Sie sind auf Seite 1von 88

The Atiyah-Singer Index Theorem and

Fractal Geometry

Thesis submitted in partial fulfillment of the requirement for

Honors in

Mathematics

Aaron Ackbarali

Adviser, Jonathan Forde April 12, 2016


Preface

This text is meant for a reader who is very familiar with basic constructions
in algebraic topology, algebra, and analysis. Specifically we shall assume a
background knowledge of homology and cohomology of simplicial, CW, and
singular varieties, as well as a working knowledge of category theory, familiarity
with rings and modules, and a comprehensive understanding of calculus on
manifolds. The last of these is perhaps most important as we will always be
using high-level language to discuss the theory of manifolds. For example, we
will never get into the rigorous details of stitching and gluing together smooth
objects via a partition of unity subordinate to some trivialization. The reader
should also be familiar with the theory of the Riemann zeta function as this will
be essential in understanding the final chapters.

In this text we use many conventions in order to keep the discussion focused
on the most important material.

Whenever we say manifold it is meant smooth manifold unless otherwise


specified.

Only equations that will be referred to will be labelled.

N will include 0 when convenient.

Whenever a proof is at the very least influenced, even only in spirit, by

1
another author, a citation will be present at the beginning of that proof.
It is worth noting that at the time of writing, the content of (7.1) is novel
work.

Unless otherwise stated or obvious, maps will always be homomorphisms.


An example of when it is obvious that the reader is not to assume this is
when we are explicitly showing that a map is a homomorphism.

Only when nontrivial will we show that isomorphisms are well-defined and
satisfy the homomorphism property.

When possible, simple tensors will be used, taking advantage of the fact
that all of the theory presented extends bilinearly.

Indices will be suppressed when understood. For example, x1 , . . . , xn will


be written simply as x when working with local coordinates.

2
Contents

Introduction 6

I Bundles and Cohomology 8

1 Bundles 9

1.1 Fiber Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.2 Vector Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.3 Operations and Bundle Maps . . . . . . . . . . . . . . . . . . . . 15

1.4 Orientation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Cohomology 20

2.1 Differential Forms . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.2 de Rham Cohomology . . . . . . . . . . . . . . . . . . . . . . . . 23

2.3 Poincare Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.4 Leray-Hirsch Theorem . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Characteristic classes 34

3
3.1 Orienting Vector Bundles . . . . . . . . . . . . . . . . . . . . . . 35

3.2 Thom and Euler Class . . . . . . . . . . . . . . . . . . . . . . . . 36

3.3 Chern Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

II K-Theory 41

4 K-theory 42

4.1 K(X) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.2 K n (X) and the Thom Isomorphism . . . . . . . . . . . . . . . . 46

4.3 Topological Index . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5 The Atiyah-Singer Index Theorem 53

5.1 Elliptic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5.2 Analytic Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.3 Proof of the theorem . . . . . . . . . . . . . . . . . . . . . . . . . 60

III Fractal Geometry 66

6 Fractal Geometry 67

6.1 Fractal dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6.2 The Inverse Spectral Problem . . . . . . . . . . . . . . . . . . . . 70

7 Fractal Cohomology 73

7.1 Spectral Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

7.2 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 76

4
Appendices 77

A Homological Algebra 78

A.1 Five Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

A.2 Euler Characteristic . . . . . . . . . . . . . . . . . . . . . . . . . 79

B Computing Cohomology 81

B.1 Mayer-Vietoris Sequence . . . . . . . . . . . . . . . . . . . . . . . 81

References 85

5
Introduction

The fundamental theorem of arithmetic says that every number can be written
as the product of prime numbers. In this way, prime numbers are the fundamental
building blocks for mathematics, and understanding them is inherently valuable
to the field. Mathematicians have tried, and mostly succeeded, to unravel the
mysteries of prime numbers for centuries, but one mystery still eludes them.
Mathematicians have not yet figured out if the prime numbers are distributed
randomly or not. This is why Riemann hypothesis is perhaps the greatest open
question of all history. It claims not only that the prime numbers are not
random, but even gives a formula for how often they occur. The issue is that
this formula relies on knowing that the zeroes of the Riemann zeta function all
have real part 1/2, and proving this has defeated the greatest mathematicians
of the past three centuries. Defeat has not been without reward though.

The mathematician Paul Erdos said A problem worthy of attack, proves


its worth by fighting back! Fighting with the Riemann hypothesis has led
mathematicians to discover incredible tools and theorems. One tool that has
been unexpectedly instrumental in this fight is cohomology. In 1974, Pierre
Deligne proved several conjectures about the behavior of zeta functions over
algebraic varieties, specifically the Weil conjectures, using etale cohomology.
More recently Michel L. Lapidus proved that the Weyl-Berry conjecture, also
known as the inverse spectrum problem, is equivalent to the Riemann Hypothesis

6
when extended to fractals. He and collaborator Machiel van Frankenhuijsen
then proposed in several publications that a cohomological theory, similar etale
cohomology, could provide great insight into the nature of fractals and their
spectrum. It is with this motivation that we explore one of the fundamental
results involving cohomology, the Atiyah-Singer index theorem, and begin to
construct the pieces to an analogous construction for fractals.

The Atiyah-Singer index theorem is itself one of the greatest results of the
20th century. It bridges the gap from algebraic topology to differential equations,
a task that seems more miraculous the more you understand these subjects. In
order to understand this deep theorem, we will first introduce the concepts
essential to the theory of vector bundles and then develop de Rham cohomology
with vector bundles in mind. We will next cover the basics of topological
K-theory and tie this into our discussion of de Rham cohomology. Then, we will
introduce the bare essentials of differential operator theory as this theory does
not play a role in the direction we investigate fractals. We will conclude the
non-fractal discussion by giving a K-theoretic proof of the Atiyah-Singer index
theorem. Finally, we will introduce the concept of fractals, begin to develop a
theory similar in spirit to etale cohomology, concluding with a discussion of how
this work fits in to the bigger picture.

7
Part I

Bundles and Cohomology

8
Chapter 1

Bundles

One of the more frequently used ideas in math is that of local triviality. Smooth
manifolds, topological manifolds, fibrations, and jet bundles are all examples
of the concept being applied with the intention of generalizing fundamental
properties. When speaking generally, we will choose to denote a space with
such a set of fundamental properties as 1. Often the notion of generalization
is achieved by replacing equality with the isomorphisms in some category with
information relevant to 1. Both fiber bundles and vector bundles make use of the
idea of local triviality to generalize the structures we need when we do calculus.
Intuitively, up close a bundle resembles the product of a small neighborhood of
euclidean space with another small neighborhood of a different euclidean space.
A fiber bundle permits both spaces to be globally non-euclidean, so long as the
second space has topological structure. A vector bundle restricts this slightly
by requiring the second space to be a vector space isomorphic to Rn .

9
1.1 Fiber Bundles

Definition 1.1.1. A morphism of topological spaces, : X Y , is locally


trivial if for all points y Y there exists an open neighborhood U of y such that
we can find an isomorphism, called a trivialization, : 1 (U ) 1, where 1 is
trivial relative to U by definition.

This is equivalent to requiring that the following diagram commute.


1 (U ) 1

id|U
U

Our definition of local triviality is extremely general in that both the trivial
structure and category of isomorphism are left unspecified. We shall not make
full use of the generality of this definition, rather we will encounter specific cases
of it. For example, in the case of fiber bundles the structure we define to be
trivial relative to U is the product structure U F ' 1. Furthermore, should
we be working with bundles of a particular smoothness, then the isomorphism
category may be restricted to maps of an equivalent degree of smoothness, i.e.
 
C k hom Mank .

Definition 1.1.2. A fiber bundle is a quadruple (E, B, F, ) where {E, B, F }


Top, and is a continuous surjection from E onto B satisfying the local
triviality condition where the trivialization is a homeomorphism onto the product
space U F .

This is a special case of the commutative diagram for local triviality,

10
hom(Top)
1 (U ) U F


proj1

It is common to refer to B as the base space, F as the fiber, and E as either


the total space or as the bundle when is understood. An immediate result
of this definition being quantified over all points (1.1.1) of the base space is
that we can find an open cover of the base space where every element of the
cover has an associated trivialization. The agreement of these trivializations on
intersections of open sets is not a given for any collection of trivializations. In
fact, unless E is the trivial bundle, there is going to be some open set where
trivializations disagree, however, one can always find a map in the appropriate
category which resolves disagreement, simply as a result of both trivializations
being isomorphisms onto isomorphic spaces.

Definition 1.1.3. Let G be a topological group with faithful action1 on the fiber
F of a bundle E with base B. A G-atlas, A, is a collection of pairs (U , )
where U U, U an open cover of B, and a trivialization of U . This
collection must satisfy the agreement condition that the transition functions are
continuous with values in G, g (x) = 1 |{x}F G.

One important characteristic to notice is that the structure group is not


unique. In fact every fiber bundle has structure group Iso(F ). Rarely is it the
case that one works with all of Iso(F ). If we can identify the maps necessary to
the action of a particular bundle with elements of a subgroup H of Iso(F ), then
we say that the structure group can be reduced to H. The key here is that we
cannot reduce the structure group of every bundle to the trivial group. A bundle
1 Recall that an action is faithful if only the identity has trivial action.

11
Figure 1.1: The classical Mobius strip as a fiber bundle where
the base manifold S 1 is shown in bold.

is said to be trivial is we can reduce its structure group to te trivial group. Often
when constructing a bundle we work with the most reduced structure group.

We can think of this less abstractly by working in the category of smooth


manifolds thus defining G as a subgroup of the diffeomorphism group on F ,
Diff(F ), where to each transition function we can associate one specific element
of G. This unique association is made possible since the action of G is faithful
which forces the subgroup of elements acting trivially on F to be the trivial
group. Intuitively, G does the twisting to a bundle which would otherwise be
a trivial bundle. We can see this by constructing the Mobius strip as a fiber
bundle over S 1 and comparing it to the trivial bundle over S 1 . First we must
know how to take a group G, which we will refer to as the structure group, and
construct a bundle out of a fiber and base space. This is done by taking the
quotient of the trivial bundle by the equivalence relation (x, y) (x, g (x)y)
where x U U , and y and g (x)y are the fiber coordinates for U F and
U F respectively.
a 
E= U F /(x, y) (x, g (x)y)

For the trivial bundle over S 1 with fiber R the structure group is simply the
trivial group, making the bundle clearly isomorphic to a cylinder. To make the
Mobius strip we again take S 1 as our base space and R as the fiber, however,

12
this time we let the structure group be minimally reduced to Z/2Z.2 We can
construct a G-atlas by covering S 1 with two open sets S 1 {p} and S 1 {q},
where p 6= q, and defining the transition maps to be trivial and multiplication by
1 in the upper and lower intersections respectively. This forces the fibers of the
bundle to twist smoothly3 relative to each other on the intersection associated
with the nontrivial transition map.

1.2 Vector Bundles

Vector bundles intuitively formalize two notions in geometry and algebra. We


can understand vector bundles geometrically as parameterizations of vector
spaces over enriched curves called connections, which describe how to transform
a vector through the vector spaces over the curve. Alternatively, we consider
vector bundles as an algebraic answer to the question, Given a manifold, in
what space do the vectors over a point exist?

Definition 1.2.1. A vector bundle, V , is a fiber bundle where the fiber is a


vector space. If the fiber is specifically isomorphic to Rn or Cn then V is a real
or complex vector bundle, respectively.

One of the trickier points about manifolds is where the manifold lives.
Often when visualizing manifolds, we unthinkingly embed them, usually in some
Euclidean space. In truth it is a mistake to think about where manifolds live; a
manifold is the whole space, it does not need to live anywhere. Since a manifold
is a place where things live, rather than something which lives in a place, the
2 Precisely, we mean the group formed by the identity on R and the map : x 7 x where
the group operation is composition. In the algebraic fields of math, it is common practice
to work with groups via an isomorphic cyclic integer group, should one exist. The structure
group here is naturally isomorphic to Z/2Z.
3 If the twist were not smooth then at least one of the trivializations would not be isomorphic

to R2 .

13
natural question to ask is what are the properties of its inhabitants, specifically
curves and their tangent vectors. Awkwardly, while curves live comfortably
inside a manifold, their tangent vectors require a space of their own.

Definition 1.2.2. Given a smooth manifold M , the tangent space, Tp M , at a


n o
point p is the vector space over R spanned by x1(p) , . . . , xn(p) , where n is
the dimension of M . The tangent bundle of M , T M , is the bundle where the
fiber over a point is its tangent space.

We also define the dual structure, the cotangent bundle, whose importance
will be made apparent in the discussion of cohomology.

Definition 1.2.3. The cotangent bundle T M of a smooth manifold M is the


algebraic dual to T M . Specifically, the fiber over a point p is the dual vector
space of Tp M .

Intuitively, we can think of the tangent bundle as gluing a copy of Rn to every


point on M , thus lending familiarity to the vector space, as opposed to the more
abstract vector space of partial differential operators. On the one hand, this is
more than legitimate since the two vector spaces are isomorphic by definition,
however, this local replacement should not be thought of as implying that the
tangent bundle is globally trivial. For example, the tangent bundle of S 2 can
be shown to be nontrivial as a famous result of the hairy ball theorem. Given a
manifold, if the tangent bundle is trivial we say the manifold is parallelizable.
Noticing that we can locally treat the cotangent bundle the same way, one may
ask if they are the same bundle. To answer this we need a formal equivalence
for vector bundles.

Definition 1.2.4. Given two vector bundles V and W over the same base space
B, if there exists f : V W such that for all b in B, (f V1 )(b) ' W
1
(b),
then V ' W .

14
Simply put, two vector bundles over the same base space are isomorphic
if there is a homeomorphism that is a vector space isomorphism on the fiber
above each point. If f only induces a homomorphism on the fibers, then it is
simply a vector bundle homomorphism. As vector bundles, the tangent and
cotangent bundles are indeed isomorphic, however, it should be noted that they
are distinguishable by other properties.4

1.3 Operations and Bundle Maps

As is often the case in topology, what we are really interested in are the different
maps between objects rather than the objects themselves. One important
consideration is what the morphisms should be in order to make a collection
of bundles into a category. To this end, the structure relevant to a bundle is
the relationship between the fiber on a base space and the total space. This
motivates the following definition.

Definition 1.3.1. A bundle morphism between two bundles, E and E 0 , is a


pair of maps (u, f ) where u : E E 0 and f : B B 0 such that the following
diagram commutes.

u
E E0
0
f
B B0

When the base space is unchanged we end up with just the map u, thus
in definition (1.2.4), the map f is a bundle morphism, specifically a bundle
isomorphism. Thus we can form the category Bun and the subcategory BunB
4 One such distinguishing feature is that the cotangent bundle has a canonical symplectic
structure whereas the tangent bundle does not.

15
over a particular base space. Another task for maps, other than defining a
category, is defining a relation. We can make use of bundle isomorphisms as an
equivalence relation and define Vect(B) to be the set of all vector bundles over
B up to isomorphism. Vectn (B) is the subset of Vect(B) of vector bundles with
rank n.5

Definition 1.3.2. Given two vector bundles E and E 0 over the same base B,
we define their direct sum, or Whitney sum, E E 0 as the bundle over B where
the fiber over a point p is 1 (p) 01 (p).

The pair (Vect(B), ) forms a semigroup which is naturally abelian since


the direct sum of vector spaces is abelian. As we shall see later, associating this
semigroup to an abelian group with certain properties is the main content of
K-theory. Similar to the Whitney sum, the definitions for dual bundle and tensor
product are defined on the fiber, specifically the dual bundle has fiber the dual
spaces of the original fiber and the tensor product has fiber the tensor product of
the original fibers. As one gets with these operations, we will denote the exterior
Vk N
algebra as E = k E. These bundle operations lead to a very useful lemma
which allows us to make sense of the homomorphism bundle Hom(E, E 0 ).

Lemma 1.3.1. E E 0 is isomorphic to Hom(E, E 0 ) .

Proof. The universal property of tensor products states that when is bilinear,
there is a unique linear map which makes the following diagram commute.

h
E E0 E E0

Hom(E, E 0 )
5 The rank of a vector bundle is the dimension of the fiber with respect to the category of
the fiber. For example, a bundle with fiber R and a bundle with fiber C both have rank 1.

16
Thus, we need only find a bilinear surjection6 : E E 0 Hom(E, E 0 ).
Let us define as the map (, v 0 ) 7 [v 7 (v)v 0 ] where v E, v 0 E 0 , and
E . is clearly bilinear, and we can see it is surjective by noticing that
takes basis elements of E to basis elements of E 0 when we let be the Kronecker
delta. Thus there is a set in the image of which spans Hom(E, E 0 ).

Considering not just bundle morphisms, but also morphisms on their bases,
we define an important bundle construction.

Definition 1.3.3. Given a smooth map f : M N and a bundle E over N , the


bundle induced by f , or the pullback bundle f E, is defined to be the subbundle
{(m, e) M E|f (m) = (e)}, i.e. the bundle such that the following diagram
commutes.
2
f E E
1

f
M N

Intuitively, the pullback bundle, or the bundle induced by f , is the bundle


over M where the fiber over a particular point is the fiber over the image of that
point under f . The induced bundle construction is extremely important as it
lets us make a claim about homotopic spaces, which is essential to understanding
any topological theory.

Theorem 1.3.4. Given a vector bundle E with compact base B,7 if f0 : X B


and f1 : X B are homotopic maps for some compact manifold X, then
f0 E ' f1 E.

Proof. [8](6.8) Let f : X I B be a homotopy between f0 and f1 , and let


I : X I X be the projection for the trivial bundle over X. Suppose,
6 dim(E E 0 ) = dim(Hom(E, E 0 )), thus need only be a surjection to be bijective.
7 This theorem also holds for more general CW spaces. For a proof of this see [25] (1.11).

17
for some t0 in I, that ft0 E is isomorphic to a bundle E 0 over X. By the
previous lemma we know that Hom(f E, I E 0 ) is well defined as a bundle.
Let Iso(f E, I E 0 ) be the subbundle of isomorphisms. Since the fibers of this
bundle are all isomorphisms between f Ep and I Ep0 , any global section is
an isomorphism between the bundles. Iso(f E, I E 0 ) has a section over the
submanifold X t0 X I by our hypothesis (f E, I E 0 ) ' E 0 . We will
denote this isomorphism by . Since X is compact, we can find a cover C of
X with finitely many elements of the trivialization of Hom(f E, I E 0 ). Since
Hom(f E, I E 0 ) is well defined as a vector bundle, the fibers are isomorphic to
Euclidean space, meaning on C we can linearly extend our isomorphism to
an isomorphism on C. By the compactness of I we can repeat this process,
covering I with finitely many trivial sets and extending linearly, thus extending
to an isomorphism across all t I as well. As {0, 1} I, it follows that
f0 E ' f1 .

Corollary 1.3.5. For homotopic compact manifolds X and Y , Vect(X) '


Vect(Y ).

The intuition behind this proof is that we showed that along the time variable
of the homotopy, two induced bundles that are near enough are isomorphic.
Since I is compact, it follows by induction on a finite cover that the bundles
induced by the boundaries of the interval are isomorphic. Furthermore, the
corollary that homotopic spaces have isomorphic Vect() semigroups, follows
immediately from the definition of homotopy type.

Thus far, we have only worked with bundles by taking fibers to points, maps
which take points to elements of their fiber are called sections.

Definition 1.3.6. s : B E is a section of E if s = 1B .

We denote the space of all smooth sections of a bundle E as E. A trivial


but important example of a section is the zero section s0 which takes a point

18
in the base space x X and pairs it with the zero vector in its respective fiber
(x, 0). The zero section is clearly a natural left inverse of for any bundle, and
not an inverse on the right. In (2.3) we will prove the Poincare lemma which
will require showing that the zero section composed with on the left or right
is homotopic to the identity in cohomology.

1.4 Orientation

Up to this point, the reader might be under the impression that a vector
bundle is quite simply a bigger notion of a manifold. While it is true that
many manifolds are expressible as vector bundles, one of the properties which
distinguishes the two is orientability. Traditionally, a manifold is orientable
if there is a globally consistent choice of whether a neighborhood is locally
right-handed or left-handed. This notion can be formalized by specifying
that the transition functions of the manifolds atlas have a positive Jacobian.
Equivalently,8 a manifold is orientable if and only if there exists a non-vanishing
global volume form.9 The former is more readily generalizable to vector bundles.

8A proof of this can be found in [8] (3.2).


9A volume form is an n-form where n is the dimension of the manifold.

19
Chapter 2

Cohomology

Originally constructed by Poincare to prove his duality formula, cohomology


is much more than simply the dual construction to homology. One of the key
differences which motivates its study is that, unlike homology, cohomology has
a natural ring structure. This ring structure opens the door to many powerful
tools and theorems, not the least of which is the theory of characteristic classes.
The goal of this chapter is to introduce some of the general ideas of cohomology,
its computation, and the background material necessary to understanding the
Thom, Euler, and Chern classes.

2.1 Differential Forms

Intuitively, we would like to have some way to discuss the behavior of some
smooth function on a smooth manifold in relation to that manifolds geometry.
Differential k-forms are operators that take in a functions tangent vectors and
map them into R. We can then put all functions that bend a certain way into
the same class by specifying a particular value that a class of k-forms map those

20
functions to. This is motivated in particular by the beautiful algebraic structure
that is created from classes of k-forms: differential cohomology. Since we will
only be dealing with smooth forms, we will use the terms differential k-forms
and k-forms interchangeably.

Definition 2.1.1. Given a manifold M , a k-form is a smooth section of the


Vk
exterior algebra of the cotangent bundle, : T M R.

Given a manifold of dimension n, we can form an algebra over R from


the vector space of differentials, , with basis {1, dxi , dxi dxj , . . . , dx1 . . . dxn }
where the indices are ordered, together with a formal product satisfying the
relations,
dxi dxi = 0 (2.1)

dxi dxj = dxj dxi (2.2)

A k-form can be expressed as the formal sum of elements of the set k (M ) =


{fi1 ...ik dxi1 ...ik | fi1 ...ik C }. The collection of all formal sums of k (M )
forms the cochain algebra k (M ). The double index is necessary since we can
have k-forms that operate on different elements of the cotangent bundle. For
example, for M = R3 ; x2 dx, xy dz, and x2 dx + xy dz are all in 1 (R3 ). When
all of the k (M ) are formally summed together we get the naturally graded
Ln
algebra (M ) = 0 k (M ). Furthermore, the product defined on extends
naturally to a product on (M ), called the wedge product.

Definition 2.1.2. Given = fI dxI and = gJ dxJ in (M ), we define


the wedge product as = fI gJ dxI dxJ .1

The smoothness condition in (2.1.1) manifests itself in the smoothness of the


functions fI . This smoothness is similar to the traditional notion of smoothness
1 The wedge in dxI dxJ is suppressed both for clarity as it is not to be confused with
the product being defined, and since it is sensible to be understood as the product previously
defined on .

21
for functions on a manifold. For a given atlas, one can either require that
the transition maps act as the identity on function restricted to any given
intersection of charts, or one can force the matter by taking a partition of unity
subordinate to the atlas and piecing together the functions. Both are equivalent
and we shall make use of both notions noting that for the latter procedure, in
the countable case, it readily lends itself to induction.

The use of standard C functions as coefficients of the differentials means


that integration of differential forms is extremely similar to ordinary integration
on manifolds. Let = f (x1 , . . . , xn ) dx1 . . . dxn be an n-form on Rn with
compact support. Integrating is done simply by taking the Riemann integral
with the caveat that the order of the differentials is unsigned, that is
Z Z Z
= ... f (x1 , . . . , xn ) |dx1 . . . dxn |
Rn Rn

Furthermore, suppose M is an oriented n-manifold with atlas A = (U , )


and is a smooth n-form, where smoothness is satisfied by a partition of
unity U = (U , ) subordinate to the cover S = U . Since is stitched
together from components defined on each element of the partition of unity,
we define integration simple as the sum of the stitched together components,
each integrated in Rn after pulling back by the trivializations.

Z XZ
= (1
) ( )
M Rn

Finally, given a vector bundle E with coordinates x = (x1 , . . . , xn ) in the


base and coordinates t = (t1 , . . . , tn ) along the fiber, we define integrating
vertically, or along the fiber, as the integral of the fiber component. We will
denote this particular integration map by . More precisely, given some form
((E)), and = ( ())f (x, t)dtI ,
Z
: f (x, t) dtI
F

22
Observe that since we are integrating over the whole fiber, the integral is
only well-defined when f has compact support. This form of integration will
allow us to partition forms on vector bundles into two classes, those that have
all the differentials of the fiber and those that do not. This distinction is made
apparent by the fact that if any of the fiber differentials are missing from a form,
maps that form to 0 by standard integration.

2.2 de Rham Cohomology

Considering that differential forms are by definition smooth, it is natural to ask


how to associate two forms where the coefficient of one is the derivative of the
other. For example, xy 2 dxI and 2xy dxJ + y 2 dxK . Notice that this derivative
operator should raise the degree of the form to target a specific variable in each
component, analogous to how the classical derivative targets a specific variables
and separates functions. For example, the gradient of a function f : R3 R
D E
splits the function by variable into the vector f f f
x y z . Furthermore, as all
, ,
differential structures do, differential form differentiation should satisfy Leibnizs
law.

Definition 2.2.1. The exterior derivative d : k (M ) k+1 (M ) is defined


recursively as follows,

f
1. for f 0 (M ), df = xi
dxi

2. for = fI dxI , d = dfI dxI .

Theorem 2.2.2. d satisfies the Leibniz rule.

Proof. [8](1.3) By linearity, we only need to check this for monomials. Let

23
= fI dxI and = gJ dxJ .

d(fI gJ )dxI dxJ = d(fI )gJ dxI dxJ + fI d(gJ )dxI dxJ

d( ) = (d ) + (1)deg (d)

The naming of d is motivated by the fact that it is exterior to each level


of the graded algebra (M ). As a simple example of exterior differentiation,
consider a general 2-form on R2 , = f1 dxdy.

P
d = dfI dxI

= (df1 ) dxdy
 
f f
= x dx + y dy dxdy

f f
= x dxdxdy + y dydxdy

f
=0 y dydydx

=0

Most of the time differential forms end up being zero because of the internal
algebra rules (2.1) and (2.2). If the exterior derivative of a differential form is 0,
the we call that form closed. Some forms are more destined for 0 than others,
specifically those that are already the exterior derivative of a form.

Theorem 2.2.3. d2 = 0

Proof. [8](1.4) We only really need to check this in the case of 0-forms, as all
other forms inductively reduce to this case by the definition of d.
X  XX
2 f f
d f =d dxi = dxi dxj
xi xi xj

24
By Clairauts theorem the mixed partials are symmetric, and by (2.2) the
differentials are antisymmetric. Thus, all of the terms in the double sum
cancel.

An important characteristic of differential forms is the notion of a pullback.


Given a smooth map f : M N , f induces a map f : k (N ) k (M ) by
composition on the right. That is, for some k (N ), f () = f . It is quite
a simple idea, which is intuitively straightforward, resulting from fact that the
composition of differentiable functions is differentiable, thus the pullback can
be thought of as a coordinate transformation from the tangent bundle of N to
the tangent bundle of M . Deceptively simple as it is, the pullback is a defining
feature of forms of any theory. For example, a saying that has been handed
down through at least three known generations of topology classes is: forms
can be pulled back, measures can be pushed forward.

Lemma 2.2.1. [8](2.1) Let f : M N be a smooth map. f , the pullback of


f , commutes with d.

Proof.

df (gI dxi1 . . . dxiq ) = d((gI f )dfi1 . . . dfiq ) = d(gI f )dfi1 . . . dfiq

n
!

X gI
f d(gI dxi1 . . . dxiq ) = f dxi dxi1 . . . dxiq
i=1
xi
n   
X gI
= f dfi dfi1 . . . dfiq
i=1
xi

= d(gI f )dfi1 . . . dfiq

If there exists a (k 1)-form such that d = , then is called exact.


By (2.2.3) it follows that every exact form is also a closed form. de Rham

25
cohomology is an explicit answer to the question of which forms are closed but
not exact. In order to define de Rham cohomology, we must first build a general
framework for any differential-type cohomology.

C k together
L
Definition 2.2.4. A cochain complex is a graded algebra C = kZ

with a set of homomorphisms dk : C k C k+1 such that for all k, dk+1 dk = 0.2

When (M ) is paired with the defined differential operator d, we have the


de Rham complex, which we choose to also denote with (M ). Suppressing the
d in the notation should be clear since from this point on the de Rham complex
is the only object which will be referred to by (M ). With the de Rham
complex well-defined, we are now able to define several varieties of differential
cohomology.

Definition 2.2.5. Given a manifold M of dimension n, we may construct its


de Rham complex (M ). From this cochain complex we can construct the
sequence
d1 d dn1 d
0 0
0
. . . n1 n
n
0

The i-th cohomology group is defined as

H i (M ) = ker (di )/Im (di1 )

Definition 2.2.6. Given a manifold M of dimension n, if we consider only


differential forms with compact support we can define a new cochain complex
0 (M ) = {fI dxI (M ) | fI C0 (M )}, then we can construct the
sequence,
d1 d dn1 d
0 00
0
. . . n1
0 n0
n
0

The i-th compact cohomology group is defined as

H0i (M ) = ker (di )/Im (di1 )


2Z is motivated by preserving k = 0 to be assigned to d0 : C 0 C 1 while permitting
the inclusion of a map d1 : 0 C 0 . Notice, choosing the natural numbers would force our
module of 0-forms to live in C 1 .

26
Definition 2.2.7. Given a vector bundle E of rank n and base M , if we consider
only differential forms with compact support along the fiber cv (E) = {
(E) | K K(M ), 1 (K) Supp()}, then we can construct the sequence,

d1 d dn1 d
0 0cv
0
. . . n1
cv ncv
n
0

The i-th compact vertical cohomology group is defined as

i
Hcv (M ) = ker (di )/Im (di1 )

The beauty of cohomology isnt the variety of theories, nor is it the way
computing it reduces to simple arithmetic.3 The real beauty of cohomology is
that, unlike homology, there is a natural ring structure. In singular cohomology
one uses the cup product as the ring product, however, for de Rham cohomology,
the cochain complex already has a preferred product.

Lemma 2.2.2. The wedge product on forms, induces a well-defined product in


cohomology [] [ ] = [ ].

Proof. Let , ker(d) and Im(d). We choose ( + ) as a representative


for [].
( + ) = +

We must now verify that is exact. Since is exact, let d = .

d( ) = (d) + (1)deg() (d ) = + 0

Theorem 2.2.8. Paired with formal sum and wedge product with respect to
equivalence classes, de Rham cohomology satisfies the ring axioms.

Proof. Since the additive operation is formal sum, it is trivial that (H (M ), +)


is an abelian group. In terms of the multiplicative operation, (H (M ), ), we
3 See Appendix (B)

27
have 1 = thus the class [1] serves as the multiplicative identity. It also
follows that [] ([ ] []) = ([] [ ]) [] since the wedge product on forms
is associative, and the induced product on classes is well-defined. Likewise,
distributivity follows due to the distributivity of the wedge product on forms,
and the product on classes being well-defined.

This ring structure is defined and holds likewise for both compact and
vertically compact cohomology, and the proofs run identically. As an example of
a de Rham cohomology ring, consider the 2-sphere. The de Rham cohomology
groups are H 0 (S 2 ) = H 2 (S 2 ) = R, generated by 0 and 2 respectively, and
H 1 (S 2 ) = 0.4 For 0 we can always choose the 0-form 1 from the class [1] as
a representative. Furthermore, for any we have 1 = thus, 0 is our
multiplicative identity. Considering this, our ring looks like graded products of
the form 1, 1 2 , 2 2 = 0 with coefficients in R. Thus, H (S 2 ) ' R[x]/(x2 )
where x is a generator for H 2 (S 2 ).

2.3 Poincare Lemma

In this section, we will prove that de Rham cohomology is homotopy invariant.


For a background in homotopy, one should consult [12] chapter 0. It is worth
noting that this lemma is not true for compact cohomology.

Lemma 2.3.1 (Poincare Lemma).



H0 = R
H (Rn ) = H (point) =
H n = 0, for n > 0

Proof. [8](4.1)The proof of the Poincare Lemma follows by induction based on


4 For a computation of this see Appendix (B)

28
the cohomology of the trivial real line bundle5 over Rn and the cohomology
of the base, Rn , being isomorphic. To establish this fact we must show that
the bundles projection map and the zero section s0 induce isomorphisms in
cohomology. This is equivalent to showing s0 = 1 and s0 = 1. The
former is trivial since s0 = 1; however, to prove the latter we will define an
operator K such that 1 s0 = (dK Kd). Notice that if we let be
a closed form, then d(K) is exact since it is d of K and K(d) = K(0)
is trivially exact. Thus, dK Kd maps closed forms to exact forms, inducing
the zero map in cohomology. This implies that s0 is equivalent to the
identity. This construction is used fairly often, as such operators like K are
called homotopy operators, and this particular type of equivalence is called a
chain homotopy.

As discussed in (2.1), the forms on a vector bundle can be separated into


those which have all the differentials of the fiber, and those which do not. We
make that distinction here by defining a form to be of type 1 if = ( )f (x, t),
and type 2 if = ( )f (x, t)dt. We define K as the operator that maps forms
Rt
of type 1 to 0, and forms of type 2 to ( ) 0 f (x, t). The proof now reduces
to computation, verifying 1 s0 = (dK Kd) for forms of type 1 and
then type 2. One subtlety of these computations is that any map induced on
forms technically induces two further maps: one which acts as a pullback on
the variables in the coefficient function, and one which acts as a pullback on the
variables in the differential. This considered, s0 (dt) = 0 thus s0 (f (x, t)dt) = 0.

We now verify the equality 1 s0 = (dK Kd). Let be a form of


5 The line bundle of a space X is a vector bundle E with base X and fiber F of dimension
1. Naturally the trivial real line bundle is the vector bundle X R.

29
type 1.

(1 s0 ) () = ( ) f (x, t) ( ) f (x, 0)
(dK Kd) () = Kd ()
  
deg()
= K (d ) f (x, t) + (1) f
x dx +
f
t dt
t
= (1)deg()1 0 f
R
t dt

= (1)deg()1 (f (x, t) f (x, 0))

Let be a form of type 2.

(1 s0 ) () = 0
Rt R  
t f
dK () = ( d) 0 f (x, t)dt + (1)deg()1 ( ) 0 x
dt dx + f (x, t)dt
Rt R 
deg()1 t f
Kd () = ( d) 0 f (x, t)dt + (1) ( ) 0 x dt dx
(dK Kd) () = (1)deg()1 ( d) f (x, t)dt = (1)deg()1

Given some manifold M , it trivially holds that H (M R) = H (M ) since


the above lemma holds on every chart of a given atlas.

Lemma 2.3.2. Homotopic maps induce equivalent maps in cohomology.

Proof. [8](4.1.2)Let F : M R N be a homotopy of the maps f, g : M N


such that F (x, t) = f (x) for t 1 and F (x, t) = g(x) for t 0, and s1 : h(x) 7
h(x, 1) be the 1-section. By definition, f = F s1 and g = F s0 . Thus the
induced maps are f = s1 F and g = s0 F . By the previous lemma, we
know that the 0 and 1-sections both invert which implies s1 = s0 . Thus,
f = g .

Theorem 2.3.1. de Rham cohomology is homotopy invariant.

Proof. Suppose M and N have the same homotopy type. Then there exist
maps f : M N and g : N M such that f g and g f are homotopic

30
to the identities idM and idN , respectively. Thus, by the previous lemma,
H (M ) ' H (N ).

2.4 Leray-Hirsch Theorem

The Leray-Hirsch Theorem is a generalization of the Poincare lemma in the sense


that it expresses the cohomology of a fiber bundle as a product involving the
cohomology of the base and the cohomology of the fiber. The Kunneth Theorem
is a special case of the Leray-Hirsch Theorem when the bundle is trivial.

Theorem 2.4.1 (Kunneth Theorem). Let (E, M, F, ) be a fiber bundle such


that there exist global forms e1 , . . . , er which, when restricted to F , generate
H (F ), and M has a finite cover C such that all nonempty finite intersections
are diffeomorphic to Rn .6

H (E) ' H (M ) H (F )

Proof. [8](5.11) Let : E F be the natural projection map onto the fiber.
Define : H (M ) H (F ) H (E) as : [] [] 7 [ ], where
(M ) and (F ). Let U and V be open sets on M and let n N.
Since the tensor product is right exact, tensoring in the cohomology of the fiber
to a Mayer-Vietoris sequence preserves exactness. This must be done carefully
in order to preserve the degree of the cohomology group, thus for any element of
the sequence H p ( ) we will tensor with H np (F ) and then sum the product over
p from 1 to n. Dimension is additive over this sum, meaning the dimensions on
the left of the following diagram match the dimensions on the right. We aim to
6 Note that the supposition implies each element of the cover is itself diffeomorphic to Rn .
Any cover with the properties of C is called a good cover.

31
show that this diagram commutes.

... ...

n
[H p (U V ) H np (F )]
L
H n ( 1 (U V ))
p=0

n
[(H p (U ) H np (F ))
L
H n (U F )
p=0

(H p (V ) H np (F ))] H n (V F )

n
[H p (U V ) H np (F )]
L
H n ((U V F ))
p=0
d d
n
[H p+1 (U V ) H np (F )]
L
H n+1 ( 1 (U V ))
p=0

... ...

We begin by checking commutativity in the first square. Let be in


n
[H p (U V ) H np (F )].
L
p=0

( ) = ((|U , |V )) = ( |U , |V )

( ) = ( ) = ( |U , |V )

It should be noted that since H p (U V ) is defined globally on U V , the


restriction map : H n ( 1 (U V )) H n (U F ) H n (V F ) is well defined.
Likewise, to check commutativity in the second square let ( , ) be in
n
[(H p (U ) H np (F )) (H p (V ) H np (F ))].
L
p=0

(( , )) = (( ) ) = ( )

(( , )) = (( , )) = ( ) (2.3)

Note that in (2.3) we can simplify to ( ) because the wedge product


distributes. Finally, we check commutativity of the third and trickiest square.

32
n
[H p (U V ) H np (F )]. Note that the map
L
Let be in
p=0

n
M n
M
d : [H p (U V ) H np (F )] [H p+1 (U V ) H np (F )]
p=0 p=0

is defined as d : 7 (d ) . This is because, if d acted on , then it


n
[H p+1 (U V )
L
would raise its degree which would change the codomain to
p=0
H n+1p (F )]. However, H np (F ) is the space we tensored the Mayer-Vietoris
sequence with, thus d cannot act on the fiber. Furthermore, let {U , V } be a
partition of unity subordinate to {U, V }. On (U V ) F we have

d ( ) = d(( U ) )

= (d (U ))

= (d)

Since we are working in cohomology, is at least a closed form, thus what would
be the left side of the Leibniz rule vanishes in the above equations. Therefore,
the third square can be shown to commute.

d ( ) = ((d ) ) = (d )

d ( ) = d ( ) = (d )

Since U , V , and U V are all diffeomorphic to Rn , their respective maps in the


above diagram are isomorphisms by the Poincare lemma. By the 5-lemma,7 it
follows that for U V is an isomorphism. Assume now that the theorem holds
when C has p elements. Suppose now that C = {U0 , . . . , Up+1 } for M has p + 1
Sp
elements. We can find a cover D of ( n=0 Un ) Up+1 with the same properties
as C. D is precisely {Un Up+1 }pn=0 . By hypothesis, the theorem holds for
Sp Sp
n=0 Un , Up+1 , and ( n=0 Un ) Up+1 , thus it holds for C. The theorem now

follows by induction on the cardinality of C.

7 See Appendix (A)

33
Chapter 3

Characteristic classes

The ideas of characteristic classes were around long before cohomology theory.
The Euler characteristic, a classical invariant on polyhedra noticed by Descartes
as early as 1639, is perhaps the first characteristic class. The core principle of
a characteristic class is to distinguish objects based on their local features. The
Euler characteristic, predating algebraic topology, did exactly this by keeping
track of vertices, edges, and faces, distinguishing spherical polyhedra1 into the
class 2. With the foundation of algebraic topology laid by Poincare, modern
characteristic classes were constructed to deal with a more rigorous notion of
locality. The problem was how to overcome the obstructions of extending a
local continuous map to a global one. The primary obstruction is a space
being twisted, which forces any continuous map to also locally twist, potentially
inducing a map which doesnt line up globally. The classes presented here
were all birthed as solutions to keeping track of this local twisting via a globally
defined function.
1 Spherical polyhedra can be thought of as a generalization of the class of polyhedron
containing the Platonic solids.

34
3.1 Orienting Vector Bundles

Definition 3.1.1. Let E be a vector bundle. E is orientable if its structure


group can be reduced to SO(n).

Loosely, this means that the transition functions for E can be chosen such
that all of their determinants are positive. Any trivialization of E that satisfies
this property is called oriented. Two oriented trivializations are equal if their
charts agree, specifically det( 1
) > 0 where {U , } and {V , } are

charts of different trivializations such that U V . A particular equivalence


class of orientations defines a particular orientation. Interestingly, a complex
vector bundle with its complex structure dropped2 is always orientable since
U(n) SO(2n).

Since we are working with a vector bundle, it follows from the fibers being
Vdeg F
vector spaces that the space F is one dimensional. Choosing a nonzero
Vdeg F
element of F is equivalent to choosing an orientation on the vector space
F . Choosing an orientation on E can thus be viewed as a consistent choice of
orientation on the fibers which varies smoothly. Being nonzero and smooth, it
follows that a vector bundle is orientable if there exists a nonvanishing volume
form. Such a form is called an orientation form.

Considering an orientation form locally, the local orientation form measures


the angular twisting between local coordinate systems. Note that these forms
are not locally equivalent, the orientation form is of top degree whereas the local
orientation form is a 0-form, however they are intimately related as we will see
in the construction of the Euler class in rank 2. Being a function, we more
often work with its exterior derivative since this trivially satisfies the cocycle
condition,3 and thus behaves equivalently to the transition functions up to the
2 By dropping complex structure, we simply mean replacing the fibers Cn with R2n .
3A set of local functions between charts satisfies the cocycle condition if =

35
specific angular rotation between charts.

3.2 Thom and Euler Class

Characteristic classes are invariants by which one can make rigorous the notion
of how much a bundle is twisted. The most straightforward way one might
approach the problem of measuring a bundles twistedness is, loosely speaking,
to record the amount of rotation that occurs as one moves between neighboring
fibers. This is intuitively equivalent to keeping track of how the tangent space of
a fiber changes as one moves over the base space, i.e. the global variations in the
cohomology of the fiber. The Thom class is exactly such a tracking mechanism.

Definition 3.2.1. Let E be an oriented rank n vector bundle. The Thom class is
n
the cohomology class in Hcv (E) which, when restricted to the fiber, generates
H0n (F ).

The existence of the Thom class is rather trivial to demonstrate, but it


requires material we will not develop until much later. This being said, we
shall prove, in a very general setting, the Thom isomorphism induced by the
Thom class. The Thom class itself can be quite cumbersome to work with if
one is trying to understand the twistedness of any bundle over a space M rather
than one particular bundle. This is because, while it is influenced by a bundles
projection map , it also requires explicit knowledge of the orientation induced
on the fiber. While the Thom class is extremely important and a powerful tool,
to be able to work with orientation and global cohomology via one class, we
are more interested in sections of bundles rather than their specific orientations.
Towards this goal we will prefer a class which is multiplicatively additive and be
on U U U . It is trivially the case that the transition functions of a vector bundle satisfy
the cocycle condition.

36
satisfied with navely treating the orientation information of a bundle to reach
this goal.

Definition 3.2.2. Let E be an oriented vector bundle. The Euler class e(E) is
the pullback of the Thom class by the zero section.

As an example of the Euler class at its finest, we will calculate the Euler
class of the trivial bundle over Rm . Let E be the trivial bundle Rm Rn . The
Poincare lemma for compact cohomology [8] (4.7.1) tells us that for E, H0n (F ) =
H0n (Rn ) = R. This implies that (E) is a class in Hcv n
(E) containing forms of
the type = f (x1 , . . . , xm , t1 , . . . , tn )dt1 . . . dtn such that F |F dtN = 1. As
R

a representative of this class, we choose the coefficient f to be the function


which takes (x1 , . . . , xn ) to 0 and is a bump function on the fiber coordinates
with total integral 1. Pulling back this form by the zero section maps all of
the fiber coordinates to 0 as well. Thus the Euler class for any vector bundle
over Rn is trivial. In fact, we can extend this to any trivial bundle over any
manifold by simply applying this calculation to a partition of unity subordinate
to a cover where every set and nonempty finite intersection is contractible.4 We
can extend this result further: any vector bundle with a non-vanishing section
has trivial Euler class.5 We can calculate the Euler class of several important
bundles, one case being the bundle LL . Employing (1.3.1), since Hom(L, L )
has a non-vanishing section given by the identity, it follows that e(L L ) = 0.
We will use this result later in the next section to derive an important identity
for the top Chern class.

Another important result we will be making use of is the explicit formula


for the Euler class of an oriented rank 2 vector bundle. The construction of
the formula is similar to the stitching done above, however we need to be more
4 For a proof that we can find such a cover for any manifold, see [8] (5.1).
5 Proving this extension is most naturally done when the Euler class is defined using relative
cohomology theory. For a relative cohomological proof of this extension see [1] (11.7.19).

37
careful with transition maps in the general case. Let E be an oriented rank 2
vector bundle. Since it is rank 2, we can specify that the structure group can
be reduced to SO(2). Due to this reduction, we can identify transition maps
g : U U SO(2) with complex-valued functions. This identification
induces an angular twisting on intersections of (1/i) log g . Since E is oriented,
we can find a local orientation form to identify with this local twisting d =
(1/i) log g . Letting {U , } be some partition of unity
! !
1/ d = 1/ d log g 1/2i
X X
2 2i d log g

We can cleverly concatenate this and express the Euler class of an oriented
rank 2 vector bundle as6

1 X
e(E) = d ( d log g ) (3.1)
2i

This directly implies the naturality of the Euler class, e f 1 E = f e (E)




where f is a smooth map from a manifold M to the base of E and f 1 E is the


induced bundle over N .

3.3 Chern Class

To define the Chern class we must first define several special bundles one can
construct when given a complex vector bundle.

Definition 3.3.1. Let E be a complex vector bundle with base M and transition
functions gab : Ua Ub GL(n, C). The projectivization of E, P (E), is the
bundle whose fiber above a point x in M is the projective space P ( 1 (x)), and
transition functions gab : Ua Ub P GL(n, C) induced by gab .
6 For details on the insertion of d in these formulas, as well as details on how this is
equivalent to pulling back the Thom class by the zero section, see [8] (70-75).

38
Definition 3.3.2. Let : E M be a complex vector bundle with base M ,
and : P (E) M be its projectivization. Let x be a point in M and lp be an
element of 1 (x), the pullback bundle 1 E is the bundle with base P (E) and
fiber 1 (X) at point lp .

Definition 3.3.3. Let E be a complex vector bundle and 1 E be the pullback


of its projectivization, also known as the tautological bundle. The universal
subbundle S is defined as

S = {(lp , v) 1 E|v lp }

Noteworthy in these definitions is first that the fiber above a point x in 1 E


is the trivial bundle 1 (x) 1 (x), and secondly that the universal subbundle
is a line bundle over 1 E. Given any complex vector bundle, we can always
reduce its structure group to the unitary group U (n).7 For a line bundle, we
have the isomorphism : U (1) SO(2) given by

cos sin
: ei 7
sin cos

Thus there is a bijective correspondence between a complex line bundle and


real oriented rank 2 bundles. We exploit this correspondence to define the top
Chern class of a complex line bundle L over M as c1 (L) = e(LR ) H 2 (M ),
where LR is the bundle L with the complex structure dropped from the fiber.
By the formula for the Euler class (3.1)8 and the triviality of L L we derive
7 For a proof of this see [24] (2.2.5)
8 The specific tool used in the derivation is the old arithmetic trick of log() splitting a
product into a sum.

39
the following property of the top Chern class of a line bundle.

c1 (L L0 ) = c1 (L) + c1 (L0 ) (3.2)

c1 (L L ) = c1 (L) + c1 (L ) (3.3)

0 = c1 (L) + c1 (L ) (3.4)

c1 (L) = c1 (L ) (3.5)

Let x = c1 (S ) H 2 (P (E)).9 Restricting S to the fiber over a point b in the


base M , P ( 1 (b)), is the same as taking the universal subbundle over that
particular fiber, i.e. S = {(lp , v) 1 1 (b)|v lp }. Note that S ' CP n1
by definition. CP n has a cellular decomposition with one even dimensional
cell per dimension, thus we can easily compute the cohomology ring to be
H (CP n ) = R[x]/(xn+1 ) where x is a generator for H 2 (CP n ).10 This implies
that H (P ( 1 (b)) = R[x]/(xn ), implying that 1, x, . . . , xn1 is a basis. The
naturality of the Euler class implies that considering the fiber is equivalent
to restricting x to P ( 1 (b)). Thus the classes 1, x, . . . , xn1 are global
classes which, when restricted, freely generate the cohomology of the fiber for
P (E). By the Leray-Hirsch theorem, H (P (E)) is a free module on M with
basis {1, . . . , xn1 }. Being a free module, we can write xn as a unique linear
combination of this basis where the coefficients are classes in H (M ).

Definition 3.3.4. The i-th Chern class of a complex vector bundle E with base
M , is defined to be the i-th coefficient of the expression representing c1 (S ).

c1 (E)xn1 + . . . + cn (E) = xn , ci (E) H 2i (M )

9 It is tradition to let the generator be the top Chern class of S .


10 For a more precise computation of this see [21] (18.4).

40
Part II

K-Theory

41
Chapter 4

K-theory

K-theory is, loosely speaking, the study of the algebraic properties of the types
of functors that map some category with commutative objects into the category
of abelian groups. Here we will be working exclusively in the category of
isomorphism classes of complex vector bundles. This means that whenever
we mention a vector bundle or the functor Vect( ), unless otherwise stated1 ,
we mean the complex type. The study of the particular K-functor that maps
Vect( ) into Ab is referred to as topological K-theory. Working with real vector
bundles is so different that it has its own title, KO-theory. As we shall see,
the beauty of topological K-theory is that not only does it have a natural
cohomology theory, but its relationship to other cohomology theories often stems
from powerful isomorphisms.
1 We consider it implicitly stated that certain bundles are necessarily real, like the tangent
and cotangent bundles.

42
4.1 K(X)

Let X be a compact manifold. Earlier we noted that Vect(X) is an abelian


semigroup when paired with the Whitney sum . In fact, by observing that the
isomorphism class of the bundle X {0} is an additive identity up to bundle
isomorphism, Vect(X) is an abelian monoid. To any abelian monoid M we can
associate an abelian group K(M ) known as its Grothendieck group. Since the
only monoid we will work with is Vect(X), we will suppress Vect( ) from the
notation and write K(X) for K(Vect(X)).

Definition 4.1.1. Let M be an abelian monoid. The Grothendieck group of M ,


K(M ), is the abelian group satisfying the universal property that there exists a
monoid homomorphism i : M K such that for any monoid homomorphism
f : M A to an abelian group, there exists a group homomorphism g : K A
such that the following diagram commutes.

i
M K
g
f
A

As universal definitions tend to be fairly unintuitive, another way of thinking


of K(M ) is as the free group completion of M . This alternative definition gives
us a way to explicitly construct K(M ) as the quotient of the free abelian group
generated by M and E(M ) = {a + a0 (a A a0 )|a, a0 A}, F (M )/E(M ).
The most fundamental example of this is K((N, +)) = Z. While this may seem
trivial, interestingly K(S 1 ) is also Z since Vect(S 1 ) is isomorphic to (N, +).
More importantly, Vect(pt) is also isomorphic to (N, +) thus K(pt) = Z.

For topological K-theory, K(X) has more structure than being only an
abelian group. We observe that Vect(X) has the product of tensoring fiberwise.
Examining the sequence 0 E(Vect(X)) , F (Vect(X))  K(X) 0 we

43
can view K(X) as the element making it exact. In this way, it is clear that
induces a product on K(X) making it into a commutative ring. Furthermore,
it is trivial that a map f : X Y which induces a bundle, also induces a ring
homomorphism f : K(Y ) K(X). Thus, functorially, K is a contravariant
functor from the category of smooth compact manifolds to commutative rings.
Making use of this contravariance we define the reduced K-group, which will be
fundamental to defining the K-group when X is only locally compact.

Definition 4.1.2. Let i : pt , X be the natural inclusion. The reduced


K-group of X, K(X), is defined as ker i .

As was the case when working with the projectivization of a bundle in


developing the Chern forms, we would like to also work with bundles over other
bundles in K-theory, specifically bundles over the cotangent bundle. However,
it is necessary that the fiber be compact for a fiber bundle to be compact.
Having fiber isomorphic to Rn , the cotangent bundle is never compact, but it
is locally compact. Thus we extend our definitions supposing now that X is
locally compact and X is the Alexandroff one-point compactification.2

Definition 4.1.3. When X is locally compact, its Grothendieck group is defined


as K(X) = K(X).

Intuitively, pt represents the one point in the Alexandroff extension and we


are quotienting out by its K-group. The category we choose to work in, of locally
compact spaces, necessarily has proper maps as its morphisms such that our
bundle constructions are all still well defined. One problem with this definition is
that if we want to work in Z, the usual reduction by i is trivially 0. We know via
embedding theorems that, at least for the cotangent bundle of a compact space,
we can embed it in sufficiently high dimension Euclidean space. If K(T Rn )
were Z, then we could construct a map i! : K(T X) K(T Rn ) by extending
2 When X is compact, X = X pt. pt is traditionally denoted by the symbol .

44
suitable inclusions. With this in mind, let us establish that K(T Rn ) ' Z.

Theorem 4.1.4. K(T Rn ) ' Z

Proof. Since T Rn is naturally isomorphic to R2n and R2n is contractible, it


follows from (1.3.5) that K(R2n ) = K(pt).

We also need the following special bundle to construct i! .

Definition 4.1.5. Let X and Y be manifolds and i : X Y be an embedding.


The normal bundle of X in Y is the bundle over X such that the following
sequence is exact.
0 T X T Y |i(X) N X 0

Intuitively, we can think of vectors in the fiber of N X as being the tangent


vectors in Y which were orthogonal to the tangent vectors of X. By the tubular
neighborhood theorem,3 can find a neighborhood N of X in T Y diffeomorphic
to N X. By the same theorem, it follows that T N is diffeomorphic to N (T X) in
T Y . To work with it in topological K-theory, we need to now equip T N with
complex structure. To do this, notice that for any bundle E over X, the tangent
bundle T E admits the following decomoposition.

T E = T X E

This follows from the fact that T E is given locally by the product of the base
and fiber tangent bundles, Tx U T Rm = Tx X Ex . Letting E be T X and
T Y |X , we obtain the following decompositions.

T (T X) = T X T X

T (T Y )|T X = (T Y |X ) (T Y |X )
3 See [13] (4.5).

45
Thus, the normal bundle admits the following decomposition which trivially
admits a complex structure.

T N = N N

Since K is a contravariant functor, the map h : T Y T N, given by trivially


extending the map T Y |i(X) N, induces the map h : K(T N) K(T Y ).
In the next section we will prove the Thom Isomorphism with some generality,
which will imply that the map : K(T X) K(T N) is an isomorphism. We
now construct i! : K(T X) K(T Y ) by simply composing these two maps.
Letting Y be Rn and noting the obvious bundle isomorphism4 T X ' T X
completes the reduction. We will call the general map i! : K(T X) K(T Y )
the induced pushforward map.

4.2 K n (X) and the Thom Isomorphism

We would now like to make the connection between the topological K-theory
we are developing and de Rham cohomology. Quite obviously, to do this we
must first construct a cohomology theory from topological K-theory.

Definition 4.2.1. In topological K-theory for n 0 we define K n (X) =


K(S n X) where S n X is the n-fold reduced suspension of X.

Noticing that the double suspension of a point is S 2 leads to the famous


Bott periodicity theorem , the proof of which is far to lengthy to include here
but can be found in [2](2.2).

Theorem 4.2.2 (Bott Periodicity Theorem). K n (X) ' K n2 (X)

The Bott periodicity theorem means that K (X) is the Z/2Z graded ring
K 0 (X) K 1 (X). To make the connection between K-theory and de Rham
4 This is an immediate result of the existence of a Riemannian metric, which itself exists
for any paracompact manifold, compact manifolds thus being a special case.

46
cohomology, we need a map that will take bundles to forms in a way that
respects our ring operations. As it so happens, we have already seen that Chern
classes are not only capable of expressing the cohomology ring of X, but also
satisfy the homomorphism property for additive and multiplicative operations
on line bundles. Any vector bundle can be decomposed into a sum of line
bundles5 as one would expect considering vector spaces are entirely determined
by dimension. Letting E = Li be the line bundle decomposition of a rank n
complex vector bundle we can construct a map into cohomology in the following
way.

Definition 4.2.3. We define the Chern character of a line bundle L as the


exponential of the top Chern class ch(L) = ec1 (L) . Letting ti = c1 (Li ) we can
extend this map to a complex vector bundle E by defining the Chern character
as follows.
X tk1 + . . . tkn
ch(E) = ch(Li ) = n + (t1 + . . . + tn ) + . . . +
k!

We now extend ch(E) to K-theory, by defining it on the vector bundle


differences used in our K-theory construction, and get a ring homomorphism
K(X) H even as we know from our earlier discussion that Chern classes reside
in even dimensions. We can extend the Chern character to odd dimensions, ch :
K 1 (X) H odd , by composing it with 1 where : H odd (X) H even (SX)
is defined by : ( ) 7 ( ) where generates H 1 (S 1 ). While the Chern
character is a homomorphism it does not define an isomorphism since K (X)
is not necessarily torsion free. The good news is that we can make it into an
isomorphism by tensoring in any torsion free unital module, traditionally Q,
giving ch : K (X) Q H (X; Q). Using R is equally as good and gives an
isomorphism into de Rham cohomology. The trouble is that in order to prove
this one relies on spectral theory which we have omitted from this text. As
5 We omit proving this here since it requires defining and exploring the properties of flag
bundles, however, for a proof of this splitting principle see [2](2.7.1).

47
such, a proof can be found [4](2). Thus, while (4.2.3) is the definition of ch, one
often abuses the notation and means the isomorphism ch.

Via this isomorphism it is clear that as a cohomology theory, K-theory


satisfies all of the same properties. The primary one we will use is the following.

Corollary 4.2.4. K is homotopy invariant

With K-theory and de Rham cohomology tied together, we would now like
to show that the fundamental map given by multiplying in by the Thom class
defines an isomorphism, the Thom isomorphism. To do this we must first recall
the definition of relative cohomology groups.

Definition 4.2.5. Let (X, A) be a pair such that A X and let G be an abelian
group. The relative chain complex Cn (X, A) can be defined as the complex
making the sequence 0 Cn (A) Cn (X) Cn (X, A) 0 exact. The
relative cohomology group with coefficients in G is defined as the homology of
the complex dualized by Hom(, G), i.e. the long exact sequence of cohomology
groups induced by the following short exact sequence.

0 Hom(Cn (A), G)) Hom(Cn (X), G)) Hom(Cn (X, A), G)) 0

To prove the Thom isomorphism, that it is in fact an isomorphism, in the


most general context possible, we concern ourselves here not with any specific
cohomology theory, rather any cohomology theory satisfying the generalized
Eilenberg-Steenrod axioms. We recall the axioms as follows.

Definition 4.2.6 (Eilenberg-Steenrod Cohomology Axioms). Let (X, A)


be a pair of topological spaces such that A , X is subspace inclusion, and let X
refer to (X, ). We denote the category of all such pairs Top, . A cohomology
theory is a functor H : (Top, )op AbZ together with natural transformation
: H (A, ) H +1 (X, A) such that:

48
1. (Exactness) For A , X, the induced sequence


. . . H n (X, A) H n (X) H n (A) H n+1 (X, A) . . .

is a long exact sequence of abelian groups.

2. (Homotopy Invariance) For a weak homotopy equivalence f : X Y ,


H (f ) : H (Y ) H (X) is an isomorphism.

3. (Additivity) If (X, A) = ti (Xi , Ai ) then H n (X, A) = ti H n (Xi , Ai ).

4. (Excision) For S , A , X, the natural inclusion of a pair i : (X


S, A S) , (X, A) induces an isomorphism H n (i) : H n (X S, A S)
H n (X, A).

We now define an important subbundle inspired by relative structures on


subspaces.

Definition 4.2.7. Let : E B be a fiber bundle with fiber F . (E, E 0 ) is a


fiber bundle pair if E 0 is a subspace of E such that : E 0 B is a fiber bundle
with fiber F 0 a subspace of F .

We shall also make use of the mapping cylinder and thus recall its definition
as well.

Definition 4.2.8. Given a map f : X Y , the mapping cylinder Mf is the


space (X I) t Y quotient the relation (x, 1) f (x).

Theorem 4.2.9. Let (F, F 0 ) (E, E 0 ) B be a fiber bundle pair such that
H (F, F 0 ; R) is a free R-module and finitely generated in each dimension. If
there exist classes cj H (E, E 0 ; R) whose restrictions to the fiber form a basis
for H (F, F 0 ; R), then H (E, E 0 ; R) is a free H (B; R)-module with basis {cj }.

Proof. [12](4D.8) We construct the bundle E B from E by attaching the


mapping cylinder M of : E 0 B to E by identifying the subspaces E 0 E

49
and E 0 M . Via the excision axiom and the deformation retract of E B onto
E, we have H (E, M ; R) ' H (E B, M B; R) ' H (E, E 0 ; R). Furthermore,
since we trivially have that B is a deformation retract of M , it follows that
H (E, M ; R) ' H (E, B; R). Along with the knowledge that trivially B is a
deformation retract of E, this gives us the natural splitting of H (E; R) into
H (B; R)-modules H (E, B; R) H (B; R). Letting cj H (E; R) correspond
to cj H (E, E 0 ; R), it follows that {1, cj } restricts to a basis for the fiber
cohomology H (F ; R). By the Leray-Hirsch theorem, {cj } is a basis for the free
H (B; R)-module H (E, E 0 ; R).

This proof, while intimidating in appearance, is actually simply the rigorous


validation of the claim that when relative cohomology groups split naturally
their generators do as well, as we would expect given the Leray-Hirsch theorem
(and as we would hope considering bundles as generalizations of products).
The simplicity of this claim is what makes the Thom isomorphism so beautiful.
The Thom isomorphism follows as a special case of this claim when we restrict
ourselves to only the case of fiber bundles with disks for fibers.


Corollary 4.2.10 (Thom Isomorphism). Let (Dn , S n1 ) (E, E 0 ) B be
the disk bundle. Suppose there exists an element H n (E, E 0 ; R) such that
restricting to the fiber of (Dn , S n1 ) generates H n (Dn , S n1 ; R). The map
: H i (B; R) H i+n (E, E 0 ; R) defined as : b 7 (b) is an isomorphism
for all i 0. When it exists, the Thom class is denoted by , and the induced
Thom isomorphism is denoted .

The space E/E 0 is called the Thom space of E, denoted T h(E). Furthermore
we have an important existence claim.

Theorem 4.2.11. Every orientable vector bundle has a Thom class.

Proof. This follows immediately since the top cohomology of the Thom space

50
of a vector space is one dimensional, making it trivially finitely generated.

Proceeding from the calculation of the Euler class we can calculate exactly
what the Thom class will be. We start by finding a trivializing cover and
a partition of unity subordinate to it, then sum up the generator along the
partition of unity while respecting orientation. We can now generalize our
discussion of the Euler class and explicitly write down the Thom class. Letting
be such a partition of unity, be a choice of local polar coordinates, and
g be our transition functions, the Thom class can be expressed explicitly as
follows.6
  !
d 1
X
= d (r) + d (r) d log (g )
2 2

Applying this to our cohomology theories we have the following similarly


named theorems.

Theorem 4.2.12 (Thom Isomorphism for de Rham cohomology). Let


: V M be an orientable real vector bundle. Then,


Hcv (V ) ' H n (M )

Theorem 4.2.13 (Thom Isomorphism for K-Theory). Let V be a complex


vector bundle7 with compact base B. Since we naturally have that T h(V ) ' V ,
it follows that the map : K(B) K(V ), defined by exterior multiplication
with a Thom class, is an isomorphism.
6 Note that although this construction is in de Rham cohomology, by the de Rham theorem,
de Rham cohomology is isomorphic to singular cohomology.
7 Recall that for every complex vector bundle, dropping the complex structure gives a real

vector bundle, which is trivially orientable by choosing orientation consistent with the complex
structure.

51
4.3 Topological Index

Observe that for the inclusion map j : pt Rn , the induced pushforward map
j! : K(T pt) K(T Rn ) is an isomorphism. This follows immediately as a
consequence of the Thom isomorphism : K(pt) K(Cn ).

Definition 4.3.1. Let X be a compact manifold and n be large enough such


that i : X Rn is an embedding. Let j : pt Rn be the inclusion of the origin.
The topological index -ind(X) : K(T X) Z is defined as the composition of
i! and j!1 .
i j 1
-ind : K(T X) ! K(T Rn )
!
K(T pt) ' Z

As we shall see later, the following two properties fundamentally characterize


the topological index. The direction we shall take in proving the Atiyah-Singer
index theorem will be to show that the analytic index also satisfies these two
properties.

Theorem 4.3.2. -ind(pt) : Z Z is the identity map.

Proof. Since our space is pt, it follows that the induced pushforward i! is the map
j! . As we already established, j! is an isomorphism, thus it follows immediately
that -ind = j!1 j! is the identity.

Theorem 4.3.3. -ind commutes with i! .

i i j 1
Proof. i! -ind = K(T X) ! K(T Rn ) ! K(T Rn )
!
K(T pt)
i j 1
= K(T X) ! K(T Rn )
!
K(T pt)
i j 1 Id
= K(T X) ! K(T Rn )
!
K(T pt) K(T pt)
i j 1 j!
= K(T X) ! K(T Rn )
!
K(T pt) K(T Rn )

= -ind i!

52
Chapter 5

The Atiyah-Singer Index


Theorem

Thus far, we have told a topological story of smooth manifolds and various
constructions dealing with their vector bundles. However, the Atiyah-Singer
index theorem is not an answer to a topological question. It is part of the
answer to the larger analytic question of computing the number of solutions to
system of differential equations. In practice, it tends to be extremely difficult
to manipulate differential operators and integral equations in such a way that
the space of their solutions becomes clear., much less the solutions themselves
On the other hand, topology and its tools are usually so general that once one
has established some properties, computation becomes fairly simple. Of course
there are exceptions to this. For example, computing higher homotopy groups is
extremely challenging, but this is a decidedly different challenge than producing
tomes of integral equations when solving one set of differential equations. The
Atiyah-Singer index theorem is the theorem which translates so much analytic
complexity into topological simplicity. If it were simply a computational tool

53
though, it would not be modern mathematics greatest achievement. It does
more than serve as a tool; it is a bridge between the topological and analytic
worlds. Its content is intrinsically mathematically beautiful, regardless of its
applications. We endeavor in this section to present the analytic side of this
story and finally prove the greatest theorem of the 20th century.

Remark. We use multi-index notation in this chapter as is conventional practice


when working with differential operators.

5.1 Elliptic Operators

The theory of partial differential equations is divided into the study of three
classes of operators: elliptic, hyperbolic, and parabolic. Elliptic differential
equations, and elliptic differential operators, are arguably the top of the three.
This is for two reasons: firstly, elliptic operators generalize the Laplace operator
which is the king of operators in classical analysis, as we know. Secondly,
solutions to elliptic equations can never have discontinuous derivatives, whereas
the other two classes are degenerate cases. In fact, solutions to hyperbolic and
parabolic equations that are not degenerate in this sense are often solutions to
elliptic equations.

We begin our analytic story by generalizing differential operators from Rn


to smooth manifolds in the expected locally Euclidean way.

Definition 5.1.1. Let E and F be complex vector bundles with base B a compact
manifold. D : E F is a partial differential operator if there exists local
trivializations such that in local coordinates D can be expressed as
X ||
D= A (x)
x
||m

where A (x) is a rank(F ) rank(E)-matrix of smooth complex functions such


that there exists || = m making A 6= 0.

54
We are interested in the global behavior of differential operators, and as such
we would like to concern ourselves only with the differentials which determine
the behavior. These are going to be the highest order differentials, thus we
would like to work with only top degree differentials. To do this we will rewrite
the operator as a polynomial and remove all low order terms, which will let us
work with the top degree differentials.

Definition 5.1.2. The symbol of a differential operator D is the polynomial


||
constructed by replacing x with i .
X
p(x, ) = A (x) i
||m

The principal symbol, (D) is the highest order terms of p(x, ).


X
(D) = im A (x)
||=m

Now that we can isolate the highest order derivatives, we can classify the
behavior of an operator and decide when it is well-behaved enough to be elliptic.

Definition 5.1.3. A differential operator D is elliptic if its principal symbol is


invertible when 6= 0.

The Laplace operator, which elliptic operators seek to generalize, seen can
be seen to be elliptic as its symbol is simple paraboloid.
X 2f
f =
n
x2n
X
() = n2
n

We seek to generalize this further. Notice that the inversion of the Fourier
transform of a function f is the map which takes the differentials of a differential
operator to polynomial indeterminates.1
1 For a discussion of this representation of the Fourier transform see [26] (1).

55
Z

f (x) = f () eihx,i d
xj xj
Rn
Z
= ij f () eihx,i d
Rn

More generally
Z
Df (x) = p (x, ) f () eihx,i d
Rn

We make use of this integral formulation to generalize differential operators,


first to Rn then manifolds, by allowing p(x, ) to be any function with suitable
growth conditions permitting differentiation under the integral.

Definition 5.1.4. Let P : C0 (Rn ) C (Rn ) be linear. We define P to be a


 
pseudodifferential operator of order m if it can be expressed as p x, i/x
where p(x, ) is a smooth and satisfies the following growth condition.

m||
x p (x, ) c, (1 + ||)

Here c, are positive constants for each , . We also require that the following
limit exist.
p (x, )
lim
m
This limit defines the symbol of P , m (x, ).

Note that when P is a differential operator, as previously defined, we have the


confusing bit of nomenclature that the pseudodifferential symbol corresponds
to the differential principal symbol (P ) = m (x, ). We now generalize this
to manifolds by looking at a trivializing cover.

Definition 5.1.5. Let E and F be complex vector bundles with base B a compact
manifold. A linear operator P : 0 E F is a pseudodifferential operator
of degree m on B if given a cover {Ui } of B trivializing E, F , and T B,

56
then for all Ui restricting P to functions with compact support on Ui gives
a pseudodifferential operator Pi of degree m on Rdim B .

Definition 5.1.6. Let P be a pseudodifferential operator between complex vector


bundles E and F with compact base B. Let : T B B be the projection of the
cotangent bundle. Let (P ) be a smooth section of Hom( E, F ) canonically
identifying dx to . (P ) is the global symbol of P if there exists a cover of
B trivializing E and F such that restricting (P ) restricted to Ui is the symbol
of Pi .

The existence of the symbol of P can be seen as natural considering that


(P ) is a smooth section of the bundle Hom( E, F ) under the canonical
isomorphism, which is well defined as a vector bundle. Thus the stitching for
Pi and its symbol are well-defined. Another approach is to formally construct a
pseudodifferential operator P for every possible symbol in a way that respects
the natural algebra of symbols. For the details of this specific construction see
[28].

Definition 5.1.7. A pseudodifferential operator P on complex vector bundles


E and F is elliptic if there exists c > 0 such that (P ) is invertible for || > c.

We denote the space of pseudodifferential operators DOm and the space of


elliptic pseudodifferential operators EDOm of order m. We can now begin to
relate this to the topological theory we have developed. To make this connection
we will need the following insight.

Theorem 5.1.8. Let E and F be vector bundles over a manifold X and let
a : E F be an isomorphism. We define C(X) to be the homotopy classes of
all such a. C(X) is isomorphic to K(X).

Proof. Notice that two pairs (E1 , E2 ) and (F1 , F2 ) in C(X) are equivalent if
there exists bundles Q and P such that E1 Q ' F1 P and E2 Q ' F2 P .

57
This means that the map (E1 , E2 ) 7 [E2 ] [E1 ] is an isomorphism quite simply
because this map is how we construct K(X).

In fact, we can define a more general complex Cn (X) of bundle isomorphisms


which form a sequence of length n. It turns out that for any n greater than 0,
Cn (X) is isomorphic to K(X). Proving this rigorously turns out to be quite
tedious due to the precise nature of products the ring K(X) and how this
translates to products in sequences. To avoid delving into algebraic K-theory
rather than analysis we direct the reader to the proof [2](2.6.1) in lieu of
reproducing it here. Instead we proceed to draw the link between our analytic
and topological stories.

Theorem 5.1.9. For elliptic pseudodifferential operators on complex vector


bundles with compact base X, the symbol map is a surjection onto K(T X).

Proof. [19](3.3) Let X be a compact manifold and a K(T X). Let E and F
be the complex vector bundles over X such that a Hom( E, F ) where :
T X X is the projection of the cotangent bundle. Let C be a cover of X that
trivializes E, F , and T X. In local coordinates, let () be a smooth function
such that there exists an open set U with the following properties: U contains
the origin, there exists V U containing the origin such that is 0 on V and
1 on U c , and U is a compact proper subset of Tx X. Let p(x, ) = ()a(x, ).
Define the operator P = p(x, D). It follows immediately from the definition
of elliptic pseudodifferential operator that P is an elliptic pseudodifferential
operator. This shows that the map : EDO K(T X) is surjective.

5.2 Analytic Index

Let X be a compact manifold and E and F be complex vector bundles over


X. While not easy to show, it does follow that for an elliptic pseudodifferential

58
operator P EDOm (E, F ), there exists an operator Q EDOm (F, E)
such that P Q 1 and QP 1 have compact kernel. This follows essentially
from the symbol maps behavior (P ) = (P ) where P is the adjoint of P ,
and the fact that the symbol is invertible when P is elliptic.2 This compactness
result immediately implies that on a compact manifold X, a pseudodifferential
operator is Fredholm, meaning it has finite dimensional kernel and cokernel.
Thus the index inherited from P being Fredholm is well defined.

Definition 5.2.1. Let X be a compact manifold. The index of an elliptic


pseudodifferential operator P on complex vector bundles E and F with base
X is ind(P ) = dim ker(P ) dim Cok(P ).

With this in mind, we define the analytic index as follows.

Definition 5.2.2. Let X be a compact manifold. The analytic index a-ind :


K(T X) Z of an elliptic pseudodifferential operator P on complex vector
bundles E and F with base X is defined by the map x 7 ind(P ) where x = (P ).

Concerning whether the analytic index is well-defined as a map, we already


know that the symbol map is a surjection onto K(T X), however, we dont
know that the symbol map is pseudo-injective, i.e. two operators with the
same symbol have the same index. Proving this brings us to the reason the
Atiyah-Singer index theorem was ever posed as a question.

Theorem 5.2.3. ind is homotopy invariant.

We omit a proof of this as it requires substantial background knowledge of


functional analysis that is both undeveloped in this text and not in the spirit
this work, however a proof can be found in [2] (Appendix I). In his paper On
Elliptic Equations, Israel Gelfand noticed that the index for elliptic operators
2 This equivalence only holds up to a compact operator. For a proof and an in depth
discussion of this see [14] (19).

59
was homotopy invariant. This led him to pose the question of whether it could
be calculated by topological means. This motivation, together with the growing
importance of index theory in theoretical physics, spurred the discovery of the
index theorem. History aside, it follows immediately from this that the symbol
map is pseudo-injective.

Corollary 5.2.4. Let P and P 0 be elliptic pseudodifferential operators between


complex vector bundles with a compact base such that (P ) = (P 0 ), it follows
that ind(P ) = ind(P 0 ).

5.3 Proof of the theorem

As alluded to earlier, the direction we take to prove the Atiyah-Singer index


theorem is to show that the analytic index satisfies the same properties as the
topological index. We now define what it means to be an index function at
all, and then the necessary theorem that any index that does satisfy the same
properties as the topological index is an equivalent index.

Definition 5.3.1. A collection of homomorphisms IndexX : K(T X) Z


for all compact manifolds X is an index function if, for any diffeomorphism
f : X Y , for all y in K(T Y ) it follows that IndexX f y = IndexY y where
f is the induced map f : K(T Y ) K(T X).

Theorem 5.3.2. If Index is an index function and satisfies (4.3.2) and (4.3.3),
then Index = -ind.

Proof. [19](3.1) Let Index be an index function, X be a compact manifold,


i : X Rn be an embedding, and j : pt Rn be the inclusion of the origin.
By (4.3.2) and (4.3.3), the following diagram commutes.

60
Index
K(T X) Z

-ind
i!

j!1
K(T Rn ) K(T pt)

As an equation, this diagram reads,

Index(X) = Index(pt) j!1 i! = j!1 i! = -ind(X)

Theorem 5.3.3 (Atiyah-Singer Index Theorem).

-ind = a-ind

Proof. [3], [19](3.4) It is trivial that a-ind is natural in the functorial sense
with respect to diffeomorphisms, thus this proof follows naturally from the
previous theorem. That is, we aim to show that a-ind satisfies (4.3.2) and
(4.3.3). Verifying (4.3.2) is quite simple. If we let X be pt, then any bundles
E and F over X are complex vector spaces determined uniquely by dimension,
thus any P EDO(E, F ) is a linear transformation, meaning its symbol is of
order 0 and the symbol space is Z. More precisely (P ) = dim(E) dim(F ).
The index of P , ind(P ) = dim ker(P )dim Cok(P ) can be rewritten as ind(P ) =
dim(E)dim Im(P )(dim(F )dim Im(P )) = (P ). Thus, a-ind is the identity
map. To show that a-ind commutes with i! it suffices to show that it commutes
with the Thom isomorphism and the extension map for tubular neighborhoods
h . We begin with h .

Suppose h : U X is the inclusion of an open set U into X. Let : T X


X be the projection map for the cotangent bundle. Choose K(T U ) such

61
that is degree 0,3 and there exists a compact set C U such that is the
identity outside of C. We can trivially extend to the following map.

, on T U
0 =
id, on T X T U

We define h as the map 7 0 K(T X). Let P and P 0 be the operators


with symbol and 0 respectively. It follows naturally that h (P ) = (P 0 ).
Furthermore it follows that P and P 0 have identical support since they agree on
U . This follows for their adjoints as well, thus ind(P ) = ind(P 0 ). This shows
that h commutes with a-ind.

Substantially more effort is required to show commutativity with . Let


: E X be a complex vector bundle with compact base X, : K(X)
K(E) be the Thom isomorphism given by exterior multiplication with the Thom
class E . Recalling that E = (1), commutativity with a-ind is expressed as
n
a-ind( x E ) = a-ind(x) a-indR
O(n) (Cn ) = a-ind(x). To show this we show

that a more general multiplication formula holds. With the Thom class, we must
be careful about the twisting of our vector bundle. Because of this, we observe
that the multiplication formula is a map of type K(T X) KO(n) (T Rn )
K(T E). KO(n) (T Rn ) is the Grothendieck group of vector bundles over T Rn
with a smooth left action O(n) T Rn T Rn .4 Bundles equipped with such
an action are called principal bundles. We do not require their full theory, simply
their commutativity, which follows by definition. Note that elliptic operators
commute with this action. We can rewrite our multiplicative formula, denoting
respective domains by superscripts, as follows.

n
a-indE (a b) = a-indX (a) a-indR
O(n) (b) (5.1)
3 We can choose of any degree, since a direct implication of (5.2.4) is that index is
independent of order.
4 For a general group G and principal G-bundles this is called equivariant K-theory. We

are only concerned with the group O(n) since this is a structure for any vector bundle.

62
Letting : T X X be the projection map for the cotangent bundle, we
choose K(T X) of degree 1 as a representative for a. Letting A be the
elliptic pseudodifferential with symbol , we lift A to A on E by a partition
of unity subordinate to some cover of X which trivializes E. We apply the
same process to an element of KO(n) (T Rn ) to get an elliptic pseudodifferential
operator B on E. To check (5.1) we multiply externally, giving us the following
operator.
A B
D=

B A
To calculate the index of D we make use of the following operators: P0 =
A A + B B, P1 = AA + B B, Q0 = AA + B B , P1 = A A + B B . Note that
by the commutativity of B and its adjoint with A and its adjoint we have,

P 0 0 P 1 0
D D = , DD =
0 Q0 0 Q1

Thus, ind(D) = (dim ker(P0 ) dim ker(P1 )) + (dim ker(Q0 ) dim ker(Q1 )).
For the operator P0 , it follows from the definition of adjoint that hP0 u, ui =
hAu, Aui + hBu, Bui, which implies that ker(P0 ) = ker(A) ker(B). Letting
P be a bundle such that X = P/O(n), since B is a fiberwise extension of
B, it follows that its kernel is an extension of ker(B) as well. Specifically,
ker(B) = P O(n) ker(B), which we will denote kB . Thus, A induces an
operator C where (C) = id(kB ). It follows that (C) = a[kB ] K(T X)
where [kB ] is the class of kB in K(X).

While it may seem complicated, this is a quite natural decomposition of


ker(P0 ) when we consider it splitting over the fiber up to a twist from O(n).
We can repeat this process mutatis mutandis for P1 and A deriving C . This
permits us to rewrite the left summand of ind(D) as dim ker(C) dim ker(C )
which is simply ind(C) = a-indX (a[kB ]). We repeat this process and decompose
the cokernel giving us the right summand of ind(D) as the index of an operator

63
induced on the bundle lB = P O(n) Cok(B). Therefore ind(D) = a-indX (a
([kB ] [lB ])). Since we are working with a complex vector bundle it follows
n
that for our B-fiber operator [kB ] [lB ] = a-indR
O(n) (b), since the fiber is just a

vector space. Since a-ind is a homomorphism, this completes the verification of


n
(5.1). All that remains now is to verify that a-indR
O(n) (Cn ) = 1.

Let j : pt Rn be the inclusion of the origin, inducing j! : KO(n) (T pt)


n
KO(n) (T Rn ). We must show that a-indO(N ) (j! (1)) = a-indR 5
O(n) (n ) = 1. Since

O(n) has a representation given by G1 . . . Gn where Gi is O(1) or SO(2),


we only need to check the cases n = 1 and n = 2.

Let n = 2. We dont need to work with all of O(2). By (3.1.1) we


only need to consider SO(2). Since the one-point compactification of R2 is
isomorphic to S 2 , we look to utilize KSO(2) (T S 2 ), which we can associate
to the Thom class. Quite trivially we can identify the symbol of a differential
operator to a differential complex which is naturally elliptic since d2 = 0.
L n 2
Here the complex is the complexified de Rham complex, (T S ) C.
To calculate the index of associated to this complex, one takes the Euler
characteristic6 of the sequence. Utilizing the homotopy invariance of de Rham
cohomology, (S 2 ) = dim(H 0 (S 2 )) dim(H 1 (S 2 )) + dim(H 2 (S 2 )) = 2.7 Thus
n
a-indR
SO(2) () = 2. We now make use of Lemma (3.2) from [3] which states that

for S n , = j! (1) j! (1), where is multiplication by 1 on cotangent vectors.


2 2 2
It follows that a-indR R R
SO(2) () = 2 a-indSO(2) (n ), thus a-indSO(2) (n ) = 1.

Letting n = 1, similarly to above, were working with KO(1) (T S 1 ) and the


operator d in the complexified de Rham complex. This is much simpler since
the cohomology groups for S 1 are nontrivial only for dimensions 0 and 1. Being
sequential, the kernel and cokernel of d are easy to work with since the only
5 Notice here that j! is the Thom isomorphism, thus the first equivalence follows by
definition.
6 See Appendix (A). It is trivial that this index aligns with the Fredholm index.
7 See Appendix (B) for a computation of the cohomology groups of S 2 .

64
nontrivial homomorphism is d0 : 0 1 . Checking this one homomorphism
we see that the kernel of d is the set of constant functions, [1] in cohomology, and
the cokernel is generated by dx. The group action of O(1) is multiplication by
1 on the cokernel of d and trivial on the kernel, which contains only constant
functions. It follows immediately from this that a-indR
O(1) (n ) = 1.

This result, together with our multiplication formula gives us the following
set of equivalences which completes the proof.

a-indE (a) = a-indE ( a T E ) (5.2)


n
= a-indX (a) a-indR
O(n) (n ) (5.3)

= a-ind(a) (5.4)

65
Part III

Fractal Geometry

66
Chapter 6

Fractal Geometry

Fractals are, intuitively, objects that have some property that makes them
too complex a structure to be fully understood in their topological dimension.
Historically, the property we are interested in capturing, the one that defined
fractals, is the self-similar nature of a space. Self-similarity is not as simple as
a space looking like itself when you zoom in on a small neighborhood, rather a
space is self-similar if you can continue this process indefinitely. That is, you
can continue to zoom in on the space, and you continue to see copies of the
whole space. We are motivated to study these objects since not only are they
intrinsically interesting, but they also prove valuable to other fields of math,
such as analytic number theory.

6.1 Fractal dimension

To define fractals, we first need to rework the notion of topological dimension.


Let us first recall the topological dimension.

Definition 6.1.1. Let X be a topological space and C be a cover of X. C has

67
order n+1 if there is a point x in X such that x is in n+1 elements of the cover
and no point is in more than n + 1 elements of the cover. X has topological
dimension n if n is the minimum value such that every cover has a refinement
of order n + 1 or below.

As an example of this, consider S 1 . As a compact manifold we can always


find a refinement of a cover of S 1 with a cover consisting of neighborhoods
homeomorphic to the closed interval. Two small lemmas which we shall not
give proof of, but which are intuitively obvious,1 are that the closed interval has
topological dimension 1 and any space written as the union of closed subspace
has dimension equal to the maximum of the collection of subspace dimensions.
Thus, S 1 has topological dimension 1. We will not be generalizing this to
include permit any purely topological space to have non-integer dimension. We
are interested in how scaling affects a space, since this is key to self-similarity.
Thus, we will only be working with metric spaces, so we can make sense of scale.
We can now begin to define the fractal dimension of a space as a non-integer
dimension which captures self-similar structure.

The idea is that for well-behaved spaces, measuring distance with a straight
line is invariant under scaling. For example, measuring the interval in R with
a ruler of length 1 or two rulers of length 1/2 will both return 1. The fractal
dimension of a space X given a ruler of length is log N where N is the number
of -rulers it takes to cover X. We can navely capture the self-similar nature of
a space by allowing our ruler to vary in scale. This bring us to the Minkowski
dimension.

Definition 6.1.2. Let A be a subset of the metric space (Rn , d) and define
V () = voln {x A|d(x, A) < } where voln is the n-dimensional Lebesgue
measure and A is the boundary of A. The Minkowski dimension of A is defined
1 Short proofs can be found in [23](50.1).

68
as follows.
D(A) = inf{ 0|V () = O(n ) as 0+ }

V () is referred to as the inner tubular -neighborhood of A. The inner


tubular neighborhood formalizes our notion of measuring with rulers and the
Minkowski dimension capture self-similarity by scaling these rulers towards 0.
We are now well equipped to define precisely what we mean by a fractal.

Definition 6.1.3. A subset A of the metric space (Rn , d) is a fractal if D(A)


is non-integer.

There are several alternative characterizations of fractal dimension, the most


well-known of which is probably the Hausdorff dimension. Our motivation for
using the Minkowski dimension comes from the inverse spectral problem. Quite
simply put, Can one hear the shape of a drum? More precisely, if one had
a comprehensive list of the fundamental tone and overtones produced by a
perturbed surface, could one reconstruct the boundary of that surface. From
basic classical mechanics, we know that fundamental frequencies and overtones
are proportional, by a volume constant, to the objects spectrum. In 1912,
Hermann Weyl showed something incredible [27]. Let be a nonempty bounded
open subset of Rn with boundary . Letting be the Laplacian and u be a
distribution,2 we have the Dirichlet problem,

u = u x 0
(6.1)
u=0 x

Weyl showed that for sufficiently well-behaved , the eigenvalues, ordered by


multiplicity, are asymptotically dependent only on the dimension . Specifically,
letting Bn be the Lebesgue volume of the unit ball in Rn and Cn be the constant
(2 2 )(Bn )2/n , Weyl showed that i Cn (i/voln ())2/n as i . This lets
2 By distribution we mean in the Sobolev sense, that is u is in the completion of C0 ()
with respect to the Sobolev norm.

69
us deduce that the behavior of the eigenvalue counting function, defined as
N () = #{i|0 < i }, is asymptotically N () (2)n Bn voln ()n/2 as
. However, Weyls classification of well-behaved boundary did not include
all smooth manifolds, much less fractals. In 1979, Michael Berry conjectured [5]
that for a fractal, the eigenvalue counting function had asymptotic behavior
N () = (2)n Bn voln ()n/2 c(n,H) HH ()H/2 + o(H/2 ) as , where
H is the Hausdorff dimension of and HH ( ) is a function related to Hausdorff
dimension called the Hausdorff content. The problem with Berrys conjecture
is that the Hausdorff dimension is not the fractal dimension which satisfies
this equation. Brossard and Cremona[7] found a counterexample for which the
asymptotic formula failed, and showed that the Minkowski dimension is the
suitable notion of dimension. They did not resolve the Weyl-Berry conjecture,
as the asymptotic formula for fractal is named, rather they showed that the
issue with the Hausdorff dimension is that it is not invariant under perturbation,
i.e. the drum actually resonating, whereas the Minkowski dimension is. For
this reason, with the goal of understanding the spectra of spaces with fractal
boundary, we use the Minkowski dimension.

6.2 The Inverse Spectral Problem

We are interested in fractals that have Minkowski dimension in the open interval,
(0, 1). These are called fractal strings and are all bounded open subsets of R.
All bounded open subsets of R are collections of countably many disjoint open
intervals. For a fractal string we shall denote the set as . This should not
lead us to think they are simple objects. For example, the complement of the
ternary Cantor set in [0, 1] is a fractal string that is clearly not a simple object.
Since fractal strings are uniquely characterized by the lengths and multiplicities
of the countably many open intervals comprising them, thus we shall define L

70
to be the more structured ordered set of the interval lengths, `1 `2 . . . > 0.

It is not a coincidence that the open interval also defines the critical strip
for the Riemann zeta function. The inverse spectral problem provides context
for rephrasing the Riemann hypothesis. Let us now formulate this problem
precisely. To do this we will need the following definition.

Definition 6.2.1. A fractal string, L, is Minkowski measurable if its Minkowski


content exists in (0, ), where Minkowski content is defined as follows.

M(D; L) = lim+ V () (1D)


0

The Minkowski content of a fractal string L is intimately tied to what we can


think of as geometric oscillations, or the geometric spectrum of L. The inverse
spectral problem asks whether the geometric spectrum behaves the same way as
the classical spectrum. By classical spectrum we mean exactly the normalized

eigenvalues of , /, on L as a bounded open subset of R in the distributional
sense as in (6.1). Since we are dealing with copies of the open interval, the
frequencies in the spectrum of L are the integral multiples of the reciprocal
lengths, n `1
j , where n is in N. The spectrum of a fractal string lets us define

an important function.

Definition 6.2.2. Let L be a fractal string with frequencies f . The spectral


zeta function of L is defined as,
X
(s) = f s
f

Along this line of analysis, we can construct a zeta function from the interval
lengths themselves with the intent of describing a spectrum derived from the
geometry instead of the harmonics.

Definition 6.2.3. Let L = {`j } be a fractal string, where w` is the multiplicity

71
of the length `.3 The geometric zeta function of L is defined as follows.

X X
L (s) = `sj = w` `s
j=1 `

This gives us the natural relationship = L , where the final zeta function
is the Riemann zeta function. In order to figure out the relationship of these
functions over C, we need to find the poles of their meromorphic extension. A
small issue is that L may not have an extension to all of C; for this reason
Lapidus and Frankenhuijsen introduced the notion of a window.

Definition 6.2.4. Let L be a fractal string, and let s(t) : R [, D(L)] be


continuous. We define the screen S to be the contour s(t) + i t, and define the
window W to be the set {z C|Re(z) s(Im(z))}.

If L has a meromorphic extension to a neighborhood of a window W with


no pole on S, we call the set of poles of L the visible complex dimensions
of L and denote this set D. If there is a meromorphic extension to all of C then
D is the set of poles of this extension and is called the complex dimensions of
L. In theorem (8.15) of [18], the authors establish a criterion for Minkowski
measurability which equates the properties of the complex dimensions, and
thus L , to Minkowski measurability. We would like to know if the complex
dimensions can be put in relation to the spectral zeta functions, the spectrum
of L. So, the inverse spectral problem is formulated as: If the spectrum of a
fractal string, L, has no d D periodic terms, does it follow that L is Minkowski
measurable?

3 We can explicitly express w` = #{j|`j = l}.

72
Chapter 7

Fractal Cohomology

Motivated by the inverse spectral problem, we seek to understand how a fractals


geometry relates to its spectrum. Rather than the approach taken by Lapidus et
al., we seek to understand this not by the analytic number theoretic properties of
the complex plane, but rather the topological information. Such an inquiry has
proven fruitful in other settings, such as Pierre Delignes use of etale cohomology
to prove several extremely important number theory theorems known as the Weil
conjectures.

7.1 Spectral Class

In this section we will present some ideas for capturing the length of an interval
` = (a, b) in cohomology. In order to capture length in cohomology, we can
consider a subset of (`) which has some property tied to the length of `. One
possible approach is to simply let the metric on ` (R, dR ) be a generator
for this collection of functions. To do this in one variable, let us fix the

73
left boundary point a,1 and consider the collection of functions generated by
dR (a, x). This approach, while intuitive, is fundamentally flawed since, because
constant functions are excluded, it generates a trivial subring of de Rham
cohomology, i.e. the complexes dR (a, x) R under the differential operator
d. In fact, no collection of non-compactly supported functions that exclude
constant functions will generate something nontrivial. Considering that we
are trying to capture length we might as well consider any ring generated by
constant functions trivial. With this in mind we propose a different approach.

It is clear we must consider functions that are compactly supported. We can


in fact modify our previous collection of functions to be compactly supported.
Rather than fixing one point, we can fix the entire boundary, in the interval case
we have a leftmost point a and a rightmost point b, and consider the function
d(x) = min{dR (a, x), dR (b, x)}. This function is not necessarily smooth, but
we can model its key properties using smooth functions. We first capture the
symmetry of d about the midpoint of the interval. Secondly, we would like some
property of the our model functions value at any point, x, to be proportional
to d(x).

Definition 7.1.1. Let I be an open interval of R. We define C


Sym (I) to be the

collection of smooth symmetric functions with compact support on I.

Note that for a suitable translation of I, C


Sym (I) is simply the collection of

even functions, thus it is closed under all of the expected algebraic operations.
This means that (C
Sym (I) R , d) defines a cohomology theory.

Definition 7.1.2. Let X be a smooth manifold2 , and let A be a submanifold of



rank n.The symmetric cohomology of A, HSym (A), is defined as the cohomology
induced by the differential complex (C n
Sym (I ) R , d).
1 Note that ` is an open interval, thus a depends on the closure of ` relative to dR .
2 Since every manifold is metrizable, we do not need to specify a metric to generalize our
notions of fractal dimension.

74

For the interval, symmetric cohomology HSym (I) is isomorphic to compactly
supported de Rham cohomology. The fact that the two theories are isomorphic
is mostly trivial since, by the Poincare lemma for compact cohomology, only the
top dimension, that is H01 , will be nontrivial. Not only will it be nontrivial, but
by the algebra on forms it will be one dimensional.3 Thus, since C
Sym C0 , as
1
long as the induced cohomology HSym is nontrivial the two cohomology groups
are isomorphic. Finally, any even bump function will serve as a nontrivial
1
element of HSym .

The isomorphism is not necessarily natural since the space of symmetric


functions is a proper subspace of compactly supported functions, thus the
representatives for preferred generators may be off by an exact form. More
importantly, given a fiber bundle E over I, the Thom class will necessarily have
a symmetric analog in HSym , via a potentially unnatural isomorphism.

Definition 7.1.3. Let X be a smooth manifold, A be a rank n submanifold


n
of X, : E A be a fiber bundle. Suppose : HSym (A) H0n (A) is an
isomorphism. The spectral class of E is defined as = 1 (), where is the
Thom class of E.

What is interesting about the spectral class is that it induces the Thom

isomorphism via . This means that, while neither HSym nor H0 is not homotopy
invariant, the spectral class is, making it well behaved in cohomology. This may
seem trivial due to the fact that, in cohomology, being off by an exact form
amounts to equivalence. However, the spectral class offers a natural choice of
representative which could be fundamentally connected to the Lebesgue measure
of the space. I conjecture that there is a restriction of C
Sym such that the natural

choice of representative for the spectral class has the property that for all a in
A, the curvature of at a is proportional to d up to a polynomial. Were this
3 In general, for a smooth orientable manifold X, the top cohomology group of X is one
dimensional when X is connected.

75
conjecture true, we could topologically investigate the properties of L and in
a way similar to the Atiyah-Singer index theorem, but with a focus on analytic
number theory.

7.2 Concluding Remarks

The greater goal of this research is to fill in as many pieces of the following
diagram as possible.

Atiyah-Singer
H , K EDO

Inverse Spectral
Problem
, L

Naturally, the most important link is the one between L and , as this
would be equivalent to showing the Riemann hypothesis. Perhaps the most
interesting link though is the one investigated in the previous section, L and
cohomology theories. Fractals are such a new subject, and their topology is so
unexplored, that there is so much we dont know.

Many of the pieces that could be filled in are not even shown, as lines would
not adequately describe their relationships. The numerous analogies between
the two contexts is why the Atiyah-Singer index theorem was chosen for study.
The Atiyah-Singer index theorem tends to have many ties to any problem,
since it is such a deep theorem. However, for this problem in particular, the
abundance of potential is encouraging.

76
Appendices

77
Appendix A

Homological Algebra

A.1 Five Lemma

Theorem A.1.1. Consider the following commutative diagram of modules and


homomorphisms with exact rows,

f2 f1 f0 f1
A2 A1 A0 A1 A2

h2 h1 h0 h1 h2

B2 g2
B1 g1
B0 g0
B1 g1
B2

1. If Cok(h2 ) = ker(h1 ) = ker(h1 ) = 0 then ker(h0 ) = 0

2. If Cok(h1 ) = Cok(h1 ) = ker(h2 ) = 0 then Cok(h0 ) = 0

Proof. Assume Cok(h2 ) = ker(h1 ) = ker(h1 ) = 0. If a ker(h0 ) then a


ker(f0 ) since (g0 h0 )(a) = 0 implies (h1 f0 )(a) = 0 and ker(h1 ) = 0. By
exactness a0 A1 such that f1 (a0 ) = a. Furthermore, by exactness, b B2

78
such that h1 (a0 ) = g2 (b) since (g1 h1 )(a0 ) = (h0 f1 )(a0 ) = h0 (a) = 0. It
follows from the assumption Cok(h2 ) = 0 that a00 A2 such that h2 (a00 ) = b.
By commutativity of the diagram it follows that (h1 f2 )(a00 ) = (g2 h2 )(a00 ) =
g2 (b) = h1 (a0 ), which can be done since we assumed ker(h1 ) = 0. By exactness
we know (f1 f2 )(a00 ) = 0. From this we can conclude (f1 f2 )(a00 ) = f1 (a0 ) =
a = 0, which completes the proof of (1).

Now we assume Cok(h1 ) = Cok(h1 ) = ker(h2 ) = 0. If b B0 then


a A1 such that h1 (a) = g0 (b), since by assumption Cok(h1 ) = 0. By
exactness we have that (g1 g0 )(b) = 0 thus by the assumption ker(h2 ) = 0
and commutativity it follows that a ker(f1 ). Further by exactness it follows
that a0 A0 such that f0 (a0 ) = a. Thus by commutativity (g0 h0 )(a0 ) = g0 (b).
By the homomorphism property it follows that bh0 (a0 ) ker(g0 ). By exactness
we have that b0 B1 such that g1 (b0 ) = b h0 (a0 ), which, by assumption that
Cok(h1 ) = 0, implies that a00 A1 such that (g1 h1 )(a00 ) = b h0 (a0 ).
Finally, by commutativity we have that b h0 (a0 ) = (h0 f1 )(a00 ), which, by
the homomorphism property, implies that b = h0 (a0 + f1 (a00 )).

A.2 Euler Characteristic

Given a finite sequence,

f0 fk2 fk1 fk fk+1


S = 0 V0 . . . Vk1 Vk Vk+1 . . . Vn 0

we define the Euler Characteristic of the sequence as the alternating sum of the
dimensions.1
n
X k
(S) = (1) Dim (Vk ) (A.1)
k=0

Theorem A.2.1. If S is exact then (S) = 0


1 By starting the sum at k=0 we mean that the sum begin with V0 , not to be confused with
the trivial space 0 which begins S.

79
Proof. Naturally, we can write any element of the sequence in terms of the maps
incident to it.

Vk ' Im (fk1 ) Cok (fk1 ) ' ker (fk ) Coim (fk ) (A.2)

By exactness we can rewrite any element of S as Vk ' Im (fk1 )Coim (fk ).


Furthermore, as an immediate consequence of the rank-nullity theorem which
states Im (f ) ' Coim (f ), we can swap the operators on our maps resulting in
Vk ' Coim (fk1 ) Im (fk ). Since we can write each element of S as the sum
of parts of the element before it and after it, summing over the even elements
is isomorphic to summing over the odd elements.

M M
Vk ' Vk (A.3)
k even k odd

The dimension operator is additive over direct sums which completes the proof.
X X
Dim (Vk ) = Dim (Vk ) (A.4)
k even k odd
X k
(1) Dim (Vk ) = 0 (A.5)
k=0

80
Appendix B

Computing Cohomology

B.1 Mayer-Vietoris Sequence

One of the most important tools for computing and working with cohomology
is the Mayer-Vietoris sequence.

Definition B.1.1. Let M be a manifold. Given {U, V }, an open cover of M ,


we can construct the following exact sequence.

0 (X) (U ) (V ) (U V ) 0
(B.1)
(, ) 7

This short exact sequence naturally induces the following long exact sequence in
cohomology called the Mayer-Vietoris sequence.
i
H k+1 (M )
d

i j
H k (M ) H k (U ) H k (V ) H k (U V )
d
j
H k1 (U V )

81
For a proof of the exactness of (B.1), see [8](2.3). The exactness of the
Mayer-Vietoris sequence allows us to apply our previous theorem (A.2.1). What
this means is that computing the cohomology of a manifold covered by two sets
of known cohomology amounts to using exact sequence tricks and basic algebra
to fill in a table. To demonstrate this, let us compute the cohomology of S 1 and
S2.

To cover S 1 we let U and V be disks centered at the north and south pole of
S 1 respectively such that they extend past the equator of S 1 . Thus, U and V
are diffeomorphic to R and their intersection is diffeomorphic to R S 0 . Since
cohomology is homotopy invariant, the cohomology of U V is isomorphic to
the cohomology of S 0 . Thus, our Mayer-Vietoris sequence is as follows.

0 H 0 S 1 H 0 (pt pt) H 0 (R R) H 1 S 1 0
 

Note that every Mayer-Vietoris sequence terminates at the dimension of the


manifold since wedging in any form to a top degree form will always induce 0.
We can fill in this sequence for the spaces of which we know their cohomology
groups as follows.

0 H 0 S1 R R R R H 1 S1 0
 

By exactness we know that H 1 (S 1 ) is isomorphic to the cokernel of the map


j : H 0 (U t V ) H 0 (U V ), thus H 1 (S 1 ) is R. Also by exactness, we know
that H 0 (S 1 ) is isomorphic to the kernel of the map j : H 0 (U tV ) H 0 (U V )
which is simply the diagonal of R R thus H 0 (S 1 ) is R. We now move on to
S2.

To cover S 2 , we let U and V be disks centered at the north and south pole
of S 2 respectively such that they extend past the equator of S 2 . Thus, U and V
are diffeomorphic to R2 and their intersection is diffeomorphic to R S 1 . Since
cohomology is homotopy invariant, the cohomology of U V is isomorphic to

82
the cohomology of S 1 . We have shown above that H 0 (S 1 ) = H 1 (S 1 ) = R, thus
we only need to fill in the first column of the following table.

S2 U tV U V
H2 ? 0 0
H1 ? 0 R
H0 ? RR R

By exactness we know that H 2 (S 2 ) is isomorphic to the cokernel of the map


j : H 1 (U t V ) H 1 (U V ), i.e. j : 0 R, thus H 2 (S 2 ) is R. Also
by exactness, we know that H 0 (S 2 ) is isomorphic to the kernel of the map
j : H 0 (U t V ) H 0 (U V ) which is simply the diagonal of R R thus
H 0 (S 2 ) is R. We can now apply theorem (A.2.1) converting the long exact
sequence to the equation 1 2 + 1 x + 0 1 + 1 = 0. Amazingly calculating
H 1 (S 2 ) has simplified to elementary algebra. Solving for x yields H 1 (S 2 ) = 0.

It is not coincidental that the computation of the cohomology of S 1 and


S 2 look almost identical. Using the Mayer-Vietoris sequence, one can show
inductively that H 0 (S n ) ' H n (S n ) ' R, and H i (S n ) ' 0 for all i not equal to
0 or n. To do this we cover S n as we covered S 1 and S 2 , with two hemispheres
whose intersection is homotopy equivalent to S n1 . One then proceeds by doing
induction on n over the Mayer-Vietoris sequence for S n .

83
List of Symbols

#X The cardinality of the set X.

X The Alexandroff one-point compactification/Alexandroff extension


of X.

Coim() For f : X Y , Coim(f ) = X/ ker(f )

Cok() For f : X Y , Cok(f ) = Y / Im(f )

Hom(X, Y ) The class (set) of morphisms from X to Y

hom() The class (set) of morphisms in a category (small category)

Mank The category of k-differentiable manifolds

Top The category of topological spaces

K(M ) The set (class) of all compact sets in M

C (M ) The set of all smooth functions on M

C0 (M ) The set of smooth functions on M with compact support

pt One point space

X The algebraic dual space of X

X0 The interior of a set X.

84
References

[1] Aguilar, M. A., Gitler, S., & Prieto, C. (2002). Algebraic Topology from a
Homotopical Viewpoint. New York: Springer.

[2] Atiyah, M. F. (1967). K-Theory. New York: W.A. Benjamin.

[3] Atiyah, M. F., & Singer, I. M. (1968). The Index of Elliptic Operators: I.
The Annals of Mathematics, 87(3), 484530.

[4] Atiyah, M. F., & Singer, I. M. (1963). The Index of Elliptic Operators on
Compact Manifolds. Bulletin of the American Mathematical Society, 69,
422433.

[5] Berry, M.V. (1980). Some Geometric Aspects of Wave Motion: Wavefront
Dislocations, Diffraction Catastrophes, Diffractals. Proceedings of Symposia
in Pure Mathematics, 36, 13-28.

[6] Bredon, G. E. (1993). Topology and Geometry. New York: Springer-Verlag.

[7] Brossard, J. M., & Cremona, R. A. (1986). Can one hear the dimension of
a fractal? Communications in Mathematical Physics, 104, 103-122.

[8] Bott, R., & Tu, L. W. (1982). Differential Forms in Algebraic Topology. New
York: Springer-Verlag.

85
[9] Cartan, H., & Eilenberg, S. (1956). Homological Algebra. Princeton:
Princeton University Press.

[10] Dummit, D. S., & Foote, R. M. (2004). Abstract Algebra (3rd ed.). John
Wiley & Sons.

[11] Fuchs, D. B., & Viro, O. Y. (2004). Topology II: Homotopy and Homology.
Classical Manifolds. Berlin: Springer-Verlag.

[12] Hatcher, A. (2002). Algebraic Topology. Cambridge: Cambridge University


Press.

[13] Hirsch, M. W. (1976). Differential Topology. New York: Springer-Verlag.

[14] Hormander, L. V. (1985). The Analysis of Linear Partial Differential


Operators Vol. 3: Pseudo-Diffferential Operators. Berlin: Springer-Verlag.

[15] Husemoller, D. H. (1966). Fibre Bundles. New York: McGraw-Hill.

[16] Lapidus, M. L. (1991). Fractal Drum, Inverse Spectral Problems for


Elliptic Operators and a Partial Resolution of the Weyl-Berry Conjecture.
Transactions of the American Mathematical Society, 325(2), 465-529.

[17] Lapidus, M. L. (1995). The Riemann Hypothesis and Inverse Spectral


Problems for Fractal Strings. Journal of the London Mathematical Society,
52(2), 15-34.

[18] Lapidus, M. L., & Frankenhuysen, M. V. (2006). Fractal Geometry,


Complex Dimensions and Zeta Functions: Geometry and Spectra of Fractal
Strings. Dordrecht: Springer.

[19] Landweber, G. D. (2005, April). K-Theory and Elliptic Operators. Cornell


University Library. http://arxiv.org/abs/math/0504555v1

[20] Lawson, H. B., & Michelsohn, M.-L. (1989). Spin Geometry. Princeton, NJ:
Princeton University Press.

86
[21] May, J. P. (1999). A Concise Course in Algebraic Topology. Chicago:
University of Chicago Press.

[22] Milnor, J. W., & Stasheff, J. D. (1974). Characteristic Classes. Princeton,


NJ: Princeton University Press.

[23] Munkres, J. R. (2000). Topology. Upper Saddle River, NJ: Prentice Hall,
Inc.

[24] Osborn, H. (1982). Vector Bundles. New York: Academic Press.

[25] Steenrod, N. E. (1951). The Topology of Fibre Bundles. Princeton:


Princeton University Press.

[26] Taylor, M. E. (1981). Pseudodifferential Operators. Princeton, NJ:


Princeton University Press.

[27] Weyl, H. K. (1912). Das asymptotische Verteilungsgesetz der Eigenwerte


Linearer Partieller Differentialgleichungen. Mathematische Annalen, 71,
441-479.

[28] Widom, H. (1980). A Complete Symbolic Calculus for Pseudodifferential


Operators. Bulletin des Sciences Mathematiques, 104, 1963.

87

Das könnte Ihnen auch gefallen