Está en la página 1de 59

Proyecto Fin de Master en Investigacion Matematica

Facultad de Ciencias Matematicas


Universidad Complutense de Madrid

Morse theory
Author:
Jos
e Alberto Iglesias Martnez

Supervisor:
Vicente Mu
noz Vel
azquez

2009-2010

ABSTRACT
In Morse theory, one studies the topology of a finite-dimensional smooth manifold
through the critical points of a sufficiently well-behaved real valued function on it. Here
we present an introduction to Morse homology from a modern standpoint, with techniques similar to the ones used in Floer homology for infinite-dimensional manifolds.
First, we start defining Morse functions, proving existence theorems for them, and
considering the trajectory spaces of the gradient flow of those functions. We also introduce the concept of a Morse-Smale pair, which is the one that is actually useful for
Morse homology.
The modern approach of Morse homology mentioned is based on giving said trajectory spaces a Banach manifold structure, and through analysis of Fredholm operators,
obtain compactifications of the corresponding moduli spaces. Some further work in this
line allows to define a manifold with corners structure on these moduli spaces, and coherent orientations on their boundaries. This way of tackling the technical difficulties
can be generalized for Floer homology theories, unlike the easier classical dynamical
approaches to Morse theory.
In a slight detour, we give a brief introduction to sheaf cohomology, prove the de
Rham theorem, and show how homology can be calculated from currents. While not
strictly needed for the rest, this approach does simplify later work and gives insight into
the different ways in which the same (co)homology arises, for nice spaces.
The compactification theorems are then used to define a chain complex formed from
the critical points of a Morse function, and proving that its homology is isomorphic
to the singular homology of the underlying manifold, the so-called Morse homology
theorem.
Finally, some applications are presented, and extensions of this theory are hinted at.

Key words: Morse functions, trajectory spaces, Morse homology, sheaf cohomology, de Rham theorem

MSC2000: primary 58E05; secondary 55N30, 47A13, 37C10, 58A12

RESUMEN
La teora de Morse consiste en el estudio de la topologa de una variedad finito-dimensional
a traves de los puntos crticos de una funcion real definida en dicha variedad, de comportamiento suficientemente bueno. Aqu se pretende presentar una introduccion a la misma
desde un punto de vista moderno, con tecnicas similares a las usadas en la homologa
Floer para variedades de dimensi
on infinita.
Primero, se da la definici
on de funciones de Morse, teoremas de existencia para las
mismas, y se consideran los espacios de trayectorias del flujo dado por el gradiente de
dichas funciones. Tambien se introduce el concepto de par Morse-Smale, el cual es el
principal para la homologa de Morse.
El punto de vista moderno mencionado se basa en dar una estructura de variedad
Banach a los espacios de trayectorias, y, mediante el analisis de operadores de Fredholm,
obtener compactificaciones de los espacios de moduli correspondientes. Con algo mas
de trabajo se le da a estos u
ltimos espacios una estructura de variedad con esquinas,
y se orientan los bordes correspondientes. Esta forma de trabajar se puede generalizar
a homologa Floer, no como los enfoques mas clasicos, aunque sean estos u
ltimos mas
sencillos.
Dando un cierto rodeo, tambien se presenta una introduccion a la cohomologa de
haces, se demuestra con ella el teorema de De Rham, y se muestra como es posible calcular homologa con corrientes. Si bien esto no es necesario para los objetivos principales,
simplifica lo que sigue, y ayuda a comprender como la misma homologa puede surgir de
diferentes formas, en espacios suficientemente buenos.
Despues de esto, se usan los teoremas de compactificacion para definir un complejo
de cadenas desde los puntos puntos crticos de una funcion de Morse, y probar que la
homologa de este complejo es isomorfa a la homologa singular de la variedad sobre la
que se trabaja.
Finalmente se presentan algunas aplicaciones, y se dan ideas sobre las generalizaciones existentes.

Palabras clave: funciones de Morse, espacios de orbitas, homologa de Morse,


cohomologa de haces, teorema de De Rham

El/la abajo firmante, matriculado/a en el Master en Investigacion Matematica de


la Facultad de Ciencias Matematicas, autoriza a la Universidad Complutense de
Madrid (UCM) a difundir y utilizar con fines academicos, no comerciales y mencionando expresamente a su autor el presente Trabajo Fin de Master: Morse theory,
realizado durante el curso academico 2009-2010 bajo la direccion de Vicente Mu
noz
Velazquez en el Departamento de Geometra y topologa, y a la Biblioteca de la
UCM a depositarlo en el Archivo Institucional E-Prints Complutense con el objeto
de incrementar la difusion, uso e impacto del trabajo en Internet y garantizar su
preservacion y acceso a largo plazo.

Fdo: Jose Alberto Iglesias Martnez

Supervisado y autorizado: Vicente Mu


noz Velazquez

Contents
1 Introduction
1.1 The classical approach . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Modern approaches. Advantages . . . . . . . . . . . . . . . . . . . . .

2
2
2

2 Basic definitions and results. Morse functions

3 The trajectory spaces


3.1 Trajectory spaces. Banach manifold structure .
3.2 Fredholm operators. Finite-dimensionality . . .
3.2.1 Non-trivial bundles . . . . . . . . . . . .
3.3 Transversality. Manifold structure . . . . . . . .
3.3.1 Genericity of the Morse-Smale condition
3.4 Compactification . . . . . . . . . . . . . . . . .
3.4.1 The space of unparametrized trajectories
3.4.2 Compactification result . . . . . . . . . .
3.5 Gluing . . . . . . . . . . . . . . . . . . . . . . .
3.6 Orientation . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

8
8
11
19
19
22
22
22
25
29
29

.
.
.
.
.

31
31
32
34
42
43

Morse homology theorem


The Morse chain complex . . . . . . . . . . . . . . . . . . . . . . . .
The chain homotopy. Morse homology theorem . . . . . . . . . . . .
The Morse inequalities . . . . . . . . . . . . . . . . . . . . . . . . . .

46
46
47
51

4 Sheaves, cohomology and currents


4.1 Basics. Presheaves and sheaves .
4.2 Resolutions of sheaves . . . . . .
4.3 Sheaf cohomology . . . . . . . . .
4.4 de Rham theorem . . . . . . . . .
4.5 Currents and homology . . . . . .
5 The
5.1
5.2
5.3

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

6 Generalizations
53
6.1 Morse-Bott theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Novikov homology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
References

55

Introduction

In Morse theory, one studies the topology of a finite-dimensional smooth manifold


through the critical points of a sufficiently well-behaved real valued function on it.
Here we present an introduction to it from a modern standpoint, with techniques
similar to the ones used in Floer homology for infinite-dimensional manifolds. Somewhere in between, with the purpose of presenting different ways in which the usual
singular homology of a manifold arises, we take a detour in order to define sheaf
cohomology, prove the de Rham theorem and see how homology arises from sheaves
of currents. I do not claim originality for any of the results given.
It is quite easy to find motivation for an approach of this kind. Just looking at
level sets of a height function on a surface, its easy to notice that when passing
a critical level, the topology of what is left behind changes. For example, in a
terrain-like surface S (i.e. the graph of a function f ), one would consider the sets
S a = {x S | f (x) a}.
Now, topologically, going through a valley (minimum) is equivalent to adjoining a
disjoint disk to S a , passing a peak (maximum) is the same as gluing a disk to S a
along their boundaries, and for a saddle point one attaches a handle.

1.1

The classical approach

The classical approach to Morse theory, as illustrated in the excellent text [Mi],
consists in treating the matter quite directly, with rather elementary methods.
The first step is to restrict the investigations to the most adequate kind of functions, i.e., the ones whose critical points are all isolated. Its not hard to prove that
these always exist using Sards theorem.
Then, one tries to reduce the variety of cases to treat, realizing that for any
critical point, only its index will be relevant (the so-called Morse lemma). In more
precise terms, it is possible to find a chart in a neighborhood of a critical point, so
that the function near there is in a standard form only depending on the index of
said critical point.
With that, by exactly the same method used in the example above, it is quite
straightforward to prove that any compact manifold has the homotopy type of a
CW-complex. This is already an interesting result.
For more delicate investigations into the behavior of Morse functions, the next
step is to consider their gradient vector field and corresponding flow, in order to know
how their values increase or decrease. Treating this flow like a dynamical system,
and applying a number of results from dynamical systems theory, it is eventually
possible to prove the Morse homology theorem: homology of the manifold can be
calculated as homology of a chain complex formed from the critical points of a Morse
function, with boundaries arising from the flow lines connecting critical points.

1.2

Modern approaches. Advantages

But the above method is not the only one to get insight about the behavior of Morse
functions and prove the Morse homology theorem. A more modern approach than
2

the above, and the one we try to introduce in this paper, is to use analytical methods
to analyze the trajectories of the gradient flow.
The starting point is the same, Morse functions and their gradient flows, and
one defines the sets
M(p, q)
of flow lines connecting two critical points (it is straightforward to see that for a
compact manifolds a flow line must always start and end on critical points). These,
as before, arise as solutions to the differential equation
= f
But instead of a dynamical approach to this equation and its flow, like before,
the idea is to apply analytical methods to get properties of the solution spaces, to
eventually get enough information about these trajectory spaces so as to get the
same Morse homology theorem. That will be our purpose.
This approach could seem heavy-handed, since just defining appropiate spaces
for these trajectories means getting into Banach manifold territory, and quite some
analysis is needed before everything comes together. The reason for approaching
the theory in this way is mostly pedadogical: while harder and more elaborate, it
gives deeper insight about the trajectory spaces, and it more or less easily generalizes into other settings, like that of infinite-dimensional starting manifolds (Floer
homologies).

Basic definitions and results. Morse functions

Definition 2.1. Let M be a smooth manifold, and f : M R a smooth function.


(i) A point p M is a critical point for f if dp f = 0. A point which is not critical
is called a regular point. Images of critical and regular points through f are
called critical and regular values, respectively.
(ii) The Hessian of f at p, Hp (f ) : Tp M Tp M is defined by Hp (f )(v) = v (df ),
for an arbitrary connection . df vanishing at p guarantees that this doesnt
depend on the choice of connection, since any two connections differ by a
tensor.
(iii) A critical point is called nondegenerate, if the Hessian there has null kernel.
(iv) The index of a non-degenerate critical point is defined as the number of negative eigenvalues of the Hessian.
Remark 2.2. One can define the Hessian without resorting to connections. This
is done by identifying the Hessian with a tensor, Hp (f ) : Tp M Tp M R, and
defining:
f ))(p) = Vp (W
f)
Hp (f )(v, w) = (V (W
defined on an open neighborhood of p, such that Vp = v and
For vector fields V , W

Wp = w. It is easy to check that this doesnt depend on the extensions used, and in
fact is equivalent to the previous definition.
Remark 2.3. From this last characterization of the Hessian it is also easy to theck,
that if : U Rm is a chart, with U an open neighborhood of p, and (p) = 0,
The matrix of Hp (f ) with respect to the basis x 1 , . . . , xm has the following second
derivative expression:
 2

(f 1 )
(p)
Mp (f ) =
xi xj
ij
Definition 2.4 (Morse function). We say a real valued function on a smooth manifold is a Morse function, if all of its critical points are nondegenerate.
From these definitions we can get some easy consequences directly:
Lemma 2.5. Non-degenerate critical points are isolated. In particular, a Morse
function on a compact manifold has a finite number of critical points.
Proof. Consider a chart around p as in the above remark. Define the map g :
(U ) Rm given by:


(f 1 )
(f 1 )
(x), . . . ,
(x)
g(x) =
x1
xm
Then g(0) = 0, and d0 g = Mp (f ) is nonsingular. By the inverse function theorem,
g is a diffeomorphism of some neighborhood V of 0, in particular injective in such a
neighborhood. This means g(x) 6= 0 for 0 6= x V , and x is not a critical point for
f.
4

The starting point for classical Morse theory is the following lemma, for which
we dont give a proof since it wont be needed for our discussions (one can find a
proof in [Mi], for example). However, it does provide some geometrical intuition
about the behaviour of functions near a non-degenerate critical point.
Lemma 2.6 (Morse lemma). Let p M be a non-degenerate critical point of index
k of a smooth function f : M R. Then there exists a chart : U Rm , with U
an open neighborhood of p, and (p) = 0, such that, in local coordinates:
(f 1 )(x1 , . . . , xm ) = f (p) x21 . . . x2k + x2k+1 + . . . + x2m
We will want to study not only critical points but relations between them, seeing
how the values of f increase or decrease. In euclidean space, the obvious choice would
be to use the gradient vector field of f . On an arbitrary manifold, we need to get a
vector field out of df . This is archieved through a metric on the manifold.
Definition 2.7 (Gradient). Let M be a smooth manifold, g a metric on M and
f : M R a smooth function. The gradient of f with respect to g is defined
as the image of df through the canonical isomorphism g : T M T M , given by
g(v)(w) = g(v, w). We denote the gradient of f on (M, g) by f .
Our goal, very roughly speaking, will be to understand the behavior of points of
M in the dynamical system defined by the gradient.
Remark 2.8. Throughout this paper we will use the negative gradient f , as is
customary in the literature.
Proposition 2.9. Let (M, g) be a Riemannian manifold, f : M R a smooth
function on f , and t : M M the local 1-parameter group of diffeomorphisms
generated by f . For an arbitrary x M , denote by x : (a, b) M the integral
curve given by x (t) = t (x). Then limt+ x (t) and limt x (t) both exist and
are critical points for f .
Proof. Let x M and x (t) the corresponding flow line of the negative gradient.
Since M is compact, x (t) is defined for all r R, and the image of f x is bounded
in R. Then, since f decreases along flow lines (very easy to see), we must have:
d
(f (x (t))) = 0
t dt
lim

Let tn be a sequence of real numbers such that tn . Then {x (tn )} has an


accumulation point q since M is compact. Now q must be a critical point of f by
the preceding limit. Pick a neighborhood U of q in which q is the only critical point.
Now assume limt x (t) 6= q. Then there is another sequence sn and another
neighboorhood V U of q such that sn , but x (sn ) U V . Then, as
above, x (sn ) must have an accumulation point, which must be a critical point of
f . But this contradicts that q is the only critical point in U .
Definition 2.10 (Stable and unstable manifolds). Let p M be a non-degenerate
critical point of f.
5

(1) The unstable manifold of p is defined to be:




u
W (p) = x M | lim x (t) = p
t

(2) The stable manifold of p is defined to be:




s
W (p) = x M | lim x (t) = p
t+

That these sets are actually manifolds will be proved in the next section. We
will also call them the descending manifold of p and the ascending manifold of p,
for W u (p) and W s (p) respectively.
We have defined Morse functions, but so far we dont even have an existence result
for them. We obtain such a result now. The proof is based on the following wellknown theorem, for which a proof can be found in [GuiPo]
Theorem 2.11 (Sards theorem). Let f : M N be a smooth map between manifolds M and N , and C the set of critical points of f . Then f (C) has measure zero
in N.
Lets first prove our theorem in an euclidean setting (i.e., locally):
Lemma 2.12. Let f : U R be a smooth function defined on an open set U o f
m
Rm . Then
Pm for almost all a = (a1 , . . . , am ) R , the function defined by fa (x) =
f (x) i=1 ai xi is a Morse function.
Proof. Consider the function g given by:


f
f
(x), . . . ,
(x)
g(x) =
x1
xm
And let a be a regular value of g. By Sards theorem, this is true for almost all a.
Now let fa be as in the statement, and let p Rm be a critical point of fa . Then:
dp fa = g(p) a = 0
Which means that g(p) = a, a regular value. Then dp g is surjective, and hence
invertible. But the Hessian Hp (fa ) is precisely dp g. So p is nondegenerate.
For our result, we need another of the basic theorems in differential topology. A
proof can be found in [Br].
Theorem 2.13 (Whitney embedding theorem). If M is a smooth manifold of dimension n, there exists a smooth embedding g : M R2n+1
So we can assume our manifold is embedded in euclidean space, which enables
us to use the lemma.
Theorem 2.14 (Existence of Morse functions). Let M be a smooth manifold emr
r
bedded in R
P.r For almost all a = (a1 , . . . , ar ) R , the function f : M R given
by f (x) = j=1 aj xj is a Morse function.
6

Proof. Let x = (x1 , . . . , xr ) M Rr . Then, there is a neighborhood U of x in


which a subset of the coordinates, say xj1 , . . . xjm form local coordinates for M . To
see this, one just needs to note that since Tx M , Tx Rr is injective, its dual Tx Rr
Tx M is surjective. Hence, for some neighborhood U of x, Tx M is generated by a
linearly independent set dxj1 , . . . , dxjm . Then xj1 , . . . , xjm are linearly independent
and hence form a coordinate system on U.
Now cover M by open sets Uj of this kind, and let (x1 , . . . , xm ) be a coordinate
system on Uj (assuming, without loss of generality,
jk = k). Choose (am+1 , . . . , ar )
Pr
rm
R
, and consider the function f (x) = j=m+1 aj xj . By the lemma, the function:
fa (x) = f (x)

r
X

aj x j

j=1

is a Morse function
for almost all a = (a1 , . . . , am ), fa . Hence, for almost all
Pr
r
(a1 , . . . , ar ),
a
x
j=1 j j is Morse on Uj . Let Aj be the points a R for which
this last
S function is not Morse on Uj . Each of these has measure zero, and hence
A = j Aj also has measure zero.
Remark 2.15. In fact, a number of stronger statements than the ones above can
be proved, such that the set of Morse functions is generic among smooth functions
on M (in a Baire category sense), and that there are Morse functions arbitrarily
uniformly close to a given smooth function.
Actually, one needs to require a bit more from our functions for them to be
useful in homology calculations, namely, transversality of intersections of stable and
unstable mainfolds.
Definition 2.16 (Morse-Smale function). Given a Riemannian manifold (M, g), a
smooth function f : M R is a Morse-Smale function if for all p, q Crit(f ),
W u (q) t W s (p), where Crit(f ) denotes the set of critical points of f .
Remark 2.17. Note that this transversality condition depends on the metric g, since
it is defined through the negative gradient field of f . Hence, since we want to prove
topological results, it suffices to find a metric in which the above condition holds. It
will be proved in the next section that this is the case for a large class of Riemannian
metrics on M , for an arbitrary Morse function.

The trajectory spaces

We start by presenting a theorem that is more powerful than we need, and consequently more difficult than we will be able to prove. The objective of this full section
will be to prove, with varying levels of detail, the part of it that is actually useful for
the Morse homology theorem, that is, finiteness of the number of trajectories when
the relative index is one, and coherent orientations for that case.
Theorem 3.1. If M is a closed smooth manifold, and (f, g) is a Morse-Smale pair
on M , then for any two critical points p, q M , the trajectory space of flow lines
connecting them, M(p, q) has a natural compactification to a smooth manifold with
corners M(p, q), whose codimension k stratum is
[
M(p, q)k =
M(p, r1 ) M(r1 , r2 ) . . . M(rk , q)
r1 ,...,rk Crit(f )

with r1 , . . . , rk , q all different. For the case k = 1, the boundary of M(p, q), as
oriented manifolds, we have:
[
M(p, q) =
(1)ind p+ind r+1 M(p, r) M(r, q)
rCrit(f )

Some of these terms havent even been defined yet, but the idea is clear. What
one wants to show, is that the trajectory spaces associated to the gradient flow of
f are compact up to broken trajectories, of which each part is a flow curve itself.
Moreover, one can give orientations to these boundaries in a coherent way.
To do this, first one has to define adequate topologies and structure on curve
spaces, which in this case will turn out to be that of infinite-dimensional Banach
manifolds. The next step is, through analysis of certain Fredholm operators, to
obtain a finite-dimensional manifold structure on the trajectory spaces themselves.
After that, one can work towards the compactness result itself, again with suitable
modes of convergence for the trajectory spaces. Then, it will be possible to prove
gluing results to perturb broken flow lines to actual flow lines, in such a way that
the process can be parametrized and forced to converge to the original broken flow
line, so as to obtain the mentioned manifold-with-corners structure. The final last
step will be to obtain coherent orientations for all these spaces.
This whole section follows the work done in [Sch].

3.1

Trajectory spaces. Banach manifold structure

In this section we define the trajectory spaces we need to work with. To mimic the
behavior of flow lines, we start compactifying R in the following way:
Definition 3.2. Let R = R {}, equipped with the structure of a bounded
manifold, by the requirement that:
h : R [1, 1]
t
t
1 + t2
8

be a diffeomorphism. Additionally, for x, y M we define the set of smooth,

by:
compact curves Cx,y

Cx,y
= Cx,y
(R, M ) = {u C (R, M ) | u() = x, u(+) = y}

Directly from this definition, we can get a characterization of the asymptotic


decrease of C (R) functions imposed by the differentiable structure of R.
Lemma 3.3. For each f C 1 (R, R) there is a constant c(f ) > 0, such that the
following estimate holds:
c(f )
|f 0 (t)|
3
(1 + t2 ) 2
Actually, the following useful estimates are also true:
Corollary 3.4. Given A C 1 (R, GL(n, R)), there is a constant c(A) > 0 such that
the estimate
kAsk1,2 c(A)ksk1,2
holds for all s H 1,2 (R, Rn ), and where (As)(t) = A(t) s(t), t R.
Corollary 3.5. Let f C 1 (R, R) satisfy the condition f () = 0. Then f
H 1,2 (R, R).
Definition 3.6. Now let Vec R be a C (R)-smooth, finite-dimensional vector

=
bundle on R, and let : R Rn be a smooth trivialization. Then using the
induced one-to-one mapping between the associated vector spaces of sections, we
can define:
1,2
1,2
(R, Rn )) = {1
(R, Rn )}
HR1,2 () = 1
(H
(s) | s H

Now, the first of the corollaries above implies that induces a banach space topology on the vector space HR1,2 () in a way which is independent of the particular choice
of trivialization , since the change from one trivialization to another is represented
by some A C (R, GL(n, R)).
It is convenient to note that this space hasnt been constructed from a measure
on R, but from isomorphisms with H 1,2 (R, Rn ).
Now, consider the exponential map on the complete Riemannian manifold M ,
exp : T M D M
where D is a convex neighborhood of the zero section where the exponential is
defined everywhere. We denote by h D the induced open and convex neighborhood
of the zero section in the pullback bundle h T M , for a smooth, compact curve
h C (R, M ).
1,2
, C 0 , we can define
Definition 3.7. From the Sobolev embedding Hloc

exph : HR1,2 (h D) C (R, M )


s exp s
where (exp s) = exph(t) s(t), which is well defined for h C (R, M ). Thus we can
define:
1,2
1,2

(R, M ) = {exp s C 0 (R, M ) | s HR1,2 (h D), h Cx,y


Px,y
= Px,y
(R, M )}

A proof for the three key propositions below can be found in [Sch], appendix 1:
1,2
0
Proposition 3.8. The set of curves Px,y
Cx,y
(R, M ) with given endpoints is
equipped with a Banach manifold structure via the atlas of charts

{HR1,2 (h D), exph }hCx,y


(R,M )
Additionally, the following inclusions hold:
dense

dense

1,2
0
Cx,y
(R, M ) Px,y
(R, M ) Cx,y
(R, M )

Moreover, there is a countable sub-atlas.


We shall use the following representation of the tangent space
[
1,2
1,2
HR1,2 (s T M )
T Px,y
= HR1,2 (Px,y
T M) =
1,2
xPx,y

1,2
with H 1,2 (R, Rn ) as characteristic fiber. By
Now this is a Banach bundle on Px,y
1,2
analogy to HR we can define a section functor

L2R : VecC (R) Ban


which is endowed with a Banach space topology given by L2 (R, Rn ). This directs
us to the Banach bundle
[
1,2
L2R (Px,y
T M) =
L2R (s T M )
1,2
sPx,y

Taking the above bundle structure into account, the second key proposition is:
Proposition 3.9. Let f C (M, R) be an arbitrary smooth real function on M .
Then, given critical points x, y Crit(f ) as endpoints, the gradient field f induces
a smooth section in the L2 -Banach bundle,
1,2
1,2
F : Px,y
L2R (Px,y
T M)

s s + f s
Actually, the study of the trajectory spaces is founded upon exactly this section
F , as shown by the following proposition:
1,2
1,2
Proposition 3.10. The zeroes of the section F : Px,y
L2R (Px,y
T M ) are exactly
the smooth curves which solve the differential equation s = s + f s, and also
satisfy the conditions limt s(t) = x and limt+ s(t) = y.

Definition 3.11 (Trajectory spaces). We denote the above trajectory spaces of


solutions to
s = s + f s,
subject to limt s(t) = x and limt+ s(t) = y, by M(x, y).
10

We also have the following useful lemma, which can be proved in an elementary
but somewhat involved way:
Lemma 3.12. Let X : U (0) Rn be a C 1 vectori field defined on a neghborhood
of 0 R, and let 0 be a critical point of X such that the linearization DX(0) is
non-degenerate and symmetric. Then there is an  > 0 such that the solutions to
s = X(s) with lim s(t) = 0
t

satisfy the following estimate: There are constants c > 0 and t0 R depending on
s, satisfying
|s(t)| cest for all t t0
Note that this will be the case for the gradient vector field and corresponding
flows, so we have some estimate about the asymptotic behaviour of curves in the
trajectory spacces.

3.2

Fredholm operators. Finite-dimensionality

The next step consists in showing that the above operator F is Fredholm. First,
we will prove the analogous result considering linear operators on the trivial bundle
R Rn , of the type
(FA s)(t) = s(t)
+ A(t) s(t)
with s H 1,2 (R, Rn ) and A Cb0 (R, End(Rn )). Lets settle some notation first.
We say an operator A End(Rn ) is conjugated self-adjoint if it is self-adjoint with
respect to some scalar product. Also, we will denote:
X = H 1,2 (R, Rn ), Y = L2 (R, Rn )
S = {A GL(n, R) | A is conjugated self-adjoint}
A = {A C 0 (R, End(Rn )) | A = A() S}
The set S is the set of Hessians appearing in Morse theory. So we can define:
(A) = #((A) R ) for A S
to be the so-called Morse index of A, where (A) is the spectrum of A. We can also
regard A as a normed space, with respect to k k .
We wish to study the following map:
Cb0 (R, End(Rn )) 3 A (FA : X Y )
Where (Fa s)(t) = s(t)
+ A(t) s(t). We denote this map by F . This is an affine
map, continuous by the estimate:
Z
k(FA FB )(s)k0 =

 12
|(A B)s| dt
kA Bk ksk0 kA Bk ksk1
2

Hence kFA FB kL(X,Y ) kA Bk . The central result of this section is the


following
11

Proposition 3.13. For all A A, the linear operator FA : X Y is Fredholm.


To prove it, we will need the following lemma (a semi-Fredholm operator is
one with either finite dimensional kernel, or finite dimensional cokernel, but not
necessarily both):
Lemma 3.14. Let X, Y, Z be Banach spaces and F L(X, Y ), K K(X, Z) and
c > 0 with
kxkX c(kF xkY + kKxkZ ), for all x X
Then F is a semi-Fredholm operator.
Proof. Define the set
S = {x ker F | kxkX = 1}
and consider a sequence (xk ) S. Then, by our assumption, and the identity
Fxk = 0 for all k N we get
kXk xl kX ckK(xk xl )kZ , for allk, l N
Given that K is compact, the sequence (Kxk ), and therefore (xk ) have convergent
subsequences. Thus S is also compact, and so ker F is of finite dimension.
Now, from the Hahn-Banach theorem and this finite dimension, there exists a
closed subspace X0 X satisfying
ker F X0 = X
Consider now a sequence (F xk ) R(F ) converging to Y , so without loss of generality we can assume
(xk )kN X0 , F xk y Y
Assume (xk ) is unbounded. Then we switch to the sequence
assume (maybe passing to a subsequence) that
(xk )kN , kxk kX = 1, kF xk kY

xk
.
kxk k

Thus we can

1
for all k N
k

Using the hypothesis again, there is a convergent subsequence of (xk ), since (F xk )


converges and K is compact. Thus, denoting the subsequence again by (xk ),
xk x, kxkX = 1 and F x = 0
in contradiction to the construction of X0 . Therefore, the sequence (xk ) must be
bounded, so that the same argument yields a convergent subsequence. This implies
the identity
y = F x, for some x X0
hence R(F ) is closed in Y .
Proof of the proposition. It will consist in four parts, formulated as lemmas below.

12

Lemma 3.15 (Part 1). Let A A be a constant map from R to S. Then there is
a constant c > 0, such that the estimate
ksk1 ckFA sk0 holds for all s X
Proof. By F : Y Y we denote the Fourier isometry, where we use the notations
X = H 1,2 (R, Cn ) and Y = L2 (R, Cn ) throught the proof. In particular,

F(S)(t)
= itF(s)(t), t R for all s X
Also, let : F(X) Y be the operator (s)(t) = t s(t), t R so that we obtain
the identity
F(FA (s)) = (i + A) F(s)
and therefore
Fa = F 1 (i + A) F : X Y

(3.1)

So let us assume A S, so that 0 = min |(A)| > 0 holds. We consider


B1 (), B2 () : Cn Cn ,
1

B1 () x = (1 + 2 ) 2 x, R,
B2 () = i + A, R
and hence
0
/ (B2 ()) = i + (A)
This implies that the inverse B2 ()1 exists and stisfies
kB21 ()k =

sup

|1 | =

(B2 ())

1
1
=p 2
0 + 2
(A)+i ||
sup

Summing we obtain
s
kB1 () B21 ()k

1
1 + 2
max( , 1) = c(A)
2
2
0 +
0

So we regard B1 , B21 C 0 (R, End(Cn )) as multiplication operators in L(X, Y ),


obtaining the inequality
kB1 B21 k c(A)
(3.2)
Putting all this together we get the needed estimation, combining

ksk21 = k 1 + 2 Fsk20 = kF 1 B1 B21 FF 1 B2 Fsk20


with
hF 1 B1 B2 F, i 0 = hB1 B21 F, Fi0 kB1 B21 k kk0 kk0
By the inequality (3.2) this yields
kF 1 B1 B21 FkL(X,Y ) c(A),
so that we conclude the estimate
ksk1 c(A)kF 1 B2 Fsk0
and taking (3.1) into account we get the inequality of the lemma statement.
13

Lemma 3.16 (Part 2). Given any A A, there are constants T > 0, c(T ) > 0,
that satisfy
ksk1 c(T )kFA sk0 for all s X, s|[T,T ] = 0
Proof. From step 1 we have
ksk1 c(A )kFA sk0 for all s X
We therefore define c = max(c(A+ ), c(A )). Given  > 0and A A, there is a
T > 0 large enough such that
kA A(t)k  for all t T
kA+ A(t)k  for all t T
+
Restricting ourselves to s X with s|
[T ,) = 0 and s|(,T ] = 0, we get

kFA s k0 kFA s k0 + k(FA FA )s k0 kFA s k0 + ks k0


Now consider s X such that s|[T ,T ] = 0. Then we find s as above, fulfilling
s = s+ s such that
ksk1 = ks k1 + ks+ k1 c(kFA s k0 + kFA+ s+ k0 )
c(kFA s k0 + kFA s+ k0 ) + c(ks k0 + ks+ k0 )
= ckFA sk0 + cksk0
And this finally implies
ksk1 ckFA sk0 + cksk1
and for an  < 1/c and appropiate T ()
ksk1

1
kFA sk0 , for all s X with s|[T (),T ()] = 0
1 c

Lemma 3.17 (Part 3). For any A A, there is Banach space Z and a K
K(X, Z), c > 0 satisfying
kxkX c(kFA kY + kKxkZ ) for all x X
In fact, FA is a semi-Fredholm operator.
Proof. Let T be the constant obtained in step 2. Then
Z T
Z T

2
|s + As| dt =
|s|
2 + 2hs,
Asi + kAsk2 dt
T
T

Z T 
1 2
2
|s|
|As| dt
=
2
T

14

due to the computation


1 2
|As|2
|s|
2 + 2hs,
Asi + |As|2 |s|
2
from |s + 2As|2 0. Therefore, using |A(t) s(t)| kA(t)k |s(t)| and setting
c = max[T,T ] kA(t)k, we conclude
Z

1
|s + As| dt
2
2

|s|
c

|s|2 dt

Hence, there is a c > 0 satisfying


Z
Z T

2
2
|s|
+ |s| dt c
T


|s|2 + |s + As|2 dt.

(3.3)

Now, defining a cut-off function C (R, [0, 1]) with the properties
(
0, |t| T + 1
6= 0 for |t| (T, T + 1)
(t) =
and (t)
1, |t| T
we can combine the estimate (3.3) with step 2 and get
ksk1 = ks + (1 )sk1 ksk1 + k(1 )sk1
c(ksk0 + kFA (s)k0 + kFA ((1 )s)k0 )
for c > 0 big enough. That is



ksk1 c ksk0 + 2ksk0 + kFA sk0 + k(1 )FA sk0



c kskL2 ([T 1,T +1]) + kFA sk0
And considerng the following composition of a continous restruction map and the
Rellich-Kondrachov compact embedding:
K : H 1,2 (R, Rn ) H 1,2 ([T 1, T + 1], Rn ) , L2 ([T 1, T + 1], Rn ) = Z
we obtain
K:XZ
s s|[T 1,T +1] L2 ([T 1, T + 1])
as a compact operator. Thus lemma (3.14) completes the proof
Lemma 3.18 (Part 4). FA is a Fredholm operator.
Proof. Up to now, we know that FA has finite-dimensional kernel, and that its
arange is closed in Y . So the cokernel of FA is a banach space satisfying coker FA
=
R(FA ) , because Y is a Hilbert space. Thus, let r R(FA ) , that is
hr, s + Asi0 = 0for any s H 1,2 .
15

In particular, we can deduce


hr, i
0 = hAt r, i0 , for all C0 (R, Rn )
But, by definition, this means r is weakly differentiable, with r = At r L2 . Hence
r W 1,2 .
Then we know that r X and r ker FAt , beacuse At A. Therefore, there
is an isomorphism
coker FA
= ker FAt
And working analogously to the proof sof step 3, we have dim coker FAt < . This
concludes the proof.
Now our aim is to express the associated Fredholm index in terms of the Morse
indices of the critical points. At this point, only the changes of sign of eigenvalues
of A A is important. The proof will consist in transforming A in such a way that
the Fredholm index doesnt change, but can be easily analyzed.
Definition 3.19. From the proposition, we consider the subset
= F (A) = {FA L(X, Y ) | A A} F(X, Y )
and denote the equivalence class of operators from with respect to the relation
B = A by
FA = {FB | B = A }, A A
The folowing lemma will be crucial:
Lemma 3.20. Given F , the class F is contractible within as a subspace of
F(X, Y ).
Proof. Let F = FA0 be arbitrary and define = F . Then we study the map
(just like a linear homotopy):
: [0, 1] with (, FA ) = FA( )
A( ) = (1 ) A + A0 , that is
FA( ) for all [0, 1], since A( ) = A = A
0
It is clear that
(0, ) = id and (1, ) = AA
Thus, we must show the continuity of in both variables, so let us start with
lim n = and lim FAn = FA

that is
lim kFAn FA kL(X,Y )=0 .

so we must check that


lim kFA( ) FAn (n ) kL(X,Y )=0 .

16

(3.4)

So, let  > 0 and nk a subsequence (nk )kN satisfying


kFA( ) FAnk (nk ) kL(X,Y ) 
Taking into account equation (3.4) above,
k(A( ) Ank (nk )) unk k0
=k[(1 ) A + A0 (1 nk ) Ank nk A0 ] unk k0
=k(A Ank ) unk + ( nk )A0 unk
+ (nk A A + nk Ank nk A) unk k0
kFA Fnk kL(X,Y ) + | nk |kA0 kL(X,Y )
+ | nk |kAkL(X,Y ) + |nk |kFA FAnk kL(X,Y ) 0
which is a contradiction.
This last lemma implies that the index map ind : Z must be constant when
restricted to one of the FA , and this means that ind FA is determined uniquely by
the endpoints A S. Knowing this, we carry out an appropiate conjugation of
FA :

Lemma 3.21. For any A A, we can find a D A of the form diag(


1 , . . . , n )
with
ind FD = ind FA

Proof. Since A A we have A S. Then there is C GL(n, R) such that

C A (C )1 = diag(
1 , . . . , n )

so that the eigenvalues are ordered by sign, that is, sign


i sign i+1 .

Also, the ends C can be chosen in order to satisfy det C > 0. so that both
+
C and C lie in the same pathwise connected component of GL(n, R). Thus there
is a curve
C C (R, GL(n, R)) with ends C() = C

which is also eventually constant. This means


(
C +, t T
C(t) =
for some T > 0
C , t T
We shall henceforth denote by C both multiplication operators
CX : X X and CY : Y Y
sC s
Its obvious that C represents a linear isomorphism and that the following identities
hold:

1
(CY FA CX
)(s)(t) = C(t) ( + A(t)) (C 1 (t) s(t))
t

1
1
= s(t)
+ C(t) (C )(t) + C(t)A(t)C (t) s(t)
t


= FC +CAC 1 s (t)
t

17

Here,

(C 1 )(t)
t

= 0 holds for |t| T . Thus we compute the ends as




1
1

C (C ) + CAC
= diag(
1 , . . . , n ) = D
t

obtaining the relation


1
CY F A CX
D

Now, given that CX : X X and CY : Y Y are isomorphisms, the identity


1
ind(CY FA CX
) = ind FA

Follow from the composition rule for Fredholm indices.


And we finish with:
Proposition 3.22. Given any A A, the Fredholm index of FA equals the relative
Morse index,
ind FA = (A ) (A+ )
Proof. It is sufficient to compute ker FA and coker FA when A is of the shape:
A(t) = diag(1 (t), . . . , n (t)), and
sign i () sign i+1 (), i = 1, . . . , n 1
and i (t) constant for |t| 1
Now, s H 1,2 (R, Rn ), s(t) = (s1 (t), . . . , sn (t)) is in ker FA if and only if it represents
a global solution of the sysem of differential equations:
s t = i (t) si (t), i = 1, . . . , n
with the bounded functions i : R R. Now, by explicit calculation, for large
times we have
s i (t) = i (t) si (t), t R and 0 6= si H 1,2 (R, R)
if and only if
(
ei t , t < 1
+

<
0
and

>
0,
as
s
(t)
=
+
i
i
i
ei t , > 1
And, as the eigenvalues are ordered by sign, we compute
+
dim ker FA = #{k {1, . . . , n} |
k < 0 and k > 0}
= max((A ) (A+ ), 0)

Analogously we compute
dim coker FA = max((A ) (A+ ), 0) = max((A+ ) (A ), 0)
from the isomorphism coker FA
= ker FAt
= ker FA . Now, the difference between
the two dimensiones above gives ind FA = (A ) (A+ ).
18

3.2.1

Non-trivial bundles

The case that is actually useful for Morse theory is that of a smooth vector bundle
on R endowed with a Riemannian metric, so that H 1,2 () and L2 () are well defined,
as in the previous section. The principal obstruction is that one has to generalize
the time derivation to some covariant derivation that might disturb the Fredholm
property.
The key to all this turns out to lie in that the Christoffel symbols associated
to the connection and any trivialization should vanish asymptotically (i.e, at ).
If this happens, the covariant derivation is called Fredholm-admissible, and the
corresponding operators are, as expected, Fredholm. It is obvious that the bundles
and connections arising in our context satisfy this condition. More details can be
found in [Sch].

3.3

Transversality. Manifold structure

We now focus on endowing the trajectory spaces a finite-dimensional manifold structure, building upon the results of the previous one together with the Morse-Smale
condition. The result we are aiming for is the following:
Theorem 3.23. If (f,g) satisfy the Morse-Smale condition, the trajectory spaces
1,2
of finite dimension (x) (y).
M(x, y) are closed submanifolds of Px,y
The proof of this result will be quite technical, so we need to refresh some
definitions first.
Definition 3.24 (Genericity). In a Baire space X, we call a subset X a G-set,
if it is a countable intersection of open and dense subsets. A subset G X is generic
with respect to some condition if the condition holds for a G-set G.
It is now time to enunciate the above theorem in a more general (and technical)
fashion, which hints at the way its proven:
Proposition 3.25. Let G and M be Banach manifolds and : E M a Banach
bundle on M with fiber E. Additionally, let : G M E be a smooth Gparameter section, i.e. smooth and (g, ) a smooth section in E for each g G.
Furthermore, let satisfy the condition:

There is a countable trivialization {(U, )}, : E|U U EU , such that for each
(U, ):
(a) 0 is a regular value of pr2 : G U EU
(b) pr2 g : U EU is a Fredholm map with index r for al g G.
Then there is a G-set G such that the set
Zg = 1
g (0) = {m M | g (m) = 0}
is a closed submanifold of M for al g .
19

Remark 3.26. Let us translate the result into our starting Morse theory terms:
Our G will be a Banach space of metrics to be defined later (not trivial), the Gparameter section will be the one from the previous sections, that is, for a starting
F
1,2
1,2
manifold N , M = Px,y
, E = L2 (Px,y
T N ) amd g = F where s s + f s is
the one from proposition 3.9, which implicitly depends on the metric through the
definition of the gradient.
As for the hypotheses (a) and (b), they are part of the Fredholm results in
the previous section, when specialized to this situation, subject to the countable
sub-atlas of proposition 3.8.
And the result itself provides us with both the manifold structure of the trajectory spaces, and the Morse-Smale condition for a generic set of metrics G.
For the proof, we will need the following somewhat-standard Banach space
lemma, which we dont prove here:
Lemma 3.27. Let be a bounded, linear map of Banach spaces, which is onto and
of the form
: E F G, (e, f ) = 1 (e) + 2 (f )
and such that 1 , 2 are continuously linear, and let 2 : F G be a Fredholm
operator. Then there is a decomposition
E F = ker H
such that H is closed in E F .
We will also need the Sard-Smale theorem, which is a generalization of Sards
theorem for Fredholm maps on Banach manifolds:
Theorem 3.28. Let f : M V be a C q Fredholm map with q > max(ind f, 0).
Then the regular values of f is a G-set in V .
Proof. Actually not too complicated, can be found in [Sm].
Proof of proposition 3.25. Owing to the countable trivialization of E we can assume
without loss of generality that E is a trivial bundle. This is to say that there is a
smooth map
:GM E
with respect to the Banach space E, such that 0 E is a regular value and the maps
g : M E are T
Fredholm maps of index r for all g G. This assumption can be
made since = U U is again a G-set if we consider the countable trivialization
{(U, U )}. Thus
Z = 1 (0) G M
is a smooth Banach manifold with associated tangent spaces given by
Tz Z = ker D(z) for all z = (g, m) Z
as follows from local coordinate charts, lemma 3.27 and the implicit function theorem.
20

Now we need to conclude that the restriction of the projection map to this
Banach manifold Z,
:ZG
is a Fredholm map endowed with the same index as g , if we consider z = (g, m) Z.
First, let us focus on the kernels of the maps D(x) : Tz Z E and D2 (z) : Tm M
E. If we assume D2 (z) Fredholm, we obtain the identity (of finite dimensional
vector spaces)
ker D(z) = Tz Z Tm M = ker D2 (z)
(3.5)
Next, we consider the following homomorphism between the quotient vector spaces:
g
D
1 : Tg G/R(D) E/R(D2 )
[v]R(D) [D1 v]R(D2 )
having in mind the identities
v = D (v, w) and D1 v = D2 w
for each pair (v, w) Tz Z Tg G Tm M , we deduce the isomorphism property of
g
D
1 from D(z) being onto. Now, knowing that coker D2 (z) is of finite dimension,
g
D1 appears to be an isomorphism of finite-dimensional cokernels,

=
g
D
1 : coker(D(z)) coker D2 (z)

(3.6)

Hence, 3.5 and 3.6 imply the Fredholm property of the projection map . Now we
can apply the Sard-Smale theorem to this map and get a G-set of regular values of
.
Now, to see D(z) onto, let b . If (b, m) does not vanish for any m M ,
0 is trivially a regular value of b . Therefore, let us consider an m M with
(b, m) = 0. Then, we have to show the operator
Db (m) = D2 (b, m) : Tm M E
to be onto. Let E be arbitrary. Due to regularity of b with respecto to , there
is a pair (, ) Tb G Tm M satisfying
= D(b, m) (, ) = D1 + D2

(3.7)

But b is also regular with respecto to : Z G, that is, for any Tb G one can
find an (0 , 0 ) T(b,m) Z such that it holds
0 = D(b, m) (0 , 0 ) =
Hence (0 , 0 ) ker D(b, m), that is
0 = D(b, m) (0 , 0 ) = D1 + D2 0

(3.8)

Substractiong 3.8 from 3.7, we find a preimage of under D2 (b, m) of the form
= D2 (b, m) ( 0 )
and we are finished.
21

3.3.1

Genericity of the Morse-Smale condition

Given a Morse function, there is a set of metrics generic with respect to the MorseSmale condition. This will be derived from proposition 3.25, where the only thing
remaining is to define a suitable Banach manifold of metrics.
We will start with a fixed, arbitrary metric g0 and find a generic set of variations
with respect to g0 . By this we mean a set of smooth sections A End(T M ) which
are fiberwise self-adjoint with respect to g0 , positive definite and which differ from
the identity section by a small amount.
Such a set can be explicitly given the structure of a Banach manifold, by defining
a certain kind of norms on sections of any vector bundle. The details are rather
technical, and can be found in [Sch] and [Flo].

3.4

Compactification

In this section, we want to study the compactness properties of the trajectory spaces.
The implicit function theorem implies that the finite-dimensional manifolds M(p, q)
are manifolds without boundary, since they arise from the differential equation
= f
3.4.1

The space of unparametrized trajectories

Before going into the main part of the section, it will be useful to obtain some more
information about the trajectory manifold M(x, y). We want to analyze one of its
essential properties, that is, its symmetry with respect to time-shifting, or more
formally, additive reparametrization of the trajectories.
First, lets ground some notation. Let : R M be a solution of the differential
equation = f . Then it is clear that the shifted curve = = ( + )
is also a solution, since

( ) = (
+ )
t
Moreover, it holds that (R) = ( )(R) for all M(x, y), that is, the image (R)
remains unchanged. It is in this sense, of image sets, that we will mean geometrical
behavior of the trajectories.
Proposition 3.29 (Equivalence of the trajectory space definitions. Group action).
The additive group R acts smoothly, freely and properly on the manifold M(x, y) by
R M(x, y) M(x, y)
(, )
provided that x 6= y, so that M(x, y) consists in non-constant trajectories.
Proof. We shall make use of the identification M(x, y) W u (x) W s (y). To prove
it, consider the smooth evaluation map:
1,2
E0 : Px,y
(R, M ) M

22

(0)

Smoothness follow immediately from the representation in suitable local coordinates,


E0,loc : HR1,2 (h D) U (h(0)) Th(0) M



exp1
E

exp
0
h () = (0)
h(0)
where E0,loc is continously linear. Restricting the evaluation map E0 to the trajectory
space M(x, y), we obtain an embedding
E0 : M(x, y) , M
because the differential
DE0 () : T M(x, y) T(0) M
DE0 () = (0)
is injective at each M(x, y). To see that, let us consider T M(x, y) =
ker DF (), satisfying (0) = 0. As analyzed in the Fredholm section, may be
treated as the solution of a linear ordinary differential equation, so that uniqueness
of solutions gives us 0. So the evaluation map leads to a diffeomorphism between
the manifolds

E0 : M(x, y) W u (x) W s (y)

(0)
Now, it is obvious that E0 identifies the group action (, ) with the negative
gradient flow,
R (W u (x) W s (Y )) W u (x) W s (y)
(t, p)

t (p)
This gradient flow represents a smooth, free and proper R-action. Hence the equivalence = (E01 t E0 )() finishes the proof.
Now the following definition makes sense:
Definition 3.30 (Unparametrized trajectory spaces). Given (M, g) a riemannian
manifold, f a Morse function on it, and p, q Crit(f ), we define the space of
unparametrized trajectories by the formula:
\q) = M(p, q)/
M(p,
where stands for the relation induced from the action of the group R by translation
of the time parameter. In view of the previous results, this quotient is a well-defined
manifold of dimension ind p ind q 1.
Now, define the smooth function = f E0 : M(x, y) R with d() =
df ((0)) (0) = hf ((0)), (0)i.

23

Definition 3.31. Let f (y) < a < f (x), such that (0) is not critical for any
1 (a). Then a is a regular value of and
Ma (x, y) = 1 (a)
is a (ind(x) ind(y) 1)-dimensional submanifold of M(x, y).
The following proposition gives us an interesting description of the unparametrized
trajectory spaces:
Proposition 3.32. The map
a : R Ma (x, y) M(x, y)
(, )
represents an R-equivariant diffeomorphism, with respect with the trivial action on
the left and the above one on the right.
Proof. The R-equivariance a ( + , ) = a (, ) is obvious from proposition
3.29. It is also easy to verify bijectiviy, with 3.29 and the uniqueness of solutions of
ordinary differential equations. Thus, it is enough to prove that
D(, ) : R T Ma (x, y) T M(x, y)
is an isomorphism. Let (s, ) R T Ma (x, y). Then the definition of above
gives us
df ((0)) (0) = 0
(3.9)
and given (s, ) ker Da (0, ), the equation
0 = Da (0, ) (s, ) = s +
follows. From = f we obtain (0) = s f ((0)), so that (3.9) implies
(0) = 0. So we conclude the injectivity of Da (0, ), and the surjectivity follows
by a dimension argument. Now, the identity
a (, ) = (a (0, ))
finishes the proof, since by proposition 3.29, is a diffeomorphism.
Remark 3.33. This means a is a diffeomorphism
\y) M(x, y) Ma (x, y)
M(x,
\y) can be viewed as the transversal intersection of the surface associated
so M(x,
to a regular level of f with W u (x) W s (y).

24

3.4.2

Compactification result

We now turn our attention to the main compactness result. For that, it will be
convenient to introduce some a new condition for f .
Definition 3.34 (Palais-Smale condition). A function f fulfills the Palais-Smale
condition, if every sequence (xn )nN M , such that (|f (xn )|)nN is bounded and
|f (xn )| 0, has a convergent subsequence.
Remark 3.35. Note that a Morse function on a compact manifold will automatically
satisfy this condition.
The precise sense of compactness we will be talking about is the following:
c y)
Definition 3.36 (Compactness up to broken trajectories). A subset K M(x,
is called compact up to broken trajectorioes of order , or up to ( 1)-broken
trajectories exactly if, for all (
un )nN K, either:
(1) There exists a convergent subsequence (
unk ), or
(2) There are critical points
x = y0 , . . . , yi = y Crit(f ), 2 i
and connecting trajectories together with associated reparametrization times
vj Myj ,yj+1 , (n,j )nN R, j = 0, . . . , i 1
such that for some subsequence (nk )kN , the following convergence holds
C

loc
vj
unk nk ,j

Definition 3.37 (Geometrical convergence). The convergence of trajectories above,


C

loc

w in Cx,y
(R, M ) will be called weak convergence in what follows, and the weak
wj
convergence of unparametrized trajectories subject to suitable reparametrization
times will be called geometrical convergence.

The main result of this section is the following


Theorem 3.38. [Compactification result]If f satisfies the Palais- Smale condition,
\y) is compact up to broken trajectories of order (x) (y).
the manifold M(x,
This immediately implies the following fact, which will be crucial for constructing
Morse homology:
\y) is a finite set.
Corollary 3.39. If, moreover, ind x ind y = 1, then M(x,
Now we state an easy consequence of the Palais-Smale condition, which we will
need to prove the compactness result.

25

Proposition 3.40. If f satisfies the Palais-Smale condition above, there is an  > 0


depending merely on f , x and y, such that the set
Kx,y = {z M | f (y) f (z) f (x), kf (z)k }
is relatively compact.
Lets get started. As it is often the case with compactness results, an aplication
of the Arzela and Ascoli theorem is at the heart of the argument:
Lemma 3.41. Every sequence (un )nN M(x, y) has a weakly convergent subsequence
C

loc
unk
v C (R, M )

Proof. The strategy will be to transition from the H 1,2 topology of the Hilbert
0
submanifold M(x, y) to the Cloc
topology. it will be crucial to have an uniform
12
bound for the H -norm of trajectories with fixed endpoints x, y Crit(f ). Once

0
will, in principle, follow from elliptic regularity.
-convergence, the Cloc
we have Cloc
Now, the fact that the trajectories un arise from the negative gradient flow, with
fixed endpoints x, y gives us the estimate
Z t
Z t
2
hu n , f un id f (x) f (y), for all s t
(3.10)
|u n ( )| d =
s

Now let d be the Riemannian metric on M . We get


sZ
Z t
t
p
p
p
|u n ( )|d |t s|
|u n ( )|2 d |t s| f (x) f (y)
d(un (t), un (s))
s

where the first inequality is by Holders inequality, and the second by 3.10. This in
turn means equicontinuity of the un in C 0 (R, M ). To be able to apply the theorem of
Arzela-Ascoli, we need to obtain pointwise convergence. For now we have a uniform
L2 bound on the derivatives, and the Palais-Smale condition. Choose a fixed t0 R.
By proposition 3.40, we can assume without loss of generality that
|(f un )(t0 )|  for all n N
And due to the asymptotic decrease lim|t| |f (un (t))| = 0 we can find a sequence
(tn )nN satisfying:
|f (un (tn ))| =  and |f (un (s))|  for all s [tn , t0 ]

(3.11)

and thus the points un (tn ) Kx,y . In other words, there is an r0 > 0 such that
d(un (tn ), x) < r0 holds for all n N. Now, using the inequality
Z s
d(x, un (t0 )) < r0 + ln (t0 ) with ln (s) =
|u n ( )|d
tn

we see that the only thing needed for pointwise convergence is a bound for ln (t0 ).
Calculating
dln
( ) = |f un ( )| 
ds
26

d(f un )
( ) = |f un ( )|2
ds
with 3.11, for tn t0 gives us the estimate
 
dln
1 d
( )
(f un )( )
ds
 ds
hence we obtain the desired upper bound from
Z t0
Z
dln
1 tn d(f un )
f (un (tn )) f (un (t0 ))
( )d
( )d =
ln (t0 ) =
 t0
ds

tn ds
f (x) f (y)


and, since (un (t))nN is a compact set for each fixed t R, by the Arzela-Ascoli
theorem we get a subsequence (unk )kN converging on compact intervals. That is,
there exists v C 0 (R, M ) with v(t) = limk unk (t), such that
unk |[R,R]

C 0 ([R,R])

v|[R,R] for all R 0

And the fact the trajectories are solutions of the differential equation u = f u
implies the C k -convergence,
unk |[R,R]

C k ([R,R])

v|[R,R]

iteratively for all k N.


Lemma 3.42. Let (un )nN M(x, y) be a weakly convergent sequence,
C

loc
un
v C (R, M )

such that v M(x, y), that is, v is in the same trajectory space. Then the sequence
(un ) is also H 1,2 -convergent,
1,2
Px,y

un v M(x, y)
Proof. It consists on showing that the elements un of the sequence converge uniformly toward y and x when t respectively. If that is true, proposition 3.12
about the asymptotical behavior of the gradient trajectories implies uniform expo
nential asymptotic convergence, which combined with the Cloc
convergence eventu1,2
ally implies H -convergence.
The details are more technical than interesting, and can be found in [Sch].
Proof of theorem 3.38. Now the objective is to pass to the unparametrized trajectory spaces, and explore the obstructions to strong H 1,2 convergence there. Lemma
3.41 gives us weak convergence to a curve v C , and thus to a trajectory of the
negative gradient flow, that is v + f v = 0, but it need not be the case that this

27

new trajectory is in the same trajectory space. Given any convergent subsequence
obtained from lemma 3.41, local convergence:
C

loc
unk
v C (R, M )

(3.12)

f (v(t)) [f (y), f (x)] for all t R

(3.13)

implies the estimate:

so the possible trajectory manifolds where such trajetories lie, of the form M(x0 , y 0 )
must satisfy:
f (y) f (y 0 ) f (x0 ) f (x).
Now we deduce that v is actually an element as element of these spaces, from 3.12
and the Palais-Smale condition, as follows:
First, by the above restriction for the values f (v(T )), and the equation v =
f v:
Z T
2
|v(s)|

ds = f (v(T )) f (v(T )) f (x) f (y) for all t R


T

so kvk
L2 , which in turn implies (being v continous) that
lim f (v(t)) = lim v(t)
=0

(3.14)

And finally, by combining the Palais-Smale condition of f with 3.13 and 3.14, we
deduce the relation v M(x0 , y 0 ), where x0 , y 0 satisfy f (y) f (y 0 ) f (x0 ) f (x).
Now, we have to look separately at the different cases for x0 and y 0 . If v
M(x, y), the previous lemma gives us strong convegrence and we are finished. The
other cases which lead to splitting up into broken trajectories, that is, without loss
of generality,
v M(x, y 0 ) with f (y) < f (y 0 )
can be trated by suitable reparametrizations. Let us choose a regular level a of f ,
with f (y) < a < f (y 0 ) and a reparametrization of unk ,
uk = unk k = unk ( + k ), (k )kN R
such that the identity f ((unk k )(0)) = a holds for all k N. Using lemma 3.41
C

loc
again on this sequence (
uk ), we get another weakly convergent subsequence ukl
v,
which satisfies
f (y) f (
v (+)) f (
v ()) f (y 0 )

Now, iterating these procedures of sorting by the values of f at the ends of the trajectories and reparametrizating either provides us, step by step, with subsequences

H 1,2 -convergent or Cloc


-convergent to constant trajectories, or recovers new critical
points. But, by the results of the previous section, the indices at the ends of the
non-constant trajectories must be different:
ind(v (j) ()) ind(v (j) (+)) > 0
so this iterative process of reparametrization must stop at broken trajectories of
order at most ind(x) ind(y). The theorem follows.
28

3.5

Gluing

In the last section, we saw how the only obstructions to compactness of the trajectory
spaces are broken trajectories, joining the critical points in question, but passing
through other critical points, in a non-smooth fashion. Now we want to introduce
a gluing operation to be able to analyze the behavior of critical lines lying near the
boundary. Specifically, one defines a map # that maps simply broken trajectories,
that is, pairs (u, v) M(x, y) M(y, z) into the trajectory space M(x, z) in a way
depending on the positive parameter , so that when + the glued trajectory
approaches the original one in a certain way.
This is made precise in the following theorem:
Theorem 3.43. Given a compact set of simply broken trajectories K M(x, y)
M(y, z), there is a lower bound K 0 and a smooth map
# : K [K , +) M(x, z)
(u, v, ) u# v
satisfying: The map # : K , M(x, z) is an embedding for each gluing parameter
\y) M(y,
\z) of unparametrized
M(x,
K . Moreover, given a compact set K
trajectories, # induces an smooth embedding
\z)
:K
[ , ) , M(x,
#
K
such that we obtain weak convergence towards the simply broken trajectory
C

loc
v
(
u, v)
u#

as +. Conversely, any sequence of unparametrized trajectories converging to

a simply broken trajectory lies within the range of such a gluing map #.
Here we wont go into the details of this construction, since it is quite involved,
and we already have almost everything we need for constructing a Morse homology
theory, most importantly, finiteness of the unparametrized trajectory sets of relative
index one (by compactness), and the orientations to be constructed below.
The gluing operation tells us two things: that every simply broken trajectory that
could appear as a boundary actually does (note that this wasnt clear from the compactification result), since we can glue them, and that we can embed those broken
trajectories smoothly into the trajectory spaces, hence generating a manifold-withcorners structure. This, combined with coherent orientations would finish theorem
3.1, at least the codimension one part, which is the interesting one for our homology.

3.6

Orientation

Right now, we would be ready to build a version of the Morse homology (that is,
homology calculations from critical points) with coefficients in Z2 , by counting flow
lines from a critical point p to critical points q with ind q = ind p 1. However, to
be able to admit arbitrary coefficient groups (i.e., coefficients in Z, by the universal
29

coefficient theorem) we need to get orientation results for the trajectory spaces
involved.
Here we will depart from the analytical point of view, using more geometrical
methods, in order to shorten the exposition. The disadvantage is that these methods
are only valid in a finite-dimensional setting, and although that is enough for us, it
leaves little room for generalization.
The starting point is the diffeomorphism of proposition 3.29,
M(x, y) W u (x) W s (y)
The main result that enables us to define orientations is the following:
Proposition 3.44. Let (M, g) and f satisfy the Morse-Smale condition, and let
x, y Crit(f ), M(x, y) be given. Under the above identification, and choosing an orientation for W u (x), for any point z (R), we have an isomorphism,
canonical at the level of orientations:
T W u (x)
= Tz (W u (x) W s (y)) (Tz M/Tz W s (y))
\y) Tz Ty W u (y)

= T M(x,
Proof. The first isomorphism comes from the Morse-Smale condition.
For the second line, the isomorphism
\y) Tz
Tz (W u (x) W s (y))
= T M(x,
is clear from proposition 3.29 and its corollaries. And the isomorphism
Tz M/Tz W s (y)
= Ty W u (y)
is obtained by translating the subspace Ty W u (y) Tq M along while keeping it
complementary to T W s (y).
Definition 3.45 (Orientation of M(x, y)). We define an orientation in M(x, y) so
that the isomorphism above is orientation-preserving.
Note that this is dependent on which orientation one starts with, but as long as
the picks are all consistent, (e.g. letting the W u (x) be all oriented as submanifolds
of M , assuming M orientable), it wont cause problems.
Now, for the case of interest to us, that is, when the critical points have relative
index one (it will be the only one used later), defining an orientation is equivalent
\y). This leads to the following definition:
to choosing a sign for each point in M(x,
Definition 3.46. In the above situation, that is, trajectory space orientations inherited from an orientation for M , we define the integer
#M(x, y)
\y) according to the
as the sum of the numbers (1) assigned to each u M(x,
orientation.
Remark 3.47. For our purposes, the orientation chosen does not matter since a global
sign change will eventually give the same results.
This will later form the basis for the Morse differential operator, for the Morse
complex with groups defined as the free abelian groups with basis the critical points
with given index.
30

Sheaves, cohomology and currents

Here we investigate connections between the topology of a smooth manifold M, and


the algebra of differential forms on it. Specifically, de Rham theorem relating singular and de Rham cohomology, and the isomorphism between de Rham cohomology
and the cohomology of the current sheaf on M.
This is all done through basic sheaf theory, which provides a systematic way to
track local data defined on the manifold, and its relation with the global data. The
treatment here is essentially that of [We], chapter II.

4.1

Basics. Presheaves and sheaves

Definition 4.1 (Presheaf). A presheaf F over a topological space X, with values


on a category C consists of:
(a) For each open subset U X, an object F(U ) in C.
(b) If V U , a morphism rVU : F(U ) F(V ). These are called restriction
morphisms.
Such that the following conditions hold:
(p1) rUU is the identity morphism for all U .
U
V
.
rVU = rW
(p2) If W V U , then rW

Definition 4.2 (Morphism of preseaves). If F and G are presheaves over X, a


morphism of presheaves h : F G, consists of morphisms hU : F(U ) G(U ), such
that the following diagram commutes, if V U X:
F(U ) G(U )
hU

rU
r U
yV
yV
F(V ) G(V )
hV

Definition 4.3 (Sheaf). SA presheaf is called a sheaf if for each covering Ui of an


open set U , that is, U = i Ui , the following hold:
(s1) If s, t F(U ), and rUUi (s) = rUUi (t) for all i, then s = t.
(s2) If si F(Ui ), and if for Ui Uj 6= we have:
U

rUUiiUj (si ) = rUijUj (sj )


For all i, then there is an s F(U ) such that, for all i, rUUi (s) = si .
Thus, informally, a presheaf is a way to coherently assign objects in a category to
open subsets of a topological space, and a presheaf is a sheaf when such local objects
can be globalized in a certain sense.
31

Example 4.4 (Constant sheaves). Let X be a topological space and G an abelian


group. If we assign G to each connected open set U X, we trivially get a sheaf
on X. This sheaf will be called the constant sheaf G on X, denoted also by G.
Example 4.5 (Sheaves of differential forms on a differentiable manifold). On a
differentiable manifold M we define the sheaf of differential forms or order k on it,
assigning to each open set U M the set k (U ). It is easy to check that this
definition fits the conditions to be a presheaf, and a sheaf.
Example 4.6 (A presheaf which is not a sheaf). Define, on the complex plane C,
the presheaf F which assigns to each U C the algebra of bounded holomorphic
functions with domain U . Now consider the open disks Uj = {z C | |z| < j},
which cover the complex plane. Now let fj F(Uj ) be defined by fj (z) = z. But
then there can be no f F(C) such that f |Uj = fj , and thus the second axiom
for sheaves doesnt hold. In fact, by Liouvilles theorem, F(C) consists of constant
functions only.
Given a presheaf F over X, a point p X, and a neighborhood basis at p
denoted by B(p), then it is obvious the images of the sets in said basis, along with
the correspondent restriction maps, form a direct system in C. This motivates the
following definition, when C is a category with algebraic structure preserved by
direct limits, such as abelian groups or commutative rings:
Definition 4.7 (Stalk. Germ). Let F be a presheaf over X, p X. Then we
define the stalk of F at p as the direct limit of the sets F(U ) with x U , with
respect to the restriction maps rVU . Given s F(U ), we define its germ at x to be
sx = rxU (s), where rxU : F(U ) Fx is the natural homomorphism taking an element
to its equivalence class in the direct limit.
Note that in the usual cases, like the sheaf of differentiable real functions on a
manifold, this is equivalent to the usual definition of germs.

4.2

Resolutions of sheaves

We now turn to the sequences of sheaves that will allow us to compute cohomology.
First we will look at a natural way to make presheaves a natural object, which in
turn will allow us to get a sheaf out of an arbitrary presheaf, even if not a sheaf by
itself.
e space. Section). An etale space over a topological space X
Definition 4.8 (Etal
is a topological space Y , together with a continous and surjective map : Y X
which is also a local homeomorphism. Given an etale space : Y X, and an
open set U X, a continous map f : U Y such that f = 1U is called a section
over U . The set of such maps is denoted (U, Y ).
Remark 4.9. Its easy to see that these sets of sections form a sheaf over X, which
in fact is a subsheaf of the sheaf of continous functions from X to Y .

32

Now we will see a way to associate an etale space F to a presheaf F. This will
be done in such a way that if F is also a sheaf, then the sheaf of sections of the etale
space F is another model for F.
The idea is to take the germs
S of elements of the sheaf to be the points of the
new space, that is, to take F =
Fx . Then the map : F X taking each germ
xX

to its basepoint is obviously surjective.


It remains to define a suitable topology on this space. For each s F(U ), define
an application s : U F by s(x) = sx . Note that s = 1U . We take the sets
{
s(U ) | s F(U ), U X open} to be a basis for the topology in F (it is obvious
that this is a base for a topology). Then all the functions s are continous, and in
fact this is the final topology for this set of functions.
One then defines the sheaf on X which assigns, to each open set U X, the
which inherits the algebraic structure of the sets F(U )
set of local sections (U, F),
(given that such structure is preserved by direct limits). That this is actually a sheaf
is seen in the following proposition:
Proposition 4.10 (The associated sheaf). For any presheaf F, the assignment
defined above is a sheaf. Moreover, if F is a sheaf, then it is isomorphic to F.
with
Proof. Now define the map : F F by the functions U : F(U ) (U, F),
U (s) = s as defined above. It suffices to check that U is bijective for each U. For
injectivity, suppose that there are s1 , s1 F(U ) with U (s1 ) = U (s2 ).
Then, for all x U , U (s1 )(x) = U (s2 )(x), that is, rxU (s1 ) = rxU (s2 ). But
then, by the definition of direct limit, there exists a neighborhood V of x, such that
rVU (s1 ) = rVU (s2 ).
S
But this is true for each x, so we can produce an open cover ni=1 Ui = X, such
that rUUi (s1 ) = rUUi (s2 ) holds for each i. Since F was a sheaf to begin with, we must
have s1 = s2 .

Now lets check surjectivity. Let (U, F).


Then for x U there is a
neighborhood V of x, and s F(V ) such that (x) = sx = V (s)(x). By the
definition of section of an etale space, if any two sections agree on a point, they
agree on a neghborhood of that point. So let x X. We have, for some W a
neighborhood of x,
V
|W = V (s)|W = W (rW
(s))
Sn
Since x was arbitrary, we get an open cover i=1 Ui = U , such that for each i there
exists si F(Ui ), and |Ui = Ui (si ). Also, we have
Ui (si ) = Uj (sj ) on Ui Uj
U

using the injectivity proved above, this means rUUiiUj (si ) = rUijUj (sj ). Now, since
F is a sheaf and the Ui cover U , we can patch together the si to get s F(U ) such
sthat rUUi = si . Then
U (s)|Ui = Ui (rUUi (s)) = Ui (si ) = |Ui
which in turn gives us U (s) = .
33

In the rest of our discussion, we will assume all sheaves to be of abelian groups,
maybe with additional structure (e.g. sheaves of rings).
Definition 4.11 (Exact sequence of sheaves). For A, B, C sheaves of abelian groups,
the sequence:
ABC
is exact if the induced stalk sequence:
Ax Bx Cx
is exact for every x. Equivalently, for each x there is a neghboorhood x Ux such
that the sequence A(Ux ) B(Ux ) C(Ux ) is exact. That is, exactness needs to
hold only locally.
There is an specific type of exact sequence of sheaves which will be particularly
useful for cohomology:
Definition 4.12 (Sheaf resolution). A resolution of a sheaf F is an exact sequence
of sheaves of the form:
0 F F0 F1 . . . Fn . . .
The following fairly simple case will be of special interest to us:
Lemma 4.13 (Poincare lemma). Let A Rn is an open and star shaped, that is,
there is a point p A such that A contains all lines joining p to any other point.
Then every closed differential form on A is exact.
Proof. Elementary but somewhat involved application of calculus. Can be found in
[Spi].
Remark 4.14. The classical Poincare lemma guarantees that the following is a resolution of the constant R sheaf on a differentiable manifold M :
d

0 R 0M 1M . . . nM 0
Where n = dim(M ), and kM denotes the sheaf differential forms on M of order k,
that is kM (U ) = k (U ).
In what follows, we will often consider the groups F(X), which in analogy to the
associated etale space, we call the group of global sections of the sheaf F.

4.3

Sheaf cohomology

We will now look at the problem of global exactness of sequences of sheaves, that
is, for an exact sequence of sheaves 0 A B C 0, when is the induced
sequence
0 A(X) B(X) C(X) 0 exact?
Part of this is always true, as the following easy proposition shows:
34

Proposition 4.15 (Left exactness of global sections). Given an exact sequence of


g
gX
h
sheaves 0 A B C 0, the induced global section sequence 0 A(X)
hX
B(X)
C(X) 0 is exact at A(X) and B(X).
Proof. To check exactness at A(X), it suffices to note that if there were to be a
non-null section s over X in ker gX , then there would be p X with s(p) 6= 0, and
so the germ sp 6= 0 (and sp ker gp by naturality), contradicting the fact that the
first sequence is exact.
Now let b gX (AX ). Then, for all x X, the germ bx = rxX (b) is in ker hx , by
assumption. Since x was arbitrary, this means we can find an open cover {Ui } of
X, such that for each i, rUXi (hX (b)) = hUi (rUXi (b)) = 0, and using the first axiom for
sheaves, necessarily hX (b) = 0.
The reverse inclusion is proved similarly: one only needs to note that injectivity
of the gx makes the choice of local preimages unique, so that they can be glued
together with the second sheaf axiom.
Example 4.16. To see that global sections of sheaves need not be right exact,
consider the following sequence:
i

exp

0 Z O O 0
Where O is the sheaf of holomorphic functions on C {0}, and O is the sheaf
of nonvanishing holomorphic functions on this same set. The map i is the natural
inclusion, and exp : O O is defined by expU (f )(z) = exp 2if (z). Note that the
group structure for sections of O is additive, whereas for sections of O it must be
multiplicative for the above to make sense.
We claim this is an exact sequence of sheaves (that is, at germ level).
To check exactness at O , let gx be a point at some point x, take a representative
g of gx defined on a sufficiently small, simply-connected neighborhood U of x. Then
1
log g)x , for some branch of the logarithm function log, as a
we can choose fx = ( 2i
preimage of gx under exp.
For the middle term, expx (fx ) = 0 (where 0 is the identity element of the group

O ) implies, for a representative f of fx in some connected open neighborhood U of


x,
exp 2if (z) = 1, for z U
Which implies that f is constant, and in fact, an integer. Thus ker(expx ) = Z.
exp
On the other hand, for global sections, the map O(X) O (X) is clearly not
surjective, since there cannot be an holomorphic logarithm function on the whole
C {0}, so the global section sequence 0 Z O(X) O (X) 0 is not exact.
This example is already a hint that the obstructions to exactness for global
sections are often of a topological nature.
It is convenient to extend sections of sheaves to closed sets. This is done through
direct limits, just as for germs. Let F be a sheaf over X, and S X a closed subset.
The we define:
F(S) = lim (U )

SU

35

this can be identified with


From the point of view of the associated etale space F,
1
S = (S). In this line, one can use the notation
the set of continous sections of F|
(S, F). The direct limit construction also gives us restriction maps F(U ) F(S),
just like for stalks.
Definition 4.17 (Soft sheaf). A sheaf F over X is called soft, if for each closed
set S X, the restriction map F(X) F(S) is surjective. In other words, if any
section over a closed set can be extended to a section over the whole space.
Theorem 4.18. If A is a soft sheaf, and
g

0ABC0
is an exact sequence of sheaves, then the induced sequence:
g

X
X
0 A(X)
B(X)
C(X) 0

is also exact.
Proof. Let c C(X). We need to produce b B(X) such that hX (b) = c. Since the
sequence of sheaves is exact, for each x X there is a neighborhood U of x, and
some b B(U ), such that hU (b) = c|U . Since x was arbitrary, it follows that we can
find an open cover {Ui } of X and preimages of the restrictions of c to the open sets
of the cover. The question is if those sections can be glued together to form a global
section, b B(X).
Being X is paracompact, we can find a refinement {Si } of {Ui }, locally finite
and such that each Si is a closed set. Consider the set of all pairs (b, S), where S is
a union of some of the Si and b is such that hS (b) = c|S . This set can be partially
ordered by the relation (b, S) < (b0 , S 0 ) if S S 0 and b0 |S = b. The second axiom
for sheaves implies that in this order, every linearly ordered chain has a maximal
element. Thus, by Zorns lemma, we can find a maximal pair (b, S), such that
hS (B) = c|S .
Suppose that S 6= X. Then there is some Sj {Si } such that Sj 6 S. But
h(bj b) = c c = 0 in Sj S. So, by exactness of the global section sequence at
the term B(X) (proved in the above proposition), there is a A(Sj S) such that
g(a) b bj . Since A is soft, we can extend a to the whole space X. Using the
same notation for the extension, we can define b B(S Sj ) by setting
(
on S
b = b
bj + g(a) on Sj
Then, h(b) = c|SSj , hence S was not maximal and we get a contradiction. This
finishes the proof.
There is another relevant class of sheaves in which we will be interested:
Definition 4.19 (Fine sheaf). A sheaf F over a paracompact topological space X is
fine, if for any locally finite open cover Ui of X, there is a family of sheaf morphisms
i : F F, such that:
36

(a)

i = 1.

(b) i (Fx ) = 0 for x in some neighborhood of X Ui .


We call this family {i } a partition of unity of F subordinate to the cover {Ui }.
Note that this is equivalent to the usual definition for real-valued partitions of
unity on a manifold, for example.
Proposition 4.20. Fine sheaves are soft.
Proof. Let F be a fine sheaf over X, and S X a closed set. Let s F(S). Then,
there is a covering {Ui } of S, each Ui open in X, and sections si F(Ui ) such that
s|SUi = si|SUi .
Now let Uo = X S, s0 = 0, to extend {Ui } to covering of X. Since X is
paracompat, we can assume {Ui } to be locally finite, so there is a partition of unity
{i } subordinate to {Ui }. Then i (si ) is a section on Ui which is identically zero on
a neighborhood of the boundary of Ui , so it may be extended by zero to all of X.
Thus we define:
X
s =
i (si )
i

which is the desired extension of s.


Remark 4.21. Note that on differentiable manifolds, which is the case of interest
to us, most of the natural sheaves are fine, and therefore soft, by the existence of
the usual partitions of unity. This will prove to be a very good situation for our
purposes.
Corollary 4.22. If A and B are soft, and
0ABC0
is exact, then C is soft.
Proof. Let S be a closed set, and s a section of C over S. Restrict the above
sequence to S. By softness of A we can apply the theorem to get exactness of
0 A(S) B(S) C(S) 0. Hence, we can find a preimage for s and extend it
to X by softness of B. Taking its image again will give us a suitable extension of
s.
Proposition 4.23 (Global sections of soft sheaves preserve exactness). If
0 S0 S1 . . . Sn . . .
is an exact sequence of soft sheaves, then the induced global section sequence:
0 S0 (X) S1 (X) . . . Sn (X) . . .
is also exact.

37

Proof. Let Ki = ker(Si Si+1 ) so we can write short exact sequences


0 Ki Si Ki+1 0
For i = 1, K1 = imS0 , and S0 is soft. Thus, by the theorem, we have the exact
sequence
0 K1 (X) S1 (X) K2 (X) 0
By induction, the corollary above shows that Ki is soft for all i, so we obtain exact
section sequences
0 Ki (X) Si (X) Ki+1 (X) 0
which we can splice together to get the desired exact sequence.
We now turn to the definition of the cohomology groups of a sheaf. Here the
construction will be done through a canonical soft resolution of the original sheaft.
There are other equivalent ways to define these groups, for example as right derived
functors of the global section functor, but the one used here is probably the most
elementary.
Definition 4.24 (Sheaf of discontinous sections). Let S be a sheaf over a topological

space X, and S X its associated etale space. We define the presheaf C 0 (S) as:
n
o
C 0 (S)(U ) = f : U S | f = 1U
That is, the same condition as for sections over U , but without requiring continuity.
It is trivial that this is actually a sheaf. Moreover, there is a natural injection
S C 0 (S).
Now we can use this construction to get information about S and X. Let
F 1 (S) = C 0 (S)/S and C 1 (S) = C 0 (F 1 (S)). Now iterate the process by defining:
F i (S) = C i1 (S)/F i1 (S) and C i (S) = C 0 (F i (S)).
By the very definition, we get exact sequences of the form:
0 S C 0 (S) F 1 (S) 0
0 F i (S) C i (S) F i+1 (S) 0
which we can splice together to get the following long exact sequence:
0 S C 0 (S) C 1 (S) . . . C k (S) . . .
Definition 4.25 (Canonical resolution). Given a sheaf S over a topological space
X we call the above sequence the canonical resolution of S. We will also use the
notation
0 S C (S).

38

Now we may take global sections in this canonical resolution, to get a sequence:
0 (X, S) (X, C 0 (S)) (X, C 1 (S)) . . . (X, C k (S)) . . .
which, as has already been seen, need not be exact. It is precisely the possible lack
of exactness of this sequence that gives rise to the sheaf cohomology groups, as seen
in the following definition.
Remark 4.26. Given the cumbersomeness of the above notation, we will denote
C k (S) = (X, C k (S)). The above sequence is then written in the form 0
(X, S) C (S).
Definition 4.27 (Cohomology groups of a space with coefficients in a sheaf). Given
a sheaf S over a topological space X, the cohomology groups of X with coefficients
in S, denoted H k (x, S) are defined as:
H k (X, S) =

ker C k C k+1
im C k1 C k

That is, the cohomology groups of the sequence of global sections of the canonical
resolution.
Note that softness is preserved through the whole construction, so if S is soft,
the global section sequence would be exact everywhere, and the cohomology groups
would vanish.
Theorem 4.28 (Basic properties and functoriality of the cohomology groups). Let
X be a paracompact Hausdorff space. Then the following hold:
(a) For any sheaf S, H 0 (X, S) = S(X). If S is soft, H q (X, S) = 0 for q > 0.
(b) (functoriality) for any sheaf morphism h : A B, there is for each q a group
homomorphism hq : H q (X, A) H q (X, B) with the following properties:
(i) h0 = hX : A(X) B(X).
(ii) hq is the identity map for all q > 0 if h is the identity map
(iii) gq hq = (g h)q for all q 0, for g and h sheaf morphisms.
(c) (exact sequences) For any exact sequence of sheaves
0ABC0
for each q 0 there is a group homomorphism
q : H q (X, C) H q+1 (X, A)
such that
(i) The following induced sequence is exact:
0

0 H 0 (X, A) H 0 (X, B) H 0 (X, C) H 1 (X, A) . . .


q

. . . H q (X, A) H q (X, B) H q (X, C) . . .


39

(ii) A commutative diagram


0

0 A0 B 0 C 0 0
Gives rise to a commutative diagram:
0 H 0 (X, A) H 0 (X, B) H 0 (X, C) H 1 (X, A) . . .

y
y
y
y
y
0 H 0 (X, A0 ) H 0 (X, B 0 ) H 0 (X, C 0 ) H 1 (X, A0 ) . . .
Proof. For part (a)(i), note that the resolution
0 (X, S) C 0 (X, S) C 1 (X, S) . . .
is exact at C 0 (X, S), so (X, S) = ker(C 0 (X, S) C 1 (X, S)) = H 0 (X, S).
Part (a)(ii) follows easily from proposition 4.23.
For (b) and (c) we will first show that a map h : A B induces a natural cochain
map h : C(A) C(B), where C(A) = C(X, A) and C(B) = C(X, B). First we
define a map
h0 : C 0 (A) C 0 (B)
by letting h0 (sx ) = (hs)x , for s a discontinous section of A. This induces a quotient
map:
0 : C 0 (A)/A C 0 (B)/B
h
given that C 0 (A)/A = F 1 (A) and C 0 (B)/B = F 1 (B), repeating the above construction gives us a map:
h1 : C 0 (F 1 (A)) C 0 (F 1 (B))
1 , and keeping on in this way we get, for
wich again induces a map of quotients h
each q 0, a map
hq : C q (A) C q (B)
so the induced section maps give us the desired map h . Given the way it was
constructed, this map is clearly functorial.
Now if
0ABC0
is exact, then
0 C (A) C (B) C (C) 0
is an exact sequence of complexes of sheaves. But these sheaves are all soft, so the
induced global section maps make
0 C (A) C (B) C (C) 0
an exact sequence of cochain complexes of abelian groups. We can then derive
(via the zig-zag lemma of basic homological algebra) a long exact sequence for the
corresponding cohomology groups:
q

. . . H q (C (A)) H q (C (B)) H q (C (C)) H q+1 (C (A)) . . .


40

where q is constructed in the usual way. Now one just needs to note that the
construction is natural to get parts (b) and (c).
Definition 4.29 (Acyclic resolution). A resolution of a sheaf S over a space X,
of the form 0 S A is called acyclic, if for q > 0 and p 0, all the groups
H q (X, Ap ) vanish.
The following theorem is the core result for our purposes, and a very powerful
tool in calculation of (co)homology.
Theorem 4.30 (Abstract de Rham theorem). Let S be a sheaf over a topological space X, and let 0 S A be a resolution of S. Then there are natural
homomorphisms
p : H p ((X, A )) H p (X, S)
where H p ((X, A )) are the cohomology groups of the cochain complex (X, A ),
where the maps are the global section maps derived from the resolution. Also, if the
resolution is acyclic, the maps p are isomorphisms.
Proof. Let Kp = ker(Ap Ap+1 ) = im(Ap1 Ap ) (note that these are sheaves,
not global sections, so equality holds). This way K0 = S. We have short exact
sequences:
0 Kp1 Ap Kp 0
which induce, by the properties of the cohomology groups, exact sequences of the
form:
0 (X, Kp1 ) (X, Ap ) (X, Kp ) H 1 (X, Kp1 ) H 1 (X, Ap ) . . .
Note that, by definition of the Kp , we have (X, Kp )
= ker((X, Ap ) (X, Ap+1 ),
and in turn
H p ((X, A ))
= (X, Kp )/ im((X, Ap1 ) (X, Kp ))
So we can define, through the exact sequence above, a map 1p : H p ((X, A ))
H 1 (X, Kp1 ) which is, in fact, injective. Moreover, if the resolution is acyclic,
H 1 (X, Ap1 ) = 0 and, again by the long exact sequence, 1p is an isomorphism.
Corollary 4.31. If we have an homomorphism
form:
0 S

f
y
y

of resolutions of sheaves, of the


A

g
y

0 J B
then there is an induced homomorphism
gp

H p ((X, A )) H p ((X, B ))
which if f is an isomorphism of sheaves and the resolution are acyclic, is an isomorphism.
This means that we can calculate sheaf cohomology of a space X with coefficientes in a sheaf S using any acyclic resolution of S, in particular any soft resolution.
41

4.4

de Rham theorem

We now use the last result of the previous section, abstract de Rham theorem, to
relate different notions of cohomology of a manifold, through soft resolutions and
sheaf cohomology. We will start by formulating singular cohomology in terms of
sheaves.
Definition 4.32 (Sheaves of singular cochains). Let M be a differentiable manifold,
and for each open set U M , consider the group of singular cochains defined in U
with real coefficients S p (U, R) = HomZ (Sp (U, Z), R), where (Sp (U, Z) is the usual
group of singular chains of degree p with integer coefficients. Then the assignment
U S p (U, R) is easily seen to be a presheaf. We denote the sheaf generated by this
presheaf by S p (M, R).
Now note that the sequence of sheaves

0 G S 0 (M, R) S 1 (M, R) . . .
is exact, since singular cohomology for an euclidean ball is zero for p > 0, similarly
to the Poincare lemma mentioned above.
Remark 4.33. Its not hard to see that in our setting, that is, smooth manifolds, we
can use C cochains and get the same results above. This is what we will use from
p
(M, R)
now on, Denoting these sheaves S
Lemma 4.34 (Sheaves of modules over soft sheaves of rings are soft). If M is a
sheaf of modules over a soft sheaf of rings R, then M is itself soft.
Proof. Let s (K, M), where K X is a closed subset. Then, by the direct
limit definition of such sections, s extends to an open set U K. Now define
(K (X U ), R) be defined by:
(
1 on K
=
0 on X U
Then, since R is soft, we can extend to a section over the whole of X, and keeping
the same notation for the extension, s is the desired extension of s.
Which inmediately gives us what we were looking for, taking into account that
the constant sheaf R is obviously soft:
Corollary 4.35 (Softness of the singular cochain sheaf). Let M be a differentable
q
manifold. Then the sheaves of C singular cochains, S
(M, R), are soft.

Remark 4.36. One can define an homomorphism I : M S


(M, R), and, in
turn an homomorphism of the corresponding resolutions of the sheaf R, By integration of differential forms over smooth chains. That is, by defining IU : M (U )

S
(M, R)(U ) by the equation:
Z
IU ()(c) =
c

where c is a C singular chain, so IU () is a C singular cochain on U. Moreover,


Stokes theorem implies that I commutes with the corresponding differentials.
42

We have already seen that the differential forms on a differentiable manifold M


form a soft resolution of the constant sheaf R. Therefore, we have obtained the
following well-known theorem.
Theorem 4.37 (de Rham theorem). Let M be a differentiable manifold. Then the

(M, R)) induced by integration of differennatural maps I : H p (M (M )) H p (S

tial forms over C singular cochains with real coefficients are all isomorphisms.
Proof. That the resolutions induce isomorphisms of their cohomology groups with
the cohomology groups of M with coefficients in R follows from the abstract de
Rham theorem 4.30 above, and the fact that the composite isomorphism is induced
by the integration maps follows from corollary 4.31.

4.5

Currents and homology

Now we turn our attention to a certain kind of duals of the differential form groups.
By analogy to singular homology and cohomology, and looking at the de Rham
theorem above, one would expect this new groups to be useful in calculating singular
homology of the underlying manifold. In fact, even more is true, as we shall see
below.
Definition 4.38 (Currents). Let M be a smooth manifold and let kc (M ) be the
space of differentiable k-forms with compact support on it. A k-current is a linear
functional on kc (M ), which is continous in the sense of distributions. The space of
k-currents on M will be denoted Dk (M ), endowed with the weak-* topology.
Example 4.39. Let M be a differentiable manifold, m = dim(M ), k (M ).
Then we can associate to a m k-current defined by
Z
T () =

and the map T is injective.


Since it will not be needed for our purposes, we wont get into analytical details
about currents. As expected, one can form a presheaf of currents by taking currents
defined on open sets U M (note that the compact support requirement still
applies), that is, the assignment U Dk (U ). It is not hard to check that this is
also a sheaf, by standard arguments. This sheaf will be denoted DkM .
Now, analogously to the differential form and singular cases, one considers the maps:

n1

n
M
DnM
Dn1
. . . D0M 0

(where n = dim(M )) dual to the exterior derivative of differential forms, that is,
defined by k T () = T (d). Thus, one gets immediately k k1 = 0, that is, the
above is a differential complex of sheaves.
Definition 4.40 (Pushforward of currents). Let M1 , M2 be differentiable manifolds
of dimensions m1 and m2 respectively, and F : M1 M2 a smooth map. Then
there is a pullback morphism of forms, F : kc (M2 ) k (M1 ) which is continous
43

in the C topology, but for kc (M2 ), Supp F F 1 (Supp ) need not be


compact.
If T Dk (M1 ) is such that for every compact subset K of M2 , Supp T F 1 (K)
is compact, then T (F ) is well defined and continuous on kc (M2 ), so there is
a unique current F T defined by the equation
F T () = T (F )
which we call the direct image or pushforward of T by F .
We will need the following easy properties:
Proposition 4.41 (Properties of pushforwards of currents). Let M1 , M2 , M3 be difF
G
ferentiable manifolds, and M1 M2 M3 smooth maps. Then, for suitable currents in the sense of the previous definition, the following hold:
(a) (F T ) = F T
(b) (G F ) T = G (F T )
Proposition 4.42 (Poincare lemma for currents). Let U Rm be star-shaped and
open, and T Dq (U ) be such that q T = 0. Then there is a (q + 1)-current
S Dq+1 (U ) such that q+1 S = T .
Proof. The trick is to reduce it to the standard Poincare lemma for differential forms,
by standard distribution techniques (in essence, if a current is regular enough it can
be represented by a form, just like for functions and distributions). Details can be
found in [De], albeit with a somewhat different notation.
Remark 4.43. The above proposition shows that one can form a resolution of the
constant sheaf R of the form:
n1

n
M
0 R DnM
Dn1
. . . D0M 0

Proposition 4.44 (Resolution with currents is acyclic). Let M be a compact smooth


manifold. Then the resolution above is acyclic.
Proof. It suffices to prove that each DkM is soft. But partitions of unity on M ensure
that these sheaves are fine, hence soft.
We have finally arrived to the main theorem we wanted to prove in this section:
Theorem 4.45 (Singular homology from currents). Let M be a compact smooth
orientable manifold. Then the singular homology groups of M are isomorphic to the
homology groups of the complex

n1

n
M
DnM (M )
Dn1
(M ) . . . D0M (M ) 0

of global sections of the sheaves of currents on M . That is, Hk (M )


= Hk (DM (M )).

44

Proof. One just needs to note that


n1

n
M
. . . D0M 0
0 R DnM
Dn1

is an acyclic resolution of the constant sheaf R, so that the complex in the statement
of the theorem calculates singular cohomology, by the abstract de Rham theorem
4.30, and then apply Poincare duality.
A very geometrical explanation and proof of Poincare duality can be found in
the last chapter of [Mu]. A quick introduction to currents can be found on [De], and
further material in [dR], for example.

45

The Morse homology theorem

In this section we put together the material from the previous ones, and prove the
theorem we have been aiming for, that is, that the singular homology groups of a
smooth manifold are isomorphic to the homology groups of a certain chain complex,
constructed with the free abelian groups with basis the critical points of each index.
The analytic constructions in section 3 already pointed out how the boundary maps
should be constructed. This construction is formalized below.
Remark 5.1. Unless otherwise noted, in this section all differentiable manifolds will
be regarded as compact and orientable.

5.1

The Morse chain complex

Definition 5.2. Let (N, g) a riemannian manifold, f : N R a Morse function,


and Criti (f ) the
L set of index i critical points of f . Then we define the groups
CiMorse (f, g) = pCriti (f ) Z, and the corresponding differential Morse given by the
following equation:
X
Morse (p) =
#M(p, q) q
qCriti1 (f )

where #M(p, q) denotes the (signed according to orientation) number of components


\q), which must be finite
of M(p, q), or alternatively, the number of points in M(p,
since ind p ind q = 1.
Now, all the work in section 3 already starts to be fruitful.
Proposition 5.3. The groups and maps defined above form a chain complex. That
is, ( Morse )2 = 0.
Proof. This is an easy consequence of theorem 3.1. One just needs to compute:
X
h( Morse )2 p, qi =
h Morse p, rih Morse r, qi =
rCriti1 (f )

=#

M(p, r) M(r, q)

rCriti1 (f )

= #M(p, q)
= 0.
Where the second equality is by definition, the third by theorem 3.1, and the fourth
because the boundary of a compact manifold of dimension 1 will always have zero
points, counted with sign.
In analogy with the notation above, we will denote the homology groups of this
complex HkM (f, g).

46

5.2

The chain homotopy. Morse homology theorem

The goal of this part (and the essence of the whole text) is to prove the following
theorem:
Theorem 5.4 (The Morse homology theorem). Let (N, g) a compact riemannian
manifold and f : M R a Morse function, such that (f, g) is a Morse-smale pair.
Then there is a canonical isomorphism
HkM (f, g)
= Hk (N, Z)
where Hk (N, Z) denotes the k-th singular homology group of N with coefficients in
Z.
The proof will consist in producing a chain homotopy between the two chain
complexes, which will in turn induce homology isomorphisms. The forward map
will consist in taking a critical point to its descending (unstable) manifold, and the
backwards one in flowing a simplex with the gradient flow and taking it to the sum
of the critical points it reaches at . To make this all rigorous we will need to
define some concepts and prove a number of things about them first.
Definition 5.5. An i-simplex : i N is said to be generic, if is smooth,
and each face of is transverse to the ascending manifolds of all the critical points
of f . We define the groups Ci (N ) to be the Z-module with basis consisting on the
i-currents on N generated by generic i-simplices.
Remark 5.6. What we mean above by equating simplices and currents, is the inclusion j : Sk (N, R) Dk (N ) defined by
Z
j()() =

For a singular simplex Sk (N, Z) and then extending to chains, where Sk (N, R)
is the usual singular chain group with R coefficients, and k (N ). Often we will
identify and j().
Remark 5.7. Analogously to the above, one can consider, for an oriented submanifold
L N of dimension k, the k-current [L] defined by
Z
[L]() =

Proposition 5.8. (a) In the groups Ck (N ), the standard simplex boundary defines
the same maps as the current boundaries k : Dk Dk1 .
(b) Considering the chain complex (C (N ), ), there is a canonical isomorphism
of its homology groups with the singular homology groups of N
Hk (C (N ))
= Hk (N, Z)

47

Proof. Part (a) is a direct consequence of Stokes theorem.


For part (b), one could consider the groups C k (U ) = hom(Ck (U ), Z) (where the
transversality condition is kept by restricting f to U ) and follow the same steps
used to prove the de Rham theorem from the abstract de Rham theorem. That

was done with the sheaves S


(M, R).
In this case, one would start defining presheaves U C k (U ), then taking the
sheaves generated from those presheaves (the transversality condition is a local one
so it does not interfere), and considering the corresponding resolution of the constant
sheaf Z. It will indeed be a resolution, given that exactness wont be affected by
the small perturbations that might be needed to fulfill the transversality conditions
(we wont give further details here). Softness of these sheaves is obvious by their
definition. One would then have H k (C (N ))
= H k (N, Z). This in turn implies
Hk (C (N ))
= Hk (N, Z).
For our purposes, we will need the following generalized version of theorem 3.1.
We dont give the details, but it can be proved analogously to the latter.
Theorem 5.9 (Compactification of descending manifolds). For a closed smooth
manifold N , (f, g) a Morse-Smale pair on N , and a critical point p M , the
descending manifold W u (p) has a natural compactification to a smooth manifold
with corners W u (p), whose codimension k stratum is
[
W u (p)k =
M(p, q1 ) M(q1 , q2 ) . . . W u (qk )
q1 ,...,qk Crit(f )

with r1 , . . . , rk , q all different. For the case k = 1, as oriented manifolds, we have:


[
W u (p) =
(1)ind p+ind q+1 M(p, q) W u (q)
p6=q Crit(f )

and the maps W u (p)k N given by projecting to W u (qk ) patch together to a smooth
map:
e : W u (p) N
extending the inclusion W u (p) N
Definition 5.10 (D map). We define the map D : CkMorse (f, g) Ck (N ), by letting,
for p Critk (f ), D(p) = e [W u (p)], where e is the embedding map above, so e is
the corresponding push-forward of currents.
Remark 5.11. Note that when using currents, signs are implicitly included, since
integration is done on oriented submanifolds.
Proposition 5.12. D is a chain map, that is, D = D Morse
Proof. Let p Criti (f ). By the last compactification theorem, we have:
[
W u (p) =
(1)ind(p)+ind(q)+1 M(p, q) W u (q)
qCrit f

q6=p

48

Therefore
D(p) =

(1)ind(p)+ind(q)+1 e [M(p, q) W u (q)] Ci1 (N )

qCrit f

q6=p

But if ind(q) > i 1, being f Morse-Smale, M(p, q) is empty, and if ind(q) < i 1,
then the contribution to the right hand side is zero in Ci1 (N ), because e would
map M(p, q) D(q) top the support of D(q), which is a current of dimension less
than i 1. We then have:
X
D(p) =
#M(p, q) e [W u (q)]
qCriti1 f

For the inverse map, we must consider flow lines originating from a given simplex,
so one defines


M(, q) = x N | lim x (t) = q and for some t0 , x (t0 )
t+

for a simplex and a critical point q. Note that the support of the simplex has
been denoted in the same way, to simplify notation. Notice that by the Morse-Smale
condition there is an isomorphism, for any such x
Tx (t0 )
= T M(, q) Tq W u (q)
and so we give an orientation to T M(, q) in such a way that this isomorphism is
orientation-preserving. We require yet another compactification theorem for these
sets, which again can be proved in the same way as the previous ones.
Theorem 5.13 (Compactification of M(, q)). With the notation above, M(p, q)
has a natural compactification to a smooth manifold with corners M(, q), with
codimension k stratum
M(, q)k =

k
[
j=0

M(kj , p1 ) M(p1 , p2 ) . . . M(pj1 , pj ) M(pj , q)

p1 ,...,pk Crit(f )

p1 ,...,pk ,q distinct

where j denotes the codimension j stratum of . When k = 1, as oriented manifolds


we have:
[
M(, q) = M(, q)
(1)i+ind(q) M(, p) M(p, q)
pCrit(f )

p6=q

Taking into accout the above theorem, and that dim M(, q) = i ind(p), one
can define a map A : C (N ) CMorse (f, g) through the formula
X
A() =
#M(, p) p
pCriti (f )

49

Proposition 5.14 (A is a chain map). A = Morse A


Proof. Like before, this follows from the above compactness result.
Let q Criti1 (f ). Then:
[
#M(, q) = #M(, q) #
M(, p) M(p, q)
q6=pCrit(f )

= #M(, q) #

M(, p) M(p, q)

qCriti f

= hA(), qi h

Morse

A(, q)i

Lemma 5.15. A D = id : CiMorse CiMorse


Proof. If p Criti (f ), then M(D(p), p) contains one point, the constant flow line,
oriented positively. On the other hand M(D(p), q) is empty if q is any other index i
critical point, seeing the definition of D(p), and by the Morse-Smale condition.
We will need a last compactification theorem:
Theorem 5.16 (Compactification of forward orbits). For a generic i-simplex, we
define its forward orbit to be the set
F() = [0, +)
together with the map e : F() N defined by
e(s, x) = s ((x))
This forward orbit has a natural compactification to a smooth manifold with corners
F(), whose codimension k stratum, for k > 2, is
F()k = F(k )

k
[
j=0

M(kj , r1 )M(r1 , r2 ). . .M(rj1 , rj )W u (rj )

r1 ,...,rk Crit(f )

r1 ,...,rk distinct

For k = 1, as oriented manifold one has


[

F() = F()

M(, r) W u (r)

rCrit(f )

And the map e extends to this compatification as a smooth map which projects to
W u (rj ) N .
And this theorem enables us to get what we were looking for.
Proof of the Morse homology theorem. Define F : Ci (N ) Ci+1 (N ) by
F () = e [F()]
Then the above compactification result implies that F is a chain homotopy between
the identity and D A,
F + F = D A id
and by basic homological algebra, we are done.
50

5.3

The Morse inequalities

We now present some applications of the Morse homology theorem, most notably
the so-called Morse inequalities. It is worth noting that most or all of these results
could be proved more directly without needing the full power of the Morse homology
theorem, but here we will deduce them from it.
First, we will settle some notation. Let N be a compact smooth manifold and
let bk denote the k-th Betti number of N , that is, the rank of the finitely generated Z-module Hk (N, Z). Now, for a Morse function f : N R, let k (f ) =
Card(Critk (f )). Note that for chain complexes and boundaries we will use the same
symbols as in the previous part, but with different meanings. Hopefully the meaning
will be clear.
An obvious consequence of the Morse homology theorem is the inequality k (f )
bk , since the Morse complex with the critical points as generators has the same
homology. These are called the weak Morse inequalities. Somewhat more interesting
are the following stronger inequalities:
Theorem 5.17 (The Morse inequalities). With the above notation,
P
Pn
k+n
k (f ) nk=0 (1)k+n bk
(a)
k=0 (1)
Pn
Pn
k
k
(b)
k=0 (1) bk
k=0 (1) k (f ) =
where n = dim(N ).
To prove this, we will need the following:
Theorem 5.18 (Euler-Poincare theorem). Let (C , ) be a finitely generated chain
complex, for which Ck = 0 if k > n, for a certain n > 0. Let ck = Rank(Ck ),
bk = Rank(Hk (C )), for k = 0, . . . , n. Then the following equation holds:
m
m
X
X
k
(1)k bk
(1) ck =
k=0

k=0

Proof. The exactness of

k
0 ker k Ck
im k 0

shows that ck = Rank(ker k ) + Rank(im k ) for each k. Similarly,


0 im k+1 ker k Hk (C ) 0
being exact gives us Rank(ker k ) = Rank(im k+1 ) + bk , hence
Rank(ker k ) = ck Rank(im k ) = Rank(im k+1 ) + bk
which in turn implies:
m
m
X
X
k
(1) (ck Rank(im k )) =
(1)k (Rank(im k+1 ) + bk )
k=0

k=0

which in turn implies


m
m
X
X
k
(1) ck =
(1)k bk
k=0

k=0

51

Proof (of the Morse inequalities). Since by the Morse homology theorem, the Morse
complex has Hk (N, Z) as homology, we can apply the above theorem to immediately
get the second part of the theorem. For the first part, let m < n and consider the
(m)
truncated Morse chain complexes (C , k ) given by:
(
CkM (f, g) if k m
(m)
Ck =
0
if k > m
where a Morse function f and suitable Riemannian metric g are assumed. Then,
applying the Euler-Poincare theorem we have
m

(1)

m
X

(1)

(m)
Rank(Ck )

= (1)

k=0

m
X

(1)k Rank(Hk (C(m) ))

k=0
(m)

(m)

Now, since k (f ) = Rank(Ck ) for k m, bk = Rank(Hk (C )) for k m 1,


(m)
and Hm (CM (f, g)) is a quotient of Hm (C ), we end up with
m m1 + . . . + (1)m 0 bm bm1 + . . . + (1)m b0
as desired.
An easy application of the Morse inequalities is the following proposition:
Proposition 5.19. The Euler characteristic (N ) of a compact smooth manifold
N of odd dimension n is zero.
Proof. Let f : N R be a Morse function (which we know always exist). Trivially,
k (f ) = nk (f ). Then we have the following:
n
n
X
X
k
(N ) =
(1) k (f ) =
(1)k nk (f ) =
k=0

= (1)n

k=0
n
X

n
X

k=0

k=0

(1)nk nk (f ) = (1)n

And since n is odd, (N ) = 0.

52

k (f ) = (1)n (N )

Generalizations

Here we give some ideas (without any sort of proofs) about some generalizations of
the theory presented.

6.1

Morse-Bott theory

The fact that a small perturbation will make any smooth function Morse seen in
section 2 sounds good in principle. But in an specific situation, there is no obvious
way of choosing a Morse function for a given manifold, generic as they may be.
Especially since choosing a function with obvious symmetries will often work against
us, like the height function of a lying torus, which has critical circles, instead of
isolated points.
One way to go around this is to extend the theory to Morse-Bott functions. A
Morse-Bott function is a smooth function M R such that Crit(f ) is a union of
submanifolds, and everywhere on those submanifolds the kernel of the Hessian is
exactly the tangent space to such critical submanifolds.
The theory then follows in much the same lines, except that one has to consider
more complex moduli spaces (informally, one has to flow down bigger subsets).
Using appropiate metrics and defining a similar chain complex with the critical
submanifolds, and taking extra care of the indices involved, the homology of M is
recovered, just like in the Morse function case.
This theory is presented, for example, in [La], with a method based on currents
and spectral sequences, rather than analysis. Some of the development can also be
found in [Hu], using the familiar compactification approach.

6.2

Novikov homology

Another way for generalization is to note that after the first definitions, everything
is based on the closed 1-form df , rather than on the actual function f . This suggests
replacing df by a different closed 1-form , and trying to repeat the process. This,
actually, turns out to be very fruitful.
The process of constructing such an homology is esentially the same, one defines
Morse forms to be the ones with nondegenerate critical points (that is, zeroes) and
an analog of the gradient for such a form raising the index of through a Riemannian
metric g. The compacntess theorems can be then proved with very similar arguments
to the Morse function case, the main difference being that one has to mod out the
(de Rham) cohomology class of throughout. Once one obtains finite counts of the
relevant flow lines, the construction of the chain complex is essentially the same.
Now, the interesting part is that, unlike in the original case of a real Morse function, the forms used might have non-zero homology classes. This can be most easily
seen if one starts with functions f : M S 1 , in what is called circle-valued Morse
homology. To understand what happens in these cases, it is useful to remember the
Galois correspondence of covers of M and subgroups of its fundamental group, and
the relation between the fundamental group and H1 (M, Z).
M in which
In the end, what one gets is the homology of a covering space M
the lift of is exact, with coefficients in a certain (Novikov) ring defined based on
53

the algebraic properties of the class [] H 1 (M, Z). This is useful in a number of
situations, incluiding the study of periodic orbits of the gradient flow, which in the
case of circle-valued Morse functions, are not excluded.
A good reference for this theory and many of its consequences is [Paj].

54

References
[BaHu] Banyaga, Agustin; Hurtubise, David. Lectures On Morse Homology, Kluwer
Texts in the Mathematical Sciences Vol. 29, Kluwer Acad. Publisher, 2004.
[BoTu] Bott, Raoul; Tu, Loring W. Differential forms in algebraic topology, Graduate Texts in Mathematics, 82. Springer-Verlag 1982.
[Br] Bredon, Glen E. Topology and geometry, Graduate Texts in Mathematics 139,
Springer-Verlag 1993.
[Br2] Bredon, Glen E. Sheaf theory, Graduate Texts in Mathematics 170, 2ed.,
Springer-Verlag 1997.
[De] Demailly, J.P Complex analytic and differential geometry, Unpublished,
http://www-fourier.ujf-grenoble.fr/demailly/manuscripts/agbook.pdf.
[dR] de Rham, Georges. Differentiable manifolds, Springer-Verlag 1984.
[Flo] Floer, A. The unrgularized gradient flow of the symplectic action, Comm. Pure
Appl. Math. 41 (1988), 393-407.
[GuiPo] Guillemin V., Pollack A. Differential Topology, Prentice Hall, 1974.
[Hu] Hutchings, Michael. Lecture notes on Morse homology (with an eye towards
Floer theory and pseudoholomorphic curves),
http://math.berkeley.edu/hutching/teach/276/mfp.ps
[La] Latschev, Janko. Gradient flows of Morse-Bott functions, Math. Ann. 318
(2000), no. 4, 731-759.
[Lan] Lang, S. Differential and Riemannian Manifolds, Graduate Texts in Mathematics, 160. Springer-Verlag 1995.
[Mi] Milnor, John. Morse Theory, Princeton University Press 1963.
[Mu] Munkres, James R. Elements of algebraic topology, Perseus 1984.
[Paj] Pajitnov, Andrei V. Circle-valued Morse theory, de Gruyter 2006.
[Sch] Schwarz, Matthias. Morse Homology, Birkhauser 1993.
[Sm] Smale, Stephen. An infinite dimensional version of Sards theorem, American
Journal of Mathematics, Vol. 87, No. 4 (Oct., 1965), pp. 861-866
[Spi] Spivak, M. Calculus on manifolds, Addison-Wesley 1965.
[We] Wells, Raymond O. Jr. Differential Analysis on Complex Manifolds, Graduate
Texts in Mathematics. 3ed., Springer-Verlag 2008.

55

También podría gustarte