (h) = M ∩ {z = h} does not contain a point where the differential of f is
zero, so that most level sets f −1 (h) are smooth submanifolds of M of codimension 1, i.e.,
smooth curves on M .
We can ask a more general question. Given two smooth manifolds X, Y , a smooth map
f : X → Y , is it true that for “most” y ∈ Y the level set f −1 (y) is a smooth submanifold of
X of codimension dim Y ? This question has a positive answer, known as Sard’s theorem.
Definition 2.1.14. Suppose that Y is a smooth, connected manifold of dimension m.

(a) We say that a subset S ⊂ Y is negligible if, for any coordinate chart of Y , Ψ : U →
Rm , the set Ψ(S ∩ U ) ⊂ Rm has Lebesgue measure zero in Rm .

(b) Suppose F : X → Y is a smooth map, where X is a smooth manifold. A point

x ∈ X is called a critical point of F , if the differential Dx F : Tx X → TF (x) Y is not
surjective.

We denote by Cr F the set of critical points of F , and by ∆F ⊂ Y its image via F . We

will refer to ∆F as the discriminant set of F . The points in ∆F are called the critical
values of F . ⊔
⊓
32 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Exercise 2.1.15. Define


Z := (x, a, b, c) ∈ R4 ; ax2 + bx + c = 0; a 6= 0 .

(a) Prove that Z is a smooth submanifold of R4 .

π
(b) Define π : Z → R3 by (x, a, b, c) 7−→ (a, b, c). Describe explicitly the discriminant set
of π. ⊔
⊓

Definition 2.1.16. Suppose U, V are finite dimensional real Euclidean vector spaces,
O ⊂ U is an open subset, and F : O → V is a smooth map. Then a point u ∈ O is a
critical point of F if and only if


rank (Du F : U → V ) < min dim U, dim V . ⊔
⊓

Exercise 2.1.17. Show that if F : M → N is a smooth map and dim N ≤ dim M , hen for
every q ∈ N \ ∆F the fiber F −1 (q) is either empty, or a submanifold of M of codimension
dim N . ⊔
⊓

Theorem 2.1.18 (Sard). Suppose U, V are finite dimensional real Euclidean vector spaces,
O ⊂ U is an open subset, and F : O → V is a smooth map. Assume dim V ≤ dim U .
Then the discriminant set ∆F is negligible.

Proof. We follow the elegant approach of J. Milnor [97] and L. Pontryagin [111]. Set
n = dim U , and m = dim V . We will argue inductively on the dimension n.
For every positive integer k we denote by Cr kF ⊂ Cr F the set of points u ∈ O such
that all the partial derivatives of F up to order k vanish at u. We obtain a decreasing
filtration of closed sets
Cr F ⊃ Cr 1F ⊃ Cr 2F ⊃ · · · .
The case n = 0 is trivially true so we may assume n > 0, and the statement is true for
any n′ < n, and any m ≤ n′ . The inductive step is divided into three intermediary steps.
Step 1. The set F (Cr F \ Cr 1F ) is negligible.
Step 2. The set F (Cr kF \ Cr k+1 F ) is negligible for all k ≥ 1.
k
Step 3. The set F (Cr F ) is negligible for some sufficiently large k.
Step
 1. Set Cr ′F := Cr F \ Cr 1F . We
will show that there exists a countable open cover
Oj j≥1 of Cr ′F such that F (Oj ∩Cr ′F is negligible for all j ≥ 1. Since Cr ′F is separable,
it suffices to
prove that every point u ∈ Cr ′F admits an open neighborhood N such that
F N ∩ Cr ′F is negligible.
Suppose u0 ∈ Cr ′F . Assume first that there exist local coordinates (x1 , . . . , , xn )
defined in a neighborhood N of u0 , and local coordinates (y 1 , . . . , y m ) near v0 = F (u0 )
such that,
xi (u0 ) = 0, ∀i = 1, . . . n, y j (v0 ) = 0, ∀j = 1, . . . m,
and the restriction of F to N is described by functions y j = y j (x1 , . . . , xm ) such that
y 1 = x1 .
For every t ∈ R we set

Nt := (x1 , . . . , xn ) ∈ N; x1 = t ,
2.1. THE TANGENT BUNDLE 33

and we define


Gt : Nt → Rm−1 , (t, x2 , . . . , xn ) 7→ y 2 (t, x2 , . . . , xn ), . . . , y m (t, x2 , . . . , xn ) .

Observe that
[
F N ∩ Cr ′F = {t} × Gt (Cr Gt ).
t

The inductive assumption implies that the sets Gt (Cr Gt ) have trivial (m − 1)-dimensional
Lebesgue measure. Using Cavalieri’s principle or Fubini’s theorem we deduce that F (N ∩
Cr ′F ) has trivial m-dimensional Lebesgue measure.
To conclude Step 1 is suffices to prove that the above simplifying assumption concerning
the existence of nice coordinates is always fulfilled. To see this, choose local coordinates
(s1 , . . . , sn ) near u0 and coordinates (y 1 , . . . , y m ) near v0 such that

si (u0 ) = 0, ∀i = 1, . . . n, y j (v0 ) = 0, ∀j = 1, . . . m,

The map F is then locally described by a collection of functions y j (s1 , . . . , sn ), j =

1, . . . , n. Since u ∈ Cr ′F , we can assume, after an eventual re-labelling of coordinates,
1
that ∂y ∂s1
(u0 ) 6= 0. Now define

x1 = y 1 (s1 , . . . , sn ), xi = si , ∀i = 2, . . . , n.

The implicit function theorem shows that the collection of functions (x1 , . . . , xn ) defines
a coordinate system in a neighborhood of u0 . We regard y j as functions of xi . From the
definition we deduce y 1 = x1 .
(k) (k)
Step 2. Set Cr F := Cr kF \ Cr k+1 F . Since u0 ∈ Cr F , we can find local coordinates
(s1 , . . . , sn ) near u0 and coordinates (y 1 , . . . , y m ) near v0 such that

si (u0 ) = 0, ∀i = 1, . . . n, y j (v0 ) = 0, ∀j = 1, . . . m,

∂ j y1
(u0 ) = 0, ∀j = 1, . . . , k,
∂(s1 )j
and
∂ k+1 y 1
(u0 ) 6= 0.
∂(s1 )k+1
Define
∂ k y1
x1 (s) = ,
∂(s1 )k
and set xi := si , ∀i = 2, . . . , n.
Then the collection (xi ) defines smooth local coordinates on an open neighborhood N
of u0 , and Cr kF ∩N is contained in the hyperplane {x1 = 0}. Define

G : N ∩ {x1 = 0} → V ; G(x2 , . . . , xm ) = F (0, x2 , . . . , xn ).

Then
Cr kG ∩N = Cr kG , F (Cr kF ∩N) = G(Cr kG ),
34 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

(k)
and the induction assumption implies that G(Cr kG ) is negligible. By covering Cr F with
a countably many open neighborhood {Nℓ }ℓ>1 such that F (Cr kF ∩Nℓ ) is negligible we
(k)
conclude that F (Cr F ) is negligible.
n
Step 3. Suppose k > m . We will prove that F (Cr kF ) is negligible. More precisely, we
will show that, for every compact subset S ⊂ O, the set F (S ∩ Cr kF ) is negligible.
From the Taylor expansion around points in Cr kF ∩S we deduce that there exist num-
bers r0 ∈ (0, 1) and λ0 > 0, depending only on S, such that, if C is a cube with edge
r < r0 which intersects Cr kF ∩S, then

diam F (C) < λ0 r k ,

where for every set A ⊂ V we define


diam(A) := sup |a1 − a2 |; a1 , a2 ∈ A .

In particular, if µm denotes the m-dimensional Lebesgue measure on V , and µn denotes

the n-dimensional Lebesgue measure on U , we deduce that there exists a constant λ1 > 0
such that


µm (F (C) ≤ λ1 r mk = λ1 µn (C)mk/n .

Cover Cr kF ∩S by finitely many cubes {Cℓ }1≤ℓ≤L , of edges < r0 , such that their interiors
are disjoint. For every positive integer P , subdivide each of the cubes Cℓ into P n sub-cubes
Cℓσ of equal sizes. For every sub-cube Cℓσ which intersects Cr kF we have

λ1
µm ( F (Cℓσ ) ) ≤ λ1 µn (Cℓσ )mk/n = µn (Cℓ ).
P mk
We deduce that

X
µm F (Cℓ ∩ Cr kF ) ≤ µm ( F (Cℓσ ∩ Cr kF ) ) ≤ P n−mk µn (C).
σ

n
If we let P → ∞ in the above inequality, we deduce that when k > m we have

µm ( F (Cℓ ∩ Cr kF ) ) = 0, ∀ℓ = 1, . . . , L. ⊔
⊓

Theorem 2.1.18 admits the following immediate generalization.

Corollary 2.1.19 (Sard). Suppose F : X → Y is a smooth map between two smooth
connected, separable, manifolds. Then its discriminant set is negligible. ⊔
⊓

Definition 2.1.20. A smooth map f : M → N is called immersion (resp. submersion)

if for every p ∈ M the differential Dp f : Tp M → Tf (p) N is injective (resp. surjective). A
smooth map f : M → N is called an embedding if it is an injective immersion. ⊔
⊓

Suppose F : M → N is a smooth map, and dim M ≥ dim N . Then F is a submersion

if and only if the discriminant set ∆F is empty.
2.1. THE TANGENT BUNDLE 35

Exercise 2.1.21. Suppose F : M → N is a smooth map, and S ⊂ N is a smooth

submanifold of N . We say that F is transversal to S if for every x ∈ F −1 (S) we have

TF (x) N = TF (x) S + Dx F (Tx M ).

Prove that if F is transversal to S, then F −1 (S) is a submanifold of M whose codimension

is equal to the codimension of S in N . ⊔
⊓

Exercise 2.1.22. Suppose Λ, X, Y are smooth, connected manifolds, and F : Λ × X → Y

is a smooth map
Λ × X ∋ (λ, x) 7→ Fλ (x) ∈ Y.

Suppose S is a submanifold of Y such that F is transversal to S. Define

ZS := F −1 (S) ⊂ Λ × X.

Prove that if λ0 ∈ Λ is a regular value of the natural map

ZS ⊂ Λ × X → Λ,

then the map Fλ0 : X → Y is transversal to S. ⊔

⊓

2.1.4 Vector bundles

The tangent bundle T M of a manifold M has some special features which makes it a
very particular type of manifold. We list now the special ingredients which enter into
this special structure of T M since they will occur in many instances. Set for brevity
E := T M , and F := Rm (m = dim M ). We denote by Aut(F ) the Lie group GL(n, R) of
linear automorphisms of F . Then

(a) E is a smooth manifold, and there exists a surjective submersion π : E → M . For

every U ⊂ M we set E |U := π −1 (U ).

(b) From (2.1.4) we deduce that there exists a trivializing cover, i.e., an open cover U
of M , and for every U ∈ U a diffeomorphism

ΨU : E |U → U × F, v 7→ (p = π(v), ΦU
p (v))

(b1) Φp is a diffeomorphism Ep → F for any p ∈ U .

(b2) If U, V ∈ U are two trivializing neighborhoods with non empty overlap U ∩ V
then, for any p ∈ U ∩ V , the map ΦV U (p) = ΦVp ◦ (ΦU
p)
−1 : F → F is a linear

isomorphism, and moreover, the map

p 7→ ΦV U (p) ∈ Aut(F )

is smooth.
36 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

In our special case, the map ΦV U (p) is explicitly defined by the matrix (2.1.4)
 j 
∂y
A(p) = (p) .
∂xi 1≤i,j≤m

In the above formula, the functions (xi ) are local coordinates on U , and the functions (y j )
are local coordinates on V .
The properties (a) and (b) make no mention of the special relationship between E =
T M and M . There are many other quadruples (E, π, M, F ) with these properties and
they deserve a special name.
Definition 2.1.23. A vector bundle is a quadruple (E, π, M, F ) such that

• E, M are smooth manifolds,

• π : E → M is a surjective submersion,

• F is a vector space over the field K = R, C, and

• the conditions (a) and (b) above are satisfied.

The manifold E is called the total space, and M is called the base space. The vector
space F is called the standard fiber, and its dimension (over the field of scalars K) is called
the rank of the bundle. A line bundle is a vector bundle of rank one. ⊔
⊓

Roughly speaking, a vector bundle is a smooth family of vector spaces. Note that the
properties (b1) and (b2) imply that the fibers π −1 (p) of a vector bundle have a natural
structure of linear space. In particular, one can add elements in the same fiber. Moreover,
the addition and scalar multiplication operations on π −1 (p) depend smoothly on p. The
smoothness of the addition operation this means that the addition is a smooth map

+ : E ×M E = (u, v) ∈ E × E; π(u) = π(v) → E.

The smoothness of the scalar multiplication means that it is smooth map

R × E → E.

There is an equivalent way of defining vector bundles. To describe it, let us introduce
a notation. For any vector space F over the field K = R, C we denote by GLK (F ), (or
simply GL(F ) if there is no ambiguity concerning the field of scalars K) the Lie group of
linear automorphisms F → F .
According to Definition 2.1.23, we can find an open cover (Uα ) of M such that each of
the restrictions Eα = E |Uα is isomorphic to a product Ψα : Eα ∼ = F × Uα . Moreover, on
the overlaps Uα ∩ Uβ , the transition maps gαβ = Ψα Ψ−1
β can be viewed as smooth maps

gαβ : Uα ∩ Uβ → GL(F ).

They satisfy the cocycle condition

2.1. THE TANGENT BUNDLE 37

(a) gαα = 1F

(b) gαβ gβγ gγα = 1F over Uα ∩ Uβ ∩ Uγ .

Conversely, given an open cover (Uα ) of M , and a collection of smooth maps

gαβ : Uα ∩ Uβ → GL(F )

satisfying the cocycle condition, we can reconstruct a vector bundle by gluing the product
bundles Eα = F × Uα on the overlaps Uα ∩ Uβ according to the gluing rules


the point (v, x) ∈ Eα is identified with the point gβα (x)v, x ∈ Eβ ∀x ∈ Uα ∩ Uβ .

The details are carried out in the exercise below.

We will say that the map gβα is the transition from the α-trivialization to the β-
trivialization, and we will refer to the collection of maps (gβα ) satisfying the cocycle
condition as a gluing cocycle. We will refer to the cover (Uα ) as above as a trivializing
cover.
Exercise 2.1.24. Consider a smooth manifold M , a vector space V , an open cover (Uα ),
and smooth maps
gαβ : Uα ∩ Uβ → GL(V )
satisfying the cocycle condition. Set
[
X := V × Uα × {α}.
α

We topologize X as the disjoint union of the topological spaces Uα × V , and we define a

relation ∼⊂ X × X by
def
V × Uα × {α} ∋ (u, x, α) ∼ (v, x, β) ∈ V × Uβ × {β} ⇐⇒ x = y, v = gβα (x)u.

(a) Show that ∼ is an equivalence relation, and E = X/ ∼ equipped with the quotient
topology has a natural structure of smooth manifold.
(b) Show that the projection π : X → M , (u, x, α) 7→ x descends to a submersion E → M .
(c) Prove that (E, π, M, V ) is naturally a smooth vector bundle. ⊔
⊓

Definition 2.1.25. A description of a vector bundle in terms of a trivializing cover, and

a gluing cocycle is called a gluing cocycle description of that vector bundle. ⊔
⊓

⊔
Exercise 2.1.26. Find a gluing cocycle description of the tangent bundle of the 2-sphere.⊓

In the sequel, we will prefer to think of vector bundles in terms of gluing cocycles.
π
Definition 2.1.27. (a) A section in a vector bundle E → M defined over the open subset
U ⊂ M is a smooth map s : U → E such that

s(p) ∈ Ep = π −1 (p), ∀p ∈ U ⇐⇒ π ◦ s = 1U .
38 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

The space of smooth sections of E over U will be denoted by Γ(U, E) or C ∞ (U, E). Note
that Γ(U, E) is naturally a vector space. We will use the simpler notation C ∞ (E) when
referring to the space of sections of E over M .
(b) A section of the tangent bundle of a smooth manifold is called a vector field on that
manifold. The space of vector fields over on open subset U of a smooth manifold is denoted
by Vect (U ). ⊔
⊓

Proposition 2.1.28. Suppose E → M is a smooth vector bundle with standard fiber F ,

defined by an open cover (Uα )α∈A , and gluing cocycle

gβα : Uαβ → GL(F ).

Then there exists a natural bijection

 between the vector space of smooth sections of E, and
the set of families of smooth maps sα : Uα → F ; α ∈ A , satisfying the following gluing
condition on the overlaps

sα (x) = gαβ (x)sβ (x), ∀x ∈ Uα ∩ Uβ . ⊔

⊓

Exercise 2.1.29. Prove the above proposition. ⊔

⊓
π
Definition 2.1.30. (a) Let E i →i Mi be two smooth vector bundles. A vector bundle map
consists of a pair of smooth maps f : M1 → M2 and F : E 1 → E 2 satisfying the following
properties.

• The map F covers f , i.e., F (Ep1 ) ⊂ Ef2(p) , ∀p ∈ M1 . Equivalently, this means that
the diagram below is commutative

E1
F
w E2
π1 π2

u u
M1
f
w M2
• The induced map F : Ep1 → Ef2(p) is linear.

The composition of bundle maps is defined in the obvious manner and so is the identity
morphism so that one can define the notion of bundle isomorphism in the standard way.
(b) If E and F are two vector bundles over the same manifold, then we denote by
Hom(E, F ) the space of bundle maps E → F which cover the identity 1M . Such bundle
maps are called bundle morphisms. ⊔
⊓

For example, the differential Df of a smooth map f : M → N is a bundle map

Df : T M → T N covering f .
π
Definition 2.1.31. Let E → M be a smooth vector bundle. A bundle endomorphism of
E is a bundle morphism F : E → E. An automorphism (or gauge transformation) is an
invertible endomorphism. ⊔
⊓
2.1. THE TANGENT BUNDLE 39

Example 2.1.32. Consider the trivial vector bundle RnM → M over the smooth manifold
M . A section of this vector bundle is a smooth map u : M → Rn . We can think of u as a
smooth family of vectors (u(x) ∈ Rn )x∈M .
An endomorphism of this vector bundle is a smooth map A : M → EndR (Rn ). We can
think of A as a smooth family of n × n matrices
 1 
a1 (x) a12 (x) · · · a1n (x)
 a2 (x) a2 (x) · · · a2 (x) 
 1 2 n 
Ax =  .. .. .. ..  , aij (x) ∈ C ∞ (M ).
 . . . . 
an1 (x) an2 (x) · · · ann (x)
The map A is a gauge transformation if and only if det Ax 6= 0, ∀x ∈ M . ⊔
⊓

Exercise 2.1.33. Suppose E1 , E2 → M are two smooth vector bundles over the smooth
manifold with standard fibers F1 , and respectively F2 . Assume that both bundles are
defined by a common trivializing cover (Uα )α∈A and gluing cocycles
gβα : Uαβ → GL(F1 ), hβα : Uαβ → GL(F2 ).
Prove that there exists a bijection between the vector space of bundle morphisms Hom(E, F ),
and the set of families of smooth maps

Tα : Uα → Hom(F1 , F2 ); α ∈ A ,
satisfying the gluing conditions
−1
Tβ (x) = hβα (x)Tα (x)(x)gβα , ∀x ∈ Uαβ . ⊔
⊓
Exercise 2.1.34. Let V be a vector space, M a smooth manifold, {Uα } an open cover
of M , and gαβ , hαβ : Uα ∩ Uβ → GL(V ) two collections of smooth maps satisfying the
cocycle conditions. Prove the two collections define isomorphic vector bundles if and only
they are cohomologous, i.e., there exist smooth maps φα : Uα → GL(V ) such that
hαβ = φα gαβ φ−1
β . ⊔
⊓

2.1.5 Some examples of vector bundles

In this section we would like to present some important examples of vector bundles and
then formulate some questions concerning the global structure of a bundle.
Example 2.1.35. (The tautological line bundle over RPn and CPn ). First, let us
recall that a rank one vector bundle is usually called a line bundle. We consider only the
complex case. The total space of the tautological or universal line bundle over CPn is the
space

Un = UC n = (z, L) ∈ C
n+1
× CPn ; z belongs to the line L ⊂ Cn+1 .
Let π : UC n
n → CP denote the projection onto the second component. Note that for every
L ∈ CPn , the fiber through π −1 (L) = UC
n,L coincides with the one-dimensional subspace
in Cn+1 defined by L. ⊔
⊓
40 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Example 2.1.36. (The tautological vector bundle over a Grassmannian). We

consider here for brevity only complex Grassmannian Grk (Cn ). The real case is completely
similar. The total space of this bundle is

Uk,n = UC n n
k,n = (z, L) ∈ C × Grk (C ) ; z belongs to the subspace L ⊂ C
n .

If π denotes the natural projection π : Uk,n → Grk (Cn ), then for each L ∈ Grk (Cn ) the
fiber over L coincides with the subspace in Cn defined by L. Note that UC C
n−1 = U1,n . ⊔
⊓

Exercise 2.1.37. Prove that UC C

n and Uk,n are indeed smooth vector bundles. Describe a
C
gluing cocycle for Un . ⊔
⊓

Example 2.1.38. A complex line bundle over a smooth manifold M is described by an

open cover (Uα )α∈A , and smooth maps

gβα : Uα ∩ Uβ → GL(1, C) ∼
= C∗ ,

satisfying the cocycle condition

gγα (x) = gγβ (x) · gβα (x), ∀x ∈ Uα ∩ Uβ ∩ Uγ .

Consider for example the manifold M = S 2 ⊂ R3 . Denote as usual by N and S the North
and respectively South pole. We have an open cover

S 2 = U0 ∪ U∞ , U0 = S 2 \ {S}, U1 = S 2 \ {N }.

In this case, we have only a single nontrivial overlap, UN ∩ US . Identify U0 with the
complex line C, so that the North pole becomes the origin z = 0.
For every n ∈ Z we obtain a complex line bundle Ln → S 2 , defined by the open cover
{U0 , U1 } and gluing cocycle

g10 (z) = z −n , ∀z ∈ C∗ = U0 \ {0}.

A smooth section of this line bundle is described by a pair of smooth functions

u0 : U0 → C, u1 : U1 → C,

which along the overlap U0 ∩ U1 satisfy the equality u1 (z) = z −n u0 (z). For example, if
n ≥ 0, the pair of functions

u0 (z) = z n , u1 (p) = 1, ∀p ∈ U1

defines a smooth section of Ln . ⊔

⊓

Exercise 2.1.39. We know that CP1 is diffeomorphic to S 2 . Prove that the universal
line bundle Un → CP1 is isomorphic with the line bundle L−1 constructed in the above
example. ⊔
⊓
2.1. THE TANGENT BUNDLE 41

Exercise 2.1.40. Consider the incidence set


I := (x, L) ∈ ( Cn+1 \ {0} ) × CPn ; z ∈ L .

Prove that the closure of I in Cn+1 × CPn is a smooth manifold diffeomorphic to the total
space of the universal line bundle Un → CPn . This manifold is called the complex blowup
of Cn+1 at the origin. ⊔
⊓

The family of vector bundles is very large. The following construction provides a very
powerful method of producing vector bundles.
Definition 2.1.41. Let f : X → M be a smooth map, and E a vector bundle over M
defined by an open cover (Uα ) and gluing cocycle (gαβ ). The pullback of E by f is the
vector bundle f ∗ E over X defined by the open cover f −1 (Uα ), and the gluing cocycle
( gαβ ◦ f ). ⊔
⊓

One can check easily that the isomorphism class of the pullback of a vector bundle E is
independent of the choice of gluing cocycle describing E. The pullback operation defines
a linear map between the space of sections of E and the space of sections of f ∗ E.
More precisely, if s ∈ Γ(E) is defined by the open cover (Uα ), and the collection of
smooth maps ( sα ), then the pullback f ∗ s is defined by the open cover f −1 (Uα ), and
the smooth maps ( sα ◦ f ). Again, there is no difficulty to check the above definition is
independent of the various choices.
Exercise 2.1.42. For every positive integer k consider the map

pk : CP1 → CP1 , pk ([z0 , z1 ]) = [z0k , z1k ].

Show that p∗k Ln ∼

= Lkn , where Ln is the complex line bundle Ln → CP1 defined in Example
2.1.38. ⊔
⊓

Exercise 2.1.43. Let E → X be a rank k (complex) smooth vector bundle over the
manifold X. Assume E is ample, i.e. there exists a finite family s1 , . . . , sN of smooth
sections of E such that, for any x ∈ X, the collection {s1 (x), . . . , sN (x)} spans Ex . For
each x ∈ X we set n X o
Sx := v ∈ CN ; v i si (x) = 0 .
i

Note that dim Sx = N − k. We have a map F : X → Grk (CN ) defined by x 7→ Sx⊥ .

(a) Prove that F is smooth.
(b) Prove that E is isomorphic with the pullback F ∗ Uk,N . ⊔
⊓

Exercise 2.1.44. Show that any vector bundle over a smooth compact manifold is ample.
Thus any vector bundle over a compact manifold is a pullback of some tautological bundle!
⊔
⊓

The notion of vector bundle is trickier than it may look. Its definition may suggest
that a vector bundle looks like a direct product manif old×vector space since this happens
at least locally. We will denote by KnM the bundle Kr × M → M .
42 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

π
Definition 2.1.45. A rank r vector bundle E → M (over the field K = R, C) is called
trivial or trivializable if there exists a bundle isomorphism E ∼ = KrM . A bundle isomor-
phism E → KrM is called a trivialization of E, while an isomorphism Kr → E is called a
framing of E.
A pair (trivial vector bundle, trivialization) is called a trivialized, or framed bundle. ⊔
⊓
r
Remark 2.1.46. Let us explain why we refer to a bundle isomorphism ϕ : KM → E as
a framing.
Denote by (e1 , . . . , er ) the canonical basis of Kr . We can also regard the vectors
ei as constant maps M → Kr , i.e., as (special) sections of KrM . The isomorphism ϕ
determines sections fi = ϕ(ei ) of E with the property that for every x ∈ M the collection
(f1 (x), . . . , fr (x) ) is a frame of the fiber Ex .
This observation shows that we can regard any framing of a bundle E → M of rank r
as a collection of r sections u1 , . . . , ur which are pointwise linearly independent. ⊔
⊓

One can naively ask the following question. Is every vector bundle trivial? We can
even limit our search to tangent bundles. Thus we ask the following question.
Is it true that for every smooth manifold M the tangent bundle T M is trivial (as a
vector bundle)?
Let us look at some positive examples.
Example 2.1.47. T S 1 ∼ = RS 1 Let θ denote the angular coordinate on the circle. Then
∂θ is a globally defined, nowhere vanishing vector field on S 1 . We thus get a map

RS 1 → T S 1 , (s, θ) 7→ (s∂θ , θ) ∈ Tθ S 1

which is easily seen to be a bundle isomorphism.

Let us carefully analyze this example. Think of S 1 as a Lie group (the group of complex
numbers of norm 1). The tangent space at z = 1, i.e., θ = 0, coincides with the subspace
∂
Re z = 0, and ∂θ |1 is the unit vertical vector j.
Denote by Rθ the counterclockwise rotation by an angle θ. Clearly Rθ is a diffeomor-
phism, and for each θ we have a linear isomorphism

Dθ |θ=0 Rθ : T1 S 1 → Tθ S 1 .

Moreover,
∂θ = Dθ |θ=0 Rθ j.
The existence of the trivializing vector field ∂θ is due to to our ability to “move freely and
coherently” inside S 1 . One has a similar freedom inside a Lie group as we are going to see
in the next example. ⊔
⊓

Example 2.1.48. For any Lie group G the tangent bundle T G is trivial.
To see this let n = dim G, and consider e1 , . . . , en a basis of the tangent space at the
origin, T1 G. We denote by Rg the right translation (by g) in the group defined by

Rg : x 7→ x · g, ∀x ∈ G.
2.2. A LINEAR ALGEBRA INTERLUDE 43

Rg is a diffeomorphism with inverse Rg−1 so that the differential DRg defines a linear
isomorphism DRg : T1 G → Tg G. Set
Ei (g) = DRg (ei ) ∈ Tg G, i = 1, · · · , n.
Since the multiplication G × G → G, (g, h) 7→ g · h is a smooth map we deduce that the
vectors Ei (g) define smooth vector fields over G. Moreover, for every g ∈ G, the collection
{E1 (g), . . . , En (g)} is a basis of Tg G so we can define without ambiguity a map
X
Φ : RnG → T G, (g; X 1 , . . . X n ) 7→ (g; X i Ei (g)).
One checks immediately that Φ is a vector bundle isomorphism and this proves the claim.
In particular T S 3 is trivial since the sphere S 3 is a Lie group (unit quaternions). (Using
the Cayley numbers one can show that T S 7 is also trivial; see [112] for details). ⊔
⊓

We see that the tangent bundle T M of a manifold M is trivial if and only if there exist
vector fields X1 , . . . , Xm (m = dim M ) such that for each p ∈ M , X1 (p), . . . , Xm (p) span
Tp M . This suggests the following more refined question.
Problem Given a manifold M , compute v(M ), the maximum number of pointwise linearly
independent vector fields over M . Obviously 0 ≤ v(M ) ≤ dim M and T M is trivial if and
only if v(M ) = dim M . A special instance of this problem is the celebrated vector field
problem: compute v(S n ) for any n ≥ 1.
We have seen that v(S n ) = n for n =1,3 and 7. Amazingly, these are the only cases
when the above equality holds. This is a highly nontrivial result, first proved by J.F.Adams
in [3] using very sophisticated algebraic tools. This fact is related to many other natural
questions in algebra. For a nice presentation we refer to [91].
The methods we will develop in this book will not suffice to compute v(S n ) for any n,
but we will be able to solve “half” of this problem. More precisely we will show that
v(S n ) = 0 if and only if n is an even number.
In particular, this shows that T S 2n is not trivial. In odd dimensions the situation is far
more elaborate (a complete answer can be found in [3]).
Exercise 2.1.49. v(S 2k−1 ) ≥ 1 for any k ≥ 1. ⊔
⊓

The quantity v(M ) can be viewed as a measure of nontriviality of a tangent bundle.

Unfortunately, its computation is highly nontrivial. In the second part of this book we will
describe more efficient ways of measuring the extent of nontriviality of a vector bundle.

2.2 A linear algebra interlude

We collect in this section some classical notions of linear algebra. Most of them might be
familiar to the reader, but we will present them in a form suitable for applications in differ-
ential geometry. This is perhaps the least glamorous part of geometry, and unfortunately
cannot be avoided.
☞ Convention. All the vector spaces in this section will tacitly be assumed finite dimen-
sional, unless otherwise stated.
44 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

2.2.1 Tensor products

Let E, F be two vector spaces over the field K (K = R, C). Consider the (infinite) direct
sum M
T(E, F ) = K.
(e,f )∈E×F

Equivalently, the vector space T(E, F ) can be identified with the space of functions c :
E ×F → K with finite support. The space T(E, F ) has a natural basis consisting of “Dirac
functions” 
1 if (x, y) = (e, f )
δe,f : E × F → K, (x, y) 7→
0 if (x, y) 6= (e, f )
In particular, we have an injection1

δ : E × F → T(E, F ), (e, f ) 7→ δe,f .

Inside T(E, F ) sits the linear subspace R(E, F ) spanned by

λδe,f − δλe,f , λδe,f − δe,λf , δe+e′ ,f − δe,f − δe′ ,f , δe,f +f ′ − δe,f − δe,f ′ ,

where e, e′ ∈ E, f, f ′ ∈ F , and λ ∈ K. Now define

E ⊗K F := T(E, F )/R(E, F ),

and denote by π the canonical projection π : T(E, F ) → E ⊗ F . Set

e ⊗ f := π(δe,f ).

We get a natural map

ι : E × F → E ⊗ F, e × f 7→ e ⊗ f.
Obviously ι is bilinear. The vector space E ⊗K F is called the tensor product of E and F
over K. Often, when the field of scalars is clear from the context, we will use the simpler
notation E ⊗ F . The tensor product has the following universality property.
Proposition 2.2.1. For any bilinear map φ : E × F → G there exists a unique linear
map Φ : E ⊗ F → G such that the diagram below is commutative.

'' w E ⊗ F
ι
E×F
'' Φ
φ )' u . ⊔
⊓

The proof of this result is left to the reader as an exercise. Note that if (ei ) is a basis
of E, and (fj ) is a basis of F , then (ei ⊗ fj ) is a basis of E ⊗ F , and therefore

dimK E ⊗K F = (dimK E) · (dimK F ).

1
A word of caution: δ is not linear!
2.2. A LINEAR ALGEBRA INTERLUDE 45

Exercise 2.2.2. Using the universality property of the tensor product prove that there
exists a natural isomorphism E ⊗ F ∼
= F ⊗ E uniquely defined by e ⊗ f 7→ f ⊗ e. ⊔
⊓

The above construction can be iterated. Given three vector spaces E1 , E2 , E3 over
the same field of scalars K we can construct two triple tensor products:

(E1 ⊗ E2 ) ⊗ E3 and E1 ⊗ (E2 ⊗ E3 ).

Exercise 2.2.3. Prove there exists a natural isomorphism of K-vector spaces

(E1 ⊗ E2 ) ⊗ E3 ∼
= E1 ⊗ (E2 ⊗ E3 ). ⊔
⊓

The above exercise implies that there exists a unique (up to isomorphism) triple tensor
product which we denote by E1 ⊗ E2 ⊗ E3 . Clearly, we can now define multiple tensor
products: E1 ⊗ · · · ⊗ En .

Definition 2.2.4. (a) For any two vector spaces U, V over the field K we denote by
Hom(U, V ), or HomK (U, V ) the space of K-linear maps U → V .
(b) The dual of a K-linear space E is the linear space E ∗ defined as the space HomK (E, K)
of K-linear maps E → K. For any e∗ ∈ E ∗ and e ∈ E we set

he∗ , ei := e∗ (e). ⊔
⊓

The above constructions are functorial. More precisely, we have the following result.

Proposition 2.2.5. Suppose Ei , Fi , Gi , i = 1, 2 are K-vector spaces. Let Ti ∈ Hom (Ei , Fi ),

Si ∈ Hom(Fi , Gi ), i = 1, 2, be two linear operators. Then they naturally induce a linear
operator
T = T1 ⊗ T2 : E1 ⊗ E2 → F1 ⊗ F2 , S1 ⊗ S2 : F1 ⊗ F2
uniquely defined by

T1 ⊗ T2 (e1 ⊗ e2 ) = (T1 e1 ) ⊗ (T2 e2 ), ∀ei ∈ Ei ,

and satisfying
(S1 ⊗ S2 ) ◦ (T1 ⊗ T2 ) = (S1 ◦ T1 ) ⊗ (S2 ◦ T2 ).
(b) Any linear operator A : E → F induces a linear operator A† : F ∗ → E ∗ uniquely
defined by
hA† f ∗ , ei = hf ∗ , Aei, ∀e ∈ E, f ∗ ∈ F ∗ .
The operator A† is called the transpose or adjoint of A. Moreover,

(A ◦ B)† = B † ◦ A† , ∀A ∈ Hom(F, G), B ∈ Hom(E, F ). ⊔

⊓

Exercise 2.2.6. Prove the above proposition. ⊔

⊓
46 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Remark 2.2.7. Any basis (ei )1≤i≤n of the n-dimensional K-vector space determines a
basis (ei )1≤i≤n of the dual vector space V ∗ uniquely defined by the conditions
(
1 i=j
hei , ej i = δji =
0 i 6= j.

We say that the basis (ei ) is dual to the basis (ej ). The quantity (δji ) is called the
Kronecker symbol.
A vector v ∈ V admits a decomposition
n
X
v= v i ei ,
i=1

while a vector v ∗ ∈ V ∗ admits a decomposition

n
X
∗
v = vi∗ ei .
i=1

Moreover,
n
X
hv ∗ , vi = vi∗ v i .
i=1
Classically, a vector v in V is represented by a one-column matrix
 1 
v
 .. 
v =  . ,
vn

while a vector v ∗ is represented by a one-row matrix

v ∗ = [v1∗ . . . vn∗ ].

Then  
v1
 
hv ∗ , vi = [v1∗ . . . vn∗ ] ·  ...  ,
vn
where the · denotes the multiplication of matrices. ⊔
⊓

Using the functoriality of the tensor product and of the dualization construction one
proves easily the following result.
Proposition 2.2.8. (a) There exists a natural isomorphism

E∗ ⊗ F ∗ ∼
= (E ⊗ F )∗ ,

uniquely defined by

E ∗ ⊗ F ∗ ∋ e∗ ⊗ f ∗ 7−→ Le∗ ⊗f ∗ ∈ (E ⊗ F )∗ ,
2.2. A LINEAR ALGEBRA INTERLUDE 47

where
hLe∗ ⊗f ∗ , x ⊗ yi = he∗ , xihf ∗ , yi, ∀x ∈ E, y ∈ F.
In particular, this shows E ∗ ⊗F ∗ can be naturally identified with the space of bilinear maps
E × F → K.
(b) The adjunction morphism E ∗ ⊗ F → Hom (E, F ), given by

E ∗ ⊗ F ∋ e∗ ⊗ f 7−→ Te∗ ⊗f ∈ Hom(E, F ),

where
Te∗ ⊗f (x) := he∗ , xif, ∀x ∈ E,
is an isomorphism.2 ⊔
⊓

Exercise 2.2.9. Prove the above proposition. ⊔

⊓

Let V be a vector space. For r, s ≥ 0 set

Tsr (V ) := V ⊗r ⊗ (V ∗ )⊗s ,

where by definition V ⊗0 = (V ∗ )⊗0 = K. An element of Tsr is called tensor of type (r,s).

Example 2.2.10. According to Proposition 2.2.8 a tensor of type (1, 1) can be identified
with a linear endomorphism of V , i.e.,

T11 (V ) ∼
= End (V ),

while a tensor of type (0, k) can be identified with a k-linear map

V · · × V} → K.
| × ·{z ⊔
⊓
k

A tensor of type (r, 0) is called contravariant, while a tensor of type (0, s) is called
covariant. The tensor algebra of V is defined to be
M
T(V ) := Tsr (V ).
r,s

We use the term algebra since the tensor product induces bilinear maps
′ r+r ′
⊗ : Tsr × Tsr′ → Ts+s ′.

The elements of T(V ) are called tensors.

Exercise 2.2.11. Show that T(V ), +, ⊗ is an associative algebra. ⊔
⊓
2
The finite dimensionality of E is absolutely necessary. This adjunction formula is known in Bourbaki
circles as “formule d’adjonction chêr à Cartan”.
48 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Example 2.2.12. It is often useful to represent tensors using coordinates. To achieve

this pick a basis (ei ) of V , and let (ei ) denote the dual basis in V ∗ uniquely defined by

i i 1 if i = j
he , ej i = δj = .
0 if i 6= j

We then obtain a basis of Tsr (V )

{ei1 ⊗ · · · ⊗ eir ⊗ ej1 ⊗ · · · ⊗ ejs / 1 ≤ iα , jβ ≤ dim V }.

Any element T ∈ Tsr (V ) has a decomposition

T = Tji11...j
...ir
e ⊗ · · · ⊗ eir ⊗ ej1 ⊗ · · · , ⊗ejs ,
s i1

where we use Einstein convention to sum over indices which appear twice, once as a
superscript, and the second time as a subscript.
Using the adjunction morphism in Proposition 2.2.8, we can identify the space T11 (V )
with the space End(V ) a linear isomorphisms. Using the bases (ei ) and (ej ), and Einstein’s
convention, the adjunction identification can be described as the correspondence which
associates to the tensor A = aij ei ⊗ ej ∈ T11 (V ), the linear operator LA : V → V which
maps the vector v = v j ej to the vector LA v = aij v j ei . ⊔
⊓

On the tensor algebra there is a natural contraction (or trace) operation

r−1
tr : Tsr → Ts−1

uniquely defined by

tr (v1 ⊗ · · · ⊗ vr ⊗ u1 ⊗ · · · ⊗ us ) := hu1 , v1 iv2 ⊗ · · · vr ⊗ u2 ⊗ · · · ⊗ us ,

∀vi ∈ V, uj ∈ V ∗ .
In the coordinates determined by a basis (ei ) of V , the contraction can be described
as  
(tr T )ji22...i
...js
r
= T ii2 ...ir
ij2 ...js ,

where again we use Einstein’s convention. In particular, we see that the contraction
coincides with the usual trace on T11 (V ) ∼
= End (V ).

2.2.2 Symmetric and skew-symmetric tensors

Let V be a vector space over K = R, C. We set T r (V ) := T0r (V ), and we denote by Sr the
group of permutations of r objects. When r = 0 we set S0 := {1}.
Every permutation σ ∈ Sr determines a linear map T r (V ) → T r (V ), uniquely deter-
mined by the correspondences

v1 ⊗ · · · ⊗ vr 7−→ vσ(1) ⊗ · · · ⊗ vσ(r) , ∀v1 , . . . , vr ∈ V.

We denote this action of σ ∈ Sr on an arbitrary element t ∈ T r (V ) by σt.

2.2. A LINEAR ALGEBRA INTERLUDE 49

In this subsection we will describe two subspaces invariant under this action. These
are special instances of the so called Schur functors. (We refer to [46] for more general
constructions.) Define
1 X
S r : T r (V ) → T r (V ), S r (t) := σt,
r!
σ∈Sr

and  P
1
r r r! σ∈Srǫ(σ)σt if r ≤ dim V
Ar : T (V ) → T (V ), Ar (t) := .
0 if r > dim V
Above, we denoted by ǫ(σ) the signature of the permutation σ. Note that
A0 = S 0 = 1K .
The following results are immediate. Their proofs are left to the reader as exercises.
Lemma 2.2.13. The operators Ar and S r are projectors of T r (V ), i.e.,
S 2r = S r , A2r = Ar .
Moreover,
σS r (t) = S r (σt) = S r (t), σAr (t) = Ar (σt) = ǫ(σ)Ar (t), ∀t ∈ T r (V ). ⊔
⊓
Definition 2.2.14. A tensor T ∈ T r (V ) is called symmetric (respectively skew-symmetric)
if
S r (T ) = T (respectively Ar (T ) = T ).
The nonnegative integer r is called the degree of the (skew-)symmetric tensor.
The space of symmetric tensors (respectively skew-symmetric ones) of degree r will be
denoted by S r V (and respectively Λr V ). ⊔
⊓

Set M M
S • V := S r V, and Λ• V := Λr V.
r≥0 r≥0
Definition 2.2.15. The exterior product is the bilinear map
∧ : Λr V × Λs V → Λr+s V,
defined by
(r + s)!
ω r ∧ η s := Ar+s (ω ⊗ η), ∀ω r ∈ Λr V, η s ∈ Λs V. ⊔
⊓
r!s!
Proposition 2.2.16. The exterior product has the following properties.
(a) (Associativity)
(α ∧ β) ∧ γ = α ∧ (β ∧ γ), ∀α, β γ ∈ Λ• V.
In particular,
X
v1 ∧ · · · ∧ vk = k!Ak (v1 ⊗ ... ⊗ vk ) = ǫ(σ)vσ(1) ⊗ ... ⊗ vσ(k) , ∀vi ∈ V.
σ∈Sk

(b) (Super-commutativity)
ω r ∧ η s = (−1)rs η s ∧ ω r , ∀ω r ∈ Λr V, ω s ∈ Λs V.
50 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Proof. We first define a new product “∧1 ” by

ω r ∧1 η s := Ar+s (ω ⊗ η),

which will turn out to be associative and will force ∧ to be associative as well.
L of ∧r1 consider the quotient
To prove the associativity algebra Q∗ = T ∗ /I∗ , where T ∗
is the associative algebra ( r≥0 T (V ), +, ⊗), and I∗ is the bilateral ideal generated by
the set of squares {v ⊗ v/ v ∈ V }. Denote the (obviously associative) multiplication in Q∗
by ∪. The natural projection π : T ∗ → Q∗ induces a linear map π : Λ• V → Q∗ . We will
complete the proof of the proposition in two steps.
Step 1. We will prove that the map π : Λ• V → Q∗ is a linear isomorphism, and moreover

π(α∧1 β) = π(α) ∪ π(β). (2.2.1)

In particular, ∧1 is an associative product.

The crucial observation is

π(T ) = π(Ar (T )), ∀t ∈ T r (V ). (2.2.2)

It suffices to check (2.2.2) on monomials T = e1 ⊗ · · · ⊗ er , ei ∈ V . Since

(u + v)⊗2 ∈ I∗ , ∀u, v ∈ V

we deduce u ⊗ v = −v ⊗ u(mod I∗ ). Hence, for any σ ∈ Sr

π(e1 ⊗ · · · ⊗ er ) = ǫ(σ)π(eσ(1) ⊗ · · · ⊗ eσ(r) ) (2.2.3)

When we sum over σ ∈ Sr in (2.2.3) we obtain (2.2.2).

To prove the injectivity of π note first that A∗ (I∗ ) = 0. If π(ω) = 0 for some ω ∈ Λ• V ,
then ω ∈ ker π = I∗ ∩ Λ• V so that

ω = A∗ (ω) = 0.

The surjectivity of π follows immediately from (2.2.2). Indeed, any π(T ) can be alterna-
tively described as π(ω) for some ω ∈ Λ• V . It suffices to take ω = A∗ (T ).
To prove (2.2.1) it suffices to consider only the special cases when α and β are mono-
mials:
α = Ar (e1 ⊗ · · · ⊗ er ), β = As (f1 ⊗ · · · ⊗ fs ).
We have 

π(α∧1 β) = π Ar+s Ar (e1 ⊗ · · · ⊗ er ) ⊗ As (f1 ⊗ · · · ⊗ fs )

(2.2.2)
 
= π Ar (e1 ⊗ · · · ⊗ er ) ⊗ As (f1 ⊗ · · · ⊗ fs )

def


= π Ar (e1 ⊗ · · · ⊗ er ) ∪ π As (f1 ⊗ · · · ⊗ fs ) = π(α) ∪ π(β).
Thus ∧1 is associative.
2.2. A LINEAR ALGEBRA INTERLUDE 51

Step 2. The product ∧ is associative. Consider α ∈ Λr V , β ∈ Λs V and γ ∈ Λt V . We

have  
(r + s)! (r + s)! (r + s + t)!
(α ∧ β) ∧ γ = α∧1 β ∧ γ = (α∧1 β)∧1 γ
r!s! r!s! (r + s)!t!
(r + s + t)! (r + s + t)!
= (α∧1 β)∧1 γ = α∧1 (β∧1 γ) = α ∧ (β ∧ γ).
r!s!t! r!s!t!
The associativity of ∧ is proved. The computation above shows that
e1 ∧ · · · ∧ ek = k!Ak (e1 ⊗ · · · ⊗ ek ).
(b) The supercommutativity of ∧ follows from the supercommutativity of ∧1 (or ∪). To
prove the latter one uses (2.2.2). The details are left to the reader. ⊔
⊓

Exercise 2.2.17. Finish the proof of part (b) in the above proposition. ⊔
⊓

The space Λ• V is called the exterior algebra of V . The operation ∧ is called the
exterior product. The exterior algebra is a Z-graded algebra, i.e.,
(Λr V ) ∧ (Λs V ) ⊂ Λr+s V, ∀r, s.
Note that Λr V = 0 for r > dim V (pigeonhole principle).
Definition 2.2.18. Let V be an n-dimensional K-vector space. The one dimensional
vector space Λn V is called thedeterminant line of V , and it is denoted by det V . ⊔
⊓

There exists a natural injection ιV : V ֒→ Λ• V , ιV (v) = v, such that

ιV (v) ∧ ιV (v) = 0, ∀v ∈ V.
This map enters crucially into the formulation of the following universality property.
Proposition 2.2.19. Let V be a vector space over K. For any K-algebra A, and any
linear map φ : V → A such that φ(x)2 = 0, there exists an unique morphism of K-algebras
Φ : Λ• V → A such that the diagram below is commutative

[y w Λ• V
ιV
V
[[
φ ℄
[ uΦ ,

A
i.e., Φ ◦ ιV = φ.
Exercise 2.2.20. Prove Proposition 2.2.19. ⊔
⊓

The space of symmetric tensors S • V can be similarly given a structure of associative

algebra with respect to the product
α · β := S r+s (α ⊗ β), ∀α ∈ S r V, β ∈ S s V.
The symmetric product “·” is also commutative.
52 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Exercise 2.2.21. Formulate and prove the analogue of Proposition 2.2.19 for the algebra
S• V . ⊔
⊓

It is often convenient to represent (skew-)symmetric tensors in coordinates. If e1 , ... , en

is a basis of the vector space V then, for any 1 ≤ r ≤ n, the family
{ei1 ∧ · · · ∧ eir / 1 ≤ i1 < · · · < ir ≤ n}
is a basis for Λr V so that any degree r skew-symmetric tensor ω can be uniquely repre-
sented as X
ω= ω i1 ···ir ei1 ∧ · · · ∧ eir .
1≤i1 <···<ir ≤n

Symmetric tensors can be represented in a similar way. To any multi-index I = (i1 , . . . , in ) ∈

Z≥0 we can associate the symmetric monomial
eI := ei11 · · · einn
and the resulting collection (eI ) is a basis of S • V .
The Λ• and the S • constructions are functorial in the following sense.
Proposition 2.2.22. Any linear map L : V → W induces a natural morphisms of algebras
Λ• L : Λ• V → Λ• W, S • L : S • V → S • W
uniquely defined by their actions on monomials
Λ• L(v1 ∧ · · · ∧ vr ) = (Lv1 ) ∧ · · · ∧ (Lvr ),
and
S • L(v1 , . . . , vr ) = (Lv1 ) · · · (Lvr ).
A B
Moreover, if U −→ V −→ W are two linear maps, then
Λr (BA) = (Λr B)(Λr A), S r (BA) = (S r B)(S r A). ⊔
⊓
Exercise 2.2.23. Prove the above proposition. ⊔
⊓

In particular, if n = dimK V , then any linear endomorphism L : V → V defines an

endomorphism
Λn L : det V = Λn V → det V.
Since the vector space det V is 1-dimensional, the endomorphism Λn L can be identified
with a scalar det L, the determinant of the endomorphism L.
Definition 2.2.24. Suppose V is a finite dimensional vector space and A : V → V is
an endomorphism of V . For every positive integer r we denote by σr (A) the trace of the
induced endomorphism
Λr A : Λr V → Λr V,
and by ψr (A) the trace of the endomorphism Ar : V → V . We define
σ0 (A) = 1, ψ0 (A) = dim V. ⊔
⊓
2.2. A LINEAR ALGEBRA INTERLUDE 53

Exercise 2.2.25. Suppose V is a complex n-dimensional vector space, and A is an endo-

morphism of V .
(a) Prove that if A is diagonalizable and its eigenvalues are a1 , . . . , an , then
X
σr (A) = ai1 · · · air , ψr (A) = ar1 + · · · + arn .
1≤i1 <···<ir ≤n

(b) Prove that for every sufficiently small z ∈ C we have the equalities
X d X
det(1V + zA) = σr (A)z r , − log det(1 − zA) = ψr (A)z r−1 . ⊔
⊓
dz
j≥0 r≥1

The functors Λ• and S • have an exponential like behavior, i.e., for any finite dimen-
sional vectors spaces V, W , there exists natural isomorphisms of vector spaces

Λ• V ⊗ Λ• W ∼
= Λ• (V ⊕ W ). (2.2.4a)

S• V ⊗ S• W ∼
= S • (V ⊕ W ). (2.2.4b)
To define the isomorphism in (2.2.4a) consider the bilinear map

φ : Λ• V × Λ• W → Λ• (V ⊕ W ),

uniquely determined by

φ(v1 ∧ · · · ∧ vr , w1 ∧ · · · ∧ ws ) = v1 ∧ · · · ∧ vr ∧ w1 ∧ · · · ∧ ws .

The universality property of the tensor product implies the existence of a linear map

Φ : Λ• V ⊗ Λ• W → Λ• (V ⊕ W ),

such that Φ ◦ ι = φ, where ι is the inclusion of Λ• V × Λ• W in Λ• V ⊗ Λ• W . This map is

clearly surjective and it is an isomorphism since

dim Λ• V × Λ• W = dim Λ• (V ⊕ W ) = 2dim V +dim W .

This proves (2.2.4a).

To prove (2.2.4b) observe as above that there exists a linear map

φ : S • V ⊗ S • W → S • (V ⊕ W )

uniquely determined by the equality

φ(v1 · · · vj , w1 · · · wk ) = v1 · · · vj · w1 · · · wk .

To prove that it is an isomorphism, fix basis (ei )1≤i≤m of V and basis (fj )1≤j≤n of W .
Note that the collection  I
e ⊗ f J ; I ∈ Zm n
≥0 , J ∈ Z≥0
54 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

is a basis of S • V ⊗ S • W that is mapped bijectively by φ to the collection

 I J
e · f ; I ∈ Zm n
≥0 , J ∈ Z≥0

which is a basis of S • (V ⊕ W ).
We want to mention a few general facts about Z-graded vector spaces, i.e., vector
spaces equipped with a direct sum decomposition
M
V = Vn .
n∈Z

(We will always assume that each Vn is finite dimensional.) The vectors in Vn are said
to be homogeneous, of degree n. For example, the ring of polynomials K[x] is a K-graded
vector space. The spaces Λ• V and S • V are Z-graded vector spaces.
The direct sum of two Z-graded vector spaces V and W is a Z-graded vector space
with
(V ⊕ W )n := Vn ⊕ Wn .
The tensor product of two Z-graded vector spaces V and W is a Z-graded vector space
with M
(V ⊗ W )n := Vr ⊗ W s .
r+s=n

To any Z-graded vector space V one can naturally associate a formal series PV (t) ∈
Z[[t, t−1 ]] by X
PV (t) := (dimK Vn )tn .
n∈Z

The series PV (t) is called the Poincaré series of V .

Example 2.2.26. The Poincaré series of K[x] is
1
PK[x] (t) = 1 + t + t2 + · · · + tn−1 + · · · = . ⊔
⊓
1−t
Exercise 2.2.27. Let V and W be two Z-graded vector spaces. Prove the following
statements are true (whenever they make sense).
(a) PV ⊕W (t) = PV (t) + PW (t).
(b) PV ⊗W (t) = PV (t) · PW (t).
(c) dim V = PV (1). ⊔
⊓

Definition 2.2.28. Let V be a Z-graded vector space. The Euler characteristic of V,

denoted by χ(V ), is defined by
X
χ(V ) := PV (−1) = (−1)n dim Vn ,
n∈Z

whenever the sum on the right-hand side makes sense. ⊔

⊓

Remark 2.2.29. If we try to compute χ(K[x]) using the first formula in Definition 2.2.28
we get χ(K[x]) = 1/2, while the second formula makes no sense (divergent series). ⊔
⊓
2.2. A LINEAR ALGEBRA INTERLUDE 55

Proposition 2.2.30. Let V be a K-vector space of dimension n. Then

PΛ• V (t) = (1 + t)n ,

and  n  n−1  
1 1 d 1
PS •
V (t) = = .
1−t (n − 1)! dt 1−t
In particular, dim Λ• V = 2n and χ(Λ• V ) = 0.

Proof. From (2.2.4a) and (2.2.4b) we deduce using Exercise 2.2.27 that for any vector
spaces V and W we have
PΛ• (V ⊕W ) (t) = PΛ• V (t) · PΛ• W (t) and PS • (V ⊕W ) (t) = PS • V (t) · PS • W (t).

In particular, if V has dimension n, then V ∼

= Kn so that
PΛ• V (t) = (PΛ• K (t))n and PS • V (t) = (PS • K (t))n .
The proposition follows using the equalities
1
PΛ• K (t) = 1 + t, and PS • K (t) = PK[x] (t) = . ⊔
⊓
1−t

2.2.3 The “super” slang

The aim of this very brief section is to introduce the reader to the “super” terminology.
We owe the “super” slang to the physicists. In the quantum world many objects have a
special feature not present in the Newtonian world. They have parity (or chirality), and
objects with different chiralities had to be treated differently.
The “super” terminology provides an algebraic formalism which allows one to deal
with the different parities on an equal basis. From a strictly syntactic point of view, the
“super” slang adds the attribute super to most of the commonly used algebraic objects.
In this book, the prefix “s-” will abbreviate the word “super”
Definition 2.2.31. (a) A s-space is a Z2 -graded vector space, i.e., a vector space V
equipped with a direct sum decomposition V = V0 ⊕ V1 .
(b) A s-algebra over K is a Z2 -graded K-algebra, i.e., a K-algebra A together with a direct
sum decomposition A = A0 ⊕ A1 such that Ai · Aj ⊂ Ai+j (mod 2) . The elements in Ai are
called homogeneous of degree i. For any a ∈ Ai we denote its degree (mod 2) by |a|. The
elements in A0 are said to be even while the elements in A1 are said to be odd.
(c) The supercommutator in a s-algebra A = A0 ⊕ A1 is the bilinear map

[•, •]s : A × A → A,
defined on homogeneous elements ω i ∈ Ai , η j ∈ Aj by
[ω i , η j ]s := ω i η j − (−1)ij η j ω j .
An s-algebra is called s-commutative, if the suppercommutator is trivial, [•, •]s ≡ 0. ⊔
⊓

.
56 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Example 2.2.32. Let E = E 0 ⊕ E 1 be a s-space. Any linear endomorphism T ∈ End (E)

has a block decomposition
 
T00 T01
T = ,
T10 T11

where Tji ∈ End (E i , E j ). We can use this block decomposition to describe a structure of
s-algebra on End (E). The even endomorphisms have the form
 
T00 0
,
0 T11

while the odd endomorphisms have the form

 
0 T01
. ⊔
⊓
T10 0

Example 2.2.33. Let V be a finite dimensional space. The exterior algebra Λ• V is

naturally a s-algebra. The even elements are gathered in
M
Λeven V = Λr V,
r even

while the odd elements are gathered in

M
Λodd V = Λr V.
r odd

The s-algebra Λ• V is s-commutative. ⊔

⊓

Definition 2.2.34. Let A = A0 ⊕ A1 be a s-algebra. An s-derivation on A is a linear

operator on D ∈ End (A) such that, for any x ∈ A,

[D, Lx ]End(A)
s = LDx , (2.2.5)

End(A)
where [ , ]s denotes the supercommutator in End (A) (with the s-structure defined
in Example 2.2.32), while for any z ∈ A we denoted by Lz the left multiplication operator
a 7→ z · a.
An s-derivation is called even (respectively odd), if it is even (respectively odd) as an
element of the s-algebra End (A). ⊔
⊓

Remark 2.2.35. The relation (2.2.5) is a super version of the usual Leibniz formula.
Indeed, assuming D is homogeneous (as an element of the s-algebra End (A)) then equality
(2.2.5) becomes
D(xy) = (Dx)y + (−1)|x||D| x(Dy),

for any homogeneous elements x, y ∈ A. ⊔

⊓
2.2. A LINEAR ALGEBRA INTERLUDE 57

Example 2.2.36. Let V be a vector space. Any u∗ ∈ V ∗ defines an odd s-derivation of

Λ• V denoted by iu∗ uniquely determined by its action on monomials.
r
X
i (v0 ∧ v1 ∧ · · · ∧ vr ) =
u∗ (−1)i hu∗ , vi iv0 ∧ v1 ∧ ... ∧ v̂i ∧ · · · ∧ vr .
i=0

As usual, a hat indicates a missing entry. The derivation iu∗ is called the interior derivation
by u∗ or the contraction by u∗ . Often, one uses the alternate notation u∗ to denote this
derivation. ⊔
⊓

Exercise 2.2.37. Prove the statement in the above example. ⊔

⊓

Definition 2.2.38. Let A = (A0 ⊕ A1 , +, [ , ]) be an s-algebra over K, not necessarily

associative. For any x ∈ A we denote by Rx the right multiplication operator, a 7→ [a, x].
A is called an s-Lie algebra if it is s-anticommutative, i.e.,

[x, y] + (−1)|x||y| [y, x] = 0, for all homogeneous elements x, y ∈ A,

and ∀x ∈ A, Rx is a s-derivation.
When A is purely even, i.e., A1 = {0}, then A is called simply a Lie algebra. The
multiplication in a (s-) Lie algebra is called the (s-)bracket. ⊔
⊓

The above definition is highly condensed. In down-to-earth terms, the fact that Rx is
a s-derivation for all x ∈ A is equivalent with the super Jacobi identity

[[y, z], x] = [[y, x], z] + (−1)|x||y| [y, [z, x]], (2.2.6)

for all homogeneous elements x, y, z ∈ A. When A is a purely even K-algebra, then A is

a Lie algebra over K if [ , ] is anticommutative and satisfies (2.2.6), which in this case is
equivalent with the classical Jacobi identity,

[[x, y], z] + [[y, z], x] + [[z, x], y] = 0, ∀x, y, z ∈ A. (2.2.7)

Example 2.2.39. Let E be a vector space (purely even). Then A = End (E) is a Lie
algebra with bracket given by the usual commutator: [a, b] = ab − ba. ⊔
⊓

Proposition 2.2.40. Let A = A0 ⊕ A1 be a s-algebra, and denote by Ders (A) the vector
space of s-derivations of A.
(a) For any D ∈ Ders (A), its homogeneous components D 0 , D 1 ∈ End (A) are also
s-derivations.
End(A)
(b) For any D, D ′ ∈ Ders (A), the s-commutator [D, D ′ ]s is again an s-derivation.
x
(c) ∀x ∈ A the bracket B : a 7→ [a, x]s is a s-derivation called the bracket derivation
determined by x. Moreover

[B x , B y ]End
s
(A)
= B [x,y]s , ∀x, y ∈ A. ⊔
⊓
58 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Exercise 2.2.41. Prove Proposition 2.2.40. ⊔

⊓

Definition 2.2.42. Let E = E 0 ⊕ E 1 and F = F 0 ⊕ F 1 be two s-spaces. Their s-tensor

product is the s-space E ⊗ F with the Z2 -grading,
M
(E ⊗ F )ǫ := E i ⊗ F j , ǫ = 0, 1.
i+j≡ǫ (2)

b instead
To emphasize the super-nature of the tensor product we will use the symbol “⊗”
of the usual “⊗”. ⊔
⊓

Exercise 2.2.43. Show that there exists a natural isomorphism of s-spaces

b •V ∼
V ∗ ⊗Λ = Ders (Λ• V ),
∗ ⊗ω ∗ ⊗ω
uniquely determined by v ∗ × ω 7→ D v , where D v is s-derivation defined by
∗ ⊗ω
Dv (v) = hv ∗ , viω, ∀v ∈ V.

Notice in particular that any s-derivation of Λ• V is uniquely determined by its action on

∗
Λ1 V . (When ω = 1, D v ⊗1 coincides with the internal derivation discussed in Example
2.2.36.) ⊔
⊓

Let A = A0 ⊕ A1 be an s-algebra over K = R, C. A supertrace on A is a K-linear map

τ : A → K such that,
τ ([x, y]s ) = 0 ∀x, y ∈ A.
If we denote by [A, A]s the linear subspace of A spanned by the supercommutators

[x, y]s ; x, y ∈ A ,

then we see that the space of s-traces is isomorphic with the dual of the quotient space
A/[A, A]s .

Proposition 2.2.44. Let E = E0 ⊕ E1 be a finite dimensional s-space, and denote by

A the s-algebra of endomorphisms of E. Then there exists a canonical s-trace trs on A
uniquely defined by
trs 1E = dim E0 − dim E1 .
In fact, if T ∈ A has the block decomposition
 
T00 T01
T = ,
T10 T11

then
trs T = tr T00 − tr T11 . ⊔
⊓
Exercise 2.2.45. Prove the above proposition. ⊔
⊓
2.2. A LINEAR ALGEBRA INTERLUDE 59

2.2.4 Duality
Duality is a subtle and fundamental concept which permeates all branches of mathematics.
This section is devoted to those aspects of the atmosphere called duality which are relevant
to differential geometry. In the sequel, all vector spaces will be tacitly assumed finite
dimensional, and we will use Einstein’s convention without mentioning it. K will denote
one of the fields R or C.
Definition 2.2.46. A pairing between two K-vector spaces V and W is a bilinear map
B : V × W → K. ⊔
⊓

Any pairing B : V × W → K defines a linear map

IB : V → W ∗ , v 7→ B(v, • ) ∈ W ∗ ,
called the adjunction morphism associated to the pairing.
Conversely, any linear map L : V → W ∗ defines a pairing
BL : V × W → K, B(v, w) = (Lv)(w), ∀v ∈ V, w ∈ W.
Observe that IBL = L. A pairing B is called a duality if the adjunction map IB is an
isomorphisms.
Example 2.2.47. The natural pairing h•, •i : V ∗ × V → K is a duality. One sees that
Ih•,•i = 1V ∗ : V ∗ → V ∗ . This pairing is called the natural duality between a vector space
and its dual. ⊔
⊓

Example 2.2.48. Let V be a finite dimensional real vector space. Any symmetric non-
degenerate quadratic form (•, •) : V × V → R defines a (self)duality, and in particular a
natural isomorphism
L := I(•,•) : V → V ∗ .
When (•, •) is positive definite, then the operator L is called metric duality or lowering-
the-indices map. This operator can be nicely described in coordinates as follows. Pick a
basis (ei ) of V , and set
gij := (ei , ej ).
Let (ej ) denote the dual basis of V ∗ defined by
hej , ei i = δij , ∀i, j.
The action of L is then
Lei = gij ej . ⊔
⊓
Example 2.2.49. Consider V a real vector space and ω : V × V → R a skew-symmetric
bilinear form on V . The form ω is said to be symplectic if this pairing is a duality. In
this case, the induced operator Iω : V → V ∗ is called symplectic duality. ⊔
⊓

Exercise 2.2.50. Suppose that V is a real vector space, and ω : V ×V → R is a symplectic

duality. Prove the following.
(a) The V has even dimension.
(b) If (ei ) is a basis of V , and ωij := ω(ei , ej ), then det(ωij )1≤i,j≤dim V 6= 0. ⊔
⊓
60 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

The notion of duality is compatible with the functorial constructions introduced so

far.
Proposition 2.2.51. Let Bi : Vi × Wi → R (i = 1, 2) be two pairs of spaces in duality.
Then there exists a natural duality

B = B1 ⊗ B2 : (V1 ⊗ V2 ) × (W1 ⊗ W2 ) → R,

uniquely determined by

IB1 ⊗B2 = IB1 ⊗ IB2 ⇐⇒ B(v1 ⊗ v2 , w1 ⊗ w2 ) = B1 (v1 , w1 ) · B2 (v2 , w2 ). ⊔

⊓

Exercise 2.2.52. Prove Proposition 2.2.51. ⊔

⊓

Proposition 2.2.51 implies that given two spaces in duality B : V × W → K there is a

naturally induced duality
B ⊗n : V ⊗r × W ⊗r → K.
This defines by restriction a pairing

Λr B : Λr V × Λr W → K

uniquely determined by

Λr B (v1 ∧ · · · ∧ vr , w1 ∧ · · · ∧ wr ) := det (B (vi , wj ))1≤i,j≤r .

Exercise 2.2.53. Prove the above pairing is a duality. ⊔

⊓

In particular, the natural duality h • , • i : V ∗ × V → K induces a duality

h • , • i : Λr V ∗ × Λr V → R,

and thus defines a natural isomorphism

Λr V ∗ ∼ ∗
= (Λr V ) .

This shows that we can regard the elements of Λr V ∗ as skew-symmetric r-linear forms
V r → K.
A duality B : V × W → K naturally induces a duality B † : V ∗ × W ∗ → K by

B † (v ∗ , w∗ ) : =hv ∗ , I−1 ∗
B w i,

where IB : V → W ∗ is the adjunction isomorphism induced by the duality B.

Now consider a (real) Euclidean vector space V . Denote its inner product by ( • , • ).
The self-duality defined by ( • , • ) induces a self-duality

( • , • ) : Λr V × Λr V → R,
2.2. A LINEAR ALGEBRA INTERLUDE 61

determined by

(v1 ∧ · · · ∧ vr , w1 ∧ · · · ∧ wr ) := det ( (vi , wj ) )1≤i,j≤r . (2.2.8)

The symmetric r × r matrix

 
(v1 , v1 (v1 , v2 ) · · · (v1 , vr )
 (v2 , v1 ) (v2 , v2 ) · · · (v2 , vr ) 
 
G(v1 , . . . , vr ) :=  .. .. .. .. 
 . . . . 
(vr , v1 ) (vr , v2 ) · · · (vr , vr )

is called the Grammian of the vectors v1 , . . . , vr . It is positive semidefinite and, if the

vectors v1 , . . . , vr are linearly independent, then it is also nonsingular. This shows that
the bilinear form in (2.2.8) is symmetric and positive definite. We have thus proved the
following result.
Corollary 2.2.54. An inner product on a real vector space V naturally induces an inner
product on the tensor algebra T(V ), and in the exterior algebra Λ• V . ⊔
⊓

In a Euclidean vector space V the inner product induces the metric duality L : V → V ∗ .
r−1
This induces an operator L : Tsr (V ) → Ts+1 (V ) defined by

L(v1 ⊗ . . . ⊗ vr ⊗ u1 ⊗ · · · ⊗ us ) = (v2 ⊗ · · · ⊗ vr ) ⊗ ((Lv1 ⊗ u1 ⊗ · · · ⊗ us ). (2.2.9)

The operation defined in (2.2.9) is classically referred to as lowering the indices.

The reason for this nomenclature comes from the coordinate description of this oper-
ation. If T ∈ Tsr (V ) is given by

T = Tji11...j
...ir
e ⊗ · · · ⊗ eir ⊗ ej1 ⊗ · · · ⊗ ejs ,
s i1

then
i2 ...ir
(LT )jj 1 ...jr
= gij Tjii12...j
...ir
s
,

where gij = (ei , ej ). The inverse of the metric duality L−1 : V ∗ → V induces a linear
r+1
operation Tsr (V ) → Ts−1 (V ) called raising the indices.
Exercise 2.2.55 (Cartan). Let V be an Euclidean vector space. For any v ∈ V denote by
ev (resp. iv ) the linear endomorphism of Λ∗ V defined by ev ω = v ∧ ω (resp. iv = ıv∗ where
ıv∗ denotes the interior derivation defined by v ∗ ∈ V ∗ -the metric dual of v; see Example
2.2.36). Show that for any u, v ∈ V

[ev , iu ]s = ev iu + iu ev = (u, v)1Λ• V . ⊔

⊓

Definition 2.2.56. Let V be a real vector space. A volume form on V is a nontrivial

linear form on the determinant line of V , µ : det V → R. ⊔
⊓
62 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Equivalently, a volume form on V is a nontrivial element of det V ∗ = Λn V ∗ , n = dim V .

Since det V is 1-dimensional, a choice of a volume form corresponds to a choice of a basis
of det V .
Definition 2.2.57. (a) An orientation on a vector space V is a continuous, surjective
map
or : det V \ {0} → {±1}.
We denote by Or(V ) the set of orientations of V . Observe that Or(V ) consists of precisely
two elements.
(b) A pair (V, or), where V is a vector space, and or is an orientation on V is called an
oriented vector space.
(c) Suppose or ∈ Or(V ). A basis ω of det V is said to be positively oriented if or(ω) > 0.
Otherwise, the basis is said to be negatively oriented. ⊔
⊓

There is an equivalent way of looking at orientations. To describe it, note that any
nontrivial volume form µ on V uniquely specifies an orientation or µ given by

or µ (ω) := sign µ(ω), ∀ω ∈ det V \ {0}.

We define an equivalence relation on the space of nontrivial volume forms by declaring

µ1 ∼ µ2 ⇐⇒ µ1 (ω)µ2 (ω) > 0, ∀ω ∈ det V \ {0}.

Then
µ1 ∼ µ2 ⇐⇒ or µ1 = or µ2 .
To every orientation or we can associate an equivalence class [µ]or of volume forms such
that
µ(ω)or(ω) > 0, ∀ω ∈ det V \ {0}.
Thus, we can identify the set of orientations with the set of equivalence classes of nontrivial
volume forms.
Equivalently, to specify an orientation on V it suffices to specify a basis ω of det V .
The associated orientation or ω is uniquely characterized by the condition

or ω (ω) = 1.

To any basis {e1 , ..., en } of V one can associate a basis e1 ∧ · · · ∧ en of det V . Note that a
permutation of the indices 1, . . . , n changes the associated basis of det V by a factor equal
to the signature of the permutation. Thus, to define an orientation on a vector space, it
suffices to specify a total ordering of a given basis of the space.
An ordered basis of an oriented vector space (V, or) is said to be positively oriented
if so is the associated basis of det V .
Definition 2.2.58. Given two orientations or 1 , or 2 on the vector space V we define

or 1 /or 2 ∈ {±1}

to be
or 1 /or 2 := or 1 (ω)or 2 (ω), ∀ω ∈ det V \ {0}.
2.2. A LINEAR ALGEBRA INTERLUDE 63

We will say that or 1 /or 2 is the relative signature of the pair of orientations or 1 , or 2 . ⊔
⊓

Assume now that V is an Euclidean space. Denote the Euclidean inner product by
g( • , • ). The vector space det V has an induced Euclidean structure, and in particular,
there exist exactly two length-one-vectors in det V . If we fix one of them, call it ω, and
we think of it as a basis of det V , then we achieve two things.
• First, it determines a volume form µg defined by
µg (λω) = λ.

• Second, it determines an orientation on V .

Conversely, an orientation or ∈ Or(V ) uniquely selects a length-one-vector ω = ωor
in det V , which determines a volume form µg = µor
g . Thus, we have proved the following
result.
Proposition 2.2.59. An orientation or on an Euclidean vector space (V, g) canonically
selects a volume form on V , henceforth denoted by Detg = Detor
g . ⊔
⊓

Exercise 2.2.60. Let (V, g) be an n-dimensional Euclidean vector space, and or an

orientation on V . Show that, for any basis v1 , . . . vn of V , we have
q
Detor
g (v 1 ∧ · · · ∧ v n ) = or(v 1 ∧ · · · ∧ v n ) (det g(vi , vj )).

If V = R2 with its standard metric, and the orientation given by e1 ∧ e2 , prove that
| Detor
g (v1 ∧ v2 )|

is the area of the parallelogram spanned by v1 and v2 . ⊔

⊓

Definition 2.2.61. Let (V, g, or) be an oriented, Euclidean space and denote by Detorg
the associated volume form. The Berezin integral or ( berezinian) is the linear form
Z~
: Λ• V → R,
g

defined on homogeneous elements by

Z~ 
0 if deg ω < dim V
ω= or . ⊔
⊓
g Det g ω if deg ω = dim V
Definition 2.2.62. Let ω ∈ Λ2 V , where (V, g, or) is an oriented, Euclidean space. We
define its pfaffian as
Z~ 
or 0 if dim V is odd
P f (ω) = P f g (ω) := exp ω = 1 or ∧n ,
g n! Detg (ω ) if dim V = 2n
where exp ω denotes the exponential in the (nilpotent) algebra Λ• V ,
X ωk
exp ω := . ⊔
⊓
k!
k≥0
64 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

For any Euclidean space (V, g) we denote End− (V ) = End− (V, g) the space of skew-
symmetric linear maps V → V . a natural isomorphism
End− (V ) → Λ2 V ∗ , |; End− (V ) ∋ A 7→ ωA ∈ Λ2 V,
(2.2.10)
ωA (v1 , v2 ) = g(Av1 , v2 ), ∀v1 , v2 ∈ V.

If (e1 , . . . , eN ) is an orthonormal basis of V then

X 1X 1X
ωA = g(Aei , ej )ei ∧ ej = g(Aei , ej )ei ∧ ej = − aij ei ∧ ej ,
2 2
i<j i,j i<j

where aij are the entris of the N × N matrix describing the endomorphism A in the basis
(ei ),
aij = g(Aej , ei ) = g(ei , Aej ).
Given an orientation or on (V, g) we define the pfaffian of A ∈ End− (V ) by

P f (A) = P f or
g (A) := P f(ωA ).

Example 2.2.63. Let V = R2 denote the standard Euclidean space oriented by e1 ∧ e2 ,

where e1 e2 denotes the standard basis. If
 
0 −θ
A= ,
θ 0

then ωA = θe1 ∧ e2 so that P f(A) = θ. ⊔

⊓

Exercise 2.2.64. Let A : V → V be a skew-symmetric endomorphism of an oriented

Euclidean space V . Prove that P f (A)2 = det A. ⊔
⊓

Exercise 2.2.65. Let (V, g, or) be an oriented Euclidean space of dimension 2n. Consider
A : V → V a skewsymmetric endomorphism and a positively oriented orthonormal frame
e1 , . . . , e2n . Prove that
(−1)n X
P f or
g (A) = ǫ(σ)aσ(1)σ(2) · · · aσ(2n−1)σ(2n)
2n n!
σ∈S2n
X
= (−1)n ǫ(σ)aσ(1)σ(2) · · · aσ(2n−1)σ(2n) ,
σ∈S′2n

where aij = g(ei , Aej ) is the (i, j)-th entry in the matrix representing A in the basis (ei ),
and S′2n denotes the set of permutations σ ∈ S2n satisfying

σ(2k − 1) < min{σ(2k), σ(2k + 1)}, ∀k. ⊔

⊓

Let (V, g, or) be an n-dimensional, oriented, real Euclidean vector space. The metric
duality Lg : V → V ∗ induces both a metric, and an orientation on V ∗ . In the sequel we
will continue to use the same notation Lg to denote the metric duality Tsr (V ) → Tsr (V ∗ ) ∼
=
Tsr (V ).
2.2. A LINEAR ALGEBRA INTERLUDE 65

Definition 2.2.66. Suppose (V, g, or) oriented Euclidean space, and r is a nonnegative
integer, r ≤ dim V . The r-th Hodge pairing is the pairing

Ξ = Ξrg,or : Λr V ∗ × Λn−r V → R,

defined by


Ξ(ω r , η n−r ) := Detor
g L−1 r
g ω ∧η
n−r
, ω r ∈ Λr V ∗ , η n−r ∈ Λn−r V. ⊔
⊓

Exercise 2.2.67. Prove that the Hodge pairing is a duality. ⊔

⊓

Definition 2.2.68. The Hodge ∗-operator is the adjunction isomorphism

∗ = IΞ : Λr V ∗ → Λn−r V ∗

induced by the Hodge duality. ⊔

⊓

The above definition obscures the meaning of the ∗-operator. We want to spend some
time clarifying its significance.
Let α ∈ Λr V ∗ so that ∗α ∈ Λn−r V ∗ . Denote by h•, •i the standard pairing

Λn−r V ∗ × Λn−r V → R,

and by (•, •) the induced metric on Λn−r V ∗ . Then, by definition, for every β ∈ Λn−s V ∗
the operator ∗ satisfies

Detg (L−1 −1 −1
g α ∧ Lg β) = ∗α, Lg β = (∗α, β), ∀β ∈ Λ
n−r ∗
V . (2.2.11)

Let ω denote the unit vector in det V defining the orientation. Then (2.2.11) can be
rewritten as
hα ∧ β, ωi = (∗α, β), ∀β ∈ Λn−r V ∗ .

Thus
r ∗ n−r ∗
α ∧ β = (∗α, β) Det or
g , ∀α ∈ Λ V , ∀β ∈ Λ V . (2.2.12)

Equality (2.2.12) uniquely determines the action of ∗.

Example 2.2.69. Let V be the standard Euclidean space R3 with standard basis e1 , e2 , e3 ,
and orientation determined by e1 ∧ e2 ∧ e3 . Then

∗e1 = e2 ∧ e3 , ∗e2 = e3 ∧ e1 , ∗e3 = e1 ∧ e2 ,

∗1 = e1 ∧ e2 ∧ e3 , ∗(e1 ∧ e2 ∧ e3 ) = 1,

∗(e2 ∧ e3 ) = e1 , ∗(e3 ∧ e1 ) = e2 , ∗(e1 ∧ e2 ) = e3 . ⊔

⊓

The following result is left to the reader as an exercise.

66 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Proposition 2.2.70. Suppose (V, g, or) is an oriented, real Euclidean space of dimension
n. Then the associated Hodge ∗-operator satisfies

∗(∗ω) = (−1)p(n−p) ω ∀ω ∈ Λp V ∗ ,

Detg (∗1) = 1,

and
α ∧ ∗β = (α, β) ∗ 1, ∀α ∈ Λk V ∗ , ∀β ∈ Λn−k V ∗ . ⊔
⊓

Exercise 2.2.71. Let (V, g, ε) be an n-dimensional, oriented, Euclidean space. For every
t > 0 Denote by gt the rescaled metric gt = t2 g. If ∗ is the Hodge operator corresponding
to the metric g and orientation or, and ∗t is the Hodge operator corresponding to the
metric gt and the same orientation, show that

n
Detor or
tg = t Detg ,

and
∗t ω = tn−2p ∗ ω ∀ω ∈ Λp V ∗ . ⊔
⊓

We conclude this subsection with a brief discussion of densities.

Definition 2.2.72. Let V be a real vector space. For any r ≥ 0 we define an r-density
to be a function f : det V → R such that

f (λu) = |λ|r f (u), ∀u ∈ det V \ {0}, ∀λ 6= 0. ⊔

⊓

The linear space of r-densities on V will be denoted by |Λ|rV . When r = 1 we set

|Λ|V := |Λ|1V , and we will refer to 1-densities simply as densities.

Example 2.2.73. Any Euclidean metric g on V defines a canonical 1-density | Detg | ∈

|Λ|1V which associated to each ω ∈ det V its length, |ω|g . ⊔
⊓

Observe that an orientation or ∈ Or(V ) defines a natural linear isomorphism

ıor : det V ∗ → |Λ|V det V ∗ ∋ µ 7→ ıor µ ∈ |Λ|V , (2.2.13)

where
ıor µ(ω) = or(ω)µ(ω), ∀ω ∈ det V \ {0}.

In particular, an orientation in a Euclidean vector space canonically identifies |Λ|V with

R.
2.2. A LINEAR ALGEBRA INTERLUDE 67

2.2.5 Some complex linear algebra

It is convenient to have a coordinate description of the abstract objects introduced above.
Let V be an n-dimensional complex vector space and h a Hermitian metric on it. Pick an
unitary basis e1 , ..., en of V , i.e., n = dimC V , and h(ei , ej ) = δij . For each j, we denote
by fj the vector iej . Then the collection e1 , f1 , . . . , en , fn is an R-basis of V . Denote by
e1 , f 1 , . . . , en , f n the dual R-basis in V ∗ . Then

Re h (ei , ej ) = δij = Re h(fi , fj ) and Re h (ei , fj ) = 0,

i.e., X
Re h = (ei ⊗ ei + f i ⊗ f i ).
i

Also
ω(ei , fj ) = −Im h(ei , iej ) = δij , ω(ei , ej ) = ω(fi , fj ) = 0 ∀i, j,
which shows that X
ω = −Im h = ei ∧ f i .
i

Any complex
√ space V can be also thought of as a real vector space. The multiplication
by i = −1 defines a real linear operator which we denote by J. Obviously J satisfies
J 2 = −1V .
Conversely, if V is a real vector space then any real operator J : V → V as above
defines a complex structure on V by

(a + bi)v = av + bJv, ∀v ∈ V, a + bi ∈ C.

We will call an operator J as above a complex structure .

Let V be a real vector space with a complex structure J on it. The operator J has no
eigenvectors on V . The natural extension of J to the complexification of V , VC = V ⊗ C,
has two eigenvalues ±i, and we have a splitting of VC as a direct sum of complex vector
spaces (eigenspaces)
VC = ker (J − i) ⊕ (ker J + i).
Exercise 2.2.74. Prove that we have the following isomorphisms of complex vector spaces

V ∼
= ker (J − i) V ∼
= (ker J + i). ⊔
⊓

Set
V 1,0 := ker(J − i) ∼
=C V V 0,1 := ker(J + i) ∼
=C V .
= V 1,0 ⊕ V 0,1 ∼
Thus VC ∼ = V ⊕ V . We obtain an isomorphism of Z-graded complex vector
spaces
Λ• VC ∼
= Λ• V 1,0 ⊗ Λ• V 0,1 .
If we set

Λp,q V := Λp V 1,0 ⊗C Λq V 0,1 ,

68 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

then the above isomorphism can be reformulated as

M
Λk VC ∼
= Λp,q V. (2.2.14)
p+q=k

Note that the complex structure J on V induces by duality a complex structure J ∗ on

Vr∗ , and we have an isomorphism of complex vector spaces

Vc∗ = (V, J)∗c ∼

= (Vr∗ , J ∗ ).

We can define similarly Λp,q V ∗ as the Λp,q -construction applied to the real vector space
Vr∗ equipped with the complex structure J ∗ . Note that

Λ1,0 V ∗ ∼
= (Λ1,0 V )∗c ,

Λ0,1 V ∗ ∼
= (Λ0,1 V )∗c ,
and, more generally
Λp,q V ∗ ∼
= (Λp,q V )∗c .
If h is a Hermitian metric on the complex vector space (V, J), then we have a natural
isomorphism of complex vector spaces

Vc∗ ∼
= (Vr∗ , J ∗ ) ∼
=C (V, −J) ∼
=C V ,

so that
Λp,q V ∗ ∼
=C Λq,p V.
The Euclidean metric g = Re h, and the associated 2-form ω = −Im h are related by

g(u, v) = ω(u, Jv), ω(u, v) = g(Ju, v), ∀u, v ∈ V. (2.2.15)

Moreover, ω is a (1, 1)-form. To see this it suffices to pick a unitary basis (ei ) of V , and
construct as usual the associated real orthonormal basis {e1 , f1 , · · · , en , fn } (fi = Jei ).
Denote by {ei , f i ; i = 1, . . . , n} the dual orthonormal basis in Vr∗ . Then J ∗ ei = −f i , and
if we set
1 1
εi := √ (ei + if i ), ε̄j := √ (ej − if j ),
2 2
then
Λ1,0 V ∗ = spanC {εi } Λ0,1 = spanC {ε̄j },

and X
ω=i εi ∧ ε̄i .

Let V be a complex vector space, and e1 , . . . , en be a basis of V over C. This is not a real
basis of V since dimR V = 2 dimC V . We can however complete this to a real basis. More
precisely, the vectors e1 , ie1 , . . . , en , ien form a real basis of V .
2.2. A LINEAR ALGEBRA INTERLUDE 69

Proposition 2.2.75. Suppose (e1 , . . . , en ) and (f1 , . . . , fn ) are two complex bases of V ,
and Z = (zkj )1≤j,k is the complex matrix describing the transition from the basis e to the
basis f , i.e., X j
fk = zk ej , ∀1 ≤ k ≤ n.
j

Then
f1 ∧ if1 ∧ · · · ∧ fn ∧ ifn = | det Z|2 e1 ∧ ie1 ∧ · · · ∧ en ∧ ien .

Proof. We write
zkj = xjk + iykj , xjk , ykj ∈ R, êj = iej , fˆk = ifk .
Then X j X j X j
fk = (xk + iykj )ej = xk ej + yk êj ,
j j j

and X X
fˆk = − ykj ej + xjk êj .
j j

Then, if we set ǫn = (−1)n(n−1)/2 , we deduce

f1 ∧ if1 ∧ · · · ∧ fn ∧ ifn = ǫn f1 ∧ · · · ∧ fn ∧ fˆ1 ∧ · · · ∧ fˆn

b 1 ∧ ie1 ∧ ... ∧ en ∧ ien ,
= ǫn (det Ẑ)e1 ∧ · · · ∧ en ∧ ê1 ∧ · · · ∧ ên = (det Z)e
where Zb is the 2n × 2n real matrix
   
b Re Z − Im Z X −Y
Z= = ,
Im Z Re Z Y X

and X, Y denote the n × n real matrices with entries (xjk ) and respectively (ykj ).
We want to show that

det Zb = | det Z|2 , ∀Z ∈ EndC (Cn ). (2.2.16)

Let 
A := Z ∈ EndC (Cn ); | det Z|2 = det Zb .
We will prove that A = EndC (Cn ).
The set A is nonempty since it contains all the diagonal matrices. Clearly A is a closed
subset of EndC (Cn ), so it suffices to show that A is dense in EndC (Cn ).
Observe that the correspondence

EndC (Cn ) ∋ Z → Zb ∈ EndR (R)

is an endomorphism of R-algebras. Then, for every Z ∈ EndC (Cn ), and any complex
linear automorphism T ∈ AutC (Cn ), we have

T\
ZT −1 = TbZ
bTb−1 .
70 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Hence
Z ∈ A ⇐⇒ T ZT −1 ∈ A, ∀T ∈ AutC (Cn ).
In other words, if a complex matrix Z satisfies (2.2.16), then so will any of its conjugates.
In particular, A contains all the diagonalizable n × n complex matrices, and these form a
dense subset of EndC (Cn ) (see Exercise 2.2.76). ⊔
⊓

Exercise 2.2.76. Prove that the set of diagonalizable n × n complex matrices form a
dense subset of the vector space EndC (Cn ). ⊔
⊓

Definition 2.2.77. The canonical orientation of a complex vector space V , dimC V = n,

is the orientation defined by e1 ∧ ie1 ∧ ... ∧ en ∧ ien ∈ Λ2n
R V , where {e1 , . . . , en } is any
complex basis of V . ⊔
⊓

Suppose h is a Hermitian metric on the complex vector space V . Then g = Re h defines

is real, Euclidean metric on V regarded as a real vector space. The canonical orientation
or c on V , and the metric g define a volume form Detor g
c
∈ Λ2n ∗
R Vr , n = dimC V , and a
pffafian
1
P f h = P f or
h
c
: Λ2R V ∗ → R, Λ2 Vr∗ ∋ η 7→ g(η n , Detg ).
n!
If ω = − Im h is real 2-form associated with the Hermitian metric h, then

1 n
Detg = Detor
g
c
= ω ,
n!
and we conclude
P f h ω = 1.

2.3 Tensor fields

2.3.1 Operations with vector bundles
We now return to geometry, and more specifically, to vector bundles.
Let K denote one of the fields R or C, and let E → M be a rank r K-vector bundle
over the smooth manifold M . According to the definition of a vector bundle, we can find
an open cover (Uα ) of M such that each restriction E |Uα is trivial: E |Uα ∼ = V × Uα ,
where V is an r-dimensional vector space over the field K. The bundle E is obtained by
gluing these trivial pieces on the overlaps Uα ∩ Uβ using a collection of transition maps
gαβ : Uα ∩ Uβ → GL(V ) satisfying the cocycle condition.
Conversely, a collection of gluing maps as above satisfying the cocycle condition
uniquely defines a vector bundle. In the sequel, we will exclusively think of vector bundles
in terms of gluing cocycles.
Let E, F be two vector bundles over the smooth manifold M with standard fibers VE
and respectively VF , given by a (common) open cover (Uα ), and gluing cocycles

gαβ : Uαβ → GL(VE ), and respectively hαβ : Uαβ → GL(VF ).

2.3. TENSOR FIELDS 71

Then the collections

gαβ ⊕ hαβ : Uαβ → GL(VE ⊕ VF ), gαβ ⊗ hαβ : Uαβ → GL(VE ⊗ VF ),

† −1
(gαβ ) : Uαβ → GL( VE∗ ), Λr gαβ : Uαβ → GL(Λr VE ),

where † denotes the transpose of a linear map, satisfy the cocycle condition, and therefore
define vector bundles which we denote by E ⊕ F , E ⊗ F , E ∗ , and respectively Λr E. In
particular, if r = rankK E, the bundle Λr E has rank 1. It is called the determinant line
bundle of E, and it is denoted by det E. Given the adjunction isomorphism VE∗ ⊗ VF ∼ =
Hom(VE , VF ), we set
Hom(E, F ) := E ∗ ⊗ F.
We set
End(E) := Hom(E, E).
The reader can check easily that these vector bundles are independent of the choices of
transition maps used to characterize E and F (use Exercise 2.1.34). The bundle E ∗ is
called the dual of the vector bundle E. The direct sum E ⊕ F is also called the Whitney
sum of vector bundles. All the functorial constructions on vector spaces discussed in the
previous section have a vector bundle correspondent. (Observe that a vector space can be
thought of as a vector bundle over a point.)
These above constructions are natural in the following sense. Let E ′ and F ′ be vector
bundles over the same smooth manifold M ′ . Any bundle maps S : E → E ′ and T : F →
F ′ , both covering the same diffeomorphism φ : M → M ′ , induce bundle morphisms

S ⊕ T : E ⊕ F → E′ ⊕ T ′, S ⊗ T : E ⊗ F → E′ ⊗ F ′,

covering φ, a morphism
S † : (E ′ )∗ → E ∗ ,
covering φ−1 etc.
Exercise 2.3.1. Prove the assertion above. ⊔
⊓

Example 2.3.2. Let E, F , E ′ and F ′ be vector bundles over a smooth manifold M . Con-
sider bundle isomorphisms S : E → E ′ and T : F → F ′ covering the same diffeomorphism
of the base, φ : M → M . Then (S −1 )† : E ∗ → (E ′ )∗ is a bundle isomorphism covering φ,
so that we get an induced map (S −1 )† ⊗ T : E ∗ ⊗ F → (E ′ )∗ ⊗ F ′ . Note that we have a
natural identification


C ∞ Hom(E, F ) = C ∞ (E ∗ ⊗ F ) ∼
= Hom(E, F ),

where we recall that Hom(E, F ) denotes the space of smooth bundle morphisms E → F .⊓
⊔

Definition 2.3.3. Let E → M be a K-vector bundle over M . A metric on E is a section

h of E ∗ ⊗K E ∗ (E = E if K = R) such that, for any m ∈ M , h(m) defines a metric on Em
(Euclidean if K = R or Hermitian if K = C). ⊔
⊓
72 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

2.3.2 Tensor fields

We now specialize the previous considerations to the special situation when E is the
tangent bundle of M , E ∼
= T M . The cotangent bundle is then

T ∗ M := (T M )∗ .

We define the tensor bundles of M

Tsr (M ) : =Tsr (T M ) = (T M )⊗r ⊗ (T ∗ M )⊗s .

Definition 2.3.4. (a) A tensor field of type (r, s) over the open set U ⊂ M is a section
of Tsr (M ) over U .
(b) A degree r differential form (r-form for brevity) is a section of Λr (T ∗ M ). The space
of (smooth) r-forms over M is denoted by Ωr (M ). We set
M
Ω• (M ) := Ωr (M ).
r≥0

(c) A Riemannian metric on a manifold M is a metric on the tangent bundle. More

precisely, it is a symmetric (0, 2)-tensor g, such that for every x ∈ M , the bilinear map

gx : Tx M × Tx M → R

defines a Euclidean metric on Tx M . ⊔

⊓

If we view the tangent bundle as a smooth family of vector spaces, then a tensor field
can be viewed as a smooth selection of a tensor in each of the tangent spaces. In particular,
a Riemann metric defines a smoothly varying procedure of measuring lengths of vectors
in tangent spaces.
Example 2.3.5. It is often very useful to have a local description of these objects.
If (x1 , . . . , xn ) are local coordinates on an open set U ⊂ M , then the vector fields
( ∂x∂ 1 , . . . , ∂x∂ n ) trivialize T M |U , i.e., they define a framing of the restriction of T M to
U . We can form a dual framing of T ∗ M |U , using the 1-forms dxi , i = 1, . . . , n. They
satisfy the duality conditions
hdxi , ∂xj i = δji , ∀i, j.
A basis in Tsr (Tx M ) is given by

∂xi1 ⊗ ... ⊗ ∂xir ⊗ dxj1 ⊗ ... ⊗ dxjs ; 1 ≤ i1 , ... , ir ≤ n, 1 ≤ j1 , ... , js ≤ n .

Hence, any tensor T ∈ Tsr (M ) has a local description

T = Tji11...j
...ir
s
∂xi1 ⊗ ... ⊗ ∂xir ⊗ dxj1 ⊗ ... ⊗ dxjs .

In the above equality we have used Einstein’s convention. In particular, an r-form ω has
the local description
X

ω= ωi1 ...ir dxi1 ∧ · · · ∧ dxir , ωi1 ...ir = ω ∂xi1 , · · · , ∂xir ,
1≤i1 <···<ir ≤n
2.3. TENSOR FIELDS 73

while a Riemann metric g has the local description

X

g= gij dxi ⊗ dxj , gij = gji = g ∂xi , ∂xj . ⊔
⊓
i,j

Remark 2.3.6. (a) A covariant tensor field, i.e., a (0, s)-tensor field S, naturally defines
a C ∞ (M )-multilinear map
s
M
S: Vect (M ) → C ∞ (M ),
1


(X1 , ... , Xs ) 7→ p 7→ Sp X1 (p), . . . , Xs (p) ∈ C ∞ (M ).
Conversely, any such map uniquely defines a (0, s)-tensor field. In particular, an r-form η
can be identified with a skew-symmetric C ∞ (M )-multilinear map
r
M
η: Vect (M ) → C ∞ (M ).
1

Notice that the wedge product in the exterior algebras induces an associative product in
Ω• (M ) which we continue to denote by ∧.
(b) Let f ∈ C ∞ (M ). Its Fréchet derivative Df : T M → T R ∼
= R × R is naturally a 1-form.
Indeed, we get a smooth C ∞ (M )-linear map df : Vect (M ) → C ∞ (M ) defined by

df (X)m := Df (X)f (m) ∈ Tf (m) R ∼

= R.

In the sequel we will always regard the differential of a smooth function f as a 1-form and
to indicate this we will use the notation df (instead of the usual Df ). ⊔
⊓

Any diffeomorphism f : M → N induces bundle isomorphisms Df : T M → T N and

(Df −1 )† : T ∗ M → T ∗ N covering f . Thus, a diffeomorphism f induces a linear map

f∗ : Tsr (M ) → Tsr (N ), (2.3.1)

called the push-forward map. In particular, the group of diffeomorphisms of M acts

naturally (and linearly) on the space of tensor fields on M .
Example 2.3.7. Suppose f : M → N is a diffeomorphism, and S is a (0, k)-tensor field
on M , which we regard as a C ∞ (M )-multilinear map

S : Vect (M ) × · · · × Vect (M ) → C ∞ (M ).
| {z }
k

Then f∗ S is a (0, k) tensor field on N . Let q ∈ N , and set p := f −1 (q). Then, for any
Y1 , . . . , Yk ∈ Tq N , we have


(f∗ S)q (Y1 , . . . , Yk ) = Sp f∗−1 Y1 , . . . , (f∗ )−1 Yk


= Sp (Dp f )−1 Y1 , . . . , (Dp f )−1 Yk . ⊔
⊓.
74 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

For covariant tensor fields a more general result is true. More precisely, any smooth
map f : M → N defines a linear map

f ∗ : Ts0 (N ) → Ts0 (M ),

called the pullback by f . Explicitly, if S is such a tensor defined by a C ∞ (M )-multilinear

map
S : (Vect (N ))s → C ∞ (N ),
then f ∗ S is the covariant tensor field defined by


(f ∗ S)p (X1 (p), . . . , Xs (p)) := Sf (p) Dp f (X1 ), . . . , Dp f (Xs ) ,

∀X1 , . . . , Xs ∈ Vect (M ), p ∈ M . Note that when f is a diffeomorphism we have

f ∗ = (f∗−1 )† ,

where f∗ is the push-forward map defined in (2.3.1).

Example 2.3.8. Consider the map

F∗ : (0, ∞) × (0, 2π) → R2 , (r, θ) 7→ (x = r cos θ, y = r sin θ).

The map F defines the usual polar coordinates. It is a diffeomorphism onto the open
subset
U := R2 \ {(x, 0); x ≥ 0}.
∂ ∂
For simplicity, we will write ∂r , ∂x instead of ∂r , ∂x etc. We have

F ∗ dx = d(r cos θ) = cos θdr − rsinθdθ, F ∗ dy = d(r sin θ) = sinθdr + r cos θdθ,

F ∗ (dx ∧ dy) = d(r cos θ) ∧ d(rsinθ) = (cos θdr − rsinθdθ) ∧ (sinθdr + r cos θdθ)
= r(cos2 θ + sin2 θ)dr ∧ dθ = rdr ∧ dθ.
To compute F∗ ∂r and F∗ ∂θ we use the chain rule which implies
∂x ∂y 1
F∗ ∂r = ∂x + ∂y = cos θ∂x + sin θ∂y = (x∂x + y∂y )
∂r ∂r r
1
= (x∂x + y∂y ).
(x2 + y 2 )1/2
The Euclidean metric is described over U by the symmetric (0, 2)–tensor g = dx2 + dy 2 .
The pullback of g by F is the symmetric (0, 2)–tensor

2
2
F ∗ (dx2 + dy 2 ) = d(r cos θ) + d(r sin θ)

= (cos θdr − rsinθdθ)2 + (sinθdr + r cos θdθ)2 = dr 2 + r 2 dθ 2 .

To compute F∗ dr we need to express r as a function of x and y, r = (x2 + y 2 )1/2 , and then
we have

F∗ dr = (F −1 )∗ dr = d(x2 + y 2 )1/2 = x(x2 + y 2 )−1/2 dx + y(x2 + y 2 )−1/2 dy. ⊔

⊓
2.3. TENSOR FIELDS 75

All the operations discussed in the previous section have natural extensions to ten-
sor fields. There exists a tensor multiplication, a Riemann metric defines a duality
L : Vect (M ) → Ω1 (M ) etc. In particular, there exists a contraction operator
r+1
tr : Ts+1 (M ) → Tsr (M )

defined by

tr (X0 ⊗ · · · ⊗ Xr ) ⊗ (ω0 ⊗ · · · ⊗ ωs ) = ω0 (X0 )(X1 ⊗ · · · Xr ⊗ ω1 ⊗ · · · ⊗ ωs ),

∀Xi ∈ Vect (M ), ∀ωj ∈ Ω1 (M ). In local coordinates the contraction has the form
n  o
tr Tji00...j
...ir
s
= {Tijii11...j
...ir
s
}.

Let us observe that a Riemann metric g on a manifold M induces metrics in all the
associated tensor bundles Tsr (M ). If we choose local coordinates (xi ) on an open set U
then, as explained above, the metric g can be described as
X
g= gij dxi ⊗ dxj ,
i,j

while a tensor field T of type (r, s) can be described as

T = Tji11...j
...ir
∂ i ⊗ ... ⊗ ∂xir ⊗ dxj1 ⊗ ... ⊗ dxjs .
s x 1

If we denote by (gij ) the inverse of the matrix (gij ), then, for every point p ∈ U , the length
of T (p) ∈ Tsr (M )p is the number |T (p)|g defined by

|T (p)|g = gi1 k1 . . . gir kr gj1 ℓ1 · · · gjs ℓs Tji11...j

...ir k1 ...kr
T
s ℓ1 ...ℓs
,

where in the above equalities we have used Einstein’s convention.

The exterior product defines an exterior product on the space of smooth differential
forms
∧ : Ω• (M ) × Ω• (M ) → Ω• (M ).
The space Ω• (M ) is then an associative algebra with respect to the operations + and ∧.
Proposition 2.3.9. Let f : M → N be a smooth map. The pullback by f defines a
morphism of associative algebras f ∗ : Ω• (N ) → Ω• (M ). ⊔
⊓

Exercise 2.3.10. Prove the above proposition. ⊔

⊓

2.3.3 Fiber bundles

We consider useful at this point to bring up the notion of fiber bundle. There are several
reasons to do this.
On one hand, they arise naturally in geometry, and they impose themselves as worth
studying. On the other hand, they provide a very elegant and concise language to describe
many phenomena in geometry.
76 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

We have already met examples of fiber bundles when we discussed vector bundles.
These were “smooth families of vector spaces”. A fiber bundle wants to be a smooth
family of copies of the same manifold. This is a very loose description, but it offers a first
glimpse at the notion about to be discussed.
The model situation is that of direct product X = F × B, where B and F are smooth
manifolds. It is convenient to regard this as a family of manifolds (Fb )b∈B . The manifold
B is called the base, F is called the standard (model) fiber, and X is called the total
space. This is an example of trivial fiber bundle.
In general, a fiber bundle is obtained by gluing a bunch of trivial ones according to
a prescribed rule. The gluing may encode a symmetry of the fiber, and we would like to
spend some time explaining what do we mean by symmetry.
Definition 2.3.11. (a) Let M be a smooth manifold, and G a Lie group. We say the
group G acts on M from the left (respectively right), if there exists a smooth map

Φ : G × M → M, (g, m) 7→ Φg m,

such that Φ1 ≡ 1M and

Φg (Φh p) = Φgh p (respectively Φg (Φh p) = Φhg p) ∀g, h ∈ G, p ∈ M.

| {z } | {z }
left action right action

In particular, we deduce that ∀g ∈ G the map Φg is a diffeomorphism of M . For any

p ∈ M the set
G · p = {Φg p; g ∈ G}
is called the orbit through p of the action of G.
(b) Let G act on M . The action is called free if ∀g ∈ G \ {1}, and ∀p ∈ M , Φg p 6= p. The
action is called effective if, ∀g ∈ G \ {1}, Φg 6= 1M . ⊔
⊓

It is useful to think of a Lie group action on a manifold as encoding a symmetry of

that manifold.
Example 2.3.12. Consider the unit 2-dimensional sphere

S 2 = (x, y, z) ∈ R3 ; x2 + y 2 + z 2 = 1 }.

Then the counterclockwise rotations about the z-axis define a smooth left action of S 1 on
S 2 . More formally, if we use cylindrical coordinates (r, θ, z),

x = r cos θ, y = r sin θ, z = 0,

then for every ϕ ∈ R mod 2π ∼

= S 1 we define Rϕ : S 2 → S 2 by

Rϕ (r, θ, z) = (r, (θ + ϕ) mod 2π, z).

The resulting map R : S 1 × S 2 → S 2 , (ϕ, p) 7→ Rϕ (p) defines a left action of S 1 on S 2

encoding the rotational symmetry of S 2 about the z-axis. ⊔
⊓
2.3. TENSOR FIELDS 77

Example 2.3.13. Let G be a Lie group. A linear representation of G on a vector space

V is a left action Φ : G × V → V of G on V such that Φg is a linear map for any g ∈ G.
One says V is a G-module. For example, the tautological linear action of SO(n) on Rn
defines a linear representation of SO(n). ⊔
⊓

Example 2.3.14. Let G be a Lie group. For any g ∈ G denote by Lg (resp. Rg ) the left
(resp right) translation by g. In this way we get the tautological left (resp. right) action
of G on itself. ⊔
⊓

Definition 2.3.15. A smooth fiber bundle is an object composed of the following:

(a) a smooth manifold E called the total space;

(b) a smooth manifold F called the standard fiber ;

(c) a smooth manifold B called the base;

(d) a surjective submersion π : E → B called the natural projection;

(e) a collection of local trivializations, i.e., an open cover (Uα ) of the base B, and dif-
feomorphisms Ψα : F × Uα → π −1 (Uα ) such that

π ◦ Ψα (f, b) = b, ∀(f, b) ∈ F × Uα ,

i.e., the diagram below is commutative.

F × Uα
'')
Ψα
w π−1(Uα)
)' ^[^[[π
Uα

We can form the transition (gluing) maps Ψαβ : F × Uαβ → F × Uαβ , where Uαβ =
Uα ∩ Uβ , defined by Ψαβ = ψα−1 ◦ ψβ . According to (e), these maps can be written
as
ψαβ (f, b) = (Tαβ (b)f, b),
where Tαβ (b) is a diffeomorphism of F depending smoothly upon b ∈ Uαβ .

We will denote this fiber bundle by (E, π, F, B).

If G is a Lie group, then the bundle (E, π, F, B) is called a G-fiber bundle if it satisfies
the following additional conditions.

(f) There exists an effective left action of the Lie group G on F ,

G × F ∋ (g, x) 7→ g · x = Tg x ∈ F.

The group G is called the symmetry group of the bundle.

78 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

(g) There exist smooth maps gαβ : Uαβ → G satisfying the cocycle condition

gαα = 1 ∈ G, gγα = gγβ · gβα , ∀α, β, γ,

and such that

Tαβ (b) = Tgαβ (b) .

We will denote a G-fiber bundle by (E, π, F, B, G). ⊔

⊓

The choice of an open cover (Uα ) in the above definition is a source of arbitrariness
since there is no natural prescription on how to perform this choice. We need to describe
when two such choices are equivalent.
Two open covers (Uα ) and (Vi ), together with the collections of local trivializations

Φα : F × Uα → π −1 (Uα ) and Ψi : F × Vi → π −1 (Vi )

are said to be equivalent if, for all α, i, there exists a smooth map

Tαi : Uα ∩ Vi → G,

such that, for any x ∈ Uα ∩ Vi , and any f ∈ F , we have

Φ−1
α Ψi (f, x) = (Tαi (x)f, x).

A G-bundle structure is defined by an equivalence class of trivializing covers.

As in the case of vector bundles, a collection of gluing data determines a G-fiber bundle.
Indeed, if we are given a cover (Uα )α∈A of the base B, and a collection of transition maps
gαβ : Uα ∩ Uβ → G satisfying the cocycle condition, then we can get a bundle by gluing
the trivial pieces Uα × F along the overlaps.
More precisely, if b ∈ Uα ∩ Uβ , then the element (f, b) ∈ F × Uα is identified with the
element (gβα (b) · f, b) ∈ F × Uβ .
π
Definition 2.3.16. Let E → B be a G-fiber bundle. A G-automorphism of this bundle is
a diffeomorphism T : E → E such that π ◦ T = π, i.e., T maps fibers to fibers, and for any
trivializing cover (Uα ) (as in Definition 2.3.15) there exists a smooth map gα : Uα → G
such that
Ψ−1
α T Ψα (f, b) = ( gα (b)f, b ), ∀b, f. ⊔
⊓

Definition 2.3.17. (a) A fiber bundle is an object defined by conditions (a)-(d) and (f) in
the above definition. (One can think the structure group is the group of diffeomorphisms
of the standard fiber).
(b) A section of a fiber bundle E → B is a smooth map s : B → E such that π ◦ s = 1B ,
π
−1
i.e., s(b) ∈ π (b), ∀ b ∈ B. ⊔
⊓

Example 2.3.18. A rank r vector bundle (over K = R, C) is a GL(r, K)-fiber bundle with
standard fiber Kr , and where the group GL(r, K) acts on Kr in the natural way. ⊔
⊓
2.3. TENSOR FIELDS 79

Example 2.3.19. Let G be a Lie group. A principal G-bundle is a G-fiber bundle with
fiber G, where G acts on itself by left translations. Equivalently, a principal G-bundle
over a smooth manifold M can be described by an open cover U of M and a G-cocycle,
i.e., a collection of smooth maps

gU V : U ∩ V → G U, V ∈ U,

such that ∀x ∈ U ∩ V ∩ W (U, V, W ∈ U)

gU V (x)gV W (x)gW U (x) = 1 ∈ G. ⊔

⊓

Exercise 2.3.20 (Alternate definition of a principal bundle). Let P be a fiber bundle

with fiber a Lie group G. Prove the following are equivalent.
(a) P is a principal G-bundle.
(b) There exists a free, right action of G on G,

P × G → P, (p, g) 7→ p · g,

such that its orbits coincide with the fibers of the bundle P , and there exists a trivializing
cover 
Ψα : G × Uα → π −1 (Uα ) ,
such that
Ψα (hg, u) = Ψα (h, u) · g, ∀g, h ∈ G, u ∈ Uα . ⊔
⊓
Exercise 2.3.21 (The frame bundle of a manifold). Let M n be a smooth manifold.
Denote by F (M ) the set of frames on M , i.e.,

F (M ) = {(m; X1 , ... , Xn ); m ∈ M, Xi ∈ Tm M and span(X1 , ... , Xn ) = Tm M } .

(a) Prove that F (M ) can be naturally organized as a smooth manifold such that the
natural projection p : F (M ) → M , (m; X1 , ... , Xn ) 7→ m is a submersion.
(b) Show F (M ) is a principal GL(n, R)-bundle. The bundle F (M ) is called the frame
bundle of the manifold M .
Hint: A matrix T = (Tji ) ∈ GL(n, K) acts on the right on F (M ) by

(m; X1 , ..., Xn ) 7→ (m; (T −1 )i1 Xi , ..., (T −1 )in Xi ). ⊔

⊓

Example 2.3.22 (Associated fiber bundles). Let π : P → G be a principal G-bundle.

Consider a trivializing cover (Uα )α∈A , and denote by gαβ : Uα ∩ Uβ → G a collection of
gluing maps determined by this cover. Assume G acts (on the left) on a smooth manifold
F
τ : G × F → F, (g, f ) 7→ τ (g)f.
The collection ταβ = τ (gαβ ) : Uαβ → Diffeo (F ) satisfies the cocycle condition and can be
used (exactly as we did for vector bundles) to define a G-fiber bundle with fiber F . This
new bundle is independent of the various choices made (cover (Uα ) and transition maps
gαβ ). (Prove this!) It is called the bundle associated to P via τ and is denoted by P ×τ F .
⊔
⊓
80 CHAPTER 2. NATURAL CONSTRUCTIONS ON MANIFOLDS

Exercise 2.3.23. Prove that the tangent bundle of a manifold M n is associated to F (M )

via the natural action of GL(n, R) on Rn .] ⊔
⊓

Exercise 2.3.24 (The Hopf bundle). If we identify the unit odd dimensional sphere S 2n−1
with the submanifold

(z1 , ... , zn ) ∈ Cn ; |z0 |2 + · · · + |zn |2 = 1

then we detect an S 1 -action on S 2n−1 given by

eiθ · (z1 , ..., zn ) = (eiθ z1 , ..., eiθ zn ).

The space of orbits of this action is naturally identified with the complex projective space
CPn−1 .
(a) Prove that p : S 2n−1 → CPn−1 is a principal S 1 bundle called Hopf bundle. (p is the
obvious projection). Describe one collection of transition maps.
(b) Prove that the tautological line bundle over CPn−1 is associated to the Hopf bundle
via the natural action of S 1 on C1 . ⊔
⊓

Exercise 2.3.25. Let E be a vector bundle over the smooth manifold M . Any metric h
on E (euclidian or Hermitian) defines a submanifold S(E) ⊂ E by

S(E) = {v ∈ E; |v|h = 1}.

Prove that S(E) is a fibration over M with standard fiber a sphere of dimension rank E −1.
The bundle S(E) is usually called the sphere bundle of E. ⊔
⊓
Chapter 3

Calculus on Manifolds

This chapter describes the “kitchen” of differential geometry. We will discuss how one
can operate with the various objects we have introduced so far. In particular, we will
introduce several derivations of the various algebras of tensor fields, and we will also
present the inverse operation of integration.

3.1 The Lie derivative

3.1.1 Flows on manifolds
The notion of flow should be familiar to anyone who has had a course in ordinary differ-
ential equations. In this section we only want to describe some classical analytic facts in a
geometric light. We strongly recommend [6, 12] for more details, and excellent examples.
A neighborhood N of {0} × M in R × M is called balanced if, ∀m ∈ M , there exists
r ∈ (0, ∞] such that
( R × {m} ) ∩ N = (−r, r) × {m}.
Note that any continuous function f : M → (0, ∞) defines a balanced open

Nf := (t, m) ∈ R × M ; |t| < f (m) .

Definition 3.1.1. A local flow is a smooth map Φ : N → M , (t, m) 7→ Φt (m), where N

is a balanced neighborhood of {0} × M in R × M , such that

(a) Φ0 (m) = m, ∀m ∈ M .

(b) Φt (Φs (m)) = Φt+s (m) for all s, t ∈ R, m ∈ M such that

(s, m), (s + t, m), (t, Φs (m)) ∈ N.

When N = R × M , Φ is called a flow. ⊔

⊓

The conditions (a) and (b) above show that a flow is nothing but a left action of the
additive (Lie) group (R, +) on M .

81
82 CHAPTER 3. CALCULUS ON MANIFOLDS

Example 3.1.2. Let A be a n × n real matrix. It generates a flow ΦtA on Rn by

∞ k
!
X t
ΦtA x = etA x = Ak x. ⊔
⊓
k!
k=0

Definition 3.1.3. Let Φ : N → M be a local flow on M . The infinitesimal generator of

Φ is the vector field X on M defined by

d
X(p) = XΦ (p) := |t=0 Φt (p), ∀p ∈ M,
dt
i.e., X(p) is the tangent vector to the smooth path t 7→ Φt (p) at t = 0. This path is called
the flow line through p. ⊔
⊓

Exercise 3.1.4. Show that XΦ is a smooth vector field. ⊔

⊓

Example 3.1.5. Consider the flow etA on Rn generated by a n×n matrix A. Its generator
is the vector field XA on Rn defined by

d
XA (u) = |t=0 etA u = Au. ⊔
⊓
dt
Proposition 3.1.6. Let M be a smooth n-dimensional manifold. The map

X : {Local flows on M } → Vect (M ), Φ 7→ XΦ ,

is a surjection. Moreover, if Φi : Ni → M (i=1,2) are two local flows such that XΦ1 = XΦ2 ,
then Φ1 = Φ2 on N1 ∩ N2 .

Proof. Surjectivity. Let X be a vector field on M . An integral curve for X is a smooth

curve γ : (a, b) → M such that
γ̇(t) = X(γ(t)).
In local coordinates (xi ) over on open subset U ⊂ M this condition can be rewritten as


ẋi (t) = X i x1 (t), ..., xn (t) , ∀i = 1, . . . , n, (3.1.1)
∂
where γ(t) = (x1 (t), ..., xn (t)), and X = X i ∂x i . The above equality is a system of ordinary
differential equations. Classical existence results (see e.g. [12, 60]) show that, for any
precompact open subset K ⊂ U , there exists ε > 0 such that, for all x ∈ K, there exists
a unique integral curve for X, γx : (−ε, ε) → M satisfying

γx (0) = x. (3.1.2)

Moreover, as a consequence of the smooth dependence upon initial data we deduce that
the map
ΦK : NK = (−ε, ε) × K → M, (x, t) 7→ γx (t),
is smooth.
3.1. THE LIE DERIVATIVE 83

Now we can cover M by open, precompact, local coordinate neighborhoods (Kα )α∈A ,
and as above, we get smooth maps Φα : Nα = (−εα , εα ) × Kα → M solving the initial
value problem (3.1.1-2). Moreover, by uniqueness, we deduce

Φα = Φα on Nα ∩ Nβ .

Define [
N := Nα ,
α∈A

and set Φ : N → M , Φ = Φα on Nα .
The uniqueness of solutions of initial value problems for ordinary differential equations
implies that Φ satisfies all the conditions in the definition of a local flow. Tautologically,
X is the infinitesimal generator of Φ. The second part of the proposition follows from the
uniqueness in initial value problems. ⊔
⊓

The family of local flows on M with the same infinitesimal generator X ∈ Vect(M ) is
naturally ordered according to their domains,

(Φ1 : N1 → M ) ≺ (Φ2 : N2 → M )

if and only if N1 ⊂ N2 . This family has a unique maximal element which is called the
local flow generated by X, and it is denoted by ΦX .
Exercise 3.1.7. Consider the unit sphere

S 2 = (x, y, z) ∈ R3 ; x2 + y 2 + z 2 = 1 .

For every point p ∈ S 2 we denote by X(p) ∈ Tp R3 , the orthogonal projection of the vector
k = (0, 0, 1) onto Tp S 2 .
(a) Prove that p 7→ X(p) is a smooth vector field on S 2 , and then describe it in cylindrical
coordinates (z, θ), where

x = r cos θ, y = r sin θ, r = (x2 + y 2 )1/2 .

(b) Describe explicitly the flow generated by X. ⊔

⊓

3.1.2 The Lie derivative

Let X be a vector field on the smooth n-dimensional manifold M and denote by Φ =
ΦX the local flow it generates. For simplicity, we assume Φ is actually a flow so its
domain is R × M . The local flow situation is conceptually identical, but notationally more
complicated.
For each t ∈ R, the map Φt is a diffeomorphism of M and so it induces a push-forward
map on the space of tensor fields. If S is a tensor field on M we define its Lie derivative
along the direction given by X as
1

LX Sm := − lim (Φt∗ S)m − Sm ∀m ∈ M. (3.1.3)
t→0 t
84 CHAPTER 3. CALCULUS ON MANIFOLDS

Intuitively, LX S measures how fast is the flow Φ changing1 the tensor S.

If the limit in (3.1.3) exists, then one sees that LX S is a tensor of the same type as S.
To show that the limit exists, we will provide more explicit descriptions of this operation.
Lemma 3.1.8. For any X ∈ Vect (M ) and f ∈ C ∞ (M ) we have

Xf := LX f = hdf, Xi = df (X).

Above, h•, •i denotes the natural duality between T ∗ M and T M ,

h•, •i : C ∞ (T ∗ M ) × C ∞ (T M ) → C ∞ (M ),

C ∞ (T ∗ M ) × C ∞ (T M ) ∋ (α, X) 7→ α(X) ∈ C ∞ (M ).
In particular, LX is a derivation of C ∞ (M ).

Proof. Let Φt = ΦtX be the local flow generated by X. Assume for simplicity that it is
defined for all t. The map Φt acts on C ∞ (M ) by the pullback of its inverse, i.e.

Φt∗ = (Φ−t )∗ .

Hence, for point p ∈ M we have

1 d
LX f (p) = lim (f (p) − f (Φ−t p)) = − |t=0 f (Φ−t p) = hdf, Xip . ⊔
⊓
t→0 t dt
Exercise 3.1.9. Prove that any derivation of the algebra C ∞ (M ) is of the form LX for
some X ∈ Vect(M ), i.e.
Der (C ∞ (M )) ∼
= Vect (M ). ⊔
⊓

Lemma 3.1.10. Let X, Y ∈ Vect (M ). Then the Lie derivative of Y along X is a new
vector field LX Y which, viewed as a derivation of C ∞ (M ), coincides with the commutator
of the two derivations of C ∞ (M ) defined by X and Y i.e.

LX Y f = [X, Y ]f, ∀ f ∈ C ∞ (M ).

The vector field [X, Y ] = LX Y is called the Lie bracket of X and Y . In particular the Lie
bracket induces a Lie algebra structure on Vect (M ).

Proof. We will work in local coordinates (xi ) near a point m ∈ M so that

X = X i ∂xi and Y = Y j ∂xj .

We first describe the commutator [X, Y ]. If f ∈ C ∞ (M ), then

∂f ∂f
[X, Y ]f = (X i ∂xi )(Y j j
) − (Y j ∂xj )(X i i )
∂x ∂x
1
Arnold refers to the Lie derivative LX as the “fisherman’s derivative”. Here is the intuition behind
this very suggestive terminology. We place an observer (fisherman) at a fixed point p ∈ M , and we let him
keep track of the the sizes of the tensor S carried by the flow at the point p. The Lie derivatives measures
the rate of change in these sizes.
3.1. THE LIE DERIVATIVE 85
   
i ∂2f
j
j
i ∂Y ∂f
2
i j ∂ f
i
j ∂X ∂f
= XY + X − XY +Y ,
∂xi ∂xj ∂xi ∂xj ∂xi ∂xj ∂xj ∂xi
so that the commutator of the two derivations is the derivation defined by the vector field
 k k

i ∂Y j ∂X
[X, Y ] = X −Y ∂xk . (3.1.4)
∂xi ∂xj

Note in particular that [∂xi , ∂xj ] = 0, i.e., the basic vectors ∂xi commute as derivations.
So far we have not proved the vector field in (3.1.4) is independent of coordinates. We
will achieve this by identifying it with the intrinsically defined vector field LX Y .
Set γ(t) = Φt m so that we have a parametrization γ(t) = (xi (t)) with ẋi = X i . Then


Φ−t m = γ(−t) = γ(0) − γ̇(0)t + O(t2 ) = xi0 − tX i + O(t2 ) ,

and
j ∂Y j
Yγ(−t) + O(t2 ).
= Ymj − tX i (3.1.5)
∂xi
Note that Φ−t
∗ : Tγ(0) M → Tγ(−t) M is the linearization of the map


(xi ) 7→ xi − tX i + O(t2 ) ,

so it has a matrix representation

 
∂X i
Φ−t
∗ =1−t + O(t2 ). (3.1.6)
∂xj i,j

In particular, using the geometric series

(1 − A)−1 = 1 + A + A2 + · · · ,

where A is a matrix of operator norm strictly less than 1, we deduce that the differential

Φt∗ = (Φ−t
∗ )
−1
: Tγ(−t) M → Tγ(0) M,

has the matrix form  

∂X i
Φt∗ =1+t + O(t2 ). (3.1.7)
∂xj i,j

Using (3.1.7) in (3.1.5) we deduce

 

k i ∂Y
k
j ∂X
k
Ymk − Φt∗ YΦ−t m =t X −Y + O(t2 ).
∂xi ∂xj
This concludes the proof of the lemma. ⊔
⊓

Lemma 3.1.11. For any differential form ω ∈ Ω1 (M ), and any vector fields X, Y ∈
V ect(M ) we have
(LX ω)(Y ) = LX (ω(Y )) − ω([X, Y ]), (3.1.8)
where X · ω(Y ) denotes the (Lie) derivative of the function ω(Y ) along the vector field X.
86 CHAPTER 3. CALCULUS ON MANIFOLDS

Proof. Denote by Φt the local flow generated by X. We have Φt∗ ω = (Φ−t )∗ ω, i.e., for any
p ∈ M , and any Y ∈ Vect (M ), we have


(Φt∗ ω)p (Yp ) = ωΦ−t p Φ−t
∗ Yp .

Fix a point p ∈ M , and choose local coordinates (xi ) near p. Then

X X X
ω= ω i dxi , X = X i ∂xi , Y = Y i ∂xi .
i i i

Denote by γ(t) the path t 7→ Φt (p). We set ωi (t) = ωi (γ(t)), X0i = X i (p), and Y0i = Y i (p).
Using (3.1.6) we deduce
X  X ∂X i j 
(Φt∗ ω)p (Yp ) = ωi (−t) · Y0i − t Y 0 + O(t 2
) .
∂xj
i j

Hence
d X X ∂X i
−(LX ω)Y = |t=0 (Φt∗ ω)p (Yp ) = − ω̇i (0)Y0i − ωi (0) j Y0j .
dt ∂x
i i,j

One the other hand, we have

d X X X ∂Y i
X · ω(Y ) = |t=0 ωi (t)Y i (t) = ω̇i (0)Y0i + ωi (0)X0j j .
dt ∂x
i i i,j

We deduce that
X  i ∂X i j 
j ∂Y
X · ω(Y ) − (LX ω)Y = ωi (0) X0 j − Y = ωp ([X, Y ]p ). ⊔
⊓
∂x ∂xj 0
i,j

Observe that if S, T are two tensor fields on M such that both LX S and LX T exist,
then using (3.1.3) we deduce that LX (S ⊗ T ) exists, and

LX (S ⊗ T ) = LX S ⊗ T + S ⊗ LX T. (3.1.9)

Since any tensor field is locally a linear combination of tensor monomials of the form

X1 ⊗ · · · ⊗ Xr ⊗ ω1 ⊗ · · · ⊗ ωs , Xi ∈ Vect (M ), ωj ∈ Ω1 (M ),

we deduce that the Lie derivative LX S exists for every X ∈ Vect (M ), and any smooth
tensor field S. We can now completely describe the Lie derivative on the algebra of tensor
fields.
Proposition 3.1.12. Let X be a vector field on the smooth manifold M . Then the Lie
derivative LX is the unique derivation of T∗∗ (M ) with the following properties.
(a) LX f = hdf, Xi = Xf , ∀f ∈ C ∞ (M ).
(b) LX Y = [X, Y ], ∀X, Y ∈ Vect (M ).
r+1
(c) LX commutes with the contraction tr : Ts+1 (M ) → Tsr (M ).
Moreover, LX is a natural operation, i.e., for any diffeomorphism φ : M → N we have
φ∗ ◦ LX = Lφ∗ X ◦ φ∗ , ∀X ∈ Vect (M ), i.e., φ∗ (LX ) = Lφ∗ X .
3.1. THE LIE DERIVATIVE 87

Proof. The fact that LX is a derivation follows from (3.1.9). Properties (a) and (b) were
proved above. As for part (c), in its simplest form, when T = Y ⊗ ω, where Y ∈ Vect (M ),
and ω ∈ Ω1 (M ), the equality
LX tr T = tr LX T
is equivalent to
LX (ω(Y )) = (LX ω)(Y ) + ω(LX (Y ), (3.1.10)
which is precisely (3.1.8).
Since LX is a derivation of the algebra of tensor fields, its restriction to C ∞ (M ) ⊕
Vect (M ) ⊕ Ω1 (M ) uniquely determines the action on the entire algebra of tensor fields
which is generated by the above subspace. The reader can check easily that the general
case of property (c) follow from this observation coupled with the product rule (3.1.9).
The naturality of LX is another way of phrasing the coordinate independence of this
operation. We leave the reader to fill in the routine details. ⊔
⊓

Corollary 3.1.13. For any X, Y ∈ Vect (M ) we have

[LX , LY ] = L[X,Y ] ,

as derivations of the algebra of tensor fields on M . In particular, this says that Vect (M )
as a space of derivations of T∗∗ (M ) is a Lie subalgebra of the Lie algebra of derivations.

Proof. The commutator [LX , LY ] is a derivation (as a commutator of derivations). By

Lemma 3.1.10, [LX , LY ] = L[X,Y ] on C ∞ (M ). Also, a simple computation shows that

[LX , LY ]Z = L[X,Y ] Z, ∀Z ∈ Vect (M ),

so that [LX , LY ] = L[X,Y ] on Vect (M ). Finally, since the contraction commutes with both
LX and LY it obviously commutes with LX LY − LY LX . The corollary is proved. ⊔
⊓

Exercise 3.1.14. Prove that the map

D : Vect (M ) ⊕ End (T M ) → Der(T∗∗ (M ))

given by D(X, S) = LX + S is well defined and is a linear isomorphism. Moreover,

[D(X1 , S1 ), D(X2 , S2 )] = D([X1 , X2 ], [S1 , S2 ]). ⊔

⊓

The Lie derivative LX is a derivation of T∗∗ with the remarkable property

LX (Ω∗ (M )) ⊂ Ω∗ (M ).

The wedge product makes Ω∗ (M ) a s-algebra, and it is natural to ask whether LX is an

s-derivation with respect to this product.
Proposition 3.1.15. The Lie derivative along a vector field X is an even s-derivation of
Ω∗ (M ), i.e.
LX (ω ∧ η) = (LX ω) ∧ η + ω ∧ (LX η), ∀ω, η ∈ Ω∗ (M ).
88 CHAPTER 3. CALCULUS ON MANIFOLDS

Proof. As in Subsection 2.2.2, denote by A the anti-symmetrization operator A : (T ∗ M )⊗k →

Ωk (M ). The statement in the proposition follows immediately from the straightforward
observation that the Lie derivative commutes with this operator (which is a projector).
We leave the reader to fill in the details. ⊔
⊓

Exercise 3.1.16. Let M be a smooth manifold, and suppose that Φ, Ψ : R × M → M

are two smooth flows on M with infinitesimal generators X and respectively Y . We say
that the two flows commute if

Φt ◦ Ψs = Ψs ◦ Ψt , ∀s, t ∈ R.

Prove that if
 
p ∈ M ; X(p) = 0 = p ∈ M ; Y (p) = 0 .
the two flows commute if and only if [X, Y ] = 0. ⊔
⊓

3.1.3 Examples
Example 3.1.17. Let ω = ωi dxi be a 1-form on Rn . If X = X j ∂xj is a vector field on
Rn , then LX ω = (LX ω)k dxk is defined by
 
∂X i
(LX ω)k = (LX ω)(∂xk ) = Xω(∂xk ) − ω(LX ∂xk ) = X · ωk + ω ∂ i .
∂xk x

Hence  
∂ωk ∂X j
LX ω = Xj j + ωj k dxk .
∂x ∂x
P
In particular, if X = ∂xi = j δij ∂xj , then

n
X ∂ωk
LX ω = L∂xi ω = dxk .
∂xi
k=1
P i∂ ,
If X is the radial vector field X = ix xi then
X
LX ω = (X · ωk + ωk )dxk . ⊔
⊓
k

Example 3.1.18. Consider a smooth vector field X = F ∂x + G∂y + H∂z on R3 . We want

to compute LX dv, where dv is the volume form on R3 , dv = dx ∧ dy ∧ dz. Since LX is an
even s-derivation of Ω∗ (M ), we deduce

LX (dx ∧ dy ∧ dz) = (LX dx) ∧ dy ∧ dz + dx ∧ (LX dy) ∧ dz + dx ∧ dy ∧ (LX dz).

Using the computation in the previous example we get

∂F ∂F ∂F
LX (dx) = dF := dx + dy + dz, LX (dy) = dG, LX (dz) = dH
∂x ∂y ∂z
3.1. THE LIE DERIVATIVE 89

so that  
∂F ∂G ∂H
LX (dv) = + + dv = (div X)dv.
∂x ∂y ∂z
In particular, we deduce that if div X = 0, the local flow generated by X preserves the
form dv. We will get a better understanding of this statement once we learn integration
on manifolds, later in this chapter. ⊔
⊓

Example 3.1.19. (The exponential map of a Lie group). Consider a Lie group
G. Any element g ∈ G defines two diffeomorphisms of G: the left (Lg ), and the right
translation (Rg ) on G,

Lg (h) = g · hy, Rg (h) = h · g, ∀h ∈ G.

A tensor field T on G is called left (respectively right) invariant if for any g ∈ G (Lg )∗ T =
T (respectively (Rg )∗ T = T ). The set of left invariant vector fields on G is denoted by
LG .
The naturality of the Lie bracket implies

(Lg )∗ [X, Y ] = [(Lg )∗ X, (Lg )∗ Y ],

so that ∀X, Y ∈ LG , [X, Y ] ∈ LG . Hence LG is a Lie subalgebra of Vect (G). It is called

called the Lie algebra of the group G.
Fact 1. dim LG = dim G. Indeed, the left invariance implies that the restriction map
LG → T1 G, X 7→ X1 is an isomorphism (Exercise). We will often find it convenient to
identify the Lie algebra of G with the tangent space at 1.
Fact 2. Any X ∈ LG defines a local flow ΦtX on G that is is defined for all t ∈ R. In
other words, ΦtX is a flow. (Exercise) Set

exp(tX) : =ΦtX (1).

We thus get a map

exp : T1 G ∼
= LG → G, X 7→ exp(X)
called the exponential map of the group G.
Fact 3. ΦtX (g) = g · exp (tX), i.e.,

ΦtX = Rexp (tX) .

Indeed, it suffices to check that

d
|t=0 (g · exp (tX)) = Xg .
dt
We can write (g · exp(tX)) = Lg exp(tX) so that

d  d 
|t=0 (Lg exp(tX)) = (Lg )∗ |t=0 exp(tX) = (Lg )∗ X = Xg , (left invariance).
dt dt
90 CHAPTER 3. CALCULUS ON MANIFOLDS

The reason for the notation exp (tX) is that when G = GL(n, K), then the Lie algebra of G
is the Lie algebra gl(n, K) of all n × n matrices with the bracket given by the commutator
of two matrices, and for any X ∈ LG we have (Exercise)
X 1
exp (X) = eX = Xk. ⊔
⊓
k!
k≥0

Exercise 3.1.20. Prove the statements left as exercises in the example above. ⊔
⊓

Exercise 3.1.21. Let G be a matrix Lie group, i.e., a Lie subgroup of some general linear
group GL(N, K). This means the tangent space T1 G can be identified with a linear space
of matrices. Let X, Y ∈ T1 G, and denote by exp(tX) and exp(tY ) the 1-parameter groups
with they generate, and set

g(s, t) = exp(sX) exp(tY ) exp(−sX) exp(−tY ).

1. Show that
gs,t = 1 + [X, Y ]alg st + O((s2 + t2 )3/2 ) as s, t → 0,
where the bracket [X, Y ]alg (temporarily) denotes the commutator of the two ma-
trices X and Y .
2. Denote (temporarily) by [X, Y ]geom the Lie bracket of X and Y viewed as left in-
variant vector fields on G. Show that at 1 ∈ G

[X, Y ]alg = [X, Y ]geom .

3. Show that o(n) ⊂ gl(n, R) (defined in Section 1.2.2) is a Lie subalgebra with respect
to the commutator [ · , · ]. Similarly, show that u(n), su(n) ⊂ gl(n, C) are real Lie
subalgebras of gl(n, C), while su(n, C) is even a complex Lie subalgebra of gl(n, C).
4. Prove that we have the following isomorphisms of real Lie algebras,

= o(n), LU (n) ∼
LO(n) ∼ = u(n), LSU (n) ∼
= su(n) and LSL(n,C) ∼
= sl(n, C).

⊔
⊓

Remark 3.1.22. In general, in a non-commutative matrix Lie group G, the traditional

equality
exp(tX) exp(tY ) = exp(t(X + Y ))
no longer holds. Instead, one has the Campbell-Hausdorff formula


exp(tX) · exp(tY ) = exp td1 (X, Y ) + t2 d2 (X, Y ) + t3 d3 (X, Y ) + · · · ,
where dk are homogeneous polynomials of degree k in X, and Y with respect to the
multiplication between X and Y given by their bracket. The dk ’s are usually known as
Dynkin polynomials. For example,
1
d1 (X, Y ) = X + Y, d2 (X, Y ) = [X, Y ],
2
3.2. DERIVATIONS OF Ω• (M ) 91

1
d3 (X, Y ) = ([X, [X, Y ]] + [Y, [Y, X]]) etc.
12
For more details we refer to [58, 113]. ⊔
⊓

3.2 Derivations of Ω•(M)

3.2.1 The exterior derivative
The super-algebra of exterior forms on a smooth manifold M has additional structure,
and in particular, its space of derivations has special features. This section is devoted
precisely to these new features.
The Lie derivative along a vector field X defines an even derivation in Ω• (M ). The
vector field X also defines, via the contraction map, an odd derivation iX , called the
interior derivation along X, or the contraction by X,

iX ω : =tr (X ⊗ ω), ∀ω ∈ Ωr (M ).

More precisely, iX ω is the (r − 1)-form determined by

(iX ω)(X1 , . . . , Xr−1 ) = ω(X, X1 , . . . , Xr−1 ), ∀X1 , . . . , Xr−1 ∈ Vect (M ).

The fact that iX is an odd s-derivation is equivalent to

iX (ω ∧ η) = (iX ω) ∧ η + (−1)deg ω ω ∧ (iX η), ∀ω, η ∈ Ω∗ (M ).

Often the contraction by X is denoted by X .

Exercise 3.2.1. Prove that the interior derivation along a vector field is a s-derivation.
⊔
⊓

Proposition 3.2.2. (a) [iX , iY ]s = iX iY + iY iX = 0.

(b) The super-commutator of LX and iY as s-derivations of Ω∗ (M ) is given by

[LX , iY ]s = LX iY − iY LX = i[X,Y ] . ⊔
⊓

The proof uses the fact that the Lie derivative commutes with the contraction operator,
and it is left to the reader as an exercise.
The above s-derivations by no means exhaust the space of s-derivations of Ω• (M ). In
fact we have the following fundamental result.
Proposition 3.2.3. There exists an odd s-derivation d on the s-algebra of differential
forms Ω• ( · ) uniquely characterized by the following conditions.

(a) For any smooth function f ∈ Ω0 (M ), df coincides with the differential of f .

(b) d2 = 0.

(c) d is natural, i.e., for any smooth function φ : N → M , and for any form ω on M ,
we have
dφ∗ ω = φ∗ dω(⇐⇒ [φ∗ , d] = 0).
92 CHAPTER 3. CALCULUS ON MANIFOLDS

The derivation d is called the exterior derivative.

Proof. Uniqueness. Let U be a local coordinate chart on M n with local coordinates

(x1 , ... , xn ). Then, over U , any r-form ω can be described as
X
ω= ωi1 ...ir dxi1 ∧ ... ∧ dxir .
1≤i1 <···<ir ≤n

Since d is an s-derivation, and d(dxi ) = 0, we deduce that, over U

X
dω = (dωi1 ...ir ) ∧ (dxi1 ∧ · · · ∧ dxir )
1≤i1 <···<ir ≤n

X  
∂ωi1 ...ir i
= dx ∧ (dxi1 ∧ · · · ∧ dxir ). (3.2.1)
∂xi
1≤i1 <···<ir ≤n

Thus, the form dω is uniquely determined on any coordinate neighborhood, and this
completes the proof of the uniqueness of d.

Existence. Consider an r-form ω. For each coordinate neighborhood U we define dω |U as

in (3.2.1). To prove that this is a well defined operation we must show that, if U , V are
two coordinate neighborhoods, then

dω |U = dω |V on U ∩ V.

Denote by (x1 , ... , xn ) the local coordinates on U , and by (y 1 , ... , y n ) the local coordinates
along V , so that on the overlap U ∩ V we can describe the y’s as functions of the x’s. Over
U we have X
ω= ωi1 ...ir dxi1 ∧ ... ∧ dxir
1≤i1 <···<ir ≤n
X  
∂ωi1 ...ir i
dω = dx ∧ (dxi1 ∧ ... ∧ dxir ),
∂xi
1≤i1 <···<ir ≤n

while over V we have

X
ω= bj1 ...jr dy j1 ∧ · · · ∧ dy jr
ω
1≤j1 <···<jr ≤n

X  
bj1 ...jr j
∂ω
dω = dy (dy j1 ∧ · · · ∧ dy jr ).
∂y j
1≤j1 <···<jr ≤n

bj1 ...jr are skew-symmetric, i.e., ∀σ ∈ Sr

The components ωi1 ...ir and ω

ωiσ(1) ...iσ(r) = ǫ(σ)ωi1 ...ir ,

b ′ s. Since ω |U = ω |V over U ∩ V we deduce

and similarly for the ω

∂y j1 ∂y jr
ωi1 ...ir = · · · bj ...j .
ω
∂xi1 ∂xir 1 r
3.2. DERIVATIONS OF Ω• (M ) 93

Hence
X r  
∂ωi1 ...ir ∂y j1 ∂ 2 y jk ∂y jr ∂y j1 ∂y jr ∂ ω
bj1 ...jr
= · · · · · · b
ω j 1 ...j r + · · · ,
∂xi ∂x i 1 i
∂x ∂x i k ∂x i r ∂x i 1 ∂x i r ∂xi
k=1

where in the above equality we also sum over the indices j1 , ..., jr according to Einstein’s
convention. We deduce
X X ∂ωi1 ...ir i
dx ∧ dxi1 ∧ ... ∧ dxir
∂xi
i 1≤i1 <···<ir ≤n

r
XX ∂y j1 ∂ 2 y jk ∂y jr
= ··· · · · bj ...j dxi ∧ dxi1 ∧ ... ∧ dxir
ω
∂xi1 ∂xi ∂xik ∂xir 1 r
i k=1
r
XX ∂y j1 ∂y jr ∂ ω
bj1 ...jr i
+ ··· dx ∧ dxi1 ∧ ... ∧ dxir . (3.2.2)
∂xi1 ∂x i r ∂xi
i k=1

Notice that
∂2 ∂2
= ,
∂xi ∂xik ∂xik ∂xi
while dxi ∧ dxik = −dxik ∧ dxi so that the first term in the right hand side of (3.2.2)
vanishes. Consequently on U ∩ V

∂ωi1 ...ir i i1 ir ∂y j1 ∂y jr ∂ ω
bj1 ...jr
i
dx ∧ dx · · · ∧ dx = i
· · · i
∧ dxi ∧ dxi1 · · · dxir
∂x ∂x 1 ∂x r ∂xi
   j1   jr 
∂ωbj1 ...jr i ∂y i1 ∂y ir
= dx ∧ dx ∧ ··· ∧ dx
∂xi ∂xi1 ∂xir
bj1 ...jr j
∂ω
ωj1 ...jr ) ∧ dy j1 ∧ · · · ∧ dy jr =
= (db dy ∧ dy j1 ∧ · · · ∧ dy jr .
∂y j
This proves dω |U = dω |V over U ∩ V . We have thus constructed a well defined linear map

d : Ω• (M ) → Ω•+1 (M ).

To prove that d is an odd s-derivation it suffices to work in local coordinates and show
that the (super)product rule on monomials.
Let θ = f dxi1 ∧ · · · ∧ dxir and ω = gdxj1 ∧ · · · ∧ dxjs . We set for simplicity

dxI := dxi1 ∧ · · · ∧ dxir and dxJ := dxj1 ∧ · · · ∧ dxjs .

Then
d(θ ∧ ω) = d(f gdxI ∧ dxJ ) = d(f g) ∧ dxI ∧ dxJ
= (df · g + f · dg) ∧ dxI ∧ dxJ
= df ∧ dxI ∧ dxJ + (−1)r (f ∧ dxI ) ∧ (dg ∧ dxJ )
= dθ ∧ ω + (−1)deg θ θ ∧ dω.
94 CHAPTER 3. CALCULUS ON MANIFOLDS

We now prove d2 = 0. We check this on monomials f dxI as above.

d2 (f dxI ) = d(df ∧ dxI ) = (d2 f ) ∧ dxI .

Thus, it suffices to show d2 f = 0 for all smooth functions f . We have

∂f 2
d2 f = dxi ∧ dxj .
∂xi ∂xj
The desired conclusion follows from the identities

∂f 2 ∂f 2
= and dxi ∧ dxj = −dxj ∧ dxi .
∂xi ∂xj ∂xj ∂xi
P
Finally, let φ be a smooth map N → M and ω = I ωI dxI be an r-form on M . Here I
runs through all ordered multi-indices 1 ≤ i1 < · · · < ir ≤ dim M . We have
X

dN (φ∗ ω) = dN (φ∗ ωI ) ∧ φ∗ (dxI ) + φ∗ ω I ∧ d(φ∗ dxI ) .
I

For functions, the usual chain rule gives dN (φ∗ ωI ) = φ∗ (dM ωI ). In terms of local coor-
dinates (xi ) the map φ looks like a collection of smooth n functions φi = xi ◦ φ = φ∗ (xi )
and since the pullback is a morphism with respect to the exterior multiplication we get

φ∗ (dxI ) = dφI = φ∗ dxi1 ∧ · · · ∧ φ∗ dxir .

dN φi1 ∧ · · · ∧ dN φir .
In particular, dN (dφI ) = 0. We put all the above together and we deduce

dN (φ∗ ω) = φ∗ (dM ω I ) ∧ dφI = φ∗ (dM ω I ) ∧ φ∗ dxI = φ∗ (dM ω).

The proposition is proved. ⊔

⊓

Proposition 3.2.4. The exterior derivative satisfies the following relations.

1.
[d, d]s = 2d2 = 0.

2. (Cartan’s homotopy formula)

[d, iX ]s = diX + iX d = LX , ∀X ∈ Vect (M ).

3.
[d, LX ]s = dLX − LX d = 0, ∀X ∈ Vect (M ).

Above, [ , ]s denotes the super-commutator in the s-algebra of real endomorphisms of

Ω• (M )
3.2. DERIVATIONS OF Ω• (M ) 95

Proof. To prove the homotopy formula set

D := [d, iX ]s = diX + iX d.

The operator D is an even s-derivation of Ω∗ (M ). It is a local s-derivation, i.e., if ω ∈

Ω∗ (M ) vanishes on some open set U then Dω vanishes on that open set as well. The
reader can check easily by direct computation that Dω = LX ω, ∀ω ∈ Ω0 (M ) ⊕ Ω1 (M ).
The homotopy formula is now a consequence of the following technical result left to the
reader as an exercise.
Lemma 3.2.5. Let D, D′ be two local s-derivations of Ω• (M ) which have the same parity,
i.e., they are either both even or both odd. If D = D′ on Ω0 (M ) ⊕ Ω1 (M ), then D = D′
on Ω• (M ). ⊔
⊓

Part (iii) of the proposition is proved in a similar way. ⊔

⊓

An immediate consequence of the homotopy formula is the following invariant descrip-

tion of the exterior derivative:
r
X
(dω)(X0 , X1 , ... , Xr ) = (−1)i Xi (ω(X0 , ..., X̂i , ..., Xr ))
i=0 (3.2.3)
X
i+j
+ (−1) ω([Xi , Xj ], X0 , ..., X̂i , ..., X̂j , ..., Xr ).
0≤i<j≤r

Above, the hat indicates that the corresponding entry is missing.

Equality (3.2.3) is a simple consequence of the homotopy formula. We prove it in two
special case r = 1 and r = 2.
The case r = 1. Let ω be an 1-form and let X, Y ∈ Vect (M ). We deduce from the
homotopy formula

dω(X, Y ) = (iX dω)(Y ) = (LX ω)(Y ) − (dω(X))(Y ).

On the other hand, since LX commutes with the contraction operator, we deduce

Xω(Y ) = LX (ω(Y )) = (LX ω)(Y ) + ω([X, Y ]).

Hence

dω(X, Y ) = Xω(Y ) − ω([X, Y ]) − (dω(X))(Y ) = Xω(Y ) − Y ω(X) − ω([X, Y ]).

This proves (3.2.3) in the case r = 1.

The case r = 2. Consider a 2-form ω and three vector fields X, Y and Z. We deduce
from the homotopy formula

(dω)(X, Y, Z) = (iX dω)(Y, Z) = (LX − diX )ω(Y, Z). (3.2.4)

Since LX commutes with contractions we deduce

(LX ω)(Y, X) = X(ω(Y, Z)) − ω([X, Y ], Z) − ω(Y, [X, Z]). (3.2.5)

96 CHAPTER 3. CALCULUS ON MANIFOLDS

We substitute (3.2.5) into (3.2.4) and we get

(dω)(X, Y, Z) = X(ω(Y, Z)) − ω([X, Y ], Z) − ω(Y, [X, Z]) − d(iX ω)(Y, X). (3.2.6)

We apply now (3.2.3) for r = 1 to the 1-form iX ω. We get

d(iX ω)(Y, X) = Y (iX ω(Z)) − Z(iX ω(Y )) − (iX ω)([Y, Z])

= Y ω(X, Z) − Zω(X, Y ) − ω(X, [Y, Z]). (3.2.7)

If we use (3.2.7) in (3.2.6) we deduce

(dω)(X, Y, Z) = Xω(Y, Z) − Y ω(X, Z) + Zω(X, Y )

−ω([X, Y ], Z) + ω([X, Z], Y ) − ω([Y, Z], X). (3.2.8)

The general case in (3.2.3) can be proved by induction.
Exercise 3.2.6. Prove Lemma 3.2.5. ⊔
⊓

Exercise 3.2.7. Finish the proof of (3.2.3) in the general case. ⊔

⊓

3.2.2 Examples
Example 3.2.8. (The exterior derivative in R3 ).
(a) Let f ∈ C ∞ (R3 ). Then

∂f ∂f ∂f
df = dx + dy + dz.
∂x ∂y ∂z

The differential df looks like the gradient of f .

(b) Let ω ∈ Ω1 (R3 ), ω = P dx + Qdy + Rdz. Then

dω = dP ∧ dx + dQ ∧ dy + dR ∧ dz
     
∂Q ∂P ∂R ∂Q ∂P ∂R
= − dx ∧ dy + − dy ∧ dz + − dz ∧ dx,
∂x ∂y ∂y ∂z ∂z ∂x
so that dω looks very much like a curl.
(c) Let ω = P dy ∧ dz + Qdz ∧ dx + Rdx ∧ dy ∈ Ω2 (R3 ). Then
 
∂P ∂Q ∂R
dω = + + dx ∧ dy ∧ dz.
∂x ∂y ∂z

This looks very much like a divergence. ⊔

⊓

Example 3.2.9. Let G be a connected Lie group. In Example 3.1.19 we defined the Lie
algebra LG of G as the space of left invariant vector fields on G. Set

Ωrlef t (G) = left invariant r-forms on G.

3.3. CONNECTIONS ON VECTOR BUNDLES 97

= Ω1lef t (G). If we identify L∗G ∼

In particular, L∗G ∼ = T1∗ G, then we get a natural isomorphism

Ωrlef t (G) ∼
= Λr L∗G .

The exterior derivative of a form in Ω∗lef t can be described only in terms of the algebraic
structure of LG .
Indeed, let ω ∈ L∗G = Ω1lef t (G). For X, Y ∈ L∗G we have (see (3.2.3) )

dω(X, Y ) = Xω(Y ) − Y ω(X) − ω([X, Y ]).

Since ω, X and Y are left invariant, the scalars ω(X) and ω(Y ) are constants. Thus, the
first two terms in the above equality vanish so that

dω(X, Y ) = −ω([X, Y ]).

More generally, if ω ∈ Ωrlef t , then the same arguments applied to (3.2.3) imply that for all
X0 , ..., Xr ∈ LG we have
X
dω(X0 , X1 , ..., Xr ) = (−1)i+j ω([Xi , Xj ], X1 , ..., X̂i , ..., X̂j , ..., Xr ). (3.2.9)
0≤i<j≤r

⊔
⊓

3.3 Connections on vector bundles

3.3.1 Covariant derivatives
We learned several methods of differentiating tensor objects on manifolds. However, the
tensor bundles are not the only vector bundles arising in geometry, and very often one is
interested in measuring the “oscillations” of sections of vector bundles.
Let E be a K-vector bundle over the smooth manifold M (K = R, C). Technically, one
would like to have a procedure of measuring the rate of change of a section u of E along a
direction described by a vector field X. For such an arbitrary E, we encounter a problem
which was not present in the case of tensor bundles. Namely, the local flow generated by
the vector field X on M no longer induces bundle homomorphisms.
For tensor fields, the transport along a flow was a method of comparing objects in
different fibers which otherwise are abstract linear spaces with no natural relationship
between them.
To obtain something that looks like a derivation we need to formulate clearly what
properties we should expect from such an operation.

1. It should measure how fast is a given section changing along a direction given by a
vector field X. Hence it has to be an operator

∇ : Vect (M ) × C ∞ (E) → C ∞ (E), (X, u) 7→ ∇X u,

where we recall (see Definition 2.1.27) that C ∞ (E) denotes the space of smooth
sections of E over M .
98 CHAPTER 3. CALCULUS ON MANIFOLDS

2. If we think of the usual directional derivative, we expect that after “rescaling” the
direction X the derivative along X should only rescale by the same factor, i.e.,

∀f ∈ C ∞ (M ) : ∇f X u = f ∇X u.

3. Since ∇ is to be a derivation, it has to satisfy a sort of (Leibniz) product rule.

The only product that exists on an abstract vector bundle is the multiplication of a
section with a smooth function. Hence we require

∇X (f u) = (Xf )u + f ∇X u, ∀f ∈ C ∞ (M ), u ∈ C ∞ (E).

The conditions (a) and (b) can be rephrased as follows: for any u ∈ C ∞ (E), the map

∇u : Vect (M ) → C ∞ (E), X 7→ ∇X u,

is C ∞ (M )-linear so that it defines a bundle morphism (see Definition 2.1.30(b) and Ex-
ample 2.3.2)
∇u ∈ Hom(T M, E) ∼ = C ∞ (T ∗ M ⊗ E).
Summarizing, we can formulate the following definition.
Definition 3.3.1. A covariant derivative (or linear connection) on E is a K-linear map

∇ : C ∞ (E) → C ∞ (T ∗ M ⊗ E),

such that, ∀f ∈ C ∞ (M ), and ∀u ∈ C ∞ (E), we have

∇(f u) = df ⊗ u + f ∇u. ⊔
⊓

Example 3.3.2. Let KrM ∼ = Kr ×M be the rank r trivial vector bundle over M . The space
∞
C (KM ) of smooth sections coincides with the space C ∞ (M, Kr ) of Kr -valued smooth
functions on M . We can define

∇0 : C ∞ (M, Kr ) → C ∞ (M, T ∗ M ⊗ Kr )

∇0 (f1 , ..., fr ) = (df1 , ..., dfr ).

One checks easily that ∇ is a connection. This is called the trivial connection. ⊔
⊓

Remark 3.3.3. Let ∇0 , ∇1 be two connections on a vector bundle E → M . Then for

any α ∈ C ∞ (M ) the map

∇ = α∇1 + (1 − α)∇0 : C ∞ (E) → C ∞ (T ∗ ⊗ E)

is again a connection. ⊔
⊓

✍Notation. For any vector bundle F over M we set

Ωk (F ) : =C ∞ (Λk T ∗ M ⊗ F ).

We will refer to these sections as differential k-forms with coefficients in the vector bundle
F. ⊔
⊓
3.3. CONNECTIONS ON VECTOR BUNDLES 99

Proposition 3.3.4. Let E be a vector bundle. The space A(E) of linear connections on
E is an affine space modeled on Ω1 (End (E)).

Proof. We first prove that A (E) is not empty. To see this, choose an open cover {Uα } of
M such that E |Uα is trivial ∀α. Next, pick a smooth partition of unity (µβ ) subordinated
to this cover.
Since E |Uα is trivial, it admits at least one connection, the trivial one, as in the above
example. Denote such a connection by ∇α . Now define
X
∇ := µα ∇α .
α

One checks easily that ∇ is a connection so that A(E) is nonempty. To check that A (E)
is an affine space, consider two connections ∇0 and ∇1 . Their difference A = ∇1 − ∇0 is
an operator
A : C ∞ (E) → C ∞ (T ∗ M ⊗ E),
satisfying A(f u) = f A(u), ∀u ∈ C ∞ (E). Thus,



A ∈ C ∞ Hom(E, T ∗ M ⊗ E) ∼ = C ∞ (T ∗ M ⊗ E ∗ ⊗ E) ∼
= Ω1 (E ∗ ⊗ E) ∼
= Ω1 End(E) .


Conversely, given ∇0 ∈ A(E) and A ∈ Ω1 End(E) one can verify that the operator

∇A = ∇0 + A : C ∞ (E) → Ω1 (E).

is a linear connection. This concludes the proof of the proposition. ⊔

⊓

The tensorial operations on vector bundles extend naturally to vector bundles with
connections. The guiding principle behind this fact is the product formula.

• If Ei (i = 1, 2) are two bundles with connections ∇i , then E1 ⊗ E2 has a naturally

induced connection ∇E1 ⊗E2 uniquely determined by the product rule,

∇E1 ⊗E2 (u1 ⊗ u2 ) = (∇1 u1 ) ⊗ u2 + u1 ⊗ ∇2 u2 .

• A connection on a bundle E induced a connection ∇∗ on the dual bundle E ∗ deter-

mined by the identity

Xhv, ui = h∇∗X v, ui + hv, ∇X ui, ∀u ∈ C ∞ (E), v ∈ C ∞ (E ∗ ), X ∈ Vect (M ),

where
h•, •i : C ∞ (E ∗ ) × C ∞ (E) → C ∞ (M )
is the pairing induced by the natural duality between the fibers of E ∗ and E.

• Any connection ∇E on a vector bundle E induces a connection ∇End(E) on End (E) ∼

=
E ∗ ⊗ E defined by

(∇End(E) T )(u) = ∇E (T u) − T (∇E u) = [∇E , T ]u, (3.3.1)

∀T ∈ End(E), u ∈ C ∞ (E).
100 CHAPTER 3. CALCULUS ON MANIFOLDS

A metric h on a real vector bundle E → M is a section of E ∗ ⊗ E ∗ . A connection ∇

on E is said to be compatible with a metric h on E if
∗ ⊗E ∗
∇E h = 0.

If ∇ is compatible with a metric h then, for any u, v ∈ C ∞ (E) and any X ∈ Vect (M ) we
have
X · h(u, v) = h(∇X Du, v) + h(u, ∇X v). (3.3.2)
It is often useful to have a local description of a covariant derivative. This can be obtained
using Cartan’s moving frame method.
Let E → M be a K-vector bundle of rank r over the smooth manifold M . Pick a
coordinate neighborhood U such E |U is trivial. A moving frame 2 is a bundle isomorphism
φ : KrU = Kr × M → E |U .
Consider the sections eα = φ(δα ), α = 1, ..., r, where δα are the natural basic sections
of the trivial bundle KrU . As x moves in U , the collection (e1 (x), ..., er (x)) describes a basis
of the moving fiber Ex , whence the terminology moving frame. A section u ∈ C ∞ (E |U )
can be written as a linear combination

u = uα eα uα ∈ C ∞ (U, K).

Hence, if ∇ is a covariant derivative in E, we have

∇u = duα ⊗ eα + uα ∇eα .

Thus, the covariant derivative is completely described by its action on a moving frame.
To get a more concrete description, pick local coordinates (xi ) over U . Then ∇eα ∈
Ω1 (E |U ) so that we can write

∇eα = Γβiα dxi ⊗ eβ , Γβiα ∈ C ∞ (U, K).

Thus, for any section uα eα of E |U we have

∇u = duα ⊗ eα + Γβiα uα dxi ⊗ eβ . (3.3.3)

 
It is convenient to view Γβiα as an r × r-matrix valued 1-form, and we write this as
 
Γβiα = dxi ⊗ Γi .

The form Γ = dxi ⊗ Γi is called the connection 1-form associated to the choice of local
gauge. A moving frame allows us to identify sections of E |U with Kr -valued functions on
U , and we can rewrite (3.3.3) as
∇u = du + Γu. (3.3.4)
If the connection ∇ is compatible with a metric h and we choose the local moving frame
(eα ) defined to be h-orthonormal, then (3.3.2) implies that for every. vector field CX we
have
h(∇X eα , eβ ) = −h(eα , ∇X eβ ).
2
A moving frame is what physicists call a choice of local gauge.
3.3. CONNECTIONS ON VECTOR BUNDLES 101

Choosing local coordinates (xi ) we deduce that if ∇ is compatible with a metric h and
(eα ) is an orthonormal moving frame, then the associated connection 1-form
X
Γ= dxi ⊗ Γi
i

satisfies the additional property that Γi are skew-symmetric matrices.

A natural question arises: how does the connection 1-form changes with the change of
the local gauge?
Let f = (fα ) be another moving frame of E |U . The correspondence eα 7→ fα defines
an automorphism of E |U . Using the local frame e we can identify this correspondence
with a smooth map g : U → GL(r; K). The map g is called the local gauge transformation
relating e to f .
Let Γ̂ denote the connection 1-form corresponding to the new moving frame, i.e.,

∇fα = Γ̂βα fβ .

Consider a section σ of E |U . With respect to the local frame (eα ) the section σ has a
decomposition
σ = uα eα ,
while with respect to (fβ ) it has a decomposition

σ = ûβ fβ .

The two decompositions are related by

u = gû. (3.3.5)

Now, we can identify the E-valued 1-form ∇σ with a Kr -valued 1-form in two ways: either
using the frame e, or using the frame f . In the first case, the derivative ∇σ is identified
with the Kr -valued 1-form
du + Γu,
while in the second case it is identified with

dû + Γ̂û.

These two identifications are related by the same rule as in (3.3.5):

du + Γu = g(dû + Γ̂û).

Using (3.3.5) in the above equality we get

(dg)û + gdû + Γgû = gdû + gΓ̂û.

Hence
Γ̂ = g −1 dg + g−1 Γg.
The above relation is the transition rule relating two local gauge descriptions of the same
connection.
102 CHAPTER 3. CALCULUS ON MANIFOLDS

☞A word of warning. The identification

{moving f rames} ∼
= {local trivialization}

should be treated carefully. These are like an object and its image in a mirror, and there
is a great chance of confusing the right hand with the left hand.
∼
= ∼
=
More concretely, if tα : Eα −→ Kr × Uα (respectively tβ : Eβ −→ Kr × Uβ ) is a
trivialization of a bundle E over an open set Uα (respectively Uβ ), then the transition
map “from α to β” over Uα ∩ Uβ is gβα = tβ ◦ t−1 r
α . The standard basis in K , denoted by
(δi ), induces two local moving frames on E:

eα,i = t−1 −1
α (δi ) and eβ,i = tβ (δi ).

On the overlap Uα ∩ Uβ these two frames are related by the local gauge transformation

−1
eβ,i = gβα eα,i = gαβ eα,i .

This is precisely the opposite way the two trivializations are identified. ⊔
⊓

The above arguments can be reversed producing the following global result.

Proposition 3.3.5. Let E → M be a rank r smooth vector bundle, and (Uα ) a trivializing
cover with transition maps gαβ : Uα ∩Uβ → GL(r; K). Then any collection of matrix valued
1-forms Γα ∈ Ω1 (End KrUα ) satisfying

−1 −1 −1 −1
Γβ = (gαβ dgαβ ) + gαβ Γα gαβ = −(dgβα )gβα + gβα Γα gβα over Uα ∩ Uβ , (3.3.6)

uniquely defines a covariant derivative on E. ⊔

⊓

Exercise 3.3.6. Prove the above proposition. ⊔

⊓

We can use the local description in Proposition 3.3.5 to define the notion of pullback
of a connection. Suppose we are given the following data.

• A smooth map f : N → M .

• A rank r K-vector bundle E → M defined by the open cover (Uα ), and transition
maps gβα : Uα ∩ Uβ → GL(Kr ).


• A connection ∇ on E defined by the 1-forms Γα ∈ Ω1 End(KrUα ) satisfying the
gluing conditions (3.3.6).

Then, these data define a connection f ∗ ∇ on f ∗ E described by the open cover f −1 (Uα ),
transition maps gβα ◦ f and 1-forms f ∗ Γα . This connection is independent of the various
choices and it is called the pullback of ∇ by f
3.3. CONNECTIONS ON VECTOR BUNDLES 103

Example 3.3.7. (Complex line bundles). Let L → M be a complex line bundle

over the smooth manifold M . Let {Uα } be a trivializing cover with transition maps
zαβ : Uα ∩ Uβ → C∗ = GL(1, C). The bundle of endomorphisms of L, End (L) ∼ = L∗ ⊗ L
−1
is trivial since it can be defined by transition maps (zαβ ) ⊗ zαβ = 1. Thus, the space of
connections on L, A (L) is an affine space modelled by the linear space of complex valued
1-forms. A connection on L is simply a collection of C-valued 1-forms ω α on Uα related
on overlaps by
dzαβ
ωβ = + ω α = d log zαβ + ω α . ⊔
⊓
zαβ

3.3.2 Parallel transport

As we have already pointed out, the main reason we could not construct natural derivations
on the space of sections of a vector bundle was the lack of a canonical procedure of
identifying fibers at different points. We will see in this subsection that such a procedure
is all we need to define covariant derivatives. More precisely, we will show that once a
covariant derivative is chosen, it offers a simple way of identifying different fibers.
Let E → M be a rank r K-vector bundle and ∇ a covariant derivative on E. For any
smooth path γ : [0, 1] → M we will define a linear isomorphism Tγ : Eγ(0) → Eγ(1) called
the parallel transport along γ. More exactly, we will construct an entire family of linear
isomorphisms
Tt : Eγ(0) → Eγ(t) .
One should think of this Tt as identifying different fibers. In particular, if u0 ∈ Eγ(0)
then the path t 7→ ut = Tt u0 ∈ Eγ(t) should be thought of as a “constant” path. The
rigorous way of stating this “constancy” is via derivations: a quantity is “constant” if its
derivatives are identically 0. Now, the only way we know how to derivate sections is via
∇, i.e., ut should satisfy
d
∇ d ut = 0, where = γ̇.
dt dt
The above equation suggests a way of defining Tt . For any u0 ∈ Eγ(0) , and any t ∈ [0, 1],
define Tt u0 as the value at t of the solution of the initial value problem
(
∇ d u(t) = 0
dt . (3.3.7)
u(0) = u0

The equation (3.3.7) is a system of linear ordinary differential equations in disguise.

To see this, let us make the simplifying assumption that γ(t) lies entirely in some
coordinate neighborhood U with coordinates (x1 , ..., xn ), such that E |U is trivial. This
is always happening, at least on every small portion of γ. Denote by (eα )1≤α≤r a local
moving frame trivializing E |U so that u = uα eα . The connection 1-form corresponding to
this moving frame will be denoted by Γ ∈ Ω1 (End (Kr )). Equation (3.3.7) becomes (using
Einstein’s convention)
 duα
dt + Γαtβ uβ = 0
, (3.3.8)
uα (0) = uα0
104 CHAPTER 3. CALCULUS ON MANIFOLDS

where
d

Γt = Γ = γ̇ Γ ∈ Ω0 End(Kr ) = End(Kr ).
dt
More explicitly, if the path γ(t) is given by the smooth map


t 7→ γ(t) = x1 (t), . . . , xn (t) ,

then Γt is the endomorphism given by

Γt eβ = ẋi Γαiβ eα .

The system (3.3.8) can be rewritten as

 duα
 dt + Γαiβ ẋi uβ = 0
, (3.3.9)

uα (0) = uα0

This is obviously a system of linear ordinary differential equations whose solutions

exist for any t. We deduce
u̇(0) = −Γt u0 . (3.3.10)
This gives a geometric interpretation for the connection 1-form Γ: for any vector field X,
the contraction

−iX Γ = −Γ(X) ∈ C ∞ End(E)
describes the infinitesimal parallel transport along the direction prescribed by the vector
field X, in the non-canonical identification of nearby fibers via a local moving frame.
In more intuitive terms, if γ(t) is an integral curve for X, and Tt denotes the parallel
transport along γ from Eγ(0) to Eγ(t) , then, given a local moving frame for E in a neigh-
borhood of γ(0), Tt is identified with a t-dependent matrix which has a Taylor expansion
of the form
Tt = 1 − Γ0 t + O(t2 ), t very small, (3.3.11)
with Γ0 = (iX Γ) |γ(0) .
If the connection ∇ is also compatible with a metric h, then (3.3.2) shows the parallel
transport
Tt : Eγ(0) → Eγ(t) .
is an isometry.

3.3.3 The curvature of a connection

Consider a rank k smooth K-vector bundle E → M over the smooth manifold M , and let
∇ : Ω0 (E) → Ω1 (E) be a covariant derivative on E.
Proposition 3.3.8. The connection ∇ has a natural extension to an operator

d∇ : Ωr (E) → Ωr+1 (E)

uniquely defined by the requirements,

3.3. CONNECTIONS ON VECTOR BUNDLES 105

1.
d∇ |Ω0 (E) = ∇,

2. ∀ω ∈ Ωr (M ), η ∈ Ωs (E)

d∇ (ω ∧ η) = dω ∧ η + (−1)r ω ∧ d∇ η.

Outline of the proof Existence. For ω ∈ Ωr (M ), u ∈ Ω0 (E) set

d∇ (ω ⊗ u) = dω ⊗ u + (−1)r ω∇u. (3.3.12)

Using a partition of unity one shows that any η ∈ Ωr (E) is a locally finite combination of
monomials as above so the above definition induces an operator Ωr (E) → Ωr+1 (E). We
let the reader check that this extension satisfies conditions (a) and (b) above.
Uniqueness. Any operator with the properties (a) and (b) acts on monomials as in (3.3.12)
so it has to coincide with the operator described above using a given partition of unity. ⊔
⊓

Example 3.3.9. The trivial bundle KM has a natural connection ∇0 - the trivial con-
nection. This coincides with the usual differential d : Ω0 (M ) ⊗ K → Ω1 (M ) ⊗ K. The
0
extension d∇ is the usual exterior derivative. ⊔
⊓

There is a major difference between the usual exterior derivative d, and an arbitrary
d∇ . In the former case we have d2 = 0, which is a consequence of the commutativity
[∂xi , ∂xj ] = 0, where (xi ) are local coordinates on M . In the second case, the equality
(d∇ )2 = 0 does not hold in general. Still, something very interesting happens.

Lemma 3.3.10. For any smooth function f ∈ C ∞ (M ), and any ω ∈ Ωr (E) we have

(d∇ )2 (f ω) = f {(d∇ )2 ω}.

Hence (d∇ )2 is a bundle morphism Λr T ∗ M ⊗ E → Λr+2 T ∗ M ⊗ E.

Proof. We compute
(d∇ )2 (f ω) = d∇ (df ∧ ω + f d∇ ω)

= −df ∧ d∇ ω + df ∧ d∇ ω + f (d∇ )2 ω = f (d∇ )2 ω. ⊔

⊓

As a map Ω0 (E) → Ω2 (E), the operator (d∇ )2 can be identified with a section of

HomK (E, Λ2 T ∗ M ⊗R E) ∼
= E ∗ ⊗ Λ2 T ∗ M ⊗R E ∼
= Λ2 T ∗ M ⊗R EndK (E).

Thus, (d∇ )2 is an EndK (E)-valued 2-form.

Definition 3.3.11. For any connection ∇ on a smooth vector bundle E → M , the object
(d∇ )2 ∈ Ω2 (EndK (E)) is called the curvature of ∇, and it is usually denoted by F (∇). ⊔
⊓
106 CHAPTER 3. CALCULUS ON MANIFOLDS

Example 3.3.12. Consider the trivial bundle KrM . The sections of this bundle are smooth
Kr -valued functions on M . The exterior derivative d defines the trivial connection on KrM ,
and any other connection differs from d by a Mr (K)-valued 1-form on M . If A is such a
form, then the curvature of the connection d + A is the 2-form F (A) defined by

F (A)u = (d + A)2 u = (dA + A ∧ A)u, ∀u ∈ C ∞ (M, Kr ).

The ∧-operation above is defined for any vector bundle E as the bilinear map

Ωj (End (E)) × Ωk (End (E)) → Ωj+k (End (E)),

uniquely determined by


(ω r ⊗ A) ∧ (η s ⊗ B) = ω j ∧ η k ⊗ AB, A, B ∈ C ∞ End(E) . ⊔
⊓

We conclude this subsection with an alternate description of the curvature which hope-
fully will shed some light on its analytical significance.
Let E → M be a smooth vector bundle on M and ∇ a connection on it. Denote its
curvature by F = F (∇) ∈ Ω2 (End (E)). For any X, Y ∈ Vect (M ) the quantity F (X, Y )
is an endomorphism of E. In the remaining part of this section we will give a different
description of this endomorphism.
For any vector field Z, we denote by iZ : Ωr (E) → Ωr−1 (E) the C ∞ (M ) − linear
operator defined by

iZ (ω ⊗ u) = (iZ ω) ⊗ u, ∀ω ∈ Ωr (M ), u ∈ Ω0 (E).

For any vector field Z, the covariant derivative ∇Z : C ∞ (E) → C ∞ (E) extends naturally
as a linear operator Ωr (E) → Ωr (E), which we continue to denote by ∇Z , uniquely defined
by the requirements
∇Z (ω ⊗ u) := (LZ ω) ⊗ u + ω ⊗ ∇Z u.
The operators d∇ , iZ , ∇Z : Ω• (E) → Ω• (E) satisfy the usual super-commutation identi-
ties.
iZ d∇ + d∇ iZ = ∇Z . (3.3.13)
iX iY + iY iX = 0. (3.3.14)
∇X iY − iY ∇X = i[X,Y ] . (3.3.15)
For any u ∈ Ω0 (E) we compute using (3.3.13)-(3.3.15)

F (X, Y )u = iY iX (d∇ )2 u = iY (iX d∇ )∇u

= iY (∇X − d∇ iX )∇u = (iY ∇X )∇u − (iY d∇ )∇X u

= (∇X iY − i[X,Y ] )∇u − ∇Y ∇X u = (∇X ∇Y − ∇Y ∇X − ∇[X,Y ] )u.
Hence
F (X, Y ) = [∇X , ∇Y ] − ∇[X,Y ] . (3.3.16)
3.3. CONNECTIONS ON VECTOR BUNDLES 107

If in the above formula we take X = ∂xi and Y = ∂xj , where (xi ) are local coordinates on
M , and we set ∇i := ∇∂xi , ∇j = ∇∂xj , then we deduce


Fij = −Fji := F ∂xi , ∂xj = [∇i , ∇j ]. (3.3.17)

Thus, the endomorphism Fij measures the extent to which the partial derivatives ∇i , ∇j
fail to commute. This is in sharp contrast with the classical calculus and an analytically
oriented reader may object to this by saying we were careless when we picked the con-
nection. Maybe an intelligent choice will restore the classical commutativity of partial
derivatives so we should concentrate from the very beginning to covariant derivatives ∇
such that F (∇) = 0.
Definition 3.3.13. A connection ∇ such that F (∇) = 0 is called flat. ⊔
⊓

Proposition 3.3.14. If ∇ is a connection on
a real vector bundle that is compatible

with a metric h, then F (∇) ∈ Ω2 End− (E) where End− (E) denotes the bundle of
endomorphisms of E that are skew-symmetric with respect to h.

Proof. Let X, Y ∈ Vect (M ) and u, v ∈ C ∞ (M ). Using (3.3.2) we deduce

XY h(u, v) = Xh(∇Y u, v) + Xh(u, ∇Y v)

= h(∇X ∇Y u, v) + h(∇Y u, ∇X v) + h(∇X u, ∇Y v) + h(u, ∇X ∇Y v).

Hence
[X, Y ]h(u, v) = XY h(u, v) − Y Xh(uv)
(3.3.18)
= h([∇X , ∇Y ]u, v) + h(u, [∇X , ∇Y ]v).

On the other hand,

[X, Y ]h(u, v) = h(∇[X,Y ] u, v) + h(u, ∇[X,Y ] v). (3.3.19)

The desired conclusion is reached after subtracting (3.3.19) from (3.3.18). ⊔

⊓

A natural question arises: given an arbitrary vector bundle E → M do there exist

flat connections on E? If E is trivial then the answer is obviously positive. In general,
the answer is negative, and this has to do with the global structure of the bundle. In the
second half of this book we will discuss in more detail this fact.

3.3.4 Holonomy
The reader may ask a very legitimate question: why have we chosen to name curvature,
the deviation from commutativity of a given connection. In this subsection we describe the
geometric meaning of curvature, and maybe this will explain this choice of terminology.
Throughout this subsection we will use Einstein’s convention.
Let E → M be a rank r smooth K-vector bundle, and ∇ a connection on it. Consider
local coordinates (x1 , ..., xn ) on an open subset U ⊂ M such that E |U is trivial. Pick a
108 CHAPTER 3. CALCULUS ON MANIFOLDS

moving frame (e1 , ..., er ) of E over U . The connection 1-form associated to this moving
frame is
Γ = Γi dxi = (Γαiβ )dxi , 1 ≤ α, β ≤ r.
It is defined by the equalities (∇i := ∇∂xi )
∇i eβ = Γαiβ eα . (3.3.20)
Using (3.3.17) we compute
Fij eβ = (∇i ∇j − ∇j ∇i )eβ = ∇i (Γj eβ ) − ∇j (Γi eβ )
 α   
∂Γjβ ∂Γαjβ γ α γ α
= − eα + Γ Γ
jβ iγ − Γ iβ jγ eα ,
Γ
∂xi ∂xj
so that  
∂Γj ∂Γi
Fij = − + Γ Γ
i j − Γ Γ
j i . (3.3.21)
∂xi ∂xj
Though the above equation looks very complicated it will be the clue to understanding
the geometric significance of curvature.

p p
3 2

p p
0 1

Figure 3.1: Parallel transport along a coordinate parallelogram.

Assume for simplicity that the point of coordinates (0, ..., 0) lies in U . Denote by
T1s the parallel transport (using the connection ∇) from (x1 , ..., xn ) to (x1 + s, x2 , ..., xn )
along the curve τ 7→ (x1 + τ, x2 , ..., xn ). Define T2t in a similar way using the coordinate
x2 instead of x1 .
Look at the parallelogram Ps,t in the “plane” (x1 , x2 ) described in Figure 3.1, where
p0 = (0, . . . , 0), p1 = (s, 0, . . . , 0), p2 = (s, t, 0, . . . , 0), p3 = (0, t, 0, . . . , 0).
We identify E|U with the trivial bundle E0 × U where E0 is the fiber of E over p0 . We now
perform the counterclockwise parallel transport along the boundary of Ps,t . The outcome
is a linear map Ts,t : E0 → E0 . Set F12 := F ( ∂x∂ 1 ∂x∂ 2 ) |(0,...,0) . F12 is also an endomorphism
of E0 .
Proposition 3.3.15. We have
∂2
F12 = − Ts,t .
∂s∂t
3.3. CONNECTIONS ON VECTOR BUNDLES 109

We see that the parallel transport of an element u ∈ E0 along a closed path may not
return it to itself. The curvature is an infinitesimal measure of this deviation.

Proof. The parallel transport along ∂Ps,t can be described as

Ts,t = T2−t T1−s T2t T1s .

The parallel transport T1s : E0 → Ep1 can be approximated using (3.3.10)

T1s = 1 − sΓ1 (p0 ) + C1 s2 + O(s3 ). (3.3.22)

The term C1 is an endomorphism of E0 whose exact form is not relevant to our computa-
tions. In the sequel the letter C (eventually indexed) will denote constants.


T2t T1s = 1 − tΓ2 (p1 ) + C2 t2 + O(t3 ) T1s



= 1 − tΓ2 (p1 ) + C2 t2 + O(t3 ) 1 − sΓ1 (p0 ) + C1 s2 + O(s3 )


= 1 − sΓ1 (p0 ) − tΓ2 (p1 ) + tsΓ2 (p1 )Γ1 (p0 ) + C1 s2 + C2 t2 + O(3).
Above and in the sequel O(k) denotes an error ≤ C(s2 + t2 )k/2 as s, t → 0. Now use the
approximation
∂Γ2
Γ2 (p1 ) = Γ2 (p0 ) + s 1 (p0 ) + O(2),
∂x
to deduce
( )
 ∂Γ 
T2t T1s = 1 − sΓ1 − tΓ2 − st
2
− Γ2 Γ1 |p0
∂x1 (3.3.23)
+C1 s2 + C2 t2 + O(3).

Similarly, we have


T1−s T2t T1s = 1 + sΓ1 (p2 ) + C3 s2 + O(3) T2t T1s .
The Γ-term in the right-hand-side of the above equality can be approximated as
∂Γ1 ∂Γ1
Γ1 (p2 ) = Γ1 (p0 ) + s 1
(p0 ) + t 2 (p0 ) + O(2).
∂x ∂x
Using (3.3.23) we get after an elementary computation
  
∂Γ1 ∂Γ2
T1 T2 T1 = 1 − tΓ2 + st
−s t s
− + Γ2 Γ1 − Γ1 Γ2 |p0
∂x2 ∂x1 (3.3.24)
2 2
+C4 s + C5 t + O(3).

Finally, we have


T2−t T1−s T2t T1s = 1 + tΓ2 (p3 ) + C6 t2 + O(3) T1−s T2t T1s ,
with
∂Γ2
Γ2 (p3 ) = Γ2 (p0 ) + t (p0 ) + C7 t2 + O(3).
∂x2
110 CHAPTER 3. CALCULUS ON MANIFOLDS

Using (3.3.24) we get

 
∂Γ1 ∂Γ2
T2−t T1−s T2t T1s = 1 + st − + Γ2 Γ1 − Γ1 Γ2 |p0
∂x2 ∂x1
+C8 s2 + C9 t2 + O(3)

= 1 − stF12 (p0 ) + C8 s2 + C9 t2 + O(3).

∂ 2 u4
Clearly ∂s∂t = −F12 (p0 ) as claimed. ⊔
⊓

Remark 3.3.16. If we had kept track of the various constants in the above computation
we would have arrived at the conclusion that C8 = C9 = 0, i.e.,

Ts,t = 1 − stF12 + O(3). (3.3.25)

Alternatively, the constant C8 is the second order correction in the Taylor expansion of
s 7→ Ts,0 = 1, ∀s, so C8 = 0. The same goes for C9 . Thus we have

dTs,t dT√s,√s
−F12 = = .
dareaPs,t ds

Loosely speaking, the last equality states that the curvature is the the “amount of holon-
omy per unit of area”. ⊔
⊓

The result in the above proposition is usually formulated in terms of holonomy.

Definition 3.3.17. Let E → M be a vector bundle with a connection ∇. The holonomy
of ∇ along a closed path γ is the parallel transport along γ. ⊔
⊓

We see that the curvature measures the holonomy along infinitesimal parallelograms.
A connection can be viewed as an analytic way of trivializing a bundle. We can do so
along paths starting at a fixed point, using the parallel transport, but using different paths
ending at the same point we may wind up with trivializations which differ by a twist. The
curvature provides an infinitesimal measure of that twist.
Exercise 3.3.18. Prove that any vector bundle E over the Euclidean space Rn is trivial-
izable.
Hint: Use the parallel transport defined by a connection on the vector bundle E to
produce a bundle isomorphism E → E0 × Rn , where E0 is the fiber of E over the origin.
⊔
⊓

3.3.5 The Bianchi identities

Consider a smooth K-vector bundle E → M equipped with a connection ∇ = ∇E . We
have seen that the associated exterior derivative d∇ : Ωp (E) → Ωp+1 (E) does not satisfy
the usual (d∇ )2 = 0, and the curvature is to blame for this. The Bianchi identity describes
one remarkable algebraic feature of the curvature.
3.3. CONNECTIONS ON VECTOR BUNDLES 111

Recall that ∇E induces a connection in any tensor bundle constructed from E. In

particular, it induces a connection in E ∗ ⊗ E ∼
= End (E) which we denote by ∇End(E) .
This extends to an “exterior derivative”
End(E)
D E = d∇ : Ωp (End (E)) → Ωp+1 (End (E)).
Proposition 3.3.19 (Bianchi identity). Let E → M be a K-vector bundle on M and ∇E
a connection on E. Then
D E F (∇E ) = 0.
Roughly speaking, the Bianchi identity states that (d∇ )3 is 0.

Proof. We will use the identities (3.3.13)–(3.3.15). For any vector fields X, Y , Z we have
End(E)
iX D E = ∇X − D E iX .
Hence,
End(E)
(D E F )(X, Y, Z) = iZ iY iX D E F = iZ iY (∇X − D E iX )F
End(E) End(E)
= iZ (∇X iY − i[X,Y ] )F − iZ (∇Y − D E iY )iX F
End(E)

= ∇X iZ iY − i[X,Z] iY − iZ i[X,Y ] F
End(E) End(E)

− ∇Y iZ iX − i[Y,Z]iX − ∇Z iY iX F


= i[X,Y ] iZ + i[Y,Z] iX + i[Z,X] iY F
End(E) End(E) End(E)

− ∇X iY iZ + ∇Y iZ iX + ∇Z iX iY F
= A(X, Y, Z) + A(Y, Z, X) + A(Z, X, Y ),
where

End(E)
A(X, Y, Z) := i[X,Y ] iZ − ∇X iY iZ F, ∀X, Y, Z ∈ Vect (M ).
We compute immediately
h i
i[X,Y ] iZ F = F (Z, [X, Y ]) = ∇E E E
Z , ∇[X,Y ] − ∇[Z,[X,Y ]] .

Also for any u ∈ Ω0 (E) we have

End(E)  E 
(∇X iY iZ F )u = ∇E E
X (F (Z, Y )u) − F (Z, Y )∇X u = ∇X , F (Z, Y ) u
h i  E  E E 
= ∇E X , ∇ E
[Y,Z] u − ∇X , ∇Y , ∇Z u.
Hence   E E 
A(X, Y, Z) = ∇E
X , ∇Y , ∇Z .
The Bianchi identity now follows from the classical Jacobi identity for commutators. ⊔
⊓

Example 3.3.20. Let K be the trivial line bundle over a smooth manifold M . Any
connection on K has the form ∇ω = d + ω, where d is the trivial connection, and ω is a
K-valued 1-form on M . The curvature of this connection is
F (ω) = dω.
The Bianchi identity is in this case precisely the equality d2 ω = 0. ⊔
⊓
112 CHAPTER 3. CALCULUS ON MANIFOLDS

3.3.6 Connections on tangent bundles

The tangent bundles are very special cases of vector bundles so the general theory of
connections and parallel transport is applicable in this situation as well. However, the
tangent bundles have some peculiar features which enrich the structure of a connection.
Recall that, when looking for a local description for a connection on a vector bundle,
we have to first choose local coordinates on the manifolds, and then a local moving frame
for the vector bundle. For an arbitrary vector bundle there is no correlation between these
two choices.
For tangent bundles it happens that, once local coordinates (xi ) are chosen, they
automatically define a moving frame of the tangent bundle, (∂i = ∂xi ), and it is thus very
natural to work with this frame. Hence, let ∇ be a connection on T M . With the above
notations we set
∇i ∂j = Γkij ∂k (∇i = ∇∂i ).

The coefficients Γkij are usually known as the Christoffel symbols of the connection. As
usual we construct the curvature tensor


F (X, Y ) = [∇X , ∇Y ] − ∇[X,Y ] ∈ C∞ End(T M ) .

Still, this is not the only tensor naturally associated to ∇.

Lemma 3.3.21. For X, Y ∈ Vect (M ) consider

T (X, Y ) = ∇X Y − ∇Y X − [X, Y ] ∈ Vect (M ).

Then ∀f ∈ C ∞ (M )
T (f X, Y ) = T (X, f Y ) = f T (X, Y ),
so that T (•, •) is a tensor T ∈ Ω2 (T M ), i.e., a 2-form whose coefficients are vector fields
on M . The tensor T is called the torsion of the connection ∇. ⊔
⊓

The proof of this lemma is left to the reader as an exercise. In terms of Christoffel
symbols, the torsion has the description

T (∂i , ∂j ) = (Γkij − Γkji )∂k .

Definition 3.3.22. A connection on T M is said to be symmetric if T = 0. ⊔

⊓

We guess by now the reader is wondering how the mathematicians came up with
this object called torsion. In the remaining of this subsection we will try to sketch the
geometrical meaning of torsion.
To find such an interpretation, we have to look at the finer structure of the tangent
space at a point x ∈ M . It will be convenient to regard Tx M as an affine space modeled
by Rn , n = dim M . Thus, we will no longer think of the elements of Tx M as vectors, but
instead we will treat them as points. The tangent space Tx M can be coordinatized using
affine frames. These are pairs (p; e), where p is a point in Tx M , and e is a basis of the
3.3. CONNECTIONS ON VECTOR BUNDLES 113

underlying vector space. A frame allows one to identify Tx M with Rn , where p is thought
of as the origin.
Suppose that A, B are two affine spaces, both modelled by Rn , and (p; e) , (p; f ) are
affine frames of A and respectively B. Denote by (xi ) the coordinates in A induced by
the frame (p; e), and by (y j ) the coordinates in B induced by the frame (q; f ). An affine
isomorphism T : A → B can then be described using these coordinates as
T : Rnx → Rny x 7→ y = Sx + v,

where v is a vector in Rn , and S is an invertible n × n real matrix. Thus, an affine map is

described by a “rotation” S, followed by a translation v. This vector measures the “drift”
of the origin. We write T = S +̂x
If now (xi ) are local coordinates on M , then they define an affine frame Ax at each
x ∈ M : (Ax = (0; (∂i )). Given a connection ∇ on T M , and a smooth path γ : I → M , we
will construct a family of affine isomorphisms Tt : Tγ(0) → Tγ(t) called the affine transport
of ∇ along γ. In fact, we will determine Tt by imposing the initial condition T0 = 1, and
then describing Ṫt .
This is equivalent to describing the infinitesimal affine transport at a given point
x0 ∈ M along a direction given by a vector X = X i ∂i ∈ Tx0 M . The affine frame of Tx0 M
is Ax0 = (0; (∂i )).
If xt is a point along the integral curve of X, close to x0 then its coordinates satisfy
xit = xi0 + tX i + O(t2 ).

This shows the origin x0 of Ax0 “drifts” by tX + O(t2 ). The frame (∂i ) suffers a parallel
transport measured as usual by 1 − tiX Γ + O(t2 ). The total affine transport will be

Tt = (1 − tiX Γ)+̂tX + O(t2 ).

The holonomy of ∇ along a closed path will be an affine transformation and as such it has
two components: a “rotation”” and a translation. As in Proposition 3.3.15 one can show
the torsion measures the translation component of the holonomy along an infinitesimal
parallelogram, i.e., the “amount of drift per unit of area”. Since we will not need this fact
we will not include a proof of it.
Exercise 3.3.23. Consider the vector valued 1-form ω ∈ Ω1 (T M ) defined by
ω(X) = X ∀X ∈ Vect (M ).

Show that if ∇ is a linear connection on T M , then d∇ ω = T ∇ , where T ∇ denotes the

torsion of ∇. ⊔
⊓

Exercise 3.3.24. Consider a smooth vector bundle E → M over the smooth manifold
M . We assume that both E and T M are equipped with connections and moreover the
connection on T M is torsionless. Denote by ∇ ˆ the induced connection on Λ2 T ∗ M ⊗
End (E). Prove that ∀X, Y, Z ∈ Vect (M )




ˆ X F (Y, Z) + ∇
∇ ˆ Y F (Z, X) + ∇ˆ Z F (X, Y ) = 0. ⊔
⊓
114 CHAPTER 3. CALCULUS ON MANIFOLDS

3.4 Integration on manifolds

3.4.1 Integration of 1-densities
We spent a lot of time learning to differentiate geometrical objects but, just as in classical
calculus, the story is only half complete without the reverse operation, integration.
Classically, integration requires a background measure, and in this subsection we will
describe the differential geometric analogue of a measure, namely the notion of 1-density
on a manifold.
Let E → M be a rank k, smooth real vector bundle over a manifold M defined by
an open cover (Uα ) and transition maps gβα : Uαβ → GL(k, R) satisfying the cocycle
condition. For any r ∈ R we can form the real line bundle |Λ|r (E) defined by the same
open cover and transition maps

tβα := | det gβα |−r = | det gαβ |r : Uαβ → R>0 ֒→ GL(1, R).

The fiber at p ∈ M of this bundle consists of r-densities on Ep (see Subsection 2.2.4).

Definition 3.4.1. Let M be a smooth manifold and r ≥ 0. The bundle of r-densities on
M is
|Λ|rM : =|Λ|r (T M ).
When r = 1 we will use the notation |Λ|M = |Λ|1M . We call |Λ|M the density bundle of
M. ⊔
⊓

Denote by C ∞ (|Λ|M ) the space of smooth sections of |Λ|M , and by C0∞ (|Λ|M ) its
subspace consisting of compactly supported densities.
It helps to have local descriptions of densities. To this aim, pick an open cover of M
consisting of coordinate neighborhoods (Uα ). Denote the local coordinates on Uα by (xiα ).
This choice of a cover produces a trivializing cover of T M with transition maps
!
∂xiα
Tαβ = ,
∂xjβ
1≤i,j≤n

where n is the dimension of M . Set δαβ = | det Tαβ |. A 1-density on M is then a collection
of functions µα ∈ C ∞ (Uα ) related by
−1
µα = δαβ µβ .

It may help to think that for each point p ∈ Uα the basis ∂x∂ 1 , ..., ∂x∂n of Tp M spans an
α α
infinitesimal parallelepiped and µα (p) is its “volume”. A change in coordinates should
be thought of as a change in the measuring units. The gluing rules describe how the
numerical value of the volume changes from one choice of units to another.
The densities on a manifold resemble in many respects the differential forms of maximal
degree. Denote by det T M = Λdim M T M the determinant line bundle of T M . A density
is a map
µ : C ∞ (det T M ) → C ∞ (M ),
3.4. INTEGRATION ON MANIFOLDS 115

such that µ(f e) = |f |µ(e), for all smooth functions f : M → R, and all e ∈ C ∞ (det T M ).
In particular, any smooth map φ : M → N between manifolds of the same dimension
induces a pullback transformation

φ∗ : C ∞ (|Λ|N ) → C ∞ (|Λ|M ),

described by


(φ∗ µ)(e) = µ (det φ∗ ) · e = | det φ∗ | µ(e), ∀e ∈ C ∞ (det T M ).

Example 3.4.2. (a) Consider the special case M = Rn . Denote by e1 , ..., en the canonical
basis. This extends to a trivialization of T Rn and, in particular, the bundle of densities
comes with a natural trivialization. It has a nowhere vanishing section |dvn | defined by

|dvn |(e1 ∧ ... ∧ en ) = 1.

In this case, any smooth density on Rn takes the form µ = f |dvn |, where f is some smooth
function on Rn . The reader should think of |dvn | as the standard Lebesgue measure on
Rn .
If φ : Rn → Rn is a smooth map, viewed as a collection of n smooth functions

φ1 = φ1 (x1 , ..., xn ), . . . , φn = φn (x1 , ..., xn ),

then, !
∂φi
φ∗ (|dvn |) = det · |dvn |.
∂xj
(b) Suppose M is a smooth manifold of dimension m. Then any top degree form ω ∈
Ωm (M ) defines a density |ω| on M which associates to each section e ∈ C ∞ (det T M ) the
smooth function
x 7→ |ωx ( e(x) )|
Observe that |ω| = | − ω|, so this map Ωm (M ) → C ∞ (|Λ|M ) is not linear.
(c) Suppose g is a Riemann metric on the smooth manifold M . The volume density
defined by g is the density denoted by |dVg | which associates to each e ∈ C ∞ (det T M )
the pointwise length
x 7→ |e(x)|g .
If (Uα , (xiα ) ) is an atlas of M , then on each Uα we have top degree forms

dxα := dx1α ∧ · · · ∧ dxm

α,

to which we associate the density |dxα |. In the coordinates (xiα ) the metric g can be
described as X
g= gα;ij dxiα ⊗ dxjα .
i,j

We denote by |gα | the determinant of the symmetric matrix gα = (gα;ij )1≤i,j≤m . Then the
restriction of |dVg | to Uα has the description
p
|dVg | = |gα | |dxα |. ⊔
⊓
116 CHAPTER 3. CALCULUS ON MANIFOLDS

The importance of densities comes from the fact that they are exactly the objects that
can be integrated. More precisely, we have the following abstract result.
Proposition 3.4.3. There exists a natural way to associate to each smooth manifold M
a linear map Z
: C0∞ (|Λ|M ) → R
M
uniquely
R defined by the following conditions.
(a) M is invariant under diffeomorphisms, i.e., for any smooth manifolds M , N of the
same dimension n, any diffeomorphism φ : M → N , and any µ ∈ C0∞ (|Λ|M ), we have
Z Z
∗
φ µ= µ.
M N
R
(b) M is a local operation, i.e., for any open set U ⊂ M , and any µ ∈ C0∞ (|Λ|M ) with
supp µ ⊂ U , we have Z Z
µ= µ.
M U
(c) For any ρ ∈ C0∞ (Rn ) we have
Z Z
ρ|dvn | = ρ(x)dx,
Rn Rn

where inR the right-hand-side stands the Lebesgue integral of the compactly supported func-
tion ρ. M is called the integral on M .

Proof. To establish the existence of an integral we associate to each manifold M a collec-

tion of data as follows.

(i) A smooth partition of unity A ⊂ C0∞ (M ) such that ∀α ∈ A the support supp α lies
entirely in some precompact coordinate neighborhood Uα , and such that the cover
(Uα ) is locally finite.

(ii) For each Uα we pick a collection of local coordinates (xiα ), and we denote by |dxα |
(n = dim M ) the density on Uα defined by
 
∂ ∂
|dxα | ∧ ... ∧ n = 1.
∂x1α ∂xα

For any µ ∈ C ∞ (|Λ|), the product αµ is a density supported in Uα , and can be written
as
αµ = µα |dxα |,
where µα is some smooth function compactly supported on Uα . The local coordinates
allow us to interpret µα as a function on Rn . Under this identification |dxiα | corresponds
to the Lebesgue measure |dvn | on Rn , and µα is a compactly supported, smooth function.
We set Z Z
αµ := µα |dxα |.
Uα Rn
3.4. INTEGRATION ON MANIFOLDS 117

Finally, define
Z A Z XZ
def
µ= µ = αµ.
M M α∈A Uα

The above sum contains only finitely many nonzero terms since supp µ is compact, and
thus it intersects only finitely many of the Uα′ s which form a locally finite cover.
To prove property (a) we will first prove that the integral defined as above is indepen-
dent of the various choices: the partition of unity A ⊂ C0∞ (M ), and the local coordinates
(xiα )α∈A .
• Independence of coordinates. Fix the partition of unity A, and consider a new
collection of local coordinates (yαi ) on each Uα . These determine two densities |dxiα | and
respectively |dyαj |. For each µ ∈ C0∞ (|Λ|M ) we have

αµ = αµxα |dxα | = αµyα |dyα |,

where µxα , µyα ∈ C0∞ (Uα ) are related by

 
∂xiα
µyα = det µxα .
∂yαj
The equality Z Z
µxα |dxα | = µyα |dyα |
Rn Rn
is the classical change in variables formula for the Lebesgue integral.
• Independence of the partition of unity. Let A, B ⊂ C0∞ (M ) two partitions of unity
on M . We will show that Z A Z B
= .
M M
Form the partition of unity

A ∗ B := αβ ; (α, β) ∈ A × B ⊂ C0∞ (M ).

Note that supp αβ ⊂ Uαβ = Uα ∩ Uβ . We will prove

Z A Z A∗B Z B
= = .
M M M

Let µ ∈ C0∞ (|Λ|M ). We can view αµ as a compactly supported function on Rn . We have

Z XZ XZ
αµ = βαµ = αβµ. (3.4.1)
Uα β Uα ⊂Rn β Uαβ

Similarly Z XZ
βµ = αβµ. (3.4.2)
Uβ α Uαβ

Summing (3.4.1) over α and (3.4.2) over β we get the desired conclusion.
118 CHAPTER 3. CALCULUS ON MANIFOLDS

To prove property (a) for a diffeomorphism φ : M → N , consider a partition of unity

A ⊂ C0∞ (N ). From the classical change in variables formula we deduce that, for any
coordinate neighborhood Uα containing the support of α ∈ A, and any µ ∈ C0∞ (|Λ|N ) we
have Z Z
(φ∗ α)φ∗ µ = αµ.
φ−1 (Uα ) Uα
 
The collection φ∗ α = α ◦ φ forms a partition of unity on M . Property (a) now
α∈A
follows by summing over α the above equality, and using the independence of the integral
on partitions of unity.
To prove property (b) on the local character of the integral, pick U ⊂ M , and then
choose a partition of unity B ⊂ C0∞ (U ) subordinated to the open cover (Vβ )β∈B . For any
partition of unity A ⊂ C0∞ (M ) with associated cover (Vα )α∈A we can form a new partition
of unity A ∗ B of U with associated cover Vαβ = Vα ∩ Vβ . We use this partition of unity
to compute integrals over U . For any density µ on M supported on U we have
Z XZ X XZ X Z Z
µ= αµ = αβµ = αβµ = µ.
M α Vα α∈A β∈B Vαβ αβ∈A∗B Vαβ U

Property (c) is clear since, for M = Rn , we can assume that all the local coordinates chosen
are Cartesian. The uniqueness of the integral is immediate, and we leave the reader to fill
in the details. ⊔
⊓

3.4.2 Orientability and integration of differential forms

Under some mild restrictions on the manifold, the calculus with densities can be replaced
with the richer calculus with differential forms. The mild restrictions referred to above
have a global nature. More precisely, we have to require that the background manifold is
oriented.
Roughly speaking, the oriented manifolds are the “2-sided manifolds”, i.e., one can
distinguish between an “inside face” and an “outside face” of the manifold. (Think of a
2-sphere in R3 (a soccer ball) which is naturally a “2-faced” surface.)
The 2-sidedness feature is such a frequent occurrence in the real world that for many
years it was taken for granted. This explains the “big surprise” produced by the famous
counter-example due to Möbius in the first half of the 19th century. He produced a 1-sided
surface nowadays known as the Möbius band using paper and glue. More precisely, he
glued the opposite sides of a paper rectangle attaching arrow to arrow as in Figure 3.2.
The 2-sidedness can be formulated rigorously as follows.
Definition 3.4.4. A smooth manifold M is said to be orientable if the determinant line
bundle det T M (or equivalently det T ∗ M ) is trivializable. ⊔
⊓

We see that det T ∗ M is trivializable if and only if it admits a nowhere vanishing section.
Such a section is called a volume form on M . We say that two volume forms ω1 and ω2
are equivalent if there exists f ∈ C ∞ (M ) such that

ω2 = ef ω1 .
3.4. INTEGRATION ON MANIFOLDS 119

Figure 3.2: The Mobius band.

This is indeed an equivalence relation, and an equivalence class of volume forms will be
called an orientation of the manifold. We denote by Or(M ) the set of orientations on
the smooth manifold M . A pair (orientable manifold, orientation) is called an oriented
manifold.
Let us observe that if M is orientable, and thus Or(M ) 6= ∅, then for every point
p ∈ M we have a natural map

Or(M ) → Or(Tp M ), Or(M ) ∋ or 7→ or p ∈ Or(Tp M ),

defined as follows. If the orientation or on M is defined by a volume form ω, then

ωp ∈ det Tp∗ M is a nontrivial volume form on Tp M , which canonically defines an orientation
or ω,p on Tp M . It is clear that if ω1 and ω2 are equivalent volume form then or ω1 ,p = or ω2 ,p .
This map is clearly a surjection because or −ω,p = −or ω,p , for any volume form ω.
Proposition 3.4.5. If M is a connected, orientable smooth manifold M , then for every
p ∈ M the map
Or(M ) ∋ or 7→ or p ∈ Or(Tp M )
is a bijection.

Proof. Suppose or and or ′ are two orientations on M such that or p = or ′p . The function

M ∋ q 7→ ǫ(q) = or ′q /or q ∈ {±1}

is continuous, and thus constant. In particular, ǫ(q) = ǫ(p) = 1, ∀q ∈ M .

If or is given by the volume form ω and or ′ is given by the volume form ω ′ , then there
exists a nowhere vanishing smooth function ρ : M → R such that ω ′ = ρω. We deduce

sign ρ(q) = ǫ(q), ∀q ∈ M.

This shows that the two forms ω ′ and ω are equivalent and thus or = or ′ . ⊔
⊓
120 CHAPTER 3. CALCULUS ON MANIFOLDS

Figure 3.3: The normal line bundle to the round sphere.

The last proposition shows that on a connected, orientable manifold, a choice of an

orientation of one of its tangent spaces uniquely determines an orientation of the manifold.
A natural question arises.
How can one decide whether a given manifold is orientable or not.
We see this is just a special instance of the more general question we addressed in Chap-
ter 2: how can one decide whether a given vector bundle is trivial or not. The orientability
question can be given a very satisfactory answer using topological techniques. However,
it is often convenient to decide the orientability issue using ad-hoc arguments. In the
remaining part of this section we will describe several simple ways to detect orientability.
Example 3.4.6. If the tangent bundle of a manifold M is trivial, then clearly T M is
orientable. In particular, all Lie groups are orientable. ⊔
⊓

Example 3.4.7. Suppose M is a manifold such that the Whitney sum RkM ⊕ T M is
trivial. Then M is orientable. Indeed, we have

det(Rk ⊕ T M ) = det Rk ⊗ det T M.

Both det Rk and det(Rk ⊕ T M ) are trivial. We deduce det T M is trivial since

det T M ∼
= det(Rk ⊕ T M ) ⊗ (det Rk )∗ .

This trick works for example when M ∼ = S n . Indeed, let ν denote the normal line bundle.
The fiber of ν at a point p ∈ S n is the 1-dimensional space spanned by the position vector
of p as a point in Rn ; (see Figure 3.3). This is clearly a trivial line bundle since it has a
tautological nowhere vanishing section p 7→ p ∈ νp . The line bundle ν has a remarkable
feature:
ν ⊕ T S n = Rn+1 .
Hence all spheres are orientable. ⊔
⊓
3.4. INTEGRATION ON MANIFOLDS 121

☞Important convention. The canonical orientation on Rn is the orientation defined

by the volume form dx1 ∧· · ·∧dxn , where x1 , ..., xn are the canonical Cartesian coordinates.
The unit sphere S n ⊂ Rn+1 is orientable. In the sequel we will exclusively deal with
its canonical orientation. To describe this orientation it suffices to describe a positively
oriented basis of det Tp M for some p ∈ S n . To this aim we will use the relation

Rn+1 ∼
= νp ⊕ Tp S n .

An element ω ∈ det Tp S n defines the canonical orientation if p~ ∧ ω ∈ det Rn+1 defines the
canonical orientation of Rn+1 . Above, by ~p we denoted the position vector of p as a point
inside the Euclidean space Rn+1 . We can think of p~ as the “outer” normal to the round
sphere. We call this orientation outer normal first. When n = 1 it coincides with the
counterclockwise orientation of the unit circle S 1 . ⊔
⊓

Lemma 3.4.8. A smooth manifold M is orientable if and only if there exists an open
cover (Uα )α∈A , and local coordinates (x1α , ..., xnα ) on Uα such that
!
∂xiα
det > 0 on Uα ∩ Uβ . (3.4.3)
∂xjβ

Proof. 1. We assume that there exists an open cover with the properties in the lemma,
and we will prove that det T ∗ M is trivial by proving that there exists a volume form.
Consider a partition of unity B ⊂ C0∞ (M ) subordinated to the cover (Uα )α∈A , i.e.,
there exists a map ϕ : B → A such that

supp β ⊂ Uϕ(β) ∀β ∈ B.

Define X
ω := βωϕ(β) ,
β

where for all α ∈ A we define ωα := dx1α ∧ · · · ∧ dxnα . The form ω is nowhere vanishing
since condition (3.4.3) implies that on an overlap Uα1 ∩ · · · ∩ Uαm the forms ωα1 , ..., ωαm
differ by a positive multiplicative factor.
2. Conversely, let ω be a volume form on M and consider an atlas (Uα ; (xiα )). Then

ω |Uα = µα dx1α ∧ · · · ∧ dxnα ,

where the smooth functions µα are nowhere vanishing, and on the overlaps they satisfy
the gluing condition !
∂xiα µβ
∆αβ = det j = .
∂xβ µα

A permutation ϕ of the variables x1α , ..., xnα will change dx1α ∧ · · · ∧ dxnα by a factor ǫ(ϕ) so
we can always arrange these variables in such an order so that µα > 0. This will insure
the positivity condition
∆αβ > 0.
122 CHAPTER 3. CALCULUS ON MANIFOLDS

The lemma is proved. ⊔

⊓

We can rephrase the result in the above lemma in a more conceptual way using the
notion of orientation bundle. Suppose E → M is a real vector bundle of rank r on the
smooth manifold M described by the open cover (Uα )α∈A , and gluing cocycle

gβα : Uαβ → GL(r, R).

The orientation bundle associated to E is the real line bundle Θ(E) → M described by
the open cover (Uα )α∈A , and gluing cocycle

ǫβα := sign det gβα : Uαβ → R∗ = GL(1, R).

We define orientation bundle ΘM of a smooth manifold M as the orientation bundle

associated to the tangent bundle of M , ΘM := Θ(T M ).
The statement in Lemma 3.4.8 can now be rephrased as follows.
Corollary 3.4.9. A smooth manifold M is orientable if and only if the orientation bundle
ΘM is trivializable. ⊔
⊓

From Lemma 3.4.8 we deduce immediately the following consequence.

Proposition 3.4.10. The connected sum of two orientable manifolds is an orientable
manifold. ⊔
⊓

Exercise 3.4.11. Prove the above result. ⊔

⊓

Using Lemma 3.4.8 and Proposition 2.2.75 we deduce the following result.
Proposition 3.4.12. Any complex manifold is orientable. In particular, the complex
Grassmannians Grk (Cn ) are orientable. ⊔
⊓

Exercise 3.4.13. Supply the details of the proof of Proposition 3.4.12. ⊔

⊓

The reader can check immediately that the product of two orientable manifolds is
again an orientable manifold. Using connected sums and products we can now produce
many examples of manifolds. In particular, the connected sums of g tori is an orientable
manifold.
By now the reader may ask where does orientability interact with integration. The
answer lies in Subsection 2.2.4 where we showed that an orientation or on a vector space
V induces a canonical, linear isomorphism ıor : det V ∗ → |Λ|V ; see (2.2.13).
Similarly, an orientation or on a smooth manifold M defines an isomorphism

ıor : C ∞ (det T ∗ M ) → C ∞ (|Λ|M ).

For any compactly supported differential form ω on M of maximal degree we define its
integral by Z Z
ω := ıor ω.
M M
3.4. INTEGRATION ON MANIFOLDS 123

ϕ P
r
y
O

Figure 3.4: Spherical coordinates.

We want to emphasize that this definition depends on the choice of orientation.

We ought to pause and explain in more detail the isomorphism ıor : C ∞ (det ∗

T M) →
∞ i
C (|Λ|M ). Since M is oriented we can choose a coordinate atlas Uα , (xα ) such that

h ∂xi i
β
det j 1≤,i,j≤M
> 0, n = dim M, (3.4.4)
∂xα
and on each coordinate patch Uα the orientation is given by the top degree form dxα =
dx1α ∧ · · · ∧ dxnα .
A top degree differential form ω is described by a collection of forms

ωα = ρα dxα , ρα ∈ C ∞ (Uα ),

and due to the condition (3.4.4) the collection of densities

µα = ρα |dxα | ∈ C ∞ (Uα , | |Λ| |M )

satisfy µα = µβ on the overlap Uαβ . Thus they glue together to a density on M , which is
precisely ıor ω.
Example 3.4.14. Consider Rthe 2-form on R3 , ω = xdy ∧ dz, and let S 2 denote the unit
sphere. We want to compute S 2 ω |S 2 , where S 2 has the canonical orientation. To compute
this integral we will use spherical coordinates (r, ϕ, θ). These are defined by (see Figure
3.4.) 
 x = r sin ϕ cos θ
y = r sin ϕ sin θ .

z = r cos ϕ
At the point p = (1, 0, 0) we have

∂r = ∂x = p~, ∂θ = ∂y ∂ϕ = −∂z ,
124 CHAPTER 3. CALCULUS ON MANIFOLDS

so that the standard orientation on S 2 is given by dϕ ∧ dθ. On S 2 we have r ≡ 1 and

dr ≡ 0 so that

xdy ∧ dz |S 2 = sin ϕ cos θ (cos θ sin ϕdθ + sin θ cos ϕdϕ)) ∧ (− sin ϕ)dϕ

= sin3 ϕ cos2 θdϕ ∧ dθ.

The standard orientation associates to this form de density sin3 ϕ cos2 θ|dϕdθ|, and we
deduce
Z Z Z π Z 2π
3 2 3
ω= sin ϕ cos θ|dϕdθ| = sin ϕdϕ · cos2 θdθ
S2 [0,π]×[0,2π] 0 0

4π
= = volume of the unit ball B 3 ⊂ R3 .
3
As we will see in the next subsection the above equality is no accident. ⊔
⊓

Example 3.4.15 (Invariant integration on compact Lie groups). Let G be a compact,

connected Lie group. Fix once and for all an orientation on the Lie algebra LG . Consider
a positively oriented volume element ω ∈ det L∗G . We can extend ω by left translations
to a left-invariant volume form on G which we continue to denote by ω. This defines an
orientation, and in particular, by integration, we get a positive scalar
Z
c= ω.
G

Set dVG = 1c ω so that Z

dVG = 1. (3.4.5)
G
The differential form dVG is the unique left-invariant n-form (n = dim G) on G satisfying
(3.4.5) (assuming a fixed orientation on G). We claim dVG is also right invariant.
To prove this, consider the modular function G ∋ h 7→ ∆(h) ∈ R defined by

Rh∗ (dVG ) = ∆(h)dVG .

The quantity ∆(h) is independent of h because Rh∗ dVG is a left invariant form, so it has
to be a scalar multiple of dVG . Since (Rh1 h2 )∗ = (Rh2 Rh1 )∗ = Rh∗ 1 Rh∗ 2 we deduce

∆(h1 h2 ) = ∆(h1 )∆(h2 ) ∀h1 , h2 ∈ G.

Hence h 7→ ∆h is a smooth morphism

G → (R \ {0}, ·).

Since G is connected ∆(G) ⊂ R+ , and since G is compact, the set ∆(G) is bounded.
If there exists x ∈ G such that ∆(x) 6= 1, then either ∆(x) > 1, or ∆(x−1 ) > 1, and
in particular, we would deduce the set (∆(xn ))n∈Z is unbounded. Thus ∆ ≡ 1 which
establishes the right invariance of dVG .
3.4. INTEGRATION ON MANIFOLDS 125

The invariant measure dVG provides a very simple way of producing invariant objects
on G. More precisely, if T is tensor field on G, then for each x ∈ G define
Z
ℓ
T x = ((Lg )∗ T )x dVG (g).
G

Then x 7→ T x defines a smooth tensor field on G. We claim that T is left invariant.

Indeed, for any h ∈ G we have
Z Z
ℓ
(Lh )∗ T = (Lh )∗ ((Lg )∗ T )dVG (g) = ((Lhg )∗ T )dVG (g)
G G
Z
u=hg ℓ
= (Lu )∗ T L∗h−1 dVG (u) = T ( L∗h−1 dVG = dVG ).
G
If we average once more on the right we get a tensor
Z
ℓ

G ∋ x 7→ (Rg )∗ T x dVG ,
G

which is both left and right invariant. ⊔

⊓

Exercise 3.4.16. Let G be a Lie group. For any X ∈ LG denote by ad(X) the linear
map LG → LG defined by
LG ∋ Y 7→ [X, Y ] ∈ LG .
(a) If ω denotes a left invariant volume form prove that ∀X ∈ LG

LX ω = − tr ad(X)ω.

(b) Prove that if G is a compact Lie group, then tr ad(X) = 0, for any X ∈ LG . ⊔
⊓

3.4.3 Stokes’ formula

The Stokes’ formula is the higher dimensional version of the fundamental theorem of
calculus (Leibniz-Newton formula)
Z b
df = f (b) − f (a),
a

where f : [a, b] → R is a smooth function and df = f ′ (t)dt. In fact, the higher dimensional
formula will follow from the simplest 1-dimensional situation.
We will spend most of the time finding the correct formulation of the general version,
and this requires the concept of manifold with boundary. The standard example is the
lower half-space
Hn− = { (x1 , ..., xn ) ∈ Rn ; x1 ≤ 0 }.
Definition 3.4.17. A smooth manifold with boundary of dimension n is a topological
space with the following properties.

f that contains M as a closed subset.

(a) There exists a smooth n-dimensional manifold M
126 CHAPTER 3. CALCULUS ON MANIFOLDS

(b) The interior of M , denoted by M 0 , is non empty.

(c) For each point p ∈ ∂M := M \ M 0 , there exist smooth local coordinates (x1 , ..., xn )
defined on an open neighborhood N of p in Mf such that

(c1 ) M 0 ∩ N = q ∈ N; x1 (q) < 0, .
(c2 ) ∂M ∩ N = { x1 = 0 }.

The set ∂M is called the boundary of M . A manifold with boundary M is called

orientable if its interior M 0 is orientable. ⊔
⊓

Example 3.4.18. (a) A closed interval I = [a, b] is a smooth 1-dimensional manifold with
boundary ∂I = {a, b}. We can take M f = R.
(b) The closed unit ball B ⊂ R is an orientable manifold with boundary ∂B 3 = S 2 .
3 3

We can take M f = R3 .
(c) Suppose X is a smooth manifold, and f : X → R is a smooth function such that
0 ∈ R is a regular value of f , i.e.,

f (x) = 0 =⇒ df (x) 6= 0.

Define M := x ∈ X; f (x) ≤ 0 . A simple application of the implicit function theorem
shows that the pair (X, M ) defines a manifold with boundary. Note that examples (a)
and (b) are special cases of this construction. In the case (a) we take X = R, and
f (x) = (x−a)(x−b), while in the case (b) we take X = R3 and f (x, y, z) = (x2 +y 2 +z 2 )−1.
⊔
⊓

Definition 3.4.19. Two manifolds with boundary M1 ⊂ M f1 , and M2 ⊂ M

f2 are said to
fi ,
be diffeomorphic if, for every i = 1, 2 there exists an open neighborhood Ui of Mi in M
and a diffeomorphism F : U1 → U2 such that F (M1 ) = M2 . ⊔
⊓

Exercise 3.4.20. Prove that any manifold with boundary is diffeomorphic to a manifold
with boundary constructed via the process described in Example 3.4.18(c). ⊔
⊓

Proposition 3.4.21. Let M be a smooth manifold with boundary. Then its boundary
∂M is also a smooth manifold of dimension dim ∂M = dim M − 1. Moreover, if M is
orientable, then so is its boundary. ⊔
⊓

The proof is left to the reader as an exercise.

☞Important convention. Let M be an orientable manifold with boundary. There is
a (non-canonical) way to associate to an orientation on M 0 an orientation on the boundary
∂M . This will be the only way in which we will orient boundaries throughout this book.
If we do not pay attention to this convention then our results may be off by a sign.
We now proceed to described this induced orientation on ∂M . For any p ∈ ∂M choose
local coordinates (x1 , ..., xn ) as in Definition 3.4.17. Then the induced orientation of Tp ∂M
is defined by
ǫdx2 ∧ · · · ∧ dxn ∈ det Tp ∂M, ǫ = ±1,
3.4. INTEGRATION ON MANIFOLDS 127

where ǫ is chosen so that for x1 < 0, i.e., inside M , the form

ǫdx1 ∧ dx2 ∧ · · · ∧ dxn

is positively oriented. The differential dx1 is usually called an outer conormal since
x1 increases as we go towards the exterior of M . −dx1 is then the inner conormal for
analogous reasons. The rule by which we get the induced orientation on the boundary can
be rephrased as

{outer conormal} ∧ {orientation on boundary} = {orientation in the interior}.

We may call this rule “outer normal first” for obvious reasons. ⊔
⊓

Example 3.4.22. The canonical orientation on S n ⊂ Rn+1 coincides with the induced
orientation of S n+1 as the boundary of the unit ball B n+1 . ⊔
⊓

Exercise 3.4.23. Consider the hyperplane Hi ⊂ Rn defined by the equation {xi = 0}.
Prove that the induced orientation of Hi as the boundary of the half-space Hn,i i
+ = {x ≥ 0}
ci ∧ · · · dxn where, as usual, the hat indicates
is given by the (n − 1)-form (−1)i dx1 ∧ · · · ∧ dx
a missing term. ⊔
⊓

Theorem 3.4.24 (Stokes formula). Let M be an oriented n-dimensional manifold with

boundary ∂M and ω ∈ Ωn−1 (M ) a compactly supported form. Then
Z Z
dω = ω.
M0 ∂M

In the above formula d denotes the exterior derivative, and ∂M has the induced orientation.

Proof. Via partitions of unity the verification is reduced to the following two situations.
Case 1. The (n − 1)-form ω is compactly supported in Rn . We have to show
Z
dω = 0.
Rn

It suffices to consider only the special case

ω = f (x)dx2 ∧ · · · ∧ dxn ,

where f (x) is a compactly supported smooth function. The general case is a linear com-
bination of these special situations. We compute
Z Z Z Z 
∂f 1 n ∂f
dω = 1
dx ∧ · · · ∧ dx = 1
1
dx dx2 ∧ · · · ∧ dxn = 0,
Rn Rn ∂x Rn−1 R ∂x

since Z
∂f
dx1 = f (∞, x2 , ..., xn ) − f (−∞, x2 , ..., xn ),
R ∂x1
and f has compact support.
128 CHAPTER 3. CALCULUS ON MANIFOLDS

Case 2. The (n − 1)-form ω is compactly supported in Hn− . Let

X
ω= ci ∧ · · · ∧ dxn .
fi (x)dx1 ∧ · · · ∧ dx
i

Then !
X ∂f
dω = (−1)i+1 i dx1 ∧ · · · ∧ dxn .
∂x
i
One verifies as in Case 1 that
Z
∂f 1
i
dx ∧ · · · ∧ dxn = 0 for i 6= 1.
n
H− ∂x

For i = 1 we have
Z Z Z 0 
∂f 1 n ∂f
1
dx ∧ · · · ∧ dx = 1
1
dx dx2 ∧ · · · ∧ dxn
n
H+ ∂x R n−1 −∞ ∂x
Z


= f (0, x2 , ..., xn ) − f (−∞, x2 , ..., xn ) dx2 ∧ · · · ∧ dxn
Rn−1
Z Z
2 n 2 n
= f (0, x , ..., x )dx ∧ · · · ∧ dx = ω.
Rn−1 ∂Hn
−

The last equality follows from the fact that the induced orientation on ∂Hn− is given by
dx2 ∧ · · · ∧ dxn . This concludes the proof of the Stokes formula. ⊔
⊓

Remark 3.4.25.R Stokes formula illustrates an interesting global phenomenon. It shows

that the integral M dω is independent of the behavior of ω inside M . It only depends on
the behavior of ω on the boundary. ⊔
⊓

Example 3.4.26.
Z Z
4π
xdy ∧ dz = dx ∧ dy ∧ dz = vol (B 3 ) = . ⊔
⊓
S2 B3 3
Remark 3.4.27. The above considerations extend easily to more singular situations.
For example, when M is the cube [0, 1]n its topological boundary is no longer a smooth
manifold. However, its singularities are inessential as far as integration is concerned. The
Stokes formula continues to hold
Z Z
dω = ω ∀ω ∈ Ωn−1 (I n ).
In ∂I

The boundary is smooth outside a set of measure zero and is given the induced orientation:
“ outer (co)normal first”. The above equality can be used to give an explanation for the
terminology “exterior derivative” we use to call d. Indeed if ω ∈ Ωn−1 (Rn ) and Ih = [0, h]
then we deduce Z
−n
dω |x=0 = lim h ω. (3.4.6)
h→0 ∂Ihn

When n = 1 this is the usual definition of the derivative. ⊔

⊓
3.4. INTEGRATION ON MANIFOLDS 129

Example 3.4.28. We now have sufficient technical background to describe an example

of vector bundle which admits no flat connections, thus answering the question raised at
the end of Section 3.3.3.
Consider the complex line bundle Ln → S 2 constructed in Example 2.1.38. Recall that
Ln is described by the open cover

S 2 = U0 ∪ U1 , U0 = S 2 \ {South pole}, U1 = S 2 \ {North pole},

and gluing cocycle

g10 : U0 ∩ U1 → C∗ , g10 (z) = z −n = g01 (z)−1 ,

where we identified the overlap U0 ∩ U1 with the punctured complex line C∗ .

A connection on Ln is a collection of two complex valued forms ω0 ∈ Ω1 (U1 ) ⊗ C,
ω1 ∈ Ω1 (U1 ) ⊗ C, satisfying a gluing relation on the overlap (see Example 3.3.7)

dg10 dz
ω1 = − + ω0 = n + ω0 .
g10 z

If the connection is flat, then

dω0 = 0 on U0 and dω1 = 0 on U1 .

Let E + be the Equator equipped with the induced orientation as the boundary of the
northern hemisphere, and E − the equator with the opposite orientation, as the boundary
of the southern hemisphere. The orientation of E + coincides with the orientation given
by the form dθ, where z = exp(iθ).
We deduce from the Stokes formula (which works for complex valued forms as well)
that Z Z Z
ω0 = 0 ω1 = − ω1 = 0.
E+ E+ E−
On the other hand, over the Equator, we have

dz
ω1 − ω0 = n = nidθ,
z
from which we deduce
Z Z
0= ω0 − ω1 = ni dθ = 2nπi !!!
E+ E+

Thus there exist no flat connections on the line bundle Ln , n 6= 0, and at fault is the
gluing cocycle defining L. In a future chapter we will quantify the measure in which the
gluing data obstruct the existence of flat connections. ⊔
⊓

3.4.4 Representations and characters of compact Lie groups

The invariant integration on compact Lie groups is a very powerful tool with many uses.
Undoubtedly, one of the most spectacular application is Hermann Weyl’s computation
130 CHAPTER 3. CALCULUS ON MANIFOLDS

of the characters of representations of compact semi-simple Lie groups. The invariant

integration occupies a central place in his solution to this problem.
We devote this subsection to the description of the most elementary aspects of the
representation theory of compact Lie groups.
Let G be a Lie group. Recall that a (linear) representation of G is a left action on a
(finite dimensional) vector space V

G×V →V (g, v) 7→ T (g)v ∈ V,

such that the map T (g) is linear for any g. One also says that V has a structure of
G-module. If V is a real (respectively complex) vector space, then it is said to be a real
(respectively complex) G-module.
Example 3.4.29. Let V = Cn . Then G = GL(n, C) acts linearly on V in the tautological
manner. Moreover V ∗ , V ⊗k , Λm V and S ℓ V are complex G-modules. ⊔
⊓

Example 3.4.30. Suppose G is a Lie group with Lie algebra LG − T1 G. For every g ∈ G,
the conjugation
C g : G → G, h 7→ C g (h) = ghg −1 ,
sends the identity 1 ∈ G to itself. We denote by Adg : LG → L the differential of
h 7→ C g (h) at h = 1. The operator Adg is a linear isomorphism of LG , and the resulting
map
G 7→ Aut(LG ), g 7→ Adg ,
is called adjoint representation of G.
For example, if G = SO(n), then LG = so(n), and the adjoint representation Ad :
SO(n) → Aut(so(n)), can be given the more explicit description

Ad(g)
so(n) ∋ X 7−→ gXg −1 ∈ so(n), ∀g ∈ SO(n). ⊔
⊓

Exercise 3.4.31. Let G be a Lie group, with Lie algebra LG . Prove that for any X, Y ∈
LG we have
d
|t=0 Adexp(tX) (Y ) = ad(X)Y = [X, Y ]. ⊔
⊓
dt
Definition 3.4.32. A morphism of G-modules V1 and V2 is a linear map L : V1 → V2
such that, for any g ∈ G, the diagram below is commutative, i.e., T2 (g)L = LT1 (g).

V1
L
w V2
T1 (g) T2.(g)
u u
V1
L
w V2
The space of morphisms of G-modules is denoted by HomG (V1 , V2 ). The collection of
isomorphisms classes of complex G-modules is denoted by G − M od. ⊔
⊓
3.4. INTEGRATION ON MANIFOLDS 131

If V is a G-module, then an invariant subspace (or submodule) is a subspace U ⊂ V

such that T (g)(U ) ⊂ U , ∀g ∈ G. A G-module is said to be irreducible if it has no invariant
subspaces other than {0} and V itself.

Proposition 3.4.33. The direct sum “⊕”, and the tensor product “⊗” define a structure
of semi-ring with 1 on G − M od. 0 is represented by the null representation {0}, while 1
is represented by the trivial module G → Aut (C), g 7→ 1. ⊔
⊓

The proof of this proposition is left to the reader.

Example 3.4.34. Let Ti : G → Aut (Ui ) (i = 1, 2) be two complex G-modules. Then U1∗
is a G-module given by (g, u∗ ) 7→ T1 (g−1 )† u∗ . Hence Hom (U1 , U2 ) is also a G-module.
Explicitly, the action of g ∈ G is given by

(g, L) 7−→ T2 (g)LT1 (g −1 ), ∀L ∈ Hom(U1 , U2 ).

We see that HomG (U1 , U2 ) can be identified with the linear subspace in Hom (U1 , U2 )
consisting of the linear maps U1 → U2 unchanged by the above action of G. ⊔
⊓

Proposition 3.4.35 (Weyl’s unitary trick). Let G be a compact Lie group, and V a
complex G-module. Then there exists a Hermitian metric h on V which is G-invariant,
i.e., h(gv1 , gv2 ) = h(v1 , v2 ), ∀v1 , v2 ∈ V .

Proof. Let h be an arbitrary Hermitian metric on V . Define its G-average by

Z
h(u, v) := h(gu, gv)dVG (g),
G

where dVG (g) denotes the normalized bi-invariant measure on G. One can now check
easily that h is G-invariant. ⊔
⊓

In the sequel, G will always denote a compact Lie group.

Proposition 3.4.36. Let V be a complex G-module and h a G-invariant Hermitian met-

ric. If U is an invariant subspace of V then so is U ⊥ , where “⊥” denotes the orthogonal
complement with respect to h.

Proof. Since h is G-invariant it follows that, ∀g ∈ G, the operator T (g) is unitary,

T (g)T ∗ (g) = 1V . Hence, T ∗ (g) = T −1 (g) = T (g −1 ), ∀g ∈ G.
If x ∈ U ⊥ , then for all u ∈ U , and ∀g ∈ G

h(T (g)x, u) = h((x, T ∗ (g)u) = h(x, T (g −1 )u) = 0.

Thus T (g)x ∈ U ⊥ , so that U ⊥ is G-invariant. ⊔

⊓

Corollary 3.4.37. Every G-module V can be decomposed as a direct sum of irreducible

ones. ⊔
⊓
132 CHAPTER 3. CALCULUS ON MANIFOLDS

If we denoted by Irr(G) the collection of isomorphism classes of irreducible G-modules,

then we can rephrase the above corollary by saying that Irr(G) generates the semigroup
(G − M od, ⊕).
To gain a little more insight we need to use the following remarkable trick due to Isaac
Schur.
Lemma 3.4.38 (Schur lemma). Let V1 , V2 be two irreducible complex G-modules. Then

1 if V1 ∼ = V2
dimC HomG (V1 , V2 ) = ∼ .
0 if V1 6= V2
Proof. Let L ∈ HomG (V1 , V2 ). Then ker L ⊂ V1 is an invariant subspace of V1 . Similarly,
Range (L) ⊂ V2 is an invariant subspace of V2 . Thus, either ker L = 0 or ker L = V1 .
The first situation forces Range L 6= 0 and, since V2 is irreducible, we conclude
Range L = V2 . Hence, L has to be in isomorphism of G-modules. We deduce that, if
V1 and V2 are not isomorphic as G-modules, then HomG (V1 , V2 ) = {0}.
Assume now that V1 ∼ = V2 and S : V1 → V2 is an isomorphism of G-modules. According
to the previous discussion, any other nontrivial G-morphism L : V1 → V2 has to be an
isomorphism. Consider the automorphism T = S −1 L : V1 → V1 . Since V1 is a complex
vector space T admits at least one (non-zero) eigenvalue λ.
The map λ1V1 − T is an endomorphism of G-modules, and ker (λ1V1 − T ) 6= 0.
Invoking again the above discussion we deduce T ≡ λ1V1 , i.e. L ≡ λS. This shows
dim HomG (V1 , V2 ) = 1. ⊔
⊓

Schur’s lemma is powerful enough to completely characterize S 1 − M od, the repre-

sentations of S 1 .
Example 3.4.39. (The irreducible (complex) representations of S 1 ). Let V be
a complex irreducible S 1 -module
S 1 × V ∋ (eiθ , v) 7−→ Tθ v ∈ V,
where Tθ1 ·Tθ2 = Tθ1 +θ2 mod 2π , In particular, this implies that each Tθ is an S 1 -automorphism
since it obviously commutes with the action of this group. Hence Tθ = λ(θ)1V which
shows that dim V = 1 since any 1-dimensional subspace of V is S 1 -invariant. We have
thus obtained a smooth map
λ : S 1 → C∗ ,
such that
λ(eiθ · eiτ ) = λ(eiθ )λ(eiθ ).
Hence λ : S 1 → C∗ is a group morphism. As in the discussion of the modular function we
deduce that |λ| ≡ 1. Thus, λ looks like an exponential, i.e., there exists α ∈ R such that
(verify!)
λ(eiθ ) = exp(iαθ), ∀θ ∈ R.
Moreover, exp(2πiα) = 1, so that α ∈ Z.
Conversely, for any integer n ∈ Z we have a representation
ρn
S 1 → Aut (C) (eiθ , z) 7→ einθ z.
The exponentials exp(inθ) are called the characters of the representations ρn . ⊔
⊓
3.4. INTEGRATION ON MANIFOLDS 133

Exercise 3.4.40. Describe the irreducible representations of T n -the n-dimensional torus.

⊔
⊓

Definition 3.4.41. (a) Let V be a complex G-module, g 7→ T (g) ∈ Aut (V ). The

character of V is the smooth function

χV : G → C, χV (g) := tr T (g).

(b) A class function is a continuous function f : G → C such that

f (hgh−1 ) = f (g) ∀g, h ∈ G.

(The character of a representation is an example of class function). ⊔

⊓

Theorem 3.4.42. Let G be a compact Lie group, U1 , U2 complex G-modules and χUi
their characters. Then the following hold.
(a)χU1 ⊕U2 = χU1 + χU2 , χU1 ⊗U2 = χU1 · χU1 .
(b) χUi (1) = dim Ui .
(c) χUi∗ = χUi -the complex conjugate of χUi .
(d) Z
χUi (g)dVG (g) = dim UiG ,
G

where UiG denotes the space of G-invariant elements of Ui ,

UiG = {x ∈ Ui ; x = Ti (g)x ∀g ∈ G}.

(e) Z
χU1 (g) · χU2 (g)dVG (g) = dim HomG (U2 , U1 ).
G

Proof. The parts (a) and (b) are left to the reader. To prove (c), fix an invariant Hermitian
metric on U = Ui . Thus, each T (g) is a unitary operator on U . The action of G on U ∗ is
given by T (g−1 )† . Since T (g) is unitary, we have T (g −1 )† = T (g). This proves (c).
Proof of (d). Consider
Z
P : U → U, P u = T (g)u dVG (g).
G

Note that P T (h) = T (h)P , ∀h ∈ G, i.e., P ∈ HomG (U, U ). We now compute

Z Z
T (h)P u = T (hg)u dVG (g) = T (γ)u Rh∗ −1 dVG (γ),
G G
Z
T (γ)u dVG (γ) = P u.
G
Thus, each P u is G-invariant. Conversely, if x ∈ U is G-invariant, then
Z Z
Px = T (g)xdg = x dVG (g) = x,
G G
134 CHAPTER 3. CALCULUS ON MANIFOLDS

i.e., U G = Range P . Note also that P is a projector, i.e., P 2 = P . Indeed,

Z Z
P 2u = T (g)P u dVG (g) = P u dVG (g) = P u.
G G

Hence P is a projection onto U G , and in particular

Z Z
G
dimC U = tr P = tr T (g) dVG (g) = χU (g) dVG (g).
G G

Proof of (e).
Z Z Z Z
χU1 · χU2 dVG (g) = χU1 · χU2∗ dVG (g) = χU1 ⊗U2∗ dVG (g) = χHom(U2 ,U1 )
G G G G
G
= dimC (Hom (U2 , U1 )) = dimC HomG (U2 , U1 ),

since HomG coincides with the space of G-invariant morphisms. ⊔

⊓

Corollary 3.4.43. Let U , V be irreducible G-modules. Then

Z 
1 , U∼
=V
(χU , χV ) = χU · χV dg = δU V = .
G 0 , U∼
6= V

Proof. Follows from Theorem 3.4.42 using Schur’s lemma. ⊔

⊓

Corollary 3.4.44. Let U , V be two G-modules. Then U ∼

= V if and only if χU = χV .

Proof. Decompose U and V as direct sums of irreducible G-modules

U = ⊕m
1 (mi Ui ) V = ⊕ℓ1 (nj Vj ).
P P
Hence χU = mi χVi and χV = nj χVj . The equivalence “representation” ⇐⇒ “char-
acters” stated by this corollary now follows immediately from Schur’s lemma and the
previous corollary. ⊔
⊓

Thus, the problem of describing the representations of a compact Lie group boils down
to describing the characters of its irreducible representations. This problem was completely
solved by Hermann Weyl, but its solution requires a lot more work that goes beyond the
scope of this book. We will spend the remaining part of this subsection analyzing the
equality (d) in Theorem 3.4.42.
Describing the invariants of a group action was a very fashionable problem in the
second half of the nineteenth century. Formula (d) mentioned above is a truly remarkable
result. It allows (in principle) to compute the maximum number of linearly independent
invariant elements.
Let V be a complex G-module and denote by χV its character. The complex exterior
algebra Λ•c V ∗ is a complex G-module, as the space of complex multi-linear skew-symmetric
maps
V × · · · × V → C.
3.4. INTEGRATION ON MANIFOLDS 135

Denote by bck (V ) the complex dimension of the space of G-invariant elements in Λkc V ∗ .
One has the equality Z
bck (V ) = χΛkc V ∗ dVG (g).
G

These facts can be presented coherently by considering the Z-graded vector space
M
I•c (V ) := Λkinv V ∗ .
k

Its Poincaré polynomial is

X Z
PI•c (V ) (t) = tk bck (V )= tk χΛkc V ∗ dVG (g).
G

To obtain a more concentrated formulation of the above equality we need to recall some
elementary facts of linear algebra.
For each endomorphism A of V denote by σk (A) the trace of the endomorphism

Λk A : Λk V → Λk V.

Equivalently, (see Exercise 2.2.25) the number σk (A) is the coefficient of tk in the charac-
teristic polynomial
σt (A) = det(1V + tA).
Explicitly, σk (A) is given by the sum
X

σk (A) = det aiα iβ (n = dim V ).
1≤i1 ···ik ≤n

If g ∈ G acts on V by T (g), then g acts on Λk V ∗ by Λk T (g−1 )† = Λk T (g). (We

implicitly assumed that each T (g) is unitary with respect to some G-invariant metric on
V ). Hence


χΛkc V ∗ = σk T (g) . (3.4.7)
We conclude that
Z X Z



PI•c (V ) (t) = k
t σk T (g) dVG (g) = det 1V + t T (g) dVG (g). (3.4.8)
G G

Consider now the following situation. Let V be a complex G-module. Denote by Λ•r V the
space of R-multi-linear, skew-symmetric maps

V × · · · × V → R.

The vector space Λ•r V ∗ is a real G-module. We complexify it, so that Λ•r V ⊗ C is the space
of R-multi-linear, skew-symmetric maps

V × · · · × V → C,
136 CHAPTER 3. CALCULUS ON MANIFOLDS

and as such, it is a complex G-module. The real dimension of the subspace Ikr (V ) of
G-invariant elements in Λkr V ∗ will be denoted by brk (V ), so that the Poincaré polynomial
of I•r (V ) = ⊕k Ikr is X
PI•r (V ) (t) = tk brk (V ).
k

On the other hand, ) is equal to the complex dimension of Λkr V ∗ ⊗ C. Using the
brk (V
results of Subsection 2.2.5 we deduce
∗ M ∗

Λ•r V ⊗ C ∼
= Λ•c V ∗ ⊗C Λ•c V = ⊕i+j=k Λic V ∗ ⊗ Λjc V . (3.4.9)
k

Each of the above summands is a G-invariant subspace. Using (3.4.7) and (3.4.9) we
deduce
XZ X


PI•r (V ) (t) = σi T (g) σj T (g) ti+j dVG (g)
k G i+j=k
Z



= det 1V + tT (g) det 1V + t T (g) dVG (g) (3.4.10)
ZG


= det 1V + tT (g) 2
dVG (g).
G

We will have the chance to use this result in computing topological invariants of man-
ifolds with a “high degree of symmetry” like, e.g., the complex Grassmannians.

3.4.5 Fibered calculus

In the previous section we have described the calculus associated to objects defined on a
single manifold. The aim of this subsection is to discuss what happens when we deal with
an entire family of objects parameterized by some smooth manifold. We will discuss only
the fibered version of integration. The exterior derivative also has a fibered version but its
true meaning can only be grasped by referring to Leray’s spectral sequence of a fibration
and so we will not deal with it. The interested reader can learn more about this operation
from [55], Chapter 3, Sec.5.
Assume now that, instead of a single manifold F , we have an entire (smooth) family of
them (Fb )b∈B . In more rigorous terms this means that we are given a smooth fiber bundle
p : E → B with standard fiber F .
On the total space E we will always work with split coordinates (xi ; y j ), where (xi )
are local coordinates on the standard fiber F , and (y j ) are local coordinates on the base
B (the parameter space).
The model situation is the bundle
p p
E = Rk × Rm → Rm = B, (x, y) 7−→ y.

We will first define a fibered version of integration. This requires a fibered version of
orientability.
Definition 3.4.45. Let p : E → B be a smooth bundle with standard fiber F . The
bundle is said to be orientable if the following hold.
3.4. INTEGRATION ON MANIFOLDS 137

(a) The manifold F is orientable;

Ψ
(b) There exists an open cover (Uα ), and trivializations p−1 (Uα ) −→
α
F × Uα , such that
the gluing maps

Ψβ ◦ Ψ−1
α : F × Uαβ → F × Uαβ (Uαβ = Uα ∩ Uβ )

are fiberwise orientation preserving, i.e., for each y ∈ Uαβ , the diffeomorphism

F ∋ f 7→ Ψαβ (f, y) ∈ F

preserves any orientation on F . ⊔

⊓

Exercise 3.4.46. If the base B of an orientable bundle p : E → B is orientable, then so

is the total space E (as an abstract smooth manifold). ⊔
⊓

☞Important convention. Let p : E → B be an orientable bundle with oriented basis

B. The natural orientation of the total space E is defined as follows.
If E = F × B then the orientation of the tangent space T(f,b) E is given by ΩF ×
ωB , where ωF ∈ det Tf F (respectively ωB ∈ det Tb B) defines the orientation of Tf F
(respectively Tb B).
The general case reduces to this one since any bundle is locally a product, and the
gluing maps are fiberwise orientation preserving. This convention can be briefly described
as
orientation total space = orientation fiber ∧ orientation base.
The natural orientation can thus be called the fiber-first orientation. In the sequel all
orientable bundles will be given the fiber-first orientation. ⊔
⊓

Let p : E → B be an orientable fiber bundle with standard fiber F .

Proposition 3.4.47. There exists a linear operator

Z
p∗ = : Ω•cpt(E) → Ω•−r
cpt (B), r = dim F,
E/B

uniquely defined by its action on forms supported on domains D of split coordinates

p
D∼
= Rr × Rm → Rm , (x; y) 7→ y.

If ω = f dxI ∧ dy J , f ∈ Ccpt
∞ (Rr+m ), then

Z 
R
0 , |I| =
6 r
= .
E/B Rr f dxI dy J, |I| = r
R
The operator E/B is called the integration-along-fibers operator. ⊔
⊓
138 CHAPTER 3. CALCULUS ON MANIFOLDS

The proof goes exactly as in the non-parametric case, i.e., when B is a point. One
shows using partitions of unity that these local definitions can be patched together to
produce a well defined map
Z
: Ω•cpt (E) → Ω•−r
cpt (B).
E/B

The details are left to the reader.

Proposition 3.4.48. Let p : E → B be an orientable bundle with an r-dimensional

standard fiber F . Then for any ω ∈ Ω∗cpt(E) and η ∈ ωcpt
∗ (B) such that deg ω + deg η =

dim E we have Z Z
dE ω = (−1)r dB ω.
E/B E/B

If B is oriented and ω η ar as above then

Z Z Z !
ω ∧ p∗ (η) = ω ∧ η. (Fubini)
E B E/B

The last equality implies immediately the projection formula

p∗ (ω ∧ p∗ η) = p∗ ω ∧ η. (3.4.11)

Proof. It suffices to consider only the model case

p
p : E = Rr × Rm → Rm = B, (x; y) → y,

and ω = f dxI ∧ dy J . Then

X ∂f X ∂f
dE ω = i
dxi ∧ dxI ∧ dy J + (−1)|I| dxI ∧ dy j ∧ dy J .
∂x ∂y j
i j

!  
Z Z X ∂f Z
X ∂f
dE ω = dxi ∧ dxI + (−1)|I|  dxI  ∧ dy j ∧ dy J .
E/B Rr ∂xi Rr ∂y j
i j

The above integrals are defined to be zero if the corresponding forms do not have degree
r. Stokes’ formula shows that the first integral is always zero. Hence
 
Z Z X Z
∂  |I| I j J r
dE ω = (−1) dx ∧ dy ∧ dy = (−1) dB ω.
E/B ∂y j Rr E/B
j

The second equality is left to the reader as an exercise. ⊔

⊓
3.4. INTEGRATION ON MANIFOLDS 139

Exercise 3.4.49 (Gelfand–Leray). Suppose that p : E → B is an oriented fibration, ωE

is a volume form on E, and ωB is a volume form on B.
(a) Prove that, for every b ∈ B, there exists a unique volume form ωE/B on Eb = p−1 (b)
with the property that, for every x ∈ Eb , we have

ωE (x) = ωE/B (x) ∧ (p∗ ωB )(x) ∈ Λdim E Tx∗ E.

This form is called the Gelfand–Leray residue of ωE rel p.

(b) Prove that for every compactly supported smooth function f : E → R we have
Z  Z Z Z 


p∗ (f ωE ) b = f ωE/B ωB (b), ∀b ∈ B, f ωE = f ωE/B ωB .
Eb E B Eb

p
(c) Consider the fibration R2 → R, (x, y) 7→ t = ax + by, a2 + b2 6= 0. Compute the
Gelfand–Leray residue dx∧dy
dt along the fiber p(x, y) = 0. ⊔
⊓

Definition 3.4.50. A ∂-bundle is a collection (E, ∂E, p, B) consisting of the following.

(a) A smooth manifold E with boundary ∂E.

(ii) A smooth map p : E → B such that the restrictions p : Int E → B and p : ∂E → B

are smooth bundles.

The standard fiber of p : Int E → B is called the interior fiber. ⊔

⊓

One can think of a ∂-bundle as a smooth family of manifolds with boundary.

Example 3.4.51. The projection

p : [0, 1] × M → M (t; m) 7→ m,

defines a ∂-bundle. The interior fiber is the open interval (0, 1). The fiber of p : ∂(I ×M ) →
M is the disjoint union of two points. ⊔
⊓

Standard Models. A ∂-bundle is obtained by gluing two types of local models.

• Interior models. Rr × Rm → Rm

• Boundary models. Hr+ × Rm → Rm , where


Hr+ := (x1 , · · · , xr ) ∈ Rr ; x1 ≥ 0 .

Remark 3.4.52. Let p : (E, ∂E) → B be a ∂-bundle. If p : Int E → B is orientable and

the basis B is oriented as well, then on ∂E one can define two orientations.

(i) The fiber-first orientation as the total space of an oriented bundle ∂E → B.

(ii) The induced orientation as the boundary of E.

140 CHAPTER 3. CALCULUS ON MANIFOLDS

These two orientations coincide. ⊔

⊓

Exercise 3.4.53. Prove that the above orientations on ∂E coincide. ⊔

⊓

Theorem 3.4.54. Let p : (E, ∂E) → B be an orientable ∂-bundle with an r-dimensional

interior fiber. Then for any ω ∈ Ω•cpt(E) we have
Z Z Z
r
ω= dE ω − (−1) dB ω (Homotopy formula). ⊔
⊓
∂E/B E/B E/B

The last equality can be formulated as

Z Z Z
r
= dE − (−1) dB .
∂E/B E/B E/B

This is “the mother of all homotopy formulæ”. It will play a crucial part in Chapter 7
when we embark on the study of DeRham cohomology.
Exercise 3.4.55. Prove the above theorem. ⊔
⊓
Chapter 4

Riemannian Geometry

Now we can finally put to work the abstract notions discussed in the previous chapters.
Loosely speaking, the Riemannian geometry studies the properties of surfaces (manifolds)
“made of canvas”. These are manifolds with an extra structure arising naturally in many
instances. On such manifolds one can speak of the length of a curve, and the angle between
two smooth curves. In particular, we will study the problem formulated in Chapter 1: why
a plane (flat) canvas disk cannot be wrapped in an one-to-one fashinon around the unit
sphere in R3 . Answering this requires the notion of Riemann curvature which will be the
central theme of this chapter.

4.1 Metric properties

4.1.1 Definitions and examples
To motivate our definition we will first try to formulate rigorously what do we mean by a
“canvas surface”.
A “canvas surface” can be deformed in many ways but with some limitations: it cannot
be stretched as a rubber surface because the fibers of the canvas are flexible but not elastic.
Alternatively, this means that the only operations we can perform are those which do not
change the lengths of curves on the surface. Thus, one can think of “canvas surfaces” as
those surfaces on which any “reasonable” curve has a well defined length.
Adapting a more constructive point of view, one can say that such surfaces are endowed
with a clear procedure of measuring lengths of piecewise smooth curves.
Classical vector analysis describes one method of measuring lengths of smooth paths
in R3 . If γ : [0, 1] → R3 is such a paths, then its length is given by
Z 1
length (γ) = |γ̇(t)|dt,
0

where |γ̇(t)| is the Euclidean length of the tangent vector γ̇(t).

We want to do the same thing on an abstract manifold, and we are clearly faced with
one problem: how do we make sense of the length |γ̇(t)|? Obviously, this problem can be
solved if we assume that there is a procedure of measuring lengths of tangent vectors at
any point on our manifold. The simplest way to do achieve this is to assume that each

141
142 CHAPTER 4. RIEMANNIAN GEOMETRY

tangent space is endowed with an inner product (which can vary from point to point in a
smooth way).
Definition 4.1.1. (a) A Riemann manifold is a pair (M, g) consisting of a smooth
manifold M and a metric g on the tangent bundle, i.e., a smooth, symmetric positive
definite (0, 2)–tensor field on M . The tensor g is called a Riemann metric on M .
(b) Two Riemann manifolds (Mi , gi ) (i = 1, 2) are said to be isometric if there exists a
diffeomorphism φ : M1 → M2 such that φ∗ g2 = g1 . ⊔
⊓

If (M, g) is a Riemann manifold then, for any x ∈ M , the restriction

gx : Tx M × Tx M → R

is an inner product on the tangent space Tx M . We will frequently use the alternative
notation (•, •)x = gx (•, •). The length of a tangent vector v ∈ Tx M is defined as usual,

|v|x := gx (v, v)1/2 .

If γ : [a, b] → M is a piecewise smooth path, then we define its length by

Z b
l(γ) = |γ̇(t)|γ(t) dt.
a

If we choose local coordinates (x1 , . . . , xn ) on M , then we get a local description of g as

g = gij dxi dxj , gij = g(∂xi , ∂xj ).

Proposition 4.1.2. Let M be a smooth manifold, and denote by RM the set of Riemann
metrics on M . Then RM is a non-empty convex cone in the linear space of symmetric
(0, 2)–tensors.

Proof. The only thing that is not obvious is that RM is non-empty. We will use again
partitions of unity. Cover M by coordinate neighborhoods (Uα )α∈A . Let (xiα ) be a collec-
tion of local coordinates on Uα . Using these local coordinates we can construct by hand
the metric gα on Uα by
gα = (dx1α )2 + · · · + (dxnα )2 .
Now, pick a partition of unity B ⊂ C0∞ (M ) subordinated to the cover (Uα )α∈A , i.e., there
exits a map φ : B → A such that ∀β ∈ B supp β ⊂ Uφ(β) . Then define
X
g= βgφ(β) .
β∈B

The reader can check easily that g is well defined, and it is indeed a Riemann metric on
M. ⊔
⊓

Example 4.1.3 (The Euclidean space). The space Rn has a natural Riemann metric

g0 = (dx1 )2 + · · · + (dxn )2 .

The geometry of (Rn , g0 ) is the classical Euclidean geometry. ⊔

⊓
4.1. METRIC PROPERTIES 143

Example 4.1.4 (Induced metrics on submanifolds). Let (M, g) be a Riemann manifold

and S ⊂ M a submanifold. If ı : S → M denotes the natural inclusion then we obtain by
pull back a metric on S
gS = ı∗ g = g |S .
For example, any invertible symmetric n × n matrix defines a quadratic hypersurface in
Rn by 
HA = x ∈ Rn ; (Ax, x) = 1, ,
where (•, •) denotes the Euclidean inner product on Rn . HA has a natural metric induced
by the Euclidean metric on Rn . For example, when A = In , then HIn is the unit sphere
in Rn , and the induced metric is called the round metric of S n−1 . ⊔
⊓

Figure 4.1: The unit sphere and an ellipsoid look “different”.

Figure 4.2: A plane sheet and a half cylinder are “not so different”.

Remark 4.1.5. On any manifold there exist many Riemann metrics, and there is no
natural way of selecting one of them. One can visualize a Riemann structure as defining
a “shape” of the manifold. For example, the unit sphere x2 + y 2 + z 2 = 1 is diffeomorphic
2 2 2
to the ellipsoid x12 + y22 + z32 = 1, but they look “different ” (see Figure 4.1). However,
appearances may be deceiving. In Figure 4.2 it is illustrated the deformation of a sheet
144 CHAPTER 4. RIEMANNIAN GEOMETRY

of paper to a half cylinder. They look different, but the metric structures are the same
since we have not changed the lengths of curves on our sheet. The conclusion to be drawn
from these two examples is that we have to be very careful when we use the attribute
“different”. ⊔
⊓

Example 4.1.6 (The hyperbolic plane). The Poincaré model of the hyperbolic plane is
the Riemann manifold (D, g) where D is the unit open disk in the plane R2 and the metric
g is given by
1
g= (dx2 + dy 2 ). ⊔
⊓
1 − x2 − y 2
Exercise 4.1.7. Let H denote the upper half-plane

H = {(u, v) ∈ R2 ; v > 0},

endowed with the metric

1
h= (du2 + dv 2 ).
4v 2
Show that the Cayley transform

z+i
z = x + iy 7→ w = −i = u + iv
z−i
establishes an isometry (D, g) ∼
= (H, h). ⊔
⊓

Example 4.1.8 (Left invariant metrics on Lie groups). Consider a Lie group G, and
denote by LG its Lie algebra. Then any inner product h·, ·i on LG induces a Riemann
metric h = h·, ·ig on G defined by

hg (X, Y ) = hX, Y ig = h(Lg−1 )∗ X, (Lg−1 )∗ Y i, ∀g ∈ G, X, Y ∈ Tg G,

where (Lg−1 )∗ : Tg G → T1 G is the differential at g ∈ G of the left translation map Lg−1 .

One checks easily that the correspondence

G ∋ g 7→ h·, ·ig

is a smooth tensor field, and it is left invariant, i.e.,

L∗g h = h ∀g ∈ G.

If G is also compact, we can use the averaging technique of Subsection 3.4.2 to produce
metrics which are both left and right invariant. ⊔
⊓

4.1.2 The Levi–Civita connection

To continue our study of Riemann manifolds we will try to follow a close parallel with
classical Euclidean geometry. The first question one may ask is whether there is a notion
of “straight line” on a Riemann manifold.
In the Euclidean space R3 there are at least two ways to define a line segment.
4.1. METRIC PROPERTIES 145

(i) A line segment is the shortest path connecting two given points.

(ii) A line segment is a smooth path γ : [0, 1] → R3 satisfying

γ̈(t) = 0. (4.1.1)

Since we have not said anything about calculus of variations which deals precisely with
problems of type (i), we will use the second interpretation as our starting point. We will
soon see however that both points of view yield the same conclusion.
Let us first reformulate (4.1.1). As we know, the tangent bundle of R3 is equipped
with a natural trivialization, and as such, it has a natural trivial connection ∇0 defined
by
∇0i ∂j = 0 ∀i, j, where ∂i := ∂xi , ∇i := ∇∂i ,
i.e., all the Christoffel symbols vanish. Moreover, if g0 denotes the Euclidean metric, then


∇0i g0 (∂j , ∂k ) = ∇0i δjk − g0 (∇0i ∂j , ∂k ) − g0 (∂j , ∇0i ∂k ) = 0,

i.e., the connection is compatible with the metric. Condition (4.1.1) can be rephrased as

∇0γ̇(t) γ̇(t) = 0, (4.1.2)

so that the problem of defining “lines” in a Riemann manifold reduces to choosing a

“natural” connection on the tangent bundle.
Of course, we would like this connection to be compatible with the metric, but even so,
there are infinitely many connections to choose from. The following fundamental result
will solve this dilemma.
Proposition 4.1.9. Consider a Riemann manifold (M, g). Then there exists a unique
symmetric connection ∇ on T M compatible with the metric g, i.e.,

T (∇) = 0, ∇g = 0.

The connection ∇ is usually called the Levi–Civita connection associated to the metric g.

Proof. Uniqueness. We will achieve this by producing an explicit description of a connec-

tion with the above two mproperties.
Let ∇ be such a connection, i.e.,

∇g = 0 and ∇X Y − ∇Y X = [X, Y ], ∀X, Y ∈ Vect (M ).

For any X, Y, Z ∈ Vect (M ) we have

Zg(X, Y ) = g(∇Z X, Y ) + g(X, ∇Z Y )

since ∇g = 0. Using the symmetry of the connection we compute

Zg(X, Y ) − Y g(Z, X) + Xg(Y, Z) = g(∇Z X, Y ) − g(∇Y Z, X) + g(∇X Y, Z)

146 CHAPTER 4. RIEMANNIAN GEOMETRY

+g(X, ∇Z Y ) − g(Z, ∇Y X) + g(Y, ∇X Z)

= g([Z, Y ], X) + g([X, Y ], Z) + g([Z, X], Y ) + 2g(∇X Z, Y ).
We conclude that
1
g(∇X Z, Y ) = {Xg(Y, Z) − Y g(Z, X) + Zg(X, Y )
2
−g([X, Y ], Z) + g([Y, Z], X) − g([Z, X], Y )}. (4.1.3)
The above equality establishes the uniqueness of ∇.
Using local coordinates (x1 , . . . , xn ) on M we deduce from (4.1.3), with X = ∂i = ∂xi ,
Y = ∂k = ∂xk , Z = ∂j = ∂xj ), that

1
g(∇i ∂j , ∂k ) = gkℓ Γℓij = (∂i gjk − ∂k gij + ∂j gik ) .
2
Above, the scalars Γℓij denote the Christoffel symbols of ∇ in these coordinates, i.e.,

∇∂i ∂j = Γℓij ∂ℓ .

If (giℓ ) denotes the inverse of (giℓ ) we deduce

1
Γℓij = gkℓ (∂i gjk − ∂k gij + ∂j gik ) . (4.1.4)
2
Existence. It boils down to showing that (4.1.3) indeed defines a connection with the
required properties. The routine but rather tedious details are left to the reader. ⊔
⊓

We can now define the notion of “straight line” on a Riemann manifold.

Definition 4.1.10. A geodesic on a Riemann manifold (M, g) is a smooth path

γ : (a, b) → M,

satisfying
∇γ̇(t) γ̇(t) = 0, (4.1.5)
where ∇ is the Levi–Civita connection. ⊔
⊓

Using local coordinates (x1 , ..., xn ) with respect to which the Christoffel symbols are
(Γkij ),
and the path γ is described by γ(t) = (x1 (t), ..., xn (t)), we can rewrite the geodesic
equation as a second order, nonlinear system of ordinary differential equations. Set

d
= γ̇(t) = ẋi ∂i .
dt
Then,
∇ d γ̇(t) = ẍi ∂i + ẋi ∇ d ∂i = ẍi ∂i + ẋi ẋj ∇j ∂i
dt dt

k
= ẍ ∂k + Γkji ẋi ẋj ∂k (Γkij = Γkji ),
4.1. METRIC PROPERTIES 147

so that the geodesic equation is equivalent to

ẍk + Γkij ẋi ẋj = 0 ∀k = 1, ..., n. (4.1.6)

Since the coefficients Γkij = Γkij (x) depend smoothly upon x, we can use the classical
Banach-Picard theorem on existence in initial value problems (see e.g. [6]). We deduce
the following local existence result.
Proposition 4.1.11. Let (M, g) be a Riemann manifold. For any compact subset K ⊂
T M there exists ε > 0 such that for any (x, X) ∈ K there exists a unique geodesic
γ = γx,X : (−ε, ε) → M such that γ(0) = x, γ̇(0) = X. ⊔
⊓

One can think of a geodesic as defining a path in the tangent bundle t 7→ (γ(t), γ̇(t)).
The above proposition shows that the geodesics define a local flow Φ on T M by

Φt (x, X) = (γ(t), γ̇(t)) γ = γx,X .

Definition 4.1.12. The local flow defined above is called the geodesic flow of the Riemann
manifold (M, g). When the geodesic glow is a global flow, i.e., any γx,X is defined at each
moment of time t for any (x, X) ∈ T M , then the Riemann manifold is called geodesically
complete. ⊔
⊓

The geodesic flow has some remarkable properties.

Proposition 4.1.13 (Conservation of energy). If the path γ(t) is a geodesic, then the
length of γ̇(t) is independent of time. ⊔
⊓

Proof. We have
d d
|γ̇(t)|2 = g(γ̇(t), γ̇(t)) = 2g(∇γ̇(t) , γ̇(t)) = 0. ⊔
⊓
dt dt
Thus, if we consider the sphere bundles

Sr (M ) = {X ∈ T M ; |X| = r},

we deduce that Sr (M ) are invariant subsets of the geodesic flow.

Exercise 4.1.14. Describe the infinitesimal generator of the geodesic flow. ⊔
⊓

Example 4.1.15. Let G be a connected Lie group, and let LG be its Lie algebra. Any
X ∈ LG defines an endomorphism ad(X) of LG by

ad(X)Y := [X, Y ].

The Jacobi identity implies that

ad([X, Y ]) = [ad(X), ad(Y )],

where the bracket in the right hand side is the usual commutator of two endomorphisms.
148 CHAPTER 4. RIEMANNIAN GEOMETRY

Assume that there exists an inner product h·, ·i on LG such that, for any X ∈ LG , the
operator ad(X) is skew-adjoint, i.e.,

h[X, Y ], Zi = −hY, [X, Z]i. (4.1.7)

We can now extend this inner product to a left invariant metric h on G. We want to
describe its geodesics.
First, we have to determine the associated Levi–Civita connection. Using (4.1.3) we
get
1
h(∇X Z, Y ) = {Xh(Y, Z) − Y (Z, X) + Zh(X, Y )
2
−h([X, Y ], Z) + h([Y, Z], X) − h([Z, X], Y )}.
If we take X, Y, Z ∈ LG , i.e., these vector fields are left invariant, then h(Y, Z) = const.,
h(Z, X) = const., h(X, Y ) = const. so that the first three terms in the above formula
vanish. We obtain the following equality (at 1 ∈ G)

1
h∇X Z, Y i = {−h[X, Y ], Zi + h[Y, Z], Xi − h[Z, X], Y i}.
2
Using the skew-symmetry of ad(X) and ad(Z) we deduce

1
h∇X Z, Y i = h[X, Z], Y i,
2
so that, at 1 ∈ G, we have
1
∇X Z = [X, Z] ∀X, Z ∈ LG . (4.1.8)
2
This formula correctly defines a connection since any X ∈ Vect (G) can be written as a
linear combination X
X= αi Xi αi ∈ C ∞ (G) Xi ∈ LG .
P
If γ(t) is a geodesic, we can write γ̇(t) = γi Xi , so that
X 1X
0 = ∇γ̇(t) γ̇(t) = γ̇i Xi + γi γj [Xi , Xj ].
2
i i,j

Since [Xi , Xj ] = −[Xj , Xi ], we deduce γ̇i = 0, i.e.,

X
γ̇(t) = γi (0)Xi = X.

This means that γ is an integral curve of the left invariant vector field X so that the
geodesics through the origin with initial direction X ∈ T1 G are

γX (t) = exp(tX). ⊔
⊓
4.1. METRIC PROPERTIES 149

Exercise 4.1.16. Let G be a Lie group and h a bi-invariant metric on G. Prove that its
restriction to LG satisfies (4.1.7). In particular, on any compact Lie groups there exist
metrics satisfying (4.1.7). ⊔
⊓

Definition 4.1.17. Let L be a finite dimensional real Lie algebra. The Killing pairing or
form is the bilinear map

κ : L × L → R, κ(x, y) := −tr (ad(x) ad(y)) x, y ∈ L.

The Lie algebra L is said to be semisimple if the Killing pairing is a duality. A Lie group
G is called semisimple if its Lie algebra is semisimple. ⊔
⊓

Exercise 4.1.18. Prove that SO(n) and SU (n) and SL(n, R) are semisimple Lie groups,
but U (n) is not. ⊔
⊓

Exercise 4.1.19. Let G be a compact Lie group. Prove that the Killing form is positive
semi-definite1 and satisfies (4.1.7).
Hint: Use Exercise 4.1.16. ⊔
⊓

Exercise 4.1.20. Show that the parallel transport of X along exp(tY ) is

(Lexp( t Y ) )∗ (Rexp( t Y ) )∗ X. ⊔
⊓
2 2

Example 4.1.21. (Geodesics on flat tori, and on SU (2)). The n-dimensional torus
Tn ∼
= S 1 × · · · × S 1 is an Abelian, compact Lie group. If (θ 1 , ..., θ n ) are the natural angular
coordinates on T n , then the flat metric is defined by

g = (dθ 1 )2 + · · · + (dθ n )2 .

The metric g on T n is bi-invariant, and obviously, its restriction to the origin satisfies the
skew-symmetry condition (4.1.7) since the bracket is 0. The geodesics through 1 will be
the exponentials
γα1 ,...,αn (t) t 7→ (eiα1 t , ..., eiαn t ) αk ∈ R.
If the numbers αk are linearly dependent over Q, then obviously γα1 ,...,αn (t) is a closed
curve. On the contrary, when the α’s are linearly independent over Q then a classical
result of Kronecker (see e.g. [59]) states that the image of γα1 ,...,αn is dense in T n !!! (see
also Section 7.4 to come)
The special unitary group SU (2) can also be identified with the group of unit quater-
nions 
a + bi + cj + dk ; a2 + b2 + c2 + d2 = 1 ,
so that SU (2) is diffeomorphic with the unit sphere S 3 ⊂ R4 . The round metric on S 3 is
bi-invariant with respect to left and right (unit) quaternionic multiplication (verify this),
1
The converse of the above exercise is also true, i.e., any semisimple Lie group with positive definite
Killing form is compact. This is known as Weyl’s theorem. Its proof, which will be given later in the book,
requires substantially more work.
150 CHAPTER 4. RIEMANNIAN GEOMETRY

and its restriction to (1, 0, 0, 0) satisfies (4.1.7). The geodesics of this metric are the 1-
parameter subgroups of S 3 , and we let the reader verify that these are in fact the great
circles of S 3 , i.e., the circles of maximal diameter on S 3 . Thus, all the geodesics on S 3 are
closed. ⊔
⊓

4.1.3 The exponential map and normal coordinates

We have already seen that there are many differences between the classical Euclidean
geometry and the the general Riemannian geometry in the large. In particular, we have
seen examples in which one of the basic axioms of Euclidean geometry no longer holds: two
distinct geodesic (read lines) may intersect in more than one point. The global topology
of the manifold is responsible for this “failure”.
Locally however, things are not “as bad”. Local Riemannian geometry is similar in
many respects with the Euclidean geometry. For example, locally, all of the classical
incidence axioms hold.
In this section we will define using the metric some special collections of local coordi-
nates in which things are very close to being Euclidean.
Let (M, g) be a Riemann manifold and U an open coordinate neighborhood with
coordinates (x1 , ..., xn ). We will try to find a local change in coordinates (xi ) 7→ (y j )
in which the expression of the metric is as close as possible to the Euclidean metric
g0 = δij dy i dy j .
Let q ∈ U be the point with coordinates (0, ..., 0). Via a linear change in coordinates
we may as well assume that
gij (q) = δij .
We can formulate this by saying that (gij ) is Euclidean up to order zero.
We would like to “spread” the above equality to an entire neighborhood of q. To
achieve this we try to find local coordinates (y j ) near q such that in these new coordinates
the metric is Euclidean up to order one, i.e.,
∂gij ∂δij
gij (q) = δij k
(q) = (q) = 0. ∀i, j, k.
∂y ∂y k
We now describe a geometric way of producing such coordinates using the geodesic flow.
Denote as usual the geodesic from q with initial direction X ∈ Tq M by γq,X (t). Note
the following simple fact.
∀s > 0 γq,sX (t) = γq,X (st).
Hence, there exists a small neighborhood V of 0 ∈ Tq M such that, for any X ∈ V , the
geodesic γq,X (t) is defined for all |t| ≤ 1. We define the exponential map at q by

expq : V ⊂ Tq M → M, X 7→ γq,X (1).

The tangent space Tq M is a Euclidean space, and we can define Dq (r) ⊂ Tq M , the open
“disk” of radius r centered at the origin. We have the following result.
Proposition 4.1.22. Let (M, g) and q ∈ M as above. Then there exists r > 0 such that
the exponential map
expq : Dq (r) → M
4.1. METRIC PROPERTIES 151

is a diffeomorphism onto. The supremum of all radii r with this property is denoted by
ρM (q). ⊔
⊓

Definition 4.1.23. The positive real number ρM (q) is called the injectivity radius of M
at q. The infimum
ρM = inf ρM (q)
q

is called the injectivity radius of M . ⊔

⊓

The proof of Proposition 4.1.22 relies on the following key fact.

Lemma 4.1.24. The Fréchet differential at 0 ∈ Tq M of the exponential map

D0 expq : Tq M → Texpq (0) M = Tq M

is the identity Tq M → Tq M .

Proof. Consider X ∈ Tq M . It defines a line t 7→ tX in Tq M which is mapped via the

exponential map to the geodesic γq,X (t). By definition

(D0 expq )X = γ̇q,X (0) = X. ⊔

⊓

Proposition 4.1.22 follows immediately from the above lemma using the inverse func-
tion theorem. ⊔
⊓

Now choose an orthonormal frame (e1 , ..., en ) of Tq M , and denote by (x1 , ..., xn ) the
resulting cartesian coordinates in Tq M . For 0 < r < ρM (q), any point p ∈ expq (Dq (r))
can be uniquely written as
p = expq (xi ei ),
so that the collection (x1 , ..., xn ) provides a coordinatization of the open set expq (Dq (r)) ⊂
M . The coordinates thus obtained are called normal coordinates at q, the open set
expq (Dq (r)) is called a normal neighborhood, and will be denoted by Br (q) for reasons
that will become apparent a little later.
Proposition 4.1.25. Let (xi ) be normal coordinates at q ∈ M , and denote by g ij the
expression of the metric tensor in these coordinates. Then we have
∂g ij
gij (q) = δij and (q) = 0 ∀i, j, k.
∂xk
Thus, the normal coordinates provide a first order contact between g and the Euclidean
metric.
∂
Proof. By construction, the vectors ei = ∂x i form an orthonormal basis of Tq M and this
proves the first equality. To prove the second equality we need the following auxiliary
result.
Lemma 4.1.26. In normal coordinates (xi ) (at q) the Christoffel symbols Γijk vanish at
q.
152 CHAPTER 4. RIEMANNIAN GEOMETRY

Proof. For any (m1 , ..., mn ) ∈ Rn the curve t 7→ xi (t) 1≤i≤n , xi (t) = mi t, ∀i is the
P i ∂

geodesic t 7→ expq m t ∂xi so that

Γijk (x(t))mj mk = 0.

In particular,
Γijk (0)mj mk = 0 ∀mj ∈ Rn
from which we deduce the lemma. ⊔
⊓

The result in the above lemma can be formulated as

 
∂ ∂
g ∇ ∂ ,
i ∂xk
= 0, ∀i, j, k
∂xj ∂x

so that,
∂
∇ ∂ = 0 at q, ∀i, j. (4.1.9)
∂xj ∂xi
∂g ij
Using ∇g = 0 we deduce ∂xk
(q) = ( ∂x∂ k g ij ) |q = 0. ⊔
⊓

The reader may ask whether we can go one step further, and find local coordinates
which produce a second order contact with the Euclidean metric. At this step we are in for
a big surprise. This thing is in general not possible and, in fact, there is a geometric way
of measuring the “second order distance” between an arbitrary metric and the Euclidean
metric. This is where the curvature of the Levi–Civita connection comes in, and we will
devote an entire section to this subject.

4.1.4 The length minimizing property of geodesics

We defined geodesics via a second order equation imitating the second order equation
defining lines in an Euclidean space. As we have already mentioned, this is not the unique
way of extending the notion of Euclidean straight line to arbitrary Riemann manifolds.
One may try to look for curves of minimal length joining two given points. We will
prove that the geodesics defined as in the previous subsection do just that, at least locally.
We begin with a technical result which is interesting in its own. Let (M, g) be a Riemann
manifold.
Lemma 4.1.27. For each q ∈ M there exists 0 < r < ρM (q), and ε > 0 such that,
∀m ∈ Br (q), we have ε < ρM (m) and Bε (m) ⊃ Br (q). In particular, any two points of
Br (q) can be joined by a unique geodesic of length < ε.
We must warn the reader that the above result does not guarantee that the postulated
connecting geodesic lies entirely in Br (q). This is a different ball game.

Proof. Using the smooth dependence upon initial data in ordinary differential equations
we deduce that there exists an open neighborhood V of (q, 0) ∈ T M such that expm X is
well defined for all (m, X) ∈ V . We get a smooth map

F : V → M × M (m, X) 7→ (m, expm X).

4.1. METRIC PROPERTIES 153

We compute the differential of F at (q, 0). First, using normal coordinates (xi ) near q we
get coordinates (xi ; X j ) near (q, 0) ∈ T M . The partial derivatives of F at (q, 0) are

∂ ∂ ∂ ∂ ∂
D(q,0) F ( )= + i
, D(q,0) F ( i
)= .
∂x i ∂x i
∂X ∂X ∂X i
Thus, the matrix defining D(q,0) F has the form
 
1 0
,
∗ 1

and in particular, it is nonsingular.

It follows from the implicit function theorem that F maps some neighborhood V of
(q, 0) ∈ T M diffeomorphically onto some neighborhood U of (q, q) ∈ M × M . We can
choose V to have the form {(m, X) ; |X|m < ε, m ∈ Bδ (q)} for some sufficiently small ε
and δ. Choose 0 < r < min(ε, ρM (q)) such that

m1 , m2 ∈ Br (q) =⇒ (m1 , m2 ) ∈ U ⇐⇒m2 = expm1 X, |X|m1 < ε.

In particular, we deduce that, for any m ∈ Br (q), the map expm : Dε (m) ⊂ Tm M → M
is a diffeomorphism onto its image, and

Bε (m) = expm (Dε (m)) ⊃ Br (q).

Clearly, for any m ∈ M , the curve t 7→ expm (tX) is a geodesic of length < ε joining m to
expm (X). It is the unique geodesic with this property since F : V → U is injective. ⊔
⊓

We can now formulate the main result of this subsection.

Theorem 4.1.28. Let q, r and ε as in the previous lemma, and consider the unique
geodesic γ : [0, 1] → M of length < ε joining two points of Br (q). If ω : [0, 1] → M is a
piecewise smooth path with the same endpoints as γ then
Z 1 Z 1
|γ̇(t)|dt ≤ |ω̇(t)|dt.
0 0

with equality if and only if ω([0, 1]) = γ([0, 1]). Thus, γ is the shortest path joining its
endpoints.

The proof relies on two lemmata. Let m ∈ M be an arbitrary point, and assume
0 < R < ρM (m).

Lemma 4.1.29 (Gauss). In BR (m) ⊂ M , the geodesics through m are orthogonal to the
hypersurfaces

Σδ = expq (Sδ (0)) = expm (X); |X| = δ , 0 < δ < R.
154 CHAPTER 4. RIEMANNIAN GEOMETRY

Proof. Let t 7→ X(t), 0 ≤ t ≤ 1 denote a smooth curve in Tm M such that |X(t)|m = 1,

i.e., X(t) is a curve on the unit sphere S1 (0) ⊂ Tm M . We assume that the map t 7→
X(t) ∈ S1 (0) is an embedding, i.e., it is injective, and its differential is nowhere zero. We
have to prove that the curve t 7→ expm (δX(t)) are orthogonal to the radial geodesics

s 7→ expm (sX(t)), 0 ≤ s ≤ R.

Consider the smooth map

f : [0, R] × [0, 1] → M, f (s, t) = expm (sX(t)) (s, t) ∈ (0, R) × (0, 1).

If we use normal coordinates on BR (m) we can express f as the embedding

(0, R) × (0, 1) → Tm M, (s, t) 7→ sX(t).

Set  
∂
∂s := f∗ ∈ Tf (s,t) M.
∂s
Define ∂t similarly. The objects ∂s , and ∂t are sections of the restriction of T M to the
image of f . Using the normal coordinates on BR (m) we can think of ∂s as a vector field
on a region in Tm M , and as such we have the equality ∂s = X(t). We have to show

h∂s , ∂t i = 0 ∀(s, t),

where h·, ·i denotes the inner product defined by g.

Using the metric compatibility of the Levi–Civita connection along the image of f we
compute
∂s h∂s , ∂t i = h∇∂s ∂s , ∂t i + h∂s , ∇∂s ∂t i.
Since the curves s 7→ f (s, t = const.) are geodesics, we deduce

∇∂s ∂s = 0.

On the other hand, since [∂s , ∂t ] = 0, we deduce (using the symmetry of the Levi–Civita
connection)
1
h∂s , ∇∂s ∂t i = h∂s , ∇∂t ∂s i = ∂t |∂s |2 = 0,
2
since |∂s | = |X(t)| = 1. We conclude that the quantity h∂s , ∂t i is independent of s. For
s = 0 we have f (0, t) = expm (0) so that ∂t |s=0 = 0, and therefore

h∂s , ∂t i = 0 ∀(s, t),

as needed. ⊔
⊓

Now consider any continuous, piecewise smooth curve

ω : [a, b] → BR (m) \ {m}.

Each ω(t) can be uniquely expressed in the form

ω(t) = expm (ρ(t)X(t)) |X(t)| = 1 0 < |ρ(t)| < R.

4.1. METRIC PROPERTIES 155

Lemma 4.1.30. The length of the curve ω(t) is ≥ |ρ(b) − ρ(a)|. The equality holds if
and only if X(t) = const and ρ̇(t) ≥ 0. In other words, the shortest path joining two
concentrical shells Σδ is a radial geodesic.

Proof. Let f (ρ, t) := expm (ρX(t)), so that ω(t) = f (ρ(t), t). Then
∂f ∂f
ω̇ = ρ̇ + .
∂ρ ∂t
∂f ∂f
Since the vectors ∂ρ and ∂t are orthogonal, and since

∂f
= |X(t)| = 1
∂ρ
we get
2
∂f
|ω̇|2 = |ρ̇|2 + ≥ |ρ̇|2 .
∂t
∂f
The equality holds if and only if ∂t = 0, i.e. Ẋ = 0. Thus
Z b Z b
|ω̇|dt ≥ |ρ̇|dt ≥ |ρ(b) − ρ(a)|.
a a

Equality holds if and only if ρ(t) is monotone, and X(t) is constant. This completes the
proof of the lemma. ⊔
⊓

The proof of Theorem 4.1.28 is now immediate. Let m0 , m1 ∈ Br (q), and consider a
geodesic γ : [0, 1] → M of length < ε such that γ(i) = mi , i = 0, 1. We can write

γ(t) = expm0 (tX) X ∈ Dε (m0 ).

Set R = |X|. Consider any other piecewise smooth path ω : [a, b] → M joining m0 to
m1 . For any δ > 0 this path must contain a portion joining the shell Σδ (m0 ) to the shell
ΣR (m0 ) and lying between them. By the previous lemma the length of this segment will
be ≥ R − δ. Letting δ → 0 we deduce

l(ω) ≥ R = l(γ).

If ω([a, b]) does not coincide with γ([0, 1]) then we obtain a strict inequality. ⊔
⊓

Any Riemann manifold has a natural structure of metric space. More precisely, we set
n o
d(p, q) = inf l(ω); ω : [0, 1] → M piecewise smooth path joining p to q

A piecewise smooth path ω connecting two points p, q such that l(ω) = d(p, q) is said to
be minimal. From Theorem 4.1.28 we deduce immediately the following consequence.
Corollary 4.1.31. The image of any minimal path coincides with the image of a geodesic.
In other words, any minimal path can be reparametrized such that it satisfies the geodesic
equation.
156 CHAPTER 4. RIEMANNIAN GEOMETRY

Exercise 4.1.32. Prove the above corollary. ⊔

⊓

Theorem 4.1.28 also implies that any two nearby points can be joined by a unique
minimal geodesic. In particular we have the following consequence.
Corollary 4.1.33. Let q ∈ M . Then for all r > 0 sufficiently small

expq (Dr (0))( = Br (q) ) = p ∈ M d(p, q) < r . (4.1.10)

Corollary 4.1.34. For any q ∈ M we have the equality


ρM (q) = sup r; r satisfies (4.1.10) .

Proof. The same argument used in the proof of Theorem 4.1.28 shows that for any 0 <
r < ρM (q) the radial geodesics expq (tX) are minimal. ⊔
⊓

Definition 4.1.35. A subset U ⊂ M is said to be convex if any two points in U can be

joined by a unique minimal geodesic which lies entirely inside U . ⊔
⊓

Proposition 4.1.36. For any q ∈ M there exists 0 < R < ιM (q) such that for any r < R
the ball Br (q) is convex.

Proof. Choose ax 0 < ε < 21 ρM (q), and 0 < R < ε such that any two points m0 , m1
in BR (q) can be joined by a unique minimal geodesic [0, 1] ∋ t 7→ γm0 ,m1 (t) of length
< ε, not necessarily contained in BR (q). We will prove that ∀m0 , m1 ∈ BR (q) the map
t 7→ d(q, γm0 ,m1 (t)) is convex and thus it achieves its maxima at the endpoints t = 0, 1.
Note that
d(q, γ(t)) < R + ε < ρM (q).
The geodesic γm0 ,m1 (t) can be uniquely expressed as

γm0 ,m1 (t) = expq (r(t)X(t)) X(t) ∈ Tq M with r(t) = d(q, γm0 ,m1 (t)).
d 2
2
It suffices to show dt2 (r ) ≥ 0 for t ∈ [0, 1] if d(q, m0 ) and d(q, m1 ) are sufficiently small.
At this moment it is convenient to use normal coordinates (xi ) near q. The geodesic
γm0 ,m1 takes the form (xi (t)), and we have

r 2 = (x1 )2 + · · · + (xn )2 .

We compute easily
d2 2
(r ) = 2r 2 (ẍ1 + · · · + ẍn ) + |ẋ|2 (4.1.11)
dt2
P
where ẋ(t) = ẋi ei ∈ Tq M . The path γ satisfies the equation

ẍi + Γijk (x)ẋj ẋk = 0.

Since Γijk (0) = 0 (normal coordinates), we deduce that there exists a constant C > 0
(depending only on the magnitude of the second derivatives of the metric at q) such that

|Γijk (x)| ≤ C|x|.

4.1. METRIC PROPERTIES 157

Using the geodesic equation we obtain

ẍi ≥ −C|x||ẋ|2 .

We substitute the above inequality in (4.1.11) to get

d2 2

2
(r ) ≥ 2|ẋ|2 1 − nC|x|3 . (4.1.12)
dt
If we chose from the very beginning

R + ε ≤ (nC)−1/3 ,

then, because along the geodesic we have |x| ≤ R + ε, the right-hand-side of (4.1.12) is
nonnegative. This establishes the convexity of t 7→ r 2 (t) and concludes the proof of the
proposition. ⊔
⊓

In the last result of this subsection we return to the concept of geodesic completeness.
We will see that this can be described in terms of the metric space structure alone.
Theorem 4.1.37 (Hopf-Rinow). Let M be a Riemann manifold and q ∈ M . The following
assertions are equivalent:
(a) expq is defined on all of Tq M .
(b) The closed and bounded (with respect to the metric structure) sets of M are compact.
(c) M is complete as a metric space.
(d) M is geodesically complete. S
(e) There exists a sequence of compact sets Kn ⊂ M , Kn ⊂ Kn+1 and n Kn = M such
that if pn 6∈ Kn then d(q, pn ) → ∞.
Moreover, on a (geodesically) complete manifold any two points can be joined by a
minimal geodesic. ⊔
⊓

Exercise 4.1.38. Prove the Hopf-Rinow theorem. ⊔

⊓

Remark 4.1.39. (a) Every smooth manifold admits complete Riemann metrics. We refer
to [108] for a proof and more refined statements.
(b) On a complete manifold there could exist points (sufficiently far apart) which can be
joined by more than one minimal geodesic. Think for example of a manifold where there
exist closed geodesic, e.g., the tori T n . ⊔
⊓

4.1.5 Calculus on Riemann manifolds

The classical vector analysis extends nicely to Riemann manifolds. We devote this sub-
section to describing this more general “vector analysis”.
Let (M, g) be an oriented Riemann manifold. We now have two structures at our
disposal: a Riemann metric, and an orientation, and we will use both of them to construct
a plethora of operations on tensors.
First, using the metric we can construct by duality the lowering-the-indices isomor-
phism L : Vect (M ) → Ω1 (M ); see Example 2.2.48.
158 CHAPTER 4. RIEMANNIAN GEOMETRY

Example 4.1.40. Let M = R3 with the Euclidean metric. A vector field V on M has
the form
∂ ∂ ∂
V =P +Q +R .
∂x ∂y ∂z
Then
W = LV = P dx + Qdy + Rdz.
If we think of V as a field of forces in the space, then W is the infinitesimal work of V . ⊔
⊓

On a Riemann manifold there is an equivalent way of describing the exterior derivative.

Proposition 4.1.41. Let

ε : C ∞ (T ∗ M ⊗ Λk T ∗ M ) → C ∞ (Λk+1 T ∗ M )

denote the exterior multiplication operator

ε(α ⊗ β) = α ∧ β, ∀α ∈ Ω1 (M ), β ∈ Ωk (M ).

Then the exterior derivative d is related to the Levi–Civita on Λk T ∗ M connection via the
equality d = ε ◦ ∇.

Proof. We will use a strategy useful in many other situations. Our discussion about normal
coordinates will payoff. Denote temporarily by D the operator ε ◦ ∇.
The equality d = D is a local statement, and it suffices to prove it in any coordinate
neighborhood. Choose (xi ) normal coordinates at an arbitrary point p ∈ M , and set
∂
∂i := ∂x i . Note that
X
D= dxi ∧ ∇i , ∇i = ∇∂i .
i

Let ω ∈ Ωk (M ). Near p it can be written as

X
ω= ωI dxI ,
I

where as usual, for any ordered multi-index I: (1 ≤ i1 < · · · < ik ≤ n), we set

dxI := dxi1 ∧ · · · ∧ dxik .

In normal coordinates at p we have (∇i ∂i ) |p = 0 from which we get the equalities

(∇i dxj ) |p (∂k ) = −(dxj (∇i ∂k )) |p = 0.

Thus, at p, X
Dω = dxi ∧ ∇i (ωI dxI )
I
X X
= dxi ∧ (∂j ωI dxI + ωI ∇i (dxI )) = dxi ∧ ∂j ωI = dω.
I I

Since the point p was chosen arbitrarily this completes the proof of Proposition 4.1.41. ⊔
⊓
4.1. METRIC PROPERTIES 159

Exercise 4.1.42. Show that for any k-form ω on the Riemann manifold (M, g) the exterior
derivative dω can be expressed by
k
X
dω(X0 , X1 , . . . , Xk ) = (−1)i (∇Xi ω)(X0 , . . . , X̂i , . . . , Xk ),
i=0

for all X0 , . . . , Xk ∈ Vect (M ). (∇ denotes the Levi–Civita connection.) ⊔

⊓

The Riemann metric defines a metric in any tensor bundle Tsr (M ) which we continue
to denote by g. Thus, given two tensor fields T1 , T2 of the same type (r, s) we can form
their pointwise scalar product

M ∋ p 7→ g(T, S)p = gp (T1 (p), T2 (p)).

In particular, any such tensor has a pointwise norm

M ∋ p 7→ |T |g,p = (T, T )1/2

p .

Using the orientation we can construct (using the results in subsection 2.2.4) a natural
volume form on M which we denote by dVg , and we call it the metric volume. This is the
positively oriented volume form of pointwise norm ≡ 1.
If (x1 , ..., xn ) are local coordinates such that dx1 ∧ · · · ∧ dxn is positively oriented, then
p
dVg = |g|dx1 ∧ · · · ∧ dxn ,

where |g| := det(gij ). In particular, we can integrate (compactly supported) functions on

M by Z Z
def
f = f dVg ∀f ∈ C0∞ (M ).
(M,g) M

We have the following not so surprising result.

Proposition 4.1.43. ∇X dVg = 0, ∀X ∈ Vect (M ).

Proof. We have to show that for any p ∈ M

(∇X dVg )(e1 , ..., en ) = 0, (4.1.13)

∂
where e1 , ..., ep is a basis of Tp M . Choose normal coordinates (xi ) near p. Set ∂i = ∂x i,
g ij = g(∂i , ∂k ), and ei = ∂i |p . Since the expression in (4.1.13) is linear in X, we may as
well assume X = ∂k , for some k = 1, ..., n. We compute

(∇X dVg )(e1 , ..., en ) = X(dVg (∂1 , ..., ∂n )) |p

X
− dvg (e1 , ..., (∇X ∂i ) |p , ..., ∂n ). (4.1.14)
i

We consider each term separately. Note first that dVg (∂1 , ..., ∂n ) = (det(g ij ))1/2 , so that

X(det(g ij ))1/2 |p = ∂k (det(g ij ))1/2 |p

160 CHAPTER 4. RIEMANNIAN GEOMETRY

is a linear combination of products in which each product has a factor of the form ∂k g ij |p .
Such a factor is zero since we are working in normal coordinates. Thus, the first term
in(4.1.14) is zero. The other terms are zero as well since in normal coordinates at p we
have the equality
∇X ∂i = ∇∂k ∂i = 0.
Proposition 11.3.3 is proved. ⊔
⊓

Once we have an orientation, we also have the Hodge ∗-operator

∗ : Ωk (M ) → Ωn−k (M ),

uniquely determined by

α ∧ ∗β = (α, β)dVg , ∀α β ∈ Ωk (M ). (4.1.15)

In particular, ∗1 = dVg .
Example 4.1.44. To any vector field F = P ∂x + Q∂y + R∂z on R3 we associated its
infinitesimal work
WF = L(F ) = P dx + Qdy + Rdz.
The infinitesimal energy flux of F is the 2-form

ΦF = ∗WF = P dy ∧ dz + Qdz ∧ dx + Rdx ∧ dy.

The exterior derivative of WF is the infinitesimal flux of the vector field curl F

dWF = (∂y R − ∂z Q)dy ∧ dz + (∂z P − ∂x R)dz ∧ dx + (∂x Q − ∂y P )dx ∧ dy

= Φcurl F = ∗Wcurl F .
The divergence of F is the scalar defined as

div F = ∗d ∗ WF = ∗dΦF

= ∗ {(∂x P + ∂y Q + ∂z R)dx ∧ dy ∧ dz} = ∂x P + ∂y Q + ∂z R.

If f is a function on R3 , then we compute easily

∗d ∗ df = ∂x2 f + ∂y2 f + ∂x2 f = ∆f. ⊔

⊓

Definition 4.1.45. (a) For any smooth function f on the Riemann manifold (M, g) we
denote by grad f , or gradg f , the vector field g-dual to the 1-form df . In other words

g(grad f, X) = df (X) = X · f ∀X ∈ Vect (M ).

(b) If (M, g) is an oriented Riemann manifold, and X ∈ Vect (M ), we denote by div X,

or div g X, the smooth function defined by the equality

LX dVg = (div X)dVg . ⊔

⊓
4.1. METRIC PROPERTIES 161

Exercise 4.1.46. Consider the unit sphere


S 2 = (x, y, z) ∈ R3 ; x2 + y 2 + z 2 = 1 ,
and denote by g the Riemann metric on S 2 induced by the Euclidean metric g0 = dx2 +
dy 2 + dz 2 on R3 .
(a) Express g in the spherical coordinates (r, θ, ϕ) defined as in Example 3.4.14.
(b) Denote by h the restriction to S 2 of the function ĥ(x, y, z) = z. Express gradg f in
spherical coordinates. ⊔
⊓

Proposition 4.1.47. Let X be a vector field on the oriented Riemann manifold (M, g),
and denote by α the 1-form dual to X. Then
(a) div X = tr (∇X), where we view ∇X as an element of C ∞ (End (T M )) via the iden-
tifications
∇X ∈ Ω1 (T M ) ∼
= C ∞ (T ∗ M ⊗ T M ) ∼
= C ∞ (End (T M )).
(b) div X = ∗d ∗ α.
(c) If (x1 , ..., xn ) are local coordinates such that dx1 ∧ · · · ∧ dxn is positively oriented, then
1 p
div X = p ∂i ( |g|X i ),
|g|
where X = X i ∂i .
The proof will rely on the following technical result which is interesting in its own. For
simplicity, we will denote the inner products by (•, •), instead of the more precise g(•, •).
Lemma 4.1.48. Denote by δ the operator
δ = ∗d∗ : Ωk (M ) → Ωk−1 (M ).
Let α be a (k − 1)-form and β a k-form such that at least one of them is compactly
supported. Then Z Z
νn,k
(dα, β)dVg = (−1) (α, δβ)dVg ,
M M
where νn,k = nk + n + 1.

Proof. We have
Z Z Z Z
(dα, β)dVg = dα ∧ ∗β = d(α ∧ ∗β) + (−1)k α ∧ d ∗ β.
M M M M
The first integral in the right-hand-side vanishes by the Stokes formula since α ∧ ∗β has
compact support. Since
d ∗ β ∈ Ωn−k+1 (M ) and ∗2 = (−1)(n−k+1)(k−1) on Ωn−k+1 (M )
we deduce Z Z
(dα, β)dVg = (−1)k+(n−k+1)(n−k) α ∧ ∗δβ.
M M
This establishes the assertion in the lemma since
(n − k + 1)(k − 1) + k ≡ νn,k (mod 2). ⊔
⊓
162 CHAPTER 4. RIEMANNIAN GEOMETRY

Definition 4.1.49. Define d∗ : Ωk (M ) → Ωk−1 (M ) by

d∗ := (−1)νn,k δ = (−1)νn,k ∗ d ∗ . ⊔
⊓

Proof of the proposition. Set Ω := dVg , and let (X1 , ..., Xn ) be a local moving frame
of T M in a neighborhood of some point. Then
X
(LX Ω)(X1 , ..., Xn ) = X(Ω(X1 , ..., Xn )) − Ω(X1 , ..., [X, Xi ], ..., Xn ). (4.1.16)
i

Since ∇Ω = 0 we get
X
X · (Ω(X1 , ..., Xn )) = Ω(X1 , ..., ∇X Xi , ..., Xn ).
i

Using the above equality in (4.1.16) we deduce from ∇X Y − [X, Y ] = ∇Y X that

X
(LX Ω)(X1 , ..., Xn ) = Ω(X1 , ..., ∇Xi X, ..., Xn ). (4.1.17)
i

Over the neighborhood where the local moving frame is defined, we can find smooth
functions fij , such that
∇Xi X = fij Xj ⇒ tr (∇X) = fii .
Part (a) of the proposition follows after we substitute the above equality in (4.1.17).
Proof of (b) For any f ∈ C0∞ (M ) we have

LX (f ω) = (Xf )Ω + f (div X)Ω.

On the other hand,

LX (f Ω) = (iX d + d iX )(f Ω) = d iX (f Ω).
Hence
{(Xf ) + f (div X)} dVg = d(iX f Ω).
Since the form f Ω is compactly supported we deduce from Stokes formula
Z
d(iX f Ω) = 0.
M

We have thus proved that for any compactly supported function f we have the equality
Z Z Z
− f (div X)dVg = (Xf )dVg = df (X)dVg
M M M
Z Z
= (grad f, X)dVg = (df, α)dVg .
M M
Using Lemma 4.1.48 we deduce
Z Z
− f (div X)dVg = − f δαdVg ∀f ∈ C0∞ (M ).
M M
4.1. METRIC PROPERTIES 163

This completes the proof of (b).

Proof of (c) We use the equality
p p
LX ( |g|dx1 ∧ · · · ∧ dxn ) = div(X)( |g|dx1 ∧ · · · ∧ dxn ).
The desired formula follows derivating in the left-hand-side. One uses the fact that LX is
an even s-derivation and the equalities
LX dxi = ∂i X i dxi (no summation),
proved in Subsection 3.1.3. ⊔
⊓

Exercise 4.1.50. Let (M, g) be a Riemann manifold and X ∈ Vect (M ). Show that the
following conditions on X are equivalent.
(a) LX g = 0.
(b) g(∇Y X, Z) + g(Y, ∇Z X) = 0 for all Y, Z ∈ Vect (M ).
(A vector field X satisfying the above equivalent conditions is called a Killing vector field).
⊔
⊓

Exercise 4.1.51. Consider a Killing vector field X on the oriented Riemann manifold
(M, g), and denote by η the 1-form dual to X. Show that δη = 0, i.e., div(X) = 0. ⊔
⊓

Definition 4.1.52. Let (M, g) be an oriented Riemann manifold (possibly with bound-
ary). For any k-forms α, β define
Z Z
hα, βi = hα, βiM = (α, β)dVg = α ∧ ∗β,
M M
whenever the integrals in the right-hand-side are finite. ⊔
⊓

Let (M, g) be an oriented Riemann manifold with boundary ∂M . By definition, M is

a closed subset of a boundary-less manifold M̃ of the same dimension. Along ∂M we have
a vector bundle decomposition
(T M̃ ) |∂M = T (∂M ) ⊕ n
where n = (T ∂M )⊥ is the orthogonal complement of T ∂M in (T M ) |∂M . Since both M
and ∂M are oriented manifolds it follows that ν is a trivial line bundle. Indeed, over the
boundary
det T M = det(T ∂M ) ⊗ n
so that
n∼
= det T M ⊗ det(T ∂M )∗ .
In particular, n admits nowhere vanishing sections, and each such section defines an
orientation in the fibers of n.
An outer normal is a nowhere vanishing section σ of n such that, for each x ∈ ∂M ,
and any positively oriented ωx ∈ det Tx ∂M , the product σx ∧ ωx is a positively oriented
element of det Tx M . Since n carries a fiber metric, we can select a unique outer normal
of pointwise length ≡ 1. This will be called the unit outer normal, and will be denoted by
~ν . Using partitions of unity we can extend ~ν to a vector field defined on M .
164 CHAPTER 4. RIEMANNIAN GEOMETRY

Proposition 4.1.53 (Integration by parts). Let (M, g) be a compact, oriented Riemann

manifold with boundary, α ∈ Ωk−1 (M ) and β ∈ Ωk (M ). Then
Z Z Z
(dα, β)dVg = (α ∧ ∗β) |∂M + (α, d∗ β)dVg
M ∂M M
Z Z
= α |∂M ∧ˆ∗(i~ν β) |∂M + (α, d∗ β)dVg
∂M M
where ˆ∗ denotes the Hodge ∗-operator on ∂M with the induced metric ĝ and orientation.
Using the h·, ·i notation of Definition 4.1.52 we can rephrase the above equality as
hdα, βiM = hα, i~ν βi∂M + hα, d∗ βiM .
Proof. As in the proof of Lemma 4.1.48 we have
(dα, β)dVg = dα ∧ ∗β = d(α ∧ ∗β) + (−1)k α ∧ d ∗ β.
The first part of the proposition follows from Stokes formula arguing precisely as in Lemma
4.1.48. To prove the second part we have to check that
(α ∧ ∗β) |∂M = α |∂M ∧ˆ∗(i~ν β) |∂M .
This is a local (even a pointwise) assertion so we may as well assume
M = Hn+ = {(x1 , ..., xn ) ∈ Rn ; x1 ≥ 0},
and that the metric is the Euclidean metric. Note that ~ν = −∂1 . Let I be an ordered
(k − 1)-index, and J be an ordered k-index. Denote by J c the ordered (n − k)-index
complementary to J so that (as sets) J ∪ J c = {1, ..., n}. By linearity, it suffices to
consider only the cases α = dxI , β = dxJ . We have
c
∗dxJ = ǫJ dxJ (ǫJ = ±1) (4.1.18)
and 
J 0 , 1 6∈ J
i~ν dx = ′ ,
−dxJ , 1∈J
where J ′ = J \ {1}. Note that, if 1 6∈ J, then 1 ∈ J c so that
(α ∧ ∗β) |∂M = 0 = α |∂M ∧ˆ∗(i~ν β) |∂M ,
and therefor, the only nontrivial situation left to be discussed is 1 ∈ J. On the boundary
we have the equality
′ c
ˆ ∗(dxJ ) = −ǫ′J dxJ
∗(i~ν dxJ ) = −ˆ (ǫ′J = ±1). (4.1.19)
We have to compare the two signs ǫJ and ǫ′J . in (4.1.18) and (4.1.19). The sign ǫJ is
the signature of the permutation J ~ ∪J c of {1, ..., n} obtained by writing the two increasing
multi-indices one after the other, first J and then J c . Similarly, since the boundary ∂M
has the orientation −dx2 ∧ · · · ∧ dxn , we deduce that the sign ǫ′J is (−1)×(the signature
~ J c of {2, ..., n}). Obviously
of the permutation J ′ ∪
~ J c ),
sign (J ~∪J c ) = sign (J ′ ∪
so that ǫJ = −ǫ′J . The proposition now follows from (4.1.18) and (4.1.19). ⊔
⊓
4.1. METRIC PROPERTIES 165

Corollary 4.1.54 (Gauss). Let (M, g) be a compact, oriented Riemann manifold with
boundary, and X a vector field on M . Then
Z Z
div(X)dVg = (X, ~ν )dvg∂ ,
M ∂M

where g∂ = g |∂M .

Proof. Denote by α the 1-form dual to X. We have

Z Z Z Z
div(X)dVg = 1 ∧ ∗d ∗ αdVg = (1, ∗d ∗ α)dVg = − (1, d∗ α)dVg
M M M M
Z Z
= α(~ν )dvg∂ = (X, ~ν )dvg∂ . ⊔
⊓
∂M ∂M
Remark 4.1.55. The compactness assumption on M can be replaced with an integrability
condition on the forms α, β so that the previous results hold for noncompact manifolds as
well provided all the integrals are finite. ⊔
⊓

Definition 4.1.56. Let (M, g) be an oriented Riemann manifold. The geometric Lapla-
cian is the linear operator ∆M : C ∞ (M ) → C ∞ (M ) defined by

∆M = d∗ df = − ∗ d ∗ df = − div(grad f ).

A smooth function f on M satisfying the equation ∆M f = 0 is called harmonic. ⊔

⊓

Using Proposition 4.1.47 we deduce that in local coordinates (x1 , ..., xn ), the geometer’s
Laplacian takes the form
1 p 
∆M = − p ∂i |g|g ij ∂j ,
|g|

where (gij ) denotes as usual the matrix inverse to (gij ). Note that when g is the Euclidean
metric, then the geometers’ Laplacian is

∆0 = −(∂i2 + · · · + ∂n2 ),

which differs from the physicists’ Laplacian by a sign.

Corollary 4.1.57 (Green). Let (M, g) as in Proposition 4.1.53, and f, g ∈ C ∞ (M ).
Then
∂g
hf, ∆M giM = hdf, dgiM − hf, i∂M ,
∂~ν
and
∂f ∂g
hf, ∆M giM − h∆M f, giM = h , gi∂M − hf, i∂M .
∂~ν ∂~ν
Proof. The first equality follows immediately from the integration by parts formula (Propo-
sition 4.1.53), with α = f , and β = dg. The second identity is now obvious. ⊔
⊓
166 CHAPTER 4. RIEMANNIAN GEOMETRY

Exercise 4.1.58. (a) Prove that the only harmonic functions on a compact oriented
Riemann manifold M are the constant ones. R
(b) If u, f ∈ C ∞ (M ) are such that ∆M u = f show that M f = 0. ⊔
⊓

Exercise 4.1.59. Denote by (u1 , . . . , un ) the coordinates on the round sphere S n ֒→ Rn+1
obtained via the stereographic projection from the south pole.
(a) Show that the round metric g0 on S n is given in these coordinates by
4 
g0 = (du1 )2 + · · · + (dun )2 ,
1 + r2
where r 2 = (u1 )2 + · · · + (un )2 .
(b) Show that the n-dimensional “area” of S n is
Z
2π (n+1)/2
σn = dvg0 = ,
Sn Γ( n+1
2 )

where Γ is Euler’s Gamma function

Z ∞
Γ(s) = ts−1 e−t dt. ⊔
⊓
0

Hint: Use the “doubling formula”

π 1/2 Γ(2s) = 22s−1 Γ(s)Γ(s + 1/2),

and the classical Beta integrals (see [54], or [134], Chapter XII)
Z ∞
r n−1 (Γ(n/2))2
2 n
dr = . ⊔
⊓
0 (1 + r ) 2Γ(n)

Exercise 4.1.60. Consider the Killing form on su(2) (the Lie algebra of SU (2)) defined
by
hX, Y i = −tr X · Y.
(a) Show that the Killing form defines a bi-invariant metric on SU (2), and then compute
the volume of the group with respect to this metric. The group SU (2) is given the
orientation defined by e1 ∧ e2 ∧ e3 ∈ Λ3 su(2), where ei ∈ su(2) are the Pauli matrices
     
i 0 0 1 0 i
e1 = e2 = e3 =
0 −i −1 0 i 0
(b) Show that the trilinear form on su(2) defined by

B(X, Y, Z) = h[X, Y ], Zi,

is skew-symmetric. In particular, B ∈ Λ3 su(2)∗ .

(c) B has a unique extension as a left-invariant 3-form on SUR (2) known as the Cartan
form on SU (2)) which we continue to denote by B. Compute SU (2) B.
Hint: Use the natural diffeomorphism SU (2) ∼ = S 3 , and the computations in the previous
exercise. ⊔
⊓
4.2. THE RIEMANN CURVATURE 167

4.2 The Riemann curvature

Roughly speaking, a Riemann metric on a manifold has the effect of “giving a shape” to
the manifold. Thus, a very short (in diameter) manifold is different from a very long one,
and a large (in volume) manifold is different from a small one. However, there is a lot
more information encoded in the Riemann manifold than just its size. To recover it, we
need to look deeper in the structure, and go beyond the first order approximations we
have used so far.
The Riemann curvature tensor achieves just that. It is an object which is very rich
in information about the “shape” of a manifold, and loosely speaking, provides a second
order approximation to the geometry of the manifold. As Riemann himself observed, we
do not need to go beyond this order of approximation to recover all the information.
In this section we introduce the reader to the Riemann curvature tensor and its as-
sociates. We will describe some special examples, and we will conclude with the Gauss–
Bonnet theorem which shows that the local object which is the Riemann curvature has
global effects.
☞ Unless otherwise indicated, we will use Einstein’s summation convention.

4.2.1 Definitions and properties

Let (M, g) be a Riemann manifold, and denote by ∇ the Levi–Civita connection.
Definition 4.2.1. The Riemann curvature is the tensor R = R(g), defined as

R(g) = F (∇),

where F (∇) is the curvature of the Levi–Civita connection. ⊔

⊓

The Riemann curvature is a tensor R ∈ Ω2 (End (T M )) explicitly defined by

R(X, Y )Z = [∇X , ∇Y ]Z − ∇[X,Y ] Z.

In local coordinates (x1 , . . . , xn ) we have the description

ℓ
Rijk ∂ℓ = R(∂j , ∂k )∂i .

In terms of the Christoffel symbols we have

ℓ
Rijk = ∂j Γℓik − ∂k Γℓij + Γℓmj Γm ℓ m
ik − Γmk Γij .

Lowering the indices we get a new tensor

m



Rijkℓ := gim Rjkℓ = g R(∂k , ∂ℓ )∂j , ∂i = g ∂i , R(∂k , ∂ℓ )∂j .

Theorem 4.2.2 (The symmetries of the curvature tensor). The Riemann curvature tensor
R satisfies the following identities (X, Y, Z, U, V ∈ Vect (M )).

1. g(R(X, Y )U, V ) = −g(R((Y, X), U, V ).

168 Riemannian geometry

2. g(R(X, Y )U, V ) = −g(R(X, Y )V, U ).

3. (The 1st Bianchi identity)
R(X, Y )Z + R(Z, X)Y + R(Y, Z)X = 0.

4.
g(R(X, Y )U, V ) = g(R(U, V )X, Y ).

5. (The 2nd Bianchi identity)

(∇X R)(Y, Z) + (∇Y R)(Z, X) + (∇Z R)(X, Y ) = 0

In local coordinates the above identities have the form

Rijkℓ = −Rjikℓ = −Rijℓk , (4.2.1a)
Rijkℓ = Rkℓij , (4.2.1b)
i i i
Rjkℓ + Rℓjk + Rkℓj = 0, (4.2.1c)
(∇i R)jmkℓ + (∇ℓ R)jmik + (∇k R)jmℓi = 0. (4.2.1d)

Proof. (a) It follows immediately from the definition of R as an End(T M )-valued skew-
symmetric bilinear map (X, Y ) 7→ R(X, Y ).
(b) This is a special case of Proposition 3.3.14 about the curvature of connections com-
patible with a metric.
(c) This is a local statement so we suffices to prove it holds on each coordinate patch. The
identity is also C ∞ (M )-multilinear so it suffices to prove it only in the special case when
the vector fields X, Y , Z pairwise commute. The 1st Bianchi identity is then equivalent
to
∇X ∇Y Z − ∇Y ∇X Z + ∇Z ∇X Y − ∇X ∇Z Y + ∇Y ∇Z X − ∇Z ∇Y X = 0.
The identity now follows from the symmetry of the connection: ∇X Y = ∇Y X etc.
(d) We will use the following algebraic lemma ([80], Chapter 5).
Lemma 4.2.3. Let R : E × E × E × E → R be a quadrilinear map on a real vector space
E. Define
S(X1 , X2 , X3 , X4 ) = R(X1 , X2 , X3 , X4 ) + R(X2 , X3 , X1 , X4 ) + R(X3 , X1 , X2 , X4 ).
If R satisfies the symmetry conditions
R(X1 , X2 , X3 , X4 ) = −R(X2 , X1 , X3 , X4 )
R(X1 , X2 , X3 , X4 ) = −R(X1 , X2 , X4 , X3 ),
then

R(X1 , X2 , X3 , X4 ) − R(X3 , X4 , X1 , X2 )
1n
= S(X1 , X2 , X3 , X4 ) − S(X2 , X3 , X4 , X1 )
2 o
− S(X3 , X4 , X1 , X2 ) + S(X4 , X3 , X1 , X2 ) .
The Riemann curvature 169

The proof of the lemma is a straightforward (but tedious) computation

which is left
to the reader. The Riemann curvature R = g X3 , R(X1 , X2 )X4 satisfies the symmetries
required in the lemma and moreover, the 1st Bianchi identity shows that the associated
form S is identically zero. This concludes the proof of (d).
(e) This is the Bianchi identity we established for any linear connection (see Exercise
3.3.24). ⊔
⊓

Exercise 4.2.4. Denote by Cn of n-dimensional curvature tensors, i.e., tensors (Rijkl ) ∈

(Rn )⊗4 satisfying the conditions,

Rijkℓ = Rkℓij = −Rjikℓ, Rijkℓ + Riℓjk + Rikℓj = 0, ∀i, j, k, ℓ.

Prove that  n
        

2 +1 n 1 n n n
dim Cn = − = +1 − .
2 4 2 2 2 4
(Hint: Consult [16], Chapter 1, Section G.) ⊔
⊓

Remark 4.2.5. The Riemann tensor R of a metric g is naturally a section of Λ2 T ∗ M ⊗

End− (T M ). The tensor R† obtained from R by lowering the indices


R† (X, Y, U V ) = g X, R(U, V )Y

is a section of Λ2 T ∗ M ⊗ T ∗ M . Using the isomorphism End− T ∗ M → Λ2 T ∗ M defined in

(2.2.10) we obtain another tensor ωR ∈ C ∞ (Λ2 T ∗ M ⊗ Λ2 T ∗ M ). The tensors R† and ωR
are related by the equality
R† = −2ωR (4.2.2)


Using the metric g we can identify Λ2 T ∗ M ⊗ Λ2 T ∗ M with the bundle End Λ2 T ∗ M . The
identity (d) in Theorem 4.2.2 states that the tensor R† is symmetric when regarded as an
endomorphism of the bundle Λ2 T ∗ M . ⊔
⊓

The Riemann curvature tensor is the source of many important invariants associated
to a Riemann manifold. We begin by presenting the simplest ones.
Definition 4.2.6. Let (M, g) be a Riemann manifold with curvature tensor R. Any two
vector fields X, Y on M define an endomorphism of T M by

U 7→ R(U, X)Y.

The Ricci curvature is the trace of this endomorphism, i.e.,

Ric (X, Y ) = tr (U 7→ R(U, X)Y ).

We view it as a (0,2)-tensor (X, Y ) 7→ Ric (X, Y ) ∈ C ∞ (M ). ⊔

⊓

If (x1 , . . . , xn ) are local coordinates on M and the curvature R has the local expression
ℓ ) then the Ricci curvature has the local description
R = (Rkij
X
ℓ
Ric = (Ricij ) = Rjℓi .
ℓ
170 Riemannian geometry

The symmetries of the Riemann curvature imply that Ric is a symmetric (0,2)-tensor (as
the metric).
Definition 4.2.7. The scalar curvature s of a Riemann manifold is the trace of the Ricci
tensor. In local coordinates, the scalar curvature is described by

s = gij Ricij = gij Riℓj

ℓ
, (4.2.3)

where (gij ) is the inverse matrix of (gij ). ⊔

⊓

Let (M, g) be a Riemann manifold and p ∈ M . For any linearly independent X, Y ∈

Tp M set
(R(X, Y )Y, X)
Kp (X, Y ) = ,
|X ∧ Y |
where |X ∧ Y | denotes the Gramm determinant

(X, X) (X, Y )
|X ∧ Y | = ,
(Y, X) (Y, Y )

which is non-zero since X and Y are linearly independent. (|X ∧ Y |1/2 measures the area
of the parallelogram in Tp M spanned by X and Y .)
If we choose local coordinates (xi ) near p than we set

Rijij
Kij (p) = Kp (∂xi , ∂xj ) = 2 .
gii gjj − gij

Remark 4.2.8. (a) The sectional curvature determines the Riemann tensor. For a proof
we refer to Thm. 3.8 in [48].
(b) Given a metric g on a smooth manifold M , and a constant λ > 0, we obtained a new,
rescaled metric gλ = λ2 g. A simple computation shows that the Christoffel symbols and
Riemann tensor of gλ are equal with the Christoffel symbols and the Riemann tensor of
the metric g. In particular, this implies

Ricgλ = Ricg .

However, the sectional curvatures are sensitive to metric rescaling.

For example, if g is the canonical metric on the 2-sphere of radius 1 in R3 , then gλ is
the induced metric on the 2-sphere of radius λ in R3 . Intuitively, the larger the constant
λ, the less curved is the corresponding sphere.
In general, for any two linearly independent vectors X, Y ∈ Vect (M ) we have

Kgλ (X, Y ) = λ−2 Kg (X, Y ).

In particular, the scalar curvature changes by the same factor upon rescaling the metric.
If we thing of the metric as a quantity measured in meter2 , then the sectional curvatures
are measured in meter−2 . ⊔
⊓
The Riemann curvature 171

Exercise 4.2.9. Let X, Y, Z, W ∈ Tp M such that span(X,Y)= span(Z,W) is a 2-dimensional

subspace of Tp M prove that Kp (X, Y ) = Kp (Z, W ). ⊔
⊓

According to the above exercise the quantity Kp (X, Y ) depends only upon the 2-plane
in Tp M generated by X and Y . Thus Kp is in fact a function on Gr2 (p) the Grassmannian
of 2-dimensional subspaces of Tp M .
Definition 4.2.10. The function Kp : Gr2 (p) → R defined above is called the sectional
curvature of M at p. ⊔
⊓

Exercise 4.2.11. Prove that

Gr2 (M ) = disjoint union of Gr2 (p) p ∈ M

can be organized as a smooth fiber bundle over M with standard fiber Gr2 (Rn ), n =
dim M , such that, if M is a Riemann manifold, Gr2 (M ) ∋ (p; π) 7→ Kp (π) is a smooth
map. ⊔
⊓

4.2.2 Examples
Example 4.2.12. Consider again the situation discussed in Example 4.1.15. Thus, G is
a Lie group, and h•, •i is a metric on the Lie algebra LG satisfying

had(X)Y, Zi = −hY, ad(X)Zi.

In other words, h•, •i is the restriction of a bi-invariant metric m on G. We have shown

that the Levi–Civita connection of this metric is
1
∇X Y = [X, Y ], ∀X, Y ∈ LG .
2
We can now easily compute the curvature
1 1
R(X, Y )Z = {[X, [Y, Z]] − [Y, [X, Z]]} − [[X, Y ], Z]
4 2
1 1 1 1
(Jacobi identity) = [[X, Y ], Z] + [Y, [X, Z]] − [Y, [X, Z]] − [[X, Y ], Z]
4 4 4 2
1
= − [[X, Y ], Z].
4
We deduce
1 1
hR(X, Y )Z, W i = − h[[X, Y ], Z], W i = had(Z)[X, Y ], W i
4 4
1 1
= − h[X, Y ], ad(Z)W i = − h[X, Y ], [Z, W ]i.
4 4
Now let π ∈ Gr2 (Tg G) be a 2-dimensional subspace of Tg G, for some g ∈ G. If (X, Y ) is
an orthonormal basis of π, viewed as left invariant vector fields on G, then the sectional
curvature along π is
1
Kg (π) = [X, Y ], [X, Y ] ≥ 0.
4
172 Riemannian geometry

Denote the Killing form by κ(X, Y ) = −tr ad(X) ad(Y ) . To compute the Ricci curva-
ture we pick an orthonormal basis E1 , . . . , En of LG . For any X = X i Ei , Y = Y j Ej ∈ LG
we have
1
Ric (X, Y ) = tr (Z 7→ [[X, Z], Y ])
4
1X 1X
= h[[X, Ei ], Y ], Ei )i = − had(Y )[X, Ei ], Ei i
4 4
i i

1X 1X
= h[X, Ei ], [Y, Ei ]i = had(X)Ei , ad(Y )Ei i
4 4
i i

1X 1
1
=− had(Y ) ad(X)Ei , Ei i = − tr ad(Y ) ad(X) = κ(X, Y ).
4 4 4
i

In particular, on a compact semisimple Lie group the Ricci curvature is a symmetric

positive definite (0, 2)-tensor, and more precisely, it is a scalar multiple of the Killing
metric.
We can now easily compute the scalar curvature. Using the same notations as above
we get
1X 1X
s= Ric (Ei , Ei ) = κ(Ei , Ei ).
4 4
i

In particular, if G is a compact semisimple group and the metric is given by the Killing
form then the scalar curvature is
1
s(κ) = dim G. ⊔
⊓
4

Remark 4.2.13. Many problems in topology lead to a slightly more general situation
than the one discussed in the above example namely to metrics on Lie groups which are
only left invariant. Although the results are not as “crisp” as in the bi-invariant case many
nice things do happen. For details we refer to [96]. ⊔
⊓

Example 4.2.14. Let M be a 2-dimensional Riemann manifold (surface), and consider

local coordinates on M , (x1 , x2 ). Due to the symmetries of R,

Rijkl = −Rijlk = Rklij ,

we deduce that the only nontrivial component of the Riemann tensor is R = R1212 . The
sectional curvature is simply a function on M

1 1
K= R1212 = s(g), where |g| = det(gij ).
|g| 2

In this case, the scalar K is known as the total curvature or the Gauss curvature of the
surface.
The Riemann curvature 173

In particular, if M is oriented, and the form dx1 ∧ dx2 defines the orientation, we can
construct a 2-form
1 1 1
ε(g) = Kdvg = s(g)dVg = p R1212 dx1 ∧ dx2 .
2π 4π 2π |g|
The 2-form ε(g) is called the Euler form associated to the metric g. We want to emphasize
that this form is defined only when M is oriented.
We can rewrite this using the pfaffian construction of Subsection 2.2.4. The curvature
R is a 2-form with coefficients in the bundle of skew-symmetric endomorphisms of T M so
we can write  
1 0 R1212
R = A ⊗ dVg , A = p
|g| R2112 0
Assume for simplicity that (x1 , x2 ) are normal coordinates at a point q ∈ M . Thus at q,
|g| = 1 since ∂1 , ∂2 is an orthonormal basis of Tq M . Hence, at q, dVg = dx1 ∧ dx2 , and
1
1
ε(g) = g R(∂1 , ∂2 ∂2 , ∂1 )dx1 ∧ dx2 = R1212 dx1 ∧ dx2 .
2π 2π
Hence we can write
1 1
ε(g) = P f g (−A)dvg =: P f g (−R).
2π 2π
The Euler form has a very nice interpretation in terms of holonomy. Assume √ as before
√
that (x1 , x2 ) are normal coordinates at q, and consider the square St = [0, t] × [0, t]
in the (x1 , x2 ) plane. Denote the (counterclockwise) holonomy along ∂St by Tt . This is
an orthogonal transformation of Tq M , and with respect to the orthogonal basis (∂1 , ∂2 ) of
Tq M , it has a matrix description as
 
cos θ(t) − sin θ(t)
Tt = .
sin θ(t) cos θ(t)
The result in Subsection 3.3.4 can be rephrased as
sin θ(t) = −tg(R(∂1 , ∂2 )∂2 , ∂1 ) + O(t2 ),
so that
R1212 = θ̇(0).
Hence R1212 is simply the infinitesimal angle measuring the infinitesimal rotation suffered
by ∂1 along St . We can think of the Euler form as a “density” of holonomy since it
measures the holonomy per elementary parallelogram. ⊔
⊓

4.2.3 Cartan’s moving frame method

This method was introduced by Élie Cartan at the beginning of the 20th century. Cartan’s
insight was that the local properties of a manifold equipped with a geometric structure can
be very well understood if one knows how the frames of the tangent bundle (compatible
with the geometric structure) vary from one point of the manifold to another. We will
begin our discussion with the model case of Rn . Throughout this subsection we will use
Einstein’s convention.
174 Riemannian geometry

Example 4.2.15. Consider an orthonormal moving frame on Rn , Xα = Xαi ∂i , α = 1, ..., n,

∂
where (x1 , . . . , xn ) are the usual Cartesian coordinates, and ∂i := ∂x α
i . Denote by (θ ) the
∗ n
dual coframe, i.e., the moving frame of T R defined by

θ α (Xβ ) = δβα .

The 1-forms θ α measure the infinitesimal displacement of a point P with respect to the
frame (Xα ). Note that the T M -valued 1-form θ = θ α Xα is the differential of the identity
map 1 : Rn → Rn expressed using the given moving frame.
Introduce the 1-forms ωβα defined by

dXβ = ωβα Xα , (4.2.4)

where we set  
∂Xαi
dXα := dxj ⊗ ∂i .
∂xj
We can form the matrix valued 1-form ω = (ωβα ) which measures the infinitesimal rotation
suffered by the moving frame (Xα ) following the infinitesimal displacement x 7→ x + dx.
In particular, ω = (ωβα ) is a skew-symmetric matrix since

0 = dhXα , Xβ i = hω · Xα , Xβ i + hXα , ω · Xβ i.

Since θ = d1 we deduce

0 = d2 1 = dθ = dθ α ⊗ Xα − θ β ⊗ dXβ = (dθ α − θ β ∧ ωβα ) ⊗ Xα ,

and we can rewrite this as

dθ α = θ β ∧ ωβα , or dθ = −ω ∧ θ. (4.2.5)

Above, in the last equality we interpret ω as a n × n matrix whose entries are 1-forms,
and θ as a column matrix, or a n × 1 matrix whose entries are 1-forms.
Using the equality d2 Xβ = 0 in (4.2.4) we deduce

dωβα = −ωγα ∧ ωβγ , or equivalently, dω = −ω ∧ ω. (4.2.6)

The equations (4.2.5)–(4.2.6) are called the structural equations of the Euclidean space.
The significance of these structural equations will become evident in a little while. ⊔
⊓

We now try to perform the same computations on an arbitrary Riemann manifold

(M, g), dim M = n. We choose a local orthonormal moving frame (Xα )1≤α≤n , and we
construct similarly its dual coframe (θ α )1≤α≤n . Unfortunately, there is no natural way to
define dXα to produce the forms ωβα entering the structural equations. We will find them
using a different (dual) search strategy.
Proposition 4.2.16 (E. Cartan). There exists a collection of 1-forms (ωβα )1≤α,β≤n uniquely
defined by the requirements

(a) ωβα = −ωαβ , ∀α, β.

The Riemann curvature 175

(b) dθ α = θ β ∧ ωβα , ∀α.

Proof. Since the collection of two forms (θ α ∧ θ β )1≤α<β≤n defines a local frame of Λ2 T ∗ Rn ,
α , uniquely determined by the conditions
there exist functions gβγ

1 α β
dθ α = g θ ∧ θ γ , gβγ
α α
= −gγβ .
2 βγ
Uniqueness. Suppose that there exist forms ωβα satisfy the conditions (a)&(b) above. Then
α such that
there exist functions fβγ
α γ
ωβα = fβγ θ .

Then the condition (a) is equivalent to

α = −f β ,
(a1) fβγ αγ
while (b) gives
α − f α = gα
(b1) fβγ γβ βγ
The above two relations uniquely determine the f ’s in terms of the g’s via a cyclic
permutation of the indices α, β, γ

α 1 α β γ
fβγ = (gβγ + gγα − gαβ ) (4.2.7)
2
α θ γ , where the f ’s are given by (4.2.7). We let the reader check
Existence. Define ωβα = fβγ
α
that the forms ωβ satisfy both (a) and (b). ⊔
⊓

The reader may now ask why go through all this trouble. What have we gained by
constructing the forms ω, and after all, what is their significance?
To answer these questions, consider the Levi–Civita connection ∇. Define ω̂βα by

∇Xβ = ω̂βα Xα .

Hence
∇Xγ Xβ = ω̂βα (Xγ )Xα .
Since ∇ is compatible with the Riemann metric, we deduce in standard manner that
ω̂βα = −ω̂αβ .
The differential of θ α can be computed in terms of the Levi–Civita connection (see
Subsection 4.1.5), and we have

dθ α (Xβ , Xγ ) = Xβ θ α (Xγ ) − Xγ θ α (Xβ ) − θ α (∇Xβ Xγ ) + θ α (∇Xγ Xβ )

(use θ α (Xβ ) = δβα = const) = −θ α(ω̂γδ (Xβ )Xδ ) + θ α(ω̂βδ (Xγ )Xδ )
= ω̂βα (Xγ ) − ω̂γα (Xβ ) = (θ β ∧ ω̂βα )(Xβ , Xγ ).
Thus the ω̂’s satisfy both conditions (a) and (b) of Proposition 4.2.16 so that we must
have
ω̂βα = ωβα .
176 Riemannian geometry

In other words, the matrix valued 1-form (ωβα ) is the 1-form associated to the Levi–Civita
connection in this local moving frame. In particular, using the computation in Example
3.3.12 we deduce that the 2-form

Ω = (dω + ω ∧ ω)

is the Riemannian curvature of g. The Cartan structural equations of a Riemann manifold

take the form
dθ = −ω ∧ θ, dω + ω ∧ ω = Ω. (4.2.8)
Comparing these with the Euclidean structural equations we deduce another interpretation
of the Riemann curvature: it measures “the distance” between the given Riemann metric
and the Euclidean one”. We refer to [123] for more details on this aspect of the Riemann
tensor.
The technique of orthonormal frames is extremely versatile in concrete computations.
Example 4.2.17. We will use the moving frame method to compute the curvature of the
hyperbolic plane, i.e., the upper half space

H+ = {(x, y) ; y > 0}

endowed with the metric g = y −2 (dx2 + dy 2 ).

The pair (y∂x , y∂y ) is an orthonormal moving frame, and (θ x = y1 dx, θ y = y1 dy) is its
dual coframe. We compute easily
1 1 1
dθ x = d( dx) = 2 dx ∧ dy = ( dx) ∧ θ y ,
y y y

1 1
dθ y = d( dy) = 0 = (− dx) ∧ θ x .
y y
Thus the connection 1-form in this local moving frame is
" #
0 − y1
ω= 1 dx.
y 0

Note that ω ∧ ω = 0. Using the structural equations (4.2.8) we deduce that the Riemann
curvature is " #  
1
0 2 0 −1
Ω = dω = y
dy ∧ dx = θx ∧ θy .
− y12 0 1 0

The Gauss curvature is

1 1
K= g(Ω(∂x , ∂y )∂y , ∂x ) = y 4 (− 4 ) = −1. ⊔
⊓
|g| y

Exercise 4.2.18. Suppose (M, g) is a Riemann manifold, and u ∈ C ∞ (M ). Define a new

metric gu := e2f g. Using the moving frames method, describe the scalar curvature of gu
in terms of u and the scalar curvature of g. ⊔
⊓
The Riemann curvature 177

4.2.4 The geometry of submanifolds

We now want to apply Cartan’s method of moving frames to discuss the the local geometry
of submanifolds of a Riemann manifold.
Let (M, g) be a Riemann manifold of dimension m, and S a k-dimensional submanifold
in M . The restriction of g to S induces a Riemann metric gS on S. We want to analyze
the relationship between the Riemann geometry of M (assumed to be known) and the
geometry of S with the induced metric.
Denote by ∇M (respectively ∇S ) the Levi–Civita connection of (M, g) (respectively of
(S, gS ). The metric g produces an orthogonal splitting of vector bundles

(T M ) |S ∼
= T S ⊕ NS S.

The NS is called the normal bundle of S ֒→ M , and its is the orthogonal complement of
T S in (T M ) |S . Thus, a section of (T M ) |S , that is a vector field X of M along S, splits
into two components: a tangential component X τ , and a normal component, X ν .
Now choose a local orthonormal moving frame (X1 , ..., Xk ; Xk+1 , ..., Xm ) such that the
first k vectors (X1 , ..., Xk ) are tangent to S. Denote the dual coframe by (θ α )1≤α≤m . Note
that
θ α |S = 0 for α > k.
Denote by (µαβ ), (1 ≤ α, β ≤ m) the connection 1-forms associated to ∇M by this frame,
and let σβα , (1 ≤ α, β ≤ k) be the connection 1-forms of ∇S corresponding to the frame
(X1 , . . . , Xk ). We will analyze the structural equations of M restricted to S ֒→ M .

dθ α = θ β ∧ µαβ 1 ≤ α, β ≤ m. (4.2.9)

We distinguish two situations.

A. 1 ≤ α ≤ k. Since θ β |S = 0 for β > k the equality (4.2.9) yields

k
X
dθ α = θ β ∧ µαβ , µαβ = −µβα 1 ≤ α, β ≤ k.
β=1

The uniqueness part of Proposition 4.2.16 implies that along S

σβα = µαβ 1 ≤ α, β ≤ k.

This can be equivalently rephrased as

∇SX Y = (∇M
XY)
τ
∀X, Y ∈ Vect (S). (4.2.10)

B. k < α ≤ m. We deduce
k
X
0= θ β ∧ µαβ .
β=1

At this point we want to use the following elementary result.

178 Riemannian geometry

Lemma 4.2.19 (Cartan). Let V be a d-dimensional real vector space and consider p
linearly independent elements ω1 , ... , ωp ∈ Λ1 V , p ≤ d. If θ1 , ... , θp ∈ Λ1 V are such that
p
X
θi ∧ ωi = 0,
i=1

then there exist scalars Aij , 1 ≤ i, j ≤ p such that Aij = Aji and
p
X
θi = Aij ωj .
j=1

Proof. Extend the collection (ωi )1≤i≤p to a basis (ωk )1≤k≤d of Λ1 V . Then

d
X
θi = Aik ωk
k=1

and
X p X
X d
0= θi ∧ ωi = Aik ωk ∧ ωi
i i=1 k=1

X p X
X


= Aik − Aki ωk ∧ ωi + Aik ωk ∧ ωi .
1≤k<i≤p i=1 k>p

This proves that Aik = 0 for k > p and Aik = Aki for 1 ≤ i, k ≤ p. ⊔
⊓

λ , λ > k, 1 ≤ β, γ ≤ k satisfying
Using Cartan lemma we can find smooth functions fβγ

λ λ
fβγ = fγβ , and µλβ = fβγ
λ γ
θ .

Now define
λ β
N := fβγ θ ⊗ θ γ ⊗ Xλ .
We can view N as a symmetric bilinear map

Vect (S) × Vect (S) → C ∞ (NS ).

If U, V ∈ Vect (S)
U = U β Xβ = θ β (U )Xβ 1 ≤ β ≤ k,
and
V = V γ Xγ = θ γ (V )Xγ 1 ≤ γ ≤ k,
then    
X k
X X X
N(U, V ) =  λ β
fβγ θ (U )θ γ (V ) X λ =  µλβ (V )U β  Xλ .
λ>k β,γ=1 λ>k β
The Riemann curvature 179

The last term is precisely the normal component of ∇M

V U . We have thus proved the
following equality.

ν
ν
∇M
V U = N(U, V ) = N(V, U ) = ∇M
U V . (4.2.11)

The map N is called the 2nd fundamental form 2 of S ֒→ M .

There is an alternative way of looking at N. Choose

U, V ∈ Vect (S), N ∈ C ∞ (NS ).

If we write g(•, •) = •, • , then

ν
N(U, V ), N = ∇M
U V ,N = ∇M
U V, N

τ
= U · V, N − V, ∇M
U N = − V, ∇M
U N .
We have thus established

τ
τ
− V, ∇M
U N = N(U, V ), N = N(V, U ), N = − U, ∇M
V N . (4.2.12)

The second fundamental form can be used to determine a relationship between the curva-
ture of M and that of S. More precisely we have the following celebrated result.

T heorema Egregium (Gauss). Let RM (resp. RS ) denote the Riemann curvature

of (M, g) (resp. (S, g |S ). Then for any X, Y, Z, T ∈ Vect (S) we have

RM (X, Y )Z, T = RS (X, Y )Z, T

(4.2.13)
+ N(X, Z) , N(Y, T ) − N(X, T ) , N(Y, Z) .

Proof. Note that

∇M S
X Y = ∇X Y + N(X, Y ).

We have
RM (X, Y )X = [∇M M M
X , ∇Y ]Z − ∇[X,Y ] Z



= ∇M
X ∇SY Z + N(Y, Z) − ∇M S S
Y ∇X Z + N(X, Z) − ∇[X,Y ] Z − N([X, Y ], Z).

Take the inner product with T of both sides above. Since N(•, •) is NS -valued, we deduce
using (4.2.10)-(4.2.12)

RM (X, Y )Z, T = ∇M S
X ∇Y Z, T + ∇M
X N(Y, Z), T

−h∇M S
Y ∇X Z, T − ∇M
Y N(X, Z), T − ∇S[X,Y ] Z, T

= [∇SX , ∇SY ]Z, T − N(Y, Z), N(X, T ) + N(X, Z), N(Y, T ) − ∇S[X,Y ] Z, T .
This is precisely the equality (4.2.13). ⊔
⊓
2
The first fundamental form is the induced metric.
180 Riemannian geometry

The above result is especially interesting when S is a transversally oriented hypersur-

face, i.e., S is a a codimension 1 submanifold such that the normal bundle NS is trivial3 .
Pick an orthonormal frame n of NS , i.e., a length 1 section of NS , and choose an orthonor-
mal moving frame (X1 , ..., Xm−1 ) of T S.
Then (X1 , ..., Xm−1 , n) is an orthonormal moving frame of (T M ) |S , and the second
fundamental form is completely described by

Nn (X, Y ) := N(X, Y ), n .

Nn is a bona-fide symmetric bilinear form, and moreover, according to (4.2.12) we have

Nn (X, Y ) = − ∇M
X n, Y = − ∇M
Y n, X . (4.2.14)

In this case, Gauss formula becomes

Nn (X, Z) Nn (X, T )
RS (X, Y )Z, T = RM (X, Y )Z, T − .
Nn (Y, Z) Nn (Y, T )

Let us further specialize, and assume M = Rm , so S is a hypersurface in Rm . Then

Nn (X, T ) Nn (X, Z)
RS (X, Y )Z, T = . (4.2.15)
Nn (Y, T ) Nn (Y, Z)

In particular, the sectional curvature along the plane spanned by X, Y is

RS (X, Y )Y, X = Nn (X, X) · Nn (Y, Y ) − |Nn (X, Y )|2 .

Remark 4.2.20. (a) The equality (4.2.15) is a truly remarkable result. On the right-
hand-side we have an extrinsic term (it depends on the “space surrounding S”), while in
the left-hand-side we have a purely intrinsic term (which is defined entirely in terms of the
internal geometry of S). Historically, the extrinsic term was discovered first (by Gauss),
and very much to Gauss surprise (?!?) one does not need to look outside S to compute
it. This marked the beginning of a new era in geometry. It changed dramatically the way
people looked at manifolds and thus it fully deserves the name of The Golden (egregium)
Theorem of Geometry.
We can now explain rigorously why we cannot wrap a plane canvas around the sphere.
Notice that, when we deform a plane canvas, the only thing that changes is the extrinsic
geometry, while the intrinsic geometry is not changed since the lengths of the“fibers” stays
the same. Thus, any intrinsic quantity is invariant under “bending”. In particular, no
matter how we deform the plane canvas we will always get a surface with Gauss curvature
0 which cannot be wrapped on a surface of constant positive curvature! Gauss himself
called the total curvature a“bending invariant”. ⊔
⊓

Example 4.2.21. Suppose that S is a co-oriented hypersurface of Rm . The choice of

co-orienting unit normal vector field n can be viewed a smooth map n : S → S m−1 where
3
Locally, all hypersurfaces are transversally oriented since NS is locally trivial by definition.
The Riemann curvature 181

S m−1 denotes the unit sphere in Rm . Observe that, for any p ∈ S, we have an equality of
subspaces in Rm
Tp S = Tn(p) S m−1 .
The second fundamental form N of S is symmetric bilinear form on Tp S. Using the metric
on Tp S we can identify it with a symmetric operator

N : Tp S → Tp S = Tn(p) S m−1 , hN X, Y i = N(X, Y ).

Using the equality (4.2.14) we deduce

m m
hN X, Y i = Nn (X, Y ) = − ∇R
Y n, X = − ∇R
X n, Y .

This shows that −N is the differential of the map

G : S → S m−1 , S ∋ p 7→ n(p) ∈ S m−1 .

The map G is called the Gauss map of the hypersurface S ⊂ Rm . The differential of the
Gauss map is known as the shape operator of the hypersurface. . We will have more to
say about the Gauss map in Section 9.2.1.
Suppose now that S is the sphere of radius r in Rm centered at the origin, and n is
the our unit normal vector field along S. In this case the Gauss map is

1
S ∋ x 7→ x ∈ S m−1 .
r
Thus
1
Dx G = 1T S .
r x
1
We deduce from (4.2.15) that the sectional curvature of S along any 2-plane is r2 . ⊔
⊓

Example 4.2.22. (Quadrics in R3 ). Let A : R3 → R3 be a self-adjoint, invertible

linear operator with at least one positive eigenvalue. This implies the quadric

QA = {u ∈ R3 ; Au, u = 1},

is nonempty and smooth (use the implicit function theorem to check this). Let u0 ∈ QA .
Then
Tu0 QA = {x ∈ R3 ; Au0 , x = 0} = (Au0 )⊥ .
QA is a transversally oriented hypersurface in R3 since the map QA ∋ u 7→ Au defines a
1
nowhere vanishing section of the normal bundle. Set n = |Au| Au.
3
Consider an orthonormal frame (e0 , e1 , e2 ) of R such that e0 = n(u0 ). Denote the
Cartesian coordinates in R3 with respect to this frame by (x0 , x1 , x2 ), and set ∂i := ∂xi .
Extend (e1 , e2 ) to a local moving frame of T QA near u0 .
The second fundamental form of QA at u0 is

Nn (∂i , ∂j ) = ∂i n, ∂j |u0 .
182 Riemannian geometry

We compute  
Au −1/2 1
∂i n = ∂i = ∂i ( Au, Au )Au + A∂i u
|Au| |Au|
∂i Au, Au 1
=− Au + ∂i Au.
|Au|3/2 |Au|
Hence
1
Nn (∂i , ∂j ) |u0 = A∂i u, ej |u0
|Au0 |
1 1
= ∂i u, Aej |u0 = ei , Aej . (4.2.16)
|Au0 | |Au0 |
We can now compute the Gaussian curvature at u0 .

1 Ae1 , e1 Ae1 , e2
Ku0 = .
|Au0 |2 Ae2 , e1 Ae2 , e2

In particular, when A = r −2 I so that QA is the round sphere of radius r we deduce

1
Ku = ∀|u| = r.
r2
Thus, the round sphere has constant positive curvature. ⊔
⊓

Example 4.2.23. (Gauss). Let Σ be a transversally oriented, compact surface in R3 ,

e.g., a connected sum of a finite number of tori. Note that the Whitney sum NΣ ⊕ T Σ is
the trivial bundle R3Σ . We orient NΣ such that

orientation NΣ ∧ orientation T Σ = orientation R3 .

Let n be the unit section of NΣ defining the above orientation. The Gauss map

G : Σ → S 2 = {u ∈ R3 ; |u| = 1}, Σ ∋ x 7→ n(x) ∈ S 2 ,

depends on how Σ is embedded in R3 so it is an extrinsic object. Denote by Nn the second

fundamental form of Σ ֒→ R3 , and let (x1 , x2 ) be normal coordinates at q ∈ Σ such that

orientation Tq Σ = ∂1 ∧ ∂2 .

Consider the Euler form εΣ on Σ with the metric induced by the Euclidean metric in R3 .
Then, taking into account our orientation conventions, we have
 
Nn (∂1 , ∂1 ) Nn (∂1 , ∂2 )
2πεΣ (∂1 , ∂2 ) = R1212 = det . (4.2.17)
Nn (∂2 , ∂1 ) Nn (∂2 , ∂2 )

Now notice that

∂i n = −Nn (∂i , ∂1 )∂1 − Nn (∂i , ∂2 )∂2 .
The Riemann curvature 183

We can think of n, ∂1 |q and ∂2 |q as defining a (positively oriented) frame of R3 . The last

equality can be rephrased by saying that the derivative of the Gauss map

G∗ : Tq Σ → Tn(q) S 2

acts according to
∂i 7→ −Nn (∂i , ∂1 )∂1 − Nn (∂i , ∂2 )∂2 .
Hence  
∗ −Nn (∂1 , ∂1 ) −Nn (∂1 , ∂2 )
G dvS 2 = (det G∗ )dvΣ = det dvΣ
−Nn (∂2 , ∂1 ) −Nn (∂2 , ∂2 )
 
Nn (∂1 , ∂1 ) Nn (∂1 , ∂2 )
= det dvΣ = 2πεΣ .
Nn (∂2 , ∂1 ) Nn (∂2 , ∂2 )
Hence
1 ∗
εΣ =
G dv 2 . (4.2.18)
2π Σ S
This is one form of the celebrated Gauss–Bonnet theorem. We will have more to say about
it in Subsection 4.2.6
Note that the last equality offers yet another interpretation of the Gauss curvature.
From this point of view the curvature is a “distortion factor”. The Gauss map “stretches”
an infinitesimal parallelogram to some infinitesimal region on the unit sphere. The Gauss
curvature describes by what factor the area of this parallelogram was changed. In Chap-
ter 9 we will investigate in greater detail the Gauss map of arbitrary submanifolds of a
Euclidean space. ⊔
⊓

Example 4.2.24. Suppose that S is a k-dimensional submanifold of Rm equipped with

the induced metric. Fix a point p0 ∈ S and (u1 , . . . , uk ) are normal coordinates on S at p0 .
The point on S with coordinates (u1 , . . . , uk ) can be identified with a point x = x(u) ∈ Rm .
Classically x(u) is called the position vector of the corresponding point in S.
The tangent spaces Tp S can be identified with a subspaces of Rm . As
such, the frame
moving frame (∂ui )1≤i≤k along S can be identified with the collection x′ui 1≤i≤k of vectors
in Rm .
Since the coordinates u are normal at p0 we deduce that

∇S∂ i ∂ui = 0 at p0 .
u

In other words, at p0 , the vectors x′′ij := x′′ui uj (0) are normal to Tp0 S. Hence

Np0 (∂ui , ∂uj ) = x′′ij .

Using (4.2.13) we deduce

Rijkℓ (p0 ) = ∂k , Rp0 (∂i , ∂j )∂ℓ = hx′′ik , x′′jℓ i − hx′′iℓ , x′′jk i.

If we define
K : S × S → R, K(p, q) := hp, qi,
184 Riemannian geometry

then in a neighborhood of (p0 , p0 ) we can view K as a function of two pairs of variables

u = (u1 , . . . , uk ) and v = (v 1 , . . . , v k ) and we can rewrite the last equality as

Rijkℓ = ∂u4i uk vj vℓ K(u, v) u=v=0

− ∂u4i uℓ vj vk K(u, v) u=v=0
. (4.2.19)

⊔
⊓

4.2.5 Correlators and their geometry

The concept of correlator has its origin in probability (see Remark 4.2.30) but we will
not dwell too much on this aspect. This concept sheds a new light on the results about
the geometry of submanifolds discussed in the previous subsection but, unfortunately, few
geometers are familiar with it. This subsection aims to introduce the audience to this
concept and some of its applications. We follow the presentation in [104] and [105].
First some terminology. Given smooth bundles Ei → Mi , i = 0, 1, over the smooth
manifolds M0 , M1 we denote by E0 ⊞ E1 and E0 ⊠ E1 the bundles

π0∗ (E0 ) ⊕ π1∗ (E1 ), π0∗ (E0 ) ⊗ π1∗ (E1 ) → M0 × M1 ,

where πi : M0 × M1 → Mi denotes the canonical projection.

2 f the section of T ∗ M ⊠ T ∗ M defined as
For f ∈ C ∞ (M0 × M1 ) we denote by ∂x,y 0 1
follows. For p0 ∈ M0 and p1 ∈ U1 choose local coordinates x near p0 and y near p1 . Then
X
2
∂x,y f U0 ×U1 = ∂x2i yj f dxi ⊗ dy j .
i,j

While the above description is local, we leave the reader to convince him/herself that the
resulting object is a well defined global section of T ∗ M0 ⊠ T ∗ M1 .
Definition 4.2.25. A correlator on the smooth manifold M is a smooth function

K :M ×M →R

satisfying the following conditions.

1. The function K is symmetric, i.e., K(p, q) = K(q, p), ∀(p, q) ∈ M × M .

2 K)
2. The symmetric bilinear form (∂x,y ∈ Tp∗ M ⊗ Tp∗ M is positive definite for
p,q p=q
any p ∈ M .

⊔
⊓

Definition 4.2.26. Let M be smooth manifold and V ⊂ C ∞ (M ) a vector space of smooth

functions on M . We say that V is ample if, for any p ∈ M the linear map

V ∋ v 7→ dv(p) ∈ Tp∗ M

is onto. ⊔
⊓
The Riemann curvature 185

Example 4.2.27. Suppose that we are given smooth functions

Ψ1 , . . . , Ψn : M → R

such that span Ψ1 , . . . , ΨN is an ample subspace of C ∞ (M ). Define K : M × M → R
N
X
K(p, q) = Ψi (p)Ψi (q), ∀p, q ∈ M.
i=1

Clearly, condition (i) is satisfied. Next, observe that

N
X
2
∂x,y Kp,q = dΨi (p) ⊗ dΨi (q),
i=1

and, for any p ∈ M and any X ∈ Tp M , we have

N
X
2
∂x,y Kp,p (X, X) = dΨi (X)2 ≥ 0.
i=1

Since the map 

span Ψ1 , . . . , ΨN ∋ Ψ 7→ dΨ(p) ∈ Tp∗ M
is onto, for any X ∈ Tp M \ {0} there exists i ∈ {1, . . . , N } such that dΨi (X) 6= 0. This
proves that ∂x,y 2 K
p,p is positive definite.
As a special case of this construction suppose that M is embedded in RN . Denote by
e1 , . . . , eN the canonical basis of Hom(RN , R). We denote by Ψi the restriction of ei to
M . In this case
N
X N
X
K(p, q) = Ψi (p)Ψi (q) = ei (p)ei (q) = (p, q)RN , (4.2.20)
i=1 i=1

where (−, −)RN denotes the canonical inner product in RN . The ampleness follows from
the fact that the natural map Tp∗ Rn → Tp∗ M is onto, ∀p ∈ M .
We will refer to the correlator constructed in this fashion as the correlator associated
to an embedding. More generally if F : M → Rn is an immersion, then the function


K : M × M → R, K(p, q) = F (p), F (q)

is also a correlator. ⊔
⊓

Suppose that M is a smooth m-dimensional manifold and K is a correlator on M . For

simplicity we denote by E the tangent bundle T M . We set
2
C := ∂x,y K ∈ C ∞ (E ∗ ⊠ E ∗ ).

From the definition of a correlator we see that C defines a metric (−, −)C on E. Note
that when K is the correlator determined by an embedding, the metric defined by C is
none other than the induced metric.
186 Riemannian geometry

For any x, y ∈ M we can identify Cx,y ∈ Ex∗ ⊗ Ey∗ with an element of

Sx,y ∈ Hom(Ey , Ex∗ ).

Since the bilinear form Cx,x : Ex × Ex → R is nondegenerate, the induced map

Sx,x : Ex → Ex∗

is an isomorphism. We set
−1
T (x, y) := Sx,x Sx,y ∈ Hom(Ey , Ex )

We will refer to T (x, y) as the tunneling map associated to the correlator K. Note that
Tx,x = 1Ex .
Fix a point p0 ∈ M and local coordinates (xi )1≤i≤m in a neighborhood O of p0 in
M . Suppose that e(x) = (eα (x))1≤α≤m is a local frame of E|O . We regard it as an
isomorphism of bundles Φ(e) : RrO → E|O . The tunnelings define a smooth map

Te : O × O → End(Rm ), Te (x, y) = Φx (e)−1 Tx,y Φy (e).

If we denote by (eα (x)) the
dual moving frame of E ∗ |O , then Cx,y is described by the
matrix Se (x, y) = sαβ (x, y) 1≤α,β≤m
X
Cx,y = sαβ (x, y)eα (x) ⊗ eβ (y),
α,β

sαβ (x, y) = ∂eα (x) ∂eβ (y) K(x, y).

The fact that Cx,x is a nondegenerate symmetric bilinear form implies that the matrix
Se (x) = Se (x, x) is invertible. Let


Se (x, x)−1 = sαβ (x) 1≤α,β≤m .

Then Te (x, y) is represented by the matrix with entries

X
Tβα (x, y) = sαγ (x)sλβ (x, y).
γ,λ

If we choose the local frame e to be orthonormal with respect to the metric (−, −)C , then

sα,β (x, x) = δαβ .

Hence, in this case Te (x, y) is represented by the matrix sαβ (x, y) 1≤α,β≤m .
∗
If we denote by Tx,y ∈ Hom(Ey , Ex ) the adjoint of Tx,y with respect to the metric
(−, −)C , then the symmetry of K implies that

sαβ (x, y) = sβα (y, x),

i.e.,
∗
Ty,x = Tx,y ,
The Riemann curvature 187

Lemma 4.2.28. Let e be a frame of E|O orthonormal with respect to the correlator metric
(−, −)C . Then, for any i = 1, . . . , m, the operator

∂xi Te (x, y)|x=y : Rm m

y → Rx ,

is skew-symmetric.

Proof. We identify O ×O with an open neighborhood of (0, 0) ∈ Rm ×Rm with coordinates

(xi , y j ). Introduce new coordinates z i := xi − y i , sj := xj + y j , so that ∂xi = ∂z i + ∂si .
We view the map T (e) as depending on the variables z, s. Note that

Te (0, s) = 1, Te (−z, s) = Te (z, s)∗ , ∀z, s.

We deduce that
∂si Te (0, s) = ∂si Te (0, s)∗ = 0,
∂xi Te (0, s) = ∂z i Te (0, s) + ∂si Te (0, s) = ∂z i Te (0, s),

∗
∂xi Te (0, s) = ∂xi Te (0, s)∗ = −∂zi Te (0, s) + ∂si Te (0, s) = −∂xi Te (0, s).
⊔
⊓

Given a coordinate neighborhood with coordinates (xi ) and a local frame e defining
and isomorphism vector bundles Φ(e) : Rm
O → E|O as above, we define the endomorphisms

Γi (e) : Rm m
O → RO , i = 1, . . . , m = dim M,
(4.2.21)
Γi (e)y = −∂xi Te (x, y)|x=y .
P
We obtain a 1-form with matrix coefficients Γ(e) := i Γi (e)dy i . The operator

∇e = d + Γ(e) (4.2.22)

is then a connection on Rm O . When e is orthonormal this connection is also compatible

with the natural metric on this trivial bundle. The isomorphism Φ(e) induces a connection
Φ(e)∗ ∇e on E|O .
Suppose that Φ(f ) : RrO → E|O is another isomorphism determined by the frame f of
EO related to e via a transition map

g : O → GLm (R), f = e · g.

Then
Tf (x, y) = g −1 (x)Te (x, y)g(y).
We denote by dx the differential with respect to the x variable. We deduce

Γ(f )y = −dx T (f )x,y |x=y

= − dx g −1 (x) · Te (y,
x=y | {z }
y) ·g(y) − g −1
(x) dx Te (x, y) |x=y g(y)
=1
−1
=g (y)dg(y)g−1 (y) · g(y) + g −1 (y)Γ(e)y g(y)
188 Riemannian geometry

= g(y)−1 dg(y) + g−1 (y)Γ(e)y g(y).

Thus
Γ(e · g) = g −1 dg + g−1 Γ(e)g.
Proposition 3.3.5 shows that the collection of locally defined 1-forms Γ(e), e local frame of
E, defines a connection ∇C on E. This connection is compatible with the metric (−, −)C
because the local connections ∇e are such when e is an orthonormal frame.
Proposition 4.2.29. Any correlator K on M induces a canonical metric (−, −)C on
T M . The associated Levi–Civita connection coincides with the connection ∇C described
in a local frame e by X
∇C = d + Γi (e)dxi ,
i

where the m × m-matrix Γi (e) is given by (4.2.21). ⊔

⊓

Proof. It suffices to prove that the connection ∇C is symmetric, i.e., its torsion is 0. Once
we choose local coordinates (xi ) on O we have a natural frame of E|O

eα = ∂α := ∂xα .

With these choices Cx,y is represented by the matrix Se (x, y) with entries

sαβ (x, y) = ∂x2α yβ K(x, y).

We denote by sαβ (x) the inverse of (sαβ (x, x)). Then Γi (x) is an m × m matrix


Γi (x) = Γαβi (x) 1≤α,β≤m
.
 
Γi (e, x) = −∂xi Se (x, x)−1 Se (x, y)
x=y
 
= −∂xi Se (x, x)−1 Se (x, x) − Se (x, x)−1 ∂xi Se (x, y) y=x

= Se (x, x)−1 ∂xi Se (x, x) − Se (x, x)−1 ∂xi Se (x, y) y=x

 
= Se (x, x)−1 ∂xi Se (x, y) + ∂yi Se (x, y) − Se (x, x)−1 ∂xi Se (x, y) y=x
x=y
−1
= Se (x, x) ∂yi Se (x, y) y=x
.

The last equality shows that

X
Γαβi (x) = sαγ (x)∂x3γ yβ yi K(x, y)|x=y . (4.2.23)
γ

Note that this implies that

Γαβi = Γαiβ ,

which shows that ∇C is a torsion free connection. ⊔

⊓
The Riemann curvature 189

Let us compute curvature RC of ∇C at a point p0 . Fix a small coordinate neighborhood

O with coordinates (xi ) such that xi (p0 ) = 0, ∀i = 1, . . . , m. Using a local frame e we
can write X
RC = Rij (e, x)dxi ∧ dxj ∈ End(Rm ) ⊗ Ω2 (O),
1≤i<j≤m

where
X
α α
Rij (e, x) = Rβij (e, x) 1≤α,β≤m
, Rβij eα = Rij eβ ,
α
is given by  
Rij (e, y) = ∂xi Γj (e, x) − ∂xj Γi (e, x) + Γi (e, x), Γj (e, x) .
Using the local frame e we represent Se (x, y) as an m × m-matrix



Se (x, y) = sαβ (x, y) 1≤α,β≤m , sαβ (x, y) = Cx,y eα (x), eβ (y) .

Then Te (x, y) = Se (x, x)−1 Se (x, y), and

Γi (e, x) = Se (x, x)−1 ∂yi Se (x, y) y=x

,

 
∂xi Γj (e, x) = ∂xi Se (x, x)−1 ∂yj Se (x, y) y=x
 
= −Se (x, x)−1 ∂xi Se (x, x) Se (x, x)−1 ∂yj Se (x, y) x=y
 
+Se (x, x)−1 ∂xi ∂yj Se (x, y) x=y
 
= −Se (x, x)−1 ∂xi Se (x, y) + ∂yi Se (x, y) Se (x, x)−1 ∂yj Se (x, y) x=y
x=y
+Se (x, x)−1 ∂x2i yj Se (x, y) x=y + Se (x, x)−1 ∂y2i yj Se (x, y) x=y
.

Suppose now that eα = ∂xα and e is orthonormal at p0 where x = y = 0. More precisely,

assume that
C0,0 (∂xα , ∂xβ ) = δαβ , L > 0.
Then
S(0, 0)−1 = 1,
and
∂xi Γj (e, 0) − ∂xj Γi (e, 0)
 
= − ∂xi Se (x, y) ∂ j S (x, y)
x=y=0 y e x=y=0
+ ∂ x j Se (x, y)
x=y=0
∂y i Se (x, y)
x=y=0
 
2 2
+ ∂xi yj Se (x, y) x=y=0 − ∂xj yi Se (x, y) x=y=0 ,
and
 
Γi (e, 0), Γj (e, 0) = ∂yi Se (x, y) ∂ j S (x, y) x=y=0
x=y=0 y e
− ∂yj Se (x, y) ∂ i.
x=y=0 y

If, additionally, the frame (∂xα ) is synchronous at p0 , i.e.,

∇C
∂ i ∂xα (p0 ) = 0, ∀i, α,
x
190 Riemannian geometry

then
∂yj Se (x, y) x=y=0
= Γi (e, 0) = 0,
so that
Rij (e, 0) = ∂x2i yj Se (x, y) x=y=0
− ∂x2j yi Se (x, y) x=y=0
and
 
γ
Rαβij = δαγ Rβij = ∂x4α xi yβ yj K(x, y) x=y=0
− ∂x4α xj yβ yi K(x, y) x=y=0
. (4.2.24)

This should be compared with (4.2.19).

Remark 4.2.30. We cannot end this section without giving the reader an idea about the
probabilistic flavor of the concept of correlator. To keep things simple suppose that

K :M ×M →R

is the correlator determined by an embedding M ֒→ RN .

Denote by e1 , . . . , eN the canonical basis of the dual Hom(RN , R). Denote by Ψα
the restriction to M of the linear function eα : RN → R. Now choose N independent
random variables A1 , . . . , AN with mean 0 and variance 14 and form the random linear
combination
XN
Ψ= Aα Ψα
α=1
The random object Ψ is known as a random function or random field on M . The value
of Ψ at a fixed point p is the random variable
X
Ψ(p) = Aα Ψα (p).
α

Its mean or expectation is

  X 
E Ψ(p) = Ψk (p)E[ Aα = 0.
α

The correlation or covariance kernel of the random function Ψ is the function

 
K : M × M → R, K(p, q) = E Ψ(p)Ψ(q) .

In other words, K(p, q) computes the covariance of the random variables Ψ(p) and Ψ(q).
Note that
" #
X  X 
K(p, q) = E Aα Ψα (p) Aβ Ψβ (q)
α β
XX X
= Ψα (p)Ψβ (p) E[ Aα Aβ ] = Ψα (q)Ψα (q).
| {z } α
α,β α,β
=δαβ

4
This implies that E Ai Aj = δij , 1 ≤ i, j ≤ N , where E X denotes the expectation of the random
   

variable X.
The Riemann curvature 191

In other words, K(p, q) coincides with what we dubbed the correlator determined by the
embedding M ֒→ RN .
Given two vector fields X, Y ∈ Vect (M ) we obtain two new random functions X · Ψ
and Y · Ψ. The metric g defined by the correlator K has the probabilistic description



gp (Xp , Yp ) = E XΨ(p) · Y Ψ(p) ].

To see this it suffices to work in local coordinates (xi ) near p such that xi (p) = 0 and
assume X = ∂xi , Y = ∂xj . Then



E (XΨ(p) · Y Ψ(p)
" ! !#
X X
=E Aα ∂xi Ψα (0) Aβ ∂xi Ψβ (0)
α α
X
= ∂xi Ψα (0)∂xj Ψβ (0)E[ Aα Aβ ]
α,β
X
= ∂xi Ψα (0)∂xj Ψα (0) = ∂x2i yk K(x, y) x=y=0
.
α

The curvature formula (4.2.24) also has a probabilistic interpretation, but it is a bit
more elaborate. To describe it we need to introduce an appropriate language. Given an
m-dimensional vector space V and integers 0 ≤ p, q ≤ m we set

Λp,q V ∗ := Λp V ∗ ⊗ Λq V ∗ .

There exists a natural bilinear product

∧ : Λp0 ,q0 V ∗ × Λp1 ,q1 V ∗ → Λp0 +p1 ,q0+q1 V ∗

uniquely determined by the requirement

(α0 ⊗ β0 ) ∧ (α1 ⊗ β1 ) := (α0 ∧ α1 ) ⊗ (β0 ∧ β1 ).

For example Λ1,1 V ∗ is the space of bilinear maps V × V → R. If e1 , . . . , em is a. basis of

V , e1 , . . . , em is the dual basis of V ∗ and
X
B= bij ei ⊗ ej .
i,j

Then
!  
X X
B∧B= bik dxi ⊗ dxk ∧ bjℓ dxj ⊗ dxℓ 
ik j,ℓ
X
= bik bjℓ dxi ∧ dx ⊗ dx ∧ dxℓ ∈ Λ2,2 V ∗ .
j k

i,j,k,ℓ
192 Riemannian geometry

The Hessian of a function f on an arbitrary manifold is intrinsically well defined only

at the critical points of f . However, on a Riemann manifold (M, g) with Levi–Civita
connection ∇, we can define the Hessian of the function f to be the map

Hf : Vect (M ) × Vect (M ) → C∞(M ), Hf (X, Y ) := XY f − (∇X Y )f.

The symmetry of the Levi–Civita connection implies that Hf is indeed a symmetric tensor.
If (xi ) are normal coordinates at a point p, then in these coordinates, the Hessian of f at
p is given by the familiar formula
X
Hfp = ∂x2i xj f (p)dxi ⊗ dxj ∈ Λ1,1 Tp∗ M.
i,j

Note that X
Hfp ∧ Hfp = ∂x2i xk f (0)∂x2j xℓ f (0)dxi ∧ dxj ⊗ dxk ∧ dxℓ . (4.2.25)
i,j,k,ℓ

If we specialize to the case when g is the metric associated to the correlator

X
K(p, q) = Ψα (p)Ψα (q)
α

we deduce from (4.2.25) that

X p p
X
HΨ α
∧ HΨ α
= ∂x4i xk yj yℓ K(x, y) dxi ∧ dxj ⊗ dxk ∧ dxℓ
x=y=0
α i,j,k,ℓ
X  
= ∂x4i xk yj yℓ K(x, y) − ∂x4j xk yi yℓ K(x, y) dxi ∧ dxj ⊗ dxk ∧ dxℓ
x=y=0 x=y=0
i<j,k,ℓ
(4.2.24 ) X (4.2.1a ) X
= Rkℓij dxi ∧ dxj ⊗ dxk ∧ dxℓ = 2 Rkℓij dxi ∧ dxj ⊗ dxk ∧ dxℓ
i<j,k,ℓ i<j,k<ℓ
(4.2.1b ) X X
= 2 Rijkl dxi ∧ dxj ⊗ dxk ∧ dxℓ = −2 ωRkℓ dxk ∧ dxℓ ,
i<j,k<ℓ k<ℓ
| {z }
=:ωR

where Rk,ℓ = R(∂xk , ∂xℓ ) ∈ End− (T M ) and A 7→ ωA is the isomorphisms

End− (T M ) → Λ2 Tp∗ M, ωA (X, Y ) = gp (AX, Y )

described in (2.2.10).
Now observe that the Hessian at p of the random function Ψ is the random bilinear.
form X
p p
HΨ = Aα HΨ α
.
α

A simple computation shows that

 p p
X p p
E HΨ ∧ HΨ = H Ψα ∧ H Ψ α
α
The Riemann curvature 193

and we deduce the following equality

 p p
E HΨ ∧ HΨ = −2ωR . (4.2.26)
This is a special case of a much more general result of R. Adler and J. Taylor [2, Lemma
12.2.1]. ⊔
⊓

Example 4.2.31 (Reproducing kernels). We have seen that any immersion of a smooth
manifold into a Euclidean space produces a correlator on that manifold. It turns out that
all correlators satisfying a stronger positivity conditions are obtained in this fashion. This
follows using the concept of reproducing kernel Hilbert space that plays an important role
in varied areas of mathematics,[110].
Consider a smooth submanifold M ⊂ RN . It is equipped with a natural correlator
(4.2.20)
K : M × M → R, K(p, q) = (p, q). (4.2.27)
For any positive integer n and any p1 , . . . , pn ∈ M define the symmetric n × n matrix
 
K(p1 , p1 ) K(p1 , p2 ) · · · · · · K(p1 , pn )
 
 
 K(p2 , p1 ) K(p2 , p2 ) · · · · · · K(p2 , pn ) 
 
 
CK (p1 , . . . , pn ) :=  .
 .. .. .. .. .. 
 . . . . . 
 
 
K(pn , p1 ) K(pn , p2 ) · · · · · · K(pn , pn )
Note that if K is defined by (4.2.27) then, for any n ∈ N and any p1 , . . . , pn ∈ M the
symmetric matrix CK (p1 , . . . , pn ) is positive semidefinite.
A correlator K on a smooth manifold M is called a reproducing kernel if for any
p1 , . . . , pn ∈ M the symmetric matrix CK (p1 , . . . , pn ) is a positive semidefinite. The
reason for this terminology stems from the following remarkable reproducing property.
For each q ∈ M define
Kq ∈ C ∞ (M ), ; Kq (p) = K(p, q)
Denote by W ⊂ C ∞ (M ) the linear span of the family

Kq ; q ∈ M
Then the following hold (see [110, Chap.2]).
1. If X X
u= ui Kq i ∈ W , v = vi Kqi ∈ W,
i∈I i∈I
then the quantity X
(u, v)K := ui vi K(q i , q j )
i,j∈I
is independent of the decomposition of u and v as linear combination of functions
Kq . Moreover, it defines an inner product on W . Denote by H the completion of
W with respect to this inner product.
194 Riemannian geometry

2. For any p ∈ M , the linear functional

ev p : W → R, ev p (w) = w(p),

is continuous with continuous with respect to the norm on W induced by the inner
product (−, −)K and thus defines an element ev p in the dual H∗ of H.

3. The resulting map

M ∋ p 7→ ev p ∈ H∗
is continuous. Moreover

K(p, q) = ev p , ev q H∗
.

Since K is smooth the above map is smooth. Moreover, the condition (ii) in Definition
4.2.26 of a correlator shows that this map is also an immersion. ⊔
⊓

4.2.6 The Gauss–Bonnet theorem for oriented surfaces

We conclude this chapter with one of the most beautiful results in geometry. Its meaning
reaches deep inside the structure of a manifold and can be viewed as the origin of many
fertile ideas.
Recall one of the questions we formulated at the beginning of our study: explain
unambiguously why a sphere is “different” from a torus. This may sound like forcing our
way in through an open door since everybody can “see” they are different. Unfortunately
this is not a conclusive explanation since we can see only 3-dimensional things and possibly
there are many ways to deform a surface outside our tight 3D Universe.
The elements of Riemann geometry we discussed so far will allow us to produce an
invariant powerful enough to distinguish a sphere from a torus. But it will do more than
that.
Theorem 4.2.32 (Gauss–Bonnet Theorem. Preliminary version). Let S be a compact
oriented surface without boundary. If g0 and g1 are two Riemann metrics on S and εgi (S)
(i = 0, 1) are the corresponding Euler forms then
Z Z
εg0 (S) = εg1 (S).
S S
R
Hence the quantity S εg (S) is independent of the Riemann metric g so that it really
depends only on the topology of S!!!
The idea behind the proof is very natural. Denote by gt the metric gt = g0 + t(g1 − g0 ).
We will show Z
d
εg = 0 ∀t ∈ [0, 1].
dt S t
It is convenient to consider a more general problem.
Definition 4.2.33. Let M be a compact oriented manifold. For any Riemann metric g
on E define Z
EM (M, g) = s(g)dVg ,
M
The Riemann curvature 195

where s(g) denotes the scalar curvature of (M, g). The functional g 7→ E(g) is called the
Hilbert-Einstein functional. ⊔
⊓

We have the following remarkable result.

Lemma 4.2.34. Let M be a compact oriented manifold without boundary and g t = (gij
t )

be a 1-parameter family of Riemann metrics on M depending smoothly upon t ∈ R. Then

Z
d 1
E(gt ) = − Ricgt − s(gt )gt , ġt t dVgt , ∀t.
dt M 2

In the above formula h·, ·it denotes the inner product induced by g t on the space of sym-
metric (0,2)-tensors while the dot denotes the t-derivative.

Definition 4.2.35. A Riemann manifold (M, g) of dimension n is said to be Einstein if

the metric g satisfies Einstein’s equation

s(x)
Ricg = g,
n
where s(x) denotes the scalar curvature.

Example 4.2.36. Observe that if the Riemann metric g satisfies the condition

Ricg (x) = λ(x)g(x) (4.2.28)

for some smooth function λ ∈ C ∞ (M ), then by taking the traces of both sides of the
above equality we deduce

s(g)(x)
λ(x) = , n := dim M.
n
Thus, the Riemann manifold is Einstein if and only if it satisfies (4.2.28).
Using the computations in Example 4.2.12 we deduce that a certain constant multiple
of the Killing metric on a compact semisimple Lie group is an Einstein metric.
We refer to [16] for an in depth study of the Einstein manifolds. ⊔
⊓

The (0, 2)-tensor

1
Eij := Rij (x) − s(x)gij (x)
2
is called the Einstein tensor of (M, g).

Exercise 4.2.37. Consider a 3-dimensional Riemann manifold (M, g). Show that
s
Rijkℓ = Eik gjℓ − Eiℓ gjk + Ejℓ gik − Ejk giℓ + (giℓ gjk − gik gjℓ ).
2
In particular, this shows that on a Riemann 3-manifold the full Riemann tensor is com-
pletely determined by the Einstein tensor. ⊔
⊓
196 Riemannian geometry

Exercise 4.2.38. (Schouten-Struik, [117]). Prove that the scalar curvature of an

Einstein manifold of dimension ≥ 3 is constant.
Hint: Use the 2nd Bianchi identity. ⊔
⊓

Notice that when (S, g) is a compact oriented Riemann surface two very nice things
happen.
(i) (S, g) is Einstein.
R (Recall that only R1212 is nontrivial).
(ii) E(g) = 2 S εg .
Theorem 4.2.32 is thus an immediate consequence of Lemma 4.2.34.
Proof of the lemma We will produce a very explicit description of the integrand
d  d
d

s(gt )dVgt = s(gt ) dVgt + s(gt ) dVgt (4.2.29)
dt dt dt
d
of dt E(g t ). We will adopt a “roll up your sleeves, and just do it” strategy reminiscent to
the good old days of the tensor calculus frenzy. By this we mean that we will work in a
nicely chosen collection of local coordinates, and we will keep track of the zillion indices
we will encounter. As we will see, the computations are not as hopeless as they may seem
to be.
We will study the integrand (4.2.29) at t = 0. The general case is entirely analogous.
For typographical reasons we will be forced to introduce new notations. Thus, ĝ will
denote (gt ) for t = 0, while g t will be denoted simply by g. A hat over a quantity means
we think of that quantity at t = 0, while a dot means differentiation with respect to t at
t = 0.
Let q be an arbitrary point on S, and denote by (x1 , . . . , xn ) a collection of ĝ-normal
coordinates at q. Denote by ∇ the Levi–Civita connection of g and let Γijk denote its
Christoffel symbols in the coordinates (xi ).
Many nice things happen at q, and we list a few of them which will be used later.

ĝij = ĝij = δij , ∂k ĝij = 0. (4.2.30)

ˆ i ∂j = 0, Γ̂i = 0.
∇ (4.2.31)
jk

If α = αi dxi is a 1-form then, at q,

X
δĝ α = ∂i αi . where δ = ∗d ∗ . (4.2.32)
i

In particular, for any smooth function u we have

X
(∆M,ĝ u)(q) = − ∂i 2 u. (4.2.33)
i

Set
h = (hij ) := (ġ) = (ġij ).
The tensor h is a symmetric (0, 2)-tensor. Its ĝ-trace is the scalar

trĝ h = ĝij hij = tr L−1 (h),

The Riemann curvature 197

where L is the lowering the indices operator defined by ĝ. In particular, at q

X
trĝ h = hii . (4.2.34)
i

The curvature of g is given by

ℓ ℓ
Rikj = −Rijk = ∂k Γℓij − ∂j Γℓik + Γℓmk Γm ℓ m
ij − Γmj Γik .

The Ricci tensor is

k
Rij = Rikj = ∂k Γkij − ∂j Γkik + Γkmk Γm k m
ij − Γmj Γik .

Finally, the scalar curvature is

 
s = trg Rij = g ij Rij = gij ∂k Γkij − ∂j Γkik + Γkmk Γm
ij − Γ k m
mj ik .
Γ

Differentiating s at t = 0, and then evaluating at q we obtain

   
ṡ = ġij ∂k Γ̂kij − ∂j Γ̂kik + δij ∂k Γ̇kij − ∂j Γ̇kik
X  (4.2.35)
= ġ ij R̂ij + ∂k Γ̇kii − ∂i Γ̇kik .
i

The term ġ ij can be computed by derivating the equality gik gjk = δki at t = 0. We get

ġik ĝjk + ĝik hjk = 0,

so that
ġij = −hij . (4.2.36)
To evaluate the derivatives Γ̇’s we use the known formulæ
1 km
Γm
ij = g (∂i gjk − ∂k gij + ∂j gik ) ,
2
which, upon differentiation at t = 0, yield
1 1
Γ̇m
ij = (∂i ĝjk − ∂k ĝij + ∂j ĝik ) + ĝkm (∂i hjk − ∂k hij + ∂j hik )
2 2 (4.2.37)
1
= (∂i hjm − ∂m hij + ∂j him ) .
2
We substitute (4.2.36) -(4.2.37) in (4.2.35), and we get, at q

X 1X 1X 2
ṡ = − hij R̂ij + (∂k ∂i hik − ∂k 2 hii + ∂k ∂i hik ) − (∂i hkk − ∂i ∂k hik )
2 2
i,j i,k i,k
X X X
=− hij R̂ij − ∂i 2 hkk + ∂i ∂k hik
i,j i,k i,k
X
d ġiĝ + ∆M,ĝ trĝ ġ +
= −hRic, ∂i ∂k hik . (4.2.38)
i,k
198 Riemannian geometry

To get a coordinate free description of the last term note that, at q,

ˆ k h)(∂i , ∂m ) = ∂k him .
(∇
ˆ is a (0, 3)-tensor. Using the ĝ-trace we can construct a
The total covariant derivative ∇h
(0, 1)-tensor, i.e., a 1-form
ˆ = tr(L−1 ∇h),
trĝ (∇h) ˆ
ĝ

where L−1 i
ĝ is the raising the indices operator defined by ĝ. In the local coordinates (x )
we have
ˆ = ĝ ij (∇
trĝ (∇h) ˆ i h)jk dxk .
Using (4.2.30), and (4.2.32) we deduce that the last term in (4.2.38) can be rewritten (at
q) as
ˆ = δtrĝ (∇
δtrĝ (∇h) ˆ ġ).
We have thus established that
d ġiĝ + ∆M,ĝ trĝ ġ + δtrĝ (∇
ṡ = −hRic, ˆ ġ). (4.2.39)

The second term of the integrand (4.2.29) is a lot easier to compute.

p
dVg = ± |g|dx1 ∧ · · · ∧ dxn ,

so that
1 d
dV̇g = ± |ĝ|−1/2 |g|dx1 ∧ · · · ∧ dxn .
2 dt
At q the metric is Euclidean, ĝij = δij , and
d X
|g| = ġii = |ĝ| · trĝ (ġ) = h ĝ, ġ iĝ |ĝ|.
dt
i

Hence
Z * ! + Z  
1 ˆ ġ) dVĝ .
Ė(g) = s(ĝ)ĝ − Ricĝ , ġ dVĝ + ∆M,ĝ trĝ ġ + δtrĝ (∇
M 2 M
ĝ

Green’s formula shows the last two terms vanish and the proof of the Lemma is concluded.
⊔
⊓

Definition 4.2.39. Let S be a compact, oriented surface without boundary. We define

its Euler characteristic as the number
Z
1
χ(S) = ε(g),
2π S
where g is an arbitrary Riemann metric on S. The number
1
g(S) = (2 − χ(S))
2
is called the genus of the surface. ⊔
⊓
The Riemann curvature 199

Remark 4.2.40. According to the theorem we have just proved, the Euler characteristic is
independent of the metric used to define it. Hence, the Euler characteristic is a topological
invariant of the surface. The reason for this terminology will become apparent when we
discuss DeRham cohomology, a Z-graded vector space naturally associated to a surface
whose Euler characteristic coincides with the number defined above. So far we have no
idea whether χ(S) is even an integer. ⊔
⊓

Proposition 4.2.41.
χ(S 2 ) = 2 and χ(T 2 ) = 0.

Proof. To compute χ(S 2 ) we use the round metric g0 for which K = 1 so that
Z
2 1 1
χ(S ) = dvg0 = areag0 (S 2 ) = 2.
2π S2 2π

To compute the Euler characteristic of the torus we think of it as an Abelian Lie group with
a bi-invariant metric. Since the Lie bracket is trivial we deduce from the computations in
Subsection 4.2.2 that its curvature is zero. This concludes the proof of the proposition. ⊔⊓

Proposition 4.2.42. If Si (i=1,2) are two compact oriented surfaces without boundary
then
χ(S1 #S2 ) = χ(S1 ) + χ(S2 ) − 2.

Thus upon iteration we get

k
X
χ(S1 # · · · #Sk ) = χ(Si ) − 2(k − 1),
i=1

for any compact oriented surfaces S1 , . . . , Sk . In terms of genera, the last equality can be
rephrased as
Xk
g(S1 # · · · #Sk ) = g(Si ).
i=1

In the proof of this proposition we will use special metrics on connected sums of surfaces
which require a preliminary analytical discussion.
Consider f : (−4, 4) → (0, ∞) a smooth, even function such that

1. f (x) = 1 for |x| ≤ 2.

p
2. f (x) = 1 − (x + 3)2 for x ∈ [−4, −3.5].
p
3. f (x) = 1 − (x − 3)2 for x ∈ [3.5, 4].

4. f is non-decreasing on [−4, 0].

200 Riemannian geometry

S
f

- +
S S
f f

Figure 4.3: Generating a hot-dog-shaped surface

S
1
D1

S2

D
2

Figure 4.4: Special metric on a connected sum

The Riemann curvature 201

One such function is graphed in Figure 4.3

Denote by Sf the surface inside R3 obtained by rotating the graph of f about the x-
axis. Because of properties (i)-(iv), Sf is a smooth surface diffeomorphic5 to S 2 . We denote
by g the metric on Sf induced by the Euclidean metric in R3 . Since Sf is diffeomorphic
to a sphere Z
Kg dVg = 2πχ(S 2 ) = 4π.
Sf

Set
Sf± := Sf ∩ {±x > 0}, Sf±1 := Sf ∩ {±x > 1}
Since f is even we deduce
Z Z
1
Kg dvg = Kg dvg = 2π. (4.2.40)
±
Sf 2 Sf

On the other hand, on the neck C = {|x| ≤ 2} the metric g is locally Euclidean g =
dx2 + dθ 2 , so that over this region Kg = 0. Hence
Z
Kg dvg = 0. (4.2.41)
C

Proof of the Proposition 4.2.42 Let Di ⊂ Si (i = 1, 2) be a local coordinate neigh-

borhood diffeomorphic with a disk in the plane. Pick a metric gi on Si such that (D1 , g1 ) is
isometric with Sf+ and (D2 , g2 ) is isometric to Sf− . The connected sum S1 #S2 is obtained
by chopping off the regions Sf1 from D1 and Sf−1 from D2 and (isometrically) identifying
the remaining cylinders Sf± ∩ {|x| ≤ 1} = C and call O the overlap created by gluing (see
Figure 4.4). Denote the metric thus obtained on S1 #S2 by ĝ. We can now compute
Z
1
χ(S1 #S2 ) = Kĝ dvĝ
2π S1 #S2
Z Z Z
1 1 1
= Kg dVg1 + Kg dVg2 + Kg dVg
2π S1 \D1 1 2π S2 \D2 2 2π O
Z Z
(4.2.41) 1 1
= Kg dVg1 − Kg dVg
2π S1 1 2π D1
Z Z
1 1 (4.2.40)
+ Kg2 dVg2 − Kg dVg = χ(S1 ) + χ(S2 ) − 2.
2π S2 2π D2

This completes the proof of the proposition. ⊔

⊓

Corollary 4.2.43 (Gauss–Bonnet). Let Σg denote the connected sum of g-tori. (By
definition Σ0 = S 2 . Then

χ(Σg ) = 2 − 2g and g(Σg ) = g.

In particular, a sphere is not diffeomorphic to a torus.

5
One such diffeomorphism can be explicitly constructed projecting along radii starting at the origin.
202 Riemannian geometry

Remark 4.2.44. It is a classical result that the only compact oriented surfaces are the
connected sums of g-tori (see [90]), so that the genus of a compact oriented surface is a
complete topological invariant. ⊔
⊓
Chapter 5

Elements of the Calculus of

Variations

This is a very exciting subject lieing at the frontier between mathematics and physics. The
limited space we will devote to this subject will hardly do it justice, and we will barely
touch its physical significance. We recommend to anyone looking for an intellectual feast
the Chapter 16 in vol.2 of “The Feynmann Lectures on Physics” [47], which in our opinion
is the most eloquent argument for the raison d’être of the calculus of variations.

5.1 The least action principle

5.1.1 The 1-dimensional Euler-Lagrange equations
From a very “dry” point of view, the fundamental problem of the calculus of variations
can be easily formulated as follows.
Consider a smooth manifold M , and let L : R × T M → R by a smooth function
called the lagrangian. Fix two points p0 , p1 ∈ M . The action of a piecewise smooth path
γ : [0, 1] → M connecting these points is the real number S(γ) = SL (γ) defined by
Z 1
S(γ) = SL (γ) := L(t, γ̇(t), γ(t))dt.
0
In the calculus of variations one is interested in those paths as above with minimal action.
Example 5.1.1. Given a smooth function U : R3 → R called the potential, we can form
the lagrangian
L(q̇, q) : R3 × R3 ∼
= T R3 → R,
given by
1
L = Q − U = kinetic energy − potential energy = m|q̇|2 − U (q).
2
The scalar m is called the mass. The action of a path (trajectory) γ : [0, 1] → R3 is a
quantity called the Newtonian action.Note that, as a physical quantity, the Newtonian
action is measured in the same units as the energy. ⊔
⊓

203
204 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

Example 5.1.2. To any Riemann manifold (M, g) one can naturally associate two la-
grangians L1 , L2 : T M → R defined by

L1 (v, q) = gq (v, v)1/2 (v ∈ Tq M ),

and
1
L2 (v, q) = gq (v, v).
2
We see that the action defined by L1 coincides with the length of a path. The action
defined by L2 is called the energy of a path. ⊔
⊓

Before we present the main result of this subsection we need to introduce a bit of
notation.
Tangent bundles are very peculiar manifolds. Any collection (q 1 , . . . , q n ) of local co-
ordinates on a smooth manifold M automatically induces local coordinates on T M . Any
point in T M can be described by a pair (v, q), where q ∈ M , v ∈ Tq M . Furthermore, v
has a decomposition
∂
v = v i ∂i , where ∂i := i .
∂q
We set q̇ i := v i so that
v = q̇ i ∂i .
The collection (q̇ 1 , . . . , q̇ n ; q 1 , . . . , q n ) defines local coordinates on T M . These are said to
be holonomic local coordinates on T M . This will be the only type of local coordinates we
will ever use.
Theorem 5.1.3 (The least action principle). Let L : R × T M → R be a lagrangian, and
p0 , p1 ∈ M two fixed points. Suppose γ : [0, 1] → M is a smooth path such that the
following hold.

(i) γ(i) = pi , i = 0, 1.

(ii) SL (γ) ≤ SL (γ̃), for any smooth path γ̃ : [0, 1] → M joining p0 to p1 .

Then the path γ satisfies the Euler-Lagrange equations

d ∂ ∂
L(t, γ̇, γ) = L(t, γ̇, γ).
dt ∂ γ̇ ∂γ

More precisely, if (q̇ j , q i ) are holonomic local coordinates on T M such that γ(t) = (q i (t)),
and γ̇ = (q̇ j (t)), then γ is a solution of the system of nonlinear ordinary differential
equations
d ∂L ∂L
k
(t, q̇ j , q i ) = k (t, q̇ j , q i ), k = 1, . . . , n = dim M.
dt ∂ q̇ ∂q

Definition 5.1.4. A path γ : [0, 1] → M satisfying the Euler-Lagrange equations with

respect to some lagrangian L is said to be an extremal of L. ⊔
⊓
5.1. THE LEAST ACTION PRINCIPLE 205

To get a better feeling of these equations consider the special case discussed in Example
5.1.1
1
L = m|q̇|2 − U (q).
2
Then
∂ ∂
L = mq̇, L = −∇U (q),
∂ q̇ ∂q
and the Euler-Lagrange equations become

mq̈ = −∇U (q). (5.1.1)

These are precisely Newton’s equation of the motion of a particle of mass m in the force
field −∇U generated by the potential U .
In the proof of the least action principle we will use the notion of variation of a path.
Definition 5.1.5. Let γ : [0, 1] → M be a smooth path. A variation of γ is a smooth
map
α = αs (t) : (−ε, ε) × [0, 1] → M,
such that α0 (t) = γ(t). If moreover, αs (i) = pi ∀s, i = 0, 1, then we say that α is a
variation rel endpoints. ⊔
⊓

Proof of Theorem 5.1.3. Let αs be a variation of γ rel endpoints. Then

SL (α0 ) ≤ SL (αs ) ∀s,

so that
d
|s=0 SL (αs ) = 0.
ds
Assume for simplicity that the image of γ is entirely contained in some open coordinate
neighborhood U with coordinates (q 1 , . . . , q n ). Then, for very small |s|, we can write

dαs
αs (t) = (q i (s, t)) and = (q̇ i (s, t)).
dt
Following the tradition, we set

∂α ∂q i ∂ dαs ∂ q̇ j ∂
δα := |s=0 = ∂i δα̇ := |s=0 = .
∂s ∂s ∂s dt ∂s ∂ q̇ j

The quantity δα is a vector field along γ called infinitesimal variation (see Figure 5.1). In
fact, the pair (δα; δα̇) ∈ T (T M ) is a vector field along t 7→ (γ(t), γ̇(t)) ∈ T M . Note that
d
δα̇ = dt δα, and at endpoints δα = 0.
Exercise 5.1.6. Prove that if t 7→ X(t) ∈ Tγ(t) M is a smooth vector field along γ, such
that X(t) = 0 for t = 0, 1 then there exists at least one variation rel endpoints α such that
δα = X.
Hint: Use the exponential map of some Riemann metric on M . ⊔
⊓
206 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

γ
p
q

Figure 5.1: Deforming a path rel endpoints

We compute (at s = 0)
Z 1 Z 1 Z 1
d d ∂L i ∂L j
0 = SL (αs ) = L(t, α˙s , αs ) = δα dt + δα̇ dt. (5.1.2)
ds ds 0 0 ∂q i 0 ∂ q̇ j s
Integrating by parts in the second term in the right-hand-side we deduce
Z 1  
∂L d ∂L
i
− i
δαi dt. (5.1.3)
0 ∂q dt ∂ q̇
The last equality holds for any variation α. From Exercise 5.1.6 we deduce that it holds
for any vector field δαi ∂i along γ. At this point we use the following classical result of
analysis.
If f (t) is a continuous function on [0,1] such that
Z 1
f (t)g(t)dt = 0 ∀g ∈ C0∞ (0, 1),
0

then f is identically zero.

Using this result in (5.1.3) we deduce the desired conclusion. ⊔
⊓

Remark 5.1.7. (a) In the proof of the least action principle we used a simplifying assump-
tion, namely that the image of γ lies in a coordinate neighborhood. This is true locally,
and for the above arguments to work it suffices to choose only a special type of variations,
localized on small intervals of [0, 1]. In terms of infinitesimal variations this means we
need to look only at vector fields along γ supported in local coordinate neighborhoods.
We leave the reader fill in the details.
(b) The Euler-Lagrange equations were described using holonomic local coordinates.
The minimizers of the action, if any, are objects independent of any choice of local coordi-
nates, so that the Euler-Lagrange equations have to be independent of such choices. We
check this directly.
5.1. THE LEAST ACTION PRINCIPLE 207

If (xi ) is another collection of local coordinates on M and (ẋj , xi ) are the coordinates
induced on T M , then we have the transition rules

∂xj k
xi = xi (q 1 , . . . , q n ), ẋj = q̇ ,
∂q k

so that
∂ ∂xj ∂ ∂ 2 xj k ∂
= + q̇
∂q i ∂q i ∂xj ∂q k ∂q i ∂ ẋj
∂ ∂ ẋi ∂ ∂xj ∂
= =
∂ q̇ j ∂ q̇ j ∂ ẋj ∂q i ∂ ẋj
Then
∂L ∂xj ∂L ∂ 2 xj k ∂L
= + q̇ .
∂q i ∂q i ∂xj ∂q k ∂q i ∂ ẋj
     
d ∂L d ∂xj ∂L ∂ 2 xj k ∂L ∂xj d ∂L
= = k i q̇ + i .
dt ∂ q̇ i dt ∂q i ∂ ẋj ∂q ∂q ∂ ẋj ∂q dt ∂ ẋj
We now see that the Euler-Lagrange equations in the q-variables imply the Euler-Lagrange
in the x-variable, i.e., these equations are independent of coordinates.
The æsthetically conscious reader may object to the way we chose to present the
Euler-Lagrange equations. These are intrinsic equations we formulated in a coordinate
dependent fashion. Is there any way of writing these equation so that the intrinsic nature
is visible “on the nose”?
If the lagrangian L satisfies certain nondegeneracy conditions there are two ways of
achieving this goal. One method is to consider a natural nonlinear connection ∇L on T M
as in [100] . The Euler-Lagrange equations for an extremal γ(t) can then be rewritten as
a “geodesics equation”
∇Lγ̇ γ̇.

The example below will illustrate this approach on a very special case when L is the
lagrangian L2 defined in Example 5.1.2, in which the extremals will turn out to be precisely
the geodesics on a Riemann manifold.
Another far reaching method of globalizing the formulation of the Euler-Lagrange
equation is through the Legendre transform, which again requires a nondegeneracy condi-
tion on the lagrangian. Via the Legendre transform the Euler-Lagrange equations become
a system of first order equations on the cotangent bundle T ∗ M known as Hamilton equa-
tions.
These equations have the advantage that can be formulated on manifolds more general
than the cotangent bundles, namely on symplectic manifolds. These are manifolds carrying
a closed 2-form whose restriction to each tangent space defines a symplectic duality (see
Subsection 2.2.4.)
Much like the geodesics equations on a Riemann manifold, the Hamilton equations
carry a lot of information about the structure of symplectic manifolds, and they are
currently the focus of very intense research. For more details and examples we refer to
the monographs [7, 38]. ⊔
⊓
208 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

Example 5.1.8. Let (M, g) be a Riemann manifold. We will compute the Euler-Lagrange
equations for the lagrangians L1 , L2 in Example 5.1.2.
1
L2 (q̇, q) = gij (q)q̇ i q̇ j ,
2
so that
∂L2 ∂L2 1 ∂gij i j
= gjk q̇ j = q̇ q̇ .
∂ q̇ k ∂q k 2 ∂q k
The Euler-Lagrange equations are
∂gjk i j 1 ∂gij i j
q̈ j gjk + i
q̇ q̇ = q̇ q̇ . (5.1.4)
∂q 2 ∂q k
Since g km gjm = δjm , we get
 
m km ∂gjk 1 ∂gij
q̈ + g i
− q̇ i q̇ j = 0. (5.1.5)
∂q 2 ∂q k
When we derivate with respect to t the equality gik q̇ i = gjk q̇ j we deduce
 
km ∂gjk i j 1 km ∂gik ∂gjk
g q̇ q̇ = g + q̇ i q̇ j .
∂q i 2 ∂q j ∂q i
We substitute this equality in (5.1.5), and we get
 
m 1 km ∂gik ∂gjk ∂gij
q̈ + g + − q̇ i q̇ j = 0. (5.1.6)
2 ∂q j ∂q i ∂q k
The coefficient of q̇ i q̇ j in (5.1.6) is none other than the Christoffel symbol Γm
ij so this
equation is precisely the geodesic equation. ⊔
⊓

Example 5.1.9. Consider now the lagrangian L1 (q̇, q) = (gij q̇ i q̇ j )1/2 . Note that the
action Z p1


SL q(t) = L(q̇, q)dt,
p0
is independent of the parametrization t 7→ q(t) since it computes the length of the path
t 7→ q(t). Thus, when we express the Euler-Lagrange equations for a minimizer γ0 of
this action, we may as well assume it is parameterized by arclength, i.e., |γ̇0 | = 1. The
Euler-Lagrange equations for L1 are
∂g
ij i j
d gkj q̇ j ∂q k
q̇ q̇
p = p .
dt gij q̇ i q̇ j 2 gij q̇ i q̇ j
Along the extremal we have gij q̇ i q̇ j = 1 (arclength parametrization) so that the previous
equations can be rewritten as
d
1 ∂gij i j
gkj q̇ j = q̇ q̇ .
dt 2 ∂q k
We recognize here the equation (5.1.4) which, as we have seen, is the geodesic equation in
disguise. This fact almost explains why the geodesics are the shortest paths between two
nearby points. ⊔
⊓
5.1. THE LEAST ACTION PRINCIPLE 209

5.1.2 Noether’s conservation principle

his subsection is intended to offer the reader a glimpse at a fascinating subject touching
both physics and geometry. We first need to introduce a bit of traditional terminology
commonly used by physicists.
Consider a smooth manifold M . The tangent bundle T M is usually referred to as the
space of states or the lagrangian phase space. A point in T M is said to be a state. A
lagrangian L : R × T M → R associates to each state several meaningful quantities.
∂L
• The generalized momenta: pi = ∂ q̇ i
.

• The energy: H = pi q̇ i − L.
∂L
• The generalized force: F = ∂q i
.

This terminology can be justified by looking at the lagrangian of a classical particle in

a potential force field, F = −∇U ,
1
L = m|q̇|2 − U (q).
2
The momenta associated to this lagrangian are the usual kinetic momenta of the Newto-
nian mechanics
pi = mq̇ i ,
while H is simply the total energy
1
H= m|q̇|2 + U (q).
2
It will be convenient to think of an extremal for an arbitrary lagrangian L(t, q̇, q) as
describing the motion of a particle under the influence of the generalized force.
Proposition 5.1.10 (Conservation of energy). Let γ(t) be an extremal of a time inde-
pendent lagrangian L = L(q̇, q). Then the energy is conserved along γ, i.e.,
d
H(γ, γ̇) = 0.
dt
Proof. By direct computation we get
 
d d i d ∂L ∂L ∂L ∂L
H(γ, γ̇) = (pi q̇ − L) = i
q̇ i + i q̈ i − i q̇ i − i q̈ i
dt dt dt ∂ q̇ ∂ q̇ ∂q ∂ q̇
   
d ∂L ∂L
= − q̈ i = 0 (by Euler − Lagrange). ⊔
⊓
dt ∂ q̇ i ∂q i
At the beginning of the 20th century (1918), Emmy Noether discovered that many of the
conservation laws of the classical mechanics had a geometric origin: they were, most of
them, reflecting a symmetry of the lagrangian!!!
This became a driving principle in the search for conservation laws, and in fact, con-
servation became almost synonymous with symmetry. It eased the leap from classical to
210 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

quantum mechanics, and one can say it is a very important building block of quantum
physics in general. In the few instances of conservation laws where the symmetry was not
apparent the conservation was always “blamed” on a “hidden symmetry”. What is then
this Noether principle?
To answer this question we need to make some simple observations.
Let X be a vector field on a smooth manifold M defining a global flow Φs . This flow
induces a flow Ψs on the tangent bundle T M defined by


Ψs (v, x) = Φs∗ (v), Φs (x) .

One can think of Ψs as defining an action of the additive group R on T M . Alterna-

tively, physicists say that X is an infinitesimal symmetry of the given mechanical system
described by the lagrangian L.
Example 5.1.11. Let M be the unit round sphere S 2 ⊂ R3 . The rotations about the
∂
z-axis define a 1-parameter group of isometries of S 2 generated by ∂θ , where θ is the
2
longitude on S . ⊔
⊓

Definition 5.1.12. Let L be a lagrangian on T M , and X a vector field on M . The

lagrangian L is said to be X- invariant if

L ◦ Ψs = L, ∀s. ⊔
⊓

Denote by X ∈ Vect (T M ) the infinitesimal generator of Ψs , and by LX the Lie deriva-

tive on T M along X. We see that L is X-invariant if and only if

LX L = 0.

We describe this derivative using the local coordinates (q̇ j , q i ). Set

q̇ j (s), q i (s) := Ψs (q̇ j , q i ).

Then
d
|s=0 q i (s) = X k δki .
ds
d
To compute ds |s=0 q̇ j (s) ∂q∂ j we use the definition of the Lie derivative on M
 
d ∂ ∂ ∂ q̇ j k ∂X
j ∂ i ∂X
j ∂
− q̇ j j = LX (q̇ i i ) = X k
− q̇ = − q̇ ,
ds ∂q ∂q ∂q k ∂q k ∂q j ∂q i ∂q j

since ∂ q̇ j /∂q i = 0 on T M . Hence

∂ j ∂
k ∂X
X = Xi + q̇ .
∂q i ∂q k ∂ q̇ j
Corollary 5.1.13. The lagrangian L is X-invariant if and only if
∂L j
k ∂X ∂L
Xi + q̇ = 0. (5.1.7)
∂q i ∂q k ∂ q̇ j
⊔
⊓
5.1. THE LEAST ACTION PRINCIPLE 211

y
O

Figure 5.2: A surface of revolution

Theorem 5.1.14 (E. Noether). If the lagrangian L is X-invariant, then the quantity

∂L
PX = X i = X i pi
∂ q̇ i

is conserved along the extremals of L.

Proof. Consider an extremal γ = γ(q i (t)) of L. We compute

   
d d i ∂L ∂X i k ∂L i d ∂L
PX (γ, γ̇) = X (γ(t)) i = q̇ +X
dt dt ∂ q̇ ∂q k ∂ q̇ i dt ∂ q̇ i

Euler-Lagrange ∂X i k ∂L ∂L (5.1.7)
= k
q̇ i
+ X i i = 0. ⊔
⊓
∂q ∂ q̇ ∂q
The classical conservation-of-momentum law is a special consequence of Noether’s theo-
rem.
∂L
Corollary 5.1.15. Consider a lagrangian L = L(t, q̇, q, ) on Rn . If ∂q i
= 0, i.e., the i-th
dpi
component of the force is zero, then dt = 0 along any extremal, i.e., the i-th component
of the momentum is conserved).
∂
Proof. Take X = ∂q i
in Noether’s conservation law. ⊔
⊓

The conservation of momentum has an interesting application in the study of geodesics.

Example 5.1.16. (Geodesics on surfaces of revolution). Consider a surface of
revolution S in R3 obtained by rotating about the z-axis the curve y = f (z) situated in
the yz plane. If we use cylindrical coordinates (r, θ, z) we can describe S as r = f (z).
In these coordinates, the Euclidean metric in R3 has the form

ds2 = dr 2 + dz 2 + r 2 dθ 2 .
212 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

We can choose (z, θ) as local coordinates on S, then the induced metric has the form

gS = {1 + (f ′ (z))2 }dz 2 + f 2 (z)dθ 2 = A(z)dz 2 + r 2 dθ 2 , r = f (z).

The lagrangian defining the geodesics on S is

1 2 
L= Aż + r 2 θ̇ 2 .
2
∂L
We see that L is independent of θ: ∂θ = 0, so that the generalized momentum
∂L
= r 2 θ̇
∂ θ̇
is conserved along the geodesics. ⊔
⊓

This fact can be given a nice geometric interpretation. Consider a geodesic γ(t) =
(z(t), θ(t)) on our surface of revolution S. The angle φ between the geodesic and the
parallels z = const can be determined from the equality

hγ̇, ∂θ i r 2 θ̇
cos φ = = ,
|γ̇| · |∂θ | r|γ̇|

i.e., r cos φ = r 2 θ̇|γ̇|−1 . The conservation of energy implies that |γ̇|2 = 2L = H is constant
along the geodesics. We deduce the following classical result.
Theorem 5.1.17 (Clairaut). On a a surface of revolution r = f (z) the quantity r cos φ
is constant along any geodesic, where φ ∈ (−π, π) is the oriented angle the geodesic makes
with the parallels z = const. ⊔
⊓

Exercise 5.1.18. Describe the geodesics on the round sphere S 2 , and on the cylinder
{x2 + y 2 = 1} ⊂ R3 . ⊔
⊓

5.2 The variational theory of geodesics

We have seen that the paths of minimal length between two points on a Riemann manifold
are necessarily geodesics.
However, given a geodesic joining two points q0 , q1 it may happen that it is not a
minimal path. This should be compared with the situation in calculus, when a critical
point of a function f may not be a minimum or a maximum.
To decide this issue one has to look at the second derivative. This is precisely what
we intend to do in the case of geodesics. This situation is a bit more complicated since
the action functional Z
1
S= |γ̇|2 dt
2
is not defined on a finite dimensional manifold. It is a function defined on the “space of all
paths” joining the two given points. With some extra effort this space can be organized
as an infinite dimensional manifold. We will not attempt to formalize these prescriptions,
but rather follow the ad-hoc, intuitive approach of [94].
5.2. THE VARIATIONAL THEORY OF GEODESICS 213

5.2.1 Variational formulæ

Let M be a connected Riemann manifold, and consider p, q ∈ M . Denote by Ωp,q =
Ωp,q (M ) the space of all continuous, piecewise smooth paths γ : [0, 1] → M connecting p
to q.
An infinitesimal variation of a path γ ∈ Ωp,q is a continuous, piecewise smooth
vector field V along γ such that V (0) = 0 and V (1) = 0 and

lim V (t ± h)
hց0

exists for every t ∈ [0, 1], they are vectors V (t)± ∈ Tγ (t)M , and V (t)+ = V ( t)− for all
but finitely many t-s. The space of infinitesimal variations of γ is an infinite dimensional
linear space denoted by Tγ = Tγ Ωp,q .
Definition 5.2.1. Let γ ∈ Ωp,q . A variation of γ is a continuous map

α = αs (t) : (−ε, ε) × [0, 1] → M

such that

1. ∀s ∈ (−ε, ε), αs ∈ Ωp,q .

2. There exists a partition 0 = t0 < t1 · · · < tk−1 < tk = 1 of [0,1] such that the
restriction of α to each (−ε, ε) × (ti−1 , ti ) is a smooth map. ⊔
⊓

Every variation α of γ defines an infinitesimal variation

∂αs
δα := |s=0 .
∂s
Exercise 5.2.2. Given V ∈ Tγ construct a variation α such that δα = V . ⊔
⊓

Consider now the energy functional

Z 1
1
E : Ωp,q → R, E(γ) = |γ̇(t)|2 dt.
2 0

Fix γ ∈ Ωp,q , and let α be a variation of γ. The velocity γ̇(t) has a finite number of
discontinuities, so that the quantity

∆t γ̇ = lim (γ̇(t + h) − γ̇(t − h))

h→0+

is nonzero only for finitely many t’s.

Theorem 5.2.3 (The first variation formula).
X Z 1
d
E∗ (δα) := |s=0 E(αs ) = − h(δα)(t), ∆t γ̇i − hδα, ∇ d γ̇idt, (5.2.1)
ds t 0 dt

where ∇ denotes the Levi–Civita connection. (Note that the right-hand-side depends on
α only through δα so it is really a linear function on Tγ .)
214 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

∂αs
Proof. Set α̇s = ∂t . We differentiate under the integral sign using the equality

∂
|α̇s |2 = 2h∇ ∂ α̇s , α̇s i,
∂s ∂s

and we get Z 1
d
|s=0 E(αs ) = h∇ ∂ α̇s , α̇s i |s=0 dt.
ds 0 ∂s

∂ ∂
Since the vector fields ∂s and ∂t commute we have ∇ ∂ ∂α
∂t = ∇ ∂t∂
∂α
∂s .
∂s
Let 0 = t0 < t2 < · · · < tk = 1 be a partition of [0,1] as in Definition 5.2.1. Since
αs = γ for s = 0 we conclude
k Z
X ti
E∗ (δα) = h∇ ∂ δα, γ̇i.
∂t
i=1 ti−1

We use the equality

∂
hδα, γ̇i = h∇ ∂ δα, γ̇i + hδα, ∇ ∂ γ̇i
∂t ∂t ∂t

to integrate by parts, and we obtain

k
X k Z
X ti
ti
E∗ (δα) = hδα, γ̇i ti−1 − hδα, ∇ ∂ γ̇idt.
∂t
i=1 i=1 ti−1

This is precisely equality (5.2.1). ⊔

⊓

Definition 5.2.4. A path γ ∈ Ωp,q is called critical if

E∗ (V ) = 0, ∀V ∈ Tγ . ⊔
⊓

Corollary 5.2.5. A path γ ∈ Ωp,q is critical if and only if it is a geodesic. ⊔

⊓

Exercise 5.2.6. Prove the above corollary. ⊔

⊓

Remark 5.2.7. Note that, a priori, a critical path may have a discontinuous first deriva-
tive. The above corollary shows that this is not the case: the criticality also implies
smoothness. This is a manifestation of a more general analytical phenomenon called el-
liptic regularity. We will have more to say about it in Chapter 11. ⊔
⊓

The map E∗ : Tγ → R, δα 7→ E∗ (δα) is called the first derivative of E at γ ∈ Ωp,q .

We want to define a second derivative of E in order to address the issue raised at the
beginning of this section. We will imitate the finite dimensional case which we now briefly
analyze.
Let f : X → R be a smooth function on the finite dimensional smooth manifold X. If
x0 is a critical point of f , i.e., df (x0 ) = 0, then we can define the Hessian at x0

f∗∗ : Tx0 X × Tx0 X → R

5.2. THE VARIATIONAL THEORY OF GEODESICS 215

as follows. Given V1 , V2 ∈ Tx0 X, consider a smooth map (s1 , s2 ) 7→ α(s1 , s2 ) ∈ X such

that
∂α
α(0, 0) = x0 and (0, 0) = Vi , i = 1, 2. (5.2.2)
∂si
Now set
∂ 2 f ( α(s1 , s2 ) )
f∗∗ (V1 , V2 ) = |(0,0) .
∂s1 ∂s2
Note that since x0 is a critical point of f , the Hessian f∗∗ (V1 , V2 ) is independent of the
function α satisfying (5.2.2).
We now return to our energy functional E : Ωp,q → R. Let γ ∈ Ωp,q be a critical path.
Consider a 2-parameter variation of γ

α := αs1 ,s2 : (−ε, ε) × (−ε, ε) × [0, 1] → M, (s1 , s2 , t) 7→ αs1 ,s2 (t).

Set U := (−ε, ε) × (−ε, ε) ⊂ R2 and γ := α0,0 . The map α is continuous, and has
second order derivatives everywhere except maybe on finitely many “coordinate” planes
∂α
si = const, or t = const. Set δi α := ∂s |
i (0,0)
, i = 1, 2. Note that δi α ∈ Tγ .
Exercise 5.2.8. Given V1 , V2 ∈ Tγ construct a 2-parameter variation α such that Vi = δi α.
⊔
⊓

We can now define the Hessian of E at γ by

∂ 2 E(αs1 ,s2 )
E∗∗ (δ1 α, δ2 α) := |(0,0) .
∂s1 ∂s2
Theorem 5.2.9 (The second variation formula).
X Z 1
∂
E∗∗ (δ1 α, δ2 α) = − hδ2 α, ∆t δ1 αi − hδ2 α, ∇2 δ1 α − R(γ̇, δ1 α)γ̇idt, (5.2.3)
t 0 ∂t

where R denotes the Riemann curvature. In particular, E∗∗ is a bilinear functional on

Tγ .

Proof. According to the first variation formula we have

X Z 1
∂E ∂α ∂α
=− hδ2 α, ∆t i− hδ2 α, ∇ ∂ idt.
∂s2 t
∂t 0 ∂t ∂t

Hence  
∂2E X X ∂α
=− h∇ ∂ δ2 α, ∆1 γ̇i − hδ2 α, ∇ ∂ ∆t i
∂s1 ∂s2 t
∂s1
t
∂s1 ∂t
Z 1 Z 1
∂α
− h∇ ∂ δ2 α, ∇ ∂ γ̇idt − hδ2 α, ∇ ∂ ∇ ∂ idt. (5.2.4)
∂s1 ∂t ∂s1 ∂t ∂t
0 0
Since γ is a geodesic, we have

∆t γ̇ = 0 and ∇ ∂ γ̇ = 0.
∂t
216 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

∂
Using the commutativity of ∂t with ∂s∂ 1 we deduce
     
∂α ∂α
∇ ∂ ∆t = ∆t ∇ ∂ = ∆t ∇ ∂ δ1 α .
∂s1 ∂t ∂s1 ∂t ∂t

Finally, the definition of the curvature implies

∇ ∂ ∇∂ =∇∂ ∇ ∂ + R(δ1 α, γ̇).

∂s1 ∂t ∂t ∂s1

Putting all the above together we deduce immediately the equality (5.2.3). ⊔
⊓

Corollary 5.2.10.
E∗∗ (V1 , V2 ) = E∗∗ (V2 , V1 ), ∀V1 , V2 ∈ Tγ . ⊔
⊓

5.2.2 Jacobi fields

In this subsection we will put to work the elements of calculus of variations presented so
far. Let (M, g) be a Riemann manifold and p, q ∈ M .
Definition 5.2.11. Let γ ∈ Ωp,q be a geodesic. A geodesic variation of γ is a smooth
map αs (t) : (−ε, ε) × [0, 1] → M such that, α0 = γ, and t 7→ αs (t) is a geodesic for all s.
We set as usual δα = ∂α ∂s |s=0 . ⊔
⊓

Proposition 5.2.12. Let γ ∈ Ωp,q be a geodesic and (αs ) a geodesic variation of γ. Then
the infinitesimal variation δα satisfies the Jacobi equation

∇2t δα = R(γ̇, δα)γ̇ (∇t = ∇ ∂ ).

∂t

Proof.    
∂α ∂α
∇2t δα = ∇t ∇t = ∇t ∇s
∂s ∂t
 
∂α ∂α ∂α
= ∇s ∇t + R(γ̇, δα) = R(γ̇, δα) . ⊔
⊓
∂t ∂t ∂t
Definition 5.2.13. A smooth vector field J along a geodesic γ is called a Jacobi field if
it satisfies the Jacobi equation
∇2t J = R(γ̇, J)γ̇. ⊔
⊓
Exercise 5.2.14. Show that if J is a Jacobi field along a geodesic γ, then there exists a
geodesic variation αs of γ such that J = δα. ⊔
⊓

Exercise 5.2.15. Let γ ∈ Ωp,q , and J a vector field along γ.

(a) Prove that J is a Jacobi field if and only if

E∗∗ (J, V ) = 0, ∀V ∈ Tγ .

(b) Show that a vector field J along γ which vanishes at endpoints, is a Jacobi field if any
only if E∗∗ (J, W ) = 0, for all vector fields W along γ. ⊔
⊓
5.2. THE VARIATIONAL THEORY OF GEODESICS 217

γ
N

Figure 5.3: The poles are conjugate along meridians.

Exercise 5.2.16. Let γ ∈ Ωp,q be a geodesic. Define Jp to be the space of Jacobi fields
V along γ such that V (p) = 0. Show that dim Jp = dim M , and moreover, the evaluation
map
evq : Jp → Tp M V 7→ ∇t V (p)

is a linear isomorphism. ⊔
⊓

Definition 5.2.17. Let γ(t) be a geodesic. Two points γ(t1 ) and γ(t2 ) on γ are said to
be conjugate along γ if there exists a nontrivial Jacobi field J along γ such that J(ti ) = 0,
i = 1, 2. ⊔
⊓

Example 5.2.18. Consider γ : [0, 2π] → S 2 a meridian on the round sphere connection
the poles. One can verify easily (using Clairaut’s theorem) that γ is a geodesic. The
counterclockwise rotation by an angle θ about the z-axis will produce a new meridian,
hence a new geodesic γθ . Thus (γθ ) is a geodesic variation of γ with fixed endpoints. δγ
is a Jacobi field vanishing at the poles. We conclude that the poles are conjugate along
any meridian (see Figure 5.3). ⊔
⊓

Definition 5.2.19. A geodesic γ ∈ Ωp,q is said to be nondegenerate if q is not conjugated

to p along γ. ⊔
⊓

The following result (partially) explains the geometric significance of conjugate points.

Theorem 5.2.20. Let γ ∈ Ωp,q be a nondegenerate, minimal geodesic. Then p is con-

jugate with no point on γ other than itself. In particular, a geodesic segment containing
conjugate points cannot be minimal !
218 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

Proof. We argue by contradiction. Let p1 = γ(t1 ) be a point on γ conjugate with p.

Denote by Jt a Jacobi field along γ |[0,t1 ] such that J0 = 0 and Jt1 = 0. Define V ∈ Tγ by
(
Jt , t ∈ [0, t1 ]
Vt =
0, t ≥ t1 .

We will prove that Vt is a Jacobi field along γ which contradicts the nondegeneracy of γ.
Step 1.
E∗∗ (U, U ) ≥ 0 ∀U ∈ Tγ . (5.2.5)
Indeed, let αs denote a variation of γ such that δα = U . One computes easily that

d2
E(αs2 ) = 2E∗∗ (U, U ).
ds2
Since γ is minimal for any small s we have length (αs2 ) ≥ length (α0 ) so that
Z 1 2
1 1 1
E(αs2 ) ≥ |α̇s2 |dt = length (αs2 )2 ≥ length (α0 )2
2 0 2 2

1
= length (γ)2 = E(α0 ).
2
Hence
d2
|s=0 E(αs2 ) ≥ 0.
ds2
This proves (5.2.5).
Step 2. E∗∗ (V, V ) = 0. This follows immediately from the second variation formula and
the fact that the nontrivial portion of V is a Jacobi field.
Step 3.
E∗∗ (U, V ) = 0, ∀U ∈ Tγ .
From (5.2.5) and Step 2 we deduce

0 = E∗∗ (V, V ) ≤ E∗∗ (V + τ U, V + τ U ) = fU (τ ) ∀τ.

Thus, τ = 0 is a global minimum of fU (τ ) so that

fU′ (0) = 0.

Step 3 follows from the above equality using the bilinearity and the symmetry of E∗∗ . The
final conclusion (that V is a Jacobi field) follows from Exercise 5.2.15. ⊔
⊓

Exercise 5.2.21. Let γ : R → M be a geodesic. Prove that the set


t ∈ R ; γ(t) is conjugate to γ(0)

is discrete. ⊔
⊓
5.2. THE VARIATIONAL THEORY OF GEODESICS 219

Definition 5.2.22. Let γ ∈ Ωp,q be a geodesic. We define its index, denoted by ind (γ),
as the cardinality of the set

Cγ = t ∈ (0, 1) ; is conjugate to γ(0)

which by Exercise 5.2.21 is finite. ⊔

⊓

Theorem 5.2.20 can be reformulated as follows: the index of a nondegenerate minimal

geodesic is zero.
The index of a geodesic obviously depends on the curvature of the manifold. Often,
this dependence is very powerful.

Theorem 5.2.23. Let M be a Riemann manifold with non-positive sectional curvature,

i.e.,
hR(X, Y )Y, Xi ≤ 0 ∀X, Y ∈ Tx M ∀x ∈ M. (5.2.6)

Then for any p, q ∈ M and any geodesic γ ∈ Ωp,q , ind (γ) = 0.

Proof. It suffices to show that for any geodesic γ : [0, 1] → M the point γ(1) is not
conjugated to γ(0).
Let Jt be a Jacobi field along γ vanishing at the endpoints. Thus

∇2t J = R(γ̇, J)γ̇,

so that Z Z Z
1 1 1
h∇2t J, Jidt = hR(γ̇, J)γ̇, Jidt = − hR(J, γ̇)γ̇, Jidt.
0 0 0

We integrate by parts the left-hand-side of the above equality, and we deduce

Z 1 Z 1
h∇t J, Ji|10 − |∇t J|2 dt = − hR(J, γ̇)γ̇, Jidt.
0 0

Since J(τ ) = 0 for τ = 0, 1, we deduce using (5.2.6)

Z 1
|∇t J|2 dt ≤ 0.
0

This implies ∇t J = 0 which coupled with the condition J(0) = 0 implies J ≡ 0. The
proof is complete. ⊔
⊓

The notion of conjugacy is intimately related to the behavior of the exponential map.

Theorem 5.2.24. Let (M, g) be a connected, complete, Riemann manifold and q0 ∈ M .

A point q ∈ M is conjugated to q0 along some geodesic if and only if it is a critical value
of the exponential map
expq0 : Tq0 M → M.
220 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

Proof. Let q = expq0 v, v ∈ Tq0 M . Assume first that q is a critical value for expq0 , and v
is a critical point. Then Dv expq0 (X) = 0, for some X ∈ Tv (Tq0 M ). Let v(s) be a path
in Tq0 M such that v(0) = v and v̇(0) = X. The map (s, t) 7→ expq0 (tv(s)) is a geodesic
variation of the radial geodesic γv : t 7→ expq0 (tv). Hence, the vector field

∂
W = |s=0 expq0 (tv(s))
∂s
is a Jacobi field along γv . Obviously W (0) = 0, and moreover
∂
W (1) = |s=0 expq0 (v(s)) = Dv expq0 (X) = 0.
∂s
On the other hand this is a nontrivial field since
∂
∇t W = ∇s |s=0 expq0 (tv(s)) = ∇s v(s) |s=0 6= 0.
∂t
This proves q0 and q are conjugated along γv .
Conversely, assume v is not a critical point for expq0 . For any X ∈ Tv (Tq0 M ) denote
by JX the Jacobi field along γv such that

JX (q0 ) = 0. (5.2.7)

The existence of such a Jacobi field follows from Exercise 5.2.16. As in that exercise,
denote by Jq0 the space of Jacobi fields J along γv such that J(q0 ) = 0. The map

Tv (Tq0 M ) → Jq0 X 7→ JX

is a linear isomorphism. Thus, a Jacobi field along γv vanishing at both q0 and q must
have the form JX , where X ∈ Tv (Tq0 M ) satisfies Dv expq0 (X) = 0. Since v is not a critical
point, this means X = 0 so that JX ≡ 0. ⊔
⊓

Corollary 5.2.25. On a complete Riemann manifold M with non-positive sectional cur-

vature the exponential map expq has no critical values for any q ∈ M . ⊔
⊓

We will see in the next chapter that this corollary has a lot to say about the topology
of M .
Consider now the following experiment. Stretch the round two-dimensional sphere of
radius 1 until it becomes “very long”. A possible shape one can obtain may look like
in Figure 5.4. The long tube is very similar to a piece of cylinder so that the total (=
scalar) curvature is very close to zero, in other words is very small. The lesson to learn
from this intuitive experiment is that the price we have to pay for lengthening the sphere
is decreasing the curvature. Equivalently, a highly curved surface cannot have a large
diameter. Our next result offers a more quantitative description of this phenomenon.
Theorem 5.2.26 (Myers). Let M be an n-dimensional complete Riemann manifold. If
for all X ∈ Vect (M )
(n − 1)
Ric (X, X) ≥ |X|2 ,
r2
5.2. THE VARIATIONAL THEORY OF GEODESICS 221

Figure 5.4: Lengthening a sphere.

then every geodesic of length ≥ πr has conjugate points and thus is not minimal. Hence

diam (M ) = sup{dist (p, q) ; p, q ∈ M } ≤ πr,

and in particular, Hopf-Rinow theorem implies that M must be compact. ⊔

⊓

Proof. Fix a minimal geodesic γ : [0, ℓ] → M of length ℓ, and let ei (t) be an orthonormal
basis of vector fields along γ such that en (t) = γ̇(t). Set q0 = γ(0), and q1 = γ(ℓ). Since
γ is minimal we deduce
E∗∗ (V, V ) ≥ 0 ∀V ∈ Tγ .
Set Wi = sin(πt/ℓ)ei . Then
Z ℓ
E∗∗ (Wi , Wi ) = − hWi , ∇t Wi + R(Wi , γ̇)γ̇idt
0
Z ℓ

= sin2 (πt/ℓ) π 2 /ℓ2 − hR(ei , γ̇)γ̇, ei i dt.
0

We sum over i = 1, . . . , n − 1, and we obtain

n−1
X Z ℓ

E∗∗ (Wi , Wi ) = sin2 πt/ℓ (n − 1)π 2 /ℓ2 − Ric (γ̇, γ̇) dt ≥ 0.
i=1 0

If ℓ > πr, then

(n − 1)π 2 /ℓ2 − Ric (γ̇, γ̇) < 0,
so that,
n−1
X
E∗∗ (Wi , Wi ) < 0.
i=1
222 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

Hence, at least for some Wi , we have E∗∗ (Wi , Wi ) < 0. This contradicts the minimality
of γ. The proof is complete. ⊔
⊓

We already know that the Killing form of a compact, semisimple Lie group is positive
definite; see Exercise 4.1.19. The next result shows that the converse is also true.
Corollary 5.2.27. A semisimple Lie group G with positive definite Killing pairing is
compact.

Proof. The Killing form defines in this case a bi-invariant Riemann metric on G. Its
geodesics through the origin 1 ∈ G are the 1-parameter subgroups exp(tX) which are
defined for all t ∈ R. Hence, by Hopf-Rinow theorem G has to be complete.
On the other hand, we have computed the Ricci curvature of the Killing metric, and
we found
1
Ric (X, Y ) = κ(X, Y ) ∀X, Y ∈ LG .
4
The corollary now follows from Myers’ theorem. ⊔
⊓

Exercise 5.2.28. Let M be a Riemann manifold and q ∈ M . For the unitary vectors
X, Y ∈ Tq M consider the family of geodesics

γs (t) = expq t(X + sY ).

Denote by Wt = δγs the associated Jacobi field along γ0 (t). Form f (t) = |Wt |2 . Prove the
following.
(a) Wt = DtX expq (Y ) = Frechet derivative of v 7→ expq (v).
(b) f (t) = t2 − 13 hR(Y, X)X, Y iq t4 + O(t5 ).
(c) Denote by xi a collection of normal coordinates at q. Prove that

1
gkℓ (x) = δkℓ − Rkijℓ xi xj + O(3).
3
1
det gij (x) = 1 − Rij xi xj + O(3).
3
(d) Let
Dr (q) = {x ∈ Tq M ; |x| ≤ r}.
Prove that if the Ricci curvature is negative definite at q then

vol0 (Dr (q)) ≤ volg (expq (Dr (q))

for all r sufficiently small. Above, vol0 denotes the Euclidean volume in Tq M while volg
denotes the volume on the Riemann manifold M . ⊔
⊓

Remark 5.2.29. The interdependence “curvature-topology” on a Riemann manifold has

deep reaching ramifications which stimulate the curiosity of many researchers. We refer
to [35] or [94] and the extensive references therein for a presentation of some of the most
attractive results in this direction. ⊔
⊓
5.2. THE VARIATIONAL THEORY OF GEODESICS 223

5.2.3 The Hamilton-Jacobi equations

Suppose that (M, g) is a smooth, m-dimensional manifold, p0 ∈ M , U is an open, geodesi-
cally convex neighborhood of p0 and (x1 , . . . , xm ) are normal coordinates on U centered
at p0 . We have a smooth function

η : U × U → [0, ∞), η(p, q) = S(p, q)2 , S(p, q) := distg (p, q)

We want to investigate the partial derivatives of η at (p0 , p0 ). Using the above normal
coordinates we regard η as a function η = η(x, y) defined in an open neighborhood of
(0, 0) ∈ Rm × Rm . The next result is classical and it is the main tool that will help us find
the Taylor expansion of η(x, y) at (0, 0).
Theorem 5.2.30. Fix y ∈ U The function U ∋ x 7→ η(x, y) satisfies the Hamilton-Jacobi
equation.
∂η(x, y) ∂η(x, y)
gkℓ (x) = 4η(x, y), ∀x, y. (5.2.8)
∂xk ∂xℓ
Proof. Our approach is inspired from [114, p.167-170]. Denote by γx,y = γx,y (t) the unique
geodesic that connects x to y parametrized such that
Z 1
S(x, y) = distg (x, y) = |γ̇x,y (t)|dt.
0

Since t 7→ |γ̇x,y (t)| is constant along a geodesic we have

Z 1 Z 1
2 2
η(x, y) = S(x, y) = |γ̇x,y (t)| dt = L(γx,y , γ̇x,y )dt,
0 0

where X
L(q, q̇) = gij (q)q̇ i q̇ j .
i,j

We now let the point x vary along an arbitrary smooth path x = x(s), |s| ≪ 1 in U . Set
x0 = x(0) ∈ U . We write
d
us (t) = γx(s),y (t), δu(t) = u (t)
s=0 s
ds
We have (using Einstein’s conventions)
Z 1
i∂η d d
δu (x0 , y) = S(x(s), y) = L(us , u̇s )ds
∂xi ds s=0 ds s=0 0
Z 1 Z 1
(5.1.2) ∂L i ∂L j
= δu dt + δu̇ dt
0 ∂q i
0 ∂ q̇ j s
(integrate by parts)
Z 1 
∂L d ∂L ∂L ∂L
= − δui dt − i δui = − i δui
∂q i dt ∂ q̇ i ∂ q̇ ∂ q̇
|0 {z }
=0
224 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

Since the path x(s) was arbitrary we deduce

∂S ∂L(x, ẋ)
i
=− , ∀i.
∂x ∂ ẋi
We set
1 ∂L 1
pi := i
= gik ẋk ,
2 ∂ ẋ 2
and we introduce the Hamiltonian
1 1
H(x, p) = gij pi pj = gij gik ẋk pj = δkj ẋk pj
2 2
1 j 1
= ẋ pj = gij ẋi ẋj = L(x, ẋ).
2 2
The quantity L(x, ẋ) is constant along the geodesic γx,y and, at the endpoint x, we have

1

η(x, y) = L(x, ẋ) = H(x, p) = H xi , ∂xj η(x, y) .
4
Expanding we get
1
η(x, y) = gij ∂xi η(x, y)∂xj η(x, y). ⊔
⊓
4
Besides their intrinsic interest, the Hamilton-Jacobi equations can be used to produce
an explicit description of the Taylor expansion for the square of the geodesic distance
η(x, y). This Taylor expansion will play an important part in Subsection 11.4.5.
With this goal in mind we introduce some notation. If f = f (t1 , . . . , tN ) is a smooth
function defined in a neighborhood of 0 ∈ RN and k is a nonnegative integer, then we
denote by [f ]k the degree k-homogeneous part in the Taylor expansion of f at 0, i.e.,

1 X α
[f ]k = ∂t f t=0
tα ∈ R[t1 , . . . , tN ].
k!
|α|=k

In the coordinates (xi ) the metric g has the form (using Einstein’s summation convention
throughout)
g = gij dxi dxj ,
where gij satisfy the estimates (see Exercise 5.2.28)

1
gkℓ = δkℓ − Rikjℓ (0)xi xj + O(|x|3 ). (5.2.9)
3
For any n × n matrix A that is small enough we have

(1 − A)−1 = 1 + A + A2 + · · · .

We deduce that
1
gkℓ = δkℓ + Rikjℓ (0)xi xj + O(|x|3 ). (5.2.10)
3
5.2. THE VARIATIONAL THEORY OF GEODESICS 225

The function η satisfies a Hamilton-Jacobi equation (5.2.8) and the obvious symmetry
conditions
m
X
η(x, y) = η(y, x), η(0, x) = η(x, 0) = |x|2 := (xi )2 , η(x, x) = 0. (5.2.11)
i=1

The degree 2 part of the Taylor expansion of η satisfies similar symmetries. This implies
m
X
[η]2 = |x − y|2 = (xi − y i )2 . (5.2.12)
i=1

The symmetries (5.2.11) suggest the introduction of new coordinates (u, v) on U × U ,

ui = xi − y i , vj = xj + y j .

Then
1 1
xi = (ui + vi ), y j = (vj − uj ), ∂xi = ∂ui + ∂vi .
2 2
The equality (5.2.10) can be rewritten as
1 X
gkℓ (x) = δkℓ + Rikjℓ (ui + vi )(uj + vj ) + O(3). (5.2.13)
12
i,j

The symmetry relations (5.2.11) become

η(u, v) = η(−u, v), η(u, u) = |u|2 , (5.2.14)

while (5.2.12) changes to

[η]1 = 0, [η]2 = |u|2 . (5.2.15)
The equality (5.2.8) can be rewritten
X
′

gkl (x) ηu′ k + ηv′ k ηuℓ + ηv′ ℓ = 4η. (5.2.16)
k,l
| {z }| {z }
=:Ak =:Aℓ

Note that
[Ak ]0 = [Aℓ ]0 = [gkℓ ]1 = 0, (5.2.17)
while (5.2.15) implies that
[Ak ]1 = 2uk .
We deduce
X
X X
4[η]3 = [gkl ]0 [Ak ]1 [Aℓ ]2 + [Ak ]2 [Aℓ ]1 = 2[Ak ]2 [Ak ]1 = 4 uk [Ak ]2 .
k,ℓ k k

We can rewrite this last equality as a differential equation for [η]3 namely
X
[η]3 = uk (∂uk + ∂vk )[η]3 .
k
226 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

We set P := [η]3 so that P is a homogeneous polynomial of degree 3 in the variables u, v.

Moreover, according to (5.2.14), the polynomial P is even in u and P (u, u) = 0. Thus P
has the form X
P = Ci (u)vi + P0 (v),
|i {z }
=:P2

where Ci (u) is a homogeneous polynomial of degree 2 in the variables u, and P0 (v) is

homogeneous of degree 3 in the variables v.
We have
X X X X
uk ∂vk P2 = Ck (u)uk , Q1 := uk ∂vk P0 , uk ∂uk P0 = 0.
k
|k {z } k k
=:Q3

The classical Euler equations imply

X
uk ∂uk P2 = 2P2 .
k

We deduce
P = 2P2 + Q3 + Q1 ,
where the polynomials Q3 and Q1 are odd in the variable u. Since P is even in the variable
u we deduce
Q3 + Q1 = 0,
so that P2 + P0 = P = 2P2 . Hence P2 = P0 = 0 and thus

[η]3 = 0. (5.2.18)

In particular
[Ak ]2 = 0, ∀k. (5.2.19)
Going back to (5.2.16) and using (5.2.17) and (5.2.19) we deduce
X X

4[η]4 = [g kℓ ]2 [Ak ]1 [Aℓ ]1 + [gkℓ ]0 [Ak ]1 [Aℓ ]3 + [Ak ]3 [Aℓ ]1
k,ℓ k,ℓ
X X (5.2.20)
=4 [gkℓ ]2 uk uℓ + 2 uk [Ak ]3 .
k,ℓ k

We set P := [η]4 . The polynomial P is homogeneous of degree 4 in the variables u, v, and

it is even in the variable u. We can write P = P0 + P2 + P4 , where
X X
P4 = cijkl ui uj uk uℓ , P2 = Qij (u)vi vj ,
k i,j

and P0 is homogeneous of degree 4 in the variables v, Qij (u) is a homogeneous quadratic

polynomial in the variables u. We have
X X
uk [Ak ]3 = uk (∂uk + ∂vk )P.
k k
5.2. THE VARIATIONAL THEORY OF GEODESICS 227

We have X X
uk ∂uk P2ν = 2νP2ν , ν = 0, 1, 2, uk ∂vk P4 = 0,
k k
X X X

uk ∂vk P2 = uk Qij (δki vj + δkj vi ) = Qkj uk vj + Qjk vj uk
k k,i,j k,j

Using these equalities in (5.2.20) we deduce

X X
4P4 + 4P2 + 4P0 = 4 [gkℓ ]2 uk uℓ + 4P4 + 2P2 + uk ∂vk P0
k,ℓ k
X 
+ Qjk + Qkj uk vj .
k,j

This implies P0 = 0 so that P = P4 + P2 , and we can then rewrite the above equality as
X X 
P2 = 2 [gkℓ ]2 uk uℓ + Qjk + Qkj uk vj . (5.2.21)
k,ℓ k,j

Note that the equality r(u, u) = |u|2 implies P (u, u) = 0 so that

P4 (u) = P4 (u, u) = −P2 (u, u).

Therefore it suffices to determine P2 . This can be achieved using the equality (5.2.13) in
(5.2.21). We have
X 1 X
2 [g kℓ ]2 uk uℓ = Rikjℓ (ui + vi )(uj + vj )uk uℓ
6
k,ℓ i,j,k,ℓ
1 X X  X
= Rikjℓ uk uℓ vi vj + Sj (u)vj ,
6
i,j k,ℓ j
| {z }
b ij (u)
Q

where Sj (u) denotes a homogeneous polynomial of degree 3 in u. The equality (5.2.21)

can now be rewritten as
X 1X b X 1 X 
Qij (u)vi vj = Qij (u)vi vj + Sj (u)vj + Qjk + Qkj uk vj .
6 2
i,j i,j j k,j

From this we read easily

1b 1X
Qij (u) = Q ij (u) = Rikjℓ uk uℓ .
6 6
k,ℓ

This determines P2 .
1X b
P2 (u, v) = Qij (u)vi vj . (5.2.22)
6
i,j
228 CHAPTER 5. ELEMENTS OF THE CALCULUS OF VARIATIONS

As we have indicated above P2 determines P4 .

1 X
P4 (u) = −P2 (u, u) = − Rikjℓui uj uk uℓ . (5.2.23)
6
i,j,k,ℓ

The skew symmetries of the Riemann tensor imply that P4 = 0 so that

1X b b ij (u) =
X
[η]4 (u, v) = Qij (u)vi vj , Q Rikjℓ uk uℓ . (5.2.24)
6
i,j k,ℓ

Example 5.2.31. Suppose that M is a surface, i.e., m = 2. Set

K = R1212 = R2121 = −R1221 .

Note that K is the Gaussian curvature of the surface. Then

X X
b11 =
Q b 22 =
R1k1ℓ uk uℓ = Ku22 , Q R2k2ℓ uk uℓ = Ku21 ,
k,ℓ k,ℓ
X
b 12 =
Q b 21 .
R1k2ℓ uk uℓ = −Ku1 u2 = Q
k,ℓ

Hence
K 2 2 K
P2 (u, v) = (u2 v1 + u21 v22 − 2u1 u2 v1 v2 ) = (u1 v2 − u2 v1 )2 .
6 6
In higher dimensions, if we restrict (5.2.24) to the surface
n o
Σij := x ∈ M ; xk = xℓ = 0, ∀k, ℓ 6∈ {i, j }

we deduce
Q̂ii Σij
= Kij u2j , Q̂jj Σij
= Kij u2i ,
Kij
2
Q̂ij Σij
= ui vj − uj vi .
6
⊔
⊓
Chapter 6

The Fundamental Group and

Covering Spaces

In the previous chapters we almost exclusively studied local properties of manifolds. This
study is interesting only if some additional structure is present since otherwise all manifolds
are locally alike.
We noticed an interesting phenomenon: the global “shape” (topology) of a manifold
restricts the types of structures that can exist on a manifold. For example, the Gauss–
Bonnet theorem implies that on a connected sum of two tori there cannot exist metrics
with curvature everywhere positive because the integral of the curvature is a negative
universal constant.
We used the Gauss–Bonnet theorem in the opposite direction, and we deduced the
intuitively obvious fact that a sphere is not diffeomorphic to a torus because they have
distinct genera. The Gauss–Bonnet theorem involves a heavy analytical machinery which
may obscure the intuition. Notice that S 2 has a remarkable property which distinguishes
it from T 2 : on the sphere any closed curve can be shrunk to a point while on the torus
there exist at least two “independent” unshrinkable curves (see Figure 6.1). In particular,
this means the sphere is not diffeomorphic to a torus.
This chapter will set the above observations on a rigorous foundation.

Figure 6.1: Looking for unshrinkable loops.

229
230 CHAPTER 6. THE FUNDAMENTAL GROUP AND COVERING SPACES

6.1 The fundamental group

6.1.1 Basic notions
In the sequel all topological spaces will be locally path connected spaces.
Definition 6.1.1. (a) Let X and Y be two topological spaces. Two continuous maps
f0 , f1 : X → Y are said to be homotopic if there exists a continuous map

F : [0, 1] × X → Y (t, x) 7→ Ft (x),

such that, Fi ≡ fi for i = 0, 1. We write this as f0 ≃ f1 .

(b) Two topological spaces X, Y are said to be homotopy equivalent if there exist maps
f : X → Y and g : Y → X such that f ◦ g ≃ 1Y and g ◦ f ≃ 1X . We write this X ≃ Y .
⊔
(c) A topological space is said to be contractible if it is homotopy equivalent to a point.⊓

Example 6.1.2. The unit disk in the plane is contractible. The annulus {1 ≤ |z| ≤ 2} is
homotopy equivalent to the unit circle. ⊔
⊓

Definition 6.1.3. (a) Let X be a topological space and x0 ∈ X. A loop based at x0 is a

continuous map
γ : [0, 1] → X, such that γ(0) = γ(1) = x0 .
The space of loops in X based at x0 is denoted by Ω(X, x0 ).
(b) Two loops γ0 , γ1 : I → X based at x0 are said to be homotopic rel x0 if there exists a
continuous map
Γ : [0, 1] × I → X, (t, s) 7→ Γt (s),
such that
Γi (s) = γi (s) i = 0, 1,
and
(s 7→ Γt (s)) ∈ Ω(X, x0 ) ∀t ∈ [0, 1].
We write this as γ0 ≃x0 γ1 . ⊔
⊓

Note that a loop is more than a closed curve; it is a closed curve + a description of a
motion of a point around the closed curve.
Example 6.1.4. The two loops γ1 , γ2 : I → C, γk (t) = exp(2kπt), k = 1, 2 are different
though they have the same image. ⊔
⊓

Definition 6.1.5. (a) Let γ1 , γ2 be two loops based at x0 ∈ X. The product of γ1 and γ2
is the loop 
γ1 (2s) , 0 ≤ s1/2
γ1 ∗ γ2 (s) = .
γ2 (2s − 1) , 1/2 ≤ s ≤ 1
The inverse of a based loop γ is the based loop γ − defined by

γ − (s) = γ(1 − s).

(c) The identity loop is the constant loop ex0 (s) ≡ x0 . ⊔

⊓
6.1. THE FUNDAMENTAL GROUP 231

Intuitively, the product of two loops γ1 and γ2 is the loop obtained by first following
γ1 (twice faster), and then γ2 (twice faster).
The following result is left to the reader as an exercise.
Lemma 6.1.6. Let α0 ≃x0 α1 , β0 ≃x0 β1 and γ0 ≃x0 γ1 be three pairs of homotopic based
loops.. Then
(a) α0 ∗ β0 ≃x0 α1 ∗ β1 .
(b) α0 ∗ α−0 ≃x0 ex0 .
(c) α0 ∗ ex0 ≃x0 α0 .
(d) (α0 ∗ β0 ) ∗ γ0 ≃x0 α0 ∗ (β0 ∗ γ0 ). ⊔
⊓

Hence, the product operation descends to an operation “·” on Ω(X, x0 )/≃x0 , the set
of homotopy classes of based loops. Moreover the induced operation is associative, it has
a unit and each element has an inverse. Hence (Ω(X, x0 )/≃x0 , ·) is a group.
Definition 6.1.7. The group (Ω(X, x0 )/≃x0 , ·) is called the fundamental group (or the
Poincaré group) of the topological space X, and it is denoted by π1 (X, x0 ). The image of
a based loop γ in π1 (X, x0 ) is denoted by [γ]. ⊔
⊓

The elements of π1 (X, x0 ) are the “unshrinkable loops” discussed at the beginning of
this chapter.
The fundamental group π1 (X, x0 ) “sees” only the connected component of X which
contains x0 . To get more information about X one should study all the groups {π1 (X, x)}x∈X .
Proposition 6.1.8. Let X and Y be two topological spaces. Fix two points, x0 ∈ X and
y0 ∈ Y . Then any continuous map f : X → Y such that f (x0 ) = y0 induces a morphism
of groups
f∗ : π1 (X, x0 ) → π1 (Y, y0 ),
satisfying the following functoriality properties.
(a) (1X )∗ = 1π1 (X,x0 ) .
(b) If
f g
(X, x0 ) → (Y, y0 ) → (Z, z0 )
are continuous maps, such that f (x0 ) = y0 and g(y0 ) = z0 , then (g ◦ f )∗ = g∗ ◦ f∗ .
(c) Let f0 , f1 : (X, x0 ) → (Y, y0 ) be two base-point-preserving continuous maps. Assume f0
is homotopic to f1 rel x0 , i.e., there exists a continuous map F : I ×X → Y , (t, x) 7→ Ft (x)
such that, Fi (x) ≡ fi (x). for i = 0, 1 and Ft (x0 ) ≡ y0 . Then (f0 )∗ = (f1 )∗ .

Proof. Let γ ∈ Ω(X, x0 ), Then f (γ) ∈ Ω(Y, y0 ), and one can check immediately that

γ≃x0 γ ′ ⇒ f (γ) ≃y0 f (γ ′ ).

Hence the correspondence

Ω(X, x0 ) ∋ γ 7→ f (γ) ∈ Ω(Y, y0 )

descends to a map f : π1 (X, x0 ) → π1 (Y, y0 ). This is clearly a group morphism. The

statements (a) and (b) are now obvious. We prove (c).
232 CHAPTER 6. THE FUNDAMENTAL GROUP AND COVERING SPACES

x −
0 α
x
γ 1
α

Figure 6.2: Connecting base points.

Let f0 , f1 : (X, x0 ) → (Y, y0 ) be two continuous maps, and Ft a homotopy rel x0

connecting them. For any γ ∈ Ω(X, x0 ) we have

β0 = f0 (γ) ≃y0 f1 (γ) = β1 .

The above homotopy is realized by Bt = Ft (γ). ⊔

⊓

A priori, the fundamental group of a topological space X may change as the base point
varies and it almost certainly does if X has several connected components. However, if
X is connected, and thus path connected since it is locally so, all the fundamental groups
π1 (X, x), x ∈ X are isomorphic.

Proposition 6.1.9. Let X be a connected topological space. Any continuous path α :

[0, 1] → X joining x0 to x1 induces an isomorphism

α∗ : π1 (X, x0 ) → π1 (X, x1 ),

defined by, α∗ ([γ]) := [α− ∗ γ ∗ α]; see Figure 6.2. ⊔

⊓

Exercise 6.1.10. Prove Proposition 6.1.9. ⊔

⊓

Thus, the fundamental group of a connected space X is independent of the base point
modulo some isomorphism. We will write π1 (X, pt) to underscore this weak dependence
on the base point.

Corollary 6.1.11. Two homotopically equivalent connected spaces have isomorphic fun-
damental groups.

Example 6.1.12. (a) π1 (Rn , pt) ∼ π1 (pt, pt) = {1}.

(b) π1 (annulus) ∼ π1 (S 1 ). ⊔
⊓

Definition 6.1.13. A connected space X such that π1 (X, pt) = {1} is said to be simply
connected. ⊔
⊓

Exercise 6.1.14. Prove that the spheres of dimension ≥ 2 are simply connected. ⊔
⊓
6.1. THE FUNDAMENTAL GROUP 233

Exercise 6.1.15. Let G be a connected Lie group. Define a new operation “⋆ ” on Ω(G, 1)
by
(α ⋆ β)(s) = α(s) · β(s),
where · denotes the group multiplication.
(a) Prove that α ⋆ β ≃1 α ∗ β.
(b) Prove that π1 (G, 1) is Abelian. ⊔
⊓

Exercise 6.1.16. Let E → X be a rank r complex vector bundle over the smooth manifold
X, and let ∇ be a flat connection on E, i.e., F (∇) = 0. Pick x0 ∈ X, and identify the fiber
Ex0 with Cr . For any continuous, piecewise smooth γ ∈ Ω(X, x0 ) denote by Tγ = Tγ (∇)
the parallel transport along γ, so that Tγ ∈ GL (Cr ).
(a) Prove that α≃x0 β ⇒ Tα = Tβ .
(b) Tβ∗α = Tα ◦ Tβ .
Thus, any flat connection induces a group morphism

T : π1 (X, x0 ) → GL(Cr ) γ 7→ Tγ−1 .

This morphism (representation) is called the monodromy of the connection. ⊔

⊓

Example 6.1.17. We want to compute the fundamental group of the complex projective
space CPn . More precisely, we want to show it is simply connected. We will establish this
by induction.
For n = 1, CP1 ∼ = S 2 , and by Exercise 6.1.14, the sphere S 2 is simply connected. We
k
next assume CP is simply connected for k < n and prove the same is true for n.
Notice first that the natural embedding Ck+1 ֒→ Cn+1 induces an embedding CPk ֒→
n
CP . More precisely, in terms of homogeneous coordinates this embedding is given by

[z0 , . . . : zk ] 7→ [z0 , . . . , zk , 0, . . . , 0] ∈ CPn .

Choose as base point pt = [1, 0, . . . , 0] ∈ CPn , and let γ ∈ Ω(CPn , pt). We may assume γ
avoids the point P = [0, . . . , 0, 1] since we can homotop it out of any neighborhood of P .
We now use a classical construction of projective geometry. We project γ from P to the
hyperplane H = CPn−1 ֒→ CPn . More precisely, if ζ = [z0 , . . . , zn ] ∈ CPn , we denote by
π(ζ) the intersection of the line P ζ with the hyperplane H . In homogeneous coordinates

π(ζ) = [z0 (1 − zn ), . . . , zn−1 (1 − zn ), 0] (= [z0 , . . . , zn−1 , 0] when zn 6= 1).

Clearly π is continuous. For t ∈ [0, 1] define

πt (ζ) = [z0 (1 − tzn ), . . . , zn−1 (1 − tzn ), (1 − t)zn ].

Geometrically, πt flows the point ζ along the line P ζ until it reaches the hyperplane H .
Note that πt (ζ) = ζ, ∀t, and ∀ζ ∈ H . Clearly, πt is a homotopy rel pt connecting
γ = π0 (γ) to a loop γ1 in H ∼
= CPn−1 based at pt. Our induction hypothesis shows that
γ1 can be shrunk to pt inside H . This proves that CPn is simply connected. ⊔
⊓
234 CHAPTER 6. THE FUNDAMENTAL GROUP AND COVERING SPACES

6.1.2 Of categories and functors

The considerations in the previous subsection can be very elegantly presented using the
language of categories and functors. This brief subsection is a minimal introduction to
this language. The interested reader can learn more about it from the monograph [74, 89].
A category is a triplet C = (Ob(C), Hom(C), ◦) satisfying the following conditions.

(i) Ob(C) is a set whose elements are called the objects of the category.

(ii) Hom(C) is a family of sets Hom (X, Y ), one for each pair of objects X and Y . The
elements of Hom (X, Y ) are called the morphisms (or arrows) from X to Y .

(iii) ◦ is a collection of maps

◦ : Hom(X, Y ) × Hom(Y, Z) → Hom(X, Z), (f, g) 7→ g ◦ f,

which satisfies the following conditions.

(C1) For any object X, there exists a unique element 1X ∈ Hom(X, X) such that,
f ◦ 1X = f g ◦ 1X = g ∀f ∈ Hom(X, Y ), ∀g ∈ Hom(Z, X).

(C2) ∀f ∈ Hom(X, Y ), g ∈ Hom(Y, Z), h ∈ Hom(Z, W )

h ◦ (g ◦ f ) = (h ◦ g) ◦ f.

Example 6.1.18. • Top is the category of topological spaces. The objects are topological
spaces and the morphisms are the continuous maps. Here we have to be careful about one
foundational issue. Namely, the collection of all topological spaces is not a set. To avoid
this problem we need to restrict to topological spaces whose subjacent sets belong to a
certain Universe. For more about this foundational issue we refer to [74].
• (Top, ∗) is the category of marked topological spaces. The objects are pairs (X, ∗),
where X is a topological space, and ∗ is a distinguished point of X. The morphisms
f
(X, ∗) → (Y, ⋄)

are the continuous maps f : X → Y such that f (∗) = ⋄.

• F Vect is the category of vector spaces over the field F. The morphisms are the F-linear
maps.
• Gr is the category of groups, while Ab denotes the category of Abelian groups. The
morphisms are the obvious ones.
• R Mod denotes the category of left R-modules, where R is some ring. ⊔
⊓

Definition 6.1.19. Let C1 and C2 be two categories. A covariant (respectively contravari-

ant) functor is a map

F : Ob(C1 ) × Hom(C1 ) → Ob(C2 ) × Hom(C2 ),

6.2. COVERING SPACES 235

(X, f ) 7→ (F(X), F(f )),

f F(f ) F(f )
such that, if X → Y , then F(X) −→ F(Y ) (respectively F(X) ←− F(Y )), and
(i) F(1X ) = 1F(X) ,
(ii) F(g) ◦ F(f ) = F(g ◦ f ) (respectively F(f ) ◦ F(g) = F(g ◦ f )). ⊔
⊓

Example 6.1.20. Let V be a real vector space. Then the operation of right tensoring
with V is a covariant functor
L L⊗1V
⊗V : R Vect → R Vect, U U ⊗ V, (U1 → U2 ) (U1 ⊗ V → U2 ⊗ V ).

On the other hand, the operation of taking the dual defines a contravariant functor,

∗ L Lt
: R Vect → R Vect, V V ∗ , (U → V ) (V ∗ → U ∗ ).

The fundamental group construction of the previous is a covariant functor

π1 : (Top, ∗) → Gr. ⊔
⊓

In Chapter 7 we will introduce other functors very important in geometry. For more
information about categories and functors we refer to [74, 89].

6.2 Covering Spaces

6.2.1 Definitions and examples
As in the previous section we will assume that all topological spaces are locally path
connected.
Definition 6.2.1. (a) A continuous map π : X → Y is said to be a covering map if, for
any y ∈ Y , there exists an open neighborhood U of y in Y , such that π −1 (U ) is a disjoint
union of open sets Vi ⊂ X each of which is mapped homeomorphically onto U by π. Such
a neighborhood U is said to be an evenly covered neighborhood. The sets Vi are called
the sheets of π over U .
(b) Let Y be a topological space. A covering space of Y is a topological space X, together
with a covering map π : X → Y .
(c) If π : X → Y is a covering map, then for any y ∈ Y the set π −1 (y) is called the fiber
over y. ⊔
⊓

Example 6.2.2. Let D be a discrete set. Then, for any topological space X, the product
X × D is a covering of X with covering projection π(x, d) = x. This type of covering space
is said to be trivial. ⊔
⊓

Exercise 6.2.3. Show that a fibration with standard fiber a discrete space is a covering.
⊔
⊓

Example 6.2.4. The exponential map exp : R → S 1 , t 7→ exp(2πit) is a covering map.

However, its restriction to (0, ∞) is no longer a a covering map. (Prove this!). ⊔
⊓
236 CHAPTER 6. THE FUNDAMENTAL GROUP AND COVERING SPACES

Exercise 6.2.5. Let (M, g) and (M̃ , g̃) be two Riemann manifolds of the same dimension
such that (M̃ , g̃) is complete. Let φ : M̃ → M be a surjective local isometry i.e. Φ is
smooth and
|v|g = |Dφ(v)|g̃ ∀v ∈ T M̃ .
Prove that φ is a covering map. ⊔
⊓

The above exercise has a particularly nice consequence.

Theorem 6.2.6 (Cartan-Hadamard). Let (M, g) be a complete Riemann manifold with
non-positive sectional curvature. Then for every point q ∈ M , the exponential map

expq : Tq M → M

is a covering map.

Proof. The pull-back h = exp∗q (g) is a symmetric non-negative definite (0, 2)-tensor field
on Tq M . It is in fact positive definite since the map expq has no critical points due to the
non-positivity of the sectional curvature.
The lines t 7→ tv through the origin of Tq M are geodesics of h and they are defined
for all t ∈ R. By Hopf-Rinow theorem we deduce that (Tq M, h) is complete. The theorem
now follows from Exercise 6.2.5. ⊔
⊓

Exercise 6.2.7. Let G̃ and G be two Lie groups of the same dimension and φ : G̃ → G
a smooth, surjective group morphism. Prove that φ is a covering map. In particular, this
explains why exp : R → S 1 is a covering map. ⊔
⊓

Exercise 6.2.8. Identify S 3 ⊂ R4 with the group of unit quaternions

S 3 = {q ∈ H ; |q| = 1}.

The linear space R3 can be identified with the space of purely imaginary quaternions

R3 = Im H = {xi + yj + zk}.

(a) Prove that qxq −1 ∈ Im H, ∀q ∈ S 3 .

(b) Prove that for any q ∈ S 3 the linear map

Tq : Im H → ImH x 7→ qxq −1

is an isometry so that Tq ∈ SO(3). Moreover, the map

S 3 ∋ q 7→ Tq ∈ SO(3)

is a group morphism.
(c) Prove the above group morphism is a covering map. ⊔
⊓
6.2. COVERING SPACES 237

Example 6.2.9. Let M be a smooth manifold. A Riemann metric on M induces a metric

on the determinant line bundle det T M . The sphere bundle of det T M (with respect to
this metric) is a 2 : 1 covering space of M called the orientation cover of M . ⊔
⊓
π π
Definition 6.2.10. Let X1 →1 Y and X2 →2 Y be two covering spaces of Y . A morphism
of covering spaces is a continuous map F : X1 → X2 such that π2 ◦ F = π1 , i.e., the
diagram below is commutative.

X1
[℄[ w X2
F

π 1 π 2

If F is also a homeomorphism we say F is an isomorphism of covering spaces.

π
Finally, if X → Y is a covering space then its automorphisms are called deck transfor-
mations. The deck transformations form a group denoted by Deck (X, π). ⊔
⊓
exp
Exercise 6.2.11. Show that Deck (R → S 1 ) ∼
= Z. ⊔
⊓

Exercise 6.2.12. (a) Prove that the natural projection S n → RPn is a covering map.
(b) Denote by UR n
1 the tautological (real) line bundle over RP . Using a metric on this
line bundle form the associated sphere bundle S(UR n
1 ) → RP . Prove that this fiber bundle
defines a covering space isomorphic with the one described in part (a). ⊔
⊓

6.2.2 Unique lifting property

.
π
Definition 6.2.13. Let X → Y be a covering space and F : Z → Y a continuous map.
A lift of f is a continuous map F : Z → X such that π ◦ F = f , i.e. the diagram below is
commutative.
X

 π
F
⊔
⊓
f u
Z wY
π
Proposition 6.2.14 (Unique Path Lifting). Let X → Y be a covering map, γ : [0, 1] → Y
a path in Y and x0 a point in the fiber over y0 = γ(0), x0 ∈ π −1 (y0 ). Then there exists at
most one lift of γ, Γ : [0, 1] → Y such that Γ(0) = x0 .

Proof. We argue by contradiction. Assume there exist two such lifts, Γ1 , Γ2 : [0, 1] → X.
Set 
S := t ∈ [0, 1]; Γ1 (t) = Γ2 (t) .
The set S is nonempty since 0 ∈ S. Obviously S is closed so it suffices to prove that it is also
open. We will prove that for every s ∈ S, there exists ε > 0 such that [s−ε, s+ε]∩[0, 1] ⊂ S.
For simplicity, we consider the case s = 0. The general situation is entirely similar.
238 CHAPTER 6. THE FUNDAMENTAL GROUP AND COVERING SPACES

We will prove that there exists r0 > 0 such that [0, r0 ] ⊂ S. Pick a small open
neighborhood U of x0 such
that π restricts to a homeomorphism onto π(U ). There exists
r0 > 0 such that γi [0, r0 ] ⊂ U , i = 1, 2. Since π ◦ Γ1 = π ◦ Γ2 , we deduce Γ1 |[0,r0 ] =
Γ2 |[0,r0 ] . The proposition is proved. ⊔
⊓

π
Theorem 6.2.15. Let X → Y be a covering space, and f : Z → Y be a continuous map,
where Z is a connected space. Fix z0 ∈ Z, and x0 ∈ π −1 (y0 ), where y0 = f (z0 ). Then
there exists at most one lift F : Z → X of f such that F (z0 ) = x0 .

Proof. For each z ∈ Z let αz be a continuous path connecting z0 to z. If F1 , F2 are two

lifts of f such that F1 (z0 ) = F2 (z0 ) = x0 then, for any z ∈ Z, the paths Γ1 = Fi (αz ),
and Γ2 = F2 (αz ) are two lifts of γ = f (αz ) starting at the same point. From Proposition
6.2.14 we deduce that Γ1 ≡ Γ2 , i.e., F1 (z) = F2 (z), for any z ∈ Z. ⊔
⊓

6.2.3 Homotopy lifting property

π
Theorem 6.2.16 (Homotopy lifting property). Let X → Y be a covering space, f : Z →
Y be a continuous map, and F : Z → X be a lift of f . If

h : [0, 1] × Z → Y (t, z) 7→ ht (z)

is a homotopy of f (h0 (z) ≡ f (z)), then there exists a unique lift of h

H : [0, 1] × Z → X (t, z) 7→ Ht (z),

such that H0 (z) ≡ F (z).

Proof. For each z ∈ Z we can find an open neighborhood Uz of z ∈ Z, and a partition

0 = t0 < t1 < . . . < tn = 1, depending on z, such that h maps [ti−1 , ti ] × Uz into an evenly
covered neighborhood of hti−1 (z). Following this partition, we can now successively lift
h |I×Uz to a continuous map H = H z : I × Uz → X such that H0 (ζ) = F (ζ), ∀ζ ∈ Uz . By
unique lifting property, the liftings on I × Uz1 and I × Uz2 agree on I × (Uz1 ∩ Uz2 ), for
any z1 , z2 ∈ Z, and hence we can glue all these local lifts together to obtain the desired
lift H on I × Z. ⊔
⊓
π
Corollary 6.2.17 (Path lifting property). Suppose that X → Y is a covering map, y0 ∈
Y , and γ : [0, 1] → Y is a continuous path starting at y0 . Then, for every x0 ∈ π −1 (y0 )
there exists a unique lift Γ : [0, 1] → X of γ starting at x0 .

Proof. Use the previous theorem with f : {pt} → Y , f (pt) = γ(0), and ht (pt) = γ(t). ⊔
⊓
π
Corollary 6.2.18. Let X → Y be a covering space, and y0 ∈ Y . If γ0 , γ1 ∈ Ω(Y, y0 ) are
homotopic rel y0 , then any lifts Γ0 , Γ1 which start at the same point also end at the same
point, i.e. Γ0 (1) = Γ1 (1).

Proof. Lift the homotopy connecting γ0 to γ1 to a homotopy in X. By unique lifting

property this lift connects Γ0 to Γ1 . We thus get a continuous path Γt (1) inside the fiber
6.2. COVERING SPACES 239

π −1 (y0 ) which connects Γ0 (1) to Γ1 (1). Since the fibers are discrete, this path must be
constant. ⊔
⊓
π
Let X → Y be a covering space, and y0 ∈ Y . Then, for every x ∈ π −1 (y0 ), and any
γ ∈ Ω( Y, y0 ), denote by Γx the unique lift of γ starting at x. Set

x · γ := Γx (1).

By Corollary 6.2.18, if ω ∈ Ω(Y, y0 ) is homotopic to γ rel y0 , then

x · γ = x · ω.

Hence, x · γ depends only upon the equivalence class [γ] ∈ π1 (Y, y0 ). Clearly

x · ([γ] · [ω]) = (x · [γ]) · [ω],

and
x · ey0 = x,
so that the correspondence

π −1 (y0 ) × π1 (Y, y0 ) ∋ (x, γ) 7→ x · γ ∈ x · γ ∈ π −1 (y0 )

defines a right action of π1 (Y, y0 ) on the fiber π −1 (y0 ). This action is called the monodromy
of the covering. The map x 7→ x · γ is called the monodromy along γ. Note that when Y
is simply connected, the monodromy is trivial. The map π induces a group morphism

π∗ : π1 (X, x0 ) → π1 (Y, y0 ) x0 ∈ π −1 (y0 ).

Proposition 6.2.19. π∗ is injective.

Proof. Indeed, let γ ∈ Ω(X, x0 ) such that π(γ) is trivial in π1 (Y, y0 ). The homotopy
connecting π(γ) to ey0 lifts to a homotopy connecting γ to the unique lift of ey0 at x0 ,
which is ex0 . ⊔
⊓

6.2.4 On the existence of lifts

π
Theorem 6.2.20. Let X → Y be a covering space, x0 ∈ X, y0 = π(x0 ) ∈ Y , f : Z → Y
a continuous map and z0 ∈ Z such that f (z0 ) = y0 . Assume the spaces Y and Z are
connected (and thus path connected). f admits a lift F : Z → X such that F (z0 ) = x0 if
and only if
f∗ (π1 (Z, z0 )) ⊂ π∗ (π1 (X, x0 )) . (6.2.1)

Proof. Necessity. If F is such a lift then, using the functoriality of the fundamental
group construction, we deduce f∗ = π∗ ◦ F∗ . This implies the inclusion (6.2.1).
Sufficiency. For any z ∈ Z, choose a path γz from z0 to z. Then αz = f (γz ) is a
path from y0 to y = f (z). Denote by Az the unique lift of αz starting at x0 , and set
F (z) = Az (1). We claim that F is a well defined map.
240 CHAPTER 6. THE FUNDAMENTAL GROUP AND COVERING SPACES

Indeed, let ωz be another path in Z connecting z0 to z. Set λz := f (ωz ), and denote

by Λz its unique lift in X starting at x0 . We have to show that Λz (1) = Az (1). Construct
the loop based at z0
βz = ωz ∗ γz− .
Then f (βz ) is a loop in Y based at y0 . From (6.2.1) we deduce that the lift Bz of f (βz )
at x0 ∈ X is a closed path, i.e., the monodromy along f (βz ) is trivial. We now have

Λz (1) = Bz (1/2) = Az (0) = Az (1).

This proves that F is a well defined map. We still have to show that this map is also
continuous.
Pick z ∈ Z. Since f is continuous, for every arbitrarily small, evenly covered neigh-
borhood U of f (z) ∈ Y there exists a path connected neighborhood V of z ∈ Z such that
f (V ) ⊂ U . For any ζ ∈ V pick a path σ = σζ in V connecting z to ζ. Let ω denote
the path ω = γz ∗ σζ (go from z0 to z along γz , and then from z to ζ along σζ ). Then
F (ζ) = Ω(1), where Ω is the unique lift of f (ω) starting at x0 . Since (f (ζ) ∈ U , we deduce
that Ω(1) belongs to the local sheet Σ, containing F (z), which homeomorphically covers
U . We have thus proved z ∈ V ⊂ F −1 (Σ). The proof is complete since the local sheets Σ
form a basis of neighborhoods of F (z). ⊔
⊓
π
Definition 6.2.21. Let Y be a connected space. A covering space X → Y is said to be
universal if X is simply connected. ⊔
⊓
pi
Corollary 6.2.22. Let X1 → Y (i=0,1) be two covering spaces of Y . Fix xi ∈ Xi such
that p0 (x0 ) = p(x1 ) = y0 ∈ Y . If X0 is universal, then there exists a unique covering
morphism F : X0 → X1 such that F (x0 ) = x1 .

Proof. A bundle morphism F : X0 → X1 can be viewed as a lift of the map p0 : X0 → Y

to the total space of the covering space defined by p1 . The corollary follows immediately
from Theorem 6.2.20 and the unique lifting property. ⊔
⊓

Corollary 6.2.23. Every space admits at most one universal covering space (up to iso-
morphism).
Theorem 6.2.24. Let Y be a connected, locally path connected space such that each of
its points admits a simply connected neighborhood. Then Y admits an (essentially unique)
universal covering space.
Sketch of proof. Assume for simplicity that Y is a metric space. Fix y0 ∈ Y . Let Py0
denote the collection of continuous paths in Y starting at y0 . It can be topologized using
the metric of uniform convergence in Y . Two paths in Py0 are said to be homotopic rel
endpoints if they we can deform one to the other while keeping the endpoints fixed. This
defines an equivalence relation on Py0 . We denote the space of equivalence classes by Ỹ ,
and we endow it with the quotient topology. Define p : Ỹ → Y by

p([γ]) = γ(1) ∀γ ∈ Py0 .

Then (Ỹ , p) is a universal covering space of Y . ⊔

⊓
6.2. COVERING SPACES 241

Exercise 6.2.25. Finish the proof of the above theorem. ⊔

⊓
exp
Example 6.2.26. The map R → S 1 is the universal cover of S 1 . More generally, exp :
Rn → T n √
(t1 , . . . , tn ) 7→ (exp(2πit1 ), . . . , exp(2πitn )), i = −1,
is the universal cover of T n . The natural projection p : S n → RPn is the universal cover
of RPn . ⊔
⊓

Example 6.2.27. Let (M, g) be a complete Riemann manifold with non-positive sectional
curvature. By Cartan-Hadamard theorem, the exponential map expq : Tq M → M is a
covering map. Thus, the universal cover of such a manifold is a linear space of the same
dimension. In particular, the universal covering space is contractible!!!
We now have another explanation why exp : Rn → T n is a universal covering space of
the torus: the sectional curvature of the (flat) torus is zero. ⊔
⊓

Exercise 6.2.28. Let (M, g) be a complete Riemann manifold and p : M̃ → M its

universal covering space.
(a) Prove that M̃ has a natural structure of smooth manifold such that p is a local
diffeomorphism.
(b) Prove that the pullback p∗ g defines a complete Riemann metric on M̃ locally isometric
with g. ⊔
⊓

Example 6.2.29. Let (M, g) be a complete Riemann manifold such that

Ric (X, X) ≥ const.|X|2g , (6.2.2)
where const denotes a strictly positive constant. By Myers theorem M is compact. Using
the previous exercise we deduce that the universal cover M̃ is a complete Riemann manifold
locally isometric with (M, g). Hence the inequality (6.2.2) continues to hold on the covering
M̃ . Myers theorem implies again that the universal cover M̃ is compact!! In particular,
the universal cover of a semisimple, compact Lie group is compact!!! ⊔
⊓

6.2.5 The universal cover and the fundamental group

p
Theorem 6.2.30. Let X̃ → X be the universal cover of a space X. Then
π1 (X, pt) ∼
= Deck (X̃ → X).
Proof. Fix ξ0 ∈ X̃ and set x0 = p(ξ0 ). There exists a bijection
Ev : Deck (X̃) → p−1 (x0 ),
given by the evaluation
Ev (F ) := F (ξ0 ).
For any ξ ∈ π −1 (x
0 ), let γξ be a path connecting ξ0 to ξ. Any two such paths are homotopic
rel endpoints since X̃ is simply connected (check this). Their projections on the base X
determine identical elements in π1 (X, x0 ). We thus have a natural map
Ψ : Deck (X̃) → π1 (X, x0 ) F 7→ p(γF (ξ0 ) ).
242 CHAPTER 6. THE FUNDAMENTAL GROUP AND COVERING SPACES

The map Ψ is clearly a group morphism. (Think monodromy!) The injectivity and the
surjectivity of Ψ are consequences of the lifting properties of the universal cover. ⊔
⊓

Corollary 6.2.31. If the space X has a compact universal cover, then π1 (X, pt) is finite.

Proof. Indeed, the fibers of the universal cover have to be both discrete and compact.
⊔
Hence they must be finite. The map Ev in the above proof is a bijection onto Deck (X̃).⊓

Corollary 6.2.32 (H. Weyl). The fundamental group of a compact semisimple group is
finite.

Proof. Indeed, we deduce from Example 6.2.29 that the universal cover of such a group is
compact. ⊔
⊓

Example 6.2.33. From Example 6.2.11 we deduce that π1 (S 1 ) ∼

= (Z, +). ⊔
⊓

Exercise 6.2.34. (a) Prove that π1 (RPn , pt) ∼

= Z2 , ∀n ≥ 2.
n ∼ n
(b) Prove that π1 (T ) = Z . ⊔
⊓

Exercise 6.2.35. Show that the natural inclusion U (n − 1) ֒→ U (n) induces an isomor-
phism between the fundamental groups. Conclude that the map

det : U (n) → S 1

induces an isomorphism
π1 (U (n), 1) ∼
= π1 (S 1 , 1) ∼
= Z. ⊔
⊓
Chapter 7

Cohomology

7.1 DeRham cohomology

7.1.1 Speculations around the Poincaré lemma
To start off, consider the following partial differential equation in the plane. Given two
smooth functions P and Q, find a smooth function u such that

∂u ∂u
= P, = Q. (7.1.1)
∂x ∂y

As is, the formulation is still ambiguous since we have not specified the domains of the
functions u, P and Q. As it will turn out, this aspect has an incredible relevance in
geometry.
Equation (7.1.1) has another interesting feature: it is overdetermined, i.e., it imposes
too many conditions on too few unknowns. It is therefore quite natural to impose some
additional restrictions on the data P , Q just like the zero determinant condition when
solving overdetermined linear systems.
To see what restrictions one should add it is convenient to introduce the 1-form α =
P dx + Qdy. The equality (7.1.1) can be rewritten as

du = α. (7.1.2)

If (7.1.2) has at least one solution u, then 0 = d2 u = dα, so that a necessary condition for
existence is
dα = 0, (7.1.3)
i.e.,
∂P ∂Q
= .
∂y ∂x
A form satisfying (7.1.3) is said to be closed. Thus, if the equation du = α has a solution
then α is necessarily closed. Is the converse also true?
Let us introduce a bit more terminology. A form α such that the equation (7.1.2) has
a solution is said to be exact. The motivation for this terminology comes from the fact

243
244 CHAPTER 7. COHOMOLOGY

that sometimes the differential form du is called the exact differential of u. We thus have
an inclusion of vector spaces

{exact forms} ⊂ {closed forms}.

Is it true that the opposite inclusion also holds?

Amazingly, the answer to this question depends on the domain on which we study
(7.1.2). The Poincaré lemma comes to raise our hopes. It says that this is always true, at
least locally.
Lemma 7.1.1 (Poincaré lemma). Let C be an open convex set in Rn and α ∈ Ωk (C).
Then the equation
du = α (7.1.4)
has a solution u ∈ Ωk−1 (C) if and only if α is closed, dα = 0.

Proof. The necessity is clear. We prove the sufficiency. We may as well assume that 0 ∈ C.
Consider the radial vector field on C

~r = xi ∂xi ,

and denote by Φt the flow it generates. More explicitly, Φt is the linear flow

Φt (x) = et x, x ∈ Rn .

The flow lines of Φt are half-lines, and since C is convex, for every x ∈ C, and any t ≤ 0
we have Φt (x) ∈ C.
We begin with an a priori study of (7.1.4). Let u satisfy du = α. Using the homotopy
formula,
L~r = di~r + i~r d,
we get
dL~r u = d(di~r + i~r d)u = di~r α ⇒ d(L~r u − i~r α) = 0.
This suggests looking for solutions ϕ of the equation

L~r ϕ = i~r α, ϕ ∈ Ωk−1 (C). (7.1.5)

If ϕ is a solution of this equation, then

L~r dϕ = dL~r ϕ = di~r α = L~r α − iR~ dα = L~r α.

Hence the form ϕ also satisfies

L~r (dϕ − α) = 0.
P
Set ω := dϕ − α = I ωI dxI . Using the computations in Subsection 3.1.3 we deduce

L~r dxi = dxi ,

so that X
L~r ω = (L~r ωI )dxI = 0.
I
7.1. DERHAM COHOMOLOGY 245

We deduce that L~r ωI = 0, and consequently, the coefficients ωI are constants along the
flow lines of Φt which all converge at 0. Thus,

ωI = cI = const.

Each monomial cI dxI is exact, i.e., there exist ηI ∈ Ωk−1 (C) such that dηI = cI dxI . For
example, when I = 1 < 2 < · · · < k

dx1 ∧ dx2 ∧ · · · ∧ dxk = d(x1 dx2 ∧ · · · ∧ dxk ).

Thus, the equality L~r ω = 0 implies ω is exact. Hence there exists η ∈ Ωk−1 (C) such that

d(ϕ − η) = α,

i.e., the differential form u := ϕ − η solves (7.1.4). Conclusion: any solution of (7.1.5)
produces a solution of (7.1.4).
We now proceed to solve (7.1.5), and to this aim, we use the flow Φt .Define
Z 0
u= (Φt )∗ (i~r α)dt. (7.1.6)
−∞

Here the convexity assumption on C enters essentially since it implies that

Φt (C) ⊂ C ∀t ≤ 0,

so that if the above integral is convergent, then u is a form on C. If we write

X
(Φt )∗ (i~r α) = ηIt (x)dxI ,
|I|=k−1

then
X Z 0 
u(x) = ηIt (x)dt dxI . (7.1.7)
I −∞

We have to check two things.

A. The integral in (7.1.7) is well defined. To see this, we first write
X
α= αJ dxJ ,
|J|=k

and then we set

A(x) := max |αJ (τ x)|.
J;0≤τ ≤1

Then !
X
(Φt )∗ (i~r α) = et i~r αJ (et x)dxJ ,
J

so that
|ηIt (x)| ≤ Cet |x|A(x) ∀t ≤ 0.
246 CHAPTER 7. COHOMOLOGY

This proves the integral in (7.1.7) converges.

B. The differential form u defined by (7.1.6) is a solution of (7.1.5). Indeed,
1
L~r u = lim ( (Φs )∗ u − u )
s→0 s
 Z 0 Z 0 
1 s ∗ t ∗ t ∗
= lim (Φ ) (Φ ) (i~r α)dt − (Φ ) (i~r α)dt
s→0 s −∞ −∞
Z s Z 0 
= lim (Φt )∗ (i~r α)dt − (Φt )∗ (i~r α)dt
s→0
Z s−∞ −∞

= lim (Φt )∗ (i~r α)dt = (Φ0 )∗ (i~r α) = i~r α.

s→0 0

The Poincaré lemma is proved. ⊔

⊓

The local solvability does not in any way implies global solvability. Something happens
when one tries to go from local to global.
Example 7.1.2. Consider the form dθ on R2 \{0} where (r, θ) denote the polar coordinates
in the punctured plane. To write it in cartesian coordinates (x, y) we use the equality
y
tan θ =
x
so that
y dy y2 −ydx + xdy
(1 + tan2 θ)dθ = − 2 dx + and (1 + 2 )dθ = ,
x x x x2
i.e.,
−ydx + xdy
dθ = = α.
x2 + y 2
Obviously, dα = d2 θ = 0 on R2 \ {0} so that α is closed on the punctured plane. Can we
find a smooth function u on R2 \ {0} such that du = α?
We know that we can always do this locally. However, we cannot achieve this globally.
Indeed, if this was possible, then
Z Z Z
du = α= dθ = 2π.
S1 S1 S1

On the other hand, using polar coordinates u = u(r, θ) we get

Z Z Z 2π
∂u ∂u
du = dθ = dθ = u(1, 2π) − u(1, 0) = 0.
S1 S 1 ∂θ 0 ∂θ
Hence on R2 \ {0}
{exact forms} =
6 {closed forms}.
We see what a dramatic difference a point can make: R2 \ {point} is structurally very
different from R2 . ⊔
⊓

The artifice in the previous example simply increases the mystery. It is still not clear
what makes it impossible to patch-up local solutions. The next subsection describes two
ways to deal with this issue.
7.1. DERHAM COHOMOLOGY 247

7.1.2 Čech vs. DeRham

Let us try to analyze what prevents the “spreading” of local solvability of (7.1.4) to global
solvability. We will stay in the low degree range.

The Čech approach. Consider a closed 1-form ω on a smooth manifold. To solve the
equation du = ω, u ∈ C ∞ (M ) we first cover M by open sets (Uα ) which are geodesically
convex with respect to some fixed Riemann metric.
Poincaré lemma shows that we can solve du = ω on each open set Uα so that we can
find a smooth function fα ∈ C ∞ (Uα ) such that dfα = ω. We get a global solution if and
only if
fαβ = fα − fβ = 0 on each Uαβ = Uα ∩ Uβ 6= ∅.
For fixed α, the solutions of the equation du = ω on Uα differ only by additive constants,
i.e., closed 0-forms.
The quantities fαβ satisfy dfαβ = 0 on the (connected) overlaps Uαβ so they are con-
stants. Clearly they satisfy the conditions

fαβ + fβγ + fγα = 0 on every Uαβγ := Uα ∩ Uβ ∩ Uγ 6= ∅. (7.1.8)

On each Uα we have, as we have seen, several choices of solutions. Altering a choice is

tantamount to adding a constant fα → fα + cα . The quantities fαβ change according to

fαβ → fαβ + cα − cβ .

Thus, the global solvability issue leads to the following situation.

Pick a collection of local solutions fα . The equation du = ω is globally solvable if we
can alter each fα by a constant cα such that

fαβ = (cβ − cα ) ∀α, β such that Uαβ 6= ∅. (7.1.9)

We can start the alteration at some open set Uα , and work our way up from one such
open set to its neighbors, always trying to implement (7.1.9). It may happen that in the
process we might have to return to an open set whose solution was already altered. Now
we are in trouble. (Try this on S 1 , and ω = dθ.) After several attempts one can point the
finger to the culprit: the global topology of the manifold may force us to always return to
some already altered local solution.
Notice that we replaced the partial differential equation du = ω with a system of linear
equations (7.1.9), where the constants fαβ are subject to the constraints (7.1.8). This is
no computational progress since the complexity of the combinatorics of this system makes
it impossible solve in most cases.
The above considerations extend to higher degree, and one can imagine that the com-
plexity increases considerably. This is however the approach Čech adopted in order to
study the topology of manifolds, and although it may seem computationally hopeless, its
theoretical insights are invaluable.

The DeRham approach. This time we postpone asking why the global solvability
is not always possible. Instead, for each smooth manifold M one considers the Z-graded
248 CHAPTER 7. COHOMOLOGY

vector spaces
M
M
B • (M ) = B k (M ), B 0 (M ) := {0} Z • (M ) := Z k (M ),
k≥0 k≥0

where 
B k (M ) = dω ∈ Ωk (M ); ω ∈ Ωk−1 (M ) = exact k − forms,
and 
Z k (M = η ∈ Ωk (M ); dη = 0 = closed k − forms.
Clearly B k ⊂ Z k . We form the quotients,

H k (M ) := Z k (M )/B k (M ).

Intuitively, this space consists of those closed k-forms ω for which the equation du = ω
has no global solution u ∈ Ωk−1 (M ). Thus, if we can somehow describe these spaces, we
may get an idea “who” is responsible for the global nonsolvability.
Definition 7.1.3. For any smooth manifold M the vector space H k (M ) is called the k-th
DeRham cohomology group of M . ⊔
⊓

Clearly H k (M ) = 0 for k > dim M .

Example 7.1.4. The Poincaré lemma shows that H k (Rn ) = 0 for k > 0. The discussion
in Example 7.1.2 shows that H 1 (R2 \ {0}) 6= 0. ⊔
⊓

Proposition 7.1.5. For any smooth manifold M

dim H 0 (M ) = number of connected components of M.

Proof. Indeed 
H 0 (M ) = Z 0 (M ) = f ∈ C ∞ (M ) ; df = 0 .
Thus, H 0 (M ) coincides with the linear space of locally constant functions. These are
constant on the connected components of M . ⊔
⊓

We see that H 0 (M ), the simplest of the DeRham groups, already contains an important
topological information. Obviously the groups H k are diffeomorphism invariants, and its
is reasonable to suspect that the higher cohomology groups may contain more topological
information.
Thus, to any smooth manifold M we can now associate the graded vector space
M
H • (M ) := H k (M ).
k≥0

A priori, the spaces H k (M ) may be infinite dimensional. The Poincaré polynomial of M ,

denoted by PM (t), is defined by
X
PM (t) = tk dim H k (M ),
k≥0
7.1. DERHAM COHOMOLOGY 249

every time the right-hand-side expression makes sense. The number dim H k (M ) is usually
denoted by bk (M ), and it is called the k-th Betti number of M . Hence
X
PM (t) = bk (M )tk .
k

The alternating sum X

χ(M ) := (−1)k bk (M ),
k
is called the Euler characteristic of M .
Exercise 7.1.6. Show that PS 1 (t) = 1 + t. ⊔
⊓

We will spend the remaining of this chapter trying to understand what is that these
groups do and which, if any, is the connection between the two approaches outlined above.

7.1.3 Very little homological algebra

At this point it is important to isolate the common algebraic skeleton on which both
the DeRham and the Čech approaches are built. This requires a little terminology from
homological algebra.
☞ In the sequel all rings will be assumed commutative with 1.

Definition 7.1.7. (a) Let R be a ring, and let

M M
C• = C n, D• = Dn
n∈Z n∈Z

be two Z-graded left R-modules. A degree k-morphism φ : C • → D • is an R-module

morphism such that
φ(C n ) ⊂ D n+k ∀n ∈ Z.
(b) Let M
C• = Cn
n∈Z
be a Z-graded R-module. A boundary (respectively coboundary) operator is a degree −1
(respectively a degree 1) endomorphism d : C • → C • such that d2 = 0.
A chain (respectively cochain) complex over R is a pair (C • , d), where C • is a Z-graded
R-module, and d is a boundary (respectively a coboundary) operator. ⊔
⊓

In this book we will be interested mainly in cochain complexes so in the remaining

part of this subsection we will stick to this situation. In this case cochain complexes are
usually described as (C • = ⊕n∈Z C n , d). Moreover, we will consider only the case C n = 0
for n < 0.
Traditionally, a cochain complex is represented as a long sequence of R-modules and
morphisms of R-modules
dn−1 d
(C • , d) : · · · → C n−1 −→ C n −→
n
C n+1 → · · · ,
such that range (dn−1 ) ⊂ ker (dn ), i.e., dn dn−1 = 0.
250 CHAPTER 7. COHOMOLOGY

Definition 7.1.8. Let

dn−1 d
· · · → C n−1 −→ C n −→
n
C n+1 → · · ·

be a cochain complex of R-modules. Set

Z n (C) := ker dn B n (C) := range (dn−1 ).

The elements of Z n (C) are called cocycles, and the elements of B n (C) are called cobound-
aries.
Two cocycles c, c′ ∈ Z n (C) are said to be cohomologous if c−c′ ∈ B n (C). The quotient
module
H n (C) := Z n (C)/B n (C)
is called the n-th cohomology group (module) of C. It can be identified with the set of
equivalence classes of cohomologous cocycles. A cochain complex complex C is said to be
acyclic if H n (C) = 0 for all n > 0. ⊔
⊓

For a cochain complex (C • , d) one usually writes

M
H • (C) = H • (C, d) = H n (C).
n≥0

Example 7.1.9. (The DeRham complex). Let M be an m-dimensional smooth

manifold. Then the sequence
d d d
0 → Ω0 (M ) → Ω1 → · · · → Ωm (M ) → 0

(where d is the exterior derivative) is a cochain complex of real vector spaces called the
DeRham complex. Its cohomology groups are the DeRham cohomology groups of the
manifold. ⊔
⊓

Example 7.1.10. Let (g, [·, ·]) be a real Lie algebra. Define

d : Λk g∗ → Λk+1 g∗ ,

by
X
(dω)(X0 , X1 , . . . , Xk ) := (−1)i+j ω([Xi , Xj ], X0 , . . . , X̂i , . . . , X̂j , . . . , Xk ),
0≤i<j≤k

where, as usual, the hatˆindicates a missing argument.

According to the computations in Example 3.2.9 the operator d is a coboundary op-
erator, so that (Λ• g∗ , d) is a cochain complex. Its cohomology is called the Lie algebra
cohomology, and it is denoted by H • (g). ⊔
⊓

Exercise 7.1.11. (a) Let g be a real Lie algebra. Show that

H 1 (g) ∼
= (g/[g, g])∗ ,
7.1. DERHAM COHOMOLOGY 251


where [g, g] = span [X, Y ] ; X, Y ∈ g .
(b) Compute H 1 (gl(n, R)), where gl(n, R) denotes the Lie algebra of n × n real matrices
with the bracket given by the commutator.
(c) (Whitehead) Let g be a semisimple Lie algebra, i.e., its Killing pairing is nondegenerate.
Prove that H 1 (g) = {0}. (Hint: Prove that [g, g]⊥ = 0, where ⊥ denotes the orthogonal
complement with respect to the Killing pairing.) ⊔
⊓

Proposition 7.1.12. Let

dn−1 d
(C • , d) : · · · → C n−1 −→ C n −→
n
C n+1 → · · · ,
be a cochain complex of R-modules. Assume moreover that C is also a Z-graded R-algebra,
i.e., there exists an associative multiplication such that
C n · C m ⊂ C m+n ∀m, n.
If d is a quasi-derivation, i.e.,
d(x · y) = ±(dx) · y ± x · (dy) ∀x, y ∈ C,
then H • (C) inherits a structure of Z-graded R-algebra.
A cochain complex as in the above proposition is called a differential graded algebra
or DGA.
Proof. It suffices to show Z • (C) · Z • (C) ⊂ Z • (C), and B • (C) · B • (C) ⊂ B • (C).
If dx = dy = 0, then d(xy) = ±(dx)y ± x(dy) = 0. If x = dx′ and y = dy ′ then, since
d2 = 0, we deduce xy = ±(dx′ dy ′ ). ⊔
⊓

Corollary 7.1.13. The DeRham cohomology of a smooth manifold has an R-algebra struc-
ture induced by the exterior multiplication of differential forms. ⊔
⊓

Definition 7.1.14. Let (A• , d) and (B • , δ) be two cochain complexes of R-modules.

(a) A cochain morphism, or morphism of cochain complexes is a degree 0 morphism φ :
A• → B • such that φ ◦ d = δ ◦ φ, i.e., the diagram below is commutative for any n.

An
dn
w An+1
φn φn+1

un u
B
δn
w B n+1
(b) Two cochain morphisms φ, ψ : A• → B • are said to be cochain homotopic , and we
write this φ ≃ ψ, if there exists a degree −1 morphism χ : {An → B n−1 } such that
φ(a) − ψ(a) = ±δ ◦ χ(a) ± χ ◦ d(a).
(c) Two cochain complexes (A• , d), and (B • , δ) are said to be homotopic, if there exist
cochain morphism
φ : A → B and ψ : B → A,
such that ψ ◦ φ ≃ 1A , and φ ◦ ψ ≃ 1B . ⊔
⊓
252 CHAPTER 7. COHOMOLOGY

Example 7.1.15. The commutation rules in Subsection 3.2.1, namely [LX , d] = 0, and
[iX , d]s = LX , show that for each vector field X on a smooth manifold M , the Lie derivative
along X, LX : Ω• (M ) → Ω• (M ) is a cochain morphism homotopic with the trivial map
(≡ 0). The interior derivative iX is the cochain homotopy achieving this. ⊔
⊓

Proposition 7.1.16. (a) Any cochain morphism φ : (A• , d) → (B • , δ) induces a degree

zero morphism in cohomology

φ∗ : H • (A) → H • (B).

(b) If the cochain maps φ, ψ : A → B are cochain homotopic, then they induce identical
morphisms in cohomology, φ∗ = ψ∗ .
φ ψ
(c) (1A )∗ = 1H • (A) , and if (A•0 , d0 ) −→ (A•1 , d1 ) −→ (A•2 , d2 ) are cochain morphisms, then
(ψ ◦ φ)∗ = ψ∗ ◦ φ∗ .

Proof. (a) It boils down to checking the inclusions

φ(Z n (A)) ⊂ Z n (B) and φ(B n (A)) ⊂ B n (B).

These follow immediately from the definition of a cochain map.

(b) We have to show that φ(cocycle) − ψ(cocycle) = coboundary. Let da = 0. Then

φ(a) − ψ(a) = ±δ(χ(a) ± χ(da) = δ(±χ(a)) = coboundary in B.

(c) Obvious. ⊔
⊓

Corollary 7.1.17. If two cochain complexes (A• , d) and (B • , δ) are cochain homotopic,
then their cohomology modules are isomorphic. ⊔
⊓

Proposition 7.1.18. Let

φ ψ
0 → (A• , dA ) −→ (B • , dB ) −→ (C • , dC ) → 0

be a short exact sequence of cochain complexes of R-modules. This means that we have a
commutative diagram
7.1. DERHAM COHOMOLOGY 253

.. .. ..
u
. u
. u
.

dA dB dC

0 w An+1
u
φn+1
w B n+1
u
ψn+1
w C n+1
u w0
dA dB dC

0 w Au n φn
w Bu n ψn
w Cu n w0 (7.1.10)

dA dB dC

0 w An−1
u
ψn−1
w B n−1
u
ψn−1
w C n−1
u w0
dA dB dC

.. .. ..
. . .

in which the rows are exact. Then there exists a long exact sequence
∂n−1 φ∗ ψ∗ ∂
· · · → H n−1 (C) → H n (A) → H n (B) → H n (C) →
n
H n+1 (A) → · · · . (7.1.11)

We will not include a proof of this proposition. We believe this is one proof in ho-
mological algebra that the reader should try to produce on his/her own. We will just
indicate the construction of the connecting maps ∂n . This construction, and in fact the
entire proof, relies on a simple technique called diagram chasing.
Start with x ∈ H n (C). The cohomology class x can be represented by some cocycle
c ∈ Z n (C). Since ψn is surjective there exists b ∈ B n such that c = ψn (b). From the
commutativity of the diagram (7.1.10) we deduce 0 = dC ψn (b) = ψn+1 dB (b), i.e., dB (b) ∈
ker ψn+1 = range φn+1 . In other words, there exists a ∈ An+1 such that φn+1 (a) = dB n b.
We claim that a is a cocycle. Indeed,

φn+2 dA B B B
n+1 a = dn+1 φn+1 a = dn+1 dn b = 0.

Since φn+2 is injective we deduce dA

n+1 a = 0, i.e., a is a cocycle.
If we trace back the path which lead us from c ∈ Z n (C) to a ∈ Z n+1 (A), we can write

a = φ−1 B −1 −1 B
n+1 ◦ d ◦ ψn (c) = φn+1 ◦ d b.

This description is not entirely precise since a depends on various choices. We let the
reader check that the correspondence Z n (C) ∋ c 7→ a ∈ Z n+1 (A) above induces a well
254 CHAPTER 7. COHOMOLOGY

defined map in cohomology, ∂n : H n (C) → H n+1 (A) and moreover, the sequence (7.1.11)
is exact.
Exercise 7.1.19. 1 Suppose R is a commutative ring with 1. For any cochain complex
(K • , dK ) of R-modules, and any integer n we denote by K[n]• the complex defined by
K[n]m = K n+m , dK[n] = (−1)n dK . We associate to any cochain map f : K • → L• two
new cochain complexes:
 
(a) The cone C(f )• , dC(f ) where
     i+1 
• • • ki+1 −dK 0 k
C(f ) = K[1] ⊕ L , dC(f ) = · .
ℓi f dL ℓi
 
(b) The cylinder Cyl(f ), dCyl(f )
     
ki dK −1K i+1 0 ki
Cyl(f )• ∼
= K • ⊕ C(f )• , dCyl(f )  ki+1  =  0 −dK 0  ·  ki+1  .
ℓi 0 f dL ℓi

We have canonical inclusions α : L• → Cyl(f ), f¯ : K • → Cyl(f )• , a canonical projections

β : Cyl(f )• → L• , δ = δ(f ) : C(f )• → K[1]• , and π : Cyl(f ) → C(f ).

(i) Prove that α, β f¯, δ(f ) are cochain maps, β ◦ α = 1L and α ◦ β is cochain homotopic to
1Cyl(f ) .
(ii) Show that we have the following commutative diagram of cochain complexes, where
the rows are exact.

0 w L• π̄
w C(f )• δ(f )
w K[1]• w0
α 1C(f )

u u
0 w K• f¯
w Cyl(f ) π
w C(f ) w0
1K β

u u
K•
f
w L•
(iii) Show that the connecting morphism in the long exact sequence corresponding to the
short exact sequence
f¯ π
0 → K • −→ Cyl(f ) −→ C(f ) → 0
coincides with the morphism induced in cohomology by δ(f ) : C(f ) → K[1]• .
(iv) Prove that f induces an isomorphism in cohomology if and only if the cone of f is
acyclic. ⊔
⊓
1
This exercise describes additional features of the long exact sequence in cohomology. They are partic-
ularly useful in the study of derived categories.
7.1. DERHAM COHOMOLOGY 255

L
Exercise 7.1.20. (Abstract Morse inequalities). Let C • = n≥0 C n be a cochain
complex of vector spaces over the field F. Assume each of the vector spaces C n is finite
dimensional. Form the Poincaré series
X
PC (t) = tn dimF C n ,
n≥0

and X
PH • (C) (t) = tn dimF H n (C).
n≥0

Prove that there exists a formal series R(t) ∈ Z[[t]] with non-negative coefficients such
that
PC • (t) = PH • (C) (t) + (1 + t)R(t).
In particular, whenever it makes sense, the graded spaces C ∗ and H ∗ have identical Euler
characteristics
χ(C • ) = PC (−1) = PH • (C) (−1) = χ(H • (C)). ⊔
⊓
Exercise 7.1.21. (Additivity of Euler characteristic). Let

0 → A• → B • → C • → 0

be a short exact sequence of cochain complexes of vector spaces over the field F. Prove
that if at least two of the cohomology modules H • (A), H • (B) and H • (C) have finite
dimension over F, then the same is true about the third one, and moreover

χ(H • (B)) = χ(H • (A)) + χ(H • (C)). ⊔

⊓

Exercise 7.1.22. (Finite dimensional Hodge theory). Let

M 
(V • , d) := V n , dn ,
n≥0

be a cochain complex of real vector spaces such that dim V n < ∞, for all n. Assume that
each V n is equipped with a Euclidean metric, and denote by d∗n : V n+1 → V n the adjoint
of dn . We can now form the Laplacians

∆n : V n → V n , ∆n := d∗n dn + dn−1 d∗n−1 .

(a) Prove that ⊕n≥0 ∆n = (d + d∗ )2 .

(b) Prove that ∆n c = 0 if and only if dn c = 0 and d∗n−1 c = 0. In particular, ker ∆n ⊂
Z n (V • ).
(c) Let c ∈ Z n (V ). Prove that there exists a unique c ∈ Z n (V ) cohomologous to c such
that 
|c| = min |c′ |; c − c′ ∈ B n (C) ,
where | · | denotes the Euclidean norm in V n .
256 CHAPTER 7. COHOMOLOGY

(d) Prove that c determined in part (c) satisfies ∆n c = 0. Deduce from all the above that
the natural map
ker ∆n → H n (V • )
is a linear isomorphism. ⊔
⊓

Exercise 7.1.23. Let V be a finite dimensional real vector space, and v0 ∈ V . Define

dk = dk (v0 ) : Λk V → Λk+1 V, ω 7−→ v0 ∧ ω.

(a) Prove that

dk−1 d dk+1
· · · −→ Λk V −→
k
Λk+1 V −→ · · ·
is a cochain complex. (It is known as the Koszul complex.)
(b) Use the Cartan identity in Exercise 2.2.55 and the finite dimensional Hodge theory
described in previous exercise to prove that the Koszul complex is acyclic if v0 6= 0, i.e.,


H k Λ• V, d(v0 ) = 0, ∀k ≥ 0. ⊔
⊓

7.1.4 Functorial properties of the DeRham cohomology

Let M and N be two smooth manifolds. For any smooth map φ : M → N the pullback

φ∗ : Ω• (N ) → Ω• (M )

is a cochain morphism, i.e., φ∗ dN = dM φ∗ , where dM and respectively dN denote the

exterior derivative on M , and respectively N . Thus, φ∗ induces a morphism in cohomology
which we continue to denote by φ∗ ;

φ∗ : H • (N ) → H • (M ).

In fact, we have a more precise statement.

Proposition 7.1.24. The DeRham cohomology construction is a contravariant functor
from the category of smooth manifolds and smooth maps to the category of Z-graded vector
spaces with degree zero morphisms. ⊔
⊓

Note that the pullback is an algebra morphism φ∗ : Ω• (N ) → Ω• (M ),

φ∗ (α ∧ β) = (φ∗ α) ∧ (φ∗ β), ∀α, β ∈ Ω• (N ),

and the exterior differentiation is a quasi-derivation, so that the map it induces in coho-
mology will also be a ring morphism.
Definition 7.1.25. (a) Two smooth maps φ0 , φ1 : M → N are said to be (smoothly)
homotopic, and we write this φ0 ≃sh φ1 , if there exists a smooth map

Φ : I × M → N (t, m) 7→ Φt (m),

such that Φi = φi , for i = 0, 1.

7.1. DERHAM COHOMOLOGY 257

(b) A smooth map φ : M → N is said to be a (smooth) homotopy equivalence if there

exists a smooth map ψ : N → M such that φ ◦ ψ ≃sh 1N , and ψ ◦ φ ≃sh 1M .
(c) Two smooth manifolds M and N are said to be homotopy equivalent if there exists a
homotopy equivalence φ : M → N . ⊔
⊓

Proposition 7.1.26. Let φ0 , φ1 : M → N be two homotopic smooth maps. Then they

induce identical maps in cohomology

φ∗0 = φ∗1 : H • (N ) → H • (M ).

Proof. According to the general results in homological algebra, it suffices to show that the
pullbacks
φ∗0 , φ∗1 : Ω• (N ) → Ω• (M )
are cochain homotopic. Thus, we have to produce a map

χ : Ω• (N ) → Ω∗•−1 (M ),

such that
φ∗1 (ω) − φ∗0 (ω) = ±χ(dω) ± dχω, ∀ω ∈ Ω• (M ).
At this point, our discussion on the fibered calculus of Subsection 3.4.5 will pay off.
The projection Φ : I × M → M defines an oriented ∂-bundle with standard fiber I.
For any ω ∈ Ω• (N ) we have the equality
Z
∗ ∗ ∗ ∗
φ1 (ω) − φ0 (ω) = Φ (ω) |1×M −Φ (ω) |0×M = Φ∗ (ω).
(∂I×M )/M

We now use the homotopy formula in Theorem 3.4.54 of Subsection 3.4.5, and we deduce
Z Z Z
∗ ∗
Φ (ω) = dI×M Φ (ω) − dM Φ∗ (ω)
(∂I×M )/M (I×M )/M (I×M )/M
Z Z
= Φ∗ (dN ω) − dM Φ∗ (ω).
(I×M )/M (I×M )/M

Thus Z
χ(ω) := Φ∗ (ω)
(I×M )/M

is the sought for cochain homotopy. ⊔

⊓

⊔
Corollary 7.1.27. Two homotopy equivalent spaces have isomorphic cohomology rings.⊓

Consider a smooth manifold M , and U , V two open subsets of M such that M = U ∪V .

Denote by ıU (respectively ıV ) the inclusions U ֒→ M (resp. V ֒→ M ). These induce the
restriction maps
ı∗U : Ω• (M ) → Ω• (U ), ω 7→ ω |U ,
and
ı∗V : Ω• (M ) → Ω• (V ), ω 7→ ω |V .
258 CHAPTER 7. COHOMOLOGY

We get a cochain morphism

r : Ω• (M ) → Ω• (U ) ⊕ Ω• (V ), ω 7→ (ı∗U ω, ı∗V ω).

There exists another cochain morphism

δ : Ω• (U ) ⊕ Ω• (V ) → Ω• (U ∩ V ), (ω, η) 7→ −ω |U ∩V +η |U ∩V .

Lemma 7.1.28. The short Mayer–Vietoris sequence

r δ
0 → Ω• (M ) −→ Ω• (U ) ⊕ Ω• (V ) −→ Ω∗ (U ∩ V ) → 0

is exact.

Proof. Obviously r is injective. The proof of the equality Range r = ker δ can be safely
left to the reader. The surjectivity of δ requires a little more effort.
The collection {U, V } is an open cover of M , so we can find a partition of unity
{ϕU , ϕV } ⊂ C ∞ (M ) subordinated to this cover, i.e.,

supp ϕU ⊂ U, supp ϕV ⊂ V, 0 ≤ ϕU , ϕV ≤ 1, ϕU + ϕV = 1.

Note that for any ω ∈ Ω∗ (U ∩ V ) we have

supp ϕV ω ⊂ supp ϕV ⊂ V,

and thus, upon extending ϕV ω by 0 outside V , we can view it as a form on U . Similarly,

ϕU ω ∈ Ω∗ (V ). Note that

δ(−ϕV ω, ϕU ω) = (ϕV + ϕU )ω = ω.

This establishes the surjectivity of δ. ⊔

⊓

Using the abstract results in homological algebra we deduce from the above lemma
the following fundamental result.
Theorem 7.1.29 (Mayer–Vietoris). Let M = U ∪ V be an open cover of the smooth
manifold M. Then there exists a long exact sequence
r δ ∂
· · · → H k (M ) → H k (U ) ⊕ H k (V ) → H k (U ∩ V ) → H k+1 (M ) → · · · ,

called the long Mayer–Vietoris sequence. ⊔

⊓

The connecting morphisms ∂ can be explicitly described using the prescriptions fol-
lowing Proposition 7.1.18 in the previous subsection. Let us recall that construction.
Start with ω ∈ Ωk (U ∩ V ) such that dω = 0. Writing as before

ω = ϕV ω + ϕU ω,

we deduce
d(ϕV ω) = d(−ϕU ω) on U ∩ V.
7.1. DERHAM COHOMOLOGY 259

Thus, we can find η ∈ Ωk+1 (M ) such that

η |U = d(ϕV ω) η |V = d(−ϕU ω).

Then
∂ω := η.
The reader can prove directly that the above definition is independent of the various
choices.
The Mayer–Vietoris sequence has the following functorial property.
Proposition 7.1.30. Let φ : M → N be a smooth map and {U, V } an open cover of N .
Then U ′ = φ−1 (U ), V ′ = φ−1 (V ) form an open cover of M and moreover, the diagram
below is commutative.

H k (N )
r
w H k (U ) ⊕ H k (V ) δ
w H k (U ∩ V ) ∂
w H k+1(N )
φ∗ φ∗ φ∗ φ∗ ⊔
⊓
u u u u
H k (M )
r′
w H k (U ′) ⊕ H k (V ′) δ′
w H k (U ′ ∩ V ′) ∂′
w H k+1(M )
Exercise 7.1.31. Prove the above proposition. ⊔
⊓

7.1.5 Some simple examples

The Mayer–Vietoris theorem established in the previous subsection is a very powerful tool
for computing the cohomology of manifolds. In principle, it allows one to recover the
cohomology of a manifold decomposed into simpler parts, knowing the cohomologies of its
constituents. In this subsection we will illustrate this principle on some simple examples.
Example 7.1.32. (The cohomology of spheres). The cohomology of S 1 can be easily
computed using the definition of DeRham cohomology. We have H 0 (S 1 ) = R since S 1 is
connected. A 1-form η ∈ Ω1 (S 1 ) is automatically closed, and it is exact if and only if
Z
η = 0.
S1

Indeed, if η = dF , where F : R → R is a smooth 2π-periodic function, then

Z
η = F (2π) − F (0) = 0.
S1

Conversely if η = f (θ)dθ, where f : R → R is a smooth 2π-periodic function and

Z 2π
f (θ)dθ = 0,
0

then the function Z t

F (t) = f (s)ds
0
260 CHAPTER 7. COHOMOLOGY

is smooth, 2π-periodic, and dF = η.

Thus, the map Z Z
1 1
: Ω (S ) → R, η 7→ η,
S1 S1

induces an isomorphism H 1 (S 1 ) → R, and we deduce

PS 1 (t) = 1 + t.

To compute the cohomology of higher dimensional spheres we use the Mayer–Vietoris

theorem.
The (n + 1)-dimensional sphere S n+1 can be covered by two open sets

Usouth = S n+1 \ {North pole} and Unorth = S n+1 \ {South pole}.

Each is diffeomorphic to Rn+1 . Note that the overlap Unorth ∩ Usouth is homotopically
equivalent with the equator S n . The Poincaré lemma implies that

H k+1 (Unorth ) ⊕ H k+1 (Usouth ) ∼

=0

for k ≥ 0. The Mayer–Vietoris sequence gives

H k (Unorth ) ⊕ H k (Usouth ) → H k (Unorth ∩ Usouth ) → H k+1 (S n+1 ) → 0.

For k > 0 the group on the left is also trivial, so that we have the isomorphisms

H k (S n ) ∼
= H k (Unorth ∩ Usouth ) ∼
= H k+1 (S n+1 ) k > 0.

Denote by Pn (t) the Poincaré polynomial of S n and set Qn (t) = Pn (t) − Pn (0) = Pn (t) − 1.
We can rewrite the above equality as

Qn+1 (t) = tQn (t) n > 0.

Since Q1 (t) = t we deduce Qn (t) = tn , i.e.,

PS n (t) = 1 + tn . ⊔
⊓

Example 7.1.33. Let {U, V } be an open cover of the smooth manifold M. We assume
that all the Betti numbers of U , V and U ∩ V are finite. Using the Mayer–Vietoris short
exact sequence, and the Exercise 7.1.21 in Subsection 7.1.3, we deduce that all the Betti
numbers of M are finite, and moreover

χ(M ) = χ(U ) + χ(V ) − χ(U ∩ V ). (7.1.12)

This resembles very much the classical inclusion-exclusion principle in combinatorics. We

will use this simple observation to prove that the Betti numbers of a connected sum of g
tori is finite, and then compute its Euler characteristic.
Let Σ be a surface with finite Betti numbers. From the decomposition

Σ = (Σ \ disk) ∪ disk,
7.1. DERHAM COHOMOLOGY 261

we deduce (using again Exercise 7.1.21)

χ(Σ) = χ(Σ \ disk) + χ(disk) − χ ( (Σ \ disk) ∩ (disk) ) .

Since (Σ \ disk) ∩ disk is homotopic to a circle, and χ(disk) = 1, we deduce

χ(Σ) = χ(Σ \ disk) + 1 − χ(S 1 ) = χ(Σ \ disk) + 1.

If now Σ1 and Σ2 are two surfaces with finite Betti numbers, then

Σ1 #Σ2 = (Σ1 \ disk) ∪ (Σ2 \ disk),

where the two holed surfaces intersect over an entire annulus, which is homotopically a
circle. Thus
χ(Σ1 #Σ2 ) = χ(Σ1 \ disk) + χ(Σ2 \ disk) − χ(S 1 )
= χ(Σ1 ) + χ(Σ2 ) − 2.
This equality is identical with the one proved in Proposition 4.2.42 of Subsection 4.2.5.
We can decompose a torus as a union of two cylinders. The intersection of these
cylinders is the disjoint union of two annuli so homotopically, this overlap is a disjoint
union of two circles. In particular, the Euler characteristic of the intersection is zero.
Hence
χ(torus) = 2χ(cylinder) = 2χ(circle) = 0.
We conclude as in Proposition 4.2.42 that

χ(connected sum of g tori) = 2 − 2g.

This is a pleasant, surprising connection with the Gauss–Bonnet theorem. And the story
is not over. ⊔
⊓

7.1.6 The Mayer–Vietoris principle

We describe in this subsection a “patching” technique which is extremely versatile in estab-
lishing general homological results about arbitrary manifolds building up from elementary
ones.
Definition 7.1.34. A smooth manifold M is said to be of finite type if it can be covered
by finitely many open sets U1 , . . . , Um such that any nonempty intersection Ui1 ∩ · · · ∩ Uik
(k ≥ 1) is diffeomorphic to Rdim M . Such a cover is said to be a good cover. ⊔
⊓

Example 7.1.35. (a) All compact manifolds are of finite type. To see this, it suffices
to cover such a manifold by finitely many open sets which are geodesically convex with
respect to some Riemann metric.
(b) If M is a finite type manifold, and U ⊂ M is a closed subset homeomorphic with
the closed unit ball in Rdim M , then M \ U is a finite type non-compact manifold. (It
suffices to see that Rn \ closed ball is of finite type).
(c) The connected sums, and the direct products of finite type manifolds are finite
type manifolds. ⊔
⊓
262 CHAPTER 7. COHOMOLOGY

Proposition 7.1.36. Let p : E → B be a smooth vector bundle. If the base B is of finite

type, then so is the total space E.

In the proof of this proposition we will use the following fundamental result.

Lemma 7.1.37. Let p : E → B be a smooth vector bundle such that B is diffeomorphic

to Rn . Then p : E → B is a trivializable bundle. ⊔
⊓

Proof of Proposition 7.1.36. Denote by F the standard fiber of E. The fiber F

is a vector space. Let (Ui )1≤i≤ν be a good cover of B. For each ordered multi-index
I := {i1 < · · · < ik } denote by UI the multiple overlap Ui1 ∩ · · · ∩ Uik . Using the previous
lemma we deduce that each EI = E |UI is a product F × Ui , and thus it is diffeomorphic
with some vector space. Hence (Ei ) is a good cover. ⊔
⊓

Exercise 7.1.38. Prove Lemma 7.1.37.

Hint: Assume that E is a vector bundle over the unit open ball B ⊂ Rn . Fix a connection
∇ on E, and then use the ∇-parallel transport along the half-lines Lx , [0, ∞) ∋ t 7→ tx ∈
Rn , x ∈ Rn \ {0}. ⊔
⊓

✍ We denote by Mn the category finite type smooth manifolds of dimension n. The

morphisms of this category are the smooth embeddings, i.e., the one-to-one immersions
M1 ֒→ M2 , Mi ∈ Mn .

Definition 7.1.39. Let R be a commutative ring with 1. A contravariant Mayer–Vietoris

functor (or M V -functor for brevity) is a contravariant functor from the category Mn , to
the category of Z-graded R-modules
M
F = ⊕n∈Z Fn → GradR Mod, M 7→ Fn (M ),
n

with the following property. If {U, V } is a M V -cover of M ∈ Mn , i.e., U, V, U ∩ V ∈ Mn ,

then there exist morphisms of R-modules

∂n : Fn (U ∩ V ) → Fn+1 (M ),

such that the sequence below is exact

r∗ δ ∂
· · · → Fn (M ) → Fn (U ) ⊕ Fn (V ) → Fn (U ∩ V ) →
n
Fn+1 (M ) → · · · ,

where r ∗ is defined by
r ∗ = F(ıU ) ⊕ F(ıV ),
and δ is defined by
δ(x ⊕ y) = F(ıU ∩V )(y) − F(ıU ∩V )(x).
(The maps ı• denote natural embeddings.) Moreover, if N ∈ Mn is an open submanifold
of N , and {U, V } is an M V -cover of M such that {U ∩ N, V ∩ N } is an M V -cover of N ,
7.1. DERHAM COHOMOLOGY 263

then the diagram below is commutative.

Fn (U ∩ V )
∂n
w Fn+1(M )

u u
Fn (U ∩ V ∩ N )
∂n
w Fn+1(N )
The vertical arrows are the morphisms F(ı• ) induced by inclusions. ⊔
⊓

The covariant M V -functors are defined in the dual way, by reversing the orientation
of all the arrows in the above definition.
Definition 7.1.40. Let F, G be two contravariant M V -functors,

F, G : Mn → GradR Mod.

A correspondence 2 between these functors is a collection of R-module morphisms

M M M
φM = φnM : Fn (M ) −→ Gn (M ),
n∈Z

ϕ
one morphism for each M ∈ Mn , such that, for any embedding M1 ֒→ M2 , the diagram
below is commutative
Fn (M2 )
F(ϕ)
Fn (M1 ) w
φM 2 φM 1 ,
u u
Gn (M2 )
G(ϕ)
w Gn(M1)
and, for any M ∈ Mn and any M V -cover {U, V } of M , the diagram below is commutative.

Fn (U ∩ V )
∂n
w Fn+1(M )
φU ∩V φM

u u
n
G (U ∩ V )
∂n
wG n+1
(M )

The correspondence is said to be a natural equivalence if all the morphisms φM are iso-
morphisms. ⊔
⊓

Theorem 7.1.41 (Mayer–Vietoris principle). Let F, G be two (contravariant) Mayer–

Vietoris functors on Mn and φ : F → G a correspondence. If

φkRn : Fk (Rn ) → Gk (Rn )

is an isomorphism for any k ∈ Z, then φ is a natural equivalence. ⊔

⊓
2
Our notion of correspondence corresponds to the categorical notion of morphism of functors.
264 CHAPTER 7. COHOMOLOGY

Proof. The family of finite type manifolds Mn has a natural filtration

M1n ⊂ M2n ⊂ · · · ⊂ Mrn ⊂ · · · ,
where Mrn is the collection of all smooth manifolds which admit a good cover consisting
of at most r open sets. We will prove the theorem using an induction over r.
The theorem is clearly true for r = 1 by hypothesis. Assume φkM is an isomorphism
for all M ∈ Mnr−1 . Let M ∈ Mrn and consider a good cover {U1 , . . . , Ur } of M . Then
{U = U1 ∪ · · · ∪ Ur−1 , Ur }
is an M V -cover of M . We thus get a commutative diagram

Fn (U ) ⊕ Fn (Ur ) w Fn (U ∩ Ur ) ∂
w Fn+1(M ) w Fn+1(U ) ⊕ Gn+1(Ur )
u u u u n+1
Gn (U ) ⊕ Gn (Ur ) w Gn(U ∩ Ur ) ∂
w Gn+1(M ) wF n+1
(U ) ⊕ G (Ur )

The vertical arrows are defined by the correspondence φ. Note the inductive assumption
implies that in the above infinite sequence only the morphisms φM may not be isomor-
phisms. At this point we invoke the following technical result.
Lemma 7.1.42. (The five lemma.) Consider the following commutative diagram of
R-modules.
A−2 A−1 A0 wA1 w
A2 w w
u u u u u
f−2 f−1 f0 f1 f2

B−2 w B−1 w B0 w B1 w B2
If fi is an isomorphism for any i 6= 0, then so is f0 . ⊔
⊓

Exercise 7.1.43. Prove the five lemma. ⊔

⊓

The five lemma applied to our situation shows that the morphisms φM must be iso-
morphisms. ⊔
⊓

Remark 7.1.44. (a) The Mayer–Vietoris principle is true for covariant M V -functors as
well. The proof is obtained by reversing the orientation of the horizontal arrows in the
above proof.
(b) The Mayer–Vietoris principle can be refined a little bit. Assume that F and
G are functors from Mn to the category of Z-graded R-algebras, and φ : F → G is a
correspondence compatible with the multiplicative structures, i.e., each of the R-module
morphisms φM are in fact morphisms of R-algebras. Then, if φRn are isomorphisms of
Z-graded R-algebras , then so are the φM ’s, for any M ∈ Mn .
(c) Assume R is a field. The proof of the Mayer–Vietoris principle shows that if F is
a MV-functor and dimR F∗ (Rn ) < ∞ then dim F∗ (M ) < ∞ for all M ∈ Mn .
(d) The Mayer–Vietoris principle is a baby case of the very general technique in alge-
braic topology called the acyclic models principle. ⊔
⊓
7.1. DERHAM COHOMOLOGY 265

Corollary 7.1.45. Any finite type manifold has finite Betti numbers. ⊔
⊓

7.1.7 The Künneth formula

We learned in principle how to compute the cohomology of a “union of manifolds”. We will
now use the Mayer–Vietoris principle to compute the cohomology of products of manifolds.
Theorem 7.1.46 (Künneth formula). Let M ∈ Mm and N ∈ Mn . Then there exists a
natural isomorphism of graded R-algebras
!
M M
H (M × N ) ∼
• • ∗
= H (M ) ⊗ H (N ) = p q
H (M ) ⊗ H (N ) .
n≥0 p+q=n

In particular, we deduce
PM ×N (t) = PM (t) · PN (t).

Proof. We construct two functors

F, G : Mm → GradR Alg,
( )
M M M
F : M 7→ Fr (M ) = H p (M ) ⊗ H q (N ) ,
r≥0 r≥0 p+q=r

and M M
G : M 7→ Gr (M ) = H r (M × N ),
r≥0 r≥0

where !
M M
F(f ) = f ∗ |H p (M2 ) ⊗1H q (N ) , ∀f : M1 ֒→ M2 ,
r≥0 p+q=r

and M
G(f ) = (f × 1N )∗ |H r (M2 ×N ) , ∀f : M1 ֒→ M2 .
r≥0

We let the reader check the following elementary fact.

Exercise 7.1.47. F and G are contravariant M V -functors. ⊔
⊓

For M ∈ MM , define φM : F(M ) → G(N ) by

∗ ∗
φM (ω ⊗ η) = ω × η := πM ω ∧ πN η ( ω ∈ H • (M ), η ∈ H • (M ) ),

where πM (respectively πN ) are the canonical projections M × N → M (respectively

M × N → N ). The operation

× : H • (M ) ⊗ H • (N ) → H • (M × N ) (ω ⊗ η) 7→ ω × η,

is called the cross product. The Künneth formula is a consequence of the following lemma.
266 CHAPTER 7. COHOMOLOGY

Lemma 7.1.48. (a) φ is a correspondence of MV-functors.

(b) φRm is an isomorphism.

Proof. The only nontrivial thing to prove is that for any MV-cover {U, V } of M ∈ Mm
the diagram below is commutative.

⊕p+q=r H p (U ∩ V ) ⊗ H q (N )
∂
w ⊕p+q=r H p+1(M ) ⊗ H q (N )
φU ∩V φM
u u
H r ((U × N ) ∩ (V × N ))
∂′
w H r+1(M × N )
We briefly recall the construction of the connecting morphisms ∂ and ∂ ′ .
One considers a partition of unity {ϕU , ϕV } subordinated to the cover {U, V }. Then,
the functions ψU = πM ∗ ϕ and ψ = π ∗ ϕ form a partition of unity subordinated to the
U V M V
cover 
U × N, V × N
of M × N . If ω ⊗ η ∈ H • (U ∩ V ) ⊗ H • (N ),

∂(ω ⊗ η) = ω̂ ⊗ η,

where
ω̂ |U = −d(ϕV ω) ω̂ = d(ϕU ω).
On the other hand, φU ∩V (ω ⊗ η) = ω × η, and ∂ ′ (ω × η) = ω̂ × η. This proves (a).
To establish (b), note that the inclusion

 : N ֒→ Rm × N, x 7→ (0, x)

is a homotopy equivalence, with πN a homotopy inverse. Hence, by the homotopy invari-

ance of the DeRham cohomology we deduce

G(Rm ) ∼
= H • (N ).

Using the Poincaré lemma and the above isomorphism we can identify the morphism φRm
with 1Rm . ⊔
⊓

Example 7.1.49. Consider the n-dimensional torus, T n . By writing it as a direct product

of n circles we deduce from Künneth formula that

PT n (t) = {PS 1 (t)}n = (1 + t)n .

Thus  
n n
bk (T ) = , dim H ∗ (T n ) = 2n ,
k
and χ(T n ) = 0.
One can easily describe a basis of H ∗ (T n ). Choose angular coordinates (θ 1 , . . . , θ n ) on
n
T . For each ordered multi-index I = (1 ≤ i1 < · · · < ik ≤ n) we have a closed, non-exact
7.2. THE POINCARÉ DUALITY 267

form dθ I . These monomials are linearly independent (over R) and there are 2n of them.
Thus, they form a basis of H ∗ (T n ). In fact, one can read the multiplicative structure
using this basis. We have an isomorphism of R-algebras

H • (T n ) ∼
= Λ• Rn . ⊔
⊓

Exercise 7.1.50. Let M ∈ Mm and N ∈ Mn . Show that for any ωi ∈ H ∗ (M ), ηj ∈

H ∗ (N ), i, j = 0, 1, the following equality holds.

(ω0 × η0 ) ∧ (ω1 × η1 ) = (−1)deg η0 deg ω1 (ω0 ∧ ω1 ) × (η0 ∧ η1 ). ⊔

⊓

Exercise 7.1.51. (Leray-Hirsch). Let p : E → M be smooth bundle with standard

fiber F . We assume the following:

(a) Both M and F are of finite type.

(b) There exist cohomology classes e1 , . . . , er ∈ H • (E) such that their restrictions to
any fiber generate the cohomology algebra of that fiber.

The projection p induces a H • (M )-module structure on H • (E) by

ω · η = p∗ ω ∧ η, ∀ω ∈ H • (M ), η ∈ H • (E).

Show that H • (E) is a free H • (M )-module with generators e1 , . . . , er . ⊔

⊓

7.2 The Poincaré duality

7.2.1 Cohomology with compact supports
Let M be a smooth n-dimensional manifold. Denote by Ωkcpt(M ) the space of smooth
compactly supported k-forms. Then
d d
0 → Ω0cpt (M ) → · · · → Ωncpt(M ) → 0

is a cochain complex. Its cohomology is denoted by Hcpt• (M ), and it is called the DeR-

ham cohomology with compact supports. Note that when M is compact this cohomology
coincides with the usual DeRham cohomology.
Although it looks very similar to the usual DeRham cohomology, there are many
important differences. The most visible one is that if φ : M → N is a smooth map,
and ω ∈ Ω•cpt (N ), then the pull-back φ∗ ω may not have compact support, so this new
construction is no longer a contravariant functor from the category of smooth manifolds
and smooth maps, to the category of graded vector spaces.
On the other hand, if dim M = dim N , and φ is an embedding, we can identify M with
an open subset of N , and then any η ∈ Ω•cpt(M ) can be extended by 0 outside M ⊂ N .
This extension by zero defines a push-forward map

φ∗ : Ω•cpt(M ) → Ω•cpt (N ).
268 CHAPTER 7. COHOMOLOGY

One can verify easily that φ∗ is a cochain map so that it induces a morphism
• •
φ∗ : Hcpt (M ) → Hcpt (N ).
• is a covariant functor from the category M ,
In terms of our category Mn we see that Hcpt n
to the category of graded real vector spaces. As we will see, it is a rather nice functor.
• is a covariant M V -functor, and moreover
Theorem 7.2.1. Hcpt

k 0 , k<n
Hcpt (Rn ) = .
R , k=n

The last assertion of this theorem is usually called the Poincaré lemma for compact
supports.
We first prove the Poincaré lemma for compact supports. The crucial step is the
following technical result that we borrowed from [14].
p
Lemma 7.2.2. Let E → B be a rank r real vector bundle which is orientable in the sense
described in Subsection 3.4.5. Denote by p∗ the integration-along-fibers map

p∗ : Ω•cpt(E) → Ω•−r
cpt (B).

Then there exists a smooth bilinear map

m : Ωicpt(E) × Ωjcpt(E) → Ωi+j−r−1

cpt (E)

such that,

p∗ p∗ α ∧ β − α ∧ p∗ p∗ β = (−1)r d(m(α, β)) − m(dα, β) + (−1)deg α m(α, dβ).

Proof. Consider the ∂-bundle

π : E = I × (E ⊕ E) → E, π : (t; v0 , v1 ) 7→ (t; v0 + t(v1 − v0 )).

Note that
∂E = ( {0} × (E ⊕ E) ) ⊔ ({1} × (E ⊕ E)) .
Define π t : E ⊕ E → E as the composition
π
E⊕E ∼
= {t} × E ⊕ E ֒→ I × (E ⊕ E) → E.

Observe that
∂π = π |∂E = (−π 0 ) ⊔ π 1 .
For (α, β) ∈ Ω•cpt(E) × Ω•cpt(E) define α ⊙ β ∈ Ω•cpt(E ⊕ E) by

α ⊙ β := (π 0 )∗ α ∧ (π 1 )∗ β.
7.2. THE POINCARÉ DUALITY 269

(Verify that the support of α ⊙ β is indeed compact.) For α ∈ Ωicpt(E), and β ∈ Ωjcpt(E)
we have the equalities

p∗ p∗ α ∧ β = π∗1 (α ⊙ β) ∈ Ωi+j−r
cpt (E), α ∧ p∗ p∗ β = π∗0 (α ⊙ β) ∈ Ωi+j−r
cpt (E).

Hence
Z
∗ ∗
D(α, β) = p p∗ α ∧ β − α ∧ p p∗ β = π∗1 (α ⊙ β) − π∗0 (α ⊙ β) = α ⊙ β.
∂E/E

We now use the fibered Stokes formula to get

Z Z
∗ r
D(α, β) = dE T (α ⊙ β) + (−1) dE T ∗ (α ⊙ β),
E/E E/E

where T is the natural projection E = I × (E ⊕ E) → E ⊕ E. The lemma holds with

Z
m(α, β) = T ∗ (α ⊙ β). ⊔
⊓
E/E

Proof of the Poincaré lemma for compact supports. Consider δ ∈ C0∞ (Rn ) such
that Z
0 ≤ δ ≤ 1, δ(x)dx = 1.
Rn
Define the, closed, compactly supported n-form

τ := δ(x)dx1 ∧ · · · ∧ dxn .

We want to use Lemma 7.2.2 in which E is the rank n bundle over a point, i.e., E =
p
{pt} × Rn → {pt}. The integration along fibers is simply the integration map.

0 , deg ω < n
p∗ : Ωcpt(R ) → R, ω 7→ p∗ ω = R
• n
Rn ω , deg ω = n.

If now ω is a closed, compactly supported form on Rn , we have

ω = (p∗ p∗ τ ) ∧ ω.

Using Lemma 7.2.2 we deduce

ω − τ ∧ p∗ p∗ ω = (−1)n dm(τ, ω).

Thus any closed, compactly supported form ω on Rn is cohomologous to τ ∧ p∗ p∗ ω. The

latter
R is always zero if deg ω < n. When deg ω = n we deduce that ω is cohomologous to
( ω)τ . This completes the proof of the Poincaré lemma. ⊔
⊓

To finish the proof of Theorem 7.2.1 we must construct a Mayer–Vietoris sequence.

Let M be a smooth manifold decomposed as an union of two open sets M = U ∪ V , The
sequence of inclusions
U ∩ V ֒→ U, V ֒→ M,
270 CHAPTER 7. COHOMOLOGY

induces a short sequence

i j
0 → Ω•cpt (U ∩ V ) → Ω•cpt (U ) ⊕ Ω•cpt (V ) → Ω•cpt(M ) → 0,

where
i(ω) = (ω̂, ω̂), j(ω, η) = η̂ − ω̂.
The hat ˆ denotes the extension by zero outside the support. This sequence is called the
Mayer–Vietoris short sequence for compact supports.
Lemma 7.2.3. The above Mayer–Vietoris sequence is exact.

Proof. The morphism i is obviously injective. Clearly, Range (i) = ker(j). We have to
prove that j is surjective.
Let (ϕU , ϕV ) a partition of unity subordinated to the cover {U, V }. Then, for any
η ∈ Ω∗cpt(M ), we have
ϕU η ∈ Ω•cpt(U ) and ϕV η ∈ Ω•cpt (V ).
In particular, η = j(−ϕU η, ϕV η), which shows that j is surjective. ⊔
⊓

We get a long exact sequence called the long Mayer–Vietoris sequence for compact
supports.

k k k k k+1 δ
· · · → Hcpt (U ∩ V ) → Hcpt (U ) ⊕ Hcpt (V ) → Hcpt (M ) → Hcpt → ···

The connecting homomorphism can be explicitly described as follows. If ω ∈ Ωkcpt(M ) is

a closed form then
d(ϕU ω) = d(−ϕV ω) on U ∩ V.
We set δω := d(ϕU ω). The reader can check immediately that the cohomology class of δω
is independent of all the choices made.
If φ : N ֒→ M is a morphism of Mn then for any MV-cover of {U, V } of M {φ−1 (U ), φ−1 (V )}
is an MV-cover of N . Moreover, we almost tautologically get a commutative diagram

k
Hcpt (N )
δ
w Hcpt
k+1 −1
(φ (U ∩ V ))

φ∗
u u
φ∗

k
Hcpt (M )
δ
w Hcpt
k+1
(U ∩ V )

• is a covariant Mayer–Vietoris sequence.

This proves Hcpt ⊔
⊓

Remark 7.2.4. To be perfectly honest (from a categorical point of view) we should have
considered the chain complex M M
Ω̃k = Ω−k
cpt ,
k≤0 k≤0

and correspondingly the associated homology

−•
H̃• := Hcpt .
7.2. THE POINCARÉ DUALITY 271

This makes the sequence

0 ← Ω̃−n ← · · · ← Ω̃−1 ← Ω̃0 ← 0
a chain complex ,and its homology H̃• is a bona-fide covariant Mayer–Vietoris functor
since the connecting morphism δ goes in the right direction H̃• → H̃•−1 . However, the
simplicity of the original notation is worth the small formal ambiguity, so we stick to our
upper indices. ⊔
⊓

From the proof of the Mayer–Vietoris principle we deduce the following.

k (M ) < ∞.
Corollary 7.2.5. For any M ∈ Mn , and any k ≤ n we have dim Hcpt ⊔
⊓

7.2.2 The Poincaré duality

Definition 7.2.6. Denote by M+ n the category of n-dimensional, finite type, oriented
manifolds. The morphisms are the embeddings of such manifold. The M V functors on
M+
n are defined exactly as for Mn . ⊔
⊓

Given M ∈ M+
n , there is a natural pairing
n−k
h•, •iκ : Ωk (M ) × Ωcpt (M ) → R,
defined by Z
hω, ηiκ := ω ∧ η.
M
This pairing is called the Kronecker pairing. We can extend this pairing to any (ω, η) ∈
Ω• × Ω•cpt as

0 , deg ω + deg η 6= n
hω, ηiκ = R .
M ω ∧ η , deg ω + deg η = n
The Kronecker pairing induces maps
n−k
D = Dk : Ωk (M ) → (Ωcpt (M ))∗ , hD(ω), ηi = hω, ηiκ .
Above, h•, •i denotes the natural pairing between a vector space V and its dual V ∗ ,
h•, •i : V ∗ × V → R.
n−k
If ω is closed, the restriction of D(ω) to the space Zcpt of closed, compactly supported
n−k
(n − k)-forms vanishes on the subspace Bcpt = dΩn−k−1
cpt (M ). Indeed, if η = dη ′ , η ′ ∈
Ωn−k−1
cpt (M ) then
Z Z
Stokes
hD(ω), ηi = ω ∧ dη ′ = ± dω ∧ η ′ = 0.
M M
n−k
Thus, if ω is closed, the linear functional D(ω) defines an element of (Hcpt (M ))∗ . If
n−k
moreover ω is exact, a computation as above shows that D(ω) = 0 ∈ (Hcpt (M ))∗ . Hence
D descends to a map in cohomology
D : H k (M ) → (Hcpt
n−k
(M ))∗ .
272 CHAPTER 7. COHOMOLOGY

Equivalently, this means that the Kronecker pairing descends to a pairing in cohomology,

h•, •iκ : H k (M ) × Hcpt

n−k
(M ) → R.

Theorem 7.2.7 (Poincaré duality). The Kronecker pairing in cohomology is a duality for
all M ∈ M+
n.

Proof. The functor M+

n → Graded Vector Spaces defined by
M M
∗
n−k
M→ H̃ k (M ) = Hcpt (M )
k k

is a contravariant M V -functor. (The exactness of the Mayer–Vietoris sequence is pre-

served by transposition. This is where the fact that all the cohomology groups are finite
dimensional vector spaces plays a very important role).
For purely formal reasons which will become apparent in a little while, we re-define
the connecting morphism
d˜
H̃ k (U ∩ V ) → H̃ k+1 (U ∪ V ),
n−k−1 n−k δ
to be (−1)k δ† , where Hcpt (U ∩ V ) → Hcpt (U ∪ V ) denotes the connecting morphism
in the DeRham cohomology with compact supports, and δ† denotes its transpose.
The Poincaré lemma for compact supports can be rephrased

k n R , k=0
H̃ (R ) = .
0 , k>0

The Kronecker pairing induces linear maps

DM : H k (M ) → H̃ k (M ).
L
Lemma 7.2.8. k Dk is a correspondence of M V functors.

Proof. We have to check two facts.

φ
Fact A. Let M ֒→ N be a morphism in M+
n . Then the diagram below is commutative.

H k (N )
φ∗
w H k (M )
DN
u DM
u .
H̃ k (N )
φ̃∗
w H̃ k (M )
Fact B. If {U, V } is an M V -cover of M ∈ M+
n , then the diagram bellow is commutative

H k (U ∩ V )
∂
w H k+1(M )
.
u u
DU ∩V DM

H̃ k (U ∩ V )
(−1)k δ†
w H̃ k+1(M )
7.2. THE POINCARÉ DUALITY 273

Proof of Fact A. Let ω ∈ H k (N ). Denoting by h•, •i the natural duality between a

n−k
vector space and its dual we deduce that for any η ∈ Hcpt (M ) we have

φ̃∗ ◦ DN (ω), η = (φ∗ )† DN (ω), η = DN (ω), φ∗ η

Z Z Z
= ω ∧ φ∗ η = ω |M ∧η = φ∗ ω ∧ η = hDM (φ∗ ω), ηi.
N M ֒→N M

Hence φ̃∗ ◦ DN = DM ◦ φ∗ .
Proof of Fact B. Let ϕU , ϕV be a partition of unity subordinated to the M V -cover
{U, V } of M ∈ M+ k
n . Consider a closed k-form ω ∈ Ω (U ∩ V ). Then the connecting
morphism in usual DeRham cohomology acts as

d(−ϕV ω) on U
∂ω = .
d(ϕU ω) on V

Choose η ∈ Ωn−k−1
cpt (M ) such that dη = 0. We have
Z Z Z Z
hDM ∂ω, ηi = ∂ω ∧ η = ∂ω ∧ η + ∂ω ∧ η − ∂ω ∧ η
M U V U ∩V
Z Z Z
=− d(ϕV ω) ∧ η + dϕU ω) ∧ η + d(ϕV ω) ∧ η.
U V U ∩V
Note that the first two integrals vanish. Indeed, over U we have the equality


d(ϕV ω) ∧ η = d ( ϕV ω ∧ η ) ,

and the vanishing now follows from Stokes formula. The second term is dealt with in a
similar fashion. As for the last term, we have
Z Z Z
deg ω
d(ϕV ω) ∧ η = dϕV ∧ ω ∧ η = (−1) ω ∧ (dϕV ∧ η)
U ∩V U ∩V U ∩V
Z
= (−1)k ω ∧ δη = (−1)k hDU ∩V ω, δηi = h(−1)k δ† DU ∩V ω, ηi.
U ∩V
This concludes the proof of Fact B. The Poincaré duality now follows from the Mayer–
Vietoris principle. ⊔
⊓

Remark 7.2.9. Using the Poincaré duality we can associate to any smooth map f : M →
N between compact oriented manifolds of dimensions m and respectively n a natural push-
forward or Gysin map

f∗ : H • (M ) → H •+q (N ), q := dim N − dim M = n − m)

defined by the composition

D (f ∗ )† D −1
H • (M ) →
M
(H m−• (M ))∗ → (H m−• (N ))∗ →
N
H •+n−m (N ),

where (f ∗ )† denotes the transpose of the pullback morphism. ⊔

⊓
274 CHAPTER 7. COHOMOLOGY

Corollary 7.2.10. If M ∈ M+
n then

•
Hcpt (M ) ∼
= (H n−k (M ))• .

k (M ) is finite dimensional, the transpose

Proof. Since Hcpt

DtM : (Hcpt
n−k
(M ))∗∗ → (H k (M ))∗

is an isomorphism. On the other hand, for any finite dimensional vector space there exists
a natural isomorphism
V ∗∗ ∼
= V. ⊔
⊓

Corollary 7.2.11. Let M be a compact, connected, oriented, n-dimensional manifold.

Then the pairing Z
k n−k
H (M ) × H (M ) → R (ω, η) 7→ ω∧η
M

is a duality. In particular, bk (M ) = bn−k (M ), ∀k. ⊔

⊓

If M is connected H 0 (M ) ∼= H n (M ) ∼
= R so that H n (M ) is generated by any volume
form defining the orientation.
The symmetry of Betti numbers can be translated in the language of Poincaré poly-
nomials as
1
tn PM ( ) = PM (t). (7.2.1)
t
Example 7.2.12. Let Σg denote the connected sum of g tori. We have shown that

χ(Σg ) = b0 − b1 + b2 = 2 − 2g.

Since Σg is connected, the Poincaré duality implies b2 = b0 = 1. Hence b1 = 2g i.e.

PΣg (t) = 1 + 2gt + t2 . ⊔

⊓

Consider now a compact oriented smooth manifold such that dim M = 2k. The Kro-
necker pairing induces a non-degenerate bilinear form
Z
k k
I : H (M ) × H (M ) → R I(ω, η) = ω ∧ η.
M

The bilinear form I is called the cohomological intersection form of M .

When k is even (so that n is divisible by 4) I is a symmetric form. Its signature is
called the signature of M , and it is denoted by σ(M ).
When k is odd, I is skew-symmetric, i.e., it is a symplectic form. In particular,
H 2k+1 (M ) must be an even dimensional space.

Corollary 7.2.13. For any compact manifold M ∈ M+

4k+2 the middle Betti number
b2k+1 (M ) is even. ⊔
⊓
7.3. INTERSECTION THEORY 275

Exercise 7.2.14. (a) Let P ∈ Z[t] be an odd degree polynomial with non-negative integer
coefficients such that P (0) = 1. Show that if P satisfies the symmetry condition (7.2.1)
there exists a compact, connected, oriented manifold M such that PM (t) = P (t).
(b) Let P ∈ Z[t] be a polynomial of degree 2k with non-negative integer coefficients.
Assume P (0) = 1 and P satisfies (7.2.1). If the coefficient of tk is even then there exists
a compact connected manifold M ∈ M+ 2k such that PM (t) = P (t).
Hint: Describe the Poincaré polynomial of a connect sum in terms of the polynomials of
its constituents. Combine this fact with the Künneth formula. ⊔
⊓

Remark 7.2.15. The result in the above exercise is sharp. Using his intersection theorem
F. Hirzebruch showed that there exist no smooth manifolds M of dimension 12 or 20 with
Poincaré polynomials 1 + t6 + t12 and respectively 1 + t10 + t20 . Note that in each of these
cases both middle Betti numbers are odd. For details we refer to J. P. Serre, “Travaux de
Hirzebruch sur la topologie des variétés”, Seminaire Bourbaki 1953/54,n◦ 88. ⊔
⊓

7.3 Intersection theory

7.3.1 Cycles and their duals
Suppose that M is a smooth manifold.
Definition 7.3.1. A k-dimensional cycle in M is a pair (S, φ), where S is a compact,
oriented k-dimensional manifold without boundary, and φ : S → M is a smooth map. We
denote by Ck (M ) the set of k-dimensional cycles in M . ⊔
⊓

Definition 7.3.2. (a) Two cycles (S0 , φ0 ), (S1 , φ1 ) ∈ Ck (M ) are said to be cobordant,
and we write this (S0 , φ0 ) ∼c (S1 , φ1 ), if there exists a compact, oriented manifold with
boundary Σ, and a smooth map Φ : Σ → M such that the following hold.
(a1) ∂Σ = (−S0 ) ⊔ S1 where −S0 denotes the oriented manifold S0 equipped with the
opposite orientation, and “⊔” denotes the disjoint union.
(a2) Φ |Si = φi , i = 0, 1.
(b) A cycle (S, φ) ∈ Ck (M ) is called trivial if there exists a (k + 1)-dimensional,
oriented manifold Σ with (oriented) boundary S, and a smooth map Φ : Σ → M such
that Φ|∂Σ = φ. We denote by Tk (M ) the set of trivial cycles.
(c) A cycle (S, φ) ∈ Ck (M ) is said to be degenerate if it is cobordant to a constant
cycle, i.e., a cycle (S ′ , φ′ ) such that φ′ is map constant on the components of S ′ . We
denote by Dk (M ) the set of degenerate cycles. ⊔
⊓

Exercise 7.3.3. Let (S0 , φ0 ) ∼c (S1 , φ1 ). Prove that (−S0 ⊔ S1 , φ0 ⊔ φ1 ) is a trivial cycle.
⊔
⊓

The cobordism relation on Ck (M ) is an equivalence relation3 , and we denote by Zk (M )

the set of equivalence classes. For any cycle (S, φ) ∈ Ck (M ) we denote by [S, φ] its image
in Zk (M ). Since M is connected, all the trivial cycles are cobordant, and they define an
element in Zk (M ) which we denote by [0].
3
We urge the reader to supply a proof of this fact.
276 CHAPTER 7. COHOMOLOGY

S0
S
1

Figure 7.1: A cobordism in R3

Given [S0 , φ0 ], [S1 , φ1 ] ∈ Zk (M ) we define

[S0 , φ0 ] + [S1 , φ1 ] := [S0 ⊔ S1 , φ0 ⊔ φ1 ],

where “⊔” denotes the disjoint union.

Proposition 7.3.4. Suppose M is a smooth manifold.
(a) Let [Si , φi ], [Si′ , φ′i ] ∈ Ck (M ), i = 0, 1. If [Si , φi ] ∼c [Si′ , φ′i ], for i = 0, 1, then

[S0 ⊔ S1 , φ0 ⊔ φ1 ] ∼c [S0′ ⊔ S1′ , φ′0 ⊔ φ′1 ]

so that the above map + : Zk (M ) × Zk (M ) → Zk (M ) is well defined.

(b) The binary operation + induces a structure of Abelian group on Zk (M ). The trivial
element is represented by the trivial cycles. Moreover,

−[S, φ] = [−S, φ] ∈ Hk (M ). ⊔
⊓

Exercise 7.3.5. Prove the above proposition. ⊔

⊓

We denote by Hk (M ) the quotient of Zk (M ) modulo the subgroup generated by the

degenerate cycles. Let us point out that any trivial cycle is degenerate, but the converse
is not necessarily true.
Suppose M ∈ M+ k
n . Any k-cycle (S, φ) defines a linear map H (M ) → R given by
Z
k
H (M ) ∋ ω 7→ φ∗ ω.
S

Stokes formula shows that this map is well defined, i.e., it is independent of the closed
form representing a cohomology class.
Indeed, if ω is exact, i.e., ω = dω ′ , then
Z Z
φ∗ dω ′ = dφ∗ ω ′ = 0.
S S
7.3. INTERSECTION THEORY 277

In other words, each cycle defines an element in ( H k (M ) )∗ . Via the Poincaré duality we
n−k n−k
identify the vector space ( H k (M ) )∗ with Hcpt (M ). Thus, there exists δS ∈ Hcpt (M )
such that Z Z
ω ∧ δS = φ∗ ω ∀ω ∈ H k (M ).
M S
The compactly supported cohomology class δS is called the Poincaré dual of (S, φ).
n−k
There exist many closed forms η ∈ Ωcpt (M ) representing δS . When there is no risk
of confusion, we continue denote any such representative by δS .

Figure 7.2: The dual of a point is Dirac’s distribution

Example 7.3.6. Let M = Rn , and S is a point, S = {pt} ⊂ Rn . pt is canonically a

0-cycle. Its Poincaré dual is a compactly supported n-form ω such that for any constant
λ (i.e. closed 0-form) Z Z
λω = λ = λ,
Rn pt
i.e., Z
ω = 1.
Rn
Thus δpt can be represented by any compactly supported n-form with integral 1. In
particular, we can choose representatives with arbitrarily small supports. Their “profiles”
look like in Figure 7.2. “At limit” they approach Dirac’s delta distribution. ⊔
⊓

Example 7.3.7. Consider an n-dimensional, compact, connected, oriented manifold M .

We denote by [M ] the cycle (M, 1M ). Then δ[M ] = 1 ∈ H 0 (M ). ⊔
⊓

✍ For any differential form ω, we set (for typographical reasons), |ω| := deg ω.

Example 7.3.8. Consider the manifolds M ∈ M+ +

m and N ∈ Mn . (The manifolds M and
N need not be compact.) To any pair of cycles (S, φ) ∈ Cp (M ), and (T, ψ) ∈ Cq (N ) we
can associate the cycle (S × T, φ × ψ) ∈ Cp+q (M × N ). We denote by πM (respectively
πN ) the natural projection M × N → M (respectively M × N → N ). We want to prove
the equality
δS×T = (−1)(m−p)q δS × δT , (7.3.1)
278 CHAPTER 7. COHOMOLOGY

where
∗ ∗
ω × η := πM ω ∧ πN η, ∀(ω, η) ∈ Ω• (M ) × Ω• (N ).
Pick (ω, η) ∈ Ω• (M ) × Ω• (N ) such that, deg ω + deg η = (m + n) − (p + q). Then, using
Exercise 7.1.50
Z Z
(m−p)|η|
(ω × η) ∧ (δS × δT ) = (−1) (ω ∧ δS ) × (η ∧ δT ).
M ×N M ×N

The above integral should be understood in the generalized sense of Kronecker pairing.
The only time when this pairing does not vanish is when |ω| = p and |η| = q. In this case
the last term equals
Z  Z  Z  Z 
q(m−p) q(m−p) ∗ ∗
(−1) ω ∧ δS η ∧ δT = (−1) φ ω ψ η
S T S T
Z
= (−1)q(m−p) (φ × ψ)∗ (ω ∧ η).
S×T
This establishes the equality (7.3.1). ⊔
⊓

Example 7.3.9. Consider a compact manifold M ∈ M+ •

n . Fix a basis (ωi ) of H (M ) such
that each ωi is homogeneous of degree |ωi | = di , and denote by ω i the basis of H • (M )
dual to (ωi ) with respect to the Kronecker pairing, i.e.,
i |·|ω
hω i , ωj iκ = (−1)|ω j|
hωj , ω i iκ = δji .

In M × M there exists a remarkable cycle, the diagonal

∆ = ∆M : M → M × M, x 7→ (x, x).

We claim that the Poincaré dual of this cycle is

X i
δ∆ = δ M = (−1)|ω | ω i × ωi . (7.3.2)
i

Indeed, for any homogeneous forms α, β ∈ Ω• (M ) such that |α| + |β| = n, we have
Z X Z
|ω i |
(α × β) ∧ δ M = (−1) (α × β) ∧ (ω i × ωi )
M ×M i M ×M

X Z
i i
= (−1)|ω | (−1)|β|·|ω | (α ∧ ω i ) × (β ∧ ωi )
i M ×M

X Z  Z 
i i
= (−1)|ω | (−1)|β|·|ω | α∧ω i
β ∧ ωi .
i M M

The i-th summand is nontrivial only when |β| = |ω i |, and |α| = |ωi |. Using the equality
|ω i | + |ω i |2 ≡ 0(mod 2) we deduce
Z X Z  Z  X
i
(α × β) ∧ δ M = α∧ω β ∧ ωi = hα, ω i iκ hβ, ωi iκ .
M ×M i M M i
7.3. INTERSECTION THEORY 279

From the equalities

X X
α= ωi hω i , αiκ and β = hβ, ωj , iκ ω j
i j

we conclude
Z Z Z X X
∗
∆ (α × β) = α∧β = ( ωi hω i , αiκ ) ∧ ( hβ, ωj iκ ω j )
M M M i j

Z X X
= hω i , αiκ hβ, ωj iκ ωi ∧ ω j = hω i , αiκ hβ, ωj iκ hωi , ω j iκ
M i,j i,j
X X
= (−1)|ωi |(n−|ωi |) hα, ω i iκ (−1)|ωi |(n−|ωi |) hβ, ωi iκ = hα, ω i iκ hβ, ωi iκ .
i i

Equality (7.3.2) is proved. ⊔

⊓

Proposition 7.3.10. Let M ∈ M+ n be a manifold, and suppose that (Si , φi ) ∈ Ck (M )

(i = 0, 1) are two k-cycles in M .
n−k
(a) If (S0 , φ0 ) ∼c (S1 , φ1 ), then δS0 = δS1 in Hcpt (M ).
n−k
(b) If (S0 , φ0 ) is trivial, then δS0 = 0 in Hcpt (M ).
n−k
(c) δS0 ⊔S1 = δS0 + δS1 in Hcpt (M ).
n−k
(d) δ−S0 = −δS0 in Hcpt (M ).

Proof. (a) Consider a compact manifold Σ with boundary ∂Σ = −S0 ⊔ S1 and a smooth
map Φ : Σ → M such that Φ |∂Σ = φ0 ⊔ φ1 . For any closed k-form ω ∈ Ωk (M ) we have
Z Z Z Z Z
∗ ∗ Stokes ∗
0= Φ (dω) = dΦ ω = ω= φ1 ω − φ∗0 ω.
Σ Σ ∂Σ S1 S0

Part (b) is left to the reader. Part (c) is obvious. To prove (d) consider Σ = [0, 1] × S0
and
Φ : [0, 1] × S0 → M, Φ(t, x) = φ0 (x), ∀(t, x) ∈ Σ.
Note that ∂Σ = (−S0 ) ⊔ S0 so that

δ−S0 + δS0 = δ−S0 ⊔S0 = δ∂Σ = 0. ⊔

⊓

The above proposition shows that the correspondence

n−k
Ck (M ) ∋ (S, φ) 7→ δS ∈ Hcpt (M )

descends to a map
n−k
δ : Hk (M ) → Hcpt (M ).
This is usually called the homological Poincaré duality. We are not claiming that δ is an
isomorphism.
280 CHAPTER 7. COHOMOLOGY

2 C
1
1

C
2

-1 1
1

Figure 7.3: The intersection number of the two cycles on T 2 is 1

7.3.2 Intersection theory

Consider M ∈ M+ n , and a k-dimensional compact oriented submanifold S of M . We
denote by ı : S ֒→ M inclusion map so that (S, ı) is a k-cycle.
Definition 7.3.11. A smooth map φ : T → M from an (n − k)-dimensional, oriented
manifold T is said to be transversal to S, and we write this S ⋔ φ, if the following hold.
(a) φ−1 (S) is a finite subset of T;
(b) for every x ∈ φ−1 (S) we have
φ∗ (Tx T ) + Tφ(x) S = Tφ(x) M (direct sum).

If S ⋔ Φ, then for each x ∈ φ−1 (S) we define the local intersection number at x to be
(or = orientation)

1 , or (Tφ(x) S) ∧ or(φ∗ Tx T ) = or(Tφ(x) M )
ix (S, T ) = .
−1 , or(Tφ(x) S) ∧ or(φ∗ Tx T ) = −or(Tφ(x) M )
Finally, we define the intersection number of S with T to be
X
S · T := ix (S, T ). ⊔
⊓
x∈φ−1 (S)

Our next result offers a different description of the intersection number indicating how
one can drop the transversality assumption from the original definition.
Proposition 7.3.12. Let M ∈ M+ n . Consider a compact, oriented, k-dimensional sub-
manifold S ֒→ M , and (T, φ) ∈ Cn−k (M ) a (n − k)-dimensional cycle intersecting S
transversally, i.e., S ⋔ φ. Then
Z
S ·T = δS ∧ δT , (7.3.3)
M
7.3. INTERSECTION THEORY 281

where δ• denotes the Poincaré dual of •.

The proof of the proposition relies on a couple of technical lemmata of independent
interest.
Lemma 7.3.13 (Localization lemma). Let M ∈ M+ n and (S, φ) ∈ Ck (M ). Then, for any
n−k
open neighborhood N of φ(S) in M there exists δSN ∈ Ωcpt (M ) such that
N n−k
(a) δS represents the Poincaré dual δS ∈ Hcpt (M );
(b) supp δSN ⊂ N.

Proof. Fix a Riemann metric on M . Each point p ∈ φ(S) has a geodesically convex open
neighborhood entirely contained in N. Cover φ(S) by finitely many such neighborhoods,
+
and denote by N their union. Then N ∈ M+ n , and (S, φ) ∈ Ck (N ).
N
Denote by δS the Poincaré dual of S in N . It can be represented by a closed form in
n−k
Ωcpt (N ) which we continue to denote by δSN . If we pick a closed form ω ∈ Ωk (M ), then
ω |N is also closed, and Z Z Z
ω ∧ δSN = ω ∧ δSN = φ∗ ω.
M N S
n−k
Hence, δSN represents the Poincaré dual of S in Hcpt (M ), and moreover, supp δSN ⊂ N. ⊔
⊓

Definition 7.3.14. Let M ∈ M+ n , and S ֒→ M a compact, k-dimensional, oriented

submanifold of M. A local transversal at p ∈ S is an embedding

φ : B ⊂ Rn−k → M, B = open ball centered at 0 ∈ Rm−k ,

such that S ⋔ φ and φ−1 (S) = {0}. ⊔

⊓

Lemma 7.3.15. Let M ∈ M+ n , S ֒→ M a compact, k-dimensional, oriented submanifold

of M and (B, φ) a local transversal at p ∈ S. Then for any sufficiently “thin” closed
neighborhood N of S ⊂ M we have
Z
S · (B, φ) = φ∗ δSN .
B

Proof. Using the transversality S ⋔ φ, the implicit function theorem, and eventually
restricting φ to a smaller ball, we deduce that, for some sufficiently “thin” neighborhood
N of S, there exist local coordinates (x1 , . . . , xn ) defined on some neighborhood U of
p ∈ M diffeomorphic with the cube

{|xi | < 1, ∀i}

such that the following hold.

(i) S ∩ U = {xk+1 = · · · = xn = 0}, p = (0, . . . , 0).
(ii) The orientation of S ∩ U is defined by dx1 ∧ · · · ∧ dxk .
n−k
(iii) The map φ : B ⊂ R(y 1 ,...,y n−k ) → M is expressed in these coordinates as

x1 = 0, . . . , xk = 0, xk+1 = y 1 , . . . , xn = y n−k .
282 CHAPTER 7. COHOMOLOGY

(iv) N ∩ U = {|xj | ≤ 1/2 ; j = 1, . . . , n}.

Let ǫ = ±1 such that ǫdx1 ∧ . . . ∧ dxn defines the orientation of T M . In other words,

ǫ = S · (B, φ).

For each ξ = (x1 , . . . , xk ) ∈ S ∩ U denote by Pξ the (n − k)-“plane”

Pξ = {(ξ; xk+1 , . . . , xn ) ; |xj | < 1 j > k}.

We orient each Pξ using the (n − k)-form dxk+1 ∧ · · · ∧ dxn , and set

Z
v(ξ) := δSN .
Pξ

Equivalently, Z
v(ξ) = φ∗ξ δSN ,
B
where φξ : B → M is defined by

φξ (y 1 , . . . , y n−k ) = (ξ; y 1 , . . . , y n−k ).

To any function ϕ = ϕ(ξ) ∈ C ∞ (S ∩ U ) such that

supp ϕ ⊂ {|xi | ≤ 1/2 ; i ≤ k},

we associate the k-form

ωϕ := ϕdx1 ∧ · · · ∧ dxk = ϕdξ ∈ Ωkcpt(S ∩ U ).

Extend the functions x1 , . . . , xk ∈ C ∞ (U ∩ N) to smooth compactly supported functions

x̃i ∈ C0∞ (M ) → [0, 1].

The form ωϕ is then the restriction to U ∩ S of the closed compactly supported form

ω̃ϕ = ϕ(x̃1 , . . . , x̃k )dx̃1 ∧ · · · ∧ dx̃k .

We have Z Z Z Z
ω̃ϕ ∧ δSN = ωϕ ∧ δSN = ωϕ = ϕ(ξ)dξ. (7.3.4)
M U S Rk
The integral over U can be evaluated using the Fubini theorem. Write

δSN = f dxk+1 ∧ · · · ∧ dxn + ̺,

where ̺ is an (n − k)-form not containing the monomial dxk+1 ∧ · · · ∧ dxn . Then

Z Z
ωϕ ∧ δSN = f ϕdx1 ∧ · · · ∧ dxn
U U
Z
=ǫ f ϕ|dx1 ∧ · · · ∧ dxn | (|dx1 ∧ · · · ∧ dxn | = Lebesgue density)
U
7.3. INTERSECTION THEORY 283

Z Z !
F ubini k+1 n
= ǫ ϕ(ξ) f |dx ∧ · · · ∧ dx | |dξ|
S∩U Pξ

Z Z ! Z
=ǫ ϕ(ξ) δSN |dξ| = ǫ ϕ(ξ)v(ξ)|dξ|.
S Pξ S

Comparing with (7.3.4), and taking into account that ϕ was chosen arbitrarily, we deduce
that Z
φ∗ δSN = v(0) = ǫ.
B
The local transversal lemma is proved. ⊔
⊓

Proof of Proposition 7.3.12 Let

φ−1 (S) = {p1 , . . . , pm }.

The transversality assumption implies that each pi has an open neighborhood Bi diffeo-
morphic to an open ball such that φi = φ |Bi is a local transversal at yi = φ(pi ). Moreover,
we can choose the neighborhoods Bi to be mutually disjoint. Then
X
S·T = S · (Bi , φi ).
i

The compact set K := φ(T \ ∪Bi ) does not intersect S so that we can find a “thin”, closed
neighborhood” N of S ֒→ M such that K ∩ N = ∅. Then, φ∗ δSN is compactly supported in
the union of the Bi ’s and
Z Z XZ
δSN ∧ δT = φ∗ δSN = φ∗i δSN .
M T i Bi

From the local transversal lemma we get

Z
φ∗ δSN = S · (Bi , φ) = ipi (S, T ). ⊔
⊓
Bi

Equality (7.3.3) has a remarkable feature. Its right-hand-side is an integer which is

defined only for cycles S, T such that S is embedded and S ⋔ T .
The left-hand-side makes sense for any cycles of complementary dimensions, but a
priori it may not be an integer. In any event, we have a remarkable consequence.
Corollary 7.3.16. Let (Si , φi ) ∈ Ck (M ) and ((Ti , ψi ) ∈ Cn−k (M ) where M ∈ M+
n, i =
0, 1. If

(a) S0 ∼c S1 , T0 ∼c T1 ,

(b) the cycles Si are embedded, and

(c) Si ⋔ Ti ,
284 CHAPTER 7. COHOMOLOGY

then
S0 · T0 = S1 · T1 . ⊔
⊓
Definition 7.3.17. The homological intersection pairing is the Z-bilinear map

I = IM : Hk (M ) × Hn−k (M ) → R,

(M ∈ M+
n ) defined by Z
I(S, T ) = δS ∧ δT . ⊔
⊓
M

We have proved that in some special instances I(S, T ) ∈ Z. We want to prove that,
when M is compact, this is always the case.
Theorem 7.3.18. Let M ∈ M+ n be compact manifold. Then, for any (S, T ) ∈ Hk (M ) ×
Hn−k (M ), the intersection number I(S, T ) is an integer.
The theorem will follow from two lemmata. The first one will show that it suffices to
consider only the situation when one of the two cycles is embedded. The second one will
show that, if one of the cycles is embedded, then the second cycle can be deformed so that
it intersects the former transversally. (This is called a general position result.)
Lemma 7.3.19 (Reduction-to-diagonal trick). Let M , S and T as in Theorem 7.3.18.
Then Z
n−k
I(S, T ) = (−1) δS×T ∧ δ∆ ,
M ×M

where ∆ is the diagonal cycle ∆ : M → M × M , x 7→ (x, x). (It is here where the
compactness of M is essential, since otherwise ∆ would not be a cycle).

Proof. We will use the equality (7.3.1)

δS×T = (−1)n−k δS × δT .

Then Z Z
(−1)n−k = δS×T ∧ δ∆ = (δS × δT ) ∧ δ∆
M ×M M ×M
Z Z
∗
= ∆ (δS × δT ) = δS ∧ δT . ⊔
⊓
M M

Lemma 7.3.20 (Moving lemma). Let S ∈ Ck (M ), and T ∈ Cn−k (M ) be two cycles in

M ∈ M+
n . If S is embedded, then T is cobordant to a cycle T̃ such that S ⋔ T̃ .

The proof of this result relies on Sard’s theorem. For details we refer to [62], Chapter
3.
Proof of Theorem 7.3.18 Let (S, T ) ∈ Ck (M ) × Cn−k (M ). Then,

I(S, T ) = (−1)n−k I(S × T, ∆).

Since ∆ is embedded we may assume by the moving lemma that (S × T ) ⋔ ∆, so that

I(S × T, ∆) ∈ Z. ⊔
⊓
7.3. INTERSECTION THEORY 285

7.3.3 The topological degree

Consider two compact, connected, oriented smooth manifolds M , N having the same
dimension n. Any smooth map F : M → N canonically defines an n-dimensional cycle in
M ×N
ΓF : M → M × N x 7→ (x, F (x)).
ΓF is called the graph of F.
Any point y ∈ N defines an n-dimensional cycle M × {y}. Since N is connected all
these cycles are cobordant so that the integer ΓF · (M × {y}) is independent of y.
Definition 7.3.21. The topological degree of the map F is defined by

deg F := (M × {y}) · ΓF . ⊔
⊓

Note that the intersections of ΓF with M × {y} correspond to the solutions of the
equation F (x) = y. Thus the topological degree counts these solutions (with sign).
Proposition 7.3.22. Let F : M → N as above. Then for any n-form ω ∈ Ωn (N )
Z Z
∗
F ω = deg F ω.
M N

Remark 7.3.23. The map F induces a morphism

F ∗
R∼
= H n (N ) → H n (M ) ∼
=R
which can be identified with a real number. The above proposition guarantees that this
number is an integer. ⊔
⊓

Proof of the proposition Note that if ω ∈ Ωn (N ) is exact, then

Z Z
ω= F ∗ ω = 0.
N M

Thus, to prove the proposition it suffices to check it for any particular form which generates
H n (N ). Our candidate will be the Poincaré dual δy of a point y ∈ N . We have
Z
δy = 1,
N

while equality (7.3.1) gives

δM ×{y} = δM × δy = 1 × δy .

We can then compute the degree of F using Theorem 7.3.18

Z Z Z Z
deg F = deg F δy = (1 × δy ) ∧ δΓF = Γ∗F (1 × δy ) = F ∗ δy . ⊔
⊓
N M ×N M M

Corollary 7.3.24 (Gauss–Bonnet). Consider a connected sum of g-tori Σ = Σg embedded

in R3 and let GΣ : Σ → S 2 be its Gauss map. Then

deg GΣ = χ(Σ) = 2 − 2g. ⊔

⊓
286 CHAPTER 7. COHOMOLOGY

This corollary follows immediately from the considerations at the end of Subsection
4.2.4.

Exercise 7.3.25. Consider the compact, connected manifolds M0 , M1 , N ∈ M+

n and
the smooth maps Fi : Mi → N , i = 0, 1. Show that if F0 is cobordant to F1 then
deg F0 = deg F1 . In particular, homotopic maps have the same degree. ⊔
⊓

Exercise 7.3.26. Let A be a nonsingular n × n real matrix. It defines a smooth map

Ax
FA : S n−1 → S n−1 , x 7→ .
|Ax|

Prove that deg FA = sign det A.

Hint: Use the polar decomposition A = P · O (where P is a positive symmetric matrix
and O is an orthogonal one) to deform A inside GL(n, R) to a diagonal matrix. ⊔
⊓

F
Exercise 7.3.27. Let M → N be a smooth map (M , N are smooth, compact oriented of
dimension n). Assume y ∈ N is a regular value of F , i.e., for all x ∈ F −1 (y) the derivative

Dx F : Tx M → Ty N

is invertible. For x ∈ F −1 (y) define


1 , Dx F preserves orientations
deg(F, x) =
−1 , otherwise

Prove that
X
deg F = deg(F, x). ⊔
⊓
F (x)=y

Exercise 7.3.28. Let M denote a compact oriented manifold, and consider a smooth
map F : M → M Regard H • (M ) as a superspace with the obvious Z2 -grading

H • (M ) = H even (M ) ⊕ H odd (M ),

and define the Lefschetz number λ(F ) of F as the supertrace of the pull back

F ∗ : H • (M ) → H • (M ).

Prove that λ(F ) = ∆ · ΓF , and deduce from this the Lefschetz fixed point theorem:
λ(F ) 6= 0 ⇒ F has a fixed point. ⊔
⊓

Remark 7.3.29. For an elementary approach to degree theory, based only on Sard’s
theorem, we refer to the beautiful book of Milnor, [97]. ⊔
⊓
7.3. INTERSECTION THEORY 287

7.3.4 The Thom isomorphism theorem

Let p : E → B be an orientable fiber bundle (in the sense of Definition 3.4.45) with
standard fiber F and compact, oriented basis B. Let dim B = m and dim F = r. The total
space E is a compact orientable manifold which we equip with the fiber-first orientation.
In this subsection we will extensively use the techniques of fibered calculus described
in Subsection 3.4.5. The integration along fibers
Z
= p∗ : Ω•cpt (E) → Ω•−r (B)
E/B

satisfies
p∗ dE = (−1)r dB p∗ ,
so that it induces a map in cohomology
•
p∗ : Hcpt (E) → H •−r (B).
This induced map in cohomology is sometimes called the Gysin map.
Remark 7.3.30. If the standard fiber F is compact, then the total space E is also
compact. Using Proposition 3.4.48 we deduce that for any ω ∈ Ω• (E), and any η ∈ Ω• (B)
such that
deg ω + deg η = dim E,
we have
hp∗ (ω), ηiκ = hω, p∗ (η)ik
If we denote by DM : Ω• (M ) → Ωdim M −• (M )∗ the Poincaré duality isomorphism on a
compact oriented smooth manifold M , then we can rewrite the above equality as
hDB p∗ (ω), ηi = hDE ω, p∗ ηi = h(p∗ )† DE ω, ηi =⇒ DB p∗ (ω) = (p∗ )† DE ω.
Hence
p∗ = D−1 ∗ †
B (p ) DE ,
so that, in this case, the Gysin map coincides with the pushforward (Gysin) map defined
in Remark 7.2.9. ⊔
⊓

Exercise 7.3.31. Consider a smooth map f : M → N between compact, oriented man-

ifolds M , N of dimensions m and respectively n. Denote by if the embedding of M in
M × N as the graph of f
M ∋ x 7→ (x, f (x)) ∈ M × N.
The natural projection M × N → N allows us to regard M × N as a trivial fiber bundle
over N .
Show that the push-forward map f∗ : H • (M ) → H •+n−m (N ) defined in Remark 7.2.9
can be equivalently defined by
f∗ = π∗ ◦ (if )∗ ,
where π∗ : H • (M × N ) → H •−m (N ) denotes the integration along fibers while
(if )∗ : H • (M ) → H •+n (M × N ),
is the pushforward morphism defined by the embedding if . ⊔
⊓
288 CHAPTER 7. COHOMOLOGY

Let us return to the fiber bundle p : E → B. Any smooth section σ : B → E defines

an embedded cycle in E of dimension m = dim B. Denote by δσ its Poincaré dual in
r (E).
Hcpt
Using the properties of the integration along fibers we deduce that, for any ω ∈ Ωm (B),
we have !
Z Z Z
δσ ∧ p∗ ω = δσ ω.
E B E/B

On the other hand, by Poincaré duality we get

Z Z
∗ rm
δσ ∧ p ω = (−1) p∗ ω ∧ δσ
E E
Z Z Z
rm ∗ ∗ rm ∗ rm
= (−1) σ p ω = (−1) (pσ) ω = (−1) ω.
B B B

Hence Z
p∗ δσ = δσ = (−1)rm ∈ Ω0 (B).
E/B

Proposition 7.3.32. Let p : E → B a bundle as above. If it admits at least one section,

then the Gysin map
•
p∗ : Hcpt (E) → H •−r (B)
is surjective.

Proof. Denote by τσ the map

•+r
τσ : H • (B) → Hcpt (E) ω 7→ (−1)rm δσ ∧ p∗ ω = p∗ ω ∧ δσ .

Then τσ is a right inverse for p∗ . Indeed

ω = (−1)rm p∗ δσ ∧ ω = (−1)rm p∗ (δσ ∧ p∗ ω) = p∗ (τσ ω). ⊔

⊓

The map p∗ is not injective in general. For example, if (S, φ) is a k-cycle in F , then
it defines a cycle in any fiber π −1 (b), and consequently in E. Denote by δS its Poincaré
m+r−k
dual in Hcpt (E). Then for any ω ∈ Ωm−k (B) we have
Z Z Z Z
(p∗ δS ) ∧ ω = δS ∧ p∗ ω = ± φ∗ p∗ ω = (p ◦ φ)∗ ω = 0,
B E D S

since p ◦ φ is constant. Hence p∗ δS = 0, and we conclude that that if F carries nontrivial

cycles ker p∗ may not be trivial.
The simplest example of standard fiber with only trivial cycles is a vector space.
Definition 7.3.33. Let p : E → B be an orientable vector bundle over the compact
oriented manifold B (dim B = m, rank (E) = r). The Thom class of E, denoted by τE is
the Poincaré dual of the cycle defined by the zero section ζ0 : B → E, b 7→ 0 ∈ Eb . Note
r (E).
that τE ∈ Hcpt ⊔
⊓
7.3. INTERSECTION THEORY 289

Theorem 7.3.34 (Thom isomorphism). Let p : E → B as in the above definition. Then

the map
•+r
τ : H • (B) → Hcpt (E) ω 7→ τE ∧ p∗ ω

is an isomorphism called the Thom isomorphism. Its inverse is the Gysin map

(−1)rm p∗ : Hcpt
•
(E) → H •−r (B).

Proof. We have already established that p∗ τ = (−1)rm . To prove the reverse equality,
τ p∗ = (−1)rm , we will use Lemma 7.2.2 of Subsection 7.2.1. For β ∈ Ω•cpt (E) we have

(p∗ p∗ τE ) ∧ β − τE ∧ (p∗ p∗ β) = (−1)r d(m(τE , β)),

where m(τE , β) ∈ Ω•cpt (E). Since p∗ p∗ τE = (−1)rm , we deduce

(−1)rm β = (τE ∧ p∗ (p∗ β) + exact form ⇒ (−1)rm β = τE ◦ p∗ (β) in Hcpt

•
(E). ⊔
⊓

•+dim M
Exercise 7.3.35. Show that τE = ζ∗ 1, where ζ• : H • (M ) → Hcpt (E) is the push-
forward map defined by a section ζ : M → E. ⊔
⊓

Remark 7.3.36. Suppose M ∈ M+ n , and S ֒→ M is a compact, oriented, k-dimensional

manifold. Fix a Riemann metric g on M , and denote by NS the normal bundle of the
embedding S ֒→ M , i.e., the orthogonal complement of T N in (T M )|S .
The exponential map defined by the metric g defines a smooth map

exp : NS → M,

which induces a diffeomorphism from an open neighborhood O of S in NS , to an open

(tubular) neighborhood N of S in M . Fix a closed form δSN ∈ Ωn−k (M ) with compact
support contained in N, and representing the Poincaré dual of (S, i), where i : S ֒→ M
denotes the canonical inclusion.
Then exp∗ δSN is the Poincaré dual of the cycle (S, ζ) in NS , where ζ : S → NS denotes
the zero section. This shows that exp∗ δSN is a compactly supported form representing the
Thom class of the normal bundle NS .
Using the identification between O and N we obtain a natural submersive projection
π : N → S corresponding to the natural projection NS → S. In more intuitive terms, π
associates to each x ∈ N the unique point in S which closest to x. One can prove that the
Gysin map
•+(n−k)
i∗ : H • (S) → Hcpt (M ),

is given by,
•+(n−k)
H • (S) ∋ ω 7−→ exp∗ δSN ∧ π ∗ ω ∈ Hcpt (M ). (7.3.5)

⊔
⊓

Exercise 7.3.37. Prove the equality (7.3.5). ⊔

⊓
290 CHAPTER 7. COHOMOLOGY

7.3.5 Gauss–Bonnet revisited

We now examine a very special type of vector bundle: the tangent bundle of a compact,
oriented, smooth manifold M . Note first the following fact.
Exercise 7.3.38. Prove that M is orientable if and only if T M is orientable as a bundle.
⊔
⊓

Definition 7.3.39. Let E → M be a real orientable vector bundle over the compact,
n (E) the Thom class of
oriented, n-dimensional, smooth manifold M . Denote by τE ∈ Hcpt
E. The Euler class of E is defined by

e(E) := ζ0∗ τE ∈ H n (M ),

where ζ0 : M → E denotes the zero section. e(T M ) is called the Euler class of M , and it
is denoted by e(M ). ⊔
⊓

Note that the sections of T M are precisely the vector fields on M . Moreover, any such
section σ : M → T M tautologically defines an n-dimensional cycle in T M , and in fact,
any two such cycles are homotopic: try a homotopy, affine along the fibers of T M .
Any two sections σ0 , σ1 : M → T M determine cycles of complementary dimension,
and thus the intersection number σ0 · σ1 is a well defined integer, independent of the two
sections. It is a number reflecting the topological structure of the manifold.
Proposition 7.3.40. Let σ0 , σ1 : M → T M be two sections of T M . Then
Z
e(M ) = σ0 · σ1 .
M

In particular, if dim M is odd then

Z
e(M ) = 0.
M
n (T M ) is the Thom
Proof. The section σ0 , σ1 are cobordant, and their Poincaré dual in Hcpt
class τM . Hence Z Z
σ0 · σ1 = δσ0 ∧ δσ1 = τM ∧ τM
TM TM
Z Z Z
∗
= τM ∧ δζ0 = ζ0 τM = e(M ).
TM M M
If dim M is odd then
Z Z
e(M ) = σ0 · σ1 = −σ1 · σ0 = − e(M ). ⊔
⊓
M M

Theorem 7.3.41. Let M be a compact oriented n-dimensional manifold, and denote by

e(M ) ∈ H n (M ) its Euler class. Then the integral of e(M ) over M is equal to the Euler
characteristic of M ,
Z X n
e(M ) = χ(M ) = (−1)k bk (M ).
M k=0
7.3. INTERSECTION THEORY 291

In the proof we will use an equivalent description of χ(M ).

Lemma 7.3.42. Denote by ∆ the diagonal cycle ∆ ∈ Hn (M × M ). Then χ(M ) = ∆ · ∆.

n (M × M ) the Poincaré dual of ∆. Consider a basis (ω ) of

Proof. Denote by δ M ∈ Hcpt j
H (M ) consisting of homogeneous elements. We denote by (ω i ) the dual basis, i.e.,
•

hω i , ωj iκ = δji . ∀i, j.
According to (7.3.2), we have
X i
δM = (−1)|ω | ω i × ωi .
i

Similarly, if we start first with the basis (ω i ), then its dual basis is
i |·|ω
(−1)|ω i|
ωi ,
so, taking into account that |ωj | + |ω j | · |ωj | ≡ n|ωj | (mod 2), we also have
X
δM = (−1)n·|ωj | ωj × ω j .
i

Using Exercise 7.1.50 we deduce

!  
Z Z X X
i
∆·∆= δM ∧ δM = (−1)|ω | ω i × ωi ∧  (−1)n|ωj | ωj × ω j 
M ×M M ×M i j
 
Z X i

=  (−1)|ω |+n·|ωj | (−1)|ωi |·|ωj | ω i ∧ ωj  × ωi × ω j
M ×M i,j
X i |+n·|ω
= (−1)|ω j|
(−1)|ωi |·|ωj | hω i , ωj iκ hωi , ω j iκ .
i,j
In the last expression we now use the duality equations
j
hωi , ω j iκ = (−1)|ωi |·|ω | δij ,
and the congruences
n≡0 mod 2, |ω i | + n|ωi | + |ωi |2 + |ωi | · |ω i | ≡ |ω i | mod 2,
to conclude that X i
∆·∆= (−1)|ω | = χ(M ). ⊔
⊓
ωi
Proof of theorem 7.3.41 The tangent bundle of M × M restricts to the diagonal ∆
as a rank 2n vector bundle. If we choose a Riemann metric on M × M then we get an
orthogonal splitting
T (M × M ) |∆ = N∆ ⊕ T ∆.
The diagonal map M → M × M identifies M with ∆ so that T ∆ ∼ = T M . We now have
the following remarkable result.
292 CHAPTER 7. COHOMOLOGY

Lemma 7.3.43. N∆ ∼
= T M.

Proof Use the isomorphisms

T (M × M ) |∆ ∼
= T ∆ ⊕ N∆ ∼
= T M ⊕ N∆

and
T (M × M ) |∆ = T M ⊕ T M. ⊔
⊓

From this lemma we immediately deduce the equality of Thom classes

τ N∆ = τ M . (7.3.6)

At this point we want to invoke a technical result whose proof is left to the reader as an
exercise in Riemann geometry.

Lemma 7.3.44. Denote by exp the exponential map of a Riemann metric g on M × M .

Regard ∆ as a submanifold in N∆ via the embedding given by the zero section. Then there
exists an open neighborhood U of ∆ ⊂ N∆ ⊂ T (M × M ) such that

exp |U : U → M × M

is an embedding. ⊔
⊓

Let U be a neighborhood of ∆ ⊂ N∆ as in the above lemma, and set N := exp(U).

U the Poincaré dual of ∆ in U, δ U ∈ H n (U), and by δ N the Poincaré dual of
Denote by δ∆ ∆ cpt ∆
N n (N). Then
∆ in N, δ∆ ∈ Hcpt
Z Z Z
U U N N
δ∆ ∧ δ∆ = δ∆ ∧ δ∆ = δ M ∧ δ M = χ(M ).
U N M ×M

U is the Thom class of the bundle N

The cohomology class δ∆ ∆ → ∆, which in view of
U
(7.3.6) means that δ∆ = τN∆ = τM . We get
Z Z Z Z Z
U
δ∆ U
∧ δ∆ = U
δ∆ = ζ0∗ τN∆ = ζ0∗ τM = e(M ).
U ∆ ∆ M M

Hence Z
e(M ) = χ(M ). ⊔
⊓
M

If M is a connected sum of g tori then we can rephrase the Gauss–Bonnet theorem as

follows.

Corollary 7.3.45. For any Riemann metric h on a connected sum of g-tori Σg we have

1 1
ε(h) = sh dvh = e(Σg ) in H ∗ (Σg ). ⊔
⊓
2π 4π
7.3. INTERSECTION THEORY 293

The remarkable feature of the Gauss–Bonnet theorem is that, once we choose a metric,
we can explicitly describe a representative of the Euler class in terms of the Riemann
curvature.
The same is true for any compact, oriented, even-dimensional Riemann manifold. In
this generality, the result is known as Gauss–Bonnet–Chern, and we will have more to say
about it in the next two chapters.
We now have a new interpretation of the Euler characteristic of a compact oriented
manifold M .
Given a smooth vector field X on M , its “graph” in T M ,
ΓX = {(x, X(x)) ∈ Tx M ; x ∈ M },
is an n-dimensional submanifold of T M . The Euler characteristic is then the intersection
number
χ(M ) = ΓX · M,
where we regard M as a submanifold in T M via the embedding given by the zero section.
In other words, the Euler characteristic counts (with sign) the zeroes of the vector fields
on M . For example if χ(M ) 6= 0 this means that any vector field on M must have a zero
! We have thus proved the following result.
Corollary 7.3.46. If χ(M ) 6= 0 then the tangent bundle T M is nontrivial. ⊔
⊓

The equality χ(S 2n ) = 2 is particularly relevant in the vector field problem discussed
in Subsection 2.1.4. Using the notations of that subsection we can write
v(S 2n ) = 0.
We have thus solved “half” the vector field problem.
Exercise 7.3.47. Let X be a vector field over the compact oriented manifold M . A point
x ∈ M is said to be a non-degenerate zero of X if X(0) = 0 and
 
∂Xi
det |x=x0 6= 0.
∂xj
for some local coordinates (xi ) near x0 such that the orientation of Tx∗0 M is given by
dx1 ∧ · · · ∧ dxn . Prove that the local intersection number of ΓX with M at x0 is given by
 
∂Xi
ix0 (ΓX , M ) = sign det |x=x0 .
∂xj
This is sometimes called the local index of X at x0 , and it is denoted by i(X, x0 ). ⊔
⊓

From the above exercise we deduce the following celebrated result.

Corollary 7.3.48 (Poincaré-Hopf). If X is a vector field along a compact, oriented man-
ifold M , with only non-degenerated zeros x1 , . . . , xk , then
X
χ(M ) = i(X, xj ). ⊔
⊓
j
294 CHAPTER 7. COHOMOLOGY

Exercise 7.3.49. Let X be a vector field on Rn and having a non-degenerate zero at the
origin.
(a) prove that for all r > 0 sufficiently small X has no zeros on Sr = {|x| = r}.
(b) Consider Fr : Sr → S n−1 defined by
1
Fr (x) = X(x).
|X(x)|
Prove that i(X, 0) = deg Fr for all r > 0 sufficiently small.
Hint: Deform X to a linear vector field. ⊔
⊓

7.4 Symmetry and topology

The symmetry properties of a manifold have a great impact on its global (topological)
structure. We devote this section to a more in depth investigation of the correlation
symmetry-topology.

7.4.1 Symmetric spaces

Definition 7.4.1. A homogeneous space is a smooth manifold M acted transitively by
a Lie group G called the symmetry group. ⊔
⊓

Recall that a smooth left action of a Lie group G, on a smooth manifold M

G × M → M (g, m) 7→ g · m,

is called transitive if, for any m ∈ M , the map

Ψm : G ∋ g 7→ g · m ∈ M

is surjective. For any point x of a homogeneous space M we define the isotropy group at
x by
Ix = {g ∈ G ; g · x = x}.
Lemma 7.4.2. Let M be a homogeneous space with symmetry group G and x, y ∈ M .
Then
(a) Ix is a closed subgroup of G;
(b) Ix ∼
= Iy .

Proof. (a) is immediate. To prove (b), choose g ∈ G such that y = g · x. Then note that
Iy = gIx g−1 . ⊔
⊓

Remark 7.4.3. It is worth mentioning some fundamental results in the theory of Lie
groups which will shed a new light on the considerations of this section. For proofs we
refer to [61, 128].
Fact 1. Any closed subgroup of a Lie group is also a Lie group (see Remark 1.2.31). In
particular, the isotropy groups Ix of a homogeneous space are all Lie groups. They are
smooth submanifolds of the symmetry group.
7.4. SYMMETRY AND TOPOLOGY 295

Fact 2. Let G be a Lie group, and H a closed subgroup. Then the space of left cosets,

G/H := g · H; g ∈ G ,
can be given a smooth structure such that the map
G × (G/H) → G/H (g1 , g2 H) 7→ (g1 g2 ) · H
is smooth. The manifold G/H becomes a homogeneous space with symmetry group G.
All the isotropy groups subgroups of G conjugate to H.
Fact 3. If M is a homogeneous space with symmetry group G, and x ∈ M , then M is
equivariantly diffeomorphic to G/Ix , i.e., there exists a diffeomorphism
φ : M → G/Ix ,
such that φ(g · y) = g · φ(y). ⊔
⊓

We will be mainly interested in a very special class of homogeneous spaces.

Definition 7.4.4. A symmetric space is a collection of data (M, h, G, σ, i) satisfying the
following conditions.
(a) (M, h) is a Riemann manifold.
(b) G is a connected Lie group acting isometrically and transitively on M
G × M ∋ (g, m) 7→ g · m ∈ M.
(c) σ : M × M → M is a smooth map (m1 , m2 ) 7→ σm1 (m2 ) such that the following hold.

(c1) ∀m ∈ M σm : M → M is an isometry, and σm (m) = m.

(c2) σm ◦ σm = 1M .
(c3) Dσm |Tm M = −1Tm M .
(c4) σgm = gσm g −1 .

(d) i : M × G → G, (m, g) 7→ im g is a smooth map such that the following hold.

(d1) ∀m ∈ M , im : G → G is a homomorphism of G.
(d2) im ◦ im = 1G .
(d3) igm = gim g−1 , ∀(m, g) ∈ M × G.
−1 (x) = σ gσ (x) = i (g) · x, ∀m, x ∈ M , g ∈ G.
(e) σm gσm ⊔
⊓
m m m

Remark 7.4.5. This may not be the most elegant definition of a symmetric space, and
certainly it is not the minimal one. As a matter of fact, a Riemann manifold (M, g) is
a symmetric space if and only if there exists a smooth map σ : M × M → M satisfying
the conditions (c1),(c2) and (c3). We refer to [61, 70] for an extensive presentation of this
subject, including a proof of the equivalence of the two descriptions. Our definition has
one academic advantage: it lists all the properties we need to establish the topological
results of this section. ⊔
⊓
296 CHAPTER 7. COHOMOLOGY

The next exercise offers the reader a feeling of what symmetric spaces are all about.
In particular, it describes the geometric significance of the family of involutions σm .
Exercise 7.4.6. Let (M, h) be a symmetric space. Denote by ∇ the Levi–Civita connec-
tion, and by R the Riemann curvature tensor.
(a) Prove that ∇R = 0.
(b) Fix m ∈ M , and let γ(t) be a geodesic of M such that γ(0) = m. Show that

σm γ(t) = γ(−t). ⊔
⊓

Example 7.4.7. Perhaps the most popular example of symmetric space is the round
sphere S n ⊂ Rn+1 . The symmetry group is SO(n + 1), the group of orientation preserving
“rotations” of Rn+1 . For each m ∈ S n+1 we denote by σm the orthogonal reflection through
the 1-dimensional space determined by the radius Om. We then set im (T ) = σm T σm −1 ,

∀T ∈ SO(n + 1). We let the reader check that σ and i satisfy all the required axioms. ⊔ ⊓

Example 7.4.8. Let G be a connected Lie group and m a bi-invariant Riemann metric
on G. The direct product G × G acts on G by

(g1 , g2 ) · h = g1 hg2−1 .

This action is clearly transitive and since m is bi-invariant its is also isometric. Define

σ : G × G → G σg h = gh−1 g−1

and
i : G × (G × G) → G × G ig (g1 , g2 ) = (gg1 g−1 , gg2 g−1 ).
We leave the reader to check that these data do indeed define a symmetric space structure
on (G, m). The symmetry group is G × G. ⊔
⊓

Example 7.4.9. Consider the complex Grassmannian M = Grk (Cn ). Recall that in
Example 1.2.22 we described Grk (Cn ) as a submanifold of End+ (V ) – the linear space
of self-adjoint n × n complex matrices, via the map which associates to each complex
subspace S, the orthogonal projection PS : Cn → Cn onto S.
The linear space End+ (Cn ) has a natural metric g0 (A, B) = 21 Re tr (AB ∗ ) that restricts
to a Riemann metric g on M . The unitary group U (n) acts on End+ (Cn ) by conjugation,

U (n) × End+ (Cn ) ∋ (T, A) 7−→ T ⋆ A := T AT ∗ .

Note that U (n) ⋆ M = M , and g0 is U (n)-invariant. Thus U (n) acts transitively, and
isometrically on M .
For each subspace S ∈ M define

RS := PS − PS ⊥ = 2PS − 1.

The operator RS is the orthogonal reflection through S ⊥ . Note that RS ∈ U (n), and
RS2 = 2. The map A 7→ RS ⋆ A is an involution of End+ (Cn ). It descends to an involution
of M . We thus get an entire family of involutions

σ : M × M → M, (PS1 , PS2 ) 7→ RS1 ⋆ PS2 .

7.4. SYMMETRY AND TOPOLOGY 297

Define
i : M × U (n) → U (n), iS T = RS T RS .
We leave the reader to check that the above collection of data defines a symmetric space
structure on Grk (Cn ). ⊔
⊓

Exercise 7.4.10. Fill in the details left out in the above example. ⊔
⊓

7.4.2 Symmetry and cohomology

Definition 7.4.11. Let M be a homogeneous space with symmetry group G. A differen-
tial form ω ∈ Ω∗ (M ) is said to be (left) invariant if ℓ∗g ω = ω ∀g ∈ G, where we denoted
by
ℓ∗g : Ω• (M ) → Ω• (M )
the pullback defined by the left action by g: m 7→ g · m. ⊔
⊓

Proposition 7.4.12. Let M be a compact homogeneous space with compact, connected

symmetry group G. Then any cohomology class of M can be represented by a (not neces-
sarily unique) invariant form.

Proof. Denote by dVG (g) the normalized bi-invariant volume form on G. For any form
ω ∈ Ω∗ (M ) we define its G-average by
Z
ω= ℓ∗g ω dVG (g).
G

The form ω is an invariant form on M . The proposition is a consequence of the following

result.
Lemma 7.4.13. If ω is a closed form on M then ω is closed and cohomologous to ω.

Proof. The form ω is obviously closed so we only need to prove it is cohomologous to ω.

Consider a bi-invariant Riemann metric m on G. Since G is connected, the exponential
map
exp : LG → G X 7→ exp(tX)
is surjective. Choose r > 0 sufficiently small so that the map

exp : Dr = |X|m = r ; X ∈ LG → G

is an embedding. Set Br = exp Dr . We can select finitely many g1 , . . . , gm ∈ G such that

m
[
G= Bj , Bj := gj Br .
j=1

Now pick a partition of unity (αj ) ⊂ C ∞ (G) subordinated to the cover (Bj ), i.e.,
X
0 ≤ αj ≤ 1, supp αj ⊂ Bj , αj = 1.
j
298 CHAPTER 7. COHOMOLOGY

Set Z
aj := αj dVG (g).
G
P P
Since the volume of G is normalized to 1, and j αj = 1, we deduce that j aj = 1. For
any j = 1, . . . , m define Tj : Ω∗ (M ) → Ω∗ (M ) by
Z
Tj ω := αj (g)ℓ∗g ω dVG (g).
G

Note that X
ω= Tj ω and dTj ω = Tj dω.
j

Each Tj is thus a cochain morphism. It induces a morphism in cohomology which we

continue to denote by Tj . The proof of the lemma will be completed in several steps.
Step 1. ℓ∗g = 1 on H ∗ (M ), for all g ∈ G. Let X ∈ LG such that g = exp X. Define

f : I × M → M ft (m) := exp(tX) · m = ℓexp(tX) m.

The map f is a homotopy connecting 1M with ℓ∗g . This concludes Step 1.

Step 2.
Tj = aj 1H ∗ (M ) .
For t ∈ [0, 1], consider φj,t : Bj → G defined as the composition

gj−1 exp−1 t·− exp gj

Bj −→ Br −→ Dr −→ Dtr −→ Btr ֒→ G.

Define Tj,t : Ω∗ (M ) → Ω∗ (M ) by
Z
Tj,t ω = αj (g)ℓ∗φj,t (t) ω dVG (g) = Tj φ∗j,t ω. (7.4.1)
G

We claim that Tj,0 is cochain homotopic to Tj,1 .

To verify this claim set t := es , −∞ < s ≤ 0 and

gs = exp(es exp−1 (g)) ∀g ∈ Br .

Then Z Z
def
Us ω = Tj,es ω = αj (gj g)ℓ∗gj gs ω = α(gj g)ℓ∗gs ℓ∗gj ω dVG (g).
Br Br

For each g ∈ Br the map (s, m) 7→ Ψs (m) = gs (m) defines a local flow on M . We denote
by Xg its infinitesimal generator. Then
Z
d
(Us ω) = αj (gj g)LXg ℓ∗gs ℓ∗gj ω dVG (g)
ds Br
Z
= αj (gj g)(dM iXg + iXg dM )(ℓ∗gs ℓ∗gj ω) dVG (g).
Br
7.4. SYMMETRY AND TOPOLOGY 299

Consequently,
Tj,0 ω − Tj,1 ω = U−∞ ω − U0 ω
Z 0 Z 
=− αj (gj g)(diXg + iXg d)(ℓ∗gs ℓ∗gj ω) dVG (g) ds.
−∞ Br

(An argument entirely similar to the one we used in the proof of the Poincaré lemma
shows that the above improper integral is pointwise convergent). From the above formula
we immediately read a cochain homotopy χ : Ω• (M ) → Ω•−1 (M ) connecting U−∞ to U0 .
More precisely
Z 0 Z n o 
{χ(ω)} |x∈M := − αj (gj g) iXg ℓ∗gs ℓ∗gj ω |x dVG (g) ds.
−∞ Br

Now notice that Z !

Tj,0 ω = αj (g)dVG (g) = aj ℓ∗gj ω,
Bj

while Tj,1 ω = Tj ω. Taking into account Step 1, we deduce Tj,0 = aj · 1. Step 2 is

completed.
The lemma and hence the proposition follow from the equality
X X
1H ∗ (M ) = aj 1H ∗ (M ) = Tj = G − average. ⊔
⊓
j j

The proposition we have just proved has a greater impact when M is a symmetric space.
Proposition 7.4.14. Let (M, h) be an, oriented symmetric space with symmetry group
G. Then the following are true.
(a) Every invariant form on M is closed.
(b) If moreover M is compact, then the only invariant form cohomologous to zero is the
trivial one.
∗ ω. We claim ω̂
Proof. (a) Consider an invariant k-form ω. Fix m0 ∈ M and set ω̂ = σm0
is invariant. Indeed, ∀g ∈ G

ℓ∗g ω̂ = ℓ∗g σm
∗
0
ω = (σm0 g)∗ ω = (gig σm0 g)∗ ω = σm
∗
ℓ ∗ ω = σm
0 i(g)
∗
0
ω = ω̂.

Since Dσm0 |Tm0 M = −1Tm0 M , we deduce that, at m0 ∈ M , we have ω̂ = (−1)k ω.

Both ω and ω̂ are G-invariant, and we deduce that the above equality holds at any
point m = g · m0 . Invoking the transitivity of the G-action we conclude that

ω̂ = (−1)k ω on M.

In particular, dω̂ = (−1)k dω on M .

The (k + 1)-forms dω̂ = σm c and dω are both invariant and, arguing as above,
∗ dω = dω,
0
we deduce
c = (−1)k+1 dω.
dω̂ = dω
The last two inequalities imply dω = 0.
300 CHAPTER 7. COHOMOLOGY

(b) Let ω be an invariant form cohomologous to zero, i.e. ω = dα. Denote by ∗ the Hodge
∗-operator corresponding to the invariant metric h. Since G acts by isometries, the form
η = ∗ω is also invariant, so that dη = 0. We can now integrate (M is compact), and use
Stokes theorem to get
Z Z Z
ω ∧ ∗ω = dα∧ = ± α ∧ dη = 0.
M M M
This forces ω ≡ 0. ⊔
⊓

From Proposition 7.4.12 and the above theorem we deduce the following celebrated
result of Élie Cartan ([25])
Corollary 7.4.15 (É. Cartan). Let (M, h) be a compact, oriented symmetric space with
compact, connected symmetry group G. Then the cohomology algebra H ∗ (M ) of M is
isomorphic with the graded algebra Ω∗inv (M ) of invariant forms on M . ⊔
⊓

In the coming subsections we will apply this result to the symmetric spaces discussed
in the previous subsection: the Lie groups and the complex Grassmannians.

7.4.3 The cohomology of compact Lie groups

Consider a compact, connected Lie group G, and denote by LG its Lie algebra. According
to Proposition 7.4.12, in computing its cohomology, it suffices to restrict our considerations
to the subcomplex consisting of left invariant forms. This can be identified with the
exterior algebra Λ• L∗G . We deduce the following result.
Corollary 7.4.16. H • (G) ∼= H • (LG ) ∼
= Λ•inv LG , where Λ•inv LG denotes the algebra of
•
bi-invariant forms on G, while H (LG ) denotes the Lie algebra cohomology introduced in
Example 7.1.10. ⊔
⊓

Using the Exercise 7.1.11 we deduce the following consequence.

Corollary 7.4.17. If G is a compact semisimple Lie group then H 1 (G) = 0. ⊔
⊓

Proposition 7.4.18. Let G be a compact semisimple Lie group. Then H 2 (G) = 0.

Proof. A closed bi-invariant 2-form ω on G is uniquely defined by its restriction to LG ,

and satisfies the following conditions.
dω = 0 ⇐⇒ ω([X0 , X1 ], X2 ) − ω([X0 , X2 ], X1 ) + ω([X1 , X2 ], X0 ) = 0,
and (right-invariance)
(LX0 ω)(X1 , X2 ) = 0 ∀X0 ∈ LG ⇐⇒ ω([X0 , X1 ], X2 ) − ω([X0 , X2 ], X1 ) = 0.
Thus
ω([X0 , X1 ], X2 ) = 0 ∀X0 , X1 , X2 ∈ LG .
On the other hand, since H 1 (LG ) = 0 we deduce (see Exercise 7.1.11) LG = [LG , LG ], so
that the last equality can be rephrased as
ω(X, Y ) = 0 ∀X, Y ∈ LG . ⊔
⊓
7.4. SYMMETRY AND TOPOLOGY 301

Definition 7.4.19. A Lie algebra is called simple if it has no nontrivial ideals. A Lie
group is called simple if its Lie algebra is simple. ⊔
⊓

Exercise 7.4.20. Prove that SU (n) and SO(m) are simple. ⊔

⊓

Proposition 7.4.21. Let G be a compact, simple Lie group. Then H 3 (G) ∼

= R. Moreover,
H 3 (G) is generated by the Cartan form
α(X, Y, Z) = κ([X, Y ], Z),
where κ denotes the Killing pairing. ⊔
⊓

The proof of the proposition is contained in the following sequence of exercises.

Exercise 7.4.22. Prove that a simple Lie algebra is necessarily semi-simple. ⊔
⊓

Exercise 7.4.23. Let ω be a closed, bi-invariant 3-form on a Lie group G. Then

ω(X, Y, [Z, T ]) = ω([X, Y ], Z, T ) ∀X, Y, Z, T ∈ LG . ⊔
⊓
Exercise 7.4.24. Let ω be a closed, bi-invariant 3-form on a compact, semisimple Lie
group.
(a) Prove that for any X ∈ LG there exists a unique left-invariant form ηX ∈ Ω1 (G) such
that
(iX ω)(Y, Z) = ηX ([Y, Z]).
Moreover, the correspondence X 7→ ηX is linear. Hint: Use H 1 (G) = H 2 (G) = 0.
(b) Denote by A the linear operator LG → LG defined by
κ(AX, Y ) = ηX (Y ).
Prove that A is self-adjoint with respect to the Killing metric.
(c) Prove that the eigenspaces of A are ideals of LG . Use this to prove Proposition 7.4.21.
⊔
⊓

Exercise 7.4.25. Compute Z Z

α and α,
SU (2) SO(3)
where α denotes the Cartan form. (These groups are oriented by their Cartan forms.)
Hint: Use the computation in the Exercise 4.1.60 and the double cover SU (2) → SO(3)
described in the Subsection 6.2.1. Pay very much attention to the various constants. ⊔
⊓

7.4.4 Invariant forms on Grassmannians and Weyl’s integral formula

We will use the results of Subsection 7.4.2 to compute the Poincaré polynomial of the
complex Grassmannian Grk (Cn ). Set ℓ = n − k.
As we have seen in the previous subsection, the Grassmannian Grk (Cn ) is a symmet-
ric space with symmetry group U (n). It is a complex manifold so that it is orientable
(cf. Exercise 3.4.13). Alternatively, the orientability of Grk (Cn ) is a consequence of the
following fact.
302 CHAPTER 7. COHOMOLOGY

Exercise 7.4.26. If M is a homogeneous space with connected isotropy groups, then M

is orientable. ⊔
⊓

We have to describe the U (n)-invariant forms on Grk (Cn ). These forms are completely
determined by their values at a particular point in the Grassmannian. We choose this point
to correspond to the subspace S0 determined by the canonical inclusion Ck ֒→ Cn .
The isotropy of S0 is the group H = U (k) × U (ℓ). The group H acts linearly on the
tangent space V0 = TS0 Grk (Cn ). If ω is an U (n)-invariant form, then its restriction to V0
is an H-invariant skew-symmetric, multilinear map

V0 × · · · × V0 → R.

Conversely, any H-invariant element of Λ• V0∗ extends via the transitive action of U (n)
to an invariant form on Grk (Cn ). Denote by Λ•inv the space of H-invariant elements of
Λ• V0∗ . We have thus established the following result.
Proposition 7.4.27. There exists an isomorphism of graded R-algebras:

H • (Grk (Cn )) ∼
= Λ•inv . ⊔
⊓

We want to determine the Poincaré polynomial of the complexified Z-graded space,

Λ•inv ⊗C X
Pk,ℓ (t) = tj dimC Λjinv ⊗ C = PGrk (Cn ) (t).
j

Denote the action of H on V0 by

H ∋ h 7→ Th ∈ Aut (V0 ).

Using the equality (3.4.10) of Subsection 3.4.4 we deduce

Z
Pk,ℓ (t) = | det(1V0 + tTh )|2 dVH (h), (7.4.2)
H

where dVH denotes the normalized bi-invariant volume form on H.

At this point, the above formula may look hopelessly complicated. Fortunately, it can
be dramatically simplified using a truly remarkable idea of H. Weyl.
Note first that the function

H ∋ h 7→ ϕ(h) = | det(1V0 + tTh )|2

is a class function, i.e., ϕ(ghg −1 ) = ϕ(h), ∀g, h ∈ H.

Inside H sits the maximal torus T = Tk × Tℓ , where

Tk = diag (eiθ1 , . . . , eiθk ) ∈ U (k) andbT ℓ = {diag (eiφ1 , . . . , eiφℓ ) ∈ U (ℓ)}.

Each h ∈ U (k) × U (ℓ) is conjugate to diagonal unitary matrix, i.e., there exists g ∈ H
such that ghg −1 ∈ T.
We can rephrase this fact in terms of the conjugation action of H on itself

C : H × H → H (g, h) 7→ C g (h) = ghg −1 .

7.4. SYMMETRY AND TOPOLOGY 303

The class functions are constant along the orbits of this conjugation action, and each
such orbit intersects the maximal torus T. In other words, a class function is completely
determined by its restriction to the maximal torus. Hence, it is reasonable to expect
that we ought to be able to describe the integral in (7.4.2) as an integral over T. This is
achieved in a very explicit manner by the next result.
Define ∆k : Tk → C by
Y i j √
∆k (θ 1 , . . . , θ k ) = (eiθ − eiθ ), i= −1.
1≤i<j≤n

On the unitary group U (k) we fix the bi-invariant metric m = mk such that at T1 U (k) we
have
m(X, Y ) := Re tr(XY ∗ ).
If we think of X, Y as k × k matrices, then
X
m(X, Y ) = Re xij ȳij .
i,j

We denote by dvk the volume form induced by this metric, and by Vk the total volume
Z
Vk := dvk .
U (k)

Theorem 7.4.28 (Weyl’s integration formula). Consider a class function ϕ on the group
G = U (k1 ) × · · · × U (ks ), denote by dg the volume form

dg = dvk1 ∧ · · · ∧ dvks ,

and by V the volume of G, V = Vk1 · · · Vks . Then

Z Z s
Y
1 1 1
ϕ(g)dg = Qs ϕ(θ 1 , . . . , θ s ) |∆kj (θ j )|2 dθ 1 ∧ · · · ∧ dθ s ,
V G k
j=1 j ! vol(T) T j=1

Z s
Y
1
= Qs kj k !
ϕ(θ 1 , . . . , θ s ) |∆kj (θ j )|2 dθ 1 ∧ · · · ∧ dθ s
j=1 (2π) j T j=1

Above, for every j = 1, . . . , s, we denoted by θ j the angular coordinates on Tkj ,

k
θ j := (θj1 , . . . , θj j ),

while dθ j denotes the bi-invariant volume on Tkj ,

k
dθ j := dθj1 ∧ · · · ∧ dθj j .
304 CHAPTER 7. COHOMOLOGY

The remainder of this subsection is devoted to the proof of this proposition. The
reader may skip this part at the first lecture, and go directly to Subsection 7.4.5 where
this formula is used to produce an explicit description of the Poincaré polynomial of a
complex Grassmannian.
Proof of Weyl’s integration formula To keep the main ideas as transparent as possi-
ble, we will consider only the case s = 1. The general situation is entirely similar. Thus,
G = U (k) and T = Tk . In this case, the volume form dg is the volume form associated to
metric mk .
We denote by Ad : G → Aut(LG ) the adjoint representation of G; see Example 3.4.30.
In this concrete case, Ad can be given the explicit description

Adg (X) = gXg −1 , ∀g ∈ U (k), X ∈ u(k).

Denote the angular coordinates θ on T by θ = (θ 1 , . . . , θ k ). The restriction of the metric

m to T is described in these coordinates by

m|T = (dθ 1 )2 + · · · + (dθ k )2 .

To any class function ϕ : G → C we associate the complex valued differential form

ωϕ := ϕ(g)dg.

Consider the homogeneous space G/T, and the smooth map

q : T × G/T → G (t, gT ) = gtg −1 .

Note that if g1 T = g2 T, then g1 tg1−1 = g1 tg2−1 , so the map q is well defined.

We have a m-orthogonal splitting of the Lie algebra LG

LG = LT ⊕ L⊥
G.

The tangent space to G/T at 1 · T ∈ G/T can be identified with L⊥ T . For this reason we
⊥
will write LG/T instead of LT .
Fix x ∈ G/T. Any g ∈ G defines a linear map Lg : Tx G/T → Tgx G/T. Moreover, if
gx = hx = y, then Lg and Lh differ by an element in the stabilizer of x ∈ G/T. This
stabilizer is isomorphic to T, and in particular it is connected.
Hence, if ω ∈ det Tx G/T, then Lg ω ∈ det Ty G/T and Lh ω ∈ det Ty G/T define the
same orientation of Ty G/T. In other words, an orientation in one of the tangent spaces of
G/T “spreads” via the action of G to an orientation of the entire manifold. Thus, we can
orient G/T by fixing an orientation on LG/T .
We fix an orientation on LG , and we orient LT using the form dθ = dθ 1 ∧ · · · ∧ dθ k .
The orientation on LG/T will be determined by the condition

or(LG ) = or(LT ) ∧ or(LG/T ).

The proof of Weyl’s integration formula will be carried out in three steps.
7.4. SYMMETRY AND TOPOLOGY 305

Step 1. Z Z
1
ωϕ = q ∗ ωϕ ∀ω.
G k! T×G/T

Step 2. We prove that there exists a positive constant Ck such that for any class
function ϕ on G we have
Z Z Z
k! ωϕ = q ωϕ = Ck ϕ(θ)|∆k (θ)|2 dθ.
∗
G T×G/T T

Step 3. We prove that

Vk Vk
Ck = =
vol (T) (2π)k

Step 1. We use the equality

Z Z
q ∗ ωϕ = deg q ωϕ ,
T×G/T G

so it suffices to compute the degree of q.

Denote by N (T) the normalizer of T in G, i.e.,

N (T) = g ∈ G ; gTg −1 ⊂ T ,

and then form the Weyl group

W := N (T)/T.
Lemma 7.4.29. The Weyl group W is isomorphic to the group Sk of permutations of k
symbols.

Proof. This is a pompous rephrasing of the classical statement in linear algebra that two
unitary matrices are similar if and only if they have the same spectrum, multiplicities
included. The adjoint action of N (T) on T= diagonal unitary matrices simply permutes
the entries of a diagonal unitary matrix. This action descends to an action on the quotient
W so that W ⊂ Sk .
Conversely, any permutation of the entries of a diagonal matrix can be achieved by a
conjugation. Geometrically, this corresponds to a reordering of an orthonormal basis. ⊔ ⊓
1 k
Lemma 7.4.30. Let α1 , . . . , αk ) ∈ Rk such that 1, α2π , . . . , α2π are linearly independent
over Q. Set τ := (exp(iα1 ), . . . , exp(iαk )) ∈ Tk . Then the sequence (τ n )n∈Z is dense in
Tk . (The element τ is said to be a generator of Tk .)
For the sake of clarity, we defer the proof of this lemma to the end of this subsection.
Lemma 7.4.31. Let τ ∈ Tk ⊂ G be a generator of Tk . Then q −1 (τ ) ⊂ T × G/T consists
of |W| = k! points.

Proof.
q(s, gT ) = τ ⇐⇒ gsg −1 = τ ⇐⇒ gτ g −1 = s ∈ T.
306 CHAPTER 7. COHOMOLOGY

In particular, gτ n g −1 = sn ∈ T, ∀n ∈ Z. Since (τ n ) is dense in T, we deduce

gT g−1 ⊂ T ⇒ g ∈ N (T).

Hence 
q −1 (τ ) = (g−1 τ g, gT ) ∈ T × G/T ; g ∈ N (T) ,
and thus q −1 (τ ) has the same cardinality as the Weyl group W. ⊔
⊓

The metric m on LG/T extends to a G-invariant metric on G/T. It defines a left-

invariant volume form dµ on G/T. Observe that the volume form on T induced by the
metric m is precisely
dθ = θ 1 ∧ · · · ∧ dθ k .
Lemma 7.4.32. q ∗ dg = |∆k (θ)|2 dθ ∧ dµ. In particular, any generator τ of Tk ⊂ G is a
regular value of q since ∆k (σ) 6= 0, ∀σ ∈ q −1 (τ ).

Proof. Fix a point x0 = (t0 , g0 T) ∈ T×G/T. We can identify Tx0 (T×G/T) with LT ⊕LG/T
using the isometric action of T × G on T × G/T.
Fix X ∈ LT , Y ∈ LG/T . For every real number s consider


hs = q t0 exp(sX) , g0 exp(sY )T = g0 exp(sY )t0 exp(sX) exp(−sY )g0−1 ∈ G.

We want to describe
d
|s=0 h−1
0 hs ∈ T1 G = LG .
ds
Using the Taylor expansions

exp(sX) = 1 + sX + O(s2 ), and exp(sY ) = 1 + sY + O(s2 ),

we deduce
h−1 −1 −1
0 hs = g0 t0 (1 + sY )t0 (1 + sX)(1 − sY )g0 + O(s )
2


= 1 + s g0 t−1 −1 −1
0 Y t0 g0 + g0 Xg0 − g0 Y g0
−1
+ O(s2 ).
Hence, the differential

Dx0 : Tx0 (T × G/T) ∼

= LT ⊕ LG/T → LT ⊕ LG/T ∼
= LG ,

can be rewritten as

Dx0 q(X ⊕ Y ) = Adg0 (Adt−1 − 1)Y + Adg0 X,

0

or, in block form, " #

1LT 0
Dx0 q = Adg0 .
0 Adt−1 − 1LG/T
0

The linear operator Adg is an m-orthogonal endomorphism of LG , so that det Adg = ±1.
On the other hand, since G = U (k) is connected, det Adg = det Ad1 = 1. Hence,

det Dx0 q = det(Adt−1 − 1LG/T ).

7.4. SYMMETRY AND TOPOLOGY 307

Now observe that LG/T = L⊥

T is equal to

X ∈ u(k) ; Xjj = 0 ∀j = 1, . . . , k ⊂ u(k) = LG .

Given t = diag (exp(iθ 1 ), . . . , exp(iθ k )) ∈ Tk ⊂ U (k), we can explicitly compute the

eigenvalues of Adt−1 acting on LG/T . More precisely, they are

exp(−i(θi − θj )) ; 1 ≤ i 6= j ≤ k .

Consequently
det Dx0 q = det(Adt−1 − 1) = |∆k (t)|2 . ⊔
⊓

Lemma 7.4.32 shows that q is an orientation preserving map. Using Lemma 7.4.31 and
Exercise 7.3.27 we deduce deg q = |W| = k!. Step 1 is completed.
Step 2 follows immediately from Lemma 7.4.32. More precisely, we deduce
Z Z !Z
1
ωϕ = dµ ϕ(θ)∆k (θ)dθ.
G k! G/T Tk
| {z }
=:Ck

To complete Step 3, that is, to find the constant Ck , we apply the above equality in the
special case ϕ = 1. so that ωϕ = dvk . We deduce
Z
Ck
Vk = ∆k (θ)dθ,
k! Tk

so that
Vk k!
Ck = R .
Tk ∆k (θ)dθ

Thus, we have to show that

Z
∆k (θ)dθ = k!vol (T) = (2π)k k!.
Tk

To compute the last integral we observe that ∆k (θ) can be expressed as a Vandermonde
determinant
1 1 ··· 1
1 2 k
eiθ eiθ ··· eiθ
1 2 k
∆k (θ) = e2iθ e2iθ ··· e2iθ
.. .. .. ..
. . . .
i(k−1)θ 1 i(k−1)θ 2 i(k−1)θ k
e e ··· e
This shows that we can write ∆k as a trigonometric polynomial
X
∆k (θ) = ǫ(σ)eσ (θ),
σ∈Sk
308 CHAPTER 7. COHOMOLOGY

where, for any permutation σ of {1, 2, . . . , k} we denoted by ǫ(σ) its signature, and by
eσ (θ) the trigonometric monomial
Qk iσ(j)θ j
1 k j=1 e
eσ(θ) = eσ (θ , . . . , θ ) = Qk .
iθ j
j=1 e

The monomials eσ are orthogonal with respect to the L2 -metric on Tk , and we deduce
Z XZ k
2
|∆k (θ)| dθ = |eσ (θ)|2 dθ = (2π)k k!.
Tk σ∈Sk T

This completes the proof of Weyl’s integration formula. ⊔

⊓

Proof of Lemma 7.4.30 We follow Weyl’s original approach ([129, 130]) in a modern
presentation.
Let X = C(T, C) denote the Banach space of continuous complex valued functions on
T. We will prove that
n Z
1 X j
lim f (τ ) = f dt, ∀f ∈ X. (7.4.3)
n→∞ n + 1 T
j=0

If U ⊂ T is an open subset and f is a continuous, non-negative function supported in U

(f 6≡ 0) then, for very large n,
n Z
1 X
f (τ j ) ≈ f dt 6= 0.
n+1 T
j=0

This means that f (τ j )

6= 0, i.e., τj
∈ U for some j.
To prove the equality (7.4.3) consider the continuous linear functionals Ln , L : X → C
n Z
1 X
Ln (f ) = f (τ j ), and L(f ) = f dt.
n+1 T
j=0

We have to prove that

lim Ln (f ) = L(f ) f ∈ X. (7.4.4)
n→∞
It suffices to establish (7.4.4) for any f ∈ S, where S is a subset of X spanning a dense
subspace. We let S be the subset consisting of the trigonometric monomials
eζ (θ 1 , . . . , θ k ) = exp(iζ1 θ 1 ) · · · exp(iζk θ k ), ζ = (ζ1 , . . . , ζk ) ∈ Zk .
The Weierstrass approximation theorem guarantees that this S spans a dense subspace.
We compute easily
n
1 X 1 eζ (α)n+1 − 1
Ln (eζ ) = ejζ (α) = .
n+1 n + 1 eζ (α) − 1
j=0
1 1
Since 1, 2π α1 , . . . , 2π αk are linearly independent over Q we deduce that eζ (α) 6= 1 for all
k
ζ ∈ Z . Hence Z
lim Ln (eζ ) = 0 = eζ dt = L(eζ ).
n→∞ T
Lemma 7.4.30 is proved. ⊔
⊓
7.4. SYMMETRY AND TOPOLOGY 309

7.4.5 The Poincaré polynomial of a complex Grassmannian

After this rather long detour we can continue our search for the Poincaré polynomial of
Grk (Cn ).
Let S0 denote the canonical subspace Ck ֒→ Cn . The tangent space of Grk (Cn ) at S0
can be identified with the linear space E of complex linear maps Ck → Cℓ , ℓ = n − k. The
isotropy group at S0 is H = U (k) × U (ℓ).
Exercise 7.4.33. Prove that the isotropy group H acts on E = {L : Ck → Cℓ } by

(T, S) · L = SLT ∗ ∀L ∈ E, T ∈ U (k), S ∈ U (ℓ). ⊔

⊓

Consider the maximal torus T k × T ℓ ⊂ H formed by the diagonal unitary matrices.

We will denote the elements of T k by
α
ε := (ε1 , . . . , εk ), εα = e2πiτ ,

and the elements of T ℓ by

e := (e1 , . . . , eℓ ), ej = exp(2πiθ j ),

The normalized measure on T k is then

dτ = dτ 1 ∧ · · · ∧ dτ k ,

and the normalized measure on T ℓ is

dθ = dθ 1 ∧ · · · ∧ dθ ℓ .

The element (ε, e) ∈ T k × T ℓ viewed as a linear operator on E has eigenvalues

{εα ej ; 1 ≤ α ≤ k 1 ≤ j ≤ ℓ}.

Using the Weyl integration formula we deduce that the Poincaré polynomial of Grk (Cn )
is Z
1 Y
Pk,ℓ (t) = |1 + tεα ej |2 |∆k (ε(τ ))|2 |∆ℓ (e(θ))|2 dτ ∧ dθ
k!ℓ! T k ×T ℓ
α,j
Z Y
1
= |εα + tej |2 |∆k (ε(τ ))|2 |∆ℓ (e(θ))|2 dτ ∧ dθ.
k!ℓ! T k ×T ℓ α,j

We definitely need to analyze the integrand in the above formula. Set

Y
Ik,ℓ (t) := (εα + tej )∆k (ε)∆ℓ (e),
α,j

so that Z
1
Pk,ℓ (t) = Ik,ℓ (t)Ik,ℓ (t)dτ ∧ dθ. (7.4.5)
k!ℓ! T k ×T ℓ
310 CHAPTER 7. COHOMOLOGY

6 5
5 4
5 4
3 3
1 3
1

Figure 7.4: The conjugate of (6,5,5,3,1) is (5,4,4,3,3,1)

We will study in great detail the formal expression

Y
Jk,ℓ (t; x; y) = (xα + tyj ).
α,j

The Weyl group W = Sk × Sℓ acts on the variables (x; y) by separately permuting the
x-components and the y-components. If (σ, ϕ) ∈ W then

Jk,ℓ (t; σ(x); ϕ(y)) = Jk,ℓ (t; x; y).

Thus we can write Jk,ℓ as a sum

X
Jk,ℓ (t) = td Qd (x)Rd (y)
d≥0

where Qd (x) and Rd (y) are symmetric polynomials in x and respectively y.

To understand the nature of these polynomials we need to introduce a very useful class
of symmetric polynomials, namely the Schur polynomials. This will require a short trip in
the beautiful subject of symmetric polynomials. An extensive presentation of this topic is
contained in the monograph [88].
A partition is a compactly supported, decreasing function

λ : {1, 2, . . .} → {0, 1, 2, . . .}.

We will describe a partition by an ordered finite collection (λ1 , λ2 , . . . , λn ), where λ1 ≥

· · · ≥ λn ≥ λn+1 = 0. The length of a partition λ is the number

L(λ) := max{n ; λn 6= 0}.

The weight of a partition λ is the number

X
|λ| := λn .
n≥1

Traditionally, one can visualize a partition using Young diagrams. A Young diagram
is an array of boxes arranged in left justified rows (see Figure 7.4). Given a partition
(λ1 ≥ · · · ≥ λn ), its Young diagram will have λ1 boxes on the first row, λ2 boxes on the
second row etc.
7.4. SYMMETRY AND TOPOLOGY 311

Any partition λ has a conjugate λ̂ defined by

λ̂n := #{j ≥ 0 ; λj ≥ n}.

The Young diagram of λ̂ is the transpose of the Young diagram of λ (see Figure 7.4).
A strict partition is a partition which is strictly decreasing on its support. Denote
by Pn the set of partitions of length ≤ n, and by P∗n the set of strict partitions λ of
length n − 1 ≤ L(λ) ≤ n. Clearly P∗n ⊂ Pn . Denote by δ = δn ∈ P∗n the partition
(n − 1, n − 2, . . . , 1, 0, . . .).
Remark 7.4.34. The correspondence Pn ∋ λ 7→ λ + δn ∈ P∗n . is a bijection. ⊔
⊓

To any λ ∈ P∗n we can associate a skew-symmetric polynomial

X
aλ (x1 , . . . , xn ) := det(xλj i ) = ǫ(σ)xλσ(i)
i
.
σ∈Sn

Note that aδn is the Vandermonde determinant

Y
aδ (x1 , . . . , xn ) = det(xin−1−i ) = (xi − xj ) = ∆n (x).
i<j

For each λ ∈ Pn we have λ + δ ∈ P∗n , so that aλ+δ is well defined and nontrivial.
Note that aλ+δ vanishes when xi = xj , so that the polynomial aλ+δ (x) is divisible by
each of the differences (xi − xj ) and consequently, it is divisible by aδ . Hence

aλ+δ (x)
Sλ (x) :=
aδ (x)
is a well defined polynomial. It is a symmetric polynomial since each of the quantities
aλ+δ and aδ is skew-symmetric in its arguments. The polynomial Sλ (x) is called the
Schur polynomial corresponding to the partition λ. Note that each Schur polynomial Sλ
is homogeneous of degree |λ|. We have the following remarkable result.
Lemma 7.4.35. X
Jk,ℓ (t) = t|λ| Sλ̂ (x)Sλ (y),
λ∈Pk,ℓ

where 
Pk,ℓ := λ ; λ1 ≤ k L(λ) ≤ ℓ .
For each λ ∈ Pk,ℓ we denoted by λ the complementary partition

λ = (k − λℓ , k − λℓ−1 , . . . , k − λ1 ).

Geometrically, the partitions in Pk,ℓ are precisely those partitions whose Young dia-
grams fit inside a ℓ × k rectangle. If λ is such a partition then the Young diagram of the
complementary of λ is (up to a 180◦ rotation) the complementary of the diagram of λ in
the ℓ × k rectangle (see Figure 7.5).
312 CHAPTER 7. COHOMOLOGY

6 7
4 6
4 5
2 3
1 3
0 1
Figure 7.5: The complementary of (6, 4, 4, 2, 1, 0) ∈ P7,6 is (7, 6, 5, 3, 3, 1) ∈ P6,7

For a proof of Lemma 7.4.35 we refer to [88], Section I.4, Example 5. The true essence
of the Schur polynomials is however representation theoretic, and a reader with a little
more representation theoretic background may want to consult the classical reference
[87], Chapter VI, Section 6.4, Theorem V, for a very exciting presentation of the Schur
polynomials, and the various identities they satisfy, including the one in Lemma 7.4.35.
Using (7.4.5), Lemma 7.4.35, and the definition of the Schur polynomials, we can
describe the Poincaré polynomial of Grk (Cn ) as
2
Z X
1
Pk,ℓ (t) = t|λ| aλ̂+δk (ε)aλ+δℓ (e) dτ ∧ dθ. (7.4.6)
k!ℓ! T k ×T ℓ λ∈Pk,ℓ

The integrand in (7.4.6) is a linear combination of trigonometric monomials εr11 · · · εrkk ·

es11 · · · esℓ ℓ , where the r-s and s-s are nonnegative integers.
Note that if λ, µ ∈ Pk,ℓ are distinct partitions, then the terms aλ̂+δ (ε) and aµ̂+δ (ε)
have no monomials in common. Hence
Z
aλ̂+δ (ε)aµ̂+δ (ε) dτ = 0.
Tk
Similarly, Z
aλ+δ (e)aµ+δ (e)dθ = 0, if λ 6= µ.
Tℓ
On the other hand, a simple computation shows that
Z
|aλ̂+δ (ε)|2 dτ = k!,
Tk
and Z
|aλ+δ (e)|2 dθ = ℓ!.
Tℓ
In other words, the terms
 
1 1/2
aλ̂+δ (ε)aλ+δ (e)
k!ℓ!
form an orthonormal system in the space of trigonometric (Fourier) polynomials endowed
with the L2 inner product. We deduce immediately from (7.4.6) that
X
Pk,ℓ (t) = t2|λ| . (7.4.7)
λ∈Pk,ℓ
7.4. SYMMETRY AND TOPOLOGY 313

k+1
R l+1 R
l+1

k+1
R
k+1
Figure 7.6: The rectangles Rℓ+1 and Rk+1 are “framed” inside Rℓ+1

The map
Pk,ℓ ∋ λ 7→ λ ∈ Pℓ,k ,
is a bijection so that X
Pk,ℓ (t) = t2|λ| = Pℓ,k (t). (7.4.8)
λ∈Pℓ,k

Computing the Betti numbers, i.e., the number of partitions in Pk,ℓ with a given weight
is a very complicated combinatorial problem, and currently there are no exact general
formulæ. We will achieve the next best thing, and rewrite the Poincaré polynomial as a
“fake” rational function.
Denote by bk,ℓ (w) the number of partitions λ ∈ Pℓ,k with weight |λ| = w. Hence

kℓ
X
Pk,ℓ (t) = bk,ℓ (w)t2w .
w=1

Alternatively, bk,ℓ (w) is the number of Young diagrams of weight w which fit inside a k × ℓ
rectangle.
Lemma 7.4.36.
bk+1,ℓ+1 (w) = bk,ℓ+1 (w) + bk+1,ℓ (w − ℓ − 1).
k+1
Proof. Look at the (k + 1) × ℓ rectangle Rk+1 inside the (k + 1) × (l + 1)-rectangle Rℓ+1
(see Figure 7.6). Then
 k+1
bk+1,ℓ+1 (w) = # diagrams of weight w which fit inside Rℓ+1

= # diagrams which fit inside Rk+1

+# diagrams which do not fit inside Rk+1 .
On the other hand,

bk+1,ℓ = # diagrams which fit inside Rk+1 .
314 CHAPTER 7. COHOMOLOGY

If a diagram does not fit inside Rk+1 this means that its first line consists of ℓ + 1 boxes.
When we drop this line, we get a diagram of weight w − ℓ − 1 which fits inside the
k × (ℓ + 1) rectangle Rℓ+1 of Figure 7.6. Thus, the second contribution to bk+1,ℓ+1 (w) is
bk,ℓ+1 (w − ℓ − 1). ⊔
⊓

The result in the above lemma can be reformulated as

Pk+1,ℓ+1 (t) = Pk+1,ℓ (t) + t2(ℓ+1) Pk,ℓ+1 (t).

Because the roles of k and ℓ are symmetric (cf. (7.4.8)) we also have

Pk+1,ℓ+1 (t) = Pk,ℓ+1 (t) + t2(k+1) Pk+1,ℓ (t).

These two equality together yield

Pk,ℓ+1 (1 − t2(ℓ+1 ) = Pk+1,ℓ (1 − t2(k+1) ).

Let m = k + ℓ + 1 and set Qd,m (t) = Pd,m−d (t) = PGd (m,C) (t). The last equality can be
rephrased as
1 − tm−k
Qk+1,m (t) = Qk,m · ,
1 − t2(k+1)
so that
1 − t2(m−k) 1 − t2(m−k+1) 1 − t2(n−1)
Qk+1,m(t) = · · · · .
1 − t2(k+1) 1 − t2k 1 − t4
Now we can check easily that b1,m−1 (w) = 1, i.e.,

1 − t2m
Q1,m (t) = 1 + t2 + t4 + · · · + t2(m−1) = .
1 − t2
Hence
(1 − t2(m−k+1) ) · · · (1 − t2m )
PGk (m,C) (t) = Qk,m(t) =
(1 − t2 ) · · · (1 − t2k )
(1 − t2 ) · · · (1 − t2m )
= .
(1 − t2 ) · · · (1 − t2k )(1 − t2 ) · · · (1 − t2(m−k) )
Remark 7.4.37. (a) The invariant theoretic approach in computing the cohomology
of Grk (Cn ) was used successfully for the first time by C. Ehresmann[41]. His method
was then extended to arbitrary compact, oriented symmetric spaces by H. Iwamoto [69].
However, we followed a different avenue which did not require Cartan’s maximal weight
theory.
(b) We borrowed the idea of using the Weyl’s integration formula from Weyl’s classical
monograph [133]. In turn, Weyl attributes this line of attack to R. Brauer. Our strategy
is however quite different from Weyl’s. Weyl uses an equality similar to (7.4.5) to produce
an upper estimate for the Betti numbers (of U (n) in his case) and then produces by hand
sufficiently many invariant forms. The upper estimate is then used to established that
these are the only ones. We refer also to [123] for an explicit description of the invariant
forms on Grassmannians. ⊔
⊓
7.5. ČECH COHOMOLOGY 315

Exercise 7.4.38. Show that the cohomology algebra of CPn is isomorphic to the truncated
ring of polynomials
R[x]/(xn+1 ),
where x is a formal variable of degree 2 while (xn+1 ) denotes the ideal generated by xn+1 .
Hint: Describe Λ•inv CPn explicitly. ⊔
⊓

7.5 Čech cohomology

In this last section we return to the problem formulated in the beginning of this chapter:
what is the relationship between the Čech and the DeRham approach. We will see that
these are essentially two equivalent facets of the same phenomenon. Understanding this
equivalence requires the introduction of a new and very versatile concept namely, the
concept of sheaf. This is done in the first part of the section. The second part is a fast
paced introduction to Čech cohomology. A concise yet very clear presentation of these
topics can be found in [63]. For a very detailed presentation of the classical aspects of
subject we refer to [52]. For a modern presentation, from the point of view of derived
categories we refer to the monographs [68, 73].

7.5.1 Sheaves and presheaves

Consider a topological space X. The topology TX on X, i.e., the collection of open subsets,
can be organized as a category. The morphisms are the inclusions U ֒→ V . A presheaf of
Abelian groups on X is a contravariant functor S : TX → Ab.
In other words, S associates to each open set an Abelian group S(U ), and to each
inclusion U ֒→ V , a group morphism rVU : S(V ) → S(U ) such that, if U ֒→ V ֒→ W , then
U = r U ◦ r V . If s ∈ S(V ) then, for any U ֒→ V , we set
rW V W

s |U := rVU (s) ∈ S(U ).

If f ∈ S(U ) then we define dom f := U .

The presheaves of rings, modules, vector spaces are defined in an obvious fashion.
Example 7.5.1. Let X be a topological space. For each open set U ⊂ X we denote by
C(U ) the space of continuous functions U → R. The assignment U 7→ C(U ) defines a
presheaf of R-algebras on X. The maps rVU are determined by the restrictions |U : C(V ) →
C(U ).
If X is a smooth manifold we get another presheaf U 7→ C ∞ (U ). More generally, the
differential forms of degree k can be organized in a presheaf Ωk (•). If E is a smooth vector
bundle, then the E-valued differential forms of degree k can be organized as a presheaf of
vector spaces
U 7→ ΩkE (U ) = Ωk (E |U ).
If G is an Abelian group equipped with the discrete topology, then the G-valued continuous
functions C(U, G) determine a presheaf called the constant G-presheaf which is denoted
by GX . ⊔
⊓
316 CHAPTER 7. COHOMOLOGY

Definition 7.5.2. A presheaf S on a topological space X is said to be a sheaf if the

following hold.

(a) If (Uα ) is an open cover of the open set U , and f, g ∈ S(U ) satisfy f |Uα = g |Uα , ∀α,
then f = g.

(b) If (Uα ) is an open cover of the open set U , and fα ∈ S(Uα ) satisfy

fα |Uα ∩Uβ = fβ |Uα ∩Uβ ∀Uα ∩ Uβ 6= ∅

then there exists f ∈ S(U ) such that, f |Uα = fα , ∀α. ⊔

⊓

Example 7.5.3. All the presheaves discussed in Example 7.5.1 are sheaves. ⊔
⊓

Example 7.5.4. Consider the presheaf S over R defined by

S(U ) := continuous, bounded functions f : U → R.

We let the reader verify this is not a sheaf since the condition (b) is violated. The reason
behind this violation is that in the definition of this presheaf we included a global condition
namely the boundedness assumption. ⊔
⊓

Definition 7.5.5. Let X be a topological space, and R a commutative ring with 1. We

equip R with the discrete topology
(a) A space of germs over X (“espace étalé” in the French literature) is a topological space
E, together with a continuous map π : E → X satisfying the following conditions.

(a1) The map π is a local homeomorphism that is, each point e ∈ E has a neighborhood
U such that π |U is a homeomorphism onto the open subset π(U ) ⊂ X. For every
x ∈ X, the set Ex := π −1 (x), is called the stalk at x.

(a2) There exist continuous maps

· : R × E → E,
and

+ : (u, v) ∈ E × E; π(u) = π(v) → E,
such that

∀r ∈ R, ∀x ∈ X, ∀u, v ∈ Ex we have r · u ∈ Ex , u + v ∈ Ex ,

and with respect to the operations +, ·, the stalk Ex is an R-module, ∀x ∈ X.

(b) A section of a space of germs π : E → X over a subset Y ⊂ X is a continuous function

s : Y → E such that s(y) ∈ Ey , ∀y ∈ Y . The spaces of sections defined over Y will be
denoted by E(Y ). ⊔
⊓
7.5. ČECH COHOMOLOGY 317

Example 7.5.6. (The space of germs associated to a presheaf ). Let S be a

presheaf of Abelian groups over a topological space X. For each x ∈ X define an equiva-
lence relation ∼x on
G
S(U ), U runs through the open neighborhoods of X,
U ∋x

by
f ∼x g ⇐⇒ ∃ open U ∋ x such thatf |U = g |U .
F
The equivalence class of f ∈ U ∋x S(U ) is denoted by [f ]x , and it is called the germ of f
at x. Set G
Sx := {[f ]x ; dom f ∋ x}, and bS := Sx .
x∈X

There exists a natural projection π : b

S → X which maps [f ]x to x. The “fibers” of this
map are π −1 (x) = Sx – the germs at x ∈ X. Any f ∈ S(U ) defines a subset

U f = [f ]u ; u ∈ U ⊂ bS.

We can define a topology in b S by indicating a basis of neighborhoods. A basis of open

neighborhoods of [f ]x ∈ b
S is given by the collection
 g
U ; U ∋ x, g ∈ S(U ) [g]x = [f ]x .

We let the reader check that this collection of sets satisfies the axioms of a basis of
neighborhoods as discussed e.g. in [77].
With this topology, each f ∈ S(U ) defines a continuous section of π over U

[f ] : U ∋ u 7→ [f ]u ∈ Su .

Note that each fiber Sx has a well defined structure of Abelian group

[f ]x + [g]x = [(f + g) |U ]x U ∋ x is open and U ⊂ dom f ∩ dom g.

(Check that this addition is independent of the various choices.) Since π : f (U ) → U is a

homeomorphism, it follows that π : bS → X is a space of germs. It is called the space of
germs associated to the presheaf S. ⊔
⊓

If the space of germs associated to a sheaf S is a covering space, we say that S is a

sheaf of locally constant functions (valued in some discrete Abelian group). When the
covering is trivial, i.e., it is isomorphic to a product X × discrete set , then the sheaf is
really the constant sheaf associated to a discrete Abelian group.
Example 7.5.7. (The sheaf associated to a space of germs). Consider a space of
π
germs E → X over the topological space X. For each open subset U ⊂ X, we denote by
E(U ) the space of continuous sections U → E. The correspondence U 7→ E(U ) clearly a
sheaf. E is called the sheaf associated to the space of germs. ⊔
⊓
318 CHAPTER 7. COHOMOLOGY

π
Proposition 7.5.8. (a) Let E → X be a space of germs. Then E b = E.

(b) A presheaf S over the topological space X is a sheaf if and only if b

S = S. ⊔
⊓

Exercise 7.5.9. Prove the above proposition. ⊔

⊓

Definition 7.5.10. If S is a presheaf over the topological space X, then the sheaf b
S is
called the sheaf associated to S, or the sheafification of S. ⊔
⊓

Definition 7.5.11. (a) Let A morphism between the (pre)sheaves of Abelian groups
(modules etc.) S and S̃ over the topological space X is a collection of morphisms of
Abelian groups (modules etc.) hU : S(U ) → S̃(U ), one for each open set U ⊂ X, such
that, when V ⊂ U , we have hV ◦ rUV = r̃UV ◦ hU . Above, rUV denotes the restriction
morphisms of S, while the r̃UV denotes the restriction morphisms of S̃. A morphism h is
said to be injective if each hU is injective.
(b) Let S be a presheaf over the topological space X. A sub-presheaf of S is a pair (T, ı),
where T is a presheaf over X, and ı : T → S is an injective morphism. The morphism ı is
called the canonical inclusion of the sub-presheaf. ⊔
⊓

Let h : S → T be a morphism of presheaves. The correspondence

U 7→ ker hU ⊂ S(U ),

defines a presheaf called the kernel of the morphism h. It is a sub-presheaf of S.

Proposition 7.5.12. Let h : S → T be a morphism of presheaves. If both S and T are
sheaves, then so is the kernel of h. ⊔
⊓

The proof of this proposition is left to the reader as an exercise.

π
Definition 7.5.13. (a) Let Ei →i X (i = 0, 1) be two spaces of germs over the same
topological space X. A morphism of spaces of germs is a continuous map h : E0 → E1
satisfying the following conditions.

(a1) π1 ◦ h = π0 , i.e., h(π0−1 (x)) ⊂ π1−1 (x), ∀x ∈ X.

(a2) For any x ∈ X, the induced map hx : π0−1 (x) → π1−1 (x) is a morphism of Abelian
groups (modules etc.).

The morphism h is called injective if each hx is injective.

π
(b) Let E → X be a space of germs. A subspace of germs is a pair (F, ), where F is a
space of germs over X and  : F → E is an injective morphism. ⊔
⊓

Proposition 7.5.14. (a) Let h : E0 → E1 be a morphism between two spaces of germs

π
over X. Then h(E0 ) →1 X is a space of germs over X called the image of h. It is denoted
by Im h, and it is a subspace of E1 . ⊔
⊓

Exercise 7.5.15. Prove the above proposition. ⊔

⊓
7.5. ČECH COHOMOLOGY 319

Lemma 7.5.16. Consider two sheaves S and T, and let h : S → T be a morphism. Then
h induces a morphism between the associated spaces of germs b b
h : Ŝ → T. ⊔
⊓

The definition of ĥ should be obvious. If f ∈ SU , and x ∈ U , then ĥ([f ]x ) = [h(f )]x ,

where h(f ) is now an element of T(U ). We let the reader check that ĥ is independent of
the various choices, and that it is a continuous map Ŝ → T̂ with respect to the topologies
described in Example 7.5.6.
The sheaf associated to the space of germs Im ĥ is a subsheaf of T called the image of
h, and denoted by Im h.

Exercise 7.5.17. Consider a morphism of sheaves over X, h : S → T. Let U ⊂ X be an

open set. Show that a section g ∈ T(U ) belongs to (Im h)(U ) if an only if, for every x ∈ X,
there exists an open neighborhood Vx ⊂ U , such that g |Vx = h(f ), for some f ∈ S(Vx ). ⊔ ⊓

☞ Due to Proposition 7.5.8(a), in the sequel we will make no distinction between sheaves
and spaces of germs.

Definition 7.5.18. (a) A sequence of sheaves and morphisms of sheaves,

nh hn+1
· · · → Sn → Sn+1 → Sn+2 → · · · ,

is said to be exact if Im hn = ker hn+1 , ∀n.

(b) Consider a sheaf S over the space X. A resolution of S is a long exact sequence

ı d d d
0 ֒→ S ֒→ S0 →0 S1 →1 · · · → Sn →
n
Sn+1 → · · · . ⊔
⊓

Exercise 7.5.19. Consider a short exact sequence of sheaves

0 → S−1 → S0 → S1 → 0.

For each open set U define S(U ) = S0 (U )/S−1 (U ).

(a) Prove that U 7→ S(U ) is a presheaf.
¯
(b) Prove that S1 ∼
= Ŝ = the sheaf associated to the presheaf S. ⊔
⊓

Example 7.5.20. (The DeRham resolution). Let M be a smooth n-dimensional

manifold. Using the Poincaré lemma and the Exercise 7.5.17 we deduce immediately that
the sequence
d d d
0 ֒→ RM ֒→ Ω0M → Ω1M → · · · → ΩnM → 0

is a resolution of the constant sheaf RM . ΩkM denotes the sheaf of k-forms on M while d
denotes the exterior differentiation. ⊔
⊓
320 CHAPTER 7. COHOMOLOGY

7.5.2 Čech cohomology

Let U 7→ S(U ) be a presheaf of Abelian groups over a topological space X. Consider
an open
 cover U = (Uα )α∈A of X. A simplex associated to U is a nonempty subset
A = α0 , . . . , αq ⊂ A, such that
q
\
UA := Uαi 6= ∅.
0

We define dim A to be one less the cardinality of A, dim A = |A| − 1. The set of all
q-dimensional simplices is denoted by Nq (U). Their union,
[
Nq (U)
q

is denoted by N(U), and it is called the nerve of the cover.

For every nonnegative integer q we set ∆q := {0, . . . , q}, and we define an ordered
q-simplex to be a map σ : ∆q → A with the property that σ(∆q ) is a simplex of N(U),
possibly of dimension < q. We will denote by N ~ q (U) the set of ordered q-simplices, and
we will use the symbol (α0 , . . . , αq ) to denote an ordered q-simplex σ : ∆q → A such that
σ(k) = αk .
Define Y
C q (S, U) := Sσ , Sσ := S(Uσ(∆q ) ).
~ q (U)
σ∈N

The elements of C q (S, U) are called Čech q-cochains subordinated to the cover U. In other
words, a q-cochain c associates to each ordered q-simplex σ an element hc, σi ∈ Sσ .
We have face operators
~ q (U) → N
∂j = ∂j : N ~ q−1 (U), j = 0, . . . , q,

where for a ordered q-simplex σ = (α0 , . . . , αq ), we set

∂ j σ = ∂qj σ = (α0 , . . . , α̂j , . . . , αq ) ∈ U(q−1) ,

and where a hat “ˆ” indicates a missing entry.

i ∂ j = ∂ j−1 ∂ i , for all j > i.
Exercise 7.5.21. Prove that ∂q−1 ⊔
⊓
q q−1 q

We can now define an operator

δ : C q−1 (S, U) → C q (S, U),

which assigns to each (q −1)-cochain c, a q-cochain δc whose value on an ordered q-simplex

σ is given by
Xq
hδc, σi := (−1)j hc, ∂ j σi |Uσ .
j=0

Using Exercise 7.5.21 above we deduce immediately the following result.

7.5. ČECH COHOMOLOGY 321

Lemma 7.5.22. δ2 = 0 so that

δ δ δ δ
0 ֒→ C 0 (S, U) → C 1 (S, U) → · · · → C q (S, U) → · · ·

is a cochain complex.
The cohomology of this cochain complex is called the Čech cohomology of the cover U
with coefficients in the pre-sheaf S.
Example 7.5.23. Let U and S as above. A 0-cochain is a correspondence which associates
to each open set Uα ∈ U an element cα ∈ S(Uα ). It is a cocycle if, for any pair (α, β) such
that Uα ∩ β 6= ∅, we have
cβ − cα = 0.
Observe that if S is a sheaf, then the collection (cα ) determines a unique section of S.
A 1-cochain associates to each pair (α, β) such that Uαβ 6= ∅ an element

cαβ ∈ S(Uαβ ).

This correspondence is a cocycle if, for any ordered 2-simplex (α, β, γ) we have

cβγ − cαγ + cαβ = 0.

For example, if X is a smooth manifold, and U is a good cover, then we can associate to
each closed 1-form ω ∈ Ω1 (M ) a Čech 1-cocycle valued in RX as follows.
First, select for each Uα a solution fα ∈ C ∞ (Uα ) of

dfα = ω.

Since d(fα − fβ ) ≡ 0 on Uαβ , we deduce there exist constants cαβ such that fα − fβ = cαβ .
Obviously this is a cocycle, and it is easy to see that its cohomology class is independent
of the initial selection of local solutions fα . Moreover, if ω is exact, this cocycle is a
coboundary. In other words we have a natural map


H 1 (X) → H 1 N(U), RX .

We will see later this is an isomorphism. ⊔

⊓

Definition 7.5.24. Consider two open covers U = (Uα )α∈A , and V = (Vβ )β∈B of the same
topological space X. We say V is finer than U, and we write this U ≺ V, if there exists a
map ̺ : B → A, such that
Vβ ⊂ U̺(β) ∀β ∈ B.
The map ̺ is said to be a refinement map. ⊔
⊓

Proposition 7.5.25. Consider two open covers U = (Uα )α∈A and V = (Vβ )β∈B of the
same topological space X such that U ≺ V. Fix a sheaf of Abelian groups S. Then the
following are true.
(a) Any refinement map ̺ induces a cochain morphism
M M
̺∗ : C q (U, S) → C q (V, S).
q q
322 CHAPTER 7. COHOMOLOGY

(b) If r : A → B is another refinement map, then ̺∗ is cochain homotopic to r∗ . In

particular, any relation U ≺ V defines a unique morphism in cohomology

ıVU : H • (U, S) → H • (V, S).

(c) If U ≺ V ≺ W, then
ıW W V
U = ıV ◦ ıU .

Proof. (a) We define ̺∗ : C q (U) → C q (V) by

S(Vσ ) ∋ h̺∗ (c), σi := hc, ̺(σ)i |Vσ ∀c ∈ C q (U) σ ∈ V(q)) ,

where, by definition, ̺(σ) ∈ Aq is the ordered q-simplex (̺(β0 ), . . . , ̺(βq )). The fact that
̺∗ is a cochain morphism follows immediately from the obvious equality

̺ ◦ ∂qj = ∂qj ◦ ̺.

~ q−1 (V) → N
(b) We define hj : N ~ q (U) by

hj (β0 , . . . , βq−1 ) = (̺(β0 ), . . . , ̺(βj ), r(βj ), · · · , r(βq−1 )).

The reader should check that hj (σ) is indeed an ordered simplex of U for any ordered
simplex σ of V. Note that Vσ ⊂ Uhj (σ) ∀j. Now define

χ := χq : C q (U) → C q−1 (V)

by
q−1
X
hχq (c), σi := (−1)j hc, hj (σ)i |Vσ ∀c ∈ C q (U) ∀σ ∈ V(q−1) .
j=0

We will show that

δ ◦ χq (c) + χq+1 ◦ δ(σ) = ̺∗ (c) − r∗ (c) ∀c ∈ C q (U).

~ q (V), and set

Let σ = (β0 , . . . , βq ) ∈ N

̺(σ) := (λ0 , . . . , λq ), r(σ) = (µ0 , . . . , µq ) ∈ U(q) ,

so that,
hj (σ) = (λ0 , . . . , λj , µj , . . . , µq ).
Then X
hχ ◦ δ(c), σi = (−1)j hδc, hj (σ)i |Vσ
q
X q+1
X
= (−1) ( (−1)k hc, ∂q+1
j k
hj (σ)i |Vσ )
j=0 k=0
q
X j
X
= (−1)j (−1)k hc, (λ0 , . . . , λ̂j , . . . , λj , µj , . . . , µq )i |Vσ
j=0 k=0
7.5. ČECH COHOMOLOGY 323

q
X
+ (−1)ℓ+1 hc, (λ0 , . . . , λj , µj , . . . , µ̂k , . . . , µq )i |Vσ 
ℓ=j
q
X j−1
X
= (−1)j (−1)k hc, (λ0 , . . . , λ̂j , . . . , λj , µj , . . . , µq )i |Vσ
j=0 k=0

q
X
+ (−1)ℓ+1 hc, (λ0 , . . . , λj , µj , . . . , µ̂k , . . . , µq )i |Vσ 
ℓ=j+1
q
X 
+ (−1)j hc, (λ0 , . . . , λj−1 , µj , . . . , µq )i |Vσ +hc, (λ0 , . . . , λj , µj+1 , . . . , µq )i |Vσ .
j=0
The last term is a telescopic sum which is equal to
hc, (λ0 , . . . , λq i |Vσ −hc, (µ0 , . . . , µq )i |Vσ = h̺∗ c, σi − hr∗ c, σi.
If we change the order of summation in the first two term we recover the term h−δχc, σi.
Part (c) is left to the reader as an exercise. ⊔
⊓

We now have a collection of graded groups

 •
H (U, S) ; U − open cover of X ,
and morphisms n o
ıVU : H • (U, S) → H • (V, S); U ≺ V ,
such that,
U = 1 and ıU = ıV ◦ ıU ,
ıU W W V

whenever U ≺ V ≺ W. We can thus define the inductive (direct) limit

H • (X, S) := lim H • (U, S).
U

The group H • (X, S) is called the Čech cohomology of the space X with coefficients in the
pre-sheaf S.
Let us briefly recall the definition of the direct limit. One defines an equivalence
relation on the disjoint union a
H • (U, S),
U
by
H • (U) ∋ f ∼ g ∈ H • (V) ⇐⇒ ∃W ≻ U, V : ıW W
U f = ıV g.
We denote the equivalence class of f by f . Then
!
a
• •
lim H (U) = H (U, S) / ∼ .
U
U

Note that we have canonical morphisms,

ıU : H • (U, S) → H • (X, S).
324 CHAPTER 7. COHOMOLOGY

Example 7.5.26. Let S be a sheaf over the space X. For any open cover U = (Uα ),
a 0-cycle subordinated to U is a collection of sections fα ∈ S(Uα ) such that, every time
Uα ∩ Uβ 6= ∅, we have
fα |Uαβ = fβ |Uαβ .
According to the properties of a sheaf, such a collection defines a unique global section
f ∈ S(X). Hence, H 0 (X, S) ∼
= S(X). ⊔
⊓

Proposition 7.5.27. Any morphism of pre-sheaves h : S0 → S1 over X induces a mor-

phism in cohomology
h∗ : H • (X, S0 ) → H • (X, S1 ).
Sketch of proof. Let U be an open cover of X. Define

h∗ : C q (U, S0 ) → C q (U, S1 ),

by
hh∗ c, σi = hU (hc, σi) ∀c ∈ C q (U, S0 ) σ ∈ U( q).
The reader can check easily that h∗ is a cochain map so it induces a map in cohomology
• •
hU
∗ : H (U, S0 ) → H (U, S1 )

which commutes with the refinements ıVU . The proposition follows by passing to direct
limits. ⊔
⊓

Theorem 7.5.28. Let

 p
0 → S−1 → S0 → S1 → 0,
be an exact sequence of sheaves over a paracompact space X. Then there exists a natural
long exact sequence
∗ p∗ δ
· · · → H q (X, S−1 ) → H q (X, S0 ) → H q (X, S1 ) →
∗
H q+1 (X, S−1 ) → · · · .

Sketch of proof. For each open set U ⊂ X define S(U ) := S0 (U )/S−1 (U ). Then the
correspondence U 7→ S(U ) defines a pre-sheaf on X. Its associated sheaf is isomorphic
with S1 (see Exercise 7.5.19). Thus, for each open cover U we have a short exact sequence
 π
0 → C q (U, S−1 ) → C q (U, S0 ) → C q (U, S) → 0.

We obtain a long exact sequence in cohomology

· · · → H q (U, S−1 ) → H q (U, S0 ) → H q (U, S) → H q+1 (U, X) → · · · .

Passing to direct limits we get a long exact sequence

· · · → H q (X, S−1 ) → H q (X, S0 ) → H q (X, S) → H q+1 (X, S−1 ) → · · · .

To conclude the proof of the proposition we invoke the following technical result. Its proof
can be found in [119].
7.5. ČECH COHOMOLOGY 325

Lemma 7.5.29. If two pre-sheaves S, S′ over a paracompact topological space X have

isomorphic associated sheaves then

H • (X, S) ∼
= H • (X, S′ ). ⊔
⊓

Definition 7.5.30. A sheaf S is said to be fine if, for any locally finite open cover U =
(Uα )α∈A there exist morphisms hα : S → S with the following properties.

(a) For any α ∈ A there exists a closed set Cα ⊂ Uα , and hα (Sx ) = 0 for x 6∈ Cα , where
Sx denotes the stalk of S at x ∈ X.
P
(b) α hα = 1S . This sum is well defined since the cover U is locally finite. ⊔
⊓

Example 7.5.31. Let X be a smooth manifold. Using partitions of unity we deduce that
the sheaf ΩkX of smooth k-forms is fine. More generally, if E is a smooth vector bundle
over X then the space ΩkE of E-valued k-forms is fine. ⊔
⊓

Proposition 7.5.32. Let S be a fine sheaf over a paracompact space X. Then H q (X, S) ∼
=
0 for q ≥ 1.

Proof. Because X is paracompact, any open cover admits a locally finite refinement. Thus,
it suffices to show that, for each locally finite open cover U = (Uα )α∈A , the cohomology
groups H q (U, S) are trivial for q ≥ 1. We will achieve this by showing that the identity
map C q (U, S) → C q (U, S) is cochain homotopic with the trivial map. We thus need to
produce a map
χq : C q (U, S) → C q−1 (U, S),
such that
χq+1 δq + δq−1 χq = 1. (7.5.1)
Consider the morphisms hα : S → S associated to the cover U postulated by the definition
of a fine sheaf. For every α ∈ A, σ ∈ N ~ q (U), and every f ∈ C q (U, S), we construct
htα (f ), σi ∈ S(Uσ ) as follows. Consider the open cover of Uσ

V = Uα ∩ Uσ , W := Uσ \ Cα , (supp hα ⊂ Cα ).

Note that hα f |V ∩W = 0 and, according to the axioms of a sheaf, the section hα (f |V ) can
be extended by zero to a section htα (f ), σi ∈ S(Uσ ). Now, for every f ∈ C q (U, S), define
χq f ∈ C q−1 (U, S) by X
hχq (f ), σi := htα (f ), σi.
α
The above sum is well defined since the cover U is locally finite. We let the reader check
that χq satisfies (7.5.1). ⊔
⊓

Definition 7.5.33. Let S be a sheaf over a space X. A fine resolution is a resolution

d d
0 → S ֒→ S0 → S1 → · · · ,

such that each of the sheaves Sj is fine. ⊔

⊓
326 CHAPTER 7. COHOMOLOGY

Theorem 7.5.34 (Abstract DeRham theorem). Let

d0 d1
0 ֒→ S → S0 → S1 → · · ·

be a fine resolution of the sheaf S over the paracompact space X. Then

d d
0 → S0 (X) →0 S1 (X) →1 · · ·

is a cochain complex, and there exists a natural isomorphism

H q (X, S) ∼
= H q (Sq (X)).

Proof. The first statement in the theorem can be safely left to the reader. For q ≥ 1,
denote by Zq the kernel of the sheaf morphism dq . We set for uniformity Z0 := S. We get
a short exact sequence of sheaves

0 → Zq → Sq → Zq+1 → 0 q ≥ 0. (7.5.2)

We use the associated long exact sequence in which H k (X, Sq ) = 0, for k ≥ 1, since Sq is
a fine sheaf. This yields the isomorphisms

H k−1 (X, Zq+1 ) ∼

= H k (X, Zq ) k ≥ 2.

We deduce inductively that

H m (X, Z0 ) ∼
= H 1 (X, Zm−1 ) m ≥ 1. (7.5.3)

Using again the long sequence associated to (7.5.2), we get an exact sequence

dm−1
H 0 (X, Zm−1 ) ∗→ H 0 (X, Zm ) → H 1 (X, Zm−1 ) → 0.

We apply the computation in Example 7.5.26, and we get

H 1 (X, Zm−1 ) ∼
= Zm (X)/dm−1
∗ Sm−1 (X) .

This is precisely the content of the theorem. ⊔

⊓

Corollary 7.5.35. Let M be a smooth manifold. Then

H • (M, RM ) ∼
= H • (M ).

Proof. The manifold M is paracompact. We conclude using the fine resolution

d
0 → RM ֒→ Ω0M → Ω1M → · · · . ⊔
⊓

Exercise 7.5.36. Describe explicitly the isomorphisms

H 1 (M ) ∼
= H 1 (M, RM ) and H 2 (M ) ∼
= H 2 (M, RM ). ⊔
⊓
7.5. ČECH COHOMOLOGY 327

Remark 7.5.37. The above corollary has a surprising implication. Since the Čech co-
homology is obviously a topological invariant, so must be the DeRham cohomology which
is defined in terms of a smooth structure. Hence if two smooth manifolds are homeo-
morphic they must have isomorphic DeRham groups, even if the manifolds may not be
diffeomorphic.
Such exotic situations do exist. In a celebrated paper [93], John Milnor has constructed
a family of nondiffeomorphic manifolds all homeomorphic to the sphere S 7 . More recently,
the work of Simon Donaldson in gauge theory was used by Michael Freedman to construct
a smooth manifold homeomorphic to R4 but not diffeomorphic with R4 equipped with the
natural smooth structure. (This is possible only for 4-dimensional vector spaces!) These
three mathematicians, J. Milnor, S. Donaldson and M. Freedman were awarded Fields
medals for their contributions. ⊔
⊓

Theorem 7.5.38. (Leray) Let M be a smooth manifold and U = (Uα )α∈A a good cover
of M , i.e.,
Uσ ∼
= Rdim M .
Then
H • (U, RM ) ∼
= H • (M ).

Proof. Let Zk denote the sheaf of closed k-forms on M . Using the Poincaré lemma we
deduce
Zk (Uσ ) = dΩk−1 (Uσ ).
We thus have a short exact sequence
d
0 → C q (U, Zk ) → C q (U, Ωk−1 ) →
∗
C q (U, Zk ) → 0.

Using the associated long exact sequence and the fact that Ωk−1 is a fine sheaf, we deduce
as in the proof of the abstract DeRham theorem that

H q (U, RM ) ∼
= H q−1 (U, Z1 ) ∼
= ···
∼
= H 1 (U, Zq−1 ) ∼
= H 0 (U, Zq )/d∗ H 0 (U, Ωq−1 ) ∼
= Zq (M )/dΩq−1 (M ) ∼
= H • (M ). ⊔
⊓
Remark 7.5.39. The above result is a special case of a theorem of Leray: if S is a sheaf
on a paracompact space X, and U is an open cover such that

H q (UA , S) = 0, ∀q ≥ 1 A ∈ N(U),

then H • (U, S) = H • (X, S). For a proof we refer to [52].

When GM is a constant sheaf, where G is an arbitrary Abelian group, we have a
Poincaré lemma (see [42], Chapter IX, Thm. 5.1.),

H q (Rn , G) = 0 q ≥ 1.

Hence, for any good cover U

H • (M, GM ) = H • (U, GM ). ⊔
⊓
328 CHAPTER 7. COHOMOLOGY

Remark 7.5.40. A combinatorial simplicial complex, or a simplicial scheme, is a collec-

tion K of nonempty, finite subsets of a set V with the property that any nonempty subset
B of a subset A ∈ K must also belong to the collection K, i.e.,
A ∈ K, B ⊂ A, B 6= ∅ =⇒ B ∈ K.
The set V is called the vertex set of K, while the subsets in K are called the (open) faces)
of the simplicial scheme. The nerve of an open cover is an example of simplicial scheme.
To any simplicial scheme K, with vertex set V , we can associate a topological space
|K| in the following way.
Denote by RV the vector space of all maps f : V → R with the property that f (v) = 0,
for all but finitely many v’s. In other words,
M
RV = R.
v∈V

Note that RV has a canonical basis given by the Dirac functions

(
1 u=v
δv : V → R, δv (u) =
0 u 6= v.

We topologize RV by declaring a subset C ⊂ RV closed if the intersection of C with any

finite dimensional subspace of RV is a closed subset with respect to the Euclidean topology
of that finite dimensional subspace. 
For any face F ⊂ K we denote by ∆F ⊂ RV the convex hull of the set δv ; v ∈ F .
Note that ∆F is a simplex of dimension |F | − 1. Now set
[
|K| = ∆F .
F ∈K

We will say that |K| is the geometric realization of the simplicial scheme K.
A result of Borsuk-Weil (see [19]) states that if U = (Uα )α∈A is a good cover of a com-
pact manifold M , then the geometric realization of the nerve N(U) is homotopy equivalent
to M . Thus, the nerve of a good cover contains all the homotopy information about the
manifold. Since the DeRham cohomology is homotopy invariant, Leray’s theorem comes
as no suprise. What is remarkable is the explicit way in which one can extract the coho-
mological information from the combinatorics of the nerve.
While at this point, we should remark that any compact manifold admits finite good
covers. This shows that the homotopy type is determined by a finite, albeit very large,
set of data. ⊔
⊓

Example 7.5.41. Let M be a smooth manifold and U = (Uα )α∈A a good cover of M . A
1-cocycle of RM is a collection of real numbers fαβ - one for each pair (α, β) ∈ A2 such
that Uαβ 6= ∅ satisfying
fαβ + fβγ + fγα = 0,
whenever Uαβγ 6= ∅. The collection is a coboundary if there exist the constants fα such
that fαβ = fβ − fα . This is precisely the situation encountered in Subsection 7.1.2. The
abstract DeRham theorem explains why the Čech approach is equivalent with the DeRham
approach. ⊔
⊓
7.5. ČECH COHOMOLOGY 329

Remark 7.5.42. Often, in concrete applications it is convenient to work with skew-

symmetric Čech cochains. A cochain c ∈ C q (S, U) is skew-symmetric if for any ordered
q-simplex σ = (α0 , . . . , αq ), and for any permutation ϕ of ∆q we have

hc, (α0 , . . . , αq )i = ǫ(ϕ)hc, (αϕ(0) , . . . , αϕ(q) )i.

One can then define a “skew-symmetric” Čech cohomology following the same strategy.
The resulting cohomology coincides with the cohomology described in this subsection. For
a proof of this fact we refer to [119]. ⊔
⊓

Exercise 7.5.43. A sheaf of Abelian groups S on a paracompact space X is called soft if

for every closed subset C ⊂ X the restriction map S(X) → S(C) is surjective. The sheaf
S is called flabby if, for any open subset U ⊂ X, the restriction map S(X) → S(U ) is
surjective.
(a) Prove that any flabby sheaf is also soft.
(b) Suppose 0 → S0 → S1 → S2 → 0 is a short exact sequence of sheaves on X. Show that
if S0 is a flabby sheaf, then the sequence of Abelian groups

0 → S0 (X) → S1 (X) → S2 (X) → 0

is exact. ⊔
⊓

Exercise 7.5.44. Suppose X is a topological space and S is a sheaf of Abelian groups on

X.
(a) Prove that S admits a flabby resolution, i.e., there exists a sequence of flabby sheaves
over X, Sk , k ≥ 0, and morphisms of sheaves fk : Sk−1 → Sk , k = 0, 1, . . . , S−1 = S, such
that the sequence below is exacft
f0 f1
0 → S −→ S0 −→ S1 −→ · · · .

(b) Prove that if

f0 f1
0 → S −→ S0 −→ S1 −→ · · ·
is a flabby resolution of S, then the cohomology groups H k (X, S) are isomorphic to the
cohomology groups of the cochain complex
f0
S0 (X) −→ S1 (X) −→ · · · . ⊔
⊓
330 CHAPTER 7. COHOMOLOGY
Chapter 8

Characteristic Classes

We now have sufficient background to approach a problem formulated in Chapter 2: find a

way to measure the “extent of nontriviality” of a given vector bundle. This is essentially a
topological issue but, as we will see, in the context of smooth manifolds there are powerful
differential geometric methods which will solve a large part of this problem. Ultimately,
only topological techniques yield the best results.

8.1 Chern–Weil theory

8.1.1 Connections in principal G-bundles
connection.
All the Lie groups we will consider will be assumed to be matrix Lie groups, i.e., Lie
subgroups of a general linear group GL(n, K) = GL(Kn ).
This restriction is neither severe, nor necessary. It is not severe since, according to
a nontrivial result (Peter-Weyl theorem), any compact Lie group is isomorphic with a
matrix Lie group, and these groups are sufficient for most applications in geometry. It is
not necessary since all the results of this subsection are true for any Lie group. We stick
with this assumption since most proofs are easier to “swallow” in this context.
The Lie algebra g of a matrix Lie group G is a Lie algebra of matrices in which the
bracket is the usual commutator.
Let M be a smooth manifold. Recall that a principal G-bundle P over M can be
defined by an open cover (Uα ) of M and a gluing cocycle

gαβ : Uαβ → G.

The Lie group G operates on its Lie algebra g via the adjoint action

Ad : G → GL(g), g 7→ Ad(g) ∈ GL(g),

where
Ad(g)X := gXg−1 , ∀X ∈ g, g ∈ G.
We denote by Ad(P ) the vector bundle with standard fiber g associated to P via the
adjoint representation. In other words, Ad(P ) is the vector bundle defined by the open

331
332 CHAPTER 8. CHARACTERISTIC CLASSES

cover (Uα ), and gluing cocycle

Ad(gαβ ) : Uαβ → GL(g).

The bracket operation in the fibers of Ad(P ) induces a bilinear map

[·, ·] : Ωk Ad(P ) × Ωℓ Ad(P ) → Ωk+ℓ Ad(P ) ,

defined by
[ω k ⊗ X, η ℓ ⊗ Y ] := (ω k ∧ η ℓ ) ⊗ [X, Y ], (8.1.1)


for all ω k ∈ Ωk (M ), η ℓ ∈ Ωℓ (M ), and X, Y ∈ Ω0 Ad(P ) .
Exercise 8.1.1. Prove that for any ω, η, φ ∈ Ω∗ (Ad(P )) the following hold.

[ω, η] = −(−1)|ω|·|η| [η, ω], (8.1.2)

[ [ω, η], φ] = [ [ω, φ], η] + (−1)|ω|·|φ| [ω, [η, φ] ]. (8.1.3)

In other words, Ω• Ad(P ) , [ , ] is a super Lie algebra. ⊔
⊓

Using Proposition 3.3.5 as inspiration we introduce the following fundamental concept.

Definition 8.1.2. (a) A connection on the principal bundle P defined by the open cover
U = (Uα )α∈A , and the gluing cocycle gβα : Uαβ → G is a collection

Aα ∈ Ω1 (Uα ) ⊗ g,

satisfying the transition rules

−1 −1
Aβ (x) = gαβ (x)dgαβ (x) + gαβ (x)Aα (x)gαβ (x)
−1 −1
= −( dgβα (x) )gβα (x) + gβα (x)Aα (x)gβα (x), ∀x ∈ Uαβ .

We will denote by A(U, g•• ) the set of connections defined by the open cover U and the
gluing cocycle g•• : U•• → G.
(b) The curvature of a connection A ∈ A(U, g•• ) is defined as the collection Fα ∈ Ω2 (Uα )⊗g
where
1
Fα = dAα + [Aα , Aα ]. ⊔
⊓
2
Remark 8.1.3. Given an open cover (Uα ) of M , then two gluing cocycles

gβα , hβα : Uαβ → G,

define isomorphic principal bundles if and only if there exists smooth maps

Tα : Uα → G,

such that
hβα (x) = Tβ (x)gβα (x)Tα (x)−1 , , ∀α, β, ∀x ∈ Uαβ .
If this happens, we say that the cocycles g•• and h•• are cohomologous.
8.1. CHERN–WEIL THEORY 333

Suppose (gαβ ), (hαβ ) are two such cohomologous cocycles. Then the maps Tα induce
a well defined correspondence

T : A(U, g•• ) → A(U, h•• )

given by
T

A(U, g•• ) ∋ (Aα ) 7−→ Bα := −(dTα )Tα−1 + Tα Aα Tα−1 ∈ A(U, h•• ).

The correspondence T is a bijection.

Suppose U = (Uα )α∈A and V = (Vi )i∈I are open covers, and gβα : Uαβ → G is a gluing
cocycle. Suppose that the open cover (Vi )i∈I is finer that the cover (Uα )α∈A , i.e., there
exists a map ϕ : I → A such that

Vi ⊂ Uϕ(i) , ∀i ∈ I.
ϕ
We obtain a new gluing cocycle gij : Vij → G given by
ϕ
gij (x) = gϕ(i)ϕ(j) (x), ∀i, j ∈ I, x ∈ Vij .

This new cocycle defines a principal bundle isomorphic to the principal bundle defined by
the cocycle (gβα ). If (Aα ) is a connection defined by the open cover (Uα ) and the cocycle
(gβα ), then it induces a connection

Aϕ
i = Aϕ(i) |Vi

defined by the open cover (Vi ), and the cocycle gij . We say that the connection Aϕ is the
V-refinement of the connection A. The correspondence
ϕ
TVU : A(U, g•• ) → A(V, g•• ), A 7→ Aϕ

is also a bijection.
Finally, given two pairs (open cover, gluing cocycle), (U, g•• ) and (V, h•• ), both de-
scribing the principal bundle P → M , there exists an open cover W, finer that both U
and W, and cohomologous gluing cocycles ḡ•• , h̄•• : W•• → G refining g•• and respectively
h•• , such that the induced maps

A(U, g•• ) −→ A(W, ḡ•• ) −→ A(W, h̄•• ) ←− A(V, h•• )

are bijections.
For simplicity, we will denote by A(P ), any of these isomorphic spaces A(U, g•• ).
☞ All the constructions that we will perform in the remainder of this section are com-
patible with the above isomorphisms but, in order to keep the presentation as transparent
as possible, in our proofs we will not keep track of these isomorphisms. We are convinced
that the reader can easily supply the obvious, and repetitious missing details. ⊔
⊓

Proposition 8.1.4. Suppose P → M is a principal G-bundle described by an open cover

U = (Uα )α∈A , and gluing cocycle gβα : Uαβ → G.
(a) The set A(P ) of connections on P is an affine space modelled by Ω1 ( Ad(P ) ).
334 CHAPTER 8. CHARACTERISTIC CLASSES

(b) For any connection A = (Aα ) ∈ A(U, g•• ), the collection (Fα ) defines a global Ad(P )-
valued 2-form. We will denote it by FA or F (A), and we will refer to it as the curvature
of the connection A.
(c) (The Bianchi identity.)

dFα + [Aα , Fα ] = 0, ∀α. (8.1.4)

Proof. (a) If (Aα )α∈A , (Bα )α∈A ∈ A(P ), then their difference Cα = Aα − Bα satisfies the
gluing rules
−1
Cβ = gβα Cα gβα ,
so that it
defines an element of Ω1 (Ad(P )). Conversely, if (Aα ) ∈ A(P ), and ω ∈
Ω1 Ad(P ) , then ω is described by a collection of g-valued 1-forms ωα ∈ Ω1 (Uα ) ⊗ g,
satisfying the gluing rules
−1
ωβ |Uαβ = gβα ωα |Uαβ gβα .
The collection A′α := Aα + ωα is then a connection on P . This proves that if A(P ) is
nonempty, then it is an affine space modelled by Ω1 ( Ad(P ) ).
To prove that A(P ) 6= ∅ we consider a partition of unity subordinated to the cover
Uα . More precisely, we consider a family of nonnegative smooth functions uα : M → R,
α ∈ A, such that supp uα ⊂ Uα , ∀α, and
X
uα = 1.
α

For every α ∈ A we set

X
−1
Bα := uγ gγα dgγα ∈ Ω1 (Uα ) ⊗ g.
γ

−1 we deduce g −1 dg −1
Since gγα = gαγ γα γα = −(dgαγ )gαγ , so that
X
−1
Bα = − uγ (dgαγ )gαγ .
γ

Then, on the overlap Uαβ , we have the equality

X X
−1 −1 −1 −1
Bβ − gβα Bα gβα =− uγ (dgβγ )gβγ + uγ gβα (dgαγ )gαγ gβα
γ γ

The cocycle condition implies that

dgβγ = (dgβα )gαγ + gβα (dgαγ ),

so that
−1
(dgβγ )gβγ = (dgβγ )gγβ = (dgβα )gαβ + gβα (dgαγ )gγβ .
Hence X X
−1
− uγ (dgβγ )gβγ = −(dgβα )gαβ − uγ gβα (dgαγ )gγβ .
γ γ
8.1. CHERN–WEIL THEORY 335

−1 g −1 = g , so that
Using the cocycle condition again, we deduce gαγ βα γβ
X X
−1 −1
uγ gβα (dgαγ )gαγ gβα = uγ gβα (dgαγ )gγβ .
γ γ

Hence, on the overlap Uαβ we have

−1 −1 −1
Bβ = −(dgβα )gβα + gβα Bα gβα = gαβ dgαβ + gαβ Bα gαβ .

This shows that the collection (Bα ) defines a connection on P .

(b) We need to check that the forms Fα satisfy the gluing rules

Fβ = g −1 Fα g,
−1
where g = gαβ = gβα . We have

1
Fβ = dAβ + [Aβ , Aβ ]
2
1
= d(g −1 dg + g−1 Aα g) + [g−1 dg + g −1 Aα g, g−1 dg + g−1 Aα g].
2
−1
Set ̟ := g dg. Using (8.1.2) we get
1
Fβ = d̟ + [̟, ̟]
2
1
+d(g−1 Aα g) + [̟, g−1 Aα g] + [g−1 Aα g, g −1 Aα g]. (8.1.5)
2
We will check two things.
A. The Maurer-Cartan structural equations.
1
d̟ + [̟, ̟] = 0.
2
B.
d(g −1 Aα g) + [̟, g −1 Aα g] = g −1 (dAα )g.
Proof of A. Let us first introduce a new operation. Let gl(n, K) denote the associative
algebra of K-valued n × n matrices. There exists a natural operation

∧ : Ωk (Uα ) ⊗ gl(n, K) × Ωℓ (Uα ) ⊗ gl(n, K) → Ωk+ℓ (Uα ) ⊗ gl(n, K),

uniquely defined by

(ω k ⊗ A) ∧ (η ℓ ⊗ B) = (ω k ∧ η ℓ ) ⊗ (A · B), (8.1.6)

where ω k ∈ Ωk (Uα ), η ℓ ∈ Ωℓ (Uα ) and A, B ∈ gl(n, K) (see also Example 3.3.12). The
space gl(n, K) is naturally a Lie algebra with respect to the commutator of two matrices.
This structure induces a bracket

[•, •] : Ωk (Uα ) ⊗ gl(n, K) × Ωℓ (Uα ) ⊗ gl(n, K) → Ωk+ℓ (Uα ) ⊗ gl(n, K)

336 CHAPTER 8. CHARACTERISTIC CLASSES

defined as in (8.1.1). A very simple computation yields the following identity.

1
ω ∧ η = [ω, η] ∀ω, η ∈ Ω1 (Uα ) ⊗ gl(n, K). (8.1.7)
2
The Lie group lies inside GL(n, K), so that its Lie algebra g lies inside gl(n, K). We can
think of the map gαβ as a matrix valued map, so that we have

d̟ =d(g −1 dg) = (dg −1 ) ∧ dg

= − (g −1 · dg · g−1 )dg = −(g −1 dg) ∧ (g−1 dg)
(8.1.7) 1
= − ̟ ∧ ̟ = − [̟, ̟].
2
Proof of B. We compute

d(g −1 Aα g) = (dg −1 Aα ) · g + g −1 (dAα )g + g−1 Aα dg

= −g−1 · dg · g−1 ∧ Aα · g + g −1 (dAα )g + (g−1 Aα g) ∧ g −1 dg
= −̟ ∧ g−1 Aα g + g −1 Aα g ∧ ̟ + g−1 (dAα )g
(8.1.7 ) 1 1
= − [̟, g−1 Aα g] + [̟, g1 Aα g] + g −1 (dAα )g
2 2
(8.1.2 )
= −[̟, g−1 Aα g] + g−1 (dAα )g.

Part (b) of the proposition now follows from A, B and (8.1.5).

(c) First, we let the reader check the following identity

d[ω, η] = [dω, η] + (−1)|ω| [ω, dη], (8.1.8)

where ω, η ∈ Ω∗ (Uα ) ⊗ g. Using the above equality we get

1 (8.1.2)
d(Fα ) = {[dAα , Aα ] − [Aα , dAα ]} = [dAα , Aα ]
2
1 (8.1.3)
=[Fα , Aα ] − [[Aα , Aα ], Aα ] = [Fα , Aα ].
2
The proposition is proved. ⊔
⊓

Exercise 8.1.5. Let ωα ∈ Ωk (Uα ) ⊗ g satisfy the gluing rules

−1
ωβ = gβα ωα gβα on Uαβ .

In other words, the collection ωα defines a global k-form ω ∈ Ωk (Ad(P )). Prove that the
collection
dωα + [Aα , ωα ],
defines a global Ad(P )-valued (k + 1)-form on M which we denote by dA ω. Thus, the
Bianchi identity can be rewritten as dA F (A) = 0, for any A ∈ A(P ). ⊔
⊓
8.1. CHERN–WEIL THEORY 337

Remark 8.1.6. Suppose P → M is a principal G-bundle given by the open cover U = (Uα )
and gluing cocycle gβα : Uαβ → G. A gauge transformation of P is by definition, a
collection of smooth maps
Tα : Uα → G
satisfying the gluing rules

Tβ (x) = gβα (x)Tα (x)gβα (x)−1 , ∀x ∈ U αβ.

Observe that a gauge transformation is a section of the G-fiber bundle

G ֒→ C(P ) ։ M,

(see Definition 2.3.15) with standard fiber G, symmetry group G, where the symmetry
group G acts on itself by conjugation
C
G × G ∋ (g, h) 7−→ Cg (h) := ghg 1 ∈ G.

The set of gauge transformations forms a group with respect to the operation

(Sα ) · (Tα ) := (Sα Tα ).

We denote by G(U, g•• ) this group. One can verify that if P is described the (open cover,
gluing cocycle)-pair (V, h•• ), then the groups G(U, g•• ) and G(V), h•• ) are isomorphic. We
denote by G(P ) the isomorphism class of all these groups.
The group G(U, g•• ) acts on A(U, g•• ) according to


G(P ) × A(P ) ∋ (T, A) 7→ T AT −1 := −(dTα )Tα−1 + Tα Aα Tα−1 ∈ A(P ).


The group G(P ) also acts on the vector spaces Ω• Ad(P ) and, for any A ∈ A(P ),
T ∈ G(P ) we have
FT AT −1 = T FA T −1 .
We say that two connections A0 , A1 ∈ A(P ) are gauge equivalent if there exists T ∈ G(P )
such that A1 = T A0 T −1 . ⊔
⊓

8.1.2 G-vector bundles

Definition 8.1.7. Let G be a Lie group, and E → M a vector bundle with standard fiber
a vector space V . A G-structure on E is defined by the following collection of data.

(a) A representation ρ : G → GL(V ).

(b) A principal G-bundle P over M such that E is associated to P via ρ. In other words,
there exists an open cover (Uα ) of M , and a gluing cocycle gαβ : Uαβ → G, such
that the vector bundle E can be defined by the cocycle

ρ(gαβ ) : Uαβ → GL(V ).

338 CHAPTER 8. CHARACTERISTIC CLASSES

We denote a G-structure by the pair (P, ρ).

Two G-structures (Pi , ρi ) on E, i = 1, 2, are said to be isomorphic, if the representa-
tions ρi are isomorphic, and the principal G-bundles Pi are isomorphic. ⊔
⊓

Example 8.1.8. Let E → M be a rank r real vector bundle over a smooth manifold
M . A metric on E allows us to talk about orthonormal moving frames. They are easily
produced from arbitrary ones via the Gramm-Schimdt orthonormalization technique. In
particular, two different orthonormal local trivializations are related by a transition map
valued in the orthogonal group O(r), so that a metric on a bundle allows one to replace an
arbitrary collection of gluing data by an equivalent (cohomologous) one with transitions in
O(r). In other words, a metric on a bundle induces an O(r) structure. The representation
ρ is in this case the natural injection O(r) ֒→ GL(r, R).
Conversely, an O(r) structure on a rank r real vector bundle is tantamount to choosing
a metric on that bundle.
Similarly, a Hermitian metric on a rank k complex vector bundle defines an U (k)-
structure on that bundle. ⊔
⊓

Let E = (P, ρ, V ) be a G-vector bundle. Assume P is defined by an open cover (Uα ),

and gluing cocycle
gαβ : Uαβ → G.
If the collection {Aα ∈ Ω1 (Uα )⊗g} defines a connection on the principal bundle P , then the
collection ρ∗ (Aα ) defines a connection on the vector bundle E. Above, ρ∗ : g → End(V )
denotes the derivative of ρ at 1 ∈ G. A connection of E obtained in this manner is
said to be compatible with the G-structure. Note that if F (Aα ) is the curvature of the
connection on P , then the collection ρ∗ (F (Aα )) coincides with the curvature F (ρ∗ (Aα ))
of the connection ρ∗ (Aα ).
For example, a connection compatible with some metric on a vector bundle is com-
patible with the orthogonal/unitary structure of that bundle. The curvature of such a
connection is skew-symmetric which shows the infinitesimal holonomy is an infinitesimal
orthogonal/unitary transformation of a given fiber.

8.1.3 Invariant polynomials

Let V be a vector space over K = R, C. Consider the symmetric power
S k (V ∗ ) ⊂ (V ∗ )⊗k ,
which consists of symmetric, multilinear maps
ϕ : V × · · · × V → K.
Note that any ϕ ∈ S k (V ∗ ) is completely determined by
Pϕ (v) = ϕ(v, . . . , v).
This follows immediately from the polarization formula
1 ∂k
ϕ(v1 , . . . , vk ) = Pϕ (t1 v1 + · · · + tk vk ). (8.1.9)
k! ∂t1 · · · ∂tk
8.1. CHERN–WEIL THEORY 339

If dim V = n then, fixing a basis of V , we can identify S k (V ∗ ) with the space of degree k
homogeneous polynomials in n variables.
Assume now that A is a K-algebra with 1. Starting with ϕ ∈ S k (V ∗ ) we can produce
a K-multilinear map

ϕ = ϕA : (A ⊗ V ) × · · · × (A ⊗ V ) → A,

uniquely determined by

ϕ(a1 ⊗ v1 , . . . , ak ⊗ vk ) = ϕ(v1 , . . . , vk )a1 a2 · · · ak ∈ A.

If moreover the algebra A is commutative, then ϕA is uniquely determined by the polyno-

mial
Pϕ (x) = ϕA (x, . . . , x) x ∈ A ⊗ V.
Remark 8.1.9. Let us emphasize that when A is not commutative, then the above func-
tion is not symmetric in its variables. For example, if a1 a2 = −a2 a1 , then

P (a1 X1 , a2 X2 , · · · ) = −P (a2 X2 , a1 X1 , · · · ).

For applications to geometry, A will be the algebra Ω• (M ) of complex valued differential

forms on a smooth manifold M . When restricted to the commutative subalgebra
M
Ωeven (M ) = Ω2k (M ) ⊗ C,
k≥0

we do get a symmetric function. ⊔

⊓

Example 8.1.10. Let V = gl(n, C). For each matrix T ∈ V we denote by ck (T ) the
coefficient of λk in the characteristic polynomial
  X
λ √
cλ (T ) := det 1 − T = ck (T )λk , (i = −1).
2πi
k≥0

Then, ck (T ) is a degree k homogeneous polynomial in the entries of T . For example,

 
1 1 n
c1 (T ) = − tr T, cn (T ) = − det T.
2πi 2πi

Via polarization, ck (T ) defines an element of S k (gl(n, C)∗ ).

If A is a commutative C-algebra with 1, then A ⊗ gl(n, C) can be identified with the
space gl(n, A) of n × n matrices with entries in A. For each T ∈ gl(n, A) we have
 
λ
det 1 − T ∈ A[λ],
2πi

and ck (T ) continues to be the coefficient of λk in the above polynomial. ⊔

⊓
340 CHAPTER 8. CHARACTERISTIC CLASSES

Consider now a matrix Lie group G. The adjoint action of G on its Lie algebra g
induces an action on S k (g∗ ) still denoted by Ad. We denote by I k (G) the Ad-invariant
elements of S k (g∗ ). It consists of those ϕ ∈ S k (g∗ ) such that

ϕ(gX1 g−1 , . . . , gXk g−1 ) = ϕ(X1 , . . . , Xk ),

for all X1 , . . . , Xk ∈ g. Set

M Y
I • (G) := I k (G) and I •• (G) := I k (G).
k≥0 k≥0

The elements of I • (G) are usually called invariant polynomials. The space I •• (G) can
be identified (as vector space) with the space of Ad-invariant formal power series with
variables from g∗ .
Example 8.1.11. Let G = GL(n, C), so that g = gl(n, C). The map

gl(n, C) ∋ X 7→ tr exp(X),

defines an element of I •• (GL(n, C)). To see this, we use the “Taylor expansion”
X 1
exp(X) = Xk,
k!
k≥0

which yields
X 1
tr exp(X) = tr X k .
k!
k≥0

For each k, tr X k ∈ I k (GL(n, C)) since

tr (gXg −1 )k = tr gX k g−1 = tr X k . ⊔
⊓

Proposition 8.1.12. Let ϕ ∈ I k (G). Then for any X, X1 , . . . , Xk ∈ g we have

ϕ([X, X1 ], X2 , . . . , Xk ) + · · · + ϕ(X1 , X2 , . . . , [X, Xk ]) = 0. (8.1.10)

Proof. The proposition follows immediately from the equality

d
|t=0 ϕ(etX X1 e−tX , . . . , etX Xk e−tX ) = 0. ⊔
⊓
dt

Let us point out a useful identity. If P ∈ I k (g), U is an open subset of Rn , and

Fi = ωi ⊗ Xi ∈ Ωdi (U ) ⊗ g, A = ω ⊗ X ∈ Ωd (U ) ⊗ g

then
P (F1 , . . . , Fi−1 , [A, Fi ], Fi+1 . . . , Fk )
= (−1)d(d1 +···di−1 ) ωω1 · · · ωk P (X1 , · · · , [X, Xi ], · · · Xk ).
8.1. CHERN–WEIL THEORY 341

In particular, if F1 , . . . , Fk−1 have even degree, we deduce that for every i = 1, . . . , k we

have

P (F1 , . . . , Fi−1 , [A, Fi ], Fi+1 , . . . , Fk ) = ωω1 · · · ωk P (X1 , . . . , [X, Xi ], . . . , Xk ).

Summing over i, and using the Ad-invariance of the polynomial P , we deduce

k
X
P (F1 , . . . , Fi−1 , [A, Fi ], Fi+1 , . . . , Fk ) = 0, (8.1.11)
i=1

∀F1 , . . . , Fk−1 ∈ Ωeven (U ) ⊗ g, Fk , A ∈ Ω• (U ) ⊗ g.

8.1.4 The Chern–Weil Theory

Let G be a matrix Lie group with Lie algebra g, and P → M be a principal G-bundle over
the smooth manifold M .
Assume P is defined by an open cover (Uα ), and a gluing cocycle

gαβ : Uαβ → G.

Pick A ∈ A(P ) defined by the collection Aα ∈ Ω1 (Uα ) ⊗ g. Its curvature is then defined
by the collection
1
Fα = dAα + [Aα , Aα ].
2
Given φ ∈ I k (G), we can define as in the previous section (with A = Ωeven (Uα ), V = g)

Pφ (Fα ) := φ(Fα , . . . , Fα ) ∈ Ω2k (Uα ).

−1
Because φ is Ad-invariant, and Fβ = gβα Fα gβα , we deduce

Pφ (Fα ) = Pφ (Fβ ) on Uαβ ,

so that the locally defined forms Pφ (Fα ) patch-up to a global 2k-form on M which we
denote by φ(FA ).
Theorem 8.1.13 (Chern–Weil). (a) The form φ(FA ) is closed, ∀A ∈ A(P ).
(b) If A0 , A1 ∈ A(P ), then the forms φ(FA0 ), and φ(FA1 ) are cohomologous. In other
words, the closed form φ(FA ) defines a cohomology class in H 2k (M ) which is independent
of the connection A ∈ A(P ).

Proof. We use the Bianchi identity dFα = −[Aα , Fα ]. The Leibniz’ rule yields

dφ(Fα , . . . , Fα ) = φ(dFα , Fα , . . . , Fα ) + · · · + φ(Fα , . . . , Fα , dFα )

(8.1.10 )
= −φ([Aα , Fα ], Fα , . . . , Fα ) − · · · − φ(Fα , . . . , Fα , [Aα , Fα ]) = 0.
(b) Let Ai ∈ A(P ) (i = 0, 1) be defined by the collections

Aiα ∈ Ω1 (Uα ) ⊗ g.
342 CHAPTER 8. CHARACTERISTIC CLASSES

Set Cα := A1α − A0α . For 0 ≤ t ≤ 1 we define Atα ∈ Ω1 (Uα ) ⊗ g by Atα := A0α + tCα .
The collection (Atα ) defines a connection At ∈ A(P ), and t 7→ At ∈ A(P ) is an (affine)
path connecting A0 to A1 . Note that

C = (Cα ) = Ȧt .

We denote by F t = (Fαt ) the curvature of At . A simple computation yields

t2
Fαt = Fα0 + t(dCα + [A0α , Cα ]) + [Cα , Cα ]. (8.1.12)
2
Hence,
Ḟαt = dCα + [A0α , Cα ] + t[Cα , Cα ] = dCα + [Atα , Cα ].
Consequently,
Z 1n o
φ(Fα1 ) − φ(Fα0 ) = φ(Ḟαt , Fαt , . . . , Fαt ) + · · · + φ(Fαt , . . . , Fαt , Ḟαt ) dt
0
Z 1
= φ(dCα , Fαt , . . . , Fαt ) + · · · + φ(Fαt , . . . , Fαt , dCα ) dt
0
Z 1
+ φ([Atα , Cα ], Fαt , . . . , Fαt ) + · · · + φ(Fαt , . . . , Fαt , [Atα , Cα ]) dt.
0

Because the algebra Ωeven (Uα ) is commutative, we deduce

φ(ωσ(1) , . . . , ωσ(k) ) = φ(ω1 , . . . , ωk ),

for all σ ∈ Sk and any ω1 , . . . , ωk ∈ Ωeven (Uα ) ⊗ g. Hence

Z 1
φ(Fα1 ) − φ(Fα0 ) =k φ(Fαt , . . . , Fαt , dCα + [Atα Cα ])dt.
0

We claim that

φ( Fαt , . . . , Fαt , dCα + [Atα , Cα ] ) = dφ( Fαt , . . . , Fαt , Cα ).

Using the Bianchi identity we get

dφ( Fαt , . . . , Fαt , Cα )

= φ(Fαt , · · · , Fαt , dCα ) + φ(dFαt , · · · , Fαt , Cα ) + · · · + φ(Fαt , · · · , dFαt , Cα )
= φ( Fαt , . . . , Fαt , dCα )
−φ( Cα , [Atα , Fαt ], Fαt , . . . , Fαt ) − · · · − φ( Cα , Fαt , . . . , Fαt , [Atα , Fαt ] )
= φ( Fαt , . . . , Fαt , dCα + [Atα , Cα ] ) − φ( Fαt , . . . , Fαt , [Atα , Cα ] )
−φ( [Atα , Fαt ], Fαt , . . . , Fαt , Cα ) − · · · − φ( Fαt , . . . , Fαt , [Atα , Fαt ], Cα )
(8.1.11 )
= φ( Fαt , . . . , Fαt , dCα + [Atα , Cα ] ) = φ( dCα + [Atα , Cα ], Fαt , . . . , Fαt ).
8.1. CHERN–WEIL THEORY 343

Hence Z 1
φ(Fα1 ) − φ(Fα0 ) =d kφ(Ȧtα , Fαt , . . . , Fαt )dt. (8.1.13)
0
We set Z 1
Tφ (A1α , A0α ) := kφ(Ȧtα , Fαt , . . . , Fαt )dt.
0
−1 −1
Since Cβ = gβα Cα gβα , and Fβ = gβα Fα gβα
on Uαβ , we conclude from the Ad-invariance of
1 0
φ that the collection Tφ (Aα , Aα ) defines a global (2k − 1)-form on M which we denote by
Tφ (A1 , A0 ), and we name it the φ-transgression from A0 to A1 . We have thus established
the transgression formula

φ(FA1 ) − φ(FA0 ) ) = d Tφ (A1 , A0 ). (8.1.14)

The Chern–Weil theorem is proved. ⊔

⊓

Remark 8.1.14. Observe that for every Ad-invariant polynomial φ ∈ I k (g), any principal
G-bundle P , any connection A ∈ A(P ), and any gauge transformation T ∈ G(P ) we have



φ FA = φ FT AT −1 .

We say that the Chern–Weil construction is gauge invariant. ⊔

⊓

Example 8.1.15. Consider a matrix Lie group G with Lie algebra g, and denote by P0
the trivial principal G-bundle over G, P0 = G × G. Denote by ̟ the tautological 1-form
̟ = g−1 dg ∈ Ω1 (G) ⊗ g. Note that for every left invariant vector field X ∈ g we have

̟(X) = X.

Denote by d the trivial connection on P0 . Clearly d is a flat connection. Moreover, the

Maurer-Cartan equation implies that1 d + ̟ is also a flat connection. Thus, for any
φ ∈ I k (G)
φ(Fd ) = φ(F (d+̟ ) = 0.
The transgression formula implies that the form
Z 1
τφ = Tφ (d + ̟, d) = k φ(̟, Fd+t̟ , . . . , Fd+t̟ )dt ∈ Ω2k−1 (G)
0

is closed.
A simple computation using the Maurer-Cartan equations shows that
Z 1 
k
τφ = k−1 (t2 − 1)k−1 dt · φ(̟, [̟, ̟], · · · , [̟, ̟])
2 0
k 22k−1 k!(k − 1)!
= (−1)k−1 φ(̟, [̟, ̟], · · · , [̟, ̟])
2k−1 (2k)!
2k
= (−1)k−1 2k
· φ(̟, [̟, ̟], · · · , [̟, ̟]).
k
1
The connections d and d + ̟ are in fact gauge equivalent.
344 CHAPTER 8. CHARACTERISTIC CLASSES

We thus have a natural map τ : I • (G) → H odd (G) called transgression. The elements
in the range of τ are called transgressive. When G is compact and connected, then a
nontrivial result due to the combined efforts of H. Hopf, C. Chevalley, H. Cartan, A.
Weil and L. Koszul states that the cohomology of G is generated as an R-algebra by the
transgressive elements. We refer to [26] for a beautiful survey of this subject. ⊔
⊓

Exercise 8.1.16. Let G = SU (2). The Killing form κ is a degree 2 Ad-invariant polyno-
mial on su(2). Describe τκ ∈ Ω3 (G) and then compute
Z
τκ .
G

Compare this result with the similar computations in Subsection 7.4.3. ⊔

⊓

Let us now analyze the essentials of the Chern–Weil construction.

Input: (a) A principal G-bundle P over a smooth manifold M , defined by an open cover
(Uα ) ,and gluing cocycle gαβ : Uαβ → G.
(b) A connection A ∈ A(P ) defined by the collection

Aα ∈ Ω1 (Uα ) ⊗ g,

satisfying the transition rules

−1 −1
Aβ = gαβ dgαβ + gαβ Aα gαβ on Uαβ .

(c) φ ∈ I k (G).
Output: A closed form φ(F (A)) ∈ Ω2k (M ), whose cohomology class is independent of
the connection A. We denote this cohomology class by φ(P ).
Thus, the principal bundle P defines a map, called the Chern–Weil correspondence

cw P : I • (G) → H • (M ) φ 7→ φ(P ).

One can check easily the map cw P is a morphism of R-algebras.

Definition 8.1.17. Let M and N be two smooth manifolds, and F : M → N be a
smooth map. If P is a principal G-bundle over N defined by an open cover (Uα ), and
gluing cocycle gαβ : Uαβ → G, then the pullback of P by F is the principal bundle F ∗ (P )
over M defined by the open cover F −1 (Uα ), and gluing cocycle
F gαβ
F −1 (Uαβ ) → Uαβ → G. ⊔
⊓

The pullback of a connection on P is defined similarly. The following result should be

obvious.
Proposition 8.1.18. (a) If P is a trivial G-bundle over the smooth manifold M , then
φ(P ) = 0 ∈ H • (M ), for any φ ∈ I • (G).
F
(b) Let M → N be a smooth map between the smooth manifolds M and N . Then, for
every principal G-bundle over N , and any φ ∈ I • (G) we have

φ(F ∗ (P )) = F ∗ (φ(P )).

8.2. IMPORTANT EXAMPLES 345

Equivalently, this means the diagram below is commutative.

''w H (N )
cwP
I • (G) •

cwF • (P )
)' u F • ⊔
⊓
H • (M )

Denote by PG the collection of smooth principal G-bundles (over smooth manifolds).

For each P ∈ PG we denote by BP the base of P . Finally, we denote by F a contravariant
functor from the category of smooth manifolds (and smooth maps) to the category of
Abelian groups.

Definition 8.1.19. An F-valued G-characteristic class is a correspondence

PG ∋ P 7→ c(P ) ∈ F(BP ),

such that the following hold.

(a) c(P ) = 0 if P is trivial.

(b) F(F )(c(P )) = c(F ∗ (P )), for any smooth map F : M → N , and any principal G-
bundle P → N . ⊔
⊓

Hence, the Chern–Weil construction is just a method of producing G-characteristic

classes valued in the DeRham cohomology.

Remark 8.1.20. (a) We see that each characteristic class provides a way of measuring
the nontriviality of a principal G-bundle.
(b) A very legitimate question arises. Do there exist characteristic classes (in the DeRham
cohomology) not obtainable via the Chern–Weil construction?
The answer is negative, but the proof requires an elaborate topological technology
which is beyond the reach of this course. The interested reader can find the details in the
monograph [98] which is the ultimate reference on the subject of characteristic classes.
(b) There exist characteristic classes valued in contravariant functors other then the
DeRham cohomology. E.g., for each Abelian group A, the Čech cohomology with coeffi-
cients in the constant sheaf A defines a contravariant functor H • (−, A), and using topo-
logical techniques, one can produce H • (−, A)-valued characteristic classes. For details we
refer to [98], or the classical [124]. ⊔
⊓

8.2 Important examples

We devote this section to the description of some of the most important examples of
characteristic classes. In the process we will describe the invariants of some commonly
encountered Lie groups.
346 CHAPTER 8. CHARACTERISTIC CLASSES

8.2.1 The invariants of the torus T n

The n-dimensional torus T n = U (1) × · · · × U (1) is an Abelian Lie group, so that the
adjoint action on its Lie algebra tn is trivial. Hence

I • (T n ) = S • ((tn )∗ ).

In practice one uses a more explicit description obtained as follows. Pick angular coordi-
nates 0 ≤ θ i ≤ 2π, 1 ≤ i ≤ n, and set

1
xj := − dθ j .
2πi
The x′j s form a basis of (tn )∗ , and now we can identify

I • (T n ) ∼
= R[x1 , . . . , xn ].

8.2.2 Chern classes

Let E be a rank r complex vector bundle over the smooth manifold M . We have seen
that a Hermitian metric on E induces an U (r)-structure (P, ρ), where ρ is the tautological
representation
ρ : U (r) ֒→ GL(r, C).
Exercise 8.2.1. Prove that different Hermitian metrics on E define isomorphic U (r)-
structures. ⊔
⊓

Thus, we can identify such a bundle with the tautological principal U (r)-bundle of uni-
tary frames. A connection on this U (r)-bundle is then equivalent with a linear connection
∇ on E compatible with a Hermitian metric h•, •i, i.e.,

∇X hλu, vi = λ{h∇X u, vi + hu, ∇X vi},

∀λ ∈ C, u, v ∈ C ∞ (E), X ∈ Vect (M ).
The characteristic classes of E are by definition the characteristic classes of the tauto-
logical principal U (r)-bundle. To describe these characteristic classes we need to elucidate
the structure of the ring of invariants I • (U (r)).
The ring I • (U (r)) consists of symmetric, r-linear maps

φ : u(r) × · · · × u(r) → R,

invariant with respect to the adjoint action

u(r) ∋ X 7→ T XT −1 ∈ u(r), T ∈ U (r).

It is convenient to identify such a map with its polynomial form

Pφ (X) = φ(X, . . . , X).

8.2. IMPORTANT EXAMPLES 347

The Lie algebra u(r) consists of r × r complex skew-hermitian matrices. Classical results
of linear algebra show that, for any X ∈ u(r), there exists T ∈ U (r), such that T XT −1 is
diagonal
T XT −1 = i diag(λ1 , . . . , λr ).
The set of diagonal matrices in u(r) is called the Cartan algebra of u(r), and we will denote
it by Cu(r) . It is a (maximal) Abelian Lie subalgebra of u(r). Consider the stabilizer

SU (r) := T ∈ U (r) ; T XT −1 = X, ∀X ∈ Cu(r) , ,

and the normalizer


NU (r) := T ∈ U (r); T Cu(r) T −1 ⊂ Cu(r) , .

The stabilizer SU (r) is a normal subgroup of NU (r) , so we can form the quotient

WU (r) := NU (r) /SU (r)

called the Weyl group of U (r). As in Subsection 7.4.4, we see that the Weyl group is
isomorphic with the symmetric group Sr because two diagonal skew-Hermitian matrices are
unitarily equivalent if and only if they have the same eigenvalues, including multiplicities.
We see that Pφ is Ad-invariant if and only if its restriction to the Cartan algebra is
invariant under the action of the Weyl group.
The Cartan algebra is the Lie algebra of the (maximal) torus T n consisting of diagonal
unitary matrices. As in the previous subsection we introduce the variables
1
xj := − dθj .
2πi
The restriction of Pφ to Cu(r) is a polynomial in the variables x1 , . . . , xr . The Weyl
group Sr permutes these variables, so that Pφ is Ad-invariant if and only if Pφ (x1 , . . . , xr )
is a symmetric polynomial in its variables. According to the fundamental theorem of
symmetric polynomials, the ring of these polynomials is generated (as an R-algebra) by
the elementary ones P
c1 = x
P j j
c2 = i<j xi xj
.. .. ..
. . .
cr = x1 · · · xr
Thus
I • (U (r)) = R[c1 , c2 , . . . , cr ].
In terms of matrices X ∈ u(r) we have
X  
t
ck (X)t = det 1 −
k
X ∈ I • (U (r))[t].
2πi
k

The above polynomial is known as the universal rank r Chern polynomial, and its coeffi-
cients are called the universal, rank r Chern classes.
348 CHAPTER 8. CHARACTERISTIC CLASSES

Returning to our rank r vector bundle E, we obtain the Chern classes of E

ck (E) := ck (F (∇)) ∈ H 2k (M ),

and the Chern polynomial of E

 
t
ct (E) := det 1 − F (∇) ∈ H • (M )[t].
2πi

Above, ∇ denotes a connection compatible with a Hermitian metric h•, •i on E, while

F (∇) denotes its curvature.
Remark 8.2.2. The Chern classes produced via the Chern–Weil method capture only
a part of what topologists usually refer to characteristic classes of complex bundles. To
give the reader a feeling of what the Chern–Weil construction is unable to capture we will
sketch a different definition of the first Chern class of a complex line bundle. The following
facts are essentially due to Kodaira and Spencer [81]; see also [55] for a nice presentation.
Let L → M be a smooth complex Hermitian line bundle over the smooth manifold M .
Upon choosing a good open cover (Uα ) of M we can describe L by a collection of smooth
maps zαβ : Uαβ → U (1) ∼ = S 1 satisfying the cocycle condition

zαβ zβγ zγα = 1 ∀α, β, γ. (8.2.1)

If we denote by C ∞ (·, S 1 ) the sheaf of multiplicative groups of smooth S 1 -valued functions,

we see that the family of complex line bundles on M can be identified with the Čech
group H 1 (M, C ∞ (·, S 1 )). This group is called the smooth Picard group of M . The group
multiplication is precisely the tensor product of two line bundles. We will denote it by
Pic∞ (M ).
If we write zαβ = exp(2πiθαβ ) (θβα = −θαβ ∈ C ∞ (Uαβ , R)) we deduce from (8.2.1)
that ∀Uαβγ 6= ∅
θαβ + θβγ + θγα = nαβγ ∈ Z.
It is not difficult to see that, ∀Uαβγδ 6= ∅, we have

nβγδ − nαγδ + nαβδ − nαβγ = 0.

In other words, the collection nαβγ defines a Čech 2-cocycle of the constant sheaf Z.
On a more formal level, we can capture the above cocycle starting from the exact
sequence of sheaves
exp(2πi·)
0 → Z ֒→ C ∞ (·, R) −→ C ∞ (·, S 1 ) → 0.

The middle sheaf is a fine sheaf so its cohomology vanishes in positive dimensions. The
long exact sequence in cohomology then gives
δ
0 → Pic∞ (M ) → H 2 (M, Z) → 0.

The cocycle ( nαβγ ) represents precisely the class δ(L).

8.2. IMPORTANT EXAMPLES 349

The class δ(L), L ∈ Pic∞ (M ) is called the topological first Chern class and is denoted
by ctop
1 (L). This terminology is motivated by the following result of Kodaira and Spencer,
[81]:
The image of ctop
1 (L) in the DeRham cohomology via the natural morphism

H ∗ (M, Z) → H ∗ (M, R) ∼ ∗
= HDR (M )

coincides with the first Chern class obtained via the Chern–Weil procedure.
The Chern–Weil construction misses precisely the torsion elements in H 2 (M, Z). For
example, if a line bundle admits a flat connection then its first Chern class is trivial. This
may not be the case with the topological one, because line bundle may not be topologically
trivial. ⊔
⊓

8.2.3 Pontryagin classes

Let E be a rank r real vector bundle over the smooth manifold M . An Euclidean metric
on E induces an O(r) structure (P, ρ). The representation ρ is the tautological one

ρ : O(r) ֒→ GL(r, R).

Exercise 8.2.3. Prove that two metrics on E induce isomorphic O(r)-structures. ⊔

⊓

Hence, exactly as in the complex case, we can naturally identify the rank r-real vector
bundles equipped with metric with principal O(r)-bundles. A connection on the principal
bundle can be viewed as a metric compatible connection in the associated vector bundle.
To describe the various characteristic classes we need to understand the ring of invariants
I • (O(r)).
As usual, we will identify the elements of I k (O(r)) with the degree k, Ad-invariant
polynomials on the Lie algebra o(r) consisting of skew-symmetric r × r real matrices. Fix
P ∈ I k (O(r)). Set m = [r/2], and denote by J the 2 × 2 matrix
 
0 −1
J := .
1 0

Consider the Cartan algebra


{λ1 J ⊕ · · · ⊕ λm J ∈ o(r) ; λj ∈ R} , r = 2m
Co(r) =
{λ1 J ⊕ · · · ⊕ λm J ⊕ 0 ∈ o(r) ; λj ∈ R} , r = 2m + 1

The Cartan algebra Co(r) is the Lie algebra of the (maximal) torus

m Rθ1 ⊕ · · · ⊕ Rθm ∈ O(r) , r = 2m
T = ,
Rθ1 ⊕ · · · ⊕ Rθm ⊕ 1R ∈ O(r) , r = 2m + 1

where for each θ ∈ [0, 2π] we denoted by Rθ the 2 × 2 rotation

 
cos θ − sin θ
Rθ := .
sin θ cos θ
350 CHAPTER 8. CHARACTERISTIC CLASSES

As in Subsection 8.2.1 we introduce the variables

1
xj = − dθj .
2π
Using standard results concerning the normal Jordan form of a skew-symmetric matrix,
we deduce that, for every X ∈ o(r), there exists T ∈ O(r) such that T XT −1 ∈ Co(r) .
Consequently, any Ad-invariant polynomial on o(r) is uniquely defined by its restriction
to the Cartan algebra.
Following the approach in the complex case, we consider

SO(r) = T ∈ O(r); T XT −1 = X, ∀X ∈ Co(r) , ,

NO(r) = T ∈ O(r); T Co(r) T −1 ⊂ Co(r) , .
The stabilizer SO(r) is a normal subgroup in N(O(r)), so we can form the Weyl group
WO(r) := NO(r) /SO(r) .
Exercise 8.2.4. Prove that WO(r) is the subgroup of GL(m, R) generated by the involu-
tions
σij : (x1 , . . . , xi , . . . , xj , . . . , xm ) 7→ (x1 , . . . , xj , . . . , xi , . . . , xm )
εj : (x1 , . . . , xj , . . . , xm ) 7→ (x1 , . . . , −xj , . . . , xm ). ⊔
⊓
The restriction of P ∈ I k (O(r)) to Co(r) is a degree k homogeneous polynomial in the
variables x1 , . . . , xm invariant under the action of the Weyl group. Using the above exercise
we deduce that P must be a symmetric polynomial P = P (x1 , . . . , xm ), separately even
in each variable. Invoking once again the fundamental theorem of symmetric polynomials
we conclude that P must be a polynomial in the elementary symmetric ones
P 2
p1 = x
P j 2j 2
p2 = i<j i xj
x
.. .. .. .
. . .
pm = x21 · · · x2m
Hence
I • (O(r)) = R[p1 , . . . , p⌊r/2⌋ ].
In terms of X ∈ o(r) we have
X  
t
pt (X) = pj (X)t = det 1 −
2j
X ∈ I • (O(r))[t].
2π
j

The above polynomial is called the universal rank r Pontryagin polynomial, while its
coefficients pj (X) are called the universal rank r Pontryagin classes.
The Pontryagin classes p1 (E), . . . , pm (E) of our real vector bundle E are then defined
by the equality
X  
t
pt (E) = pj (E)t = det 1 −
2j
F (∇) ∈ H • (M )[t],
2π
j

where ∇ denotes a connection compatible with some (real) metric on E, while F (∇)
denotes its curvature. Note that pj (E) ∈ H 4j (M ). The polynomial pt (E) ∈ H • (M )[t] is
called the Pontryagin polynomial of E.
8.2. IMPORTANT EXAMPLES 351

8.2.4 The Euler class

Let E be a rank r, real oriented vector bundle. A metric on E induces an O(r)-structure,
but the existence of an orientation implies the existence of a finer structure, namely an
SO(r)-symmetry.
The groups O(r) and SO(r) share the same Lie algebra so(r) = o(r). The inclusion
ı : SO(r) ֒→ O(r),
induces a morphism of R-algebras
ı∗ : I • (O(r)) → I • (SO(r)).
Because so(r) = o(r) one deduces immediately that ı∗ is injective.
Lemma 8.2.5. When r is odd then ı∗ : I • (O(r)) → I • (SO(r)) is an isomorphism. ⊔
⊓

Exercise 8.2.6. Prove the above lemma. ⊔

⊓

The situation is different when r is even, r = 2m. To describe the ring of invariants
I • (SO(2m)), we need to study in greater detail the Cartan algebra

Co(2m) = λ1 J ⊕ · · · ⊕ λm J ∈ o(2m)
and the corresponding Weyl group action. The Weyl group WSO(2m) , defined as usual as
the quotient
WSO(2m) = NSO(2m) /SSO(2m) ,
is isomorphic to the subgroup of GL(Co(2m) ) generated by the involutions
σij : (λ1 , . . . , λi , . . . , λj , . . . , λm ) 7→ (λ1 , . . . , λj , . . . , λi , . . . , λm ),
and
ε : (λ1 , . . . , λm ) 7→ (ε1 λ1 , . . . , εm λm ),
where ε1 , . . . , εm = ±1, and ε1 · · · εm = 1. (Check this!)
Set as usual xi := −λi /2π. The Pontryagin O(2m)-invariants
X
pj (x1 , . . . , xm ) = (xi1 · · · xij )2 ,
1≤i1 <···<ij ≤m

continue to be WSO(2m) invariants. There is however a new invariant,

Y
∆(x1 , . . . , xm ) := xj .
j

In terms of
X = λ1 J ⊕ · · · ⊕ λm J ∈ CSO(2m) ,
we can write  
−1 m
∆(X) = P f (X),
2π
where P f denotes the pfaffian of the skewsymmetric matrix X viewed as a linear map
R2m → R2m , when R2m is endowed with the canonical orientation. Note that pm = ∆2 .
352 CHAPTER 8. CHARACTERISTIC CLASSES

Proposition 8.2.7.

I • (SO(2m)) ∼
= R[Z1 , Z2 , . . . , Zm ; Y ]/(Y 2 − Zm ) (Zj = pj , Y = ∆)

where (Y 2 − Zm ) denotes the ideal generated by the polynomial Y 2 − Zm .

Proof. We follow the approach used by H. Weyl in describing the invariants of the alternate
group ([133], Sec. II.2). The isomorphism will be established in two steps.
Step 1. The space I • (SO(2m)) is generated as an R-algebra by the polynomials
p1 , · · · , pm , ∆.
Step 2. The kernel of the morphism
ψ
R[Z1 , . . . , Zm , ; Y ] → I • (SO(2m))

defined by Zj 7→ pj , Y 7→ ∆ is the ideal (Y 2 − Zm ).

Proof of Step 1. Note that WSO(2m) has index 2 as a subgroup in WO(2m) . Thus
WSO(2m) is a normal subgroup, and

G = WO(2m) /WSO(2m) ∼
= Z2 .

The group G = {1, e} acts on I • (SO(2m)) by

(eF )(x1 , x2 , . . . , xm ) = F (−x1 , x2 , . . . , xm ) = · · · = F (x1 , x2 , . . . , −xm ),

and moreover,
I • (O(2m)) = ker(1 − e).
For each F ∈ I • (SO(2m)) we define F + := (1 + e)F , and we observe that F + ∈ ker(1 − e).
Hence
F + = P (p1 , . . . , pm ).
On the other hand, the polynomial F − := (1 −e)F is separately odd in each of its variables.
Indeed,
F − (−x1 , x2 , . . . , xm ) = eF − (x1 , . . . , xm )
= e(1 − e)F (x1 , . . . , xm ) = −(1 − e)F (x1 , . . . , xm ) = −F − (x1 , . . . , xm ).
Hence, F − vanishes when any of its variables vanishes so that F − is divisible by their
product ∆ = x1 · · · xm ,
F − = ∆ · G.
Since eF − = −F − , and e∆ = −∆ we deduce eG = G, i.e., G ∈ I • (O(2m)). Consequently,
G can be written as
G = Q(p1 , . . . , pm ),
so that
F − = ∆ · Q(p1 , . . . , pm ).
Step 1 follows from
1 + 1
F = (F + F − ) = (P (p1 , . . . , pm ) + ∆ · Q(p1 , . . . , pm )).
2 2
8.2. IMPORTANT EXAMPLES 353

Proof of Step 2. From the equality

det X = P f (X)2 ∀X ∈ so(2m),

we deduce
(Y 2 − Zm ) ⊂ ker ψ,

so that we only need to establish the opposite inclusion.

Let P = P (Z1 , Z2 , . . . , Zm ; Y ) ∈ ker ψ. Consider P as a polynomial in Y with coef-
ficients in R[Z1 , . . . , Zm ]. Divide P by the quadratic polynomial (in Y ) Y 2 − Zm . The
remainder is linear
R = A(Z1 , . . . , Zm )Y + B(Z1 , . . . , Zm ).

Since Y 2 − Zm , P ∈ ker ψ, we deduce R ∈ ker ψ. Thus

A(p1 , . . . , pm )∆ + B(p1 , . . . , pm ) = 0.

Applying the morphism e we get

−A(p1 , . . . , pm )∆ + B(p1 , . . . , pm ) = 0.

Hence A ≡ B ≡ 0 so that P is divisible by Y 2 − Zm . ⊔

⊓

Let E be a rank 2m, real, oriented vector bundle over the smooth manifold M . As in
the previous subsection we deduce that we can use a metric to naturally identify E with a
principal SO(2m)-bundle and in fact, this principal bundle is independent of the metric.
Finally, choose a connection ∇ compatible with some metric on E.

Definition 8.2.8. (a) The universal Euler class is defined by

1
e = e(X) = P f (−X) ∈ I m (SO(2m)).
(2π)m

(b) The Euler class of E, is the cohomology class e(E) ∈ H 2m (M ) represented by the
Euler form
1

e(∇) = P f − F (∇) ∈ Ω2m (M ).
(2π)m
(According to the Chern–Weil theorem this cohomology class is independent of the metric
and the connection.) ⊔
⊓

Example 8.2.9. Let (Σ, g) be a compact, oriented, Riemann surface and denote by ∇g
the Levi–Civita connection. The the Euler form
1
ε(g) = s(g)dvg
4π

coincides with the Euler form e(∇g ) obtained via the Chern–Weil construction. ⊔
⊓
354 CHAPTER 8. CHARACTERISTIC CLASSES

Remark 8.2.10. Let E be a rank 2m, real, oriented vector bundle over the smooth,
compact, oriented manifold M . We now have two apparently conflicting notions of Euler
classes.
A topological Euler class etop (E) ∈ H 2m (M ) defined as the pullback of the Thom class
via an arbitrary section of E.
A geometric Euler class egeom (E) ∈ H 2m (M ) defined via the Chern–Weil construction.
The most general version of the Gauss–Bonnet theorem, which will be established later
in this chapter, will show that these two notions coincide! ⊔
⊓

8.2.5 Universal classes

In each of the situations discussed so far we defined characteristic classes for vector bundles
with a given rank. In this subsection we show how one can coherently present these facts
all at once, irrespective of rank. The algebraic machinery which will achieve this end is
called inverse limit. We begin by first describing a special example of inverse limit.
A projective sequence of rings is a a sequence of rings {Rn }n≥0 , together with a sequence
φn
of ring morphisms Rn ←− Rn+1 . The inverse limit of a projective system (Rn , φn ) is the
subring Y
lim Rn ⊂ Rn
←
n≥0

consisting of the sequences (x1 , x2 , . . .) such that φn (xn+1 ) = xn , ∀n ≥ 0.

Example 8.2.11. Let Rn = R[[X1 , . . . , Xn ]] be the ring of formal power series in n
variables with coefficients in the commutative ring with unit R. (R0 = R.) Denote by
φn : Rn+1 → Rn the natural morphism defined by setting Xn+1 = 0. The inverse limit of
this projective system is denoted by R[[X1 , X2 , . . .]].
Given a sequence Fn ∈ R[[X]] (n ≥ 1) such that Fn (0) = 1, we can form the sequence
of products
(1, F1 (X1 ), F1 (X1 )F2 (X2 ), . . . , F1 (X1 ) · · · Fn (Xn ), . . .),
which defines an element in R[[X1 , X2 , . . .]] denoted by F1 (X1 )F2 (X2 ) · · · . When F1 =
F2 = · · · Fn = · · · = F the corresponding elements is denoted by (F )∞ .
Exercise 8.2.12. (a) Let R[[x]]♭ denote the set of formal power series F ∈ R[[x]] such
that F (0) = 1. Prove that (R[[x]]♭ , ·) is an abelian group.
(b) Prove that ∀F, G ∈ R[[x]]♭

(F · G)∞ = (F )∞ · (G)∞ . ⊔
⊓

Similarly, given Gn ∈ R[[x]], (n ≥ 1) such that Fn (0) = 0, we can from the sequence
of sums
(0, G1 (X1 ), G1 (X1 ) + G2 (X2 ), . . . , G1 (X1 ) + · · · ⊕ Gn (Xn ), . . .),
which defines an element in R[[X1 , X2 , . . .]] denoted by G1 (X1 ) + G2 (X2 ) + · · · . When
G1 = G2 = · · · = Fn = · · · = F we denote the corresponding element (G)∞ . ⊔
⊓
8.2. IMPORTANT EXAMPLES 355

In dealing with characteristic classes of vector bundles one naturally encounters the
increasing sequences
U (1) ֒→ U (2) ֒→ · · · (8.2.2)
(in the complex case) and (in the real case)

O(1) ֒→ O(2) ֒→ · · · . (8.2.3)

We will discuss these two situations separately.

The complex case. The sequence in (8.2.2) induces a projective sequence of rings

R ← I •• (U (1)) ← I •• (U (2)) ← · · · . (8.2.4)

We know that I •• (U (n)) = Sn [[xj ]]= the ring of symmetric formal power series in n
variables with coefficients in R. Set

S∞ [[xj ]] := lim Sn .
←

Given a rank n vector bundle E over a smooth manifold M , and φ ∈ I •• (U (n)), the
characteristic class φ(E) is well defined since udim M +1 = 0, for any u ∈ Ω∗ (M ). Thus
we can work with the ring I •• (U (n)) rather than I • (U (n)) as we have done so far. An
element
φ = (φ1 , φ2 , . . .) ∈ I •• (U (∞)) := lim I •• (U (n)) = S∞ [[xj ]]
←

is called universal characteristic class.

If E is a complex vector bundle, we set φ(E)” = φr (E), where r = rank E. More
precisely, to define φr (E) we need to pick a connection ∇ compatible with some Hermitian
metric on E, and then set
φ(E) = φr (F (∇)).
(n)
Example 8.2.13. We denote by ck (n ≥ k) the elementary symmetric polynomial in n
variables X
(n)
ck = x i1 · · · x ik .
1≤i1 <···<ik ≤n

Then the sequence

(k) (k+1)
(0, . . . , 0, ck , ck , . . .)
defines an element in S∞ [[xj ]] denoted by ck , which we call the universal k-th Chern class.
Formally, we can write X
ck = x i1 x i2 · · · x ik .
1≤i1 <···<ik <∞

We can present the above arguments in a more concise form as follows. Consider the
function F (x) = (1 + tx) ∈ R[[x]], where R = R[[t]] is the ring of formal power series in
the variable t. Then (F )∞ defines an element in R[[X1 , X2 , . . .]]. One sees immediately
that in fact, (F )∞ ∈ S∞ [[xj ]] [[t]] and moreover

(F )∞ (x1 , x2 , . . .) = (1 + tx1 )(1 + tx2 ) · · · = 1 + c1 t + c2 t2 + · · · .

356 CHAPTER 8. CHARACTERISTIC CLASSES

We see that Y
ct := (1 + txj )
j≥1

is the universal Chern polynomial. ⊔

⊓

We can perform a similar operation with any F ∈ R[[x]] such that F (0) = 1. We get
a semigroup morphism

(R[[x]]♭ , ·) ∋ F 7→ (F )∞ ∈ (S∞ [[xj ]], ·).

(see Exercise 8.2.12.) One very important example is

X ∞
x 1 1 2 1 Bk 2k
F (x) = −x
= 1 + x + x + · · · = 1 + x + (−1)k−1 x ∈ R[[x]].
1−e 2 12 2 (2k)!
k=1

The coefficients Bk are known as the Bernoulli numbers. The product

   
∞ x1 x2
(F ) = · ··· ,
1 − e−x1 1 − e−x2

defines an element in S∞ [[xj ]] called the universal Todd class. We denote it by Td. Using
the fundamental theorem of symmetric polynomials we can write

Td = 1 + Td1 + Td2 + · · · ,

where Tdn ∈ S∞ [[xj ]] is an universal symmetric, homogeneous “polynomial” of degree n,

hence expressible as a combination of the elementary symmetric “polynomials” c1 , c2 , . . ..
By an universal“polynomial” we understand element in the inverse limit

R[x1 , x2 , . . .] = lim R[x1 , x2 , . . .].

←

An universal “polynomial” P is said to be homogeneous of degree d, if it can be represented

as a sequence
P = (P1 , P2 , . . .),
where Pm is a homogeneous polynomial of degree d in m variables, and

Pm+1 (x1 , x2 , . . . , xm , 0) = Pm (x1 , . . . xm ).

For example,
1 1 1
Td1 = c1 , Td2 = (c21 + c2 ), Td3 = c1 c2 etc.
2 12 24
Analogously, any function G ∈ R[[x]] such that G(0) = 0 defines an element

(G)∞ = G(x1 ) + G(x2 ) + · · · ∈ S∞ [[xj ]].

We have two examples in mind. First, consider G(x) = xk . We get the symmetric function

sk = (xk )∞ = (G)∞ = xk1 + xk2 + · · · ,

8.2. IMPORTANT EXAMPLES 357

called the universal k-th power sum. We will denote these by sk .

Next, consider
X xm
G(x) = ex − 1 = .
m!
m≥1

We have X 1 X 1
(ex − 1)∞ = (xm )∞ = sm ∈ S∞ [[xj ]]
m! m=1
m!
m≥1

Given a complex vector bundle E we define

ch(E) = rank E + (ex − 1)∞ (E).

The cohomology class ch(E) is called the Chern character of the bundle E. If ∇ is a
connection on E compatible with some Hermitian metric, then we can express the Chern
character of E as
  ∞
X 1
ch(E) = tr eF (∇) = rank (E) + tr (F (∇)∧k ),
k!
k=1

where ∧ is the bilinear map

Ωi (End (E)) × Ωj (End (E)) → Ωi+j (End (E))

defined in Example 3.3.12.

Proposition 8.2.14. Consider two complex vector bundles E1 , E2 over the same manifold
M . Then
ch(E1 ⊕ E2 ) = ch(E1 ) + ch(E2 ),
and
ch(E1 ⊗ E2 ) = ch(E1 ) ∧ ch(E2 ) ∈ H • (M ).

Proof. Consider a connection ∇i on Ei compatible with some Hermitian metric hi , i = 1, 2.

Then ∇1 ⊕∇2 is a connection on E1 ⊕E2 compatible with the metric h1 ⊕h2 , and moreover

F (∇1 ⊕ ∇2 ) = F (∇1 ) ⊕ F (∇2 ).

Hence
exp(F (∇1 ⊕ ∇2 )) = exp(F (∇1 )) ⊕ exp(F (∇2 )),
from which we deduce the first equality.
As for the second equality, consider the connection ∇ on E1 ⊗ E2 uniquely defined by
the product rule

∇(s1 ⊗ s2 ) = (∇1 s1 ) ⊗ s2 + s1 ⊗ (∇2 s2 ), si ∈ C ∞ (Ei ),

where the operation

⊗ : Ωk (E1 ) × Ωℓ (E2 ) → Ωk+ℓ (E1 ⊗ E2 )
358 CHAPTER 8. CHARACTERISTIC CLASSES

is defined by
(ω k ⊗ s1 ) ⊗ (η ℓ s2 ) = (ω k ∧ η ℓ ) ⊗ (s1 ⊗ s2 ) (8.2.5)
∀si ∈ C ∞ (Ei ), ∀ω k ∈ Ωk (M ), and ∀η ℓ ∈ Ωℓ (M ).
We compute the curvature of ∇ using the equality F (∇) = (d∇ )2 . If si ∈ C ∞ (Ei ),
then
F (∇)(s1 ⊗ s2 ) = d∇ {(∇1 s1 ) ⊗ s2 + s1 ⊗ (∇2 s2 )}

= (F (∇1 )s1 ) ⊗ s2 − (∇1 s1 ) ⊗ (∇2 s2 ) + (∇1 s1 ) ⊗ (∇2 s2 ) + s1 ⊗ (F (∇2 ))

= F (∇1 ) ⊗ 1E2 + 1E1 ⊗ F (∇2 ).

The second equality in the proposition is a consequence the following technical lemma.

Lemma 8.2.15. Let A (respectively B) be a skew-adjoint, n × n (respectively m × m)

complex matrix. Then

tr (exp(A ⊗ 1Cm + 1Cn ⊗ B)) = tr (exp(A)) · tr (exp(B)).

Proof. Pick an orthonormal basis (ei ) of Cn and an orthonormal basis (fj ) of Cm ) such
that, with respect to these bases A = diag(λ1 , . . . , λn ) and B = diag(µ1 , . . . , µm )). Then,
with respect to the basis (ei ⊗ fj ) of Cn ⊗ Cm , we have

(A ⊗ 1Cm + 1Cn ⊗ B) = diag( λi + µj ).

Hence
exp(A ⊗ 1Cm + 1Cn ⊗ B) = diag(eλi eµj ),
so that, X
tr (A ⊗ 1Cm + 1Cn ⊗ B) = eλi eµj = tr (eA ) · tr (eB ). ⊔
⊓

The proposition is proved. ⊔

⊓

Exercise 8.2.16. (Newton’s formulæ.) Consider the symmetric polynomials

X
ck = xi1 · · · xik ∈ R[x1 , . . . , xn ],
1≤i1 ≤···ik ≤n

and (r ≥ 0)
X
sr = xrj ∈ R[x1 , . . . , xn ].
j

Set
n
Y
f (t) := (1 − xj t).
j=1

(a) Show that

f ′ (t) X
=− sr tr−1 .
f (t) r
8.2. IMPORTANT EXAMPLES 359

(b) Prove the Newton formulæ

r
X
(−1)j sr−j cj = 0 ∀1 ≤ r ≤ n.
j=1

(c) Deduce from the above formulae the following identities between universal symmetric
polynomials.
s1 = c1 , s2 = c21 − 2c2 , s3 = c31 − 3c1 c2 + 3c3 . ⊔
⊓
The real case. The sequence (8.2.3) induces a projective system
I •• (O(1)) ← I •• (O(2)) ← · · · .
We have proved that I •• (O(n)) = S⌊n/2⌋ [[x2j ]] = the ring of even, symmetric power series
in ⌊n/2⌋ variables. The inverse limit of this system is
I • • (O(∞) = lim I •• (O(r)) = S∞ [[x2j ]] := lim Sm [[x2j ]].
← ←
As in the complex case, any element of this ring is called a universal characteristic class. In
fact, for any real vector bundle E and any φ ∈ S∞ [[x2j ]] there is a well defined characteristic
class φ(E) which can be expressed exactly as in the complex case, using metric compatible
connections. If F ∈ R[[x]]♭ then (F (x2 ))∞ defines an element of S∞ [[x2j ]].
In topology, the most commonly encountered situations are the following.
A. √
x/2 X 22k−1 − 1
F (x) = √ =1+ (−1)k 2k−1 Bk xk .
sinh( x/2) 2 (2k)!
k≥1

The universal characteristic class (F (x))∞

is denoted by A,b and it is called the A-genus.
b
We can write
Ab =1+A
b1 + Ab2 + ··· ,
where Ak are universal, symmetric, even, homogeneous “polynomials”, and as such they
can be described using universal Pontryagin classes
X
pm = x2j1 · · · x2jm .
1≤j1 <···jm

The first couple of terms are

b 1 = − p1 , A
A b2 = 1
(−4p2 + 7p21 ) etc.
24 2 · 32 · 5
7

B. Consider
√ X
x 1 1 22k
F (x) = √ = 1 + x + x2 + · · · = 1 + (−1)k−1 Bk xk .
tanh x 3 45 (2k)!
k≥1

The universal class (F (x2 ))∞ is denoted by L, and it is called the L-genus. As before, we
can write
L = 1 + L1 + L2 + · · · ,
where the Lj ’s are universal, symmetric, even, homogeneous “polynomials”. They can be
expressed in terms of the universal Pontryagin classes. The first few terms are
1 1
L1 = p1 , L2 = (7p2 − p21 ) etc.
3 45
360 CHAPTER 8. CHARACTERISTIC CLASSES

8.3 Computing characteristic classes

The theory of characteristic classes is as useful as one’s ability to compute them. In this
section we will describe some methods of doing this.
Most concrete applications require the aplication of a combination of techniques from
topology, differential and algebraic geometry and Lie group theory that go beyond the
scope of this book. We will discuss in some detail a few invariant theoretic methods and
we will present one topological result more precisely the Gauss–Bonnet–Chern theorem.

8.3.1 Reductions
In applications, the symmetries of a vector bundle are implicitly described through topo-
logical or geometric properties.
For example, if a rank r complex vector bundle E splits as a Whitney sum E = E1 ⊕E2
with rank Ei = ri , then E, which has a natural U (r)- symmetry, can be given a finer
structure of U (r1 ) × U (r2 ) vector bundle.
More generally, assume that a given rank r complex vector bundle E admits a G-
structure (P, ρ), where G is a Lie group, P is a principal G bundle, and ρ : G → U (r) is a
representation of G. Then we can perform two types of Chern–Weil constructions: using
the U (r) structure, and using the G structure. In particular, we obtain two collections
of characteristic classes associated to E. One natural question is whether there is any
relationship between them.
In terms of the Whitney splitting E = E1 ⊕ E2 above, the problem takes a more
concrete form: compute the Chern classes of E in terms of the Chern classes of E1 and
E2 . Our next definition formalizes the above situations.
Definition 8.3.1. Let ϕ : H → G be a smooth morphism of (matrix) Lie groups.
(a) If P is a principal H-bundle over the smooth manifold M defined by the open cover
(Uα ), and gluing cocycle
hαβ : Uαβ → H,
then the principal G-bundle defined by the gluing cocycle
gαβ = ϕ ◦ hαβ : Uαβ → G
is said to be the ϕ-associate of P , and it is denoted by ϕ(P ).
(b) A principal G-bundle Q over M is said to be ϕ-reducible, if there exists a principal
H-bundle P → M such that Q = ϕ(P ). ⊔
⊓

The morphism ϕ : H → G in the above definition induces a morphism of R-algebras

ϕ∗ : I • (G) → I • (H).
The elements of ker ϕ∗ ⊂ I • (G) are called universal identities.
The following result is immediate.
Proposition 8.3.2. Let P be a principal G-bundle which can be reduced to a principal
H-bundle Q. Then for every η ∈ ker ϕ∗ we have
η(P ) = 0 in H ∗ (M ).
8.3. COMPUTING CHARACTERISTIC CLASSES 361

Proof. Denote by LG (respectively LH ) the Lie algebra of G (respectively H), and by ϕ∗

the differential of ϕ at 1 ∈ H,
ϕ∗ : LH → LG .
Pick a connection (Aα ) on Q, and denote by (Fα ) its curvature. Then the collection
ϕ∗ (Aα ) defines a connection on P with curvature ϕ∗ (Fα ). Now

η(ϕ∗ (Fα )) = (ϕ∗ η)(Fα ) = 0. ⊔

⊓

The above result should be seen as a guiding principle in proving identities between
characteristic classes, rather than a rigid result. What is important about this result is
the simple argument used to prove it.
We conclude this subsection with some simple, but very important applications of the
above principle.

Example 8.3.3. Let E and F be two complex vector bundles over the same smooth
manifold M of ranks r and respectively s. Then the Chern polynomials of E, F and
E ⊕ F are related by the identity

ct (E ⊕ F ) = ct (E) · ct (F ), (8.3.1)

where the “·” denotes the ∧-multiplication in H even (M ). Equivalently, this means
X
ck (E ⊕ F ) = ci (E) · cj (F ).
i+j=k

To check this, pick a Hermitian metric g on E and a Hermitian metric h on F . g ⊕ h is

a Hermitian metric on E ⊕ F . Hence, E ⊕ F has an U (r + s) structure reducible to an
U (r) × U (s) structure.
The Lie algebra of U (r) × U (s) is the direct sum u(r) ⊕ u(s). Any element X in this
algebra has a block decomposition XE
 
Xr 0
X= = Xr ⊕ Xs ,
0 Xs

where Xr (respectively Xs ) is an r ×r (respectively s×s) complex, skew-hermitian matrix.

(ν)
Let ı denote the natural inclusion u(r) ⊕ u(s) ֒→ u(r + s) and denote by ct ∈ I ∗ (U (ν))[t]
the Chern polynomial.
We have  
t
∗ (r+s)
ı (ct )(Xr ⊕ Xs ) = det 1r+s − Xr ⊕ Xs
2πi
   
t t
= det 1r − Xr · det 1s −
(r) (s)
Xs = ct · ct (s).
2πi 2πi
The equality (8.3.1) now follows using the argument in the proof of Proposition 8.3.2. ⊔
⊓
362 CHAPTER 8. CHARACTERISTIC CLASSES

Remark 8.3.4. Consider the Grassmannian Grk (Cn ) of complex k-dimensional subspaces
in Cn . The universal complex vector bundle Uk,n → Grk (Cn ) is a subbundle of the trivial
bundle Cn → Grk (Cn ). The trivial bundle Cn is equipped with a canonical Hermitian
metric. We denote by Qk,n the orthogonal complement of Uk,n in Cn so that we have an
isomorphism
Cn ∼= Uk,n ⊕ Qk,n .
From the above example we deduce that



ct Uk,n ct Qk,n = ct (Cn ) = 1.

Denote by uj the j-th Chern class of Uk,n and by vℓ the ℓ-th Chern class of Qk,n . Then
k

X
n−k
X
ct Uk,n = uj tj , ct Qk,n = vℓ t ℓ .
j=0 ℓ=0


One can then prove that the cohomology ring H • Grk (Cn ), R) is generated by the classes
ui , vj , which are subject to the single relation above. More formally

. X X 
H • Grk (Cn ), R) ∼ = R[u ,
i jv ; 0 ≤ i ≤ k, 0 ≤ j ≤ n − k] ui )( v j ) = 1 ,
i j

deg ui = 2i, deg vj = 2j. The proof requires a more sophisticated topological machinery.
For details we refer to [99], Chapter v3, Section 6. ⊔
⊓

Exercise 8.3.5. Let E and F be two complex vector bundles over the same manifold M .
Show that
Td(E ⊕ F ) = Td(E) · Td(F ). ⊔
⊓
Exercise 8.3.6. Let E and F be two real vector bundles over the same manifold M .
Prove that
pt (E ⊕ F ) = pt (E) · pt (F ), (8.3.2)
where pt denotes the Pontryagin polynomials. ⊔
⊓

Exercise 8.3.7. Let E and F be two real vector bundles over the same smooth manifold
M . Show that
L(E ⊕ F ) = L(E) · L(F )
b
A(E b
⊕ F ) = A(E) b ).
· A(F ⊔
⊓
Example 8.3.8. The natural inclusion Rn ֒→ Cn induces an embedding ı : O(n) ֒→ U (n).
(An orthogonal map T : Rn → Rn extends by complexification to an unitary map TC :
Cn → Cn ). This is mirrored at the Lie algebra level by an inclusion

o(n) ֒→ u(n).

We obtain a morphism ı∗ : I • (U (n)) → I • (O(n)), and we claim that

i∗ (c2k+1 ) = 0, and ı∗ (c2k ) = (−1)k pk .

8.3. COMPUTING CHARACTERISTIC CLASSES 363

Indeed, for X ∈ o(n) we have

 2k+1 X
∗ 1
ı (c2k+1 )(X) = − λi1 (X) · · · λi2k+1 (X),
2πi
1≤i1 <···<i2k+1 ≤n

where λj (X) are the eigenvalues of X over C. Since X is, in effect, a real skew-symmetric
matrix we have
λj (X) = λj (X) = −λj (X).
Consequently,
 2k+1 X
∗ ∗ 1
ı (c2k+1 )(X) = ı (c2k+1 (X)) = − λi1 (X) · · · λi2k+1 (X)
2πi
1≤i1 <···<i2k+1 ≤n

 2k+1 X
2k+1 1
= (−1) − λi1 (X) · · · λi2k+1 (X) = −ı∗ (c2k+1 )(X).
2πi
1≤i1 <···<i2k+1 ≤n

The equality ı∗ (c2k ) = (−1)k pk is proved similarly. ⊔

⊓

From the above example we deduce immediately the following consequence.

Proposition 8.3.9. If E → M is a real vector bundle and E ⊗ C is its complexification

then
pk (E) = (−1)k c2k (E ⊗ C), k = 1, 2, . . . . (8.3.3)
In a more concentrated form, this means

pt (E) = p−t (E) = cit (E ⊗ C). ⊔

⊓

Exercise 8.3.10. Let E → M be a complex vector bundle of rank r.

(a) Show that ck (E ∗ ) = (−1)k ck (E), i. e.,

ct (E ∗ ) = c−t (E).

(b) One can also regard E as a real, oriented vector bundle ER . Prove that
X
pit (ER ) = (−1)k t2k pk (ER ) = ct (E) · c−t (E),
k

and
cr (E) = e(ER ). ⊔
⊓

Exercise 8.3.11. (a) The natural morphisms I •• (U (r)) → I •• (O(r)) described above
induce a morphism
Φ∞ : I •• (U (∞)) → I •• (O(∞)).
364 CHAPTER 8. CHARACTERISTIC CLASSES

As we already know, for any F ∈ R[[x]]♭ , the product (F )∞ is an element of I •• (U (∞)),

and if moreover F is even, then (F )∞ can be regarded as an element of I ∗∗ (O(∞)). Show
that for every F ∈ R[[x]]♭

Φ∞ ((F )∞ ) = (F · F − )∞ ∈ I •• (O(∞)),

where F − (x) = F (−x).

(b) Let E be a real vector bundle. Deduce from part (a) that

b
Td(E ⊗ C) = A(E)2
. ⊔
⊓

Exercise 8.3.12. For every square matrix X we set2

∞
X    
1/2 r r(r − 1) · · · (r − k + 1)
(1 + X) 1/2
= k
(−1) k
X , := ,
k k k!
k=0

and det1/2 (1 + X) := det(1 + X)1/2 .

Suppose that E → M is a rank r real vector bundle, equipped with a metric and a
compatible connection ∇. The connection ∇ induces a Hermitian connection ∇c on the
complexification Ec := E ⊗R C. Prove that
! !
i c i c)
1/2 2π F (∇ ) b 1/2 4π F (∇
L(∇) = det i
, A(∇) = det i
. ⊔
⊓
tanh 2π F (∇c ) sinh( 4π F (∇c )

Example 8.3.13. Consider the inclusion

ı : SO(2k) × SO(2ℓ)) ֒→ SO(2k + 2ℓ).

This induces a ring morphism

ı∗ : I • (SO(2k + 2ℓ) → I • (SO(2k) × SO(2ℓ)).

Note that
I • (SO(2k) × SO(2ℓ)) ∼
= I • (SO(2k)) ⊗R I • (SO(2ℓ).
Denote by e(ν) the Euler class in I • (SO(2ν)). We want to prove that

ı∗ (e(k+ℓ) ) = e(k) ⊗ e(ℓ) .

Let X = Xk ⊕ Xℓ ∈ so(2k) ⊕ so(2ℓ). Modulo a conjugation by (S, T ) ∈ SO(2k) × SO(2ℓ)

we may assume that

Xk = λ1 J ⊕ · · · ⊕ λk J and Xℓ = µ1 J ⊕ · · · ⊕ µℓ J,
2
We are not worried about convergence issues because the matrices for which we intend to apply the
formula have nilpotent entries.
8.3. COMPUTING CHARACTERISTIC CLASSES 365

where J denotes the 2 × 2 matrix

 
0 −1
J= .
1 0

We have
 k+ℓ
1
ı∗ (e(k+ℓ) )(X) = − λ1 · · · λk · µ1 · · · µℓ = e(k) (Xk ) · e(ℓ) (Xℓ )
2π

= e(k) ⊗ e(ℓ) (Xk ⊕ Xℓ ). ⊔

⊓

The above example has an interesting consequence.

Proposition 8.3.14. Let E and F be two real, oriented vector bundle of even ranks over
the same manifold M . Then

e(E ⊕ F ) = e(E) · e(F ),

where · denotes the ∧-multiplication in H even (M ). ⊔

⊓

Example 8.3.15. Let E be a rank 2k real, oriented vector bundle over the smooth
manifold M . We claim that if E admits a nowhere vanishing section ξ then e(E) = 0.
To see this, fix an Euclidean metric on E so that E is now endowed with an SO(2k)-
structure. Denote by L the real line subbundle of E generated by the section ξ. Clearly,
L is a trivial line bundle, and E splits as an orthogonal sum

E = L ⊕ L⊥ .

The orientation on E, and the orientation on L defined by ξ induce an orientation on L⊥ ,

so that L⊥ has an SO(2k − 1)-structure.
In other words, the SO(2k) structure of E can be reduced to an SO(1) × SO(2k − 1) ∼
=
SO(2k − 1)-structure. Denote by ı∗ the inclusion induced morphism

I ∗ (SO(2k)) → I ∗ (SO(2k − 1)).

Since ı∗ (e(k) ) = 0, we deduce from the Proposition 8.3.2 that e(E) = 0. ⊔

⊓

The result proved in the above example can be reformulated more suggestively as
follows.

Corollary 8.3.16. Let E be a real oriented vector bundle of even rank over the smooth
manifold M . If e(E) 6= 0, then any section of E must vanish somewhere on M ! ⊔
⊓
366 CHAPTER 8. CHARACTERISTIC CLASSES

8.3.2 The Gauss–Bonnet–Chern theorem

If E → M is a real oriented vector bundle over a smooth, compact, oriented manifold M
then there are two apparently conflicting notions of Euler class naturally associated to E.
• The topological Euler class
etop (E) = ζ0∗ τE ,
where τE is the Thom class of E and ζ0 : M → E is the zero section.
• The geometric Euler class
 1
egeom (E) = (2π)r P f (−F (∇)) if rank (E) is even ,
0 if rank (E) is odd

where ∇ is a connection on E compatible with some metric and 2r = rank (E). The next
result, which generalizes the Gauss–Bonnet theorem, will show that these two notions of
Euler class coincide.
π
Theorem 8.3.17 (Gauss–Bonnet–Chern). Let E → M be a real, oriented vector bundle
over the compact oriented manifold M . Then

etop (E) = egeom (E).

Proof. We will distinguish two cases.

A. rank (E) is odd. Consider the automorphism of E

i : E → E u 7→ −u ∀u ∈ E.

Since the fibers of E are odd dimensional, we deduce that i reverses the orientation in the
fibers. In particular, this implies

π∗ i∗ τE = −π∗ τE = π∗ (−τE ),

where π∗ denotes the integration along fibers. Since π∗ is an isomorphism (Thom isomor-
phism theorem), we deduce
i∗ τE = −τE .
Hence
etop (E) = −ζ0∗ i∗ τE . (8.3.4)
On the other hand, notice that
ζ0∗ i∗ = ζ0 .
Indeed,
ζ0∗ i∗ = (iζ0 )∗ = (−ζ0 )∗ = (ζ)∗ (ζ0 = −ζ0 ).
The equality etop = egeom now follows from (8.3.4).
B. rank (E) = 2k. We will use a variation of the original argument of Chern, [29]. Let
∇ denote a connection on E compatible with a metric g. The strategy of proof is very
simple. We will explicitly construct a closed form ω ∈ Ω2k
cpt (E) such that
8.3. COMPUTING CHARACTERISTIC CLASSES 367

(i) π∗ ω = 1 ∈ Ω0 (M ).

(ii) ζ0∗ ω = e(∇) = (2π)−k P f (−F (∇)).

The Thom isomorphism theorem coupled with (i) implies that ω represents the Thom
2k (E). The condition (ii) simply states the sought for equality e
class in Hcpt top = egeom .
Denote by S(E) the unit sphere bundle of E,

S(E) = {u ∈ E ; |u|g = 1}.

Then S(E) is a compact manifold, and

dim S(E) = dim M + 2k − 1.

Denote by π0 the natural projection S(E) → M , and by π0∗ (E) → S(E) the pullback
of E to S(E) via the map π0 . The vector bundle π0∗ (E) has an SO(2k)-structure, and
moreover, it admits a tautological, nowhere vanishing section

Υ : S(Ex ) ∋ e 7→ e ∈ Ex ≡ (π0∗ (E)x )e (x ∈ M ).

Thus, according to Example 8.3.15 we must have

egeom (π0∗ (E)) = 0 ∈ H 2k (S(E)),

where egeom (π0∗ E) denotes the differential form

1
egeom (π0∗ ∇) = P f (−F (π0∗ ∇)).
(2π)k

Hence there must exist ψ ∈ Ω2k−1 (S(E)) such that

dψ = egeom (π0∗ E).

The decisive step in the proof of Gauss–Bonnet–Chern theorem is contained in the follow-
ing lemma.
Lemma 8.3.18. There exists Ψ = Ψ(∇) ∈ Ω2k−1 (S(E)) such that

dΨ(∇) = egeom (π0∗ (E)) (8.3.5)

and Z
Ψ(∇) = −1 ∈ Ω0 (M ). (8.3.6)
S(E)/M

The form Ψ(∇) is sometimes referred to as the global angular form of the pair (E, ∇).
For the clarity of the exposition we will conclude the proof of the Gauss–Bonnet–Chern
theorem assuming Lemma 8.3.18 which will be proved later on.
Denote by r : E → R+ the norm function

E ∋ e 7→ |e|g
368 CHAPTER 8. CHARACTERISTIC CLASSES

If we set E 0 = E \ { zero section }, then we can identify

1
E0 ∼
= (0, ∞) × S(E) e 7→ (|e|, e).
|e|
Consider the smooth cutoff function

ρ = ρ(r) : [0, ∞) → R,

such that ρ(r) = −1 for r ∈ [0, 1/4], and ρ(r) = 0 for r ≥ 3/4. Finally, define

ω = ω(∇) = −ρ′ (r)dr ∧ Ψ(∇) − ρ(r)π ∗ (e(∇)).

The differential form ω is well defined since ρ′ (r) ≡ 0 near the zero section. Obviously ω
has compact support on E, and satisfies the condition (ii) since

ζ0∗ ω = −ρ(0)ζ0∗ π ∗ e(∇) = e(∇).

From the equality Z

ρ(r)π ∗ e(∇) = 0,
E/M

we deduce
Z Z Z ∞ Z
′ ′
ω=− ρ (r)dr ∧ Ψ(∇) = − ρ (r)dr · Ψ(∇)
E/M E/M 0 S(E)/M
Z
(8.3.6)
= −(ρ(1) − ρ(0)) Ψ(∇) = 1.
S(E)/M

To complete the program outlined at the beginning of the proof we need to show that ω
is closed.
dω = ρ′ (r)dr ∧ dΨ(∇) − ρ′ (r) ∧ π ∗ e(∇)
(8.3.5 )
= ρ′ (r)dr ∧ {π0∗ e(∇) − π ∗ e(∇)}.
The above form is identically zero since π0∗ e(∇) = π ∗ e(∇) on the support of ρ′ . Thus ω
is closed and the theorem is proved. ⊔
⊓

Proof of Lemma 8.3.18 We denote by ∇ the pullback of ∇ to π0∗ E. The tautological

section Υ : S(E) → π0∗ E can be used to produce an orthogonal splitting

π0∗ E = L ⊕ L⊥ ,

where L is the real line bundle spanned by Υ, while L⊥ is its orthogonal complement in
π0∗ E with respect to the pullback metric g. Denote by

P : π0∗ E → π0∗ E

the orthogonal projection onto L⊥ . Using P , we can produce a new metric compatible
ˆ on π ∗ E by
connection ∇ 0

ˆ = (trivial connection on L) ⊕ P ∇P.

∇
8.3. COMPUTING CHARACTERISTIC CLASSES 369

We have an equality of differential forms

1
π0∗ e(∇) = e(∇) = P f (−F (∇)).
(2π)k
ˆ splits as a direct sum
Since the curvature of ∇
ˆ = 0 ⊕ F ′ (∇),
F (∇) ˆ

ˆ denotes the curvature of ∇

where F ′ (∇) ˆ |L⊥ , we deduce

ˆ = 0.
P f (F (∇))

We denote by ∇t the connection ∇ ˆ + t(∇ − ∇),

ˆ so that ∇0 = ∇, ˆ and ∇1 = ∇. If F t is
t
the curvature of ∇ , we deduce from the transgression formula (8.1.13) that
(  Z )
k 1
ˆ =d −1 ˆ F t , . . . , F t )dt .
π0∗ e(∇) = e(∇) − e(∇) k P f (∇ − ∇,
2π 0

We claim that the form

 k Z 1
−1 ˆ F t , . . . , F t )dt
Ψ(∇) = k P f (∇ − ∇,
2π 0

satisfies all the conditions in Lemma 8.3.18.

By construction,
dΨ(∇) = π0∗ e(∇),
so all that we need to prove is
Z
Ψ(∇) = −1 ∈ Ω0 (M ).
S(E)/M

It suffices to show that for each fiber Ex of E we have

Z
Ψ(∇) = −1.
Ex

Along this fiber π0∗ E is naturally isomorphic with a trivial bundle

π0∗ E |Ex ∼
= (Ex × Ex → Ex ).

Moreover, the connection ∇ restricts as the trivial connection. By choosing an orthonormal

basis of Ex we can identify π0∗ E |Ex with the trivial bundle R2k over R2k . The unit sphere
S(Ex ) is identified with the unit sphere S 2k−1 ⊂ R2k . The splitting L ⊕ L⊥ over S(E)
restricts over S(Ex ) as the splitting

R2k = ν ⊕ T S 2k−1 ,
ˆ is then the direct
where ν denotes the normal bundle of S 2k−1 ֒→ R2k . The connection ∇
sum between the trivial connection on ν, and the Levi–Civita connection on T S 2k−1 .
370 CHAPTER 8. CHARACTERISTIC CLASSES

Fix a point p ∈ S 2k−1 , and denote by (x1 , . . . , x2k−1 ) a collection of normal coordinates
near p, such that the basis (∂xi |p ) is positively oriented. Set ∂i := ∂xi for i = 1, . . . , 2k − 1.
Denote the unit outer normal vector field by ∂0 . For α = 0, 1, . . . , 2k − 1 set f α = ∂α |p .
The vectors f α form a positively oriented orthonormal basis of R2k .
We will use Latin letters to denote indices running from 1 to 2k − 1, and Greek letters
indices running from 0 to 2k − 1.

∇i ∂α = (∇i ∂α )ν + (∇i ∂α )τ ,

where the superscript ν indicates the normal component, while the superscript τ indicates
the tangential component. Since at p
ˆ i ∂j = (∇i ∂j )τ ,
0=∇

we deduce
∇i ∂j = (∇i ∂j )ν at p.
Hence
∇i ∂j = (∇i ∂j , ∂0 )∂0 = −(∂j , ∇i ∂0 )∂0 .
Recalling that ∇ is the trivial connection in R2k , we deduce
 
∂
∇i ∂0 |p = ∂0 |p = f i = ∂i |p .
∂fi
Consequently,
∇i ∂j = −δji ∂0 , at p.
If we denote by θ i the local frame of T ∗ S 2k−1 dual to (∂i ), then we can rephrase the above
equality as
∇∂j = −(θ 1 + · · · + θ 2k−1 ) ⊗ ∂0 .
On the other hand, ∇i ∂0 = ∂i , i.e.,

∇∂0 = θ 1 ⊗ ∂1 + · · · + θ 2k−1 ⊗ ∂2k−1 .

Since (x1 , · · · , x2k−1 ) are normal coordinates with respect to the Levi–Civita connection
ˆ we deduce that ∇∂
∇, ˆ α = 0, ∀α so that
 
0 −θ 1 · · · −θ 2k−1
 θ1 0 ··· 0 
 
A = (∇ − ∇)ˆ |p = 
 θ
2 0 ··· 0 
.
 .. .. .. .. 
 . . . . 
θ 2k−1 0 ··· 0

Denote by F 0 the curvature of ∇0 = ∇ˆ at p. Then F 0 = 0 ⊕ R, where R denotes the

Riemann curvature of ∇ˆ at p.
The computations in Example 4.2.22 show that the second fundamental form of the
embedding
S 2k−1 ֒→ R2k ,
8.3. COMPUTING CHARACTERISTIC CLASSES 371

coincides with the induced Riemann metric (which is the first fundamental form). Using
Teorema Egregium we get

hR(∂i , ∂j )∂k , ∂ℓ i = δiℓ δjk − δik δjℓ .

In matrix format we have

F 0 = 0 ⊕ (Ωij ),
ˆ + tA can be computed using the
where Ωij = θ i ∧ θ j . The curvature F t at p of ∇t = ∇
equation (8.1.12) of subsection 8.1.4, and we get

F t = F 0 + t2 A ∧ A = 0 ⊕ (1 − t2 )F 0 .

We can now proceed to evaluate Ψ(∇).

 k Z 1
−1
Ψ(∇) |p = k P f (A, (1 − t2 )F 0 , · · · , (1 − t2 )F 0 )dt
2π 0
 k Z 1 
−1
= k (1 − t ) dt P f (A, F 0 , F 0 , . . . , F 0 ).
2 k−1
(8.3.7)
2π 0
We need to evaluate the pfaffian in the right-hand-side of the above formula. Set for
simplicity F := F 0 .
Using the polarization formula and Exercise 2.2.65 in Subsection 2.2.4, we get

(−1)k X
P f (A, F, F, · · · , F ) = ǫ(σ)Aσ0 σ1 Fσ2 σ3 · · · Fσ2k−2 σ2k−1 .
2k k!
σ∈S2k

For i = 0, 1, define
Si = {σ ∈ S2k ; σi = 0}.
We deduce
X
2k k! P f (A, F, · · · , F ) = (−1)k ǫ(σ)(−θ σ1 ) ∧ θ σ2 ∧ θ σ3 ∧ · · · ∧ θ σ2k−2 ∧ θ σ2k−1
σ∈S0
X
+(−1)k ǫ(σ)θ σ0 ∧ θ σ2 ∧ θ σ3 ∧ · · · ∧ θ σ2k−2 ∧ θ σ2k−1 .
σ∈S1

For each σ ∈ S0 we get a permutation

φ : (σ1 , σ2 , · · · , σ2k−1 ) ∈ S2k−1 ,

such that ǫ(σ) = ǫ(φ). Similarly, for σ ∈ S1 , we get a permutation

φ = (σ0 , σ2 , · · · , σ2k−1 ) ∈ S2k−1 ,

such that ǫ(σ) = ǫ(φ). Hence

X
2k k! P f (A, F, · · · , F ) = 2(−1)k+1 ǫ(φ)θ φ1 ∧ · · · · · · θ φ2k−1
φ∈S2k−1
372 CHAPTER 8. CHARACTERISTIC CLASSES

= 2(−1)k+1 (2k − 1)!θ 1 ∧ · · · ∧ θ 2k−1 = 2(−1)k+1 dVS 2k−1 ,

where dVS 2k−1 denotes the Riemannian volume form on the unit sphere S 2k−1 .Using the
last equality in (8.3.7) we get
 k Z 1 
−1
Ψ(∇) |p = k (1 − t ) dt · 2(−1)k+1 (2k − 1)!dVS 2k−1
2 k−1
2π 0
Z 1 
(2k)! 2 k−1
=− (1 − t ) dt dVS 2k−1 .
(4π)k k! 0
Using the Exercise 4.1.59 we get that
Z
2π k
ω 2k−1 = dVS 2k−1 = ,
S 2k−1 (k − 1)!
and consequently
Z Z 1 
(2k)! 2 k−1
Ψ(∇) = −ω 2k−1 (1 − t ) dt
S 2k−1 (4π)k k! 0
Z 1 
(2k)! 2 k−1
= − 2k−1 (1 − t ) dt .
2 k!(k − 1)! 0

The above integral can be evaluated inductively using the substitution t = cos ϕ. We have
Z 1  Z π/2
Ik = (1 − t2 )k−1 dt = (cos ϕ)2k−1 dϕ
0 0
Z π/2
π/2
= (cos ϕ)2k−2 sin ϕ|0 + (2k − 2) (cos ϕ)2k−3 (sin ϕ)2 dϕ
0
= (2k − 1)Ik−1 − (2k − 2)Ik
so that
2k − 3
Ik = Ik−1 .
2k − 2
One now sees immediately that
Z
Ψ(∇) = −1.
S 2k−1

Lemma 8.3.18 is proved. ⊔

⊓

Corollary 8.3.19 (Chern). Let (M, g) be a compact, oriented Riemann Manifold of

dimension 2n. If R denotes the Riemann curvature then
Z
1
χ(M ) = P f (−R). ⊔
⊓
(2π)n M
The next exercises provide another description of the Euler class of a real oriented
vector bundle E → M over the compact oriented manifold M in terms of the homological
Poincaré duality. Set r = rank E, and let τE be a compactly supported form representing
the Thom class.
8.3. COMPUTING CHARACTERISTIC CLASSES 373

Exercise 8.3.20. Let Φ : N → M be a smooth map, where N is compact and oriented.

We denote by Φ# the bundle map Φ∗ E → E induced by the pullback operation. Show
that Φ∗ τE := (Φ# )∗ τE ∈ Ωr (Φ∗ E) is compactly supported, and represents the Thom class
of Φ∗ E. ⊔
⊓

In the next exercise we will also assume M is endowed with a Riemann structure.
Exercise 8.3.21. Consider a nondegenerate smooth section s of E, i.e.,

(i) the set Z = s−1 (0) is a codimension r smooth submanifold of M , and

(ii) there exists a connection ∇0 on E such that

∇0X s 6= 0 along Z

for any vector field X normal to Z (along Z).

(a) (Adjunction formula.) Let ∇ be an arbitrary connection on E, and denote by NZ

the normal bundle of the embedding Z ֒→ M . Show that the adjunction map

a∇ : NZ → E |Z ,

defined by
X 7→ ∇X s X ∈ C ∞ (NZ ),
is a bundle isomorphism. Conclude that Z is endowed with a natural orientation.
(b) Show that there exists an open neighborhood N of Z ֒→ NZ ֒→ T M such that

exp : N → exp(N)

is a diffeomorphism. (Compare with Lemma 7.3.44.) Deduce that the Poincaré dual of Z
in M can be identified (via the above diffeomorphism) with the Thom class of NZ .
(c) Prove that the Euler class of E coincides with the Poincaré dual of Z . ⊔
⊓

The part (c) of the above exercise generalizes the Poincaré-Hopf theorem (see Corollary
7.3.48). In that case, the section was a nondegenerate vector field, and its zero set was a
finite collection of points. The local index of each zero measures the difference between two
orientations of the normal bundle of this finite collection of points: one is the orientation
obtained if one tautologically identifies this normal bundle with the restriction of T M to
this finite collection of points while, and the other one is obtained via the adjunction map.
In many instances one can explicitly describe a nondegenerate section and its zero set,
and thus one gets a description of the Euler class which is satisfactory for most topological
applications.
Example 8.3.22. Let U denote the tautological line bundle over the complex projective
space CPn . According to Exercise 8.3.10

c1 (U) = e(U),
374 CHAPTER 8. CHARACTERISTIC CLASSES

when we view U as a rank 2 oriented, real vector bundle. Denote by [H] the (2n − 2)-cycle
defined by the natural inclusion

ı : CPn−1 ֒→ CPn , [z0 , . . . , zn−1 ] 7→ [z0 , . . . , zn−1 , 0] ∈ CPn .

We claim that the (homological) Poincaré dual of c1 (τn ) is −[H]. We will achieve this by
showing that c1 (U∗ ) = −c1 (U) is the Poincaré dual of [H].
Let P be a degree 1 homogeneous polynomial P ∈ C[z0 , . . . , zn ]. For each complex line
L ֒→ Cn+1 , the polynomial P defines a complex linear map L → C, and hence an element
of L∗ , which we denote by P |L . We thus have a well defined map

CPn ∋ L 7→ P |L ∈ L∗ = U∗ |L ,

and the reader can check easily that this is a smooth section of U∗ , which we denote by
[P ]. Consider the special case P0 = zn . The zero set of the section [P0 ] is precisely the
image of [H]. We let the reader keep track of all the orientation conventions in Exercise
8.3.21, and conclude that c1 (U∗ ) is indeed the Poincaré dual of [H]. ⊔
⊓

Exercise 8.3.23. Let US = S 2 \ {north pole}, and UN = S 2 \ {south pole} ∼ = C. The

∗
overlap UN ∩ US is diffeomorphic with the punctured plane C = C \ {0}. For each map
g : C∗ → U (1) ∼
= S 1 denote by Lg the line bundle over S 2 defined by the gluing map

gN S : UN ∩ US → U (1), gN S (z) = g(z).

Show that Z
c1 (Lg ) = deg g,
S2

where deg g denotes the degree of the smooth map g |S 1 ⊂C∗ → S 1 . ⊔

⊓
Chapter 9

Classical Integral Geometry

Ultimately, mathematics is about solving problems, and we can trace the roots of most
remarkable achievements in mathematics to attempts of solving concrete problems which,
for various reasons, were deemed very interesting by the mathematical community.
In this chapter, we will present some very beautiful applications of the techniques
developed so far. In a sense, we are going against the natural course of things, since the
problems we will solve in this chapter were some of the catalysts for the discoveries of
many of the geometric results discussed in the previous chapters.
Integral geometry, also known as geometric probability, is a mixed breed subject.
The question it addresses have purely geometric formulations, but the solutions borrow
ideas from many mathematical areas, such as representation theory and probability. This
chapter is intended to wet the reader’s appetite for unusual questions, and make him/her
appreciate the power of the technology developed in the previous chapters.

9.1 The integral geometry of real Grassmannians

9.1.1 Co-area formulæ
As Gelfand and his school pointed out, the main trick of classical integral geometry is a
very elementary one, namely the change of order of summation. Let us explain the bare
bones version of this trick, unencumbered by various technical assumptions.
Consider a “roof”
[β
 ℄
X
α

A B

α β
where X → A and X → B are maps between finite sets. One should think of α and β as
defining two different fibrations with the same total space X.
Suppose we are given a function f : X → R. We define its “average” over X as the
“integral”
X
hf iX := f (x).
x∈X

375
376 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

We can compute the average hf iX in two different ways, either summing first over the
fibers of α, or summing first over the fibers of β, i.e.
X X  X X 
f (x) = hf iX = f (x) . (9.1.1)
a∈A α(x)=a b∈B β(x)=b

We can reformulate the above equality in a more conceptual way by introducing the
functions X
α∗ (f ) : A → C, α∗ (f )(a) = f (x),
α(x)=a

and X
β∗ (f ) : B → C, β∗ (f )(b) = f (x).
β(x)=b

The equality (9.1.1) implies that

α∗ (f ) A
= β∗ (f ) B
.

All the integral geometric results that we will prove in this chapter will be of this form.
The reason such formulæ have captured the imagination of geometers comes from the fact
that the two sides hα∗ (f )iA and hβ∗ (f )iB have rather unrelated interpretations.
The geometric situation we will investigate is a bit more complicated than the above
baby case, because in the geometric case the sets X, A, B are smooth manifolds, the maps
α and β define smooth fibrations, and the function f is not really a function, but an
object which can be integrated over X, i.e., a density. It turns out that the push forward
operations do make sense for densities resulting in formulas of the type

h f iX = α∗ (f ) A
.

These formulæ generalize the classical Fubini theorem and they are often referred to as
coarea formulæ for reasons which will become apparent a bit later. To produce such results
we need to gain a deeper understanding of the notion of density introduced in Subsection
3.4.1.
Suppose V is a finite dimensional real vector space. We denote by det V the top
exterior power of V . Given a real number s we define an s-density on V to be a map
λ : det V → R such that

λ(tΩ) = |t|s λ(Ω), ∀t ∈ R∗ , Ω ∈ det V.

We denote by |Λ|s (V ) the one dimensional space of s-densities. Note that we have a
canonical identification |Λ|0 (E) = R. We will refer to 1-densities simply as densities, and
we denote the corresponding space by |Λ|(V ).
We say that an s-density λ : det V → R is positive if

λ(det V \ 0) ⊂ (0, ∞).

We denote by |Λ|+
s (V ) the cone of positive densities.
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 377

Note that any basis (v1 , . . . , vn ) of V defines linear isomorphisms

|Λ|s V → R, λ 7→ λ(v1 ∧ · · · ∧ vn ).
In particular, we have a canonical identification
|Λ|s (Rn ) ∼
= R, λ 7→ λ(e1 ∧ · · · ∧ en ),
where (e1 , . . . , en ) is the canonical basis of Rn .
If V0 and V1 are vector spaces of the same dimension n, and g : V0 → V1 is a linear
isomorphism then we get a linear map
g ∗ : |Λ|s (V1 ) → |Λ|s (V0 ), |Λ|s (V1 ) ∋ λ 7→ g ∗ λ,
where

(g ∗ λ) (∧i vi ) = λ ∧i (gvi ) , ∀v1 , · · · , vn ∈ V0 .
If V0 = V1 = V so that g ∈ Aut(V ), then
g∗ λ = | det g|s λ.
For every g, h ∈ Aut(V ) we gave (gh)∗ = h∗ g∗ , and thus we have a left action of Aut(V )
on |Λ|s (V )
Aut(V ) × |Λ|s (V ) → |Λ|s (V ),
Aut(V ) × |Λ|s (V ) ∋ (g, λ) 7→ g∗ λ = (g−1 )∗ λ = | det g|−s λ.
We have bilinear maps
|Λ|s (V ) ⊗ |Λ|t (V ) → |Λ|s+t (V ), (λ, µ) 7→ λ · µ.
To any short exact sequence of vector spaces
α β
0 → U → V → W → 0, dim U = m, dim V = m, dim W = p
we can associate maps
\ : |Λ|+
s (U ) × |Λ|s (V ) → |Λ|s (W ),
/ : |Λ|s (V ) × |Λ|+
s (W ) → |Λ|s (U ),
and
× : |Λ|s (U ) × |Λ|s (W ) → |Λ|s (V )
as follows.

• Let µ ∈ |Λ|+ s (U ), λ ∈ |Λ|s (V ), and suppose (wj )1≤j≤p is a basis of W . Now choose
lifts vj ∈ V of wj , such that β(vj ) = wj , and a basis (ui )1≤i≤m of U such that

α(u1 ), . . . , α(um ), v1 , . . . , vp ,
is a basis of V . We set



λ (∧i α(ui ) ) ∧ (∧j vj )
(µ\λ) ∧j wj :=
.
µ ∧ i ui
It is easily seen that the above definition is independent of the choices of v’s and u’s.
378 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

• Let λ ∈ |Λ|s (V ) and ν ∈
|Λ|+

s (W ). Given a basis (ui )1≤i≤m of U , extend the linearly
independent set α(ui ) ⊂ V to a basis

α(u1 ), . . . , α(um ), v1 , · · · , vp ,

of V and now define

λ (∧i α(ui )) ∧ (∧j vj )
(λ/ν) ∧i ui :=
. (9.1.2)
ν ∧j β(vj )

Again it is easily verified that the above definition is independent of the various
choices.

• Let µ ∈ |Λ|s (U ) and ν ∈ |Λ|s (W ). To define µ × ν : det V → R it suffices to indicate

its value on a single nonzero vector of the line det V . Fix a basis (ui )1≤i≤m of U and
a basis (wj )1≤j≤p of W . Choose lifts (vj ) of wj to V . Then we set


(µ × ν) (∧i α(ui )) ∧ (∧j wj ) = µ( ∧i ui )ν(∧j vj ).

Again one can check that this is independent of the various bases (ui ) and (wj ).

Exercise 9.1.1. Prove that the above constructions are indeed independent of the various
choices of bases. ⊔
⊓

Remark 9.1.2. The constructions \, /, and × associated to a short exact sequence of

vector spaces
α β
0 → U −→ V −→ W → 0,
do depend on the maps α and β! For example if we replace α by αt = tα and β by βτ = τ β,
t, τ > 0, and if
µ ∈ |Λ|s (U ), λ ∈ |Λ|s (V ), ν ∈ |Λ|s (W ),
then
µ\αt ,βτ λ = (t/τ )s µ\α,β λ, λ/αt ,βτ ν = (t/τ )s λ/α,β ν,
λ ×αt ,βτ ν = (t/τ )−s µ ×α,β ν.
The next example illustrates this. ⊔
⊓

Example 9.1.3. Consider the short exact sequence

α β
0 → U = R −→ V = R2 −→ W = R → 0

given by
α(s) = (4s, 10s), β(x, y) = 5x − 2y.
Denote by e the canonical basis of U , by (e1 , e2 ) the canonical basis of V and by f the
canonical basis of W . We obtain canonical densities λU on U , λV on V and λW on W
given by
λU (e) = λV (e1 ∧ e2 ) = λW (f ) = 1.
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 379

We would like to describe the density λV /β ∗ λW on V . Set

f 1 = α(e) = (4, 10).

We choose f 2 ∈ V such that β(f 2 ) = f , for example, f 2 = (1, 2). Then

 
∗ 4 1
λV /β λW (e) = λV (f 1 ∧ f 2 )/λW (f ) = det = 2.
10 2

Hence λV /β ∗ λW = 2λU . ⊔
⊓

Suppose now that E → M is a real vector bundle of rank n over the smooth manifold
M . Assume that it is given by the open cover (Uα ) and gluing cocycle

gβα : Uαβ → Aut(V ),

where V is a fixed real vector space of dimension n. Then the bundle of s-densities
associated to E is the real line bundle |Λ|s E given by the open cover (Uα ) and gluing
cocycle
| det gβα |−s : Uαβ → Aut( |Λ|s (V ) ) ∼
= R∗ .
We denote by C ∞ (|Λ|s E ) the space of smooth sections of |Λ|s E. Such a section is given
by a collection of smooth functions λα : Uα → |Λ|s (V ) satisfying the gluing conditions
−1 ∗
λβ (x) = (gβα ) λα (x) = | det gβα |−s λα (x), ∀α, β, x ∈ Uαβ .

Let us point out that if V = Rn , then we have a canonical identification |Λ|s (Rn ) → R,
and in this case a density can be regarded as a collection of smooth functions λα : Uα → R
satisfying the above gluing conditions.
An s-density λ ∈ C ∞ (|Λ|s E) is called positive if for every x ∈ M we have λ(x) ∈
+
|Λ|s (Ex ).
If φ : N → M is a smooth map, and E → M is a smooth real vector bundle, then we
obtain the pullback bundle π ∗ E → N . We have canonical isomorphisms

|Λ|s π ∗ E ∼
= π ∗ |Λ|s E,

and a natural pullback map

φ∗ : C ∞ (|Λ|s E) → C ∞ (π ∗ |Λ|s E) ∼
= C ∞ (|Λ|s π ∗ E).

Given a short exact sequence of vector bundles

0 → E0 → E1 → E2 → 0

over M , we obtain maps

\ : C ∞ (|Λ|+ ∞ ∞
s E0 ) × C (|Λ|s E1 ) → C (|Λ|s E2 ),

/ : C ∞ (|Λ|s E1 ) × C ∞ (|Λ|+ ∞
s E2 ) → C (|Λ|s E0 ),
380 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

and
× : C ∞ (|Λ|s E0 ) × C ∞ (|Λ|s E2 ) → C ∞ (|Λ|s E1 ).

Observe that for every positive smooth function f : M → (0, ∞) we have

(f µ)\λ = (f −1 (µ\λ), λ/(f ν) = (f −1 )(µ/λ).

In the sequel we will almost exclusively use a special case of the above construction, when
E is the tangent bundle of the smooth manifold M . We will denote by |Λ|s (M ) the line
bundle |Λ|s (T M ), and we will refer to its sections as (smooth) s-densities on M . When
s = 1, we will use the simpler notation |Λ|M to denote |Λ|1 (M ).
As explained

in Subsection 3.4.1, an s-density on M is described by a coordinate atlas
i
Uα , (xα ) , and smooth functions λα : Uα → R satisfying the conditions
!
∂xjβ
λβ = |dβα |−s λα , where dβα = det , n = dim M. (9.1.3)
∂xiα
1≤i,j≤n

We deduce that the smooth 0-densities on M are precisely the smooth functions.

Example 9.1.4. (a) Suppose ω ∈ Ωn (M ) is a top degree differential form on M . Then

in a coordinate atlas (Uα , (xiα ) ) this form is described by a collection of forms

ωα = λα dx1α ∧ · · · ∧ dxnα .

The functions λα satisfy the gluing conditions λβ = d−1 βα λα , and we conclude that the
collection of functions |λα |s defines an s-density on M which we will denote by |ω|s .
Because of this fact, the s-densities are traditionally described as collections

λα |dxα |s , dxα := dx1α ∧ · · · ∧ dxnα .

(b) Suppose M is an orientable manifold. By fixing an orientation we choose an atlas

(Uα , (xiα ) ) so that all the determinants dβα are positive. If ω is a top dimensional form
on M described locally by forms ωα = λα dxα , then the collection of functions λα defines
a density on M . Thus, a choice of orientation produces a linear map

Ωn (M ) → C ∞ (|Λ|(M ) ).

As explained in Subsection 3.4.2, this map is a bijection.

(c) Any Riemann metric g on M defines a canonical density on M denoted by |dVg |
and called the volume density. It is locally described by the collection
p
|gα ||dxα |,

where |gα | denotes the determinant of the symmetric matrix representing the metric g in
the coordinates (xiα ). ⊔
⊓
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 381

The densities on a manifold serve one major purpose: they can be integrated. We
denote by C0 (|Λ|(M ) ) the space of continuous densities with compact support. Consider
the integration map defined in Subsection 3.4.1
Z Z
: C0 (|Λ|(M ) ) → R, |dµ| 7→ |dµ|.
M M

Note that if f is a continuous, compactly supported function on M , and |dµ| is a density,

then f |dµ| is a continuous compactly supported density, and thus there is a well defined
integral Z
f |dµ|.
M
We obtain in this fashion a natural pairing
Z
C0 (M ) × C(|Λ|M ), (f, |dµ|) 7→ f |dµ|.
M

Let us observe that if |dρ| and |dτ | are two positive densities, then there exists a positive
function f such that
|dρ| = f |dτ |.
The existence of this function follows from the Radon-Nycodim theorem. For every x ∈ M
we have R
|dρ
f (x) = lim R U ,
U →{x} U |dτ |

where the above limit is taken over open sets shrinking to x. We will use the notation
|dρ|
f := ,
|dτ |
and we will refer to f as the Jacobian of |dρ| relative to |dτ |. If φ : M → N is a
diffeomorphism and |dρ| = (Uα , ρα |dyα |) is a density on N , then we define the pullback of
|dρ| by φ to be the density φ∗ |dρ| on M defined by
φ∗ |dρ| = φ−1 (Uα ), ρα |dyα |), yαi = xiα ◦ φ.
The classical change in variables formula now takes the form
Z Z
|dρ| = φ∗ |dρ|.
N M

Example 9.1.5. Suppose φ : M → N is a diffeomorphism between two smooth m-

dimensional manifolds, ω ∈ Ωm (M ), and |ω| is the associated density. Then
φ∗ |ω| = |φ∗ ω|. ⊔
⊓
Suppose a Lie group G acts smoothly on M . Then for every g ∈ G, and any density
|dρ|, we get a new density g∗ |dρ|. The density |dρ| is called G-invariant if
g∗ |dρ| = |dρ|, ∀g ∈ G.
Note that a density is G-invariant if and only if the associated Borel measure is G-invariant.
∗ |dρ|
A positive density is invariant if the Jacobian g|dρ| is identically equal to 1, ∀g ∈ G.
382 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Proposition 9.1.6. Suppose |dρ| and |dτ | are two G-invariant positive densities. Then
|dρ|
the Jacobian J = |dτ | is a G-invariant smooth, positive function on G.

Proof. Let x ∈ M and g ∈ G. Then for every open neighborhood U of x we have

Z Z Z Z R R
|dτ | g(U ) |dτ |
|dρ| = |dρ|, |dτ | = |dτ | =⇒ RU =R ,
U g(U ) U g(U ) U |dρ| g(U ) |dρ|

and then letting U → {x} we deduce

J(x) = J(gx), ∀x ∈ M, g ∈ G. ⊔
⊓

Corollary 9.1.7. If G acts smoothly and transitively on the smooth manifold M then, up
to a positive multiplicative constant there exists at most one invariant positive density. ⊔
⊓

Suppose Φ : M → B is a submersion. The kernels of the differentials of Φ form a

vector subbundle T V M ֒→ T M consisting of the planes tangent to the fibers of Φ. We
will refer to it as the vertical tangent bundle. Since Φ is a submersion, we have a short
exact sequence of bundles over M .
DΦ
0 → T V M ֒→ T M −→ Φ∗ T B → 0.

Observe that any (positive) density |dν| on B defines by pullback a (positive) density
Φ∗ |dν| associated to the bundle Φ∗ T B → M . If λ is a density associated to the vertical
tangent bundle T V M , then we obtain a density λ × Φ∗ |dν| on M .
Suppose |dµ| is a density on M such that Φ is proper on the support of |dµ|. Set
k := dim B, r := dim M − dim B. We would like to describe a density Φ∗ |dµ| on B called
the pushforward of |dµ| by Φ. Intuitively, Φ∗ |dµ| is the unique density on B such that for
any open subset U ⊂ B we have
Z Z
Φ∗ |dµ| = |dµ|.
U Φ−1 (U )

Proposition 9.1.8 (The pushforward of a density). There exists a smooth density Φ∗ |dµ|
on B uniquely characterized by the following condition. For every density |dν| on B we
have
Φ∗ |dµ| = Vν |dν|,
where Vν ∈ C ∞ (B) is given by
Z
Vν (b) := |dµ|/Φ∗ |dν|.
Φ−1 (b)

Proof. Fix a positive density
|dν| on B. Along every fiber Mb = Φ−1 (b) we have a density
|dµ|b /Φ∗ |dν| ∈ C ∞ |Λ|(Mb ) corresponding to the short exact sequence

DΦ
0 → T Mb → (T M )|Mb −→ (Φ∗ T B)|Mb → 0.
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 383

To understand this density fix x ∈ Mb . Then we can find local coordinates (yj )1≤j≤k near
b ∈ B, and smooth functions (xi )1≤i≤r defined in a neighborhood V of x in M such that
the collection of functions (xi , y j ) defines local coordinates near x on M , and in these
coordinates the map Φ is given by the projection (x, y) 7→ y.
In the coordinates y on B we can write

|dν| = ρB (y)|dy| and |dµ| = ρM (x, y)|dx ∧ dy|.

Then along the fibers y = const we have

ρM (x, y)
|dµ|b /Φ∗ |dν| = |dx|.
ρB (y)
We set Z
Vν (b) := |dµ|b /Φ∗ |dν|
Mb

The function Vν : B → R is smooth function. We can associate to Vν the density Vν |dν|

which a priori depends on ν.
Observe that if |dν̂| is another density, then there exists a positive smooth function
w : B → R such that
|dν̂| = w|dν|.
Then
|dµ|b /Φ∗ |dν̂| = w−1 |dµ|b /Φ∗ |dν|, Vν̂ = w−1 Vν
so that
Vν̂ |dν̂| = Vν |dν|.
In other words, the density Vν |dν| on B is independent on ν. It depends only on |dµ|. ⊔
⊓

Using partitions of unity and the classical Fubini theorem we obtain the coarea formula
for densities .
Theorem 9.1.9 (Coarea formula). Suppose the Φ : M → B is a submersion, |dν| is a
density on B and |dµ| is a density on M such that Φ is proper on the support of |dµ|.
Z Z Z Z !
|dµ| = Φ∗ |dµ| = |dµ|/Φ∗ |dν| |dν|. (9.1.4)
M B B Φ−1 (b)

Remark 9.1.10. Very often the submersion Φ : M → B satisfies the following condition.
For every point on the base b ∈ B there exist an open neighborhood U of b in B, a
nowhere vanishing form ω ∈ Ωk (U ), a nowhere vanishing form Ω ∈ Ωk+r (MU ), (MU :=
Φ−1 (U )), and a form η ∈ Ωr (MU ) such that

Ω = η ∧ π ∗ ω.

Then we can write |dµ| = ρ|Ω|, for some ρ ∈ C ∞ (MU ). The form ω defines a density |ω|
on U , and we conclude that

|dµ|/f ∗ |ω| = ρ|η|, Φ∗ |dµ| = f |ω|,

384 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

where for every u ∈ U ⊂ B we have

Z Z Z
∗ ∗
f (u) = |dµ|/f |ω| = ρ|Ω|/|f ω| = ρ|η|.
Mu Mu Mu

In particular, using (9.1.4) we obtain the generalized coarea formula

Z Z Z Z Z 
|dµ| = ρ|η ∧ π ∗ ω| = f (u)|ω| = ρ|η| |ω|. (9.1.5)
MU MU U U Mu

⊔
⊓

Example 9.1.11. (The classical coarea formula). Suppose (M, g) is a Riemann

manifold of dimension m + 1, and f : M → R is a smooth function without critical points.
For simplicity, we also assume that the level sets Mt = f −1 (t) are compact.
On M we have a volume density |dVg |. We would like to compute the pushforward
density f∗ |dVg | on R. We seek f∗ |dVg | of the form

f∗ |dVg | = v(t)|dt|,

where |dt| is the Euclidean volume density on R, and v is a smooth function.

For t ∈ R, the fiber Mt is a compact codimension 1 submanifold of M . We denote by
|dVt | the volume density on Mt defined by the induced metric gt := g|Mt . We denote by
1
∇f the g-gradient of f , and we set n := |∇f | ∇f .
Fix t0 ∈ R. For every point p ∈ Mt0 we have df (p) 6= 0, and from the implicit func-
tion theorem we deduce that we can find an open neighborhood U , and smooth function
x1 , . . . , xm such that (f, x1 , . . . , xm ) are local coordinates on U . Then along U we can
write
|dVg | = ρ|df ∧ dx1 ∧ · · · ∧ dxm |,
where ρ is a smooth function. The vector field n is a unit normal vector field along
Mt0 ∩ U , so that we have the equality

|dVt0 | |U ∩Mt0 = ρ n (df ∧ dx1 ∧ · · · ∧ dxm ) |U ∩Mt0

= ρ|∇f | · (dx1 ∧ · · · ∧ dxm ) |U ∩Mt0

Now observe that along U we have

ρ(df ∧ dx1 ∧ · · · ∧ dxm ) = ρ(f ∗ dt ∧ dx1 ∧ · · · ∧ dxm ).

Hence
|dVg |/f ∗ |dt| = ρ|(dx1 ∧ · · · ∧ dxm )|,
so that
1
|dVt0 | |U ∩Mt0 = |∇f | · |dVg |/f ∗ |dt| and |dVg |/f ∗ |dt| = |dVt0 | |U ∩Mt0 .
|∇f |
Hence Z
1
f∗ |dVg | = v(t)|dt|, v(t) = |dVt |, (9.1.6)
Mt |∇f |
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 385

and we obtain in this fashion the classical coarea formula

Z Z Z !
1
|dVg | = |dVt | |dt|. (9.1.7)
M R Mt |∇f |

To see how this works in practice, consider the unit sphere S n ⊂ Rn+1 . We denote the
coordinates in Rn+1 by (t, x1 , . . . , xn ). We let P± ∈ S n denote the poles given t = ±1.
We denote by |dVn | the volume density on S n , and by π : S n → R the natural projection
given by
(t, x1 , . . . , xn ) 7→ t.
The map π is a submersion on the complement of the poles, M = S n \ {P± }, and π(M ) =
(−1, 1). We want to compute π∗ |dVn |.

p 1
θ
θ t
1

0 t p'

Figure 9.1: Slicing a sphere by hyperplanes

Observe that π −1 (t) is the (n − 1)-dimensional sphere of radius (1 − t2 )1/2 . To find

the gradient ∇π observe that for every p ∈ S n the tangent vector ∇π(p) is the projection
of the vector ∂t on the tangent space Tp S n , because ∂t is the gradient with respect to the
Euclidean metric on Rn+1 of the linear function π : Rn+1 → R, π(t, xi ) = t.
We denote by θ the angle between ∂t and Tp S n , set p′ = π(p), and by t the coordinate
of p′ (see Figure 9.1). Then θ is equal to the angle at p between the radius [0, p] and the
segment [p, p′ ]. We deduce
cos θ = length [p, p′ ] = (1 − t2 )1/2 .
Hence
|∇π(p)| = (1 − t2 )1/2 ,
and consequently,
Z Z
1 n−2
|dVt | = (1 − t2 )−1/2 |dVt | = σn−1 (1 − t2 ) 2 ,
π −1 (t) |∇π| π −1 (t)

where σ m denotes the m-dimensional area of the unit m-dimensional sphere S m . The last
formula implies
Z 1 Z 1  
2 n−2 n−2
−1/2 1 n
σ n = 2σ n (1 − t ) 2 |dt| = σ n−1 (1 − s) 2 s |ds| = σ n−1 B , ,
0 0 2 2
386 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

where B denotes the Beta function

Z 1
B(p, q) := sp−1 (1 − s)q−1 ds, p, q > 0. (9.1.8)
0

It is known (see [134], Section 12.41) that

Z ∞
xp dx Γ(p)Γ(q)
B(p, q) = p+q
· = , (9.1.9)
0 (1 + x) x Γ(p + q)
where Γ(x) denotes Euler’s Gamma function
Z ∞
Γ(x) = e−t tx−1 dt.
0

We deduce
2Γ( 12 )n+1
σn = . (9.1.10)
Γ( n+12 )
⊔
⊓

The equality (9.1.7) can be further generalized. We start with a simple linear algebra
result.
Lemma 9.1.12. Suppose that U and V are Euclidean spaces of dimensions n + k and
respectively k, n ≥ 0, and A : U → V is a surjective linear map. Then there exist Eu-
clidean coordinates x1 , . . . , xn+k on U , Euclidean coordinates y 1 , . . . , y k on V and positive
numbers µ1 , . . . , µk such that, in these coordinates the operator A is described by

y j = µj xj , 1 ≤ j ≤ k.

The numbers µ21 , . . . , µ2k are the eigenvalues of the positive symmetric operator AA∗ : V →
V so that √
µ1 · · · µk = det AA∗ .

Proof. Let W denote the orthogonal complement of ker A in U . Denote by A0 the re-
striction of A to W so that A0 : W → V is a linear isomorphism. Note that W coincides
with the range of the adjoint operator A∗ : V → U so that

A0 A∗0 = AA∗ .

We want to find a linear isometry R : V → W such that the operator

B = A0 R : V → V

is symmetric. Note that since R is an isometry we have R−1 = R∗ . Moreover we have a

commutative diagram
W
A0
V u w u
R 1V

V B
wV
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 387

Note that A0 A∗ : V → V is positive and symmetric. We define

R := A∗0 (A0 A∗0 )−1/2 : V → W .
Let us show that R is indeed an isometry. Indeed, for any v ∈ V we have



(Rv, Rv) = A∗0 (A0 A∗0 )−1/2 v, A∗0 (A0 A∗0 )−1/2 v = (A0 A∗0 )−1/2 v, A0 A∗0 (A0 A∗0 )−1/2 v


= (A0 A∗0 )−1/2 v, (A0 A∗0 )1/2 v = (v, v).
Clearly A0 R = A0 A∗0 (A0 A∗0 )−1/2 = (A0 A∗0 )1/2 is symmetric. Now choose an orthonormal
basis that diagonalizes B. Transport it via R to an orthonormal basis of W . With respect
to these bases of W and V the operator A is described by a diagonal matrix with entries
consisting of the eigenvalues of A0 R = (A0 A∗0 )1/2 .
⊔
⊓

Suppose now that A : U 1 → U 2 , dim U 1 = n + k, dim U 2 = k is a surjective linear

map between Euclidean spaces. Let µi ∈ |Λ|(U i ) denote the Euclidean density on U i ,
i = 1, 2. Set U 0 := ker A and denote by µ0 the Euclidean density on U 0 .
Using Lemma 9.1.12 we can find Euclidean coordinates x1 , . . . , xn+k on U 1 , Euclidean
coordinates y 1 , . . . , y n on U 2 , and positive numbers µ1 , . . . , µn such that A is described
the equations,
y j = µj , j = 1, . . . , n.
Then
µ0 = |dxn+1 ∧ · · · ∧ dxn+k |, µ1 = |dx1 ∧ · · · ∧ dxn+k |,
µ2 = |dy 1 ∧ · · · ∧ dy n |.
Consider now the short exact sequence of vector spaces
i A
0 → U 0 ֒→ U 1 −→ U 2 → 0,
where i is the canonical inclusion. Using (9.1.2) we deduce
1 1
µ2 /A∗ µ1 = |dxn+1 ∧ · · · ∧ dxn+k | = µ0 . (9.1.11)
µ1 · · · µn JA
Suppose now that (M, gM ), (B, gB ) are Riemannian manifolds, Φ : M → B is a submersion
and f : M → R is a continuous function such that Φ is proper on the support of of f .
Denote by |dVM | the metric volume density on M and by dVB . We define its Jacobian to
be the function p
JΦ : M → R, JΦ (p) = det(DΦ(x))∗ DΦ(p),
where DΦ(p) : Tx M → TΦ(x) B is the differential of Φ at p. From (9.1.11) we deduce that
1
|dVM |/Φ∗ |dVN | Φ−1 (b)
= |dVΦ−1 (b) |,
JΦ
where |dVΦ−1 (b) | denotes the metric volume density along the fiber Φ−1 (b), b ∈ B.
From the coarea formula (9.1.4) we deduce that metric coarea formula
Z Z Z !
f
f |dVM | = |dVΦ−1 (b) | |dVB |. (9.1.12)
M B Φ−1 (b) JΦ
388 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

9.1.2 Invariant measures on linear Grassmannians

Suppose that V is an Euclidean real vector space of dimension n. We denote by • the inner
product on V . For every subspace U ⊂ V we denote by PU the orthogonal projection onto
U . We would like to investigate a certain natural density on Grk (V ), the Grassmannian
of k dimensional subspaces of V . We are forced into considering densities rather than
forms because very often the Grassmannians Grk (V ) are not orientable. More precisely,
Grk (V ) is orientable if and only if dim V is even.
In Example 1.2.22 we have shown that, via the (orthogonal) projection map U 7→ PU ,
we can regard Grk (V ) as a submanifold of End+ (V ), the vector space of symmetric
endomorphisms of V . The Euclidean metric (1.2.2) on End(V ), defined by

1
hA, Bi = tr(AB † ), ∀A, B ∈ End(V ),
2
induces a metric h = hn,k on Grk (V ).
Denote by O(V ) the group of orthogonal transformations of V . The group O(V ) acts
smoothly and transitively on Grk (V )

O(V ) × Grk (V ) ∋ (g, L) 7→ g(L) ∈ Grk (V ).

Note that
PgL = gPL g−1 .
The action of O(V ) on End+ (V ) by conjugation preserves the inner product on End+ (V ),
and we deduce that the action of O(V ) on Grk (V ) preserves the metric h.
We would like to express this metric in the graph coordinates introduced in Example
1.2.22. Consider L ∈ Grk (V ), and S ∈ Hom(L, L⊥ ). Then, for every t ∈ R, we denote by
Ut the graph of the map tS : L → L⊥ ,

Ut := ΓtS ∈ Grk (V, L).

If U̇0 denotes the tangent to the path t 7→ Ut at t = 0, then

1 d
h(U̇0 , U̇0 ) = tr(Ṗ02 ), Ṗ0 := |t=0 PΓtS ,
2 dt
If we write Pt := PΓtS we deduce from (1.2.5) that
 
(1L + t2 S ∗ S)−1 t(1L + t2 S ∗ S)−1 S ∗
Pt = .
tS(1L + t2 S ∗ S)−1 t2 S(1L + t2 S ∗ S)−1 S ∗

Hence  
0 S∗
Ṗt=0 = = S ∗ PL⊥ + SPL , (9.1.13)
S 0
so that
1

h(U̇0 , U̇0 ) = tr(SS ∗ ) + tr(S ∗ S) = tr(SS ∗ ).
2
We can be even more concrete.
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 389

Let us choose an orthonormal basis (~

ei )1≤i≤k of L, and an orthonormal basis (eα )k<α≤n
of L . With respect to these bases, the map S : L → L⊥ is described by a matrix
⊥

(sαi )1≤i≤k<α≤n , and then

X
tr(SS ∗ ) = tr(S ∗ S) = |sαi |2 .
i,α

We can think of the collection (sαi ) as defining local coordinates on the open subset
Grk (V, L) ⊂ Grk (V ) consisting of k-planes intersecting L⊥ transversally. Hence
X
h(U̇0 , U̇0 ) = |sαi |2 . (9.1.14)
i,α

In integral geometric computations we will find convenient to relate the above coor-
dinates to the classical language of moving frames. In the sequel we make the following
notational conventions.

• We will use lower case Latin letters i, j, ... to denote indices in the range {1, ..., k}.

• We will use lower case Greek letters α, β, γ, ... to denote indices in the range
{k + 1, . . . , n}.

• We will use upper case Latin letters A, B, C, ... to denote indices in the range
{1, · · · , n}.

Suppose we have a smooth 1-parameter family of orthonormal frames

(eA ) = (eA (t)), |t| ≪ 1.

This defines a smooth path

t 7→ Lt = span (ei (t) ) ∈ Grk (V ).

We would like to compute ĥ(L̇0 , L̇0 ).

Observe that we have a smooth path t 7→ gt ∈ O(V ), defined by

gt eA (0) = eA (t).

With respect to the fixed frame ( eA (0) ), the orthogonal transformation gt is given by a
matrix ( sAB (t) ), where

sAB = eA (0) • gt eB (0) .
Observe that g0 = 1V . Let Pt denote the projection onto Lt . Then

Pt = gt P0 gt−1 ,
d
so that, if we set X = dt |t=0 gt , we have

Ṗ0 = [X, P0 ].
390 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

With respect to the decomposition V = L0 + L⊥0 the projector P0 has the block decom-
position  
1L0 0
P0 = .
0 0
The operator X is represented by a skew-symmetric matrix with entries

xAB = ṡAB = eA • ėB ,

which has the block form " #

XL0 ,L0 −XL∗ ⊥ ,L
X= 0 0
,
XL⊥ ,L0 XL⊥ ,L⊥
0 0 0

where XL⊥ ,L0 denotes a map L0 → L⊥

0 etc. We deduce
0
" #
0 XL∗ ⊥ ,L
[X, P0 ] = 0 0
.
XL⊥ ,L0 0
0

Hence X
1
h(L̇0 , L̇0 ) = tr(Ṗ0 , Ṗ0 ) = tr(XL⊥ ,L0 XL∗ ⊥ ,L0 ) = |ṡαi |2 . (9.1.15)
2 0 0
α,i

We want to interpret this in the language of moving frames.

Suppose M is a smooth m-dimensional manifold and L : M → Grk (V ) is a smooth
map. Fix a point p0 ∈ M and local coordinates (ui )1≤i≤m near p0 such that ui (p0 ) = 0.
The map Lt can be described near p0 via a moving frame, i.e., an orthonormal frame
(eA ) of V , varying smoothly with (ui ), such that

L(u) = span ( ei (u) ).

The above computations show that the differential of L at p0 is described by the (n−k)×k
matrix of 1-forms on M
Dp0 L = (θαi ), θαi := eα • dei .
More precisely, this means that if X = (u̇a ) ∈ Tp0 M , and we let “ ˙ ” denote the differen-
tiation along the flow lines of X, then
X
Dp0 L(X) = (xαi ) ∈ TL(0) Grk (V ), xαi = eα • ėi = eα • dei (X). (9.1.16)
a

If M happens to be an open subset of Grk (V ), then we can use the forms θαi to describe
the metric h. More precisely, the equalities (9.1.14) and (9.1.15) show that
X
h= θαi ⊗ θαi . (9.1.17)
α,i

In other words, the collection (θαi ) is an orthonormal frame of the cotangent bundle
T ∗ Grk (V ).
The metric h = hn,k is O(V )-invariant so that the associated Riemannian volume
defines an invariant density. We will denote this invariant density by |dγn,k |, where n =
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 391

dim V , and we will refer to it as the kinematic density on Grk (V ). Since the action of
the group O(V ) is transitive, we deduce that any other invariant density is equivalent to
a constant multiple of this metric density. We would like to give a local description of
|dγn,k |.
Set n := dim V . If O is a sufficiently small open subset of Grk (V ) then we can find
smooth maps
eA : O → V, A = 1, · · · , n,
with the following properties.


• For every L ∈ O the collection eA (L) 1≤A≤n is an orthonormal frame of V .


• For every L ∈ O the collection ei (L) 1≤i≤k is an orthonormal frame of L.
For every 1 ≤ i ≤ k, and every k + 1 ≤ α ≤ 1, we have a 1-form
θαi ∈ Ω1 (O), θαi = eα • dei = −deα • ei .
As explained above, the metric h is described along O by the symmetric tensor
X
h= θαi ⊗ θαi ,
α,i

and the associated volume density is described by

Y ^
|dγn,k | = θαi := θαi .
α,i α,i

Example 9.1.13. To understand the above construction it is helpful to consider the

special case of the Grassmannian Gr1 (R2 ) of lines through the origin in R2 . This space is
diffeomorphic to the real projective line RP1 , which in turn, it is diffeomorphic to a circle.
A line L in R2 is uniquely determined by the angle θ ∈ [0, π] that it forms with the
x-axis. For such an angle θ, we denote by Lθ the corresponding line. The line Lθ is also
represented by the the orthonormal frame
e1 (θ) = (cos θ, sin θ), e2 (θ) = (− sin θ, cos θ), Lθ = span ( e1 (θ) ),
where the first vector e1 (θ) gives the direction of the line. Then θ21 = e2 • de1 = dθ, and
|dγ2,1 | = |dθ|. ⊔
⊓

We would like to compute the volumes of the Grassmannians Grk (V ), dim V = n with
respect to the kinematic density |dγn,k |, i.e., we would like to compute
Z
Cn,k := |dγn,k |.
Grk (V )

Denote by ω n the volume of the unit ball B n ⊂ Rn , and by σn−1 the (n − 1)-dimensional
“surface area” of the unit sphere S n−1 , so that
 k

 π

 , n = 2k,
Γ(1/2)n  k!
σn−1 = nω n , and ω n = = ,
Γ(1 + n/2)   2k+1 π k k!

 2
 , n = 2k + 1,
(2k + 1)!
392 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

where Γ(x) is the Gamma function. We list below the values of ω n for small n.

n 0 1 2 3 4
4π π2 .
ωn 1 2 π 3 2

To compute the volume of the Grassmannians we need to use yet another description of
the Grassmannians, as homogeneous spaces.
Note first that the group O(V ) acts transitively on Grk (V ). Fix L0 ∈ Grk (V ).
Then the stabilizer of L0 with respect to the action of O(V ) on Grk (V ) is the sub-
group O(L0 ) × O(L⊥ 0 ) and thus we can identify Grk (V ) with the homogeneous space of
left cosets O(V )/O(L0 ) × O(L⊥0 ).
The computation of Cn,k is carried out in three steps.
Step 1. We equip the orthogonal groups O(Rn ) with a canonical invariant density |dγn |
called the kinematic density on O(n). Set
Z
Cn := |dγn |.
O(Rn )

Step 2. We show that

Cn
Cn,k = .
Ck Cn−k
Step 3. We show that
1 nω n
Cn,1 = σ n−1 = ,
2 2
and then compute Cn inductively using the recurrence relation from Step 2

Cn+1 = (C1 Cn,1 )Cn ,

and the initial condition

C2 = vol ( O(2) ) = 2σ 1 .

Step 1. The group O(V ) is a submanifold of End(V ) consisting of endomorphisms S

satisfying SS ∗ = S ∗ S = 1V . We equip End(V ) with the inner product
1
hA, Bi = tr(AB ∗ ).
2
This metric induces an invariant metric h on O(V ). We would like to give a more concrete
description of this metric.
Denote by End− (V ) the subspace of End(V ) consisting of skew-symmetric operators.
For any S0 ∈ O(V ) we have a map

expS0 : End− (V ) → O(V ), End− (V ) 7−→ S0 · exp(X).

This defines a diffeomorphism from a neighborhood of 0 in End− (V ) to a neighborhood

of S0 in O(V ). Two skew-symmetric endomorphisms X, Y ∈ End− (V ) define paths

γX , γY : R → O(V ), γX (t) = S0 exp(tX), γY (t) = S0 exp(tY ),

9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 393

originating at S0 . We set

Ẋ := γ̇X (0) ∈ TS0 O(V ) ⊂ End(V ), Ẏ := γ̇Y (0) ∈ TS0 O(V ) ⊂ End(V ).

Then Ẋ = S0 X, Ẏ = S0 Y , and

1
1
1

h(Ẋ, Ẏ ) = tr (S0 X)(S0 Y )∗ = tr S0 XY ∗ S0∗ = tr S0∗ S0 XY ∗
2 2 2
1

= tr XY ∗
2
If we choose an orthonormal basis (eA ) of V so that X and Y are given by the skew
symmetric matrices (xAB ), (yAB ), then we deduce
X
h(Ẋ, Ẏ ) = xAB yAB .
A>B

If we set f A (t) := exp(tX)eA , then we deduce

xAB = eA • ḟ B (0) = f A (0) • ḟ B (0).

More generally, if we define

f A : O(V ) → V, f A (S) := SeA ,

then we obtain the angular forms θAB := f A • df B . The above metric metric has the
description X
h= θAB ⊗ θAB .
A>B

The associated volume density is

^
|dγn | = θAB .
A>B

Step 2. Fix an orthonormal frame (eA ) of V such that L0 = span (ei ; 1 ≤ i ≤ k). We
can identify V with Rn , O(V ) with O(n), and L0 with the subspace

Rk ⊕ 0n−k ⊂ Rn .

An orthogonal n × n matrix T is uniquely determined by the orthonormal frame (T eA )

via the equalities
TAB = eA • T eB .
Define
p : O(n) → Grk (Rn ), p(T ) = T (L0 ).
More explicitly, we have
p(T ) = span (T ei )1≤i≤k .
394 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

We will prove that we have a principal fibration

O(k) × O(n − k) y w O(n)

upn ,
Grk (R )

and that
p∗ |dγn | = Ck Cn−k |dγn,k |.
Once we have this we deduce from the Fubini formula for densities (9.1.5) that

Cn = Ck Cn−k Cn,k .

Let us prove the above facts.

For every sufficiently small open subset U ⊂ Grk (V ) we can find a smooth section

φ : U → O(n)

of the map p : O(n) → Grk (Rn ). The section can be identified with a smooth family of
orthonormal frames (φA (L), L ∈ U )1≤A≤n of Rn , such that

L = span (φi (L); 1 ≤ i ≤ k).

To such a frame


orthogonal matrix φ(L) ∈ O(n) which maps the fixed
we associate the
frame eA to the frame φB . It is a given by a matrix with entries

φ(L)AB = eA • φB .

Then we have a smooth map

Ψ : O(k) × O(n − k) × U → p−1 (U ),

defined as follows.

• Given (s, t, L) ∈ O(k) × O(n − k) × U , express s as a k × k matrix s = (sij ), and t

as a (n − k) × (n − k) matrix (tαβ ).

• Define the orthonormal frame of Rn

( f A ) := (φB ) ∗ (s, t),

via the equalities

X
f i = f i (s, L) = sji φj (L) ∈ L, 1 ≤ i ≤ k, (9.1.18)
j

X
f α = f α (t, L) = tβα φβ (L) ∈ L⊥ , k + 1 ≤ α ≤ n. (9.1.19)
β
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 395

• Now define Ψ = Ψ(s, t, L) to be the orthogonal transformation of Rn which maps

the frame (eA ) to the frame (f B ), i.e.

f A = ΨeA , ∀A.

The map Ψ is a homeomorphism with inverse

O(n) ∋ T 7→ Ψ−1 (T ) = (s, t; L) ∈ O(k) × O(n − k) × L

defined as follows. We set f A = f A (T ) := T eA , 1 ≤ A ≤ n. Then

L = LT = span (f i )1≤i≤k ,

while the matrices (sij ) and (tαβ ) are obtained via (9.1.18) and (9.1.19). More precisely, we
have
sij = φi (LT ) • f j , sαβ = φα (LT ) • f β .
Observe that, ∀s0 , s1 ∈ O(k), t0 , t1 ∈ O(n − k), we have


(φB ) ∗ (s0 , t0 ) ∗ (s1 , t1 ) = (φB ) ∗ (s0 s1 , t0 t1 ).

This means that Ψ is equivariant with respect to the right actions of O(k) × O(n − k) on
O(k) × O(n − k) × U and O(n). We have a commutative diagram

O(k) × O(n − k) × U
hhh
Ψ
' w p−1(U )
πhj *'''p'
U ⊂ Grk (Rn )

In particular, this shows that p defines a principal O(k) × O(n − k)-bundle.

Observe now that p∗ |dγn | is an invariant density on Grk (Rn ), and thus there exists a
constant c such that
p∗ |dγn | = c|dγn,k |.
This constant is given by the integral of the density |dγn |/p∗ |dγn,k | along the fiber p−1 (L0 ).
Hence
Z Z Z
Cn = |dγn | = p∗ |dγn | = c |dγn,k | = cCn,k .
O(n) Grk (Rn ) Grk (Rn )

Recall that if we define f A : O(n) → Rn by f A (T ) := T eA , and θAB := f A • df B , then

^
|dγn | = | θAB |.
A>B

We write this as ^  ^  ^ 
θij ∧ θαβ ∧ θα,i .
i>j α>β α,i
396 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY
V 
The form α,i θα,i is the pullback of a nowhere vanishing form defined in a neighborhood
of L0 in Grk (Rn ), whose associated density is |dγn,k |. We now find ourselves in the
situation described in Remark 9.1.10. We deduce
Z ^   ^ 
c= θij ∧ θαβ
p−1 (L0 ) i>j α>β
Z Z 
= |dγk | |dγn−k | = Ck Cn−k .
O(k) O(n−k)

Hence
Cn
Cn,k = .
Ck Cn−k
n−1
Step 3. Fix an orthonormal basis {eA } of V , and denote by S + the open hemisphere
n−1

S+ = ~v ∈ V ; |~v | = 1, ~v • e1 > 0 .
Note that Gr1 (V ) ∼ = RPn−1 is the Grassmannian of lines in V . The set of lines that do
n−1
not intersect S + is a smooth hypersurface of Gr1 (V ) diffeomorphic to RPn−2 and thus
has kinematic measure zero. We denote by Gr1 (V )∗ the open subset consisting of lines
n−1
intersecting S + . We thus have a map
ψ : Gr∗1 (V ) → S +
n−1 n−1
, ℓ 7→ ℓ ∩ S + .
This map is a diffeomorphism, and we have
Z Z Z
Cn,1 = |dγn,1 | = |dγn,1 | = (ψ −1 )∗ |dγn,1 |.
Gr1 (V ) Gr∗1 (V ) S n−1
+

Now observe that ψ is in fact an isometry, and thus we deduce

1 σn−1 nω n
Cn,1 = area (S n−1 ) = = .
2 2 2
Hence
Cn+1 = Cn C1 Cn+1,1 = σ n Cn ,
which implies inductively that
n−1
Y n−1
Y
Cn = σn−1 · · · σ 2 C2 = 2 σk = σj .
k=1 j=0

In particular, we deduce the following result.

Proposition 9.1.14. For every 1 ≤ k < n we have
Z Qn−1
j=0 σ j
Cn,k = |dγn,k | = Qk−1
Qn−k−1

Grk (Rn ) i=0 σ i · j=0 σj
  Qn
n j=1 ω j
= Qk
Qn−k
.
k i=1 ω i · j=1 ω j
⊔
⊓
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 397

Following [79], we set

n
Y
1 σ n−1 nω n ω n n!
[n] := = , [n]! := [k] = n ,
2 ω n−1 2ω n−1 2
k=1
   
n [n]! n ωn
:= = . (9.1.20)
k ([k]!)([n − k]!) k ω k ω n−k
Denote by |dνn,k | the unique invariant density on Grk (V ), dim V = n such that
Z  
n
|dνn,k | = . (9.1.21)
Grk (V ) k

We have n
k
|dνn,k | = |dγnk |.
Cn,k
Example 9.1.15. Using the computation in Example 9.1.13 we deduce

|dγ2,1 | = |dθ|, 0 ≤ θ < π,

which implies that Z π

ω2
C2,1 = |dθ| = 2 ,
0 ω12
as predicted by Proposition 9.1.14. We have
 
2 ω2
= 2 2 = C2,1 ,
1 ω1

so that |dν2,1 | = |dγ2,1 |. ⊔

⊓

9.1.3 Affine Grassmannians

We denote by Graff k (V ) the set of k-dimensional affine subspaces of V . We would like
to describe a natural structure of smooth manifold on Graff k (V ).
We have the tautological vector bundle U = Un,k → Grk (V ). It is naturally a subbun-
dle of the trivial vector bundle V = V × Grk (V ) → Grk (V ) whose fiber over L ∈ Grk (V )
is the vector subspace L. The trivial vector bundle V is equipped with a natural metric,
and we denote by U⊥ → Grk (V ) the orthogonal complement of U in V .
The fiber of U⊥ over L ∈ Grk (V ) is canonically identified with the orthogonal com-
plement L⊥ of L in V . The points of U⊥ are pairs (~c, L), where L ∈ Grk (V ), and ~c is a
vector in L⊥ .
We have a natural map A : U⊥ → Graff k (V ) given by

(~c, L) 7−→ ~c + L.

This map is a bijection with inverse

Graff k (V ) ∋ S 7→ (S ∩ [S]⊥ , [S]),

398 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

where [S] ∈ Grk (V ) denotes the affine subspace through the origin parallel to S,

[S] = S − S = s1 − s2 ; s1 , s2 ∈ S .

We set
~c(S) := S ∩ [S]⊥ ,
and we say that ~c(S) is the center of the affine plane S.
We equip Graff k (V ) with the structure of smooth manifold which makes A a diffeo-
morphism. Thus, we identify Graff k (V ) with a vector subbundle of the trivial bundle
V × Grk (V ) described by

U⊥ = (~c, L) ∈ V × Grk (V ); PL~c = 0 ,

where PL denotes the orthogonal projection onto L.

The projection π : U⊥ → Grk (V ) is a submersion. The fiber of this submersion over
L ∈ Grk (V ) is canonically identified with the vector subspace L⊥ ⊂ V . As such, it
is equipped with a volume density |dVL⊥ |. We obtain in this fashion a density |dVL⊥ |
associated to the vertical subbundle ker Dπ ⊂ T U⊥ .
The base Grk (V ) of the submersion π : Graff k (V ) → Grk (V ) is equipped with a
density |dγn,k |, and thus we obtain a density π ∗ |dγn,k | associated to the bundle

π ∗ T Grk (V ) → Graff k (V ).

Using the short exact sequence of vector bundles over Graff k (V ),

0 → ker Dπ −→ T Graff k (V ) −→ π ∗ T Grk (V ) → 0,

we obtain a density |e
γn,k | on Graff k (V ) given by

γn,k | = |dVL⊥ | × π ∗ |dγn,k |.

|de

Let us provide a local description for this density.

e its preimage in Graff k (V ) via
Fix a small open subset O ⊂ Grk (V ), and denote by O
the projection π. Then we can find smooth maps

e→V
eA : O → V, ~r : O

with the following properties.

e (eA (S) ) is an orthonormal frame of V , and

• For every S ∈ O,


[L] = span ei ([L]) .

e we have
• For every S ∈ O
S = ~r(S) + [S].
9.1. THE INTEGRAL GEOMETRY OF REAL GRASSMANNIANS 399

We rewrite the last equality as

S = S(~r, ei ).

Observe that the center of this affine plane is the projection of ~r onto [S]⊥
X
~c(S) = ( eα • ~r )eα .
α

Following the tradition we introduce the (locally defined) 1-forms

θα := eα • d~r, θαi := eα • dei .

For fixed L ∈ Grk (V ) the density on the fiber U⊥ ⊥

L = L is given by
^
|dVL⊥ | = θα .

e by
The volume density on Graff k (V ) is described along O
^  ^ 
|de
γn,k | = θα ∧ θαi .
α α,i

The coarea formula (9.1.4) implies the following result.

Theorem 9.1.16. Suppose f : Graff k (V ) → R is a compactly supported |dγ̂n,k |-integrable

function. Then
Z Z Z !
f (S)|de
γn,k (S)| = f (p + L)|dVL⊥ (p)| |dγn,k (L)|,
Graff k (V ) Grk (V ) L⊥

where |dVL⊥ | denotes the Euclidean volume density on L⊥ . ⊔

⊓

Denote by Iso(V ) the group of affine isometries of V , i.e., the subgroup of the group of
affine transformations generated by translations, and by the rotations about a fixed point.
Any affine isometry T : V → V is described by a unique pair (t, S) ∈ V × O(V ) so that

T (v) = Sv + t, ∀v ∈ V.

The group Iso(V ) acts in an obvious fashion on Graff k (V ), and a simple computation
shows that the associated volume density |de γn,k | is Iso(V ) invariant.
If instead of the density |dγn,k | on Grk (V ) we use the density |dνn,k |, we obtain a
density |de
νn,k | on Graff k (V ) which is a constant multiple of |de γn,k |.
n
k
|de
νn,k | = |de
γnk |. (9.1.22)
Cn,k
400 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Example 9.1.17. Let us unravel the above definition in the special case Graff 1 (R2 ), the
Grassmannians of affine lines in R2 . Such a line L is determined by two quantities: the
angle θ ∈ [0, π) is makes with the x-axis, and the signed distance ρ ∈ (−∞, ∞) from the
origin. More precisely, for every ρ ∈ R and θ ∈ [0, π) we denote by Lθ,ρ the line is given
in Euclidean coordinates by the equation

x sin θ − y cos θ = ρ

As a manifold, the Grassmannian Graff 1 (R2 ) is diffeomorphic to the interior of the Möbius
band. The Fubini formula in Theorem 9.1.16 can now be rewritten
Z Z ∞ Z π 
f (L) |de
γ2,1 |(L) = f (Lθ,ρ )|dθ| |dρ|,
Graff 1 (R2 ) −∞ 0

∞ (Graff (R2 ) ).
∀f ∈ Ccpt ⊔
⊓
1

9.2 Gauss–Bonnet again?!?

The Grassmannians enter into the study of the geometry of a submanifold via a fundamen-
tal observation going back to Gauss. To understand the shape of a submanifold M ֒→ Rn ,
m = dim M , it is very productive to understand how a tangent plane Tp M varies, as p
varies along M .
Observe that if M is flat, i.e., it is an open subset of an affine subspace, then the
tangent plane Tp M does not vary as p moves around M . However, if M is curved, this
tangent plane changes its location in the Grassmannian Grm (Rn ), and in fact, “the more
curved is M , the faster will the tangent plane move inside this Grassmannian”.
In Example 4.2.23 we have seen this heuristic principle at work in the case of a hy-
persurface in R3 , where the variation of the unit normal along the surface detects the
Riemann curvature of the surface. Note that the tangent plane is uniquely determined by
the unit normal.
The goal of this section is to analyze in great detail the map

M ∋ p 7→ Tp M ∈ Grm (Rn ),

which, for obvious reasons, is named the Gauss map of the submanifold M .

9.2.1 The shape operator and the second fundamental form

Suppose M is a smooth submanifold M of Rn , and set m := dim M . To minimize the
notational burden we will use the following conventions.

• We will use lower case Latin letters i, j, . . . to denote indices in the range

1 ≤ i, j, . . . ≤ m = dim M.

• We will use lower case Greek letters α, β, . . . to denote indices in the range

m < α, β, . . . ≤ n.
9.2. GAUSS–BONNET AGAIN?!? 401

• We will use the capital Latin letters A, B, C to denote indices in the range

1 ≤ A, B, C ≤ n.

• We will use Einstein’s summation convention.

The Euclidean metric of Rn induces a metric g on M . In Subsection 4.2.4 we ex-

plained how to determine the Levi–Civita connection ∇M of g using Cartan’s moving
frame method. Let us recall this construction using notations more appropriate for the
applications in this chapter.
Denote by D the Levi–Civita connection of the Euclidean metric on Rn . Every vector
field on Rn can be regarded as an n-uple of functions
 1 
X
 .. 
X =  . .
Xn

Then  
dX 1
 
D X =  ...  = dX.
dX n
The restriction to M of the tangent bundle T Rn admits an orthogonal decomposition

(T Rn )|M = T M ⊕ (T M )⊥ .

Correspondingly, a section X of (T Rn )|M decomposes into a tangential part X τ , and a

normal part X ν . Fix a a point p0 ∈ M , an open neighborhood U of p0 in Rn , and a
local orthonormal frame (eA ) of T Rn along U such that the collection (ei |M ) is a local
orthonormal frame of T M along U ∩ M . We denote by (θ A ) the dual coframe of (eA ),
i.e.,
θ A (eB ) = eA • eB = δAB .
If X is a section of T Rn M
then

X τ = θ i (X)ei , X ν = θ α (X)eα .

We denote by ΘA
B the 1-forms associated to D by the frame (eA ). They satisfy Cartan’s
equations
dθ A = −ΘA B A A
B ∧ θ , DeB = ΘB eA , ΘB = −ΘA .
B

If we write
φA := θ A |M , ΦA A
B := ΘB |M ,

then, as explained Subsection 4.2.4, we deduce that (Φij ) are the 1-forms associated to the
Levi–Civita connection ∇M by the local orthonormal frame (ei |M ). This implies that

∇M ej = Φij ei = the tangential component of ΦA

j eA = D ej ,
402 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

and thus,
∇M τ
X Y = (D X Y ) , ∀X, Y ∈ Vect (M ).
Consider the Gauss map
G = GM : M → Grm (Rn ), x 7→ Tx M.
The shape operator of the submanifold M ֒→ Rn is, by definition, the differential of the
Gauss map. Intuitively, the shape operator measures how fast the tanget plane changes
its location as a point in a Grassmannian.
We denote the shape operator by S M , and we would like to relate it to the structural
coefficients ΦAB.
As explained in Subsection 9.1.2, in the neighborhood U of p0 , the “moving plane”
x 7→ Tx M can be represented by the orthonormal frame (eA ) which has the property
that the first m vectors e1 (x), . . . , em (x) span Tx M . The differential of the Gauss map at
x ∈ U ∩ M is a linear map
DG : Tx M → TG(x) Grm (Rm ) = Hom(Tx M, (Tx M )⊥ ).
As explained in (9.1.16), this differential described by the (n − m) × m matrix of 1-forms



eα • dej , α,i = eα • D ej , α,i

On the other hand, D ej = ΦA α

j eA , so that eα • D ej = Φj . Define

ΦA
ij := ei ΦA 0
j ∈ Ω (M ∩ U ),

so that
ΦA A i
j = Φij ∧ φ .
We have thus obtained the following result.
Proposition 9.2.1. The shape operator of M , that is the differential of the Gauss map,
is locally described by the matrix of 1-forms (Φαi )1≤i≤m<α<n . More precisely the operator
S M (ei ) ∈ Hom(Tx M, (Tx M )⊥ )
is given by
S M (ei )ej = (D ei ej )ν = Φαij eα , ∀i, j. ⊔
⊓

In Subsection 4.2.4 we have defined the second fundamental form of the submanifold
M to be the C ∞ (M )-bilinear and symmetric map


SM : Vect (M ) × Vect (M ) → C ∞ (T M )⊥ , (X, Y ) 7→ (D X Y )ν .
Note that
SM (ei , ej ) = φα (D ei ej )eα = Φαij eα = S M (ei )ej . (9.2.1)
The equality (4.2.12) can be rewritten as
ej • (Dei eα ) = −(D ei ej ) • eα = −eα • SM (ei , ej ), ∀i, j, α. (9.2.2)
Theorema Egregium states that the Riemann curvature tensor of M is completely deter-
mined by the second fundamental form. We recall this result below.
9.2. GAUSS–BONNET AGAIN?!? 403

Theorem 9.2.2. Suppose M is a submanifold of Rn . We denote by g the induced metric

on M and by SM the second fundamental form of the embedding M ֒→ Rn . Denote by R
the Riemann curvature of M with the induced metric. Then for any X1 , . . . , X4 ∈ Vect (M )
we have

g(X1 , R(X3 , X4 )X2 ) = SM (X3 , X1 ) • SM (X4 , X2 ) − SM (X3 , X2 ) • SM (X4 , X1 ),

where • denotes the inner product in Rn . ⊔

⊓

9.2.2 The Gauss–Bonnet theorem for hypersurfaces

of an Euclidean space.
The results in the previous subsection have very surprising consequences. Suppose M is a
compact, orientable 1 hypersurface of Rm+1 , i.e., an orientable submanifold of codimension
1. If we fix an orientation on M , then we obtain a normal vector field

n : M → Rm+1 , n(x) ⊥ Tx M, |n(x)| = 1, ∀x ∈ M.

If we choose an oriented, local orthonormal frame e1 , . . . , em of T M , then the ordered set

{n(x), e1 (x), . . . , em } is an oriented, local orthonormal frame of Rm+1 . In this case we
can identify the second fundamental form with a genuine symmetric bilinear form


SM ∈ C ∞ T ∗ M ⊗2 , SM (X, Y ) = n • (D X Y ).

The Gauss map GM : M → Grm (Rm+1 ) of the embedding M ֒→ Rm+1 can be given the
description,

M ∋ x 7→ hn(x)i⊥ := the vector subspace orthogonal to n(x).

On the other hand, we have an oriented Gauss map

~GM : M → S m , x → n(x),

and a double cover

π : S m → Grm (Rm+1 ), S m ∋ ~u 7→ h~ui⊥ ,
so that the diagram below is commutative

'' w S m
~G
M
G
)' u π
Grm (Rm+1 )

We fix an oriented orthonormal frame (f~0 , f~1 , . . . , f~m ) of Rm+1 , and we orient the unit
sphere S m ⊂ Rm+1 so that the orientation of Tf~0 S m is given by the ordered frame

(f~1 , . . . , f~m ).
1
The orientability assumption is superfluous. It follows from the Alexander duality theorem with Z/2
coefficients that a compact hypersurface of a vector space is the boundary of a compact domain, and in
particular, it is orientable.
404 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

In the remainder of this subsection we will assume that m is even, m = 2h.

Denote by dAm ∈ Ωm (S m ) the “area” form on the unit m-dimensional sphere S m .
Recall that σm denotes the “area” of S m ,

σ 2h 22h π h h!
σ 2h = (2h + 1)ω 2r+1 =⇒ = . (9.2.3)
2 (2h)!

Hence Z
1
dAm = 1,
S m σm
so that Z
1 ~G∗ dAm = deg GM .
M
σm M
Denote by g the induced metric on M , and by R the curvature of g. We would like to
prove that the integrand ~G∗M dAm has the form

~G∗ dAm = P (RM )dVM ,

M

where dVM denotes the metric volume form on M and P (RM ) is a universal polynomial
of degree m2 in the curvature R of M .
Fix a positively oriented orthonormal frame ~e = (e1 , . . . , em ) of T M defined on some
open set U ⊂ M , and denote by ~θ = (θ 1 , . . . , θ m ) the dual coframe. Observe that

dVM = θ 1 ∧ · · · ∧ θ m .

We set

Sij := SM (ei , ej ) • n, Rijkℓ := g ei , R(ek , eℓ )ej .
Theorem 9.2.2 implies that

Sik Siℓ
Rijkℓ = Sik Sjℓ − Siℓ Sjk = . (9.2.4)
Sjk Sjℓ

Observe that Rijkℓ 6= 0 =⇒ i 6= j, k 6= ℓ, and in this case the matrix

 
Sik Siℓ
Sjk Sjℓ

is the 2 × 2 submatrix of S = (Sij )1≤i,j≤m lieing at the intersection of the i, j-rows with
the k, ℓ-columns. We can rephrase the equality (9.2.4) in a more convenient form.
First, we regard the curvature Rijkℓ at a point x ∈ M as a linear map
X
Λ2 Tx M → Λ2 Tx∗ M, R(ek ∧ eℓ ) = Rijkℓ θ i ∧ θ j .
i<j

Next, regard S as a linear map

S : Tx M → T ∗ M, Sej = Sij θ i .
9.2. GAUSS–BONNET AGAIN?!? 405

Then S induces linear maps

Λp S : Λk Tx M → Λp T ∗ M,

defined by

S(ei1 ∧ · · · ∧ eip ) = (Sei1 ) ∧ · · · ∧ (Seip ), ∀1 ≤ i1 < · · · < ip ≤ m.

The equality (9.2.4) can now be rephrased as

R = Λ2 S. (9.2.5)

Along U we have the equality

(~G∗M dAM ) |U = Λm S(e1 ∧ · · · ∧ em ) = (det S)θ 1 ∧ · · · ∧ θ m = (det S)dVM .

We want to prove that det S can be described in terms of Λ2 S. To see this observe that

(Λ2h S)(e1 ∧ e2 ∧ · · · e2h−1 ∧ e2h ) = (Λ2 S)(e1 ∧ e2 ) ∧ · · · ∧ (Λ2 S)(e2h−1 ∧ e2h )

h X
^  X Y
h 
= Rij,2s−1,2s θ i ∧ θ j = ǫ(ϕ) Rϕ(2s−1)ϕ(2s),2s−1,2s dVM ,
s=1 i<j ϕ∈S′m s=1

where S′m denotes the set of permutations ϕ of {1, 2, . . . , m = 2h} such that

ϕ(1) < ϕ(2), . . . , ϕ(2h − 1) < ϕ(2h),

and ǫ(ϕ) = ±1 denotes the signature of a permutation. Observe that

     
′ 2h 2h − 2 2 (2h)!
#Sm = · ··· = h . (9.2.6)
2 2 2 2

We would like to give an alternate description of det S using the concept of pfaffian.
The Riemann curvature, like the curvature of any metric connection, can be regarded
as a 2-form with coefficients in the bundle of skew-symmetric endomorphisms of T M . It
is locally described by a m × m skew-symmetric matrix


Θ = Θg := Θij 1≤i,j≤m

whose entries are 2-forms on U ,

X 1X
Θij = Rijkℓ θ k ∧ θ ℓ = Rijkℓ θ k ∧ θ ℓ = Λ2 S(ei ∧ ej ) ∈ Ω2 (U ).
2
k<ℓ k,ℓ

Recall from Subsection 8.2.4 that the pfaffian of Θ is the differential form

(−1)h X
P f (Θ) := ǫ(ϕ)Θϕ(1)ϕ(2) ∧ · · · ∧ Θϕ(2h−1)ϕ(2h) ∈ Ω2h (U ),
2h h!
ϕ∈Sm
406 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

where Sm denotes the group of permutations of {1, 2, . . . , m}. Observe that

(−1)h X
P f (Θ) = ǫ(ϕ)Θϕ(1)ϕ(2) ∧ · · · ∧ Θϕ(2h−1)ϕ(2h) ,
h! ′ϕ∈Sm

We can simplify this some more if we introduce the set S′′m consisting of permutations
ϕ ∈ S′m such that
ϕ(1) < ϕ(3) < · · · < ϕ(2h − 1).
Observe that
S′M (2h)!
#S′m = (#S′′m )h! =⇒ #S′′m = = h = 1 · 3 · · · (2h − 1) =: γ(2h).
h! 2 h!
Then X
P f (Θ) = (−1)h ǫ(ϕ)Θϕ(1)ϕ(2) ∧ · · · ∧ Θϕ(2h−1)ϕ(2h) .
ϕ∈S′′
m

We have

1 X Y
h 
P f (−Θ) = ǫ(σϕ) Rϕ(2j−1)ϕ(2j)σ(2j−1)σ(2j) dVM .
h!
(σ,ϕ)∈S′m ×S′m j=1

On the other hand,

(Λ2h S)(e1 ∧ e2 ∧ · · · ∧ e2h−1 ∧ e2h )
1 X
= ǫ(ϕ)Λm S(eϕ(1) ∧ eϕ(2) ∧ · · · ∧ eϕ(2h−1) ∧ eϕ(2h) )
#S′m ′ϕ∈Sm

1 X
= ǫ(ϕ)Θϕ(1)ϕ(2) ∧ · · · ∧ Θϕ(2h−1)ϕ(2h)
#S′m ′ ϕ∈Sm

1 X Y
h  h!
= ǫ(σϕ) Rϕ(2j−1)ϕ(2j)σ(2j−1)σ(2j) dVM = P f (−Θ).
(#S′m ) #S′m
(σ,ϕ)∈S′m ×S′m j=1

Hence
~G∗ dAM = (Λ2h S)(e1 ∧ e2 ∧ · · · ∧ e2h−1 ∧ e2h ) = h!
M P f (−Θ),
#S′m
so that
2
2 h! (9.2.3) h! (2h)!
~G∗ dAM = P f (−Θ) = P f (−Θ)
M ′
σ 2h σ2h #Sm #Sm 2 π h h!
′ 2h

(9.2.6) 1
= P f (−Θ)
(2π)h
The last form is the Euler form e(∇M ) associated to the Levi–Civita connection ∇M .
Using the Gauss–Bonnet–Chern theorem we obtain the following result.
9.2. GAUSS–BONNET AGAIN?!? 407

Theorem 9.2.3. If M 2h ⊂ R2h+1 is a compact, oriented hypersurface, and g denotes the

induced metric, then

2 ~∗ 1
GM dAm = e(∇M ) = P f (−Θg ),
σm (2π)h

and Z
2 deg ~GM = e(∇M ) = χ(M ).
M

Moreover, if (e1 , . . . , e2h ) is a local, positively oriented orthonormal frame of T M , then

1 X Y
h 
P f (−Θg ) = ǫ(σϕ) Rϕ(2j−1)ϕ(2j)σ(2j−1)σ(2j) dVM (9.2.7a)
h!
(σ,ϕ)∈S′2h ×S′m j=1

X
= ǫ(ϕ)Θϕ(1)ϕ(2) ∧ · · · ∧ Θϕ(2h−1)ϕ(2h) , (9.2.7b)
ϕ∈S′′
2h

where S′2h denotes the set of permutations ϕ of {1, . . . , 2h} such that

ϕ(2j − 1) < ϕ(2j), ∀1 ≤ j ≤ h,

S′′m denotes the set of permutations ϕ ∈ S′m such that

ϕ(1) < ϕ(3) < · · · < ϕ(2h − 1),

and X
Θij = Rijkℓ θ k ∧ θ ℓ . ⊔
⊓
k<ℓ

Example 9.2.4. (a) If dim M = 2 then

P f (−Θg ) = R1212 dVM = (the Gaussian curvature of M ) × dVM

(b) If dim M = 4, then S′′4 consists of 3 permutations

1, 2, 3, 4, ǫ = 1 −→ Θ1212 Θ3434 ,
1, 3, 2, 4, ǫ = −1 −→ −Θ1312 Θ2434
1, 4, 2, 3, ǫ = 1 −→ Θ1412 Θ2334

We deduce
P f (−Θg ) = Θ12 ∧ Θ34 − Θ13 ∧ Θ24 + Θ14 ∧ Θ23
1 1
= (R12ij θ i ∧ θ j ) ∧ (R34kℓ θ k ∧ θ ℓ ) − (R13ij θ i ∧ θ j ) ∧ (R24kℓ θ k ∧ θ ℓ )
4 4
1
+ (R14ij θ i ∧ θ j ) ∧ (R23kℓ θ k ∧ θ ℓ )
4
408 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Each of the three exterior products above in involves only six nontrivial terms, because
θ i ∧ θ j = 0, when i = j. Thus the total number of terms in the above expression is 36.
However, there are many repetitions due to the symmetries of the Riemann tensor

Rijkℓ = Rkℓij = −Rijℓk .

We have
(R12ij θ i ∧ θ j ) ∧ (R34kℓ θ k ∧ θ ℓ )

= R1212 R3434 + R1213 R3442 + R1214 R3423

+R1223 R3414 + R1242 R3413 + R1234 R3412 dVM , etc.

(c) We can choose the positively oriented local orthonormal frame (e1 , . . . , em ) so that it
diagonalizes SM at a given point x ∈ M . Then the eigenvalues of SM at x are called the
principal curvatures at x and are denoted by κ1 (x), . . . , κm (x). Then

P f(−Θ) = ρdVM , ρ ∈ C ∞ (M ),

where
m
Y
ρ(x) = (2h − 1)!! κi (x), ∀x ∈ M. ⊔
⊓
k=1

9.2.3 Gauss–Bonnet theorem for domains in an Euclidean space

Suppose D is a relatively compact open subset of an Euclidean space Rm+1 with smooth
boundary ∂D. We denote by n the outer normal vector field along the boundary. It
defines an oriented Gauss map
~GD : ∂D → S m .

We denote by dAm the area form on the unit sphere S m so that

Z
~ 1 ~G∗ dAm .
deg GD =
σ m ∂D D
If m is even then the Gauss–Bonnet theorem for the hypersurface ∂D implies
Z
1 ~G∗ dAm = 1 χ(∂D).
σ m ∂D D 2
Using the Poincaré duality for the oriented manifold with boundary D we deduce χ(∂D) =
2χ(D), so that Z
1 ~G∗ dAm = χ(D), m ∈ 2Z.
σm ∂D D
We want to prove that the above equality holds even when m is odd. Therefore in the
remainder of this section we assume m is odd.
We first describe the integrand ~G∗D dAm . Let SD denote the second fundamental form
of the hypersurface

SD (X, Y ) = n • (D X Y ) = −X • (D Y n), ∀X, Y ∈ Vect (∂D).

9.2. GAUSS–BONNET AGAIN?!? 409

We deduce
1 ~∗ 1
GD dAm = det(−SD )dV∂D ,
σm σm
where dV∂D denotes the volume form on ∂D.
A smooth vector field on D̄,
X : D̄ → Rm+1
is called admissible if along the boundary it points towards the exterior of D,

X • n > 0, on ∂D.

For an admissible vector field X we define

1
X̄ : ∂D → S m , X̄(p) := X(p), ∀p ∈ ∂D.
|X(p)|

Let us observe that the map X̄ is homotopic to the map ~GD . Indeed, for t ∈ [0, 1], define

1
Yt : ∂D → S m , Yt (p) = ((1 − t)n + tX̄
|(1 − t)n + tX̄|

Observe that this map is well defined since

|(1 − t)n + tX̄|2 = t2 + (1 − t)2 + 2t(1 − t)(n • X̄) > 0.

Hence
deg ~GD = deg X̄,
for any admissible vector field X.
Suppose X is a nondegenerate admissible vector field. This means that X has a finite
number of stationary points,

ZX = {p1 , . . . , pν }, X(pi ) = 0,

and all of them are nondegenerate, i.e., for any p ∈ ZX , the linear map

AX,p : Tp Rm+1 → Tp Rm+1 , Tp Rm+1 ∋ v 7→ (D v X)(p)

is invertible. Define
ǫX : ZX → {±1}, ǫ(p) = sign det AX,p .
For any ε > 0 sufficiently small the closed balls of radius ε centered at the points in ZX
are disjoint. Set [
Dε = D \ Bε (p).
p∈ZX

The vector field X does not vanish on Dε , and we obtain a map

1
X̄ : D̄ε → S m−1 , X̄ = X.
|X|
410 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Set
1 ∗
Ω := X̄ dAm .
σm
Observe that
1 ∗
dΩ = X̄ d(dAm ) = 0 on Dε .
σm
Stokes’ theorem then implies that
Z Z Z X Z
Ω= ~
dΩ = 0 =⇒ deg GD = Ω= Ω,
∂Dε Dε ∂D p∈ZX ∂Bε (p)

where the spheres ∂Bε (p) are oriented as boundaries of the balls Bε (p). If we let ε → 0
we deduce X
deg ~GD = ǫX (p), (9.2.8)
p∈ZX

for any nondegenerate admissible vector field X.

To give an interpretation of the right-hand side of the above equality, consider the
double of D. This is the smooth manifold D b obtained by gluing D along ∂D to a copy of
itself equipped with the opposite orientation,
b = D ∪∂D (−D).
D

The manifold without boundary D b is equipped with an orientation reversing involution

ϕ : D̂ → D̂ whose fixed point set is ∂D. In particular, along ∂D ⊂ D̂ we have a ϕ-invariant
decomposition
b ∂D = T ∂D ⊕ L,
T D|
where L is a real line bundle along which the differential of ϕ acts as −1L . The normal
vector field n defines a basis of L. If X is a vector field on D which is equal to n along
∂D, then we obtain a vector field X b on D̂ by setting
(
b X on D
X :=
−ϕ∗ (X) on −D.

b where the nondegeneracy of X

If X is nondegenerate, then so is X, b is defined in terms of
an arbitrary connection on T D̂. More precisely, if ∇ is a connection on T D̂, and q ∈ ZX̂ ,
then q is nondegenerate if the map
b b b
b : Tq D → Tq D, v 7→ ∇v X
AX,q

is an isomorphism. This map is independent of the connection ∇, and we denote by ǫX̂ (q)
the sign of its determinant. Moreover

ZXb = ZX ∪ ϕ(ZX ),

and, because m is odd, the map

ǫXb : ZXb → {±1}
9.3. CURVATURE MEASURES 411

satisfies
ǫX (p) = ǫXb (ϕ(p)).
Hence
X X (9.2.8)
ǫXb (q) = 2 ǫX (q) = 2 deg ~GD .
q∈ZX
b p∈ZX

On the other hand, the general Poincaré-Hopf theorem (see Corollary 7.3.48) implies that
X
ǫX̂ (q) = χ(D̂).
q∈ZX̂

Using the Mayer–Vietoris theorem we deduce

b = 2χ(D) − χ(∂D).
χ(D)
Since ∂D is odd dimensional and oriented we deduce that χ(∂D) = 0, and therefore
X
2χ(D) = χ(D)b =2 ǫX (q) = 2 deg ~GD .
p∈ZX

We have thus proved the following result.

Theorem 9.2.5 (Gauss–Bonnet for domains). Suppose D is a relatively compact open
subset of Rm+1 with smooth boundary ∂D. We denote by ~GD the oriented Gauss map
~GD : ∂D → S m , ∂D ∋ p 7→ n(p) = unit outer normal,

and by SD the second fundamental form of ∂D,

SD (X, Y ) = n • (D X Y ), ∀X, Y ∈ Vect (∂D).
Then Z
1
det(−SD )dV∂D = deg ~GD = χ(D). ⊔
⊓
σm ∂D

9.3 Curvature measures

In this section we introduce the main characters of classical integral geometry, the so called
curvature measures, and we describe several probabilistic interpretations of them known
under the common name of Crofton formulæ.
We will introduce these quantities trough the back door, via the beautiful tube formula
of Weyl, which describes the volume of a tube of small radius r around a compact sub-
manifold of an Euclidean space as a polynomial in r whose coefficients are these geometric
measures. Surprisingly, these coefficients depend only on the Riemann curvature of the
induced metric on the submanifold.
The Crofton Formulæ state that these curvature measures can be computed by slicing
the submanifold with affine subspaces of various codimensions, and averaging the Euler
characteristics of such slices. To keep the geometric ideas as transparent as possible, and
the analytical machinery to a minimum, we chose to describe these facts in the slightly
more restrictive context of tame geometry.
412 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

9.3.1 Tame geometry

The category of tame spaces and tame maps is sufficiently large to include all the compact
triangulable spaces, yet sufficiently restrictive to rule out pathological situations such as
Cantor sets, Hawaiian rings, or nasty functions such as sin(1/t).
We believe that the subject of tame geometry is one mathematical gem which should
be familiar to a larger geometric audience, but since this is not yet the case, we devote
this section to a brief introduction to this topic. Unavoidably, we will have to omit many
interesting details and contributions, but we refer to [31, 36, 37] for more systematic
presentations. For every set X we will denote by P(X) the collection of all subsets of X.
An R-structure 2 is a collection S = {Sn }n≥1 , Sn ⊂ P(Rn ), with the following properties.

E1 . Sn contains all the real algebraic subsets of Rn , i.e., the subsets described by finitely
many polynomial equations.

E2 . Sn contains all the closed affine half-spaces of Rn .

P1 . Sn is closed under boolean operations, ∪, ∩ and complement.

P2 . If A ∈ Sm , and B ∈ Sn , then A × B ∈ Sm+n .

P3 . If A ∈ Sm , and T : Rm → Rn is an affine map, then T (A) ∈ Sn .

Example 9.3.1. (Semialgebraic sets). Denote by Snalg the collection of semialgebraic

subsets of Rn , i.e., the subsets S ⊂ Rn which are finite unions

S = S1 ∪ · · · ∪ Sν ,

where each Sj is described by finitely many polynomial inequalities. The Tarski-Seidenberg

theorem (see [20]) states that Salg is a structure.
To appreciate the strength of this theorem we want to discuss one of its many conse-
quences. Consider the real algebraic set

Zn := (x, a0 , . . . , an−1 ) ∈ Rn+1 ; a0 + a1 x + · · · an−1 xn−1 + xn = 0 .

The Tarski-Seidenberg theorem implies that the projection of Zn on the subspace with
coordinates (ai ) is a semialgebraic set

Rn := ~a = (a0 , . . . , an−1 ) ∈ Rn ; ∃x ∈ R, P~a (x) = 0 ,

where
P~a (x) = a0 + a1 x + · · · an−1 xn−1 + xn .
In other words, the polynomial P~a has a real root if and only if the coefficients ~a satisfy at
least one system of polynomial inequalities from a finite, universal, collection of systems
of polynomial inequalities.
2
This is a highly condensed and special version of the traditional definition of structure. The model
theoretic definition allows for ordered fields, other than R, such as extensions of R by “infinitesimals”.
This can come in handy even if one is interested only in the field R.
9.3. CURVATURE MEASURES 413

For example, when n = 2, the resolvent set R2 is described by the well known inequality
a21
− 4a0 ≥ 0.
When n = 3, we can obtain a similar conclusion using Cardano’s formulæ. For n ≥ 5,
we know that there cannot exist any algebraic formulæ describing the roots of a degree n
polynomial, yet we can find algebraic inequalities which can decide if such a polynomial
has at least one real root. ⊔
⊓

Suppose S is a structure. We say that a set is S-definable if it belongs to one of the

Sn ’s. If A, B are S-definable then a function f : A → B is called S-definable if its graph

Γf := (a, b) ∈ A × B; b = f (a)

is S-definable. The reason these sets are called definable has to do with mathematical
logic.
A formula 3 is a property defining a certain set. For example, the two different looking
formulas  
x ∈ R; x ≥ 0}, x ∈ R; ∃y ∈ R : x = y 2 },
describe the same set, [0, ∞).
Given a collection of formulas, we can obtain new formulas, using the logical operations
∧, ∨, ¬, and quantifiers ∃, ∀. If we start with a collection of formulas, each describing
an S-definable set, then any formula obtained from them by applying the above logical
transformations will describe a definable set.
To see this, observe that the operators ∧, ∨, ¬ correspond to the boolean operations,
∩, ∪, and taking the complement. The existential quantifier corresponds to taking a
projection. For example, suppose we are given a formula φ(a, b), (a, b) ∈ A × B, A, B
definable, describing a definable set C ⊂ A × B. Then the formula

a ∈ A; ∃b ∈ B : φ(a, b)

describes the image of the subset C ⊂ A × B via the canonical projection A × B → A. If

A ⊂ Rm , B ⊂ Rn , then the projection A × B → A is the restriction to A × B of the linear
projection Rm × Rn → Rm and P3 implies that the image of C is also definable. Observe
that the universal quantifier can be replaced with the operator ¬∃¬.
f g
Example 9.3.2. (a) The composition of two definable functions A → B → C is a definable
function because

Γg◦f = (a, c) ∈ A × C; ∃b ∈ B : (a, b) ∈ Γf , (b, c) ∈ Γg .

Note that any polynomial with real coefficients is a definable function.

(b) The image and the preimage of a definable set via a definable function is a definable
set.
(c) Observe that Salg is contained in any structure S. In particular, the Euclidean norm
X
n 1/2
n
| • | : R → R, |(x1 , . . . , xn )| = x2i
i=1
3
We are deliberately vague on the meaning of formula.
414 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

is S-definable.
Observe that any Grassmannian Grk (Rn ) is a semialgebraic subset of the Euclidean
space End+ (R) of symmetric operators Rn → Rn because it is defined by the system of
algebraic (in)equalities

Grk (Rn ) = P ∈ End+ (Rn ); P 2 = P, Λk P 6= 0, Λm P = 0, ∀m > k ,

where Λj P denotes the endomorphism of Λj Rn induced by the linear map P ∈ End(Rn ).

In the above description, to the orthogonal projection P of rank k, we associate its range,
which is a vector subspace of dimension k.
Similarly, the affine Grassmannian Graff k (Rn ) is a semialgebraic set because its is
described by 
Graff k (Rn ) = (x, P ) ∈ Rn × Grk (Rn ); P x = 0 .
In the above description, to the pair (x, P ) we associate the affine plane x + Range (P ).
(d) Any compact, affine simplicial complex K ⊂ Rn is S-definable because it is a finite
union of affine images of finite intersections of half-spaces.
(e) Suppose A ⊂ Rn is S-definable. Then its closure cl(A) is described by the formula

x ∈ Rn ; ∀ε > 0, ∃a ∈ A : |x − a| < ε ,

and we deduce that cl(A) is also S-definable. Let us examine the correspondence between
the operations on formulas and operations on sets on this example.
We rewrite this formula as
 
∀ε (ε > 0) ⇒ ∃a(a ∈ A) ∧ (x ∈ Rn ) ∧ (|x − a| < ε) .

In the above formula we see one free variable x, and the set described by this formula
consists of those x for which that formula is a true statement.
The above formula is made of the “atomic” formulæ,

(a ∈ A), (x ∈ Rn ), (|x − a| < ε), (ε > 0),

which all describe definable sets. The logical connector ⇒ can be replaced by ∨¬. Finally,
we can replace the universal quantifier to rewrite the formula as a transform of atomic
formulas via the basic logical operations.
n  o
¬ ∃ε¬ (ε > 0) ⇒ ∃a(a ∈ A) ∧ (x ∈ Rn ) ∧ (|x − a| < ε) .

Arguing in a similar fashion we deduce that the interior of an S-definable set is an S-

definable set. ⊔
⊓

Given an R-structure S, and a collection A = (An )n≥1 , An ⊂ P(Rn ), we can form a

new structure S(A), which is the smallest structure containing S and the sets in An . We
say that S(A) is obtained from S by adjoining the collection A.
Definition 9.3.3. An R-structure S is called tame, or o-minimal (order minimal) if it
satisfies the property
9.3. CURVATURE MEASURES 415

T. Any set A ∈ S1 is a finite union of open intervals (a, b), −∞ ≤ a < b ≤ ∞, and
singletons {r}. ⊔
⊓

Example 9.3.4. (a) The collection of real semialgebraic sets Salg is a tame structure.
(b)(Gabrielov-Hironaka-Hardt) A restricted real analytic function is a function f : Rn → R
with the property that there exists a real analytic function f˜ defined in an open neigh-
borhood U of the cube Cn := [−1, 1]n such that
(
f˜(x) x ∈ Cn
f (x) =
0 x ∈ R n \ Cn .

we denote by San the structure obtained from Salg by adjoining the graphs of all the
restricted real analytic functions. For example, all the compact, real analytic submanifolds
of Rn belong to San . The structure San is tame.
(c)(Wilkie, van den Dries, Macintyre, Marker ) The structure obtained by adjoining to
San the graph of the exponential function R → R, t 7→ et , is a tame structure.
(d)(Khovanski-Speissegger ) There exists a tame structure S′ with the following properties
(d1 ) San ⊂ S′ .

(d2 ) If U ⊂ Rn is open, connected and S′ -definable, F1 , . . . , Fn : U × R → R are S′ -

definable and C 1 , and f : U → R is a C 1 function satisfying
∂f
= Fi (x, f (x)), ∀x ∈ R, , i = 1, . . . , n, (9.3.1)
∂xi
then f is S′ -definable.
The smallest structure satisfying the above two properties, is called the pfaffian closure 4
of San , and we will denote it by b
San .
Observe that if f : (a, b) → R is C 1 , b
San -definable, and x0 ∈ (a, b) then the antideriva-
tive F : (a, b) → R Z x
F (x) = f (t)dt, x ∈ (a, b),
x0

is also b
San -definable. ⊔
⊓

The definable sets and functions of a tame structure have rather remarkable tame
behavior which prohibits many pathologies. It is perhaps instructive to give an example
of function which is not definable in any tame structure. For example, the function
x 7→ sin x is not definable in a tame structure because the intersection of its graph with
the horizontal axis is the countable set πZ which violates the o-minimality condition T.
We will list below some of the nice properties of the sets and function definable in a
tame structure S. Their proofs can be found in [31, 36].
• (Piecewise smoothness of tame functions and sets.) Suppose A is an S-definable set, p is
a positive integer, and f : A → R is a definable function. Then A can be partitioned into
4
Our definition of pfaffian closure is more restrictive than the original one in [78, 122], but it suffices
for many geometric applications.
416 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

finitely many S definable sets S1 , . . . , Sk , such that each Si is a C p -manifold, and each of
the restrictions f |Si is a C p -function. The dimension of A is then defined as max dim Si .
• (Dimension of the boundary.) If A is an S-definable set, then dim(cl(A) \ A) < dim A.
• (Closed graph theorem.) Suppose X is a tame set and f : X → Rn is a tame bounded
function. Then f is continuous if and only if its graph is closed in X × Rn .
• (Curve selection.) If A is an S-definable set, k > 0 an integer, and x ∈ cl(A) \ A, then
there exists an S-definable C k -map γ : (0, 1) → A such that x = limt→0 γ(t).
• (Triangulability.) For every compact definable set A, and any finite collection of defin-
able subsets {S1 , . . . , Sk }, there exists a compact simplicial complex K, and a definable
homeomorphism
Φ:K→A
such that all the sets Φ−1 (Si ) are unions of relative interiors of faces of K.
• Any definable set has finitely many connected components, and each of them is definable.
• (Definable selection.) Suppose A, Λ are S-definable. Then a definable family of subsets
of A parameterized by Λ is a subset

S ⊂ A × Λ.

We set 
Sλ := a ∈ A; (a, λ) ∈ S ,
and we denote by ΛS the projection of S on Λ. Then there exists a definable function
s : ΛS → S such that
s(λ) ∈ Sλ , ∀λ ∈ ΛS .
• (Definability of dimension.) If (Sλ )λ∈Λ is a definable family of definable sets, then the
function
Λ ∋ λ 7→ dim Sλ ∈ R
is definable. In particular, its range must be a finite subset of Z.
• (Definability of Euler characteristic.) Suppose (Sλ )λ∈Λ is a definable family of compact
tame sets. Then the map

Λ ∋ λ 7→ χ(Sλ ) = the Euler characteristic of Sλ ∈ Z


is definable. In particular, the set χ(Sλ ); λ ∈ Λ, ⊂ Z is finite.
•(Scissor equivalence.) If A and B are two compact definable sets, then there exists a
definable bijection ϕ : A → B if and only if A and B have the same dimensions and the
same Euler characteristics. (The map ϕ need not be continuous.)
• (Definable triviality of tame maps.) We say that a tame map Φ : X → S is definably
trivial if there exists a definable set F , and a definable homeomorphism τ : X → F × S
such that the diagram below is commutative

X [
τ
w S×F
[℄ π
Φ  S
.
S
9.3. CURVATURE MEASURES 417

If Ψ : X → Y is a definable map, and p is a positive integer, then there exists a partition

of Y into definable C p -manifolds Y1 , . . . , Yk such that each the restrictions

Ψ : Ψ−1 (Yk ) → Yk

is definably trivial. ⊔
⊓

Definition 9.3.5. A subset A of some Euclidean space Rn is called tame if it is definable

within a tame structure S. ⊔
⊓

Exercise 9.3.6. Suppose S is a tame structure and ϕ : [0, 1] → R2 is an S-definable map.

(a) Prove that ϕ is differentiable on the complement of a finite subset of [0, 1].
(b) Prove that the set

(t, m) ∈ [0, 1] × R2 ; ϕ is differentiable at t and m = ϕ′ (t)

is S-definable.
(c) Prove that the curve 
C = (t, ϕ(t) ); t ∈ [0, 1]
has finitely many components, and each of them has finite length. ⊔
⊓

9.3.2 Invariants of the orthogonal group

In the proof of the tube formula we will need to use H.Weyl’s characterization of polyno-
mials invariant under the orthogonal group.
Suppose V is a finite dimensional Euclidean space with metric (−, −). We denote by
h−, −i the canonical pairing

h−, −i : V ∗ × V → R, hλ, vi = λ(v), ∀v ∈ V, λ ∈ V ∗ = Hom(V, R).

We denote by O(V ) the group of orthogonal transformations of the Euclidean space V .

Recall that an O(V )-module is a pair (E, ρ), where E is a finite dimensional real vector
space, while ρ is a group morphism

ρ : O(V ) → Aut(E), g 7→ ρ(g).

A morphism of O(V )-modules (Ei , ρi ), i = 0, 1, is a linear map A : E0 → E1 such that for

every g ∈ O(V ) the diagram below is commutative

E0
A
w E1
ρ0 (g)
u u ρ (g)
1

E0
A
w E1
We will denote by HomO(V ) (E0 , E1 ) the spaces of morphisms of O(V )-modules.
The vector space V has a tautological structure of O(V )-module given by

τ : O(V ) → Aut(V ), τ (g)v = gv, ∀g ∈ O(V ), v ∈ V.

418 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

It induces a structure of O(V )-module on V ∗ = Hom(V, R) given by

ρ† : O(V ) → Aut(V ∗ ), g 7→ ρ† (g),

where
hρ† (g)λ, vi = hλ, g −1 vi, ∀λ ∈ V ∗ , v ∈ V.
Equivalently, ρ† (g) = ρ(g −1 )† , where ρ(h)† denotes the transpose of ρ(h). We obtain an
action on (V ∗ )⊗n ,


(ρ† )⊗n : O(V ) → Aut (V ∗ )⊗n , g 7→ ρ† (g)⊗n .

We denote by (V ∗ )⊗n
O(V ) the subspace consisting of invariant tensors,

⊗n
ω ∈ (V ∗ )⊗n
O(V ) ⇐⇒ ρ† (g) ω = ω, ∀g ∈ O(V ).

Observe that (V ∗ )⊗n can be identified with the vector space of multi-linear maps

ω :Vn =V · · × V} → R,
| × ·{z
n

so that (V ∗ )⊗nO(V )
can be identified with the subspace of O(V )-invariant multilinear maps
n
V → R.
Hermann Weyl has produced in his classic monograph [133] an explicit description
of (V ∗ )⊗n
O(V ) . We would like to present here, without proof, this beautiful result of Weyl
since it will play an important role in the future. We follow the elegant and more modern
presentation in Appendix I of [8] to which we refer for proofs.
Observe first that the metric duality defines a natural isomorphism of vector spaces

D : V → V ∗ , v 7→ v † ,

defined by
hv † , ui = (v, u), ∀u, v ∈ V.
This isomorphism induces an isomorphism of O(V )-modules

D : (V, ρ) → (V ∗ , ρ† ).

We conclude that for ever nonnegative integers r, s we have isomorphisms of G-modules

(V ∗ )⊗(r+s) ∼
= (V ∗ ⊗r) ⊗ V ⊗s ∼
= Hom(V ⊗r , V ⊗s )

In particular,



(V ∗ )⊗(r+s) ∼
= Hom(V ⊗r , V ⊗s ) = HomO(V ) (V ⊗r , V ⊗s ).
O(V ) O(V )

Let us observe that if we denote by Sr the group of permutations of {1, . . . , r}, then for
every ϕ ∈ Sr we obtain a morphism of O(V )-modules

Tφ ∈ HomO(V ) (V ⊗r , V ⊗r ), Tϕ (v1 ⊗ · · · ⊗ vr ) = vϕ(1) ⊗ · · · ⊗ vϕ(r) .

9.3. CURVATURE MEASURES 419

Weyl’s First Main Theorem of Invariant Theory states that

HomO(V ) (V ⊗r , V ⊗s ) 6= 0 ⇐⇒ r = s,

and that nX o
HomO(V ) (V ⊗r , V ⊗r ) = R[Sr ] := cϕ Tϕ ; cϕ ∈ R, .
ϕ∈Sr

We can translate this result in terms of invariant multi-linear forms. Thus

(V ∗ )⊗n
O(V ) 6= 0 ⇐⇒ n = 2r, r ∈ Z≥0 ,

and (V ∗ )⊗2r
O(V ) is spanned by the multilinear forms

Pϕ : V 2r → R, (ϕ ∈ Sr ),

defined by


Pϕ (u1 , . . . , ur , v 1 , . . . , v r ) = u1 , v ϕ(1) · · · ur , v ϕ(r) .
The above result has an immediate consequence. Suppose we have a map

· · × V} → R, (v 1 , . . . , v n ) 7→ f (v 1 , . . . , v n ),
| × ·{z
f :V
n

which is a homogeneous polynomial of degree di in the variable v i , ∀i = 1, . . . , n. This

form determines a multilinear form

Polf : V d1 × · · · × V dn → R

obtained by polarization in each variables separately,

Polf (u11 , . . . , ud11 ; . . . ; v 1n , . . . , v dnn )

= the coefficient of the monomial t11 t12 · · · t1d1 · · · tn1 · · · tndn in the polynomial
X
d1 dn
X 
Pf (t11 , t12 , . . . , t1d1 , . . . , tn1 , . . . , tndn ) = f tij uj1 , . . . , tnj ujn .
j=1 j=1

Observe that
f (v 1 , . . . , v n ) = Polf (v 1 , . . . , v 1 , . . . , v n , . . . , v n ),
| {z } | {z }
d1 dn

and f is O(V )-invariant if and only if Polf is O(V ) invariant.

Note that every function
f :V · · × V} → R
| × ·{z
n
which is polynomial in each of the variables is a linear combination of functions which is
polynomial and homogeneous in each of the variables. For every 1 ≤ i ≤ j ≤ n we define

· · × V} → R, qij (v 1 , . . . , v n ) := (v i , v j ).
| × ·{z
qij : V
n
420 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Theorem 9.3.7 (Weyl). If f : V × · · · × V → R is a polynomial
map then f is O(V )-

n+1
invariant if and only if there exists a polynomial P in the 2 variables qij such that

f (v1 , . . . , v n ) = P ( qij (v 1 , . . . , v n )1≤i≤j≤n ). ⊔

⊓

Example 9.3.8. (a) Consider the space E = V ⊗k . Observe that a degree n homogeneous
polynomial P on E can by identified with an element in the symmetric tensor product

Symd (E ∗ ) ⊂ (V ∗ )⊗2kn .

The polynomial P is called a degree n orthogonal invariant of tensors T ∈ V ⊗k if it is

invariant as an element of (V ∗ )⊗kn . For example, Weyl’s theorem implies that the only
degree 1 invariant of a tensor
X
T = Tij ei ⊗ ej ∈ V ⊗2
i,j

is the trace X X
tr(T ) = Ti,j ei ⊗ ej = Tii .
i,j i

The space of degree 2 invariants is spanned by the polynomials

X X
(tr(T ))2 , Q(T ) = Tij2 , Q̃(T ) = Tij Tji . ⊔
⊓
ij i,j

Here is briefly how to find a basis of the space of degree k invariant polynomials

P : V ⊗h → R.

First of all, such polynomials exist if and only if hk is even, hk = 2m. Fix an orthonormal
basis { e1 , · · · , ed } of V . A tensor T ∈ V ⊗h decomposes as
X
T = Ti1 ...ih ei1 ⊗ · · · ⊗ eih .
1≤i1 ,...,ih ≤d

A degree k polynomial P : V ⊗h → R is then a polynomial in the dh variables Ti1 ...ih .

We will construct a bijection between a basis of invariant polynomials and certain
combinatorial structures called matchings.
A matching on the set Ihk := {1, . . . , hk} is an equivalence relation ∼ with the property
that each equivalence class consists of two elements. Thus to produce a matching we need
to choose a subset A ⊂ Ihk of cardinality m = hk 2 , and then a bijection ϕ : A → Ihk \ A.
Observing that the matching associated to (A, ϕ) is the same as the matching associated
to (Ihk \ A, ϕ−1 ), we deduce that the number of matchings is
 
1 2m (2m)!
(m!) · = .
2 m 2(m!)
9.3. CURVATURE MEASURES 421

Denote by F∼ the set of functions

µ : Ihk → {1, 2, . . . , d}

such that i ∼ j =⇒ µ(i) = µ(j). Equivalently, F∼ , can be identified with the set of
functions µ : Ihk / ∼→ {1, . . . , d}, so there are dm such functions.
For any matching ∼ define
X
P∼ (T ) = Tµ(1)...µ(h) · · · Tµ(hk−h+1)...µ(hk) .
µ∈F∼

The collection 
P∼ ; ∼ is a matching of Ihk
is a basis of the space of degree k invariant polynomials V ⊗h → R.
For example, the space of degree 1 invariant polynomials in tensors of order 4 has
4!
dimension 2(2!) = 6. Each polynomial in a basis constructed as above is a sum of d2
terms. We see that things are getting “hairy” pretty fast. ⊔
⊓

9.3.3 The tube formula and curvature measures

Suppose that M is an m-dimensional submanifold of Rn . We set

c := codim M = n − m.

In this section we will assume that M is compact and without boundary, but we will not
assume that it is orientable.
For r > 0 we define the tube of radius r around M to be the closed set

Tr (M ) := x ∈ M ; dist (x, M ) ≤ r ,

and we denote by V (M, r) its volume.

Note that a first approximation for V (M, r) is

V (M, r) ≈ vol (M ) · ω c r c ,

where ω c r c is the volume of a ball of dimension c and radius r.

We want to prove that there exists a polynomial

PM (r) = pc r c + · · · + pn r n ,

such that, for all sufficiently small r, we have

• V (M, r) = PM (r),

• pc = ω c · vol (M ),

• all the coefficients of PM are described, up to some universal multiplicative constants,

by certain integral, intrinsic geometric invariants of M .
422 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Let N(M ) denote the orthogonal complement of T M in (T Rn )|M , and we will call it
the normal bundle of M ֒→ Rn . We define

Dr (Rn ) := {(v, p); p ∈ Rn , v ∈ Tp Rn , |v| ≤ r ⊂ T Rn ,

and we set
Nr (M ) := N(M ) ∩ Dr (Rn ).
The manifold with boundary Dr (Rn ) is a bundle of n-dimensional disks over Rn , while
Nr (M ) is a bundle of c-dimensional disks over M .
The exponential map E : T Rn → Rn restricts to an exponential map

EM : N(M ) → Rn .

Observe that because M is compact, there exists r0 = r0 (M ) > 0 such that, for every
r ∈ (0, r0 ), the exponential map EM induces a diffeomorphism

EM : Nr (M ) → Tr (M ).

If we denote by |dVn | the Euclidean volume density on Rn we deduce

Z Z


V (M, r) = vol Tr (M ) = |dVn | = E∗M |dVn |.
Tr (M ) Nr (M )

If π : Nr (M ) → M denotes the canonical projection, then we deduce from Fubini’s theorem

that Z
V (M, r) = π∗ E∗M |dVn |. (9.3.2)
M
We want to give a more explicit description of the density π∗ E∗M |dVn |. We will continue
to use the indexing conventions we have used in Subsection 9.2.1.
Fix a local orthonormal frame (eA ) of (T Rn )|M defined in a neighborhood U ⊂ M of
a point p0 ∈ M such that for all 1 ≤ i ≤ m vector field ei is tangent to U . We define
 X
Drc := ~t = (tα ) = (tm+1 , . . . , tn ) ∈ Rc ; |tα |2 ≤ r .
α

Note that we have a diffeomorphism

Drc × U −→ Nr (U ) := Nr (M )|U , (~t, x) 7→ (tα eα (x), x) ∈ Nr (M ),

and thus we can identify Drc × U with the open subset π −1 (U ) ⊂ Nr (M ), and we can use
x ∈ M and ~t ∈ Drc as local coordinates on π −1 (U ). Define

Tr (U ) := EM ( Nr (U ) ) ⊂ Rn ,

and
ẽA : Tr (U ) → Rn by ẽA (x + tα eα ) = eA (x).
We have thus extended in a special way the local frame (eA ) of (T Rn )|U to a local frame
of (T Rn )|Tr (U ) so that
D ẽα ẽA = 0, ∀α, A. (9.3.3)
9.3. CURVATURE MEASURES 423

We denote by (θ A ) the coframe of Tr (U ) dual to ẽA .

Over Drc × U we have a local frame (∂tα , ei ) with dual coframe (φA ) defined by

φi = π ∗ θ i , φα = dtα .


Consider the 1-forms ΘA 1
B ∈ Ω Tr (U ) associated to the Levi–Civita connection D by
the frame (ẽA ) on Tr (U ), and set

ΘA
CB = ẽC ΘA
B , ∀i,

so that
D eC eB = ΘA
CB eA .

Using (9.3.3) we deduce

ΘA A A i
αB = 0, ∀α =⇒ ΘB = ΘiB θ . (9.3.4)
Finally set

ΦA ∗ A 1 A ∗ A ∞
B = π (ΘB |M ) ∈ Ω (Nr (U )), ΦiB := π (ΘiB |U ) ∈ C (Nr (U ) ).

The equalities (9.3.4) imply

ΦA A i
B = ΦiB φ .

On Nr (M )|U we use (~t, x) as coordinates and we have

EM (tα eα (x), x) = x + tα eα .

We have X X
E∗M θ A = (eA • D ei EM )φi + (eA • ∂tα EM )dtα
i α
X X
= eA • (ei + tα Dei eα )φi + δAα dtα = δAi φi + tα ΦA i α
iα φ + δAα dt .
i α

Hence X
E∗M θ j = φj + tα Φjiα φi = φj − tα Φαij φi , E∗M θ β = dtβ .
α

We find it convenient to set

Φij = (Φm+1
ij , . . . , Φnij ) : U → Rc ,

so that X
E∗M θ j = φj − (~t • Φij )φi .
i
Define the m × m symmetric matrix


S := S(~t, x) = ~t • Φij (x) 1≤i,j≤m

Note that the volume density on Rn is

|dVn | = |θ m+1 ∧ · · · ∧ θ n ∧ θ 1 ∧ · · · ∧ θ m .|
424 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

E∗M dVn = | det(1 − S(~t, x) | |d~t ∧ dφ| = det(1 − S(~t, x) |d~t ∧ dφ|, (9.3.5)
d~t = dtm+1 ∧ · · · ∧ dtn , dφ = dφ1 ∧ · · · ∧ dφm .
Recalling that | ∧i dθ i |M | is the metric volume density |dVM | on M , we deduce


E∗M |dVn | = det(1 − S(~t, x) |d~t| × π ∗ |dVM |,

where |d~t| denotes the volume density on Rc . For simplicity we write |dVM | instead of
π ∗ |dVM |. Now set
1
ρ := |~t|, ω := ~t,
ρ
and denote by |dω| the area density on the unit sphere in Rc . Then


E∗M |dVn | = det(1 − ρS(ω, x) ρc−1 |dρ| × |dω| × |dVM |. (9.3.6)

Observe that
m
X


det(1 − ρS(ω, x) = (−1)ν ρν Pν Φij (x) • ω ,
ν=0

where Pν denotes a homogeneous polynomial of degree ν in the m2 variables

uij ∈ Rc , 1 ≤ i, j ≤ m.

We set Z



P ν Φij (x) := Pν Φij (x) • ω |dω|.
S c−1


Above, P ν uij is an O(c)-invariant, homogeneous polynomial of degree ν in the variables
uij ∈ Rc , 1 ≤ i, j ≤ m. We conclude,
m
X (−1)ν

π∗ E∗M |dVn | = r c+ν P ν Φij (x) |dVM (x)|. (9.3.7)
c+ν
ν=0


We would like to determine the invariant polynomials P ν uij .
Theorem 9.3.7 on invariants of the orthogonal group O(c) implies that P ν must be a
polynomial in the quantities
qi,j,k,ℓ := uik • ujℓ .
Because these quantities are homogeneous of degree 2 in the variables uij we deduce
P ν = 0 if ν is odd. Assume therefore ν = 2h, h ∈ Z≥0 .
For every ~t ∈ Rc = span (em+1 , . . . , em+c ) we form the linear operator

U (~t) = U (uij , ~t) : Rm → Rm

given by the m × m matrix ( uij • ~t )1≤i,j≤m . We deduce that

(−1)ν Pν ( uij • ~t ) = tr Λν U (uij , ~t)

= the sum of all the ν × ν minors of U (~t) symmetric with respect to the diagonal.
9.3. CURVATURE MEASURES 425

These minors are parametrized by the subsets I ⊂ {1, . . . , m} of cardinality #I = ν. For

every ω ∈ Rc we denote by µI ( uij • ω ) the corresponding minor of U (ω), and by µI its
average, Z
µI (uij ) := µI ( uij • ω )dω.
S c−1
Note that µI is an O(c)-invariant polynomial in the variables {uij }i,j∈I .
Let
I = {1 ≤ i1 < i2 < · · · < i2h ≤ m} ⊂ {1, . . . , m},
and denote by SI the group of permutations of I. For ϕ ∈ SI we set

ϕj := ϕ(ij ), ∀j = 1, . . . , 2h.

For any σ, ϕ ∈ SI we denote by ǫ(σ, ϕ) the signature of the permutation σ ◦ ϕ−1 , and by
Qσ,ϕ the invariant polynomial
h
Y h
Y
QI,σ,ϕ = qϕ2j−1 ,ϕ2j ,σ2j−1 ,σ2j = uϕ2j−1 σ2j−1 • uϕ2j σ2j .
j=1 j=1

Lemma 9.3.9. There exists a constant ξ = ξm,ν,c depending only on m, ν and c such that
X
µ̄I = ξQI , QI := ǫ(σ, ϕ)QI,σ,ϕ .
ϕ,σ∈SI

Proof. We regard µI as a function on the vector space of (2h) × (2h) matrices U with
entries in Rc
U = [uij ]i,j∈I .
We observe that µI satisfies the following determinant like properties.

• µI changes sign if we switch two rows (or columns).

• µI is separately linear in each of the variables uij .

• µI is a homogeneous polynomial of degree h in the variables qi,j,k,ℓ.

We deduce that µI is a linear combination of monomials of the form

qk1 ,k2 ,ℓ1 ,ℓ2 · · · qk2h−1 k2h ,ℓ2h−1 ,ℓ2h ,

where
{k1 , . . . , k2h } and {ℓ1 , . . . , ℓ2h }
are permutations of I. The skew-symmetry of µI with respect to the permutations of rows
and columns now implies that µI must be a multiple of QI .
⊔
⊓

The constant ξ satisfies

µI (uij )
ξm,ν,c = , ∀uij ∈ Rc
QI (uij )
426 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

so it suffices to compute the numerator and denominator of the above fraction for some
special values of uij . We can assume I = {1, 2, . . . , 2h} and we choose
 
1
 0 
 
uij =  ..  ∈ Rc .
 . 
0

Then, if we set
 
t1
 t2  1
~t = 
 ..

 ∈ Rc , ω = ~t
 .  |~t|
tc
we deduce  
t1 0 · · · 0
 0 t1 · · · 0 
 
U (uij , ~t) =  .. .. . . ..  , µI (uij • ~t) = |t1 |ν .
 . . . . 
0 0 ··· t1
Hence Z
µI (uij ) = |ω 1 |2h |dω|. (9.3.8)
S c−1

On the other hand, we have

h
Y
QI,σ,ϕ = uϕ2j−1 σ2j−1 • uϕ2j σ2j ,
j=1

which is nonzero if and only if σ = ϕ. We conclude that for this particular choice of uij
we have
QI = (2h)!.

Hence Z
1
ξm,ν,c = |ω 1 |2h |dω|.
(2h)! S c−1

At this point we invoke the following result whose proof is deferred to the end of this
subsection.

Lemma 9.3.10. For any even, nonnegative integers 2h1 , . . . , 2hc we have
Z
2Γ( 2h12+1 ) · · · Γ( 2hc2+1 )
|ω 1 |2h1 · · · |ω c |2hc |dω| = ,
S c−1 Γ( c+2h
2 )

where h = h1 + · · · + hc .
9.3. CURVATURE MEASURES 427

We deduce
2Γ( 2h+1
2 )Γ(1/2)
c−1
ξm,2h,c = , (9.3.9)
(2h)!Γ( c+2h
2 )
and X
P ν (uij ) = ξm,2h,c QI (uij ). (9.3.10)
#I=ν

We denote by S2 the group of permutations of a linearly ordered set with two elements.
We observe that every element

τ = (τ1 , . . . , τh ) ∈ G = S2 × · · · × S2
| {z }
h

defines a permutation of I by regarding τ1 as a permutation of {i1 , i2 }, τ2 as a permutation

of {i3 , i4 } etc. Thus G is naturally a subgroup of SI . The space Sk /G of left cosets of
this group can be identified with the subset S′I ⊂ SI consisting of bijections ϕ : I → I
satisfying the conditions

ϕ1 < ϕ2 , ϕ3 < ϕ4 , . . . , ϕ2h−1 < ϕ2h .

We deduce that if uij = Sij (x), then for every σ ∈ SI we have

X X
ǫ(σ, ϕ)QI,σ,ϕ = ǫ(σ, ϕτ )QI,σ,ϕτ
ϕ∈SI ϕ∈S′I ,τ ∈G

X h
Y

= ǫ(σ, ϕ) qϕ2j−1 ,ϕ2j ,σ2j−1 ,σ2j − qϕ2j ,ϕ2j−1 ,σ2j−1 ,σ2j
ϕ∈S′I j=1

X h
Y
= ǫ(σ, ϕ) Rϕ2j−1 ϕ2j σ2j−1 σ2j
ϕ∈S′I j=1

Using the skew-symmetry Rijkℓ = −Rijℓk we deduce

X X X h
Y
ǫ(σ, ϕ)QI,σ,ϕ = ǫ(σ, ϕ) Rϕ2j−1 ϕ2j σ2j−1 σ2j
σ,ϕ∈SI σ∈SI ϕ∈S′I j=1

X h
Y
= ǫ(στ, ϕ) Rϕ2j−1 ϕ2j σ2j−1 σ2j
σ∈S′I ,τ ∈G j=1

X h
Y
h
=2 ǫ(σ, ϕ) Rϕ2j−1 ϕ2j σ2j−1 σ2j
σ,ϕ∈S′I j=1
| {z }
=:QI (R)

We conclude that X
P ν (ψij ) = 2h ξm,2h,c QI (R). (9.3.11)
#I=2h
428 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Using (9.3.2) and (9.3.7) we deduce that

⌊m/2⌋
X Z
2h ξm,2h,c
V (M, r) = vol (Tr (M )) = ω c+2h r c+2h Qh (R)|dVM |.
(c + 2h)ω c+2h M
h=0

Let us observe that the constant

2h ξm,2h,c
(c + 2h)ω c+2h
is independent of the codimension c. It depends only on h. Indeed, we have

2h ξm,2h,c 2h ξm,2h,c 2h 2Γ( 2h+1

2 )Γ(1/2)
c−1
= =
(c + 2h)ω c+2h σ c+2h−1 σ c+2h−1 (2h)!Γ( c+2h
2 )

2h Γ(h + 1/2) γ(2h) 1

= = h = .
Γ(1/2)1+2h (2h)! π (2h)! (2π)h h!
We have thus obtained the following celebrated result of Hermann Weyl, [132].
Theorem 9.3.11 (Tube formula). Suppose M is a closed, compact submanifold of Rn ,
dim M = m, c = n − m. Denote by R the Riemann curvature of the induced metric on
M . Then for all r > 0 sufficiently small we have
⌊m/2⌋
X
V (M, r) = vol (Tr (M )) = ω c+2h r c+2h µm−2h (M ),
h=0
Z
1
µm−2h (M ) = Qh (R)|dVM |,
(2π)h h! M
where Qh is a polynomial of degree h in the curvature. By choosing a local, orthonormal
frame (e1 , . . . , em ) of T M we can express the polynomial Qh (R) as
X
Qh (R) := QI (R),
#I=2h

where for every I = {i1 < i2 < · · · < i2h } ⊂ {1, . . . , m} we define

X h
Y
QI (R) = ǫ(σ, ϕ) Rϕ(i2j−1 )ϕ(i2j )σ(i2j−1 )σ(i2j ) . ⊔
⊓
σ,ϕ∈S′I j=1

Example 9.3.12. (a) If h = 0 then

µm (M ) = vol (M ).

(b) Assume now that m is even, m = 2h, and oriented. Then c + 2h = m + c = n and
Z
1 1
µ0 (M ) = h
Qh (R)dVM .
(2π) M h!
9.3. CURVATURE MEASURES 429

Comparing the definition of Qh (R) with (9.2.7a) we deduce that the top dimensional form

1 1
Qh (R)dVM ∈ Ω2h (M )
(2π)h h!

is precisely the Euler form associated with the orientation of M and the induced metric,
so that Z
µ0 (M, g) = e(M, g). (9.3.12)
M

(c) Suppose now that M is a hypersurface in Rm+1 . Consider the second fundamental
form
S = (Sij )1≤i,j≤m , Sij = (D ei ej ) • em+1 ,
where we recall that em+1 is in fact the oriented unit normal vector field along M . Fix
a point x0 ∈ M and assume that at this point the frame (e1 , . . . , em ) diagonalizes the
second fundamental form so that
Sij = κi δij .
The eigenvalues κ1 , . . . , κm are the principal curvatures at the point x0 . We denote by
cν (κ) the elementary symmetric polynomial of degree ν in the variables κi . In this case
c = 1, and we have
m
X
E∗M dVRm+1 = det(1 − tS)dt ∧ dVM = (−1)ν tν cν (κ)dt ∧ dVM
ν=0

m Z
X r  ⌊m/2⌋ 2h+1
X r
π∗ E∗M dVRm+1 = tν dt cν (κ)dVM = 2 c2h (κ)dVM
ν=0 −r 2h + 1
h=0

so that
⌊m/2⌋
X X r 2h+1 Z
⌊m/2⌋
ω 1+2h r 1+2h µm−2h (M ) = V (M, r) = 2 c2h (κ)dVM .
2h + 1 M
h=0 h=0

We conclude that
Z
2
µm−2h (M ) = c2h (κ)dVM , σ 2h = (2h + 1)ω 1+2h .
σ 2h M

If M = S m ֒→ Rm+1 is the unit sphere then κi = 1 and we deduce that

 
m σm m
µm−2h (S ) = 2 . (9.3.13)
σ 2h 2h

(d) Using the definition of the scalar curvature we deduce that for any m-dimensional
submanifold M m ֒→ Rn we have
Z
µm−2 (M, g) = constm sg |dVg |,
M
430 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

where s denotes the scalar curvature of the induced metric g, and constm is an universal
constant, depending only on m. We see that the map g → µm−2 (M, g) is precisely the
Hilbert-Einstein functional we discussed in Subsection 4.2.6, Definition 4.2.33.
To find the constm we compute µm−2 (M ) when M = S m . Using (9.3.13) we deduce
  Z
2σ m m
= constm sround |dVS m |,
σ2 2 M

where sround denotes the scalar curvature of the round metric on the unit sphere.
Observe that the Grassmannian of oriented codimension one subspaces of Rm+1 can be
identified with the unit sphere S m . Hence, the oriented Gauss map of the unit sphere S m
is the identity. In particular, the shape operator of S m ֒→ Rm+1 is the identity operator
From Theorema Egregium we deduce that all the sectional curvatures of S n are equal to
one. Using the equalities (4.2.3) we deduce
X X  
m
sround = Rijij = 1=2 .
2
i,j i,j

Hence    
2σ m m m
= 2σ m constm
σ2 2 2
Z
1 1 1
=⇒ constm = = =⇒ µm−2 (M, g) = sg |dVg |.
σ2 4π 4π M
(e) The polynomial Q2 still has a “reasonable form”
X
Q2 (R) = QI
#I=4

Then #S′I = 6 and X

QI = ǫ(σ, ϕ)Rσ1 σ2 ϕ1 ϕ2 Rσ3 σ4 ϕ3 ϕ4
σ,ϕ∈S′I
X X
= Rσ1 σ2 σ1 σ2 Rσ3 σ4 σ3 σ4 + ǫ(σ, ϕ)Rσ1 σ2 ϕ1 ϕ2 Rσ3 σ4 ϕ3 ϕ4 .
σ∈S′I σ6=ϕ∈S′I

The first sum has only three
different monomials, each of them appearing twice is them
sum. The second sum has 62 different monomials (corresponding to subsets of cardinality
2 of S′I ) and each of them appears twice. ⊔
⊓

Definition 9.3.13. If (M, g) is a closed, compact, oriented, Riemann manifold, m =

dim M , and w is nonnegative integer. If m − w is odd we set

µw (M ) = 0.

If m − w is an even, nonnegative integer, m − w = 2h, then we set

Z
1
µw (M, g) = Qh (R)|dVM |.
(2π)h h! M
9.3. CURVATURE MEASURES 431

We will say that µw (M, g) is the weight w curvature measure of (M, g). We set

1
|dµw | := Qh (R)|dVM |,
(2π)h h!

and we will refer to it as the (weight w) curvature density. ⊔

⊓

Remark 9.3.14. (a) Let us observe that for any Riemann manifold M , orientable or not,
the quantities |dµw | are indeed well defined, i.e. independent of the choice of local frames
used in their definition. The fastest way to argue this is by invoking Nash embedding
theorem which implies that any compact manifold is can be isometrically embedded in an
Euclidean space. For submanifolds of Rn , the proof of the tube formula then implies that
these densities are indeed well defined.
We can prove this by more elementary means by observing that, for any finite set I,
the relative signature ǫ(σ, ϕ) of two permutations ϕ, σ : I → I is defined by choosing a
linear ordering on I, but it is independent of this choice.
(b) The weight of the curvature density has a very intuitive meaning. Namely, we should
think of µw (M, g) as a quantity measured in meter w . ⊔
⊓

Proof of Lemma 9.3.10. Consider the integral

Z
~2
I(h1 , . . . , hc ) = e−|t| |t1 |2h1 · · · |tc |2hc |dt|.
Rc

We have e−|~t| = e−|t+1| · · · e−|t

2 2 c |2
so that
c Z
Y ∞  c Z
Y ∞ 
−s2 2hj c 2
I(h1 , . . . , hc ) = e s ds = 2 e−s s2hj ds
j=1 −∞ j=1 0

(u = s2 )
c Z
Y ∞  Yc  2h + 1 
j
= e−u thj −1/2 du = Γ .
0 2
j=1 j=1

On the other hand, using spherical coordinates, ρ = |~t|, ω = |1~t| ~t, and recalling that
h = h1 + · · · + hc , we deduce that
Z Z ∞ 2

1 2h1 c 2hc
I(h1 , . . . , hc ) = |ω | · · · |ω | dω e−ρ ρ2h+c−1 dρ
S c−1 0

(u = ρ2 )
Z Z ∞
1 1 2h1 c 2hc c+2h
= |ω | · · · |ω | dω e−u u 2 −1 du
2 S c−1 0
 Z 
1 c + 2h
= Γ |ω 1 |2h1 · · · |ω c |2hc dω . ⊔
⊓
2 2 S c−1
432 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

9.3.4 Tube formula =⇒ Gauss–Bonnet formula

for arbitrary submanifolds of an Euclidean space
Suppose M m ⊂ Rn is a closed, compact submanifold of Rn . We do not assume that M is
orientable. As usual, set c = n − m, and we denote by g the induced metric on M . For
every sufficiently small positive real number r we set

Mr := {x ∈ Rn ; dist (x, M ) = r} = ∂Tr (M ).

The closed set Mr is a compact hypersurface of Rn , and we denote by gr the induced

metric. Observe that for r and ε sufficiently small we have

Tε (Mr ) = Tr+ε (M ) − Tr−ε (M )

so that,
V (Mr , ε) = V (M, r + ε) − V (M, r − ε),
which implies that X
ω 1+2h ε1+2h µn−1−2h (Mr , gr )
h≥0
X n o
= ω c+2k (r + ε)c+2k − (r − ε)c+2k µn−c−2k (M, g).
k≥0

We deduce
X  
2 c + 2k c−1+2k−2h
µn−1−2h (Mr , gr ) = ω c+2k r µn−c−2k (M, g).
ω 1+2h 1 + 2h
k≥0

We make a change in variables. We set

p := n − 1 − 2h, w := n − c − 2k = m − 2k.

Then c + 2k = n − w, 1 + 2h = n − p, c + 2k − 1 − 2h = p − w, so that we can rewrite the

above formula as
m 
X 
n − w ω n−w p−w
µp (Mr , gr ) = 2 r µw (M, g). (9.3.14)
n − p ω n−p
w=0

In the above equality it is understood that µw (M ) = 0 if m − w is odd. In particular, we

deduce that if the codimension of M is odd then

lim µp (Mr , gr ) = 2µp (M, g), ∀0 ≤ p ≤ m = dim M. (9.3.15)

r→0

If in the formula (9.3.14) we assume that the manifold M is a point, then we deduce that
Mr is the (n − 1)-dimensional sphere of radius r, Mr = Srn−1 , and we conclude that
   
n−1 n ωn p n p
µp (Sr ) = 2 r = 2ω p r , n − p ≡ 1 mod 2, (9.3.16)
p ω n−p p
9.3. CURVATURE MEASURES 433

 
where np is defined by (9.1.20). The last equality agrees with our previous computation
(9.3.13).
If in the formula (9.3.14) we let p = 0 we deduce

µ0 (Mr , gr ) = 2µ0 (M, g), ∀0 < r ≪ 1, if codim M is odd.

Observe that the tube Tr (M ) is naturally oriented, even though the manifold M may not
be orientable. The Gauss–Bonnet theorem for oriented hypersurfaces implies

χ(Mr ) = µ0 (Mr , gr )

so that
1
µ0 (M, g) = χ(Mr ), ∀0 < r ≪ 1. (9.3.17)
2
Theorem 9.3.15 (Gauss–Bonnet). Suppose M is a closed, compact submanifold of an
Euclidean space Rn . Denote by g the induced metric. Then

µ0 (M, g) = χ(M ).

Proof. If m = dim M is odd, then both χ(M ) and µ0 (M ) are equal to zero and the identity
is trivial. Assume therefore that m is even. If the dimension n of the ambient space is
odd, then the Poincaré duality for the oriented n-dimensional manifold with boundary
Tr (M ) implies
χ(Mr ) = χ( ∂Tr (M ) ) = 2χ( Tr (M ) ) = 2χ(M ),
and the theorem follows from (9.3.17).
If n is even, we apply the above argument to the embedding

M ֒→ Rn ֒→ Rn+1 ,

where we observe that the metric induced by the embedding M ֒→ Rn+1 coincides with
the metric induced by the original embedding M ֒→ Rn . ⊔
⊓

Remark 9.3.16. (a) We want emphasize again that in the above theorem we did not
require that M be orientable which is the traditional assumption in the Gauss–Bonnet
theorem.
(b) By invoking the Nash embedding theorem, we deduce that the tube formula implies
the Gauss–Bonnet formula for any compact Riemann manifold, orientable or not. ⊔
⊓

Let us record for later use the following corollary of the above proof.

Corollary 9.3.17. For every closed compact, smooth submanifold M of an Euclidean

space V such that dim V − dim M is odd we have

χ(M ) = 2χ( ∂Tr (M ) ), ∀0 < r ≪ 1. ⊔

⊓
434 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

9.3.5 Curvature measures of domains in an Euclidean space

The second fundamental form of a submanifold is in fact a bilinear form with values in
the normal bundle. If the submanifold happens to be the boundary of a domain, then the
normal bundle admits a canonical trivialization, and the second fundamental form will be
a scalar valued form. The next definition formalizes this observation.
Definition 9.3.18. For any relatively compact open subset D of an Euclidean space V ,
with smooth boundary ∂D, we define the co-oriented second fundamental form of D to
be the symmetric bilinear map

SD : Vect (∂D) × Vect (∂D) → C ∞ (∂D),

SD (X, Y ) = (D X Y ) • n, X, Y ∈ Vect (∂D),

where n : ∂D → V denotes the outer unit normal vector field along ∂D. ⊔
⊓

Suppose D ⊂ Rm+1 is an open, relatively compact subset with smooth boundary

M := ∂D. We denote by n the unit outer normal vector field along M := ∂D, and by
S = SD the co-oriented second fundamental form of D. For every symmetric bilinear form
B on an Euclidean space V we define trj (B) the j-th elementary symmetric polynomial
in the eigenvalues of B, i.e.,
X
z j trj (B) = det(1V + zB).
j≥0

Equivalently,

trj B = tr Λj B : Λj V → Λj V .
We define the tube of radius r around D to be

Tr (D) := x ∈ Rm+1 ; dist (x, D) ≤ r .

We denote by EM the exponential map

EM : (T Rm+1 )|M → Rm+1 ,

(X, p) 7−→ EM (X, p) = p + X, p ∈ M, X ∈ Tp Rm+1 .

For r > 0 we denote by ∆r ⊂ (T Rm+1 )|M the closed set

∆r := (tn(p), p); p ∈ M, t ∈ [0, r] .

For sufficiently small r the map EM defines a diffeomorphism

EM : ∆r → Tr (D) \ D,

so that Z
vol (Tr (D) ) = vol (D) + E∗M dVRm+1 .
∆r
9.3. CURVATURE MEASURES 435

Fix p0 ∈ M and a local, positively oriented local orthonormal frame

(e1 , . . . , em )

of T M defined in a neighborhood U of p0 in M , such that, for every p ∈ U , the collection

(n(p), e1 (p), . . . , em (p))

is a positively oriented, orthonormal frame of Rn . We obtain a dual coframe θ, θ 1 , . . . , θ n .

As in the previous section, the pullback of E∗M dVRm+1 to ∆r has the description


E∗M dVRm+1 = det 1 − tSM dt ∧ dθ 1 ∧ · · · ∧ θ m
 
m
X
= trj (−SM )tj  dt ∧ dθ 1 ∧ · · · ∧ θ m .
j=1

We deduce Z Z 
m
X r j+1
E∗M dVRm+1 = trj (−SM )dVM .
∆r j+1 M
j=0

Define Z 
1
µm−j (D) := trj (−SM )dVM , 0 ≤ j ≤ m,
σj M
and
µm+1 (D) := vol (D)
so that using the equality σ j = (j + 1)ω j+1 we deduce the tube formula for domains,

m+1
X
vol Tr (D) = ω m+1−k r m+1−k µk (D). (9.3.18)
k=0

Theorem 9.2.5 shows that, just as in the case of submanifolds, we have µ0 (D) = χ(D).
Definition 9.3.19. Suppose D is a relatively compact domain with smooth boundary of
an Euclidean space V , dim V = n. Then the curvature densities of D are the densities
|dµj | on ∂D defined by
1
|dµj | := trn−j (−SD )|dV∂D |,
σn−j
where |dV∂D | denotes the volume density on ∂D induced by the Euclidean metric on V ,
and SD denotes the co-oriented second fundamental form of ∂D. ⊔
⊓

We denote by Dm+1
r the ball of radius r in Rm+1 . Then,

Tε (Dm+1
r ) = Dm+1
r+ε ,

so that X
ω m+1 (r + ε)m+1 = ω m+1−k εm+1−k µk (Dm+1
r ).
k≥0
436 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

We conclude    
ω m+1 m + 1 k m+1 k
µk (Dm+1
r )= r = ωk r . (9.3.19)
ω m+1−k k k
Suppose X ֒→ Rm+1 is a closed, compact smooth submanifold. Then for every suffi-
ciently small r > 0, the tube Dr := Tr (X) is a compact domain with smooth boundary
and
Tε (Dr ) = Tr+ε (X).
The tube formula for X implies that
X X
ω j εj µm+1−j (Dr ) = ω k r + ε)k µm+1−k (X).
j≥0 k≥0

We deduce that  
1 X k k−j
µm+1−j (Dr ) = ωk r µm+1−k (X)
ωj j
k≥j

We set n := m + 1 and we make the change in variables

p := n − j, w := n − k.

Then k − j = p − w, and we obtain the following generalization of the tube formula

 

1 X n − w p−w
µp Tr (X) = ω n−w r µw (X)
ω n−p w n−p
X   (9.3.20)
n − w p−w
= ω p−w r µw (X).
w
p−w

In particular, we have


lim µp Tr (X) = µp (X), ∀0 ≤ p ≤ dim X. (9.3.21)
r→0

9.3.6 Crofton formulæ for domains of an Euclidean space

Suppose D is an open, relatively compact subset of the Euclidean space Rn with smooth
boundary M = ∂D. We denote by g the induced metric on M , by Grc the Grassmannian
of linear subspaces of Rn of codimension c, and by Graff c the affine Grassmannian of
codimension c affine subspaces of Rn .
Recall that on Grc we have a natural metric with volume density |dγc | and total volume
Qn−1
j=0 σ j
Vc := Qc−1
Qm−1−c
.
i=0 σ i · j=0 σj

We rescale this volume density as in (9.1.21) to obtain a new volume density |dνc | with
total volume Z  
n
|dνc | = . (9.3.22)
Grc c
As explained in Subsection 9.1.3, these two densities produce two invariant densities |de
γc |
νc | on Graff c which differ by a multiplicative constant.
and |de
9.3. CURVATURE MEASURES 437

Theorem 9.3.20 (Crofton Formula). Let 1 ≤ p ≤ n − c, and consider the function

f : Graff c → R, f (L) = µp (L ∩ D).

If the function f is |de

ν |-integrable, then
  Z
p+c
µp+c (D) = µp (L ∩ D)|de
νc (L)|.
p Graff c

Proof. For simplicity, we set V = Rn , m = n − 1 = dim M . We will carry out the proof
in several steps.
Step 1. We will prove that there exists a constant ξm,c,p, depending only on m, c, and p
such that Z
ξn,c,pµp+c (D) = µp (L ∩ D)|deνc (L)|.
Graff c
 
Step 2. We will show that the constant ξ is equal to p+c p by explicitly computing both
sides of the above equality in the special case D = D m+1 .
Step 1. We will rely on a basic trick in integral geometry. For every S ∈ Graff c we
denote by [S] ∈ Grc the parallel translate of S containing the origin, [S] = S − S. We
introduce the incidence relation

I = (v, S) ∈ V × Graff c ; v ∈ S ⊂ V × Graff c .

Observe that we have a diffeomorphism

I → V × Grc , I ∋ (v, S) 7−→ (v, [S]) ∈ V × Grc ,

with inverse
V × Grc (V ) ∋ (v, L) 7−→ (v, v + L) ∈ I.
We obtain a double fibration

')'r
^[[
ℓ
I

V Graff c

and we set 
I(M ) := ℓ−1 (M ) = (v, S) ∈ V × Graff c ; v ∈ S ∩ M .
Since dim I = dim V + dim Grc = n + c(n − c) we deduce

dim I(M ) = n + c(n − c) − codim M = n + c(n − c).

Again we have a diagram

I(M )
'')r
^[[
ℓ

M Graff c
438 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

The map r need not be a submersion. Fortunately, Sard’s theorem shows that r fails to
be a surjection on a rather thin set.
Denote by Graff c (M ) the set of codimension c affine planes which intersect M transver-
sally. The set Graff c (M ) is an open subset of Graff c , and Sard’s theorem implies that
its complement has measure zero. We set


I(M )∗ := r −1 Graff c (M ) .

The set I(M )∗ is an open subset of I(M ), and we obtain a double fibration

I(M )∗

DAAA r
ℓ (9.3.23)
c
M Graff (M )

The fiber of r over L ∈ Graff c is the slice ML := L ∩ M which is the boundary of the
domain DL := (L ∩ D) ⊂ L.
The vertical bundle of the fibration r : I∗ (M ) → Graff c (M ) is equipped with a natural
density given along a fiber L ∩ M by the curvature density |dµk | of the domain DL . We
will denote this density by |dµL k |. As explained in Subsection 9.1.1, using the pullback
r ∗ |de
γc | we obtain a density
|dλ| := |dµL ∗
p | × r |de
γc |
on I∗ (M ) satisfying,
Z Z Z ! Z
|dλ| = |dµL
p| |de
γc (L)| = µp (L ∩ D)|de
γc (L)|.
I∗ (M ) Graff c (M ) L∩M Graff c

To complete Step 1 in our strategy it suffices to prove that there exists a constant ξ,
depending only on m and c such that

ℓ∗ |dλ| = ξ|dµc |,

where the curvature density is described in Definition 9.3.13.

Set h := (m − c). The points in I(M ) are pairs (x, L), where x ∈ M , and L is an affine
plane of dimension h + 1. Suppose (x0 , L0 ) ∈ I∗ (M ). Then we can parametrize a small
open neighborhood of (x0 , L0 ) in I∗ (M ) by a family

(x, e0 (S), e1 (S), . . . , eh (S), eh+1 (S), . . . , em (S)),

where x runs in a small neighborhood of x0 ∈ M , S runs in a small neighborhood U0 of

[L0 ] in Grc so that the following hold for every S.

• The collection
{e0 (S), e1 (S), . . . , eh (S), eh+1 (S), . . . , em (S)}
is an orthonormal frame of Rn .

• The collection {e0 , e1 , . . . , eh } is a basis of S.

9.3. CURVATURE MEASURES 439

• The collection span {e1 , . . . , eh } is a basis of Tx0 M ∩ S.

• The above condition imply that e0 is a normal vector to the boundary of the domain
D ∩ L ⊂ L. We require that e0 is the outer normal.

A neighborhood of (x0 , L0 ) in I is parametrized by the family

~r, e0 (S), e1 (S), . . . , eh (S), eh+1 (S), . . . , em (S) ,

where ~r runs in a neighborhood of x0 in the ambient space V .

We denote by SD , the co-oriented second fundamental form of D, by SL the co-oriented
second fundamental form of DL ⊂ L, and by |dVL∩M | the metric volume density on L∩M .
Then, if we set k = dim L − p = m − c − p, we deduce
1
|dµL
p| = trk (−SL )|dVL∩M |.
σk
In the sequel we will use the following conventions.

• i, j, k denote indices running in the set {0, . . . , h}.

• α, β, γ denote indices running in the set {h + 1, . . . , m}.
• A, B, C denote indices running in the set {0, 1, . . . , m}.

We denote by (θ A ) the dual coframe of (eA ), and set

θAB := eA • (D eB ).

Then, the volume density of the natural metric on Grc is

^
|dγc | = θαi .
α,i

Then ^ ^
|de
γc | = D ~r • eα × |dγc | = θ α × |dγc |,
α α
and ^
1
|dλ| = |dµL
p | × |de
γc | = det(−SL∩M )|dVL∩M | × θ α × |dγc |. (9.3.24)
σk α
The fiber of ℓ : I(M ) → M over x0 is described by

Gx0 := (~r, eA (S)) ∈ I(M ), ~r = x0 .

We set
G∗x0 := Gx0 ∩ I∗ (M ).
G∗x0 (M ) can be identified with the space of linear subspaces S of codimension c such that
Tx0 M + S = V , i.e., the affine subspace x0 + S intersects M transversally at x0 .
Denote by n a smooth unit normal vector field defined in a neighborhood of x0 in M ,
i.e.,
n(x) ⊥ Tx M, |n(x)| = 1.
440 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Lemma 9.3.21. Suppose x0 + S intersects M transversally at x0 . We set eA := eA (S).

Then at the point x0 ∈ M we have

|(n • e0 )| · |dVM | = |θ 1 ∧ · · · ∧ θ m |,

i.e., for any X1 , . . . , Xm ∈ Tx0 M we have

|(n • e0 )| · |dVM |(X1 , . . . , Xm ) = |θ 1 ∧ · · · ∧ θ m |(X1 , . . . , Xm ).

Proof. It suffices to verify this for one basis X1 , . . . , Xm of Tx0 M which we can choose to
consists of the orthogonal projections f 1 , . . . , f m of e1 , . . . , em . These projections form a
basis since S intersects Tx0 M transversally.
Observe that
f i = ei , ∀1 ≤ i ≤ h, f α = eα − (eα • n)n.
Then
|dVM |(f 1 , . . . , f m )2 = det(f A • f B )1≤A,B≤m
We observe that
f i • f j = δij , f i • f α = 0, ∀1 ≤ i, j ≤ 2h < α
f α • f β = δαβ − nα nβ , nα := n • eα .
We deduce
|dVM |(f 1 , . . . , f m )2 = det(1 − A),
where A denotes the c × c symmetric matrix with entries nα nβ , h + 1 ≤ α, β ≤ m = h + c.
If we denote by ~u the vector  
nh+1
 
~u =  ...  ∈ Rc ,
nh+c
which we also regard as a c × 1 matrix, then we deduce that A = ~u~u† .
This matrix has a c − 1 dimensional kernel corresponding to vectors orthogonal to ~u.
The vector ~u itself is an eigenvector of A, and the corresponding eigenvalue λ is obtained
from the equality
X
λ~u = |~u|2 ~u =⇒ λ = |~u|2 = n2α = |n|2 − |n • e0 |2 = 1 − |n • e0 |2 .
α

We conclude that

det(1 − A) = |n • e0 |2 =⇒ |dVM |(f 1 , . . . , f m ) = |n • e0 |.

On the other hand,

|θ 1 ∧ · · · ∧ θ m |(f 1 , . . . , f m ) = | det(eA • f B )1≤A,B≤m |.

We have again
ei • f j = δij , ei • f α = 0, ∀1 ≤ i, j ≤ h < α,
9.3. CURVATURE MEASURES 441

eα • f β = δαβ − nα nβ ,
so that
|θ 1 ∧ · · · ∧ θ m |(f 1 , . . . , f m ) = |n • e0 |2 .
The lemma is now proved. ⊔
⊓

Lemma 9.3.22 (Euler-Meusnier). Suppose L ∈ Graff c intersects M transversally, and

x0 ∈ L ∩ M . If n is a unit vector perpendicular to Tx0 M , then

SL = (n • e0 )SD |Tx0 ∩[L] ,

that is,
SL (ei , ej ) = (n • e0 )SD (ei , ej ), ∀1 ≤ i, j ≤ 2h.

Proof. We have
SL (ei , ej ) = e0 • (D ei ej ).
Let us now observe that the vector (D ei ej ) is parallel with the plane L because the vectors
ei and ej lie in this plane. Thus, D ei ej decomposes into two components, one component
parallel to e0 , and another component, (Deei ej )τ , tangent to L ∩ M . Hence
h
X
k
D ei ej = SL (ei , ej )e0 + Sij ek .
k=1

Taking the inner product with n we deduce

SD (ei , ej ) = (D ei ej ) • n = SL (ei , ej )(e0 • n). ⊔

⊓

From the above lemma we deduce

trk (−SL |x0 ) = |n • e0 |k trk (−SD |Tx0 M ∩[L] ),

for any (x0 , L) ∈ I∗ (M ). In a neighborhood of (x0 , L0 ) ∈ I∗ (M ) we have

1 ^
|dλ|(x, L) = trk (−SL )|dVL∩M | × θ α × |dγc |
σk α

m
1
^
= |n • e0 |k trk (−SD |Tx M ∩[L] ) θ A × |dγc |
σk
A=1

(use Lemma 9.3.21)

1

= |n • e0 |k+1 trk (−SD |Tx M ∩[L] ) |dVM | × |dγc |.
σk
This proves that along the fiber G∗x0 we have
1

|dλ|/|dVM | = |n • e0 |k+1 trk (−SD |Tx0 M ∩[L] ) |dγc |([L]).
σk
442 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

If we denote by θ([L], Tx0 M ) the angle between [L] and the hyperplane Tx0 M we deduce
1

|dλ|/|dVM | = · | cos θ([L], Tx0 M )|k+1 trk (−SD |Tx0 M ∩[L] ) |dγc |([L]).
σk
The map G∗x0 ∋ (x0 , L) 7−→ [L] ∈ Grc identifies G∗x0 with an open subset of Grc whose
complement has measure zero. We now have the following result.
Lemma 9.3.23. Suppose that V is an Euclidean space, dim V = m + 1, H ⊂ V is a
hyperplane through the origin, and B : H × H → R a symmetric bilinear map. Denote by
O(H) the subgroup of orthogonal transformations of V which map H to itself and suppose
that
f : Grc → R
is an O(H)-invariant function. Define

GrcH := S ∈ Grc ; S intersects H trasversally .

Then, for every 0 ≤ k ≤ m − c, there exists a constant ξ = ξm,c,k , depending only on m,

c and k, such that
Z Z
Ik (f, B) := f (S) trk (B|H∩S )|dγc |(S) = ξm,c,k trk (B) f |dγc |(S).
GrcH Grc

Proof. Observe that, for fixed f , the map B →

7 Ik (f, B) is an O(H)-invariant homogeneous
polynomial of degree k in the entries of B. We can therefore express it as a polynomial


Ik (f, B) = Pf tr1 (B), . . . , trk (B)


= ξf trk (B) + Qf tr1 (B), . . . , trk−1 (B) .

Let us prove that Qf ≡ 0. To do this, we apply the above formula to a symmetric bilinear
form B such that
dim ker B > m − k.
Thus, at least m − k + 1 of the m eigenvalues of B vanish, so that trk (B) = 0. For such
forms we have

Ik (f, B) = Q tr1 (B), . . . , trk−1 (B) .
On the other hand, for almost all S ∈ GrcH we have

dim S ∩ ker B > m − c − k.

The restriction of B to S ∩ H has m − c eigenvalues, and from the above inequality we

deduce that at least m − c − k of them are trivial. Hence

Ik (f, B) = 0, ∀B, dim ker B > m − k =⇒ Qf = 0.

Now choose B to be the bilinear form corresponding to the inner product on H. Then
   
m m−c
trk (B) = and trk (B|H∩S ) = , ∀S ∈ GrcH ,
k k
9.3. CURVATURE MEASURES 443

and we conclude that Z   

m−c m
f |dγc |(S) = ξf . ⊔
⊓
k Gr c k
Now apply the above lemma in the special case
1
H = Tx0 M, B = −SD , f (S) = | cos θ(S, H)|k+1
σk
to conclude that
ℓ∗ |dλ| = ξ trk (−SD )|dVM | = ξ|dµp+c |
so that
Z Z
ξµp+c (D) = ℓ∗ |dλ| = |dλ|
M I∗ (M )
Z Z
= r∗ |dλ| = µp (L ∩ D)|de
γc |(L).
Graff c (M ) Graff c (M )

Thus, rescaling |de

γc | to |de
νc |, we deduce that there exists a constant ξ depending only on
m and c such that Z
ξµp+c (D) = µp (L ∩ D)|deνc |(L).
Graff c (M )
Step 2. To determine the constant ξ in the above equality we apply it in the special case
D = Dm+1 . Using (9.3.19) we deduce
   
m+1 ω m+1 m+1
ξµp+c (Dm+1 ) = ξω p+c =ξ .
p+c ω m+1−c−p p + c
Now observe that for L ∈ Graff c we set r = r(L) = dist (L, 0). Then L ∩ Dm+1 is empty
if r > 1, and it is a disk of dimension (m + 1 − c) = dim L and radius (1 − r 2 )1/2 if r < 1.
We conclude that
(
m+1 m+1−c (1 − r 2 )p/2 r < 1
µp (L ∩ D ) = µp (D )×
0 p > 1.
We set
   
m+1−c m+1−c ω m+1−c m + 1 − c
µm,c,p := µp (D ) = ωp = .
p ω m+1−c−p p
Using Theorem 9.1.16 we deduce
Z
µp (L ∩ Dm+1 )|de
νc |(L)
Graff c
Z Z !
m+1


= µp D ∩ (x + [L]) |dV[L]⊥ |(x) |dνc |([L])
Grc [L]⊥
Z Z !
2 p/2
= µm,c,p (1 − |x| ) |dV[L]⊥ |(x) |dνc |([L])
Grc x∈[L]⊥ , |x|<1
| {z }
=:Ic,p
Z  
(9.3.22 ) m+1
= µm,c,p Ic,p |dνc (S)| = µm,c,p Ic,p .
Grc c
444 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Hence      
ω m+1 m+1 ω m+1−c m + 1 − c m+1
ξ = Ic,p .
ω m+1−c−p p+c ω m+1−c−p p c
Using spherical coordinates on Rc we deduce
Z Z 1
2 p/2
Ic,p = (1 − |x| ) dVRc = σ c−1 r c−1 (1 − r 2 )p/2 dr
Rc 0
Z
σ c−1 1 c−2
s=r 2
= s 2 (1 − s)p/2 ds
2 0
  (9.1.9) σ c p
(9.1.8) σ c−1 c p c−1 Γ( 2 )Γ(1 + 2 )
= B , +1 =
2 2 2 2 Γ(1 + 2c + p2 )
(9.1.10) Γ(1 + 2p )
c ω p+c
= Γ(1/2) = .
Γ(1 + 2c + 2p ) ωp
Hence     
m+1 ω m+1−c ω p+c m + 1 m + 1 − c
ξω m+1 =
p+c ωp c p
  
ω m+1 ω p+c m + 1 m + 1 − c
= .
ωpωc c p
We deduce
m+1
   
ω p+c c ω p+c p + c p+c
ξ=
ω p ωc m+1
m+1−c
=
ω p ωc p
=
p
.
p+c p
⊔
⊓

We now describe a simple situation when the function Graff c ∋ L 7→ µ0 (L ∩ M ) is

integrable.
Proposition 9.3.24. If the bounded domain D ⊂ Rn with smooth boundary is tame, then
the function
Graff c ∋ L 7→ χ(L ∩ D) = µ0 (L ∩ D)
is bounded and has compact support. In particular, it is integrable and
Z
µc (D) = χ(L ∩ D) |de ν |(L).
Graff c

Proof. We know that there exists a tame structure S such that D ∈ Sn . In particular,
M = ∂D ∈ Sn . Since the Grassmannian Graff c is semialgebraic, we deduce that Graff c
is also S-definable. Hence, the incidence set

I = ((x, L) ∈ M × Graff c ; x ∈ L

is also S-definable, and so is the map

π : I → Graff c , (x, L) 7→ L.
9.3. CURVATURE MEASURES 445

The fiber of π over L is the intersection L∩M . From the definability of Euler characteristic
we deduce that the map
Graff c ∋ L 7→ χ(L ∩ M ) ∈ Z
is definable. Its range is a definable subset of Z, i.e., a finite set. To see that it has
compact support it suffices to observe that since M is compact, an affine plane which is
too far from the origin cannot intersect M . ⊔
⊓

Definition 9.3.25. For any compact tame subset S ⊂ Rn , and for every integer 0 ≤ p ≤ n
we define Z
µ̂p (S) := χ(L ∩ D) |deν |(L). ⊔
⊓
Graff p (Rn )

The proof of Proposition 9.3.24 shows that the integral in the above definition is well
defined and finite. This proposition also shows that

µ̂p (D) = µp (D)

if D ⊂ Rn is a tame domain with smooth boundary.

The above definition has a “problem”: a priori, the quantity µ̂p (S) depends on the
dimension of the ambient space Rn ⊃ S. The next exercise asks the reader to prove that
this is not the case.
Exercise 9.3.26. Fix tame structure S, and suppose S ⊂ Rn is a compact, S-definable
set. Let N > n and suppose i : Rn → RN is an affine, isometric embedding.
(a) Prove that
Z Z
χ(L ∩ S) |de
ν |(L) = χ(U ∩ i(S) ) |de
ν |(U ),
Graff p (Rn ) Graff p (RN )

so that µ̂p (S) is independent of the dimension of the ambient space.

(b) Prove that for every real number λ we have

µ̂p (λS) = |λ|p µ̂p (S).

(c) Suppose S1 , S2 ⊂ Rn are two compact S-definable sets. Prove that

µ̂p (S1 ∪ S2 ) = µ̂p (S1 ) + µ̂2 (S2 ) − µ̂p (S1 ∩ S2 ).

(d) Suppose T ⊂ Rn is a triangle, i.e., the convex hull of three non collinear points.
Compute µ̂p (T ), ∀p ≥ 0. ⊔
⊓

Proposition 9.3.27. Fix a tame structure S. Suppose K ⊂ Rn is a compact, S-definable

set, Λ is an arbitrary S-definable set, and A ⊂ K × Λ is a closed S-definable subset of
K × Λ. For every λ ∈ Λ we set

Aλ := a ∈ K; (a, λ) ∈ A .

Then the function Λ ∋ λ 7−→ fp (λ) := µ̂p (Aλ ) ∈ R is bounded and Borel measurable.
446 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Proof. Denote by Graff p (K) the set of affine planes of codimension p in Rn which intersect
K. This is a compact, definable subset of Graff p . Define

F : Λ × Graff p (K) → Z, (λ, L) 7→ χ(Aλ ∩ L) ∈ Z.


Since the family Aλ ∩ L λ,L is definable, we deduce that the function F is definable.
Hence, its range must be a finite (!?!) subset of Z. In particular, this also shows that F
is measurable and bounded.
Now observe that Z
fp (λ) = F (λ, L) |de
ν |(L).
Graff p (K)

The measurability follows from the classical Fubini theorem. The boundedness is obvious.
⊔
⊓

The above result has the following immediate corollary.

Corollary 9.3.28. Suppose the bounded domain D ⊂ Rn is tame and has smooth bound-
ary. Denote by Graff cD the set of affine planes of codimension c which intersect ∂D
transversally. Then for any integer, 0 ≤ p ≤ n the function

Graff cD ∋ L 7→ µp (L ∩ D)

is bounded, measurable, and has compact support. In particular, it is integrable, and

  Z Z
p+c
µp+c (D) = µp (L ∩ D) |de
ν (L)| = µ̂p (L ∩ D)|de
ν |(L).
c Graff cD Graff c

Proof. The set Graff cD is a tame open and dense subset of Graff c so that its complement
has measure zero. Moreover

µp (L ∩ D) = µ̂p (L ∩ D), ∀L ∈ Graff cD ,

while, by Proposition 9.3.27, the function L →

7 µ̂p (L ∩ D) is bounded and measurable.
This function has compact support because the planes too far from the origin cannot
intersect D. ⊔
⊓

9.3.7 Crofton formulæ for submanifolds of an Euclidean space

In this last subsection we fix a tame structure S, and we assume that M is a closed,
compact, smooth, S-definable submanifold of dimension m of the Euclidean space Rn . We
continue to denote by Graff c the Grassmannian of affine planes in Rn of codimension c.
We want to prove the following result.
Theorem 9.3.29 (Crofton Formula. Part 1.). Denote by Graff c (M ) the subset of Graff c
consisting of affine planes which intersect M transversally. Then
Z Z
µc (M ) = µ0 (L ∩ M )|de
ν |(L) = µ0 (L ∩ M )|de
ν (L)|,
Graff c (M ) Graff c
9.3. CURVATURE MEASURES 447

that is,
µ̂c (M ) = µc (M ).

Proof. We can assume that k = codim M < 1. For every x ∈ Rn set d(x) := dist (x, M ).
Fix R > 0 such that for any x such that d(x) ≤ R there exists a unique point x̄ ∈ M such
that
|x − x̄| = d(x).
For every r < R consider the tube of radius r, around M ,

Dr := Tr (M ),

and set Mr = ∂Tr (M ). For uniformity, we set D0 = M . From (9.3.15) we deduce

µc (M ) = lim µc (Dr ).
r→0

The tube Dr is a tame domain with smooth boundary, and Theorem 9.3.20 implies
Z
µc (Dr ) = χ(L ∩ Dr )|dν|(L).
Graff c

Thus it suffices to show that

Z Z
lim χ(L ∩ Dr )|dν|(L) = χ(L ∩ M ).
r→0 Graff c Graff c

For r ∈ [0, R) we define

fr : Graff c → R, fr (L) = χ(L ∩ Dr ).

Lemma 9.3.30. There exists C > 0 such that

|fr (L)| ≤ C, ∀L ∈ Graff c , r ∈ [0, R).

Proof. This follows from Proposition 9.3.27. ⊔

⊓

Let 
Graff c (Mr ) := L ⊂ Graff c ; L intersects Mr transversally .
Observe that Graff c (M ) is an open subset of Graff c with negligible complement. For
every r > 0 we set

Xr = L ∈ Graff c (M ); L ∈ Graff c (Ms ); χ(L ∩ Ds ) = χ(L ∩ M ), ∀s ∈ (0, r]

Observe that
Graff c (M, r1 ) ⊂ Graff c (M, r0 ), ∀r1 ≥ r0 .
To proceed further we need the following technical result, whose proof will presented at
the end of this subsection.
448 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Lemma 9.3.31. The sets Xr are measurable in Graff c and

[
Xr = Graff c (M ).
r>0

Set 
Graff c∗ := L ∈ Graff c ; L ∩ DR 6= ∅ .
The region Graff c∗ (M ) is a relatively compact subset of Graff c , and thus it has finite
measure. Define
X∗r := Xr ∩ Graff c∗ , Y∗r := Graff c∗ \X∗r .
For 0 < r < R we have
Z Z
µc (Dr ) = νc |(L) =
fr (L)|de νc |(L)
fr (L)|de
Graff c Grc∗
Z Z Z Z
= νc | +
fr (L)|de νc | = 2
fr (L)|de νc | +
f0 (L)|de νc |
fr (L)|de
X∗r Y∗r X∗r Y∗r

Hence
Z Z
µc (Mr ) − νc | ≤
f0 (L)|de νc | ≤ Cvol (Y∗r ).
|fr (L)||de
X∗r Y∗r

We now let r → 0, and since vol (Y∗r ) → 0 we conclude that

Z Z
µc (M ) = lim µc (Dr ) = lim νc | =
f0 (L)|de νc |.
f0 (L)|de
r→0 r→0 X∗ Graff c
r

This concludes the proof of Theorem 9.3.29. ⊔

⊓

Proof of Lemma 9.3.31. We will prove that for any given L0 ∈ Graff c (M ) there exists
and ρ = ρ(L0 ) such that
L0 ∈ Xρ .
The measurability follows from the fact that Xr is described using countably many boolean
operations on measurable sets.
Consider the normal bundle

N := (T M )⊥ → M.

For x in M we denote by Nx the fiber of N over x.

Let y ∈ L0 ∩ M . We denote by Ny0 the orthogonal complement in L0 of Ty (L0 ∩ M ),

⊥
Ny0 = L0 ∩ Ty (L0 ∩ M ) .

We think of Ny0 as an affine subspace of Rn containing y. Because L0 intersects M

transversally we have
dim Ny = dim Ny0 = k = codim M.
9.3. CURVATURE MEASURES 449

0
Ny(r) L0

Dr

Figure 9.2: Slicing the tube Dr around the submanifold M by a plane L0 .

For every r > 0 we set Ny0 (r) := Ny0 ∩ Dr ; see Figure 9.2.
The collection (Ny0 )y∈L0 ∩M forms a vector subbundle N 0 → L0 ∩ M of (T Rn )|L0 ∩M .
The exponential map on T Rn restricts to a map

EL0 ∩M : N 0 → L0 .

Denote by δ the pullback to N 0 of the distance function x 7→ d(x) = dist (x, M ),

δL0 = d ◦ EL0 ∩M : N 0 → R.

For every y ∈ L0 ∩ M , the restriction to Ny0 of the Hessian of d at y is positive definite.

Hence, there exists ρ = ρ(L0 ) > 0 such that the map

EL0 ∩M : x ∈ N 0 ; δ(x) ≤ ρ → L0 ∩ Dρ

is a diffeomorphism. We deduce that we have a natural projection

π : L0 ∩ Dρ → L0 ∩ M,

which is continuous and defines a locally trivial fibration with fibers Ny0 (ρ).
For every y ∈ L0 ∩ M , the fiber Ny0 (ρ) is homeomorphic to a disk of dimension k.
Thus, L0 ∩ Dρ is homeomorphic to a tube in L0 around L0 ∩ M ⊂ L0 , so that

χ(L0 ∩ Dρ ) = χ(L0 ∩ M ).

The downward gradient flow of the restriction to L0 ∩ Dρ of the distance function d(x)
produces diffeomorphisms of manifolds with boundary

L0 ∩ Dρ ∼
= L0 ∩ Dr , ∀r ∈ (0, ρ).

Hence
χ(L0 ∩ Dr ) = χ(L0 ∩ Dρ ) = χ(L0 ∩ M ), ∀r ∈ (0, ρ].
450 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Since the restriction to L0 ∩ Dρ of the distance function d(x) has no critical points other
than the minima y ∈ L0 ∩ M , we deduce that L0 is transversal to the level sets

{ d(x) = r } = Mr , ∀r ∈ (0, ρ].

This proves L0 ∈ Xρ . ⊔
⊓

Corollary 9.3.32. Suppose C ⊂ R2 is a smooth, closed, compact tame curve. For every
line L ∈ Gr1 (R2 ) = Gr1 (R2 ) we set

nC (L) := #(L ∩ C).

Then the function L 7→ nC (L) belongs to L∞ (Gr1 (R2 ), |de

ν |2,1 ), has compact support and
Z
length (C) = nC (L)|de ν2,1 |(L). ⊔
⊓
Gr1 (R2 )

Remark 9.3.33. Theorem 9.3.29 offers a strange interpretation to the Hilbert-Einstein

functional. Suppose M is a tame, compact, orientable submanifold of the Euclidean space
Rn . Let m = dim M , and denote by h the induced metric on M . Then
1
µm−2 (M, h) = EM (h),
2π
where EM (h) is the Hilbert-Einstein functional
Z
EM (h) = s(h)dVM (h).
M

Denote by Graff m−2 (M ) = Graff (M ) the set of codimension (m − 2) affine planes in Rn

which intersect M along a nonempty subset, and by Graff m−2 ∗ (M ) ⊂ Graff m−2 (M ) the
subset consisting of those planes intersecting M transversally. For every L ∈ Graff m−2
∗ (M ),
the intersection L ∩ M is a, possible disconnected, smooth, orientable Riemann surface.
By Sard’s theorem the complement of Graff m−2 ∗ (M ) in Graff m−2 (M ) has zero measure.
For every R > 0 we denote by Graff m−2 (R) the set of codimension (m − 2) planes
in Rn which intersect the disk DnR of radius R and centered at 0. Note that since M is
compact, there exists R0 > 0 such that

Graff m−2 (M ) ⊂ Graff m−2 (R), ∀R > R0 .

We set Z Z
c(R) = |de
ν (L)|, c(M ) := |de
ν (L)|.
Graff m−2 (R) Graff m−2 (M )

Using Crofton’s formula in Proposition 9.3.24 and the simple observation that

χ(L ∩ DnR ) = 1, ∀L ∈ Graff m−2 (R),

we deduce that  
n
c(R) = µm−2 (DnR ) = ωm−2 Rm−2 .
m−2
9.3. CURVATURE MEASURES 451

1
Observe now that c(R) ν (L)| is a probability measure on Graff m−2 (R), and we can regard
|de
the correspondence
Graff m−2 (R) ∋ L 7−→ χ(L ∩ M ) ∈ Z
as a random variable ξR whose expectation is

1 1
hξR i := µm−2 (M ) = EM (h).
c(R) 2πc(R)

We deduce
   
n m−2 n
EM (h) = 2π ωm−2 R hξR i = 2π ωm−2 lim Rm−2 hξR i. ⊔
⊓
m−2 m−2 R→∞

Theorem 9.3.34 (Crofton Formula. Part 2.). Denote by Graff c (M ) the subset of Graff c
consisting of affine planes which intersect M transversally. Then, for every 0 ≤ p ≤ m − c
we have
  Z Z
p+c
µp+c (M ) = µp (L ∩ M )|de
ν |(L) = µ̂p (L ∩ M )|de
ν (L)|.
p Graff c (M ) Graff c

Proof. We continue to use the same notations we used in the proof of Theorem 9.3.29.
From (9.3.15) we deduce
µp+c (M ) = lim µp+c (Dr ).
r→0

The set Dr is a tame domain with smooth boundary, and Corollary 9.3.28 implies
  Z Z
p+c
µp+c (Dr ) = µp (L ∩ Dr )|de
ν |(L) = µ̂p (L ∩ Dr )|de
ν |(L).
c Graff c Graff c

Thus it suffices to show that

Z Z
lim µ̂p (L ∩ Dr )|de
ν (L)| = µ̂p (L ∩ M )|de
ν (L)|.
r→0 Graff c Graff c

For r ∈ [0, R) we define

gr : Graff c → R, gr (L) = µ̂p (L ∩ Dr ).

From Proposition 9.3.27 we deduce that

∃C > 0 : |gr (L)| ≤ C, ∀L ∈ Graff c , r ∈ [0, R). (9.3.25)

Lemma 9.3.35.
lim gr (L) = g0 (L),
rց0

for all L ∈ Graff c (M ).

452 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY

Proof. Let L ∈ Graff c (M ) and set Sr := L ∩ Dr . We have to prove that

lim µ̂p (Sr ) = µ̂p (S0 ).

rց0

Denote by Graff p (S0 ) the set of codimension p affine planes in Rn which intersect S0
transversally. The set Graff p (S0 ) is an open subset of Graff p , and its complement has
measure zero. Then,
Z Z
p
gr (L) = µ̂(Sr ) = χ(U ∩ Sr )|deν (U )| = χ(U ∩ Sr )|de ν p (U )|. (9.3.26)
Graff p Graff p (S0 )

From Proposition 9.3.27 we deduce that the function

Graff p ×[0, R) ∋ (U, r) 7→ χ(U ∩ Sr ) ∈ Z

is definable, and thus bounded.

Arguing exactly as in the proof of Lemma 9.3.31 we deduce that, for every U ∈
Graff p (S0 ), there exists ρ ∈ (0, R) such that, for every r ∈ (0, ρ), the set U ∩ Sr is
homotopy equivalent to U ∩ S0 . Hence

lim χ(U ∩ Sr ) = χ(U ∩ S0 ), ∀U ∈ Graff p (S0 ).

rց0

If we let r go to zero in (9.3.26), we deduce from the dominated convergence theorem that
Z
lim gr (L) = ν p (U )| = µ̂p (S0 ).
χ(U ∩ S0 )|de
rց0 Graff p

Using the fact that S0 is a smooth, compact tame manifold we deduce from Theorem
9.3.29 that
µ̂p (S0 ) = µp (S0 ) = g0 (L). ⊔
⊓
Using (9.3.25) and Lemma 9.3.35, we deduce from the dominated convergence theorem
that Z Z
lim µ̂p (L ∩ Dr )|de
ν (L)| = µ̂p (L ∩ M )|de
ν (L)|
rց0 Graff c Graff c
Z
= µp (L ∩ M )|de
ν (L)|. ⊔
⊓
Graff c (M )

Remark 9.3.36. (a) The tameness assumption in the Crofton formulaæ is not really
needed, and they are true in much more general situations; see [45], Sect. 2.10, 3.2.
We have chosen to work in this more restrictive context for two reasons. First, the
tameness assumption eliminates the need for sophisticated analytical arguments. Second,
we believe that geometers should become more familiar with tame context in which the
geometric intuition is not distracted by analytical caveats.
(b) The subject of integral geometry owes a great deal to the remarkable work of S.S.
Chern who was the first to have reformulated the main problems from a modern point of
view, and push our understanding of this subject to remarkable heights.
It was observed by Chern that the Crofton formulæ that we have proved in this section
are only special cases of the so called kinematic formulæ which explain how to recover
9.3. CURVATURE MEASURES 453

information about a submanifold of a homogeneous space such as Rn , by intersecting it

with a large family of simpler shapes. In the Euclidean space, these other simpler shapes
could be affine planes, spheres of a given radius, balls of a given radius.
For example, given a compact, m-dimensional submanifold M ⊂ Rn , and a radius R,
we can ask what is the value of the integral
Z
χ(M ∩ BR (x) )|dx|,
Rn

where BR (x) denotes the ball BR (x) = { y ∈ Rn ; |y − x| ≤ R }. One of the kinematic

formulæ states that
Z Xn
χ(M ∩ BR (x) ) = cm,n,i µi (M )Rn−i ,
Rn i=0

where cm,n,i are universal constants, independent of M .

The above equality is strikingly similar to the tube formula. To see the origin of this
similarity note that BR (x) ∩ M 6= ∅ ⇐⇒ x ∈ TR (M ). If R is sufficiently small, the above
formula does indeed specialize to the tube formula, but the kinematic formula extends
beyond the range of applicability of the tube formula.
A detailed presentation of these ideas will send us too far, and instead of pursuing
further this line of thought, we recommend to the reader to open the classical, yet very
much alive monograph of L. Santaló [115], for a more in depth look at this subject. ⊔
⊓
454 CHAPTER 9. CLASSICAL INTEGRAL GEOMETRY
Chapter 10

Elliptic Equations on Manifolds

Almost all the objects in differential geometry are defined by expressions involving partial
derivatives. The curvature of a connection is the most eloquent example.
Many concrete geometric problems lead to studying such objects with specific prop-
erties. For example, we inquired whether on a given vector bundle there exist flat con-
nections. This situation can be dealt with topologically, using the Chern–Weil theory of
characteristic classes.
Very often, topological considerations alone are not sufficient, and one has look into
the microstructure of the problem. This is where analysis comes in, and more specifically,
one is led to the study of partial differential equations. Among them, the elliptic ones
play a crucial role in modern geometry.
This chapter is an introduction to this vast and dynamic subject which has numerous
penetrating applications in geometry and topology.

10.1 Partial differential operators: algebraic aspects

10.1.1 Basic notions
We first need to introduce the concept of partial differential operator (p.d.o. for brevity)
on a smooth manifold M . To understand the forthcoming formalism, it is best that we
begin with the simplest of the situations, M = RN .
Perhaps the best known partial differential operator is the Laplacian,
X
∆ : C ∞ (RN ) → C ∞ (RN ), ∆u := − ∂i 2 u,
i

where, as usual, ∂i := ∂xi . This is a scalar operator in the sense that it acts on scalar
valued functions. Note that our definition of the Laplacian differs from the usual one by
a sign. The Laplacian defined as above is sometimes called the geometers’ Laplacian.
Next in line is the exterior derivative

d : Ωk (RN ) → Ωk+1 (RN ).

This is a vectorial operator in the sense it acts on

vector valued functions. A degree k
form ω on RN can be viewed as a collection of Nk smooth functions or equivalently, as a

455
456 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

N
smooth function ω : RN → R( k ) . Thus d can be viewed as an operator
 N
  N

d : C ∞ RN , R( k ) → C ∞ RN , R(k+1) .

It is convenient to think of C ∞ (RN , Rν ) as the space of smooth sections of the trivial

bundle Rν over RN .
For any smooth K = R, C-vector bundles E, F over a smooth manifold M we denote
by Op(E, F ) the space of K-linear operators

C ∞ (E) → C ∞ (F ).

The space Op(E, E) is an associative K-algebra.

The spaces of smooth sections C ∞ (E) and C ∞ (F ) are more than just K-vector spaces.
They are modules over the ring of smooth functions C ∞ (M ). The partial differential
operators are elements of Op which interact in a special way with the above C ∞ (M )-
module structures. First define

P DO 0 (E, F ) := Hom(E, F ).

Given T ∈ P DO 0 , u ∈ C ∞ (E), and f ∈ C ∞ (M ), we have T (f u) − f (T u) = 0 or, in terms

of commutators,
[T, f ]u = T (f u) − f (T u) = 0. (10.1.1)
Each f ∈ C ∞ (M ) defines a map

ad(f ) : Op (E, F ) → Op (E, F ),

by
ad(f )T := T ◦ f − f ◦ T = [T, f ], ∀T ∈ Op(E, F ).
Above, f denotes the C ∞ (M )-module multiplication by f . We can rephrase the equality
(10.1.1) as

P DO 0 (E, F ) = {T ∈ Op (E, F ), ad(f )T = 0 ∀f ∈ C ∞ (M )} =: ker ad .

Define

P DO(m) (E, F ) := ker adm+1


= T ∈ Op(E, F ); T ∈ ker ad(f0 ) ad(f1 ) · · · ad(fm ), ∀fi ∈ C ∞ (M ) .

The elements of P DO (m) are called partial differential operators of order ≤ m. We set
[
P DO(E, F ) := P DO(m) (E, F ).
m≥0

Remark 10.1.1. Note that we could have defined P DO(m) inductively as

P DO (m) = {T ∈ Op ; [T, f ] ∈ P DO (m−1) , ∀f ∈ C ∞ (M )}.

This point of view is especially useful in induction proofs. ⊔

⊓
10.1. PARTIAL DIFFERENTIAL OPERATORS: ALGEBRAIC ASPECTS 457

Example 10.1.2. Denote by R the trivial line bundle over RN . The sections of R are
precisely the real functions on
RN . We want

to analyze P DO(1) = P DO(1) (R, R).
Observe first that P DO0 R, R ∼ = C ∞ RN .
Let L ∈ P DO(1) , u, f ∈ C ∞ (RN ). Then [L, f ]u = σ(f ) · u, for some smooth function
σ(f ) ∈ C ∞ (RN ). On the other hand, for any f, g ∈ C ∞ (RN )

σ(f g)u = [L, f g]u = [L, f ](gu) + f ([L, g]u) = σ(f )g · u + f σ(g) · u.

Hence σ(f g) = σ(f )g + f σ(g). In other words, the map f 7→ σ(f ) is a derivation of
C ∞ (RN ), and consequently (see Exercise 3.1.9) there exists a smooth vector field X on
RN such that
σ(f ) = X · f, ∀f ∈ C ∞ (RN ).
Let µ := L(1) ∈ C ∞ (RN ). Then, for all u ∈ C ∞ (RN ), we have

Lu = L(u · 1) = [L, u] · 1 + u · L(1) = X · u + µ · u. ⊔

⊓

Lemma 10.1.3. Any L ∈ P DO (m) (E, F ) is a local operator, i.e., ∀u ∈ C ∞ (E),

supp Lu ⊂ supp u.

Proof. We argue by induction over m. For m = 0 the result is obvious. Let L ∈

P DO(m+1) , and u ∈ C ∞ (E). For every f ∈ C ∞ (M ) we have

L(f u) = [L, f ]u + f Lu.

Since [L, f ] ∈ P DO(m) we deduce by induction

supp L(f u) ⊂ supp u ∪ supp f, ∀f ∈ C ∞ (M ).

For any open set O such that O ⊃ supp u we can find f ≡ 1 on supp u and f ≡ 0 outside
O so that f u ≡ u). This concludes the proof of the lemma. ⊔
⊓

The above lemma shows that, in order to analyze the action of a p.d.o., one can work
in local coordinates. Thus, understanding the structure of an arbitrary p.d.o. boils down
to understanding the action of a p.d.o. in P DO(m) (Kp , Kq ), where Kp , and Kq are trivial
K-vector bundles over RN . This is done in the exercises at the end of this subsection.
Proposition 10.1.4. Let E, F, G be smooth K-vector bundles over the same manifold
M . If P ∈ P DO (m) (F, G) and Q ∈ P DO(n) (E, F ) then P ◦ Q ∈ P DO (m+n) (E, G).

Proof. We argue by induction over m + n. For m + n = 0 the result is obvious. In general,

if f ∈ C ∞ (M ), then
[P ◦ Q, f ] = [P, f ] ◦ Q + P ◦ [Q, f ].
By induction, the operators in the right-hand-side have orders ≤ m+n−1. The proposition
is proved. ⊔
⊓
458 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Corollary 10.1.5. The operator

X
L= aα (x)∂ α : C ∞ (RN ) → C ∞ (RN ),
|α|≤m
X αN
α = (α1 , . . . , αN ) ∈ ZN
≥0 , |α| = αi , ∂ α = ∂1α1 · · · ∂N ,
i
is a p.d.o. of order ≤ m.

Proof. According to the computation in Example 10.1.2, each partial derivative ∂i is a 1st
order p.d.o. According to the above proposition, multiple compositions of such operators
are again p.d.o.’s. ⊔
⊓

Lemma 10.1.6. Let E, F → M be two smooth vector bundles over the smooth manifold
M . Then for any P ∈ P DO (E, F ), and any f, g ∈ C ∞ (M )

ad(f ) · (ad(g)P ) = ad(g) · (ad(f )P ).

Proof.  
ad(f ) · (ad(g)P ) = [P, g], f
     
= [P, f ], g + P, [f, g] = [P, f ], g = ad(g) · (ad(f )P ).
⊔
⊓

From the above lemma we deduce that if P ∈ P DO(m) then, for any f1 , . . . , fm ∈
C ∞ (M ), the bundle morphism

ad(f1 ) ad(f2 ) · · · ad(fm )P

does not change if we permute the f ’s.

Proposition 10.1.7. Let P ∈ P DO (m) (E, F ), fi , gi ∈ C ∞ (M ) (i = 1, . . . , m) such that,
at a point x0 ∈ M ,
dfi (x0 ) = dgi (x0 ) ∈ Tx∗0 M, ∀i = 1, . . . , m.
Then
{ad(f1 ) ad(f2 ) · · · ad(fm )P } |x0 = {ad(g1 ) ad(g2 ) · · · ad(gm )P } |x0 .
In the proof we will use the following technical result which we leave to the reader as an
exercise.
Lemma 10.1.8. For each x0 ∈ M consider the ideals of C ∞ (M )

mx0 = f ∈ C ∞ (M ); f (x0 ) = 0 ,

Jx0 = f ∈ C ∞ (M ); f (x0 ) = 0, df (x0 ) = 0,
Then Jx0 = m2x0 , i.e., any function f which vanishes at x0 together with its derivatives
can be written as X
f= gj hj gj , hj ∈ mx0 . ⊔
⊓
j
10.1. PARTIAL DIFFERENTIAL OPERATORS: ALGEBRAIC ASPECTS 459

Exercise 10.1.9. Prove the above lemma. ⊔

⊓

Proof. Let P ∈ P DO(m) and fi , gi ∈ C ∞ (M ) such that

dfi (x0 ) = dgi (x0 ) ∀i = 1, . . . , m.
Since ad(const) = 0, we may assume (eventually altering the f ’s and the g’s by additive
constants) that
fi (x0 ) = gi (x0 ) ∀i.
We will show that
{ ad(f1 ) ad(f2 ) · · · ad(fm )P } |x0 = { ad(g1 ) ad(f2 ) · · · ad(fm )P } |x0 .
Iterating we get the desired conclusion.
Let φ = f1 − g1 , and set Q := ad(f2 ) · · · ad(fm )P ∈ P DO(1) . We have to show that
{ad(φ)Q} |x0 = 0. (10.1.2)
Note that φ ∈ Jx0 so, according to the above lemma, we can write
X
φ= αj βj , αj , βj ∈ mx0 .
j

We have
nX o X X
{ad(φ)Q} |x0 = ad(αj βj )Q |x0 = {[Q, αj ]βj } |x0 + {αj [Q, βj ]} |x0 = 0.
j j j

This proves the equality (10.1.2), and hence the proposition. ⊔

⊓

The proposition we have just proved has an interesting consequence. Given P ∈

P DO(m) (E, F ), x0 ∈ M , and fi ∈ C ∞ (M ) (i = 1, . . . , m), the linear map
1
{ ad(f1 ) · · · ad(fm )P } |x0 : Ex0 → Fx0
m!
depends only on the quantities ξi := dfi (x0 ) ∈ Tx∗0 M . Hence, for any ξi ∈ Tx∗0 M (i =
1, . . . , m), the above expression unambiguously induces a linear map
σ(P )(ξ1 , . . . , ξm ) : Ex0 → Fx0 ,
that is symmetric in the variables ξi . Using the polarization formula (8.1.9) of Chapter 8,
we see that this map is uniquely determined by the polynomial
σ(P )(ξ) := σm (P )(ξ) = σ(P )(ξ, . . . , ξ ).
| {z }
m

If we denote by π : T ∗ M → M the natural projection then, for each P ∈ P DO (m) (E, F ),

we have a well defined map
σm (P )(·) ∈ Hom (π ∗ E, π ∗ F ),
where π ∗ E and π ∗ F denote the pullbacks of E and F to T ∗ M via π. Along the fibers
of T ∗ M the map ξ 7→ σm (P )(ξ) looks like a degree m homogeneous “polynomial” with
coefficients in Hom (Ex0 , Fx0 ).
460 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Proposition 10.1.10. Let P ∈ P DO (m) (E, F ) and Q ∈ P DO(n) (F, G). Then

σm+n (Q ◦ P ) = σn (Q) ◦ σm (P ). ⊔
⊓

Exercise 10.1.11. Prove the above proposition. ⊔

⊓

Definition 10.1.12. A p.d.o. P ∈ P DO(m) is said to have order m if σm (P ) 6≡ 0. In

this case σm (P ) is called the (principal) symbol of P . The set of p.d.o.’s of order m will
be denoted by P DO m . ⊔
⊓

Definition 10.1.13. The operator P ∈ P DOm (E, F ) is said to be elliptic if, for any
x ∈ M , and any ξ ∈ Tx∗ M \ {0}, the map

σm (P )(ξ) : Ex → Fx

is a linear isomorphism. ⊔
⊓

The following exercises provide a complete explicit description of the p.d.o.’s on RN .

Exercise 10.1.14. Consider the scalar p.d.o. on RN described in Corollary 10.1.5
X
L= aα (x)∂ α .
|α|≤m

Show that X X αN
σm (L)(ξ) = aα (x)ξ α = aα (x)ξ1α1 · · · ξN . ⊔
⊓
|α|=m |α|=m

Exercise 10.1.15. Let L : C ∞ (RN ) → C ∞ (RN ) be a p.d.o. of order m. Its principal

symbol has the form X
σm (L)(ξ) = aα (x)ξ α .
|α|=m
P
Show that L − |α|=m aα (x)∂ α
is a p.d.o. of order ≤ m − 1, and conclude that the only
N
scalar p.d.o.-s on R are those indicated in Corollary 10.1.5. ⊔
⊓

Exercise 10.1.16. Let L ∈ P DO (m) (E, F ) and u ∈ C ∞ (E). Show the operator

C ∞ (M ) ∋ f 7−→ [L, f ]u ∈ C ∞ (F )

belongs to P DO (m−1) (RM , F ). ⊔

⊓

Exercise 10.1.17. Let Kp and Kq denote the trivial K-vector bundles over RN of rank p
and respectively q. Show that any L ∈ P DO(m) (Kp , Kq ) has the form
X
L= Aα (x)∂ α ,
|α|≤m

where Aα ∈ C ∞ (RN , Hom (Kp , Kq )) for any α .

Hint: Use the previous exercise to reduce the problem to the case p = 1. ⊔
⊓
10.1. PARTIAL DIFFERENTIAL OPERATORS: ALGEBRAIC ASPECTS 461

10.1.2 Examples
At a first glance, the notions introduced so far may look too difficult to “swallow”. To
help the reader get a friendlier feeling towards them, we included in this subsection a
couple of classical examples which hopefully will ease this process. More specifically, we
will compute the principal symbols of some p.d.o.’s that we have been extensively using
in this book.
In the sequel, we will use the notation “ f · ” to denote the operation of multiplication
by the smooth scalar function f .
Example 10.1.18. (The Euclidean Laplacian). This is the second order p.d.o.,
X
∆ : C ∞ (RN ) → C ∞ (RN ), ∆ = − ∂i 2 , ∂i = ∂xi .
i

Let f ∈ C ∞ (RN ). Then

X X
ad(f )(∆) = − ad(f ) (∂i )2 = − {ad(f )(∂i ) ◦ ∂i + ∂i ◦ ad(f )(∂i )}
i i
X X
=− {fxi · ∂i + ∂i (fxi ·)} = − {fxi · ∂i + fxi xi · +fxi · ∂i }
i i
X
= (∆f ) · −2 fxi · ∂i .
i

Hence
X X
ad(f )2 (∆) = ad(f )(∆f ) · −2 fxi · ad(f ) (∂i ) = −2 (fxi )2 · = −2|df |2 · .
i i

If we set ξ := df in the above equality, we deduce

σ2 (∆)(ξ) = −|ξ|2 · .

In particular, this shows that the Laplacian ∆ is an elliptic operator. ⊔

⊓

Example 10.1.19. (Covariant derivatives). Consider a vector bundle E → M over

the smooth manifold M , and a connection ∇ on E. We can view ∇ as a p.d.o.

∇ : C ∞ (E) → C ∞ (T ∗ M ⊗ E).

Its symbol can be read from ad(f )∇, f ∈ C ∞ (M ). For any u ∈ C ∞ (E)

(ad(f )∇)u = ∇(f u) − f (∇u) = df ⊗ u.

By setting ξ = df we deduce σ1 (∇)(ξ) = ξ⊗, i.e., the symbol is the tensor multiplication
by ξ. ⊔
⊓

Example 10.1.20. (The exterior derivative.) Let M be a smooth manifold. The

exterior derivative
d : Ω• (M ) → Ω•+1 (M )
462 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

is a first order p.d.o. To compute its symbol, consider ω ∈ Ωk (M ), and f ∈ C ∞ (M ).

Then

ad(f )d ω = d(f ω) − f dω = df ∧ ω.
If we set ξ = df , we deduce σ1 (d) = e(ξ) = the left exterior multiplication by ξ. ⊔
⊓

Example 10.1.21. Consider an oriented, n-dimensional Riemann manifold (M, g). As in

Chapter 4, we can produce an operator

δ = ∗d∗ : Ω• (M ) → Ω•−1 (M ),

where ∗ is the Hodge ∗-operator

∗ : Ω• (M ) → Ωn−• (M ).

The operator δ is a first order p.d.o., and moreover

σ1 (δ) = ∗σ1 (d)∗ = ∗e(ξ) ∗ .

This description can be further simplified.

Fix ξ ∈ Tx∗ M , and denote by ξ ∗ ∈ Tx M its metric dual. For simplicity we assume
|ξ| = |ξ ∗ | = 1. Include ξ in an oriented orthonormal basis (ξ 1 , . . . , ξ n ) of Tx∗ M , ξ 1 = ξ,
and denote by ξi the dual basis of Tx M .
Consider a monomial ω = ξ I ∈ Λk Tx∗ M , where I = (i1 , . . . , ik ) denotes as usual an
ordered multi-index. Note that if 1 6∈ I, then

σ1 (ξ 1 )ω = 0. (10.1.3)

If 1 ∈ I, e.g., I = (1, . . . , k), then

∗e(ξ)(∗ω) = ∗(ξ 1 ∧ ξ k+1 ∧ · · · ∧ ξ n ) = (−1)(n−k)(k−1) ξ 2 ∧ · · · ∧ ξ k

= −(−1)νn,k i(ξ ∗ )ω, (10.1.4)

where νn,k = nk + n + 1 is the exponent introduced in Subsection 4.1.5 while i(ξ ∗ ) denotes
the interior derivative along ξ ∗ . Putting together (10.1.3) and (10.1.4) we deduce

σ1 (δ)(ξ) = −(−1)νn,k i(ξ ∗ ). (10.1.5)

⊔
⊓

Example 10.1.22. (The Hodge–DeRham operator). Let (M, g) be as in the above

example. The Hodge–DeRham operator is

d + d∗ : Ω• (M ) → Ω• (M ),

where d∗ = (−1)νn,k δ. Hence d + d∗ is a first order p.d.o., and moreover,

σ(d + d∗ )(ξ) = σ(d)(ξ) + σ(d∗ )(ξ) = e(ξ) − i(ξ ∗ ).

10.1. PARTIAL DIFFERENTIAL OPERATORS: ALGEBRAIC ASPECTS 463

The Hodge Laplacian is the operator (d + d∗ )2 . We call it Laplacian since

 2
2
σ (d + d∗ )2 (ξ) = σ(d + d∗ )(ξ) = e(ξ) − i(ξ ∗ ) ,

while Exercise 2.2.55 shows

2

e(ξ) − i(ξ ∗ ) = − e(ξ)i(ξ ∗ ) + i(ξ ∗ )e(ξ) = −|ξ|2g · .

Notice that d + d∗ is elliptic since the square of its symbol is invertible. ⊔

⊓

Definition 10.1.23. Let E → M be a smooth vector bundle over the Riemann manifold
(M, g). A second order p.d.o.

L : C ∞ (E) → C ∞ (E)

is called a generalized Laplacian if σ2 (L)(ξ) = −|ξ|2g . ⊔

⊓

☞ Observe that all the generalized Laplacians are elliptic operators.

10.1.3 Formal adjoints

For simplicity, all the vector bundles in this subsection will be assumed complex, unless
otherwise indicated.
Let E1 , E2 → M be vector bundles over a smooth oriented manifold. Fix a Riemann
metric g on M and Hermitian metrics h·, ·ii on Ei , i = 1, 2. We denote by dVg = ∗1 the
volume form on M defined by the metric g. Finally, C0∞ (Ei ) denotes the space of smooth,
compactly supported sections of Ei .
Definition 10.1.24. Let P ∈ P DO(E1 , E2 ). The operator Q ∈ P DO(E2 , E1 ) is said to
be a formal adjoint of P if, ∀u ∈ C0∞ (E1 ), and ∀v ∈ C0∞ (E2 ) we have
Z Z
hP u, vi2 dVg = hu, Qvi1 dVg . ⊔
⊓
M M

Lemma 10.1.25. Any P ∈ P DO (E1 , E2 ) admits at most one formal adjoint.

Proof. Let Q1 , Q2 be two formal adjoints of P . Then, ∀v ∈ C0∞ (E2 ), we have

Z
hu, (Q1 − Q2 )vi1 dVg = 0 ∀u ∈ C0∞ (E1 ).
M

This implies (Q1 − Q2 )v = 0 ∀v ∈ C0∞ (E2 ). If now v ∈ C ∞ (E2 ) is not necessarily

compactly supported then, choosing a partition of unity (α) ⊂ C0∞ (M ), we conclude
using the locality of Q = Q1 − Q2 that
X
Qv = Q(αv) = 0. ⊔
⊓

The formal adjoint of a p.d.o. P ∈ P DO (E1 , E2 ), whose existence is not yet guaran-
teed, is denoted by P ∗ . It is worth emphasizing that P ∗ depends on the choices of g and
h·, ·ii .
464 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Proposition 10.1.26. (a) Let L0 ∈ P DO (E0 , E1 ) and L1 ∈ P DO (E1 , E2 ) admit for-

mal adjoints L∗i ∈ P DO (Ei+1 , Ei ) (i = 0, 1), with respect to a metric g on the base, and
metrics h·, ·ij on Ej , j = 0, 1, 2. Then L1 L0 admits a formal adjoint, and

(L1 L0 )∗ = L∗0 L∗1 .

(b) If L ∈ P DO (m) (E0 , E1 ), then L∗ ∈ P DO (m) (E1 , E0 ).

Proof. (a) For any ui ∈ C0∞ (Ei ) we have

Z Z Z
hL1 L0 u0 , u2 i2 dVg = hL0 u0 , L∗1 u0 i1 dVg = hu0 , L∗0 L∗1 u2 i0 dVg .
M M M

(b) Let f ∈ C ∞ (M ). Then

(ad(f )L)∗ = (L ◦ f − f ◦ L)∗ = −[L∗ , f ] = − ad(f )L∗ .

Thus
ad(f0 ) ad(f1 ) · · · ad(fm )L∗ = (−1)m+1 (ad(f0 )ad(f1 ) · · · ad(fm )L)∗ = 0. ⊔
⊓

The above computation yields the following result.

Corollary 10.1.27. If P ∈ P DO(m) admits a formal adjoint, then

σm (P ∗ ) = (−1)m σm (P )∗ ,

where the ∗ in the right-hand-side denotes the conjugate transpose of a linear map. ⊔
⊓

Let E be a Hermitian vector bundle over the oriented Riemann manifold (M, g).

Definition 10.1.28. A p.d.o. L ∈ P DO (E, E) is said to be formally self-adjoint if

L = L∗ . ⊔
⊓

The above notion depends clearly on the various metrics

Example 10.1.29. Using the integration by parts formula of Subsection 4.1.5, we deduce
that the Hodge–DeRham operator

d + d∗ : Ω• (M ) → Ω• (M ),

on an oriented Riemann manifold (M, g) is formally self-adjoint with respect with the
metrics induced by g in the various intervening bundles. In fact, d∗ is the formal adjoint
of d. ⊔
⊓

Proposition 10.1.30. Let (Ei , h·, ·ii ), i = 1, 2, be two arbitrary Hermitian vector bundle
over the oriented Riemann manifold (M, g). Then any L ∈ P DO (E1 , E2 ) admits at least
(and hence exactly) one formal adjoint L∗ .
10.1. PARTIAL DIFFERENTIAL OPERATORS: ALGEBRAIC ASPECTS 465

Sketch of proof We prove this result only in the case when E1 and E2 are trivial vector
bundles over RN . However, we do not assume that the Riemann metric over RN is the
Euclidean one. The general case can be reduced to this via partitions of unity and we
leave the reader to check it for him/her-self.
Let E1 = Cp and E2 = Cq . By choosing orthonormal moving frames, we can assume
that the metrics on Ei are the Euclidean ones. According to the exercises at the end of
Subsection 10.1.1, any L ∈ P DO(m) (E1 , E2 ) has the form
X
L= Aα (x)∂ α ,
|α|≤m

where Aα ∈ C ∞ (RN , Mq×p (C)). Clearly, the formal adjoint of Aα is the conjugate trans-
pose
t
A∗α = Aα .
To prove the proposition it suffices to show that each of the operators ∂i ∈ P DO1 (E1 , E1 )
admits a formal adjoint. It is convenient to consider the slightly more general situation.
Lemma 10.1.31. Let X = X i ∂i ∈ Vect (RN ), and denote by ∇X the first order p.d.o.

∇X u = X i ∂i u u ∈ C0∞ (Cp ).

Then
∇X = −∇X − div g (X),
where div g (X) denotes the divergence of X with respect to the metric g, i.e., the scalar
defined by the equality
LX (dVg ) = div g (X) · dVg .

Proof. Let u, v ∈ C0∞ (RN , Cp ). Choose R ≫ 0 such that the Euclidean ball BR of radius
R centered at the origin contains the supports of both u and v. From the equality

X · hu, vi = h∇X u, vi + hu, ∇X vi,

we deduce Z Z
h∇X u, vidVg = h∇X u, vidVg
RN BR
Z Z
= X · hu, vidVg − hu, ∇X vidVg . (10.1.6)
BR BR

Set f := hu, vi ∈ C0∞ (RN , C), and denote by α ∈ Ω1 (RN ) the 1-form dual to X with
respect to the Riemann metric g, i.e.,

(α, β)g = β(X) ∀β ∈ Ω1 (RN ).

Equivalently,
α = gij X i dxj .
466 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

The equality (10.1.6) can be rewritten

Z Z Z
h∇X u, vidVg = df (X)dVg − hu, ∇X vidVg
BR BR RN
Z Z
= (df, α)g dVg − hu, ∇X vidVg .
BR RN

The integration by parts formula of Subsection 4.1.5 yields

Z Z Z
(df, α)g dVg = (df ∧ ∗g α) |∂BR + f d∗ αdVg .
BR ∂BR BR

Since f ≡ 0 on a neighborhood of ∂BR , we get

Z Z
(df, α)g dvg = hu, vid∗ α dVg .
BR BR

Since d∗ α = − div g (X), (see Subsection 4.1.5) we deduce

Z Z Z
X · hu, vidVg = − hu, vi div g (X)dVg = hu, −div g (X)vidVg .
BR BR BR

Putting together all of the above we get

Z Z
h∇X u, vidVg = hu, (−∇X − div g (X))vidVg ,
RN RN

i.e.,
1 X p
∇∗X = −∇X − div g (X) = −∇X − p ∂i ( |g|X i ). (10.1.7)
|g| i
The lemma and consequently the proposition is proved. ⊔
⊓

Example 10.1.32. Let E be a rank r smooth vector bundle over the oriented Riemann
manifold (M, g), dim M = m. Let h·, ·i denote a Hermitian metric on E, and consider a
connection ∇ on E compatible with this metric. The connection ∇ defines a first order
p.d.o.
∇ : C ∞ (E) → C ∞ (T ∗ M ⊗ E).
The metrics g and h·, ·i induce a metric on T ∗ M ⊗ E. We want to describe the formal
adjoint of ∇ with respect to these choices of metrics.
As we have mentioned in the proof of the previous proposition, this is an entirely local
issue. We fix x0 ∈ M , and we denote by (x1 , . . . , xm ) a collection of g-normal coordinates
on a neighborhood U of x0 . We set

∇i := ∇∂xi .

Next, we pick a local synchronous frame of E near x0 , i.e., a local orthonormal frame (eα )
such that, at x0 , we have
∇i eα = 0 ∀i = 1, . . . , m.
10.1. PARTIAL DIFFERENTIAL OPERATORS: ALGEBRAIC ASPECTS 467

The adjoint of the operator

dxk ⊗ : C ∞ (E |U ) → C ∞ (T ∗ U ⊗ E |U ),

is the interior derivative (contraction) along the vector field g-dual to the 1-form dxk , i.e.,

(dxk ⊗)∗ = C k := g jk · i∂xj ).

Since ∇ is a metric connection, we deduce as in the proof of Lemma 10.1.31 that

∇k ∗ = −∇k − div g (∂xk ).

Hence, X X

∇∗ = ∇k ∗ ◦ C k = −∇k − div g (∂xk ) ◦ C k
k k
X p

=− (∇k + ∂xk log( |g|) ◦ C k . (10.1.8)
k

In particular, since at x0 we have ∂xk g = 0, we get

X
∇∗ |x0 = − ∇k ◦ C k .
k

The covariant Laplacian is the second order p.d.o.

∆ = ∆∇ : C ∞ (E) → C ∞ (E), ∆ = ∇∗ ∇.

To justify the attribute Laplacian we will show that ∆ is indeed a generalized Laplacian.
Using (10.1.8) we deduce that over U (chosen as above) we have
( )  
X p
 X 
∆=− ∇k + ∂xk log |g| ◦ C k ◦ dxj ⊗ ∇j
 
k j
nX p
o  
=− ∇k + ∂xk log |g| ◦ g kj · ∇j
k
X p

=− gkj ∇k + ∂xk gkj + gkj ∂xk log |g| ◦ ∇j
k,j
X 1 p

=− gkj ∇k ∇j + p ∂xk |g|gkj · ∇j .
k,j
|g|

The symbol of ∆ can be read easily from the last equality. More precisely,

σ2 (∆)(ξ) = −g jk ξj ξk = −|ξ|2g .

Hence ∆ is indeed a generalized Laplacian. ⊔

⊓

In the following exercise we use the notations in the previous example.

468 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Exercise 10.1.33. (a) Show that

1 p
gkℓ Γikℓ = − p ∂xk ( |g|gik ),
|g|

where Γikℓ denote the Christoffel symbols of the Levi–Civita connection associated to the
metric g.
(b) Show that
∗
∆∇ = − tr2g (∇T M ⊗E ∇E ),
∗
where ∇T M ⊗E is the connection on T ∗ M ⊗ E obtained by tensoring the Levi–Civita
connection on T ∗ M and the connection ∇E on E, while

tr2g : C ∞ (T ∗ M ⊗2 ⊗ E) → C ∞ (E)

denotes the double contraction by g,

tr2g (Sij ⊗ u) = g ij Sij u. ⊔

⊓

Proposition 10.1.34. Let E and M as above, and suppose that L ∈ P DO 2 (E) is a

generalized Laplacian. Then, there exists a unique metric connection ∇ on E and R =
R(L) ∈ End (E) such that
L = ∇∗ ∇ + R.
The endomorphism R is known as the Weitzenböck remainder of the Laplacian L.
Exercise 10.1.35. Prove the above proposition.
Hint Try ∇ defined by
f
∇f grad(h) u = {(∆g h)u − (ad(h)L)u} f, h ∈ C ∞ (M ), u ∈ C ∞ (E). ⊔
⊓
2
Exercise 10.1.36. (General Green’s formula). Consider a compact Riemannian man-
ifold (M, g) with boundary ∂M . Denote by ~n the unit outer normal along ∂M . Let
E, F → M be Hermitian vector bundles over M and suppose L ∈ PDOk (E, F ). Set
g0 = g |∂M , E0 = E |∂M and F0 = F |∂M . The Green formula states that there exists a
sesquilinear map
BL : C ∞ (E) × C ∞ (F ) → C ∞ (∂M ),
such that Z Z Z
hLu, vidv(g) = BL (u, v)dv(g0 ) + hu, L∗ vidv(g).
M ∂M M
Prove the following.
(a) If L is a zeroth order operator, then BL = 0.
(b) If L1 ∈ P DO(F, G), and L2 ∈ P DO(E, F ), then

BL1 L2 (u, v) = BL1 (L2 u, v) + BL2 (u, L∗1 v).

(c)
BL∗ (v, u) = −BL (u, v).
10.2. FUNCTIONAL FRAMEWORK 469

(d) Suppose ∇ is a Hermitian connection on E, and X ∈ Vect (M ). Then

B∇X (u, v) = hu, vig(X, ~n), B∇ (u, v) = hu, i~n viE ,

where i~n denotes the contraction by ~n.

(e) Denote by ~ν the section of T ∗ M |∂M g-dual to ~n. Suppose L is a first order p.d.o., and
set J := σL (~ν ). Then
BL (u, v) = hJu, viF . (10.1.9)
(f) Using (a)-(e) show that, for all u ∈ C ∞ (E), v ∈ C ∞ (F ), and any x0 ∈ ∂M , the
quantity BL (u, v)(x0 ) depends only on the jets of u, v at x0 of order at most k − 1. In
other words, if all the partial derivatives of u up to order (k − 1), with respect to some
connection, vanish along the boundary, then BL (u, v) = 0. ⊔
⊓

10.2 Functional framework

The partial differential operators are linear operators in infinite dimensional spaces, and
this feature requires special care in dealing with them. Linear algebra alone is not suffi-
cient. This is where functional analysis comes in.
In this section we introduce a whole range of functional spaces which are extremely
useful for most geometric applications.The presentation assumes the reader is familiar with
some basic principles of functional analysis. As a reference for these facts we recommend
the excellent monograph [23], or the very comprehensive [40, 137].

10.2.1 Sobolev spaces in RN

Let D denote an open subset of RN . We denote by L1loc (D) the space of locally integrable
real functions on D, i.e., Lebesgue measurable functions f : RN → R such that, ∀α ∈
C0∞ (RN ), the function αf is Lebesgue integrable, αf ∈ L1 (RN ).
Definition 10.2.1. Let f ∈ L1loc (D), and 1 ≤ k ≤ N . A function g ∈ L1loc (D) is said to
be the weak k-th partial derivative of f , and we write this g = ∂k f weakly, if
Z Z
gϕdx = − f ∂k ϕdx, ∀ϕ ∈ C0∞ (D). ⊔
⊓
D D

Lemma 10.2.2. Any f ∈ L1loc (D) admits at most one weak partial derivative. ⊔
⊓

The proof of this lemma is left to the reader.

☞ Not all locally integrable functions admit weak derivatives.
Exercise 10.2.3. Let f ∈ C ∞ (D), and 1 ≤ k ≤ N . Prove that the classical partial
derivative ∂k f is also its weak k-th derivative. ⊔
⊓

Exercise 10.2.4. Let H ∈ L1loc (R) denote the Heaviside function, H(t) ≡ 1, for t ≥ 0,
H(t) ≡ 0 for t < 0. Prove that H is not weakly differentiable. ⊔
⊓

Exercise 10.2.5. Let f1 , f2 ∈ L1loc (D). If ∂k fi = gi ∈ L1loc weakly, then ∂k (f1 + f2 ) =

g1 + g2 weakly. ⊔
⊓
470 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

The definition of weak derivative can be generalized to higher order derivatives as

follows. Consider a scalar p.d.o. L : C ∞ (D) → C ∞ (D), and f, g ∈ L1loc (D). Then we say
that Lf = g weakly if
Z Z
gϕdx = f L∗ ϕdx ∀ϕ ∈ C0∞ (D).
D D

Above, L∗ denotes the formal adjoint of L with respect to the Euclidean metric on RN .
Exercise 10.2.6. Let f, g ∈ C ∞ (D). Prove that

Lf = g classically ⇐⇒ Lf = g weakly. ⊔
⊓

Definition 10.2.7. Let k ∈ Z+ , and p ∈ [1, ∞]. The Sobolev space Lk,p (D) consists of
all the functions f ∈ Lp (D) such that, for any multi-index α satisfying |α| ≤ k, the mixed
partial derivative ∂ α f exists weakly, and moreover, ∂ α f ∈ Lp (D). For every f ∈ Lk,p(D)
we set  1/p
X Z
kf kk,p := kf kk,p,D =  |∂ α f |p dx ,
|α|≤k D

if p < ∞, while if p = ∞ X
kf kk,∞ = ess sup |∂ α f |.
|α|≤k

When k = 0 we write kf kp instead of kf k0,p . ⊔

⊓

Definition 10.2.8. Let k ∈ Z+ and p ∈ [1, ∞]. Set


Lk,p 1 k,p ∞
loc (D) := f ∈ Lloc (D) ; ϕf ∈ L (D) ∀ϕ ∈ C0 (D) . ⊔
⊓

Exercise 10.2.9. Let f (t) = |t|α , t ∈ R, α > 0. Show that for every p > 1 such that
α > 1 − 1p , we have
f (t) ∈ L1,p
loc (R). ⊔
⊓
Theorem 10.2.10. Let k ∈ Z+ and 1 ≤ p ≤ ∞. Then
(a) (Lk,p (RN ), k · kk,p ) is a Banach space.
(b) If 1 ≤ p < ∞ the subspace C0∞ (RN ) is dense in Lk,p (RN ).
(c) If 1 < p < ∞ the Sobolev space Lk,p(RN ) is reflexive.
The proof of this theorem relies on a collection of basic techniques frequently used in
the study of partial differential equations. This is why we choose to cover the proof of
this theorem in some detail. We will consider only the case p < ∞, leaving the p = ∞
situation to the reader.

Proof. Using the Exercise 10.2.5, we deduce that Lk,p is a vector space. From the classical
Minkowski inequality,
ν
!1/p ν
!1/p ν
!1/p
X X X
p p p
|xi + yi | ≤ |xi | + |yi | ,
i=1 i=1 i=1
10.2. FUNCTIONAL FRAMEWORK 471

we deduce that k · kk,p is a norm. To prove that Lk,p is complete, we will use the well
established fact that Lp is complete.
✍ To simplify the notations, throughout this chapter all the extracted subsequences will
be denoted by the same symbols as the sequences they originate from.
Let (fn ) ⊂ Lk,p (RN ) be a Cauchy sequence, i.e.,

lim kfm − fn kk,p = 0.

m,n→∞

In particular, for each multi-index |α| ≤ k, the sequence (∂ α fn ) is Cauchy in Lp (RN ), and
thus,
Lp
∂ α fn → gα , ∀|α| ≤ k.
Set f = limn fn . We claim that ∂ α f = gα weakly.
Indeed, for any ϕ ∈ C0∞ (RN ) we have
Z Z
α |α|
∂ fn ϕ dx = (−1) fn ∂ α ϕ.
RN RN

Since ∂ α fn → gα , fn → f in Lp , and ϕ ∈ Lq (RN ), where 1/q = 1 − 1/p, we conclude

Z Z
gα · ϕ dx = lim ∂ α fn · ϕ dx
RN n RN
Z Z
|α| α |α|
= lim(−1) fn · ∂ ϕ dx = (−1) f · ∂ α ϕ dx.
n RN RN
Part (a) is proved.
To prove that C0∞ (RN ) is dense we will use mollifiers. Their definition uses the
operation of convolution. Given f, g ∈ L1 (RN ), we define
Z
(f ∗ g)(x) := f (x − y)g(y)dy.
RN

We let the reader check that f ∗ g is well defined, i.e., y 7→ f (x − y)g(y) ∈ L1 , for almost
all x.
Exercise 10.2.11. (Young’s Inequality). If f ∈ Lp (RN ), and g ∈ Lq (RN ), 1 ≤ p, q ≤
∞, then f ∗ g is well defined, f ∗ g ∈ Lr (RN ), 1r = 1p + 1q − 1, and

kf ∗ gkr ≤ kf kp · kgkq . (10.2.1)

⊔
⊓

To define the mollifiers one usually starts with a function ρ ∈ C0∞ (RN ), such that
Z

ρ ≥ 0, supp ρ ⊂ |x| < 1 , and ρ dx = 1.
RN

Next, for each δ > 0, we define

ρδ (x) := δ−N ρ(x/δ).

472 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Note that Z

supp ρδ ⊂ |x| < δ and ρδ dx = 1.
RN
The sequence (ρδ ) is called a mollifying sequence. The next result describes the main use
of this construction.
Lemma 10.2.12. (a) For any f ∈ L1loc (RN ) the convolution ρδ ∗ f is a smooth1 function!
(b) If f ∈ Lp (RN ) (1 ≤ p < ∞) then
Lp
ρδ ∗ f → f as δ → 0.

Proof. Part (a) is left to the reader as an exercise in the differentiability of integrals with
parameters.
To establish part (b), we will use the fact that C0∞ (RN ) is dense in Lp (RN ). Fix ε > 0,
and choose g ∈ C0∞ (RN ) such that,

kf − gkp ≤ ε/3.

We have
kρδ ∗ f − f kp ≤ kρδ ∗ (f − g)kp + kρδ ∗ g − gkp + kg − f kp .
Using the inequality (10.2.1) we deduce

kρδ ∗ f − f kp ≤ 2kf − gkp + kρδ ∗ g − gkp . (10.2.2)

We need to estimate kρδ ∗ g − gkp . Note that

Z Z
ρδ ∗ g (x) = ρ(z)g(x − δz) dz and g(x) = ρ(z)g(x)dz.
RN RN

Hence Z
|ρδ ∗ g (x) − g(x)| ≤ ρ(z)|g(x − δz) − g(x)|dz ≤ δ sup |dg|.
RN
Since 
supp ρδ ∗ g ⊂ x + z ; x ∈ supp g ; z ∈ supp ρδ ,
there exists a compact set K ⊂ RN such that,

supp(ρδ ∗ g − g) ⊂ K ∀δ ∈ (0, 1).

We conclude that
Z 1/p
p p
kρδ ∗ g − gkp ≤ δ (sup |dg|) dx = vol (K)1/p δ sup |dg|.
K

If now we pick δ such that

vol (K)1/p δ sup |dg| ≤ ε/3,
1
This explains the term mollifier: ρδ smoothes out the asperities.
10.2. FUNCTIONAL FRAMEWORK 473

we conclude from (10.2.2) that

kρδ ∗ f − f kp ≤ ε.
The lemma is proved. ⊔
⊓

The next auxiliary result describes another useful feature of the mollification technique,
especially versatile in as far as the study of partial differential equations is concerned.

Lemma 10.2.13. Let f, g ∈ L1loc (RN ) such that ∂k f = g weakly. Then ∂k (ρδ ∗ f ) = ρδ ∗ g.
More generally, if X
L= aα ∂ α
|α|≤m

is a p.d.o with constant coefficients aα ∈ R, and Lf = g weakly, then

L(ρδ ∗ f ) = ρδ ∗ g classically.

Proof. It suffices to prove only the second part. We will write Lx to emphasize that L
acts via derivatives with respect to the variables x = (x1 , . . . , xN ). Note that
Z
L(ρδ ∗ f ) = (Lx ρδ (x − y))f (y)dy. (10.2.3)
RN

Since
∂
∂xi ρδ (x − y) = − ρδ (x − y),
∂y i
and ∂i ∗ = −∂i , we deduce from (10.2.3) that
Z Z
∗
L(ρδ ∗ f ) = (Ly ρδ (x − y))f (y)dy = ρδ (x − y)Ly f (y) dy
RN RN
Z
= ρδ (x − y)g(y)dy = ρδ ∗ g. ⊔
⊓
RN

Remark 10.2.14. The above lemma is a commutativity result. It shows that if L is a

p.d.o. with constant coefficients, then

[L, ρδ ∗]f = L(ρδ ∗ f ) − ρδ ∗ (Lf ) = 0.

This fact has a fundamental importance in establishing regularity results for elliptic op-
erators. ⊔
⊓

After this rather long detour, we return to the proof of Theorem 10.2.10. Let f ∈
Lk,p (RN ). We will construct fn ∈ C0∞ (RN ) such that fn → f in Lk,p using two basic
techniques: truncation and mollification (or smoothing).
Truncation. The essentials of this technique are contained in the following result.
474 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Lemma 10.2.15. Let f ∈ Lk,p (RN ). Consider for each R > 0 a smooth function ηR ∈
C0∞ (RN ) such that η(x) ≡ 1 for |x| ≤ R, ηR (x) ≡ 0 for |x| ≥ R + 1, and |dηR (x)| ≤ 2 ∀x.
Then ηR · f ∈ Lk,p (RN ), ∀R ≥ 0, and moreover
Lk,p
ηR · f −→ f as R → ∞.

Proof. We consider only the case k = 1. The general situation can be proved by induction.
We first prove that ∂i (ηR f ) exists weakly and as expected

∂i (ηR · f ) = (∂i ηR ) · f + ηR · ∂i f.

Let ϕ ∈ C0∞ (RN ). Since ηR ϕ ∈ C0∞ (RN ), we have

Z Z Z
(∂i ηR )ϕf + ηR · ∂i ϕf dx = ∂i (ηR ϕ)f = − ηR ϕ∂i f dx.
RN RN RN

This confirms our claim. Clearly (∂i ηR )f + ηR ∂i f ∈ Lp (RN ), so that ηR f is in L1,p (RN ).
Note that ∂i ηR ≡ 0 for |x| ≤ R and |x| ≥ R + 1. In particular, we deduce

(∂i ηR )f → 0 a.e.

Clearly |(∂i ηR )f | ≤ 2|f (x)|, so that by the dominated convergence theorem we conclude
Lp
(∂i ηR ) · f → 0 as R → ∞.

Similarly
ηR ∂i f → ∂i f a.e.,
and |ηR ∂i f | ≤ |∂i f |, which implies ηR ∂i f → ∂i f in Lp . The lemma is proved. ⊔
⊓

According to the above lemma, the space of compactly supported Lk,p -functions is
dense in Lk,p (RN ). Hence, it suffices to show that any such function can be arbitrarily
well approximated in the Lk,p -norm by smooth, compactly supported functions.
Let f ∈ Lk,p (RN ), and assume

ess supp f ⊂ {|x| ≤ R}.

Mollification. The sequence ρδ ∗ f converges to f in the norm of Lk,p as δ ց 0.

Note that each ρδ ∗ f is a smooth function, supported in { |x| ≤ R + δ }. According to
Lemma 10.2.13, we have

∂ α (ρδ ∗ f ) = ρδ ∗ (∂ α f ) ∀|α| ≤ k.

The desired conclusion now follows using Lemma 10.2.12.

To conclude the proof of Theorem 10.2.10, we need to show that Lk,p is reflexive if
1 < p < ∞. We will use the fact that Lp is reflexive for p in this range.
Note first that Lk,p (RN ) can be viewed as a closed subspace of the direct product
Y
Lp (RN ),
|α|≤k
10.2. FUNCTIONAL FRAMEWORK 475

via the map,

Y
T : Lk,p (RN ) → Lp (RN ), f 7→ (∂ α f )|α|≤k ,
|α|≤k

Lp
which is continuous, one-to-one, and has closed range. Indeed, if ∂ α fn → fα , then, arguing
as in the proof of completeness, we deduce that

fα = ∂ α f0 weakly,

where f0 = lim fn . Hence (fα ) = T f0 . We now conclude that Lk,p (RN ) is reflexive as a
closed subspace of a reflexive space. Theorem 10.2.10 is proved. ⊔
⊓

Remark 10.2.16. (a) For p = 2 the spaces Lk,2 (RN ) are in fact Hilbert spaces. The
inner product is given by
 
Z X
hu, vik =  ∂ α u · ∂ α vdx dx.
RN |α|≤k

(b) If D ⊂ RN is open, then Lk,p (D) is a Banach space, reflexive if 1 < p < ∞. However
C0∞ (D), is no longer dense in Lk,p(D). The closure of C0∞ (D) in Lk,p (D) is denoted by
Lk,p k,p k,p
0 (D). Intuitively, L0 (D) consists of the functions u ∈ L (D) such that

∂j u
= 0 on ∂D, ∀j = 0, 1, . . . , k − 1,
∂ν j

where ∂/∂ν denotes the normal derivative along the boundary. The above statement
should be taken with a grain of salt since at this point it is not clear how one can define
u |∂D when u is defined only almost everywhere. We refer to [4] for a way around this
issue.
The larger space C ∞ (D) ∩ Lk,p (D) is dense in Lk,p (D), provided that the boundary of
D is sufficiently regular. We refer again to [4] for details. ⊔
⊓

Exercise 10.2.17. Prove that the following statements are equivalent.

(a) u ∈ L1,p (RN ).
(b) There exists a constant C > 0 such that, for all ϕ ∈ C ∞ (RN ), we have
Z
∂ϕ
u ≤ CkϕkLp′ , ∀i = 1, . . . , N,
RN ∂xi

where p′ = p/(p − 1).

(c) There exists C > 0 such that, for all h ∈ RN , we have

k∆h ukLp ≤ C|h|,

where ∆h u(x) = u(x + h) − u(x). ⊔

⊓
476 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Exercise 10.2.18. Let f ∈ L1,p (RN ), and φ ∈ C ∞ (R), such that

|dφ| ≤ const,

and φ(f ) ∈ Lp . Then φ(f ) ∈ L1,p (RN ), and

∂i φ(f ) = φ′ (f ) · ∂i f. ⊔
⊓

Exercise 10.2.19. Let f ∈ L1,p (RN ). Show that |f | ∈ L1,p (RN ), and

∂i f a.e. on {f ≥ 0}
∂i |f | =
−∂i f a.e. on {f < 0}

Hint: Show that fε := (ε2 + f 2 )1/2 converges to f in L1,p as ε → 0. ⊔

⊓

10.2.2 Embedding theorems: integrability properties

The embedding theorems describe various inclusions between the Sobolev spaces Lk,p (RN ).
Define the “strength” of the Sobolev space Lk,p (RN ) as the quantity

σ(k, p) := σN (k, p) = k − N/p.

The “strength” is a measure of the size of a Sobolev size. Loosely speaking, the bigger the
strength, the more regular are the functions in that space, and thus it consists of “fewer”
functions.
Remark 10.2.20. The origin of the quantities σN (k, p) can be explained by using the
notion of conformal weight. A function u : RN → R can be thought of as a dimensionless
physical quantity. Its conformal weight is 0. Its partial derivatives ∂i u are physical quanti-
ties measured in meter −1 = variation per unit of distance, and they have conformal weight
−1. More generally, a mixed partial ∂ α u has conformal weight −|α|. The quantities |∂ α |p
have conformal weight −p|α|. The volume form dx is assigned conformal weight N : the
volume is measured in meter N . The integral of a quantity of conformal weight w is a
quantity of conformal weight w + N . For example, the quantity
Z
|∂ α u|p dx
RN

has conformal weight N − p|α|. In particular the quantity

 1/p
Z n X o
 |∂ α u|p dx
|α|=k

has conformal weight (N − kp)/p = −σN (k, p). Geometrically, the conformal weight is
captured by the behavior under the rescalings x = λy.
If we replace the Euclidean metric g on Rn with the metric gλ = λ2 g, λ > 0, then,

dVgλ = λn dVg , |du|gλ = λ−1 |du|g , ∀u ∈ C ∞ (RN ).

10.2. FUNCTIONAL FRAMEWORK 477

Equivalently, if we introduce the new linear variables y i related to the canonical Euclidean
variables xi by xi = λy i , then

∂xi = λ−1 ∂yi , dx1 ∧ · · · ∧ dxN = λN dy 1 ∧ · · · ∧ dy N . ⊔

⊓

Theorem 10.2.21 (Sobolev). If

σN (k, p) = σN (m, q) < 0 and k > m

then,
Lk,p (RN ) ֒→ Lm,q (RN ),
and the natural inclusion is continuous, i.e., there exists C = C(N, k, m, p, q) > 0 such
that
kf km,q ≤ Ckf kk,p ∀f ∈ Lk,p (RN ).

Proof. We follow the approach of [107] which relies on the following elementary, but inge-
nious lemma.
Lemma 10.2.22 (Gagliardo-Nirenberg). Let N ≥ 2 and f1 , . . . , fN ∈ LN −1 (RN −1 ). For
each x ∈ RN , and 1 ≤ i ≤ N define

ξi = (x1 , . . . , x̂i , . . . , , xN ) ∈ RN −1 .

Then
f (x) = f1 (ξ1 )f2 (ξ2 ) · · · fN (ξN ) ∈ L1 (RN ),
and moreover,
N
Y
kf k1 ≤ kfi kN −1 . ⊔
⊓
i=1

Exercise 10.2.23. Prove Lemma 10.2.22. ⊔

⊓

We first prove the theorem in the case k = 1, p = 1, which means m = 0, and

q = N/(N − 1). We will show that

∃C > 0 : kukN/(N −1) ≤ Ck |du| k1 ∀u ∈ C0∞ (RN ),

where |du|2 = |∂1 u|2 + · · · + |∂N u|2 . This result then extends by density to any u ∈
L1,1 (RN ).
We have
Z x1
|u(x1 , . . . , xN )| ≤ ∂i u(x1 , . . . , xi−1 , t, xi+1 , . . . , xN ) dt
−∞

Z x1
≤ ∂i u(x1 , . . . , xi−1 , t, xi+1 , . . . , xN ) dt =: gi (ξi ).
−∞
478 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Note that gi ∈ L1 (RN −1 ), so that

fi (ξi ) = gi (ξi )1/(N −1) ∈ LN −1 (RN −1 ).

Since
|u(x)|N/(N −1) ≤ f1 (ξ1 ) · · · fN (ξN ),
we conclude from Lemma 10.2.22 that u(x) ∈ LN/(N −1) (RN ), and

N
!1/N N
!1/N
Y Y
kukN/(N −1) ≤ kgi (ξi )k1 = k∂i uk1 .
1 1

Using the classical arithmetic-geometric means inequality, we conclude

N N
1 X const. X
kukN/(N −1) ≤ k∂i uk1 ≤ k |du| k1 .
N 1 N 1

We have thus proved that L1,1 (RN ) embeds continuously in LN/(N −1) (RN ).
Now let 1 < p < ∞ such that σN (1, p) = 1 − N/p < 0, i.e., p < N . We have to show
∗
that L1,p (RN ) embeds continuously in Lp (RN ), where
Np
p∗ = .
N −p
Let u ∈ C0∞ (RN ). Set v := |u|r−1 u, where r > 1 will be specified later. The inequality

N
!1/N
Y
kvkN/(N −1) ≤ ∂i v 1
1

implies
N
!1/N
Y
kukrrN/(N −1) ≤ r |u|r−1 ∂i u 1
.
1
If q = p/(p − 1) is the conjugate exponent of p, then using the Hölder inequality we get
r−1
|u|r−1 ∂i u ≤ ukq(r−1) k∂i u p
.

Consequently,
N
!1/N
Y
kukrrN/(N −1) ≤ r−1
rkukq(r−1) bigk ∂i u p
.
1
Now choose r such that rN/(N − 1) = q(r − 1). This gives
N −1
r = p∗ ,
N
and we get
N
!1/N
Y
kukp∗ ≤ r ∂i u p
≤ C(N, p)k |du| kp .
1
10.2. FUNCTIONAL FRAMEWORK 479

∗
This shows L1,p ֒→ Lp if 1 ≤ p < N . The general case

Lk,p ֒→ Lm,q if σN (k, p) = σN (m, q) < 0 k > m

follows easily by induction over k. We let the reader fill in the details. ⊔
⊓

Theorem 10.2.24 (Rellich-Kondrachov). Let (k, p), (m, q) ∈ Z+ × [1, ∞) such that

k > m and 0 > σN (k, p) > σN (m, q).

Then any bounded sequence (un ) ⊂ Lk,p (RN ) supported in a ball BR (0), R > 0 has a
subsequence strongly convergent in Lm,q (RN ).

Proof. We discuss only the case k = 1, so that the condition σN (1, p) < 0 imposes 1 ≤
p < N . We follow the elegant presentation in [23, Chap. IX]. The proof will be carried
out in several steps.
Step 1. For every h ∈ RN we define the translation operator

τh : Lq (RN ) → Lq (Rn ), τh f (x) = f (x + h), ∀f ∈ Lq (RN ), x ∈ RN .

Fix q ∈ [1, p∗ ), p∗ = N p/(N − p) and a radius R > 0. Let α ∈ (0, 1] be defined by

1 1−α
=α+ . (10.2.4)
q p∗
Then there exists a positive constant C, depending only on R, p, q, N such that

∀|h| ≤ 1, ∀u ∈ L1,p (Rn ) : kτh u − ukq,BR ≤ C|h|α kukL1,p (RN ) . (10.2.5)

Indeed, if u is a smooth function then

Z 1
|τh u(x) − u(x)| ≤ |h| |du(x + th)|dt.
0

Hence Z
kτh u − ukL1 (BR ) = τh u(x) − u(x) dx
BR
Z 1 Z 
≤ |h| |du(x + th)|dx dt
0 BR
Z 1 Z 1/p
= |h| |du(y)|p dy vol(BR+1 )1−1/p ≤ C|h|kuk1,p,RN ,
0 th+BR

where C = C(R, p, q, N ). Finally, using (10.2.4) and Hölder’s inequality we deduce

kτh u − ukLq (BR ) ≤ kτh u − ukαL1 (BR ) kτh u − uk1−α

Lp∗ (B
≤ C|h|α kuk1,p,RN .
R)

This proves the inequality (10.2.5) for smooth functions. The general case is obtained by
passing to the limit in the L1,p -norm.
480 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Step 2. Let U ⊂ L1,p (RN ) be a bounded subset a such that

ess supp u ⊂ {|x| ≤ R} ∀u ∈ U.

We will prove that, for every ε > 0 there exists 0 < δ = δ(ε) < 1, such that

kρδ ∗ u − ukq,BR ≤ ε, ∀δ ≤ δ(ε), ∀u ∈ U. (10.2.6)

Let u ∈ U. We have
Z
|ρδ ∗ u(x) − u(x)| ≤ |τ−y u(x) − u(x)|ρδ (y)dy
Bδ

Using the convexity of the function s 7→ sq , s ≥ 0 and the fact that ρδ (y)dy is a probability
measure we deduce
Z q  Z 
q
|τ−y u(x) − u(x)|ρδ (y)dy ≤ |τ−y u(x) − u(x)| ρδ (y)dy
Bδ Bδ

we deduce Z
|ρδ ∗ u(x) − u(x)|q ≤ |τ−y u(x) − u(x)|q ρδ (y)dy,
Bδ

so that
Z 1/q (10.2.5)
kρδ ∗ un − un kq,BR ≤ kτ−y un − un kqq,BR ρδ (y)dy ≤ Ckuk1,p,Rn δα .
Bδ

We now choose δ(ε) such that

Cδ(ε)α sup kuk1,p,Rn < ε.

u∈U

Step 3. Conclusion Suppose U ⊂ L1,p (Rn ) is a bounded of functions with supports

contained in BR . We need to prove that for any ε > 0 we can find a finite cover of U with
Lq (BR )-balls of radius ε.
To prove this, we set
Uδ = {uδ := ρδ ∗ u; u ∈ U},
and we show that Uδ is precompact in C 0 (BR+1 ). To achieve this we invoke the Arzéla-
Ascoli theorem and we will show that for any δ > 0 there exists C = C(δ) > 0, such
that
uδ (x) < C ∀u ∈ U, ∀|x| ≤ R + 1,
|uδ (x1 ) − uδ (x2 )| ≤ C|x1 − x2 | ∀u ∈ U, ∀|x1 |, |x2 | ≤ R + 1.
Indeed,
Z   Z
−N x−y −N
|ρδ ∗ u (x)| ≤ δ ρ |u(y)|dy ≤ δ |u(y)|dy
|y−x|≤δ δ Bδ (x)

∗ −1)/p∗
≤ C(N, p)δ−N kukp∗ · vol (Bδ )(p
10.2. FUNCTIONAL FRAMEWORK 481

≤ C(δ)kuk1,p by the Sobolev embedding theorem).

Similarly, Z
|uδ (x1 ) − uδ (x2 )| ≤ |ρδ (x1 − y) − ρδ (x2 − y)| · |u(y)|dy
BR+1
Z
≤ C(δ) · |x1 − x2 | |u(y)|dy ≤ C(δ) · |x1 − x2 | · kuk1,p .
BR+1

Now let δ ≤ δ(ε/2), where δ(ε/2) is defined by (10.2.6). Then the family Uδ is relatively
compact in C 0 (Br+1 ) and thus it can be covered by finitely many Lq (BR+1 ) balls of radius
ε/2. Using (10.2.6) we see that we can cover U with finitely many Lq (BR+1 )-balls of radius
ε. The compactness theorem is proved. ⊔
⊓

10.2.3 Embedding theorems: differentiability properties

A priori, the functions in the Sobolev spaces Lk,p are only measurable, and are defined
only almost everywhere. However, if the strength σN (k, p) is sufficiently large, then the
functions of Lk,p have a built-in regularity: each can be modified on a negligible set to
become continuous, and even differentiable.
To formulate our next results we must introduce another important family of Banach
spaces, namely the spaces of Hölder continuous functions.
Let α ∈ (0, 1). A function u : D ⊂ RN → R is said to be α-Hölder continuous if

[u]α := sup R−α osc (u; BR (z) ∩ D) < ∞,

0<R<1,z∈D

where for any set S ⊂ D, we denoted by osc (u; S) the oscillation of u on S, i.e.,

osc (u; S) := sup {|u(x) − u(y)| ; x, y ∈ S}.

Set
kuk∞,D := sup |u(x)|,
x∈D

and define 
C 0,α (D) := u : D → R ; kuk0,α,D := kuk∞ + [u]α < ∞ .
More generally, for every integer k ≥ 0, define

C k,α(D) := u ∈ C k (D); ∂ β u ∈ C 0,α (D), ∀|β| ≤ k .

The space C k,α(D) is a Banach space with respect to the norm,

X X
kukk,α := k∂ β uk∞,D + [∂ β u]α,D .
|β|≤k |β|=k

Define the strength of the Hölder space C k,α as the quantity

σ(k, α) := k + α.
482 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Theorem 10.2.25 (Morrey). Consider (m, p) ∈ Z+ × [1, ∞] and (k, α) ∈ Z+ × (0, 1)

such that m > k and σN (m, p) = σ(k, α) > 0. Then Lm,p (RN ) embeds continuously in
C k,α(RN ).

Proof. We consider only the case k = 1, and (necessarily) m = 0. The proof relies on the
following elementary observation.

Lemma 10.2.26. Let u ∈ C ∞ (BR ) ∩ L1,1 (BR ) and set

Z
1
u := u(x)dx.
vol (BR ) BR

Then Z
2N |du(y)|
|u(x) − u| ≤ dy. (10.2.7)
σ N −1 BR |x − y|N −1
In the above inequality σ N −1 denotes the “area” of the unit (N − 1)-dimensional round
sphere S N −1 ⊂ RN .

Proof.
Z |x−y|
∂ x−y
u(x) − u(y) = − u(x + rω)dr (ω = − ).
0 ∂r |x − y|
Integrating the above equality with respect to y we get,
Z Z |x−y|
∂
vol (BR )(u(x) − u) = − dy u(x + rω)dr.
BR 0 ∂r

If we set |∂r u(x + rω)| = 0 for |x + rω| > R, then

Z Z ∞
vol (BR )|u(x) − u| ≤ dy |∂r u(x + rω)|dr.
|x−y|≤2R 0

If we use polar coordinates (ρ, ω) centered at x, then dy = ρN −1 dρdω, where ρ = |x − y|,

and dω denotes the Euclidean “area” form on the unit round sphere. We deduce
Z ∞ Z Z 2R
vol (BR )|u(x) − u| ≤ dr dω |∂r u(x + rω)|ρN −1 dρ
0 S N−1 0
Z Z
(2R)N ∞
= dr |∂r (x + rω)|dωdr
N 0 S N−1
Z Z
(2R)N ∞ N −1 1
= r dr |∂ u(x + rω)|dω
N −1 r
N 0 S N−1 r
Z Z
(2R)N |∂r u(z)| (2R)N |du(z)|
(z = x + rω) = N −1
dz ≤ dz. ⊔
⊓
N BR |x − z| N BR |x − z|N −1
We want to make two simple observations.
10.2. FUNCTIONAL FRAMEWORK 483

1. In the above lemma, we can replace the round ball BR centered at origin by any
other ball, centered at any other point. In the sequel for any R > 0, and any x0 ∈ RN
we set Z
1
ux0 ,R := u(y)dy.
vol (BR (x0 )) BR (x0 )

2. The inequality (10.2.7) can be extended by density to any u ∈ L1,1 (BR ).

We will complete the proof of Morrey’s theorem in three steps.

Step 1. L∞ -estimates. We will show there exists C > 0 such that, ∀u ∈ L1,p (RN ) ∩
C 1 (RN )
kuk∞ ≤ Ckuk1,p .
For each x ∈ RN , denote by B(x) the unit ball centered at x, and set ux := ux,1 . Using
(10.2.7) we deduce
Z
|du(y)|
|u(x)| ≤ |ux | + CN N −1
dy
B(x) |x − y|
Z ! (10.2.8)
|du(y)|
≤ C kukp + N −1
dy .
B(x) |x − y|

Since σN (1, p) > 0, we deduce that p > N , so that its conjugate exponent q satisfies
p N
q= < .
p−1 N −1

7 |x − y|−(N −1) lies in Lq (B(x)), and

In particular, the function y →
Z
1
q(N −1)
dy ≤ C(N, q),
B(x) |x − y|

where C(N, q) is a universal constant depending only on N , and q. Using the Hölder
inequality in (10.2.8) we conclude

|u(x)| ≤ Ckuk1,p ∀x.

Step 2. Oscillation estimates. We will show there exists C > 0 such that for all u ∈
L1,p (RN ) ∩ C 1 (RN )
[u]α ≤ Ckuk1,p .
Indeed, from the inequality (10.2.7) we deduce that, for any ball BR (x0 ), and any x ∈
BR (x0 ), Z
|du(y)|
|u(x) − ux0 ,R | ≤ C N −1
BR (x0 ) |x − y|
p
(q = p−1 )

Z !1/q
(Hölder) 1
≤ Ck |du| kp · dy ≤ Ck |du| kp Rν ,
BR (x0 ) |x − y|q(N −1)
484 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

where
1
N
ν= N − q(N − 1) = 1 − = α.
q p
Hence
|u(x) − ux0 ,R | ≤ CRα ∀x ∈ BR (x0 ),
and consequently,
osc (u; BR (x0 )) ≤ CRα .
Step 2 is completed.
Step 3. Conclusion. Given u ∈ L1,p (RN ) we can find (un ) ∈ C0∞ (RN ) such that

un → u in L1,p and almost everywhere.

The estimates established at Step 1 and 2 are preserved as n → ∞. In fact, these es-
timates actually show that the sequence (un ) converges in the C 0,α -norm to a function
v ∈ C 0,α (RN ) which agrees almost everywhere with u. ⊔
⊓

Exercise 10.2.27. Let u ∈ L1,1 (RN ) satisfy a (q, ν)-energy estimate, i.e.
Z
1
∃C > 0 : N |du(y)|q dy ≤ C1 r −ν ∀x ∈ RN , 0 < r < 2,
r Br (x)

where 0 ≤ ν < q and q > 1. Show that (up to a change on a negligible set) the function
u is α-Hölder continuous, α = 1 − ν/q, and moreover

[u]α ≤ C2 ,

where the constant C2 depends only on N, q, ν and C1 .

Hint: Prove that Z
|du(y)|
N −1
≤ Cr α ,
Br (x) |x − y|
and then use the inequality (10.2.7). ⊔
⊓

Remark 10.2.28. The result in the above exercise has a suggestive interpretation. If u
satisfies the (q, ν)-energy estimate then although |du| may not be bounded, on average, it
“explodes” no worse that r α−1 as r → 0. Thus
Z |x|
|u(x) − u(0)| ≈ C tα−1 dt ≈ C|x|α .
0

The energy estimate is a very useful tool in the study of nonlinear elliptic equations. ⊔
⊓

The Morrey embedding theorem can be complemented by a compactness result. Let

(k, p) ∈ Z+ × [1, ∞] and (m, α) ∈ Z+ × (0, 1) such that

σN (k, p) > σ(m, α) k > m.

10.2. FUNCTIONAL FRAMEWORK 485

Then a simple application of the Arzela-Ascoli theorem shows that any bounded sequence
in Lk,p(RN ) admits a subsequence which converges in the C m,α norm on any bounded
open subset of RN .
The last results we want to discuss are the interpolation inequalities. They play an
important part in applications, but we chose not to include their long but elementary
proofs since they do not use any concept we will need later. The interested reader may
consult [4] or [15] for details.
Theorem 10.2.29 (Interpolation inequalities). For each R > 0 choose a smooth, cutoff
function ηR ∈ C0∞ (RN ) such that

ηR ≡ 1 if |x| ≤ R,

ηR ≡ 0 if |x| ≥ R + 1,
and
|dηR (x)| ≤ 2 ∀x ∈ RN .
Fix (m, p) ∈ Z+ × [1, ∞), and (k, α) ∈ Z+ × (0, 1).
(a) For every 0 < r ≤ R+1, there exists C = C(r, R, m, p) such that, for every 0 ≤ j < m,
ε > 0 and for all u ∈ Lm,p (RN ), we have

kηR ukj,p,RN ≤ CεkηR ukm,p,RN + Cε−j(m−j) kηR ukp,Br .

(b) For every 0 < r ≤ R + 1, there exists C = C(r, R, k, α) such that, for every 0 ≤ j < k,
ε > 0, and for all u ∈ C k,α(RN ), we have

kηR ukj,α,RN ≤ CεkηR ukk,α,RN + Cε−j(m−j) kηR uk0,α,Br . ⊔

⊓

The results in this, and the previous section extend verbatim to slightly more general
situations, namely to functions f : RN → H, where H is a finite dimensional complex
Hermitian space.

10.2.4 Functional spaces on manifolds

The Sobolev and the Hölder spaces can be defined over manifolds as well. To define these
spaces we need three things: an oriented Riemann manifold (M, g), a K-vector bundle
π : E → M endowed with a metric h = h•, •i, and a connection ∇ = ∇E compatible with
h. The metric g = (•, •) defines two important objects.

(i) The Levi–Civita connection ∇g .

(ii) A volume form dVg = ∗1. In particular, dVg defines a Borel measure on M . We de-
note by Lp (M, K) the space of K-valued p-integrable functions on (M, dVg ) (modulo
the equivalence relation of equality almost everywhere).

Definition 10.2.30. Let p ∈ [1, ∞]. An Lp -section of E is a Borel measurable map

ψ : M → E, i.e., ψ −1 (U ) is Borel measurable for any open subset U ⊂ E) such that the
following hold.
486 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

(a) π ◦ ψ(x) = x, for almost all x ∈ M except possibly a negligible set.

(b) The function x 7→ |ψ(x)|ph is integrable with respect to the measure defined by dVg .
⊔
⊓

The space of Lp -sections of E (modulo equality almost everywhere) is denoted by

Lp (E). We let the reader check the following fact.
Proposition 10.2.31. Lp (E) is a Banach space with respect to the norm
( R
1/p
p
kψkp,E = M |ψ(x)| dvg (x) if p < ∞
. ⊔
⊓
ess supx |ψ(x)| if p = ∞

Note that if p, q ∈ [1, ∞] are conjugate, 1/p+1/q = 1, then the metric h : E ×E → KM

defines a continuous pairing

h•, •i : Lp (E) × Lq (E) → L1 (M, K),

i.e., Z
hψ, φidVg ≤ kψkp,E · kφkq,E
M
This follows immediately from the Cauchy inequality


h ψ(x), φ(x) ≤ |ψ(x)| · |φ(x)| a.e. on M,

and the usual Hölder inequality.

Exercise 10.2.32. Let Ei → M (i = 1, . . . , k) be vector bundles with metrics and consider
a multilinear bundle map
Ξ : E1 × · · · × Ek → KM .
We regard Ξ as a section of E1∗ ⊗ · · · ⊗ Ek∗ . If

Ξ ∈ Lp0 (E1∗ ⊗ · · · ⊗ Ek∗ ),

then for every p1 , . . . , pk ∈ [1, ∞], such that

1 − 1/p0 = 1/p1 + · · · + 1/pk ,

and ∀ψj ∈ Lpj (Ej ), j = 1, . . . , k

Z
Ξ(ψ1 , . . . , ψk ) dVg ≤ kΞkp0 · kψ1 kp1 · · · kψk kpk . ⊔
⊓
M

For each m = 1, 2, . . ., define ∇m as the composition

∗ M ⊗E
∇E ∇T ∇
∇m : C ∞ (E) → C ∞ (T ∗ M ⊗ E) −→ · · · → C ∞ (T ∗ M ⊗m ⊗ E),

where we used the symbol ∇ to generically denote the connections in the tensor products
T ∗ M ⊗j ⊗E induced by ∇g and ∇E
The metrics g and h induce metrics in each of the tensor bundles T ∗ M ⊗m ⊗ E, and in
particular, we can define the spaces Lp (T ∗ M ⊗m ⊗ E).
10.2. FUNCTIONAL FRAMEWORK 487

Definition 10.2.33. (a) Let u ∈ L1loc (E) and v ∈ L1loc (T ∗ M ⊗m ⊗ E). We say that
∇m u = v weakly if
Z Z
hv, φidVg = hu, (∇m )∗ φidVg ∀u ∈ C0∞ (T ∗ M ⊗m ⊗ E).
M

(b) Define Lm,p (E) as the space of sections u ∈ Lp (E) such that, ∀j = 1, . . . , m, there
exist vj ∈ Lp (T ∗ M ⊗j ⊗ E), such that ∇j u = vj weakly. We set
m
X
kukm,p := kukm,p,E = k∇j ukp . ⊔
⊓
j=1

☞ A word of warning. The Sobolev space Lm,p (E) introduced above depends on several
choices: the metrics on M and E and the connection on E. When M is non-compact this
dependence is very dramatic and has to be seriously taken into consideration.
Example 10.2.34. Let (M, g) be the space RN endowed with the Euclidean metric. The
trivial line bundle E = RM is naturally equipped with the trivial metric and connection.
Then, Lp (RM ) = Lp (M, R).
Denote by D the Levi–Civita connection. Then, for every u ∈ C ∞ (M ), and m ∈ Z+ ,
we have X
Dmu = dx⊗α ⊗ ∂ α u,
|α|=m

where for every multi-index α we denoted by dx⊗α the monomial

dxα1 ⊗ · · · ⊗ dxαN .
The length of D m u(x) is
 1/2
X
 |∂ α u(x)|2  .
|α|=m

The space Lm,p (R M ) coincides, as a set, with the Sobolev space L

m,p (RN ). The norm

k • km,p,RM is equivalent with the norm k • km,p,RN introduced in the previous sections. ⊔
⊓

Proposition 10.2.35. (Lk,p (E), k · kk,p,E ) is a Banach space which is reflexive if 1 < p <
∞. ⊔
⊓

The proof of this result is left to the reader as an exercise.

The Hölder spaces can be defined on manifolds as well. If (M, g) is a Riemann manifold,
then g canonically defines a metric space structure on M , (see Chapter 4) and in particular,
we can talk about the oscillation of a function u : M → K. On the other hand, defining
the oscillation of a section of some bundle over M requires a little more work.
Let (E, h, ∇) as before. We assume that the injectivity radius ρM of M is positive. Set
ρ0 := min{1, ρM }. If x, y ∈ M are two points such that distg (x, y) ≤ ρ0 , then they can
be joined by a unique minimal geodesic γx,y starting at x, and ending at y. We denote by
Tx,y : Ey → Ex the ∇E -parallel transport along γx,y . For each ξ ∈ Ex , and η ∈ Ey we set
|ξ − η| := |ξ − Tx,y η|x = |η − Ty,x ξ|y .
488 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

If u : M → E is a section of E, and S ⊂ M has the diameter < ρ0 , we define


osc (u; S) := sup |u(x) − u(y)|; x, y ∈ S .

Finally set 
[u]α,E := sup r −α osc (u; Br (x)); 0 < r < ρ0 , x ∈ M .
For any k ≥ 0, we define
k
X
kukk,α,E := k∇j uk∞,E + [∇m u]α,T ∗ M ⊗m ⊗E ,
j=0

and we set 
C k,α(E) := u ∈ C k (E); kukk,α < ∞ .
Theorem 10.2.36. Let (M, g) be a compact, N -dimensional, oriented Riemann mani-
fold, and E a vector bundle over M equipped with a metric h, and compatible connection
∇. Then the following are true.
(a) The Sobolev space Lm,p (E), and the Hölder spaces C k,α(E) do not depend on the
metrics g, h and on the connection ∇. More precisely, if g1 is a different metric on M ,
and ∇1 is another connection on E compatible with some metric h1 then

Lm,p (E, g, h, ∇) = Lm,p (E, g1 , h1 , ∇1 ) as sets of sections,

and the identity map between these two spaces is a Banach space is continuous. A similar
statement is true for the Hölder spaces.
(b) If 1 ≤ p < ∞, then C ∞ (E) is dense in Lk,p (E).
(c) If (ki , pi ) ∈ Z+ × [1, ∞) (i = 0, 1) are such that

k0 ≥ k1 and σN (k0 , p0 ) = k0 − N/p0 ≥ k1 − N/p1 = σN (k1 , p1 ),

then Lk0 ,p0 (E) embeds continuously in Lk1 ,p1 (E). If moreover,

k0 > k1 and k0 − N/p0 > k1 − N/p1 ,

then the embedding Lk0 ,p0 (E) ֒→ Lk1 ,p1 (E) is compact, i.e., any bounded sequence of
Lk0 ,p0 (E) admits a subsequence convergent in the Lk1 ,p1 -norm.
(d) If (m, p) ∈ Z+ × [1, ∞), (k, α) ∈ Z+ × (0, 1) and

m − N/p ≥ k + α,

then Lm,p (E) embeds continuously in C k,α (E). If moreover

m − N/p > k + α,

then the embedding is also compact. ⊔

⊓

We developed all the tools needed to prove this theorem, and we leave this task to the
reader. The method can be briefly characterized by two key phrases: partition of unity
and interpolation inequalities. We will see them at work in the next section.
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS489

10.3 Elliptic partial differential operators: analytic aspects

This section represents the analytical heart of this chapter. We discuss two notions that
play a pivotal role in the study of elliptic partial differential equations. More precisely, we
will introduce the notion of weak solution, and a priori estimates.
Consider the following simple example. Suppose we want to solve the partial differen-
tial equation
∆u + u = f ∈ L2 (S 2 , R), (10.3.1)
where ∆ denotes the Laplace-Beltrami operator on the round sphere, ∆ = d∗ d. Riemann
suggested that one should look for the minima of the energy functional
Z n o
1
u 7−→ E(u) := (|du|2 + u2 ) − f u dVg .
S2 2

If u0 is a minimum of E, i.e.,
E(u0 ) ≤ E(u), ∀u,
then
Z
d 
0 = |t=0 E(u0 + tv) = (du0 , dv)g + u0 · v − f · v dVg ∀v. (10.3.2)
dt S2

Integrating by parts we get

Z
(d∗ du0 + u0 − f ) · v dVg = 0 ∀v,
S2

so that necessarily,
∆u0 + u0 = f.
There are a few grey areas in this approach, and Weierstrass was quick to point them
out: what is the domain of E, u0 may not exist, and if it does, it may not be C 2 , so that
the integration by parts is illegal etc. This avenue was abandoned until the dawns of the
twentieth century, when Hilbert reintroduced them into the spotlight, and emphasized the
need to deal with these issues.
His important new point of view was that the approach suggested by Riemann does
indeed produce a solution of (10.3.1) “provided if need be that the notion of solution be
suitable extended ”. The suitable notion of solution is precisely described in (10.3.2).
Naturally, one asks when this extended notion of solution coincides with the classical one.
Clearly, it suffices that u of (10.3.2) be at least C 2 , so that everything boils down to a
question of regularity.
Riemann’s idea was first rehabilitated in Weyl’s famous treatise [131] on Riemann
surfaces. It took the effort of many talented people to materialize Hilbert’s program
formulated as his 19th and 20th problem in the famous list of 27 problems he presented
at the Paris conference at the beginning of the 20th century. We refer the reader to [1]
for more details.
This section takes up the issues raised in the above simple example. The key fact which
will allow us to legitimize Riemann’s argument is the ellipticity of the partial differential
operator involved in this equation.
490 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

10.3.1 Elliptic estimates in RN

Let E = Cr denote the trivial vector bundle over RN . Denote by h•, •i the natural
Hermitian metric on E, and by ∂, the trivial connection. The norm of Lk,p (E) (defined
using the Euclidean volume) will be denoted by k • kk,p . Consider an elliptic operator of
order m X
L= Aα (x)∂xα : C ∞ (E) → C ∞ (E).
|α|≤m

For any k ∈ Z+ , and any R > 0 define

X
kLkk,R = sup k∂xβ (Aα (x))k.
|α|≤m,|β|≤k+m−|α| BR (0)

In this subsection we will establish the following fundamental result.

Theorem 10.3.1 (Interior elliptic estimates). (a) Let (k, p) ∈ Z+ × (1, ∞), and R > 0.
Then, there exists C = C(kLkk+1,R , k, p, N, R) > 0 such that, ∀u ∈ C0∞ (E |BR (0) ), we have


kukk+m,p ≤ C kLukk,p + kukp . (10.3.3)

(b) Let (k, α) ∈ Z+ ×(0, 1), and R > 0. Then, there exists C = C(kLkk+1,R , k, α, N, R) > 0
such that, ∀u ∈ C0∞ (E |BR (0) ), we have


kukk+m,α ≤ C kLukk,α + kuk0,α . (10.3.4)

Proof. The proof consists of two conceptually distinct parts. In the first part we establish
the result under the supplementary assumption that L has constant coefficients. In the sec-
ond part, the general result is deduced from the special case using perturbation techniques
in which the interpolation inequalities play an important role. Throughout the proof we
will use the same letter C to denote various constants C = C(kLkk+1,R , k, p, N, R) > 0
Step 1. Constant coefficients case. We assume L has the form
X
L= Aα ∂xα ,
|α|=m

where Aα are r × r complex matrices , independent of x ∈ RN . We set

X
kLk := kAα k.

We will prove the conclusions of the theorem hold in this special case. We will rely on a
very deep analytical result whose proof goes beyond the scope of this book.
For each f ∈ L1 (RN , C) denote by fˆ(ξ) its Fourier transform
Z
fˆ(ξ) := exp(−ix · ξ)dx.
RN
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS491

Theorem 10.3.2 (Calderon-Zygmund). Let m : S N −1 → C be a smooth function, and

define
m(ξ) : RN \ {0} → C
by  
ξ
m(ξ) = m .
|ξ|
Then the following hold.
(a) ThereR exist Ω ∈ C ∞ (S N −1 , C), and c ∈ C, such that
(a1 ) S N−1 ΩdvS N−1 = 0.
(a2 ) For any u ∈ C0∞ (RN ), the limit
Z
Ω(y)
(T u)(x) = lim u(x − y)dy
εց0 |y|≥ε |y|N

exists for almost every x ∈ RN , and moreover

Z
1
cu(x) + T u(x) = exp(ix · ξ)m(ξ)û(ξ)dξ.
(2π)N RN

(b) For every 1 < p < ∞ there exists C = C(p, kmk∞ ) > 0 such that,

kT ukp ≤ Ckukp ∀u ∈ C0∞ (RN ).

(c)(Korn-Lichtenstein). For every 0 < α < 1, and any R > 0, there exists C =
C(α, kmk0,α , R) > 0 such that ,∀u ∈ C00,α (BR )

[T u]α ≤ Ckuk0,α,BR . ⊔
⊓

For a proof of part (a) and (b) we refer to [125], Chap II §4.4. Part (c) is “elementary”
and we suggest the reader to try and prove it. In any case a proof, of this inequality can
be found in [15], Part II.5.
Let us now return to our problem. Assuming L has the above special form we will
prove (10.3.3). The proof of (10.3.4) is entirely similar and is left to the reader. We discuss
first the case k = 0.
Let u ∈ C0∞ (E |BR ). The function u can be viewed as a collection

u(x) = (u1 (x), . . . , ur (x)),

of smooth functions compactly supported in BR . Define

û(ξ) := (û1 (ξ), . . . , ûr (ξ)).

If we set v = Lu, then for any multi-index β such that |β| = m, we have

L∂ β u = ∂ β Lu = ∂ β v,
492 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

because L has constant coefficients. We Fourier transform the above equality, and we get
X
(−i)m Aα ξ α ∂d
β u(ξ) = (−i)m ξ β v̂ ( ξ). (10.3.5)
|α|=m

Note that since L is elliptic, the operator

X
σ(L)(ξ) = A(ξ) = Aα ξ α : Cr → Cr
|α|=m

is invertible for any ξ 6= 0. From (11.1.7) we deduce

∂\ ˆ ∀ξ 6= 0,
β u(ξ) = ξ β B(ξ)v(ξ)

where B(ξ) = A(ξ)−1 . Note that B(ξ) is homogeneous of degree −m, so that M (ξ) =
ξ β B(ξ) is homogeneous of degree 0. Thus, we can find functions mij (ξ) ∈ C ∞ (Rn \ {0}),
which are homogeneous of degree 0, and such that
X
∂d
β ui (ξ) = mij (ξ)v̂ j (ξ).
j

Using Theorem 10.3.2 (a) and (b) we deduce

k∂ β ukp ≤ Ckvkp = CkLukp .

This proves (10.3.3) when L has this special form.

Step 2. The general case. Suppose now that L is an arbitrary elliptic operator of order
m. Let r > 0 sufficiently small (to be specified later). Cover BR by finitely many balls
Br (xν ), and consider ην ∈ C0∞ (Br (xν )), such that each point in BR is covered by at most
10N of these balls, and X
ην ≥ 0, ην = 1,
ν

k∂ β ην k ≤ Cr −|β| ∀|β| ≤ m.
If u ∈ C0∞ (BR ), then
!
X X
v = Lu = L ην u = L(ην u).
ν ν

Set uν := ην u, and X
Lν := Aα (xν )∂ α .
|α|=m

def
We rewrite the equality vν = Luν as
X X
Lν uν = (Lν − L)uν + vν = εα,ν (x)∂ α uν − Aβ (x)∂ β uν + vν ,
|α|=m |β|<m
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS493

where εα,ν (x) = Aα (xν ) − Aα (x). Using (10.3.3) we deduce

X
kuν km,p ≤ C(kuν kp,Br (xν ) + kvν kp,Br (xν ) + kεα,ν (x)∂ α ukp,Br (xν ) + kukm−1,p,Br (xν ) ).
|α|=m

Since |εα,ν (x)| ≤ Cr on Br (xν ), where C = C(kLk1,R ), we deduce

kuν km,p,Br (xν ) ≤ C(rkuν kp,Br (xν ) + kuν kp,Br (xν ) + kvν kp,Br (xν ) ).

We can now specify r > 0 such that Cr < 1/2 in the above inequality. Hence

kuν km,p,Br (xν ) ≤ C(kuν kp,Br (xν ) + kvν kp,Br (xν ) ).

We need to estimate kvν kp . We use the equality

vν = L(ην u) = ην Lu + [L, ην ]u,

in which [L, ην ] = ad(ην )L is a p.d.o. of order m − 1, so that

k[L, ην ]ukp,Br (xν ) ≤ Cr −(m−1) kukm−1,p,Br (xν ) .

Hence
kvν kp,Br (xν ) ≤ C(kην Lukp,Br (xν ) + r −(m−1) kukm−1,p,Br (xν ) ),
so that
kuν km,p,Br (xν ) ≤ C(kLukp,BR + r −(m−1) kukm−1,p,BR ).
We sum over ν, and taking into account that the number of spheres Br (xν ) is O((R/r)N )
we deduce
X
kukm,p,BR ≤ kuν km,p ≤ CRN (r −N kLukp,BR + r −(m+N −1) kukm−1,p,BR ).
ν

Note that r depends only on R, p, kLk1,R , so that

kukm,p,BR ≤ C(kLukp,BR + kukm−1,p,BR ),

where C is as in the statement of Theorem 10.3.1.

We still need to deal with the term kukm−1,p,BR in the above inequality. It is precisely
at this point that the interpolation inequalities enter crucially.
View u as a section of C0∞ (E |B2R ). If we pick η ∈ C0∞ (B2R ) such that η ≡ 1 on BR ,
we deduce from the interpolation inequalities that there exists C > 0 such that

kukm−1,BR ≤ εkukm,p,R + Cε−(m−1) kukp,BR .

Hence
kukm,p,BR ≤ C(kLukp,BR + εkukm,p,BR + ε−(m−1) kukp,BR ).
If now we choose ε > 0 sufficiently small we deduce the case k = 0 of the inequality
(10.3.3).
494 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

To establish this inequality for arbitrary k we argue by induction. Consider a multi-

index |β| = k. If u ∈ C0∞ (E |BR ), and Lu = v then

L(∂ β u) = ∂ β Lu + [L, ∂ β ]u = ∂ β v + [L, ∂ β ]u.

The crucial observation is that [L, ∂ β ] is a p.d.o. of order ≤ m + k − 1. Indeed,

σm+k ([L, ∂ β ]) = [σm (L), σk (∂ β )] = 0.

Using (10.3.3) with k = 0, we deduce

k∂ β ukm,p,BR ≤ C(k∂ β vkp,BR + k[L, ∂ β ]ukp,BR + k∂ β ukp,BR )

≤ C(kvkk,p,BR + kukm+k−1,p,BR + kukk,p,BR ).

The term kukm+k−1,p,BR + kukk,p,BR can be estimated from above by

εkukm+k,p,BR + Ckukp,BR ,

using the interpolation inequalities as before. The inequality (10.3.3) is completely proved.
The Hölder case is entirely similar. It is left to the reader as an exercise. Theorem 10.3.1
is proved. ⊔
⊓

Using the truncation technique and the interpolation inequalities we deduce the fol-
lowing consequence.
Corollary 10.3.3. Let L as in Theorem 10.3.1, and fix 0 < r < R. Then, for every
k ∈ Z+ , 1 < p < ∞, and α ∈ (0, 1), there exists C = C(k, p, α, N, kLkk+1,R , R, r) > 0 such
that, ∀u ∈ C ∞ (E)

kukk+m,p,Br ≤ C kLukk,p,BR + kukp,BR ,
and

kukk+m,α,Br ≤ C kLukk,α,BR + kuk0,α,BR . ⊔
⊓

Exercise 10.3.4. Prove the above corollary. ⊔

⊓

10.3.2 Elliptic regularity

In this subsection we continue to use the notations of the previous subsection.
Definition 10.3.5. Let u, v : RN → Cr be measurable functions.
(a) The function u is a classical solution of the partial differential equation
X
Lu = Aα (x)∂ α u (x) = v(x) (10.3.6)
|β|≤m

0,α m,α
if there exists α ∈ (0, 1) such that, v ∈ Cloc (RN ), u ∈ Cloc (RN ), and (10.3.6) holds
everywhere.
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS495

(b) The function u is said to be an Lp -strong solution of (10.3.6) if u ∈ Lm,p N

loc (R ), v ∈
p
Lloc (RN ), and (10.3.6) hold almost everywhere. (The partial derivatives of u should be
understood in generalized sense.)
(c) The function u is said to be an Lp -weak solution if u, v ∈ Lploc (RN ) and
Z Z
∗
hu, L φidx = hv, φi dx, ∀φ ∈ C0∞ (E). ⊔
⊓
RN RN

Note the following obvious inclusion

{Lp − weak solutions} ⊃ {Lp − strong solutions}.

The principal result of this subsection states that when L is elliptic the above inclusion is
an equality.
Theorem 10.3.6. Let p ∈ (1, ∞) and suppose that L : C ∞ (E) → C ∞ (E) is an elliptic
operator of order m, with smooth coefficients. If u ∈ Lploc and Lu ∈ Lploc then u ∈ Lm,p
loc (E).

Remark 10.3.7. Loosely speaking the above theorem says that if a “clever” (i.e. elliptic)
combination of mixed partial derivatives can be defined weakly, then any mixed partial
derivative (up to a certain order) can be weakly defined as well. ⊔
⊓

Proof. 2 We prove only two special special cases of the theorem. Namely we assume that
L has either constant coefficients or it is a first order operator. The essential ingredient in
the proof is the technique of mollification. For each δ > 0 we denote by Tδ the operator
of convolution with ρδ . Fix a radius R > 0 and a partial differential operator S of order
m > 0. We first prove that there exists a constant C > 0 depending only on R, N, p, S
such that
k [S, Tδ ]u kp,BR+δ ≤ Ckukm−1,p,BR+1 , ∀u ∈ C0∞ (BR+1 ). (10.3.7)
Using the equality
[S1 S2 , Tδ ] = [S1 , Tδ ]S2 + S1 [S2 , Tδ ]
we deduce that if (10.3.7) holds for the operators S1 and S2 of positive orders then it also
holds for their product S1 S2 . Clearly (10.3.7) is trivially satisfied for operators S with
constant coefficients because in this case [S, Tδ ] = 0. Thus, it suffices to verify (10.3.7) for
first order operators of the form a(x)∂i + b(x).
Denote by Mb : C0∞ → C0∞ the operator of multiplication of by the smooth bundle
morphism b(x). Then
Z
[Mb , Tδ ]u(x) = ρδ (x − y)( b(x) − b(y) )u(y)dy,
RN

and the smoothness of b implies that there exists a constant C > 0 independent of δ > 0
such that
k [Mb , Tδ ] kp,BR+δ ≤ Ckukp,BR+1 , ∀u ∈ C0∞ (BR+1 ).
2
I would like to thank my friend Weimin Chen for pointing out a subtle error in the original proof.
496 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Next observe that

Z Z
[a∂i , Tδ ]u(x) = a(x)∂xi ρδ (x − y)u(u)dy − ρδ (x − y)a(y)∂yi u(y)dy
Rn Rn
Z  
1 x−y
= ∂ iρ a(x)u(y)dy
δN +1 Rn x δ
Z   Z
1 x−y
+ ∂yi ρ a(x)u(y)dy + ρδ (x − y)∂yi a(y)u(y)dy
δN +1 RN δ RN

(∂xi ρ(x − y) = −∂yi ρ(x − y))

Z  
1 x−y

= ∂yi ρ a(x) − a(y) u(y)dy
δN +1 |x−y|<δ δ
Z
+ ρδ (x − y)∂yi a(y)u(y)dy.
|x−y|<δ

The inequality (10.3.7) follows by observing that |a(x) − a(y)| = O(δ) when x, y ∈ BR+1 ,
|x − y| < δ.
We deduce that if L is a first order operator or L has constant coefficients there exists
a constant C > 0 depending only on R, N, q, L such that

k [Tδ , L∗ ]u kq,BR+δ ≤ Ckukq,BR+1 , ∀u ∈ C0∞ (BR+1 ). (10.3.8)

Suppose that u ∈ Lploc and v = Lu ∈ Lploc . Then for any φ ∈ C0∞ (BR+1 ) we have
Z Z
hLTδ u, φ idV = hTδ u, L∗ φidV
BR+1 BR+1
Z Z Z
∗ ∗
= hu, Tδ L φidV = hu, L Tδ φidV + u, [Tδ , L∗ ]φ
BR+1 BR+1 BR+1
Z Z
= hv, Tδ φidV + u, [Tδ , L∗ ]φ .
BR+1 BR+1
1 1
Let q ∈ (1, ∞)such that p + q = 1. Using (10.3.8) and Hölder inequality we deduce
Z
hLTδ u, φdV ≤ kvkp,BR+1 kTδ φkq,BR+1 + Ckukq,BR+1 kφkq,BR+1
BR+1

≤ C(kukp,BR+1 + kvkp,BR+1 kkφkq,BR+1 ,
for some constant C depending only on R, N, q, L. This implies that LTδ u is bounded in
Lp (BR+1 ). From the elliptic estimates we deduce that the familiy Tδ u it is bounded in
Lm,p (BR ) as δ → 0. Since the space Lm,p (BR ) is reflexive we deduce that a subsequence
Tδn u converges weakly as δn → 0 to a section u0 ∈ Lm,p (BR ). On the other hand, from the
Rellich-Kondrachov compactness theorem we deduce that Tδn u converges in the Lp -norm
to u0 . This implies that u0 = u so that u ∈ Lm,p
loc . ⊔
⊓
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS497

Remark 10.3.8. The proof of the general case of the regularity theorem requires consid-
erable more analytic work than we are willing to include in a geometry book. However,
the general idea is easy to describe.
Using the theory of pseudo-differential operators, one can construct a parametrix for
any elliptic operator L on RN T . This is an integral operator T : C0∞ (RN ) → C ∞ (RN )
which serves as a sort of inverse for L, i.e., the operators Sr = LT − 1 and Sℓ = T L − 1 are
smoothing operators. This means that they map locally integrable functions (and more
generally, distributions) to smooth functions. The convolution with ρδ is a prime example
of smoothing operator, while the convolution with the function |x|2−N is the parametrix
of the (geometers’) Laplacian on RN .
If Lu = v weakly then
T v = T Lu = u + Sℓ u.
Thus T v − u is a smooth function, and thus it suffices to prove that T v ∈ Lm,p
loc . This is
achieved, again using the theory of pseudo-differential operators, the Calderon-Zygmund
inequality, and the special case of Theorem 10.3.6 when L is a constant coefficients oper-
ator. For details we refer to [126]. ⊔
⊓

Corollary 10.3.9. If u ∈ Lploc (E) is weak Lp -solution of Lu = v, 1 < p < ∞, and

v ∈ Lk,p k+m,p
loc (E), then u ∈ Lloc (E), and for every 0 < r < R we have

kukm+k,p,Br ≤ C(kvkk,p,BR + kukp,BR ),

where as usual C = C(kLkk+m+1,R , R, r, . . .).

k+m,p
Proof. To prove that u ∈ Lloc (E) we argue by induction on k. The case k = 0 is
settled by Theorem 10.3.6. For the inductive step assume that u ∈ Lploc and Lu ∈ Lloc
k−1,p
k−1+m,p
implies that u ∈ Lloc . We want to prove that u ∈ Lploc and Lu ∈ Lk,ploc implies that
k+m,p p
u ∈ Lloc . Set v = Lu. The elliptic regularity implies that ∂i u ∈ Lloc . Moreover we
have the weak equality
∂i v = ∂i Lu = L∂i u + [∂i , L]u,
i.e.,
h∂i v, φi = hu, (L∂+i )∗ φi + hu, [∂i , L]∗ φi, ∀φ ∈ C0∞ .
To put it differently, the section ∂i u is a Lp -weak solution the equation

L(∂i u) = ∂i v − [∂i , L]u.

Observe that [∂i , L] is a p.d.o. of order ≤ m which means that the right hand side of
k−1,p k−1+m,p
the above equality is in Lloc since the inductive assumption implies that u ∈ Lloc .
k−1+m,p
Invoking the inductive assumption again we deduce that ∂i u ∈ Lloc , ∀i, i.e., u ∈
k+m,p
Lloc .
To prove the elliptic estimate pick a sequence un ∈ C0∞ (E) such that
k+m,p
un → u strongly in Lloc (E).
498 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Then
k+m,p
Lun → Lu strongly in Lloc (E),
and
kun km+k,p,Br ≤ C(kun k0,p,BR + kLun kk,p,BR ).
The desired estimate is obtained by letting n → ∞ in the above inequality. ⊔
⊓

Corollary 10.3.10 (Weyl Lemma). If u ∈ Lploc (E) is an Lp -weak solution of Lu = v, and

v is smooth, then u must be smooth.

Proof. Since v is smooth we deduce v ∈ Lk,p k+m,p

loc ∀k ∈ Z+ . Hence u ∈ Lloc ∀k. Using
m,α
Morrey embedding theorem we deduce that u ∈ Cloc ∀m ≥ 0. ⊔
⊓

The results in this, and the previous subsection are local, and so extend to the more
general case of p.d.o. on manifolds. They take a particularly nice form for operators on
compact manifolds.
Let (M, g) be a compact, oriented Riemann manifold and E, F → M be two metric
vector bundles with compatible connections. Denote by L ∈ P DOm (E, F ) an elliptic
operator of order m.

Theorem 10.3.11. (a) Let u ∈ Lp (E) and v ∈ Lk,p (F ), 1 < p < ∞, such that
Z Z
hv, φiF dvg = hu, L∗ φiE dVg ∀φ ∈ C0∞ (F ).
M M

Then u ∈ Lk+m,p (E), and

kukk+mp,E ≤ C(kvkk,p,F + kukp,E ),

where C = C(L, k, p).

(b) If u ∈ C m,α (E) and v ∈ C k,α (F ) (0 < α < 1) are such that Lu = v, then u ∈
C m+k,α(E), and
kukk+m,α,E ≤ C(kuk0,α,E + kvkk,α,F ),
where C = C(L, k, α).

Remark 10.3.12. The regularity results and the a priori estimates we have established
so far represent only the minimal information one needs to become an user of the elliptic
theory. Regrettably, we have mentioned nothing about two important topics: equations
with non-smooth coefficients, and boundary value problems. For generalized Laplacians
these topics are discussed in great detail in [51] and [102]. The boundary value problems
for first order elliptic operators require a more delicate treatment. We refer to [21] for a
very nice presentation of this subject. ⊔
⊓

Exercise 10.3.13. (Kato’s inequalities). Let (M, g) denote a compact oriented

Riemann manifold without boundary. Consider a metric vector bundle E → M equipped
with a compatible connection ∇.
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS499

(a) Show that for every u ∈ L1,2 (E), the function x 7→ |u(x)| is in L1,2 (M ), and moreover

|d |u(x)| | ≤ |∇u(x)|,

for almost all x ∈ M .

(b) Set ∆E = ∇∗ ∇, and denote by ∆M the scalar Laplacian. Show that, for all u ∈
L2,2 (E), we have
∆M (|u|2 ) = 2h∆E u, uiE − 2|∇u|2 .
Conclude that ∀φ ∈ C ∞ (M ) such that φ ≥ 0, we have
Z Z
(d|u|, d(φ|u|) )g dVg ≤ h∆E u, uiE φ dVg ,
M M

i.e.,
|u(x)|∆M (|u(x)|) ≤ h∆E u(x), u(x)iE weakly. ⊔
⊓
Exercise 10.3.14. (Local estimates). Suppose (M, g) is a connected, oriented, Rie-
mann manifolds of dimension m, not necessarily, E0 , E1 → M are two smooth hermitian
vector bundles on M , and ∇i , i = 0, 1 are Hermitian connection on Ei .
Suppose that L : C ∞ (E0 ) → C ∞ (E1 ) is an elliptic operator of order ν with smooth
coefficients. Prove that for any compact subset K ⊂ M , any relatively compact open
neighborhood U of K, and any p ∈ (1, ∞), ℓ ∈ Z≥0 , there exists a positive constat C
depending only on K, U, p, ℓ, ∇0 , ∇1 , and the coefficients of L such that, ∀u ∈ C ∞ (E0 ),
we have

kukLℓ+ν,p (K,E0) ≤ C kLukLℓ,p (U,E1 ) + kukLp (U,E0 ) ,
where the above Sobolev norms are defined in terms of the connections ∇i . ⊔
⊓

10.3.3 An application: prescribing the curvature of surfaces

In this subsection we will illustrate the power of the results we proved so far by showing how
they can be successfully used to prove an important part of the celebrated uniformization
theorem. In the process we will have the occasion to introduce the reader to some tricks
frequently used in the study of nonlinear elliptic equations. We will consider a slightly
more general situation than the one required by the uniformization theorem.
Let (M, g) be a compact, connected, oriented Riemann manifold of dimension N .
Denote by ∆ = d∗ d : C ∞ (M ) → C ∞ (M ) the scalar Laplacian. We assume for simplicity
that Z
volg (M ) = dVg = 1,
M
so that the average of any integrable function ϕ is defined by
Z
ϕ= ϕ(x)dVg (x).
M

We will study the following partial differential equation.

∆u + f (u) = s(x), (10.3.9)

500 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

where f : R → R, and s ∈ C ∞ (M ) satisfy the following conditions.

(C1 ) The function f is smooth and strictly increasing.
(C2 ) There exist a > 0 and b ∈ R, such that
f (t) ≥ at + b ∀t ∈ R.
Set Z u
F (u) = f (t)dt.
0
We assume
(C3 ) lim|t|→∞ (F (t) − st) = ∞, where s denotes the average of s,
Z
s= s(x)dVg .
M
Theorem 10.3.15. Let f and s(x) satisfy the conditions (C1 ), (C2 ), C3 ). Then there
exists a unique u ∈ C ∞ (M ) such that
∆u (x) + f (u(x)) = s(x) ∀x ∈ M.
Proof. The proof of this theorem will be carried out in two steps.
Step 1. Existence of a weak solution. A weak solution of (10.3.9) is a function u ∈ L1,2 (M )
such that f (u(x)) ∈ L2 (M ), and
Z Z
{(du, dφ) + f (u)φ}dVg = s(x)φ(x)dVg (x), ∀φ ∈ L1,2 (M ).
M M
Step 2. Regularity. We show that a weak solution is in fact a classical solution.
Proof of Step 1. We will use the direct method of the calculus of variations outlined at
the beginning of the current section. Consider the energy functional
Z
1
I : L1,2 (M ) → R, I(u) = |du(x)|2 + F (u(x)) − g(x)u(x) dVg (x).
M 2
This functional is not quite well defined, since there is no guarantee that F (u) ∈ L1 (M )
for all u ∈ L1,2 (M ).
Leaving this issue aside for a moment, we can perform a formal computation à la
Riemann. Assume u is a minimizer of I, i.e.,
I(u) ≤ I(v), ∀v ∈ L1,2 (M ).
Thus, for all φ ∈ L1,2 (M )
I(u) ≤ I(u + tφ) ∀t ∈ R.
Hence t = 0 is a minimum of hφ (t) = I(u+tφ), so that h′φ (0) = 0 ∀φ. A simple computation
shows that Z
h′φ (0) = {(du, dφ) + f (u)φ − s(x)φ(x)}dVg = 0,
M
so that a minimizer of I is a weak solution provided that we deal with the integrability
issue raised at the beginning of this discussion. Anyway, the lesson we learn from this
formal computation is that minimizers of I are strong candidates for solutions of (10.3.9).
We will circumvent the trouble with the possible non-integrability of F (u) by using a
famous trick in elliptic partial differential equations called the maximum principle.
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS501

Lemma 10.3.16. Let h : R → R be a continuous, strictly increasing function and u, v ∈

L1,2 (M ) such that
(i) h(u), h(v) ∈ L2 (M ).
(ii) ∆u + h(u) ≥ ∆v + h(v) weakly i.e.
Z Z
{(du, dφ) + h(u)φ}dVg ≥ {dv, dφ) + h(v)φ}dVg (10.3.10)
M M

∀φ ∈ L1,2 (M ) such that φ ≥ 0 a.e. M . Then u ≥ v a.e. on M .

Proof. Let
1
(u − v)− = min{u − v, 0) =
{(u − v) − |u − v|}.
2
According to the Exercise 10.2.19 we have (u − v)− ∈ L1,2 (M ), and

− d(u − v) a.e. on {u < v}
d(u − v) =
0 a.e. on {u ≥ v}

Using φ = −(u − v) in (10.3.10) we deduce

Z



−
− d(u − v), d(u − v)− + h(u) − h(v) u − v dVg ≥ 0.
M

Clearly

d(u − v), d(u − v)− = |d(u − v)− |2 ,
and since h is nondecreasing,

(h(u) − h(v))(u − v)− ≥ 0.

Hence, Z Z
− 2
|d(u − v) | dVg ≤ − ( h(u) − h(v) )(u − v)− dVg ≤ 0,
M M
so that
|d(u − v)− | ≡ 0.
Since M is connected, this means (u − v)− ≡ c ≤ 0. If c < 0, then

(u − v) ≡ (u − v)− ≡ c,

so that u = v + c < v. Since h is strictly increasing we conclude

Z Z
h(u) dVg < h(v)dVg .
M M

On the other hand, using φ ≡ 1 in (10.3.10), we deduce

Z Z
h(u)dVg ≥ h(v)dVg !
M M

Hence c cannot be negative, so that (u − v)− ≡ 0 which is another way of saying u ≥ v.

The maximum principle is proved. ⊔
⊓
502 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Exercise 10.3.17. Assume h in the above lemma is only non-decreasing, but u and v
satisfy the supplementary condition
u ≥ v.
Show the conclusion of Lemma 10.3.16 continues to hold. ⊔
⊓

We now return to the equation (10.3.9). Note first that if u and v are two weak
solutions of this equation, then

∆u + f (u) ≥ (≤) ∆v + h(v) weakly,

so that by the maximum principle u ≥ (≤) v. This shows the equation (10.3.9) has at
most one weak solution.
To proceed further we need the following a priori estimate.
Lemma 10.3.18. Let u be a weak solution of (10.3.9). If C is a positive constant such
that
f (C) ≥ sup s(x),
M

then u(x) ≤ C a.e. on M .

Proof. The equality f (C) ≥ sup s(x) implies

∆C + f (C) ≥ s(x) = ∆u + f (u) weakly

so the conclusion follows from the maximum principle. ⊔

⊓

Fix C0 > 0 such that f (C0 ) ≥ sup s(x). Consider a strictly increasing C 2 -function
f˜ : R → R such that
f˜(u) = f (u) for u ≤ C0 ,
f˜(u) is linear for u ≥ C0 + 1.
The conditions (C1 ) and (C2 ) imply that there exist A, B > 0 such that

|f˜(u)| ≤ A|u| + B. (10.3.11)

Lemma 10.3.19. If u is a weak solution of

∆u + f˜(u) = s(x) (10.3.12)

then u is also a weak solution of (10.3.9).

Proof. We deduce as in the proof of Lemma 10.3.18 that u ≤ C0 , which is precisely the
range where f coincides with f˜. ⊔
⊓

The above lemma shows that instead of looking for a weak solution of (10.3.9), we
should try to find a weak solution of (10.3.12). We will use the direct method of the
calculus of variations on a new functional
Z
1
˜
I(u) = |du(x)|2 + F̃ (u(x)) − s(x)u(x) dVg (x),
M 2
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS503

where Z u
F̃ (u) := f˜(t)dt.
0

The advantage we gain by using this new functional is clear. The inequality (10.3.11)
shows F̃ has at most quadratic growth, so that F̃ (u) ∈ L1 (M ), for all u ∈ L2 (M ). The
existence of a minimizer is a consequence of the following fundamental principle of the
calculus of variations.
Proposition 10.3.20. Let X be a reflexive Banach space and J : X → R a convex,
weakly lower semi-continuous, coercive functional, i.e., the sublevel sets

J c = {x ; J(x) ≤ c}

are respectively convex, weakly closed and bounded in X. Then J admits a minimizer, i.e.,
there exists x0 ∈ X such that

J(x0 ) ≤ J(x) ∀x ∈ X.

Proof. Note that

inf J(x) = infc J(x)
X J

Consider xn ∈ J c such that

lim J(xn ) = inf J.
n

Since X is reflexive, and Jc is convex, weakly closed and bounded in the norm of X, we
deduce that J c is weakly compact. Hence a (generalized) subsequence (xν ) of xn converges
weakly to some x0 ∈ J c . Using the lower semi-continuity of J we deduce

J(x0 ) ≤ lim inf J(xν ) = inf J.

ν

Hence x0 is a minimizer of J. The proposition is proved. ⊔

⊓

The next result will conclude the proof of Step 1.

Lemma 10.3.21. I˜ is convex, weakly lower semi-continuous and coercive (with respect to
the L1,2 -norm).

Proof. The convexity is clear since F̃ is convex on account that f˜ is strictly increasing.
The functionals
Z Z
1 2
u 7→ |du| dVg and u 7→ − s(x)u(x)dVg (x)
2 M M

are clearly weakly lower semi-continuous. We need to show the functional

Z
u 7→ F̃ (u)dVg
M

is also weakly lower semi-continuous.

504 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Since F̃ is convex, we can find α > 0, β ∈ R such that

F̃ (u) − αu − β ≥ 0.

If un → u strongly in L1,2 (M ) we deduce using the Fatou lemma that

Z Z
{F̃ (u) − αu − β}dVg ≤ lim inf {F̃ (un ) − αun − β}dVg .
M n→∞ M

On the other hand, Z Z

lim αun + βdVg = αu + βdVg ,
n M M

which shows that Z Z

F̃ (u)dVg ≤ lim inf F̃ (un )dVg .
M n

This means that the functional

Z
1,2
L (M ) ∋ u 7→ F̃ (u)dVg
M

is strongly lower semi-continuous. Thus, the sublevel sets

n Z o
u; F̃ (u)dVg ≤ c ,
M

are both convex and strongly closed. The Hahn-Banach separation principle can now be
invoked to conclude these level sets are also weakly closed. We have thus established that
I˜ is convex and weakly lower semi-continuous.

Remark 10.3.22. We see that the lower semi-continuity, and the convexity conditions
are very closely related. In some sense they are almost equivalent. We refer to [33] for
a presentation of the direct method of the calculus of variations were the lower semi-
continuity issue is studied in great detail. ⊔
⊓

The coercivity will require a little more work. The key ingredient will be a Poincaré
inequality. We first need to introduce some more terminology.
For any u ∈ L2 (M ) we denoted by u its average. Now set

u⊥ (x) = u(x) − u.

Note the average of u⊥ is 0. This choice of notation is motivated by the fact that u⊥ is
perpendicular (with respect to the L2 (M )-inner product) to the kernel of the Laplacian
∆ which is the 1-dimensional space spanned by the constant functions.

Lemma 10.3.23 (Poincaré inequality). There exists C > 0 such that

Z Z
|du|2 dVg ≥ C |u⊥ |2 dVg , ∀u ∈ L1,2 (M ).
M M
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS505

Proof. We argue by contradiction. Assume that for any ε > 0 there exists uε ∈ L1,2 (M ),
such that Z Z
|u⊥ε |2
dV g = 1, and |duε |2 dVg ≤ ε.
M M

The above two conditions imply that the family (u⊥

ε) is bounded in L1,2 (M ). Since L1,2 (M )
is reflexive, we deduce that, on a subsequence

u⊥ 1,2
ε ⇀ v, weakly in L (M ).

The inclusion L1,2 (M ) ֒→ L2 (M ) is compact, (Rellich-Kondrachov) so that, on a subse-

quence
u⊥ 2
ε → v strongly in L (M ).

This implies v 6= 0, since

Z Z
2
|v| dVg = lim |u⊥ 2
ε | dVg = 1.
M ε M

On the other hand,

duε = du⊥ 2
ε → 0, strongly in L .

We conclude
Z Z
(dv, dφ) dVg = lim (duε , dφ) dVg = 0 ∀φ ∈ L1,2 (M ).
M ε M

In particular, Z
(dv, dv) dVg = 0,
M
so that dv ≡ 0. Since M is connected, we deduce v ≡ c = const. Moreover,
Z Z
c= v(x)dVg (s) = lim u⊥ε dVg = 0.
M ε M

This contradicts the fact that v 6= 0. The Poincaré inequality is proved. ⊔

⊓

˜
We can now establish the coercivity of I(u). Let κ > 0, and u ∈ L1,2 (M ) such that
Z n o
1
|du(x)|2 + F̃ (u(x)) − s(x)u(x) dVg (x) ≤ κ. (10.3.13)
M 2

Since Z Z Z Z
2 ⊥ 2 2 2
|du| dVg = |du | dVg and |u| = |u| + |u⊥ |2 dVg ,
M M M M
it suffices to show that the quantities
Z Z
⊥ 2
u, |u | dVg , |du⊥ |dVg
M M

are bounded. In view of the Poincaré inequality, the boundedness of

Z
|du⊥ |2 dVg
M
506 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

implies the boundedness of ku⊥ k2,M , so that we should concentrate only on u and kdu⊥ k2 .
The inequality (10.3.13) can be rewritten as
Z n o
1 ⊥2
|du | + F̃ (u) − s⊥ u⊥ dVg − s · u ≤ c.
M 2

The Poincaré and Cauchy inequalities imply

Z
Cku⊥ k22 ⊥ ⊥
− ks k2 · ku k2 + F̃ (u)dVg − s · u ≤ κ.
M

Since volg (M ) = 1, and F̃ is convex, we have a Jensen-type inequality

Z  Z
F̃ udVg ≤ F̃ (u)dVg ,
M M

so that
Cku⊥ k22 − ks⊥ k2 · ku⊥ k2 + F̃ (u) − s · u ≤ κ. (10.3.14)
Set P (t) := Ct2 − ks⊥ k2 t, and let m = inf P (t). From the inequality (10.3.14) we deduce

F̃ (u) − s · u ≤ κ − m.

Using condition (C3 ) we deduce that |u| must be bounded. Feed this information back
in (10.3.14). We conclude that P (ku⊥ k2 ) must be bounded. This forces ku⊥ k2 to be
bounded. Thus I˜ is coercive, and Lemma 10.3.21 is proved. ⊔
⊓

Step 2. The regularity of the minimizer. We will use a technique called bootstrapping,
which blends the elliptic regularity theory, and the Sobolev embedding theorems, to grad-
ually improve the regularity of the weak solution.
Let u be the weak solution of (10.3.12). Then u(x) is a weak L2 solution of

∆u = h(x) on M,

where h(x) = −f˜(u(x)) − s(x). Note that since the growth of f˜ is at most linear, we have
f˜(u(x)) ∈ L2 (M ). The elliptic regularity theory implies that u ∈ L2,2 (M ). Using Sobolev
(or Morrey) embedding theorem we can considerably improve the integrability of u. We
deduce that
(i) either u ∈ Lq (M ) if −N/q ≤ 2 − N/2 ≤ 0 (dim M = N ),
(ii) or u is Hölder continuous if 2 − N/2 > 0.
In any case, this shows u(x) ∈ Lq1 (M ), for some q1 > 2, which implies h(x) ∈ Lq1 (M ).
Using again elliptic regularity we deduce u ∈ L2,q1 and Sobolev inequality implies that
u ∈ Lq2 (M ) for some q2 > q1 .
After a finite number of steps we conclude that h(x) ∈ Lq (M ), for all q > 1. Elliptic
regularity implies u ∈ L2,q (M ), for all q > 1. This implies h(x) ∈ L2,q (M ) ,for all q > 1.
Invoking elliptic regularity again we deduce that u ∈ L4,q (M ), for any q > 1. (At this
point it is convenient to work with f , rather than with f˜ which was only C 2 ). Feed this
back in h(x), and regularity theory improves the regularity of u two orders at a time. In
10.3. ELLIPTIC PARTIAL DIFFERENTIAL OPERATORS: ANALYTIC ASPECTS507

view of Morrey embedding theorem the conclusion is clear: u ∈ C ∞ (M ). The proof of

Theorem 10.3.15 is complete. ⊔
⊓

From the theorem we have just proved we deduce immediately the following conse-
quence.
Corollary 10.3.24. Let (M, g) be a compact, connected, oriented Riemann manifold and
s(x) ∈ CR∞ (M ). Assume volg (M ) = 1. Then the following two conditions are equivalent.
(a) s = M s(x)dVg > 0

(b) For every λ > 0 there exists a unique u = uλ ∈ C ∞ (M ) such that

∆u + λeu = s(x). (10.3.15)

Proof. (a) ⇒ (b) follows from Theorem 10.3.15.

(b) ⇒ (a) follows by multiplying (10.3.15) with v(x) ≡ 1, and then integrating by parts
so that Z
s̄ = λ eu(x) dVg (x) > 0. ⊔
⊓
M
Although the above corollary may look like a purely academic result, it has a very
nice geometrical application. We will use it to prove a special case of the celebrated
uniformization theorem.
Definition 10.3.25. Let M be a smooth manifold. Two Riemann metrics g1 and g2 are
said to be conformal if there exists f ∈ C ∞ (M ), such that g2 = ef g1 . ⊔
⊓

Exercise 10.3.26. Let (M, g) be an oriented Riemann manifold of dimension N and

f ∈ C ∞ (M ). Denote by g̃ the conformal metric g̃ = ef g. If s(x) is the scalar curvature of
g and s̃ is the scalar curvature of g̃ show that
 
−f (N − 1)(N − 2) 2
s̃(x) = e s(x) + (N − 1)∆g f − |df (x)|g , , (10.3.16)
4
where ∆g denotes the scalar Laplacian of the metric g while | · |g denotes the length
measured in the metric g. ⊔
⊓

Remark 10.3.27. A long standing problem in differential geometry, which was only
relatively recently solved, is the Yamabe problem:
“If (M, g) is a compact oriented Riemann manifold, does there exist a metric conformal
to g whose scalar curvature is constant? ”
In dimension 2 this problem is related to the uniformization problem in complex anal-
ysis.
The solution of the Yamabe problem in its complete generality is due to the combined
efforts of T. Aubin, [10, 11] and R. Schoen [118]. For a very beautiful account of its proof
we recommend the excellent survey of J. Lee and T.Parker, [85]. One can formulate a
more general question than the Yamabe problem.
Given a compact oriented Riemann manifold (M, g) decide whether a smooth function
s(x) on M is the scalar curvature of some metric on M conformal to g.
508 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

This problem is known as the Kazdan-Warner problem. The case dim M = 2 is com-
pletely solved in [76]. The higher dimensional situation dim M > 2 is far more complicated,
both topologically and analytically. ⊔
⊓

Theorem 10.3.28 (Uniformization Theorem). Let (Σ, g) be a compact, oriented Riemann

manifold of dimension 2. Assume volg (Σ) = 1. If χ(Σ) < 0 (or equivalently if its genus
is ≥ 2) then there exists a unique metric g̃ conformal to g such that

s(g̃) ≡ −1.

Proof. We look for g̃ of the form g̃ = eu g. Using Exercise 10.3.26 we deduce that u should
satisfy
−1 = e−u {s(x) + ∆u},

i.e.
∆u + eu = −s(x),

where s(x) is the scalar curvature of the metric g. The Gauss–Bonnet theorem implies
that
s = 4πχ(Σ) < 0

so that the existence of u is guaranteed by Corollary 10.3.24. The uniformization theorem

is proved. ⊔
⊓

On a manifold of dimension 2 the scalar curvature coincides up to a positive factor

with the sectional curvature. The uniformization theorem implies that the compact ori-
ented surfaces of negative Euler characteristic admit metrics of constant negative sectional
curvature. Now, using the Cartan-Hadamard theorem we deduce the following topological
consequence.

Corollary 10.3.29. The universal cover of a compact, oriented surface Σ of negative

Euler characteristic is diffeomorphic to C. ⊔
⊓

Remark 10.3.30. The “inverse” of the covering projection π : C → Σ is classically called

a uniformizing parameter. It is very similar to the angular coordinate θ on the circle
S 1 . This is a uniformizing parameter on the circle, which is an “inverse” of the universal
covering map R ∋ θ 7→ eiθ ∈ S 1 . ⊔
⊓

In the following exercises (M, g) denotes a compact, oriented Riemann manifold with-
out boundary.

Exercise 10.3.31. Fix c > 0. Show that for every f ∈ L2 (M ) the equation

∆u + cu = f

has an unique solution u ∈ L2,2 (M ). ⊔

⊓
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 509

Exercise 10.3.32. Consider a smooth function f : M × R → R such that for every x ∈ M

the function u 7→ f (x, u) is increasing. Assume that the equation

∆g u = f (x, u) (10.3.17)

admits a pair of comparable sub/super-solutions, i.e., there exist functions u0 , U0 ∈ L1,2 (M )∩

L∞ (M ) such that
U0 (x) ≥ u0 (x), a.e. on M,
and
∆g U0 ≥ f (x, U0 (x)) ≥ f (x, u0 (x)) ≥ ∆g u0 , weakly in L1,2 (M ).
Fix c > 0, and define (un )n≥1 ⊂ L2,2 (M ) inductively as the unique solution of the equation

∆un (x) + cun (x) = cun−1 (x) + f (x, un−1 (x)).

(a) Show that

u0 (x) ≤ u1 (x) ≤ u2 (x) ≤ · · · ≤ un (x) ≤ · · · ≤ U0 (x) ∀x ∈ M.

(b) Show that un converges uniformly on M to a solution u ∈ C ∞ (M ) of (10.3.17) satis-

fying
u0 ≤ u ≤ U0 .
(c) Prove that the above conclusions continue to hold, even if the monotonicity assumption
on f is dropped. ⊔
⊓

10.4 Elliptic operators on compact manifolds

The elliptic operators on compact manifolds behave in many respects as finite dimensional
operators. It is the goal of this last section to present the reader some fundamental analytic
facts which will transform the manipulation with such p.d.o. into a less painful task.
The main reason why these operators are so “friendly” is the existence of a priori
estimates. These estimates, coupled with the Rellich-Kondratchov compactness theorem,
are the keys which will open many doors.

10.4.1 Fredholm theory

Throughout this section we assume that the reader is familiar with some fundamental facts
about unbounded linear operators. We refer to [23] Ch.II for a very concise presentation
of these notions. An exhaustive presentation of this subject can be found in [75].
Definition 10.4.1. (a) Let X, Y be two Hilbert spaces over K = R, C, and

T : D(T ) ⊂ X → Y

be a linear operator, not necessarily continuous, defined on the linear subspace D(T ) ⊂ X.
The operator T is said to be densely defined if its domain D(T ) is dense in X.
510 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

The operator T is said to be closed if its graph,


ΓT := (x, T x) ∈ D(T ) × Y ⊂ X × Y ,
is a closed subspace in X × Y .
(b) Let T : D(T ) ⊂ X → Y be a closed, densely defined linear operator. The adjoint of
T is the operator T ∗ : D(T ∗ ) ⊂ Y → X defined by its graph
ΓT ∗ = {(y ∗ , x∗ ) ∈ Y × X ; hx∗ , xi = hy ∗ , T xi ∀x ∈ D(T ),
where h·, ·i : Z × Z → K denotes the inner product3 in a Hilbert space Z.
(c) A closed, densely defined operator T : D(T ) ⊂ X → X is said to be self-adjoint if
T = T ∗. ⊔
⊓

Remark 10.4.2. (a) In more concrete terms, the operator T : D(T ) ⊂ X → Y is closed if,
for any sequence (xn ) ⊂ D(T ), such that (xn , T xn ) → (x, y), it follows that (i) x ∈ D(T ),
and (ii) y = T x.
(b) If T : X → Y is a closed operator, then T is bounded (closed graph theorem).
Also note that if T : D(T ) ⊂ X → Y is a closed, densely defined operator, then ker T is a
closed subspace of X.
(c) One can show that the adjoint of any closed, densely defined operator, is a closed,
densely defined operator.
(d) The closed, densely defined operator T : D(T ) ⊂ X → X is self-adjoint if the
following two conditions hold.
• hT x, yi = hx, T yi for all x, y ∈ D(T ) and

• D(T ) = y ∈ X; ∃C > 0, |hT x, yi| ≤ C|x|, ∀x ∈ D(T ) .
If only the first condition is satisfied the operator T is called symmetric. ⊔
⊓

Let (M, g) be a compact, oriented Riemann manifold, E, F two metric vector bundles
with compatible connections, and L ∈ P DOk = P DOk (E, F ), a k-th order elliptic p.d.o.
We will denote the various L2 norms by k · k, and the Lk,2 -norms by k · kk .
Definition 10.4.3. The analytical realization of L is the linear operator
La : D(La ) ⊂ L2 (E) → L2 (E), D(La ) = Lk,2 (E),
given by u 7→ Lu for all u ∈ Lk,2 (E). ⊔
⊓

Proposition 10.4.4. (a) The analytical realization La of L is a closed, densely defined

linear operator.
(b) If L∗ : C ∞ (F ) → C ∞ (E) is the formal adjoint of L then
(L∗ )a = (La )∗
In particular, the analytical realization La is self-adjoint if and only if L is formally self-
adjoint.
3
When K = C, we use the convention that the inner product h ·, · i is conjugate linear in the second
variable, i.e., hz1 , λz2 i = λ̄hz1 , z2 i, ∀λ ∈ C, z1 , z2 ∈ Z.
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 511

Proof. (a) Since C ∞ (E) ⊂ D(La ) = Lk,2 (E) is dense in L2 (E) , we deduce that La is
densely defined. To prove that La is also closed, consider a sequence (un ) ⊂ Lk,2 (E) such
that
un → u strongly in L2 (E), and Lun → v strongly in L2 (F ).
From the elliptic estimates we deduce


kun − um kk ≤ C kLun − Lum k + kun − um k → 0 as m, n → ∞.

Hence (un ) is a Cauchy sequence in Lk,2 (E), so that un → u in Lk,2 (E). It is now clear
that v = Lu.
(b) From the equality
Z Z
hLu, vidVg = hu, L∗ vidVg , ∀u ∈ Lk,2 (E), v ∈ Lk,2 (F ),
M M

we deduce


D (La )∗ ⊃ D (L∗ )a = Lk,2 (F ),
and (La )∗ = (L∗ )a on Lk,2 (F ). To prove that



D (La )∗ ⊂ D (L∗ )a = Lk,2 (F )

we need to show that if v ∈ L2 (F ) is such that ∃C > 0 with the property that
Z
hLu, vidVg ≤ Ckuk, ∀u ∈ Lk,2 (E),
M

then v ∈ Lk,2 (F ). Indeed, the above inequality shows that the functional
Z
u 7→ hLu, vidVg
M

extends to a continuous linear functional on L2 (E). Hence, there exists φ ∈ L2 (E) such
that Z Z
∗ ∗
(L ) u, v dVg = hu, φidVg ∀u ∈ Lk,2 (E).
M M

In other words, v is a L2 -weaksolution of the elliptic equation L∗ v = φ. Using elliptic

regularity theory we deduce v ∈ Lk,2 (M ). The proposition is proved. ⊔
⊓

Following the above result we will not make any notational distinction between an
elliptic operator (on a compact manifold) and its analytical realization.
Definition 10.4.5. (a) Let X and Y be two Hilbert spaces over K = R, C and suppose
that T : D(T ) ⊂ X → Y is a closed, densely defined linear operator. The operator T is
said to be semi-Fredholm if the following hold.

(i) dim ker T < ∞ and

(ii) The range R (T ) of T is closed.

512 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

(b) The operator T is called Fredholm if both T and T ∗ are semi-Fredholm. In this case,
the integer
ind T := dimK ker T − dimK ker T ∗

is called the Fredholm index of T . ⊔

⊓

Remark 10.4.6. The above terminology has its origin in the work of Ivar Fredholm at
the twentieth century. His result, later considerably generalized by F. Riesz, states that
if K : H → H is a compact operator from a Hilbert space to itself, then 1H + K is a
Fredholm operator of index 0. ⊔
⊓

Consider again the elliptic operator L of Proposition 10.4.4.

Theorem 10.4.7. The operator La : D(La ) ⊂ L2 (E) → L2 (F ) is Fredholm.

Proof. The Fredholm property is a consequence of the following compactness result.


Lemma 10.4.8. Any sequence (un )n≥0 ⊂ Lk,2 (E) such that kun k + kLun k n≥0 is
bounded contains a subsequence strongly convergent in L2 (E).

Proof. Using elliptic estimates we deduce that

kun kk ≤ C(kLun k + kun k) ≤ const.

Hence (un ) is also bounded in Lk,2 (E). On the other hand, since M is compact, the space
Lk,2 (E) embeds compactly in L2 (E). The lemma is proved. ⊔
⊓

We will first show that dim ker L < ∞. In the proof we will rely on the classical result
of F. Riesz which states that a Banach space is finite dimensional if and only if its bounded
subsets are precompact (see [23], Chap. VI).
Note first that, according to Weyl’s lemma ker L ⊂ C ∞ (E). Next, notice that ker L is
a Banach space with respect to the L2 -norm since according to Remark 10.4.2 (a) ker L
is closed in L2 (E). We will show that any sequence (un ) ⊂ ker L which is also bounded
in the L2 -norm contains a subsequence convergent in L2 . This follows immediately from
Lemma 10.4.8 since kun k + kLun k = kun k is bounded.
To prove that the range R (T ) is closed, we will rely on the following very useful
inequality, a special case of which we have seen at work in Subsection 9.3.3.

Lemma 10.4.9 (Poincaré inequality). There exists C > 0 such that

kuk ≤ CkLuk,

for all u ∈ Lk,2 (E) which are L2 -orthogonal to ker L, i.e.,

Z
hu, φidVg = 0 ∀φ ∈ ker L.
M
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 513

Proof. We will argue by contradiction. Denote by X ⊂ Lk,2 (E) the subspace consisting
of sections L2 -orthogonal to ker L. Assume that for any n > 0 there exists un ∈ X such
that
kun k = 1 and kLun k ≤ 1/n.
Thus, kLun k → 0, and in particular, kun k + kLun k is bounded. Using Lemma 10.4.8 we
deduce that a subsequence of (un ) is convergent in L2 (E) to some u. Note that kuk = 1.
It is not difficult to see that in fact u ∈ X. We get a sequence

(un , Lun ) ⊂ ΓL = the graph of L,

such that (un , Lun ) → (u, 0). Since L is closed, we deduce u ∈ D(L) and Lu = 0. Hence
u ∈ ker L ∩ X = {0}. This contradicts the condition kuk = 1. ⊔
⊓

We can now conclude the proof of Theorem 10.4.7. Consider a sequence (vn ) ⊂ R (L)
such that vn → v in L2 (F ). We want to show v ∈ R (L).
For each vn we can find a unique un ∈ X = (ker L)⊥ such that

Lun = vn .

Using the Poincaré inequality we deduce

kun − um k ≤ Ckvn − vm k.

When we couple this inequality with the elliptic estimates we get

kun − um kk ≤ C(kLun − Lum k + kun − um k) ≤ Ckvn − vm k → 0 as m, n → ∞.

Hence (un ) is a Cauchy sequence in Lk,2 (E) so that un → u in Lk,2 (E). Clearly Lu = v,
so that v ∈ R (L).
We have so far proved that ker L is finite dimensional, and R (L) is closed, i.e. La is
semi-Fredholm. Since (La )∗ = (L∗ )a , and L∗ is also an elliptic operator, we deduce (La )∗
is also semi-Fredholm. This completes the proof of Theorem 10.4.7. ⊔
⊓

Using the closed range theorem of functional analysis we deduce the following impor-
tant consequence.
Corollary 10.4.10 (Abstract Hodge decomposition). Any k-th order elliptic operator L :
C ∞ (E) → C ∞ (F ) over the compact manifold M defines natural orthogonal decompositions
of L2 (E) and L2 (F ),

L2 (E) = ker L ⊕ R (L∗ ) and L2 (F ) = ker L∗ ⊕ R (L). ⊔

⊓

Corollary 10.4.11. If ker L∗ = 0, then for every v ∈ L2 (F ) the partial differential

equation Lu = v admits at least one weak L2 -solution u ∈ L2 (E). ⊔
⊓

Corollary 10.4.12. Suppose that L is additionally formally self-adjoint. If ker L = 0,

then for any j ∈ Z≥0 , the induced linear map Lk+j,2 (E) → Lj,2 (F ) is a topological iso-
morphism. ⊔
⊓
514 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

The last corollary is really unusual. It states the equation Lu = v has a solution
provided the dual equation L∗ v = 0 has no nontrivial solution. A nonexistence hypothesis
implies an existence result! This partially explains the importance of the vanishing results
in geometry, i.e., the results to the effect that ker L∗ = 0. With an existence result in
our hands presumably we are more capable of producing geometric objects. In the next
chapter we will describe one powerful technique of producing vanishing theorems based
on the so called Weitzenböck identities.
Corollary 10.4.13. Over a compact manifold

ker L = ker L∗ L and ker L∗ = ker LL∗ .

Proof. Clearly ker L ⊂ ker L∗ L. Conversely, let ψ ∈ C ∞ (E) such that L∗ Lψ = 0. Then
Z Z
kLψk2 = hLψLψidVg = hL∗ Lψ, ψidVg = 0. ⊔
⊓
M M

The Fredholm property of an elliptic operator has very deep topological ramifications
culminating with one of the most beautiful results in mathematics: the Atiyah-Singer
index theorems. Unfortunately, this would require a lot more extra work to include it
here. However, in the remaining part of this subsection we will try to unveil some of the
natural beauty of elliptic operators. We will show that the index of an elliptic operator
has many of the attributes of a topological invariant.
We stick to the notations used so far. Denote by Ellk (E, F ) the space of elliptic
operators C ∞ (E) → C ∞ (F ) of order k. By using the attribute space when referring to
Ellk we implicitly suggested that it carries some structure. It is not a vector space, it
is not an affine space, it is not even a convex set. It is only a cone in the linear space
P DO(m) , but it carries a natural topology which we now proceed to describe.
Let L1 , L2 ∈ Ellk (E, F ). We set

δ(L1 , L2 ) = sup kL1 u − L2 uk; kukk = 1, .

Define

d(L1 , L2 ) = max δ(L1 , L2 ), δ(L∗1 , L∗2 ), .
We let the reader check that (Ellk , d) is indeed a metric space. A continuous family of
elliptic operators (Lλ )λ∈Λ , where Λ is a topological space, is then a continuous map

Λ ∋ λ 7→ Lλ ∈ Ellk .

Roughly speaking, this means that the coefficients of Lλ , and their derivatives up to order
k depend continuously upon λ.
Theorem 10.4.14. The index map

ind : Ellk (E, F ) → Z, L 7→ ind (L)

is continuous.
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 515

The proof relies on a very simple algebraic trick which requires some analytical foun-
dation.
Let X, Y be two Hilbert spaces. For any Fredholm operator L : D(L) ⊂ X → Y denote
by ıL : ker L → X (respectively by PL : X → ker L) the natural inclusion ker L ֒→ X
(respectively the orthogonal projection X → ker L). If Li : D(Li ) ⊂ X → Y (i = 0, 1) are
two Fredholm operators define

RL0 (L1 ) : D(L1 ) ⊕ ker L∗0 ⊂ X ⊕ ker L∗0 → Y ⊕ ker L0

by
RL0 (L1 )(u, φ) = (L1 u + ıL∗0 φ, PL0 u), u ∈ D(L1 ), φ ∈ ker L∗0 .
In other words, the operator RL0 (L1 ) is given by the block decomposition
 
L1 ıL∗0
RL0 (L1 ) = .
PL0 0

We will call RL0 (L1 ) is the regularization of L1 at L0 . The operator L0 is called the center
of the regularization. For simplicity, when L0 = L1 = L, we write

RL = RL (L).

The result below lists the main properties of the regularization.

Lemma 10.4.15. (a) RL0 (L1 ) is a Fredholm operator.

(b) R∗L0 (L1 ) = RL∗0 (L∗1 ).
(c) RL0 is invertible (with bounded inverse). ⊔
⊓

Exercise 10.4.16. Prove the above lemma. ⊔

⊓

We strongly recommend the reader who feels less comfortable with basic arguments of
functional analysis to try to provide the no-surprise proof of the above result. It is a very
good “routine booster”.
Proof of Theorem 10.4.14 Let L0 ∈ Ellk (E, F ). We have to find r > 0 such that,
∀L ∈ Ellk (E, F ) satisfying d(L0 , L) ≤ r, we have

ind (L) = ind (L0 ).

We will achieve this in two steps.

Step 1. We will find r > 0 such that, for any L satisfying d(L0 , L) < r, the regularization
of L at L0 is invertible, with bounded inverse.
Step 2. We will conclude that if d(L, L0 ) < r, where r > 0 is determined at Step 1, then
ind (L) = ind (L0 ).
Step 1. Since RL0 (L) is Fredholm, it suffices to show that both RL0 (L), and RL∗0 (L∗ )
are injective, if L is sufficiently close to L0 . We will do this only for RL0 (L), since the
remaining case is entirely similar.
516 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

We argue by contradiction. Assume that there exists a sequence (un , φn ) ⊂ Lk,2 (E) ×
ker L∗0 , and a sequence (Ln ) ⊂ Ellk (E, F ) such that

kun kk + kφn k = 1, (10.4.1)

RL0 (Ln )(un , φn ) = (0, 0), (10.4.2)

and
d(L0 , Ln ) ≤ 1/n. (10.4.3)
From (10.4.1) we deduce that (φn ) is a bounded sequence in the finite dimensional space
ker L∗0 . Hence it contains a subsequence strongly convergent in L2 , and in fact in any
Sobolev norm. Set φ := lim φn . Note that kφk = limn kφn k. Using (10.4.2) we deduce

Ln un = −φn ,

i.e., the sequence (Ln un ) is strongly convergent to −φ in L2 (F ). The condition (10.4.3)

now gives
kL0 un − Ln un k ≤ 1/n,
i.e.,
lim Ln un = lim L0 un = −φ ∈ L2 (F ).
n n

Since un ⊥ ker L0 (by (10.4.2)) we deduce from the Poincaré inequality combined with
the elliptic estimates that

kun − um kk ≤ CkL0 un − L0 um k → 0 as m, n → ∞.

Hence the sequence un strongly converges in Lk,2 to some u. Moreover,

lim kun kk = kukk and kukk + kφk = 1.

n

Putting all the above together we conclude that there exists a pair (u, φ) ∈ Lk,2 (E)×ker L∗0 ,
such that
kukk + kφk = 1,
L0 u = −φ and u ⊥ ker L. (10.4.4)
This contradicts the abstract Hodge decomposition which coupled with (10.4.4) implies
u = 0 and φ = 0. Step 1 is completed.
Step 2. Let r > 0 as determined at Step 1, and L ∈ Ellk (E, F ). Hence
 
L ıL∗0
RL0 (L) =
PL0 0

is invertible. We will use the invertibility of this operator to produce an injective operator

ker L∗ ⊕ ker L0 ֒→ ker L ⊕ ker L∗0 .

This implies dim ker L∗ + dim ker L0 ≤ dim ker L + dim ker L∗0 , i.e.,

ind (L0 ) ≤ ind(L).

10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 517

A dual argument, with L replaced by L∗ , and L0 replaced by L∗0 , will produce the opposite
inequality, and thus finish the proof of Theorem 10.4.14. Now let us provide the details.
First, we orthogonally decompose

L2 (E) = (ker L)⊥ ⊕ ker L and L2 (F ) = (ker L∗ )⊥ ⊕ ker L∗ .

Set U := ker L ⊕ ker L∗0 , and V := ker L∗ ⊕ ker L0 . We will regard RL0 (L) as an operator

RL0 (L) : (ker L)⊥ ⊕ U → (ker L∗ )⊥ ⊕ V.

As such, it has a block decomposition

 
T A
RL0 (L) = ,
B C

where
T : Lk,2 (E) ∩ (ker L)⊥ ⊂ (ker L)⊥ → (ker L∗ )⊥ = Range (L)
denotes the restriction of L to (ker L)⊥ . The operator T is invertible and its inverse is
bounded.
Since RL0 (L) is invertible, for any v ∈ V we can find a unique pair (φ, u) ∈ (ker L)⊥ ⊕
U , such that    
φ 0
RL0 (L) = .
u v
This means
T φ + Au = 0 and Bφ + Cu = v.
We can view both φ and u as (linear) functions of v, φ = φ(v) and u = u(v). We claim
the map v 7→ u = u(v) is injective.
Indeed, if u(v) = 0 for some v, then T φ = 0, and since T is injective φ must be zero.
From the equality v = Bφ + Cu we deduce v = 0. We have thus produced the promised
injective map V ֒→ U . Theorem 10.4.14 is proved. ⊔
⊓

The theorem we have just proved has many topological consequences. We mention
only one of them.
Corollary 10.4.17. Let L0 , L1 ∈ Ellk (E, F ) if σk (L0 ) = σk (L1 ) then ind (L0 ) = ind (L1 ).

Proof. For every t ∈ [0, 1] Lt = (1 − t)L0 + tL1 is a k-th order elliptic operator depend-
ing continuously on t. (Look at the symbols). Thus ind (Lt ) is an integer depending
continuously on t so it must be independent of t. ⊔
⊓

This corollary allows us to interpret the index as as a continuous map from the elliptic
symbols to the integers. The analysis has vanished! This is (almost) a purely algebraic-
topologic object. There is one (major) difficulty. These symbols are “polynomials with
coefficients in some spaces of endomorphism”.
The deformation invariance of the index provides a very powerful method for com-
puting it by deforming a “complicated” situation to a “simpler” one. Unfortunately, our
deformation freedom is severely limited by the “polynomial” character of the symbols.
518 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

There aren’t that many polynomials around. Two polynomial-like elliptic symbols may
be homotopic in a larger classes of symbols which are only positively homogeneous along
the fibers of the cotangent bundle).
At this point one should return to analysis, and try to conceive some operators that
behave very much like elliptic p.d.o. and have more general symbols. Such objects exist,
and are called pseudo-differential operators. We refer to [83, 121] for a very efficient
presentation of this subject. We will not follow this path, but we believe the reader who
reached this point can complete this journey alone.
Exercise 10.4.18. Let L ∈ Ellk (E, F ). A finite dimensional subspace V ⊂ L2 (F ) is
called a stabilizer of L if the operator

SL,V : Lk,2 (E) ⊕ V → L2 (F ) SL,V (u ⊕ v) = Lu + v

is surjective.
(a) Show that any subspace V ⊂ L2 (F ) containing ker L∗ is a stabilizer of L. More gener-
ally, any finite dimensional subspace of L2 (F ) containing a stabilizer is itself a stabilizer.
Conclude that if V is a stabilizer, then

ind L = dim ker SL,V − dim V. ⊔

⊓

Exercise 10.4.19. Consider a compact manifold Λ and L : Λ → Ellk (E, F ) a continuous

family of elliptic operators.
(a) Show that the family L admits an uniform stabilizer, i.e., there exists a finite dimen-
sional subspace V ⊂ L2 (F ), such that V is a stabilizer of each operator Lλ in the family
L.
(b) Show that if V is an uniform stabilizer of the family L, then the family of subspaces
ker SLλ ,V defines a vector bundle over Λ.
(c) Show that if V1 and V2 are two uniform stabilizers of the family L, then we have a
natural isomorphism vector bundles

ker SL,V1 ⊕ V 2 ∼
= ker SL,V2 ⊕ V1 .

In particular, we have an isomorphism of line bundles

det ker SL,V1 ⊗ det V1∗ ∼

= det ker SL,V2 ⊗ det V2∗ .

Thus the line bundle det ker SL,V ⊗ det V ∗ → Λ is independent of the uniform stabilizer
V . It is called the determinant line bundle of the family L and is denoted by det ind(L).
⊔
⊓

10.4.2 Spectral theory

We mentioned at the beginning of this section that the elliptic operators on compact
manifold behave very much like matrices. Perhaps nothing illustrates this feature better
than their remarkable spectral properties. This is the subject we want to address in this
subsection.
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 519

Consider as usual, a compact, oriented Riemann manifold (M, g), and a complex vec-
tor bundle E → M endowed with a Hermitian metric h•, •i and compatible connection.
Throughout this subsection L will denote a k-th order, formally self-adjoint elliptic oper-
ator L : C ∞ (E) → C ∞ (E). Its analytical realization

La : Lk,2 (E) ⊂ L2 (E) → L2 (E),

is a self-adjoint, elliptic operator, so its spectrum spec(L) is an unbounded closed subset

of R. Note that for any λ ∈ R the operator λ − La is the analytical realization of the
elliptic p.d.o. λ1E − L, and in particular, the operator λ − La is a Fredholm, so that

λ ∈ spec(L) ⇐⇒ ker(λ − L) 6= 0.

Thus, the spectrum of L consists only of eigenvalues of finite multiplicities. The main
result of this subsection states that one can find an orthonormal basis of L2 (E) which
diagonalizes La .
Theorem 10.4.20. Let L ∈ Ellk (E) be a formally self-adjoint elliptic operator. Then the
following are true.
(a) The spectrum spec(L) is real, spec(L) ⊂ R, and for each λ ∈ spec(L) , the subspace
ker(λ − L) is finite dimensional and consists of smooth sections.
(b) The spectrum spec(L) is a closed, countable, discrete, unbounded set.
(c) There exists an orthogonal decomposition
M
L2 (E) = ker(λ − L).
λ∈spec(L)

(d) Denote by Pλ the orthogonal projection onto ker(λ − L). Then

 X
Lk,2 (E) = D(La ) = ψ ∈ L2 (E) ; λ2 kPλ ψk2 < ∞ .
λ

Part (c) of this theorem allows one to write

X X
1= Pλ and L = λPλ .
λ λ

The first identity is true over the entire L2 (E) while part (d) of the theorem shows the
domain of validity of the second equality is precisely the domain of L.

Proof. (a) We only need to show that ker(λ − L) consists of smooth sections. In view of
Weyl’s lemma this is certainly the case since λ − L is an elliptic operator.
(b)&(c) We first show spec(L) is discrete.
More precisely, given λ0 ∈ spec(L), we will find ε > 0 such that ker(λ − L) = 0,
∀|λ − λ0 |, ε, λ 6= λ0 . Assume for simplicity λ0 = 0.
We will argue by contradiction. Thus there exist λn → 0 and un ∈ C ∞ (E), such that

Lun = λn un , kun k = 1.
520 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Clearly un ∈ R (L) = (ker L∗ )⊥ = (ker L)⊥ , so that the Poincaré inequality implies

1 = kun k ≤ CkLun k = Cλn → 0.

Thus spec(L) must be a discrete set.

Now consider t0 ∈ R \ σ(L). Thus t0 − L has a bounded inverse

T = (t0 − L)−1 .

Obviously T is a self-adjoint operator. We claim that T is also a compact operator.

Assume (vn ) is a bounded sequence in L2 (E). We have to show un = T vn admits a
subsequence which converges in L2 (E). Note that un is a solution of the partial differential
equation
(t0 − L)un = t0 un − Lun = vn .
Using the elliptic estimates we deduce

kun kk ≤ C(kun k + kvn k).

Obviously un = T vn is bounded in L2 (E), so the above inequality implies kun kk is

also bounded. The desired conclusion follows from the compactness of the embedding
Lk,2 (E) → L2 (E).
Thus T is a compact, self-adjoint operator. We can now use the spectral theory of such
well behaved operators as described for example in [23], Chap. 6. The spectrum of T is a
closed, bounded, countable set with one accumulation point, µ = 0. Any µ ∈ spec(T )\{0}
is an eigenvalue of T with finite multiplicity, and since ker T = 0,
M
L2 (E) = ker(µ − T ).
µ∈spec(T )\{0}

Using the equality L = t0 − T −1 we deduce

σ(L) = {t0 − µ−1 ; µ ∈ spec(T ) \ {0}}.

This proves (b)&(c).

To prove (d), note that if ψ ∈ Lk,2 (E), then Lψ ∈ L2 (E), i.e.,
X X
k λPλ ψk2 = λ2 kPλ ψk2 < ∞.
λ λ

Conversely, if X
λ2 kPλ ψk2 < ∞,
λ

consider the sequence of smooth sections

X
φn = λPλ ψ
|λ|≤n
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 521

which converges in L2 (E) to

X
φ= λPλ ψ.
λ

On the other hand, φn = Lψn , where

X
ψn = Pλ ψ
|λ|≤n

converges in L2 (E) to ψ. Using the elliptic estimates we deduce

kψn − ψm kk ≤ C(kψn − ψm k + kφn − φm k) → 0 as n, m → ∞.

Hence ψ ∈ Lk,2 (E) as a Lk,2 -limit of smooth sections. The theorem is proved. ⊔
⊓

Example 10.4.21. Let M = S 1 , E = CM and

∂
L = −i : C ∞ (S 1 , C) → C ∞ (S 1 , C).
∂θ
The operator L is clearly a formally self-adjoint elliptic p.d.o. The eigenvalues and the
eigenvectors of L are determined from the periodic boundary value problem

∂u
−i = λu, u(0) = u(2π),
∂θ
which implies
u(θ) = C exp(iλθ) and exp(2πλi) = 1.
Hence
spec(L) = Z and ker(n − L) = spanC {exp(inθ)}.
The orthogonal decomposition
M
L2 (S 1 ) = ker(n − L)
n

is the usual Fourier decomposition of periodic functions. Note that

X
u(θ) = un exp(inθ) ∈ L1,2 (S 1 )
n

if and only if X
(1 + n2 )|un |2 < ∞. ⊔
⊓
n∈Z

The following exercises provide a variational description of the eigenvalues of a formally

self-adjoint elliptic operator L ∈ Ellk (E) which is bounded from below i.e.
nZ o
inf hLu, uidvg ; u ∈ Lk,2 (E), kuk = 1 > −∞.
M
522 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Exercise 10.4.22. Let V ∈ Lk,2 (E) be a finite dimensional invariant subspace of L, i.e.
L(V ) ⊂ V . Show that

1. The space V consists only of smooth sections.

2. The quantity
nZ o
⊥
λ(V ) = inf hLu, uidvg ; u ∈ Lk,2 (E) ∩ V ⊥ , kuk = 1
M

is an eigenvalue of L. (V ⊥ denotes the orthogonal complement of V in L2 (E)).

⊔
⊓

Exercise 10.4.23. Set V0 = 0, and denote λ0 = λ(V0⊥ ), According to the previous

exercise λ0 is an eigenvalue of L. Pick φ0 an eigenvector corresponding to λ0 such that
kφ0 k = 1 and form V1 = V0 ⊕ span {φ0 }. Set λ1 = λ(V1⊥ ) and iterate the procedure. After
m steps we have produced m + 1 vectors φ0 , φ1 , . . . , φm corresponding to m + 1 eigenvalues
λ0 , λ1 ≤ · · · λm of L. Set

Vm+1 := spanC {φ0 , φ1 , . . . , φm }

⊥ ) etc.
and λm+1 := λ(Vm+1

1. Prove that the set

{φ1 , . . . , φm , . . .}
is a Hilbert basis of L2 (E), and

spec(L) = λ1 ≤ · · · ≤ λm · · · .

2. Denote by Grm the Grassmannian of m-dimensional subspaces of Lk,2 (E). Show

that nZ o
λm = inf max hLu, uidvg ; u ∈ V kuk = 1 .
V ∈Grm M

⊔
⊓

Exercise 10.4.24. Use the results in the above exercises to show that if L is a bounded
from below, k-th order formally self-adjoint elliptic p.do. over an N -dimensional manifold
then
λm (L) = O(mk/N ) as m → ∞,
and
d(Λ) = dim ⊕λ≤Λ ker(λ − L) = O(ΛN/k ) as Λ → ∞. ⊔
⊓
Remark 10.4.25. (a) When L is a formally self-adjoint generalized Laplacian then the
result in the above exercise can be considerably sharpened. More precisely H.Weyl showed
that
rank (E) · volg (M )
lim Λ−N/2 d(Λ) = . (10.4.5)
Λ→∞ (4π)N/2 Γ(N/2 + 1)
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 523

The very ingenious proof of this result relies on another famous p.d.o., namely the heat
operator ∂t + L on R × M . For details we refer to [14], or [121].
In the next chapter we will investigate the spectrum of the Laplacian ∆ of a compact
Riemann manifold using another famous p.d.o., the wave operator  = ∂t2 + ∆ on R × M .
In particular, we will obtain a refinement of (10.4.5).
(b) Assume L is the scalar Laplacian ∆ on a compact Riemann manifold (M, g) of
dimension M . Weyl’s formula shows that the asymptotic behavior of the spectrum of ∆
contains several geometric informations about M : we can read the dimension and the
volume of M from it. If we think of M as the elastic membrane of a drum then the
eigenvalues of ∆ describe all the frequencies of the sounds the “drum” M can produce.
Thus “we can hear” the dimension and the volume of a drum. This is a special case of
a famous question raised by V.Kac in [71]: can one hear the shape of a drum? In more
rigorous terms this question asks how much of the geometry of a Riemann manifold can
be recovered from the spectrum of its Laplacian. This is what spectral geometry is all
about.
It has been established recently that the answer to Kac’s original question is negative.
We refer to [53] and the references therein for more details. In the next chapter we can
reconstruct the geometry of the manifold if besides the spectrum we also have information
about the eigenfunctions of the Laplace operator. ⊔
⊓

Exercise 10.4.26. Compute the spectrum of the scalar Laplacian on the torus T 2 equipped
with the flat metric, and then use this information to prove the above Weyl asymptotic
formula in this special case. ⊔
⊓

Exercise 10.4.27. Denote by ∆S n−1 the scalar Laplacian on the unit sphere S n−1 ⊂ Rn
equipped with the induced metric. We assume n ≥ 2. A polynomial p = p(x1 , . . . , xn ) is
called harmonic if ∆Rn p = 0.
(a) Show that if p is a homogenous harmonic polynomial of degree k, and p̄ denotes its
restriction to the unit sphere S n−1 centered at the origin, then

∆S n−1 p̄ = n(n + k − 2)p̄.

(b) Show that a function ϕ : S n → R is an eigenfunction of ∆S n−1 if and only if there

exists a homogeneous harmonic polynomial p such that ϕ = p̄.
(c) Set λk := n(n + k − 2). Show that
   

k+n−1 k+n−3
dim ker λk − ∆S n−1 = − .
n n−2

(d) Prove Weyl’s asymptotic estimate (10.4.5) in the special case of the operator ∆S n−1 .
⊔
⊓

10.4.3 Hodge theory

We now have enough theoretical background to discuss the celebrated Hodge theorem. It
is convenient to work in a slightly more general context than Hodge’s original theorem.
524 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

Definition 10.4.28. Let (M, g) be an oriented Riemann manifold. An elliptic complex

is a sequence of first order p.d.o.’s

D D Dm−1
0 → C ∞ (E0 ) →0 C ∞ (E1 ) →1 · · · → C ∞ (Em ) → 0

satisfying the following conditions.

(i) (C ∞ (Ei ), Di ) is a cochain complex, i.e., Di Di−1 = 0, ∀1 ≤ i ≤ m.

(ii) For each (x, ξ) ∈ T ∗ M \ {0} the sequence of principal symbols

σ(D0 )(x,ξ) σ(Dm−1 )(x,ξ)

0 → (E0 )x −→ (E1 )x → · · · −→ (Em )x → 0

is exact. ⊔
⊓

Example 10.4.29. The DeRham complex (Ω∗ (M ), d) is an elliptic complex. In this case,
the associated sequence of principal symbols is (e(ξ) = exterior multiplication by ξ)

e(ξ) e(ξ) e(ξ)

0 → R −→ Tx∗ M −→ · · · −→ det(Tx∗ M ) → 0

is the Koszul complex of Exercise 7.1.23 of Subsection 7.1.3 where it is shown to be exact.
Hence, the DeRham complex is elliptic. We will have the occasion to discuss another
famous elliptic complex in the next chapter. ⊔
⊓

Consider an elliptic complex (C ∞ (E· , D· )) over a compact oriented Riemann manifold

(M, g). Denote its cohomology by H • (E· , D· ). A priori these may be infinite dimensional
spaces. We will see that the combination ellipticity + compactness prevents this from
happening. Endow each Ei with a metric and compatible connection. We can now talk
about Sobolev spaces and formal adjoints Di∗ . Form the operators

∆i = Di∗ Di + Di−1 Di−1

∗
: C ∞ (Ei ) → C ∞ (Ei ).

We can now state and prove the celebrated Hodge theorem.

Theorem 10.4.30 (Hodge). Assume that M is compact and oriented. Then the following
are true.

1. H i (E· , D· ) ∼
= ker ∆i ⊂ C ∞ (Ei ), ∀i.

2. dim H i (E· , D· ) < ∞ ∀i.

3. (Hodge decomposition). There exists an orthogonal decomposition

L2 (Ei ) = ker ∆i ⊕ R (Di−1 ) ⊕ R (Di∗ ),

where we view both Di−1 , and Di as bounded operators L1,2 → L2 .

10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 525

Proof. Set

E = ⊕Ei , D = ⊕Di , D ∗ = ⊕Di∗ , ∆ = ⊕∆i , D̂ = D + D ∗ .

Thus D, D ∗ , ∆ ∈ P DO(E, E). Since Di Di−1 = 0, we deduce D 2 = (D ∗ )2 = 0 which

implies
∆ = D ∗ D + DD ∗ = (D + D ∗ )2 = D̂ 2 .
We now invoke the following elementary algebraic fact which is a consequence of the
exactness of the symbol sequence.
Exercise 10.4.31. The operators D̂ and ∆ are elliptic, formally self-adjoint p.d.o. (Hint:
Use Exercise 7.1.22 in Subsection 7.1.3.) ⊔
⊓

Note that according to Corollary 10.4.13 we have ker ∆ = ker D̂, so that we have an
orthogonal decomposition
L2 (E) = ker ∆ ⊕ R (D̂). (10.4.6)
This is precisely part (c) of Hodge’s theorem.
For each i denote by Pi the orthogonal projection L2 (Ei ) → ker ∆i . Set

Z i = u ∈ C ∞ (Ei ); Di u = 0 and B i = Di−1 (C ∞ (Ei−1 )),

so that
H i (E· , D· ) = Z i /B i .
We claim that the map Pi : Z i → ker ∆i descends to an isomorphism H i (E· , D· ) → ker ∆i .
This will complete the proof of Hodge theorem. The above claim is a consequence of
several simple facts.
Fact 1. ker ∆i ⊂ Z i . This follows from the equality ker ∆ = ker D̂.
Fact 2. If u ∈ Zi then u − Pi u ∈ Bi . Indeed, using the decomposition (10.4.6) we have

u = Pi u + D̂ψ ψ ∈ L1,2 (E).

Since u − Pi u ∈ C ∞ (Ei ), we deduce from Weyl’s lemma that ψ ∈ C ∞ (E). Thus, there
exist v ∈ C ∞ (Ei−1 ) and w ∈ C ∞ (Ei ) such that ψ = v ⊕ w and

u = Pi u + Di−1 v + Di∗ w.

Applying Di on both sides of this equality we get

0 = Di u = Di Pi u + Di Di−1 u + Di Di∗ w = Di Di∗ w.

Since ker Di∗ = ker Di Di∗ , the above equalities imply

u − Pi u = Di−1 v ∈ B i .

We conclude that Pi descends to a linear map ker ∆i → H i (E· , D· ). Thus

B i ⊂ R (D̂) = (ker ∆)⊥ ,

526 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS

and we deduce that no two distinct elements in ker ∆i are cohomologous since otherwise
their difference would have been orthogonal to ker ∆i . Hence, the induced linear map
Pi : ker ∆i → H i (E· , D· ) is injective. Fact 1 shows it is also surjective so that

ker ∆i ∼
= H i (E· , D· ).

Hodge theorem is proved. ⊔

⊓

Let us apply Hodge theorem to the DeRham complex on a compact oriented Riemann
manifold
d d d
0 → Ω0 (M ) −→ Ω1 (M ) −→ · · · −→ Ωn (M ) → 0, n = dim M.
We know that the formal adjoint of d : Ωk (M ) → Ωk+1 (M ) is

d∗ = (−1)νn,k ∗ d∗,

where νn,k = nk + n + 1, and ∗ denotes the Hodge ∗-operator defined by the Riemann
metric g and the fixed orientation on M . Set

∆ = dd∗ + d∗ d.

Corollary 10.4.32 (Hodge). Any smooth k-form ω ∈ Ωk (M ) decomposes uniquely as

ω = [ω]g + dη + d∗ ζ, η ∈ Ωk−1 (M ), ζ ∈ Ωk+1 (M ),

and [ω]g ∈ Ωk (M ) is g-harmonic, i.e.,

∆[ω]g = 0 ⇐⇒ d[ω]g = 0 and d∗ [ω]g = 0.

The form [ω]g is called the g-harmonic part of ω.

If moreover ω is closed, then d∗ ζ = 0, and this means any cohomology class [z] ∈
k
H (M ) is represented by a unique harmonic k-form. ⊔
⊓

Denote by Hk (M, g) the space of g-harmonic k-forms on M . The above corollary

shows
Hk (M, g) ∼
= H k (M )
for any metric g.
Corollary 10.4.33. The Hodge ∗-operator defines a bijection

∗ : Hk (M, g) → Hn−k (M, g).

Proof. If ω is g-harmonic, then so is ∗ω since

d ∗ ω = ± ∗ d ∗ ω = 0,

and
∗d ∗ (∗ω) = ± ∗ (dω) = 0.
The Hodge ∗-operator is bijective since ∗2 = (−1)k(n−k) . ⊔
⊓
10.4. ELLIPTIC OPERATORS ON COMPACT MANIFOLDS 527

Using the L2 -inner product on Ω• (M ) we can identify Hn−k (M, g) with its dual, and
thus we can view ∗ as an isomorphism
∗
Hk → (Hn−k )∗ .

On the other hand, the Poincaré duality described in Chapter 7 induces another isomor-
phism
PD
Hk −→ (Hn−k )∗ ,
defined by Z
hP D(ω), ηi0 = ω ∧ η,
M
where h·, ·i0 denotes the natural pairing between a vector space and its dual.
Proposition 10.4.34. P D = ∗, i.e.
Z Z
h∗ω, ηig dvg = ω ∧ η ∀ω ∈ Hk η ∈ Hn−k .
M M

Proof. We have

h∗ω, ηig dvg = hη, ∗ωig dvg = η ∧ ∗2 ω = (−1)k(n−k) η ∧ ω = ω ∧ η. ⊔

⊓

Exercise 10.4.35. Let ω0 ∈ Ωk (M ) be a harmonic k form and denote by Cω0 its coho-
mology class. Show that
Z Z
|ω0 |2g dvg ≤ |ω|2g dvg ∀ω ∈ Cω0 ,
M M

with equality if and only if ω = ω0 . ⊔

⊓

Exercise 10.4.36. Let G denote a compact connected Lie group equipped with a bi-
invariant Riemann metric h. Prove that a differential form on G is h-harmonic if and only
if it is bi-invariant. ⊔
⊓
528 CHAPTER 10. ELLIPTIC EQUATIONS ON MANIFOLDS
Chapter 11

Spectral Geometry

On a compact Riemann manifold the Laplacian produces a wealth of data: its spectrum
and the corresponding eigenfunctions. We will refer to these as the spectral data of a
Riemann manifold. Clearly the spectral data depend on the metric. In this chapter we
discuss how much information about the metric we can extract from the spectral data.
In approaching this problem we rely on a strategy pioneered by L. Hörmander [65] and
based on the wave equation.

11.1 Generalized functions and currents

We will rely heavily on the theory of distributions or generalized functions. The goal of
this section is to describe, mostly without proofs, several basic facts and, in greater detail
several more exotic the aspects of this theory. For more details we refer [39, 49, 50, 66,
118, 127] which were our main sources of inspiration.

11.1.1 Generalized functions and operations withy them

For any open set O ⊂ Rn and any compact set K ⊂ O we denote by DK (O) the space of
complex valued smooth functions on O whose support is contained in K. We set
[
D(O) := DK (O).
K⊂O

We will refer to the functions in D(O) as (compactly supported) test functions.

Definition 11.1.1. A generalized function or distribution on O is a linear functional

u : D(O) → C, D(O) ∋ ϕ 7→ hu, ϕi,

satisfying the following continuity property: for any compact set K ⊂ O there exist
m = mK ∈ Z≥0 and C = CK > 0 such that, for any function ϕ ∈ DK (O)
X
hu, ϕi| ≤ C sup |∂xα ϕ(x)|. (11.1.1)
x∈K
|α|≤mK

529
530 CHAPTER 11. SPECTRAL GEOMETRY

We denote by D ′ (O) or C −∞ (O) the space of generalized functions on O.

A generalized function u is said to have order ≤ m if in (11.1.1) we can chose mK ≤ m
for any compact K ⊂ O. ⊔
⊓

Remark 11.1.2. Technically, a generalized function is really a linear functional on the

space of smooth densities on O with compact support. However, the open set O is equipped
with a canonical volume density defined by the Euclidean metric and this allows us to
identify the space smooth densities with compact support with D(O).
This distinction becomes very relevant when working on manifolds where there is no
canonical volume density. We refer to [56] for an approach to generalized functions that
takes consistently into account this aspect. ⊔
⊓

Example 11.1.3. (a) Any complex valued measure on the σ-algebra of Borel subset of
O defines a generalized function on O. In particular L1loc (O) ⊂ C −∞ (O). A function
f ∈ L1loc (O) defines a distribution uf by setting
Z
huf , ϕi := f (x)ϕ(x)dx, ∀ϕ ∈ D(O).
O

(b) For any nonzero multi-index α ∈ ZN

≥0 any point x0 ∈ O the linear map

D(O) ∋ ϕ 7→ ∂xα ϕ(x0 ) ∈ C

is a generalized function. When |α| ≥ 1 this generalized function is not determined by any
complex measure on O. When α = 0 this is the Dirac function (measure) concentrated
at x0 and it is denoted by δx0 . ⊔
⊓

Definition 11.1.4. Let O be an open subset of Rn . A sequence of generalized functions

uk ∈ C −∞ (O), k ≥ 0, is said to converge weakly or in the sense of distributions to
u ∈ C −∞ (O) if
lim huk , ϕi = hu, ϕi, ∀ϕ ∈ D(O). ⊔
⊓
k→∞
The above definition can be strengthened by the following nontrivial result. For a
proof we refer to [50, Sec. I.8] or [127, Sec. 33].
Theorem 11.1.5 (Uniform boundedness principle). Let O be an open subset of Rn . Sup-
pose that uj ∈ C −∞ (O), j ≥ 0, is a sequence of generalized functions such that, ∀ϕ ∈ D(O)
the sequence
huj , ϕi
has a finite limit L(ϕ) ∈ C. Then the following hold.

1. The correspondence D(O) ∋ ϕ 7→ L(ϕ) ∈ C is continuous in the sense of Definition

11.1.1 and thus defines a generalized function.
2. If K ⊂ O is a compact set and (ϕj ) ∈ DK (O) is a sequence of smooth functions
converging uniformly togethere with all their derivatives to a smooth function ϕ ∈
DK (O), then
lim huj , ϕj i = L(ϕ).
j→∞
11.1. GENERALIZED FUNCTIONS AND CURRENTS 531

⊔
⊓

Definition 11.1.6. Any p.d.o. L : D(O) → D(O) extends to a linear operator

L : C −∞ (O) → C −∞ (O),

according to the rule

Lu, ϕ̄i := u, L∗ ϕ , ∀u ∈ C −∞ (O), ϕ ∈ D(O),

where ϕ̄ denotes the complex conjugate of ϕ. In particular, for any multi-index α ∈ Zn≥0
and any u ∈ C −∞ (O) the distributional derivative ∂ α u of u is determined by the equality

∂ α u, ϕ := (−1)|α| u, ∂ α ϕ , ∀ϕ ∈ D(O).

Example 11.1.7. Let δ0 ∈ C −∞ (Rn ) be the Dirac function concentrated at 0. Then

∂ α δ0 , ϕ = (−1)α ∂ α ϕ(0).

Consider the Heaviside function

(
1, x > 0,
H : R → R, H(x) =
0, x ≤ 0.

Being a locally integrable function, H defines a generalized function on R. Moreover

d
H = δ0 .
dx
Consider the continuous function f : R → R
(
x, x > 0,
f (x) =
0, x ≤ 0.

df
Then dx = H as generalized functions on R. ⊔
⊓

If O0 , O1 are open subsets of Rn , such that O0 ⊂ O1 , then we have an extension-by-zero

inclusion D(O0 ) ֒→ D(O1 ). This induces a restriction map

C −∞ (O1 ) ∋ u 7→ u O0
∈ C −∞ (O0 ),

u O0
, ϕ = u, ϕ , ∀u ∈ D(O0 ) ⊂ D(O1 ).
Definition 11.1.8. Let u ∈ C −∞ (O), O ⊂ Rn .

1. The support of u, denoted by supp u, is the complement in O of the largest open

subset Ω ⊂ O such that u Ω = 0.

2. The singular support of u, denoted by sing supp u, is the complement in O of the

largest open set Ω ⊂ O such that u Ω ∈ C ∞ (Ω).
532 CHAPTER 11. SPECTRAL GEOMETRY

⊔
⊓

Remark 11.1.9. Suppose that u ∈ C −∞ (O) and ϕ ∈ C ∞ (O) is such that

supp ϕu ⊂ O \ sing supp u.

Then supp ϕu ∈ O \ sing supp u and sing supp ϕu = ∅, i.e., the generalized function
coincides with a smooth function. More precisely, it coincides with the extension by zero
of the smooth function (ϕu) O\sing supp u . ⊔
⊓

Example 11.1.10. (a) If H denotes the Heaviside function then

supp H = [0, ∞) sing supp H = {0}.

(b) For any multi-index α ∈ Zn≥0 the generalized function ∂xα δ0 satisfies

supp ∂xα δ0 = sing supp ∂xα δ0 = {0}. ⊔

⊓

Fix a compact set K ⊂ O and denote by C-∞ (O)K the set of generalized function u
on O such that supp u ⊂ K. Observe that any u ∈ C-∞ (O)K induces a linear function
u : C ∞ (O) → R as follows: fix η ∈ D(O) such that η = 1 in a neighborhood of K. For
φ ∈ C ∞ (O) we set
hu, φi∗ := hu, ηφi. (11.1.2)
Note that hu, ηφi is independent of the choice of η.
Indeed, if η0 , η1 are two test functions equal to 1 on some neighborhood of K, then
η0 φ − η1 φ ∈ D(O \ K) so that
hu, η0 φ − η1 φi = 0.
We denote by C-∞ (O)cpt the space of distributions with compact support.
Note that the distributional derivatives of continuous functions are generalized func-
tions of finite order. We have the following converse, [118, Chap.III, Thm. XXI]
Theorem 11.1.11. Let u ∈ C −∞ (Rn ). Then for any open set Ω ⊂ Rn with compact
closure Ω̄ and for any open neighborhood O of Ω̄ we can find a continuous function

Φ : Rn → R

and a multi-index α ∈ Zn≥0 such that

supp Φ ⊂ O and (∂ α Φ)|Ω = u |Ω .

In particular, any compactly supported distribution has finite order. ⊔

⊓

The distributions with support consisting of a single point have a particularly simple
structure, [118, Chap. III, Thm. XXXV].
Theorem 11.1.12. Let u ∈ C −∞ (Rn ). If supp u = {0}, then there exists an integer
m ≥ 0 and complex constants cα , α ∈ Zn≥0 , |α| ≤ m, such that
X
u= cα ∂ α δ0 . ⊔
⊓
|α|≤m
11.1. GENERALIZED FUNCTIONS AND CURRENTS 533

Note that if O0 , O1 ⊂ Rn are open sets and Ψ : O0 → O1 is a diffeomorphism, then we

have a pullback operation Ψ∗ : D(O1 ) → D(O0 ) and a dual, pushforward operation

Ψ∗ : C-∞(O0 ) → C-∞ (O1 ),

hΨ∗ u, ϕi = hu, Ψ∗ ϕi, ∀u ∈ C-∞ (O0 ), ϕ ∈ D(O1 ).

Given u ∈ C ∞ (Rn ) and v ∈ D(Rn ) we define their convolution u ∗ v ∈ C ∞ (Rm ) by
Z Z
y→x−z
u ∗ v(x) = u(x − y)v(y) dy = u(z)v(x − z)dz = v ∗ u(x).
Rn Rn

If we regard u ∗ v as a generalized function, then, for ϕ ∈ D(Rn ) we have

Z Z 
hu ∗ v, ϕi = u(x − y)v(y) dy ϕ(x) dx
Rn Rn
Z Z 
= v(y) u(x − y)ϕ(x) dx dy
Rn Rn

(z = x − y, x = z + y)
Z Z  Z Z 
= v(y) u(z)ϕ(z + y) dz dy = v(y) u(z)Ty∗ ϕ(z) dz dy,
Rn Rn Rn Rn

where Ty : Rn → Rn denotes the translation z 7→ z + y. If we now think of u, v as

generalized functions we have
D E
hu ∗ v, ϕi = v, u, Ty∗ ϕ .

Observe that for any u ∈ C-∞ (Rn ) and any ϕ ∈ D(Rn ) the function

Rn ∋ y 7→ ψ(y) := u, Ty∗ ϕ ∈ R

is smooth. Since v has compact support we can pair it with ψ as in (11.1.2).

Definition 11.1.13. For u ∈ C-∞ (Rn ) and v ∈ C-∞ (Rn )cpt we define u ∗ v ∈ C-∞ (Rn ) by
the equality D E D E
u ∗ v, ϕ := v, u, Ty∗ ϕ = u, v, Ty∗ ϕ . (11.1.3)
∗
⊔
⊓

Exercise 11.1.14. Let u ∈ C-∞ (R). Show that for any k ∈ Z≥0 we have
(k)
δ0 ∗ u = u(k) in the sense of distributions. ⊔
⊓

Remark 11.1.15. More generally, suppose that C ⊂ Rn is a closed convex cone satisfying
the properness condition

C ∩ (K − C) is compact for any compact subset K ⊂ C. (11.1.4)

534 CHAPTER 11. SPECTRAL GEOMETRY

The condition (11.1.4) is satisfied if for example if the map

C × C ∋ (x, y) 7→ x + y ∈ Rn

is proper. Note that if u, v ∈ C ∞ (Rn ) are such that

supp u, supp v ⊂ C,

then, ∀x ∈ Rn the function

Rn ∋ y 7→ u(x − y)v(y) ∈ C

has compact support so the convolution of these functions.

Suppose now that ϕ ∈ D(Rn ), K := supp ϕ, and
Z
ψ(y) = u(z)ϕ(z + y)dz, ∀y ∈ Rn .
Rn

Note that
u(z)ϕ(z + y) 6= 0 =⇒ z ∈ C, y + z ∈ K =⇒ y ∈ K − C
so that
supp ψ ∈ K − C.
In particular, supp v(y)ψ(y) ⊂ K ∩ (K − C) is compact so the integral
Z
v(y)ψ(y)dy
Rn

is well defined.
More generally if u, v ∈ C −∞ (Rn ) are supported in a closed convex cone satisfying the
properness condition (11.1.4), then the convolution u ∗ v is well defined. ⊔
⊓

11.1.2 Currents
For any open subset O ⊂ Rn and any k ≤ n we denote by Ωk (O) the space of k-dimensional
currents on O, i.e., the topological dual of the space Ωkcpt (O) of complex valued, degree
k-forms on O with compact support. More precisely, a linear functional

u : Ωkcpt (O) → C, ϕ 7→ hϕ, ui

belongs to Ωk (O) if it satisfies the continuity condition (11.1.1).

Example 11.1.16. (a) The bijection

D(O) ∋ ϕ 7→ ϕdt ∧ dx1 ∧ · · · ∧ dxm ∈ Ωncpt (O)

determines a natural bijection

C −∞ (O) ∋ u 7→ [u] ∈ Ωn (O),

hϕ ∧ dx1 ∧ · · · ∧ dxm , [u]i = hu, ϕi, ∀ϕ ∈ D(O).

11.1. GENERALIZED FUNCTIONS AND CURRENTS 535

In particular, we have a bijection C −∞ (R) → Ω1 (R).

(b) If S ⊂ O is a properly embedded1 k-dimensional oriented submanifold, possibly with
boundary, then integration along S defines a k-dimensional current [S],
Z
hϕ, [S]i := ϕ, ∀ϕ ∈ Ωkcpt (O).
S

We say that [S] is the integration current or current of integration defined by S.

(c) A form η ∈ Ωn−k O) defines a k-dimensional current [η], by setting
Z
hϕ, [η]i = ϕ ∧ η, ∀ϕ ∈ Ωkcpt (O). ⊔
⊓
O

Definition 11.1.17 (Boundary operator). The boundary operator

∂ : Ωk (O) → Ωk−1 (O)

k−1
is the dual of the exterior derivative d : Ωcpt (O) → Ωkcpt (O), i.e., for any u ∈ Ωk (O) we
have
k−1
hϕ, ∂ui := hdϕ, ui, ∀ϕ ∈ Ωcpt (O).
The current ∂u is called the boundary of u. ⊔
⊓

Example 11.1.18. (a) If S ֒→ O is a smooth, properly embedded, oriented k-dimensional

submanifold with boundary, then Stokes’ formula implies that

∂[S] = [∂S],

where [∂S] is the integration current defined by the boundary ∂S equipped with the
outer=-normal-first orientation.
(b) Let η ∈ Ωn−k (O) and denote by [η] the associated k-dimensional current. from the
equality
d ϕ ∧ η) = dϕ ∧ η + (−1)k−1 ϕ ∧ η
and Stokes’ formula we deduce that

∂[η] = (−1)k [dη] = (−1)dim[η] [dη]. ⊔

⊓

11.1.3 Temperate distributions and the Fourier transform

We begin by recalling the definition of the Schwartz space Sn = S (Rn ). Roughly speak-
ing, this is the subspace of C ∞ (Rn ) consisting of functions with fast decay at ∞. Here is
the precise definition.
For x ∈ Rn we set p
hxi := 1 + |x|2 .
1
A subset S ⊂ O is properly embedded if the inclusion i : S ֒→ O is a proper map.
536 CHAPTER 11. SPECTRAL GEOMETRY

For k ∈ Z≥0 we define

X
pk : C ∞ (Rn ) → [0, ∞], pk (ϕ) = sup hxik |∂xα ϕ(x)|.
x∈Rn
|α|≤k

Then 
Sn = S (Rn ) := ϕ ∈ C ∞ (Rn ); pk (ϕ) < ∞, ∀k ∈ Z≥0 .
The functions in Sn are called Schwartz functions. Let us observe that
2
D(Rn ) ⊂ Sn , e−|x| ∈ Sn .

The space Sn is equipped with a natural topology defined by the metric

X 1

dist(ϕ, ψ) = k
min 1, pk (ϕ − ψ) .
2
k≥0

Thus ϕν → ϕ as ν → ∞ iff

lim pk (ϕν − ϕ) = 0, ∀k ≥ 0.
ν→∞

We have the following result, [66, Lemma 7.1.8].

Lemma 11.1.19. The subspace D(Rn ) is dense in Sn . ⊔
⊓

Definition 11.1.20. A temperate distribution on Rn is a continuous linear functional

u : Sn → C,

i.e., there exists k ∈ Z≥0 and C > 0 such that

hu, ϕi ≤ Cpk (ϕ), ∀ϕ ∈ Sn .

We denote by Sn′ the space of temperate distributions. ⊔

⊓

We have strict inclusions

−∞
Ccpt (Rn ) ⊂ Sn′ ⊂ C-∞ (Rn ).

We denote by (−, −) the canonical inner product on Rn .

Definition 11.1.21. The Fourier transform on Rn is the linear map

F : Sn → C ∞ (Rn ), Sn ∋ ϕ 7→ ϕ b = F [ϕ],
Z
b
ϕ(ξ) = e−i(ξ,x) ϕ(x)dx, ∀ξ ∈ Rn . ⊔
⊓
Rn
We list below a few fundamental results concerning the Fourier transform. For a proof
we refer to [39, Thm. 14.8].
11.1. GENERALIZED FUNCTIONS AND CURRENTS 537

Theorem 11.1.22. (a) For any ϕ ∈ Sn , we have ϕ

b ∈ Sn and the resulting linear map

F : Sn → Sn

is continuous.
(b) For any ϕ, ψ ∈ Sn
Z Z
b
ϕ(x)ψ(x)dx = b
ϕ(x)ψ(x)dx. (11.1.5)
Rn Rn

(c) (Parseval formula) we have

1
ϕ, ψ L2 (Rn )
= b ψb
ϕ, L2 (Rn )
,
(2π)n

i.e., Z Z
1 b
ϕ(x)ψ(x)dx = b ψ(ξ)dξ.
ϕ(ξ)
Rn (2π)n Rn

(d) (Fourier inversion formula) For any ϕ ∈ Sn we have

Z
1
ϕ(x) = ei(ξ,x) ϕ(ξ)dξ.
b
(2π)n Rn

(d) For any ϕ, ψ ∈ Sn , their convolution ϕ ∗ ψ is also a Schwartz function and

ϕ[
∗ ψ(ξ) = ϕ(ξ) b
b · ψ(ξ). ⊔
⊓

Using the equality (11.1.5) above we can extend the Fourier transform to temperate
distributions.

Definition 11.1.23. The Fourier transform of a temperate distribution u ∈ Sn′ is the

b ∈ Sn′ defined by
temperate distribution u

b, ψ := u, ψb , ∀ψ ∈ Sn i.
u ⊔
⊓

11.1.4 Linear differential equations with distributional data

We gather here a few basic facts about generalized function solutions of ordinary differen-
tial equations. Our presentation follows closely [66, Chap. III]. We begin with the simplest
differential equation.

Theorem 11.1.24. Suppose I ⊂ R is an open interval. Then the only solutions of the
differential equation
du
= 0, u ∈ C −∞ (I)
dt
are the functions constant on I.
538 CHAPTER 11. SPECTRAL GEOMETRY

Proof. This means that

hu, φ̇i = 0, ∀φ ∈ D(I).
Consider the generalized function µI ∈ C −∞ (I),
Z
hµI , ψi = ψ(s)ds, ∀ψ ∈ C0∞ (I).
I

In other words, µI is the generalized function determined by the indicator function of I.

Observe that if Z
ψ = hµI , ψi = 0,
I

then the function Z t

φ(t) = ψ(s)ds
−∞

is supported on I and,
hu, ψi = hu, φ̇i = −hu̇, φi = 0.
Hence
hµI , ψi = 0 ⇒ hu, ψi = 0.
Now choose ψ0 ∈ C0∞ (I) such that

hµI , ψ0 i = 1.

Set c0 := hu, ψ0 i.Then u = c0 µI . Indeed, if φ ∈ D(I) and

Z
c = φ(s)ds, ψ = φ − cψ0
I

then
0 = hu, ψi = hu, φi − chu, ψ0 i,
so
hu, φi = cc0 = c0 hµI , φi.
⊔
⊓

Corollary 11.1.25. Suppose that I ⊂ R is an open interval,

n ∈ Z>0 , A ∈ C ∞ I, End(Rn ) , f ∈ C 0 (I, Rn ).

If u ∈ C −∞ (I)n satisfies the linear differential equation

du
+ A(t)u = f (t), (11.1.6)
dt

then u ∈ C 1 (I, Rn ), i.e., u is a classical solution of (11.1.6).

11.1. GENERALIZED FUNCTIONS AND CURRENTS 539

Proof. Choose

E ∈ C ∞ I, End(Rn )

such that Ė = EA, and E(t) is invertible for any t ∈ I. We deduce that

d
( E(t)u(t) ) = E(t)f (t).
dt
If v ∈ C 1 (I, Rn ) satisfies v ′ = E(t)f (t), then we deduce

d
(E(t)u(t) − v(t)) = 0.
dt
Theorem 11.1.24 implies that there exists c ∈ Rn such that

E(t)u(t) = v(t) + c.

Hence u(t) = E(t)−1 v(t)+ E(t)−1 c is a classical solution of the above differential equation.
⊔
⊓

Corollary 11.1.26. Suppose that I ⊂ R is an open interval, a0 , . . . , am−1 ∈ C ∞ (I),

f (t) ∈ C 0 (I). If v ∈ C −∞ (I) satisfies the differential equation
m−1
dm u X dj v
+ aj j = f (t),
dtm dt
j=0

then v ∈ C m (I), i.e., v is a classical solution of the above equation.

Proof. Observe that the (vector valued) generalized function

u(t) = v(t), v ′ (t), . . . , v (m−1) (t) ∈ C −∞ (I)m

satisfies a linear first order ordinary differential equation with smooth coefficients and
continuous nonhomogeneous term. Corollary 11.1.25 implies that u is a classical solution.
⊔
⊓

Corollary 11.1.27. Suppose that Ω ⊂ Rn is an open set and I ⊂ R is an open interval.

We denote by x1 , . . . , xn the Cartesian coordinates on Ω and by t the Cartesian coordinate
on I. If u ∈ C −∞ (I × Ω) is a generalized function satisfying the differential equation
∂t u = 0, then there exists u0 ∈ C −∞ (Ω) such that
Z
hu, φ(t, x)i = hu0 (x), φ(t, x)idt, ∀φ ∈ D(I × Ω). (11.1.7)
I

Proof. For any ψ ∈ D(Ω) define uψ ∈ C −∞ (I) by setting

huψ , ϕi = hu, ψϕi, ∀ϕ ∈ D(I).

540 CHAPTER 11. SPECTRAL GEOMETRY

d
Then dt uψ = 0. Using Theorem 11.1.24 we deduce that uψ is a constant function. The
map
D(Ω) ∋ ψ 7→ uψ ∈ C −∞ (I)
is obviously linear and continuous from the usual topology of D(Ω) to the weak topology
of C −∞ (I). The image of this map is the 1-dimensional subspace consisting of constant
functions so that the resulting map

D(Ω) ∋ ψ 7→ uψ ∈ R

is continuous. This defines a generalized function u0 ∈ C −∞ (Ω) satisfying (11.1.7) for

φ ∈ D(I × Ω) of the form φ(t, x) = ϕ(t)ψ(x), i.e.
Z Z
hu, ϕψi = uψ ϕ(t)dt = hu0 (x), ϕ(t)ψ(x)idt.
I I

The desired conclusion follows by observing that the linear span of functions of the form
ϕ(t)ψ(x), ϕ ∈ D(I), ψ ∈ D(Ω) is dense in D(I × Ω).
⊔
⊓

For any open interval I ⊂ R, any open set O ⊂ Rn we denote by C 0,−∞ (I × O) the
space of (weakly) continuous maps I ∋ t → ut ∈ C −∞ (O), i.e.,

t 7→ hut , ϕi

is continuous for any ϕ ∈ D(O). Note that we have inclusions inclusion

C 0 (I × O) ⊂ C 0,−∞ (I × O) ⊂ C −∞ (I × O).

The first inclusion associates to each continuous function u ∈ C 0 (I × O) the path t 7→ ut

of continuous functions on O, ut (x) = u(t, x), ∀x ∈ O. Moreover

∀u ∈ C 0,−∞ (I × O), ∀α ∈ Zn≥0 , t 7→ ∂xα ut ∈ C 0,−∞ (I × O). (11.1.8)

Let t denote the Cartesian coordinate along I. For k > 0 we define


C k,−∞(I × O) := u ∈ C 0,−∞ (I × O); ∂tj u ∈ C 0,−∞ (I × O), 1 ≤ j ≤ k .

We have the following result, [66, Thm. 4.4.8].

Proposition 11.1.28. Suppose that u ∈ C −∞ (R × Rm ) and its restriction to R>0 × Rm
satisfies the differential equation
ν−1
X
∂tν u + Aj (∂x )∂tj u = F, (11.1.9)
j=0

where Aj (∂) : C ∞ (O) → C ∞ (O) are partial differential operators with smooth coefficients
independent of t. If F ∈ C 0,−inf ty (Rm ), then

u ∈ C ν,−∞ (I × O).
11.1. GENERALIZED FUNCTIONS AND CURRENTS 541

Moreover, if
lim Ft
tց0

exists in C −∞ (Rm ), then the limits

lim ∂tj ut , 1 ≤ j ≤ ν − 1
tց0

exist in C −∞ (Rm ).

Proof. It suffices to show that, for any r > 0 and any bounded open set Ω ⊂ Rm , we have

u|(0,r)×Ω ∈ C ν,−∞ (J × Ω).

Fix such a Ω.
Observe next that we can reduce (11.1.9) to a first order (in ∂t ) system of partial
differential equation by replacing u with the vector-valued generalized function

v = (∂tj u)0≤j≤ν−1 .

Thus we can assume that ν = 1, i.e., (11.1.9) has the form

∂t u + A(∂x )u = F. (11.1.10)

Fix ε > 0 and set J := (−ε, r). Using Theorem 11.1.11 we deduce that we can find an
integer µ ≥ 0, multiindices αk ∈ Zm
≥0 , 0 ≤ k ≤ µ, and generalized functions



uk ∈ C 0 J × Ω ⊂ C 0,−∞ J × Ω , 0 ≤ k ≤ µ

such that !
m
X
u|J×Ω = ∂tk ∂xαk ∈ uk |J×Ω .
k=0

From (11.1.8) we deduce

u|J×Ω ∈ C 0,−∞ (J × Ω).
We say that the t-order of u is ≤ µ, and we write this ordt (u) ≤ µ. We will prove by
induction on µ that u ∈ C 1,−∞ (J × Ω).
If µ = 0, so that the t-order of u is ≤ 0, then

u = ∂xα0 u0 ∈ C 0 (J × Ω).

Using (11.1.10) we deduce

∂t u = −A0 (∂x )∂xα0 u0 (t) + F (t) ∈ C 0,−∞ ((0, r) × Ω) ⇒ u ∈ C 1,−∞((0, r) × Ω).

| {z }
G

Moreover, for any ϕ ∈ D(Ω) we have

Z t0
limhu(t), ϕi = hu(t0 ), ϕi − G(t), ϕ dt.
tց0 0
542 CHAPTER 11. SPECTRAL GEOMETRY

The last integral is finite since the function

[0, t0 ] 7→ hG(t), ϕi
is continuous.
Suppose now that ordt (u) ≤ µ, µ > 0. Then
m
X α
∂t u = −A0 (∂x )u = ∂tj ∂x j uj + F
j=0
 
m
X α
= −∂t  ∂tj−1 A0 (∂x )∂x j uj  + F − A0 (∂x )∂xα0 u0 .
j=1

We set
m
X α
v =u+ ∂tj−1 A0 (∂x )∂x j uj ,
j=1

so that
∂t v = F − A0 (∂x )∂xα0 u0 ∈ C 0,−∞ ((0, r) × Ω).
| {z }
G
Now choose w ∈ C 1,−∞ ((0, r) × Ω), (0, r) ∋ t 7→ wt ∈ C −∞ (Ω) such that ∂t w = G. E.g.,
Z t
hwt , ϕi = hGs , ϕi, ∀ϕ ∈ D(Ω).
t0

Clearly
lim wt , ϕ
tց0

exists ∀ϕ ∈ D(Ω) since the limit

lim G(t), ϕ
tց0

exists ∀ϕ ∈ D(Ω). Then

∂t (v − w) = 0 in C −∞ (0, r) × Ω .
Invoking Corollary 11.1.27 we conclude that v − w is a constant map (0, r) → C −∞ (Ω).
Thus


ordt (v) = ordt (v − w) + w ≤ max ordt (v − w), ordt w ≤ 0.
Hence
   
m
X m
X
ordt u = ordt  v − ∂tj−1 A0 (∂x )uj  ≤ ordt  ∂tj−1 A0 (∂x )uj  ≤ m − 1.
j=1 j=1

We can now conclude by induction that u ∈ C 1,−∞ (J × Ω). By construction, the limit
lim ut , ϕ
tց0

exists ∀ϕ ∈ D(Ω).
⊔
⊓
11.2. IMPORTANT FAMILIES OF GENERALIZED FUNCTIONS 543

11.2 Important families of generalized functions

We want to spend some time discussing in some detail several important examples of
generalized functions that are rarely discussed in a traditional graduate course. We will
make use of these generalized functions when we investigate the spectral data of a Riemann
manifold. Throughout the following discussion Euler’s Gamma function
Z ∞
Γ(a) = e−t ts−1 ds, Re s > 0,
0

plays an important role. In particular, we will use less familiar properties of this function.
They can be found in [84, 134].
An important concept we will use in the sequel is that of holomorphic family of gen-
eralized functions. A family of generalized functions (ua )a∈A ∈ C −∞ (R) parametrized by
an open subset A ⊂ C is called holomorphic if, for any ϕ ∈ D(R) the function

A ∋ a 7→ hua , ϕi ∈ C

is holomorphic. The concept of meromorphic family is defined in a similar fashion.

11.2.1 Some classical generalized functions on the real axis

If a ∈ C, Re a > −1, then the function xa+ : R → C defined by
(
xa , x > 0,
xa+ :=
0, x ≤ 0,

is locally integrable and thus defines a generalized function xa+ ∈ C −∞ (R). A simple
computation shows that
x · xa+ = xa+1
+ , ∀ Re a > −1, (11.2.1a)
d a
x = axa−1 + , ∀ Re a > 0. (11.2.1b)
dx +
We want to extend the definition of xa+ to all a ∈ C. For φ ∈ D(R) the function
Z ∞

Re a > −1 ∋ a 7→ Ia (φ) := hxa+ , φi = xa ϕ(x)dx
0

is holomorphic. Using (11.2.1b) iteratively we deduce

(−1)k

Ia (φ) = Ia+k φ(k) , ∀ Re a > −1, k ∈ Z≥0 . (11.2.2)
(a + 1) · · · (a + k)
If a ∈ C \ Z<0 then we can define

(−1)k

Ia (ϕ) := Ia+k φ(k) , ∀k > max(0, −1 − Re a).
(a + 1) · · · (a + k)
The equality (11.2.2) shows that the right hand side of the above equality is independent
of the choice of k > max(0, −1 − Re a).
544 CHAPTER 11. SPECTRAL GEOMETRY

This gives a meromorphic extension of the function a 7→ Ia (ϕ) with simple poles located
at Z<0 and thus allows us to define the generalized function xa+ for any a ∈ C\Z<0 . Observe
that
(−1)k
1
lim (a + k)Ia (ϕ) = I0 φ(k) = φ(k−1) (0), (11.2.3)
a→−k (1 − k)(2 − k) · · · (−1)! (k − 1)!

so that
(−1)k−1 (k−1)
lim (a + k)xa+ = δ .
a→−k (k − 1)! 0
More explicitly, for a ∈ C \ Z<0 , Re a > −n − 1, n ∈ Z>0 we have
Z 1  X xk φ(k) (0) 
n−1
hxa+ , φi = xa φ(x) − dx
0 k!
k=0
Z ∞ Xn
φ(k−1) (0)
+ xa φ(x)dx + .
1 (k − 1)!(a + k)
k=1

To eliminate the poles of a 7→ xa+ we will divide xa+ by a meromorphic function with simple
poles at Z<0 .
For any Re s > 0 Euler’s Gamma-function satisfies

Γ(s + 1) = sΓ(s).

We can use this equality to extend Γ(s) to a meromorphic function on C with simple poles
at Z≤0 . Note that the residue at −k ∈ Z≤0 can be determined from the equality

Γ(s + k + 1) Γ(1) (−1)k

lim (s + k)Γ(s) = lim = (−1)k = . (11.2.4)
s→−k s→−k s(s + 1) · · · (s + k − 1) k! k!

Using the equality

π
Γ(a)Γ(1 − a) = , Re a ∈ (0, 1)
sin πa
we deduce that Γ has no zeros in Z. Set
1
χa+ (x) := xa , Re a > −1. (11.2.5)
Γ(a + 1) +

The correspondence 
Re a > −1 ∋ a 7→ χa+ ∈ C −∞ (R)
and in this half-plane it satisfies the equality
d a+1
χ = χa+ .
dx +
We can use this equality to extend χa+ analytically to all a ∈ C,

dn a+n
χa+ = χ
dxn +
11.2. IMPORTANT FAMILIES OF GENERALIZED FUNCTIONS 545

where n is a positive integer such that n + Re a > −1. Since χ0+ is the Heaviside function
we deduce
−k−1 (k) dk
χ+ = δ0 = k δ0 , ∀k ∈ Z≥0 . (11.2.6)
dx
Similarly, for Re a > −1 we can define
(
a 0, x ≥ 0,
x− = a
|x| , x < 0,

and for any a

(
1 0, x≥0
χa− (x) = χa+ (−x) = a
Re a > −1.
Γ(a + 1) |x| , x < 0,

Equivalently
hχa− , ϕi = hχa+ , φ̌i, ∀φ ∈ D(R),
where
φ̌(x) = φ(−x), ∀x ∈ R.
Note that
dn a+n
χ = (−1)n χa− .
dxn −
Observe that for Re a > −1 we have

xa+ + xa− = |x|a .

Let φ ∈ D(R). The function a 7→ hxa+ , φi is meromorphic on C with simple poles at

a = −1, −2, . . . , −k, . . .

According to (11.2.3) the residue at a = −k is

1 (−1)(k−1) (k−1)
φ(k−1) (0) = hδ , φi.
(k − 1) (k − 1)! 0

Similarly, the function a 7→ hxa− , φi is meromorphic on C with simple poles at

a = −1, −2, . . . , −k, . . .

and the residue at a = −k is

1 1 (k−1)
φ̌(k−1) (0) = hδ , φi.
(k − 1) (k − 1)! 0
Thus the function
a 7→ Gϕ (a) := h|x|a , φi
is meromorphic on C with simple poles at

a = −1, −3, . . . , −(2j + 1), . . . , j ∈ Z≥0 .

546 CHAPTER 11. SPECTRAL GEOMETRY

The residue at a = −(2j + 1) is

2 (2j)
hδ0 , φi.
(2j)!
1
The function a 7→ Γ( a+1 )
is holomorphic on C with simple zeroes at
2

a = −1, −3, . . . , −(2j + 1), . . . , j ∈ Z≥0 .

The equality (11.2.4) shows that the derivative of this function at a = −(2j + 1) is
(−1)j+1 (j + 1)!. We conclude that, for any ϕ ∈ D(R), the function

1
a 7→ Fϕ (a) := Gϕ (a)
Γ( a+1
2 )

is holomorphic on C. We denote by |χ|a the generalized function defined by the equality

h|χ|a , ϕi := Fϕ (a), ∀ϕ ∈ D(R).

Less precisely, but more intuitively, this means

1
|χ|a (x) := a
a+1 |x| . (11.2.7)
Γ( 2 )

Note that
(−1)j+1 2(j + 1)! (2j)
|χ|−(2j+1) = δ0 . (11.2.8)
(2j)!
For y 6= 0 and a ∈ C we define


(k−1)
(x + iy)a := ea log |x+iy|+i arg(x+iy)
hδ0 , φi,

where
arg(x + iy) ∈ (−π, π), x + iy = |x + iy|ei arg(x+iy) .
We deduce that for Re a > −1 we have

(x ± i0)a := lim (x + iy)a = xa+ + e±πia xa− . (11.2.9)

y→0±

We have a more precise result.

Theorem 11.2.1. Consider a holomorphic function f defined on the strip

Z := z = x + iy ∈ C; 0 < y < γ .

If there exists constant C > 0 and an integer N ∈ Z≥0 such that

|f (x + iy)| ≤ C|y|−N , ∀x + iy ∈ Z,

then f (• + iy) has a limit f0 ∈ C −∞ (R) when y → 0. Moreover, this limit is a generalized
function of order at most N + 1. We denote it by f (x + i0).
11.2. IMPORTANT FAMILIES OF GENERALIZED FUNCTIONS 547

Proof. Fix z0 ∈ Z and consider

Z z
F (z) = f (ζ)dζ.
z0

Then (
C0 |y|1−N , N >1
|F (z)| ≤ .
−C0 log y + C1 , N = 1

If N = 1 it follows that the integral of F is continuous on the closure of Z since log t is

integrable on the interval [0, γ].
We deduce that there exists a continuous function G : Z̄ → C whose restriction to Z
is holomorphic and such that
dN +1
F (z) = N +1 G(z).
dz
This implies that for fixed y we have

dN +1
f (x + iy) = G(x + iy).
dxN +1
Since G(• + iy) converges in the sense of generalized functions to a continuous function
G(x + i0), we deduce that f (x + iy) converges as y ց 0 to the generalized function
dN+1
dxn+1
G(x + i0) which has order ≤ N + 1. ⊔
⊓

Theorem 11.2.1 implies that (11.2.9) holds for all a ∈ C \ Z<0 and that the family

a 7→ (x ± i0)a

is a holomorphic extension to C of the, a priori meromorphic, family

a 7→ xa+ + e±aπi xa− .

In particular, we deduce that

1  
xa+ = eaπi (x − i0)a − e−aπi (x + i0)a , (11.2.10a)
2i sin πa

1

xa+ = (x + i0)a − (x − i0)a . (11.2.10b)
2i sin πa
Let us observe that
d
(x + i0)a = a(x + i0)a−1 .
dx
Proposition 11.2.2.
iπ(a+1)
χc
a (ξ) = e∓
±
2 (ξ ∓ i0)−(a+1) (11.2.11a)

da (ξ) = √π2a+1 |χ|−(a+1) (ξ), ∀a ∈ C

|χ| (11.2.11b)
548 CHAPTER 11. SPECTRAL GEOMETRY

Proof. Let Re a > −1 and ε > 0. Set fε (x) := e−εx xa+ so that fε → f0 = xa+ in the space
S ′ of Schwartz generalized functions. The Fourier transform of fε is
Z ∞
b
fε (ξ) = e−x(ε+iξ) xa dx.
0

Set ζ := ε + iξ so that Z ∞
fbε (ξ) = ζ −a e−xζ (xζ)a dx
0
(z := xζ)
Z
= ζ −a−1 e−z z a dz = ζ −(a+1) Γ(a + 1) = Γ(a + 1)(ǫ + iξ)−(a+1)
R>0 ·ζ

iπ(a+1)
= e− 2 (ξ − iε)−(a+1) .
Hence iπ(a+1)
fb0 (ξ) = Γ(a + 1)e− 2 (ξ − i0)−(a+1) .
Arguing in a similar fashion we deduce that for Re a > −1 the Fourier transform of xa− is
iπ(a+1)
Γ(a + 1)e 2 (ξ + i0)−(a+1) .

Hence the Fourier transform of νa (x) = |x|a is

iπ(a+1) iπ(a+1)

νba (ξ) = Γ(a + 1) e− 2 (ξ − i0)−(a+1) + e 2 (ξ + i0)−(a+1) .

Assuming that a 6∈ Z we deduce from (11.2.9) that

1 iπ(a+1) iπ(a+1)
νba (ξ) = e− 2 (ξ − i0)−(a+1) + e 2 (ξ + i0)−(a+1)
Γ(a + 1)
iπ(a+1)
  iπ(a+1)
 
−(a+1) −(a+1) −(a+1) −(a+1)
= e− 2 ξ+ + e(a+1)πi ξ− +e 2 ξ+ + e−(a+1)πi ξ−
   
π(a + 1) −(a+1) −aπ
= 2 cos |ξ| = 2 sin |ξ|−(a+1) .
2 2
At this point we invoke the reflection formula for the Gamma function
 
−aπ π
sin =
2 Γ(− 2 )Γ(1 + a2 )
a

to conclude that
1 π 1 −(a+1) π
νba (ξ) = a a |ξ| = |χ|−(a+1) (ξ).
Γ(a + 1) Γ(1 + 2 ) Γ(− 2 ) Γ(1 + a2 )

Thus
da (ξ) = 2πΓ(a + 1) −(a+1)
|χ| a |χ| (ξ).
Γ( a+1
2 )Γ(1 + 2 )
11.2. IMPORTANT FAMILIES OF GENERALIZED FUNCTIONS 549

Using the duplication formula we deduce

  
√ a a+1 a
πΓ(a + 1) = 2 Γ Γ 1+ .
2 2

Hence
da (ξ) = √π2a+1 |χ|−(a+1) (ξ), ∀ Re a > −1, a 6∈ Z.
|χ|
The equality (11.2.11b) now follows by unique continuation. ⊔
⊓

Corollary 11.2.3.
χa+ ∗ χb+ = χa+b+1
+ , ∀a, b.

Proof. Pass to Fourier transforms and then use (11.2.11a). ⊔

⊓

If we define
−(1+α)
I+
α := χ+ , α ∈ C,
then we deduce
(k)
I+
k = δ0 , ∀k ∈ Z≥0 ,

and
+ +
I+
α ∗ I β = I α+β , ∀α, β ∈ C.

Noting that
(k)
I+
k ∗ f = δ0 ∗ f = f
(k)
, ∀f ∈ D(R),
we could conclude that the operator

D(R) ∋ u 7→ I + ∞
α ∗ u ∈ C (R)

can be viewed as a fractional derivative of order α. The distributions I −

α , are defined as
− +
I α := Ǐ α , and satisfy similar properties.

11.2.2 Homogeneous generalized functions

Suppose Y ⊂ Rn , X ⊂ RN are open sets and Ψ : X → Y is a submersion. Denote by
y = (y 1 , . . . , y n ) the Euclidean coordinates on Y and by x = (x1 , . . . , xN ) the Euclidean
cordinates on X. We have a natural pushforward map

Ψ∗ : D(X) → D(Y )

motivated follows. The co-area formula (9.1.12) implies that for any ϕ ∈ D(X), u ∈ D(Y )
we have
Z Z Z 
1
(Ψ∗ u)(x)ϕ(x)|dx| = u(y) ϕ(x)|dVΨ−1 (y) (x)| |dy|,
X Y Ψ−1 (y) |∇Ψ|(x)
550 CHAPTER 11. SPECTRAL GEOMETRY

where |∇Ψ|(x) denotes the Jacobian of Ψ at the point x ∈ X, and |dVΨ−1 (y) (x)| denotes
the Euclidean volume density on the fiber Ψ−1 (x). More precisely,
p
|∇Ψ|(x) = det(DΨ)(x)∗ (DΨ)(x)

where DΨ(x) : RN → Rn denotes the differential of Ψ at x. Note that if n = 1, then

|∇Ψ|(x) is indeed the length of the gradient of Ψ at x.
For any ϕ ∈ D(X) we set
Z
1
Ψ∗ ϕ(y) := ϕ(x) |dVΨ−1 (y) (x)|. (11.2.12)
Ψ−1 (y) |∇Ψ|(x)

More conceptually, the density Ψ∗ ϕ(y)|dy| on Y is the pushforward of the density ϕ(x) |dx|
on X. Note that Ψ∗ ϕ ∈ D(Y ). We define the pullback

Ψ∗ : C −∞ (Y ) → C −∞ (X)

where for u ∈ C −∞ (Y ) we define Ψ∗ u ∈ C −∞ (X) via the equality

hΨ∗ u, ϕi := hu, Ψ∗ ϕi, ∀ϕ ∈ D(X). (11.2.13)

Note that if u ∈ C ∞ (Y ) ⊂ C −∞ (Y ), then Ψ∗ u ∈ C ∞ (X) and, more precisely

Ψ∗ u(x) = u Ψ(x) .

Indeed, using the coarea formula, we deduce

Z Z Z !

1
u Ψ(x) ϕ(x) |dVX (x)| = u(y) ϕ(x)|dVΨ−1 (y)| (x)| |dVY (y)|
X Y Ψ−1 (y) |∇Ψ(x)

More generally, if Ψ is not a submersion, but its singular values are outside an open neigh-
borhood of U of sing supp u, then we can define Ψ∗ u via the following cutoff construction.
Fix a smooth function η ∈ D(U ) such that η = 1 on a neighborhood V of sing supp u
in U . The pullback Ψ∗ (ηu) is well defined via the above procedure since the restriction of
Ψ to Ψ−1 (U ) is a submersion.
According to Remark 11.1.9, the distribution (1 − η)u ∈ C −∞ ∗ Y ) is smooth and we
define Ψ∗ ( (1 − η)u) as the usual pullback of a smooth function via a smooth map. We
then set
Ψ∗ u := Ψ∗ (ηu) + Ψ∗ ( (1 − η)u ).
One can verify that the above definition is independent of the choice of η with the above
properties.
Indeed, if η0 , η1 are two such functions, then


Ψ∗ (η0 u) = Ψ∗ (η1 η0 u) + Ψ∗ (1 − η1 )η0 u ,



Ψ∗ (1 − η0 )u = Ψ∗ (η1 (1 − η0 )u) + Ψ∗ (1 − η1 )(1 − η0 )u .
By adding the above equalities and using the linearity of Ψ∗ we deduce



Ψ∗ (η0 u) + Ψ∗ (1 − η0 )u = Ψ∗ (η1 u) + Ψ∗ (1 − η1 )u .
11.2. IMPORTANT FAMILIES OF GENERALIZED FUNCTIONS 551

Example 11.2.4. Let c 6= 0 and consider the map

Qc : R → R, Qc (t) = t2 − c2 .

Note that only the point s0 = −c2 is a critical value of this map, and Q−1 (s0 ) = {0}. For
s > s0 we have p
Q−1c (s) = ± s + c .
2

If ϕ ∈ D(R) is such that supp ϕ ∩ {0} = ∅ we have

  √
√


 √ 1
ϕ s + c2 + ϕ − s + c2 , s > −c2
 2 s+c2
(Qc )∗ ϕ(s) = .


0, s ≤ −c2

Hence
1

hQ∗c δ0 , ϕ) = ϕ(c) + ϕ(−c) .
2c
Using the cutoff construction of the pullback we deduce
1
Q∗c δ0 = (δc + δ−c ). ⊔
⊓
2c

For t > 0 we denote by Rt the diffeomorphism

Rt : Rn → Rn , x 7→ tx.

Note that for ϕ ∈ D(Rn ) we have

1

(Rt )∗ ϕ(y) = −1
ϕ(t−1 y) = t−n ϕ t−1 y .
|∇Rt |(t y)
Suppose that X is a conical open subset of Rn , i.e.,

Rt (X) = X, ∀t > 0.

A generalized function u ∈ C −∞ (X) is said to be homogeneous of degree a ∈ C if

Rt∗ u = ta u, ∀t > 0.

Equivalently, this means that for any ϕ ∈ D(X) we have

ta hu, ϕi = hu, (Rt )∗ ϕi, ∀ϕ ∈ D(X). (11.2.14)

If we differentiate the last equality with respect to t at t = 1 we deduce the Euler identity
~
(a + n)hu, ϕi + hu, Eϕi = 0, ∀ϕ ∈ D(X), (11.2.15)
~ is the Euler vector field
where E
n
X
~ =
E xj ∂xj .
j=1
552 CHAPTER 11. SPECTRAL GEOMETRY

Conversely, the equality (11.2.15) implies (11.2.14). Indeed, for t ∈ R define the continuous
linear operator
St : C −∞ (X) → C −∞ (X), St u = e−ta Re∗t u.
Then
S0 = 1, St0 ◦ St0 = St0 +t1 , ∀t0 , t1 ∈ R.
The equality (11.2.15) is equivalent to
d
|t=0 St u = 0.
dt
so that
1
lim (Sh − S0 )u = 0,
h→0 h
in the sense of generalized functions. Hence
 
1 1
0 = St0 lim (Sh − S0 )u = lim (St0 +h − St0 )u.
h→0 h h→0 h

Hence
d
|t=t0 St u = 0, ∀t0 .
dt
Hence the function t 7→ St u ∈ C −∞ (X) is constant, i.e., u is homogeneous of degree a.
Example 11.2.5. (a) The generalized functions xa± , (x ± i0)a and χa+ are homogeneous
of degree a on R \ 0.
(b) The Dirac generalized function δ0 ∈ C −∞ (Rn ) is homogeneous of degree −n. Indeed
hδ0 , (Rt )∗ ϕi = t−n hδ0 , ϕ(tx)i = t−n ϕ(0) = t−n hδ0 , ϕi.
⊔
⊓

Theorem 11.2.6. Suppose that u ∈ C −∞ (R>0 ) is a homogeneous generalized function of

degree a ∈ C. Then there exists c ∈ C such that u = cta+ .

Proof. The generalized function u satisfies the Euler equation

u̇ + (a + m)u = 0.
Now invoke Corollary 11.1.25 to reach the desired conclusion. ⊔
⊓

Theorem 11.2.7. If u ∈ C −∞ (Rn \ 0) is a homogeneous generalized function of degree

e ∈ C −∞ (Rn ) which is homogeneous of degree
a ∈ C\Z≤−n , then u has a unique extention u
a. If P is a homogeneous polynomial then
]
(P u) = P · u
e.
Moreover, if a ∈ C \ Z≤−(n−1) we have

^
(∂ e, ∀j = 1, . . . , n.
xj u) = ∂xj u (11.2.16)
11.2. IMPORTANT FAMILIES OF GENERALIZED FUNCTIONS 553

Proof. Uniqueness. Suppose that that ũ and ū are two homogeneous extensions of u to
Rn . Then ũ − ū is supported at the origin. Hence it is a linear combinations of derivatives
of the Dirac generalized function. Such a linear combination cannot be homogeneous of
degree ∈ C \ Z≤−n . This proves the uniqueness of the extension.
Existence. For any φ ∈ D(Rn ) we define
a+n−1
T a φ(x) := t+ , φx (t) R
, φx (t) := φ(tx), x ∈ Rn \ 0,

where h−, −iR denotes the pairing between a generalized function and a test function on
R. Note that T a φ ∈ C ∞ (Rn \ 0).
Lemma 11.2.8. For any a ∈ C \ Z≤−n and any ϕ ∈ D(Rn ) the function T a φ is homoge-
neous of degree −a − n.

Proof. For a > −n and r > 0 we have

Z ∞
a+n−1
T a φ(rx) = ht+ , φ(rtx)i = ta+n−1 φ(rtx)dt
0
Z ∞
s:=rt −a−n
= r sa+n−1 φ(sx)ds = r −a−n T a φ(x).
0
By unique continuation, the last equality continues to hold for any a ∈ C \ Z≤−n , so that,
for any a ∈ C \ Z≤−n the function T a φ(x) is homogeneous of degree −a − n. ⊔
⊓

The key moment in the existence proof is the following auxiliary result whose proof
we postpone.
Lemma 11.2.9. If u ∈ C −∞ (Rn \ 0) is a homogeneous generalized function of degree a,
then
hu, ηi = 0 if η ∈ D(Rn \ 0) and T a η(x) = 0, ∀x ∈ Rn \ 0.
⊔
⊓

Fix a function ψ ∈ D(Rn \ 0) satisfying

Z ∞
dt
ψ(tx) = 1, ∀x 6= 0. (11.2.17)
0 t
Then ψT a φ ∈ D(Rn \ 0) for any φ ∈ D(Rn ) and
Z ∞
T a (ψT a φ)(x) = ta+n−1 ψ(tx)T a φ(tx)dt
0

(T a φ(tx) = t−a−n T a φ(x))

Z ∞
dt
= T a φ(x) ψ(tx) = T a φ(x).
0 t
We rewrite the above equality as


T a φ − ψT a φ = 0,
554 CHAPTER 11. SPECTRAL GEOMETRY

and Lemma 11.2.9 implies that

hu, φi = hu, ψT a φi, ∀φ ∈ C −∞ (Rn \ 0).

The sought for extension of u can now be easily described. More precisely, we set

u, φi := hu, ψT a φi, ∀φ ∈ D(Rn ).

he

the right-hand side of the above equality is well defined since ψT a φ ∈ D(Rn \ 0), ∀φ ∈
D(Rn ).
If P is a homogeneous polynomial, then P u is a homogeneous generalized function of
degree a + deg P ∈ C \ Z≤−n . Clearly Pfu coincides with P u
e on Rn \ 0 and the uniqueness
of the extension implies that
Pfu = P u
e on Rn .

The equality
^
(∂ e
xj u) = ∂xj u

when a ∈ C \ Z≤−(n−1) is proved in a similar fashion. ⊔

⊓

Proof of Lemma 11.2.9. We will prove that

~ + (a + n)φ = η.
η ∈ D(Rn \ 0), T a η = 0 ⇒ ∃φ ∈ D(Rn \ 0) : Eφ (11.2.18)

Clearly (11.2.18) implies Lemma 11.2.9 since

(11.2.15 )
~ + (a + n)φi
hu, ηi = hu, Eφ = 0.

To prove we observe that for φ ∈ C ∞ (Rn \ 0) we have

d 1~
φ(tx) = Eφ(tx), ∀|x| = 1.
dt t
Hence
d a+n
 
t ~
φ(tx) = ta+n−1 Eφ(tx) + (a + n)φ(tx) .
dt
Thus (11.2.18) is equivalent to

d a+n

t φ(tx) = ta+n−1 η(tx), ∀|x| = 1.
dt
The function φ defined by
Z ∞
1
φ(tx) = − τ a+n−1 η(τ x)dτ
ta+n t

will do the trick. ⊔

⊓
11.3. THE WAVE EQUATION 555

11.3 The wave equation

The wave operator of a compact Riemann manifold (M, g) is the p.d.o.

M : C ∞ (R × M ) → C ∞ (R × M ), M u = ∂t2 u + ∆M u.

In this section we consider initial value problems of the type

M u = 0, u(0, −) = 0, u′t (0, −) = ϕ ∈ C ∞ (M )

As we will see later in this section, this initial value problem has a unique solution which
we denote by Ut (ϕ).
Using a technique pioneered by J. Hadamard [57] and clarified by M. Riesz [114], we
will produce asymptotic expansions Ut (ϕ) for |t| small. Our approach follows Hörmander’s
modern presentation in [67] of the ideas of Hadamard and Riesz.2 As Hörmander explained
in his groundbreaking work [65], these short term asymptotics can be used to obtain rather
precise information about the spectrum of ∆M .

11.3.1 Fundamental solutions of the wave operator

Let x1 , . . . , xm denote the Euclidean coordinates in Rm . We set
m
X
 := Rm = ∂t2 − ∂x2k .
k=1

A fundamental solution of  is a generalized function E ∈ C −∞ (Rm+1 ) such that

E = δ0 .

A Cauchy fundamental solution for  is a C 2 -map

(0, ∞) ∋ t 7→ Et ∈ C −∞ (Rm ),

such that
m
X
∂t2 Et − ∂x2j Et = 0, t > 0, (11.3.1a)
j=1

lim Et = 0, (11.3.1b)
tց0

lim ∂t Et = δ0 ∈ C −∞ (Rm ). (11.3.1c)

tց0

To a Cauchy fundamental solution (Et )t>0 we associate a generalized function Ee ∈

−∞
C (R m+1 ) defined by
Z ∞
e
hE, ϕi = Et , ϕ(t, −) dt, ∀ϕ ∈ D(Rm+1 ). (11.3.2)
0
2
Hörmander’s initial doctoral adviser was M. Riesz who retired. Hörmander then completed his disser-
tation under the direction of L. Garding.
556 CHAPTER 11. SPECTRAL GEOMETRY

The above integral is well defined since the function

[0, ∞) ∋ t 7→ Et , ϕ(t, −) ∈ R

is continuous and compactly supported. The Uniform Boundedness Principle, Theorem

e defined by (11.3.2) is indeed a generalized function. Note that
11.1.5, implies that E
e ⊂ Rm+1 .
supp E t≥0

Proposition 11.3.1. If (Et )t>0 is a Cauchy fundamental solution for the operator L,
e is a fundamental solution of L. We say that E
then the associated generalized function E e
is the fundamental solution associated to the Cauchy fundamental solution (Et )t≥0 .

Proof. For ϕ ∈ D(Rm+1 )

Z ∞
e ϕ = E,
E, e ϕ = lim Et , (ϕ)(t, −) dt
εց0 ε
Z ∞ Z ∞
= lim Et , (∆ϕ)(t, −) dt + lim Et , (∂t2 ϕ)(t, −) dt
εց0 ε εց0 ε
Z ∞ Z ∞
= lim ∆Et , ϕ(t, −) dt − lim Eε , ∂t ϕ(ε, 0) − lim ∂t Et , (∂t ϕ)(t, −) dt
εց0 ε εց0 εց0 ε
Z ∞ Z ∞
(11.3.1b )
= lim ∆Et , ϕ(t, −) dt + lim ∂t Et , ϕ(t, −) + lim ∂t2 Et , ϕ(t, −) dt
εց0 ε εց0 εց0 ε
Z ∞
(11.3.1c ) (11.3.1a )
= lim ∂t2 Et + ∆Et , ϕ(t, −) dt + ϕ(0, 0) = ϕ(0, 0) = δRm+1 , ϕ .
εց0 ε
⊔
⊓

Suppose that n = m + 1, and we denote the canonical Euclidean coordinates in Rm+1

by (t, x1 , . . . , xm ). We consider the quadratic from

q : R1+m → R, q(t, x) = t2 − |x|2 , x ∈ Rm , (11.3.3)

we denote by C+ the forward light cone

 
C+ := (t, x); q(t, x) ≥ 0, t ≥ 0 = (t, x); t ≥ |x| ,

and by C− the backward light cone

 
C− := (t, x); q(t, x) ≥ 0, t ≤ 0 = (t, x); −t ≥ |x| .

The map
C+ × C+ ∋ (x, y) → x + y ∈ Rn
is obviously proper, so the cone C+ satisfies the properness condition (11.1.4). We have
the following important uniqueness result.
Proposition 11.3.2. The wave operator  admits at most one fundamental solution
supported in the forward light cone C+ .
11.3. THE WAVE EQUATION 557

Proof. Suppose that E, E ′ are two fundamental solutions with support in C+ . Then
E ′ = E + u, where u is a generalized function with support in C+ such that
u = (E ′ − E) = 0.
Since the cone C+ satisfies the properness condition (11.1.4) we deduce from Remark
11.1.15 that the convolution E ∗ u is well defined and we have
0 = (E ∗ u) = (E ∗ u) = (E) ∗ u = δ ∗ u = u.
⊔
⊓

11.3.2 The wave family

We denote by Lom the Lie group of special Lorentz rotations consisting of linear operators
A : R1+m → R1+m
such that
A∗ q = q, AC+ = C+ , det A = 1.
Note that  is Lom -invariant. In this section we plan to solve a more general problem
namely, we want to describe all the homogeneous, Lom -invariant generalized functions on
R1+m that are supported on the cone C+ .
Let us start with a simpler question. Suppose that f : Rm+1 → R is a continuous
Lom -invariant function on Rm+1 . Observe that the orbits of Lom consist of the level sets
of q. Hence there exists a continuous f¯ : R → R such that


f (t, x) = f¯ q(t, x) .
Moreover, if f is supported on C+ , then f¯ is supported on [0, ∞). Finally, if f is homo-
geneous of degree d > 0, then Theorem 11.2.6 shows that f must have the form
d
f = const.q ∗ χ+2 .
To deal with the more general case when f is a generalized function we follow the strategy
in [92].
Set n := 1 + m. For any open set O ⊂ R1,m we have a cap product
∩ : Ωj (O) × Ωk (O) → Ωk−j (O),
k−j
hβ, α ∩ T i = hβ ∧ α, T i, ∀T ∈ Ωk (O), α ∈ Ωj (O), β ∈ Ωcpt (O).
Observe that if O is a Lom -invariant open subset of R1,m , then a generalized function
u ∈ C −∞ (O) is Lom -invariant if and only if the associated n-dimensional current is such.
Proposition 11.3.3. Suppose that O ⊂ R1,m is Lom -invariant.
(a) If u ∈ C −∞ (O) is Lom -invariant, then
du ∧ dq = 0. (11.3.4)
(b) If T ∈ Ωn (O) is Lom -invariant, then
dq ∩ ∂T = 0, (11.3.5)
where ∂ : Ωn (O) → Ωn−1 (O) is the boundary operator, dual to the exterior derivative.
558 CHAPTER 11. SPECTRAL GEOMETRY

Proof. (a) The Lie algebra of Lom can be tautologically identified with a subspace of the
space of linear vector fields on R1,m . A basis is given by the vector fields

Xi = xi ∂t + t∂xi , Xij = xi ∂xj − xj ∂xi , 1 ≤ i, j ≤ m.

If u is Lom -invariant, then

Xi · u = Xij · u = 0, ∀1 ≤ i, j ≤ m.

These equalities can be rewritten in the compact form du ∧ dq = 0. Part (b) follows easily
from (a). ⊔
⊓

Denote by O+ the open set


O+ := R1,m \ C− = (t, x) ∈ Rn+1 ; t > −|x| ,

and by q + the restriction of q to O+ . The set O+ is Lom -invariant. We have a pushforward

map
q∗+ : Ωk (O+ ) → Ωk (R), k ≤ 1.
On the other hand, the map q + : O+ → R is a submersion. Thus there is a push-forward
map
q!+ : Ωncpt (O+ ) → Ω1cpt (R)
given by a formula very similar, in fact equivalent, to (11.2.12). More precisely
Z
η
q!+ (η) = η̄q dq, η¯q := ,
+
q =q dq
η
where dq is the so-called Gelfand–Leray residue of η along the level set {q + = q}. This
differential form is represented by the restriction to this level set of any (n − 1)-form ω
defined in a neighborhood U of this fiber of q and satisfying the equality

η = dq ∧ ω

The above equation has many solutions ω, but all their restrictions to the fiber are the
same.
In particular, we have a dual pullback map

(q + )! : Ω1 (R) → Ωn (O+ ).

The image of (q + )! consists of Lom -invariant currents. We want to prove that, in fact,
any invariant current on O+ lies in the image of (q + )! .
Let T+ = {t > 0} ⊂ R1,m . On this open set the functions (q, x1 , . . . , xm ) define a
global coordinate system. More precisely, T+ can be identified with the open set in the
(q, x1 , . . . , xm )-space defined by the inequality

|x|2 + q > 0.
11.3. THE WAVE EQUATION 559

Fix a function α ∈ C ∞ (R1,m ), such that supp α ⊂ T+ and for any λ ∈ R we have

supp α ∩ {q = λ} is compact,
Z
α(λ, x1 , . . . , xm )dx1 ∧ · · · ∧ dxm = 1.
Rm
Now set
γ = αdx1 ∧ · · · ∧ dxm ∈ Ωn−1 (T+ ).
We have a projection formula


S = q∗ (q + )! γ ∩ S , ∀S ∈ Ω1 (R) (11.3.6)

whose proof is an immediate consequence of the definitions of the various intervening

operations. We have the following result, [92, p. 233-234].
Proposition 11.3.4. If T ∈ Ωn (O+ ) is Lom -invariant, then there exists S ∈ Ω1 (R) such
that
T = (q + )! S. (11.3.7)

Proof. Observe that if (11.3.7) were true, than the current S is uniquely determined by T
via the projection formula
S = q∗+ (γ ∩ T ).
We will prove that


T = (q + )! q∗+ (γ ∩ T ) on T+ (11.3.8)
The equality (11.3.7) follows from (11.3.8) by invoking the invariance of T and the fact
that the collection A · T+ , A ∈ Lom is an open cover of O+ .
We set  
S := q∗+ (γ ∩ T ), T0 := T − (q + )! S |T+ .

For any η ∈ Ωncpt (T+ ) we have

hη, T0 i = hη, T i − q!+ η T = hη, T i − (q + )∗ (q! η), γ ∩ T

= η − (q + )∗ (q! η) ∧ γ , T . (11.3.9)
| {z }
=:β

We can write
β = b(q, x1 , . . . , xm )dq ∧ dx1 ∧ · · · ∧ dxm ∧ dq.
From the equality


0 = q!+ η − (q + )∗ (q!+ η) ∧ γ
we deduce Z
b(λ, x1 , . . . , xm )dx1 ∧ · · · ∧ dxm = 0, ∀λ ∈ R.
560 CHAPTER 11. SPECTRAL GEOMETRY

For any fixed λ the function

(x1 , . . . , xm ) 7→ b(λ, x1 , . . . , xm )

has compact support. Using the Poincaré lemma for compact supports Theorem 7.2.1, or
rather its proof, we can find a family of forms Bq ∈ Ωm−1 m
cpt (R ), depending smoothly on
q ∈ R, such that, ∀q ∈ R we have

supp Bq ⊂ |x|2 + q > 0 ,
b(q, x)dx1 ∧ · · · ∧ dxm = dx Bq ,

where dx denotes the exterior derivative on the space Rm with coordinates x = (x1 , . . . , xm ).
We can regard the family (Bq ) as defining an (m − 1)-form B on the space T+ with coor-
dinates (q, x) and as such we have

β = d(B ∧ dq)

and
(11.3.9 ) (11.3.5 )
hη, T0 i = hd(B ∧ dq), T i = hB ∧ dq, ∂T i = hB, dq ∩ ∂T i = 0.
⊔
⊓

Similarly, if we set O− = R1,m \ C+ and we denote by q − the restriction of q to O− ,

then we deduce that T ∈ Ωn (O− ) is Lom -invariant if and only if there exists S ∈ Ω1 (R)
such that
T = (q − )! S.
Corollary 11.3.5. Let T ∈ Ωn (R1,m \ 0). Then T is Lom -invariant if and only if there
exist S± ∈ Ω1 (R) such that
S+ |R<0 = S− |R<0 (11.3.10)
and
T |O± = (q ± )! S± .
We denote by u(S+ , S− ) ∈ C −∞ (R1,m \ 0) the invariant generalized function associated to
the currents S± satisfying the compatibility condition (11.3.10). ⊔
⊓

Suppose that u ∈ C −∞ (R1,m \ 0) is both Lom -invariant and supported in C+ \ 0. The

above corollary shows there exists a generalized function v ∈ C −∞ (R) such that

v|R<0 = 0, u|O+ = q ∗ v, u|O− = 0. (11.3.11)

✍ In the sequel we will use the notation u = (q + )∗ v when referring to the generalized
function u on the punctured space R1,m \ 0 defined by (11.3.11), where v ∈ C −∞ (R),
supp v ⊂ R≥0 . Informally
(q + )∗ v = v(t2+ − |x|2 ).

A simple computation shows that

 
(q + )∗ v = (q + )∗ Pn v ,
11.3. THE WAVE EQUATION 561

where Pn : C −∞ (R) → C −∞ (R) is the differential operator

d2 d
Pn = 4q + 2n .
dq 2 dq

The above discussion leads to the following conclusion.

Corollary 11.3.6. Suppose that u ∈ C −∞ (R1,m \ 0) is Lom -invariant, homogeneous of

degree d ∈ C \ 2Z<0 and supp u ⊂ C+ . Then
d d

u = const.(q + )∗ χ+2 = constχ+2 t2+ − |x|2 ,

and
d d−2
(q + )∗ χ+2 = (2d + 2m − 2)(q + )∗ χ+2 , on R1,m \ 0. (11.3.12)

Proof. We know that u = (q + )∗ v for some v ∈ C −∞ (R) with support in [0, ∞). Since
u is homogeneous of degree d and q is homogeneous of degree 2 we deduce that v is
homogeneous of degree d2 ∈ R \Z<0 . Theorem 11.2.7 implies that v is uniquely determined
by its restriction to (0, ∞), while Theorem 11.2.6 shows that the restriction of v to (0, ∞)
d
is equal to constχ+2 .
To prove the second statement observe that

d d2 d−2 d d−2 d − 2 d−2

χ+ = χ+2 , q χ+2 = χ+2 ,
dq dq 2
d−2
d−2
since χ+2 is homogeneous of degree 2 . Hence

d d−2 d−2
Pn χ+2 = (2d + 2n − 4)χ+2 = (2d + 2m − 2)χ+2 .

⊔
⊓

For Re a > 0 we consider the measurable Lom -invariant function

( a−n
a q(t, x) 2 , (t, x) ∈ int C+
R+ := cm (a) × ,
0, (t, x) ∈ R1,m \ C+

where
Γ( a+1
2 )
n := m + 1, cm (a) := m .
π Γ(a)Γ( a−m+1
2
2 )
a is homogeneous of degree a − n.
Observe that R+
a is locally integrable if a > m − 1.
Lemma 11.3.7. Let m > 1. The function R+
562 CHAPTER 11. SPECTRAL GEOMETRY

Proof. As we have seen in the proof of Proposition 11.3.4 on the open set O+ we can use
as coordinates the functions q, x1 , . . . , xm . In these coordinates we have
p
t = |x|2 + q
so that
m
X xi 1
dt = p dxi + p dq
i=1 |x|2 +q 2 |x|2 + q
and
1
dt ∧ dx1 ∧ · · · ∧ dxm = p dq ∧ dx1 ∧ · · · ∧ dxm .
2 |x|2 + q
In the coordinates (q, x) the cone C+ becomes the half-plane q ≥ 0 and the function
a−n
q 2
p
|x|2 + q
is locally integrable on this half-plane because
a−n Re a−n
q 2 q 2
p ≤ , ∀q ≥ 0, x ∈ Rm \ 0,
|x|2 + q |x|
and the function is the right hand side is locally integrable if m > 1 and
Re a − n
> −1⇐⇒ Re a > m − 1.
2
⊔
⊓

☞ In the remainder of this section we will assume that m > 1.

Let Re a > m − 1. Then R+ a is locally integrable and defines a Lo -invariant gener-
m
1,m
alized function on R . We denote by R+ a this generalized function. It is homogeneous

of degree a − n > −2 and its support is contained in C+ . Corollary 11.3.6 shows that R+ a

is the unique extension to R 1,m of the homogeneous generalized function

a−n

a−n
dm (a)χ+2 t2+ − |x|2 = cm (a) t2+ − |x|2 +2 ∈ C −∞ (R1,m \ 0),
where n = m + 1,  
a−n Γ( a+1 )
dm (a) := Γ + 1 cm (a) = m 2 .
2 π 2 Γ(a)
In particular
1
dm (2) =
m−1 .
2π 2
Observe that for any a, b > m−1 the convolution R+ a ∗Rb of the locally integrable functions
∗
R+a , Rb is well defined. It is a locally integrable, Lo -invariant function, homogeneous of
+ m
degree (a + b) − n with support in C+ + C+ = C+ . We deduce from Corollary 11.3.6 that
there exists a constant c(a, b) such that
a b a+b
R+ ∗ R+ = c(a, b)R+ .
We can be more precise.
11.3. THE WAVE EQUATION 563

Proposition 11.3.8.
a b a+b
R+ ∗ R+ = R+ , ∀a, b > m − 1.

Proof. We follow the approach in [39, p. 160]. Observe that

Z Z
−t a b a+b
e R+ ∗ R+ (t, x)dtdx = R+ (t, x)e−t dtdx.
R1,m R1,m

If we set Z
T (a) := e−t R+
a
(t, x)dtdx,
R1,m

then the identity et = et−s es implies that

Z
e−t R+
a b
∗ R+ (t, x)dtdx = T (a)T (b)
R1,m

so that
T (a)T (b)
c(a, b) = .
T (a + b)
Now observe that
Z Z !
T (a) a−m−1
= e−t (t2 − |x|2 ) 2 dx dt
cm (a) R>0 |x|≤t

Z Z t 
a−m−1
−t 2 2 m−1
= σ m−1 e (t − r ) 2 r dr dt,
R>0 0

where σ m denotes the area of the unit sphere in S m ⊂ Rm+1 ,

m+1
2π 2
σm = .
Γ( m+1
2 )

Now observe that

Z t Z
2 2 a−m−1
m−1 ta−1 1 a−m−1 m−2
(t − r ) 2 r dr = (1 − s) 2 s 2 ds
0 2 0
 
ta−1 m a−m+1
= B , .
2 2 2
Hence  Z ∞
T (a) σ m−1 m a−m+1
= B , e−t ta−1 dt
cm (a) 2 2 2 0

σ m−1 Γ(a)Γ( m
2 )Γ(
a−m+1
2 ) 1
= a+1 = .
2 Γ( 2 ) cm (a)
Thence T (a) = 1 and c(a, b) = 1. ⊔
⊓
564 CHAPTER 11. SPECTRAL GEOMETRY

a , Re a > m − 1 satisfy the

The above result show that the generalized functions R+
convolution equalities
a b a+b
R+ ∗ R+ = R+ , ∀ Re a, Re b > m − 1.

Using the equality (11.3.12) and the definition of dm (a) we deduce that
a a−2
R+ = R+ on R1,m \ 0, ∀ Re a > m + 1.

Invoking (11.2.16) we deduce that the above equality holds on R1,m . We have thus proved
the following result.
Proposition 11.3.9.
a b a+b
R+ ∗ R+ = R+ , ∀ Re a, Re b > m − 1, (11.3.13a)
a a−2
R+ = R+ on R1,m , ∀ Re a > m + 1. (11.3.13b)

We can now use the above equalities to define R+ a ∈ C −∞ (R1,m ) for any a ∈ C by

setting
a
R+ := k R+a+2k
, (11.3.14)
where k is a positive integer such that Re a + 2k > m − 1. The family
a
C ∋ a 7→ R+ ∈ C −∞ (R1,m )
a is homogeneous of
is holomorphic. Moreover, for any a ∈ C the generalized function R+
degree a, it is Lom -invariant and
a
supp R+ ⊂ C+ .

Moreover, the equalities (11.3.13a) and (11.3.13b) extend by unique continuation to all
a, b ∈ C. The family (R+a) a
a∈ C is called the wave family and the generalized functions R+
are the Riesz generalized functions, [114].
Proposition 11.3.10.
a a−2
(a − 2)∂xi R+ = −xi R+ , ∀a ∈ C, 1 ≤ i ≤ m. (11.3.15)

Proof. To prove (11.3.15) we observe that both sides of this equality are generalized func-
tions depending holomorphically on a. Hence it suffices to prove it only for Re a ≫ 0. In
this case a−n a−2−n
a
∂xi R+ = dm (a)∂xi (q + )∗ χ+2 = dm (a)(∂xi q)(q + )∗ χ 2
dm (a) a−2−n dm (a)
= −2xi dm (a − 2)(q + )∗ χ 2 = −2xi R a−2
dm (a − 2) dm (a − 2) +
Γ( a+1
2 )Γ(a − 2) a−2
a−1
a−2 xi
= −2xi R+ = −2x i
2
R+ =− R a−2 .
a−1
Γ(a)Γ( 2 ) (a − 1)(a − 2) a−2 +
⊔
⊓
11.3. THE WAVE EQUATION 565

Theorem 11.3.11. The unique fundamental solution of  on R1,m , m > 1, with support
2.
in C+ is the Riesz generalized function R+

Proof. Observe first that the function

Γ( a+1
2 )
a 7→ dm (a) = m , Re a > m − 1
π 2 Γ(a)

extends to a holomorphic map on C.

1
Indeed, the function a 7→ Γ(a) defines a holomorphic function on C with zero set Z≤0 .
1
Each of these zeroes is a simple zero. Moreover, the derivative of Γ(a) at a = −k ∈ Z≤0 is
(−1)k k!.
The function a 7→ Γ( a+1
2 ) admits a meromorphic extension to C. All its poles are
simple and are located at −1 + 2Z≤0 . Thus the function dm (a) extends to a holomorphic
function on C. It has only simple zeros situated at 2Z≤0 . We obtain in this fashion a
holomorphic family of generalized functions
a−n
a
C ∋ a 7→ R̄+ := dm (a)(q + )∗ χ+2 ∈ C −∞ (R1,m \ 0).

Note that
a a
R̄+ = R+ |R1,m \0 , ∀ Re a > m − 1.
By unique continuation we deduce that the above equality holds for all a ∈ C. Using
(11.3.13b) we deduce that
a a−2
R̄+ = R̄+ on R1,m \ 0, ∀a ∈ C.

Since dm (a) = 0 for a ∈ 2Z≤0 , we deduce

0
R̄+ = 0 on R1,m \ 0.

Hence
0
supp R+ = {0}.
0 is homogeneous of degree −n we deduce that R 0 = cδ , for some constant c.
Since R+ + 0
Then
a a 0 a a
R+ = R+ ∗ R+ = cR+ ∗ δ0 = cR+ .
This shows that c = 1. ⊔
⊓

2 . Using Proposition 11.1.28 with f = 0 we deduce that there exists a

We set E := R+
C 2 -path
(0, ∞) ∋ t 7→ E(t) ∈ C −∞ (Rm )
such that the limits
E0 := lim E(t), E0′ := lim ∂t E(t)
tց0 tց0

exist in C −∞ (Rm ).
566 CHAPTER 11. SPECTRAL GEOMETRY

Proposition 11.3.12. The family ( E(t) )t>0 is a Cauchy fundamental solution of the
wave operator, i.e., it satisfies the conditions (11.3.1a), (11.3.1b) and (11.3.1c).

Proof. We follow the approach in [66, Sec. 6.2].The equality (11.3.1a) follows from the
fact that E = δ0 . We have to show that E0 = 0 and E0′ = δ0 . To achive this we first
need to have a clearer idea on the nature of the generalized functions E(t).
Let us compute the restriction of
1 2−n 1 1−m

(m−1)/2
q ∗ χ+2 = (m−1)/2
q ∗ χ+ 2
2π 2π
to the punctured space R1+m \ 0. Note that
p
|∇q(t, x)| = 2 t2 + |x|2 .

For t > 0 we have

 p
q −1 (s) ∩ C+ \ 0 = (t, x); t = s + |x|2 , s + |x| > 0 , s ≥ 0.

Thus q −1 (s) is the graph of the function

p
us : Rm → [0, ∞), us (x) = s + |x|2 .

The volume element on this graph is

s s
p |x|2 s + 2|x|2
|dVq−1 (s) | = 1 + |∇us |2 |dx| = 1+ |dx| = .
s + |x|2 s + |x|2
p
Note that along q −1 (s) we have |∇q| = 2 s + 2|x|2 so
1 1
|dVq−1 (s) | = p |dx|.
|∇q| 2 s + |x|2

If ϕ ∈ D(R1+m \ 0), then

Z
1

q∗ ϕ(s) = ϕ (s + |x|2 )1/2 , x (s + |x|2 )−1/2 |dx|.
2 Rm

Using spherical coordinates (r, ω) we have

p p rdr rdr
x = rω t = s + r2, r = t2 − s, .dt = 1/2
=
t (s + r 2 )1/2
We set Z
ϕ̃(t, r) := r m−2 ϕ(t, rω)|dω|
|ω|=1

and we get
Z ∞ Z
1 rdr 1

q∗ ϕ(s) = ϕ̃(t, r) 2 1/2
= ϕ̃ t, (t2 − s)1/2 dt.
2 0 (s + r ) 2 t2 >s
11.3. THE WAVE EQUATION 567

Hence
Z
1 (1−m)/2

E+ , ϕ = χ+ , ϕ̃ t, (t2 − ·)1/2 dt, ∀ϕ ∈ D(R1+m \ 0). (11.3.16)
4π (m−1)/2 s<t2
m
For any ψ ∈ D R we set
Z
m−2
ψ̃(r) := r ψ(rω)|dω|.
|ω|=1

The equality (11.3.16) shows that for t > 0 and ψ ∈ D(Rm ) we have
1 D 1−m
E
E+ (t), ψ = (m−1)/2 χ+ 2 (s), ψ̃ (t2 − s)1/2 . (11.3.17)
4π
We rewrite (11.3.16) as
Z ∞
E+ , ϕ = E+ (t), ϕ(t, x) dt, ∀ϕ ∈ D(R1+m \ 0).
0
We claim that Z ∞
E+ , ϕ = E+ (t), ϕ(t, x) dt, ∀ϕ ∈ D(R1+m ). (11.3.18)
0
To see this observe that since limtց0 E(t) exists weakly we deduce from the uniform
boundedness principle that the right-hand side of (11.3.18) exists and defines a homoge-
neous distribution of degree 1 − n on Rn = R1+m . This coincides with the homogeneous
distribution E+ on the punctured space Rn \ 0. We conclude from Proposition that they
coincide on Rn .
To show that E0 = 0 and E0′ = δ0 we argue as in the proof of Proposition 11.3.1. Let
ϕ = ϕ(t, x) ∈ D(R1+m ). We have
ϕ(0, 0) = hδ0,0 , ϕi = E+ , ϕ = E+ , ϕ
Z ∞
(11.3.18 )
= lim E+ (t), (ϕ)(t, −) dt
εց0 ε
Z ∞ Z ∞
= lim E+ (t), (∆ϕ)(t, −) dt + lim E+ (t), (∂t2 ϕ)(t, −) dt
εց0 ε εց0 ε
Z ∞ Z ∞
= lim ∆Et , ϕ(t, −) dt − lim E+ (ε), ∂t ϕ(ε, −) − lim ∂t Et , (∂t ϕ)(t, −) dt
εց0 ε εց0 εց0 ε
Z ∞
= − E0 , ∂t ϕ(0, −) + lim ∆Et , ϕ(t, −) dt
εց0 ε
Z ∞
+ lim ∂t Eε , ϕ(ε, −) + lim ∂t2 Et , ϕ(t, −) dt
εց0 εց0 ε
Z ∞
′
= − E0 , ∂t ϕ(0, −) + E0 , ϕ(0, −) + lim ∂t2 Et + ∆Et , ϕ(t, −) dt
εց0 ε

= − E0 , ∂t ϕ(0, −) + E0′ , ϕ(0, −) .

Since the map


D(R1+m ) ∋ ϕ(t, x) 7→ ϕ(0, x), ϕ′t (0, x) D(Rm ) × D(Rm )
is onto, we deduce E0 = 0 and E0 = δ0 . ⊔
⊓
568 CHAPTER 11. SPECTRAL GEOMETRY

2 is a Cauchy fundamental solution and so

Remark 11.3.13. The generalized function R+
that it is determined by a smooth family
2
(0, ∞) ∋ t 7→ R+ (t) ∈ C −∞ (Rm )

satisfying

(∂t2 + ∆)R+
2 2 +
(t) = 0, t > 0, R+ 2 +
(0 ) = 0, ∂t R+ (0 ) = δ0 . (11.3.19)

Thus R+ 2 describes the propagation of a wave generated by a disturbance at t = 0 localized

near (or at) the origin. If x ∈ Rm \ 0, then R+2 (t, x) = 0 for |x| > t. In other words, the

disturbance at 0 takes t = |x| seconds to reach the point x.

If m is odd, m = 2k + 1, k > 0, then
2−n
(k−1)
χ+2 = χ−k
+ = δ0

so that
2
R+ = dm (2)χ−k 2 2
+ (t+ − |x| )

has support on ∂C+ . In this case disturbance leaves the point x immediately after it
reaches this point. This is sometimes referred to as Huygens’ Principle.
If m is even, the disturbance will linger at x forever, with diminishing strength. ⊔
⊓
2.
Here is a simple but important application of the fundamental solution R+
Theorem 11.3.14. For arbitrary ϕ0 , ϕ1 ∈ Rm and f ∈ C ∞ ([0, ∞) × Rm ) the Cauchy
problem

u = f, on (0, ∞) × Rm ,
(11.3.20)
u(0, x) = ϕ0 (x), ∂t u(0, x) = ϕ1 (x), ∀x ∈ Rm ,

has a unique solution u ∈ C ∞ ([0, ∞) × Rm ) given by

Z t
2 d 2
2
u(t, −) = R+ (t) ∗ ϕ1 + R (t) ∗ ϕ1 + R+ (t − s) ∗ f (s)ds. (11.3.21)
dt + 0

Proof. Existence. From Proposition 11.3.12 we deduce that R+ 2 is a Cauchy fundamental

solution for the wave operator. Using Proposition 11.1.28 we deduce that the correspon-
dence
2
(0, ∞) ∋ t 7→ R+ (t) ∈ C −∞ (Rm )
is smooth. Moreover the support of R+ 2 (t) is contained in the ball {|x| ≤ t}. Thus,

the expression in the right-hand side of (11.3.21) is a well defined smooth function on
[0, ∞) × Rm . The fact that u given by (11.3.21) is a solution of the initial value problem
(11.3.20) follows by direct computation using the equalities

2 d 2
R+ = 0 in (0, ∞) × Rm , R+
2
(0+) = 0, R (0+) = δ0 .
dt +
11.3. THE WAVE EQUATION 569

Uniqueness. There are many approaches. Ours is based on finite speed propagation of
solutions. We follow closely the presentation in [32, VI.6.1].
For every T > 0 we consider the dependence cone (see Figure 11.1)
T

C− := (t, x); t ∈ [0, T ], |x| ≤ (T − t) .

C_T

x2

x1

T.
Figure 11.1: The dependence cone C−

T the base of
We denote by ∂+ C∗T conical part of the boundary of the cone and by ∂0 C−
the cone; see Figure 11.1. The unit outer normal along the (smooth part of the) boundary
of C−T has the form

Xn
n = a0 ∂t + ai ∂xi .
i=1

Along the conical part of the boundary, described by the quadratic equation

(T − t)2 = |x|2 ,

the components of the outer normal satisfy the conditions

m
X
a20 = a2i , a0 > 0. (11.3.22)
i=1

Along the base of the cone, described by {t = 0, |x| ≤ T }, the unit other normal is
n0 = −∂t . For any smooth function u = u(t, x) we set
m
X m
X
ˆ := ut dt +
du ˆ 2 = u2 +
uxi dxi , |du| u2xi .
t
i=1 i=1

We have
  X
m
ˆ 2 +
2ut u = ∂t |du| ∂xi (−2ut uxi ).
i=1
570 CHAPTER 11. SPECTRAL GEOMETRY

In other words, ut u is the divergence of the vector field on R1+m ,

  m
X
Xu = ˆ 2 ∂t − 2
|du| (ut uxi )∂xi .
i=1

Hence Z Z Z
(ut u)dtdx = div X u dtdx = (X u , n) dS
T
C− T
C− T
∂C−

Z Z m m
!
X X
=− ˆ
|du(0, x)|2 dx + a0 u2xi − 2 ai ut uxi dS
|x|≤T T
∂∗ C− i=0 i=1
Z Z m
!
(11.3.22 ) 1 X
= − ˆ
|du(0, x)|2 dx + (a0 uxi − ai ut )2 dS.
|x|≤T T
∂∗ C− a0
i=1

We deduce that if u = 0, then

Z Z m
!
1 X
ˆ
|du(0, x)|2 dx = (a0 uxi − ai ut )2 dS. (11.3.23)
|x|≤T T
∂∗ C− a0
i=1

If additionally
ˆ
du(0, x) = 0,
T we have
then we deduce that along ∂∗ C−

a0 uxi − ai ut = 0, ∀i = 1, . . . , m. (11.3.24)

The vector fields

Vi = −ai ∂t + a0 ∂xi , i = 1, . . . m
T and they are linearly independent there. In other words, they define
are tangent to ∂∗ C−
a framing of the tangent bundle T ∂+ C−T . We can rewrite the equalities (11.3.24) as

Vi · u = 0, ∀i = 1, . . . , m.

They imply that u is constant along ∂∗ C0T . If we assume additionally that u(0, x) = 0,
then we conclude that u ≡ 0 on ∂C− T , so that u(T, 0) = 0, ∀T > 0. Applying the same

argument using instead the translate C+ T (x ) = x + C T we obtain the desired uniqueness

0 0 +
result. ⊔
⊓

Remark 11.3.15. (a) The above proof of the uniqueness result justifies the name depen-
dence cone we gave to C−T . If u, v are smooth solutions of the wave equation satisfying

the equalities
u(0, x) − v(0, x) = ut (0, x) − vt (0, x) = 0, ∀|x| ≤ T,
then u(0, T ) = v(0, T ). In other words, the value at (T, 0) is only determined by the initial
data inside the dependence cone.
11.3. THE WAVE EQUATION 571

(b) There is an alternate proof of the uniqueness, based on the so called energy method
and in which the dependence cone also play a central role. Suppose the u is a C 2 solution
of the wave equation. For t ≤ T we define
Z
1 ˆ
Eu (t) = |du(t, x)|2 dx.
2 |x|≤T −t

An elementary computation (see [44, §2.4.3]) shows that

0 ≤ Eu (t) ≤ Eu (0), ∀t ∈ (0, T ).

If u(0, x) = ut (0, x) = 0, then Eu (0) = 0 so that Eu (t) = 0, and

ˆ
du(t, x) = 0, ∀|x| ≤ T − t. ⊔
⊓

Remark 11.3.16. The space of solutions of harmonic oscillator equation ẍ + ω 2 x = 0 is

spanned by two solutions
sin(ωt)
a(t) = cos(ωt), b(t) = .
ω
Note that a(t) satisfies the initial conditions a(0) = 1, dota(0) = 0, b(0) = 0, ḃ(0) = 1.
We can think of the wave equation ∂t2 u + ∆u = 0 as an operatorial harmonic oscillator
equation √
2
ü + ∆ u = 0.
(Recall that our Laplacian is the geometers’ Laplacian so it is nonnegative.)
The analogy with the harmonic oscillator becomes even more clear if we take the
Fourier transform in the space variables
Z
e(t, ξ) =
u e−ihξ,xiu(t, x)dx.
Rm

e(t, ξ) satisfies the harmonic oscillator equation

Then, for every ξ, the function u

∂t2 u
e + |ξ|2 u
e = 0.

One can show that (see [39, Ex. 18.7])

g 2 sin(|ξ|t+ )
R+ (t, ξ) = .
|ξ|
⊔
⊓

Consider the generalized functions

χa+ (t2+ − |x|2 ) ∈ C −∞ (R1,m \ 0).

Fix x0 ∈ Rm \ 0 and consider the open set

 1
U0 := (t, x) ∈ R1,m ; |x − x0 | < |x0 | ⊂ R1,m \ 0
2
572 CHAPTER 11. SPECTRAL GEOMETRY

C+

U0
x0 x

Lx
0

Figure 11.2: Behavior near a temporal line.

containing the line

Lx0 = {(t, x0 ); t ∈ R} ⊂ R1,m \ 0.
Note that n o
1
supp χa+ (t2+ − |x|2 )|U0 ⊂ X0 := t > |x0 | ∩ U0
2
In the open set X0 we can use as coordinates the variables q, x1 , . . . , xm .
The map q : X0 → R is a submersion with image an open interval I0 containing the
support [0, ∞) of χa+ . We have


χa+ t2+ − |x|2 |X0 = q ∗ (χa+ |I0 )


We want to define the restriction of χa+ t2+ − |x|2 to the line Lx0 .
p
On X0 we have t+ = q + |x|2 . For Re a ≫ 0 and ϕ ∈ D(U0 ) we have
Z p

1 a ϕ( q + |x|2 , x)
χa+ t2+ 2
− |x| |X0 , ϕ(t, x) = q+ p dqdx
Γ(a + 1) X0 2 q + |x|2
Z Z  p
ϕ( q + |x|2 , x)
= χa+ (q), ψx (q)dq dx, ψx (q) = p ∈ D(R).
|x−x0 |< 21 |x0 | R 2 q + |x|2

This shows that for Re a ≫ 0 the restriction of χa+ (t2 − |x|2 ) to Lx0 is given by
D E D 1 p E
χa+ (t2 − |x|2 )|Lx0 , ϕ = χa+ (q), p ϕ( q + |x0 |2 ) , (11.3.25)
2 q + |x0 |2

∀ϕ ∈ D(R). The right-hand side of the above equality depends holomorphically on a and
we define χa+ (t2+ − |x|2 ) to the line Lx0 for any a ∈ C by the same equality (11.3.25) using
unique continuation.
11.3. THE WAVE EQUATION 573

Remark 11.3.17. This is admittedly a rather ad-hoc definition of the restriction of a

generalized function to a smooth submanifold. There is a more conceptual definition
of the restriction that works for a large class of generalized functions satisfying certain
microlocal conditions; see [66]. Our ad-hoc definition agrees with the more conceptual
one, but we will not need this fact in the sequel. ⊔
⊓

Proposition 11.3.18. Let n := m + 1. Then

 a− n a− n

lim ∂t χ+ 2 (t2+ − |x|2 ) − χ+ 2 (t2− − |x|2 ) = 2|χ|2a−2−m , ∀a ∈ C. (11.3.26)
x→0

where |χ|a ∈ C −∞ (R) is defined by (11.2.7)

1
|χ|a (t) := a
a+1 |t| .
Γ( 2 )

a− n
Proof. For x0 6= 0 fixed we set c = c(x0 ) = |x0 | and we denote by χ+ 2
(t2+ − c2 ) the
a− n
restriction of χ+ 2 (t2+ − |x|2 ) to the temporal line

Lx0 = {(t, x0 ) ∈ R1,m ; t ∈ R .

For Re a ≫ 0 we have
a− n a−1− n
∂t χ+ 2
(t2+ − c2 ) = 2t+ χ+ 2
(t2+ − c2 ) ∈ C −∞ (R),

 a− n a− n

∂t χ+ 2 (t2+ − c2 ) − χ+ 2 (t2− − c2 )
a−1− n a−1− n
= 2t+ χ+ 2
(t2+ − c2 ) + 2t− χ+ 2
(t2− − c2 )
a−1− n2
= 2|t|χ+ (t2 − c2 ).

By unique continuation we conclude that the above equality holds for any a, where the
multiplications |t| · χa+ (t2 − c2 ) are well defined since the singular supports of |t| and
χa+ (t2 − c2 ) are disjoint. We deduce that for any ϕ ∈ D(R) we have
D a−1− n
E D a−1− n
E
|t|χ+ 2
(t2 − c2 ), ϕ(t) = t+ χ+ 2
(t2 − c2 ), ϕ(t) + ϕ(−t)
p
(q = t2+ − c2 , t+ = q + c2 )
* +
(11.3.25 ) p a−1− n 1 p p

= ( q + c2 )χ+ 2
(q), p ϕ( q + c2 ) + ϕ(− q + c2 )
2 q + c2
 
a−1− n 1 p 2

p
2


= χ+ 2
(q), ϕ q+c +ϕ − q+c .
2
574 CHAPTER 11. SPECTRAL GEOMETRY

Observe that 12 (ϕ(t) + ϕ(−t)) is even so that according to Whitney3

1
(ϕ(t) + ϕ(−t)) = η(t2 ), (11.3.27)
2
for some η ∈ D(R). Hence

1 p
p

ϕ q + c2 + ϕ − q + c2 = η(q + c2 ),
2
D a− n a− n
E D a−1− n E
lim ∂t χ+ 2 (t2+ − c2 ) − χ+ 2 (t2− − c2 ) , ϕ(t) = χ+ 2
(q), η(q) .
c→0

We want to express hχa+ (q), η(q)i in terms of the function ϕ(t) satisfying (11.3.27). Note
that the function
a 7→ hχa+ (q), η(q)i
is holomorphic, so it suffices to determine it for a in an open set in the complex plane.
We assume Re a > 0. In this case
Z ∞
a 1
hχ+ (q), η(q)i = q a η(q)dq
Γ(a + 1) 0
Z ∞
q=s2 1
= s2a+1 (ϕ(s) + ϕ(−s))ds
Γ(a + 1) 0
Z
1
= |s|2a+1 ϕ(s)ds = |χ|2a+1 , ϕ .
Γ(a + 1) R
The map a 7→ h|χ|2a+1 , ϕ is holomorphic and we can invoke unique continuation to
conclude the validity of the obove equality for any a. The equality (11.3.26) follows from
the above since  n
2 a−1− + 1 = 2a − 1 − n = 2a − 2 − m.
2
⊔
⊓

Remark 11.3.19. The equality (11.3.26) implies that

2k 2k
∂t R+ − Ř+ )(t, 0) = 2dm (2k)|χ|2k−1−n (t), (11.3.28)

where √ Qk 2j−1
Γ( 2k+1
2 )
π j=1 2 1
dm (2k) = m = m = m−1 ,
π Γ(2k)
2 π (2k − 1)!
2 2π 2 (k − 1)!
so that
2k 1 k− n
R+ = m−1 χ+ 2
(t2+ − |x|2 ). (11.3.29)
π 2 (k − 1)!
3
We follow [56, p. 343]. Let n an tn be the Taylor expansion of ϕ0 (t) = 12 (ϕ(t)+ϕ(−t)) at t = 0. Since
P
P 2n
ϕ0 is even we deduce an = 0 for n odd so that ϕ0 (t) = a2n t . ByPBorel’s theorem, [66, Thm. 1.2.6],
there exists a smooth function u(s) whose Taylor expansion
√ at s = 0 is n a2n sn . Then v(t) = ϕ 2
√0 (t)−u(t )
vanishes with all its derivatives at t = 0 and hence v( t) is smooth. We set η(s) := u(s) + v( s).
11.3. THE WAVE EQUATION 575

We deduce that when k = 1 we have

2 2 1
∂t R+ − Ř+ )(t, 0) = 2dm (2)|χ|−m (t) = m−1 |χ|−m (t). (11.3.30)
π 2

Using (11.2.11b) we deduce that the Fourier transform of ∂t R+ 2 − Ř 2 )(t, 0) with respect
+
to the t-variable is
√
 2 2
 2 π (m−1) 2π
Ft ∂t R+ − Ř+ )(t, 0) (τ ) = m−1 |χ| (τ ) = m
m
|τ |m−1
m m
2 π 2 Γ( 2 )
2 π 2
m
2π 2 m−1 σ m−1 m−1 mω m m−1
=π m m |τ | =π m
|τ | =π |τ | .
(2π) Γ( 2 ) (2π) (2π)m
Hence !
1  2 2
 1 d ωmτ m
Ft ∂t R+ − Ř+ )(t, 0) (τ ) = sign τ . (11.3.31)
2π 2 dτ (2π)m
⊔
⊓

11.3.3 Local parametrices for the wave equation with variable coeffi-
cients
We want to extend the identities (11.3.13b) and (11.3.15) to slightly more general situa-
tions.
Consider a new linear change of coordinates on Rm , x = T y, where T : Rm → Rm is a
linear isomorphism. More explicitly,
X
xi = tij y j .
j

We denote by (gjk ) the matrix associated to the operator (T ∗ T ) in the y-coordinates and
by (gjk ) its inverse. More precisely
X
gjk := tij tik .
i

We set
|g| := det(gjk ) = det(T ∗ T ).
Then in the y-coordinates we have


∆ = ∆y := −∂yj gjk ∂yk .

Observe that
X
|x|2 = |y|2g := y j y k , q(t, x) = t2 − |x|2 = qg (t, y) = t2 − |y|2g .
jk

Then
(∂t2 + ∆y )R+
a a−2
(t, T y) = R+ (t, T y), (11.3.32)
576 CHAPTER 11. SPECTRAL GEOMETRY

and X
a−2
(a − 2) a
gjk ∂yk R+ (t, T y) = −y j R+ (t, T y). (11.3.33)
k
If we multiply both sides of the above equality with gij and then we sum over j we deduce
 
X
(a − 2)∂yi R+ a
(t, T y) = −  gij y j  R+
a−2
(t, T y). (11.3.34)
j

Let us observe that if δ0 (x) ∈ C −∞ (Rm ) denotes the Dirac generalized function concen-
trated at 0, then
1
δ0 (T y) = T ∗ δ0 = |g|− 2 δ0 (y). (11.3.35)
a (t, |y| ) ∈ C −∞ (R1,m \ 0) using the technique in (11.3.11)
For Re a ≫ m we define R̄+ g

a−n
a
R̄+ (t, |y|g ) := dm (a)(qg+ )∗ χ 2 .

The generalized function R̄+a (t, |y| ) admits a unique homogeneous extension R a (t, |y| )
g + g
to R 1,m which coincides with R+ a (t, T y). We can then use (11.3.32) to show that

a a
R̄+ (t, |y|g ) = R+ (t, T y)|R1,m, \0 , ∀a.

For this reason we will use the more suggestive notation R+ a (t, |y| ) when referring to
g
a
R+ (t, T y).
More generally, consider a Riemann metric g = (gij ) on Rm ,
X
g= gij dxi dxj .
i,j

The Laplacian in this metric is the differential operator

X
X j
∆g = − ∂xj gjk ∂xk + b ∂xj ,
j,k j
!
1 1 X
j
b = − ∇j log |g| = − jk
g ∂xk log |g| .
2 2
k

Suppose that (xi )

are geodesic coordinates near 0. This means that the rays through 0
are geodesics and thus, if we set |x|g = distg (x, 0), then
 1
2
X
|x|g =  gjk (0)x x  .
j k
(11.3.36)
j,k

Moreover4 X X
gjk (x)xk = gjk (0)xk , (11.3.37)
k k
4
The equality (11.3.37)
P is an algebraic rephrasing of the Gauss Lemma, Lemma 4.1.29, stating that the
the radial vector field j xj ∂xj is the gradient of the function 12 |x|2g .
11.3. THE WAVE EQUATION 577

gjk (x) − gjk (0) = O |x|2 , ∀j, k. (11.3.38)
We set
X (11.3.37 ) 1

h(x) := bj (x)gjk (0)xk = − g ∇ log |g|, x . (11.3.39)
2
j,k

Observe that if f ∈ C 1 , then (11.3.37) implies that

!
X X
jk
g (x)∂xk f (|x|2g ) = g jk
(x)∂xk |x|2g f ′ (|x|2g )
k k
  !
X X
= gjk (x)gik (x)xi  f ′ (|x|2g ) = δij xi f ′ (|x|2g )
i,j i
 
X X
= gjk (0)gik (0)xi  f ′ (|x|2g ) = gjk (0)∂xk f (|x|2g ).
i,j k

We deduce from (11.3.32) and the above identity that for Re a ≫ 0 we have
 
X

∂t2 − ∂xj gjk ∂xk  R+ a a−2
(t, |x|g ) = R+ (t, |x|g ). (11.3.40)
j,k

By unique continuation we conclude that the above equality holds for all a ∈ C.
More generally, if u = u(x) is a t-independent smooth function, then we deduce
from(11.3.33), (11.3.34), (11.3.39) and (11.3.40) that for any a 6= 2 we have


a−2
∂t2 + ∆g u(x)R+ a
(t, |x|g ) = uR+ + (∆g u)R+a
 
1  X (11.3.41)
− hu − 2 xj ∂xj u R+a−2
.
a−2
j

When a = 2 we have
m−1 1− n
R 2 (t, |x|g ) = (qg+ )∗ f2 , f2 = π − 2 χ+ 2
, (t, x) 6= 0.
1− n
Since χ+ 2 is homogeneous of degree 2 − n on R1,m \ 0 it admits a unique extension to
R1,m as a homogeneous generalized function of degree 2 − n. The generalized function
−n
xj (qg+ )∗ χ+ 2 ∈ C −∞ (R1,m \ 0)

is homogeneous of degree 1 − n and thus it has a unique homogeneous extension to R1,m .

The equality
n n
∂xj (qg+ )χ1− 2 = −2xj (qg+ )∗ χ− 2
over the entire space R1,m is a consequence of (11.2.16). We deduce
m−1 n
∂xj R 2 (t, |x|g ) = −2π − 2 xj (qg+ )∗ χ− 2 . (11.3.42)
578 CHAPTER 11. SPECTRAL GEOMETRY

1
∂t2 + ∆g u(x)R+2
(t, |x|g ) = u(0)|g0 |− 2 δ0 + (∆g u)R+ 2
 
m−1 X −n (11.3.43)
−2π − 2 hu − 2 xj ∂xj u (qg+ )∗ χ+ 2 .
j

Thus, if u1 , . . . , uN are smooth functions defined in a neighborhood of 0 ∈ Rm , then

N N

X

X
∂t2 + ∆g 2k
uk (x)R+ 2
= ∂t2 + ∆g u1 R+ + ∂t2 + ∆ 2k
uk (x)R+
k=1 k=2
N
X N
X −1  
2k 2k 1 1
= (∆g uk )R+ + R+ uk+1 − huk+1 + x · ∇uk+1
2k k
k=2 k=1
m−1
− 21 2 n
+u1 (0)|g0 | δ0 + (∆g u1 )R+ − 2π − 2 (hu1 − 2x · ∇u1 ) (qg+ )∗ χ− 2
1
= u1 (0)|g0 |− 2 δ0 + (∆g uN )R+
2N

N
X −1  
2k 1 1
+ R+ ∆g uk + uk+1 − huk+1 + x · ∇uk+1
2k k
k=1
m−1 n
−2π − 2 (hu1 − 2x · ∇u1 ) (qg+ )∗ χ− 2 .

Proposition 11.3.20. For any smooth function h ∈ C ∞ (Rm ) such that h(0) = 0 there
exists a unique sequence of smooth functions U1 , U2 , . . . , UN , . . . defined on a small neigh-
borhood of 0 ∈ Rm and satisfying the differential equations

U1 (0) = 1, 2x · ∇U1 = hU1 , (U 1,h )

 
1 1
x · ∇Uk+1 + 1 − h Uk+1 = −∆g Uk , ∀k ≥ 1. (U k,h )
k 2k

Proof. If we work in spherical coordinates (r, θ) in Rm \ 0, r > 0, θ ∈ S m−1 , then

x · ∇ = r∂r . The equation 2x · ∇U1 = hU1 then reads

1
∂r U1 = hU1 .
2r
Since h(0) = 0, we deduce that for any θ ∈ Rm , |θ| = 1 we have
Z 1
h(rθ) = r ∂θ h(trθ)dt = rg(rθ),
0

where g is a smooth function in a neighborhood of 0. Thus

Z r  Z r 
1 h(trθ)
U1 (rθ) = U1 (0) exp g(tθ)dt = exp dt . (11.3.44)
0 2 0 2t
11.3. THE WAVE EQUATION 579

Lemma 11.3.21. For any k ∈ Z>0 , and any smooth function w defined in a neighborhood
of 0 ∈ Rm , the differential equation
 
h
x · ∇U + k − U =w
2

has a unique solution defined in a neighborhood of 0.

Proof. To see this we rewrite the above equation as

 
h
r∂r U + k − U = w,
2

and, if we set, Rrh

V := U e− 0 2t dt ,
then we deduce Rrh
r∂r V + kV = we− 0 2t dt .
Multipliying the above equation with r k−1 we deduce
Rr h
∂r (r k V ) = r k−1 we− 0 2t
dt
,

so that Z r  
1 Rρ h(tρθ)
V (rθ) = k ρk−1 w(ρθ)e− 0 2t
dt
dρ.
r 0
⊔
⊓

Proposition 11.3.20 now follows from the above lemma. ⊔

⊓

If we choose U1 , . . . , UN as in Proposition 11.3.20, where h is defined by (11.3.39), then

we deduce
N
2

X 2k 1
∂t + ∆g Uk (x)R+ = |g0 |− 2 δ0 + (∆g UN )R+
2N
. (11.3.45)
k=1

Observe that for any positive integer ν we can find N = N (ν) such that
2N
(∆g UN )R+ ∈ Cν.
n
More precisely, since χN − 2 is homogeneous of degree N − n2 , we deduce that

2N
2N − n > 2ν ⇒ R+ ∈ Cν.

Suppose now that (M, g) is a smooth compact, connected, m-dimensional Riemann

manifold. Let |dVg | be the canonical volume density determined by the Riemann metric.
In local coordinates (x1 , . . . , xm )
p
|dVg | = |g||dx1 · · · dxm |.
580 CHAPTER 11. SPECTRAL GEOMETRY

Remark 11.3.22. We want to discuss a rather confusing fact that unfortunately perme-
ates the more analytically and less geometrically oriented texts.
A metric g on M induces a natural bijection


C ∞ (M ) → C ∞ |Λ|M , f 7→ f |dVg |.

Invariantly,
the generalized functions on M are continuous linear functionals on the space
C ∞ |Λ|M of smooth densities on M .
The Dirac measure concentrated at a point x0 ∈ M is the continuous linear map

δx0 : C(M ) → R, δc ∞
x0 (f ) = f (x0 ), ∀f ∈ C (M ).

The Dirac generalized function on M concentrated at x0 is the linear functional

δxM0 : C ∞ (|Λ|M ) → R, δxM0 f |dVg | = f (x0 ).

In local coordinates around x0 , the Dirac measure can be identified with a multiple of
the Dirac generalized function. More precisely, we have the equality of locally defined
generalized functions
1
δxM0 = p δx0 .
|g|
To understand this equality, consider a family of nonnegative smooth functions

φε : M → [0, ∞), ε > 0,

such that, Z
φε |dVg | = 1, supp φε ⊂ Bε (x0 ), ∀ε > 0.
M

These functions converge to the Dirac generalized function δxM0 .

If we choose local coordinates (x1 , . . . , xm ), then we can think of φε as functions φε (x)
defined on Rm and supports contained in tiny balls centered at the origin. As such, they
satisfy Z p
|g|(x)φε (x) |dx| = 1
p
The probability measures |g|(x)φε (x) |dx| converge weakly to the Dirac measure δ0 . ⊔
⊓

We denote by ∆ = ∆g the Laplacian associated to the Riemann metric g. More

explicitly, if we fix a point p ∈ M and local coordinates (x1 , . . . , xm ) defined in a small
geodesic ball Up centered at p such that

xi (p) = 0, i = 1, . . . , m.

In these coordinates, the Laplacian is described by

1 X √ jk
X
X j
∆ = −p ∂xk gg ∂xk = − ∂xj gjk ∂xk + b ∂xj ,
|g| j,k j,k j
11.3. THE WAVE EQUATION 581

where
1
bj = − ∇j log |g|.
2
Let
s : M × M → [0, ∞), (x, y) 7→ sy (x) = distg (x, y)

be the geodesic distance function determined by g. For y ∈ M we denote by ~ry the radial
vector field on Ty M . Using the exponential map exp : Ty M → M we can regard ~ry as
a vector field defined on a small geodesic ball Uy centered at y. For fixed y we obtain a
generalized function


(t, x) 7→ R a t, sy (x) , distg (x, y) < c
a on
which in normal coordinates at y coincides with the wave function R+

R × Ty M = R × Rm .

By choosing an orthonormal frame of T M over Up0 we can identify Ty M with Rm and then
we can regard the correspondence y 7→ R+ 2a (t, s (x)) as a smooth family of generalized
y
functions defined on a neighborhood origin in R × Rm .
Equivalently, we can identify a neighborhood of the diagonal in M × M with a neigh-
borhood of the zero section in T M . For fixed y ∈ M the generalized function R+ 2a (t, s (x))
y
is defined on Ty M .
Identifying Up0 with a neighborhood of 0 ∈ Rm , we can further view this family as a
generalized function on a neighborhood of the origin in R × Rm × Rm ,

2a


R+ (t, sy (x)) ∈ C −∞ (−c, c) × Up0 × Up0 .

Moreover, this generalized function depends holomorphically on a ∈ C. Observe that if

Re a ≫ 0, then we can identify R+ 2a (t, s (x)) with the classical locally integrable function
y

n

dm (2a)χa− 2 t2+ − distg (x, y)2 .

Thus
2a 2a
R+ (t, sy (x)) = R+ (t, sx (y)), Re a ≫ 0

and by unique continuation we deduce that the above symmetry holds for any a,

2a 2a
R+ (t, sy (x)) = R+ (t, sx (y)), ∀a ∈ C. (11.3.46)

Using Proposition 11.3.20 we deduce the following result.

Proposition 11.3.23. There exist

• a positive number c = c(M ) > 0 and

• a sequence of smooth functions Uν ∈ C ∞ (M × M ), ν = 1, 2, . . .

582 CHAPTER 11. SPECTRAL GEOMETRY

with the following property:

for any y ∈ M , the functions x 7→ Uνy (x) := Uν (x, y) satisfy the equations (U 1,h ) and
(U k,h ) for sy (x) < c, where

∂
h = g ∇ log |g|, ~ry = log |g|. (11.3.47)
∂~ry

We deduce that for (t, x, y) ∈ (−∞, c)×M ×M we have the equality of generalized functions
N

X

∂t2 + ∆x 2k
Uk (x, y)R+ t, sy (x)
k=1 (11.3.48)
− 12 2N


= |gy | δy + (∆x UN (x, y) )R+ t, sy (x) .
1
Above |gy |− 2 δy ∈ C −∞ (R × M ) is the Dirac-delta generalized function on R × M con-
centrated at (0, y) ∈ R × M . More precisely (see Remark 11.3.22) for any function
u ∈ C0∞ (R × M ) we have
1
|gy |− 2 δy , u|dVg | = u(0, y).
⊔
⊓

11.4 Spectral geometry

11.4.1 The spectral function of the Laplacian on a compact manifold
Suppose that (M, g) is a smooth compact, connected, m-dimensional Riemann manifold.
We denote by |dVg | the volume density on M determined by g and by ∆ the associated
Laplacian. The L2 -inner product of two smooth complex valued functions u, v : M → C
is defined by Z
(u, v) = u(x)v̄(x)|dVg (x)|.
M
Note that it is conjugate linear in the second variable.
Fix an orthonormal Hilbert basis of L2 (M ) consisting of real eigenfunctions (Ψn )n≥0
of M such that if
∆Ψk = λk Ψk ,
and
0 = λ0 < λ1 ≤ λ2 ≤ · · ·
We define
X
E = EM,g : [0, ∞) × M × M → R, E (τ, x, y) := Ψk (x)Ψk (y).
λk ≤τ

The function EM,g is called the spectral function of the Riemann manifold (M, g). If we
denote by Hτ the space 
Hτ := span Ψk ; λk ≤ τ ,
11.4. SPECTRAL GEOMETRY 583

then the function

M × M ∋ (x, y) 7→ Eτ (x, y) = E (τ, x, y)
is the Schwarz kernel of the orthogonal projection Pτ : L2 (M ) → Hτ , i.e.,
Z
(Pτ u)(x) = E (τ, x, y)u(y)|dVg (y)|, ∀u ∈ L2 (M, g).
M

We set
X XZ
N (τ ) = Ng (τ ) := dim Hτ = 1= Ψk (x)2 |dVg (x)|
λk ≤τ λk ≤τ M
Z (11.4.49)
= Eτ (x, x)|dVg (x)|.
M

For any positive integer k, any p ∈ [0, ∞), and any u ∈ C ∞ (M ) we set
 1
XZ
p

kuk Lpk (M ) :=  |∇ u| |dVg (x)|  .

j p

j≤k M

Theorem 11.4.1. For any ν ∈ Z≥0 there exists Cν > 0 such that
m
kEτ kC ν (M ×M ) ≤ Cν τ ν+ 2 , ∀τ > 0. (11.4.50)

Proof. The proof uses a refined Sobolev estimate.

Lemma 11.4.2. If k > ν + m 2 , then there exists C = C(k, ν) > 0 such that for any λ > 1
and any u ∈ C ∞ (M ) we have
m

λk−ν− 2 kuk2C ν (M ) ≤ C kuk2L2 (M ) + λk kuk2L2 (M ) . (11.4.51)
k

Proof of the lemma. We will prove that there exists a constant C > 0 such that for
any u ∈ C0∞ (Rm ) and any λ > 1 we have
 
Z X  Z
m X
λk−ν− 2 sup |∂xα u(x)|2 ≤ C  |∂xα u(x)|2 |dx| + λk |u(x)|2 |dx| .
x Rm |α|=k Rm
|α|=ν
(11.4.52)
The inequality (11.4.51) follows from (11.4.52) via partitions of unity, so it suffices to prove
(11.4.52) .
We make the substitution
1
x := λ− 2 y.
Then
∂y 1 m
∂x = ∂y = λ 2 y , |dx| = λ− 2 |dy|,
∂x
584 CHAPTER 11. SPECTRAL GEOMETRY

and if we set
1
v(y) = u(x) = u(λ− 2 y),
then we see that it suffices to prove (11.4.52) only in the case λ = 1 for the function v.
From the Morrey inequalities we deduce
X Z X 
α 2
sup |∂ u(x)| ≤ C |∂ α u(x)|2 |dx|.
x Rm |α|≤k
|α|=ν

Using Parseval’s formula we deduce that there exists a constant C ′ > 0 such that
 
Z X  Z X  Z
|∂ α u(x)|2 |dx| ≤ C ′  |∂ α u(x)|2 |dx| + |u(x)|2 |dx| ,
Rm |α|≤k Rm |α|=k Rm

for any u ∈ C0∞ (Rm ). ⊔

⊓

Consider now the operator L := ∆ + 1. This is a self-adjoint elliptic operator with

trivial kernel. We deduce from Corollary 10.4.12 that, for any k ≥ 0, it induces bounded,
bijective, linear operators
L : L2k+2 (M ) → L2k (M )
with bounded inverses. Hence, for any k ≥ 0 there exists a constant Ck > 0 such that

kukL2 ≤ Ck kLukL2 (M ) , ∀u ∈ C ∞ (M ). (11.4.53)

k+2 (M ) k

Let u ∈ C ∞ (M ), and set uτ := Pτ u. Then

kLk uτ kL2 (M ) ≤ (τ + 1)k kukL2 (M ) ,

so that
kuτ kL2 ≤ C(τ + 1)k kukL2 (M ) ,
2k (M )

where C is the product of the constants C0 , C2 , . . . C2k−2 defined by (11.4.53). Using

(11.4.51) we deduce that if 2k > ν + m
2 we have
m m
kuτ k2C ν (M ) ≤ C(τ + 1)ν+ 2 kuτ k2L2 (M ) ≤ C(τ + 1)ν+ 2 kuk2L2 (M ) . (11.4.54)

Fix a point p0 ∈ M and geodesic coordinates x at p0 . Observe that for any multi-index
α such that |α| = ν we have

∂xα uτ (p0 ) = (u, Vα,p0 )L2 (M ) ,

where Vα,p0 : M → R is given by

X

Vα,p0 (q) = ∂xα Eτ (p0 , q) = ∂xα Ψj (p0 ) Ψj (q).
λj ≤τ

The above discussion shows that for any u ∈ C ∞ (M ) we have

ν m
(u, Vα,p0 )L2 (M ) ≤ C(τ + 1) 2 + 4 kukL2 (M ) .
11.4. SPECTRAL GEOMETRY 585

Hence
ν m
kVα,p0 kL2 (M ) ≤ C(τ + 1) 2 + 4 .
Now observe that Vα,p0 ∈ Hτ so that

kLk Vα,p0 kL2 (M ) ≤ C(τ + 1)k kVα,p0 kL2 (M ) .

ν
Now choose k > 2. Arguing as in the proof of (11.4.54) we can invoke (11.4.51) and
deduce that
m
kVα,p0 kC ν (M ) ≤ C ′′ (τ + 1)ν+ 2 ,
where the constant C ′′ is independent of the point p0 . The above inequality is precisely
(11.4.50).
⊔
⊓

Using the inequality (11.4.50) for ν = 0 we deduce that

m

N (τ ) = O τ 2 . (11.4.55)

This can be substantially improved. Let us preview the strategy.

Fix p, q ∈ M and consider the function
R ∋ τ 7→ E(τ ) = Ep,q (τ ) := (sign τ )E (τ 2 , p, q)
X
= (sign τ ) Ψk (p)Ψk (q) ∈ R. (11.4.56)
λk ≤τ 2

Note that Z
2
N (τ ) = Ep,p (τ ) |dVg (p)|.
M
We will obtain sharp estimates for N (τ ) via sharp estimates for Ep,p (τ ).
Theorem 11.4.1 shows that, for fixed p, Ep,p (τ ) is a temperate generalized function in
the τ -variable. Its τ -derivative is the generalized function
1 d 1X  
Ep,q (τ ) = Ψk (p)Ψk (q) δ√λk (τ ) + δ−√λk (τ ) ∈ C −∞ (R).
2 dτ 2
k

Our good fortune is that we can obtain quite a bit of information about the Fourier
transform of the generalized function
1 d
τ 7→ Ep,q (τ ).
2 dτ
More precisely
h1 d i 1 X −it√λk √ 
Fτ Ep,q (τ ) (t) = Kt := e + eit λk Ψk (p)Ψk (q)
2 dτ 2
k
X p

= cos t λk Ψk (p)Ψk (q).
k
We will devote the next subsection to a more in-depth investigation of Kt . But, before we
do so, we want to give an explicit description of the spectral function in a special case.
586 CHAPTER 11. SPECTRAL GEOMETRY

Example 11.4.3. (Spectral function of a flat torus.) Denote by Tm the flat torus

Tm := Rm /(2πZ)m

equipped with the flat metric

m
X
g := (dθ j )2 .
j=1

Note that vol(Tm ) = (2π)m .

The eigenvalues of the corresponding Laplacian ∆m are

|~k|2 , ~k = (k1 , . . . , km ) ∈ Zm .

For θ~ = (θ 1 , . . . , θ m ) ∈ R and ~k ∈ Zm we set

m
X
h~k, θi
~ := kj θ j .
j=1

Denote by ≺ the lexicographic order on Zm . An orthornormal basis of L2 (Tm ) is given by

the functions 
 ~k = ~0
1 1,
1
Ψ~k (~
θ) = m 2 2 sinh~k, ~θi, ~k ≻ 0,
(2π) 2 
 12
2 cosh~k, θi,~ ~k < 0.

We have
1 2 X 2 X
ER2 (~θ, ϕ
~) = + sinh~k, θi
~ sinh~k, ϕ
~ i + cosh~k, ~θi cosh~k, ϕ
~i
(2π)m (2π)m (2π)m
|~k|≤R |~k|≤R
~k≻~0 ~k≻~0

1 2 X 1 X ~ ~
= + cosh~k, θ~ − ϕ
~ i = eihk,θ−~ϕi .
(2π)m (2π)m (2π)m
|~k|≤R |~k|≤R
~k≻~0

Hence X
1 ~ ~
ER2 (~
θ, 0) = eihk,θi .
(2π)m
|~k|≤R

Consider the rescaled function

1 1~
1 X 1~ ~ 1 X ~
E¯R (θ) = m ER2 θ, 0, = m
eih R k,θi = eih~ν ,θi .
R R (2πR) (2πR)m 1
|~k|≤R ~
ν ∈( R Z)m
|~
ν |≤1

Observe that as R → ∞ we have

Z  
1 1
E¯R (θ) = eihξ,θi
|dξ| + O , (11.4.57)
(2π)m |ξ|≤1 R
11.4. SPECTRAL GEOMETRY 587

uniformly on the compact sets of the ~θ-space.

Let us consider a more general, but simpler situation. Suppose that w ∈ S (R). Define
!
X |~k|2
~ ϕ
Ew,R (θ, ~ ) := w Ψ~k (~θ)Ψ~k (~
ϕ).
R2
~k∈Zm

Arguing as above we deduce that

!
1 X |~k|2 ~ ~
Ew,R (~
θ, ϕ
~) = m
w eihk,θ−~ϕi . (11.4.58)
(2π) R2
~k∈Zm

Again, we form the rescaled function

!
1 1~ 1 X |~k|2 1~ ~
E¯w,R (~
θ) = m Ew,R ( θ, 0) = w eih R k,θi .
R R (2πR)m R2
~k∈Zm

~ and ~x ∈ Rm we define
For every θ,
~
fθ~ (x) = w(|~x|2 )eih~x,θi
so that fθ~ ∈ S (Rm ). Hence
X  
1 1~
E¯w,R (~θ) = f θ~ R
k .
(2πR)m
~k∈Zm

At this point we want to invoke the Poisson’s summation formula.

Lemma 11.4.4 (Poisson sumation). For any φ ∈ S (Rm ) and any a > 0 we have
X  2π   a m X
φ ~k = b ν ),
φ(a~ (11.4.59)
a 2π m
~k∈Zm ~
ν ∈Z

where φb is the Fourier transform of φ.

Proof. To prove Poisson’s formula is suffices to consider the case a = 1. The general case
follows from this case applied to the Schwartz function
 
1
φa (~x) = φ ~x .
a
Consider the (2πZ)m -periodic function
X

Φ : Rm → C, Φ(~x) = φ ~x + 2π~k .
~k∈Zm

This can be viewed as a smooth function on the torus Tm and, as such, it admits a Fourier
(spectral) decomposition
X Z
1 ~
Φ(~x) = m
C~k e ν ,~
ih~ xi
, C~ν = Φ(~θ)e−ih~ν ,θi |dεθ|.
(2π) m [0,2π]m
~
ν ∈Z
588 CHAPTER 11. SPECTRAL GEOMETRY

The above series converges in C ∞ . The sum in the left-hand side of (11.4.59) is Φ(0). We
have
1 X
Φ(0) = C~ν .
(2π)m m ~
ν ∈Z

Now observe that

X Z ~
Z
~
C~ν = φ(~
θ)e−ih~ν ,θi |dθ| = φ(~θ)e−ih~ν ,θi |d~θ| = φ(~ν ).
~k∈Zm [0,2π]m −~k Rm

⊔
⊓

If in Poisson’s formula we choose a = 2πR, then we deduce

 
1 X 1~ X
b
φ k = φ(2πR~
ν ).
Rm m
R m
~
ν ∈Z ~
ν ∈Z

Now let
~
φ(~x) = eih~x,θi w(|~x|2 ) .
| {z }
=:u(~
x)

We deduce that
b =u
φ(ξ) ~
b(ξ − θ).
Hence
1 X
E¯w,R (~
θ) = b(2πR~ν − ~θ)
u
(2π)m m~
ν ∈Z

1 1 X
= b(−~
u θ) + u ~
b(2πR~ν − θ).
(2π)m (2π)m
ν ∈Zm \0
~

b ∈ S (Rm ) we deduce that for any p > 0 we have

Since u

1
E¯w,R (~
θ) = b(~θ) + O(R−p ), as R → ∞,
u (11.4.60)
(2π)m

uniformly on the compacts of the ~ θ-spaces. Note that (11.4.57) is formally a special case
of (11.4.60) in which u is the characteristic function of the unit ball in the ~x-space.

11.4.2 Short time asymptotics for the wave kernel

Let (M, g) be as in the previous section. For t ∈ R we consider the bounded operator
√

cos t ∆ : L2 (M ) → L2 (M ),

defined as follows. Pick u ∈ L2 (M ). Then u has a Fourier decomposition

X
u= uk Ψk , uk = (u, Ψk ),
k≥0
11.4. SPECTRAL GEOMETRY 589

where X
kuk2L2 (M ) = |uk |2 .
k≥0

We set √
X p

cos t ∆ u = cos t λk uk Ψk . (11.4.61)
k≥0

Note that √ X p
k cos(t ∆)uk2L2 (M ) = cos2 (t λk )|uk |2 ≤ kuk2L2 (M )
k≥0
√
so that the operator cos(t ∆) is indeed well defined and bounded L2 (M ) → L2 (M ).
Note that if u ∈ C ∞ (M ), then for any N ∈ Z≥0 we have
X
∆N u = λN
k uk Ψ k
k≥0

and we deduce
√ √ X p
cos(t ∆)∆N u = ∆N cos(t ∆)u = λN
k cos(t λk )uk Ψk (p).
k≥0

This shows that

√

k∆N cos t ∆ ukL2 (M ) ≤ k∆N ukL2 (M ) , ∀N ∈ Z≥0 .

If we formally differentiate (11.4.61) we respect to t, we deduce

√
X p
∂t2j cos(t ∆)u (p) = (−1)j λjk cos(t λk )uk Ψk (p).
k≥0

The above discussion shows that if u ∈ C ∞ (M ), then

√
cos(t ∆)u ∈ C ∞ (R × M ).

For any u, v ∈ C ∞ (M ) we define u ⊠ v ∈ C ∞ (M × M ) by the equality

(u ⊠ v)(p, q) = u(p)v(q), ∀p, q ∈ M.

√
The Schwarz kernel Kt (p, q) of cos(t ∆) is the generalized function
X p
Kt (p, q) = cos(t λk )Ψk (p)Ψk (q) ∈ C −∞ (M × M ).
k

This means that for any u, v ∈ C ∞ (M ) we have

√
cos(t ∆)u, v|dVg | = Kt , (u ⊠ v)|dVg×g |
X p
:= cos(t λk i(u, Ψk )L2 (v, Ψk )L2 .
k≥0
590 CHAPTER 11. SPECTRAL GEOMETRY

The family of generalized function Kt (p, q) ∈ C −∞ (M × M ), t ∈ R is called the wave

kernel of the Riemann manifold (M, g).
For fixed p, q ∈ M we obtain a temperate generalized function on R, t 7→ Kt (p, q),
defined by
X Z p 
hKt , ηi = η(t) cos(t λk )dt Ψk (p)Ψk (q), ∀η ∈ S (R).
k R

Theorem 11.4.1 shows that the above series is absolutely convergent for any η ∈ S (R).
This generalized function is the Fourier transform of the temperate generalized function
1 d 1X

Ep,q (τ ) = Ψk (p)Ψk (q) δ√λk (τ ) + δ−√λk (τ ) ∈ S ′ (R). (11.4.62)
2 dτ 2
k

Thus, an approximation for Kt will yield an approximation for the Fourier transform of
the derivative of the modified spectral function
1
τ 7→ (sign τ )E (τ 2 , p, q).
2
We will spend the rest of this section constructing an approximation for the wave kernel
Kt . √
If f ∈ C ∞ (M ), then the function u(t, p) = cos(t ∆)f ∈ C ∞ (R × M ) satisfies the
initial value problem
(∂t2 + ∆)u = 0, on (0, ∞) × M, (11.4.63a)
u(0, p) = f (p), ut (0, p) = 0, ∀p ∈ M. (11.4.63b)
We deduce that the kernel Kt defines a smooth map
(0, ∞) × M ∋ (t, p) 7→ Kt (p, q) ∈ C −∞ (M ).

The a priori estimates described in our next result will play a key role in our search for
an approximation for Kt .
Proposition 11.4.5. Let v ∈ C ∞ ([0, 1] × M ) be a solution of the non-homogeneous wave
equation
(∂t2 + ∆)v = h on [0, 1] × M
(11.4.64)
v(0, p) = vt (0, p) = 0, ∀p ∈ M.
Assume that
∂tj h(0, p) = 0, ∀p ∈ M, j < k. (11.4.65)
Then there exists Ck > 0 independent of h such that for any t ∈ (0, 1) we have
k+1
X
∂tk+1−j v(t, −) L2j (M )
≤
j=0
  (11.4.66)
Z t k−1
X
Ck  ∂sk h(s, −) L2 (M ) ds + ∂tk−1−j h(t, −) L2 (M )  .
0 j=0
11.4. SPECTRAL GEOMETRY 591

Proof. For any t ∈ ([0, 1] we have

1d  
kvt k2L2 (M ) + (∆v, v)L2 (M ) = (vtt , vt )L2 (M ) + (∆v, vt )L2 (M )
2 dt
= (h, ut )L2 (M ) .
Observing that
(∆v, v)L2 (M ) = kdvk2L2 (M ) ,
we deduce by integrating that
Z t
k∂t v(t, −)k2L2 (M ) + kdv(t, −)k2L2 (M ) =2 (h, ∂s v)L2 (M ) ds
0
Z t Z 1
≤ 2K kh(s, −)kL2 (M )ds, K = kvt kL2 (M ) ds.
0 0
Since Z 1
2
K ≤ k∂t v(s, −)k2L2 (M ) |ds|,
0
we deduce Z 1
K 2 ≤ 2K kh(s, −)kL2 ds,
0
so that Z 1
k(∂t v(t, −)kL2 + k(dv(t, −)kL2 ≤ C kh(s, −)kL2 ds.
0
In particular,
Z t Z t
kv(t, −)kL2 ≤ k∂s v(s, −)kL2 ds ≤ Ct kh(s, −)kL2 ds
0 0
Z t (11.4.67)
≤C kh(s, −)kL2 ds.
0

This proves (11.4.66) for k = 0. We can now argue inductively.

Assume that we have proved (11.4.66) for k and we prove it for k + 1. We consider
the function u := ∂t v. Note that
u(0, p) = 0, ∀p ∈ M,
and
ut (0, p) = ∂t2 v(0, p) = h(0, p) − ∆v(0, p) = h(0, p) = 0, ∀p ∈ M.
Thus u satisfies an initial value problem of the type (11.4.64), where h is replaced with
∂t h. By induction we deduce from (11.4.66), where h is replaced with ∂t h, that
k+1
X k+1
X
(∂tk+2−j v(t, −) L2j
= (∂tk+1−j u(t, −) L2j
j=0 j=0
  (11.4.68)
Z t k−1
X
≤ Ck  (∂sk+1 h(s, −) L2 ds + (∂tk−j h(s, −) L2  .
0 j=0
592 CHAPTER 11. SPECTRAL GEOMETRY

Now observe that

kv(t, −)kL2 ≤ C k∆v(t, −)kL2 + kv(t, −)kL2
k+2 k

Z t 
(11.4.67 )
= C kh(s, −)kL2 ds + k∂t2 v(t, −)kL2 + kh(t, −)kL2 .
k k
0

Using this in (11.4.68) we obtain the sought for inequality.

⊔
⊓

Using the functions Uk ∈ C ∞ (M × M ), k = 1, 2, . . . constructed in Proposition 11.3.23

we deduce that there exists a constant c ∈ (0, 1) such that for any N > 0, any y ∈ M and
any t < c we have

N

X
1

∂t2 +∆x 2k
Uk (x, y)R+ t, sy (x) = |gy |− 2 δy +(∆x UN (x, y) )R+
2N
t, sy (x) . (11.4.69)
|k=1 {z }
=:EN (t,x,y)

The generalized function EN is an “approximate” fundamental solution of the wave op-

erator on R × M .√We want to prove that ∂t EN (t, −, −) is a good approximation for the
kernel Kt of cos(t ∆).
Using Theorem 11.3.14 as inspiration we associate to each f ∈ C ∞ (M ) and each large
positive integer N the function
Z
Wf (t, x) = WfN (t, x, y) := EN (t, x)f (y)|dVg (y)|.
M

We need to explain the above construction. The generalized function EN defines a smooth
map
(0, ∞) × M ∋ (t, x) 7→ EN (t, x, −) ∈ C −∞ (M ),
and thus
WfN (t, x) = EN (t, x, −), f |dVg | .
Using (11.3.19) we deduce that
Z
Wf (0+ , x) = 0, ∂t Wf (0+ , x) = 2 +
U1 (x, y)∂t R+ (0 , sy (x) )f (y)|dVg (y)|
M

(11.3.46 ) 1
= |gx |− 2 δx , U1 (x, −)f (−)|dVg | = U1 (x, x)f (x) = f (x).
Observe that Z
(∂t2 + ∆)Wf (t, x, y) = SN (t, x, y)f (y)|dVg (y)|,
M

where
2N


SN (t, x, y) = (∆x UN (x, y) )R+ t, sy (x) .
11.4. SPECTRAL GEOMETRY 593

For N large SN (t, x, y) = 0 for t = 0 and we deduce that

Z
∂t2 Wf (0+ , x) = SN (0, x, y)f (y)|dVg (y)| −∆W (0+ , x) = 0.
|M {z }
=0

Thus,√the function ∂t Wf satisfies the same initial conditions (11.4.63b) as the function
cos(t ∆)f . Consider their difference, i.e., the function
√
Vf (t, x) = VfN (t, x) = cos(t ∆)f − ∂t WfN (t, x).

The function Vf is the unique solution of the initial value problem

(∂t2 + ∆)Vf = −(∂t2 + ∆)∂t Wf = hf ,

Vf (0, x) = ∂t Vf (0, x) = 0,

where Z
hf = hN
f (t, x) := − ∂t SN (t, x, y)f (y)|dVg (y)|
M
Z
= ∂t SN (t, x, y)f (y)|dVg (y)|.
dist(y,x)<t

The distribution ∂t R+2N is homogeneous of degree 2N − n − 1 = 2N − m − 2. If we

2N ∈ C k+1 and thus for any
choose N sufficiently large, e.g., 2N > n + 2(k + 1), then R+
|α| ≤ k, and any x, y such that sy (x) ≤ t ≤ c we have
α 2N
∂t,x,y ∂t R+ (t, x, y) ≤ Ct2N −n−1−|α| .

Hence for any t ∈ (0, c) we have

Z t k−1
X m
k∂sk hf (s, −)kL2 ds + k∂tk−1−j hf (t, −)kL2 ≤ Ct2N − 2 −2−k kf kL2 . (11.4.70)
j
0 j=0

Proposition 11.4.5 now implies that for t ∈ (0, c)

k+1
X m
kVf (t, −)kL2 + k∂t Vf (t, −)kL2 ≤ ∂tk+1−j V (t, −) L2j
≤ Ct2N − 2 −2−k kf kL2 .
k+1 k
j=0

The estimate (11.4.67) implies

m
kVf (t, −)kL2 ≤ Ct2N − 2 −2 kf kL2 .
m
Using Lemma 11.4.2 we deduce that if k > ν + 2, so that N > ν + m + 23 , then for any
λ > 1 we have
m

λk+1−ν− 2 kVf (t, −)k2C ν ≤ C kVf (t, −)k2L2 + λk kVf (t, −)k2L2
k+1

 
≤ C t4N −m−4−2k + λk t4N −m−4 kf k2L2 .
594 CHAPTER 11. SPECTRAL GEOMETRY

If we let λ = t−2 we deduce

t2ν+m−2k−2 kVf (t, −)k2C ν ≤ Ct4N −m−2k−4 kf k2L2 ,

so that
kVfN (t, −)kC ν ≤ t2N −n−1−ν kf kL2 , ∀t ∈ [0, c], ∀N ≫ 0.
We deduce similarly that

k∂t VfN (t, −)kC ν ≤ t2N −n−2−ν kf kL2 , ∀t ∈ [0, c], ∀N ≫ 0.

Using the equality

∂t2 Vf = hf − ∆Vf
we deduce gradually from the above estimates that
α
k∂t,x VfN (t, −)kC ν ≤ Ct2N −n−1−|α| kf kL2 , ∀t ∈ [0, c], ∀N ≫ 0. (11.4.71)

Similarly, the function V∆j f satisfies the initial value problem

(∂t2 + ∆)V∆j f = −(∂t2 + ∆)∂t U∆j f = h∆j f ,

V∆j f (0, x) = ∂t V∆j f (0, x) = 0,

where
Z Z
j
h∆j f := − ∂t SN (t, x, y)∆ f (y)|dVg (y)| = − ∆jy ∂t SN (t, x, y)f (y)|dVg (y)|
M M
Z
=− ∆jy ∂t SN (t, x, y)f (y)|dVg (y)|.
dist(y,x)<t

Arguing as before we deduce that for any k, j ∈ Z≥0 and N sufficiently large,

kV∆Nj f (t, −)kC ν (M ) ≤ t2N −n−1−2j−ν kf kL2 , ∀t ∈ [0, c],

k∂tk V∆Nj f (t, −)kC ν (M ) ≤ t2N −n−1−k−2j−ν kf kL2 , ∀t ∈ [0, c].

Now observe that we can write

VfN (t, x) = Kt (x, y) − ∂t EN (t, x, y) , f (y)|dVg (y)| ,

| {z }
=:TN (t,x,y)

where we regard the tail TN (t, x, y) as a smooth map

(0, c) × M ∋ (t, x) 7→ TN (t, x, y) ∈ C −∞ (M ).

We deduce that for any multi-index α and any j ≥ 0 there exist N = N (α, j) and C > 0
such that for any N ≥ N (α, j), f ∈ C ∞ (M ) and any t ∈ (0, c) we have
α
sup ∂t,x TN (t, x, y) , ∆jy f (y)|dVg (y)| ≤ Ct2N −n−1−2j−|α|kf kL2 . (11.4.72)
x∈M
11.4. SPECTRAL GEOMETRY 595

From the above inequality we deduce that for any (t, x) the generalized function ∆y TN (t, x, y)
is represented by a function in L2 (M ) for N sufficiently large. From elliptic regularity we
deduce that for N sufficiently large and any (t, x) the function y 7→ TN (t, x, y) belongs to
the Sobolev space L2j,2 and
α
k∂t,x ∆jy TN (t, x, y)kL2 ≤ Ct2N −n−1−2j−|α| .

Using elliptic estimates for ∆y and Lemma 11.4.2 we deduce that for any k, ν ∈ Z≥0 there
exist N (k, ν) and C > 0 such that, for any N ≥ N (ν, k), any x ∈ M , t ∈ [0, c], and any
multi-index |α| ≤ k we have
α
k∂t,x Tn (t, x, −)kC ν ≤ Ct2N −n−1−|α|−ν .

This implies that for any µ ∈ Z≥0 there exists N = N (µ) and C > 0 such that for any
N > N (µ) we have TN (t, x, y) ∈ C µ ( (0, c) × M × M ) and for any j ≤ µ we have

k∂tj TN (t, −, −)kC µ−j (M ×M ) ≤ Ct2N −n−1−µ . (11.4.73)

We can rewrite this as

n o
∂tj Kt (−, −) − ∂t EN (t, −, −) ≤ Ct2N −n−1−µ , t ∈ [0, c]. (11.4.74)
C µ−j (M ×M )

Since Kt is even in t we deduce that for |t| ≤ c, N ≥ N (µ) we have

n
o
∂tj Kt (−, −)−∂t EN (t, −, −)−EN (−t, −, −) µ−j
≤ Ct2N −n−1−µ . (11.4.75)
C (M ×M )

We set
Ee(t, x, y) := EN (t, x, y) − EN (−t, x, y),
eN (t, x, y) := Kt (x, y) − ∂t EeN (t, x, y).
T
We can rewrite (11.4.75) as

eN (t, −, −)
∂tj T ≤ C|t|2N −n−1−µ , |t| ≤ c, j ≤ µ, N ≥ N (µ). (11.4.76)
C µ−j (M ×M )

Set
n := 1 + m = 1 + dim M, η(x, y) := dist(x, y)2 ,
 a− n
a− n

H2a (t, x, y) := dm (2a)∂t χ+ 2 t2+ − η(x, y) − χ+ 2 t2− − η(x, y)


 (11.4.77)
2a 2a
= ∂t R+ t, sy (x) − R+ −t, sy (x) .

We have thus obtained the following result.

Theorem 11.4.6. We have an asymptotic expansion
∞
X
Kt (x, y) ∼ Uk (x, y)H2k (t, x, y), |t| ≤ c, (11.4.78)
k=1
596 CHAPTER 11. SPECTRAL GEOMETRY

where the functions Uk ∈ C ∞ (M ×M ), k = 1, 2, . . . are constructed in Proposition 11.3.23

.
More precisely, (11.4.78) means that, for any positive integer µ, there exists a positive
integer N (µ) so that, for any N ≥ N (µ), the tail
N
X
eN (t, x, y) := Kt (x, y) −
T Uk (x, y)H2k (t, x, y)
k=1


belongs to C µ (−c, c) × M × M and satisfies the estimates (11.4.76). ⊔
⊓

For N ≥ N (1) and y ∈ M fixed we obtain a C 1 -function

eN (t, y, y) ∈ R.
(−c, c) → T

Using Proposition 11.3.18 we deduce that

N
X 


2k 2k
Uk (x, y)∂t R+ t, sy (x) − R+ −t, sy (x)
x=y
k=1

N
X
=2 Uk (y, y)dm (2k)|χ|2k−1−n (t).
k=1
N
2 X Γ(k − n2 )
= m Uk (y, y)|χ|2k−1−n (t).
π 2 k=1 Γ(2k − 1 − n)
We have thus proved the following result.
Corollary 11.4.7. For any y ∈ M the restriction of Kt (x, y) to the line

R × {(y, y)} ⊂ R × M × M

is a well defined generalized function Kt = Kt (y, y) ∈ C −∞ (R) satisfying the asymptotic

expansion
X∞
Kt ∼ 2 Uk (y, y)dm (2k)|χ|2k−1−n (t), |t| < c (11.4.79)
k=1

meaning that for any µ > 0 there exists N (µ) > 0 such that for N ≥ N (µ) we have
N
X
eN (t, y, y) = Kt (y, y) − 2
T Uk (y, y)dm (2k)|χ|2k−1−n (t) ∈ C µ ((−c, c), ×M ).
k=1

and
eN (t, y, y) = Cj |t|2N −n−1−j , |t| ≤ c, j ≤ µ; |α| ≤ µ − j
∂tj ∂yα T

for some constant Cj > 0 independent of y. ⊔

⊓
11.4. SPECTRAL GEOMETRY 597

11.4.3 Spectral function asymptotics

It is now time to reap the benefits of all the previous theoretical work. Recall that
X
Ng (R2 ) := dim ker(λ − ∆g ).
λ∈[0,R2 ]

Theorem 11.4.8. There exists a constant K > 0 such that, for any y ∈ M we have

X 1
Ψk (y)Ψk (y) − ω m τ m ≤ Kτ m−1 , as τ → ∞. (11.4.80)
(2π)m
λk ≤τ 2

In particular
ω m Rm

Ng (R2 ) = volg (M ) + O R m−1
as R → ∞. (11.4.81)
(2π)m

Proof. Consider the increasing functions

1
µ(τ ) = Ey,y (τ ), ν(τ ) = ω m |τ |m sign(τ ),
(2π)m

where Ex,y (τ ) is the modified spectral function defined in (11.4.56). We will prove that


µ(τ ) − ν(τ ) = O τ m−1 as τ → ∞. (11.4.82)

We will obtain information about Ey,y (τ ) by recalling that, for every x, y, the Fourier
transform in the temporal variable of 21 dµ
dt is the wave kernel Kt (x, y).
On the other hand, according to (11.3.31) the Fourier transform 12 dτ dν
is the first term
in the asymptotic expansion (11.4.79) of Kt for t near zero.
To put this fortunate coincidence to good use requires some technical skill and imagi-
nation. What follows is inspired from [67, Sec.17.5].
Choose a positive function φ(τ ) ∈ S (R) such that

b > 0, ∀|t| ≤ c , φ(t)

φ(t) b = 0, ∀|t| ≥ c,
2Z
(11.4.83)
b = 1.
φ(τ )dτ = φ(0)
R

Such a function is easy to construct. We define φ as a convolution

φ = |ψ|2 ∗ |ψ|2 ,

where

b is supported in the interval |t| ≤ c .

• The Fourier transform ψ(t) 2
R
• R |ψ(τ )|2 dτ = 1.
598 CHAPTER 11. SPECTRAL GEOMETRY

Observe that 2πψ(−τ ) is the Fourier transform of the function ψb ∈ C0∞ (R). We deduce
that it is real analytic5 thus its zero set is discrete. In particular

|ψ|2 ∗ |ψ|2 (τ ) > 0, ∀τ.

Then  d 2
b = |ψ|
φ(t) 2 (t) ≥0
and Z 2
φ(0) d2 (0)2 =
b = |ψ| 2
|ψ(τ )| dτ = 1.
R
Fix a cut-off function φ satisfying (11.4.83). For any a ≫ 0 we set
1 τ 
φa (τ ) := φ
a a
ca (t) = φ(at)
so that φ b has support in [− ac , ac ]. We deduce that for any a > 0 we have
∞
X
ca (t)Kt ∼ 2
φ ca (t)|χ|2k−1−n (t),
Uk (y, y)dm (2k)φ (11.4.84)
k=1

More precisely for N ≫ 0 the remainder

N
X
φ eN (t, y, y) = φ
ca (t)T ca (t)Kt − 2φ
ca (t) Uk (y, y)dm (2k)|χ|2k−1−n (t)
k=1

is a continuous function supported on the interval [− ac , ac ] and satisfying the estimate

eN (t, y, y) ≤ C|t|2N −n−1 , |t| ≤ c,

ca (t)T
φ

where the constant C is independent of a. Here we can be even more specific. It suffices
to choose N such that
m l mm
2N − 1 − n ≥ 0⇐⇒N ≥ 1 + ⇐⇒N ≥ N (m) := 1 + .
2 2
Hence
eN (t, y, y) ≤ C|t|2N (m)−n−1 ≤ C|t|m , |t| ≤ min(1, c).
ca (t)T
φ
As we know, the inverse Fourier transform of Kt is the generalized function
1 d X

Ep,q (τ ) = Ψk (p)Ψk (q) δ√λk (τ ) + δ−√λk (τ ) ∈ C −∞ (R).
2 dτ
k

For this reason we plan to take the inverse Fourier transform of the equality
X
ca (t)Kt = 2
φ ca (t)|χ|2k−1−n (t)
Uk (y, y)dm (2k)φ
2≤2k<n+1
ca (t)T
+φ eN (m) (t, y, y).
5
The Fourier-Laplace transform of a smooth compactly supported function is an entire function.
11.4. SPECTRAL GEOMETRY 599

Using the identity (11.2.11b) we deduce that the inverse Fourier transform of |χ|2k−1−n (t)
is the function
22k−1−n n−2k 22k−1−n m+1−2k
√ |χ| (τ ) = √ |χ| (τ ).
π π
Hence
1 d X 22k−m−1 dm (2k)
(φa ∗ Ey,y (τ ) ) = Uk (y, y) √ φa ∗ |χ|m+1−2k (τ )
2 dτ π
2≤2k≤n (11.4.85)
h i
+ F −1 φ eN (m) (t, y, y) ,
ca (t)T

where F −1 denotes the inverse Fourier transform

Z
1
F −1 [f (t)] = eiτ t f (t)dt.
2π R
Using (11.3.31) and the equality U1 (y, y) = 1 we see that the first term of the sum in
(11.4.85) above is
1 d ω m |τ |m
sign τ.
2 dτ (2π)m
We can rewrite (11.4.85) as
!
1 d 1

φa ∗ Ey,y (τ ) − m
ω m |τ |m sign τ
2 dτ (2π)
X 22k−m−1 dm (2k)
= Uk (y, y) √ φa ∗ |χ|m+1−2k (τ )
π (11.4.86)
2<2k≤n
 
c e 
+F −1 φ a (t)TN (m) (t, y, y) .
| {z }
=:ρa (t)

Let us estimate the terms on the right-hand side of (11.4.86).

The remainder ρa (t) is supported on the interval [− ac , ac ] and satisfies the estimate

|ρa (t)| ≤ C|t|m ,

for some constant C independent of a. Note that

Z c Z τc
1 a 1 a
F −1 [ρa ] = eiτ t ρa (t)dt = eis ρa (s/τ )ds
2π − c 2π|τ | − τ c
a a

Z τc
C a C
≤ sm ds ≤ .
|τ |m+1 0 am+1
Next, we need to estimate terms of the form φa ∗ |χ|p (τ ), p ∈ Z>0 . In other words we need
to estimate functions of the form
Z
1 τ − s
p
fp (τ ) := φ |s| ds.
a R a
600 CHAPTER 11. SPECTRAL GEOMETRY

Note that Z Z
p p p
fp (aτ ) = a φ(τ − s)|s| ds = a |τ − s|p φ(s)ds.
R R
Using the inequality


|τ − s|p ≤ (|τ | + |s|)p ≤ 2p−1 |τ |p + |s|p ,

we deduce  
Z Z
 
|fp (aτ )| ≤ 2p−1 
(a|τ |)
p
φ(s)ds +ap |s|p φ(s)ds
.
R R
| {z }
=1

so that
p
|fp (τ )| ≤ Cp |τ | + a , ∀τ. (11.4.87)
Using (11.4.86) and (11.4.87) we deduce that there exist positive constants K0 , K1 , inde-
pendent of y, such that
dν
≤ K0 |τ |m−1 , ∀τ, (11.4.88a)
dt
d
m−3
φa ∗ (µ − ν)(τ ) ≤ K1 |τ | + a , ∀τ. (11.4.88b)
dt
Arguing as in the proof of (11.4.87) we deduce

d
m−1
φa ∗ ν(τ ) ≤ K2 |τ | + a .
dt

Set
ε0 := min φ(τ ) > 0.
|τ |≤ 21

The function µ is nondecreasing and defines a Lebesgue-Stiltjes measure dµ on R. We

observe that
Z Z  
d
1 t−s
φa ∗ µ (τ ) = dµ ∗ φa (τ ) = φa (τ − s)dµ(s) = φ dµ(s).
dτ R a R a

Hence Z τ + a2
−1
ε0 a dµ ≤ dµ ∗ φa (τ ) = (dµ − dν) ∗ φa (τ ) + dν ∗ φa (τ )
τ − a2

m−3
m−1
≤ K1 |τ | + a + K2 |τ | + a .
We deduce that for any b ∈ [0, a] we have
Z τ
µ(τ ) − µ(τ − b) ≤ µ(τ ) − µ(τ − a) ≤ dµ
τ −a
 m−3  m−1
3a 3a
≤ K1 |τ | + + K2 |τ | + .
2 2
11.4. SPECTRAL GEOMETRY 601

For s > 0, we divide (0, s) into ν ≤ s + 1 intervals of length ≤ 1 and we deduce from the
above inequality that
    !
3a m−3 3a m−1
µ(τ ) − µ(τ − as) ≤ (s + 1) K1 |τ | + + K2 |τ | + .
2 2

Similarly
 m−3  m−1 !
3a 3a
µ(τ + as) − µ(τ ) ≤ (s + 1) K1 |τ + as| + + K2 |τ + as| +
2 2

We deduce that for any s ∈ R we have

|µ(τ ) − µ(τ − as)|

 
≤ (|s| + 1) K3 ( |τ | + a(|s| + 1) )m−3 + K4 (|τ | + |a|(|s| + 1) )m−1 .

Observe that
Z   Z
1 τ −s
µ ∗ φa (τ ) = µ(s)φ ds = µ(τ − as)φ(s)ds.
a R a R

Hence Z
|µ(τ ) − µ ∗ φa (τ )| = (µ(τ ) − µ(τ − as))φ(s)ds
R
Z  
≤ (|s| + 1) K3 ( |τ | + a(|s| + 1) )m−3 + K4 (|τ | + |a|(|s| + 1) )m−1 φ(s)ds
R



≤ K5 |τ |m−3 + |τ |m−1 + am−3 + am−1 = O |τ |m−1 .
Arguing similarly we deduce that



|ν(τ ) − ν ∗ φa (τ )| ≤ K6 |τ |m−1 + am−1 = O |τ |m−1 .

Hence

|µ(τ ) − ν(τ )| ≤ |µ(τ ) − µ ∗ φa (τ )| + |ν(τ ) − ν ∗ φa (τ )| + |(µ − ν) ∗ φa (τ )|

Z τ
m−3 m−1 m−3 m−1


≤ K7 |τ | + |τ | +a +a + |φa ∗ d(µ − ν)(s)|
0
Z τ


≤ K7 |τ |m−3 + |τ |m−1 + am−3 + am−1 + K2 (|s| + a)m−3 ds
0
m−3 m−2 m−1 m−3 m−2



≤ K8 |τ | + |τ | + |τ | +a +a + am−1 = O |τ |m−1 .
This proves (11.4.82)
⊔
⊓
602 CHAPTER 11. SPECTRAL GEOMETRY

11.4.4 Spectral estimates of smoothing operators

For every even function w : R → R such that w(x) decays sufficiently fast as |x| → ∞ we
set
wε (x) := w(εx), ∀x ∈ R
and we define E ε = Ewε ∈ C ∞ (M × M ),
X p

Ewε (p, q) := wε λk Ψk (p)Ψk (q).
k≥0

The function Ewε is the Schwartz kernel of the bounded operator

√
√ p

wε ∆ : L2 (M ) → L2 (M ), wε ( ∆)Ψk = wε λk Ψ k .

If w is the indicator function of [−1, 1],

(
1, x ∈ [−1, 1],
w(x) =
0, otherwise,

1/τ
then Ew coincides with the spectral
√ function we have investigated earlier. Note also that
2
when w(x) = e−x , then w√t ( ∆) = e−t∆ . This is the so called heat operator.
In this section, building on the results developed so far we present in some detail
information about the behavior along the diagonal of this kernel as ε ց 0 when w is a
Schwartz function.
Theorem 11.4.9. Suppose that w ∈ S (R) is an even, nonnegative Schwartz function,
and (M, g) is a smooth, compact, connected m-dimensional Riemann manifold. We define
X p
E ε = Ewε : M × M → R, Ewε (p, q) = w(ε λk )Ψk (p)Ψk (q),
k≥0

where (Ψk )k≥1 is an orthonormal basis of L2 (M, g) consisting of eigenfunctions of ∆g .

Fix a point p0 ∈ M and normal coordinates at p0 defined in an open neighborhood O0
of p0 . The restriction of E ε to O0 × O0 can be viewed as a function E ε (x, y) defined in
an open neighborhood of (0, 0) in Rm × Rm . Fix multi-indices α, β ∈ (Z≥0 )m . Then, as
ε ց 0, we have

|α|−|β| Z 
−m−2d(α,β) i
∂xα ∂yβ E ε (x, y)|x=y=0 =ε w(|x|)x α+β 2
dx + O(ε ) , (11.4.89)
(2π)m Rm

where 
|α + β|
d(α, β) := .
2
Moreover, the constant implied by the symbol O(ε) in (11.4.89) uniformly bounded with
respect to p0 .
11.4. SPECTRAL GEOMETRY 603

Proof. Observe that for any smooth f : M → R we have

Z Z   √
√ 1 √
it ∆ 1 t
wε ( ∆)f = cε (t)e
w f dt = b
w eit ∆ f dt. (11.4.90)
2π R 2πε R ε
The Fourier transform w(t)b b
is a Schwartz function so w(t/ε) is really small for t outside
a small interval
√ around 0 and ε sufficiently small. Thus, a good understanding of the
kernel of eit ∆ for t sufficiently small could potentially lead to a good understanding of
√
the Schwartz kernel of wε ( ∆).
Fix a point p0 ∈ M and normal coordinates at p0 defined in a neighborhood O0 of p0 .
Then we can identify a point (p, q) ∈ O0 × O0 with a point (x, y) in a neighborhood of
(0, 0) in Rm × Rm .
Using (11.4.90) we deduce
 E
α β ε 1D α β t
∂x ∂y E (x, y)|x=y = b
∂x ∂y Kt (x, y)|x=y , w . (11.4.91)
ε | {z } ε
=:Ktα,β

Choose an even, nonnegative cutoff function ρ ∈ C0∞ (R) such that

(
1, |t| ≤ 4c ,
ρ(t) =
0, |t| ≥ 2c ,

where c > 0 is the constant in Theorem 11.4.6. Then

 E  E
α β ε 1 D α,β t 1 D α,β
t
∂x ∂y E (x, y)|x=y = b
Kt , ρ(t)w + Kt , 1 − ρ(t) wb .
ε ε ε ε
Let us observe that, for any N > 0, we have
 E
1 D α,β
t
Kt , 1 − ρ(t) w b = O(εN ) as ε → 0
ε ε
Thus
 E
1 D α,β t
∀N > 0 ∂xα ∂yβ E ε (x, y)|x=y ∼ b
Kt , ρ(t)w + O(εN ), ε → 0. (11.4.92)
ε ε
From (11.4.78) we deduce
∞
X
∂xα ∂yβ Kt (x, y) ∼ ∂xα ∂yβ { Uk (x, y)H2k (t, x, y) } . (11.4.93)
k=1

On the other hand, Theorem 11.4.6 implies that

∞
X
∂xα ∂yβ Kt (x, y)|x=y=0 ∼ Am,α,β,k |χ|−m−|α+β|+2k (t), (11.4.94)
k=0

where Am,α,β,0 is a universal constant depending only on m, α, β, which is equal to 0 if

|α + β| is odd.
604 CHAPTER 11. SPECTRAL GEOMETRY

Lemma 11.4.10. (a) For any r ∈ Z and any N > 0 we have

1  
wε = εr |χ|r , w
|χ|r , ρc b + O(εN ) as ε → 0.
ε
(b) For every positive integer r we have
√ 1−r Z
−r π2
h|χ| , wi
b = |τ |r−1 w(τ )dτ.
Γ( 2r ) R

Proof. (a) For transparency we will use the integral notation for the pairing between a
generalized function and a test function. We have
Z Z
r 1 r
|χ| , ηcwε = |χ| (t)ρ(t)w(t/ε)dt
b = |χ|r (εt)ρ(εt)w(t)dt
b
ε R R
Z
r
=ε |χ|r (t)ρ(εt)w(t)dt
b = εr h|χ|r , ρε wi,
b ρε (t) = ρ(εt).
R
b−w
Now observe that ρε w b = w(ρ
b ε − 1) → 0 in S (R). More precisely for k ≥ 0 we have

∂k
(ρε − 1) = O(εN tN ) as ε → 0.
∂k
This implies that
|χ|r , w(ρ
b ε − 1) = O(εN ) as ε → 0,
so that

h|χ|r , ρε wi
b = h|χ|r , wi+
b |χ|r , w(ρ
b ε − 1) = h|χ|r , wi
b + O(εN ) as ε → 0.

(b) We have

[ (11.2.11b ) √ 1−r
h|χ|−r , wi
b = |χ|−r , w = π2 |χ|r−1 (τ ), w(τ )
√ Z
π21−r
= |τ |r−1 w(τ )dτ.
Γ( 2r ) R
⊔
⊓

Using (11.4.92) and the above lemma we deduce that as ε ց 0 we have

 
∂xα ∂yβ E ε (x, y)|x=y = Dm,α,β ε−m−2d(α,β) + O ε−m−2d(α,β)+2 , (11.4.95)

where Dm,α,β is a universal constant that depends only on m, α, β which is = 0 if |α + β|

is odd,
√ 1−r Z
π2
Dm,α,β = Am,α,β,0 |τ |r−1 w(τ )dτ, r = m + 2d(α, β). (11.4.96)
Γ( 2r ) R
11.4. SPECTRAL GEOMETRY 605

To determine the constant Dm,α,β it suffices to compute it for one particular m-

dimensional Riemann manifold. Assume that (M, g) is the torus T m equipped with the
flat metric
Xm
g= (dθ i )2 , 0 ≤ θ i ≤ 2π.
i=1

The eigenvalues of the corresponding Laplacian ∆m are

|~k|2 , ~k = (k1 , . . . , km ) ∈ Zm .

Denote by ≺ the lexicographic order on Zm . For ~θ = (θ 1 , . . . , θ m ) ∈ R and ~k ∈ Zm we set

m
X
h~k, θi
~ := kj θ j .
j=1

A real orthornormal basis of L2 (Tm ) is given by the functions


 ~k = ~0
1 1,
1
Ψ~k (~
θ) = m 2 2 sinh~k, θi,
~ ~k ≻ ~0,
(2π) 2 
 12
2 cosh~k, θi,
~ ~k ≺ ~0.

Then X
1 ~ ~
E ε (~
θ, ϕ
~) = m
w(ε|~k|)eihk,θ−~ϕi ,
(2π)
~k∈Zm

so that
i|α|−|β| X ~ ~
∂θ~α ∂ϕβ~ E ε (~
θ, 0) = m
wε (|~k|)~kα+β eihk,θi .
(2π)
~k∈Zm

Define
Wm , uε : Rm → R, Wm (x) = w(|x|), uε (x) = Wm (εx)xα+β .
Using Poisson’s summation formula (11.4.59) we deduce

i|α|−|β| X
∂θ~α ∂ϕβ~ E ε (0, 0) = bε (2π~ν ).
u
(2π)m m~
ν ∈Z

Observe that
Z Z 
−ihξ,xi α+β α+β −ihξ,xi
bε (ξ) =
u e w(ε|x|)x dx = (i∂ξ ) e Wm (εx)dx
Rm Rm
Z   
−m α+β −ih 1ε ξ,yi −m cm 1
=ε (i∂ξ ) e Wm (y)dy =ε (i∂ξ )α+β W ξ .
Rm ε
Hence   
i|α|−|β| X 1
∂θ~α ∂ϕβ~ E ε (~
θ, 0) = m
(i∂ξ )α+β c
Wm ξ .
(2πε) m
ε ξ=2π~
ν
~
ν ∈Z
606 CHAPTER 11. SPECTRAL GEOMETRY

As ε → 0 we have

i|α|−|β|  
∂θ~α ∂ϕβ~ E ε (0, 0) = ε−m−|α+β| (i∂ξ )α+β c
W m (0) + O(εN
) , ∀N.
(2π)m

Now observe that Z

cm (0) =
(i∂ξ )α+β W w(|x|)xα+β dx.
Rm
so that
Z 
i|α|−|β|
∂θ~α ∂ϕβ~ E ε (0, 0) = ε−m−|α+β| w(|x|)x α+β
dx + O(ε ) N
, ∀N. (11.4.97)
(2π)m Rm

This shows that Z

i|α|−|β|
Dm,α,β = w(|x|)xα+β dx. (11.4.98)
(2π)m Rm
This completes the proof of Theorem 11.4.9. ⊔
⊓

Remark 11.4.11. Note that

Z Z ! Z 
∞
α+β α+β m+|α+β|−1
w(|x|)x dx = x dA(x) w(r)r dr .
Rm |x|=1 0
| {z }
=:Im,α,β (w)

On the other hand, according to Lemma 9.3.10 we have

 Qk α +β +1
Z  2 i=1 Γ( i 2 i ) , α + β ∈ (2Z )m ,
≥0
xα+β dA(x) = Zm,α,β
m+|α+β|
:= Γ( 2
) (11.4.99)
|x|=1 
0, otherwise.

We can now rewrite (11.4.98) as

i|α|−|β|Zm,α,β
Dm,α,β = ε−m−|α+β| Im,α,β (w). (11.4.100)
(2π)m
⊔
⊓

11.4.5 Spectral perestroika

It is time to reap some geometric benefits from our analytic toil. Let (M, g) be a com-
pact, connected, Riemann m-dimensional manifold. Denote by ∆g the associated Laplace
operator.
Definition 11.4.12. The spectral data of (M, g) consist of the following.

1. The spectrum of ∆g
spec(∆g ) ⊂ [0, ∞)
11.4. SPECTRAL GEOMETRY 607

2. The orthogonal projections

P λ : L2 (M, g) → ker(λ − ∆g ), λ ∈ spec(∆g ).

⊔
⊓

To use a terminology dear to the late great Vadimir Arnold (1937–2010), the spectral
perestroika question asks how much of the geometry of the Riemann manifold (M, g) can
be recovered from its spectral data.
Note that the multiplicity n(λ) of λ ∈ spec(∆g ) is computable from the spectral data
as the dimension of the range of P λ . The operator P λ is an integral operator whose kernel
K λ (p, q) is given by
n(λ)
X
K λ (p, q) = Φj (p)Φj (q),
j=1

where (Ψj )1≤j≤n(λ) is any orthonormal basis of ker(λ − ∆g ) consisting of real valued
eigenfunctions. This means that
Z
(P λ u)(p) = K λ (p, q)u(q) |dVg (q)|.
M

As in the previous subsections we describe the spectrum of ∆g as a nondecreasing sequence

0 = λ 0 < λ1 ≤ λ 2 ≤ · · · ,

where each eigenvalue appears exactly as many times as its multiplicity. Fix an orthonor-
mal basis (Ψk )k≥0 of L2 (M ) such that

∆Ψk = λk Ψk , ∀k ≥ 0

Then X
K λ (p, q) = Ψk (p)Ψk (q).
λk =λ

Fix an even smooth function w : R → [0, 1] such that

(
1, |t| ≤ 1/2,
w(t) =
0, |t| > 1.

We can think of w(t) as a smooth approximation of the indicator function of the interval
[−1, 1]; see Figure 11.3.
As in the previous subsection we set
X p
X √

Ewε (p, q) := w ε λk Ψk (p)Ψk (q) = w ε λ K λ (p, q).
k≥0 λ∈spec(∆)

The above sum consists of finitely may terms since it has only contributions from eigen-
values < ε−2 so it has O(ε−m ) terms.
608 CHAPTER 11. SPECTRAL GEOMETRY

Figure 11.3: A smooth approximation of the indicator function of [−1, 1].

Fix a point p ∈ M and normal coordinates (xi ) near p. From (11.4.89) with |α| =
|β| = 1 we deduce
Z 
ε −m−2 1 i j 2
∂xi ∂yj E (x, y)|x=y=0 = ε w(|x|)x x dx + O(ε )
(2π)m Rm
 

δij Z 
−m−2  
=ε  w(|x|)(x1 )2 dx +O(ε2 )  .
(2π)m  Rm 
| {z }
=:Bm (w)

The positive constant Bm (w) can be determined using (11.4.100). More generally, for two
vector fields X1 , X2 ∈ Vect (M ) we have
X
X1 (x)X2 (y)E ε (x, y)|x=y=0 = X1i ∂xi X2j (y)∂yj E ε (x, y)|x=y=0
i,j
1 X  
j
= ε−m−2 δij X i
1 (0)X 2 (0) B m (w) + O(ε2
) (11.4.101)
(2π)m
i,j
1
 
= ε−m−2 gp X 1 (p), X 2 (p) B m (w) + O(ε2
) .
(2π)m
For X1 , X2 ∈ Vect (M ) fixed, the above estimate is uniform in p ∈ M . We set
X
D(ε) := dim ker(λ − ∆g ).
λ<ε−2

According to (11.4.81) we have



D(ε) = volg (M )ωm ε−m 1 + O ε as ε ց 0.

Observe that the sums in (11.4.101) consist of D(ε) terms.

Proposition 11.4.13. For ε > 0 sufficiently small the map F ε : M → RD(ε) ,
s !
Fε εm+2 (2π)m p

p 7−→ w ε λk Ψk (p) ∈ RD(ε) , (11.4.102)
Bm (w) −2λk <ε

is an embedding and the metric induced by F ε converges uniformly to g as ε > 0.

11.4. SPECTRAL GEOMETRY 609

Proof. Denote by • the canonical inner product in RD(ε) . For p ∈ M and X ∈ Tp M we

denote by DX F ε (p) the value of the differential of F ε at p on the tangent vector X(p).
Explicitly,
d

DX F ε (p) = F ε γ(t) ∈ RD(ε) ,
dt t=0
where γ : (−ε, ε) → M is any smooth path such that γ(0) = p and γ̇(0) = X(p).
The estimates (11.4.101) show that, for any X ∈ Vect (M ) and any p ∈ M , we have

 
DX F ε (p) • DX F ε (p) = gp X(p), X(p) 1 + O(ε2 ) as ε ց 0,

uniformly in p ∈ M . This proves that for ε > 0 sufficiently small, the map F ε is an
immersion and the metric gε on M induced by F ε is C 0 -close to g.
To prove that F ε is injective for ε > 0 sufficiently small we argue by contradiction.
Assume that this is not true. This means that for any n ∈ N there exist points pn , q n ∈ M
such that
Ψk (pn ) = Ψk (q n ), ∀k = 1, . . . , n.
After eventually extracting subsequence we can assume that the sequences (pn )n∈N and
(q n )n∈N are convergent
p := lim pn , q := lim q n .
n→∞ n→∞
Note that
Ψk (p) = Ψk (q), ∀k ∈ N. (11.4.103)
The vector space spanned by (Ψk )k∈N is dense in C 0 (M ) so (11.4.103) implies

f (p) = f (q), ∀f ∈ C 0 (M ).

In particular, this means p = q.

Set δn := distg (pn , q n ), so that δn → 0 as n → ∞. For n large there exists a unique
unit vector Vn ∈ Tpn M such that q n = exp(δn Vn ). After passing to subsequences we can
assume that Vn converges to a unit vector V ∈ Tp M .
Since F ε is an immersion for any ε > 0 sufficiently small we deduce that, given such
a small ε0 > 0, there exists k ∈ {1, . . . , D(ε0 )} such that

dΨk (V )p 6= 0. (11.4.104)

Denote by γn the smooth path

γn : [0, δn ] → M, γn (t) = exppn (tVn ).

Note that, for large n, γ n is the unique unit speed minimal geodesic connecting pn to q n .
Applying the mean value theorem to the function


fk (t) = Ψk γn (t) , t ∈ [0, δn ]

we deduce that, ∀n ∈ N, there exists τn ∈ [0, δn ] such that

dΨk γ̇(τn ) = 0. (11.4.105)
610 CHAPTER 11. SPECTRAL GEOMETRY

Note that as n → ∞ we have

γn (τn ) → p and γ̇(τn ) → V.

Hence, letting n → ∞ in (11.4.105) we deduce

dΨk (V )p = 0.

This contradicts (11.4.104) and completes the proof of Proposition 11.4.13. ⊔

⊓

Proposition 11.4.13 shows that the function

εm+2 (2π)m ε
K ε = Kwε : M × M → R, K ε (p, q) = Ew (p, q) (11.4.106)
Bm (w)

is the correlator associated to the embedding F ε : M → RD(ε) ; see Example 4.2.27.

Moreover, as ε > 0, the metric gε determined by this correlator converges to the original
metric in C 0 (M ). We state this conclusion as a corollary first observed by S. Zelditch
[136].
Corollary 11.4.14. Using the spectral data we can produce arbitrarily accurate C 0 -
approximations of the original metric. ⊔
⊓

From the above corollary we deduce that the volume measure dVgε converges weakly
to the volume measure dVg . We can say a bit more. Fix g-normal coordinates x near a
point p ∈ M . In these coordinates the metric gε is described by the matrix
ε
gij = δij + O(ε2 ). (11.4.107)

Using (11.4.89) with |α| = 1 and |β| = 2 we deduce

Z 
∂xα ∂yβ K ε (x, y) = const. w(x)xα+β dx + O(ε2 ) = O(ε2 ).
Rm

We denote by ∇ε the Levi–Civita connection of gε and by ∇ the Levi–Civita connection

of the original metric g. Their difference is a tensor


Aε = ∇ε − ∇ ∈ Ω1 End(T M ) .

Denote by Γε the Christoffel symbols of ∇ε in the coordinates x and by Γ the Christoffel

symbols of ∇ in the same coordinates. Since the coordinates where chosen normal at p
we have Γ(0) = 0 and deduce

Tp∗ M ⊗ End(Tp M ) ∋ Aε (p) = Γε (0) − Γ(0) = Γε (0).

The formula (4.2.23) describing the Christoffel symbols in terms of correlators coupled
with (11.4.107) implies that
kAε (p)k = O(ε2 )
11.4. SPECTRAL GEOMETRY 611

and this estimate is uniform in p. Hence

kAε kC 0 = O(ε2 ). (11.4.108)

The continuous dependence on data of solutions of ordinary differential equations [12]

implies that the geodesics of gε C 1 -approximate the geodesics of g
Can we determine more geometric features of (M, g) from the geometry of (M, gε ) from
what we know so far? The next simple example shows that, given what we know at this
point, we cannot expect that the curvature of gε approximates the curvature of g.

Example 11.4.15. Let us assume that dim M = 2 and in the neighborhood of a point
the metric g is Euclidean. Suppose that on the same neighborhood gε is conformal to g
p
gε = efε (r) g, r = x2 + y 2 , fε ∈ C ∞ (R).

According to (10.3.16) the scalar curvature of sε is

 
d2 1 d
sε = e−fε ∆g fε = −e−fε f ε − f ε .
dr 2 r dr

Consider the function w in Figure 11.3. We set

Z 1
1
I := w(t)dt ≥ .
0 2

For t ∈ [−1, 1] set

Z t Z t/ε Z t
gε (t) := w(s/ε)ds = ε w(τ )dτ, fε (t) = gε (s)ds.
0 0 0

We deduce Z Z
ε t/ε
fε (ε) = ε dt w(ds)ds ≤ Iε2 ,
0 0
Z r/ε Z r/ε
ε 1
∆g fε = −w(r/ε) + w(s)ds = w(s)ds − w(r/ε) ≥ 0.
r 0 r/ε 0

Observe that the scalar curvature function sε is radially symmetric and

2

sε |r=ε = e−fε (ε) I ≥ e−Iε I ≥ 1 − Iε2 I.

This shows that the scalar curvature function sε does not converge uniformly to 0, the
scalar curvature of the flat Euclidean metric.
As a curiosity, note that
ε
sε (r) = 0, ∀r ≤ .
2
This shows that the metric gε is flat on the disk of radius ε/2 centered at the origin, but
becomes very curved along tiny circle of radius ε centered at the origin. ⊔
⊓
612 CHAPTER 11. SPECTRAL GEOMETRY

We want to investigate the behavior of the Riemann curvature tensor of the metric
gε . Fix a point p and g-normal coordinates (xi ) at p. Then, the local frame ∂xi is nearly
gε -normal in the following sense
(11.4.107 )
gε (∂xi , ∂xj )p = δij + O(ε)2 ,
(11.4.108 )
∇ε∂ i ∂xj p
= O(ε2 ).
x

Denote by Rε the Riemann curvature tensor of gε . Using the above estimates and (4.2.24)
we deduce

Rkℓij (p) = ∂x4k xi yℓ yj K ε (x, y) x=y=0

− ∂x4k xj yℓ yi K(x, y) x=y=0
+ O(ε2 ).

Theorem 11.4.16. Fix a point p ∈ M and normal coordinates (xi ) near p. For any
multi-induces α, β ∈ (Z≥0 )m we set
ε
Eα;β := ∂xα ∂yβ E ε (x, y)|x=y=0 .

Then there exists a universal constant Zm that depends only on the dimension of M and
the weight w such that
ε ε


Eii;jj − Eij;ij = Zm Rijij (p)ε−m−2 1 + O(ε2 ) as ε → 0. (11.4.109)

Proof. Using (11.4.91) we deduce

*  +
1 t
ε
Eii;jj ε
− Eij;ij ∼ Ktii,jj − Ktij,ij , η(t)w
b + O(εN ), ε → 0 (11.4.110)
ε ε

On the other hand from (11.4.93) we conclude

∞
X  
Ktii,jj − Ktij,ij ∼ dm (2k) ∂x2i ∂y2j − ∂x2i xj ∂y2i yj Uk (x, y)Hk (t, x, y) |x=y=0 ,
k=1
(11.4.111)

where we recall that

η(x, y) := dist(x, y)2 ,

a− n
a− n

Ha (t, x, y) := ∂t χ+ 2 t2+ − η(x, y) − χ+ 2 t2− − η(x, y) ,



2k 2k
= ∂t R+ t, sy (x) − R+ −t, sy (x) .
Let us introduce some notation.

• For a positive integer k we denote by ∂ k a generic mixed-partial derivative of order

k in the space variables xi , y j . We will write ∂ instead of ∂ 1 .

• We denote by ∂ k η the collection of k-th order derivatives of η(x, y).

11.4. SPECTRAL GEOMETRY 613

• Pi (X) will denote a homogeneous polynomial of degree i in the variables X.

• Pk (X)Pℓ (Y ) will denote a polynomial which is homogeneous of degree k in the

variables X and homogeneous of degree ℓ in the variables Y .

We then have the equalities

∂Ha = P1 (∂η)Ha−1 , (11.4.112)

∂ 2 Ha = P2 (∂η)Ha−2 + P1 (∂ 2 η)Ha−1 , (11.4.113)
3 2 3
∂ Ha = P3 (∂η)Ha−3 + P1 (∂η)P1 (∂ η)Ha−2 + P1 (∂ η)Ha−1 , (11.4.114)


∂ 4 Ha = P4 (∂η)Ha−4 + P2 (∂η)P1 (∂ 2 η) Ha−3

(11.4.115)
+ P2 (∂ 2 η) + P1 (∂η)P1 (∂ 3 η) Ha−2 + P1 (∂ 4 η)Ha−1 .
To simplify the presentation we will assume that in (11.4.109) we have i = 1, j =
2. Also, we will denote by O(1) a function f (x, y) such that f (x, y)|x=y=0 = 0. The
computations in Section 5.2.3 show that

Pj (∂η) = Pk (∂ 3 η) = O(1). (11.4.116)

In particular, the above equalities show that the 1st and 3rd order derivatives of Ha are
O(1). We have


∂x21 ∂y22 (Uk Hk ) = ∂x21 ∂y22 Uk Hk + 2∂y2 Uk ∂y2 Hk + Uk ∂y22 Hk




= ∂x21 ∂y22 Uk Hk + ∂y22 Uk ∂x21 Hk ) + ∂x21 Uk ∂y22 Hk ) (11.4.117)
+4(∂x21 y2 Uk )(∂x21 y2 Hk ) + Uk ∂x21 ∂y22 Hk + O(1),

∂x21 x2 ∂y21 y2 (Uk Hk )



= ∂x21 x2 ∂y21 y2 Uk Hk + ∂y1 Uk ∂y2 Hk + ∂y2 Uk ∂y1 Hk + Uk ∂y21 y2 Hk




= ∂x21 x2 ∂y21 y2 Uk Hk + ∂y21 y2 Uk ∂x21 x2 Hk
+ ∂x22 y1 Uk ∂x21 y2 Hk + ∂x21 y1 Uk ∂x22 y2 Hk + ∂x22 y2 Uk ∂x1 y1 Hk + ∂x21 y2 Uk ∂x22 y1 Hk

+ ∂x21 x2 Uk ∂y21 y2 Hk + Uk ∂x21 x2 ∂y21 y2 Hk + O(1). (11.4.118)

Using (11.4.112)-(11.4.115) we deduce that

  4
X
∂x21 ∂y22 − ∂x21 x2 ∂y21 y2 (Uk Hk )x=y=0 = Tkj Hk−j |x=y=0 ,
j=0

where the coefficients Tkj are degree j polynomials in the derivatives of Uk and η at the
point (x, y) = (0, 0). Using (11.4.116) we deduce

Tk4 = Tk3 = 0 .
614 CHAPTER 11. SPECTRAL GEOMETRY

Moreover, in view to (11.4.116), the terms in Tk2 are due only to the 4-th order derivatives of
Hk . Upon inspecting (11.4.117) and (11.4.118) (pay attention to the boxed terms in these
equalities) we see that the 4th order derivatives of Hk are multiplied by Uk . According to
(U k,h ) the function Uk is O(1) if k > 1. Hence Tk2 = 0 for k > 1. We deduce
∞
X

Ktii,jj − Ktij,ij ∼ dm (2k) Tk0 Hk + Tk1 Hk−1 + Tk2 Hk−2 |x=y=0
k=1

= B−1 H−1 |x=y=0 + B0 H0 |x=y=0 + B1 H1 |x=y=0 + · · · ,

where
B−1 = dm (2)T12 , B0 = dm (2)T11 , B1 = dm (2)T10 + dm (4)T21 , . . . .
The term B−1 can be alternatively described as

B−1 = Am,ii;jj,0 − Am,ij;ij,0 ,

where the coefficients Am,α,β,0 are defined as in (11.4.94). Using (11.4.96) and (11.4.98)
we deduce
B−1 = 0.
To compute T11 we observe first that
X
η(x − y) = (xi − y i )2 + higer order terms. (11.4.119)
i

Using (11.4.113) we can simplify (11.4.117) and (11.4.118) in the case k = 1 as follows.


∂x21 ∂y22 (U1 H1 ) = ∂x21 ∂y22 U1 H1 + U1 ∂x21 ∂y22 H1 + O(1), (11.4.120)


∂x21 x2 ∂y21 y2 (U1 H1 ) = ∂x21 x2 ∂y21 y2 U1 H1 + ∂x21 y1 U1 ∂x22 y2 H1
(11.4.121)
+∂x22 y2 U1 ∂x1 y1 H1 + U1 ∂x21 x2 ∂y21 y2 H1 + O(1).
Using (11.4.113), (11.4.115) and (11.4.119) we deduce that
 
T11 = ∂x21 ∂y22 − ∂x21 x2 ∂y21 y2 η|(0,0)
   
+2 ∂x21 U1 + ∂y22 U1 |(0,0) + 2 ∂x21 y1 U1 + ∂x22 y2 U1 |(0,0) .

Using the transport equation (U 1,h ) with h given by (11.3.47) we deduce that

U1 (x, y) = U1 (y, x), U1 (x, x) ≡ 1.

These conditions imply that the Hessian of U1 (x, y) at (0, 0) is a quadratic form in the
variables ui = (xi − y i ) so that

∂x21 U1 (0, 0) + ∂x21 y1 U1 (0, 0) = ∂y22 U1 (0, 0) + ∂x22 y2 U1 (0, 0) = 0. (11.4.122)

Hence  
T11 = ∂x21 ∂y22 − ∂x21 x2 ∂y21 y2 η|(0,0) .
11.4. SPECTRAL GEOMETRY 615

Using (5.2.24) we conclude that

T11 = ZR1212 ,
where Z is a universal constant, independent of (M, g). Hence
X
Ktii,jj − Ktij,ij ∼ dm (2)ZR1212 (p)H0 |x=y=0 + Bk Hk |x=y=0 .
k≥1

Using (11.4.110) we deduce

*  +
ε ε dm (2)ZR 1212 (p) t
Eii;jj − Eij;ij ∼ H0 |x=y=0 , η(t)w
b
ε ε
*   +
1 X t
+ Bk Hk |x=y=0 , η(t)wb + O(εN ).
ε ε
k≥1

The equality (11.4.109) now follows from the above equality by invoking (11.3.26), the
definition (11.4.77) of Hk and Lemma 11.4.10. ⊔
⊓

We can now tie all the loose ends together. Fix a point p ∈ M and g-normal coordinates
(xi ) at p. We denote by Rε the Riemann curvature of the metric gε determined by the
correlator (11.4.106) and by Kij ε the sectional curvature of the metric g along the plane
ε
spanned by ∂xi , ∂xj ,
ε (p)
Rijji
ε
Kij (p) = .
areagε (∂xi , ∂xj )
From Proposition 11.4.13 and Theorem 11.4.16 we deduce that there exists a constant
L = Lm (w) that depends on the dimension m and the function w, but it is independent
of the manifold M and the metric g such that
ε
lim Kij (p) = Lm Kij (p).
εց0

If there existed some manifold M and some metric g on M the convergence of gε to g is

C 2 , then would could conclude that Lm = 1 since the curvature is a continuous function
in gij and its derivatives of order ≤ 2.
Suppose now that M = S m is the unit sphere in Rm+1 and g is the induced metric. The
metric is SO(m+1)-invariant and so is the associated Laplacian operator. In particular, the
eigenspaces are SO(m + 1)-invariants. The orthogonal projections onto these eigenspaces
are SO(m + 1)-equivariant operators. From the description
X √

Ewε (p, q) = w ε λ K λ (p, q).
λ∈spec(∆)

deduce that the correlator K ε (p, q) is invariant with respect to the diagonal action of
SO(m + 1) on S m × S m . This shows that the metric gε is SO(m + 1)-invariant. Since
the action of SO(m + 1) on S m is transitive we deduce that there exists a constant cε > 0
such that
gε = cε g.
616 CHAPTER 11. SPECTRAL GEOMETRY

Since gε converges uniformly to g we deduce that cε → 1. In particular, we deduce that

the curvature of gε converges to that of g so Lm (w) = 1. The next result summarizes the
facts outlined above.
Corollary 11.4.17. The spectral data determine the sectional curvature of the metric.
More precisely, the sectional curvatures of the correlator metric gε converge uniformly to
the sectional curvatures of the original metric g. ⊔
⊓

We can say a bit more. We want to show that the spectral data determine the
whole Riemann curvature tensor of g. Fix a point p ∈ M and denote by Gr2 (Tp M )
the Grassmannian of 2-planes in Tp M and by S+ 2 (T M ) the cone of inner products on
p
Tp M . For any Riemann metric h on M the sectional curvature at p is a continuous
function Kph : Gr2 (Tp M ) → R.
There exists a continuous function


R : C 0 Gr2 (Tp M ) × S+ 2
(Tp M ) → Λ2 Tp∗ M ⊗ Λ2 Tp∗ M, (K, h) 7→ R(K, h),

described by an explicit formula6 in [28, Eq. (1.12)], such that, for any Riemann metric g
on M , the quantity
R Kpg , gp ) ∈ Λ2 Tp∗ M ⊗ Λ2 Tp∗ M
coincides with the Riemann curvature tensor of g at p.
The sectional curvatures Kpε of the metric gε at p converges to the sectional curvature
g p p ε
p and gε → g . Denote by Rp the Riemann curvature of gε at p. We deduce that

Rpε = R Kpε , gεp ) → R Kp , g p ) = Rp .

Corollary 11.4.18. The spectral data determine the Riemann curvature of the original
metric g. ⊔
⊓

Remark 11.4.19. (a) Using the very deep results in [5, 109] one can show that the C 0 -
convergence of the metrics gε coupled with the convergence of sectional curvatures imply
that gε converges to g in C ∞ . We can rephrase this by saying that, for ε > 0 sufficiently
small, the map F ε : M → RD(ε) defined in (11.4.102) is a nearly isometric embedding.
(b) Let us observe that for any even, nonnegative temperate function w ∈ S (R) we can
define a map
s
Fw X εm+2 (2π)m p

M ∋ p 7−→ε
w ε λk Ψk (p)Ψk ∈ L2 (M, g)
Bm (w)
k≥0

When w has compact support, the range of this map is contained in a finite dimensional
subspace.
2
The case w(r) = e−r was first investigated in [17] where, among many other things,
the authors show that for small ε > 0 the map F w
ε is a nearly isometric embedding of M
in the Hilbert space L2 (M, g). The image of this embedding is not be contained in any
finite dimensional subspace.
6
The explicit formula of R is rather complicated, but it is polynomial in K and h so it is continuous.
11.4. SPECTRAL GEOMETRY 617

Observe that in this case, if we set t := ε2 we have

√ X
Ew t (p, q) = Ht (x, y) := e−tλk Ψk (p)Ψk (q). (11.4.123)
k≥0

This is so called heat kernel of the Riemann manifold.

(c) Let us point out that the above results solve more than the original spectral perestroika
problem.

The geometry of the compact Riemann manifold (M, g) is completely de-

termined by the associated heat kernel. More precisely, the metric and its
curvature are determined by the behavior as t ց 0 of the jets of order 4 along
the diagonal of the heat kernel Ht (x, y).

This is a rather unexpected conclusion: we can learn the “shape” of the manifold by
investigating how the heat propagates or, equivalently, how the diffusion takes place. ⊔
⊓

Remark 11.4.20. (a) As explained in Remark 4.2.30, the correlators associated to em-
beddings have probabilistic interpretations. For example, when w is as in Figure 11.3, the
correlator Ewε (p, q) is associated to random functions on M of the form
X
Uε := Xkε Ψk ,
k≥0

where (Xkε )k≥ are independent random variables with mean zero and variances

εm+2 (2π)m p
variance (Xkε ) = w(ε λk ).
Bm (w)

As ε ց 0 this random function converges in a precise sense to the “perfect chaos”, the
so called called Gaussian white-noise. A plastic way of phrasing the conclusions of the
previous results is that there is a pattern in devolution to chaos which is very sensitive to
the background geometry.
(b) For every p ∈ M we have a random symmetric bilinear form

Gε := dU ε (p) ⊗ dU ε (p) : Tp M × Tp M → R.

For ε > 0 small, its expectation is the metric gε induced by the embedding F ε defined
in (11.4.102). We are inclined to believe that the random variable Gε converges almost
surely and L1 to the (deterministic) g. In other words Gε is highly concentrated around
its mean as ε ց 0.
(c) Here is another little gem hidden under the probability cloak. Assume M oriented
and even dimensional m = dim M = 2h. The differential of the random function Uε is a
random section dU ε of the cotangent bundle. One can show that, with probability 1, dUε
intersects the zero section of T ∗ M transversally, i.e., dUε has nondegenerate zeros. To
618 CHAPTER 11. SPECTRAL GEOMETRY

such a nondegenerate zero p we can associate a local index ǫ(Uε , p) = ±1 (see Exercise
7.3.47) and from the Poincaré-Hopf formula in Corollary 7.3.48 we deduce that
X
ǫ(Uε , p) = χ(M ). (11.4.124)
dUε (p)=0

In fact, above we have produced quite a bit more, namely a 0-dimensional current
X
ZdUε = ǫ(Uε , p)δp ∈ Ω0 (M ).
dUε (p)=0

The ”integral” of a given 0-form f over this random current is the random variable
X
f, ZdUε := ǫ(Uε , p)f (p).
dUε (p)=0

The expectation of this random variable has a rather striking geometric description [106].
More precisely we have Z
 
E f, ZdUε = f e(∇gε ), (11.4.125)
M
where e(∇gε ) is the Euler form associated to the Levi–Civita connection of the metric gε ,
1

e(∇gε ) = Pf − Rε ,
(2π)m/2

where Rε is the Riemann curvature tensor of the metric gε . Thus

Z
  1 ε


E f, ZdUε = f P f − R , (11.4.126)
(2π)m/2 M

Note that when f = 1, then

X (11.4.124 )
1, ZdUε = ǫ(Uε , p) = χ(M ).
dUε (p)=0

This is no longer a random quantity. Using this in (11.4.126) we deduce

Z
1

χ(M ) = m/2
P f − Rε .
(2π) M

This is the Gauss–Bonnet theorem for the metric gε . On the other hand, we know that Rε
converges uniformly to the Riemann curvature R of the original metric g. If we let ε ց 0
we obtain the Gauss–Bonnet theorem for the original metric g!
This “accident” is a manifestation of a more general phenomenon. In particular, the
general Gauss–Bonnet theorem has a probabilistic interpretation. For details we refer to
[105, 106]. ⊔
⊓
Chapter 12

Dirac Operators

We devote this last chapter to a presentation of a very important class of first order
elliptic operators which have numerous applications in modern geometry. We will first
describe their general features, and then we will spend the remaining part discussing some
frequently encountered examples.

12.1 The structure of Dirac operators

12.1.1 Basic definitions and examples
Consider a Riemann manifold (M, g), and a smooth vector bundle E → M .
Definition 12.1.1. A Dirac operator is a first order p.d.o.

D : C ∞ (E) → C ∞ (E),

such that D 2 is a generalized Laplacian, i.e.,

σ(D 2 )(x, ξ) = −|ξ|2g 1Ex ∀(x, ξ) ∈ T ∗ M.

The Dirac operator is said to be graded if E splits as E = E0 ⊕ E1 , and D(C ∞ (Ei )) ⊂

C ∞ (E(i+1) mod 2 ). In other words, D has a block decomposition
 
0 A
D= . ⊔
⊓
B 0

Note that the Dirac operators are first order elliptic p.d.o.-s.
Example 12.1.2. (Hamilton–Floer). Denote by E the trivial vector bundle R2n over
the circle S 1 . Thus C ∞ (E) can be identified with the space of smooth functions

u : S 1 → R2n .

Let J : C ∞ (E) → C ∞ (E) denote the endomorphism of E which has the block decompo-
sition  
0 −1Rn
J=
1Rn 0

619
620 CHAPTER 12. DIRAC OPERATORS

with respect to the natural splitting R2n = Rn ⊕ Rn . We define the Hamilton–Floer

operator
du
F : C ∞ (E) → C ∞ (E), Fu = J ∀u ∈ C ∞ (E).
dθ
d2
Clearly, F2 = − dθ 2 is a generalized Laplacian. ⊔
⊓

Example 12.1.3. (Cauchy–Riemann). Consider the trivial bundle C2 over the com-
plex plane C equipped with the standard Euclidean metric. The Cauchy–Riemann oper-
ator is the first order p.d.o. D : C2 → C2 defined by
     
u 0 −∂z u
7−→ 2 · ,
v ∂z̄ 0 v

where z = x + yi, and

1 1
∂z = (∂x − i∂y ), ∂z̄ = (∂x + i∂y ).
2 2
A simple computation shows that D is a Dirac operator. ⊔
⊓

Example 12.1.4. Suppose E is a vector bundle over the Riemann manifold (M, g), and
D : C ∞ (M ) → C ∞ (M ) is a Dirac operator. Denote by M̂ the cylinder R × M , and by ĝ
the cylindrical metric ĝ = dt2 + g on M̂ , where t denotes the coordinate along the factor
R.
We have a natural projection π : M̂ → M , (t, x) 7→ x, and we set Ê := π ∗ E, the
pullback of E to M̂ via π. A section û of Ê is then a smooth 1-parameter family of
sections u(t) of E, t ∈ R. In particular, we can define unambiguously
1

∂t û(t0 , x0 ) = lim u(t0 + h, x0 ) − u(t0 , x0 ) ,
h→0 h

where the above limit is in the t-independent vector space Ex0 . Now define

D̂ : C ∞ (Ê ⊕ Ê) → C ∞ (Ê ⊕ Ê),

by     
u(t) 0 −∂t + D u(t)
= .
v(t) ∂t + D 0 v(t)
Then D̂ is a Dirac operator called the suspension of D. Note that the Cauchy–Riemann
operator is the suspension of the Hamilton–Floer operator. ⊔
⊓

Hamilton–Floer
Example 12.1.5. (Hodge–DeRham). Let (M, g) be an oriented Riemann manifold.
Then, according to the computations in the previous chapter, the Hodge–DeRham oper-
ator
d + d∗ : Ω• (M ) → Ω• (M ).
is a Dirac operator. ⊔
⊓
12.1. THE STRUCTURE OF DIRAC OPERATORS 621

Let D : C ∞ (E) → C ∞ (E) be a Dirac operator over the oriented Riemann manifold
(M, g). Its symbol is an endomorphism σ(D) : π ∗ E → π ∗ E, where π : T ∗ M → M denotes
the natural projection. Thus, for any x ∈ M , and any ξ ∈ Tx∗ M , the operator

c(ξ) := σ(D)(ξ, x)

is an endomorphism of Ex depending linearly upon ξ. Since D 2 is a generalized Laplacian

we deduce that
c(ξ)2 = σ(D 2 )(x, ξ) = −|ξ|2g 1Ex .
To summarize, we see that each Dirac operator induces a bundle morphism

c : T ∗ M ⊗ E → E (ξ, e) 7→ c(ξ)e,

such that c(ξ)2 = −|ξ|2 . From the equality

c(ξ + η)2 = −|ξ + η|2 ∀ξ, η ∈ Tx∗ M, x ∈ M

we conclude that
{c(ξ), c(η)} = −2g(ξ, η)1Ex ,
where for any linear operators A, B we denoted by {A, B} their anticommutator

{A, B} := AB + BA.

Definition 12.1.6. (a) A Clifford structure on a vector bundle E over a Riemann man-
ifold (M, g) is a smooth bundle morphism c : T ∗ M ⊗ E → E, such that

{c(ξ), c(η)} = −2g(ξ, η)1E , ∀ξ, η ∈ Ω1 (M ),

where for every 1-form α we denoted by c(α) the bundle morphism c(α) : E → E given
by
c(α)u = c(α, u), ∀u ∈ C ∞ (E).
The morphism c is usually called the Clifford multiplication of the (Clifford) structure. A
pair (vector bundle, Clifford structure) is called a Clifford bundle.
(b) A Z2 -grading of a Clifford bundle E → M is a splitting E = E + ⊕ E − such that,
∀α ∈ Ω1 (M ), the Clifford multiplication by α is an odd endomorphism of the superspace
C ∞ (E+ ) ⊕ C ∞ (E− ), i.e., c(α)C ∞ (E ± ) ⊂ C ∞ (E ∓ ). ⊔
⊓

Proposition 12.1.7. Let E → M be a smooth vector bundle over the Riemann manifold
(M, g). Then the following conditions are equivalent.
(a) There exists a Dirac operator D : C ∞ (E) → C ∞ (E).
(b) The bundle E admits a Clifford structure.

Proof. We have just seen that (a)⇒(b). To prove the reverse implication let

c : T ∗M ⊗ E → E

be a Clifford multiplication. Then, for every connection

∇ : C ∞ (E) → C ∞ (T ∗ M ⊗ E),
622 CHAPTER 12. DIRAC OPERATORS

the composition
∇ c
D = c ◦ ∇ : C ∞ (E) → C ∞ (T ∗ M ⊗ E) → C ∞ (E)

is a first order p.d.o. with symbol c. Clearly D is a Dirac operator. ⊔

⊓

Example 12.1.8. Let (M, g) be a Riemann manifold. For each x ∈ M , and ξ ∈ Tx∗ M
define
c(ξ) : Λ• Tx∗ M → Λ• Tx∗ M
by
c(ξ)ω = (eξ − iξ )ω,
where eξ denotes the (left) exterior multiplication by ξ, while iξ denotes the interior
differentiation along ξ ∗ ∈ Tx M - the metric dual of ξ. The Exercise 2.2.55 of Section 2.2.4
shows that c defines a Clifford multiplication on Λ• T ∗ M . If ∇ denotes the Levi–Civita
connection on Λ• T ∗ M , then the Dirac operator c ◦ ∇ is none other than the Hodge–
DeRham operator. ⊔
⊓

Exercise 12.1.9. Prove the last assertion in the above example. ⊔

⊓

The above proposition reduces the problem of describing which vector bundles admit
Dirac operators to an algebraic-topological one: find the bundles admitting a Clifford
structure. In the following subsections we will address precisely this issue.

12.1.2 Clifford algebras

The first thing we want to understand is the object called Clifford multiplication.
Consider a real finite dimensional, Euclidean space (V, g). A Clifford multiplication
associated with (V, g) is then a pair (E, ρ), where E is a K-vector space, and ρ : V →
End (E) is an R-linear map such that

{ρ(u), ρ(v)} = −2g(u, v)1E , ∀u, v ∈ V.

If (ei ) is an orthonormal basis of V , then ρ is completely determined by the linear operators

ρi = ρ(ei ) which satisfy the anti-commutation rules

{ρi , ρj } = −2δij 1E .

The collection (ρi ) generates an associative subalgebra in End (E), and it is natural to try
to understand its structure. We will look at the following universal situation.
Definition 12.1.10. Let V be a real, finite dimensional vector space, and

q :V ×V →R

a symmetric bilinear form. The Clifford algebra Cl(V, q) is the associative R–algebra with
unit, generated by V , and subject to the relations

{u, v} = uv + vu = −2q(u, v) · 1 ∀u, v ∈ V. ⊔

⊓
12.1. THE STRUCTURE OF DIRAC OPERATORS 623

Proposition 12.1.11. The Clifford algebra Cl(V, q) exists, and is uniquely defined by
its universality property: for every linear map  : V → A such that A is an associative
R-algebra with unit, and {(u), (v)} = −2q(u, v) · 1, there exists an unique morphism of
algebras Φ : Cl(V, q) → A such that the diagram below is commutative.

[ w Cl(V, q)
ı
V
[[
 ℄[ uΦ
A

ı denotes the natural inclusion V ֒→ Cl(V, q).

Sketch of proof Let A = ⊕k≥0 V ⊗k (V ⊗0 = R) denote the free associative R–algebra
with unit generated by V . Set
Cl(V, q) = A/I,
where I is the ideal generated by

u ⊗ v + v ⊗ u + 2q(u, v) ⊗ 1; u, v ∈ V .

The map ı is the composition V ֒→ A → Cl(V, q) where the second arrow is the natural
projection. We let the reader check the universality property. ⊔
⊓

Exercise 12.1.12. Prove the universality property. ⊔

⊓

Remark 12.1.13. (a) When q ≡ 0 then Cl(V, 0) is the exterior algebra Λ• V .

(b) In the sequel the inclusion V ֒→ Cl(V, q) will be thought of as being part of the
definition of a Clifford algebra. This makes a Clifford algebra a structure richer than
merely an abstract R–algebra: it is an algebra with a distinguished real subspace. Thus
when thinking of automorphisms of this structure one should really concentrate only on
those automorphisms of R–algebras preserving the distinguished subspace. ⊔
⊓

Corollary 12.1.14. Let (Vi , qi ) (i = 1, 2) be two real, finite dimensional vector spaces
endowed with quadratic forms qi : V → R. Then, any linear map T : V1 → V2 such
that q2 (T v) = q1 (v), ∀v ∈ V1 induces a unique morphism of algebras T# : Cl(V1 , q1 ) →
Cl(V2 , q2 ) such that T# (V1 ) ⊂ V2 , where we view Vi as a linear subspace in Cl(Vi , qi ).
The correspondence T 7→ T# constructed above is functorial, i.e., (1Vi )# = 1Cl(Vi ,qi ) ,
and (S ◦ T )# = S# ◦ T# , for all admissible S, and T .
The above corollary shows that the algebra Cl(V, q) depends only on the isomorphism
class of the pair (V, q)= vector space + quadratic form. It is known from linear algebra
that the isomorphism classes of such pairs are classified by some simple invariants:

(dim V, rank q, sign q).

We will be interested in the special case when dim V = rank q = sign q = n, i.e., when q
is an Euclidean metric on the n-dimensional space V . In this case, the Clifford algebra
Cl(V, q) is usually denoted by Cl(V ), or Cln . If (ei ) is an orthonormal basis of V , then we
624 CHAPTER 12. DIRAC OPERATORS

can alternatively describe Cln as the associative R–algebra with 1 generated by (ei ), and
subject to the relations
ei ej + ej ei = −2δij .
Using the universality property of Cln , we deduce that the map

V → Cl(V ), v 7→ −v ∈ Cl(V )

extends to an automorphism of algebras α : Cl(V ) → Cl(V ). Note that α is involutive,

i.e., α2 = 1. Set
Cl0 (V ) := ker(α − 1), Cl1 (V ) = ker(α + 1).
Note that Cl(V ) = Cl0 (V ) ⊕ Cl1 (V ), and moreover

Clε (V ) · Clη (V ) ⊂ Cl(ε+η) mod 2 (V ),

i.e., the automorphism α naturally defines a Z2 -grading of Cl(V ). In other words, the
Clifford algebra Cl(V ) is naturally a super-algebra.
Let (C̃l(V ), +, ∗) denote the opposite algebra of Cl(V ). Then C̃l(V ) coincides with
Cl(V ) as a vector space, but its multiplication ∗ is defined by

x ∗ y := y · x ∀x, y ∈ C̃l(V ),

where “·” denotes the usual multiplication in Cl(V ). Note that for any u, v ∈ V

u · v + v · u = u ∗ v + v ∗ u,

so that, using the universality property of Clifford algebras, we conclude that the natural
injection V ֒→ C̃l(V ) extends to a morphism of algebras Cl(V ) → C̃l(V ). This may as
well be regarded as an antimorphism Cl(V ) → Cl(V ) which we call the transposition map,
x 7→ x♭ . Note that
(u1 · u2 · · · ur )♭ = ur · · · u1 , ∀ui ∈ V.
For x ∈ Cl(V ) we set
x† := (α(x))♭ = α(x♭ ).
The element x† is called the adjoint of x.
For each v ∈ V define c(v) ∈ End (Λ• V ) by

c(v)ω := (ev − iv )ω ∀ω ∈ Λ• V,

where as usual ev denotes the (left) exterior multiplication by v, while iv denotes the
interior derivative along the metric dual of v. Invoking again the Exercise 2.2.55 we
deduce
c(v)2 = −|v|2g ,
so that, by the universality property of the Clifford algebras, the map c extends to a
morphism of algebras c : Cl(V ) → End (Λ• V ).
12.1. THE STRUCTURE OF DIRAC OPERATORS 625

Exercise 12.1.15. Prove that ∀x ∈ Cl(V ) we have

c(x† ) = c(x)∗ ,

where the ∗ in the right-hand-side denotes the adjoint of c(x) viewed as a linear operator
on the linear space Λ• V endowed with the metric induced by the metric on V . ⊔
⊓

For each x ∈ Cl(V ), we set σ(x) := c(x)1 ∈ Λ• V . This is an element of Λ• V , called

the symbol of x. The resulting linear map

Cl(V ) ∋ x 7→ σ(x) ∈ Λ• V,

is called the symbol map. If (ei ) is an orthonormal basis, then

σ(ei1 · · · eik ) = ei1 ∧ · · · ∧ eik ∀eij .

This shows that the symbol map is bijective since the ordered monomials

{ei1 · · · eik ; 1 ≤ i1 , · · · ik ≤ dim V }

form a basis of Cl(V ). The inverse of the symbol map is called the quantization map and
is denoted by q : Λ• V → Cl(V ).

Exercise 12.1.16. Show that q(Λeven/odd V ) = Cleven/odd (V ). ⊔

⊓

Definition 12.1.17. (a) A K(=R-,C)-vector space E is said to be a K- Clifford module

if there exists a morphism of R–algebras

ρ : Cl(V ) → EndK (E).

(b) A K-superspace E is said to be a K-Clifford s-module if there exists a morphism of

s-algebras
d K (E).
ρ : Cl(V ) → End
(c) Let (E, ρ) be a K-Clifford module, ρ : Cl(V ) → EndK (E). The module (E, ρ) is said
to be self-adjoint if there exists a metric on E (Euclidean if K = R, Hermitian if K = C)
such that
ρ(x† ) = ρ(x)∗ ∀x ∈ Cl(V ). ⊔
⊓

We now see that what we originally called a Clifford structure is precisely a Clifford
module.

Example 12.1.18. Λ• V is a self-adjoint, real Cl(V ) super-module. ⊔

⊓

In the following two subsections we intend to describe the complex Clifford modules.
The real theory is far more elaborate. For more information we refer the reader to the
excellent monograph [83].
626 CHAPTER 12. DIRAC OPERATORS

12.1.3 Clifford modules: the even case

In studying complex Clifford modules it is convenient to work with the complexified Clif-
ford algebras
Cln = Cln ⊗R C.
The (complex) representation theory of Cln depends on the parity of n so that we will
discuss each case separately. The reader may want to refresh his/her memory of the
considerations in Subsection 2.2.5.
Let n = 2k, and consider an n-dimensional Euclidean space (V, g) . The decisive step
in describing the complex Cl(V )-modules is the following.
Proposition 12.1.19. There exists a complex Cl(V )-module S = S(V ) such that

Cl(V ) ∼
= EndC (S) as C−algebras.

(The above isomorphism is not natural; it depends on several auxiliary choices.)

Proof. Consider a complex structure on V , i.e., a skew-symmetric operator J : V → V

such that J 2 = −1V . Such a J exists since V is even dimensional. Let

{e1 , f1 , . . . , ek , fk }

be an orthonormal basis of V such that Jei = fi , ∀i.

Extend J by complex linearity to V ⊗R C. We can now decompose V ⊗ C into the
eigenspaces of J
V = V 1,0 ⊕ V 0,1 ,
where V 1,0 = ker(i − J) and V 0,1 = ker(i + J). Alternatively,

V 1,0 = spanC (ej − ifj ), V 0,1 = spanC (ej + ifj ).

The metric on V defines a Hermitian metric on the complex vector space (V, J)

h(u, v) = g(u, v) + ig(u, Jv),

(see (2.2.15) of Subsection 2.2.5) which allows us to identify

=C (V, J) ∼
V 0,1 ∼ =C Vc∗ ∼
=C (V 1,0 )∗ .

(Above, Vc∗ denotes the complex dual of the complex space (V, J)). With respect to this
Hermitian metric the collection
 1
εj := √ (ej − ifj ); 1 ≤ j ≤ k
2

is an orthonormal basis of V 1,0 , while the collection

 1
εj = √ (ej + ifj ); 1 ≤ j ≤ k
2
12.1. THE STRUCTURE OF DIRAC OPERATORS 627

is an orthonormal basis of V 0,1 . Set

S2k := Λ• V 1,0 = Λ•,0 V.

A morphism of algebras c : Cl(V ) → End (S2k ) is uniquely defined by its restriction to

V ⊗ C = V 1,0 ⊕ V 0,1 .

Hence, we have to specify the action of each of the components V 1,0 and V 0,1 .
The elements w ∈ V 1,0 will act by exterior multiplication
√ √
c(w)ω = 2e(w)ω = 2w ∧ ω, ∀ω ∈ Λ•,0 V.

The elements w ∈ V 0,1 can be identified with complex linear functionals on V 1,0 , and as
such, they will act by interior differentiation
√
c(w)w1 ∧ · · · ∧ wℓ = − 2i(w)(w1 ∧ · · · ∧ wℓ )
ℓ
√ X
= 2 (−1)j gC (wj , w)w1 ∧ · · · ∧ ŵj ∧ · · · ∧ wℓ ,
j=1

where gC denotes the extension of g to (V ⊗ C) × (V ⊗ C) by complex linearity.

To check that the above constructions do indeed define an action of Cl(V ) we need to
check that, ∀v ∈ V , we have
c(v)2 = −1S2k .
This boils down to verifying the anticommutation rules

{c(ei ), c(fj )} = 0, {c(ei ), c(ej )} = −2δij = {c(fi ), c(fj )}.

We have
1 i
ei = √ (εi + εi ), fj = √ (εj − εj ),
2 2
so that,
c(ei ) = e(εi ) − i(εi ), c(fj ) = i(e(εj ) + i(εj )).
The anti-commutation rules follow as in the Exercise 2.2.55 using the equalities

gC (εi , εj ) = δij .

This shows S2k is naturally a Cl(V )-module. Note that dimC S2k = 2k so that

dimC EndC (S) = 2n = dimC Cl(V ).

A little work (left to the reader) shows the morphism c is injective. This completes the
proof of the proposition. ⊔
⊓

Definition 12.1.20. The Cl(V )-module S(V ) is known as the (even) complex spinor
module. ⊔
⊓
628 CHAPTER 12. DIRAC OPERATORS

Using basic algebraic results about the representation theory of the algebra of endo-
morphisms of a vector space we can draw several useful consequences. (See [120] for a
very nice presentation of these facts.)

Corollary 12.1.21. There exists an unique (up to isomorphism) irreducible complex Cl2k -
module and this is the complex spinor module S2k . ⊔
⊓

Corollary 12.1.22. Any complex Cl2k -module has the form S2k ⊗ W , where W is an
arbitrary complex vector space. The action of Cl2k on S2k ⊗ W is defined by

v · (s ⊗ w) = c(v)s ⊗ w.

The vector space W is called the twisting space of the given Clifford module. ⊔
⊓

Remark 12.1.23. Given a complex Cl2k -module E, its twisting space can be recovered
as the space of morphisms of Clifford modules

W = HomCl2k (S2k , E). ⊔

⊓

Assume now that (V, g) is an oriented, 2k-dimensional Euclidean space. For any posi-
tively oriented orthonormal basis e1 , . . . , e2k we can form the element

Γ = ik e1 · · · e2k ∈ Cl(V ).

One can check easily this element is independent of the oriented orthonormal basis, and
thus it is an element intrinsically induced by the orientation. It is called the chirality
operator defined by the orientation. Note that

Γ2 = 1 and Γx = (−1)deg x Γ ∀x ∈ Cl0 (V ) ∪ Cl1 (V ).

Let S = S(V ) denote the spinor module of Cl(V ). The chirality operator defines an
involutive endomorphism of S, and thus defines a Z2 -grading on S

S = S+ ⊕ S− (S± = ker(±1 − Γ)),

and hence a Z2 -grading of End(S). Since {v, Γ} = 0, ∀v ∈ V , we deduce the Clifford mul-
tiplication by v is an odd endomorphism of S. This means that any algebra isomorphism
Cl(V ) → End (S(V )) is an isomorphism of Z2 -graded algebras.

Exercise 12.1.24. Let J be a complex structure on V . This produces two things: it

defines an orientation on V and identifies S = S(V ) ∼
= Λ∗,0 V . Prove that with respect to
these data the chiral grading of S is

S+/− ∼
= Λeven/odd,0 V. ⊔
⊓
12.1. THE STRUCTURE OF DIRAC OPERATORS 629

The above considerations extend to arbitrary Clifford modules. The chirality operator
introduces a Z2 -grading in any complex Clifford module which we call the chiral grading.
However this does not exhaust the family of Clifford s-modules. The family of s-modules
can be completely described as

ˆ ; W complex s − space ,
S⊗W

where ⊗ˆ denotes the s-tensor product. The modules endowed with the chiral grading form
the subfamily in which the twisting s-space W is purely even.
Example 12.1.25. Let (V, g) be a 2k-dimensional, oriented, Euclidean space. Then
Λ•R V ⊗ C is naturally a Clifford module. Thus, it has the form

Λ•R V ⊗ C ∼
= S ⊗ W.

To find the twisting space, we pick a complex structure on V whose induced orientation
agrees with the given orientation of V . This complex structure produces an isomorphism

Λ•R V ⊗R C ∼
= Λ•,0 V ⊗ Λ0,• V ∼
= S ⊗ (Λ•,0 V )∗ ∼
= S ⊗R S ∗ .

This shows the twisting space is S∗ .

On the other hand, the chirality operator defines Z2 gradings on both S, and S∗ , so
that Λ•R V ⊗ C can be given two different s-structures: the chiral superstructure, in which
the grading of S∗ is forgotten, and the grading as a super-tensor product S⊗S ˆ ∗ . Using
Exercise 12.1.24 we deduce that the second grading is precisely the degree grading

Λ•R V ⊗ C = Λeven V ⊗ C ⊕ Λodd V ⊗ C.

To understand the chiral grading we need to describe the action of the chiral operator on
Λ•R V ⊗ C. This can be done via the Hodge ∗-operator. More precisely we have

Γ · ω = ik+p(p−1) ∗ ω ∀ω ∈ Λp V ⊗ C. (12.1.1)

⊔
⊓

Exercise 12.1.26. Prove the equality 12.1.1. ⊔

⊓

In order to formulate the final result of this subsection we need to extend the automor-
phism α, and the anti-automorphism ♭ to the complexified Clifford algebra Cl(V ). The
automorphism α can be extended by complex linearity

α(x ⊗ z) = α(x) ⊗ z, ∀x ∈ Cl(V ),

while ♭ extends according to the rule

(x ⊗ z)♭ = x♭ ⊗ z.

As in the real case, we set y † := α(y ♭ ) = α(y)♭ , ∀y ∈ Cl(V ).

630 CHAPTER 12. DIRAC OPERATORS

Proposition 12.1.27. Let S(V ) denote the spinor module of the 2k-dimensional Eu-
clidean space (V, g). Then for every isomorphism of algebras

ρ : Cl(V ) → EndC (S(V ))

there exists a Hermitian metric on S(V ) such that

ρ(y † ) = ρ(y)∗ , ∀y ∈ Cl(V ).

Moreover, this metric is unique up to a multiplicative constant.

Sketch of proof Choose an orthonormal basis {e1 , . . . , e2k } of V , and denote by G the
finite subgroup of Cl(V ) generated by the basic vectors ei .
Pick a Hermitian metric h on S(V ) and, for each g ∈ G, denote by hg the pulled-back
metric
hg (s1 , s2 ) := h(ρ(g)s1 , ρ(g)s2 ), ∀s1 , s2 ∈ S.
We can now form the averaged metric,
1 X
hG := hg .
|G|
g∈G

Each ρ(g) is an unitary operator with respect to this metric. We leave the reader to check
that this is the metric we are after. The uniqueness follows from the irreducibility of S(V )
using Schur’s lemma. ⊔
⊓

Corollary 12.1.28. Let (V, g) as above, and ρ : Cl(V ) → EndC (E) be a complex Clifford
module. Then E admits at least one Hermitian metric with respect to which ρ is self-
adjoint.

Proof. Decompose E as S ⊗ W , and ρ as ∆ ⊗ 1W , for some isomorphism of algebras

∆ : Cl(V ) → End(S). The sought for metric is a tensor product of the canonical metric
on S, and some metric on W . ⊔
⊓

12.1.4 Clifford modules: the odd case

The odd dimensional situation can be deduced using the facts we have just established
concerning the algebras Cl2k . The bridge between these two situations is provided by the
following general result.
Lemma 12.1.29. Clm ∼
= Cleven
m+1 .

Proof. Pick an orthonormal basis {e0 , e1 , . . . , em } of the standard Euclidean space Rm+1 .
These generate the algebra Clm+1 . We view Clm as the Clifford algebra generated by
{e1 , . . . , em }. Now define

Ψ : Clm → Cleven 0 1 0 1
m+1 , Ψ(x + x ) = x + e0 · x ,

where x0 ∈ Cleven 1 odd

m , and x ∈ Clm . We leave the reader to check that this is indeed an
isomorphism of algebras. ⊔
⊓
12.1. THE STRUCTURE OF DIRAC OPERATORS 631

Proposition 12.1.30. Let (V, g) be a (2k + 1)-dimensional Euclidean space. Then there
exist two complex, irreducible Cl(V )-modules S+ (V ) and S− (V ) such that

Cl(V ) ∼
= EndC (S+ ) ⊕ EndC (S− ) as ungraded algebras.

The direct sum S(V ) = S+ (V ) ⊕ S− (V ) is called the (odd) spinor module.

Proof. Fix an orientation on V and a positively oriented orthonormal basis e1 , e2 , . . . , e2k+1 .

Denote by S2k+2 the spinor module of Cl(V ⊕ R), where V ⊕ R is given the direct sum
Euclidean metric and the orientation

or(V ⊕ R) = or ∧ or(R).

Choose an isomorphism
ρ : Cl(V ⊕ R) → EndC (S2k+2 ).
Then S2k+2 becomes naturally a super Cl2k+2 -module

S2k+2 = S+ −
2k+2 ⊕ S2k+2 ,

and we thus we get the isomorphisms of algebras

Cl(V ) ∼
= Cl(V ⊕ R)even ∼
= Endeven (S2k+2 ) ∼
= End (S+ −
2k+2 ) ⊕ End (S2k+2 ). ⊔
⊓

As in the previous section, we can choose Hermitian metrics on S± (V ) such that the
morphism

ρ : Cl(V ) → EndC S(V )
is self-adjoint, i.e.,
ρ(u† ) = ρ(u)∗ , ∀u ∈ Cl(V ).
The above result can be used to describe the complex (super)modules of Cl2k+1 . We will
not present the details since the applications we have in mind do not require these facts.
For more details we refer to [83].

12.1.5 A look ahead

In this heuristic section we interrupt a little bit the flow of arguments to provide the
reader a sense of direction. The next step in our story is to glue all the pointwise data
presented so far into smooth families (i.e. bundles). To produce a Dirac operator on an
n-dimensional Riemann manifold (M, g) one needs several things.

(a) A bundle of Clifford algebras C → M such that Cx ∼

= Cln or Cln , ∀x ∈ M .

(b) A fiberwise injective morphism of vector bundles ı : T ∗ M ֒→ C such that, ∀x ∈ M

{ı(u), ı(v)}C = −2g(u, v), ∀u, v ∈ Tx∗ M.

(c) A bundle of Clifford modules, i.e. a vector bundle E → M , together with a morphism
c : C → End (E) whose restrictions to the fibers are morphisms of algebras.
632 CHAPTER 12. DIRAC OPERATORS

(d) A connection on E.

The above collection of data can be constructed from bundles associated to a common
principal bundle. The symmetry group of this principal bundle has to be a Lie group with
several additional features which we now proceed to describe.
Let (V, g) denote the standard fiber of T ∗ M , and denote by AutV the group of auto-
morphisms φ of Cl(V ) such that φ(V ) ⊂ V . The group AutcV is defined similarly, using
the complexified algebra Cl(V ) instead of Cl(V ). For brevity, we discuss only the real
case. We need the following objects.

• A Lie group G which admits a smooth morphism ρ : G → AutV . Tautologically, ρ

defines a representation ρ : G → GL(V ), which we assume is orthogonal.
• A Clifford module c : Cl(V ) → End (E).

• A representation µ : G → GL(E), such that, for every v ∈ V , and any g ∈ G, the

diagram below is commutative.
E
c(v)
E w
g
u ug (12.1.2)
E wE
c(g·v)

The above commutativity can be given an invariant theoretic interpretation as follows.

View the Clifford multiplication c : V → End (E) as an element c ∈ V ∗ ⊗ E ∗ ⊗ E. The
group G acts on this tensor product, and the above commutativity simply means that c
is invariant under this action.
In concrete applications E comes with a metric, and we also need to require that µ is
an orthogonal/unitary representation.
To produce all the data (a)-(d) all we now need is a principal G-bundle P → M such
that the associated bundle P ×ρ V is isomorphic with T ∗ M . (This may not be always
feasible due to possible topological obstructions). Any connection ∇ on P induces by
association metric connections ∇M on1 T ∗ M and ∇E on the bundle of Clifford modules
E = P ×µ E. With respect to these connections, the Clifford multiplication is covariant
constant, i.e.,

∇E (c(α)u) = c(∇M α) + c(α)∇E u, ∀α ∈ Ω1 (M ), u ∈ C ∞ (E).

This follows from the following elementary invariant theoretic result.

Lemma 12.1.31. Let G be a Lie group, and ρ : G → Aut (E) be a linear representation of
G. Assume that there exists e0 ∈ E such that ρ(g)e0 = e0 , ∀g ∈ G. Consider an arbitrary
principal G-bundle P → X, and an arbitrary connection ∇ on P . Then e0 canonically
determines a section u0 on P ×ρ E which is covariant constant with respect to the induced
connection ∇E = ρ∗ (∇), i.e.,
∇E u0 = 0. ⊔
⊓

1
In practice one requires a little more namely that ∇M is precisely the Levi–Civita connection on T ∗ M .
This leads to significant simplifications in many instances.
12.1. THE STRUCTURE OF DIRAC OPERATORS 633

Exercise 12.1.32. Prove the above lemma. ⊔

⊓

Apparently, the chances that a Lie group G with the above properties exists are very
slim. The very pleasant surprise is that all these, and even more, happen in many geo-
metrically interesting situations.
Example 12.1.33. Let (V, g) be an oriented Euclidean space. Using the universality
property of Clifford algebras we deduce that each g ∈ SO(V ) induces an automorphism
of Cl(V ) preserving V ֒→ Cl(V ). Moreover, it defines an orthogonal representation on the
canonical Clifford module
c : Cl(V ) → End (Λ• V ),
such that

c(g · v)(ω) = g · (c(v)(g −1 · ω)), ∀g ∈ SO(V ), v ∈ V, ω ∈ Λ• V,

i.e., SO(V ) satisfies the equivariance property (12.1.2).

If (M, g) is an oriented Riemann manifold, we can now build our bundle of Clifford
modules starting from the principal SO bundle of its oriented orthonormal coframes. As
connections we can now pick the Levi–Civita connection and its associates. The corre-
sponding Dirac operator is the Hodge–DeRham operator. ⊔
⊓

The next two sections discuss two important examples of Lie groups with the above
properties. These are the spin groups Spin(n) and its “complexification” Spinc (n). It
turns out that all the groups one needs to build Dirac operators are these three classes:
SO, Spin, and Spinc .

12.1.6 The spin group

Let (V, g) be a finite dimensional Euclidean space. The group of automorphisms of the
Clifford algebra Cl(V ) contains a very rich subgroup consisting of the interior ones. These
have the form

ϕx : Cl(V ) → Cl(V ) u 7→ ϕx (u) = x · u · x−1 , ∀u ∈ Cl(V ),

where x is some invertible element in Cl(V ).

The candidates for the Lie groups with the properties outlined in the previous subsec-
tion will be sought for amongst subgroups of interior automorphisms. It is thus natural
to determine the subgroup

x ∈ Cl(V )⋆ ; x · V · x−1 ⊂ V ,

where Cl(V )⋆ denotes the group of invertible elements. We will instead try to understand
the Clifford group

Γ(V ) = x ∈ Cl⋆ (V ) α(x) · V · x−1 ⊂ V ,

where α : Cl(V ) → Cl(V ) denotes the involutive automorphism of Cl(V ) defining its
Z2 -grading.
634 CHAPTER 12. DIRAC OPERATORS

In general, the map

ρx
Cl(V ) ∋ u 7−→ α(x)ux−1 ∈ Cl(V )
is not an automorphism of algebras, but as we will see by the end of this subsection, if
x ∈ Γ(V ), then ρx = ±ϕx , and hence a posteriori this alteration has no impact. Its impact
is mainly on the æsthetics of the presentation which we borrowed from the elegant paper
[9].
By construction, the Clifford group Γ(V ) comes equipped with a tautological repre-
sentation
ρ : Γ(V ) → GL(V ) ρ(x) : v 7→ α(x) · v · x−1 .
Proposition 12.1.34. ker ρ = (R∗ , ·) ⊂ Cl(V )⋆ .

Proof. Clearly R∗ ⊂ ker ρ. To establish the opposite inclusion choose an orthonormal

basis (ei ) of V , and let x ∈ ker ρ. The element x decomposes into even/odd components

x = x0 + x1 ,

and the condition α(x)ei x−1 = ei translates into

(x0 − x1 )ei = ei (x0 + x1 ), ∀i.

This is equivalent with the following two conditions

[x0 , ei ] = x0 ei − ei x0 = x1 ei + ei x1 = {x1 , ei } = 0, ∀i.

In terms of the s-commutator, the above two equalities can be written as one

[ei , x]s = 0 ∀i.

Since [·, x] is a superderivation of Cl(V ) we conclude that

[y, x]s = 0, ∀y ∈ Cl(V ).

In particular, the even part x0 lies in the center of Cl(V ). We let the reader check the
following elementary fact.
Lemma 12.1.35. The center of the Clifford algebra is the field of scalars R ⊂ Cl(V ). ⊔
⊓

Note that since {x1 , ei } = 0, then x1 should be a linear combination of elementary

monomials ej1 · · · ejs none of which containing ei as a factor. Since this should happen for
every i this means x1 = 0, and this concludes the proof of the proposition. ⊔
⊓

Definition 12.1.36. The spinorial norm is the map

N : Cl(V ) → Cl(V ) N (x) = x♭ x. ⊔

⊓

Proposition 12.1.37. The following hold.

1. N (Γ(V )) ⊂ R∗ .
12.1. THE STRUCTURE OF DIRAC OPERATORS 635

2. The map N : Γ(V ) → R∗ is a group morphism.

Proof. Let x ∈ Γ(V ). We first prove that x♭ ∈ Γ(V ). Since α(x)vx−1 ∈ V , ∀v ∈ V , we

deduce that
α({α(x) · v · x−1 }♭ ) = −α(x) · v · x−1 ∈ V.
Using the fact that x 7→ x♭ is an anti-automorphism, we deduce

α( (x♭ )−1 · v · α(x♭ ) ) ∈ V,

so that,
α((x♭ )−1 ) · v · x♭ ∈ V,
that is, (x♭ )−1 ∈ Γ(V ). Hence x† ∈ Γ(V ), ∀x ∈ Γ(V ). In particular, since α(Γ(V )) ⊂ Γ(V ),
we deduce N (Γ(V )) ⊂ Γ(V ).
For any v ∈ V we have

α(N (x)) · v · (N (x))−1 = α(x♭ x) · v · (x♭ x)−1 = α(x♭ ) · {α(x) · v · x−1 } · (x♭ )−1 .

On the other hand, y = α(x) · v · x−1 is an element in V , which implies y † = α(y ♭ ) = −y.
Hence
α(N (x)) · v · (N (x))−1 = −α(x♭ ) · y † · (x♭ )−1
= −α(x♭ ) · α(x♭ )−1 · v † · x♭ · (x♭ )−1 = −v † = v.
This means N (x) ∈ ker ρ = R∗ .
(b) If x, y ∈ Γ(V ), then

N (x · y) = (xy)♭ (xy) = y ♭ x♭ xy = y ♭ N (x)y = N (x)y ♭ y = N (x)N (y). ⊔

⊓

Theorem 12.1.38. (a) For every x ∈ Γ(V ), the transformation ρ(x) of V is orthogonal.
(b) There exists a short exact sequence of groups
ρ
1 → R∗ ֒→ Γ(V ) → O(V ) → 1.

(c) Every x ∈ Γ(V ) can be written (in a non-unique way) as a product x = v1 · · · vk ,

vj ∈ V . In particular, every element of Γ(V ) is Z2 -homogeneous, i.e., it is either purely
even, or purely odd.

Proof. (a) Note that, ∀v ∈ V , we have N (v) = −|v|2g . For every x ∈ Γ(V ) we get

N (ρ(x)(v)) = N (α(x)vx−1 ) = N (α(x))N (v)N (x−1 ) = N (α(x))N (x)−1 N (v).

On the other hand, x2 = α(x2 ) = α(x)2 so we conclude that

N (x)2 = N (α(x))2 .

Hence, N (ρ(x)(v)) = ±N (v). Since both N (v) and N (ρx (v)) are negative numbers, we
deduce that the only possible choice of signs in the above equality is +. This shows that
ρ(x) is an orthogonal transformation.
636 CHAPTER 12. DIRAC OPERATORS

(b) & (c) We only need to show ρ(Γ(V )) = O(V ). For x ∈ V with |x|g = 1 we have

α(x) = −x = x−1 .

If we decompose v ∈ V as λx + u, where λ ∈ R and u ⊥ x, then we deduce

ρ(x)v = −λx + u.

In other words, ρ(x) is the orthogonal reflection in the hyperplane through origin which
is perpendicular to x. Since any orthogonal transformation of V is a composition of such
reflections (Exercise), we deduce that for each T ∈ O(V ) we can find v1 , . . . , vk ∈ V such
that
T = ρ(v1 ) · · · ρ(vk ).
Incidentally, this also establishes (c). ⊔
⊓

Exercise 12.1.39. Suppose V is a finite dimensional Euclidean space. Prove that any
orthogonal operator T ∈ O(V ) is a product of at most dim V reflections. ⊔
⊓

Set
Γ0 (V ) := Γ(V ) ∩ Cleven .
Note that 
ρ(Γ0 (V )) ⊂ SO(V ) = T ∈ O(V ); det T = 1 .
Hence, we have a short exact sequence
ρ
1 → R∗ ֒→ Γ0 (V ) → SO(V ) → 1.

Definition 12.1.40. Set


Pin(V ) := x ∈ Γ(V ); |N (x)| = 1 ,

and 
Spin(V ) := x ∈ Γ(V ); N (x) = 1 = Pin(V ) ∩ Γ0 (V ). ⊔
⊓
The results we proved so far show that Spin(V ) can be alternatively described by the
following “friendlier” equality

Spin(V ) = {v1 · · · v2k ; k ≥ 0, vi ∈ V, |vi | = 1, ∀i = 1. . . . 2k}.

In particular, this shows that Spin(V ) is a compact topological group. Observe that
Spin(V ) is a closed subgroup of the Lie group GL( Cl(V ) ). This implies (see Remark
1.2.31 and [61, 128]) that Spin(V ) is in fact a Lie group, and the map Spin(V ) ֒→ Cl(V )
is a smooth embedding. The proof of the following result is left to the reader.
Proposition 12.1.41. There exist short exact sequences

1 → Z2 → Pin(V ) → O(V ) → 1

1 → Z2 → Spin(V ) → SO(V ) → 1. ⊔
⊓
12.1. THE STRUCTURE OF DIRAC OPERATORS 637

Proposition 12.1.42. The morphism ρ : Spin(V ) → SO(V ) is a covering map. More-

over, the group Spin(V ) is connected if dim V ≥ 2, and simply connected if dim V ≥ 3. In
particular, Spin(V ) is the universal cover of SO(V ), when dim V ≥ 3.

Proof. The fact that ρ : Spin(V ) → SO(V ) is a covering map is an elementary consequence
of the following simple observations.

(i) The map ρ is a group morphism ;

(ii) The map ρ is continuous, and proper;

(iii) The subgroup ker ρ is discrete.

Since SO(V ) is connected if dim V ≥ 2, the fact that Spin(V ) is connected would
follow if we showed that any points in the same fiber of ρ can be connected by arcs. It
suffices to look at the fiber ρ−1 (1) = {−1, 1}.
If u, v ∈ V are such that |u| = |v| = 1, u ⊥ v, then the path

γ(t) = (u cos t + v sin t)(u cos t − v sin t), 0 ≤ t ≤ π/2

lies inside Spin(V ), because |u cos t ± v sin t| = 1, and moreover

γ(0) = −1, γ(π/2) = 1.

To prove that Spin(V ) is simply connected, we argue by induction on dim V .

If dim V = 3, then Spin(V ) is isomorphic to the group of unit quaternions (see Example
12.1.55), and in particular, it is homeomorphic to the sphere S 3 which is simply connected.
Note that if V is an Euclidean space, and U is a subspace, the natural inclusion U ֒→ V
induces morphisms

Cl( U ) ֒→ Cl(V ), SO(U ) ֒→ SO(V ), Spin(U ) ֒→ Spin(V ),

such that the diagrams below are commutative

u
Spin(V )
ρ
w SO(V
u ) u
Spin(V ) y w Cl(Vu )
i
y y , y y
Spin(U ) ρ
w SO(U ) Spin(U ) w Cl(U )
Hence, it suffices to show that if dim V > 3, then for every smooth, closed path

û : [0, 1] → Spin(V ), û(0) = û(1) = 1,

there exists a codimension one subspace U ֒→ V such that û is homotopic to a loop in

Spin(U ). Fix a unit vector e ∈ V , set u(t) := ρû(t) ∈ SO(V ), and define v(t) := u(t)e ∈ V .
The correspondence t 7→ v(t) is a smooth, closed path on the unit sphere S dim V −1 ⊂ V .
Using Sard’s theorem we deduce that there exists a vector x ∈ S dim V −1 , such that

v(t) 6= ±x, ∀t ∈ [0, 1].

638 CHAPTER 12. DIRAC OPERATORS

In particular, the vectors v(t) and x are linearly independent, for any t. Denote by f (t)
the vector obtained by applying the Gramm-Schmidt orthonormalization process to the
ordered linearly independent set {x, v(t)}. In other words,

1
f (t) := f0 (t), f0 (t) := v(t) − (v(t) • x)x,
|f0 (t)|

where we denoted the inner product in V by •. We can find a smooth map

θ : [0, 1] → [0, 2π),

such that,
θ(0) = θ(1) = 0, v(t) = x cos 2θ(t) + f (t) sin 2θ(t).
Set
σ(t) := cos θ + xf sin θ ∈ Cl(V ).
Note that  θ θ  θ θ
σ(t) = x cos − f sin −x cos − f sin ,
2 2 2 2
so that σ(t) ∈ Spin(V ).
On the other hand,
σ(t)−1 (t) = cos θ − xf sin θ,
and

σ(t)xσ(t)−1 = (x cos θ + f sin θ)(cos θ − xf sin θ) = x cos 2θ + f sin 2θ = e0 .

In other words,
ρσ(t) x = v(t).
For every s ∈ [0, 1] define

σs (t) := ûs (t) = cos sθ + xf sin sθ, ûs (t) := σs (t)−1 û(t)σs (t),

and
us (t) := ρ−1
σs (t) u(t)ρσs (t) ∈ SO(V ).

Observe that for s = 0 we have û0 (t) = û(t). On the other hand, for s = 1 we have

u1 (t)x = ρ−1
σs (t) u(t)ρσs (t) x = ρσs (t)−1 v(t) = x.

If we denote by U the orthogonal complement of x in V , we deduce that u1 (t) ∈ Spin(U ),

∀t ∈ [0, 1]. ⊔
⊓

We can view Spin(V ) as a submanifold of Cl(V ), and as such we can identify its Lie
algebra spin(V ) with a linear subspace of Cl(V ). The next result offers a more precise
description.
12.1. THE STRUCTURE OF DIRAC OPERATORS 639

Proposition 12.1.43. Consider the quantization map q : Λ∗ V → Cl(V ). Then

spin(V ) = q(Λ2 V ).

The Lie bracket is given by the commutator in Cl(V ).

Proof. The group Γ(V ) is a Lie group as a closed subgroup of the group of linear trans-
formations of Cl(V ). Since the elements of Γ(V ) are either purely even, or purely odd, we
deduce that the tangent space at 1 ∈ Γ(V ) is a subspace of

E= x ∈ Cleven (V ); xv − vx ∈ V, ∀v ∈ V .

Fix x ∈ E, and let e1 , . . . , en be an orthonormal basis of V . We can decompose x as

x = x0 + e1 x1 ,

where x0 ∈ Cl(V )even , and x1 ∈ Cl(V )odd are linear combinations of monomials involving
only the vectors e2 , . . . , en . Since [x0 , e1 ] = 0, and {x1 , e1 } = 0, we deduce

1 1
e1 x1 = [e1 , x1 ] = [e1 , x] ∈ V.
2 2
In particular this means x1 ∈ R ⊕ V ∈ Cl(V ).
Repeating the same argument with every vector ei we deduce that

E ⊂ R ⊕ span{ei · ej ; 1 ≤ i < j ≤ dim V } = R ⊕ q(Λ2 V ).

Thus,
T1 Γ0 (V ) ⊂ R ⊕ q(Λ2 V ).
The tangent space to Spin(V ) satisfies a further restriction obtained by differentiating the
condition N (x) = 1. This gives

spin(V ) ⊂ {x ∈ R ⊕ q(Λ2 V ) ; x♭ + x = 0} = q(Λ2 V ).

Since dim spin(V ) = dim so(V ) = dim Λ2 V we conclude that the above inclusion is in fact
an equality of vector spaces.
Now consider two smooth paths x, y : (−ε, ε) → Spin(V ) such that x(0) = y(0) = 1.
The Lie bracket of ẋ(0) and ẏ(0) is then found (using the Exercise 3.1.21) from the equality

x(t)y(t)x(t)−1 y(t)−1 = 1 + [ẋ(0), ẏ(0)]t2 + O(t3 ) (as t → 0),

where the above bracket is the commutator of ẋ(0) and ẏ(0) viewed as elements in the
associative algebra Cl(V ). ⊔
⊓

To get a more explicit picture of the induced morphism of Lie algebras

ρ∗ : spin(V ) → so(V ),
640 CHAPTER 12. DIRAC OPERATORS

we fix an orientation on V , and then choose a positively oriented orthonormal basis

{e1 , . . . , en } of V , (n = dim V ). For every x ∈ spin(V ), the element ρ∗ (x) ∈ spin(V )
acts on V according to
ρ∗ (x)v = x · v − v · x.
If X
x= xij ei ej ,
i<j

then X
ρ∗ (x)ej = −2 xij ei , (xij = −xji ).
i

Note the following often confusing fact. If we identify as usual so(V ) ∼

= Λ2 V by
X X
so(n) ∋ A 7→ ωA = g(Aei , ej )ei ∧ ej = − g(ei , Aej )ei ∧ ej ,
i<j i<j

then the Lie algebra morphism ρ∗ takes the form

X X
ωρ∗ (x) = − g(ei , ρ∗ (x)ej )ei ∧ ej = 2 xij ei ∧ ej = 2σ(x) ∈ Λ2 V,
i<j i<j

where σ : Cl(V ) → Λ• V is the symbol map, ei ej 7→ ei ∧ ej . In particular, this shows ρ∗ is

an isomorphism, and that the map ρ : Spin(V ) → SO(V ) is also a submersion.

☞ A word of warning. If A ∈ so(V ) has the matrix description

X
Aej = aij ei
i

with respect to an oriented orthonormal basis {e1 , . . . , en }, (n = dim V ), then the 2-form
associated to A has the form X
ωA = − aij ei ∧ ej ,
i<j

so that
1X 1X
ρ−1
∗ (A) = − aij ei ej = − aij ei ej . (12.1.3)
2 4
i<j i,j

The above negative sign is essential, and in many concrete problems it makes a world of
difference. ⊔
⊓

Any Clifford module φ : Cl(V ) → EndK (E) defines by restriction a representation

φ : Spin(V ) → GLK (E).

The (complex) representation theory described in the previous sections can be used to
determine the representations of Spin(V ).
12.1. THE STRUCTURE OF DIRAC OPERATORS 641

Example 12.1.44. (The complex spinor representations). Consider a finite dimen-

sional oriented Euclidean space (V, g). Assume first that dim V is even. The orientation
on V induces a Z2 -grading on the spinor module S = S+ ⊕ S− . Since Spin(V ) ⊂ Cleven (V )
we deduce that each of the spinor spaces S± is a representation space for Spin(V ). They
are in fact irreducible, nonisomorphic complex Spin(V )-modules. They are called the
positive/negative complex spin representations.
Assume next that dim V is odd. The spinor module S(V )

S(V ) = S+ (V ) ⊕ S− (V ).

is not irreducible as a Cl(V ) modules. Each of the modules S± (V ) is a representation

space for Spin(V ). They are irreducible but also isomorphic as Spin(V )-modules. If we
pick an oriented orthonormal basis e1 , · · · , e2n+1 of V then the Clifford multiplication by
ω = e1 · · · e2n+1 intertwines the ± components. This is a Spin(V ) isomorphism since ω
lies in the center of Cl(V ). ⊔
⊓

✍ Convention. For each positive integer n we will denote by Sn the Spin(n)-module

defined by
S2k ∼
= S+ (R2k ) ⊕ S− (R2k ) if n = 2k,

and
S2k+1 ∼
= S+ (R2k+1 ) ∼
=Spin(2k+1) S− (R2k+1 ) if n = 2k + 1.

The module Sn will be called the fundamental complex spinor module of Spin(n). ⊔
⊓

Exercise 12.1.45. Let (V, g) be a 2k-dimensional Euclidean space and J : V → V a

complex structure compatible with the metric g. Thus we have an explicit isomorphism

∆ : Cl(V ) → End (Λ•,0 V ).

Choose u ∈ V such that |u|g = 1 and then for each t ∈ R set

q = q(t) = cos t + u · v sin t, v = Jv.

Note that q ∈ Spin(V ) so that ∆(q) preserves the parities when acting on Λ∗,0 (V ). Hence
∆(q) = ∆+ (q) ⊕ ∆− (q) where ∆+/− (q) acts on Λeven/odd,0 V . Compute tr (∆+/− (q)) and
then conclude that Λeven,/odd,0 V are non-isomorphic Spin(V )-modules. ⊔
⊓

Proposition 12.1.46. Let φ : Cl(V ) → End (E) be a self-adjoint Clifford module. Then
the induced representation of Spin(V ) is orthogonal (unitary).

Exercise 12.1.47. Prove the above proposition. ⊔

⊓

Exercise 12.1.48. Prove that the group Spin(V ) satisfies all the conditions discussed in
Subsection 10.1.5. ⊔
⊓
642 CHAPTER 12. DIRAC OPERATORS

12.1.7 The complex spin group

The considerations in the previous case have a natural extension to the complexified
Clifford algebra Cln . The canonical involutive automorphism

α : Cln → Cln

extends by complex linearity to an automorphism of Cln , while the anti-automorphism

♭: Cln → Cln extends to Cln according to the rule

(v ⊗ z)♭ = v ⊗ z.

As in the real case set x† := α(x)♭ , and N (x) = x♭ · x.

Let (V, g) be an Euclidean space. The complex Clifford group Γc (V ) is defined by

Γc (V ) = {x ∈ Cl(V )⋆ ; α(x) · v · x−1 ∈ V ∀v ∈ V }.

We denote by ρc the tautological representation ρc : Γc (V ) → GL(V, R). As in the real

case one can check that
ρc (Γc (V )) = O(V ), ker ρc = C∗ .
The spinorial norm N (x) determines a group morphism N : Γc (V ) → C∗ . Define

Pinc (V ) = {x ∈ Γc (V ) ; |N (x)| = 1}.

We let the reader check the following result.

Proposition 12.1.49. There exists a short exact sequence

1 → S 1 → Pinc (V ) → O(V ) → 1. ⊔
⊓

Corollary 12.1.50. There exists a natural isomorphism

Pinc (V ) ∼
= (Pin(V ) × S 1 )/ ∼,
where “∼” is the equivalence relation

(x, z) ∼ (−x, −z) ∀(x, z) ∈ Pin(V ) × S 1 .

Proof. The inclusions Pin(V ) ⊂ Cl(V ), S 1 ⊂ C induce an inclusion

(Pin(V ) × S 1 )/ ∼→ Cl(V ).

The image of this morphism lies obviously in Γc (V ) ∩ {|N | = 1} so that (Pin(V ) × S 1 )/ ∼

can be viewed as a subgroup of Pinc (V ). The sought for isomorphism now follows from
the exact sequence
1 → S 1 → (Pin(V ) × S 1 )/ ∼→ O(V ) → 1. ⊔
⊓
We define Spinc (V ) as the inverse image of SO(V ) via the morphism

ρc : Pinc (V ) → O(V ).

Arguing as in the above corollary we deduce

Spinc (V ) ∼
= (Spin(V ) × S 1 )/ ∼∼
= (Spin(V ) × S 1 )/Z2 .
12.1. THE STRUCTURE OF DIRAC OPERATORS 643

Exercise 12.1.51. Prove Spinc (V ) satisfies all the conditions outlined in Subsection
10.1.5. ⊔
⊓

Assume now dim V is even. Then, any any isomorphism mCl(V ) ∼

= EndC (S(V ))
induces a complex unitary representation

Spinc (V ) → Aut (S(V ))

called the complex spinorial representation of Spinc . It is not irreducible since, once we fix
an orientation on V , the space End (S(V )) has a natural superstructure and, by definition,
Spinc acts through even automorphism. As in the real case, S(V ) splits into a direct sum
of irreducible representations S± (V ).
Fix a complex structure J on V . This complex structure determines two things.

(i) A canonical orientation on V .

(ii) A natural subgroup

U (V, J) = {T ∈ SO(V ) ; [T, J] = 0} ⊂ SO(V ).

Denote by ıJ : U (V, J) → SO(V ) the inclusion map.

ξ
J
Proposition 12.1.52. There exists a natural group morphism U (V, J) −→ Spinc (V ) such
that the diagram below is commutative.

''ı w Spin (V )
ξJ c
U (V, J)
')' ρ J

u
c

SO(V )

Proof. Let ω ∈ U (V ), and consider a path γ : [0, 1] → U (V ) connecting 1 to ω. Via the

inclusion U (V ) ֒→ SO(V ) we may regard γ as a path in SO(V ). As such, it admits a
unique lift γ̃ : [0, 1] → Spin(V ) such that γ̃(0) = 1.
Using the double cover S 1 → S 1 , z 7→ z 2 , we can find a path t 7→ δ(t) ∈ S 1 such that

δ(0) = 1 and δ2 (t) = det γ(t).

Define ξ(ω) to be the image of (γ̃(1), δ(1)) in Spinc (V ). We have to check two things.

(i) The map ξ is well defined.

(ii) The map σ is a smooth group morphism.

To prove (i), we need to show that if η : [0, 1] → U (V ) is a different path connecting

1 to ω, and λ : [0, 1] → S 1 is such that λ(0) = 1 and λ(t) = det η(t)2 , then
(η̃(1), λ(1)) = (γ̃(1), δ(1)) in Spinc (V ).
644 CHAPTER 12. DIRAC OPERATORS

ρ
The elements γ̃(1) and η̃(1) lie in the same fiber of the covering Spin(V ) → SO(V ) so
that they differ by an element in ker ρ. Hence

γ̃(1) = ǫη̃(1), ǫ = ±1.

We can identify ǫ as the holonomy of the covering Spin(V ) → SO(V ) along the loop γ ∗ η −
which goes from 1 to ω along γ, and then back to 1 along η − (t) = η(1 − t).
The map det : U (V ) → S 1 induces an isomorphism between the fundamental groups
(see Exercise 6.2.35 of Subsection 6.2.5). Hence, in describing the holonomy ǫ it suffices
to replace the loop γ ∗ η − ⊂ U (V ) by any loop ν(t) such that

det ν(t) = det(γ ∗ η − ) = ∆(t) ∈ S 1 .

Such a loop will be homotopic to γ ∗ η −1 in U (V ), and thus in SO(V ) as well. Select ν(t)
of the form
ν(t)e1 = ∆(t)e1 , ν(t)ei = ei ∀i ≥ 2,
where (ei ) is a complex, orthonormal basis of (V, J). Set fi = Jei . With respect to the
real basis (e1 , f1 , e2 , f2 , . . . ) the operator ν(t) (viewed as an element of SO(V )) has the
matrix description  
cos θ(t) − sin θ(t) · · ·
 sin θ(t) cos θ(t) · · · 
 ,
.. ..
. . 1
where θ : [0, 1] → R is a continuous map such that ∆(t) = eiθ(t) .
The lift of ν(t) to Spin(V ) has the form
θ(t) θ(t)

ν̃(t) = cos − e1 f1 sin .
2 2
We see that the holonomy defined by ν̃(t) is nontrivial if and only if the holonomy of the
z2
loop t 7→ δ(t) in the double cover S 1 → S 1 is nontrivial. This means that δ(1) and λ(1)
differ by the same element of Z2 as γ̃(1) and η̃(1). This proves ξ is well defined. We leave
the reader to check that ξ is indeed a smooth morphism of groups. ⊔
⊓

12.1.8 Low dimensional examples

In low dimensions the objects discussed in the previous subsections can be given more
suggestive interpretations. In this subsection we will describe some of these interpretations.
Example 12.1.53. (The case n = 1). The Clifford algebra Cl1 is isomorphic with
the field of complex numbers C. The Z2 -grading is Re C ⊕ Im C. The group Spin(1) is
isomorphic with Z2 . ⊔
⊓

Example 12.1.54. (The case n = 2). The Clifford algebra Cl2 is isomorphic with the
algebra of quaternions H. This can be seen by choosing an orthonormal basis {e1 , e2 } in
R2 . The isomorphism is given by

1 7→ 1, e1 7→ i, e2 7→ j, e1 e2 7→ k,
12.1. THE STRUCTURE OF DIRAC OPERATORS 645

where i, j, and k are the imaginary units in H. Note that

Spin(2) = {a + bk ; a, b ∈ R, a2 + b2 = 1} ∼
= S 1.

The natural map Spin(1) → SO(2) ∼

= S 1 takes the form eiθ 7→ e2iθ . ⊔
⊓

Example 12.1.55. (The case n = 3). The Clifford algebra Cl3 is isomorphic, as an
ungraded algebra, to the direct sum H ⊕ H. More relevant is the isomorphism Cleven ∼
=
3
Cl2 ∼
= H given by
1 7→ 1, e1 e2 7→ i, e2 e3 7→ j, e3 e1 7→ k,
where {e1 , e2 , e3 } is an orthonormal basis in R3 . Under this identification the operation
x 7→ x♭ coincides with the conjugation in H

x = a + bi + cj + dk 7→ x = a − bi − cj − dk.

In particular, the spinorial norm coincides with the usual norm on H

N (a + bi + cj + dk) = a2 + b2 + c2 + d2 .

Thus, any x ∈ Cleven

3 \{0} is invertible, and

1
x−1 = x♭ .
N (x)

Moreover, a simple computation shows that xR3 x−1 ⊂ R3 , ∀x ∈ Cleven

3 \{0}, so that

Γ0 (R3 ) ∼
= H \ {0}.

Hence
Spin(3) ∼
= { x ∈ H; |x| = 1 } ∼
= SU (2).
The natural map Spin(3) → SO(3) is precisely the map described in the Exercise 6.2.8 of
Subsection 6.2.1.
The isomorphism Spin(3) ∼
= SU (2) can be visualized by writing each q = a+bi+cj +dk
as
q = u + jv, u = a + bi, v = (c − di) ∈ C.
To a quaternion q = u + jv one associates the 2 × 2 complex matrix
 
u −v̄
Sq = ∈ SU (2).
v ū

Note that Sq̄ = Sq∗ , ∀q ∈ H. For each quaternion q ∈ H we denote by Lq (respectively Rq )

the left (respectively right) multiplication. The right multiplication by i defines a complex
structure on H. Define T : H → C2 by
 
u
q = u + jv 7→ T q = .
v
646 CHAPTER 12. DIRAC OPERATORS

A simple computation shows that T (Ri q) = iT q, i.e., T is a complex linear map. Moreover,
∀q ∈ Spin(3) ∼
= S 3 , the matrix Sq is in SU (2), and the diagram below is commutative.

H
T
w C2
L q
u T S u2
q

H wC
In other words, the representation

Spin(3) ∋ q 7→ Lq ∈ GLC (H)

of Spin(3) is isomorphic with the tautological representation of SU (2) on C2 .

On the other hand, the correspondences

e1 e2 7→ Si ⊕ Si ∈ End (C2 ) ⊕ End (C2 )

e2 e3 7→ Sj ⊕ Sj ∈ End (C2 ) ⊕ End (C2 )

e3 e1 7→ Sk ⊕ Sk ∈ End (C2 ) ⊕ End (C2 )
e1 e2 e3 7→ R = 1C2 ⊕ (−1C2 ) ∈ End (C2 ) ⊕ End (C2 )
extend to an isomorphism of algebras Cl3 → End (C2 ) ⊕ End (C2 ). This proves that the
tautological representation of SU (2) is precisely the complex spinorial representation S3 .
From the equalities
[Rj , Lq ] = {Rj , Ri } = 0,
we deduce that Rj defines an isomorphism of Spin(3)-modules

Rj : S3 → S3 .

This implies there exists a Spin(3)-invariant bilinear map

β : S3 × S3 → C.

This plays an important part in the formulation of the recently introduced Seiberg-Witten
equations (see [135]). ⊔
⊓

Exercise 12.1.56. The left multiplication by i introduces a different complex structure

on H. Prove the representation Spin(3) ∋ q 7→ (Rq : H → H) is

(i) complex with respect to the above introduced complex structure on H, and

(ii) it is isomorphic with the complex spinorial representation described by the left mul-
tiplication.

⊔
⊓
12.1. THE STRUCTURE OF DIRAC OPERATORS 647

Example 12.1.57. (The case n = 4). The Clifford algebra Cl4 can be realized as the
algebra of 2 × 2 matrices with entries in H. To describe this isomorphism we have to start
from a natural embedding R4 ֒→ M2 (H) given by the correspondence
 
∼ 4 0 −x
H = R ∋ x 7→
x 0
A simple computation shows that the conditions in the universality property of a Clifford
algebra are satisfied, and this correspondence extends to a bona-fide morphism of algebras
Cl4 → M2 (H). We let the reader check this morphism is also injective. A dimension count
concludes it must also be surjective. ⊔
⊓

Proposition 12.1.58. Spin(4) ∼

= SU (2) × SU (2).
Proof. We will use the description of Spin(4) as the universal (double-cover) of SO(4) so
we will explicitly produce a smooth 2 : 1 group morphism SU (2) × SU (2) → SO(4).
Again we think of SU (2) as the group of unit quaternions. Thus each pair (q1 , q2 ) ∈
SU (2) × SU (2) defines a real linear map

Tq1 ,q2 : H → H, x 7→ Tq1 ,q2 x = q1 xq 2 .

Clearly |x| = |q1 | · |x| · |q2 | = |Tq1 ,q2 x| ∀x ∈ H, so that each Tq1 ,q2 is an orthogonal
transformation of H. Since SU (2) × SU (2) is connected, all the operators Tq1 ,q2 belong to
the component of O(4) containing 1, i.e,. T defines an (obviously smooth) group morphism

T : SU (2) × SU (2) → SO(4).

Note that ker T = {1, −1}, so that T is 2 : 1. In order to prove T is a double cover it
suffices to show it is onto. This follows easily by noticing T is a submersion (verify this! ),
so that its range must contain an entire neighborhood of 1 ∈ SO(4). Since the range
of T is closed (verify this!) we conclude that T must be onto because the closure of the
subgroup (algebraically) generated by an open set in a connected Lie group coincides with
the group itself (see Subsection 1.2.3). ⊔
⊓

The above result shows that

so(4) ∼
= spin(4) ∼
= su(2) ⊕ su(2) ∼
= so(3) ⊕ so(3).

Exercise 12.1.59. Using the identification Cl4 ∼ = M2 (H) show that Spin(4) corresponds
to the subgroup 
diag (p, q); p, q ∈ H, |p| = |q| = 1 ⊂ M2 (H). ⊔
⊓
Exercise 12.1.60. Let {e1 , e2 , e3 , e4 } be an oriented orthonormal basis of R4 . Let ∗ denote
the Hodge operator defined by the canonical metric and the above chosen orientation. Note
that
∗ : Λ2 R4 → Λ2 R4
is involutive, ∗2 = 1, so that we can split Λ2 into the ±1 eigenspaces of ∗

Λ2 R4 = Λ2+ R4 ⊕ Λ2− R4 .
648 CHAPTER 12. DIRAC OPERATORS

(a) Show that 

Λ2± = spanR η1± , η2± , η3± ,
where
1 1
η1± = √ (e1 ∧ e2 ± e3 ∧ e4 ), η2± = √ (e1 ∧ e3 ± e4 ∧ e2 ),
2 2
1
η3± = √ (e1 ∧ e4 ± e2 ∧ e3 ).
2
(b) Show that the above splitting of Λ R4 corresponds to the splitting so(4) = so(3)⊕so(3)
2

under the natural identification Λ2 R4 ∼= so(4). ⊔

⊓

To obtain an explicit realization of the complex spinorial representations S± 4 we need

to describe a concrete realization of the complexification Cl4 . We start from the morphism
of R-algebras  
u −v̄
H ∋ x = u + jv 7→ Sx =
v u
This extends by complexification to an isomorphism of C-algebras
H ⊗R C ∼
= M2 (C).
(Verify this!) We now use this isomorphism to achieve the identification.
EndH (H ⊕ H) ⊗R C ∼
= M2 (H) ⊗R C ∼
= EndC (C2 ⊕ C2 ).
The embedding R4 → Cl4 now takes the form
 
∼ 4 0 −Sx̄
H = R ∋ x 7→ Tx = ∈ End (C2 ⊕ C2 ). (12.1.4)
Sx 0
Note that the chirality operator Γ = −e1 e2 e3 e4 is represented by the canonical involution
Γ 7→ 1C2 ⊕ (−1C2 ).
We deduce the S±
4 representations of Spin(4) = SU (2) × SU (2) are given by

S+
4 : SU (2) × SU (2) ∋ (p, q) 7→ p ∈ GL(2; C),

S−
4 : SU (2) × SU (2) ∋ (p, q) 7→ q ∈ GL(2; C).
Exactly as in the case of Spin(3), these representations can be given quaternionic descrip-
tions.
Exercise 12.1.61. The space R4 ∼ = H has a canonical complex structure defined by Ri
which defines (following the prescriptions in §10.1.3) an isomorphism
c : Cl4 → End (Λ• C2 ).
Identify C2 in the obvious way with Λ1 C2 Λodd C2 , and with Λeven C2 via the map
e1 7→ 1 ∈ Λ0 C2 , e2 7→ e1 ∧ e2 .
Show that under these identifications we have
c(x) = Tx ∀x ∈ R4 ∼
= (H, Ri ) ∼
= C2 ,
where Tx is the odd endomorphism of C2 ⊕ C2 defined in (12.1.4). ⊔
⊓
12.1. THE STRUCTURE OF DIRAC OPERATORS 649

Exercise 12.1.62. Let V be a 4-dimensional, oriented Euclidean space. Denote by

q : Λ• V → Cl(V )

the quantization map, and fix an isomorphism ∆ : Cl(V ) → End (S(V )) of Z2 -graded alge-
bras. Show that for any η ∈ Λ2+ (V ), the image ∆ ◦ q(η) ∈ End (S(V )) is an endomorphism
of the form T ⊕ 0 ∈ End (S+ (V )) ⊕ End (S− (V )). ⊔
⊓

Exercise 12.1.63. Denote by V a 4-dimensional oriented Euclidean space.

(a) Show that the representation S+ +
4 ⊗ S4 of Spin(4) descends to a representation of SO(4)
and moreover
S+ (V ) ⊗C S+ (V ) ∼
= (Λ0 (V ) ⊕ Λ2+ (V )) ⊗R C
as SO(4) representations.
(b) Show that S+ (V ) ∼
= S̄+ (V ) as Spin(4) modules.
(c)The above isomorphism defines an element φ ∈ S+ (V ) ⊗C S+ (V ). Show that via the
correspondence at (a) the isomorphism φ spans Λ0 (V ). ⊔
⊓

12.1.9 Dirac bundles

In this subsection we discuss a distinguished class of Clifford bundles which is both fre-
quently encountered in applications and it is rich in geometric information. We will touch
only the general aspects. The special characteristics of the most important concrete exam-
ples are studied in some detail in the following section. In the sequel all Clifford bundles
will be assumed to be complex.
Definition 12.1.64. Let E → M be a Clifford bundle over the oriented Riemann manifold
(M, g). We denote by c : Ω1 (M ) → End(E) the Clifford multiplication.
(a) A Dirac structure on E is a pair (h, ∇) consisting of a Hermitian metric h on E,
and a Clifford connection, i.e., a connection ∇ on E compatible with h and satisfying the
following conditions.

(a1) For any α ∈ Ω1 (M ), the Clifford multiplication by α is a skew-Hermitian endomor-

phism of E.

(a2) For any α ∈ Ω1 (M ), X ∈ Vect (M ), u ∈ C ∞ (E)

∇X (c(α)u) = c(∇M
X α)u + c(α)(∇X u),

where ∇M denotes the Levi–Civita connection on T ∗ M . (This condition means that

the Clifford multiplication is covariant constant).

A pair (Clifford bundle, Dirac structure) will be called a Dirac bundle.

(b) A Z2 -grading on a Dirac bundle (E, h, ∇) is a Z2 grading of the underlying Clifford
structure E = E0 ⊕ E1 , such that h = h0 ⊕ h1 , and ∇ = ∇0 ⊕ ∇1 , where hi (respectively
∇i ) is a metric (respectively a metric connection) on Ei . ⊔
⊓

The next result addresses the fundamental consistency question: do there exist Dirac
bundles?
650 CHAPTER 12. DIRAC OPERATORS

Proposition 12.1.65. Let E → M be a Clifford bundle over the oriented Riemann man-
ifold (M, g). Then there exist Dirac structures on E.

Proof. Denote by D the (possible empty) family of Dirac structure on E. Note that if
(hi , ∇i ) ∈ D, i = 1, 2, and f ∈ C ∞ (M ), then

(f h1 + (1 − f )h2 , f ∇1 + (1 − f )∇2 ) ∈ D.

This elementary fact shows that the existence of Dirac structures is essentially a local
issue: local Dirac structures can be patched-up via partitions of unity.
Thus, it suffices to consider only the case when M is an open subset of Rn , and E is
a trivial vector bundle. On the other hand, we cannot assume that the metric g is also
trivial (Euclidean) since the local obstructions given by the Riemann curvature cannot be
removed. We will distinguish two cases.
A. n = dim M is even. The proof will be completed in three steps.
Step 1. A special example. Consider the (complex) spinor module (representation)

c : Cln → End (Sn ),

We can assume that c is self-adjoint, i.e.,

c(u† ) = c(u)∗ , ∀u ∈ Cln .

We will continue to denote by c the restriction of the Clifford multiplication to Spin(n) ֒→

Cln .
Fix a global, oriented, orthonormal frame (ei ) of T M , and denote by (ej ) its dual
coframe. Denote by ω = (ωij ) the connection 1-form of the Levi–Civita connection on
T ∗ M with respect to this moving frame, i.e.,
X
∇ej = ωej = ωij ⊗ ei , ω ∈ Ω1 (M ) ⊗ so(n).
i

Using the canonical isomorphism ρ∗ : spin(n) → so(n) we define

1 X ij i j
ω̃ := ρ−1
∗ (ω) = − ω e · e ∈ Ω1 (M ) ⊗ spin(n).
2
i<j

This defines a connection ∇S on the trivial vector bundle SM = Sn × M given by

1X
∇S u = du − ωij ⊗ c(ei ) · c(ej )u, ∀u ∈ C ∞ (SM ).
2
i<j

The considerations in §10.1.5 show that ∇S is indeed a Clifford connection so that SM is

a Dirac bundle.
Step 2. Constructing general Dirac bundles. Fix a Dirac bundle (E, hE , ∇E ) over M . For
any Hermitian vector bundle (W, hW ) equipped with a Hermitian connection ∇W we can
construct the tensor product E ⊗ W equipped with the metric hE ⊗ hW and the product
12.1. THE STRUCTURE OF DIRAC OPERATORS 651

connection ∇E⊗W . These two data define a Dirac structure on E ⊗ W (Exercise 12.1.68
at the end of this subsection).
The representation theory of the Clifford algebra Cln with n even shows that any
Clifford bundle over M must be a twisting SM ⊗ W of the spinor bundle SM . This
completes the proof of the proposition when dim M is even.

Step 3 Conclusion. The odd case is dealt with similarly using the different representation
theory of the Cl2k+1 . Now instead of one generating model S there are two, but the proof
is conceptually identical. The straightforward details are left to the reader. ⊔
⊓

Denote by (E, h, ∇) a Dirac bundle over the oriented Riemann manifold (M, g). There
exists a Dirac operator on E canonically associated to this structure given by
∇ c
D = c ◦ ∇ : C ∞ (E) → C ∞ (T ∗ M ⊗ E) → C ∞ (E).

A Dirac operator associated to a Dirac structure is said to be a geometric Dirac operator.

Proposition 12.1.66. Any geometric Dirac operator is formally self-adjoint.

Proof. The assertion in the above proposition is local so we can work with local orthonor-
mal moving frames. Fix x0 ∈ M , and denote by (xi ) a collection of normal coordinates
near x0 . Set ei = ∂xi . Denote by (ei ) the dual coframe of (ei ). If (h, ∇) is a Dirac structure
on the Clifford bundle E, then at x0 the associated Dirac operator can be described as
X
D= c(ei )∇i (∇i = ∇ei ).
i

We deduce
D ∗ = (∇i )∗ c(ei )∗ .
Since the connection ∇ is compatible with h and div g (ei ) |x0 = 0, we deduce

(∇i )∗ = −∇i .

Since the Clifford multiplication is skew-Hermitian, c(ei )∗ = −c(ei ), we deduce that at x0

we have X X
D∗ = ∇i ◦ c(ei ) = [∇i , c(ei )] + D.
i i

On the other hand, the Clifford multiplication is covariant constant, and (∇M i
i e ) |x0 = 0
so we conclude
[∇i , c(ei )] |x0 = 0 ∀i. ⊔
⊓
Let D be a geometric operator associated to the Dirac bundle (E, h, ∇). By definition,
D 2 is a generalized Laplacian, and according to Proposition 10.1.34 we must have an
equality of the form
D2 = ∇ ˜ ∗∇
˜ + R,
˜ is a connection on E, and R is the Weitzenböck remainder which is an endo-
where ∇
˜ = ∇, and this remainder can be given
morphism of E. For geometric Dirac operators, ∇
652 CHAPTER 12. DIRAC OPERATORS

a very explicit description with remarkable geometric consequences. To formulate it we

need a little bit of foundational work.
Let (E, h, ∇) be a Dirac bundle over the oriented Riemann manifold (M, g). Denote
by Cl(T ∗ M M ) → M the bundle of Clifford algebras generated by (T ∗ M, g).
The curvature F (∇) of ∇ is a section of Λ2 T ∗ M ⊗ End (E). Using the quantization
map q : Λ• T ∗ M → Cl(T ∗ M ) we get a section

q(F ) ∈ Cl(T ∗ M ) ⊗ End (E).

On the other hand, the Clifford multiplication c : Cl(T ∗ M ) → End (E) defines a linear
map
Cl(T ∗ M ) ⊗ End (E) → End(E), ω ⊗ T 7→ c(ω) ◦ T.
This map associates to the element q(F ) an endomorphism of E which we denote by c(F ).
If (ei ) is a local, oriented, orthonormal moving frame for T ∗ M , then we can write
X
F (∇) = ei ∧ ej ⊗ Fij ,
i<j

and X 1X
c(F ) = c(ei )c(ej )Fij = c(ei )c(ej )Fij .
2
i<j i,j

Theorem 12.1.67 (Bochner-Weitzenböck)). Let D be the geometric Dirac operator

associated to the Dirac bundle (E, h, ∇) over the oriented Riemann manifold (M, g). Then

D 2 = ∇∗ ∇ + c(F (∇)).

Proof. Fix x ∈ M , and then choose an oriented, local orthonormal moving frame (ei ) of
T M near x such that
[ei , ej ] |x = (∇M ei ) |x = 0,
where ∇M denotes the Levi–Civita connection. Such a choice is always possible because
the torsion of the Levi–Civita connection is zero. Finally, denote by (ei ) the dual coframe
of (ei ). Then  
X X
D 2 |x = c(ei )∇i  c(ej )∇j  .
i j

Since [∇i , c(ej )] |x = 0 we deduce

X X X
D 2 |x = c(ei )c(ej )∇i ∇j = − ∇2i + c(ei )c(ej )∇i ∇j
i,j i i6=j
X X X X
=− ∇2i + c(ei )c(ej )[∇i , ∇j ] = − ∇2i + c(ei )c(ej )Fij (∇).
i i<j i i<j

We want to emphasize again the above equalities hold only at x. The theorem now follows
by observing that !
X
∗ 2
(∇ ∇) |x = − ∇i |x . ⊔
⊓
i
12.2. FUNDAMENTAL EXAMPLES 653

Exercise 12.1.68. Let (E, h, ∇) be a Dirac bundle over the oriented Riemann manifold
(M, g) with associated Dirac operator D. Consider a Hermitian bundle W → M and a
connection ∇W compatible with the Hermitian metric hW .
(a) Show that (E ⊗ W, h ⊗ hW , ∇ˆ = ∇ ⊗ 1W + 1E ⊗ ∇W ) defines a Dirac structure on
E ⊗ W in which the Clifford multiplication by α ∈ Ω1 (M ) is defined by

c(α)(e ⊗ w) = (c(α)e) ⊗ w, e ∈ C ∞ (E), w ∈ C ∞ (W ).

We denote by DW the corresponding geometric Dirac operator.

(b) Denote by cW (F (∇W )) the endomorphism of E ⊗ W defined by the sequence
q c
F (∇W ) ∈ C ∞ (Λ2 T ∗ M ⊗ End (W )) 7→ C ∞ (Cl(T ∗ M ) ⊗ End (W )) 7→ C ∞ (End (E ⊗ W )).

Show that
2
DW ˆ ∗∇
=∇ ˆ + c(F (∇)) + cW (F (∇W )). ⊔
⊓

12.2 Fundamental examples

This section is entirely devoted to the presentation of some fundamental examples of
Dirac operators. More specifically we will discuss the Hodge–DeRham operator, the Dol-
beault operator the spin and spinc Dirac. We will provide more concrete descriptions of
the Weitzenböck remainder presented in Subsection 10.1.9 and show some of its uses in
establishing vanishing theorems.

12.2.1 The Hodge–DeRham operator

Let (M, g) be an oriented Riemann manifold and set

Λ•C T ∗ M := Λ• T ∗ M ⊗ C.

For simplicity we continue to denote by Ω• (M ) the space of smooth differential forms on

M with complex coefficients. We have already seen that the Hodge–DeRham operator

d + d∗ : Ω• (M ) → Ω• (M )

is a Dirac operator. In fact, we will prove this operator is a geometric Dirac operator.
Continue to denote by g the Hermitian metric induced by the metric g on the com-
plexification Λ•C T ∗ M . We will denote by ∇g all the Levi–Civita connection on the tensor
bundles of M . When we want to be more specific about which Levi–Civita connection we
are using at a given moment we will indicate the bundle it acts on as a superscript. E.g.,
∗
∇T M is the Levi–Civita connection on T ∗ M .

Proposition 12.2.1. The pair (g, ∇g ) defines a Dirac structure on the Clifford bundle
ΛC T ∗ M and d + d∗ is the associated Dirac operator.
654 CHAPTER 12. DIRAC OPERATORS

Proof. In Subsection 4.1.5 we have proved that d can be alternatively described as the
composition
∇ ε
C ∞ (Λ• T ∗ M ) → C ∞ (T ∗ M ⊗ Λ• T ∗ M ) → C ∞ (Λ• T ∗ M ),
where ε denotes the exterior multiplication map. Thus
d + d∗ = ε ◦ ∇ + ∇∗ ◦ ε∗ .
If X1 , · · · , Xn is a local orthonormal frame of T M , and θ 1 , · · · , θ n is its dual coframe, then
for any ordered multi-index I we have
X
ε∗ (θ I ) = iXj θ I .
j

Thus, for any ω ∈ C ∞ (Λ• T ∗ M )

we have
X 
(∇∗ ◦ ε∗ )ω = − ∇Xk iXk + div g (Xk )iXk ω.
k

Fortunately, we have the freedom to choose the frame (Xk ) in any manner we find conve-
nient.
Fix an arbitrary point x0 ∈ M and choose (Xk ) such that, at x0 , we have Xk =
∂xk , where (xk ) denotes a collection of normal coordinates near x0 . With such a choice
div g (Xk ) = 0 at x0 , and thus
X
d∗ ω |x0 = − i∂k ∇∂k ω.
k

This shows d + d∗ can be written as the composition

∇ c
C ∞ (Λ• T ∗ M ) → C ∞ (T ∗ M ⊗ Λ• T ∗ M ) → C ∞ (Λ• T ∗ M ),
where c denotes the usual Clifford multiplication on an exterior algebra. We leave the
reader to verify that the Levi–Civita connection on Λ• T ∗ M is indeed a Clifford connection,
i.e., the Clifford multiplication is covariant constant. Passing to the complexification
Λ•C T ∗ M we deduce that d + d∗ is a geometric Dirac operator. ⊔
⊓

We want to spend some time elucidating the structure of the Weitzenböck remainder.
First of all, we need a better description of the curvature of ∇g viewed as a connection on
Λ• T ∗ M .
Denote by R the Riemann curvature tensor, i.e., the curvature of the Levi–Civita
connection
∇g : C ∞ (T M ) → C ∞ (T ∗ M ⊗ T M ).
Thus R is a bundle morphism
R : C ∞ (T M ) → C ∞ (Λ2 T ∗ M ⊗ T M ).
We have a dual morphism
R̃ : C ∞ (T ∗ M ) → C ∞ (Λ2 T ∗ M ⊗ T ∗ M ),
uniquely determined by the equality
(R̃(Y, Z)α)(X) = −α(R(Y, Z)X) ∀α ∈ Ω1 (M ) X, Y, Z ∈ Vect (M ).
12.2. FUNDAMENTAL EXAMPLES 655

∗M
Lemma 12.2.2. R̃ is the curvature of the Levi–Civita connection ∇T .

Proof. The Levi–Civita connection on T ∗ M is determined by the equalities

∗M
(∇TZ α)(X) = Z · α(X) − α(∇TZ M X) ∀α ∈ Ω1 (M ), X, Z ∈ Vect (M ).

Derivating along Y ∈ Vect (M ) we get

(∇Y ∇Z α)(X) = Y · (∇Z α)(X) − (∇Z α)(∇Y X)

= Y · Z · α(X) − Y · α(∇Z X) − Z · α(∇Y X) − α(∇Z ∇Y X).

Similar computations give ∇Z ∇Y α and ∇[Y,Z] and we get
∗M
(RT (Y, Z)α)(X) = −α(RT M (Y, Z)X). ⊔
⊓
∗
The Levi–Civita connection ∇g = ∇T M extends as an even derivation to a connection on
Λ• T ∗ M . More precisely, for every X ∈ Vect (M ), and any α1 , . . . , αk ∈ Ω1 (M ), we define

∇gX (α1 ∧ · · · ∧ αk ) = (∇gX α1 ) ∧ · · · ∧ αk + · · · + α1 ∧ · · · ∧ (∇gX αk ).

Lemma 12.2.3. The curvature of the Levi–Civita connection on Λ• T ∗ M is defined by

•T ∗ M
RΛ (X, Y )(α1 ∧ · · · ∧ αk )
= (R̃(X, Y )α1 ) ∧ · · · ∧ αk + · · · + α1 ∧ · · · ∧ (R̃(X, Y )αk ),

for any vector fields X, Y , and any 1-forms α1 , . . . , αk . ⊔

⊓

Exercise 12.2.4. Prove the above lemma. ⊔

⊓
• ∗
The Weitzenböck remainder of the Dirac operator d + d∗ is c(RΛ T M ). To better
understand its action we need to pick a local, oriented, orthonormal moving frame (ei ) of
T M . We denote by (ei ) its dual coframe. The Riemann curvature tensor can be expressed
as X
R= ei ∧ ej Rij ,
i<j

where Rij is the skew-symmetric endomorphism Rij = R(ei , ej ) : T M → T M . Thus

•T ∗M
X •T ∗M 1X • ∗
c(RΛ )= c(ei )c(ej )RΛ (ei , ej ) = c(ei )c(ej )RΛ T M (ei , ej ),
2
i<j i,j

where c(ej ) = e(ej ) − i(ej ).

Note that since both ∇∗ ∇, and (d + d∗ )2 preserve the Z-grading of Λ•C T ∗ M , so does
the Weitzenböck remainder, and consequently it must split as
• T ∗M
c(RΛ ) = ⊕k≥0 Rk .

Since R0 ≡ 0, the first interesting case is R1 . To understand its form, pick normal
coordinates (xi ) near x0 ∈ M . Set ei = ∂xi |x0 ∈ Tx0 M and ei = dxi |x0 ∈ Tx∗0 M .
656 CHAPTER 12. DIRAC OPERATORS

At x0 the Riemann curvature tensor R has the form

X
R= ek ∧ eℓ R(ek , eℓ )
k<ℓ

i e =R
where R(ek , eℓ )ej = Rjkℓ i ijkℓ ei . Using Lemma 12.2.2 we get

R̃(ek , eℓ )ej = Rijkℓ ei .

Using this in the expression of R1 at x0 we get

X 1 XX
R1 ( αj ej ) = αj Rijkℓ c(ek )c(eℓ )ei .
2
j k,ℓ i,j

We need to evaluate the Clifford actions in the above equality. We have

c(eℓ )ei = eℓ ∧ ei − δiℓ ,

and
c(ek )c(eℓ )ei = ek ∧ eℓ ∧ ei − δiℓ ek − δkℓ ei + δik eℓ .
Hence
X 1 X
R1 ( αj ej ) = αj Rijkℓ (ek ∧ eℓ ∧ ei − δiℓ ek − δkℓ ei + δik eℓ ).
2
j i,j,k,ℓ

Using the first Bianchi identity we deduce that

X X
Rijkℓ ek ∧ eℓ ∧ ei = − Rjikℓei ∧ ek ∧ eℓ = 0, ∀j.
i,k,ℓ i,k,ℓ

Because of the skew-symmetry Rijkℓ = −Rijℓk we conclude that

X
αj Rijkℓ δkℓ ei = 0.
ijkℓ

Hence
X 1 X
R1 ( αj ej ) = αj Rijkℓ (δik eℓ − δiℓ ek )
2
j i,j,k,ℓ

1X 1X X X
= αj Rijiℓ eℓ − αj Rijki ek = αj Rijik ek = αj Ricjk ek ,
2 2
i,j,ℓ i,j,k i,j,k jk
P
where Ricjk = i Rijik denotes the Ricci tensor at x0 . Hence

R1 = Ric . (12.2.1)

In the above equality, the Ricci tensor Ric is regarded (via the metric duality) as a
self-adjoint endomorphism of T ∗ M .
The identity (12.2.1) has a beautiful consequence.
12.2. FUNDAMENTAL EXAMPLES 657

Theorem 12.2.5 (Bochner). Let (M, g) be a compact, connected, oriented Riemann man-
ifold.
(a) If the Ricci tensor is non-negative definite, then b1 (M ) ≤ dim M .
(b) If Ricci tensor is non-negative definite, but it is somewhere strictly positive definite
then b1 (M ) = 0.
(Recall that b1 (M ) denotes the first Betti number of M ).

The above result is truly remarkable. The condition on the Ricci tensor is purely
local, but it has global consequences. We have proved a similar result using geodesics (see
Myers Theorem, Subsection 5.2.2 , 6.2.4 and 6.2.5). Under more restrictive assumptions
on the Ricci tensor (uniformly positive definite) one deduces a stronger conclusion namely
that the fundamental group is finite. If the uniformity assumption is dropped then the
conclusion of the Myers theorem no longer holds (think of the flat torus).
This result gives yet another explanation for the equality H 1 (G) = 0, where G is
a compact semisimple Lie group. Recall that in this case the Ricci curvature is 41 ×
{the Killing metric}.

Proof. (a) Let ∆1 denote the metric Laplacian

∆1 = dd∗ + d∗ d : Ω1 (M ) → Ω1 (M ).

Hodge theory asserts that b1 (M ) = dim ker ∆1 so that, to find the first Betti number, we
need to estimate the “number” of solutions of the elliptic equation

∆1 η = 0, η ∈ Ω1 .

Using the Bochner-Weitzenböck theorem and the equality (12.2.1) we deduce that if η ∈
ker ∆1 , then
∇∗ ∇η + Ric η = 0, on M.
Taking the L2 -inner product by η, and then integrating by parts, we get
Z Z
2
|∇η| dvg + (Ric η, η)dvg = 0. (12.2.2)
M M

Since Ric is non-negative definite we deduce ∇η = 0, so that any harmonic 1-form must be
covariant constant. In particular, since M is connected, the number of linearly independent
harmonic 1-forms is no greater than the rank of T ∗ M which is dim M .
(b) Using the equality (12.2.2) we deduce that any harmonic 1-form η must satisfy

(Ric(x)ηx , ηx )x = 0 ∀x ∈ M.

If the Ricci tensor is positive at some x0 ∈ M , then η(x0 ) = 0. Since η is also covariant
constant, and M is connected, we conclude that η ≡ 0. ⊔
⊓

Remark 12.2.6. For a very nice survey of some beautiful applications of this technique
we refer to [13]. ⊔
⊓
658 CHAPTER 12. DIRAC OPERATORS

12.2.2 The Hodge–Dolbeault operator

This subsection introduces the reader to the Dolbeault operator which plays a central
role in complex geometry. Since we had almost no contact with this beautiful branch
of geometry we will present only those aspects concerning the “Dirac nature” of these
operators. To define this operator we need a little more differential geometric background.
Definition 12.2.7. (a) Let E → M be a smooth real vector bundle over the smooth
manifold M . An almost complex structure on E is an endomorphism J : E → E such
that J 2 = −1E .
(b) An almost complex structure on a smooth manifold M is an almost complex structure
J on the tangent bundle. An almost complex manifold is a pair (manifold, almost complex
structure). ⊔
⊓

Note that any almost complex manifold is necessarily even dimensional and orientable,
so from the start we know that not any manifold admits almost complex structures. In
fact, the existence of such a structure is determined by topological invariants finer than
the dimension and orientability.
Example 12.2.8. (a) A complex manifold M is almost complex. Indeed, the manifold
M is locally modelled by Cn , and the transition maps are holomorphic maps Cn → Cn .
The multiplication by i defines a real endomorphism on R2n ∼ = Cn , and the differential
of a holomorphic map Cn → Cn commutes with the endomorphism of T Cn given by
multiplication by i in the fibers. Hence, this endomorphism induces the almost complex
structure on T M .
(b) For any manifold M , the total space of its tangent bundle T M can be equipped in
many different ways with an almost complex structure.
To see this, denote by E the tangent bundle of T M , E = T (T M ). The bundle E has
a natural subbundle V, the vertical subbundle, which is the kernel of the differential of the
canonical projection π : T M → M .
If x ∈ M , v ∈ Tx M , then a tangent vector in T(v,x) T M is vertical if and only if it
is tangent to a smooth path which entirely the fiber Tx M ⊂ T M . Note that we have a
canonical identification
V(v,x) ∼
= Tx M.
Fix a Riemann metric h on the manifold T M , and denote by H ⊂ E, the subbundle of E
which is the orthogonal complement of V in E. For every (v, x) ∈ T M , the differential of
π induces an isomorphism
π∗ : H(v,x) → Tx M ∼
= Vv,x .
Thus, π∗ induces a bundle isomorphism A : H → V. Using the direct sum decomposition
E = H ⊕ V we define the endomorphism J of E in the block form
 
0 −A−1
J= . ⊔
⊓
A 0

Exercise 12.2.9. Prove that a smooth manifold M of dimension 2n admits an almost

complex structure if an only if there exists a 2-form ω ∈ Ω2 (M ) such that the top exterior
power ω n is a volume form on M . ⊔
⊓
12.2. FUNDAMENTAL EXAMPLES 659

Let (M, J) be an almost complex manifold. Using the results of Subsection 2.2.5 we
deduce that the complexified tangent bundle T M ⊗ C splits as

T M ⊗ C = (T M )1,0 ⊗ (T M )0,1 .

The complex bundle T M 1,0 is isomorphic (over C) with (T M, J).

By duality, the operator J induces an almost complex structure in the cotangent bundle
T ∗ M , and we get a similar decomposition

T ∗ M ⊗ C = (T ∗ M )1,0 ⊕ (T ∗ M )0,1 .

In turn, this defines a decomposition

M
Λ•C T ∗ M = Λp,q T ∗ M.
p,q

We set Ωp,q (M ) := C ∞ (Λp,q T ∗ M ).

Example 12.2.10. Let M be a complex manifold. If (z j = xj +iy j ) are local holomorphic
coordinates on M , then (T M )1,0 is generated (locally) by the complex tangent vectors
1

∂z j = ∂xj − i∂yj ,
2
while (T ∗ M )1,0 is locally generated by the complex 1-forms dz j = dxj + iy j .
The bundle (T M )1,0 is also known as the holomorphic tangent space, while (T ∗ M )1,0 is
called the holomorphic cotangent space. The space (T ∗ M )0,1 is called the anti-holomorphic
cotangent space, and it is locally spanned by the complex 1-forms dz̄ j = dxj − iy j .
Note that any (p, q)-form η ∈ Ωp,q (M ) can be locally described as
X
η= ηA,B dz A ∧ dz̄ B , ηA,B ∈ C ∞ (M, C),
|A|=p,|B|=q

where we use capital Latin letters A, B, C . . . to denote ordered multi-indices, and for each
such index A we set
dz A =: dz A1 ∧ · · · ∧ dz Ap .
The differential form dz̄ B is defined similarly. ⊔
⊓

Exercise 12.2.11. Let (M, J) be an arbitrary smooth almost complex manifold.

(a) Prove that

dΩp,q (M ) ⊂ Ωp+2,q−1 (M ) ⊕ Ωp+1,q (M ) ⊕ Ωp,q+1 (M )Ωp−1,q+2 (M ).

(b) Show that if M is a complex manifold then

dΩp,q (M ) ⊂ Ωp+1,q (M ) ⊕ Ωp,q+1 (M ). (12.2.3)

(The converse is also true, and it is known as the Newlander-Nirenberg theorem. Its proof
is far from trivial. For details we refer to the original paper [103]). ⊔
⊓
660 CHAPTER 12. DIRAC OPERATORS

Definition 12.2.12. An almost complex structure on a smooth manifold is called in-

tegrable if it can derived from a holomorphic atlas, i.e. an atlas in which the transition
maps are holomorphic. ⊔
⊓

Thus, the Newlander-Nirenberg theorem mentioned above states that the condition
(12.2.3) is necessary and sufficient for an almost complex structure to be integrable.

Exercise 12.2.13. Assuming the Newlander-Nirenberg theorem prove that an almost

complex structure J on the smooth manifold M is integrable if and only if the Nijenhuis
tensor N ∈ C ∞ (T ∗ M ⊗2 ⊗ T M ) defined by

N (X, Y ) = [JX, JY ] − [X, Y ] − J[X, JY ] − J[JX, Y ] ∀X, Y ∈ Vect (M )

vanishes identically. ⊔
⊓

In the sequel we will exclusively consider only complex manifolds. Let M be such a
manifold. The above exercise shows that the exterior derivative

d : ΩkC (M ) → Ωk+1
C (M )

splits as a direct sum d = ⊕p+q=k dp,q , where

dp,q = {d : Ωp,q (M ) → Ωp+1,q (M ) ⊕ Ωp,q+1 (M )}.

The component Ωp,q → Ωp+1,q is denoted by ∂ = ∂ p,q , while the component Ωp,q → Ωp,q+1
is denoted by ∂¯ = ∂¯p,q .

Example 12.2.14. In local holomorphic coordinates (z i ) the action of the operator ∂¯ is

described by
 
X X ∂ηAB A
∂¯  ηAB dz A ∧ dz̄ B  = (−1)|A| dz ∧ dz̄ j ∧ dz̄ B . ⊔
⊓
∂ z̄ j
A,B j,A,B

It is not difficult to see that

∂¯p,q+1 ◦ ∂¯p,q = 0 ∀p, q.

In other words, for any 0 ≤ p ≤ dimC M , the sequence

∂¯ ∂¯
0 → Ωp,0 (M ) −→ Ωp,1 (M ) −→ · · ·

is a cochain complex known as the p-th Dolbeault complex of the complex manifold M .
Its cohomology groups are denoted by H∂p,q
¯ (M ).

Lemma 12.2.15. The Dolbeault complex is an elliptic complex.

12.2. FUNDAMENTAL EXAMPLES 661

Proof. The symbol of ∂¯p,q is very similar to the symbol of the exterior derivative. For any
x ∈ M and any ξ ∈ T ∗ X

σ(∂¯p,q )(ξ) : Λp,q Tx∗ M → Λp,q+1 Tx∗ M

is (up to a multiplicative constant) the (left) exterior multiplication by ξ 0,1 , where ξ 0,1
denotes the (0, 1) component of ξ viewed as an element of the complexified tangent space.
More precisely,
1
ξ 0,1 = (ξ + iJ0 ξ),
2
where J0 : Tx M → Tx M denotes the canonical complex structure induced on Tx∗ M by the
∗ ∗

holomorphic charts. The sequence of symbols is the cochain complex

1⊗(−1)p ξ 0,1 ∧ 1⊗(−1)p ξ 0,1 ∧

0 → Λp,0 ⊗ Λ0,0 −→ Λp,0 ⊗ Λp,0 ⊗ Λ0,1 −→ ··· .

This complex is the (Z-graded) tensor product of the trivial complex

1
0 → Λp,0 → Λp,0 → 0,

with the Koszul complex

(−1)p ξ 0,1 ∧ (−1)p ξ 0,1 ∧

0 → Λ0,0 −→ Λ0,1 −→ ··· .

Since ξ 0,1 6= 0, for any ξ 6= 0, the Koszul complex is exact (see Subsection 7.1.3) . This
proves that the Dolbeault complex is elliptic. ⊔
⊓

To study the Dirac nature of this complex we need to introduce a Hermitian metric
h on T M . Its real part is a Riemann metric g on M , and the canonical almost complex
structure on T M is a skew-symmetric endomorphism with respect to this real metric. The
associated 2-form Ωh = −Im h is nondegenerate in the sense that Ωn (n = dimC M ) is
a volume form on M . According to the results of Subsection 2.2.5 the orientation of M
defined by Ωn coincides with the orientation induced by the complex structure.
We form the Hodge–Dolbeault operator

∂¯ + ∂¯∗ : Ωp,• (M ) → Ωp,• (M ).

√
Proposition 12.2.16. The Hodge–Dolbeault operator 2(∂¯ + ∂¯∗ ) is a Dirac operator.

Proof. We need to show that

2 1
¯
σ(∂)(ξ) ¯
− σ(∂)(ξ)∗
= − |ξ|2 1, ∀ξ ∈ T ∗ M.
2
Denote by J the canonical complex structure on T ∗ M and set η = Jξ. Note that ξ ⊥ η
and |ξ| = |η|. Then

¯ 1 1
σ(∂)(ξ) = (−1)p e(ξ + iη) = (e(ξ) + ie(η)),
2 2
662 CHAPTER 12. DIRAC OPERATORS

¯
where as usual e(·) denotes the (left) exterior multiplication. The adjoint of σ(∂)(ξ) is

¯ 1
(−1)p σ(∂)(ξ)∗
= ξ∗ −iη ∗ ),
2
where ξ ∗ (respectively η ∗ ) denotes the metric dual of ξ (respectively η), and ξ ∗ denotes
the contraction by ξ ∗ . We set

c(ξ) := e(ξ) − ξ ∗ , c̃(η) := e(η) + η ∗ ,

and we deduce

2 1
¯
σ(∂)(ξ) ¯
− σ(∂)(ξ)∗
= { e(ξ) + ie(η) − ξ ∗ +iη ∗ }2
4
1

2 1 2
= e(ξ) − ξ ∗ + i e(η) + η ∗ = c(ξ) + ic̃(η)
4 4
1

= c(ξ)2 − c̃(η)2 + i c(ξ)c̃(η) + c(η)c(ξ) .
4
Note that

c(ξ)2 = − e(ξ)ξ ∗ +ξ ∗ e(ξ) = −|ξ|2 ,
and
c̃(η)2 = e(η)η ∗ +η ∗ e(η) = |η|2 .
On the other hand, since ξ ⊥ η, we deduce as above that

c(ξ)c̃(η) + c̃(η)c(ξ) = 0.

(Verify this!) Hence

 2 1 1
¯
σ(∂)(ξ) ¯
− σ(∂)(ξ)∗
= − (|ξ|2 + |η|2 ) = − |η|2 . ⊔
⊓
4 2
A natural question arises as to when the above operator is a geometric Dirac operator.
Note first that the Clifford multiplication is certainly skew-adjoint since it is the symbol
of a formally self-adjoint operator. Thus all we need to inquire is when the Clifford
multiplication is covariant constant. Since

(−1)p

c(ξ) = √ e(ξ) + ie(η) − ξ ∗ −iη ∗ ,
2
where η = Jξ, we deduce the Clifford multiplication is covariant constant if ∇g J = 0.
Definition 12.2.17. Let M be a complex manifold, and h a Hermitian metric on T M
(viewed as a complex bundle). Then h is said to be a Kähler metric if ∇J = 0, where ∇
denotes the Levi–Civita connection associated to the Riemann metric Re h, and J is the
canonical almost complex structure on T M . A pair (complex manifold, Kähler metric) is
called a Kähler manifold. ⊔
⊓
12.2. FUNDAMENTAL EXAMPLES 663

Exercise 12.2.18. Let (M, J) be an almost complex manifold and h a Hermitian metric
on T M . Let Ω = −Im h. Using the exercise 12.2.13 show that if dΩh = 0, then the
almost complex is integrable, and the metric h is Kähler.
Conversely, assuming that J is integrable, and h is Kähler show that dΩ = 0. ⊔
⊓

We see that on a Kähler manifold the above Clifford multiplication is covariant con-
stant. In fact, a more precise statement is true.

Proposition 12.2.19. Let M be a complex manifold and h a Kähler metric on T M . Then

the Levi–Civita induced connection on Λ0,• T ∗ M is a Clifford connection with√respect to
the above Clifford multiplication, and moreover, the Hodge–Dolbeault operator 2(∂¯ + ∂¯∗ )
is the geometric Dirac operator associated to this connection. ⊔
⊓

Exercise 12.2.20. Prove the above proposition. ⊔

⊓

Example 12.2.21. Let (M, g) be an oriented 2-dimensional Riemann manifold (surface).

The Hodge ∗-operator defines an endomorphism

∗ : T M → T M,

satisfying ∗2 = −1T M , i.e., the operator ∗ is an almost complex structure on M . Us-

ing the Exercise 12.2.18 we deduce this almost complex structure is integrable since, by
dimensionality
dΩ = 0,

where Ω is the natural 2-form Ω(X, Y ) = g(∗X, Y ) X, Y ∈ Vect (M ). This complex

structure is said to be canonically associated to the metric. ⊔
⊓

Example 12.2.22. Perhaps the favorite example of Kähler manifold is the complex
projective space Cn . To describe this structure consider the tautological line bundle
L1 → CPn . It can be naturally viewed as a subbundle of the trivial bundle Cn+1 → CPn .
Denote by h0 the canonical Hermitian metric on Cn+1 , and by ∇0 the trivial connec-
tion. If we denote by P : Cn+1 → L1 the orthogonal projection, then ∇ = P ◦ ∇0 |L1
defines a connection on L1 compatible with h1 = h |L1 . Denote by ω the first Chern form
associated to this connection
i
ω= F (∇),
2π
and set
hF S (X, Y )x := −ωx (X, JY ) + iω(X, Y ) ∀x ∈ M, X, Y ∈ Tx M.

Then hF S is a Hermitian metric on CPn (verify!) called the Fubini-Study metric. It is

clearly a Kähler metric since (see the Exercise 12.2.18) dΩh = −dω = −dc1 (∇) = 0. ⊔
⊓

Exercise 12.2.23. Describe hF S in projective coordinates, and then prove that h is indeed
a Hermitian metric, i.e. it is positive definite). ⊔
⊓
664 CHAPTER 12. DIRAC OPERATORS

Remark 12.2.24. Any complex submanifold of a Kähler manifold is obviously Kähler.

In particular, any complex submanifold of CPn is automatically Kähler.
A celebrated result of Chow states that any complex submanifold of CPn is auto-
matically algebraic, i.e., it can be defined as the zero set of a family of homogeneous
polynomials. Thus, all complex nonsingular algebraic varieties admit a natural Kähler
structure. It is thus natural to ask whether there exist Kähler manifolds which are not
algebraic.
The answer is positive, and a very thorough resolution of this problem is contained in
the famous Kodaira embedding theorem which provides a simple necessary and sufficient
condition for a compact complex manifold to be algebraic.
For this work, Kodaira was awarded the Fields medal in 1954. His proofs rely essen-
tially on some vanishing results deduced from the Weitzenböck formulæ for the Dolbeault
operator ∂¯ + ∂¯∗ , and its twisted versions. A very clear presentation of this subject can be
found in the beautiful monograph [55]. ⊔
⊓

12.2.3 The spin Dirac operator

Like the Dolbeault operator, the spin Dirac operator exists only on manifolds with a bit
of extra structure. We will first describe this new structure.
Let (M n , g) be an n-dimensional, oriented Riemann manifold. In other words, the
tangent bundle T M admits an SO(n) structure so that it can be defined by an open cover
(Uα ), and transition maps
gαβ : Uαβ → SO(n)
satisfying the cocycle condition.
The manifold is said to spinnable if there exist smooth maps
g̃αβ : Uαβ → Spin(n)
satisfying the cocycle condition, and such that
ρ(g̃αβ ) = gαβ ∀α, β,
where ρ : Spin(n) → SO(n) denotes the canonical double cover. The collection g̃αβ as
above is called a spin structure. A pair (manifold, spin structure) is called a spin manifold.
Not all manifolds are spinnable. To understand what can go wrong, let us start with
a trivializing cover U = (Uα ) for T M , with transition maps gαβ , and such that all the
multiple intersection Uαβ···γ are contractible. In other words, U is a good cover.
Since each of the overlaps Uαβ is contractible, each map gαβ : Uαβ → SO(n) admits
at least one lift
g̃αβ : Uαβ → Spin(n).
From the equality ρ(g̃αβ g̃βγ g̃γα ) = gαβ gβγ gγα = 1 we deduce
ǫαβγ = g̃αβ g̃βγ g̃γα ∈ ker ρ = Z2 .
Thus, any lift of the gluing data gαβ to Spin(n) produces a degree 2 Čech cochain of the
trivial sheaf Z2 , namely the 2-cochain
(ǫ• ) : Uαβγ 7→ ǫαβγ .
12.2. FUNDAMENTAL EXAMPLES 665

Note that for any α, β, γ, δ such that Uαβγδ 6= ∅, we have

ǫβγδ − ǫαγδ + ǫαβδ − ǫαβγ = 0 ∈ Z2 .

In other words, ǫ• defines a Čech 2-cocycle, and thus defines an element in the Čech
cohomology group H 2 (M, Z2 ).
It is not difficult to see that this element is independent of the various choices: the
cover U, the gluing data gαβ , and the lifts g̃αβ . This element is intrinsic to the tangent
bundle T M . It is called the second Stiefel-Whitney class of M , and it is denoted by
w2 (M ). We see that if w2 (M ) 6= 0 then M cannot admit a spin structure. In fact, the
converse is also true.
Proposition 12.2.25. An oriented Riemann manifold M admits a spin structure if and
only if w2 (M ) = 0. ⊔
⊓

Exercise 12.2.26. Prove the above result. ⊔

⊓

Remark 12.2.27. The usefulness of the above proposition depends strongly on the ability
of computing w2 . This is a good news/bad news situation. The good news is that algebraic
topology has produced very efficient tools for doing this. The bad news is that we will not
mention them since it would lead us far astray. See [83] and [98] for more details. ⊔
⊓

Remark 12.2.28. The definition of isomorphism of spin structures is rather subtle (see
[95]). More precisely, two spin structures defined by the cocycles g̃•• and h̃•• are isomorphic
if there exists a collection εα ∈ Z2 ⊂ Spin(n) such that the diagram below is commutative
for all x ∈ Uαβ
Spin(n)
εα
w Spin(n)
u g̃ βα (x)
u h̃ (x)
βα

Spin(n)
εβ
w Spin(n)
The group H 1 (M, Z2 ) acts on Spin(M ) as follows. Take an element ε ∈ H 1 (M, Z2 )
represented by a Čech cocycle, i.e., a collection of continuous maps εαβ : Uαβ → Z2 ⊂
Spin(n) satisfying the cocycle condition

εαβ · εβγ · εγα = 1.

Then the collection ε•• · g̃•• is a Spin(n) gluing cocycle defining a spin structure that we
denote by ε · σ.
It is easy to check that the isomorphism class of ε · σ is independent of the various
choice, i.e, the Čech representatives for ε and σ. Clearly, the correspondence

H 1 (M, Z2 ) × Spin(M ) ∋ (ε, σ) 7→ ε · σ ∈ Spin(M )

defines a left action of H 1 (M, Z2 ) on Spin(M ). This action is transitive and free. ⊔
⊓

Exercise 12.2.29. Prove the statements in the above remark. ⊔

⊓
666 CHAPTER 12. DIRAC OPERATORS

Exercise 12.2.30. Describe the only two spin structures on S 1 . ⊔

⊓

Example 12.2.31. (a) A simply connected Riemann manifold M of dimension ≥ 5 is

spinnable if and only if every compact orientable surface embedded in M has trivial
normal bundle.
(b) A simply connected four-manifold M is spinnable if and only if the normal bundle NΣ
of any embedded compact, orientable surface Σ has even Euler class, i.e.,
Z
e(NΣ )
Σ

is an even integer.
(c) Any compact oriented surface is spinnable. Any sphere S n admits a unique spin
structure. The product of two spinnable manifolds is canonically a spinnable manifold.
(d) w2 (RPn ) = 0 if and only if n ≡ 3 (mod 4), while CPn admits spin structures if and
only if n is odd. ⊔
⊓

Let (M n , g) be a spin manifold. Assume the tangent bundle T M is defined by the

open cover (Uα ), and transition maps

gαβ : Uαβ → SO(n).

Moreover assume the spin structure is given by the lifts

g̃αβ : Uαβ → Spin(n).

As usual, we regard the collection gαβ as defining the principal SO(n)-bundle of oriented
frames of T M . We call this bundle PSO(M ) .
The collection g̃αβ defines a principal Spin(n)-bundle which we denote by PSpin(M ) .
We can regard PSO(M ) as a bundle associated to PSpin(M ) via ρ : Spin(n) → SO(n). Using
the unitary spinorial representation

∆n : Spin(n) → Aut (Sn )

we get a Hermitian vector bundle

S(M ) = PSpin(M ) ×∆n Sn

called the complex spinor bundle. Its sections are called (complex) spinors. Let us
pont out that when M is even dimensional, the spinor bundle is equipped with a natural
Z/2-grading
S(M ) = S+ (M ) ⊕ S− (M ).
From Lemma 12.1.31 we deduce that S(M ) is naturally a bundle of Cl(T M ) modules.
Then, the natural isomorphism Cl(T M ) ∼ = Cl(T ∗ M ) induces on S(M ) a structure of
Cl(T ∗ M )-module,
c : Cl(T M ) → EndC ( S(M ) ).
12.2. FUNDAMENTAL EXAMPLES 667

Moreover, when M is even dimensional, the above morphism is compatible with the Z/2-
gradings of Cl(T M ) and EndC ( S(M ) ) As it turns out, the bundle S(M ) has a natural
Dirac structure whose associated Dirac operator is the spin Dirac operator on M . We will
denote it by D.
Since ∆n is a unitary representation, we can equip the spinor bundle S(M ) with a
natural metric with respect to which the Clifford multiplication

c : Cl(T M ) → EndC ( S(M ) )

is self-adjoint, i.e.,
c(u† ) = c(u)∗ , ∀u ∈ C ∞ (Cl(T M )).
Since Cl(T M ) is locally generated as an algebra by Vect (M ), we can equivalently rewrite
the self-adjointness condition as

c(X) = −c(X)∗ , ∀X ∈ Vect (M ) ⊂ C ∞ ( Cl(T ) ).

All we now need is to describe a natural connection on S(M ) with respect to which the
Clifford multiplication is covariant constant.
We start with the Levi–Civita connection ∇g which we can regard as a connection on
the principal bundle PSO(M ) . Alternatively, ∇g can be defined by a collection of so(n)-
valued 1-forms ωα ∈ Ω1 (Uα ) ⊗ so(n) such that
−1 −1
ωβ = gαβ dgαβ + gαβ ωα gαβ on Uαβ .

Consider the canonical isomorphism of Lie algebras

ρ∗ : spin(n) → so(n).

Then the collection ω̃α = ρ−1 ˆ

∗ (ωα ) defines a connection ∇ on the principal bundle PSpin(M ) ,
and thus via the representation ∆n it defines a connection ∇ = ∇S on the spinor bundle
S(M ).
The above construction can be better visualized if we work in local coordinates. Choose
a local, oriented orthonormal frame (ei ) of T M |Uα , and denote by (ej ) is dual coframe.
The Levi–Civita connection has the form

∇ej = ek ⊗ ωkj
i
ei ,

j j
so that ωα = ek ⊗(ωki ), where for each k the collection (ωki )i,j is a skew-symmetric matrix.
−1
Using the concrete description of ρ∗ given in (12.1.3) we deduce
X 1X i 1X i k
ω̃α = − ek ⊗ ( ωkj ei ej ) = − ωkj e ⊗ ei ej .
2 4
k i<j i,j,k

Lemma 12.2.32. The connection ∇ ˆ is a Clifford connection on S(M ) so that (S(M ), ∇)

ˆ
is a Dirac bundle called the bundle of complex spinors.
668 CHAPTER 12. DIRAC OPERATORS

Proof. Use Lemma 12.1.31. ⊔

⊓

We denote by D = DM the geometric Dirac operator associated to the bundle of

complex spinors S. When the manifold M is even dimensional, this operator is graded.
We recall that this means that, with respect to the decomposition S = S+ ⊕ S− , the Dirac
operator D has the block form  
0 D6 ∗
D= ,
6
D 0
where D6 : C ∞ (S+ ) → C ∞ (S− ) is a first order operator whose symbol is given by the
Clifford multiplication.
We now want to understand the structure of the Weitzenböck remainder of the geo-
metric Dirac operator D. If
X
R= ek ∧ eℓ Rkℓ Rkℓ = (Rjkℓ
i
) = (Rijkℓ )
k<ℓ

is the Riemann curvature tensor, (in the above trivializations over Uα ) we deduce that the
ˆ is
curvature of ∇
X 1 XX k
R̃ = ek ∧ eℓ ⊗ ρ−1
∗ (Rkℓ ) = − e ∧ eℓ ⊗ Rijkℓ ei ej .
4
k<ℓ k<ℓ ij

From this we obtain

ˆ =− 1X
c(F (∇)) Rijkℓ c(ei )c(ej )c(ek )c(eℓ ).
8
ijkℓ

In the above sum, the terms corresponding to indices (i, j, k, ℓ) such that i = j or k = ℓ
vanish due to the corresponding skew-symmetry of the Riemann tensor. Thus, we can
write XX
ˆ = −1
c(F (∇)) Rijkℓ c(ei )c(ej )c(ek )c(eℓ )
8
i6=j k6=ℓ

Using the equalities c(ei )c(ej )+c(ej )c(ei )

= −2δij we deduce that the monomial c(ei )c(ej )
anti-commutes with c(ek )c(eℓ ) if the two sets{i, j} and {k, ℓ} have a unique element in
common. Such pairs of monomials will have no contributions in the above sum due to the
curvature symmetry
Rijkℓ = Rkℓij .
Thus, we can split the above sum into two parts
X X
ˆ = −1
c(F (∇)) Rijij c(ei )c(ej )c(ei )c(ej ) + Rijkl c(ei )c(ej )c(ek )c(eℓ ).
4
i,j i,j,k,ℓ distinct

Using the first Bianchi identity we deduce that the second sum vanishes. The first sum is
equal to
1X 1X s
− Rijij (c(ei )c(ej ))2 = Rijij = ,
4 4 4
i,j i,j

where s denotes the scalar curvature of M . We have thus proved the following result, [86].
12.2. FUNDAMENTAL EXAMPLES 669

Theorem 12.2.33 (Lichnerowicz).

1
D2 = ∇∗ ∇ + s. ⊔
⊓
4
A section ψ ∈ C ∞ (S(M )) such that D2 ψ = 0 is called a harmonic spinor. Lichnerowicz
theorem shows that a compact spin manifold with positive scalar curvature admits no
harmonic spinors.
Exercise 12.2.34. Consider an oriented 4-dimensional Riemann spin manifold (M, g)
and W → M a Hermitian vector bundle equipped with a Hermitian connection ∇. Form
the twisted Dirac operator



DW : C ∞ (S+ (M ) ⊕ S− (M )) ⊗ W → C ∞ (S+ (M ) ⊕ S− (M )) ⊗ W

defined in Exercise 12.1.68. The operator DW is Z2 -graded, and hence it has a block
decomposition  
0 6 ∗W,+
D
DW = ,
6 W,+
D 0
6 W,+ : C ∞ (S+ (M ) ⊗ W ) → C ∞ (S− (M ) ⊗ W ). Show that
where D


6 ∗W,+D
D 6 W,+ = ∇ ˜ + s + c F+ (∇) ,
˜ ∗∇
4
˜ denotes the product connection of S(M ) ⊗ W , and
where ∇
1

F+ (∇) = F (∇) + ∗F (∇) ,
2
denotes the self-dual part of the curvature of the bundle (W, ∇). ⊔
⊓

Exercise 12.2.35. The torus T 3 = S 1 × S 1 × S 1 equipped with the product metric (of
volume (2π)3 ) has eight spin structures. Since T 3 is also a Lie group, its tangent bundle
admits a canonical trivialization. We denote by σ0 the spin structure determined by this
trivialization. Using the free and transitive action of H 1 (T 3 , Z2 ) on Spin(T 3 ) we deduce
that we have a canonical bijection

H 1 (T 3 , Z2 ) → Spin(T 3 ), H 1 (T 3 , Z2 ) ∋ ǫ 7→ σǫ := ǫ · σ0 .

Compute the spectrum of the spin Dirac operator Dǫ determined by the spin structure
σǫ . ⊔
⊓

12.2.4 The spinc Dirac operator

Our last example of Dirac operator generalizes both the spin Dirac operator, and the
Hodge–Dolbeault operator. The common ingredient behind both these examples is the
notion of spinc structure. We begin by introducing it to the reader.
Let (M n , g) be an oriented, n-dimensional Riemann manifold. As in the previous
section we can regard the tangent bundle as associated to the principal bundle PSO(M ) of
670 CHAPTER 12. DIRAC OPERATORS

oriented orthonormal frames. Assume PSO(M ) is defined by a good open cover U = (Uα )
and transition maps
gαβ : Uαβ → SO(n).
The manifold M is said to posses a spinc structure (or complex spin structure) if there
exists a principal Spinc (n) -bundle PSpinc such that PSO(M ) is associated to PSpinc via the
natural morphism ρc : Spinc (n) → SO(n):

PSO(M ) = PSpinc ×ρc SO(n).

Equivalently, this means there exist smooth maps g̃αβ : Uαβ → Spinc (n), satisfying the
cocycle condition, such that
ρc (g̃αβ ) = gαβ .
As for spin structures, there are obstructions to the existence spinc structures, but they are
less restrictive. Let us try to understand what can go wrong. We stick to the assumption
that all the overlaps Uαβ···γ are contractible.
Since Spinc (n) = (Spin(n) × S 1 )/Z2 , lifting the SO(n) structure (gαβ ) reduces to
finding smooth maps

hαβ : Uαβ → Spin(n) and zαβ : Uαβ → S 1 ,

such that
ρ(hαβ ) = gαβ ,
and


(ǫαβγ , ζαβγ ) := hαβ hβγ hγα , zαβ zβγ zγα ∈ (−1, −1), (1, 1) . (12.2.4)
2 :U 1
If we set λαβ := zαβ αβ → S , we deduce from (12.2.4) that the collection (λαβ ) should
satisfy the cocycle condition. In particular, it defines a principal S 1 -bundle over M , or
equivalently, a complex Hermitian line bundle2 L. This line bundle should be considered
as part of the data defining a spinc structure.
The collection (ǫαβγ ) is an old acquaintance: it is a Čech 2-cocycle representing the
second Stieffel-Whitney class.
As in Subsection 8.2.2, we can represent the cocycle λαβ as

λαβ = exp(iθαβ ).

The collection
1

nαβγ = θαβ + θβγ + θγα ,
2π
defines a 2-cocycle of the constant sheaf Z representing the topological first Chern class
of L. The condition (12.2.4) shows that

nαβγ = ǫαβγ (mod 2).

2
In this subsection, by complex Hermitian line bundle we understand a complex line bundle equipped
with a U (1)-structure.
12.2. FUNDAMENTAL EXAMPLES 671

To summarize, we see that the existence of a spinc structure implies the existence of a
complex Hermitian line bundle L such that

ctop
1 (L) = w2 (M ) (mod 2).

It is not difficult to prove the above condition is also sufficient. In fact, one can be more
precise.
Denote by Spinc (M ) the collection of isomorphism classes of spinc structures on the
manifold M . Any σ ∈ Spinc (M ) is defined by a lift (hαβ , zαβ ) as above. We denote by
det σ the complex Hermitian line bundle defined by the gluing data (zαβ ). We have seen
that
ctop
1 (det σ) ≡ w2 (M ) (mod 2).

Denote by LM ⊂ H 2 (M, Z) the “affine ” subspace consisting of those cohomology classes

satisfying the above congruence modulo 2. We thus have a map

Spinc (M ) → LM , σ 7→ ctop
1 (det σ).

Proposition 12.2.36. The above map is a surjection. ⊔

⊓

Exercise 12.2.37. Complete the proof of the above proposition. ⊔

⊓

The smooth Picard group Pic∞ (M ) of isomorphisms classes of complex line bundles
(with group operation given by the tensor product) acts on Spinc (M ) by

Spinc (M ) × Pic∞ (M ) ∋ (σ, L) 7→ σ ⊗ L.

More precisely, if σ ∈ Spinc (M ) is given by the cocycle

σ = [hαβ , zαβ ] : Uαβ → Spin (n) × S 1 / ∼,

and L is given by the S 1 cocycle ζαβ : Uαβ → S 1 , then σ ⊗ L is given by the cocycle
[hαβ , zαβ ζαβ ]. Note that
det(σ ⊗ L) = det σ ⊗ L2 ,
so that

ctop
1 det(σ ⊗ L) = ctop top
1 (det σ) + 2c1 (L).

Proposition 12.2.38. The above action of Pic∞ (M ) on Spinc (M ) is free and transitive.

Proof. Consider two spinc structures σ 1 and σ 2 defined by the good cover (Uα ), and the
gluing cocycles
(i) (i)
[hαβ , zαβ ], i = 1, 2.
(1) (2) (2)
Since ρc (hαβ ) = ρc (hαβ ) = gαβ , we can assume (eventually modifying the maps hαβ by a
sign) that
(1) (2)
hαβ = hαβ .
This implies that the collection
(2) (1)
ζαβ = zαβ /zαβ
672 CHAPTER 12. DIRAC OPERATORS

is an S 1 -cocycle defining a complex Hermitian line bundle L. Obviously σ 2 = σ 1 ⊗ L.

This shows the action of Pic∞ (M ) is transitive. We leave the reader verify this action is
indeed free. The proposition is proved. ⊔
⊓

Given two spinc structures σ1 and σ2 we can define their “difference” σ2 /σ1 as the
unique complex Hermitian line bundle L such that σ2 = σ1 ⊗ L. This shows that the
collection of spinc structures is (non-canonically) isomorphic with H 2 (M, Z) ∼= Pic∞ (M ).
2
It is a sort of affine space modelled on H (M, Z) in the sense that the “difference ”
between two spinc structures is an element in H 2 (M, Z), but there is no distinguished
origin of this space. A structure as above is usually called a H 2 (M, Z)-torsor.
The set Spinc (M ) is equipped with a natural involution
Spinc (M ) ∋ σ 7→ σ̄ ∈ Spinc (M ),
defined as follows. If σ is defined by the cocycle [g̃αβ , zαβ ], then σ̄ is defined by the cocycle
[g̃αβ , zαβ −1 ]. Observe that
σ/σ̄ = det σ.
Without sufficient background in algebraic topology the above results may look of very
little help in detecting spinc structures. This is not the case, and to convince the reader
we will list below (without proofs) some examples of spinc manifolds.
Example 12.2.39. (a) Any spin manifold admits a spinc structure.
(b) Any almost complex manifold has a natural spinc structure.
(c) (Hirzebruch-Hopf, [64]; see also [102]) Any oriented manifold of dimension ≤ 4 admits
a spinc structure. ⊔
⊓

Let us analyze the first two examples above. If M is a spin manifold, then the lift
g̃αβ : Uαβ → Spin(n)
of the SO-structure to a spin structure canonically defines a spinc structure via the trivial
morphism
Spin(n) → Spinc (n) ×Z2 S 1 , g 7→ (g, 1) mod the Z2 − action.
We see that in this case the associated complex line bundle is the trivial bundle. This is
called the canonical spinc structure of a spin manifold. We thus have a map
Spin(M ) → Spinc (M ).
Suppose we have fixed a spin structure on M given by a Spin-lift g̃βα of an SO-gluing
cocycle gβα .
To any complex Hermitian line bundle L defined by the S 1 -cocycle (zαβ ) we can
associate the spinc structure σ ⊗ L defined by the gluing data

(g̃αβ , zαβ ) .
The complex Hermitian line bundle det σ⊗L associated to this structure is det σ⊗L = L⊗2 .
Since the topological Picard group Pic∞ (M ) acts freely and transitively on Spinc (M ), we
deduce that to any pair (ǫ, σ) ∈ Spin(M ) × Spinc (M ) we can canonically associate a
complex line bundle L = Lǫ,σ such that L⊗2 ∼= det σ, i.e., Lǫ,σ is a “square root of det σ.
12.2. FUNDAMENTAL EXAMPLES 673

Exercise 12.2.40. Show that for any σ ∈ Spinc (M ) there exists a natural bijection
between the set Spin(M ), and the set of isomorphisms of complex line bundles L such
that L⊗2 ∼
= det σ. ⊔
⊓

Exercise 12.2.41. Prove that the image of the natural map Spin(M ) → Spinc (M ) coin-
cides with the fixed point set of the involution σ 7→ σ̄. ⊔
⊓

Exercise 12.2.42. The torus T 3 is the base of a principal bundle

Z3 ֒→ R3 → T 3 ,

so that, to any group morphism ρ : Z3 → U (1) we can associate a complex line bundle
Lρ → T 3 .
(a) Prove that Lρ1 ·ρ2 ∼
= Lρ1 ⊗ Lρ2 , ∀ρ1 , ρ1 ∈ Hom(Z3 , S 1 ), and use this fact to describe
the first Chern class of Lρ by explicitly producing a Hermitian connection on Lρ .
(b) Prove that any complex line bundle on T 3 is isomorphic to a line bundle of the form
Lρ , for some morphism ρ ∈ Hom(Z3 , S 1 ). .
(c) Show that the image of the map Spin(T 3 ) → Spinc (T 3 ) consists of a single point. ⊔ ⊓

Exercise 12.2.43. Prove that Spin(RP3 ) consists of precisely two isomorphisms classes
of spin structures and moreover, the natural map

Spin(RP3 ) → Spinc (RP3 )

is a bijection. ⊔
⊓

To understand why an almost complex manifold admits a canonical spinc structure it

suffices to recall the natural morphism U (k) → SO(2k) factors through a morphism

ξ : U (k) → Spinc (2k).

The U (k)-structure of T M , defined by the gluing data

hαβ : Uαβ → U (k),

induces a spinc structure defined by the gluing data ξ(hαβ ). Its associated line bundle is
given by the S 1 -cocycle
detC (hαβ ) : Uαβ → S 1 ,
and it is precisely the determinant line bundle

detC T 1,0 M = Λk,0 T M.

The dual of this line bundle, detC (T ∗ M )1,0 = Λk,0 T ∗ M plays a special role in algebraic
geometry. It usually denoted by KM , and it is called the canonical line bundle. Thus the
−1 def
line bundle associated to this spinc structure is KM = KM ∗ .

From the considerations in Subsection 11.1.5 and 11.1.7 we see that many (complex)
vector bundles associated to the principal Spinc bundle of a spinc manifold carry natural
674 CHAPTER 12. DIRAC OPERATORS

Clifford structures, and in particular, one can speak of Dirac operators. We want to discuss
in some detail a very important special case.
Assume that (M, g) is an oriented, n-dimensional Riemann manifold. Fix σ ∈ Spinc (M )
(assuming there exist spinc structures). Denote by (gαβ ) a collection of gluing data defin-
ing the SO structure PSO(M ) on M with respect to some good open cover (Uα ). Moreover,
we assume σ is defined by the data

hαβ : Uαβ → Spinc (n).

Denote by ∆cn the fundamental complex spinorial representation defined in Subsection

11.1.6,
∆n : Spinc (n) → Aut (Sn ).
We obtain a complex Hermitian vector bundle

Sσ (M ) = PSpinc ×∆n Sn ,

which has a natural Clifford structure. This is called the bundle of complex spinors
associated to σ.
Example 12.2.44. (a) Assume M is a spin manifold. We denote by σ0 the spinc structure
corresponding to the fixed spin structure. The corresponding bundle of spinors S0 (M )
coincides with the bundle of pure spinors defined in the previous section. Moreover for
any complex line bundle L we have

SL := Sσ ∼
= S0 ⊗ L,
where σ = σ0 ⊗ L. Note that in this case L2 = det σ so one can write

Sσ ∼
= S0 ⊗ (det σ)1/2 .
(b) Assume M is an almost complex manifold. The bundle of complex spinors associated
−1
to the canonical spinc structure σ (such that det σ = KM ) is denoted by SC (M ). Note
that
SC (M ) ∼
= Λ0,• T ∗ M. ⊔
⊓
We will construct a natural family of Dirac structures on the bundle of complex spinors
associated to a spinc structure.
Consider the warm-up case when T M is trivial. Then we can assume gαβ ≡ 1, and

hαβ = (1, zαβ ) : Uαβ → Spin(n) × S 1 → Spinc (n).

The S 1 -cocycle (zαβ

2 ) defines the complex Hermitian line bundle det σ. In this case some-

thing more happens.

The collection (zαβ ) is also an S 1 -cocycle defining a complex Hermitian line bundle L̂,
such that L̂2 ∼= det σ. Traditionally, L̂ is denoted by (det σ)1/2 , though the square root
may not be uniquely defined.
We can now regard Sσ (M ) as a bundle associated to the trivial Spin(n)-bundle PSpin ,
and as such, there exists an isomorphism of complex Spin(n) vector bundles

Sσ (M ) ∼
= S(M ) ⊗ (det σ)1/2 .
12.2. FUNDAMENTAL EXAMPLES 675

As in the exercise 12.1.68 of Subsection 11.1.9, we deduce that twisting the canonical
connection on the bundle of pure spinors S0 (M ) with any U (1)-connection on det σ 1/2 we
obtain a Clifford connection on Sσ (M ). Notice that if the collection

ωα ∈ u(1) ⊗ Ω1 (Uα )

defines a connection on det σ, i.e.,

2
dzαβ
ωβ = 2 + ωα over Uαβ ,
zαβ

then the collection

1
ω̂α = ωα
2
defines a Hermitian connection on L̂ = det σ 1/2 . Moreover, if F denotes the curvature of
(ω· ), then the curvature of (ω̂· ) is given by

1
F̂ = F. (12.2.5)
2
Hence any connection on det σ defines in a unique way a Clifford connection on Sσ (M ).
Assume now that T M is not necessarily trivial. We can however cover M by open sets
(Uα ) such that each T Uα is trivial. If we pick from the start a Hermitian connection on
det σ, this induces a Clifford connection on each Sσ (Uα ). These can be glued back to a
Clifford connection on Sσ (M ) using partitions of unity. We let the reader check that the
connection obtained in this way is independent of the various choices.
Here is an equivalent way of associating a Clifford connection on Sσ to any Hermitian
connection on det σ. Fix an open cover (Uα ) of M consisting of geodesically convex
open sets. The restriction of T M to any Uα is trivializable. Fixing such (orthogonal)
trivializations leading to the gluing cocycle

gβα : Uαβ → SO(n), n := dim M.

We can describe the Levi–Civita connection on T M as a collection

Aα ∈ Ω1 (Uα ) ⊗ so(n),

satisfying the transition rules

−1 −1
Aβ = −(dgβα )gβα + gβα Aα gβα .

The spinc structure σ is described by a gluing cocycle

hβα : Uαβ → Spinc (n), hβα = [ĝβα , zβα ] ∈ Spin(n) × U (1)/{±1},

such that, if ρ : Spin(n) → SO(n) denotes the natural 2 : 1 morphism, then

gβα = ρ(ĝβα ).
676 CHAPTER 12. DIRAC OPERATORS

Define Âα ∈ Ω1 (Uα ) ⊗ spin(n) by setting

Âα (x) = ρ−1
∗ Aα (x) ,

where ρ∗ : spin(n) → so(n) is the natural Lie algebra isomorphism described in (12.1.3).
The collection (Âα ) satisfies the transition rules
−1 −1
Âβ = −(dĝβα )ĝβα + ĝβα Âα ĝβα .

Observe that although ĝ•• is only defined up to a ±1 ambiguity, this ambiguity is lost in
the above equality.
Consider now a connection ω on the Hermitian line bundle det σ defined by the gluing
2 ). It is defined by a collection
cocycle (z••

ωα ∈ Ω1 (Uα ) ⊗ u(1) = Ω1 (Uα ) ⊗ iR,

satisfying the transition rules

dzβα
ωβ = −2 + ωα .
zβα
The collection
1

Aωα := Âα ⊕ ωα ∈ Ω1 (Uα ) ⊗ spin(n) ⊕ u(1) = Ω1 (Uα ) ⊗ spinc (n)
2
defines a connection Aω on the principal Spinc (n)-bundle PSpin
σ
c → M determined by the
c ω
spin -structure σ. In induces a connection ∇ on the associated vector bundle Sσ , and as
in the previous subsection one can verify that ∇ω is a Clifford connection.
Example 12.2.45. Assume (M, g) is both complex and spin. Then a choice of a spin
−1 −1
structure canonically selects a square root KM −1/2 of the line bundle KM because KM
c
is the line bundle associated to the spin structure determined by the complex structure
on M . Then
SC ∼
= S0 ⊗ KM −1/2 .
Any Hermitian connection on KM induces a connection of SC . If M happens to be Kähler,
then the Levi–Civita connection induces a complex Hermitian connection KM and thus a
Clifford connection on SC (M ) ∼
= Λ0,• T ∗ M . ⊔
⊓

Let ω be a connection on det σ. Denote by ∇ω the Clifford connection it induces

on Sσ (M ) and by Dω the associated geometric Dirac operator. Since the Weitzenböck
remainder of this Dirac operator is a local object so to determine its form we may as well
assume Sσ = S ⊗ det σ 1/2 . Using the computation of Exercise 12.1.68 and the form of the
Weitzenböck remainder for the spin operator we deduce
1 1
D2σ,ω = (∇ω )∗ ∇ω + s + c( Fσ (ω) ), (12.2.6)
4 2
where Fσ denotes the curvature of the connection ω on det σ, and c(Fσ ) denotes the
Clifford multiplication by the purely imaginary 2-form Fσ .
12.2. FUNDAMENTAL EXAMPLES 677

Exercise 12.2.46. Consider the torus T 3 = S 1 × S 1 × S 1 equipped with the product

round metric
g = (dθ 1 )2 + (dθ 2 )2 + (dθ 3 )2 ,
where for j = 1, 2, 3, θ j ∈ [0, 2π) denotes the angular coordinate on the j-th circle.
Denote by σ0 the unique spinc structure on T 3 , such that det σ0 is a trivial line bundle.
As in the Exercise 12.2.42, for every morphism ρ : Z3 → S 1 , we denote by Lρ the associated
complex line bundle. We set σρ := σ ⊗ Lρ , and we fix a Hermitian metric on Lρ . The
spinc structure σρ , and a Hermitian connection ∇ on Lρ determines spinc Dirac operator
Dρ (∇).
(a) Prove that if is a Hermitian a unitary automorphism of Lρ then

Dρ (γ∇γ −1 ) = γDρ (∇)γ −1 .

In particular, the operators Dρ (γ∇γ −1 ) and Dρ (∇) have the same spectrum.
(b)∗ We say that two Hermitian connections ∇, ∇′ on Lρ are gauge equivalent if and only
if there exists a unitary automorphism γ of L such that ∇′ = γ∇γ −1 . Prove that two
connections ∇ and ∇′ are gauge equivalent if and only if
Z
1

dθ j ∧ F (∇′ ) − F (∇) ∈ 2πiZ, ∀j = 1, 2, 3.
2π T 3

(c) Fix a Hermitian connection ∇ on Lρ . Describe the spectrum of Dρ (∇) in terms of the
real numbers Z
1
sj = dθ j ∧ F (∇), j = 1, 2, 3. ⊔
⊓
2πi T 3
678 CHAPTER 12. DIRAC OPERATORS
Bibliography

[1] Mathematical Developments Arising from Hilbert Problems, Proceedings of Sym. in

Pure Math., vol . 28, Amer. Math. Soc. Providence RI, 1976.

[2] R. Adler, R. J. E. Taylor: Random Fields and Geometry, Springer Monographs in

Mathematics, Springer Verlag, 2007.

[3] J. F. Adams: Vector fields on spheres, Ann.of Math 75 (1960), p.603-632.

[4] R. A. Adams: Sobolev Spaces, Academic Press, New York, 1975.

[5] M.T. Anderson, J. Cheeger: C α -compactness for manifolds with Ricci curvature and
injectivity radius bounded below, J. Diff. Geom., 35(1992), 265-281.

[6] V. I. Arnold: Ordinary Differential Equations, Springer Texts, Springer-Verlag, 1992.

[7] V. I. Arnold: Mathematical Methods of Classical Mechanics, Springer Texts, Springer

Verlag, 1989.

[8] M. F. Atiyah, R. Bott, V. K. Patodi: On the heat equation and the index theorem,
Invent. Math., 19(1973),279-330.

[9] M. F. Atiyah, R. Bott, A. Shapiro: Clifford modules, Topology 3, Suppl.1 (1964), p.

3-38.

[10] T. Aubin: Problèmes isopérimétriques et espaces de Sobolev, J. Diff. Geom. 11

(1976), p. 533-598.

[11] T. Aubin: Équations différentielles non linéaires et problème de Yamabe concernant

la courbure scalaire, J. Math. Pures et Appl., 55 (1976), p. 269-296.

[12] V. Barbu: Differential Equations, Springer Verlag, 2016.

[13] P. Bérard: From vanishing theorems to estimating theorems: the Bochner technique
revisited, Bull. A.M.S. 19(1988), p.371-406.

[14] N. Berline, E. Getzler, M. Vergne: Heat Kernels and Dirac Aperators, Springer Verlag,
1992.

[15] L. Bers, F. John, M. Schechter: Partial Differential Equations, Lectures in Applied

Mathematics, vol. 3A, Amer. Math. Soc., Providence RI, 1964.

679
680 BIBLIOGRAPHY

[16] A. Besse: Einstein Manifolds, Ergebnisse der Mathematik, Bd.10, Springer Verlag,
1987.

[17] P. Bérard, G. Besson, S. Gallot: Embedding Riemannian manifolds by their heat

kernel, Geom. Funct. Anal., 4(1994), 373-398.

[18] X. Bin: Derivatives of the spectral function and Sobolev norms of eigenfunctions on a
closed Riemannian manifold, Ann. Global. Analysis an Geometry, 26(2004), 231-252.

[19] A. Bjorner: Topological methods, Hadbook of Combinatorics, R. Graham, M.

Grötschel and L. Lovász, (Eds), North-Holland, Amsterdam, 1995, p. 1819-1872.

[20] J. Bochnak, M. Coste, M.-F. Roy: Real Algebraic Geometry, Springer Verlag, 1998.

[21] B. Boos, K. Wojciechowski: Elliptic Boundary Problems for Dirac Operators,

Birkhauser, 1993.

[22] R. Bott, L. Tu: Differential Forms in Algebraic Topology, Springer-Verlag, 1982.

[23] H. Brezis: Functional Analysis, Sobolev Spaces and Partial Differential Equations,
Universitext, Springer Verlag, 2011.

[24] T. Bröcker, K. Jänich : Introduction To Differential Topology, Cambridge University

Press, 1982.

[25] E. Cartan: Sur les invariants integraux des certains espaces homogènes et les pro-
priétés topologiques de ces espaces, Ann. Soc. Pol. Math, 8(1929), p.181-225.

[26] H. Cartan: La transgression dans un group de Lie et dans un espace fibré principal,
p.1268-1282, vol. III of H. Cartan’s “ Collected Works”, Springer-Verlag, 1979.

[27] I. Chavel: Eigenvalues in Riemannian Geometry, Academic Press, 1984.

[28] J. Cheeger, D. G. Ebin: Comparison Theorems in Riemann Geometry, AMS Chelsea

Publishing, 2008.

[29] S. S. Chern: A simple intrinsic proof of the Gauss-Bonnet formula for closed Riemann
manifolds, Ann. Math, 45(1946), p.747-752.

[30] J. B. Conway: A Course in Point Set Topology, Undergraduate Texts in Mathematics,

Springer Verlag, 2014

[31] M. Coste: An Introduction to o-minimal geometry, Real Algebraic and Analytic Ge-
ometry Network,
http://www.ihp-raag.org/publications.php

[32] R. Courant, D. Hilbert: Methods of Mathematicsl Physics, vol. II, John Wiley &
Sons, 1989.

[33] B. Dacorogna: Direct Methods in the Calculus of Variations, Applied Mathematical

Sciences, vol. 78, Springer-Verlag, 1989.
BIBLIOGRAPHY 681

[34] J. Dieudonne: Foundations of Modern Analysis, Academic Press, 1963.

[35] M. P. DoCarmo: Riemannian Geometry, Birkhäuser, 1992.

[36] L. van der Dries: Tame Topology and o-minimal Structures, London Math. Soc. Series,
vol. 248, Cambridge University Press, 1998.

[37] L. van der Dries, C. Miller: Geometric categories and o-minimal structures Duke
Math. J. 84(1996), 497-540.

[38] B. A. Dubrovin, A. T. Fomenko, S.P Novikov: Modern Geometry - Methods And

Applications, Vol.1-3, Springer Verlag, 1984, 1985, 1990.

[39] J. J. Duistermaat, J. A. C. Kolk: Distributions. Theory and Applications, Birkhäuser,

2010.

[40] R. E. Edwards: Functional Analysis. Theory and Applications, Dover Publications,

1995.

[41] C. Ehresmann: Sur la topologie de certains espaces homogènes, Ann. of Math.,

35(1934), p.396-443.

[42] S. Eilenberg, N. Steenrod: Foundations of Algebraic Topology, Princeton University

Press, 1952.

[43] J. D. Elworthy, Y. LeJan, X.-M. Lin: On the Geometry of Diffusion Operators and
Stochastic Flows, Lect. Notes in Math., vol. 1720, Springer Verlag, 1999.

[44] L. C. Evans: Partial Differential Equations, 2nd Edition, Grad. Studies in Math., vol.
19, Amer. Math. Soc., 2010

[45] H. Federer: Geometric Measure Theory, Springer Verlag, 1969.

[46] W. Fulton, J. Harris: Representation Theory, Springer-Verlag 1991.

[47] R. P. Feynman, R. B. Leighton, M. Sands: The Feynamn Lectures on Physics, Addi-

son Wesley, 1964.

[48] S. Gallot, D. Hulin, J. Lafonte: Riemannian Geometry, 3rd Edition, Springer Verlag,
2004.

[49] I.M. Gelfand, G.E. Shilov: Generalized Functions. Properties and Operations, vol. 1,
Academic Press, 1964.

[50] I.M. Gelfand, G.E. Shilov: Generalized Functions. Spaces of Fundamental and Gen-
eralized functions, vol. 2, Academic Press, 1968.

[51] D. Gilbarg, N. S. Trudinger: Elliptic Partial Differential Equations of Second Order,

Springer Verlag, 1983.

[52] R. Godement: Topologie Algébrique et Théorie des Faisceaux, Hermann et Cie, Paris,
1964.
682 BIBLIOGRAPHY

[53] C. Gordon. D. L. Webb, S. Wolpert: One cannot hear the shape of a drum, Bull.
A.M.S., 27(1992), p.134-138.

[54] I. S. Gradshteyn, I. M. Ryzhik: Tables of Integrals, Series and Products, Academic

Press, 1994.

[55] P. Griffiths, J. Harris: Principles of Algebraic Geometry, John Wiley&Sons, 1978.

[56] V. Guillemin, S. Sternberg: Geometric Asymptotics, Math. Surveys and Monographs,

vol. 14, Amer. Math. Soc., 1990.

[57] J. Hadamard: Lectures on Cauchy’s Problem in Linear Partial Differential Equations,

Yale University Press, 1923.

[58] B. Hall: Lie Groups, Lie Algebras, and Reprsentations. An Elementary Introduction,
Springer Verlag 2003.

[59] G. H. Hardy, E.M Wright: An Introduction to the Theory of Numbers, Oxford Uni-
versity Press, 1962.

[60] P. Hartman: Ordinary Differential Equations, 2nd Edition, Classics in Applied Math-
ematics, SIAM 2002.

[61] S. Helgason: Differential Geometry, Lie Groups, and Symmetric Spaces, Graduate
Studies in Math. v. 34, Amer. Math. Soc., 2001.

[62] M.Hirsch: Differential Topology, Springer-Verlag, 1976.

[63] F. Hirzebruch: Topological Methods in Algebraic Geometry, Springer Verlag, New

York, 1966.

[64] F. Hirzebruch, H. Hopf: Felder von Flachenelementen in 4-dimensionalen Manig-

faltigkeiten Math. Ann. 136(1958).

[65] L. Hörmander: On the spectral function of an elliptic operator, Acta Math.

121(1968), 193-218.

[66] L. Hörmander: The Analysis of Linear Partial Differential Operators I., Springer
Verlag, 2003.

[67] L. Hörmander: The Analysis of Linear Partial Differential Operators III., Springer
Verlag, 1985.

[68] B. Iversen: Cohomology of Sheaves, Springer Verlag, 1986.

[69] H. Iwamoto: On integral invariants and Betti numbers of symmetric Riemannian

manifolds, I, J. of the Math. Soc. of Japan, 1(1949), p.91-110.

[70] J. Jost: Riemannian Goemtry and Geometric Analysis, Universitext, Springer Verlag,
1998.
BIBLIOGRAPHY 683

[71] M. Kac: Can one hear the shape of a drum?, Amer. Math. Monthly, 73(1966) Part
II, p.1-23.

[72] M. Karoubi: K-Theory, Springer-Verlag, 1978.

[73] M. Kashiwara, P. Schapira: Sheaves on Manifolds, Springer Verlag, 1990.

[74] M. Kashiwara, P. Schapira: Categories and Sheaves, Springer Verlag, 2006.

[75] T. Kato: Perturbation Theory for Linear Operators, Springer Verlag, 1984.

[76] J. Kazdan, F. Warner: Curvature functions for compact 2-manifolds, Ann. Math.
99(1974), p.14-47.

[77] J. L. Kelley: General Topology, Van Nostrand, 1955.

[78] A. Khovanskii: Fewnomials, Transl. Math. Monogr., 88, Amer.Math. Soc., 1991.

[79] D. A. Klain, G.-C. Rota: Introduction to Geometric Probability, Cambridge University

Press, 1997.

[80] S. Kobayashi, K. Nomizu: Foundations of Differential Geometry, Interscience Pub-

lishers, New York, 1963.

[81] K. Kodaira, D. C. Spencer: Groups of complex line bundles over compact Kähler va-
rieties. Divisor class group on algebraic varieties, Proc. Nat. Acad. Sci. USA 39(1953),
p. 868-877.

[82] J. Kubarski, N. Teleman: N. Linear Direct Connections, Proceedings 7th Conference

on ”Geometry and Topology of Mani- folds - The Mathematical Legacy of Charles
Ehresmann”, Betlewo, May 2005.

[83] H. B. Lawson, M.-L. Michelson: Spin Geometry, Princeton University Press, 1989.

[84] N. N. Lebedev: Special Functions and Their Applications, Dover, 1972.

[85] J. M. Lee, T. H. Parker: The Yamabe problem, Bull. A.M.S. 17 (1987), p. 37-91.

[86] A. Lichnerowicz: Laplacien sur une variété riemannienne et spineure, Atti Accad.
Naz. dei Lincei, Rendicotti 33(1962), 187-191.

[87] D. E. Littlewood: The Theory of Group Characters and Matrix Representations of

Groups, Clarendon Press, Oxford, 1940.

[88] I. G. Macdonald: Symmetric Functions and Hall Polynomials, Clarendon Press, Ox-
ford, 1979.

[89] S. MacLane: Categories for the Working Mathematician, Springer Verlag, 1971.

[90] W. S. Massey: A Basic Course in Algebraic Topology, Springer Verlag, 1991.

[91] W. S. Massey: Cross products of vectors in higher dimensional euclidian spaces, Amer.
Math. Monthly, 93(1983), p.697-700.
684 BIBLIOGRAPHY

[92] P. -D. Mehtée: Sur les distributions invariantes dans le groupe des rotations de
Lorentz, Comm. Math. Helv. 28(1954), 225-269.

[93] J. Milnor: On manifolds homeomorphic to the 7-sphere, Ann. of Math. 64 (1956),

p.399-405.

[94] J. W. Milnor: Morse Theory, Princeton University Press, 1963.

[95] J. W. Milnor: Spin structures on manifolds, L’Enseignement Math., 9(1963), 198-203.

[96] J. W. Milnor: Curvatures of left invariant metrics on Lie groups, Adv. in Math.,
21(1976), p.293-329.

[97] J. W. Milnor: Topology from the Differentiable Viewpoint, Princeton University Press,
1965.

[98] J. W. Milnor, J. D. Stasheff: Characteristic Classes, Ann. Math. Studies 74, Princeton
University Press, Princeton, 1974.

[99] M. Mimura, H. Toda: Topology of Lie Groups, I and II, Translations of Mathematical
Monographs, vol. 91, Amer. Math. Soc., 1991.

[100] R. Miron, M. Anastasiei: The Geometry of Lagrange Spaces: Theory and Applica-
tions, Kluwer Academic Publishers, 1994.

[101] J. W. Morgan: The Seiberg-Witten Equations and Applications to the Topology of

Smooth Manifolds, Mathematical Notes, Princeton University Press, 1996.

[102] C. Morrey: Multiple Integrals in the Calculus of Variations, Springer Verlag, 1966.

[103] A. Newlander, L. Nirenberg: Complex analytic coordinates on almost complez man-

ifolds, Ann. Math., 54(1954), p.391-404.

[104] L. I. Nicolaescu: A new method of constructing connections on vector bundles,

Liviu’s Math Blog, October 31,2014.

[105] L. I. Nicolaescu: A stochastic Gauss-Bonnet–Chern formula, arXiv: 1408.5746,

Prob. Th. & Related Fields, 165(2016), 235-265.

[106] L. I. Nicolaescu, N. Savale:The Gauss-Bonnet-Chern theorem: a probabilistic per-

spective, Trans. A.M.S., 369(2017), 2951-2986.

[107] L. Nirenberg: On elliptic partial differential equations, Ann. Scuola Norm. Sup.
Pisa, 13(1959), p.1-48.

[108] K. Nomizu, H. Ozeki: The existence of complete Riemann metrics, Proc. A.M.S. ,
12(1961), 889-891.

[109] S. Peters: Convergence of Riemannian manifolds, Compositio Math., 62(1987), 3-16.

[110] V. I. Paulsen, M. Raghupati: Introduction to the Theory of Reproducing Kernel

Hilbert Spaces, Cambridge Univ. Press, 2016.
BIBLIOGRAPHY 685

[111] L. S. Pontryagin: Smooth manifolds and their applications in homotopy theory,

Amer. Math. Soc. Trans., II(1959), 1-114.

[112] I. R. Porteous: Topological Geometry, Cambridge University Press, 1981.

[113] M. M. Postnikov: Lectures in Geometry. Semester V: Lie Groups and Lie Algebras,
Mir Publishers, Moscow, 1986.

[114] M. Riesz: L’ integrale de Riemann-Liouville et le problème de Cauchy, Acta Math.,

81(1949), 1-223.

[115] L. Santaló: Integral Geometry and Geometric Probability, Cambridge University

Press, 2004.

[116] R. Schoen: Conformal deformations of a Riemannian metric to constant curvature,

J. Diff. Geom. 20 (1984), p. 479-495.

[117] J. A. Schouten, D. J. Struik: On some properties of general manifolds relating to

Einstein’s theory of gravitation, Amer. J. Math 43(1921), p. 213-216.

[118] L. Schwartz: Théorie des Distributions, Hermann, 1966.

[119] J. P. Serre: Faisceaux algébriques cohérentes, Ann Math. 61(1955), p.197-278.

[120] I. R. Shafarevitch: Algebra I. Basic Notions, Encyclopedia of Mathematical Sciences,

vol. 11, Springer Verlag, 1990.

[121] M. A. Shubin: Pseudodifferential Operators and Spectral Theory, Springer Verlag,

2001.

[122] P. Speissegger: The Pfaffian closure of an o-minimal structure, J. Reine. Angew.

Math., 508(1999), 189-211.

[123] M. Spivak: A Comprehensive Introduction to Differential Geometry, vol. 1-5, Publish

or Perish, 1979.

[124] N. E. Steenrod: The Topology of Fibre bundles, Princeton University Press, 1951.

[125] E. M. Stein: Singular Integrals and Differentiability Properties of Functions, Prince-

ton University Press, 1970.

[126] M. Taylor: Pseudodifferential Operators, Princeton University Press, 1981.

[127] F. Trèves: Topological Vector Spaces, Distributions and Kernels, Dover, 2006.

[128] F. W. Warner: Foundations of Differentiable Manifolds and Lie Groups, Springer

Verlag, 1983.

[129] H.Weyl: Nachr. Akad. Wiss Göttingen, 1914, 235-236.

[130] H.Weyl: Math. Ann., 77 (1916), p.313-315.

686 BIBLIOGRAPHY

[131] H.Weyl: Die Idee Der Riemannschen Fläche B.G. Tuebner, Berlin 1913.

[132] H. Weyl: On the volume of tubes, Amer. J. Math., 61(1939), 461-472.

[133] H. Weyl: The Classical Groups. Their Invariants and Representations, Princeton
University Press, 1939.

[134] E.T. Whittaker, G.N. Watson: A Course of Modern Analysis, Cambridge Univ.
Press, 1996.

[135] E. Witten: Monopoles and four-manifolds, Math.Res. Letters, 1(1994), 769-796.

[136] S. Zelditch: Real and complex zeros of Riemannian random waves, Spectral analysis
in geometry and number theory, Contemp. Math., 484, Amer. Math. Soc., Providence,
RI, 2009. arXiv:0803.433v1

[137] K. Yosida: Functional Analysis, Springer Verlag, 1974.

Index

(x ± i0)a , 547 I±α , 549

A† , 45 e(E), 290
C ∞ (E), 38 χa+ , 544
C ∞ (U, E), 38 χa− , 545
C −∞ (O)), 530 δji , 48
C ∞ (M ), 9 det V , 51
C+ , 556 det ind(L), 518
C− , 556 div X, 160
E0 ⊠ E1 , 184 D(O), 529
I • (g), 340 DK (O), 529
Lk,p , 470 EM,g , 582
R+ a , 561 R+ a , 562

Adg , 130 S (Rn ), 535

Cr F , 31 Sn , 535, 536
∆F , 31 Sn′ , 536
∆M , 165 expq , 150
End− (V ), 64 EM (g), 195
GL(Kn ), 20 Un , 39
GL(n, K), 20 Uk,n , 40, 362
Grk (V ), 15, 388 grad f , 160
Λ• V , 51 µ̂p , 445
Λp,q V , 68 R , 91
Lom , 557 RM , 116
Ωk (F ), 98 E/B , 137
Ωk (O), 534 µw , 430, 435
Pin(V ), 636 1S , 5
Spin(V ), 636 ∂-bundle, 139
Vect , 38, 73 π1 (X, x0 ), 231
kukLp (M ) , 583 ρM (q), 151
k
Sn , 641 σj (A), 52
Tm , 586 σ(M ), 274
Θ(E), 122 ≃, 230
ΘM , 122 spec(L), 519
L-genus, 359 sing supp u, 531
γ(2h), 406 tr , 48
or 1 /or 2 , 63 trj , 434
G(v1 , . . . , vr ), 61 Hom(E, F ), 38

687
688 INDEX

|χ|a , 546, 573 determinant line, 114, 118

b
A-genus, 359 Dirac, 649
{·, ·}, 621 fiber, 75, 77
bk (M ), 249 frame, 79
d∗ , 162 Hopf, 80
dm (a), 562, 565 line, 36, 39
r-density, 66 tautological, 39, 373
sy (x), 581 universal, 39
xa+ , 543 principal, 79, 331
K , 673 connection on, 332
R~ M
g , 63
pullback of a, 344
R[[x]]♭ , 354 tangent, 30
Pic∞ , 348, 671 vector, 36, 39
ind (γ), 219 rank of a, 36
trs , 58 ample, 41
Tsr (V ), 47 automorphism, 38
gl(n), 21, 90 base of a, 36
o(n), 21, 90 dual, 71
sl(n), 21, 90 endomorphism, 38
su(n), 21, 90 framed, 42
u(n), 21, 90 map, 38
~ν , 163 morphism, 38
pullback, 41
adjoint, 45 section, 37
adjoint representation, 130, 304, 331 tautological, 40
adjunction isomorphism, 47 total space of a, 36
adjunction morphism, 59 trivial, 42, 110
ample subspace, 184 trivializable, 42, 118
antipodal points, 15 trivialized, 42
atlas, 8 universal, 40
maximal, 9
Cartan
berezinian, 63 algebra, 347, 349, 351
Bernoulli numbers, 356 form, 166, 301
Beta function, 386 Lemma, 178
Betti number, 249 structural equations, 176
Bianchi identity, 110, 168 categories, 234
bootstrap, 506 Cayley transform, 20, 144
bundle, 29 character, 133
G-fiber, 77 characteristic class, 345
canonical line -, 673 2nd Stiefel-Whitney, 665
determinant line, 71 universal, 355, 359
Clifford, 621 charts, 8
complex spinor, 666 Chern
cotangent, 72 classes, 348
INDEX 689

universal, 355 inner, 127

universal rank r, 347 outer, 127
polynomial, 348, 356 contraction, 48, 57, 75, 91
universal rank r, 347 convolution, 471
Chern–Weil correspondence, 344 coordinates, 3
Christoffel symbols, 112, 146, 167 Cartesian system, 4
circle, 23 local, 8
class function, 133 polar, 4
Clifford coordinatization, 3, 5
algebra, 622 correlator, 184, 610
connection, 649 associated to an embedding, 185
module, 625 tunneling associated to a, 186
multiplication, 621, 622 covariant derivative, see connection
s-module, 625 covering space, 235
structure, 621, 649 sheets of a, 235
coboundary, 250 universal, 240
cochain homotopy, 251 critical
cochain morphism, 251 point, 31
cocycle, 250 value, 31
cohomologous, 250 critical point, 32
condition, 36 current, 534
cohomology boundary of, 535
Čech, 323 of integration, 535
Čech , 321 curvature, 105, 334
DeRham, 248 Gauss, 172
compact supports, 267 Ricci, 169, 220, 656
complex Riemann, 167
chain, 249 scalar, 170
cochain, 249 sectional, 171, 219
acyclic, 250 cycle, 275
DeRham, 250 degenerate, 275
Dolbeault, 660 transversal, 280
Koszul, 661 cycles
complex structure, 67 cobordant, 275
configuration space, 5
conjugate points, see geodesic deck transformation, 237
connection, 98, 332 densities, 114
compatible with a metric, 145 Jacobian, 381
compatible with a metric, 100, 107 pullback of, 381
flat, 107 pushforward of, 382
gauge equivalence, 337 density bundle, 114
Levi–Civita, 145 dependence cone, 569
symmetric, 112 DeRham
connection 1-form, 100 cohomology, 250
conormal, 127 complex, 250
690 INDEX

derivative, 6 algebra, 51
Fréchet, 6 derivative, 92
determinant line, 51, 61 product, 51
diffeomorphism, 9 extremal, 204
of manifolds with boundary, 126
differential, 30 five lemma, 264
differential form, 72 flow, 81
closed, 243 infinitesimal generator of a, 82
local, 81, 83
exact, 243
formula
Dirac function, 530
duplication, 549
distribution, 529
Parseval, 584
temperate, 536
Poisson summation, 587
distributional derivative, 531
polarization, 338
duality, 59
reflection, 548
Hodge, 65
co-area, 384, 385
metric, 59
coarea, 383
natural, 59
Crofton, 437, 446, 451
symplectic, 59
Euler-Meusnier , 441
duplication formula, 549
Gauss, 179
Dynkin polynomials, 90
Green, 165, 468
Künneth, 265
Einstein
Parseval, 537
convention, 48
Stokes, 127
equation, 195
Weyl integration, 303
manifold, 195
Fourier transform, 490, 536
tensor, 195
Fréchet derivative, 5
elliptic complex, 524
framing, 42, 72
embedding, 34
Fredholm index, 512
energy method, 571
function, 5
equation
C k, 5
Einstein, 195
Fréchet differentiable, 5
Euler-Lagrange, 204
harmonic, 165
Hamilton-Jacobi, 223
smooth, 5, 9
Jacobi, 216
functors, 234
Euler
contravariant, 234
characteristic, 54, 198, 249, 255, 260,
covariant, 234
290
fundamental solution, 555
class, 290, 353 Cauchy, 555
geometric, 354
topological, 354 Gamma function, 166, 386, 392
universal, 353 gauge equivalence, see connection
form, 173, 182, 353 gauge transformation, 38, 337
Euler-Lagrange equation, 204 Gauss
exponential map see manifold 150 curvature, 172, 183
exterior, 51 lemma, 153
INDEX 691

map, 181, 182, 402 Gysin map, 273, 287

Gauss Lemma, 576
Gelfand–Leray residue, 139, 558 Hamilton-Jacobi equation, 223
generalized heat kernel, 617
function Heaviside function, 531
Riesz, 564 Hessian, 192, 214
generalized function, 529 Hilbert-Einstein functional, 195, 430, 450
of order ≤ m, 530 Hodge ∗-operator, see operator
singular support of, 531 holonomy, 110
homogeneous space, 294
support of, 531
homotopy, 256
genus, 198
geodesic, 146 immersion, 34
flow, 147 injectivity radius, 151
conjugate points, 217 integral curve, 82
index of a, 219 integration current, 535
nondegenerate, 217 interior derivation, 57, 91
global angular form, 367 intersection
gluing cocycle, 37, 332 form, 274
gluing condition, 38 number, 280
Grammian, 61 invariant
Grassmannian, 15, 40, 301 form, 297
complex, 15, 296, 301 forms, 300, 302
real, 15 integration, 124
group, 20, 42 polynomials, 340
action, 76 tensor field, 89
effective, 76 isometry, 142
free, 76 isotropy, 294
orbit, 76
Clifford, 633 Jacobi
complex Clifford, 642 equation, 216
connected, 22 field, 216, 219
fundamental, 231 identity, 57
left translation in a, 20 Jacobian
Lie, 20, 42, 89, 130, 147, 171, 300, 331, relative, 381
341
Killing
exponential map of a, 89
pairing, 149, 251, 301
representation, 77 vector field, 163
linear representation, 130 kinematic density, 392
irreducible, 131 Kronecker
orthogonal, 20 pairing, 271
Picard, 348 symbol, 46
right translation in a, 42
special orthogonal, 21 lagrangian, 203
special unitary, 21 invariant, 210
unitary, 21 Laplacian, 165, 463
692 INDEX

covariant, 467 lift of a, 237

least action, see Euler-Lagrange quantization, 639
Lefschetz number, 286 transition, 8, 36
Lie cohomologous, 39
bracket, 84 transversal, 35
derivative, 83 matching, 420
group, see group matrix, 6
Lie algebra, 57, 87, 89, 124 invertible, 6
simple, 301 Jacobian, 6
semisimple, 149, 251, 301 orthogonal, 20
light cone skew-Hermitian, 21
backward, 556 skew-symmetric, 20
forward, 556 unitary, 21
maximum principle, 500
Möbius band, 118 Mayer–Vietoris
manifold cover, 262
with boundary, 125 functors, 262
almost complex, 658 correspondence of, 263
complex, 15 long sequence, 258, 270
connect sum, 23 principle, 263
connected sum, 13 short sequence, 258, 270
direct product, 13 metric, 71
finite type, 261 conformal, 507
Kähler, 662 Kähler, 662
orientable, 23, 118 Riemann, 72
orientation of a, 119 metric volume, 159
oriented, 118, 119 module
Riemann, 142 G-, 130
convex sets in a, 156 irreducible, 131
exponential map of a, 150 morphism, 130
geodesically complete, 147 mollifier, 471
smooth, 8 mollifying sequence, 472
spinnable, 664 monodromy, 233, 239
manifolds, 5, 12 moving frame, 100
with boundary
negligible set, 31
diffeomorphism of, 126
nerve, 320
map, see function
Newton formulæ, 359
differential of, 30 normal coordinates, 151
gluing, 15
cochain operator, 5
cone of a, 254 Hodge ∗, 65, 160, 526, 663
cylinder of a, 254 boundary, 249
covering, 235 Cauchy–Riemann, 620
degree of a, 285 chirality, 628
homotopic, 230 closed, 510
INDEX 693

Fredholm, 512 polynomial, 248

self-adjoint, 510 series, 54
semi-Fredholm, 511 polarization formula, 338, 459
symmetric, 510 polynomial
coboundary, 249 Ad-invariant, 340
Dirac, 619 Chern, 348
suspension of, 620 Pontryagin, 350
Fredholm, 512 Pontryagin
geometric Dirac, 651 classes, 350
Hamilton–Floer, 620 universal rank r, 350
Hodge–DeRham, 620 polynomial
Hodge–Dolbeault, 661 universal rank r, 350
linear, 5 presheaf, 315
transpose of a, 45 projection, 12
partial differential, see p.d.o. stereographic, 12
orientation, 62 projection formula, 138
bundle, 122 projective plane, see space
canonical, 70 pullback, 74
cover, 237 pushforward, 273, 287
relative signature, 63
quantization map, 625
orientation cover, 237
quaternions, 149
orientationsee vector space 62
outer normal, 163 reflection formula, 548
representation, 130
p.d.o., 455, 456
reproducing kernel
elliptic, 460
Hilbert space, 193
formal adjoint, 463 Ricci tensor, see curvature
formally self-adjoint, 464 Riemann metric, 142
symbol of a , 460 Riesz, 564
pairing, 59
Hodge, 65 Schur
parallel transport, 103 lemma, 132
parametrix, 497 polynomial, 311
partition, 310 Schwartz
conjugate of a, 311 function, 536
length of a, 310 second fundamental form, see submanifold
weight of a, 310 shape operator, 181, 402, 430
partition of unity, 11 sheaf, 316
pfaffian, 63, 351 fine, 325
Picard group, 348, 671 resolution, 319
Poincaré fine, 325
duality, 272, 279 sheafification, 318
inequality, 504, 512 Shwartz
lemma, 244 space, 535
for compact supports, 268 simplicial scheme, 328
694 INDEX

smooth, see function doughnut-shaped, 23

structure, 8 symbol map, 625
space symmetric space, 295
Banach, 5 symmetry, 20
complex projective, 15
locally compact, 8 tame geometry, 412
real projective, 13 tensor, 44
second countable, 8 product, 44
simply connected, 232 algebra, 47
Sobolev, 470 contravariant, 47
tangent, 25, 29 covariant, 47
space of germs, 316 field, 72
morphism, 318 left invariant, 89
section of, 316 right invariant, 89
stalk, 316 product
spectral data, 529, 606 universality property of, 44
spectral function, 582, 597 skew-symmetric, 49
spectral perestroika, 607 exterior product of, 49
sphere, 12 symmetric, 49
sphere bundle, 80 type (r,s), 47
spin structure, 664 test function, 529
spinc structure, 670 theorem, 5
spinor, 666 Fubini, 33, 376
bundle, 666 Nash embedding, 431, 433
harmonic, 669 Banach fixed point, 7
module, 627, 631, 650 Bochner vanishing, 657
split coordinates, 136 Bochner-Weitzenböck, 652
structural equations, 174 Calderon-Zygmund, 491
structure, 412 Cartan, 300
o-minimal, 412 Cartan-Hadamard, 236
tame, 412 Chern–Weil, 341
submanifold, 11 Clairaut, 212
2nd fundamental form, 179, 402 dominated convergence, 452
codimension of a, 11 Fubini, 138
submersion, 34 Gauss–Bonnet, 183, 194, 261, 286, 406,
super, 55 433, 618
algebra, 55 Gauss–Bonnet–Chern, 366, 406, 618
commutative, 55 Hodge, 523, 524, 526
commutator, 55 Hopf-Rinow, 157
derivation, 56 implicit function, 7
Lie algebra, 57 inverse function, 6
space, 55 Leray-Hirsch, 267
tensor product, 58 Mayer–Vietoris, 258
supertrace, 58 Morrey, 482
surface, 23 Myers, 220
INDEX 695

Noether, 211 family, 564

Poincaré-Hopf, 293 kernel, 590
Rellich-Kondrachov compactness, 479 weak derivative, 469
Sard, 32, 284, 438, 637 Weitzenböck
Sobolev embedding, 477 remainder, 468, 654, 668
Stokes, 127 Weyl
Tarski-Seidenberg, 412 group, 305, 347, 350
uniformization, 508 integration formula, 303
Weyl, 242 lemma, 498
Theorema egregium, 179, 402 unitary trick, 131
Thom class, 288 Whitney sum, 71
Todd class
universal, 356 Young diagram, 310
torsion, 112
torsor, 672
torus, 13, 20
maximal, 302
trace, 48, 52
transform, 20
Fourier, 490
transpose, 45
transversal, 35
trivializing cover, 35
tube formula, 428
tunneling, 186

uniform boundedness principle, 530

unit outer normal, 163

variation, 205, 213

geodesic, 216
infinitesimal, 213
vector, 26
field, 38
space, 27, 43
Z-graded, 54
determinant line of a, 51, 61
dual, 45
orientation of a, 62
oriented, 62
volume form in a, 61
tangent , 26
vector bundle, see bundle
volume form, 118

wave