Morse theory

Shintaro Fushida-Hardy
381B Sloan Hall, Stanford University, CA

This document contains Morse theory notes, largely following Audin and Damian [AD14].
The focus is on developing Morse homology and exploring some applications (such as
the Morse inequalities). Some solutions to exercises are also given here. At the end of
these notes we give a proof outline of the h-cobordism theorem (and prove the generalised
Poincaré conjecture) following Milnor’s lecture notes [Mil65]. Finally we explore the status
of the generalised Poincaré conjecture and h-cobordism theorem (for each dimension) in
several categories of manifolds (Man, Man∞ , etc).
Contents

1 Introduction to Morse theory 2

1.1 Introduction to the introduction . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Morse functions: existence and genericness . . . . . . . . . . . . . . . . . . . 3
1.3 The Morse lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Examples and exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Pseudo-gradients, topology, and the Smale condition 9

2.1 Existence of pseudo-gradients . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Trajectories and stable/unstable manifolds . . . . . . . . . . . . . . . . . . . 11
2.3 Critical values topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Smale condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Morse homology fundamentals 16

3.1 Morse homology modulo 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Integral Morse homology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Well-definedness of the Morse complex . . . . . . . . . . . . . . . . . . . . . 23
3.4 Morse-Smale pair invariance of the Morse homology . . . . . . . . . . . . . 26
3.5 Morse homology is singular homology . . . . . . . . . . . . . . . . . . . . . 28

4 Morse homology applications 31

4.1 The Morse inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Morse functions and simple connectedness . . . . . . . . . . . . . . . . . . . 33
4.3 Poincaré duality realised in Morse homology . . . . . . . . . . . . . . . . . . 34

5 The h-cobordism theorem 38

5.1 Smale’s original proof of the generalised Poincaré conjecture . . . . . . . . . 38
5.2 Proving the Poincaré conjecture from the h-cobordism theorem . . . . . . . 39
5.3 Proof outline of the h-cobordism theorem . . . . . . . . . . . . . . . . . . . 42
5.4 Poincaré conjecture and h-cobordism theorem in different categories . . . . 45

1
Chapter 1

Introduction to Morse theory

1.1 Introduction to the introduction

The fundamental idea in Morse theory is the following:

A well chosen map f : M → R encodes a lot of information about M .

For example, consider the “height function” h : T2 → R of a torus as depicted in figure

1.1:

3
4

1
4

Figure 1.1: A torus with its height function next to it.

This height function has exactly four critical points, i.e. points where dh vanishes. We
find that these critical points correspond to changes in the topology of the level sets of the
function. Precisely, the level starts empty for negative values of a. Then at a = 0, there
is a bifurcation and the level set is a point. As a continues to increase, the level set is a

2
circle, until we reach a = 41 . Again there is a bifurcation, and the level set at this point is
a figure 8. As we continue to increase a, the level set is now two disjoint circles and so on.
We find that the changes in topology of the level sets occurs precisely at the critical points
of h. On the other hand, when a is not a critical point, the submanifold theorem ensures
that the level set h−1 (a) is a submanifold of the torus. This agrees with our observations
above.
The goal of Morse theory is to find invariants of manifolds by counting critical points
of well chosen functions. The notion of a “well chosen function” is formalised to mean a
Morse function.

Definition 1.1.1. A map f : M → R is a Morse function if its critical points are non-
degenerate. That is, if the Hessian of f at each critical point is non-singular.

A motivation for the existence of useful invariants of manifolds arising from Morse
functions is Reeb’s theorem.

Theorem 1.1.2 (Reeb’s theorem). Let M be a compact manifold. Suppose there exists
a Morse function on M with exactly two critical points. Then M is homeomorphic to a
sphere.

This theorem shows that a “choice” of Morse function can give results about the under-
lying space that are independent of the choice of Morse function. Eventually we generalise
this idea and develop Morse homology. This is a homology theory constructed by count-
ing critical points of Morse functions, which we show depends only on the diffeomorphism
class of the manifold. The first section of these notes will culminate in the famous Morse
inequalities.

1.2 Morse functions: existence and genericness

Definition 1.2.1. Let M be a smooth manifold, and f : M → R a smooth map. Then
any x ∈ M such that dfx = 0 is a critical point. If M is compact, smooth functions always
have critical points since they must attain their maxima and minima.

While first derivatives exist, second derivatives (Hessians) do not exist on smooth man-
ifolds in general. However, they are well defined on critical points.

Definition 1.2.2. Let x ∈ M , and f : M → R. Suppose dfx = 0. The Hessian at x is

defined by
d2 fx (X, Y ) = (X(Ye f ))(x),
where X, Y are tangent vectors in Tx M , and Ye is any local extension of Y .

3
 We must verify that the Hessian is a well defined symmetric bilinear form. Suppose Yb
is any other extension of Y , and let X
b be an extension of X. Then

(X(Ye f ))(x) − (Y (Xf

e ))(x) = [X,
e Ye ]x f = dfx ([X,
e Ye ]x ).

But the last term vanishes since dfx = 0 by assumption. Moreover, this calculation shows
that the map is well defined, since

(X(Ye f ))(x) = (Y (Xf

e ))(x) = (X(Yb f ))(x).

A second approach to defining the Hessian is to use local charts as in the exercise 1 of
Audin-Damian:

Exercise 1.2.3. (A-D, exercise 1) Let U be an open subset of Rn and let f : U → R be

smooth. Let V be another open subset of Rn , and ϕ : V → U a diffeomorphism. Compute
d2 (f ◦ ϕ)y for y ∈ V . Let M be a manifold and g : M → R a function. Show that (d2 g)x
is well defined on ker(dg)x ⊂ Tx M .

Solution: Here (d2 f )x denotes the usual Hessian of f at x, defined by

∂2f i j
(Hfx )(u, v) = ((d2 f )x )ij ui v j = uv .
∂xi ∂xj

Since ϕ is a diffeomorphism, it can be expressed as a smooth change of coordinates

(y1 , . . . , yn ) 7→ (ϕ1 (y1 , . . . , yn ), . . . , ϕn (y1 , . . . , yn )) = (x1 , . . . , xn ). Then the Hessian of
f ◦ ϕ is given by

∂ 2 (f ◦ ϕ) i j
(H(f ◦ ϕ)y )(u, v) = uv
∂yi ∂yj
 ∂x ∂  ∂x ∂f 
k l
= ui v j
∂yi ∂xk ∂yj ∂xl
 ∂x ∂  ∂x  ∂f 
k l ∂ 2 f  ∂xk i  ∂xl j 
= ui v j + u v .
∂yi ∂xk ∂yj ∂xl ∂xk ∂xl ∂yi ∂yj

In fact, this calculation shows that the Hessian of g : M → R at x is well defined on the
kernel of dgx , since that is where the first term in the above formula vanishes. (Observe
that ϕ corresponds to the choice of local chart on a manifold, and the second term is
chart-invariant.) 4

Some examples and non-examples of Morse functions are explored in the exercises at
the end of this section. We next prove that morse functions exist, and in fact, there are
many of them!

4
Proposition 1.2.4. Let M ⊂ Rn be a submanifold. For almost any p ∈ Rn , the function

fp : M → R, x 7→ kx − pk2

is a Morse function.

Remark. By the Whitney embedding theorem, it follows that Morse functions exist on
all smooth manifolds.

Proof. Let fp be as above. The derivative of fp is given by

dfp,x (v) = 2(x − p, v).

Therefore the critical points occur exactly when Tx M is normal to x − p. (Such a p

can always be found if n > dim M , so critical points exist.) Choose local coordinates
(u1 , . . . , ud ) for M , so that

∂fp ∂x ∂ 2 fp  ∂x ∂x ∂2x 
= 2(x − p) · , =2 + (x − p) · .
∂ui ∂ui ∂ui ∂uj ∂ui ∂uj ∂ui ∂uj

Therefore (by definition) x is a non-degenerate critical point if and only if x − p is normal

to Tx M , and the matrix on the right is non-degenerate, i.e. has rank d. Recall that Sard’s
theorem states that the set of critical points of a map g : M → N has measure zero in N ,
where a critical point is any x ∈ M such that dgx does not have maximal rank (i.e. rank
equal to min{dim M, dim N }). Therefore by Sard’s theorem, it suffices to show that the
p ∈ Rn such that x − p is normal to Tx M and the matrix on the right is singular, are the
critical points of a smooth map.
To this end, consider the normal bundle of M in Rn ,

N M = {(x, v) ∈ Tx Rn : v ∈ Tx M ⊥ }.

Define the map E : N M → Rn by E(x, v) = x + v. It can then be verified that p = x + v ∈

Rn is a critical point of E if and only if
 ∂x ∂x ∂2x 
2 +v·
∂ui ∂uj ∂ui ∂uj

is singular. Therefore the set of all fp (with p varying in Rn ) which are not Morse functions
corresponds to a subset of the critical points of E, which by Sard’s theorem, has measure
zero in Rn . Thus for almost all p, fp is Morse.

This shows that manifolds have many Morse functions. However, it is not immediate
that the Morse functions are generic in the sense that any function is approximated by a
Morse function. This turns out to be the case too!

5
Proposition 1.2.5. Let M be a manifold, and f : M → R smooth. Let k ∈ N. Then on
any compact subset of M , f can be approximated by a Morse function in C k -norm.

Proof. This follows from the previous proposition, with details given in A-D. The idea is
to choose an embedding into Rn , and then use the previous proposition explicitly (that is,
the proof makes use of the functions fp ).

An alternative but similar result is the following, which relies on transversality results
and no embedding.

Proposition 1.2.6. Let M be a compact manifold. Then the set of Morse functions on
M is a dense open subset of C ∞ (M ).

1.3 The Morse lemma

We know from Taylor’s theorem that f near a critical point is approximated by its second
derivative in the sense that
1
f (x) ≈ f (c) + (d2 f )c (x − c, x − c).
2
The Morse lemma states that in an appropriate chart, we have equality.

Theorem 1.3.1 (Morse lemma). Let f : M → R be a Morse function. Suppose c is a

critical point of f . Then there is a local chart (x1 , . . . , xn ) (called a Morse chart) containing
c such that, on this chart,
i
X n
X
f (x) = f (x1 , . . . , xn ) = f (c) − x2j + x2j .
j=1 j=i+1

The integer i depends only on the critical point, and is called the index of the critical point.

Remark. If i is the index of c, then (n − i, i) is the signature of the bilinear form (d2 f )c .

Corollary 1.3.2. Critical points of morse functions are isolated, by observing that on
a Morse chart of c, df only vanishes at c. It follows that Morse functions on compact
manifolds have finitely many critical points.

In figure 1.1 we inspected the height function of the torus. This is a Morse function
with four critical points. Starting from the bottom, we see that the critical points have
index 0, 1, 1, 2. In general a local maximum has full index n, while a local minimum has
index 0. Saddle points have index strictly between 0 and n.
The above corollary does not in fact require the full power of the Morse lemma. Another
proof is as follows.

6
Exercise 1.3.3. (A-D, exercise 2.) Characterise non-degenerate critical points of f : M →
R in terms of transversality of df : M → T ∗ M . Deduce that non-degenerate critical points
are isolated.

Solution: Let f : M → R. Suppose c is a critical point of f . Then dfc = 0. Interpreting

df as a map M → T ∗ M , this means that dfc (V ) = 0 for all V ∈ Tc M , so df intersects the
zero section Z = {α ∈ T ∗ M : αx = 0 for all x ∈ M } ∼ = M ,→ T ∗ M at c. Recall that c is
a non-degenerate critical point if and only if d2 fc is a non-degenerate bilinear form, i.e. it
has full rank. Thus c is non-degenerate if and only if d(df )c : Tc M → Tdf (c) (Tc∗ M ) defined
by d(df )c : Vc 7→ ((c, dfc ), d(df )c Vc ) is an isomorphism. Equivalently, the image of d(df )c
is Tdf (c) (Tc∗ M ). This happens if and only if im d(df )c + Tdfc Z = Tdf (c) (T ∗ M ). Therefore
non-degenerate critical points of f are precisely those c ∈ M such that df intersects Z
transversely.
We next prove that non-degenerate critical points are isolated. Recall that the inter-
section of two transverse submanifolds is itself a submanifold, with codimension given by
the sum of the codimensions of the two submanifolds. Moreover, im df is an embedded
submanifold of T ∗ M . (One can readily show, using the definition of a section, that df is
an injective immersion. Using continuity of π : T ∗ M → M , one can conclude that df is
proper, so it is an embedding.) Since the zero section and im df each have codimension n
in T ∗ M , the non-degenerate critical points must be an embedded 0-manifold. Therefore
the non-degenerate critical points are isolated, as required. 4

1.4 Examples and exercises

Arguably the most important Morse functions are height functions and distance-to-a-point
functions. The former was introduced in the introduction to the introduction, while the
latter was introduced in the proof of the abundance of Morse functions. We now see more
examples via some exercises.

Exercise 1.4.1. (A-D, exercise 3.) Monkey saddle: investigate f : R2 → R, defined by

(x, y) 7→ x3 − 3xy 2 .

Solution: Observe that ∂x f = 3x2 − 3y 2 , while ∂y f = −6xy. Therefore there is a unique

critical point, at x = y = 0. This is a degenerate critical point, since the Hessian also
vanishes. This is in keeping with visual intuition, since f does not have a true saddle
at (0, 0): instead the level set f −1 (0) is three intersecting lines, showing that the critical
point is not “primitive” in a sense. Writing f (x) = x(x2 − 3y 2 ) and perturbing it to give
(x − α)(x2 − 3y 2 ) separates the critical point into a non-degenerate saddle at (0, 0) and a
degenerate critical point (with f −1 (0) a line) at x = α. 4

7
Exercise 1.4.2. (A-D, exercise 4.) Show that if f : M → R and g : N → R are Morse,
then f + g : M × N → R is Morse, and the critical points are pairs of critical points of f
and g.

Solution: Explicitly, f +g is defined by (f +g)(x, y) = f (x)+g(y). Suppose (x, y) ∈ M ×N

is a critical point. Then x is necessarily a critical point of f , and y is necessarily a critical
point of g. To see this, observe that d(f + g) = df + dg, but dgM ≡ 0, so whenever
d(f + g) = 0, dfM must also vanish. Thus x is a critical point of f , and similarly for g. The
converse also holds, so the critical points of f + g are exactly the pairs of critical points of
f and g. Similarly the Hessian of f + g is the sum of the Hessians, which vanishes on the
critical points. Therefore f + g is Morse. 4

8
Chapter 2

Pseudo-gradients, topology, and

the Smale condition

2.1 Existence of pseudo-gradients

If f : Rn → R is a smooth function, then its gradient is the vector field grad f defined by
 ∂f ∂f 
gradx f = (x), . . . , (x) .
∂x1 ∂xn
Equivalently, it is the vector field defined by

g(grad f, Y ) = df (Y )

for all vector fields Y on Rn . Here g is the Euclidean metric on Rn . This idea generalises
to Riemannian manifolds.
Definition 2.1.1. Let f : M → R, (M, g) a Riemannian manifold. The gradient of f is
the vector field grad f defined by

g(grad f, Y ) = df (Y )

for all vector fields Y .

The two key properties of gradients are the following:
1. (Since metrics are non-degenerate), the gradient vanishes if and only if df = 0, i.e. it
vanishes precisely on critical points.
2. (Since metrics are positive-definite), f decreases along integral curves of f . More
precisely, let ϕ be the flow of − grad f . Then for any non-critical x,
d
(f (ϕt (x))) = (df )ϕt (x) (− gradϕt (x) f ) = −g(gradϕt (x) f, gradϕt (x) f ) < 0.
dt

9
Using these properties we construct pseudo-gradient fields whose integral curves connect
critical points of Morse functions. These allow the notions of stable and unstable manifolds
of critical points, which later become significant. In general we do not have a Riemannian
metric lying around, so with these two key properties in mind, we define pseudo-gradients.

Definition 2.1.2. Let f : M → R. A vector field X is a pseudo-gradient adapted to f if

1. (df )x (X) ≤ 0, and equality holds if and only if x is a critical point of f .

2. In a Morse chart around a critical point x, X agrees with − grad f (for the canonical
metric of Rn ).

We now establish some notation that will be used hereafter. Let f : M → R be a

function and c a critical point of index i. Then there is a Morse chart in a neighbourhood
of c in which f is of the form
i
X n
X
f (x) = f (c) − x2j + x2j = f (c) + Q(x).
j=1 j=i+1

Let V− be the span of x1 , . . . , xi , and V+ the span of xi+1 , . . . , xn . Then V = V− ⊕ V+ , and

Q is negative definite on V− while positive definite on V+ . For each ε, η > 0, the “standard
balls” are defined by

U (ε, η) = {x ∈ Rn : −ε < Q(x) < ε, kx− k2 kx+ k2 ≤ η(ε + η)}.

Since Q : Rn → R, it has a gradient,

− grad(x− ,x+ ) Q = 2(x− , −x+ ).

The boundary of U (ε, η) consists of three pieces:

• ∂+ U = {x ∈ Rn : Q(x) = ε, kx− k2 ≤ η},

• ∂− U = {x ∈ Rn : Q(x) = −ε, kx+ k2 ≤ η},

• ∂0 U = {x ∈ Rn : |Q(x)| ≤ ε, kx− k2 kx+ k2 = η(ε + η)}.

The first two pieces bound sublevel sets of Q, and the last piece is made up of segments of
integral curves of the gradient of Q.
In summary, given a critical point c, there is a chart U = U (ε, η) ⊂ Rn so that its
image Ω(c) ⊂ M under some diffeomorphism h is a neighbourhood of c. We also denote
the boundaries of the neighbourhood by ∂± Ω(c) = h(∂± U ) and ∂0 Ω(c) = h(∂0 U ). We
consistently try to denote neighbourhoods in M by Ω, and the model spaces (charts in Rn )
by U .

10
Theorem 2.1.3. Given any Morse function f : M → R, M compact, there is a pseudo-
gradient adapted to f .
Proof. We give a proof outline. One approach is to use the existence of Riemannian metrics
on manifolds. A more elementary approach is to use partitions of unity, which we describe
here.
1. f has finitely many critical points, c1 , . . . , cn . These have disjoint Morse charts
(U1 , h1 ), . . . , (Un , hn ). This extends to a finite atlas {Ui : i ∈ I}, so that each cj is
contained in exactly one Ui .
2. On each Ωi , define Xi to be the pullback of the vector field − grad f ◦ hi on Ui . Let
ϕj be a partition of unity subordinate to {Ωi : i ∈ I}. Define
X
X= ϕi (x)Xi (x).
i∈I

3. One can verify that X is a pseudo-gradient adapted to f . The key observation is

that if X vanishes at x, x must be a critical point: otherwise every ϕi (x) vanishes,
which is absurd.

2.2 Trajectories and stable/unstable manifolds

Let f : M → R, and let X be a pseudo-gradient field. The vector flows of X are called
trajectories of X, denoted ϕt . The most important property of trajectories is that they
are guaranteed to connect critical points. We begin this section by defining defining stable
and unstable manifolds of critical points, which are collections of trajectories that tend to
(or from) the critical point.
Definition 2.2.1. Let c be a critical point of f : M → R. The stable manifold of c is
W s (c) = {x ∈ M : lim ϕt (x) = c}.
t→∞

The unstable manifold of c is

W u (c) = {x ∈ M : lim ϕt (x) = c}.
t→−∞

With notation as established in the previous section, if our manifold if U = U (ε, η), we
have W s (0) = U ∩ V+ , W u (0) = U ∩ V− .
Proposition 2.2.2. The stable and unstable manifolds of a critical point c are submani-
folds. Moreover, they are diffeomorphic to open disks, and
dim W u (c) = codim W s (c) = ind(c)

11
 This loosely says that the trajectories belonging to stable and unstable manifolds de-
scribe the critical points. But it turns out that all trajectories belong to a stable and
unstable manifold.

Proposition 2.2.3. Let M be compact, and ϕt (x) a trajectory of a pseudo-gradient field

X of f . Then there are critical points c, d of f such that

lim ϕt (x) = c, lim ϕt (x) = d.

t→∞ t→−∞

Proof. We give a proof outline for the case limt→∞ ϕt (x). First suppose that limt→∞ ϕt (x)
exists. Then the limit is necessarily a critical point, since X vanishes exactly on critical
points. Therefore it suffices to show that the limit exists.
Suppose for a contradiction that the limit does not exist. By the definition of pseudo-
gradients, f is then strictly decreasing along ϕt , so if it enters any Morse chart it must
leave in finite time. Let t0 be the time at which it leaves all Morse charts. By compactness
of M , there exists ε0 such that for all t > t0 , (df )x (X) ≤ ε0 . Therefore the limit of f (ϕt (x))
as t → ∞ is −∞, which is impossible by the compactness of M .

2.3 Critical values topology

The two important theorems of this section establish connections between critical values
and topology. First, in the case where no critical points are crossed, the topology is
unchanged. Second, in the case where a critical point of index k is crossed, the topology
changes by the attachment of a k-cell.

Definition 2.3.1. Let M be a manifold, and f : M → R. It is well known that if a is

a regular value of f , then the level set f −1 (a) is an embedded submanifold. The same
holds for sublevel sets: define M a = f −1 ((−∞, a]). This is a submanifold with boundary.
a
Similarly superlevel sets are denoted M .

Theorem 2.3.2. Let f : M → R. Suppose a, b ∈ R, f −1 ([a, b]) is compact, and f has no

critical points in f −1 ([a, b]). Then M a is diffeomorphic to M b .

Proof. We give a proof outline. The idea is to flow along a pseudo-gradient to retract M b
to M a . Consider a function ρ : M → R satisfying
(
− (df )x1 (X) x ∈ f −1 ([a, b])
ρ(x) =
0 outside of a compact neighbourhood of f −1 ([a, b]).

Let Y be the vector field ρX, and ψ t the flow of Y . Then for each t, ψ t : M → M is a
diffeomorphism, and in particular ψ b−a maps M b onto M a .

12
Corollary 2.3.3 (Reeb’s theorem). Suppose a closed manifold M admits a Morse function
with exactly two critical points. Then M is homeomorphic to a sphere.
Proof. Let f : M → R be a Morse function with two critical points, with M an n-manifold.
Since M is compact, im f = [a, b] for some a < b. Then f −1 (a) is necessarily a maximum
and f −1 (b) is a minimum. By the Morse lemma, for ε sufficiently small, M a+ε = f −1 ([a, a+
b−ε
ε]) and M = f −1 ([b−ε, b]) are diffeomorphic to n-disks. By the previous theorem, M a+ε
b−ε
is diffeomorphic to M b−ε . But now M is equal to M b−ε ∪ M , i.e. two n-disks glued
along their boundary. This is homeomorphic to an n-sphere.

Remark. The above result is still true if f is not Morse, that is, if the critical points are
degenerate. Note also that the final conclusion is only true up to homeomorphism, since
gluing two disks can result in exotic spheres.
In the above we discussed the special case of traversing to different sublevel sets without
crossing any critical points. Next we investigate the case when we cross a critical point (of
index k).
Theorem 2.3.4. Let f : M → R. Suppose a, b ∈ R, f −1 ([a, b]) is compact, and f has
exactly one critical point α in f −1 ([a, b]), of index k. Then M b is homotopy equivalent to
M a with a k-cell attached. (More explicitly, M b is homotopic to M a ∪ W u (α).)
Rather than giving a proof, we give two examples (as in figure 2.1): In this figure,

c
β

b
α

Figure 2.1: Examples of changing topologies of level sets.

M a has the homotopy type of a point, and M b has the homotopy type of a circle. We
also observe that α is the unique critical point in f −1 ([a, b]), and has an index of 1. The
corresponding unstable manifold is shown in the figure, and are the two curves following
a downward path from α shown in red. Thus M a ∪ W u (α) is “M a with a one-dimensional
handle attached”, and has the homotopy type of a circle as required.

13
 Next observe that M c once again has the homotopy type of a point. There is a unique
critical point β lying between b and c. This has index 2, and we see that the corresponding
unstable manifold is a disk lying below β, shown in blue. Thus M b ∪ W u (β) is a “cylinder
with one end capped”, and hence has the homotopy type of a point as required.
We now give closer attention to the stable and unstable manifolds, exploring the Smale
condition.

2.4 Smale condition

Definition 2.4.1. Let f : M → R be Morse. A pseudo-gradient adapted to f satisfies the
Smale condition if all stable and unstable manifolds of f meet transversely, that is, for any
a, b critical, W u (a) t W s (b).

We later find that the Smale condition ensures some combinatorial properties that tell
us how to compare the index of distinct critical points. But first, some examples:

Example. The critical points of figure 2.1 consist of three extrema and one saddle. The
unstable and stable manifolds of extrema are either n-dimensional submanifolds or points.
Therefore whenever a stable or unstable manifold of an extremum intersects an unstable
or stable manifold of another critical point, we find that the manifold corresponding to
the extremum is n-dimensional, so the manifolds are transverse. But the remaining cases
are intersections of stable and unstable manifolds of a fixed critical point. These always
meet transversely, so this shows that the “wobbly sphere” (figure 2.1) satisfies the Smale
condition.

Example. Two special cases (one of which was used explicitly above):

• By the Morse lemma, given any critical point c, W s (c) t W u (c).

• By the definition of pseudo-gradients, the unstable manifold always lies below a

critical point, and the stable manifold above. Therefore whenever a, b are distinct
critical points with f (a) ≤ f (b), then W s (b) ∩ W u (a) is empty. In particular, they
are transverse.

Recall that for any critical point c, dim W u (c) = codim W s (c) = ind(c). But now If
a, b are any two critical points of a pseudo-gradient satisfying the Smale condition, then
W s (a) t W u (b), so

dim(W s (a) ∩ W u (b)) = n − codim(W s (a) ∩ W u (b))

= n − (codim W s (a) + codim W u (b))
= n − (ind(a) + (n − ind(b))) = ind(b) − ind(a).

14
Therefore the differences in indices is exactly the dimension of some submanifold of M .
But what is this submanifold? It consists exactly of the trajectories of the pseudo-gradient
connecting b to a!

M(b, a) := W s (a) ∩ W u (b) = {x ∈ M : lim ϕt (x) = a, lim ϕt (x) = b}.

t→∞ t→−∞

In particular, if M(b, a) is non-empty, it consists of at least one trajectory and has dimen-
sion at least one. Therefore indices of critical points always decrease along trajectories.

Theorem 2.4.2 (Kupka-Smale theorem). Let M be a manifold (possibly with boundary).

Let f be a Morse function on M with distinct critical values. Fix Morse charts about
each critical point, and denote their union by Ω. Let X be a pseudo-gradient adapted to
f , transverse to ∂M . Then there is a pseudo-gradient X 0 satisfying the Smale condition,
arbitrarily close to X (in C 1 -norm), and equal to X on Ω.

Remark. All approximations in this section use the C 1 -norm. We hereafter say that f is
approximated by g to mean there exist arbitrarily good C 1 approximations g of f .

Remark. 1. Any Morse function f : M → R can be approximated by Morse functions fe

with distinct critical values. Explicitly, perturb f by an appropriate function h which is
constant on Morse charts, and has sufficiently small |dh|.
2. It is not true in general that every Morse function on a manifold with boundary has
a pseudo-gradient transverse to the boundary. However, one can start by defining a vector
field transverse to the boundary, and then define a Morse function for which an extension
of this vector field is a pseudo-gradient.
Therefore the Smale theorem proves that on compact manifolds pairs (f, X) such that
f is a Morse function whose critical points take distinct values and X is a pseudo-gradient
adapted to f satisfying the Smale-condition exist and are generic.

Proof. A proof of the Smale theorem can be found in Audin and Damian.

15
Chapter 3

Morse homology fundamentals

3.1 Morse homology modulo 2

In the first chapter we established that given a compact manifold, Morse functions exist
and are generic. In the second chapter we established moreover that pairs (f, X) where f
is Morse and X is a pseudo-gradient adapted to f satisfying the Smale condition exist and
are generic. Such (f, X) are said to be Morse-Smale.
Let M be a compact manifold, and (f, X) Morse-Smale on M . In this chapter we
define the Morse complex on M using (f, X). We then show that the Morse complex is
independent of the choice of (f, X), so it is an invariant of M . Finally we show that Morse
homology is isomorphic to Singular homology.
To this end, we start by defining an appropriate space of coefficients via the quotient
manifold theorem.

Proposition 3.1.1. Let G be a Lie group acting smoothly, freely, and properly on a
smooth manifold M . Then M/G is a topological manifold of dimension dim M − dim G,
with a unique smooth structure such that π : M → M/G is a smooth submersion.

Given critical points a, b of a Morse function f , we defined M(b, a) to be the collection

of points lying on trajectories from b to a. Recall that M(b, a) is an ind(b) − ind(a)
dimensional submanifold of M . The Lie group R acts on M(b, a) by translations in time:

t · x = ϕt (x).

The action is smooth since ϕt is smooth. In fact, ϕt is a diffeomorphism for any fixed t, so
the action is also proper. To apply the quotient-manifold theorem, it remains to verify that
the translation action is free. This follows from the fact that M(b, a) contains no critical
points, so f (ϕt (x)) is a strictly decreasing function of t. This shows that the quotient
manifold theorem applies, giving the following definition:

16
Definition 3.1.2. Let a, b be critical points of f . Then L(b, a) := M(b, a)/t is the space
of trajectories from b to a. By the quotient manifold theorem, L(b, a) is a smooth manifold
of dimension ind(b) − ind(a) − 1.

Let M be compact, equipped with a Morse-Smale pair (f, X). For any i, let ci denote
a critical point of index i. For integral Morse homology, we use the signed cardinalities
NX (ci+1 , ci ) ∈ Z of L(ci+1 , ci ) as coefficients. For easier calculation ignoring orientation,
we consider coefficients in Z/2Z. In other words, we settle for the cardinalities computed
modulo 2, denoted nX (ci+1 , ci ) ∈ Z/2Z.

Remark. By the previous definition, L(ci+1 , ci ) is a 0-dimensional manifold for any i.

For the following definitions to be well defined, we require that L(ci+1 , ci ) is always finite
(equivalently, compact). This is indeed true, and will be shown in a subsequent section.

Definition 3.1.3. For each k, let Critk (f ) denote the set of critical points of f of index
k. For any ring R, Ck (f, R) denotes the free R-module of formal sums
n X o
Ck (f, R) := ac c : ac ∈ R .
c∈Critk (f )

The vector spaces Ck (f, Z/2Z) will be the terms appearing in the mod 2 Morse complex.

Definition 3.1.4. Given any k, the boundary map ∂ = ∂k+1 : Ck+1 (f, Z/2Z) → Ck (f, Z/2Z)
is defined on critical points ck+1 by
X
∂(ck+1 ) = nX (ck+1 , ck )ck .
ck ∈Critk (f )

This uniquely extends to a linear map on Ck+1 (f, Z/2Z).

Definition 3.1.5. The Morse complex is defined to be the chain complex

∂k+1 k ∂
· · · → Ck+1 −−−→ Ck −→ Ck−1 → · · · .

Remark. Well-definedness of the Morse complex now rests on two results, both of which
are shown in a following section. Namely,

1. The boundary maps are well defined, i.e. L(ck+1 , ck ) is finite for each ck+1 , ck .

2. The complex is truly a complex, i.e. ∂ 2 = 0.

We conclude this section by exploring some examples.

17
Example. Spheres: the usual height function, and the height function on the wobbly
sphere (as seen in figure 2.1).
We start by computing the Morse complex and corresponding homologies for the height
function h on the usual sphere, Sn , with n ≥ 2. This has exactly two critical points, one
of index 0 and one of index n. The Morse complex is then

· · · → 0 → Cn (h, Z/2Z) = Z/2Z → 0 → · · · → 0 → C0 (h, Z/2Z) = Z/2Z → 0 → · · · .

Each boundary map is necessarily the zero-map, forcing the Morse homologies to be
(
Z/2Z k ∈ {0, n}
Hk (h, Z/2Z) =
0 otherwise.

Next we consider the 2-sphere equipped with the Morse function f corresponding to the
height function of the wobbly sphere, figure 3.1. This has four critical points, one of index
0 (a), one of index 1 (b), and two of index 2 (c, d). Therefore the Morse complex is

C• (f, Z/2Z) = · · · → 0 → (Z/2Z)2 → Z/2Z → Z/2Z → 0 → · · · .

Inspecting the diagram, the boundary map ∂1 : C1 (f, Z/2Z) → C0 (f, Z/2Z) sends b to 0
(since there are two trajectories, so nX (b, a) = 0.) It follows that ∂1 is the zero map. Next
we observe from the diagram that nX (c, b) = nX (d, b) = 1. It follows that ∂2 is surjective.
Thus
im ∂1 = 0, ker ∂1 = Z/2Z, im ∂2 = Z/2Z, ker ∂2 = Z/2Z.
Computing the Morse homology, we find that
(
Z/2Z k ∈ {0, n}
Hk (f, Z/2Z) =
0 otherwise.

This shows that the Morse homologies of S2 calculated using f and h agree. In fact, we later
show that the Morse homology is independent of the choice of Morse-Smale pair (f, X).

Example. A tilted torus and a tilted Klein bottle. These examples are notable, since
(T2 , h) (where h is the usual height function) does not canonically give a Morse-Smale pair
(h, X); the two index 1 critical points are joined by two trajectories, which is forbidden
by the Smale condition. Therefore the torus (and the Klein bottle) must be tilted slightly,
as shown in figures 3.2, 3.3. Let h denoted the tilted height function of the torus. By
inspecting the figure, h has one critical point of index 0 (a), two of index 1 (b, c), and one
of index 2 (d). Therefore the Morse complex is

C• (h, Z/2Z) = · · · → 0 → Z/2Z → (Z/2Z)2 → Z/2Z → 0 → · · · .

18
 d

Figure 3.1: Wobbly sphere Morse complex.

From figure 3.2, we see that nX (d, c) = nX (d, b) = 0 since there are two trajectories in
each case (shown in red). Therefore ∂2 is the zero map. Similarly nX (c, a) = nX (b, a) = 0,
so ∂1 is also the zero map. It follows that the Morse homology is given by

2
(Z/2Z) k = 1

Hk (h, Z/2Z) = Z/2Z k ∈ {0, 2}

0 otherwise.


Observe that this agrees with the singular homology of the torus.

Figure 3.2: Tilted torus Morse complex.

Next we compute the Morse homology of the tilted Klein bottle, shown in figure 3.3.
Let h0 denote the tilted height function. We find that the Klein bottle has the same critical-
point profile as the tilted torus: one critical point of index 0 (a0 ), two of index 1 (b0 , c0 ),

19
and one of index 2 (d0 ). Therefore the Morse complex has the same objects as in the case
of the torus;

C• (h0 , Z/2Z) = · · · → 0 → Z/2Z → (Z/2Z)2 → Z/2Z → 0 → · · · .

In fact, counting trajectories modulo 2, we find that both ∂1 and ∂2 vanish in C• (h0 , Z/2Z).
Therefore 
2
(Z/2Z) k = 1

0
Hk (h , Z/2Z) = Z/2Z k ∈ {0, 2}

0 otherwise.


This shows that Morse homology (mod 2) cannot distinguish a Klein bottle from a torus.
Of course, this is expected, since Singular homology mod 2 also does not distinguish the
two.

Figure 3.3: Tilted Klein bottle Morse complex.

As a corollary of the above examples, we have shown that a sphere is not diffeomorphic
to a torus or a Klein bottle, but we still haven’t shown that a torus is not diffeomorphic
to a Klein bottle. For this we use Morse homology over the integers.

3.2 Integral Morse homology

Next we define Morse homology with coefficients in Z. The main technicality is keeping
track of orientations (signs). In fact, the mod 2 Morse homology in the previous section
section follows exactly from the construction we carry out here, since signed counting is
not detected modulo 2. Again, proofs of well-definedness are pushed into a subsequent
section, here we just define the complex and look at some examples.

20
 As defined in the previous section, the objects in the integral Morse complex are
n X o
Ck (f, Z) = ac c : ac ∈ Z .
c∈Critk (f )

The corresponding boundary maps are defined to be ∂k+1 : Ck+1 (f, Z) → Ck (f, Z),
X
∂(ck+1 ) = NX (ck+1 , ck )ck ,
ck ∈Critk (f )

where NX (ck+1 , ck ) is the signed count of trajectories from ck+1 to ck . The integral Morse
complex is
∂k+1 k ∂
· · · → Ck+1 −−−→ Ck −→ Ck−1 → · · · .
This is a well defined chain complex for the same reason that the mod 2 Morse complex is
a chain complex, which we soon show.
We now describe the process of signed counting, and compute some examples. The aim
is to induce orientations on each L(ck+1 , ck ). Since these are zero dimensional manifolds,
orientations are exactly choices of sign for each point.
Start by choosing an orientation for each stable manifold W s (c). These are homeomor-
phic to disks (of some dimension), so they are orientable. Choose any x ∈ M(ck+1 , ck ).
There is a short exact sequence

0 → Tx M(ck+1 , ck ) → Tx W s (ck ) → Nx W u (ck+1 ) → 0.

But Nx W u (ck+1 ) is canonically isomorphic to Tx W s (ck+1 ), giving a short exact sequence

0 → Tx M(ck+1 , ck ) → Tx W s (ck ) → Tx W s (ck+1 ) → 0.

Since the middle and right term are oriented, there is an induced orientation on Tx M(ck+1 , ck ).
But we also have a short exact sequence

0 → R → Tx M(ck+1 , ck ) → Tx L(ck+1 , ck ) → 0,

where R is oriented by time. This induces an orientation on Tx L(ck+1 , ck ) as required.

Remark. Although choices are being made when orienting the stable manifolds, reversing
the orientation of a given stable manifold W s (c) corresponds to multiplying NX (d, c) and
NX (c, b) by −1, for any d and b. Therefore the isomorphism classes of the integral Morse
homologies are independent of the choice of orientation.
Example. Integral Morse homology of a torus. Let h denoted the tilted height function
of the torus. By inspecting figure 3.4, h has one critical point of index 0 (a), two of index
1 (b, c), and one of index 2 (d). Therefore the Morse complex is

C• (h, Z) = · · · → 0 → Z → Z2 → Z → 0 → · · · .

21
Next we determine the boundary maps. The arrows on figure 3.4 denote the chosen orienta-
tions of stable manifolds. The stable manifold of d is a point, so it assigned the orientation
+. Working through the exact sequences above, we find that the two trajectories in L(d, c)
have orientations + and −, so NX (d, c) = 0. It turns out that all of the NX vanish (with
trajectory orientations as shown in the figure), so every boundary map is trivial. Therefore
the integral Morse homology of the torus is equal to

2
Z k = 1

Hk (h, Z) = Z k ∈ {0, 2}

0 otherwise.


d
−
+
+
c
+
−

−
b

−
+ a

Figure 3.4: Integral (signed) torus Morse complex.

Example. Integral Morse homology of a Klein bottle. This is again similar to the torus.
Let h denoted the height function of the Klein bottle. By inspecting figure 3.5, h has one
critical point of index 0 (a), two of index 1 (b, c), and one of index 2 (d). (These are not
labelled on the figure to prevent clutter; the labels a, . . . , d are in height-ascending order.)
The Morse complex is

C• (h, Z) = · · · → 0 → Z → Z2 → Z → 0 → · · · .

Next we determine the boundary maps. For the most part, the signed counting ends up
being identical to that of the torus. The only difference is the trajectories from d to b,

22
which both end up having negative sign. One can verify that a “tubular neighbourhood”
of the two trajectories from d to b is an embedded Mobius strip. All signs are shown in
figure 3.5, from which we conclude that ∂1 is trivial, and ∂2 is defined by
∂1 (xd) = −2xb
for all xd ∈ C2 (h, Z). It follows that im ∂2 ∼
= 2Z, and ker ∂2 is trivial. Therefore the integral
Morse homology of the Klein bottle is equal to

Z ⊕ Z/2Z k = 1

Hk (h, Z) = Z k=0

0 otherwise.


−
+

− −
+
−

−
+

Figure 3.5: Integral (signed) Klein bottle Morse complex.

In summary, while mod 2 Morse homology could not detect orientation, the integral
Morse homology can distinguish a Klein bottle from the torus. As a corollary, a torus and
Klein bottle are not diffeomorphic. Observe that again the Morse and singular homologies
agree.

3.3 Well-definedness of the Morse complex

The goals of this section are to show that L(b, a) is finite (when ind(b) = ind(a) + 1),
and that the Morse complex is truly a complex. To achieve this, we construct the notion

23
of broken trajectories. Recall that for any critical points a, b of a Morse function, L(b, a)
consists of the trajectories from b to a, and is a manifold of dimension ind(b) − ind(a) − 1.
The space of broken trajectories from b to a is a certain compactification of L(b, a):
Definition 3.3.1. Let a, b be critical points of a Morse function f . Then
[
L(b, a) := L(b, c1 ) × · · · × L(cq , a)
ci ∈Crit(f )

is the space of broken trajectories from b to a.

Each factor L(x, y) is endowed with the quotient of the subspace topology, and each
term in the union is equipped with the product topology. However, to make sense of the
union, we must define an appropriate topology on the whole space. The above definition
can be motivated by visualising the trajectories on a torus, and observing that if a and b
respectively denote the minimum and maximum (of the usual height function), then L(b, a)
is a disjoint union of four open intervals. However, L(b, a) should be a figure-eight.
A description of the topology of L(b, a) is given at the start of section 3.2 in Audin and
Damian. They describe a neighbourhood system as follows, to define the topology:
1. Let λ = (λ1 , . . . , λq ) ∈ L(b, a) be a broken trajectory. We will define a neighbourhood
W(λ, U − , U + ) of λ.

2. Each λi is a trajectory which exists some Morse neighbourhood Ω(ci−1 ) and enters
Ω(ci ). More specifically, the exit point xi of λi in Ω(ci−1 ) has a neighbourhood Ui−
contained in Ω(ci−1 ) ∩ f −1 (xi ). Similarly, there are neighbourhoods Ui+ of entry
points. Let U − denote the family of Ui− , and similarly for U + .

3. The collection of W(λ, U − , U + ) now defines a neighbourhood system by declaring

that η = (η1 , . . . , ηp ) ∈ W(λ, U − , U + ) whenever

(a) the ηj belong to L(cij , cij+1 ), where cij is a subsequence of the critical points
occurring in λ, and
(b) each ηj exists from within the corresponding neighbourhood in U − , and enters
the corresponding neighbourhood in U + .

With this topology, L(b, a) is a subspace of L(b, a). Moreover, the following key result
holds:
Theorem 3.3.2. The space L(b, a) of broken trajectories is compact.
Proof. We give a proof outline. Observe that the neighbourhood system given above con-
tains a countable neighbourhood system, using compactness and second-countability of
the ambient manifold M . Therefore to prove compactness it suffices to prove sequential
compactness.

24
 A further reduction can be made since a sequence in a finite product converges if and
only if it converges pointwise. Therefore it suffices to prove that any sequence in L(b, a) has
a convergence subsequence. Let (`n ) be a sequence of trajectories in L(b, a). Each `n exits
Ω(b) at some `− +
n , and enters Ω(a) at some `n . These form sequences in compact subsets
of M , so they have convergent subsequences. Recalling that a trajectory is a solution of
a differential equation, and these are unique given initial conditions, we conclude that `n
has a convergent subsequence.

This result completes one of the goals of this section:

Proposition 3.3.3. Suppose b and a are critical points with ind(b) = ind(a) + 1. Then
L(b, a) is finite. In particular, the boundary maps defined for the mod 2 and integral Morse
complex are well defined.

Proof. Given the premise, L(b, a) is a 0 dimensional manifold. Therefore it suffices to prove
compactness of L(b, a). But this follows from the general fact that L(b, a) is compact.

Our next goal is proving that the Morse complex is indeed a complex. Specifically, it
remains to show that ∂ 2 = 0. We give an outline, the details of which are in Audin and
Damian. The key result that must be proven is the following:

Theorem 3.3.4. Let a, b be critical points of f with ind(b) − ind(a) = 2. Then L(b, a) is
a compact one dimensional manifold with boundary.

To see why this gives the desired result, fix a, b as above. We prove that ∂ 2 b = 0 for
both integral Morse homology and mod 2 Morse homology.
Let i + 2 be the index of b. By definition, we have
X
∂b = NX (b, ci+1 )ci+1 ,
ci+1 ∈Criti+1 (f )
X X
2
∂ b= NX (∂b, ci )ci = NX (b, ci+1 )NX (ci+1 , ci )ci .
ci ∈Criti (f ) ci ∈Criti (f )
ci+1 ∈Criti+1 (f )

To show that ∂ 2 = 0, it suffices to show that ci NX (b, ci+1 )NX (ci+1 , a) is zero. On the
P
other hand, using the previous theorem we know that
[
L(b, a) = L(b, a) t ∂L(b, a) = L(b, a) t L(b, ci )L(ci , a).
ci
P
Therefore the expression ci NX (b, ci+1 )NX (ci+1 , a) is in fact the cardinality of ∂L(b, a);
i.e. the cardinality of the boundary of a compact one dimensional manifold with boundary.
By the classification of one dimensional manifolds with boundary, this is cardinality is
always 0 modulo 2. Similarly when the manifold is oriented, the signed count of boundary

25
 P
points is always 0. Therefore ci NX (b, ci+1 )NX (ci+1 , a) = 0. Since a, b were arbitrary, it
follows that ∂ 2 = 0. Therefore the Morse complex is truly a complex, as required.
We now explore the key theorem:

Proof. We give a proof outline that L(b, a) is a compact one dimensional manifold with
boundary, provided ind(b) − ind(a) = 2. We already know that L(b, a) is a one dimensional
manifold, and we know that L(b, a) is a compact (metrisable) topological space. Therefore
it is sufficient to prove the following result:
Let M be compact, and (f, X) a Morse-Smale pair on M . Fix k and let b, c, a be
critical points of index k + 2, k + 1, and k respectively. Let λ1 , λ2 be trajectories from
b to c and c to a respectively. Then there exists a continuous embedding ψ from [0, δ)
onto a neighbourhood of (λ1 , λ2 ) in L(b, a) that is differentiable on (0, δ), and satisfies
ψ(0) = (λ1 , λ2 ), ψ(s) ∈ L(b, a) for s 6= 0. Moreover, if (`n ) is a sequence in L(b, a) that
tends to (λ1 , λ2 ), then `n is contained in the image of ψ for sufficiently large n.
Proving the above result turns out to be fairly technical, but all of the details are given
in Audin and Damian.

3.4 Morse-Smale pair invariance of the Morse homology

Earlier in the chapter we computed the Morse homology of a 2-sphere using the standard
height function as well as the wobbly height function. In each case we observed that the
Morse homologies were unchanged! This is in fact a general result: the Morse homology
does not depend on the choice of Morse function or pseudo-gradient field. That is, the
Morse homology depends only on the smooth manifold.
In an ideal world we could interpolate between two Morse functions with Morse func-
tions: given f0 , f1 : M → R, we can define a homotopy F : M × I → R from f0 and
f1 . However, F is not generally Morse, if the number of critical points changes. The idea
is that we can actually bypass this issue, provided we can construct a morphism of com-
plexes inducing isomorphisms on homology which do not refer to the degenerate points.
We describe a proof outline, but full details are given in Audin and Damian.

Theorem 3.4.1. Let M be compact, and (f0 , X0 ), (f1 , X1 ) two Morse-Smale pairs on M .
Then there exists a morphism of complexes Φ∗ : C• (f0 , X0 ) → C• (f1 , X1 ) which induces
isomorphisms on homology.

Proof. We somewhat categorify the proof. Let F be any interpolation between f0 and f1
that is constant on [0, 1/3] ∪ [2/3, 1]. More explicitly, suppose F : M × [0, 1] → R is a
smooth function such that

Ft |[0,1/3] ≡ f0 , Ft |[2/3,1] ≡ f1 .

26
We call such an F an end-constant interpolation. Let EndConstInt(M ) denote the cate-
gory of Morse-Smale pairs on M , where morphisms between Morse-Smale pairs are equiv-
alence classes of end-constant interpolations. (Two end-constant interpolations are equiva-
lent if they take the same constant values, i.e. if they have the same domain and codomain.)
Composition of morphisms is given by
(
Ft t ∈ [0, 1/3]
G◦F =
Gt t ∈ [2/3, 1],

and given any f , the identity morphism is given by Ft ≡ f . One can readily verify that
EndConstInt(M ) is a category.
On the other hand, there is also a category of Morse complexes, which we denote
MoCplx(M ). The objects are C• (f, X) for (f, X) a Morse-Smale pair on M , and mor-
phisms are chain maps.
Suppose there is a functor Φ : EndConstInt(M ) → MoCplx(M ), and let F be an end-
constant interpolation between (f0 , X0 ) and (f1 , X1 ), and G an end-constant interpolation
between (f1 , X1 ) and (f0 , X0 ). Then ΦF ◦ΦG and ΦG ◦ΦF must both be the identity maps on
their respective complexes, so ΦF induces an isomorphism on homologies. This shows that
to prove the theorem, it suffices to find a functor Φ : EndConstInt(M ) → MoCplx(M ).
This process is broken into two parts:

1. For a given end-constant interpolation F , define ΦF : C• (f0 ) → C• (f1 ). Show that

ΦF depends only on the equivalence class of F .

2. Verify that the induced morphism of complexes is functorial:

(a) Show that if I is an identity morphism in EndConstInt, then ΦI = id.

(b) Show that for morphisms F, G, H in EndConstInt (with compatible domains
and codomains), ΦG ◦ ΦF = ΦH .

We describe the construction of Φ, but skip the proof of 2 which verifies that our choice of
Φ is indeed a functor.
1. Let F : M × I → R be an end constant interpolation from f0 to f1 . Extend F to
[−1/3, 4/3] by keeping the ends constant. Although F is not in general Morse, we define
a new function using F which is guaranteed to be Morse. Specifically, choose g : R → R
to be Morse, with two critical points 0 and 1 (the maximum and minimum respectively),
such that g is decreasing sufficiently rapidly between 0 and 1. Precisely, we require
∂F
(x, t) + g 0 (t) < 0
∂t
for all x ∈ M and t ∈ (0, 1). This can be achieved by compactness of M , by allowing the
critical value of 0 to be very large. Then F + g : M × [−1/3, 4/3] → R is Morse, with

27
critical points exactly being

Crit(f0 ) × {0} ∪ Crit(f1 ) × {1}.

Moreover, the critical points (b, 0) have index ind(b) + 1, and the critical points (a, 1) have
index ind(a). Using a partition of unity, there exists a pseudo-gradient field X adapted
to F + g which coincides with X0 − grad g on M × [−1/3, 1/3], and with X1 − grad g on
M × [2/3, 4/3]. (F + g, X) is not necessarily Morse-Smale, but by genericness of Morse-
Smale pairs, there is an approximation X e of X which is Morse-Smale. Since (F + g, X)
restricted to M × [−1/3, 1/3] or M × [2/3, 4/3] is indeed Morse-Smale, in these instances
the small-perturbation-invariance of the Morse complex gives the following:

C• (F + g|M ×[−1/3,1/3] , X)
e = C•+1 (f0 , X0 )
C• (F + g|M ×[2/3,4/3] , X)
e = C• (f1 , X1 ).

e = Ck (f0 , X0 ) ⊕ Ck+1 (f1 , X1 ). But now the

This gives a decomposition Ck+1 (F + g, X)
boundary map ∂Xe decomposes as
 
∂X0 0
∂Xe = .
ΦF ∂X1

One can show that ΦF is the desired chain map, and moreover F 7→ ΦF is a functor.

3.5 Morse homology is singular homology

The most standard proof of the Morse-Smale pair invariance of Morse homology is to
simply show that the Morse homology is canonically isomorphic to the cellular homology
(and hence singular homology). This proves not only that Morse homology is independent
of the Morse-Smale pair used to define it, but further that Morse homology depends only
on the topological structure of the manifold, and not its smooth structure.
Theorem 3.5.1. Let M be a manifold, and (f, X) a Morse-Smale pair on M . Let C• (M )
denote the associated Morse complex. There is a cellular decomposition of M (with asso-
ciated cellular complex K• (M )), and an isomorphism

F : K• (M ) → C• (M ).

That is, a map which is an isomorphism in each degree, with F ◦ ∂ = ∂X ◦ F .

It follows that the Morse and singular homologies of a manifold are isomorphic. In the
above it was not clarified whether the homology was integral, mod 2, or something else -
this does now matter as the theorem holds for any coefficient ring.
An amazingly brief overview of the proof is as follows:

28
 1. Show that the Morse-Smale pair (f, X) induces a cellular decomposition of M ; the
cells are the unstable manifolds of each critical point.

2. Show that the corresponding complexes are isomorphic.

A full proof is given in Audin and Damian. Here we compute two examples: the usual
2-sphere and the wobbly 2-sphere. These examples should show that the isomorphism of
complexes is not difficult to prove, and the real difficulty lies in proving that a Morse-Smale
pair truly induces a cellular decomposition.

Example. The usual 2-sphere equipped with the height function. The height function has
two critical points, a of index 0 and b of index 2. The corresponding unstable manifolds
are a 0-cell W u (a), and a 2-cell W u (b). The cellular decomposition has no 1-cells, so the
cellular complex is
··· → 0 → G → 0 → G → 0 → ···
where G is the coefficient ring. The boundary maps are all automatically trivial, so the
cellular homology is
( )
G if k ∈ {0, 2},
HkCell (S2 , G) = = HkMorse (S2 , G)
0 otherwise.

as required.

Example. The wobbly 2-sphere equipped with the height function. The height function
has four critical points, a of index 0, b of index 1, and c and d of index 2. The correspond-
ing unstable manifolds are a 0-cell W u (a) (blue), a 1-cell W u (b) (red), and two 2-cells
W u (c), W u (d) (white) as shown in figure 3.6.

c
b W u (d)
W u (c)

W u (b)
W u (a)
a

Figure 3.6: Induced cellular decomposition of the wobbly sphere.

29
 Therefore the cellular complex is

· · · → 0 → G2 → G → G → 0 → · · ·

where G is the coefficient ring.

The two potentially-nontrivial boundary maps ∂2 and ∂1 , which we now compute. On
each 2-cell, ∂2 is defined by ∂2 e2 = N (e2 , W u (b))W u (b) where N (e2 , W u (b)) is the degree
of the induced map
S1 → M 1 → M 1 /M 0 → S1
where M i is the i-skeleton of M . (M −1 is taken to be the empty set.) In this case the map
is a composition of identity maps, so

N (W u (c), W u (b)) = N (W u (d), W u (b)) = g

where g is generator of G. Next for the case of ∂1 , we see that N (W u (b), W u (a)) = 0.
Intuitively this is because the one-cell W u (b) forms a cycle. In terms of degrees, consider
the induced map f : S0 → M 0 /∅ = ∗ with the two points in the domain signed by −1 and
1. The degree of f is 0 since the preimage of ∗ contains both points.
These calculations are remarkable in that the boundary maps of the cellular complex
are identical to the boundary maps in the Morse complex, and even the computations to
determine the boundary maps are similar.
It follows that the cellular homology is
( )
G if k ∈ {0, 2},
HkCell (S2 , G) = = HkMorse (S2 , G)
0 otherwise.

as required.

30
Chapter 4

Morse homology applications

4.1 The Morse inequalities

As we have established that Morse homology is isomorphic to singular homology, we can
define the Betti numbers of a manifold using Morse homology and obtain an equivalent
definition as in the singular (or de Rham) cases.

Definition 4.1.1. The Betti numbers bk (M ) of a manifold M are the ranks of the k-th
homology groups;
bk (M ) := rank Hk (M, Z).

The rank of a Z-module A is the dimension of the Q-vector space Q ⊗ A. By flatness of

Q, the rank-nullity theorem from linear algebra passes over to Z-modules (more generally
modules over a PID) in the following sense: given any short exact sequence 0 → A → B →
C → 0 of Z-modules, rank B = rank A + rank C. Using this fact we can easily derive the
Morse inequalities. (Note that Audin and Damian only derive the weak Morse inequalities.
Here we derive the strong Morse inequalities, which are what are usually referred to as the
Morse inequalities.)

Theorem 4.1.2 (Strong Morse inequalities). Let M be a manifold, and f a Morse function
on M . Let Ni denote the number of index i critical points of f . Then for any k ≥ 0,
k
X k
X
(−1)k−i Ni ≥ (−1)k−i bi (M ).
i=0 i=0

Proof. Recall that any Morse function f can be perturbed so that it has the same critical
points, but admits a pseudo-gradient satisfying the Smale property. Therefore without
loss of generality, f belongs to a Morse-Smale pair (f, X), induing the Morse complex
C• (M, Z). Recall that Ci (M, Z) is the free Z-module generated by critical points of index

31
i, so Ni = rank Ci (M, Z). Moreover, by the first isomorphism theorem, rank Ci (M, Z) =
rank ker ∂i + rank im ∂i . Therefore the left hand side of the inequality we wish to derive is
k
X
(−1)k−i Ni = Nk − Nk−1 + · · · + (−1)k N0
i=0
= rank ker ∂k + rank im ∂k
− rank ker ∂k−1 − rank im ∂k−1
+ ···
+ (−1)k rank ker ∂0 + (−1)k rank im ∂0 .

Since C−1 (M, Z) = 0, rank im ∂0 = 0. On the other hand, rank im ∂k+1 ≥ 0. By regrouping
terms, this gives an inequality
k
X
(−1)k−i Ni ≥ − rank im ∂k+1 + rank ker ∂k
i=0
+ rank im ∂k − rank ker ∂k−1
+ ···
− (−1)k rank im ∂1 + (−1)k rank ker ∂0 .

By by definition, bi (M ) = rank Hi (M, Z) = − dim im ∂i+1 +dim ker ∂i . Therefore the above
inequality is exactly the inequality we set out to prove.

In the above proof, rank im ∂k+1 ≥ 0 was the only inequality contributing to the in-
equality in the final result. In the case where k = n, ∂k+1 = 0. Therefore in the top
dimensional case the Morse inequalities are an equality.

Corollary 4.1.3. Let f be a Morse function on an n dimensional manifold M . Then

n
X n
X
(−1)i Ni = (−1)i bi (M ) = χ(M ).
i=0 i=0

Theorem 4.1.4 (Weak Morse inequalities). Let M be a manifold, and f a Morse function
on M . Let Nk denote the number of index k critical points of f . Then

Nk ≥ bk (M ).

Proof. This is immediate from the strong Morse inequalities by fixing k, and adding
Pk−1
the strong Morse inequality with i=0 (−1)k−i−1 Ni to the strong Morse inequality with
Pk k−i N .
i=0 (−1) i

32
Example. The wobbly sphere is equipped with a height function with one critical point
of index 0, one of index 1, and two of index 2. The alternating sum of the Ni s is therefore
1 − 1 + 2 = 2 = χ(S2 ). The Klein bottle is equipped with a height function with one critical
point of index 0, two of index 1, and one of index 2, giving 1 − 2 + 1 = 0 = χ(K2 ).
Example. Suppose M is a closed n-manifold admitting a Morse function with two critical
points. These are necessarily a maximum and minimum, i.e. index 0 and index n. It
then follows from the Morse inequalities that for each i, rank Hi (M ) = rank Hi (Sn ). Since
H0 (M ) and Hn (M ) are always free, it follows that M has isomorphic homology to the n-
sphere. By Whitehead’s theorem, M is a homotopy n-sphere. By the Poincaré conjecture,
M is homeomorphic to the n-sphere. This gives a very clean but completely unnecessary
(and probably circular) proof of the Reeb sphere theorem.
An immediate corollary of the weak Morse inequalities is the following:
Corollary 4.1.5. Let f be a Morse function on M . Then f has at least as many critical
points as the sum of the ranks of the homology groups of M .
In summary the Morse inequalities allow us to gain a lot of topological insight from
finding good functions on a manifold, following the theme of Reeb’s theorem and the
section concerning changes in topology passing from one sublevel set to another. One more
interesting theorem which will not be proven here is a result from discrete Morse theory.
Theorem 4.1.6. Let M be a manifold equipped with a cellular decomposition. Let mk
denote the number of k-cells in the decomposition. Then for each k,

mk ≥ bk (M ).

This is the weak version of the discrete Morse inequality. This doesn’t quite follow
from our proof that the Morse and cellular homologies of a manifold are equal, as we have
not shown that every cellular decomposition induces a suitable Morse-Smale pair. As a
corollary, it follows that there are no cellular decompositions of a torus into fewer than 4
cells. From our (standard) version of the Morse inequalities, we conclude that any Morse
function on a torus has at least four critical points.

4.2 Morse functions and simple connectedness

Unfortunately Homology does not in general detect simple connectedness. A well known
aspect of Hurewicz’s theorem is that there is an isomorphism
π1 (M ) ∼
π1ab (M ) := = H1 (M ).
[π1 (M ), π1 (M )]
Since non-trivial perfect groups exist, i.e. non-trivial groups that are equal to their derived
subgroups (such as A5 ), there are manifolds with vanishing first homology but non-trivial

33
fundamental group. Without further assumptions about the manifold, one cannot conclude
from a vanishing first homology that a manifold is simply connected.

Theorem 4.2.1. If a closed n-manifold M admits a Morse function with no critical points
of index 1, M is simply connected.

Remark. By the Morse inequalities, H1 (M ) has rank zero. This doesn’t generally imply
that H1 (M ) = 0, but even if it did, by the above observation we still cannot conclude that
M is simply connected.

Proof. The proof does not use Homology. Fix any minimum a ∈ M of f to be the base
point of M . Let γ be a loop in M based at a. We can assume without loss of generality
that γ is smooth.
Now let b be any critical point of index greater than 0. By assumption, b must have
index at least 2, so its stable manifold has dimension at most n − 2. On the other hand,
γ is a smooth map, so by Sard’s theorem it is homotopic to a function which is transverse
to W s (b). This argument can be repeated for the finitely many critical points of index at
least 2, with all homotopies fixing the basepoint. Since dγ has dimension 1, transversality
is equivalent to the statement that γ does not meet any of the stable manifolds of critical
points of index at least 2.
Let X be a pseudo-gradient adapted to f , and let x be a point in the image of γ. The
limit limt→∞ ϕtX (x) is a critical point, so M is a union of the stable manifolds of critical
points of f . Since γ has empty intersection with stable manifolds of critical points of index
at least 2, and there are no critical points of index 1, γ lies in the union of stable manifolds
of index 0. These are all disjoint, so γ is contained in one stable manifold. Stable manifolds
are diffeomorphic to disks, so γ is contractible.

4.3 Poincaré duality realised in Morse homology

Since we have established that Morse homology is isomorphic to Singular homology, we
suddenly have a lot of results such as the following:

• Poincaré duality

• Excision theorem

• Homology long exact sequence, Mayer–Vietoris sequence

• Künneth formula

• Hurewicz theorem

• Universal coefficient theorem

34
These can be established purely using Morse theory, rather than transporting the result
from singular or cellular homology. The first four bulleted results are established in this
manner in Audin and Damian. Here we describe a version of Poincaré duality, as I found
the idea to be particularly clean. In Audin and Damian, Poincaré duality is proven in the
original form as stated by Poincaré:
Theorem 4.3.1. Let M be a closed n-manifold. Then there is an isomorphism Hk (M, Z/2Z) ∼
=
Hn−k (M, Z/2Z). Moreover if M is oriented, then bk (M ) = bn−k (M ).
In these notes we obtain a stronger result which looks more similar to the usual general
statement of Poincaré duality for singular homology.
Theorem 4.3.2. Let M be a closed n-manifold. Then there is an isomorphism
t : C• (M, Z/2Z) → C • (M, Z/2Z), tk : Ck (M, Z/2Z) → C n−k (M, Z/2Z)
of chain complexes. In particular, for each k there is an isomorphism Hk (M, Z/2Z) ∼
=
H n−k (M, Z/2Z). Moreover if M is oriented, then there is an isomorphism
t : C• (M, Z) → C • (M, Z), tk : Ck (M, Z) → C n−k (M, Z).
In particular, for each k there is an isomorphism Hk (M, Z) ∼
= H n−k (M, Z).
First we must make sense of cohomology in the sense of Morse complexes. In singular
homology, the chain complex is directly dualised to give singular cohomology. That is,
each Ck is replaced with Ck∗ = HomR (Ck , R), and the boundary maps are replaced with
their adjoints. For Morse homology, it is cleaner to define the Morse cohomology by first
negating the Morse function as follows:
Let f be a Morse function on M . Then the index k critical points of f are precisely
the index n − k critical points of −f , and −f is itself a Morse function on M . For each k,
this gives an isomorphism
Ck (f, R) → Cn−k (−f, R).
We then define the co-complex C • (f, R) to be the dual complex Cn−k (−f, R)∗ . In summary,
given a coefficient ring R, we have the following diagram where each tk is an isomorphism:
∂ n−k+1
··· C n−k (f, R) C n−k+1 (f, R) ···

tk tk−1

δn−k+1
··· Cn−k (−f, R) Cn−k+1 (−f, R) ···

∂k
··· Ck (f, R) Ck−1 (f, R) ···

35
To prove theorem 4.3.2, it remains to show that t is a chain map (in each of the two cases).
Let X be a pseudo-gradient adapted to f , satisfying the Smale condition. The key is
to compare the number of trajectories mod 2 for the mod 2 homology case; nX (b, a) to
n−X (a, b), and signed trajectories NX (b, a) to N−X (a, b) for the integral homology case.
Here we simply discuss the counting in the integral case, as mod 2 immediately follows.
Recall that NX (b, a) is a signed count. Given any trajectory γ ∈ LX (b, a), its sign is
determined by the orientation of Tγ(t) MX (b, a), which is the difference of the orientations
of Tγ(t) W s (b) and Tγ(t) W s (a). By assuming that M is oriented, orientations of stable
manifolds induce orientations of unstable manifolds. In particular, the short exact sequence

0 → Tγ(t) M−X (a, b) → Tγ(t) W u (a) → Tγ(t) W u (b) → 0

induces an orientation on Tγ(t) M−X (a, b). Moreover, the induced orientation did not de-
pend on γ at any point, and the signed count of trajectories NX (b, a) agrees with N−X (a, b).
More explicitly, one can show that the orientations of Tγ LX (b, a) and Tγ L−X (a, b) agree
by “orientation chasing” the following diagram:

R Tγ(t) MX (b, a) Tγ LX (b, a)

Tγ(t) W u (b) Tγ(t) M Tγ(t) W s (b)

Tγ(t) W u (a) Tγ(t) M Tγ(t) W s (a)

Tγ L−X (a, b) Tγ(t) M−X (a, b) R

We are now ready to prove the celebrated Poincaré duality.

Proof of theorem 4.3.2. As remarked earlier, it suffices to show that t is a chain map. That
is, we must show that for any k,

tk−1 ◦ ∂k = ∂ n−k+1 ◦ tk .

To this end, fix b ∈ Ck (f ) and a ∈ Cn−k+1 (−f ). On one hand, we have

 X 
((tk−1 ◦ ∂k )(b))(a) = tk−1 NX (b, ai )ai (a)
ai ∈Critk−1 (f )
X
= NX (b, ai )a∗i (a) = N (b, a).
ai ∈Critk−1 (f )

36
On the other hand,

((∂ n−k+1 ◦ tk )(b))(a) = ∂ n−k+1 (b∗ )(a)

= (b∗ ◦ δn−k+1 )(a)
X
= b∗ N−X (a, bi )bi = N−X (a, b).
bi ∈Critk (f )

As shown above, the orientation of M ensures that NX (b, a) = N−X (a, b). Therefore t is a
chain map, so the isomorphisms tk extend to an isomorphism of chain complexes.

37
Chapter 5

The h-cobordism theorem

5.1 Smale’s original proof of the generalised Poincaré con-

jecture
The Poincaré conjecture states that every simply connected, closed 3-manifold is homeo-
morphic to the 3-sphere. This was open for a long time and was even a Millennium problem,
but was settled in 2006 by Grigori Perelman. The proof was analytic, following Hamilton’s
Ricci flow programme.
An equivalent statement to the Poincaré conjecture is that every 3-manifold with the
homotopy type of the 3-sphere is homeomorphic to the 3-sphere. To see this, we prove the
following proposition.

Proposition 5.1.1. Let M be a closed simply connected 3-manifold. Then M is a homo-

topy 3-sphere.

Proof. By Whitehead’s theorem, if there is a map f : S3 → M inducing isomorphisms

on all homotopy groups, then f is a homotopy equivalence. Therefore we wish to find
such a map. But S3 and M are simply connected, so by the relative form of Hurewicz’s
theorem, any f : S3 → M inducing isomorphisms on homology will induce isomorphisms
on homotopy. Therefore to prove the proposition, it suffices to find a map f : S3 → M
inducing isomorphisms on homology.
Since M is simply connected, H1 (M ) is trivial, so H2 (M ) is also trivial by Poincaré
duality. Since M is 2-connected, Hurewicz’s theorem applies: π3 (M ) ∼= H3 (M ) ∼= Z. Thus
3
let f : S → M be any generator of π3 (M ). Then f induces isomorphisms on all homology
groups as required. Therefore M is a homotopy 3-sphere.

This gives a version of the Poincaré conjecture that easily generalises to higher dimen-
sions:

38
Conjecture 5.1.2. Let an n-manifold M be a homotopy n-sphere. Then M is homeo-
morphic to Sn .
As of 2006, this conjecture is settled for all dimensions. Dimensions n ∈ {0, 1, 2} are
trivial by the classification of n-manifolds for small n. Dimension 3 is the classical case
proven by Perelman. Dimension 4 was solved by Michael Freedman in 1982. Dimensions
5 and above were settled by Stephen Smale in the 1960s.
We now describe Smale’s beautifully simple original proof outline.
Theorem 5.1.3. For n ≥ 5, an n-manifold homotopic to the n-sphere is homeomorphic
to the n-sphere.
Proof. Most of the work goes into proving theorem C (from Smale’s paper [Sma61]).
Let M be an n-manifold homotopic to the n-sphere, with n ≥ 5. Then the Morse
homology of M is isomorphic to that of the n-sphere. In particular, the sum of the ranks
of its homology groups is 2. By the Morse inequalities, any Morse function on M has
at least two critical points. By Theorem C this inequality is sharp: M admits a Morse
function with two critical points. By Reeb’s theorem, M is therefore homeomorphic to an
n-sphere.

5.2 Proving the Poincaré conjecture from the h-cobordism

theorem
Remark. In this section we consider manifolds that are not smooth. (This is the only
section of these notes in which we do this.) Therefore we explicitly say smooth manifold if
a manifold is equipped with a smooth structure.
Soon after the proof that was geared towards the generalised Poincaré conjecture, Smale
observed that his results could be made a lot more general by stating them in terms of
cobordisms.
Definition 5.2.1. A cobordism is a compact manifold with boundary W whose boundary
decomposes as ∂W = V0 tV1 , where V0 and V1 are themselves embedded smooth manifolds.
An example of a cobordism is a pair of pants with a mysterious hole as in figure 5.1.
Definition 5.2.2. Two n-manifolds V0 , V1 (without boundary) are said to be cobordant if
there is a cobordism W such that the boundary of W is the disjoint union of V1 and V2 .
Observe that any cobordism W is equipped with natural inclusion maps ι0 : V0 ,→
W, ι1 : V1 ,→ W .
Definition 5.2.3. A cobordism W between V0 and V1 is said to be an h-cobordism if ι0 , ι1
are homotopy equivalences. The h stands for homotopy.

39
 V1
V0

Figure 5.1: Example of a cobordism.

Example. The cobordism in figure 5.1 is clearly not an h-cobordism, since V0 and V1 do
not have the same homotopy type. However, the cylinder S1 × [0, 1] is a cobordism between
two circles which is easily seen to be an h-cobordism.

Theorem 5.2.4 (h-cobordism theorem). Let n be at least 6, and W a compact n-dimensional

simply connected smooth h-cobordism between simply connected smooth (n − 1)-manifolds
V0 and V1 . Then W is diffeomorphic to V0 × [0, 1].

In the next section, we will gain some insight as to why it is necessary that the dimension
of W be at least 6. For now we assume that the h-cobordism theorem holds, and prove
the Poincaré conjecture for dimensions at least 6.

Theorem 5.2.5 (Smooth Poincaré conjecture, n ≥ 6). For n ≥ 6, a smooth n-manifold

homotopic to the n-sphere is homeomorphic to the n-sphere.

Proof. Assume the h-cobordism theorem. The proof follows figure 5.2.
Suppose M is a smooth n-manifold (n ≥ 6) with the homotopy type of an n-sphere.
Any two distinct points are contained in disjoint disks D0n and D1n . By cutting along the
boundary of the disks, we obtain a decomposition of M as shown in figure 5.2. Precisely,
we write M = D0n ∪ W ∪ D1n , where W = M \ int(D0n t D1n ).
Observe that W is a cobordism between spheres Sn−1 0 and S1n−1 . We prove in the
subsequent lemma that ι0 : Sn−10 ,→ W is a homotopy equivalence. (The same result holds
for Sn−1
1 ). Therefore by the h-cobordism theorem, W is diffeomorphic (and in particular
homeomorphic) to S n−1 × [0, 1], with the homeomorphism denoted by f in the figure.
f restricts to homeomorphisms on the boundary, e.g. g0 : S0n−1 → Sn−1 as shown in the
figure. But any homeomorphism of a sphere induces a homeomorphism of disks D0n → Dn
by the Alexander trick. (One can simply take the radial extension of the homeomorphism.)
Therefore we have homeomorphisms g0 , g1 : D0n , D1n → Dn which agree with f on overlaps.
The map M → Dn ∪ (Sn−1 × [0, 1]) ∪ Dn ∼ = Sn defined piecewise by g0 , f , and g is therefore
a homeomorphism.

40
 g0
D0n

ι0
Sn−1
0

M Sn
f
W
Sn−1 × [0, 1]
Sn−1
1

D1n

Figure 5.2: Proof of the Poincaré conjecture (in dimensions at least 6).

In the above proof we left out a crucial and (in my opinion subtle) step. We wish to
argue that the inclusion ι0 factors through Sn−1 × [0, 1] via f , so ι0 is a composition of
two homotopy equivalences and hence a homotopy equivalence. This is not true since f
will not preserve injectivity. To conclude that ι0 is a homotopy equivalence, we use the
Excision theorem for homology.
Lemma 5.2.6. Let M be a topological n-manifold homotopic to Sn . Let W be as in figure
5.2, i.e. a cobordism between Sn−1
0 and S1n−1 obtained by removing the interiors of two
n-disks. Then Sn−1
0 ,→ W is a homotopy equivalence.
Proof. Both W and Sn−1 0 are simply connected. Therefore (as remarked at the start of
this chapter) the following version of Whitehead’s theorem holds: Any map ι : S0n−1 → W
inducing isomorphisms on homology is a homotopy equivalence. Therefore we prove that
the inclusion map induces isomorphisms on homology.
For notational brevity we hereafter write S0 , S1 , D0 , and D1 to denote S0n−1 and so on.
Fix j ≥ 0. To prevent complications when j = 0, in this argument H denotes the reduced
homology. By the homology long exact sequence, it suffices to show that Hj (W, S0 ) = 0. By
the excision theorem, Hj (W, S0 ) ∼= Hj (W ∪ D0 , D0 ). By the homology long exact sequence
of Hj (W ∪ D0 , D0 ), since Hj (D0 ) is trivial, Hj (W ∪ D0 , D0 ) ∼
= Hj (W ∪ D0 ). Again by the
∼
excision theorem, Hj (M, W ∪ D0 ) = Hj (D1 , S1 ). By the homology long exact sequence,
∼
= Hj (D1 , S1 ) ∼
= Hj (Sn ) ∼
= Hj (M ). But with Hj (M, W ∪ D0 ) isomorphic to Hj (M ), the
homology long exact sequence implies that Hj (W ∪ D0 ) is trivial, and hence Hj (W, S0 ) is
trivial.

41
5.3 Proof outline of the h-cobordism theorem
We now give an overview of the proof of the (smooth) h-cobordism theorem, following
Milnor’s famous lecture notes [Mil65].
Definition 5.3.1. The Morse number µ(M ) of a manifold M is the minimum of the
number of critical points of a Morse function on M .
By the Morse inequalities, for closed manifolds this is bounded below by the topological
complexity (homology groups) of M , and is always at least 2. For a compact manifold with
boundary, the global extrema need not be critical points (in the sense of having vanishing
derivative), so manifolds with boundary may have Morse number 0.
Theorem 5.3.2. Let (W ; V0 , V1 ) be a cobordism. If µ(W ) = 0, then W is a product
cobordism, i.e. W ∼
= V0 × [0, 1].
The above theorem is related to a theorem from section two, in which we showed that
moving from one sub-level set M a to another M b doesn’t change the diffeomorphism class
provided there are no critical points in f −1 ([0, 1]). The proof is also similar. We make use
of this theorem to prove the h-cobordism theorem, by showing that for sufficiently high
dimensions, the Morse number of a simply connected cobordism between simply connected
manifolds is zero. More practically, the goal is to start with a Morse function on W , and
continue to modify it to eliminate critical points.
Some motivation for how we might eliminate critical points is the observation that
critical points of index λ and λ + 1 might cancel out, as in 5.3. In this example, we see

0 a b c 1

Figure 5.3: Example of cancellation.

that the sublevel set at a and c are diffeomorphic, even though they are not diffeomorphic
to the sublevel set at b. The idea is that the index 1 critical point between a and b has
been cancelled by the index 2 critical point between b and c. A precise statement for when
this occurs is the first cancellation theorem:

42
Theorem 5.3.3 (First cancellation theorem). Suppose W is a cobordism equipped with a
Morse-Smale pair (f, X) with exactly two critical points c and c0 , of index k and k + 1,
such that f (c) < f (c0 ). If L(c0 , c) consists of a single point, then the cobordism is a product
cobordism. If fact, the pseudo-gradient field X can be modified on an arbitrarily small
neighbourhood of the trajectory from c0 to c to produce a new pseudo-gradient field, which
corresponds to a new Morse function f 0 on W with no critical points that agrees with f
near ∂W .

In order to make use of this theorem, we must be able to guarantee that L(c0 , c) consists
of a single point along with the hypothesis of critical points occurring in the correct order.
The second of these is the rearrangement theorem, which comes in two forms.

Theorem 5.3.4 (Rearrangement theorem, version one). Any cobordism W of dimension

n can be expressed as a composition of cobordisms W = U0 U1 · · · Un , where each cobordism
Uk admits a Morse function with only one critical level, and all critical points of index k.

This can be phrased without reference to decompositions of cobordisms, and instead in

terms of self-indexing Morse functions.

Theorem 5.3.5 (Rearrangement theorem, version two). Any cobordism (W, V0 , V1 ) can be
equipped with a self-indexing Morse function. Explicitly, this means a Morse function f
satisfying

• f (V0 ) = −1/2, f (V1 ) = n + 1/2.

• f (c) = ind(c), for each critical point c.

We are now in good shape, as all that remains to apply the first cancellation theorem is
to show that under certain conditions, L(c0 , c) is a singleton. Unfortunately this is a very
difficult condition to guarantee. This is where the condition of W being simply connected
and has dimension at least 6 becomes necessary.

Theorem 5.3.6 (Second cancellation theorem). Suppose (W, V0 , V1 ) is a cobordism with

W, V0 , V1 simply connected and dim W = n. Suppose W is equipped with a Morse-Smale
pair (f, X) with exactly two critical points, of index λ and λ + 1. If 3 ≤ λ ≤ n − 3, and
the signed count NX (c0 , c) (as used in the Morse homology) is ±1, then W ∼ = V0 × [0, 1].
More explicitly, given these hypotheses, the pseudo-gradient field of f can be altered between
the critical points so that the un-signed count is 1 (so that the first cancellation theorem
applies).

Observe that a corollary of this theorem is that the analogous result holds with 2 ≤
λ ≤ n − 4, by replacing f with −f . In each case, we observe that n must be at least 6 for
the result to hold. We use the second cancellation theorem to eliminate critical points “of
middle index”.

43
Theorem 5.3.7 (Elimination of critical points of middle index). Suppose (W ; V0 , V1 ) is a
cobordism with dim W = n ≥ 6, equipped with a Morse-Smale pair (f, X) with no critical
points of index 0, 1, n − 1, or n. Assume moreover that W, V0 , and V1 are simply connected,
and H∗ (W, V0 ; Z) = 0. Then (W ; V0 , V1 ) is a product cobordism.
Proof. We give a proof outline, using the Morse complex along with the rearrangement
and second cancellation theorems. The idea is that we use the rearrangement theorem to
decompose W as
W = U2 U3 · · · Un−2 ,
where each Uk consists only of critical points of index k, all with the same level. Now
perturb two critical points c and c0 , of index k and k + 1 respectively, so that
Uk Uk+1 ∼ = U 0 Uc Uc0 U 0 ,
k k+1

where Uc and Uc0 consist of single critical points, c and c0 . Now the condition on homology
ensures exactness of the Morse complex, from which we can conclude that the intersection
number NX (c0 , c) is ±1 (by inspecting the definition of the boundary map). By applying
the second cancellation theorem to Uc Uc0 , we conclude that Uc Uc0 is a product cobordism.
Inductively we conclude that W is a product cobordism.
It now remains to prove that critical points of low and high index can be eliminated.
Again by replacing f with −f (i.e. Poincaré duality), it suffices to eliminate critical points
of middle index.
Theorem 5.3.8 (Elimination of critical points of low index). Consider a cobordism (W ; V0 , V1 ),
with dim W = n, equipped with a self-indexing Morse-Smale pair (f, X). The following
hold:
• If H0 (W, V0 ; Z) = 0, then critical points of index 0 can be cancelled against an equal
number of critical points of index 1.
• Suppose W, V0 are simply connected, and n ≥ 5. If there are no critical points of
index 0, one can insert an auxiliary critical point of index 2 and 3 for each critical
point of index 1, so that the critical point of index 2 cancels the critical point of index
1. (Effectively, one can trade critical points of index 1 for those of index 3.)
Once this theorem is established, the h-cobordism theorem is almost immediate:
Theorem 5.3.9 (h-cobordism theorem). Suppose (W ; V0 , V1 ) is an h-cobordism with W, V0 , V1
simply connected, and dim W ≥ 6. Then W is a product cobordism.
Proof. Since W is an h-cobordism, H∗ (W, V0 ; Z) vanishes. Equip W with a self-indexing
Morse-Smale pair (f, X). By the elimination of critical points of low index, we first elim-
inate all critical points of index 0. If there are any critical points of index 1, these are
traded for critical points of index 3. Replacing f with −f , we similarly eliminate critical
points of index n and n − 1. Finally by elimination of critical points of middle index, the
desired result follows.

44
5.4 Poincaré conjecture and h-cobordism theorem in differ-
ent categories
It is interesting to study the Poincaré conjecture in different categories. We have noted that
the Poincaré conjecture holds for smooth manifolds, in the sense that a smooth homotopy
n-sphere is homeomorphic to an n-sphere. However, we have not discussed the conjecture
strictly in the category of topological, piecewise linear or smooth manifolds.

Definition 5.4.1. A piecewise linear (PL) manifold (also called a combinatorial manifold )
is a topological manifold equipped with a PL structure. That is, the transition maps are
piece-wise linear. Details are given here: [Bry01].

We hereafter denote the category of topological manifolds by Man, PL manifolds by

ManPL , smooth manifolds by Man∞ , and analytic manifolds by Manω . Precisely, the
objects in these categories are the corresponding types of manifolds, and the morphisms
are continuous, PL, smooth maps, and analytic maps respectively.

Theorem 5.4.2. Man∞ and Manω are equivalent. More precisely, every smooth manifold
admits a unique analytic structure.

This result is due to Grauert and Morrey, see [Shi64]. It follows that the validity of the
Poincaré conjecture or h-cobordism theorems in Manω is the same as that in Man∞ , and
for this reason we no longer mention Manω (until the end of the section where results are
summarised).

Theorem 5.4.3. The h-cobordism theorem (for dimensions at least 6) holds in Man, ManPL ,
and Man∞ .

Proof. The h-cobordism theorem in Man∞ is the subject of this chapter, and is the cele-
brated result proven in [Mil65]. The h-cobordism theorem holds in ManPL , with a proof
given in [RS82]. Finally the h-cobordism theorem holds in Man as discussed in [KS77].

Corollary 5.4.4. The Poincaré conjecture for dimensions at least 6 holds in ManPL and
Man.

Proof. Recall the proof that smooth homotopy n-spheres are homeomorphic to n-spheres.
Since the h-cobordism theorem holds in Man, an identical proof applies. For ManPL , the
key is that the Alexander trick still holds (as radially extending a piecewise linear map
gives a piecewise linear map).
However, this proof does not hold in Man∞ , as the centre of the disk is manifestly
singular.

Theorem 5.4.5. In dimension 5, the h-cobordism theorem holds in Man but not in Man∞
or ManPL .

45
Proof. The affirmative result due to Freedman is discussed in his book, [FQ90]. Similarly
the negative result due to Donaldson is discussed in [DK97]. (Moreover, in dimension 4,
ManPL and Man∞ are equivalent.)

Theorem 5.4.6. In dimensions at most 3, Man, ManPL , and Man∞ are equivalent.

More precisely, this follows from the result that every topological manifold of dimension
at most 3 admits a smooth triangulation, unique up to diffeomorphism. (This is due to
Bing and Moise, [Moi77].) By Perelman’s proof of the Poincaré conjecture for dimension 3
in Man∞ , it follows that the Poincaré conjecture holds in all three categories. Similarly it
holds in all three categories in lower dimensions, and likewise for the h-cobordism theorem.
By a theorem of Whitehead, every smooth manifold is canonically a PL manifold.
Moreover, every topological n-sphere admits a smooth structure. Therefore if a topolog-
ical n-sphere admits a unique smooth structure, the Poincaré conjecture in Man implies
the result for both ManPL and Man∞ . On the other hand, whenever multiple smooth
structures exist on the n-sphere, the Poincaré conjecture is false in Man∞ for dimension
n. It is conjectured that all spheres beyond dimension 61 admit at least one exotic smooth
structure, which would guarantee that the smooth Poincaré conjecture is always false in
dimensions above 61 (See [WX17]).
Since the 5-sphere has a unique smooth structure, the above theorem proves the 5-
dimensional Poincaré conjecture in all three categories in consideration, even though the
h-cobordism theorem fails in two of them.
In dimension 4, the Poincaré conjecture and h-cobordism theorem are equivalent to
each other (in each of Man, ManPL , and Man∞ ). In particular in ManPL and Man∞ ,
they are equivalent to the existence of an exotic smooth structure on the 4-sphere. This is
a difficult problem, and is still open.

Proposition 5.4.7. Fix a category C consisting of 4-manifolds in either Man, ManPL ,

or Man∞ . Then the Poincaré conjecture in C is equivalent to the h-cobordism theorem
in C.

Proof. Since ManPL and Man∞ are equivalent in dimension 4, without loss of generality
suppose C consists of 4-dimensional topological manifolds or PL manifolds.
Suppose the h-cobordism theorem holds in C. The earlier proof of the Poincaré con-
jecture for dimensions at least 6 holds in C, so the Poincaré conjecture holds in C.
Next suppose the Poincaré conjecture holds in C. The proof will proceed in a similar
way to the proof of the Poincaré conjecture assuming the h-cobordism theorem. Let W
be a 4 dimensional compact simply connected h-cobordism between simply connected 3-
manifolds V0 and V1 . Then V0 and V1 are necessarily closed and simply connected, so by
the Poincaré conjecture in dimension 3, they are C-isomorphic to 3-spheres. It follows
that there are 4-cells D0 and D1 that can be glued along V0 and V1 respectively; let
M = D0 tV0 W tV1 D1 . By Whitehead’s theorem, if we can prove that M has the same

46
homology as a 4-sphere, then M is homotopy equivalent to the 4-sphere. By the Poincaré
conjecture in C, it will follow that M is C-isomorphic to the 4-sphere. Therefore by
removing V0 and V1 , the h-cobordism theorem will follow.
It remains to prove that for each i, Hi (M ) = Hi (S4 ). Here H denotes the reduced
homology, to avoid a complication when i = 1. By the long exact sequence of homology,
Hi (M ) ∼= Hi (M, D1 ). By excision, the latter is isomorphic to Hi (W ∪ D0 , V1 ). Next
observe that Hi (W ∪ D0 ) ∼
= Hi (W ∪ D0 , D0 ) again by the long exact sequence of homology.
Another application of excision gives Hi (W ∪ D0 , D0 ) ∼
= Hi (W, V0 ). By virtue of W being
an h-cobordism, Hi (W, V0 ) must vanish. This establishes that Hi (W ∪ D0 ) vanishes, so
by the long exact sequence of homology, we find that Hi+1 (M ) = Hi (V1 ) = Hi+1 (S4 ) as
required.

Finally, recall that Freedman famously proved the Poincaré conjecture for Man in
dimension 4, for which he was awarded a Field’s medal. In summary, as of the time of
writing (the era of COVID-19 lockdowns), the status of the h-cobordism theorem and
Poincaré conjecture in various dimensions is as follows:

Poincaré conjecture h-cobordism theorem

Dim Man ManPL Man∞ , Manω Dim Man ManPL Man∞ , Manω
1,2,3 True True True 1,2,3 True True True
4 True Open Open 4 True Open Open
5,6 True True True 5 True False False
7+ True True Usually false 6+ True True True

Finally we remark that it is not particularly interesting to consider this question in the
categories of almost-complex or complex manifolds, since the the only spheres that exist
in these categories are S2 and S6 for the former, and S2 for the latter. (It is not currently
known if S6 admits a complex structure.) Exposition on almost complex spheres and
complex spheres can be found in [KP17].

47
References

[AD14] Michèle Audin and Mihai Damian. Morse Theory and Floer Homology. Universi-
text. Springer-Verlag London, 2014.
[Bry01] J. Bryant. Piecewise linear topology. Handbook of Geometric Topology, pages
219–259, 12 2001.
[DK97] S.K. Donaldson and P.B. Kronheimer. The Geometry of four-manifolds. Oxford
mathematical monographs. Oxford press, 1997.
[FQ90] Michael H. Freedman and Frank Quinn. Topology of 4-Manifolds. Princeton
Mathematical Seires. Princeton University Press, 1990.
[KP17] Panagiotis Konstantis and Maurizio Parton. Almost complex structures on
spheres. Differential Geometry and its Applications, 07 2017.
[KS77] Robion C. Kirby and Laurence C. Siebenmann. Foundational Essays on Topolog-
ical Manifolds, Smoothings, and Triangulations. Annals of Mathematics Studies.
Princeton University Press, 1977.
[Mil65] John Milnor. Lectures on the H-Cobordism Theorem. Princeton Legacy Library.
Princeton University Press, 1965.
[Moi77] Edwin E. Moise. Geometric Topology in Dimensions 2 and 3. Graduate Texts in
Mathematics. Springer-Verlag New York, 1977.
[RS82] Colin P. Rourke and B.J. Sanderson. Introduction to Piecewise-Linear Topology.
Springer-Verlag Berlin Heidelberg, 1982.
[Shi64] Kôji Shiga. Some aspects of real-analytic manifolds and differentiable manifolds.
J. Math. Soc. Japan, 16(2):128–142, 04 1964.
[Sma61] Stephen Smale. Generalized poincaré’s conjecture in dimensions greater than four.
Annals of Mathematics, 74(2):391–406, 1961.
[WX17] Guozhen Wang and Zhouli Xu. The triviality of the 61-stem in the stable homo-
topy groups of spheres. Annals of Mathematics, 186(2):501–580, 2017.

48